pathmtu/pathmtu.c: Place `addreq' with the rest of the `raw' strategy.

[tripe] / pathmtu / pathmtu.c

/* -*-c-*-
 *
 * Report MTU on path to specified host
 *
 * (c) 2008 Straylight/Edgeware
 */

/*----- Licensing notice --------------------------------------------------*
 *
 * This file is part of Trivial IP Encryption (TrIPE).
 *
 * TrIPE 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 3 of the License, or (at your
 * option) any later version.
 *
 * TrIPE 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 TrIPE.  If not, see <https://www.gnu.org/licenses/>.
 */

/*----- Header files ------------------------------------------------------*/

#include "config.h"

#include <assert.h>
#include <errno.h>
#include <stddef.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <time.h>

#include <sys/types.h>
#include <sys/time.h>
#include <unistd.h>

#include <sys/socket.h>
#include <netinet/in.h>
#include <arpa/inet.h>
#include <netdb.h>

#include <netinet/in_systm.h>
#include <netinet/ip.h>
#include <netinet/ip_icmp.h>
#include <netinet/ip6.h>
#include <netinet/icmp6.h>
#include <netinet/udp.h>

#include <net/if.h>
#include <ifaddrs.h>
#include <sys/ioctl.h>

#include <mLib/alloc.h>
#include <mLib/bits.h>
#include <mLib/dstr.h>
#include <mLib/hex.h>
#include <mLib/mdwopt.h>
#include <mLib/quis.h>
#include <mLib/report.h>
#include <mLib/tv.h>

/*----- Static variables --------------------------------------------------*/

static unsigned char buf[65536];

#define POLY 0x1002d

/*----- Utility functions -------------------------------------------------*/

/* Step a value according to a simple LFSR. */
#define STEP(q)								\
  do (q) = ((q) & 0x8000) ? ((q) << 1) ^ POLY : ((q) << 1); while (0)

/* Fill buffer with a constant but pseudorandom string.  Uses a simple
 * LFSR.
 */
static void fillbuffer(unsigned char *p, size_t sz)
{
  unsigned int y = 0xbc20;
  const unsigned char *l = p + sz;
  int i;

  while (p < l) {
    *p++ = y & 0xff;
    for (i = 0; i < 8; i++) STEP(y);
  }
}

/* Convert a string to floating point. */
static double s2f(const char *s, const char *what)
{
  double f;
  char *q;

  errno = 0;
  f = strtod(s, &q);
  if (errno || *q) die(EXIT_FAILURE, "bad %s", what);
  return (f);
}

/* Convert a floating-point value into a struct timeval. */
static void f2tv(struct timeval *tv, double t)
  { tv->tv_sec = t; tv->tv_usec = (t - tv->tv_sec)*MILLION; }

union addr {
  struct sockaddr sa;
  struct sockaddr_in sin;
  struct sockaddr_in6 sin6;
};

/* Check whether an address family is even slightly supported. */
static int addrfamok(int af)
{
  switch (af) {
    case AF_INET: case AF_INET6: return (1);
    default: return (0);
  }
}

/* Return the size of a socket address. */
static size_t addrsz(const union addr *a)
{
  switch (a->sa.sa_family) {
    case AF_INET: return (sizeof(a->sin));
    case AF_INET6: return (sizeof(a->sin6));
    default: abort();
  }
}

/*----- Main algorithm skeleton -------------------------------------------*/

struct param {
  unsigned f;				/* Various flags */
#define F_VERBOSE 1u			/*   Give a running commentary */
  double retx;				/* Initial retransmit interval */
  double regr;				/* Retransmit growth factor */
  double timeout;			/* Retransmission timeout */
  int seqoff;				/* Offset to write sequence number */
  const struct probe_ops *pops;		/* Probe algorithm description */
  union addr a;				/* Destination address */
};

struct probestate {
  const struct param *pp;
  unsigned q;
};

struct probe_ops {
  const char *name;
  const struct probe_ops *next;
  size_t statesz;
  int (*setup)(void *, int, const struct param *);
  void (*finish)(void *);
  void (*selprep)(void *, int *, fd_set *);
  int (*xmit)(void *, int);
  int (*selproc)(void *, fd_set *, struct probestate *);
};

#define OPS_CHAIN 0

enum {
  RC_FAIL = -99,
  RC_OK = 0,
  RC_LOWER = -1,
  RC_HIGHER = -2,
  RC_NOREPLY = -3
  /* or a positive MTU upper-bound */
};

/* Add a file descriptor FD to the set `fd_in', updating `*maxfd'. */
#define ADDFD(fd) \
  do { FD_SET(fd, fd_in); if (*maxfd < fd) *maxfd = fd; } while (0)

