xref: /freebsd/tools/tools/netmap/nmreplay.c (revision bdd1243d)

/*
 * Copyright (C) 2016 Universita` di Pisa. All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 *   1. Redistributions of source code must retain the above copyright
 *      notice, this list of conditions and the following disclaimer.
 *   2. Redistributions in binary form must reproduce the above copyright
 *      notice, this list of conditions and the following disclaimer in the
 *    documentation and/or other materials provided with the distribution.
 *
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
 * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
 * SUCH DAMAGE.
 *
 * $FreeBSD$
 */


/*
 * This program implements NMREPLAY, a program to replay a pcap file
 * enforcing the output rate and possibly random losses and delay
 * distributions.
 * It is meant to be run from the command line and implemented with a main
 * control thread for monitoring, plus a thread to push packets out.
 *
 * The control thread parses command line arguments, prepares a
 * schedule for transmission in a memory buffer and then sits
 * in a loop where it periodically reads traffic statistics from
 * the other threads and prints them out on the console.
 *
 * The transmit buffer contains headers and packets. Each header
 * includes a timestamp that determines when the packet should be sent out.
 * A "consumer" thread cons() reads from the queue and transmits packets
 * on the output netmap port when their time has come.
 *
 * The program does CPU pinning and sets the scheduler and priority
 * for the "cons" threads. Externally one should do the
 * assignment of other threads (e.g. interrupt handlers) and
 * make sure that network interfaces are configured properly.
 *
 * --- Main functions of the program ---
 * within each function, q is used as a pointer to the queue holding
 * packets and parameters.
 *
 * pcap_prod()
 *
 *	reads from the pcap file and prepares packets to transmit.
 *	After reading a packet from the pcap file, the following information
 *	are extracted which can be used to determine the schedule:
 *
 *   	q->cur_pkt	points to the buffer containing the packet
 *	q->cur_len	packet length, excluding CRC
 *	q->cur_caplen	available packet length (may be shorter than cur_len)
 *	q->cur_tt	transmission time for the packet, computed from the trace.
 *
 *  The following functions are then called in sequence:
 *
 *  q->c_loss (set with the -L command line option) decides
 *	whether the packet should be dropped before even queuing.
 *	This is generally useful to emulate random loss.
 *	The function is supposed to set q->c_drop = 1 if the
 *	packet should be dropped, or leave it to 0 otherwise.
 *
 *   q->c_bw (set with the -B command line option) is used to
 *      enforce the transmit bandwidth. The function must store
 *	in q->cur_tt the transmission time (in nanoseconds) of
 *	the packet, which is typically proportional to the length
 *	of the packet, i.e. q->cur_tt = q->cur_len / <bandwidth>
 *	Variants are possible, eg. to account for constant framing
 *	bits as on the ethernet, or variable channel acquisition times,
 *	etc.
 *	This mechanism can also be used to simulate variable queueing
 *	delay e.g. due to the presence of cross traffic.
 *
 *   q->c_delay (set with the -D option) implements delay emulation.
 *	The function should set q->cur_delay to the additional
 *	delay the packet is subject to. The framework will take care of
 *	computing the actual exit time of a packet so that there is no
 *	reordering.
 */

// debugging macros
#define NED(_fmt, ...)	do {} while (0)
#define ED(_fmt, ...)						\
	do {							\
		struct timeval _t0;				\
		gettimeofday(&_t0, NULL);			\
		fprintf(stderr, "%03d.%03d %-10.10s [%5d] \t" _fmt "\n", \
		(int)(_t0.tv_sec % 1000), (int)_t0.tv_usec/1000, \
		__FUNCTION__, __LINE__, ##__VA_ARGS__);     \
	} while (0)

/* WWW is for warnings, EEE is for errors */
#define WWW(_fmt, ...)	ED("--WWW-- " _fmt, ##__VA_ARGS__)
#define EEE(_fmt, ...)	ED("--EEE-- " _fmt, ##__VA_ARGS__)
#define DDD(_fmt, ...)	ED("--DDD-- " _fmt, ##__VA_ARGS__)

#define _GNU_SOURCE	// for CPU_SET() etc
#include <errno.h>
#include <fcntl.h>
#include <libnetmap.h>
#include <math.h> /* log, exp etc. */
#include <pthread.h>
#ifdef __FreeBSD__
#include <pthread_np.h> /* pthread w/ affinity */
#include <sys/cpuset.h> /* cpu_set */
#endif /* __FreeBSD__ */
#include <signal.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h> /* memcpy */
#include <stdint.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include <sys/poll.h>
#include <sys/resource.h> // setpriority
#include <sys/time.h>
#include <unistd.h>

/*
 *
 * A packet in the queue is q_pkt plus the payload.
 *
 * For the packet descriptor we need the following:
 *
 *  -	position of next packet in the queue (can go backwards).
 *	We can reduce to 32 bits if we consider alignments,
 *	or we just store the length to be added to the current
 *	value and assume 0 as a special index.
 *  -	actual packet length (16 bits may be ok)
 *  -	queue output time, in nanoseconds (64 bits)
 *  -	delay line output time, in nanoseconds
 *	One of the two can be packed to a 32bit value
 *
 * A convenient coding uses 32 bytes per packet.
 */

struct q_pkt {
	uint64_t	next;		/* buffer index for next packet */
	uint64_t	pktlen;		/* actual packet len */
	uint64_t	pt_qout;	/* time of output from queue */
	uint64_t	pt_tx;		/* transmit time */
};


/*
 * The header for a pcap file
 */
struct pcap_file_header {
    uint32_t magic;
	/*used to detect the file format itself and the byte
    ordering. The writing application writes 0xa1b2c3d4 with it's native byte
    ordering format into this field. The reading application will read either
    0xa1b2c3d4 (identical) or 0xd4c3b2a1 (swapped). If the reading application
    reads the swapped 0xd4c3b2a1 value, it knows that all the following fields
    will have to be swapped too. For nanosecond-resolution files, the writing
    application writes 0xa1b23c4d, with the two nibbles of the two lower-order
    bytes swapped, and the reading application will read either 0xa1b23c4d
    (identical) or 0x4d3cb2a1 (swapped)*/
    uint16_t version_major;
    uint16_t version_minor; /*the version number of this file format */
    int32_t thiszone;
	/*the correction time in seconds between GMT (UTC) and the
    local timezone of the following packet header timestamps. Examples: If the
    timestamps are in GMT (UTC), thiszone is simply 0. If the timestamps are in
    Central European time (Amsterdam, Berlin, ...) which is GMT + 1:00, thiszone
    must be -3600*/
    uint32_t stampacc; /*the accuracy of time stamps in the capture*/
    uint32_t snaplen;
	/*the "snapshot length" for the capture (typically 65535
    or even more, but might be limited by the user)*/
    uint32_t network;
	/*link-layer header type, specifying the type of headers
    at the beginning of the packet (e.g. 1 for Ethernet); this can be various
    types such as 802.11, 802.11 with various radio information, PPP, Token
    Ring, FDDI, etc.*/
};

