mlx4: move to drivers/net/
authorBruce Richardson <bruce.richardson@intel.com>
Fri, 15 May 2015 15:56:54 +0000 (16:56 +0100)
committerThomas Monjalon <thomas.monjalon@6wind.com>
Fri, 22 May 2015 14:06:22 +0000 (16:06 +0200)
move mlx4 PMD to drivers/net directory

Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Acked-by: John McNamara <john.mcnamara@intel.com>
Acked-by: Thomas Monjalon <thomas.monjalon@6wind.com>
12 files changed:
MAINTAINERS
doc/guides/prog_guide/source_org.rst
drivers/net/Makefile
drivers/net/mlx4/Makefile [new file with mode: 0644]
drivers/net/mlx4/mlx4.c [new file with mode: 0644]
drivers/net/mlx4/mlx4.h [new file with mode: 0644]
drivers/net/mlx4/rte_pmd_mlx4_version.map [new file with mode: 0644]
lib/Makefile
lib/librte_pmd_mlx4/Makefile [deleted file]
lib/librte_pmd_mlx4/mlx4.c [deleted file]
lib/librte_pmd_mlx4/mlx4.h [deleted file]
lib/librte_pmd_mlx4/rte_pmd_mlx4_version.map [deleted file]

index 767eab5..04f61e4 100644 (file)
@@ -236,7 +236,7 @@ F: drivers/net/fm10k/
 
 Mellanox mlx4
 M: Adrien Mazarguil <adrien.mazarguil@6wind.com>
-F: lib/librte_pmd_mlx4/
+F: drivers/net/mlx4/
 F: doc/guides/nics/mlx4.rst
 
 RedHat virtio
index 8e3f4ee..0ac517b 100644 (file)
@@ -79,7 +79,6 @@ The lib directory contains::
     +-- librte_mempool      # memory pool manager (fixedsized objects)
     +-- librte_meter        # QoS metering library
     +-- librte_net          # various IP-related headers
-    +-- librte_pmd_mlx4     # Mellanox ConnectX-3 poll mode driver
     +-- librte_pmd_pcap     # PCAP poll mode driver
     +-- librte_pmd_ring     # ring poll mode driver
     +-- librte_pmd_virtio   # virtio poll mode driver
@@ -108,6 +107,7 @@ The drivers directory has a net subdirectory which contains::
     +-- fm10k              # Host interface PMD driver for FM10000 Series
     +-- i40e               # 40GbE poll mode driver
     +-- ixgbe              # 10GbE poll mode driver
+    +-- mlx4               # Mellanox ConnectX-3 poll mode driver
 
 Applications
 ------------
index 156a16f..45e06a6 100644 (file)
@@ -38,6 +38,7 @@ DIRS-$(CONFIG_RTE_LIBRTE_ENIC_PMD) += enic
 DIRS-$(CONFIG_RTE_LIBRTE_FM10K_PMD) += fm10k
 DIRS-$(CONFIG_RTE_LIBRTE_I40E_PMD) += i40e
 DIRS-$(CONFIG_RTE_LIBRTE_IXGBE_PMD) += ixgbe
+DIRS-$(CONFIG_RTE_LIBRTE_MLX4_PMD) += mlx4
 
 include $(RTE_SDK)/mk/rte.sharelib.mk
 include $(RTE_SDK)/mk/rte.subdir.mk
diff --git a/drivers/net/mlx4/Makefile b/drivers/net/mlx4/Makefile
new file mode 100644 (file)
index 0000000..97b364a
--- /dev/null
@@ -0,0 +1,121 @@
+#   BSD LICENSE
+#
+#   Copyright 2012-2015 6WIND S.A.
+#   Copyright 2012 Mellanox.
+#
+#   Redistribution and use in source and binary forms, with or without
+#   modification, are permitted provided that the following conditions
+#   are met:
+#
+#     * Redistributions of source code must retain the above copyright
+#       notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above copyright
+#       notice, this list of conditions and the following disclaimer in
+#       the documentation and/or other materials provided with the
+#       distribution.
+#     * Neither the name of 6WIND S.A. nor the names of its
+#       contributors may be used to endorse or promote products derived
+#       from this software without specific prior written permission.
+#
+#   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+#   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+#   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+#   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+#   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+#   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+#   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+#   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+#   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+#   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+#   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+include $(RTE_SDK)/mk/rte.vars.mk
+
+# Library name.
