Adding upstream version 5.10.209.upstream/5.10.209

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

15 files changed, 17213 insertions, 0 deletions

diff --git a/drivers/infiniband/hw/mlx4/Kconfig b/drivers/infiniband/hw/mlx4/Kconfig
new file mode 100644
index 000000000..f30ce9dd0
--- /dev/null
+++ b/drivers/infiniband/hw/mlx4/Kconfig
@@ -0,0 +1,11 @@
+# SPDX-License-Identifier: GPL-2.0-only
+config MLX4_INFINIBAND
+	tristate "Mellanox ConnectX HCA support"
+	depends on NETDEVICES && ETHERNET && PCI && INET
+	select NET_VENDOR_MELLANOX
+	select MLX4_CORE
+	help
+	  This driver provides low-level InfiniBand support for
+	  Mellanox ConnectX PCI Express host channel adapters (HCAs).
+	  This is required to use InfiniBand protocols such as
+	  IP-over-IB or SRP with these devices.
diff --git a/drivers/infiniband/hw/mlx4/Makefile b/drivers/infiniband/hw/mlx4/Makefile
new file mode 100644
index 000000000..7b6757b02
--- /dev/null
+++ b/drivers/infiniband/hw/mlx4/Makefile
@@ -0,0 +1,4 @@
+# SPDX-License-Identifier: GPL-2.0-only
+obj-$(CONFIG_MLX4_INFINIBAND)	+= mlx4_ib.o
+
+mlx4_ib-y :=	ah.o cq.o doorbell.o mad.o main.o mr.o qp.o srq.o mcg.o cm.o alias_GUID.o sysfs.o
diff --git a/drivers/infiniband/hw/mlx4/ah.c b/drivers/infiniband/hw/mlx4/ah.c
new file mode 100644
index 000000000..7321d6ab5
--- /dev/null
+++ b/drivers/infiniband/hw/mlx4/ah.c
@@ -0,0 +1,234 @@
+/*
+ * Copyright (c) 2007 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
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <rdma/ib_addr.h>
+#include <rdma/ib_cache.h>
+
+#include <linux/slab.h>
+#include <linux/inet.h>
+#include <linux/string.h>
+#include <linux/mlx4/driver.h>
+
+#include "mlx4_ib.h"
+
+static void create_ib_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr)
+{
+	struct mlx4_ib_ah *ah = to_mah(ib_ah);
+	struct mlx4_dev *dev = to_mdev(ib_ah->device)->dev;
+
+	ah->av.ib.port_pd = cpu_to_be32(to_mpd(ib_ah->pd)->pdn |
+			    (rdma_ah_get_port_num(ah_attr) << 24));
+	ah->av.ib.g_slid  = rdma_ah_get_path_bits(ah_attr);
+	ah->av.ib.sl_tclass_flowlabel =
+			cpu_to_be32(rdma_ah_get_sl(ah_attr) << 28);
+	if (rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) {
+		const struct ib_global_route *grh = rdma_ah_read_grh(ah_attr);
+
+		ah->av.ib.g_slid   |= 0x80;
+		ah->av.ib.gid_index = grh->sgid_index;
+		ah->av.ib.hop_limit = grh->hop_limit;
+		ah->av.ib.sl_tclass_flowlabel |=
+			cpu_to_be32((grh->traffic_class << 20) |
+				    grh->flow_label);
+		memcpy(ah->av.ib.dgid, grh->dgid.raw, 16);
+	}
+
+	ah->av.ib.dlid = cpu_to_be16(rdma_ah_get_dlid(ah_attr));
+	if (rdma_ah_get_static_rate(ah_attr)) {
+		u8 static_rate = rdma_ah_get_static_rate(ah_attr) +
+					MLX4_STAT_RATE_OFFSET;
+
+		while (static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
+		       !(1 << static_rate & dev->caps.stat_rate_support))
+			--static_rate;
+		ah->av.ib.stat_rate = static_rate;
+	}
+}
+
+static int create_iboe_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr)
+{
+	struct mlx4_ib_dev *ibdev = to_mdev(ib_ah->device);
+	struct mlx4_ib_ah *ah = to_mah(ib_ah);
+	const struct ib_gid_attr *gid_attr;
+	struct mlx4_dev *dev = ibdev->dev;
+	int is_mcast = 0;
+	struct in6_addr in6;
+	u16 vlan_tag = 0xffff;
+	const struct ib_global_route *grh = rdma_ah_read_grh(ah_attr);
+	int ret;
+
+	memcpy(&in6, grh->dgid.raw, sizeof(in6));
+	if (rdma_is_multicast_addr(&in6))
+		is_mcast = 1;
+
+	memcpy(ah->av.eth.mac, ah_attr->roce.dmac, ETH_ALEN);
+	eth_zero_addr(ah->av.eth.s_mac);
+
+	/*
+	 * If sgid_attr is NULL we are being called by mlx4_ib_create_ah_slave
+	 * and we are directly creating an AV for a slave's gid_index.
+	 */
+	gid_attr = ah_attr->grh.sgid_attr;
+	if (gid_attr) {
+		ret = rdma_read_gid_l2_fields(gid_attr, &vlan_tag,
+					      &ah->av.eth.s_mac[0]);
+		if (ret)
+			return ret;
+
+		ret = mlx4_ib_gid_index_to_real_index(ibdev, gid_attr);
+		if (ret < 0)
+			return ret;
+		ah->av.eth.gid_index = ret;
+	} else {
+		/* mlx4_ib_create_ah_slave fills in the s_mac and the vlan */
+		ah->av.eth.gid_index = ah_attr->grh.sgid_index;
+	}
+
+	if (vlan_tag < 0x1000)
+		vlan_tag |= (rdma_ah_get_sl(ah_attr) & 7) << 13;
+	ah->av.eth.port_pd = cpu_to_be32(to_mpd(ib_ah->pd)->pdn |
+					 (rdma_ah_get_port_num(ah_attr) << 24));
+	ah->av.eth.vlan = cpu_to_be16(vlan_tag);
+	ah->av.eth.hop_limit = grh->hop_limit;
+	if (rdma_ah_get_static_rate(ah_attr)) {
+		ah->av.eth.stat_rate = rdma_ah_get_static_rate(ah_attr) +
+					MLX4_STAT_RATE_OFFSET;
+		while (ah->av.eth.stat_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
+		       !(1 << ah->av.eth.stat_rate & dev->caps.stat_rate_support))
+			--ah->av.eth.stat_rate;
+	}
+	ah->av.eth.sl_tclass_flowlabel |=
+			cpu_to_be32((grh->traffic_class << 20) |
+				    grh->flow_label);
+	/*
+	 * HW requires multicast LID so we just choose one.
+	 */
+	if (is_mcast)
+		ah->av.ib.dlid = cpu_to_be16(0xc000);
+
+	memcpy(ah->av.eth.dgid, grh->dgid.raw, 16);
+	ah->av.eth.sl_tclass_flowlabel |= cpu_to_be32(rdma_ah_get_sl(ah_attr)
+						      << 29);
+	return 0;
+}
+
+int mlx4_ib_create_ah(struct ib_ah *ib_ah, struct rdma_ah_init_attr *init_attr,
+		      struct ib_udata *udata)
+{
+	struct rdma_ah_attr *ah_attr = init_attr->ah_attr;
+
+	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
+		if (!(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH))
+			return -EINVAL;
+		/*
+		 * TBD: need to handle the case when we get
+		 * called in an atomic context and there we
+		 * might sleep.  We don't expect this
+		 * currently since we're working with link
+		 * local addresses which we can translate
+		 * without going to sleep.
+		 */
+		return create_iboe_ah(ib_ah, ah_attr);
+	}
+
+	create_ib_ah(ib_ah, ah_attr);
+	return 0;
+}
+
+int mlx4_ib_create_ah_slave(struct ib_ah *ah, struct rdma_ah_attr *ah_attr,
+			    int slave_sgid_index, u8 *s_mac, u16 vlan_tag)
+{
+	struct rdma_ah_attr slave_attr = *ah_attr;
+	struct rdma_ah_init_attr init_attr = {};
+	struct mlx4_ib_ah *mah = to_mah(ah);
+	int ret;
+
+	slave_attr.grh.sgid_attr = NULL;
+	slave_attr.grh.sgid_index = slave_sgid_index;
+	init_attr.ah_attr = &slave_attr;
+	ret = mlx4_ib_create_ah(ah, &init_attr, NULL);
+	if (ret)
+		return ret;
+
+	ah->type = ah_attr->type;
+
+	/* get rid of force-loopback bit */
+	mah->av.ib.port_pd &= cpu_to_be32(0x7FFFFFFF);
+
+	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE)
+		memcpy(mah->av.eth.s_mac, s_mac, 6);
+
+	if (vlan_tag < 0x1000)
+		vlan_tag |= (rdma_ah_get_sl(ah_attr) & 7) << 13;
+	mah->av.eth.vlan = cpu_to_be16(vlan_tag);
+
+	return 0;
+}
+
+int mlx4_ib_query_ah(struct ib_ah *ibah, struct rdma_ah_attr *ah_attr)
+{
+	struct mlx4_ib_ah *ah = to_mah(ibah);
+	int port_num = be32_to_cpu(ah->av.ib.port_pd) >> 24;
+
+	memset(ah_attr, 0, sizeof *ah_attr);
+	ah_attr->type = ibah->type;
+
+	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
+		rdma_ah_set_dlid(ah_attr, 0);
+		rdma_ah_set_sl(ah_attr,
+			       be32_to_cpu(ah->av.eth.sl_tclass_flowlabel)
+			       >> 29);
+	} else {
+		rdma_ah_set_dlid(ah_attr, be16_to_cpu(ah->av.ib.dlid));
+		rdma_ah_set_sl(ah_attr,
+			       be32_to_cpu(ah->av.ib.sl_tclass_flowlabel)
+			       >> 28);
+	}
+
+	rdma_ah_set_port_num(ah_attr, port_num);
+	if (ah->av.ib.stat_rate)
+		rdma_ah_set_static_rate(ah_attr,
+					ah->av.ib.stat_rate -
+					MLX4_STAT_RATE_OFFSET);
+	rdma_ah_set_path_bits(ah_attr, ah->av.ib.g_slid & 0x7F);
+	if (mlx4_ib_ah_grh_present(ah)) {
+		u32 tc_fl = be32_to_cpu(ah->av.ib.sl_tclass_flowlabel);
+
+		rdma_ah_set_grh(ah_attr, NULL,
+				tc_fl & 0xfffff, ah->av.ib.gid_index,
+				ah->av.ib.hop_limit,
+				tc_fl >> 20);
+		rdma_ah_set_dgid_raw(ah_attr, ah->av.ib.dgid);
+	}
+
+	return 0;
+}
diff --git a/drivers/infiniband/hw/mlx4/alias_GUID.c b/drivers/infiniband/hw/mlx4/alias_GUID.c
new file mode 100644
index 000000000..cca414ecf
--- /dev/null
+++ b/drivers/infiniband/hw/mlx4/alias_GUID.c
@@ -0,0 +1,905 @@
+/*
+ * Copyright (c) 2012 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+ /***********************************************************/
+/*This file support the handling of the Alias GUID feature. */
+/***********************************************************/
+#include <rdma/ib_mad.h>
+#include <rdma/ib_smi.h>
+#include <rdma/ib_cache.h>
+#include <rdma/ib_sa.h>
+#include <rdma/ib_pack.h>
+#include <linux/mlx4/cmd.h>
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/errno.h>
+#include <rdma/ib_user_verbs.h>
+#include <linux/delay.h>
+#include "mlx4_ib.h"
+
+/*
+The driver keeps the current state of all guids, as they are in the HW.
+Whenever we receive an smp mad GUIDInfo record, the data will be cached.
+*/
+
+struct mlx4_alias_guid_work_context {
+	u8 port;
+	struct mlx4_ib_dev     *dev ;
+	struct ib_sa_query     *sa_query;
+	struct completion	done;
+	int			query_id;
+	struct list_head	list;
+	int			block_num;
+	ib_sa_comp_mask		guid_indexes;
+	u8			method;
+};
+
+struct mlx4_next_alias_guid_work {
+	u8 port;
+	u8 block_num;
+	u8 method;
+	struct mlx4_sriov_alias_guid_info_rec_det rec_det;
+};
+
+static int get_low_record_time_index(struct mlx4_ib_dev *dev, u8 port,
+				     int *resched_delay_sec);
+
+void mlx4_ib_update_cache_on_guid_change(struct mlx4_ib_dev *dev, int block_num,
+					 u8 port_num, u8 *p_data)
+{
+	int i;
+	u64 guid_indexes;
+	int slave_id;
+	int port_index = port_num - 1;
+
+	if (!mlx4_is_master(dev->dev))
+		return;
+
+	guid_indexes = be64_to_cpu((__force __be64) dev->sriov.alias_guid.
+				   ports_guid[port_num - 1].
+				   all_rec_per_port[block_num].guid_indexes);
+	pr_debug("port: %d, guid_indexes: 0x%llx\n", port_num, guid_indexes);
+
+	for (i = 0; i < NUM_ALIAS_GUID_IN_REC; i++) {
+		/* The location of the specific index starts from bit number 4
+		 * until bit num 11 */
+		if (test_bit(i + 4, (unsigned long *)&guid_indexes)) {
+			slave_id = (block_num * NUM_ALIAS_GUID_IN_REC) + i ;
+			if (slave_id >= dev->dev->num_slaves) {
+				pr_debug("The last slave: %d\n", slave_id);
+				return;
+			}
+
+			/* cache the guid: */
+			memcpy(&dev->sriov.demux[port_index].guid_cache[slave_id],
+			       &p_data[i * GUID_REC_SIZE],
+			       GUID_REC_SIZE);
+		} else
+			pr_debug("Guid number: %d in block: %d"
+				 " was not updated\n", i, block_num);
+	}
+}
+
+static __be64 get_cached_alias_guid(struct mlx4_ib_dev *dev, int port, int index)
+{
+	if (index >= NUM_ALIAS_GUID_PER_PORT) {
+		pr_err("%s: ERROR: asked for index:%d\n", __func__, index);
+		return (__force __be64) -1;
+	}
+	return *(__be64 *)&dev->sriov.demux[port - 1].guid_cache[index];
+}
+
+
+ib_sa_comp_mask mlx4_ib_get_aguid_comp_mask_from_ix(int index)
+{
+	return IB_SA_COMP_MASK(4 + index);
+}
+
+void mlx4_ib_slave_alias_guid_event(struct mlx4_ib_dev *dev, int slave,
+				    int port,  int slave_init)
+{
+	__be64 curr_guid, required_guid;
+	int record_num = slave / 8;
+	int index = slave % 8;
+	int port_index = port - 1;
+	unsigned long flags;
+	int do_work = 0;
+
+	spin_lock_irqsave(&dev->sriov.alias_guid.ag_work_lock, flags);
+	if (dev->sriov.alias_guid.ports_guid[port_index].state_flags &
+	    GUID_STATE_NEED_PORT_INIT)
+		goto unlock;
+	if (!slave_init) {
+		curr_guid = *(__be64 *)&dev->sriov.
+			alias_guid.ports_guid[port_index].
+			all_rec_per_port[record_num].
+			all_recs[GUID_REC_SIZE * index];
+		if (curr_guid == cpu_to_be64(MLX4_GUID_FOR_DELETE_VAL) ||
+		    !curr_guid)
+			goto unlock;
+		required_guid = cpu_to_be64(MLX4_GUID_FOR_DELETE_VAL);
+	} else {
+		required_guid = mlx4_get_admin_guid(dev->dev, slave, port);
+		if (required_guid == cpu_to_be64(MLX4_GUID_FOR_DELETE_VAL))
+			goto unlock;
+	}
+	*(__be64 *)&dev->sriov.alias_guid.ports_guid[port_index].
+		all_rec_per_port[record_num].
+		all_recs[GUID_REC_SIZE * index] = required_guid;
+	dev->sriov.alias_guid.ports_guid[port_index].
+		all_rec_per_port[record_num].guid_indexes
+		|= mlx4_ib_get_aguid_comp_mask_from_ix(index);
+	dev->sriov.alias_guid.ports_guid[port_index].
+		all_rec_per_port[record_num].status
+		= MLX4_GUID_INFO_STATUS_IDLE;
+	/* set to run immediately */
+	dev->sriov.alias_guid.ports_guid[port_index].
+		all_rec_per_port[record_num].time_to_run = 0;
+	dev->sriov.alias_guid.ports_guid[port_index].
+		all_rec_per_port[record_num].
+		guids_retry_schedule[index] = 0;
+	do_work = 1;
+unlock:
+	spin_unlock_irqrestore(&dev->sriov.alias_guid.ag_work_lock, flags);
+
+	if (do_work)
+		mlx4_ib_init_alias_guid_work(dev, port_index);
+}
+
+/*
+ * Whenever new GUID is set/unset (guid table change) create event and
+ * notify the relevant slave (master also should be notified).
+ * If the GUID value is not as we have in the cache the slave will not be
+ * updated; in this case it waits for the smp_snoop or the port management
+ * event to call the function and to update the slave.
+ * block_number - the index of the block (16 blocks available)
+ * port_number - 1 or 2
+ */
+void mlx4_ib_notify_slaves_on_guid_change(struct mlx4_ib_dev *dev,
+					  int block_num, u8 port_num,
+					  u8 *p_data)
+{
+	int i;
+	u64 guid_indexes;
+	int slave_id, slave_port;
+	enum slave_port_state new_state;
+	enum slave_port_state prev_state;
+	__be64 tmp_cur_ag, form_cache_ag;
+	enum slave_port_gen_event gen_event;
+	struct mlx4_sriov_alias_guid_info_rec_det *rec;
+	unsigned long flags;
+	__be64 required_value;
+
+	if (!mlx4_is_master(dev->dev))
+		return;
+
+	rec = &dev->sriov.alias_guid.ports_guid[port_num - 1].
+			all_rec_per_port[block_num];
+	guid_indexes = be64_to_cpu((__force __be64) dev->sriov.alias_guid.
+				   ports_guid[port_num - 1].
+				   all_rec_per_port[block_num].guid_indexes);
+	pr_debug("port: %d, guid_indexes: 0x%llx\n", port_num, guid_indexes);
+
+	/*calculate the slaves and notify them*/
+	for (i = 0; i < NUM_ALIAS_GUID_IN_REC; i++) {
+		/* the location of the specific index runs from bits 4..11 */
+		if (!(test_bit(i + 4, (unsigned long *)&guid_indexes)))
+			continue;
+
+		slave_id = (block_num * NUM_ALIAS_GUID_IN_REC) + i ;
+		if (slave_id >= dev->dev->persist->num_vfs + 1)
+			return;
+
+		slave_port = mlx4_phys_to_slave_port(dev->dev, slave_id, port_num);
+		if (slave_port < 0) /* this port isn't available for the VF */
+			continue;
+
+		tmp_cur_ag = *(__be64 *)&p_data[i * GUID_REC_SIZE];
+		form_cache_ag = get_cached_alias_guid(dev, port_num,
+					(NUM_ALIAS_GUID_IN_REC * block_num) + i);
+		/*
+		 * Check if guid is not the same as in the cache,
+		 * If it is different, wait for the snoop_smp or the port mgmt
+		 * change event to update the slave on its port state change
+		 */
+		if (tmp_cur_ag != form_cache_ag)
+			continue;
+
+		spin_lock_irqsave(&dev->sriov.alias_guid.ag_work_lock, flags);
+		required_value = *(__be64 *)&rec->all_recs[i * GUID_REC_SIZE];
+
+		if (required_value == cpu_to_be64(MLX4_GUID_FOR_DELETE_VAL))
+			required_value = 0;
+
+		if (tmp_cur_ag == required_value) {
+			rec->guid_indexes = rec->guid_indexes &
+			       ~mlx4_ib_get_aguid_comp_mask_from_ix(i);
+		} else {
+			/* may notify port down if value is 0 */
+			if (tmp_cur_ag != MLX4_NOT_SET_GUID) {
+				spin_unlock_irqrestore(&dev->sriov.
+					alias_guid.ag_work_lock, flags);
+				continue;
+			}
+		}
+		spin_unlock_irqrestore(&dev->sriov.alias_guid.ag_work_lock,
+				       flags);
+		mlx4_gen_guid_change_eqe(dev->dev, slave_id, port_num);
+		/*2 cases: Valid GUID, and Invalid Guid*/
+
+		if (tmp_cur_ag != MLX4_NOT_SET_GUID) { /*valid GUID*/
+			prev_state = mlx4_get_slave_port_state(dev->dev, slave_id, port_num);
+			new_state = set_and_calc_slave_port_state(dev->dev, slave_id, port_num,
+								  MLX4_PORT_STATE_IB_PORT_STATE_EVENT_GID_VALID,
+								  &gen_event);
+			pr_debug("slave: %d, port: %d prev_port_state: %d,"
+				 " new_port_state: %d, gen_event: %d\n",
+				 slave_id, port_num, prev_state, new_state, gen_event);
+			if (gen_event == SLAVE_PORT_GEN_EVENT_UP) {
+				pr_debug("sending PORT_UP event to slave: %d, port: %d\n",
+					 slave_id, port_num);
+				mlx4_gen_port_state_change_eqe(dev->dev, slave_id,
+							       port_num, MLX4_PORT_CHANGE_SUBTYPE_ACTIVE);
+			}
+		} else { /* request to invalidate GUID */
+			set_and_calc_slave_port_state(dev->dev, slave_id, port_num,
+						      MLX4_PORT_STATE_IB_EVENT_GID_INVALID,
+						      &gen_event);
+			if (gen_event == SLAVE_PORT_GEN_EVENT_DOWN) {
+				pr_debug("sending PORT DOWN event to slave: %d, port: %d\n",
+					 slave_id, port_num);
+				mlx4_gen_port_state_change_eqe(dev->dev,
+							       slave_id,
+							       port_num,
+							       MLX4_PORT_CHANGE_SUBTYPE_DOWN);
+			}
+		}
+	}
+}
+
+static void aliasguid_query_handler(int status,
+				    struct ib_sa_guidinfo_rec *guid_rec,
+				    void *context)
+{
+	struct mlx4_ib_dev *dev;
+	struct mlx4_alias_guid_work_context *cb_ctx = context;
+	u8 port_index ;
+	int i;
+	struct mlx4_sriov_alias_guid_info_rec_det *rec;
+	unsigned long flags, flags1;
+	ib_sa_comp_mask declined_guid_indexes = 0;
+	ib_sa_comp_mask applied_guid_indexes = 0;
+	unsigned int resched_delay_sec = 0;
+
+	if (!context)
+		return;
+
+	dev = cb_ctx->dev;
+	port_index = cb_ctx->port - 1;
+	rec = &dev->sriov.alias_guid.ports_guid[port_index].
+		all_rec_per_port[cb_ctx->block_num];
+
+	if (status) {
+		pr_debug("(port: %d) failed: status = %d\n",
+			 cb_ctx->port, status);
+		rec->time_to_run = ktime_get_boottime_ns() + 1 * NSEC_PER_SEC;
+		goto out;
+	}
+
+	if (guid_rec->block_num != cb_ctx->block_num) {
+		pr_err("block num mismatch: %d != %d\n",
+		       cb_ctx->block_num, guid_rec->block_num);
+		goto out;
+	}
+
+	pr_debug("lid/port: %d/%d, block_num: %d\n",
+		 be16_to_cpu(guid_rec->lid), cb_ctx->port,
+		 guid_rec->block_num);
+
+	rec = &dev->sriov.alias_guid.ports_guid[port_index].
+		all_rec_per_port[guid_rec->block_num];
+
+	spin_lock_irqsave(&dev->sriov.alias_guid.ag_work_lock, flags);
+	for (i = 0 ; i < NUM_ALIAS_GUID_IN_REC; i++) {
+		__be64 sm_response, required_val;
+
+		if (!(cb_ctx->guid_indexes &
+			mlx4_ib_get_aguid_comp_mask_from_ix(i)))
+			continue;
+		sm_response = *(__be64 *)&guid_rec->guid_info_list
+				[i * GUID_REC_SIZE];
+		required_val = *(__be64 *)&rec->all_recs[i * GUID_REC_SIZE];
+		if (cb_ctx->method == MLX4_GUID_INFO_RECORD_DELETE) {
+			if (required_val ==
+			    cpu_to_be64(MLX4_GUID_FOR_DELETE_VAL))
+				goto next_entry;
+
+			/* A new value was set till we got the response */
+			pr_debug("need to set new value %llx, record num %d, block_num:%d\n",
+				 be64_to_cpu(required_val),
+				 i, guid_rec->block_num);
+			goto entry_declined;
+		}
+
+		/* check if the SM didn't assign one of the records.
+		 * if it didn't, re-ask for.
+		 */
+		if (sm_response == MLX4_NOT_SET_GUID) {
+			if (rec->guids_retry_schedule[i] == 0)
+				mlx4_ib_warn(&dev->ib_dev,
+					     "%s:Record num %d in  block_num: %d was declined by SM\n",
+					     __func__, i,
+					     guid_rec->block_num);
+			goto entry_declined;
+		} else {
+		       /* properly assigned record. */
+		       /* We save the GUID we just got from the SM in the
+			* admin_guid in order to be persistent, and in the
+			* request from the sm the process will ask for the same GUID */
+			if (required_val &&
+			    sm_response != required_val) {
+				/* Warn only on first retry */
+				if (rec->guids_retry_schedule[i] == 0)
+					mlx4_ib_warn(&dev->ib_dev, "%s: Failed to set"
+						     " admin guid after SysAdmin "
+						     "configuration. "
+						     "Record num %d in block_num:%d "
+						     "was declined by SM, "
+						     "new val(0x%llx) was kept, SM returned (0x%llx)\n",
+						      __func__, i,
+						     guid_rec->block_num,
+						     be64_to_cpu(required_val),
+						     be64_to_cpu(sm_response));
+				goto entry_declined;
+			} else {
+				*(__be64 *)&rec->all_recs[i * GUID_REC_SIZE] =
+					sm_response;
+				if (required_val == 0)
+					mlx4_set_admin_guid(dev->dev,
+							    sm_response,
+							    (guid_rec->block_num
+							    * NUM_ALIAS_GUID_IN_REC) + i,
+							    cb_ctx->port);
+				goto next_entry;
+			}
+		}
+entry_declined:
+		declined_guid_indexes |= mlx4_ib_get_aguid_comp_mask_from_ix(i);
+		rec->guids_retry_schedule[i] =
+			(rec->guids_retry_schedule[i] == 0) ?  1 :
+			min((unsigned int)60,
+			    rec->guids_retry_schedule[i] * 2);
+		/* using the minimum value among all entries in that record */
+		resched_delay_sec = (resched_delay_sec == 0) ?
+				rec->guids_retry_schedule[i] :
+				min(resched_delay_sec,
+				    rec->guids_retry_schedule[i]);
+		continue;
+
+next_entry:
+		rec->guids_retry_schedule[i] = 0;
+	}
+
+	applied_guid_indexes =  cb_ctx->guid_indexes & ~declined_guid_indexes;
+	if (declined_guid_indexes ||
+	    rec->guid_indexes & ~(applied_guid_indexes)) {
+		pr_debug("record=%d wasn't fully set, guid_indexes=0x%llx applied_indexes=0x%llx, declined_indexes=0x%llx\n",
+			 guid_rec->block_num,
+			 be64_to_cpu((__force __be64)rec->guid_indexes),
+			 be64_to_cpu((__force __be64)applied_guid_indexes),
+			 be64_to_cpu((__force __be64)declined_guid_indexes));
+		rec->time_to_run = ktime_get_boottime_ns() +
+			resched_delay_sec * NSEC_PER_SEC;
+	} else {
+		rec->status = MLX4_GUID_INFO_STATUS_SET;
+	}
+	spin_unlock_irqrestore(&dev->sriov.alias_guid.ag_work_lock, flags);
+	/*
+	The func is call here to close the cases when the
+	sm doesn't send smp, so in the sa response the driver
+	notifies the slave.
+	*/
+	mlx4_ib_notify_slaves_on_guid_change(dev, guid_rec->block_num,
+					     cb_ctx->port,
+					     guid_rec->guid_info_list);
+out:
+	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
+	spin_lock_irqsave(&dev->sriov.alias_guid.ag_work_lock, flags1);
+	if (!dev->sriov.is_going_down) {
+		get_low_record_time_index(dev, port_index, &resched_delay_sec);
+		queue_delayed_work(dev->sriov.alias_guid.ports_guid[port_index].wq,
+				   &dev->sriov.alias_guid.ports_guid[port_index].
+				   alias_guid_work,
+				   msecs_to_jiffies(resched_delay_sec * 1000));
+	}
+	if (cb_ctx->sa_query) {
+		list_del(&cb_ctx->list);
+		kfree(cb_ctx);
+	} else
+		complete(&cb_ctx->done);
+	spin_unlock_irqrestore(&dev->sriov.alias_guid.ag_work_lock, flags1);
+	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
+}
+
+static void invalidate_guid_record(struct mlx4_ib_dev *dev, u8 port, int index)
+{
+	int i;
+	u64 cur_admin_val;
+	ib_sa_comp_mask comp_mask = 0;
+
+	dev->sriov.alias_guid.ports_guid[port - 1].all_rec_per_port[index].status
+		= MLX4_GUID_INFO_STATUS_SET;
+
+	/* calculate the comp_mask for that record.*/
+	for (i = 0; i < NUM_ALIAS_GUID_IN_REC; i++) {
+		cur_admin_val =
+			*(u64 *)&dev->sriov.alias_guid.ports_guid[port - 1].
+			all_rec_per_port[index].all_recs[GUID_REC_SIZE * i];
+		/*
+		check the admin value: if it's for delete (~00LL) or
+		it is the first guid of the first record (hw guid) or
+		the records is not in ownership of the sysadmin and the sm doesn't
+		need to assign GUIDs, then don't put it up for assignment.
+		*/
+		if (MLX4_GUID_FOR_DELETE_VAL == cur_admin_val ||
+		    (!index && !i))
+			continue;
+		comp_mask |= mlx4_ib_get_aguid_comp_mask_from_ix(i);
+	}
+	dev->sriov.alias_guid.ports_guid[port - 1].
+		all_rec_per_port[index].guid_indexes |= comp_mask;
+	if (dev->sriov.alias_guid.ports_guid[port - 1].
+	    all_rec_per_port[index].guid_indexes)
+		dev->sriov.alias_guid.ports_guid[port - 1].
+		all_rec_per_port[index].status = MLX4_GUID_INFO_STATUS_IDLE;
+
+}
+
+static int set_guid_rec(struct ib_device *ibdev,
+			struct mlx4_next_alias_guid_work *rec)
+{
+	int err;
+	struct mlx4_ib_dev *dev = to_mdev(ibdev);
+	struct ib_sa_guidinfo_rec guid_info_rec;
+	ib_sa_comp_mask comp_mask;
+	struct ib_port_attr attr;
+	struct mlx4_alias_guid_work_context *callback_context;
+	unsigned long resched_delay, flags, flags1;
+	u8 port = rec->port + 1;
+	int index = rec->block_num;
+	struct mlx4_sriov_alias_guid_info_rec_det *rec_det = &rec->rec_det;
+	struct list_head *head =
+		&dev->sriov.alias_guid.ports_guid[port - 1].cb_list;
+
+	memset(&attr, 0, sizeof(attr));
+	err = __mlx4_ib_query_port(ibdev, port, &attr, 1);
+	if (err) {
+		pr_debug("mlx4_ib_query_port failed (err: %d), port: %d\n",
+			 err, port);
+		return err;
+	}
+	/*check the port was configured by the sm, otherwise no need to send */
+	if (attr.state != IB_PORT_ACTIVE) {
+		pr_debug("port %d not active...rescheduling\n", port);
+		resched_delay = 5 * HZ;
+		err = -EAGAIN;
+		goto new_schedule;
+	}
+
+	callback_context = kmalloc(sizeof *callback_context, GFP_KERNEL);
+	if (!callback_context) {
+		err = -ENOMEM;
+		resched_delay = HZ * 5;
+		goto new_schedule;
+	}
+	callback_context->port = port;
+	callback_context->dev = dev;
+	callback_context->block_num = index;
+	callback_context->guid_indexes = rec_det->guid_indexes;
+	callback_context->method = rec->method;
+
+	memset(&guid_info_rec, 0, sizeof (struct ib_sa_guidinfo_rec));
+
+	guid_info_rec.lid = ib_lid_be16(attr.lid);
+	guid_info_rec.block_num = index;
+
+	memcpy(guid_info_rec.guid_info_list, rec_det->all_recs,
+	       GUID_REC_SIZE * NUM_ALIAS_GUID_IN_REC);
+	comp_mask = IB_SA_GUIDINFO_REC_LID | IB_SA_GUIDINFO_REC_BLOCK_NUM |
+		rec_det->guid_indexes;
+
+	init_completion(&callback_context->done);
+	spin_lock_irqsave(&dev->sriov.alias_guid.ag_work_lock, flags1);
+	list_add_tail(&callback_context->list, head);
+	spin_unlock_irqrestore(&dev->sriov.alias_guid.ag_work_lock, flags1);
+
+	callback_context->query_id =
+		ib_sa_guid_info_rec_query(dev->sriov.alias_guid.sa_client,
+					  ibdev, port, &guid_info_rec,
+					  comp_mask, rec->method, 1000,
+					  GFP_KERNEL, aliasguid_query_handler,
+					  callback_context,
+					  &callback_context->sa_query);
+	if (callback_context->query_id < 0) {
+		pr_debug("ib_sa_guid_info_rec_query failed, query_id: "
+			 "%d. will reschedule to the next 1 sec.\n",
+			 callback_context->query_id);
+		spin_lock_irqsave(&dev->sriov.alias_guid.ag_work_lock, flags1);
+		list_del(&callback_context->list);
+		kfree(callback_context);
+		spin_unlock_irqrestore(&dev->sriov.alias_guid.ag_work_lock, flags1);
+		resched_delay = 1 * HZ;
+		err = -EAGAIN;
+		goto new_schedule;
+	}
+	err = 0;
+	goto out;
+
+new_schedule:
+	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
+	spin_lock_irqsave(&dev->sriov.alias_guid.ag_work_lock, flags1);
+	invalidate_guid_record(dev, port, index);
+	if (!dev->sriov.is_going_down) {
+		queue_delayed_work(dev->sriov.alias_guid.ports_guid[port - 1].wq,
+				   &dev->sriov.alias_guid.ports_guid[port - 1].alias_guid_work,
+				   resched_delay);
+	}
+	spin_unlock_irqrestore(&dev->sriov.alias_guid.ag_work_lock, flags1);
+	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
+
+out:
+	return err;
+}
+
+static void mlx4_ib_guid_port_init(struct mlx4_ib_dev *dev, int port)
+{
+	int j, k, entry;
+	__be64 guid;
+
+	/*Check if the SM doesn't need to assign the GUIDs*/
+	for (j = 0; j < NUM_ALIAS_GUID_REC_IN_PORT; j++) {
+		for (k = 0; k < NUM_ALIAS_GUID_IN_REC; k++) {
+			entry = j * NUM_ALIAS_GUID_IN_REC + k;
+			/* no request for the 0 entry (hw guid) */
+			if (!entry || entry > dev->dev->persist->num_vfs ||
+			    !mlx4_is_slave_active(dev->dev, entry))
+				continue;
+			guid = mlx4_get_admin_guid(dev->dev, entry, port);
+			*(__be64 *)&dev->sriov.alias_guid.ports_guid[port - 1].
+				all_rec_per_port[j].all_recs
+				[GUID_REC_SIZE * k] = guid;
+			pr_debug("guid was set, entry=%d, val=0x%llx, port=%d\n",
+				 entry,
+				 be64_to_cpu(guid),
+				 port);
+		}
+	}
+}
+void mlx4_ib_invalidate_all_guid_record(struct mlx4_ib_dev *dev, int port)
+{
+	int i;
+	unsigned long flags, flags1;
+
+	pr_debug("port %d\n", port);
+
+	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
+	spin_lock_irqsave(&dev->sriov.alias_guid.ag_work_lock, flags1);
+
+	if (dev->sriov.alias_guid.ports_guid[port - 1].state_flags &
+		GUID_STATE_NEED_PORT_INIT) {
+		mlx4_ib_guid_port_init(dev, port);
+		dev->sriov.alias_guid.ports_guid[port - 1].state_flags &=
+			(~GUID_STATE_NEED_PORT_INIT);
+	}
+	for (i = 0; i < NUM_ALIAS_GUID_REC_IN_PORT; i++)
+		invalidate_guid_record(dev, port, i);
+
+	if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down) {
+		/*
+		make sure no work waits in the queue, if the work is already
+		queued(not on the timer) the cancel will fail. That is not a problem
+		because we just want the work started.
+		*/
+		cancel_delayed_work(&dev->sriov.alias_guid.
+				      ports_guid[port - 1].alias_guid_work);
+		queue_delayed_work(dev->sriov.alias_guid.ports_guid[port - 1].wq,
+				   &dev->sriov.alias_guid.ports_guid[port - 1].alias_guid_work,
+				   0);
+	}
+	spin_unlock_irqrestore(&dev->sriov.alias_guid.ag_work_lock, flags1);
+	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
+}
+
+static void set_required_record(struct mlx4_ib_dev *dev, u8 port,
+				struct mlx4_next_alias_guid_work *next_rec,
+				int record_index)
+{
+	int i;
+	int lowset_time_entry = -1;
+	int lowest_time = 0;
+	ib_sa_comp_mask delete_guid_indexes = 0;
+	ib_sa_comp_mask set_guid_indexes = 0;
+	struct mlx4_sriov_alias_guid_info_rec_det *rec =
+			&dev->sriov.alias_guid.ports_guid[port].
+			all_rec_per_port[record_index];
+
+	for (i = 0; i < NUM_ALIAS_GUID_IN_REC; i++) {
+		if (!(rec->guid_indexes &
+			mlx4_ib_get_aguid_comp_mask_from_ix(i)))
+			continue;
+
+		if (*(__be64 *)&rec->all_recs[i * GUID_REC_SIZE] ==
+				cpu_to_be64(MLX4_GUID_FOR_DELETE_VAL))
+			delete_guid_indexes |=
+				mlx4_ib_get_aguid_comp_mask_from_ix(i);
+		else
+			set_guid_indexes |=
+				mlx4_ib_get_aguid_comp_mask_from_ix(i);
+
+		if (lowset_time_entry == -1 || rec->guids_retry_schedule[i] <=
+			lowest_time) {
+			lowset_time_entry = i;
+			lowest_time = rec->guids_retry_schedule[i];
+		}
+	}
+
+	memcpy(&next_rec->rec_det, rec, sizeof(*rec));
+	next_rec->port = port;
+	next_rec->block_num = record_index;
+
+	if (*(__be64 *)&rec->all_recs[lowset_time_entry * GUID_REC_SIZE] ==
+				cpu_to_be64(MLX4_GUID_FOR_DELETE_VAL)) {
+		next_rec->rec_det.guid_indexes = delete_guid_indexes;
+		next_rec->method = MLX4_GUID_INFO_RECORD_DELETE;
+	} else {
+		next_rec->rec_det.guid_indexes = set_guid_indexes;
+		next_rec->method = MLX4_GUID_INFO_RECORD_SET;
+	}
+}
+
+/* return index of record that should be updated based on lowest
+ * rescheduled time
+ */
+static int get_low_record_time_index(struct mlx4_ib_dev *dev, u8 port,
+				     int *resched_delay_sec)
+{
+	int record_index = -1;
+	u64 low_record_time = 0;
+	struct mlx4_sriov_alias_guid_info_rec_det rec;
+	int j;
+
+	for (j = 0; j < NUM_ALIAS_GUID_REC_IN_PORT; j++) {
+		rec = dev->sriov.alias_guid.ports_guid[port].
+			all_rec_per_port[j];
+		if (rec.status == MLX4_GUID_INFO_STATUS_IDLE &&
+		    rec.guid_indexes) {
+			if (record_index == -1 ||
+			    rec.time_to_run < low_record_time) {
+				record_index = j;
+				low_record_time = rec.time_to_run;
+			}
+		}
+	}
+	if (resched_delay_sec) {
+		u64 curr_time = ktime_get_boottime_ns();
+
+		*resched_delay_sec = (low_record_time < curr_time) ? 0 :
+			div_u64((low_record_time - curr_time), NSEC_PER_SEC);
+	}
+
+	return record_index;
+}
+
+/* The function returns the next record that was
+ * not configured (or failed to be configured) */
+static int get_next_record_to_update(struct mlx4_ib_dev *dev, u8 port,
+				     struct mlx4_next_alias_guid_work *rec)
+{
+	unsigned long flags;
+	int record_index;
+	int ret = 0;
+
+	spin_lock_irqsave(&dev->sriov.alias_guid.ag_work_lock, flags);
+	record_index = get_low_record_time_index(dev, port, NULL);
+
+	if (record_index < 0) {
+		ret = -ENOENT;
+		goto out;
+	}
+
+	set_required_record(dev, port, rec, record_index);
+out:
+	spin_unlock_irqrestore(&dev->sriov.alias_guid.ag_work_lock, flags);
+	return ret;
+}
+
+static void alias_guid_work(struct work_struct *work)
+{
+	struct delayed_work *delay = to_delayed_work(work);
+	int ret = 0;
+	struct mlx4_next_alias_guid_work *rec;
+	struct mlx4_sriov_alias_guid_port_rec_det *sriov_alias_port =
+		container_of(delay, struct mlx4_sriov_alias_guid_port_rec_det,
+			     alias_guid_work);
+	struct mlx4_sriov_alias_guid *sriov_alias_guid = sriov_alias_port->parent;
+	struct mlx4_ib_sriov *ib_sriov = container_of(sriov_alias_guid,
+						struct mlx4_ib_sriov,
+						alias_guid);
+	struct mlx4_ib_dev *dev = container_of(ib_sriov, struct mlx4_ib_dev, sriov);
+
+	rec = kzalloc(sizeof *rec, GFP_KERNEL);
+	if (!rec)
+		return;
+
+	pr_debug("starting [port: %d]...\n", sriov_alias_port->port + 1);
+	ret = get_next_record_to_update(dev, sriov_alias_port->port, rec);
+	if (ret) {
+		pr_debug("No more records to update.\n");
+		goto out;
+	}
+
+	set_guid_rec(&dev->ib_dev, rec);
+out:
+	kfree(rec);
+}
+
+
+void mlx4_ib_init_alias_guid_work(struct mlx4_ib_dev *dev, int port)
+{
+	unsigned long flags, flags1;
+
+	if (!mlx4_is_master(dev->dev))
+		return;
+	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
+	spin_lock_irqsave(&dev->sriov.alias_guid.ag_work_lock, flags1);
+	if (!dev->sriov.is_going_down) {
+		/* If there is pending one should cancel then run, otherwise
+		  * won't run till previous one is ended as same work
+		  * struct is used.
+		  */
+		cancel_delayed_work(&dev->sriov.alias_guid.ports_guid[port].
+				    alias_guid_work);
+		queue_delayed_work(dev->sriov.alias_guid.ports_guid[port].wq,
+			   &dev->sriov.alias_guid.ports_guid[port].alias_guid_work, 0);
+	}
+	spin_unlock_irqrestore(&dev->sriov.alias_guid.ag_work_lock, flags1);
+	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
+}
+
+void mlx4_ib_destroy_alias_guid_service(struct mlx4_ib_dev *dev)
+{
+	int i;
+	struct mlx4_ib_sriov *sriov = &dev->sriov;
+	struct mlx4_alias_guid_work_context *cb_ctx;
+	struct mlx4_sriov_alias_guid_port_rec_det *det;
+	struct ib_sa_query *sa_query;
+	unsigned long flags;
+
+	for (i = 0 ; i < dev->num_ports; i++) {
+		det = &sriov->alias_guid.ports_guid[i];
+		cancel_delayed_work_sync(&det->alias_guid_work);
+		spin_lock_irqsave(&sriov->alias_guid.ag_work_lock, flags);
+		while (!list_empty(&det->cb_list)) {
+			cb_ctx = list_entry(det->cb_list.next,
+					    struct mlx4_alias_guid_work_context,
+					    list);
+			sa_query = cb_ctx->sa_query;
+			cb_ctx->sa_query = NULL;
+			list_del(&cb_ctx->list);
+			spin_unlock_irqrestore(&sriov->alias_guid.ag_work_lock, flags);
+			ib_sa_cancel_query(cb_ctx->query_id, sa_query);
+			wait_for_completion(&cb_ctx->done);
+			kfree(cb_ctx);
+			spin_lock_irqsave(&sriov->alias_guid.ag_work_lock, flags);
+		}
+		spin_unlock_irqrestore(&sriov->alias_guid.ag_work_lock, flags);
+	}
+	for (i = 0 ; i < dev->num_ports; i++) {
+		flush_workqueue(dev->sriov.alias_guid.ports_guid[i].wq);
+		destroy_workqueue(dev->sriov.alias_guid.ports_guid[i].wq);
+	}
+	ib_sa_unregister_client(dev->sriov.alias_guid.sa_client);
+	kfree(dev->sriov.alias_guid.sa_client);
+}
+
+int mlx4_ib_init_alias_guid_service(struct mlx4_ib_dev *dev)
+{
+	char alias_wq_name[15];
+	int ret = 0;
+	int i, j;
+	union ib_gid gid;
+
+	if (!mlx4_is_master(dev->dev))
+		return 0;
+	dev->sriov.alias_guid.sa_client =
+		kzalloc(sizeof *dev->sriov.alias_guid.sa_client, GFP_KERNEL);
+	if (!dev->sriov.alias_guid.sa_client)
+		return -ENOMEM;
+
+	ib_sa_register_client(dev->sriov.alias_guid.sa_client);
+
+	spin_lock_init(&dev->sriov.alias_guid.ag_work_lock);
+
+	for (i = 1; i <= dev->num_ports; ++i) {
+		if (dev->ib_dev.ops.query_gid(&dev->ib_dev, i, 0, &gid)) {
+			ret = -EFAULT;
+			goto err_unregister;
+		}
+	}
+
+	for (i = 0 ; i < dev->num_ports; i++) {
+		memset(&dev->sriov.alias_guid.ports_guid[i], 0,
+		       sizeof (struct mlx4_sriov_alias_guid_port_rec_det));
+		dev->sriov.alias_guid.ports_guid[i].state_flags |=
+				GUID_STATE_NEED_PORT_INIT;
+		for (j = 0; j < NUM_ALIAS_GUID_REC_IN_PORT; j++) {
+			/* mark each val as it was deleted */
+			memset(dev->sriov.alias_guid.ports_guid[i].
+				all_rec_per_port[j].all_recs, 0xFF,
+				sizeof(dev->sriov.alias_guid.ports_guid[i].
+				all_rec_per_port[j].all_recs));
+		}
+		INIT_LIST_HEAD(&dev->sriov.alias_guid.ports_guid[i].cb_list);
+		/*prepare the records, set them to be allocated by sm*/
+		if (mlx4_ib_sm_guid_assign)
+			for (j = 1; j < NUM_ALIAS_GUID_PER_PORT; j++)
+				mlx4_set_admin_guid(dev->dev, 0, j, i + 1);
+		for (j = 0 ; j < NUM_ALIAS_GUID_REC_IN_PORT; j++)
+			invalidate_guid_record(dev, i + 1, j);
+
+		dev->sriov.alias_guid.ports_guid[i].parent = &dev->sriov.alias_guid;
+		dev->sriov.alias_guid.ports_guid[i].port  = i;
+
+		snprintf(alias_wq_name, sizeof alias_wq_name, "alias_guid%d", i);
+		dev->sriov.alias_guid.ports_guid[i].wq =
+			alloc_ordered_workqueue(alias_wq_name, WQ_MEM_RECLAIM);
+		if (!dev->sriov.alias_guid.ports_guid[i].wq) {
+			ret = -ENOMEM;
+			goto err_thread;
+		}
+		INIT_DELAYED_WORK(&dev->sriov.alias_guid.ports_guid[i].alias_guid_work,
+			  alias_guid_work);
+	}
+	return 0;
+
+err_thread:
+	for (--i; i >= 0; i--) {
+		destroy_workqueue(dev->sriov.alias_guid.ports_guid[i].wq);
+		dev->sriov.alias_guid.ports_guid[i].wq = NULL;
+	}
+
+err_unregister:
+	ib_sa_unregister_client(dev->sriov.alias_guid.sa_client);
+	kfree(dev->sriov.alias_guid.sa_client);
+	dev->sriov.alias_guid.sa_client = NULL;
+	pr_err("init_alias_guid_service: Failed. (ret:%d)\n", ret);
+	return ret;
+}
diff --git a/drivers/infiniband/hw/mlx4/cm.c b/drivers/infiniband/hw/mlx4/cm.c
new file mode 100644
index 000000000..4aff1c829
--- /dev/null
+++ b/drivers/infiniband/hw/mlx4/cm.c
@@ -0,0 +1,589 @@
+/*
+ * Copyright (c) 2012 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <rdma/ib_mad.h>
+
+#include <linux/mlx4/cmd.h>
+#include <linux/rbtree.h>
+#include <linux/idr.h>
+#include <rdma/ib_cm.h>
+
+#include "mlx4_ib.h"
+
+#define CM_CLEANUP_CACHE_TIMEOUT  (30 * HZ)
+
+struct id_map_entry {
+	struct rb_node node;
+
+	u32 sl_cm_id;
+	u32 pv_cm_id;
+	int slave_id;
+	int scheduled_delete;
+	struct mlx4_ib_dev *dev;
+
+	struct list_head list;
+	struct delayed_work timeout;
+};
+
+struct rej_tmout_entry {
+	int slave;
+	u32 rem_pv_cm_id;
+	struct delayed_work timeout;
+	struct xarray *xa_rej_tmout;
+};
+
+struct cm_generic_msg {
+	struct ib_mad_hdr hdr;
+
+	__be32 local_comm_id;
+	__be32 remote_comm_id;
+	unsigned char unused[2];
+	__be16 rej_reason;
+};
+
+struct cm_sidr_generic_msg {
+	struct ib_mad_hdr hdr;
+	__be32 request_id;
+};
+
+struct cm_req_msg {
+	unsigned char unused[0x60];
+	union ib_gid primary_path_sgid;
+};
+
+
+static void set_local_comm_id(struct ib_mad *mad, u32 cm_id)
+{
+	if (mad->mad_hdr.attr_id == CM_SIDR_REQ_ATTR_ID) {
+		struct cm_sidr_generic_msg *msg =
+			(struct cm_sidr_generic_msg *)mad;
+		msg->request_id = cpu_to_be32(cm_id);
+	} else if (mad->mad_hdr.attr_id == CM_SIDR_REP_ATTR_ID) {
+		pr_err("trying to set local_comm_id in SIDR_REP\n");
+		return;
+	} else {
+		struct cm_generic_msg *msg = (struct cm_generic_msg *)mad;
+		msg->local_comm_id = cpu_to_be32(cm_id);
+	}
+}
+
+static u32 get_local_comm_id(struct ib_mad *mad)
+{
+	if (mad->mad_hdr.attr_id == CM_SIDR_REQ_ATTR_ID) {
+		struct cm_sidr_generic_msg *msg =
+			(struct cm_sidr_generic_msg *)mad;
+		return be32_to_cpu(msg->request_id);
+	} else if (mad->mad_hdr.attr_id == CM_SIDR_REP_ATTR_ID) {
+		pr_err("trying to set local_comm_id in SIDR_REP\n");
+		return -1;
+	} else {
+		struct cm_generic_msg *msg = (struct cm_generic_msg *)mad;
+		return be32_to_cpu(msg->local_comm_id);
+	}
+}
+
+static void set_remote_comm_id(struct ib_mad *mad, u32 cm_id)
+{
+	if (mad->mad_hdr.attr_id == CM_SIDR_REP_ATTR_ID) {
+		struct cm_sidr_generic_msg *msg =
+			(struct cm_sidr_generic_msg *)mad;
+		msg->request_id = cpu_to_be32(cm_id);
+	} else if (mad->mad_hdr.attr_id == CM_SIDR_REQ_ATTR_ID) {
+		pr_err("trying to set remote_comm_id in SIDR_REQ\n");
+		return;
+	} else {
+		struct cm_generic_msg *msg = (struct cm_generic_msg *)mad;
+		msg->remote_comm_id = cpu_to_be32(cm_id);
+	}
+}
+
+static u32 get_remote_comm_id(struct ib_mad *mad)
+{
+	if (mad->mad_hdr.attr_id == CM_SIDR_REP_ATTR_ID) {
+		struct cm_sidr_generic_msg *msg =
+			(struct cm_sidr_generic_msg *)mad;
+		return be32_to_cpu(msg->request_id);
+	} else if (mad->mad_hdr.attr_id == CM_SIDR_REQ_ATTR_ID) {
+		pr_err("trying to set remote_comm_id in SIDR_REQ\n");
+		return -1;
+	} else {
+		struct cm_generic_msg *msg = (struct cm_generic_msg *)mad;
+		return be32_to_cpu(msg->remote_comm_id);
+	}
+}
+
+static union ib_gid gid_from_req_msg(struct ib_device *ibdev, struct ib_mad *mad)
+{
+	struct cm_req_msg *msg = (struct cm_req_msg *)mad;
+
+	return msg->primary_path_sgid;
+}
+
+/* Lock should be taken before called */
+static struct id_map_entry *
+id_map_find_by_sl_id(struct ib_device *ibdev, u32 slave_id, u32 sl_cm_id)
+{
+	struct rb_root *sl_id_map = &to_mdev(ibdev)->sriov.sl_id_map;
+	struct rb_node *node = sl_id_map->rb_node;
+
+	while (node) {
+		struct id_map_entry *id_map_entry =
+			rb_entry(node, struct id_map_entry, node);
+
+		if (id_map_entry->sl_cm_id > sl_cm_id)
+			node = node->rb_left;
+		else if (id_map_entry->sl_cm_id < sl_cm_id)
+			node = node->rb_right;
+		else if (id_map_entry->slave_id > slave_id)
+			node = node->rb_left;
+		else if (id_map_entry->slave_id < slave_id)
+			node = node->rb_right;
+		else
+			return id_map_entry;
+	}
+	return NULL;
+}
+
+static void id_map_ent_timeout(struct work_struct *work)
+{
+	struct delayed_work *delay = to_delayed_work(work);
+	struct id_map_entry *ent = container_of(delay, struct id_map_entry, timeout);
+	struct id_map_entry *found_ent;
+	struct mlx4_ib_dev *dev = ent->dev;
+	struct mlx4_ib_sriov *sriov = &dev->sriov;
+	struct rb_root *sl_id_map = &sriov->sl_id_map;
+
+	spin_lock(&sriov->id_map_lock);
+	if (!xa_erase(&sriov->pv_id_table, ent->pv_cm_id))
+		goto out;
+	found_ent = id_map_find_by_sl_id(&dev->ib_dev, ent->slave_id, ent->sl_cm_id);
+	if (found_ent && found_ent == ent)
+		rb_erase(&found_ent->node, sl_id_map);
+
+out:
+	list_del(&ent->list);
+	spin_unlock(&sriov->id_map_lock);
+	kfree(ent);
+}
+
+static void sl_id_map_add(struct ib_device *ibdev, struct id_map_entry *new)
+{
+	struct rb_root *sl_id_map = &to_mdev(ibdev)->sriov.sl_id_map;
+	struct rb_node **link = &sl_id_map->rb_node, *parent = NULL;
+	struct id_map_entry *ent;
+	int slave_id = new->slave_id;
+	int sl_cm_id = new->sl_cm_id;
+
+	ent = id_map_find_by_sl_id(ibdev, slave_id, sl_cm_id);
+	if (ent) {
+		pr_debug("overriding existing sl_id_map entry (cm_id = %x)\n",
+			 sl_cm_id);
+
+		rb_replace_node(&ent->node, &new->node, sl_id_map);
+		return;
+	}
+
+	/* Go to the bottom of the tree */
+	while (*link) {
+		parent = *link;
+		ent = rb_entry(parent, struct id_map_entry, node);
+
+		if (ent->sl_cm_id > sl_cm_id || (ent->sl_cm_id == sl_cm_id && ent->slave_id > slave_id))
+			link = &(*link)->rb_left;
+		else
+			link = &(*link)->rb_right;
+	}
+
+	rb_link_node(&new->node, parent, link);
+	rb_insert_color(&new->node, sl_id_map);
+}
+
+static struct id_map_entry *
+id_map_alloc(struct ib_device *ibdev, int slave_id, u32 sl_cm_id)
+{
+	int ret;
+	struct id_map_entry *ent;
+	struct mlx4_ib_sriov *sriov = &to_mdev(ibdev)->sriov;
+
+	ent = kmalloc(sizeof (struct id_map_entry), GFP_KERNEL);
+	if (!ent)
+		return ERR_PTR(-ENOMEM);
+
+	ent->sl_cm_id = sl_cm_id;
+	ent->slave_id = slave_id;
+	ent->scheduled_delete = 0;
+	ent->dev = to_mdev(ibdev);
+	INIT_DELAYED_WORK(&ent->timeout, id_map_ent_timeout);
+
+	ret = xa_alloc_cyclic(&sriov->pv_id_table, &ent->pv_cm_id, ent,
+			xa_limit_32b, &sriov->pv_id_next, GFP_KERNEL);
+	if (ret >= 0) {
+		spin_lock(&sriov->id_map_lock);
+		sl_id_map_add(ibdev, ent);
+		list_add_tail(&ent->list, &sriov->cm_list);
+		spin_unlock(&sriov->id_map_lock);
+		return ent;
+	}
+
+	/*error flow*/
+	kfree(ent);
+	mlx4_ib_warn(ibdev, "Allocation failed (err:0x%x)\n", ret);
+	return ERR_PTR(-ENOMEM);
+}
+
+static struct id_map_entry *
+id_map_get(struct ib_device *ibdev, int *pv_cm_id, int slave_id, int sl_cm_id)
+{
+	struct id_map_entry *ent;
+	struct mlx4_ib_sriov *sriov = &to_mdev(ibdev)->sriov;
+
+	spin_lock(&sriov->id_map_lock);
+	if (*pv_cm_id == -1) {
+		ent = id_map_find_by_sl_id(ibdev, slave_id, sl_cm_id);
+		if (ent)
+			*pv_cm_id = (int) ent->pv_cm_id;
+	} else
+		ent = xa_load(&sriov->pv_id_table, *pv_cm_id);
+	spin_unlock(&sriov->id_map_lock);
+
+	return ent;
+}
+
+static void schedule_delayed(struct ib_device *ibdev, struct id_map_entry *id)
+{
+	struct mlx4_ib_sriov *sriov = &to_mdev(ibdev)->sriov;
+	unsigned long flags;
+
+	spin_lock(&sriov->id_map_lock);
+	spin_lock_irqsave(&sriov->going_down_lock, flags);
+	/*make sure that there is no schedule inside the scheduled work.*/
+	if (!sriov->is_going_down && !id->scheduled_delete) {
+		id->scheduled_delete = 1;
+		schedule_delayed_work(&id->timeout, CM_CLEANUP_CACHE_TIMEOUT);
+	} else if (id->scheduled_delete) {
+		/* Adjust timeout if already scheduled */
+		mod_delayed_work(system_wq, &id->timeout, CM_CLEANUP_CACHE_TIMEOUT);
+	}
+	spin_unlock_irqrestore(&sriov->going_down_lock, flags);
+	spin_unlock(&sriov->id_map_lock);
+}
+
+#define REJ_REASON(m) be16_to_cpu(((struct cm_generic_msg *)(m))->rej_reason)
+int mlx4_ib_multiplex_cm_handler(struct ib_device *ibdev, int port, int slave_id,
+		struct ib_mad *mad)
+{
+	struct id_map_entry *id;
+	u32 sl_cm_id;
+	int pv_cm_id = -1;
+
+	if (mad->mad_hdr.attr_id == CM_REQ_ATTR_ID ||
+	    mad->mad_hdr.attr_id == CM_REP_ATTR_ID ||
+	    mad->mad_hdr.attr_id == CM_MRA_ATTR_ID ||
+	    mad->mad_hdr.attr_id == CM_SIDR_REQ_ATTR_ID ||
+	    (mad->mad_hdr.attr_id == CM_REJ_ATTR_ID && REJ_REASON(mad) == IB_CM_REJ_TIMEOUT)) {
+		sl_cm_id = get_local_comm_id(mad);
+		id = id_map_get(ibdev, &pv_cm_id, slave_id, sl_cm_id);
+		if (id)
+			goto cont;
+		id = id_map_alloc(ibdev, slave_id, sl_cm_id);
+		if (IS_ERR(id)) {
+			mlx4_ib_warn(ibdev, "%s: id{slave: %d, sl_cm_id: 0x%x} Failed to id_map_alloc\n",
+				__func__, slave_id, sl_cm_id);
+			return PTR_ERR(id);
+		}
+	} else if (mad->mad_hdr.attr_id == CM_REJ_ATTR_ID ||
+		   mad->mad_hdr.attr_id == CM_SIDR_REP_ATTR_ID) {
+		return 0;
+	} else {
+		sl_cm_id = get_local_comm_id(mad);
+		id = id_map_get(ibdev, &pv_cm_id, slave_id, sl_cm_id);
+	}
+
+	if (!id) {
+		pr_debug("id{slave: %d, sl_cm_id: 0x%x} is NULL! attr_id: 0x%x\n",
+			 slave_id, sl_cm_id, be16_to_cpu(mad->mad_hdr.attr_id));
+		return -EINVAL;
+	}
+
+cont:
+	set_local_comm_id(mad, id->pv_cm_id);
+
+	if (mad->mad_hdr.attr_id == CM_DREQ_ATTR_ID)
+		schedule_delayed(ibdev, id);
+	return 0;
+}
+
+static void rej_tmout_timeout(struct work_struct *work)
+{
+	struct delayed_work *delay = to_delayed_work(work);
+	struct rej_tmout_entry *item = container_of(delay, struct rej_tmout_entry, timeout);
+	struct rej_tmout_entry *deleted;
+
+	deleted = xa_cmpxchg(item->xa_rej_tmout, item->rem_pv_cm_id, item, NULL, 0);
+
+	if (deleted != item)
+		pr_debug("deleted(%p) != item(%p)\n", deleted, item);
+
+	kfree(item);
+}
+
+static int alloc_rej_tmout(struct mlx4_ib_sriov *sriov, u32 rem_pv_cm_id, int slave)
+{
+	struct rej_tmout_entry *item;
+	struct rej_tmout_entry *old;
+	int ret = 0;
+
+	xa_lock(&sriov->xa_rej_tmout);
+	item = xa_load(&sriov->xa_rej_tmout, (unsigned long)rem_pv_cm_id);
+
+	if (item) {
+		if (xa_err(item))
+			ret =  xa_err(item);
+		else
+			/* If a retry, adjust delayed work */
+			mod_delayed_work(system_wq, &item->timeout, CM_CLEANUP_CACHE_TIMEOUT);
+		goto err_or_exists;
+	}
+	xa_unlock(&sriov->xa_rej_tmout);
+
+	item = kmalloc(sizeof(*item), GFP_KERNEL);
+	if (!item)
+		return -ENOMEM;
+
+	INIT_DELAYED_WORK(&item->timeout, rej_tmout_timeout);
+	item->slave = slave;
+	item->rem_pv_cm_id = rem_pv_cm_id;
+	item->xa_rej_tmout = &sriov->xa_rej_tmout;
+
+	old = xa_cmpxchg(&sriov->xa_rej_tmout, (unsigned long)rem_pv_cm_id, NULL, item, GFP_KERNEL);
+	if (old) {
+		pr_debug(
+			"Non-null old entry (%p) or error (%d) when inserting\n",
+			old, xa_err(old));
+		kfree(item);
+		return xa_err(old);
+	}
+
+	schedule_delayed_work(&item->timeout, CM_CLEANUP_CACHE_TIMEOUT);
+
+	return 0;
+
+err_or_exists:
+	xa_unlock(&sriov->xa_rej_tmout);
+	return ret;
+}
+
+static int lookup_rej_tmout_slave(struct mlx4_ib_sriov *sriov, u32 rem_pv_cm_id)
+{
+	struct rej_tmout_entry *item;
+	int slave;
+
+	xa_lock(&sriov->xa_rej_tmout);
+	item = xa_load(&sriov->xa_rej_tmout, (unsigned long)rem_pv_cm_id);
+
+	if (!item || xa_err(item)) {
+		pr_debug("Could not find slave. rem_pv_cm_id 0x%x error: %d\n",
+			 rem_pv_cm_id, xa_err(item));
+		slave = !item ? -ENOENT : xa_err(item);
+	} else {
+		slave = item->slave;
+	}
+	xa_unlock(&sriov->xa_rej_tmout);
+
+	return slave;
+}
+
+int mlx4_ib_demux_cm_handler(struct ib_device *ibdev, int port, int *slave,
+			     struct ib_mad *mad)
+{
+	struct mlx4_ib_sriov *sriov = &to_mdev(ibdev)->sriov;
+	u32 rem_pv_cm_id = get_local_comm_id(mad);
+	u32 pv_cm_id;
+	struct id_map_entry *id;
+	int sts;
+
+	if (mad->mad_hdr.attr_id == CM_REQ_ATTR_ID ||
+	    mad->mad_hdr.attr_id == CM_SIDR_REQ_ATTR_ID) {
+		union ib_gid gid;
+
+		if (!slave)
+			return 0;
+
+		gid = gid_from_req_msg(ibdev, mad);
+		*slave = mlx4_ib_find_real_gid(ibdev, port, gid.global.interface_id);
+		if (*slave < 0) {
+			mlx4_ib_warn(ibdev, "failed matching slave_id by gid (0x%llx)\n",
+				     be64_to_cpu(gid.global.interface_id));
+			return -ENOENT;
+		}
+
+		sts = alloc_rej_tmout(sriov, rem_pv_cm_id, *slave);
+		if (sts)
+			/* Even if this fails, we pass on the REQ to the slave */
+			pr_debug("Could not allocate rej_tmout entry. rem_pv_cm_id 0x%x slave %d status %d\n",
+				 rem_pv_cm_id, *slave, sts);
+
+		return 0;
+	}
+
+	pv_cm_id = get_remote_comm_id(mad);
+	id = id_map_get(ibdev, (int *)&pv_cm_id, -1, -1);
+
+	if (!id) {
+		if (mad->mad_hdr.attr_id == CM_REJ_ATTR_ID &&
+		    REJ_REASON(mad) == IB_CM_REJ_TIMEOUT && slave) {
+			*slave = lookup_rej_tmout_slave(sriov, rem_pv_cm_id);
+
+			return (*slave < 0) ? *slave : 0;
+		}
+		pr_debug("Couldn't find an entry for pv_cm_id 0x%x, attr_id 0x%x\n",
+			 pv_cm_id, be16_to_cpu(mad->mad_hdr.attr_id));
+		return -ENOENT;
+	}
+
+	if (slave)
+		*slave = id->slave_id;
+	set_remote_comm_id(mad, id->sl_cm_id);
+
+	if (mad->mad_hdr.attr_id == CM_DREQ_ATTR_ID ||
+	    mad->mad_hdr.attr_id == CM_REJ_ATTR_ID)
+		schedule_delayed(ibdev, id);
+
+	return 0;
+}
+
+void mlx4_ib_cm_paravirt_init(struct mlx4_ib_dev *dev)
+{
+	spin_lock_init(&dev->sriov.id_map_lock);
+	INIT_LIST_HEAD(&dev->sriov.cm_list);
+	dev->sriov.sl_id_map = RB_ROOT;
+	xa_init_flags(&dev->sriov.pv_id_table, XA_FLAGS_ALLOC);
+	xa_init(&dev->sriov.xa_rej_tmout);
+}
+
+static void rej_tmout_xa_cleanup(struct mlx4_ib_sriov *sriov, int slave)
+{
+	struct rej_tmout_entry *item;
+	bool flush_needed = false;
+	unsigned long id;
+	int cnt = 0;
+
+	xa_lock(&sriov->xa_rej_tmout);
+	xa_for_each(&sriov->xa_rej_tmout, id, item) {
+		if (slave < 0 || slave == item->slave) {
+			mod_delayed_work(system_wq, &item->timeout, 0);
+			flush_needed = true;
+			++cnt;
+		}
+	}
+	xa_unlock(&sriov->xa_rej_tmout);
+
+	if (flush_needed) {
+		flush_scheduled_work();
+		pr_debug("Deleted %d entries in xarray for slave %d during cleanup\n",
+			 cnt, slave);
+	}
+
+	if (slave < 0)
+		WARN_ON(!xa_empty(&sriov->xa_rej_tmout));
+}
+
+/* slave = -1 ==> all slaves */
+/* TBD -- call paravirt clean for single slave.  Need for slave RESET event */
+void mlx4_ib_cm_paravirt_clean(struct mlx4_ib_dev *dev, int slave)
+{
+	struct mlx4_ib_sriov *sriov = &dev->sriov;
+	struct rb_root *sl_id_map = &sriov->sl_id_map;
+	struct list_head lh;
+	struct rb_node *nd;
+	int need_flush = 0;
+	struct id_map_entry *map, *tmp_map;
+	/* cancel all delayed work queue entries */
+	INIT_LIST_HEAD(&lh);
+	spin_lock(&sriov->id_map_lock);
+	list_for_each_entry_safe(map, tmp_map, &dev->sriov.cm_list, list) {
+		if (slave < 0 || slave == map->slave_id) {
+			if (map->scheduled_delete)
+				need_flush |= !cancel_delayed_work(&map->timeout);
+		}
+	}
+
+	spin_unlock(&sriov->id_map_lock);
+
+	if (need_flush)
+		flush_scheduled_work(); /* make sure all timers were flushed */
+
+	/* now, remove all leftover entries from databases*/
+	spin_lock(&sriov->id_map_lock);
+	if (slave < 0) {
+		while (rb_first(sl_id_map)) {
+			struct id_map_entry *ent =
+				rb_entry(rb_first(sl_id_map),
+					 struct id_map_entry, node);
+
+			rb_erase(&ent->node, sl_id_map);
+			xa_erase(&sriov->pv_id_table, ent->pv_cm_id);
+		}
+		list_splice_init(&dev->sriov.cm_list, &lh);
+	} else {
+		/* first, move nodes belonging to slave to db remove list */
+		nd = rb_first(sl_id_map);
+		while (nd) {
+			struct id_map_entry *ent =
+				rb_entry(nd, struct id_map_entry, node);
+			nd = rb_next(nd);
+			if (ent->slave_id == slave)
+				list_move_tail(&ent->list, &lh);
+		}
+		/* remove those nodes from databases */
+		list_for_each_entry_safe(map, tmp_map, &lh, list) {
+			rb_erase(&map->node, sl_id_map);
+			xa_erase(&sriov->pv_id_table, map->pv_cm_id);
+		}
+
+		/* add remaining nodes from cm_list */
+		list_for_each_entry_safe(map, tmp_map, &dev->sriov.cm_list, list) {
+			if (slave == map->slave_id)
+				list_move_tail(&map->list, &lh);
+		}
+	}
+
+	spin_unlock(&sriov->id_map_lock);
+
+	/* free any map entries left behind due to cancel_delayed_work above */
+	list_for_each_entry_safe(map, tmp_map, &lh, list) {
+		list_del(&map->list);
+		kfree(map);
+	}
+
+	rej_tmout_xa_cleanup(sriov, slave);
+}
diff --git a/drivers/infiniband/hw/mlx4/cq.c b/drivers/infiniband/hw/mlx4/cq.c
new file mode 100644
index 000000000..e9b5a4d57
--- /dev/null
+++ b/drivers/infiniband/hw/mlx4/cq.c
@@ -0,0 +1,973 @@
+/*
+ * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
+ * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/mlx4/cq.h>
+#include <linux/mlx4/qp.h>
+#include <linux/mlx4/srq.h>
+#include <linux/slab.h>
+
+#include "mlx4_ib.h"
+#include <rdma/mlx4-abi.h>
+#include <rdma/uverbs_ioctl.h>
+
+static void mlx4_ib_cq_comp(struct mlx4_cq *cq)
+{
+	struct ib_cq *ibcq = &to_mibcq(cq)->ibcq;
+	ibcq->comp_handler(ibcq, ibcq->cq_context);
+}
+
+static void mlx4_ib_cq_event(struct mlx4_cq *cq, enum mlx4_event type)
+{
+	struct ib_event event;
+	struct ib_cq *ibcq;
+
+	if (type != MLX4_EVENT_TYPE_CQ_ERROR) {
+		pr_warn("Unexpected event type %d "
+		       "on CQ %06x\n", type, cq->cqn);
+		return;
+	}
+
+	ibcq = &to_mibcq(cq)->ibcq;
+	if (ibcq->event_handler) {
+		event.device     = ibcq->device;
+		event.event      = IB_EVENT_CQ_ERR;
+		event.element.cq = ibcq;
+		ibcq->event_handler(&event, ibcq->cq_context);
+	}
+}
+
+static void *get_cqe_from_buf(struct mlx4_ib_cq_buf *buf, int n)
+{
+	return mlx4_buf_offset(&buf->buf, n * buf->entry_size);
+}
+
+static void *get_cqe(struct mlx4_ib_cq *cq, int n)
+{
+	return get_cqe_from_buf(&cq->buf, n);
+}
+
+static void *get_sw_cqe(struct mlx4_ib_cq *cq, int n)
+{
+	struct mlx4_cqe *cqe = get_cqe(cq, n & cq->ibcq.cqe);
+	struct mlx4_cqe *tcqe = ((cq->buf.entry_size == 64) ? (cqe + 1) : cqe);
+
+	return (!!(tcqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK) ^
+		!!(n & (cq->ibcq.cqe + 1))) ? NULL : cqe;
+}
+
+static struct mlx4_cqe *next_cqe_sw(struct mlx4_ib_cq *cq)
+{
+	return get_sw_cqe(cq, cq->mcq.cons_index);
+}
+
+int mlx4_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
+{
+	struct mlx4_ib_cq *mcq = to_mcq(cq);
+	struct mlx4_ib_dev *dev = to_mdev(cq->device);
+
+	return mlx4_cq_modify(dev->dev, &mcq->mcq, cq_count, cq_period);
+}
+
+static int mlx4_ib_alloc_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int nent)
+{
+	int err;
+
+	err = mlx4_buf_alloc(dev->dev, nent * dev->dev->caps.cqe_size,
+			     PAGE_SIZE * 2, &buf->buf);
+
+	if (err)
+		goto out;
+
+	buf->entry_size = dev->dev->caps.cqe_size;
+	err = mlx4_mtt_init(dev->dev, buf->buf.npages, buf->buf.page_shift,
+				    &buf->mtt);
+	if (err)
+		goto err_buf;
+
+	err = mlx4_buf_write_mtt(dev->dev, &buf->mtt, &buf->buf);
+	if (err)
+		goto err_mtt;
+
+	return 0;
+
+err_mtt:
+	mlx4_mtt_cleanup(dev->dev, &buf->mtt);
+
+err_buf:
+	mlx4_buf_free(dev->dev, nent * buf->entry_size, &buf->buf);
+
+out:
+	return err;
+}
+
+static void mlx4_ib_free_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int cqe)
+{
+	mlx4_buf_free(dev->dev, (cqe + 1) * buf->entry_size, &buf->buf);
+}
+
+static int mlx4_ib_get_cq_umem(struct mlx4_ib_dev *dev, struct ib_udata *udata,
+			       struct mlx4_ib_cq_buf *buf,
+			       struct ib_umem **umem, u64 buf_addr, int cqe)
+{
+	int err;
+	int cqe_size = dev->dev->caps.cqe_size;
+	int shift;
+	int n;
+
+	*umem = ib_umem_get(&dev->ib_dev, buf_addr, cqe * cqe_size,
+			    IB_ACCESS_LOCAL_WRITE);
+	if (IS_ERR(*umem))
+		return PTR_ERR(*umem);
+
+	shift = mlx4_ib_umem_calc_optimal_mtt_size(*umem, 0, &n);
+	err = mlx4_mtt_init(dev->dev, n, shift, &buf->mtt);
+
+	if (err)
+		goto err_buf;
+
+	err = mlx4_ib_umem_write_mtt(dev, &buf->mtt, *umem);
+	if (err)
+		goto err_mtt;
+
+	return 0;
+
+err_mtt:
+	mlx4_mtt_cleanup(dev->dev, &buf->mtt);
+
+err_buf:
+	ib_umem_release(*umem);
+
+	return err;
+}
+
+#define CQ_CREATE_FLAGS_SUPPORTED IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION
+int mlx4_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
+		      struct ib_udata *udata)
+{
+	struct ib_device *ibdev = ibcq->device;
+	int entries = attr->cqe;
+	int vector = attr->comp_vector;
+	struct mlx4_ib_dev *dev = to_mdev(ibdev);
+	struct mlx4_ib_cq *cq = to_mcq(ibcq);
+	struct mlx4_uar *uar;
+	void *buf_addr;
+	int err;
+	struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context(
+		udata, struct mlx4_ib_ucontext, ibucontext);
+
+	if (entries < 1 || entries > dev->dev->caps.max_cqes)
+		return -EINVAL;
+
+	if (attr->flags & ~CQ_CREATE_FLAGS_SUPPORTED)
+		return -EINVAL;
+
+	entries      = roundup_pow_of_two(entries + 1);
+	cq->ibcq.cqe = entries - 1;
+	mutex_init(&cq->resize_mutex);
+	spin_lock_init(&cq->lock);
+	cq->resize_buf = NULL;
+	cq->resize_umem = NULL;
+	cq->create_flags = attr->flags;
+	INIT_LIST_HEAD(&cq->send_qp_list);
+	INIT_LIST_HEAD(&cq->recv_qp_list);
+
+	if (udata) {
+		struct mlx4_ib_create_cq ucmd;
+
+		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
+			err = -EFAULT;
+			goto err_cq;
+		}
+
+		buf_addr = (void *)(unsigned long)ucmd.buf_addr;
+		err = mlx4_ib_get_cq_umem(dev, udata, &cq->buf, &cq->umem,
+					  ucmd.buf_addr, entries);
+		if (err)
+			goto err_cq;
+
+		err = mlx4_ib_db_map_user(udata, ucmd.db_addr, &cq->db);
+		if (err)
+			goto err_mtt;
+
+		uar = &context->uar;
+		cq->mcq.usage = MLX4_RES_USAGE_USER_VERBS;
+	} else {
+		err = mlx4_db_alloc(dev->dev, &cq->db, 1);
+		if (err)
+			goto err_cq;
+
+		cq->mcq.set_ci_db  = cq->db.db;
+		cq->mcq.arm_db     = cq->db.db + 1;
+		*cq->mcq.set_ci_db = 0;
+		*cq->mcq.arm_db    = 0;
+
+		err = mlx4_ib_alloc_cq_buf(dev, &cq->buf, entries);
+		if (err)
+			goto err_db;
+
+		buf_addr = &cq->buf.buf;
+
+		uar = &dev->priv_uar;
+		cq->mcq.usage = MLX4_RES_USAGE_DRIVER;
+	}
+
+	if (dev->eq_table)
+		vector = dev->eq_table[vector % ibdev->num_comp_vectors];
+
+	err = mlx4_cq_alloc(dev->dev, entries, &cq->buf.mtt, uar, cq->db.dma,
+			    &cq->mcq, vector, 0,
+			    !!(cq->create_flags &
+			       IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION),
+			    buf_addr, !!udata);
+	if (err)
+		goto err_dbmap;
+
+	if (udata)
+		cq->mcq.tasklet_ctx.comp = mlx4_ib_cq_comp;
+	else
+		cq->mcq.comp = mlx4_ib_cq_comp;
+	cq->mcq.event = mlx4_ib_cq_event;
+
+	if (udata)
+		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof (__u32))) {
+			err = -EFAULT;
+			goto err_cq_free;
+		}
+
+	return 0;
+
+err_cq_free:
+	mlx4_cq_free(dev->dev, &cq->mcq);
+
+err_dbmap:
+	if (udata)
+		mlx4_ib_db_unmap_user(context, &cq->db);
+
+err_mtt:
+	mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
+
+	ib_umem_release(cq->umem);
+	if (!udata)
+		mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
+
+err_db:
+	if (!udata)
+		mlx4_db_free(dev->dev, &cq->db);
+err_cq:
+	return err;
+}
+
+static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
+				  int entries)
+{
+	int err;
+
+	if (cq->resize_buf)
+		return -EBUSY;
+
+	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_KERNEL);
+	if (!cq->resize_buf)
+		return -ENOMEM;
+
+	err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
+	if (err) {
+		kfree(cq->resize_buf);
+		cq->resize_buf = NULL;
+		return err;
+	}
+
+	cq->resize_buf->cqe = entries - 1;
+
+	return 0;
+}
+
+static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
+				   int entries, struct ib_udata *udata)
+{
+	struct mlx4_ib_resize_cq ucmd;
+	int err;
+
+	if (cq->resize_umem)
+		return -EBUSY;
+
+	if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
+		return -EFAULT;
+
+	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_KERNEL);
+	if (!cq->resize_buf)
+		return -ENOMEM;
+
+	err = mlx4_ib_get_cq_umem(dev, udata, &cq->resize_buf->buf,
+				  &cq->resize_umem, ucmd.buf_addr, entries);
+	if (err) {
+		kfree(cq->resize_buf);
+		cq->resize_buf = NULL;
+		return err;
+	}
+
+	cq->resize_buf->cqe = entries - 1;
+
+	return 0;
+}
+
+static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
+{
+	u32 i;
+
+	i = cq->mcq.cons_index;
+	while (get_sw_cqe(cq, i))
+		++i;
+
+	return i - cq->mcq.cons_index;
+}
+
+static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
+{
+	struct mlx4_cqe *cqe, *new_cqe;
+	int i;
+	int cqe_size = cq->buf.entry_size;
+	int cqe_inc = cqe_size == 64 ? 1 : 0;
+
+	i = cq->mcq.cons_index;
+	cqe = get_cqe(cq, i & cq->ibcq.cqe);
+	cqe += cqe_inc;
+
+	while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
+		new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
+					   (i + 1) & cq->resize_buf->cqe);
+		memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), cqe_size);
+		new_cqe += cqe_inc;
+
+		new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
+			(((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
+		cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
+		cqe += cqe_inc;
+	}
+	++cq->mcq.cons_index;
+}
+
+int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
+{
+	struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
+	struct mlx4_ib_cq *cq = to_mcq(ibcq);
+	struct mlx4_mtt mtt;
+	int outst_cqe;
+	int err;
+
+	mutex_lock(&cq->resize_mutex);
+	if (entries < 1 || entries > dev->dev->caps.max_cqes) {
+		err = -EINVAL;
+		goto out;
+	}
+
+	entries = roundup_pow_of_two(entries + 1);
+	if (entries == ibcq->cqe + 1) {
+		err = 0;
+		goto out;
+	}
+
+	if (entries > dev->dev->caps.max_cqes + 1) {
+		err = -EINVAL;
+		goto out;
+	}
+
+	if (ibcq->uobject) {
+		err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
+		if (err)
+			goto out;
+	} else {
+		/* Can't be smaller than the number of outstanding CQEs */
+		outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
+		if (entries < outst_cqe + 1) {
+			err = -EINVAL;
+			goto out;
+		}
+
+		err = mlx4_alloc_resize_buf(dev, cq, entries);
+		if (err)
+			goto out;
+	}
+
+	mtt = cq->buf.mtt;
+
+	err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
+	if (err)
+		goto err_buf;
+
+	mlx4_mtt_cleanup(dev->dev, &mtt);
+	if (ibcq->uobject) {
+		cq->buf      = cq->resize_buf->buf;
+		cq->ibcq.cqe = cq->resize_buf->cqe;
+		ib_umem_release(cq->umem);
+		cq->umem     = cq->resize_umem;
+
+		kfree(cq->resize_buf);
+		cq->resize_buf = NULL;
+		cq->resize_umem = NULL;
+	} else {
+		struct mlx4_ib_cq_buf tmp_buf;
+		int tmp_cqe = 0;
+
+		spin_lock_irq(&cq->lock);
+		if (cq->resize_buf) {
+			mlx4_ib_cq_resize_copy_cqes(cq);
+			tmp_buf = cq->buf;
+			tmp_cqe = cq->ibcq.cqe;
+			cq->buf      = cq->resize_buf->buf;
+			cq->ibcq.cqe = cq->resize_buf->cqe;
+
+			kfree(cq->resize_buf);
+			cq->resize_buf = NULL;
+		}
+		spin_unlock_irq(&cq->lock);
+
+		if (tmp_cqe)
+			mlx4_ib_free_cq_buf(dev, &tmp_buf, tmp_cqe);
+	}
+
+	goto out;
+
+err_buf:
+	mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
+	if (!ibcq->uobject)
+		mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
+				    cq->resize_buf->cqe);
+
+	kfree(cq->resize_buf);
+	cq->resize_buf = NULL;
+
+	ib_umem_release(cq->resize_umem);
+	cq->resize_umem = NULL;
+out:
+	mutex_unlock(&cq->resize_mutex);
+
+	return err;
+}
+
+int mlx4_ib_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
+{
+	struct mlx4_ib_dev *dev = to_mdev(cq->device);
+	struct mlx4_ib_cq *mcq = to_mcq(cq);
+
+	mlx4_cq_free(dev->dev, &mcq->mcq);
+	mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
+
+	if (udata) {
+		mlx4_ib_db_unmap_user(
+			rdma_udata_to_drv_context(
+				udata,
+				struct mlx4_ib_ucontext,
+				ibucontext),
+			&mcq->db);
+	} else {
+		mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
+		mlx4_db_free(dev->dev, &mcq->db);
+	}
+	ib_umem_release(mcq->umem);
+	return 0;
+}
+
+static void dump_cqe(void *cqe)
+{
+	__be32 *buf = cqe;
+
+	pr_debug("CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
+	       be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
+	       be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
+	       be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
+}
+
+static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
+				     struct ib_wc *wc)
+{
+	if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
+		pr_debug("local QP operation err "
+		       "(QPN %06x, WQE index %x, vendor syndrome %02x, "
+		       "opcode = %02x)\n",
+		       be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
+		       cqe->vendor_err_syndrome,
+		       cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
+		dump_cqe(cqe);
+	}
+
+	switch (cqe->syndrome) {
+	case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
+		wc->status = IB_WC_LOC_LEN_ERR;
+		break;
+	case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
+		wc->status = IB_WC_LOC_QP_OP_ERR;
+		break;
+	case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
+		wc->status = IB_WC_LOC_PROT_ERR;
+		break;
+	case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
+		wc->status = IB_WC_WR_FLUSH_ERR;
+		break;
+	case MLX4_CQE_SYNDROME_MW_BIND_ERR:
+		wc->status = IB_WC_MW_BIND_ERR;
+		break;
+	case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
+		wc->status = IB_WC_BAD_RESP_ERR;
+		break;
+	case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
+		wc->status = IB_WC_LOC_ACCESS_ERR;
+		break;
+	case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
+		wc->status = IB_WC_REM_INV_REQ_ERR;
+		break;
+	case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
+		wc->status = IB_WC_REM_ACCESS_ERR;
+		break;
+	case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
+		wc->status = IB_WC_REM_OP_ERR;
+		break;
+	case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
+		wc->status = IB_WC_RETRY_EXC_ERR;
+		break;
+	case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
+		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
+		break;
+	case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
+		wc->status = IB_WC_REM_ABORT_ERR;
+		break;
+	default:
+		wc->status = IB_WC_GENERAL_ERR;
+		break;
+	}
+
+	wc->vendor_err = cqe->vendor_err_syndrome;
+}
+
+static int mlx4_ib_ipoib_csum_ok(__be16 status, u8 badfcs_enc, __be16 checksum)
+{
+	return ((badfcs_enc & MLX4_CQE_STATUS_L4_CSUM) ||
+		((status & cpu_to_be16(MLX4_CQE_STATUS_IPOK)) &&
+		 (status & cpu_to_be16(MLX4_CQE_STATUS_TCP |
+				       MLX4_CQE_STATUS_UDP)) &&
+		 (checksum == cpu_to_be16(0xffff))));
+}
+
+static void use_tunnel_data(struct mlx4_ib_qp *qp, struct mlx4_ib_cq *cq, struct ib_wc *wc,
+			    unsigned tail, struct mlx4_cqe *cqe, int is_eth)
+{
+	struct mlx4_ib_proxy_sqp_hdr *hdr;
+
+	ib_dma_sync_single_for_cpu(qp->ibqp.device,
+				   qp->sqp_proxy_rcv[tail].map,
+				   sizeof (struct mlx4_ib_proxy_sqp_hdr),
+				   DMA_FROM_DEVICE);
+	hdr = (struct mlx4_ib_proxy_sqp_hdr *) (qp->sqp_proxy_rcv[tail].addr);
+	wc->pkey_index	= be16_to_cpu(hdr->tun.pkey_index);
+	wc->src_qp	= be32_to_cpu(hdr->tun.flags_src_qp) & 0xFFFFFF;
+	wc->wc_flags   |= (hdr->tun.g_ml_path & 0x80) ? (IB_WC_GRH) : 0;
+	wc->dlid_path_bits = 0;
+
+	if (is_eth) {
+		wc->slid = 0;
+		wc->vlan_id = be16_to_cpu(hdr->tun.sl_vid);
+		memcpy(&(wc->smac[0]), (char *)&hdr->tun.mac_31_0, 4);
+		memcpy(&(wc->smac[4]), (char *)&hdr->tun.slid_mac_47_32, 2);
+		wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
+	} else {
+		wc->slid        = be16_to_cpu(hdr->tun.slid_mac_47_32);
+		wc->sl          = (u8) (be16_to_cpu(hdr->tun.sl_vid) >> 12);
+	}
+}
+
+static void mlx4_ib_qp_sw_comp(struct mlx4_ib_qp *qp, int num_entries,
+			       struct ib_wc *wc, int *npolled, int is_send)
+{
+	struct mlx4_ib_wq *wq;
+	unsigned cur;
+	int i;
+
+	wq = is_send ? &qp->sq : &qp->rq;
+	cur = wq->head - wq->tail;
+
+	if (cur == 0)
+		return;
+
+	for (i = 0;  i < cur && *npolled < num_entries; i++) {
+		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
+		wc->status = IB_WC_WR_FLUSH_ERR;
+		wc->vendor_err = MLX4_CQE_SYNDROME_WR_FLUSH_ERR;
+		wq->tail++;
+		(*npolled)++;
+		wc->qp = &qp->ibqp;
+		wc++;
+	}
+}
+
+static void mlx4_ib_poll_sw_comp(struct mlx4_ib_cq *cq, int num_entries,
+				 struct ib_wc *wc, int *npolled)
+{
+	struct mlx4_ib_qp *qp;
+
+	*npolled = 0;
+	/* Find uncompleted WQEs belonging to that cq and return
+	 * simulated FLUSH_ERR completions
+	 */
+	list_for_each_entry(qp, &cq->send_qp_list, cq_send_list) {
+		mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 1);
+		if (*npolled >= num_entries)
+			goto out;
+	}
+
+	list_for_each_entry(qp, &cq->recv_qp_list, cq_recv_list) {
+		mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 0);
+		if (*npolled >= num_entries)
+			goto out;
+	}
+
+out:
+	return;
+}
+
+static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
+			    struct mlx4_ib_qp **cur_qp,
+			    struct ib_wc *wc)
+{
+	struct mlx4_cqe *cqe;
+	struct mlx4_qp *mqp;
+	struct mlx4_ib_wq *wq;
+	struct mlx4_ib_srq *srq;
+	struct mlx4_srq *msrq = NULL;
+	int is_send;
+	int is_error;
+	int is_eth;
+	u32 g_mlpath_rqpn;
+	u16 wqe_ctr;
+	unsigned tail = 0;
+
+repoll:
+	cqe = next_cqe_sw(cq);
+	if (!cqe)
+		return -EAGAIN;
+
+	if (cq->buf.entry_size == 64)
+		cqe++;
+
+	++cq->mcq.cons_index;
+
+	/*
+	 * Make sure we read CQ entry contents after we've checked the
+	 * ownership bit.
