initial version, corresponding to ipfw3-2012

[ipfw-google.git] / sys / netinet / ipfw / ip_dn_io.c

/*-
 * Copyright (c) 2010 Luigi Rizzo, Riccardo Panicucci, 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.
 */

/*
 * Dummynet portions related to packet handling.
 */
#include <sys/cdefs.h>
__FBSDID("$FreeBSD: user/luigi/ipfw3-head/sys/netinet/ipfw/ip_dn_io.c 203321 2010-01-31 21:39:25Z luigi $");

#include "opt_inet6.h"

#include <sys/param.h>
#include <sys/systm.h>
#include <sys/malloc.h>
#include <sys/mbuf.h>
#include <sys/kernel.h>
#include <sys/lock.h>
#include <sys/module.h>
#include <sys/priv.h>
#include <sys/proc.h>
#include <sys/rwlock.h>
#include <sys/socket.h>
#include <sys/time.h>
#include <sys/sysctl.h>

#include <net/if.h>     /* IFNAMSIZ, struct ifaddr, ifq head, lock.h mutex.h */
#include <net/netisr.h>
#include <net/vnet.h>

#include <netinet/in.h>
#include <netinet/ip.h>         /* ip_len, ip_off */
#include <netinet/ip_var.h>     /* ip_output(), IP_FORWARDING */
#include <netinet/ip_fw.h>
#include <netinet/ipfw/ip_fw_private.h>
#include <netinet/ipfw/dn_heap.h>
#include <netinet/ip_dummynet.h>
#include <netinet/ipfw/ip_dn_private.h>
#include <netinet/ipfw/dn_sched.h>

#include <netinet/if_ether.h> /* various ether_* routines */

#include <netinet/ip6.h>       /* for ip6_input, ip6_output prototypes */
#include <netinet6/ip6_var.h>

/*
 * We keep a private variable for the simulation time, but we could
 * probably use an existing one ("softticks" in sys/kern/kern_timeout.c)
 * instead of dn_cfg.curr_time
 */

struct dn_parms dn_cfg;
//VNET_DEFINE(struct dn_parms, _base_dn_cfg);

static long tick_last;          /* Last tick duration (usec). */
static long tick_delta;         /* Last vs standard tick diff (usec). */
static long tick_delta_sum;     /* Accumulated tick difference (usec).*/
static long tick_adjustment;    /* Tick adjustments done. */
static long tick_lost;          /* Lost(coalesced) ticks number. */
/* Adjusted vs non-adjusted curr_time difference (ticks). */
static long tick_diff;

static unsigned long    io_pkt;
static unsigned long    io_pkt_fast;
static unsigned long    io_pkt_drop;

/*
 * We use a heap to store entities for which we have pending timer events.
 * The heap is checked at every tick and all entities with expired events
 * are extracted.
 */
  
MALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap");

extern  void (*bridge_dn_p)(struct mbuf *, struct ifnet *);

#ifdef SYSCTL_NODE

SYSBEGIN(f4)

SYSCTL_DECL(_net_inet);
SYSCTL_DECL(_net_inet_ip);
SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet, CTLFLAG_RW, 0, "Dummynet");

/* wrapper to pass dn_cfg fields to SYSCTL_* */
//#define DC(x) (&(VNET_NAME(_base_dn_cfg).x))
#define DC(x)   (&(dn_cfg.x))
/* parameters */
SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, hash_size,
    CTLFLAG_RW, DC(hash_size), 0, "Default hash table size");
SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, pipe_slot_limit,
    CTLFLAG_RW, DC(slot_limit), 0,
    "Upper limit in slots for pipe queue.");
SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, pipe_byte_limit,
    CTLFLAG_RW, DC(byte_limit), 0,
    "Upper limit in bytes for pipe queue.");
SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, io_fast,
    CTLFLAG_RW, DC(io_fast), 0, "Enable fast dummynet io.");
SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, debug,
    CTLFLAG_RW, DC(debug), 0, "Dummynet debug level");

/* RED parameters */
SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth,
    CTLFLAG_RD, DC(red_lookup_depth), 0, "Depth of RED lookup table");
SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size,
    CTLFLAG_RD, DC(red_avg_pkt_size), 0, "RED Medium packet size");
SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size,
    CTLFLAG_RD, DC(red_max_pkt_size), 0, "RED Max packet size");

