dpdk0716 rc3

18 files changed, 9980 insertions, 0 deletions

diff --git a/src/dpdk/drivers/net/mlx5/mlx5.c b/src/dpdk/drivers/net/mlx5/mlx5.c
new file mode 100644
index 00000000..d96a9aff
--- /dev/null
+++ b/src/dpdk/drivers/net/mlx5/mlx5.c
@@ -0,0 +1,765 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2015 6WIND S.A.
+ *   Copyright 2015 Mellanox.
+ *
+ *   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.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <stddef.h>
+#include <unistd.h>
+#include <string.h>
+#include <assert.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include <errno.h>
+#include <net/if.h>
+
+/* Verbs header. */
+/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <infiniband/verbs.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+/* DPDK headers don't like -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <rte_malloc.h>
+#include <rte_ethdev.h>
+#include <rte_pci.h>
+#include <rte_common.h>
+#include <rte_kvargs.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+#include "mlx5.h"
+#include "mlx5_utils.h"
+#include "mlx5_rxtx.h"
+#include "mlx5_autoconf.h"
+#include "mlx5_defs.h"
+
+/* Device parameter to enable RX completion queue compression. */
+#define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
+
+/* Device parameter to configure inline send. */
+#define MLX5_TXQ_INLINE "txq_inline"
+
+/*
+ * Device parameter to configure the number of TX queues threshold for
+ * enabling inline send.
+ */
+#define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
+
+/* Device parameter to enable multi-packet send WQEs. */
+#define MLX5_TXQ_MPW_EN "txq_mpw_en"
+
+/**
+ * Retrieve integer value from environment variable.
+ *
+ * @param[in] name
+ *   Environment variable name.
+ *
+ * @return
+ *   Integer value, 0 if the variable is not set.
+ */
+int
+mlx5_getenv_int(const char *name)
+{
+	const char *val = getenv(name);
+
+	if (val == NULL)
+		return 0;
+	return atoi(val);
+}
+
+/**
+ * DPDK callback to close the device.
+ *
+ * Destroy all queues and objects, free memory.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ */
+static void
+mlx5_dev_close(struct rte_eth_dev *dev)
+{
+	struct priv *priv = mlx5_get_priv(dev);
+	unsigned int i;
+
+	priv_lock(priv);
+	DEBUG("%p: closing device \"%s\"",
+	      (void *)dev,
+	      ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
+	/* In case mlx5_dev_stop() has not been called. */
+	priv_dev_interrupt_handler_uninstall(priv, dev);
+	priv_special_flow_disable_all(priv);
+	priv_mac_addrs_disable(priv);
+	priv_destroy_hash_rxqs(priv);
+
+	/* Remove flow director elements. */
+	priv_fdir_disable(priv);
+	priv_fdir_delete_filters_list(priv);
+
+	/* Prevent crashes when queues are still in use. */
+	dev->rx_pkt_burst = removed_rx_burst;
+	dev->tx_pkt_burst = removed_tx_burst;
+	if (priv->rxqs != NULL) {
+		/* XXX race condition if mlx5_rx_burst() is still running. */
+		usleep(1000);
+		for (i = 0; (i != priv->rxqs_n); ++i) {
+			struct rxq *rxq = (*priv->rxqs)[i];
+			struct rxq_ctrl *rxq_ctrl;
+
+			if (rxq == NULL)
+				continue;
+			rxq_ctrl = container_of(rxq, struct rxq_ctrl, rxq);
+			(*priv->rxqs)[i] = NULL;
+			rxq_cleanup(rxq_ctrl);
+			rte_free(rxq_ctrl);
+		}
+		priv->rxqs_n = 0;
+		priv->rxqs = NULL;
+	}
+	if (priv->txqs != NULL) {
+		/* XXX race condition if mlx5_tx_burst() is still running. */
+		usleep(1000);
+		for (i = 0; (i != priv->txqs_n); ++i) {
+			struct txq *txq = (*priv->txqs)[i];
+			struct txq_ctrl *txq_ctrl;
+
+			if (txq == NULL)
+				continue;
+			txq_ctrl = container_of(txq, struct txq_ctrl, txq);
+			(*priv->txqs)[i] = NULL;
+			txq_cleanup(txq_ctrl);
+			rte_free(txq_ctrl);
+		}
+		priv->txqs_n = 0;
+		priv->txqs = NULL;
+	}
+	if (priv->pd != NULL) {
+		assert(priv->ctx != NULL);
+		claim_zero(ibv_dealloc_pd(priv->pd));
+		claim_zero(ibv_close_device(priv->ctx));
+	} else
+		assert(priv->ctx == NULL);
+	if (priv->rss_conf != NULL) {
+		for (i = 0; (i != hash_rxq_init_n); ++i)
+			rte_free((*priv->rss_conf)[i]);
+		rte_free(priv->rss_conf);
+	}
+	if (priv->reta_idx != NULL)
+		rte_free(priv->reta_idx);
+	priv_unlock(priv);
+	memset(priv, 0, sizeof(*priv));
+}
+
+static const struct eth_dev_ops mlx5_dev_ops = {
+	.dev_configure = mlx5_dev_configure,
+	.dev_start = mlx5_dev_start,
+	.dev_stop = mlx5_dev_stop,
+	.dev_set_link_down = mlx5_set_link_down,
+	.dev_set_link_up = mlx5_set_link_up,
+	.dev_close = mlx5_dev_close,
+	.promiscuous_enable = mlx5_promiscuous_enable,
+	.promiscuous_disable = mlx5_promiscuous_disable,
+	.allmulticast_enable = mlx5_allmulticast_enable,
+	.allmulticast_disable = mlx5_allmulticast_disable,
+	.link_update = mlx5_link_update,
+	.stats_get = mlx5_stats_get,
+	.stats_reset = mlx5_stats_reset,
+	.dev_infos_get = mlx5_dev_infos_get,
+	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
+	.vlan_filter_set = mlx5_vlan_filter_set,
+	.rx_queue_setup = mlx5_rx_queue_setup,
+	.tx_queue_setup = mlx5_tx_queue_setup,
+	.rx_queue_release = mlx5_rx_queue_release,
+	.tx_queue_release = mlx5_tx_queue_release,
+	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
+	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
+	.mac_addr_remove = mlx5_mac_addr_remove,
+	.mac_addr_add = mlx5_mac_addr_add,
+	.mac_addr_set = mlx5_mac_addr_set,
+	.mtu_set = mlx5_dev_set_mtu,
+	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
+	.vlan_offload_set = mlx5_vlan_offload_set,
+	.reta_update = mlx5_dev_rss_reta_update,
+	.reta_query = mlx5_dev_rss_reta_query,
+	.rss_hash_update = mlx5_rss_hash_update,
+	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
+	.filter_ctrl = mlx5_dev_filter_ctrl,
+};
+
+static struct {
+	struct rte_pci_addr pci_addr; /* associated PCI address */
+	uint32_t ports; /* physical ports bitfield. */
+} mlx5_dev[32];
+
+/**
+ * Get device index in mlx5_dev[] from PCI bus address.
+ *
+ * @param[in] pci_addr
+ *   PCI bus address to look for.
+ *
+ * @return
+ *   mlx5_dev[] index on success, -1 on failure.
+ */
+static int
+mlx5_dev_idx(struct rte_pci_addr *pci_addr)
+{
+	unsigned int i;
+	int ret = -1;
+
+	assert(pci_addr != NULL);
+	for (i = 0; (i != RTE_DIM(mlx5_dev)); ++i) {
+		if ((mlx5_dev[i].pci_addr.domain == pci_addr->domain) &&
+		    (mlx5_dev[i].pci_addr.bus == pci_addr->bus) &&
+		    (mlx5_dev[i].pci_addr.devid == pci_addr->devid) &&
+		    (mlx5_dev[i].pci_addr.function == pci_addr->function))
+			return i;
+		if ((mlx5_dev[i].ports == 0) && (ret == -1))
+			ret = i;
+	}
+	return ret;
+}
+
+/**
+ * Verify and store value for device argument.
+ *
+ * @param[in] key
+ *   Key argument to verify.
+ * @param[in] val
+ *   Value associated with key.
+ * @param opaque
+ *   User data.
+ *
+ * @return
+ *   0 on success, negative errno value on failure.
+ */
+static int
+mlx5_args_check(const char *key, const char *val, void *opaque)
+{
+	struct priv *priv = opaque;
+	unsigned long tmp;
+
+	errno = 0;
+	tmp = strtoul(val, NULL, 0);
+	if (errno) {
+		WARN("%s: \"%s\" is not a valid integer", key, val);
+		return errno;
+	}
+	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
+		priv->cqe_comp = !!tmp;
+	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
+		priv->txq_inline = tmp;
+	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
+		priv->txqs_inline = tmp;
+	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
+		priv->mps = !!tmp;
+	} else {
+		WARN("%s: unknown parameter", key);
+		return -EINVAL;
+	}
+	return 0;
+}
+
+/**
+ * Parse device parameters.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param devargs
+ *   Device arguments structure.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+mlx5_args(struct priv *priv, struct rte_devargs *devargs)
+{
+	const char **params = (const char *[]){
+		MLX5_RXQ_CQE_COMP_EN,
+		MLX5_TXQ_INLINE,
+		MLX5_TXQS_MIN_INLINE,
+		MLX5_TXQ_MPW_EN,
+		NULL,
+	};
+	struct rte_kvargs *kvlist;
+	int ret = 0;
+	int i;
+
+	if (devargs == NULL)
+		return 0;
+	/* Following UGLY cast is done to pass checkpatch. */
+	kvlist = rte_kvargs_parse(devargs->args, params);
+	if (kvlist == NULL)
+		return 0;
+	/* Process parameters. */
+	for (i = 0; (params[i] != NULL); ++i) {
+		if (rte_kvargs_count(kvlist, params[i])) {
+			ret = rte_kvargs_process(kvlist, params[i],
+						 mlx5_args_check, priv);
+			if (ret != 0)
+				return ret;
+		}
+	}
+	rte_kvargs_free(kvlist);
+	return 0;
+}
+
+static struct eth_driver mlx5_driver;
+
+/**
+ * DPDK callback to register a PCI device.
+ *
+ * This function creates an Ethernet device for each port of a given
+ * PCI device.
+ *
+ * @param[in] pci_drv
+ *   PCI driver structure (mlx5_driver).
+ * @param[in] pci_dev
+ *   PCI device information.
+ *
+ * @return
+ *   0 on success, negative errno value on failure.
+ */
+static int
+mlx5_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
+{
+	struct ibv_device **list;
+	struct ibv_device *ibv_dev;
+	int err = 0;
+	struct ibv_context *attr_ctx = NULL;
+	struct ibv_device_attr device_attr;
+	unsigned int sriov;
+	unsigned int mps;
+	int idx;
+	int i;
+
+	(void)pci_drv;
+	assert(pci_drv == &mlx5_driver.pci_drv);
+	/* Get mlx5_dev[] index. */
+	idx = mlx5_dev_idx(&pci_dev->addr);
+	if (idx == -1) {
+		ERROR("this driver cannot support any more adapters");
+		return -ENOMEM;
+	}
+	DEBUG("using driver device index %d", idx);
+
+	/* Save PCI address. */
+	mlx5_dev[idx].pci_addr = pci_dev->addr;
+	list = ibv_get_device_list(&i);
+	if (list == NULL) {
+		assert(errno);
+		if (errno == ENOSYS) {
+			WARN("cannot list devices, is ib_uverbs loaded?");
+			return 0;
+		}
+		return -errno;
+	}
+	assert(i >= 0);
+	/*
+	 * For each listed device, check related sysfs entry against
+	 * the provided PCI ID.
+	 */
+	while (i != 0) {
+		struct rte_pci_addr pci_addr;
+
+		--i;
+		DEBUG("checking device \"%s\"", list[i]->name);
+		if (mlx5_ibv_device_to_pci_addr(list[i], &pci_addr))
+			continue;
+		if ((pci_dev->addr.domain != pci_addr.domain) ||
+		    (pci_dev->addr.bus != pci_addr.bus) ||
+		    (pci_dev->addr.devid != pci_addr.devid) ||
+		    (pci_dev->addr.function != pci_addr.function))
+			continue;
+		sriov = ((pci_dev->id.device_id ==
+		       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF) ||
+		      (pci_dev->id.device_id ==
+		       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF));
+		/* Multi-packet send is only supported by ConnectX-4 Lx PF. */
+		mps = (pci_dev->id.device_id ==
+		       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX);
+		INFO("PCI information matches, using device \"%s\""
+		     " (SR-IOV: %s, MPS: %s)",
+		     list[i]->name,
+		     sriov ? "true" : "false",
+		     mps ? "true" : "false");
+		attr_ctx = ibv_open_device(list[i]);
+		err = errno;
+		break;
+	}
+	if (attr_ctx == NULL) {
+		ibv_free_device_list(list);
+		switch (err) {
+		case 0:
+			WARN("cannot access device, is mlx5_ib loaded?");
+			return 0;
+		case EINVAL:
+			WARN("cannot use device, are drivers up to date?");
+			return 0;
+		}
+		assert(err > 0);
+		return -err;
+	}
+	ibv_dev = list[i];
+
+	DEBUG("device opened");
+	if (ibv_query_device(attr_ctx, &device_attr))
+		goto error;
+	INFO("%u port(s) detected", device_attr.phys_port_cnt);
+
+	for (i = 0; i < device_attr.phys_port_cnt; i++) {
+		uint32_t port = i + 1; /* ports are indexed from one */
+		uint32_t test = (1 << i);
+		struct ibv_context *ctx = NULL;
+		struct ibv_port_attr port_attr;
+		struct ibv_pd *pd = NULL;
+		struct priv *priv = NULL;
+		struct rte_eth_dev *eth_dev;
+		struct ibv_exp_device_attr exp_device_attr;
+		struct ether_addr mac;
+		uint16_t num_vfs = 0;
+
+		exp_device_attr.comp_mask =
+			IBV_EXP_DEVICE_ATTR_EXP_CAP_FLAGS |
+			IBV_EXP_DEVICE_ATTR_RX_HASH |
+			IBV_EXP_DEVICE_ATTR_VLAN_OFFLOADS |
+			IBV_EXP_DEVICE_ATTR_RX_PAD_END_ALIGN |
+			0;
+
+		DEBUG("using port %u (%08" PRIx32 ")", port, test);
+
+		ctx = ibv_open_device(ibv_dev);
+		if (ctx == NULL)
+			goto port_error;
+
+		/* Check port status. */
+		err = ibv_query_port(ctx, port, &port_attr);
+		if (err) {
+			ERROR("port query failed: %s", strerror(err));
+			goto port_error;
+		}
+
+		if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
+			ERROR("port %d is not configured in Ethernet mode",
+			      port);
+			goto port_error;
+		}
+
+		if (port_attr.state != IBV_PORT_ACTIVE)
+			DEBUG("port %d is not active: \"%s\" (%d)",
+			      port, ibv_port_state_str(port_attr.state),
+			      port_attr.state);
+
+		/* Allocate protection domain. */
+		pd = ibv_alloc_pd(ctx);
+		if (pd == NULL) {
+			ERROR("PD allocation failure");
+			err = ENOMEM;
+			goto port_error;
+		}
+
+		mlx5_dev[idx].ports |= test;
+
+		/* from rte_ethdev.c */
+		priv = rte_zmalloc("ethdev private structure",
+				   sizeof(*priv),
+				   RTE_CACHE_LINE_SIZE);
+		if (priv == NULL) {
+			ERROR("priv allocation failure");
+			err = ENOMEM;
+			goto port_error;
+		}
+
+		priv->ctx = ctx;
+		priv->device_attr = device_attr;
+		priv->port = port;
+		priv->pd = pd;
+		priv->mtu = ETHER_MTU;
+		priv->mps = mps; /* Enable MPW by default if supported. */
+		priv->cqe_comp = 1; /* Enable compression by default. */
+		err = mlx5_args(priv, pci_dev->devargs);
+		if (err) {
+			ERROR("failed to process device arguments: %s",
+			      strerror(err));
+			goto port_error;
+		}
+		if (ibv_exp_query_device(ctx, &exp_device_attr)) {
+			ERROR("ibv_exp_query_device() failed");
+			goto port_error;
+		}
+
+		priv->hw_csum =
+			((exp_device_attr.exp_device_cap_flags &
+			  IBV_EXP_DEVICE_RX_CSUM_TCP_UDP_PKT) &&
+			 (exp_device_attr.exp_device_cap_flags &
+			  IBV_EXP_DEVICE_RX_CSUM_IP_PKT));
+		DEBUG("checksum offloading is %ssupported",
+		      (priv->hw_csum ? "" : "not "));
+
+		priv->hw_csum_l2tun = !!(exp_device_attr.exp_device_cap_flags &
+					 IBV_EXP_DEVICE_VXLAN_SUPPORT);
+		DEBUG("L2 tunnel checksum offloads are %ssupported",
+		      (priv->hw_csum_l2tun ? "" : "not "));
+
+		priv->ind_table_max_size = exp_device_attr.rx_hash_caps.max_rwq_indirection_table_size;
+		/* Remove this check once DPDK supports larger/variable
+		 * indirection tables. */
+		if (priv->ind_table_max_size > (unsigned int)RSS_INDIRECTION_TABLE_SIZE)
+			priv->ind_table_max_size = RSS_INDIRECTION_TABLE_SIZE;
+		DEBUG("maximum RX indirection table size is %u",
+		      priv->ind_table_max_size);
+		priv->hw_vlan_strip = !!(exp_device_attr.wq_vlan_offloads_cap &
+					 IBV_EXP_RECEIVE_WQ_CVLAN_STRIP);
+		DEBUG("VLAN stripping is %ssupported",
+		      (priv->hw_vlan_strip ? "" : "not "));
+
+		priv->hw_fcs_strip = !!(exp_device_attr.exp_device_cap_flags &
+					IBV_EXP_DEVICE_SCATTER_FCS);
+		DEBUG("FCS stripping configuration is %ssupported",
+		      (priv->hw_fcs_strip ? "" : "not "));
+
+		priv->hw_padding = !!exp_device_attr.rx_pad_end_addr_align;
+		DEBUG("hardware RX end alignment padding is %ssupported",
+		      (priv->hw_padding ? "" : "not "));
+
+		priv_get_num_vfs(priv, &num_vfs);
+		priv->sriov = (num_vfs || sriov);
+		if (priv->mps && !mps) {
+			ERROR("multi-packet send not supported on this device"
+			      " (" MLX5_TXQ_MPW_EN ")");
+			err = ENOTSUP;
+			goto port_error;
+		}
+		/* Allocate and register default RSS hash keys. */
+		priv->rss_conf = rte_calloc(__func__, hash_rxq_init_n,
+					    sizeof((*priv->rss_conf)[0]), 0);
+		if (priv->rss_conf == NULL) {
+			err = ENOMEM;
+			goto port_error;
+		}
+		err = rss_hash_rss_conf_new_key(priv,
+						rss_hash_default_key,
+						rss_hash_default_key_len,
+						ETH_RSS_PROTO_MASK);
+		if (err)
+			goto port_error;
+		/* Configure the first MAC address by default. */
+		if (priv_get_mac(priv, &mac.addr_bytes)) {
+			ERROR("cannot get MAC address, is mlx5_en loaded?"
+			      " (errno: %s)", strerror(errno));
+			goto port_error;
+		}
+		INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
+		     priv->port,
+		     mac.addr_bytes[0], mac.addr_bytes[1],
+		     mac.addr_bytes[2], mac.addr_bytes[3],
+		     mac.addr_bytes[4], mac.addr_bytes[5]);
+		/* Register MAC address. */
+		claim_zero(priv_mac_addr_add(priv, 0,
+					     (const uint8_t (*)[ETHER_ADDR_LEN])
+					     mac.addr_bytes));
+		/* Initialize FD filters list. */
+		err = fdir_init_filters_list(priv);
+		if (err)
+			goto port_error;
+#ifndef NDEBUG
+		{
+			char ifname[IF_NAMESIZE];
+
+			if (priv_get_ifname(priv, &ifname) == 0)
+				DEBUG("port %u ifname is \"%s\"",
+				      priv->port, ifname);
+			else
+				DEBUG("port %u ifname is unknown", priv->port);
+		}
+#endif
+		/* Get actual MTU if possible. */
+		priv_get_mtu(priv, &priv->mtu);
+		DEBUG("port %u MTU is %u", priv->port, priv->mtu);
+
+		/* from rte_ethdev.c */
+		{
+			char name[RTE_ETH_NAME_MAX_LEN];
+
+			snprintf(name, sizeof(name), "%s port %u",
+				 ibv_get_device_name(ibv_dev), port);
+			eth_dev = rte_eth_dev_allocate(name, RTE_ETH_DEV_PCI);
+		}
+		if (eth_dev == NULL) {
+			ERROR("can not allocate rte ethdev");
+			err = ENOMEM;
+			goto port_error;
+		}
+
+		/* Secondary processes have to use local storage for their
+		 * private data as well as a copy of eth_dev->data, but this
+		 * pointer must not be modified before burst functions are
+		 * actually called. */
+		if (mlx5_is_secondary()) {
+			struct mlx5_secondary_data *sd =
+				&mlx5_secondary_data[eth_dev->data->port_id];
+			sd->primary_priv = eth_dev->data->dev_private;
+			if (sd->primary_priv == NULL) {
+				ERROR("no private data for port %u",
+						eth_dev->data->port_id);
+				err = EINVAL;
+				goto port_error;
+			}
+			sd->shared_dev_data = eth_dev->data;
+			rte_spinlock_init(&sd->lock);
+			memcpy(sd->data.name, sd->shared_dev_data->name,
+				   sizeof(sd->data.name));
+			sd->data.dev_private = priv;
+			sd->data.rx_mbuf_alloc_failed = 0;
+			sd->data.mtu = ETHER_MTU;
+			sd->data.port_id = sd->shared_dev_data->port_id;
+			sd->data.mac_addrs = priv->mac;
+			eth_dev->tx_pkt_burst = mlx5_tx_burst_secondary_setup;
+			eth_dev->rx_pkt_burst = mlx5_rx_burst_secondary_setup;
+		} else {
+			eth_dev->data->dev_private = priv;
+			eth_dev->data->rx_mbuf_alloc_failed = 0;
+			eth_dev->data->mtu = ETHER_MTU;
+			eth_dev->data->mac_addrs = priv->mac;
+		}
+
+		eth_dev->pci_dev = pci_dev;
+		rte_eth_copy_pci_info(eth_dev, pci_dev);
+		eth_dev->driver = &mlx5_driver;
+		priv->dev = eth_dev;
+		eth_dev->dev_ops = &mlx5_dev_ops;
+
+		TAILQ_INIT(&eth_dev->link_intr_cbs);
+
+		/* Bring Ethernet device up. */
+		DEBUG("forcing Ethernet interface up");
+		priv_set_flags(priv, ~IFF_UP, IFF_UP);
+		continue;
+
+port_error:
+		if (priv) {
+			rte_free(priv->rss_conf);
+			rte_free(priv);
+		}
+		if (pd)
+			claim_zero(ibv_dealloc_pd(pd));
+		if (ctx)
+			claim_zero(ibv_close_device(ctx));
+		break;
+	}
+
+	/*
+	 * XXX if something went wrong in the loop above, there is a resource
+	 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
+	 * long as the dpdk does not provide a way to deallocate a ethdev and a
+	 * way to enumerate the registered ethdevs to free the previous ones.
+	 */
+
+	/* no port found, complain */
+	if (!mlx5_dev[idx].ports) {
+		err = ENODEV;
+		goto error;
+	}
+
+error:
+	if (attr_ctx)
+		claim_zero(ibv_close_device(attr_ctx));
+	if (list)
+		ibv_free_device_list(list);
+	assert(err >= 0);
+	return -err;
+}
+
+static const struct rte_pci_id mlx5_pci_id_map[] = {
+	{
+		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
+			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
+	},
+	{
+		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
+			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
+	},
+	{
+		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
+			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
+	},
+	{
+		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
+			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
+	},
+	{
+		.vendor_id = 0
+	}
+};
+
+static struct eth_driver mlx5_driver = {
+	.pci_drv = {
+		.name = MLX5_DRIVER_NAME,
+		.id_table = mlx5_pci_id_map,
+		.devinit = mlx5_pci_devinit,
+		.drv_flags = RTE_PCI_DRV_INTR_LSC,
+	},
+	.dev_private_size = sizeof(struct priv)
+};
+
+/**
+ * Driver initialization routine.
+ */
+static int
+rte_mlx5_pmd_init(const char *name, const char *args)
+{
+	(void)name;
+	(void)args;
+	/*
+	 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
+	 * huge pages. Calling ibv_fork_init() during init allows
+	 * applications to use fork() safely for purposes other than
+	 * using this PMD, which is not supported in forked processes.
+	 */
+	setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
+	ibv_fork_init();
+	rte_eal_pci_register(&mlx5_driver.pci_drv);
+	return 0;
+}
+
+static struct rte_driver rte_mlx5_driver = {
+	.type = PMD_PDEV,
+	.init = rte_mlx5_pmd_init,
+};
+
+PMD_REGISTER_DRIVER(rte_mlx5_driver, mlx5);
+DRIVER_REGISTER_PCI_TABLE(mlx5, mlx5_pci_id_map);
diff --git a/src/dpdk/drivers/net/mlx5/mlx5.h b/src/dpdk/drivers/net/mlx5/mlx5.h
new file mode 100644
index 00000000..3a866098
--- /dev/null
+++ b/src/dpdk/drivers/net/mlx5/mlx5.h
@@ -0,0 +1,267 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2015 6WIND S.A.
+ *   Copyright 2015 Mellanox.
+ *
+ *   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.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#ifndef RTE_PMD_MLX5_H_
+#define RTE_PMD_MLX5_H_
+
+#include <stddef.h>
+#include <stdint.h>
+#include <limits.h>
+#include <net/if.h>
+#include <netinet/in.h>
+
+/* Verbs header. */
+/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <infiniband/verbs.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+/* DPDK headers don't like -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <rte_ether.h>
+#include <rte_ethdev.h>
+#include <rte_spinlock.h>
+#include <rte_interrupts.h>
+#include <rte_errno.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+#include "mlx5_utils.h"
+#include "mlx5_rxtx.h"
+#include "mlx5_autoconf.h"
+#include "mlx5_defs.h"
+
+#if !defined(HAVE_VERBS_IBV_EXP_CQ_COMPRESSED_CQE) || \
+	!defined(HAVE_VERBS_MLX5_ETH_VLAN_INLINE_HEADER_SIZE)
+#error Mellanox OFED >= 3.3 is required, please refer to the documentation.
+#endif
+
+enum {
+	PCI_VENDOR_ID_MELLANOX = 0x15b3,
+};
+
+enum {
+	PCI_DEVICE_ID_MELLANOX_CONNECTX4 = 0x1013,
+	PCI_DEVICE_ID_MELLANOX_CONNECTX4VF = 0x1014,
+	PCI_DEVICE_ID_MELLANOX_CONNECTX4LX = 0x1015,
+	PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF = 0x1016,
+};
+
+struct priv {
+	struct rte_eth_dev *dev; /* Ethernet device. */
+	struct ibv_context *ctx; /* Verbs context. */
+	struct ibv_device_attr device_attr; /* Device properties. */
+	struct ibv_pd *pd; /* Protection Domain. */
+	/*
+	 * MAC addresses array and configuration bit-field.
+	 * An extra entry that cannot be modified by the DPDK is reserved
+	 * for broadcast frames (destination MAC address ff:ff:ff:ff:ff:ff).
+	 */
+	struct ether_addr mac[MLX5_MAX_MAC_ADDRESSES];
+	BITFIELD_DECLARE(mac_configured, uint32_t, MLX5_MAX_MAC_ADDRESSES);
+	uint16_t vlan_filter[MLX5_MAX_VLAN_IDS]; /* VLAN filters table. */
+	unsigned int vlan_filter_n; /* Number of configured VLAN filters. */
+	/* Device properties. */
+	uint16_t mtu; /* Configured MTU. */
+	uint8_t port; /* Physical port number. */
+	unsigned int started:1; /* Device started, flows enabled. */
+	unsigned int promisc_req:1; /* Promiscuous mode requested. */
+	unsigned int allmulti_req:1; /* All multicast mode requested. */
+	unsigned int hw_csum:1; /* Checksum offload is supported. */
+	unsigned int hw_csum_l2tun:1; /* Same for L2 tunnels. */
+	unsigned int hw_vlan_strip:1; /* VLAN stripping is supported. */
+	unsigned int hw_fcs_strip:1; /* FCS stripping is supported. */
+	unsigned int hw_padding:1; /* End alignment padding is supported. */
+	unsigned int sriov:1; /* This is a VF or PF with VF devices. */
+	unsigned int mps:1; /* Whether multi-packet send is supported. */
+	unsigned int cqe_comp:1; /* Whether CQE compression is enabled. */
+	unsigned int pending_alarm:1; /* An alarm is pending. */
+	unsigned int txq_inline; /* Maximum packet size for inlining. */
+	unsigned int txqs_inline; /* Queue number threshold for inlining. */
+	/* RX/TX queues. */
+	unsigned int rxqs_n; /* RX queues array size. */
+	unsigned int txqs_n; /* TX queues array size. */
+	struct rxq *(*rxqs)[]; /* RX queues. */
+	struct txq *(*txqs)[]; /* TX queues. */
+	/* Indirection tables referencing all RX WQs. */
+	struct ibv_exp_rwq_ind_table *(*ind_tables)[];
+	unsigned int ind_tables_n; /* Number of indirection tables. */
+	unsigned int ind_table_max_size; /* Maximum indirection table size. */
+	/* Hash RX QPs feeding the indirection table. */
+	struct hash_rxq (*hash_rxqs)[];
+	unsigned int hash_rxqs_n; /* Hash RX QPs array size. */
+	/* RSS configuration array indexed by hash RX queue type. */
+	struct rte_eth_rss_conf *(*rss_conf)[];
+	uint64_t rss_hf; /* RSS DPDK bit field of active RSS. */
+	struct rte_intr_handle intr_handle; /* Interrupt handler. */
+	unsigned int (*reta_idx)[]; /* RETA index table. */
+	unsigned int reta_idx_n; /* RETA index size. */
+	struct fdir_filter_list *fdir_filter_list; /* Flow director rules. */
+	rte_spinlock_t lock; /* Lock for control functions. */
+};
+
+/* Local storage for secondary process data. */
+struct mlx5_secondary_data {
+	struct rte_eth_dev_data data; /* Local device data. */
+	struct priv *primary_priv; /* Private structure from primary. */
+	struct rte_eth_dev_data *shared_dev_data; /* Shared device data. */
+	rte_spinlock_t lock; /* Port configuration lock. */
+} mlx5_secondary_data[RTE_MAX_ETHPORTS];
+
+/**
+ * Lock private structure to protect it from concurrent access in the
+ * control path.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ */
+static inline void
+priv_lock(struct priv *priv)
+{
+	rte_spinlock_lock(&priv->lock);
+}
+
+/**
+ * Unlock private structure.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ */
+static inline void
+priv_unlock(struct priv *priv)
+{
+	rte_spinlock_unlock(&priv->lock);
+}
+
+/* mlx5.c */
+
+int mlx5_getenv_int(const char *);
+
+/* mlx5_ethdev.c */
+
+struct priv *mlx5_get_priv(struct rte_eth_dev *dev);
+int mlx5_is_secondary(void);
+int priv_get_ifname(const struct priv *, char (*)[IF_NAMESIZE]);
+int priv_ifreq(const struct priv *, int req, struct ifreq *);
+int priv_get_num_vfs(struct priv *, uint16_t *);
+int priv_get_mtu(struct priv *, uint16_t *);
+int priv_set_flags(struct priv *, unsigned int, unsigned int);
+int mlx5_dev_configure(struct rte_eth_dev *);
+void mlx5_dev_infos_get(struct rte_eth_dev *, struct rte_eth_dev_info *);
+const uint32_t *mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev);
+int mlx5_link_update(struct rte_eth_dev *, int);
+int mlx5_dev_set_mtu(struct rte_eth_dev *, uint16_t);
+int mlx5_dev_get_flow_ctrl(struct rte_eth_dev *, struct rte_eth_fc_conf *);
+int mlx5_dev_set_flow_ctrl(struct rte_eth_dev *, struct rte_eth_fc_conf *);
+int mlx5_ibv_device_to_pci_addr(const struct ibv_device *,
+				struct rte_pci_addr *);
+void mlx5_dev_link_status_handler(void *);
+void mlx5_dev_interrupt_handler(struct rte_intr_handle *, void *);
+void priv_dev_interrupt_handler_uninstall(struct priv *, struct rte_eth_dev *);
+void priv_dev_interrupt_handler_install(struct priv *, struct rte_eth_dev *);
+int mlx5_set_link_down(struct rte_eth_dev *dev);
+int mlx5_set_link_up(struct rte_eth_dev *dev);
+struct priv *mlx5_secondary_data_setup(struct priv *priv);
+void priv_select_tx_function(struct priv *);
+void priv_select_rx_function(struct priv *);
+
+/* mlx5_mac.c */
+
+int priv_get_mac(struct priv *, uint8_t (*)[ETHER_ADDR_LEN]);
+void hash_rxq_mac_addrs_del(struct hash_rxq *);
+void priv_mac_addrs_disable(struct priv *);
+void mlx5_mac_addr_remove(struct rte_eth_dev *, uint32_t);
+int hash_rxq_mac_addrs_add(struct hash_rxq *);
+int priv_mac_addr_add(struct priv *, unsigned int,
+		      const uint8_t (*)[ETHER_ADDR_LEN]);
+int priv_mac_addrs_enable(struct priv *);
+void mlx5_mac_addr_add(struct rte_eth_dev *, struct ether_addr *, uint32_t,
+		       uint32_t);
+void mlx5_mac_addr_set(struct rte_eth_dev *, struct ether_addr *);
+
+/* mlx5_rss.c */
+
+int rss_hash_rss_conf_new_key(struct priv *, const uint8_t *, unsigned int,
+			      uint64_t);
+int mlx5_rss_hash_update(struct rte_eth_dev *, struct rte_eth_rss_conf *);
+int mlx5_rss_hash_conf_get(struct rte_eth_dev *, struct rte_eth_rss_conf *);
+int priv_rss_reta_index_resize(struct priv *, unsigned int);
+int mlx5_dev_rss_reta_query(struct rte_eth_dev *,
+			    struct rte_eth_rss_reta_entry64 *, uint16_t);
+int mlx5_dev_rss_reta_update(struct rte_eth_dev *,
+			     struct rte_eth_rss_reta_entry64 *, uint16_t);
+
+/* mlx5_rxmode.c */
+
+int priv_special_flow_enable(struct priv *, enum hash_rxq_flow_type);
+void priv_special_flow_disable(struct priv *, enum hash_rxq_flow_type);
+int priv_special_flow_enable_all(struct priv *);
+void priv_special_flow_disable_all(struct priv *);
+void mlx5_promiscuous_enable(struct rte_eth_dev *);
+void mlx5_promiscuous_disable(struct rte_eth_dev *);
+void mlx5_allmulticast_enable(struct rte_eth_dev *);
+void mlx5_allmulticast_disable(struct rte_eth_dev *);
+
+/* mlx5_stats.c */
+
+void mlx5_stats_get(struct rte_eth_dev *, struct rte_eth_stats *);
+void mlx5_stats_reset(struct rte_eth_dev *);
+
+/* mlx5_vlan.c */
+
+int mlx5_vlan_filter_set(struct rte_eth_dev *, uint16_t, int);
+void mlx5_vlan_offload_set(struct rte_eth_dev *, int);
+void mlx5_vlan_strip_queue_set(struct rte_eth_dev *, uint16_t, int);
+
+/* mlx5_trigger.c */
+
+int mlx5_dev_start(struct rte_eth_dev *);
+void mlx5_dev_stop(struct rte_eth_dev *);
+
+/* mlx5_fdir.c */
+
+int fdir_init_filters_list(struct priv *);
+void priv_fdir_delete_filters_list(struct priv *);
+void priv_fdir_disable(struct priv *);
+void priv_fdir_enable(struct priv *);
+int mlx5_dev_filter_ctrl(struct rte_eth_dev *, enum rte_filter_type,
+			 enum rte_filter_op, void *);
+
+#endif /* RTE_PMD_MLX5_H_ */
diff --git a/src/dpdk/drivers/net/mlx5/mlx5_defs.h b/src/dpdk/drivers/net/mlx5/mlx5_defs.h
new file mode 100644
index 00000000..cc2a6f3e
--- /dev/null
+++ b/src/dpdk/drivers/net/mlx5/mlx5_defs.h
@@ -0,0 +1,82 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2015 6WIND S.A.
+ *   Copyright 2015 Mellanox.
+ *
+ *   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.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#ifndef RTE_PMD_MLX5_DEFS_H_
+#define RTE_PMD_MLX5_DEFS_H_
+
+#include "mlx5_autoconf.h"
+
+/* Reported driver name. */
+#define MLX5_DRIVER_NAME "librte_pmd_mlx5"
+
+/* Maximum number of simultaneous MAC addresses. */
+#define MLX5_MAX_MAC_ADDRESSES 128
+
+/* Maximum number of simultaneous VLAN filters. */
+#define MLX5_MAX_VLAN_IDS 128
+
+/* Maximum number of special flows. */
+#define MLX5_MAX_SPECIAL_FLOWS 4
+
+/*
+ * Request TX completion every time descriptors reach this threshold since
+ * the previous request. Must be a power of two for performance reasons.
+ */
+#define MLX5_TX_COMP_THRESH 32
+
+/* RSS Indirection table size. */
+#define RSS_INDIRECTION_TABLE_SIZE 256
+
+/*
+ * Maximum number of cached Memory Pools (MPs) per TX queue. Each RTE MP
+ * from which buffers are to be transmitted will have to be mapped by this
+ * driver to their own Memory Region (MR). This is a slow operation.
+ *
+ * This value is always 1 for RX queues.
+ */
+#ifndef MLX5_PMD_TX_MP_CACHE
+#define MLX5_PMD_TX_MP_CACHE 8
+#endif
+
+/*
+ * If defined, only use software counters. The PMD will never ask the hardware
+ * for these, and many of them won't be available.
+ */
+#ifndef MLX5_PMD_SOFT_COUNTERS
+#define MLX5_PMD_SOFT_COUNTERS 1
+#endif
+
+/* Alarm timeout. */
+#define MLX5_ALARM_TIMEOUT_US 100000
+
+#endif /* RTE_PMD_MLX5_DEFS_H_ */
diff --git a/src/dpdk/drivers/net/mlx5/mlx5_ethdev.c b/src/dpdk/drivers/net/mlx5/mlx5_ethdev.c
new file mode 100644
index 00000000..130e15d5
--- /dev/null
+++ b/src/dpdk/drivers/net/mlx5/mlx5_ethdev.c
@@ -0,0 +1,1417 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2015 6WIND S.A.
+ *   Copyright 2015 Mellanox.
+ *
+ *   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.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <stddef.h>
+#include <assert.h>
+#include <unistd.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <string.h>
+#include <stdlib.h>
+#include <errno.h>
+#include <dirent.h>
+#include <net/if.h>
+#include <sys/ioctl.h>
+#include <sys/socket.h>
+#include <netinet/in.h>
+#include <linux/ethtool.h>
+#include <linux/sockios.h>
+#include <fcntl.h>
+
+/* DPDK headers don't like -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <rte_atomic.h>
+#include <rte_ethdev.h>
+#include <rte_mbuf.h>
+#include <rte_common.h>
+#include <rte_interrupts.h>
+#include <rte_alarm.h>
+#include <rte_malloc.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+#include "mlx5.h"
+#include "mlx5_rxtx.h"
+#include "mlx5_utils.h"
+
+/**
+ * Return private structure associated with an Ethernet device.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ *
+ * @return
+ *   Pointer to private structure.
+ */
+struct priv *
+mlx5_get_priv(struct rte_eth_dev *dev)
+{
+	struct mlx5_secondary_data *sd;
+
+	if (!mlx5_is_secondary())
+		return dev->data->dev_private;
+	sd = &mlx5_secondary_data[dev->data->port_id];
+	return sd->data.dev_private;
+}
+
+/**
+ * Check if running as a secondary process.
+ *
+ * @return
+ *   Nonzero if running as a secondary process.
+ */
+inline int
+mlx5_is_secondary(void)
+{
+	return rte_eal_process_type() != RTE_PROC_PRIMARY;
+}
+
+/**
+ * Get interface name from private structure.
+ *
+ * @param[in] priv
+ *   Pointer to private structure.
+ * @param[out] ifname
+ *   Interface name output buffer.
+ *
+ * @return
+ *   0 on success, -1 on failure and errno is set.
+ */
+int
+priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE])
+{
+	DIR *dir;
+	struct dirent *dent;
+	unsigned int dev_type = 0;
+	unsigned int dev_port_prev = ~0u;
+	char match[IF_NAMESIZE] = "";
+
+	{
+		MKSTR(path, "%s/device/net", priv->ctx->device->ibdev_path);
+
+		dir = opendir(path);
+		if (dir == NULL)
+			return -1;
+	}
+	while ((dent = readdir(dir)) != NULL) {
+		char *name = dent->d_name;
+		FILE *file;
+		unsigned int dev_port;
+		int r;
+
+		if ((name[0] == '.') &&
+		    ((name[1] == '\0') ||
+		     ((name[1] == '.') && (name[2] == '\0'))))
+			continue;
+
+		MKSTR(path, "%s/device/net/%s/%s",
+		      priv->ctx->device->ibdev_path, name,
+		      (dev_type ? "dev_id" : "dev_port"));
+
+		file = fopen(path, "rb");
+		if (file == NULL) {
+			if (errno != ENOENT)
+				continue;
+			/*
+			 * Switch to dev_id when dev_port does not exist as
+			 * is the case with Linux kernel versions < 3.15.