+LIB = librte_pmd_mlx4.a
+
+# Sources.
+SRCS-$(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 += $(WERROR_FLAGS)
+
+# A few warnings cannot be avoided in external headers.
+CFLAGS += -Wno-error=cast-qual
+
+EXPORT_MAP := rte_pmd_mlx4_version.map
+LIBABIVER := 1
+
+# DEBUG which is usually provided on the command-line may enable
+# CONFIG_RTE_LIBRTE_MLX4_DEBUG.
+ifeq ($(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
+
+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
+
+ifndef V
+AUTOCONF_OUTPUT := >/dev/null
+endif
+
+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 $(AUTOCONF_OUTPUT)
+       $Q sh -- '$<' '$@' \
+               INLINE_RECV \
+               infiniband/verbs.h \
+               enum IBV_EXP_DEVICE_ATTR_INLINE_RECV_SZ $(AUTOCONF_OUTPUT)
+       $Q sh -- '$<' '$@' \
+               SEND_RAW_WR_SUPPORT \
+               infiniband/verbs.h \
+               type 'struct ibv_send_wr_raw' $(AUTOCONF_OUTPUT)
+
+mlx4.o: mlx4_autoconf.h
+
+clean_mlx4: FORCE
+       $Q rm -f -- mlx4_autoconf.h
+
+clean: clean_mlx4
diff --git a/drivers/net/mlx4/mlx4.c b/drivers/net/mlx4/mlx4.c
new file mode 100644 (file)
index 0000000..f915bc1
--- /dev/null
@@ -0,0 +1,4686 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2012-2015 6WIND S.A.
+ *   Copyright 2012 Mellanox.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+/*
+ * 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>
+
+/* Verbs header. */
+/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <infiniband/verbs.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-pedantic"
+#endif
+
+/* DPDK headers don't like -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-pedantic"
+#endif
+#include <rte_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
+
+/* 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. */
+       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));
+
+       rte_free(elts_linear);
+       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));
+
+       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 occurred, 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
+       };
+       ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod,
+                               (IBV_EXP_QP_STATE | IBV_EXP_QP_PORT));
+       if (ret) {
+               ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
+                     (void *)dev, strerror(ret));
+               goto error;
+       }
+       ret = txq_alloc_elts(&tmpl, desc);
+       if (ret) {
+               ERROR("%p: TXQ allocation failed: %s",
+                     (void *)dev, strerror(ret));
+               goto error;
+       }
+       attr.mod = (struct ibv_exp_qp_attr){
+               .qp_state = IBV_QPS_RTR
+       };
+       ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
+       if (ret) {
+               ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
+                     (void *)dev, strerror(ret));
+               goto error;
+       }
+       attr.mod.qp_state = IBV_QPS_RTS;
+       ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
+       if (ret) {
+               ERROR("%p: QP state to IBV_QPS_RTS failed: %s",
+                     (void *)dev, strerror(ret));
+               goto error;
+       }
+       /* 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)
+{
+#ifndef NDEBUG
+       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);
+       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 ? htons(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 = htons(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);
+       /* Create related flow. */
+       errno = 0;
+       flow = ibv_exp_create_flow(rxq->qp, attr);
+       if (flow == NULL) {
+               int err = errno;
+
+               /* 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) ||
+                   (rxq_promiscuous_enable(rxq) == 0)) {
+                       if (rxq->promisc_flow != NULL)
+                               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
+       };
+
+       DEBUG("%p: enabling allmulticast mode", (void *)rxq);
+       if (rxq->allmulti_flow != NULL)
+               return EBUSY;
+       errno = 0;
+       flow = ibv_exp_create_flow(rxq->qp, &attr);
+       if (flow == NULL) {
+               /* It's not clear whether errno is always set in this case. */
+               ERROR("%p: flow configuration failed, errno=%d: %s",
+                     (void *)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)
+{
+       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
+       };
+
+       if (rxq->priv->vf)
+               return 0;
+       DEBUG("%p: enabling promiscuous mode", (void *)rxq);
+       if (rxq->promisc_flow != NULL)
+               return EBUSY;
+       errno = 0;
+       flow = ibv_exp_create_flow(rxq->qp, &attr);
+       if (flow == NULL) {
+               /* It's not clear whether errno is always set in this case. */
+               ERROR("%p: flow configuration failed, errno=%d: %s",
+                     (void *)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)
+{
+       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 };
+       err = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
+       if (err) {
+               ERROR("%p: cannot reset QP: %s", (void *)dev, strerror(err));
+               assert(err > 0);
+               return err;
+       }
+       err = ibv_resize_cq(tmpl.cq, desc_n);
+       if (err) {
+               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
+       };
+       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));
+       if (err) {
+               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. */
+       err = ibv_post_recv(tmpl.qp,
+                           (tmpl.sp ?