/* Check whether a buffer contains a packet from our current probe. */
static int mypacketp(struct probestate *ps,
		     const unsigned char *p, size_t sz)
{
  const struct param *pp = ps->pp;

  return (sz >= pp->seqoff + 2 && LOAD16(p + pp->seqoff) == ps->q);
}

/* See whether MTU is an acceptable MTU value.  Return an appropriate
 * RC_... code or a new suggested MTU.
 */
static int probe(struct probestate *ps, void *st, int mtu)
{
  const struct param *pp = ps->pp;
  fd_set fd_in;
  struct timeval tv, now, when, done;
  double timer = pp->retx;
  int rc, maxfd;

  /* Set up the first retransmit and give-up timers. */
  gettimeofday(&now, 0);
  f2tv(&tv, pp->timeout); TV_ADD(&done, &now, &tv);
  f2tv(&tv, timer); TV_ADD(&when, &now, &tv);
  if (TV_CMP(&when, >, &done)) when = done;

  /* Send the initial probe. */
  if (pp->f & F_VERBOSE)
    moan("sending probe of size %d (seq = %04x)", mtu, ps->q);
  STEP(ps->q);
  STORE16(buf + pp->seqoff, ps->q);
  if ((rc = pp->pops->xmit(st, mtu)) != RC_OK) return (rc);

  for (;;) {

    /* Wait for something interesting to happen. */
    maxfd = 0; FD_ZERO(&fd_in);
    pp->pops->selprep(st, &maxfd, &fd_in);
    TV_SUB(&tv, &when, &now);
    if (select(maxfd + 1, &fd_in, 0, 0, &tv) < 0) return (RC_FAIL);
    gettimeofday(&now, 0);

    /* See whether the probe method has any answers for us. */
    if ((rc = pp->pops->selproc(st, &fd_in, ps)) != RC_OK) return (rc);

    /* If we've waited too long, give up.  If we should retransmit, do
     * that.
     */
    if (TV_CMP(&now, >, &done))
      return (RC_NOREPLY);
    else if (TV_CMP(&now, >, &when)) {
      if (pp->f & F_VERBOSE) moan("re-sending probe of size %d", mtu);
      if ((rc = pp->pops->xmit(st, mtu)) != RC_OK) return (rc);
      do {
	timer *= pp->regr; f2tv(&tv, timer); TV_ADD(&when, &when, &tv);
      } while (TV_CMP(&when, <, &now));
      if (TV_CMP(&when, >, &done)) when = done;
    }
  }
}

/* Discover the path MTU to the destination address. */
static int pathmtu(const struct param *pp)
{
  int sk;
  int mtu, lo, hi;
  int rc, droppy = -1;
  void *st;
  struct probestate ps;

  /* Build and connect a UDP socket.  We'll need this to know the local port
   * number to use if nothing else.  Set other stuff up.
   */
  if ((sk = socket(pp->a.sa.sa_family, SOCK_DGRAM, IPPROTO_UDP)) < 0)
    goto fail_0;
  if (connect(sk, &pp->a.sa, addrsz(&pp->a))) goto fail_1;
  st = xmalloc(pp->pops->statesz);
  if ((mtu = pp->pops->setup(st, sk, pp)) < 0) goto fail_2;
  ps.pp = pp; ps.q = rand() & 0xffff;
  switch (pp->a.sa.sa_family) {
    case AF_INET: lo = 576; break;
    case AF_INET6: lo = 1280; break;
    default: abort();
  }
  hi = mtu;
  if (hi < lo) { errno = EMSGSIZE; return (-1); }

  /* And now we do a thing which is sort of like a binary search, except that
   * we also take explicit clues as establishing a new upper bound, and we
   * try to hug that initially.
   */
  for (;;) {
    assert(lo <= mtu && mtu <= hi);
    if (pp->f & F_VERBOSE) moan("probe: %d <= %d <= %d", lo, mtu, hi);
    rc = probe(&ps, st, mtu);
    switch (rc) {

      case RC_FAIL:
	if (pp->f & F_VERBOSE) moan("probe failed");
	goto fail_3;

      case RC_NOREPLY:
	/* If we've not seen a dropped packet before then we don't know what
	 * this means yet -- in particular, we don't know which bit of the
	 * network is swallowing packets.  Send a minimum-size probe.  If
	 * that doesn't come back then assume that the remote host is
	 * swallowing our packets.  If it does, then we assume that dropped
	 * packets are a result of ICMP fragmentation-needed reports being
	 * lost or suppressed.
	 */
	if (pp->f & F_VERBOSE) moan("gave up: black hole detected");
	if (droppy == -1) {
	  if (pp->f & F_VERBOSE) moan("sending minimum-size probe");
	  switch (probe(&ps, st, lo)) {
	    case RC_FAIL:
	      goto fail_3;
	    case RC_NOREPLY:
	      if (pp->f & F_VERBOSE) {
		moan("no reply from min-size probe: "
		     "assume black hole at target");
	      }
	      droppy = 1;
	      break;
	    case RC_HIGHER:
	      if (pp->f & F_VERBOSE) {
		moan("reply from min-size probe OK: "
		     "assume black hole in network");
	      }
	      droppy = 0;
	      break;
	    default:
	      if (pp->f & F_VERBOSE)
		moan("unexpected return code from probe");
	      errno = ENOTCONN;
	      goto fail_3;
	  }
	}