/* time adjustment */
SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta,
    CTLFLAG_RD, &tick_delta, 0, "Last vs standard tick difference (usec).");
SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta_sum,
    CTLFLAG_RD, &tick_delta_sum, 0, "Accumulated tick difference (usec).");
SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_adjustment,
    CTLFLAG_RD, &tick_adjustment, 0, "Tick adjustments done.");
SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_diff,
    CTLFLAG_RD, &tick_diff, 0,
    "Adjusted vs non-adjusted curr_time difference (ticks).");
SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_lost,
    CTLFLAG_RD, &tick_lost, 0,
    "Number of ticks coalesced by dummynet taskqueue.");

/* Drain parameters */
SYSCTL_UINT(_net_inet_ip_dummynet, OID_AUTO, expire,
    CTLFLAG_RW, DC(expire), 0, "Expire empty queues/pipes");
SYSCTL_UINT(_net_inet_ip_dummynet, OID_AUTO, expire_cycle,
    CTLFLAG_RD, DC(expire_cycle), 0, "Expire cycle for queues/pipes");
SYSCTL_UINT(_net_inet_ip_dummynet, OID_AUTO, expire_object,
    CTLFLAG_RW, DC(expire_object), 0, "Min # of objects before start drain routine");
SYSCTL_UINT(_net_inet_ip_dummynet, OID_AUTO, object_idle_tick,
    CTLFLAG_RD, DC(object_idle_tick), 0, "Time (in ticks) to cosiderer an object as idle");
SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, drain_ratio,
    CTLFLAG_RD, DC(drain_ratio), 0, "% of dummynet_task() to dedicate to drain routine");

/* statistics */
SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, schk_count,
    CTLFLAG_RD, DC(schk_count), 0, "Number of schedulers");
SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, si_count,
    CTLFLAG_RD, DC(si_count), 0, "Number of scheduler instances");
SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, fsk_count,
    CTLFLAG_RD, DC(fsk_count), 0, "Number of flowsets");
SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, queue_count,
    CTLFLAG_RD, DC(queue_count), 0, "Number of queues");
SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt,
    CTLFLAG_RD, &io_pkt, 0,
    "Number of packets passed to dummynet.");
SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_fast,
    CTLFLAG_RD, &io_pkt_fast, 0,
    "Number of packets bypassed dummynet scheduler.");
SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_drop,
    CTLFLAG_RD, &io_pkt_drop, 0,
    "Number of packets dropped by dummynet.");
#undef DC
SYSEND

#endif

static void     dummynet_send(struct mbuf *);

/*
 * Packets processed by dummynet have an mbuf tag associated with
 * them that carries their dummynet state.
 * Outside dummynet, only the 'rule' field is relevant, and it must
 * be at the beginning of the structure.
 */
struct dn_pkt_tag {
        struct ipfw_rule_ref rule;      /* matching rule        */

        /* second part, dummynet specific */
        int dn_dir;             /* action when packet comes out.*/
                                /* see ip_fw_private.h          */
        uint64_t output_time;   /* when the pkt is due for delivery*/
        struct ifnet *ifp;      /* interface, for ip_output     */
        struct _ip6dn_args ip6opt;      /* XXX ipv6 options     */
};

/*
 * Return the mbuf tag holding the dummynet state (it should
 * be the first one on the list).
 */
static struct dn_pkt_tag *
dn_tag_get(struct mbuf *m)
{
        struct m_tag *mtag = m_tag_first(m);
        KASSERT(mtag != NULL &&
            mtag->m_tag_cookie == MTAG_ABI_COMPAT &&
            mtag->m_tag_id == PACKET_TAG_DUMMYNET,
            ("packet on dummynet queue w/o dummynet tag!"));
        return (struct dn_pkt_tag *)(mtag+1);
}

static inline void
mq_append(struct mq *q, struct mbuf *m)
{
        if (q->head == NULL)
                q->head = m;
        else
                q->tail->m_nextpkt = m;
        q->tail = m;
        m->m_nextpkt = NULL;
}