+			 */
+try_dev_id:
+			match[0] = '\0';
+			if (dev_type)
+				break;
+			dev_type = 1;
+			dev_port_prev = ~0u;
+			rewinddir(dir);
+			continue;
+		}
+		r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
+		fclose(file);
+		if (r != 1)
+			continue;
+		/*
+		 * Switch to dev_id when dev_port returns the same value for
+		 * all ports. May happen when using a MOFED release older than
+		 * 3.0 with a Linux kernel >= 3.15.
+		 */
+		if (dev_port == dev_port_prev)
+			goto try_dev_id;
+		dev_port_prev = dev_port;
+		if (dev_port == (priv->port - 1u))
+			snprintf(match, sizeof(match), "%s", name);
+	}
+	closedir(dir);
+	if (match[0] == '\0')
+		return -1;
+	strncpy(*ifname, match, sizeof(*ifname));
+	return 0;
+}
+
+/**
+ * Read from sysfs entry.
+ *
+ * @param[in] priv
+ *   Pointer to private structure.
+ * @param[in] entry
+ *   Entry name relative to sysfs path.
+ * @param[out] buf
+ *   Data output buffer.
+ * @param size
+ *   Buffer size.
+ *
+ * @return
+ *   0 on success, -1 on failure and errno is set.
+ */
+static int
+priv_sysfs_read(const struct priv *priv, const char *entry,
+		char *buf, size_t size)
+{
+	char ifname[IF_NAMESIZE];
+	FILE *file;
+	int ret;
+	int err;
+
+	if (priv_get_ifname(priv, &ifname))
+		return -1;
+
+	MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
+	      ifname, entry);
+
+	file = fopen(path, "rb");
+	if (file == NULL)
+		return -1;
+	ret = fread(buf, 1, size, file);
+	err = errno;
+	if (((size_t)ret < size) && (ferror(file)))
+		ret = -1;
+	else
+		ret = size;
+	fclose(file);
+	errno = err;
+	return ret;
+}
+
+/**
+ * Write to sysfs entry.
+ *
+ * @param[in] priv
+ *   Pointer to private structure.
+ * @param[in] entry
+ *   Entry name relative to sysfs path.
+ * @param[in] buf
+ *   Data buffer.
+ * @param size
+ *   Buffer size.
+ *
+ * @return
+ *   0 on success, -1 on failure and errno is set.
+ */
+static int
+priv_sysfs_write(const struct priv *priv, const char *entry,
+		 char *buf, size_t size)
+{
+	char ifname[IF_NAMESIZE];
+	FILE *file;
+	int ret;
+	int err;
+
+	if (priv_get_ifname(priv, &ifname))
+		return -1;
+
+	MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
+	      ifname, entry);
+
+	file = fopen(path, "wb");
+	if (file == NULL)
+		return -1;
+	ret = fwrite(buf, 1, size, file);
+	err = errno;
+	if (((size_t)ret < size) || (ferror(file)))
+		ret = -1;
+	else
+		ret = size;
+	fclose(file);
+	errno = err;
+	return ret;
+}
+
+/**
+ * Get unsigned long sysfs property.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param[in] name
+ *   Entry name relative to sysfs path.
+ * @param[out] value
+ *   Value output buffer.
+ *
+ * @return
+ *   0 on success, -1 on failure and errno is set.
+ */
+static int
+priv_get_sysfs_ulong(struct priv *priv, const char *name, unsigned long *value)
+{
+	int ret;
+	unsigned long value_ret;
+	char value_str[32];
+
+	ret = priv_sysfs_read(priv, name, value_str, (sizeof(value_str) - 1));
+	if (ret == -1) {
+		DEBUG("cannot read %s value from sysfs: %s",
+		      name, strerror(errno));
+		return -1;
+	}
+	value_str[ret] = '\0';
+	errno = 0;
+	value_ret = strtoul(value_str, NULL, 0);
+	if (errno) {
+		DEBUG("invalid %s value `%s': %s", name, value_str,
+		      strerror(errno));
+		return -1;
+	}
+	*value = value_ret;
+	return 0;
+}
+
+/**
+ * Set unsigned long sysfs property.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param[in] name
+ *   Entry name relative to sysfs path.
+ * @param value
+ *   Value to set.
+ *
+ * @return
+ *   0 on success, -1 on failure and errno is set.
+ */
+static int
+priv_set_sysfs_ulong(struct priv *priv, const char *name, unsigned long value)
+{
+	int ret;
+	MKSTR(value_str, "%lu", value);
+
+	ret = priv_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1));
+	if (ret == -1) {
+		DEBUG("cannot write %s `%s' (%lu) to sysfs: %s",
+		      name, value_str, value, strerror(errno));
+		return -1;
+	}
+	return 0;
+}
+
+/**
+ * Perform ifreq ioctl() on associated Ethernet device.
+ *
+ * @param[in] priv
+ *   Pointer to private structure.
+ * @param req
+ *   Request number to pass to ioctl().
+ * @param[out] ifr
+ *   Interface request structure output buffer.
+ *
+ * @return
+ *   0 on success, -1 on failure and errno is set.
+ */
+int
+priv_ifreq(const struct priv *priv, int req, struct ifreq *ifr)
+{
+	int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
+	int ret = -1;
+
+	if (sock == -1)
+		return ret;
+	if (priv_get_ifname(priv, &ifr->ifr_name) == 0)
+		ret = ioctl(sock, req, ifr);
+	close(sock);
+	return ret;
+}
+
+/**
+ * Return the number of active VFs for the current device.
+ *
+ * @param[in] priv
+ *   Pointer to private structure.
+ * @param[out] num_vfs
+ *   Number of active VFs.
+ *
+ * @return
+ *   0 on success, -1 on failure and errno is set.
+ */
+int
+priv_get_num_vfs(struct priv *priv, uint16_t *num_vfs)
+{
+	/* The sysfs entry name depends on the operating system. */
+	const char **name = (const char *[]){
+		"device/sriov_numvfs",
+		"device/mlx5_num_vfs",
+		NULL,
+	};
+	int ret;
+
+	do {
+		unsigned long ulong_num_vfs;
+
+		ret = priv_get_sysfs_ulong(priv, *name, &ulong_num_vfs);
+		if (!ret)
+			*num_vfs = ulong_num_vfs;
+	} while (*(++name) && ret);
+	return ret;
+}
+
+/**
+ * Get device MTU.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param[out] mtu
+ *   MTU value output buffer.
+ *
+ * @return
+ *   0 on success, -1 on failure and errno is set.
+ */
+int
+priv_get_mtu(struct priv *priv, uint16_t *mtu)
+{
+	unsigned long ulong_mtu;
+
+	if (priv_get_sysfs_ulong(priv, "mtu", &ulong_mtu) == -1)
+		return -1;
+	*mtu = ulong_mtu;
+	return 0;
+}
+
+/**
+ * Set device MTU.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param mtu
+ *   MTU value to set.
+ *
+ * @return
+ *   0 on success, -1 on failure and errno is set.
+ */
+static int
+priv_set_mtu(struct priv *priv, uint16_t mtu)
+{
+	uint16_t new_mtu;
+
+	if (priv_set_sysfs_ulong(priv, "mtu", mtu) ||
+	    priv_get_mtu(priv, &new_mtu))
+		return -1;
+	if (new_mtu == mtu)
+		return 0;
+	errno = EINVAL;
+	return -1;
+}
+
+/**
+ * Set device flags.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param keep
+ *   Bitmask for flags that must remain untouched.
+ * @param flags
+ *   Bitmask for flags to modify.
+ *
+ * @return
+ *   0 on success, -1 on failure and errno is set.
+ */
+int
+priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
+{
+	unsigned long tmp;
+
+	if (priv_get_sysfs_ulong(priv, "flags", &tmp) == -1)
+		return -1;
+	tmp &= keep;
+	tmp |= (flags & (~keep));
+	return priv_set_sysfs_ulong(priv, "flags", tmp);
+}
+
+/**
+ * Ethernet device configuration.
+ *
+ * Prepare the driver for a given number of TX and RX queues.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+dev_configure(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	unsigned int rxqs_n = dev->data->nb_rx_queues;
+	unsigned int txqs_n = dev->data->nb_tx_queues;
+	unsigned int i;
+	unsigned int j;
+	unsigned int reta_idx_n;
+
+	priv->rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
+	priv->rxqs = (void *)dev->data->rx_queues;
+	priv->txqs = (void *)dev->data->tx_queues;
+	if (txqs_n != priv->txqs_n) {
+		INFO("%p: TX queues number update: %u -> %u",
+		     (void *)dev, priv->txqs_n, txqs_n);
+		priv->txqs_n = txqs_n;
+	}
+	if (rxqs_n > priv->ind_table_max_size) {
+		ERROR("cannot handle this many RX queues (%u)", rxqs_n);
+		return EINVAL;
+	}
+	if (rxqs_n == priv->rxqs_n)
+		return 0;
+	INFO("%p: RX queues number update: %u -> %u",
+	     (void *)dev, priv->rxqs_n, rxqs_n);
+	priv->rxqs_n = rxqs_n;
+	/* If the requested number of RX queues is not a power of two, use the
+	 * maximum indirection table size for better balancing.
+	 * The result is always rounded to the next power of two. */
+	reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
+				     priv->ind_table_max_size :
+				     rxqs_n));
+	if (priv_rss_reta_index_resize(priv, reta_idx_n))
+		return ENOMEM;
+	/* When the number of RX queues is not a power of two, the remaining
+	 * table entries are padded with reused WQs and hashes are not spread
+	 * uniformly. */
+	for (i = 0, j = 0; (i != reta_idx_n); ++i) {
+		(*priv->reta_idx)[i] = j;
+		if (++j == rxqs_n)
+			j = 0;
+	}
+	return 0;
+}
+
+/**
+ * DPDK callback for Ethernet device configuration.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ *
+ * @return
+ *   0 on success, negative errno value on failure.
+ */
+int
+mlx5_dev_configure(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	int ret;
+
+	if (mlx5_is_secondary())
+		return -E_RTE_SECONDARY;
+
+	priv_lock(priv);
+	ret = dev_configure(dev);
+	assert(ret >= 0);
+	priv_unlock(priv);
+	return -ret;
+}
+
+/**
+ * DPDK callback to get information about the device.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param[out] info
+ *   Info structure output buffer.
+ */
+void
+mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
+{
+	struct priv *priv = mlx5_get_priv(dev);
+	unsigned int max;
+	char ifname[IF_NAMESIZE];
+
+	priv_lock(priv);
+	/* FIXME: we should ask the device for these values. */
+	info->min_rx_bufsize = 32;
+	info->max_rx_pktlen = 65536;
+	/*
+	 * Since we need one CQ per QP, the limit is the minimum number
+	 * between the two values.
+	 */
+	max = ((priv->device_attr.max_cq > priv->device_attr.max_qp) ?
+	       priv->device_attr.max_qp : priv->device_attr.max_cq);
+	/* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
+	if (max >= 65535)
+		max = 65535;
+	info->max_rx_queues = max;
+	info->max_tx_queues = max;
+	info->max_mac_addrs = RTE_DIM(priv->mac);
+	info->rx_offload_capa =
+		(priv->hw_csum ?
+		 (DEV_RX_OFFLOAD_IPV4_CKSUM |
+		  DEV_RX_OFFLOAD_UDP_CKSUM |
+		  DEV_RX_OFFLOAD_TCP_CKSUM) :
+		 0);
+	if (!priv->mps)
+		info->tx_offload_capa = DEV_TX_OFFLOAD_VLAN_INSERT;
+	if (priv->hw_csum)
+		info->tx_offload_capa |=
+			(DEV_TX_OFFLOAD_IPV4_CKSUM |
+			 DEV_TX_OFFLOAD_UDP_CKSUM |
+			 DEV_TX_OFFLOAD_TCP_CKSUM);
+	if (priv_get_ifname(priv, &ifname) == 0)
+		info->if_index = if_nametoindex(ifname);
+	/* FIXME: RETA update/query API expects the callee to know the size of
+	 * the indirection table, for this PMD the size varies depending on
+	 * the number of RX queues, it becomes impossible to find the correct
+	 * size if it is not fixed.
+	 * The API should be updated to solve this problem. */
+	info->reta_size = priv->ind_table_max_size;
+	info->speed_capa =
+			ETH_LINK_SPEED_1G |
+			ETH_LINK_SPEED_10G |
+			ETH_LINK_SPEED_20G |
+			ETH_LINK_SPEED_25G |
+			ETH_LINK_SPEED_40G |
+			ETH_LINK_SPEED_50G |
+			ETH_LINK_SPEED_56G |
+			ETH_LINK_SPEED_100G;
+	priv_unlock(priv);
+}
+
+const uint32_t *
+mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
+{
+	static const uint32_t ptypes[] = {
+		/* refers to rxq_cq_to_pkt_type() */
+		RTE_PTYPE_L3_IPV4,
+		RTE_PTYPE_L3_IPV6,
+		RTE_PTYPE_INNER_L3_IPV4,
+		RTE_PTYPE_INNER_L3_IPV6,
+		RTE_PTYPE_UNKNOWN
+
+	};
+
+	if (dev->rx_pkt_burst == mlx5_rx_burst)
+		return ptypes;
+	return NULL;
+}
+
+/**
+ * DPDK callback to retrieve physical link information (unlocked version).
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param wait_to_complete
+ *   Wait for request completion (ignored).
+ */
+static int
+mlx5_link_update_unlocked(struct rte_eth_dev *dev, int wait_to_complete)
+{
+	struct priv *priv = mlx5_get_priv(dev);
+	struct ethtool_cmd edata = {
+		.cmd = ETHTOOL_GSET
+	};
+	struct ifreq ifr;
+	struct rte_eth_link dev_link;
+	int link_speed = 0;
+
+	(void)wait_to_complete;
+	if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
+		WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
+		return -1;
+	}
+	memset(&dev_link, 0, sizeof(dev_link));
+	dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
+				(ifr.ifr_flags & IFF_RUNNING));
+	ifr.ifr_data = (void *)&edata;
+	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
+		WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
+		     strerror(errno));
+		return -1;
+	}
+	link_speed = ethtool_cmd_speed(&edata);
+	if (link_speed == -1)
+		dev_link.link_speed = 0;
+	else
+		dev_link.link_speed = link_speed;
+	dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
+				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
+	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
+			ETH_LINK_SPEED_FIXED);
+	if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
+		/* Link status changed. */
+		dev->data->dev_link = dev_link;
+		return 0;
+	}
+	/* Link status is still the same. */
+	return -1;
+}
+
+/**
+ * DPDK callback to retrieve physical link information.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param wait_to_complete
+ *   Wait for request completion (ignored).
+ */
+int
+mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
+{
+	struct priv *priv = mlx5_get_priv(dev);
+	int ret;
+
+	priv_lock(priv);
+	ret = mlx5_link_update_unlocked(dev, wait_to_complete);
+	priv_unlock(priv);
+	return ret;
+}
+
+/**
+ * DPDK callback to change the MTU.
+ *
+ * Setting the MTU affects hardware MRU (packets larger than the MTU cannot be
+ * received). Use this as a hint to enable/disable scattered packets support
+ * and improve performance when not needed.
+ * Since failure is not an option, reconfiguring queues on the fly is not
+ * recommended.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param in_mtu
+ *   New MTU.
+ *
+ * @return
+ *   0 on success, negative errno value on failure.
+ */
+int
+mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
+{
+	struct priv *priv = dev->data->dev_private;
+	int ret = 0;
+	unsigned int i;
+	uint16_t (*rx_func)(void *, struct rte_mbuf **, uint16_t) =
+		mlx5_rx_burst;
+	unsigned int max_frame_len;
+	int rehash;
+	int restart = priv->started;
+
+	if (mlx5_is_secondary())
+		return -E_RTE_SECONDARY;
+
+	priv_lock(priv);
+	/* Set kernel interface MTU first. */
+	if (priv_set_mtu(priv, mtu)) {
+		ret = errno;
+		WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
+		     strerror(ret));
+		goto out;
+	} else
+		DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
+	/* Temporarily replace RX handler with a fake one, assuming it has not
+	 * been copied elsewhere. */
+	dev->rx_pkt_burst = removed_rx_burst;
+	/* Make sure everyone has left mlx5_rx_burst() and uses
+	 * removed_rx_burst() instead. */
+	rte_wmb();
+	usleep(1000);
+	/* MTU does not include header and CRC. */
+	max_frame_len = ETHER_HDR_LEN + mtu + ETHER_CRC_LEN;
+	/* Check if at least one queue is going to need a SGE update. */
+	for (i = 0; i != priv->rxqs_n; ++i) {
+		struct rxq *rxq = (*priv->rxqs)[i];
+		unsigned int mb_len;
+		unsigned int size = RTE_PKTMBUF_HEADROOM + max_frame_len;
+		unsigned int sges_n;
+
+		if (rxq == NULL)
+			continue;
+		mb_len = rte_pktmbuf_data_room_size(rxq->mp);
+		assert(mb_len >= RTE_PKTMBUF_HEADROOM);
+		/*
+		 * Determine the number of SGEs needed for a full packet
+		 * and round it to the next power of two.
+		 */
+		sges_n = log2above((size / mb_len) + !!(size % mb_len));
+		if (sges_n != rxq->sges_n)
+			break;
+	}
+	/*
+	 * If all queues have the right number of SGEs, a simple rehash
+	 * of their buffers is enough, otherwise SGE information can only
+	 * be updated in a queue by recreating it. All resources that depend
+	 * on queues (flows, indirection tables) must be recreated as well in
+	 * that case.
+	 */
+	rehash = (i == priv->rxqs_n);
+	if (!rehash) {
+		/* Clean up everything as with mlx5_dev_stop(). */
+		priv_special_flow_disable_all(priv);
+		priv_mac_addrs_disable(priv);
+		priv_destroy_hash_rxqs(priv);
+		priv_fdir_disable(priv);
+		priv_dev_interrupt_handler_uninstall(priv, dev);
+	}
+recover:
+	/* Reconfigure each RX queue. */
+	for (i = 0; (i != priv->rxqs_n); ++i) {
+		struct rxq *rxq = (*priv->rxqs)[i];
+		struct rxq_ctrl *rxq_ctrl =
+			container_of(rxq, struct rxq_ctrl, rxq);
+		int sp;
+		unsigned int mb_len;
+		unsigned int tmp;
+
+		if (rxq == NULL)
+			continue;
+		mb_len = rte_pktmbuf_data_room_size(rxq->mp);
+		assert(mb_len >= RTE_PKTMBUF_HEADROOM);
+		/* Toggle scattered support (sp) if necessary. */
+		sp = (max_frame_len > (mb_len - RTE_PKTMBUF_HEADROOM));
+		/* Provide new values to rxq_setup(). */
+		dev->data->dev_conf.rxmode.jumbo_frame = sp;
+		dev->data->dev_conf.rxmode.max_rx_pkt_len = max_frame_len;
+		if (rehash)
+			ret = rxq_rehash(dev, rxq_ctrl);
+		else
+			ret = rxq_ctrl_setup(dev, rxq_ctrl, rxq->elts_n,
+					     rxq_ctrl->socket, NULL, rxq->mp);
+		if (!ret)
+			continue;
+		/* Attempt to roll back in case of error. */
+		tmp = (mb_len << rxq->sges_n) - RTE_PKTMBUF_HEADROOM;
+		if (max_frame_len != tmp) {
+			max_frame_len = tmp;
+			goto recover;
+		}
+		/* Double fault, disable RX. */
+		break;
+	}
+	/*
+	 * Use a safe RX burst function in case of error, otherwise mimic
+	 * mlx5_dev_start().
+	 */
+	if (ret) {
+		ERROR("unable to reconfigure RX queues, RX disabled");
+		rx_func = removed_rx_burst;
+	} else if (restart &&
+		 !rehash &&
+		 !priv_create_hash_rxqs(priv) &&
+		 !priv_rehash_flows(priv)) {
+		if (dev->data->dev_conf.fdir_conf.mode == RTE_FDIR_MODE_NONE)
+			priv_fdir_enable(priv);
+		priv_dev_interrupt_handler_install(priv, dev);
+	}
+	priv->mtu = mtu;
+	/* Burst functions can now be called again. */
+	rte_wmb();
+	dev->rx_pkt_burst = rx_func;
+out:
+	priv_unlock(priv);
+	assert(ret >= 0);
+	return -ret;
+}
+
+/**
+ * DPDK callback to get flow control status.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param[out] fc_conf
+ *   Flow control output buffer.
+ *
+ * @return
+ *   0 on success, negative errno value on failure.
+ */
+int
+mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
+{
+	struct priv *priv = dev->data->dev_private;
+	struct ifreq ifr;
+	struct ethtool_pauseparam ethpause = {
+		.cmd = ETHTOOL_GPAUSEPARAM
+	};
+	int ret;
+
+	if (mlx5_is_secondary())
+		return -E_RTE_SECONDARY;
+
+	ifr.ifr_data = (void *)&ethpause;
+	priv_lock(priv);
+	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
+		ret = errno;
+		WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
+		     " failed: %s",
+		     strerror(ret));
+		goto out;
+	}
+
+	fc_conf->autoneg = ethpause.autoneg;
+	if (ethpause.rx_pause && ethpause.tx_pause)
+		fc_conf->mode = RTE_FC_FULL;
+	else if (ethpause.rx_pause)
+		fc_conf->mode = RTE_FC_RX_PAUSE;
+	else if (ethpause.tx_pause)
+		fc_conf->mode = RTE_FC_TX_PAUSE;
+	else
+		fc_conf->mode = RTE_FC_NONE;
+	ret = 0;
+
+out:
+	priv_unlock(priv);
+	assert(ret >= 0);
+	return -ret;
+}
+
+/**
+ * DPDK callback to modify flow control parameters.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param[in] fc_conf
+ *   Flow control parameters.
+ *
+ * @return
+ *   0 on success, negative errno value on failure.
+ */
+int
+mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
+{
+	struct priv *priv = dev->data->dev_private;
+	struct ifreq ifr;
+	struct ethtool_pauseparam ethpause = {
+		.cmd = ETHTOOL_SPAUSEPARAM
+	};
+	int ret;
+
+	if (mlx5_is_secondary())
+		return -E_RTE_SECONDARY;
+
+	ifr.ifr_data = (void *)&ethpause;
+	ethpause.autoneg = fc_conf->autoneg;
+	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
+	    (fc_conf->mode & RTE_FC_RX_PAUSE))
+		ethpause.rx_pause = 1;
+	else
+		ethpause.rx_pause = 0;
+
+	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
+	    (fc_conf->mode & RTE_FC_TX_PAUSE))
+		ethpause.tx_pause = 1;
+	else
+		ethpause.tx_pause = 0;
+
+	priv_lock(priv);
+	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
+		ret = errno;
+		WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
+		     " failed: %s",
+		     strerror(ret));
+		goto out;
+	}
+	ret = 0;
+
+out:
+	priv_unlock(priv);
+	assert(ret >= 0);
+	return -ret;
+}
+
+/**
+ * Get PCI information from struct ibv_device.
+ *
+ * @param device
+ *   Pointer to Ethernet device structure.
+ * @param[out] pci_addr
+ *   PCI bus address output buffer.
+ *
+ * @return
+ *   0 on success, -1 on failure and errno is set.
+ */
+int
+mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
+			    struct rte_pci_addr *pci_addr)
+{
+	FILE *file;
+	char line[32];
+	MKSTR(path, "%s/device/uevent", device->ibdev_path);
+
+	file = fopen(path, "rb");
+	if (file == NULL)
+		return -1;
+	while (fgets(line, sizeof(line), file) == line) {
+		size_t len = strlen(line);
+		int ret;
+
+		/* Truncate long lines. */
+		if (len == (sizeof(line) - 1))
+			while (line[(len - 1)] != '\n') {
+				ret = fgetc(file);
+				if (ret == EOF)
+					break;
+				line[(len - 1)] = ret;
+			}
+		/* Extract information. */
+		if (sscanf(line,
+			   "PCI_SLOT_NAME="
+			   "%" SCNx16 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
+			   &pci_addr->domain,
+			   &pci_addr->bus,
+			   &pci_addr->devid,
+			   &pci_addr->function) == 4) {
+			ret = 0;
+			break;
+		}
+	}
+	fclose(file);
+	return 0;
+}
+
+/**
+ * Link status handler.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param dev
+ *   Pointer to the rte_eth_dev structure.
+ *
+ * @return
+ *   Nonzero if the callback process can be called immediately.
+ */
+static int
+priv_dev_link_status_handler(struct priv *priv, struct rte_eth_dev *dev)
+{
+	struct ibv_async_event event;
+	int port_change = 0;
+	int ret = 0;
+
+	/* Read all message and acknowledge them. */
+	for (;;) {
+		if (ibv_get_async_event(priv->ctx, &event))
+			break;
+
+		if (event.event_type == IBV_EVENT_PORT_ACTIVE ||
+		    event.event_type == IBV_EVENT_PORT_ERR)
+			port_change = 1;
+		else
+			DEBUG("event type %d on port %d not handled",
+			      event.event_type, event.element.port_num);
+		ibv_ack_async_event(&event);
+	}
+
+	if (port_change ^ priv->pending_alarm) {
+		struct rte_eth_link *link = &dev->data->dev_link;
+
+		priv->pending_alarm = 0;
+		mlx5_link_update_unlocked(dev, 0);
+		if (((link->link_speed == 0) && link->link_status) ||
+		    ((link->link_speed != 0) && !link->link_status)) {
+			/* Inconsistent status, check again later. */
+			priv->pending_alarm = 1;
+			rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
+					  mlx5_dev_link_status_handler,
+					  dev);
+		} else
+			ret = 1;
+	}
+	return ret;
+}
+
+/**
+ * Handle delayed link status event.
+ *
+ * @param arg
+ *   Registered argument.
+ */
+void
+mlx5_dev_link_status_handler(void *arg)
+{
+	struct rte_eth_dev *dev = arg;
+	struct priv *priv = dev->data->dev_private;
+	int ret;
+
+	priv_lock(priv);
+	assert(priv->pending_alarm == 1);
+	ret = priv_dev_link_status_handler(priv, dev);
+	priv_unlock(priv);
+	if (ret)
+		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC);
+}
+
+/**
+ * Handle interrupts from the NIC.
+ *
+ * @param[in] intr_handle
+ *   Interrupt handler.
+ * @param cb_arg
+ *   Callback argument.
+ */
+void
+mlx5_dev_interrupt_handler(struct rte_intr_handle *intr_handle, void *cb_arg)
+{
+	struct rte_eth_dev *dev = cb_arg;
+	struct priv *priv = dev->data->dev_private;
+	int ret;
+
+	(void)intr_handle;
+	priv_lock(priv);
+	ret = priv_dev_link_status_handler(priv, dev);
+	priv_unlock(priv);
+	if (ret)
+		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC);
+}
+
+/**
+ * Uninstall interrupt handler.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param dev
+ *   Pointer to the rte_eth_dev structure.
+ */
+void
+priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
+{
+	if (!dev->data->dev_conf.intr_conf.lsc)
+		return;
+	rte_intr_callback_unregister(&priv->intr_handle,
+				     mlx5_dev_interrupt_handler,
+				     dev);
+	if (priv->pending_alarm)
+		rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
+	priv->pending_alarm = 0;
+	priv->intr_handle.fd = 0;
+	priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
+}
+
+/**
+ * Install interrupt handler.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param dev
+ *   Pointer to the rte_eth_dev structure.
+ */
+void
+priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
+{
+	int rc, flags;
+
+	if (!dev->data->dev_conf.intr_conf.lsc)
+		return;
+	assert(priv->ctx->async_fd > 0);
+	flags = fcntl(priv->ctx->async_fd, F_GETFL);
+	rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
+	if (rc < 0) {
+		INFO("failed to change file descriptor async event queue");
+		dev->data->dev_conf.intr_conf.lsc = 0;
+	} else {
+		priv->intr_handle.fd = priv->ctx->async_fd;
+		priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
+		rte_intr_callback_register(&priv->intr_handle,
+					   mlx5_dev_interrupt_handler,
+					   dev);
+	}
+}
+
+/**
+ * Change the link state (UP / DOWN).
+ *
+ * @param priv
+ *   Pointer to Ethernet device structure.
+ * @param up
+ *   Nonzero for link up, otherwise link down.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+priv_set_link(struct priv *priv, int up)
+{
+	struct rte_eth_dev *dev = priv->dev;
+	int err;
+
+	if (up) {
+		err = priv_set_flags(priv, ~IFF_UP, IFF_UP);
+		if (err)
+			return err;
+		priv_select_tx_function(priv);
+		priv_select_rx_function(priv);
+	} else {
+		err = priv_set_flags(priv, ~IFF_UP, ~IFF_UP);
+		if (err)
+			return err;
+		dev->rx_pkt_burst = removed_rx_burst;
+		dev->tx_pkt_burst = removed_tx_burst;
+	}
+	return 0;
+}
+
+/**
+ * DPDK callback to bring the link DOWN.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+int
+mlx5_set_link_down(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	int err;
+
+	priv_lock(priv);
+	err = priv_set_link(priv, 0);
+	priv_unlock(priv);
+	return err;
+}
+
+/**
+ * DPDK callback to bring the link UP.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+int
+mlx5_set_link_up(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	int err;
+
+	priv_lock(priv);
+	err = priv_set_link(priv, 1);
+	priv_unlock(priv);
+	return err;
+}
+
+/**
+ * Configure secondary process queues from a private data pointer (primary
+ * or secondary) and update burst callbacks. Can take place only once.
+ *
+ * All queues must have been previously created by the primary process to
+ * avoid undefined behavior.
+ *
+ * @param priv
+ *   Private data pointer from either primary or secondary process.
+ *
+ * @return
+ *   Private data pointer from secondary process, NULL in case of error.
+ */
+struct priv *
+mlx5_secondary_data_setup(struct priv *priv)
+{
+	unsigned int port_id = 0;
+	struct mlx5_secondary_data *sd;
+	void **tx_queues;
+	void **rx_queues;
+	unsigned int nb_tx_queues;
+	unsigned int nb_rx_queues;
+	unsigned int i;
+
+	/* priv must be valid at this point. */
+	assert(priv != NULL);
+	/* priv->dev must also be valid but may point to local memory from
+	 * another process, possibly with the same address and must not
+	 * be dereferenced yet. */
+	assert(priv->dev != NULL);
+	/* Determine port ID by finding out where priv comes from. */
+	while (1) {
+		sd = &mlx5_secondary_data[port_id];
+		rte_spinlock_lock(&sd->lock);
+		/* Primary process? */
+		if (sd->primary_priv == priv)
+			break;
+		/* Secondary process? */
+		if (sd->data.dev_private == priv)
+			break;
+		rte_spinlock_unlock(&sd->lock);
+		if (++port_id == RTE_DIM(mlx5_secondary_data))
+			port_id = 0;
+	}
+	/* Switch to secondary private structure. If private data has already
+	 * been updated by another thread, there is nothing else to do. */
+	priv = sd->data.dev_private;
+	if (priv->dev->data == &sd->data)
+		goto end;
+	/* Sanity checks. Secondary private structure is supposed to point
+	 * to local eth_dev, itself still pointing to the shared device data
+	 * structure allocated by the primary process. */
+	assert(sd->shared_dev_data != &sd->data);
+	assert(sd->data.nb_tx_queues == 0);
+	assert(sd->data.tx_queues == NULL);
+	assert(sd->data.nb_rx_queues == 0);
+	assert(sd->data.rx_queues == NULL);
+	assert(priv != sd->primary_priv);
+	assert(priv->dev->data == sd->shared_dev_data);
+	assert(priv->txqs_n == 0);
+	assert(priv->txqs == NULL);
+	assert(priv->rxqs_n == 0);
+	assert(priv->rxqs == NULL);
+	nb_tx_queues = sd->shared_dev_data->nb_tx_queues;
+	nb_rx_queues = sd->shared_dev_data->nb_rx_queues;
+	/* Allocate local storage for queues. */
+	tx_queues = rte_zmalloc("secondary ethdev->tx_queues",
+				sizeof(sd->data.tx_queues[0]) * nb_tx_queues,
+				RTE_CACHE_LINE_SIZE);
+	rx_queues = rte_zmalloc("secondary ethdev->rx_queues",
+				sizeof(sd->data.rx_queues[0]) * nb_rx_queues,
+				RTE_CACHE_LINE_SIZE);
+	if (tx_queues == NULL || rx_queues == NULL)
+		goto error;
+	/* Lock to prevent control operations during setup. */
+	priv_lock(priv);
+	/* TX queues. */
+	for (i = 0; i != nb_tx_queues; ++i) {
+		struct txq *primary_txq = (*sd->primary_priv->txqs)[i];
+		struct txq_ctrl *primary_txq_ctrl;
+		struct txq_ctrl *txq_ctrl;
+
+		if (primary_txq == NULL)
+			continue;
+		primary_txq_ctrl = container_of(primary_txq,
+						struct txq_ctrl, txq);
+		txq_ctrl = rte_calloc_socket("TXQ", 1, sizeof(*txq_ctrl), 0,
+					     primary_txq_ctrl->socket);
+		if (txq_ctrl != NULL) {
+			if (txq_ctrl_setup(priv->dev,
+					   primary_txq_ctrl,
+					   primary_txq->elts_n,
+					   primary_txq_ctrl->socket,
+					   NULL) == 0) {
+				txq_ctrl->txq.stats.idx =
+					primary_txq->stats.idx;
+				tx_queues[i] = &txq_ctrl->txq;
+				continue;
+			}
+			rte_free(txq_ctrl);
+		}
+		while (i) {
+			txq_ctrl = tx_queues[--i];
+			txq_cleanup(txq_ctrl);
+			rte_free(txq_ctrl);
+		}
+		goto error;
+	}
+	/* RX queues. */
+	for (i = 0; i != nb_rx_queues; ++i) {
+		struct rxq_ctrl *primary_rxq =
+			container_of((*sd->primary_priv->rxqs)[i],
+				     struct rxq_ctrl, rxq);
+
+		if (primary_rxq == NULL)
+			continue;
+		/* Not supported yet. */
+		rx_queues[i] = NULL;
+	}
+	/* Update everything. */
+	priv->txqs = (void *)tx_queues;
+	priv->txqs_n = nb_tx_queues;
+	priv->rxqs = (void *)rx_queues;
+	priv->rxqs_n = nb_rx_queues;
+	sd->data.rx_queues = rx_queues;
+	sd->data.tx_queues = tx_queues;
+	sd->data.nb_rx_queues = nb_rx_queues;
+	sd->data.nb_tx_queues = nb_tx_queues;
+	sd->data.dev_link = sd->shared_dev_data->dev_link;
+	sd->data.mtu = sd->shared_dev_data->mtu;
+	memcpy(sd->data.rx_queue_state, sd->shared_dev_data->rx_queue_state,
+	       sizeof(sd->data.rx_queue_state));
+	memcpy(sd->data.tx_queue_state, sd->shared_dev_data->tx_queue_state,
+	       sizeof(sd->data.tx_queue_state));
+	sd->data.dev_flags = sd->shared_dev_data->dev_flags;
+	/* Use local data from now on. */
+	rte_mb();
+	priv->dev->data = &sd->data;
+	rte_mb();
+	priv_select_tx_function(priv);
+	priv_select_rx_function(priv);
+	priv_unlock(priv);
+end:
+	/* More sanity checks. */
+	assert(priv->dev->data == &sd->data);
+	rte_spinlock_unlock(&sd->lock);
+	return priv;
+error:
+	priv_unlock(priv);
+	rte_free(tx_queues);
+	rte_free(rx_queues);
+	rte_spinlock_unlock(&sd->lock);
+	return NULL;
+}
+
+/**
+ * Configure the TX function to use.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ */
+void
+priv_select_tx_function(struct priv *priv)
+{
+	priv->dev->tx_pkt_burst = mlx5_tx_burst;
+	/* Display warning for unsupported configurations. */
+	if (priv->sriov && priv->mps)
+		WARN("multi-packet send WQE cannot be used on a SR-IOV setup");
+	/* Select appropriate TX function. */
+	if ((priv->sriov == 0) && priv->mps && priv->txq_inline) {
+		priv->dev->tx_pkt_burst = mlx5_tx_burst_mpw_inline;
+		DEBUG("selected MPW inline TX function");
+	} else if ((priv->sriov == 0) && priv->mps) {
+		priv->dev->tx_pkt_burst = mlx5_tx_burst_mpw;
+		DEBUG("selected MPW TX function");
+	} else if (priv->txq_inline && (priv->txqs_n >= priv->txqs_inline)) {
+		priv->dev->tx_pkt_burst = mlx5_tx_burst_inline;
+		DEBUG("selected inline TX function (%u >= %u queues)",
+		      priv->txqs_n, priv->txqs_inline);
+	}
+}
+
+/**
+ * Configure the RX function to use.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ */
+void
+priv_select_rx_function(struct priv *priv)
+{
+	priv->dev->rx_pkt_burst = mlx5_rx_burst;
+}
diff --git a/src/dpdk/drivers/net/mlx5/mlx5_fdir.c b/src/dpdk/drivers/net/mlx5/mlx5_fdir.c
new file mode 100644
index 00000000..73eb00ec
--- /dev/null
+++ b/src/dpdk/drivers/net/mlx5/mlx5_fdir.c
@@ -0,0 +1,974 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2015 6WIND S.A.
+ *   Copyright 2015 Mellanox.
+ *
+ *   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.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <stddef.h>
+#include <assert.h>
+#include <stdint.h>
+#include <string.h>
+#include <errno.h>
+
+/* Verbs header. */
+/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <infiniband/verbs.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+/* DPDK headers don't like -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <rte_ether.h>
+#include <rte_malloc.h>
+#include <rte_ethdev.h>
+#include <rte_common.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+#include "mlx5.h"
+#include "mlx5_rxtx.h"
+
+struct fdir_flow_desc {
+	uint16_t dst_port;
+	uint16_t src_port;
+	uint32_t src_ip[4];
+	uint32_t dst_ip[4];
+	uint8_t	mac[6];
+	uint16_t vlan_tag;
+	enum hash_rxq_type type;
+};
+
+struct mlx5_fdir_filter {
+	LIST_ENTRY(mlx5_fdir_filter) next;
+	uint16_t queue; /* Queue assigned to if FDIR match. */
+	struct fdir_flow_desc desc;
+	struct ibv_exp_flow *flow;
+};
+
+LIST_HEAD(fdir_filter_list, mlx5_fdir_filter);
+
+/**
+ * Convert struct rte_eth_fdir_filter to mlx5 filter descriptor.
+ *
+ * @param[in] fdir_filter
+ *   DPDK filter structure to convert.
+ * @param[out] desc
+ *   Resulting mlx5 filter descriptor.
+ * @param mode
+ *   Flow director mode.
+ */
+static void
+fdir_filter_to_flow_desc(const struct rte_eth_fdir_filter *fdir_filter,
+			 struct fdir_flow_desc *desc, enum rte_fdir_mode mode)
+{
+	/* Initialize descriptor. */
+	memset(desc, 0, sizeof(*desc));
+
+	/* Set VLAN ID. */
+	desc->vlan_tag = fdir_filter->input.flow_ext.vlan_tci;
+
+	/* Set MAC address. */
+	if (mode == RTE_FDIR_MODE_PERFECT_MAC_VLAN) {
+		rte_memcpy(desc->mac,
+			   fdir_filter->input.flow.mac_vlan_flow.mac_addr.
+				addr_bytes,
+			   sizeof(desc->mac));
+		desc->type = HASH_RXQ_ETH;
+		return;
+	}
+
+	/* Set mode */
+	switch (fdir_filter->input.flow_type) {
+	case RTE_ETH_FLOW_NONFRAG_IPV4_UDP:
+		desc->type = HASH_RXQ_UDPV4;
+		break;
+	case RTE_ETH_FLOW_NONFRAG_IPV4_TCP:
+		desc->type = HASH_RXQ_TCPV4;
+		break;
+	case RTE_ETH_FLOW_NONFRAG_IPV4_OTHER:
+		desc->type = HASH_RXQ_IPV4;
+		break;
+	case RTE_ETH_FLOW_NONFRAG_IPV6_UDP:
+		desc->type = HASH_RXQ_UDPV6;
+		break;
+	case RTE_ETH_FLOW_NONFRAG_IPV6_TCP:
+		desc->type = HASH_RXQ_TCPV6;
+		break;
+	case RTE_ETH_FLOW_NONFRAG_IPV6_OTHER:
+		desc->type = HASH_RXQ_IPV6;
+		break;
+	default:
+		break;
+	}
+
+	/* Set flow values */
+	switch (fdir_filter->input.flow_type) {
+	case RTE_ETH_FLOW_NONFRAG_IPV4_UDP:
+	case RTE_ETH_FLOW_NONFRAG_IPV4_TCP:
+		desc->src_port = fdir_filter->input.flow.udp4_flow.src_port;
+		desc->dst_port = fdir_filter->input.flow.udp4_flow.dst_port;
+	case RTE_ETH_FLOW_NONFRAG_IPV4_OTHER:
+		desc->src_ip[0] = fdir_filter->input.flow.ip4_flow.src_ip;
+		desc->dst_ip[0] = fdir_filter->input.flow.ip4_flow.dst_ip;
+		break;
+	case RTE_ETH_FLOW_NONFRAG_IPV6_UDP:
+	case RTE_ETH_FLOW_NONFRAG_IPV6_TCP:
+		desc->src_port = fdir_filter->input.flow.udp6_flow.src_port;
+		desc->dst_port = fdir_filter->input.flow.udp6_flow.dst_port;
+		/* Fall through. */
+	case RTE_ETH_FLOW_NONFRAG_IPV6_OTHER:
+		rte_memcpy(desc->src_ip,
+			   fdir_filter->input.flow.ipv6_flow.src_ip,
+			   sizeof(desc->src_ip));
+		rte_memcpy(desc->dst_ip,
+			   fdir_filter->input.flow.ipv6_flow.dst_ip,
+			   sizeof(desc->dst_ip));
+		break;
+	default:
+		break;
+	}
+}
+
+/**
+ * Check if two flow descriptors overlap according to configured mask.
+ *
+ * @param priv
+ *   Private structure that provides flow director mask.
+ * @param desc1
+ *   First flow descriptor to compare.