+	 */
+	rmb();
+
+	is_send  = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
+	is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
+		MLX4_CQE_OPCODE_ERROR;
+
+	/* Resize CQ in progress */
+	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_CQE_OPCODE_RESIZE)) {
+		if (cq->resize_buf) {
+			struct mlx4_ib_dev *dev = to_mdev(cq->ibcq.device);
+
+			mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
+			cq->buf      = cq->resize_buf->buf;
+			cq->ibcq.cqe = cq->resize_buf->cqe;
+
+			kfree(cq->resize_buf);
+			cq->resize_buf = NULL;
+		}
+
+		goto repoll;
+	}
+
+	if (!*cur_qp ||
+	    (be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) != (*cur_qp)->mqp.qpn) {
+		/*
+		 * We do not have to take the QP table lock here,
+		 * because CQs will be locked while QPs are removed
+		 * from the table.
+		 */
+		mqp = __mlx4_qp_lookup(to_mdev(cq->ibcq.device)->dev,
+				       be32_to_cpu(cqe->vlan_my_qpn));
+		*cur_qp = to_mibqp(mqp);
+	}
+
+	wc->qp = &(*cur_qp)->ibqp;
+
+	if (wc->qp->qp_type == IB_QPT_XRC_TGT) {
+		u32 srq_num;
+		g_mlpath_rqpn = be32_to_cpu(cqe->g_mlpath_rqpn);
+		srq_num       = g_mlpath_rqpn & 0xffffff;
+		/* SRQ is also in the radix tree */
+		msrq = mlx4_srq_lookup(to_mdev(cq->ibcq.device)->dev,
+				       srq_num);
+	}
+
+	if (is_send) {
+		wq = &(*cur_qp)->sq;
+		if (!(*cur_qp)->sq_signal_bits) {
+			wqe_ctr = be16_to_cpu(cqe->wqe_index);
+			wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
+		}
+		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
+		++wq->tail;
+	} else if ((*cur_qp)->ibqp.srq) {
+		srq = to_msrq((*cur_qp)->ibqp.srq);
+		wqe_ctr = be16_to_cpu(cqe->wqe_index);
+		wc->wr_id = srq->wrid[wqe_ctr];
+		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
+	} else if (msrq) {
+		srq = to_mibsrq(msrq);
+		wqe_ctr = be16_to_cpu(cqe->wqe_index);
+		wc->wr_id = srq->wrid[wqe_ctr];
+		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
+	} else {
+		wq	  = &(*cur_qp)->rq;
+		tail	  = wq->tail & (wq->wqe_cnt - 1);
+		wc->wr_id = wq->wrid[tail];
+		++wq->tail;
+	}
+
+	if (unlikely(is_error)) {
+		mlx4_ib_handle_error_cqe((struct mlx4_err_cqe *) cqe, wc);
+		return 0;
+	}
+
+	wc->status = IB_WC_SUCCESS;
+
+	if (is_send) {
+		wc->wc_flags = 0;
+		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
+		case MLX4_OPCODE_RDMA_WRITE_IMM:
+			wc->wc_flags |= IB_WC_WITH_IMM;
+			fallthrough;
+		case MLX4_OPCODE_RDMA_WRITE:
+			wc->opcode    = IB_WC_RDMA_WRITE;
+			break;
+		case MLX4_OPCODE_SEND_IMM:
+			wc->wc_flags |= IB_WC_WITH_IMM;
+			fallthrough;
+		case MLX4_OPCODE_SEND:
+		case MLX4_OPCODE_SEND_INVAL:
+			wc->opcode    = IB_WC_SEND;
+			break;
+		case MLX4_OPCODE_RDMA_READ:
+			wc->opcode    = IB_WC_RDMA_READ;
+			wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
+			break;
+		case MLX4_OPCODE_ATOMIC_CS:
+			wc->opcode    = IB_WC_COMP_SWAP;
+			wc->byte_len  = 8;
+			break;
+		case MLX4_OPCODE_ATOMIC_FA:
+			wc->opcode    = IB_WC_FETCH_ADD;
+			wc->byte_len  = 8;
+			break;
+		case MLX4_OPCODE_MASKED_ATOMIC_CS:
+			wc->opcode    = IB_WC_MASKED_COMP_SWAP;
+			wc->byte_len  = 8;
+			break;
+		case MLX4_OPCODE_MASKED_ATOMIC_FA:
+			wc->opcode    = IB_WC_MASKED_FETCH_ADD;
+			wc->byte_len  = 8;
+			break;
+		case MLX4_OPCODE_LSO:
+			wc->opcode    = IB_WC_LSO;
+			break;
+		case MLX4_OPCODE_FMR:
+			wc->opcode    = IB_WC_REG_MR;
+			break;
+		case MLX4_OPCODE_LOCAL_INVAL:
+			wc->opcode    = IB_WC_LOCAL_INV;
+			break;
+		}
+	} else {
+		wc->byte_len = be32_to_cpu(cqe->byte_cnt);
+
+		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
+		case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
+			wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
+			wc->wc_flags	= IB_WC_WITH_IMM;
+			wc->ex.imm_data = cqe->immed_rss_invalid;
+			break;
+		case MLX4_RECV_OPCODE_SEND_INVAL:
+			wc->opcode	= IB_WC_RECV;
+			wc->wc_flags	= IB_WC_WITH_INVALIDATE;
+			wc->ex.invalidate_rkey = be32_to_cpu(cqe->immed_rss_invalid);
+			break;
+		case MLX4_RECV_OPCODE_SEND:
+			wc->opcode   = IB_WC_RECV;
+			wc->wc_flags = 0;
+			break;
+		case MLX4_RECV_OPCODE_SEND_IMM:
+			wc->opcode	= IB_WC_RECV;
+			wc->wc_flags	= IB_WC_WITH_IMM;
+			wc->ex.imm_data = cqe->immed_rss_invalid;
+			break;
+		}
+
+		is_eth = (rdma_port_get_link_layer(wc->qp->device,
+						  (*cur_qp)->port) ==
+			  IB_LINK_LAYER_ETHERNET);
+		if (mlx4_is_mfunc(to_mdev(cq->ibcq.device)->dev)) {
+			if ((*cur_qp)->mlx4_ib_qp_type &
+			    (MLX4_IB_QPT_PROXY_SMI_OWNER |
+			     MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
+				use_tunnel_data(*cur_qp, cq, wc, tail, cqe,
+						is_eth);
+				return 0;
+			}
+		}
+
+		g_mlpath_rqpn	   = be32_to_cpu(cqe->g_mlpath_rqpn);
+		wc->src_qp	   = g_mlpath_rqpn & 0xffffff;
+		wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
+		wc->wc_flags	  |= g_mlpath_rqpn & 0x80000000 ? IB_WC_GRH : 0;
+		wc->pkey_index     = be32_to_cpu(cqe->immed_rss_invalid) & 0x7f;
+		wc->wc_flags	  |= mlx4_ib_ipoib_csum_ok(cqe->status,
+					cqe->badfcs_enc,
+					cqe->checksum) ? IB_WC_IP_CSUM_OK : 0;
+		if (is_eth) {
+			wc->slid = 0;
+			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 13;
+			if (be32_to_cpu(cqe->vlan_my_qpn) &
+					MLX4_CQE_CVLAN_PRESENT_MASK) {
+				wc->vlan_id = be16_to_cpu(cqe->sl_vid) &
+					MLX4_CQE_VID_MASK;
+			} else {
+				wc->vlan_id = 0xffff;
+			}
+			memcpy(wc->smac, cqe->smac, ETH_ALEN);
+			wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
+		} else {
+			wc->slid = be16_to_cpu(cqe->rlid);
+			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 12;
+			wc->vlan_id = 0xffff;
+		}
+	}
+
+	return 0;
+}
+
+int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
+{
+	struct mlx4_ib_cq *cq = to_mcq(ibcq);
+	struct mlx4_ib_qp *cur_qp = NULL;
+	unsigned long flags;
+	int npolled;
+	struct mlx4_ib_dev *mdev = to_mdev(cq->ibcq.device);
+
+	spin_lock_irqsave(&cq->lock, flags);
+	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
+		mlx4_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
+		goto out;
+	}
+
+	for (npolled = 0; npolled < num_entries; ++npolled) {
+		if (mlx4_ib_poll_one(cq, &cur_qp, wc + npolled))
+			break;
+	}
+
+	mlx4_cq_set_ci(&cq->mcq);
+
+out:
+	spin_unlock_irqrestore(&cq->lock, flags);
+
+	return npolled;
+}
+
+int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
+{
+	mlx4_cq_arm(&to_mcq(ibcq)->mcq,
+		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
+		    MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
+		    to_mdev(ibcq->device)->uar_map,
+		    MLX4_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->uar_lock));
+
+	return 0;
+}
+
+void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
+{
+	u32 prod_index;
+	int nfreed = 0;
+	struct mlx4_cqe *cqe, *dest;
+	u8 owner_bit;
+	int cqe_inc = cq->buf.entry_size == 64 ? 1 : 0;
+
+	/*
+	 * First we need to find the current producer index, so we
+	 * know where to start cleaning from.  It doesn't matter if HW
+	 * adds new entries after this loop -- the QP we're worried
+	 * about is already in RESET, so the new entries won't come
+	 * from our QP and therefore don't need to be checked.
+	 */
+	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
+		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
+			break;
+
+	/*
+	 * Now sweep backwards through the CQ, removing CQ entries
+	 * that match our QP by copying older entries on top of them.
+	 */
+	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
+		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
+		cqe += cqe_inc;
+
+		if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
+			if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
+				mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
+			++nfreed;
+		} else if (nfreed) {
+			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
+			dest += cqe_inc;
+
+			owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
+			memcpy(dest, cqe, sizeof *cqe);
+			dest->owner_sr_opcode = owner_bit |
+				(dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
+		}
+	}
+
+	if (nfreed) {
+		cq->mcq.cons_index += nfreed;
+		/*
+		 * Make sure update of buffer contents is done before
+		 * updating consumer index.
+		 */
+		wmb();
+		mlx4_cq_set_ci(&cq->mcq);
+	}
+}
+
+void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
+{
+	spin_lock_irq(&cq->lock);
+	__mlx4_ib_cq_clean(cq, qpn, srq);
+	spin_unlock_irq(&cq->lock);
+}
diff --git a/drivers/infiniband/hw/mlx4/doorbell.c b/drivers/infiniband/hw/mlx4/doorbell.c
new file mode 100644
index 000000000..d41f03ccb
--- /dev/null
+++ b/drivers/infiniband/hw/mlx4/doorbell.c
@@ -0,0 +1,99 @@
+/*
+ * Copyright (c) 2007 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
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/slab.h>
+#include <rdma/uverbs_ioctl.h>
+
+#include "mlx4_ib.h"
+
+struct mlx4_ib_user_db_page {
+	struct list_head	list;
+	struct ib_umem	       *umem;
+	unsigned long		user_virt;
+	int			refcnt;
+};
+
+int mlx4_ib_db_map_user(struct ib_udata *udata, unsigned long virt,
+			struct mlx4_db *db)
+{
+	struct mlx4_ib_user_db_page *page;
+	int err = 0;
+	struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context(
+		udata, struct mlx4_ib_ucontext, ibucontext);
+
+	mutex_lock(&context->db_page_mutex);
+
+	list_for_each_entry(page, &context->db_page_list, list)
+		if (page->user_virt == (virt & PAGE_MASK))
+			goto found;
+
+	page = kmalloc(sizeof *page, GFP_KERNEL);
+	if (!page) {
+		err = -ENOMEM;
+		goto out;
+	}
+
+	page->user_virt = (virt & PAGE_MASK);
+	page->refcnt    = 0;
+	page->umem = ib_umem_get(context->ibucontext.device, virt & PAGE_MASK,
+				 PAGE_SIZE, 0);
+	if (IS_ERR(page->umem)) {
+		err = PTR_ERR(page->umem);
+		kfree(page);
+		goto out;
+	}
+
+	list_add(&page->list, &context->db_page_list);
+
+found:
+	db->dma = sg_dma_address(page->umem->sg_head.sgl) + (virt & ~PAGE_MASK);
+	db->u.user_page = page;
+	++page->refcnt;
+
+out:
+	mutex_unlock(&context->db_page_mutex);
+
+	return err;
+}
+
+void mlx4_ib_db_unmap_user(struct mlx4_ib_ucontext *context, struct mlx4_db *db)
+{
+	mutex_lock(&context->db_page_mutex);
+
+	if (!--db->u.user_page->refcnt) {
+		list_del(&db->u.user_page->list);
+		ib_umem_release(db->u.user_page->umem);
+		kfree(db->u.user_page);
+	}
+
+	mutex_unlock(&context->db_page_mutex);
+}
diff --git a/drivers/infiniband/hw/mlx4/mad.c b/drivers/infiniband/hw/mlx4/mad.c
new file mode 100644
index 000000000..8bd164747
--- /dev/null
+++ b/drivers/infiniband/hw/mlx4/mad.c
@@ -0,0 +1,2396 @@
+/*
+ * Copyright (c) 2007 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
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <rdma/ib_mad.h>
+#include <rdma/ib_smi.h>
+#include <rdma/ib_sa.h>
+#include <rdma/ib_cache.h>
+
+#include <linux/random.h>
+#include <linux/mlx4/cmd.h>
+#include <linux/gfp.h>
+#include <rdma/ib_pma.h>
+#include <linux/ip.h>
+#include <net/ipv6.h>
+
+#include <linux/mlx4/driver.h>
+#include "mlx4_ib.h"
+
+enum {
+	MLX4_IB_VENDOR_CLASS1 = 0x9,
+	MLX4_IB_VENDOR_CLASS2 = 0xa
+};
+
+#define MLX4_TUN_SEND_WRID_SHIFT 34
+#define MLX4_TUN_QPN_SHIFT 32
+#define MLX4_TUN_WRID_RECV (((u64) 1) << MLX4_TUN_SEND_WRID_SHIFT)
+#define MLX4_TUN_SET_WRID_QPN(a) (((u64) ((a) & 0x3)) << MLX4_TUN_QPN_SHIFT)
+
+#define MLX4_TUN_IS_RECV(a)  (((a) >>  MLX4_TUN_SEND_WRID_SHIFT) & 0x1)
+#define MLX4_TUN_WRID_QPN(a) (((a) >> MLX4_TUN_QPN_SHIFT) & 0x3)
+
+ /* Port mgmt change event handling */
+
+#define GET_BLK_PTR_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.block_ptr)
+#define GET_MASK_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.tbl_entries_mask)
+#define NUM_IDX_IN_PKEY_TBL_BLK 32
+#define GUID_TBL_ENTRY_SIZE 8	   /* size in bytes */
+#define GUID_TBL_BLK_NUM_ENTRIES 8
+#define GUID_TBL_BLK_SIZE (GUID_TBL_ENTRY_SIZE * GUID_TBL_BLK_NUM_ENTRIES)
+
+struct mlx4_mad_rcv_buf {
+	struct ib_grh grh;
+	u8 payload[256];
+} __packed;
+
+struct mlx4_mad_snd_buf {
+	u8 payload[256];
+} __packed;
+
+struct mlx4_tunnel_mad {
+	struct ib_grh grh;
+	struct mlx4_ib_tunnel_header hdr;
+	struct ib_mad mad;
+} __packed;
+
+struct mlx4_rcv_tunnel_mad {
+	struct mlx4_rcv_tunnel_hdr hdr;
+	struct ib_grh grh;
+	struct ib_mad mad;
+} __packed;
+
+static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num);
+static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num);
+static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
+				int block, u32 change_bitmap);
+
+__be64 mlx4_ib_gen_node_guid(void)
+{
+#define NODE_GUID_HI	((u64) (((u64)IB_OPENIB_OUI) << 40))
+	return cpu_to_be64(NODE_GUID_HI | prandom_u32());
+}
+
+__be64 mlx4_ib_get_new_demux_tid(struct mlx4_ib_demux_ctx *ctx)
+{
+	return cpu_to_be64(atomic_inc_return(&ctx->tid)) |
+		cpu_to_be64(0xff00000000000000LL);
+}
+
+int mlx4_MAD_IFC(struct mlx4_ib_dev *dev, int mad_ifc_flags,
+		 int port, const struct ib_wc *in_wc,
+		 const struct ib_grh *in_grh,
+		 const void *in_mad, void *response_mad)
+{
+	struct mlx4_cmd_mailbox *inmailbox, *outmailbox;
+	void *inbox;
+	int err;
+	u32 in_modifier = port;
+	u8 op_modifier = 0;
+
+	inmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
+	if (IS_ERR(inmailbox))
+		return PTR_ERR(inmailbox);
+	inbox = inmailbox->buf;
+
+	outmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
+	if (IS_ERR(outmailbox)) {
+		mlx4_free_cmd_mailbox(dev->dev, inmailbox);
+		return PTR_ERR(outmailbox);
+	}
+
+	memcpy(inbox, in_mad, 256);
+
+	/*
+	 * Key check traps can't be generated unless we have in_wc to
+	 * tell us where to send the trap.
+	 */
+	if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_MKEY) || !in_wc)
+		op_modifier |= 0x1;
+	if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_BKEY) || !in_wc)
+		op_modifier |= 0x2;
+	if (mlx4_is_mfunc(dev->dev) &&
+	    (mad_ifc_flags & MLX4_MAD_IFC_NET_VIEW || in_wc))
+		op_modifier |= 0x8;
+
+	if (in_wc) {
+		struct {
+			__be32		my_qpn;
+			u32		reserved1;
+			__be32		rqpn;
+			u8		sl;
+			u8		g_path;
+			u16		reserved2[2];
+			__be16		pkey;
+			u32		reserved3[11];
+			u8		grh[40];
+		} *ext_info;
+
+		memset(inbox + 256, 0, 256);
+		ext_info = inbox + 256;
+
+		ext_info->my_qpn = cpu_to_be32(in_wc->qp->qp_num);
+		ext_info->rqpn   = cpu_to_be32(in_wc->src_qp);
+		ext_info->sl     = in_wc->sl << 4;
+		ext_info->g_path = in_wc->dlid_path_bits |
+			(in_wc->wc_flags & IB_WC_GRH ? 0x80 : 0);
+		ext_info->pkey   = cpu_to_be16(in_wc->pkey_index);
+
+		if (in_grh)
+			memcpy(ext_info->grh, in_grh, 40);
+
+		op_modifier |= 0x4;
+
+		in_modifier |= ib_lid_cpu16(in_wc->slid) << 16;
+	}
+
+	err = mlx4_cmd_box(dev->dev, inmailbox->dma, outmailbox->dma, in_modifier,
+			   mlx4_is_master(dev->dev) ? (op_modifier & ~0x8) : op_modifier,
+			   MLX4_CMD_MAD_IFC, MLX4_CMD_TIME_CLASS_C,
+			   (op_modifier & 0x8) ? MLX4_CMD_NATIVE : MLX4_CMD_WRAPPED);
+
+	if (!err)
+		memcpy(response_mad, outmailbox->buf, 256);
+
+	mlx4_free_cmd_mailbox(dev->dev, inmailbox);
+	mlx4_free_cmd_mailbox(dev->dev, outmailbox);
+
+	return err;
+}
+
+static void update_sm_ah(struct mlx4_ib_dev *dev, u8 port_num, u16 lid, u8 sl)
+{
+	struct ib_ah *new_ah;
+	struct rdma_ah_attr ah_attr;
+	unsigned long flags;
+
+	if (!dev->send_agent[port_num - 1][0])
+		return;
+
+	memset(&ah_attr, 0, sizeof ah_attr);
+	ah_attr.type = rdma_ah_find_type(&dev->ib_dev, port_num);
+	rdma_ah_set_dlid(&ah_attr, lid);
+	rdma_ah_set_sl(&ah_attr, sl);
+	rdma_ah_set_port_num(&ah_attr, port_num);
+
+	new_ah = rdma_create_ah(dev->send_agent[port_num - 1][0]->qp->pd,
+				&ah_attr, 0);
+	if (IS_ERR(new_ah))
+		return;
+
+	spin_lock_irqsave(&dev->sm_lock, flags);
+	if (dev->sm_ah[port_num - 1])
+		rdma_destroy_ah(dev->sm_ah[port_num - 1], 0);
+	dev->sm_ah[port_num - 1] = new_ah;
+	spin_unlock_irqrestore(&dev->sm_lock, flags);
+}
+
+/*
+ * Snoop SM MADs for port info, GUID info, and  P_Key table sets, so we can
+ * synthesize LID change, Client-Rereg, GID change, and P_Key change events.
+ */
+static void smp_snoop(struct ib_device *ibdev, u8 port_num, const struct ib_mad *mad,
+		      u16 prev_lid)
+{
+	struct ib_port_info *pinfo;
+	u16 lid;
+	__be16 *base;
+	u32 bn, pkey_change_bitmap;
+	int i;
+
+
+	struct mlx4_ib_dev *dev = to_mdev(ibdev);
+	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
+	     mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
+	    mad->mad_hdr.method == IB_MGMT_METHOD_SET)
+		switch (mad->mad_hdr.attr_id) {
+		case IB_SMP_ATTR_PORT_INFO:
+			if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV)
+				return;
+			pinfo = (struct ib_port_info *) ((struct ib_smp *) mad)->data;
+			lid = be16_to_cpu(pinfo->lid);
+
+			update_sm_ah(dev, port_num,
+				     be16_to_cpu(pinfo->sm_lid),
+				     pinfo->neighbormtu_mastersmsl & 0xf);
+
+			if (pinfo->clientrereg_resv_subnetto & 0x80)
+				handle_client_rereg_event(dev, port_num);
+
+			if (prev_lid != lid)
+				handle_lid_change_event(dev, port_num);
+			break;
+
+		case IB_SMP_ATTR_PKEY_TABLE:
+			if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV)
+				return;
+			if (!mlx4_is_mfunc(dev->dev)) {
+				mlx4_ib_dispatch_event(dev, port_num,
+						       IB_EVENT_PKEY_CHANGE);
+				break;
+			}
+
+			/* at this point, we are running in the master.
+			 * Slaves do not receive SMPs.
+			 */
+			bn  = be32_to_cpu(((struct ib_smp *)mad)->attr_mod) & 0xFFFF;
+			base = (__be16 *) &(((struct ib_smp *)mad)->data[0]);
+			pkey_change_bitmap = 0;
+			for (i = 0; i < 32; i++) {
+				pr_debug("PKEY[%d] = x%x\n",
+					 i + bn*32, be16_to_cpu(base[i]));
+				if (be16_to_cpu(base[i]) !=
+				    dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32]) {
+					pkey_change_bitmap |= (1 << i);
+					dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32] =
+						be16_to_cpu(base[i]);
+				}
+			}
+			pr_debug("PKEY Change event: port=%d, "
+				 "block=0x%x, change_bitmap=0x%x\n",
+				 port_num, bn, pkey_change_bitmap);
+
+			if (pkey_change_bitmap) {
+				mlx4_ib_dispatch_event(dev, port_num,
+						       IB_EVENT_PKEY_CHANGE);
+				if (!dev->sriov.is_going_down)
+					__propagate_pkey_ev(dev, port_num, bn,
+							    pkey_change_bitmap);
+			}
+			break;
+
+		case IB_SMP_ATTR_GUID_INFO:
+			if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV)
+				return;
+			/* paravirtualized master's guid is guid 0 -- does not change */
+			if (!mlx4_is_master(dev->dev))
+				mlx4_ib_dispatch_event(dev, port_num,
+						       IB_EVENT_GID_CHANGE);
+			/*if master, notify relevant slaves*/
+			if (mlx4_is_master(dev->dev) &&
+			    !dev->sriov.is_going_down) {
+				bn = be32_to_cpu(((struct ib_smp *)mad)->attr_mod);
+				mlx4_ib_update_cache_on_guid_change(dev, bn, port_num,
+								    (u8 *)(&((struct ib_smp *)mad)->data));
+				mlx4_ib_notify_slaves_on_guid_change(dev, bn, port_num,
+								     (u8 *)(&((struct ib_smp *)mad)->data));
+			}
+			break;
+
+		case IB_SMP_ATTR_SL_TO_VL_TABLE:
+			/* cache sl to vl mapping changes for use in
+			 * filling QP1 LRH VL field when sending packets
+			 */
+			if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV &&
+			    dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_SL_TO_VL_CHANGE_EVENT)
+				return;
+			if (!mlx4_is_slave(dev->dev)) {
+				union sl2vl_tbl_to_u64 sl2vl64;
+				int jj;
+
+				for (jj = 0; jj < 8; jj++) {
+					sl2vl64.sl8[jj] = ((struct ib_smp *)mad)->data[jj];
+					pr_debug("port %u, sl2vl[%d] = %02x\n",
+						 port_num, jj, sl2vl64.sl8[jj]);
+				}
+				atomic64_set(&dev->sl2vl[port_num - 1], sl2vl64.sl64);
+			}
+			break;
+
+		default:
+			break;
+		}
+}
+
+static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
+				int block, u32 change_bitmap)
+{
+	int i, ix, slave, err;
+	int have_event = 0;
+
+	for (slave = 0; slave < dev->dev->caps.sqp_demux; slave++) {
+		if (slave == mlx4_master_func_num(dev->dev))
+			continue;
+		if (!mlx4_is_slave_active(dev->dev, slave))
+			continue;
+
+		have_event = 0;
+		for (i = 0; i < 32; i++) {
+			if (!(change_bitmap & (1 << i)))
+				continue;
+			for (ix = 0;
+			     ix < dev->dev->caps.pkey_table_len[port_num]; ix++) {
+				if (dev->pkeys.virt2phys_pkey[slave][port_num - 1]
+				    [ix] == i + 32 * block) {
+					err = mlx4_gen_pkey_eqe(dev->dev, slave, port_num);
+					pr_debug("propagate_pkey_ev: slave %d,"
+						 " port %d, ix %d (%d)\n",
+						 slave, port_num, ix, err);
+					have_event = 1;
+					break;
+				}
+			}
+			if (have_event)
+				break;
+		}
+	}
+}
+
+static void node_desc_override(struct ib_device *dev,
+			       struct ib_mad *mad)
+{
+	unsigned long flags;
+
+	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
+	     mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
+	    mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
+	    mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
+		spin_lock_irqsave(&to_mdev(dev)->sm_lock, flags);
+		memcpy(((struct ib_smp *) mad)->data, dev->node_desc,
+		       IB_DEVICE_NODE_DESC_MAX);
+		spin_unlock_irqrestore(&to_mdev(dev)->sm_lock, flags);
+	}
+}
+
+static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, const struct ib_mad *mad)
+{
+	int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
+	struct ib_mad_send_buf *send_buf;
+	struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
+	int ret;
+	unsigned long flags;
+
+	if (agent) {
+		send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
+					      IB_MGMT_MAD_DATA, GFP_ATOMIC,
+					      IB_MGMT_BASE_VERSION);
+		if (IS_ERR(send_buf))
+			return;
+		/*
+		 * We rely here on the fact that MLX QPs don't use the
+		 * address handle after the send is posted (this is
+		 * wrong following the IB spec strictly, but we know
+		 * it's OK for our devices).
+		 */
+		spin_lock_irqsave(&dev->sm_lock, flags);
+		memcpy(send_buf->mad, mad, sizeof *mad);
+		if ((send_buf->ah = dev->sm_ah[port_num - 1]))
+			ret = ib_post_send_mad(send_buf, NULL);
+		else
+			ret = -EINVAL;
+		spin_unlock_irqrestore(&dev->sm_lock, flags);
+
+		if (ret)
+			ib_free_send_mad(send_buf);
+	}
+}
+
+static int mlx4_ib_demux_sa_handler(struct ib_device *ibdev, int port, int slave,
+							     struct ib_sa_mad *sa_mad)
+{
+	int ret = 0;
+
+	/* dispatch to different sa handlers */
+	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
+	case IB_SA_ATTR_MC_MEMBER_REC:
+		ret = mlx4_ib_mcg_demux_handler(ibdev, port, slave, sa_mad);
+		break;
+	default:
+		break;
+	}
+	return ret;
+}
+
+int mlx4_ib_find_real_gid(struct ib_device *ibdev, u8 port, __be64 guid)
+{
+	struct mlx4_ib_dev *dev = to_mdev(ibdev);
+	int i;
+
+	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
+		if (dev->sriov.demux[port - 1].guid_cache[i] == guid)
+			return i;
+	}
+	return -1;
+}
+
+
+static int find_slave_port_pkey_ix(struct mlx4_ib_dev *dev, int slave,
+				   u8 port, u16 pkey, u16 *ix)
+{
+	int i, ret;
+	u8 unassigned_pkey_ix, pkey_ix, partial_ix = 0xFF;
+	u16 slot_pkey;
+
+	if (slave == mlx4_master_func_num(dev->dev))
+		return ib_find_cached_pkey(&dev->ib_dev, port, pkey, ix);
+
+	unassigned_pkey_ix = dev->dev->phys_caps.pkey_phys_table_len[port] - 1;
+
+	for (i = 0; i < dev->dev->caps.pkey_table_len[port]; i++) {
+		if (dev->pkeys.virt2phys_pkey[slave][port - 1][i] == unassigned_pkey_ix)
+			continue;
+
+		pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][i];
+
+		ret = ib_get_cached_pkey(&dev->ib_dev, port, pkey_ix, &slot_pkey);
+		if (ret)
+			continue;
+		if ((slot_pkey & 0x7FFF) == (pkey & 0x7FFF)) {
+			if (slot_pkey & 0x8000) {
+				*ix = (u16) pkey_ix;
+				return 0;
+			} else {
+				/* take first partial pkey index found */
+				if (partial_ix == 0xFF)
+					partial_ix = pkey_ix;
+			}
+		}
+	}
+
+	if (partial_ix < 0xFF) {
+		*ix = (u16) partial_ix;
+		return 0;
+	}
+
+	return -EINVAL;
+}
+
+static int get_gids_from_l3_hdr(struct ib_grh *grh, union ib_gid *sgid,
+				union ib_gid *dgid)
+{
+	int version = ib_get_rdma_header_version((const union rdma_network_hdr *)grh);
+	enum rdma_network_type net_type;
+
+	if (version == 4)
+		net_type = RDMA_NETWORK_IPV4;
+	else if (version == 6)
+		net_type = RDMA_NETWORK_IPV6;
+	else
+		return -EINVAL;
+
+	return ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
+					 sgid, dgid);
+}
+
+static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
+{
+	int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
+
+	return (qpn >= proxy_start && qpn <= proxy_start + 1);
+}
+
+int mlx4_ib_send_to_slave(struct mlx4_ib_dev *dev, int slave, u8 port,
+			  enum ib_qp_type dest_qpt, struct ib_wc *wc,
+			  struct ib_grh *grh, struct ib_mad *mad)
+{
+	struct ib_sge list;
+	struct ib_ud_wr wr;
+	const struct ib_send_wr *bad_wr;
+	struct mlx4_ib_demux_pv_ctx *tun_ctx;
+	struct mlx4_ib_demux_pv_qp *tun_qp;
+	struct mlx4_rcv_tunnel_mad *tun_mad;
+	struct rdma_ah_attr attr;
+	struct ib_ah *ah;
+	struct ib_qp *src_qp = NULL;
+	unsigned tun_tx_ix = 0;
+	int dqpn;
+	int ret = 0;
+	u16 tun_pkey_ix;
+	u16 cached_pkey;
+	u8 is_eth = dev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
+
+	if (dest_qpt > IB_QPT_GSI) {
+		pr_debug("dest_qpt (%d) > IB_QPT_GSI\n", dest_qpt);
+		return -EINVAL;
+	}
+
+	tun_ctx = dev->sriov.demux[port-1].tun[slave];
+
+	/* check if proxy qp created */
+	if (!tun_ctx || tun_ctx->state != DEMUX_PV_STATE_ACTIVE)
+		return -EAGAIN;
+
+	if (!dest_qpt)
+		tun_qp = &tun_ctx->qp[0];
+	else
+		tun_qp = &tun_ctx->qp[1];
+
+	/* compute P_Key index to put in tunnel header for slave */
+	if (dest_qpt) {
+		u16 pkey_ix;
+		ret = ib_get_cached_pkey(&dev->ib_dev, port, wc->pkey_index, &cached_pkey);
+		if (ret) {
+			pr_debug("unable to get %s cached pkey for index %d, ret %d\n",
+				 is_proxy_qp0(dev, wc->src_qp, slave) ? "SMI" : "GSI",
+				 wc->pkey_index, ret);
+			return -EINVAL;
+		}
+
+		ret = find_slave_port_pkey_ix(dev, slave, port, cached_pkey, &pkey_ix);
+		if (ret) {
+			pr_debug("unable to get %s pkey ix for pkey 0x%x, ret %d\n",
+				 is_proxy_qp0(dev, wc->src_qp, slave) ? "SMI" : "GSI",
+				 cached_pkey, ret);
+			return -EINVAL;
+		}
+		tun_pkey_ix = pkey_ix;
+	} else
+		tun_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
+
+	dqpn = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave + port + (dest_qpt * 2) - 1;
+
+	/* get tunnel tx data buf for slave */
+	src_qp = tun_qp->qp;
+
+	/* create ah. Just need an empty one with the port num for the post send.
+	 * The driver will set the force loopback bit in post_send */
+	memset(&attr, 0, sizeof attr);
+	attr.type = rdma_ah_find_type(&dev->ib_dev, port);
+
+	rdma_ah_set_port_num(&attr, port);
+	if (is_eth) {
+		union ib_gid sgid;
+		union ib_gid dgid;
+
+		if (get_gids_from_l3_hdr(grh, &sgid, &dgid))
+			return -EINVAL;
+		rdma_ah_set_grh(&attr, &dgid, 0, 0, 0, 0);
+	}
+	ah = rdma_create_ah(tun_ctx->pd, &attr, 0);
+	if (IS_ERR(ah))
+		return -ENOMEM;
+
+	/* allocate tunnel tx buf after pass failure returns */
+	spin_lock(&tun_qp->tx_lock);
+	if (tun_qp->tx_ix_head - tun_qp->tx_ix_tail >=
+	    (MLX4_NUM_TUNNEL_BUFS - 1))
+		ret = -EAGAIN;
+	else
+		tun_tx_ix = (++tun_qp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
+	spin_unlock(&tun_qp->tx_lock);
+	if (ret)
+		goto end;
+
+	tun_mad = (struct mlx4_rcv_tunnel_mad *) (tun_qp->tx_ring[tun_tx_ix].buf.addr);
+	if (tun_qp->tx_ring[tun_tx_ix].ah)
+		rdma_destroy_ah(tun_qp->tx_ring[tun_tx_ix].ah, 0);
+	tun_qp->tx_ring[tun_tx_ix].ah = ah;
+	ib_dma_sync_single_for_cpu(&dev->ib_dev,
+				   tun_qp->tx_ring[tun_tx_ix].buf.map,
+				   sizeof (struct mlx4_rcv_tunnel_mad),
+				   DMA_TO_DEVICE);
+
+	/* copy over to tunnel buffer */
+	if (grh)
+		memcpy(&tun_mad->grh, grh, sizeof *grh);
+	memcpy(&tun_mad->mad, mad, sizeof *mad);
+
+	/* adjust tunnel data */
+	tun_mad->hdr.pkey_index = cpu_to_be16(tun_pkey_ix);
+	tun_mad->hdr.flags_src_qp = cpu_to_be32(wc->src_qp & 0xFFFFFF);
+	tun_mad->hdr.g_ml_path = (grh && (wc->wc_flags & IB_WC_GRH)) ? 0x80 : 0;
+
+	if (is_eth) {
+		u16 vlan = 0;
+		if (mlx4_get_slave_default_vlan(dev->dev, port, slave, &vlan,
+						NULL)) {
+			/* VST mode */
+			if (vlan != wc->vlan_id)
+				/* Packet vlan is not the VST-assigned vlan.
+				 * Drop the packet.
+				 */
+				goto out;
+			 else
+				/* Remove the vlan tag before forwarding
+				 * the packet to the VF.
+				 */
+				vlan = 0xffff;
+		} else {
+			vlan = wc->vlan_id;
+		}
+
+		tun_mad->hdr.sl_vid = cpu_to_be16(vlan);
+		memcpy((char *)&tun_mad->hdr.mac_31_0, &(wc->smac[0]), 4);
+		memcpy((char *)&tun_mad->hdr.slid_mac_47_32, &(wc->smac[4]), 2);
+	} else {
+		tun_mad->hdr.sl_vid = cpu_to_be16(((u16)(wc->sl)) << 12);
+		tun_mad->hdr.slid_mac_47_32 = ib_lid_be16(wc->slid);
+	}
+
+	ib_dma_sync_single_for_device(&dev->ib_dev,
+				      tun_qp->tx_ring[tun_tx_ix].buf.map,
+				      sizeof (struct mlx4_rcv_tunnel_mad),
+				      DMA_TO_DEVICE);
+
+	list.addr = tun_qp->tx_ring[tun_tx_ix].buf.map;
+	list.length = sizeof (struct mlx4_rcv_tunnel_mad);
+	list.lkey = tun_ctx->pd->local_dma_lkey;
+
+	wr.ah = ah;
+	wr.port_num = port;
+	wr.remote_qkey = IB_QP_SET_QKEY;
+	wr.remote_qpn = dqpn;
+	wr.wr.next = NULL;
+	wr.wr.wr_id = ((u64) tun_tx_ix) | MLX4_TUN_SET_WRID_QPN(dest_qpt);
+	wr.wr.sg_list = &list;
+	wr.wr.num_sge = 1;
+	wr.wr.opcode = IB_WR_SEND;
+	wr.wr.send_flags = IB_SEND_SIGNALED;
+
+	ret = ib_post_send(src_qp, &wr.wr, &bad_wr);
+	if (!ret)
+		return 0;
+ out:
+	spin_lock(&tun_qp->tx_lock);
+	tun_qp->tx_ix_tail++;
+	spin_unlock(&tun_qp->tx_lock);
+	tun_qp->tx_ring[tun_tx_ix].ah = NULL;
+end:
+	rdma_destroy_ah(ah, 0);
+	return ret;
+}
+
+static int mlx4_ib_demux_mad(struct ib_device *ibdev, u8 port,
+			struct ib_wc *wc, struct ib_grh *grh,
+			struct ib_mad *mad)
+{
+	struct mlx4_ib_dev *dev = to_mdev(ibdev);
+	int err, other_port;
+	int slave = -1;
+	u8 *slave_id;
+	int is_eth = 0;
+
+	if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
+		is_eth = 0;
+	else
+		is_eth = 1;
+
+	if (is_eth) {
+		union ib_gid dgid;
+		union ib_gid sgid;
+
+		if (get_gids_from_l3_hdr(grh, &sgid, &dgid))
+			return -EINVAL;
+		if (!(wc->wc_flags & IB_WC_GRH)) {
+			mlx4_ib_warn(ibdev, "RoCE grh not present.\n");
+			return -EINVAL;
+		}
+		if (mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_CM) {
+			mlx4_ib_warn(ibdev, "RoCE mgmt class is not CM\n");
+			return -EINVAL;
+		}
+		err = mlx4_get_slave_from_roce_gid(dev->dev, port, dgid.raw, &slave);
+		if (err && mlx4_is_mf_bonded(dev->dev)) {
+			other_port = (port == 1) ? 2 : 1;
+			err = mlx4_get_slave_from_roce_gid(dev->dev, other_port, dgid.raw, &slave);
+			if (!err) {
+				port = other_port;
+				pr_debug("resolved slave %d from gid %pI6 wire port %d other %d\n",
+					 slave, grh->dgid.raw, port, other_port);
+			}
+		}
+		if (err) {
+			mlx4_ib_warn(ibdev, "failed matching grh\n");
+			return -ENOENT;
+		}
+		if (slave >= dev->dev->caps.sqp_demux) {
+			mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
+				     slave, dev->dev->caps.sqp_demux);
+			return -ENOENT;
+		}
+
+		if (mlx4_ib_demux_cm_handler(ibdev, port, NULL, mad))
+			return 0;
+
+		err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
+		if (err)
+			pr_debug("failed sending %s to slave %d via tunnel qp (%d)\n",
+				 is_proxy_qp0(dev, wc->src_qp, slave) ? "SMI" : "GSI",
+				 slave, err);
+		return 0;
+	}
+
+	/* Initially assume that this mad is for us */
+	slave = mlx4_master_func_num(dev->dev);
+
+	/* See if the slave id is encoded in a response mad */
+	if (mad->mad_hdr.method & 0x80) {
+		slave_id = (u8 *) &mad->mad_hdr.tid;
+		slave = *slave_id;
+		if (slave != 255) /*255 indicates the dom0*/
+			*slave_id = 0; /* remap tid */
+	}
+
+	/* If a grh is present, we demux according to it */
+	if (wc->wc_flags & IB_WC_GRH) {
+		if (grh->dgid.global.interface_id ==
+			cpu_to_be64(IB_SA_WELL_KNOWN_GUID) &&
+		    grh->dgid.global.subnet_prefix == cpu_to_be64(
+			atomic64_read(&dev->sriov.demux[port - 1].subnet_prefix))) {
+			slave = 0;
+		} else {
+			slave = mlx4_ib_find_real_gid(ibdev, port,
+						      grh->dgid.global.interface_id);
+			if (slave < 0) {
+				mlx4_ib_warn(ibdev, "failed matching grh\n");
+				return -ENOENT;
+			}
+		}
+	}
+	/* Class-specific handling */
+	switch (mad->mad_hdr.mgmt_class) {
+	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
+	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
+		/* 255 indicates the dom0 */
+		if (slave != 255 && slave != mlx4_master_func_num(dev->dev)) {
+			if (!mlx4_vf_smi_enabled(dev->dev, slave, port))
+				return -EPERM;
+			/* for a VF. drop unsolicited MADs */
+			if (!(mad->mad_hdr.method & IB_MGMT_METHOD_RESP)) {
+				mlx4_ib_warn(ibdev, "demux QP0. rejecting unsolicited mad for slave %d class 0x%x, method 0x%x\n",
+					     slave, mad->mad_hdr.mgmt_class,
+					     mad->mad_hdr.method);
+				return -EINVAL;
+			}
+		}
+		break;
+	case IB_MGMT_CLASS_SUBN_ADM:
+		if (mlx4_ib_demux_sa_handler(ibdev, port, slave,
+					     (struct ib_sa_mad *) mad))
+			return 0;
+		break;
+	case IB_MGMT_CLASS_CM:
+		if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad))
+			return 0;
+		break;
+	case IB_MGMT_CLASS_DEVICE_MGMT:
+		if (mad->mad_hdr.method != IB_MGMT_METHOD_GET_RESP)
+			return 0;
+		break;
+	default:
+		/* Drop unsupported classes for slaves in tunnel mode */
+		if (slave != mlx4_master_func_num(dev->dev)) {
+			pr_debug("dropping unsupported ingress mad from class:%d "
+				 "for slave:%d\n", mad->mad_hdr.mgmt_class, slave);
+			return 0;
+		}
+	}
+	/*make sure that no slave==255 was not handled yet.*/
+	if (slave >= dev->dev->caps.sqp_demux) {
+		mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
+			     slave, dev->dev->caps.sqp_demux);
+		return -ENOENT;
+	}
+
+	err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
+	if (err)
+		pr_debug("failed sending %s to slave %d via tunnel qp (%d)\n",
+			 is_proxy_qp0(dev, wc->src_qp, slave) ? "SMI" : "GSI",
+			 slave, err);
+	return 0;
+}
+
+static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
+			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
+			const struct ib_mad *in_mad, struct ib_mad *out_mad)
+{
+	u16 slid, prev_lid = 0;
+	int err;
+	struct ib_port_attr pattr;
+
+	slid = in_wc ? ib_lid_cpu16(in_wc->slid) : be16_to_cpu(IB_LID_PERMISSIVE);
+
+	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
+		forward_trap(to_mdev(ibdev), port_num, in_mad);
+		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
+	}
+
+	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
+	    in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
+		if (in_mad->mad_hdr.method   != IB_MGMT_METHOD_GET &&
+		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_SET &&
+		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_TRAP_REPRESS)
+			return IB_MAD_RESULT_SUCCESS;
+
+		/*
+		 * Don't process SMInfo queries -- the SMA can't handle them.
+		 */
+		if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
+			return IB_MAD_RESULT_SUCCESS;
+	} else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
+		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1   ||
+		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2   ||
+		   in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
+		if (in_mad->mad_hdr.method  != IB_MGMT_METHOD_GET &&
+		    in_mad->mad_hdr.method  != IB_MGMT_METHOD_SET)
+			return IB_MAD_RESULT_SUCCESS;
+	} else
+		return IB_MAD_RESULT_SUCCESS;
+
+	if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
+	     in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
+	    in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
+	    in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
+	    !ib_query_port(ibdev, port_num, &pattr))
+		prev_lid = ib_lid_cpu16(pattr.lid);
+
+	err = mlx4_MAD_IFC(to_mdev(ibdev),
+			   (mad_flags & IB_MAD_IGNORE_MKEY ? MLX4_MAD_IFC_IGNORE_MKEY : 0) |
+			   (mad_flags & IB_MAD_IGNORE_BKEY ? MLX4_MAD_IFC_IGNORE_BKEY : 0) |
+			   MLX4_MAD_IFC_NET_VIEW,
+			   port_num, in_wc, in_grh, in_mad, out_mad);
+	if (err)
+		return IB_MAD_RESULT_FAILURE;
+
+	if (!out_mad->mad_hdr.status) {
+		smp_snoop(ibdev, port_num, in_mad, prev_lid);
+		/* slaves get node desc from FW */
+		if (!mlx4_is_slave(to_mdev(ibdev)->dev))
+			node_desc_override(ibdev, out_mad);
+	}
+
+	/* set return bit in status of directed route responses */
+	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
+		out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
+
+	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
+		/* no response for trap repress */
+		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
+
+	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
+}
+
+static void edit_counter(struct mlx4_counter *cnt, void *counters,
+			 __be16 attr_id)
+{
+	switch (attr_id) {
+	case IB_PMA_PORT_COUNTERS:
+	{
+		struct ib_pma_portcounters *pma_cnt =
+			(struct ib_pma_portcounters *)counters;
+
+		ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_data,
+				     (be64_to_cpu(cnt->tx_bytes) >> 2));
+		ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_data,
+				     (be64_to_cpu(cnt->rx_bytes) >> 2));
+		ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_packets,
+				     be64_to_cpu(cnt->tx_frames));
+		ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_packets,
+				     be64_to_cpu(cnt->rx_frames));
+		break;
+	}
+	case IB_PMA_PORT_COUNTERS_EXT:
+	{
+		struct ib_pma_portcounters_ext *pma_cnt_ext =
+			(struct ib_pma_portcounters_ext *)counters;
+
+		pma_cnt_ext->port_xmit_data =
+			cpu_to_be64(be64_to_cpu(cnt->tx_bytes) >> 2);
+		pma_cnt_ext->port_rcv_data =
+			cpu_to_be64(be64_to_cpu(cnt->rx_bytes) >> 2);
+		pma_cnt_ext->port_xmit_packets = cnt->tx_frames;
+		pma_cnt_ext->port_rcv_packets = cnt->rx_frames;
+		break;
+	}
+	}
+}
+
+static int iboe_process_mad_port_info(void *out_mad)
+{
+	struct ib_class_port_info cpi = {};
+
+	cpi.capability_mask = IB_PMA_CLASS_CAP_EXT_WIDTH;
+	memcpy(out_mad, &cpi, sizeof(cpi));
+	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
+}
+
+static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
+			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
+			const struct ib_mad *in_mad, struct ib_mad *out_mad)
+{
+	struct mlx4_counter counter_stats;
+	struct mlx4_ib_dev *dev = to_mdev(ibdev);
+	struct counter_index *tmp_counter;
+	int err = IB_MAD_RESULT_FAILURE, stats_avail = 0;
+
+	if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
+		return -EINVAL;
+
+	if (in_mad->mad_hdr.attr_id == IB_PMA_CLASS_PORT_INFO)
+		return iboe_process_mad_port_info((void *)(out_mad->data + 40));
+
+	memset(&counter_stats, 0, sizeof(counter_stats));
+	mutex_lock(&dev->counters_table[port_num - 1].mutex);
+	list_for_each_entry(tmp_counter,
+			    &dev->counters_table[port_num - 1].counters_list,
+			    list) {
+		err = mlx4_get_counter_stats(dev->dev,
+					     tmp_counter->index,
+					     &counter_stats, 0);
+		if (err) {
+			err = IB_MAD_RESULT_FAILURE;
+			stats_avail = 0;
+			break;
+		}
+		stats_avail = 1;
+	}
+	mutex_unlock(&dev->counters_table[port_num - 1].mutex);
+	if (stats_avail) {
+		switch (counter_stats.counter_mode & 0xf) {
+		case 0:
+			edit_counter(&counter_stats,
+				     (void *)(out_mad->data + 40),
+				     in_mad->mad_hdr.attr_id);
+			err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
+			break;
+		default:
+			err = IB_MAD_RESULT_FAILURE;
+		}
+	}
+
+	return err;
+}
+
+int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
+			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
+			const struct ib_mad *in, struct ib_mad *out,
+			size_t *out_mad_size, u16 *out_mad_pkey_index)
+{
+	struct mlx4_ib_dev *dev = to_mdev(ibdev);
+	enum rdma_link_layer link = rdma_port_get_link_layer(ibdev, port_num);
+
+	/* iboe_process_mad() which uses the HCA flow-counters to implement IB PMA
+	 * queries, should be called only by VFs and for that specific purpose
+	 */
+	if (link == IB_LINK_LAYER_INFINIBAND) {
+		if (mlx4_is_slave(dev->dev) &&
+		    (in->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT &&
+		     (in->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS ||
+		      in->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS_EXT ||
+		      in->mad_hdr.attr_id == IB_PMA_CLASS_PORT_INFO)))
+			return iboe_process_mad(ibdev, mad_flags, port_num,
+						in_wc, in_grh, in, out);
+
+		return ib_process_mad(ibdev, mad_flags, port_num, in_wc, in_grh,
+				      in, out);
+	}
+
+	if (link == IB_LINK_LAYER_ETHERNET)
+		return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
+					in_grh, in, out);
+
+	return -EINVAL;
+}
+
+static void send_handler(struct ib_mad_agent *agent,
+			 struct ib_mad_send_wc *mad_send_wc)
+{
+	if (mad_send_wc->send_buf->context[0])
+		rdma_destroy_ah(mad_send_wc->send_buf->context[0], 0);
+	ib_free_send_mad(mad_send_wc->send_buf);
+}
+
+int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
+{
+	struct ib_mad_agent *agent;
+	int p, q;
+	int ret;
+	enum rdma_link_layer ll;
+
+	for (p = 0; p < dev->num_ports; ++p) {
+		ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
+		for (q = 0; q <= 1; ++q) {
+			if (ll == IB_LINK_LAYER_INFINIBAND) {
+				agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
+							      q ? IB_QPT_GSI : IB_QPT_SMI,
+							      NULL, 0, send_handler,
+							      NULL, NULL, 0);
+				if (IS_ERR(agent)) {
+					ret = PTR_ERR(agent);
+					goto err;
+				}
+				dev->send_agent[p][q] = agent;
+			} else
+				dev->send_agent[p][q] = NULL;
+		}
+	}
+
+	return 0;
+
+err:
+	for (p = 0; p < dev->num_ports; ++p)
+		for (q = 0; q <= 1; ++q)
+			if (dev->send_agent[p][q])
+				ib_unregister_mad_agent(dev->send_agent[p][q]);
+
+	return ret;
+}
+
+void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
+{
+	struct ib_mad_agent *agent;
+	int p, q;
+
+	for (p = 0; p < dev->num_ports; ++p) {
+		for (q = 0; q <= 1; ++q) {
+			agent = dev->send_agent[p][q];
+			if (agent) {
+				dev->send_agent[p][q] = NULL;
+				ib_unregister_mad_agent(agent);
+			}
+		}
+
+		if (dev->sm_ah[p])
+			rdma_destroy_ah(dev->sm_ah[p], 0);
+	}
+}
+
+static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num)
+{
+	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
+
+	if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
+		mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
+					    MLX4_EQ_PORT_INFO_LID_CHANGE_MASK);
+}
+
+static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
+{
+	/* re-configure the alias-guid and mcg's */
+	if (mlx4_is_master(dev->dev)) {
+		mlx4_ib_invalidate_all_guid_record(dev, port_num);
+
+		if (!dev->sriov.is_going_down) {
+			mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
+			mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
+						    MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK);
+		}
+	}
+
+	/* Update the sl to vl table from inside client rereg
+	 * only if in secure-host mode (snooping is not possible)
+	 * and the sl-to-vl change event is not generated by FW.
+	 */
+	if (!mlx4_is_slave(dev->dev) &&
+	    dev->dev->flags & MLX4_FLAG_SECURE_HOST &&
+	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_SL_TO_VL_CHANGE_EVENT)) {
+		if (mlx4_is_master(dev->dev))
+			/* already in work queue from mlx4_ib_event queueing
+			 * mlx4_handle_port_mgmt_change_event, which calls
+			 * this procedure. Therefore, call sl2vl_update directly.
