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 /**
  74  * Pick a random link local IP address on 169.254/16, except that
  75  * the first and last 256 addresses are reserved.
  76  */
  77 static void pick(struct in_addr *ip)
  78 {
  79         unsigned tmp;
  80
  81         /* use cheaper math than lrand48() mod N */
  82         do {
  83                 tmp = (lrand48() >> 16) & IN_CLASSB_HOST;
  84         } while (tmp > (IN_CLASSB_HOST - 0x0200));
  85         ip->s_addr = htonl((LINKLOCAL_ADDR + 0x0100) + tmp);
  86 }
  87
  88 /* TODO: we need a flag to direct bb_[p]error_msg output to stderr. */
  89
  90 /**
  91  * Broadcast an ARP packet.
  92  */
  93 static void arp(int fd, struct sockaddr *saddr, int op,
  94         const struct ether_addr *source_addr, struct in_addr source_ip,
  95         const struct ether_addr *target_addr, struct in_addr target_ip)
  96 {
  97         struct arp_packet p;
  98         memset(&p, 0, sizeof(p));
  99
 100         // ether header
 101         p.hdr.ether_type = htons(ETHERTYPE_ARP);
 102         memcpy(p.hdr.ether_shost, source_addr, ETH_ALEN);
 103         memset(p.hdr.ether_dhost, 0xff, ETH_ALEN);
 104
 105         // arp request
 106         p.arp.arp_hrd = htons(ARPHRD_ETHER);
 107         p.arp.arp_pro = htons(ETHERTYPE_IP);
 108         p.arp.arp_hln = ETH_ALEN;
 109         p.arp.arp_pln = 4;
 110         p.arp.arp_op = htons(op);
 111         memcpy(&p.arp.arp_sha, source_addr, ETH_ALEN);
 112         memcpy(&p.arp.arp_spa, &source_ip, sizeof(p.arp.arp_spa));
 113         memcpy(&p.arp.arp_tha, target_addr, ETH_ALEN);
 114         memcpy(&p.arp.arp_tpa, &target_ip, sizeof(p.arp.arp_tpa));
 115
 116         // send it
 117         if (sendto(fd, &p, sizeof(p), 0, saddr, sizeof(*saddr)) < 0) {
 118                 bb_perror_msg("sendto");
 119                 //return -errno;
 120         }
 121         // Currently all callers ignore errors, that's why returns are
 122         // commented out...
 123         //return 0;
 124 }
 125
 126 /**
 127  * Run a script. argv[2] is already NULL.
 128  */
 129 static int run(char *argv[3], const char *intf, struct in_addr *ip)
 130 {
 131         int status;
 132
 133         VDBG("%s run %s %s\n", intf, argv[0], argv[1]);
 134
 135         if (ip) {
 136                 char *addr = inet_ntoa(*ip);
 137                 setenv("ip", addr, 1);
 138                 bb_info_msg("%s %s %s", argv[1], intf, addr);
 139         }
 140
 141         status = wait4pid(spawn(argv));
 142         if (status < 0) {
 143                 bb_perror_msg("%s %s", argv[1], intf);
 144                 return -errno;
 145         }
 146         if (status != 0)
 147                 bb_error_msg("script %s %s failed, exitcode=%d", argv[0], argv[1], status);
 148         return status;
 149 }
 150
 151 /**
 152  * Return milliseconds of random delay, up to "secs" seconds.
 153  */
 154 static unsigned ATTRIBUTE_ALWAYS_INLINE ms_rdelay(unsigned secs)
 155 {
 156         return lrand48() % (secs * 1000);
 157 }
 158
 159 /**
 160  * main program
 161  */
 162 int zcip_main(int argc, char **argv);
 163 int zcip_main(int argc, char **argv)
 164 {
 165         int state = PROBE;
 166         struct ether_addr eth_addr;
 167         const char *why;
 168         int fd;
 169         char *r_opt;
 170         unsigned opts;
 171
 172         /* Ugly trick, but I want these zeroed in one go */
 173         struct {
 174                 const struct in_addr null_ip;
 175                 const struct ether_addr null_addr;
 176                 struct sockaddr saddr;
 177                 struct in_addr ip;
 178                 struct ifreq ifr;
 179                 char *intf;
 180                 char *script_av[3];
 181                 suseconds_t timeout; // milliseconds
 182                 unsigned conflicts;
 183                 unsigned nprobes;
 184                 unsigned nclaims;
 185                 int ready;
 186                 int verbose;
 187         } L;
 188 #define null_ip   (L.null_ip  )
 189 #define null_addr (L.null_addr)
 190 #define saddr     (L.saddr    )
 191 #define ip        (L.ip       )
 192 #define ifr       (L.ifr      )
 193 #define intf      (L.intf     )
 194 #define script_av (L.script_av)
 195 #define timeout   (L.timeout  )
 196 #define conflicts (L.conflicts)
 197 #define nprobes   (L.nprobes  )
 198 #define nclaims   (L.nclaims  )
 199 #define ready     (L.ready    )
 200 #define verbose   (L.verbose  )
 201
 202         memset(&L, 0, sizeof(L));
 203
 204 #define FOREGROUND (opts & 1)
 205 #define QUIT       (opts & 2)
 206         // parse commandline: prog [options] ifname script
 207         // exactly 2 args; -v accumulates and implies -f
 208         opt_complementary = "=2:vv:vf";
 209         opts = getopt32(argc, argv, "fqr:v", &r_opt, &verbose);
 210         if (!FOREGROUND) {
 211                 /* Do it early, before all bb_xx_msg calls */
 212                 openlog(applet_name, 0, LOG_DAEMON);
