source: lib/trace_parallel.c @ 17a3dff

/*
 * This file is part of libtrace
 *
 * Copyright (c) 2007,2008,2009,2010 The University of Waikato, Hamilton,
 * New Zealand.
 *
 * 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
 *
 * $Id$
 *
 */


#define _GNU_SOURCE
#include "common.h"
#include "config.h"
#include <assert.h>
#include <errno.h>
#include <fcntl.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/stat.h>
#include <sys/types.h>
#ifndef WIN32
#include <sys/socket.h>
#endif
#include <stdarg.h>
#include <sys/param.h>

#ifdef HAVE_LIMITS_H
#  include <limits.h>
#endif

#ifdef HAVE_SYS_LIMITS_H
#  include <sys/limits.h>
#endif

#ifdef HAVE_NET_IF_ARP_H
#  include <net/if_arp.h>
#endif

#ifdef HAVE_NET_IF_H
#  include <net/if.h>
#endif

#ifdef HAVE_NETINET_IN_H
#  include <netinet/in.h>
#endif

#ifdef HAVE_NET_ETHERNET_H
#  include <net/ethernet.h>
#endif

#ifdef HAVE_NETINET_IF_ETHER_H
#  include <netinet/if_ether.h>
#endif

#include <time.h>
#ifdef WIN32
#include <sys/timeb.h>
#endif

#include "libtrace.h"
#include "libtrace_int.h"

#ifdef HAVE_PCAP_BPF_H
#  include <pcap-bpf.h>
#else
#  ifdef HAVE_NET_BPF_H
#    include <net/bpf.h>
#  endif
#endif


#include "libtrace_int.h"
#include "format_helper.h"
#include "rt_protocol.h"
#include "hash_toeplitz.h"

#include <pthread.h>
#include <signal.h>


extern int libtrace_parallel;

struct multithreading_stats {
        uint64_t full_queue_hits;
        uint64_t wait_for_fill_complete_hits;
} contention_stats[1024];


/**
 * @return True if the format supports parallel threads.
 */
static inline bool trace_supports_parallel(libtrace_t *trace)
{
        assert(trace);
        assert(trace->format);
        if (trace->format->pstart_input)
                return true;
        else
                return false;
        //return trace->format->pstart_input;
}

DLLEXPORT void print_contention_stats(libtrace_t *libtrace) {
        int i;
        struct multithreading_stats totals = {0};
        for (i = 0; i < libtrace->perpkt_thread_count ; i++) {
                printf("\nStats for perpkt thread#%d\n", i);
                printf("\tfull_queue_hits: %"PRIu64"\n", contention_stats[i].full_queue_hits);
                totals.full_queue_hits += contention_stats[i].full_queue_hits;
                printf("\twait_for_fill_complete_hits: %"PRIu64"\n", contention_stats[i].wait_for_fill_complete_hits);
                totals.wait_for_fill_complete_hits += contention_stats[i].wait_for_fill_complete_hits;
        }
        printf("\nTotals for perpkt threads\n");
        printf("\tfull_queue_hits: %"PRIu64"\n", totals.full_queue_hits);
        printf("\twait_for_fill_complete_hits: %"PRIu64"\n", totals.wait_for_fill_complete_hits);

        return;
}

inline void libtrace_zero_thread(libtrace_thread_t * t) {
        t->trace = NULL;
        t->ret = NULL;
        t->type = THREAD_EMPTY;
        libtrace_zero_ringbuffer(&t->rbuffer);
        libtrace_zero_vector(&t->vector);
        libtrace_zero_deque(&t->deque);
        t->recorded_first = false;
        t->perpkt_num = -1;
}

// Ints are aligned int is atomic so safe to read and write at same time
// However write must be locked, read doesn't (We never try read before written to table)
libtrace_thread_t * get_thread_table(libtrace_t *libtrace) {
        int i = 0;
        pthread_t tid = pthread_self();

        for (;i<libtrace->perpkt_thread_count ;++i) {
                if (pthread_equal(tid, libtrace->perpkt_threads[i].tid))
                        return &libtrace->perpkt_threads[i];
        }
        return NULL;
}

int get_thread_table_num(libtrace_t *libtrace) {
        int i = 0;
        pthread_t tid = pthread_self();
        for (;i<libtrace->perpkt_thread_count; ++i) {
                if (pthread_equal(tid, libtrace->perpkt_threads[i].tid))
                        return i;
        }
        return -1;
}

static libtrace_thread_t * get_thread_descriptor(libtrace_t *libtrace) {
        libtrace_thread_t *ret;
        if (!(ret = get_thread_table(libtrace))) {
                pthread_t tid = pthread_self();
                // Check if we are reducer or something else
                if (pthread_equal(tid, libtrace->reducer_thread.tid))
                        ret = &libtrace->reducer_thread;
                else if (pthread_equal(tid, libtrace->hasher_thread.tid))
                        ret = &libtrace->hasher_thread;
                else
                        ret = NULL;
        }
        return ret;
}

/**
 * Holds threads in a paused state, until released by broadcasting
 * the condition mutex.
 */
static void trace_thread_pause(libtrace_t *trace) {
        printf("Pausing thread #%d\n", get_thread_table_num(trace));
        assert(pthread_mutex_lock(&trace->libtrace_lock) == 0);
        trace->perpkts_pausing++;
        pthread_cond_broadcast(&trace->perpkt_cond);
        while (!trace->started) {
                assert(pthread_cond_wait(&trace->perpkt_cond, &trace->libtrace_lock) == 0);
        }
        trace->perpkts_pausing--;
        pthread_cond_broadcast(&trace->perpkt_cond);
        assert(pthread_mutex_unlock(&trace->libtrace_lock) == 0);
        printf("Releasing thread #%d\n", get_thread_table_num(trace));
}

/**
 * The is the entry point for our packet processing threads.
 */
static void* perpkt_threads_entry(void *data) {
        libtrace_t *trace = (libtrace_t *)data;
        libtrace_thread_t * t;
        libtrace_message_t message;
        libtrace_packet_t *packet = NULL;

        assert(pthread_mutex_lock(&trace->libtrace_lock) == 0);
        t = get_thread_table(trace);
        assert(t);
        //printf("Yay Started perpkt thread #%d\n", (int) get_thread_table_num(trace));
        assert(pthread_mutex_unlock(&trace->libtrace_lock) == 0);

        /* ~~~~~~~~~~~ Setup complete now we loop ~~~~~~~~~~~~~~~ */
        // Send a message to say we've started

        message.code = MESSAGE_STARTED;
        message.sender = t;
        message.additional = NULL;

        // Let the per_packet function know we have started
        (*trace->per_pkt)(trace, NULL, &message, t);


        for (;;) {
                int psize;

                if (libtrace_message_queue_try_get(&t->messages, &message) != LIBTRACE_MQ_FAILED) {
                        switch (message.code) {
                                case MESSAGE_PAUSE:
                                        trace_thread_pause(trace);
                                        break;
                                case MESSAGE_STOP:
                                        goto stop;
                        }
                        (*trace->per_pkt)(trace, NULL, &message, t);
                        continue;
                }

                if (trace->perpkt_thread_count == 1) {
                        if (!packet) {
                                if (!libtrace_ringbuffer_try_sread_bl(&trace->packet_freelist, (void **) &packet))
                                        packet = trace_create_packet();
                        }
                        assert(packet);
                        if ((psize = trace_read_packet(trace, packet)) <1) {
                                break;
                        }
                } else {
                        psize = trace_pread_packet(trace, &packet);
                }

                if (psize > 0) {
                        packet = (*trace->per_pkt)(trace, packet, NULL, t);
                        continue;
                }

                if (psize == -2)
                        continue; // We have a message

                if (psize < 1) { // consider sending a message
                        break;
                }

        }


stop:
        /* ~~~~~~~~~~~~~~ Trace is finished do tear down ~~~~~~~~~~~~~~~~~~~~~ */
        // Let the per_packet function know we have stopped
        message.code = MESSAGE_STOPPED;
        message.sender = message.additional = NULL;
        (*trace->per_pkt)(trace, NULL, &message, t);

        // And we're at the end free the memories
        assert(pthread_mutex_lock(&trace->libtrace_lock) == 0);
        t->state = THREAD_FINISHED;
        assert(pthread_mutex_unlock(&trace->libtrace_lock) == 0);

        // Notify only after we've defiantly set the state to finished
        message.code = MESSAGE_PERPKT_ENDED;
        message.additional = NULL;
        trace_send_message_to_reducer(trace, &message);

        pthread_exit(NULL);
};