	if (droppy) goto higher; else goto lower;

      case RC_HIGHER:
      higher:
	if (droppy == -1) {
	  if (pp->f & F_VERBOSE)
	    moan("probe returned: remote host is not a black hole");
	  droppy = 0;
	}
	if (mtu == hi) {
	  if (pp->f & F_VERBOSE) moan("probe returned: found correct MTU");
	  goto done;
	}
	lo = mtu;

	/* Now we must make a new guess, between lo and hi.  We know that lo
	 * is good; but we're not so sure about hi here.  We know that hi >
	 * lo, so this will find an approximate midpoint, greater than lo and
	 * no more than hi.
	 */
	if (pp->f & F_VERBOSE) moan("probe returned: guessing higher");
	mtu += (hi - lo + 1)/2;
	break;

      case RC_LOWER:
      lower:
	/* If this didn't work, and we're already at the bottom of our
	 * possible range, then something has gone horribly wrong.
	 */
	assert(lo < mtu);
	hi = mtu - 1;
	if (lo == hi) {
	  if (pp->f & F_VERBOSE) moan("error returned: found correct MTU");
	  mtu = lo;
	  goto done;
	}

	/* We must make a new guess, between lo and hi.  We're probably
	 * fairly sure that lo will succeed, since either it's the minimum
	 * MTU or we've tested it already; but we're not quite sure about hi,
	 * so we want to aim high.
	 */
	if (pp->f & F_VERBOSE) moan("error returned: guessing lower");
	mtu -= (hi - lo + 1)/2;
	break;

      default:
	if (pp->f & F_VERBOSE) moan("error returned with new MTU estimate");
	mtu = hi = rc;
	break;
    }
  }

done:
  /* Clean up and return our result. */
  pp->pops->finish(st);
  xfree(st);
  close(sk);
  return (mtu);

fail_3:
  pp->pops->finish(st);
fail_2:
  xfree(st);
fail_1:
  close(sk);
fail_0:
  return (-1);
}

/*----- Doing it the hard way ---------------------------------------------*/

#if defined(linux) || defined(__OpenBSD__)
#  define IPHDR_SANE
#endif

#ifdef IPHDR_SANE
#  define sane_htons htons
#  define sane_htonl htonl
#else
#  define sane_htons
#  define sane_htonl
#endif

static int rawicmp = -1, rawudp = -1, rawerr = 0;
static int rawicmp6 = -1, rawudp6 = -1, rawerr6 = 0;

#define IPCK_INIT 0xffff

/* Compare two addresses.  Maybe compare the port numbers too. */
#define AEF_PORT 1u
static int addreq(const union addr *a, const union addr *b, unsigned f)
{
  switch (a->sa.sa_family) {
    case AF_INET:
      return (a->sin.sin_addr.s_addr == b->sin.sin_addr.s_addr &&
	      (!(f&AEF_PORT) || a->sin.sin_port == b->sin.sin_port));
    case AF_INET6:
      return (!memcmp(a->sin6.sin6_addr.s6_addr,
		      b->sin6.sin6_addr.s6_addr, 16) &&
	      (!(f&AEF_PORT) || a->sin6.sin6_port == b->sin6.sin6_port));
    default:
      abort();
  }
}

/* Compute an IP checksum over some data.  This is a restartable interface:
 * initialize A to `IPCK_INIT' for the first call.
 */
static unsigned ipcksum(const void *buf, size_t n, unsigned a)
{
  unsigned long aa = a ^ 0xffff;
  const unsigned char *p = buf, *l = p + n;

  while (p < l - 1) { aa += LOAD16_B(p); p += 2; }
  if (p < l) { aa += (unsigned)(*p) << 8; }
  do aa = (aa & 0xffff) + (aa >> 16); while (aa >= 0x10000);
  return (aa == 0xffff ? aa : aa ^ 0xffff);
}