+			 */
+			mlx4_ib_sl2vl_update(dev, port_num);
+		else
+			mlx4_sched_ib_sl2vl_update_work(dev, port_num);
+	}
+	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
+}
+
+static void propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
+			      struct mlx4_eqe *eqe)
+{
+	__propagate_pkey_ev(dev, port_num, GET_BLK_PTR_FROM_EQE(eqe),
+			    GET_MASK_FROM_EQE(eqe));
+}
+
+static void handle_slaves_guid_change(struct mlx4_ib_dev *dev, u8 port_num,
+				      u32 guid_tbl_blk_num, u32 change_bitmap)
+{
+	struct ib_smp *in_mad  = NULL;
+	struct ib_smp *out_mad  = NULL;
+	u16 i;
+
+	if (!mlx4_is_mfunc(dev->dev) || !mlx4_is_master(dev->dev))
+		return;
+
+	in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
+	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
+	if (!in_mad || !out_mad)
+		goto out;
+
+	guid_tbl_blk_num  *= 4;
+
+	for (i = 0; i < 4; i++) {
+		if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
+			continue;
+		memset(in_mad, 0, sizeof *in_mad);
+		memset(out_mad, 0, sizeof *out_mad);
+
+		in_mad->base_version  = 1;
+		in_mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
+		in_mad->class_version = 1;
+		in_mad->method        = IB_MGMT_METHOD_GET;
+		in_mad->attr_id       = IB_SMP_ATTR_GUID_INFO;
+		in_mad->attr_mod      = cpu_to_be32(guid_tbl_blk_num + i);
+
+		if (mlx4_MAD_IFC(dev,
+				 MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
+				 port_num, NULL, NULL, in_mad, out_mad)) {
+			mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
+			goto out;
+		}
+
+		mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
+						    port_num,
+						    (u8 *)(&((struct ib_smp *)out_mad)->data));
+		mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
+						     port_num,
+						     (u8 *)(&((struct ib_smp *)out_mad)->data));
+	}
+
+out:
+	kfree(in_mad);
+	kfree(out_mad);
+	return;
+}
+
+void handle_port_mgmt_change_event(struct work_struct *work)
+{
+	struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
+	struct mlx4_ib_dev *dev = ew->ib_dev;
+	struct mlx4_eqe *eqe = &(ew->ib_eqe);
+	u8 port = eqe->event.port_mgmt_change.port;
+	u32 changed_attr;
+	u32 tbl_block;
+	u32 change_bitmap;
+
+	switch (eqe->subtype) {
+	case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
+		changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
+
+		/* Update the SM ah - This should be done before handling
+		   the other changed attributes so that MADs can be sent to the SM */
+		if (changed_attr & MSTR_SM_CHANGE_MASK) {
+			u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
+			u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
+			update_sm_ah(dev, port, lid, sl);
+		}
+
+		/* Check if it is a lid change event */
+		if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
+			handle_lid_change_event(dev, port);
+
+		/* Generate GUID changed event */
+		if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
+			if (mlx4_is_master(dev->dev)) {
+				union ib_gid gid;
+				int err = 0;
+
+				if (!eqe->event.port_mgmt_change.params.port_info.gid_prefix)
+					err = __mlx4_ib_query_gid(&dev->ib_dev, port, 0, &gid, 1);
+				else
+					gid.global.subnet_prefix =
+						eqe->event.port_mgmt_change.params.port_info.gid_prefix;
+				if (err) {
+					pr_warn("Could not change QP1 subnet prefix for port %d: query_gid error (%d)\n",
+						port, err);
+				} else {
+					pr_debug("Changing QP1 subnet prefix for port %d. old=0x%llx. new=0x%llx\n",
+						 port,
+						 (u64)atomic64_read(&dev->sriov.demux[port - 1].subnet_prefix),
+						 be64_to_cpu(gid.global.subnet_prefix));
+					atomic64_set(&dev->sriov.demux[port - 1].subnet_prefix,
+						     be64_to_cpu(gid.global.subnet_prefix));
+				}
+			}
+			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
+			/*if master, notify all slaves*/
+			if (mlx4_is_master(dev->dev))
+				mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
+							    MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK);
+		}
+
+		if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
+			handle_client_rereg_event(dev, port);
+		break;
+
+	case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
+		mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
+		if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
+			propagate_pkey_ev(dev, port, eqe);
+		break;
+	case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
+		/* paravirtualized master's guid is guid 0 -- does not change */
+		if (!mlx4_is_master(dev->dev))
+			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
+		/*if master, notify relevant slaves*/
+		else if (!dev->sriov.is_going_down) {
+			tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
+			change_bitmap = GET_MASK_FROM_EQE(eqe);
+			handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
+		}
+		break;
+
+	case MLX4_DEV_PMC_SUBTYPE_SL_TO_VL_MAP:
+		/* cache sl to vl mapping changes for use in
+		 * filling QP1 LRH VL field when sending packets
+		 */
+		if (!mlx4_is_slave(dev->dev)) {
+			union sl2vl_tbl_to_u64 sl2vl64;
+			int jj;
+
+			for (jj = 0; jj < 8; jj++) {
+				sl2vl64.sl8[jj] =
+					eqe->event.port_mgmt_change.params.sl2vl_tbl_change_info.sl2vl_table[jj];
+				pr_debug("port %u, sl2vl[%d] = %02x\n",
+					 port, jj, sl2vl64.sl8[jj]);
+			}
+			atomic64_set(&dev->sl2vl[port - 1], sl2vl64.sl64);
+		}
+		break;
+	default:
+		pr_warn("Unsupported subtype 0x%x for "
+			"Port Management Change event\n", eqe->subtype);
+	}
+
+	kfree(ew);
+}
+
+void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
+			    enum ib_event_type type)
+{
+	struct ib_event event;
+
+	event.device		= &dev->ib_dev;
+	event.element.port_num	= port_num;
+	event.event		= type;
+
+	ib_dispatch_event(&event);
+}
+
+static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
+{
+	unsigned long flags;
+	struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
+	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
+	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
+	if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
+		queue_work(ctx->wq, &ctx->work);
+	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
+}
+
+static void mlx4_ib_wire_comp_handler(struct ib_cq *cq, void *arg)
+{
+	unsigned long flags;
+	struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
+	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
+
+	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
+	if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
+		queue_work(ctx->wi_wq, &ctx->work);
+	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
+}
+
+static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
+				  struct mlx4_ib_demux_pv_qp *tun_qp,
+				  int index)
+{
+	struct ib_sge sg_list;
+	struct ib_recv_wr recv_wr;
+	const struct ib_recv_wr *bad_recv_wr;
+	int size;
+
+	size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
+		sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
+
+	sg_list.addr = tun_qp->ring[index].map;
+	sg_list.length = size;
+	sg_list.lkey = ctx->pd->local_dma_lkey;
+
+	recv_wr.next = NULL;
+	recv_wr.sg_list = &sg_list;
+	recv_wr.num_sge = 1;
+	recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
+		MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
+	ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
+				      size, DMA_FROM_DEVICE);
+	return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
+}
+
+static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
+		int slave, struct ib_sa_mad *sa_mad)
+{
+	int ret = 0;
+
+	/* dispatch to different sa handlers */
+	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
+	case IB_SA_ATTR_MC_MEMBER_REC:
+		ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
+		break;
+	default:
+		break;
+	}
+	return ret;
+}
+
+int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
+			 enum ib_qp_type dest_qpt, u16 pkey_index,
+			 u32 remote_qpn, u32 qkey, struct rdma_ah_attr *attr,
+			 u8 *s_mac, u16 vlan_id, struct ib_mad *mad)
+{
+	struct ib_sge list;
+	struct ib_ud_wr wr;
+	const struct ib_send_wr *bad_wr;
+	struct mlx4_ib_demux_pv_ctx *sqp_ctx;
+	struct mlx4_ib_demux_pv_qp *sqp;
+	struct mlx4_mad_snd_buf *sqp_mad;
+	struct ib_ah *ah;
+	struct ib_qp *send_qp = NULL;
+	unsigned wire_tx_ix = 0;
+	u16 wire_pkey_ix;
+	int src_qpnum;
+	int ret;
+
+	sqp_ctx = dev->sriov.sqps[port-1];
+
+	/* check if proxy qp created */
+	if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
+		return -EAGAIN;
+
+	if (dest_qpt == IB_QPT_SMI) {
+		src_qpnum = 0;
+		sqp = &sqp_ctx->qp[0];
+		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
+	} else {
+		src_qpnum = 1;
+		sqp = &sqp_ctx->qp[1];
+		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
+	}
+
+	send_qp = sqp->qp;
+
+	ah = rdma_zalloc_drv_obj(sqp_ctx->pd->device, ib_ah);
+	if (!ah)
+		return -ENOMEM;
+
+	ah->device = sqp_ctx->pd->device;
+	ah->pd = sqp_ctx->pd;
+
+	/* create ah */
+	ret = mlx4_ib_create_ah_slave(ah, attr,
+				      rdma_ah_retrieve_grh(attr)->sgid_index,
+				      s_mac, vlan_id);
+	if (ret)
+		goto out;
+
+	spin_lock(&sqp->tx_lock);
+	if (sqp->tx_ix_head - sqp->tx_ix_tail >=
+	    (MLX4_NUM_WIRE_BUFS - 1))
+		ret = -EAGAIN;
+	else
+		wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_WIRE_BUFS - 1);
+	spin_unlock(&sqp->tx_lock);
+	if (ret)
+		goto out;
+
+	sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
+	kfree(sqp->tx_ring[wire_tx_ix].ah);
+	sqp->tx_ring[wire_tx_ix].ah = ah;
+	ib_dma_sync_single_for_cpu(&dev->ib_dev,
+				   sqp->tx_ring[wire_tx_ix].buf.map,
+				   sizeof (struct mlx4_mad_snd_buf),
+				   DMA_TO_DEVICE);
+
+	memcpy(&sqp_mad->payload, mad, sizeof *mad);
+
+	ib_dma_sync_single_for_device(&dev->ib_dev,
+				      sqp->tx_ring[wire_tx_ix].buf.map,
+				      sizeof (struct mlx4_mad_snd_buf),
+				      DMA_TO_DEVICE);
+
+	list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
+	list.length = sizeof (struct mlx4_mad_snd_buf);
+	list.lkey = sqp_ctx->pd->local_dma_lkey;
+
+	wr.ah = ah;
+	wr.port_num = port;
+	wr.pkey_index = wire_pkey_ix;
+	wr.remote_qkey = qkey;
+	wr.remote_qpn = remote_qpn;
+	wr.wr.next = NULL;
+	wr.wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
+	wr.wr.sg_list = &list;
+	wr.wr.num_sge = 1;
+	wr.wr.opcode = IB_WR_SEND;
+	wr.wr.send_flags = IB_SEND_SIGNALED;
+
+	ret = ib_post_send(send_qp, &wr.wr, &bad_wr);
+	if (!ret)
+		return 0;
+
+	spin_lock(&sqp->tx_lock);
+	sqp->tx_ix_tail++;
+	spin_unlock(&sqp->tx_lock);
+	sqp->tx_ring[wire_tx_ix].ah = NULL;
+out:
+	kfree(ah);
+	return ret;
+}
+
+static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
+{
+	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
+		return slave;
+	return mlx4_get_base_gid_ix(dev->dev, slave, port);
+}
+
+static void fill_in_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
+				    struct rdma_ah_attr *ah_attr)
+{
+	struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
+	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
+		grh->sgid_index = slave;
+	else
+		grh->sgid_index += get_slave_base_gid_ix(dev, slave, port);
+}
+
+static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
+{
+	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
+	struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
+	int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
+	struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
+	struct mlx4_ib_ah ah;
+	struct rdma_ah_attr ah_attr;
+	u8 *slave_id;
+	int slave;
+	int port;
+	u16 vlan_id;
+	u8 qos;
+	u8 *dmac;
+	int sts;
+
+	/* Get slave that sent this packet */
+	if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
+	    wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
+	    (wc->src_qp & 0x1) != ctx->port - 1 ||
+	    wc->src_qp & 0x4) {
+		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
+		return;
+	}
+	slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
+	if (slave != ctx->slave) {
+		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
+			     "belongs to another slave\n", wc->src_qp);
+		return;
+	}
+
+	/* Map transaction ID */
+	ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
+				   sizeof (struct mlx4_tunnel_mad),
+				   DMA_FROM_DEVICE);
+	switch (tunnel->mad.mad_hdr.method) {
+	case IB_MGMT_METHOD_SET:
+	case IB_MGMT_METHOD_GET:
+	case IB_MGMT_METHOD_REPORT:
+	case IB_SA_METHOD_GET_TABLE:
+	case IB_SA_METHOD_DELETE:
+	case IB_SA_METHOD_GET_MULTI:
+	case IB_SA_METHOD_GET_TRACE_TBL:
+		slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
+		if (*slave_id) {
+			mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
+				     "class:%d slave:%d\n", *slave_id,
+				     tunnel->mad.mad_hdr.mgmt_class, slave);
+			return;
+		} else
+			*slave_id = slave;
+	default:
+		/* nothing */;
+	}
+
+	/* Class-specific handling */
+	switch (tunnel->mad.mad_hdr.mgmt_class) {
+	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
+	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
+		if (slave != mlx4_master_func_num(dev->dev) &&
+		    !mlx4_vf_smi_enabled(dev->dev, slave, ctx->port))
+			return;
+		break;
+	case IB_MGMT_CLASS_SUBN_ADM:
+		if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
+			      (struct ib_sa_mad *) &tunnel->mad))
+			return;
+		break;
+	case IB_MGMT_CLASS_CM:
+		if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
+			      (struct ib_mad *) &tunnel->mad))
+			return;
+		break;
+	case IB_MGMT_CLASS_DEVICE_MGMT:
+		if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
+		    tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
+			return;
+		break;
+	default:
+		/* Drop unsupported classes for slaves in tunnel mode */
+		if (slave != mlx4_master_func_num(dev->dev)) {
+			mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
+				     "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
+			return;
+		}
+	}
+
+	/* We are using standard ib_core services to send the mad, so generate a
+	 * stadard address handle by decoding the tunnelled mlx4_ah fields */
+	memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
+	ah.ibah.device = ctx->ib_dev;
+
+	port = be32_to_cpu(ah.av.ib.port_pd) >> 24;
+	port = mlx4_slave_convert_port(dev->dev, slave, port);
+	if (port < 0)
+		return;
+	ah.av.ib.port_pd = cpu_to_be32(port << 24 | (be32_to_cpu(ah.av.ib.port_pd) & 0xffffff));
+	ah.ibah.type = rdma_ah_find_type(&dev->ib_dev, port);
+
+	mlx4_ib_query_ah(&ah.ibah, &ah_attr);
+	if (rdma_ah_get_ah_flags(&ah_attr) & IB_AH_GRH)
+		fill_in_real_sgid_index(dev, slave, ctx->port, &ah_attr);
+	dmac = rdma_ah_retrieve_dmac(&ah_attr);
+	if (dmac)
+		memcpy(dmac, tunnel->hdr.mac, ETH_ALEN);
+	vlan_id = be16_to_cpu(tunnel->hdr.vlan);
+	/* if slave have default vlan use it */
+	if (mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
+					&vlan_id, &qos))
+		rdma_ah_set_sl(&ah_attr, qos);
+
+	sts = mlx4_ib_send_to_wire(dev, slave, ctx->port,
+				   is_proxy_qp0(dev, wc->src_qp, slave) ?
+				   IB_QPT_SMI : IB_QPT_GSI,
+				   be16_to_cpu(tunnel->hdr.pkey_index),
+				   be32_to_cpu(tunnel->hdr.remote_qpn),
+				   be32_to_cpu(tunnel->hdr.qkey),
+				   &ah_attr, wc->smac, vlan_id, &tunnel->mad);
+	if (sts)
+		pr_debug("failed sending %s to wire on behalf of slave %d (%d)\n",
+			 is_proxy_qp0(dev, wc->src_qp, slave) ? "SMI" : "GSI",
+			 slave, sts);
+}
+
+static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
+				 enum ib_qp_type qp_type, int is_tun)
+{
+	int i;
+	struct mlx4_ib_demux_pv_qp *tun_qp;
+	int rx_buf_size, tx_buf_size;
+	const int nmbr_bufs = is_tun ? MLX4_NUM_TUNNEL_BUFS : MLX4_NUM_WIRE_BUFS;
+
+	if (qp_type > IB_QPT_GSI)
+		return -EINVAL;
+
+	tun_qp = &ctx->qp[qp_type];
+
+	tun_qp->ring = kcalloc(nmbr_bufs,
+			       sizeof(struct mlx4_ib_buf),
+			       GFP_KERNEL);
+	if (!tun_qp->ring)
+		return -ENOMEM;
+
+	tun_qp->tx_ring = kcalloc(nmbr_bufs,
+				  sizeof (struct mlx4_ib_tun_tx_buf),
+				  GFP_KERNEL);
+	if (!tun_qp->tx_ring) {
+		kfree(tun_qp->ring);
+		tun_qp->ring = NULL;
+		return -ENOMEM;
+	}
+
+	if (is_tun) {
+		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
+		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
+	} else {
+		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
+		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
+	}
+
+	for (i = 0; i < nmbr_bufs; i++) {
+		tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
+		if (!tun_qp->ring[i].addr)
+			goto err;
+		tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
+							tun_qp->ring[i].addr,
+							rx_buf_size,
+							DMA_FROM_DEVICE);
+		if (ib_dma_mapping_error(ctx->ib_dev, tun_qp->ring[i].map)) {
+			kfree(tun_qp->ring[i].addr);
+			goto err;
+		}
+	}
+
+	for (i = 0; i < nmbr_bufs; i++) {
+		tun_qp->tx_ring[i].buf.addr =
+			kmalloc(tx_buf_size, GFP_KERNEL);
+		if (!tun_qp->tx_ring[i].buf.addr)
+			goto tx_err;
+		tun_qp->tx_ring[i].buf.map =
+			ib_dma_map_single(ctx->ib_dev,
+					  tun_qp->tx_ring[i].buf.addr,
+					  tx_buf_size,
+					  DMA_TO_DEVICE);
+		if (ib_dma_mapping_error(ctx->ib_dev,
+					 tun_qp->tx_ring[i].buf.map)) {
+			kfree(tun_qp->tx_ring[i].buf.addr);
+			goto tx_err;
+		}
+		tun_qp->tx_ring[i].ah = NULL;
+	}
+	spin_lock_init(&tun_qp->tx_lock);
+	tun_qp->tx_ix_head = 0;
+	tun_qp->tx_ix_tail = 0;
+	tun_qp->proxy_qpt = qp_type;
+
+	return 0;
+
+tx_err:
+	while (i > 0) {
+		--i;
+		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
+				    tx_buf_size, DMA_TO_DEVICE);
+		kfree(tun_qp->tx_ring[i].buf.addr);
+	}
+	i = nmbr_bufs;
+err:
+	while (i > 0) {
+		--i;
+		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
+				    rx_buf_size, DMA_FROM_DEVICE);
+		kfree(tun_qp->ring[i].addr);
+	}
+	kfree(tun_qp->tx_ring);
+	tun_qp->tx_ring = NULL;
+	kfree(tun_qp->ring);
+	tun_qp->ring = NULL;
+	return -ENOMEM;
+}
+
+static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
+				     enum ib_qp_type qp_type, int is_tun)
+{
+	int i;
+	struct mlx4_ib_demux_pv_qp *tun_qp;
+	int rx_buf_size, tx_buf_size;
+	const int nmbr_bufs = is_tun ? MLX4_NUM_TUNNEL_BUFS : MLX4_NUM_WIRE_BUFS;
+
+	if (qp_type > IB_QPT_GSI)
+		return;
+
+	tun_qp = &ctx->qp[qp_type];
+	if (is_tun) {
+		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
+		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
+	} else {
+		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
+		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
+	}
+
+
+	for (i = 0; i < nmbr_bufs; i++) {
+		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
+				    rx_buf_size, DMA_FROM_DEVICE);
+		kfree(tun_qp->ring[i].addr);
+	}
+
+	for (i = 0; i < nmbr_bufs; i++) {
+		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
+				    tx_buf_size, DMA_TO_DEVICE);
+		kfree(tun_qp->tx_ring[i].buf.addr);
+		if (tun_qp->tx_ring[i].ah)
+			rdma_destroy_ah(tun_qp->tx_ring[i].ah, 0);
+	}
+	kfree(tun_qp->tx_ring);
+	kfree(tun_qp->ring);
+}
+
+static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
+{
+	struct mlx4_ib_demux_pv_ctx *ctx;
+	struct mlx4_ib_demux_pv_qp *tun_qp;
+	struct ib_wc wc;
+	int ret;
+	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
+	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
+
+	while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
+		tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
+		if (wc.status == IB_WC_SUCCESS) {
+			switch (wc.opcode) {
+			case IB_WC_RECV:
+				mlx4_ib_multiplex_mad(ctx, &wc);
+				ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
+							     wc.wr_id &
+							     (MLX4_NUM_TUNNEL_BUFS - 1));
+				if (ret)
+					pr_err("Failed reposting tunnel "
+					       "buf:%lld\n", wc.wr_id);
+				break;
+			case IB_WC_SEND:
+				rdma_destroy_ah(tun_qp->tx_ring[wc.wr_id &
+					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah, 0);
+				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
+					= NULL;
+				spin_lock(&tun_qp->tx_lock);
+				tun_qp->tx_ix_tail++;
+				spin_unlock(&tun_qp->tx_lock);
+
+				break;
+			default:
+				break;
+			}
+		} else  {
+			pr_debug("mlx4_ib: completion error in tunnel: %d."
+				 " status = %d, wrid = 0x%llx\n",
+				 ctx->slave, wc.status, wc.wr_id);
+			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
+				rdma_destroy_ah(tun_qp->tx_ring[wc.wr_id &
+					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah, 0);
+				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
+					= NULL;
+				spin_lock(&tun_qp->tx_lock);
+				tun_qp->tx_ix_tail++;
+				spin_unlock(&tun_qp->tx_lock);
+			}
+		}
+	}
+}
+
+static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
+{
+	struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
+
+	/* It's worse than that! He's dead, Jim! */
+	pr_err("Fatal error (%d) on a MAD QP on port %d\n",
+	       event->event, sqp->port);
+}
+
+static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
+			    enum ib_qp_type qp_type, int create_tun)
+{
+	int i, ret;
+	struct mlx4_ib_demux_pv_qp *tun_qp;
+	struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
+	struct ib_qp_attr attr;
+	int qp_attr_mask_INIT;
+	const int nmbr_bufs = create_tun ? MLX4_NUM_TUNNEL_BUFS : MLX4_NUM_WIRE_BUFS;
+
+	if (qp_type > IB_QPT_GSI)
+		return -EINVAL;
+
+	tun_qp = &ctx->qp[qp_type];
+
+	memset(&qp_init_attr, 0, sizeof qp_init_attr);
+	qp_init_attr.init_attr.send_cq = ctx->cq;
+	qp_init_attr.init_attr.recv_cq = ctx->cq;
+	qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
+	qp_init_attr.init_attr.cap.max_send_wr = nmbr_bufs;
+	qp_init_attr.init_attr.cap.max_recv_wr = nmbr_bufs;
+	qp_init_attr.init_attr.cap.max_send_sge = 1;
+	qp_init_attr.init_attr.cap.max_recv_sge = 1;
+	if (create_tun) {
+		qp_init_attr.init_attr.qp_type = IB_QPT_UD;
+		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
+		qp_init_attr.port = ctx->port;
+		qp_init_attr.slave = ctx->slave;
+		qp_init_attr.proxy_qp_type = qp_type;
+		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
+			   IB_QP_QKEY | IB_QP_PORT;
+	} else {
+		qp_init_attr.init_attr.qp_type = qp_type;
+		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
+		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
+	}
+	qp_init_attr.init_attr.port_num = ctx->port;
+	qp_init_attr.init_attr.qp_context = ctx;
+	qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
+	tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
+	if (IS_ERR(tun_qp->qp)) {
+		ret = PTR_ERR(tun_qp->qp);
+		tun_qp->qp = NULL;
+		pr_err("Couldn't create %s QP (%d)\n",
+		       create_tun ? "tunnel" : "special", ret);
+		return ret;
+	}
+
+	memset(&attr, 0, sizeof attr);
+	attr.qp_state = IB_QPS_INIT;
+	ret = 0;
+	if (create_tun)
+		ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
+					      ctx->port, IB_DEFAULT_PKEY_FULL,
+					      &attr.pkey_index);
+	if (ret || !create_tun)
+		attr.pkey_index =
+			to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
+	attr.qkey = IB_QP1_QKEY;
+	attr.port_num = ctx->port;
+	ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
+	if (ret) {
+		pr_err("Couldn't change %s qp state to INIT (%d)\n",
+		       create_tun ? "tunnel" : "special", ret);
+		goto err_qp;
+	}
+	attr.qp_state = IB_QPS_RTR;
+	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
+	if (ret) {
+		pr_err("Couldn't change %s qp state to RTR (%d)\n",
+		       create_tun ? "tunnel" : "special", ret);
+		goto err_qp;
+	}
+	attr.qp_state = IB_QPS_RTS;
+	attr.sq_psn = 0;
+	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
+	if (ret) {
+		pr_err("Couldn't change %s qp state to RTS (%d)\n",
+		       create_tun ? "tunnel" : "special", ret);
+		goto err_qp;
+	}
+
+	for (i = 0; i < nmbr_bufs; i++) {
+		ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
+		if (ret) {
+			pr_err(" mlx4_ib_post_pv_buf error"
+			       " (err = %d, i = %d)\n", ret, i);
+			goto err_qp;
+		}
+	}
+	return 0;
+
+err_qp:
+	ib_destroy_qp(tun_qp->qp);
+	tun_qp->qp = NULL;
+	return ret;
+}
+
+/*
+ * IB MAD completion callback for real SQPs
+ */
+static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
+{
+	struct mlx4_ib_demux_pv_ctx *ctx;
+	struct mlx4_ib_demux_pv_qp *sqp;
+	struct ib_wc wc;
+	struct ib_grh *grh;
+	struct ib_mad *mad;
+
+	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
+	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
+
+	while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
+		sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
+		if (wc.status == IB_WC_SUCCESS) {
+			switch (wc.opcode) {
+			case IB_WC_SEND:
+				kfree(sqp->tx_ring[wc.wr_id &
+				      (MLX4_NUM_WIRE_BUFS - 1)].ah);
+				sqp->tx_ring[wc.wr_id & (MLX4_NUM_WIRE_BUFS - 1)].ah
+					= NULL;
+				spin_lock(&sqp->tx_lock);
+				sqp->tx_ix_tail++;
+				spin_unlock(&sqp->tx_lock);
+				break;
+			case IB_WC_RECV:
+				mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
+						(sqp->ring[wc.wr_id &
+						(MLX4_NUM_WIRE_BUFS - 1)].addr))->payload);
+				grh = &(((struct mlx4_mad_rcv_buf *)
+						(sqp->ring[wc.wr_id &
+						(MLX4_NUM_WIRE_BUFS - 1)].addr))->grh);
+				mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
+				if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
+							   (MLX4_NUM_WIRE_BUFS - 1)))
+					pr_err("Failed reposting SQP "
+					       "buf:%lld\n", wc.wr_id);
+				break;
+			default:
+				break;
+			}
+		} else  {
+			pr_debug("mlx4_ib: completion error in tunnel: %d."
+				 " status = %d, wrid = 0x%llx\n",
+				 ctx->slave, wc.status, wc.wr_id);
+			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
+				kfree(sqp->tx_ring[wc.wr_id &
+				      (MLX4_NUM_WIRE_BUFS - 1)].ah);
+				sqp->tx_ring[wc.wr_id & (MLX4_NUM_WIRE_BUFS - 1)].ah
+					= NULL;
+				spin_lock(&sqp->tx_lock);
+				sqp->tx_ix_tail++;
+				spin_unlock(&sqp->tx_lock);
+			}
+		}
+	}
+}
+
+static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
+			       struct mlx4_ib_demux_pv_ctx **ret_ctx)
+{
+	struct mlx4_ib_demux_pv_ctx *ctx;
+
+	*ret_ctx = NULL;
+	ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
+	if (!ctx)
+		return -ENOMEM;
+
+	ctx->ib_dev = &dev->ib_dev;
+	ctx->port = port;
+	ctx->slave = slave;
+	*ret_ctx = ctx;
+	return 0;
+}
+
+static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
+{
+	if (dev->sriov.demux[port - 1].tun[slave]) {
+		kfree(dev->sriov.demux[port - 1].tun[slave]);
+		dev->sriov.demux[port - 1].tun[slave] = NULL;
+	}
+}
+
+static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
+			       int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
+{
+	int ret, cq_size;
+	struct ib_cq_init_attr cq_attr = {};
+	const int nmbr_bufs = create_tun ? MLX4_NUM_TUNNEL_BUFS : MLX4_NUM_WIRE_BUFS;
+
+	if (ctx->state != DEMUX_PV_STATE_DOWN)
+		return -EEXIST;
+
+	ctx->state = DEMUX_PV_STATE_STARTING;
+	/* have QP0 only if link layer is IB */
+	if (rdma_port_get_link_layer(ibdev, ctx->port) ==
+	    IB_LINK_LAYER_INFINIBAND)
+		ctx->has_smi = 1;
+
+	if (ctx->has_smi) {
+		ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
+		if (ret) {
+			pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
+			goto err_out;
+		}
+	}
+
+	ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
+	if (ret) {
+		pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
+		goto err_out_qp0;
+	}
+
+	cq_size = 2 * nmbr_bufs;
+	if (ctx->has_smi)
+		cq_size *= 2;
+
+	cq_attr.cqe = cq_size;
+	ctx->cq = ib_create_cq(ctx->ib_dev,
+			       create_tun ? mlx4_ib_tunnel_comp_handler : mlx4_ib_wire_comp_handler,
+			       NULL, ctx, &cq_attr);
+	if (IS_ERR(ctx->cq)) {
+		ret = PTR_ERR(ctx->cq);
+		pr_err("Couldn't create tunnel CQ (%d)\n", ret);
+		goto err_buf;
+	}
+
+	ctx->pd = ib_alloc_pd(ctx->ib_dev, 0);
+	if (IS_ERR(ctx->pd)) {
+		ret = PTR_ERR(ctx->pd);
+		pr_err("Couldn't create tunnel PD (%d)\n", ret);
+		goto err_cq;
+	}
+
+	if (ctx->has_smi) {
+		ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
+		if (ret) {
+			pr_err("Couldn't create %s QP0 (%d)\n",
+			       create_tun ? "tunnel for" : "",  ret);
+			goto err_pd;
+		}
+	}
+
+	ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
+	if (ret) {
+		pr_err("Couldn't create %s QP1 (%d)\n",
+		       create_tun ? "tunnel for" : "",  ret);
+		goto err_qp0;
+	}
+
+	if (create_tun)
+		INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
+	else
+		INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
+
+	ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
+	ctx->wi_wq = to_mdev(ibdev)->sriov.demux[port - 1].wi_wq;
+
+	ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
+	if (ret) {
+		pr_err("Couldn't arm tunnel cq (%d)\n", ret);
+		goto err_wq;
+	}
+	ctx->state = DEMUX_PV_STATE_ACTIVE;
+	return 0;
+
+err_wq:
+	ctx->wq = NULL;
+	ib_destroy_qp(ctx->qp[1].qp);
+	ctx->qp[1].qp = NULL;
+
+
+err_qp0:
+	if (ctx->has_smi)
+		ib_destroy_qp(ctx->qp[0].qp);
+	ctx->qp[0].qp = NULL;
+
+err_pd:
+	ib_dealloc_pd(ctx->pd);
+	ctx->pd = NULL;
+
+err_cq:
+	ib_destroy_cq(ctx->cq);
+	ctx->cq = NULL;
+
+err_buf:
+	mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
+
+err_out_qp0:
+	if (ctx->has_smi)
+		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
+err_out:
+	ctx->state = DEMUX_PV_STATE_DOWN;
+	return ret;
+}
+
+static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
+				 struct mlx4_ib_demux_pv_ctx *ctx, int flush)
+{
+	if (!ctx)
+		return;
+	if (ctx->state > DEMUX_PV_STATE_DOWN) {
+		ctx->state = DEMUX_PV_STATE_DOWNING;
+		if (flush)
+			flush_workqueue(ctx->wq);
+		if (ctx->has_smi) {
+			ib_destroy_qp(ctx->qp[0].qp);
+			ctx->qp[0].qp = NULL;
+			mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
+		}
+		ib_destroy_qp(ctx->qp[1].qp);
+		ctx->qp[1].qp = NULL;
+		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
+		ib_dealloc_pd(ctx->pd);
+		ctx->pd = NULL;
+		ib_destroy_cq(ctx->cq);
+		ctx->cq = NULL;
+		ctx->state = DEMUX_PV_STATE_DOWN;
+	}
+}
+
+static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
+				  int port, int do_init)
+{
+	int ret = 0;
+
+	if (!do_init) {
+		clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
+		/* for master, destroy real sqp resources */
+		if (slave == mlx4_master_func_num(dev->dev))
+			destroy_pv_resources(dev, slave, port,
+					     dev->sriov.sqps[port - 1], 1);
+		/* destroy the tunnel qp resources */
+		destroy_pv_resources(dev, slave, port,
+				     dev->sriov.demux[port - 1].tun[slave], 1);
+		return 0;
+	}
+
+	/* create the tunnel qp resources */
+	ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
+				  dev->sriov.demux[port - 1].tun[slave]);
+
+	/* for master, create the real sqp resources */
+	if (!ret && slave == mlx4_master_func_num(dev->dev))
+		ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
+					  dev->sriov.sqps[port - 1]);
+	return ret;
+}
+
+void mlx4_ib_tunnels_update_work(struct work_struct *work)
+{
+	struct mlx4_ib_demux_work *dmxw;
+
+	dmxw = container_of(work, struct mlx4_ib_demux_work, work);
+	mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
+			       dmxw->do_init);
+	kfree(dmxw);
+	return;
+}
+
+static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
+				       struct mlx4_ib_demux_ctx *ctx,
+				       int port)
+{
+	char name[12];
+	int ret = 0;
+	int i;
+
+	ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
+			   sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
+	if (!ctx->tun)
+		return -ENOMEM;
+
+	ctx->dev = dev;
+	ctx->port = port;
+	ctx->ib_dev = &dev->ib_dev;
+
+	for (i = 0;
+	     i < min(dev->dev->caps.sqp_demux,
+	     (u16)(dev->dev->persist->num_vfs + 1));
+	     i++) {
+		struct mlx4_active_ports actv_ports =
+			mlx4_get_active_ports(dev->dev, i);
+
+		if (!test_bit(port - 1, actv_ports.ports))
+			continue;
+
+		ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
+		if (ret) {
+			ret = -ENOMEM;
+			goto err_mcg;
+		}
+	}
+
+	ret = mlx4_ib_mcg_port_init(ctx);
+	if (ret) {
+		pr_err("Failed initializing mcg para-virt (%d)\n", ret);
+		goto err_mcg;
+	}
+
+	snprintf(name, sizeof(name), "mlx4_ibt%d", port);
+	ctx->wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
+	if (!ctx->wq) {
+		pr_err("Failed to create tunnelling WQ for port %d\n", port);
+		ret = -ENOMEM;
+		goto err_wq;
+	}
+
+	snprintf(name, sizeof(name), "mlx4_ibwi%d", port);
+	ctx->wi_wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
+	if (!ctx->wi_wq) {
+		pr_err("Failed to create wire WQ for port %d\n", port);
+		ret = -ENOMEM;
+		goto err_wiwq;
+	}
+
+	snprintf(name, sizeof(name), "mlx4_ibud%d", port);
+	ctx->ud_wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
+	if (!ctx->ud_wq) {
+		pr_err("Failed to create up/down WQ for port %d\n", port);
+		ret = -ENOMEM;
+		goto err_udwq;
+	}
+
+	return 0;
+
+err_udwq:
+	destroy_workqueue(ctx->wi_wq);
+	ctx->wi_wq = NULL;
+
+err_wiwq:
+	destroy_workqueue(ctx->wq);
+	ctx->wq = NULL;
+
+err_wq:
+	mlx4_ib_mcg_port_cleanup(ctx, 1);
+err_mcg:
+	for (i = 0; i < dev->dev->caps.sqp_demux; i++)
+		free_pv_object(dev, i, port);
+	kfree(ctx->tun);
+	ctx->tun = NULL;
+	return ret;
+}
+
+static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
+{
+	if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
+		sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
+		flush_workqueue(sqp_ctx->wq);
+		if (sqp_ctx->has_smi) {
+			ib_destroy_qp(sqp_ctx->qp[0].qp);
+			sqp_ctx->qp[0].qp = NULL;
+			mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
+		}
+		ib_destroy_qp(sqp_ctx->qp[1].qp);
+		sqp_ctx->qp[1].qp = NULL;
+		mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
+		ib_dealloc_pd(sqp_ctx->pd);
+		sqp_ctx->pd = NULL;
+		ib_destroy_cq(sqp_ctx->cq);
+		sqp_ctx->cq = NULL;
+		sqp_ctx->state = DEMUX_PV_STATE_DOWN;
+	}
+}
+
+static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
+{
+	int i;
+	if (ctx) {
+		struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
+		mlx4_ib_mcg_port_cleanup(ctx, 1);
+		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
+			if (!ctx->tun[i])
+				continue;
+			if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
+				ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
+		}
+		flush_workqueue(ctx->wq);
+		flush_workqueue(ctx->wi_wq);
+		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
+			destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
+			free_pv_object(dev, i, ctx->port);
+		}
+		kfree(ctx->tun);
+		destroy_workqueue(ctx->ud_wq);
+		destroy_workqueue(ctx->wi_wq);
+		destroy_workqueue(ctx->wq);
+	}
+}
+
+static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
+{
+	int i;
+
+	if (!mlx4_is_master(dev->dev))
+		return;
+	/* initialize or tear down tunnel QPs for the master */
+	for (i = 0; i < dev->dev->caps.num_ports; i++)
+		mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
+	return;
+}
+
+int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
+{
+	int i = 0;
+	int err;
+
+	if (!mlx4_is_mfunc(dev->dev))
+		return 0;
+
+	dev->sriov.is_going_down = 0;
+	spin_lock_init(&dev->sriov.going_down_lock);
+	mlx4_ib_cm_paravirt_init(dev);
+
+	mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
+
+	if (mlx4_is_slave(dev->dev)) {
+		mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
+		return 0;
+	}
+
+	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
+		if (i == mlx4_master_func_num(dev->dev))
+			mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
+		else
+			mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
+	}
+
+	err = mlx4_ib_init_alias_guid_service(dev);
+	if (err) {
+		mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
+		goto paravirt_err;
+	}
+	err = mlx4_ib_device_register_sysfs(dev);
+	if (err) {
+		mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
+		goto sysfs_err;
+	}
+
+	mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
+		     dev->dev->caps.sqp_demux);
+	for (i = 0; i < dev->num_ports; i++) {
+		union ib_gid gid;
+		err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
+		if (err)
+			goto demux_err;
+		dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
+		atomic64_set(&dev->sriov.demux[i].subnet_prefix,
+			     be64_to_cpu(gid.global.subnet_prefix));
+		err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
+				      &dev->sriov.sqps[i]);
+		if (err)
+			goto demux_err;
+		err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
+		if (err)
+			goto free_pv;
+	}
+	mlx4_ib_master_tunnels(dev, 1);
+	return 0;
+
+free_pv:
+	free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
+demux_err:
+	while (--i >= 0) {
+		free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
+		mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
+	}
+	mlx4_ib_device_unregister_sysfs(dev);
+
+sysfs_err:
+	mlx4_ib_destroy_alias_guid_service(dev);
+
+paravirt_err:
+	mlx4_ib_cm_paravirt_clean(dev, -1);
+
+	return err;
+}
+
+void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
+{
+	int i;
+	unsigned long flags;
+
+	if (!mlx4_is_mfunc(dev->dev))
+		return;
+
+	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
+	dev->sriov.is_going_down = 1;
+	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
+	if (mlx4_is_master(dev->dev)) {
+		for (i = 0; i < dev->num_ports; i++) {
+			flush_workqueue(dev->sriov.demux[i].ud_wq);
+			mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
+			kfree(dev->sriov.sqps[i]);
+			dev->sriov.sqps[i] = NULL;
+			mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
+		}
+
+		mlx4_ib_cm_paravirt_clean(dev, -1);
+		mlx4_ib_destroy_alias_guid_service(dev);
+		mlx4_ib_device_unregister_sysfs(dev);
+	}
+}
diff --git a/drivers/infiniband/hw/mlx4/main.c b/drivers/infiniband/hw/mlx4/main.c
new file mode 100644
index 000000000..c62cdd645
--- /dev/null
+++ b/drivers/infiniband/hw/mlx4/main.c
@@ -0,0 +1,3360 @@
+/*
+ * Copyright (c) 2006, 2007 Cisco Systems, Inc. All rights reserved.
+ * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/slab.h>
+#include <linux/errno.h>
+#include <linux/netdevice.h>
+#include <linux/inetdevice.h>
+#include <linux/rtnetlink.h>
+#include <linux/if_vlan.h>
+#include <linux/sched/mm.h>
+#include <linux/sched/task.h>
+
+#include <net/ipv6.h>
+#include <net/addrconf.h>
+#include <net/devlink.h>
+
+#include <rdma/ib_smi.h>
+#include <rdma/ib_user_verbs.h>
+#include <rdma/ib_addr.h>
+#include <rdma/ib_cache.h>
+
+#include <net/bonding.h>
+
+#include <linux/mlx4/driver.h>
+#include <linux/mlx4/cmd.h>
+#include <linux/mlx4/qp.h>
+
+#include "mlx4_ib.h"
+#include <rdma/mlx4-abi.h>
+
+#define DRV_NAME	MLX4_IB_DRV_NAME
+#define DRV_VERSION	"4.0-0"
+
+#define MLX4_IB_FLOW_MAX_PRIO 0xFFF
+#define MLX4_IB_FLOW_QPN_MASK 0xFFFFFF
+#define MLX4_IB_CARD_REV_A0   0xA0
+
+MODULE_AUTHOR("Roland Dreier");
+MODULE_DESCRIPTION("Mellanox ConnectX HCA InfiniBand driver");
+MODULE_LICENSE("Dual BSD/GPL");
+
+int mlx4_ib_sm_guid_assign = 0;
+module_param_named(sm_guid_assign, mlx4_ib_sm_guid_assign, int, 0444);
+MODULE_PARM_DESC(sm_guid_assign, "Enable SM alias_GUID assignment if sm_guid_assign > 0 (Default: 0)");
+
+static const char mlx4_ib_version[] =
+	DRV_NAME ": Mellanox ConnectX InfiniBand driver v"
+	DRV_VERSION "\n";
+
+static void do_slave_init(struct mlx4_ib_dev *ibdev, int slave, int do_init);
+static enum rdma_link_layer mlx4_ib_port_link_layer(struct ib_device *device,
+						    u8 port_num);
+
+static struct workqueue_struct *wq;
+
+static void init_query_mad(struct ib_smp *mad)
+{
+	mad->base_version  = 1;
+	mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
+	mad->class_version = 1;
+	mad->method	   = IB_MGMT_METHOD_GET;
+}
+
+static int check_flow_steering_support(struct mlx4_dev *dev)
+{
+	int eth_num_ports = 0;
+	int ib_num_ports = 0;
+
+	int dmfs = dev->caps.steering_mode == MLX4_STEERING_MODE_DEVICE_MANAGED;
+
+	if (dmfs) {
+		int i;
+		mlx4_foreach_port(i, dev, MLX4_PORT_TYPE_ETH)
+			eth_num_ports++;
+		mlx4_foreach_port(i, dev, MLX4_PORT_TYPE_IB)
+			ib_num_ports++;
+		dmfs &= (!ib_num_ports ||
+			 (dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_DMFS_IPOIB)) &&
+			(!eth_num_ports ||
+			 (dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_FS_EN));
+		if (ib_num_ports && mlx4_is_mfunc(dev)) {
+			pr_warn("Device managed flow steering is unavailable for IB port in multifunction env.\n");
+			dmfs = 0;
+		}
+	}
+	return dmfs;
+}
+
+static int num_ib_ports(struct mlx4_dev *dev)
+{
+	int ib_ports = 0;
+	int i;
+
+	mlx4_foreach_port(i, dev, MLX4_PORT_TYPE_IB)
+		ib_ports++;
+
+	return ib_ports;
+}
+
+static struct net_device *mlx4_ib_get_netdev(struct ib_device *device, u8 port_num)
+{
+	struct mlx4_ib_dev *ibdev = to_mdev(device);
+	struct net_device *dev;
+
+	rcu_read_lock();
+	dev = mlx4_get_protocol_dev(ibdev->dev, MLX4_PROT_ETH, port_num);
+
+	if (dev) {
+		if (mlx4_is_bonded(ibdev->dev)) {
+			struct net_device *upper = NULL;
+
+			upper = netdev_master_upper_dev_get_rcu(dev);
+			if (upper) {
+				struct net_device *active;
+
+				active = bond_option_active_slave_get_rcu(netdev_priv(upper));
+				if (active)
+					dev = active;
+			}
+		}
+	}
+	if (dev)
+		dev_hold(dev);
+
+	rcu_read_unlock();
+	return dev;
+}
+
+static int mlx4_ib_update_gids_v1(struct gid_entry *gids,
+				  struct mlx4_ib_dev *ibdev,
+				  u8 port_num)
+{
+	struct mlx4_cmd_mailbox *mailbox;
+	int err;
+	struct mlx4_dev *dev = ibdev->dev;
+	int i;
+	union ib_gid *gid_tbl;
+
+	mailbox = mlx4_alloc_cmd_mailbox(dev);
+	if (IS_ERR(mailbox))
+		return -ENOMEM;
+
+	gid_tbl = mailbox->buf;
+
+	for (i = 0; i < MLX4_MAX_PORT_GIDS; ++i)
+		memcpy(&gid_tbl[i], &gids[i].gid, sizeof(union ib_gid));
+
+	err = mlx4_cmd(dev, mailbox->dma,
+		       MLX4_SET_PORT_GID_TABLE << 8 | port_num,
+		       1, MLX4_CMD_SET_PORT, MLX4_CMD_TIME_CLASS_B,
+		       MLX4_CMD_WRAPPED);
+	if (mlx4_is_bonded(dev))
+		err += mlx4_cmd(dev, mailbox->dma,
+				MLX4_SET_PORT_GID_TABLE << 8 | 2,
+				1, MLX4_CMD_SET_PORT, MLX4_CMD_TIME_CLASS_B,
+				MLX4_CMD_WRAPPED);
+
+	mlx4_free_cmd_mailbox(dev, mailbox);
+	return err;
+}
+
+static int mlx4_ib_update_gids_v1_v2(struct gid_entry *gids,
+				     struct mlx4_ib_dev *ibdev,
+				     u8 port_num)
+{
+	struct mlx4_cmd_mailbox *mailbox;
+	int err;
+	struct mlx4_dev *dev = ibdev->dev;
+	int i;
+	struct {
+		union ib_gid	gid;
+		__be32		rsrvd1[2];
+		__be16		rsrvd2;
+		u8		type;
+		u8		version;
+		__be32		rsrvd3;
+	} *gid_tbl;
+
+	mailbox = mlx4_alloc_cmd_mailbox(dev);
+	if (IS_ERR(mailbox))
+		return -ENOMEM;
+
+	gid_tbl = mailbox->buf;
+	for (i = 0; i < MLX4_MAX_PORT_GIDS; ++i) {
+		memcpy(&gid_tbl[i].gid, &gids[i].gid, sizeof(union ib_gid));
+		if (gids[i].gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
+			gid_tbl[i].version = 2;
+			if (!ipv6_addr_v4mapped((struct in6_addr *)&gids[i].gid))
+				gid_tbl[i].type = 1;
+		}
+	}
+
+	err = mlx4_cmd(dev, mailbox->dma,
+		       MLX4_SET_PORT_ROCE_ADDR << 8 | port_num,
+		       1, MLX4_CMD_SET_PORT, MLX4_CMD_TIME_CLASS_B,
+		       MLX4_CMD_WRAPPED);
+	if (mlx4_is_bonded(dev))
+		err += mlx4_cmd(dev, mailbox->dma,
+				MLX4_SET_PORT_ROCE_ADDR << 8 | 2,
+				1, MLX4_CMD_SET_PORT, MLX4_CMD_TIME_CLASS_B,
+				MLX4_CMD_WRAPPED);
+
+	mlx4_free_cmd_mailbox(dev, mailbox);
+	return err;
+}
+
+static int mlx4_ib_update_gids(struct gid_entry *gids,
+			       struct mlx4_ib_dev *ibdev,
+			       u8 port_num)
+{
+	if (ibdev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2)
+		return mlx4_ib_update_gids_v1_v2(gids, ibdev, port_num);
+
+	return mlx4_ib_update_gids_v1(gids, ibdev, port_num);
+}
+
+static void free_gid_entry(struct gid_entry *entry)
+{
+	memset(&entry->gid, 0, sizeof(entry->gid));
+	kfree(entry->ctx);
+	entry->ctx = NULL;
+}
+
+static int mlx4_ib_add_gid(const struct ib_gid_attr *attr, void **context)
+{
+	struct mlx4_ib_dev *ibdev = to_mdev(attr->device);
+	struct mlx4_ib_iboe *iboe = &ibdev->iboe;
+	struct mlx4_port_gid_table   *port_gid_table;
+	int free = -1, found = -1;
+	int ret = 0;
+	int hw_update = 0;
+	int i;
+	struct gid_entry *gids = NULL;
+	u16 vlan_id = 0xffff;
+	u8 mac[ETH_ALEN];
+
+	if (!rdma_cap_roce_gid_table(attr->device, attr->port_num))
+		return -EINVAL;
+
+	if (attr->port_num > MLX4_MAX_PORTS)
+		return -EINVAL;
+
+	if (!context)
+		return -EINVAL;
+
+	ret = rdma_read_gid_l2_fields(attr, &vlan_id, &mac[0]);
+	if (ret)
+		return ret;
+	port_gid_table = &iboe->gids[attr->port_num - 1];
+	spin_lock_bh(&iboe->lock);
+	for (i = 0; i < MLX4_MAX_PORT_GIDS; ++i) {
+		if (!memcmp(&port_gid_table->gids[i].gid,
+			    &attr->gid, sizeof(attr->gid)) &&
+		    port_gid_table->gids[i].gid_type == attr->gid_type &&
+		    port_gid_table->gids[i].vlan_id == vlan_id)  {
+			found = i;
+			break;
+		}
+		if (free < 0 && rdma_is_zero_gid(&port_gid_table->gids[i].gid))
+			free = i; /* HW has space */
+	}
+
+	if (found < 0) {
+		if (free < 0) {
+			ret = -ENOSPC;
+		} else {
+			port_gid_table->gids[free].ctx = kmalloc(sizeof(*port_gid_table->gids[free].ctx), GFP_ATOMIC);
+			if (!port_gid_table->gids[free].ctx) {
+				ret = -ENOMEM;
+			} else {
+				*context = port_gid_table->gids[free].ctx;
+				memcpy(&port_gid_table->gids[free].gid,
+				       &attr->gid, sizeof(attr->gid));
+				port_gid_table->gids[free].gid_type = attr->gid_type;
+				port_gid_table->gids[free].vlan_id = vlan_id;
+				port_gid_table->gids[free].ctx->real_index = free;
+				port_gid_table->gids[free].ctx->refcount = 1;
+				hw_update = 1;
+			}
+		}
+	} else {
+		struct gid_cache_context *ctx = port_gid_table->gids[found].ctx;
+		*context = ctx;
+		ctx->refcount++;
+	}
+	if (!ret && hw_update) {
+		gids = kmalloc_array(MLX4_MAX_PORT_GIDS, sizeof(*gids),
+				     GFP_ATOMIC);
+		if (!gids) {
+			ret = -ENOMEM;
+			*context = NULL;
+			free_gid_entry(&port_gid_table->gids[free]);
+		} else {
+			for (i = 0; i < MLX4_MAX_PORT_GIDS; i++) {
+				memcpy(&gids[i].gid, &port_gid_table->gids[i].gid, sizeof(union ib_gid));
+				gids[i].gid_type = port_gid_table->gids[i].gid_type;
+			}
+		}
+	}
+	spin_unlock_bh(&iboe->lock);
+
+	if (!ret && hw_update) {
+		ret = mlx4_ib_update_gids(gids, ibdev, attr->port_num);
+		if (ret) {
+			spin_lock_bh(&iboe->lock);
+			*context = NULL;
+			free_gid_entry(&port_gid_table->gids[free]);
+			spin_unlock_bh(&iboe->lock);
+		}
+		kfree(gids);
+	}
+
+	return ret;
+}
+
+static int mlx4_ib_del_gid(const struct ib_gid_attr *attr, void **context)
+{
+	struct gid_cache_context *ctx = *context;
+	struct mlx4_ib_dev *ibdev = to_mdev(attr->device);
+	struct mlx4_ib_iboe *iboe = &ibdev->iboe;
+	struct mlx4_port_gid_table   *port_gid_table;
+	int ret = 0;
+	int hw_update = 0;
+	struct gid_entry *gids = NULL;
+
+	if (!rdma_cap_roce_gid_table(attr->device, attr->port_num))
+		return -EINVAL;
+
+	if (attr->port_num > MLX4_MAX_PORTS)
+		return -EINVAL;
+
+	port_gid_table = &iboe->gids[attr->port_num - 1];
+	spin_lock_bh(&iboe->lock);
+	if (ctx) {
+		ctx->refcount--;
+		if (!ctx->refcount) {
+			unsigned int real_index = ctx->real_index;
+
+			free_gid_entry(&port_gid_table->gids[real_index]);
+			hw_update = 1;
+		}
+	}
+	if (!ret && hw_update) {
+		int i;
+
+		gids = kmalloc_array(MLX4_MAX_PORT_GIDS, sizeof(*gids),
+				     GFP_ATOMIC);
+		if (!gids) {
+			ret = -ENOMEM;
+		} else {
+			for (i = 0; i < MLX4_MAX_PORT_GIDS; i++) {
+				memcpy(&gids[i].gid,
+				       &port_gid_table->gids[i].gid,
+				       sizeof(union ib_gid));
+				gids[i].gid_type =
+				    port_gid_table->gids[i].gid_type;
+			}
+		}
+	}
+	spin_unlock_bh(&iboe->lock);
+
+	if (!ret && hw_update) {
+		ret = mlx4_ib_update_gids(gids, ibdev, attr->port_num);
+		kfree(gids);
+	}
+	return ret;
+}
+
+int mlx4_ib_gid_index_to_real_index(struct mlx4_ib_dev *ibdev,
+				    const struct ib_gid_attr *attr)
+{
+	struct mlx4_ib_iboe *iboe = &ibdev->iboe;
+	struct gid_cache_context *ctx = NULL;
+	struct mlx4_port_gid_table   *port_gid_table;
+	int real_index = -EINVAL;
+	int i;
+	unsigned long flags;
+	u8 port_num = attr->port_num;
+
+	if (port_num > MLX4_MAX_PORTS)
+		return -EINVAL;
+
+	if (mlx4_is_bonded(ibdev->dev))
+		port_num = 1;
+
+	if (!rdma_cap_roce_gid_table(&ibdev->ib_dev, port_num))
+		return attr->index;
+
+	spin_lock_irqsave(&iboe->lock, flags);
+	port_gid_table = &iboe->gids[port_num - 1];
+
+	for (i = 0; i < MLX4_MAX_PORT_GIDS; ++i)
+		if (!memcmp(&port_gid_table->gids[i].gid,
+			    &attr->gid, sizeof(attr->gid)) &&
+		    attr->gid_type == port_gid_table->gids[i].gid_type) {
+			ctx = port_gid_table->gids[i].ctx;
+			break;
+		}
+	if (ctx)
+		real_index = ctx->real_index;
+	spin_unlock_irqrestore(&iboe->lock, flags);
+	return real_index;
+}
+
+static int mlx4_ib_query_device(struct ib_device *ibdev,
+				struct ib_device_attr *props,
+				struct ib_udata *uhw)
+{
+	struct mlx4_ib_dev *dev = to_mdev(ibdev);
+	struct ib_smp *in_mad  = NULL;
+	struct ib_smp *out_mad = NULL;
+	int err;
+	int have_ib_ports;
+	struct mlx4_uverbs_ex_query_device cmd;
+	struct mlx4_uverbs_ex_query_device_resp resp = {};
+	struct mlx4_clock_params clock_params;
+
+	if (uhw->inlen) {
+		if (uhw->inlen < sizeof(cmd))
+			return -EINVAL;
+
+		err = ib_copy_from_udata(&cmd, uhw, sizeof(cmd));
+		if (err)
+			return err;
+
+		if (cmd.comp_mask)
+			return -EINVAL;
+
+		if (cmd.reserved)
+			return -EINVAL;
+	}
+
+	resp.response_length = offsetof(typeof(resp), response_length) +
+		sizeof(resp.response_length);
+	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
+	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
+	err = -ENOMEM;
+	if (!in_mad || !out_mad)
+		goto out;
+
+	init_query_mad(in_mad);
+	in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
+
+	err = mlx4_MAD_IFC(to_mdev(ibdev), MLX4_MAD_IFC_IGNORE_KEYS,
+			   1, NULL, NULL, in_mad, out_mad);
+	if (err)
+		goto out;
+
+	memset(props, 0, sizeof *props);
+
+	have_ib_ports = num_ib_ports(dev->dev);
+
+	props->fw_ver = dev->dev->caps.fw_ver;
+	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
+		IB_DEVICE_PORT_ACTIVE_EVENT		|
+		IB_DEVICE_SYS_IMAGE_GUID		|
+		IB_DEVICE_RC_RNR_NAK_GEN		|
+		IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
+	if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_BAD_PKEY_CNTR)
+		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
+	if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_BAD_QKEY_CNTR)
+		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
+	if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_APM && have_ib_ports)
+		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
+	if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_UD_AV_PORT)
+		props->device_cap_flags |= IB_DEVICE_UD_AV_PORT_ENFORCE;
+	if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_IPOIB_CSUM)
+		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
+	if (dev->dev->caps.max_gso_sz &&
+	    (dev->dev->rev_id != MLX4_IB_CARD_REV_A0) &&
+	    (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_BLH))
+		props->device_cap_flags |= IB_DEVICE_UD_TSO;
+	if (dev->dev->caps.bmme_flags & MLX4_BMME_FLAG_RESERVED_LKEY)
+		props->device_cap_flags |= IB_DEVICE_LOCAL_DMA_LKEY;
+	if ((dev->dev->caps.bmme_flags & MLX4_BMME_FLAG_LOCAL_INV) &&
+	    (dev->dev->caps.bmme_flags & MLX4_BMME_FLAG_REMOTE_INV) &&
+	    (dev->dev->caps.bmme_flags & MLX4_BMME_FLAG_FAST_REG_WR))
+		props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
+	if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC)
+		props->device_cap_flags |= IB_DEVICE_XRC;
+	if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_MEM_WINDOW)
+		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW;
+	if (dev->dev->caps.bmme_flags & MLX4_BMME_FLAG_TYPE_2_WIN) {
+		if (dev->dev->caps.bmme_flags & MLX4_BMME_FLAG_WIN_TYPE_2B)
+			props->device_cap_flags |= IB_DEVICE_MEM_WINDOW_TYPE_2B;
+		else
+			props->device_cap_flags |= IB_DEVICE_MEM_WINDOW_TYPE_2A;
+	}
+	if (dev->steering_support == MLX4_STEERING_MODE_DEVICE_MANAGED)
+		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
+
+	props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
+
+	props->vendor_id	   = be32_to_cpup((__be32 *) (out_mad->data + 36)) &
+		0xffffff;
+	props->vendor_part_id	   = dev->dev->persist->pdev->device;
+	props->hw_ver		   = be32_to_cpup((__be32 *) (out_mad->data + 32));
+	memcpy(&props->sys_image_guid, out_mad->data +	4, 8);
+
+	props->max_mr_size	   = ~0ull;
+	props->page_size_cap	   = dev->dev->caps.page_size_cap;
+	props->max_qp		   = dev->dev->quotas.qp;
+	props->max_qp_wr	   = dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE;
+	props->max_send_sge =
+		min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg);
+	props->max_recv_sge =
+		min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg);
+	props->max_sge_rd = MLX4_MAX_SGE_RD;
+	props->max_cq		   = dev->dev->quotas.cq;
+	props->max_cqe		   = dev->dev->caps.max_cqes;
+	props->max_mr		   = dev->dev->quotas.mpt;
+	props->max_pd		   = dev->dev->caps.num_pds - dev->dev->caps.reserved_pds;
+	props->max_qp_rd_atom	   = dev->dev->caps.max_qp_dest_rdma;
+	props->max_qp_init_rd_atom = dev->dev->caps.max_qp_init_rdma;
+	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
+	props->max_srq		   = dev->dev->quotas.srq;
+	props->max_srq_wr	   = dev->dev->caps.max_srq_wqes - 1;
+	props->max_srq_sge	   = dev->dev->caps.max_srq_sge;
+	props->max_fast_reg_page_list_len = MLX4_MAX_FAST_REG_PAGES;
+	props->local_ca_ack_delay  = dev->dev->caps.local_ca_ack_delay;
+	props->atomic_cap	   = dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_ATOMIC ?
+		IB_ATOMIC_HCA : IB_ATOMIC_NONE;
+	props->masked_atomic_cap   = props->atomic_cap;
+	props->max_pkeys	   = dev->dev->caps.pkey_table_len[1];
+	props->max_mcast_grp	   = dev->dev->caps.num_mgms + dev->dev->caps.num_amgms;
+	props->max_mcast_qp_attach = dev->dev->caps.num_qp_per_mgm;
+	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
+					   props->max_mcast_grp;
+	props->hca_core_clock = dev->dev->caps.hca_core_clock * 1000UL;
+	props->timestamp_mask = 0xFFFFFFFFFFFFULL;
+	props->max_ah = INT_MAX;
+
+	if (mlx4_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET ||
+	    mlx4_ib_port_link_layer(ibdev, 2) == IB_LINK_LAYER_ETHERNET) {
+		if (dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RSS) {
+			props->rss_caps.max_rwq_indirection_tables =
+				props->max_qp;
+			props->rss_caps.max_rwq_indirection_table_size =
+				dev->dev->caps.max_rss_tbl_sz;
+			props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
+			props->max_wq_type_rq = props->max_qp;
+		}
+
+		if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_FCS_KEEP)
+			props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
+	}
+
+	props->cq_caps.max_cq_moderation_count = MLX4_MAX_CQ_COUNT;
+	props->cq_caps.max_cq_moderation_period = MLX4_MAX_CQ_PERIOD;
+
+	if (uhw->outlen >= resp.response_length + sizeof(resp.hca_core_clock_offset)) {
+		resp.response_length += sizeof(resp.hca_core_clock_offset);
+		if (!mlx4_get_internal_clock_params(dev->dev, &clock_params)) {
+			resp.comp_mask |= MLX4_IB_QUERY_DEV_RESP_MASK_CORE_CLOCK_OFFSET;
+			resp.hca_core_clock_offset = clock_params.offset % PAGE_SIZE;
+		}
+	}
+
+	if (uhw->outlen >= resp.response_length +
+	    sizeof(resp.max_inl_recv_sz)) {
+		resp.response_length += sizeof(resp.max_inl_recv_sz);
+		resp.max_inl_recv_sz  = dev->dev->caps.max_rq_sg *
+			sizeof(struct mlx4_wqe_data_seg);
+	}
+
+	if (offsetofend(typeof(resp), rss_caps) <= uhw->outlen) {
+		if (props->rss_caps.supported_qpts) {
+			resp.rss_caps.rx_hash_function =
+				MLX4_IB_RX_HASH_FUNC_TOEPLITZ;
+
+			resp.rss_caps.rx_hash_fields_mask =
+				MLX4_IB_RX_HASH_SRC_IPV4 |
+				MLX4_IB_RX_HASH_DST_IPV4 |
+				MLX4_IB_RX_HASH_SRC_IPV6 |
+				MLX4_IB_RX_HASH_DST_IPV6 |
+				MLX4_IB_RX_HASH_SRC_PORT_TCP |
+				MLX4_IB_RX_HASH_DST_PORT_TCP |
+				MLX4_IB_RX_HASH_SRC_PORT_UDP |
+				MLX4_IB_RX_HASH_DST_PORT_UDP;
+
+			if (dev->dev->caps.tunnel_offload_mode ==
+			    MLX4_TUNNEL_OFFLOAD_MODE_VXLAN)
+				resp.rss_caps.rx_hash_fields_mask |=
+					MLX4_IB_RX_HASH_INNER;
+		}
+		resp.response_length = offsetof(typeof(resp), rss_caps) +
+				       sizeof(resp.rss_caps);
+	}
+
+	if (offsetofend(typeof(resp), tso_caps) <= uhw->outlen) {
+		if (dev->dev->caps.max_gso_sz &&
+		    ((mlx4_ib_port_link_layer(ibdev, 1) ==
+		    IB_LINK_LAYER_ETHERNET) ||
+		    (mlx4_ib_port_link_layer(ibdev, 2) ==
+		    IB_LINK_LAYER_ETHERNET))) {
+			resp.tso_caps.max_tso = dev->dev->caps.max_gso_sz;
+			resp.tso_caps.supported_qpts |=
+				1 << IB_QPT_RAW_PACKET;
+		}
+		resp.response_length = offsetof(typeof(resp), tso_caps) +
+				       sizeof(resp.tso_caps);
+	}
+
+	if (uhw->outlen) {
+		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
+		if (err)
+			goto out;
+	}
+out:
+	kfree(in_mad);
+	kfree(out_mad);
+
+	return err;
+}
+
+static enum rdma_link_layer
+mlx4_ib_port_link_layer(struct ib_device *device, u8 port_num)
+{
+	struct mlx4_dev *dev = to_mdev(device)->dev;
+
+	return dev->caps.port_mask[port_num] == MLX4_PORT_TYPE_IB ?
+		IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
+}
+
+static int ib_link_query_port(struct ib_device *ibdev, u8 port,
+			      struct ib_port_attr *props, int netw_view)
+{
+	struct ib_smp *in_mad  = NULL;
+	struct ib_smp *out_mad = NULL;
+	int ext_active_speed;
+	int mad_ifc_flags = MLX4_MAD_IFC_IGNORE_KEYS;
+	int err = -ENOMEM;
+
+	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
+	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
+	if (!in_mad || !out_mad)
+		goto out;
+
+	init_query_mad(in_mad);
+	in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
+	in_mad->attr_mod = cpu_to_be32(port);
+
+	if (mlx4_is_mfunc(to_mdev(ibdev)->dev) && netw_view)
+		mad_ifc_flags |= MLX4_MAD_IFC_NET_VIEW;
+
+	err = mlx4_MAD_IFC(to_mdev(ibdev), mad_ifc_flags, port, NULL, NULL,
+				in_mad, out_mad);
+	if (err)
+		goto out;
+
+
+	props->lid		= be16_to_cpup((__be16 *) (out_mad->data + 16));
+	props->lmc		= out_mad->data[34] & 0x7;
+	props->sm_lid		= be16_to_cpup((__be16 *) (out_mad->data + 18));
+	props->sm_sl		= out_mad->data[36] & 0xf;
+	props->state		= out_mad->data[32] & 0xf;
+	props->phys_state	= out_mad->data[33] >> 4;
+	props->port_cap_flags	= be32_to_cpup((__be32 *) (out_mad->data + 20));
+	if (netw_view)
+		props->gid_tbl_len = out_mad->data[50];
+	else
+		props->gid_tbl_len = to_mdev(ibdev)->dev->caps.gid_table_len[port];
+	props->max_msg_sz	= to_mdev(ibdev)->dev->caps.max_msg_sz;
+	props->pkey_tbl_len	= to_mdev(ibdev)->dev->caps.pkey_table_len[port];
+	props->bad_pkey_cntr	= be16_to_cpup((__be16 *) (out_mad->data + 46));
+	props->qkey_viol_cntr	= be16_to_cpup((__be16 *) (out_mad->data + 48));
+	props->active_width	= out_mad->data[31] & 0xf;
+	props->active_speed	= out_mad->data[35] >> 4;
+	props->max_mtu		= out_mad->data[41] & 0xf;
+	props->active_mtu	= out_mad->data[36] >> 4;
+	props->subnet_timeout	= out_mad->data[51] & 0x1f;
+	props->max_vl_num	= out_mad->data[37] >> 4;
+	props->init_type_reply	= out_mad->data[41] >> 4;
+
+	/* Check if extended speeds (EDR/FDR/...) are supported */
+	if (props->port_cap_flags & IB_PORT_EXTENDED_SPEEDS_SUP) {
+		ext_active_speed = out_mad->data[62] >> 4;
+
+		switch (ext_active_speed) {
+		case 1:
+			props->active_speed = IB_SPEED_FDR;
+			break;
+		case 2:
+			props->active_speed = IB_SPEED_EDR;
+			break;
+		}
+	}
+
+	/* If reported active speed is QDR, check if is FDR-10 */
+	if (props->active_speed == IB_SPEED_QDR) {
+		init_query_mad(in_mad);
+		in_mad->attr_id = MLX4_ATTR_EXTENDED_PORT_INFO;
+		in_mad->attr_mod = cpu_to_be32(port);
+
+		err = mlx4_MAD_IFC(to_mdev(ibdev), mad_ifc_flags, port,
+				   NULL, NULL, in_mad, out_mad);
+		if (err)
+			goto out;
+
+		/* Checking LinkSpeedActive for FDR-10 */
+		if (out_mad->data[15] & 0x1)
+			props->active_speed = IB_SPEED_FDR10;
+	}
+
+	/* Avoid wrong speed value returned by FW if the IB link is down. */
+	if (props->state == IB_PORT_DOWN)
+		 props->active_speed = IB_SPEED_SDR;
+
+out:
+	kfree(in_mad);
+	kfree(out_mad);
+	return err;
+}
+
+static u8 state_to_phys_state(enum ib_port_state state)
+{
+	return state == IB_PORT_ACTIVE ?
+		IB_PORT_PHYS_STATE_LINK_UP : IB_PORT_PHYS_STATE_DISABLED;
+}
+
+static int eth_link_query_port(struct ib_device *ibdev, u8 port,
+			       struct ib_port_attr *props)
+{
+
+	struct mlx4_ib_dev *mdev = to_mdev(ibdev);
+	struct mlx4_ib_iboe *iboe = &mdev->iboe;
+	struct net_device *ndev;
+	enum ib_mtu tmp;
+	struct mlx4_cmd_mailbox *mailbox;
+	int err = 0;
+	int is_bonded = mlx4_is_bonded(mdev->dev);
+
+	mailbox = mlx4_alloc_cmd_mailbox(mdev->dev);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+
+	err = mlx4_cmd_box(mdev->dev, 0, mailbox->dma, port, 0,
+			   MLX4_CMD_QUERY_PORT, MLX4_CMD_TIME_CLASS_B,
+			   MLX4_CMD_WRAPPED);
+	if (err)
+		goto out;
+
+	props->active_width	=  (((u8 *)mailbox->buf)[5] == 0x40) ||
+				   (((u8 *)mailbox->buf)[5] == 0x20 /*56Gb*/) ?
+					   IB_WIDTH_4X : IB_WIDTH_1X;
+	props->active_speed	=  (((u8 *)mailbox->buf)[5] == 0x20 /*56Gb*/) ?
+					   IB_SPEED_FDR : IB_SPEED_QDR;
+	props->port_cap_flags	= IB_PORT_CM_SUP;
+	props->ip_gids = true;
+	props->gid_tbl_len	= mdev->dev->caps.gid_table_len[port];
+	props->max_msg_sz	= mdev->dev->caps.max_msg_sz;
+	if (mdev->dev->caps.pkey_table_len[port])
+		props->pkey_tbl_len = 1;
+	props->max_mtu		= IB_MTU_4096;
+	props->max_vl_num	= 2;
+	props->state		= IB_PORT_DOWN;
+	props->phys_state	= state_to_phys_state(props->state);
+	props->active_mtu	= IB_MTU_256;
+	spin_lock_bh(&iboe->lock);
+	ndev = iboe->netdevs[port - 1];
+	if (ndev && is_bonded) {
+		rcu_read_lock(); /* required to get upper dev */
+		ndev = netdev_master_upper_dev_get_rcu(ndev);
+		rcu_read_unlock();
+	}
+	if (!ndev)
+		goto out_unlock;
+
+	tmp = iboe_get_mtu(ndev->mtu);
+	props->active_mtu = tmp ? min(props->max_mtu, tmp) : IB_MTU_256;
+
+	props->state		= (netif_running(ndev) && netif_carrier_ok(ndev)) ?
+					IB_PORT_ACTIVE : IB_PORT_DOWN;
+	props->phys_state	= state_to_phys_state(props->state);
+out_unlock:
+	spin_unlock_bh(&iboe->lock);
+out:
+	mlx4_free_cmd_mailbox(mdev->dev, mailbox);
+	return err;
+}
+
+int __mlx4_ib_query_port(struct ib_device *ibdev, u8 port,
+			 struct ib_port_attr *props, int netw_view)
+{
+	int err;
+
+	/* props being zeroed by the caller, avoid zeroing it here */
+
+	err = mlx4_ib_port_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND ?
+		ib_link_query_port(ibdev, port, props, netw_view) :
+				eth_link_query_port(ibdev, port, props);
+
+	return err;
+}
+
+static int mlx4_ib_query_port(struct ib_device *ibdev, u8 port,
+			      struct ib_port_attr *props)
+{
+	/* returns host view */
+	return __mlx4_ib_query_port(ibdev, port, props, 0);
+}
+
+int __mlx4_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
+			union ib_gid *gid, int netw_view)
+{
+	struct ib_smp *in_mad  = NULL;
+	struct ib_smp *out_mad = NULL;
+	int err = -ENOMEM;
+	struct mlx4_ib_dev *dev = to_mdev(ibdev);
+	int clear = 0;
+	int mad_ifc_flags = MLX4_MAD_IFC_IGNORE_KEYS;
+
+	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
+	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
+	if (!in_mad || !out_mad)
+		goto out;
+
+	init_query_mad(in_mad);
+	in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
+	in_mad->attr_mod = cpu_to_be32(port);
+
+	if (mlx4_is_mfunc(dev->dev) && netw_view)
+		mad_ifc_flags |= MLX4_MAD_IFC_NET_VIEW;
+
+	err = mlx4_MAD_IFC(dev, mad_ifc_flags, port, NULL, NULL, in_mad, out_mad);
+	if (err)
+		goto out;
+
+	memcpy(gid->raw, out_mad->data + 8, 8);
+
+	if (mlx4_is_mfunc(dev->dev) && !netw_view) {
+		if (index) {
+			/* For any index > 0, return the null guid */
+			err = 0;
+			clear = 1;
+			goto out;
+		}
+	}
+
+	init_query_mad(in_mad);
+	in_mad->attr_id  = IB_SMP_ATTR_GUID_INFO;
+	in_mad->attr_mod = cpu_to_be32(index / 8);
+
+	err = mlx4_MAD_IFC(dev, mad_ifc_flags, port,
+			   NULL, NULL, in_mad, out_mad);
+	if (err)
+		goto out;
+
+	memcpy(gid->raw + 8, out_mad->data + (index % 8) * 8, 8);
+
+out:
+	if (clear)
+		memset(gid->raw + 8, 0, 8);
+	kfree(in_mad);
+	kfree(out_mad);
+	return err;
+}
+
+static int mlx4_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
+			     union ib_gid *gid)
+{
+	if (rdma_protocol_ib(ibdev, port))
+		return __mlx4_ib_query_gid(ibdev, port, index, gid, 0);
+	return 0;
+}
+
+static int mlx4_ib_query_sl2vl(struct ib_device *ibdev, u8 port, u64 *sl2vl_tbl)
+{
+	union sl2vl_tbl_to_u64 sl2vl64;
+	struct ib_smp *in_mad  = NULL;
+	struct ib_smp *out_mad = NULL;
+	int mad_ifc_flags = MLX4_MAD_IFC_IGNORE_KEYS;
+	int err = -ENOMEM;
+	int jj;
+
+	if (mlx4_is_slave(to_mdev(ibdev)->dev)) {
+		*sl2vl_tbl = 0;
+		return 0;
+	}
+
+	in_mad  = kzalloc(sizeof(*in_mad), GFP_KERNEL);
+	out_mad = kmalloc(sizeof(*out_mad), GFP_KERNEL);
+	if (!in_mad || !out_mad)
+		goto out;
+
+	init_query_mad(in_mad);
+	in_mad->attr_id  = IB_SMP_ATTR_SL_TO_VL_TABLE;
+	in_mad->attr_mod = 0;
+
+	if (mlx4_is_mfunc(to_mdev(ibdev)->dev))
+		mad_ifc_flags |= MLX4_MAD_IFC_NET_VIEW;
+
+	err = mlx4_MAD_IFC(to_mdev(ibdev), mad_ifc_flags, port, NULL, NULL,
+			   in_mad, out_mad);
+	if (err)
+		goto out;
+
+	for (jj = 0; jj < 8; jj++)
+		sl2vl64.sl8[jj] = ((struct ib_smp *)out_mad)->data[jj];
+	*sl2vl_tbl = sl2vl64.sl64;
+
+out:
+	kfree(in_mad);
+	kfree(out_mad);
+	return err;
+}
+
+static void mlx4_init_sl2vl_tbl(struct mlx4_ib_dev *mdev)
+{
+	u64 sl2vl;
+	int i;
+	int err;
+
+	for (i = 1; i <= mdev->dev->caps.num_ports; i++) {
+		if (mdev->dev->caps.port_type[i] == MLX4_PORT_TYPE_ETH)
+			continue;
+		err = mlx4_ib_query_sl2vl(&mdev->ib_dev, i, &sl2vl);
+		if (err) {
+			pr_err("Unable to get default sl to vl mapping for port %d.  Using all zeroes (%d)\n",
+			       i, err);
+			sl2vl = 0;
+		}
+		atomic64_set(&mdev->sl2vl[i - 1], sl2vl);
+	}
+}
+
+int __mlx4_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
+			 u16 *pkey, int netw_view)
+{
+	struct ib_smp *in_mad  = NULL;
+	struct ib_smp *out_mad = NULL;
+	int mad_ifc_flags = MLX4_MAD_IFC_IGNORE_KEYS;
+	int err = -ENOMEM;
+
+	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
+	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
+	if (!in_mad || !out_mad)
+		goto out;
+
+	init_query_mad(in_mad);
+	in_mad->attr_id  = IB_SMP_ATTR_PKEY_TABLE;
+	in_mad->attr_mod = cpu_to_be32(index / 32);
+
+	if (mlx4_is_mfunc(to_mdev(ibdev)->dev) && netw_view)
+		mad_ifc_flags |= MLX4_MAD_IFC_NET_VIEW;
+
+	err = mlx4_MAD_IFC(to_mdev(ibdev), mad_ifc_flags, port, NULL, NULL,
+			   in_mad, out_mad);
+	if (err)
+		goto out;
+
+	*pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
+
+out:
+	kfree(in_mad);
+	kfree(out_mad);
+	return err;
+}
+
+static int mlx4_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index, u16 *pkey)
+{
+	return __mlx4_ib_query_pkey(ibdev, port, index, pkey, 0);
+}
+
+static int mlx4_ib_modify_device(struct ib_device *ibdev, int mask,
+				 struct ib_device_modify *props)
+{
+	struct mlx4_cmd_mailbox *mailbox;
+	unsigned long flags;
+
+	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
+		return -EOPNOTSUPP;
+
+	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
+		return 0;
+
+	if (mlx4_is_slave(to_mdev(ibdev)->dev))
+		return -EOPNOTSUPP;
+
+	spin_lock_irqsave(&to_mdev(ibdev)->sm_lock, flags);
+	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
+	spin_unlock_irqrestore(&to_mdev(ibdev)->sm_lock, flags);
+
+	/*
+	 * If possible, pass node desc to FW, so it can generate
+	 * a 144 trap.  If cmd fails, just ignore.