/** True if trace has dedicated hasher thread otherwise false */
inline int trace_has_dedicated_hasher(libtrace_t * libtrace);
inline int trace_has_dedicated_hasher(libtrace_t * libtrace)
{
        return libtrace->hasher_thread.type == THREAD_HASHER;
}

/**
 * The start point for our single threaded hasher thread, this will read
 * and hash a packet from a data source and queue it against the correct
 * core to process it.
 */
static void* hasher_start(void *data) {
        libtrace_t *trace = (libtrace_t *)data;
        libtrace_thread_t * t;
        int i;
        libtrace_packet_t * packet;

        assert(trace_has_dedicated_hasher(trace));
        /* Wait until all threads are started and objects are initialised (ring buffers) */
        assert(pthread_mutex_lock(&trace->libtrace_lock) == 0);
        t = &trace->hasher_thread;
        assert(t->type == THREAD_HASHER && pthread_equal(pthread_self(), t->tid));
        printf("Hasher Thread started\n");
        assert(pthread_mutex_unlock(&trace->libtrace_lock) == 0);
        int pkt_skipped = 0;
        /* Read all packets in then hash and queue against the correct thread */
        while (1) {
                int thread;
                if (!pkt_skipped && !libtrace_ringbuffer_try_sread_bl(&trace->packet_freelist, (void **) &packet))
                        packet = trace_create_packet();
                assert(packet);

                if (libtrace_halt) // Signal to die has been sent - TODO
                        break;

                if ((packet->error = trace_read_packet(trace, packet)) <1 /*&& psize != LIBTRACE_MESSAGE_WAITING*/) {
                        break; /* We are EOF or error'd either way we stop  */
                }

                /* We are guaranteed to have a hash function i.e. != NULL */
                trace_packet_set_hash(packet, (*trace->hasher)(packet, trace->hasher_data));
                thread = trace_packet_get_hash(packet) % trace->perpkt_thread_count;
                /* Blocking write to the correct queue - I'm the only writer */
                if (trace->perpkt_threads[thread].state != THREAD_FINISHED) {
                        libtrace_ringbuffer_write(&trace->perpkt_threads[thread].rbuffer, packet);
                        pkt_skipped = 0;
                } else {
                        pkt_skipped = 1; // Reuse that packet no one read it
                }
        }

        /* Broadcast our last failed read to all threads */
        for (i = 0; i < trace->perpkt_thread_count; i++) {
                libtrace_packet_t * bcast;
                printf("Broadcasting error/EOF now the trace is over\n");
                if (i == trace->perpkt_thread_count - 1) {
                        bcast = packet;
                } else {
                        bcast = trace_create_packet();
                        bcast->error = packet->error;
                }
                assert(pthread_mutex_lock(&trace->libtrace_lock) == 0);
                if (trace->perpkt_threads[i].state != THREAD_FINISHED) {
                        assert(pthread_mutex_unlock(&trace->libtrace_lock) == 0);
                        // Unlock early otherwise we could deadlock
                        libtrace_ringbuffer_write(&trace->perpkt_threads[i].rbuffer, NULL);
                } else {
                        assert(pthread_mutex_unlock(&trace->libtrace_lock) == 0);
                }
        }
        // We dont need to free packet

        // And we're at the end free the memories
        t->state = THREAD_FINISHED;

        // Notify only after we've defiantly set the state to finished
        libtrace_message_t message;
        message.code = MESSAGE_PERPKT_ENDED;
        message.additional = NULL;
        trace_send_message_to_reducer(trace, &message);

        // TODO remove from TTABLE t sometime
        pthread_exit(NULL);
};

/**
 * Moves src into dest(Complete copy) and copies the memory buffer and
 * its flags from dest into src ready for reuse without needing extra mallocs.
 */
static inline void swap_packets(libtrace_packet_t *dest, libtrace_packet_t *src) {
        // Save the passed in buffer status
        assert(dest->trace == NULL); // Must be a empty packet
        void * temp_buf = dest->buffer;
        buf_control_t temp_buf_control = dest->buf_control;
        // Completely copy StoredPacket into packet
        memcpy(dest, src, sizeof(libtrace_packet_t));
        // Set the buffer settings on the returned packet
        src->buffer = temp_buf;
        src->buf_control = temp_buf_control;
        src->trace = NULL;
}

/* Our simplest case when a thread becomes ready it can obtain a exclusive
 * lock to read a packet from the underlying trace.
 */
inline static int trace_pread_packet_first_in_first_served(libtrace_t *libtrace, libtrace_packet_t **packet)
{
        // We need this to fill the 'first' packet table
        libtrace_thread_t *t = get_thread_table(libtrace);
        if (!*packet) {
                if (!libtrace_ringbuffer_try_sread_bl(&libtrace->packet_freelist, (void **) packet))
                        *packet = trace_create_packet();
        }
        assert(*packet);
        assert(pthread_mutex_lock(&libtrace->libtrace_lock) == 0);
        /* Read a packet */
        (*packet)->error = trace_read_packet(libtrace, *packet);
        // Doing this inside the lock ensures the first packet is always
        // recorded first
        store_first_packet(libtrace, *packet, t);

        assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);
        return (*packet)->error;
}

/**
 * For the case that we have a dedicated hasher thread
 * 1. We read a packet from our buffer
 * 2. Move that into the packet provided (packet)
 */
inline static int trace_pread_packet_hasher_thread(libtrace_t *libtrace, libtrace_packet_t **packet)
{
        int this_thread = get_thread_table_num(libtrace); // Could be worth caching ... ?
        libtrace_thread_t* t = &libtrace->perpkt_threads[this_thread];

        if (*packet) // Recycle the old get the new
                if (!libtrace_ringbuffer_try_swrite_bl(&libtrace->packet_freelist, (void *) *packet))
                        trace_destroy_packet(*packet);
        *packet = libtrace_ringbuffer_read(&t->rbuffer);

        if (*packet) {
                return 1;
        } else {
                printf("Got a NULL packet the trace is over\n");
                return -1; // We are done for some reason
        }
}

/**
 * Tries to read from our queue and returns 1 if a packet was retrieved
 */
static inline int try_waiting_queue(libtrace_t *libtrace, libtrace_thread_t * t, libtrace_packet_t **packet, int * ret)
{
        libtrace_packet_t* retrived_packet;

        /* Lets see if we have one waiting */
        if (libtrace_ringbuffer_try_read(&t->rbuffer, (void **) &retrived_packet)) {
                /* Copy paste from trace_pread_packet_hasher_thread() except that we try read (non-blocking) */
                assert(retrived_packet);

                if (*packet) // Recycle the old get the new
                        if (!libtrace_ringbuffer_try_swrite_bl(&libtrace->packet_freelist, (void *) *packet))
                                trace_destroy_packet(*packet);
                *packet = retrived_packet;
                *ret = (*packet)->error;
                return 1;
        }
        return 0;
}

/**
 * Allows us to ensure all threads are finished writing to our threads ring_buffer
 * before returning EOF/error.
 */
inline static int trace_handle_finishing_perpkt(libtrace_t *libtrace, libtrace_packet_t **packet, libtrace_thread_t * t)
{
        /* We are waiting for the condition that another thread ends to check
         * our queue for new data, once all threads end we can go to finished */
        bool complete = false;
        int ret;

        do {
                // Wait for a thread to end
                assert(pthread_mutex_lock(&libtrace->libtrace_lock) == 0);

                // Check before
                if (libtrace->perpkts_finishing == libtrace->perpkt_thread_count) {
                        complete = true;
                        assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);
                        continue;
                }

                assert(pthread_cond_wait(&libtrace->perpkt_cond, &libtrace->libtrace_lock) == 0);

                // Check after
                if (libtrace->perpkts_finishing == libtrace->perpkt_thread_count) {
                        complete = true;
                        assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);
                        continue;
                }

                assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);

                // Always trying to keep our buffer empty for the unlikely case more threads than buffer space want to write into our queue
                if(try_waiting_queue(libtrace, t, packet, &ret))
                        return ret;
        } while (!complete);

        // We can only end up here once all threads complete
        try_waiting_queue(libtrace, t, packet, &ret);

        return ret;
        // TODO rethink this logic fix bug here
}

/**
 * Expects the libtrace_lock to not be held
 */
inline static int trace_finish_perpkt(libtrace_t *libtrace, libtrace_packet_t **packet, libtrace_thread_t * t)
{
        assert(pthread_mutex_lock(&libtrace->libtrace_lock) == 0);
        t->state = THREAD_FINISHING;
        libtrace->perpkts_finishing++;
        pthread_cond_broadcast(&libtrace->perpkt_cond);
        assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);
        return trace_handle_finishing_perpkt(libtrace, packet, t);
}

