mdw@git.distorted.org.uk Git - tripe/blob - pathmtu/pathmtu.c

   1 /* -*-c-*-
   2  *
   3  * Report MTU on path to specified host
   4  *
   5  * (c) 2008 Straylight/Edgeware
   6  */
   7
   8 /*----- Licensing notice --------------------------------------------------*
   9  *
  10  * This file is part of Trivial IP Encryption (TrIPE).
  11  *
  12  * TrIPE is free software: you can redistribute it and/or modify it under
  13  * the terms of the GNU General Public License as published by the Free
  14  * Software Foundation; either version 3 of the License, or (at your
  15  * option) any later version.
  16  *
  17  * TrIPE is distributed in the hope that it will be useful, but WITHOUT
  18  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  19  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  20  * for more details.
  21  *
  22  * You should have received a copy of the GNU General Public License
  23  * along with TrIPE.  If not, see <https://www.gnu.org/licenses/>.
  24  */
  25
  26 /*----- Header files ------------------------------------------------------*/
  27
  28 #include "config.h"
  29
  30 #include <assert.h>
  31 #include <errno.h>
  32 #include <stddef.h>
  33 #include <stdio.h>
  34 #include <stdlib.h>
  35 #include <string.h>
  36 #include <time.h>
  37
  38 #include <sys/types.h>
  39 #include <sys/time.h>
  40 #include <unistd.h>
  41
  42 #include <sys/socket.h>
  43 #include <netinet/in.h>
  44 #include <arpa/inet.h>
  45 #include <netdb.h>
  46
  47 #include <netinet/in_systm.h>
  48 #include <netinet/ip.h>
  49 #include <netinet/ip_icmp.h>
  50 #include <netinet/udp.h>
  51
  52 #include <net/if.h>
  53 #include <ifaddrs.h>
  54 #include <sys/ioctl.h>
  55
  56 #include <mLib/alloc.h>
  57 #include <mLib/bits.h>
  58 #include <mLib/dstr.h>
  59 #include <mLib/hex.h>
  60 #include <mLib/mdwopt.h>
  61 #include <mLib/quis.h>
  62 #include <mLib/report.h>
  63 #include <mLib/tv.h>
  64
  65 /*----- Static variables --------------------------------------------------*/
  66
  67 static unsigned char buf[65536];
  68
  69 #define POLY 0x1002d
  70
  71 /*----- Utility functions -------------------------------------------------*/
  72
  73 /* Step a value according to a simple LFSR. */
  74 #define STEP(q)                                                         \
  75   do (q) = ((q) & 0x8000) ? ((q) << 1) ^ POLY : ((q) << 1); while (0)
  76
  77 /* Fill buffer with a constant but pseudorandom string.  Uses a simple
  78  * LFSR.
  79  */
  80 static void fillbuffer(unsigned char *p, size_t sz)
  81 {
  82   unsigned int y = 0xbc20;
  83   const unsigned char *l = p + sz;
  84   int i;
  85
  86   while (p < l) {
  87     *p++ = y & 0xff;
  88     for (i = 0; i < 8; i++) STEP(y);
  89   }
  90 }
  91
  92 /* Convert a string to floating point. */
  93 static double s2f(const char *s, const char *what)
  94 {
  95   double f;
  96   char *q;
  97
  98   errno = 0;
  99   f = strtod(s, &q);
 100   if (errno || *q) die(EXIT_FAILURE, "bad %s", what);
 101   return (f);
 102 }
 103
 104 /* Convert a floating-point value into a struct timeval. */
 105 static void f2tv(struct timeval *tv, double t)
 106   { tv->tv_sec = t; tv->tv_usec = (t - tv->tv_sec)*MILLION; }
 107
 108 union addr {
 109   struct sockaddr sa;
 110   struct sockaddr_in sin;
 111   struct sockaddr_in6 sin6;
 112 };
 113
 114 /* Check whether an address family is even slightly supported. */
 115 static int addrfamok(int af)
 116 {
 117   switch (af) {
 118     case AF_INET: case AF_INET6: return (1);
 119     default: return (0);
 120   }
 121 }
 122
 123 /* Return the size of a socket address. */
 124 static size_t addrsz(const union addr *a)
 125 {
 126   switch (a->sa.sa_family) {
 127     case AF_INET: return (sizeof(a->sin));
 128     case AF_INET6: return (sizeof(a->sin6));
 129     default: abort();
 130   }
 131 }
 132
 133 /* Compare two addresses.  Maybe compare the port numbers too. */