+	 */
+	mailbox = mlx4_alloc_cmd_mailbox(to_mdev(ibdev)->dev);
+	if (IS_ERR(mailbox))
+		return 0;
+
+	memcpy(mailbox->buf, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
+	mlx4_cmd(to_mdev(ibdev)->dev, mailbox->dma, 1, 0,
+		 MLX4_CMD_SET_NODE, MLX4_CMD_TIME_CLASS_A, MLX4_CMD_NATIVE);
+
+	mlx4_free_cmd_mailbox(to_mdev(ibdev)->dev, mailbox);
+
+	return 0;
+}
+
+static int mlx4_ib_SET_PORT(struct mlx4_ib_dev *dev, u8 port, int reset_qkey_viols,
+			    u32 cap_mask)
+{
+	struct mlx4_cmd_mailbox *mailbox;
+	int err;
+
+	mailbox = mlx4_alloc_cmd_mailbox(dev->dev);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+
+	if (dev->dev->flags & MLX4_FLAG_OLD_PORT_CMDS) {
+		*(u8 *) mailbox->buf	     = !!reset_qkey_viols << 6;
+		((__be32 *) mailbox->buf)[2] = cpu_to_be32(cap_mask);
+	} else {
+		((u8 *) mailbox->buf)[3]     = !!reset_qkey_viols;
+		((__be32 *) mailbox->buf)[1] = cpu_to_be32(cap_mask);
+	}
+
+	err = mlx4_cmd(dev->dev, mailbox->dma, port, MLX4_SET_PORT_IB_OPCODE,
+		       MLX4_CMD_SET_PORT, MLX4_CMD_TIME_CLASS_B,
+		       MLX4_CMD_WRAPPED);
+
+	mlx4_free_cmd_mailbox(dev->dev, mailbox);
+	return err;
+}
+
+static int mlx4_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
+			       struct ib_port_modify *props)
+{
+	struct mlx4_ib_dev *mdev = to_mdev(ibdev);
+	u8 is_eth = mdev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
+	struct ib_port_attr attr;
+	u32 cap_mask;
+	int err;
+
+	/* return OK if this is RoCE. CM calls ib_modify_port() regardless
+	 * of whether port link layer is ETH or IB. For ETH ports, qkey
+	 * violations and port capabilities are not meaningful.
+	 */
+	if (is_eth)
+		return 0;
+
+	mutex_lock(&mdev->cap_mask_mutex);
+
+	err = ib_query_port(ibdev, port, &attr);
+	if (err)
+		goto out;
+
+	cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
+		~props->clr_port_cap_mask;
+
+	err = mlx4_ib_SET_PORT(mdev, port,
+			       !!(mask & IB_PORT_RESET_QKEY_CNTR),
+			       cap_mask);
+
+out:
+	mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
+	return err;
+}
+
+static int mlx4_ib_alloc_ucontext(struct ib_ucontext *uctx,
+				  struct ib_udata *udata)
+{
+	struct ib_device *ibdev = uctx->device;
+	struct mlx4_ib_dev *dev = to_mdev(ibdev);
+	struct mlx4_ib_ucontext *context = to_mucontext(uctx);
+	struct mlx4_ib_alloc_ucontext_resp_v3 resp_v3;
+	struct mlx4_ib_alloc_ucontext_resp resp;
+	int err;
+
+	if (!dev->ib_active)
+		return -EAGAIN;
+
+	if (ibdev->ops.uverbs_abi_ver ==
+	    MLX4_IB_UVERBS_NO_DEV_CAPS_ABI_VERSION) {
+		resp_v3.qp_tab_size      = dev->dev->caps.num_qps;
+		resp_v3.bf_reg_size      = dev->dev->caps.bf_reg_size;
+		resp_v3.bf_regs_per_page = dev->dev->caps.bf_regs_per_page;
+	} else {
+		resp.dev_caps	      = dev->dev->caps.userspace_caps;
+		resp.qp_tab_size      = dev->dev->caps.num_qps;
+		resp.bf_reg_size      = dev->dev->caps.bf_reg_size;
+		resp.bf_regs_per_page = dev->dev->caps.bf_regs_per_page;
+		resp.cqe_size	      = dev->dev->caps.cqe_size;
+	}
+
+	err = mlx4_uar_alloc(to_mdev(ibdev)->dev, &context->uar);
+	if (err)
+		return err;
+
+	INIT_LIST_HEAD(&context->db_page_list);
+	mutex_init(&context->db_page_mutex);
+
+	INIT_LIST_HEAD(&context->wqn_ranges_list);
+	mutex_init(&context->wqn_ranges_mutex);
+
+	if (ibdev->ops.uverbs_abi_ver == MLX4_IB_UVERBS_NO_DEV_CAPS_ABI_VERSION)
+		err = ib_copy_to_udata(udata, &resp_v3, sizeof(resp_v3));
+	else
+		err = ib_copy_to_udata(udata, &resp, sizeof(resp));
+
+	if (err) {
+		mlx4_uar_free(to_mdev(ibdev)->dev, &context->uar);
+		return -EFAULT;
+	}
+
+	return err;
+}
+
+static void mlx4_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
+{
+	struct mlx4_ib_ucontext *context = to_mucontext(ibcontext);
+
+	mlx4_uar_free(to_mdev(ibcontext->device)->dev, &context->uar);
+}
+
+static void mlx4_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
+{
+}
+
+static int mlx4_ib_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
+{
+	struct mlx4_ib_dev *dev = to_mdev(context->device);
+
+	switch (vma->vm_pgoff) {
+	case 0:
+		return rdma_user_mmap_io(context, vma,
+					 to_mucontext(context)->uar.pfn,
+					 PAGE_SIZE,
+					 pgprot_noncached(vma->vm_page_prot),
+					 NULL);
+
+	case 1:
+		if (dev->dev->caps.bf_reg_size == 0)
+			return -EINVAL;
+		return rdma_user_mmap_io(
+			context, vma,
+			to_mucontext(context)->uar.pfn +
+				dev->dev->caps.num_uars,
+			PAGE_SIZE, pgprot_writecombine(vma->vm_page_prot),
+			NULL);
+
+	case 3: {
+		struct mlx4_clock_params params;
+		int ret;
+
+		ret = mlx4_get_internal_clock_params(dev->dev, &params);
+		if (ret)
+			return ret;
+
+		return rdma_user_mmap_io(
+			context, vma,
+			(pci_resource_start(dev->dev->persist->pdev,
+					    params.bar) +
+			 params.offset) >>
+				PAGE_SHIFT,
+			PAGE_SIZE, pgprot_noncached(vma->vm_page_prot),
+			NULL);
+	}
+
+	default:
+		return -EINVAL;
+	}
+}
+
+static int mlx4_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
+{
+	struct mlx4_ib_pd *pd = to_mpd(ibpd);
+	struct ib_device *ibdev = ibpd->device;
+	int err;
+
+	err = mlx4_pd_alloc(to_mdev(ibdev)->dev, &pd->pdn);
+	if (err)
+		return err;
+
+	if (udata && ib_copy_to_udata(udata, &pd->pdn, sizeof(__u32))) {
+		mlx4_pd_free(to_mdev(ibdev)->dev, pd->pdn);
+		return -EFAULT;
+	}
+	return 0;
+}
+
+static int mlx4_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
+{
+	mlx4_pd_free(to_mdev(pd->device)->dev, to_mpd(pd)->pdn);
+	return 0;
+}
+
+static int mlx4_ib_alloc_xrcd(struct ib_xrcd *ibxrcd, struct ib_udata *udata)
+{
+	struct mlx4_ib_dev *dev = to_mdev(ibxrcd->device);
+	struct mlx4_ib_xrcd *xrcd = to_mxrcd(ibxrcd);
+	struct ib_cq_init_attr cq_attr = {};
+	int err;
+
+	if (!(dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
+		return -EOPNOTSUPP;
+
+	err = mlx4_xrcd_alloc(dev->dev, &xrcd->xrcdn);
+	if (err)
+		return err;
+
+	xrcd->pd = ib_alloc_pd(ibxrcd->device, 0);
+	if (IS_ERR(xrcd->pd)) {
+		err = PTR_ERR(xrcd->pd);
+		goto err2;
+	}
+
+	cq_attr.cqe = 1;
+	xrcd->cq = ib_create_cq(ibxrcd->device, NULL, NULL, xrcd, &cq_attr);
+	if (IS_ERR(xrcd->cq)) {
+		err = PTR_ERR(xrcd->cq);
+		goto err3;
+	}
+
+	return 0;
+
+err3:
+	ib_dealloc_pd(xrcd->pd);
+err2:
+	mlx4_xrcd_free(dev->dev, xrcd->xrcdn);
+	return err;
+}
+
+static int mlx4_ib_dealloc_xrcd(struct ib_xrcd *xrcd, struct ib_udata *udata)
+{
+	ib_destroy_cq(to_mxrcd(xrcd)->cq);
+	ib_dealloc_pd(to_mxrcd(xrcd)->pd);
+	mlx4_xrcd_free(to_mdev(xrcd->device)->dev, to_mxrcd(xrcd)->xrcdn);
+	return 0;
+}
+
+static int add_gid_entry(struct ib_qp *ibqp, union ib_gid *gid)
+{
+	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
+	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
+	struct mlx4_ib_gid_entry *ge;
+
+	ge = kzalloc(sizeof *ge, GFP_KERNEL);
+	if (!ge)
+		return -ENOMEM;
+
+	ge->gid = *gid;
+	if (mlx4_ib_add_mc(mdev, mqp, gid)) {
+		ge->port = mqp->port;
+		ge->added = 1;
+	}
+
+	mutex_lock(&mqp->mutex);
+	list_add_tail(&ge->list, &mqp->gid_list);
+	mutex_unlock(&mqp->mutex);
+
+	return 0;
+}
+
+static void mlx4_ib_delete_counters_table(struct mlx4_ib_dev *ibdev,
+					  struct mlx4_ib_counters *ctr_table)
+{
+	struct counter_index *counter, *tmp_count;
+
+	mutex_lock(&ctr_table->mutex);
+	list_for_each_entry_safe(counter, tmp_count, &ctr_table->counters_list,
+				 list) {
+		if (counter->allocated)
+			mlx4_counter_free(ibdev->dev, counter->index);
+		list_del(&counter->list);
+		kfree(counter);
+	}
+	mutex_unlock(&ctr_table->mutex);
+}
+
+int mlx4_ib_add_mc(struct mlx4_ib_dev *mdev, struct mlx4_ib_qp *mqp,
+		   union ib_gid *gid)
+{
+	struct net_device *ndev;
+	int ret = 0;
+
+	if (!mqp->port)
+		return 0;
+
+	spin_lock_bh(&mdev->iboe.lock);
+	ndev = mdev->iboe.netdevs[mqp->port - 1];
+	if (ndev)
+		dev_hold(ndev);
+	spin_unlock_bh(&mdev->iboe.lock);
+
+	if (ndev) {
+		ret = 1;
+		dev_put(ndev);
+	}
+
+	return ret;
+}
+
+struct mlx4_ib_steering {
+	struct list_head list;
+	struct mlx4_flow_reg_id reg_id;
+	union ib_gid gid;
+};
+
+#define LAST_ETH_FIELD vlan_tag
+#define LAST_IB_FIELD sl
+#define LAST_IPV4_FIELD dst_ip
+#define LAST_TCP_UDP_FIELD src_port
+
+/* Field is the last supported field */
+#define FIELDS_NOT_SUPPORTED(filter, field)\
+	memchr_inv((void *)&filter.field  +\
+		   sizeof(filter.field), 0,\
+		   sizeof(filter) -\
+		   offsetof(typeof(filter), field) -\
+		   sizeof(filter.field))
+
+static int parse_flow_attr(struct mlx4_dev *dev,
+			   u32 qp_num,
+			   union ib_flow_spec *ib_spec,
+			   struct _rule_hw *mlx4_spec)
+{
+	enum mlx4_net_trans_rule_id type;
+
+	switch (ib_spec->type) {
+	case IB_FLOW_SPEC_ETH:
+		if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
+			return -ENOTSUPP;
+
+		type = MLX4_NET_TRANS_RULE_ID_ETH;
+		memcpy(mlx4_spec->eth.dst_mac, ib_spec->eth.val.dst_mac,
+		       ETH_ALEN);
+		memcpy(mlx4_spec->eth.dst_mac_msk, ib_spec->eth.mask.dst_mac,
+		       ETH_ALEN);
+		mlx4_spec->eth.vlan_tag = ib_spec->eth.val.vlan_tag;
+		mlx4_spec->eth.vlan_tag_msk = ib_spec->eth.mask.vlan_tag;
+		break;
+	case IB_FLOW_SPEC_IB:
+		if (FIELDS_NOT_SUPPORTED(ib_spec->ib.mask, LAST_IB_FIELD))
+			return -ENOTSUPP;
+
+		type = MLX4_NET_TRANS_RULE_ID_IB;
+		mlx4_spec->ib.l3_qpn =
+			cpu_to_be32(qp_num);
+		mlx4_spec->ib.qpn_mask =
+			cpu_to_be32(MLX4_IB_FLOW_QPN_MASK);
+		break;
+
+
+	case IB_FLOW_SPEC_IPV4:
+		if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
+			return -ENOTSUPP;
+
+		type = MLX4_NET_TRANS_RULE_ID_IPV4;
+		mlx4_spec->ipv4.src_ip = ib_spec->ipv4.val.src_ip;
+		mlx4_spec->ipv4.src_ip_msk = ib_spec->ipv4.mask.src_ip;
+		mlx4_spec->ipv4.dst_ip = ib_spec->ipv4.val.dst_ip;
+		mlx4_spec->ipv4.dst_ip_msk = ib_spec->ipv4.mask.dst_ip;
+		break;
+
+	case IB_FLOW_SPEC_TCP:
+	case IB_FLOW_SPEC_UDP:
+		if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask, LAST_TCP_UDP_FIELD))
+			return -ENOTSUPP;
+
+		type = ib_spec->type == IB_FLOW_SPEC_TCP ?
+					MLX4_NET_TRANS_RULE_ID_TCP :
+					MLX4_NET_TRANS_RULE_ID_UDP;
+		mlx4_spec->tcp_udp.dst_port = ib_spec->tcp_udp.val.dst_port;
+		mlx4_spec->tcp_udp.dst_port_msk = ib_spec->tcp_udp.mask.dst_port;
+		mlx4_spec->tcp_udp.src_port = ib_spec->tcp_udp.val.src_port;
+		mlx4_spec->tcp_udp.src_port_msk = ib_spec->tcp_udp.mask.src_port;
+		break;
+
+	default:
+		return -EINVAL;
+	}
+	if (mlx4_map_sw_to_hw_steering_id(dev, type) < 0 ||
+	    mlx4_hw_rule_sz(dev, type) < 0)
+		return -EINVAL;
+	mlx4_spec->id = cpu_to_be16(mlx4_map_sw_to_hw_steering_id(dev, type));
+	mlx4_spec->size = mlx4_hw_rule_sz(dev, type) >> 2;
+	return mlx4_hw_rule_sz(dev, type);
+}
+
+struct default_rules {
+	__u32 mandatory_fields[IB_FLOW_SPEC_SUPPORT_LAYERS];
+	__u32 mandatory_not_fields[IB_FLOW_SPEC_SUPPORT_LAYERS];
+	__u32 rules_create_list[IB_FLOW_SPEC_SUPPORT_LAYERS];
+	__u8  link_layer;
+};
+static const struct default_rules default_table[] = {
+	{
+		.mandatory_fields = {IB_FLOW_SPEC_IPV4},
+		.mandatory_not_fields = {IB_FLOW_SPEC_ETH},
+		.rules_create_list = {IB_FLOW_SPEC_IB},
+		.link_layer = IB_LINK_LAYER_INFINIBAND
+	}
+};
+
+static int __mlx4_ib_default_rules_match(struct ib_qp *qp,
+					 struct ib_flow_attr *flow_attr)
+{
+	int i, j, k;
+	void *ib_flow;
+	const struct default_rules *pdefault_rules = default_table;
+	u8 link_layer = rdma_port_get_link_layer(qp->device, flow_attr->port);
+
+	for (i = 0; i < ARRAY_SIZE(default_table); i++, pdefault_rules++) {
+		__u32 field_types[IB_FLOW_SPEC_SUPPORT_LAYERS];
+		memset(&field_types, 0, sizeof(field_types));
+
+		if (link_layer != pdefault_rules->link_layer)
+			continue;
+
+		ib_flow = flow_attr + 1;
+		/* we assume the specs are sorted */
+		for (j = 0, k = 0; k < IB_FLOW_SPEC_SUPPORT_LAYERS &&
+		     j < flow_attr->num_of_specs; k++) {
+			union ib_flow_spec *current_flow =
+				(union ib_flow_spec *)ib_flow;
+
+			/* same layer but different type */
+			if (((current_flow->type & IB_FLOW_SPEC_LAYER_MASK) ==
+			     (pdefault_rules->mandatory_fields[k] &
+			      IB_FLOW_SPEC_LAYER_MASK)) &&
+			    (current_flow->type !=
+			     pdefault_rules->mandatory_fields[k]))
+				goto out;
+
+			/* same layer, try match next one */
+			if (current_flow->type ==
+			    pdefault_rules->mandatory_fields[k]) {
+				j++;
+				ib_flow +=
+					((union ib_flow_spec *)ib_flow)->size;
+			}
+		}
+
+		ib_flow = flow_attr + 1;
+		for (j = 0; j < flow_attr->num_of_specs;
+		     j++, ib_flow += ((union ib_flow_spec *)ib_flow)->size)
+			for (k = 0; k < IB_FLOW_SPEC_SUPPORT_LAYERS; k++)
+				/* same layer and same type */
+				if (((union ib_flow_spec *)ib_flow)->type ==
+				    pdefault_rules->mandatory_not_fields[k])
+					goto out;
+
+		return i;
+	}
+out:
+	return -1;
+}
+
+static int __mlx4_ib_create_default_rules(
+		struct mlx4_ib_dev *mdev,
+		struct ib_qp *qp,
+		const struct default_rules *pdefault_rules,
+		struct _rule_hw *mlx4_spec) {
+	int size = 0;
+	int i;
+
+	for (i = 0; i < ARRAY_SIZE(pdefault_rules->rules_create_list); i++) {
+		union ib_flow_spec ib_spec = {};
+		int ret;
+
+		switch (pdefault_rules->rules_create_list[i]) {
+		case 0:
+			/* no rule */
+			continue;
+		case IB_FLOW_SPEC_IB:
+			ib_spec.type = IB_FLOW_SPEC_IB;
+			ib_spec.size = sizeof(struct ib_flow_spec_ib);
+
+			break;
+		default:
+			/* invalid rule */
+			return -EINVAL;
+		}
+		/* We must put empty rule, qpn is being ignored */
+		ret = parse_flow_attr(mdev->dev, 0, &ib_spec,
+				      mlx4_spec);
+		if (ret < 0) {
+			pr_info("invalid parsing\n");
+			return -EINVAL;
+		}
+
+		mlx4_spec = (void *)mlx4_spec + ret;
+		size += ret;
+	}
+	return size;
+}
+
+static int __mlx4_ib_create_flow(struct ib_qp *qp, struct ib_flow_attr *flow_attr,
+			  int domain,
+			  enum mlx4_net_trans_promisc_mode flow_type,
+			  u64 *reg_id)
+{
+	int ret, i;
+	int size = 0;
+	void *ib_flow;
+	struct mlx4_ib_dev *mdev = to_mdev(qp->device);
+	struct mlx4_cmd_mailbox *mailbox;
+	struct mlx4_net_trans_rule_hw_ctrl *ctrl;
+	int default_flow;
+
+	if (flow_attr->priority > MLX4_IB_FLOW_MAX_PRIO) {
+		pr_err("Invalid priority value %d\n", flow_attr->priority);
+		return -EINVAL;
+	}
+
+	if (mlx4_map_sw_to_hw_steering_mode(mdev->dev, flow_type) < 0)
+		return -EINVAL;
+
+	mailbox = mlx4_alloc_cmd_mailbox(mdev->dev);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	ctrl = mailbox->buf;
+
+	ctrl->prio = cpu_to_be16(domain | flow_attr->priority);
+	ctrl->type = mlx4_map_sw_to_hw_steering_mode(mdev->dev, flow_type);
+	ctrl->port = flow_attr->port;
+	ctrl->qpn = cpu_to_be32(qp->qp_num);
+
+	ib_flow = flow_attr + 1;
+	size += sizeof(struct mlx4_net_trans_rule_hw_ctrl);
+	/* Add default flows */
+	default_flow = __mlx4_ib_default_rules_match(qp, flow_attr);
+	if (default_flow >= 0) {
+		ret = __mlx4_ib_create_default_rules(
+				mdev, qp, default_table + default_flow,
+				mailbox->buf + size);
+		if (ret < 0) {
+			mlx4_free_cmd_mailbox(mdev->dev, mailbox);
+			return -EINVAL;
+		}
+		size += ret;
+	}
+	for (i = 0; i < flow_attr->num_of_specs; i++) {
+		ret = parse_flow_attr(mdev->dev, qp->qp_num, ib_flow,
+				      mailbox->buf + size);
+		if (ret < 0) {
+			mlx4_free_cmd_mailbox(mdev->dev, mailbox);
+			return -EINVAL;
+		}
+		ib_flow += ((union ib_flow_spec *) ib_flow)->size;
+		size += ret;
+	}
+
+	if (mlx4_is_master(mdev->dev) && flow_type == MLX4_FS_REGULAR &&
+	    flow_attr->num_of_specs == 1) {
+		struct _rule_hw *rule_header = (struct _rule_hw *)(ctrl + 1);
+		enum ib_flow_spec_type header_spec =
+			((union ib_flow_spec *)(flow_attr + 1))->type;
+
+		if (header_spec == IB_FLOW_SPEC_ETH)
+			mlx4_handle_eth_header_mcast_prio(ctrl, rule_header);
+	}
+
+	ret = mlx4_cmd_imm(mdev->dev, mailbox->dma, reg_id, size >> 2, 0,
+			   MLX4_QP_FLOW_STEERING_ATTACH, MLX4_CMD_TIME_CLASS_A,
+			   MLX4_CMD_NATIVE);
+	if (ret == -ENOMEM)
+		pr_err("mcg table is full. Fail to register network rule.\n");
+	else if (ret == -ENXIO)
+		pr_err("Device managed flow steering is disabled. Fail to register network rule.\n");
+	else if (ret)
+		pr_err("Invalid argument. Fail to register network rule.\n");
+
+	mlx4_free_cmd_mailbox(mdev->dev, mailbox);
+	return ret;
+}
+
+static int __mlx4_ib_destroy_flow(struct mlx4_dev *dev, u64 reg_id)
+{
+	int err;
+	err = mlx4_cmd(dev, reg_id, 0, 0,
+		       MLX4_QP_FLOW_STEERING_DETACH, MLX4_CMD_TIME_CLASS_A,
+		       MLX4_CMD_NATIVE);
+	if (err)
+		pr_err("Fail to detach network rule. registration id = 0x%llx\n",
+		       reg_id);
+	return err;
+}
+
+static int mlx4_ib_tunnel_steer_add(struct ib_qp *qp, struct ib_flow_attr *flow_attr,
+				    u64 *reg_id)
+{
+	void *ib_flow;
+	union ib_flow_spec *ib_spec;
+	struct mlx4_dev	*dev = to_mdev(qp->device)->dev;
+	int err = 0;
+
+	if (dev->caps.tunnel_offload_mode != MLX4_TUNNEL_OFFLOAD_MODE_VXLAN ||
+	    dev->caps.dmfs_high_steer_mode == MLX4_STEERING_DMFS_A0_STATIC)
+		return 0; /* do nothing */
+
+	ib_flow = flow_attr + 1;
+	ib_spec = (union ib_flow_spec *)ib_flow;
+
+	if (ib_spec->type !=  IB_FLOW_SPEC_ETH || flow_attr->num_of_specs != 1)
+		return 0; /* do nothing */
+
+	err = mlx4_tunnel_steer_add(to_mdev(qp->device)->dev, ib_spec->eth.val.dst_mac,
+				    flow_attr->port, qp->qp_num,
+				    MLX4_DOMAIN_UVERBS | (flow_attr->priority & 0xff),
+				    reg_id);
+	return err;
+}
+
+static int mlx4_ib_add_dont_trap_rule(struct mlx4_dev *dev,
+				      struct ib_flow_attr *flow_attr,
+				      enum mlx4_net_trans_promisc_mode *type)
+{
+	int err = 0;
+
+	if (!(dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_DMFS_UC_MC_SNIFFER) ||
+	    (dev->caps.dmfs_high_steer_mode == MLX4_STEERING_DMFS_A0_STATIC) ||
+	    (flow_attr->num_of_specs > 1) || (flow_attr->priority != 0)) {
+		return -EOPNOTSUPP;
+	}
+
+	if (flow_attr->num_of_specs == 0) {
+		type[0] = MLX4_FS_MC_SNIFFER;
+		type[1] = MLX4_FS_UC_SNIFFER;
+	} else {
+		union ib_flow_spec *ib_spec;
+
+		ib_spec = (union ib_flow_spec *)(flow_attr + 1);
+		if (ib_spec->type !=  IB_FLOW_SPEC_ETH)
+			return -EINVAL;
+
+		/* if all is zero than MC and UC */
+		if (is_zero_ether_addr(ib_spec->eth.mask.dst_mac)) {
+			type[0] = MLX4_FS_MC_SNIFFER;
+			type[1] = MLX4_FS_UC_SNIFFER;
+		} else {
+			u8 mac[ETH_ALEN] = {ib_spec->eth.mask.dst_mac[0] ^ 0x01,
+					    ib_spec->eth.mask.dst_mac[1],
+					    ib_spec->eth.mask.dst_mac[2],
+					    ib_spec->eth.mask.dst_mac[3],
+					    ib_spec->eth.mask.dst_mac[4],
+					    ib_spec->eth.mask.dst_mac[5]};
+
+			/* Above xor was only on MC bit, non empty mask is valid
+			 * only if this bit is set and rest are zero.
+			 */
+			if (!is_zero_ether_addr(&mac[0]))
+				return -EINVAL;
+
+			if (is_multicast_ether_addr(ib_spec->eth.val.dst_mac))
+				type[0] = MLX4_FS_MC_SNIFFER;
+			else
+				type[0] = MLX4_FS_UC_SNIFFER;
+		}
+	}
+
+	return err;
+}
+
+static struct ib_flow *mlx4_ib_create_flow(struct ib_qp *qp,
+					   struct ib_flow_attr *flow_attr,
+					   struct ib_udata *udata)
+{
+	int err = 0, i = 0, j = 0;
+	struct mlx4_ib_flow *mflow;
+	enum mlx4_net_trans_promisc_mode type[2];
+	struct mlx4_dev *dev = (to_mdev(qp->device))->dev;
+	int is_bonded = mlx4_is_bonded(dev);
+
+	if (flow_attr->port < 1 || flow_attr->port > qp->device->phys_port_cnt)
+		return ERR_PTR(-EINVAL);
+
+	if (flow_attr->flags & ~IB_FLOW_ATTR_FLAGS_DONT_TRAP)
+		return ERR_PTR(-EOPNOTSUPP);
+
+	if ((flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP) &&
+	    (flow_attr->type != IB_FLOW_ATTR_NORMAL))
+		return ERR_PTR(-EOPNOTSUPP);
+
+	if (udata &&
+	    udata->inlen && !ib_is_udata_cleared(udata, 0, udata->inlen))
+		return ERR_PTR(-EOPNOTSUPP);
+
+	memset(type, 0, sizeof(type));
+
+	mflow = kzalloc(sizeof(*mflow), GFP_KERNEL);
+	if (!mflow) {
+		err = -ENOMEM;
+		goto err_free;
+	}
+
+	switch (flow_attr->type) {
+	case IB_FLOW_ATTR_NORMAL:
+		/* If dont trap flag (continue match) is set, under specific
+		 * condition traffic be replicated to given qp,
+		 * without stealing it
+		 */
+		if (unlikely(flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)) {
+			err = mlx4_ib_add_dont_trap_rule(dev,
+							 flow_attr,
+							 type);
+			if (err)
+				goto err_free;
+		} else {
+			type[0] = MLX4_FS_REGULAR;
+		}
+		break;
+
+	case IB_FLOW_ATTR_ALL_DEFAULT:
+		type[0] = MLX4_FS_ALL_DEFAULT;
+		break;
+
+	case IB_FLOW_ATTR_MC_DEFAULT:
+		type[0] = MLX4_FS_MC_DEFAULT;
+		break;
+
+	case IB_FLOW_ATTR_SNIFFER:
+		type[0] = MLX4_FS_MIRROR_RX_PORT;
+		type[1] = MLX4_FS_MIRROR_SX_PORT;
+		break;
+
+	default:
+		err = -EINVAL;
+		goto err_free;
+	}
+
+	while (i < ARRAY_SIZE(type) && type[i]) {
+		err = __mlx4_ib_create_flow(qp, flow_attr, MLX4_DOMAIN_UVERBS,
+					    type[i], &mflow->reg_id[i].id);
+		if (err)
+			goto err_create_flow;
+		if (is_bonded) {
+			/* Application always sees one port so the mirror rule
+			 * must be on port #2
+			 */
+			flow_attr->port = 2;
+			err = __mlx4_ib_create_flow(qp, flow_attr,
+						    MLX4_DOMAIN_UVERBS, type[j],
+						    &mflow->reg_id[j].mirror);
+			flow_attr->port = 1;
+			if (err)
+				goto err_create_flow;
+			j++;
+		}
+
+		i++;
+	}
+
+	if (i < ARRAY_SIZE(type) && flow_attr->type == IB_FLOW_ATTR_NORMAL) {
+		err = mlx4_ib_tunnel_steer_add(qp, flow_attr,
+					       &mflow->reg_id[i].id);
+		if (err)
+			goto err_create_flow;
+
+		if (is_bonded) {
+			flow_attr->port = 2;
+			err = mlx4_ib_tunnel_steer_add(qp, flow_attr,
+						       &mflow->reg_id[j].mirror);
+			flow_attr->port = 1;
+			if (err)
+				goto err_create_flow;
+			j++;
+		}
+		/* function to create mirror rule */
+		i++;
+	}
+
+	return &mflow->ibflow;
+
+err_create_flow:
+	while (i) {
+		(void)__mlx4_ib_destroy_flow(to_mdev(qp->device)->dev,
+					     mflow->reg_id[i].id);
+		i--;
+	}
+
+	while (j) {
+		(void)__mlx4_ib_destroy_flow(to_mdev(qp->device)->dev,
+					     mflow->reg_id[j].mirror);
+		j--;
+	}
+err_free:
+	kfree(mflow);
+	return ERR_PTR(err);
+}
+
+static int mlx4_ib_destroy_flow(struct ib_flow *flow_id)
+{
+	int err, ret = 0;
+	int i = 0;
+	struct mlx4_ib_dev *mdev = to_mdev(flow_id->qp->device);
+	struct mlx4_ib_flow *mflow = to_mflow(flow_id);
+
+	while (i < ARRAY_SIZE(mflow->reg_id) && mflow->reg_id[i].id) {
+		err = __mlx4_ib_destroy_flow(mdev->dev, mflow->reg_id[i].id);
+		if (err)
+			ret = err;
+		if (mflow->reg_id[i].mirror) {
+			err = __mlx4_ib_destroy_flow(mdev->dev,
+						     mflow->reg_id[i].mirror);
+			if (err)
+				ret = err;
+		}
+		i++;
+	}
+
+	kfree(mflow);
+	return ret;
+}
+
+static int mlx4_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
+{
+	int err;
+	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
+	struct mlx4_dev	*dev = mdev->dev;
+	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
+	struct mlx4_ib_steering *ib_steering = NULL;
+	enum mlx4_protocol prot = MLX4_PROT_IB_IPV6;
+	struct mlx4_flow_reg_id	reg_id;
+
+	if (mdev->dev->caps.steering_mode ==
+	    MLX4_STEERING_MODE_DEVICE_MANAGED) {
+		ib_steering = kmalloc(sizeof(*ib_steering), GFP_KERNEL);
+		if (!ib_steering)
+			return -ENOMEM;
+	}
+
+	err = mlx4_multicast_attach(mdev->dev, &mqp->mqp, gid->raw, mqp->port,
+				    !!(mqp->flags &
+				       MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK),
+				    prot, &reg_id.id);
+	if (err) {
+		pr_err("multicast attach op failed, err %d\n", err);
+		goto err_malloc;
+	}
+
+	reg_id.mirror = 0;
+	if (mlx4_is_bonded(dev)) {
+		err = mlx4_multicast_attach(mdev->dev, &mqp->mqp, gid->raw,
+					    (mqp->port == 1) ? 2 : 1,
+					    !!(mqp->flags &
+					    MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK),
+					    prot, &reg_id.mirror);
+		if (err)
+			goto err_add;
+	}
+
+	err = add_gid_entry(ibqp, gid);
+	if (err)
+		goto err_add;
+
+	if (ib_steering) {
+		memcpy(ib_steering->gid.raw, gid->raw, 16);
+		ib_steering->reg_id = reg_id;
+		mutex_lock(&mqp->mutex);
+		list_add(&ib_steering->list, &mqp->steering_rules);
+		mutex_unlock(&mqp->mutex);
+	}
+	return 0;
+
+err_add:
+	mlx4_multicast_detach(mdev->dev, &mqp->mqp, gid->raw,
+			      prot, reg_id.id);
+	if (reg_id.mirror)
+		mlx4_multicast_detach(mdev->dev, &mqp->mqp, gid->raw,
+				      prot, reg_id.mirror);
+err_malloc:
+	kfree(ib_steering);
+
+	return err;
+}
+
+static struct mlx4_ib_gid_entry *find_gid_entry(struct mlx4_ib_qp *qp, u8 *raw)
+{
+	struct mlx4_ib_gid_entry *ge;
+	struct mlx4_ib_gid_entry *tmp;
+	struct mlx4_ib_gid_entry *ret = NULL;
+
+	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
+		if (!memcmp(raw, ge->gid.raw, 16)) {
+			ret = ge;
+			break;
+		}
+	}
+
+	return ret;
+}
+
+static int mlx4_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
+{
+	int err;
+	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
+	struct mlx4_dev *dev = mdev->dev;
+	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
+	struct net_device *ndev;
+	struct mlx4_ib_gid_entry *ge;
+	struct mlx4_flow_reg_id reg_id = {0, 0};
+	enum mlx4_protocol prot =  MLX4_PROT_IB_IPV6;
+
+	if (mdev->dev->caps.steering_mode ==
+	    MLX4_STEERING_MODE_DEVICE_MANAGED) {
+		struct mlx4_ib_steering *ib_steering;
+
+		mutex_lock(&mqp->mutex);
+		list_for_each_entry(ib_steering, &mqp->steering_rules, list) {
+			if (!memcmp(ib_steering->gid.raw, gid->raw, 16)) {
+				list_del(&ib_steering->list);
+				break;
+			}
+		}
+		mutex_unlock(&mqp->mutex);
+		if (&ib_steering->list == &mqp->steering_rules) {
+			pr_err("Couldn't find reg_id for mgid. Steering rule is left attached\n");
+			return -EINVAL;
+		}
+		reg_id = ib_steering->reg_id;
+		kfree(ib_steering);
+	}
+
+	err = mlx4_multicast_detach(mdev->dev, &mqp->mqp, gid->raw,
+				    prot, reg_id.id);
+	if (err)
+		return err;
+
+	if (mlx4_is_bonded(dev)) {
+		err = mlx4_multicast_detach(mdev->dev, &mqp->mqp, gid->raw,
+					    prot, reg_id.mirror);
+		if (err)
+			return err;
+	}
+
+	mutex_lock(&mqp->mutex);
+	ge = find_gid_entry(mqp, gid->raw);
+	if (ge) {
+		spin_lock_bh(&mdev->iboe.lock);
+		ndev = ge->added ? mdev->iboe.netdevs[ge->port - 1] : NULL;
+		if (ndev)
+			dev_hold(ndev);
+		spin_unlock_bh(&mdev->iboe.lock);
+		if (ndev)
+			dev_put(ndev);
+		list_del(&ge->list);
+		kfree(ge);
+	} else
+		pr_warn("could not find mgid entry\n");
+
+	mutex_unlock(&mqp->mutex);
+
+	return 0;
+}
+
+static int init_node_data(struct mlx4_ib_dev *dev)
+{
+	struct ib_smp *in_mad  = NULL;
+	struct ib_smp *out_mad = NULL;
+	int mad_ifc_flags = MLX4_MAD_IFC_IGNORE_KEYS;
+	int err = -ENOMEM;
+
+	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
+	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
+	if (!in_mad || !out_mad)
+		goto out;
+
+	init_query_mad(in_mad);
+	in_mad->attr_id = IB_SMP_ATTR_NODE_DESC;
+	if (mlx4_is_master(dev->dev))
+		mad_ifc_flags |= MLX4_MAD_IFC_NET_VIEW;
+
+	err = mlx4_MAD_IFC(dev, mad_ifc_flags, 1, NULL, NULL, in_mad, out_mad);
+	if (err)
+		goto out;
+
+	memcpy(dev->ib_dev.node_desc, out_mad->data, IB_DEVICE_NODE_DESC_MAX);
+
+	in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
+
+	err = mlx4_MAD_IFC(dev, mad_ifc_flags, 1, NULL, NULL, in_mad, out_mad);
+	if (err)
+		goto out;
+
+	dev->dev->rev_id = be32_to_cpup((__be32 *) (out_mad->data + 32));
+	memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8);
+
+out:
+	kfree(in_mad);
+	kfree(out_mad);
+	return err;
+}
+
+static ssize_t hca_type_show(struct device *device,
+			     struct device_attribute *attr, char *buf)
+{
+	struct mlx4_ib_dev *dev =
+		rdma_device_to_drv_device(device, struct mlx4_ib_dev, ib_dev);
+	return sprintf(buf, "MT%d\n", dev->dev->persist->pdev->device);
+}
+static DEVICE_ATTR_RO(hca_type);
+
+static ssize_t hw_rev_show(struct device *device,
+			   struct device_attribute *attr, char *buf)
+{
+	struct mlx4_ib_dev *dev =
+		rdma_device_to_drv_device(device, struct mlx4_ib_dev, ib_dev);
+	return sprintf(buf, "%x\n", dev->dev->rev_id);
+}
+static DEVICE_ATTR_RO(hw_rev);
+
+static ssize_t board_id_show(struct device *device,
+			     struct device_attribute *attr, char *buf)
+{
+	struct mlx4_ib_dev *dev =
+		rdma_device_to_drv_device(device, struct mlx4_ib_dev, ib_dev);
+
+	return sprintf(buf, "%.*s\n", MLX4_BOARD_ID_LEN,
+		       dev->dev->board_id);
+}
+static DEVICE_ATTR_RO(board_id);
+
+static struct attribute *mlx4_class_attributes[] = {
+	&dev_attr_hw_rev.attr,
+	&dev_attr_hca_type.attr,
+	&dev_attr_board_id.attr,
+	NULL
+};
+
+static const struct attribute_group mlx4_attr_group = {
+	.attrs = mlx4_class_attributes,
+};
+
+struct diag_counter {
+	const char *name;
+	u32 offset;
+};
+
+#define DIAG_COUNTER(_name, _offset)			\
+	{ .name = #_name, .offset = _offset }
+
+static const struct diag_counter diag_basic[] = {
+	DIAG_COUNTER(rq_num_lle, 0x00),
+	DIAG_COUNTER(sq_num_lle, 0x04),
+	DIAG_COUNTER(rq_num_lqpoe, 0x08),
+	DIAG_COUNTER(sq_num_lqpoe, 0x0C),
+	DIAG_COUNTER(rq_num_lpe, 0x18),
+	DIAG_COUNTER(sq_num_lpe, 0x1C),
+	DIAG_COUNTER(rq_num_wrfe, 0x20),
+	DIAG_COUNTER(sq_num_wrfe, 0x24),
+	DIAG_COUNTER(sq_num_mwbe, 0x2C),
+	DIAG_COUNTER(sq_num_bre, 0x34),
+	DIAG_COUNTER(sq_num_rire, 0x44),
+	DIAG_COUNTER(rq_num_rire, 0x48),
+	DIAG_COUNTER(sq_num_rae, 0x4C),
+	DIAG_COUNTER(rq_num_rae, 0x50),
+	DIAG_COUNTER(sq_num_roe, 0x54),
+	DIAG_COUNTER(sq_num_tree, 0x5C),
+	DIAG_COUNTER(sq_num_rree, 0x64),
+	DIAG_COUNTER(rq_num_rnr, 0x68),
+	DIAG_COUNTER(sq_num_rnr, 0x6C),
+	DIAG_COUNTER(rq_num_oos, 0x100),
+	DIAG_COUNTER(sq_num_oos, 0x104),
+};
+
+static const struct diag_counter diag_ext[] = {
+	DIAG_COUNTER(rq_num_dup, 0x130),
+	DIAG_COUNTER(sq_num_to, 0x134),
+};
+
+static const struct diag_counter diag_device_only[] = {
+	DIAG_COUNTER(num_cqovf, 0x1A0),
+	DIAG_COUNTER(rq_num_udsdprd, 0x118),
+};
+
+static struct rdma_hw_stats *mlx4_ib_alloc_hw_stats(struct ib_device *ibdev,
+						    u8 port_num)
+{
+	struct mlx4_ib_dev *dev = to_mdev(ibdev);
+	struct mlx4_ib_diag_counters *diag = dev->diag_counters;
+
+	if (!diag[!!port_num].name)
+		return NULL;
+
+	return rdma_alloc_hw_stats_struct(diag[!!port_num].name,
+					  diag[!!port_num].num_counters,
+					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
+}
+
+static int mlx4_ib_get_hw_stats(struct ib_device *ibdev,
+				struct rdma_hw_stats *stats,
+				u8 port, int index)
+{
+	struct mlx4_ib_dev *dev = to_mdev(ibdev);
+	struct mlx4_ib_diag_counters *diag = dev->diag_counters;
+	u32 hw_value[ARRAY_SIZE(diag_device_only) +
+		ARRAY_SIZE(diag_ext) + ARRAY_SIZE(diag_basic)] = {};
+	int ret;
+	int i;
+
+	ret = mlx4_query_diag_counters(dev->dev,
+				       MLX4_OP_MOD_QUERY_TRANSPORT_CI_ERRORS,
+				       diag[!!port].offset, hw_value,
+				       diag[!!port].num_counters, port);
+
+	if (ret)
+		return ret;
+
+	for (i = 0; i < diag[!!port].num_counters; i++)
+		stats->value[i] = hw_value[i];
+
+	return diag[!!port].num_counters;
+}
+
+static int __mlx4_ib_alloc_diag_counters(struct mlx4_ib_dev *ibdev,
+					 const char ***name,
+					 u32 **offset,
+					 u32 *num,
+					 bool port)
+{
+	u32 num_counters;
+
+	num_counters = ARRAY_SIZE(diag_basic);
+
+	if (ibdev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_DIAG_PER_PORT)
+		num_counters += ARRAY_SIZE(diag_ext);
+
+	if (!port)
+		num_counters += ARRAY_SIZE(diag_device_only);
+
+	*name = kcalloc(num_counters, sizeof(**name), GFP_KERNEL);
+	if (!*name)
+		return -ENOMEM;
+
+	*offset = kcalloc(num_counters, sizeof(**offset), GFP_KERNEL);
+	if (!*offset)
+		goto err_name;
+
+	*num = num_counters;
+
+	return 0;
+
+err_name:
+	kfree(*name);
+	return -ENOMEM;
+}
+
+static void mlx4_ib_fill_diag_counters(struct mlx4_ib_dev *ibdev,
+				       const char **name,
+				       u32 *offset,
+				       bool port)
+{
+	int i;
+	int j;
+
+	for (i = 0, j = 0; i < ARRAY_SIZE(diag_basic); i++, j++) {
+		name[i] = diag_basic[i].name;
+		offset[i] = diag_basic[i].offset;
+	}
+
+	if (ibdev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_DIAG_PER_PORT) {
+		for (i = 0; i < ARRAY_SIZE(diag_ext); i++, j++) {
+			name[j] = diag_ext[i].name;
+			offset[j] = diag_ext[i].offset;
+		}
+	}
+
+	if (!port) {
+		for (i = 0; i < ARRAY_SIZE(diag_device_only); i++, j++) {
+			name[j] = diag_device_only[i].name;
+			offset[j] = diag_device_only[i].offset;
+		}
+	}
+}
+
+static const struct ib_device_ops mlx4_ib_hw_stats_ops = {
+	.alloc_hw_stats = mlx4_ib_alloc_hw_stats,
+	.get_hw_stats = mlx4_ib_get_hw_stats,
+};
+
+static int mlx4_ib_alloc_diag_counters(struct mlx4_ib_dev *ibdev)
+{
+	struct mlx4_ib_diag_counters *diag = ibdev->diag_counters;
+	int i;
+	int ret;
+	bool per_port = !!(ibdev->dev->caps.flags2 &
+		MLX4_DEV_CAP_FLAG2_DIAG_PER_PORT);
+
+	if (mlx4_is_slave(ibdev->dev))
+		return 0;
+
+	for (i = 0; i < MLX4_DIAG_COUNTERS_TYPES; i++) {
+		/* i == 1 means we are building port counters */
+		if (i && !per_port)
+			continue;
+
+		ret = __mlx4_ib_alloc_diag_counters(ibdev, &diag[i].name,
+						    &diag[i].offset,
+						    &diag[i].num_counters, i);
+		if (ret)
+			goto err_alloc;
+
+		mlx4_ib_fill_diag_counters(ibdev, diag[i].name,
+					   diag[i].offset, i);
+	}
+
+	ib_set_device_ops(&ibdev->ib_dev, &mlx4_ib_hw_stats_ops);
+
+	return 0;
+
+err_alloc:
+	if (i) {
+		kfree(diag[i - 1].name);
+		kfree(diag[i - 1].offset);
+	}
+
+	return ret;
+}
+
+static void mlx4_ib_diag_cleanup(struct mlx4_ib_dev *ibdev)
+{
+	int i;
+
+	for (i = 0; i < MLX4_DIAG_COUNTERS_TYPES; i++) {
+		kfree(ibdev->diag_counters[i].offset);
+		kfree(ibdev->diag_counters[i].name);
+	}
+}
+
+#define MLX4_IB_INVALID_MAC	((u64)-1)
+static void mlx4_ib_update_qps(struct mlx4_ib_dev *ibdev,
+			       struct net_device *dev,
+			       int port)
+{
+	u64 new_smac = 0;
+	u64 release_mac = MLX4_IB_INVALID_MAC;
+	struct mlx4_ib_qp *qp;
+
+	read_lock(&dev_base_lock);
+	new_smac = mlx4_mac_to_u64(dev->dev_addr);
+	read_unlock(&dev_base_lock);
+
+	atomic64_set(&ibdev->iboe.mac[port - 1], new_smac);
+
+	/* no need for update QP1 and mac registration in non-SRIOV */
+	if (!mlx4_is_mfunc(ibdev->dev))
+		return;
+
+	mutex_lock(&ibdev->qp1_proxy_lock[port - 1]);
+	qp = ibdev->qp1_proxy[port - 1];
+	if (qp) {
+		int new_smac_index;
+		u64 old_smac;
+		struct mlx4_update_qp_params update_params;
+
+		mutex_lock(&qp->mutex);
+		old_smac = qp->pri.smac;
+		if (new_smac == old_smac)
+			goto unlock;
+
+		new_smac_index = mlx4_register_mac(ibdev->dev, port, new_smac);
+
+		if (new_smac_index < 0)
+			goto unlock;
+
+		update_params.smac_index = new_smac_index;
+		if (mlx4_update_qp(ibdev->dev, qp->mqp.qpn, MLX4_UPDATE_QP_SMAC,
+				   &update_params)) {
+			release_mac = new_smac;
+			goto unlock;
+		}
+		/* if old port was zero, no mac was yet registered for this QP */
+		if (qp->pri.smac_port)
+			release_mac = old_smac;
+		qp->pri.smac = new_smac;
+		qp->pri.smac_port = port;
+		qp->pri.smac_index = new_smac_index;
+	}
+
+unlock:
+	if (release_mac != MLX4_IB_INVALID_MAC)
+		mlx4_unregister_mac(ibdev->dev, port, release_mac);
+	if (qp)
+		mutex_unlock(&qp->mutex);
+	mutex_unlock(&ibdev->qp1_proxy_lock[port - 1]);
+}
+
+static void mlx4_ib_scan_netdevs(struct mlx4_ib_dev *ibdev,
+				 struct net_device *dev,
+				 unsigned long event)
+
+{
+	struct mlx4_ib_iboe *iboe;
+	int update_qps_port = -1;
+	int port;
+
+	ASSERT_RTNL();
+
+	iboe = &ibdev->iboe;
+
+	spin_lock_bh(&iboe->lock);
+	mlx4_foreach_ib_transport_port(port, ibdev->dev) {
+
+		iboe->netdevs[port - 1] =
+			mlx4_get_protocol_dev(ibdev->dev, MLX4_PROT_ETH, port);
+
+		if (dev == iboe->netdevs[port - 1] &&
+		    (event == NETDEV_CHANGEADDR || event == NETDEV_REGISTER ||
+		     event == NETDEV_UP || event == NETDEV_CHANGE))
+			update_qps_port = port;
+
+		if (dev == iboe->netdevs[port - 1] &&
+		    (event == NETDEV_UP || event == NETDEV_DOWN)) {
+			enum ib_port_state port_state;
+			struct ib_event ibev = { };
+
+			if (ib_get_cached_port_state(&ibdev->ib_dev, port,
+						     &port_state))
+				continue;
+
+			if (event == NETDEV_UP &&
+			    (port_state != IB_PORT_ACTIVE ||
+			     iboe->last_port_state[port - 1] != IB_PORT_DOWN))
+				continue;
+			if (event == NETDEV_DOWN &&
+			    (port_state != IB_PORT_DOWN ||
+			     iboe->last_port_state[port - 1] != IB_PORT_ACTIVE))
+				continue;
+			iboe->last_port_state[port - 1] = port_state;
+
+			ibev.device = &ibdev->ib_dev;
+			ibev.element.port_num = port;
+			ibev.event = event == NETDEV_UP ? IB_EVENT_PORT_ACTIVE :
+							  IB_EVENT_PORT_ERR;
+			ib_dispatch_event(&ibev);
+		}
+
+	}
+	spin_unlock_bh(&iboe->lock);
+
+	if (update_qps_port > 0)
+		mlx4_ib_update_qps(ibdev, dev, update_qps_port);
+}
+
+static int mlx4_ib_netdev_event(struct notifier_block *this,
+				unsigned long event, void *ptr)
+{
+	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
+	struct mlx4_ib_dev *ibdev;
+
+	if (!net_eq(dev_net(dev), &init_net))
+		return NOTIFY_DONE;
+
+	ibdev = container_of(this, struct mlx4_ib_dev, iboe.nb);
+	mlx4_ib_scan_netdevs(ibdev, dev, event);
+
+	return NOTIFY_DONE;
+}
+
+static void init_pkeys(struct mlx4_ib_dev *ibdev)
+{
+	int port;
+	int slave;
+	int i;
+
+	if (mlx4_is_master(ibdev->dev)) {
+		for (slave = 0; slave <= ibdev->dev->persist->num_vfs;
+		     ++slave) {
+			for (port = 1; port <= ibdev->dev->caps.num_ports; ++port) {
+				for (i = 0;
+				     i < ibdev->dev->phys_caps.pkey_phys_table_len[port];
+				     ++i) {
+					ibdev->pkeys.virt2phys_pkey[slave][port - 1][i] =
+					/* master has the identity virt2phys pkey mapping */
+						(slave == mlx4_master_func_num(ibdev->dev) || !i) ? i :
+							ibdev->dev->phys_caps.pkey_phys_table_len[port] - 1;
+					mlx4_sync_pkey_table(ibdev->dev, slave, port, i,
+							     ibdev->pkeys.virt2phys_pkey[slave][port - 1][i]);
+				}
+			}
+		}
+		/* initialize pkey cache */
+		for (port = 1; port <= ibdev->dev->caps.num_ports; ++port) {
+			for (i = 0;
+			     i < ibdev->dev->phys_caps.pkey_phys_table_len[port];
+			     ++i)
+				ibdev->pkeys.phys_pkey_cache[port-1][i] =
+					(i) ? 0 : 0xFFFF;
+		}
+	}
+}
+
+static void mlx4_ib_alloc_eqs(struct mlx4_dev *dev, struct mlx4_ib_dev *ibdev)
+{
+	int i, j, eq = 0, total_eqs = 0;
+
+	ibdev->eq_table = kcalloc(dev->caps.num_comp_vectors,
+				  sizeof(ibdev->eq_table[0]), GFP_KERNEL);
+	if (!ibdev->eq_table)
+		return;
+
+	for (i = 1; i <= dev->caps.num_ports; i++) {
+		for (j = 0; j < mlx4_get_eqs_per_port(dev, i);
+		     j++, total_eqs++) {
+			if (i > 1 &&  mlx4_is_eq_shared(dev, total_eqs))
+				continue;
+			ibdev->eq_table[eq] = total_eqs;
+			if (!mlx4_assign_eq(dev, i,
+					    &ibdev->eq_table[eq]))
+				eq++;
+			else
+				ibdev->eq_table[eq] = -1;
+		}
+	}
+
+	for (i = eq; i < dev->caps.num_comp_vectors;
+	     ibdev->eq_table[i++] = -1)
+		;
+
+	/* Advertise the new number of EQs to clients */
+	ibdev->ib_dev.num_comp_vectors = eq;
+}
+
+static void mlx4_ib_free_eqs(struct mlx4_dev *dev, struct mlx4_ib_dev *ibdev)
+{
+	int i;
+	int total_eqs = ibdev->ib_dev.num_comp_vectors;
+
+	/* no eqs were allocated */
+	if (!ibdev->eq_table)
+		return;
+
+	/* Reset the advertised EQ number */
+	ibdev->ib_dev.num_comp_vectors = 0;
+
+	for (i = 0; i < total_eqs; i++)
+		mlx4_release_eq(dev, ibdev->eq_table[i]);
+
+	kfree(ibdev->eq_table);
+	ibdev->eq_table = NULL;
+}
+
+static int mlx4_port_immutable(struct ib_device *ibdev, u8 port_num,
+			       struct ib_port_immutable *immutable)
+{
+	struct ib_port_attr attr;
+	struct mlx4_ib_dev *mdev = to_mdev(ibdev);
+	int err;
+
+	if (mlx4_ib_port_link_layer(ibdev, port_num) == IB_LINK_LAYER_INFINIBAND) {
+		immutable->core_cap_flags = RDMA_CORE_PORT_IBA_IB;
+		immutable->max_mad_size = IB_MGMT_MAD_SIZE;
+	} else {
+		if (mdev->dev->caps.flags & MLX4_DEV_CAP_FLAG_IBOE)
+			immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE;
+		if (mdev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2)
+			immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE |
+				RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
+		immutable->core_cap_flags |= RDMA_CORE_PORT_RAW_PACKET;
+		if (immutable->core_cap_flags & (RDMA_CORE_PORT_IBA_ROCE |
+		    RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP))
+			immutable->max_mad_size = IB_MGMT_MAD_SIZE;
+	}
+
+	err = ib_query_port(ibdev, port_num, &attr);
+	if (err)
+		return err;
+
+	immutable->pkey_tbl_len = attr.pkey_tbl_len;
+	immutable->gid_tbl_len = attr.gid_tbl_len;
+
+	return 0;
+}
+
+static void get_fw_ver_str(struct ib_device *device, char *str)
+{
+	struct mlx4_ib_dev *dev =
+		container_of(device, struct mlx4_ib_dev, ib_dev);
+	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%d",
+		 (int) (dev->dev->caps.fw_ver >> 32),
+		 (int) (dev->dev->caps.fw_ver >> 16) & 0xffff,
+		 (int) dev->dev->caps.fw_ver & 0xffff);
+}
+
+static const struct ib_device_ops mlx4_ib_dev_ops = {
+	.owner = THIS_MODULE,
+	.driver_id = RDMA_DRIVER_MLX4,
+	.uverbs_abi_ver = MLX4_IB_UVERBS_ABI_VERSION,
+
+	.add_gid = mlx4_ib_add_gid,
+	.alloc_mr = mlx4_ib_alloc_mr,
+	.alloc_pd = mlx4_ib_alloc_pd,
+	.alloc_ucontext = mlx4_ib_alloc_ucontext,
+	.attach_mcast = mlx4_ib_mcg_attach,
+	.create_ah = mlx4_ib_create_ah,
+	.create_cq = mlx4_ib_create_cq,
+	.create_qp = mlx4_ib_create_qp,
+	.create_srq = mlx4_ib_create_srq,
+	.dealloc_pd = mlx4_ib_dealloc_pd,
+	.dealloc_ucontext = mlx4_ib_dealloc_ucontext,
+	.del_gid = mlx4_ib_del_gid,
+	.dereg_mr = mlx4_ib_dereg_mr,
+	.destroy_ah = mlx4_ib_destroy_ah,
+	.destroy_cq = mlx4_ib_destroy_cq,
+	.destroy_qp = mlx4_ib_destroy_qp,
+	.destroy_srq = mlx4_ib_destroy_srq,
+	.detach_mcast = mlx4_ib_mcg_detach,
+	.disassociate_ucontext = mlx4_ib_disassociate_ucontext,
+	.drain_rq = mlx4_ib_drain_rq,
+	.drain_sq = mlx4_ib_drain_sq,
+	.get_dev_fw_str = get_fw_ver_str,
+	.get_dma_mr = mlx4_ib_get_dma_mr,
+	.get_link_layer = mlx4_ib_port_link_layer,
+	.get_netdev = mlx4_ib_get_netdev,
+	.get_port_immutable = mlx4_port_immutable,
+	.map_mr_sg = mlx4_ib_map_mr_sg,
+	.mmap = mlx4_ib_mmap,
+	.modify_cq = mlx4_ib_modify_cq,
+	.modify_device = mlx4_ib_modify_device,
+	.modify_port = mlx4_ib_modify_port,
+	.modify_qp = mlx4_ib_modify_qp,
+	.modify_srq = mlx4_ib_modify_srq,
+	.poll_cq = mlx4_ib_poll_cq,
+	.post_recv = mlx4_ib_post_recv,
+	.post_send = mlx4_ib_post_send,
+	.post_srq_recv = mlx4_ib_post_srq_recv,
+	.process_mad = mlx4_ib_process_mad,
+	.query_ah = mlx4_ib_query_ah,
+	.query_device = mlx4_ib_query_device,
+	.query_gid = mlx4_ib_query_gid,
+	.query_pkey = mlx4_ib_query_pkey,
+	.query_port = mlx4_ib_query_port,
+	.query_qp = mlx4_ib_query_qp,
+	.query_srq = mlx4_ib_query_srq,
+	.reg_user_mr = mlx4_ib_reg_user_mr,
+	.req_notify_cq = mlx4_ib_arm_cq,
+	.rereg_user_mr = mlx4_ib_rereg_user_mr,
+	.resize_cq = mlx4_ib_resize_cq,
+
+	INIT_RDMA_OBJ_SIZE(ib_ah, mlx4_ib_ah, ibah),
+	INIT_RDMA_OBJ_SIZE(ib_cq, mlx4_ib_cq, ibcq),
+	INIT_RDMA_OBJ_SIZE(ib_pd, mlx4_ib_pd, ibpd),
+	INIT_RDMA_OBJ_SIZE(ib_srq, mlx4_ib_srq, ibsrq),
+	INIT_RDMA_OBJ_SIZE(ib_ucontext, mlx4_ib_ucontext, ibucontext),
+};
+
+static const struct ib_device_ops mlx4_ib_dev_wq_ops = {
+	.create_rwq_ind_table = mlx4_ib_create_rwq_ind_table,
+	.create_wq = mlx4_ib_create_wq,
+	.destroy_rwq_ind_table = mlx4_ib_destroy_rwq_ind_table,
+	.destroy_wq = mlx4_ib_destroy_wq,
+	.modify_wq = mlx4_ib_modify_wq,
+
+	INIT_RDMA_OBJ_SIZE(ib_rwq_ind_table, mlx4_ib_rwq_ind_table,
+			   ib_rwq_ind_tbl),
+};
+
+static const struct ib_device_ops mlx4_ib_dev_mw_ops = {
+	.alloc_mw = mlx4_ib_alloc_mw,
+	.dealloc_mw = mlx4_ib_dealloc_mw,
+
+	INIT_RDMA_OBJ_SIZE(ib_mw, mlx4_ib_mw, ibmw),
+};
+
+static const struct ib_device_ops mlx4_ib_dev_xrc_ops = {
+	.alloc_xrcd = mlx4_ib_alloc_xrcd,
+	.dealloc_xrcd = mlx4_ib_dealloc_xrcd,
+
+	INIT_RDMA_OBJ_SIZE(ib_xrcd, mlx4_ib_xrcd, ibxrcd),
+};
+
+static const struct ib_device_ops mlx4_ib_dev_fs_ops = {
+	.create_flow = mlx4_ib_create_flow,
+	.destroy_flow = mlx4_ib_destroy_flow,
+};
+
+static void *mlx4_ib_add(struct mlx4_dev *dev)
+{
+	struct mlx4_ib_dev *ibdev;
+	int num_ports = 0;
+	int i, j;
+	int err;
+	struct mlx4_ib_iboe *iboe;
+	int ib_num_ports = 0;
+	int num_req_counters;
+	int allocated;
+	u32 counter_index;
+	struct counter_index *new_counter_index = NULL;
+
+	pr_info_once("%s", mlx4_ib_version);
+
+	num_ports = 0;
+	mlx4_foreach_ib_transport_port(i, dev)
+		num_ports++;
+
+	/* No point in registering a device with no ports... */
+	if (num_ports == 0)
+		return NULL;
+
+	ibdev = ib_alloc_device(mlx4_ib_dev, ib_dev);
+	if (!ibdev) {
+		dev_err(&dev->persist->pdev->dev,
+			"Device struct alloc failed\n");
+		return NULL;
+	}
+
+	iboe = &ibdev->iboe;
+
+	if (mlx4_pd_alloc(dev, &ibdev->priv_pdn))
+		goto err_dealloc;
+
+	if (mlx4_uar_alloc(dev, &ibdev->priv_uar))
+		goto err_pd;
+
+	ibdev->uar_map = ioremap((phys_addr_t) ibdev->priv_uar.pfn << PAGE_SHIFT,
+				 PAGE_SIZE);
+	if (!ibdev->uar_map)
+		goto err_uar;
+	MLX4_INIT_DOORBELL_LOCK(&ibdev->uar_lock);
+
+	ibdev->dev = dev;
+	ibdev->bond_next_port	= 0;
+
+	ibdev->ib_dev.node_type		= RDMA_NODE_IB_CA;
+	ibdev->ib_dev.local_dma_lkey	= dev->caps.reserved_lkey;
+	ibdev->num_ports		= num_ports;
+	ibdev->ib_dev.phys_port_cnt     = mlx4_is_bonded(dev) ?
+						1 : ibdev->num_ports;
+	ibdev->ib_dev.num_comp_vectors	= dev->caps.num_comp_vectors;
+	ibdev->ib_dev.dev.parent	= &dev->persist->pdev->dev;
+
+	ibdev->ib_dev.uverbs_cmd_mask	=
+		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
+		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
+		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
+		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
+		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
+		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
+		(1ull << IB_USER_VERBS_CMD_REREG_MR)		|
+		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
+		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
+		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
+		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
+		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
+		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
+		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
+		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
+		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
+		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
+		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
+		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
+		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
+		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
+		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
+		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
+		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
+
+	ib_set_device_ops(&ibdev->ib_dev, &mlx4_ib_dev_ops);
+	ibdev->ib_dev.uverbs_ex_cmd_mask |=
+		(1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ) |
+		(1ull << IB_USER_VERBS_EX_CMD_CREATE_QP);
+
+	if ((dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RSS) &&
+	    ((mlx4_ib_port_link_layer(&ibdev->ib_dev, 1) ==
+	    IB_LINK_LAYER_ETHERNET) ||
+	    (mlx4_ib_port_link_layer(&ibdev->ib_dev, 2) ==
+	    IB_LINK_LAYER_ETHERNET)))
+		ib_set_device_ops(&ibdev->ib_dev, &mlx4_ib_dev_wq_ops);
+
+	if (dev->caps.flags & MLX4_DEV_CAP_FLAG_MEM_WINDOW ||
+	    dev->caps.bmme_flags & MLX4_BMME_FLAG_TYPE_2_WIN) {
+		ibdev->ib_dev.uverbs_cmd_mask |=
+			(1ull << IB_USER_VERBS_CMD_ALLOC_MW) |
+			(1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
+		ib_set_device_ops(&ibdev->ib_dev, &mlx4_ib_dev_mw_ops);
+	}
+
+	if (dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC) {
+		ibdev->ib_dev.uverbs_cmd_mask |=
+			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
+			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
+		ib_set_device_ops(&ibdev->ib_dev, &mlx4_ib_dev_xrc_ops);
+	}
+
+	if (check_flow_steering_support(dev)) {
+		ibdev->steering_support = MLX4_STEERING_MODE_DEVICE_MANAGED;
+		ib_set_device_ops(&ibdev->ib_dev, &mlx4_ib_dev_fs_ops);
+	}
+
+	if (!dev->caps.userspace_caps)
+		ibdev->ib_dev.ops.uverbs_abi_ver =
+			MLX4_IB_UVERBS_NO_DEV_CAPS_ABI_VERSION;
+
+	mlx4_ib_alloc_eqs(dev, ibdev);
+
+	spin_lock_init(&iboe->lock);
+
+	if (init_node_data(ibdev))
+		goto err_map;
+	mlx4_init_sl2vl_tbl(ibdev);
+
+	for (i = 0; i < ibdev->num_ports; ++i) {
+		mutex_init(&ibdev->counters_table[i].mutex);
+		INIT_LIST_HEAD(&ibdev->counters_table[i].counters_list);
+		iboe->last_port_state[i] = IB_PORT_DOWN;
+	}
+
+	num_req_counters = mlx4_is_bonded(dev) ? 1 : ibdev->num_ports;
+	for (i = 0; i < num_req_counters; ++i) {
+		mutex_init(&ibdev->qp1_proxy_lock[i]);
+		allocated = 0;
+		if (mlx4_ib_port_link_layer(&ibdev->ib_dev, i + 1) ==
+						IB_LINK_LAYER_ETHERNET) {
+			err = mlx4_counter_alloc(ibdev->dev, &counter_index,
+						 MLX4_RES_USAGE_DRIVER);
+			/* if failed to allocate a new counter, use default */
+			if (err)
+				counter_index =
+					mlx4_get_default_counter_index(dev,
+								       i + 1);
+			else
+				allocated = 1;
+		} else { /* IB_LINK_LAYER_INFINIBAND use the default counter */
+			counter_index = mlx4_get_default_counter_index(dev,
+								       i + 1);
+		}
+		new_counter_index = kmalloc(sizeof(*new_counter_index),
+					    GFP_KERNEL);
+		if (!new_counter_index) {
+			if (allocated)
+				mlx4_counter_free(ibdev->dev, counter_index);
+			goto err_counter;
+		}
+		new_counter_index->index = counter_index;
+		new_counter_index->allocated = allocated;
+		list_add_tail(&new_counter_index->list,
+			      &ibdev->counters_table[i].counters_list);
+		ibdev->counters_table[i].default_counter = counter_index;
+		pr_info("counter index %d for port %d allocated %d\n",
+			counter_index, i + 1, allocated);
+	}
+	if (mlx4_is_bonded(dev))
+		for (i = 1; i < ibdev->num_ports ; ++i) {
+			new_counter_index =
+					kmalloc(sizeof(struct counter_index),
+						GFP_KERNEL);
+			if (!new_counter_index)
+				goto err_counter;
+			new_counter_index->index = counter_index;
+			new_counter_index->allocated = 0;
+			list_add_tail(&new_counter_index->list,
+				      &ibdev->counters_table[i].counters_list);
+			ibdev->counters_table[i].default_counter =
+								counter_index;
+		}
+
+	mlx4_foreach_port(i, dev, MLX4_PORT_TYPE_IB)
+		ib_num_ports++;
+
+	spin_lock_init(&ibdev->sm_lock);
+	mutex_init(&ibdev->cap_mask_mutex);
+	INIT_LIST_HEAD(&ibdev->qp_list);
+	spin_lock_init(&ibdev->reset_flow_resource_lock);
+
+	if (ibdev->steering_support == MLX4_STEERING_MODE_DEVICE_MANAGED &&
+	    ib_num_ports) {
+		ibdev->steer_qpn_count = MLX4_IB_UC_MAX_NUM_QPS;
+		err = mlx4_qp_reserve_range(dev, ibdev->steer_qpn_count,
+					    MLX4_IB_UC_STEER_QPN_ALIGN,
+					    &ibdev->steer_qpn_base, 0,
+					    MLX4_RES_USAGE_DRIVER);
+		if (err)
+			goto err_counter;
+
+		ibdev->ib_uc_qpns_bitmap =
+			kmalloc_array(BITS_TO_LONGS(ibdev->steer_qpn_count),
+				      sizeof(long),
+				      GFP_KERNEL);
+		if (!ibdev->ib_uc_qpns_bitmap)
+			goto err_steer_qp_release;
+
+		if (dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_DMFS_IPOIB) {
+			bitmap_zero(ibdev->ib_uc_qpns_bitmap,
+				    ibdev->steer_qpn_count);
+			err = mlx4_FLOW_STEERING_IB_UC_QP_RANGE(
+					dev, ibdev->steer_qpn_base,
+					ibdev->steer_qpn_base +
+					ibdev->steer_qpn_count - 1);
+			if (err)
+				goto err_steer_free_bitmap;
+		} else {
+			bitmap_fill(ibdev->ib_uc_qpns_bitmap,
+				    ibdev->steer_qpn_count);
+		}
+	}
+
+	for (j = 1; j <= ibdev->dev->caps.num_ports; j++)
+		atomic64_set(&iboe->mac[j - 1], ibdev->dev->caps.def_mac[j]);
+
+	if (mlx4_ib_alloc_diag_counters(ibdev))
+		goto err_steer_free_bitmap;
+
+	rdma_set_device_sysfs_group(&ibdev->ib_dev, &mlx4_attr_group);
+	if (ib_register_device(&ibdev->ib_dev, "mlx4_%d",
+			       &dev->persist->pdev->dev))
+		goto err_diag_counters;
+
+	if (mlx4_ib_mad_init(ibdev))
+		goto err_reg;
+
+	if (mlx4_ib_init_sriov(ibdev))
+		goto err_mad;
+
+	if (!iboe->nb.notifier_call) {
+		iboe->nb.notifier_call = mlx4_ib_netdev_event;
+		err = register_netdevice_notifier(&iboe->nb);
+		if (err) {
+			iboe->nb.notifier_call = NULL;
+			goto err_notif;
+		}
+	}
+	if (dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
+		err = mlx4_config_roce_v2_port(dev, ROCE_V2_UDP_DPORT);
+		if (err)
+			goto err_notif;
+	}
+
+	ibdev->ib_active = true;
+	mlx4_foreach_port(i, dev, MLX4_PORT_TYPE_IB)
+		devlink_port_type_ib_set(mlx4_get_devlink_port(dev, i),
+					 &ibdev->ib_dev);
+
+	if (mlx4_is_mfunc(ibdev->dev))
+		init_pkeys(ibdev);
+
+	/* create paravirt contexts for any VFs which are active */
+	if (mlx4_is_master(ibdev->dev)) {
+		for (j = 0; j < MLX4_MFUNC_MAX; j++) {
+			if (j == mlx4_master_func_num(ibdev->dev))
+				continue;
+			if (mlx4_is_slave_active(ibdev->dev, j))
+				do_slave_init(ibdev, j, 1);
+		}
+	}
+	return ibdev;
+
+err_notif:
+	if (ibdev->iboe.nb.notifier_call) {
+		if (unregister_netdevice_notifier(&ibdev->iboe.nb))
+			pr_warn("failure unregistering notifier\n");
+		ibdev->iboe.nb.notifier_call = NULL;
+	}
+	flush_workqueue(wq);
+
+	mlx4_ib_close_sriov(ibdev);
+
+err_mad:
+	mlx4_ib_mad_cleanup(ibdev);
+
+err_reg:
+	ib_unregister_device(&ibdev->ib_dev);
+
+err_diag_counters:
+	mlx4_ib_diag_cleanup(ibdev);
+
+err_steer_free_bitmap:
+	kfree(ibdev->ib_uc_qpns_bitmap);
+
+err_steer_qp_release:
+	mlx4_qp_release_range(dev, ibdev->steer_qpn_base,
+			      ibdev->steer_qpn_count);
+err_counter:
+	for (i = 0; i < ibdev->num_ports; ++i)
+		mlx4_ib_delete_counters_table(ibdev, &ibdev->counters_table[i]);
+
+err_map:
+	mlx4_ib_free_eqs(dev, ibdev);
+	iounmap(ibdev->uar_map);
+
+err_uar:
+	mlx4_uar_free(dev, &ibdev->priv_uar);
+
+err_pd:
+	mlx4_pd_free(dev, ibdev->priv_pdn);
+
+err_dealloc:
+	ib_dealloc_device(&ibdev->ib_dev);
+
+	return NULL;
+}
+
+int mlx4_ib_steer_qp_alloc(struct mlx4_ib_dev *dev, int count, int *qpn)
+{
+	int offset;
+
+	WARN_ON(!dev->ib_uc_qpns_bitmap);
+
+	offset = bitmap_find_free_region(dev->ib_uc_qpns_bitmap,
+					 dev->steer_qpn_count,
+					 get_count_order(count));
+	if (offset < 0)
+		return offset;
+
+	*qpn = dev->steer_qpn_base + offset;
+	return 0;
+}
+
+void mlx4_ib_steer_qp_free(struct mlx4_ib_dev *dev, u32 qpn, int count)
+{
+	if (!qpn ||
+	    dev->steering_support != MLX4_STEERING_MODE_DEVICE_MANAGED)
+		return;
+
+	if (WARN(qpn < dev->steer_qpn_base, "qpn = %u, steer_qpn_base = %u\n",
+		 qpn, dev->steer_qpn_base))
+		/* not supposed to be here */
+		return;
+
+	bitmap_release_region(dev->ib_uc_qpns_bitmap,
+			      qpn - dev->steer_qpn_base,
+			      get_count_order(count));
+}
+
+int mlx4_ib_steer_qp_reg(struct mlx4_ib_dev *mdev, struct mlx4_ib_qp *mqp,
+			 int is_attach)
+{
+	int err;
+	size_t flow_size;
+	struct ib_flow_attr *flow = NULL;
+	struct ib_flow_spec_ib *ib_spec;
+
+	if (is_attach) {
+		flow_size = sizeof(struct ib_flow_attr) +
+			    sizeof(struct ib_flow_spec_ib);
+		flow = kzalloc(flow_size, GFP_KERNEL);
+		if (!flow)
+			return -ENOMEM;
+		flow->port = mqp->port;
+		flow->num_of_specs = 1;
+		flow->size = flow_size;
+		ib_spec = (struct ib_flow_spec_ib *)(flow + 1);
+		ib_spec->type = IB_FLOW_SPEC_IB;
+		ib_spec->size = sizeof(struct ib_flow_spec_ib);
+		/* Add an empty rule for IB L2 */
+		memset(&ib_spec->mask, 0, sizeof(ib_spec->mask));
+
+		err = __mlx4_ib_create_flow(&mqp->ibqp, flow, MLX4_DOMAIN_NIC,
+					    MLX4_FS_REGULAR, &mqp->reg_id);
+	} else {
+		err = __mlx4_ib_destroy_flow(mdev->dev, mqp->reg_id);
+	}
+	kfree(flow);
+	return err;
+}
+
+static void mlx4_ib_remove(struct mlx4_dev *dev, void *ibdev_ptr)
+{
+	struct mlx4_ib_dev *ibdev = ibdev_ptr;
+	int p;
+	int i;
+
+	mlx4_foreach_port(i, dev, MLX4_PORT_TYPE_IB)
+		devlink_port_type_clear(mlx4_get_devlink_port(dev, i));
+	ibdev->ib_active = false;
+	flush_workqueue(wq);
+
+	if (ibdev->iboe.nb.notifier_call) {
+		if (unregister_netdevice_notifier(&ibdev->iboe.nb))
+			pr_warn("failure unregistering notifier\n");
+		ibdev->iboe.nb.notifier_call = NULL;
+	}
+
+	mlx4_ib_close_sriov(ibdev);
+	mlx4_ib_mad_cleanup(ibdev);
+	ib_unregister_device(&ibdev->ib_dev);
+	mlx4_ib_diag_cleanup(ibdev);
+
+	mlx4_qp_release_range(dev, ibdev->steer_qpn_base,
+			      ibdev->steer_qpn_count);
+	kfree(ibdev->ib_uc_qpns_bitmap);
+
+	iounmap(ibdev->uar_map);
+	for (p = 0; p < ibdev->num_ports; ++p)
+		mlx4_ib_delete_counters_table(ibdev, &ibdev->counters_table[p]);
+
+	mlx4_foreach_port(p, dev, MLX4_PORT_TYPE_IB)
+		mlx4_CLOSE_PORT(dev, p);
+
+	mlx4_ib_free_eqs(dev, ibdev);
+
+	mlx4_uar_free(dev, &ibdev->priv_uar);
+	mlx4_pd_free(dev, ibdev->priv_pdn);
+	ib_dealloc_device(&ibdev->ib_dev);
+}
+
+static void do_slave_init(struct mlx4_ib_dev *ibdev, int slave, int do_init)
+{
+	struct mlx4_ib_demux_work **dm = NULL;
+	struct mlx4_dev *dev = ibdev->dev;
+	int i;
+	unsigned long flags;
+	struct mlx4_active_ports actv_ports;
+	unsigned int ports;
+	unsigned int first_port;
+
+	if (!mlx4_is_master(dev))
+		return;
+
+	actv_ports = mlx4_get_active_ports(dev, slave);
+	ports = bitmap_weight(actv_ports.ports, dev->caps.num_ports);
+	first_port = find_first_bit(actv_ports.ports, dev->caps.num_ports);
+
+	dm = kcalloc(ports, sizeof(*dm), GFP_ATOMIC);
+	if (!dm)
+		return;
+
+	for (i = 0; i < ports; i++) {
+		dm[i] = kmalloc(sizeof (struct mlx4_ib_demux_work), GFP_ATOMIC);
+		if (!dm[i]) {
+			while (--i >= 0)
+				kfree(dm[i]);
+			goto out;
+		}
+		INIT_WORK(&dm[i]->work, mlx4_ib_tunnels_update_work);
+		dm[i]->port = first_port + i + 1;
+		dm[i]->slave = slave;
+		dm[i]->do_init = do_init;
+		dm[i]->dev = ibdev;
+	}
+	/* initialize or tear down tunnel QPs for the slave */
+	spin_lock_irqsave(&ibdev->sriov.going_down_lock, flags);
+	if (!ibdev->sriov.is_going_down) {
+		for (i = 0; i < ports; i++)
+			queue_work(ibdev->sriov.demux[i].ud_wq, &dm[i]->work);
+		spin_unlock_irqrestore(&ibdev->sriov.going_down_lock, flags);
+	} else {
+		spin_unlock_irqrestore(&ibdev->sriov.going_down_lock, flags);
+		for (i = 0; i < ports; i++)
+			kfree(dm[i]);
+	}
+out:
+	kfree(dm);
+	return;
+}
+
+static void mlx4_ib_handle_catas_error(struct mlx4_ib_dev *ibdev)
+{
+	struct mlx4_ib_qp *mqp;
+	unsigned long flags_qp;
+	unsigned long flags_cq;
+	struct mlx4_ib_cq *send_mcq, *recv_mcq;
+	struct list_head    cq_notify_list;
+	struct mlx4_cq *mcq;
+	unsigned long flags;
+
+	pr_warn("mlx4_ib_handle_catas_error was started\n");
+	INIT_LIST_HEAD(&cq_notify_list);
+
+	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
+	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
+
+	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
+		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
+		if (mqp->sq.tail != mqp->sq.head) {
+			send_mcq = to_mcq(mqp->ibqp.send_cq);
+			spin_lock_irqsave(&send_mcq->lock, flags_cq);
+			if (send_mcq->mcq.comp &&
+			    mqp->ibqp.send_cq->comp_handler) {
+				if (!send_mcq->mcq.reset_notify_added) {
+					send_mcq->mcq.reset_notify_added = 1;
+					list_add_tail(&send_mcq->mcq.reset_notify,
+						      &cq_notify_list);
+				}
+			}
+			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
+		}
+		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
+		/* Now, handle the QP's receive queue */
+		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
+		/* no handling is needed for SRQ */
+		if (!mqp->ibqp.srq) {
+			if (mqp->rq.tail != mqp->rq.head) {
+				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
+				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
+				if (recv_mcq->mcq.comp &&
+				    mqp->ibqp.recv_cq->comp_handler) {
+					if (!recv_mcq->mcq.reset_notify_added) {
+						recv_mcq->mcq.reset_notify_added = 1;
+						list_add_tail(&recv_mcq->mcq.reset_notify,
+							      &cq_notify_list);
+					}
+				}
+				spin_unlock_irqrestore(&recv_mcq->lock,
+						       flags_cq);
+			}
+		}
+		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
+	}
+
+	list_for_each_entry(mcq, &cq_notify_list, reset_notify) {
+		mcq->comp(mcq);
+	}
+	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
+	pr_warn("mlx4_ib_handle_catas_error ended\n");
+}
+
+static void handle_bonded_port_state_event(struct work_struct *work)
+{
+	struct ib_event_work *ew =
+		container_of(work, struct ib_event_work, work);
+	struct mlx4_ib_dev *ibdev = ew->ib_dev;
+	enum ib_port_state bonded_port_state = IB_PORT_NOP;
+	int i;
+	struct ib_event ibev;
+
+	kfree(ew);
+	spin_lock_bh(&ibdev->iboe.lock);
+	for (i = 0; i < MLX4_MAX_PORTS; ++i) {
+		struct net_device *curr_netdev = ibdev->iboe.netdevs[i];
+		enum ib_port_state curr_port_state;
+
+		if (!curr_netdev)
+			continue;
+
+		curr_port_state =
+			(netif_running(curr_netdev) &&
+			 netif_carrier_ok(curr_netdev)) ?