/*
 * Dispose a list of packet. Use a functions so if we need to do
 * more work, this is a central point to do it.
 */
void dn_free_pkts(struct mbuf *mnext)
{
        struct mbuf *m;
    
        while ((m = mnext) != NULL) {
                mnext = m->m_nextpkt;
                FREE_PKT(m);
        }
}

static int
red_drops (struct dn_queue *q, int len)
{
        /*
         * RED algorithm
         *
         * RED calculates the average queue size (avg) using a low-pass filter
         * with an exponential weighted (w_q) moving average:
         *      avg  <-  (1-w_q) * avg + w_q * q_size
         * where q_size is the queue length (measured in bytes or * packets).
         *
         * If q_size == 0, we compute the idle time for the link, and set
         *      avg = (1 - w_q)^(idle/s)
         * where s is the time needed for transmitting a medium-sized packet.
         *
         * Now, if avg < min_th the packet is enqueued.
         * If avg > max_th the packet is dropped. Otherwise, the packet is
         * dropped with probability P function of avg.
         */

        struct dn_fsk *fs = q->fs;
        int64_t p_b = 0;

        /* Queue in bytes or packets? */
        uint32_t q_size = (fs->fs.flags & DN_QSIZE_BYTES) ?
            q->ni.len_bytes : q->ni.length;

        /* Average queue size estimation. */
        if (q_size != 0) {
                /* Queue is not empty, avg <- avg + (q_size - avg) * w_q */
                int diff = SCALE(q_size) - q->avg;
                int64_t v = SCALE_MUL((int64_t)diff, (int64_t)fs->w_q);

                q->avg += (int)v;
        } else {
                /*
                 * Queue is empty, find for how long the queue has been
                 * empty and use a lookup table for computing
                 * (1 - * w_q)^(idle_time/s) where s is the time to send a
                 * (small) packet.
                 * XXX check wraps...
                 */
                if (q->avg) {
                        u_int t = div64((dn_cfg.curr_time - q->q_time), fs->lookup_step);

                        q->avg = (t < fs->lookup_depth) ?
                            SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0;
                }
        }

        /* Should i drop? */
        if (q->avg < fs->min_th) {
                q->count = -1;
                return (0);     /* accept packet */
        }
        if (q->avg >= fs->max_th) {     /* average queue >=  max threshold */
                if (fs->fs.flags & DN_IS_GENTLE_RED) {
                        /*
                         * According to Gentle-RED, if avg is greater than
                         * max_th the packet is dropped with a probability
                         *       p_b = c_3 * avg - c_4
                         * where c_3 = (1 - max_p) / max_th
                         *       c_4 = 1 - 2 * max_p
                         */
                        p_b = SCALE_MUL((int64_t)fs->c_3, (int64_t)q->avg) -
                            fs->c_4;
                } else {
                        q->count = -1;
                        return (1);
                }
        } else if (q->avg > fs->min_th) {
                /*
                 * We compute p_b using the linear dropping function
                 *       p_b = c_1 * avg - c_2
                 * where c_1 = max_p / (max_th - min_th)
                 *       c_2 = max_p * min_th / (max_th - min_th)
                 */
                p_b = SCALE_MUL((int64_t)fs->c_1, (int64_t)q->avg) - fs->c_2;
        }

        if (fs->fs.flags & DN_QSIZE_BYTES)
                p_b = div64((p_b * len) , fs->max_pkt_size);
        if (++q->count == 0)
                q->random = random() & 0xffff;
        else {
                /*
                 * q->count counts packets arrived since last drop, so a greater
                 * value of q->count means a greater packet drop probability.
                 */
                if (SCALE_MUL(p_b, SCALE((int64_t)q->count)) > q->random) {
                        q->count = 0;
                        /* After a drop we calculate a new random value. */
                        q->random = random() & 0xffff;
                        return (1);     /* drop */
                }
        }
        /* End of RED algorithm. */

        return (0);     /* accept */

}

/*
 * Enqueue a packet in q, subject to space and queue management policy
 * (whose parameters are in q->fs).
 * Update stats for the queue and the scheduler.
 * Return 0 on success, 1 on drop. The packet is consumed anyways.
 */
int
dn_enqueue(struct dn_queue *q, struct mbuf* m, int drop)
{   
        struct dn_fs *f;
        struct dn_flow *ni;     /* stats for scheduler instance */
        uint64_t len;