/* TCP/UDP pseudoheader structure. */
struct phdr {
  struct in_addr ph_src, ph_dst;
  uint8_t ph_z, ph_p;
  uint16_t ph_len;
};
struct phdr6 {
  struct in6_addr ph6_src, ph6_dst;
  uint32_t ph6_len;
  uint8_t ph6_z0, ph6_z1, ph6_z2, ph6_nxt;
};

struct raw_state {
  union addr me, a;
  int sk, rawicmp, rawudp;
  uint16_t srcport, dstport;
  unsigned q;
};

static int raw_setup(void *stv, int sk, const struct param *pp)
{
  struct raw_state *st = stv;
  socklen_t sz;
  int i, mtu = -1;
  struct ifaddrs *ifa, *ifaa, *ifap;
  struct ifreq ifr;
  struct icmp6_filter f6;

  /* Check that the address is OK, and that we have the necessary raw
   * sockets.
   *
   * For IPv6, also set the filter so we don't get too many useless wakeups.
   */
  switch (pp->a.sa.sa_family) {
    case AF_INET:
      if (rawerr) { errno = rawerr; goto fail_0; }
      st->rawicmp = rawicmp; st->rawudp = rawudp; st->sk = sk;
      /* IPv4 filtering is available on Linux but isn't portable. */
      break;
    case AF_INET6:
      if (rawerr6) { errno = rawerr6; goto fail_0; }
      st->rawicmp = rawicmp6; st->rawudp = rawudp6; st->sk = sk;
      ICMP6_FILTER_SETBLOCKALL(&f6);
      ICMP6_FILTER_SETPASS(ICMP6_PACKET_TOO_BIG, &f6);
      ICMP6_FILTER_SETPASS(ICMP6_DST_UNREACH, &f6);
      if (setsockopt(st->rawicmp, IPPROTO_ICMPV6, ICMP6_FILTER,
		     &f6, sizeof(f6))) {
	die(EXIT_FAILURE, "failed to set icmpv6 filter: %s",
	    strerror(errno));
      }
      break;
    default:
      errno = EPFNOSUPPORT; goto fail_0;
  }

  /* Initialize the sequence number. */
  st->q = rand() & 0xffff;

  /* Snaffle the local and remote address and port number. */
  st->a = pp->a;
  sz = sizeof(st->me);
  if (getsockname(sk, &st->me.sa, &sz))
    goto fail_0;

  /* Only now do some fiddling because Linux doesn't like port numbers in
   * IPv6 raw destination addresses...
   */
  switch (pp->a.sa.sa_family) {
    case AF_INET:
      st->srcport = st->me.sin.sin_port; st->me.sin.sin_port = 0;
      st->dstport =  st->a.sin.sin_port;  st->a.sin.sin_port = 0;
      break;
    case AF_INET6:
      st->srcport = st->me.sin6.sin6_port; st->me.sin6.sin6_port = 0;
      st->dstport =  st->a.sin6.sin6_port;  st->a.sin6.sin6_port = 0;
      break;
    default:
      abort();
  }

  /* There isn't a portable way to force the DF flag onto a packet through
   * UDP, or even through raw IP, unless we write the entire IP header
   * ourselves.  This is somewhat annoying, especially since we have an
   * uphill struggle keeping track of which systems randomly expect which
   * header fields to be presented in host byte order.  Oh, well.
   */
  i = 1;
  if (setsockopt(rawudp, IPPROTO_IP, IP_HDRINCL, &i, sizeof(i))) goto fail_0;

  /* Find an upper bound on the MTU.  Do two passes over the interface
   * list.  If we can find matches for our local address then use the
   * highest one of those; otherwise do a second pass and simply take the
   * highest MTU of any network interface.
   */
  if (getifaddrs(&ifaa)) goto fail_0;
  for (i = 0; i < 2; i++) {
    for (ifap = 0, ifa = ifaa; ifa; ifa = ifa->ifa_next) {
      if (!(ifa->ifa_flags & IFF_UP) || !ifa->ifa_addr ||
	  ifa->ifa_addr->sa_family != st->me.sa.sa_family ||
	  (i == 0 &&
	   !addreq((union addr *)ifa->ifa_addr, &st->me, 0)) ||
	  (i == 1 && ifap && strcmp(ifap->ifa_name, ifa->ifa_name) == 0) ||
	  strlen(ifa->ifa_name) >= sizeof(ifr.ifr_name))
	continue;
      ifap = ifa;
      strcpy(ifr.ifr_name, ifa->ifa_name);
      if (ioctl(sk, SIOCGIFMTU, &ifr)) goto fail_1;
      if (mtu < ifr.ifr_mtu) mtu = ifr.ifr_mtu;
    }
    if (mtu > 0) break;
  }
  if (mtu < 0) { errno = ENOTCONN; goto fail_1; }
  freeifaddrs(ifaa);