+			IB_PORT_ACTIVE : IB_PORT_DOWN;
+
+		bonded_port_state = (bonded_port_state != IB_PORT_ACTIVE) ?
+			curr_port_state : IB_PORT_ACTIVE;
+	}
+	spin_unlock_bh(&ibdev->iboe.lock);
+
+	ibev.device = &ibdev->ib_dev;
+	ibev.element.port_num = 1;
+	ibev.event = (bonded_port_state == IB_PORT_ACTIVE) ?
+		IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
+
+	ib_dispatch_event(&ibev);
+}
+
+void mlx4_ib_sl2vl_update(struct mlx4_ib_dev *mdev, int port)
+{
+	u64 sl2vl;
+	int err;
+
+	err = mlx4_ib_query_sl2vl(&mdev->ib_dev, port, &sl2vl);
+	if (err) {
+		pr_err("Unable to get current sl to vl mapping for port %d.  Using all zeroes (%d)\n",
+		       port, err);
+		sl2vl = 0;
+	}
+	atomic64_set(&mdev->sl2vl[port - 1], sl2vl);
+}
+
+static void ib_sl2vl_update_work(struct work_struct *work)
+{
+	struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
+	struct mlx4_ib_dev *mdev = ew->ib_dev;
+	int port = ew->port;
+
+	mlx4_ib_sl2vl_update(mdev, port);
+
+	kfree(ew);
+}
+
+void mlx4_sched_ib_sl2vl_update_work(struct mlx4_ib_dev *ibdev,
+				     int port)
+{
+	struct ib_event_work *ew;
+
+	ew = kmalloc(sizeof(*ew), GFP_ATOMIC);
+	if (ew) {
+		INIT_WORK(&ew->work, ib_sl2vl_update_work);
+		ew->port = port;
+		ew->ib_dev = ibdev;
+		queue_work(wq, &ew->work);
+	}
+}
+
+static void mlx4_ib_event(struct mlx4_dev *dev, void *ibdev_ptr,
+			  enum mlx4_dev_event event, unsigned long param)
+{
+	struct ib_event ibev;
+	struct mlx4_ib_dev *ibdev = to_mdev((struct ib_device *) ibdev_ptr);
+	struct mlx4_eqe *eqe = NULL;
+	struct ib_event_work *ew;
+	int p = 0;
+
+	if (mlx4_is_bonded(dev) &&
+	    ((event == MLX4_DEV_EVENT_PORT_UP) ||
+	    (event == MLX4_DEV_EVENT_PORT_DOWN))) {
+		ew = kmalloc(sizeof(*ew), GFP_ATOMIC);
+		if (!ew)
+			return;
+		INIT_WORK(&ew->work, handle_bonded_port_state_event);
+		ew->ib_dev = ibdev;
+		queue_work(wq, &ew->work);
+		return;
+	}
+
+	if (event == MLX4_DEV_EVENT_PORT_MGMT_CHANGE)
+		eqe = (struct mlx4_eqe *)param;
+	else
+		p = (int) param;
+
+	switch (event) {
+	case MLX4_DEV_EVENT_PORT_UP:
+		if (p > ibdev->num_ports)
+			return;
+		if (!mlx4_is_slave(dev) &&
+		    rdma_port_get_link_layer(&ibdev->ib_dev, p) ==
+			IB_LINK_LAYER_INFINIBAND) {
+			if (mlx4_is_master(dev))
+				mlx4_ib_invalidate_all_guid_record(ibdev, p);
+			if (ibdev->dev->flags & MLX4_FLAG_SECURE_HOST &&
+			    !(ibdev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_SL_TO_VL_CHANGE_EVENT))
+				mlx4_sched_ib_sl2vl_update_work(ibdev, p);
+		}
+		ibev.event = IB_EVENT_PORT_ACTIVE;
+		break;
+
+	case MLX4_DEV_EVENT_PORT_DOWN:
+		if (p > ibdev->num_ports)
+			return;
+		ibev.event = IB_EVENT_PORT_ERR;
+		break;
+
+	case MLX4_DEV_EVENT_CATASTROPHIC_ERROR:
+		ibdev->ib_active = false;
+		ibev.event = IB_EVENT_DEVICE_FATAL;
+		mlx4_ib_handle_catas_error(ibdev);
+		break;
+
+	case MLX4_DEV_EVENT_PORT_MGMT_CHANGE:
+		ew = kmalloc(sizeof *ew, GFP_ATOMIC);
+		if (!ew)
+			return;
+
+		INIT_WORK(&ew->work, handle_port_mgmt_change_event);
+		memcpy(&ew->ib_eqe, eqe, sizeof *eqe);
+		ew->ib_dev = ibdev;
+		/* need to queue only for port owner, which uses GEN_EQE */
+		if (mlx4_is_master(dev))
+			queue_work(wq, &ew->work);
+		else
+			handle_port_mgmt_change_event(&ew->work);
+		return;
+
+	case MLX4_DEV_EVENT_SLAVE_INIT:
+		/* here, p is the slave id */
+		do_slave_init(ibdev, p, 1);
+		if (mlx4_is_master(dev)) {
+			int i;
+
+			for (i = 1; i <= ibdev->num_ports; i++) {
+				if (rdma_port_get_link_layer(&ibdev->ib_dev, i)
+					== IB_LINK_LAYER_INFINIBAND)
+					mlx4_ib_slave_alias_guid_event(ibdev,
+								       p, i,
+								       1);
+			}
+		}
+		return;
+
+	case MLX4_DEV_EVENT_SLAVE_SHUTDOWN:
+		if (mlx4_is_master(dev)) {
+			int i;
+
+			for (i = 1; i <= ibdev->num_ports; i++) {
+				if (rdma_port_get_link_layer(&ibdev->ib_dev, i)
+					== IB_LINK_LAYER_INFINIBAND)
+					mlx4_ib_slave_alias_guid_event(ibdev,
+								       p, i,
+								       0);
+			}
+		}
+		/* here, p is the slave id */
+		do_slave_init(ibdev, p, 0);
+		return;
+
+	default:
+		return;
+	}
+
+	ibev.device	      = ibdev_ptr;
+	ibev.element.port_num = mlx4_is_bonded(ibdev->dev) ? 1 : (u8)p;
+
+	ib_dispatch_event(&ibev);
+}
+
+static struct mlx4_interface mlx4_ib_interface = {
+	.add		= mlx4_ib_add,
+	.remove		= mlx4_ib_remove,
+	.event		= mlx4_ib_event,
+	.protocol	= MLX4_PROT_IB_IPV6,
+	.flags		= MLX4_INTFF_BONDING
+};
+
+static int __init mlx4_ib_init(void)
+{
+	int err;
+
+	wq = alloc_ordered_workqueue("mlx4_ib", WQ_MEM_RECLAIM);
+	if (!wq)
+		return -ENOMEM;
+
+	err = mlx4_ib_mcg_init();
+	if (err)
+		goto clean_wq;
+
+	err = mlx4_register_interface(&mlx4_ib_interface);
+	if (err)
+		goto clean_mcg;
+
+	return 0;
+
+clean_mcg:
+	mlx4_ib_mcg_destroy();
+
+clean_wq:
+	destroy_workqueue(wq);
+	return err;
+}
+
+static void __exit mlx4_ib_cleanup(void)
+{
+	mlx4_unregister_interface(&mlx4_ib_interface);
+	mlx4_ib_mcg_destroy();
+	destroy_workqueue(wq);
+}
+
+module_init(mlx4_ib_init);
+module_exit(mlx4_ib_cleanup);
diff --git a/drivers/infiniband/hw/mlx4/mcg.c b/drivers/infiniband/hw/mlx4/mcg.c
new file mode 100644
index 000000000..5e4ec9786
--- /dev/null
+++ b/drivers/infiniband/hw/mlx4/mcg.c
@@ -0,0 +1,1257 @@
+/*
+ * Copyright (c) 2012 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <rdma/ib_mad.h>
+#include <rdma/ib_smi.h>
+#include <rdma/ib_cache.h>
+#include <rdma/ib_sa.h>
+
+#include <linux/mlx4/cmd.h>
+#include <linux/rbtree.h>
+#include <linux/delay.h>
+
+#include "mlx4_ib.h"
+
+#define MAX_VFS		80
+#define MAX_PEND_REQS_PER_FUNC 4
+#define MAD_TIMEOUT_MS	2000
+
+#define mcg_warn(fmt, arg...)	pr_warn("MCG WARNING: " fmt, ##arg)
+#define mcg_error(fmt, arg...)	pr_err(fmt, ##arg)
+#define mcg_warn_group(group, format, arg...) \
+	pr_warn("%s-%d: %16s (port %d): WARNING: " format, __func__, __LINE__,\
+	(group)->name, group->demux->port, ## arg)
+
+#define mcg_debug_group(group, format, arg...) \
+	pr_debug("%s-%d: %16s (port %d): WARNING: " format, __func__, __LINE__,\
+		 (group)->name, (group)->demux->port, ## arg)
+
+#define mcg_error_group(group, format, arg...) \
+	pr_err("  %16s: " format, (group)->name, ## arg)
+
+
+static union ib_gid mgid0;
+
+static struct workqueue_struct *clean_wq;
+
+enum mcast_state {
+	MCAST_NOT_MEMBER = 0,
+	MCAST_MEMBER,
+};
+
+enum mcast_group_state {
+	MCAST_IDLE,
+	MCAST_JOIN_SENT,
+	MCAST_LEAVE_SENT,
+	MCAST_RESP_READY
+};
+
+struct mcast_member {
+	enum mcast_state state;
+	uint8_t			join_state;
+	int			num_pend_reqs;
+	struct list_head	pending;
+};
+
+struct ib_sa_mcmember_data {
+	union ib_gid	mgid;
+	union ib_gid	port_gid;
+	__be32		qkey;
+	__be16		mlid;
+	u8		mtusel_mtu;
+	u8		tclass;
+	__be16		pkey;
+	u8		ratesel_rate;
+	u8		lifetmsel_lifetm;
+	__be32		sl_flowlabel_hoplimit;
+	u8		scope_join_state;
+	u8		proxy_join;
+	u8		reserved[2];
+} __packed __aligned(4);
+
+struct mcast_group {
+	struct ib_sa_mcmember_data rec;
+	struct rb_node		node;
+	struct list_head	mgid0_list;
+	struct mlx4_ib_demux_ctx *demux;
+	struct mcast_member	func[MAX_VFS];
+	struct mutex		lock;
+	struct work_struct	work;
+	struct list_head	pending_list;
+	int			members[3];
+	enum mcast_group_state	state;
+	enum mcast_group_state	prev_state;
+	struct ib_sa_mad	response_sa_mad;
+	__be64			last_req_tid;
+
+	char			name[33]; /* MGID string */
+	struct device_attribute	dentry;
+
+	/* refcount is the reference count for the following:
+	   1. Each queued request
+	   2. Each invocation of the worker thread
+	   3. Membership of the port at the SA
+	*/
+	atomic_t		refcount;
+
+	/* delayed work to clean pending SM request */
+	struct delayed_work	timeout_work;
+	struct list_head	cleanup_list;
+};
+
+struct mcast_req {
+	int			func;
+	struct ib_sa_mad	sa_mad;
+	struct list_head	group_list;
+	struct list_head	func_list;
+	struct mcast_group	*group;
+	int			clean;
+};
+
+
+#define safe_atomic_dec(ref) \
+	do {\
+		if (atomic_dec_and_test(ref)) \
+			mcg_warn_group(group, "did not expect to reach zero\n"); \
+	} while (0)
+
+static const char *get_state_string(enum mcast_group_state state)
+{
+	switch (state) {
+	case MCAST_IDLE:
+		return "MCAST_IDLE";
+	case MCAST_JOIN_SENT:
+		return "MCAST_JOIN_SENT";
+	case MCAST_LEAVE_SENT:
+		return "MCAST_LEAVE_SENT";
+	case MCAST_RESP_READY:
+		return "MCAST_RESP_READY";
+	}
+	return "Invalid State";
+}
+
+static struct mcast_group *mcast_find(struct mlx4_ib_demux_ctx *ctx,
+				      union ib_gid *mgid)
+{
+	struct rb_node *node = ctx->mcg_table.rb_node;
+	struct mcast_group *group;
+	int ret;
+
+	while (node) {
+		group = rb_entry(node, struct mcast_group, node);
+		ret = memcmp(mgid->raw, group->rec.mgid.raw, sizeof *mgid);
+		if (!ret)
+			return group;
+
+		if (ret < 0)
+			node = node->rb_left;
+		else
+			node = node->rb_right;
+	}
+	return NULL;
+}
+
+static struct mcast_group *mcast_insert(struct mlx4_ib_demux_ctx *ctx,
+					struct mcast_group *group)
+{
+	struct rb_node **link = &ctx->mcg_table.rb_node;
+	struct rb_node *parent = NULL;
+	struct mcast_group *cur_group;
+	int ret;
+
+	while (*link) {
+		parent = *link;
+		cur_group = rb_entry(parent, struct mcast_group, node);
+
+		ret = memcmp(group->rec.mgid.raw, cur_group->rec.mgid.raw,
+			     sizeof group->rec.mgid);
+		if (ret < 0)
+			link = &(*link)->rb_left;
+		else if (ret > 0)
+			link = &(*link)->rb_right;
+		else
+			return cur_group;
+	}
+	rb_link_node(&group->node, parent, link);
+	rb_insert_color(&group->node, &ctx->mcg_table);
+	return NULL;
+}
+
+static int send_mad_to_wire(struct mlx4_ib_demux_ctx *ctx, struct ib_mad *mad)
+{
+	struct mlx4_ib_dev *dev = ctx->dev;
+	struct rdma_ah_attr	ah_attr;
+	unsigned long flags;
+
+	spin_lock_irqsave(&dev->sm_lock, flags);
+	if (!dev->sm_ah[ctx->port - 1]) {
+		/* port is not yet Active, sm_ah not ready */
+		spin_unlock_irqrestore(&dev->sm_lock, flags);
+		return -EAGAIN;
+	}
+	mlx4_ib_query_ah(dev->sm_ah[ctx->port - 1], &ah_attr);
+	spin_unlock_irqrestore(&dev->sm_lock, flags);
+	return mlx4_ib_send_to_wire(dev, mlx4_master_func_num(dev->dev),
+				    ctx->port, IB_QPT_GSI, 0, 1, IB_QP1_QKEY,
+				    &ah_attr, NULL, 0xffff, mad);
+}
+
+static int send_mad_to_slave(int slave, struct mlx4_ib_demux_ctx *ctx,
+			     struct ib_mad *mad)
+{
+	struct mlx4_ib_dev *dev = ctx->dev;
+	struct ib_mad_agent *agent = dev->send_agent[ctx->port - 1][1];
+	struct ib_wc wc;
+	struct rdma_ah_attr ah_attr;
+
+	/* Our agent might not yet be registered when mads start to arrive */
+	if (!agent)
+		return -EAGAIN;
+
+	rdma_query_ah(dev->sm_ah[ctx->port - 1], &ah_attr);
+
+	if (ib_find_cached_pkey(&dev->ib_dev, ctx->port, IB_DEFAULT_PKEY_FULL, &wc.pkey_index))
+		return -EINVAL;
+	wc.sl = 0;
+	wc.dlid_path_bits = 0;
+	wc.port_num = ctx->port;
+	wc.slid = rdma_ah_get_dlid(&ah_attr);  /* opensm lid */
+	wc.src_qp = 1;
+	return mlx4_ib_send_to_slave(dev, slave, ctx->port, IB_QPT_GSI, &wc, NULL, mad);
+}
+
+static int send_join_to_wire(struct mcast_group *group, struct ib_sa_mad *sa_mad)
+{
+	struct ib_sa_mad mad;
+	struct ib_sa_mcmember_data *sa_mad_data = (struct ib_sa_mcmember_data *)&mad.data;
+	int ret;
+
+	/* we rely on a mad request as arrived from a VF */
+	memcpy(&mad, sa_mad, sizeof mad);
+
+	/* fix port GID to be the real one (slave 0) */
+	sa_mad_data->port_gid.global.interface_id = group->demux->guid_cache[0];
+
+	/* assign our own TID */
+	mad.mad_hdr.tid = mlx4_ib_get_new_demux_tid(group->demux);
+	group->last_req_tid = mad.mad_hdr.tid; /* keep it for later validation */
+
+	ret = send_mad_to_wire(group->demux, (struct ib_mad *)&mad);
+	/* set timeout handler */
+	if (!ret) {
+		/* calls mlx4_ib_mcg_timeout_handler */
+		queue_delayed_work(group->demux->mcg_wq, &group->timeout_work,
+				msecs_to_jiffies(MAD_TIMEOUT_MS));
+	}
+
+	return ret;
+}
+
+static int send_leave_to_wire(struct mcast_group *group, u8 join_state)
+{
+	struct ib_sa_mad mad;
+	struct ib_sa_mcmember_data *sa_data = (struct ib_sa_mcmember_data *)&mad.data;
+	int ret;
+
+	memset(&mad, 0, sizeof mad);
+	mad.mad_hdr.base_version = 1;
+	mad.mad_hdr.mgmt_class = IB_MGMT_CLASS_SUBN_ADM;
+	mad.mad_hdr.class_version = 2;
+	mad.mad_hdr.method = IB_SA_METHOD_DELETE;
+	mad.mad_hdr.status = cpu_to_be16(0);
+	mad.mad_hdr.class_specific = cpu_to_be16(0);
+	mad.mad_hdr.tid = mlx4_ib_get_new_demux_tid(group->demux);
+	group->last_req_tid = mad.mad_hdr.tid; /* keep it for later validation */
+	mad.mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_MC_MEMBER_REC);
+	mad.mad_hdr.attr_mod = cpu_to_be32(0);
+	mad.sa_hdr.sm_key = 0x0;
+	mad.sa_hdr.attr_offset = cpu_to_be16(7);
+	mad.sa_hdr.comp_mask = IB_SA_MCMEMBER_REC_MGID |
+		IB_SA_MCMEMBER_REC_PORT_GID | IB_SA_MCMEMBER_REC_JOIN_STATE;
+
+	*sa_data = group->rec;
+	sa_data->scope_join_state = join_state;
+
+	ret = send_mad_to_wire(group->demux, (struct ib_mad *)&mad);
+	if (ret)
+		group->state = MCAST_IDLE;
+
+	/* set timeout handler */
+	if (!ret) {
+		/* calls mlx4_ib_mcg_timeout_handler */
+		queue_delayed_work(group->demux->mcg_wq, &group->timeout_work,
+				msecs_to_jiffies(MAD_TIMEOUT_MS));
+	}
+
+	return ret;
+}
+
+static int send_reply_to_slave(int slave, struct mcast_group *group,
+		struct ib_sa_mad *req_sa_mad, u16 status)
+{
+	struct ib_sa_mad mad;
+	struct ib_sa_mcmember_data *sa_data = (struct ib_sa_mcmember_data *)&mad.data;
+	struct ib_sa_mcmember_data *req_sa_data = (struct ib_sa_mcmember_data *)&req_sa_mad->data;
+	int ret;
+
+	memset(&mad, 0, sizeof mad);
+	mad.mad_hdr.base_version = 1;
+	mad.mad_hdr.mgmt_class = IB_MGMT_CLASS_SUBN_ADM;
+	mad.mad_hdr.class_version = 2;
+	mad.mad_hdr.method = IB_MGMT_METHOD_GET_RESP;
+	mad.mad_hdr.status = cpu_to_be16(status);
+	mad.mad_hdr.class_specific = cpu_to_be16(0);
+	mad.mad_hdr.tid = req_sa_mad->mad_hdr.tid;
+	*(u8 *)&mad.mad_hdr.tid = 0; /* resetting tid to 0 */
+	mad.mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_MC_MEMBER_REC);
+	mad.mad_hdr.attr_mod = cpu_to_be32(0);
+	mad.sa_hdr.sm_key = req_sa_mad->sa_hdr.sm_key;
+	mad.sa_hdr.attr_offset = cpu_to_be16(7);
+	mad.sa_hdr.comp_mask = 0; /* ignored on responses, see IBTA spec */
+
+	*sa_data = group->rec;
+
+	/* reconstruct VF's requested join_state and port_gid */
+	sa_data->scope_join_state &= 0xf0;
+	sa_data->scope_join_state |= (group->func[slave].join_state & 0x0f);
+	memcpy(&sa_data->port_gid, &req_sa_data->port_gid, sizeof req_sa_data->port_gid);
+
+	ret = send_mad_to_slave(slave, group->demux, (struct ib_mad *)&mad);
+	return ret;
+}
+
+static int check_selector(ib_sa_comp_mask comp_mask,
+			  ib_sa_comp_mask selector_mask,
+			  ib_sa_comp_mask value_mask,
+			  u8 src_value, u8 dst_value)
+{
+	int err;
+	u8 selector = dst_value >> 6;
+	dst_value &= 0x3f;
+	src_value &= 0x3f;
+
+	if (!(comp_mask & selector_mask) || !(comp_mask & value_mask))
+		return 0;
+
+	switch (selector) {
+	case IB_SA_GT:
+		err = (src_value <= dst_value);
+		break;
+	case IB_SA_LT:
+		err = (src_value >= dst_value);
+		break;
+	case IB_SA_EQ:
+		err = (src_value != dst_value);
+		break;
+	default:
+		err = 0;
+		break;
+	}
+
+	return err;
+}
+
+static u16 cmp_rec(struct ib_sa_mcmember_data *src,
+		   struct ib_sa_mcmember_data *dst, ib_sa_comp_mask comp_mask)
+{
+	/* src is group record, dst is request record */
+	/* MGID must already match */
+	/* Port_GID we always replace to our Port_GID, so it is a match */
+
+#define MAD_STATUS_REQ_INVALID 0x0200
+	if (comp_mask & IB_SA_MCMEMBER_REC_QKEY && src->qkey != dst->qkey)
+		return MAD_STATUS_REQ_INVALID;
+	if (comp_mask & IB_SA_MCMEMBER_REC_MLID && src->mlid != dst->mlid)
+		return MAD_STATUS_REQ_INVALID;
+	if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_MTU_SELECTOR,
+				 IB_SA_MCMEMBER_REC_MTU,
+				 src->mtusel_mtu, dst->mtusel_mtu))
+		return MAD_STATUS_REQ_INVALID;
+	if (comp_mask & IB_SA_MCMEMBER_REC_TRAFFIC_CLASS &&
+	    src->tclass != dst->tclass)
+		return MAD_STATUS_REQ_INVALID;
+	if (comp_mask & IB_SA_MCMEMBER_REC_PKEY && src->pkey != dst->pkey)
+		return MAD_STATUS_REQ_INVALID;
+	if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_RATE_SELECTOR,
+				 IB_SA_MCMEMBER_REC_RATE,
+				 src->ratesel_rate, dst->ratesel_rate))
+		return MAD_STATUS_REQ_INVALID;
+	if (check_selector(comp_mask,
+				 IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME_SELECTOR,
+				 IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME,
+				 src->lifetmsel_lifetm, dst->lifetmsel_lifetm))
+		return MAD_STATUS_REQ_INVALID;
+	if (comp_mask & IB_SA_MCMEMBER_REC_SL &&
+			(be32_to_cpu(src->sl_flowlabel_hoplimit) & 0xf0000000) !=
+			(be32_to_cpu(dst->sl_flowlabel_hoplimit) & 0xf0000000))
+		return MAD_STATUS_REQ_INVALID;
+	if (comp_mask & IB_SA_MCMEMBER_REC_FLOW_LABEL &&
+			(be32_to_cpu(src->sl_flowlabel_hoplimit) & 0x0fffff00) !=
+			(be32_to_cpu(dst->sl_flowlabel_hoplimit) & 0x0fffff00))
+		return MAD_STATUS_REQ_INVALID;
+	if (comp_mask & IB_SA_MCMEMBER_REC_HOP_LIMIT &&
+			(be32_to_cpu(src->sl_flowlabel_hoplimit) & 0x000000ff) !=
+			(be32_to_cpu(dst->sl_flowlabel_hoplimit) & 0x000000ff))
+		return MAD_STATUS_REQ_INVALID;
+	if (comp_mask & IB_SA_MCMEMBER_REC_SCOPE &&
+			(src->scope_join_state & 0xf0) !=
+			(dst->scope_join_state & 0xf0))
+		return MAD_STATUS_REQ_INVALID;
+
+	/* join_state checked separately, proxy_join ignored */
+
+	return 0;
+}
+
+/* release group, return 1 if this was last release and group is destroyed
+ * timout work is canceled sync */
+static int release_group(struct mcast_group *group, int from_timeout_handler)
+{
+	struct mlx4_ib_demux_ctx *ctx = group->demux;
+	int nzgroup;
+
+	mutex_lock(&ctx->mcg_table_lock);
+	mutex_lock(&group->lock);
+	if (atomic_dec_and_test(&group->refcount)) {
+		if (!from_timeout_handler) {
+			if (group->state != MCAST_IDLE &&
+			    !cancel_delayed_work(&group->timeout_work)) {
+				atomic_inc(&group->refcount);
+				mutex_unlock(&group->lock);
+				mutex_unlock(&ctx->mcg_table_lock);
+				return 0;
+			}
+		}
+
+		nzgroup = memcmp(&group->rec.mgid, &mgid0, sizeof mgid0);
+		if (nzgroup)
+			del_sysfs_port_mcg_attr(ctx->dev, ctx->port, &group->dentry.attr);
+		if (!list_empty(&group->pending_list))
+			mcg_warn_group(group, "releasing a group with non empty pending list\n");
+		if (nzgroup)
+			rb_erase(&group->node, &ctx->mcg_table);
+		list_del_init(&group->mgid0_list);
+		mutex_unlock(&group->lock);
+		mutex_unlock(&ctx->mcg_table_lock);
+		kfree(group);
+		return 1;
+	} else {
+		mutex_unlock(&group->lock);
+		mutex_unlock(&ctx->mcg_table_lock);
+	}
+	return 0;
+}
+
+static void adjust_membership(struct mcast_group *group, u8 join_state, int inc)
+{
+	int i;
+
+	for (i = 0; i < 3; i++, join_state >>= 1)
+		if (join_state & 0x1)
+			group->members[i] += inc;
+}
+
+static u8 get_leave_state(struct mcast_group *group)
+{
+	u8 leave_state = 0;
+	int i;
+
+	for (i = 0; i < 3; i++)
+		if (!group->members[i])
+			leave_state |= (1 << i);
+
+	return leave_state & (group->rec.scope_join_state & 0xf);
+}
+
+static int join_group(struct mcast_group *group, int slave, u8 join_mask)
+{
+	int ret = 0;
+	u8 join_state;
+
+	/* remove bits that slave is already member of, and adjust */
+	join_state = join_mask & (~group->func[slave].join_state);
+	adjust_membership(group, join_state, 1);
+	group->func[slave].join_state |= join_state;
+	if (group->func[slave].state != MCAST_MEMBER && join_state) {
+		group->func[slave].state = MCAST_MEMBER;
+		ret = 1;
+	}
+	return ret;
+}
+
+static int leave_group(struct mcast_group *group, int slave, u8 leave_state)
+{
+	int ret = 0;
+
+	adjust_membership(group, leave_state, -1);
+	group->func[slave].join_state &= ~leave_state;
+	if (!group->func[slave].join_state) {
+		group->func[slave].state = MCAST_NOT_MEMBER;
+		ret = 1;
+	}
+	return ret;
+}
+
+static int check_leave(struct mcast_group *group, int slave, u8 leave_mask)
+{
+	if (group->func[slave].state != MCAST_MEMBER)
+		return MAD_STATUS_REQ_INVALID;
+
+	/* make sure we're not deleting unset bits */
+	if (~group->func[slave].join_state & leave_mask)
+		return MAD_STATUS_REQ_INVALID;
+
+	if (!leave_mask)
+		return MAD_STATUS_REQ_INVALID;
+
+	return 0;
+}
+
+static void mlx4_ib_mcg_timeout_handler(struct work_struct *work)
+{
+	struct delayed_work *delay = to_delayed_work(work);
+	struct mcast_group *group;
+	struct mcast_req *req = NULL;
+
+	group = container_of(delay, typeof(*group), timeout_work);
+
+	mutex_lock(&group->lock);
+	if (group->state == MCAST_JOIN_SENT) {
+		if (!list_empty(&group->pending_list)) {
+			req = list_first_entry(&group->pending_list, struct mcast_req, group_list);
+			list_del(&req->group_list);
+			list_del(&req->func_list);
+			--group->func[req->func].num_pend_reqs;
+			mutex_unlock(&group->lock);
+			kfree(req);
+			if (memcmp(&group->rec.mgid, &mgid0, sizeof mgid0)) {
+				if (release_group(group, 1))
+					return;
+			} else {
+				kfree(group);
+				return;
+			}
+			mutex_lock(&group->lock);
+		} else
+			mcg_warn_group(group, "DRIVER BUG\n");
+	} else if (group->state == MCAST_LEAVE_SENT) {
+		if (group->rec.scope_join_state & 0xf)
+			group->rec.scope_join_state &= 0xf0;
+		group->state = MCAST_IDLE;
+		mutex_unlock(&group->lock);
+		if (release_group(group, 1))
+			return;
+		mutex_lock(&group->lock);
+	} else
+		mcg_warn_group(group, "invalid state %s\n", get_state_string(group->state));
+	group->state = MCAST_IDLE;
+	atomic_inc(&group->refcount);
+	if (!queue_work(group->demux->mcg_wq, &group->work))
+		safe_atomic_dec(&group->refcount);
+
+	mutex_unlock(&group->lock);
+}
+
+static int handle_leave_req(struct mcast_group *group, u8 leave_mask,
+			    struct mcast_req *req)
+{
+	u16 status;
+
+	if (req->clean)
+		leave_mask = group->func[req->func].join_state;
+
+	status = check_leave(group, req->func, leave_mask);
+	if (!status)
+		leave_group(group, req->func, leave_mask);
+
+	if (!req->clean)
+		send_reply_to_slave(req->func, group, &req->sa_mad, status);
+	--group->func[req->func].num_pend_reqs;
+	list_del(&req->group_list);
+	list_del(&req->func_list);
+	kfree(req);
+	return 1;
+}
+
+static int handle_join_req(struct mcast_group *group, u8 join_mask,
+			   struct mcast_req *req)
+{
+	u8 group_join_state = group->rec.scope_join_state & 0xf;
+	int ref = 0;
+	u16 status;
+	struct ib_sa_mcmember_data *sa_data = (struct ib_sa_mcmember_data *)req->sa_mad.data;
+
+	if (join_mask == (group_join_state & join_mask)) {
+		/* port's membership need not change */
+		status = cmp_rec(&group->rec, sa_data, req->sa_mad.sa_hdr.comp_mask);
+		if (!status)
+			join_group(group, req->func, join_mask);
+
+		--group->func[req->func].num_pend_reqs;
+		send_reply_to_slave(req->func, group, &req->sa_mad, status);
+		list_del(&req->group_list);
+		list_del(&req->func_list);
+		kfree(req);
+		++ref;
+	} else {
+		/* port's membership needs to be updated */
+		group->prev_state = group->state;
+		if (send_join_to_wire(group, &req->sa_mad)) {
+			--group->func[req->func].num_pend_reqs;
+			list_del(&req->group_list);
+			list_del(&req->func_list);
+			kfree(req);
+			ref = 1;
+			group->state = group->prev_state;
+		} else
+			group->state = MCAST_JOIN_SENT;
+	}
+
+	return ref;
+}
+
+static void mlx4_ib_mcg_work_handler(struct work_struct *work)
+{
+	struct mcast_group *group;
+	struct mcast_req *req = NULL;
+	struct ib_sa_mcmember_data *sa_data;
+	u8 req_join_state;
+	int rc = 1; /* release_count - this is for the scheduled work */
+	u16 status;
+	u8 method;
+
+	group = container_of(work, typeof(*group), work);
+
+	mutex_lock(&group->lock);
+
+	/* First, let's see if a response from SM is waiting regarding this group.
+	 * If so, we need to update the group's REC. If this is a bad response, we
+	 * may need to send a bad response to a VF waiting for it. If VF is waiting
+	 * and this is a good response, the VF will be answered later in this func. */
+	if (group->state == MCAST_RESP_READY) {
+		/* cancels mlx4_ib_mcg_timeout_handler */
+		cancel_delayed_work(&group->timeout_work);
+		status = be16_to_cpu(group->response_sa_mad.mad_hdr.status);
+		method = group->response_sa_mad.mad_hdr.method;
+		if (group->last_req_tid != group->response_sa_mad.mad_hdr.tid) {
+			mcg_warn_group(group, "Got MAD response to existing MGID but wrong TID, dropping. Resp TID=%llx, group TID=%llx\n",
+				be64_to_cpu(group->response_sa_mad.mad_hdr.tid),
+				be64_to_cpu(group->last_req_tid));
+			group->state = group->prev_state;
+			goto process_requests;
+		}
+		if (status) {
+			if (!list_empty(&group->pending_list))
+				req = list_first_entry(&group->pending_list,
+						struct mcast_req, group_list);
+			if (method == IB_MGMT_METHOD_GET_RESP) {
+					if (req) {
+						send_reply_to_slave(req->func, group, &req->sa_mad, status);
+						--group->func[req->func].num_pend_reqs;
+						list_del(&req->group_list);
+						list_del(&req->func_list);
+						kfree(req);
+						++rc;
+					} else
+						mcg_warn_group(group, "no request for failed join\n");
+			} else if (method == IB_SA_METHOD_DELETE_RESP && group->demux->flushing)
+				++rc;
+		} else {
+			u8 resp_join_state;
+			u8 cur_join_state;
+
+			resp_join_state = ((struct ib_sa_mcmember_data *)
+						group->response_sa_mad.data)->scope_join_state & 0xf;
+			cur_join_state = group->rec.scope_join_state & 0xf;
+
+			if (method == IB_MGMT_METHOD_GET_RESP) {
+				/* successfull join */
+				if (!cur_join_state && resp_join_state)
+					--rc;
+			} else if (!resp_join_state)
+					++rc;
+			memcpy(&group->rec, group->response_sa_mad.data, sizeof group->rec);
+		}
+		group->state = MCAST_IDLE;
+	}
+
+process_requests:
+	/* We should now go over pending join/leave requests, as long as we are idle. */
+	while (!list_empty(&group->pending_list) && group->state == MCAST_IDLE) {
+		req = list_first_entry(&group->pending_list, struct mcast_req,
+				       group_list);
+		sa_data = (struct ib_sa_mcmember_data *)req->sa_mad.data;
+		req_join_state = sa_data->scope_join_state & 0xf;
+
+		/* For a leave request, we will immediately answer the VF, and
+		 * update our internal counters. The actual leave will be sent
+		 * to SM later, if at all needed. We dequeue the request now. */
+		if (req->sa_mad.mad_hdr.method == IB_SA_METHOD_DELETE)
+			rc += handle_leave_req(group, req_join_state, req);
+		else
+			rc += handle_join_req(group, req_join_state, req);
+	}
+
+	/* Handle leaves */
+	if (group->state == MCAST_IDLE) {
+		req_join_state = get_leave_state(group);
+		if (req_join_state) {
+			group->rec.scope_join_state &= ~req_join_state;
+			group->prev_state = group->state;
+			if (send_leave_to_wire(group, req_join_state)) {
+				group->state = group->prev_state;
+				++rc;
+			} else
+				group->state = MCAST_LEAVE_SENT;
+		}
+	}
+
+	if (!list_empty(&group->pending_list) && group->state == MCAST_IDLE)
+		goto process_requests;
+	mutex_unlock(&group->lock);
+
+	while (rc--)
+		release_group(group, 0);
+}
+
+static struct mcast_group *search_relocate_mgid0_group(struct mlx4_ib_demux_ctx *ctx,
+						       __be64 tid,
+						       union ib_gid *new_mgid)
+{
+	struct mcast_group *group = NULL, *cur_group, *n;
+	struct mcast_req *req;
+
+	mutex_lock(&ctx->mcg_table_lock);
+	list_for_each_entry_safe(group, n, &ctx->mcg_mgid0_list, mgid0_list) {
+		mutex_lock(&group->lock);
+		if (group->last_req_tid == tid) {
+			if (memcmp(new_mgid, &mgid0, sizeof mgid0)) {
+				group->rec.mgid = *new_mgid;
+				sprintf(group->name, "%016llx%016llx",
+						be64_to_cpu(group->rec.mgid.global.subnet_prefix),
+						be64_to_cpu(group->rec.mgid.global.interface_id));
+				list_del_init(&group->mgid0_list);
+				cur_group = mcast_insert(ctx, group);
+				if (cur_group) {
+					/* A race between our code and SM. Silently cleaning the new one */
+					req = list_first_entry(&group->pending_list,
+							       struct mcast_req, group_list);
+					--group->func[req->func].num_pend_reqs;
+					list_del(&req->group_list);
+					list_del(&req->func_list);
+					kfree(req);
+					mutex_unlock(&group->lock);
+					mutex_unlock(&ctx->mcg_table_lock);
+					release_group(group, 0);
+					return NULL;
+				}
+
+				atomic_inc(&group->refcount);
+				add_sysfs_port_mcg_attr(ctx->dev, ctx->port, &group->dentry.attr);
+				mutex_unlock(&group->lock);
+				mutex_unlock(&ctx->mcg_table_lock);
+				return group;
+			} else {
+				struct mcast_req *tmp1, *tmp2;
+
+				list_del(&group->mgid0_list);
+				if (!list_empty(&group->pending_list) && group->state != MCAST_IDLE)
+					cancel_delayed_work_sync(&group->timeout_work);
+
+				list_for_each_entry_safe(tmp1, tmp2, &group->pending_list, group_list) {
+					list_del(&tmp1->group_list);
+					kfree(tmp1);
+				}
+				mutex_unlock(&group->lock);
+				mutex_unlock(&ctx->mcg_table_lock);
+				kfree(group);
+				return NULL;
+			}
+		}
+		mutex_unlock(&group->lock);
+	}
+	mutex_unlock(&ctx->mcg_table_lock);
+
+	return NULL;
+}
+
+static ssize_t sysfs_show_group(struct device *dev,
+		struct device_attribute *attr, char *buf);
+
+static struct mcast_group *acquire_group(struct mlx4_ib_demux_ctx *ctx,
+					 union ib_gid *mgid, int create)
+{
+	struct mcast_group *group, *cur_group;
+	int is_mgid0;
+	int i;
+
+	is_mgid0 = !memcmp(&mgid0, mgid, sizeof mgid0);
+	if (!is_mgid0) {
+		group = mcast_find(ctx, mgid);
+		if (group)
+			goto found;
+	}
+
+	if (!create)
+		return ERR_PTR(-ENOENT);
+
+	group = kzalloc(sizeof(*group), GFP_KERNEL);
+	if (!group)
+		return ERR_PTR(-ENOMEM);
+
+	group->demux = ctx;
+	group->rec.mgid = *mgid;
+	INIT_LIST_HEAD(&group->pending_list);
+	INIT_LIST_HEAD(&group->mgid0_list);
+	for (i = 0; i < MAX_VFS; ++i)
+		INIT_LIST_HEAD(&group->func[i].pending);
+	INIT_WORK(&group->work, mlx4_ib_mcg_work_handler);
+	INIT_DELAYED_WORK(&group->timeout_work, mlx4_ib_mcg_timeout_handler);
+	mutex_init(&group->lock);
+	sprintf(group->name, "%016llx%016llx",
+			be64_to_cpu(group->rec.mgid.global.subnet_prefix),
+			be64_to_cpu(group->rec.mgid.global.interface_id));
+	sysfs_attr_init(&group->dentry.attr);
+	group->dentry.show = sysfs_show_group;
+	group->dentry.store = NULL;
+	group->dentry.attr.name = group->name;
+	group->dentry.attr.mode = 0400;
+	group->state = MCAST_IDLE;
+
+	if (is_mgid0) {
+		list_add(&group->mgid0_list, &ctx->mcg_mgid0_list);
+		goto found;
+	}
+
+	cur_group = mcast_insert(ctx, group);
+	if (cur_group) {
+		mcg_warn("group just showed up %s - confused\n", cur_group->name);
+		kfree(group);
+		return ERR_PTR(-EINVAL);
+	}
+
+	add_sysfs_port_mcg_attr(ctx->dev, ctx->port, &group->dentry.attr);
+
+found:
+	atomic_inc(&group->refcount);
+	return group;
+}
+
+static void queue_req(struct mcast_req *req)
+{
+	struct mcast_group *group = req->group;
+
+	atomic_inc(&group->refcount); /* for the request */
+	atomic_inc(&group->refcount); /* for scheduling the work */
+	list_add_tail(&req->group_list, &group->pending_list);
+	list_add_tail(&req->func_list, &group->func[req->func].pending);
+	/* calls mlx4_ib_mcg_work_handler */
+	if (!queue_work(group->demux->mcg_wq, &group->work))
+		safe_atomic_dec(&group->refcount);
+}
+
+int mlx4_ib_mcg_demux_handler(struct ib_device *ibdev, int port, int slave,
+			      struct ib_sa_mad *mad)
+{
+	struct mlx4_ib_dev *dev = to_mdev(ibdev);
+	struct ib_sa_mcmember_data *rec = (struct ib_sa_mcmember_data *)mad->data;
+	struct mlx4_ib_demux_ctx *ctx = &dev->sriov.demux[port - 1];
+	struct mcast_group *group;
+
+	switch (mad->mad_hdr.method) {
+	case IB_MGMT_METHOD_GET_RESP:
+	case IB_SA_METHOD_DELETE_RESP:
+		mutex_lock(&ctx->mcg_table_lock);
+		group = acquire_group(ctx, &rec->mgid, 0);
+		mutex_unlock(&ctx->mcg_table_lock);
+		if (IS_ERR(group)) {
+			if (mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP) {
+				__be64 tid = mad->mad_hdr.tid;
+				*(u8 *)(&tid) = (u8)slave; /* in group we kept the modified TID */
+				group = search_relocate_mgid0_group(ctx, tid, &rec->mgid);
+			} else
+				group = NULL;
+		}
+
+		if (!group)
+			return 1;
+
+		mutex_lock(&group->lock);
+		group->response_sa_mad = *mad;
+		group->prev_state = group->state;
+		group->state = MCAST_RESP_READY;
+		/* calls mlx4_ib_mcg_work_handler */
+		atomic_inc(&group->refcount);
+		if (!queue_work(ctx->mcg_wq, &group->work))
+			safe_atomic_dec(&group->refcount);
+		mutex_unlock(&group->lock);
+		release_group(group, 0);
+		return 1; /* consumed */
+	case IB_MGMT_METHOD_SET:
+	case IB_SA_METHOD_GET_TABLE:
+	case IB_SA_METHOD_GET_TABLE_RESP:
+	case IB_SA_METHOD_DELETE:
+		return 0; /* not consumed, pass-through to guest over tunnel */
+	default:
+		mcg_warn("In demux, port %d: unexpected MCMember method: 0x%x, dropping\n",
+			port, mad->mad_hdr.method);
+		return 1; /* consumed */
+	}
+}
+
+int mlx4_ib_mcg_multiplex_handler(struct ib_device *ibdev, int port,
+				  int slave, struct ib_sa_mad *sa_mad)
+{
+	struct mlx4_ib_dev *dev = to_mdev(ibdev);
+	struct ib_sa_mcmember_data *rec = (struct ib_sa_mcmember_data *)sa_mad->data;
+	struct mlx4_ib_demux_ctx *ctx = &dev->sriov.demux[port - 1];
+	struct mcast_group *group;
+	struct mcast_req *req;
+	int may_create = 0;
+
+	if (ctx->flushing)
+		return -EAGAIN;
+
+	switch (sa_mad->mad_hdr.method) {
+	case IB_MGMT_METHOD_SET:
+		may_create = 1;
+		fallthrough;
+	case IB_SA_METHOD_DELETE:
+		req = kzalloc(sizeof *req, GFP_KERNEL);
+		if (!req)
+			return -ENOMEM;
+
+		req->func = slave;
+		req->sa_mad = *sa_mad;
+
+		mutex_lock(&ctx->mcg_table_lock);
+		group = acquire_group(ctx, &rec->mgid, may_create);
+		mutex_unlock(&ctx->mcg_table_lock);
+		if (IS_ERR(group)) {
+			kfree(req);
+			return PTR_ERR(group);
+		}
+		mutex_lock(&group->lock);
+		if (group->func[slave].num_pend_reqs > MAX_PEND_REQS_PER_FUNC) {
+			mutex_unlock(&group->lock);
+			mcg_debug_group(group, "Port %d, Func %d has too many pending requests (%d), dropping\n",
+					port, slave, MAX_PEND_REQS_PER_FUNC);
+			release_group(group, 0);
+			kfree(req);
+			return -ENOMEM;
+		}
+		++group->func[slave].num_pend_reqs;
+		req->group = group;
+		queue_req(req);
+		mutex_unlock(&group->lock);
+		release_group(group, 0);
+		return 1; /* consumed */
+	case IB_SA_METHOD_GET_TABLE:
+	case IB_MGMT_METHOD_GET_RESP:
+	case IB_SA_METHOD_GET_TABLE_RESP:
+	case IB_SA_METHOD_DELETE_RESP:
+		return 0; /* not consumed, pass-through */
+	default:
+		mcg_warn("In multiplex, port %d, func %d: unexpected MCMember method: 0x%x, dropping\n",
+			port, slave, sa_mad->mad_hdr.method);
+		return 1; /* consumed */
+	}
+}
+
+static ssize_t sysfs_show_group(struct device *dev,
+		struct device_attribute *attr, char *buf)
+{
+	struct mcast_group *group =
+		container_of(attr, struct mcast_group, dentry);
+	struct mcast_req *req = NULL;
+	char pending_str[40];
+	char state_str[40];
+	ssize_t len = 0;
+	int f;
+
+	if (group->state == MCAST_IDLE)
+		sprintf(state_str, "%s", get_state_string(group->state));
+	else
+		sprintf(state_str, "%s(TID=0x%llx)",
+				get_state_string(group->state),
+				be64_to_cpu(group->last_req_tid));
+	if (list_empty(&group->pending_list)) {
+		sprintf(pending_str, "No");
+	} else {
+		req = list_first_entry(&group->pending_list, struct mcast_req, group_list);
+		sprintf(pending_str, "Yes(TID=0x%llx)",
+				be64_to_cpu(req->sa_mad.mad_hdr.tid));
+	}
+	len += sprintf(buf + len, "%1d [%02d,%02d,%02d] %4d %4s %5s     ",
+			group->rec.scope_join_state & 0xf,
+			group->members[2], group->members[1], group->members[0],
+			atomic_read(&group->refcount),
+			pending_str,
+			state_str);
+	for (f = 0; f < MAX_VFS; ++f)
+		if (group->func[f].state == MCAST_MEMBER)
+			len += sprintf(buf + len, "%d[%1x] ",
+					f, group->func[f].join_state);
+
+	len += sprintf(buf + len, "\t\t(%4hx %4x %2x %2x %2x %2x %2x "
+		"%4x %4x %2x %2x)\n",
+		be16_to_cpu(group->rec.pkey),
+		be32_to_cpu(group->rec.qkey),
+		(group->rec.mtusel_mtu & 0xc0) >> 6,
+		group->rec.mtusel_mtu & 0x3f,
+		group->rec.tclass,
+		(group->rec.ratesel_rate & 0xc0) >> 6,
+		group->rec.ratesel_rate & 0x3f,
+		(be32_to_cpu(group->rec.sl_flowlabel_hoplimit) & 0xf0000000) >> 28,
+		(be32_to_cpu(group->rec.sl_flowlabel_hoplimit) & 0x0fffff00) >> 8,
+		be32_to_cpu(group->rec.sl_flowlabel_hoplimit) & 0x000000ff,
+		group->rec.proxy_join);
+
+	return len;
+}
+
+int mlx4_ib_mcg_port_init(struct mlx4_ib_demux_ctx *ctx)
+{
+	char name[20];
+
+	atomic_set(&ctx->tid, 0);
+	sprintf(name, "mlx4_ib_mcg%d", ctx->port);
+	ctx->mcg_wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
+	if (!ctx->mcg_wq)
+		return -ENOMEM;
+
+	mutex_init(&ctx->mcg_table_lock);
+	ctx->mcg_table = RB_ROOT;
+	INIT_LIST_HEAD(&ctx->mcg_mgid0_list);
+	ctx->flushing = 0;
+
+	return 0;
+}
+
+static void force_clean_group(struct mcast_group *group)
+{
+	struct mcast_req *req, *tmp
+		;
+	list_for_each_entry_safe(req, tmp, &group->pending_list, group_list) {
+		list_del(&req->group_list);
+		kfree(req);
+	}
+	del_sysfs_port_mcg_attr(group->demux->dev, group->demux->port, &group->dentry.attr);
+	rb_erase(&group->node, &group->demux->mcg_table);
+	kfree(group);
+}
+
+static void _mlx4_ib_mcg_port_cleanup(struct mlx4_ib_demux_ctx *ctx, int destroy_wq)
+{
+	int i;
+	struct rb_node *p;
+	struct mcast_group *group;
+	unsigned long end;
+	int count;
+
+	for (i = 0; i < MAX_VFS; ++i)
+		clean_vf_mcast(ctx, i);
+
+	end = jiffies + msecs_to_jiffies(MAD_TIMEOUT_MS + 3000);
+	do {
+		count = 0;
+		mutex_lock(&ctx->mcg_table_lock);
+		for (p = rb_first(&ctx->mcg_table); p; p = rb_next(p))
+			++count;
+		mutex_unlock(&ctx->mcg_table_lock);
+		if (!count)
+			break;
+
+		usleep_range(1000, 2000);
+	} while (time_after(end, jiffies));
+
+	flush_workqueue(ctx->mcg_wq);
+	if (destroy_wq)
+		destroy_workqueue(ctx->mcg_wq);
+
+	mutex_lock(&ctx->mcg_table_lock);
+	while ((p = rb_first(&ctx->mcg_table)) != NULL) {
+		group = rb_entry(p, struct mcast_group, node);
+		if (atomic_read(&group->refcount))
+			mcg_debug_group(group, "group refcount %d!!! (pointer %p)\n",
+					atomic_read(&group->refcount), group);
+
+		force_clean_group(group);
+	}
+	mutex_unlock(&ctx->mcg_table_lock);
+}
+
+struct clean_work {
+	struct work_struct work;
+	struct mlx4_ib_demux_ctx *ctx;
+	int destroy_wq;
+};
+
+static void mcg_clean_task(struct work_struct *work)
+{
+	struct clean_work *cw = container_of(work, struct clean_work, work);
+
+	_mlx4_ib_mcg_port_cleanup(cw->ctx, cw->destroy_wq);
+	cw->ctx->flushing = 0;
+	kfree(cw);
+}
+
+void mlx4_ib_mcg_port_cleanup(struct mlx4_ib_demux_ctx *ctx, int destroy_wq)
+{
+	struct clean_work *work;
+
+	if (ctx->flushing)
+		return;
+
+	ctx->flushing = 1;
+
+	if (destroy_wq) {
+		_mlx4_ib_mcg_port_cleanup(ctx, destroy_wq);
+		ctx->flushing = 0;
+		return;
+	}
+
+	work = kmalloc(sizeof *work, GFP_KERNEL);
+	if (!work) {
+		ctx->flushing = 0;
+		return;
+	}
+
+	work->ctx = ctx;
+	work->destroy_wq = destroy_wq;
+	INIT_WORK(&work->work, mcg_clean_task);
+	queue_work(clean_wq, &work->work);
+}
+
+static void build_leave_mad(struct mcast_req *req)
+{
+	struct ib_sa_mad *mad = &req->sa_mad;
+
+	mad->mad_hdr.method = IB_SA_METHOD_DELETE;
+}
+
+
+static void clear_pending_reqs(struct mcast_group *group, int vf)
+{
+	struct mcast_req *req, *tmp, *group_first = NULL;
+	int clear;
+	int pend = 0;
+
+	if (!list_empty(&group->pending_list))
+		group_first = list_first_entry(&group->pending_list, struct mcast_req, group_list);
+
+	list_for_each_entry_safe(req, tmp, &group->func[vf].pending, func_list) {
+		clear = 1;
+		if (group_first == req &&
+		    (group->state == MCAST_JOIN_SENT ||
+		     group->state == MCAST_LEAVE_SENT)) {
+			clear = cancel_delayed_work(&group->timeout_work);
+			pend = !clear;
+			group->state = MCAST_IDLE;
+		}
+		if (clear) {
+			--group->func[vf].num_pend_reqs;
+			list_del(&req->group_list);
+			list_del(&req->func_list);
+			kfree(req);
+			atomic_dec(&group->refcount);
+		}
+	}
+
+	if (!pend && (!list_empty(&group->func[vf].pending) || group->func[vf].num_pend_reqs)) {
+		mcg_warn_group(group, "DRIVER BUG: list_empty %d, num_pend_reqs %d\n",
+			       list_empty(&group->func[vf].pending), group->func[vf].num_pend_reqs);
+	}
+}
+
+static int push_deleteing_req(struct mcast_group *group, int slave)
+{
+	struct mcast_req *req;
+	struct mcast_req *pend_req;
+
+	if (!group->func[slave].join_state)
+		return 0;
+
+	req = kzalloc(sizeof *req, GFP_KERNEL);
+	if (!req)
+		return -ENOMEM;
+
+	if (!list_empty(&group->func[slave].pending)) {
+		pend_req = list_entry(group->func[slave].pending.prev, struct mcast_req, group_list);
+		if (pend_req->clean) {
+			kfree(req);
+			return 0;
+		}
+	}
+
+	req->clean = 1;
+	req->func = slave;
+	req->group = group;
+	++group->func[slave].num_pend_reqs;
+	build_leave_mad(req);
+	queue_req(req);
+	return 0;
+}
+
+void clean_vf_mcast(struct mlx4_ib_demux_ctx *ctx, int slave)
+{
+	struct mcast_group *group;
+	struct rb_node *p;
+
+	mutex_lock(&ctx->mcg_table_lock);
+	for (p = rb_first(&ctx->mcg_table); p; p = rb_next(p)) {
+		group = rb_entry(p, struct mcast_group, node);
+		mutex_lock(&group->lock);
+		if (atomic_read(&group->refcount)) {
+			/* clear pending requests of this VF */
+			clear_pending_reqs(group, slave);
+			push_deleteing_req(group, slave);
+		}
+		mutex_unlock(&group->lock);
+	}
+	mutex_unlock(&ctx->mcg_table_lock);
+}
+
+
+int mlx4_ib_mcg_init(void)
+{
+	clean_wq = alloc_ordered_workqueue("mlx4_ib_mcg", WQ_MEM_RECLAIM);
+	if (!clean_wq)
+		return -ENOMEM;
+
+	return 0;
+}
+
+void mlx4_ib_mcg_destroy(void)
+{
+	destroy_workqueue(clean_wq);
+}
diff --git a/drivers/infiniband/hw/mlx4/mlx4_ib.h b/drivers/infiniband/hw/mlx4/mlx4_ib.h
new file mode 100644
index 000000000..58df06492
--- /dev/null
+++ b/drivers/infiniband/hw/mlx4/mlx4_ib.h
@@ -0,0 +1,941 @@
+/*
+ * Copyright (c) 2006, 2007 Cisco Systems.  All rights reserved.
+ * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef MLX4_IB_H
+#define MLX4_IB_H
+
+#include <linux/compiler.h>
+#include <linux/list.h>
+#include <linux/mutex.h>
+#include <linux/idr.h>
+
+#include <rdma/ib_verbs.h>
+#include <rdma/ib_umem.h>
+#include <rdma/ib_mad.h>
+#include <rdma/ib_sa.h>
+
+#include <linux/mlx4/device.h>
+#include <linux/mlx4/doorbell.h>
+#include <linux/mlx4/qp.h>
+#include <linux/mlx4/cq.h>
+
+#define MLX4_IB_DRV_NAME	"mlx4_ib"
+
+#ifdef pr_fmt
+#undef pr_fmt
+#endif
+#define pr_fmt(fmt)	"<" MLX4_IB_DRV_NAME "> %s: " fmt, __func__
+
+#define mlx4_ib_warn(ibdev, format, arg...) \
+	dev_warn((ibdev)->dev.parent, MLX4_IB_DRV_NAME ": " format, ## arg)
+
+enum {
+	MLX4_IB_SQ_MIN_WQE_SHIFT = 6,
+	MLX4_IB_MAX_HEADROOM	 = 2048
+};
+
+#define MLX4_IB_SQ_HEADROOM(shift)	((MLX4_IB_MAX_HEADROOM >> (shift)) + 1)
+#define MLX4_IB_SQ_MAX_SPARE		(MLX4_IB_SQ_HEADROOM(MLX4_IB_SQ_MIN_WQE_SHIFT))
+
+/*module param to indicate if SM assigns the alias_GUID*/
+extern int mlx4_ib_sm_guid_assign;
+
+#define MLX4_IB_UC_STEER_QPN_ALIGN 1
+#define MLX4_IB_UC_MAX_NUM_QPS     256
+
+enum hw_bar_type {
+	HW_BAR_BF,
+	HW_BAR_DB,
+	HW_BAR_CLOCK,
+	HW_BAR_COUNT
+};
+
+struct mlx4_ib_ucontext {
+	struct ib_ucontext	ibucontext;
+	struct mlx4_uar		uar;
+	struct list_head	db_page_list;
+	struct mutex		db_page_mutex;
+	struct list_head	wqn_ranges_list;
+	struct mutex		wqn_ranges_mutex; /* protect wqn_ranges_list */
+};
+
+struct mlx4_ib_pd {
+	struct ib_pd		ibpd;
+	u32			pdn;
+};
+
+struct mlx4_ib_xrcd {
+	struct ib_xrcd		ibxrcd;
+	u32			xrcdn;
+	struct ib_pd	       *pd;
+	struct ib_cq	       *cq;
+};
+
+struct mlx4_ib_cq_buf {
+	struct mlx4_buf		buf;
+	struct mlx4_mtt		mtt;
+	int			entry_size;
+};
+
+struct mlx4_ib_cq_resize {
+	struct mlx4_ib_cq_buf	buf;
+	int			cqe;
+};
+
+struct mlx4_ib_cq {
+	struct ib_cq		ibcq;
+	struct mlx4_cq		mcq;
+	struct mlx4_ib_cq_buf	buf;
+	struct mlx4_ib_cq_resize *resize_buf;
+	struct mlx4_db		db;
+	spinlock_t		lock;
+	struct mutex		resize_mutex;
+	struct ib_umem	       *umem;
+	struct ib_umem	       *resize_umem;
+	int			create_flags;
+	/* List of qps that it serves.*/
+	struct list_head		send_qp_list;
+	struct list_head		recv_qp_list;
+};
+
+#define MLX4_MR_PAGES_ALIGN 0x40
+
+struct mlx4_ib_mr {
+	struct ib_mr		ibmr;
+	__be64			*pages;
+	dma_addr_t		page_map;
+	u32			npages;
+	u32			max_pages;
+	struct mlx4_mr		mmr;
+	struct ib_umem	       *umem;
+	size_t			page_map_size;
+};
+
+struct mlx4_ib_mw {
+	struct ib_mw		ibmw;
+	struct mlx4_mw		mmw;
+};
+
+#define MAX_REGS_PER_FLOW 2
+
+struct mlx4_flow_reg_id {
+	u64 id;
+	u64 mirror;
+};
+
+struct mlx4_ib_flow {
+	struct ib_flow ibflow;
+	/* translating DMFS verbs sniffer rule to FW API requires two reg IDs */
+	struct mlx4_flow_reg_id reg_id[MAX_REGS_PER_FLOW];
+};
+
+struct mlx4_ib_wq {
+	u64		       *wrid;
+	spinlock_t		lock;
+	int			wqe_cnt;
+	int			max_post;
+	int			max_gs;
+	int			offset;
+	int			wqe_shift;
+	unsigned		head;
+	unsigned		tail;
+};
+
+enum {
+	MLX4_IB_QP_CREATE_ROCE_V2_GSI = IB_QP_CREATE_RESERVED_START
+};
+
+enum mlx4_ib_qp_flags {
+	MLX4_IB_QP_LSO = IB_QP_CREATE_IPOIB_UD_LSO,
+	MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK = IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK,
+	MLX4_IB_QP_NETIF = IB_QP_CREATE_NETIF_QP,
+	MLX4_IB_QP_SCATTER_FCS = IB_QP_CREATE_SCATTER_FCS,
+
+	/* Mellanox specific flags start from IB_QP_CREATE_RESERVED_START */
+	MLX4_IB_ROCE_V2_GSI_QP = MLX4_IB_QP_CREATE_ROCE_V2_GSI,
+	MLX4_IB_SRIOV_TUNNEL_QP = 1 << 30,
+	MLX4_IB_SRIOV_SQP = 1 << 31,
+};
+
+struct mlx4_ib_gid_entry {
+	struct list_head	list;
+	union ib_gid		gid;
+	int			added;
+	u8			port;
+};
+
+enum mlx4_ib_qp_type {
+	/*
+	 * IB_QPT_SMI and IB_QPT_GSI have to be the first two entries
+	 * here (and in that order) since the MAD layer uses them as
+	 * indices into a 2-entry table.