        if (q->fs == NULL || q->_si == NULL) {
                printf("%s fs %p si %p, dropping\n",
                        __FUNCTION__, q->fs, q->_si);
                FREE_PKT(m);
                return 1;
        }
        f = &(q->fs->fs);
        ni = &q->_si->ni;
        len = m->m_pkthdr.len;
        /* Update statistics, then check reasons to drop pkt. */
        q->ni.tot_bytes += len;
        q->ni.tot_pkts++;
        ni->tot_bytes += len;
        ni->tot_pkts++;
        if (drop)
                goto drop;
        if (f->plr && random() < f->plr)
                goto drop;
        if (f->flags & DN_IS_RED && red_drops(q, m->m_pkthdr.len))
                goto drop;
        if (f->flags & DN_QSIZE_BYTES) {
                if (q->ni.len_bytes > f->qsize)
                        goto drop;
        } else if (q->ni.length >= f->qsize) {
                goto drop;
        }
        mq_append(&q->mq, m);
        if (q->ni.length == 0) {        /* queue was idle */
                dn_cfg.idle_queue--;
                if (ni->length == 0)    /* scheduler was idle */
                        dn_cfg.idle_si--;
        }
        q->ni.length++;
        q->ni.len_bytes += len;
        ni->length++;
        ni->len_bytes += len;
        return 0;

drop:
        io_pkt_drop++;
        q->ni.drops++;
        ni->drops++;
        FREE_PKT(m);
        return 1;
}

/*
 * Fetch packets from the delay line which are due now. If there are
 * leftover packets, reinsert the delay line in the heap.
 * Runs under scheduler lock.
 */
static void
transmit_event(struct mq *q, struct delay_line *dline, uint64_t now)
{
        struct mbuf *m;
        struct dn_pkt_tag *pkt = NULL;

        dline->oid.subtype = 0; /* not in heap */
        while ((m = dline->mq.head) != NULL) {
                pkt = dn_tag_get(m);
                if (!DN_KEY_LEQ(pkt->output_time, now))
                        break;
                dline->mq.head = m->m_nextpkt;
                mq_append(q, m);
        }
        if (m != NULL) {
                dline->oid.subtype = 1; /* in heap */
                heap_insert(&dn_cfg.evheap, pkt->output_time, dline);
        }
}

/*
 * Convert the additional MAC overheads/delays into an equivalent
 * number of bits for the given data rate. The samples are
 * in milliseconds so we need to divide by 1000.
 */
static uint64_t
extra_bits(struct mbuf *m, struct dn_schk *s)
{
        int index;
        uint64_t bits;
        struct dn_profile *pf = s->profile;

        if (!pf || pf->samples_no == 0)
                return 0;
        index  = random() % pf->samples_no;
        bits = div64((uint64_t)pf->samples[index] * s->link.bandwidth, 1000);
        if (index >= pf->loss_level) {
                struct dn_pkt_tag *dt = dn_tag_get(m);
                if (dt)
                        dt->dn_dir = DIR_DROP;
        }
        return bits;
}

/*
 * Send traffic from a scheduler instance due by 'now'.
 * Return a pointer to the head of the queue.
 */
static struct mbuf *
serve_sched(struct mq *q, struct dn_sch_inst *si, uint64_t now)
{
        struct mq def_q;
        struct dn_schk *s = si->sched;
        struct mbuf *m = NULL;
        int delay_line_idle = (si->dline.mq.head == NULL);
        int done, bw;

        if (q == NULL) {
                q = &def_q;
                q->head = NULL;
        }

        bw = s->link.bandwidth;
        si->kflags &= ~DN_ACTIVE;

        if (bw > 0)
                si->credit += (now - si->sched_time) * bw;
        else
                si->credit = 0;
        si->sched_time = now;
        done = 0;
        while (si->credit >= 0 && (m = s->fp->dequeue(si)) != NULL) {
                uint64_t len_scaled;

                /*
                 * Some schedulers might want wake up the scheduler later.
                 * To suppor this the caller returns an mbuf with len < 0
                 * this will result in a new wake up of the scheduler
                 * instance between m->m_pkthdr.len ticks.
                 */
                if (m->m_pkthdr.len < 0) {
                        si->kflags |= DN_ACTIVE;
                        heap_insert(&dn_cfg.evheap, now - m->m_pkthdr.len, si);
                        if (delay_line_idle && done)
                                transmit_event(q, &si->dline, now);
                        return NULL;
                }