 134 #define AEF_PORT 1u
 135 static int addreq(const union addr *a, const union addr *b, unsigned f)
 136 {
 137   switch (a->sa.sa_family) {
 138     case AF_INET:
 139       return (a->sin.sin_addr.s_addr == b->sin.sin_addr.s_addr &&
 140               (!(f&AEF_PORT) || a->sin.sin_port == b->sin.sin_port));
 141     case AF_INET6:
 142       return (!memcmp(a->sin6.sin6_addr.s6_addr,
 143                       b->sin6.sin6_addr.s6_addr, 16) &&
 144               (!(f&AEF_PORT) || a->sin6.sin6_port == b->sin6.sin6_port));
 145     default:
 146       abort();
 147   }
 148 }
 149
 150 /*----- Main algorithm skeleton -------------------------------------------*/
 151
 152 struct param {
 153   unsigned f;                           /* Various flags */
 154 #define F_VERBOSE 1u                    /*   Give a running commentary */
 155   double retx;                          /* Initial retransmit interval */
 156   double regr;                          /* Retransmit growth factor */
 157   double timeout;                       /* Retransmission timeout */
 158   int seqoff;                           /* Offset to write sequence number */
 159   const struct probe_ops *pops;         /* Probe algorithm description */
 160   union addr a;                         /* Destination address */
 161 };
 162
 163 struct probestate {
 164   const struct param *pp;
 165   unsigned q;
 166 };
 167
 168 struct probe_ops {
 169   const char *name;
 170   const struct probe_ops *next;
 171   size_t statesz;
 172   int (*setup)(void *, int, const struct param *);
 173   void (*finish)(void *);
 174   void (*selprep)(void *, int *, fd_set *);
 175   int (*xmit)(void *, int);
 176   int (*selproc)(void *, fd_set *, struct probestate *);
 177 };
 178
 179 #define OPS_CHAIN 0
 180
 181 enum {
 182   RC_FAIL = -99,
 183   RC_OK = 0,
 184   RC_LOWER = -1,
 185   RC_HIGHER = -2,
 186   RC_NOREPLY = -3
 187   /* or a positive MTU upper-bound */
 188 };
 189
 190 /* Add a file descriptor FD to the set `fd_in', updating `*maxfd'. */
 191 #define ADDFD(fd) \
 192   do { FD_SET(fd, fd_in); if (*maxfd < fd) *maxfd = fd; } while (0)
 193
 194 /* Check whether a buffer contains a packet from our current probe. */
 195 static int mypacketp(struct probestate *ps,
 196                      const unsigned char *p, size_t sz)
 197 {
 198   const struct param *pp = ps->pp;
 199
 200   return (sz >= pp->seqoff + 2 && LOAD16(p + pp->seqoff) == ps->q);
 201 }
 202
 203 /* See whether MTU is an acceptable MTU value.  Return an appropriate
 204  * RC_... code or a new suggested MTU.
 205  */
 206 static int probe(struct probestate *ps, void *st, int mtu)
 207 {
 208   const struct param *pp = ps->pp;
 209   fd_set fd_in;
 210   struct timeval tv, now, when, done;
 211   double timer = pp->retx;
 212   int rc, maxfd;
 213
 214   /* Set up the first retransmit and give-up timers. */
 215   gettimeofday(&now, 0);
 216   f2tv(&tv, pp->timeout); TV_ADD(&done, &now, &tv);
 217   f2tv(&tv, timer); TV_ADD(&when, &now, &tv);
 218   if (TV_CMP(&when, >, &done)) when = done;
 219
 220   /* Send the initial probe. */
 221   if (pp->f & F_VERBOSE)
 222     moan("sending probe of size %d (seq = %04x)", mtu, ps->q);
 223   STEP(ps->q);
 224   STORE16(buf + pp->seqoff, ps->q);
 225   if ((rc = pp->pops->xmit(st, mtu)) != RC_OK) return (rc);
 226
 227   for (;;) {
 228
 229     /* Wait for something interesting to happen. */
 230     maxfd = 0; FD_ZERO(&fd_in);
 231     pp->pops->selprep(st, &maxfd, &fd_in);
 232     TV_SUB(&tv, &when, &now);
 233     if (select(maxfd + 1, &fd_in, 0, 0, &tv) < 0) return (RC_FAIL);
 234     gettimeofday(&now, 0);
 235
 236     /* See whether the probe method has any answers for us. */
 237     if ((rc = pp->pops->selproc(st, &fd_in, ps)) != RC_OK) return (rc);
 238
 239     /* If we've waited too long, give up.  If we should retransmit, do
 240      * that.
