source: lib/format_dpdk.c @ 6b98325

/*
 * This file is part of libtrace
 *
 * Copyright (c) 2007,2008,2009,2010 The University of Waikato, Hamilton,
 * New Zealand.
 *
 * Author: Richard Sanger
 *
 * All rights reserved.
 *
 * This code has been developed by the University of Waikato WAND
 * research group. For further information please see http://www.wand.net.nz/
 *
 * libtrace is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 2 of the License, or
 * (at your option) any later version.
 *
 * libtrace is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with libtrace; if not, write to the Free Software
 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
 *
 * $Id: format_dpdk.c 1805 2013-03-08 02:01:35Z salcock $
 *
 */

/* This format module deals with using the Intel Data Plane Development
 * Kit capture format.
 *
 * Intel Data Plane Development Kit is a LIVE capture format.
 *
 * This format also supports writing which will write packets out to the
 * network as a form of packet replay. This should not be confused with the
 * RT protocol which is intended to transfer captured packet records between
 * RT-speaking programs.
 */

#define _GNU_SOURCE

#include "config.h"
#include "libtrace.h"
#include "libtrace_int.h"
#include "format_helper.h"
#include "libtrace_arphrd.h"
#include "hash_toeplitz.h"

#ifdef HAVE_INTTYPES_H
#  include <inttypes.h>
#else
# error "Can't find inttypes.h"
#endif

#include <stdlib.h>
#include <assert.h>
#include <unistd.h>
#include <endian.h>
#include <string.h>

#if HAVE_LIBNUMA
#include <numa.h>
#endif

/* We can deal with any minor differences by checking the RTE VERSION
 * Typically DPDK backports some fixes (typically for building against
 * newer kernels) to the older version of DPDK.
 *
 * These get released with the rX suffix. The following macros where added
 * in these new releases.
 *
 * Below this is a log of version that required changes to the libtrace
 * code (that we still attempt to support).
 *
 * DPDK v1.7.1 is recommended.
 * However 1.5 to 1.8 are likely supported.
 */
#include <rte_eal.h>
#include <rte_version.h>
#ifndef RTE_VERSION_NUM
#       define RTE_VERSION_NUM(a,b,c,d) ((a) << 24 | (b) << 16 | (c) << 8 | (d))
#endif
#ifndef RTE_VER_PATCH_RELEASE
#       define RTE_VER_PATCH_RELEASE 0
#endif
#ifndef RTE_VERSION
#       define RTE_VERSION RTE_VERSION_NUM(RTE_VER_MAJOR,RTE_VER_MINOR, \
        RTE_VER_PATCH_LEVEL, RTE_VER_PATCH_RELEASE)
#endif

/* 1.6.0r2 :
 *      rte_eal_pci_set_blacklist() is removed
 *      device_list is renamed to pci_device_list
 *      In the 1.7.0 release rte_eal_pci_probe is called by rte_eal_init
 *      as such we do apply the whitelist before rte_eal_init.
 *      This also works correctly with DPDK 1.6.0r2.
 *
 * Replaced by:
 *      rte_devargs (we can simply whitelist)
 */
#if RTE_VERSION <= RTE_VERSION_NUM(1, 6, 0, 1)
#       define DPDK_USE_BLACKLIST 1
#else
#       define DPDK_USE_BLACKLIST 0
#endif

/*
 * 1.7.0 :
 *      rte_pmd_init_all is removed
 *
 * Replaced by:
 *      Nothing, no longer needed
 */
#if RTE_VERSION < RTE_VERSION_NUM(1, 7, 0, 0)
#       define DPDK_USE_PMD_INIT 1
#else
#       define DPDK_USE_PMD_INIT 0
#endif

/* 1.7.0-rc3 :
 *
 * Since 1.7.0-rc3 rte_eal_pci_probe is called as part of rte_eal_init.
 * Somewhere between 1.7 and 1.8 calling it twice broke so we should not call
 * it twice.
 */
#if RTE_VERSION < RTE_VERSION_NUM(1, 7, 0, 3)
#       define DPDK_USE_PCI_PROBE 1
#else
#       define DPDK_USE_PCI_PROBE 0
#endif

/* 1.8.0-rc1 :
 * LOG LEVEL is a command line option which overrides what
 * we previously set it to.
 */
#if RTE_VERSION >= RTE_VERSION_NUM(1, 8, 0, 1)
#       define DPDK_USE_LOG_LEVEL 1
#else
#       define DPDK_USE_LOG_LEVEL 0
#endif

#include <rte_per_lcore.h>
#include <rte_debug.h>
#include <rte_errno.h>
#include <rte_common.h>
#include <rte_log.h>
#include <rte_memcpy.h>
#include <rte_prefetch.h>
#include <rte_branch_prediction.h>
#include <rte_pci.h>
#include <rte_ether.h>
#include <rte_ethdev.h>
#include <rte_ring.h>
#include <rte_mempool.h>
#include <rte_mbuf.h>
#include <rte_launch.h>
#include <rte_lcore.h>
#include <rte_per_lcore.h>
#include <rte_cycles.h>
#include <pthread.h>

/* The default size of memory buffers to use - This is the max size of standard
 * ethernet packet less the size of the MAC CHECKSUM */
#define RX_MBUF_SIZE 1514

/* The minimum number of memory buffers per queue tx or rx. Search for
 * _MIN_RING_DESC in DPDK. The largest minimum is 64 for 10GBit cards.
 */
#define MIN_NB_BUF 64

/* Number of receive memory buffers to use
 * By default this is limited by driver to 4k and must be a multiple of 128.
 * A modification can be made to the driver to remove this limit.
 * This can be increased in the driver and here.
 * Should be at least MIN_NB_BUF.
 */
#define NB_RX_MBUF 4096

/* Number of send memory buffers to use.
 * Same limits apply as those to NB_TX_MBUF.
 */
#define NB_TX_MBUF 1024

/* The size of the PCI blacklist needs to be big enough to contain
 * every PCI device address (listed by lspci every bus:device.function tuple).
 */
#define BLACK_LIST_SIZE 50

/* The maximum number of characters the mempool name can be */
#define MEMPOOL_NAME_LEN 20

/* For single threaded libtrace we read packets as a batch/burst
 * this is the maximum size of said burst */
#define BURST_SIZE 50

#define MBUF(x) ((struct rte_mbuf *) x)
/* Get the original placement of the packet data */
#define MBUF_PKTDATA(x) ((char *) x + sizeof(struct rte_mbuf) + RTE_PKTMBUF_HEADROOM)
#define FORMAT(x) ((struct dpdk_format_data_t*)(x->format_data))
#define PERPKT_FORMAT(x) ((struct dpdk_per_lcore_t*)(x->format_data))

#define FORMAT_DATA_HEAD(x) FORMAT(x)->per_stream->head
#define FORMAT_DATA_FIRST(x) ((dpdk_per_stream_t *)FORMAT_DATA_HEAD(x)->data)

#define TV_TO_NS(tv) ((uint64_t) tv.tv_sec*1000000000ull + \
                        (uint64_t) tv.tv_usec*1000ull)
#define TS_TO_NS(ts) ((uint64_t) ts.tv_sec*1000000000ull + \
                        (uint64_t) ts.tv_nsec)

#if RTE_PKTMBUF_HEADROOM != 128
#warning "RTE_PKT_MBUF_HEADROOM is not set to the default value of 128 - " \
         "any libtrace instance processing these packet must be have the" \
         "same RTE_PKTMBUF_HEADROOM set"
#endif

/* ~~~~~~~~~~~~~~~~~~~~~~ Advance settings ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 * THESE MAY REQUIRE MODIFICATIONS TO INTEL DPDK
 *
 * Make sure you understand what these are doing before enabling them.
 * They might make traces incompatable with other builds etc.
 *
 * These are also included to show how to do somethings which aren't
 * obvious in the DPDK documentation.
 */

/* Print verbose messages to stderr */
#define DEBUG 0

/* Use clock_gettime() for nanosecond resolution rather than gettimeofday()
 * only turn on if you know clock_gettime is a vsyscall on your system
 * overwise could be a large overhead. Again gettimeofday() should be
 * vsyscall also if it's not you should seriously consider updating your
 * kernel.
 */
#ifdef HAVE_CLOCK_GETTIME
/* You can turn this on (set to 1) to prefer clock_gettime */
#define USE_CLOCK_GETTIME 1
#else
/* DON'T CHANGE THIS !!! */
#define USE_CLOCK_GETTIME 0
#endif

/* This is fairly safe to turn on - currently there appears to be a 'bug'
 * in DPDK that will remove the checksum by making the packet appear 4bytes
 * smaller than what it really is. Most formats don't include the checksum
 * hence writing out a port such as int: ring: and dpdk: assumes there
 * is no checksum and will attempt to write the checksum as part of the
 * packet
 */
#define GET_MAC_CRC_CHECKSUM 0

/* This requires a modification of the pmd drivers (inside Intel DPDK)
 * TODO this requires updating (packet sizes are wrong TS most likely also)
 */
#define HAS_HW_TIMESTAMPS_82580 0

#if HAS_HW_TIMESTAMPS_82580
# define TS_NBITS_82580     40
/* The maximum on the +ve or -ve side that we can be, make it half way */
# define MAXSKEW_82580 ((uint64_t) (.5 * (double)(1ull<<TS_NBITS_82580)))
#define WITHIN_VARIANCE(v1,v2,var) (((v1) - (var) < (v2)) && ((v1) + (var) > (v2)))
#endif

static pthread_mutex_t dpdk_lock = PTHREAD_MUTEX_INITIALIZER;
/* Memory pools Per NUMA node */
static struct rte_mempool * mem_pools[4][RTE_MAX_LCORE] = {{0}};

/* As per Intel 82580 specification - mismatch in 82580 datasheet
 * it states ts is stored in Big Endian, however its actually Little */
struct hw_timestamp_82580 {
        uint64_t reserved;
        uint64_t timestamp; /* Little Endian only lower 40 bits are valid */
};

enum paused_state {
        DPDK_NEVER_STARTED,
        DPDK_RUNNING,
        DPDK_PAUSED,
};

struct dpdk_per_stream_t
{
        uint16_t queue_id;
        uint64_t ts_last_sys; /* System timestamp of our most recent packet in nanoseconds */
        struct rte_mempool *mempool;
        int lcore;
#if HAS_HW_TIMESTAMPS_82580
        /* Timestamping only relevent to RX */
        uint64_t ts_first_sys; /* Sytem timestamp of the first packet in nanoseconds */
        uint32_t wrap_count; /* Number of times the NIC clock has wrapped around completely */
#endif
} ALIGN_STRUCT(CACHE_LINE_SIZE);

#if HAS_HW_TIMESTAMPS_82580
#define DPDK_EMPTY_STREAM {-1, 0, NULL, -1, 0, 0}
#else
#define DPDK_EMPTY_STREAM {-1, 0, NULL, -1}
#endif

typedef struct dpdk_per_stream_t dpdk_per_stream_t;

