[dpdk-dev,2/2] mlx4: new poll mode driver

Message ID 1422544846-10697-3-git-send-email-adrien.mazarguil@6wind.com (mailing list archive)
State Superseded, archived
Headers

Commit Message

Adrien Mazarguil Jan. 29, 2015, 3:20 p.m. UTC
This PMD manages all variants of Mellanox ConnectX-3 (EN 40, EN 10, Pro EN
40) as well as their virtual functions in SR-IOV context through IB Verbs
(libibverbs) and the dedicated user-space driver (libmlx4).

It is disabled by default due to dependencies on these libraries and only
supports Linux userland at the moment partly because /sys (sysfs) support is
required.

Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Olga Shern <olgas@mellanox.com>
---
 config/common_bsdapp         |   11 +
 config/common_linuxapp       |   11 +
 lib/Makefile                 |    1 +
 lib/librte_pmd_mlx4/Makefile |  119 ++
 lib/librte_pmd_mlx4/mlx4.c   | 4739 ++++++++++++++++++++++++++++++++++++++++++
 lib/librte_pmd_mlx4/mlx4.h   |  166 ++
 mk/rte.app.mk                |    8 +
 7 files changed, 5055 insertions(+)
 create mode 100755 lib/librte_pmd_mlx4/Makefile
 create mode 100755 lib/librte_pmd_mlx4/mlx4.c
 create mode 100644 lib/librte_pmd_mlx4/mlx4.h
  

Comments

Stephen Hemminger Jan. 29, 2015, 3:37 p.m. UTC | #1
On Thu, 29 Jan 2015 16:20:46 +0100
Adrien Mazarguil <adrien.mazarguil@6wind.com> wrote:

> This PMD manages all variants of Mellanox ConnectX-3 (EN 40, EN 10, Pro EN
> 40) as well as their virtual functions in SR-IOV context through IB Verbs
> (libibverbs) and the dedicated user-space driver (libmlx4).

Are these available with an open source license or is this
just an enabler for closed source?
  
Adrien Mazarguil Jan. 29, 2015, 3:45 p.m. UTC | #2
On Thu, Jan 29, 2015 at 03:37:39PM +0000, Stephen Hemminger wrote:
> On Thu, 29 Jan 2015 16:20:46 +0100
> Adrien Mazarguil <adrien.mazarguil@6wind.com> wrote:
> 
> > This PMD manages all variants of Mellanox ConnectX-3 (EN 40, EN 10, Pro EN
> > 40) as well as their virtual functions in SR-IOV context through IB Verbs
> > (libibverbs) and the dedicated user-space driver (libmlx4).
> 
> Are these available with an open source license or is this
> just an enabler for closed source?

Yes, both libraries are BSD/GPL-licensed.
  

Patch

diff --git a/config/common_bsdapp b/config/common_bsdapp
index 9177db1..f7b3fe1 100644
--- a/config/common_bsdapp
+++ b/config/common_bsdapp
@@ -182,6 +182,17 @@  CONFIG_RTE_LIBRTE_I40E_QUEUE_NUM_PER_VM=4
 CONFIG_RTE_LIBRTE_I40E_ITR_INTERVAL=-1
 
 #
+# Compile burst-oriented Mellanox ConnectX-3 (MLX4) PMD
+#
+CONFIG_RTE_LIBRTE_MLX4_PMD=n
+CONFIG_RTE_LIBRTE_MLX4_DEBUG=n
+CONFIG_RTE_LIBRTE_MLX4_SGE_WR_N=4
+CONFIG_RTE_LIBRTE_MLX4_MAX_INLINE=0
+CONFIG_RTE_LIBRTE_MLX4_TX_MP_CACHE=8
+CONFIG_RTE_LIBRTE_MLX4_SOFT_COUNTERS=1
+CONFIG_RTE_LIBRTE_MLX4_COMPAT_VMWARE=1
+
+#
 # Compile burst-oriented Cisco ENIC PMD driver
 #
 CONFIG_RTE_LIBRTE_ENIC_PMD=y
diff --git a/config/common_linuxapp b/config/common_linuxapp
index 2f9643b..0cec157 100644
--- a/config/common_linuxapp
+++ b/config/common_linuxapp
@@ -180,6 +180,17 @@  CONFIG_RTE_LIBRTE_I40E_QUEUE_NUM_PER_VM=4
 CONFIG_RTE_LIBRTE_I40E_ITR_INTERVAL=-1
 