#if 0 /* from pcap.h */
struct pcap_file_header {
        bpf_u_int32 magic;
        u_short version_major;
        u_short version_minor;
        bpf_int32 thiszone;     /* gmt to local correction */
        bpf_u_int32 sigfigs;    /* accuracy of timestamps */
        bpf_u_int32 snaplen;    /* max length saved portion of each pkt */
        bpf_u_int32 linktype;   /* data link type (LINKTYPE_*) */
};

struct pcap_pkthdr {
        struct timeval ts;      /* time stamp */
        bpf_u_int32 caplen;     /* length of portion present */
        bpf_u_int32 len;        /* length this packet (off wire) */
};
#endif /* from pcap.h */

struct pcap_pkthdr {
    uint32_t ts_sec; /* seconds from epoch */
    uint32_t ts_frac; /* microseconds or nanoseconds depending on sigfigs */
    uint32_t caplen;
	/*the number of bytes of packet data actually captured
    and saved in the file. This value should never become larger than orig_len
    or the snaplen value of the global header*/
    uint32_t len;	/* wire length */
};


#define PKT_PAD         (32)    /* padding on packets */

static inline int pad(int x)
{
        return ((x) + PKT_PAD - 1) & ~(PKT_PAD - 1) ;
}



/*
 * wrapper around the pcap file.
 * We mmap the file so it is easy to do multiple passes through it.
 */
struct nm_pcap_file {
    int fd;
    uint64_t filesize;
    const char *data; /* mmapped file */

    uint64_t tot_pkt;
    uint64_t tot_bytes;
    uint64_t tot_bytes_rounded;	/* need hdr + pad(len) */
    uint32_t resolution; /* 1000 for us, 1 for ns */
    int swap; /* need to swap fields ? */

    uint64_t first_ts;
    uint64_t total_tx_time;
	/*
	 * total_tx_time is computed as last_ts - first_ts, plus the
	 * transmission time for the first packet which in turn is
	 * computed according to the average bandwidth
	 */

    uint64_t file_len;
    const char *cur;	/* running pointer */
    const char *lim;	/* data + file_len */
    int err;
};

static struct nm_pcap_file *readpcap(const char *fn);
static void destroy_pcap(struct nm_pcap_file *file);


#define NS_SCALE 1000000000UL	/* nanoseconds in 1s */

static void destroy_pcap(struct nm_pcap_file *pf)
{
    if (!pf)
	return;

    munmap((void *)(uintptr_t)pf->data, pf->filesize);
    close(pf->fd);
    bzero(pf, sizeof(*pf));
    free(pf);
    return;
}

// convert a field of given size if swap is needed.
static uint32_t
cvt(const void *src, int size, char swap)
{
    uint32_t ret = 0;
    if (size != 2 && size != 4) {
	EEE("Invalid size %d\n", size);
	exit(1);
    }
    memcpy(&ret, src, size);
    if (swap) {
	unsigned char tmp, *data = (unsigned char *)&ret;
	int i;
        for (i = 0; i < size / 2; i++) {
            tmp = data[i];
            data[i] = data[size - (1 + i)];
            data[size - (1 + i)] = tmp;
        }
    }
    return ret;
}

static uint32_t
read_next_info(struct nm_pcap_file *pf, int size)
{
    const char *end = pf->cur + size;
    uint32_t ret;
    if (end > pf->lim) {
	pf->err = 1;
	ret = 0;
    } else {
	ret = cvt(pf->cur, size, pf->swap);
	pf->cur = end;
    }
    return ret;
}

/*
 * mmap the file, make sure timestamps are sorted, and count
 * packets and sizes
 * Timestamps represent the receive time of the packets.
 * We need to compute also the 'first_ts' which refers to a hypotetical
 * packet right before the first one, see the code for details.
 */
static struct nm_pcap_file *
readpcap(const char *fn)
{
    struct nm_pcap_file _f, *pf = &_f;
    uint64_t prev_ts, first_pkt_time;
    uint32_t magic, first_len = 0;

    bzero(pf, sizeof(*pf));
    pf->fd = open(fn, O_RDONLY);
    if (pf->fd < 0) {
	EEE("cannot open file %s", fn);
	return NULL;
    }
    /* compute length */
    pf->filesize = lseek(pf->fd, 0, SEEK_END);
    lseek(pf->fd, 0, SEEK_SET);
    ED("filesize is %lu", (u_long)(pf->filesize));
    if (pf->filesize < sizeof(struct pcap_file_header)) {
	EEE("file too short %s", fn);
	close(pf->fd);
	return NULL;
    }
    pf->data = mmap(NULL, pf->filesize, PROT_READ, MAP_SHARED, pf->fd, 0);
    if (pf->data == MAP_FAILED) {
	EEE("cannot mmap file %s", fn);
	close(pf->fd);
	return NULL;
    }
    pf->cur = pf->data;
    pf->lim = pf->data + pf->filesize;
    pf->err = 0;
    pf->swap = 0; /* default, same endianness when read magic */

    magic = read_next_info(pf, 4);
    ED("magic is 0x%x", magic);
    switch (magic) {
    case 0xa1b2c3d4: /* native, us resolution */
	pf->swap = 0;
	pf->resolution = 1000;
	break;
    case 0xd4c3b2a1: /* swapped, us resolution */
	pf->swap = 1;
	pf->resolution = 1000;
	break;
    case 0xa1b23c4d:	/* native, ns resolution */
	pf->swap = 0;
	pf->resolution = 1; /* nanoseconds */
	break;
    case 0x4d3cb2a1:	/* swapped, ns resolution */
	pf->swap = 1;
	pf->resolution = 1; /* nanoseconds */
	break;
    default:
	EEE("unknown magic 0x%x", magic);
	return NULL;
    }

    ED("swap %d res %d\n", pf->swap, pf->resolution);
    pf->cur = pf->data + sizeof(struct pcap_file_header);
    pf->lim = pf->data + pf->filesize;
    pf->err = 0;
    prev_ts = 0;
    while (pf->cur < pf->lim && pf->err == 0) {
	uint32_t base = pf->cur - pf->data;
	uint64_t cur_ts = read_next_info(pf, 4) * NS_SCALE +
		read_next_info(pf, 4) * pf->resolution;
	uint32_t caplen = read_next_info(pf, 4);
	uint32_t len = read_next_info(pf, 4);

	if (pf->err) {
	    WWW("end of pcap file after %d packets\n",
		(int)pf->tot_pkt);
	    break;
	}
	if  (cur_ts < prev_ts) {
	    WWW("reordered packet %d\n",
		(int)pf->tot_pkt);
	}
	prev_ts = cur_ts;
	(void)base;
	if (pf->tot_pkt == 0) {
	    pf->first_ts = cur_ts;
	    first_len = len;
	}
	pf->tot_pkt++;
	pf->tot_bytes += len;
	pf->tot_bytes_rounded += pad(len) + sizeof(struct q_pkt);
	pf->cur += caplen;
    }
    pf->total_tx_time = prev_ts - pf->first_ts; /* excluding first packet */
    ED("tot_pkt %lu tot_bytes %lu tx_time %.6f s first_len %lu",
	(u_long)pf->tot_pkt, (u_long)pf->tot_bytes,
	1e-9*pf->total_tx_time, (u_long)first_len);
    /*
     * We determine that based on the
     * average bandwidth of the trace, as follows
     *   first_pkt_ts = p[0].len / avg_bw
     * In turn avg_bw = (total_len - p[0].len)/(p[n-1].ts - p[0].ts)
     * so
     *   first_ts =  p[0].ts - p[0].len * (p[n-1].ts - p[0].ts) / (total_len - p[0].len)
     */
    if (pf->tot_bytes == first_len) {
	/* cannot estimate bandwidth, so force 1 Gbit */
	first_pkt_time = first_len * 8; /* * 10^9 / bw */
    } else {
	first_pkt_time = pf->total_tx_time * first_len / (pf->tot_bytes - first_len);
    }
    ED("first_pkt_time %.6f s", 1e-9*first_pkt_time);
    pf->total_tx_time += first_pkt_time;
    pf->first_ts -= first_pkt_time;