  /* Done. */
  return (mtu);

fail_1:
  freeifaddrs(ifaa);
fail_0:
  return (-1);
}

static void raw_finish(void *stv) { ; }

static void raw_selprep(void *stv, int *maxfd, fd_set *fd_in)
  { struct raw_state *st = stv; ADDFD(st->sk); ADDFD(st->rawicmp); }

static int raw_xmit(void *stv, int mtu)
{
  struct raw_state *st = stv;
  unsigned char b[65536], *p;
  struct ip *ip;
  struct ip6_hdr *ip6;
  struct udphdr *udp;
  struct phdr ph;
  struct phdr6 ph6;
  unsigned ck;

  switch (st->a.sa.sa_family) {

    case AF_INET:

      /* Build the IP header. */
      ip = (struct ip *)b;
      ip->ip_v = 4;
      ip->ip_hl = sizeof(*ip)/4;
      ip->ip_tos = IPTOS_RELIABILITY;
      ip->ip_len = sane_htons(mtu);
      STEP(st->q); ip->ip_id = htons(st->q);
      ip->ip_off = sane_htons(0 | IP_DF);
      ip->ip_ttl = 64;
      ip->ip_p = IPPROTO_UDP;
      ip->ip_sum = 0;
      ip->ip_src = st->me.sin.sin_addr;
      ip->ip_dst = st->a.sin.sin_addr;

      /* Build a UDP packet in the output buffer. */
      udp = (struct udphdr *)(ip + 1);
      udp->uh_sport = st->srcport;
      udp->uh_dport = st->dstport;
      udp->uh_ulen = htons(mtu - sizeof(*ip));
      udp->uh_sum = 0;

      /* Copy the payload. */
      p = (unsigned char *)(udp + 1);
      memcpy(p, buf, mtu - (p - b));

      /* Calculate the UDP checksum. */
      ph.ph_src = ip->ip_src;
      ph.ph_dst = ip->ip_dst;
      ph.ph_z = 0;
      ph.ph_p = IPPROTO_UDP;
      ph.ph_len = udp->uh_ulen;
      ck = IPCK_INIT;
      ck = ipcksum(&ph, sizeof(ph), ck);
      ck = ipcksum(udp, mtu - sizeof(*ip), ck);
      udp->uh_sum = htons(ck);

      break;

    case AF_INET6:

      /* Build the IP header. */
      ip6 = (struct ip6_hdr *)b;
      STEP(st->q); ip6->ip6_flow = htonl(0x60000000 | st->q);
      ip6->ip6_plen = htons(mtu - sizeof(*ip6));
      ip6->ip6_nxt = IPPROTO_UDP;
      ip6->ip6_hlim = 64;
      ip6->ip6_src = st->me.sin6.sin6_addr;
      ip6->ip6_dst = st->a.sin6.sin6_addr;

      /* Build a UDP packet in the output buffer. */
      udp = (struct udphdr *)(ip6 + 1);
      udp->uh_sport = st->srcport;
      udp->uh_dport = st->dstport;
      udp->uh_ulen = htons(mtu - sizeof(*ip6));
      udp->uh_sum = 0;

      /* Copy the payload. */
      p = (unsigned char *)(udp + 1);
      memcpy(p, buf, mtu - (p - b));

      /* Calculate the UDP checksum. */
      ph6.ph6_src = ip6->ip6_src;
      ph6.ph6_dst = ip6->ip6_dst;
      ph6.ph6_len = udp->uh_ulen;
      ph6.ph6_z0 = ph6.ph6_z1 = ph6.ph6_z2 = 0;
      ph6.ph6_nxt = IPPROTO_UDP;
      ck = IPCK_INIT;
      ck = ipcksum(&ph6, sizeof(ph6), ck);
      ck = ipcksum(udp, mtu - sizeof(*ip6), ck);
      udp->uh_sum = htons(ck);

      break;

    default:
      abort();
  }

  /* Send the whole thing off.  If we're too big for the interface then we
   * might need to trim immediately.
   */
  if (sendto(st->rawudp, b, mtu, 0, &st->a.sa, addrsz(&st->a)) < 0) {
    if (errno == EMSGSIZE) return (RC_LOWER);
    else goto fail_0;
  }

  /* Done. */
  return (RC_OK);

fail_0:
  return (RC_FAIL);
}

static int raw_selproc(void *stv, fd_set *fd_in, struct probestate *ps)
{
  struct raw_state *st = stv;
  unsigned char b[65536];
  struct ip *ip;
  struct ip6_hdr *ip6;
  struct icmp *icmp;
  struct icmp6_hdr *icmp6;
  struct udphdr *udp;
  const unsigned char *payload;
  ssize_t n;