 #
+# Compile burst-oriented Mellanox ConnectX-3 (MLX4) PMD
+#
+CONFIG_RTE_LIBRTE_MLX4_PMD=n
+CONFIG_RTE_LIBRTE_MLX4_DEBUG=n
+CONFIG_RTE_LIBRTE_MLX4_SGE_WR_N=4
+CONFIG_RTE_LIBRTE_MLX4_MAX_INLINE=0
+CONFIG_RTE_LIBRTE_MLX4_TX_MP_CACHE=8
+CONFIG_RTE_LIBRTE_MLX4_SOFT_COUNTERS=1
+CONFIG_RTE_LIBRTE_MLX4_COMPAT_VMWARE=1
+
+#
 # Compile burst-oriented Cisco ENIC PMD driver
 #
 CONFIG_RTE_LIBRTE_ENIC_PMD=y
diff --git a/lib/Makefile b/lib/Makefile
index 0ffc982..9f1994d 100644
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -43,6 +43,7 @@  DIRS-$(CONFIG_RTE_LIBRTE_ETHER) += librte_ether
 DIRS-$(CONFIG_RTE_LIBRTE_E1000_PMD) += librte_pmd_e1000
 DIRS-$(CONFIG_RTE_LIBRTE_IXGBE_PMD) += librte_pmd_ixgbe
 DIRS-$(CONFIG_RTE_LIBRTE_I40E_PMD) += librte_pmd_i40e
+DIRS-$(CONFIG_RTE_LIBRTE_MLX4_PMD) += librte_pmd_mlx4
 DIRS-$(CONFIG_RTE_LIBRTE_ENIC_PMD) += librte_pmd_enic
 DIRS-$(CONFIG_RTE_LIBRTE_PMD_BOND) += librte_pmd_bond
 DIRS-$(CONFIG_RTE_LIBRTE_PMD_RING) += librte_pmd_ring
diff --git a/lib/librte_pmd_mlx4/Makefile b/lib/librte_pmd_mlx4/Makefile
new file mode 100755
index 0000000..9a6c891
--- /dev/null
+++ b/lib/librte_pmd_mlx4/Makefile
@@ -0,0 +1,119 @@ 
+#   BSD LICENSE
+#
+#   Copyright(c) 2012-2015 6WIND S.A.
+#   Copyright(c) 2012 Mellanox.
+#   All rights reserved.
+#
+#   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 $(RTE_SDK)/mk/rte.vars.mk
+
+# Library name.
+LIB = librte_pmd_mlx4.a
+
+# Sources.
+SRCS-$(CONFIG_RTE_LIBRTE_MLX4_PMD) += mlx4.c
+
+# Dependencies.
+DEPDIRS-$(CONFIG_RTE_LIBRTE_MLX4_PMD) += lib/librte_ether
+DEPDIRS-$(CONFIG_RTE_LIBRTE_MLX4_PMD) += lib/librte_mbuf
+DEPDIRS-$(CONFIG_RTE_LIBRTE_MLX4_PMD) += lib/librte_eal
+DEPDIRS-$(CONFIG_RTE_LIBRTE_MLX4_PMD) += lib/librte_mempool
+DEPDIRS-$(CONFIG_RTE_LIBRTE_MLX4_PMD) += lib/librte_malloc
+
+# Basic CFLAGS.
+CFLAGS += -O3
+CFLAGS += -std=gnu99 -Wall -Wextra
+CFLAGS += -g
+CFLAGS += -I.
+CFLAGS += -D_XOPEN_SOURCE=600
+CFLAGS += $(WERROR_FLAGS)
+
+# A few warnings cannot be avoided in external headers.
+CFLAGS += -Wno-error=cast-qual
+
+# DEBUG which is usually provided on the command-line may enable
+# CONFIG_RTE_LIBRTE_MLX4_DEBUG.
+ifeq ($(DEBUG),1)
+CONFIG_RTE_LIBRTE_MLX4_DEBUG := y
+endif
+
+# User-defined CFLAGS.
+ifeq ($(CONFIG_RTE_LIBRTE_MLX4_DEBUG),y)
+CFLAGS += -pedantic -UNDEBUG -DPEDANTIC
+else
+CFLAGS += -DNDEBUG -UPEDANTIC
+endif
+
+ifdef CONFIG_RTE_LIBRTE_MLX4_SGE_WR_N
+CFLAGS += -DMLX4_PMD_SGE_WR_N=$(CONFIG_RTE_LIBRTE_MLX4_SGE_WR_N)
+endif
+
+ifdef CONFIG_RTE_LIBRTE_MLX4_MAX_INLINE
+CFLAGS += -DMLX4_PMD_MAX_INLINE=$(CONFIG_RTE_LIBRTE_MLX4_MAX_INLINE)
+endif
+
+ifdef CONFIG_RTE_LIBRTE_MLX4_TX_MP_CACHE
+CFLAGS += -DMLX4_PMD_TX_MP_CACHE=$(CONFIG_RTE_LIBRTE_MLX4_TX_MP_CACHE)
+endif
+
+ifdef CONFIG_RTE_LIBRTE_MLX4_SOFT_COUNTERS
+CFLAGS += -DMLX4_PMD_SOFT_COUNTERS=$(CONFIG_RTE_LIBRTE_MLX4_SOFT_COUNTERS)
+endif
+
+ifdef CONFIG_RTE_LIBRTE_MLX4_COMPAT_VMWARE
+CFLAGS += -DMLX4_PMD_COMPAT_VMWARE=$(CONFIG_RTE_LIBRTE_MLX4_COMPAT_VMWARE)
+endif
+
+include $(RTE_SDK)/mk/rte.lib.mk
+
+# Generate and clean-up mlx4_autoconf.h.
+
+export CC CFLAGS CPPFLAGS EXTRA_CFLAGS EXTRA_CPPFLAGS
+export AUTO_CONFIG_CFLAGS = -Wno-error
+
+mlx4_autoconf.h: $(RTE_SDK)/scripts/auto-config-h.sh
+	$Q $(RM) -f -- '$@'
+	$Q sh -- '$<' '$@' \
+		RSS_SUPPORT \
+		infiniband/verbs.h \
+		enum IBV_EXP_DEVICE_UD_RSS
+	$Q sh -- '$<' '$@' \
+		INLINE_RECV \
+		infiniband/verbs.h \
+		enum IBV_EXP_DEVICE_ATTR_INLINE_RECV_SZ
+	$Q sh -- '$<' '$@' \
+		SEND_RAW_WR_SUPPORT \
+		infiniband/verbs.h \
+		type 'struct ibv_send_wr_raw'
+
+mlx4.o: mlx4_autoconf.h
+
+clean_mlx4: FORCE
+	$Q rm -f -- mlx4_autoconf.h
+
+clean: clean_mlx4
diff --git a/lib/librte_pmd_mlx4/mlx4.c b/lib/librte_pmd_mlx4/mlx4.c
new file mode 100755
index 0000000..71c75ad
--- /dev/null
+++ b/lib/librte_pmd_mlx4/mlx4.c
@@ -0,0 +1,4739 @@ 
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright(c) 2012-2015 6WIND S.A.
+ *   Copyright(c) 2012 Mellanox.
+ *   All rights reserved.
+ *
+ *   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.
+ */
+
+/*
+ * Known limitations:
+ * - Multiple RX VLAN filters can be configured, but only the first one
+ *   works properly.
+ * - RSS hash key and options cannot be modified.
+ * - Hardware counters aren't implemented.
+ */
+
+/* System headers. */
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <inttypes.h>
+#include <string.h>
+#include <errno.h>
+#include <unistd.h>
+#include <limits.h>
+#include <assert.h>
+#include <arpa/inet.h>
+#include <net/if.h>
+#include <dirent.h>
+#include <sys/ioctl.h>
+#include <sys/socket.h>
+#include <netinet/in.h>
+#include <linux/if.h>
+#include <linux/ethtool.h>
+#include <linux/sockios.h>
+
+/* DPDK headers don't like -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <rte_config.h>
+#include <rte_ether.h>
+#include <rte_ethdev.h>
+#include <rte_dev.h>
+#include <rte_mbuf.h>
+#include <rte_errno.h>
+#include <rte_mempool.h>
+#include <rte_prefetch.h>
+#include <rte_malloc.h>
+#include <rte_spinlock.h>
+#include <rte_atomic.h>
+#include <rte_version.h>
+#include <rte_log.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+/* 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
+
+/* Generated configuration header. */
+#include "mlx4_autoconf.h"
+
+/* PMD header. */
+#include "mlx4.h"
+
+/* Runtime logging through RTE_LOG() is enabled when not in debugging mode.
+ * Intermediate LOG_*() macros add the required end-of-line characters. */
+#ifndef NDEBUG
+#define INFO(...) DEBUG(__VA_ARGS__)
+#define WARN(...) DEBUG(__VA_ARGS__)
+#define ERROR(...) DEBUG(__VA_ARGS__)
+#else
+#define LOG__(level, m, ...) \
+	RTE_LOG(level, PMD, MLX4_DRIVER_NAME ": " m "%c", __VA_ARGS__)
+#define LOG_(level, ...) LOG__(level, __VA_ARGS__, '\n')
+#define INFO(...) LOG_(INFO, __VA_ARGS__)
+#define WARN(...) LOG_(WARNING, __VA_ARGS__)
+#define ERROR(...) LOG_(ERR, __VA_ARGS__)
+#endif
+
+/* 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
+
+/* Work Request ID data type (64 bit). */
+typedef union {
+	struct {
+		uint32_t id;
+		uint16_t offset;
+	} data;
+	uint64_t raw;
+} wr_id_t;
+
+#define WR_ID(o) ((wr_id_t *)&(o))->data
+
+/* Compile-time check. */
+static inline void wr_id_t_check(void)
+{
+	wr_id_t check[1 + (2 * -!(sizeof(wr_id_t) == sizeof(uint64_t)))];
+
+	(void)check;
+	(void)wr_id_t_check;
+}
+
+/* If raw send operations are available, use them since they are faster. */
+#ifdef SEND_RAW_WR_SUPPORT
+typedef struct ibv_send_wr_raw mlx4_send_wr_t;
+#define mlx4_post_send ibv_post_send_raw
+#else
+typedef struct ibv_send_wr mlx4_send_wr_t;
+#define mlx4_post_send ibv_post_send
+#endif
+
+struct mlx4_rxq_stats {
+	unsigned int idx; /**< Mapping index. */
+#ifdef MLX4_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 mlx4_txq_stats {
+	unsigned int idx; /**< Mapping index. */
+#ifdef MLX4_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. */
+};
+
+/* RX element (scattered packets). */
+struct rxq_elt_sp {
+	struct ibv_recv_wr wr; /* Work Request. */
+	struct ibv_sge sges[MLX4_PMD_SGE_WR_N]; /* Scatter/Gather Elements. */
+	struct rte_mbuf *bufs[MLX4_PMD_SGE_WR_N]; /* SGEs buffers. */
+};
+
+/* RX element. */
+struct rxq_elt {
+	struct ibv_recv_wr wr; /* Work Request. */
+	struct ibv_sge sge; /* Scatter/Gather Element. */
+	/* mbuf pointer is derived from WR_ID(wr.wr_id).offset. */
+};
+
+/* RX queue descriptor. */
+struct rxq {
+	struct priv *priv; /* Back pointer to private data. */
+	struct rte_mempool *mp; /* Memory Pool for allocations. */
+	struct ibv_mr *mr; /* Memory Region (for mp). */
+	struct ibv_cq *cq; /* Completion Queue. */
+	struct ibv_qp *qp; /* Queue Pair. */
+	/*
+	 * There is exactly one flow configured per MAC address. Each flow
+	 * may contain several specifications, one per configured VLAN ID.
+	 */
+	BITFIELD_DECLARE(mac_configured, uint32_t, MLX4_MAX_MAC_ADDRESSES);
+	struct ibv_exp_flow *mac_flow[MLX4_MAX_MAC_ADDRESSES];
+	struct ibv_exp_flow *promisc_flow; /* Promiscuous flow. */
+	struct ibv_exp_flow *allmulti_flow; /* Multicast flow. */
+	unsigned int port_id; /* Port ID for incoming packets. */
+	unsigned int elts_n; /* (*elts)[] length. */
+	union {
+		struct rxq_elt_sp (*sp)[]; /* Scattered RX elements. */
+		struct rxq_elt (*no_sp)[]; /* RX elements. */
+	} elts;
+	unsigned int sp:1; /* Use scattered RX elements. */
+	uint32_t mb_len; /* Length of a mp-issued mbuf. */
+	struct mlx4_rxq_stats stats; /* RX queue counters. */
+	unsigned int socket; /* CPU socket ID for allocations. */
+};
+
+/* TX element. */
+struct txq_elt {
+	mlx4_send_wr_t wr; /* Work Request. */
+	struct ibv_sge sges[MLX4_PMD_SGE_WR_N]; /* Scatter/Gather Elements. */
+	/* mbuf pointer is derived from WR_ID(wr.wr_id).offset. */
+};
+
+/* Linear buffer type. It is used when transmitting buffers with too many
+ * segments that do not fit the hardware queue (see max_send_sge).
+ * Extra segments are copied (linearized) in such buffers, replacing the
+ * last SGE during TX.
+ * The size is arbitrary but large enough to hold a jumbo frame with
+ * 8 segments considering mbuf.buf_len is about 2048 bytes. */
+typedef uint8_t linear_t[16384];
+
+/* TX queue descriptor. */
+struct txq {
+	struct priv *priv; /* Back pointer to private data. */
+	struct {
+		struct rte_mempool *mp; /* Cached Memory Pool. */
+		struct ibv_mr *mr; /* Memory Region (for mp). */
+		uint32_t lkey; /* mr->lkey */
+	} mp2mr[MLX4_PMD_TX_MP_CACHE]; /* MP to MR translation table. */
+	struct ibv_cq *cq; /* Completion Queue. */
+	struct ibv_qp *qp; /* Queue Pair. */
+#if MLX4_PMD_MAX_INLINE > 0
+	uint32_t max_inline; /* Max inline send size <= MLX4_PMD_MAX_INLINE. */
+#endif
+	unsigned int elts_n; /* (*elts)[] length. */
+	struct txq_elt (*elts)[]; /* TX elements. */
+	unsigned int elts_head; /* Current index in (*elts)[]. */
+	unsigned int elts_tail; /* First element awaiting completion. */
+	unsigned int elts_comp; /* Number of completion requests. */
+	struct mlx4_txq_stats stats; /* TX queue counters. */
+	linear_t (*elts_linear)[]; /* Linearized buffers. */
+	struct ibv_mr *mr_linear; /* Memory Region for linearized buffers. */
+	unsigned int socket; /* CPU socket ID for allocations. */
+};
+
+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_port_attr port_attr; /* Physical port 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[MLX4_MAX_MAC_ADDRESSES];
+	BITFIELD_DECLARE(mac_configured, uint32_t, MLX4_MAX_MAC_ADDRESSES);
+	/* VLAN filters. */
+	struct {
+		unsigned int enabled:1; /* If enabled. */
+		unsigned int id:12; /* VLAN ID (0-4095). */
+	} vlan_filter[MLX4_MAX_VLAN_IDS]; /* VLAN filters table. */
+	/* Device properties. */
+	uint16_t mtu; /* Configured MTU. */
+	uint8_t port; /* Physical port number. */
+	unsigned int started:1; /* Device started, flows enabled. */
+	unsigned int promisc:1; /* Device in promiscuous mode. */
+	unsigned int promisc_ok:1; /* Promiscuous flow is supported. */
+	unsigned int allmulti:1; /* Device receives all multicast packets. */
+	unsigned int hw_qpg:1; /* QP groups are supported. */
+	unsigned int hw_tss:1; /* TSS is supported. */
+	unsigned int hw_rss:1; /* RSS is supported. */
+	unsigned int rss:1; /* RSS is enabled. */
+#ifdef MLX4_COMPAT_VMWARE
+	unsigned int vmware:1; /* Use VMware compatibility. */
+#endif
+	unsigned int vf:1; /* This is a VF device. */
+#ifdef INLINE_RECV
+	unsigned int inl_recv_size; /* Inline recv size */
+#endif
+	unsigned int max_rss_tbl_sz; /* Maximum number of RSS queues. */
+	/* RX/TX queues. */
+	struct rxq rxq_parent; /* Parent queue when RSS is enabled. */
+	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. */
+	rte_spinlock_t lock; /* Lock for control functions. */
+};
+
+/**
+ * Lock private structure to protect it from concurrent access in the
+ * control path.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ */
+static void
+priv_lock(struct priv *priv)
+{
+	rte_spinlock_lock(&priv->lock);
+}
+
+/**
+ * Unlock private structure.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ */
+static void
+priv_unlock(struct priv *priv)
+{
+	rte_spinlock_unlock(&priv->lock);
+}
+
+/* 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__)
+
+/**
+ * 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.
+ */
+static int
+priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE])
+{
+	int ret = -1;
+	DIR *dir;
+	struct dirent *dent;
+
+	{
+		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_id;
+		int r;
+
+		if ((name[0] == '.') &&
+		    ((name[1] == '\0') ||
+		     ((name[1] == '.') && (name[2] == '\0'))))
+			continue;
+
+		MKSTR(path, "%s/device/net/%s/dev_id",
+		      priv->ctx->device->ibdev_path, name);
+
+		file = fopen(path, "rb");
+		if (file == NULL)
+			continue;
+		r = fscanf(file, "%x", &dev_id);
+		fclose(file);
+		if ((r == 1) && (dev_id == (priv->port - 1u))) {
+			snprintf(*ifname, sizeof(*ifname), "%s", name);
+			ret = 0;
+			break;
+		}
+	}
+	closedir(dir);
+	return ret;
+}
+
+/**
+ * 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.
+ */
+static 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;
+}
+
+/**
+ * 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.
+ */
+static 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)
+{
+	return priv_set_sysfs_ulong(priv, "mtu", mtu);
+}
+
+/**
+ * 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.
+ */
+static 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;
+	return priv_set_sysfs_ulong(priv, "flags", tmp);
+}
+
+/* Device configuration. */
+
+static int
+rxq_setup(struct rte_eth_dev *dev, struct rxq *rxq, uint16_t desc,
+	  unsigned int socket, const struct rte_eth_rxconf *conf,
+	  struct rte_mempool *mp);
+
+static void
+rxq_cleanup(struct rxq *rxq);
+
+/**
+ * Ethernet device configuration.
+ *
+ * Prepare the driver for a given number of TX and RX queues.
+ * Allocate parent RSS queue when several RX queues are requested.
+ *
+ * @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 tmp;
+	int ret;
+
+	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->rxqs_n)
+		return 0;
+	INFO("%p: RX queues number update: %u -> %u",
+	     (void *)dev, priv->rxqs_n, rxqs_n);
+	/* If RSS is enabled, disable it first. */
+	if (priv->rss) {
+		unsigned int i;
+
+		/* Only if there are no remaining child RX queues. */
+		for (i = 0; (i != priv->rxqs_n); ++i)
+			if ((*priv->rxqs)[i] != NULL)
+				return EINVAL;
+		rxq_cleanup(&priv->rxq_parent);
+		priv->rss = 0;
+		priv->rxqs_n = 0;
+	}
+	if (rxqs_n <= 1) {
+		/* Nothing else to do. */
+		priv->rxqs_n = rxqs_n;
+		return 0;
+	}
+	/* Allocate a new RSS parent queue if supported by hardware. */
+	if (!priv->hw_rss) {
+		ERROR("%p: only a single RX queue can be configured when"
+		      " hardware doesn't support RSS",
+		      (void *)dev);
+		return EINVAL;
+	}
+	/* Fail if hardware doesn't support that many RSS queues. */
+	if (rxqs_n >= priv->max_rss_tbl_sz) {
+		ERROR("%p: only %u RX queues can be configured for RSS",
+		      (void *)dev, priv->max_rss_tbl_sz);
+		return EINVAL;
+	}
+	priv->rss = 1;
+	tmp = priv->rxqs_n;
+	priv->rxqs_n = rxqs_n;
+	ret = rxq_setup(dev, &priv->rxq_parent, 0, 0, NULL, NULL);
+	if (!ret)