                /* a regular mbuf received */
                done++;
                len_scaled = (bw == 0) ? 0 : hz *
                        (m->m_pkthdr.len * 8 + extra_bits(m, s));
                si->credit -= len_scaled;
                /* Move packet in the delay line */
                dn_tag_get(m)->output_time = dn_cfg.curr_time + s->link.delay;
                mq_append(&si->dline.mq, m);
        }

        /*
         * If credit >= 0 the instance is idle, mark time.
         * Otherwise put back in the heap, and adjust the output
         * time of the last inserted packet, m, which was too early.
         */
        if (si->credit >= 0) {
                si->idle_time = now;
        } else {
                uint64_t t;
                KASSERT (bw > 0, ("bw=0 and credit<0 ?"));
                t = div64(bw - 1 - si->credit, bw);
                if (m)
                        dn_tag_get(m)->output_time += t;
                si->kflags |= DN_ACTIVE;
                heap_insert(&dn_cfg.evheap, now + t, si);
        }
        if (delay_line_idle && done)
                transmit_event(q, &si->dline, now);
        return q->head;
}

/*
 * Support function to read the TSC (or equivalent). We use this
 * high resolution timer to adapt the amount of work done for
 * expiring the clock.
 * Supports Linux and FreeBSD both i386 and amd64 platform
 * Supports OpenWRT mips architecture
 *
 * SMP no special works is needed in
 * - In linux 2.6 timers will always run in the same cpu that have added it.See
 * (http://book.opensourceproject.org.cn/kernel/kernel3rd/opensource/0596005652/understandlk-chp-6-sect-5.html)
 * - FreeBSD8 has a new callout_reset_on() with specify the cpu on which
 *   the timer must be run
 * - Windows runs dummynet_task() on cpu0.
 *
 * - Linux 2.4 doesn't assure to run a timer in the same cpu every time.
 */
#ifdef HAVE_TSC
uint64_t
readTSC (void)
{
        uint64_t a=0;

#ifdef __linux__
        /* Linux and openwrt have a macro to read the tsc for i386 and
         * amd64.
         * Openwrt have patched the kernel and allow use of tsc with mips
         * and other platforms
         * rdtscll() is a macro defined in include/asm-xxx/msr.h,
         * where xxx is the architecture (x86, mips).
         */
        rdtscll(a);
#elif defined(_WIN32)
        /* Microsoft recommends the use of KeQueryPerformanceCounter()
         * insteead of rdtsc().
         */
        KeQueryPerformanceCounter((PLARGE_INTEGER)&a);  //XXX not tested!
#elif defined(__FreeBSD__)
        /* FreeBSD (i386/amd64) has macro rdtsc() defined in machine/cpufunc.h.
         * We could use the macro instead of explicity assembly XXX
         */
        return rdtsc();
#endif
        return a;
}
#endif /* HAVE_TSC */

/*
 * compute avg task period.
 * We could do something more complex, possibly.
 */
static void
do_update_cycle(void)
{
#ifdef HAVE_TSC
        uint64_t tmp = readTSC();
#if defined (LINUX_24) && defined(CONFIG_SMP)
        /* on LINUX24 and SMP, we have no guarantees on which cpu runs
         * the timer callbacks. If the difference between new and
         * old value is negative, we assume that the values come from
         * different cpus so we adjust 'new' accordingly.
         */
        if (tmp <= dn_cfg.cycle_task_new)
                dn_cfg.cycle_task_new = tmp - dn_cfg.cycle_task;
#endif /* !(linux24 && SMP) */
        dn_cfg.cycle_task_old = dn_cfg.cycle_task_new;
        dn_cfg.cycle_task_new = tmp;
        dn_cfg.cycle_task = dn_cfg.cycle_task_new - dn_cfg.cycle_task_old;

        /* Update the average
         * avg = (2^N * avg + new - avg ) / 2^N * avg
         * N==4 seems to be a good compromise between clock clock change
         *      and 'spurious' cycle_task value
         */
#define DN_N    4
        dn_cfg.cycle_task_avg = (dn_cfg.cycle_task_avg << DN_N) +
                                dn_cfg.cycle_task - dn_cfg.cycle_task_avg;
        dn_cfg.cycle_task_avg = dn_cfg.cycle_task_avg >> DN_N;
#undef DN_N