 241      */
 242     if (TV_CMP(&now, >, &done))
 243       return (RC_NOREPLY);
 244     else if (TV_CMP(&now, >, &when)) {
 245       if (pp->f & F_VERBOSE) moan("re-sending probe of size %d", mtu);
 246       if ((rc = pp->pops->xmit(st, mtu)) != RC_OK) return (rc);
 247       do {
 248         timer *= pp->regr; f2tv(&tv, timer); TV_ADD(&when, &when, &tv);
 249       } while (TV_CMP(&when, <, &now));
 250       if (TV_CMP(&when, >, &done)) when = done;
 251     }
 252   }
 253 }
 254
 255 /* Discover the path MTU to the destination address. */
 256 static int pathmtu(const struct param *pp)
 257 {
 258   int sk;
 259   int mtu, lo, hi;
 260   int rc, droppy = -1;
 261   void *st;
 262   struct probestate ps;
 263
 264   /* Build and connect a UDP socket.  We'll need this to know the local port
 265    * number to use if nothing else.  Set other stuff up.
 266    */
 267   if ((sk = socket(pp->a.sa.sa_family, SOCK_DGRAM, IPPROTO_UDP)) < 0)
 268     goto fail_0;
 269   if (connect(sk, &pp->a.sa, addrsz(&pp->a))) goto fail_1;
 270   st = xmalloc(pp->pops->statesz);
 271   if ((mtu = pp->pops->setup(st, sk, pp)) < 0) goto fail_2;
 272   ps.pp = pp; ps.q = rand() & 0xffff;
 273   switch (pp->a.sa.sa_family) {
 274     case AF_INET: lo = 576; break;
 275     case AF_INET6: lo = 1280; break;
 276     default: abort();
 277   }
 278   hi = mtu;
 279   if (hi < lo) { errno = EMSGSIZE; return (-1); }
 280
 281   /* And now we do a thing which is sort of like a binary search, except that
 282    * we also take explicit clues as establishing a new upper bound, and we
 283    * try to hug that initially.
 284    */
 285   for (;;) {
 286     assert(lo <= mtu && mtu <= hi);
 287     if (pp->f & F_VERBOSE) moan("probe: %d <= %d <= %d", lo, mtu, hi);
 288     rc = probe(&ps, st, mtu);
 289     switch (rc) {
 290
 291       case RC_FAIL:
 292         if (pp->f & F_VERBOSE) moan("probe failed");
 293         goto fail_3;
 294
 295       case RC_NOREPLY:
 296         /* If we've not seen a dropped packet before then we don't know what
 297          * this means yet -- in particular, we don't know which bit of the
 298          * network is swallowing packets.  Send a minimum-size probe.  If
 299          * that doesn't come back then assume that the remote host is
 300          * swallowing our packets.  If it does, then we assume that dropped
 301          * packets are a result of ICMP fragmentation-needed reports being
 302          * lost or suppressed.
 303          */
 304         if (pp->f & F_VERBOSE) moan("gave up: black hole detected");
 305         if (droppy == -1) {
 306           if (pp->f & F_VERBOSE) moan("sending minimum-size probe");
 307           switch (probe(&ps, st, lo)) {
 308             case RC_FAIL:
 309               goto fail_3;
 310             case RC_NOREPLY:
 311               if (pp->f & F_VERBOSE) {
 312                 moan("no reply from min-size probe: "
 313                      "assume black hole at target");
 314               }
 315               droppy = 1;
 316               break;
 317             case RC_HIGHER:
 318               if (pp->f & F_VERBOSE) {
 319                 moan("reply from min-size probe OK: "
 320                      "assume black hole in network");
 321               }
 322               droppy = 0;
 323               break;
 324             default:
 325               if (pp->f & F_VERBOSE)
 326                 moan("unexpected return code from probe");
 327               errno = ENOTCONN;
 328               goto fail_3;
 329           }
 330         }
 331
 332         if (droppy) goto higher; else goto lower;
 333
 334       case RC_HIGHER:
 335       higher:
 336         if (droppy == -1) {
 337           if (pp->f & F_VERBOSE)
 338             moan("probe returned: remote host is not a black hole");
 339           droppy = 0;
 340         }
 341         if (mtu == hi) {
 342           if (pp->f & F_VERBOSE) moan("probe returned: found correct MTU");
 343           goto done;
 344         }
 345         lo = mtu;
 346
 347         /* Now we must make a new guess, between lo and hi.  We know that lo
 348          * is good; but we're not so sure about hi here.  We know that hi >
 349          * lo, so this will find an approximate midpoint, greater than lo and
 350          * no more than hi.
 351          */
 352         if (pp->f & F_VERBOSE) moan("probe returned: guessing higher");
 353         mtu += (hi - lo + 1)/2;
 354         break;
 355
 356       case RC_LOWER:
 357       lower:
 358         /* If this didn't work, and we're already at the bottom of our
 359          * possible range, then something has gone horribly wrong.