+		return 0;
+	/* Failure, rollback. */
+	priv->rss = 0;
+	priv->rxqs_n = tmp;
+	assert(ret > 0);
+	return ret;
+}
+
+/**
+ * DPDK callback for Ethernet device configuration.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ *
+ * @return
+ *   0 on success, negative errno value on failure.
+ */
+static int
+mlx4_dev_configure(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	int ret;
+
+	priv_lock(priv);
+	ret = dev_configure(dev);
+	assert(ret >= 0);
+	priv_unlock(priv);
+	return -ret;
+}
+
+/* TX queues handling. */
+
+/**
+ * Allocate TX queue elements.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param elts_n
+ *   Number of elements to allocate.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+txq_alloc_elts(struct txq *txq, unsigned int elts_n)
+{
+	unsigned int i;
+	struct txq_elt (*elts)[elts_n] =
+		rte_calloc_socket("TXQ", 1, sizeof(*elts), 0, txq->socket);
+	linear_t (*elts_linear)[elts_n] =
+		rte_calloc_socket("TXQ", 1, sizeof(*elts_linear), 0,
+				  txq->socket);
+	struct ibv_mr *mr_linear = NULL;
+	int ret = 0;
+
+	if ((elts == NULL) || (elts_linear == NULL)) {
+		ERROR("%p: can't allocate packets array", (void *)txq);
+		ret = ENOMEM;
+		goto error;
+	}
+	mr_linear =
+		ibv_reg_mr(txq->priv->pd, elts_linear, sizeof(*elts_linear),
+			   (IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE));
+	if (mr_linear == NULL) {
+		ERROR("%p: unable to configure MR, ibv_reg_mr() failed",
+		      (void *)txq);
+		ret = EINVAL;
+		goto error;
+	}
+	for (i = 0; (i != elts_n); ++i) {
+		struct txq_elt *elt = &(*elts)[i];
+		mlx4_send_wr_t *wr = &elt->wr;
+
+		/* Configure WR. */
+		WR_ID(wr->wr_id).id = i;
+		WR_ID(wr->wr_id).offset = 0;
+		wr->sg_list = &elt->sges[0];
+		wr->opcode = IBV_WR_SEND;
+		/* Other fields are updated during TX. */
+	}
+	DEBUG("%p: allocated and configured %u WRs", (void *)txq, elts_n);
+	txq->elts_n = elts_n;
+	txq->elts = elts;
+	txq->elts_head = 0;
+	txq->elts_tail = 0;
+	txq->elts_comp = 0;
+	txq->elts_linear = elts_linear;
+	txq->mr_linear = mr_linear;
+	assert(ret == 0);
+	return 0;
+error:
+	if (mr_linear != NULL)
+		claim_zero(ibv_dereg_mr(mr_linear));
+	if (elts_linear != NULL)
+		rte_free(elts_linear);
+	if (elts != NULL)
+		rte_free(elts);
+	DEBUG("%p: failed, freed everything", (void *)txq);
+	assert(ret > 0);
+	return ret;
+}
+
+/**
+ * Free TX queue elements.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ */
+static void
+txq_free_elts(struct txq *txq)
+{
+	unsigned int i;
+	unsigned int elts_n = txq->elts_n;
+	struct txq_elt (*elts)[elts_n] = txq->elts;
+	linear_t (*elts_linear)[elts_n] = txq->elts_linear;
+	struct ibv_mr *mr_linear = txq->mr_linear;
+
+	DEBUG("%p: freeing WRs", (void *)txq);
+	txq->elts_n = 0;
+	txq->elts = NULL;
+	txq->elts_linear = NULL;
+	txq->mr_linear = NULL;
+	if (mr_linear != NULL)
+		claim_zero(ibv_dereg_mr(mr_linear));
+	if (elts_linear != NULL)
+		rte_free(elts_linear);
+	if (elts == NULL)
+		return;
+	for (i = 0; (i != elemof(*elts)); ++i) {
+		struct txq_elt *elt = &(*elts)[i];
+
+		if (WR_ID(elt->wr.wr_id).offset == 0)
+			continue;
+		rte_pktmbuf_free((void *)(elt->sges[0].addr -
+					  WR_ID(elt->wr.wr_id).offset));
+	}
+	rte_free(elts);
+}
+
+
+/**
+ * Clean up a TX queue.
+ *
+ * Destroy objects, free allocated memory and reset the structure for reuse.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ */
+static void
+txq_cleanup(struct txq *txq)
+{
+	size_t i;
+
+	DEBUG("cleaning up %p", (void *)txq);
+	txq_free_elts(txq);
+	if (txq->qp != NULL)
+		claim_zero(ibv_destroy_qp(txq->qp));
+	if (txq->cq != NULL)
+		claim_zero(ibv_destroy_cq(txq->cq));
+	for (i = 0; (i != elemof(txq->mp2mr)); ++i) {
+		if (txq->mp2mr[i].mp == NULL)
+			break;
+		assert(txq->mp2mr[i].mr != NULL);
+		claim_zero(ibv_dereg_mr(txq->mp2mr[i].mr));
+	}
+	memset(txq, 0, sizeof(*txq));
+}
+
+/**
+ * Manage TX completions.
+ *
+ * When sending a burst, mlx4_tx_burst() posts several WRs.
+ * To improve performance, a completion event is only required for the last of
+ * them. Doing so discards completion information for other WRs, but this
+ * information would not be used anyway.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ *
+ * @return
+ *   0 on success, -1 on failure.
+ */
+static int
+txq_complete(struct txq *txq)
+{
+	unsigned int elts_comp = txq->elts_comp;
+	unsigned int elts_tail;
+	const unsigned int elts_n = txq->elts_n;
+	struct ibv_wc wcs[elts_comp];
+	int wcs_n;
+
+	if (unlikely(elts_comp == 0))
+		return 0;
+#ifdef DEBUG_SEND
+	DEBUG("%p: processing %u work requests completions",
+	      (void *)txq, elts_comp);
+#endif
+	wcs_n = ibv_poll_cq(txq->cq, elts_comp, wcs);
+	if (unlikely(wcs_n == 0))
+		return 0;
+	if (unlikely(wcs_n < 0)) {
+		DEBUG("%p: ibv_poll_cq() failed (wcs_n=%d)",
+		      (void *)txq, wcs_n);
+		return -1;
+	}
+	elts_comp -= wcs_n;
+	assert(elts_comp <= txq->elts_comp);
+	/*
+	 * Work Completion ID contains the associated element index in
+	 * (*txq->elts)[]. Since WCs are returned in order, we only need to
+	 * look at the last WC to clear older Work Requests.
+	 *
+	 * Assume WC status is successful as nothing can be done about it
+	 * anyway.
+	 */
+	elts_tail = WR_ID(wcs[wcs_n - 1].wr_id).id;
+	/* Consume the last WC. */
+	if (++elts_tail >= elts_n)
+		elts_tail = 0;
+	txq->elts_tail = elts_tail;
+	txq->elts_comp = elts_comp;
+	return 0;
+}
+
+/**
+ * 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 uint32_t
+txq_mp2mr(struct txq *txq, struct rte_mempool *mp)
+{
+	unsigned int i;
+	struct ibv_mr *mr;
+
+	for (i = 0; (i != elemof(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(txq->mp2mr[i].mr->lkey == txq->mp2mr[i].lkey);
+			return txq->mp2mr[i].lkey;
+		}
+	}
+	/* Add a new entry, register MR first. */
+	DEBUG("%p: discovered new memory pool %p", (void *)txq, (void *)mp);
+	mr = ibv_reg_mr(txq->priv->pd,
+		        (void *)mp->elt_va_start,
+		        (mp->elt_va_end - mp->elt_va_start),
+			(IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE));
+	if (unlikely(mr == NULL)) {
+		DEBUG("%p: unable to configure MR, ibv_reg_mr() failed.",
+		      (void *)txq);
+		return (uint32_t)-1;
+	}
+	if (unlikely(i == elemof(txq->mp2mr))) {
+		/* Table is full, remove oldest entry. */
+		DEBUG("%p: MR <-> MP table full, dropping oldest entry.",
+		      (void *)txq);
+		--i;
+		claim_zero(ibv_dereg_mr(txq->mp2mr[i].mr));
+		memmove(&txq->mp2mr[0], &txq->mp2mr[1],
+			(sizeof(txq->mp2mr) - sizeof(txq->mp2mr[0])));
+	}
+	/* Store the new entry. */
+	txq->mp2mr[i].mp = mp;
+	txq->mp2mr[i].mr = mr;
+	txq->mp2mr[i].lkey = mr->lkey;
+	DEBUG("%p: new MR lkey for MP %p: 0x%08" PRIu32,
+	      (void *)txq, (void *)mp, txq->mp2mr[i].lkey);
+	return txq->mp2mr[i].lkey;
+}
+
+/**
+ * Copy scattered mbuf contents to a single linear buffer.
+ *
+ * @param[out] linear
+ *   Linear output buffer.
+ * @param[in] buf
+ *   Scattered input buffer.
+ *
+ * @return
+ *   Number of bytes copied to the output buffer or 0 if not large enough.
+ */
+static unsigned int
+linearize_mbuf(linear_t *linear, struct rte_mbuf *buf)
+{
+	unsigned int size = 0;
+	unsigned int offset;
+
+	do {
+		unsigned int len = DATA_LEN(buf);
+
+		offset = size;
+		size += len;
+		if (unlikely(size > sizeof(*linear)))
+			return 0;
+		memcpy(&(*linear)[offset],
+		       rte_pktmbuf_mtod(buf, uint8_t *),
+		       len);
+		buf = NEXT(buf);
+	}
+	while (buf != NULL);
+	return size;
+}
+
+/**
+ * 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).
+ */
+static uint16_t
+mlx4_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+	struct txq *txq = (struct txq *)dpdk_txq;
+	mlx4_send_wr_t head;
+	mlx4_send_wr_t **wr_next = &head.next;
+	mlx4_send_wr_t *bad_wr;
+	unsigned int elts_head = txq->elts_head;
+	const unsigned int elts_tail = txq->elts_tail;
+	const unsigned int elts_n = txq->elts_n;
+	unsigned int i;
+	unsigned int max;
+	int err;
+
+	txq_complete(txq);
+	max = (elts_n - (elts_head - elts_tail));
+	if (max > elts_n)
+		max -= elts_n;
+	assert(max >= 1);
+	assert(max <= elts_n);
+	/* Always leave one free entry in the ring. */
+	--max;
+	if (max == 0)
+		return 0;
+	if (max > pkts_n)
+		max = pkts_n;
+	for (i = 0; (i != max); ++i) {
+		struct rte_mbuf *buf = pkts[i];
+		struct txq_elt *elt = &(*txq->elts)[elts_head];
+		mlx4_send_wr_t *wr = &elt->wr;
+		unsigned int segs = NB_SEGS(buf);
+#if (MLX4_PMD_MAX_INLINE > 0) || defined(MLX4_PMD_SOFT_COUNTERS)
+		unsigned int sent_size = 0;
+#endif
+		unsigned int j;
+		int linearize = 0;
+
+		/* Clean up old buffer. */
+		if (likely(WR_ID(wr->wr_id).offset != 0)) {
+			struct rte_mbuf *tmp = (void *)
+				(elt->sges[0].addr - WR_ID(wr->wr_id).offset);
+
+			/* Faster than rte_pktmbuf_free(). */
+			do {
+				struct rte_mbuf *next = NEXT(tmp);
+
+				rte_pktmbuf_free_seg(tmp);
+				tmp = next;
+			}
+			while (tmp != NULL);
+		}
+#ifndef NDEBUG
+		/* For assert(). */
+		WR_ID(wr->wr_id).offset = 0;
+		for (j = 0; ((int)j < wr->num_sge); ++j) {
+			elt->sges[j].addr = 0;
+			elt->sges[j].length = 0;
+			elt->sges[j].lkey = 0;
+		}
+		wr->next = NULL;
+		wr->num_sge = 0;
+#endif
+		/* Sanity checks, most of which are only relevant with
+		 * debugging enabled. */
+		assert(WR_ID(wr->wr_id).id == elts_head);
+		assert(WR_ID(wr->wr_id).offset == 0);
+		assert(wr->next == NULL);
+		assert(wr->sg_list == &elt->sges[0]);
+		assert(wr->num_sge == 0);
+		assert(wr->opcode == IBV_WR_SEND);
+		/* When there are too many segments, extra segments are
+		 * linearized in the last SGE. */
+		if (unlikely(segs > elemof(elt->sges))) {
+			segs = (elemof(elt->sges) - 1);
+			linearize = 1;
+		}
+		/* Set WR fields. */
+		assert(((uintptr_t)rte_pktmbuf_mtod(buf, char *) -
+			(uintptr_t)buf) <= 0xffff);
+		WR_ID(wr->wr_id).offset =
+			((uintptr_t)rte_pktmbuf_mtod(buf, char *) -
+			 (uintptr_t)buf);
+		wr->num_sge = segs;
+		/* Register segments as SGEs. */
+		for (j = 0; (j != segs); ++j) {
+			struct ibv_sge *sge = &elt->sges[j];
+			uint32_t lkey;
+
+			/* Retrieve Memory Region key for this memory pool. */
+			lkey = txq_mp2mr(txq, buf->pool);
+			if (unlikely(lkey == (uint32_t)-1)) {
+				/* MR does not exist. */
+				DEBUG("%p: unable to get MP <-> MR"
+				      " association", (void *)txq);
+				/* Clean up TX element. */
+				WR_ID(elt->wr.wr_id).offset = 0;
+#ifndef NDEBUG
+				/* For assert(). */
+				while (j) {
+					--j;
+					--sge;
+					sge->addr = 0;
+					sge->length = 0;
+					sge->lkey = 0;
+				}
+				wr->num_sge = 0;
+#endif
+				goto stop;
+			}
+			/* Sanity checks, only relevant with debugging
+			 * enabled. */
+			assert(sge->addr == 0);
+			assert(sge->length == 0);
+			assert(sge->lkey == 0);
+			/* Update SGE. */
+			sge->addr = (uintptr_t)rte_pktmbuf_mtod(buf, char *);
+			if (txq->priv->vf)
+				rte_prefetch0((volatile void *)sge->addr);
+			sge->length = DATA_LEN(buf);
+			sge->lkey = lkey;
+#if (MLX4_PMD_MAX_INLINE > 0) || defined(MLX4_PMD_SOFT_COUNTERS)
+			sent_size += sge->length;
+#endif
+			buf = NEXT(buf);
+		}
+		/* If buf is not NULL here and is not going to be linearized,
+		 * nb_segs is not valid. */
+		assert(j == segs);
+		assert((buf == NULL) || (linearize));
+		/* Linearize extra segments. */
+		if (linearize) {
+			struct ibv_sge *sge = &elt->sges[segs];
+			linear_t *linear = &(*txq->elts_linear)[elts_head];
+			unsigned int size = linearize_mbuf(linear, buf);
+
+			assert(segs == (elemof(elt->sges) - 1));
+			if (size == 0) {
+				/* Invalid packet. */
+				DEBUG("%p: packet too large to be linearized.",
+				      (void *)txq);
+				/* Clean up TX element. */
+				WR_ID(elt->wr.wr_id).offset = 0;
+#ifndef NDEBUG
+				/* For assert(). */
+				while (j) {
+					--j;
+					--sge;
+					sge->addr = 0;
+					sge->length = 0;
+					sge->lkey = 0;
+				}
+				wr->num_sge = 0;
+#endif
+				goto stop;
+			}
+			/* If MLX4_PMD_SGE_WR_N is 1, free mbuf immediately
+			 * and clear offset from WR ID. */
+			if (elemof(elt->sges) == 1) {
+				do {
+					struct rte_mbuf *next = NEXT(buf);
+
+					rte_pktmbuf_free_seg(buf);
+					buf = next;
+				}
+				while (buf != NULL);
+				WR_ID(wr->wr_id).offset = 0;
+			}
+			/* Set WR fields and fill SGE with linear buffer. */
+			++wr->num_sge;
+			/* Sanity checks, only relevant with debugging
+			 * enabled. */
+			assert(sge->addr == 0);
+			assert(sge->length == 0);
+			assert(sge->lkey == 0);
+			/* Update SGE. */
+			sge->addr = (uintptr_t)&(*linear)[0];
+			sge->length = size;
+			sge->lkey = txq->mr_linear->lkey;
+#if (MLX4_PMD_MAX_INLINE > 0) || defined(MLX4_PMD_SOFT_COUNTERS)
+			sent_size += size;
+#endif
+		}
+		/* Link WRs together for ibv_post_send(). */
+		*wr_next = wr;
+		wr_next = &wr->next;
+#if MLX4_PMD_MAX_INLINE > 0
+		if (sent_size <= txq->max_inline)
+			wr->send_flags = IBV_SEND_INLINE;
+		else
+#endif
+			wr->send_flags = 0;
+		if (++elts_head >= elts_n)
+			elts_head = 0;
+#ifdef MLX4_PMD_SOFT_COUNTERS
+		/* Increment sent bytes counter. */
+		txq->stats.obytes += sent_size;
+#endif
+	}
+stop:
+	/* Take a shortcut if nothing must be sent. */
+	if (unlikely(i == 0))
+		return 0;
+#ifdef MLX4_PMD_SOFT_COUNTERS
+	/* Increment sent packets counter. */
+	txq->stats.opackets += i;
+#endif
+	*wr_next = NULL;
+	/* The last WR is the only one asking for a completion event. */
+	containerof(wr_next, mlx4_send_wr_t, next)->
+		send_flags |= IBV_SEND_SIGNALED;
+	err = mlx4_post_send(txq->qp, head.next, &bad_wr);
+	if (unlikely(err)) {
+		unsigned int unsent = 0;
+
+		/* An error occured, completion event is lost. Fix counters. */
+		while (bad_wr != NULL) {
+			struct txq_elt *elt =
+				containerof(bad_wr, struct txq_elt, wr);
+			mlx4_send_wr_t *wr = &elt->wr;
+			mlx4_send_wr_t *next = wr->next;
+#if defined(MLX4_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
+			unsigned int j;
+#endif
+
+			assert(wr == bad_wr);
+			/* Clean up TX element without freeing it, caller
+			 * should take care of this. */
+			WR_ID(elt->wr.wr_id).offset = 0;
+#ifdef MLX4_PMD_SOFT_COUNTERS
+			for (j = 0; ((int)j < wr->num_sge); ++j)
+				txq->stats.obytes -= wr->sg_list[j].length;
+#endif
+			++unsent;
+#ifndef NDEBUG
+			/* For assert(). */
+			for (j = 0; ((int)j < wr->num_sge); ++j) {
+				elt->sges[j].addr = 0;
+				elt->sges[j].length = 0;
+				elt->sges[j].lkey = 0;
+			}
+			wr->next = NULL;
+			wr->num_sge = 0;
+#endif
+			bad_wr = next;
+		}
+#ifdef MLX4_PMD_SOFT_COUNTERS
+		txq->stats.opackets -= unsent;
+#endif
+		assert(i >= unsent);
+		i -= unsent;
+		/* "Unsend" remaining packets. */
+		elts_head -= unsent;
+		if (elts_head >= elts_n)
+			elts_head += elts_n;
+		assert(elts_head < elts_n);
+		DEBUG("%p: mlx4_post_send() failed, %u unprocessed WRs: %s",
+		      (void *)txq, unsent,
+		      ((err <= -1) ? "Internal error" : strerror(err)));
+	}
+	else
+		++txq->elts_comp;
+	txq->elts_head = elts_head;
+	return i;
+}
+
+/**
+ * Configure a TX queue.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param txq
+ *   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.
+ */
+static int
+txq_setup(struct rte_eth_dev *dev, struct txq *txq, uint16_t desc,