/**
 * This case is much like the dedicated hasher, except that we will become
 * hasher if we don't have a a packet waiting.
 *
 * Note: This is only every used if we have are doing hashing.
 *
 * TODO: Can block on zero copy formats such as ring: and dpdk: if the
 * queue sizes in total are larger than the ring size.
 *
 * 1. We read a packet from our buffer
 * 2. Move that into the packet provided (packet)
 */
inline static int trace_pread_packet_hash_locked(libtrace_t *libtrace, libtrace_packet_t **packet)
{
        int this_thread = get_thread_table_num(libtrace); // Could be worth caching ... ?
        libtrace_thread_t * t = &libtrace->perpkt_threads[this_thread];
        int thread, ret/*, psize*/;

        while (1) {
                if(try_waiting_queue(libtrace, t, packet, &ret))
                        return ret;
                // Can still block here if another thread is writing to a full queue
                assert(pthread_mutex_lock(&libtrace->libtrace_lock) == 0);

                // Its impossible for our own queue to overfill, because no one can write
                // when we are in the lock
                if(try_waiting_queue(libtrace, t, packet, &ret)) {
                        assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);
                        return ret;
                }

                // Another thread cannot write a packet because a queue has filled up. Is it ours?
                if (libtrace->perpkt_queue_full) {
                        contention_stats[this_thread].wait_for_fill_complete_hits++;
                        assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);
                        continue;
                }

                if (!*packet) {
                        if (!libtrace_ringbuffer_try_sread_bl(&libtrace->packet_freelist, (void **) packet))
                                *packet = trace_create_packet();
                }
                assert(*packet);

                // If we fail here we can guarantee that our queue is empty (and no new data will be added because we hold the lock)
                if (libtrace_halt || ((*packet)->error = trace_read_packet(libtrace, *packet)) <1 /*&& psize != LIBTRACE_MESSAGE_WAITING*/) {
                        assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);
                        if (libtrace_halt)
                                return 0;
                        else
                                return (*packet)->error;
                }

                trace_packet_set_hash(*packet, (*libtrace->hasher)(*packet, libtrace->hasher_data));
                thread = trace_packet_get_hash(*packet) % libtrace->perpkt_thread_count;
                if (thread == this_thread) {
                        // If it's this thread we must be in order because we checked the buffer once we got the lock
                        assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);
                        return (*packet)->error;
                }

                if (libtrace->perpkt_threads[thread].state != THREAD_FINISHED) {
                        while (!libtrace_ringbuffer_try_swrite_bl(&libtrace->perpkt_threads[thread].rbuffer, *packet)) {
                                libtrace->perpkt_queue_full = true;
                                assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);
                                contention_stats[this_thread].full_queue_hits++;
                                assert(pthread_mutex_lock(&libtrace->libtrace_lock) == 0);
                        }
                        *packet = NULL;
                        libtrace->perpkt_queue_full = false;
                } else {
                        /* We can get here if the user closes the thread before natural completion/or error */
                        assert (!"packet_hash_locked() The user terminated the trace in a abnormal manner");
                }
                assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);
        }
}

/**
 * This case is much like the dedicated hasher, except that we will become
 * hasher if we don't have a a packet waiting.
 *
 * TODO: You can loose the tail of a trace if the final thread
 * fills its own queue and therefore breaks early and doesn't empty the sliding window.
 *
 * TODO: Can block on zero copy formats such as ring: and dpdk: if the
 * queue sizes in total are larger than the ring size.
 *
 * 1. We read a packet from our buffer
 * 2. Move that into the packet provided (packet)
 */
inline static int trace_pread_packet_sliding_window(libtrace_t *libtrace, libtrace_packet_t **packet)
{
        int this_thread = get_thread_table_num(libtrace); // Could be worth caching ... ?
        libtrace_thread_t * t = &libtrace->perpkt_threads[this_thread];
        int ret, i, thread/*, psize*/;

        if (t->state == THREAD_FINISHING)
                return trace_handle_finishing_perpkt(libtrace, packet, t);

        while (1) {
                // Check if we have packets ready
                if(try_waiting_queue(libtrace, t, packet, &ret))
                        return ret;

                // We limit the number of packets we get to the size of the sliding window
                // such that it is impossible for any given thread to fail to store a packet
                assert(sem_wait(&libtrace->sem) == 0);
                /*~~~~Single threaded read of a packet~~~~*/
                assert(pthread_mutex_lock(&libtrace->libtrace_lock) == 0);

                /* Re-check our queue things we might have data waiting */
                if(try_waiting_queue(libtrace, t, packet, &ret)) {
                        assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);
                        assert(sem_post(&libtrace->sem) == 0);
                        return ret;
                }

                // TODO put on *proper* condition variable
                if (libtrace->perpkt_queue_full) {
                        assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);
                        assert(sem_post(&libtrace->sem) == 0);
                        contention_stats[this_thread].wait_for_fill_complete_hits++;
                        continue;
                }

                if (!*packet) {
                        if (!libtrace_ringbuffer_try_sread_bl(&libtrace->packet_freelist, (void **) packet))
                                *packet = trace_create_packet();
                }
                assert(*packet);

                if (libtrace_halt || ((*packet)->error = trace_read_packet(libtrace, *packet)) <1 /*&& psize != LIBTRACE_MESSAGE_WAITING*/) {
                        assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);
                        assert(sem_post(&libtrace->sem) == 0);
                        // Finish this thread ensuring that any data written later by another thread is retrieved also
                        if (libtrace_halt)
                                return 0;
                        else
                                return trace_finish_perpkt(libtrace, packet, t);
                }
                assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);

                /* ~~~~Multiple threads can run the hasher~~~~ */
                trace_packet_set_hash(*packet, (*libtrace->hasher)(*packet, libtrace->hasher_data));

                /* Yes this is correct opposite read lock for a write operation */
                assert(pthread_rwlock_rdlock(&libtrace->window_lock) == 0);
                if (!libtrace_slidingwindow_try_write(&libtrace->sliding_window, trace_packet_get_order(*packet), *packet))
                        assert(!"Semaphore should stop us from ever overfilling the sliding window");
                assert(pthread_rwlock_unlock(&libtrace->window_lock) == 0);
                *packet = NULL;

                // Always try read any data from the sliding window
                while (libtrace_slidingwindow_read_ready(&libtrace->sliding_window)) {
                        assert(pthread_rwlock_wrlock(&libtrace->window_lock) == 0);
                        if (libtrace->perpkt_queue_full) {
                                // I might be the holdup in which case if I can read my queue I should do that and return
                                if(try_waiting_queue(libtrace, t, packet, &ret)) {
                                        assert(pthread_rwlock_unlock(&libtrace->window_lock) == 0);
                                        return ret;
                                }
                                assert(pthread_rwlock_unlock(&libtrace->window_lock) == 0);
                                continue;
                        }
                        // Read greedily as many as we can
                        while (libtrace_slidingwindow_try_read(&libtrace->sliding_window, (void **) packet, NULL)) {
                                thread = trace_packet_get_hash(*packet) % libtrace->perpkt_thread_count;
                                if (libtrace->perpkt_threads[thread].state != THREAD_FINISHED) {
                                        while (!libtrace_ringbuffer_try_swrite_bl(&libtrace->perpkt_threads[thread].rbuffer, *packet)) {
                                                if (this_thread == thread)
                                                {
                                                        // TODO think about this case more because we have to stop early if this were to happen on the last read
                                                        // before EOF/error we might not have emptied the sliding window
                                                        printf("!~!~!~!~!~!~In this Code~!~!~!~!\n");
                                                        // Its our queue we must have a packet to read out
                                                        if(try_waiting_queue(libtrace, t, packet, &ret)) {
                                                                // We must be able to write this now 100% without fail
                                                                libtrace_ringbuffer_write(&libtrace->perpkt_threads[thread].rbuffer, *packet);
                                                                assert(sem_post(&libtrace->sem) == 0);
                                                                assert(pthread_rwlock_unlock(&libtrace->window_lock) == 0);
                                                                return ret;
                                                        } else {
                                                                assert(!"Our queue is full but I cannot read from it??");
                                                        }
                                                }
                                                // Not us we have to give the other threads a chance to write there packets then
                                                libtrace->perpkt_queue_full = true;
                                                assert(pthread_rwlock_unlock(&libtrace->window_lock) == 0);
                                                for (i = 0; i < libtrace->perpkt_thread_count-1; i++) // Release all other threads to read there packets
                                                        assert(sem_post(&libtrace->sem) == 0);