 360          */
 361         assert(lo < mtu);
 362         hi = mtu - 1;
 363         if (lo == hi) {
 364           if (pp->f & F_VERBOSE) moan("error returned: found correct MTU");
 365           mtu = lo;
 366           goto done;
 367         }
 368
 369         /* We must make a new guess, between lo and hi.  We're probably
 370          * fairly sure that lo will succeed, since either it's the minimum
 371          * MTU or we've tested it already; but we're not quite sure about hi,
 372          * so we want to aim high.
 373          */
 374         if (pp->f & F_VERBOSE) moan("error returned: guessing lower");
 375         mtu -= (hi - lo + 1)/2;
 376         break;
 377
 378       default:
 379         if (pp->f & F_VERBOSE) moan("error returned with new MTU estimate");
 380         mtu = hi = rc;
 381         break;
 382     }
 383   }
 384
 385 done:
 386   /* Clean up and return our result. */
 387   pp->pops->finish(st);
 388   xfree(st);
 389   close(sk);
 390   return (mtu);
 391
 392 fail_3:
 393   pp->pops->finish(st);
 394 fail_2:
 395   xfree(st);
 396 fail_1:
 397   close(sk);
 398 fail_0:
 399   return (-1);
 400 }
 401
 402 /*----- Doing it the hard way ---------------------------------------------*/
 403
 404 #if defined(linux) || defined(__OpenBSD__)
 405 #define IPHDR_SANE
 406 #endif
 407
 408 #ifdef IPHDR_SANE
 409 #  define sane_htons htons
 410 #  define sane_htonl htonl
 411 #else
 412 #  define sane_htons
 413 #  define sane_htonl
 414 #endif
 415
 416 static int rawicmp = -1, rawudp = -1, rawerr = 0;
 417
 418 #define IPCK_INIT 0xffff
 419
 420 /* Compute an IP checksum over some data.  This is a restartable interface:
 421  * initialize A to `IPCK_INIT' for the first call.
 422  */
 423 static unsigned ipcksum(const void *buf, size_t n, unsigned a)
 424 {
 425   unsigned long aa = a ^ 0xffff;
 426   const unsigned char *p = buf, *l = p + n;
 427
 428   while (p < l - 1) { aa += LOAD16_B(p); p += 2; }
 429   if (p < l) { aa += (unsigned)(*p) << 8; }
 430   do aa = (aa & 0xffff) + (aa >> 16); while (aa >= 0x10000);
 431   return (aa == 0xffff ? aa : aa ^ 0xffff);
 432 }
 433
 434 /* TCP/UDP pseudoheader structure. */
 435 struct phdr {
 436   struct in_addr ph_src, ph_dst;
 437   uint8_t ph_z, ph_p;
 438   uint16_t ph_len;
 439 };
 440
 441 struct raw_state {
 442   union addr me, a;
 443   int sk, rawicmp, rawudp;
 444   uint16_t srcport, dstport;
 445   unsigned q;
 446 };
 447
 448 static int raw_setup(void *stv, int sk, const struct param *pp)
 449 {
 450   struct raw_state *st = stv;
 451   socklen_t sz;
 452   int i, mtu = -1;
 453   struct ifaddrs *ifa, *ifaa, *ifap;
 454   struct ifreq ifr;
 455
 456   /* Check that the address is OK, and that we have the necessary raw
 457    * sockets.
 458    */
 459   switch (pp->a.sa.sa_family) {
 460     case AF_INET:
 461       if (rawerr) { errno = rawerr; goto fail_0; }
 462       st->rawicmp = rawicmp; st->rawudp = rawudp; st->sk = sk;
 463       break;
 464     default:
 465       errno = EPFNOSUPPORT; goto fail_0;
 466   }
 467
 468   /* Initialize the sequence number. */
 469   st->q = rand() & 0xffff;
 470
 471   /* Snaffle the local and remote address and port number. */
 472   st->a = pp->a;
 473   sz = sizeof(st->me);
 474   if (getsockname(sk, &st->me.sa, &sz))
 475     goto fail_0;
 476
 477   /* An unfortunate bodge which will make sense in the future. */
 478   st->srcport = st->me.sin.sin_port; st->me.sin.sin_port = 0;
 479   st->dstport =  st->a.sin.sin_port;  st->a.sin.sin_port = 0;
 480
 481   /* There isn't a portable way to force the DF flag onto a packet through
 482    * UDP, or even through raw IP, unless we write the entire IP header
 483    * ourselves.  This is somewhat annoying, especially since we have an
 484    * uphill struggle keeping track of which systems randomly expect which
 485    * header fields to be presented in host byte order.  Oh, well.
 486    */
 487   i = 1;
 488   if (setsockopt(rawudp, IPPROTO_IP, IP_HDRINCL, &i, sizeof(i))) goto fail_0;
 489
 490   /* Find an upper bound on the MTU.  Do two passes over the interface
 491    * list.  If we can find matches for our local address then use the
 492    * highest one of those; otherwise do a second pass and simply take the
 493    * highest MTU of any network interface.