/* Used by both input and output however some fields are not used
 * for output */
struct dpdk_format_data_t {
        int8_t promisc; /* promiscuous mode - RX only */
        uint8_t port; /* Always 0 we only whitelist a single port - Shared TX & RX */
        uint8_t nb_ports; /* Total number of usable ports on system should be 1 */
        uint8_t paused; /* See paused_state */
        uint16_t link_speed; /* Link speed 10,100,1000,10000 etc. */
        int snaplen; /* The snap length for the capture - RX only */
        /* We always have to setup both rx and tx queues even if we don't want them */
        int nb_rx_buf; /* The number of packet buffers in the rx ring */
        int nb_tx_buf; /* The number of packet buffers in the tx ring */
        int nic_numa_node; /* The NUMA node that the NIC is attached to */
        struct rte_mempool * pktmbuf_pool; /* Our packet memory pool */
#if DPDK_USE_BLACKLIST
        struct rte_pci_addr blacklist[BLACK_LIST_SIZE]; /* Holds our device blacklist */
        unsigned int nb_blacklist; /* Number of blacklist items in are valid */
#endif
        char mempool_name[MEMPOOL_NAME_LEN]; /* The name of the mempool that we are using */
        uint8_t rss_key[40]; // This is the RSS KEY
        /* To improve single-threaded performance we always batch reading
         * packets, in a burst, otherwise the parallel library does this for us */
        struct rte_mbuf* burst_pkts[BURST_SIZE];
        int burst_size; /* The total number read in the burst */
        int burst_offset; /* The offset we are into the burst */

        /* Our parallel streams */
        libtrace_list_t *per_stream;
};

enum dpdk_addt_hdr_flags {
        INCLUDES_CHECKSUM = 0x1,
        INCLUDES_HW_TIMESTAMP = 0x2, /* Used with 82580 driver */
};

/**
 * A structure placed in front of the packet where we can store
 * additional information about the given packet.
 * +--------------------------+
 * |       rte_mbuf (pkt)     | sizeof(rte_mbuf)
 * +--------------------------+
 * |       dpdk_addt_hdr      | sizeof(dpdk_addt_hdr)
 * +--------------------------+
 * |           padding        | RTE_PKTMBUF_HEADROOM-sizeof(dpdk_addt_hdr)
 * +--------------------------+
 * *   hw_timestamp_82580     * 16 bytes Optional
 * +--------------------------+
 * |       Packet data        | Variable Size
 * |                          |
 */
struct dpdk_addt_hdr {
        uint64_t timestamp;
        uint8_t flags;
        uint8_t direction;
        uint8_t reserved1;
        uint8_t reserved2;
        uint32_t cap_len; /* The size to say the capture is */
};

/**
 * We want to blacklist all devices except those on the whitelist
 * (I say list, but yes it is only the one).
 *
 * The default behaviour of rte_pci_probe() will map every possible device
 * to its DPDK driver. The DPDK driver will take the ethernet device
 * out of the kernel (i.e. no longer /dev/ethx) and cannot be used.
 *
 * So blacklist all devices except the one that we wish to use so that
 * the others can still be used as standard ethernet ports.
 *
 * @return 0 if successful, otherwise -1 on error.
 */
#if DPDK_USE_BLACKLIST
static int blacklist_devices(struct dpdk_format_data_t *format_data, struct rte_pci_addr *whitelist)
{
        struct rte_pci_device *dev = NULL;
        format_data->nb_blacklist = 0;

        memset(format_data->blacklist, 0, sizeof (format_data->blacklist));

        TAILQ_FOREACH(dev, &device_list, next) {
        if (whitelist != NULL && whitelist->domain == dev->addr.domain
            && whitelist->bus == dev->addr.bus
            && whitelist->devid == dev->addr.devid
            && whitelist->function == dev->addr.function)
            continue;
                if (format_data->nb_blacklist >= sizeof (format_data->blacklist)
                                / sizeof (format_data->blacklist[0])) {
                        fprintf(stderr, "Warning: too many devices to blacklist consider"
                                        " increasing BLACK_LIST_SIZE");
                        break;
                }
                format_data->blacklist[format_data->nb_blacklist] = dev->addr;
                ++format_data->nb_blacklist;
        }

        rte_eal_pci_set_blacklist(format_data->blacklist, format_data->nb_blacklist);
        return 0;
}
#else /* DPDK_USE_BLACKLIST */
#include <rte_devargs.h>
static int whitelist_device(struct dpdk_format_data_t *format_data UNUSED, struct rte_pci_addr *whitelist)
{
        char pci_str[20] = {0};
        snprintf(pci_str, sizeof(pci_str), PCI_PRI_FMT,
                 whitelist->domain,
                 whitelist->bus,
                 whitelist->devid,
                 whitelist->function);
        if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI, pci_str) < 0) {
                return -1;
        }
        return 0;
}
#endif

/**
 * Parse the URI format as a pci address
 * Fills in addr, note core is optional and is unchanged if
 * a value for it is not provided.
 *
 * i.e. ./libtrace dpdk:0:1:0.0 -> 0:1:0.0
 * or ./libtrace dpdk:0:1:0.1-2 -> 0:1:0.1 (Using CPU core #2)
 */
static int parse_pciaddr(char * str, struct rte_pci_addr * addr, long * core) {
        int matches;
        assert(str);
        matches = sscanf(str, "%4"SCNx16":%2"SCNx8":%2"SCNx8".%2"SCNx8"-%ld",
                         &addr->domain, &addr->bus, &addr->devid,
                         &addr->function, core);
        if (matches >= 4) {
                return 0;
        } else {
                return -1;
        }
}

/**
 * Convert a pci address to the numa node it is
 * connected to.
 *
 * This checks /sys/bus/pci/devices/XXXX:XX:XX.X/numa_node
 * so we can call it before DPDK
 *
 * @return -1 if unknown otherwise a number 0 or higher of the numa node
 */
static int pci_to_numa(struct rte_pci_addr * dev_addr) {
        char path[50] = {0};
        FILE *file;

        /* Read from the system */
        snprintf(path, sizeof(path), "/sys/bus/pci/devices/"PCI_PRI_FMT"/numa_node",
                 dev_addr->domain,
                 dev_addr->bus,
                 dev_addr->devid,
                 dev_addr->function);

        if((file = fopen(path, "r")) != NULL) {
                int numa_node = -1;
                fscanf(file, "%d", &numa_node);
                fclose(file);
                return numa_node;
        }
        return -1;
}

#if DEBUG
/* For debugging */
static inline void dump_configuration()
{
        struct rte_config * global_config;
        long nb_cpu = sysconf(_SC_NPROCESSORS_ONLN);

        if (nb_cpu <= 0) {
                perror("sysconf(_SC_NPROCESSORS_ONLN) failed."
                       " Falling back to the first core.");
                nb_cpu = 1; /* fallback to just 1 core */
        }
        if (nb_cpu > RTE_MAX_LCORE)
                nb_cpu = RTE_MAX_LCORE;

        global_config = rte_eal_get_configuration();

        if (global_config != NULL) {
                int i;
                fprintf(stderr, "Intel DPDK setup\n"
                        "---Version      : %s\n"
                        "---Master LCore : %"PRIu32"\n"
                        "---LCore Count  : %"PRIu32"\n",
                        rte_version(),
                        global_config->master_lcore, global_config->lcore_count);

                for (i = 0 ; i < nb_cpu; i++) {
                        fprintf(stderr, "   ---Core %d : %s\n", i,
                                global_config->lcore_role[i] == ROLE_RTE ? "on" : "off");
                }

                const char * proc_type;
                switch (global_config->process_type) {
                case RTE_PROC_AUTO:
                        proc_type = "auto";
                        break;
                case RTE_PROC_PRIMARY:
                        proc_type = "primary";
                        break;
                case RTE_PROC_SECONDARY:
                        proc_type = "secondary";
                        break;
                case RTE_PROC_INVALID:
                        proc_type = "invalid";
                        break;
                default:
                        proc_type = "something worse than invalid!!";
                }
                fprintf(stderr, "---Process Type : %s\n", proc_type);
        }

}
#endif

/**
 * Expects to be called from the master lcore and moves it to the given dpdk id
 * @param core (zero indexed) If core is on the physical system affinity is bound otherwise
 *               affinity is set to all cores. Must be less than RTE_MAX_LCORE
 *               and not already in use.
 * @return 0 is successful otherwise -1 on error.
 */
static inline int dpdk_move_master_lcore(libtrace_t *libtrace, size_t core) {
        struct rte_config *cfg = rte_eal_get_configuration();
        cpu_set_t cpuset;
        int i;

        assert (core < RTE_MAX_LCORE);
        assert (rte_get_master_lcore() == rte_lcore_id());

        if (core == rte_lcore_id())
                return 0;

        /* Make sure we are not overwriting someone else */
        assert(!rte_lcore_is_enabled(core));

        /* Move the core */
        cfg->lcore_role[rte_lcore_id()] = ROLE_OFF;
        cfg->lcore_role[core] = ROLE_RTE;
        lcore_config[core].thread_id = lcore_config[rte_lcore_id()].thread_id;
        rte_eal_get_configuration()->master_lcore = core;
        RTE_PER_LCORE(_lcore_id) = core;

        /* Now change the affinity, either mapped to a single core or all accepted */
        CPU_ZERO(&cpuset);

        if (lcore_config[core].detected) {
                CPU_SET(core, &cpuset);
        } else {
                for (i = 0; i < RTE_MAX_LCORE; ++i) {
                        if (lcore_config[i].detected)
                                CPU_SET(i, &cpuset);
                }
        }

        i = pthread_setaffinity_np(pthread_self(), sizeof(cpuset), &cpuset);
        if (i != 0) {
                trace_set_err(libtrace, errno, "pthread_setaffinity_np failed\n");
                return -1;
        }
        return 0;
}

/**
 * XXX This is very bad XXX
 * But we have to do something to allow getopts nesting
 * Luckly normally the format is last so it doesn't matter
 * DPDK only supports modern systems so hopefully this
 * will continue to work
 */
struct saved_getopts {
        char *optarg;
        int optind;
        int opterr;
        int optopt;
};

static void save_getopts(struct saved_getopts *opts) {
        opts->optarg = optarg;
        opts->optind = optind;
        opts->opterr = opterr;
        opts->optopt = optopt;
}

static void restore_getopts(struct saved_getopts *opts) {
        optarg = opts->optarg;
        optind = opts->optind;
        opterr = opts->opterr;
        optopt = opts->optopt;
}

static inline int dpdk_init_environment(char * uridata, struct dpdk_format_data_t * format_data,
                                        char * err, int errlen) {
        int ret; /* Returned error codes */
        struct rte_pci_addr use_addr; /* The only address that we don't blacklist */
        char cpu_number[10] = {0}; /* The CPU mask we want to bind to */
        char mem_map[20] = {0}; /* The memory name */
        long nb_cpu; /* The number of CPUs in the system */
        long my_cpu; /* The CPU number we want to bind to */
        int i;
        struct rte_config *cfg = rte_eal_get_configuration();
        struct saved_getopts save_opts;

