networking/zcip.c

   1 /* vi: set sw=4 ts=4: */
   2 /*
   3  * RFC3927 ZeroConf IPv4 Link-Local addressing
   4  * (see <http://www.zeroconf.org/>)
   5  *
   6  * Copyright (C) 2003 by Arthur van Hoff (avh@strangeberry.com)
   7  * Copyright (C) 2004 by David Brownell
   8  *
   9  * Licensed under the GPL v2 or later, see the file LICENSE in this tarball.
  10  */
  11
  12 /*
  13  * ZCIP just manages the 169.254.*.* addresses.  That network is not
  14  * routed at the IP level, though various proxies or bridges can
  15  * certainly be used.  Its naming is built over multicast DNS.
  16  */
  17
  18 //#define DEBUG
  19
  20 // TODO:
  21 // - more real-world usage/testing, especially daemon mode
  22 // - kernel packet filters to reduce scheduling noise
  23 // - avoid silent script failures, especially under load...
  24 // - link status monitoring (restart on link-up; stop on link-down)
  25
  26 #include "busybox.h"
  27 #include <syslog.h>
  28 #include <poll.h>
  29 #include <sys/wait.h>
  30 #include <netinet/ether.h>
  31 #include <net/ethernet.h>
  32 #include <net/if.h>
  33 #include <net/if_arp.h>
  34
  35 #include <linux/if_packet.h>
  36 #include <linux/sockios.h>
  37
  38
  39 struct arp_packet {
  40         struct ether_header hdr;
  41         struct ether_arp arp;
  42 } ATTRIBUTE_PACKED;
  43
  44 enum {
  45 /* 169.254.0.0 */
  46         LINKLOCAL_ADDR = 0xa9fe0000,
  47
  48 /* protocol timeout parameters, specified in seconds */
  49         PROBE_WAIT = 1,
  50         PROBE_MIN = 1,
  51         PROBE_MAX = 2,
  52         PROBE_NUM = 3,
  53         MAX_CONFLICTS = 10,
  54         RATE_LIMIT_INTERVAL = 60,
  55         ANNOUNCE_WAIT = 2,
  56         ANNOUNCE_NUM = 2,
  57         ANNOUNCE_INTERVAL = 2,
  58         DEFEND_INTERVAL = 10
  59 };
  60
  61 /* States during the configuration process. */
  62 enum {
  63         PROBE = 0,
  64         RATE_LIMIT_PROBE,
  65         ANNOUNCE,
  66         MONITOR,
  67         DEFEND
  68 };
  69
  70 #define VDBG(fmt,args...) \
  71         do { } while (0)
  72
  73 static unsigned opts;
  74 #define FOREGROUND (opts & 1)
  75 #define QUIT (opts & 2)
  76
  77 /**
  78  * Pick a random link local IP address on 169.254/16, except that
  79  * the first and last 256 addresses are reserved.
  80  */
  81 static void pick(struct in_addr *ip)
  82 {
  83         unsigned tmp;
  84
  85         /* use cheaper math than lrand48() mod N */
  86         do {
  87                 tmp = (lrand48() >> 16) & IN_CLASSB_HOST;
  88         } while (tmp > (IN_CLASSB_HOST - 0x0200));
  89         ip->s_addr = htonl((LINKLOCAL_ADDR + 0x0100) + tmp);
  90 }
  91
  92 /* TODO: we need a flag to direct bb_[p]error_msg output to stderr. */
  93
  94 /**
  95  * Broadcast an ARP packet.
  96  */
  97 static void arp(int fd, struct sockaddr *saddr, int op,
  98         const struct ether_addr *source_addr, struct in_addr source_ip,
  99         const struct ether_addr *target_addr, struct in_addr target_ip)
 100 {
 101         struct arp_packet p;
 102         memset(&p, 0, sizeof(p));
 103
 104         // ether header
 105         p.hdr.ether_type = htons(ETHERTYPE_ARP);
 106         memcpy(p.hdr.ether_shost, source_addr, ETH_ALEN);
 107         memset(p.hdr.ether_dhost, 0xff, ETH_ALEN);
 108
 109         // arp request
 110         p.arp.arp_hrd = htons(ARPHRD_ETHER);
 111         p.arp.arp_pro = htons(ETHERTYPE_IP);
 112         p.arp.arp_hln = ETH_ALEN;
 113         p.arp.arp_pln = 4;
 114         p.arp.arp_op = htons(op);
 115         memcpy(&p.arp.arp_sha, source_addr, ETH_ALEN);
 116         memcpy(&p.arp.arp_spa, &source_ip, sizeof(p.arp.arp_spa));
 117         memcpy(&p.arp.arp_tha, target_addr, ETH_ALEN);
 118         memcpy(&p.arp.arp_tpa, &target_ip, sizeof(p.arp.arp_tpa));
 119
 120         // send it
 121         if (sendto(fd, &p, sizeof(p), 0, saddr, sizeof(*saddr)) < 0) {
 122                 bb_perror_msg("sendto");
 123                 //return -errno;
 124         }
 125         // Currently all callers ignore errors, that's why returns are
 126         // commented out...