                                                contention_stats[this_thread].full_queue_hits++;
                                                assert(pthread_rwlock_wrlock(&libtrace->window_lock) == 0);
                                                // Grab these back
                                                for (i = 0; i < libtrace->perpkt_thread_count-1; i++) // Release all other threads to read there packets
                                                        assert(sem_wait(&libtrace->sem) == 0);
                                                libtrace->perpkt_queue_full = false;
                                        }
                                        assert(sem_post(&libtrace->sem) == 0);
                                        *packet = NULL;
                                } else {
                                        // Cannot write to a queue if no ones waiting (I think this is unreachable)
                                        // in the general case (unless the user ends early without proper clean up).
                                        assert (!"unreachable code??");
                                }
                        }
                        assert(pthread_rwlock_unlock(&libtrace->window_lock) == 0);
                }
                // Now we go back to checking our queue anyways
        }
}


/**
 * For the first packet of each queue we keep a copy and note the system
 * time it was received at.
 *
 * This is used for finding the first packet when playing back a trace
 * in trace time. And can be used by real time applications to print
 * results out every XXX seconds.
 */
inline void store_first_packet(libtrace_t *libtrace, libtrace_packet_t *packet, libtrace_thread_t *t)
{
        if (!t->recorded_first) {
                struct timeval tv;
                libtrace_packet_t * dup;
                // For what it's worth we can call these outside of the lock
                gettimeofday(&tv, NULL);
                dup = trace_copy_packet(packet);
                assert(pthread_spin_lock(&libtrace->first_packets.lock) == 0);
                libtrace->first_packets.packets[t->perpkt_num].packet = dup;
                //printf("Stored first packet time=%f\n", trace_get_seconds(dup));
                memcpy(&libtrace->first_packets.packets[t->perpkt_num].tv, &tv, sizeof(tv));
                // Now update the first
                libtrace->first_packets.count++;
                if (libtrace->first_packets.count == 1) {
                        // We the first entry hence also the first known packet
                        libtrace->first_packets.first = t->perpkt_num;
                } else {
                        // Check if we are newer than the previous 'first' packet
                        size_t first = libtrace->first_packets.first;
                        if (trace_get_seconds(dup) <
                                trace_get_seconds(libtrace->first_packets.packets[first].packet))
                                libtrace->first_packets.first = t->perpkt_num;
                }
                assert(pthread_spin_unlock(&libtrace->first_packets.lock) == 0);
                libtrace_message_t mesg;
                mesg.code = MESSAGE_FIRST_PACKET;
                mesg.additional = NULL;
                trace_send_message_to_reducer(libtrace, &mesg);
                t->recorded_first = true;
        }
}

/**
 * Returns 1 if it's certain that the first packet is truly the first packet
 * rather than a best guess based upon threads that have published so far.
 * Otherwise 0 is returned.
 * It's recommended that this result is stored rather than calling this
 * function again.
 */
DLLEXPORT int retrive_first_packet(libtrace_t *libtrace, libtrace_packet_t **packet, struct timeval **tv)
{
        int ret = 0;
        assert(pthread_spin_lock(&libtrace->first_packets.lock) == 0);
        if (libtrace->first_packets.count) {
                *packet = libtrace->first_packets.packets[libtrace->first_packets.first].packet;
                *tv = &libtrace->first_packets.packets[libtrace->first_packets.first].tv;
                if (libtrace->first_packets.count == libtrace->perpkt_thread_count) {
                        ret = 1;
                } else {
                        struct timeval curr_tv;
                        // If a second has passed since the first entry we will assume this is the very first packet
                        gettimeofday(&curr_tv, NULL);
                        if (curr_tv.tv_sec > (*tv)->tv_sec) {
                                if(curr_tv.tv_usec > (*tv)->tv_usec || curr_tv.tv_sec - (*tv)->tv_sec > 1) {
                                        ret = 1;
                                }
                        }
                }
        } else {
                *packet = NULL;
                *tv = NULL;
        }
        assert(pthread_spin_unlock(&libtrace->first_packets.lock) == 0);
        return ret;
}


DLLEXPORT inline uint64_t tv_to_usec(struct timeval *tv)
{
        return (uint64_t) tv->tv_sec*1000000ull + (uint64_t) tv->tv_usec;
}

inline static struct timeval usec_to_tv(uint64_t usec)
{
        struct timeval tv;
        tv.tv_sec = usec / 1000000;
        tv.tv_usec = usec % 1000000;
        return tv;
}


/**
 * Delays a packets playback so the playback will be in trace time
 */
static inline void delay_tracetime(libtrace_t *libtrace, libtrace_packet_t *packet, libtrace_thread_t *t) {
        struct timeval curr_tv, pkt_tv;
        uint64_t next_release = t->tracetime_offset_usec; // Time at which to release the packet
        uint64_t curr_usec;
        /* Tracetime we might delay releasing this packet */
        if (!t->tracetime_offset_usec) {
                libtrace_packet_t * first_pkt;
                struct timeval *sys_tv;
                int64_t initial_offset;
                int stable = retrive_first_packet(libtrace, &first_pkt, &sys_tv);
                assert(first_pkt);
                pkt_tv = trace_get_timeval(first_pkt);
                initial_offset = (int64_t)tv_to_usec(sys_tv) - (int64_t)tv_to_usec(&pkt_tv);
                if (stable)
                        // 0->1 because 0 is used to mean unset
                        t->tracetime_offset_usec = initial_offset ? initial_offset: 1;
                next_release = initial_offset;
        }
        /* next_release == offset */
        pkt_tv = trace_get_timeval(packet);
        next_release += tv_to_usec(&pkt_tv);
        gettimeofday(&curr_tv, NULL);
        curr_usec = tv_to_usec(&curr_tv);
        if (next_release > curr_usec) {
                // We need to wait
                struct timeval delay_tv = usec_to_tv(next_release-curr_usec);
                //printf("WAITING for %d.%d next=%"PRIu64" curr=%"PRIu64" seconds packettime %f\n", delay_tv.tv_sec, delay_tv.tv_usec, next_release, curr_usec, trace_get_seconds(packet));
                select(0, NULL, NULL, NULL, &delay_tv);
        }
}

/* Read one packet from the trace into a buffer. Note that this function will
 * block until a packet is read (or EOF is reached).
 *
 * @param libtrace      the libtrace opaque pointer
 * @param packet        the packet opaque pointer
 * @returns 0 on EOF, negative value on error
 *
 * Note this is identical to read_packet but calls pread_packet instead of
 * read packet in the format.
 *
 */
static inline int trace_pread_packet_wrapper(libtrace_t *libtrace, libtrace_packet_t *packet) {

        assert(libtrace && "You called trace_read_packet() with a NULL libtrace parameter!\n");
        if (trace_is_err(libtrace))
                return -1;
        if (!libtrace->started) {
                trace_set_err(libtrace,TRACE_ERR_BAD_STATE,"You must call libtrace_start() before trace_read_packet()\n");
                return -1;
        }
        if (!(packet->buf_control==TRACE_CTRL_PACKET || packet->buf_control==TRACE_CTRL_EXTERNAL)) {
                trace_set_err(libtrace,TRACE_ERR_BAD_STATE,"Packet passed to trace_read_packet() is invalid\n");
                return -1;
        }
        assert(packet);

        if (libtrace->format->read_packet) {
                do {
                        size_t ret;
                        /* Finalise the packet, freeing any resources the format module
                         * may have allocated it and zeroing all data associated with it.
                         */
                        trace_fin_packet(packet);
                        /* Store the trace we are reading from into the packet opaque
                         * structure */
                        packet->trace = libtrace;
                        ret=libtrace->format->pread_packet(libtrace,packet);
                        if (ret==(size_t)-1 || ret==(size_t)-2 || ret==0) {
                                return ret;
                        }
                        if (libtrace->filter) {
                                /* If the filter doesn't match, read another
                                 * packet
                                 */
                                if (!trace_apply_filter(libtrace->filter,packet)){
                                        ++libtrace->filtered_packets;
                                        continue;
                                }
                        }
                        if (libtrace->snaplen>0) {
                                /* Snap the packet */
                                trace_set_capture_length(packet,
                                                libtrace->snaplen);
                        }
                        trace_packet_set_order(packet, libtrace->accepted_packets);
                        ++libtrace->accepted_packets;
                        return ret;
                } while(1);
        }
        trace_set_err(libtrace,TRACE_ERR_UNSUPPORTED,"This format does not support reading packets\n");
        return ~0U;
}

/**
 * Read a packet from the parallel trace
 */
DLLEXPORT int trace_pread_packet(libtrace_t *libtrace, libtrace_packet_t **packet)
{
        int ret;
        libtrace_thread_t *t = get_thread_table(libtrace);