        /* This initialises the Environment Abstraction Layer (EAL)
         * If we had slave workers these are put into WAITING state
         *
         * Basically binds this thread to a fixed core, which we choose as
         * the last core on the machine (assuming fewer interrupts mapped here).
         * "-c" controls the cpu mask 0x1=1st core 0x2=2nd 0x4=3rd and so on
         * "-n" the number of memory channels into the CPU (hardware specific)
         *      - Most likely to be half the number of ram slots in your machine.
         *        We could count ram slots by "dmidecode -t 17 | grep -c 'Size:'"
         * Controls where in memory packets are stored such that they are spread
         * across the channels. We just use 1 to be safe.
         *
         * Using unique file prefixes mean separate memory is used, unlinking
         * the two processes. However be careful we still cannot access a
         * port that already in use.
         */
        char* argv[] = {"libtrace",
                        "-c", cpu_number,
                        "-n", "1",
                        "--proc-type", "auto",
                        "--file-prefix", mem_map,
                        "-m", "512",
#if DPDK_USE_LOG_LEVEL
#       if DEBUG
                        "--log-level", "8", /* RTE_LOG_DEBUG */
#       else
                        "--log-level", "5", /* RTE_LOG_WARNING */
#       endif
#endif
                        NULL};
        int argc = sizeof(argv) / sizeof(argv[0]) - 1;

#if DEBUG
        rte_set_log_level(RTE_LOG_DEBUG);
#else
        rte_set_log_level(RTE_LOG_WARNING);
#endif

        /* Get the number of cpu cores in the system and use the last core
         * on the correct numa node */
        nb_cpu = sysconf(_SC_NPROCESSORS_ONLN);
        if (nb_cpu <= 0) {
                perror("sysconf(_SC_NPROCESSORS_ONLN) failed."
                       " Falling back to the first core.");
                nb_cpu = 1; /* fallback to the first core */
        }
        if (nb_cpu > RTE_MAX_LCORE)
                nb_cpu = RTE_MAX_LCORE;

        my_cpu = -1;
        /* This allows the user to specify the core - we would try to do this
         * automatically but it's hard to tell that this is secondary
         * before running rte_eal_init(...). Currently we are limited to 1
         * instance per core due to the way memory is allocated. */
        if (parse_pciaddr(uridata, &use_addr, &my_cpu) != 0) {
                snprintf(err, errlen, "Failed to parse URI");
                return -1;
        }

#if HAVE_LIBNUMA
        format_data->nic_numa_node = pci_to_numa(&use_addr);
        if (my_cpu < 0) {
                /* If we can assign to a core on the same numa node */
                fprintf(stderr, "Using pci card on numa_node%d\n", format_data->nic_numa_node);
                if(format_data->nic_numa_node >= 0) {
                        int max_node_cpu = -1;
                        struct bitmask *mask = numa_allocate_cpumask();
                        assert(mask);
                        numa_node_to_cpus(format_data->nic_numa_node, mask);
                        for (i = 0 ; i < nb_cpu; ++i) {
                                if (numa_bitmask_isbitset(mask,i))
                                        max_node_cpu = i+1;
                        }
                        my_cpu = max_node_cpu;
                }
        }
#endif
        if (my_cpu < 0) {
                my_cpu = nb_cpu;
        }


        snprintf(format_data->mempool_name, MEMPOOL_NAME_LEN,
                 "libtrace_pool_%"PRIu32, (uint32_t) nb_cpu);

        if (!(my_cpu > 0 && my_cpu <= nb_cpu)) {
                snprintf(err, errlen,
                         "Intel DPDK - User defined a bad CPU number %"PRIu32" must be"
                         " between 1 and %"PRIu32, (uint32_t) my_cpu, (uint32_t) nb_cpu);
                return -1;
        }

        /* Make our mask with all cores turned on this is so that DPDK to
         * gets CPU info older versions */
        snprintf(cpu_number, sizeof(cpu_number), "%x", ~(UINT32_MAX<<MIN(31, nb_cpu)));
        //snprintf(cpu_number, sizeof(cpu_number), "%x", 0x1 << (my_cpu - 1));

#if !DPDK_USE_BLACKLIST
        /* Black list all ports besides the one that we want to use */
        if ((ret = whitelist_device(format_data, &use_addr)) < 0) {
                snprintf(err, errlen, "Intel DPDK - Whitelisting PCI device failed,"
                         " are you sure the address is correct?: %s", strerror(-ret));
                return -1;
        }
#endif

        /* Give the memory map a unique name */
        snprintf(mem_map, sizeof(mem_map), "libtrace-%d", (int) getpid());
        /* rte_eal_init it makes a call to getopt so we need to reset the
         * global optind variable of getopt otherwise this fails */
        save_getopts(&save_opts);
        optind = 1;
        if ((ret = rte_eal_init(argc, argv)) < 0) {
                snprintf(err, errlen,
                         "Intel DPDK - Initialisation of EAL failed: %s", strerror(-ret));
                return -1;
        }
        restore_getopts(&save_opts);
        // These are still running but will never do anything with DPDK v1.7 we
        // should remove this XXX in the future
        for(i = 0; i < RTE_MAX_LCORE; ++i) {
                if (rte_lcore_is_enabled(i) && i != (int) rte_get_master_lcore()) {
                        cfg->lcore_role[i] = ROLE_OFF;
                        cfg->lcore_count--;
                }
        }
        // Only the master should be running
        assert(cfg->lcore_count == 1);

        // TODO XXX TODO
        dpdk_move_master_lcore(NULL, my_cpu-1);

#if DEBUG
        dump_configuration();
#endif

#if DPDK_USE_PMD_INIT
        /* This registers all available NICs with Intel DPDK
         * These are not loaded until rte_eal_pci_probe() is called.
         */
        if ((ret = rte_pmd_init_all()) < 0) {
                snprintf(err, errlen,
                         "Intel DPDK - rte_pmd_init_all failed: %s", strerror(-ret));
                return -1;
        }
#endif

#if DPDK_USE_BLACKLIST
        /* Blacklist all ports besides the one that we want to use */
        if ((ret = blacklist_devices(format_data, &use_addr)) < 0) {
                snprintf(err, errlen, "Intel DPDK - Whitelisting PCI device failed,"
                         " are you sure the address is correct?: %s", strerror(-ret));
                return -1;
        }
#endif

#if DPDK_USE_PCI_PROBE
        /* This loads DPDK drivers against all ports that are not blacklisted */
        if ((ret = rte_eal_pci_probe()) < 0) {
                snprintf(err, errlen,
                         "Intel DPDK - rte_eal_pci_probe failed: %s", strerror(-ret));
                return -1;
        }
#endif

        format_data->nb_ports = rte_eth_dev_count();

        if (format_data->nb_ports != 1) {
                snprintf(err, errlen,
                         "Intel DPDK - rte_eth_dev_count returned %d but it should be 1",
                         format_data->nb_ports);
                return -1;
        }

        struct rte_eth_dev_info dev_info;
        rte_eth_dev_info_get(0, &dev_info);
        fprintf(stderr, "Device port=0\n\tmin_rx_bufsize=%d\n\tmax_rx_pktlen=%d\n\tmax rx queues=%d\n\tmax tx queues=%d",
                (int) dev_info.min_rx_bufsize, (int) dev_info.max_rx_pktlen, (int) dev_info.max_rx_queues, (int) dev_info.max_tx_queues);

        return 0;
}

static int dpdk_init_input (libtrace_t *libtrace) {
        dpdk_per_stream_t stream = DPDK_EMPTY_STREAM;
        char err[500];
        err[0] = 0;

        libtrace->format_data = (struct dpdk_format_data_t *)
                                malloc(sizeof(struct dpdk_format_data_t));
        FORMAT(libtrace)->port = 0; /* Always assume 1 port loaded */
        FORMAT(libtrace)->nb_ports = 0;
        FORMAT(libtrace)->snaplen = 0; /* Use default */
        FORMAT(libtrace)->nb_rx_buf = NB_RX_MBUF;
        FORMAT(libtrace)->nb_tx_buf = MIN_NB_BUF;
        FORMAT(libtrace)->nic_numa_node = -1;
        FORMAT(libtrace)->promisc = -1;
        FORMAT(libtrace)->pktmbuf_pool = NULL;
#if DPDK_USE_BLACKLIST
        FORMAT(libtrace)->nb_blacklist = 0;
#endif
        FORMAT(libtrace)->paused = DPDK_NEVER_STARTED;
        FORMAT(libtrace)->mempool_name[0] = 0;
        memset(FORMAT(libtrace)->burst_pkts, 0,
               sizeof(FORMAT(libtrace)->burst_pkts[0]) * BURST_SIZE);
        FORMAT(libtrace)->burst_size = 0;
        FORMAT(libtrace)->burst_offset = 0;

        /* Make our first stream */
        FORMAT(libtrace)->per_stream = libtrace_list_init(sizeof(struct dpdk_per_stream_t));
        libtrace_list_push_back(FORMAT(libtrace)->per_stream, &stream);

        if (dpdk_init_environment(libtrace->uridata, FORMAT(libtrace), err, sizeof(err)) != 0) {
                trace_set_err(libtrace, TRACE_ERR_INIT_FAILED, "%s", err);
                free(libtrace->format_data);
                libtrace->format_data = NULL;
                return -1;
        }
        return 0;
}

static int dpdk_init_output(libtrace_out_t *libtrace)
{
        char err[500];
        err[0] = 0;

        libtrace->format_data = (struct dpdk_format_data_t *)
                                malloc(sizeof(struct dpdk_format_data_t));
        FORMAT(libtrace)->port = 0; /* Always assume 1 port loaded */
        FORMAT(libtrace)->nb_ports = 0;
        FORMAT(libtrace)->snaplen = 0; /* Use default */
        FORMAT(libtrace)->nb_rx_buf = MIN_NB_BUF;
        FORMAT(libtrace)->nb_tx_buf = NB_TX_MBUF;
        FORMAT(libtrace)->nic_numa_node = -1;
        FORMAT(libtrace)->promisc = -1;
        FORMAT(libtrace)->pktmbuf_pool = NULL;
#if DPDK_USE_BLACKLIST
        FORMAT(libtrace)->nb_blacklist = 0;
#endif
        FORMAT(libtrace)->paused = DPDK_NEVER_STARTED;
        FORMAT(libtrace)->mempool_name[0] = 0;
        memset(FORMAT(libtrace)->burst_pkts, 0, sizeof(FORMAT(libtrace)->burst_pkts[0]) * BURST_SIZE);
        FORMAT(libtrace)->burst_size = 0;
        FORMAT(libtrace)->burst_offset = 0;

        if (dpdk_init_environment(libtrace->uridata, FORMAT(libtrace), err, sizeof(err)) != 0) {
                trace_set_err_out(libtrace, TRACE_ERR_INIT_FAILED, "%s", err);
                free(libtrace->format_data);
                libtrace->format_data = NULL;
                return -1;
        }
        return 0;
}