 127         //return 0;
 128 }
 129
 130 /**
 131  * Run a script.
 132  */
 133 static int run(const char *script, const char *arg, const char *intf, struct in_addr *ip)
 134 {
 135         int pid, status;
 136         const char *why;
 137
 138         if(1) { //always true: if (script != NULL)
 139                 VDBG("%s run %s %s\n", intf, script, arg);
 140                 if (ip != NULL) {
 141                         char *addr = inet_ntoa(*ip);
 142                         setenv("ip", addr, 1);
 143                         bb_info_msg("%s %s %s", arg, intf, addr);
 144                 }
 145
 146                 pid = vfork();
 147                 if (pid < 0) {                  // error
 148                         why = "vfork";
 149                         goto bad;
 150                 } else if (pid == 0) {          // child
 151                         execl(script, script, arg, NULL);
 152                         bb_perror_msg("execl");
 153                         _exit(EXIT_FAILURE);
 154                 }
 155
 156                 if (waitpid(pid, &status, 0) <= 0) {
 157                         why = "waitpid";
 158                         goto bad;
 159                 }
 160                 if (WEXITSTATUS(status) != 0) {
 161                         bb_error_msg("script %s failed, exit=%d",
 162                                 script, WEXITSTATUS(status));
 163                         return -errno;
 164                 }
 165         }
 166         return 0;
 167 bad:
 168         status = -errno;
 169         bb_perror_msg("%s %s, %s", arg, intf, why);
 170         return status;
 171 }
 172
 173
 174 /**
 175  * Return milliseconds of random delay, up to "secs" seconds.
 176  */
 177 static unsigned ATTRIBUTE_ALWAYS_INLINE ms_rdelay(unsigned secs)
 178 {
 179         return lrand48() % (secs * 1000);
 180 }
 181
 182 /**
 183  * main program
 184  */
 185
 186 /* Used to be auto variables on main() stack, but
 187  * most of them were zero-inited. Moving them to bss
 188  * is more space-efficient.
 189  */
 190 static  const struct in_addr null_ip; // = { 0 };
 191 static  const struct ether_addr null_addr; // = { {0, 0, 0, 0, 0, 0} };
 192
 193 static  struct sockaddr saddr; // memset(0);
 194 static  struct in_addr ip; // = { 0 };
 195 static  struct ifreq ifr; //memset(0);
 196
 197 static  char *intf; // = NULL;
 198 static  char *script; // = NULL;
 199 static  suseconds_t timeout; // = 0;    // milliseconds
 200 static  unsigned conflicts; // = 0;
 201 static  unsigned nprobes; // = 0;
 202 static  unsigned nclaims; // = 0;
 203 static  int ready; // = 0;
 204 static  int verbose; // = 0;
 205 static  int state = PROBE;
 206
 207 int zcip_main(int argc, char *argv[]);
 208 int zcip_main(int argc, char *argv[])
 209 {
 210         struct ether_addr eth_addr;
 211         const char *why;
 212         int fd;
 213
 214         // parse commandline: prog [options] ifname script
 215         char *r_opt;
 216         opt_complementary = "vv:vf"; // -v accumulates and implies -f
 217         opts = getopt32(argc, argv, "fqr:v", &r_opt, &verbose);
 218         if (!FOREGROUND) {
 219                 /* Do it early, before all bb_xx_msg calls */
 220                 logmode = LOGMODE_SYSLOG;
 221                 openlog(applet_name, 0, LOG_DAEMON);
 222         }
 223         if (opts & 4) { // -r n.n.n.n
 224                 if (inet_aton(r_opt, &ip) == 0
 225                  || (ntohl(ip.s_addr) & IN_CLASSB_NET) != LINKLOCAL_ADDR
 226                 ) {
 227                         bb_error_msg_and_die("invalid link address");
 228                 }
 229         }
 230         argc -= optind;
 231         argv += optind;
 232         if (argc != 2)
 233                 bb_show_usage();
 234         intf = argv[0];
 235         script = argv[1];
 236         setenv("interface", intf, 1);
 237
 238         // initialize the interface (modprobe, ifup, etc)
 239         if (run(script, "init", intf, NULL) < 0)
 240                 return EXIT_FAILURE;
 241
 242         // initialize saddr
 243         //memset(&saddr, 0, sizeof(saddr));
 244         safe_strncpy(saddr.sa_data, intf, sizeof(saddr.sa_data));
 245
 246         // open an ARP socket
 247         fd = xsocket(PF_PACKET, SOCK_PACKET, htons(ETH_P_ARP));
 248         // bind to the interface's ARP socket
 249         xbind(fd, &saddr, sizeof(saddr));
 250
 251         // get the interface's ethernet address
 252         //memset(&ifr, 0, sizeof(ifr));
 253         strncpy(ifr.ifr_name, intf, sizeof(ifr.ifr_name));
 254         if (ioctl(fd, SIOCGIFHWADDR, &ifr) < 0) {
 255                 bb_perror_msg_and_die("get ethernet address");
 256         }
 257         memcpy(&eth_addr, &ifr.ifr_hwaddr.sa_data, ETH_ALEN);
 258
 259         // start with some stable ip address, either a function of
 260         // the hardware address or else the last address we used.