        // Cleanup the packet passed back
        if (*packet)
                trace_fin_packet(*packet);

        if (libtrace->format->pread_packet) {
                if (!*packet)
                        *packet = trace_create_packet();
                ret = trace_pread_packet_wrapper(libtrace, *packet);
        } else  if (!libtrace->hasher) {
                /* We don't care about which core a packet goes to */
                ret =  trace_pread_packet_first_in_first_served(libtrace, packet);
        } else if (trace_has_dedicated_hasher(libtrace)) {
                ret = trace_pread_packet_hasher_thread(libtrace, packet);
        } else if (libtrace->reducer_flags & PERPKT_USE_SLIDING_WINDOW) {
                ret = trace_pread_packet_sliding_window(libtrace, packet);
        } else {
                ret = trace_pread_packet_hash_locked(libtrace, packet);
        }

        // Formats can also optionally do this internally to ensure the first
        // packet is always reported correctly
        if (ret > 0) {
                store_first_packet(libtrace, *packet, t);
                if (libtrace->tracetime)
                        delay_tracetime(libtrace, *packet, t);
        }

        return ret;
}

/* Starts perpkt threads
 * @return threads_started
 */
static inline int trace_start_perpkt_threads (libtrace_t *libtrace) {
        int i;

        for (i = 0; i < libtrace->perpkt_thread_count; i++) {
                libtrace_thread_t *t = &libtrace->perpkt_threads[i];
                assert(pthread_create(&t->tid, NULL, perpkt_threads_entry, (void *) libtrace) == 0);
        }
        return libtrace->perpkt_thread_count;
}

/* Start an input trace in a parallel fashion.
 *
 * @param libtrace      the input trace to start
 * @param global_blob some global data you can share with the new thread
 * @returns 0 on success
 */
DLLEXPORT int trace_pstart(libtrace_t *libtrace, void* global_blob, fn_per_pkt per_pkt, fn_reducer reducer)
{
        int i;
        sigset_t sig_before, sig_block_all;

        assert(libtrace);
        if (trace_is_err(libtrace))
                return -1;;
        if (libtrace->perpkts_pausing != 0) {
                printf("Restarting trace\n");
                libtrace->format->pstart_input(libtrace);
                // TODO empty any queues out here //
                assert(pthread_mutex_lock(&libtrace->libtrace_lock) == 0);
                libtrace->started = true;
                assert(pthread_cond_broadcast(&libtrace->perpkt_cond) == 0);
                assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);
                return 0;
        }

        libtrace_parallel = 1;

        // Store the user defined things against the trace
        libtrace->global_blob = global_blob;
        libtrace->per_pkt = per_pkt;
        libtrace->reducer = reducer;
        libtrace->perpkts_finishing = 0;
        // libtrace->hasher = &rand_hash; /* Hasher now set via option */

        assert(pthread_mutex_init(&libtrace->libtrace_lock, NULL) == 0);
        assert(pthread_cond_init(&libtrace->perpkt_cond, NULL) == 0);
        assert(pthread_rwlock_init(&libtrace->window_lock, NULL) == 0);
        assert(pthread_mutex_lock(&libtrace->libtrace_lock) == 0);

        // Set default buffer sizes
        if (libtrace->perpkt_buffer_size <= 0)
                libtrace->perpkt_buffer_size = 1000;

        if (libtrace->perpkt_thread_count <= 0)
                libtrace->perpkt_thread_count = 2; // XXX scale to system

        if(libtrace->packet_freelist_size <= 0)
                libtrace->packet_freelist_size = (libtrace->perpkt_buffer_size + 1) * libtrace->perpkt_thread_count;

        if(libtrace->packet_freelist_size <
                (libtrace->perpkt_buffer_size + 1) * libtrace->perpkt_thread_count)
                fprintf(stderr, "WARNING deadlocks may occur and extra memory allocating buffer sizes (packet_freelist_size) mismatched\n");

        libtrace->started=true; // Before we start the threads otherwise we could have issues
        /* Disable signals - Pthread signal handling */

        sigemptyset(&sig_block_all);

        assert(pthread_sigmask(SIG_SETMASK, &sig_block_all, &sig_before) == 0);

        // If we are using a hasher start it
        if (libtrace->hasher && libtrace->hasher_thread.type == THREAD_HASHER) {
                libtrace_thread_t *t = &libtrace->hasher_thread;
                t->trace = libtrace;
                t->ret = NULL;
                t->type = THREAD_HASHER;
                t->state = THREAD_RUNNING;
                assert(pthread_create(&t->tid, NULL, hasher_start, (void *) libtrace) == 0);
        } else {
                libtrace->hasher_thread.type = THREAD_EMPTY;
        }
        libtrace_ringbuffer_init(&libtrace->packet_freelist, libtrace->packet_freelist_size, LIBTRACE_RINGBUFFER_POLLING);
        libtrace_slidingwindow_init(&libtrace->sliding_window, libtrace->packet_freelist_size, 0);
        assert(sem_init(&libtrace->sem, 0, libtrace->packet_freelist_size) == 0);
        // This will be applied to every new thread that starts, i.e. they will block all signals
        // Lets start a fixed number of reading threads

        // For now we never have a dedicated thread for the reducer
        // i.e. This main thread is used as the reducer
        libtrace->reducer_thread.tid = pthread_self();
        libtrace->reducer_thread.type = THREAD_REDUCER;
        libtrace->reducer_thread.state = THREAD_RUNNING;
        libtrace_message_queue_init(&libtrace->reducer_thread.messages, sizeof(libtrace_message_t));

        /* Ready some storages */
        libtrace->first_packets.first = 0;
        libtrace->first_packets.count = 0;
        assert(pthread_spin_init(&libtrace->first_packets.lock, 0) == 0);
        libtrace->first_packets.packets = calloc(libtrace->perpkt_thread_count, sizeof(struct  __packet_storage_magic_type));


        /* Start all of our perpkt threads */
        libtrace->perpkt_threads = calloc(sizeof(libtrace_thread_t), libtrace->perpkt_thread_count);
        for (i = 0; i < libtrace->perpkt_thread_count; i++) {
                libtrace_thread_t *t = &libtrace->perpkt_threads[i];
                t->trace = libtrace;
                t->ret = NULL;
                t->type = THREAD_PERPKT;
                t->state = THREAD_RUNNING;
                t->user_data = NULL;
                // t->tid DONE on create
                t->perpkt_num = i;
                if (libtrace->hasher)
                        libtrace_ringbuffer_init(&t->rbuffer, libtrace->perpkt_buffer_size, LIBTRACE_RINGBUFFER_POLLING);
                // Depending on the mode vector or deque might be chosen
                libtrace_vector_init(&t->vector, sizeof(libtrace_result_t));
                libtrace_deque_init(&t->deque, sizeof(libtrace_result_t));
                libtrace_message_queue_init(&t->messages, sizeof(libtrace_message_t));
                t->tmp_key = 0;
                t->tmp_data = NULL;
                t->recorded_first = false;
                assert(pthread_spin_init(&t->tmp_spinlock, 0) == 0);
                t->tracetime_offset_usec = 0;;
        }

        int threads_started = 0;
        /* Setup the trace and start our threads */
        if (libtrace->perpkt_thread_count > 1 && libtrace->format->pstart_input) {
                printf("This format has direct support for p's\n");
                threads_started = libtrace->format->pstart_input(libtrace);
        } else {
                if (libtrace->format->start_input) {
                        threads_started=libtrace->format->start_input(libtrace);
                }
        }
        if (threads_started == 0)
                threads_started = trace_start_perpkt_threads(libtrace);


        // Revert back - Allow signals again
        assert(pthread_sigmask(SIG_SETMASK, &sig_before, NULL) == 0);
        assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);

        if (threads_started < 0)
                // Error
                return threads_started;

        // TODO fix these leaks etc
        if (libtrace->perpkt_thread_count != threads_started)
                printf("Warning started threads not equal requested s=%d r=%d", threads_started, libtrace->perpkt_thread_count);


        return 0;
}

/**
 * Pauses a trace, this should only be called by the main thread
 * 1. Set started = false 
 * 2. All perpkt threads are paused waiting on a condition var
 * 3. Then call ppause on the underlying format if found
 * 4. Return with perpkts_pausing set to perpkt_count (Used when restarting so we reuse the threads)
 * 
 * Once done you should be a able to modify the trace setup and call pstart again
 * TODO handle changing thread numbers
 */
DLLEXPORT int trace_ppause(libtrace_t *libtrace)
{
        libtrace_thread_t *t;
        int i;
        assert(libtrace);
        if (!libtrace->started) {
                trace_set_err(libtrace,TRACE_ERR_BAD_STATE, "You must call trace_start() before calling trace_ppause()");
                return -1;
        }

        t = get_thread_table(libtrace);