/**
 * Note here snaplen excludes the MAC checksum. Packets over
 * the requested snaplen will be dropped. (Excluding MAC checksum)
 *
 * I.e the maximum size of a standard ethernet packet is 1518 (Including MAC checksum)
 * So to allow packets upto 1518 this would be set to 1514 and if GET_MAC_CRC_CHECKSUM
 * is set the maximum size of the returned packet would be 1518 otherwise
 * 1514 would be the largest size possibly returned.
 *
 */
static int dpdk_config_input (libtrace_t *libtrace,
                              trace_option_t option,
                              void *data) {
        switch (option) {
        case TRACE_OPTION_SNAPLEN:
                /* Only support changing snaplen before a call to start is
                 * made */
                if (FORMAT(libtrace)->paused == DPDK_NEVER_STARTED)
                        FORMAT(libtrace)->snaplen=*(int*)data;
                else
                        return -1;
                return 0;
        case TRACE_OPTION_PROMISC:
                FORMAT(libtrace)->promisc=*(int*)data;
                return 0;
        case TRACE_OPTION_HASHER:
                switch (*((enum hasher_types *) data))
                {
                case HASHER_BALANCE:
                case HASHER_UNIDIRECTIONAL:
                        toeplitz_create_unikey(FORMAT(libtrace)->rss_key);
                        return 0;
                case HASHER_BIDIRECTIONAL:
                        toeplitz_create_bikey(FORMAT(libtrace)->rss_key);
                        return 0;
                case HASHER_CUSTOM:
                        // We don't support these
                        return -1;
                }
                break;
        case TRACE_OPTION_FILTER:
                /* TODO filtering */
        case TRACE_OPTION_META_FREQ:
        case TRACE_OPTION_EVENT_REALTIME:
                break;
        /* Avoid default: so that future options will cause a warning
         * here to remind us to implement it, or flag it as
         * unimplementable
         */
        }

        /* Don't set an error - trace_config will try to deal with the
         * option and will set an error if it fails */
        return -1;
}

/* Can set jumbo frames/ or limit the size of a frame by setting both
 * max_rx_pkt_len and jumbo_frame. This can be limited to less than
 *
 */
static struct rte_eth_conf port_conf = {
        .rxmode = {
                .mq_mode = ETH_RSS,
                .split_hdr_size = 0,
                .header_split   = 0, /**< Header Split disabled */
                .hw_ip_checksum = 0, /**< IP checksum offload disabled */
                .hw_vlan_filter = 0, /**< VLAN filtering disabled */
                .jumbo_frame    = 0, /**< Jumbo Frame Support disabled */
                .max_rx_pkt_len = 0, /**< Max frame Size if Jumbo enabled */
#if GET_MAC_CRC_CHECKSUM
/* So it appears that if hw_strip_crc is turned off the driver will still
 * take this off. See line 955ish in lib/librte_pmd_e1000/igb_rxtx.c.
 * So if .hw_strip_crc=0 a valid CRC exists 4 bytes after the end of the
 * So lets just add it back on when we receive the packet.
 */
                .hw_strip_crc   = 0, /**< CRC stripped by hardware */
#else
/* By default strip the MAC checksum because it's a bit of a hack to
 * actually read these. And don't want to rely on disabling this to actualy
 * always cut off the checksum in the future
 */
                .hw_strip_crc   = 1, /**< CRC stripped by hardware */
#endif
        },
        .txmode = {
                .mq_mode = ETH_DCB_NONE,
        },
        .rx_adv_conf = {
                .rss_conf = {
                        // .rss_key = &rss_key, // We set this per format
                        .rss_hf = ETH_RSS_IPV4_UDP | ETH_RSS_IPV6 | ETH_RSS_IPV4 | ETH_RSS_IPV4_TCP | ETH_RSS_IPV6_TCP | ETH_RSS_IPV6_UDP,
                },
        },
        .intr_conf = {
                .lsc = 1
        }
};

static const struct rte_eth_rxconf rx_conf = {
        .rx_thresh = {
                .pthresh = 8,/* RX_PTHRESH prefetch */
                .hthresh = 8,/* RX_HTHRESH host */
                .wthresh = 4,/* RX_WTHRESH writeback */
        },
        .rx_free_thresh = 0,
        .rx_drop_en = 0, /* Drop packets oldest packets if out of space */
};

static const struct rte_eth_txconf tx_conf = {
        .tx_thresh = {
                /*
                 * TX_PTHRESH prefetch
                 * Set on the NIC, if the number of unprocessed descriptors to queued on
                 * the card fall below this try grab at least hthresh more unprocessed
                 * descriptors.
                 */
                .pthresh = 36,

                /* TX_HTHRESH host
                 * Set on the NIC, the batch size to prefetch unprocessed tx descriptors.
                 */
                .hthresh = 0,

                /* TX_WTHRESH writeback
                 * Set on the NIC, the number of sent descriptors before writing back
                 * status to confirm the transmission. This is done more efficiently as
                 * a bulk DMA-transfer rather than writing one at a time.
                 * Similar to tx_free_thresh however this is applied to the NIC, where
                 * as tx_free_thresh is when DPDK will check these. This is extended
                 * upon by tx_rs_thresh (10Gbit cards) which doesn't write all
                 * descriptors rather only every n'th item, reducing DMA memory bandwidth.
                 */
                .wthresh = 4,
        },

        /* Used internally by DPDK rather than passed to the NIC. The number of
         * packet descriptors to send before checking for any responses written
         * back (to confirm the transmission). Default = 32 if set to 0)
         */
        .tx_free_thresh = 0,

        /* This is the Report Status threshold, used by 10Gbit cards,
         * This signals the card to only write back status (such as
         * transmission successful) after this minimum number of transmit
         * descriptors are seen. The default is 32 (if set to 0) however if set
         * to greater than 1 TX wthresh must be set to zero, because this is kindof
         * a replacement. See the dpdk programmers guide for more restrictions.
         */
        .tx_rs_thresh = 1,
};

/**
 * A callback for a link state change (LSC).
 *
 * Packets may be received before this notification. In fact the DPDK IGXBE
 * driver likes to put a delay upto 5sec before sending this.
 *
 * We use this to ensure the link speed is correct for our timestamp
 * calculations. Because packets might be received before the link up we still
 * update this when the packet is received.
 *
 * @param port The DPDK port
 * @param event The TYPE of event (expected to be RTE_ETH_EVENT_INTR_LSC)
 * @param cb_arg The dpdk_format_data_t structure associated with the format
 */
static void dpdk_lsc_callback(uint8_t port, enum rte_eth_event_type event,
                              void *cb_arg) {
        struct dpdk_format_data_t * format_data = cb_arg;
        struct rte_eth_link link_info;
        assert(event == RTE_ETH_EVENT_INTR_LSC);
        assert(port == format_data->port);

        rte_eth_link_get_nowait(port, &link_info);

        if (link_info.link_status)
                format_data->link_speed = link_info.link_speed;
        else
                format_data->link_speed = 0;

#if DEBUG
        fprintf(stderr, "LSC - link status is %s %s speed=%d\n",
                link_info.link_status ? "up" : "down",
                (link_info.link_duplex == ETH_LINK_FULL_DUPLEX) ?
                                          "full-duplex" : "half-duplex",
                (int) link_info.link_speed);
#endif

        /* Turns out DPDK drivers might not come back up if the link speed
         * changes. So we reset the autoneg procedure. This is very unsafe
         * we have have threads reading packets and we stop the port. */
#if 0
        if (!link_info.link_status) {
                int ret;
                rte_eth_dev_stop(port);
                ret = rte_eth_dev_start(port);
                if (ret < 0) {
                        fprintf(stderr, "Resetting the DPDK port failed : %s\n",
                                strerror(-ret));
                }
        }
#endif
}

/** Reserve a DPDK lcore ID for a thread globally.
 *
 * @param real If true allocate a real lcore, otherwise allocate a core which
 * does not exist on the local machine.
 * @param socket the prefered NUMA socket - only used if a real core is requested
 * @return a valid core, which can later be used with dpdk_register_lcore() or a
 * -1 if have run out of cores.
 *
 * If any thread is reading or freeing packets we need to register it here
 * due to TLS caches in the memory pools.
 */
static int dpdk_reserve_lcore(bool real, int socket) {
        int new_id = -1;
        int i;
        struct rte_config *cfg = rte_eal_get_configuration();

        pthread_mutex_lock(&dpdk_lock);
        /* If 'reading packets' fill in cores from 0 up and bind affinity
         * otherwise start from the MAX core (which is also the master) and work backwards
         * in this case physical cores on the system will not exist so we don't bind
         * these to any particular physical core */
        if (real) {
#if HAVE_LIBNUMA
                for (i = 0; i < RTE_MAX_LCORE; ++i) {
                        if (!rte_lcore_is_enabled(i) && numa_node_of_cpu(i) == socket) {
                                new_id = i;
                                if (!lcore_config[i].detected)
                                        new_id = -1;
                                break;
                        }
                }
#endif
                /* Retry without the the numa restriction */
                if (new_id == -1) {
                        for (i = 0; i < RTE_MAX_LCORE; ++i) {
                                if (!rte_lcore_is_enabled(i)) {
                                        new_id = i;
                                        if (!lcore_config[i].detected)
                                                fprintf(stderr, "Warning the"
                                                        " number of 'reading' "
                                                        "threads exceed cores\n");
                                        break;
                                }
                        }
                }
        } else {
                for (i = RTE_MAX_LCORE-1; i >= 0; --i) {
                        if (!rte_lcore_is_enabled(i)) {
                                new_id = i;
                                break;
                        }
                }
        }

        if (new_id != -1) {
                /* Enable the core in global DPDK structs */
                cfg->lcore_role[new_id] = ROLE_RTE;
                cfg->lcore_count++;
        }

        pthread_mutex_unlock(&dpdk_lock);
        return new_id;
}

/** Register a thread as a lcore
 * @param libtrace any error is set against libtrace on exit
 * @param real If this is a true lcore we will bind its affinty to the
 * requested core.
 * @param lcore The lcore as retrieved from dpdk_reserve_lcore()
 * @return 0, if successful otherwise -1 if an error occured (details are stored
 * in libtrace)
 *
 * @note This must be called from the thread being registered.
 */
static int dpdk_register_lcore(libtrace_t *libtrace, bool real, int lcore) {
        int ret;
        RTE_PER_LCORE(_lcore_id) = lcore;

        /* Set affinity bind to corresponding core */
        if (real) {
                cpu_set_t cpuset;
                CPU_ZERO(&cpuset);
                CPU_SET(rte_lcore_id(), &cpuset);
                ret = pthread_setaffinity_np(pthread_self(), sizeof(cpuset), &cpuset);
                if (ret != 0) {
                        trace_set_err(libtrace, errno, "Warning "
                                      "pthread_setaffinity_np failed");
                        return -1;
                }
        }

        return 0;
}

/** Allocates a new dpdk packet buffer memory pool.
 *
 * @param n The number of threads
 * @param pkt_size The packet size we need ot store
 * @param socket_id The NUMA socket id
 * @param A new mempool, if NULL query the DPDK library for the error code
 * see rte_mempool_create() documentation.
 *
 * This allocates a new pool or recycles an existing memory pool.
 * Call dpdk_free_memory() to free the memory.
 * We cannot delete memory so instead we store the pools, allowing them to be
 * re-used.
 */
static struct rte_mempool *dpdk_alloc_memory(unsigned n,
                                             unsigned pkt_size,
                                             int socket_id) {
        struct rte_mempool *ret;
        size_t j,k;
        char name[MEMPOOL_NAME_LEN];

        /* Add on packet size overheads */
        pkt_size += sizeof(struct rte_mbuf) + RTE_PKTMBUF_HEADROOM;

        pthread_mutex_lock(&dpdk_lock);

        if (socket_id == SOCKET_ID_ANY || socket_id > 4) {
                /* Best guess go for zero */
                socket_id = 0;
        }