    /* all correct, allocate a record and copy */
    pf = calloc(1, sizeof(*pf));
    *pf = _f;
    /* reset pointer to start */
    pf->cur = pf->data + sizeof(struct pcap_file_header);
    pf->err = 0;
    return pf;
}

enum my_pcap_mode { PM_NONE, PM_FAST, PM_FIXED, PM_REAL };

static int verbose = 0;

static int do_abort = 0;

#ifdef linux
#define cpuset_t        cpu_set_t
#endif

#ifdef __APPLE__
#define cpuset_t        uint64_t        // XXX
static inline void CPU_ZERO(cpuset_t *p)
{
        *p = 0;
}

static inline void CPU_SET(uint32_t i, cpuset_t *p)
{
        *p |= 1<< (i & 0x3f);
}

#define pthread_setaffinity_np(a, b, c) ((void)a, 0)
#define sched_setscheduler(a, b, c)	(1) /* error */
#define clock_gettime(a,b)      \
        do {struct timespec t0 = {0,0}; *(b) = t0; } while (0)

#define	_P64	unsigned long
#endif

#ifndef _P64

/* we use uint64_t widely, but printf gives trouble on different
 * platforms so we use _P64 as a cast
 */
#define	_P64	uint64_t
#endif /* print stuff */


struct _qs;	/* forward */
/*
 * descriptor of a configuration entry.
 * Each handler has a parse function which takes ac/av[] and returns
 * true if successful. Any allocated space is stored into struct _cfg *
 * that is passed as argument.
 * arg and arg_len are included for convenience.
 */
struct _cfg {
    int (*parse)(struct _qs *, struct _cfg *, int ac, char *av[]);  /* 0 ok, 1 on error */
    int (*run)(struct _qs *, struct _cfg *arg);         /* 0 Ok, 1 on error */
    // int close(struct _qs *, void *arg);              /* 0 Ok, 1 on error */

    const char *optarg;	/* command line argument. Initial value is the error message */
    /* placeholders for common values */
    void *arg;		/* allocated memory if any */
    int arg_len;	/* size of *arg in case a realloc is needed */
    uint64_t d[16];	/* static storage for simple cases */
    double f[4];	/* static storage for simple cases */
};


/*
 * communication occurs through this data structure, with fields
 * cache-aligned according to who are the readers/writers.
 *

The queue is an array of memory  (buf) of size buflen (does not change).

The producer uses 'tail' as an index in the queue to indicate
the first empty location (ie. after the last byte of data),
the consumer uses head to indicate the next byte to consume.

For best performance we should align buffers and packets
to multiples of cacheline, but this would explode memory too much.
Worst case memory explosion is with 65 byte packets.
Memory usage as shown below:

		qpkt-pad
	size	32-16	32-32	32-64	64-64

	64	96	96	96	128
	65	112	128	160	192


An empty queue has head == tail, a full queue will have free space
below a threshold.  In our case the queue is large enough and we
are non blocking so we can simply drop traffic when the queue
approaches a full state.

To simulate bandwidth limitations efficiently, the producer has a second
pointer, prod_tail_1, used to check for expired packets. This is done lazily.

 */
/*
 * When sizing the buffer, we must assume some value for the bandwidth.
 * INFINITE_BW is supposed to be faster than what we support
 */
#define INFINITE_BW	(200ULL*1000000*1000)
#define	MY_CACHELINE	(128ULL)
#define MAX_PKT		(9200)	/* max packet size */

#define ALIGN_CACHE	__attribute__ ((aligned (MY_CACHELINE)))

struct _qs { /* shared queue */
	uint64_t	t0;	/* start of times */

	uint64_t 	buflen;	/* queue length */
	char *buf;

	/* handlers for various options */
	struct _cfg	c_delay;
	struct _cfg	c_bw;
	struct _cfg	c_loss;

	/* producer's fields */
	uint64_t	tx ALIGN_CACHE;	/* tx counter */
	uint64_t	prod_tail_1;	/* head of queue */
	uint64_t	prod_head;	/* cached copy */
	uint64_t	prod_tail;	/* cached copy */
	uint64_t	prod_drop;	/* drop packet count */
	uint64_t	prod_max_gap;	/* rx round duration */

	struct nm_pcap_file	*pcap;		/* the pcap struct */

	/* parameters for reading from the netmap port */
	struct nmport_d *src_port;		/* netmap descriptor */
	const char *	prod_ifname;	/* interface name or pcap file */
	struct netmap_ring *rxring;	/* current ring being handled */
	uint32_t	si;		/* ring index */
	int		burst;
	uint32_t	rx_qmax;	/* stats on max queued */

	uint64_t	qt_qout;	/* queue exit time for last packet */
		/*
		 * when doing shaping, the software computes and stores here
		 * the time when the most recently queued packet will exit from
		 * the queue.
		 */

	uint64_t	qt_tx;		/* delay line exit time for last packet */
		/*
		 * The software computes the time at which the most recently
		 * queued packet exits from the queue.
		 * To avoid reordering, the next packet should exit at least
		 * at qt_tx + cur_tt
		 */

	/* producer's fields controlling the queueing */
	const char *	cur_pkt;	/* current packet being analysed */
	uint32_t	cur_len;	/* length of current packet */
	uint32_t	cur_caplen;	/* captured length of current packet */

	int		cur_drop;	/* 1 if current  packet should be dropped. */
		/*
		 * cur_drop can be set as a result of the loss emulation,
		 * and may need to use the packet size, current time, etc.
		 */

	uint64_t	cur_tt;		/* transmission time (ns) for current packet */
		/*
		 * The transmission time is how much link time the packet will consume.
		 * should be set by the function that does the bandwidth emulation,
		 * but could also be the result of a function that emulates the
		 * presence of competing traffic, MAC protocols etc.
		 * cur_tt is 0 for links with infinite bandwidth.
		 */

	uint64_t	cur_delay;	/* delay (ns) for current packet from c_delay.run() */
		/*
		 * this should be set by the function that computes the extra delay
		 * applied to the packet.
		 * The code makes sure that there is no reordering and possibly
		 * bumps the output time as needed.
		 */


	/* consumer's fields */
	const char *		cons_ifname;
	uint64_t rx ALIGN_CACHE;	/* rx counter */
	uint64_t	cons_head;	/* cached copy */
	uint64_t	cons_tail;	/* cached copy */
	uint64_t	cons_now;	/* most recent producer timestamp */
	uint64_t	rx_wait;	/* stats */

	/* shared fields */
	volatile uint64_t _tail ALIGN_CACHE ;	/* producer writes here */
	volatile uint64_t _head ALIGN_CACHE ;	/* consumer reads from here */
};

struct pipe_args {
	int		wait_link;

	pthread_t	cons_tid;	/* main thread */
	pthread_t	prod_tid;	/* producer thread */

	/* Affinity: */
	int		cons_core;	/* core for cons() */
	int		prod_core;	/* core for prod() */

	struct nmport_d *pa;		/* netmap descriptor */
	struct nmport_d *pb;

	struct _qs	q;
};