 494    */
 495   if (getifaddrs(&ifaa)) goto fail_0;
 496   for (i = 0; i < 2; i++) {
 497     for (ifap = 0, ifa = ifaa; ifa; ifa = ifa->ifa_next) {
 498       if (!(ifa->ifa_flags & IFF_UP) || !ifa->ifa_addr ||
 499           ifa->ifa_addr->sa_family != st->me.sa.sa_family ||
 500           (i == 0 &&
 501            !addreq((union addr *)ifa->ifa_addr, &st->me, 0)) ||
 502           (i == 1 && ifap && strcmp(ifap->ifa_name, ifa->ifa_name) == 0) ||
 503           strlen(ifa->ifa_name) >= sizeof(ifr.ifr_name))
 504         continue;
 505       ifap = ifa;
 506       strcpy(ifr.ifr_name, ifa->ifa_name);
 507       if (ioctl(sk, SIOCGIFMTU, &ifr)) goto fail_1;
 508       if (mtu < ifr.ifr_mtu) mtu = ifr.ifr_mtu;
 509     }
 510     if (mtu > 0) break;
 511   }
 512   if (mtu < 0) { errno = ENOTCONN; goto fail_1; }
 513   freeifaddrs(ifaa);
 514
 515   /* Done. */
 516   return (mtu);
 517
 518 fail_1:
 519   freeifaddrs(ifaa);
 520 fail_0:
 521   return (-1);
 522 }
 523
 524 static void raw_finish(void *stv) { ; }
 525
 526 static void raw_selprep(void *stv, int *maxfd, fd_set *fd_in)
 527   { struct raw_state *st = stv; ADDFD(st->sk); ADDFD(st->rawicmp); }
 528
 529 static int raw_xmit(void *stv, int mtu)
 530 {
 531   struct raw_state *st = stv;
 532   unsigned char b[65536], *p;
 533   struct ip *ip;
 534   struct udphdr *udp;
 535   struct phdr ph;
 536   unsigned ck;
 537
 538   /* Build the IP header. */
 539   ip = (struct ip *)b;
 540   ip->ip_v = 4;
 541   ip->ip_hl = sizeof(*ip)/4;
 542   ip->ip_tos = IPTOS_RELIABILITY;
 543   ip->ip_len = sane_htons(mtu);
 544   STEP(st->q); ip->ip_id = htons(st->q);
 545   ip->ip_off = sane_htons(0 | IP_DF);
 546   ip->ip_ttl = 64;
 547   ip->ip_p = IPPROTO_UDP;
 548   ip->ip_sum = 0;
 549   ip->ip_src = st->me.sin.sin_addr;
 550   ip->ip_dst = st->a.sin.sin_addr;
 551
 552   /* Build a UDP packet in the output buffer. */
 553   udp = (struct udphdr *)(ip + 1);
 554   udp->uh_sport = st->srcport;
 555   udp->uh_dport = st->dstport;
 556   udp->uh_ulen = htons(mtu - sizeof(*ip));
 557   udp->uh_sum = 0;
 558
 559   /* Copy the payload. */
 560   p = (unsigned char *)(udp + 1);
 561   memcpy(p, buf, mtu - (p - b));
 562
 563   /* Calculate the UDP checksum. */
 564   ph.ph_src = ip->ip_src;
 565   ph.ph_dst = ip->ip_dst;
 566   ph.ph_z = 0;
 567   ph.ph_p = IPPROTO_UDP;
 568   ph.ph_len = udp->uh_ulen;
 569   ck = IPCK_INIT;
 570   ck = ipcksum(&ph, sizeof(ph), ck);
 571   ck = ipcksum(udp, mtu - sizeof(*ip), ck);
 572   udp->uh_sum = htons(ck);
 573
 574   /* Send the whole thing off.  If we're too big for the interface then we
 575    * might need to trim immediately.