+	 */
+	MLX4_IB_QPT_SMI = IB_QPT_SMI,
+	MLX4_IB_QPT_GSI = IB_QPT_GSI,
+
+	MLX4_IB_QPT_RC = IB_QPT_RC,
+	MLX4_IB_QPT_UC = IB_QPT_UC,
+	MLX4_IB_QPT_UD = IB_QPT_UD,
+	MLX4_IB_QPT_RAW_IPV6 = IB_QPT_RAW_IPV6,
+	MLX4_IB_QPT_RAW_ETHERTYPE = IB_QPT_RAW_ETHERTYPE,
+	MLX4_IB_QPT_RAW_PACKET = IB_QPT_RAW_PACKET,
+	MLX4_IB_QPT_XRC_INI = IB_QPT_XRC_INI,
+	MLX4_IB_QPT_XRC_TGT = IB_QPT_XRC_TGT,
+
+	MLX4_IB_QPT_PROXY_SMI_OWNER	= 1 << 16,
+	MLX4_IB_QPT_PROXY_SMI		= 1 << 17,
+	MLX4_IB_QPT_PROXY_GSI		= 1 << 18,
+	MLX4_IB_QPT_TUN_SMI_OWNER	= 1 << 19,
+	MLX4_IB_QPT_TUN_SMI		= 1 << 20,
+	MLX4_IB_QPT_TUN_GSI		= 1 << 21,
+};
+
+#define MLX4_IB_QPT_ANY_SRIOV	(MLX4_IB_QPT_PROXY_SMI_OWNER | \
+	MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER | \
+	MLX4_IB_QPT_TUN_SMI | MLX4_IB_QPT_TUN_GSI)
+
+enum mlx4_ib_mad_ifc_flags {
+	MLX4_MAD_IFC_IGNORE_MKEY	= 1,
+	MLX4_MAD_IFC_IGNORE_BKEY	= 2,
+	MLX4_MAD_IFC_IGNORE_KEYS	= (MLX4_MAD_IFC_IGNORE_MKEY |
+					   MLX4_MAD_IFC_IGNORE_BKEY),
+	MLX4_MAD_IFC_NET_VIEW		= 4,
+};
+
+enum {
+	MLX4_NUM_TUNNEL_BUFS		= 512,
+	MLX4_NUM_WIRE_BUFS		= 2048,
+};
+
+struct mlx4_ib_tunnel_header {
+	struct mlx4_av av;
+	__be32 remote_qpn;
+	__be32 qkey;
+	__be16 vlan;
+	u8 mac[6];
+	__be16 pkey_index;
+	u8 reserved[6];
+};
+
+struct mlx4_ib_buf {
+	void *addr;
+	dma_addr_t map;
+};
+
+struct mlx4_rcv_tunnel_hdr {
+	__be32 flags_src_qp; /* flags[6:5] is defined for VLANs:
+			      * 0x0 - no vlan was in the packet
+			      * 0x01 - C-VLAN was in the packet */
+	u8 g_ml_path; /* gid bit stands for ipv6/4 header in RoCE */
+	u8 reserved;
+	__be16 pkey_index;
+	__be16 sl_vid;
+	__be16 slid_mac_47_32;
+	__be32 mac_31_0;
+};
+
+struct mlx4_ib_proxy_sqp_hdr {
+	struct ib_grh grh;
+	struct mlx4_rcv_tunnel_hdr tun;
+}  __packed;
+
+struct mlx4_roce_smac_vlan_info {
+	u64 smac;
+	int smac_index;
+	int smac_port;
+	u64 candidate_smac;
+	int candidate_smac_index;
+	int candidate_smac_port;
+	u16 vid;
+	int vlan_index;
+	int vlan_port;
+	u16 candidate_vid;
+	int candidate_vlan_index;
+	int candidate_vlan_port;
+	int update_vid;
+};
+
+struct mlx4_wqn_range {
+	int			base_wqn;
+	int			size;
+	int			refcount;
+	bool			dirty;
+	struct list_head	list;
+};
+
+struct mlx4_ib_rss {
+	unsigned int		base_qpn_tbl_sz;
+	u8			flags;
+	u8			rss_key[MLX4_EN_RSS_KEY_SIZE];
+};
+
+enum {
+	/*
+	 * Largest possible UD header: send with GRH and immediate
+	 * data plus 18 bytes for an Ethernet header with VLAN/802.1Q
+	 * tag.  (LRH would only use 8 bytes, so Ethernet is the
+	 * biggest case)
+	 */
+	MLX4_IB_UD_HEADER_SIZE		= 82,
+	MLX4_IB_LSO_HEADER_SPARE	= 128,
+};
+
+struct mlx4_ib_sqp {
+	int pkey_index;
+	u32 qkey;
+	u32 send_psn;
+	struct ib_ud_header ud_header;
+	u8 header_buf[MLX4_IB_UD_HEADER_SIZE];
+	struct ib_qp *roce_v2_gsi;
+};
+
+struct mlx4_ib_qp {
+	union {
+		struct ib_qp	ibqp;
+		struct ib_wq	ibwq;
+	};
+	struct mlx4_qp		mqp;
+	struct mlx4_buf		buf;
+
+	struct mlx4_db		db;
+	struct mlx4_ib_wq	rq;
+
+	u32			doorbell_qpn;
+	__be32			sq_signal_bits;
+	unsigned		sq_next_wqe;
+	int			sq_spare_wqes;
+	struct mlx4_ib_wq	sq;
+
+	enum mlx4_ib_qp_type	mlx4_ib_qp_type;
+	struct ib_umem	       *umem;
+	struct mlx4_mtt		mtt;
+	int			buf_size;
+	struct mutex		mutex;
+	u16			xrcdn;
+	u32			flags;
+	u8			port;
+	u8			alt_port;
+	u8			atomic_rd_en;
+	u8			resp_depth;
+	u8			sq_no_prefetch;
+	u8			state;
+	int			mlx_type;
+	u32			inl_recv_sz;
+	struct list_head	gid_list;
+	struct list_head	steering_rules;
+	struct mlx4_ib_buf	*sqp_proxy_rcv;
+	struct mlx4_roce_smac_vlan_info pri;
+	struct mlx4_roce_smac_vlan_info alt;
+	u64			reg_id;
+	struct list_head	qps_list;
+	struct list_head	cq_recv_list;
+	struct list_head	cq_send_list;
+	struct counter_index	*counter_index;
+	struct mlx4_wqn_range	*wqn_range;
+	/* Number of RSS QP parents that uses this WQ */
+	u32			rss_usecnt;
+	union {
+		struct mlx4_ib_rss *rss_ctx;
+		struct mlx4_ib_sqp *sqp;
+	};
+};
+
+struct mlx4_ib_srq {
+	struct ib_srq		ibsrq;
+	struct mlx4_srq		msrq;
+	struct mlx4_buf		buf;
+	struct mlx4_db		db;
+	u64		       *wrid;
+	spinlock_t		lock;
+	int			head;
+	int			tail;
+	u16			wqe_ctr;
+	struct ib_umem	       *umem;
+	struct mlx4_mtt		mtt;
+	struct mutex		mutex;
+};
+
+struct mlx4_ib_ah {
+	struct ib_ah		ibah;
+	union mlx4_ext_av       av;
+};
+
+struct mlx4_ib_rwq_ind_table {
+	struct ib_rwq_ind_table ib_rwq_ind_tbl;
+};
+
+/****************************************/
+/* alias guid support */
+/****************************************/
+#define NUM_PORT_ALIAS_GUID		2
+#define NUM_ALIAS_GUID_IN_REC		8
+#define NUM_ALIAS_GUID_REC_IN_PORT	16
+#define GUID_REC_SIZE			8
+#define NUM_ALIAS_GUID_PER_PORT		128
+#define MLX4_NOT_SET_GUID		(0x00LL)
+#define MLX4_GUID_FOR_DELETE_VAL	(~(0x00LL))
+
+enum mlx4_guid_alias_rec_status {
+	MLX4_GUID_INFO_STATUS_IDLE,
+	MLX4_GUID_INFO_STATUS_SET,
+};
+
+#define GUID_STATE_NEED_PORT_INIT 0x01
+
+enum mlx4_guid_alias_rec_method {
+	MLX4_GUID_INFO_RECORD_SET	= IB_MGMT_METHOD_SET,
+	MLX4_GUID_INFO_RECORD_DELETE	= IB_SA_METHOD_DELETE,
+};
+
+struct mlx4_sriov_alias_guid_info_rec_det {
+	u8 all_recs[GUID_REC_SIZE * NUM_ALIAS_GUID_IN_REC];
+	ib_sa_comp_mask guid_indexes; /*indicates what from the 8 records are valid*/
+	enum mlx4_guid_alias_rec_status status; /*indicates the administraively status of the record.*/
+	unsigned int guids_retry_schedule[NUM_ALIAS_GUID_IN_REC];
+	u64 time_to_run;
+};
+
+struct mlx4_sriov_alias_guid_port_rec_det {
+	struct mlx4_sriov_alias_guid_info_rec_det all_rec_per_port[NUM_ALIAS_GUID_REC_IN_PORT];
+	struct workqueue_struct *wq;
+	struct delayed_work alias_guid_work;
+	u8 port;
+	u32 state_flags;
+	struct mlx4_sriov_alias_guid *parent;
+	struct list_head cb_list;
+};
+
+struct mlx4_sriov_alias_guid {
+	struct mlx4_sriov_alias_guid_port_rec_det ports_guid[MLX4_MAX_PORTS];
+	spinlock_t ag_work_lock;
+	struct ib_sa_client *sa_client;
+};
+
+struct mlx4_ib_demux_work {
+	struct work_struct	work;
+	struct mlx4_ib_dev     *dev;
+	int			slave;
+	int			do_init;
+	u8			port;
+
+};
+
+struct mlx4_ib_tun_tx_buf {
+	struct mlx4_ib_buf buf;
+	struct ib_ah *ah;
+};
+
+struct mlx4_ib_demux_pv_qp {
+	struct ib_qp *qp;
+	enum ib_qp_type proxy_qpt;
+	struct mlx4_ib_buf *ring;
+	struct mlx4_ib_tun_tx_buf *tx_ring;
+	spinlock_t tx_lock;
+	unsigned tx_ix_head;
+	unsigned tx_ix_tail;
+};
+
+enum mlx4_ib_demux_pv_state {
+	DEMUX_PV_STATE_DOWN,
+	DEMUX_PV_STATE_STARTING,
+	DEMUX_PV_STATE_ACTIVE,
+	DEMUX_PV_STATE_DOWNING,
+};
+
+struct mlx4_ib_demux_pv_ctx {
+	int port;
+	int slave;
+	enum mlx4_ib_demux_pv_state state;
+	int has_smi;
+	struct ib_device *ib_dev;
+	struct ib_cq *cq;
+	struct ib_pd *pd;
+	struct work_struct work;
+	struct workqueue_struct *wq;
+	struct workqueue_struct *wi_wq;
+	struct mlx4_ib_demux_pv_qp qp[2];
+};
+
+struct mlx4_ib_demux_ctx {
+	struct ib_device *ib_dev;
+	int port;
+	struct workqueue_struct *wq;
+	struct workqueue_struct *wi_wq;
+	struct workqueue_struct *ud_wq;
+	spinlock_t ud_lock;
+	atomic64_t subnet_prefix;
+	__be64 guid_cache[128];
+	struct mlx4_ib_dev *dev;
+	/* the following lock protects both mcg_table and mcg_mgid0_list */
+	struct mutex		mcg_table_lock;
+	struct rb_root		mcg_table;
+	struct list_head	mcg_mgid0_list;
+	struct workqueue_struct	*mcg_wq;
+	struct mlx4_ib_demux_pv_ctx **tun;
+	atomic_t tid;
+	int    flushing; /* flushing the work queue */
+};
+
+struct mlx4_ib_sriov {
+	struct mlx4_ib_demux_ctx demux[MLX4_MAX_PORTS];
+	struct mlx4_ib_demux_pv_ctx *sqps[MLX4_MAX_PORTS];
+	/* when using this spinlock you should use "irq" because
+	 * it may be called from interrupt context.*/
+	spinlock_t going_down_lock;
+	int is_going_down;
+
+	struct mlx4_sriov_alias_guid alias_guid;
+
+	/* CM paravirtualization fields */
+	struct xarray pv_id_table;
+	u32 pv_id_next;
+	spinlock_t id_map_lock;
+	struct rb_root sl_id_map;
+	struct list_head cm_list;
+	struct xarray xa_rej_tmout;
+};
+
+struct gid_cache_context {
+	int real_index;
+	int refcount;
+};
+
+struct gid_entry {
+	union ib_gid	gid;
+	enum ib_gid_type gid_type;
+	struct gid_cache_context *ctx;
+	u16 vlan_id;
+};
+
+struct mlx4_port_gid_table {
+	struct gid_entry gids[MLX4_MAX_PORT_GIDS];
+};
+
+struct mlx4_ib_iboe {
+	spinlock_t		lock;
+	struct net_device      *netdevs[MLX4_MAX_PORTS];
+	atomic64_t		mac[MLX4_MAX_PORTS];
+	struct notifier_block 	nb;
+	struct mlx4_port_gid_table gids[MLX4_MAX_PORTS];
+	enum ib_port_state	last_port_state[MLX4_MAX_PORTS];
+};
+
+struct pkey_mgt {
+	u8			virt2phys_pkey[MLX4_MFUNC_MAX][MLX4_MAX_PORTS][MLX4_MAX_PORT_PKEYS];
+	u16			phys_pkey_cache[MLX4_MAX_PORTS][MLX4_MAX_PORT_PKEYS];
+	struct list_head	pkey_port_list[MLX4_MFUNC_MAX];
+	struct kobject	       *device_parent[MLX4_MFUNC_MAX];
+};
+
+struct mlx4_ib_iov_sysfs_attr {
+	void *ctx;
+	struct kobject *kobj;
+	unsigned long data;
+	u32 entry_num;
+	char name[15];
+	struct device_attribute dentry;
+	struct device *dev;
+};
+
+struct mlx4_ib_iov_sysfs_attr_ar {
+	struct mlx4_ib_iov_sysfs_attr dentries[3 * NUM_ALIAS_GUID_PER_PORT + 1];
+};
+
+struct mlx4_ib_iov_port {
+	char name[100];
+	u8 num;
+	struct mlx4_ib_dev *dev;
+	struct list_head list;
+	struct mlx4_ib_iov_sysfs_attr_ar *dentr_ar;
+	struct ib_port_attr attr;
+	struct kobject	*cur_port;
+	struct kobject	*admin_alias_parent;
+	struct kobject	*gids_parent;
+	struct kobject	*pkeys_parent;
+	struct kobject	*mcgs_parent;
+	struct mlx4_ib_iov_sysfs_attr mcg_dentry;
+};
+
+struct counter_index {
+	struct  list_head       list;
+	u32		index;
+	u8		allocated;
+};
+
+struct mlx4_ib_counters {
+	struct list_head        counters_list;
+	struct mutex            mutex; /* mutex for accessing counters list */
+	u32			default_counter;
+};
+
+#define MLX4_DIAG_COUNTERS_TYPES 2
+
+struct mlx4_ib_diag_counters {
+	const char **name;
+	u32 *offset;
+	u32 num_counters;
+};
+
+struct mlx4_ib_dev {
+	struct ib_device	ib_dev;
+	struct mlx4_dev	       *dev;
+	int			num_ports;
+	void __iomem	       *uar_map;
+
+	struct mlx4_uar		priv_uar;
+	u32			priv_pdn;
+	MLX4_DECLARE_DOORBELL_LOCK(uar_lock);
+
+	struct ib_mad_agent    *send_agent[MLX4_MAX_PORTS][2];
+	struct ib_ah	       *sm_ah[MLX4_MAX_PORTS];
+	spinlock_t		sm_lock;
+	atomic64_t		sl2vl[MLX4_MAX_PORTS];
+	struct mlx4_ib_sriov	sriov;
+
+	struct mutex		cap_mask_mutex;
+	bool			ib_active;
+	struct mlx4_ib_iboe	iboe;
+	struct mlx4_ib_counters counters_table[MLX4_MAX_PORTS];
+	int		       *eq_table;
+	struct kobject	       *iov_parent;
+	struct kobject	       *ports_parent;
+	struct kobject	       *dev_ports_parent[MLX4_MFUNC_MAX];
+	struct mlx4_ib_iov_port	iov_ports[MLX4_MAX_PORTS];
+	struct pkey_mgt		pkeys;
+	unsigned long *ib_uc_qpns_bitmap;
+	int steer_qpn_count;
+	int steer_qpn_base;
+	int steering_support;
+	struct mlx4_ib_qp      *qp1_proxy[MLX4_MAX_PORTS];
+	/* lock when destroying qp1_proxy and getting netdev events */
+	struct mutex		qp1_proxy_lock[MLX4_MAX_PORTS];
+	u8			bond_next_port;
+	/* protect resources needed as part of reset flow */
+	spinlock_t		reset_flow_resource_lock;
+	struct list_head		qp_list;
+	struct mlx4_ib_diag_counters diag_counters[MLX4_DIAG_COUNTERS_TYPES];
+};
+
+struct ib_event_work {
+	struct work_struct	work;
+	struct mlx4_ib_dev	*ib_dev;
+	struct mlx4_eqe		ib_eqe;
+	int			port;
+};
+
+struct mlx4_ib_qp_tunnel_init_attr {
+	struct ib_qp_init_attr init_attr;
+	int slave;
+	enum ib_qp_type proxy_qp_type;
+	u8 port;
+};
+
+struct mlx4_uverbs_ex_query_device {
+	__u32 comp_mask;
+	__u32 reserved;
+};
+
+static inline struct mlx4_ib_dev *to_mdev(struct ib_device *ibdev)
+{
+	return container_of(ibdev, struct mlx4_ib_dev, ib_dev);
+}
+
+static inline struct mlx4_ib_ucontext *to_mucontext(struct ib_ucontext *ibucontext)
+{
+	return container_of(ibucontext, struct mlx4_ib_ucontext, ibucontext);
+}
+
+static inline struct mlx4_ib_pd *to_mpd(struct ib_pd *ibpd)
+{
+	return container_of(ibpd, struct mlx4_ib_pd, ibpd);
+}
+
+static inline struct mlx4_ib_xrcd *to_mxrcd(struct ib_xrcd *ibxrcd)
+{
+	return container_of(ibxrcd, struct mlx4_ib_xrcd, ibxrcd);
+}
+
+static inline struct mlx4_ib_cq *to_mcq(struct ib_cq *ibcq)
+{
+	return container_of(ibcq, struct mlx4_ib_cq, ibcq);
+}
+
+static inline struct mlx4_ib_cq *to_mibcq(struct mlx4_cq *mcq)
+{
+	return container_of(mcq, struct mlx4_ib_cq, mcq);
+}
+
+static inline struct mlx4_ib_mr *to_mmr(struct ib_mr *ibmr)
+{
+	return container_of(ibmr, struct mlx4_ib_mr, ibmr);
+}
+
+static inline struct mlx4_ib_mw *to_mmw(struct ib_mw *ibmw)
+{
+	return container_of(ibmw, struct mlx4_ib_mw, ibmw);
+}
+
+static inline struct mlx4_ib_flow *to_mflow(struct ib_flow *ibflow)
+{
+	return container_of(ibflow, struct mlx4_ib_flow, ibflow);
+}
+
+static inline struct mlx4_ib_qp *to_mqp(struct ib_qp *ibqp)
+{
+	return container_of(ibqp, struct mlx4_ib_qp, ibqp);
+}
+
+static inline struct mlx4_ib_qp *to_mibqp(struct mlx4_qp *mqp)
+{
+	return container_of(mqp, struct mlx4_ib_qp, mqp);
+}
+
+static inline struct mlx4_ib_srq *to_msrq(struct ib_srq *ibsrq)
+{
+	return container_of(ibsrq, struct mlx4_ib_srq, ibsrq);
+}
+
+static inline struct mlx4_ib_srq *to_mibsrq(struct mlx4_srq *msrq)
+{
+	return container_of(msrq, struct mlx4_ib_srq, msrq);
+}
+
+static inline struct mlx4_ib_ah *to_mah(struct ib_ah *ibah)
+{
+	return container_of(ibah, struct mlx4_ib_ah, ibah);
+}
+
+static inline u8 mlx4_ib_bond_next_port(struct mlx4_ib_dev *dev)
+{
+	dev->bond_next_port = (dev->bond_next_port + 1) % dev->num_ports;
+
+	return dev->bond_next_port + 1;
+}
+
+int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev);
+void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev);
+
+int mlx4_ib_db_map_user(struct ib_udata *udata, unsigned long virt,
+			struct mlx4_db *db);
+void mlx4_ib_db_unmap_user(struct mlx4_ib_ucontext *context, struct mlx4_db *db);
+
+struct ib_mr *mlx4_ib_get_dma_mr(struct ib_pd *pd, int acc);
+int mlx4_ib_umem_write_mtt(struct mlx4_ib_dev *dev, struct mlx4_mtt *mtt,
+			   struct ib_umem *umem);
+struct ib_mr *mlx4_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
+				  u64 virt_addr, int access_flags,
+				  struct ib_udata *udata);
+int mlx4_ib_dereg_mr(struct ib_mr *mr, struct ib_udata *udata);
+int mlx4_ib_alloc_mw(struct ib_mw *mw, struct ib_udata *udata);
+int mlx4_ib_dealloc_mw(struct ib_mw *mw);
+struct ib_mr *mlx4_ib_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
+			       u32 max_num_sg);
+int mlx4_ib_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
+		      unsigned int *sg_offset);
+int mlx4_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period);
+int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata);
+int mlx4_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
+		      struct ib_udata *udata);
+int mlx4_ib_destroy_cq(struct ib_cq *cq, struct ib_udata *udata);
+int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc);
+int mlx4_ib_arm_cq(struct ib_cq *cq, enum ib_cq_notify_flags flags);
+void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq);
+void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq);
+
+int mlx4_ib_create_ah(struct ib_ah *ah, struct rdma_ah_init_attr *init_attr,
+		      struct ib_udata *udata);
+int mlx4_ib_create_ah_slave(struct ib_ah *ah, struct rdma_ah_attr *ah_attr,
+			    int slave_sgid_index, u8 *s_mac, u16 vlan_tag);
+int mlx4_ib_query_ah(struct ib_ah *ibah, struct rdma_ah_attr *ah_attr);
+static inline int mlx4_ib_destroy_ah(struct ib_ah *ah, u32 flags)
+{
+	return 0;
+}
+
+int mlx4_ib_create_srq(struct ib_srq *srq, struct ib_srq_init_attr *init_attr,
+		       struct ib_udata *udata);
+int mlx4_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
+		       enum ib_srq_attr_mask attr_mask, struct ib_udata *udata);
+int mlx4_ib_query_srq(struct ib_srq *srq, struct ib_srq_attr *srq_attr);
+int mlx4_ib_destroy_srq(struct ib_srq *srq, struct ib_udata *udata);
+void mlx4_ib_free_srq_wqe(struct mlx4_ib_srq *srq, int wqe_index);
+int mlx4_ib_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
+			  const struct ib_recv_wr **bad_wr);
+
+struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
+				struct ib_qp_init_attr *init_attr,
+				struct ib_udata *udata);
+int mlx4_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata);
+void mlx4_ib_drain_sq(struct ib_qp *qp);
+void mlx4_ib_drain_rq(struct ib_qp *qp);
+int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
+		      int attr_mask, struct ib_udata *udata);
+int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
+		     struct ib_qp_init_attr *qp_init_attr);
+int mlx4_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
+		      const struct ib_send_wr **bad_wr);
+int mlx4_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
+		      const struct ib_recv_wr **bad_wr);
+
+int mlx4_MAD_IFC(struct mlx4_ib_dev *dev, int mad_ifc_flags,
+		 int port, const struct ib_wc *in_wc, const struct ib_grh *in_grh,
+		 const void *in_mad, void *response_mad);
+int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
+			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
+			const struct ib_mad *in, struct ib_mad *out,
+			size_t *out_mad_size, u16 *out_mad_pkey_index);
+int mlx4_ib_mad_init(struct mlx4_ib_dev *dev);
+void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev);
+
+int __mlx4_ib_query_port(struct ib_device *ibdev, u8 port,
+			 struct ib_port_attr *props, int netw_view);
+int __mlx4_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
+			 u16 *pkey, int netw_view);
+
+int __mlx4_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
+			union ib_gid *gid, int netw_view);
+
+static inline bool mlx4_ib_ah_grh_present(struct mlx4_ib_ah *ah)
+{
+	u8 port = be32_to_cpu(ah->av.ib.port_pd) >> 24 & 3;
+
+	if (rdma_port_get_link_layer(ah->ibah.device, port) == IB_LINK_LAYER_ETHERNET)
+		return true;
+
+	return !!(ah->av.ib.g_slid & 0x80);
+}
+
+int mlx4_ib_mcg_port_init(struct mlx4_ib_demux_ctx *ctx);
+void mlx4_ib_mcg_port_cleanup(struct mlx4_ib_demux_ctx *ctx, int destroy_wq);
+void clean_vf_mcast(struct mlx4_ib_demux_ctx *ctx, int slave);
+int mlx4_ib_mcg_init(void);
+void mlx4_ib_mcg_destroy(void);
+
+int mlx4_ib_find_real_gid(struct ib_device *ibdev, u8 port, __be64 guid);
+
+int mlx4_ib_mcg_multiplex_handler(struct ib_device *ibdev, int port, int slave,
+				  struct ib_sa_mad *sa_mad);
+int mlx4_ib_mcg_demux_handler(struct ib_device *ibdev, int port, int slave,
+			      struct ib_sa_mad *mad);
+
+int mlx4_ib_add_mc(struct mlx4_ib_dev *mdev, struct mlx4_ib_qp *mqp,
+		   union ib_gid *gid);
+
+void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
+			    enum ib_event_type type);
+
+void mlx4_ib_tunnels_update_work(struct work_struct *work);
+
+int mlx4_ib_send_to_slave(struct mlx4_ib_dev *dev, int slave, u8 port,
+			  enum ib_qp_type qpt, struct ib_wc *wc,
+			  struct ib_grh *grh, struct ib_mad *mad);
+
+int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
+			 enum ib_qp_type dest_qpt, u16 pkey_index, u32 remote_qpn,
+			 u32 qkey, struct rdma_ah_attr *attr, u8 *s_mac,
+			 u16 vlan_id, struct ib_mad *mad);
+
+__be64 mlx4_ib_get_new_demux_tid(struct mlx4_ib_demux_ctx *ctx);
+
+int mlx4_ib_demux_cm_handler(struct ib_device *ibdev, int port, int *slave,
+		struct ib_mad *mad);
+
+int mlx4_ib_multiplex_cm_handler(struct ib_device *ibdev, int port, int slave_id,
+		struct ib_mad *mad);
+
+void mlx4_ib_cm_paravirt_init(struct mlx4_ib_dev *dev);
+void mlx4_ib_cm_paravirt_clean(struct mlx4_ib_dev *dev, int slave_id);
+
+/* alias guid support */
+void mlx4_ib_init_alias_guid_work(struct mlx4_ib_dev *dev, int port);
+int mlx4_ib_init_alias_guid_service(struct mlx4_ib_dev *dev);
+void mlx4_ib_destroy_alias_guid_service(struct mlx4_ib_dev *dev);
+void mlx4_ib_invalidate_all_guid_record(struct mlx4_ib_dev *dev, int port);
+
+void mlx4_ib_notify_slaves_on_guid_change(struct mlx4_ib_dev *dev,
+					  int block_num,
+					  u8 port_num, u8 *p_data);
+
+void mlx4_ib_update_cache_on_guid_change(struct mlx4_ib_dev *dev,
+					 int block_num, u8 port_num,
+					 u8 *p_data);
+
+int add_sysfs_port_mcg_attr(struct mlx4_ib_dev *device, int port_num,
+			    struct attribute *attr);
+void del_sysfs_port_mcg_attr(struct mlx4_ib_dev *device, int port_num,
+			     struct attribute *attr);
+ib_sa_comp_mask mlx4_ib_get_aguid_comp_mask_from_ix(int index);
+void mlx4_ib_slave_alias_guid_event(struct mlx4_ib_dev *dev, int slave,
+				    int port, int slave_init);
+
+int mlx4_ib_device_register_sysfs(struct mlx4_ib_dev *device) ;
+
+void mlx4_ib_device_unregister_sysfs(struct mlx4_ib_dev *device);
+
+__be64 mlx4_ib_gen_node_guid(void);
+
+int mlx4_ib_steer_qp_alloc(struct mlx4_ib_dev *dev, int count, int *qpn);
+void mlx4_ib_steer_qp_free(struct mlx4_ib_dev *dev, u32 qpn, int count);
+int mlx4_ib_steer_qp_reg(struct mlx4_ib_dev *mdev, struct mlx4_ib_qp *mqp,
+			 int is_attach);
+int mlx4_ib_rereg_user_mr(struct ib_mr *mr, int flags,
+			  u64 start, u64 length, u64 virt_addr,
+			  int mr_access_flags, struct ib_pd *pd,
+			  struct ib_udata *udata);
+int mlx4_ib_gid_index_to_real_index(struct mlx4_ib_dev *ibdev,
+				    const struct ib_gid_attr *attr);
+
+void mlx4_sched_ib_sl2vl_update_work(struct mlx4_ib_dev *ibdev,
+				     int port);
+
+void mlx4_ib_sl2vl_update(struct mlx4_ib_dev *mdev, int port);
+
+struct ib_wq *mlx4_ib_create_wq(struct ib_pd *pd,
+				struct ib_wq_init_attr *init_attr,
+				struct ib_udata *udata);
+int mlx4_ib_destroy_wq(struct ib_wq *wq, struct ib_udata *udata);
+int mlx4_ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
+		      u32 wq_attr_mask, struct ib_udata *udata);
+
+int mlx4_ib_create_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl,
+				 struct ib_rwq_ind_table_init_attr *init_attr,
+				 struct ib_udata *udata);
+static inline int
+mlx4_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *wq_ind_table)
+{
+	return 0;
+}
+int mlx4_ib_umem_calc_optimal_mtt_size(struct ib_umem *umem, u64 start_va,
+				       int *num_of_mtts);
+
+#endif /* MLX4_IB_H */
diff --git a/drivers/infiniband/hw/mlx4/mr.c b/drivers/infiniband/hw/mlx4/mr.c
new file mode 100644
index 000000000..811b4bb34
--- /dev/null
+++ b/drivers/infiniband/hw/mlx4/mr.c
@@ -0,0 +1,717 @@
+/*
+ * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
+ * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/slab.h>
+#include <rdma/ib_user_verbs.h>
+
+#include "mlx4_ib.h"
+
+static u32 convert_access(int acc)
+{
+	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX4_PERM_ATOMIC       : 0) |
+	       (acc & IB_ACCESS_REMOTE_WRITE  ? MLX4_PERM_REMOTE_WRITE : 0) |
+	       (acc & IB_ACCESS_REMOTE_READ   ? MLX4_PERM_REMOTE_READ  : 0) |
+	       (acc & IB_ACCESS_LOCAL_WRITE   ? MLX4_PERM_LOCAL_WRITE  : 0) |
+	       (acc & IB_ACCESS_MW_BIND	      ? MLX4_PERM_BIND_MW      : 0) |
+	       MLX4_PERM_LOCAL_READ;
+}
+
+static enum mlx4_mw_type to_mlx4_type(enum ib_mw_type type)
+{
+	switch (type) {
+	case IB_MW_TYPE_1:	return MLX4_MW_TYPE_1;
+	case IB_MW_TYPE_2:	return MLX4_MW_TYPE_2;
+	default:		return -1;
+	}
+}
+
+struct ib_mr *mlx4_ib_get_dma_mr(struct ib_pd *pd, int acc)
+{
+	struct mlx4_ib_mr *mr;
+	int err;
+
+	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
+	if (!mr)
+		return ERR_PTR(-ENOMEM);
+
+	err = mlx4_mr_alloc(to_mdev(pd->device)->dev, to_mpd(pd)->pdn, 0,
+			    ~0ull, convert_access(acc), 0, 0, &mr->mmr);
+	if (err)
+		goto err_free;
+
+	err = mlx4_mr_enable(to_mdev(pd->device)->dev, &mr->mmr);
+	if (err)
+		goto err_mr;
+
+	mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
+	mr->umem = NULL;
+
+	return &mr->ibmr;
+
+err_mr:
+	(void) mlx4_mr_free(to_mdev(pd->device)->dev, &mr->mmr);
+
+err_free:
+	kfree(mr);
+
+	return ERR_PTR(err);
+}
+
+enum {
+	MLX4_MAX_MTT_SHIFT = 31
+};
+
+static int mlx4_ib_umem_write_mtt_block(struct mlx4_ib_dev *dev,
+					struct mlx4_mtt *mtt,
+					u64 mtt_size, u64 mtt_shift, u64 len,
+					u64 cur_start_addr, u64 *pages,
+					int *start_index, int *npages)
+{
+	u64 cur_end_addr = cur_start_addr + len;
+	u64 cur_end_addr_aligned = 0;
+	u64 mtt_entries;
+	int err = 0;
+	int k;
+
+	len += (cur_start_addr & (mtt_size - 1ULL));
+	cur_end_addr_aligned = round_up(cur_end_addr, mtt_size);
+	len += (cur_end_addr_aligned - cur_end_addr);
+	if (len & (mtt_size - 1ULL)) {
+		pr_warn("write_block: len %llx is not aligned to mtt_size %llx\n",
+			len, mtt_size);
+		return -EINVAL;
+	}
+
+	mtt_entries = (len >> mtt_shift);
+
+	/*
+	 * Align the MTT start address to the mtt_size.
+	 * Required to handle cases when the MR starts in the middle of an MTT
+	 * record. Was not required in old code since the physical addresses
+	 * provided by the dma subsystem were page aligned, which was also the
+	 * MTT size.
+	 */
+	cur_start_addr = round_down(cur_start_addr, mtt_size);
+	/* A new block is started ... */
+	for (k = 0; k < mtt_entries; ++k) {
+		pages[*npages] = cur_start_addr + (mtt_size * k);
+		(*npages)++;
+		/*
+		 * Be friendly to mlx4_write_mtt() and pass it chunks of
+		 * appropriate size.
+		 */
+		if (*npages == PAGE_SIZE / sizeof(u64)) {
+			err = mlx4_write_mtt(dev->dev, mtt, *start_index,
+					     *npages, pages);
+			if (err)
+				return err;
+
+			(*start_index) += *npages;
+			*npages = 0;
+		}
+	}
+
+	return 0;
+}
+
+static inline u64 alignment_of(u64 ptr)
+{
+	return ilog2(ptr & (~(ptr - 1)));
+}
+
+static int mlx4_ib_umem_calc_block_mtt(u64 next_block_start,
+				       u64 current_block_end,
+				       u64 block_shift)
+{
+	/* Check whether the alignment of the new block is aligned as well as
+	 * the previous block.
+	 * Block address must start with zeros till size of entity_size.
+	 */
+	if ((next_block_start & ((1ULL << block_shift) - 1ULL)) != 0)
+		/*
+		 * It is not as well aligned as the previous block-reduce the
+		 * mtt size accordingly. Here we take the last right bit which
+		 * is 1.
+		 */
+		block_shift = alignment_of(next_block_start);
+
+	/*
+	 * Check whether the alignment of the end of previous block - is it
+	 * aligned as well as the start of the block
+	 */
+	if (((current_block_end) & ((1ULL << block_shift) - 1ULL)) != 0)
+		/*
+		 * It is not as well aligned as the start of the block -
+		 * reduce the mtt size accordingly.
+		 */
+		block_shift = alignment_of(current_block_end);
+
+	return block_shift;
+}
+
+int mlx4_ib_umem_write_mtt(struct mlx4_ib_dev *dev, struct mlx4_mtt *mtt,
+			   struct ib_umem *umem)
+{
+	u64 *pages;
+	u64 len = 0;
+	int err = 0;
+	u64 mtt_size;
+	u64 cur_start_addr = 0;
+	u64 mtt_shift;
+	int start_index = 0;
+	int npages = 0;
+	struct scatterlist *sg;
+	int i;
+
+	pages = (u64 *) __get_free_page(GFP_KERNEL);
+	if (!pages)
+		return -ENOMEM;
+
+	mtt_shift = mtt->page_shift;
+	mtt_size = 1ULL << mtt_shift;
+
+	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) {
+		if (cur_start_addr + len == sg_dma_address(sg)) {
+			/* still the same block */
+			len += sg_dma_len(sg);
+			continue;
+		}
+		/*
+		 * A new block is started ...
+		 * If len is malaligned, write an extra mtt entry to cover the
+		 * misaligned area (round up the division)
+		 */
+		err = mlx4_ib_umem_write_mtt_block(dev, mtt, mtt_size,
+						   mtt_shift, len,
+						   cur_start_addr,
+						   pages, &start_index,
+						   &npages);
+		if (err)
+			goto out;
+
+		cur_start_addr = sg_dma_address(sg);
+		len = sg_dma_len(sg);
+	}
+
+	/* Handle the last block */
+	if (len > 0) {
+		/*
+		 * If len is malaligned, write an extra mtt entry to cover
+		 * the misaligned area (round up the division)
+		 */
+		err = mlx4_ib_umem_write_mtt_block(dev, mtt, mtt_size,
+						   mtt_shift, len,
+						   cur_start_addr, pages,
+						   &start_index, &npages);
+		if (err)
+			goto out;
+	}
+
+	if (npages)
+		err = mlx4_write_mtt(dev->dev, mtt, start_index, npages, pages);
+
+out:
+	free_page((unsigned long) pages);
+	return err;
+}
+
+/*
+ * Calculate optimal mtt size based on contiguous pages.
+ * Function will return also the number of pages that are not aligned to the
+ * calculated mtt_size to be added to total number of pages. For that we should
+ * check the first chunk length & last chunk length and if not aligned to
+ * mtt_size we should increment the non_aligned_pages number. All chunks in the
+ * middle already handled as part of mtt shift calculation for both their start
+ * & end addresses.
+ */
+int mlx4_ib_umem_calc_optimal_mtt_size(struct ib_umem *umem, u64 start_va,
+				       int *num_of_mtts)
+{
+	u64 block_shift = MLX4_MAX_MTT_SHIFT;
+	u64 min_shift = PAGE_SHIFT;
+	u64 last_block_aligned_end = 0;
+	u64 current_block_start = 0;
+	u64 first_block_start = 0;
+	u64 current_block_len = 0;
+	u64 last_block_end = 0;
+	struct scatterlist *sg;
+	u64 current_block_end;
+	u64 misalignment_bits;
+	u64 next_block_start;
+	u64 total_len = 0;
+	int i;
+
+	*num_of_mtts = ib_umem_num_dma_blocks(umem, PAGE_SIZE);
+
+	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) {
+		/*
+		 * Initialization - save the first chunk start as the
+		 * current_block_start - block means contiguous pages.
+		 */
+		if (current_block_len == 0 && current_block_start == 0) {
+			current_block_start = sg_dma_address(sg);
+			first_block_start = current_block_start;
+			/*
+			 * Find the bits that are different between the physical
+			 * address and the virtual address for the start of the
+			 * MR.
+			 * umem_get aligned the start_va to a page boundary.
+			 * Therefore, we need to align the start va to the same
+			 * boundary.
+			 * misalignment_bits is needed to handle the  case of a
+			 * single memory region. In this case, the rest of the
+			 * logic will not reduce the block size.  If we use a
+			 * block size which is bigger than the alignment of the
+			 * misalignment bits, we might use the virtual page
+			 * number instead of the physical page number, resulting
+			 * in access to the wrong data.
+			 */
+			misalignment_bits =
+				(start_va & (~(((u64)(PAGE_SIZE)) - 1ULL))) ^
+				current_block_start;
+			block_shift = min(alignment_of(misalignment_bits),
+					  block_shift);
+		}
+
+		/*
+		 * Go over the scatter entries and check if they continue the
+		 * previous scatter entry.
+		 */
+		next_block_start = sg_dma_address(sg);
+		current_block_end = current_block_start	+ current_block_len;
+		/* If we have a split (non-contig.) between two blocks */
+		if (current_block_end != next_block_start) {
+			block_shift = mlx4_ib_umem_calc_block_mtt
+					(next_block_start,
+					 current_block_end,
+					 block_shift);
+
+			/*
+			 * If we reached the minimum shift for 4k page we stop
+			 * the loop.
+			 */
+			if (block_shift <= min_shift)
+				goto end;
+
+			/*
+			 * If not saved yet we are in first block - we save the
+			 * length of first block to calculate the
+			 * non_aligned_pages number at the end.
+			 */
+			total_len += current_block_len;
+
+			/* Start a new block */
+			current_block_start = next_block_start;
+			current_block_len = sg_dma_len(sg);
+			continue;
+		}
+		/* The scatter entry is another part of the current block,
+		 * increase the block size.
+		 * An entry in the scatter can be larger than 4k (page) as of
+		 * dma mapping which merge some blocks together.
+		 */
+		current_block_len += sg_dma_len(sg);
+	}
+
+	/* Account for the last block in the total len */
+	total_len += current_block_len;
+	/* Add to the first block the misalignment that it suffers from. */
+	total_len += (first_block_start & ((1ULL << block_shift) - 1ULL));
+	last_block_end = current_block_start + current_block_len;
+	last_block_aligned_end = round_up(last_block_end, 1ULL << block_shift);
+	total_len += (last_block_aligned_end - last_block_end);
+
+	if (total_len & ((1ULL << block_shift) - 1ULL))
+		pr_warn("misaligned total length detected (%llu, %llu)!",
+			total_len, block_shift);
+
+	*num_of_mtts = total_len >> block_shift;
+end:
+	if (block_shift < min_shift) {
+		/*
+		 * If shift is less than the min we set a warning and return the
+		 * min shift.
+		 */
+		pr_warn("umem_calc_optimal_mtt_size - unexpected shift %lld\n", block_shift);
+
+		block_shift = min_shift;
+	}
+	return block_shift;
+}
+
+static struct ib_umem *mlx4_get_umem_mr(struct ib_device *device, u64 start,
+					u64 length, int access_flags)
+{
+	/*
+	 * Force registering the memory as writable if the underlying pages
+	 * are writable.  This is so rereg can change the access permissions
+	 * from readable to writable without having to run through ib_umem_get
+	 * again
+	 */
+	if (!ib_access_writable(access_flags)) {
+		unsigned long untagged_start = untagged_addr(start);
+		struct vm_area_struct *vma;
+
+		mmap_read_lock(current->mm);
+		/*
+		 * FIXME: Ideally this would iterate over all the vmas that
+		 * cover the memory, but for now it requires a single vma to
+		 * entirely cover the MR to support RO mappings.
+		 */
+		vma = find_vma(current->mm, untagged_start);
+		if (vma && vma->vm_end >= untagged_start + length &&
+		    vma->vm_start <= untagged_start) {
+			if (vma->vm_flags & VM_WRITE)
+				access_flags |= IB_ACCESS_LOCAL_WRITE;
+		} else {
+			access_flags |= IB_ACCESS_LOCAL_WRITE;
+		}
+
+		mmap_read_unlock(current->mm);
+	}
+
+	return ib_umem_get(device, start, length, access_flags);
+}
+
+struct ib_mr *mlx4_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
+				  u64 virt_addr, int access_flags,
+				  struct ib_udata *udata)
+{
+	struct mlx4_ib_dev *dev = to_mdev(pd->device);
+	struct mlx4_ib_mr *mr;
+	int shift;
+	int err;
+	int n;
+
+	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
+	if (!mr)
+		return ERR_PTR(-ENOMEM);
+
+	mr->umem = mlx4_get_umem_mr(pd->device, start, length, access_flags);
+	if (IS_ERR(mr->umem)) {
+		err = PTR_ERR(mr->umem);
+		goto err_free;
+	}
+
+	shift = mlx4_ib_umem_calc_optimal_mtt_size(mr->umem, start, &n);
+
+	err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, virt_addr, length,
+			    convert_access(access_flags), n, shift, &mr->mmr);
+	if (err)
+		goto err_umem;
+
+	err = mlx4_ib_umem_write_mtt(dev, &mr->mmr.mtt, mr->umem);
+	if (err)
+		goto err_mr;
+
+	err = mlx4_mr_enable(dev->dev, &mr->mmr);
+	if (err)
+		goto err_mr;
+
+	mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
+	mr->ibmr.page_size = 1U << shift;
+
+	return &mr->ibmr;
+
+err_mr:
+	(void) mlx4_mr_free(to_mdev(pd->device)->dev, &mr->mmr);
+
+err_umem:
+	ib_umem_release(mr->umem);
+
+err_free:
+	kfree(mr);
+
+	return ERR_PTR(err);
+}
+
+int mlx4_ib_rereg_user_mr(struct ib_mr *mr, int flags,
+			  u64 start, u64 length, u64 virt_addr,
+			  int mr_access_flags, struct ib_pd *pd,
+			  struct ib_udata *udata)
+{
+	struct mlx4_ib_dev *dev = to_mdev(mr->device);
+	struct mlx4_ib_mr *mmr = to_mmr(mr);
+	struct mlx4_mpt_entry *mpt_entry;
+	struct mlx4_mpt_entry **pmpt_entry = &mpt_entry;
+	int err;
+
+	/* Since we synchronize this call and mlx4_ib_dereg_mr via uverbs,
+	 * we assume that the calls can't run concurrently. Otherwise, a
+	 * race exists.
+	 */
+	err =  mlx4_mr_hw_get_mpt(dev->dev, &mmr->mmr, &pmpt_entry);
+
+	if (err)
+		return err;
+
+	if (flags & IB_MR_REREG_PD) {
+		err = mlx4_mr_hw_change_pd(dev->dev, *pmpt_entry,
+					   to_mpd(pd)->pdn);
+
+		if (err)
+			goto release_mpt_entry;
+	}
+
+	if (flags & IB_MR_REREG_ACCESS) {
+		if (ib_access_writable(mr_access_flags) &&
+		    !mmr->umem->writable) {
+			err = -EPERM;
+			goto release_mpt_entry;
+		}
+
+		err = mlx4_mr_hw_change_access(dev->dev, *pmpt_entry,
+					       convert_access(mr_access_flags));
+
+		if (err)
+			goto release_mpt_entry;
+	}
+
+	if (flags & IB_MR_REREG_TRANS) {
+		int shift;
+		int n;
+
+		mlx4_mr_rereg_mem_cleanup(dev->dev, &mmr->mmr);
+		ib_umem_release(mmr->umem);
+		mmr->umem = mlx4_get_umem_mr(mr->device, start, length,
+					     mr_access_flags);
+		if (IS_ERR(mmr->umem)) {
+			err = PTR_ERR(mmr->umem);
+			/* Prevent mlx4_ib_dereg_mr from free'ing invalid pointer */
+			mmr->umem = NULL;
+			goto release_mpt_entry;
+		}
+		n = ib_umem_num_dma_blocks(mmr->umem, PAGE_SIZE);
+		shift = PAGE_SHIFT;
+
+		err = mlx4_mr_rereg_mem_write(dev->dev, &mmr->mmr,
+					      virt_addr, length, n, shift,
+					      *pmpt_entry);
+		if (err) {
+			ib_umem_release(mmr->umem);
+			goto release_mpt_entry;
+		}
+		mmr->mmr.iova       = virt_addr;
+		mmr->mmr.size       = length;
+
+		err = mlx4_ib_umem_write_mtt(dev, &mmr->mmr.mtt, mmr->umem);
+		if (err) {
+			mlx4_mr_rereg_mem_cleanup(dev->dev, &mmr->mmr);
+			ib_umem_release(mmr->umem);
+			goto release_mpt_entry;
+		}
+	}
+
+	/* If we couldn't transfer the MR to the HCA, just remember to
+	 * return a failure. But dereg_mr will free the resources.
+	 */
+	err = mlx4_mr_hw_write_mpt(dev->dev, &mmr->mmr, pmpt_entry);
+	if (!err && flags & IB_MR_REREG_ACCESS)
+		mmr->mmr.access = mr_access_flags;
+
+release_mpt_entry:
+	mlx4_mr_hw_put_mpt(dev->dev, pmpt_entry);
+
+	return err;
+}
+
+static int
+mlx4_alloc_priv_pages(struct ib_device *device,
+		      struct mlx4_ib_mr *mr,
+		      int max_pages)
+{
+	int ret;
+
+	/* Ensure that size is aligned to DMA cacheline
+	 * requirements.
+	 * max_pages is limited to MLX4_MAX_FAST_REG_PAGES
+	 * so page_map_size will never cross PAGE_SIZE.
+	 */
+	mr->page_map_size = roundup(max_pages * sizeof(u64),
+				    MLX4_MR_PAGES_ALIGN);
+
+	/* Prevent cross page boundary allocation. */
+	mr->pages = (__be64 *)get_zeroed_page(GFP_KERNEL);
+	if (!mr->pages)
+		return -ENOMEM;
+
+	mr->page_map = dma_map_single(device->dev.parent, mr->pages,
+				      mr->page_map_size, DMA_TO_DEVICE);
+
+	if (dma_mapping_error(device->dev.parent, mr->page_map)) {
+		ret = -ENOMEM;
+		goto err;
+	}
+
+	return 0;
+
+err:
+	free_page((unsigned long)mr->pages);
+	return ret;
+}
+
+static void
+mlx4_free_priv_pages(struct mlx4_ib_mr *mr)
+{
+	if (mr->pages) {
+		struct ib_device *device = mr->ibmr.device;
+
+		dma_unmap_single(device->dev.parent, mr->page_map,
+				 mr->page_map_size, DMA_TO_DEVICE);
+		free_page((unsigned long)mr->pages);
+		mr->pages = NULL;
+	}
+}
+
+int mlx4_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
+{
+	struct mlx4_ib_mr *mr = to_mmr(ibmr);
+	int ret;
+
+	mlx4_free_priv_pages(mr);
+
+	ret = mlx4_mr_free(to_mdev(ibmr->device)->dev, &mr->mmr);
+	if (ret)
+		return ret;
+	if (mr->umem)
+		ib_umem_release(mr->umem);
+	kfree(mr);
+
+	return 0;
+}
+
+int mlx4_ib_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata)
+{
+	struct mlx4_ib_dev *dev = to_mdev(ibmw->device);
+	struct mlx4_ib_mw *mw = to_mmw(ibmw);
+	int err;
+
+	err = mlx4_mw_alloc(dev->dev, to_mpd(ibmw->pd)->pdn,
+			    to_mlx4_type(ibmw->type), &mw->mmw);
+	if (err)
+		return err;
+
+	err = mlx4_mw_enable(dev->dev, &mw->mmw);
+	if (err)
+		goto err_mw;
+
+	ibmw->rkey = mw->mmw.key;
+	return 0;
+
+err_mw:
+	mlx4_mw_free(dev->dev, &mw->mmw);
+	return err;
+}
+
+int mlx4_ib_dealloc_mw(struct ib_mw *ibmw)
+{
+	struct mlx4_ib_mw *mw = to_mmw(ibmw);
+
+	mlx4_mw_free(to_mdev(ibmw->device)->dev, &mw->mmw);
+	return 0;
+}
+
+struct ib_mr *mlx4_ib_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
+			       u32 max_num_sg)
+{
+	struct mlx4_ib_dev *dev = to_mdev(pd->device);
+	struct mlx4_ib_mr *mr;
+	int err;
+
+	if (mr_type != IB_MR_TYPE_MEM_REG ||
+	    max_num_sg > MLX4_MAX_FAST_REG_PAGES)
+		return ERR_PTR(-EINVAL);
+
+	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
+	if (!mr)
+		return ERR_PTR(-ENOMEM);
+
+	err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, 0, 0, 0,
+			    max_num_sg, 0, &mr->mmr);
+	if (err)
+		goto err_free;
+
+	err = mlx4_alloc_priv_pages(pd->device, mr, max_num_sg);
+	if (err)
+		goto err_free_mr;
+
+	mr->max_pages = max_num_sg;
+	err = mlx4_mr_enable(dev->dev, &mr->mmr);
+	if (err)
+		goto err_free_pl;
+
+	mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
+	mr->umem = NULL;
+
+	return &mr->ibmr;
+
+err_free_pl:
+	mr->ibmr.device = pd->device;
+	mlx4_free_priv_pages(mr);
+err_free_mr:
+	(void) mlx4_mr_free(dev->dev, &mr->mmr);
+err_free:
+	kfree(mr);
+	return ERR_PTR(err);
+}
+
+static int mlx4_set_page(struct ib_mr *ibmr, u64 addr)
+{
+	struct mlx4_ib_mr *mr = to_mmr(ibmr);
+
+	if (unlikely(mr->npages == mr->max_pages))
+		return -ENOMEM;
+
+	mr->pages[mr->npages++] = cpu_to_be64(addr | MLX4_MTT_FLAG_PRESENT);
+
+	return 0;
+}
+
+int mlx4_ib_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
+		      unsigned int *sg_offset)
+{
+	struct mlx4_ib_mr *mr = to_mmr(ibmr);
+	int rc;
+
+	mr->npages = 0;
+
+	ib_dma_sync_single_for_cpu(ibmr->device, mr->page_map,
+				   mr->page_map_size, DMA_TO_DEVICE);
+
+	rc = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, mlx4_set_page);
+
+	ib_dma_sync_single_for_device(ibmr->device, mr->page_map,
+				      mr->page_map_size, DMA_TO_DEVICE);
+
+	return rc;
+}
diff --git a/drivers/infiniband/hw/mlx4/qp.c b/drivers/infiniband/hw/mlx4/qp.c
new file mode 100644
index 000000000..50b355e34
--- /dev/null
+++ b/drivers/infiniband/hw/mlx4/qp.c
@@ -0,0 +1,4474 @@
+/*
+ * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
+ * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/log2.h>
+#include <linux/etherdevice.h>
+#include <net/ip.h>
+#include <linux/slab.h>
+#include <linux/netdevice.h>
+
+#include <rdma/ib_cache.h>
+#include <rdma/ib_pack.h>
+#include <rdma/ib_addr.h>
+#include <rdma/ib_mad.h>
+#include <rdma/uverbs_ioctl.h>
+
+#include <linux/mlx4/driver.h>
+#include <linux/mlx4/qp.h>
+
+#include "mlx4_ib.h"
+#include <rdma/mlx4-abi.h>
+
+static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq,
+			     struct mlx4_ib_cq *recv_cq);
+static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq,
+			       struct mlx4_ib_cq *recv_cq);
+static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state,
+			      struct ib_udata *udata);
+
+enum {
+	MLX4_IB_ACK_REQ_FREQ	= 8,
+};
+
+enum {
+	MLX4_IB_DEFAULT_SCHED_QUEUE	= 0x83,
+	MLX4_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
+	MLX4_IB_LINK_TYPE_IB		= 0,
+	MLX4_IB_LINK_TYPE_ETH		= 1
+};
+
+enum {
+	MLX4_IB_MIN_SQ_STRIDE	= 6,
+	MLX4_IB_CACHE_LINE_SIZE	= 64,
+};
+
+enum {
+	MLX4_RAW_QP_MTU		= 7,
+	MLX4_RAW_QP_MSGMAX	= 31,
+};
+
+#ifndef ETH_ALEN
+#define ETH_ALEN        6
+#endif
+
+static const __be32 mlx4_ib_opcode[] = {
+	[IB_WR_SEND]				= cpu_to_be32(MLX4_OPCODE_SEND),
+	[IB_WR_LSO]				= cpu_to_be32(MLX4_OPCODE_LSO),
+	[IB_WR_SEND_WITH_IMM]			= cpu_to_be32(MLX4_OPCODE_SEND_IMM),
+	[IB_WR_RDMA_WRITE]			= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
+	[IB_WR_RDMA_WRITE_WITH_IMM]		= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
+	[IB_WR_RDMA_READ]			= cpu_to_be32(MLX4_OPCODE_RDMA_READ),
+	[IB_WR_ATOMIC_CMP_AND_SWP]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
+	[IB_WR_ATOMIC_FETCH_AND_ADD]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
+	[IB_WR_SEND_WITH_INV]			= cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
+	[IB_WR_LOCAL_INV]			= cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
+	[IB_WR_REG_MR]				= cpu_to_be32(MLX4_OPCODE_FMR),
+	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
+	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
+};
+
+enum mlx4_ib_source_type {
+	MLX4_IB_QP_SRC	= 0,
+	MLX4_IB_RWQ_SRC	= 1,
+};
+
+static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
+{
+	if (!mlx4_is_master(dev->dev))
+		return 0;
+
+	return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
+	       qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
+		8 * MLX4_MFUNC_MAX;
+}
+
+static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
+{
+	int proxy_sqp = 0;
+	int real_sqp = 0;
+	int i;
+	/* PPF or Native -- real SQP */
+	real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
+		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
+		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
+	if (real_sqp)
+		return 1;
+	/* VF or PF -- proxy SQP */
+	if (mlx4_is_mfunc(dev->dev)) {
+		for (i = 0; i < dev->dev->caps.num_ports; i++) {
+			if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy ||
+			    qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp1_proxy) {
+				proxy_sqp = 1;
+				break;
+			}
+		}
+	}
+	if (proxy_sqp)
+		return 1;
+
+	return !!(qp->flags & MLX4_IB_ROCE_V2_GSI_QP);
+}
+
+/* used for INIT/CLOSE port logic */
+static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
+{
+	int proxy_qp0 = 0;
+	int real_qp0 = 0;
+	int i;
+	/* PPF or Native -- real QP0 */
+	real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
+		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
+		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
+	if (real_qp0)
+		return 1;
+	/* VF or PF -- proxy QP0 */
+	if (mlx4_is_mfunc(dev->dev)) {
+		for (i = 0; i < dev->dev->caps.num_ports; i++) {
+			if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy) {
+				proxy_qp0 = 1;
+				break;
+			}
+		}
+	}
+	return proxy_qp0;
+}
+
+static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
+{
+	return mlx4_buf_offset(&qp->buf, offset);
+}
+
+static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
+{
+	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
+}
+
+static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
+{
+	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
+}
+
+/*
+ * Stamp a SQ WQE so that it is invalid if prefetched by marking the
+ * first four bytes of every 64 byte chunk with 0xffffffff, except for
+ * the very first chunk of the WQE.
+ */
+static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n)
+{
+	__be32 *wqe;
+	int i;
+	int s;
+	void *buf;
+	struct mlx4_wqe_ctrl_seg *ctrl;
+
+	buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
+	ctrl = (struct mlx4_wqe_ctrl_seg *)buf;
+	s = (ctrl->qpn_vlan.fence_size & 0x3f) << 4;
+	for (i = 64; i < s; i += 64) {
+		wqe = buf + i;
+		*wqe = cpu_to_be32(0xffffffff);
+	}
+}
+
+static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
+{
+	struct ib_event event;
+	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
+
+	if (type == MLX4_EVENT_TYPE_PATH_MIG)
+		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
+
+	if (ibqp->event_handler) {
+		event.device     = ibqp->device;
+		event.element.qp = ibqp;
+		switch (type) {
+		case MLX4_EVENT_TYPE_PATH_MIG:
+			event.event = IB_EVENT_PATH_MIG;
+			break;
+		case MLX4_EVENT_TYPE_COMM_EST:
+			event.event = IB_EVENT_COMM_EST;
+			break;
+		case MLX4_EVENT_TYPE_SQ_DRAINED:
+			event.event = IB_EVENT_SQ_DRAINED;
+			break;
+		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
+			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
+			break;
+		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
+			event.event = IB_EVENT_QP_FATAL;
+			break;
+		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
+			event.event = IB_EVENT_PATH_MIG_ERR;
+			break;
+		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
+			event.event = IB_EVENT_QP_REQ_ERR;
+			break;
+		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
+			event.event = IB_EVENT_QP_ACCESS_ERR;
+			break;
+		default:
+			pr_warn("Unexpected event type %d "
+			       "on QP %06x\n", type, qp->qpn);
+			return;
+		}
+
+		ibqp->event_handler(&event, ibqp->qp_context);
+	}
+}
+
+static void mlx4_ib_wq_event(struct mlx4_qp *qp, enum mlx4_event type)
+{
+	pr_warn_ratelimited("Unexpected event type %d on WQ 0x%06x. Events are not supported for WQs\n",
+			    type, qp->qpn);
+}
+
+static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
+{
+	/*
+	 * UD WQEs must have a datagram segment.
+	 * RC and UC WQEs might have a remote address segment.
+	 * MLX WQEs need two extra inline data segments (for the UD
+	 * header and space for the ICRC).
+	 */
+	switch (type) {
+	case MLX4_IB_QPT_UD:
+		return sizeof (struct mlx4_wqe_ctrl_seg) +
+			sizeof (struct mlx4_wqe_datagram_seg) +
+			((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
+	case MLX4_IB_QPT_PROXY_SMI_OWNER:
+	case MLX4_IB_QPT_PROXY_SMI:
+	case MLX4_IB_QPT_PROXY_GSI:
+		return sizeof (struct mlx4_wqe_ctrl_seg) +
+			sizeof (struct mlx4_wqe_datagram_seg) + 64;
+	case MLX4_IB_QPT_TUN_SMI_OWNER:
+	case MLX4_IB_QPT_TUN_GSI:
+		return sizeof (struct mlx4_wqe_ctrl_seg) +
+			sizeof (struct mlx4_wqe_datagram_seg);
+
+	case MLX4_IB_QPT_UC:
+		return sizeof (struct mlx4_wqe_ctrl_seg) +
+			sizeof (struct mlx4_wqe_raddr_seg);
+	case MLX4_IB_QPT_RC:
+		return sizeof (struct mlx4_wqe_ctrl_seg) +
+			sizeof (struct mlx4_wqe_masked_atomic_seg) +
+			sizeof (struct mlx4_wqe_raddr_seg);
+	case MLX4_IB_QPT_SMI:
+	case MLX4_IB_QPT_GSI:
+		return sizeof (struct mlx4_wqe_ctrl_seg) +
+			ALIGN(MLX4_IB_UD_HEADER_SIZE +
+			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
+					   MLX4_INLINE_ALIGN) *
+			      sizeof (struct mlx4_wqe_inline_seg),
+			      sizeof (struct mlx4_wqe_data_seg)) +
+			ALIGN(4 +
+			      sizeof (struct mlx4_wqe_inline_seg),
+			      sizeof (struct mlx4_wqe_data_seg));
+	default:
+		return sizeof (struct mlx4_wqe_ctrl_seg);
+	}
+}
+
+static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
+		       bool is_user, bool has_rq, struct mlx4_ib_qp *qp,
+		       u32 inl_recv_sz)
+{
+	/* Sanity check RQ size before proceeding */
+	if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
+	    cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
+		return -EINVAL;
+
+	if (!has_rq) {
+		if (cap->max_recv_wr || inl_recv_sz)
+			return -EINVAL;
+
+		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
+	} else {
+		u32 max_inl_recv_sz = dev->dev->caps.max_rq_sg *
+			sizeof(struct mlx4_wqe_data_seg);
+		u32 wqe_size;
+
+		/* HW requires >= 1 RQ entry with >= 1 gather entry */
+		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge ||
+				inl_recv_sz > max_inl_recv_sz))
+			return -EINVAL;
+
+		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
+		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
+		wqe_size = qp->rq.max_gs * sizeof(struct mlx4_wqe_data_seg);
+		qp->rq.wqe_shift = ilog2(max_t(u32, wqe_size, inl_recv_sz));
+	}
+
+	/* leave userspace return values as they were, so as not to break ABI */
+	if (is_user) {
+		cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
+		cap->max_recv_sge = qp->rq.max_gs;
+	} else {
+		cap->max_recv_wr  = qp->rq.max_post =
+			min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
+		cap->max_recv_sge = min(qp->rq.max_gs,
+					min(dev->dev->caps.max_sq_sg,
+					    dev->dev->caps.max_rq_sg));
+	}
+
+	return 0;
+}
+
+static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
+			      enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp)
+{
+	int s;
+
+	/* Sanity check SQ size before proceeding */
+	if (cap->max_send_wr  > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
+	    cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
+	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
+	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
+		return -EINVAL;
+
+	/*
+	 * For MLX transport we need 2 extra S/G entries:
+	 * one for the header and one for the checksum at the end
+	 */
+	if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
+	     type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
+	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
+		return -EINVAL;
+
+	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
+		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
+		send_wqe_overhead(type, qp->flags);
+
+	if (s > dev->dev->caps.max_sq_desc_sz)
+		return -EINVAL;
+
+	qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
+
+	/*
+	 * We need to leave 2 KB + 1 WR of headroom in the SQ to
+	 * allow HW to prefetch.
+	 */
+	qp->sq_spare_wqes = MLX4_IB_SQ_HEADROOM(qp->sq.wqe_shift);
+	qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr +
+					    qp->sq_spare_wqes);
+
+	qp->sq.max_gs =
+		(min(dev->dev->caps.max_sq_desc_sz,
+		     (1 << qp->sq.wqe_shift)) -
+		 send_wqe_overhead(type, qp->flags)) /
+		sizeof (struct mlx4_wqe_data_seg);
+
+	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
+		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
+	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
+		qp->rq.offset = 0;
+		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
+	} else {
+		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
+		qp->sq.offset = 0;
+	}
+
+	cap->max_send_wr  = qp->sq.max_post =
+		qp->sq.wqe_cnt - qp->sq_spare_wqes;
+	cap->max_send_sge = min(qp->sq.max_gs,
+				min(dev->dev->caps.max_sq_sg,
+				    dev->dev->caps.max_rq_sg));
+	/* We don't support inline sends for kernel QPs (yet) */
+	cap->max_inline_data = 0;
+
+	return 0;
+}
+
+static int set_user_sq_size(struct mlx4_ib_dev *dev,
+			    struct mlx4_ib_qp *qp,
+			    struct mlx4_ib_create_qp *ucmd)
+{
+	u32 cnt;
+
+	/* Sanity check SQ size before proceeding */
+	if (check_shl_overflow(1, ucmd->log_sq_bb_count, &cnt) ||
+	    cnt > dev->dev->caps.max_wqes)
+		return -EINVAL;
+	if (ucmd->log_sq_stride >
+		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
+	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
+		return -EINVAL;
+
+	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
+	qp->sq.wqe_shift = ucmd->log_sq_stride;
+
+	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
+		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
+
+	return 0;
+}
+
+static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
+{
+	int i;
+
+	qp->sqp_proxy_rcv =
+		kmalloc_array(qp->rq.wqe_cnt, sizeof(struct mlx4_ib_buf),
+			      GFP_KERNEL);
+	if (!qp->sqp_proxy_rcv)
+		return -ENOMEM;
+	for (i = 0; i < qp->rq.wqe_cnt; i++) {
+		qp->sqp_proxy_rcv[i].addr =
+			kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
+				GFP_KERNEL);
+		if (!qp->sqp_proxy_rcv[i].addr)
+			goto err;
+		qp->sqp_proxy_rcv[i].map =
+			ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
+					  sizeof (struct mlx4_ib_proxy_sqp_hdr),
+					  DMA_FROM_DEVICE);
+		if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) {
+			kfree(qp->sqp_proxy_rcv[i].addr);
+			goto err;
+		}
+	}
+	return 0;
+
+err:
+	while (i > 0) {
+		--i;
+		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
+				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
+				    DMA_FROM_DEVICE);
+		kfree(qp->sqp_proxy_rcv[i].addr);
+	}
+	kfree(qp->sqp_proxy_rcv);
+	qp->sqp_proxy_rcv = NULL;
+	return -ENOMEM;
+}
+
+static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
+{
+	int i;
+
+	for (i = 0; i < qp->rq.wqe_cnt; i++) {
+		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
+				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
+				    DMA_FROM_DEVICE);
+		kfree(qp->sqp_proxy_rcv[i].addr);
+	}
+	kfree(qp->sqp_proxy_rcv);
+}
+
+static bool qp_has_rq(struct ib_qp_init_attr *attr)
+{
+	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
+		return false;
+
+	return !attr->srq;
+}
+
+static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
+{
+	int i;
+	for (i = 0; i < dev->caps.num_ports; i++) {
+		if (qpn == dev->caps.spec_qps[i].qp0_proxy)
+			return !!dev->caps.spec_qps[i].qp0_qkey;
+	}
+	return 0;
+}
+
+static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev,
+				    struct mlx4_ib_qp *qp)
+{
+	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
+	mlx4_counter_free(dev->dev, qp->counter_index->index);
+	list_del(&qp->counter_index->list);
+	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
+
+	kfree(qp->counter_index);
+	qp->counter_index = NULL;
+}
+
+static int set_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_rss *rss_ctx,
+		      struct ib_qp_init_attr *init_attr,
+		      struct mlx4_ib_create_qp_rss *ucmd)
+{
+	rss_ctx->base_qpn_tbl_sz = init_attr->rwq_ind_tbl->ind_tbl[0]->wq_num |
+		(init_attr->rwq_ind_tbl->log_ind_tbl_size << 24);
+
+	if ((ucmd->rx_hash_function == MLX4_IB_RX_HASH_FUNC_TOEPLITZ) &&
+	    (dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RSS_TOP)) {
+		memcpy(rss_ctx->rss_key, ucmd->rx_hash_key,
+		       MLX4_EN_RSS_KEY_SIZE);
+	} else {
+		pr_debug("RX Hash function is not supported\n");
+		return (-EOPNOTSUPP);
+	}
+
+	if (ucmd->rx_hash_fields_mask & ~(u64)(MLX4_IB_RX_HASH_SRC_IPV4	|
+					       MLX4_IB_RX_HASH_DST_IPV4	|
+					       MLX4_IB_RX_HASH_SRC_IPV6	|
+					       MLX4_IB_RX_HASH_DST_IPV6	|
+					       MLX4_IB_RX_HASH_SRC_PORT_TCP |
+					       MLX4_IB_RX_HASH_DST_PORT_TCP |
+					       MLX4_IB_RX_HASH_SRC_PORT_UDP |
+					       MLX4_IB_RX_HASH_DST_PORT_UDP |
+					       MLX4_IB_RX_HASH_INNER)) {
+		pr_debug("RX Hash fields_mask has unsupported mask (0x%llx)\n",
+			 ucmd->rx_hash_fields_mask);
+		return (-EOPNOTSUPP);
+	}
+
+	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) &&
+	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
+		rss_ctx->flags = MLX4_RSS_IPV4;
+	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) ||
+		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
+		pr_debug("RX Hash fields_mask is not supported - both IPv4 SRC and DST must be set\n");
+		return (-EOPNOTSUPP);
+	}
+
+	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) &&
+	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
+		rss_ctx->flags |= MLX4_RSS_IPV6;
+	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) ||
+		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
+		pr_debug("RX Hash fields_mask is not supported - both IPv6 SRC and DST must be set\n");
+		return (-EOPNOTSUPP);
+	}
+
+	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) &&
+	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
+		if (!(dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_UDP_RSS)) {
+			pr_debug("RX Hash fields_mask for UDP is not supported\n");
+			return (-EOPNOTSUPP);
+		}
+
+		if (rss_ctx->flags & MLX4_RSS_IPV4)
+			rss_ctx->flags |= MLX4_RSS_UDP_IPV4;
+		if (rss_ctx->flags & MLX4_RSS_IPV6)
+			rss_ctx->flags |= MLX4_RSS_UDP_IPV6;
+		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
+			pr_debug("RX Hash fields_mask is not supported - UDP must be set with IPv4 or IPv6\n");
+			return (-EOPNOTSUPP);
+		}
+	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) ||
+		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
+		pr_debug("RX Hash fields_mask is not supported - both UDP SRC and DST must be set\n");
+		return (-EOPNOTSUPP);
+	}
+
+	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) &&
+	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
+		if (rss_ctx->flags & MLX4_RSS_IPV4)
+			rss_ctx->flags |= MLX4_RSS_TCP_IPV4;
+		if (rss_ctx->flags & MLX4_RSS_IPV6)
+			rss_ctx->flags |= MLX4_RSS_TCP_IPV6;
+		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
+			pr_debug("RX Hash fields_mask is not supported - TCP must be set with IPv4 or IPv6\n");
+			return (-EOPNOTSUPP);
+		}
+	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) ||
+		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
+		pr_debug("RX Hash fields_mask is not supported - both TCP SRC and DST must be set\n");
+		return (-EOPNOTSUPP);
+	}
+
+	if (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_INNER) {
+		if (dev->dev->caps.tunnel_offload_mode ==
+		    MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
+			/*
+			 * Hash according to inner headers if exist, otherwise
+			 * according to outer headers.
+			 */
+			rss_ctx->flags |= MLX4_RSS_BY_INNER_HEADERS_IPONLY;
+		} else {
+			pr_debug("RSS Hash for inner headers isn't supported\n");
+			return (-EOPNOTSUPP);
+		}
+	}
+
+	return 0;
+}
+
+static int create_qp_rss(struct mlx4_ib_dev *dev,
+			 struct ib_qp_init_attr *init_attr,
+			 struct mlx4_ib_create_qp_rss *ucmd,
+			 struct mlx4_ib_qp *qp)
+{
+	int qpn;
+	int err;
+
+	qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
+
+	err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn, 0, qp->mqp.usage);
+	if (err)
+		return err;
+
+	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
+	if (err)
+		goto err_qpn;
+
+	INIT_LIST_HEAD(&qp->gid_list);
+	INIT_LIST_HEAD(&qp->steering_rules);
+
+	qp->mlx4_ib_qp_type = MLX4_IB_QPT_RAW_PACKET;
+	qp->state = IB_QPS_RESET;
+
+	/* Set dummy send resources to be compatible with HV and PRM */
+	qp->sq_no_prefetch = 1;
+	qp->sq.wqe_cnt = 1;
+	qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
+	qp->buf_size = qp->sq.wqe_cnt << MLX4_IB_MIN_SQ_STRIDE;
+	qp->mtt = (to_mqp(
+		   (struct ib_qp *)init_attr->rwq_ind_tbl->ind_tbl[0]))->mtt;
+
+	qp->rss_ctx = kzalloc(sizeof(*qp->rss_ctx), GFP_KERNEL);
+	if (!qp->rss_ctx) {
+		err = -ENOMEM;
+		goto err_qp_alloc;
+	}
+
+	err = set_qp_rss(dev, qp->rss_ctx, init_attr, ucmd);
+	if (err)
+		goto err;
+
+	return 0;
+
+err:
+	kfree(qp->rss_ctx);
+
+err_qp_alloc:
+	mlx4_qp_remove(dev->dev, &qp->mqp);
+	mlx4_qp_free(dev->dev, &qp->mqp);
+
+err_qpn:
+	mlx4_qp_release_range(dev->dev, qpn, 1);
+	return err;
+}
+
+static int _mlx4_ib_create_qp_rss(struct ib_pd *pd, struct mlx4_ib_qp *qp,
+				  struct ib_qp_init_attr *init_attr,
+				  struct ib_udata *udata)
+{
+	struct mlx4_ib_create_qp_rss ucmd = {};
+	size_t required_cmd_sz;
+	int err;
+
+	if (!udata) {
+		pr_debug("RSS QP with NULL udata\n");
+		return -EINVAL;
+	}
+
+	if (udata->outlen)
+		return -EOPNOTSUPP;
+
+	required_cmd_sz = offsetof(typeof(ucmd), reserved1) +
+					sizeof(ucmd.reserved1);
+	if (udata->inlen < required_cmd_sz) {
+		pr_debug("invalid inlen\n");
+		return -EINVAL;
+	}
+
+	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
+		pr_debug("copy failed\n");
+		return -EFAULT;
+	}
+
+	if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)))
+		return -EOPNOTSUPP;
+
+	if (ucmd.comp_mask || ucmd.reserved1)
+		return -EOPNOTSUPP;
+
+	if (udata->inlen > sizeof(ucmd) &&
+	    !ib_is_udata_cleared(udata, sizeof(ucmd),
+				 udata->inlen - sizeof(ucmd))) {
+		pr_debug("inlen is not supported\n");
+		return -EOPNOTSUPP;
+	}
+
+	if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
+		pr_debug("RSS QP with unsupported QP type %d\n",
+			 init_attr->qp_type);
+		return -EOPNOTSUPP;
+	}
+
+	if (init_attr->create_flags) {
+		pr_debug("RSS QP doesn't support create flags\n");
+		return -EOPNOTSUPP;
+	}
+
+	if (init_attr->send_cq || init_attr->cap.max_send_wr) {
+		pr_debug("RSS QP with unsupported send attributes\n");
+		return -EOPNOTSUPP;
+	}
+
+	qp->pri.vid = 0xFFFF;
+	qp->alt.vid = 0xFFFF;
+
+	err = create_qp_rss(to_mdev(pd->device), init_attr, &ucmd, qp);
+	if (err)
+		return err;
+
+	qp->ibqp.qp_num = qp->mqp.qpn;
+	return 0;
+}
+
+/*
+ * This function allocates a WQN from a range which is consecutive and aligned
+ * to its size. In case the range is full, then it creates a new range and
+ * allocates WQN from it. The new range will be used for following allocations.
+ */
+static int mlx4_ib_alloc_wqn(struct mlx4_ib_ucontext *context,
+			     struct mlx4_ib_qp *qp, int range_size, int *wqn)
+{
+	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
+	struct mlx4_wqn_range *range;
+	int err = 0;
+
+	mutex_lock(&context->wqn_ranges_mutex);
+
+	range = list_first_entry_or_null(&context->wqn_ranges_list,
+					 struct mlx4_wqn_range, list);
+
+	if (!range || (range->refcount == range->size) || range->dirty) {
+		range = kzalloc(sizeof(*range), GFP_KERNEL);
+		if (!range) {
+			err = -ENOMEM;
+			goto out;
+		}
+
+		err = mlx4_qp_reserve_range(dev->dev, range_size,
+					    range_size, &range->base_wqn, 0,
+					    qp->mqp.usage);
+		if (err) {
+			kfree(range);
+			goto out;
+		}
+
+		range->size = range_size;
+		list_add(&range->list, &context->wqn_ranges_list);
+	} else if (range_size != 1) {
+		/*
+		 * Requesting a new range (>1) when last range is still open, is
+		 * not valid.
+		 */
+		err = -EINVAL;
+		goto out;
+	}
+
+	qp->wqn_range = range;
+
+	*wqn = range->base_wqn + range->refcount;
+
+	range->refcount++;
+
+out:
+	mutex_unlock(&context->wqn_ranges_mutex);
+
+	return err;
+}
+
+static void mlx4_ib_release_wqn(struct mlx4_ib_ucontext *context,
+				struct mlx4_ib_qp *qp, bool dirty_release)
+{
+	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
+	struct mlx4_wqn_range *range;
+
+	mutex_lock(&context->wqn_ranges_mutex);
+
+	range = qp->wqn_range;
+
+	range->refcount--;
+	if (!range->refcount) {
+		mlx4_qp_release_range(dev->dev, range->base_wqn,
+				      range->size);
+		list_del(&range->list);
+		kfree(range);
+	} else if (dirty_release) {
+	/*
+	 * A range which one of its WQNs is destroyed, won't be able to be
+	 * reused for further WQN allocations.
+	 * The next created WQ will allocate a new range.
+	 */
+		range->dirty = true;
+	}
+
+	mutex_unlock(&context->wqn_ranges_mutex);
+}
+
+static int create_rq(struct ib_pd *pd, struct ib_qp_init_attr *init_attr,
+		     struct ib_udata *udata, struct mlx4_ib_qp *qp)
+{
+	struct mlx4_ib_dev *dev = to_mdev(pd->device);
+	int qpn;
+	int err;
+	struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context(
+		udata, struct mlx4_ib_ucontext, ibucontext);
+	struct mlx4_ib_cq *mcq;
+	unsigned long flags;
+	int range_size;
+	struct mlx4_ib_create_wq wq;
+	size_t copy_len;
+	int shift;
+	int n;
+
+	qp->mlx4_ib_qp_type = MLX4_IB_QPT_RAW_PACKET;
+
+	spin_lock_init(&qp->sq.lock);
+	spin_lock_init(&qp->rq.lock);
+	INIT_LIST_HEAD(&qp->gid_list);
+	INIT_LIST_HEAD(&qp->steering_rules);
+
+	qp->state = IB_QPS_RESET;
+
+	copy_len = min(sizeof(struct mlx4_ib_create_wq), udata->inlen);
+
+	if (ib_copy_from_udata(&wq, udata, copy_len)) {
+		err = -EFAULT;
+		goto err;
+	}
+
+	if (wq.comp_mask || wq.reserved[0] || wq.reserved[1] ||
+	    wq.reserved[2]) {
+		pr_debug("user command isn't supported\n");
+		err = -EOPNOTSUPP;
+		goto err;
+	}
+
+	if (wq.log_range_size > ilog2(dev->dev->caps.max_rss_tbl_sz)) {
+		pr_debug("WQN range size must be equal or smaller than %d\n",
+			 dev->dev->caps.max_rss_tbl_sz);
+		err = -EOPNOTSUPP;
+		goto err;
+	}
+	range_size = 1 << wq.log_range_size;
+
+	if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS)
+		qp->flags |= MLX4_IB_QP_SCATTER_FCS;
+
+	err = set_rq_size(dev, &init_attr->cap, true, true, qp, qp->inl_recv_sz);
+	if (err)
+		goto err;
+
+	qp->sq_no_prefetch = 1;
+	qp->sq.wqe_cnt = 1;
+	qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
+	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
+		       (qp->sq.wqe_cnt << qp->sq.wqe_shift);
+
+	qp->umem = ib_umem_get(pd->device, wq.buf_addr, qp->buf_size, 0);
+	if (IS_ERR(qp->umem)) {
+		err = PTR_ERR(qp->umem);
+		goto err;
+	}
+
+	shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n);
+	err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt);
+
+	if (err)
+		goto err_buf;
+
+	err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
+	if (err)
+		goto err_mtt;
+
+	err = mlx4_ib_db_map_user(udata, wq.db_addr, &qp->db);
+	if (err)
+		goto err_mtt;
+	qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
+
+	err = mlx4_ib_alloc_wqn(context, qp, range_size, &qpn);
+	if (err)
+		goto err_wrid;
+
+	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
+	if (err)
+		goto err_qpn;
+
+	/*
+	 * Hardware wants QPN written in big-endian order (after
+	 * shifting) for send doorbell.  Precompute this value to save
+	 * a little bit when posting sends.