+ * @param desc2
+ *   Second flow descriptor to compare.
+ *
+ * @return
+ *   Nonzero if descriptors overlap.
+ */
+static int
+priv_fdir_overlap(const struct priv *priv,
+		  const struct fdir_flow_desc *desc1,
+		  const struct fdir_flow_desc *desc2)
+{
+	const struct rte_eth_fdir_masks *mask =
+		&priv->dev->data->dev_conf.fdir_conf.mask;
+	unsigned int i;
+
+	if (desc1->type != desc2->type)
+		return 0;
+	/* Ignore non masked bits. */
+	for (i = 0; i != RTE_DIM(desc1->mac); ++i)
+		if ((desc1->mac[i] & mask->mac_addr_byte_mask) !=
+		    (desc2->mac[i] & mask->mac_addr_byte_mask))
+			return 0;
+	if (((desc1->src_port & mask->src_port_mask) !=
+	     (desc2->src_port & mask->src_port_mask)) ||
+	    ((desc1->dst_port & mask->dst_port_mask) !=
+	     (desc2->dst_port & mask->dst_port_mask)))
+		return 0;
+	switch (desc1->type) {
+	case HASH_RXQ_IPV4:
+	case HASH_RXQ_UDPV4:
+	case HASH_RXQ_TCPV4:
+		if (((desc1->src_ip[0] & mask->ipv4_mask.src_ip) !=
+		     (desc2->src_ip[0] & mask->ipv4_mask.src_ip)) ||
+		    ((desc1->dst_ip[0] & mask->ipv4_mask.dst_ip) !=
+		     (desc2->dst_ip[0] & mask->ipv4_mask.dst_ip)))
+			return 0;
+		break;
+	case HASH_RXQ_IPV6:
+	case HASH_RXQ_UDPV6:
+	case HASH_RXQ_TCPV6:
+		for (i = 0; i != RTE_DIM(desc1->src_ip); ++i)
+			if (((desc1->src_ip[i] & mask->ipv6_mask.src_ip[i]) !=
+			     (desc2->src_ip[i] & mask->ipv6_mask.src_ip[i])) ||
+			    ((desc1->dst_ip[i] & mask->ipv6_mask.dst_ip[i]) !=
+			     (desc2->dst_ip[i] & mask->ipv6_mask.dst_ip[i])))
+				return 0;
+		break;
+	default:
+		break;
+	}
+	return 1;
+}
+
+/**
+ * Create flow director steering rule for a specific filter.
+ *
+ * @param priv
+ *   Private structure.
+ * @param mlx5_fdir_filter
+ *   Filter to create a steering rule for.
+ * @param fdir_queue
+ *   Flow director queue for matching packets.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+priv_fdir_flow_add(struct priv *priv,
+		   struct mlx5_fdir_filter *mlx5_fdir_filter,
+		   struct fdir_queue *fdir_queue)
+{
+	struct ibv_exp_flow *flow;
+	struct fdir_flow_desc *desc = &mlx5_fdir_filter->desc;
+	enum rte_fdir_mode fdir_mode =
+		priv->dev->data->dev_conf.fdir_conf.mode;
+	struct rte_eth_fdir_masks *mask =
+		&priv->dev->data->dev_conf.fdir_conf.mask;
+	FLOW_ATTR_SPEC_ETH(data, priv_flow_attr(priv, NULL, 0, desc->type));
+	struct ibv_exp_flow_attr *attr = &data->attr;
+	uintptr_t spec_offset = (uintptr_t)&data->spec;
+	struct ibv_exp_flow_spec_eth *spec_eth;
+	struct ibv_exp_flow_spec_ipv4 *spec_ipv4;
+	struct ibv_exp_flow_spec_ipv6 *spec_ipv6;
+	struct ibv_exp_flow_spec_tcp_udp *spec_tcp_udp;
+	struct mlx5_fdir_filter *iter_fdir_filter;
+	unsigned int i;
+
+	/* Abort if an existing flow overlaps this one to avoid packet
+	 * duplication, even if it targets another queue. */
+	LIST_FOREACH(iter_fdir_filter, priv->fdir_filter_list, next)
+		if ((iter_fdir_filter != mlx5_fdir_filter) &&
+		    (iter_fdir_filter->flow != NULL) &&
+		    (priv_fdir_overlap(priv,
+				       &mlx5_fdir_filter->desc,
+				       &iter_fdir_filter->desc)))
+			return EEXIST;
+
+	/*
+	 * No padding must be inserted by the compiler between attr and spec.
+	 * This layout is expected by libibverbs.
+	 */
+	assert(((uint8_t *)attr + sizeof(*attr)) == (uint8_t *)spec_offset);
+	priv_flow_attr(priv, attr, sizeof(data), desc->type);
+
+	/* Set Ethernet spec */
+	spec_eth = (struct ibv_exp_flow_spec_eth *)spec_offset;
+
+	/* The first specification must be Ethernet. */
+	assert(spec_eth->type == IBV_EXP_FLOW_SPEC_ETH);
+	assert(spec_eth->size == sizeof(*spec_eth));
+
+	/* VLAN ID */
+	spec_eth->val.vlan_tag = desc->vlan_tag & mask->vlan_tci_mask;
+	spec_eth->mask.vlan_tag = mask->vlan_tci_mask;
+
+	/* Update priority */
+	attr->priority = 2;
+
+	if (fdir_mode == RTE_FDIR_MODE_PERFECT_MAC_VLAN) {
+		/* MAC Address */
+		for (i = 0; i != RTE_DIM(spec_eth->mask.dst_mac); ++i) {
+			spec_eth->val.dst_mac[i] =
+				desc->mac[i] & mask->mac_addr_byte_mask;
+			spec_eth->mask.dst_mac[i] = mask->mac_addr_byte_mask;
+		}
+		goto create_flow;
+	}
+
+	switch (desc->type) {
+	case HASH_RXQ_IPV4:
+	case HASH_RXQ_UDPV4:
+	case HASH_RXQ_TCPV4:
+		spec_offset += spec_eth->size;
+
+		/* Set IP spec */
+		spec_ipv4 = (struct ibv_exp_flow_spec_ipv4 *)spec_offset;
+
+		/* The second specification must be IP. */
+		assert(spec_ipv4->type == IBV_EXP_FLOW_SPEC_IPV4);
+		assert(spec_ipv4->size == sizeof(*spec_ipv4));
+
+		spec_ipv4->val.src_ip =
+			desc->src_ip[0] & mask->ipv4_mask.src_ip;
+		spec_ipv4->val.dst_ip =
+			desc->dst_ip[0] & mask->ipv4_mask.dst_ip;
+		spec_ipv4->mask.src_ip = mask->ipv4_mask.src_ip;
+		spec_ipv4->mask.dst_ip = mask->ipv4_mask.dst_ip;
+
+		/* Update priority */
+		attr->priority = 1;
+
+		if (desc->type == HASH_RXQ_IPV4)
+			goto create_flow;
+
+		spec_offset += spec_ipv4->size;
+		break;
+	case HASH_RXQ_IPV6:
+	case HASH_RXQ_UDPV6:
+	case HASH_RXQ_TCPV6:
+		spec_offset += spec_eth->size;
+
+		/* Set IP spec */
+		spec_ipv6 = (struct ibv_exp_flow_spec_ipv6 *)spec_offset;
+
+		/* The second specification must be IP. */
+		assert(spec_ipv6->type == IBV_EXP_FLOW_SPEC_IPV6);
+		assert(spec_ipv6->size == sizeof(*spec_ipv6));
+
+		for (i = 0; i != RTE_DIM(desc->src_ip); ++i) {
+			((uint32_t *)spec_ipv6->val.src_ip)[i] =
+				desc->src_ip[i] & mask->ipv6_mask.src_ip[i];
+			((uint32_t *)spec_ipv6->val.dst_ip)[i] =
+				desc->dst_ip[i] & mask->ipv6_mask.dst_ip[i];
+		}
+		rte_memcpy(spec_ipv6->mask.src_ip,
+			   mask->ipv6_mask.src_ip,
+			   sizeof(spec_ipv6->mask.src_ip));
+		rte_memcpy(spec_ipv6->mask.dst_ip,
+			   mask->ipv6_mask.dst_ip,
+			   sizeof(spec_ipv6->mask.dst_ip));
+
+		/* Update priority */
+		attr->priority = 1;
+
+		if (desc->type == HASH_RXQ_IPV6)
+			goto create_flow;
+
+		spec_offset += spec_ipv6->size;
+		break;
+	default:
+		ERROR("invalid flow attribute type");
+		return EINVAL;
+	}
+
+	/* Set TCP/UDP flow specification. */
+	spec_tcp_udp = (struct ibv_exp_flow_spec_tcp_udp *)spec_offset;
+
+	/* The third specification must be TCP/UDP. */
+	assert(spec_tcp_udp->type == IBV_EXP_FLOW_SPEC_TCP ||
+	       spec_tcp_udp->type == IBV_EXP_FLOW_SPEC_UDP);
+	assert(spec_tcp_udp->size == sizeof(*spec_tcp_udp));
+
+	spec_tcp_udp->val.src_port = desc->src_port & mask->src_port_mask;
+	spec_tcp_udp->val.dst_port = desc->dst_port & mask->dst_port_mask;
+	spec_tcp_udp->mask.src_port = mask->src_port_mask;
+	spec_tcp_udp->mask.dst_port = mask->dst_port_mask;
+
+	/* Update priority */
+	attr->priority = 0;
+
+create_flow:
+
+	errno = 0;
+	flow = ibv_exp_create_flow(fdir_queue->qp, attr);
+	if (flow == NULL) {
+		/* It's not clear whether errno is always set in this case. */
+		ERROR("%p: flow director configuration failed, errno=%d: %s",
+		      (void *)priv, errno,
+		      (errno ? strerror(errno) : "Unknown error"));
+		if (errno)
+			return errno;
+		return EINVAL;
+	}
+
+	DEBUG("%p: added flow director rule (%p)", (void *)priv, (void *)flow);
+	mlx5_fdir_filter->flow = flow;
+	return 0;
+}
+
+/**
+ * Get flow director queue for a specific RX queue, create it in case
+ * it does not exist.
+ *
+ * @param priv
+ *   Private structure.
+ * @param idx
+ *   RX queue index.
+ *
+ * @return
+ *   Related flow director queue on success, NULL otherwise.
+ */
+static struct fdir_queue *
+priv_get_fdir_queue(struct priv *priv, uint16_t idx)
+{
+	struct rxq_ctrl *rxq_ctrl =
+		container_of((*priv->rxqs)[idx], struct rxq_ctrl, rxq);
+	struct fdir_queue *fdir_queue = &rxq_ctrl->fdir_queue;
+	struct ibv_exp_rwq_ind_table *ind_table = NULL;
+	struct ibv_qp *qp = NULL;
+	struct ibv_exp_rwq_ind_table_init_attr ind_init_attr;
+	struct ibv_exp_rx_hash_conf hash_conf;
+	struct ibv_exp_qp_init_attr qp_init_attr;
+	int err = 0;
+
+	/* Return immediately if it has already been created. */
+	if (fdir_queue->qp != NULL)
+		return fdir_queue;
+
+	ind_init_attr = (struct ibv_exp_rwq_ind_table_init_attr){
+		.pd = priv->pd,
+		.log_ind_tbl_size = 0,
+		.ind_tbl = &rxq_ctrl->wq,
+		.comp_mask = 0,
+	};
+
+	errno = 0;
+	ind_table = ibv_exp_create_rwq_ind_table(priv->ctx,
+						 &ind_init_attr);
+	if (ind_table == NULL) {
+		/* Not clear whether errno is set. */
+		err = (errno ? errno : EINVAL);
+		ERROR("RX indirection table creation failed with error %d: %s",
+		      err, strerror(err));
+		goto error;
+	}
+
+	/* Create fdir_queue qp. */
+	hash_conf = (struct ibv_exp_rx_hash_conf){
+		.rx_hash_function = IBV_EXP_RX_HASH_FUNC_TOEPLITZ,
+		.rx_hash_key_len = rss_hash_default_key_len,
+		.rx_hash_key = rss_hash_default_key,
+		.rx_hash_fields_mask = 0,
+		.rwq_ind_tbl = ind_table,
+	};
+	qp_init_attr = (struct ibv_exp_qp_init_attr){
+		.max_inl_recv = 0, /* Currently not supported. */
+		.qp_type = IBV_QPT_RAW_PACKET,
+		.comp_mask = (IBV_EXP_QP_INIT_ATTR_PD |
+			      IBV_EXP_QP_INIT_ATTR_RX_HASH),
+		.pd = priv->pd,
+		.rx_hash_conf = &hash_conf,
+		.port_num = priv->port,
+	};
+
+	qp = ibv_exp_create_qp(priv->ctx, &qp_init_attr);
+	if (qp == NULL) {
+		err = (errno ? errno : EINVAL);
+		ERROR("hash RX QP creation failure: %s", strerror(err));
+		goto error;
+	}
+
+	fdir_queue->ind_table = ind_table;
+	fdir_queue->qp = qp;
+
+	return fdir_queue;
+
+error:
+	if (qp != NULL)
+		claim_zero(ibv_destroy_qp(qp));
+
+	if (ind_table != NULL)
+		claim_zero(ibv_exp_destroy_rwq_ind_table(ind_table));
+
+	return NULL;
+}
+
+/**
+ * Enable flow director filter and create steering rules.
+ *
+ * @param priv
+ *   Private structure.
+ * @param mlx5_fdir_filter
+ *   Filter to create steering rule for.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+priv_fdir_filter_enable(struct priv *priv,
+			struct mlx5_fdir_filter *mlx5_fdir_filter)
+{
+	struct fdir_queue *fdir_queue;
+
+	/* Check if flow already exists. */
+	if (mlx5_fdir_filter->flow != NULL)
+		return 0;
+
+	/* Get fdir_queue for specific queue. */
+	fdir_queue = priv_get_fdir_queue(priv, mlx5_fdir_filter->queue);
+
+	if (fdir_queue == NULL) {
+		ERROR("failed to create flow director rxq for queue %d",
+		      mlx5_fdir_filter->queue);
+		return EINVAL;
+	}
+
+	/* Create flow */
+	return priv_fdir_flow_add(priv, mlx5_fdir_filter, fdir_queue);
+}
+
+/**
+ * Initialize flow director filters list.
+ *
+ * @param priv
+ *   Private structure.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+int
+fdir_init_filters_list(struct priv *priv)
+{
+	/* Filter list initialization should be done only once. */
+	if (priv->fdir_filter_list)
+		return 0;
+
+	/* Create filters list. */
+	priv->fdir_filter_list =
+		rte_calloc(__func__, 1, sizeof(*priv->fdir_filter_list), 0);
+
+	if (priv->fdir_filter_list == NULL) {
+		int err = ENOMEM;
+
+		ERROR("cannot allocate flow director filter list: %s",
+		      strerror(err));
+		return err;
+	}
+
+	LIST_INIT(priv->fdir_filter_list);
+
+	return 0;
+}
+
+/**
+ * Flush all filters.
+ *
+ * @param priv
+ *   Private structure.
+ */
+static void
+priv_fdir_filter_flush(struct priv *priv)
+{
+	struct mlx5_fdir_filter *mlx5_fdir_filter;
+
+	while ((mlx5_fdir_filter = LIST_FIRST(priv->fdir_filter_list))) {
+		struct ibv_exp_flow *flow = mlx5_fdir_filter->flow;
+
+		DEBUG("%p: flushing flow director filter %p",
+		      (void *)priv, (void *)mlx5_fdir_filter);
+		LIST_REMOVE(mlx5_fdir_filter, next);
+		if (flow != NULL)
+			claim_zero(ibv_exp_destroy_flow(flow));
+		rte_free(mlx5_fdir_filter);
+	}
+}
+
+/**
+ * Remove all flow director filters and delete list.
+ *
+ * @param priv
+ *   Private structure.
+ */
+void
+priv_fdir_delete_filters_list(struct priv *priv)
+{
+	priv_fdir_filter_flush(priv);
+	rte_free(priv->fdir_filter_list);
+	priv->fdir_filter_list = NULL;
+}
+
+/**
+ * Disable flow director, remove all steering rules.
+ *
+ * @param priv
+ *   Private structure.
+ */
+void
+priv_fdir_disable(struct priv *priv)
+{
+	unsigned int i;
+	struct mlx5_fdir_filter *mlx5_fdir_filter;
+	struct fdir_queue *fdir_queue;
+
+	/* Run on every flow director filter and destroy flow handle. */
+	LIST_FOREACH(mlx5_fdir_filter, priv->fdir_filter_list, next) {
+		struct ibv_exp_flow *flow;
+
+		/* Only valid elements should be in the list */
+		assert(mlx5_fdir_filter != NULL);
+		flow = mlx5_fdir_filter->flow;
+
+		/* Destroy flow handle */
+		if (flow != NULL) {
+			claim_zero(ibv_exp_destroy_flow(flow));
+			mlx5_fdir_filter->flow = NULL;
+		}
+	}
+
+	/* Run on every RX queue to destroy related flow director QP and
+	 * indirection table. */
+	for (i = 0; (i != priv->rxqs_n); i++) {
+		struct rxq_ctrl *rxq_ctrl =
+			container_of((*priv->rxqs)[i], struct rxq_ctrl, rxq);
+
+		fdir_queue = &rxq_ctrl->fdir_queue;
+		if (fdir_queue->qp != NULL) {
+			claim_zero(ibv_destroy_qp(fdir_queue->qp));
+			fdir_queue->qp = NULL;
+		}
+
+		if (fdir_queue->ind_table != NULL) {
+			claim_zero(ibv_exp_destroy_rwq_ind_table
+				   (fdir_queue->ind_table));
+			fdir_queue->ind_table = NULL;
+		}
+	}
+}
+
+/**
+ * Enable flow director, create steering rules.
+ *
+ * @param priv
+ *   Private structure.
+ */
+void
+priv_fdir_enable(struct priv *priv)
+{
+	struct mlx5_fdir_filter *mlx5_fdir_filter;
+
+	/* Run on every fdir filter and create flow handle */
+	LIST_FOREACH(mlx5_fdir_filter, priv->fdir_filter_list, next) {
+		/* Only valid elements should be in the list */
+		assert(mlx5_fdir_filter != NULL);
+
+		priv_fdir_filter_enable(priv, mlx5_fdir_filter);
+	}
+}
+
+/**
+ * Find specific filter in list.
+ *
+ * @param priv
+ *   Private structure.
+ * @param fdir_filter
+ *   Flow director filter to find.
+ *
+ * @return
+ *   Filter element if found, otherwise NULL.
+ */
+static struct mlx5_fdir_filter *
+priv_find_filter_in_list(struct priv *priv,
+			 const struct rte_eth_fdir_filter *fdir_filter)
+{
+	struct fdir_flow_desc desc;
+	struct mlx5_fdir_filter *mlx5_fdir_filter;
+	enum rte_fdir_mode fdir_mode = priv->dev->data->dev_conf.fdir_conf.mode;
+
+	/* Get flow director filter to look for. */
+	fdir_filter_to_flow_desc(fdir_filter, &desc, fdir_mode);
+
+	/* Look for the requested element. */
+	LIST_FOREACH(mlx5_fdir_filter, priv->fdir_filter_list, next) {
+		/* Only valid elements should be in the list. */
+		assert(mlx5_fdir_filter != NULL);
+
+		/* Return matching filter. */
+		if (!memcmp(&desc, &mlx5_fdir_filter->desc, sizeof(desc)))
+			return mlx5_fdir_filter;
+	}
+
+	/* Filter not found */
+	return NULL;
+}
+
+/**
+ * Add new flow director filter and store it in list.
+ *
+ * @param priv
+ *   Private structure.
+ * @param fdir_filter
+ *   Flow director filter to add.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+priv_fdir_filter_add(struct priv *priv,
+		     const struct rte_eth_fdir_filter *fdir_filter)
+{
+	struct mlx5_fdir_filter *mlx5_fdir_filter;
+	enum rte_fdir_mode fdir_mode = priv->dev->data->dev_conf.fdir_conf.mode;
+	int err = 0;
+
+	/* Validate queue number. */
+	if (fdir_filter->action.rx_queue >= priv->rxqs_n) {
+		ERROR("invalid queue number %d", fdir_filter->action.rx_queue);
+		return EINVAL;
+	}
+
+	/* Duplicate filters are currently unsupported. */
+	mlx5_fdir_filter = priv_find_filter_in_list(priv, fdir_filter);
+	if (mlx5_fdir_filter != NULL) {
+		ERROR("filter already exists");
+		return EINVAL;
+	}
+
+	/* Create new flow director filter. */
+	mlx5_fdir_filter =
+		rte_calloc(__func__, 1, sizeof(*mlx5_fdir_filter), 0);
+	if (mlx5_fdir_filter == NULL) {
+		err = ENOMEM;
+		ERROR("cannot allocate flow director filter: %s",
+		      strerror(err));
+		return err;
+	}
+
+	/* Set queue. */
+	mlx5_fdir_filter->queue = fdir_filter->action.rx_queue;
+
+	/* Convert to mlx5 filter descriptor. */
+	fdir_filter_to_flow_desc(fdir_filter,
+				 &mlx5_fdir_filter->desc, fdir_mode);
+
+	/* Insert new filter into list. */
+	LIST_INSERT_HEAD(priv->fdir_filter_list, mlx5_fdir_filter, next);
+
+	DEBUG("%p: flow director filter %p added",
+	      (void *)priv, (void *)mlx5_fdir_filter);
+
+	/* Enable filter immediately if device is started. */
+	if (priv->started)
+		err = priv_fdir_filter_enable(priv, mlx5_fdir_filter);
+
+	return err;
+}
+
+/**
+ * Update queue for specific filter.
+ *
+ * @param priv
+ *   Private structure.
+ * @param fdir_filter
+ *   Filter to be updated.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+priv_fdir_filter_update(struct priv *priv,
+			const struct rte_eth_fdir_filter *fdir_filter)
+{
+	struct mlx5_fdir_filter *mlx5_fdir_filter;
+
+	/* Validate queue number. */
+	if (fdir_filter->action.rx_queue >= priv->rxqs_n) {
+		ERROR("invalid queue number %d", fdir_filter->action.rx_queue);
+		return EINVAL;
+	}
+
+	mlx5_fdir_filter = priv_find_filter_in_list(priv, fdir_filter);
+	if (mlx5_fdir_filter != NULL) {
+		struct ibv_exp_flow *flow = mlx5_fdir_filter->flow;
+		int err = 0;
+
+		/* Update queue number. */
+		mlx5_fdir_filter->queue = fdir_filter->action.rx_queue;
+
+		/* Destroy flow handle. */
+		if (flow != NULL) {
+			claim_zero(ibv_exp_destroy_flow(flow));
+			mlx5_fdir_filter->flow = NULL;
+		}
+		DEBUG("%p: flow director filter %p updated",
+		      (void *)priv, (void *)mlx5_fdir_filter);
+
+		/* Enable filter if device is started. */
+		if (priv->started)
+			err = priv_fdir_filter_enable(priv, mlx5_fdir_filter);
+
+		return err;
+	}
+
+	/* Filter not found, create it. */
+	DEBUG("%p: filter not found for update, creating new filter",
+	      (void *)priv);
+	return priv_fdir_filter_add(priv, fdir_filter);
+}
+
+/**
+ * Delete specific filter.
+ *
+ * @param priv
+ *   Private structure.
+ * @param fdir_filter
+ *   Filter to be deleted.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+priv_fdir_filter_delete(struct priv *priv,
+			const struct rte_eth_fdir_filter *fdir_filter)
+{
+	struct mlx5_fdir_filter *mlx5_fdir_filter;
+
+	mlx5_fdir_filter = priv_find_filter_in_list(priv, fdir_filter);
+	if (mlx5_fdir_filter != NULL) {
+		struct ibv_exp_flow *flow = mlx5_fdir_filter->flow;
+
+		/* Remove element from list. */
+		LIST_REMOVE(mlx5_fdir_filter, next);
+
+		/* Destroy flow handle. */
+		if (flow != NULL) {
+			claim_zero(ibv_exp_destroy_flow(flow));
+			mlx5_fdir_filter->flow = NULL;
+		}
+
+		DEBUG("%p: flow director filter %p deleted",
+		      (void *)priv, (void *)mlx5_fdir_filter);
+
+		/* Delete filter. */
+		rte_free(mlx5_fdir_filter);
+
+		return 0;
+	}
+
+	ERROR("%p: flow director delete failed, cannot find filter",
+	      (void *)priv);
+	return EINVAL;
+}
+
+/**
+ * Get flow director information.
+ *
+ * @param priv
+ *   Private structure.
+ * @param[out] fdir_info
+ *   Resulting flow director information.
+ */
+static void
+priv_fdir_info_get(struct priv *priv, struct rte_eth_fdir_info *fdir_info)
+{
+	struct rte_eth_fdir_masks *mask =
+		&priv->dev->data->dev_conf.fdir_conf.mask;
+
+	fdir_info->mode = priv->dev->data->dev_conf.fdir_conf.mode;
+	fdir_info->guarant_spc = 0;
+
+	rte_memcpy(&fdir_info->mask, mask, sizeof(fdir_info->mask));
+
+	fdir_info->max_flexpayload = 0;
+	fdir_info->flow_types_mask[0] = 0;
+
+	fdir_info->flex_payload_unit = 0;
+	fdir_info->max_flex_payload_segment_num = 0;
+	fdir_info->flex_payload_limit = 0;
+	memset(&fdir_info->flex_conf, 0, sizeof(fdir_info->flex_conf));
+}
+
+/**
+ * Deal with flow director operations.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param filter_op
+ *   Operation to perform.
+ * @param arg
+ *   Pointer to operation-specific structure.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+priv_fdir_ctrl_func(struct priv *priv, enum rte_filter_op filter_op, void *arg)
+{
+	enum rte_fdir_mode fdir_mode =
+		priv->dev->data->dev_conf.fdir_conf.mode;
+	int ret = 0;
+
+	if (filter_op == RTE_ETH_FILTER_NOP)
+		return 0;
+
+	if (fdir_mode != RTE_FDIR_MODE_PERFECT &&
+	    fdir_mode != RTE_FDIR_MODE_PERFECT_MAC_VLAN) {
+		ERROR("%p: flow director mode %d not supported",
+		      (void *)priv, fdir_mode);
+		return EINVAL;
+	}
+
+	switch (filter_op) {
+	case RTE_ETH_FILTER_ADD:
+		ret = priv_fdir_filter_add(priv, arg);
+		break;
+	case RTE_ETH_FILTER_UPDATE:
+		ret = priv_fdir_filter_update(priv, arg);
+		break;
+	case RTE_ETH_FILTER_DELETE:
+		ret = priv_fdir_filter_delete(priv, arg);
+		break;
+	case RTE_ETH_FILTER_FLUSH:
+		priv_fdir_filter_flush(priv);
+		break;
+	case RTE_ETH_FILTER_INFO:
+		priv_fdir_info_get(priv, arg);
+		break;
+	default:
+		DEBUG("%p: unknown operation %u", (void *)priv, filter_op);
+		ret = EINVAL;
+		break;
+	}
+	return ret;
+}
+
+/**
+ * Manage filter operations.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param filter_type
+ *   Filter type.
+ * @param filter_op
+ *   Operation to perform.
+ * @param arg
+ *   Pointer to operation-specific structure.
+ *
+ * @return
+ *   0 on success, negative errno value on failure.
+ */
+int
+mlx5_dev_filter_ctrl(struct rte_eth_dev *dev,
+		     enum rte_filter_type filter_type,
+		     enum rte_filter_op filter_op,
+		     void *arg)
+{
+	int ret = -EINVAL;
+	struct priv *priv = dev->data->dev_private;
+
+	switch (filter_type) {
+	case RTE_ETH_FILTER_FDIR:
+		priv_lock(priv);
+		ret = priv_fdir_ctrl_func(priv, filter_op, arg);
+		priv_unlock(priv);
+		break;
+	default:
+		ERROR("%p: filter type (%d) not supported",
+		      (void *)dev, filter_type);
+		break;
+	}
+
+	return ret;
+}
diff --git a/src/dpdk/drivers/net/mlx5/mlx5_mac.c b/src/dpdk/drivers/net/mlx5/mlx5_mac.c
new file mode 100644
index 00000000..f6b27bb8
--- /dev/null
+++ b/src/dpdk/drivers/net/mlx5/mlx5_mac.c
@@ -0,0 +1,509 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2015 6WIND S.A.
+ *   Copyright 2015 Mellanox.
+ *
+ *   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.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <stddef.h>
+#include <assert.h>
+#include <stdint.h>
+#include <string.h>
+#include <inttypes.h>
+#include <errno.h>
+#include <netinet/in.h>
+#include <sys/ioctl.h>
+#include <arpa/inet.h>
+
+/* Verbs header. */
+/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <infiniband/verbs.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+/* DPDK headers don't like -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <rte_ether.h>
+#include <rte_ethdev.h>
+#include <rte_common.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+#include "mlx5.h"
+#include "mlx5_utils.h"
+#include "mlx5_rxtx.h"
+#include "mlx5_defs.h"
+
+/**
+ * Get MAC address by querying netdevice.
+ *
+ * @param[in] priv
+ *   struct priv for the requested device.
+ * @param[out] mac
+ *   MAC address output buffer.
+ *
+ * @return
+ *   0 on success, -1 on failure and errno is set.
+ */
+int
+priv_get_mac(struct priv *priv, uint8_t (*mac)[ETHER_ADDR_LEN])
+{
+	struct ifreq request;
+
+	if (priv_ifreq(priv, SIOCGIFHWADDR, &request))
+		return -1;
+	memcpy(mac, request.ifr_hwaddr.sa_data, ETHER_ADDR_LEN);
+	return 0;
+}
+
+/**
+ * Delete MAC flow steering rule.
+ *
+ * @param hash_rxq
+ *   Pointer to hash RX queue structure.
+ * @param mac_index
+ *   MAC address index.
+ * @param vlan_index
+ *   VLAN index to use.
+ */
+static void
+hash_rxq_del_mac_flow(struct hash_rxq *hash_rxq, unsigned int mac_index,
+		      unsigned int vlan_index)
+{
+#ifndef NDEBUG
+	const uint8_t (*mac)[ETHER_ADDR_LEN] =
+		(const uint8_t (*)[ETHER_ADDR_LEN])
+		hash_rxq->priv->mac[mac_index].addr_bytes;
+#endif
+
+	assert(mac_index < RTE_DIM(hash_rxq->mac_flow));
+	assert(vlan_index < RTE_DIM(hash_rxq->mac_flow[mac_index]));
+	if (hash_rxq->mac_flow[mac_index][vlan_index] == NULL)
+		return;
+	DEBUG("%p: removing MAC address %02x:%02x:%02x:%02x:%02x:%02x index %u"
+	      " VLAN index %u",
+	      (void *)hash_rxq,
+	      (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5],
+	      mac_index,
+	      vlan_index);
+	claim_zero(ibv_exp_destroy_flow(hash_rxq->mac_flow
+					[mac_index][vlan_index]));
+	hash_rxq->mac_flow[mac_index][vlan_index] = NULL;
+}
+
+/**
+ * Unregister a MAC address from a hash RX queue.
+ *
+ * @param hash_rxq
+ *   Pointer to hash RX queue structure.
+ * @param mac_index
+ *   MAC address index.
+ */
+static void
+hash_rxq_mac_addr_del(struct hash_rxq *hash_rxq, unsigned int mac_index)
+{
+	unsigned int i;
+
+	assert(mac_index < RTE_DIM(hash_rxq->mac_flow));
+	for (i = 0; (i != RTE_DIM(hash_rxq->mac_flow[mac_index])); ++i)
+		hash_rxq_del_mac_flow(hash_rxq, mac_index, i);
+}
+
+/**
+ * Unregister all MAC addresses from a hash RX queue.
+ *
+ * @param hash_rxq
+ *   Pointer to hash RX queue structure.
+ */
+void
+hash_rxq_mac_addrs_del(struct hash_rxq *hash_rxq)
+{
+	unsigned int i;
+
+	for (i = 0; (i != RTE_DIM(hash_rxq->mac_flow)); ++i)
+		hash_rxq_mac_addr_del(hash_rxq, i);
+}
+
+/**
+ * Unregister a MAC address.
+ *
+ * This is done for each hash RX queue.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param mac_index
+ *   MAC address index.
+ */
+static void
+priv_mac_addr_del(struct priv *priv, unsigned int mac_index)
+{
+	unsigned int i;
+
+	assert(mac_index < RTE_DIM(priv->mac));
+	if (!BITFIELD_ISSET(priv->mac_configured, mac_index))
+		return;
+	for (i = 0; (i != priv->hash_rxqs_n); ++i)
+		hash_rxq_mac_addr_del(&(*priv->hash_rxqs)[i], mac_index);
+	BITFIELD_RESET(priv->mac_configured, mac_index);
+}
+
+/**
+ * Unregister all MAC addresses from all hash RX queues.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ */
+void
+priv_mac_addrs_disable(struct priv *priv)
+{
+	unsigned int i;
+
+	for (i = 0; (i != priv->hash_rxqs_n); ++i)
+		hash_rxq_mac_addrs_del(&(*priv->hash_rxqs)[i]);
+}
+
+/**
+ * DPDK callback to remove a MAC address.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param index
+ *   MAC address index.
+ */
+void
+mlx5_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
+{
+	struct priv *priv = dev->data->dev_private;
+
+	if (mlx5_is_secondary())
+		return;
+
+	priv_lock(priv);
+	DEBUG("%p: removing MAC address from index %" PRIu32,
+	      (void *)dev, index);
+	if (index >= RTE_DIM(priv->mac))
+		goto end;
+	priv_mac_addr_del(priv, index);
+end:
+	priv_unlock(priv);
+}
+
+/**
+ * Add MAC flow steering rule.
+ *
+ * @param hash_rxq
+ *   Pointer to hash RX queue structure.
+ * @param mac_index
+ *   MAC address index to register.
+ * @param vlan_index
+ *   VLAN index to use.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+hash_rxq_add_mac_flow(struct hash_rxq *hash_rxq, unsigned int mac_index,
+		      unsigned int vlan_index)
+{
+	struct ibv_exp_flow *flow;
+	struct priv *priv = hash_rxq->priv;
+	const uint8_t (*mac)[ETHER_ADDR_LEN] =
+			(const uint8_t (*)[ETHER_ADDR_LEN])
+			priv->mac[mac_index].addr_bytes;
+	FLOW_ATTR_SPEC_ETH(data, priv_flow_attr(priv, NULL, 0, hash_rxq->type));
+	struct ibv_exp_flow_attr *attr = &data->attr;
+	struct ibv_exp_flow_spec_eth *spec = &data->spec;
+	unsigned int vlan_enabled = !!priv->vlan_filter_n;
+	unsigned int vlan_id = priv->vlan_filter[vlan_index];
+
+	assert(mac_index < RTE_DIM(hash_rxq->mac_flow));
+	assert(vlan_index < RTE_DIM(hash_rxq->mac_flow[mac_index]));
+	if (hash_rxq->mac_flow[mac_index][vlan_index] != NULL)
+		return 0;
+	/*
+	 * No padding must be inserted by the compiler between attr and spec.
+	 * This layout is expected by libibverbs.
+	 */
+	assert(((uint8_t *)attr + sizeof(*attr)) == (uint8_t *)spec);
+	priv_flow_attr(priv, attr, sizeof(data), hash_rxq->type);
+	/* The first specification must be Ethernet. */
+	assert(spec->type == IBV_EXP_FLOW_SPEC_ETH);
+	assert(spec->size == sizeof(*spec));
+	*spec = (struct ibv_exp_flow_spec_eth){
+		.type = IBV_EXP_FLOW_SPEC_ETH,
+		.size = sizeof(*spec),
+		.val = {
+			.dst_mac = {
+				(*mac)[0], (*mac)[1], (*mac)[2],
+				(*mac)[3], (*mac)[4], (*mac)[5]
+			},
+			.vlan_tag = (vlan_enabled ? htons(vlan_id) : 0),
+		},
+		.mask = {
+			.dst_mac = "\xff\xff\xff\xff\xff\xff",
+			.vlan_tag = (vlan_enabled ? htons(0xfff) : 0),
+		},
+	};
+	DEBUG("%p: adding MAC address %02x:%02x:%02x:%02x:%02x:%02x index %u"
+	      " VLAN index %u filtering %s, ID %u",
+	      (void *)hash_rxq,
+	      (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5],
+	      mac_index,
+	      vlan_index,
+	      (vlan_enabled ? "enabled" : "disabled"),
+	      vlan_id);
+	/* Create related flow. */
+	errno = 0;
+	flow = ibv_exp_create_flow(hash_rxq->qp, attr);
+	if (flow == NULL) {
+		/* It's not clear whether errno is always set in this case. */
+		ERROR("%p: flow configuration failed, errno=%d: %s",
+		      (void *)hash_rxq, errno,
+		      (errno ? strerror(errno) : "Unknown error"));
+		if (errno)
+			return errno;
+		return EINVAL;
+	}
+	hash_rxq->mac_flow[mac_index][vlan_index] = flow;
+	return 0;
+}
+
+/**
+ * Register a MAC address in a hash RX queue.
+ *
+ * @param hash_rxq
+ *   Pointer to hash RX queue structure.
+ * @param mac_index
+ *   MAC address index to register.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+hash_rxq_mac_addr_add(struct hash_rxq *hash_rxq, unsigned int mac_index)
+{
+	struct priv *priv = hash_rxq->priv;
+	unsigned int i = 0;
+	int ret;
+
+	assert(mac_index < RTE_DIM(hash_rxq->mac_flow));
+	assert(RTE_DIM(hash_rxq->mac_flow[mac_index]) ==
+	       RTE_DIM(priv->vlan_filter));
+	/* Add a MAC address for each VLAN filter, or at least once. */
+	do {
+		ret = hash_rxq_add_mac_flow(hash_rxq, mac_index, i);
+		if (ret) {
+			/* Failure, rollback. */
+			while (i != 0)
+				hash_rxq_del_mac_flow(hash_rxq, mac_index,
+						      --i);
+			return ret;
+		}
+	} while (++i < priv->vlan_filter_n);
+	return 0;
+}
+
+/**
+ * Register all MAC addresses in a hash RX queue.
+ *
+ * @param hash_rxq
+ *   Pointer to hash RX queue structure.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+int
+hash_rxq_mac_addrs_add(struct hash_rxq *hash_rxq)
+{
+	struct priv *priv = hash_rxq->priv;
+	unsigned int i;
+	int ret;
+
+	assert(RTE_DIM(priv->mac) == RTE_DIM(hash_rxq->mac_flow));
+	for (i = 0; (i != RTE_DIM(priv->mac)); ++i) {
+		if (!BITFIELD_ISSET(priv->mac_configured, i))
+			continue;
+		ret = hash_rxq_mac_addr_add(hash_rxq, i);
+		if (!ret)
+			continue;
+		/* Failure, rollback. */
+		while (i != 0)
+			hash_rxq_mac_addr_del(hash_rxq, --i);
+		assert(ret > 0);
+		return ret;
+	}
+	return 0;
+}
+
+/**
+ * Register a MAC address.
+ *
+ * This is done for each hash RX queue.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param mac_index
+ *   MAC address index to use.
+ * @param mac
+ *   MAC address to register.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+int
+priv_mac_addr_add(struct priv *priv, unsigned int mac_index,
+		  const uint8_t (*mac)[ETHER_ADDR_LEN])
+{
+	unsigned int i;
+	int ret;
+
+	assert(mac_index < RTE_DIM(priv->mac));
+	/* First, make sure this address isn't already configured. */
+	for (i = 0; (i != RTE_DIM(priv->mac)); ++i) {
+		/* Skip this index, it's going to be reconfigured. */
+		if (i == mac_index)
+			continue;
+		if (!BITFIELD_ISSET(priv->mac_configured, i))
+			continue;
+		if (memcmp(priv->mac[i].addr_bytes, *mac, sizeof(*mac)))
+			continue;
+		/* Address already configured elsewhere, return with error. */
+		return EADDRINUSE;
+	}
+	if (BITFIELD_ISSET(priv->mac_configured, mac_index))
+		priv_mac_addr_del(priv, mac_index);
+	priv->mac[mac_index] = (struct ether_addr){
+		{
+			(*mac)[0], (*mac)[1], (*mac)[2],
+			(*mac)[3], (*mac)[4], (*mac)[5]
+		}
+	};
+	if (!priv_allow_flow_type(priv, HASH_RXQ_FLOW_TYPE_MAC))
+		goto end;
+	for (i = 0; (i != priv->hash_rxqs_n); ++i) {
+		ret = hash_rxq_mac_addr_add(&(*priv->hash_rxqs)[i], mac_index);
+		if (!ret)
+			continue;
+		/* Failure, rollback. */
+		while (i != 0)
+			hash_rxq_mac_addr_del(&(*priv->hash_rxqs)[--i],
+					      mac_index);
+		return ret;
+	}
+end:
+	BITFIELD_SET(priv->mac_configured, mac_index);
+	return 0;
+}
+
+/**
+ * Register all MAC addresses in all hash RX queues.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+int
+priv_mac_addrs_enable(struct priv *priv)
+{
+	unsigned int i;
+	int ret;
+
+	if (!priv_allow_flow_type(priv, HASH_RXQ_FLOW_TYPE_MAC))
+		return 0;
+	for (i = 0; (i != priv->hash_rxqs_n); ++i) {
+		ret = hash_rxq_mac_addrs_add(&(*priv->hash_rxqs)[i]);
+		if (!ret)
+			continue;
+		/* Failure, rollback. */
+		while (i != 0)
+			hash_rxq_mac_addrs_del(&(*priv->hash_rxqs)[--i]);
+		assert(ret > 0);
+		return ret;
+	}
+	return 0;
+}
+
+/**
+ * DPDK callback to add a MAC address.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param mac_addr
+ *   MAC address to register.
+ * @param index
+ *   MAC address index.
+ * @param vmdq
+ *   VMDq pool index to associate address with (ignored).