 576    */
 577   if (sendto(st->rawudp, b, mtu, 0, &st->a.sa, addrsz(&st->a)) < 0) {
 578     if (errno == EMSGSIZE) return (RC_LOWER);
 579     else goto fail_0;
 580   }
 581
 582   /* Done. */
 583   return (RC_OK);
 584
 585 fail_0:
 586   return (RC_FAIL);
 587 }
 588
 589 static int raw_selproc(void *stv, fd_set *fd_in, struct probestate *ps)
 590 {
 591   struct raw_state *st = stv;
 592   unsigned char b[65536];
 593   struct ip *ip;
 594   struct icmp *icmp;
 595   struct udphdr *udp;
 596   const unsigned char *payload;
 597   ssize_t n;
 598
 599   /* An ICMP packet: see what's inside. */
 600   if (FD_ISSET(st->rawicmp, fd_in)) {
 601     if ((n = read(st->rawicmp, b, sizeof(b))) < 0) goto fail_0;
 602
 603     ip = (struct ip *)b;
 604     if (n < sizeof(*ip) || n < sizeof(4*ip->ip_hl) ||
 605         ip->ip_v != 4 || ip->ip_p != IPPROTO_ICMP)
 606       goto skip_icmp;
 607     n -= sizeof(4*ip->ip_hl);
 608
 609     icmp = (struct icmp *)(b + 4*ip->ip_hl);
 610     if (n < sizeof(*icmp) || icmp->icmp_type != ICMP_UNREACH)
 611       goto skip_icmp;
 612     n -= offsetof(struct icmp, icmp_ip);
 613
 614     ip = &icmp->icmp_ip;
 615     if (n < sizeof(*ip) ||
 616         ip->ip_p != IPPROTO_UDP || ip->ip_hl != sizeof(*ip)/4 ||
 617         ip->ip_id != htons(st->q) ||
 618         ip->ip_src.s_addr != st->me.sin.sin_addr.s_addr ||
 619         ip->ip_dst.s_addr != st->a.sin.sin_addr.s_addr)
 620       goto skip_icmp;
 621     n -= sizeof(*ip);
 622
 623     udp = (struct udphdr *)(ip + 1);
 624     if (n < sizeof(*udp) || udp->uh_sport != st->srcport ||
 625         udp->uh_dport != st->dstport)
 626       goto skip_icmp;
 627     n -= sizeof(*udp);
 628
 629     payload = (const unsigned char *)(udp + 1);
 630     if (!mypacketp(ps, payload, n)) goto skip_icmp;
 631
 632     if (icmp->icmp_code == ICMP_UNREACH_PORT) return (RC_HIGHER);
 633     else if (icmp->icmp_code != ICMP_UNREACH_NEEDFRAG) goto skip_icmp;
 634     else if (icmp->icmp_nextmtu) return (htons(icmp->icmp_nextmtu));
 635     else return (RC_LOWER);
 636   }
 637 skip_icmp:;
 638
 639   /* If we got a reply to the current probe then we're good.  If we got an
 640    * error, or the packet's sequence number is wrong, then ignore it.
 641    */
 642   if (FD_ISSET(st->sk, fd_in)) {
 643     if ((n = read(st->sk, b, sizeof(b))) < 0) return (RC_OK);
 644     else if (mypacketp(ps, b, n)) return (RC_HIGHER);
 645     else return (RC_OK);
 646   }
 647
 648   return (RC_OK);
 649
 650 fail_0:
 651   return (RC_FAIL);
 652 }
 653
 654 static const struct probe_ops raw_ops = {
 655   "raw", OPS_CHAIN, sizeof(struct raw_state),
 656   raw_setup, raw_finish,
 657   raw_selprep, raw_xmit, raw_selproc
 658 };
 659
 660 #undef OPS_CHAIN
 661 #define OPS_CHAIN &raw_ops
 662
 663 /*----- Doing the job on Linux --------------------------------------------*/
 664
 665 #if defined(linux)
 666
 667 #ifndef IP_MTU
 668 #  define IP_MTU 14                     /* Blech! */
 669 #endif
 670
 671 struct linux_state {
 672   int sol, so_mtu_discover, so_mtu;
 673   int sk;
 674   size_t hdrlen;
 675 };
 676
 677 static int linux_setup(void *stv, int sk, const struct param *pp)
 678 {
 679   struct linux_state *st = stv;
 680   int i, mtu;
 681   socklen_t sz;
 682
 683   /* Check that the address is OK. */
 684   switch (pp->a.sa.sa_family) {
 685     case AF_INET:
 686       st->sol = IPPROTO_IP;
 687       st->so_mtu_discover = IP_MTU_DISCOVER;
 688       st->so_mtu = IP_MTU;
 689       st->hdrlen = 28;
 690       break;
 691     case AF_INET6:
 692       st->sol = IPPROTO_IPV6;
 693       st->so_mtu_discover = IPV6_MTU_DISCOVER;
 694       st->so_mtu = IPV6_MTU;
 695       st->hdrlen = 48;
 696       break;
 697     default:
 698       errno = EPFNOSUPPORT;
 699       return (-1);
 700   }
 701
 702   /* Snaffle the UDP socket. */
 703   st->sk = sk;
 704