+	 */
+	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
+
+	qp->mqp.event = mlx4_ib_wq_event;
+
+	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
+	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
+			 to_mcq(init_attr->recv_cq));
+	/* Maintain device to QPs access, needed for further handling
+	 * via reset flow
+	 */
+	list_add_tail(&qp->qps_list, &dev->qp_list);
+	/* Maintain CQ to QPs access, needed for further handling
+	 * via reset flow
+	 */
+	mcq = to_mcq(init_attr->send_cq);
+	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
+	mcq = to_mcq(init_attr->recv_cq);
+	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
+	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
+			   to_mcq(init_attr->recv_cq));
+	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
+	return 0;
+
+err_qpn:
+	mlx4_ib_release_wqn(context, qp, 0);
+err_wrid:
+	mlx4_ib_db_unmap_user(context, &qp->db);
+
+err_mtt:
+	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
+err_buf:
+	ib_umem_release(qp->umem);
+err:
+	return err;
+}
+
+static int create_qp_common(struct ib_pd *pd, struct ib_qp_init_attr *init_attr,
+			    struct ib_udata *udata, int sqpn,
+			    struct mlx4_ib_qp *qp)
+{
+	struct mlx4_ib_dev *dev = to_mdev(pd->device);
+	int qpn;
+	int err;
+	struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context(
+		udata, struct mlx4_ib_ucontext, ibucontext);
+	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
+	struct mlx4_ib_cq *mcq;
+	unsigned long flags;
+
+	/* When tunneling special qps, we use a plain UD qp */
+	if (sqpn) {
+		if (mlx4_is_mfunc(dev->dev) &&
+		    (!mlx4_is_master(dev->dev) ||
+		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
+			if (init_attr->qp_type == IB_QPT_GSI)
+				qp_type = MLX4_IB_QPT_PROXY_GSI;
+			else {
+				if (mlx4_is_master(dev->dev) ||
+				    qp0_enabled_vf(dev->dev, sqpn))
+					qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
+				else
+					qp_type = MLX4_IB_QPT_PROXY_SMI;
+			}
+		}
+		qpn = sqpn;
+		/* add extra sg entry for tunneling */
+		init_attr->cap.max_recv_sge++;
+	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
+		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
+			container_of(init_attr,
+				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
+		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
+		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
+		    !mlx4_is_master(dev->dev))
+			return -EINVAL;
+		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
+			qp_type = MLX4_IB_QPT_TUN_GSI;
+		else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
+			 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
+					     tnl_init->port))
+			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
+		else
+			qp_type = MLX4_IB_QPT_TUN_SMI;
+		/* we are definitely in the PPF here, since we are creating
+		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
+		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
+			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
+		sqpn = qpn;
+	}
+
+	if (init_attr->qp_type == IB_QPT_SMI ||
+	    init_attr->qp_type == IB_QPT_GSI || qp_type == MLX4_IB_QPT_SMI ||
+	    qp_type == MLX4_IB_QPT_GSI ||
+	    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
+			MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
+		qp->sqp = kzalloc(sizeof(struct mlx4_ib_sqp), GFP_KERNEL);
+		if (!qp->sqp)
+			return -ENOMEM;
+	}
+
+	qp->mlx4_ib_qp_type = qp_type;
+
+	spin_lock_init(&qp->sq.lock);
+	spin_lock_init(&qp->rq.lock);
+	INIT_LIST_HEAD(&qp->gid_list);
+	INIT_LIST_HEAD(&qp->steering_rules);
+
+	qp->state = IB_QPS_RESET;
+	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
+		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
+
+	if (udata) {
+		struct mlx4_ib_create_qp ucmd;
+		size_t copy_len;
+		int shift;
+		int n;
+
+		copy_len = sizeof(struct mlx4_ib_create_qp);
+
+		if (ib_copy_from_udata(&ucmd, udata, copy_len)) {
+			err = -EFAULT;
+			goto err;
+		}
+
+		qp->inl_recv_sz = ucmd.inl_recv_sz;
+
+		if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
+			if (!(dev->dev->caps.flags &
+			      MLX4_DEV_CAP_FLAG_FCS_KEEP)) {
+				pr_debug("scatter FCS is unsupported\n");
+				err = -EOPNOTSUPP;
+				goto err;
+			}
+
+			qp->flags |= MLX4_IB_QP_SCATTER_FCS;
+		}
+
+		err = set_rq_size(dev, &init_attr->cap, udata,
+				  qp_has_rq(init_attr), qp, qp->inl_recv_sz);
+		if (err)
+			goto err;
+
+		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
+
+		err = set_user_sq_size(dev, qp, &ucmd);
+		if (err)
+			goto err;
+
+		qp->umem =
+			ib_umem_get(pd->device, ucmd.buf_addr, qp->buf_size, 0);
+		if (IS_ERR(qp->umem)) {
+			err = PTR_ERR(qp->umem);
+			goto err;
+		}
+
+		shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n);
+		err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt);
+
+		if (err)
+			goto err_buf;
+
+		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
+		if (err)
+			goto err_mtt;
+
+		if (qp_has_rq(init_attr)) {
+			err = mlx4_ib_db_map_user(udata, ucmd.db_addr, &qp->db);
+			if (err)
+				goto err_mtt;
+		}
+		qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
+	} else {
+		err = set_rq_size(dev, &init_attr->cap, udata,
+				  qp_has_rq(init_attr), qp, 0);
+		if (err)
+			goto err;
+
+		qp->sq_no_prefetch = 0;
+
+		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
+			qp->flags |= MLX4_IB_QP_LSO;
+
+		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
+			if (dev->steering_support ==
+			    MLX4_STEERING_MODE_DEVICE_MANAGED)
+				qp->flags |= MLX4_IB_QP_NETIF;
+			else {
+				err = -EINVAL;
+				goto err;
+			}
+		}
+
+		err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp);
+		if (err)
+			goto err;
+
+		if (qp_has_rq(init_attr)) {
+			err = mlx4_db_alloc(dev->dev, &qp->db, 0);
+			if (err)
+				goto err;
+
+			*qp->db.db = 0;
+		}
+
+		if (mlx4_buf_alloc(dev->dev, qp->buf_size,  PAGE_SIZE * 2,
+				   &qp->buf)) {
+			err = -ENOMEM;
+			goto err_db;
+		}
+
+		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
+				    &qp->mtt);
+		if (err)
+			goto err_buf;
+
+		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
+		if (err)
+			goto err_mtt;
+
+		qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
+					     sizeof(u64), GFP_KERNEL);
+		qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
+					     sizeof(u64), GFP_KERNEL);
+		if (!qp->sq.wrid || !qp->rq.wrid) {
+			err = -ENOMEM;
+			goto err_wrid;
+		}
+		qp->mqp.usage = MLX4_RES_USAGE_DRIVER;
+	}
+
+	if (sqpn) {
+		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
+		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
+			if (alloc_proxy_bufs(pd->device, qp)) {
+				err = -ENOMEM;
+				goto err_wrid;
+			}
+		}
+	} else {
+		/* Raw packet QPNs may not have bits 6,7 set in their qp_num;
+		 * otherwise, the WQE BlueFlame setup flow wrongly causes
+		 * VLAN insertion. */
+		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
+			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
+						    (init_attr->cap.max_send_wr ?
+						     MLX4_RESERVE_ETH_BF_QP : 0) |
+						    (init_attr->cap.max_recv_wr ?
+						     MLX4_RESERVE_A0_QP : 0),
+						    qp->mqp.usage);
+		else
+			if (qp->flags & MLX4_IB_QP_NETIF)
+				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
+			else
+				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
+							    &qpn, 0, qp->mqp.usage);
+		if (err)
+			goto err_proxy;
+	}
+
+	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
+		qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
+
+	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
+	if (err)
+		goto err_qpn;
+
+	if (init_attr->qp_type == IB_QPT_XRC_TGT)
+		qp->mqp.qpn |= (1 << 23);
+
+	/*
+	 * Hardware wants QPN written in big-endian order (after
+	 * shifting) for send doorbell.  Precompute this value to save
+	 * a little bit when posting sends.
+	 */
+	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
+
+	qp->mqp.event = mlx4_ib_qp_event;
+
+	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
+	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
+			 to_mcq(init_attr->recv_cq));
+	/* Maintain device to QPs access, needed for further handling
+	 * via reset flow
+	 */
+	list_add_tail(&qp->qps_list, &dev->qp_list);
+	/* Maintain CQ to QPs access, needed for further handling
+	 * via reset flow
+	 */
+	mcq = to_mcq(init_attr->send_cq);
+	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
+	mcq = to_mcq(init_attr->recv_cq);
+	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
+	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
+			   to_mcq(init_attr->recv_cq));
+	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
+	return 0;
+
+err_qpn:
+	if (!sqpn) {
+		if (qp->flags & MLX4_IB_QP_NETIF)
+			mlx4_ib_steer_qp_free(dev, qpn, 1);
+		else
+			mlx4_qp_release_range(dev->dev, qpn, 1);
+	}
+err_proxy:
+	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
+		free_proxy_bufs(pd->device, qp);
+err_wrid:
+	if (udata) {
+		if (qp_has_rq(init_attr))
+			mlx4_ib_db_unmap_user(context, &qp->db);
+	} else {
+		kvfree(qp->sq.wrid);
+		kvfree(qp->rq.wrid);
+	}
+
+err_mtt:
+	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
+
+err_buf:
+	if (!qp->umem)
+		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
+	ib_umem_release(qp->umem);
+
+err_db:
+	if (!udata && qp_has_rq(init_attr))
+		mlx4_db_free(dev->dev, &qp->db);
+
+err:
+	kfree(qp->sqp);
+	return err;
+}
+
+static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
+{
+	switch (state) {
+	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
+	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
+	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
+	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
+	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
+	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
+	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
+	default:		return -1;
+	}
+}
+
+static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
+	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
+{
+	if (send_cq == recv_cq) {
+		spin_lock(&send_cq->lock);
+		__acquire(&recv_cq->lock);
+	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
+		spin_lock(&send_cq->lock);
+		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
+	} else {
+		spin_lock(&recv_cq->lock);
+		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
+	}
+}
+
+static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
+	__releases(&send_cq->lock) __releases(&recv_cq->lock)
+{
+	if (send_cq == recv_cq) {
+		__release(&recv_cq->lock);
+		spin_unlock(&send_cq->lock);
+	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
+		spin_unlock(&recv_cq->lock);
+		spin_unlock(&send_cq->lock);
+	} else {
+		spin_unlock(&send_cq->lock);
+		spin_unlock(&recv_cq->lock);
+	}
+}
+
+static void del_gid_entries(struct mlx4_ib_qp *qp)
+{
+	struct mlx4_ib_gid_entry *ge, *tmp;
+
+	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
+		list_del(&ge->list);
+		kfree(ge);
+	}
+}
+
+static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
+{
+	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
+		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
+	else
+		return to_mpd(qp->ibqp.pd);
+}
+
+static void get_cqs(struct mlx4_ib_qp *qp, enum mlx4_ib_source_type src,
+		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
+{
+	switch (qp->ibqp.qp_type) {
+	case IB_QPT_XRC_TGT:
+		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
+		*recv_cq = *send_cq;
+		break;
+	case IB_QPT_XRC_INI:
+		*send_cq = to_mcq(qp->ibqp.send_cq);
+		*recv_cq = *send_cq;
+		break;
+	default:
+		*recv_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.recv_cq) :
+						     to_mcq(qp->ibwq.cq);
+		*send_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.send_cq) :
+						     *recv_cq;
+		break;
+	}
+}
+
+static void destroy_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
+{
+	if (qp->state != IB_QPS_RESET) {
+		int i;
+
+		for (i = 0; i < (1 << qp->ibqp.rwq_ind_tbl->log_ind_tbl_size);
+		     i++) {
+			struct ib_wq *ibwq = qp->ibqp.rwq_ind_tbl->ind_tbl[i];
+			struct mlx4_ib_qp *wq =	to_mqp((struct ib_qp *)ibwq);
+
+			mutex_lock(&wq->mutex);
+
+			wq->rss_usecnt--;
+
+			mutex_unlock(&wq->mutex);
+		}
+
+		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
+				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
+			pr_warn("modify QP %06x to RESET failed.\n",
+				qp->mqp.qpn);
+	}
+
+	mlx4_qp_remove(dev->dev, &qp->mqp);
+	mlx4_qp_free(dev->dev, &qp->mqp);
+	mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
+	del_gid_entries(qp);
+}
+
+static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
+			      enum mlx4_ib_source_type src,
+			      struct ib_udata *udata)
+{
+	struct mlx4_ib_cq *send_cq, *recv_cq;
+	unsigned long flags;
+
+	if (qp->state != IB_QPS_RESET) {
+		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
+				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
+			pr_warn("modify QP %06x to RESET failed.\n",
+			       qp->mqp.qpn);
+		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
+			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
+			qp->pri.smac = 0;
+			qp->pri.smac_port = 0;
+		}
+		if (qp->alt.smac) {
+			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
+			qp->alt.smac = 0;
+		}
+		if (qp->pri.vid < 0x1000) {
+			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
+			qp->pri.vid = 0xFFFF;
+			qp->pri.candidate_vid = 0xFFFF;
+			qp->pri.update_vid = 0;
+		}
+		if (qp->alt.vid < 0x1000) {
+			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
+			qp->alt.vid = 0xFFFF;
+			qp->alt.candidate_vid = 0xFFFF;
+			qp->alt.update_vid = 0;
+		}
+	}
+
+	get_cqs(qp, src, &send_cq, &recv_cq);
+
+	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
+	mlx4_ib_lock_cqs(send_cq, recv_cq);
+
+	/* del from lists under both locks above to protect reset flow paths */
+	list_del(&qp->qps_list);
+	list_del(&qp->cq_send_list);
+	list_del(&qp->cq_recv_list);
+	if (!udata) {
+		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
+				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
+		if (send_cq != recv_cq)
+			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
+	}
+
+	mlx4_qp_remove(dev->dev, &qp->mqp);
+
+	mlx4_ib_unlock_cqs(send_cq, recv_cq);
+	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
+
+	mlx4_qp_free(dev->dev, &qp->mqp);
+
+	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
+		if (qp->flags & MLX4_IB_QP_NETIF)
+			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
+		else if (src == MLX4_IB_RWQ_SRC)
+			mlx4_ib_release_wqn(
+				rdma_udata_to_drv_context(
+					udata,
+					struct mlx4_ib_ucontext,
+					ibucontext),
+				qp, 1);
+		else
+			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
+	}
+
+	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
+
+	if (udata) {
+		if (qp->rq.wqe_cnt) {
+			struct mlx4_ib_ucontext *mcontext =
+				rdma_udata_to_drv_context(
+					udata,
+					struct mlx4_ib_ucontext,
+					ibucontext);
+
+			mlx4_ib_db_unmap_user(mcontext, &qp->db);
+		}
+	} else {
+		kvfree(qp->sq.wrid);
+		kvfree(qp->rq.wrid);
+		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
+		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
+			free_proxy_bufs(&dev->ib_dev, qp);
+		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
+		if (qp->rq.wqe_cnt)
+			mlx4_db_free(dev->dev, &qp->db);
+	}
+	ib_umem_release(qp->umem);
+
+	del_gid_entries(qp);
+}
+
+static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
+{
+	/* Native or PPF */
+	if (!mlx4_is_mfunc(dev->dev) ||
+	    (mlx4_is_master(dev->dev) &&
+	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
+		return  dev->dev->phys_caps.base_sqpn +
+			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
+			attr->port_num - 1;
+	}
+	/* PF or VF -- creating proxies */
+	if (attr->qp_type == IB_QPT_SMI)
+		return dev->dev->caps.spec_qps[attr->port_num - 1].qp0_proxy;
+	else
+		return dev->dev->caps.spec_qps[attr->port_num - 1].qp1_proxy;
+}
+
+static int _mlx4_ib_create_qp(struct ib_pd *pd, struct mlx4_ib_qp *qp,
+			      struct ib_qp_init_attr *init_attr,
+			      struct ib_udata *udata)
+{
+	int err;
+	int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
+	u16 xrcdn = 0;
+
+	if (init_attr->rwq_ind_tbl)
+		return _mlx4_ib_create_qp_rss(pd, qp, init_attr, udata);
+
+	/*
+	 * We only support LSO, vendor flag1, and multicast loopback blocking,
+	 * and only for kernel UD QPs.
+	 */
+	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
+					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
+					MLX4_IB_SRIOV_TUNNEL_QP |
+					MLX4_IB_SRIOV_SQP |
+					MLX4_IB_QP_NETIF |
+					MLX4_IB_QP_CREATE_ROCE_V2_GSI))
+		return -EINVAL;
+
+	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
+		if (init_attr->qp_type != IB_QPT_UD)
+			return -EINVAL;
+	}
+
+	if (init_attr->create_flags) {
+		if (udata && init_attr->create_flags & ~(sup_u_create_flags))
+			return -EINVAL;
+
+		if ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
+						 MLX4_IB_QP_CREATE_ROCE_V2_GSI  |
+						 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) &&
+		     init_attr->qp_type != IB_QPT_UD) ||
+		    (init_attr->create_flags & MLX4_IB_SRIOV_SQP &&
+		     init_attr->qp_type > IB_QPT_GSI) ||
+		    (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI &&
+		     init_attr->qp_type != IB_QPT_GSI))
+			return -EINVAL;
+	}
+
+	switch (init_attr->qp_type) {
+	case IB_QPT_XRC_TGT:
+		pd = to_mxrcd(init_attr->xrcd)->pd;
+		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
+		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
+		fallthrough;
+	case IB_QPT_XRC_INI:
+		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
+			return -ENOSYS;
+		init_attr->recv_cq = init_attr->send_cq;
+		fallthrough;
+	case IB_QPT_RC:
+	case IB_QPT_UC:
+	case IB_QPT_RAW_PACKET:
+	case IB_QPT_UD:
+		qp->pri.vid = 0xFFFF;
+		qp->alt.vid = 0xFFFF;
+		err = create_qp_common(pd, init_attr, udata, 0, qp);
+		if (err)
+			return err;
+
+		qp->ibqp.qp_num = qp->mqp.qpn;
+		qp->xrcdn = xrcdn;
+		break;
+	case IB_QPT_SMI:
+	case IB_QPT_GSI:
+	{
+		int sqpn;
+
+		if (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI) {
+			int res = mlx4_qp_reserve_range(to_mdev(pd->device)->dev,
+							1, 1, &sqpn, 0,
+							MLX4_RES_USAGE_DRIVER);
+
+			if (res)
+				return res;
+		} else {
+			sqpn = get_sqp_num(to_mdev(pd->device), init_attr);
+		}
+
+		qp->pri.vid = 0xFFFF;
+		qp->alt.vid = 0xFFFF;
+		err = create_qp_common(pd, init_attr, udata, sqpn, qp);
+		if (err)
+			return err;
+
+		qp->port	= init_attr->port_num;
+		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 :
+			init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn : 1;
+		break;
+	}
+	default:
+		/* Don't support raw QPs */
+		return -EOPNOTSUPP;
+	}
+	return 0;
+}
+
+struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
+				struct ib_qp_init_attr *init_attr,
+				struct ib_udata *udata) {
+	struct ib_device *device = pd ? pd->device : init_attr->xrcd->device;
+	struct mlx4_ib_dev *dev = to_mdev(device);
+	struct mlx4_ib_qp *qp;
+	int ret;
+
+	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
+	if (!qp)
+		return ERR_PTR(-ENOMEM);
+
+	mutex_init(&qp->mutex);
+	ret = _mlx4_ib_create_qp(pd, qp, init_attr, udata);
+	if (ret) {
+		kfree(qp);
+		return ERR_PTR(ret);
+	}
+
+	if (init_attr->qp_type == IB_QPT_GSI &&
+	    !(init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI)) {
+		struct mlx4_ib_sqp *sqp = qp->sqp;
+		int is_eth = rdma_cap_eth_ah(&dev->ib_dev, init_attr->port_num);
+
+		if (is_eth &&
+		    dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
+			init_attr->create_flags |= MLX4_IB_QP_CREATE_ROCE_V2_GSI;
+			sqp->roce_v2_gsi = ib_create_qp(pd, init_attr);
+
+			if (IS_ERR(sqp->roce_v2_gsi)) {
+				pr_err("Failed to create GSI QP for RoCEv2 (%ld)\n", PTR_ERR(sqp->roce_v2_gsi));
+				sqp->roce_v2_gsi = NULL;
+			} else {
+				to_mqp(sqp->roce_v2_gsi)->flags |=
+					MLX4_IB_ROCE_V2_GSI_QP;
+			}
+
+			init_attr->create_flags &= ~MLX4_IB_QP_CREATE_ROCE_V2_GSI;
+		}
+	}
+	return &qp->ibqp;
+}
+
+static int _mlx4_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
+{
+	struct mlx4_ib_dev *dev = to_mdev(qp->device);
+	struct mlx4_ib_qp *mqp = to_mqp(qp);
+
+	if (is_qp0(dev, mqp))
+		mlx4_CLOSE_PORT(dev->dev, mqp->port);
+
+	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI &&
+	    dev->qp1_proxy[mqp->port - 1] == mqp) {
+		mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
+		dev->qp1_proxy[mqp->port - 1] = NULL;
+		mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
+	}
+
+	if (mqp->counter_index)
+		mlx4_ib_free_qp_counter(dev, mqp);
+
+	if (qp->rwq_ind_tbl) {
+		destroy_qp_rss(dev, mqp);
+	} else {
+		destroy_qp_common(dev, mqp, MLX4_IB_QP_SRC, udata);
+	}
+
+	kfree(mqp->sqp);
+	kfree(mqp);
+
+	return 0;
+}
+
+int mlx4_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
+{
+	struct mlx4_ib_qp *mqp = to_mqp(qp);
+
+	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
+		struct mlx4_ib_sqp *sqp = mqp->sqp;
+
+		if (sqp->roce_v2_gsi)
+			ib_destroy_qp(sqp->roce_v2_gsi);
+	}
+
+	return _mlx4_ib_destroy_qp(qp, udata);
+}
+
+static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
+{
+	switch (type) {
+	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
+	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
+	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
+	case MLX4_IB_QPT_XRC_INI:
+	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
+	case MLX4_IB_QPT_SMI:
+	case MLX4_IB_QPT_GSI:
+	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
+
+	case MLX4_IB_QPT_PROXY_SMI_OWNER:
+	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
+						MLX4_QP_ST_MLX : -1);
+	case MLX4_IB_QPT_PROXY_SMI:
+	case MLX4_IB_QPT_TUN_SMI:
+	case MLX4_IB_QPT_PROXY_GSI:
+	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
+						MLX4_QP_ST_UD : -1);
+	default:			return -1;
+	}
+}
+
+static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
+				   int attr_mask)
+{
+	u8 dest_rd_atomic;
+	u32 access_flags;
+	u32 hw_access_flags = 0;
+
+	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
+		dest_rd_atomic = attr->max_dest_rd_atomic;
+	else
+		dest_rd_atomic = qp->resp_depth;
+
+	if (attr_mask & IB_QP_ACCESS_FLAGS)
+		access_flags = attr->qp_access_flags;
+	else
+		access_flags = qp->atomic_rd_en;
+
+	if (!dest_rd_atomic)
+		access_flags &= IB_ACCESS_REMOTE_WRITE;
+
+	if (access_flags & IB_ACCESS_REMOTE_READ)
+		hw_access_flags |= MLX4_QP_BIT_RRE;
+	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
+		hw_access_flags |= MLX4_QP_BIT_RAE;
+	if (access_flags & IB_ACCESS_REMOTE_WRITE)
+		hw_access_flags |= MLX4_QP_BIT_RWE;
+
+	return cpu_to_be32(hw_access_flags);
+}
+
+static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
+			    int attr_mask)
+{
+	if (attr_mask & IB_QP_PKEY_INDEX)
+		sqp->pkey_index = attr->pkey_index;
+	if (attr_mask & IB_QP_QKEY)
+		sqp->qkey = attr->qkey;
+	if (attr_mask & IB_QP_SQ_PSN)
+		sqp->send_psn = attr->sq_psn;
+}
+
+static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
+{
+	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
+}
+
+static int _mlx4_set_path(struct mlx4_ib_dev *dev,
+			  const struct rdma_ah_attr *ah,
+			  u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
+			  struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
+{
+	int vidx;
+	int smac_index;
+	int err;
+
+	path->grh_mylmc = rdma_ah_get_path_bits(ah) & 0x7f;
+	path->rlid = cpu_to_be16(rdma_ah_get_dlid(ah));
+	if (rdma_ah_get_static_rate(ah)) {
+		path->static_rate = rdma_ah_get_static_rate(ah) +
+				    MLX4_STAT_RATE_OFFSET;
+		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
+		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
+			--path->static_rate;
+	} else
+		path->static_rate = 0;
+
+	if (rdma_ah_get_ah_flags(ah) & IB_AH_GRH) {
+		const struct ib_global_route *grh = rdma_ah_read_grh(ah);
+		int real_sgid_index =
+			mlx4_ib_gid_index_to_real_index(dev, grh->sgid_attr);
+
+		if (real_sgid_index < 0)
+			return real_sgid_index;
+		if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
+			pr_err("sgid_index (%u) too large. max is %d\n",
+			       real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
+			return -1;
+		}
+
+		path->grh_mylmc |= 1 << 7;
+		path->mgid_index = real_sgid_index;
+		path->hop_limit  = grh->hop_limit;
+		path->tclass_flowlabel =
+			cpu_to_be32((grh->traffic_class << 20) |
+				    (grh->flow_label));
+		memcpy(path->rgid, grh->dgid.raw, 16);
+	}
+
+	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
+		if (!(rdma_ah_get_ah_flags(ah) & IB_AH_GRH))
+			return -1;
+
+		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
+			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 7) << 3);
+
+		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
+		if (vlan_tag < 0x1000) {
+			if (smac_info->vid < 0x1000) {
+				/* both valid vlan ids */
+				if (smac_info->vid != vlan_tag) {
+					/* different VIDs.  unreg old and reg new */
+					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
+					if (err)
+						return err;
+					smac_info->candidate_vid = vlan_tag;
+					smac_info->candidate_vlan_index = vidx;
+					smac_info->candidate_vlan_port = port;
+					smac_info->update_vid = 1;
+					path->vlan_index = vidx;
+				} else {
+					path->vlan_index = smac_info->vlan_index;
+				}
+			} else {
+				/* no current vlan tag in qp */
+				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
+				if (err)
+					return err;
+				smac_info->candidate_vid = vlan_tag;
+				smac_info->candidate_vlan_index = vidx;
+				smac_info->candidate_vlan_port = port;
+				smac_info->update_vid = 1;
+				path->vlan_index = vidx;
+			}
+			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
+			path->fl = 1 << 6;
+		} else {
+			/* have current vlan tag. unregister it at modify-qp success */
+			if (smac_info->vid < 0x1000) {
+				smac_info->candidate_vid = 0xFFFF;
+				smac_info->update_vid = 1;
+			}
+		}
+
+		/* get smac_index for RoCE use.
+		 * If no smac was yet assigned, register one.
+		 * If one was already assigned, but the new mac differs,
+		 * unregister the old one and register the new one.
+		*/
+		if ((!smac_info->smac && !smac_info->smac_port) ||
+		    smac_info->smac != smac) {
+			/* register candidate now, unreg if needed, after success */
+			smac_index = mlx4_register_mac(dev->dev, port, smac);
+			if (smac_index >= 0) {
+				smac_info->candidate_smac_index = smac_index;
+				smac_info->candidate_smac = smac;
+				smac_info->candidate_smac_port = port;
+			} else {
+				return -EINVAL;
+			}
+		} else {
+			smac_index = smac_info->smac_index;
+		}
+		memcpy(path->dmac, ah->roce.dmac, 6);
+		path->ackto = MLX4_IB_LINK_TYPE_ETH;
+		/* put MAC table smac index for IBoE */
+		path->grh_mylmc = (u8) (smac_index) | 0x80;
+	} else {
+		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
+			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 0xf) << 2);
+	}
+
+	return 0;
+}
+
+static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
+			 enum ib_qp_attr_mask qp_attr_mask,
+			 struct mlx4_ib_qp *mqp,
+			 struct mlx4_qp_path *path, u8 port,
+			 u16 vlan_id, u8 *smac)
+{
+	return _mlx4_set_path(dev, &qp->ah_attr,
+			      mlx4_mac_to_u64(smac),
+			      vlan_id,
+			      path, &mqp->pri, port);
+}
+
+static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
+			     const struct ib_qp_attr *qp,
+			     enum ib_qp_attr_mask qp_attr_mask,
+			     struct mlx4_ib_qp *mqp,
+			     struct mlx4_qp_path *path, u8 port)
+{
+	return _mlx4_set_path(dev, &qp->alt_ah_attr,
+			      0,
+			      0xffff,
+			      path, &mqp->alt, port);
+}
+
+static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
+{
+	struct mlx4_ib_gid_entry *ge, *tmp;
+
+	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
+		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
+			ge->added = 1;
+			ge->port = qp->port;
+		}
+	}
+}
+
+static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev,
+				    struct mlx4_ib_qp *qp,
+				    struct mlx4_qp_context *context)
+{
+	u64 u64_mac;
+	int smac_index;
+
+	u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
+
+	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
+	if (!qp->pri.smac && !qp->pri.smac_port) {
+		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
+		if (smac_index >= 0) {
+			qp->pri.candidate_smac_index = smac_index;
+			qp->pri.candidate_smac = u64_mac;
+			qp->pri.candidate_smac_port = qp->port;
+			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
+		} else {
+			return -ENOENT;
+		}
+	}
+	return 0;
+}
+
+static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
+{
+	struct counter_index *new_counter_index;
+	int err;
+	u32 tmp_idx;
+
+	if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
+	    IB_LINK_LAYER_ETHERNET ||
+	    !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
+	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
+		return 0;
+
+	err = mlx4_counter_alloc(dev->dev, &tmp_idx, MLX4_RES_USAGE_DRIVER);
+	if (err)
+		return err;
+
+	new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL);
+	if (!new_counter_index) {
+		mlx4_counter_free(dev->dev, tmp_idx);
+		return -ENOMEM;
+	}
+
+	new_counter_index->index = tmp_idx;
+	new_counter_index->allocated = 1;
+	qp->counter_index = new_counter_index;
+
+	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
+	list_add_tail(&new_counter_index->list,
+		      &dev->counters_table[qp->port - 1].counters_list);
+	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
+
+	return 0;
+}
+
+enum {
+	MLX4_QPC_ROCE_MODE_1 = 0,
+	MLX4_QPC_ROCE_MODE_2 = 2,
+	MLX4_QPC_ROCE_MODE_UNDEFINED = 0xff
+};
+
+static u8 gid_type_to_qpc(enum ib_gid_type gid_type)
+{
+	switch (gid_type) {
+	case IB_GID_TYPE_ROCE:
+		return MLX4_QPC_ROCE_MODE_1;
+	case IB_GID_TYPE_ROCE_UDP_ENCAP:
+		return MLX4_QPC_ROCE_MODE_2;
+	default:
+		return MLX4_QPC_ROCE_MODE_UNDEFINED;
+	}
+}
+
+/*
+ * Go over all RSS QP's childes (WQs) and apply their HW state according to
+ * their logic state if the RSS QP is the first RSS QP associated for the WQ.
+ */
+static int bringup_rss_rwqs(struct ib_rwq_ind_table *ind_tbl, u8 port_num,
+			    struct ib_udata *udata)
+{
+	int err = 0;
+	int i;
+
+	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
+		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
+		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
+
+		mutex_lock(&wq->mutex);
+
+		/* Mlx4_ib restrictions:
+		 * WQ's is associated to a port according to the RSS QP it is
+		 * associates to.
+		 * In case the WQ is associated to a different port by another
+		 * RSS QP, return a failure.
+		 */
+		if ((wq->rss_usecnt > 0) && (wq->port != port_num)) {
+			err = -EINVAL;
+			mutex_unlock(&wq->mutex);
+			break;
+		}
+		wq->port = port_num;
+		if ((wq->rss_usecnt == 0) && (ibwq->state == IB_WQS_RDY)) {
+			err = _mlx4_ib_modify_wq(ibwq, IB_WQS_RDY, udata);
+			if (err) {
+				mutex_unlock(&wq->mutex);
+				break;
+			}
+		}
+		wq->rss_usecnt++;
+
+		mutex_unlock(&wq->mutex);
+	}
+
+	if (i && err) {
+		int j;
+
+		for (j = (i - 1); j >= 0; j--) {
+			struct ib_wq *ibwq = ind_tbl->ind_tbl[j];
+			struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
+
+			mutex_lock(&wq->mutex);
+
+			if ((wq->rss_usecnt == 1) &&
+			    (ibwq->state == IB_WQS_RDY))
+				if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET,
+						       udata))
+					pr_warn("failed to reverse WQN=0x%06x\n",
+						ibwq->wq_num);
+			wq->rss_usecnt--;
+
+			mutex_unlock(&wq->mutex);
+		}
+	}
+
+	return err;
+}
+
+static void bring_down_rss_rwqs(struct ib_rwq_ind_table *ind_tbl,
+				struct ib_udata *udata)
+{
+	int i;
+
+	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
+		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
+		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
+
+		mutex_lock(&wq->mutex);
+
+		if ((wq->rss_usecnt == 1) && (ibwq->state == IB_WQS_RDY))
+			if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET, udata))
+				pr_warn("failed to reverse WQN=%x\n",
+					ibwq->wq_num);
+		wq->rss_usecnt--;
+
+		mutex_unlock(&wq->mutex);
+	}
+}
+
+static void fill_qp_rss_context(struct mlx4_qp_context *context,
+				struct mlx4_ib_qp *qp)
+{
+	struct mlx4_rss_context *rss_context;
+
+	rss_context = (void *)context + offsetof(struct mlx4_qp_context,
+			pri_path) + MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
+
+	rss_context->base_qpn = cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz);
+	rss_context->default_qpn =
+		cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz & 0xffffff);
+	if (qp->rss_ctx->flags & (MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6))
+		rss_context->base_qpn_udp = rss_context->default_qpn;
+	rss_context->flags = qp->rss_ctx->flags;
+	/* Currently support just toeplitz */
+	rss_context->hash_fn = MLX4_RSS_HASH_TOP;
+
+	memcpy(rss_context->rss_key, qp->rss_ctx->rss_key,
+	       MLX4_EN_RSS_KEY_SIZE);
+}
+
+static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type,
+			       const struct ib_qp_attr *attr, int attr_mask,
+			       enum ib_qp_state cur_state,
+			       enum ib_qp_state new_state,
+			       struct ib_udata *udata)
+{
+	struct ib_srq  *ibsrq;
+	const struct ib_gid_attr *gid_attr = NULL;
+	struct ib_rwq_ind_table *rwq_ind_tbl;
+	enum ib_qp_type qp_type;
+	struct mlx4_ib_dev *dev;
+	struct mlx4_ib_qp *qp;
+	struct mlx4_ib_pd *pd;
+	struct mlx4_ib_cq *send_cq, *recv_cq;
+	struct mlx4_ib_ucontext *ucontext = rdma_udata_to_drv_context(
+		udata, struct mlx4_ib_ucontext, ibucontext);
+	struct mlx4_qp_context *context;
+	enum mlx4_qp_optpar optpar = 0;
+	int sqd_event;
+	int steer_qp = 0;
+	int err = -EINVAL;
+	int counter_index;
+
+	if (src_type == MLX4_IB_RWQ_SRC) {
+		struct ib_wq *ibwq;
+
+		ibwq	    = (struct ib_wq *)src;
+		ibsrq	    = NULL;
+		rwq_ind_tbl = NULL;
+		qp_type     = IB_QPT_RAW_PACKET;
+		qp	    = to_mqp((struct ib_qp *)ibwq);
+		dev	    = to_mdev(ibwq->device);
+		pd	    = to_mpd(ibwq->pd);
+	} else {
+		struct ib_qp *ibqp;
+
+		ibqp	    = (struct ib_qp *)src;
+		ibsrq	    = ibqp->srq;
+		rwq_ind_tbl = ibqp->rwq_ind_tbl;
+		qp_type     = ibqp->qp_type;
+		qp	    = to_mqp(ibqp);
+		dev	    = to_mdev(ibqp->device);
+		pd	    = get_pd(qp);
+	}
+
+	/* APM is not supported under RoCE */
+	if (attr_mask & IB_QP_ALT_PATH &&
+	    rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
+	    IB_LINK_LAYER_ETHERNET)
+		return -ENOTSUPP;
+
+	context = kzalloc(sizeof *context, GFP_KERNEL);
+	if (!context)
+		return -ENOMEM;
+
+	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
+				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
+
+	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
+		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
+	else {
+		optpar |= MLX4_QP_OPTPAR_PM_STATE;
+		switch (attr->path_mig_state) {
+		case IB_MIG_MIGRATED:
+			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
+			break;
+		case IB_MIG_REARM:
+			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
+			break;
+		case IB_MIG_ARMED:
+			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
+			break;
+		}
+	}
+
+	if (qp->inl_recv_sz)
+		context->param3 |= cpu_to_be32(1 << 25);
+
+	if (qp->flags & MLX4_IB_QP_SCATTER_FCS)
+		context->param3 |= cpu_to_be32(1 << 29);
+
+	if (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI)
+		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
+	else if (qp_type == IB_QPT_RAW_PACKET)
+		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
+	else if (qp_type == IB_QPT_UD) {
+		if (qp->flags & MLX4_IB_QP_LSO)
+			context->mtu_msgmax = (IB_MTU_4096 << 5) |
+					      ilog2(dev->dev->caps.max_gso_sz);
+		else
+			context->mtu_msgmax = (IB_MTU_4096 << 5) | 13;
+	} else if (attr_mask & IB_QP_PATH_MTU) {
+		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
+			pr_err("path MTU (%u) is invalid\n",
+			       attr->path_mtu);
+			goto out;
+		}
+		context->mtu_msgmax = (attr->path_mtu << 5) |
+			ilog2(dev->dev->caps.max_msg_sz);
+	}
+
+	if (!rwq_ind_tbl) { /* PRM RSS receive side should be left zeros */
+		if (qp->rq.wqe_cnt)
+			context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
+		context->rq_size_stride |= qp->rq.wqe_shift - 4;
+	}
+
+	if (qp->sq.wqe_cnt)
+		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
+	context->sq_size_stride |= qp->sq.wqe_shift - 4;
+
+	if (new_state == IB_QPS_RESET && qp->counter_index)
+		mlx4_ib_free_qp_counter(dev, qp);
+
+	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
+		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
+		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
+		if (qp_type == IB_QPT_RAW_PACKET)
+			context->param3 |= cpu_to_be32(1 << 30);
+	}
+
+	if (ucontext)
+		context->usr_page = cpu_to_be32(
+			mlx4_to_hw_uar_index(dev->dev, ucontext->uar.index));
+	else
+		context->usr_page = cpu_to_be32(
+			mlx4_to_hw_uar_index(dev->dev, dev->priv_uar.index));
+
+	if (attr_mask & IB_QP_DEST_QPN)
+		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
+
+	if (attr_mask & IB_QP_PORT) {
+		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
+		    !(attr_mask & IB_QP_AV)) {
+			mlx4_set_sched(&context->pri_path, attr->port_num);
+			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
+		}
+	}
+
+	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
+		err = create_qp_lb_counter(dev, qp);
+		if (err)
+			goto out;
+
+		counter_index =
+			dev->counters_table[qp->port - 1].default_counter;
+		if (qp->counter_index)
+			counter_index = qp->counter_index->index;
+
+		if (counter_index != -1) {
+			context->pri_path.counter_index = counter_index;
+			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
+			if (qp->counter_index) {
+				context->pri_path.fl |=
+					MLX4_FL_ETH_SRC_CHECK_MC_LB;
+				context->pri_path.vlan_control |=
+					MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
+			}
+		} else
+			context->pri_path.counter_index =
+				MLX4_SINK_COUNTER_INDEX(dev->dev);
+
+		if (qp->flags & MLX4_IB_QP_NETIF) {
+			mlx4_ib_steer_qp_reg(dev, qp, 1);
+			steer_qp = 1;
+		}
+
+		if (qp_type == IB_QPT_GSI) {
+			enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ?
+				IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE;
+			u8 qpc_roce_mode = gid_type_to_qpc(gid_type);
+
+			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
+		}
+	}
+
+	if (attr_mask & IB_QP_PKEY_INDEX) {
+		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
+			context->pri_path.disable_pkey_check = 0x40;
+		context->pri_path.pkey_index = attr->pkey_index;
+		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
+	}
+
+	if (attr_mask & IB_QP_AV) {
+		u8 port_num = mlx4_is_bonded(dev->dev) ? 1 :
+			attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
+		u16 vlan = 0xffff;
+		u8 smac[ETH_ALEN];
+		int is_eth =
+			rdma_cap_eth_ah(&dev->ib_dev, port_num) &&
+			rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
+
+		if (is_eth) {
+			gid_attr = attr->ah_attr.grh.sgid_attr;
+			err = rdma_read_gid_l2_fields(gid_attr, &vlan,
+						      &smac[0]);
+			if (err)
+				goto out;
+		}
+
+		if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
+				  port_num, vlan, smac))
+			goto out;
+
+		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
+			   MLX4_QP_OPTPAR_SCHED_QUEUE);
+
+		if (is_eth &&
+		    (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR)) {
+			u8 qpc_roce_mode = gid_type_to_qpc(gid_attr->gid_type);
+
+			if (qpc_roce_mode == MLX4_QPC_ROCE_MODE_UNDEFINED) {
+				err = -EINVAL;
+				goto out;
+			}
+			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
+		}
+
+	}
+
+	if (attr_mask & IB_QP_TIMEOUT) {
+		context->pri_path.ackto |= attr->timeout << 3;
+		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
+	}
+
+	if (attr_mask & IB_QP_ALT_PATH) {
+		if (attr->alt_port_num == 0 ||
+		    attr->alt_port_num > dev->dev->caps.num_ports)
+			goto out;
+
+		if (attr->alt_pkey_index >=
+		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
+			goto out;
+
+		if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
+				      &context->alt_path,
+				      attr->alt_port_num))
+			goto out;
+
+		context->alt_path.pkey_index = attr->alt_pkey_index;
+		context->alt_path.ackto = attr->alt_timeout << 3;
+		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
+	}
+
+	context->pd = cpu_to_be32(pd->pdn);
+
+	if (!rwq_ind_tbl) {
+		context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
+		get_cqs(qp, src_type, &send_cq, &recv_cq);
+	} else { /* Set dummy CQs to be compatible with HV and PRM */
+		send_cq = to_mcq(rwq_ind_tbl->ind_tbl[0]->cq);
+		recv_cq = send_cq;
+	}
+	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
+	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
+
+	/* Set "fast registration enabled" for all kernel QPs */
+	if (!ucontext)
+		context->params1 |= cpu_to_be32(1 << 11);
+
+	if (attr_mask & IB_QP_RNR_RETRY) {
+		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
+		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
+	}
+
+	if (attr_mask & IB_QP_RETRY_CNT) {
+		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
+		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
+	}
+
+	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
+		if (attr->max_rd_atomic)
+			context->params1 |=
+				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
+		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
+	}
+
+	if (attr_mask & IB_QP_SQ_PSN)
+		context->next_send_psn = cpu_to_be32(attr->sq_psn);
+
+	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
+		if (attr->max_dest_rd_atomic)
+			context->params2 |=
+				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
+		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
+	}
+
+	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
+		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
+		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
+	}
+
+	if (ibsrq)
+		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
+
+	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
+		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
+		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
+	}
+	if (attr_mask & IB_QP_RQ_PSN)
+		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
+
+	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
+	if (attr_mask & IB_QP_QKEY) {
+		if (qp->mlx4_ib_qp_type &
+		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
+			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
+		else {
+			if (mlx4_is_mfunc(dev->dev) &&
+			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
+			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
+			    MLX4_RESERVED_QKEY_BASE) {
+				pr_err("Cannot use reserved QKEY"
+				       " 0x%x (range 0xffff0000..0xffffffff"
+				       " is reserved)\n", attr->qkey);
+				err = -EINVAL;
+				goto out;
+			}
+			context->qkey = cpu_to_be32(attr->qkey);
+		}
+		optpar |= MLX4_QP_OPTPAR_Q_KEY;
+	}
+
+	if (ibsrq)
+		context->srqn = cpu_to_be32(1 << 24 |
+					    to_msrq(ibsrq)->msrq.srqn);
+
+	if (qp->rq.wqe_cnt &&
+	    cur_state == IB_QPS_RESET &&
+	    new_state == IB_QPS_INIT)
+		context->db_rec_addr = cpu_to_be64(qp->db.dma);
+
+	if (cur_state == IB_QPS_INIT &&
+	    new_state == IB_QPS_RTR  &&
+	    (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI ||
+	     qp_type == IB_QPT_UD || qp_type == IB_QPT_RAW_PACKET)) {
+		context->pri_path.sched_queue = (qp->port - 1) << 6;
+		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
+		    qp->mlx4_ib_qp_type &
+		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
+			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
+			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
+				context->pri_path.fl = 0x80;
+		} else {
+			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
+				context->pri_path.fl = 0x80;
+			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
+		}
+		if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
+		    IB_LINK_LAYER_ETHERNET) {
+			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
+			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
+				context->pri_path.feup = 1 << 7; /* don't fsm */
+			/* handle smac_index */
+			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
+			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
+			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
+				err = handle_eth_ud_smac_index(dev, qp, context);
+				if (err) {
+					err = -EINVAL;
+					goto out;
+				}
+				if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
+					dev->qp1_proxy[qp->port - 1] = qp;
+			}
+		}
+	}
+
+	if (qp_type == IB_QPT_RAW_PACKET) {
+		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
+					MLX4_IB_LINK_TYPE_ETH;
+		if (dev->dev->caps.tunnel_offload_mode ==  MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
+			/* set QP to receive both tunneled & non-tunneled packets */
+			if (!rwq_ind_tbl)
+				context->srqn = cpu_to_be32(7 << 28);
+		}
+	}
+
+	if (qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
+		int is_eth = rdma_port_get_link_layer(
+				&dev->ib_dev, qp->port) ==
+				IB_LINK_LAYER_ETHERNET;
+		if (is_eth) {
+			context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
+			optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
+		}
+	}
+
+	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
+	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
+		sqd_event = 1;
+	else
+		sqd_event = 0;
+
+	if (!ucontext &&
+	    cur_state == IB_QPS_RESET &&
+	    new_state == IB_QPS_INIT)
+		context->rlkey_roce_mode |= (1 << 4);
+
+	/*
+	 * Before passing a kernel QP to the HW, make sure that the
+	 * ownership bits of the send queue are set and the SQ
+	 * headroom is stamped so that the hardware doesn't start
+	 * processing stale work requests.
+	 */
+	if (!ucontext &&
+	    cur_state == IB_QPS_RESET &&
+	    new_state == IB_QPS_INIT) {
+		struct mlx4_wqe_ctrl_seg *ctrl;
+		int i;
+
+		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
+			ctrl = get_send_wqe(qp, i);
+			ctrl->owner_opcode = cpu_to_be32(1 << 31);
+			ctrl->qpn_vlan.fence_size =
+				1 << (qp->sq.wqe_shift - 4);
+			stamp_send_wqe(qp, i);
+		}
+	}
+
+	if (rwq_ind_tbl	&&
+	    cur_state == IB_QPS_RESET &&
+	    new_state == IB_QPS_INIT) {
+		fill_qp_rss_context(context, qp);
+		context->flags |= cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET);
+	}
+
+	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
+			     to_mlx4_state(new_state), context, optpar,
+			     sqd_event, &qp->mqp);
+	if (err)
+		goto out;
+
+	qp->state = new_state;
+
+	if (attr_mask & IB_QP_ACCESS_FLAGS)
+		qp->atomic_rd_en = attr->qp_access_flags;
+	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
+		qp->resp_depth = attr->max_dest_rd_atomic;
+	if (attr_mask & IB_QP_PORT) {
+		qp->port = attr->port_num;
+		update_mcg_macs(dev, qp);
+	}
+	if (attr_mask & IB_QP_ALT_PATH)
+		qp->alt_port = attr->alt_port_num;
+
+	if (is_sqp(dev, qp))
+		store_sqp_attrs(qp->sqp, attr, attr_mask);
+
+	/*
+	 * If we moved QP0 to RTR, bring the IB link up; if we moved
+	 * QP0 to RESET or ERROR, bring the link back down.
+	 */
+	if (is_qp0(dev, qp)) {
+		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
+			if (mlx4_INIT_PORT(dev->dev, qp->port))
+				pr_warn("INIT_PORT failed for port %d\n",
+				       qp->port);
+
+		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
+		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
+			mlx4_CLOSE_PORT(dev->dev, qp->port);
+	}
+
+	/*
+	 * If we moved a kernel QP to RESET, clean up all old CQ
+	 * entries and reinitialize the QP.
+	 */
+	if (new_state == IB_QPS_RESET) {
+		if (!ucontext) {
+			mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
+					 ibsrq ? to_msrq(ibsrq) : NULL);
+			if (send_cq != recv_cq)
+				mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
+
+			qp->rq.head = 0;
+			qp->rq.tail = 0;
+			qp->sq.head = 0;
+			qp->sq.tail = 0;
+			qp->sq_next_wqe = 0;
+			if (qp->rq.wqe_cnt)
+				*qp->db.db  = 0;
+
+			if (qp->flags & MLX4_IB_QP_NETIF)
+				mlx4_ib_steer_qp_reg(dev, qp, 0);
+		}
+		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
+			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
+			qp->pri.smac = 0;
+			qp->pri.smac_port = 0;
+		}
+		if (qp->alt.smac) {
+			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
+			qp->alt.smac = 0;
+		}
+		if (qp->pri.vid < 0x1000) {
+			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
+			qp->pri.vid = 0xFFFF;
+			qp->pri.candidate_vid = 0xFFFF;
+			qp->pri.update_vid = 0;
+		}
+
+		if (qp->alt.vid < 0x1000) {
+			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
+			qp->alt.vid = 0xFFFF;
+			qp->alt.candidate_vid = 0xFFFF;
+			qp->alt.update_vid = 0;
+		}
+	}
+out:
+	if (err && qp->counter_index)
+		mlx4_ib_free_qp_counter(dev, qp);
+	if (err && steer_qp)
+		mlx4_ib_steer_qp_reg(dev, qp, 0);
+	kfree(context);
+	if (qp->pri.candidate_smac ||
+	    (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
+		if (err) {
+			mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
+		} else {
+			if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
+				mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
+			qp->pri.smac = qp->pri.candidate_smac;
+			qp->pri.smac_index = qp->pri.candidate_smac_index;
+			qp->pri.smac_port = qp->pri.candidate_smac_port;
+		}
+		qp->pri.candidate_smac = 0;
+		qp->pri.candidate_smac_index = 0;
+		qp->pri.candidate_smac_port = 0;
+	}
+	if (qp->alt.candidate_smac) {
+		if (err) {
+			mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
+		} else {
+			if (qp->alt.smac)
+				mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
+			qp->alt.smac = qp->alt.candidate_smac;
+			qp->alt.smac_index = qp->alt.candidate_smac_index;
+			qp->alt.smac_port = qp->alt.candidate_smac_port;
+		}
+		qp->alt.candidate_smac = 0;
+		qp->alt.candidate_smac_index = 0;
+		qp->alt.candidate_smac_port = 0;
+	}
+
+	if (qp->pri.update_vid) {
+		if (err) {
+			if (qp->pri.candidate_vid < 0x1000)
+				mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
+						     qp->pri.candidate_vid);
+		} else {
+			if (qp->pri.vid < 0x1000)
+				mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
+						     qp->pri.vid);
+			qp->pri.vid = qp->pri.candidate_vid;
+			qp->pri.vlan_port = qp->pri.candidate_vlan_port;
+			qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
+		}
+		qp->pri.candidate_vid = 0xFFFF;
+		qp->pri.update_vid = 0;
+	}
+
+	if (qp->alt.update_vid) {
+		if (err) {
+			if (qp->alt.candidate_vid < 0x1000)
+				mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
+						     qp->alt.candidate_vid);
+		} else {
+			if (qp->alt.vid < 0x1000)
+				mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
+						     qp->alt.vid);
+			qp->alt.vid = qp->alt.candidate_vid;
+			qp->alt.vlan_port = qp->alt.candidate_vlan_port;
+			qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
+		}
+		qp->alt.candidate_vid = 0xFFFF;
+		qp->alt.update_vid = 0;
+	}
+
+	return err;
+}
+
+enum {
+	MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK = (IB_QP_STATE	|
+					      IB_QP_PORT),
+};
+
+static int _mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
+			      int attr_mask, struct ib_udata *udata)
+{
+	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
+	struct mlx4_ib_qp *qp = to_mqp(ibqp);
+	enum ib_qp_state cur_state, new_state;
+	int err = -EINVAL;
+	mutex_lock(&qp->mutex);
+
+	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
+	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
+
+	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
+				attr_mask)) {
+		pr_debug("qpn 0x%x: invalid attribute mask specified "
+			 "for transition %d to %d. qp_type %d,"
+			 " attr_mask 0x%x\n",
+			 ibqp->qp_num, cur_state, new_state,
+			 ibqp->qp_type, attr_mask);
+		goto out;
+	}
+
+	if (ibqp->rwq_ind_tbl) {
+		if (!(((cur_state == IB_QPS_RESET) &&
+		       (new_state == IB_QPS_INIT)) ||
+		      ((cur_state == IB_QPS_INIT)  &&
+		       (new_state == IB_QPS_RTR)))) {
+			pr_debug("qpn 0x%x: RSS QP unsupported transition %d to %d\n",
+				 ibqp->qp_num, cur_state, new_state);
+
+			err = -EOPNOTSUPP;
+			goto out;
+		}
+
+		if (attr_mask & ~MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK) {
+			pr_debug("qpn 0x%x: RSS QP unsupported attribute mask 0x%x for transition %d to %d\n",
+				 ibqp->qp_num, attr_mask, cur_state, new_state);
+
+			err = -EOPNOTSUPP;
+			goto out;
+		}
+	}
+
+	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
+		if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
+			if ((ibqp->qp_type == IB_QPT_RC) ||
+			    (ibqp->qp_type == IB_QPT_UD) ||
+			    (ibqp->qp_type == IB_QPT_UC) ||
+			    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
+			    (ibqp->qp_type == IB_QPT_XRC_INI)) {
+				attr->port_num = mlx4_ib_bond_next_port(dev);
+			}
+		} else {
+			/* no sense in changing port_num
+			 * when ports are bonded */
+			attr_mask &= ~IB_QP_PORT;
+		}
+	}
+
+	if ((attr_mask & IB_QP_PORT) &&
+	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
+		pr_debug("qpn 0x%x: invalid port number (%d) specified "
+			 "for transition %d to %d. qp_type %d\n",
+			 ibqp->qp_num, attr->port_num, cur_state,
+			 new_state, ibqp->qp_type);
+		goto out;
+	}
+
+	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
+	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
+	     IB_LINK_LAYER_ETHERNET))
+		goto out;
+
+	if (attr_mask & IB_QP_PKEY_INDEX) {
+		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
+		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
+			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
+				 "for transition %d to %d. qp_type %d\n",
+				 ibqp->qp_num, attr->pkey_index, cur_state,
+				 new_state, ibqp->qp_type);
+			goto out;
+		}
+	}
+
+	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
+	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
+		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
+			 "Transition %d to %d. qp_type %d\n",
+			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
+			 new_state, ibqp->qp_type);
+		goto out;
+	}
+
+	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
+	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
+		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
+			 "Transition %d to %d. qp_type %d\n",
+			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
+			 new_state, ibqp->qp_type);
+		goto out;
+	}
+
+	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
+		err = 0;
+		goto out;
+	}
+
+	if (ibqp->rwq_ind_tbl && (new_state == IB_QPS_INIT)) {
+		err = bringup_rss_rwqs(ibqp->rwq_ind_tbl, attr->port_num,
+				       udata);
+		if (err)
+			goto out;
+	}
+
+	err = __mlx4_ib_modify_qp(ibqp, MLX4_IB_QP_SRC, attr, attr_mask,
+				  cur_state, new_state, udata);
+
+	if (ibqp->rwq_ind_tbl && err)
+		bring_down_rss_rwqs(ibqp->rwq_ind_tbl, udata);
+
+	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
+		attr->port_num = 1;
+
+out:
+	mutex_unlock(&qp->mutex);
+	return err;
+}
+
+int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
+		      int attr_mask, struct ib_udata *udata)
+{
+	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
+	int ret;
+
+	ret = _mlx4_ib_modify_qp(ibqp, attr, attr_mask, udata);
+
+	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
+		struct mlx4_ib_sqp *sqp = mqp->sqp;
+		int err = 0;
+
+		if (sqp->roce_v2_gsi)
+			err = ib_modify_qp(sqp->roce_v2_gsi, attr, attr_mask);
+		if (err)
+			pr_err("Failed to modify GSI QP for RoCEv2 (%d)\n",
+			       err);
+	}
+	return ret;
+}
+
+static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
+{
+	int i;
+	for (i = 0; i < dev->caps.num_ports; i++) {
+		if (qpn == dev->caps.spec_qps[i].qp0_proxy ||
+		    qpn == dev->caps.spec_qps[i].qp0_tunnel) {
+			*qkey = dev->caps.spec_qps[i].qp0_qkey;
+			return 0;
+		}
+	}
+	return -EINVAL;
+}
+
+static int build_sriov_qp0_header(struct mlx4_ib_qp *qp,
+				  const struct ib_ud_wr *wr,
+				  void *wqe, unsigned *mlx_seg_len)
+{
+	struct mlx4_ib_dev *mdev = to_mdev(qp->ibqp.device);
+	struct mlx4_ib_sqp *sqp = qp->sqp;
+	struct ib_device *ib_dev = qp->ibqp.device;
+	struct mlx4_wqe_mlx_seg *mlx = wqe;
+	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
+	struct mlx4_ib_ah *ah = to_mah(wr->ah);
+	u16 pkey;
+	u32 qkey;
+	int send_size;
+	int header_size;
+	int spc;
+	int err;
+	int i;
+
+	if (wr->wr.opcode != IB_WR_SEND)
+		return -EINVAL;
+
+	send_size = 0;
+
+	for (i = 0; i < wr->wr.num_sge; ++i)
+		send_size += wr->wr.sg_list[i].length;
+
+	/* for proxy-qp0 sends, need to add in size of tunnel header */
+	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
+	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
+		send_size += sizeof (struct mlx4_ib_tunnel_header);
+
+	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, 0, 0, &sqp->ud_header);
+
+	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
+		sqp->ud_header.lrh.service_level =
+			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
+		sqp->ud_header.lrh.destination_lid =
+			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
+		sqp->ud_header.lrh.source_lid =
+			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
+	}
+
+	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
+
+	/* force loopback */
+	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
+	mlx->rlid = sqp->ud_header.lrh.destination_lid;
+
+	sqp->ud_header.lrh.virtual_lane    = 0;
+	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
+	err = ib_get_cached_pkey(ib_dev, qp->port, 0, &pkey);
+	if (err)
+		return err;
+	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
+	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
+		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
+	else
+		sqp->ud_header.bth.destination_qpn =
+			cpu_to_be32(mdev->dev->caps.spec_qps[qp->port - 1].qp0_tunnel);
+
+	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
+	if (mlx4_is_master(mdev->dev)) {
+		if (mlx4_get_parav_qkey(mdev->dev, qp->mqp.qpn, &qkey))
+			return -EINVAL;
+	} else {
+		if (vf_get_qp0_qkey(mdev->dev, qp->mqp.qpn, &qkey))
+			return -EINVAL;
+	}
+	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
+	sqp->ud_header.deth.source_qpn = cpu_to_be32(qp->mqp.qpn);
+
+	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
+	sqp->ud_header.immediate_present = 0;
+
+	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
+
+	/*
+	 * Inline data segments may not cross a 64 byte boundary.  If
+	 * our UD header is bigger than the space available up to the
+	 * next 64 byte boundary in the WQE, use two inline data
+	 * segments to hold the UD header.
+	 */
+	spc = MLX4_INLINE_ALIGN -
+	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
+	if (header_size <= spc) {
+		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
+		memcpy(inl + 1, sqp->header_buf, header_size);
+		i = 1;
+	} else {
+		inl->byte_count = cpu_to_be32(1 << 31 | spc);
+		memcpy(inl + 1, sqp->header_buf, spc);
+
+		inl = (void *) (inl + 1) + spc;
+		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
+		/*
+		 * Need a barrier here to make sure all the data is
+		 * visible before the byte_count field is set.
+		 * Otherwise the HCA prefetcher could grab the 64-byte
+		 * chunk with this inline segment and get a valid (!=
+		 * 0xffffffff) byte count but stale data, and end up
+		 * generating a packet with bad headers.
+		 *
+		 * The first inline segment's byte_count field doesn't
+		 * need a barrier, because it comes after a
+		 * control/MLX segment and therefore is at an offset
+		 * of 16 mod 64.
+		 */
+		wmb();
+		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
+		i = 2;
+	}
+
+	*mlx_seg_len =
+	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
+	return 0;
+}
+
+static u8 sl_to_vl(struct mlx4_ib_dev *dev, u8 sl, int port_num)
+{
+	union sl2vl_tbl_to_u64 tmp_vltab;
+	u8 vl;
+
+	if (sl > 15)
+		return 0xf;
+	tmp_vltab.sl64 = atomic64_read(&dev->sl2vl[port_num - 1]);
+	vl = tmp_vltab.sl8[sl >> 1];
+	if (sl & 1)
+		vl &= 0x0f;
+	else
+		vl >>= 4;
+	return vl;
+}
+
+static int fill_gid_by_hw_index(struct mlx4_ib_dev *ibdev, u8 port_num,
+				int index, union ib_gid *gid,
+				enum ib_gid_type *gid_type)
+{
+	struct mlx4_ib_iboe *iboe = &ibdev->iboe;
+	struct mlx4_port_gid_table *port_gid_table;
+	unsigned long flags;
+
+	port_gid_table = &iboe->gids[port_num - 1];
+	spin_lock_irqsave(&iboe->lock, flags);
+	memcpy(gid, &port_gid_table->gids[index].gid, sizeof(*gid));
+	*gid_type = port_gid_table->gids[index].gid_type;
+	spin_unlock_irqrestore(&iboe->lock, flags);
+	if (rdma_is_zero_gid(gid))
+		return -ENOENT;
+
+	return 0;
+}
+
+#define MLX4_ROCEV2_QP1_SPORT 0xC000
+static int build_mlx_header(struct mlx4_ib_qp *qp, const struct ib_ud_wr *wr,
+			    void *wqe, unsigned *mlx_seg_len)
+{
+	struct mlx4_ib_sqp *sqp = qp->sqp;
+	struct ib_device *ib_dev = qp->ibqp.device;
+	struct mlx4_ib_dev *ibdev = to_mdev(ib_dev);
+	struct mlx4_wqe_mlx_seg *mlx = wqe;
+	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
+	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
+	struct mlx4_ib_ah *ah = to_mah(wr->ah);
+	union ib_gid sgid;
+	u16 pkey;
+	int send_size;
+	int header_size;
+	int spc;
+	int i;
+	int err = 0;
+	u16 vlan = 0xffff;
+	bool is_eth;
+	bool is_vlan = false;
+	bool is_grh;
+	bool is_udp = false;
+	int ip_version = 0;
+
+	send_size = 0;
+	for (i = 0; i < wr->wr.num_sge; ++i)
+		send_size += wr->wr.sg_list[i].length;
+
+	is_eth = rdma_port_get_link_layer(qp->ibqp.device, qp->port) == IB_LINK_LAYER_ETHERNET;
+	is_grh = mlx4_ib_ah_grh_present(ah);
+	if (is_eth) {
+		enum ib_gid_type gid_type;
+		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
+			/* When multi-function is enabled, the ib_core gid
+			 * indexes don't necessarily match the hw ones, so
+			 * we must use our own cache */
+			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
+							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
+							   ah->av.ib.gid_index, &sgid.raw[0]);
+			if (err)
+				return err;
+		} else  {
+			err = fill_gid_by_hw_index(ibdev, qp->port,
+						   ah->av.ib.gid_index, &sgid,
+						   &gid_type);
+			if (!err) {
+				is_udp = gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
+				if (is_udp) {
+					if (ipv6_addr_v4mapped((struct in6_addr *)&sgid))
+						ip_version = 4;
+					else
+						ip_version = 6;
+					is_grh = false;
+				}
+			} else {
+				return err;
+			}
+		}
+		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
+			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
+			is_vlan = true;
+		}
+	}
+	err = ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh,
+			  ip_version, is_udp, 0, &sqp->ud_header);
+	if (err)
+		return err;
+
+	if (!is_eth) {
+		sqp->ud_header.lrh.service_level =
+			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
+		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
+		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
+	}
+
+	if (is_grh || (ip_version == 6)) {
+		sqp->ud_header.grh.traffic_class =
+			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
+		sqp->ud_header.grh.flow_label    =
+			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
+		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
+		if (is_eth) {
+			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
+		} else {
+			if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
+				/* When multi-function is enabled, the ib_core gid
+				 * indexes don't necessarily match the hw ones, so
+				 * we must use our own cache
+				 */
+				sqp->ud_header.grh.source_gid.global
+					.subnet_prefix =
+					cpu_to_be64(atomic64_read(
+						&(to_mdev(ib_dev)
+							  ->sriov
+							  .demux[qp->port - 1]
+							  .subnet_prefix)));
+				sqp->ud_header.grh.source_gid.global
+					.interface_id =
+					to_mdev(ib_dev)
+						->sriov.demux[qp->port - 1]
+						.guid_cache[ah->av.ib.gid_index];
+			} else {
+				sqp->ud_header.grh.source_gid =
+					ah->ibah.sgid_attr->gid;
+			}
+		}
+		memcpy(sqp->ud_header.grh.destination_gid.raw,
+		       ah->av.ib.dgid, 16);
+	}
+
+	if (ip_version == 4) {
+		sqp->ud_header.ip4.tos =
+			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
+		sqp->ud_header.ip4.id = 0;
+		sqp->ud_header.ip4.frag_off = htons(IP_DF);
+		sqp->ud_header.ip4.ttl = ah->av.eth.hop_limit;
+
+		memcpy(&sqp->ud_header.ip4.saddr,
+		       sgid.raw + 12, 4);
+		memcpy(&sqp->ud_header.ip4.daddr, ah->av.ib.dgid + 12, 4);
+		sqp->ud_header.ip4.check = ib_ud_ip4_csum(&sqp->ud_header);
+	}
+
+	if (is_udp) {
+		sqp->ud_header.udp.dport = htons(ROCE_V2_UDP_DPORT);
+		sqp->ud_header.udp.sport = htons(MLX4_ROCEV2_QP1_SPORT);
+		sqp->ud_header.udp.csum = 0;
+	}
+
+	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
+
+	if (!is_eth) {
+		mlx->flags |=
+			cpu_to_be32((!qp->ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
+				    (sqp->ud_header.lrh.destination_lid ==
+						     IB_LID_PERMISSIVE ?
+					     MLX4_WQE_MLX_SLR :
+					     0) |
+				    (sqp->ud_header.lrh.service_level << 8));
+		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
+			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
+		mlx->rlid = sqp->ud_header.lrh.destination_lid;
+	}
+
+	switch (wr->wr.opcode) {
+	case IB_WR_SEND:
+		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
+		sqp->ud_header.immediate_present = 0;
+		break;
+	case IB_WR_SEND_WITH_IMM:
+		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
+		sqp->ud_header.immediate_present = 1;
+		sqp->ud_header.immediate_data    = wr->wr.ex.imm_data;
+		break;
+	default:
+		return -EINVAL;
+	}
+
+	if (is_eth) {
+		struct in6_addr in6;
+		u16 ether_type;
+		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
+
+		ether_type = (!is_udp) ? ETH_P_IBOE:
+			(ip_version == 4 ? ETH_P_IP : ETH_P_IPV6);
+
+		mlx->sched_prio = cpu_to_be16(pcp);
+
+		ether_addr_copy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac);
+		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
+		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
+		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
+		memcpy(&in6, sgid.raw, sizeof(in6));
+
+
+		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
+			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
+		if (!is_vlan) {
+			sqp->ud_header.eth.type = cpu_to_be16(ether_type);
+		} else {
+			sqp->ud_header.vlan.type = cpu_to_be16(ether_type);
+			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
+		}
+	} else {
+		sqp->ud_header.lrh.virtual_lane =
+			!qp->ibqp.qp_num ?
+				15 :
+				sl_to_vl(to_mdev(ib_dev),
+					 sqp->ud_header.lrh.service_level,
+					 qp->port);
+		if (qp->ibqp.qp_num && sqp->ud_header.lrh.virtual_lane == 15)
+			return -EINVAL;
+		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
+			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
+	}
+	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
+	if (!qp->ibqp.qp_num)
+		err = ib_get_cached_pkey(ib_dev, qp->port, sqp->pkey_index,
+					 &pkey);
+	else
+		err = ib_get_cached_pkey(ib_dev, qp->port, wr->pkey_index,
+					 &pkey);
+	if (err)
+		return err;
+
+	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
+	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
+	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
+	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
+					       sqp->qkey : wr->remote_qkey);
+	sqp->ud_header.deth.source_qpn = cpu_to_be32(qp->ibqp.qp_num);
+
+	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
+
+	if (0) {
+		pr_err("built UD header of size %d:\n", header_size);
+		for (i = 0; i < header_size / 4; ++i) {
+			if (i % 8 == 0)
+				pr_err("  [%02x] ", i * 4);
+			pr_cont(" %08x",
+				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
+			if ((i + 1) % 8 == 0)
+				pr_cont("\n");
+		}
+		pr_err("\n");
+	}
+
+	/*
+	 * Inline data segments may not cross a 64 byte boundary.  If
+	 * our UD header is bigger than the space available up to the
+	 * next 64 byte boundary in the WQE, use two inline data
+	 * segments to hold the UD header.
+	 */
+	spc = MLX4_INLINE_ALIGN -
+		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
+	if (header_size <= spc) {
+		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
+		memcpy(inl + 1, sqp->header_buf, header_size);
+		i = 1;
+	} else {
+		inl->byte_count = cpu_to_be32(1 << 31 | spc);
+		memcpy(inl + 1, sqp->header_buf, spc);
+
+		inl = (void *) (inl + 1) + spc;
+		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
+		/*
+		 * Need a barrier here to make sure all the data is
+		 * visible before the byte_count field is set.
+		 * Otherwise the HCA prefetcher could grab the 64-byte
+		 * chunk with this inline segment and get a valid (!=
+		 * 0xffffffff) byte count but stale data, and end up
+		 * generating a packet with bad headers.
+		 *
+		 * The first inline segment's byte_count field doesn't
+		 * need a barrier, because it comes after a
+		 * control/MLX segment and therefore is at an offset
+		 * of 16 mod 64.
+		 */
+		wmb();
+		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
+		i = 2;
+	}
+
+	*mlx_seg_len =
+		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
+	return 0;
+}
+
+static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
+{
+	unsigned cur;
+	struct mlx4_ib_cq *cq;
+
+	cur = wq->head - wq->tail;
+	if (likely(cur + nreq < wq->max_post))
+		return 0;
+
+	cq = to_mcq(ib_cq);
+	spin_lock(&cq->lock);
+	cur = wq->head - wq->tail;
+	spin_unlock(&cq->lock);
+
+	return cur + nreq >= wq->max_post;
+}
+
+static __be32 convert_access(int acc)
+{
+	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
+		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
+	       (acc & IB_ACCESS_REMOTE_WRITE  ?
+		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
+	       (acc & IB_ACCESS_REMOTE_READ   ?
+		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
+	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
+		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
+}
+
+static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg,
+			const struct ib_reg_wr *wr)
+{
+	struct mlx4_ib_mr *mr = to_mmr(wr->mr);
+
+	fseg->flags		= convert_access(wr->access);
+	fseg->mem_key		= cpu_to_be32(wr->key);
+	fseg->buf_list		= cpu_to_be64(mr->page_map);
+	fseg->start_addr	= cpu_to_be64(mr->ibmr.iova);
+	fseg->reg_len		= cpu_to_be64(mr->ibmr.length);
+	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
+	fseg->page_size		= cpu_to_be32(ilog2(mr->ibmr.page_size));
+	fseg->reserved[0]	= 0;
+	fseg->reserved[1]	= 0;
+}
+
+static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
+{
+	memset(iseg, 0, sizeof(*iseg));
+	iseg->mem_key = cpu_to_be32(rkey);
+}
+
+static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
+					  u64 remote_addr, u32 rkey)
+{
+	rseg->raddr    = cpu_to_be64(remote_addr);
+	rseg->rkey     = cpu_to_be32(rkey);
+	rseg->reserved = 0;
+}
+
+static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg,
+			   const struct ib_atomic_wr *wr)
+{
+	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
+		aseg->swap_add = cpu_to_be64(wr->swap);
+		aseg->compare  = cpu_to_be64(wr->compare_add);
+	} else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
+		aseg->swap_add = cpu_to_be64(wr->compare_add);
+		aseg->compare  = cpu_to_be64(wr->compare_add_mask);
+	} else {
+		aseg->swap_add = cpu_to_be64(wr->compare_add);
+		aseg->compare  = 0;
+	}
+
+}
+
+static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
+				  const struct ib_atomic_wr *wr)
+{
+	aseg->swap_add		= cpu_to_be64(wr->swap);
+	aseg->swap_add_mask	= cpu_to_be64(wr->swap_mask);
+	aseg->compare		= cpu_to_be64(wr->compare_add);
+	aseg->compare_mask	= cpu_to_be64(wr->compare_add_mask);
+}
+
+static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
+			     const struct ib_ud_wr *wr)
+{
+	memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av));
+	dseg->dqpn = cpu_to_be32(wr->remote_qpn);
+	dseg->qkey = cpu_to_be32(wr->remote_qkey);
+	dseg->vlan = to_mah(wr->ah)->av.eth.vlan;
+	memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6);
+}
+
+static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
+				    struct mlx4_wqe_datagram_seg *dseg,
+				    const struct ib_ud_wr *wr,
+				    enum mlx4_ib_qp_type qpt)
+{
+	union mlx4_ext_av *av = &to_mah(wr->ah)->av;
+	struct mlx4_av sqp_av = {0};
+	int port = *((u8 *) &av->ib.port_pd) & 0x3;
+
+	/* force loopback */
+	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
+	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
+	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
+			cpu_to_be32(0xf0000000);
+
+	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
+	if (qpt == MLX4_IB_QPT_PROXY_GSI)
+		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp1_tunnel);
+	else
+		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp0_tunnel);
+	/* Use QKEY from the QP context, which is set by master */
+	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
+}
+
+static void build_tunnel_header(const struct ib_ud_wr *wr, void *wqe,
+				unsigned *mlx_seg_len)
+{
+	struct mlx4_wqe_inline_seg *inl = wqe;
+	struct mlx4_ib_tunnel_header hdr;
+	struct mlx4_ib_ah *ah = to_mah(wr->ah);
+	int spc;
+	int i;
+
+	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
+	hdr.remote_qpn = cpu_to_be32(wr->remote_qpn);
+	hdr.pkey_index = cpu_to_be16(wr->pkey_index);
+	hdr.qkey = cpu_to_be32(wr->remote_qkey);
+	memcpy(hdr.mac, ah->av.eth.mac, 6);
+	hdr.vlan = ah->av.eth.vlan;
+
+	spc = MLX4_INLINE_ALIGN -
+		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
+	if (sizeof (hdr) <= spc) {
+		memcpy(inl + 1, &hdr, sizeof (hdr));
+		wmb();
+		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
+		i = 1;
+	} else {
+		memcpy(inl + 1, &hdr, spc);
+		wmb();
+		inl->byte_count = cpu_to_be32(1 << 31 | spc);
+
+		inl = (void *) (inl + 1) + spc;
+		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
+		wmb();
+		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
+		i = 2;
+	}
+
+	*mlx_seg_len =
+		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
+}
+
+static void set_mlx_icrc_seg(void *dseg)
+{
+	u32 *t = dseg;
+	struct mlx4_wqe_inline_seg *iseg = dseg;
+
+	t[1] = 0;
+
+	/*
+	 * Need a barrier here before writing the byte_count field to
+	 * make sure that all the data is visible before the
+	 * byte_count field is set.  Otherwise, if the segment begins
+	 * a new cacheline, the HCA prefetcher could grab the 64-byte
+	 * chunk and get a valid (!= * 0xffffffff) byte count but
+	 * stale data, and end up sending the wrong data.