+                            &(*tmpl.elts.sp)[0].wr :
+                            &(*tmpl.elts.no_sp)[0].wr),
+                           &bad_wr);
+       if (err) {
+               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
+       };
+       err = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
+       if (err)
+               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
+       };
+       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));
+       if (ret) {
+               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. */
+               ret = rxq_mac_addrs_add(&tmpl);
+               if (ret) {
+                       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;
+       }
+       ret = ibv_post_recv(tmpl.qp,
+                           (tmpl.sp ?
+                            &(*tmpl.elts.sp)[0].wr :
+                            &(*tmpl.elts.no_sp)[0].wr),
+                           &bad_wr);
+       if (ret) {
+               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
+       };
+       ret = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
+       if (ret) {
+               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;
+               ret = rxq_mac_addrs_add(rxq);
+               if (!ret && priv->promisc)
+                       ret = rxq_promiscuous_enable(rxq);
+               if (!ret && priv->allmulti)
+                       ret = rxq_allmulticast_enable(rxq);
+               if (!ret)
+                       continue;
+               WARN("%p: QP flow attachment failed: %s",
+                    (void *)dev, strerror(ret));
+               /* Rollback. */
+               while (i != 0) {
+                       rxq = (*priv->rxqs)[--i];
+                       if (rxq != 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 = {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;
+       errno = ibv_query_port(priv->ctx, priv->port, &port_attr);
+       if (errno) {
+               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;
+               ret = rxq_rehash(dev, rxq);
+               if (ret) {
+                       /* 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 const 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);
+               if (errno == ENOSYS) {
+                       WARN("cannot list devices, is ib_uverbs loaded?");
+                       return 0;
+               }
+               return -errno;
+       }
+       assert(i >= 0);
+       /*
+        * For each listed device, check related sysfs entry against
+        * the provided PCI ID.
+        */
+       while (i != 0) {
+               struct rte_pci_addr pci_addr;
+
+               --i;
+               DEBUG("checking device \"%s\"", list[i]->name);
+               if (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) {
+               ibv_free_device_list(list);
+               switch (err) {
+               case 0:
+                       WARN("cannot access device, is mlx4_ib loaded?");
+                       return 0;
+               case EINVAL:
+                       WARN("cannot use device, are drivers up to date?");
+                       return 0;
+               }
+               assert(err > 0);
+               return -err;
+       }
+       ibv_dev = list[i];
+
+       DEBUG("device opened");
+       if (ibv_query_device(attr_ctx, &device_attr))
+               goto error;
+       INFO("%u port(s) detected", device_attr.phys_port_cnt);
+
+       for (i = 0; i < device_attr.phys_port_cnt; i++) {
+               uint32_t port = i + 1; /* ports are indexed from one */
+               uint32_t test = (1 << i);
+               struct ibv_context *ctx = NULL;
+               struct ibv_port_attr port_attr;
+               struct ibv_pd *pd = NULL;
+               struct priv *priv = NULL;
+               struct rte_eth_dev *eth_dev;
+#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. */
+               err = ibv_query_port(ctx, port, &port_attr);
+               if (err) {
+                       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)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 */
+
+               (void)mlx4_getenv_int;
+               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 */
+               {
+                       char name[RTE_ETH_NAME_MAX_LEN];
+
+                       snprintf(name, sizeof(name), "%s port %u",
+                                ibv_get_device_name(ibv_dev), port);
+                       eth_dev = rte_eth_dev_allocate(name, RTE_ETH_DEV_PCI);
+               }
+               if (eth_dev == NULL) {
+                       ERROR("can not allocate rte ethdev");
+                       err = ENOMEM;
+                       goto port_error;
+               }
+
+               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;
+               eth_dev->data->mtu = ETHER_MTU;
+
+               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:
+               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 const 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/drivers/net/mlx4/mlx4.h b/drivers/net/mlx4/mlx4.h
new file mode 100644 (file)
index 0000000..433aa3b
--- /dev/null
@@ -0,0 +1,157 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2012-2015 6WIND S.A.