#define NS_IN_S	(1000000000ULL)	// nanoseconds
#define TIME_UNITS	NS_IN_S
/* set the thread affinity. */
static int
setaffinity(int i)
{
        cpuset_t cpumask;
	struct sched_param p;

        if (i == -1)
                return 0;

        /* Set thread affinity affinity.*/
        CPU_ZERO(&cpumask);
        CPU_SET(i, &cpumask);

        if (pthread_setaffinity_np(pthread_self(), sizeof(cpuset_t), &cpumask) != 0) {
                WWW("Unable to set affinity: %s", strerror(errno));
        }
	if (setpriority(PRIO_PROCESS, 0, -10)) {; // XXX not meaningful
                WWW("Unable to set priority: %s", strerror(errno));
	}
	bzero(&p, sizeof(p));
	p.sched_priority = 10; // 99 on linux ?
	// use SCHED_RR or SCHED_FIFO
	if (sched_setscheduler(0, SCHED_RR, &p)) {
                WWW("Unable to set scheduler: %s", strerror(errno));
	}
        return 0;
}


/*
 * set the timestamp from the clock, subtract t0
 */
static inline void
set_tns_now(uint64_t *now, uint64_t t0)
{
    struct timespec t;

    clock_gettime(CLOCK_REALTIME, &t); // XXX precise on FreeBSD ?
    *now = (uint64_t)(t.tv_nsec + NS_IN_S * t.tv_sec);
    *now -= t0;
}



/* compare two timestamps */
static inline int64_t
ts_cmp(uint64_t a, uint64_t b)
{
	return (int64_t)(a - b);
}

/* create a packet descriptor */
static inline struct q_pkt *
pkt_at(struct _qs *q, uint64_t ofs)
{
    return (struct q_pkt *)(q->buf + ofs);
}


/*
 * we have already checked for room and prepared p->next
 */
static inline int
enq(struct _qs *q)
{
    struct q_pkt *p = pkt_at(q, q->prod_tail);

    /* hopefully prefetch has been done ahead */
    nm_pkt_copy(q->cur_pkt, (char *)(p+1), q->cur_caplen);
    p->pktlen = q->cur_len;
    p->pt_qout = q->qt_qout;
    p->pt_tx = q->qt_tx;
    p->next = q->prod_tail + pad(q->cur_len) + sizeof(struct q_pkt);
    ND("enqueue len %d at %d new tail %ld qout %.6f tx %.6f",
        q->cur_len, (int)q->prod_tail, p->next,
        1e-9*p->pt_qout, 1e-9*p->pt_tx);
    q->prod_tail = p->next;
    q->tx++;
    return 0;
}

/*
 * simple handler for parameters not supplied
 */
static int
null_run_fn(struct _qs *q, struct _cfg *cfg)
{
    (void)q;
    (void)cfg;
    return 0;
}



/*
 * put packet data into the buffer.
 * We read from the mmapped pcap file, construct header, copy
 * the captured length of the packet and pad with zeroes.
 */
static void *
pcap_prod(void *_pa)
{
    struct pipe_args *pa = _pa;
    struct _qs *q = &pa->q;
    struct nm_pcap_file *pf = q->pcap;	/* already opened by readpcap */
    uint32_t loops, i, tot_pkts;

    /* data plus the loop record */
    uint64_t need;
    uint64_t t_tx, tt, last_ts; /* last timestamp from trace */

    /*
     * For speed we make sure the trace is at least some 1000 packets,
     * so we may need to loop the trace more than once (for short traces)
     */
    loops = (1 + 10000 / pf->tot_pkt);
    tot_pkts = loops * pf->tot_pkt;
    need = loops * pf->tot_bytes_rounded + sizeof(struct q_pkt);
    q->buf = calloc(1, need);
    if (q->buf == NULL) {
	D("alloc %lld bytes for queue failed, exiting",(long long)need);
	goto fail;
    }
    q->prod_head = q->prod_tail = 0;
    q->buflen = need;

    pf->cur = pf->data + sizeof(struct pcap_file_header);
    pf->err = 0;

    ED("--- start create %lu packets at tail %d",
	(u_long)tot_pkts, (int)q->prod_tail);
    last_ts = pf->first_ts; /* beginning of the trace */

    q->qt_qout = 0; /* first packet out of the queue */

    for (loops = 0, i = 0; i < tot_pkts && !do_abort; i++) {
	const char *next_pkt; /* in the pcap buffer */
	uint64_t cur_ts;

	/* read values from the pcap buffer */
	cur_ts = read_next_info(pf, 4) * NS_SCALE +
		read_next_info(pf, 4) * pf->resolution;
	q->cur_caplen = read_next_info(pf, 4);
	q->cur_len = read_next_info(pf, 4);
	next_pkt = pf->cur + q->cur_caplen;

	/* prepare fields in q for the generator */
	q->cur_pkt = pf->cur;
	/* initial estimate of tx time */
	q->cur_tt = cur_ts - last_ts;
	    // -pf->first_ts + loops * pf->total_tx_time - last_ts;

	if ((i % pf->tot_pkt) == 0)
	   ED("insert %5d len %lu cur_tt %.6f",
		i, (u_long)q->cur_len, 1e-9*q->cur_tt);

	/* prepare for next iteration */
	pf->cur = next_pkt;
	last_ts = cur_ts;
	if (next_pkt == pf->lim) {	//last pkt
	    pf->cur = pf->data + sizeof(struct pcap_file_header);
    	    last_ts = pf->first_ts; /* beginning of the trace */
	    loops++;
	}

	q->c_loss.run(q, &q->c_loss);
	if (q->cur_drop)
	    continue;
	q->c_bw.run(q, &q->c_bw);
	tt = q->cur_tt;
	q->qt_qout += tt;
#if 0
	if (drop_after(q))
	    continue;
#endif
	q->c_delay.run(q, &q->c_delay); /* compute delay */
	t_tx = q->qt_qout + q->cur_delay;
	ND(5, "tt %ld qout %ld tx %ld qt_tx %ld", tt, q->qt_qout, t_tx, q->qt_tx);
	/* insure no reordering and spacing by transmission time */
	q->qt_tx = (t_tx >= q->qt_tx + tt) ? t_tx : q->qt_tx + tt;
	enq(q);

	q->tx++;
	ND("ins %d q->prod_tail = %lu", (int)insert, (unsigned long)q->prod_tail);
    }
    /* loop marker ? */
    ED("done q->prod_tail:%d",(int)q->prod_tail);
    q->_tail = q->prod_tail; /* publish */

    return NULL;
fail:
    if (q->buf != NULL) {
	free(q->buf);
    }
    nmport_close(pa->pb);
    return (NULL);
}


/*
 * the consumer reads from the queue using head,
 * advances it every now and then.
 */
static void *
cons(void *_pa)
{
    struct pipe_args *pa = _pa;
    struct _qs *q = &pa->q;
    int pending = 0;
    uint64_t last_ts = 0;

    /* read the start of times in q->t0 */
    set_tns_now(&q->t0, 0);
    /* set the time (cons_now) to clock - q->t0 */
    set_tns_now(&q->cons_now, q->t0);
    q->cons_head = q->_head;
    q->cons_tail = q->_tail;
    while (!do_abort) { /* consumer, infinite */
	struct q_pkt *p = pkt_at(q, q->cons_head);