+	  unsigned int socket, const struct rte_eth_txconf *conf)
+{
+	struct priv *priv = dev->data->dev_private;
+	struct txq tmpl = {
+		.priv = priv,
+		.socket = socket
+	};
+	union {
+		struct ibv_qp_init_attr init;
+		struct ibv_exp_qp_attr mod;
+	} attr;
+	int ret = 0;
+
+	(void)conf; /* Thresholds configuration (ignored). */
+	if ((desc == 0) || (desc % MLX4_PMD_SGE_WR_N)) {
+		ERROR("%p: invalid number of TX descriptors (must be a"
+		      " multiple of %d)", (void *)dev, desc);
+		return EINVAL;
+	}
+	desc /= MLX4_PMD_SGE_WR_N;
+	/* MRs will be registered in mp2mr[] later. */
+	tmpl.cq = ibv_create_cq(priv->ctx, desc, NULL, NULL, 0);
+	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_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. */
+			.max_send_sge = ((priv->device_attr.max_sge <
+					  MLX4_PMD_SGE_WR_N) ?
+					 priv->device_attr.max_sge :
+					 MLX4_PMD_SGE_WR_N),
+#if MLX4_PMD_MAX_INLINE > 0
+			.max_inline_data = MLX4_PMD_MAX_INLINE,
+#endif
+		},
+		.qp_type = IBV_QPT_RAW_PACKET,
+		/* Do *NOT* enable this, completions events are managed per
+		 * TX burst. */
+		.sq_sig_all = 0
+	};
+	tmpl.qp = ibv_create_qp(priv->pd, &attr.init);
+	if (tmpl.qp == NULL) {
+		ret = (errno ? errno : EINVAL);
+		ERROR("%p: QP creation failure: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+#if MLX4_PMD_MAX_INLINE > 0
+	/* ibv_create_qp() updates this value. */
+	tmpl.max_inline = attr.init.cap.max_inline_data;
+#endif
+	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
+	};
+	if ((ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod,
+				 (IBV_EXP_QP_STATE | IBV_EXP_QP_PORT)))) {
+		ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	if ((ret = txq_alloc_elts(&tmpl, desc))) {
+		ERROR("%p: TXQ allocation failed: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	attr.mod = (struct ibv_exp_qp_attr){
+		.qp_state = IBV_QPS_RTR
+	};
+	if ((ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE))) {
+		ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	attr.mod.qp_state = IBV_QPS_RTS;
+	if ((ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE))) {
+		ERROR("%p: QP state to IBV_QPS_RTS failed: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	/* Clean up txq in case we're reinitializing it. */
+	DEBUG("%p: cleaning-up old txq just in case", (void *)txq);
+	txq_cleanup(txq);
+	*txq = tmpl;
+	DEBUG("%p: txq updated with %p", (void *)txq, (void *)&tmpl);
+	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.
+ */
+static int
+mlx4_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];
+	int ret;
+
+	priv_lock(priv);
+	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);
+	}
+	else {
+		txq = rte_calloc_socket("TXQ", 1, sizeof(*txq), 0, socket);
+		if (txq == NULL) {
+			ERROR("%p: unable to allocate queue index %u",
+			      (void *)dev, idx);
+			priv_unlock(priv);
+			return -ENOMEM;
+		}
+	}
+	ret = txq_setup(dev, txq, desc, socket, conf);
+	if (ret)
+		rte_free(txq);
+	else {
+		txq->stats.idx = idx;
+		DEBUG("%p: adding TX queue %p to list",
+		      (void *)dev, (void *)txq);
+		(*priv->txqs)[idx] = txq;
+		/* Update send callback. */
+		dev->tx_pkt_burst = mlx4_tx_burst;
+	}
+	priv_unlock(priv);
+	return -ret;
+}
+
+/**
+ * DPDK callback to release a TX queue.
+ *
+ * @param dpdk_txq
+ *   Generic TX queue pointer.
+ */
+static void
+mlx4_tx_queue_release(void *dpdk_txq)
+{
+	struct txq *txq = (struct txq *)dpdk_txq;
+	struct priv *priv;
+	unsigned int i;
+
+	if (txq == NULL)
+		return;
+	priv = txq->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);
+			(*priv->txqs)[i] = NULL;
+			break;
+		}
+	txq_cleanup(txq);
+	rte_free(txq);
+	priv_unlock(priv);
+}
+
+/* RX queues handling. */
+
+/**
+ * Allocate RX queue elements with scattered packets support.
+ *
+ * @param rxq
+ *   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_sp(struct rxq *rxq, unsigned int elts_n,
+		  struct rte_mbuf **pool)
+{
+	unsigned int i;
+	struct rxq_elt_sp (*elts)[elts_n] =
+		rte_calloc_socket("RXQ elements", 1, sizeof(*elts), 0,
+				  rxq->socket);
+	int ret = 0;
+
+	if (elts == NULL) {
+		ERROR("%p: can't allocate packets array", (void *)rxq);
+		ret = ENOMEM;
+		goto error;
+	}
+	/* For each WR (packet). */
+	for (i = 0; (i != elts_n); ++i) {
+		unsigned int j;
+		struct rxq_elt_sp *elt = &(*elts)[i];
+		struct ibv_recv_wr *wr = &elt->wr;
+		struct ibv_sge (*sges)[(elemof(elt->sges))] = &elt->sges;
+
+		/* These two arrays must have the same size. */
+		assert(elemof(elt->sges) == elemof(elt->bufs));
+		/* Configure WR. */
+		wr->wr_id = i;
+		wr->next = &(*elts)[(i + 1)].wr;
+		wr->sg_list = &(*sges)[0];
+		wr->num_sge = elemof(*sges);
+		/* For each SGE (segment). */
+		for (j = 0; (j != elemof(elt->bufs)); ++j) {
+			struct ibv_sge *sge = &(*sges)[j];
+			struct rte_mbuf *buf;
+
+			if (pool != NULL) {
+				buf = *(pool++);
+				assert(buf != NULL);
+				rte_pktmbuf_reset(buf);
+			}
+			else
+				buf = rte_pktmbuf_alloc(rxq->mp);
+			if (buf == NULL) {
+				assert(pool == NULL);
+				ERROR("%p: empty mbuf pool", (void *)rxq);
+				ret = ENOMEM;
+				goto error;
+			}
+			elt->bufs[j] = buf;
+			/* 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);
+			/* sge->addr must be able to store a pointer. */
+			assert(sizeof(sge->addr) >= sizeof(uintptr_t));
+			if (j == 0) {
+				/* The first SGE keeps its headroom. */
+				sge->addr = (uintptr_t)rte_pktmbuf_mtod(buf,
+									char *);
+				sge->length = (buf->buf_len -
+					       RTE_PKTMBUF_HEADROOM);
+			}
+			else {
+				/* Subsequent SGEs lose theirs. */
+				assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
+				SET_DATA_OFF(buf, 0);
+				sge->addr = (uintptr_t)buf->buf_addr;
+				sge->length = buf->buf_len;
+			}
+			sge->lkey = rxq->mr->lkey;
+			/* Redundant check for tailroom. */
+			assert(sge->length == rte_pktmbuf_tailroom(buf));
+		}
+	}
+	/* The last WR pointer must be NULL. */
+	(*elts)[(i - 1)].wr.next = NULL;
+	DEBUG("%p: allocated and configured %u WRs (%zu segments)",
+	      (void *)rxq, elts_n, (elts_n * elemof((*elts)[0].sges)));
+	rxq->elts_n = elts_n;
+	rxq->elts.sp = elts;
+	assert(ret == 0);
+	return 0;
+error:
+	if (elts != NULL) {
+		assert(pool == NULL);
+		for (i = 0; (i != elemof(*elts)); ++i) {
+			unsigned int j;
+			struct rxq_elt_sp *elt = &(*elts)[i];
+
+			for (j = 0; (j != elemof(elt->bufs)); ++j) {
+				struct rte_mbuf *buf = elt->bufs[j];
+
+				if (buf != NULL)
+					rte_pktmbuf_free_seg(buf);
+			}
+		}
+		rte_free(elts);
+	}
+	DEBUG("%p: failed, freed everything", (void *)rxq);
+	assert(ret > 0);
+	return ret;
+}
+
+/**
+ * Free RX queue elements with scattered packets support.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ */
+static void
+rxq_free_elts_sp(struct rxq *rxq)
+{
+	unsigned int i;
+	unsigned int elts_n = rxq->elts_n;
+	struct rxq_elt_sp (*elts)[elts_n] = rxq->elts.sp;
+
+	DEBUG("%p: freeing WRs", (void *)rxq);
+	rxq->elts_n = 0;
+	rxq->elts.sp = NULL;
+	if (elts == NULL)
+		return;
+	for (i = 0; (i != elemof(*elts)); ++i) {
+		unsigned int j;
+		struct rxq_elt_sp *elt = &(*elts)[i];
+
+		for (j = 0; (j != elemof(elt->bufs)); ++j) {
+			struct rte_mbuf *buf = elt->bufs[j];
+
+			if (buf != NULL)
+				rte_pktmbuf_free_seg(buf);
+		}
+	}
+	rte_free(elts);
+}
+
+/**
+ * Allocate RX queue elements.
+ *
+ * @param rxq
+ *   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 *rxq, unsigned int elts_n, struct rte_mbuf **pool)
+{
+	unsigned int i;
+	struct rxq_elt (*elts)[elts_n] =
+		rte_calloc_socket("RXQ elements", 1, sizeof(*elts), 0,
+				  rxq->socket);
+	int ret = 0;
+
+	if (elts == NULL) {
+		ERROR("%p: can't allocate packets array", (void *)rxq);
+		ret = ENOMEM;
+		goto error;
+	}
+	/* For each WR (packet). */
+	for (i = 0; (i != elts_n); ++i) {
+		struct rxq_elt *elt = &(*elts)[i];
+		struct ibv_recv_wr *wr = &elt->wr;
+		struct ibv_sge *sge = &(*elts)[i].sge;
+		struct rte_mbuf *buf;
+
+		if (pool != NULL) {
+			buf = *(pool++);
+			assert(buf != NULL);
+			rte_pktmbuf_reset(buf);
+		}
+		else
+			buf = rte_pktmbuf_alloc(rxq->mp);
+		if (buf == NULL) {
+			assert(pool == NULL);
+			ERROR("%p: empty mbuf pool", (void *)rxq);
+			ret = ENOMEM;
+			goto error;
+		}
+		/* Configure WR. Work request ID contains its own index in
+		 * the elts array and the offset between SGE buffer header and
+		 * its data. */
+		WR_ID(wr->wr_id).id = i;
+		WR_ID(wr->wr_id).offset =
+			(((uintptr_t)buf->buf_addr + RTE_PKTMBUF_HEADROOM) -
+			 (uintptr_t)buf);
+		wr->next = &(*elts)[(i + 1)].wr;
+		wr->sg_list = sge;
+		wr->num_sge = 1;
+		/* 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);
+		/* sge->addr must be able to store a pointer. */
+		assert(sizeof(sge->addr) >= sizeof(uintptr_t));
+		/* SGE keeps its headroom. */
+		sge->addr = (uintptr_t)
+			((uint8_t *)buf->buf_addr + RTE_PKTMBUF_HEADROOM);
+		sge->length = (buf->buf_len - RTE_PKTMBUF_HEADROOM);
+		sge->lkey = rxq->mr->lkey;
+		/* Redundant check for tailroom. */
+		assert(sge->length == rte_pktmbuf_tailroom(buf));
+		/* Make sure elts index and SGE mbuf pointer can be deduced
+		 * from WR ID. */
+		if ((WR_ID(wr->wr_id).id != i) ||
+		    ((void *)(sge->addr - WR_ID(wr->wr_id).offset) != buf)) {
+			ERROR("%p: cannot store index and offset in WR ID",
+			      (void *)rxq);
+			sge->addr = 0;
+			rte_pktmbuf_free(buf);
+			ret = EOVERFLOW;
+			goto error;
+		}
+	}
+	/* The last WR pointer must be NULL. */
+	(*elts)[(i - 1)].wr.next = NULL;
+	DEBUG("%p: allocated and configured %u single-segment WRs",
+	      (void *)rxq, elts_n);
+	rxq->elts_n = elts_n;
+	rxq->elts.no_sp = elts;
+	assert(ret == 0);
+	return 0;
+error:
+	if (elts != NULL) {
+		assert(pool == NULL);
+		for (i = 0; (i != elemof(*elts)); ++i) {
+			struct rxq_elt *elt = &(*elts)[i];
+			struct rte_mbuf *buf;
+
+			if (elt->sge.addr == 0)
+				continue;
+			assert(WR_ID(elt->wr.wr_id).id == i);
+			buf = (void *)
+				(elt->sge.addr - WR_ID(elt->wr.wr_id).offset);
+			rte_pktmbuf_free_seg(buf);
+		}
+		rte_free(elts);
+	}
+	DEBUG("%p: failed, freed everything", (void *)rxq);
+	assert(ret > 0);
+	return ret;
+}
+
+/**
+ * Free RX queue elements.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ */
+static void
+rxq_free_elts(struct rxq *rxq)
+{
+	unsigned int i;
+	unsigned int elts_n = rxq->elts_n;
+	struct rxq_elt (*elts)[elts_n] = rxq->elts.no_sp;
+
+	DEBUG("%p: freeing WRs", (void *)rxq);
+	rxq->elts_n = 0;
+	rxq->elts.no_sp = NULL;
+	if (elts == NULL)
+		return;
+	for (i = 0; (i != elemof(*elts)); ++i) {
+		struct rxq_elt *elt = &(*elts)[i];
+		struct rte_mbuf *buf;
+
+		if (elt->sge.addr == 0)
+			continue;
+		assert(WR_ID(elt->wr.wr_id).id == i);
+		buf = (void *)(elt->sge.addr - WR_ID(elt->wr.wr_id).offset);
+		rte_pktmbuf_free_seg(buf);
+	}
+	rte_free(elts);
+}
+
+/**
+ * Unregister a MAC address from a RX queue.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ * @param mac_index
+ *   MAC address index.
+ */
+static void
+rxq_mac_addr_del(struct rxq *rxq, unsigned int mac_index)
+{
+#if defined(NDEBUG) || defined(MLX4_COMPAT_VMWARE)
+	struct priv *priv = rxq->priv;
+	const uint8_t (*mac)[ETHER_ADDR_LEN] =
+		(const uint8_t (*)[ETHER_ADDR_LEN])
+		priv->mac[mac_index].addr_bytes;
+#endif
+
+	assert(mac_index < elemof(priv->mac));
+	if (!BITFIELD_ISSET(rxq->mac_configured, mac_index)) {
+		assert(rxq->mac_flow[mac_index] == NULL);
+		return;
+	}
+	DEBUG("%p: removing MAC address %02x:%02x:%02x:%02x:%02x:%02x"
+	      " index %u",
+	      (void *)rxq,
+	      (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5],
+	      mac_index);
+#ifdef MLX4_COMPAT_VMWARE
+	if (priv->vmware) {
+		union ibv_gid gid = { .raw = { 0 } };
+
+		memcpy(&gid.raw[10], *mac, sizeof(*mac));
+		claim_zero(ibv_detach_mcast(rxq->qp, &gid, 0));
+		BITFIELD_RESET(rxq->mac_configured, mac_index);
+		return;
+	}
+#endif
+	assert(rxq->mac_flow[mac_index] != NULL);
+	claim_zero(ibv_exp_destroy_flow(rxq->mac_flow[mac_index]));
+	rxq->mac_flow[mac_index] = NULL;
+	BITFIELD_RESET(rxq->mac_configured, mac_index);
+}
+
+/**
+ * Unregister all MAC addresses from a RX queue.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ */
+static void
+rxq_mac_addrs_del(struct rxq *rxq)
+{
+	struct priv *priv = rxq->priv;
+	unsigned int i;
+
+	for (i = 0; (i != elemof(priv->mac)); ++i)
+		rxq_mac_addr_del(rxq, i);
+}
+
+static int rxq_promiscuous_enable(struct rxq *);
+static void rxq_promiscuous_disable(struct rxq *);
+
+/**
+ * Register a MAC address in a RX queue.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ * @param mac_index
+ *   MAC address index to register.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+rxq_mac_addr_add(struct rxq *rxq, unsigned int mac_index)
+{
+	struct priv *priv = rxq->priv;
+	const uint8_t (*mac)[ETHER_ADDR_LEN] =
+		(const uint8_t (*)[ETHER_ADDR_LEN])
+		priv->mac[mac_index].addr_bytes;
+	unsigned int vlans = 0;
+	unsigned int specs = 0;
+	unsigned int i, j;
+	struct ibv_exp_flow *flow;
+
+	assert(mac_index < elemof(priv->mac));
+	if (BITFIELD_ISSET(rxq->mac_configured, mac_index))
+		rxq_mac_addr_del(rxq, mac_index);
+	/* Number of configured VLANs. */
+	for (i = 0; (i != elemof(priv->vlan_filter)); ++i)
+		if (priv->vlan_filter[i].enabled)
+			++vlans;
+	specs = (vlans ? vlans : 1);
+
+	/* Allocate flow specification on the stack. */
+	struct ibv_exp_flow_attr data[1 +
+				  (sizeof(struct ibv_exp_flow_spec_eth[specs]) /
+				   sizeof(struct ibv_exp_flow_attr)) +
+				  !!(sizeof(struct ibv_exp_flow_spec_eth[specs]) %
+				  sizeof(struct ibv_exp_flow_attr))];
+
+	struct ibv_exp_flow_attr *attr = (void *)&data[0];
+	struct ibv_exp_flow_spec_eth *spec = (void *)&data[1];
+
+	/*
+	 * 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);
+	*attr = (struct ibv_exp_flow_attr){
+		.type = IBV_EXP_FLOW_ATTR_NORMAL,
+		.num_of_specs = specs,
+		.port = priv->port,
+		.flags = 0
+	};
+	*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]
+			}
+		},
+		.mask = {
+			.dst_mac = "\xff\xff\xff\xff\xff\xff",
+			.vlan_tag = (vlans ? 0xfff : 0)
+		}
+	};
+	/* Fill VLAN specifications. */
+	for (i = 0, j = 0; (i != elemof(priv->vlan_filter)); ++i) {
+		if (!priv->vlan_filter[i].enabled)
+			continue;
+		assert(j != vlans);
+		if (j)
+			spec[j] = spec[0];
+		spec[j].val.vlan_tag = priv->vlan_filter[i].id;
+		++j;
+	}
+	DEBUG("%p: adding MAC address %02x:%02x:%02x:%02x:%02x:%02x index %u"
+	      " (%u VLAN(s) configured)",
+	      (void *)rxq,
+	      (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5],
+	      mac_index,
+	      vlans);
+#ifdef MLX4_COMPAT_VMWARE
+	if (priv->vmware) {
+		union ibv_gid gid = { .raw = { 0 } };
+
+		/* Call multicast attach with unicast mac to get traffic. */
+		memcpy(&gid.raw[10], *mac, sizeof(*mac));
+		errno = 0;
+		if (ibv_attach_mcast(rxq->qp, &gid, 0)) {
+			if (errno)
+				return errno;
+			return EINVAL;
+		}
+		BITFIELD_SET(rxq->mac_configured, mac_index);
+		return 0;
+	}
+#endif
+	/* Create related flow. */
+	errno = 0;
+	if ((flow = ibv_exp_create_flow(rxq->qp, attr)) == NULL) {
+		int err = errno;
+		int disable = 0;
+
+		/* Flow creation failure is not fatal when in DMFS A0 mode.