 705   /* Turn on kernel path-MTU discovery and force DF on. */
 706   i = IP_PMTUDISC_PROBE;
 707   if (setsockopt(st->sk, st->sol, st->so_mtu_discover, &i, sizeof(i)))
 708     return (-1);
 709
 710   /* Read the initial MTU guess back and report it. */
 711   sz = sizeof(mtu);
 712   if (getsockopt(st->sk, st->sol, st->so_mtu, &mtu, &sz))
 713     return (-1);
 714
 715   /* Done. */
 716   return (mtu);
 717 }
 718
 719 static void linux_finish(void *stv) { ; }
 720
 721 static void linux_selprep(void *stv, int *maxfd, fd_set *fd_in)
 722   { struct linux_state *st = stv; ADDFD(st->sk); }
 723
 724 static int linux_xmit(void *stv, int mtu)
 725 {
 726   struct linux_state *st = stv;
 727
 728   /* Write the packet. */
 729   if (write(st->sk, buf, mtu - st->hdrlen) >= 0) return (RC_OK);
 730   else if (errno == EMSGSIZE) return (RC_LOWER);
 731   else return (RC_FAIL);
 732 }
 733
 734 static int linux_selproc(void *stv, fd_set *fd_in, struct probestate *ps)
 735 {
 736   struct linux_state *st = stv;
 737   int mtu;
 738   socklen_t sz;
 739   ssize_t n;
 740   unsigned char b[65536];
 741
 742   /* Read an answer.  If it looks like the right kind of error then report a
 743    * success.  This is potentially wrong, since we can't tell whether an
 744    * error was delayed from an earlier probe.  However, we never return
 745    * RC_LOWER from this method, so the packet sizes ought to be monotonically
 746    * decreasing and this won't cause trouble.  Otherwise update from the
 747    * kernel's idea of the right MTU.
 748    */
 749   if (FD_ISSET(st->sk, fd_in)) {
 750     n = read(st->sk, &buf, sizeof(buf));
 751     if (n >= 0 ?
 752         mypacketp(ps, b, n) :
 753         errno == ECONNREFUSED || errno == EHOSTUNREACH)
 754       return (RC_HIGHER);
 755     sz = sizeof(mtu);
 756     if (getsockopt(st->sk, st->sol, st->so_mtu, &mtu, &sz))
 757       return (RC_FAIL);
 758     return (mtu);
 759   }
 760   return (RC_OK);
 761 }
 762
 763 static const struct probe_ops linux_ops = {
 764   "linux", OPS_CHAIN, sizeof(struct linux_state),
 765   linux_setup, linux_finish,
 766   linux_selprep, linux_xmit, linux_selproc
 767 };
 768
 769 #undef OPS_CHAIN
 770 #define OPS_CHAIN &linux_ops
 771
 772 #endif
 773
 774 /*----- Help options ------------------------------------------------------*/
 775
 776 static const struct probe_ops *probe_ops = OPS_CHAIN;
 777
 778 static void version(FILE *fp)
 779   { pquis(fp, "$, TrIPE version " VERSION "\n"); }
 780
 781 static void usage(FILE *fp)
 782 {
 783   pquis(fp, "Usage: $ [-46v] [-H HEADER] [-m METHOD]\n\
 784          [-r SECS] [-g FACTOR] [-t SECS] HOST [PORT]\n");
 785 }
 786
 787 static void help(FILE *fp)
 788 {
 789   const struct probe_ops *ops;
 790
 791   version(fp);
 792   fputc('\n', fp);
 793   usage(fp);
 794   fputs("\
 795 \n\
 796 Options in full:\n\
 797 \n\
 798 -h, --help              Show this help text.\n\
 799 -V, --version           Show version number.\n\
 800 -u, --usage             Show brief usage message.\n\
 801 \n\
 802 -4, --ipv4              Restrict to IPv4 only.\n\
 803 -6, --ipv6              Restrict to IPv6 only.\n\
 804 -g, --growth=FACTOR     Growth factor for retransmit interval.\n\
 805 -m, --method=METHOD     Use METHOD to probe for MTU.\n\
 806 -r, --retransmit=SECS   Retransmit if no reply after SEC.\n\
 807 -t, --timeout=SECS      Give up expecting a reply after SECS.\n\
 808 -v, --verbose           Write a running commentary to stderr.\n\
 809 -H, --header=HEX        Packet header, in hexadecimal.\n\
 810 \n\
 811 Probe methods:\n\
 812 ", fp);
 813   for (ops = probe_ops; ops; ops = ops->next)
 814     printf("\t%s\n", ops->name);
 815 }
 816
 817 /*----- Main code ---------------------------------------------------------*/
 818
 819 int main(int argc, char *argv[])
 820 {
 821   struct param pp = { 0, 0.333, 3.0, 8.0, 0, OPS_CHAIN };