+ */
+void
+mlx5_mac_addr_add(struct rte_eth_dev *dev, struct ether_addr *mac_addr,
+		  uint32_t index, uint32_t vmdq)
+{
+	struct priv *priv = dev->data->dev_private;
+
+	if (mlx5_is_secondary())
+		return;
+
+	(void)vmdq;
+	priv_lock(priv);
+	DEBUG("%p: adding MAC address at index %" PRIu32,
+	      (void *)dev, index);
+	if (index >= RTE_DIM(priv->mac))
+		goto end;
+	priv_mac_addr_add(priv, index,
+			  (const uint8_t (*)[ETHER_ADDR_LEN])
+			  mac_addr->addr_bytes);
+end:
+	priv_unlock(priv);
+}
+
+/**
+ * DPDK callback to set primary MAC address.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param mac_addr
+ *   MAC address to register.
+ */
+void
+mlx5_mac_addr_set(struct rte_eth_dev *dev, struct ether_addr *mac_addr)
+{
+	DEBUG("%p: setting primary MAC address", (void *)dev);
+	mlx5_mac_addr_remove(dev, 0);
+	mlx5_mac_addr_add(dev, mac_addr, 0, 0);
+}
diff --git a/src/dpdk/drivers/net/mlx5/mlx5_mr.c b/src/dpdk/drivers/net/mlx5/mlx5_mr.c
new file mode 100644
index 00000000..67dfefa8
--- /dev/null
+++ b/src/dpdk/drivers/net/mlx5/mlx5_mr.c
@@ -0,0 +1,283 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2016 6WIND S.A.
+ *   Copyright 2016 Mellanox.
+ *
+ *   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.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+/* Verbs header. */
+/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <infiniband/verbs.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+/* DPDK headers don't like -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <rte_mempool.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+#include "mlx5.h"
+#include "mlx5_rxtx.h"
+
+struct mlx5_check_mempool_data {
+	int ret;
+	char *start;
+	char *end;
+};
+
+/* Called by mlx5_check_mempool() when iterating the memory chunks. */
+static void
+mlx5_check_mempool_cb(struct rte_mempool *mp,
+		      void *opaque, struct rte_mempool_memhdr *memhdr,
+		      unsigned int mem_idx)
+{
+	struct mlx5_check_mempool_data *data = opaque;
+
+	(void)mp;
+	(void)mem_idx;
+
+	/* It already failed, skip the next chunks. */
+	if (data->ret != 0)
+		return;
+	/* It is the first chunk. */
+	if (data->start == NULL && data->end == NULL) {
+		data->start = memhdr->addr;
+		data->end = data->start + memhdr->len;
+		return;
+	}
+	if (data->end == memhdr->addr) {
+		data->end += memhdr->len;
+		return;
+	}
+	if (data->start == (char *)memhdr->addr + memhdr->len) {
+		data->start -= memhdr->len;
+		return;
+	}
+	/* Error, mempool is not virtually contiguous. */
+	data->ret = -1;
+}
+
+/**
+ * Check if a mempool can be used: it must be virtually contiguous.
+ *
+ * @param[in] mp
+ *   Pointer to memory pool.
+ * @param[out] start
+ *   Pointer to the start address of the mempool virtual memory area
+ * @param[out] end
+ *   Pointer to the end address of the mempool virtual memory area
+ *
+ * @return
+ *   0 on success (mempool is virtually contiguous), -1 on error.
+ */
+static int mlx5_check_mempool(struct rte_mempool *mp, uintptr_t *start,
+	uintptr_t *end)
+{
+	struct mlx5_check_mempool_data data;
+
+	memset(&data, 0, sizeof(data));
+	rte_mempool_mem_iter(mp, mlx5_check_mempool_cb, &data);
+	*start = (uintptr_t)data.start;
+	*end = (uintptr_t)data.end;
+
+	return data.ret;
+}
+
+/**
+ * Register mempool as a memory region.
+ *
+ * @param pd
+ *   Pointer to protection domain.
+ * @param mp
+ *   Pointer to memory pool.
+ *
+ * @return
+ *   Memory region pointer, NULL in case of error.
+ */
+struct ibv_mr *
+mlx5_mp2mr(struct ibv_pd *pd, struct rte_mempool *mp)
+{
+	const struct rte_memseg *ms = rte_eal_get_physmem_layout();
+	uintptr_t start;
+	uintptr_t end;
+	unsigned int i;
+
+	if (mlx5_check_mempool(mp, &start, &end) != 0) {
+		ERROR("mempool %p: not virtually contiguous",
+		      (void *)mp);
+		return NULL;
+	}
+
+	DEBUG("mempool %p area start=%p end=%p size=%zu",
+	      (void *)mp, (void *)start, (void *)end,
+	      (size_t)(end - start));
+	/* Round start and end to page boundary if found in memory segments. */
+	for (i = 0; (i < RTE_MAX_MEMSEG) && (ms[i].addr != NULL); ++i) {
+		uintptr_t addr = (uintptr_t)ms[i].addr;
+		size_t len = ms[i].len;
+		unsigned int align = ms[i].hugepage_sz;
+
+		if ((start > addr) && (start < addr + len))
+			start = RTE_ALIGN_FLOOR(start, align);
+		if ((end > addr) && (end < addr + len))
+			end = RTE_ALIGN_CEIL(end, align);
+	}
+	DEBUG("mempool %p using start=%p end=%p size=%zu for MR",
+	      (void *)mp, (void *)start, (void *)end,
+	      (size_t)(end - start));
+	return ibv_reg_mr(pd,
+			  (void *)start,
+			  end - start,
+			  IBV_ACCESS_LOCAL_WRITE);
+}
+
+/**
+ * Register a Memory Region (MR) <-> Memory Pool (MP) association in
+ * txq->mp2mr[]. If mp2mr[] is full, remove an entry first.
+ *
+ * This function should only be called by txq_mp2mr().
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param[in] mp
+ *   Memory Pool for which a Memory Region lkey must be returned.
+ * @param idx
+ *   Index of the next available entry.
+ *
+ * @return
+ *   mr->lkey on success, (uint32_t)-1 on failure.
+ */
+uint32_t
+txq_mp2mr_reg(struct txq *txq, struct rte_mempool *mp, unsigned int idx)
+{
+	struct txq_ctrl *txq_ctrl = container_of(txq, struct txq_ctrl, txq);
+	struct ibv_mr *mr;
+
+	/* Add a new entry, register MR first. */
+	DEBUG("%p: discovered new memory pool \"%s\" (%p)",
+	      (void *)txq_ctrl, mp->name, (void *)mp);
+	mr = mlx5_mp2mr(txq_ctrl->priv->pd, mp);
+	if (unlikely(mr == NULL)) {
+		DEBUG("%p: unable to configure MR, ibv_reg_mr() failed.",
+		      (void *)txq_ctrl);
+		return (uint32_t)-1;
+	}
+	if (unlikely(idx == RTE_DIM(txq_ctrl->txq.mp2mr))) {
+		/* Table is full, remove oldest entry. */
+		DEBUG("%p: MR <-> MP table full, dropping oldest entry.",
+		      (void *)txq_ctrl);
+		--idx;
+		claim_zero(ibv_dereg_mr(txq_ctrl->txq.mp2mr[0].mr));
+		memmove(&txq_ctrl->txq.mp2mr[0], &txq_ctrl->txq.mp2mr[1],
+			(sizeof(txq_ctrl->txq.mp2mr) -
+			 sizeof(txq_ctrl->txq.mp2mr[0])));
+	}
+	/* Store the new entry. */
+	txq_ctrl->txq.mp2mr[idx].mp = mp;
+	txq_ctrl->txq.mp2mr[idx].mr = mr;
+	txq_ctrl->txq.mp2mr[idx].lkey = htonl(mr->lkey);
+	DEBUG("%p: new MR lkey for MP \"%s\" (%p): 0x%08" PRIu32,
+	      (void *)txq_ctrl, mp->name, (void *)mp,
+	      txq_ctrl->txq.mp2mr[idx].lkey);
+	return txq_ctrl->txq.mp2mr[idx].lkey;
+}
+
+struct txq_mp2mr_mbuf_check_data {
+	int ret;
+};
+
+/**
+ * Callback function for rte_mempool_obj_iter() to check whether a given
+ * mempool object looks like a mbuf.
+ *
+ * @param[in] mp
+ *   The mempool pointer
+ * @param[in] arg
+ *   Context data (struct txq_mp2mr_mbuf_check_data). Contains the
+ *   return value.
+ * @param[in] obj
+ *   Object address.
+ * @param index
+ *   Object index, unused.
+ */
+static void
+txq_mp2mr_mbuf_check(struct rte_mempool *mp, void *arg, void *obj,
+	uint32_t index __rte_unused)
+{
+	struct txq_mp2mr_mbuf_check_data *data = arg;
+	struct rte_mbuf *buf = obj;
+
+	/*
+	 * Check whether mbuf structure fits element size and whether mempool
+	 * pointer is valid.
+	 */
+	if (sizeof(*buf) > mp->elt_size || buf->pool != mp)
+		data->ret = -1;
+}
+
+/**
+ * Iterator function for rte_mempool_walk() to register existing mempools and
+ * fill the MP to MR cache of a TX queue.
+ *
+ * @param[in] mp
+ *   Memory Pool to register.
+ * @param *arg
+ *   Pointer to TX queue structure.
+ */
+void
+txq_mp2mr_iter(struct rte_mempool *mp, void *arg)
+{
+	struct txq_ctrl *txq_ctrl = arg;
+	struct txq_mp2mr_mbuf_check_data data = {
+		.ret = 0,
+	};
+	unsigned int i;
+
+	/* Register mempool only if the first element looks like a mbuf. */
+	if (rte_mempool_obj_iter(mp, txq_mp2mr_mbuf_check, &data) == 0 ||
+			data.ret == -1)
+		return;
+	for (i = 0; (i != RTE_DIM(txq_ctrl->txq.mp2mr)); ++i) {
+		if (unlikely(txq_ctrl->txq.mp2mr[i].mp == NULL)) {
+			/* Unknown MP, add a new MR for it. */
+			break;
+		}
+		if (txq_ctrl->txq.mp2mr[i].mp == mp)
+			return;
+	}
+	txq_mp2mr_reg(&txq_ctrl->txq, mp, i);
+}
diff --git a/src/dpdk/drivers/net/mlx5/mlx5_prm.h b/src/dpdk/drivers/net/mlx5/mlx5_prm.h
new file mode 100644
index 00000000..5db219b3
--- /dev/null
+++ b/src/dpdk/drivers/net/mlx5/mlx5_prm.h
@@ -0,0 +1,163 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2016 6WIND S.A.
+ *   Copyright 2016 Mellanox.
+ *
+ *   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.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#ifndef RTE_PMD_MLX5_PRM_H_
+#define RTE_PMD_MLX5_PRM_H_
+
+/* Verbs header. */
+/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <infiniband/mlx5_hw.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+/* Get CQE owner bit. */
+#define MLX5_CQE_OWNER(op_own) ((op_own) & MLX5_CQE_OWNER_MASK)
+
+/* Get CQE format. */
+#define MLX5_CQE_FORMAT(op_own) (((op_own) & MLX5E_CQE_FORMAT_MASK) >> 2)
+
+/* Get CQE opcode. */
+#define MLX5_CQE_OPCODE(op_own) (((op_own) & 0xf0) >> 4)
+
+/* Get CQE solicited event. */
+#define MLX5_CQE_SE(op_own) (((op_own) >> 1) & 1)
+
+/* Invalidate a CQE. */
+#define MLX5_CQE_INVALIDATE (MLX5_CQE_INVALID << 4)
+
+/* CQE value to inform that VLAN is stripped. */
+#define MLX5_CQE_VLAN_STRIPPED 0x1
+
+/* Maximum number of packets a multi-packet WQE can handle. */
+#define MLX5_MPW_DSEG_MAX 5
+
+/* Room for inline data in regular work queue element. */
+#define MLX5_WQE64_INL_DATA 12
+
+/* Room for inline data in multi-packet WQE. */
+#define MLX5_MWQE64_INL_DATA 28
+
+/* Subset of struct mlx5_wqe_eth_seg. */
+struct mlx5_wqe_eth_seg_small {
+	uint32_t rsvd0;
+	uint8_t	cs_flags;
+	uint8_t	rsvd1;
+	uint16_t mss;
+	uint32_t rsvd2;
+	uint16_t inline_hdr_sz;
+};
+
+/* Regular WQE. */
+struct mlx5_wqe_regular {
+	union {
+		struct mlx5_wqe_ctrl_seg ctrl;
+		uint32_t data[4];
+	} ctrl;
+	struct mlx5_wqe_eth_seg eseg;
+	struct mlx5_wqe_data_seg dseg;
+} __rte_aligned(64);
+
+/* Inline WQE. */
+struct mlx5_wqe_inl {
+	union {
+		struct mlx5_wqe_ctrl_seg ctrl;
+		uint32_t data[4];
+	} ctrl;
+	struct mlx5_wqe_eth_seg eseg;
+	uint32_t byte_cnt;
+	uint8_t data[MLX5_WQE64_INL_DATA];
+} __rte_aligned(64);
+
+/* Multi-packet WQE. */
+struct mlx5_wqe_mpw {
+	union {
+		struct mlx5_wqe_ctrl_seg ctrl;
+		uint32_t data[4];
+	} ctrl;
+	struct mlx5_wqe_eth_seg_small eseg;
+	struct mlx5_wqe_data_seg dseg[2];
+} __rte_aligned(64);
+
+/* Multi-packet WQE with inline. */
+struct mlx5_wqe_mpw_inl {
+	union {
+		struct mlx5_wqe_ctrl_seg ctrl;
+		uint32_t data[4];
+	} ctrl;
+	struct mlx5_wqe_eth_seg_small eseg;
+	uint32_t byte_cnt;
+	uint8_t data[MLX5_MWQE64_INL_DATA];
+} __rte_aligned(64);
+
+/* Union of all WQE types. */
+union mlx5_wqe {
+	struct mlx5_wqe_regular wqe;
+	struct mlx5_wqe_inl inl;
+	struct mlx5_wqe_mpw mpw;
+	struct mlx5_wqe_mpw_inl mpw_inl;
+	uint8_t data[64];
+};
+
+/* MPW session status. */
+enum mlx5_mpw_state {
+	MLX5_MPW_STATE_OPENED,
+	MLX5_MPW_INL_STATE_OPENED,
+	MLX5_MPW_STATE_CLOSED,
+};
+
+/* MPW session descriptor. */
+struct mlx5_mpw {
+	enum mlx5_mpw_state state;
+	unsigned int pkts_n;
+	unsigned int len;
+	unsigned int total_len;
+	volatile union mlx5_wqe *wqe;
+	union {
+		volatile struct mlx5_wqe_data_seg *dseg[MLX5_MPW_DSEG_MAX];
+		volatile uint8_t *raw;
+	} data;
+};
+
+/* CQ element structure - should be equal to the cache line size */
+struct mlx5_cqe {
+#if (RTE_CACHE_LINE_SIZE == 128)
+	uint8_t padding[64];
+#endif
+	struct mlx5_cqe64 cqe64;
+};
+
+#endif /* RTE_PMD_MLX5_PRM_H_ */
diff --git a/src/dpdk/drivers/net/mlx5/mlx5_rss.c b/src/dpdk/drivers/net/mlx5/mlx5_rss.c
new file mode 100644
index 00000000..639e935b
--- /dev/null
+++ b/src/dpdk/drivers/net/mlx5/mlx5_rss.c
@@ -0,0 +1,367 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2015 6WIND S.A.
+ *   Copyright 2015 Mellanox.
+ *
+ *   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.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <stddef.h>
+#include <stdint.h>
+#include <errno.h>
+#include <string.h>
+#include <assert.h>
+
+/* Verbs header. */
+/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <infiniband/verbs.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+/* DPDK headers don't like -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <rte_malloc.h>
+#include <rte_ethdev.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+#include "mlx5.h"
+#include "mlx5_rxtx.h"
+
+/**
+ * Get a RSS configuration hash key.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param rss_hf
+ *   RSS hash functions configuration must be retrieved for.
+ *
+ * @return
+ *   Pointer to a RSS configuration structure or NULL if rss_hf cannot
+ *   be matched.
+ */
+static struct rte_eth_rss_conf *
+rss_hash_get(struct priv *priv, uint64_t rss_hf)
+{
+	unsigned int i;
+
+	for (i = 0; (i != hash_rxq_init_n); ++i) {
+		uint64_t dpdk_rss_hf = hash_rxq_init[i].dpdk_rss_hf;
+
+		if (!(dpdk_rss_hf & rss_hf))
+			continue;
+		return (*priv->rss_conf)[i];
+	}
+	return NULL;
+}
+
+/**
+ * Register a RSS key.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param key
+ *   Hash key to register.
+ * @param key_len
+ *   Hash key length in bytes.
+ * @param rss_hf
+ *   RSS hash functions the provided key applies to.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+int
+rss_hash_rss_conf_new_key(struct priv *priv, const uint8_t *key,
+			  unsigned int key_len, uint64_t rss_hf)
+{
+	unsigned int i;
+
+	for (i = 0; (i != hash_rxq_init_n); ++i) {
+		struct rte_eth_rss_conf *rss_conf;
+		uint64_t dpdk_rss_hf = hash_rxq_init[i].dpdk_rss_hf;
+
+		if (!(dpdk_rss_hf & rss_hf))
+			continue;
+		rss_conf = rte_realloc((*priv->rss_conf)[i],
+				       (sizeof(*rss_conf) + key_len),
+				       0);
+		if (!rss_conf)
+			return ENOMEM;
+		rss_conf->rss_key = (void *)(rss_conf + 1);
+		rss_conf->rss_key_len = key_len;
+		rss_conf->rss_hf = dpdk_rss_hf;
+		memcpy(rss_conf->rss_key, key, key_len);
+		(*priv->rss_conf)[i] = rss_conf;
+	}
+	return 0;
+}
+
+/**
+ * DPDK callback to update the RSS hash configuration.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param[in] rss_conf
+ *   RSS configuration data.
+ *
+ * @return
+ *   0 on success, negative errno value on failure.
+ */
+int
+mlx5_rss_hash_update(struct rte_eth_dev *dev,
+		     struct rte_eth_rss_conf *rss_conf)
+{
+	struct priv *priv = dev->data->dev_private;
+	int err = 0;
+
+	priv_lock(priv);
+
+	assert(priv->rss_conf != NULL);
+
+	/* Apply configuration. */
+	if (rss_conf->rss_key)
+		err = rss_hash_rss_conf_new_key(priv,
+						rss_conf->rss_key,
+						rss_conf->rss_key_len,
+						rss_conf->rss_hf);
+	/* Store protocols for which RSS is enabled. */
+	priv->rss_hf = rss_conf->rss_hf;
+	priv_unlock(priv);
+	assert(err >= 0);
+	return -err;
+}
+
+/**
+ * DPDK callback to get the RSS hash configuration.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param[in, out] rss_conf
+ *   RSS configuration data.
+ *
+ * @return
+ *   0 on success, negative errno value on failure.
+ */
+int
+mlx5_rss_hash_conf_get(struct rte_eth_dev *dev,
+		       struct rte_eth_rss_conf *rss_conf)
+{
+	struct priv *priv = dev->data->dev_private;
+	struct rte_eth_rss_conf *priv_rss_conf;
+
+	priv_lock(priv);
+
+	assert(priv->rss_conf != NULL);
+
+	priv_rss_conf = rss_hash_get(priv, rss_conf->rss_hf);
+	if (!priv_rss_conf) {
+		rss_conf->rss_hf = 0;
+		priv_unlock(priv);
+		return -EINVAL;
+	}
+	if (rss_conf->rss_key &&
+	    rss_conf->rss_key_len >= priv_rss_conf->rss_key_len)
+		memcpy(rss_conf->rss_key,
+		       priv_rss_conf->rss_key,
+		       priv_rss_conf->rss_key_len);
+	rss_conf->rss_key_len = priv_rss_conf->rss_key_len;
+	rss_conf->rss_hf = priv_rss_conf->rss_hf;
+
+	priv_unlock(priv);
+	return 0;
+}
+
+/**
+ * Allocate/reallocate RETA index table.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @praram reta_size
+ *   The size of the array to allocate.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+int
+priv_rss_reta_index_resize(struct priv *priv, unsigned int reta_size)
+{
+	void *mem;
+	unsigned int old_size = priv->reta_idx_n;
+
+	if (priv->reta_idx_n == reta_size)
+		return 0;
+
+	mem = rte_realloc(priv->reta_idx,
+			  reta_size * sizeof((*priv->reta_idx)[0]), 0);
+	if (!mem)
+		return ENOMEM;
+	priv->reta_idx = mem;
+	priv->reta_idx_n = reta_size;
+
+	if (old_size < reta_size)
+		memset(&(*priv->reta_idx)[old_size], 0,
+		       (reta_size - old_size) *
+		       sizeof((*priv->reta_idx)[0]));
+	return 0;
+}
+
+/**
+ * Query RETA table.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param[in, out] reta_conf
+ *   Pointer to the first RETA configuration structure.
+ * @param reta_size
+ *   Number of entries.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+priv_dev_rss_reta_query(struct priv *priv,
+			struct rte_eth_rss_reta_entry64 *reta_conf,
+			unsigned int reta_size)
+{
+	unsigned int idx;
+	unsigned int i;
+	int ret;
+
+	/* See RETA comment in mlx5_dev_infos_get(). */
+	ret = priv_rss_reta_index_resize(priv, priv->ind_table_max_size);
+	if (ret)
+		return ret;
+
+	/* Fill each entry of the table even if its bit is not set. */
+	for (idx = 0, i = 0; (i != reta_size); ++i) {
+		idx = i / RTE_RETA_GROUP_SIZE;
+		reta_conf[idx].reta[i % RTE_RETA_GROUP_SIZE] =
+			(*priv->reta_idx)[i];
+	}
+	return 0;
+}
+
+/**
+ * Update RETA table.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param[in] reta_conf
+ *   Pointer to the first RETA configuration structure.
+ * @param reta_size
+ *   Number of entries.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+priv_dev_rss_reta_update(struct priv *priv,
+			 struct rte_eth_rss_reta_entry64 *reta_conf,
+			 unsigned int reta_size)
+{
+	unsigned int idx;
+	unsigned int i;
+	unsigned int pos;
+	int ret;
+
+	/* See RETA comment in mlx5_dev_infos_get(). */
+	ret = priv_rss_reta_index_resize(priv, priv->ind_table_max_size);
+	if (ret)
+		return ret;
+
+	for (idx = 0, i = 0; (i != reta_size); ++i) {
+		idx = i / RTE_RETA_GROUP_SIZE;
+		pos = i % RTE_RETA_GROUP_SIZE;
+		if (((reta_conf[idx].mask >> i) & 0x1) == 0)
+			continue;
+		assert(reta_conf[idx].reta[pos] < priv->rxqs_n);
+		(*priv->reta_idx)[i] = reta_conf[idx].reta[pos];
+	}
+	return 0;
+}
+
+/**
+ * DPDK callback to get the RETA indirection table.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param reta_conf
+ *   Pointer to RETA configuration structure array.
+ * @param reta_size
+ *   Size of the RETA table.
+ *
+ * @return
+ *   0 on success, negative errno value on failure.
+ */
+int
+mlx5_dev_rss_reta_query(struct rte_eth_dev *dev,
+			struct rte_eth_rss_reta_entry64 *reta_conf,
+			uint16_t reta_size)
+{
+	int ret;
+	struct priv *priv = dev->data->dev_private;
+
+	priv_lock(priv);
+	ret = priv_dev_rss_reta_query(priv, reta_conf, reta_size);
+	priv_unlock(priv);
+	return -ret;
+}
+
+/**
+ * DPDK callback to update the RETA indirection table.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param reta_conf
+ *   Pointer to RETA configuration structure array.
+ * @param reta_size
+ *   Size of the RETA table.
+ *
+ * @return
+ *   0 on success, negative errno value on failure.
+ */
+int
+mlx5_dev_rss_reta_update(struct rte_eth_dev *dev,
+			 struct rte_eth_rss_reta_entry64 *reta_conf,
+			 uint16_t reta_size)
+{
+	int ret;
+	struct priv *priv = dev->data->dev_private;
+
+	priv_lock(priv);
+	ret = priv_dev_rss_reta_update(priv, reta_conf, reta_size);
+	priv_unlock(priv);
+	return -ret;
+}
diff --git a/src/dpdk/drivers/net/mlx5/mlx5_rxmode.c b/src/dpdk/drivers/net/mlx5/mlx5_rxmode.c
new file mode 100644
index 00000000..8b585554
--- /dev/null
+++ b/src/dpdk/drivers/net/mlx5/mlx5_rxmode.c
@@ -0,0 +1,477 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2015 6WIND S.A.
+ *   Copyright 2015 Mellanox.
+ *
+ *   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.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <stddef.h>
+#include <errno.h>
+#include <string.h>
+
+/* Verbs header. */
+/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <infiniband/verbs.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+/* DPDK headers don't like -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <rte_ethdev.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+#include "mlx5.h"
+#include "mlx5_rxtx.h"
+#include "mlx5_utils.h"
+
+/* Initialization data for special flows. */
+static const struct special_flow_init special_flow_init[] = {
+	[HASH_RXQ_FLOW_TYPE_PROMISC] = {
+		.dst_mac_val = "\x00\x00\x00\x00\x00\x00",
+		.dst_mac_mask = "\x00\x00\x00\x00\x00\x00",
+		.hash_types =
+			1 << HASH_RXQ_TCPV4 |
+			1 << HASH_RXQ_UDPV4 |
+			1 << HASH_RXQ_IPV4 |
+			1 << HASH_RXQ_TCPV6 |
+			1 << HASH_RXQ_UDPV6 |
+			1 << HASH_RXQ_IPV6 |
+			1 << HASH_RXQ_ETH |
+			0,
+		.per_vlan = 0,
+	},
+	[HASH_RXQ_FLOW_TYPE_ALLMULTI] = {
+		.dst_mac_val = "\x01\x00\x00\x00\x00\x00",
+		.dst_mac_mask = "\x01\x00\x00\x00\x00\x00",
+		.hash_types =
+			1 << HASH_RXQ_UDPV4 |
+			1 << HASH_RXQ_IPV4 |
+			1 << HASH_RXQ_UDPV6 |
+			1 << HASH_RXQ_IPV6 |
+			1 << HASH_RXQ_ETH |
+			0,
+		.per_vlan = 0,
+	},
+	[HASH_RXQ_FLOW_TYPE_BROADCAST] = {
+		.dst_mac_val = "\xff\xff\xff\xff\xff\xff",
+		.dst_mac_mask = "\xff\xff\xff\xff\xff\xff",
+		.hash_types =
+			1 << HASH_RXQ_UDPV4 |
+			1 << HASH_RXQ_IPV4 |
+			1 << HASH_RXQ_UDPV6 |
+			1 << HASH_RXQ_IPV6 |
+			1 << HASH_RXQ_ETH |
+			0,
+		.per_vlan = 1,
+	},
+	[HASH_RXQ_FLOW_TYPE_IPV6MULTI] = {
+		.dst_mac_val = "\x33\x33\x00\x00\x00\x00",
+		.dst_mac_mask = "\xff\xff\x00\x00\x00\x00",
+		.hash_types =
+			1 << HASH_RXQ_UDPV6 |
+			1 << HASH_RXQ_IPV6 |
+			1 << HASH_RXQ_ETH |
+			0,
+		.per_vlan = 1,
+	},
+};
+
+/**
+ * Enable a special flow in a hash RX queue for a given VLAN index.
+ *
+ * @param hash_rxq
+ *   Pointer to hash RX queue structure.
+ * @param flow_type
+ *   Special flow type.
+ * @param vlan_index
+ *   VLAN index to use.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+hash_rxq_special_flow_enable_vlan(struct hash_rxq *hash_rxq,
+				  enum hash_rxq_flow_type flow_type,
+				  unsigned int vlan_index)
+{
+	struct priv *priv = hash_rxq->priv;
+	struct ibv_exp_flow *flow;
+	FLOW_ATTR_SPEC_ETH(data, priv_flow_attr(priv, NULL, 0, hash_rxq->type));
+	struct ibv_exp_flow_attr *attr = &data->attr;
+	struct ibv_exp_flow_spec_eth *spec = &data->spec;
+	const uint8_t *mac;
+	const uint8_t *mask;
+	unsigned int vlan_enabled = (priv->vlan_filter_n &&
+				     special_flow_init[flow_type].per_vlan);
+	unsigned int vlan_id = priv->vlan_filter[vlan_index];
+
+	/* Check if flow is relevant for this hash_rxq. */
+	if (!(special_flow_init[flow_type].hash_types & (1 << hash_rxq->type)))
+		return 0;
+	/* Check if flow already exists. */
+	if (hash_rxq->special_flow[flow_type][vlan_index] != NULL)
+		return 0;
+
+	/*
+	 * No padding must be inserted by the compiler between attr and spec.
+	 * This layout is expected by libibverbs.
+	 */
+	assert(((uint8_t *)attr + sizeof(*attr)) == (uint8_t *)spec);
+	priv_flow_attr(priv, attr, sizeof(data), hash_rxq->type);
+	/* The first specification must be Ethernet. */
+	assert(spec->type == IBV_EXP_FLOW_SPEC_ETH);
+	assert(spec->size == sizeof(*spec));
+
+	mac = special_flow_init[flow_type].dst_mac_val;
+	mask = special_flow_init[flow_type].dst_mac_mask;
+	*spec = (struct ibv_exp_flow_spec_eth){
+		.type = IBV_EXP_FLOW_SPEC_ETH,
+		.size = sizeof(*spec),
+		.val = {
+			.dst_mac = {
+				mac[0], mac[1], mac[2],
+				mac[3], mac[4], mac[5],
+			},
+			.vlan_tag = (vlan_enabled ? htons(vlan_id) : 0),
+		},
+		.mask = {
+			.dst_mac = {
+				mask[0], mask[1], mask[2],
+				mask[3], mask[4], mask[5],
+			},
+			.vlan_tag = (vlan_enabled ? htons(0xfff) : 0),
+		},
+	};
+
+	errno = 0;
+	flow = ibv_exp_create_flow(hash_rxq->qp, attr);
+	if (flow == NULL) {
+		/* It's not clear whether errno is always set in this case. */
+		ERROR("%p: flow configuration failed, errno=%d: %s",
+		      (void *)hash_rxq, errno,
+		      (errno ? strerror(errno) : "Unknown error"));
+		if (errno)
+			return errno;
+		return EINVAL;
+	}
+	hash_rxq->special_flow[flow_type][vlan_index] = flow;
+	DEBUG("%p: special flow %s (index %d) VLAN %u (index %u) enabled",
+	      (void *)hash_rxq, hash_rxq_flow_type_str(flow_type), flow_type,
+	      vlan_id, vlan_index);
+	return 0;
+}
+
+/**
+ * Disable a special flow in a hash RX queue for a given VLAN index.
+ *
+ * @param hash_rxq
+ *   Pointer to hash RX queue structure.
+ * @param flow_type
+ *   Special flow type.
+ * @param vlan_index
+ *   VLAN index to use.
+ */
+static void
+hash_rxq_special_flow_disable_vlan(struct hash_rxq *hash_rxq,
+				   enum hash_rxq_flow_type flow_type,
+				   unsigned int vlan_index)
+{
+	struct ibv_exp_flow *flow =
+		hash_rxq->special_flow[flow_type][vlan_index];
+
+	if (flow == NULL)
+		return;
+	claim_zero(ibv_exp_destroy_flow(flow));
+	hash_rxq->special_flow[flow_type][vlan_index] = NULL;
+	DEBUG("%p: special flow %s (index %d) VLAN %u (index %u) disabled",
+	      (void *)hash_rxq, hash_rxq_flow_type_str(flow_type), flow_type,
+	      hash_rxq->priv->vlan_filter[vlan_index], vlan_index);
+}
+
+/**
+ * Enable a special flow in a hash RX queue.
+ *
+ * @param hash_rxq
+ *   Pointer to hash RX queue structure.
+ * @param flow_type
+ *   Special flow type.
+ * @param vlan_index
+ *   VLAN index to use.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+hash_rxq_special_flow_enable(struct hash_rxq *hash_rxq,
+			     enum hash_rxq_flow_type flow_type)
+{
+	struct priv *priv = hash_rxq->priv;
+	unsigned int i = 0;
+	int ret;
+
+	assert((unsigned int)flow_type < RTE_DIM(hash_rxq->special_flow));
+	assert(RTE_DIM(hash_rxq->special_flow[flow_type]) ==
+	       RTE_DIM(priv->vlan_filter));
+	/* Add a special flow for each VLAN filter when relevant. */
+	do {
+		ret = hash_rxq_special_flow_enable_vlan(hash_rxq, flow_type, i);
+		if (ret) {
+			/* Failure, rollback. */
+			while (i != 0)
+				hash_rxq_special_flow_disable_vlan(hash_rxq,
+								   flow_type,
+								   --i);
+			return ret;
+		}
+	} while (special_flow_init[flow_type].per_vlan &&
+		 ++i < priv->vlan_filter_n);
+	return 0;
+}
+
+/**
+ * Disable a special flow in a hash RX queue.
+ *
+ * @param hash_rxq
+ *   Pointer to hash RX queue structure.
+ * @param flow_type
+ *   Special flow type.
+ */
+static void
+hash_rxq_special_flow_disable(struct hash_rxq *hash_rxq,
+			      enum hash_rxq_flow_type flow_type)
+{
+	unsigned int i;
+
+	assert((unsigned int)flow_type < RTE_DIM(hash_rxq->special_flow));
+	for (i = 0; (i != RTE_DIM(hash_rxq->special_flow[flow_type])); ++i)
+		hash_rxq_special_flow_disable_vlan(hash_rxq, flow_type, i);
+}
+
+/**
+ * Enable a special flow in all hash RX queues.
+ *
+ * @param priv
+ *   Private structure.
+ * @param flow_type
+ *   Special flow type.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+int
+priv_special_flow_enable(struct priv *priv, enum hash_rxq_flow_type flow_type)
+{
+	unsigned int i;
+
+	if (!priv_allow_flow_type(priv, flow_type))
+		return 0;
+	for (i = 0; (i != priv->hash_rxqs_n); ++i) {
+		struct hash_rxq *hash_rxq = &(*priv->hash_rxqs)[i];
+		int ret;
+
+		ret = hash_rxq_special_flow_enable(hash_rxq, flow_type);
+		if (!ret)
+			continue;
+		/* Failure, rollback. */
+		while (i != 0) {
+			hash_rxq = &(*priv->hash_rxqs)[--i];
+			hash_rxq_special_flow_disable(hash_rxq, flow_type);
+		}
+		return ret;
+	}
+	return 0;
+}
+
+/**
+ * Disable a special flow in all hash RX queues.
+ *
+ * @param priv
+ *   Private structure.
+ * @param flow_type
+ *   Special flow type.
+ */
+void
+priv_special_flow_disable(struct priv *priv, enum hash_rxq_flow_type flow_type)
+{
+	unsigned int i;
+
+	for (i = 0; (i != priv->hash_rxqs_n); ++i) {
+		struct hash_rxq *hash_rxq = &(*priv->hash_rxqs)[i];
+
+		hash_rxq_special_flow_disable(hash_rxq, flow_type);
+	}
+}
+
+/**
+ * Enable all special flows in all hash RX queues.
+ *
+ * @param priv
+ *   Private structure.
+ */
+int
+priv_special_flow_enable_all(struct priv *priv)
+{
+	enum hash_rxq_flow_type flow_type;
+
+	for (flow_type = HASH_RXQ_FLOW_TYPE_PROMISC;
+			flow_type != HASH_RXQ_FLOW_TYPE_MAC;
+			++flow_type) {
+		int ret;
+
+		ret = priv_special_flow_enable(priv, flow_type);
+		if (!ret)
+			continue;
+		/* Failure, rollback. */
+		while (flow_type)
+			priv_special_flow_disable(priv, --flow_type);
+		return ret;
+	}
+	return 0;
+}
+
+/**
+ * Disable all special flows in all hash RX queues.
+ *
+ * @param priv
+ *   Private structure.
+ */
+void
+priv_special_flow_disable_all(struct priv *priv)
+{
+	enum hash_rxq_flow_type flow_type;
+
+	for (flow_type = HASH_RXQ_FLOW_TYPE_PROMISC;
+			flow_type != HASH_RXQ_FLOW_TYPE_MAC;
+			++flow_type)
+		priv_special_flow_disable(priv, flow_type);
+}
+
+/**
+ * DPDK callback to enable promiscuous mode.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ */
+void
+mlx5_promiscuous_enable(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	int ret;
+
+	if (mlx5_is_secondary())
+		return;
+
+	priv_lock(priv);
+	priv->promisc_req = 1;
+	ret = priv_rehash_flows(priv);
+	if (ret)
+		ERROR("error while enabling promiscuous mode: %s",
+		      strerror(ret));
+	priv_unlock(priv);
+}
+
+/**
+ * DPDK callback to disable promiscuous mode.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ */
+void
+mlx5_promiscuous_disable(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	int ret;
+
+	if (mlx5_is_secondary())
+		return;
+
+	priv_lock(priv);
+	priv->promisc_req = 0;
+	ret = priv_rehash_flows(priv);
+	if (ret)
+		ERROR("error while disabling promiscuous mode: %s",
+		      strerror(ret));
+	priv_unlock(priv);
+}
+
+/**
+ * DPDK callback to enable allmulti mode.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ */
+void
+mlx5_allmulticast_enable(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	int ret;
+
+	if (mlx5_is_secondary())
+		return;
+
+	priv_lock(priv);
+	priv->allmulti_req = 1;
+	ret = priv_rehash_flows(priv);
+	if (ret)
+		ERROR("error while enabling allmulticast mode: %s",
+		      strerror(ret));
+	priv_unlock(priv);
+}
+
+/**
+ * DPDK callback to disable allmulti mode.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ */
+void
+mlx5_allmulticast_disable(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	int ret;
+
+	if (mlx5_is_secondary())
+		return;
+
+	priv_lock(priv);
+	priv->allmulti_req = 0;
+	ret = priv_rehash_flows(priv);
+	if (ret)
+		ERROR("error while disabling allmulticast mode: %s",
+		      strerror(ret));
+	priv_unlock(priv);
+}
diff --git a/src/dpdk/drivers/net/mlx5/mlx5_rxq.c b/src/dpdk/drivers/net/mlx5/mlx5_rxq.c
new file mode 100644
index 00000000..29c137cd
--- /dev/null
+++ b/src/dpdk/drivers/net/mlx5/mlx5_rxq.c
@@ -0,0 +1,1341 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2015 6WIND S.A.
+ *   Copyright 2015 Mellanox.