+		 * Ignore error if promiscuity is already enabled or can be
+		 * enabled. */
+		if (priv->promisc_ok)
+			return 0;
+		if ((rxq->promisc_flow != NULL) ||
+		    (disable = 1, rxq_promiscuous_enable(rxq) == 0)) {
+			if (disable)
+				rxq_promiscuous_disable(rxq);
+			WARN("cannot configure normal flow but promiscuous"
+			     " mode is fine, assuming promiscuous optimization"
+			     " is enabled"
+			     " (options mlx4_core log_num_mgm_entry_size=-7)");
+			priv->promisc_ok = 1;
+			return 0;
+		}
+		errno = err;
+		/* It's not clear whether errno is always set in this case. */
+		ERROR("%p: flow configuration failed, errno=%d: %s",
+		      (void *)rxq, errno,
+		      (errno ? strerror(errno) : "Unknown error"));
+		if (errno)
+			return errno;
+		return EINVAL;
+	}
+	assert(rxq->mac_flow[mac_index] == NULL);
+	rxq->mac_flow[mac_index] = flow;
+	BITFIELD_SET(rxq->mac_configured, mac_index);
+	return 0;
+}
+
+/**
+ * Register all MAC addresses in a RX queue.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+rxq_mac_addrs_add(struct rxq *rxq)
+{
+	struct priv *priv = rxq->priv;
+	unsigned int i;
+	int ret;
+
+	for (i = 0; (i != elemof(priv->mac)); ++i) {
+		if (!BITFIELD_ISSET(priv->mac_configured, i))
+			continue;
+		ret = rxq_mac_addr_add(rxq, i);
+		if (!ret)
+			continue;
+		/* Failure, rollback. */
+		while (i != 0)
+			rxq_mac_addr_del(rxq, --i);
+		assert(ret > 0);
+		return ret;
+	}
+	return 0;
+}
+
+/**
+ * Unregister a MAC address.
+ *
+ * In RSS mode, the MAC address is unregistered from the parent queue,
+ * otherwise it is unregistered from each queue directly.
+ *
+ * @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 < elemof(priv->mac));
+	if (!BITFIELD_ISSET(priv->mac_configured, mac_index))
+		return;
+	if (priv->rss) {
+		rxq_mac_addr_del(&priv->rxq_parent, mac_index);
+		goto end;
+	}
+	for (i = 0; (i != priv->dev->data->nb_rx_queues); ++i)
+		rxq_mac_addr_del((*priv->rxqs)[i], mac_index);
+end:
+	BITFIELD_RESET(priv->mac_configured, mac_index);
+}
+
+/**
+ * Register a MAC address.
+ *
+ * In RSS mode, the MAC address is registered in the parent queue,
+ * otherwise it is registered in each queue directly.
+ *
+ * @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.
+ */
+static 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 < elemof(priv->mac));
+	/* First, make sure this address isn't already configured. */
+	for (i = 0; (i != elemof(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 device isn't started, this is all we need to do. */
+	if (!priv->started) {
+#ifndef NDEBUG
+		/* Verify that all queues have this index disabled. */
+		for (i = 0; (i != priv->rxqs_n); ++i) {
+			if ((*priv->rxqs)[i] == NULL)
+				continue;
+			assert(!BITFIELD_ISSET
+			       ((*priv->rxqs)[i]->mac_configured, mac_index));
+		}
+#endif
+		goto end;
+	}
+	if (priv->rss) {
+		ret = rxq_mac_addr_add(&priv->rxq_parent, mac_index);
+		if (ret)
+			return ret;
+		goto end;
+	}
+	for (i = 0; (i != priv->rxqs_n); ++i) {
+		if ((*priv->rxqs)[i] == NULL)
+			continue;
+		ret = rxq_mac_addr_add((*priv->rxqs)[i], mac_index);
+		if (!ret)
+			continue;
+		/* Failure, rollback. */
+		while (i != 0)
+			if ((*priv->rxqs)[(--i)] != NULL)
+				rxq_mac_addr_del((*priv->rxqs)[i], mac_index);
+		return ret;
+	}
+end:
+	BITFIELD_SET(priv->mac_configured, mac_index);
+	return 0;
+}
+
+/**
+ * Enable allmulti mode in a RX queue.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+rxq_allmulticast_enable(struct rxq *rxq)
+{
+	struct ibv_exp_flow *flow;
+	struct ibv_exp_flow_attr attr = {
+		.type = IBV_EXP_FLOW_ATTR_MC_DEFAULT,
+		.num_of_specs = 0,
+		.port = rxq->priv->port,
+		.flags = 0
+	};
+
+#ifdef MLX4_COMPAT_VMWARE
+	if (rxq->priv->vmware) {
+		ERROR("%p: allmulticast mode is not supported in VMware",
+		      (void *)rxq);
+		return EINVAL;
+	}
+#endif
+	DEBUG("%p: enabling allmulticast mode", (void *)rxq);
+	if (rxq->allmulti_flow != NULL)
+		return EBUSY;
+	errno = 0;
+	if ((flow = ibv_exp_create_flow(rxq->qp, &attr)) == NULL) {
+		/* It's not clear whether errno is always set in this case. */
+		ERROR("%p: flow configuration failed, errno=%d: %s",
+		      (void *)rxq, errno,
+		      (errno ? strerror(errno) : "Unknown error"));
+		if (errno)
+			return errno;
+		return EINVAL;
+	}
+	rxq->allmulti_flow = flow;
+	DEBUG("%p: allmulticast mode enabled", (void *)rxq);
+	return 0;
+}
+
+/**
+ * Disable allmulti mode in a RX queue.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ */
+static void
+rxq_allmulticast_disable(struct rxq *rxq)
+{
+#ifdef MLX4_COMPAT_VMWARE
+	if (rxq->priv->vmware) {
+		ERROR("%p: allmulticast mode is not supported in VMware",
+		      (void *)rxq);
+		return;
+	}
+#endif
+	DEBUG("%p: disabling allmulticast mode", (void *)rxq);
+	if (rxq->allmulti_flow == NULL)
+		return;
+	claim_zero(ibv_exp_destroy_flow(rxq->allmulti_flow));
+	rxq->allmulti_flow = NULL;
+	DEBUG("%p: allmulticast mode disabled", (void *)rxq);
+}
+
+/**
+ * Enable promiscuous mode in a RX queue.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+rxq_promiscuous_enable(struct rxq *rxq)
+{
+	struct ibv_exp_flow *flow;
+	struct ibv_exp_flow_attr attr = {
+		.type = IBV_EXP_FLOW_ATTR_ALL_DEFAULT,
+		.num_of_specs = 0,
+		.port = rxq->priv->port,
+		.flags = 0
+	};
+
+#ifdef MLX4_COMPAT_VMWARE
+	if (rxq->priv->vmware) {
+		ERROR("%p: promiscuous mode is not supported in VMware",
+		      (void *)rxq);
+		return EINVAL;
+	}
+#endif
+	if (rxq->priv->vf)
+		return 0;
+	DEBUG("%p: enabling promiscuous mode", (void *)rxq);
+	if (rxq->promisc_flow != NULL)
+		return EBUSY;
+	errno = 0;
+	if ((flow = ibv_exp_create_flow(rxq->qp, &attr)) == NULL) {
+		/* It's not clear whether errno is always set in this case. */
+		ERROR("%p: flow configuration failed, errno=%d: %s",
+		      (void *)rxq, errno,
+		      (errno ? strerror(errno) : "Unknown error"));
+		if (errno)
+			return errno;
+		return EINVAL;
+	}
+	rxq->promisc_flow = flow;
+	DEBUG("%p: promiscuous mode enabled", (void *)rxq);
+	return 0;
+}
+
+/**
+ * Disable promiscuous mode in a RX queue.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ */
+static void
+rxq_promiscuous_disable(struct rxq *rxq)
+{
+#ifdef MLX4_COMPAT_VMWARE
+	if (rxq->priv->vmware) {
+		ERROR("%p: promiscuous mode is not supported in VMware",
+		      (void *)rxq);
+		return;
+	}
+#endif
+	if (rxq->priv->vf)
+		return;
+	DEBUG("%p: disabling promiscuous mode", (void *)rxq);
+	if (rxq->promisc_flow == NULL)
+		return;
+	claim_zero(ibv_exp_destroy_flow(rxq->promisc_flow));
+	rxq->promisc_flow = NULL;
+	DEBUG("%p: promiscuous mode disabled", (void *)rxq);
+}
+
+/**
+ * Clean up a RX queue.
+ *
+ * Destroy objects, free allocated memory and reset the structure for reuse.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ */
+static void
+rxq_cleanup(struct rxq *rxq)
+{
+	DEBUG("cleaning up %p", (void *)rxq);
+	if (rxq->sp)
+		rxq_free_elts_sp(rxq);
+	else
+		rxq_free_elts(rxq);
+	if (rxq->qp != NULL) {
+		rxq_promiscuous_disable(rxq);
+		rxq_allmulticast_disable(rxq);
+		rxq_mac_addrs_del(rxq);
+		claim_zero(ibv_destroy_qp(rxq->qp));
+	}
+	if (rxq->cq != NULL)
+		claim_zero(ibv_destroy_cq(rxq->cq));
+	if (rxq->mr != NULL)
+		claim_zero(ibv_dereg_mr(rxq->mr));
+	memset(rxq, 0, sizeof(*rxq));
+}
+
+static uint16_t
+mlx4_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n);
+
+/**
+ * DPDK callback for RX with scattered packets support.
+ *
+ * @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).
+ */
+static uint16_t
+mlx4_rx_burst_sp(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+	struct rxq *rxq = (struct rxq *)dpdk_rxq;
+	struct rxq_elt_sp (*elts)[rxq->elts_n] = rxq->elts.sp;
+	struct ibv_wc wcs[pkts_n];
+	struct ibv_recv_wr head;
+	struct ibv_recv_wr **next = &head.next;
+	struct ibv_recv_wr *bad_wr;
+	int ret = 0;
+	int wcs_n;
+	int i;
+
+	if (unlikely(!rxq->sp))
+		return mlx4_rx_burst(dpdk_rxq, pkts, pkts_n);
+	if (unlikely(elts == NULL)) /* See RTE_DEV_CMD_SET_MTU. */
+		return 0;
+	wcs_n = ibv_poll_cq(rxq->cq, pkts_n, wcs);
+	if (unlikely(wcs_n == 0))
+		return 0;
+	if (unlikely(wcs_n < 0)) {
+		DEBUG("rxq=%p, ibv_poll_cq() failed (wc_n=%d)",
+		      (void *)rxq, wcs_n);
+		return 0;
+	}
+	assert(wcs_n <= (int)pkts_n);
+	/* For each work completion. */
+	for (i = 0; (i != wcs_n); ++i) {
+		struct ibv_wc *wc = &wcs[i];
+		uint64_t wr_id = wc->wr_id;
+		uint32_t len = wc->byte_len;
+		struct rxq_elt_sp *elt = &(*elts)[wr_id];
+		struct ibv_recv_wr *wr = &elt->wr;
+		struct rte_mbuf *pkt_buf = NULL; /* Buffer returned in pkts. */
+		struct rte_mbuf **pkt_buf_next = &pkt_buf;
+		unsigned int seg_headroom = RTE_PKTMBUF_HEADROOM;
+		unsigned int j = 0;
+
+		/* Sanity checks. */
+		assert(wr_id < rxq->elts_n);
+		assert(wr_id == wr->wr_id);
+		assert(wr->sg_list == elt->sges);
+		assert(wr->num_sge == elemof(elt->sges));
+		/* Link completed WRs together for repost. */
+		*next = wr;
+		next = &wr->next;
+		if (unlikely(wc->status != IBV_WC_SUCCESS)) {
+			/* Whatever, just repost the offending WR. */
+			DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work completion"
+			      " status (%d): %s",
+			      (void *)rxq, wc->wr_id, wc->status,
+			      ibv_wc_status_str(wc->status));
+#ifdef MLX4_PMD_SOFT_COUNTERS
+			/* Increase dropped packets counter. */
+			++rxq->stats.idropped;
+#endif
+			goto repost;
+		}
+		/*
+		 * Replace spent segments with new ones, concatenate and
+		 * return them as pkt_buf.
+		 */
+		while (1) {
+			struct ibv_sge *sge = &elt->sges[j];
+			struct rte_mbuf *seg = elt->bufs[j];
+			struct rte_mbuf *rep;
+			unsigned int seg_tailroom;
+
+			/*
+			 * Fetch initial bytes of packet descriptor into a
+			 * cacheline while allocating rep.
+			 */
+			rte_prefetch0(seg);
+			rep = __rte_mbuf_raw_alloc(rxq->mp);
+			if (unlikely(rep == NULL)) {
+				/*
+				 * Unable to allocate a replacement mbuf,
+				 * repost WR.
+				 */
+				DEBUG("rxq=%p, wr_id=%" PRIu64 ":"
+				      " can't allocate a new mbuf",
+				      (void *)rxq, wr_id);
+				if (pkt_buf != NULL)
+					rte_pktmbuf_free(pkt_buf);
+				/* Increase out of memory counters. */
+				++rxq->stats.rx_nombuf;
+				++rxq->priv->dev->data->rx_mbuf_alloc_failed;
+				goto repost;
+			}
+#ifndef NDEBUG
+			/* Poison user-modifiable fields in rep. */
+			NEXT(rep) = (void *)((uintptr_t)-1);
+			SET_DATA_OFF(rep, 0xdead);
+			DATA_LEN(rep) = 0xd00d;
+			PKT_LEN(rep) = 0xdeadd00d;
+			NB_SEGS(rep) = 0x2a;
+			PORT(rep) = 0x2a;
+			rep->ol_flags = -1;
+#endif
+			assert(rep->buf_len == seg->buf_len);
+			assert(rep->buf_len == rxq->mb_len);
+			/* Reconfigure sge to use rep instead of seg. */
+			assert(sge->lkey == rxq->mr->lkey);
+			sge->addr = ((uintptr_t)rep->buf_addr + seg_headroom);
+			elt->bufs[j] = rep;
+			++j;
+			/* Update pkt_buf if it's the first segment, or link
+			 * seg to the previous one and update pkt_buf_next. */
+			*pkt_buf_next = seg;
+			pkt_buf_next = &NEXT(seg);
+			/* Update seg information. */
+			seg_tailroom = (seg->buf_len - seg_headroom);
+			assert(sge->length == seg_tailroom);
+			SET_DATA_OFF(seg, seg_headroom);
+			if (likely(len <= seg_tailroom)) {
+				/* Last segment. */
+				DATA_LEN(seg) = len;
+				PKT_LEN(seg) = len;
+				/* Sanity check. */
+				assert(rte_pktmbuf_headroom(seg) ==
+				       seg_headroom);
+				assert(rte_pktmbuf_tailroom(seg) ==
+				       (seg_tailroom - len));
+				break;
+			}
+			DATA_LEN(seg) = seg_tailroom;
+			PKT_LEN(seg) = seg_tailroom;
+			/* Sanity check. */
+			assert(rte_pktmbuf_headroom(seg) == seg_headroom);
+			assert(rte_pktmbuf_tailroom(seg) == 0);
+			/* Fix len and clear headroom for next segments. */
+			len -= seg_tailroom;
+			seg_headroom = 0;
+		}
+		/* Update head and tail segments. */
+		*pkt_buf_next = NULL;
+		assert(pkt_buf != NULL);
+		assert(j != 0);
+		NB_SEGS(pkt_buf) = j;
+		PORT(pkt_buf) = rxq->port_id;
+		PKT_LEN(pkt_buf) = wc->byte_len;
+		pkt_buf->ol_flags = 0;
+
+		/* Return packet. */
+		*(pkts++) = pkt_buf;
+		++ret;
+#ifdef MLX4_PMD_SOFT_COUNTERS
+		/* Increase bytes counter. */
+		rxq->stats.ibytes += wc->byte_len;
+#endif
+	repost:
+		continue;
+	}
+	*next = NULL;
+	/* Repost WRs. */
+#ifdef DEBUG_RECV
+	DEBUG("%p: reposting %d WRs starting from %" PRIu64 " (%p)",
+	      (void *)rxq, wcs_n, wcs[0].wr_id, (void *)head.next);
+#endif
+	i = ibv_post_recv(rxq->qp, head.next, &bad_wr);
+	if (unlikely(i)) {
+		/* Inability to repost WRs is fatal. */
+		DEBUG("%p: ibv_post_recv(): failed for WR %p: %s",
+		      (void *)rxq->priv,
+		      (void *)bad_wr,
+		      strerror(i));
+		abort();
+	}
+#ifdef MLX4_PMD_SOFT_COUNTERS
+	/* Increase packets counter. */
+	rxq->stats.ipackets += ret;
+#endif
+	return ret;
+}
+
+/**
+ * DPDK callback for RX.
+ *
+ * The following function is the same as mlx4_rx_burst_sp(), except it doesn't
+ * manage scattered packets. Improves performance when MRU is lower than the
+ * size of the first segment.
+ *
+ * @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).
+ */
+static uint16_t
+mlx4_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+	struct rxq *rxq = (struct rxq *)dpdk_rxq;
+	struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts.no_sp;
+	struct ibv_wc wcs[pkts_n];
+	struct ibv_recv_wr head;
+	struct ibv_recv_wr **next = &head.next;
+	struct ibv_recv_wr *bad_wr;
+	int ret = 0;
+	int wcs_n;
+	int i;
+
+	if (unlikely(rxq->sp))
+		return mlx4_rx_burst_sp(dpdk_rxq, pkts, pkts_n);
+	wcs_n = ibv_poll_cq(rxq->cq, pkts_n, wcs);
+	if (unlikely(wcs_n == 0))
+		return 0;
+	if (unlikely(wcs_n < 0)) {
+		DEBUG("rxq=%p, ibv_poll_cq() failed (wc_n=%d)",
+		      (void *)rxq, wcs_n);
+		return 0;
+	}
+	assert(wcs_n <= (int)pkts_n);
+	/* For each work completion. */
+	for (i = 0; (i != wcs_n); ++i) {
+		struct ibv_wc *wc = &wcs[i];
+		uint64_t wr_id = wc->wr_id;
+		uint32_t len = wc->byte_len;
+		struct rxq_elt *elt = &(*elts)[WR_ID(wr_id).id];
+		struct ibv_recv_wr *wr = &elt->wr;
+		struct rte_mbuf *seg =
+			(void *)(elt->sge.addr - WR_ID(wr_id).offset);
+		struct rte_mbuf *rep;
+
+		/* Sanity checks. */
+		assert(WR_ID(wr_id).id < rxq->elts_n);
+		assert(wr_id == wr->wr_id);
+		assert(wr->sg_list == &elt->sge);
+		assert(wr->num_sge == 1);
+		/* Link completed WRs together for repost. */
+		*next = wr;
+		next = &wr->next;
+		if (unlikely(wc->status != IBV_WC_SUCCESS)) {
+			/* Whatever, just repost the offending WR. */
+			DEBUG("rxq=%p, wr_id=%" PRIu32 ": bad work completion"
+			      " status (%d): %s",
+			      (void *)rxq, WR_ID(wr_id).id, wc->status,
+			      ibv_wc_status_str(wc->status));
+#ifdef MLX4_PMD_SOFT_COUNTERS
+			/* Increase dropped packets counter. */
+			++rxq->stats.idropped;
+#endif
+			goto repost;
+		}
+		/*
+		 * Fetch initial bytes of packet descriptor into a
+		 * cacheline while allocating rep.
+		 */
+		rte_prefetch0(seg);
+		rep = __rte_mbuf_raw_alloc(rxq->mp);
+		if (unlikely(rep == NULL)) {
+			/*
+			 * Unable to allocate a replacement mbuf,
+			 * repost WR.
+			 */
+			DEBUG("rxq=%p, wr_id=%" PRIu32 ":"
+			      " can't allocate a new mbuf",
+			      (void *)rxq, WR_ID(wr_id).id);