+ *   Copyright 2012 Mellanox.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#ifndef RTE_PMD_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
+
+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/drivers/net/mlx4/rte_pmd_mlx4_version.map b/drivers/net/mlx4/rte_pmd_mlx4_version.map
new file mode 100644 (file)
index 0000000..ef35398
--- /dev/null
@@ -0,0 +1,4 @@
+DPDK_2.0 {
+
+       local: *;
+};
index ffd7870..80978cd 100644 (file)
@@ -41,7 +41,6 @@ DIRS-$(CONFIG_RTE_LIBRTE_TIMER) += librte_timer
 DIRS-$(CONFIG_RTE_LIBRTE_CFGFILE) += librte_cfgfile
 DIRS-$(CONFIG_RTE_LIBRTE_CMDLINE) += librte_cmdline
 DIRS-$(CONFIG_RTE_LIBRTE_ETHER) += librte_ether
-DIRS-$(CONFIG_RTE_LIBRTE_MLX4_PMD) += librte_pmd_mlx4
 DIRS-$(CONFIG_RTE_LIBRTE_PMD_RING) += librte_pmd_ring
 DIRS-$(CONFIG_RTE_LIBRTE_PMD_PCAP) += librte_pmd_pcap
 DIRS-$(CONFIG_RTE_LIBRTE_VIRTIO_PMD) += librte_pmd_virtio
diff --git a/lib/librte_pmd_mlx4/Makefile b/lib/librte_pmd_mlx4/Makefile
deleted file mode 100644 (file)
index 97b364a..0000000
+++ /dev/null
@@ -1,121 +0,0 @@
-#   BSD LICENSE
-#
-#   Copyright 2012-2015 6WIND S.A.
-#   Copyright 2012 Mellanox.
-#
-#   Redistribution and use in source and binary forms, with or without
-#   modification, are permitted provided that the following conditions
-#   are met:
-#
-#     * Redistributions of source code must retain the above copyright
-#       notice, this list of conditions and the following disclaimer.
-#     * Redistributions in binary form must reproduce the above copyright
-#       notice, this list of conditions and the following disclaimer in
-#       the documentation and/or other materials provided with the
-#       distribution.
-#     * Neither the name of 6WIND S.A. nor the names of its
-#       contributors may be used to endorse or promote products derived
-#       from this software without specific prior written permission.
-#
-#   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-#   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-#   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-#   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-#   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-#   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-#   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-#   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-#   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-#   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-#   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-include $(RTE_SDK)/mk/rte.vars.mk
-
-# Library name.
-LIB = librte_pmd_mlx4.a
-
-# Sources.
-SRCS-$(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 += $(WERROR_FLAGS)
-
-# A few warnings cannot be avoided in external headers.
-CFLAGS += -Wno-error=cast-qual
-
-EXPORT_MAP := rte_pmd_mlx4_version.map
-LIBABIVER := 1
-
-# DEBUG which is usually provided on the command-line may enable
-# CONFIG_RTE_LIBRTE_MLX4_DEBUG.
-ifeq ($(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
-
-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
-
-ifndef V
-AUTOCONF_OUTPUT := >/dev/null
-endif
-
-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 $(AUTOCONF_OUTPUT)
-       $Q sh -- '$<' '$@' \
-               INLINE_RECV \
-               infiniband/verbs.h \
-               enum IBV_EXP_DEVICE_ATTR_INLINE_RECV_SZ $(AUTOCONF_OUTPUT)
-       $Q sh -- '$<' '$@' \
-               SEND_RAW_WR_SUPPORT \
-               infiniband/verbs.h \
-               type 'struct ibv_send_wr_raw' $(AUTOCONF_OUTPUT)
-
-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
deleted file mode 100644 (file)
index f915bc1..0000000
+++ /dev/null
@@ -1,4686 +0,0 @@
-/*-
- *   BSD LICENSE
- *
- *   Copyright 2012-2015 6WIND S.A.
- *   Copyright 2012 Mellanox.
- *
- *   Redistribution and use in source and binary forms, with or without
- *   modification, are permitted provided that the following conditions
- *   are met:
- *
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in
- *       the documentation and/or other materials provided with the
- *       distribution.