  /* An ICMP packet: see what's inside. */
  if (FD_ISSET(st->rawicmp, fd_in)) {
    if ((n = read(st->rawicmp, b, sizeof(b))) < 0) goto fail_0;

    switch (st->me.sa.sa_family) {

      case AF_INET:

	ip = (struct ip *)b;
	if (n < sizeof(*ip) || n < sizeof(4*ip->ip_hl) ||
	    ip->ip_v != 4 || ip->ip_p != IPPROTO_ICMP)
	  goto skip_icmp;
	n -= sizeof(4*ip->ip_hl);

	icmp = (struct icmp *)(b + 4*ip->ip_hl);
	if (n < sizeof(*icmp) || icmp->icmp_type != ICMP_UNREACH)
	  goto skip_icmp;
	n -= offsetof(struct icmp, icmp_ip);

	ip = &icmp->icmp_ip;
	if (n < sizeof(*ip) ||
	    ip->ip_p != IPPROTO_UDP || ip->ip_hl != sizeof(*ip)/4 ||
	    ip->ip_id != htons(st->q) ||
	    ip->ip_src.s_addr != st->me.sin.sin_addr.s_addr ||
	    ip->ip_dst.s_addr != st->a.sin.sin_addr.s_addr)
	  goto skip_icmp;
	n -= sizeof(*ip);

	udp = (struct udphdr *)(ip + 1);
	if (n < sizeof(*udp) || udp->uh_sport != st->srcport ||
	    udp->uh_dport != st->dstport)
	  goto skip_icmp;
	n -= sizeof(*udp);

	payload = (const unsigned char *)(udp + 1);
	if (!mypacketp(ps, payload, n)) goto skip_icmp;

	if (icmp->icmp_code == ICMP_UNREACH_PORT) return (RC_HIGHER);
	else if (icmp->icmp_code != ICMP_UNREACH_NEEDFRAG) goto skip_icmp;
	else if (icmp->icmp_nextmtu) return (htons(icmp->icmp_nextmtu));
	else return (RC_LOWER);

	break;

      case AF_INET6:
	icmp6 = (struct icmp6_hdr *)b;
	if (n < sizeof(*icmp6) ||
	    (icmp6->icmp6_type != ICMP6_PACKET_TOO_BIG &&
	     icmp6->icmp6_type != ICMP6_DST_UNREACH))
	  goto skip_icmp;
	n -= sizeof(*icmp6);

	ip6 = (struct ip6_hdr *)(icmp6 + 1);
	if (n < sizeof(*ip6) || ip6->ip6_nxt != IPPROTO_UDP ||
	    memcmp(ip6->ip6_src.s6_addr,
		   st->me.sin6.sin6_addr.s6_addr, 16) ||
	    memcmp(ip6->ip6_dst.s6_addr,
		   st->a.sin6.sin6_addr.s6_addr, 16) ||
	    (ntohl(ip6->ip6_flow)&0xffff) != st->q)
	  goto skip_icmp;
	n -= sizeof(*ip6);

	udp = (struct udphdr *)(ip6 + 1);
	if (n < sizeof(*udp) || udp->uh_sport != st->srcport ||
	    udp->uh_dport != st->dstport)
	  goto skip_icmp;
	n -= sizeof(*udp);

	payload = (const unsigned char *)(udp + 1);
	if (!mypacketp(ps, payload, n)) goto skip_icmp;

	if (icmp6->icmp6_type == ICMP6_PACKET_TOO_BIG)
	  return (ntohs(icmp6->icmp6_mtu));
	else switch (icmp6->icmp6_code) {
	    case ICMP6_DST_UNREACH_ADMIN:
	    case ICMP6_DST_UNREACH_NOPORT:
	      return (RC_HIGHER);
	    default:
	      goto skip_icmp;
	  }
	break;

      default:
	abort();
    }
  }

skip_icmp:;

  /* If we got a reply to the current probe then we're good.  If we got an
   * error, or the packet's sequence number is wrong, then ignore it.
   */
  if (FD_ISSET(st->sk, fd_in)) {
    if ((n = read(st->sk, b, sizeof(b))) < 0) return (RC_OK);
    else if (mypacketp(ps, b, n)) return (RC_HIGHER);
    else return (RC_OK);
  }

  return (RC_OK);

fail_0:
  return (RC_FAIL);
}

static const struct probe_ops raw_ops = {
  "raw", OPS_CHAIN, sizeof(struct raw_state),
  raw_setup, raw_finish,
  raw_selprep, raw_xmit, raw_selproc
};

#undef OPS_CHAIN
#define OPS_CHAIN &raw_ops

/*----- Doing the job on Linux --------------------------------------------*/

#if defined(linux)

#ifndef IP_MTU
#  define IP_MTU 14			/* Blech! */
#endif

struct linux_state {
  int sol, so_mtu_discover, so_mtu;
  int sk;
  size_t hdrlen;
};

static int linux_setup(void *stv, int sk, const struct param *pp)
{
  struct linux_state *st = stv;
  int i, mtu;
  socklen_t sz;