+ *
+ *   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.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <stddef.h>
+#include <assert.h>
+#include <errno.h>
+#include <string.h>
+#include <stdint.h>
+
+/* Verbs header. */
+/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <infiniband/verbs.h>
+#include <infiniband/arch.h>
+#include <infiniband/mlx5_hw.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+/* DPDK headers don't like -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <rte_mbuf.h>
+#include <rte_malloc.h>
+#include <rte_ethdev.h>
+#include <rte_common.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+#include "mlx5.h"
+#include "mlx5_rxtx.h"
+#include "mlx5_utils.h"
+#include "mlx5_autoconf.h"
+#include "mlx5_defs.h"
+
+/* Initialization data for hash RX queues. */
+const struct hash_rxq_init hash_rxq_init[] = {
+	[HASH_RXQ_TCPV4] = {
+		.hash_fields = (IBV_EXP_RX_HASH_SRC_IPV4 |
+				IBV_EXP_RX_HASH_DST_IPV4 |
+				IBV_EXP_RX_HASH_SRC_PORT_TCP |
+				IBV_EXP_RX_HASH_DST_PORT_TCP),
+		.dpdk_rss_hf = ETH_RSS_NONFRAG_IPV4_TCP,
+		.flow_priority = 0,
+		.flow_spec.tcp_udp = {
+			.type = IBV_EXP_FLOW_SPEC_TCP,
+			.size = sizeof(hash_rxq_init[0].flow_spec.tcp_udp),
+		},
+		.underlayer = &hash_rxq_init[HASH_RXQ_IPV4],
+	},
+	[HASH_RXQ_UDPV4] = {
+		.hash_fields = (IBV_EXP_RX_HASH_SRC_IPV4 |
+				IBV_EXP_RX_HASH_DST_IPV4 |
+				IBV_EXP_RX_HASH_SRC_PORT_UDP |
+				IBV_EXP_RX_HASH_DST_PORT_UDP),
+		.dpdk_rss_hf = ETH_RSS_NONFRAG_IPV4_UDP,
+		.flow_priority = 0,
+		.flow_spec.tcp_udp = {
+			.type = IBV_EXP_FLOW_SPEC_UDP,
+			.size = sizeof(hash_rxq_init[0].flow_spec.tcp_udp),
+		},
+		.underlayer = &hash_rxq_init[HASH_RXQ_IPV4],
+	},
+	[HASH_RXQ_IPV4] = {
+		.hash_fields = (IBV_EXP_RX_HASH_SRC_IPV4 |
+				IBV_EXP_RX_HASH_DST_IPV4),
+		.dpdk_rss_hf = (ETH_RSS_IPV4 |
+				ETH_RSS_FRAG_IPV4),
+		.flow_priority = 1,
+		.flow_spec.ipv4 = {
+			.type = IBV_EXP_FLOW_SPEC_IPV4,
+			.size = sizeof(hash_rxq_init[0].flow_spec.ipv4),
+		},
+		.underlayer = &hash_rxq_init[HASH_RXQ_ETH],
+	},
+	[HASH_RXQ_TCPV6] = {
+		.hash_fields = (IBV_EXP_RX_HASH_SRC_IPV6 |
+				IBV_EXP_RX_HASH_DST_IPV6 |
+				IBV_EXP_RX_HASH_SRC_PORT_TCP |
+				IBV_EXP_RX_HASH_DST_PORT_TCP),
+		.dpdk_rss_hf = ETH_RSS_NONFRAG_IPV6_TCP,
+		.flow_priority = 0,
+		.flow_spec.tcp_udp = {
+			.type = IBV_EXP_FLOW_SPEC_TCP,
+			.size = sizeof(hash_rxq_init[0].flow_spec.tcp_udp),
+		},
+		.underlayer = &hash_rxq_init[HASH_RXQ_IPV6],
+	},
+	[HASH_RXQ_UDPV6] = {
+		.hash_fields = (IBV_EXP_RX_HASH_SRC_IPV6 |
+				IBV_EXP_RX_HASH_DST_IPV6 |
+				IBV_EXP_RX_HASH_SRC_PORT_UDP |
+				IBV_EXP_RX_HASH_DST_PORT_UDP),
+		.dpdk_rss_hf = ETH_RSS_NONFRAG_IPV6_UDP,
+		.flow_priority = 0,
+		.flow_spec.tcp_udp = {
+			.type = IBV_EXP_FLOW_SPEC_UDP,
+			.size = sizeof(hash_rxq_init[0].flow_spec.tcp_udp),
+		},
+		.underlayer = &hash_rxq_init[HASH_RXQ_IPV6],
+	},
+	[HASH_RXQ_IPV6] = {
+		.hash_fields = (IBV_EXP_RX_HASH_SRC_IPV6 |
+				IBV_EXP_RX_HASH_DST_IPV6),
+		.dpdk_rss_hf = (ETH_RSS_IPV6 |
+				ETH_RSS_FRAG_IPV6),
+		.flow_priority = 1,
+		.flow_spec.ipv6 = {
+			.type = IBV_EXP_FLOW_SPEC_IPV6,
+			.size = sizeof(hash_rxq_init[0].flow_spec.ipv6),
+		},
+		.underlayer = &hash_rxq_init[HASH_RXQ_ETH],
+	},
+	[HASH_RXQ_ETH] = {
+		.hash_fields = 0,
+		.dpdk_rss_hf = 0,
+		.flow_priority = 2,
+		.flow_spec.eth = {
+			.type = IBV_EXP_FLOW_SPEC_ETH,
+			.size = sizeof(hash_rxq_init[0].flow_spec.eth),
+		},
+		.underlayer = NULL,
+	},
+};
+
+/* Number of entries in hash_rxq_init[]. */
+const unsigned int hash_rxq_init_n = RTE_DIM(hash_rxq_init);
+
+/* Initialization data for hash RX queue indirection tables. */
+static const struct ind_table_init ind_table_init[] = {
+	{
+		.max_size = -1u, /* Superseded by HW limitations. */
+		.hash_types =
+			1 << HASH_RXQ_TCPV4 |
+			1 << HASH_RXQ_UDPV4 |
+			1 << HASH_RXQ_IPV4 |
+			1 << HASH_RXQ_TCPV6 |
+			1 << HASH_RXQ_UDPV6 |
+			1 << HASH_RXQ_IPV6 |
+			0,
+		.hash_types_n = 6,
+	},
+	{
+		.max_size = 1,
+		.hash_types = 1 << HASH_RXQ_ETH,
+		.hash_types_n = 1,
+	},
+};
+
+#define IND_TABLE_INIT_N RTE_DIM(ind_table_init)
+
+/* Default RSS hash key also used for ConnectX-3. */
+uint8_t rss_hash_default_key[] = {
+	0x2c, 0xc6, 0x81, 0xd1,
+	0x5b, 0xdb, 0xf4, 0xf7,
+	0xfc, 0xa2, 0x83, 0x19,
+	0xdb, 0x1a, 0x3e, 0x94,
+	0x6b, 0x9e, 0x38, 0xd9,
+	0x2c, 0x9c, 0x03, 0xd1,
+	0xad, 0x99, 0x44, 0xa7,
+	0xd9, 0x56, 0x3d, 0x59,
+	0x06, 0x3c, 0x25, 0xf3,
+	0xfc, 0x1f, 0xdc, 0x2a,
+};
+
+/* Length of the default RSS hash key. */
+const size_t rss_hash_default_key_len = sizeof(rss_hash_default_key);
+
+/**
+ * Populate flow steering rule for a given hash RX queue type using
+ * information from hash_rxq_init[]. Nothing is written to flow_attr when
+ * flow_attr_size is not large enough, but the required size is still returned.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param[out] flow_attr
+ *   Pointer to flow attribute structure to fill. Note that the allocated
+ *   area must be larger and large enough to hold all flow specifications.
+ * @param flow_attr_size
+ *   Entire size of flow_attr and trailing room for flow specifications.
+ * @param type
+ *   Hash RX queue type to use for flow steering rule.
+ *
+ * @return
+ *   Total size of the flow attribute buffer. No errors are defined.
+ */
+size_t
+priv_flow_attr(struct priv *priv, struct ibv_exp_flow_attr *flow_attr,
+	       size_t flow_attr_size, enum hash_rxq_type type)
+{
+	size_t offset = sizeof(*flow_attr);
+	const struct hash_rxq_init *init = &hash_rxq_init[type];
+
+	assert(priv != NULL);
+	assert((size_t)type < RTE_DIM(hash_rxq_init));
+	do {
+		offset += init->flow_spec.hdr.size;
+		init = init->underlayer;
+	} while (init != NULL);
+	if (offset > flow_attr_size)
+		return offset;
+	flow_attr_size = offset;
+	init = &hash_rxq_init[type];
+	*flow_attr = (struct ibv_exp_flow_attr){
+		.type = IBV_EXP_FLOW_ATTR_NORMAL,
+		/* Priorities < 3 are reserved for flow director. */
+		.priority = init->flow_priority + 3,
+		.num_of_specs = 0,
+		.port = priv->port,
+		.flags = 0,
+	};
+	do {
+		offset -= init->flow_spec.hdr.size;
+		memcpy((void *)((uintptr_t)flow_attr + offset),
+		       &init->flow_spec,
+		       init->flow_spec.hdr.size);
+		++flow_attr->num_of_specs;
+		init = init->underlayer;
+	} while (init != NULL);
+	return flow_attr_size;
+}
+
+/**
+ * Convert hash type position in indirection table initializer to
+ * hash RX queue type.
+ *
+ * @param table
+ *   Indirection table initializer.
+ * @param pos
+ *   Hash type position.
+ *
+ * @return
+ *   Hash RX queue type.
+ */
+static enum hash_rxq_type
+hash_rxq_type_from_pos(const struct ind_table_init *table, unsigned int pos)
+{
+	enum hash_rxq_type type = HASH_RXQ_TCPV4;
+
+	assert(pos < table->hash_types_n);
+	do {
+		if ((table->hash_types & (1 << type)) && (pos-- == 0))
+			break;
+		++type;
+	} while (1);
+	return type;
+}
+
+/**
+ * Filter out disabled hash RX queue types from ind_table_init[].
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param[out] table
+ *   Output table.
+ *
+ * @return
+ *   Number of table entries.
+ */
+static unsigned int
+priv_make_ind_table_init(struct priv *priv,
+			 struct ind_table_init (*table)[IND_TABLE_INIT_N])
+{
+	uint64_t rss_hf;
+	unsigned int i;
+	unsigned int j;
+	unsigned int table_n = 0;
+	/* Mandatory to receive frames not handled by normal hash RX queues. */
+	unsigned int hash_types_sup = 1 << HASH_RXQ_ETH;
+
+	rss_hf = priv->rss_hf;
+	/* Process other protocols only if more than one queue. */
+	if (priv->rxqs_n > 1)
+		for (i = 0; (i != hash_rxq_init_n); ++i)
+			if (rss_hf & hash_rxq_init[i].dpdk_rss_hf)
+				hash_types_sup |= (1 << i);
+
+	/* Filter out entries whose protocols are not in the set. */
+	for (i = 0, j = 0; (i != IND_TABLE_INIT_N); ++i) {
+		unsigned int nb;
+		unsigned int h;
+
+		/* j is increased only if the table has valid protocols. */
+		assert(j <= i);
+		(*table)[j] = ind_table_init[i];
+		(*table)[j].hash_types &= hash_types_sup;
+		for (h = 0, nb = 0; (h != hash_rxq_init_n); ++h)
+			if (((*table)[j].hash_types >> h) & 0x1)
+				++nb;
+		(*table)[i].hash_types_n = nb;
+		if (nb) {
+			++table_n;
+			++j;
+		}
+	}
+	return table_n;
+}
+
+/**
+ * Initialize hash RX queues and indirection table.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+int
+priv_create_hash_rxqs(struct priv *priv)
+{
+	struct ibv_exp_wq *wqs[priv->reta_idx_n];
+	struct ind_table_init ind_table_init[IND_TABLE_INIT_N];
+	unsigned int ind_tables_n =
+		priv_make_ind_table_init(priv, &ind_table_init);
+	unsigned int hash_rxqs_n = 0;
+	struct hash_rxq (*hash_rxqs)[] = NULL;
+	struct ibv_exp_rwq_ind_table *(*ind_tables)[] = NULL;
+	unsigned int i;
+	unsigned int j;
+	unsigned int k;
+	int err = 0;
+
+	assert(priv->ind_tables == NULL);
+	assert(priv->ind_tables_n == 0);
+	assert(priv->hash_rxqs == NULL);
+	assert(priv->hash_rxqs_n == 0);
+	assert(priv->pd != NULL);
+	assert(priv->ctx != NULL);
+	if (priv->rxqs_n == 0)
+		return EINVAL;
+	assert(priv->rxqs != NULL);
+	if (ind_tables_n == 0) {
+		ERROR("all hash RX queue types have been filtered out,"
+		      " indirection table cannot be created");
+		return EINVAL;
+	}
+	if (priv->rxqs_n & (priv->rxqs_n - 1)) {
+		INFO("%u RX queues are configured, consider rounding this"
+		     " number to the next power of two for better balancing",
+		     priv->rxqs_n);
+		DEBUG("indirection table extended to assume %u WQs",
+		      priv->reta_idx_n);
+	}
+	for (i = 0; (i != priv->reta_idx_n); ++i) {
+		struct rxq_ctrl *rxq_ctrl;
+
+		rxq_ctrl = container_of((*priv->rxqs)[(*priv->reta_idx)[i]],
+					struct rxq_ctrl, rxq);
+		wqs[i] = rxq_ctrl->wq;
+	}
+	/* Get number of hash RX queues to configure. */
+	for (i = 0, hash_rxqs_n = 0; (i != ind_tables_n); ++i)
+		hash_rxqs_n += ind_table_init[i].hash_types_n;
+	DEBUG("allocating %u hash RX queues for %u WQs, %u indirection tables",
+	      hash_rxqs_n, priv->rxqs_n, ind_tables_n);
+	/* Create indirection tables. */
+	ind_tables = rte_calloc(__func__, ind_tables_n,
+				sizeof((*ind_tables)[0]), 0);
+	if (ind_tables == NULL) {
+		err = ENOMEM;
+		ERROR("cannot allocate indirection tables container: %s",
+		      strerror(err));
+		goto error;
+	}
+	for (i = 0; (i != ind_tables_n); ++i) {
+		struct ibv_exp_rwq_ind_table_init_attr ind_init_attr = {
+			.pd = priv->pd,
+			.log_ind_tbl_size = 0, /* Set below. */
+			.ind_tbl = wqs,
+			.comp_mask = 0,
+		};
+		unsigned int ind_tbl_size = ind_table_init[i].max_size;
+		struct ibv_exp_rwq_ind_table *ind_table;
+
+		if (priv->reta_idx_n < ind_tbl_size)
+			ind_tbl_size = priv->reta_idx_n;
+		ind_init_attr.log_ind_tbl_size = log2above(ind_tbl_size);
+		errno = 0;
+		ind_table = ibv_exp_create_rwq_ind_table(priv->ctx,
+							 &ind_init_attr);
+		if (ind_table != NULL) {
+			(*ind_tables)[i] = ind_table;
+			continue;
+		}
+		/* Not clear whether errno is set. */
+		err = (errno ? errno : EINVAL);
+		ERROR("RX indirection table creation failed with error %d: %s",
+		      err, strerror(err));
+		goto error;
+	}
+	/* Allocate array that holds hash RX queues and related data. */
+	hash_rxqs = rte_calloc(__func__, hash_rxqs_n,
+			       sizeof((*hash_rxqs)[0]), 0);
+	if (hash_rxqs == NULL) {
+		err = ENOMEM;
+		ERROR("cannot allocate hash RX queues container: %s",
+		      strerror(err));
+		goto error;
+	}
+	for (i = 0, j = 0, k = 0;
+	     ((i != hash_rxqs_n) && (j != ind_tables_n));
+	     ++i) {
+		struct hash_rxq *hash_rxq = &(*hash_rxqs)[i];
+		enum hash_rxq_type type =
+			hash_rxq_type_from_pos(&ind_table_init[j], k);
+		struct rte_eth_rss_conf *priv_rss_conf =
+			(*priv->rss_conf)[type];
+		struct ibv_exp_rx_hash_conf hash_conf = {
+			.rx_hash_function = IBV_EXP_RX_HASH_FUNC_TOEPLITZ,
+			.rx_hash_key_len = (priv_rss_conf ?
+					    priv_rss_conf->rss_key_len :
+					    rss_hash_default_key_len),
+			.rx_hash_key = (priv_rss_conf ?
+					priv_rss_conf->rss_key :
+					rss_hash_default_key),
+			.rx_hash_fields_mask = hash_rxq_init[type].hash_fields,
+			.rwq_ind_tbl = (*ind_tables)[j],
+		};
+		struct ibv_exp_qp_init_attr qp_init_attr = {
+			.max_inl_recv = 0, /* Currently not supported. */
+			.qp_type = IBV_QPT_RAW_PACKET,
+			.comp_mask = (IBV_EXP_QP_INIT_ATTR_PD |
+				      IBV_EXP_QP_INIT_ATTR_RX_HASH),
+			.pd = priv->pd,
+			.rx_hash_conf = &hash_conf,
+			.port_num = priv->port,
+		};
+
+		DEBUG("using indirection table %u for hash RX queue %u type %d",
+		      j, i, type);
+		*hash_rxq = (struct hash_rxq){
+			.priv = priv,
+			.qp = ibv_exp_create_qp(priv->ctx, &qp_init_attr),
+			.type = type,
+		};
+		if (hash_rxq->qp == NULL) {
+			err = (errno ? errno : EINVAL);
+			ERROR("Hash RX QP creation failure: %s",
+			      strerror(err));
+			goto error;
+		}
+		if (++k < ind_table_init[j].hash_types_n)
+			continue;
+		/* Switch to the next indirection table and reset hash RX
+		 * queue type array index. */
+		++j;
+		k = 0;
+	}
+	priv->ind_tables = ind_tables;
+	priv->ind_tables_n = ind_tables_n;
+	priv->hash_rxqs = hash_rxqs;
+	priv->hash_rxqs_n = hash_rxqs_n;
+	assert(err == 0);
+	return 0;
+error:
+	if (hash_rxqs != NULL) {
+		for (i = 0; (i != hash_rxqs_n); ++i) {
+			struct ibv_qp *qp = (*hash_rxqs)[i].qp;
+
+			if (qp == NULL)
+				continue;
+			claim_zero(ibv_destroy_qp(qp));
+		}
+		rte_free(hash_rxqs);
+	}
+	if (ind_tables != NULL) {
+		for (j = 0; (j != ind_tables_n); ++j) {
+			struct ibv_exp_rwq_ind_table *ind_table =
+				(*ind_tables)[j];
+
+			if (ind_table == NULL)
+				continue;
+			claim_zero(ibv_exp_destroy_rwq_ind_table(ind_table));
+		}
+		rte_free(ind_tables);
+	}
+	return err;
+}
+
+/**
+ * Clean up hash RX queues and indirection table.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ */
+void
+priv_destroy_hash_rxqs(struct priv *priv)
+{
+	unsigned int i;
+
+	DEBUG("destroying %u hash RX queues", priv->hash_rxqs_n);
+	if (priv->hash_rxqs_n == 0) {
+		assert(priv->hash_rxqs == NULL);
+		assert(priv->ind_tables == NULL);
+		return;
+	}
+	for (i = 0; (i != priv->hash_rxqs_n); ++i) {
+		struct hash_rxq *hash_rxq = &(*priv->hash_rxqs)[i];
+		unsigned int j, k;
+
+		assert(hash_rxq->priv == priv);
+		assert(hash_rxq->qp != NULL);
+		/* Also check that there are no remaining flows. */
+		for (j = 0; (j != RTE_DIM(hash_rxq->special_flow)); ++j)
+			for (k = 0;
+			     (k != RTE_DIM(hash_rxq->special_flow[j]));
+			     ++k)
+				assert(hash_rxq->special_flow[j][k] == NULL);
+		for (j = 0; (j != RTE_DIM(hash_rxq->mac_flow)); ++j)
+			for (k = 0; (k != RTE_DIM(hash_rxq->mac_flow[j])); ++k)
+				assert(hash_rxq->mac_flow[j][k] == NULL);
+		claim_zero(ibv_destroy_qp(hash_rxq->qp));
+	}
+	priv->hash_rxqs_n = 0;
+	rte_free(priv->hash_rxqs);
+	priv->hash_rxqs = NULL;
+	for (i = 0; (i != priv->ind_tables_n); ++i) {
+		struct ibv_exp_rwq_ind_table *ind_table =
+			(*priv->ind_tables)[i];
+
+		assert(ind_table != NULL);
+		claim_zero(ibv_exp_destroy_rwq_ind_table(ind_table));
+	}
+	priv->ind_tables_n = 0;
+	rte_free(priv->ind_tables);
+	priv->ind_tables = NULL;
+}
+
+/**
+ * Check whether a given flow type is allowed.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param type
+ *   Flow type to check.
+ *
+ * @return
+ *   Nonzero if the given flow type is allowed.
+ */
+int
+priv_allow_flow_type(struct priv *priv, enum hash_rxq_flow_type type)
+{
+	/* Only FLOW_TYPE_PROMISC is allowed when promiscuous mode
+	 * has been requested. */
+	if (priv->promisc_req)
+		return type == HASH_RXQ_FLOW_TYPE_PROMISC;
+	switch (type) {
+	case HASH_RXQ_FLOW_TYPE_PROMISC:
+		return !!priv->promisc_req;
+	case HASH_RXQ_FLOW_TYPE_ALLMULTI:
+		return !!priv->allmulti_req;
+	case HASH_RXQ_FLOW_TYPE_BROADCAST:
+	case HASH_RXQ_FLOW_TYPE_IPV6MULTI:
+		/* If allmulti is enabled, broadcast and ipv6multi
+		 * are unnecessary. */
+		return !priv->allmulti_req;
+	case HASH_RXQ_FLOW_TYPE_MAC:
+		return 1;
+	default:
+		/* Unsupported flow type is not allowed. */
+		return 0;
+	}
+	return 0;
+}
+
+/**
+ * Automatically enable/disable flows according to configuration.
+ *
+ * @param priv
+ *   Private structure.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+int
+priv_rehash_flows(struct priv *priv)
+{
+	enum hash_rxq_flow_type i;
+
+	for (i = HASH_RXQ_FLOW_TYPE_PROMISC;
+			i != RTE_DIM((*priv->hash_rxqs)[0].special_flow);
+			++i)
+		if (!priv_allow_flow_type(priv, i)) {
+			priv_special_flow_disable(priv, i);
+		} else {
+			int ret = priv_special_flow_enable(priv, i);
+
+			if (ret)
+				return ret;
+		}
+	if (priv_allow_flow_type(priv, HASH_RXQ_FLOW_TYPE_MAC))
+		return priv_mac_addrs_enable(priv);
+	priv_mac_addrs_disable(priv);
+	return 0;
+}
+
+/**
+ * Allocate RX queue elements.
+ *
+ * @param rxq_ctrl
+ *   Pointer to RX queue structure.
+ * @param elts_n
+ *   Number of elements to allocate.
+ * @param[in] pool
+ *   If not NULL, fetch buffers from this array instead of allocating them
+ *   with rte_pktmbuf_alloc().
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+rxq_alloc_elts(struct rxq_ctrl *rxq_ctrl, unsigned int elts_n,
+	       struct rte_mbuf *(*pool)[])
+{
+	const unsigned int sges_n = 1 << rxq_ctrl->rxq.sges_n;
+	unsigned int i;
+	int ret = 0;
+
+	/* Iterate on segments. */
+	for (i = 0; (i != elts_n); ++i) {
+		struct rte_mbuf *buf;
+		volatile struct mlx5_wqe_data_seg *scat =
+			&(*rxq_ctrl->rxq.wqes)[i];
+
+		if (pool != NULL) {
+			buf = (*pool)[i];
+			assert(buf != NULL);
+			rte_pktmbuf_reset(buf);
+			rte_pktmbuf_refcnt_update(buf, 1);
+		} else
+			buf = rte_pktmbuf_alloc(rxq_ctrl->rxq.mp);
+		if (buf == NULL) {
+			assert(pool == NULL);
+			ERROR("%p: empty mbuf pool", (void *)rxq_ctrl);
+			ret = ENOMEM;
+			goto error;
+		}
+		/* Headroom is reserved by rte_pktmbuf_alloc(). */
+		assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
+		/* Buffer is supposed to be empty. */
+		assert(rte_pktmbuf_data_len(buf) == 0);
+		assert(rte_pktmbuf_pkt_len(buf) == 0);
+		assert(!buf->next);
+		/* Only the first segment keeps headroom. */
+		if (i % sges_n)
+			SET_DATA_OFF(buf, 0);
+		PORT(buf) = rxq_ctrl->rxq.port_id;
+		DATA_LEN(buf) = rte_pktmbuf_tailroom(buf);
+		PKT_LEN(buf) = DATA_LEN(buf);
+		NB_SEGS(buf) = 1;
+		/* scat->addr must be able to store a pointer. */
+		assert(sizeof(scat->addr) >= sizeof(uintptr_t));
+		*scat = (struct mlx5_wqe_data_seg){
+			.addr = htonll(rte_pktmbuf_mtod(buf, uintptr_t)),
+			.byte_count = htonl(DATA_LEN(buf)),
+			.lkey = htonl(rxq_ctrl->mr->lkey),
+		};
+		(*rxq_ctrl->rxq.elts)[i] = buf;
+	}
+	DEBUG("%p: allocated and configured %u segments (max %u packets)",
+	      (void *)rxq_ctrl, elts_n, elts_n / (1 << rxq_ctrl->rxq.sges_n));
+	assert(ret == 0);
+	return 0;
+error:
+	assert(pool == NULL);
+	elts_n = i;
+	for (i = 0; (i != elts_n); ++i) {
+		if ((*rxq_ctrl->rxq.elts)[i] != NULL)
+			rte_pktmbuf_free_seg((*rxq_ctrl->rxq.elts)[i]);
+		(*rxq_ctrl->rxq.elts)[i] = NULL;
+	}
+	DEBUG("%p: failed, freed everything", (void *)rxq_ctrl);
+	assert(ret > 0);
+	return ret;
+}
+
+/**
+ * Free RX queue elements.
+ *
+ * @param rxq_ctrl
+ *   Pointer to RX queue structure.
+ */
+static void
+rxq_free_elts(struct rxq_ctrl *rxq_ctrl)
+{
+	unsigned int i;
+
+	DEBUG("%p: freeing WRs", (void *)rxq_ctrl);
+	if (rxq_ctrl->rxq.elts == NULL)
+		return;
+
+	for (i = 0; (i != rxq_ctrl->rxq.elts_n); ++i) {
+		if ((*rxq_ctrl->rxq.elts)[i] != NULL)
+			rte_pktmbuf_free_seg((*rxq_ctrl->rxq.elts)[i]);
+		(*rxq_ctrl->rxq.elts)[i] = NULL;
+	}
+}
+
+/**
+ * Clean up a RX queue.
+ *
+ * Destroy objects, free allocated memory and reset the structure for reuse.
+ *
+ * @param rxq_ctrl
+ *   Pointer to RX queue structure.
+ */
+void
+rxq_cleanup(struct rxq_ctrl *rxq_ctrl)
+{
+	struct ibv_exp_release_intf_params params;
+
+	DEBUG("cleaning up %p", (void *)rxq_ctrl);
+	rxq_free_elts(rxq_ctrl);
+	if (rxq_ctrl->if_wq != NULL) {
+		assert(rxq_ctrl->priv != NULL);
+		assert(rxq_ctrl->priv->ctx != NULL);
+		assert(rxq_ctrl->wq != NULL);
+		params = (struct ibv_exp_release_intf_params){
+			.comp_mask = 0,
+		};
+		claim_zero(ibv_exp_release_intf(rxq_ctrl->priv->ctx,
+						rxq_ctrl->if_wq,
+						&params));
+	}
+	if (rxq_ctrl->if_cq != NULL) {
+		assert(rxq_ctrl->priv != NULL);
+		assert(rxq_ctrl->priv->ctx != NULL);
+		assert(rxq_ctrl->cq != NULL);
+		params = (struct ibv_exp_release_intf_params){
+			.comp_mask = 0,
+		};
+		claim_zero(ibv_exp_release_intf(rxq_ctrl->priv->ctx,
+						rxq_ctrl->if_cq,
+						&params));
+	}
+	if (rxq_ctrl->wq != NULL)
+		claim_zero(ibv_exp_destroy_wq(rxq_ctrl->wq));
+	if (rxq_ctrl->cq != NULL)
+		claim_zero(ibv_destroy_cq(rxq_ctrl->cq));
+	if (rxq_ctrl->rd != NULL) {
+		struct ibv_exp_destroy_res_domain_attr attr = {
+			.comp_mask = 0,
+		};
+
+		assert(rxq_ctrl->priv != NULL);
+		assert(rxq_ctrl->priv->ctx != NULL);
+		claim_zero(ibv_exp_destroy_res_domain(rxq_ctrl->priv->ctx,
+						      rxq_ctrl->rd,
+						      &attr));
+	}
+	if (rxq_ctrl->mr != NULL)
+		claim_zero(ibv_dereg_mr(rxq_ctrl->mr));
+	memset(rxq_ctrl, 0, sizeof(*rxq_ctrl));
+}
+
+/**
+ * Reconfigure RX queue buffers.
+ *
+ * rxq_rehash() does not allocate mbufs, which, if not done from the right
+ * thread (such as a control thread), may corrupt the pool.
+ * In case of failure, the queue is left untouched.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param rxq_ctrl
+ *   RX queue pointer.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+int
+rxq_rehash(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl)
+{
+	unsigned int elts_n = rxq_ctrl->rxq.elts_n;
+	unsigned int i;
+	struct ibv_exp_wq_attr mod;
+	int err;
+
+	DEBUG("%p: rehashing queue %p with %u SGE(s) per packet",
+	      (void *)dev, (void *)rxq_ctrl, 1 << rxq_ctrl->rxq.sges_n);
+	assert(!(elts_n % (1 << rxq_ctrl->rxq.sges_n)));
+	/* From now on, any failure will render the queue unusable.
+	 * Reinitialize WQ. */
+	mod = (struct ibv_exp_wq_attr){
+		.attr_mask = IBV_EXP_WQ_ATTR_STATE,
+		.wq_state = IBV_EXP_WQS_RESET,
+	};
+	err = ibv_exp_modify_wq(rxq_ctrl->wq, &mod);
+	if (err) {
+		ERROR("%p: cannot reset WQ: %s", (void *)dev, strerror(err));
+		assert(err > 0);
+		return err;
+	}
+	/* Snatch mbufs from original queue. */
+	claim_zero(rxq_alloc_elts(rxq_ctrl, elts_n, rxq_ctrl->rxq.elts));
+	for (i = 0; i != elts_n; ++i) {
+		struct rte_mbuf *buf = (*rxq_ctrl->rxq.elts)[i];
+
+		assert(rte_mbuf_refcnt_read(buf) == 2);
+		rte_pktmbuf_free_seg(buf);
+	}
+	/* Change queue state to ready. */
+	mod = (struct ibv_exp_wq_attr){
+		.attr_mask = IBV_EXP_WQ_ATTR_STATE,
+		.wq_state = IBV_EXP_WQS_RDY,
+	};
+	err = ibv_exp_modify_wq(rxq_ctrl->wq, &mod);
+	if (err) {
+		ERROR("%p: WQ state to IBV_EXP_WQS_RDY failed: %s",
+		      (void *)dev, strerror(err));
+		goto error;
+	}
+	/* Update doorbell counter. */
+	rxq_ctrl->rxq.rq_ci = elts_n >> rxq_ctrl->rxq.sges_n;
+	rte_wmb();
+	*rxq_ctrl->rxq.rq_db = htonl(rxq_ctrl->rxq.rq_ci);
+error:
+	assert(err >= 0);
+	return err;
+}
+
+/**
+ * Initialize RX queue.
+ *
+ * @param tmpl
+ *   Pointer to RX queue control template.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static inline int
+rxq_setup(struct rxq_ctrl *tmpl)
+{
+	struct ibv_cq *ibcq = tmpl->cq;
+	struct mlx5_cq *cq = to_mxxx(cq, cq);
+	struct mlx5_rwq *rwq = container_of(tmpl->wq, struct mlx5_rwq, wq);
+	struct rte_mbuf *(*elts)[tmpl->rxq.elts_n] =
+		rte_calloc_socket("RXQ", 1, sizeof(*elts), 0, tmpl->socket);
+
+	if (cq->cqe_sz != RTE_CACHE_LINE_SIZE) {
+		ERROR("Wrong MLX5_CQE_SIZE environment variable value: "
+		      "it should be set to %u", RTE_CACHE_LINE_SIZE);
+		return EINVAL;
+	}
+	if (elts == NULL)
+		return ENOMEM;
+	tmpl->rxq.rq_db = rwq->rq.db;
+	tmpl->rxq.cqe_n = ibcq->cqe + 1;
+	tmpl->rxq.cq_ci = 0;
+	tmpl->rxq.rq_ci = 0;
+	tmpl->rxq.cq_db = cq->dbrec;
+	tmpl->rxq.wqes =
+		(volatile struct mlx5_wqe_data_seg (*)[])
+		(uintptr_t)rwq->rq.buff;
+	tmpl->rxq.cqes =
+		(volatile struct mlx5_cqe (*)[])
+		(uintptr_t)cq->active_buf->buf;
+	tmpl->rxq.elts = elts;
+	return 0;
+}
+
+/**
+ * Configure a RX queue.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param rxq_ctrl
+ *   Pointer to RX queue structure.
+ * @param desc
+ *   Number of descriptors to configure in queue.
+ * @param socket
+ *   NUMA socket on which memory must be allocated.
+ * @param[in] conf
+ *   Thresholds parameters.
+ * @param mp
+ *   Memory pool for buffer allocations.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+int
+rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
+	       uint16_t desc, unsigned int socket,
+	       const struct rte_eth_rxconf *conf, struct rte_mempool *mp)
+{
+	struct priv *priv = dev->data->dev_private;
+	struct rxq_ctrl tmpl = {
+		.priv = priv,
+		.socket = socket,
+		.rxq = {
+			.elts_n = desc,
+			.mp = mp,
+		},
+	};
+	struct ibv_exp_wq_attr mod;
+	union {
+		struct ibv_exp_query_intf_params params;
+		struct ibv_exp_cq_init_attr cq;
+		struct ibv_exp_res_domain_init_attr rd;
+		struct ibv_exp_wq_init_attr wq;
+		struct ibv_exp_cq_attr cq_attr;
+	} attr;
+	enum ibv_exp_query_intf_status status;
+	unsigned int mb_len = rte_pktmbuf_data_room_size(mp);
+	unsigned int cqe_n = desc - 1;
+	struct rte_mbuf *(*elts)[desc] = NULL;
+	int ret = 0;
+
+	(void)conf; /* Thresholds configuration (ignored). */
+	/* Enable scattered packets support for this queue if necessary. */
+	assert(mb_len >= RTE_PKTMBUF_HEADROOM);
+	if ((dev->data->dev_conf.rxmode.jumbo_frame) &&
+	    (dev->data->dev_conf.rxmode.max_rx_pkt_len >
+	     (mb_len - RTE_PKTMBUF_HEADROOM))) {
+		unsigned int size =
+			RTE_PKTMBUF_HEADROOM +
+			dev->data->dev_conf.rxmode.max_rx_pkt_len;
+		unsigned int sges_n;
+
+		/*
+		 * Determine the number of SGEs needed for a full packet
+		 * and round it to the next power of two.
+		 */
+		sges_n = log2above((size / mb_len) + !!(size % mb_len));
+		tmpl.rxq.sges_n = sges_n;
+		/* Make sure rxq.sges_n did not overflow. */
+		size = mb_len * (1 << tmpl.rxq.sges_n);
+		size -= RTE_PKTMBUF_HEADROOM;
+		if (size < dev->data->dev_conf.rxmode.max_rx_pkt_len) {
+			ERROR("%p: too many SGEs (%u) needed to handle"
+			      " requested maximum packet size %u",
+			      (void *)dev,
+			      1 << sges_n,
+			      dev->data->dev_conf.rxmode.max_rx_pkt_len);
+			return EOVERFLOW;
+		}
+	}
+	DEBUG("%p: maximum number of segments per packet: %u",
+	      (void *)dev, 1 << tmpl.rxq.sges_n);
+	if (desc % (1 << tmpl.rxq.sges_n)) {
+		ERROR("%p: number of RX queue descriptors (%u) is not a"
+		      " multiple of SGEs per packet (%u)",
+		      (void *)dev,
+		      desc,
+		      1 << tmpl.rxq.sges_n);
+		return EINVAL;
+	}
+	/* Toggle RX checksum offload if hardware supports it. */
+	if (priv->hw_csum)
+		tmpl.rxq.csum = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
+	if (priv->hw_csum_l2tun)
+		tmpl.rxq.csum_l2tun =
+			!!dev->data->dev_conf.rxmode.hw_ip_checksum;
+	/* Use the entire RX mempool as the memory region. */
+	tmpl.mr = mlx5_mp2mr(priv->pd, mp);
+	if (tmpl.mr == NULL) {
+		ret = EINVAL;
+		ERROR("%p: MR creation failure: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	attr.rd = (struct ibv_exp_res_domain_init_attr){
+		.comp_mask = (IBV_EXP_RES_DOMAIN_THREAD_MODEL |
+			      IBV_EXP_RES_DOMAIN_MSG_MODEL),
+		.thread_model = IBV_EXP_THREAD_SINGLE,
+		.msg_model = IBV_EXP_MSG_HIGH_BW,
+	};
+	tmpl.rd = ibv_exp_create_res_domain(priv->ctx, &attr.rd);
+	if (tmpl.rd == NULL) {
+		ret = ENOMEM;
+		ERROR("%p: RD creation failure: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	attr.cq = (struct ibv_exp_cq_init_attr){
+		.comp_mask = IBV_EXP_CQ_INIT_ATTR_RES_DOMAIN,
+		.res_domain = tmpl.rd,
+	};
+	if (priv->cqe_comp) {
+		attr.cq.comp_mask |= IBV_EXP_CQ_INIT_ATTR_FLAGS;
+		attr.cq.flags |= IBV_EXP_CQ_COMPRESSED_CQE;
+		cqe_n = (desc * 2) - 1; /* Double the number of CQEs. */
+	}
+	tmpl.cq = ibv_exp_create_cq(priv->ctx, cqe_n, NULL, NULL, 0,
+				    &attr.cq);
+	if (tmpl.cq == NULL) {
+		ret = ENOMEM;
+		ERROR("%p: CQ creation failure: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	DEBUG("priv->device_attr.max_qp_wr is %d",
+	      priv->device_attr.max_qp_wr);
+	DEBUG("priv->device_attr.max_sge is %d",
+	      priv->device_attr.max_sge);
+	/* Configure VLAN stripping. */
+	tmpl.rxq.vlan_strip = (priv->hw_vlan_strip &&
+			       !!dev->data->dev_conf.rxmode.hw_vlan_strip);
+	attr.wq = (struct ibv_exp_wq_init_attr){
+		.wq_context = NULL, /* Could be useful in the future. */
+		.wq_type = IBV_EXP_WQT_RQ,
+		/* Max number of outstanding WRs. */
+		.max_recv_wr = desc >> tmpl.rxq.sges_n,
+		/* Max number of scatter/gather elements in a WR. */
+		.max_recv_sge = 1 << tmpl.rxq.sges_n,
+		.pd = priv->pd,
+		.cq = tmpl.cq,
+		.comp_mask =
+			IBV_EXP_CREATE_WQ_RES_DOMAIN |
+			IBV_EXP_CREATE_WQ_VLAN_OFFLOADS |
+			0,
+		.res_domain = tmpl.rd,
+		.vlan_offloads = (tmpl.rxq.vlan_strip ?
+				  IBV_EXP_RECEIVE_WQ_CVLAN_STRIP :
+				  0),
+	};
+	/* By default, FCS (CRC) is stripped by hardware. */
+	if (dev->data->dev_conf.rxmode.hw_strip_crc) {
+		tmpl.rxq.crc_present = 0;
+	} else if (priv->hw_fcs_strip) {
+		/* Ask HW/Verbs to leave CRC in place when supported. */
+		attr.wq.flags |= IBV_EXP_CREATE_WQ_FLAG_SCATTER_FCS;
+		attr.wq.comp_mask |= IBV_EXP_CREATE_WQ_FLAGS;
+		tmpl.rxq.crc_present = 1;
+	} else {
+		WARN("%p: CRC stripping has been disabled but will still"
+		     " be performed by hardware, make sure MLNX_OFED and"
+		     " firmware are up to date",
+		     (void *)dev);
+		tmpl.rxq.crc_present = 0;
+	}
+	DEBUG("%p: CRC stripping is %s, %u bytes will be subtracted from"
+	      " incoming frames to hide it",
+	      (void *)dev,
+	      tmpl.rxq.crc_present ? "disabled" : "enabled",
+	      tmpl.rxq.crc_present << 2);
+	if (!mlx5_getenv_int("MLX5_PMD_ENABLE_PADDING"))
+		; /* Nothing else to do. */
+	else if (priv->hw_padding) {
+		INFO("%p: enabling packet padding on queue %p",
+		     (void *)dev, (void *)rxq_ctrl);
+		attr.wq.flags |= IBV_EXP_CREATE_WQ_FLAG_RX_END_PADDING;
+		attr.wq.comp_mask |= IBV_EXP_CREATE_WQ_FLAGS;
+	} else
+		WARN("%p: packet padding has been requested but is not"
+		     " supported, make sure MLNX_OFED and firmware are"
+		     " up to date",
+		     (void *)dev);
+
+	tmpl.wq = ibv_exp_create_wq(priv->ctx, &attr.wq);
+	if (tmpl.wq == NULL) {
+		ret = (errno ? errno : EINVAL);
+		ERROR("%p: WQ creation failure: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	/*
+	 * Make sure number of WRs*SGEs match expectations since a queue
+	 * cannot allocate more than "desc" buffers.
+	 */
+	if (((int)attr.wq.max_recv_wr != (desc >> tmpl.rxq.sges_n)) ||
+	    ((int)attr.wq.max_recv_sge != (1 << tmpl.rxq.sges_n))) {
+		ERROR("%p: requested %u*%u but got %u*%u WRs*SGEs",
+		      (void *)dev,
+		      (desc >> tmpl.rxq.sges_n), (1 << tmpl.rxq.sges_n),
+		      attr.wq.max_recv_wr, attr.wq.max_recv_sge);
+		ret = EINVAL;
+		goto error;
+	}
+	/* Save port ID. */
+	tmpl.rxq.port_id = dev->data->port_id;
+	DEBUG("%p: RTE port ID: %u", (void *)rxq_ctrl, tmpl.rxq.port_id);
+	attr.params = (struct ibv_exp_query_intf_params){
+		.intf_scope = IBV_EXP_INTF_GLOBAL,
+		.intf_version = 1,
+		.intf = IBV_EXP_INTF_CQ,
+		.obj = tmpl.cq,
+	};
+	tmpl.if_cq = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
+	if (tmpl.if_cq == NULL) {
+		ERROR("%p: CQ interface family query failed with status %d",
+		      (void *)dev, status);
+		goto error;
+	}
+	attr.params = (struct ibv_exp_query_intf_params){
+		.intf_scope = IBV_EXP_INTF_GLOBAL,
+		.intf = IBV_EXP_INTF_WQ,
+		.obj = tmpl.wq,
+	};
+	tmpl.if_wq = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
+	if (tmpl.if_wq == NULL) {
+		ERROR("%p: WQ interface family query failed with status %d",
+		      (void *)dev, status);
+		goto error;
+	}
+	/* Change queue state to ready. */
+	mod = (struct ibv_exp_wq_attr){
+		.attr_mask = IBV_EXP_WQ_ATTR_STATE,
+		.wq_state = IBV_EXP_WQS_RDY,
+	};
+	ret = ibv_exp_modify_wq(tmpl.wq, &mod);
+	if (ret) {
+		ERROR("%p: WQ state to IBV_EXP_WQS_RDY failed: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	ret = rxq_setup(&tmpl);
+	if (ret) {
+		ERROR("%p: cannot initialize RX queue structure: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	/* Reuse buffers from original queue if possible. */
+	if (rxq_ctrl->rxq.elts_n) {
+		assert(rxq_ctrl->rxq.elts_n == desc);
+		assert(rxq_ctrl->rxq.elts != tmpl.rxq.elts);
+		ret = rxq_alloc_elts(&tmpl, desc, rxq_ctrl->rxq.elts);
+	} else
+		ret = rxq_alloc_elts(&tmpl, desc, NULL);
+	if (ret) {
+		ERROR("%p: RXQ allocation failed: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	/* Clean up rxq in case we're reinitializing it. */
+	DEBUG("%p: cleaning-up old rxq just in case", (void *)rxq_ctrl);
+	rxq_cleanup(rxq_ctrl);
+	/* Move mbuf pointers to dedicated storage area in RX queue. */
+	elts = (void *)(rxq_ctrl + 1);
+	rte_memcpy(elts, tmpl.rxq.elts, sizeof(*elts));
+#ifndef NDEBUG
+	memset(tmpl.rxq.elts, 0x55, sizeof(*elts));
+#endif
+	rte_free(tmpl.rxq.elts);
+	tmpl.rxq.elts = elts;
+	*rxq_ctrl = tmpl;
+	/* Update doorbell counter. */
+	rxq_ctrl->rxq.rq_ci = desc >> rxq_ctrl->rxq.sges_n;
+	rte_wmb();
+	*rxq_ctrl->rxq.rq_db = htonl(rxq_ctrl->rxq.rq_ci);
+	DEBUG("%p: rxq updated with %p", (void *)rxq_ctrl, (void *)&tmpl);
+	assert(ret == 0);
+	return 0;
+error:
+	elts = tmpl.rxq.elts;
+	rxq_cleanup(&tmpl);
+	rte_free(elts);
+	assert(ret > 0);
+	return ret;
+}
+
+/**
+ * DPDK callback to configure a RX queue.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param idx
+ *   RX queue index.