- *     * Neither the name of 6WIND S.A. nor the names of its
- *       contributors may be used to endorse or promote products derived
- *       from this software without specific prior written permission.
- *
- *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
- *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
- *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
- *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
- *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
- *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
- *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
- *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
- *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
- *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- */
-
-/*
- * 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>
-
-/* Verbs header. */
-/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
-#ifdef PEDANTIC
-#pragma GCC diagnostic ignored "-pedantic"
-#endif
-#include <infiniband/verbs.h>
-#ifdef PEDANTIC
-#pragma GCC diagnostic error "-pedantic"
-#endif
-
-/* DPDK headers don't like -pedantic. */
-#ifdef PEDANTIC
-#pragma GCC diagnostic ignored "-pedantic"
-#endif
-#include <rte_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
-
-/* 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. */
-       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));
-
-       rte_free(elts_linear);
-       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));
-
-       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 occurred, 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
-       };
-       ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod,
-                               (IBV_EXP_QP_STATE | IBV_EXP_QP_PORT));
-       if (ret) {
-               ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
-                     (void *)dev, strerror(ret));
-               goto error;
-       }
-       ret = txq_alloc_elts(&tmpl, desc);
-       if (ret) {
-               ERROR("%p: TXQ allocation failed: %s",
-                     (void *)dev, strerror(ret));
-               goto error;
-       }
-       attr.mod = (struct ibv_exp_qp_attr){
-               .qp_state = IBV_QPS_RTR
-       };
-       ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
-       if (ret) {
-               ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
-                     (void *)dev, strerror(ret));
-               goto error;
-       }
-       attr.mod.qp_state = IBV_QPS_RTS;
-       ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
-       if (ret) {
-               ERROR("%p: QP state to IBV_QPS_RTS failed: %s",
-                     (void *)dev, strerror(ret));
-               goto error;
-       }
-       /* 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)
-{
-#ifndef NDEBUG
-       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);
-       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 ? htons(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 = htons(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);
-       /* Create related flow. */
-       errno = 0;
-       flow = ibv_exp_create_flow(rxq->qp, attr);
-       if (flow == NULL) {
-               int err = errno;
-
-               /* 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) ||
-                   (rxq_promiscuous_enable(rxq) == 0)) {
-                       if (rxq->promisc_flow != NULL)
-                               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
-       };
-
-       DEBUG("%p: enabling allmulticast mode", (void *)rxq);
-       if (rxq->allmulti_flow != NULL)
-               return EBUSY;
-       errno = 0;
-       flow = ibv_exp_create_flow(rxq->qp, &attr);
-       if (flow == NULL) {
-               /* It's not clear whether errno is always set in this case. */
-               ERROR("%p: flow configuration failed, errno=%d: %s",
-                     (void *)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)
-{
-       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
-       };
-
-       if (rxq->priv->vf)
-               return 0;
-       DEBUG("%p: enabling promiscuous mode", (void *)rxq);
-       if (rxq->promisc_flow != NULL)
-               return EBUSY;
-       errno = 0;
-       flow = ibv_exp_create_flow(rxq->qp, &attr);
-       if (flow == NULL) {
-               /* It's not clear whether errno is always set in this case. */
-               ERROR("%p: flow configuration failed, errno=%d: %s",
-                     (void *)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)
-{
-       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 };
-       err = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
-       if (err) {
-               ERROR("%p: cannot reset QP: %s", (void *)dev, strerror(err));
-               assert(err > 0);
-               return err;
-       }
-       err = ibv_resize_cq(tmpl.cq, desc_n);
-       if (err) {
-               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
-       };
-       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));
-       if (err) {
-               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. */
-       err = ibv_post_recv(tmpl.qp,
-                           (tmpl.sp ?
-                            &(*tmpl.elts.sp)[0].wr :
-                            &(*tmpl.elts.no_sp)[0].wr),
-                           &bad_wr);
-       if (err) {
-               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
-       };
-       err = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
-       if (err)
-               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
-       };
-       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));
-       if (ret) {
-               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. */
-               ret = rxq_mac_addrs_add(&tmpl);
-               if (ret) {
-                       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;
-       }
-       ret = ibv_post_recv(tmpl.qp,
-                           (tmpl.sp ?