	__builtin_prefetch (q->buf + p->next);

	if (q->cons_head == q->cons_tail) {	//reset record
	    ND("Transmission restarted");
	    /*
	     * add to q->t0 the time for the last packet
	     */
	    q->t0 += last_ts;
	    set_tns_now(&q->cons_now, q->t0);
	    q->cons_head = 0;	//restart from beginning of the queue
	    continue;
	}
	last_ts = p->pt_tx;
	if (ts_cmp(p->pt_tx, q->cons_now) > 0) {
	    // packet not ready
	    q->rx_wait++;
	    /* the ioctl should be conditional */
	    ioctl(pa->pb->fd, NIOCTXSYNC, 0); // XXX just in case
	    pending = 0;
	    usleep(20);
	    set_tns_now(&q->cons_now, q->t0);
	    continue;
	}
	/* XXX copy is inefficient but simple */
	if (nmport_inject(pa->pb, (char *)(p + 1), p->pktlen) == 0) {
	    RD(1, "inject failed len %d now %ld tx %ld h %ld t %ld next %ld",
		(int)p->pktlen, (u_long)q->cons_now, (u_long)p->pt_tx,
		(u_long)q->_head, (u_long)q->_tail, (u_long)p->next);
	    ioctl(pa->pb->fd, NIOCTXSYNC, 0);
	    pending = 0;
	    continue;
	}
	pending++;
	if (pending > q->burst) {
	    ioctl(pa->pb->fd, NIOCTXSYNC, 0);
	    pending = 0;
	}

	q->cons_head = p->next;
	/* drain packets from the queue */
	q->rx++;
    }
    D("exiting on abort");
    return NULL;
}

/*
 * In case of pcap file as input, the program acts in 2 different
 * phases. It first fill the queue and then starts the cons()
 */
static void *
nmreplay_main(void *_a)
{
    struct pipe_args *a = _a;
    struct _qs *q = &a->q;
    const char *cap_fname = q->prod_ifname;

    setaffinity(a->cons_core);
    set_tns_now(&q->t0, 0); /* starting reference */
    if (cap_fname == NULL) {
	goto fail;
    }
    q->pcap = readpcap(cap_fname);
    if (q->pcap == NULL) {
	EEE("unable to read file %s", cap_fname);
	goto fail;
    }
    pcap_prod((void*)a);
    destroy_pcap(q->pcap);
    q->pcap = NULL;
    a->pb = nmport_open(q->cons_ifname);
    if (a->pb == NULL) {
	EEE("cannot open netmap on %s", q->cons_ifname);
	do_abort = 1; // XXX any better way ?
	return NULL;
    }
    /* continue as cons() */
    WWW("prepare to send packets");
    usleep(1000);
    cons((void*)a);
    EEE("exiting on abort");
fail:
    if (q->pcap != NULL) {
	destroy_pcap(q->pcap);
    }
    do_abort = 1;
    return NULL;
}


static void
sigint_h(int sig)
{
	(void)sig;	/* UNUSED */
	do_abort = 1;
	signal(SIGINT, SIG_DFL);
}



static void
usage(void)
{
	fprintf(stderr,
	    "usage: nmreplay [-v] [-D delay] [-B {[constant,]bps|ether,bps|real,speedup}] [-L loss]\n"
	    "\t[-b burst] -f pcap-file -i <netmap:ifname|valeSSS:PPP>\n");
	exit(1);
}


/*---- configuration handling ---- */
/*
 * support routine: split argument, returns ac and *av.
 * av contains two extra entries, a NULL and a pointer
 * to the entire string.
 */
static char **
split_arg(const char *src, int *_ac)
{
    char *my = NULL, **av = NULL;
    const char *seps = " \t\r\n,";
    int l, i, ac; /* number of entries */

    if (!src)
	return NULL;
    l = strlen(src);
    /* in the first pass we count fields, in the second pass
     * we allocate the av[] array and a copy of the string
     * and fill av[]. av[ac] = NULL, av[ac+1]
     */
    for (;;) {
	i = ac = 0;
	ND("start pass %d: <%s>", av ? 1 : 0, my);
	while (i < l) {
	    /* trim leading separator */
	    while (i <l && strchr(seps, src[i]))
		i++;
	    if (i >= l)
		break;
	    ND("   pass %d arg %d: <%s>", av ? 1 : 0, ac, src+i);
	    if (av) /* in the second pass, set the result */
		av[ac] = my+i;
	    ac++;
	    /* skip string */
	    while (i <l && !strchr(seps, src[i])) i++;
	    if (av)
		my[i] = '\0'; /* write marker */
	}
	if (!av) { /* end of first pass */
	    ND("ac is %d", ac);
	    av = calloc(1, (l+1) + (ac + 2)*sizeof(char *));
	    my = (char *)&(av[ac+2]);
	    strcpy(my, src);
	} else {
	    break;
	}
    }
    for (i = 0; i < ac; i++) {
	NED("%d: <%s>", i, av[i]);
    }
    av[i++] = NULL;
    av[i++] = my;
    *_ac = ac;
    return av;
}


/*
 * apply a command against a set of functions,
 * install a handler in *dst
 */
static int
cmd_apply(const struct _cfg *a, const char *arg, struct _qs *q, struct _cfg *dst)
{
	int ac = 0;
	char **av;
	int i;

	if (arg == NULL || *arg == '\0')
		return 1; /* no argument may be ok */
	if (a == NULL || dst == NULL) {
		ED("program error - invalid arguments");
		exit(1);
	}
	av = split_arg(arg, &ac);
	if (av == NULL)
		return 1; /* error */
	for (i = 0; a[i].parse; i++) {
		struct _cfg x = a[i];
		const char *errmsg = x.optarg;
		int ret;

		x.arg = NULL;
		x.arg_len = 0;
		bzero(&x.d, sizeof(x.d));
		ND("apply %s to %s", av[0], errmsg);
		ret = x.parse(q, &x, ac, av);
		if (ret == 2) /* not recognised */
			continue;
		if (ret == 1) {
			ED("invalid arguments: need '%s' have '%s'",
				errmsg, arg);
			break;
		}
		x.optarg = arg;
		*dst = x;
		return 0;
	}
	ED("arguments %s not recognised", arg);
	free(av);
	return 1;
}

static struct _cfg delay_cfg[];
static struct _cfg bw_cfg[];
static struct _cfg loss_cfg[];

static uint64_t parse_bw(const char *arg);

/*
 * prodcons [options]
 * accept separate sets of arguments for the two directions
 *
 */

static void
add_to(const char ** v, int l, const char *arg, const char *msg)
{
	for (; l > 0 && *v != NULL ; l--, v++);
	if (l == 0) {
		ED("%s %s", msg, arg);
		exit(1);
	}
	*v = arg;
}

int
main(int argc, char **argv)
{
	int ch, i, err=0;

#define	N_OPTS	1
	struct pipe_args bp[N_OPTS];
	const char *d[N_OPTS], *b[N_OPTS], *l[N_OPTS], *q[N_OPTS], *ifname[N_OPTS], *m[N_OPTS];
	const char *pcap_file[N_OPTS];
	int cores[4] = { 2, 8, 4, 10 }; /* default values */

	bzero(&bp, sizeof(bp));	/* all data initially go here */
	bzero(d, sizeof(d));
	bzero(b, sizeof(b));
	bzero(l, sizeof(l));
	bzero(q, sizeof(q));
	bzero(m, sizeof(m));
	bzero(ifname, sizeof(ifname));
	bzero(pcap_file, sizeof(pcap_file));