  /* Check that the address is OK. */
  switch (pp->a.sa.sa_family) {
    case AF_INET:
      st->sol = IPPROTO_IP;
      st->so_mtu_discover = IP_MTU_DISCOVER;
      st->so_mtu = IP_MTU;
      st->hdrlen = 28;
      break;
    case AF_INET6:
      st->sol = IPPROTO_IPV6;
      st->so_mtu_discover = IPV6_MTU_DISCOVER;
      st->so_mtu = IPV6_MTU;
      st->hdrlen = 48;
      break;
    default:
      errno = EPFNOSUPPORT;
      return (-1);
  }

  /* Snaffle the UDP socket. */
  st->sk = sk;

  /* Turn on kernel path-MTU discovery and force DF on. */
  i = IP_PMTUDISC_PROBE;
  if (setsockopt(st->sk, st->sol, st->so_mtu_discover, &i, sizeof(i)))
    return (-1);

  /* Read the initial MTU guess back and report it. */
  sz = sizeof(mtu);
  if (getsockopt(st->sk, st->sol, st->so_mtu, &mtu, &sz))
    return (-1);

  /* Done. */
  return (mtu);
}

static void linux_finish(void *stv) { ; }

static void linux_selprep(void *stv, int *maxfd, fd_set *fd_in)
  { struct linux_state *st = stv; ADDFD(st->sk); }

static int linux_xmit(void *stv, int mtu)
{
  struct linux_state *st = stv;

  /* Write the packet. */
  if (write(st->sk, buf, mtu - st->hdrlen) >= 0) return (RC_OK);
  else if (errno == EMSGSIZE) return (RC_LOWER);
  else return (RC_FAIL);
}

static int linux_selproc(void *stv, fd_set *fd_in, struct probestate *ps)
{
  struct linux_state *st = stv;
  int mtu;
  socklen_t sz;
  ssize_t n;
  unsigned char b[65536];

  /* Read an answer.  If it looks like the right kind of error then report a
   * success.  This is potentially wrong, since we can't tell whether an
   * error was delayed from an earlier probe.  However, we never return
   * RC_LOWER from this method, so the packet sizes ought to be monotonically
   * decreasing and this won't cause trouble.  Otherwise update from the
   * kernel's idea of the right MTU.
   */
  if (FD_ISSET(st->sk, fd_in)) {
    n = read(st->sk, &buf, sizeof(buf));
    if (n >= 0 ?
	mypacketp(ps, b, n) :
	errno == ECONNREFUSED || errno == EHOSTUNREACH)
      return (RC_HIGHER);
    sz = sizeof(mtu);
    if (getsockopt(st->sk, st->sol, st->so_mtu, &mtu, &sz))
      return (RC_FAIL);
    return (mtu);
  }
  return (RC_OK);
}

static const struct probe_ops linux_ops = {
  "linux", OPS_CHAIN, sizeof(struct linux_state),
  linux_setup, linux_finish,
  linux_selprep, linux_xmit, linux_selproc
};

#undef OPS_CHAIN
#define OPS_CHAIN &linux_ops

#endif

/*----- Help options ------------------------------------------------------*/

static const struct probe_ops *probe_ops = OPS_CHAIN;

static void version(FILE *fp)
  { pquis(fp, "$, TrIPE version " VERSION "\n"); }

static void usage(FILE *fp)
{
  pquis(fp, "Usage: $ [-46v] [-H HEADER] [-m METHOD]\n\
	 [-r SECS] [-g FACTOR] [-t SECS] HOST [PORT]\n");
}

static void help(FILE *fp)
{
  const struct probe_ops *ops;

  version(fp);
  fputc('\n', fp);
  usage(fp);
  fputs("\
\n\
Options in full:\n\
\n\
-h, --help		Show this help text.\n\
-V, --version		Show version number.\n\
-u, --usage		Show brief usage message.\n\
\n\
-4, --ipv4		Restrict to IPv4 only.\n\
-6, --ipv6		Restrict to IPv6 only.\n\
-g, --growth=FACTOR	Growth factor for retransmit interval.\n\
-m, --method=METHOD	Use METHOD to probe for MTU.\n\
-r, --retransmit=SECS	Retransmit if no reply after SEC.\n\
-t, --timeout=SECS	Give up expecting a reply after SECS.\n\
-v, --verbose		Write a running commentary to stderr.\n\
-H, --header=HEX	Packet header, in hexadecimal.\n\
\n\
Probe methods:\n\
", fp);
  for (ops = probe_ops; ops; ops = ops->next)
    printf("\t%s\n", ops->name);
}