 261         // NOTE: the sequence of addresses we try changes only
 262         // depending on when we detect conflicts.
 263         // (SVID 3 bogon: who says that "short" is always 16 bits?)
 264         seed48( (unsigned short*)&ifr.ifr_hwaddr.sa_data );
 265         if (ip.s_addr == 0)
 266                 pick(&ip);
 267
 268         // FIXME cases to handle:
 269         //  - zcip already running!
 270         //  - link already has local address... just defend/update
 271
 272         // daemonize now; don't delay system startup
 273         if (!FOREGROUND) {
 274                 /* bb_daemonize(); - bad, will close fd! */
 275                 xdaemon(0, 0);
 276                 bb_info_msg("start, interface %s", intf);
 277         }
 278
 279         // run the dynamic address negotiation protocol,
 280         // restarting after address conflicts:
 281         //  - start with some address we want to try
 282         //  - short random delay
 283         //  - arp probes to see if another host else uses it
 284         //  - arp announcements that we're claiming it
 285         //  - use it
 286         //  - defend it, within limits
 287         while (1) {
 288                 struct pollfd fds[1];
 289                 struct timeval tv1;
 290                 struct arp_packet p;
 291
 292                 int source_ip_conflict = 0;
 293                 int target_ip_conflict = 0;
 294
 295                 fds[0].fd = fd;
 296                 fds[0].events = POLLIN;
 297                 fds[0].revents = 0;
 298
 299                 // poll, being ready to adjust current timeout
 300                 if (!timeout) {
 301                         timeout = ms_rdelay(PROBE_WAIT);
 302                         // FIXME setsockopt(fd, SO_ATTACH_FILTER, ...) to
 303                         // make the kernel filter out all packets except
 304                         // ones we'd care about.
 305                 }
 306                 // set tv1 to the point in time when we timeout
 307                 gettimeofday(&tv1, NULL);
 308                 tv1.tv_usec += (timeout % 1000) * 1000;
 309                 while (tv1.tv_usec > 1000000) {
 310                         tv1.tv_usec -= 1000000;
 311                         tv1.tv_sec++;
 312                 }
 313                 tv1.tv_sec += timeout / 1000;
 314
 315                 VDBG("...wait %ld %s nprobes=%d, nclaims=%d\n",
 316                                 timeout, intf, nprobes, nclaims);
 317                 switch (poll(fds, 1, timeout)) {
 318
 319                 // timeout
 320                 case 0:
 321                         VDBG("state = %d\n", state);
 322                         switch (state) {
 323                         case PROBE:
 324                                 // timeouts in the PROBE state mean no conflicting ARP packets
 325                                 // have been received, so we can progress through the states
 326                                 if (nprobes < PROBE_NUM) {
 327                                         nprobes++;
 328                                         VDBG("probe/%d %s@%s\n",
 329                                                         nprobes, intf, inet_ntoa(ip));
 330                                         arp(fd, &saddr, ARPOP_REQUEST,
 331                                                         &eth_addr, null_ip,
 332                                                         &null_addr, ip);
 333                                         timeout = PROBE_MIN * 1000;
 334                                         timeout += ms_rdelay(PROBE_MAX
 335                                                         - PROBE_MIN);
 336                                 }
 337                                 else {
 338                                         // Switch to announce state.