        // Set paused
        assert(pthread_mutex_lock(&libtrace->libtrace_lock) == 0);
        libtrace->started = false;
        pthread_cond_broadcast(&libtrace->perpkt_cond);
        assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);

        printf("Sending messages \n");
        // Stop threads, skip this one if its a perpkt
        for (i = 0; i < libtrace->perpkt_thread_count; i++) {
                if (&libtrace->perpkt_threads[i] != t) {
                        libtrace_message_t message;
                        message.code = MESSAGE_PAUSE;
                        message.additional = NULL;
                        trace_send_message_to_thread(libtrace, &libtrace->perpkt_threads[i], &message);
                }
        }

        // Formats must support native message handling if a message is ready
        // Approach per Perry's suggestion is a non-blocking read
        // followed by a blocking read. XXX STRIP THIS OUT

        if (t) {
                // A perpkt is doing the pausing interesting fake a extra thread paused
                // We rely on the user to not return before starting the trace again
                assert(pthread_mutex_lock(&libtrace->libtrace_lock) == 0);
                libtrace->perpkts_pausing++;
                assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);
        }

        printf("Threads are pausing\n");

        // Do a early pause to kick threads out - XXX testing for int
        if (libtrace->format->pause_input)
                        libtrace->format->pause_input(libtrace);

        // Wait for all threads to pause
        assert(pthread_mutex_lock(&libtrace->libtrace_lock) == 0);
        while (libtrace->perpkt_thread_count != libtrace->perpkts_pausing) {
                assert(pthread_cond_wait(&libtrace->perpkt_cond, &libtrace->libtrace_lock) == 0);
        }
        assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);

        printf("Threads have paused\n");

        if (trace_supports_parallel(libtrace)) {
                if (libtrace->format->ppause_input)
                        libtrace->format->ppause_input(libtrace);
                // TODO What happens if we don't have pause input??
        } else {
                printf("Trace is not parallel so we are doing a normal pause %s\n", libtrace->uridata);
                // This doesn't really work because this could be called by any thread
                // Maybe we should grab the lock here??
                if (libtrace->format->pause_input)
                        libtrace->format->pause_input(libtrace);
                // TODO What happens if we don't have pause input??
        }

        return 0;
}

/**
 * Stop trace finish prematurely as though it meet an EOF
 * This should only be called by the main thread
 * 1. Calls ppause
 * 2. Sends a message asking for threads to finish
 * 
 */
DLLEXPORT int trace_pstop(libtrace_t *libtrace)
{
        int i;
        if (!libtrace->started) {
                trace_set_err(libtrace,TRACE_ERR_BAD_STATE, "You must call trace_start() before calling trace_pstop()");
                return -1;
        }

        // Ensure all threads have paused and the underlying trace format has
        // been closed
        trace_ppause(libtrace);

        // Now send a message asking the threads to stop
        // This will be retrieved before trying to read another packet
        for (i = 0; i < libtrace->perpkt_thread_count; i++) {
                libtrace_message_t message;
                message.code = MESSAGE_STOP;
                message.additional = NULL;
                trace_send_message_to_thread(libtrace, &libtrace->perpkt_threads[i], &message);
        }

        // Now release the threads and let them stop
        assert(pthread_mutex_lock(&libtrace->libtrace_lock) == 0);
        libtrace->started = true;
        assert(pthread_cond_broadcast(&libtrace->perpkt_cond) == 0);
        assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);
        return 0;
}

/**
 * Set the hasher type along with a selected function, if hardware supports
 * that generic type of hashing it will be used otherwise the supplied
 * hasher function will be used and passed data when called.
 *
 * @return 0 if successful otherwise -1 on error
 */
DLLEXPORT int trace_set_hasher(libtrace_t *trace, enum hasher_types type, fn_hasher hasher, void *data) {
        int ret = -1;
        if (type == HASHER_HARDWARE || (type == HASHER_CUSTOM && !hasher) || (type == HASHER_BALANCE && hasher)) {
                return -1;
        }

        // Save the requirements
        trace->hasher_type = type;
        if (hasher) {
                trace->hasher = hasher;
                trace->hasher_data = hasher;
        } else {
                trace->hasher = NULL;
                // TODO consider how to handle freeing this
                trace->hasher_data = NULL;
        }

        // Try push this to hardware - NOTE hardware could do custom if
        // there is a more efficient way to apply it, in this case
        // it will simply grab the function out of libtrace_t
        if (trace->format->pconfig_input)
                ret = trace->format->pconfig_input(trace, TRACE_OPTION_SET_HASHER, &type);

        if (ret == -1) {
                // We have to deal with this ourself
                // This most likely means single threaded reading of the trace
                if (!hasher) {
                        switch (type)
                        {
                                case HASHER_CUSTOM:
                                case HASHER_BALANCE:
                                        return 0;
                                case HASHER_BIDIRECTIONAL:
                                        trace->hasher = toeplitz_hash_packet;
                                        trace->hasher_data = calloc(1, sizeof(toeplitz_conf_t));
                                        toeplitz_init_config(trace->hasher_data, 1);
                                        return 0;
                                case HASHER_UNIDIRECTIONAL:
                                        trace->hasher = toeplitz_hash_packet;
                                        trace->hasher_data = calloc(1, sizeof(toeplitz_conf_t));
                                        toeplitz_init_config(trace->hasher_data, 1);
                                        return 0;
                                case HASHER_HARDWARE:
                                        return -1;
                        }
                        return -1;
                }
        } else {
                // The hardware is dealing with this yay
                trace->hasher_type = HASHER_HARDWARE;
        }

        return 0;
}

// Waits for all threads to finish
DLLEXPORT void trace_join(libtrace_t *libtrace) {
        int i;

        /* Firstly wait for the perpkt threads to finish, since these are
         * user controlled */
        for (i=0; i< libtrace->perpkt_thread_count; i++) {
                //printf("Waiting to join with perpkt #%d\n", i);
                assert(pthread_join(libtrace->perpkt_threads[i].tid, NULL) == 0);
                //printf("Joined with perpkt #%d\n", i);
                // So we must do our best effort to empty the queue - so
                // the producer (or any other threads) don't block.
                libtrace_packet_t * packet;
                // Mark that we are no longer accepting packets
                assert(pthread_mutex_lock(&libtrace->libtrace_lock) == 0);
                libtrace->perpkt_threads[i].state = THREAD_FINISHED; // Important we are finished before we empty the buffer
                assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);
                while(libtrace_ringbuffer_try_read(&libtrace->perpkt_threads[i].rbuffer, (void **) &packet))
                        if (packet) // This could be NULL iff the perpkt finishes early
                                trace_destroy_packet(packet);
        }

        /* Now the hasher */
        // XXX signal it to stop
        if (trace_has_dedicated_hasher(libtrace)) {
                printf("Waiting to join with the hasher\n");
                pthread_join(libtrace->hasher_thread.tid, NULL);
                printf("Joined with with the hasher\n");
                libtrace->hasher_thread.state = THREAD_FINISHED;
        }

        // Now that everything is finished nothing can be touching our
        // buffers so clean them up
        for (i = 0; i < libtrace->perpkt_thread_count; i++) {
                // Its possible 1 packet got added by the reducer (or 1 per any other thread) since we cleaned up
                // if they lost timeslice before-during a write
                libtrace_packet_t * packet;
                while(libtrace_ringbuffer_try_read(&libtrace->perpkt_threads[i].rbuffer, (void **) &packet))
                        trace_destroy_packet(packet);
                if (libtrace->hasher) {
                        assert(libtrace_ringbuffer_is_empty(&libtrace->perpkt_threads[i].rbuffer));
                        libtrace_ringbuffer_destroy(&libtrace->perpkt_threads[i].rbuffer);
                }
                // Cannot destroy vector yet, this happens with trace_destroy
        }

        // Lets mark this as done for now
        libtrace->joined = true;
}

// Don't use extra overhead = :( directly place in storage structure using
// post
DLLEXPORT libtrace_result_t *trace_create_result()
{
        libtrace_result_t *result = malloc(sizeof(libtrace_result_t));
        assert(result);
        result->key = 0;
        result->value = NULL;
        // TODO automatically back with a free list!!
        return result;
}

DLLEXPORT int libtrace_thread_get_message_count(libtrace_t * libtrace)
{
        libtrace_thread_t * t = get_thread_descriptor(libtrace);
        assert(t);
        return libtrace_message_queue_count(&t->messages);
}