        /* Find a valid pool */
        for (j = 0; j < RTE_MAX_LCORE && mem_pools[socket_id][j]; ++j) {
                if (mem_pools[socket_id][j]->size >= n &&
                    mem_pools[socket_id][j]->elt_size >= pkt_size) {
                        break;
                }
        }

        /* Find the end (+1) of the list */
        for (k = j; k < RTE_MAX_LCORE && mem_pools[socket_id][k]; ++k) {}

        if (mem_pools[socket_id][j]) {
                ret = mem_pools[socket_id][j];
                mem_pools[socket_id][j] = mem_pools[socket_id][k-1];
                mem_pools[socket_id][k-1] = NULL;
                mem_pools[socket_id][j] = NULL;
        } else {
                static uint32_t test = 10;
                test++;
                snprintf(name, MEMPOOL_NAME_LEN,
                         "libtrace_pool_%"PRIu32, test);

                ret = rte_mempool_create(name, n, pkt_size,
                                         128, sizeof(struct rte_pktmbuf_pool_private),
                                         rte_pktmbuf_pool_init, NULL,
                                         rte_pktmbuf_init, NULL,
                                         socket_id, 0);
        }

        pthread_mutex_unlock(&dpdk_lock);
        return ret;
}

/** Stores the memory against the DPDK library.
 *
 * @param mempool The mempool to free
 * @param socket_id The NUMA socket this mempool was allocated upon.
 *
 * Because we cannot free a memory pool, we verify it's full (i.e. unused) and
 * store the memory shared globally against the format.
 */
static void dpdk_free_memory(struct rte_mempool *mempool, int socket_id) {
        size_t i;
        pthread_mutex_lock(&dpdk_lock);

        /* We should have all entries back in the mempool */
        rte_mempool_audit(mempool);
        if (!rte_mempool_full(mempool)) {
                fprintf(stderr, "DPDK memory pool not empty %d of %d, please "
                        "free all packets before finishing a trace\n",
                        rte_mempool_count(mempool), mempool->size);
        }

        /* Find the end (+1) of the list */
        for (i = 0; i < RTE_MAX_LCORE && mem_pools[socket_id][i]; ++i) {}

        if (i >= RTE_MAX_LCORE) {
                fprintf(stderr, "Too many memory pools, dropping this one\n");
        } else {
                mem_pools[socket_id][i] = mempool;
        }

        pthread_mutex_unlock(&dpdk_lock);
}

/* Attach memory to the port and start (or restart) the port/s.
 */
static int dpdk_start_streams(struct dpdk_format_data_t *format_data,
                              char *err, int errlen, uint16_t rx_queues) {
        int ret, i;
        struct rte_eth_link link_info; /* Wait for link */
        dpdk_per_stream_t empty_stream = DPDK_EMPTY_STREAM;

        /* Already started */
        if (format_data->paused == DPDK_RUNNING)
                return 0;

        /* First time started we need to alloc our memory, doing this here
         * rather than in environment setup because we don't have snaplen then */
        if (format_data->paused == DPDK_NEVER_STARTED) {
                if (format_data->snaplen == 0) {
                        format_data->snaplen = RX_MBUF_SIZE;
                        port_conf.rxmode.jumbo_frame = 0;
                        port_conf.rxmode.max_rx_pkt_len = 0;
                } else {
                        /* Use jumbo frames */
                        port_conf.rxmode.jumbo_frame = 1;
                        port_conf.rxmode.max_rx_pkt_len = format_data->snaplen;
                }

#if GET_MAC_CRC_CHECKSUM
                /* This is additional overhead so make sure we allow space for this */
                format_data->snaplen += ETHER_CRC_LEN;
#endif
#if HAS_HW_TIMESTAMPS_82580
                format_data->snaplen += sizeof(struct hw_timestamp_82580);
#endif

                /* Create the mbuf pool, which is the place packets are allocated
                 * from - There is no free function (I cannot see one).
                 * NOTE: RX queue requires nb_packets + 1 otherwise it fails to
                 * allocate however that extra 1 packet is not used.
                 * (I assume <= vs < error some where in DPDK code)
                 * TX requires nb_tx_buffers + 1 in the case the queue is full
                 * so that will fill the new buffer and wait until slots in the
                 * ring become available.
                 */
#if DEBUG
                fprintf(stderr, "Creating mempool named %s\n", format_data->mempool_name);
#endif
                format_data->pktmbuf_pool = dpdk_alloc_memory(format_data->nb_tx_buf*2,
                                                              format_data->snaplen,
                                                              format_data->nic_numa_node);

                if (format_data->pktmbuf_pool == NULL) {
                        snprintf(err, errlen, "Intel DPDK - Initialisation of mbuf "
                                 "pool failed: %s", strerror(rte_errno));
                        return -1;
                }
        }

        /* ----------- Now do the setup for the port mapping ------------ */
        /* Order of calls must be
         * rte_eth_dev_configure()
         * rte_eth_tx_queue_setup()
         * rte_eth_rx_queue_setup()
         * rte_eth_dev_start()
         * other rte_eth calls
         */

        /* This must be called first before another *eth* function
         * 1+ rx, 1 tx queues, port_conf sets checksum stripping etc */
        ret = rte_eth_dev_configure(format_data->port, rx_queues, 1, &port_conf);
        if (ret < 0) {
                snprintf(err, errlen, "Intel DPDK - Cannot configure device port"
                         " %"PRIu8" : %s", format_data->port,
                         strerror(-ret));
                return -1;
        }
#if DEBUG
        fprintf(stderr, "Doing dev configure\n");
#endif
        /* Initialise the TX queue a minimum value if using this port for
         * receiving. Otherwise a larger size if writing packets.
         */
        ret = rte_eth_tx_queue_setup(format_data->port,
                                     0 /* queue XXX */,
                                     format_data->nb_tx_buf,
                                     SOCKET_ID_ANY,
                                     &tx_conf);
        if (ret < 0) {
                snprintf(err, errlen, "Intel DPDK - Cannot configure TX queue"
                         " on port %"PRIu8" : %s", format_data->port,
                         strerror(-ret));
                return -1;
        }

        /* Attach memory to our RX queues */
        for (i=0; i < rx_queues; i++) {
                dpdk_per_stream_t *stream;
#if DEBUG
                fprintf(stderr, "Configuring queue %d\n", i);
#endif

                /* Add storage for the stream */
                if (libtrace_list_get_size(format_data->per_stream) <= (size_t) i)
                        libtrace_list_push_back(format_data->per_stream, &empty_stream);
                stream = libtrace_list_get_index(format_data->per_stream, i)->data;
                stream->queue_id = i;

                if (stream->lcore == -1)
                        stream->lcore = dpdk_reserve_lcore(true, format_data->nic_numa_node);

                if (stream->lcore == -1) {
                        snprintf(err, errlen, "Intel DPDK - Failed to reserve a lcore"
                                 ". Too many threads?");
                        return -1;
                }

                if (stream->mempool == NULL) {
                        stream->mempool = dpdk_alloc_memory(
                                                  format_data->nb_rx_buf*2,
                                                  format_data->snaplen,
                                                  rte_lcore_to_socket_id(stream->lcore));

                        if (stream->mempool == NULL) {
                                snprintf(err, errlen, "Intel DPDK - Initialisation of mbuf "
                                         "pool failed: %s", strerror(rte_errno));
                                return -1;
                        }
                }

                /* Initialise the RX queue with some packets from memory */
                ret = rte_eth_rx_queue_setup(format_data->port,
                                             stream->queue_id,
                                             format_data->nb_rx_buf,
                                             format_data->nic_numa_node,
                                             &rx_conf,
                                             stream->mempool);
                if (ret < 0) {
                        snprintf(err, errlen, "Intel DPDK - Cannot configure"
                                 " RX queue on port %"PRIu8" : %s",
                                 format_data->port,
                                 strerror(-ret));
                        return -1;
                }
        }

#if DEBUG
        fprintf(stderr, "Doing start device\n");
#endif
        rte_eth_stats_reset(format_data->port);
        /* Start device */
        ret = rte_eth_dev_start(format_data->port);
        if (ret < 0) {
                snprintf(err, errlen, "Intel DPDK - rte_eth_dev_start failed : %s",
                         strerror(-ret));
                return -1;
        }

        /* Default promiscuous to on */
        if (format_data->promisc == -1)
                format_data->promisc = 1;

        if (format_data->promisc == 1)
                rte_eth_promiscuous_enable(format_data->port);
        else
                rte_eth_promiscuous_disable(format_data->port);

        /* We have now successfully started/unpased */
        format_data->paused = DPDK_RUNNING;


        /* Register a callback for link state changes */
        ret = rte_eth_dev_callback_register(format_data->port,
                                            RTE_ETH_EVENT_INTR_LSC,
                                            dpdk_lsc_callback,
                                            format_data);
#if DEBUG
        if (ret)
                fprintf(stderr, "rte_eth_dev_callback_register failed %d : %s\n",
                        ret, strerror(-ret));
#endif

        /* Get the current link status */
        rte_eth_link_get_nowait(format_data->port, &link_info);
        format_data->link_speed = link_info.link_speed;
#if DEBUG
        fprintf(stderr, "Link status is %d %d %d\n", (int) link_info.link_status,
                (int) link_info.link_duplex, (int) link_info.link_speed);
#endif

        return 0;
}

static int dpdk_start_input (libtrace_t *libtrace) {
        char err[500];
        err[0] = 0;

        /* Make sure we don't reserve an extra thread for this */
        FORMAT_DATA_FIRST(libtrace)->queue_id = rte_lcore_id();

        if (dpdk_start_streams(FORMAT(libtrace), err, sizeof(err), 1) != 0) {
                trace_set_err(libtrace, TRACE_ERR_INIT_FAILED, "%s", err);
                free(libtrace->format_data);
                libtrace->format_data = NULL;
                return -1;
        }
        return 0;
}

static inline size_t dpdk_get_max_rx_queues (uint8_t port_id) {
        struct rte_eth_dev_info dev_info;
        rte_eth_dev_info_get(port_id, &dev_info);
        return dev_info.max_rx_queues;
}

static inline size_t dpdk_processor_count () {
        long nb_cpu = sysconf(_SC_NPROCESSORS_ONLN);
        if (nb_cpu <= 0)
                return 1;
        else
                return (size_t) nb_cpu;
}

static int dpdk_pstart_input (libtrace_t *libtrace) {
        char err[500];
        int i=0, phys_cores=0;
        int tot = libtrace->perpkt_thread_count;
        libtrace_list_node_t *n;
        err[0] = 0;

        if (rte_lcore_id() != rte_get_master_lcore())
                fprintf(stderr, "Warning dpdk_pstart_input should be called"
                        " from the master DPDK thread!\n");

        /* If the master is not on the last thread we move it there */
        if (rte_get_master_lcore() != RTE_MAX_LCORE - 1) {
                if (dpdk_move_master_lcore(libtrace, RTE_MAX_LCORE - 1) != 0)
                        return -1;
        }