	/* set default values */
	for (i = 0; i < N_OPTS; i++) {
	    struct _qs *qs = &bp[i].q;

	    qs->burst = 128;
	    qs->c_delay.optarg = "0";
	    qs->c_delay.run = null_run_fn;
	    qs->c_loss.optarg = "0";
	    qs->c_loss.run = null_run_fn;
	    qs->c_bw.optarg = "0";
	    qs->c_bw.run = null_run_fn;
	}

	// Options:
	// B	bandwidth in bps
	// D	delay in seconds
	// L	loss probability
	// f	pcap file
	// i	interface name
	// w	wait link
	// b	batch size
	// v	verbose
	// C	cpu placement

	while ( (ch = getopt(argc, argv, "B:C:D:L:b:f:i:vw:")) != -1) {
		switch (ch) {
		default:
			D("bad option %c %s", ch, optarg);
			usage();
			break;

		case 'C': /* CPU placement, up to 4 arguments */
			{
				int ac = 0;
				char **av = split_arg(optarg, &ac);
				if (ac == 1) { /* sequential after the first */
					cores[0] = atoi(av[0]);
					cores[1] = cores[0] + 1;
					cores[2] = cores[1] + 1;
					cores[3] = cores[2] + 1;
				} else if (ac == 2) { /* two sequential pairs */
					cores[0] = atoi(av[0]);
					cores[1] = cores[0] + 1;
					cores[2] = atoi(av[1]);
					cores[3] = cores[2] + 1;
				} else if (ac == 4) { /* four values */
					cores[0] = atoi(av[0]);
					cores[1] = atoi(av[1]);
					cores[2] = atoi(av[2]);
					cores[3] = atoi(av[3]);
				} else {
					ED(" -C accepts 1, 2 or 4 comma separated arguments");
					usage();
				}
				if (av)
					free(av);
			}
			break;

		case 'B': /* bandwidth in bps */
			add_to(b, N_OPTS, optarg, "-B too many times");
			break;

		case 'D': /* delay in seconds (float) */
			add_to(d, N_OPTS, optarg, "-D too many times");
			break;

		case 'L': /* loss probability */
			add_to(l, N_OPTS, optarg, "-L too many times");
			break;

		case 'b':	/* burst */
			bp[0].q.burst = atoi(optarg);
			break;

		case 'f':	/* pcap_file */
			add_to(pcap_file, N_OPTS, optarg, "-f too many times");
			break;
		case 'i':	/* interface */
			add_to(ifname, N_OPTS, optarg, "-i too many times");
			break;
		case 'v':
			verbose++;
			break;
		case 'w':
			bp[0].wait_link = atoi(optarg);
			break;
		}

	}

	argc -= optind;
	argv += optind;

	/*
	 * consistency checks for common arguments
	 * if pcap file has been provided we need just one interface, two otherwise
	 */
	if (!pcap_file[0]) {
		ED("missing pcap file");
		usage();
	}
	if (!ifname[0]) {
		ED("missing interface");
		usage();
	}
	if (bp[0].q.burst < 1 || bp[0].q.burst > 8192) {
		WWW("invalid burst %d, set to 1024", bp[0].q.burst);
		bp[0].q.burst = 1024; // XXX 128 is probably better
	}
	if (bp[0].wait_link > 100) {
		ED("invalid wait_link %d, set to 4", bp[0].wait_link);
		bp[0].wait_link = 4;
	}

	bp[0].q.prod_ifname = pcap_file[0];
	bp[0].q.cons_ifname = ifname[0];

	/* assign cores. prod and cons work better if on the same HT */
	bp[0].cons_core = cores[0];
	bp[0].prod_core = cores[1];
	ED("running on cores %d %d %d %d", cores[0], cores[1], cores[2], cores[3]);

	/* apply commands */
	for (i = 0; i < N_OPTS; i++) { /* once per queue */
		struct _qs *qs = &bp[i].q;
		err += cmd_apply(delay_cfg, d[i], qs, &qs->c_delay);
		err += cmd_apply(bw_cfg, b[i], qs, &qs->c_bw);
		err += cmd_apply(loss_cfg, l[i], qs, &qs->c_loss);
		if (err != 0)
			exit(1);
	}

	pthread_create(&bp[0].cons_tid, NULL, nmreplay_main, (void*)&bp[0]);
	signal(SIGINT, sigint_h);
	sleep(1);
	while (!do_abort) {
	    struct _qs olda = bp[0].q;
	    struct _qs *q0 = &bp[0].q;

	    sleep(1);
	    ED("%lld -> %lld maxq %d round %lld",
		(long long)(q0->rx - olda.rx), (long long)(q0->tx - olda.tx),
		q0->rx_qmax, (long long)q0->prod_max_gap
		);
	    ED("plr nominal %le actual %le",
		(double)(q0->c_loss.d[0])/(1<<24),
		q0->c_loss.d[1] == 0 ? 0 :
		(double)(q0->c_loss.d[2])/q0->c_loss.d[1]);
	    bp[0].q.rx_qmax = (bp[0].q.rx_qmax * 7)/8; // ewma
	    bp[0].q.prod_max_gap = (bp[0].q.prod_max_gap * 7)/8; // ewma
	}
	D("exiting on abort");
	sleep(1);

	return (0);
}

/* conversion factor for numbers.
 * Each entry has a set of characters and conversion factor,
 * the first entry should have an empty string and default factor,
 * the final entry has s = NULL.
 */
struct _sm {	/* string and multiplier */
	const char *s;
	double m;
};

/*
 * parse a generic value
 */
static double
parse_gen(const char *arg, const struct _sm *conv, int *err)
{
	double d;
	char *ep;
	int dummy;

	if (err == NULL)
		err = &dummy;
	*err = 0;
	if (arg == NULL)
		goto error;
	d = strtod(arg, &ep);
	if (ep == arg) { /* no value */
		ED("bad argument %s", arg);
		goto error;
	}
	/* special case, no conversion */
	if (conv == NULL && *ep == '\0')
		goto done;
	ND("checking %s [%s]", arg, ep);
	for (;conv->s; conv++) {
		if (strchr(conv->s, *ep))
			goto done;
	}
error:
	*err = 1;	/* unrecognised */
	return 0;

done:
	if (conv) {
		ND("scale is %s %lf", conv->s, conv->m);
		d *= conv->m; /* apply default conversion */
	}
	ND("returning %lf", d);
	return d;
}

#define U_PARSE_ERR ~(0ULL)

/* returns a value in nanoseconds */
static uint64_t
parse_time(const char *arg)
{
    struct _sm a[] = {
	{"", 1000000000 /* seconds */},
	{"n", 1 /* nanoseconds */}, {"u", 1000 /* microseconds */},
	{"m", 1000000 /* milliseconds */}, {"s", 1000000000 /* seconds */},
	{NULL, 0 /* seconds */}
    };
    int err;
    uint64_t ret = (uint64_t)parse_gen(arg, a, &err);
    return err ? U_PARSE_ERR : ret;
}


/*
 * parse a bandwidth, returns value in bps or U_PARSE_ERR if error.
 */
static uint64_t
parse_bw(const char *arg)
{
    struct _sm a[] = {
	{"", 1}, {"kK", 1000}, {"mM", 1000000}, {"gG", 1000000000}, {NULL, 0}
    };
    int err;
    uint64_t ret = (uint64_t)parse_gen(arg, a, &err);
    return err ? U_PARSE_ERR : ret;
}


/*
 * For some function we need random bits.
 * This is a wrapper to whatever function you want that returns
 * 24 useful random bits.
 */

static inline uint64_t
my_random24(void)	/* 24 useful bits */
{
	return random() & ((1<<24) - 1);
}


/*-------------- user-configuration -----------------*/

#if 0 /* start of comment block */

Here we place the functions to implement the various features
of the system. For each feature one should define a struct _cfg
(see at the beginning for definition) that refers a *_parse() function
to extract values from the command line, and a *_run() function
that is invoked on each packet to implement the desired function.