 822   hex_ctx hc;
 823   dstr d = DSTR_INIT;
 824   size_t sz;
 825   int i, err;
 826   struct addrinfo aihint = { 0 }, *ailist, *ai;
 827   const char *host, *svc = "7";
 828   unsigned f = 0;
 829
 830 #define f_bogus 1u
 831
 832   if ((rawicmp = socket(PF_INET, SOCK_RAW, IPPROTO_ICMP)) < 0 ||
 833       (rawudp = socket(PF_INET, SOCK_RAW, IPPROTO_UDP)) < 0)
 834     rawerr = errno;
 835   if (setuid(getuid()))
 836     abort();
 837
 838   ego(argv[0]);
 839   fillbuffer(buf, sizeof(buf));
 840
 841   aihint.ai_family = AF_UNSPEC;
 842   aihint.ai_protocol = IPPROTO_UDP;
 843   aihint.ai_socktype = SOCK_DGRAM;
 844   aihint.ai_flags = AI_ADDRCONFIG;
 845
 846   for (;;) {
 847     static const struct option opts[] = {
 848       { "help",         0,              0,      'h' },
 849       { "version",      0,              0,      'V' },
 850       { "usage",        0,              0,      'u' },
 851       { "ipv4",         0,              0,      '4' },
 852       { "ipv6",         0,              0,      '6' },
 853       { "header",       OPTF_ARGREQ,    0,      'H' },
 854       { "growth",       OPTF_ARGREQ,    0,      'g' },
 855       { "method",       OPTF_ARGREQ,    0,      'm' },
 856       { "retransmit",   OPTF_ARGREQ,    0,      'r' },
 857       { "timeout",      OPTF_ARGREQ,    0,      't' },
 858       { "verbose",      0,              0,      'v' },
 859       { 0,              0,              0,      0 }
 860     };
 861
 862     i = mdwopt(argc, argv, "hVu" "46H:g:m:r:t:v", opts, 0, 0, 0);
 863     if (i < 0) break;
 864     switch (i) {
 865       case 'h': help(stdout); exit(0);
 866       case 'V': version(stdout); exit(0);
 867       case 'u': usage(stdout); exit(0);
 868
 869       case 'H':
 870         DRESET(&d);
 871         hex_init(&hc);
 872         hex_decode(&hc, optarg, strlen(optarg), &d);
 873         hex_decode(&hc, 0, 0, &d);
 874         sz = d.len < 532 ? d.len : 532;
 875         memcpy(buf, d.buf, sz);
 876         pp.seqoff = sz;
 877         break;
 878
 879       case '4': aihint.ai_family = AF_INET; break;
 880       case '6': aihint.ai_family = AF_INET6; break;
 881       case 'g': pp.regr = s2f(optarg, "retransmit growth factor"); break;
 882       case 'r': pp.retx = s2f(optarg, "retransmit interval"); break;
 883       case 't': pp.timeout = s2f(optarg, "timeout"); break;
 884
 885       case 'm':
 886         for (pp.pops = OPS_CHAIN; pp.pops; pp.pops = pp.pops->next)
 887           if (strcmp(pp.pops->name, optarg) == 0) goto found_alg;
 888         die(EXIT_FAILURE, "unknown probe algorithm `%s'", optarg);
 889       found_alg:
 890         break;
 891
 892       case 'v': pp.f |= F_VERBOSE; break;
 893
 894       default:
 895         f |= f_bogus;
 896         break;
 897     }
 898   }
 899   argv += optind; argc -= optind;
 900   if ((f & f_bogus) || 1 > argc || argc > 2) {
 901     usage(stderr);
 902     exit(EXIT_FAILURE);
 903   }
 904
 905   host = argv[0];
 906   if (argv[1]) svc = argv[1];
 907   if ((err = getaddrinfo(host, svc, &aihint, &ailist)) != 0) {
 908     die(EXIT_FAILURE, "unknown host `%s' or service `%s': %s",
 909         host, svc, gai_strerror(err));
 910   }
 911   for (ai = ailist; ai && !addrfamok(ai->ai_family); ai = ai->ai_next);
 912   if (!ai) die(EXIT_FAILURE, "no supported address families for `%s'", host);
 913   assert(ai->ai_addrlen <= sizeof(pp.a));
 914   memcpy(&pp.a, ai->ai_addr, ai->ai_addrlen);
 915
 916   i = pathmtu(&pp);
 917   if (i < 0)
 918     die(EXIT_FAILURE, "failed to discover MTU: %s", strerror(errno));
 919   printf("%d\n", i);
 920   if (ferror(stdout) || fflush(stdout) || fclose(stdout))
 921     die(EXIT_FAILURE, "failed to write result: %s", strerror(errno));
 922   return (0);
 923 }
 924
 925 /*----- That's all, folks -------------------------------------------------*/