+ * @param desc
+ *   Number of descriptors to configure in queue.
+ * @param socket
+ *   NUMA socket on which memory must be allocated.
+ * @param[in] conf
+ *   Thresholds parameters.
+ * @param mp
+ *   Memory pool for buffer allocations.
+ *
+ * @return
+ *   0 on success, negative errno value on failure.
+ */
+int
+mlx5_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
+		    unsigned int socket, const struct rte_eth_rxconf *conf,
+		    struct rte_mempool *mp)
+{
+	struct priv *priv = dev->data->dev_private;
+	struct rxq *rxq = (*priv->rxqs)[idx];
+	struct rxq_ctrl *rxq_ctrl = container_of(rxq, struct rxq_ctrl, rxq);
+	int ret;
+
+	if (mlx5_is_secondary())
+		return -E_RTE_SECONDARY;
+
+	priv_lock(priv);
+	if (!rte_is_power_of_2(desc)) {
+		desc = 1 << log2above(desc);
+		WARN("%p: increased number of descriptors in RX queue %u"
+		     " to the next power of two (%d)",
+		     (void *)dev, idx, desc);
+	}
+	DEBUG("%p: configuring queue %u for %u descriptors",
+	      (void *)dev, idx, desc);
+	if (idx >= priv->rxqs_n) {
+		ERROR("%p: queue index out of range (%u >= %u)",
+		      (void *)dev, idx, priv->rxqs_n);
+		priv_unlock(priv);
+		return -EOVERFLOW;
+	}
+	if (rxq != NULL) {
+		DEBUG("%p: reusing already allocated queue index %u (%p)",
+		      (void *)dev, idx, (void *)rxq);
+		if (priv->started) {
+			priv_unlock(priv);
+			return -EEXIST;
+		}
+		(*priv->rxqs)[idx] = NULL;
+		rxq_cleanup(rxq_ctrl);
+	} else {
+		rxq_ctrl = rte_calloc_socket("RXQ", 1, sizeof(*rxq_ctrl) +
+					     desc * sizeof(struct rte_mbuf *),
+					     0, socket);
+		if (rxq_ctrl == NULL) {
+			ERROR("%p: unable to allocate queue index %u",
+			      (void *)dev, idx);
+			priv_unlock(priv);
+			return -ENOMEM;
+		}
+	}
+	ret = rxq_ctrl_setup(dev, rxq_ctrl, desc, socket, conf, mp);
+	if (ret)
+		rte_free(rxq_ctrl);
+	else {
+		rxq_ctrl->rxq.stats.idx = idx;
+		DEBUG("%p: adding RX queue %p to list",
+		      (void *)dev, (void *)rxq_ctrl);
+		(*priv->rxqs)[idx] = &rxq_ctrl->rxq;
+		/* Update receive callback. */
+		dev->rx_pkt_burst = mlx5_rx_burst;
+	}
+	priv_unlock(priv);
+	return -ret;
+}
+
+/**
+ * DPDK callback to release a RX queue.
+ *
+ * @param dpdk_rxq
+ *   Generic RX queue pointer.
+ */
+void
+mlx5_rx_queue_release(void *dpdk_rxq)
+{
+	struct rxq *rxq = (struct rxq *)dpdk_rxq;
+	struct rxq_ctrl *rxq_ctrl;
+	struct priv *priv;
+	unsigned int i;
+
+	if (mlx5_is_secondary())
+		return;
+
+	if (rxq == NULL)
+		return;
+	rxq_ctrl = container_of(rxq, struct rxq_ctrl, rxq);
+	priv = rxq_ctrl->priv;
+	priv_lock(priv);
+	for (i = 0; (i != priv->rxqs_n); ++i)
+		if ((*priv->rxqs)[i] == rxq) {
+			DEBUG("%p: removing RX queue %p from list",
+			      (void *)priv->dev, (void *)rxq_ctrl);
+			(*priv->rxqs)[i] = NULL;
+			break;
+		}
+	rxq_cleanup(rxq_ctrl);
+	rte_free(rxq_ctrl);
+	priv_unlock(priv);
+}
+
+/**
+ * DPDK callback for RX in secondary processes.
+ *
+ * This function configures all queues from primary process information
+ * if necessary before reverting to the normal RX burst callback.
+ *
+ * @param dpdk_rxq
+ *   Generic pointer to RX queue structure.
+ * @param[out] pkts
+ *   Array to store received packets.
+ * @param pkts_n
+ *   Maximum number of packets in array.
+ *
+ * @return
+ *   Number of packets successfully received (<= pkts_n).
+ */
+uint16_t
+mlx5_rx_burst_secondary_setup(void *dpdk_rxq, struct rte_mbuf **pkts,
+			      uint16_t pkts_n)
+{
+	struct rxq *rxq = dpdk_rxq;
+	struct rxq_ctrl *rxq_ctrl = container_of(rxq, struct rxq_ctrl, rxq);
+	struct priv *priv = mlx5_secondary_data_setup(rxq_ctrl->priv);
+	struct priv *primary_priv;
+	unsigned int index;
+
+	if (priv == NULL)
+		return 0;
+	primary_priv =
+		mlx5_secondary_data[priv->dev->data->port_id].primary_priv;
+	/* Look for queue index in both private structures. */
+	for (index = 0; index != priv->rxqs_n; ++index)
+		if (((*primary_priv->rxqs)[index] == rxq) ||
+		    ((*priv->rxqs)[index] == rxq))
+			break;
+	if (index == priv->rxqs_n)
+		return 0;
+	rxq = (*priv->rxqs)[index];
+	return priv->dev->rx_pkt_burst(rxq, pkts, pkts_n);
+}
diff --git a/src/dpdk/drivers/net/mlx5/mlx5_rxtx.c b/src/dpdk/drivers/net/mlx5/mlx5_rxtx.c
new file mode 100644
index 00000000..fce3381a
--- /dev/null
+++ b/src/dpdk/drivers/net/mlx5/mlx5_rxtx.c
@@ -0,0 +1,1722 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2015 6WIND S.A.
+ *   Copyright 2015 Mellanox.
+ *
+ *   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.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <assert.h>
+#include <stdint.h>
+#include <string.h>
+#include <stdlib.h>
+
+/* Verbs header. */
+/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <infiniband/verbs.h>
+#include <infiniband/mlx5_hw.h>
+#include <infiniband/arch.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+/* DPDK headers don't like -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <rte_mbuf.h>
+#include <rte_mempool.h>
+#include <rte_prefetch.h>
+#include <rte_common.h>
+#include <rte_branch_prediction.h>
+#include <rte_ether.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+#include "mlx5.h"
+#include "mlx5_utils.h"
+#include "mlx5_rxtx.h"
+#include "mlx5_autoconf.h"
+#include "mlx5_defs.h"
+#include "mlx5_prm.h"
+
+#ifndef NDEBUG
+
+/**
+ * Verify or set magic value in CQE.
+ *
+ * @param cqe
+ *   Pointer to CQE.
+ *
+ * @return
+ *   0 the first time.
+ */
+static inline int
+check_cqe64_seen(volatile struct mlx5_cqe64 *cqe)
+{
+	static const uint8_t magic[] = "seen";
+	volatile uint8_t (*buf)[sizeof(cqe->rsvd40)] = &cqe->rsvd40;
+	int ret = 1;
+	unsigned int i;
+
+	for (i = 0; i < sizeof(magic) && i < sizeof(*buf); ++i)
+		if (!ret || (*buf)[i] != magic[i]) {
+			ret = 0;
+			(*buf)[i] = magic[i];
+		}
+	return ret;
+}
+
+#endif /* NDEBUG */
+
+static inline int
+check_cqe64(volatile struct mlx5_cqe64 *cqe,
+	    unsigned int cqes_n, const uint16_t ci)
+	    __attribute__((always_inline));
+
+/**
+ * Check whether CQE is valid.
+ *
+ * @param cqe
+ *   Pointer to CQE.
+ * @param cqes_n
+ *   Size of completion queue.
+ * @param ci
+ *   Consumer index.
+ *
+ * @return
+ *   0 on success, 1 on failure.
+ */
+static inline int
+check_cqe64(volatile struct mlx5_cqe64 *cqe,
+		unsigned int cqes_n, const uint16_t ci)
+{
+	uint16_t idx = ci & cqes_n;
+	uint8_t op_own = cqe->op_own;
+	uint8_t op_owner = MLX5_CQE_OWNER(op_own);
+	uint8_t op_code = MLX5_CQE_OPCODE(op_own);
+
+	if (unlikely((op_owner != (!!(idx))) || (op_code == MLX5_CQE_INVALID)))
+		return 1; /* No CQE. */
+#ifndef NDEBUG
+	if ((op_code == MLX5_CQE_RESP_ERR) ||
+	    (op_code == MLX5_CQE_REQ_ERR)) {
+		volatile struct mlx5_err_cqe *err_cqe = (volatile void *)cqe;
+		uint8_t syndrome = err_cqe->syndrome;
+
+		if ((syndrome == MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR) ||
+		    (syndrome == MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR))
+			return 0;
+		if (!check_cqe64_seen(cqe))
+			ERROR("unexpected CQE error %u (0x%02x)"
+			      " syndrome 0x%02x",
+			      op_code, op_code, syndrome);
+		return 1;
+	} else if ((op_code != MLX5_CQE_RESP_SEND) &&
+		   (op_code != MLX5_CQE_REQ)) {
+		if (!check_cqe64_seen(cqe))
+			ERROR("unexpected CQE opcode %u (0x%02x)",
+			      op_code, op_code);
+		return 1;
+	}
+#endif /* NDEBUG */
+	return 0;
+}
+
+/**
+ * Manage TX completions.
+ *
+ * When sending a burst, mlx5_tx_burst() posts several WRs.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ */
+static void
+txq_complete(struct txq *txq)
+{
+	const unsigned int elts_n = txq->elts_n;
+	const unsigned int cqe_n = txq->cqe_n;
+	const unsigned int cqe_cnt = cqe_n - 1;
+	uint16_t elts_free = txq->elts_tail;
+	uint16_t elts_tail;
+	uint16_t cq_ci = txq->cq_ci;
+	volatile struct mlx5_cqe64 *cqe = NULL;
+	volatile union mlx5_wqe *wqe;
+
+	do {
+		volatile struct mlx5_cqe64 *tmp;
+
+		tmp = &(*txq->cqes)[cq_ci & cqe_cnt].cqe64;
+		if (check_cqe64(tmp, cqe_n, cq_ci))
+			break;
+		cqe = tmp;
+#ifndef NDEBUG
+		if (MLX5_CQE_FORMAT(cqe->op_own) == MLX5_COMPRESSED) {
+			if (!check_cqe64_seen(cqe))
+				ERROR("unexpected compressed CQE, TX stopped");
+			return;
+		}
+		if ((MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_RESP_ERR) ||
+		    (MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_REQ_ERR)) {
+			if (!check_cqe64_seen(cqe))
+				ERROR("unexpected error CQE, TX stopped");
+			return;
+		}
+#endif /* NDEBUG */
+		++cq_ci;
+	} while (1);
+	if (unlikely(cqe == NULL))
+		return;
+	wqe = &(*txq->wqes)[htons(cqe->wqe_counter) & (txq->wqe_n - 1)];
+	elts_tail = wqe->wqe.ctrl.data[3];
+	assert(elts_tail < txq->wqe_n);
+	/* Free buffers. */
+	while (elts_free != elts_tail) {
+		struct rte_mbuf *elt = (*txq->elts)[elts_free];
+		unsigned int elts_free_next =
+			(elts_free + 1) & (elts_n - 1);
+		struct rte_mbuf *elt_next = (*txq->elts)[elts_free_next];
+
+#ifndef NDEBUG
+		/* Poisoning. */
+		memset(&(*txq->elts)[elts_free],
+		       0x66,
+		       sizeof((*txq->elts)[elts_free]));
+#endif
+		RTE_MBUF_PREFETCH_TO_FREE(elt_next);
+		/* Only one segment needs to be freed. */
+		rte_pktmbuf_free_seg(elt);
+		elts_free = elts_free_next;
+	}
+	txq->cq_ci = cq_ci;
+	txq->elts_tail = elts_tail;
+	/* Update the consumer index. */
+	rte_wmb();
+	*txq->cq_db = htonl(cq_ci);
+}
+
+/**
+ * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which
+ * the cloned mbuf is allocated is returned instead.
+ *
+ * @param buf
+ *   Pointer to mbuf.
+ *
+ * @return
+ *   Memory pool where data is located for given mbuf.
+ */
+static struct rte_mempool *
+txq_mb2mp(struct rte_mbuf *buf)
+{
+	if (unlikely(RTE_MBUF_INDIRECT(buf)))
+		return rte_mbuf_from_indirect(buf)->pool;
+	return buf->pool;
+}
+
+static inline uint32_t
+txq_mp2mr(struct txq *txq, struct rte_mempool *mp)
+	__attribute__((always_inline));
+
+/**
+ * Get Memory Region (MR) <-> Memory Pool (MP) association from txq->mp2mr[].
+ * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
+ * remove an entry first.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param[in] mp
+ *   Memory Pool for which a Memory Region lkey must be returned.
+ *
+ * @return
+ *   mr->lkey on success, (uint32_t)-1 on failure.
+ */
+static inline uint32_t
+txq_mp2mr(struct txq *txq, struct rte_mempool *mp)
+{
+	unsigned int i;
+	uint32_t lkey = (uint32_t)-1;
+
+	for (i = 0; (i != RTE_DIM(txq->mp2mr)); ++i) {
+		if (unlikely(txq->mp2mr[i].mp == NULL)) {
+			/* Unknown MP, add a new MR for it. */
+			break;
+		}
+		if (txq->mp2mr[i].mp == mp) {
+			assert(txq->mp2mr[i].lkey != (uint32_t)-1);
+			assert(htonl(txq->mp2mr[i].mr->lkey) ==
+			       txq->mp2mr[i].lkey);
+			lkey = txq->mp2mr[i].lkey;
+			break;
+		}
+	}
+	if (unlikely(lkey == (uint32_t)-1))
+		lkey = txq_mp2mr_reg(txq, mp, i);
+	return lkey;
+}
+
+/**
+ * Write a regular WQE.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param wqe
+ *   Pointer to the WQE to fill.
+ * @param addr
+ *   Buffer data address.
+ * @param length
+ *   Packet length.
+ * @param lkey
+ *   Memory region lkey.
+ */
+static inline void
+mlx5_wqe_write(struct txq *txq, volatile union mlx5_wqe *wqe,
+	       uintptr_t addr, uint32_t length, uint32_t lkey)
+{
+	wqe->wqe.ctrl.data[0] = htonl((txq->wqe_ci << 8) | MLX5_OPCODE_SEND);
+	wqe->wqe.ctrl.data[1] = htonl((txq->qp_num_8s) | 4);
+	wqe->wqe.ctrl.data[2] = 0;
+	wqe->wqe.ctrl.data[3] = 0;
+	wqe->inl.eseg.rsvd0 = 0;
+	wqe->inl.eseg.rsvd1 = 0;
+	wqe->inl.eseg.mss = 0;
+	wqe->inl.eseg.rsvd2 = 0;
+	wqe->wqe.eseg.inline_hdr_sz = htons(MLX5_ETH_INLINE_HEADER_SIZE);
+	/* Copy the first 16 bytes into inline header. */
+	rte_memcpy((uint8_t *)(uintptr_t)wqe->wqe.eseg.inline_hdr_start,
+		   (uint8_t *)(uintptr_t)addr,
+		   MLX5_ETH_INLINE_HEADER_SIZE);
+	addr += MLX5_ETH_INLINE_HEADER_SIZE;
+	length -= MLX5_ETH_INLINE_HEADER_SIZE;
+	/* Store remaining data in data segment. */
+	wqe->wqe.dseg.byte_count = htonl(length);
+	wqe->wqe.dseg.lkey = lkey;
+	wqe->wqe.dseg.addr = htonll(addr);
+	/* Increment consumer index. */
+	++txq->wqe_ci;
+}
+
+/**
+ * Write a regular WQE with VLAN.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param wqe
+ *   Pointer to the WQE to fill.
+ * @param addr
+ *   Buffer data address.
+ * @param length
+ *   Packet length.
+ * @param lkey
+ *   Memory region lkey.
+ * @param vlan_tci
+ *   VLAN field to insert in packet.
+ */
+static inline void
+mlx5_wqe_write_vlan(struct txq *txq, volatile union mlx5_wqe *wqe,
+		    uintptr_t addr, uint32_t length, uint32_t lkey,
+		    uint16_t vlan_tci)
+{
+	uint32_t vlan = htonl(0x81000000 | vlan_tci);
+
+	wqe->wqe.ctrl.data[0] = htonl((txq->wqe_ci << 8) | MLX5_OPCODE_SEND);
+	wqe->wqe.ctrl.data[1] = htonl((txq->qp_num_8s) | 4);
+	wqe->wqe.ctrl.data[2] = 0;
+	wqe->wqe.ctrl.data[3] = 0;
+	wqe->inl.eseg.rsvd0 = 0;
+	wqe->inl.eseg.rsvd1 = 0;
+	wqe->inl.eseg.mss = 0;
+	wqe->inl.eseg.rsvd2 = 0;
+	wqe->wqe.eseg.inline_hdr_sz = htons(MLX5_ETH_VLAN_INLINE_HEADER_SIZE);
+	/*
+	 * Copy 12 bytes of source & destination MAC address.
+	 * Copy 4 bytes of VLAN.
+	 * Copy 2 bytes of Ether type.
+	 */
+	rte_memcpy((uint8_t *)(uintptr_t)wqe->wqe.eseg.inline_hdr_start,
+		   (uint8_t *)(uintptr_t)addr, 12);
+	rte_memcpy((uint8_t *)((uintptr_t)wqe->wqe.eseg.inline_hdr_start + 12),
+		   &vlan, sizeof(vlan));
+	rte_memcpy((uint8_t *)((uintptr_t)wqe->wqe.eseg.inline_hdr_start + 16),
+		   (uint8_t *)((uintptr_t)addr + 12), 2);
+	addr += MLX5_ETH_VLAN_INLINE_HEADER_SIZE - sizeof(vlan);
+	length -= MLX5_ETH_VLAN_INLINE_HEADER_SIZE - sizeof(vlan);
+	/* Store remaining data in data segment. */
+	wqe->wqe.dseg.byte_count = htonl(length);
+	wqe->wqe.dseg.lkey = lkey;
+	wqe->wqe.dseg.addr = htonll(addr);
+	/* Increment consumer index. */
+	++txq->wqe_ci;
+}
+
+/**
+ * Write a inline WQE.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param wqe
+ *   Pointer to the WQE to fill.
+ * @param addr
+ *   Buffer data address.
+ * @param length
+ *   Packet length.
+ * @param lkey
+ *   Memory region lkey.
+ */
+static inline void
+mlx5_wqe_write_inline(struct txq *txq, volatile union mlx5_wqe *wqe,
+		      uintptr_t addr, uint32_t length)
+{
+	uint32_t size;
+	uint16_t wqe_cnt = txq->wqe_n - 1;
+	uint16_t wqe_ci = txq->wqe_ci + 1;
+
+	/* Copy the first 16 bytes into inline header. */
+	rte_memcpy((void *)(uintptr_t)wqe->inl.eseg.inline_hdr_start,
+		   (void *)(uintptr_t)addr,
+		   MLX5_ETH_INLINE_HEADER_SIZE);
+	addr += MLX5_ETH_INLINE_HEADER_SIZE;
+	length -= MLX5_ETH_INLINE_HEADER_SIZE;
+	size = 3 + ((4 + length + 15) / 16);
+	wqe->inl.byte_cnt = htonl(length | MLX5_INLINE_SEG);
+	rte_memcpy((void *)(uintptr_t)&wqe->inl.data[0],
+		   (void *)addr, MLX5_WQE64_INL_DATA);
+	addr += MLX5_WQE64_INL_DATA;
+	length -= MLX5_WQE64_INL_DATA;
+	while (length) {
+		volatile union mlx5_wqe *wqe_next =
+			&(*txq->wqes)[wqe_ci & wqe_cnt];
+		uint32_t copy_bytes = (length > sizeof(*wqe)) ?
+				      sizeof(*wqe) :
+				      length;
+
+		rte_mov64((uint8_t *)(uintptr_t)&wqe_next->data[0],
+			  (uint8_t *)addr);
+		addr += copy_bytes;
+		length -= copy_bytes;
+		++wqe_ci;
+	}
+	assert(size < 64);
+	wqe->inl.ctrl.data[0] = htonl((txq->wqe_ci << 8) | MLX5_OPCODE_SEND);
+	wqe->inl.ctrl.data[1] = htonl(txq->qp_num_8s | size);
+	wqe->inl.ctrl.data[2] = 0;
+	wqe->inl.ctrl.data[3] = 0;
+	wqe->inl.eseg.rsvd0 = 0;
+	wqe->inl.eseg.rsvd1 = 0;
+	wqe->inl.eseg.mss = 0;
+	wqe->inl.eseg.rsvd2 = 0;
+	wqe->inl.eseg.inline_hdr_sz = htons(MLX5_ETH_INLINE_HEADER_SIZE);
+	/* Increment consumer index. */
+	txq->wqe_ci = wqe_ci;
+}
+
+/**
+ * Write a inline WQE with VLAN.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param wqe
+ *   Pointer to the WQE to fill.
+ * @param addr
+ *   Buffer data address.
+ * @param length
+ *   Packet length.
+ * @param lkey
+ *   Memory region lkey.
+ * @param vlan_tci
+ *   VLAN field to insert in packet.
+ */
+static inline void
+mlx5_wqe_write_inline_vlan(struct txq *txq, volatile union mlx5_wqe *wqe,
+			   uintptr_t addr, uint32_t length, uint16_t vlan_tci)
+{
+	uint32_t size;
+	uint32_t wqe_cnt = txq->wqe_n - 1;
+	uint16_t wqe_ci = txq->wqe_ci + 1;
+	uint32_t vlan = htonl(0x81000000 | vlan_tci);
+
+	/*
+	 * Copy 12 bytes of source & destination MAC address.
+	 * Copy 4 bytes of VLAN.
+	 * Copy 2 bytes of Ether type.
+	 */
+	rte_memcpy((uint8_t *)(uintptr_t)wqe->inl.eseg.inline_hdr_start,
+		   (uint8_t *)addr, 12);
+	rte_memcpy((uint8_t *)(uintptr_t)wqe->inl.eseg.inline_hdr_start + 12,
+		   &vlan, sizeof(vlan));
+	rte_memcpy((uint8_t *)((uintptr_t)wqe->inl.eseg.inline_hdr_start + 16),
+		   (uint8_t *)(addr + 12), 2);
+	addr += MLX5_ETH_VLAN_INLINE_HEADER_SIZE - sizeof(vlan);
+	length -= MLX5_ETH_VLAN_INLINE_HEADER_SIZE - sizeof(vlan);
+	size = (sizeof(wqe->inl.ctrl.ctrl) +
+		sizeof(wqe->inl.eseg) +
+		sizeof(wqe->inl.byte_cnt) +
+		length + 15) / 16;
+	wqe->inl.byte_cnt = htonl(length | MLX5_INLINE_SEG);
+	rte_memcpy((void *)(uintptr_t)&wqe->inl.data[0],
+		   (void *)addr, MLX5_WQE64_INL_DATA);
+	addr += MLX5_WQE64_INL_DATA;
+	length -= MLX5_WQE64_INL_DATA;
+	while (length) {
+		volatile union mlx5_wqe *wqe_next =
+			&(*txq->wqes)[wqe_ci & wqe_cnt];
+		uint32_t copy_bytes = (length > sizeof(*wqe)) ?
+				      sizeof(*wqe) :
+				      length;
+
+		rte_mov64((uint8_t *)(uintptr_t)&wqe_next->data[0],
+			  (uint8_t *)addr);
+		addr += copy_bytes;
+		length -= copy_bytes;
+		++wqe_ci;
+	}
+	assert(size < 64);
+	wqe->inl.ctrl.data[0] = htonl((txq->wqe_ci << 8) | MLX5_OPCODE_SEND);
+	wqe->inl.ctrl.data[1] = htonl(txq->qp_num_8s | size);
+	wqe->inl.ctrl.data[2] = 0;
+	wqe->inl.ctrl.data[3] = 0;
+	wqe->inl.eseg.rsvd0 = 0;
+	wqe->inl.eseg.rsvd1 = 0;
+	wqe->inl.eseg.mss = 0;
+	wqe->inl.eseg.rsvd2 = 0;
+	wqe->inl.eseg.inline_hdr_sz = htons(MLX5_ETH_VLAN_INLINE_HEADER_SIZE);
+	/* Increment consumer index. */
+	txq->wqe_ci = wqe_ci;
+}
+
+/**
+ * Ring TX queue doorbell.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ */
+static inline void
+mlx5_tx_dbrec(struct txq *txq)
+{
+	uint8_t *dst = (uint8_t *)((uintptr_t)txq->bf_reg + txq->bf_offset);
+	uint32_t data[4] = {
+		htonl((txq->wqe_ci << 8) | MLX5_OPCODE_SEND),
+		htonl(txq->qp_num_8s),
+		0,
+		0,
+	};
+	rte_wmb();
+	*txq->qp_db = htonl(txq->wqe_ci);
+	/* Ensure ordering between DB record and BF copy. */
+	rte_wmb();
+	rte_mov16(dst, (uint8_t *)data);
+	txq->bf_offset ^= txq->bf_buf_size;
+}
+
+/**
+ * Prefetch a CQE.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param cqe_ci
+ *   CQE consumer index.
+ */
+static inline void
+tx_prefetch_cqe(struct txq *txq, uint16_t ci)
+{
+	volatile struct mlx5_cqe64 *cqe;
+
+	cqe = &(*txq->cqes)[ci & (txq->cqe_n - 1)].cqe64;
+	rte_prefetch0(cqe);
+}
+
+/**
+ * Prefetch a WQE.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param  wqe_ci
+ *   WQE consumer index.
+ */
+static inline void
+tx_prefetch_wqe(struct txq *txq, uint16_t ci)
+{
+	volatile union mlx5_wqe *wqe;
+
+	wqe = &(*txq->wqes)[ci & (txq->wqe_n - 1)];
+	rte_prefetch0(wqe);
+}
+
+/**
+ * DPDK callback for TX.
+ *
+ * @param dpdk_txq
+ *   Generic pointer to TX queue structure.
+ * @param[in] pkts
+ *   Packets to transmit.
+ * @param pkts_n
+ *   Number of packets in array.
+ *
+ * @return
+ *   Number of packets successfully transmitted (<= pkts_n).
+ */
+uint16_t
+mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+	struct txq *txq = (struct txq *)dpdk_txq;
+	uint16_t elts_head = txq->elts_head;
+	const unsigned int elts_n = txq->elts_n;
+	unsigned int i = 0;
+	unsigned int j = 0;
+	unsigned int max;
+	unsigned int comp;
+	volatile union mlx5_wqe *wqe = NULL;
+
+	if (unlikely(!pkts_n))
+		return 0;
+	/* Prefetch first packet cacheline. */
+	tx_prefetch_cqe(txq, txq->cq_ci);
+	tx_prefetch_cqe(txq, txq->cq_ci + 1);
+	rte_prefetch0(*pkts);
+	/* Start processing. */
+	txq_complete(txq);
+	max = (elts_n - (elts_head - txq->elts_tail));
+	if (max > elts_n)
+		max -= elts_n;
+	do {
+		struct rte_mbuf *buf = *(pkts++);
+		unsigned int elts_head_next;
+		uintptr_t addr;
+		uint32_t length;
+		uint32_t lkey;
+		unsigned int segs_n = buf->nb_segs;
+		volatile struct mlx5_wqe_data_seg *dseg;
+		unsigned int ds = sizeof(*wqe) / 16;
+
+		/*
+		 * Make sure there is enough room to store this packet and
+		 * that one ring entry remains unused.
+		 */
+		assert(segs_n);
+		if (max < segs_n + 1)
+			break;
+		max -= segs_n;
+		--pkts_n;
+		elts_head_next = (elts_head + 1) & (elts_n - 1);
+		wqe = &(*txq->wqes)[txq->wqe_ci & (txq->wqe_n - 1)];
+		dseg = &wqe->wqe.dseg;
+		rte_prefetch0(wqe);
+		if (pkts_n)
+			rte_prefetch0(*pkts);
+		/* Retrieve buffer information. */
+		addr = rte_pktmbuf_mtod(buf, uintptr_t);
+		length = DATA_LEN(buf);
+		/* Update element. */
+		(*txq->elts)[elts_head] = buf;
+		/* Prefetch next buffer data. */
+		if (pkts_n)
+			rte_prefetch0(rte_pktmbuf_mtod(*pkts,
+						       volatile void *));
+		/* Retrieve Memory Region key for this memory pool. */
+		lkey = txq_mp2mr(txq, txq_mb2mp(buf));
+		if (buf->ol_flags & PKT_TX_VLAN_PKT)
+			mlx5_wqe_write_vlan(txq, wqe, addr, length, lkey,
+					    buf->vlan_tci);
+		else
+			mlx5_wqe_write(txq, wqe, addr, length, lkey);
+		/* Should we enable HW CKSUM offload */
+		if (buf->ol_flags &
+		    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
+			wqe->wqe.eseg.cs_flags =
+				MLX5_ETH_WQE_L3_CSUM |
+				MLX5_ETH_WQE_L4_CSUM;
+		} else {
+			wqe->wqe.eseg.cs_flags = 0;
+		}
+		while (--segs_n) {
+			/*
+			 * Spill on next WQE when the current one does not have
+			 * enough room left. Size of WQE must a be a multiple
+			 * of data segment size.
+			 */
+			assert(!(sizeof(*wqe) % sizeof(*dseg)));
+			if (!(ds % (sizeof(*wqe) / 16)))
+				dseg = (volatile void *)
+					&(*txq->wqes)[txq->wqe_ci++ &
+						      (txq->wqe_n - 1)];
+			else
+				++dseg;
+			++ds;
+			buf = buf->next;
+			assert(buf);
+			/* Store segment information. */
+			dseg->byte_count = htonl(DATA_LEN(buf));
+			dseg->lkey = txq_mp2mr(txq, txq_mb2mp(buf));
+			dseg->addr = htonll(rte_pktmbuf_mtod(buf, uintptr_t));
+			(*txq->elts)[elts_head_next] = buf;
+			elts_head_next = (elts_head_next + 1) & (elts_n - 1);
+#ifdef MLX5_PMD_SOFT_COUNTERS
+			length += DATA_LEN(buf);
+#endif
+			++j;
+		}
+		/* Update DS field in WQE. */
+		wqe->wqe.ctrl.data[1] &= htonl(0xffffffc0);
+		wqe->wqe.ctrl.data[1] |= htonl(ds & 0x3f);
+		elts_head = elts_head_next;
+#ifdef MLX5_PMD_SOFT_COUNTERS
+		/* Increment sent bytes counter. */
+		txq->stats.obytes += length;
+#endif
+		elts_head = elts_head_next;
+		++i;
+	} while (pkts_n);
+	/* Take a shortcut if nothing must be sent. */
+	if (unlikely(i == 0))
+		return 0;
+	/* Check whether completion threshold has been reached. */
+	comp = txq->elts_comp + i + j;
+	if (comp >= MLX5_TX_COMP_THRESH) {
+		/* Request completion on last WQE. */
+		wqe->wqe.ctrl.data[2] = htonl(8);
+		/* Save elts_head in unused "immediate" field of WQE. */
+		wqe->wqe.ctrl.data[3] = elts_head;
+		txq->elts_comp = 0;
+	} else {
+		txq->elts_comp = comp;
+	}
+#ifdef MLX5_PMD_SOFT_COUNTERS
+	/* Increment sent packets counter. */
+	txq->stats.opackets += i;
+#endif
+	/* Ring QP doorbell. */
+	mlx5_tx_dbrec(txq);
+	txq->elts_head = elts_head;
+	return i;
+}
+
+/**
+ * DPDK callback for TX with inline support.
+ *
+ * @param dpdk_txq
+ *   Generic pointer to TX queue structure.
+ * @param[in] pkts
+ *   Packets to transmit.
+ * @param pkts_n
+ *   Number of packets in array.
+ *
+ * @return
+ *   Number of packets successfully transmitted (<= pkts_n).
+ */
+uint16_t
+mlx5_tx_burst_inline(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+	struct txq *txq = (struct txq *)dpdk_txq;
+	uint16_t elts_head = txq->elts_head;
+	const unsigned int elts_n = txq->elts_n;
+	unsigned int i = 0;
+	unsigned int j = 0;
+	unsigned int max;
+	unsigned int comp;
+	volatile union mlx5_wqe *wqe = NULL;
+	unsigned int max_inline = txq->max_inline;
+
+	if (unlikely(!pkts_n))
+		return 0;
+	/* Prefetch first packet cacheline. */
+	tx_prefetch_cqe(txq, txq->cq_ci);
+	tx_prefetch_cqe(txq, txq->cq_ci + 1);
+	rte_prefetch0(*pkts);
+	/* Start processing. */
+	txq_complete(txq);
+	max = (elts_n - (elts_head - txq->elts_tail));
+	if (max > elts_n)
+		max -= elts_n;
+	do {
+		struct rte_mbuf *buf = *(pkts++);
+		unsigned int elts_head_next;
+		uintptr_t addr;
+		uint32_t length;
+		uint32_t lkey;
+		unsigned int segs_n = buf->nb_segs;
+		volatile struct mlx5_wqe_data_seg *dseg;
+		unsigned int ds = sizeof(*wqe) / 16;
+
+		/*
+		 * Make sure there is enough room to store this packet and
+		 * that one ring entry remains unused.
+		 */
+		assert(segs_n);
+		if (max < segs_n + 1)
+			break;
+		max -= segs_n;
+		--pkts_n;
+		elts_head_next = (elts_head + 1) & (elts_n - 1);
+		wqe = &(*txq->wqes)[txq->wqe_ci & (txq->wqe_n - 1)];
+		dseg = &wqe->wqe.dseg;
+		tx_prefetch_wqe(txq, txq->wqe_ci);
+		tx_prefetch_wqe(txq, txq->wqe_ci + 1);
+		if (pkts_n)
+			rte_prefetch0(*pkts);
+		/* Should we enable HW CKSUM offload */
+		if (buf->ol_flags &
+		    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
+			wqe->inl.eseg.cs_flags =
+				MLX5_ETH_WQE_L3_CSUM |
+				MLX5_ETH_WQE_L4_CSUM;
+		} else {
+			wqe->inl.eseg.cs_flags = 0;
+		}
+		/* Retrieve buffer information. */
+		addr = rte_pktmbuf_mtod(buf, uintptr_t);
+		length = DATA_LEN(buf);
+		/* Update element. */
+		(*txq->elts)[elts_head] = buf;
+		/* Prefetch next buffer data. */
+		if (pkts_n)
+			rte_prefetch0(rte_pktmbuf_mtod(*pkts,
+						       volatile void *));
+		if ((length <= max_inline) && (segs_n == 1)) {
+			if (buf->ol_flags & PKT_TX_VLAN_PKT)
+				mlx5_wqe_write_inline_vlan(txq, wqe,
+							   addr, length,
+							   buf->vlan_tci);
+			else
+				mlx5_wqe_write_inline(txq, wqe, addr, length);
+			goto skip_segs;
+		} else {
+			/* Retrieve Memory Region key for this memory pool. */
+			lkey = txq_mp2mr(txq, txq_mb2mp(buf));
+			if (buf->ol_flags & PKT_TX_VLAN_PKT)
+				mlx5_wqe_write_vlan(txq, wqe, addr, length,
+						    lkey, buf->vlan_tci);
+			else
+				mlx5_wqe_write(txq, wqe, addr, length, lkey);
+		}
+		while (--segs_n) {
+			/*
+			 * Spill on next WQE when the current one does not have
+			 * enough room left. Size of WQE must a be a multiple
+			 * of data segment size.
+			 */
+			assert(!(sizeof(*wqe) % sizeof(*dseg)));
+			if (!(ds % (sizeof(*wqe) / 16)))
+				dseg = (volatile void *)
+					&(*txq->wqes)[txq->wqe_ci++ &
+						      (txq->wqe_n - 1)];
+			else
+				++dseg;
+			++ds;
+			buf = buf->next;
+			assert(buf);
+			/* Store segment information. */
+			dseg->byte_count = htonl(DATA_LEN(buf));
+			dseg->lkey = txq_mp2mr(txq, txq_mb2mp(buf));
+			dseg->addr = htonll(rte_pktmbuf_mtod(buf, uintptr_t));
+			(*txq->elts)[elts_head_next] = buf;
+			elts_head_next = (elts_head_next + 1) & (elts_n - 1);
+#ifdef MLX5_PMD_SOFT_COUNTERS
+			length += DATA_LEN(buf);
+#endif
+			++j;
+		}
+		/* Update DS field in WQE. */
+		wqe->inl.ctrl.data[1] &= htonl(0xffffffc0);
+		wqe->inl.ctrl.data[1] |= htonl(ds & 0x3f);
+skip_segs:
+		elts_head = elts_head_next;
+#ifdef MLX5_PMD_SOFT_COUNTERS
+		/* Increment sent bytes counter. */
+		txq->stats.obytes += length;
+#endif
+		++i;
+	} while (pkts_n);
+	/* Take a shortcut if nothing must be sent. */
+	if (unlikely(i == 0))
+		return 0;
+	/* Check whether completion threshold has been reached. */
+	comp = txq->elts_comp + i + j;
+	if (comp >= MLX5_TX_COMP_THRESH) {
+		/* Request completion on last WQE. */
+		wqe->inl.ctrl.data[2] = htonl(8);
+		/* Save elts_head in unused "immediate" field of WQE. */
+		wqe->inl.ctrl.data[3] = elts_head;
+		txq->elts_comp = 0;
+	} else {
+		txq->elts_comp = comp;
+	}
+#ifdef MLX5_PMD_SOFT_COUNTERS
+	/* Increment sent packets counter. */
+	txq->stats.opackets += i;
+#endif
+	/* Ring QP doorbell. */
+	mlx5_tx_dbrec(txq);
+	txq->elts_head = elts_head;
+	return i;
+}
+
+/**
+ * Open a MPW session.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param mpw
+ *   Pointer to MPW session structure.
+ * @param length
+ *   Packet length.
+ */
+static inline void
+mlx5_mpw_new(struct txq *txq, struct mlx5_mpw *mpw, uint32_t length)
+{
+	uint16_t idx = txq->wqe_ci & (txq->wqe_n - 1);
+	volatile struct mlx5_wqe_data_seg (*dseg)[MLX5_MPW_DSEG_MAX] =
+		(volatile struct mlx5_wqe_data_seg (*)[])
+		(uintptr_t)&(*txq->wqes)[(idx + 1) & (txq->wqe_n - 1)];
+
+	mpw->state = MLX5_MPW_STATE_OPENED;
+	mpw->pkts_n = 0;
+	mpw->len = length;
+	mpw->total_len = 0;
+	mpw->wqe = &(*txq->wqes)[idx];
+	mpw->wqe->mpw.eseg.mss = htons(length);
+	mpw->wqe->mpw.eseg.inline_hdr_sz = 0;
+	mpw->wqe->mpw.eseg.rsvd0 = 0;
+	mpw->wqe->mpw.eseg.rsvd1 = 0;
+	mpw->wqe->mpw.eseg.rsvd2 = 0;
+	mpw->wqe->mpw.ctrl.data[0] = htonl((MLX5_OPC_MOD_MPW << 24) |
+					   (txq->wqe_ci << 8) |
+					   MLX5_OPCODE_LSO_MPW);
+	mpw->wqe->mpw.ctrl.data[2] = 0;
+	mpw->wqe->mpw.ctrl.data[3] = 0;
+	mpw->data.dseg[0] = &mpw->wqe->mpw.dseg[0];
+	mpw->data.dseg[1] = &mpw->wqe->mpw.dseg[1];
+	mpw->data.dseg[2] = &(*dseg)[0];
+	mpw->data.dseg[3] = &(*dseg)[1];
+	mpw->data.dseg[4] = &(*dseg)[2];
+}
+
+/**
+ * Close a MPW session.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param mpw
+ *   Pointer to MPW session structure.
+ */
+static inline void
+mlx5_mpw_close(struct txq *txq, struct mlx5_mpw *mpw)
+{
+	unsigned int num = mpw->pkts_n;
+
+	/*
+	 * Store size in multiple of 16 bytes. Control and Ethernet segments
+	 * count as 2.
+	 */
+	mpw->wqe->mpw.ctrl.data[1] = htonl(txq->qp_num_8s | (2 + num));
+	mpw->state = MLX5_MPW_STATE_CLOSED;
+	if (num < 3)
+		++txq->wqe_ci;
+	else
+		txq->wqe_ci += 2;
+	tx_prefetch_wqe(txq, txq->wqe_ci);
+	tx_prefetch_wqe(txq, txq->wqe_ci + 1);
+}
+
+/**
+ * DPDK callback for TX with MPW support.
+ *
+ * @param dpdk_txq
+ *   Generic pointer to TX queue structure.
+ * @param[in] pkts
+ *   Packets to transmit.
+ * @param pkts_n
+ *   Number of packets in array.
+ *
+ * @return
+ *   Number of packets successfully transmitted (<= pkts_n).