#endif /* HAVE_TSC */
}

static void
do_drain(void)
{
#ifdef HAVE_TSC
        uint64_t dt_max;
#endif
        if (!dn_cfg.expire || ++dn_cfg.expire_cycle < dn_cfg.expire)
                return;
        /* It's time to check if drain routines should be called */
        dn_cfg.expire_cycle = 0;

        dn_cfg.idle_queue_wait = 0;
        dn_cfg.idle_si_wait = 0;
        /* Do a drain cycle even if there isn't time to do it */
#ifdef HAVE_TSC
        dt_max = dn_cfg.cycle_task_avg * dn_cfg.drain_ratio;
#endif
        for (;;) {
                int done = 0;

                if (dn_cfg.idle_queue > dn_cfg.expire_object &&
                    dn_cfg.idle_queue_wait < dn_cfg.idle_queue) {
                        dn_drain_queue();
                        done = 1;
                }
                if (dn_cfg.idle_si > dn_cfg.expire_object &&
                    dn_cfg.idle_si_wait < dn_cfg.idle_si) {
                        dn_drain_scheduler();
                        done = 1;
                }
                /* time to end ? */
#ifndef HAVE_TSC
                /* If tsc does not exist, do only one drain cycle and exit */
                break;
#else
                /* Exit when nothing was done or we have consumed all time */
                if ( (done == 0) || 
                     ((readTSC() -  dn_cfg.cycle_task_new) * 100 > dt_max) )
                        break;
#endif  /* HAVE_TSC */
        }
}

/*
 * The timer handler for dummynet. Time is computed in ticks, but
 * but the code is tolerant to the actual rate at which this is called.
 * Once complete, the function reschedules itself for the next tick.
 */
void
dummynet_task(void *context, int pending)
{
        struct timeval t;
        struct mq q = { NULL, NULL }; /* queue to accumulate results */

        CURVNET_SET((struct vnet *)context);

        do_update_cycle();      /* compute avg. tick duration */

        DN_BH_WLOCK();

        /* Update number of lost(coalesced) ticks. */
        tick_lost += pending - 1;

        getmicrouptime(&t);
        /* Last tick duration (usec). */
        tick_last = (t.tv_sec - dn_cfg.prev_t.tv_sec) * 1000000 +
        (t.tv_usec - dn_cfg.prev_t.tv_usec);
        /* Last tick vs standard tick difference (usec). */
        tick_delta = (tick_last * hz - 1000000) / hz;
        /* Accumulated tick difference (usec). */
        tick_delta_sum += tick_delta;

        dn_cfg.prev_t = t;

        /*
        * Adjust curr_time if the accumulated tick difference is
        * greater than the 'standard' tick. Since curr_time should
        * be monotonically increasing, we do positive adjustments
        * as required, and throttle curr_time in case of negative
        * adjustment.
        */
        dn_cfg.curr_time++;
        if (tick_delta_sum - tick >= 0) {
                int diff = tick_delta_sum / tick;

                dn_cfg.curr_time += diff;
                tick_diff += diff;
                tick_delta_sum %= tick;
                tick_adjustment++;
        } else if (tick_delta_sum + tick <= 0) {
                dn_cfg.curr_time--;
                tick_diff--;
                tick_delta_sum += tick;
                tick_adjustment++;
        }

        /* serve pending events, accumulate in q */
        for (;;) {
                struct dn_id *p;    /* generic parameter to handler */

                if (dn_cfg.evheap.elements == 0 ||
                    DN_KEY_LT(dn_cfg.curr_time, HEAP_TOP(&dn_cfg.evheap)->key))
                        break;
                p = HEAP_TOP(&dn_cfg.evheap)->object;
                heap_extract(&dn_cfg.evheap, NULL);

                if (p->type == DN_SCH_I) {
                        serve_sched(&q, (struct dn_sch_inst *)p, dn_cfg.curr_time);
                } else { /* extracted a delay line */
                        transmit_event(&q, (struct delay_line *)p, dn_cfg.curr_time);
                }
        }
        do_drain();