-                            &(*tmpl.elts.sp)[0].wr :
-                            &(*tmpl.elts.no_sp)[0].wr),
-                           &bad_wr);
-       if (ret) {
-               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
-       };
-       ret = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
-       if (ret) {
-               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;
-               ret = rxq_mac_addrs_add(rxq);
-               if (!ret && priv->promisc)
-                       ret = rxq_promiscuous_enable(rxq);
-               if (!ret && priv->allmulti)
-                       ret = rxq_allmulticast_enable(rxq);
-               if (!ret)
-                       continue;
-               WARN("%p: QP flow attachment failed: %s",
-                    (void *)dev, strerror(ret));
-               /* Rollback. */
-               while (i != 0) {
-                       rxq = (*priv->rxqs)[--i];
-                       if (rxq != 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 = {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;
-       errno = ibv_query_port(priv->ctx, priv->port, &port_attr);
-       if (errno) {
-               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;
-               ret = rxq_rehash(dev, rxq);
-               if (ret) {
-                       /* 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 const 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);
-               if (errno == ENOSYS) {
-                       WARN("cannot list devices, is ib_uverbs loaded?");
-                       return 0;
-               }
-               return -errno;
-       }
-       assert(i >= 0);
-       /*
-        * For each listed device, check related sysfs entry against
-        * the provided PCI ID.
-        */
-       while (i != 0) {
-               struct rte_pci_addr pci_addr;
-
-               --i;
-               DEBUG("checking device \"%s\"", list[i]->name);
-               if (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) {
-               ibv_free_device_list(list);
-               switch (err) {
-               case 0:
-                       WARN("cannot access device, is mlx4_ib loaded?");
-                       return 0;
-               case EINVAL:
-                       WARN("cannot use device, are drivers up to date?");
-                       return 0;
-               }
-               assert(err > 0);
-               return -err;
-       }
-       ibv_dev = list[i];
-
-       DEBUG("device opened");
-       if (ibv_query_device(attr_ctx, &device_attr))
-               goto error;
-       INFO("%u port(s) detected", device_attr.phys_port_cnt);
-
-       for (i = 0; i < device_attr.phys_port_cnt; i++) {
-               uint32_t port = i + 1; /* ports are indexed from one */
-               uint32_t test = (1 << i);
-               struct ibv_context *ctx = NULL;
-               struct ibv_port_attr port_attr;
-               struct ibv_pd *pd = NULL;
-               struct priv *priv = NULL;
-               struct rte_eth_dev *eth_dev;
-#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. */
-               err = ibv_query_port(ctx, port, &port_attr);
-               if (err) {
-                       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)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 */
-
-               (void)mlx4_getenv_int;
-               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 */
-               {
-                       char name[RTE_ETH_NAME_MAX_LEN];
-
-                       snprintf(name, sizeof(name), "%s port %u",
-                                ibv_get_device_name(ibv_dev), port);
-                       eth_dev = rte_eth_dev_allocate(name, RTE_ETH_DEV_PCI);
-               }
-               if (eth_dev == NULL) {
-                       ERROR("can not allocate rte ethdev");
-                       err = ENOMEM;
-                       goto port_error;
-               }
-
-               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;
-               eth_dev->data->mtu = ETHER_MTU;
-
-               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:
-               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 const 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
deleted file mode 100644 (file)
index 433aa3b..0000000
+++ /dev/null
@@ -1,157 +0,0 @@
-/*-
- *   BSD LICENSE
- *
- *   Copyright 2012-2015 6WIND S.A.
- *   Copyright 2012 Mellanox.
- *
- *   Redistribution and use in source and binary forms, with or without
- *   modification, are permitted provided that the following conditions
- *   are met:
- *
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in
- *       the documentation and/or other materials provided with the
- *       distribution.
- *     * Neither the name of 6WIND S.A. nor the names of its
- *       contributors may be used to endorse or promote products derived
- *       from this software without specific prior written permission.
- *
- *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
- *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
- *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
- *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
- *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
- *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
- *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
- *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
- *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
- *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- */
-
-#ifndef RTE_PMD_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
-
-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/lib/librte_pmd_mlx4/rte_pmd_mlx4_version.map b/lib/librte_pmd_mlx4/rte_pmd_mlx4_version.map
deleted file mode 100644 (file)
index ef35398..0000000
+++ /dev/null
@@ -1,4 +0,0 @@
-DPDK_2.0 {
-
-       local: *;
-};