+			/* Increase out of memory counters. */
+			++rxq->stats.rx_nombuf;
+			++rxq->priv->dev->data->rx_mbuf_alloc_failed;
+			goto repost;
+		}
+
+		/* Reconfigure sge to use rep instead of seg. */
+		elt->sge.addr = (uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM;
+		assert(elt->sge.lkey == rxq->mr->lkey);
+		WR_ID(wr->wr_id).offset =
+			(((uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM) -
+			 (uintptr_t)rep);
+		assert(WR_ID(wr->wr_id).id == WR_ID(wr_id).id);
+
+		/* Update seg information. */
+		SET_DATA_OFF(seg, RTE_PKTMBUF_HEADROOM);
+		NB_SEGS(seg) = 1;
+		PORT(seg) = rxq->port_id;
+		NEXT(seg) = NULL;
+		PKT_LEN(seg) = len;
+		DATA_LEN(seg) = len;
+		seg->ol_flags = 0;
+
+		/* Return packet. */
+		*(pkts++) = seg;
+		++ret;
+#ifdef MLX4_PMD_SOFT_COUNTERS
+		/* Increase bytes counter. */
+		rxq->stats.ibytes += wc->byte_len;
+#endif
+	repost:
+		continue;
+	}
+	*next = NULL;
+	/* Repost WRs. */
+#ifdef DEBUG_RECV
+	DEBUG("%p: reposting %d WRs starting from %" PRIu32 " (%p)",
+	      (void *)rxq, wcs_n, WR_ID(wcs[0].wr_id).id, (void *)head.next);
+#endif
+	i = ibv_post_recv(rxq->qp, head.next, &bad_wr);
+	if (unlikely(i)) {
+		/* Inability to repost WRs is fatal. */
+		DEBUG("%p: ibv_post_recv(): failed for WR %p: %s",
+		      (void *)rxq->priv,
+		      (void *)bad_wr,
+		      strerror(i));
+		abort();
+	}
+#ifdef MLX4_PMD_SOFT_COUNTERS
+	/* Increase packets counter. */
+	rxq->stats.ipackets += ret;
+#endif
+	return ret;
+}
+
+#ifdef INLINE_RECV
+
+/**
+ * Allocate a Queue Pair in case inline receive is supported.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param cq
+ *   Completion queue to associate with QP.
+ * @param desc
+ *   Number of descriptors in QP (hint only).
+ *
+ * @return
+ *   QP pointer or NULL in case of error.
+ */
+static struct ibv_qp *
+rxq_setup_qp(struct priv *priv, struct ibv_cq *cq, uint16_t desc)
+{
+	struct ibv_exp_qp_init_attr attr = {
+		/* CQ to be associated with the send queue. */
+		.send_cq = cq,
+		/* CQ to be associated with the receive queue. */
+		.recv_cq = cq,
+		.max_inl_recv = priv->inl_recv_size,
+		.cap = {
+			/* Max number of outstanding WRs. */
+			.max_recv_wr = ((priv->device_attr.max_qp_wr < desc) ?
+					priv->device_attr.max_qp_wr :
+					desc),
+			/* Max number of scatter/gather elements in a WR. */
+			.max_recv_sge = ((priv->device_attr.max_sge <
+					  MLX4_PMD_SGE_WR_N) ?
+					 priv->device_attr.max_sge :
+					 MLX4_PMD_SGE_WR_N),
+		},
+		.qp_type = IBV_QPT_RAW_PACKET,
+		.pd = priv->pd
+	};
+
+	attr.comp_mask = IBV_EXP_QP_INIT_ATTR_PD;
+	attr.comp_mask |= IBV_EXP_QP_INIT_ATTR_INL_RECV;
+
+	return ibv_exp_create_qp(priv->ctx, &attr);
+}
+
+#else /* INLINE_RECV */
+
+/**
+ * Allocate a Queue Pair.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param cq
+ *   Completion queue to associate with QP.
+ * @param desc
+ *   Number of descriptors in QP (hint only).
+ *
+ * @return
+ *   QP pointer or NULL in case of error.
+ */
+static struct ibv_qp *
+rxq_setup_qp(struct priv *priv, struct ibv_cq *cq, uint16_t desc)
+{
+	struct ibv_qp_init_attr attr = {
+		/* CQ to be associated with the send queue. */
+		.send_cq = cq,
+		/* CQ to be associated with the receive queue. */
+		.recv_cq = cq,
+		.cap = {
+			/* Max number of outstanding WRs. */
+			.max_recv_wr = ((priv->device_attr.max_qp_wr < desc) ?
+					priv->device_attr.max_qp_wr :
+					desc),
+			/* Max number of scatter/gather elements in a WR. */
+			.max_recv_sge = ((priv->device_attr.max_sge <
+					  MLX4_PMD_SGE_WR_N) ?
+					 priv->device_attr.max_sge :
+					 MLX4_PMD_SGE_WR_N),
+		},
+		.qp_type = IBV_QPT_RAW_PACKET
+	};
+
+	return ibv_create_qp(priv->pd, &attr);
+}
+
+#endif /* INLINE_RECV */
+
+#ifdef RSS_SUPPORT
+
+/**
+ * Allocate a RSS Queue Pair.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param cq
+ *   Completion queue to associate with QP.
+ * @param desc
+ *   Number of descriptors in QP (hint only).
+ * @param parent
+ *   If nonzero, create a parent QP, otherwise a child.
+ *
+ * @return
+ *   QP pointer or NULL in case of error.
+ */
+static struct ibv_qp *
+rxq_setup_qp_rss(struct priv *priv, struct ibv_cq *cq, uint16_t desc,
+		 int parent)
+{
+	struct ibv_exp_qp_init_attr attr = {
+		/* CQ to be associated with the send queue. */
+		.send_cq = cq,
+		/* CQ to be associated with the receive queue. */
+		.recv_cq = cq,
+#ifdef INLINE_RECV
+		.max_inl_recv = priv->inl_recv_size,
+#endif
+		.cap = {
+			/* Max number of outstanding WRs. */
+			.max_recv_wr = ((priv->device_attr.max_qp_wr < desc) ?
+					priv->device_attr.max_qp_wr :
+					desc),
+			/* Max number of scatter/gather elements in a WR. */
+			.max_recv_sge = ((priv->device_attr.max_sge <
+					  MLX4_PMD_SGE_WR_N) ?
+					 priv->device_attr.max_sge :
+					 MLX4_PMD_SGE_WR_N),
+		},
+		.qp_type = IBV_QPT_RAW_PACKET,
+		.comp_mask = (IBV_EXP_QP_INIT_ATTR_PD |
+			      IBV_EXP_QP_INIT_ATTR_QPG),
+		.pd = priv->pd
+	};
+
+#ifdef INLINE_RECV
+	attr.comp_mask |= IBV_EXP_QP_INIT_ATTR_INL_RECV;
+#endif
+	if (parent) {
+		attr.qpg.qpg_type = IBV_EXP_QPG_PARENT;
+		/* TSS isn't necessary. */
+		attr.qpg.parent_attrib.tss_child_count = 0;
+		attr.qpg.parent_attrib.rss_child_count = priv->rxqs_n;
+		DEBUG("initializing parent RSS queue");
+	}
+	else {
+		attr.qpg.qpg_type = IBV_EXP_QPG_CHILD_RX;
+		attr.qpg.qpg_parent = priv->rxq_parent.qp;
+		DEBUG("initializing child RSS queue");
+	}
+	return ibv_exp_create_qp(priv->ctx, &attr);
+}
+
+#endif /* RSS_SUPPORT */
+
+/**
+ * Reconfigure a RX queue with new parameters.
+ *
+ * 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
+ *   RX queue pointer.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+rxq_rehash(struct rte_eth_dev *dev, struct rxq *rxq)
+{
+	struct priv *priv = rxq->priv;
+	struct rxq tmpl = *rxq;
+	unsigned int mbuf_n;
+	unsigned int desc_n;
+	struct rte_mbuf **pool;
+	unsigned int i, k;
+	struct ibv_exp_qp_attr mod;
+	struct ibv_recv_wr *bad_wr;
+	int err;
+	int parent = (rxq == &priv->rxq_parent);
+
+	if (parent) {
+		ERROR("%p: cannot rehash parent queue %p",
+		      (void *)dev, (void *)rxq);
+		return EINVAL;
+	}
+	DEBUG("%p: rehashing queue %p", (void *)dev, (void *)rxq);
+	/* Number of descriptors and mbufs currently allocated. */
+	desc_n = (tmpl.elts_n * (tmpl.sp ? MLX4_PMD_SGE_WR_N : 1));
+	mbuf_n = desc_n;
+	/* Enable scattered packets support for this queue if necessary. */
+	if ((dev->data->dev_conf.rxmode.jumbo_frame) &&
+	    (dev->data->dev_conf.rxmode.max_rx_pkt_len >
+	     (tmpl.mb_len - RTE_PKTMBUF_HEADROOM))) {
+		tmpl.sp = 1;
+		desc_n /= MLX4_PMD_SGE_WR_N;
+	}
+	else
+		tmpl.sp = 0;
+	DEBUG("%p: %s scattered packets support (%u WRs)",
+	      (void *)dev, (tmpl.sp ? "enabling" : "disabling"), desc_n);
+	/* If scatter mode is the same as before, nothing to do. */
+	if (tmpl.sp == rxq->sp) {
+		DEBUG("%p: nothing to do", (void *)dev);
+		return 0;
+	}
+	/* Remove attached flows if RSS is disabled (no parent queue). */
+	if (!priv->rss) {
+		rxq_allmulticast_disable(&tmpl);
+		rxq_promiscuous_disable(&tmpl);
+		rxq_mac_addrs_del(&tmpl);
+		/* Update original queue in case of failure. */
+		rxq->allmulti_flow = tmpl.allmulti_flow;
+		rxq->promisc_flow = tmpl.promisc_flow;
+		memcpy(rxq->mac_configured, tmpl.mac_configured,
+		       sizeof(rxq->mac_configured));
+		memcpy(rxq->mac_flow, tmpl.mac_flow, sizeof(rxq->mac_flow));
+	}
+	/* From now on, any failure will render the queue unusable.
+	 * Reinitialize QP. */
+	mod = (struct ibv_exp_qp_attr){ .qp_state = IBV_QPS_RESET };
+	if ((err = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE))) {
+		ERROR("%p: cannot reset QP: %s", (void *)dev, strerror(err));
+		assert(err > 0);
+		return err;
+	}
+	if ((err = ibv_resize_cq(tmpl.cq, desc_n))) {
+		ERROR("%p: cannot resize CQ: %s", (void *)dev, strerror(err));
+		assert(err > 0);
+		return err;
+	}
+	mod = (struct ibv_exp_qp_attr){
+		/* Move the QP to this state. */
+		.qp_state = IBV_QPS_INIT,
+		/* Primary port number. */
+		.port_num = priv->port
+	};
+	if ((err = ibv_exp_modify_qp(tmpl.qp, &mod,
+				 (IBV_EXP_QP_STATE |
+#ifdef RSS_SUPPORT
+				  (parent ? IBV_EXP_QP_GROUP_RSS : 0) |
+#endif /* RSS_SUPPORT */
+				  IBV_EXP_QP_PORT)))) {
+		ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
+		      (void *)dev, strerror(err));
+		assert(err > 0);
+		return err;
+	};
+	/* Reconfigure flows. Do not care for errors. */
+	if (!priv->rss) {
+		rxq_mac_addrs_add(&tmpl);
+		if (priv->promisc)
+			rxq_promiscuous_enable(&tmpl);
+		if (priv->allmulti)
+			rxq_allmulticast_enable(&tmpl);
+		/* Update original queue in case of failure. */
+		rxq->allmulti_flow = tmpl.allmulti_flow;
+		rxq->promisc_flow = tmpl.promisc_flow;
+		memcpy(rxq->mac_configured, tmpl.mac_configured,
+		       sizeof(rxq->mac_configured));
+		memcpy(rxq->mac_flow, tmpl.mac_flow, sizeof(rxq->mac_flow));
+	}
+	/* Allocate pool. */
+	pool = rte_malloc(__func__, (mbuf_n * sizeof(*pool)), 0);
+	if (pool == NULL) {
+		ERROR("%p: cannot allocate memory", (void *)dev);
+		return ENOBUFS;
+	}
+	/* Snatch mbufs from original queue. */
+	k = 0;
+	if (rxq->sp) {
+		struct rxq_elt_sp (*elts)[rxq->elts_n] = rxq->elts.sp;
+
+		for (i = 0; (i != elemof(*elts)); ++i) {
+			struct rxq_elt_sp *elt = &(*elts)[i];
+			unsigned int j;
+
+			for (j = 0; (j != elemof(elt->bufs)); ++j) {
+				assert(elt->bufs[j] != NULL);
+				pool[k++] = elt->bufs[j];
+			}
+		}
+	}
+	else {
+		struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts.no_sp;
+
+		for (i = 0; (i != elemof(*elts)); ++i) {
+			struct rxq_elt *elt = &(*elts)[i];
+			struct rte_mbuf *buf = (void *)
+				(elt->sge.addr - WR_ID(elt->wr.wr_id).offset);
+
+			assert(WR_ID(elt->wr.wr_id).id == i);
+			pool[k++] = buf;
+		}
+        }
+	assert(k == mbuf_n);
+	tmpl.elts_n = 0;
+	tmpl.elts.sp = NULL;
+	assert((void *)&tmpl.elts.sp == (void *)&tmpl.elts.no_sp);
+	err = ((tmpl.sp) ?
+	       rxq_alloc_elts_sp(&tmpl, desc_n, pool) :
+	       rxq_alloc_elts(&tmpl, desc_n, pool));
+	if (err) {
+		ERROR("%p: cannot reallocate WRs, aborting", (void *)dev);
+		rte_free(pool);
+		assert(err > 0);
+		return err;
+	}
+	assert(tmpl.elts_n == desc_n);
+	assert(tmpl.elts.sp != NULL);
+	rte_free(pool);
+	/* Clean up original data. */
+	rxq->elts_n = 0;
+	rte_free(rxq->elts.sp);
+	rxq->elts.sp = NULL;
+	/* Post WRs. */
+	if ((err = ibv_post_recv(tmpl.qp,
+				 (tmpl.sp ?
+				  &(*tmpl.elts.sp)[0].wr :
+				  &(*tmpl.elts.no_sp)[0].wr),
+				 &bad_wr))) {
+		ERROR("%p: ibv_post_recv() failed for WR %p: %s",
+		      (void *)dev,
+		      (void *)bad_wr,
+		      strerror(err));
+		goto skip_rtr;
+	}
+	mod = (struct ibv_exp_qp_attr){
+		.qp_state = IBV_QPS_RTR
+	};
+	if ((err = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE)))
+		ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
+		      (void *)dev, strerror(err));
+skip_rtr:
+	*rxq = tmpl;
+	assert(err >= 0);
+	return err;
+}
+
+/**
+ * Configure a RX queue.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param rxq
+ *   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.
+ */
+static int
+rxq_setup(struct rte_eth_dev *dev, struct rxq *rxq, 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 tmpl = {
+		.priv = priv,
+		.mp = mp,
+		.socket = socket
+	};
+	struct ibv_exp_qp_attr mod;
+	struct ibv_recv_wr *bad_wr;
+	struct rte_mbuf *buf;
+	int ret = 0;
+	int parent = (rxq == &priv->rxq_parent);
+
+	(void)conf; /* Thresholds configuration (ignored). */
+	/*
+	 * If this is a parent queue, hardware must support RSS and
+	 * RSS must be enabled.
+	 */
+	assert((!parent) || ((priv->hw_rss) && (priv->rss)));
+	if (parent) {
+		/* Even if unused, ibv_create_cq() requires at least one
+		 * descriptor. */
+		desc = 1;
+		goto skip_mr;
+	}
+	if ((desc == 0) || (desc % MLX4_PMD_SGE_WR_N)) {
+		ERROR("%p: invalid number of RX descriptors (must be a"
+		      " multiple of %d)", (void *)dev, desc);
+		return EINVAL;
+	}
+	/* Get mbuf length. */
+	buf = rte_pktmbuf_alloc(mp);
+	if (buf == NULL) {
+		ERROR("%p: unable to allocate mbuf", (void *)dev);
+		return ENOMEM;
+	}
+	tmpl.mb_len = buf->buf_len;
+	assert((rte_pktmbuf_headroom(buf) +
+		rte_pktmbuf_tailroom(buf)) == tmpl.mb_len);
+	assert(rte_pktmbuf_headroom(buf) == RTE_PKTMBUF_HEADROOM);
+	rte_pktmbuf_free(buf);
+	/* Enable scattered packets support for this queue if necessary. */
+	if ((dev->data->dev_conf.rxmode.jumbo_frame) &&
+	    (dev->data->dev_conf.rxmode.max_rx_pkt_len >
+	     (tmpl.mb_len - RTE_PKTMBUF_HEADROOM))) {
+		tmpl.sp = 1;
+		desc /= MLX4_PMD_SGE_WR_N;
+	}
+	DEBUG("%p: %s scattered packets support (%u WRs)",
+	      (void *)dev, (tmpl.sp ? "enabling" : "disabling"), desc);
+	/* Use the entire RX mempool as the memory region. */
+	tmpl.mr = ibv_reg_mr(priv->pd,
+		             (void *)mp->elt_va_start,
+			     (mp->elt_va_end - mp->elt_va_start),
+			     (IBV_ACCESS_LOCAL_WRITE |
+			      IBV_ACCESS_REMOTE_WRITE));
+	if (tmpl.mr == NULL) {
+		ret = EINVAL;
+		ERROR("%p: MR creation failure: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+skip_mr:
+	tmpl.cq = ibv_create_cq(priv->ctx, desc, NULL, NULL, 0);
+	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);
+#ifdef RSS_SUPPORT
+	if (priv->rss)
+		tmpl.qp = rxq_setup_qp_rss(priv, tmpl.cq, desc, parent);
+	else
+#endif /* RSS_SUPPORT */
+		tmpl.qp = rxq_setup_qp(priv, tmpl.cq, desc);
+	if (tmpl.qp == NULL) {
+		ret = (errno ? errno : EINVAL);
+		ERROR("%p: QP creation failure: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	mod = (struct ibv_exp_qp_attr){
+		/* Move the QP to this state. */
+		.qp_state = IBV_QPS_INIT,
+		/* Primary port number. */
+		.port_num = priv->port
+	};
+	if ((ret = ibv_exp_modify_qp(tmpl.qp, &mod,
+				     (IBV_EXP_QP_STATE |
+#ifdef RSS_SUPPORT
+				      (parent ? IBV_EXP_QP_GROUP_RSS : 0) |
+#endif /* RSS_SUPPORT */
+				      IBV_EXP_QP_PORT)))) {
+		ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	if ((parent) || (!priv->rss))  {
+		/* Configure MAC and broadcast addresses. */
+		if ((ret = rxq_mac_addrs_add(&tmpl))) {
+			ERROR("%p: QP flow attachment failed: %s",
+			      (void *)dev, strerror(ret));
+			goto error;
+		}
+	}
+	/* Allocate descriptors for RX queues, except for the RSS parent. */
+	if (parent)
+		goto skip_alloc;
+	if (tmpl.sp)
+		ret = rxq_alloc_elts_sp(&tmpl, desc, NULL);
+	else
+		ret = rxq_alloc_elts(&tmpl, desc, NULL);
+	if (ret) {
+		ERROR("%p: RXQ allocation failed: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	if ((ret = ibv_post_recv(tmpl.qp,
+				 (tmpl.sp ?
+				  &(*tmpl.elts.sp)[0].wr :
+				  &(*tmpl.elts.no_sp)[0].wr),
+				 &bad_wr))) {
+		ERROR("%p: ibv_post_recv() failed for WR %p: %s",
+		      (void *)dev,
+		      (void *)bad_wr,
+		      strerror(ret));
+		goto error;
+	}
+skip_alloc:
+	mod = (struct ibv_exp_qp_attr){
+		.qp_state = IBV_QPS_RTR
+	};
+	if ((ret = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE))) {
+		ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
+		      (void *)dev, strerror(ret));
+		goto error;
+	}
+	/* Save port ID. */
+	tmpl.port_id = dev->data->port_id;