        DN_BH_WUNLOCK();
        dn_reschedule();
        if (q.head != NULL)
                dummynet_send(q.head);
        CURVNET_RESTORE();
}

/*
 * forward a chain of packets to the proper destination.
 * This runs outside the dummynet lock.
 */
static void
dummynet_send(struct mbuf *m)
{
        struct mbuf *n;

        for (; m != NULL; m = n) {
                struct ifnet *ifp = NULL;       /* gcc 3.4.6 complains */
                struct m_tag *tag;
                int dst;

                n = m->m_nextpkt;
                m->m_nextpkt = NULL;
                tag = m_tag_first(m);
                if (tag == NULL) { /* should not happen */
                        dst = DIR_DROP;
                } else {
                        struct dn_pkt_tag *pkt = dn_tag_get(m);
                        /* extract the dummynet info, rename the tag
                         * to carry reinject info.
                         */
                        dst = pkt->dn_dir;
                        ifp = pkt->ifp;
                        tag->m_tag_cookie = MTAG_IPFW_RULE;
                        tag->m_tag_id = 0;
                }

                switch (dst) {
                case DIR_OUT:
                        SET_HOST_IPLEN(mtod(m, struct ip *));
                        ip_output(m, NULL, NULL, IP_FORWARDING, NULL, NULL);
                        break ;

                case DIR_IN :
                        /* put header in network format for ip_input() */
                        //SET_NET_IPLEN(mtod(m, struct ip *));
                        netisr_dispatch(NETISR_IP, m);
                        break;

#ifdef INET6
                case DIR_IN | PROTO_IPV6:
                        netisr_dispatch(NETISR_IPV6, m);
                        break;

                case DIR_OUT | PROTO_IPV6:
                        SET_HOST_IPLEN(mtod(m, struct ip *));
                        ip6_output(m, NULL, NULL, IPV6_FORWARDING, NULL, NULL, NULL);
                        break;
#endif

                case DIR_FWD | PROTO_IFB: /* DN_TO_IFB_FWD: */
                        if (bridge_dn_p != NULL)
                                ((*bridge_dn_p)(m, ifp));
                        else
                                printf("dummynet: if_bridge not loaded\n");

                        break;

                case DIR_IN | PROTO_LAYER2: /* DN_TO_ETH_DEMUX: */
                        /*
                         * The Ethernet code assumes the Ethernet header is
                         * contiguous in the first mbuf header.
                         * Insure this is true.
                         */
                        if (m->m_len < ETHER_HDR_LEN &&
                            (m = m_pullup(m, ETHER_HDR_LEN)) == NULL) {
                                printf("dummynet/ether: pullup failed, "
                                    "dropping packet\n");
                                break;
                        }
                        ether_demux(m->m_pkthdr.rcvif, m);
                        break;

                case DIR_OUT | PROTO_LAYER2: /* N_TO_ETH_OUT: */
                        ether_output_frame(ifp, m);
                        break;

                case DIR_DROP:
                        /* drop the packet after some time */
                        FREE_PKT(m);
                        break;

                default:
                        printf("dummynet: bad switch %d!\n", dst);
                        FREE_PKT(m);
                        break;
                }
        }
}

static inline int
tag_mbuf(struct mbuf *m, int dir, struct ip_fw_args *fwa)
{
        struct dn_pkt_tag *dt;
        struct m_tag *mtag;

        mtag = m_tag_get(PACKET_TAG_DUMMYNET,
                    sizeof(*dt), M_NOWAIT | M_ZERO);
        if (mtag == NULL)
                return 1;               /* Cannot allocate packet header. */
        m_tag_prepend(m, mtag);         /* Attach to mbuf chain. */
        dt = (struct dn_pkt_tag *)(mtag + 1);
        dt->rule = fwa->rule;
        dt->rule.info &= IPFW_ONEPASS;  /* only keep this info */
        dt->dn_dir = dir;
        dt->ifp = fwa->oif;
        /* dt->output tame is updated as we move through */
        dt->output_time = dn_cfg.curr_time;
        return 0;
}