+ */
+uint16_t
+mlx5_tx_burst_mpw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+	struct txq *txq = (struct txq *)dpdk_txq;
+	uint16_t elts_head = txq->elts_head;
+	const unsigned int elts_n = txq->elts_n;
+	unsigned int i = 0;
+	unsigned int j = 0;
+	unsigned int max;
+	unsigned int comp;
+	struct mlx5_mpw mpw = {
+		.state = MLX5_MPW_STATE_CLOSED,
+	};
+
+	if (unlikely(!pkts_n))
+		return 0;
+	/* Prefetch first packet cacheline. */
+	tx_prefetch_cqe(txq, txq->cq_ci);
+	tx_prefetch_wqe(txq, txq->wqe_ci);
+	tx_prefetch_wqe(txq, txq->wqe_ci + 1);
+	/* Start processing. */
+	txq_complete(txq);
+	max = (elts_n - (elts_head - txq->elts_tail));
+	if (max > elts_n)
+		max -= elts_n;
+	do {
+		struct rte_mbuf *buf = *(pkts++);
+		unsigned int elts_head_next;
+		uint32_t length;
+		unsigned int segs_n = buf->nb_segs;
+		uint32_t cs_flags = 0;
+
+		/*
+		 * Make sure there is enough room to store this packet and
+		 * that one ring entry remains unused.
+		 */
+		assert(segs_n);
+		if (max < segs_n + 1)
+			break;
+		/* Do not bother with large packets MPW cannot handle. */
+		if (segs_n > MLX5_MPW_DSEG_MAX)
+			break;
+		max -= segs_n;
+		--pkts_n;
+		/* Should we enable HW CKSUM offload */
+		if (buf->ol_flags &
+		    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM))
+			cs_flags = MLX5_ETH_WQE_L3_CSUM | MLX5_ETH_WQE_L4_CSUM;
+		/* Retrieve packet information. */
+		length = PKT_LEN(buf);
+		assert(length);
+		/* Start new session if packet differs. */
+		if ((mpw.state == MLX5_MPW_STATE_OPENED) &&
+		    ((mpw.len != length) ||
+		     (segs_n != 1) ||
+		     (mpw.wqe->mpw.eseg.cs_flags != cs_flags)))
+			mlx5_mpw_close(txq, &mpw);
+		if (mpw.state == MLX5_MPW_STATE_CLOSED) {
+			mlx5_mpw_new(txq, &mpw, length);
+			mpw.wqe->mpw.eseg.cs_flags = cs_flags;
+		}
+		/* Multi-segment packets must be alone in their MPW. */
+		assert((segs_n == 1) || (mpw.pkts_n == 0));
+#if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
+		length = 0;
+#endif
+		do {
+			volatile struct mlx5_wqe_data_seg *dseg;
+			uintptr_t addr;
+
+			elts_head_next = (elts_head + 1) & (elts_n - 1);
+			assert(buf);
+			(*txq->elts)[elts_head] = buf;
+			dseg = mpw.data.dseg[mpw.pkts_n];
+			addr = rte_pktmbuf_mtod(buf, uintptr_t);
+			*dseg = (struct mlx5_wqe_data_seg){
+				.byte_count = htonl(DATA_LEN(buf)),
+				.lkey = txq_mp2mr(txq, txq_mb2mp(buf)),
+				.addr = htonll(addr),
+			};
+			elts_head = elts_head_next;
+#if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
+			length += DATA_LEN(buf);
+#endif
+			buf = buf->next;
+			++mpw.pkts_n;
+			++j;
+		} while (--segs_n);
+		assert(length == mpw.len);
+		if (mpw.pkts_n == MLX5_MPW_DSEG_MAX)
+			mlx5_mpw_close(txq, &mpw);
+		elts_head = elts_head_next;
+#ifdef MLX5_PMD_SOFT_COUNTERS
+		/* Increment sent bytes counter. */
+		txq->stats.obytes += length;
+#endif
+		++i;
+	} while (pkts_n);
+	/* Take a shortcut if nothing must be sent. */
+	if (unlikely(i == 0))
+		return 0;
+	/* Check whether completion threshold has been reached. */
+	/* "j" includes both packets and segments. */
+	comp = txq->elts_comp + j;
+	if (comp >= MLX5_TX_COMP_THRESH) {
+		volatile union mlx5_wqe *wqe = mpw.wqe;
+
+		/* Request completion on last WQE. */
+		wqe->mpw.ctrl.data[2] = htonl(8);
+		/* Save elts_head in unused "immediate" field of WQE. */
+		wqe->mpw.ctrl.data[3] = elts_head;
+		txq->elts_comp = 0;
+	} else {
+		txq->elts_comp = comp;
+	}
+#ifdef MLX5_PMD_SOFT_COUNTERS
+	/* Increment sent packets counter. */
+	txq->stats.opackets += i;
+#endif
+	/* Ring QP doorbell. */
+	if (mpw.state == MLX5_MPW_STATE_OPENED)
+		mlx5_mpw_close(txq, &mpw);
+	mlx5_tx_dbrec(txq);
+	txq->elts_head = elts_head;
+	return i;
+}
+
+/**
+ * Open a MPW inline session.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param mpw
+ *   Pointer to MPW session structure.
+ * @param length
+ *   Packet length.
+ */
+static inline void
+mlx5_mpw_inline_new(struct txq *txq, struct mlx5_mpw *mpw, uint32_t length)
+{
+	uint16_t idx = txq->wqe_ci & (txq->wqe_n - 1);
+
+	mpw->state = MLX5_MPW_INL_STATE_OPENED;
+	mpw->pkts_n = 0;
+	mpw->len = length;
+	mpw->total_len = 0;
+	mpw->wqe = &(*txq->wqes)[idx];
+	mpw->wqe->mpw_inl.ctrl.data[0] = htonl((MLX5_OPC_MOD_MPW << 24) |
+					       (txq->wqe_ci << 8) |
+					       MLX5_OPCODE_LSO_MPW);
+	mpw->wqe->mpw_inl.ctrl.data[2] = 0;
+	mpw->wqe->mpw_inl.ctrl.data[3] = 0;
+	mpw->wqe->mpw_inl.eseg.mss = htons(length);
+	mpw->wqe->mpw_inl.eseg.inline_hdr_sz = 0;
+	mpw->wqe->mpw_inl.eseg.cs_flags = 0;
+	mpw->wqe->mpw_inl.eseg.rsvd0 = 0;
+	mpw->wqe->mpw_inl.eseg.rsvd1 = 0;
+	mpw->wqe->mpw_inl.eseg.rsvd2 = 0;
+	mpw->data.raw = &mpw->wqe->mpw_inl.data[0];
+}
+
+/**
+ * Close a MPW inline session.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param mpw
+ *   Pointer to MPW session structure.
+ */
+static inline void
+mlx5_mpw_inline_close(struct txq *txq, struct mlx5_mpw *mpw)
+{
+	unsigned int size;
+
+	size = sizeof(*mpw->wqe) - MLX5_MWQE64_INL_DATA + mpw->total_len;
+	/*
+	 * Store size in multiple of 16 bytes. Control and Ethernet segments
+	 * count as 2.
+	 */
+	mpw->wqe->mpw_inl.ctrl.data[1] =
+		htonl(txq->qp_num_8s | ((size + 15) / 16));
+	mpw->state = MLX5_MPW_STATE_CLOSED;
+	mpw->wqe->mpw_inl.byte_cnt = htonl(mpw->total_len | MLX5_INLINE_SEG);
+	txq->wqe_ci += (size + (sizeof(*mpw->wqe) - 1)) / sizeof(*mpw->wqe);
+}
+
+/**
+ * DPDK callback for TX with MPW inline support.
+ *
+ * @param dpdk_txq
+ *   Generic pointer to TX queue structure.
+ * @param[in] pkts
+ *   Packets to transmit.
+ * @param pkts_n
+ *   Number of packets in array.
+ *
+ * @return
+ *   Number of packets successfully transmitted (<= pkts_n).
+ */
+uint16_t
+mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
+			 uint16_t pkts_n)
+{
+	struct txq *txq = (struct txq *)dpdk_txq;
+	uint16_t elts_head = txq->elts_head;
+	const unsigned int elts_n = txq->elts_n;
+	unsigned int i = 0;
+	unsigned int j = 0;
+	unsigned int max;
+	unsigned int comp;
+	unsigned int inline_room = txq->max_inline;
+	struct mlx5_mpw mpw = {
+		.state = MLX5_MPW_STATE_CLOSED,
+	};
+
+	if (unlikely(!pkts_n))
+		return 0;
+	/* Prefetch first packet cacheline. */
+	tx_prefetch_cqe(txq, txq->cq_ci);
+	tx_prefetch_wqe(txq, txq->wqe_ci);
+	tx_prefetch_wqe(txq, txq->wqe_ci + 1);
+	/* Start processing. */
+	txq_complete(txq);
+	max = (elts_n - (elts_head - txq->elts_tail));
+	if (max > elts_n)
+		max -= elts_n;
+	do {
+		struct rte_mbuf *buf = *(pkts++);
+		unsigned int elts_head_next;
+		uintptr_t addr;
+		uint32_t length;
+		unsigned int segs_n = buf->nb_segs;
+		uint32_t cs_flags = 0;
+
+		/*
+		 * Make sure there is enough room to store this packet and
+		 * that one ring entry remains unused.
+		 */
+		assert(segs_n);
+		if (max < segs_n + 1)
+			break;
+		/* Do not bother with large packets MPW cannot handle. */
+		if (segs_n > MLX5_MPW_DSEG_MAX)
+			break;
+		max -= segs_n;
+		--pkts_n;
+		/* Should we enable HW CKSUM offload */
+		if (buf->ol_flags &
+		    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM))
+			cs_flags = MLX5_ETH_WQE_L3_CSUM | MLX5_ETH_WQE_L4_CSUM;
+		/* Retrieve packet information. */
+		length = PKT_LEN(buf);
+		/* Start new session if packet differs. */
+		if (mpw.state == MLX5_MPW_STATE_OPENED) {
+			if ((mpw.len != length) ||
+			    (segs_n != 1) ||
+			    (mpw.wqe->mpw.eseg.cs_flags != cs_flags))
+				mlx5_mpw_close(txq, &mpw);
+		} else if (mpw.state == MLX5_MPW_INL_STATE_OPENED) {
+			if ((mpw.len != length) ||
+			    (segs_n != 1) ||
+			    (length > inline_room) ||
+			    (mpw.wqe->mpw_inl.eseg.cs_flags != cs_flags)) {
+				mlx5_mpw_inline_close(txq, &mpw);
+				inline_room = txq->max_inline;
+			}
+		}
+		if (mpw.state == MLX5_MPW_STATE_CLOSED) {
+			if ((segs_n != 1) ||
+			    (length > inline_room)) {
+				mlx5_mpw_new(txq, &mpw, length);
+				mpw.wqe->mpw.eseg.cs_flags = cs_flags;
+			} else {
+				mlx5_mpw_inline_new(txq, &mpw, length);
+				mpw.wqe->mpw_inl.eseg.cs_flags = cs_flags;
+			}
+		}
+		/* Multi-segment packets must be alone in their MPW. */
+		assert((segs_n == 1) || (mpw.pkts_n == 0));
+		if (mpw.state == MLX5_MPW_STATE_OPENED) {
+			assert(inline_room == txq->max_inline);
+#if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
+			length = 0;
+#endif
+			do {
+				volatile struct mlx5_wqe_data_seg *dseg;
+
+				elts_head_next =
+					(elts_head + 1) & (elts_n - 1);
+				assert(buf);
+				(*txq->elts)[elts_head] = buf;
+				dseg = mpw.data.dseg[mpw.pkts_n];
+				addr = rte_pktmbuf_mtod(buf, uintptr_t);
+				*dseg = (struct mlx5_wqe_data_seg){
+					.byte_count = htonl(DATA_LEN(buf)),
+					.lkey = txq_mp2mr(txq, txq_mb2mp(buf)),
+					.addr = htonll(addr),
+				};
+				elts_head = elts_head_next;
+#if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
+				length += DATA_LEN(buf);
+#endif
+				buf = buf->next;
+				++mpw.pkts_n;
+				++j;
+			} while (--segs_n);
+			assert(length == mpw.len);
+			if (mpw.pkts_n == MLX5_MPW_DSEG_MAX)
+				mlx5_mpw_close(txq, &mpw);
+		} else {
+			unsigned int max;
+
+			assert(mpw.state == MLX5_MPW_INL_STATE_OPENED);
+			assert(length <= inline_room);
+			assert(length == DATA_LEN(buf));
+			elts_head_next = (elts_head + 1) & (elts_n - 1);
+			addr = rte_pktmbuf_mtod(buf, uintptr_t);
+			(*txq->elts)[elts_head] = buf;
+			/* Maximum number of bytes before wrapping. */
+			max = ((uintptr_t)&(*txq->wqes)[txq->wqe_n] -
+			       (uintptr_t)mpw.data.raw);
+			if (length > max) {
+				rte_memcpy((void *)(uintptr_t)mpw.data.raw,
+					   (void *)addr,
+					   max);
+				mpw.data.raw =
+					(volatile void *)&(*txq->wqes)[0];
+				rte_memcpy((void *)(uintptr_t)mpw.data.raw,
+					   (void *)(addr + max),
+					   length - max);
+				mpw.data.raw += length - max;
+			} else {
+				rte_memcpy((void *)(uintptr_t)mpw.data.raw,
+					   (void *)addr,
+					   length);
+				mpw.data.raw += length;
+			}
+			if ((uintptr_t)mpw.data.raw ==
+			    (uintptr_t)&(*txq->wqes)[txq->wqe_n])
+				mpw.data.raw =
+					(volatile void *)&(*txq->wqes)[0];
+			++mpw.pkts_n;
+			++j;
+			if (mpw.pkts_n == MLX5_MPW_DSEG_MAX) {
+				mlx5_mpw_inline_close(txq, &mpw);
+				inline_room = txq->max_inline;
+			} else {
+				inline_room -= length;
+			}
+		}
+		mpw.total_len += length;
+		elts_head = elts_head_next;
+#ifdef MLX5_PMD_SOFT_COUNTERS
+		/* Increment sent bytes counter. */
+		txq->stats.obytes += length;
+#endif
+		++i;
+	} while (pkts_n);
+	/* Take a shortcut if nothing must be sent. */
+	if (unlikely(i == 0))
+		return 0;
+	/* Check whether completion threshold has been reached. */
+	/* "j" includes both packets and segments. */
+	comp = txq->elts_comp + j;
+	if (comp >= MLX5_TX_COMP_THRESH) {
+		volatile union mlx5_wqe *wqe = mpw.wqe;
+
+		/* Request completion on last WQE. */
+		wqe->mpw_inl.ctrl.data[2] = htonl(8);
+		/* Save elts_head in unused "immediate" field of WQE. */
+		wqe->mpw_inl.ctrl.data[3] = elts_head;
+		txq->elts_comp = 0;
+	} else {
+		txq->elts_comp = comp;
+	}
+#ifdef MLX5_PMD_SOFT_COUNTERS
+	/* Increment sent packets counter. */
+	txq->stats.opackets += i;
+#endif
+	/* Ring QP doorbell. */
+	if (mpw.state == MLX5_MPW_INL_STATE_OPENED)
+		mlx5_mpw_inline_close(txq, &mpw);
+	else if (mpw.state == MLX5_MPW_STATE_OPENED)
+		mlx5_mpw_close(txq, &mpw);
+	mlx5_tx_dbrec(txq);
+	txq->elts_head = elts_head;
+	return i;
+}
+
+/**
+ * Translate RX completion flags to packet type.
+ *
+ * @param[in] cqe
+ *   Pointer to CQE.
+ *
+ * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
+ *
+ * @return
+ *   Packet type for struct rte_mbuf.
+ */
+static inline uint32_t
+rxq_cq_to_pkt_type(volatile struct mlx5_cqe64 *cqe)
+{
+	uint32_t pkt_type;
+	uint8_t flags = cqe->l4_hdr_type_etc;
+	uint8_t info = cqe->rsvd0[0];
+
+	if (info & IBV_EXP_CQ_RX_TUNNEL_PACKET)
+		pkt_type =
+			TRANSPOSE(flags,
+				  IBV_EXP_CQ_RX_OUTER_IPV4_PACKET,
+				  RTE_PTYPE_L3_IPV4) |
+			TRANSPOSE(flags,
+				  IBV_EXP_CQ_RX_OUTER_IPV6_PACKET,
+				  RTE_PTYPE_L3_IPV6) |
+			TRANSPOSE(flags,
+				  IBV_EXP_CQ_RX_IPV4_PACKET,
+				  RTE_PTYPE_INNER_L3_IPV4) |
+			TRANSPOSE(flags,
+				  IBV_EXP_CQ_RX_IPV6_PACKET,
+				  RTE_PTYPE_INNER_L3_IPV6);
+	else
+		pkt_type =
+			TRANSPOSE(flags,
+				  MLX5_CQE_L3_HDR_TYPE_IPV6,
+				  RTE_PTYPE_L3_IPV6) |
+			TRANSPOSE(flags,
+				  MLX5_CQE_L3_HDR_TYPE_IPV4,
+				  RTE_PTYPE_L3_IPV4);
+	return pkt_type;
+}
+
+/**
+ * Get size of the next packet for a given CQE. For compressed CQEs, the
+ * consumer index is updated only once all packets of the current one have
+ * been processed.
+ *
+ * @param rxq
+ *   Pointer to RX queue.
+ * @param cqe
+ *   CQE to process.
+ *
+ * @return
+ *   Packet size in bytes (0 if there is none), -1 in case of completion
+ *   with error.
+ */
+static inline int
+mlx5_rx_poll_len(struct rxq *rxq, volatile struct mlx5_cqe64 *cqe,
+		 uint16_t cqe_cnt)
+{
+	struct rxq_zip *zip = &rxq->zip;
+	uint16_t cqe_n = cqe_cnt + 1;
+	int len = 0;
+
+	/* Process compressed data in the CQE and mini arrays. */
+	if (zip->ai) {
+		volatile struct mlx5_mini_cqe8 (*mc)[8] =
+			(volatile struct mlx5_mini_cqe8 (*)[8])
+			(uintptr_t)(&(*rxq->cqes)[zip->ca & cqe_cnt].cqe64);
+
+		len = ntohl((*mc)[zip->ai & 7].byte_cnt);
+		if ((++zip->ai & 7) == 0) {
+			/*
+			 * Increment consumer index to skip the number of
+			 * CQEs consumed. Hardware leaves holes in the CQ
+			 * ring for software use.
+			 */
+			zip->ca = zip->na;
+			zip->na += 8;
+		}
+		if (unlikely(rxq->zip.ai == rxq->zip.cqe_cnt)) {
+			uint16_t idx = rxq->cq_ci;
+			uint16_t end = zip->cq_ci;
+
+			while (idx != end) {
+				(*rxq->cqes)[idx & cqe_cnt].cqe64.op_own =
+					MLX5_CQE_INVALIDATE;
+				++idx;
+			}
+			rxq->cq_ci = zip->cq_ci;
+			zip->ai = 0;
+		}
+	/* No compressed data, get next CQE and verify if it is compressed. */
+	} else {
+		int ret;
+		int8_t op_own;
+
+		ret = check_cqe64(cqe, cqe_n, rxq->cq_ci);
+		if (unlikely(ret == 1))
+			return 0;
+		++rxq->cq_ci;
+		op_own = cqe->op_own;
+		if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED) {
+			volatile struct mlx5_mini_cqe8 (*mc)[8] =
+				(volatile struct mlx5_mini_cqe8 (*)[8])
+				(uintptr_t)(&(*rxq->cqes)[rxq->cq_ci &
+							  cqe_cnt].cqe64);
+
+			/* Fix endianness. */
+			zip->cqe_cnt = ntohl(cqe->byte_cnt);
+			/*
+			 * Current mini array position is the one returned by
+			 * check_cqe64().
+			 *
+			 * If completion comprises several mini arrays, as a
+			 * special case the second one is located 7 CQEs after
+			 * the initial CQE instead of 8 for subsequent ones.
+			 */
+			zip->ca = rxq->cq_ci & cqe_cnt;
+			zip->na = zip->ca + 7;
+			/* Compute the next non compressed CQE. */
+			--rxq->cq_ci;
+			zip->cq_ci = rxq->cq_ci + zip->cqe_cnt;
+			/* Get packet size to return. */
+			len = ntohl((*mc)[0].byte_cnt);
+			zip->ai = 1;
+		} else {
+			len = ntohl(cqe->byte_cnt);
+		}
+		/* Error while receiving packet. */
+		if (unlikely(MLX5_CQE_OPCODE(op_own) == MLX5_CQE_RESP_ERR))
+			return -1;
+	}
+	return len;
+}
+
+/**
+ * Translate RX completion flags to offload flags.
+ *
+ * @param[in] rxq
+ *   Pointer to RX queue structure.
+ * @param[in] cqe
+ *   Pointer to CQE.
+ *
+ * @return
+ *   Offload flags (ol_flags) for struct rte_mbuf.
+ */
+static inline uint32_t
+rxq_cq_to_ol_flags(struct rxq *rxq, volatile struct mlx5_cqe64 *cqe)
+{
+	uint32_t ol_flags = 0;
+	uint8_t l3_hdr = (cqe->l4_hdr_type_etc) & MLX5_CQE_L3_HDR_TYPE_MASK;
+	uint8_t l4_hdr = (cqe->l4_hdr_type_etc) & MLX5_CQE_L4_HDR_TYPE_MASK;
+	uint8_t info = cqe->rsvd0[0];
+
+	if ((l3_hdr == MLX5_CQE_L3_HDR_TYPE_IPV4) ||
+	    (l3_hdr == MLX5_CQE_L3_HDR_TYPE_IPV6))
+		ol_flags |=
+			(!(cqe->hds_ip_ext & MLX5_CQE_L3_OK) *
+			 PKT_RX_IP_CKSUM_BAD);
+	if ((l4_hdr == MLX5_CQE_L4_HDR_TYPE_TCP) ||
+	    (l4_hdr == MLX5_CQE_L4_HDR_TYPE_TCP_EMP_ACK) ||
+	    (l4_hdr == MLX5_CQE_L4_HDR_TYPE_TCP_ACK) ||
+	    (l4_hdr == MLX5_CQE_L4_HDR_TYPE_UDP))
+		ol_flags |=
+			(!(cqe->hds_ip_ext & MLX5_CQE_L4_OK) *
+			 PKT_RX_L4_CKSUM_BAD);
+	/*
+	 * PKT_RX_IP_CKSUM_BAD and PKT_RX_L4_CKSUM_BAD are used in place
+	 * of PKT_RX_EIP_CKSUM_BAD because the latter is not functional
+	 * (its value is 0).
+	 */
+	if ((info & IBV_EXP_CQ_RX_TUNNEL_PACKET) && (rxq->csum_l2tun))
+		ol_flags |=
+			TRANSPOSE(~cqe->l4_hdr_type_etc,
+				  IBV_EXP_CQ_RX_OUTER_IP_CSUM_OK,
+				  PKT_RX_IP_CKSUM_BAD) |
+			TRANSPOSE(~cqe->l4_hdr_type_etc,
+				  IBV_EXP_CQ_RX_OUTER_TCP_UDP_CSUM_OK,
+				  PKT_RX_L4_CKSUM_BAD);
+	return ol_flags;
+}
+
+/**
+ * DPDK callback for RX.
+ *
+ * @param dpdk_rxq
+ *   Generic pointer to RX queue structure.
+ * @param[out] pkts
+ *   Array to store received packets.
+ * @param pkts_n
+ *   Maximum number of packets in array.
+ *
+ * @return
+ *   Number of packets successfully received (<= pkts_n).
+ */
+uint16_t
+mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+	struct rxq *rxq = dpdk_rxq;
+	const unsigned int wqe_cnt = rxq->elts_n - 1;
+	const unsigned int cqe_cnt = rxq->cqe_n - 1;
+	const unsigned int sges_n = rxq->sges_n;
+	struct rte_mbuf *pkt = NULL;
+	struct rte_mbuf *seg = NULL;
+	volatile struct mlx5_cqe64 *cqe =
+		&(*rxq->cqes)[rxq->cq_ci & cqe_cnt].cqe64;
+	unsigned int i = 0;
+	unsigned int rq_ci = rxq->rq_ci << sges_n;
+	int len;
+
+	while (pkts_n) {
+		unsigned int idx = rq_ci & wqe_cnt;
+		volatile struct mlx5_wqe_data_seg *wqe = &(*rxq->wqes)[idx];
+		struct rte_mbuf *rep = (*rxq->elts)[idx];
+
+		if (pkt)
+			NEXT(seg) = rep;
+		seg = rep;
+		rte_prefetch0(seg);
+		rte_prefetch0(cqe);
+		rte_prefetch0(wqe);
+		rep = rte_mbuf_raw_alloc(rxq->mp);
+		if (unlikely(rep == NULL)) {
+			while (pkt != seg) {
+				assert(pkt != (*rxq->elts)[idx]);
+				seg = NEXT(pkt);
+				rte_mbuf_refcnt_set(pkt, 0);
+				__rte_mbuf_raw_free(pkt);
+				pkt = seg;
+			}
+			++rxq->stats.rx_nombuf;
+			break;
+		}
+		if (!pkt) {
+			cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt].cqe64;
+			len = mlx5_rx_poll_len(rxq, cqe, cqe_cnt);
+			if (len == 0) {
+				rte_mbuf_refcnt_set(rep, 0);
+				__rte_mbuf_raw_free(rep);
+				break;
+			}
+			if (unlikely(len == -1)) {
+				/* RX error, packet is likely too large. */
+				rte_mbuf_refcnt_set(rep, 0);
+				__rte_mbuf_raw_free(rep);
+				++rxq->stats.idropped;
+				goto skip;
+			}
+			pkt = seg;
+			assert(len >= (rxq->crc_present << 2));
+			/* Update packet information. */
+			pkt->packet_type = 0;
+			pkt->ol_flags = 0;
+			if (rxq->csum | rxq->csum_l2tun | rxq->vlan_strip |
+			    rxq->crc_present) {
+				if (rxq->csum) {
+					pkt->packet_type =
+						rxq_cq_to_pkt_type(cqe);
+					pkt->ol_flags =
+						rxq_cq_to_ol_flags(rxq, cqe);
+				}
+				if (cqe->l4_hdr_type_etc &
+				    MLX5_CQE_VLAN_STRIPPED) {
+					pkt->ol_flags |= PKT_RX_VLAN_PKT |
+						PKT_RX_VLAN_STRIPPED;
+					pkt->vlan_tci = ntohs(cqe->vlan_info);
+				}
+				if (rxq->crc_present)
+					len -= ETHER_CRC_LEN;
+			}
+			PKT_LEN(pkt) = len;
+		}
+		DATA_LEN(rep) = DATA_LEN(seg);
+		PKT_LEN(rep) = PKT_LEN(seg);
+		SET_DATA_OFF(rep, DATA_OFF(seg));
+		NB_SEGS(rep) = NB_SEGS(seg);
+		PORT(rep) = PORT(seg);
+		NEXT(rep) = NULL;
+		(*rxq->elts)[idx] = rep;
+		/*
+		 * Fill NIC descriptor with the new buffer.  The lkey and size
+		 * of the buffers are already known, only the buffer address
+		 * changes.
+		 */
+		wqe->addr = htonll(rte_pktmbuf_mtod(rep, uintptr_t));
+		if (len > DATA_LEN(seg)) {
+			len -= DATA_LEN(seg);
+			++NB_SEGS(pkt);
+			++rq_ci;
+			continue;
+		}
+		DATA_LEN(seg) = len;
+#ifdef MLX5_PMD_SOFT_COUNTERS
+		/* Increment bytes counter. */
+		rxq->stats.ibytes += PKT_LEN(pkt);
+#endif
+		/* Return packet. */
+		*(pkts++) = pkt;
+		pkt = NULL;
+		--pkts_n;
+		++i;
+skip:
+		/* Align consumer index to the next stride. */
+		rq_ci >>= sges_n;
+		++rq_ci;
+		rq_ci <<= sges_n;
+	}
+	if (unlikely((i == 0) && ((rq_ci >> sges_n) == rxq->rq_ci)))
+		return 0;
+	/* Update the consumer index. */
+	rxq->rq_ci = rq_ci >> sges_n;
+	rte_wmb();
+	*rxq->cq_db = htonl(rxq->cq_ci);
+	rte_wmb();
+	*rxq->rq_db = htonl(rxq->rq_ci);
+#ifdef MLX5_PMD_SOFT_COUNTERS
+	/* Increment packets counter. */
+	rxq->stats.ipackets += i;
+#endif
+	return i;
+}
+
+/**
+ * Dummy DPDK callback for TX.
+ *
+ * This function is used to temporarily replace the real callback during
+ * unsafe control operations on the queue, or in case of error.
+ *
+ * @param dpdk_txq
+ *   Generic pointer to TX queue structure.
+ * @param[in] pkts
+ *   Packets to transmit.
+ * @param pkts_n
+ *   Number of packets in array.
+ *
+ * @return
+ *   Number of packets successfully transmitted (<= pkts_n).
+ */
+uint16_t
+removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+	(void)dpdk_txq;
+	(void)pkts;
+	(void)pkts_n;
+	return 0;
+}
+
+/**
+ * Dummy DPDK callback for RX.
+ *
+ * This function is used to temporarily replace the real callback during
+ * unsafe control operations on the queue, or in case of error.
+ *
+ * @param dpdk_rxq
+ *   Generic pointer to RX queue structure.
+ * @param[out] pkts
+ *   Array to store received packets.
+ * @param pkts_n
+ *   Maximum number of packets in array.
+ *
+ * @return
+ *   Number of packets successfully received (<= pkts_n).
+ */
+uint16_t
+removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+	(void)dpdk_rxq;
+	(void)pkts;
+	(void)pkts_n;
+	return 0;
+}
diff --git a/src/dpdk/drivers/net/mlx5/mlx5_rxtx.h b/src/dpdk/drivers/net/mlx5/mlx5_rxtx.h
new file mode 100644
index 00000000..f6e2cbac
--- /dev/null
+++ b/src/dpdk/drivers/net/mlx5/mlx5_rxtx.h
@@ -0,0 +1,328 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2015 6WIND S.A.
+ *   Copyright 2015 Mellanox.
+ *
+ *   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.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#ifndef RTE_PMD_MLX5_RXTX_H_
+#define RTE_PMD_MLX5_RXTX_H_
+
+#include <stddef.h>
+#include <stdint.h>
+
+/* Verbs header. */
+/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <infiniband/verbs.h>
+#include <infiniband/mlx5_hw.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+/* DPDK headers don't like -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <rte_mbuf.h>
+#include <rte_mempool.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+#include "mlx5_utils.h"
+#include "mlx5.h"
+#include "mlx5_autoconf.h"
+#include "mlx5_defs.h"
+#include "mlx5_prm.h"
+
+struct mlx5_rxq_stats {
+	unsigned int idx; /**< Mapping index. */
+#ifdef MLX5_PMD_SOFT_COUNTERS
+	uint64_t ipackets; /**< Total of successfully received packets. */
+	uint64_t ibytes; /**< Total of successfully received bytes. */
+#endif
+	uint64_t idropped; /**< Total of packets dropped when RX ring full. */
+	uint64_t rx_nombuf; /**< Total of RX mbuf allocation failures. */
+};
+
+struct mlx5_txq_stats {
+	unsigned int idx; /**< Mapping index. */
+#ifdef MLX5_PMD_SOFT_COUNTERS
+	uint64_t opackets; /**< Total of successfully sent packets. */
+	uint64_t obytes; /**< Total of successfully sent bytes. */
+#endif
+	uint64_t odropped; /**< Total of packets not sent when TX ring full. */
+};
+
+/* Flow director queue structure. */
+struct fdir_queue {
+	struct ibv_qp *qp; /* Associated RX QP. */
+	struct ibv_exp_rwq_ind_table *ind_table; /* Indirection table. */
+};
+
+struct priv;
+
+/* Compressed CQE context. */
+struct rxq_zip {
+	uint16_t ai; /* Array index. */
+	uint16_t ca; /* Current array index. */
+	uint16_t na; /* Next array index. */
+	uint16_t cq_ci; /* The next CQE. */
+	uint32_t cqe_cnt; /* Number of CQEs. */
+};
+
+/* RX queue descriptor. */
+struct rxq {
+	unsigned int csum:1; /* Enable checksum offloading. */
+	unsigned int csum_l2tun:1; /* Same for L2 tunnels. */
+	unsigned int vlan_strip:1; /* Enable VLAN stripping. */
+	unsigned int crc_present:1; /* CRC must be subtracted. */
+	unsigned int sges_n:2; /* Log 2 of SGEs (max buffers per packet). */
+	uint16_t rq_ci;
+	uint16_t cq_ci;
+	uint16_t elts_n;
+	uint16_t cqe_n; /* Number of CQ elements. */
+	uint16_t port_id;
+	volatile struct mlx5_wqe_data_seg(*wqes)[];
+	volatile struct mlx5_cqe(*cqes)[];
+	struct rxq_zip zip; /* Compressed context. */
+	volatile uint32_t *rq_db;
+	volatile uint32_t *cq_db;
+	struct rte_mbuf *(*elts)[];
+	struct rte_mempool *mp;
+	struct mlx5_rxq_stats stats;
+} __rte_cache_aligned;
+
+/* RX queue control descriptor. */
+struct rxq_ctrl {
+	struct priv *priv; /* Back pointer to private data. */
+	struct ibv_cq *cq; /* Completion Queue. */
+	struct ibv_exp_wq *wq; /* Work Queue. */
+	struct ibv_exp_res_domain *rd; /* Resource Domain. */
+	struct fdir_queue fdir_queue; /* Flow director queue. */
+	struct ibv_mr *mr; /* Memory Region (for mp). */
+	struct ibv_exp_wq_family *if_wq; /* WQ burst interface. */
+	struct ibv_exp_cq_family_v1 *if_cq; /* CQ interface. */
+	unsigned int socket; /* CPU socket ID for allocations. */
+	struct rxq rxq; /* Data path structure. */
+};
+
+/* Hash RX queue types. */
+enum hash_rxq_type {
+	HASH_RXQ_TCPV4,
+	HASH_RXQ_UDPV4,
+	HASH_RXQ_IPV4,
+	HASH_RXQ_TCPV6,
+	HASH_RXQ_UDPV6,
+	HASH_RXQ_IPV6,
+	HASH_RXQ_ETH,
+};
+
+/* Flow structure with Ethernet specification. It is packed to prevent padding
+ * between attr and spec as this layout is expected by libibverbs. */
+struct flow_attr_spec_eth {
+	struct ibv_exp_flow_attr attr;
+	struct ibv_exp_flow_spec_eth spec;
+} __attribute__((packed));
+
+/* Define a struct flow_attr_spec_eth object as an array of at least
+ * "size" bytes. Room after the first index is normally used to store
+ * extra flow specifications. */
+#define FLOW_ATTR_SPEC_ETH(name, size) \
+	struct flow_attr_spec_eth name \
+		[((size) / sizeof(struct flow_attr_spec_eth)) + \
+		 !!((size) % sizeof(struct flow_attr_spec_eth))]
+
+/* Initialization data for hash RX queue. */
+struct hash_rxq_init {
+	uint64_t hash_fields; /* Fields that participate in the hash. */
+	uint64_t dpdk_rss_hf; /* Matching DPDK RSS hash fields. */
+	unsigned int flow_priority; /* Flow priority to use. */
+	union {
+		struct {
+			enum ibv_exp_flow_spec_type type;
+			uint16_t size;
+		} hdr;
+		struct ibv_exp_flow_spec_tcp_udp tcp_udp;
+		struct ibv_exp_flow_spec_ipv4 ipv4;
+		struct ibv_exp_flow_spec_ipv6 ipv6;
+		struct ibv_exp_flow_spec_eth eth;
+	} flow_spec; /* Flow specification template. */
+	const struct hash_rxq_init *underlayer; /* Pointer to underlayer. */
+};
+
+/* Initialization data for indirection table. */
+struct ind_table_init {
+	unsigned int max_size; /* Maximum number of WQs. */
+	/* Hash RX queues using this table. */
+	unsigned int hash_types;
+	unsigned int hash_types_n;
+};
+
+/* Initialization data for special flows. */
+struct special_flow_init {
+	uint8_t dst_mac_val[6];
+	uint8_t dst_mac_mask[6];
+	unsigned int hash_types;
+	unsigned int per_vlan:1;
+};
+
+enum hash_rxq_flow_type {
+	HASH_RXQ_FLOW_TYPE_PROMISC,
+	HASH_RXQ_FLOW_TYPE_ALLMULTI,
+	HASH_RXQ_FLOW_TYPE_BROADCAST,
+	HASH_RXQ_FLOW_TYPE_IPV6MULTI,
+	HASH_RXQ_FLOW_TYPE_MAC,
+};
+
+#ifndef NDEBUG
+static inline const char *
+hash_rxq_flow_type_str(enum hash_rxq_flow_type flow_type)
+{
+	switch (flow_type) {
+	case HASH_RXQ_FLOW_TYPE_PROMISC:
+		return "promiscuous";
+	case HASH_RXQ_FLOW_TYPE_ALLMULTI:
+		return "allmulticast";
+	case HASH_RXQ_FLOW_TYPE_BROADCAST:
+		return "broadcast";
+	case HASH_RXQ_FLOW_TYPE_IPV6MULTI:
+		return "IPv6 multicast";
+	case HASH_RXQ_FLOW_TYPE_MAC:
+		return "MAC";
+	}
+	return NULL;
+}
+#endif /* NDEBUG */
+
+struct hash_rxq {
+	struct priv *priv; /* Back pointer to private data. */
+	struct ibv_qp *qp; /* Hash RX QP. */
+	enum hash_rxq_type type; /* Hash RX queue type. */
+	/* MAC flow steering rules, one per VLAN ID. */
+	struct ibv_exp_flow *mac_flow
+		[MLX5_MAX_MAC_ADDRESSES][MLX5_MAX_VLAN_IDS];
+	struct ibv_exp_flow *special_flow
+		[MLX5_MAX_SPECIAL_FLOWS][MLX5_MAX_VLAN_IDS];
+};
+
+/* TX queue descriptor. */
+struct txq {
+	uint16_t elts_head; /* Current index in (*elts)[]. */
+	uint16_t elts_tail; /* First element awaiting completion. */
+	uint16_t elts_comp; /* Counter since last completion request. */
+	uint16_t elts_n; /* (*elts)[] length. */
+	uint16_t cq_ci; /* Consumer index for completion queue. */
+	uint16_t cqe_n; /* Number of CQ elements. */
+	uint16_t wqe_ci; /* Consumer index for work queue. */
+	uint16_t wqe_n; /* Number of WQ elements. */
+	uint16_t bf_offset; /* Blueflame offset. */
+	uint16_t bf_buf_size; /* Blueflame size. */
+	uint16_t max_inline; /* Maximum size to inline in a WQE. */
+	uint32_t qp_num_8s; /* QP number shifted by 8. */
+	volatile struct mlx5_cqe (*cqes)[]; /* Completion queue. */
+	volatile union mlx5_wqe (*wqes)[]; /* Work queue. */
+	volatile uint32_t *qp_db; /* Work queue doorbell. */
+	volatile uint32_t *cq_db; /* Completion queue doorbell. */
+	volatile void *bf_reg; /* Blueflame register. */
+	struct {
+		const struct rte_mempool *mp; /* Cached Memory Pool. */
+		struct ibv_mr *mr; /* Memory Region (for mp). */
+		uint32_t lkey; /* htonl(mr->lkey) */
+	} mp2mr[MLX5_PMD_TX_MP_CACHE]; /* MP to MR translation table. */
+	struct rte_mbuf *(*elts)[]; /* TX elements. */
+	struct mlx5_txq_stats stats; /* TX queue counters. */
+} __rte_cache_aligned;
+
+/* TX queue control descriptor. */
+struct txq_ctrl {
+	struct priv *priv; /* Back pointer to private data. */
+	struct ibv_cq *cq; /* Completion Queue. */
+	struct ibv_qp *qp; /* Queue Pair. */
+	struct ibv_exp_qp_burst_family *if_qp; /* QP burst interface. */
+	struct ibv_exp_cq_family *if_cq; /* CQ interface. */
+	struct ibv_exp_res_domain *rd; /* Resource Domain. */
+	unsigned int socket; /* CPU socket ID for allocations. */
+	struct txq txq; /* Data path structure. */
+};
+
+/* mlx5_rxq.c */
+
+extern const struct hash_rxq_init hash_rxq_init[];
+extern const unsigned int hash_rxq_init_n;
+
+extern uint8_t rss_hash_default_key[];
+extern const size_t rss_hash_default_key_len;
+
+size_t priv_flow_attr(struct priv *, struct ibv_exp_flow_attr *,
+		      size_t, enum hash_rxq_type);
+int priv_create_hash_rxqs(struct priv *);
+void priv_destroy_hash_rxqs(struct priv *);
+int priv_allow_flow_type(struct priv *, enum hash_rxq_flow_type);
+int priv_rehash_flows(struct priv *);
+void rxq_cleanup(struct rxq_ctrl *);
+int rxq_rehash(struct rte_eth_dev *, struct rxq_ctrl *);
+int rxq_ctrl_setup(struct rte_eth_dev *, struct rxq_ctrl *, uint16_t,
+		   unsigned int, const struct rte_eth_rxconf *,
+		   struct rte_mempool *);
+int mlx5_rx_queue_setup(struct rte_eth_dev *, uint16_t, uint16_t, unsigned int,
+			const struct rte_eth_rxconf *, struct rte_mempool *);
+void mlx5_rx_queue_release(void *);
+uint16_t mlx5_rx_burst_secondary_setup(void *, struct rte_mbuf **, uint16_t);
+
+/* mlx5_txq.c */
+
+void txq_cleanup(struct txq_ctrl *);
+int txq_ctrl_setup(struct rte_eth_dev *, struct txq_ctrl *, uint16_t,
+		   unsigned int, const struct rte_eth_txconf *);
+int mlx5_tx_queue_setup(struct rte_eth_dev *, uint16_t, uint16_t, unsigned int,
+			const struct rte_eth_txconf *);
+void mlx5_tx_queue_release(void *);
+uint16_t mlx5_tx_burst_secondary_setup(void *, struct rte_mbuf **, uint16_t);
+
+/* mlx5_rxtx.c */
+
+uint16_t mlx5_tx_burst(void *, struct rte_mbuf **, uint16_t);
+uint16_t mlx5_tx_burst_inline(void *, struct rte_mbuf **, uint16_t);
+uint16_t mlx5_tx_burst_mpw(void *, struct rte_mbuf **, uint16_t);
+uint16_t mlx5_tx_burst_mpw_inline(void *, struct rte_mbuf **, uint16_t);
+uint16_t mlx5_rx_burst(void *, struct rte_mbuf **, uint16_t);
+uint16_t removed_tx_burst(void *, struct rte_mbuf **, uint16_t);
+uint16_t removed_rx_burst(void *, struct rte_mbuf **, uint16_t);
+
+/* mlx5_mr.c */
+
+struct ibv_mr *mlx5_mp2mr(struct ibv_pd *, struct rte_mempool *);
+void txq_mp2mr_iter(struct rte_mempool *, void *);
+uint32_t txq_mp2mr_reg(struct txq *, struct rte_mempool *, unsigned int);
+
+#endif /* RTE_PMD_MLX5_RXTX_H_ */
diff --git a/src/dpdk/drivers/net/mlx5/mlx5_stats.c b/src/dpdk/drivers/net/mlx5/mlx5_stats.c
new file mode 100644
index 00000000..2d3cb519
--- /dev/null
+++ b/src/dpdk/drivers/net/mlx5/mlx5_stats.c
@@ -0,0 +1,144 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2015 6WIND S.A.