/*----- Main code ---------------------------------------------------------*/

int main(int argc, char *argv[])
{
  struct param pp = { 0, 0.333, 3.0, 8.0, 0, OPS_CHAIN };
  hex_ctx hc;
  dstr d = DSTR_INIT;
  size_t sz;
  int i, err;
  struct addrinfo aihint = { 0 }, *ailist, *ai;
  const char *host, *svc = "7";
  unsigned f = 0;

#define f_bogus 1u

  if ((rawicmp = socket(PF_INET, SOCK_RAW, IPPROTO_ICMP)) < 0 ||
      (rawudp = socket(PF_INET, SOCK_RAW, IPPROTO_UDP)) < 0)
    rawerr = errno;
  if ((rawicmp6 = socket(PF_INET6, SOCK_RAW, IPPROTO_ICMPV6)) < 0 ||
      (rawudp6 = socket(PF_INET6, SOCK_RAW, IPPROTO_RAW)) < 0)
    rawerr6 = errno;
  if (setuid(getuid()))
    abort();

  ego(argv[0]);
  fillbuffer(buf, sizeof(buf));

  aihint.ai_family = AF_UNSPEC;
  aihint.ai_protocol = IPPROTO_UDP;
  aihint.ai_socktype = SOCK_DGRAM;
  aihint.ai_flags = AI_ADDRCONFIG;

  for (;;) {
    static const struct option opts[] = {
      { "help",		0,		0,	'h' },
      { "version",	0,		0,	'V' },
      { "usage",	0,		0,	'u' },
      { "ipv4",		0,		0,	'4' },
      { "ipv6",		0,		0,	'6' },
      { "header",	OPTF_ARGREQ,	0,	'H' },
      { "growth",	OPTF_ARGREQ,	0,	'g' },
      { "method",	OPTF_ARGREQ,	0,	'm' },
      { "retransmit",	OPTF_ARGREQ,	0,	'r' },
      { "timeout",	OPTF_ARGREQ,	0,	't' },
      { "verbose",	0,		0,	'v' },
      { 0,		0,		0,	0 }
    };

    i = mdwopt(argc, argv, "hVu" "46H:g:m:r:t:v", opts, 0, 0, 0);
    if (i < 0) break;
    switch (i) {
      case 'h': help(stdout); exit(0);
      case 'V': version(stdout); exit(0);
      case 'u': usage(stdout); exit(0);

      case 'H':
	DRESET(&d);
	hex_init(&hc);
	hex_decode(&hc, optarg, strlen(optarg), &d);
	hex_decode(&hc, 0, 0, &d);
	sz = d.len < 532 ? d.len : 532;
	memcpy(buf, d.buf, sz);
	pp.seqoff = sz;
	break;

      case '4': aihint.ai_family = AF_INET; break;
      case '6': aihint.ai_family = AF_INET6; break;
      case 'g': pp.regr = s2f(optarg, "retransmit growth factor"); break;
      case 'r': pp.retx = s2f(optarg, "retransmit interval"); break;
      case 't': pp.timeout = s2f(optarg, "timeout"); break;

      case 'm':
	for (pp.pops = OPS_CHAIN; pp.pops; pp.pops = pp.pops->next)
	  if (strcmp(pp.pops->name, optarg) == 0) goto found_alg;
	die(EXIT_FAILURE, "unknown probe algorithm `%s'", optarg);
      found_alg:
	break;

      case 'v': pp.f |= F_VERBOSE; break;

      default:
	f |= f_bogus;
	break;
    }
  }
  argv += optind; argc -= optind;
  if ((f & f_bogus) || 1 > argc || argc > 2) {
    usage(stderr);
    exit(EXIT_FAILURE);
  }

  host = argv[0];
  if (argv[1]) svc = argv[1];
  if ((err = getaddrinfo(host, svc, &aihint, &ailist)) != 0) {
    die(EXIT_FAILURE, "unknown host `%s' or service `%s': %s",
	host, svc, gai_strerror(err));
  }
  for (ai = ailist; ai && !addrfamok(ai->ai_family); ai = ai->ai_next);
  if (!ai) die(EXIT_FAILURE, "no supported address families for `%s'", host);
  assert(ai->ai_addrlen <= sizeof(pp.a));
  memcpy(&pp.a, ai->ai_addr, ai->ai_addrlen);

  i = pathmtu(&pp);
  if (i < 0)
    die(EXIT_FAILURE, "failed to discover MTU: %s", strerror(errno));
  printf("%d\n", i);
  if (ferror(stdout) || fflush(stdout) || fclose(stdout))
    die(EXIT_FAILURE, "failed to write result: %s", strerror(errno));
  return (0);
}

/*----- That's all, folks -------------------------------------------------*/