        /* Don't exceed the number of cores in the system/detected by dpdk
         * We don't have to force this but performance wont be good if we don't */
        for (i = 0; i < RTE_MAX_LCORE; ++i) {
                if (lcore_config[i].detected) {
                        if (rte_lcore_is_enabled(i)) {
#if DEBUG
                                fprintf(stderr, "Found core %d already in use!\n", i);
#endif
                        } else {
                                phys_cores++;
                        }
                }
        }
        /* If we are restarting we have already allocated some threads as such
         * we add these back to the count for this calculation */
        for (n = FORMAT_DATA_HEAD(libtrace); n; n = n->next) {
                dpdk_per_stream_t * stream = n->data;
                if (stream->lcore != -1)
                        phys_cores++;
        }

        tot = MIN(libtrace->perpkt_thread_count,
                  dpdk_get_max_rx_queues(FORMAT(libtrace)->port));
        tot = MIN(tot, phys_cores);

#if DEBUG
        fprintf(stderr, "Running pstart DPDK tot=%d req=%d phys=%d\n", tot,
                libtrace->perpkt_thread_count, phys_cores);
#endif

        if (dpdk_start_streams(FORMAT(libtrace), err, sizeof(err), tot) != 0) {
                trace_set_err(libtrace, TRACE_ERR_INIT_FAILED, "%s", err);
                free(libtrace->format_data);
                libtrace->format_data = NULL;
                return -1;
        }

        /* Make sure we only start the number that we should */
        libtrace->perpkt_thread_count = tot;
        return 0;
}

/**
 * Register a thread with the DPDK system,
 * When we start DPDK in parallel libtrace we move the 'main thread' to the
 * MAXIMUM CPU core slot (32) and remove any affinity restrictions DPDK
 * gives it.
 *
 * We then allow a mapper thread to be started on every real core as DPDK would,
 * we also bind these to the corresponding CPU cores.
 *
 * @param libtrace A pointer to the trace
 * @param reading True if the thread will be used to read packets, i.e. will
 *                call pread_packet(), false if thread used to process packet
 *                in any other manner including statistics functions.
 */
static int dpdk_pregister_thread(libtrace_t *libtrace, libtrace_thread_t *t, bool reading)
{
#if DEBUG
        char name[99];
        pthread_getname_np(pthread_self(),
                           name, sizeof(name));
#endif
        if (reading) {
                dpdk_per_stream_t *stream;
                /* Attach our thread */
                if(t->type == THREAD_PERPKT) {
                        t->format_data = libtrace_list_get_index(FORMAT(libtrace)->per_stream, t->perpkt_num)->data;
                        if (t->format_data == NULL) {
                                trace_set_err(libtrace, TRACE_ERR_INIT_FAILED,
                                              "Too many threads registered");
                                return -1;
                        }
                } else {
                        t->format_data = FORMAT_DATA_FIRST(libtrace);
                }
                stream = t->format_data;
#if DEBUG
                fprintf(stderr, "%s new id memory:%s cpu-core:%d\n", name, stream->mempool->name, rte_lcore_id());
#endif
                return dpdk_register_lcore(libtrace, true, stream->lcore);
        } else {
                int lcore = dpdk_reserve_lcore(reading, 0);
                if (lcore == -1) {
                        trace_set_err(libtrace, TRACE_ERR_INIT_FAILED, "Too many threads"
                                      " for DPDK");
                        return -1;
                }
#if DEBUG
                fprintf(stderr, "%s new id cpu-core:%d\n", name, rte_lcore_id());
#endif
                return dpdk_register_lcore(libtrace, false, lcore);
        }

        return 0;
}

/**
 * Unregister a thread with the DPDK system.
 *
 * Only previously registered threads should be calling this just before
 * they are destroyed.
 */
static void dpdk_punregister_thread(libtrace_t *libtrace UNUSED, libtrace_thread_t *t UNUSED)
{
        struct rte_config *cfg = rte_eal_get_configuration();

        assert(rte_lcore_id() < RTE_MAX_LCORE);
        pthread_mutex_lock(&dpdk_lock);
        /* Skip if master */
        if (rte_lcore_id() == rte_get_master_lcore()) {
                fprintf(stderr, "INFO: we are skipping unregistering the master lcore\n");
                pthread_mutex_unlock(&dpdk_lock);
                return;
        }

        /* Disable this core in global DPDK structs */
        cfg->lcore_role[rte_lcore_id()] = ROLE_OFF;
        cfg->lcore_count--;
        RTE_PER_LCORE(_lcore_id) = -1; // Might make the world burn if used again
        assert(cfg->lcore_count >= 1); // We cannot unregister the master LCORE!!
        pthread_mutex_unlock(&dpdk_lock);
        return;
}

static int dpdk_start_output(libtrace_out_t *libtrace)
{
        char err[500];
        err[0] = 0;

        if (dpdk_start_streams(FORMAT(libtrace), err, sizeof(err), 1) != 0) {
                trace_set_err_out(libtrace, TRACE_ERR_INIT_FAILED, "%s", err);
                free(libtrace->format_data);
                libtrace->format_data = NULL;
                return -1;
        }
        return 0;
}

static int dpdk_pause_input(libtrace_t * libtrace) {
        libtrace_list_node_t *tmp = FORMAT_DATA_HEAD(libtrace);
        /* This stops the device, but can be restarted using rte_eth_dev_start() */
        if (FORMAT(libtrace)->paused == DPDK_RUNNING) {
#if DEBUG
                fprintf(stderr, "Pausing DPDK port\n");
#endif
                rte_eth_dev_stop(FORMAT(libtrace)->port);
                FORMAT(libtrace)->paused = DPDK_PAUSED;
                /* Empty the queue of packets */
                for (; FORMAT(libtrace)->burst_offset < FORMAT(libtrace)->burst_size; ++FORMAT(libtrace)->burst_offset) {
                        rte_pktmbuf_free(FORMAT(libtrace)->burst_pkts[FORMAT(libtrace)->burst_offset]);
                }
                FORMAT(libtrace)->burst_offset = 0;
                FORMAT(libtrace)->burst_size = 0;

                for (; tmp != NULL; tmp = tmp->next) {
                        dpdk_per_stream_t *stream = tmp->data;
                        stream->ts_last_sys = 0;
#if HAS_HW_TIMESTAMPS_82580
                        stream->ts_first_sys = 0;
#endif
                }

        }
        return 0;
}

static int dpdk_write_packet(libtrace_out_t *trace,
                             libtrace_packet_t *packet){
        struct rte_mbuf* m_buff[1];

        int wirelen = trace_get_wire_length(packet);
        int caplen = trace_get_capture_length(packet);

        /* Check for a checksum and remove it */
        if (trace_get_link_type(packet) == TRACE_TYPE_ETH &&
            wirelen == caplen)
                caplen -= ETHER_CRC_LEN;

        m_buff[0] = rte_pktmbuf_alloc(FORMAT(trace)->pktmbuf_pool);
        if (m_buff[0] == NULL) {
                trace_set_err_out(trace, errno, "Cannot get an empty packet buffer");
                return -1;
        } else {
                int ret;
                memcpy(rte_pktmbuf_append(m_buff[0], caplen), packet->payload, caplen);
                do {
                        ret = rte_eth_tx_burst(0 /*queue TODO*/, FORMAT(trace)->port, m_buff, 1);
                } while (ret != 1);
        }

        return 0;
}

static int dpdk_fin_input(libtrace_t * libtrace) {
        libtrace_list_node_t * n;
        /* Free our memory structures */
        if (libtrace->format_data != NULL) {

                if (FORMAT(libtrace)->port != 0xFF)
                        rte_eth_dev_callback_unregister(FORMAT(libtrace)->port,
                                                        RTE_ETH_EVENT_INTR_LSC,
                                                        dpdk_lsc_callback,
                                                        FORMAT(libtrace));
                /* Close the device completely, device cannot be restarted */
                rte_eth_dev_close(FORMAT(libtrace)->port);

                dpdk_free_memory(FORMAT(libtrace)->pktmbuf_pool,
                                 FORMAT(libtrace)->nic_numa_node);

                for (n = FORMAT(libtrace)->per_stream->head; n ; n = n->next) {
                        dpdk_per_stream_t * stream = n->data;
                        if (stream->mempool)
                                dpdk_free_memory(stream->mempool,
                                                 rte_lcore_to_socket_id(stream->lcore));
                }

                libtrace_list_deinit(FORMAT(libtrace)->per_stream);
                /* filter here if we used it */
                free(libtrace->format_data);
        }

        return 0;
}


static int dpdk_fin_output(libtrace_out_t * libtrace) {
        /* Free our memory structures */
        if (libtrace->format_data != NULL) {
                /* Close the device completely, device cannot be restarted */
                if (FORMAT(libtrace)->port != 0xFF)
                        rte_eth_dev_close(FORMAT(libtrace)->port);
                libtrace_list_deinit(FORMAT(libtrace)->per_stream);
                /* filter here if we used it */
                free(libtrace->format_data);
        }

        return 0;
}

/**
 * Get the start of the additional header that we added to a packet.
 */
static inline struct dpdk_addt_hdr * get_addt_hdr (const libtrace_packet_t *packet) {
        assert(packet);
        assert(packet->buffer);
        /* Our header sits straight after the mbuf header */
        return (struct dpdk_addt_hdr *) ((struct rte_mbuf*) packet->buffer + 1);
}

static int dpdk_get_capture_length (const libtrace_packet_t *packet) {
        struct dpdk_addt_hdr * hdr = get_addt_hdr(packet);
        return hdr->cap_len;
}

static size_t dpdk_set_capture_length(libtrace_packet_t *packet, size_t size) {
        struct dpdk_addt_hdr * hdr = get_addt_hdr(packet);
        if (size > hdr->cap_len) {
                /* Cannot make a packet bigger */
                return trace_get_capture_length(packet);
        }

        /* Reset the cached capture length first*/
        packet->capture_length = -1;
        hdr->cap_len = (uint32_t) size;
        return trace_get_capture_length(packet);
}

static int dpdk_get_wire_length (const libtrace_packet_t *packet) {
        struct dpdk_addt_hdr * hdr = get_addt_hdr(packet);
        int org_cap_size; /* The original capture size */
        if (hdr->flags & INCLUDES_HW_TIMESTAMP) {
                org_cap_size = (int) rte_pktmbuf_pkt_len(MBUF(packet->buffer)) -
                               sizeof(struct hw_timestamp_82580);
        } else {
                org_cap_size = (int) rte_pktmbuf_pkt_len(MBUF(packet->buffer));
        }
        if (hdr->flags & INCLUDES_CHECKSUM) {
                return org_cap_size;
        } else {
                /* DPDK packets are always TRACE_TYPE_ETH packets */
                return org_cap_size + ETHER_CRC_LEN;
        }
}

static int dpdk_get_framing_length (const libtrace_packet_t *packet) {
        struct dpdk_addt_hdr * hdr = get_addt_hdr(packet);
        if (hdr->flags & INCLUDES_HW_TIMESTAMP)
                return sizeof(struct rte_mbuf) + RTE_PKTMBUF_HEADROOM +
                                sizeof(struct hw_timestamp_82580);
        else
                return sizeof(struct rte_mbuf) + RTE_PKTMBUF_HEADROOM;
}

static int dpdk_prepare_packet(libtrace_t *libtrace UNUSED,
                               libtrace_packet_t *packet, void *buffer,
                               libtrace_rt_types_t rt_type, uint32_t flags) {
        assert(packet);
        if (packet->buffer != buffer &&
            packet->buf_control == TRACE_CTRL_PACKET) {
                free(packet->buffer);
        }

        if ((flags & TRACE_PREP_OWN_BUFFER) == TRACE_PREP_OWN_BUFFER)
                packet->buf_control = TRACE_CTRL_PACKET;
        else
                packet->buf_control = TRACE_CTRL_EXTERNAL;

        packet->buffer = buffer;
        packet->header = buffer;