DLLEXPORT int libtrace_thread_get_message(libtrace_t * libtrace, libtrace_message_t * message)
{
        libtrace_thread_t * t = get_thread_descriptor(libtrace);
        assert(t);
        return libtrace_message_queue_get(&t->messages, message);
}

DLLEXPORT int libtrace_thread_try_get_message(libtrace_t * libtrace, libtrace_message_t * message)
{
        libtrace_thread_t * t = get_thread_descriptor(libtrace);
        assert(t);
        return libtrace_message_queue_try_get(&t->messages, message);
}

/**
 * Return backlog indicator
 */
DLLEXPORT int trace_post_reduce(libtrace_t *libtrace)
{
        libtrace_message_t message = {0};
        message.code = MESSAGE_POST_REDUCE;
        message.additional = NULL;
        message.sender = get_thread_descriptor(libtrace);
        return libtrace_message_queue_put(&libtrace->reducer_thread.messages, (void *) &message);
}

/**
 * Return backlog indicator
 */
DLLEXPORT int trace_send_message_to_reducer(libtrace_t * libtrace, libtrace_message_t * message)
{
        //printf("Sending message code=%d to reducer\n", message->code);
        message->sender = get_thread_descriptor(libtrace);
        return libtrace_message_queue_put(&libtrace->reducer_thread.messages, message);
}

/**
 *
 */
DLLEXPORT int trace_send_message_to_thread(libtrace_t * libtrace, libtrace_thread_t *t, libtrace_message_t * message)
{
        //printf("Sending message code=%d to reducer\n", message->code);
        message->sender = get_thread_descriptor(libtrace);
        return libtrace_message_queue_put(&t->messages, message);
}

DLLEXPORT void libtrace_result_set_key(libtrace_result_t * result, uint64_t key) {
        result->key = key;
}
DLLEXPORT uint64_t libtrace_result_get_key(libtrace_result_t * result) {
        return result->key;
}
DLLEXPORT void libtrace_result_set_value(libtrace_result_t * result, void * value) {
        result->value = value;
}
DLLEXPORT void* libtrace_result_get_value(libtrace_result_t * result) {
        return result->value;
}
DLLEXPORT void libtrace_result_set_key_value(libtrace_result_t * result, uint64_t key, void * value) {
        result->key = key;
        result->value = value;
}
DLLEXPORT void trace_destroy_result(libtrace_result_t ** result) {
        free(*result);
        result = NULL;
        // TODO automatically back with a free list!!
}

DLLEXPORT void * trace_get_global(libtrace_t *trace)
{
        return trace->global_blob;
}

DLLEXPORT void * trace_set_global(libtrace_t *trace, void * data)
{
        if (trace->global_blob && trace->global_blob != data) {
                void * ret = trace->global_blob;
                trace->global_blob = data;
                return ret;
        } else {
                trace->global_blob = data;
                return NULL;
        }
}

DLLEXPORT void * trace_get_tls(libtrace_thread_t *t)
{
        return t->user_data;
}

DLLEXPORT void * trace_set_tls(libtrace_thread_t *t, void * data)
{
        if(t->user_data && t->user_data != data) {
                void *ret = t->user_data;
                t->user_data = data;
                return ret;
        } else {
                t->user_data = data;
                return NULL;
        }
}

/**
 * Note: This function grabs a lock and expects trace_update_inprogress_result
 * to be called to release the lock.
 *
 * Expected to be used in trace-time situations to allow a result to be pending
 * a publish that can be taken by the reducer before publish if it wants to
 * publish a result. Such as publish a result every second but a queue hasn't
 * processed a packet (or is overloaded) and hasn't published yet.
 *
 * Currently this only supports a single temporary result,
 * as such if a key is different to the current temporary result the existing
 * one will be published and NULL returned.
 */
DLLEXPORT void * trace_retrive_inprogress_result(libtrace_t *libtrace, uint64_t key)
{
        int this_thread = get_thread_table_num(libtrace); // Could be worth caching ... ?
        libtrace_thread_t * t = &libtrace->perpkt_threads[this_thread];

        assert (pthread_spin_lock(&t->tmp_spinlock) == 0);
        if (t->tmp_key != key) {
                if (t->tmp_data) {
                        //printf("publising data key=%"PRIu64"\n", t->tmp_key);
                        trace_publish_result(libtrace, t->tmp_key, t->tmp_data);
                }
                t->tmp_data = NULL;
                t->tmp_key = key;
        }
        return t->tmp_data;
}

/**
 * Updates a temporary result and releases the lock previously grabbed by trace_retrive_inprogress_result
 */
DLLEXPORT void trace_update_inprogress_result(libtrace_t *libtrace, uint64_t key, void * value)
{
        int this_thread = get_thread_table_num(libtrace); // Could be worth caching ... ?
        libtrace_thread_t * t = &libtrace->perpkt_threads[this_thread];
        if (t->tmp_key != key) {
                if (t->tmp_data) {
                        printf("BAD publihsing data key=%"PRIu64"\n", t->tmp_key);
                        trace_publish_result(libtrace, t->tmp_key, t->tmp_data);
                }
                t->tmp_key = key;
        }
        t->tmp_data = value;
        assert (pthread_spin_unlock(&t->tmp_spinlock) == 0);
}

/**
 * Publish to the reduce queue, return
 */
DLLEXPORT void trace_publish_result(libtrace_t *libtrace, uint64_t key, void * value) {
        libtrace_result_t res;
        // Who am I???
        int this_thread = get_thread_table_num(libtrace); // Could be worth caching ... ?
        libtrace_thread_t * t = &libtrace->perpkt_threads[this_thread];
        // Now put it into my table
        static __thread int count = 0;


        libtrace_result_set_key_value(&res, key, value);
        /*
        if (count == 1)
                printf("My vector size is %d\n", libtrace_vector_get_size(&t->vector));
        count = (count+1) %1000;
        libtrace_vector_push_back(&t->vector, &res); // Automatically locking for us :)
        */
        /*if (count == 1)
                printf("My vector size is %d\n", libtrace_deque_get_size(&t->deque));
        count = (count+1)%1000;*/
        if (libtrace->reducer_flags & (REDUCE_SEQUENTIAL | REDUCE_ORDERED)) {
                if (libtrace_deque_get_size(&t->deque) >= 800) {
                        trace_post_reduce(libtrace);
                }
                //while (libtrace_deque_get_size(&t->deque) >= 1000)
                //      sched_yield();
                libtrace_deque_push_back(&t->deque, &res); // Automatically locking for us :)
        } else {
                //while (libtrace_vector_get_size(&t->vector) >= 1000)
                //      sched_yield();

                if (libtrace_deque_get_size(&t->deque) >= 800) {
                        trace_post_reduce(libtrace);
                }
                libtrace_vector_push_back(&t->vector, &res); // Automatically locking for us :)
        }
}


static int compareres(const void* p1, const void* p2)
{
        if (libtrace_result_get_key((libtrace_result_t *) p1) < libtrace_result_get_key((libtrace_result_t *) p2))
                return -1;
        if (libtrace_result_get_key((libtrace_result_t *) p1) == libtrace_result_get_key((libtrace_result_t *) p2))
                return 0;
        else
                return 1;
}

/* Retrieves all results with the key requested from the temporary result
 * holding zone.
 */
DLLEXPORT int trace_get_results_check_temp(libtrace_t *libtrace, libtrace_vector_t *results, uint64_t key)
{
        int i;

        libtrace_vector_empty(results);
        // Check all of the temp queues
        for (i = 0; i < libtrace->perpkt_thread_count; ++i) {
                libtrace_result_t r = {0,0};
                assert (pthread_spin_lock(&libtrace->perpkt_threads[i].tmp_spinlock) == 0);
                if (libtrace->perpkt_threads[i].tmp_key == key) {
                        libtrace_result_set_key_value(&r, key, libtrace->perpkt_threads[i].tmp_data);
                        libtrace->perpkt_threads[i].tmp_data = NULL;
                        printf("Found in temp queue\n");
                }
                assert (pthread_spin_unlock(&libtrace->perpkt_threads[i].tmp_spinlock) == 0);
                if (libtrace_result_get_value(&r)) {
                        // Got a result still in temporary
                        printf("Publishing from temp queue\n");
                        libtrace_vector_push_back(results, &r);
                } else {
                        // This might be waiting on the actual queue
                        libtrace_queue_t *v = &libtrace->perpkt_threads[i].deque;
                        if (libtrace_deque_peek_front(v, (void *) &r) &&
                                        libtrace_result_get_value(&r)) {
                                assert (libtrace_deque_pop_front(&libtrace->perpkt_threads[i].deque, (void *) &r) == 1);
                                printf("Found in real queue\n");
                                libtrace_vector_push_back(results, &r);
                        } // else no data (probably means no packets)
                        else {
                                printf("Result missing in real queue\n");
                        }
                }
        }
        //printf("Loop done yo, that means we've got #%d results to print fool!\n", libtrace_vector_get_size(results));
        return libtrace_vector_get_size(results);
}