/*
 * dummynet hook for packets.
 * We use the argument to locate the flowset fs and the sched_set sch
 * associated to it. The we apply flow_mask and sched_mask to
 * determine the queue and scheduler instances.
 *
 * dir          where shall we send the packet after dummynet.
 * *m0          the mbuf with the packet
 * ifp          the 'ifp' parameter from the caller.
 *              NULL in ip_input, destination interface in ip_output,
 */
int
dummynet_io(struct mbuf **m0, int dir, struct ip_fw_args *fwa)
{
        struct mbuf *m = *m0;
        struct dn_fsk *fs = NULL;
        struct dn_sch_inst *si;
        struct dn_queue *q = NULL;      /* default */

        int fs_id = (fwa->rule.info & IPFW_INFO_MASK) +
                ((fwa->rule.info & IPFW_IS_PIPE) ? 2*DN_MAX_ID : 0);
        DN_BH_WLOCK();
        io_pkt++;
        /* we could actually tag outside the lock, but who cares... */
        if (tag_mbuf(m, dir, fwa))
                goto dropit;
        if (dn_cfg.busy) {
                /* if the upper half is busy doing something expensive,
                 * lets queue the packet and move forward
                 */
                mq_append(&dn_cfg.pending, m);
                m = *m0 = NULL; /* consumed */
                goto done; /* already active, nothing to do */
        }
        /* XXX locate_flowset could be optimised with a direct ref. */
        fs = dn_ht_find(dn_cfg.fshash, fs_id, 0, NULL);
        if (fs == NULL)
                goto dropit;    /* This queue/pipe does not exist! */
        if (fs->sched == NULL)  /* should not happen */
                goto dropit;
        /*
         * If the scheduler supports multiple queues, find the right one
         * (otherwise it will be ignored by enqueue).
         */
        if (fs->sched->fp->flags & DN_MULTIQUEUE) {
                q = ipdn_q_find(fs, &(fwa->f_id));
                if (q == NULL)
                        goto dropit;
                /* The scheduler instance lookup is done only for new queue.
                 * The callback q_new() will create the scheduler instance
                 * if needed.
                 */
                si = q->_si;
        } else
                si = ipdn_si_find(fs->sched, &(fwa->f_id));

        if (si == NULL)
                goto dropit;
        if (fs->sched->fp->enqueue(si, q, m)) {
                /* packet was dropped by enqueue() */
                m = *m0 = NULL;
                goto dropit;
        }

        if (si->kflags & DN_ACTIVE) {
                m = *m0 = NULL; /* consumed */
                goto done; /* already active, nothing to do */
        }

        /* compute the initial allowance */
        if (si->idle_time < dn_cfg.curr_time) {
            /* Do this only on the first packet on an idle pipe */
            struct dn_link *p = &fs->sched->link;

            si->sched_time = dn_cfg.curr_time;
            si->credit = dn_cfg.io_fast ? p->bandwidth : 0;
            if (p->burst) {
                uint64_t burst = (dn_cfg.curr_time - si->idle_time) * p->bandwidth;
                if (burst > p->burst)
                        burst = p->burst;
                si->credit += burst;
            }
        }
        /* pass through scheduler and delay line */
        m = serve_sched(NULL, si, dn_cfg.curr_time);

        /* optimization -- pass it back to ipfw for immediate send */
        /* XXX Don't call dummynet_send() if scheduler return the packet
         *     just enqueued. This avoid a lock order reversal.
         *     
         */
        if (/*dn_cfg.io_fast &&*/ m == *m0 && (dir & PROTO_LAYER2) == 0 ) {
                /* fast io, rename the tag * to carry reinject info. */
                struct m_tag *tag = m_tag_first(m);

                tag->m_tag_cookie = MTAG_IPFW_RULE;
                tag->m_tag_id = 0;
                io_pkt_fast++;
                if (m->m_nextpkt != NULL) {
                        printf("dummynet: fast io: pkt chain detected!\n");
                        m->m_nextpkt = NULL;
                }
                m = NULL;
        } else {
                *m0 = NULL;
        }
done:
        DN_BH_WUNLOCK();
        if (m)
                dummynet_send(m);
        return 0;

dropit:
        io_pkt_drop++;
        DN_BH_WUNLOCK();
        if (m)
                FREE_PKT(m);
        *m0 = NULL;
        return (fs && (fs->fs.flags & DN_NOERROR)) ? 0 : ENOBUFS;
}