        /* Don't use pktmbuf_mtod will fail if the packet is a copy */
        packet->payload = (char *)buffer + dpdk_get_framing_length(packet);
        packet->type = rt_type;
        return 0;
}

/**
 * Given a packet size and a link speed, computes the
 * time to transmit in nanoseconds.
 *
 * @param format_data The dpdk format data from which we get the link speed
 *        and if unset updates it in a thread safe manner
 * @param pkt_size The size of the packet in bytes
 * @return The wire time in nanoseconds
 */
static inline uint32_t calculate_wire_time(struct dpdk_format_data_t* format_data, uint32_t pkt_size) {
        uint32_t wire_time;
        /* 20 extra bytes of interframe gap and preamble */
# if GET_MAC_CRC_CHECKSUM
        wire_time = ((pkt_size + 20) * 8000);
# else
        wire_time = ((pkt_size + 20 + ETHER_CRC_LEN) * 8000);
# endif

        /* Division is really slow and introduces a pipeline stall
         * The compiler will optimise this into magical multiplication and shifting
         * See http://ridiculousfish.com/blog/posts/labor-of-division-episode-i.html
         */
retry_calc_wiretime:
        switch (format_data->link_speed) {
        case ETH_LINK_SPEED_40G:
                wire_time /=  ETH_LINK_SPEED_40G;
                break;
        case ETH_LINK_SPEED_20G:
                wire_time /= ETH_LINK_SPEED_20G;
                break;
        case ETH_LINK_SPEED_10G:
                wire_time /= ETH_LINK_SPEED_10G;
                break;
        case ETH_LINK_SPEED_1000:
                wire_time /= ETH_LINK_SPEED_1000;
                break;
        case 0:
                {
                /* Maybe the link was down originally, but now it should be up */
                struct rte_eth_link link = {0};
                rte_eth_link_get_nowait(format_data->port, &link);
                if (link.link_status && link.link_speed) {
                        format_data->link_speed = link.link_speed;
#ifdef DEBUG
                        fprintf(stderr, "Link has come up updated speed=%d\n", (int) link.link_speed);
#endif
                        goto retry_calc_wiretime;
                }
                /* We don't know the link speed, make sure numbers are counting up */
                wire_time = 1;
                break;
                }
        default:
                wire_time /= format_data->link_speed;
        }
        return wire_time;
}

/**
 * Does any extra preperation to all captured packets
 * This includes adding our extra header to it with the timestamp,
 * and any snapping
 *
 * @param format_data The DPDK format data
 * @param plc The DPDK per lcore format data
 * @param pkts An array of size nb_pkts of DPDK packets
 */
static inline void dpdk_ready_pkts(libtrace_t *libtrace,
                                   struct dpdk_per_stream_t *plc,
                                   struct rte_mbuf **pkts,
                                   size_t nb_pkts) {
        struct dpdk_format_data_t *format_data = FORMAT(libtrace);
        struct dpdk_addt_hdr *hdr;
        size_t i;
        uint64_t cur_sys_time_ns;
#if HAS_HW_TIMESTAMPS_82580
        struct hw_timestamp_82580 *hw_ts;
        uint64_t estimated_wraps;
#else

#endif

#if USE_CLOCK_GETTIME
        struct timespec cur_sys_time = {0};
        /* This looks terrible and I feel bad doing it. But it's OK
         * on new kernels, because this is a fast vsyscall */
        clock_gettime(CLOCK_REALTIME, &cur_sys_time);
        cur_sys_time_ns = TS_TO_NS(cur_sys_time);
#else
        struct timeval cur_sys_time = {0};
        /* Also a fast vsyscall */
        gettimeofday(&cur_sys_time, NULL);
        cur_sys_time_ns = TV_TO_NS(cur_sys_time);
#endif

        /* The system clock is not perfect so when running
         * at linerate we could timestamp a packet in the past.
         * To avoid this we munge the timestamp to appear 1ns
         * after the previous packet. We should eventually catch up
         * to system time since a 64byte packet on a 10G link takes 67ns.
         *
         * Note with parallel readers timestamping packets
         * with duplicate stamps or out of order is unavoidable without
         * hardware timestamping from the NIC.
         */
#if !HAS_HW_TIMESTAMPS_82580
        if (plc->ts_last_sys >= cur_sys_time_ns) {
                cur_sys_time_ns = plc->ts_last_sys + 1;
        }
#endif

        ct_assert(RTE_PKTMBUF_HEADROOM >= sizeof(struct dpdk_addt_hdr));
        for (i = 0 ; i < nb_pkts ; ++i) {

                /* We put our header straight after the dpdk header */
                hdr = (struct dpdk_addt_hdr *) (pkts[i] + 1);
                memset(hdr, 0, sizeof(struct dpdk_addt_hdr));

#if GET_MAC_CRC_CHECKSUM
                /* Add back in the CRC sum */
                rte_pktmbuf_pkt_len(pkt) += ETHER_CRC_LEN;
                rte_pktmbuf_data_len(pkt) += ETHER_CRC_LEN;
                hdr->flags |= INCLUDES_CHECKSUM;
#endif

                hdr->cap_len = rte_pktmbuf_pkt_len(pkts[i]);

#if HAS_HW_TIMESTAMPS_82580
                /* The timestamp is sitting before our packet and is included in pkt_len */
                hdr->flags |= INCLUDES_HW_TIMESTAMP;
                hdr->cap_len -= sizeof(struct hw_timestamp_82580);
                hw_ts = (struct hw_timestamp_82580 *) MBUF_PKTDATA(pkts[i]);

                /* Taken from igb_ptp.c part of Intel Linux drivers (Good example code)
                 *
                 *        +----------+---+   +--------------+
                 *  82580 |    24    | 8 |   |      32      |
                 *        +----------+---+   +--------------+
                 *          reserved  \______ 40 bits _____/
                 *
                 * The 40 bit 82580 SYSTIM overflows every
                 *   2^40 * 10^-9 /  60  = 18.3 minutes.
                 *
                 * NOTE picture is in Big Endian order, in memory it's acutally in Little
                 * Endian (for the full 64 bits) i.e. picture is mirrored
                 */

                /* Despite what the documentation says this is in Little
                 * Endian byteorder. Mask the reserved section out.
                 */
                hdr->timestamp = le64toh(hw_ts->timestamp) &
                        ~(((~0ull)>>TS_NBITS_82580)<<TS_NBITS_82580);

                if (unlikely(plc->ts_first_sys == 0)) {
                        plc->ts_first_sys = cur_sys_time_ns - hdr->timestamp;
                        plc->ts_last_sys = plc->ts_first_sys;
                }

                /* This will have serious problems if packets aren't read quickly
                 * that is within a couple of seconds because our clock cycles every
                 * 18 seconds */
                estimated_wraps = (cur_sys_time_ns - plc->ts_last_sys)
                                  / (1ull<<TS_NBITS_82580);

                /* Estimated_wraps gives the number of times the counter should have
                 * wrapped (however depending on value last time it could have wrapped
                 * twice more (if hw clock is close to its max value) or once less (allowing
                 * for a bit of variance between hw and sys clock). But if the clock
                 * shouldn't have wrapped once then don't allow it to go backwards in time */
                if (unlikely(estimated_wraps >= 2)) {
                        /* 2 or more wrap arounds add all but the very last wrap */
                        plc->wrap_count += estimated_wraps - 1;
                }

                /* Set the timestamp to the lowest possible value we're considering */
                hdr->timestamp += plc->ts_first_sys +
                                  plc->wrap_count * (1ull<<TS_NBITS_82580);

                /* In most runs only the first if() will need evaluating - i.e our
                 * estimate is correct. */
                if (unlikely(!WITHIN_VARIANCE(cur_sys_time_ns,
                                              hdr->timestamp, MAXSKEW_82580))) {
                        /* Failed to match estimated_wraps-1 (or estimated_wraps in ==0 case) */
                        plc->wrap_count++;
                        hdr->timestamp += (1ull<<TS_NBITS_82580);
                        if (!WITHIN_VARIANCE(cur_sys_time_ns,
                                             hdr->timestamp, MAXSKEW_82580)) {
                                /* Failed to match estimated_wraps */
                                plc->wrap_count++;
                                hdr->timestamp += (1ull<<TS_NBITS_82580);
                                if (!WITHIN_VARIANCE(cur_sys_time_ns,
                                                     hdr->timestamp, MAXSKEW_82580)) {
                                        if (estimated_wraps == 0) {
                                                /* 0 case Failed to match estimated_wraps+2 */
                                                printf("WARNING - Hardware Timestamp failed to"
                                                       " match using systemtime!\n");
                                                hdr->timestamp = cur_sys_time_ns;
                                        } else {
                                                /* Failed to match estimated_wraps+1 */
                                                plc->wrap_count++;
                                                hdr->timestamp += (1ull<<TS_NBITS_82580);
                                                if (!WITHIN_VARIANCE(cur_sys_time_ns,
                                                                     hdr->timestamp, MAXSKEW_82580)) {
                                                        /* Failed to match estimated_wraps+2 */
                                                        printf("WARNING - Hardware Timestamp failed to"
                                                               " match using systemtime!!\n");
                                                }
                                        }
                                }
                        }
                }
#else

                hdr->timestamp = cur_sys_time_ns;
                /* Offset the next packet by the wire time of previous */
                calculate_wire_time(format_data, hdr->cap_len);

#endif
        }

        plc->ts_last_sys = cur_sys_time_ns;
        return;
}


static void dpdk_fin_packet(libtrace_packet_t *packet)
{
        if ( packet->buf_control == TRACE_CTRL_EXTERNAL ) {
                rte_pktmbuf_free(packet->buffer);
                packet->buffer = NULL;
        }
}

/** Reads at least one packet or returns an error
 */
static inline int dpdk_read_packet_stream (libtrace_t *libtrace,
                                           dpdk_per_stream_t *stream,
                                           libtrace_message_queue_t *mesg,
                                           struct rte_mbuf* pkts_burst[],
                                           size_t nb_packets) {
        size_t nb_rx; /* Number of rx packets we've recevied */
        while (1) {
                /* Poll for a batch of packets */
                nb_rx = rte_eth_rx_burst(FORMAT(libtrace)->port,
                                         stream->queue_id, pkts_burst, nb_packets);
                if (nb_rx > 0) {
                        /* Got some packets - otherwise we keep spining */
                        dpdk_ready_pkts(libtrace, stream, pkts_burst, nb_rx);
                        //fprintf(stderr, "Doing P READ PACKET port=%d q=%d\n", (int) FORMAT(libtrace)->port, (int) get_thread_table_num(libtrace));
                        return nb_rx;
                }
                /* Check the message queue this could be less than 0 */
                if (mesg && libtrace_message_queue_count(mesg) > 0)
                        return READ_MESSAGE;
                if (libtrace_halt)
                        return READ_EOF;
                /* Wait a while, polling on memory degrades performance
                 * This relieves the pressure on memory allowing the NIC to DMA */
                rte_delay_us(10);
        }