DLLEXPORT int trace_get_results(libtrace_t *libtrace, libtrace_vector_t * results) {
        int i;
        int flags = libtrace->reducer_flags; // Hint these aren't a changing

        libtrace_vector_empty(results);

        /* Here we assume queues are in order ascending order and they want
         * the smallest result first. If they are not in order the results
         * may not be in order.
         */
        if (flags & (REDUCE_SEQUENTIAL | REDUCE_ORDERED)) {
                int live_count = 0;
                bool live[libtrace->perpkt_thread_count]; // Set if a trace is alive
                uint64_t key[libtrace->perpkt_thread_count]; // Cached keys
                uint64_t min_key = UINT64_MAX; // XXX use max int here stdlimit.h?
                int min_queue = -1;

                /* Loop through check all are alive (have data) and find the smallest */
                for (i = 0; i < libtrace->perpkt_thread_count; ++i) {
                        libtrace_queue_t *v = &libtrace->perpkt_threads[i].deque;
                        if (libtrace_deque_get_size(v) != 0) {
                                libtrace_result_t r;
                                libtrace_deque_peek_front(v, (void *) &r);
                                live_count++;
                                live[i] = 1;
                                key[i] = libtrace_result_get_key(&r);
                                if (i==0 || min_key > key[i]) {
                                        min_key = key[i];
                                        min_queue = i;
                                }
                        } else {
                                live[i] = 0;
                        }
                }

                /* Now remove the smallest and loop - special case if all threads have joined we always flush whats left */
                while ((live_count == libtrace->perpkt_thread_count) || (live_count &&
                                ((flags & REDUCE_SEQUENTIAL && min_key == libtrace->expected_key) ||
                                libtrace->joined))) {
                        /* Get the minimum queue and then do stuff */
                        libtrace_result_t r;

                        assert (libtrace_deque_pop_front(&libtrace->perpkt_threads[min_queue].deque, (void *) &r) == 1);
                        libtrace_vector_push_back(results, &r);

                        // We expect the key we read +1 now
                        libtrace->expected_key = key[min_queue] + 1;

                        // Now update the one we just removed
                        if (libtrace_deque_get_size(&libtrace->perpkt_threads[min_queue].deque) )
                        {
                                libtrace_deque_peek_front(&libtrace->perpkt_threads[min_queue].deque, (void *) &r);
                                key[min_queue] = libtrace_result_get_key(&r);
                                if (key[min_queue] <= min_key) {
                                        // We are still the smallest, might be out of order though :(
                                        min_key = key[min_queue];
                                } else {
                                        min_key = key[min_queue]; // Update our minimum
                                        // Check all find the smallest again - all are alive
                                        for (i = 0; i < libtrace->perpkt_thread_count; ++i) {
                                                if (live[i] && min_key > key[i]) {
                                                        min_key = key[i];
                                                        min_queue = i;
                                                }
                                        }
                                }
                        } else {
                                live[min_queue] = 0;
                                live_count--;
                                min_key = UINT64_MAX; // Update our minimum
                                // Check all find the smallest again - all are alive
                                for (i = 0; i < libtrace->perpkt_thread_count; ++i) {
                                        // Still not 100% TODO (what if order is wrong or not increasing)
                                        if (live[i] && min_key >= key[i]) {
                                                min_key = key[i];
                                                min_queue = i;
                                        }
                                }
                        }
                }
        } else { // Queues are not in order - return all results in the queue
                for (i = 0; i < libtrace->perpkt_thread_count; i++) {
                        libtrace_vector_append(results, &libtrace->perpkt_threads[i].vector);
                }
                if (flags & REDUCE_SORT) {
                        qsort(results->elements, results->size, results->element_size, &compareres);
                }
        }
        return libtrace_vector_get_size(results);
}

DLLEXPORT uint64_t trace_packet_get_order(libtrace_packet_t * packet) {
        return packet->order;
}

DLLEXPORT uint64_t trace_packet_get_hash(libtrace_packet_t * packet) {
        return packet->hash;
}

DLLEXPORT void trace_packet_set_order(libtrace_packet_t * packet, uint64_t order) {
        packet->order = order;
}

DLLEXPORT void trace_packet_set_hash(libtrace_packet_t * packet, uint64_t hash) {
        packet->hash = hash;
}

DLLEXPORT int trace_finished(libtrace_t * libtrace) {
        int i;
        int b = 0;
        // TODO I don't like using this so much
        //assert(pthread_mutex_lock(&libtrace->libtrace_lock) == 0);
        for (i = 0; i < libtrace->perpkt_thread_count; i++) {
                if (libtrace->perpkt_threads[i].state == THREAD_RUNNING)
                        b++;
        }
        //assert(pthread_mutex_unlock(&libtrace->libtrace_lock) == 0);
        return !b;
}

DLLEXPORT int trace_parallel_config(libtrace_t *libtrace, trace_parallel_option_t option, void *value)
{
        int ret = -1;
        switch (option) {
                case TRACE_OPTION_SET_HASHER:
                        return trace_set_hasher(libtrace, (enum hasher_types) *((int *) value), NULL, NULL);
                case TRACE_OPTION_SET_PERPKT_BUFFER_SIZE:
                        libtrace->perpkt_buffer_size = *((int *) value);
                        return 1;
                case TRACE_OPTION_SET_PACKET_FREELIST_SIZE:
                        libtrace->packet_freelist_size = *((int *) value);
                        return 1;
                case TRACE_OPTION_SET_PERPKT_THREAD_COUNT:
                        libtrace->perpkt_thread_count = *((int *) value);
                        return 1;
                case TRACE_DROP_OUT_OF_ORDER:
                        if (*((int *) value))
                                libtrace->reducer_flags |= REDUCE_DROP_OOO;
                        else
                                libtrace->reducer_flags &= ~REDUCE_DROP_OOO;
                        return 1;
                case TRACE_OPTION_SEQUENTIAL:
                        if (*((int *) value))
                                libtrace->reducer_flags |= REDUCE_SEQUENTIAL;
                        else
                                libtrace->reducer_flags &= ~REDUCE_SEQUENTIAL;
                        return 1;
                case TRACE_OPTION_ORDERED:
                        if (*((int *) value))
                                libtrace->reducer_flags |= REDUCE_ORDERED;
                        else
                                libtrace->reducer_flags &= ~REDUCE_ORDERED;
                        return 1;
                case TRACE_OPTION_USE_DEDICATED_HASHER:
                        if (*((int *) value))
                                libtrace->hasher_thread.type = THREAD_HASHER;
                        else
                                libtrace->hasher_thread.type = THREAD_EMPTY;
                        return 1;
                case TRACE_OPTION_USE_SLIDING_WINDOW_BUFFER:
                        if (*((int *) value))
                                libtrace->reducer_flags |= PERPKT_USE_SLIDING_WINDOW;
                        else
                                libtrace->reducer_flags &= ~PERPKT_USE_SLIDING_WINDOW;
                        return 1;
                case TRACE_OPTION_TRACETIME:
                        if(*((int *) value))
                                libtrace->tracetime = 1;
                        else
                                libtrace->tracetime = 0;
                        return 0;
        }
        return 0;
}

DLLEXPORT libtrace_packet_t* trace_result_packet(libtrace_t * libtrace, libtrace_packet_t * packet) {
        libtrace_packet_t* result;
        if (!libtrace_ringbuffer_try_sread_bl(&libtrace->packet_freelist, (void **) &result))
                result = trace_create_packet();
        assert(result);
        swap_packets(result, packet); // Move the current packet into our copy
        return result;
}

DLLEXPORT void trace_free_result_packet(libtrace_t * libtrace, libtrace_packet_t * packet) {
        // Try write back the packet
        assert(packet);
        // Always release any resources this might be holding such as a slot in a ringbuffer
        trace_fin_packet(packet);
        if (!libtrace_ringbuffer_try_swrite_bl(&libtrace->packet_freelist, packet)) {
                /* We couldn't, oh well lets just destroy it - XXX consider non managed formats i.e. rings buffers loosing packets and jamming up :( */
                //assert(1 == 90);
                trace_destroy_packet(packet);
        }
}