+ *   Copyright 2015 Mellanox.
+ *
+ *   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.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+/* DPDK headers don't like -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <rte_ethdev.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+#include "mlx5.h"
+#include "mlx5_rxtx.h"
+#include "mlx5_defs.h"
+
+/**
+ * DPDK callback to get device statistics.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param[out] stats
+ *   Stats structure output buffer.
+ */
+void
+mlx5_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
+{
+	struct priv *priv = mlx5_get_priv(dev);
+	struct rte_eth_stats tmp = {0};
+	unsigned int i;
+	unsigned int idx;
+
+	priv_lock(priv);
+	/* Add software counters. */
+	for (i = 0; (i != priv->rxqs_n); ++i) {
+		struct rxq *rxq = (*priv->rxqs)[i];
+
+		if (rxq == NULL)
+			continue;
+		idx = rxq->stats.idx;
+		if (idx < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
+#ifdef MLX5_PMD_SOFT_COUNTERS
+			tmp.q_ipackets[idx] += rxq->stats.ipackets;
+			tmp.q_ibytes[idx] += rxq->stats.ibytes;
+#endif
+			tmp.q_errors[idx] += (rxq->stats.idropped +
+					      rxq->stats.rx_nombuf);
+		}
+#ifdef MLX5_PMD_SOFT_COUNTERS
+		tmp.ipackets += rxq->stats.ipackets;
+		tmp.ibytes += rxq->stats.ibytes;
+#endif
+		tmp.ierrors += rxq->stats.idropped;
+		tmp.rx_nombuf += rxq->stats.rx_nombuf;
+	}
+	for (i = 0; (i != priv->txqs_n); ++i) {
+		struct txq *txq = (*priv->txqs)[i];
+
+		if (txq == NULL)
+			continue;
+		idx = txq->stats.idx;
+		if (idx < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
+#ifdef MLX5_PMD_SOFT_COUNTERS
+			tmp.q_opackets[idx] += txq->stats.opackets;
+			tmp.q_obytes[idx] += txq->stats.obytes;
+#endif
+			tmp.q_errors[idx] += txq->stats.odropped;
+		}
+#ifdef MLX5_PMD_SOFT_COUNTERS
+		tmp.opackets += txq->stats.opackets;
+		tmp.obytes += txq->stats.obytes;
+#endif
+		tmp.oerrors += txq->stats.odropped;
+	}
+#ifndef MLX5_PMD_SOFT_COUNTERS
+	/* FIXME: retrieve and add hardware counters. */
+#endif
+	*stats = tmp;
+	priv_unlock(priv);
+}
+
+/**
+ * DPDK callback to clear device statistics.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ */
+void
+mlx5_stats_reset(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	unsigned int i;
+	unsigned int idx;
+
+	priv_lock(priv);
+	for (i = 0; (i != priv->rxqs_n); ++i) {
+		if ((*priv->rxqs)[i] == NULL)
+			continue;
+		idx = (*priv->rxqs)[i]->stats.idx;
+		(*priv->rxqs)[i]->stats =
+			(struct mlx5_rxq_stats){ .idx = idx };
+	}
+	for (i = 0; (i != priv->txqs_n); ++i) {
+		if ((*priv->txqs)[i] == NULL)
+			continue;
+		idx = (*priv->txqs)[i]->stats.idx;
+		(*priv->txqs)[i]->stats =
+			(struct mlx5_txq_stats){ .idx = idx };
+	}
+#ifndef MLX5_PMD_SOFT_COUNTERS
+	/* FIXME: reset hardware counters. */
+#endif
+	priv_unlock(priv);
+}
diff --git a/src/dpdk/drivers/net/mlx5/mlx5_trigger.c b/src/dpdk/drivers/net/mlx5/mlx5_trigger.c
new file mode 100644
index 00000000..e9b9a293
--- /dev/null
+++ b/src/dpdk/drivers/net/mlx5/mlx5_trigger.c
@@ -0,0 +1,126 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2015 6WIND S.A.
+ *   Copyright 2015 Mellanox.
+ *
+ *   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.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+/* DPDK headers don't like -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <rte_ether.h>
+#include <rte_ethdev.h>
+#include <rte_interrupts.h>
+#include <rte_alarm.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+#include "mlx5.h"
+#include "mlx5_rxtx.h"
+#include "mlx5_utils.h"
+
+/**
+ * DPDK callback to start the device.
+ *
+ * Simulate device start by attaching all configured flows.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ *
+ * @return
+ *   0 on success, negative errno value on failure.
+ */
+int
+mlx5_dev_start(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	int err;
+
+	if (mlx5_is_secondary())
+		return -E_RTE_SECONDARY;
+
+	priv_lock(priv);
+	if (priv->started) {
+		priv_unlock(priv);
+		return 0;
+	}
+	DEBUG("%p: allocating and configuring hash RX queues", (void *)dev);
+	err = priv_create_hash_rxqs(priv);
+	if (!err)
+		err = priv_rehash_flows(priv);
+	if (!err)
+		priv->started = 1;
+	else {
+		ERROR("%p: an error occurred while configuring hash RX queues:"
+		      " %s",
+		      (void *)priv, strerror(err));
+		/* Rollback. */
+		priv_special_flow_disable_all(priv);
+		priv_mac_addrs_disable(priv);
+		priv_destroy_hash_rxqs(priv);
+	}
+	if (dev->data->dev_conf.fdir_conf.mode != RTE_FDIR_MODE_NONE)
+		priv_fdir_enable(priv);
+	priv_dev_interrupt_handler_install(priv, dev);
+	priv_unlock(priv);
+	return -err;
+}
+
+/**
+ * DPDK callback to stop the device.
+ *
+ * Simulate device stop by detaching all configured flows.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ */
+void
+mlx5_dev_stop(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+
+	if (mlx5_is_secondary())
+		return;
+
+	priv_lock(priv);
+	if (!priv->started) {
+		priv_unlock(priv);
+		return;
+	}
+	DEBUG("%p: cleaning up and destroying hash RX queues", (void *)dev);
+	priv_special_flow_disable_all(priv);
+	priv_mac_addrs_disable(priv);
+	priv_destroy_hash_rxqs(priv);
+	priv_fdir_disable(priv);
+	priv_dev_interrupt_handler_uninstall(priv, dev);
+	priv->started = 0;
+	priv_unlock(priv);
+}
diff --git a/src/dpdk/drivers/net/mlx5/mlx5_txq.c b/src/dpdk/drivers/net/mlx5/mlx5_txq.c
new file mode 100644
index 00000000..6fe61c4a
--- /dev/null
+++ b/src/dpdk/drivers/net/mlx5/mlx5_txq.c
@@ -0,0 +1,599 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2015 6WIND S.A.
+ *   Copyright 2015 Mellanox.
+ *
+ *   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.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <stddef.h>
+#include <assert.h>
+#include <errno.h>
+#include <string.h>
+#include <stdint.h>
+
+/* Verbs header. */
+/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <infiniband/verbs.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+/* DPDK headers don't like -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <rte_mbuf.h>
+#include <rte_malloc.h>
+#include <rte_ethdev.h>
+#include <rte_common.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+#include "mlx5_utils.h"
+#include "mlx5_defs.h"
+#include "mlx5.h"
+#include "mlx5_rxtx.h"
+#include "mlx5_autoconf.h"
+#include "mlx5_defs.h"
+
+/**
+ * Allocate TX queue elements.
+ *
+ * @param txq_ctrl
+ *   Pointer to TX queue structure.
+ * @param elts_n
+ *   Number of elements to allocate.
+ */
+static void
+txq_alloc_elts(struct txq_ctrl *txq_ctrl, unsigned int elts_n)
+{
+	unsigned int i;
+
+	for (i = 0; (i != elts_n); ++i)
+		(*txq_ctrl->txq.elts)[i] = NULL;
+	for (i = 0; (i != txq_ctrl->txq.wqe_n); ++i) {
+		volatile union mlx5_wqe *wqe = &(*txq_ctrl->txq.wqes)[i];
+
+		memset((void *)(uintptr_t)wqe, 0x0, sizeof(*wqe));
+	}
+	DEBUG("%p: allocated and configured %u WRs", (void *)txq_ctrl, elts_n);
+	txq_ctrl->txq.elts_head = 0;
+	txq_ctrl->txq.elts_tail = 0;
+	txq_ctrl->txq.elts_comp = 0;
+}
+
+/**
+ * Free TX queue elements.
+ *
+ * @param txq_ctrl
+ *   Pointer to TX queue structure.
+ */
+static void
+txq_free_elts(struct txq_ctrl *txq_ctrl)
+{
+	unsigned int elts_n = txq_ctrl->txq.elts_n;
+	unsigned int elts_head = txq_ctrl->txq.elts_head;
+	unsigned int elts_tail = txq_ctrl->txq.elts_tail;
+	struct rte_mbuf *(*elts)[elts_n] = txq_ctrl->txq.elts;
+
+	DEBUG("%p: freeing WRs", (void *)txq_ctrl);
+	txq_ctrl->txq.elts_head = 0;
+	txq_ctrl->txq.elts_tail = 0;
+	txq_ctrl->txq.elts_comp = 0;
+
+	while (elts_tail != elts_head) {
+		struct rte_mbuf *elt = (*elts)[elts_tail];
+
+		assert(elt != NULL);
+		rte_pktmbuf_free(elt);
+#ifndef NDEBUG
+		/* Poisoning. */
+		memset(&(*elts)[elts_tail],
+		       0x77,
+		       sizeof((*elts)[elts_tail]));
+#endif
+		if (++elts_tail == elts_n)
+			elts_tail = 0;
+	}
+}
+
+/**
+ * Clean up a TX queue.
+ *
+ * Destroy objects, free allocated memory and reset the structure for reuse.
+ *
+ * @param txq_ctrl
+ *   Pointer to TX queue structure.
+ */
+void
+txq_cleanup(struct txq_ctrl *txq_ctrl)
+{
+	struct ibv_exp_release_intf_params params;
+	size_t i;
+
+	DEBUG("cleaning up %p", (void *)txq_ctrl);
+	txq_free_elts(txq_ctrl);
+	if (txq_ctrl->if_qp != NULL) {
+		assert(txq_ctrl->priv != NULL);
+		assert(txq_ctrl->priv->ctx != NULL);
+		assert(txq_ctrl->qp != NULL);
+		params = (struct ibv_exp_release_intf_params){
+			.comp_mask = 0,
+		};
+		claim_zero(ibv_exp_release_intf(txq_ctrl->priv->ctx,
+						txq_ctrl->if_qp,
+						&params));
+	}
+	if (txq_ctrl->if_cq != NULL) {
+		assert(txq_ctrl->priv != NULL);
+		assert(txq_ctrl->priv->ctx != NULL);
+		assert(txq_ctrl->cq != NULL);
+		params = (struct ibv_exp_release_intf_params){
+			.comp_mask = 0,
+		};
+		claim_zero(ibv_exp_release_intf(txq_ctrl->priv->ctx,
+						txq_ctrl->if_cq,
+						&params));
+	}
+	if (txq_ctrl->qp != NULL)
+		claim_zero(ibv_destroy_qp(txq_ctrl->qp));
+	if (txq_ctrl->cq != NULL)
+		claim_zero(ibv_destroy_cq(txq_ctrl->cq));
+	if (txq_ctrl->rd != NULL) {
+		struct ibv_exp_destroy_res_domain_attr attr = {
+			.comp_mask = 0,
+		};
+
+		assert(txq_ctrl->priv != NULL);
+		assert(txq_ctrl->priv->ctx != NULL);
+		claim_zero(ibv_exp_destroy_res_domain(txq_ctrl->priv->ctx,
+						      txq_ctrl->rd,
+						      &attr));
+	}
+	for (i = 0; (i != RTE_DIM(txq_ctrl->txq.mp2mr)); ++i) {
+		if (txq_ctrl->txq.mp2mr[i].mp == NULL)
+			break;
+		assert(txq_ctrl->txq.mp2mr[i].mr != NULL);
+		claim_zero(ibv_dereg_mr(txq_ctrl->txq.mp2mr[i].mr));
+	}
+	memset(txq_ctrl, 0, sizeof(*txq_ctrl));
+}
+
+/**
+ * Initialize TX queue.
+ *
+ * @param tmpl
+ *   Pointer to TX queue control template.
+ * @param txq_ctrl
+ *   Pointer to TX queue control.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static inline int
+txq_setup(struct txq_ctrl *tmpl, struct txq_ctrl *txq_ctrl)
+{
+	struct mlx5_qp *qp = to_mqp(tmpl->qp);
+	struct ibv_cq *ibcq = tmpl->cq;
+	struct mlx5_cq *cq = to_mxxx(cq, cq);
+
+	if (cq->cqe_sz != RTE_CACHE_LINE_SIZE) {
+		ERROR("Wrong MLX5_CQE_SIZE environment variable value: "
+		      "it should be set to %u", RTE_CACHE_LINE_SIZE);
+		return EINVAL;
+	}
+	tmpl->txq.cqe_n = ibcq->cqe + 1;
+	tmpl->txq.qp_num_8s = qp->ctrl_seg.qp_num << 8;
+	tmpl->txq.wqes =
+		(volatile union mlx5_wqe (*)[])
+		(uintptr_t)qp->gen_data.sqstart;
+	tmpl->txq.wqe_n = qp->sq.wqe_cnt;
+	tmpl->txq.qp_db = &qp->gen_data.db[MLX5_SND_DBR];
+	tmpl->txq.bf_reg = qp->gen_data.bf->reg;
+	tmpl->txq.bf_offset = qp->gen_data.bf->offset;
+	tmpl->txq.bf_buf_size = qp->gen_data.bf->buf_size;
+	tmpl->txq.cq_db = cq->dbrec;
+	tmpl->txq.cqes =
+		(volatile struct mlx5_cqe (*)[])
+		(uintptr_t)cq->active_buf->buf;
+	tmpl->txq.elts =
+		(struct rte_mbuf *(*)[tmpl->txq.elts_n])
+		((uintptr_t)txq_ctrl + sizeof(*txq_ctrl));
+	return 0;
+}
+
+/**
+ * Configure a TX queue.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param txq_ctrl
+ *   Pointer to TX queue structure.
+ * @param desc
+ *   Number of descriptors to configure in queue.
+ * @param socket
+ *   NUMA socket on which memory must be allocated.
+ * @param[in] conf
+ *   Thresholds parameters.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+int
+txq_ctrl_setup(struct rte_eth_dev *dev, struct txq_ctrl *txq_ctrl,
+	       uint16_t desc, unsigned int socket,
+	       const struct rte_eth_txconf *conf)
+{
+	struct priv *priv = mlx5_get_priv(dev);
+	struct txq_ctrl tmpl = {
+		.priv = priv,
+		.socket = socket,
+	};
+	union {
+		struct ibv_exp_query_intf_params params;
+		struct ibv_exp_qp_init_attr init;
+		struct ibv_exp_res_domain_init_attr rd;
+		struct ibv_exp_cq_init_attr cq;
+		struct ibv_exp_qp_attr mod;
+		struct ibv_exp_cq_attr cq_attr;
+	} attr;
+	enum ibv_exp_query_intf_status status;
+	int ret = 0;
+
+	if (mlx5_getenv_int("MLX5_ENABLE_CQE_COMPRESSION")) {
+		ret = ENOTSUP;
+		ERROR("MLX5_ENABLE_CQE_COMPRESSION must never be set");
+		goto error;
+	}
+	(void)conf; /* Thresholds configuration (ignored). */
+	assert(desc > MLX5_TX_COMP_THRESH);
+	tmpl.txq.elts_n = desc;
+	/* MRs will be registered in mp2mr[] later. */
+	attr.rd = (struct ibv_exp_res_domain_init_attr){
+		.comp_mask = (IBV_EXP_RES_DOMAIN_THREAD_MODEL |
+			      IBV_EXP_RES_DOMAIN_MSG_MODEL),
+		.thread_model = IBV_EXP_THREAD_SINGLE,
+		.msg_model = IBV_EXP_MSG_HIGH_BW,
+	};
+	tmpl.rd = ibv_exp_create_res_domain(priv->ctx, &attr.rd);
+	if (tmpl.rd == NULL) {
+		ret = ENOMEM;
+		ERROR("%p: RD creation failure: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	attr.cq = (struct ibv_exp_cq_init_attr){
+		.comp_mask = IBV_EXP_CQ_INIT_ATTR_RES_DOMAIN,
+		.res_domain = tmpl.rd,
+	};
+	tmpl.cq = ibv_exp_create_cq(priv->ctx,
+				    (((desc / MLX5_TX_COMP_THRESH) - 1) ?
+				     ((desc / MLX5_TX_COMP_THRESH) - 1) : 1),
+				    NULL, NULL, 0, &attr.cq);
+	if (tmpl.cq == NULL) {
+		ret = ENOMEM;
+		ERROR("%p: CQ creation failure: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	DEBUG("priv->device_attr.max_qp_wr is %d",
+	      priv->device_attr.max_qp_wr);
+	DEBUG("priv->device_attr.max_sge is %d",
+	      priv->device_attr.max_sge);
+	attr.init = (struct ibv_exp_qp_init_attr){
+		/* CQ to be associated with the send queue. */
+		.send_cq = tmpl.cq,
+		/* CQ to be associated with the receive queue. */
+		.recv_cq = tmpl.cq,
+		.cap = {
+			/* Max number of outstanding WRs. */
+			.max_send_wr = ((priv->device_attr.max_qp_wr < desc) ?
+					priv->device_attr.max_qp_wr :
+					desc),
+			/*
+			 * Max number of scatter/gather elements in a WR,
+			 * must be 1 to prevent libmlx5 from trying to affect
+			 * too much memory. TX gather is not impacted by the
+			 * priv->device_attr.max_sge limit and will still work
+			 * properly.
+			 */
+			.max_send_sge = 1,
+		},
+		.qp_type = IBV_QPT_RAW_PACKET,
+		/* Do *NOT* enable this, completions events are managed per
+		 * TX burst. */
+		.sq_sig_all = 0,
+		.pd = priv->pd,
+		.res_domain = tmpl.rd,
+		.comp_mask = (IBV_EXP_QP_INIT_ATTR_PD |
+			      IBV_EXP_QP_INIT_ATTR_RES_DOMAIN),
+	};
+	if (priv->txq_inline && priv->txqs_n >= priv->txqs_inline) {
+		tmpl.txq.max_inline = priv->txq_inline;
+		attr.init.cap.max_inline_data = tmpl.txq.max_inline;
+	}
+	tmpl.qp = ibv_exp_create_qp(priv->ctx, &attr.init);
+	if (tmpl.qp == NULL) {
+		ret = (errno ? errno : EINVAL);
+		ERROR("%p: QP creation failure: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	DEBUG("TX queue capabilities: max_send_wr=%u, max_send_sge=%u,"
+	      " max_inline_data=%u",
+	      attr.init.cap.max_send_wr,
+	      attr.init.cap.max_send_sge,
+	      attr.init.cap.max_inline_data);
+	attr.mod = (struct ibv_exp_qp_attr){
+		/* Move the QP to this state. */
+		.qp_state = IBV_QPS_INIT,
+		/* Primary port number. */
+		.port_num = priv->port
+	};
+	ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod,
+				(IBV_EXP_QP_STATE | IBV_EXP_QP_PORT));
+	if (ret) {
+		ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	ret = txq_setup(&tmpl, txq_ctrl);
+	if (ret) {
+		ERROR("%p: cannot initialize TX queue structure: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	txq_alloc_elts(&tmpl, desc);
+	attr.mod = (struct ibv_exp_qp_attr){
+		.qp_state = IBV_QPS_RTR
+	};
+	ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
+	if (ret) {
+		ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	attr.mod.qp_state = IBV_QPS_RTS;
+	ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
+	if (ret) {
+		ERROR("%p: QP state to IBV_QPS_RTS failed: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	attr.params = (struct ibv_exp_query_intf_params){
+		.intf_scope = IBV_EXP_INTF_GLOBAL,
+		.intf = IBV_EXP_INTF_CQ,
+		.obj = tmpl.cq,
+	};
+	tmpl.if_cq = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
+	if (tmpl.if_cq == NULL) {
+		ret = EINVAL;
+		ERROR("%p: CQ interface family query failed with status %d",
+		      (void *)dev, status);
+		goto error;
+	}
+	attr.params = (struct ibv_exp_query_intf_params){
+		.intf_scope = IBV_EXP_INTF_GLOBAL,
+		.intf = IBV_EXP_INTF_QP_BURST,
+		.intf_version = 1,
+		.obj = tmpl.qp,
+		/* Enable multi-packet send if supported. */
+		.family_flags =
+			((priv->mps && !priv->sriov) ?
+			 IBV_EXP_QP_BURST_CREATE_ENABLE_MULTI_PACKET_SEND_WR :
+			 0),
+	};
+	tmpl.if_qp = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
+	if (tmpl.if_qp == NULL) {
+		ret = EINVAL;
+		ERROR("%p: QP interface family query failed with status %d",
+		      (void *)dev, status);
+		goto error;
+	}
+	/* Clean up txq in case we're reinitializing it. */
+	DEBUG("%p: cleaning-up old txq just in case", (void *)txq_ctrl);
+	txq_cleanup(txq_ctrl);
+	*txq_ctrl = tmpl;
+	DEBUG("%p: txq updated with %p", (void *)txq_ctrl, (void *)&tmpl);
+	/* Pre-register known mempools. */
+	rte_mempool_walk(txq_mp2mr_iter, txq_ctrl);
+	assert(ret == 0);
+	return 0;
+error:
+	txq_cleanup(&tmpl);
+	assert(ret > 0);
+	return ret;
+}
+
+/**
+ * DPDK callback to configure a TX queue.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param idx
+ *   TX queue index.
+ * @param desc
+ *   Number of descriptors to configure in queue.
+ * @param socket
+ *   NUMA socket on which memory must be allocated.
+ * @param[in] conf
+ *   Thresholds parameters.
+ *
+ * @return
+ *   0 on success, negative errno value on failure.
+ */
+int
+mlx5_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
+		    unsigned int socket, const struct rte_eth_txconf *conf)
+{
+	struct priv *priv = dev->data->dev_private;
+	struct txq *txq = (*priv->txqs)[idx];
+	struct txq_ctrl *txq_ctrl = container_of(txq, struct txq_ctrl, txq);
+	int ret;
+
+	if (mlx5_is_secondary())
+		return -E_RTE_SECONDARY;
+
+	priv_lock(priv);
+	if (desc <= MLX5_TX_COMP_THRESH) {
+		WARN("%p: number of descriptors requested for TX queue %u"
+		     " must be higher than MLX5_TX_COMP_THRESH, using"
+		     " %u instead of %u",
+		     (void *)dev, idx, MLX5_TX_COMP_THRESH + 1, desc);
+		desc = MLX5_TX_COMP_THRESH + 1;
+	}
+	if (!rte_is_power_of_2(desc)) {
+		desc = 1 << log2above(desc);
+		WARN("%p: increased number of descriptors in TX queue %u"
+		     " to the next power of two (%d)",
+		     (void *)dev, idx, desc);
+	}
+	DEBUG("%p: configuring queue %u for %u descriptors",
+	      (void *)dev, idx, desc);
+	if (idx >= priv->txqs_n) {
+		ERROR("%p: queue index out of range (%u >= %u)",
+		      (void *)dev, idx, priv->txqs_n);
+		priv_unlock(priv);
+		return -EOVERFLOW;
+	}
+	if (txq != NULL) {
+		DEBUG("%p: reusing already allocated queue index %u (%p)",
+		      (void *)dev, idx, (void *)txq);
+		if (priv->started) {
+			priv_unlock(priv);
+			return -EEXIST;
+		}
+		(*priv->txqs)[idx] = NULL;
+		txq_cleanup(txq_ctrl);
+	} else {
+		txq_ctrl =
+			rte_calloc_socket("TXQ", 1,
+					  sizeof(*txq_ctrl) +
+					  desc * sizeof(struct rte_mbuf *),
+					  0, socket);
+		if (txq_ctrl == NULL) {
+			ERROR("%p: unable to allocate queue index %u",
+			      (void *)dev, idx);
+			priv_unlock(priv);
+			return -ENOMEM;
+		}
+	}
+	ret = txq_ctrl_setup(dev, txq_ctrl, desc, socket, conf);
+	if (ret)
+		rte_free(txq_ctrl);
+	else {
+		txq_ctrl->txq.stats.idx = idx;
+		DEBUG("%p: adding TX queue %p to list",
+		      (void *)dev, (void *)txq_ctrl);
+		(*priv->txqs)[idx] = &txq_ctrl->txq;
+		/* Update send callback. */
+		priv_select_tx_function(priv);
+	}
+	priv_unlock(priv);
+	return -ret;
+}
+
+/**
+ * DPDK callback to release a TX queue.
+ *
+ * @param dpdk_txq
+ *   Generic TX queue pointer.
+ */
+void
+mlx5_tx_queue_release(void *dpdk_txq)
+{
+	struct txq *txq = (struct txq *)dpdk_txq;
+	struct txq_ctrl *txq_ctrl;
+	struct priv *priv;
+	unsigned int i;
+
+	if (mlx5_is_secondary())
+		return;
+
+	if (txq == NULL)
+		return;
+	txq_ctrl = container_of(txq, struct txq_ctrl, txq);
+	priv = txq_ctrl->priv;
+	priv_lock(priv);
+	for (i = 0; (i != priv->txqs_n); ++i)
+		if ((*priv->txqs)[i] == txq) {
+			DEBUG("%p: removing TX queue %p from list",
+			      (void *)priv->dev, (void *)txq_ctrl);
+			(*priv->txqs)[i] = NULL;
+			break;
+		}
+	txq_cleanup(txq_ctrl);
+	rte_free(txq_ctrl);
+	priv_unlock(priv);
+}
+
+/**
+ * DPDK callback for TX in secondary processes.
+ *
+ * This function configures all queues from primary process information
+ * if necessary before reverting to the normal TX burst callback.
+ *
+ * @param dpdk_txq
+ *   Generic pointer to TX queue structure.
+ * @param[in] pkts
+ *   Packets to transmit.
+ * @param pkts_n
+ *   Number of packets in array.
+ *
+ * @return
+ *   Number of packets successfully transmitted (<= pkts_n).
+ */
+uint16_t
+mlx5_tx_burst_secondary_setup(void *dpdk_txq, struct rte_mbuf **pkts,
+			      uint16_t pkts_n)
+{
+	struct txq *txq = dpdk_txq;
+	struct txq_ctrl *txq_ctrl = container_of(txq, struct txq_ctrl, txq);
+	struct priv *priv = mlx5_secondary_data_setup(txq_ctrl->priv);
+	struct priv *primary_priv;
+	unsigned int index;
+
+	if (priv == NULL)
+		return 0;
+	primary_priv =
+		mlx5_secondary_data[priv->dev->data->port_id].primary_priv;
+	/* Look for queue index in both private structures. */
+	for (index = 0; index != priv->txqs_n; ++index)
+		if (((*primary_priv->txqs)[index] == txq) ||
+		    ((*priv->txqs)[index] == txq))
+			break;
+	if (index == priv->txqs_n)
+		return 0;
+	txq = (*priv->txqs)[index];
+	return priv->dev->tx_pkt_burst(txq, pkts, pkts_n);
+}
diff --git a/src/dpdk/drivers/net/mlx5/mlx5_utils.h b/src/dpdk/drivers/net/mlx5/mlx5_utils.h
new file mode 100644
index 00000000..a824787f
--- /dev/null
+++ b/src/dpdk/drivers/net/mlx5/mlx5_utils.h
@@ -0,0 +1,184 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2015 6WIND S.A.
+ *   Copyright 2015 Mellanox.
+ *
+ *   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.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#ifndef RTE_PMD_MLX5_UTILS_H_
+#define RTE_PMD_MLX5_UTILS_H_
+
+#include <stddef.h>
+#include <stdio.h>
+#include <limits.h>
+#include <assert.h>
+#include <errno.h>
+
+#include "mlx5_defs.h"
+
+/* Bit-field manipulation. */
+#define BITFIELD_DECLARE(bf, type, size) \
+	type bf[(((size_t)(size) / (sizeof(type) * CHAR_BIT)) + \
+		 !!((size_t)(size) % (sizeof(type) * CHAR_BIT)))]
+#define BITFIELD_DEFINE(bf, type, size) \
+	BITFIELD_DECLARE((bf), type, (size)) = { 0 }
+#define BITFIELD_SET(bf, b) \
+	(assert((size_t)(b) < (sizeof(bf) * CHAR_BIT)), \
+	 (void)((bf)[((b) / (sizeof((bf)[0]) * CHAR_BIT))] |= \
+		((size_t)1 << ((b) % (sizeof((bf)[0]) * CHAR_BIT)))))
+#define BITFIELD_RESET(bf, b) \
+	(assert((size_t)(b) < (sizeof(bf) * CHAR_BIT)), \
+	 (void)((bf)[((b) / (sizeof((bf)[0]) * CHAR_BIT))] &= \
+		~((size_t)1 << ((b) % (sizeof((bf)[0]) * CHAR_BIT)))))
+#define BITFIELD_ISSET(bf, b) \
+	(assert((size_t)(b) < (sizeof(bf) * CHAR_BIT)), \
+	 !!(((bf)[((b) / (sizeof((bf)[0]) * CHAR_BIT))] & \
+	     ((size_t)1 << ((b) % (sizeof((bf)[0]) * CHAR_BIT))))))
+
+/* Save and restore errno around argument evaluation. */
+#define ERRNO_SAFE(x) ((errno = (int []){ errno, ((x), 0) }[0]))
+
+/*
+ * Helper macros to work around __VA_ARGS__ limitations in a C99 compliant
+ * manner.
+ */
+#define PMD_DRV_LOG_STRIP(a, b) a
+#define PMD_DRV_LOG_OPAREN (
+#define PMD_DRV_LOG_CPAREN )
+#define PMD_DRV_LOG_COMMA ,
+
+/* Return the file name part of a path. */
+static inline const char *
+pmd_drv_log_basename(const char *s)
+{
+	const char *n = s;
+
+	while (*n)
+		if (*(n++) == '/')
+			s = n;
+	return s;
+}
+
+/*
+ * When debugging is enabled (NDEBUG not defined), file, line and function
+ * information replace the driver name (MLX5_DRIVER_NAME) in log messages.
+ */
+#ifndef NDEBUG
+
+#define PMD_DRV_LOG___(level, ...) \
+	ERRNO_SAFE(RTE_LOG(level, PMD, __VA_ARGS__))
+#define PMD_DRV_LOG__(level, ...) \
+	PMD_DRV_LOG___(level, "%s:%u: %s(): " __VA_ARGS__)
+#define PMD_DRV_LOG_(level, s, ...) \
+	PMD_DRV_LOG__(level, \
+		s "\n" PMD_DRV_LOG_COMMA \
+		pmd_drv_log_basename(__FILE__) PMD_DRV_LOG_COMMA \
+		__LINE__ PMD_DRV_LOG_COMMA \
+		__func__, \
+		__VA_ARGS__)
+
+#else /* NDEBUG */
+
+#define PMD_DRV_LOG___(level, ...) \
+	ERRNO_SAFE(RTE_LOG(level, PMD, MLX5_DRIVER_NAME ": " __VA_ARGS__))
+#define PMD_DRV_LOG__(level, ...) \
+	PMD_DRV_LOG___(level, __VA_ARGS__)
+#define PMD_DRV_LOG_(level, s, ...) \
+	PMD_DRV_LOG__(level, s "\n", __VA_ARGS__)
+
+#endif /* NDEBUG */
+
+/* Generic printf()-like logging macro with automatic line feed. */
+#define PMD_DRV_LOG(level, ...) \
+	PMD_DRV_LOG_(level, \
+		__VA_ARGS__ PMD_DRV_LOG_STRIP PMD_DRV_LOG_OPAREN, \
+		PMD_DRV_LOG_CPAREN)
+
+/*
+ * Like assert(), DEBUG() becomes a no-op and claim_zero() does not perform
+ * any check when debugging is disabled.
+ */
+#ifndef NDEBUG
+
+#define DEBUG(...) PMD_DRV_LOG(DEBUG, __VA_ARGS__)
+#define claim_zero(...) assert((__VA_ARGS__) == 0)
+
+#else /* NDEBUG */
+
+#define DEBUG(...) (void)0
+#define claim_zero(...) (__VA_ARGS__)
+
+#endif /* NDEBUG */
+
+#define INFO(...) PMD_DRV_LOG(INFO, __VA_ARGS__)
+#define WARN(...) PMD_DRV_LOG(WARNING, __VA_ARGS__)
+#define ERROR(...) PMD_DRV_LOG(ERR, __VA_ARGS__)
+
+/* Convenience macros for accessing mbuf fields. */
+#define NEXT(m) ((m)->next)
+#define DATA_LEN(m) ((m)->data_len)
+#define PKT_LEN(m) ((m)->pkt_len)
+#define DATA_OFF(m) ((m)->data_off)
+#define SET_DATA_OFF(m, o) ((m)->data_off = (o))
+#define NB_SEGS(m) ((m)->nb_segs)
+#define PORT(m) ((m)->port)
+
+/* Transpose flags. Useful to convert IBV to DPDK flags. */
+#define TRANSPOSE(val, from, to) \
+	(((from) >= (to)) ? \
+	 (((val) & (from)) / ((from) / (to))) : \
+	 (((val) & (from)) * ((to) / (from))))
+
+/* Allocate a buffer on the stack and fill it with a printf format string. */
+#define MKSTR(name, ...) \
+	char name[snprintf(NULL, 0, __VA_ARGS__) + 1]; \
+	\
+	snprintf(name, sizeof(name), __VA_ARGS__)
+
+/**
+ * Return nearest power of two above input value.
+ *
+ * @param v
+ *   Input value.
+ *
+ * @return
+ *   Nearest power of two above input value.
+ */
+static inline unsigned int
+log2above(unsigned int v)
+{
+	unsigned int l;
+	unsigned int r;
+
+	for (l = 0, r = 0; (v >> 1); ++l, v >>= 1)
+		r |= (v & 1);
+	return l + r;
+}
+
+#endif /* RTE_PMD_MLX5_UTILS_H_ */
diff --git a/src/dpdk/drivers/net/mlx5/mlx5_vlan.c b/src/dpdk/drivers/net/mlx5/mlx5_vlan.c
new file mode 100644
index 00000000..4719e697
--- /dev/null
+++ b/src/dpdk/drivers/net/mlx5/mlx5_vlan.c
@@ -0,0 +1,232 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2015 6WIND S.A.
+ *   Copyright 2015 Mellanox.
+ *
+ *   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.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <stddef.h>
+#include <errno.h>
+#include <assert.h>
+#include <stdint.h>
+
+/* DPDK headers don't like -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <rte_ethdev.h>
+#include <rte_common.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+#include "mlx5_utils.h"
+#include "mlx5.h"
+#include "mlx5_autoconf.h"
+
+/**
+ * Configure a VLAN filter.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param vlan_id
+ *   VLAN ID to filter.
+ * @param on
+ *   Toggle filter.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+vlan_filter_set(struct rte_eth_dev *dev, uint16_t vlan_id, int on)
+{
+	struct priv *priv = dev->data->dev_private;
+	unsigned int i;
+
+	DEBUG("%p: %s VLAN filter ID %" PRIu16,
+	      (void *)dev, (on ? "enable" : "disable"), vlan_id);
+	assert(priv->vlan_filter_n <= RTE_DIM(priv->vlan_filter));
+	for (i = 0; (i != priv->vlan_filter_n); ++i)
+		if (priv->vlan_filter[i] == vlan_id)
+			break;
+	/* Check if there's room for another VLAN filter. */
+	if (i == RTE_DIM(priv->vlan_filter))
+		return ENOMEM;
+	if (i < priv->vlan_filter_n) {
+		assert(priv->vlan_filter_n != 0);
+		/* Enabling an existing VLAN filter has no effect. */
+		if (on)
+			return 0;
+		/* Remove VLAN filter from list. */
+		--priv->vlan_filter_n;
+		memmove(&priv->vlan_filter[i],
+			&priv->vlan_filter[i + 1],
+			priv->vlan_filter_n - i);
+		priv->vlan_filter[priv->vlan_filter_n] = 0;
+	} else {
+		assert(i == priv->vlan_filter_n);
+		/* Disabling an unknown VLAN filter has no effect. */
+		if (!on)
+			return 0;
+		/* Add new VLAN filter. */
+		priv->vlan_filter[priv->vlan_filter_n] = vlan_id;
+		++priv->vlan_filter_n;
+	}
+	/* Rehash flows in all hash RX queues. */
+	priv_mac_addrs_disable(priv);
+	priv_special_flow_disable_all(priv);
+	return priv_rehash_flows(priv);
+}
+
+/**
+ * DPDK callback to configure a VLAN filter.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param vlan_id
+ *   VLAN ID to filter.
+ * @param on
+ *   Toggle filter.
+ *
+ * @return
+ *   0 on success, negative errno value on failure.
+ */
+int
+mlx5_vlan_filter_set(struct rte_eth_dev *dev, uint16_t vlan_id, int on)
+{
+	struct priv *priv = dev->data->dev_private;
+	int ret;
+
+	priv_lock(priv);
+	ret = vlan_filter_set(dev, vlan_id, on);
+	priv_unlock(priv);
+	assert(ret >= 0);
+	return -ret;
+}
+
+/**
+ * Set/reset VLAN stripping for a specific queue.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param idx
+ *   RX queue index.
+ * @param on
+ *   Enable/disable VLAN stripping.
+ */
+static void
+priv_vlan_strip_queue_set(struct priv *priv, uint16_t idx, int on)
+{
+	struct rxq *rxq = (*priv->rxqs)[idx];
+	struct rxq_ctrl *rxq_ctrl = container_of(rxq, struct rxq_ctrl, rxq);
+	struct ibv_exp_wq_attr mod;
+	uint16_t vlan_offloads =
+		(on ? IBV_EXP_RECEIVE_WQ_CVLAN_STRIP : 0) |
+		0;
+	int err;
+
+	DEBUG("set VLAN offloads 0x%x for port %d queue %d",
+	      vlan_offloads, rxq->port_id, idx);
+	mod = (struct ibv_exp_wq_attr){
+		.attr_mask = IBV_EXP_WQ_ATTR_VLAN_OFFLOADS,
+		.vlan_offloads = vlan_offloads,
+	};
+
+	err = ibv_exp_modify_wq(rxq_ctrl->wq, &mod);
+	if (err) {
+		ERROR("%p: failed to modified stripping mode: %s",
+		      (void *)priv, strerror(err));
+		return;
+	}
+
+	/* Update related bits in RX queue. */
+	rxq->vlan_strip = !!on;
+}
+
+/**
+ * Callback to set/reset VLAN stripping for a specific queue.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param queue
+ *   RX queue index.
+ * @param on
+ *   Enable/disable VLAN stripping.
+ */
+void
+mlx5_vlan_strip_queue_set(struct rte_eth_dev *dev, uint16_t queue, int on)
+{
+	struct priv *priv = dev->data->dev_private;
+
+	/* Validate hw support */
+	if (!priv->hw_vlan_strip) {
+		ERROR("VLAN stripping is not supported");
+		return;
+	}
+
+	/* Validate queue number */
+	if (queue >= priv->rxqs_n) {
+		ERROR("VLAN stripping, invalid queue number %d", queue);
+		return;
+	}
+
+	priv_lock(priv);
+	priv_vlan_strip_queue_set(priv, queue, on);
+	priv_unlock(priv);
+}
+
+/**
+ * Callback to set/reset VLAN offloads for a port.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param mask
+ *   VLAN offload bit mask.
+ */
+void
+mlx5_vlan_offload_set(struct rte_eth_dev *dev, int mask)
+{
+	struct priv *priv = dev->data->dev_private;
+	unsigned int i;
+
+	if (mask & ETH_VLAN_STRIP_MASK) {
+		int hw_vlan_strip = !!dev->data->dev_conf.rxmode.hw_vlan_strip;
+
+		if (!priv->hw_vlan_strip) {
+			ERROR("VLAN stripping is not supported");
+			return;
+		}
+
+		/* Run on every RX queue and set/reset VLAN stripping. */
+		priv_lock(priv);
+		for (i = 0; (i != priv->rxqs_n); i++)
+			priv_vlan_strip_queue_set(priv, i, hw_vlan_strip);
+		priv_unlock(priv);
+	}
+}