        /* We'll never get here - but if we did it would be bad */
        return READ_ERROR;
}

static int dpdk_pread_packets (libtrace_t *libtrace,
                                    libtrace_thread_t *t,
                                    libtrace_packet_t **packets,
                                    size_t nb_packets) {
        int nb_rx; /* Number of rx packets we've recevied */
        struct rte_mbuf* pkts_burst[nb_packets]; /* Array of pointer(s) */
        int i;
        dpdk_per_stream_t *stream = t->format_data;

        nb_rx = dpdk_read_packet_stream (libtrace, stream, &t->messages,
                                         pkts_burst, nb_packets);

        if (nb_rx > 0) {
                for (i = 0; i < nb_rx; ++i) {
                        if (packets[i]->buffer != NULL) {
                                /* The packet should always be finished */
                                assert(packets[i]->buf_control == TRACE_CTRL_PACKET);
                                free(packets[i]->buffer);
                        }
                        packets[i]->buf_control = TRACE_CTRL_EXTERNAL;
                        packets[i]->type = TRACE_RT_DATA_DPDK;
                        packets[i]->buffer = pkts_burst[i];
                        packets[i]->trace = libtrace;
                        packets[i]->error = 1;
                        dpdk_prepare_packet(libtrace, packets[i], packets[i]->buffer, packets[i]->type, 0);
                }
        }

        return nb_rx;
}

static int dpdk_read_packet (libtrace_t *libtrace, libtrace_packet_t *packet) {
        int nb_rx; /* Number of rx packets we've received */
        dpdk_per_stream_t *stream = FORMAT_DATA_FIRST(libtrace);

        /* Free the last packet buffer */
        if (packet->buffer != NULL) {
                /* The packet should always be finished */
                assert(packet->buf_control == TRACE_CTRL_PACKET);
                free(packet->buffer);
                packet->buffer = NULL;
        }

        packet->buf_control = TRACE_CTRL_EXTERNAL;
        packet->type = TRACE_RT_DATA_DPDK;

        /* Check if we already have some packets buffered */
        if (FORMAT(libtrace)->burst_size != FORMAT(libtrace)->burst_offset) {
                packet->buffer = FORMAT(libtrace)->burst_pkts[FORMAT(libtrace)->burst_offset++];
                dpdk_prepare_packet(libtrace, packet, packet->buffer, packet->type, 0);
                return 1; // TODO should be bytes read, which essentially useless anyway
        }

        nb_rx = dpdk_read_packet_stream (libtrace, stream, NULL,
                                         FORMAT(libtrace)->burst_pkts, BURST_SIZE);

        if (nb_rx > 0) {
                FORMAT(libtrace)->burst_size = nb_rx;
                FORMAT(libtrace)->burst_offset = 1;
                packet->buffer = FORMAT(libtrace)->burst_pkts[0];
                dpdk_prepare_packet(libtrace, packet, packet->buffer, packet->type, 0);
                return 1;
        }
        return nb_rx;
}

static struct timeval dpdk_get_timeval (const libtrace_packet_t *packet) {
        struct timeval tv;
        struct dpdk_addt_hdr * hdr = get_addt_hdr(packet);

        tv.tv_sec = hdr->timestamp / (uint64_t) 1000000000;
        tv.tv_usec = (hdr->timestamp % (uint64_t) 1000000000) / 1000;
        return tv;
}

static struct timespec dpdk_get_timespec (const libtrace_packet_t *packet) {
        struct timespec ts;
        struct dpdk_addt_hdr * hdr = get_addt_hdr(packet);

        ts.tv_sec = hdr->timestamp / (uint64_t) 1000000000;
        ts.tv_nsec = hdr->timestamp % (uint64_t) 1000000000;
        return ts;
}

static libtrace_linktype_t dpdk_get_link_type (const libtrace_packet_t *packet UNUSED) {
        return TRACE_TYPE_ETH; /* Always ethernet until proven otherwise */
}

static libtrace_direction_t dpdk_get_direction (const libtrace_packet_t *packet) {
        struct dpdk_addt_hdr * hdr = get_addt_hdr(packet);
        return (libtrace_direction_t) hdr->direction;
}

static libtrace_direction_t dpdk_set_direction(libtrace_packet_t *packet, libtrace_direction_t direction) {
        struct dpdk_addt_hdr * hdr = get_addt_hdr(packet);
        hdr->direction = (uint8_t) direction;
        return (libtrace_direction_t) hdr->direction;
}

static void dpdk_get_stats(libtrace_t *trace, libtrace_stat_t *stats) {
        struct rte_eth_stats dev_stats = {0};

        if (trace->format_data == NULL || FORMAT(trace)->port == 0xFF)
                return;

        /* Grab the current stats */
        rte_eth_stats_get(FORMAT(trace)->port, &dev_stats);

        stats->captured_valid = true;
        stats->captured = dev_stats.ipackets;

        /* Not that we support adding filters but if we did this
         * would work */
        stats->filtered += dev_stats.fdirmiss;

        stats->dropped_valid = true;
        stats->dropped = dev_stats.imissed;

        /* DPDK errors includes drops */
        stats->errors_valid = true;
        stats->errors = dev_stats.ierrors - dev_stats.imissed;

        stats->received_valid = true;
        stats->received = dev_stats.ipackets + dev_stats.imissed;

}

/* Attempts to read a packet in a non-blocking fashion. If one is not
 * available a SLEEP event is returned. We do not have the ability to
 * create a select()able file descriptor in DPDK.
 */
static libtrace_eventobj_t dpdk_trace_event(libtrace_t *trace,
                                            libtrace_packet_t *packet) {
        libtrace_eventobj_t event = {0,0,0.0,0};
        int nb_rx; /* Number of receive packets we've read */
        struct rte_mbuf* pkts_burst[1]; /* Array of 1 pointer(s) to rx buffers */

        do {

                /* See if we already have a packet waiting */
                nb_rx = rte_eth_rx_burst(FORMAT(trace)->port,
                                         FORMAT_DATA_FIRST(trace)->queue_id,
                                         pkts_burst, 1);

                if (nb_rx > 0) {
                        /* Free the last packet buffer */
                        if (packet->buffer != NULL) {
                                /* The packet should always be finished */
                                assert(packet->buf_control == TRACE_CTRL_PACKET);
                                free(packet->buffer);
                                packet->buffer = NULL;
                        }

                        packet->buf_control = TRACE_CTRL_EXTERNAL;
                        packet->type = TRACE_RT_DATA_DPDK;
                        event.type = TRACE_EVENT_PACKET;
                        dpdk_ready_pkts(trace, FORMAT_DATA_FIRST(trace), pkts_burst, 1);
                        packet->buffer = FORMAT(trace)->burst_pkts[0];
                        dpdk_prepare_packet(trace, packet, packet->buffer, packet->type, 0);
                        event.size = 1; // TODO should be bytes read, which essentially useless anyway

                        /* XXX - Check this passes the filter trace_read_packet normally
                         * does this for us but this wont */
                        if (trace->filter) {
                                if (!trace_apply_filter(trace->filter, packet)) {
                                        /* Failed the filter so we loop for another packet */
                                        trace->filtered_packets ++;
                                        continue;
                                }
                        }
                        trace->accepted_packets ++;
                } else {
                        /* We only want to sleep for a very short time - we are non-blocking */
                        event.type = TRACE_EVENT_SLEEP;
                        event.seconds = 0.0001;
                        event.size = 0;
                }

                /* If we get here we have our event */
                break;
        } while (1);

        return event;
}

static void dpdk_help(void) {
        printf("dpdk format module: $Revision: 1752 $\n");
        printf("Supported input URIs:\n");
        printf("\tdpdk:<domain:bus:devid.func>-<coreid>\n");
        printf("\tThe -<coreid> is optional \n");
        printf("\t e.g. dpdk:0000:01:00.1\n");
        printf("\t e.g. dpdk:0000:01:00.1-2 (Use the second CPU core)\n\n");
        printf("\t By default the last CPU core is used if not otherwise specified.\n");
        printf("\t Only a single libtrace instance of dpdk can use the same CPU core.\n");
        printf("\t Support for multiple simultaneous instances of dpdk format is currently limited.\n");
        printf("\n");
        printf("Supported output URIs:\n");
        printf("\tSame format as the input URI.\n");
        printf("\t e.g. dpdk:0000:01:00.1\n");
        printf("\t e.g. dpdk:0000:01:00.1-2 (Use the second CPU core)\n");
        printf("\n");
}

static struct libtrace_format_t dpdk = {
        "dpdk",
        "$Id$",
        TRACE_FORMAT_DPDK,
        NULL,                               /* probe filename */
        NULL,                               /* probe magic */
        dpdk_init_input,                    /* init_input */
        dpdk_config_input,                  /* config_input */
        dpdk_start_input,                   /* start_input */
        dpdk_pause_input,                   /* pause_input */
        dpdk_init_output,                   /* init_output */
        NULL,                               /* config_output */
        dpdk_start_output,                  /* start_ouput */
        dpdk_fin_input,                     /* fin_input */
        dpdk_fin_output,                    /* fin_output */
        dpdk_read_packet,                   /* read_packet */
        dpdk_prepare_packet,                /* prepare_packet */
        dpdk_fin_packet,                    /* fin_packet */
        dpdk_write_packet,                  /* write_packet */
        dpdk_get_link_type,                 /* get_link_type */
        dpdk_get_direction,                 /* get_direction */
        dpdk_set_direction,                 /* set_direction */
        NULL,                               /* get_erf_timestamp */
        dpdk_get_timeval,                   /* get_timeval */
        dpdk_get_timespec,                  /* get_timespec */
        NULL,                               /* get_seconds */
        NULL,                               /* seek_erf */
        NULL,                               /* seek_timeval */
        NULL,                               /* seek_seconds */
        dpdk_get_capture_length,            /* get_capture_length */
        dpdk_get_wire_length,               /* get_wire_length */
        dpdk_get_framing_length,            /* get_framing_length */
        dpdk_set_capture_length,            /* set_capture_length */
        NULL,                               /* get_received_packets */
        NULL,                               /* get_filtered_packets */
        NULL,                               /* get_dropped_packets */
        dpdk_get_stats,                     /* get_statistics */
        NULL,                               /* get_fd */
        dpdk_trace_event,                   /* trace_event */
        dpdk_help,                          /* help */
        NULL,                               /* next pointer */
        {true, 8},                          /* Live, NICs typically have 8 threads */
        dpdk_pstart_input,                  /* pstart_input */
        dpdk_pread_packets,                 /* pread_packets */
        dpdk_pause_input,                   /* ppause */
        dpdk_fin_input,                     /* p_fin */
        dpdk_pregister_thread,              /* pregister_thread */
        dpdk_punregister_thread,            /* punregister_thread */
        NULL                                /* get thread stats */
};

void dpdk_constructor(void) {
        register_format(&dpdk);
}