+	DEBUG("%p: RTE port ID: %u", (void *)rxq, tmpl.port_id);
+	/* Clean up rxq in case we're reinitializing it. */
+	DEBUG("%p: cleaning-up old rxq just in case", (void *)rxq);
+	rxq_cleanup(rxq);
+	*rxq = tmpl;
+	DEBUG("%p: rxq updated with %p", (void *)rxq, (void *)&tmpl);
+	assert(ret == 0);
+	return 0;
+error:
+	rxq_cleanup(&tmpl);
+	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.
+ */
+static int
+mlx4_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];
+	int ret;
+
+	priv_lock(priv);
+	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);
+	}
+	else {
+		rxq = rte_calloc_socket("RXQ", 1, sizeof(*rxq), 0, socket);
+		if (rxq == NULL) {
+			ERROR("%p: unable to allocate queue index %u",
+			      (void *)dev, idx);
+			priv_unlock(priv);
+			return -ENOMEM;
+		}
+	}
+	ret = rxq_setup(dev, rxq, desc, socket, conf, mp);
+	if (ret)
+		rte_free(rxq);
+	else {
+		rxq->stats.idx = idx;
+		DEBUG("%p: adding RX queue %p to list",
+		      (void *)dev, (void *)rxq);
+		(*priv->rxqs)[idx] = rxq;
+		/* Update receive callback. */
+		if (rxq->sp)
+			dev->rx_pkt_burst = mlx4_rx_burst_sp;
+		else
+			dev->rx_pkt_burst = mlx4_rx_burst;
+	}
+	priv_unlock(priv);
+	return -ret;
+}
+
+/**
+ * DPDK callback to release a RX queue.
+ *
+ * @param dpdk_rxq
+ *   Generic RX queue pointer.
+ */
+static void
+mlx4_rx_queue_release(void *dpdk_rxq)
+{
+	struct rxq *rxq = (struct rxq *)dpdk_rxq;
+	struct priv *priv;
+	unsigned int i;
+
+	if (rxq == NULL)
+		return;
+	priv = rxq->priv;
+	priv_lock(priv);
+	assert(rxq != &priv->rxq_parent);
+	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);
+			(*priv->rxqs)[i] = NULL;
+			break;
+		}
+	rxq_cleanup(rxq);
+	rte_free(rxq);
+	priv_unlock(priv);
+}
+
+/**
+ * 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.
+ */
+static int
+mlx4_dev_start(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	unsigned int i = 0;
+	unsigned int r;
+	struct rxq *rxq;
+
+	priv_lock(priv);
+	if (priv->started) {
+		priv_unlock(priv);
+		return 0;
+	}
+	DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
+	priv->started = 1;
+	if (priv->rss) {
+		rxq = &priv->rxq_parent;
+		r = 1;
+	}
+	else {
+		rxq = (*priv->rxqs)[0];
+		r = priv->rxqs_n;
+	}
+	/* Iterate only once when RSS is enabled. */
+	do {
+		int ret;
+
+		/* Ignore nonexistent RX queues. */
+		if (rxq == NULL)
+			continue;
+		if (((ret = rxq_mac_addrs_add(rxq)) == 0) &&
+		    ((!priv->promisc) ||
+		     ((ret = rxq_promiscuous_enable(rxq)) == 0)) &&
+		    ((!priv->allmulti) ||
+		     ((ret = rxq_allmulticast_enable(rxq)) == 0)))
+			continue;
+		WARN("%p: QP flow attachment failed: %s",
+		     (void *)dev, strerror(ret));
+		/* Rollback. */
+		while (i != 0)
+			if ((rxq = (*priv->rxqs)[--i]) != NULL) {
+				rxq_allmulticast_disable(rxq);
+				rxq_promiscuous_disable(rxq);
+				rxq_mac_addrs_del(rxq);
+			}
+		priv->started = 0;
+		return -ret;
+	}
+	while ((--r) && ((rxq = (*priv->rxqs)[++i]), i));
+	priv_unlock(priv);
+	return 0;
+}
+
+/**
+ * DPDK callback to stop the device.
+ *
+ * Simulate device stop by detaching all configured flows.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ */
+static void
+mlx4_dev_stop(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	unsigned int i = 0;
+	unsigned int r;
+	struct rxq *rxq;
+
+	priv_lock(priv);
+	if (!priv->started) {
+		priv_unlock(priv);
+		return;
+	}
+	DEBUG("%p: detaching flows from all RX queues", (void *)dev);
+	priv->started = 0;
+	if (priv->rss) {
+		rxq = &priv->rxq_parent;
+		r = 1;
+	}
+	else {
+		rxq = (*priv->rxqs)[0];
+		r = priv->rxqs_n;
+	}
+	/* Iterate only once when RSS is enabled. */
+	do {
+		/* Ignore nonexistent RX queues. */
+		if (rxq == NULL)
+			continue;
+		rxq_allmulticast_disable(rxq);
+		rxq_promiscuous_disable(rxq);
+		rxq_mac_addrs_del(rxq);
+	}
+	while ((--r) && ((rxq = (*priv->rxqs)[++i]), i));
+	priv_unlock(priv);
+}
+
+/**
+ * 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).
+ */
+static 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).
+ */
+static 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;
+}
+
+/**
+ * DPDK callback to close the device.
+ *
+ * Destroy all queues and objects, free memory.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ */
+static void
+mlx4_dev_close(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	void *tmp;
+	unsigned int i;
+
+	priv_lock(priv);
+	DEBUG("%p: closing device \"%s\"",
+	      (void *)dev,
+	      ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
+	/* Prevent crashes when queues are still in use. This is unfortunately
+	 * still required for DPDK 1.3 because some programs (such as testpmd)
+	 * never release them before closing the device. */
+	dev->rx_pkt_burst = removed_rx_burst;
+	dev->tx_pkt_burst = removed_tx_burst;
+	if (priv->rxqs != NULL) {
+		/* XXX race condition if mlx4_rx_burst() is still running. */
+		usleep(1000);
+		for (i = 0; (i != priv->rxqs_n); ++i) {
+			tmp = (*priv->rxqs)[i];
+			if (tmp == NULL)
+				continue;
+			(*priv->rxqs)[i] = NULL;
+			rxq_cleanup(tmp);
+			rte_free(tmp);
+		}
+		priv->rxqs_n = 0;
+		priv->rxqs = NULL;
+	}
+	if (priv->txqs != NULL) {
+		/* XXX race condition if mlx4_tx_burst() is still running. */
+		usleep(1000);
+		for (i = 0; (i != priv->txqs_n); ++i) {
+			tmp = (*priv->txqs)[i];
+			if (tmp == NULL)
+				continue;
+			(*priv->txqs)[i] = NULL;
+			txq_cleanup(tmp);
+			rte_free(tmp);
+		}
+		priv->txqs_n = 0;
+		priv->txqs = NULL;
+	}
+	if (priv->rss)
+		rxq_cleanup(&priv->rxq_parent);
+	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);
+	priv_unlock(priv);
+	memset(priv, 0, sizeof(*priv));
+}
+
+/**
+ * DPDK callback to get information about the device.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param[out] info
+ *   Info structure output buffer.
+ */
+static void
+mlx4_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
+{
+	struct priv *priv = dev->data->dev_private;
+	unsigned int max;
+
+	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 = elemof(priv->mac);
+	priv_unlock(priv);
+}
+
+/**
+ * DPDK callback to get device statistics.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param[out] stats
+ *   Stats structure output buffer.
+ */
+static void
+mlx4_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
+{
+	struct priv *priv = dev->data->dev_private;
+	struct rte_eth_stats tmp = { .ipackets = 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 MLX4_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 MLX4_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 MLX4_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 MLX4_PMD_SOFT_COUNTERS
+		tmp.opackets += txq->stats.opackets;
+		tmp.obytes += txq->stats.obytes;
+#endif
+		tmp.oerrors += txq->stats.odropped;
+	}
+#ifndef MLX4_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.
+ */
+static void
+mlx4_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 mlx4_rxq_stats){ .idx = idx };
+	}
+	for (i = 0; (i != priv->txqs_n); ++i) {
+		if ((*priv->txqs)[i] == NULL)
+			continue;
+		idx = (*priv->rxqs)[i]->stats.idx;
+		(*priv->txqs)[i]->stats =
+			(struct mlx4_txq_stats){ .idx = idx };
+	}
+#ifndef MLX4_PMD_SOFT_COUNTERS
+	/* FIXME: reset hardware counters. */
+#endif
+	priv_unlock(priv);
+}
+
+/**
+ * DPDK callback to remove a MAC address.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param index
+ *   MAC address index.
+ */
+static void
+mlx4_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
+{
+	struct priv *priv = dev->data->dev_private;
+
+	priv_lock(priv);
+	DEBUG("%p: removing MAC address from index %" PRIu32,
+	      (void *)dev, index);
+	if (index >= MLX4_MAX_MAC_ADDRESSES)
+		goto end;
+	/* Refuse to remove the broadcast address, this one is special. */
+	if (!memcmp(priv->mac[index].addr_bytes, "\xff\xff\xff\xff\xff\xff",
+		    ETHER_ADDR_LEN))
+		goto end;
+	priv_mac_addr_del(priv, index);
+end:
+	priv_unlock(priv);
+}
+
+/**
+ * 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).
+ */
+static void
+mlx4_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;
+
+	(void)vmdq;
+	priv_lock(priv);
+	DEBUG("%p: adding MAC address at index %" PRIu32,
+	      (void *)dev, index);
+	if (index >= MLX4_MAX_MAC_ADDRESSES)
+		goto end;
+	/* Refuse to add the broadcast address, this one is special. */
+	if (!memcmp(mac_addr->addr_bytes, "\xff\xff\xff\xff\xff\xff",
+		    ETHER_ADDR_LEN))
+		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 enable promiscuous mode.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ */
+static void
+mlx4_promiscuous_enable(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	unsigned int i;
+	int ret;
+
+	priv_lock(priv);
+	if (priv->promisc) {
+		priv_unlock(priv);
+		return;
+	}
+	/* If device isn't started, this is all we need to do. */
+	if (!priv->started)
+		goto end;
+	if (priv->rss) {
+		ret = rxq_promiscuous_enable(&priv->rxq_parent);
+		if (ret) {
+			priv_unlock(priv);
+			return;
+		}
+		goto end;
+	}
+	for (i = 0; (i != priv->rxqs_n); ++i) {
+		if ((*priv->rxqs)[i] == NULL)
+			continue;
+		ret = rxq_promiscuous_enable((*priv->rxqs)[i]);
+		if (!ret)
+			continue;
+		/* Failure, rollback. */
+		while (i != 0)
+			if ((*priv->rxqs)[--i] != NULL)
+				rxq_promiscuous_disable((*priv->rxqs)[i]);
+		priv_unlock(priv);
+		return;
+	}
+end:
+	priv->promisc = 1;
+	priv_unlock(priv);
+}
+
+/**
+ * DPDK callback to disable promiscuous mode.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ */
+static void
+mlx4_promiscuous_disable(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	unsigned int i;
+
+	priv_lock(priv);
+	if (!priv->promisc) {
+		priv_unlock(priv);
+		return;
+	}
+	if (priv->rss) {
+		rxq_promiscuous_disable(&priv->rxq_parent);
+		goto end;
+	}
+	for (i = 0; (i != priv->rxqs_n); ++i)
+		if ((*priv->rxqs)[i] != NULL)
+			rxq_promiscuous_disable((*priv->rxqs)[i]);
+end:
+	priv->promisc = 0;
+	priv_unlock(priv);
+}
+
+/**
+ * DPDK callback to enable allmulti mode.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ */
+static void
+mlx4_allmulticast_enable(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	unsigned int i;
+	int ret;
+
+	priv_lock(priv);
+	if (priv->allmulti) {
+		priv_unlock(priv);
+		return;
+	}
+	/* If device isn't started, this is all we need to do. */
+	if (!priv->started)
+		goto end;
+	if (priv->rss) {
+		ret = rxq_allmulticast_enable(&priv->rxq_parent);
+		if (ret) {
+			priv_unlock(priv);
+			return;
+		}
+		goto end;
+	}
+	for (i = 0; (i != priv->rxqs_n); ++i) {
+		if ((*priv->rxqs)[i] == NULL)
+			continue;
+		ret = rxq_allmulticast_enable((*priv->rxqs)[i]);
+		if (!ret)
+			continue;
+		/* Failure, rollback. */
+		while (i != 0)
+			if ((*priv->rxqs)[--i] != NULL)
+				rxq_allmulticast_disable((*priv->rxqs)[i]);
+		priv_unlock(priv);
+		return;
+	}
+end:
+	priv->allmulti = 1;
+	priv_unlock(priv);
+}
+
+/**
+ * DPDK callback to disable allmulti mode.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ */
+static void
+mlx4_allmulticast_disable(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	unsigned int i;
+
+	priv_lock(priv);
+	if (!priv->allmulti) {
+		priv_unlock(priv);
+		return;
+	}
+	if (priv->rss) {
+		rxq_allmulticast_disable(&priv->rxq_parent);
+		goto end;
+	}
+	for (i = 0; (i != priv->rxqs_n); ++i)
+		if ((*priv->rxqs)[i] != NULL)
+			rxq_allmulticast_disable((*priv->rxqs)[i]);
+end:
+	priv->allmulti = 0;
+	priv_unlock(priv);
+}
+
+/**
+ * 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
+mlx4_link_update_unlocked(struct rte_eth_dev *dev, int wait_to_complete)
+{
+	struct priv *priv = dev->data->dev_private;
+	struct ibv_port_attr port_attr;
+	static const uint8_t width_mult[] = {
+		/* Multiplier values taken from devinfo.c in libibverbs. */
+		0, 1, 4, 0, 8, 0, 0, 0, 12, 0
+	};
+
+	(void)wait_to_complete;
+	if ((errno = ibv_query_port(priv->ctx, priv->port, &port_attr))) {
+		WARN("port query failed: %s", strerror(errno));
+		return -1;
+	}
+	dev->data->dev_link = (struct rte_eth_link){
+		.link_speed = (ibv_rate_to_mbps(mult_to_ibv_rate
+						(port_attr.active_speed)) *
+			       width_mult[(port_attr.active_width %
+					   sizeof(width_mult))]),
+		.link_duplex = ETH_LINK_FULL_DUPLEX,
+		.link_status = (port_attr.state == IBV_PORT_ACTIVE)
+	};
+	if (memcmp(&port_attr, &priv->port_attr, sizeof(port_attr))) {
+		/* Link status changed. */
+		priv->port_attr = port_attr;
+		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).
+ */
+static int
+mlx4_link_update(struct rte_eth_dev *dev, int wait_to_complete)
+{
+	struct priv *priv = dev->data->dev_private;
+	int ret;
+
+	priv_lock(priv);
+	ret = mlx4_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.
+ */
+static int
+mlx4_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) =
+		mlx4_rx_burst;
+
+	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);
+	priv->mtu = 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 mlx4_rx_burst() and uses
+	 * removed_rx_burst() instead. */
+	rte_wmb();
+	usleep(1000);
+	/* Reconfigure each RX queue. */
+	for (i = 0; (i != priv->rxqs_n); ++i) {
+		struct rxq *rxq = (*priv->rxqs)[i];
+		unsigned int max_frame_len;
+		int sp;
+
+		if (rxq == NULL)
+			continue;
+		/* Calculate new maximum frame length according to MTU and
+		 * toggle scattered support (sp) if necessary. */
+		max_frame_len = (priv->mtu + ETHER_HDR_LEN +
+				 (ETHER_MAX_VLAN_FRAME_LEN - ETHER_MAX_LEN));
+		sp = (max_frame_len > (rxq->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 ((ret = rxq_rehash(dev, rxq))) {
+			/* Force SP RX if that queue requires it and abort. */
+			if (rxq->sp)
+				rx_func = mlx4_rx_burst_sp;
+			break;
+		}
+		/* Reenable non-RSS queue attributes. No need to check
+		 * for errors at this stage. */
+		if (!priv->rss) {
+			rxq_mac_addrs_add(rxq);
+			if (priv->promisc)
+				rxq_promiscuous_enable(rxq);
+			if (priv->allmulti)
+				rxq_allmulticast_enable(rxq);
+		}
+		/* Scattered burst function takes priority. */
+		if (rxq->sp)
+			rx_func = mlx4_rx_burst_sp;
+	}
+	/* 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.
+ */
+static int
+mlx4_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;
+
+	ifr.ifr_data = &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.
+ */
+static int
+mlx4_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;
+
+	ifr.ifr_data = &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;
+}
+
+/**
+ * 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;
+	unsigned int j = -1;
+
+	DEBUG("%p: %s VLAN filter ID %" PRIu16,
+	      (void *)dev, (on ? "enable" : "disable"), vlan_id);
+	for (i = 0; (i != elemof(priv->vlan_filter)); ++i) {
+		if (!priv->vlan_filter[i].enabled) {
+			/* Unused index, remember it. */
+			j = i;
+			continue;
+		}
+		if (priv->vlan_filter[i].id != vlan_id)
+			continue;
+		/* This VLAN ID is already known, use its index. */
+		j = i;
+		break;
+	}
+	/* Check if there's room for another VLAN filter. */
+	if (j == (unsigned int)-1)
+		return ENOMEM;
+	/*
+	 * VLAN filters apply to all configured MAC addresses, flow
+	 * specifications must be reconfigured accordingly.
+	 */
+	priv->vlan_filter[j].id = vlan_id;
+	if ((on) && (!priv->vlan_filter[j].enabled)) {