 339                                         state = ANNOUNCE;
 340                                         nclaims = 0;
 341                                         VDBG("announce/%d %s@%s\n",
 342                                                         nclaims, intf, inet_ntoa(ip));
 343                                         arp(fd, &saddr, ARPOP_REQUEST,
 344                                                         &eth_addr, ip,
 345                                                         &eth_addr, ip);
 346                                         timeout = ANNOUNCE_INTERVAL * 1000;
 347                                 }
 348                                 break;
 349                         case RATE_LIMIT_PROBE:
 350                                 // timeouts in the RATE_LIMIT_PROBE state mean no conflicting ARP packets
 351                                 // have been received, so we can move immediately to the announce state
 352                                 state = ANNOUNCE;
 353                                 nclaims = 0;
 354                                 VDBG("announce/%d %s@%s\n",
 355                                                 nclaims, intf, inet_ntoa(ip));
 356                                 arp(fd, &saddr, ARPOP_REQUEST,
 357                                                 &eth_addr, ip,
 358                                                 &eth_addr, ip);
 359                                 timeout = ANNOUNCE_INTERVAL * 1000;
 360                                 break;
 361                         case ANNOUNCE:
 362                                 // timeouts in the ANNOUNCE state mean no conflicting ARP packets
 363                                 // have been received, so we can progress through the states
 364                                 if (nclaims < ANNOUNCE_NUM) {
 365                                         nclaims++;
 366                                         VDBG("announce/%d %s@%s\n",
 367                                                         nclaims, intf, inet_ntoa(ip));
 368                                         arp(fd, &saddr, ARPOP_REQUEST,
 369                                                         &eth_addr, ip,
 370                                                         &eth_addr, ip);
 371                                         timeout = ANNOUNCE_INTERVAL * 1000;
 372                                 }
 373                                 else {
 374                                         // Switch to monitor state.
 375                                         state = MONITOR;
 376                                         // link is ok to use earlier
 377                                         // FIXME update filters
 378                                         run(script, "config", intf, &ip);
 379                                         ready = 1;
 380                                         conflicts = 0;
 381                                         timeout = -1; // Never timeout in the monitor state.
 382
 383                                         // NOTE: all other exit paths
 384                                         // should deconfig ...
 385                                         if (QUIT)
 386                                                 return EXIT_SUCCESS;
 387                                 }
 388                                 break;
 389                         case DEFEND:
 390                                 // We won!  No ARP replies, so just go back to monitor.
 391                                 state = MONITOR;
 392                                 timeout = -1;
 393                                 conflicts = 0;
 394                                 break;
 395                         default:
 396                                 // Invalid, should never happen.  Restart the whole protocol.
 397                                 state = PROBE;
 398                                 pick(&ip);
 399                                 timeout = 0;
 400                                 nprobes = 0;
 401                                 nclaims = 0;
 402                                 break;
 403                         } // switch (state)
 404                         break; // case 0 (timeout)
 405                 // packets arriving
 406                 case 1:
 407                         // We need to adjust the timeout in case we didn't receive
 408                         // a conflicting packet.
 409                         if (timeout > 0) {
 410                                 struct timeval tv2;
 411
 412                                 gettimeofday(&tv2, NULL);
 413                                 if (timercmp(&tv1, &tv2, <)) {
 414                                         // Current time is greater than the expected timeout time.
 415                                         // Should never happen.
 416                                         VDBG("missed an expected timeout\n");
 417                                         timeout = 0;
 418                                 } else {
 419                                         VDBG("adjusting timeout\n");
 420                                         timersub(&tv1, &tv2, &tv1);
 421                                         timeout = 1000 * tv1.tv_sec
 422                                                         + tv1.tv_usec / 1000;
 423                                 }
 424                         }
 425
 426                         if ((fds[0].revents & POLLIN) == 0) {
 427                                 if (fds[0].revents & POLLERR) {
 428                                         // FIXME: links routinely go down;
 429                                         // this shouldn't necessarily exit.