+	 */
+	wmb();
+
+	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
+}
+
+static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
+{
+	dseg->lkey       = cpu_to_be32(sg->lkey);
+	dseg->addr       = cpu_to_be64(sg->addr);
+
+	/*
+	 * Need a barrier here before writing the byte_count field to
+	 * make sure that all the data is visible before the
+	 * byte_count field is set.  Otherwise, if the segment begins
+	 * a new cacheline, the HCA prefetcher could grab the 64-byte
+	 * chunk and get a valid (!= * 0xffffffff) byte count but
+	 * stale data, and end up sending the wrong data.
+	 */
+	wmb();
+
+	dseg->byte_count = cpu_to_be32(sg->length);
+}
+
+static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
+{
+	dseg->byte_count = cpu_to_be32(sg->length);
+	dseg->lkey       = cpu_to_be32(sg->lkey);
+	dseg->addr       = cpu_to_be64(sg->addr);
+}
+
+static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe,
+			 const struct ib_ud_wr *wr, struct mlx4_ib_qp *qp,
+			 unsigned *lso_seg_len, __be32 *lso_hdr_sz, __be32 *blh)
+{
+	unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16);
+
+	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
+		*blh = cpu_to_be32(1 << 6);
+
+	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
+		     wr->wr.num_sge > qp->sq.max_gs - (halign >> 4)))
+		return -EINVAL;
+
+	memcpy(wqe->header, wr->header, wr->hlen);
+
+	*lso_hdr_sz  = cpu_to_be32(wr->mss << 16 | wr->hlen);
+	*lso_seg_len = halign;
+	return 0;
+}
+
+static __be32 send_ieth(const struct ib_send_wr *wr)
+{
+	switch (wr->opcode) {
+	case IB_WR_SEND_WITH_IMM:
+	case IB_WR_RDMA_WRITE_WITH_IMM:
+		return wr->ex.imm_data;
+
+	case IB_WR_SEND_WITH_INV:
+		return cpu_to_be32(wr->ex.invalidate_rkey);
+
+	default:
+		return 0;
+	}
+}
+
+static void add_zero_len_inline(void *wqe)
+{
+	struct mlx4_wqe_inline_seg *inl = wqe;
+	memset(wqe, 0, 16);
+	inl->byte_count = cpu_to_be32(1 << 31);
+}
+
+static int _mlx4_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
+			      const struct ib_send_wr **bad_wr, bool drain)
+{
+	struct mlx4_ib_qp *qp = to_mqp(ibqp);
+	void *wqe;
+	struct mlx4_wqe_ctrl_seg *ctrl;
+	struct mlx4_wqe_data_seg *dseg;
+	unsigned long flags;
+	int nreq;
+	int err = 0;
+	unsigned ind;
+	int size;
+	unsigned seglen;
+	__be32 dummy;
+	__be32 *lso_wqe;
+	__be32 lso_hdr_sz;
+	__be32 blh;
+	int i;
+	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
+
+	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
+		struct mlx4_ib_sqp *sqp = qp->sqp;
+
+		if (sqp->roce_v2_gsi) {
+			struct mlx4_ib_ah *ah = to_mah(ud_wr(wr)->ah);
+			enum ib_gid_type gid_type;
+			union ib_gid gid;
+
+			if (!fill_gid_by_hw_index(mdev, qp->port,
+					   ah->av.ib.gid_index,
+					   &gid, &gid_type))
+				qp = (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ?
+						to_mqp(sqp->roce_v2_gsi) : qp;
+			else
+				pr_err("Failed to get gid at index %d. RoCEv2 will not work properly\n",
+				       ah->av.ib.gid_index);
+		}
+	}
+
+	spin_lock_irqsave(&qp->sq.lock, flags);
+	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR &&
+	    !drain) {
+		err = -EIO;
+		*bad_wr = wr;
+		nreq = 0;
+		goto out;
+	}
+
+	ind = qp->sq_next_wqe;
+
+	for (nreq = 0; wr; ++nreq, wr = wr->next) {
+		lso_wqe = &dummy;
+		blh = 0;
+
+		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
+			err = -ENOMEM;
+			*bad_wr = wr;
+			goto out;
+		}
+
+		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
+			err = -EINVAL;
+			*bad_wr = wr;
+			goto out;
+		}
+
+		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
+		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
+
+		ctrl->srcrb_flags =
+			(wr->send_flags & IB_SEND_SIGNALED ?
+			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
+			(wr->send_flags & IB_SEND_SOLICITED ?
+			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
+			((wr->send_flags & IB_SEND_IP_CSUM) ?
+			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
+				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
+			qp->sq_signal_bits;
+
+		ctrl->imm = send_ieth(wr);
+
+		wqe += sizeof *ctrl;
+		size = sizeof *ctrl / 16;
+
+		switch (qp->mlx4_ib_qp_type) {
+		case MLX4_IB_QPT_RC:
+		case MLX4_IB_QPT_UC:
+			switch (wr->opcode) {
+			case IB_WR_ATOMIC_CMP_AND_SWP:
+			case IB_WR_ATOMIC_FETCH_AND_ADD:
+			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
+				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
+					      atomic_wr(wr)->rkey);
+				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
+
+				set_atomic_seg(wqe, atomic_wr(wr));
+				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
+
+				size += (sizeof (struct mlx4_wqe_raddr_seg) +
+					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
+
+				break;
+
+			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
+				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
+					      atomic_wr(wr)->rkey);
+				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
+
+				set_masked_atomic_seg(wqe, atomic_wr(wr));
+				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
+
+				size += (sizeof (struct mlx4_wqe_raddr_seg) +
+					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
+
+				break;
+
+			case IB_WR_RDMA_READ:
+			case IB_WR_RDMA_WRITE:
+			case IB_WR_RDMA_WRITE_WITH_IMM:
+				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
+					      rdma_wr(wr)->rkey);
+				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
+				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
+				break;
+
+			case IB_WR_LOCAL_INV:
+				ctrl->srcrb_flags |=
+					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
+				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
+				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
+				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
+				break;
+
+			case IB_WR_REG_MR:
+				ctrl->srcrb_flags |=
+					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
+				set_reg_seg(wqe, reg_wr(wr));
+				wqe  += sizeof(struct mlx4_wqe_fmr_seg);
+				size += sizeof(struct mlx4_wqe_fmr_seg) / 16;
+				break;
+
+			default:
+				/* No extra segments required for sends */
+				break;
+			}
+			break;
+
+		case MLX4_IB_QPT_TUN_SMI_OWNER:
+			err = build_sriov_qp0_header(qp, ud_wr(wr), ctrl,
+						     &seglen);
+			if (unlikely(err)) {
+				*bad_wr = wr;
+				goto out;
+			}
+			wqe  += seglen;
+			size += seglen / 16;
+			break;
+		case MLX4_IB_QPT_TUN_SMI:
+		case MLX4_IB_QPT_TUN_GSI:
+			/* this is a UD qp used in MAD responses to slaves. */
+			set_datagram_seg(wqe, ud_wr(wr));
+			/* set the forced-loopback bit in the data seg av */
+			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
+			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
+			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
+			break;
+		case MLX4_IB_QPT_UD:
+			set_datagram_seg(wqe, ud_wr(wr));
+			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
+			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
+
+			if (wr->opcode == IB_WR_LSO) {
+				err = build_lso_seg(wqe, ud_wr(wr), qp, &seglen,
+						&lso_hdr_sz, &blh);
+				if (unlikely(err)) {
+					*bad_wr = wr;
+					goto out;
+				}
+				lso_wqe = (__be32 *) wqe;
+				wqe  += seglen;
+				size += seglen / 16;
+			}
+			break;
+
+		case MLX4_IB_QPT_PROXY_SMI_OWNER:
+			err = build_sriov_qp0_header(qp, ud_wr(wr), ctrl,
+						     &seglen);
+			if (unlikely(err)) {
+				*bad_wr = wr;
+				goto out;
+			}
+			wqe  += seglen;
+			size += seglen / 16;
+			/* to start tunnel header on a cache-line boundary */
+			add_zero_len_inline(wqe);
+			wqe += 16;
+			size++;
+			build_tunnel_header(ud_wr(wr), wqe, &seglen);
+			wqe  += seglen;
+			size += seglen / 16;
+			break;
+		case MLX4_IB_QPT_PROXY_SMI:
+		case MLX4_IB_QPT_PROXY_GSI:
+			/* If we are tunneling special qps, this is a UD qp.
+			 * In this case we first add a UD segment targeting
+			 * the tunnel qp, and then add a header with address
+			 * information */
+			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe,
+						ud_wr(wr),
+						qp->mlx4_ib_qp_type);
+			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
+			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
+			build_tunnel_header(ud_wr(wr), wqe, &seglen);
+			wqe  += seglen;
+			size += seglen / 16;
+			break;
+
+		case MLX4_IB_QPT_SMI:
+		case MLX4_IB_QPT_GSI:
+			err = build_mlx_header(qp, ud_wr(wr), ctrl, &seglen);
+			if (unlikely(err)) {
+				*bad_wr = wr;
+				goto out;
+			}
+			wqe  += seglen;
+			size += seglen / 16;
+			break;
+
+		default:
+			break;
+		}
+
+		/*
+		 * Write data segments in reverse order, so as to
+		 * overwrite cacheline stamp last within each
+		 * cacheline.  This avoids issues with WQE
+		 * prefetching.
+		 */
+
+		dseg = wqe;
+		dseg += wr->num_sge - 1;
+		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
+
+		/* Add one more inline data segment for ICRC for MLX sends */
+		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
+			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
+			     qp->mlx4_ib_qp_type &
+			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
+			set_mlx_icrc_seg(dseg + 1);
+			size += sizeof (struct mlx4_wqe_data_seg) / 16;
+		}
+
+		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
+			set_data_seg(dseg, wr->sg_list + i);
+
+		/*
+		 * Possibly overwrite stamping in cacheline with LSO
+		 * segment only after making sure all data segments
+		 * are written.
+		 */
+		wmb();
+		*lso_wqe = lso_hdr_sz;
+
+		ctrl->qpn_vlan.fence_size = (wr->send_flags & IB_SEND_FENCE ?
+					     MLX4_WQE_CTRL_FENCE : 0) | size;
+
+		/*
+		 * Make sure descriptor is fully written before
+		 * setting ownership bit (because HW can start
+		 * executing as soon as we do).
+		 */
+		wmb();
+
+		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
+			*bad_wr = wr;
+			err = -EINVAL;
+			goto out;
+		}
+
+		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
+			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
+
+		/*
+		 * We can improve latency by not stamping the last
+		 * send queue WQE until after ringing the doorbell, so
+		 * only stamp here if there are still more WQEs to post.
+		 */
+		if (wr->next)
+			stamp_send_wqe(qp, ind + qp->sq_spare_wqes);
+		ind++;
+	}
+
+out:
+	if (likely(nreq)) {
+		qp->sq.head += nreq;
+
+		/*
+		 * Make sure that descriptors are written before
+		 * doorbell record.
+		 */
+		wmb();
+
+		writel_relaxed(qp->doorbell_qpn,
+			to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
+
+		stamp_send_wqe(qp, ind + qp->sq_spare_wqes - 1);
+
+		qp->sq_next_wqe = ind;
+	}
+
+	spin_unlock_irqrestore(&qp->sq.lock, flags);
+
+	return err;
+}
+
+int mlx4_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
+		      const struct ib_send_wr **bad_wr)
+{
+	return _mlx4_ib_post_send(ibqp, wr, bad_wr, false);
+}
+
+static int _mlx4_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
+			      const struct ib_recv_wr **bad_wr, bool drain)
+{
+	struct mlx4_ib_qp *qp = to_mqp(ibqp);
+	struct mlx4_wqe_data_seg *scat;
+	unsigned long flags;
+	int err = 0;
+	int nreq;
+	int ind;
+	int max_gs;
+	int i;
+	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
+
+	max_gs = qp->rq.max_gs;
+	spin_lock_irqsave(&qp->rq.lock, flags);
+
+	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR &&
+	    !drain) {
+		err = -EIO;
+		*bad_wr = wr;
+		nreq = 0;
+		goto out;
+	}
+
+	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
+
+	for (nreq = 0; wr; ++nreq, wr = wr->next) {
+		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
+			err = -ENOMEM;
+			*bad_wr = wr;
+			goto out;
+		}
+
+		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
+			err = -EINVAL;
+			*bad_wr = wr;
+			goto out;
+		}
+
+		scat = get_recv_wqe(qp, ind);
+
+		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
+		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
+			ib_dma_sync_single_for_device(ibqp->device,
+						      qp->sqp_proxy_rcv[ind].map,
+						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
+						      DMA_FROM_DEVICE);
+			scat->byte_count =
+				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
+			/* use dma lkey from upper layer entry */
+			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
+			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
+			scat++;
+			max_gs--;
+		}
+
+		for (i = 0; i < wr->num_sge; ++i)
+			__set_data_seg(scat + i, wr->sg_list + i);
+
+		if (i < max_gs) {
+			scat[i].byte_count = 0;
+			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
+			scat[i].addr       = 0;
+		}
+
+		qp->rq.wrid[ind] = wr->wr_id;
+
+		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
+	}
+
+out:
+	if (likely(nreq)) {
+		qp->rq.head += nreq;
+
+		/*
+		 * Make sure that descriptors are written before
+		 * doorbell record.
+		 */
+		wmb();
+
+		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
+	}
+
+	spin_unlock_irqrestore(&qp->rq.lock, flags);
+
+	return err;
+}
+
+int mlx4_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
+		      const struct ib_recv_wr **bad_wr)
+{
+	return _mlx4_ib_post_recv(ibqp, wr, bad_wr, false);
+}
+
+static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
+{
+	switch (mlx4_state) {
+	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
+	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
+	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
+	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
+	case MLX4_QP_STATE_SQ_DRAINING:
+	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
+	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
+	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
+	default:		     return -1;
+	}
+}
+
+static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
+{
+	switch (mlx4_mig_state) {
+	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
+	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
+	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
+	default: return -1;
+	}
+}
+
+static int to_ib_qp_access_flags(int mlx4_flags)
+{
+	int ib_flags = 0;
+
+	if (mlx4_flags & MLX4_QP_BIT_RRE)
+		ib_flags |= IB_ACCESS_REMOTE_READ;
+	if (mlx4_flags & MLX4_QP_BIT_RWE)
+		ib_flags |= IB_ACCESS_REMOTE_WRITE;
+	if (mlx4_flags & MLX4_QP_BIT_RAE)
+		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
+
+	return ib_flags;
+}
+
+static void to_rdma_ah_attr(struct mlx4_ib_dev *ibdev,
+			    struct rdma_ah_attr *ah_attr,
+			    struct mlx4_qp_path *path)
+{
+	struct mlx4_dev *dev = ibdev->dev;
+	u8 port_num = path->sched_queue & 0x40 ? 2 : 1;
+
+	memset(ah_attr, 0, sizeof(*ah_attr));
+	if (port_num == 0 || port_num > dev->caps.num_ports)
+		return;
+	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port_num);
+
+	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE)
+		rdma_ah_set_sl(ah_attr, ((path->sched_queue >> 3) & 0x7) |
+			       ((path->sched_queue & 4) << 1));
+	else
+		rdma_ah_set_sl(ah_attr, (path->sched_queue >> 2) & 0xf);
+	rdma_ah_set_port_num(ah_attr, port_num);
+
+	rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
+	rdma_ah_set_path_bits(ah_attr, path->grh_mylmc & 0x7f);
+	rdma_ah_set_static_rate(ah_attr,
+				path->static_rate ? path->static_rate - 5 : 0);
+	if (path->grh_mylmc & (1 << 7)) {
+		rdma_ah_set_grh(ah_attr, NULL,
+				be32_to_cpu(path->tclass_flowlabel) & 0xfffff,
+				path->mgid_index,
+				path->hop_limit,
+				(be32_to_cpu(path->tclass_flowlabel)
+				 >> 20) & 0xff);
+		rdma_ah_set_dgid_raw(ah_attr, path->rgid);
+	}
+}
+
+int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
+		     struct ib_qp_init_attr *qp_init_attr)
+{
+	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
+	struct mlx4_ib_qp *qp = to_mqp(ibqp);
+	struct mlx4_qp_context context;
+	int mlx4_state;
+	int err = 0;
+
+	if (ibqp->rwq_ind_tbl)
+		return -EOPNOTSUPP;
+
+	mutex_lock(&qp->mutex);
+
+	if (qp->state == IB_QPS_RESET) {
+		qp_attr->qp_state = IB_QPS_RESET;
+		goto done;
+	}
+
+	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
+	if (err) {
+		err = -EINVAL;
+		goto out;
+	}
+
+	mlx4_state = be32_to_cpu(context.flags) >> 28;
+
+	qp->state		     = to_ib_qp_state(mlx4_state);
+	qp_attr->qp_state	     = qp->state;
+	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
+	qp_attr->path_mig_state	     =
+		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
+	qp_attr->qkey		     = be32_to_cpu(context.qkey);
+	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
+	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
+	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
+	qp_attr->qp_access_flags     =
+		to_ib_qp_access_flags(be32_to_cpu(context.params2));
+
+	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
+		to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
+		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
+		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
+		qp_attr->alt_port_num	=
+			rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
+	}
+
+	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
+	if (qp_attr->qp_state == IB_QPS_INIT)
+		qp_attr->port_num = qp->port;
+	else
+		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
+
+	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
+	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
+
+	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
+
+	qp_attr->max_dest_rd_atomic =
+		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
+	qp_attr->min_rnr_timer	    =
+		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
+	qp_attr->timeout	    = context.pri_path.ackto >> 3;
+	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
+	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
+	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
+
+done:
+	qp_attr->cur_qp_state	     = qp_attr->qp_state;
+	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
+	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
+
+	if (!ibqp->uobject) {
+		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
+		qp_attr->cap.max_send_sge = qp->sq.max_gs;
+	} else {
+		qp_attr->cap.max_send_wr  = 0;
+		qp_attr->cap.max_send_sge = 0;
+	}
+
+	/*
+	 * We don't support inline sends for kernel QPs (yet), and we
+	 * don't know what userspace's value should be.
+	 */
+	qp_attr->cap.max_inline_data = 0;
+
+	qp_init_attr->cap	     = qp_attr->cap;
+
+	qp_init_attr->create_flags = 0;
+	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
+		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
+
+	if (qp->flags & MLX4_IB_QP_LSO)
+		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
+
+	if (qp->flags & MLX4_IB_QP_NETIF)
+		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
+
+	qp_init_attr->sq_sig_type =
+		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
+		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
+
+out:
+	mutex_unlock(&qp->mutex);
+	return err;
+}
+
+struct ib_wq *mlx4_ib_create_wq(struct ib_pd *pd,
+				struct ib_wq_init_attr *init_attr,
+				struct ib_udata *udata)
+{
+	struct mlx4_dev *dev = to_mdev(pd->device)->dev;
+	struct ib_qp_init_attr ib_qp_init_attr = {};
+	struct mlx4_ib_qp *qp;
+	struct mlx4_ib_create_wq ucmd;
+	int err, required_cmd_sz;
+
+	if (!udata)
+		return ERR_PTR(-EINVAL);
+
+	required_cmd_sz = offsetof(typeof(ucmd), comp_mask) +
+			  sizeof(ucmd.comp_mask);
+	if (udata->inlen < required_cmd_sz) {
+		pr_debug("invalid inlen\n");
+		return ERR_PTR(-EINVAL);
+	}
+
+	if (udata->inlen > sizeof(ucmd) &&
+	    !ib_is_udata_cleared(udata, sizeof(ucmd),
+				 udata->inlen - sizeof(ucmd))) {
+		pr_debug("inlen is not supported\n");
+		return ERR_PTR(-EOPNOTSUPP);
+	}
+
+	if (udata->outlen)
+		return ERR_PTR(-EOPNOTSUPP);
+
+	if (init_attr->wq_type != IB_WQT_RQ) {
+		pr_debug("unsupported wq type %d\n", init_attr->wq_type);
+		return ERR_PTR(-EOPNOTSUPP);
+	}
+
+	if (init_attr->create_flags & ~IB_WQ_FLAGS_SCATTER_FCS ||
+	    !(dev->caps.flags & MLX4_DEV_CAP_FLAG_FCS_KEEP)) {
+		pr_debug("unsupported create_flags %u\n",
+			 init_attr->create_flags);
+		return ERR_PTR(-EOPNOTSUPP);
+	}
+
+	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
+	if (!qp)
+		return ERR_PTR(-ENOMEM);
+
+	mutex_init(&qp->mutex);
+	qp->pri.vid = 0xFFFF;
+	qp->alt.vid = 0xFFFF;
+
+	ib_qp_init_attr.qp_context = init_attr->wq_context;
+	ib_qp_init_attr.qp_type = IB_QPT_RAW_PACKET;
+	ib_qp_init_attr.cap.max_recv_wr = init_attr->max_wr;
+	ib_qp_init_attr.cap.max_recv_sge = init_attr->max_sge;
+	ib_qp_init_attr.recv_cq = init_attr->cq;
+	ib_qp_init_attr.send_cq = ib_qp_init_attr.recv_cq; /* Dummy CQ */
+
+	if (init_attr->create_flags & IB_WQ_FLAGS_SCATTER_FCS)
+		ib_qp_init_attr.create_flags |= IB_QP_CREATE_SCATTER_FCS;
+
+	err = create_rq(pd, &ib_qp_init_attr, udata, qp);
+	if (err) {
+		kfree(qp);
+		return ERR_PTR(err);
+	}
+
+	qp->ibwq.event_handler = init_attr->event_handler;
+	qp->ibwq.wq_num = qp->mqp.qpn;
+	qp->ibwq.state = IB_WQS_RESET;
+
+	return &qp->ibwq;
+}
+
+static int ib_wq2qp_state(enum ib_wq_state state)
+{
+	switch (state) {
+	case IB_WQS_RESET:
+		return IB_QPS_RESET;
+	case IB_WQS_RDY:
+		return IB_QPS_RTR;
+	default:
+		return IB_QPS_ERR;
+	}
+}
+
+static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state,
+			      struct ib_udata *udata)
+{
+	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
+	enum ib_qp_state qp_cur_state;
+	enum ib_qp_state qp_new_state;
+	int attr_mask;
+	int err;
+
+	/* ib_qp.state represents the WQ HW state while ib_wq.state represents
+	 * the WQ logic state.
+	 */
+	qp_cur_state = qp->state;
+	qp_new_state = ib_wq2qp_state(new_state);
+
+	if (ib_wq2qp_state(new_state) == qp_cur_state)
+		return 0;
+
+	if (new_state == IB_WQS_RDY) {
+		struct ib_qp_attr attr = {};
+
+		attr.port_num = qp->port;
+		attr_mask = IB_QP_PORT;
+
+		err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, &attr,
+					  attr_mask, IB_QPS_RESET, IB_QPS_INIT,
+					  udata);
+		if (err) {
+			pr_debug("WQN=0x%06x failed to apply RST->INIT on the HW QP\n",
+				 ibwq->wq_num);
+			return err;
+		}
+
+		qp_cur_state = IB_QPS_INIT;
+	}
+
+	attr_mask = 0;
+	err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL, attr_mask,
+				  qp_cur_state,  qp_new_state, udata);
+
+	if (err && (qp_cur_state == IB_QPS_INIT)) {
+		qp_new_state = IB_QPS_RESET;
+		if (__mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL,
+					attr_mask, IB_QPS_INIT, IB_QPS_RESET,
+					udata)) {
+			pr_warn("WQN=0x%06x failed with reverting HW's resources failure\n",
+				ibwq->wq_num);
+			qp_new_state = IB_QPS_INIT;
+		}
+	}
+
+	qp->state = qp_new_state;
+
+	return err;
+}
+
+int mlx4_ib_modify_wq(struct ib_wq *ibwq, struct ib_wq_attr *wq_attr,
+		      u32 wq_attr_mask, struct ib_udata *udata)
+{
+	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
+	struct mlx4_ib_modify_wq ucmd = {};
+	size_t required_cmd_sz;
+	enum ib_wq_state cur_state, new_state;
+	int err = 0;
+
+	required_cmd_sz = offsetof(typeof(ucmd), reserved) +
+				   sizeof(ucmd.reserved);
+	if (udata->inlen < required_cmd_sz)
+		return -EINVAL;
+
+	if (udata->inlen > sizeof(ucmd) &&
+	    !ib_is_udata_cleared(udata, sizeof(ucmd),
+				 udata->inlen - sizeof(ucmd)))
+		return -EOPNOTSUPP;
+
+	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
+		return -EFAULT;
+
+	if (ucmd.comp_mask || ucmd.reserved)
+		return -EOPNOTSUPP;
+
+	if (wq_attr_mask & IB_WQ_FLAGS)
+		return -EOPNOTSUPP;
+
+	cur_state = wq_attr->curr_wq_state;
+	new_state = wq_attr->wq_state;
+
+	if ((new_state == IB_WQS_RDY) && (cur_state == IB_WQS_ERR))
+		return -EINVAL;
+
+	if ((new_state == IB_WQS_ERR) && (cur_state == IB_WQS_RESET))
+		return -EINVAL;
+
+	/* Need to protect against the parent RSS which also may modify WQ
+	 * state.
+	 */
+	mutex_lock(&qp->mutex);
+
+	/* Can update HW state only if a RSS QP has already associated to this
+	 * WQ, so we can apply its port on the WQ.
+	 */
+	if (qp->rss_usecnt)
+		err = _mlx4_ib_modify_wq(ibwq, new_state, udata);
+
+	if (!err)
+		ibwq->state = new_state;
+
+	mutex_unlock(&qp->mutex);
+
+	return err;
+}
+
+int mlx4_ib_destroy_wq(struct ib_wq *ibwq, struct ib_udata *udata)
+{
+	struct mlx4_ib_dev *dev = to_mdev(ibwq->device);
+	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
+
+	if (qp->counter_index)
+		mlx4_ib_free_qp_counter(dev, qp);
+
+	destroy_qp_common(dev, qp, MLX4_IB_RWQ_SRC, udata);
+
+	kfree(qp);
+	return 0;
+}
+
+int mlx4_ib_create_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table,
+				 struct ib_rwq_ind_table_init_attr *init_attr,
+				 struct ib_udata *udata)
+{
+	struct mlx4_ib_create_rwq_ind_tbl_resp resp = {};
+	unsigned int ind_tbl_size = 1 << init_attr->log_ind_tbl_size;
+	struct ib_device *device = rwq_ind_table->device;
+	unsigned int base_wqn;
+	size_t min_resp_len;
+	int i, err = 0;
+
+	if (udata->inlen > 0 &&
+	    !ib_is_udata_cleared(udata, 0,
+				 udata->inlen))
+		return -EOPNOTSUPP;
+
+	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
+	if (udata->outlen && udata->outlen < min_resp_len)
+		return -EINVAL;
+
+	if (ind_tbl_size >
+	    device->attrs.rss_caps.max_rwq_indirection_table_size) {
+		pr_debug("log_ind_tbl_size = %d is bigger than supported = %d\n",
+			 ind_tbl_size,
+			 device->attrs.rss_caps.max_rwq_indirection_table_size);
+		return -EINVAL;
+	}
+
+	base_wqn = init_attr->ind_tbl[0]->wq_num;
+
+	if (base_wqn % ind_tbl_size) {
+		pr_debug("WQN=0x%x isn't aligned with indirection table size\n",
+			 base_wqn);
+		return -EINVAL;
+	}
+
+	for (i = 1; i < ind_tbl_size; i++) {
+		if (++base_wqn != init_attr->ind_tbl[i]->wq_num) {
+			pr_debug("indirection table's WQNs aren't consecutive\n");
+			return -EINVAL;
+		}
+	}
+
+	if (udata->outlen) {
+		resp.response_length = offsetof(typeof(resp), response_length) +
+					sizeof(resp.response_length);
+		err = ib_copy_to_udata(udata, &resp, resp.response_length);
+	}
+
+	return err;
+}
+
+struct mlx4_ib_drain_cqe {
+	struct ib_cqe cqe;
+	struct completion done;
+};
+
+static void mlx4_ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
+{
+	struct mlx4_ib_drain_cqe *cqe = container_of(wc->wr_cqe,
+						     struct mlx4_ib_drain_cqe,
+						     cqe);
+
+	complete(&cqe->done);
+}
+
+/* This function returns only once the drained WR was completed */
+static void handle_drain_completion(struct ib_cq *cq,
+				    struct mlx4_ib_drain_cqe *sdrain,
+				    struct mlx4_ib_dev *dev)
+{
+	struct mlx4_dev *mdev = dev->dev;
+
+	if (cq->poll_ctx == IB_POLL_DIRECT) {
+		while (wait_for_completion_timeout(&sdrain->done, HZ / 10) <= 0)
+			ib_process_cq_direct(cq, -1);
+		return;
+	}
+
+	if (mdev->persist->state == MLX4_DEVICE_STATE_INTERNAL_ERROR) {
+		struct mlx4_ib_cq *mcq = to_mcq(cq);
+		bool triggered = false;
+		unsigned long flags;
+
+		spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
+		/* Make sure that the CQ handler won't run if wasn't run yet */
+		if (!mcq->mcq.reset_notify_added)
+			mcq->mcq.reset_notify_added = 1;
+		else
+			triggered = true;
+		spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
+
+		if (triggered) {
+			/* Wait for any scheduled/running task to be ended */
+			switch (cq->poll_ctx) {
+			case IB_POLL_SOFTIRQ:
+				irq_poll_disable(&cq->iop);
+				irq_poll_enable(&cq->iop);
+				break;
+			case IB_POLL_WORKQUEUE:
+				cancel_work_sync(&cq->work);
+				break;
+			default:
+				WARN_ON_ONCE(1);
+			}
+		}
+
+		/* Run the CQ handler - this makes sure that the drain WR will
+		 * be processed if wasn't processed yet.
+		 */
+		mcq->mcq.comp(&mcq->mcq);
+	}
+
+	wait_for_completion(&sdrain->done);
+}
+
+void mlx4_ib_drain_sq(struct ib_qp *qp)
+{
+	struct ib_cq *cq = qp->send_cq;
+	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
+	struct mlx4_ib_drain_cqe sdrain;
+	const struct ib_send_wr *bad_swr;
+	struct ib_rdma_wr swr = {
+		.wr = {
+			.next = NULL,
+			{ .wr_cqe	= &sdrain.cqe, },
+			.opcode	= IB_WR_RDMA_WRITE,
+		},
+	};
+	int ret;
+	struct mlx4_ib_dev *dev = to_mdev(qp->device);
+	struct mlx4_dev *mdev = dev->dev;
+
+	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
+	if (ret && mdev->persist->state != MLX4_DEVICE_STATE_INTERNAL_ERROR) {
+		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
+		return;
+	}
+
+	sdrain.cqe.done = mlx4_ib_drain_qp_done;
+	init_completion(&sdrain.done);
+
+	ret = _mlx4_ib_post_send(qp, &swr.wr, &bad_swr, true);
+	if (ret) {
+		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
+		return;
+	}
+
+	handle_drain_completion(cq, &sdrain, dev);
+}
+
+void mlx4_ib_drain_rq(struct ib_qp *qp)
+{
+	struct ib_cq *cq = qp->recv_cq;
+	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
+	struct mlx4_ib_drain_cqe rdrain;
+	struct ib_recv_wr rwr = {};
+	const struct ib_recv_wr *bad_rwr;
+	int ret;
+	struct mlx4_ib_dev *dev = to_mdev(qp->device);
+	struct mlx4_dev *mdev = dev->dev;
+
+	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
+	if (ret && mdev->persist->state != MLX4_DEVICE_STATE_INTERNAL_ERROR) {
+		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
+		return;
+	}
+
+	rwr.wr_cqe = &rdrain.cqe;
+	rdrain.cqe.done = mlx4_ib_drain_qp_done;
+	init_completion(&rdrain.done);
+
+	ret = _mlx4_ib_post_recv(qp, &rwr, &bad_rwr, true);
+	if (ret) {
+		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
+		return;
+	}
+
+	handle_drain_completion(cq, &rdrain, dev);
+}
diff --git a/drivers/infiniband/hw/mlx4/srq.c b/drivers/infiniband/hw/mlx4/srq.c
new file mode 100644
index 000000000..bf618529e
--- /dev/null
+++ b/drivers/infiniband/hw/mlx4/srq.c
@@ -0,0 +1,371 @@
+/*
+ * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
+ * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/mlx4/qp.h>
+#include <linux/mlx4/srq.h>
+#include <linux/slab.h>
+
+#include "mlx4_ib.h"
+#include <rdma/mlx4-abi.h>
+#include <rdma/uverbs_ioctl.h>
+
+static void *get_wqe(struct mlx4_ib_srq *srq, int n)
+{
+	return mlx4_buf_offset(&srq->buf, n << srq->msrq.wqe_shift);
+}
+
+static void mlx4_ib_srq_event(struct mlx4_srq *srq, enum mlx4_event type)
+{
+	struct ib_event event;
+	struct ib_srq *ibsrq = &to_mibsrq(srq)->ibsrq;
+
+	if (ibsrq->event_handler) {
+		event.device      = ibsrq->device;
+		event.element.srq = ibsrq;
+		switch (type) {
+		case MLX4_EVENT_TYPE_SRQ_LIMIT:
+			event.event = IB_EVENT_SRQ_LIMIT_REACHED;
+			break;
+		case MLX4_EVENT_TYPE_SRQ_CATAS_ERROR:
+			event.event = IB_EVENT_SRQ_ERR;
+			break;
+		default:
+			pr_warn("Unexpected event type %d "
+			       "on SRQ %06x\n", type, srq->srqn);
+			return;
+		}
+
+		ibsrq->event_handler(&event, ibsrq->srq_context);
+	}
+}
+
+int mlx4_ib_create_srq(struct ib_srq *ib_srq,
+		       struct ib_srq_init_attr *init_attr,
+		       struct ib_udata *udata)
+{
+	struct mlx4_ib_dev *dev = to_mdev(ib_srq->device);
+	struct mlx4_ib_ucontext *ucontext = rdma_udata_to_drv_context(
+		udata, struct mlx4_ib_ucontext, ibucontext);
+	struct mlx4_ib_srq *srq = to_msrq(ib_srq);
+	struct mlx4_wqe_srq_next_seg *next;
+	struct mlx4_wqe_data_seg *scatter;
+	u32 cqn;
+	u16 xrcdn;
+	int desc_size;
+	int buf_size;
+	int err;
+	int i;
+
+	/* Sanity check SRQ size before proceeding */
+	if (init_attr->attr.max_wr  >= dev->dev->caps.max_srq_wqes ||
+	    init_attr->attr.max_sge >  dev->dev->caps.max_srq_sge)
+		return -EINVAL;
+
+	mutex_init(&srq->mutex);
+	spin_lock_init(&srq->lock);
+	srq->msrq.max    = roundup_pow_of_two(init_attr->attr.max_wr + 1);
+	srq->msrq.max_gs = init_attr->attr.max_sge;
+
+	desc_size = max(32UL,
+			roundup_pow_of_two(sizeof (struct mlx4_wqe_srq_next_seg) +
+					   srq->msrq.max_gs *
+					   sizeof (struct mlx4_wqe_data_seg)));
+	srq->msrq.wqe_shift = ilog2(desc_size);
+
+	buf_size = srq->msrq.max * desc_size;
+
+	if (udata) {
+		struct mlx4_ib_create_srq ucmd;
+
+		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)))
+			return -EFAULT;
+
+		srq->umem =
+			ib_umem_get(ib_srq->device, ucmd.buf_addr, buf_size, 0);
+		if (IS_ERR(srq->umem))
+			return PTR_ERR(srq->umem);
+
+		err = mlx4_mtt_init(
+			dev->dev, ib_umem_num_dma_blocks(srq->umem, PAGE_SIZE),
+			PAGE_SHIFT, &srq->mtt);
+		if (err)
+			goto err_buf;
+
+		err = mlx4_ib_umem_write_mtt(dev, &srq->mtt, srq->umem);
+		if (err)
+			goto err_mtt;
+
+		err = mlx4_ib_db_map_user(udata, ucmd.db_addr, &srq->db);
+		if (err)
+			goto err_mtt;
+	} else {
+		err = mlx4_db_alloc(dev->dev, &srq->db, 0);
+		if (err)
+			return err;
+
+		*srq->db.db = 0;
+
+		if (mlx4_buf_alloc(dev->dev, buf_size, PAGE_SIZE * 2,
+				   &srq->buf)) {
+			err = -ENOMEM;
+			goto err_db;
+		}
+
+		srq->head    = 0;
+		srq->tail    = srq->msrq.max - 1;
+		srq->wqe_ctr = 0;
+
+		for (i = 0; i < srq->msrq.max; ++i) {
+			next = get_wqe(srq, i);
+			next->next_wqe_index =
+				cpu_to_be16((i + 1) & (srq->msrq.max - 1));
+
+			for (scatter = (void *) (next + 1);
+			     (void *) scatter < (void *) next + desc_size;
+			     ++scatter)
+				scatter->lkey = cpu_to_be32(MLX4_INVALID_LKEY);
+		}
+
+		err = mlx4_mtt_init(dev->dev, srq->buf.npages, srq->buf.page_shift,
+				    &srq->mtt);
+		if (err)
+			goto err_buf;
+
+		err = mlx4_buf_write_mtt(dev->dev, &srq->mtt, &srq->buf);
+		if (err)
+			goto err_mtt;
+
+		srq->wrid = kvmalloc_array(srq->msrq.max,
+					   sizeof(u64), GFP_KERNEL);
+		if (!srq->wrid) {
+			err = -ENOMEM;
+			goto err_mtt;
+		}
+	}
+
+	cqn = ib_srq_has_cq(init_attr->srq_type) ?
+		to_mcq(init_attr->ext.cq)->mcq.cqn : 0;
+	xrcdn = (init_attr->srq_type == IB_SRQT_XRC) ?
+		to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn :
+		(u16) dev->dev->caps.reserved_xrcds;
+	err = mlx4_srq_alloc(dev->dev, to_mpd(ib_srq->pd)->pdn, cqn, xrcdn,
+			     &srq->mtt, srq->db.dma, &srq->msrq);
+	if (err)
+		goto err_wrid;
+
+	srq->msrq.event = mlx4_ib_srq_event;
+	srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn;
+
+	if (udata)
+		if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof (__u32))) {
+			err = -EFAULT;
+			goto err_wrid;
+		}
+
+	init_attr->attr.max_wr = srq->msrq.max - 1;
+
+	return 0;
+
+err_wrid:
+	if (udata)
+		mlx4_ib_db_unmap_user(ucontext, &srq->db);
+	else
+		kvfree(srq->wrid);
+
+err_mtt:
+	mlx4_mtt_cleanup(dev->dev, &srq->mtt);
+
+err_buf:
+	if (!srq->umem)
+		mlx4_buf_free(dev->dev, buf_size, &srq->buf);
+	ib_umem_release(srq->umem);
+
+err_db:
+	if (!udata)
+		mlx4_db_free(dev->dev, &srq->db);
+
+	return err;
+}
+
+int mlx4_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
+		       enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
+{
+	struct mlx4_ib_dev *dev = to_mdev(ibsrq->device);
+	struct mlx4_ib_srq *srq = to_msrq(ibsrq);
+	int ret;
+
+	/* We don't support resizing SRQs (yet?) */
+	if (attr_mask & IB_SRQ_MAX_WR)
+		return -EINVAL;
+
+	if (attr_mask & IB_SRQ_LIMIT) {
+		if (attr->srq_limit >= srq->msrq.max)
+			return -EINVAL;
+
+		mutex_lock(&srq->mutex);
+		ret = mlx4_srq_arm(dev->dev, &srq->msrq, attr->srq_limit);
+		mutex_unlock(&srq->mutex);
+
+		if (ret)
+			return ret;
+	}
+
+	return 0;
+}
+
+int mlx4_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
+{
+	struct mlx4_ib_dev *dev = to_mdev(ibsrq->device);
+	struct mlx4_ib_srq *srq = to_msrq(ibsrq);
+	int ret;
+	int limit_watermark;
+
+	ret = mlx4_srq_query(dev->dev, &srq->msrq, &limit_watermark);
+	if (ret)
+		return ret;
+
+	srq_attr->srq_limit = limit_watermark;
+	srq_attr->max_wr    = srq->msrq.max - 1;
+	srq_attr->max_sge   = srq->msrq.max_gs;
+
+	return 0;
+}
+
+int mlx4_ib_destroy_srq(struct ib_srq *srq, struct ib_udata *udata)
+{
+	struct mlx4_ib_dev *dev = to_mdev(srq->device);
+	struct mlx4_ib_srq *msrq = to_msrq(srq);
+
+	mlx4_srq_free(dev->dev, &msrq->msrq);
+	mlx4_mtt_cleanup(dev->dev, &msrq->mtt);
+
+	if (udata) {
+		mlx4_ib_db_unmap_user(
+			rdma_udata_to_drv_context(
+				udata,
+				struct mlx4_ib_ucontext,
+				ibucontext),
+			&msrq->db);
+	} else {
+		kvfree(msrq->wrid);
+		mlx4_buf_free(dev->dev, msrq->msrq.max << msrq->msrq.wqe_shift,
+			      &msrq->buf);
+		mlx4_db_free(dev->dev, &msrq->db);
+	}
+	ib_umem_release(msrq->umem);
+	return 0;
+}
+
+void mlx4_ib_free_srq_wqe(struct mlx4_ib_srq *srq, int wqe_index)
+{
+	struct mlx4_wqe_srq_next_seg *next;
+
+	/* always called with interrupts disabled. */
+	spin_lock(&srq->lock);
+
+	next = get_wqe(srq, srq->tail);
+	next->next_wqe_index = cpu_to_be16(wqe_index);
+	srq->tail = wqe_index;
+
+	spin_unlock(&srq->lock);
+}
+
+int mlx4_ib_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
+			  const struct ib_recv_wr **bad_wr)
+{
+	struct mlx4_ib_srq *srq = to_msrq(ibsrq);
+	struct mlx4_wqe_srq_next_seg *next;
+	struct mlx4_wqe_data_seg *scat;
+	unsigned long flags;
+	int err = 0;
+	int nreq;
+	int i;
+	struct mlx4_ib_dev *mdev = to_mdev(ibsrq->device);
+
+	spin_lock_irqsave(&srq->lock, flags);
+	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
+		err = -EIO;
+		*bad_wr = wr;
+		nreq = 0;
+		goto out;
+	}
+
+	for (nreq = 0; wr; ++nreq, wr = wr->next) {
+		if (unlikely(wr->num_sge > srq->msrq.max_gs)) {
+			err = -EINVAL;
+			*bad_wr = wr;
+			break;
+		}
+
+		if (unlikely(srq->head == srq->tail)) {
+			err = -ENOMEM;
+			*bad_wr = wr;
+			break;
+		}
+
+		srq->wrid[srq->head] = wr->wr_id;
+
+		next      = get_wqe(srq, srq->head);
+		srq->head = be16_to_cpu(next->next_wqe_index);
+		scat      = (struct mlx4_wqe_data_seg *) (next + 1);
+
+		for (i = 0; i < wr->num_sge; ++i) {
+			scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length);
+			scat[i].lkey       = cpu_to_be32(wr->sg_list[i].lkey);
+			scat[i].addr       = cpu_to_be64(wr->sg_list[i].addr);
+		}
+
+		if (i < srq->msrq.max_gs) {
+			scat[i].byte_count = 0;
+			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
+			scat[i].addr       = 0;
+		}
+	}
+
+	if (likely(nreq)) {
+		srq->wqe_ctr += nreq;
+
+		/*
+		 * Make sure that descriptors are written before
+		 * doorbell record.
+		 */
+		wmb();
+
+		*srq->db.db = cpu_to_be32(srq->wqe_ctr);
+	}
+out:
+
+	spin_unlock_irqrestore(&srq->lock, flags);
+
+	return err;
+}
diff --git a/drivers/infiniband/hw/mlx4/sysfs.c b/drivers/infiniband/hw/mlx4/sysfs.c
new file mode 100644
index 000000000..6c3a23ee3
--- /dev/null
+++ b/drivers/infiniband/hw/mlx4/sysfs.c
@@ -0,0 +1,882 @@
+/*
+ * Copyright (c) 2012 Mellanox Technologies.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+/*#include "core_priv.h"*/
+#include "mlx4_ib.h"
+#include <linux/slab.h>
+#include <linux/string.h>
+#include <linux/stat.h>
+
+#include <rdma/ib_mad.h>
+/*show_admin_alias_guid returns the administratively assigned value of that GUID.
+ * Values returned in buf parameter string:
+ *	0			- requests opensm to assign a value.
+ *	ffffffffffffffff	- delete this entry.
+ *	other			- value assigned by administrator.
+ */
+static ssize_t show_admin_alias_guid(struct device *dev,
+			      struct device_attribute *attr, char *buf)
+{
+	struct mlx4_ib_iov_sysfs_attr *mlx4_ib_iov_dentry =
+		container_of(attr, struct mlx4_ib_iov_sysfs_attr, dentry);
+	struct mlx4_ib_iov_port *port = mlx4_ib_iov_dentry->ctx;
+	struct mlx4_ib_dev *mdev = port->dev;
+	__be64 sysadmin_ag_val;
+
+	sysadmin_ag_val = mlx4_get_admin_guid(mdev->dev,
+					      mlx4_ib_iov_dentry->entry_num,
+					      port->num);
+
+	return sprintf(buf, "%llx\n", be64_to_cpu(sysadmin_ag_val));
+}
+
+/* store_admin_alias_guid stores the (new) administratively assigned value of that GUID.
+ * Values in buf parameter string:
+ *	0			- requests opensm to assign a value.
+ *	0xffffffffffffffff	- delete this entry.
+ *	other			- guid value assigned by the administrator.
+ */
+static ssize_t store_admin_alias_guid(struct device *dev,
+				      struct device_attribute *attr,
+				      const char *buf, size_t count)
+{
+	int record_num;/*0-15*/
+	int guid_index_in_rec; /*0 - 7*/
+	struct mlx4_ib_iov_sysfs_attr *mlx4_ib_iov_dentry =
+		container_of(attr, struct mlx4_ib_iov_sysfs_attr, dentry);
+	struct mlx4_ib_iov_port *port = mlx4_ib_iov_dentry->ctx;
+	struct mlx4_ib_dev *mdev = port->dev;
+	u64 sysadmin_ag_val;
+	unsigned long flags;
+
+	record_num = mlx4_ib_iov_dentry->entry_num / 8;
+	guid_index_in_rec = mlx4_ib_iov_dentry->entry_num % 8;
+	if (0 == record_num && 0 == guid_index_in_rec) {
+		pr_err("GUID 0 block 0 is RO\n");
+		return count;
+	}
+	spin_lock_irqsave(&mdev->sriov.alias_guid.ag_work_lock, flags);
+	sscanf(buf, "%llx", &sysadmin_ag_val);
+	*(__be64 *)&mdev->sriov.alias_guid.ports_guid[port->num - 1].
+		all_rec_per_port[record_num].
+		all_recs[GUID_REC_SIZE * guid_index_in_rec] =
+			cpu_to_be64(sysadmin_ag_val);
+
+	/* Change the state to be pending for update */
+	mdev->sriov.alias_guid.ports_guid[port->num - 1].all_rec_per_port[record_num].status
+		= MLX4_GUID_INFO_STATUS_IDLE ;
+	mlx4_set_admin_guid(mdev->dev, cpu_to_be64(sysadmin_ag_val),
+			    mlx4_ib_iov_dentry->entry_num,
+			    port->num);
+
+	/* set the record index */
+	mdev->sriov.alias_guid.ports_guid[port->num - 1].all_rec_per_port[record_num].guid_indexes
+		|= mlx4_ib_get_aguid_comp_mask_from_ix(guid_index_in_rec);
+
+	spin_unlock_irqrestore(&mdev->sriov.alias_guid.ag_work_lock, flags);
+	mlx4_ib_init_alias_guid_work(mdev, port->num - 1);
+
+	return count;
+}
+
+static ssize_t show_port_gid(struct device *dev,
+			     struct device_attribute *attr,
+			     char *buf)
+{
+	struct mlx4_ib_iov_sysfs_attr *mlx4_ib_iov_dentry =
+		container_of(attr, struct mlx4_ib_iov_sysfs_attr, dentry);
+	struct mlx4_ib_iov_port *port = mlx4_ib_iov_dentry->ctx;
+	struct mlx4_ib_dev *mdev = port->dev;
+	union ib_gid gid;
+	ssize_t ret;
+
+	ret = __mlx4_ib_query_gid(&mdev->ib_dev, port->num,
+				  mlx4_ib_iov_dentry->entry_num, &gid, 1);
+	if (ret)
+		return ret;
+	ret = sprintf(buf, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
+		      be16_to_cpu(((__be16 *) gid.raw)[0]),
+		      be16_to_cpu(((__be16 *) gid.raw)[1]),
+		      be16_to_cpu(((__be16 *) gid.raw)[2]),
+		      be16_to_cpu(((__be16 *) gid.raw)[3]),
+		      be16_to_cpu(((__be16 *) gid.raw)[4]),
+		      be16_to_cpu(((__be16 *) gid.raw)[5]),
+		      be16_to_cpu(((__be16 *) gid.raw)[6]),
+		      be16_to_cpu(((__be16 *) gid.raw)[7]));
+	return ret;
+}
+
+static ssize_t show_phys_port_pkey(struct device *dev,
+				   struct device_attribute *attr,
+				   char *buf)
+{
+	struct mlx4_ib_iov_sysfs_attr *mlx4_ib_iov_dentry =
+		container_of(attr, struct mlx4_ib_iov_sysfs_attr, dentry);
+	struct mlx4_ib_iov_port *port = mlx4_ib_iov_dentry->ctx;
+	struct mlx4_ib_dev *mdev = port->dev;
+	u16 pkey;
+	ssize_t ret;
+
+	ret = __mlx4_ib_query_pkey(&mdev->ib_dev, port->num,
+				   mlx4_ib_iov_dentry->entry_num, &pkey, 1);
+	if (ret)
+		return ret;
+
+	return sprintf(buf, "0x%04x\n", pkey);
+}
+
+#define DENTRY_REMOVE(_dentry)						\
+do {									\
+	sysfs_remove_file((_dentry)->kobj, &(_dentry)->dentry.attr);	\
+} while (0);
+
+static int create_sysfs_entry(void *_ctx, struct mlx4_ib_iov_sysfs_attr *_dentry,
+			      char *_name, struct kobject *_kobj,
+			      ssize_t (*show)(struct device *dev,
+					      struct device_attribute *attr,
+					      char *buf),
+			      ssize_t (*store)(struct device *dev,
+					       struct device_attribute *attr,
+					       const char *buf, size_t count)
+			      )
+{
+	int ret = 0;
+	struct mlx4_ib_iov_sysfs_attr *vdentry = _dentry;
+
+	vdentry->ctx = _ctx;
+	vdentry->dentry.show = show;
+	vdentry->dentry.store = store;
+	sysfs_attr_init(&vdentry->dentry.attr);
+	vdentry->dentry.attr.name = vdentry->name;
+	vdentry->dentry.attr.mode = 0;
+	vdentry->kobj = _kobj;
+	snprintf(vdentry->name, 15, "%s", _name);
+
+	if (vdentry->dentry.store)
+		vdentry->dentry.attr.mode |= S_IWUSR;
+
+	if (vdentry->dentry.show)
+		vdentry->dentry.attr.mode |= S_IRUGO;
+
+	ret = sysfs_create_file(vdentry->kobj, &vdentry->dentry.attr);
+	if (ret) {
+		pr_err("failed to create %s\n", vdentry->dentry.attr.name);
+		vdentry->ctx = NULL;
+		return ret;
+	}
+
+	return ret;
+}
+
+int add_sysfs_port_mcg_attr(struct mlx4_ib_dev *device, int port_num,
+		struct attribute *attr)
+{
+	struct mlx4_ib_iov_port *port = &device->iov_ports[port_num - 1];
+	int ret;
+
+	ret = sysfs_create_file(port->mcgs_parent, attr);
+	if (ret)
+		pr_err("failed to create %s\n", attr->name);
+
+	return ret;
+}
+
+void del_sysfs_port_mcg_attr(struct mlx4_ib_dev *device, int port_num,
+		struct attribute *attr)
+{
+	struct mlx4_ib_iov_port *port = &device->iov_ports[port_num - 1];
+
+	sysfs_remove_file(port->mcgs_parent, attr);
+}
+
+static int add_port_entries(struct mlx4_ib_dev *device, int port_num)
+{
+	int i;
+	char buff[12];
+	struct mlx4_ib_iov_port *port = NULL;
+	int ret = 0 ;
+	struct ib_port_attr attr;
+
+	memset(&attr, 0, sizeof(attr));
+	/* get the physical gid and pkey table sizes.*/
+	ret = __mlx4_ib_query_port(&device->ib_dev, port_num, &attr, 1);
+	if (ret)
+		goto err;
+
+	port = &device->iov_ports[port_num - 1];
+	port->dev = device;
+	port->num = port_num;
+	/* Directory structure:
+	 * iov -
+	 *   port num -
+	 *	admin_guids
+	 *	gids (operational)
+	 *	mcg_table
+	 */
+	port->dentr_ar = kzalloc(sizeof (struct mlx4_ib_iov_sysfs_attr_ar),
+				 GFP_KERNEL);
+	if (!port->dentr_ar) {
+		ret = -ENOMEM;
+		goto err;
+	}
+	sprintf(buff, "%d", port_num);
+	port->cur_port = kobject_create_and_add(buff,
+				 kobject_get(device->ports_parent));
+	if (!port->cur_port) {
+		ret = -ENOMEM;
+		goto kobj_create_err;
+	}
+	/* admin GUIDs */
+	port->admin_alias_parent = kobject_create_and_add("admin_guids",
+						  kobject_get(port->cur_port));
+	if (!port->admin_alias_parent) {
+		ret = -ENOMEM;
+		goto err_admin_guids;
+	}
+	for (i = 0 ; i < attr.gid_tbl_len; i++) {
+		sprintf(buff, "%d", i);
+		port->dentr_ar->dentries[i].entry_num = i;
+		ret = create_sysfs_entry(port, &port->dentr_ar->dentries[i],
+					  buff, port->admin_alias_parent,
+					  show_admin_alias_guid, store_admin_alias_guid);
+		if (ret)
+			goto err_admin_alias_parent;
+	}
+
+	/* gids subdirectory (operational gids) */
+	port->gids_parent = kobject_create_and_add("gids",
+						  kobject_get(port->cur_port));
+	if (!port->gids_parent) {
+		ret = -ENOMEM;
+		goto err_gids;
+	}
+
+	for (i = 0 ; i < attr.gid_tbl_len; i++) {
+		sprintf(buff, "%d", i);
+		port->dentr_ar->dentries[attr.gid_tbl_len + i].entry_num = i;
+		ret = create_sysfs_entry(port,
+					 &port->dentr_ar->dentries[attr.gid_tbl_len + i],
+					 buff,
+					 port->gids_parent, show_port_gid, NULL);
+		if (ret)
+			goto err_gids_parent;
+	}
+
+	/* physical port pkey table */
+	port->pkeys_parent =
+		kobject_create_and_add("pkeys", kobject_get(port->cur_port));
+	if (!port->pkeys_parent) {
+		ret = -ENOMEM;
+		goto err_pkeys;
+	}
+
+	for (i = 0 ; i < attr.pkey_tbl_len; i++) {
+		sprintf(buff, "%d", i);
+		port->dentr_ar->dentries[2 * attr.gid_tbl_len + i].entry_num = i;
+		ret = create_sysfs_entry(port,
+					 &port->dentr_ar->dentries[2 * attr.gid_tbl_len + i],
+					 buff, port->pkeys_parent,
+					 show_phys_port_pkey, NULL);
+		if (ret)
+			goto err_pkeys_parent;
+	}
+
+	/* MCGs table */
+	port->mcgs_parent =
+		kobject_create_and_add("mcgs", kobject_get(port->cur_port));
+	if (!port->mcgs_parent) {
+		ret = -ENOMEM;
+		goto err_mcgs;
+	}
+	return 0;
+
+err_mcgs:
+	kobject_put(port->cur_port);
+
+err_pkeys_parent:
+	kobject_put(port->pkeys_parent);
+
+err_pkeys:
+	kobject_put(port->cur_port);
+
+err_gids_parent:
+	kobject_put(port->gids_parent);
+
+err_gids:
+	kobject_put(port->cur_port);
+
+err_admin_alias_parent:
+	kobject_put(port->admin_alias_parent);
+
+err_admin_guids:
+	kobject_put(port->cur_port);
+	kobject_put(port->cur_port); /* once more for create_and_add buff */
+
+kobj_create_err:
+	kobject_put(device->ports_parent);
+	kfree(port->dentr_ar);
+
+err:
+	pr_err("add_port_entries FAILED: for port:%d, error: %d\n",
+	       port_num, ret);
+	return ret;
+}
+
+static void get_name(struct mlx4_ib_dev *dev, char *name, int i, int max)
+{
+	/* pci_name format is: bus:dev:func -> xxxx:yy:zz.n
+	 * with no ARI only 3 last bits are used so when the fn is higher than 8
+	 * need to add it to the dev num, so count in the last number will be
+	 * modulo 8 */
+	snprintf(name, max, "%.8s%.2d.%d", pci_name(dev->dev->persist->pdev),
+		 i / 8, i % 8);
+}
+
+struct mlx4_port {
+	struct kobject         kobj;
+	struct mlx4_ib_dev    *dev;
+	struct attribute_group pkey_group;
+	struct attribute_group gid_group;
+	struct device_attribute	enable_smi_admin;
+	struct device_attribute	smi_enabled;
+	int		       slave;
+	u8                     port_num;
+};
+
+
+static void mlx4_port_release(struct kobject *kobj)
+{
+	struct mlx4_port *p = container_of(kobj, struct mlx4_port, kobj);
+	struct attribute *a;
+	int i;
+
+	for (i = 0; (a = p->pkey_group.attrs[i]); ++i)
+		kfree(a);
+	kfree(p->pkey_group.attrs);
+	for (i = 0; (a = p->gid_group.attrs[i]); ++i)
+		kfree(a);
+	kfree(p->gid_group.attrs);
+	kfree(p);
+}
+
+struct port_attribute {
+	struct attribute attr;
+	ssize_t (*show)(struct mlx4_port *, struct port_attribute *, char *buf);
+	ssize_t (*store)(struct mlx4_port *, struct port_attribute *,
+			 const char *buf, size_t count);
+};
+
+static ssize_t port_attr_show(struct kobject *kobj,
+			      struct attribute *attr, char *buf)
+{
+	struct port_attribute *port_attr =
+		container_of(attr, struct port_attribute, attr);
+	struct mlx4_port *p = container_of(kobj, struct mlx4_port, kobj);
+
+	if (!port_attr->show)
+		return -EIO;
+	return port_attr->show(p, port_attr, buf);
+}
+
+static ssize_t port_attr_store(struct kobject *kobj,
+			       struct attribute *attr,
+			       const char *buf, size_t size)
+{
+	struct port_attribute *port_attr =
+		container_of(attr, struct port_attribute, attr);
+	struct mlx4_port *p = container_of(kobj, struct mlx4_port, kobj);
+
+	if (!port_attr->store)
+		return -EIO;
+	return port_attr->store(p, port_attr, buf, size);
+}
+
+static const struct sysfs_ops port_sysfs_ops = {
+	.show = port_attr_show,
+	.store = port_attr_store,
+};
+
+static struct kobj_type port_type = {
+	.release    = mlx4_port_release,
+	.sysfs_ops  = &port_sysfs_ops,
+};
+
+struct port_table_attribute {
+	struct port_attribute	attr;
+	char			name[8];
+	int			index;
+};
+
+static ssize_t show_port_pkey(struct mlx4_port *p, struct port_attribute *attr,
+			      char *buf)
+{
+	struct port_table_attribute *tab_attr =
+		container_of(attr, struct port_table_attribute, attr);
+	ssize_t ret = -ENODEV;
+
+	if (p->dev->pkeys.virt2phys_pkey[p->slave][p->port_num - 1][tab_attr->index] >=
+	    (p->dev->dev->caps.pkey_table_len[p->port_num]))
+		ret = sprintf(buf, "none\n");
+	else
+		ret = sprintf(buf, "%d\n",
+			      p->dev->pkeys.virt2phys_pkey[p->slave]
+			      [p->port_num - 1][tab_attr->index]);
+	return ret;
+}
+
+static ssize_t store_port_pkey(struct mlx4_port *p, struct port_attribute *attr,
+			       const char *buf, size_t count)
+{
+	struct port_table_attribute *tab_attr =
+		container_of(attr, struct port_table_attribute, attr);
+	int idx;
+	int err;
+
+	/* do not allow remapping Dom0 virtual pkey table */
+	if (p->slave == mlx4_master_func_num(p->dev->dev))
+		return -EINVAL;
+
+	if (!strncasecmp(buf, "no", 2))
+		idx = p->dev->dev->phys_caps.pkey_phys_table_len[p->port_num] - 1;
+	else if (sscanf(buf, "%i", &idx) != 1 ||
+		 idx >= p->dev->dev->caps.pkey_table_len[p->port_num] ||
+		 idx < 0)
+		return -EINVAL;
+
+	p->dev->pkeys.virt2phys_pkey[p->slave][p->port_num - 1]
+				    [tab_attr->index] = idx;
+	mlx4_sync_pkey_table(p->dev->dev, p->slave, p->port_num,
+			     tab_attr->index, idx);
+	err = mlx4_gen_pkey_eqe(p->dev->dev, p->slave, p->port_num);
+	if (err) {
+		pr_err("mlx4_gen_pkey_eqe failed for slave %d,"
+		       " port %d, index %d\n", p->slave, p->port_num, idx);
+		return err;
+	}
+	return count;
+}
+
+static ssize_t show_port_gid_idx(struct mlx4_port *p,
+				 struct port_attribute *attr, char *buf)
+{
+	return sprintf(buf, "%d\n", p->slave);
+}
+
+static struct attribute **
+alloc_group_attrs(ssize_t (*show)(struct mlx4_port *,
+				  struct port_attribute *, char *buf),
+		  ssize_t (*store)(struct mlx4_port *, struct port_attribute *,
+				   const char *buf, size_t count),
+		  int len)
+{
+	struct attribute **tab_attr;
+	struct port_table_attribute *element;
+	int i;
+
+	tab_attr = kcalloc(1 + len, sizeof (struct attribute *), GFP_KERNEL);
+	if (!tab_attr)
+		return NULL;
+
+	for (i = 0; i < len; i++) {
+		element = kzalloc(sizeof (struct port_table_attribute),
+				  GFP_KERNEL);
+		if (!element)
+			goto err;
+		if (snprintf(element->name, sizeof (element->name),
+			     "%d", i) >= sizeof (element->name)) {
+			kfree(element);
+			goto err;
+		}
+		sysfs_attr_init(&element->attr.attr);
+		element->attr.attr.name  = element->name;
+		if (store) {
+			element->attr.attr.mode  = S_IWUSR | S_IRUGO;
+			element->attr.store	 = store;
+		} else
+			element->attr.attr.mode  = S_IRUGO;
+
+		element->attr.show       = show;
+		element->index		 = i;
+		tab_attr[i] = &element->attr.attr;
+	}
+	return tab_attr;
+
+err:
+	while (--i >= 0)
+		kfree(tab_attr[i]);
+	kfree(tab_attr);
+	return NULL;
+}
+
+static ssize_t sysfs_show_smi_enabled(struct device *dev,
+				      struct device_attribute *attr, char *buf)
+{
+	struct mlx4_port *p =
+		container_of(attr, struct mlx4_port, smi_enabled);
+	ssize_t len = 0;
+
+	if (mlx4_vf_smi_enabled(p->dev->dev, p->slave, p->port_num))
+		len = sprintf(buf, "%d\n", 1);
+	else
+		len = sprintf(buf, "%d\n", 0);
+
+	return len;
+}
+
+static ssize_t sysfs_show_enable_smi_admin(struct device *dev,
+					   struct device_attribute *attr,
+					   char *buf)
+{
+	struct mlx4_port *p =
+		container_of(attr, struct mlx4_port, enable_smi_admin);
+	ssize_t len = 0;
+
+	if (mlx4_vf_get_enable_smi_admin(p->dev->dev, p->slave, p->port_num))
+		len = sprintf(buf, "%d\n", 1);
+	else
+		len = sprintf(buf, "%d\n", 0);
+
+	return len;
+}
+
+static ssize_t sysfs_store_enable_smi_admin(struct device *dev,
+					    struct device_attribute *attr,
+					    const char *buf, size_t count)
+{
+	struct mlx4_port *p =
+		container_of(attr, struct mlx4_port, enable_smi_admin);
+	int enable;
+
+	if (sscanf(buf, "%i", &enable) != 1 ||
+	    enable < 0 || enable > 1)
+		return -EINVAL;
+
+	if (mlx4_vf_set_enable_smi_admin(p->dev->dev, p->slave, p->port_num, enable))
+		return -EINVAL;
+	return count;
+}
+
+static int add_vf_smi_entries(struct mlx4_port *p)
+{
+	int is_eth = rdma_port_get_link_layer(&p->dev->ib_dev, p->port_num) ==
+			IB_LINK_LAYER_ETHERNET;
+	int ret;
+
+	/* do not display entries if eth transport, or if master */
+	if (is_eth || p->slave == mlx4_master_func_num(p->dev->dev))
+		return 0;
+
+	sysfs_attr_init(&p->smi_enabled.attr);
+	p->smi_enabled.show = sysfs_show_smi_enabled;
+	p->smi_enabled.store = NULL;
+	p->smi_enabled.attr.name = "smi_enabled";
+	p->smi_enabled.attr.mode = 0444;
+	ret = sysfs_create_file(&p->kobj, &p->smi_enabled.attr);
+	if (ret) {
+		pr_err("failed to create smi_enabled\n");
+		return ret;
+	}
+
+	sysfs_attr_init(&p->enable_smi_admin.attr);
+	p->enable_smi_admin.show = sysfs_show_enable_smi_admin;
+	p->enable_smi_admin.store = sysfs_store_enable_smi_admin;
+	p->enable_smi_admin.attr.name = "enable_smi_admin";
+	p->enable_smi_admin.attr.mode = 0644;
+	ret = sysfs_create_file(&p->kobj, &p->enable_smi_admin.attr);
+	if (ret) {
+		pr_err("failed to create enable_smi_admin\n");
+		sysfs_remove_file(&p->kobj, &p->smi_enabled.attr);
+		return ret;
+	}
+	return 0;
+}
+
+static void remove_vf_smi_entries(struct mlx4_port *p)
+{
+	int is_eth = rdma_port_get_link_layer(&p->dev->ib_dev, p->port_num) ==
+			IB_LINK_LAYER_ETHERNET;
+
+	if (is_eth || p->slave == mlx4_master_func_num(p->dev->dev))
+		return;
+
+	sysfs_remove_file(&p->kobj, &p->smi_enabled.attr);
+	sysfs_remove_file(&p->kobj, &p->enable_smi_admin.attr);
+}
+
+static int add_port(struct mlx4_ib_dev *dev, int port_num, int slave)
+{
+	struct mlx4_port *p;
+	int i;
+	int ret;
+	int is_eth = rdma_port_get_link_layer(&dev->ib_dev, port_num) ==
+			IB_LINK_LAYER_ETHERNET;
+
+	p = kzalloc(sizeof *p, GFP_KERNEL);
+	if (!p)
+		return -ENOMEM;
+
+	p->dev = dev;
+	p->port_num = port_num;
+	p->slave = slave;
+
+	ret = kobject_init_and_add(&p->kobj, &port_type,
+				   kobject_get(dev->dev_ports_parent[slave]),
+				   "%d", port_num);
+	if (ret)
+		goto err_alloc;
+
+	p->pkey_group.name  = "pkey_idx";
+	p->pkey_group.attrs =
+		alloc_group_attrs(show_port_pkey,
+				  is_eth ? NULL : store_port_pkey,
+				  dev->dev->caps.pkey_table_len[port_num]);
+	if (!p->pkey_group.attrs) {
+		ret = -ENOMEM;
+		goto err_alloc;
+	}
+
+	ret = sysfs_create_group(&p->kobj, &p->pkey_group);
+	if (ret)
+		goto err_free_pkey;
+
+	p->gid_group.name  = "gid_idx";
+	p->gid_group.attrs = alloc_group_attrs(show_port_gid_idx, NULL, 1);
+	if (!p->gid_group.attrs) {
+		ret = -ENOMEM;
+		goto err_free_pkey;
+	}
+
+	ret = sysfs_create_group(&p->kobj, &p->gid_group);
+	if (ret)
+		goto err_free_gid;
+
+	ret = add_vf_smi_entries(p);
+	if (ret)
+		goto err_free_gid;
+
+	list_add_tail(&p->kobj.entry, &dev->pkeys.pkey_port_list[slave]);
+	return 0;
+
+err_free_gid:
+	kfree(p->gid_group.attrs[0]);
+	kfree(p->gid_group.attrs);
+
+err_free_pkey:
+	for (i = 0; i < dev->dev->caps.pkey_table_len[port_num]; ++i)
+		kfree(p->pkey_group.attrs[i]);
+	kfree(p->pkey_group.attrs);
+
+err_alloc:
+	kobject_put(dev->dev_ports_parent[slave]);
+	kfree(p);
+	return ret;
+}
+
+static int register_one_pkey_tree(struct mlx4_ib_dev *dev, int slave)
+{
+	char name[32];
+	int err;
+	int port;
+	struct kobject *p, *t;
+	struct mlx4_port *mport;
+	struct mlx4_active_ports actv_ports;
+
+	get_name(dev, name, slave, sizeof name);
+
+	dev->pkeys.device_parent[slave] =
+		kobject_create_and_add(name, kobject_get(dev->iov_parent));
+
+	if (!dev->pkeys.device_parent[slave]) {
+		err = -ENOMEM;
+		goto fail_dev;
+	}
+
+	INIT_LIST_HEAD(&dev->pkeys.pkey_port_list[slave]);
+
+	dev->dev_ports_parent[slave] =
+		kobject_create_and_add("ports",
+				       kobject_get(dev->pkeys.device_parent[slave]));
+
+	if (!dev->dev_ports_parent[slave]) {
+		err = -ENOMEM;
+		goto err_ports;
+	}
+
+	actv_ports = mlx4_get_active_ports(dev->dev, slave);
+
+	for (port = 1; port <= dev->dev->caps.num_ports; ++port) {
+		if (!test_bit(port - 1, actv_ports.ports))
+			continue;
+		err = add_port(dev, port, slave);
+		if (err)
+			goto err_add;
+	}
+	return 0;
+
+err_add:
+	list_for_each_entry_safe(p, t,
+				 &dev->pkeys.pkey_port_list[slave],
+				 entry) {
+		list_del(&p->entry);
+		mport = container_of(p, struct mlx4_port, kobj);
+		sysfs_remove_group(p, &mport->pkey_group);
+		sysfs_remove_group(p, &mport->gid_group);
+		remove_vf_smi_entries(mport);
+		kobject_put(p);
+	}
+	kobject_put(dev->dev_ports_parent[slave]);
+
+err_ports:
+	kobject_put(dev->pkeys.device_parent[slave]);
+	/* extra put for the device_parent create_and_add */
+	kobject_put(dev->pkeys.device_parent[slave]);
+
+fail_dev:
+	kobject_put(dev->iov_parent);
+	return err;
+}
+
+static int register_pkey_tree(struct mlx4_ib_dev *device)
+{
+	int i;
+
+	if (!mlx4_is_master(device->dev))
+		return 0;
+
+	for (i = 0; i <= device->dev->persist->num_vfs; ++i)
+		register_one_pkey_tree(device, i);
+
+	return 0;
+}
+
+static void unregister_pkey_tree(struct mlx4_ib_dev *device)
+{
+	int slave;
+	struct kobject *p, *t;
+	struct mlx4_port *port;
+
+	if (!mlx4_is_master(device->dev))
+		return;
+
+	for (slave = device->dev->persist->num_vfs; slave >= 0; --slave) {
+		list_for_each_entry_safe(p, t,
+					 &device->pkeys.pkey_port_list[slave],
+					 entry) {
+			list_del(&p->entry);
+			port = container_of(p, struct mlx4_port, kobj);
+			sysfs_remove_group(p, &port->pkey_group);
+			sysfs_remove_group(p, &port->gid_group);
+			remove_vf_smi_entries(port);
+			kobject_put(p);
+			kobject_put(device->dev_ports_parent[slave]);
+		}
+		kobject_put(device->dev_ports_parent[slave]);
+		kobject_put(device->pkeys.device_parent[slave]);
+		kobject_put(device->pkeys.device_parent[slave]);
+		kobject_put(device->iov_parent);
+	}
+}
+
+int mlx4_ib_device_register_sysfs(struct mlx4_ib_dev *dev)
+{
+	int i;
+	int ret = 0;
+
+	if (!mlx4_is_master(dev->dev))
+		return 0;
+
+	dev->iov_parent = kobject_create_and_add("iov", &dev->ib_dev.dev.kobj);
+	if (!dev->iov_parent) {
+		ret = -ENOMEM;
+		goto err;
+	}
+	dev->ports_parent =
+		kobject_create_and_add("ports",
+				       kobject_get(dev->iov_parent));
+	if (!dev->ports_parent) {
+		ret = -ENOMEM;
+		goto err_ports;
+	}
+
+	for (i = 1; i <= dev->ib_dev.phys_port_cnt; ++i) {
+		ret = add_port_entries(dev, i);
+		if (ret)
+			goto err_add_entries;
+	}
+
+	ret = register_pkey_tree(dev);
+	if (ret)
+		goto err_add_entries;
+	return 0;
+
+err_add_entries:
+	kobject_put(dev->ports_parent);
+
+err_ports:
+	kobject_put(dev->iov_parent);
+err:
+	pr_err("mlx4_ib_device_register_sysfs error (%d)\n", ret);
+	return ret;
+}
+
+static void unregister_alias_guid_tree(struct mlx4_ib_dev *device)
+{
+	struct mlx4_ib_iov_port *p;
+	int i;
+
+	if (!mlx4_is_master(device->dev))
+		return;
+
+	for (i = 0; i < device->dev->caps.num_ports; i++) {
+		p = &device->iov_ports[i];
+		kobject_put(p->admin_alias_parent);
+		kobject_put(p->gids_parent);
+		kobject_put(p->pkeys_parent);
+		kobject_put(p->mcgs_parent);
+		kobject_put(p->cur_port);
+		kobject_put(p->cur_port);
+		kobject_put(p->cur_port);
+		kobject_put(p->cur_port);
+		kobject_put(p->cur_port);
+		kobject_put(p->dev->ports_parent);
+		kfree(p->dentr_ar);
+	}
+}
+
+void mlx4_ib_device_unregister_sysfs(struct mlx4_ib_dev *device)
+{
+	unregister_alias_guid_tree(device);
+	unregister_pkey_tree(device);
+	kobject_put(device->ports_parent);
+	kobject_put(device->iov_parent);
+	kobject_put(device->iov_parent);
+}