+		/*
+		 * Filter is disabled, enable it.
+		 * Rehashing flows in all RX queues is necessary.
+		 */
+		if (priv->rss)
+			rxq_mac_addrs_del(&priv->rxq_parent);
+		else
+			for (i = 0; (i != priv->rxqs_n); ++i)
+				if ((*priv->rxqs)[i] != NULL)
+					rxq_mac_addrs_del((*priv->rxqs)[i]);
+		priv->vlan_filter[j].enabled = 1;
+		if (priv->started) {
+			if (priv->rss)
+				rxq_mac_addrs_add(&priv->rxq_parent);
+			else
+				for (i = 0; (i != priv->rxqs_n); ++i) {
+					if ((*priv->rxqs)[i] == NULL)
+						continue;
+					rxq_mac_addrs_add((*priv->rxqs)[i]);
+				}
+		}
+	}
+	else if ((!on) && (priv->vlan_filter[j].enabled)) {
+		/*
+		 * Filter is enabled, disable it.
+		 * Rehashing flows in all RX queues is necessary.
+		 */
+		if (priv->rss)
+			rxq_mac_addrs_del(&priv->rxq_parent);
+		else
+			for (i = 0; (i != priv->rxqs_n); ++i)
+				if ((*priv->rxqs)[i] != NULL)
+					rxq_mac_addrs_del((*priv->rxqs)[i]);
+		priv->vlan_filter[j].enabled = 0;
+		if (priv->started) {
+			if (priv->rss)
+				rxq_mac_addrs_add(&priv->rxq_parent);
+			else
+				for (i = 0; (i != priv->rxqs_n); ++i) {
+					if ((*priv->rxqs)[i] == NULL)
+						continue;
+					rxq_mac_addrs_add((*priv->rxqs)[i]);
+				}
+		}
+	}
+	return 0;
+}
+
+/**
+ * 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.
+ */
+static int
+mlx4_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;
+}
+
+static struct eth_dev_ops mlx4_dev_ops = {
+	.dev_configure = mlx4_dev_configure,
+	.dev_start = mlx4_dev_start,
+	.dev_stop = mlx4_dev_stop,
+	.dev_close = mlx4_dev_close,
+	.promiscuous_enable = mlx4_promiscuous_enable,
+	.promiscuous_disable = mlx4_promiscuous_disable,
+	.allmulticast_enable = mlx4_allmulticast_enable,
+	.allmulticast_disable = mlx4_allmulticast_disable,
+	.link_update = mlx4_link_update,
+	.stats_get = mlx4_stats_get,
+	.stats_reset = mlx4_stats_reset,
+	.queue_stats_mapping_set = NULL,
+	.dev_infos_get = mlx4_dev_infos_get,
+	.vlan_filter_set = mlx4_vlan_filter_set,
+	.vlan_tpid_set = NULL,
+	.vlan_strip_queue_set = NULL,
+	.vlan_offload_set = NULL,
+	.rx_queue_setup = mlx4_rx_queue_setup,
+	.tx_queue_setup = mlx4_tx_queue_setup,
+	.rx_queue_release = mlx4_rx_queue_release,
+	.tx_queue_release = mlx4_tx_queue_release,
+	.dev_led_on = NULL,
+	.dev_led_off = NULL,
+	.flow_ctrl_get = mlx4_dev_get_flow_ctrl,
+	.flow_ctrl_set = mlx4_dev_set_flow_ctrl,
+	.priority_flow_ctrl_set = NULL,
+	.mac_addr_remove = mlx4_mac_addr_remove,
+	.mac_addr_add = mlx4_mac_addr_add,
+	.mtu_set = mlx4_dev_set_mtu,
+	.fdir_add_signature_filter = NULL,
+	.fdir_update_signature_filter = NULL,
+	.fdir_remove_signature_filter = NULL,
+	.fdir_add_perfect_filter = NULL,
+	.fdir_update_perfect_filter = NULL,
+	.fdir_remove_perfect_filter = NULL,
+	.fdir_set_masks = NULL
+};
+
+/**
+ * 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.
+ */
+static int
+mlx4_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;
+}
+
+/**
+ * Derive MAC address from port GID.
+ *
+ * @param[out] mac
+ *   MAC address output buffer.
+ * @param port
+ *   Physical port number.
+ * @param[in] gid
+ *   Port GID.
+ */
+static void
+mac_from_gid(uint8_t (*mac)[ETHER_ADDR_LEN], uint32_t port, uint8_t *gid)
+{
+	memcpy(&(*mac)[0], gid + 8, 3);
+	memcpy(&(*mac)[3], gid + 13, 3);
+	if (port == 1)
+		(*mac)[0] ^= 2;
+}
+
+/* Support up to 32 adapters. */
+static struct {
+	struct rte_pci_addr pci_addr; /* associated PCI address */
+	uint32_t ports; /* physical ports bitfield. */
+} mlx4_dev[32];
+
+/**
+ * Get device index in mlx4_dev[] from PCI bus address.
+ *
+ * @param[in] pci_addr
+ *   PCI bus address to look for.
+ *
+ * @return
+ *   mlx4_dev[] index on success, -1 on failure.
+ */
+static int
+mlx4_dev_idx(struct rte_pci_addr *pci_addr)
+{
+	unsigned int i;
+	int ret = -1;
+
+	assert(pci_addr != NULL);
+	for (i = 0; (i != elemof(mlx4_dev)); ++i) {
+		if ((mlx4_dev[i].pci_addr.domain == pci_addr->domain) &&
+		    (mlx4_dev[i].pci_addr.bus == pci_addr->bus) &&
+		    (mlx4_dev[i].pci_addr.devid == pci_addr->devid) &&
+		    (mlx4_dev[i].pci_addr.function == pci_addr->function))
+			return i;
+		if ((mlx4_dev[i].ports == 0) && (ret == -1))
+			ret = i;
+	}
+	return ret;
+}
+
+/**
+ * Retrieve integer value from environment variable.
+ *
+ * @param[in] name
+ *   Environment variable name.
+ *
+ * @return
+ *   Integer value, 0 if the variable is not set.
+ */
+static int
+mlx4_getenv_int(const char *name)
+{
+	const char *val = getenv(name);
+
+	if (val == NULL)
+		return 0;
+	return atoi(val);
+}
+
+static struct eth_driver mlx4_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 (mlx4_driver).
+ * @param[in] pci_dev
+ *   PCI device information.
+ *
+ * @return
+ *   0 on success, negative errno value on failure.
+ */
+static int
+mlx4_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 vf;
+	int idx;
+	int i;
+
+	(void)pci_drv;
+	assert(pci_drv == &mlx4_driver.pci_drv);
+	/* Get mlx4_dev[] index. */
+	idx = mlx4_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. */
+	mlx4_dev[idx].pci_addr = pci_dev->addr;
+	list = ibv_get_device_list(&i);
+	if (list == NULL) {
+		assert(errno);
+		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 (mlx4_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;
+		vf = (pci_dev->id.device_id ==
+		      PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
+		INFO("PCI information matches, using device \"%s\" (VF: %s)",
+		     list[i]->name, (vf ? "true" : "false"));
+		attr_ctx = ibv_open_device(list[i]);
+		err = errno;
+		break;
+	}
+	if (attr_ctx == NULL) {
+		if (err == 0)
+			err = ENODEV;
+		ibv_free_device_list(list);
+		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;
+#if defined(INLINE_RECV) || defined(RSS_SUPPORT)
+		struct ibv_exp_device_attr exp_device_attr;
+#endif
+		struct ether_addr mac;
+		union ibv_gid temp_gid;
+
+#ifdef RSS_SUPPORT
+		exp_device_attr.comp_mask =
+			(IBV_EXP_DEVICE_ATTR_EXP_CAP_FLAGS |
+			 IBV_EXP_DEVICE_ATTR_RSS_TBL_SZ);
+#endif /* RSS_SUPPORT */
+
+		DEBUG("using port %u (%08" PRIx32 ")", port, test);
+
+		ctx = ibv_open_device(ibv_dev);
+		if (ctx == NULL)
+			goto port_error;
+
+		/* Check port status. */
+		if ((err = ibv_query_port(ctx, port, &port_attr))) {
+			ERROR("port query failed: %s", strerror(err));
+			goto port_error;
+		}
+		if (port_attr.state != IBV_PORT_ACTIVE)
+			WARN("bad state for port %d: \"%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;
+		}
+
+		mlx4_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_attr = port_attr;
+		priv->port = port;
+		priv->pd = pd;
+		priv->mtu = ETHER_MTU;
+#ifdef RSS_SUPPORT
+		if (ibv_exp_query_device(ctx, &exp_device_attr)) {
+			INFO("experimental ibv_exp_query_device");
+			goto port_error;
+		}
+		if ((exp_device_attr.exp_device_cap_flags & IBV_EXP_DEVICE_QPG) &&
+		    (exp_device_attr.exp_device_cap_flags & IBV_EXP_DEVICE_UD_RSS) &&
+		    (exp_device_attr.comp_mask & IBV_EXP_DEVICE_ATTR_RSS_TBL_SZ) &&
+		    (exp_device_attr.max_rss_tbl_sz > 0)) {
+			priv->hw_qpg = 1;
+			priv->hw_rss = 1;
+			priv->max_rss_tbl_sz = exp_device_attr.max_rss_tbl_sz;
+		}
+		else {
+			priv->hw_qpg = 0;
+			priv->hw_rss = 0;
+			priv->max_rss_tbl_sz = 0;
+		}
+		priv->hw_tss = !!(exp_device_attr.exp_device_cap_flags &
+				  IBV_EXP_DEVICE_UD_TSS);
+		DEBUG("device flags: %s%s%s",
+		      (priv->hw_qpg ? "IBV_DEVICE_QPG " : ""),
+		      (priv->hw_tss ? "IBV_DEVICE_TSS " : ""),
+		      (priv->hw_rss ? "IBV_DEVICE_RSS " : ""));
+		if (priv->hw_rss)
+			DEBUG("maximum RSS indirection table size: %u",
+			      exp_device_attr.max_rss_tbl_sz);
+#endif /* RSS_SUPPORT */
+
+#ifdef INLINE_RECV
+		priv->inl_recv_size = mlx4_getenv_int("MLX4_INLINE_RECV_SIZE");
+
+		if (priv->inl_recv_size) {
+			exp_device_attr.comp_mask = IBV_EXP_DEVICE_ATTR_INLINE_RECV_SZ;
+			if (ibv_exp_query_device(ctx, &exp_device_attr)) {
+				INFO("Couldn't query device for inline-receive capabilities.");
+				priv->inl_recv_size = 0;
+			} else {
+				if ((unsigned int)exp_device_attr.inline_recv_sz < priv->inl_recv_size) {
+					INFO("Max inline-receive(%d) < Requested inline-receive(%u)",
+							exp_device_attr.inline_recv_sz, priv->inl_recv_size);
+					priv->inl_recv_size = exp_device_attr.inline_recv_sz;
+				}
+			}
+			INFO("Set inline receive size to %u", priv->inl_recv_size);
+		}
+#endif /* INLINE_RECV */
+
+#ifdef MLX4_COMPAT_VMWARE
+		if (mlx4_getenv_int("MLX4_COMPAT_VMWARE"))
+			priv->vmware = 1;
+#else /* MLX4_COMPAT_VMWARE */
+		(void)mlx4_getenv_int;
+#endif /* MLX4_COMPAT_VMWARE */
+		priv->vf = vf;
+		if (ibv_query_gid(ctx, port, 0, &temp_gid)) {
+			ERROR("ibv_query_gid() failure");
+			goto port_error;
+		}
+		/* Configure the first MAC address by default. */
+		mac_from_gid(&mac.addr_bytes, port, temp_gid.raw);
+		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 and broadcast addresses. */
+		claim_zero(priv_mac_addr_add(priv, 0,
+					     (const uint8_t (*)[ETHER_ADDR_LEN])
+					     mac.addr_bytes));
+		claim_zero(priv_mac_addr_add(priv, 1,
+					     &(const uint8_t [ETHER_ADDR_LEN])
+					     { "\xff\xff\xff\xff\xff\xff" }));
+#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 */
+#if RTE_VERSION >= RTE_VERSION_NUM(1, 7, 0, 0)
+		{
+			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);
+		}
+#else
+		eth_dev = rte_eth_dev_allocate();
+#endif
+		if (eth_dev == NULL) {
+			ERROR("can not allocate rte ethdev");
+			err = ENOMEM;
+			goto port_error;
+		}
+
+		eth_dev->data->dev_private = priv;
+		eth_dev->pci_dev = pci_dev;
+		eth_dev->driver = &mlx4_driver;
+		eth_dev->data->rx_mbuf_alloc_failed = 0;
+#if RTE_VERSION >= RTE_VERSION_NUM(1, 7, 0, 0)
+		eth_dev->data->mtu = ETHER_MTU;
+#else
+		eth_dev->data->max_frame_size = ETHER_MAX_LEN;
+#endif
+
+		priv->dev = eth_dev;
+		eth_dev->dev_ops = &mlx4_dev_ops;
+		eth_dev->data->mac_addrs = priv->mac;
+
+		/* 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);
+		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 (!mlx4_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 struct rte_pci_id mlx4_pci_id_map[] = {
+	{
+		.vendor_id = PCI_VENDOR_ID_MELLANOX,
+		.device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX3,
+		.subsystem_vendor_id = PCI_ANY_ID,
+		.subsystem_device_id = PCI_ANY_ID
+	},
+	{
+		.vendor_id = PCI_VENDOR_ID_MELLANOX,
+		.device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO,
+		.subsystem_vendor_id = PCI_ANY_ID,
+		.subsystem_device_id = PCI_ANY_ID
+	},
+	{
+		.vendor_id = PCI_VENDOR_ID_MELLANOX,
+		.device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX3VF,
+		.subsystem_vendor_id = PCI_ANY_ID,
+		.subsystem_device_id = PCI_ANY_ID
+	},
+	{
+		.vendor_id = 0
+	}
+};
+
+static struct eth_driver mlx4_driver = {
+	.pci_drv = {
+		.name = MLX4_DRIVER_NAME,
+		.id_table = mlx4_pci_id_map,
+		.devinit = mlx4_pci_devinit,
+	},
+	.dev_private_size = sizeof(struct priv)
+};
+
+/**
+ * Driver initialization routine.
+ */
+static int
+rte_mlx4_pmd_init(const char *name, const char *args)
+{
+	(void)name;
+	(void)args;
+	rte_eal_pci_register(&mlx4_driver.pci_drv);
+	return 0;
+}
+
+static struct rte_driver rte_mlx4_driver = {
+	.type = PMD_PDEV,
+	.name = MLX4_DRIVER_NAME,
+	.init = rte_mlx4_pmd_init,
+};
+
+PMD_REGISTER_DRIVER(rte_mlx4_driver)
diff --git a/lib/librte_pmd_mlx4/mlx4.h b/lib/librte_pmd_mlx4/mlx4.h
new file mode 100644
index 0000000..4743a64
--- /dev/null
+++ b/lib/librte_pmd_mlx4/mlx4.h
@@ -0,0 +1,166 @@ 
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright(c) 2012-2015 6WIND S.A.
+ *   Copyright(c) 2012 Mellanox.
+ *   All rights reserved.
+ *
+ *   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_MLX4_H_
+#define RTE_PMD_MLX4_H_
+
+#include <stddef.h>
+#include <stdint.h>
+#include <limits.h>
+
+/*
+ * Maximum number of simultaneous MAC addresses supported.
+ *
+ * According to ConnectX's Programmer Reference Manual:
+ *   The L2 Address Match is implemented by comparing a MAC/VLAN combination
+ *   of 128 MAC addresses and 127 VLAN values, comprising 128x127 possible
+ *   L2 addresses.
+ */
+#define MLX4_MAX_MAC_ADDRESSES 128
+
+/* Maximum number of simultaneous VLAN filters supported. See above. */
+#define MLX4_MAX_VLAN_IDS 127
+
+/* Maximum number of Scatter/Gather Elements per Work Request. */
+#ifndef MLX4_PMD_SGE_WR_N
+#define MLX4_PMD_SGE_WR_N 4
+#endif
+
+/* Maximum size for inline data. */
+#ifndef MLX4_PMD_MAX_INLINE
+#define MLX4_PMD_MAX_INLINE 0
+#endif
+
+/*
+ * 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 MLX4_PMD_TX_MP_CACHE
+#define MLX4_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 MLX4_PMD_SOFT_COUNTERS
+#define MLX4_PMD_SOFT_COUNTERS 1
+#endif
+
+/*
+ * If defined, enable VMware compatibility code. It also requires the
+ * environment variable MLX4_COMPAT_VMWARE set to a nonzero value at runtime.
+ */
+#ifndef MLX4_COMPAT_VMWARE
+#define MLX4_COMPAT_VMWARE 1
+#endif
+
+enum {
+	PCI_VENDOR_ID_MELLANOX = 0x15b3,
+};
+
+enum {
+	PCI_DEVICE_ID_MELLANOX_CONNECTX3 = 0x1003,
+	PCI_DEVICE_ID_MELLANOX_CONNECTX3VF = 0x1004,
+	PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO = 0x1007,
+};
+
+#define MLX4_DRIVER_NAME "librte_pmd_mlx4"
+
+/* 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))))))
+
+/* Number of elements in array. */
+#define elemof(a) (sizeof(a) / sizeof((a)[0]))
+
+/* Cast pointer p to structure member m to its parent structure of type t. */
+#define containerof(p, t, m) ((t *)((uint8_t *)(p) - offsetof(t, m)))
+
+/* Branch prediction helpers. */
+#ifndef likely
+#define likely(c) __builtin_expect(!!(c), 1)
+#endif
+#ifndef unlikely
+#define unlikely(c) __builtin_expect(!!(c), 0)
+#endif
+
+/* Debugging */
+#ifndef NDEBUG
+#include <stdio.h>
+#define DEBUG__(m, ...)						\
+	(fprintf(stderr, "%s:%d: %s(): " m "%c",		\
+		 __FILE__, __LINE__, __func__, __VA_ARGS__),	\
+	 fflush(stderr),					\
+	 (void)0)
+/*
+ * Save/restore errno around DEBUG__().
+ * XXX somewhat undefined behavior, but works.
+ */
+#define DEBUG_(...)				\
+	(errno = ((int []){			\
+		*(volatile int *)&errno,	\
+		(DEBUG__(__VA_ARGS__), 0)	\
+	})[0])
+#define DEBUG(...) DEBUG_(__VA_ARGS__, '\n')
+#define claim_zero(...) assert((__VA_ARGS__) == 0)
+#define claim_nonzero(...) assert((__VA_ARGS__) != 0)
+#define claim_positive(...) assert((__VA_ARGS__) >= 0)
+#else /* NDEBUG */
+/* No-ops. */
+#define DEBUG(...) (void)0
+#define claim_zero(...) (__VA_ARGS__)
+#define claim_nonzero(...) (__VA_ARGS__)
+#define claim_positive(...) (__VA_ARGS__)
+#endif /* NDEBUG */
+
+#endif /* RTE_PMD_MLX4_H_ */
diff --git a/mk/rte.app.mk b/mk/rte.app.mk
index 4294d9a..07a1ebb 100644
--- a/mk/rte.app.mk
+++ b/mk/rte.app.mk
@@ -135,6 +135,10 @@  ifeq ($(CONFIG_RTE_LIBRTE_VHOST),y)
 LDLIBS += -lfuse
 endif
 
+ifeq ($(CONFIG_RTE_LIBRTE_MLX4_PMD),y)
+LDLIBS += -libverbs
+endif
+
 LDLIBS += --start-group
 
 ifeq ($(CONFIG_RTE_BUILD_COMBINE_LIBS),n)
@@ -219,6 +223,10 @@  ifeq ($(CONFIG_RTE_LIBRTE_E1000_PMD),y)
 LDLIBS += -lrte_pmd_e1000
 endif
 
+ifeq ($(CONFIG_RTE_LIBRTE_MLX4_PMD),y)
+LDLIBS += -lrte_pmd_mlx4
+endif
+
 ifeq ($(CONFIG_RTE_LIBRTE_PMD_RING),y)
 LDLIBS += -lrte_pmd_ring
 endif