 430                                         bb_error_msg("%s: poll error", intf);
 431                                         if (ready) {
 432                                                 run(script, "deconfig",
 433                                                                 intf, &ip);
 434                                         }
 435                                         return EXIT_FAILURE;
 436                                 }
 437                                 continue;
 438                         }
 439
 440                         // read ARP packet
 441                         if (recv(fd, &p, sizeof(p), 0) < 0) {
 442                                 why = "recv";
 443                                 goto bad;
 444                         }
 445                         if (p.hdr.ether_type != htons(ETHERTYPE_ARP))
 446                                 continue;
 447
 448 #ifdef DEBUG
 449                         {
 450                                 struct ether_addr * sha = (struct ether_addr *) p.arp.arp_sha;
 451                                 struct ether_addr * tha = (struct ether_addr *) p.arp.arp_tha;
 452                                 struct in_addr * spa = (struct in_addr *) p.arp.arp_spa;
 453                                 struct in_addr * tpa = (struct in_addr *) p.arp.arp_tpa;
 454                                 VDBG("%s recv arp type=%d, op=%d,\n",
 455                                         intf, ntohs(p.hdr.ether_type),
 456                                         ntohs(p.arp.arp_op));
 457                                 VDBG("\tsource=%s %s\n",
 458                                         ether_ntoa(sha),
 459                                         inet_ntoa(*spa));
 460                                 VDBG("\ttarget=%s %s\n",
 461                                         ether_ntoa(tha),
 462                                         inet_ntoa(*tpa));
 463                         }
 464 #endif
 465                         if (p.arp.arp_op != htons(ARPOP_REQUEST)
 466                                         && p.arp.arp_op != htons(ARPOP_REPLY))
 467                                 continue;
 468
 469                         if (memcmp(p.arp.arp_spa, &ip.s_addr, sizeof(struct in_addr)) == 0 &&
 470                                 memcmp(&eth_addr, &p.arp.arp_sha, ETH_ALEN) != 0) {
 471                                 source_ip_conflict = 1;
 472                         }
 473                         if (memcmp(p.arp.arp_tpa, &ip.s_addr, sizeof(struct in_addr)) == 0 &&
 474                                 p.arp.arp_op == htons(ARPOP_REQUEST) &&
 475                                 memcmp(&eth_addr, &p.arp.arp_tha, ETH_ALEN) != 0) {
 476                                 target_ip_conflict = 1;
 477                         }
 478
 479                         VDBG("state = %d, source ip conflict = %d, target ip conflict = %d\n",
 480                                 state, source_ip_conflict, target_ip_conflict);
 481                         switch (state) {
 482                         case PROBE:
 483                         case ANNOUNCE:
 484                                 // When probing or announcing, check for source IP conflicts
 485                                 // and other hosts doing ARP probes (target IP conflicts).
 486                                 if (source_ip_conflict || target_ip_conflict) {
 487                                         conflicts++;
 488                                         if (conflicts >= MAX_CONFLICTS) {
 489                                                 VDBG("%s ratelimit\n", intf);
 490                                                 timeout = RATE_LIMIT_INTERVAL * 1000;
 491                                                 state = RATE_LIMIT_PROBE;
 492                                         }
 493
 494                                         // restart the whole protocol
 495                                         pick(&ip);
 496                                         timeout = 0;
 497                                         nprobes = 0;
 498                                         nclaims = 0;
 499                                 }
 500                                 break;
 501                         case MONITOR:
 502                                 // If a conflict, we try to defend with a single ARP probe.
 503                                 if (source_ip_conflict) {
 504                                         VDBG("monitor conflict -- defending\n");
 505                                         state = DEFEND;
 506                                         timeout = DEFEND_INTERVAL * 1000;
 507                                         arp(fd, &saddr,
 508                                                         ARPOP_REQUEST,
 509                                                         &eth_addr, ip,
 510                                                         &eth_addr, ip);
 511                                 }
 512                                 break;
 513                         case DEFEND:
 514                                 // Well, we tried.  Start over (on conflict).
 515                                 if (source_ip_conflict) {
 516                                         state = PROBE;
 517                                         VDBG("defend conflict -- starting over\n");
 518                                         ready = 0;
 519                                         run(script, "deconfig", intf, &ip);
 520
 521                                         // restart the whole protocol
 522                                         pick(&ip);
 523                                         timeout = 0;
 524                                         nprobes = 0;
 525                                         nclaims = 0;
 526                                 }
 527                                 break;
 528                         default:
 529                                 // Invalid, should never happen.  Restart the whole protocol.
 530                                 VDBG("invalid state -- starting over\n");
 531                                 state = PROBE;
 532                                 pick(&ip);
 533                                 timeout = 0;
 534                                 nprobes = 0;
 535                                 nclaims = 0;
 536                                 break;
 537                         } // switch state
 538
 539                         break; // case 1 (packets arriving)
 540                 default:
 541                         why = "poll";
 542                         goto bad;
 543                 } // switch poll
 544         }
 545 bad:
 546         bb_perror_msg("%s, %s", intf, why);
 547         return EXIT_FAILURE;
 548 }