Examples of the two functions are below. In general:

- the *_parse() function takes argc/argv[], matches the function
  name in argv[0], extracts the operating parameters, allocates memory
  if needed, and stores them in the struct _cfg.
  Return value is 2 if argv[0] is not recosnised, 1 if there is an
  error in the arguments, 0 if all ok.

  On the command line, argv[] is a single, comma separated argument
  that follow the specific option eg -D constant,20ms

  struct _cfg has some preallocated space (e.g an array of uint64_t) so simple
  function can use that without having to allocate memory.

- the *_run() function takes struct _q *q and struct _cfg *cfg as arguments.
  *q contains all the informatio that may be possibly needed, including
  those on the packet currently under processing.
  The basic values are the following:

	char *	 cur_pkt 	points to the current packet (linear buffer)
	uint32_t cur_len;	length of the current packet
		the functions are not supposed to modify these values

	int	 cur_drop;	true if current packet must be dropped.
		Must be set to non-zero by the loss emulation function

	uint64_t cur_delay;	delay in nanoseconds for the current packet
		Must be set by the delay emulation function

   More sophisticated functions may need to access other fields in *q,
   see the structure description for that.

When implementing a new function for a feature (e.g. for delay,
bandwidth, loss...) the struct _cfg should be added to the array
that contains all possible options.

		--- Specific notes ---

DELAY emulation		-D option_arguments

    If the option is not supplied, the system applies 0 extra delay

    The resolution for times is 1ns, the precision is load dependent and
    generally in the order of 20-50us.
    Times are in nanoseconds, can be followed by a character specifying
    a different unit e.g.

	n	nanoseconds
	u	microseconds
	m	milliseconds
	s	seconds

    Currently implemented options:

    constant,t		constant delay equal to t

    uniform,tmin,tmax	uniform delay between tmin and tmax

    exp,tavg,tmin	exponential distribution with average tavg
			and minimum tmin (corresponds to an exponential
			distribution with argument 1/(tavg-tmin) )


LOSS emulation		-L option_arguments

    Loss is expressed as packet or bit error rate, which is an absolute
    number between 0 and 1 (typically small).

    Currently implemented options

    plr,p		uniform packet loss rate p, independent
			of packet size

    burst,p,lmin,lmax 	burst loss with burst probability p and
			burst length uniformly distributed between
			lmin and lmax

    ber,p		uniformly distributed bit error rate p,
			so actual loss prob. depends on size.

BANDWIDTH emulation	-B option_arguments

    Bandwidths are expressed in bits per second, can be followed by a
    character specifying a different unit e.g.

	b/B	bits per second
	k/K	kbits/s (10^3 bits/s)
	m/M	mbits/s (10^6 bits/s)
	g/G	gbits/s (10^9 bits/s)

    Currently implemented options

    const,b		constant bw, excluding mac framing
    ether,b		constant bw, including ethernet framing
			(20 bytes framing + 4 bytes crc)
    real,[scale]	use real time, optionally with a scaling factor

#endif /* end of comment block */

/*
 * Configuration options for delay
 */

/* constant delay, also accepts just a number */
static int
const_delay_parse(struct _qs *q, struct _cfg *dst, int ac, char *av[])
{
	uint64_t delay;

	(void)q;
	if (strncmp(av[0], "const", 5) != 0 && ac > 1)
		return 2; /* unrecognised */
	if (ac > 2)
		return 1; /* error */
	delay = parse_time(av[ac - 1]);
	if (delay == U_PARSE_ERR)
		return 1; /* error */
	dst->d[0] = delay;
	return 0;	/* success */
}

/* runtime function, store the delay into q->cur_delay */
static int
const_delay_run(struct _qs *q, struct _cfg *arg)
{
	q->cur_delay = arg->d[0]; /* the delay */
	return 0;
}

static int
uniform_delay_parse(struct _qs *q, struct _cfg *dst, int ac, char *av[])
{
	uint64_t dmin, dmax;

	(void)q;
	if (strcmp(av[0], "uniform") != 0)
		return 2; /* not recognised */
	if (ac != 3)
		return 1; /* error */
	dmin = parse_time(av[1]);
	dmax = parse_time(av[2]);
	if (dmin == U_PARSE_ERR || dmax == U_PARSE_ERR || dmin > dmax)
		return 1;
	D("dmin %lld dmax %lld", (long long)dmin, (long long)dmax);
	dst->d[0] = dmin;
	dst->d[1] = dmax;
	dst->d[2] = dmax - dmin;
	return 0;
}

static int
uniform_delay_run(struct _qs *q, struct _cfg *arg)
{
	uint64_t x = my_random24();
	q->cur_delay = arg->d[0] + ((arg->d[2] * x) >> 24);
#if 0 /* COMPUTE_STATS */
#endif /* COMPUTE_STATS */
	return 0;
}

/*
 * exponential delay: Prob(delay = x) = exp(-x/d_av)
 * gives a delay between 0 and infinity with average d_av
 * The cumulative function is 1 - d_av exp(-x/d_av)
 *
 * The inverse function generates a uniform random number p in 0..1
 * and generates delay = (d_av-d_min) * -ln(1-p) + d_min
 *
 * To speed up behaviour at runtime we tabulate the values
 */

static int
exp_delay_parse(struct _qs *q, struct _cfg *dst, int ac, char *av[])
{
#define	PTS_D_EXP	512
	uint64_t i, d_av, d_min, *t; /*table of values */

        (void)q;
        if (strcmp(av[0], "exp") != 0)
		return 2; /* not recognised */
        if (ac != 3)
                return 1; /* error */
        d_av = parse_time(av[1]);
        d_min = parse_time(av[2]);
        if (d_av == U_PARSE_ERR || d_min == U_PARSE_ERR || d_av < d_min)
                return 1; /* error */
	d_av -= d_min;
	dst->arg_len = PTS_D_EXP * sizeof(uint64_t);
	dst->arg = calloc(1, dst->arg_len);
	if (dst->arg == NULL)
		return 1; /* no memory */
	t = (uint64_t *)dst->arg;
	/* tabulate -ln(1-n)*delay  for n in 0..1 */
	for (i = 0; i < PTS_D_EXP; i++) {
		double d = -log2 ((double)(PTS_D_EXP - i) / PTS_D_EXP) * d_av + d_min;
		t[i] = (uint64_t)d;
		ND(5, "%ld: %le", i, d);
	}
        return 0;
}

static int
exp_delay_run(struct _qs *q, struct _cfg *arg)
{
	uint64_t *t = (uint64_t *)arg->arg;
        q->cur_delay = t[my_random24() & (PTS_D_EXP - 1)];
	RD(5, "delay %llu", (unsigned long long)q->cur_delay);
        return 0;
}


/* unused arguments in configuration */
#define TLEM_CFG_END	NULL, 0, {0}, {0}

static struct _cfg delay_cfg[] = {
	{ const_delay_parse, const_delay_run,
		"constant,delay", TLEM_CFG_END },
	{ uniform_delay_parse, uniform_delay_run,
		"uniform,dmin,dmax # dmin <= dmax", TLEM_CFG_END },
	{ exp_delay_parse, exp_delay_run,
		"exp,dmin,davg # dmin <= davg", TLEM_CFG_END },
	{ NULL, NULL, NULL, TLEM_CFG_END }
};

/* standard bandwidth, also accepts just a number */
static int
const_bw_parse(struct _qs *q, struct _cfg *dst, int ac, char *av[])
{
	uint64_t bw;

	(void)q;
	if (strncmp(av[0], "const", 5) != 0 && ac > 1)
		return 2; /* unrecognised */
	if (ac > 2)
		return 1; /* error */
	bw = parse_bw(av[ac - 1]);
	if (bw == U_PARSE_ERR) {
		return (ac == 2) ? 1 /* error */ : 2 /* unrecognised */;
	}
	dst->d[0] = bw;
	return 0;	/* success */
}


/* runtime function, store the delay into q->cur_delay */
static int
const_bw_run(struct _qs *q, struct _cfg *arg)
{
	uint64_t bps = arg->d[0];
	q->cur_tt = bps ? 8ULL* TIME_UNITS * q->cur_len / bps : 0 ;
	return 0;
}

/* ethernet bandwidth, add 672 bits per packet */
static int
ether_bw_parse(struct _qs *q, struct _cfg *dst, int ac, char *av[])
{
	uint64_t bw;

	(void)q;
	if (strcmp(av[0], "ether") != 0)
		return 2; /* unrecognised */
	if (ac != 2)
		return 1; /* error */
	bw = parse_bw(av[ac - 1]);
	if (bw == U_PARSE_ERR)
		return 1; /* error */
	dst->d[0] = bw;
	return 0;	/* success */
}


/* runtime function, add 20 bytes (framing) + 4 bytes (crc) */
static int
ether_bw_run(struct _qs *q, struct _cfg *arg)
{
	uint64_t bps = arg->d[0];
	q->cur_tt = bps ? 8ULL * TIME_UNITS * (q->cur_len + 24) / bps : 0 ;
	return 0;
}

/* real bandwidth, plus scaling factor */
static int
real_bw_parse(struct _qs *q, struct _cfg *dst, int ac, char *av[])
{
	double scale;

	(void)q;
	if (strcmp(av[0], "real") != 0)
		return 2; /* unrecognised */
	if (ac > 2) { /* second argument is optional */
		return 1; /* error */
	} else if (ac == 1) {
		scale = 1;
	} else {
		int err = 0;
		scale = parse_gen(av[ac-1], NULL, &err);
		if (err || scale <= 0 || scale > 1000)
			return 1;
	}
	ED("real -> scale is %.6f", scale);
	dst->f[0] = scale;
	return 0;	/* success */
}

static int
real_bw_run(struct _qs *q, struct _cfg *arg)
{
	q->cur_tt /= arg->f[0];
	return 0;
}

static struct _cfg bw_cfg[] = {
	{ const_bw_parse, const_bw_run,
		"constant,bps", TLEM_CFG_END },
	{ ether_bw_parse, ether_bw_run,
		"ether,bps", TLEM_CFG_END },
	{ real_bw_parse, real_bw_run,
		"real,scale", TLEM_CFG_END },
	{ NULL, NULL, NULL, TLEM_CFG_END }
};

/*
 * loss patterns
 */
static int
const_plr_parse(struct _qs *q, struct _cfg *dst, int ac, char *av[])
{
	double plr;
	int err;

	(void)q;
	if (strcmp(av[0], "plr") != 0 && ac > 1)
		return 2; /* unrecognised */
	if (ac > 2)
		return 1; /* error */
	// XXX to be completed
	plr = parse_gen(av[ac-1], NULL, &err);
	if (err || plr < 0 || plr > 1)
		return 1;
	dst->d[0] = plr * (1<<24); /* scale is 16m */
	if (plr != 0 && dst->d[0] == 0)
		ED("WWW warning,  rounding %le down to 0", plr);
	return 0;	/* success */
}

static int
const_plr_run(struct _qs *q, struct _cfg *arg)
{
	(void)arg;
	uint64_t r = my_random24();
	q->cur_drop = r < arg->d[0];
#if 1	/* keep stats */
	arg->d[1]++;
	arg->d[2] += q->cur_drop;
#endif
	return 0;
}


/*
 * For BER the loss is 1- (1-ber)**bit_len
 * The linear approximation is only good for small values, so we
 * tabulate (1-ber)**len for various sizes in bytes
 */
static int
const_ber_parse(struct _qs *q, struct _cfg *dst, int ac, char *av[])
{
	double ber, ber8, cur;
	int i, err;
	uint32_t *plr;
	const uint32_t mask = (1<<24) - 1;

	(void)q;
	if (strcmp(av[0], "ber") != 0)
		return 2; /* unrecognised */
	if (ac != 2)
		return 1; /* error */
	ber = parse_gen(av[ac-1], NULL, &err);
	if (err || ber < 0 || ber > 1)
		return 1;
	dst->arg_len = MAX_PKT * sizeof(uint32_t);
	plr = calloc(1, dst->arg_len);
	if (plr == NULL)
		return 1; /* no memory */
	dst->arg = plr;
	ber8 = 1 - ber;
	ber8 *= ber8; /* **2 */
	ber8 *= ber8; /* **4 */
	ber8 *= ber8; /* **8 */
	cur = 1;
	for (i=0; i < MAX_PKT; i++, cur *= ber8) {
		plr[i] = (mask + 1)*(1 - cur);
		if (plr[i] > mask)
			plr[i] = mask;
#if 0
		if (i>= 60) //  && plr[i] < mask/2)
			RD(50,"%4d: %le %ld", i, 1.0 - cur, (_P64)plr[i]);
#endif
	}
	dst->d[0] = ber * (mask + 1);
	return 0;	/* success */
}

static int
const_ber_run(struct _qs *q, struct _cfg *arg)
{
	int l = q->cur_len;
	uint64_t r = my_random24();
	uint32_t *plr = arg->arg;

	if (l >= MAX_PKT) {
		RD(5, "pkt len %d too large, trim to %d", l, MAX_PKT-1);
		l = MAX_PKT-1;
	}
	q->cur_drop = r < plr[l];
#if 1	/* keep stats */
	arg->d[1] += l * 8;
	arg->d[2] += q->cur_drop;
#endif
	return 0;
}

static struct _cfg loss_cfg[] = {
	{ const_plr_parse, const_plr_run,
		"plr,prob # 0 <= prob <= 1", TLEM_CFG_END },
	{ const_ber_parse, const_ber_run,
		"ber,prob # 0 <= prob <= 1", TLEM_CFG_END },
	{ NULL, NULL, NULL, TLEM_CFG_END }
};