 213                 logmode |= LOGMODE_SYSLOG;
 214         }
 215         if (opts & 4) { // -r n.n.n.n
 216                 if (inet_aton(r_opt, &ip) == 0
 217                  || (ntohl(ip.s_addr) & IN_CLASSB_NET) != LINKLOCAL_ADDR
 218                 ) {
 219                         bb_error_msg_and_die("invalid link address");
 220                 }
 221         }
 222         // On NOMMU reexec early (or else we will rerun things twice)
 223 #ifdef BB_NOMMU
 224         if (!FOREGROUND)
 225                 bb_daemonize_or_rexec(DAEMON_CHDIR_ROOT, argv);
 226 #endif
 227         argc -= optind;
 228         argv += optind;
 229
 230         intf = argv[0];
 231         script_av[0] = argv[1];
 232         setenv("interface", intf, 1);
 233
 234         // initialize the interface (modprobe, ifup, etc)
 235         script_av[1] = (char*)"init";
 236         if (run(script_av, intf, NULL))
 237                 return EXIT_FAILURE;
 238
 239         // initialize saddr
 240         //memset(&saddr, 0, sizeof(saddr));
 241         safe_strncpy(saddr.sa_data, intf, sizeof(saddr.sa_data));
 242
 243         // open an ARP socket
 244         fd = xsocket(PF_PACKET, SOCK_PACKET, htons(ETH_P_ARP));
 245         // bind to the interface's ARP socket
 246         xbind(fd, &saddr, sizeof(saddr));
 247
 248         // get the interface's ethernet address
 249         //memset(&ifr, 0, sizeof(ifr));
 250         strncpy(ifr.ifr_name, intf, sizeof(ifr.ifr_name));
 251         if (ioctl(fd, SIOCGIFHWADDR, &ifr) < 0) {
 252                 bb_perror_msg_and_die("get ethernet address");
 253         }
 254         memcpy(&eth_addr, &ifr.ifr_hwaddr.sa_data, ETH_ALEN);
 255
 256         // start with some stable ip address, either a function of
 257         // the hardware address or else the last address we used.
 258         // NOTE: the sequence of addresses we try changes only
 259         // depending on when we detect conflicts.
 260         // (SVID 3 bogon: who says that "short" is always 16 bits?)
 261         seed48( (unsigned short*)&ifr.ifr_hwaddr.sa_data );
 262         if (ip.s_addr == 0)
 263                 pick(&ip);
 264
 265         // FIXME cases to handle:
 266         //  - zcip already running!
 267         //  - link already has local address... just defend/update
 268
 269         // daemonize now; don't delay system startup
 270         if (!FOREGROUND) {
 271 #ifndef BB_NOMMU
 272                 bb_daemonize(DAEMON_CHDIR_ROOT);
 273 #endif
 274                 bb_info_msg("start, interface %s", intf);
 275         }
 276
 277         // run the dynamic address negotiation protocol,
 278         // restarting after address conflicts:
 279         //  - start with some address we want to try
 280         //  - short random delay
 281         //  - arp probes to see if another host else uses it
 282         //  - arp announcements that we're claiming it
 283         //  - use it
 284         //  - defend it, within limits
 285         while (1) {
 286                 struct pollfd fds[1];
 287                 struct timeval tv1;
 288                 struct arp_packet p;
 289
 290                 int source_ip_conflict = 0;
 291                 int target_ip_conflict = 0;
 292
 293                 fds[0].fd = fd;
 294                 fds[0].events = POLLIN;
 295                 fds[0].revents = 0;
 296
 297                 // poll, being ready to adjust current timeout
 298                 if (!timeout) {
 299                         timeout = ms_rdelay(PROBE_WAIT);
 300                         // FIXME setsockopt(fd, SO_ATTACH_FILTER, ...) to
 301                         // make the kernel filter out all packets except
 302                         // ones we'd care about.
 303                 }
 304                 // set tv1 to the point in time when we timeout
 305                 gettimeofday(&tv1, NULL);
 306                 tv1.tv_usec += (timeout % 1000) * 1000;
 307                 while (tv1.tv_usec > 1000000) {
 308                         tv1.tv_usec -= 1000000;
 309                         tv1.tv_sec++;
 310                 }
 311                 tv1.tv_sec += timeout / 1000;
 312
 313                 VDBG("...wait %ld %s nprobes=%d, nclaims=%d\n",
 314                                 timeout, intf, nprobes, nclaims);
 315                 switch (poll(fds, 1, timeout)) {
 316
 317                 // timeout
 318                 case 0:
 319                         VDBG("state = %d\n", state);
 320                         switch (state) {
 321                         case PROBE:
 322                                 // timeouts in the PROBE state mean no conflicting ARP packets
 323                                 // have been received, so we can progress through the states
 324                                 if (nprobes < PROBE_NUM) {
 325                                         nprobes++;
 326                                         VDBG("probe/%d %s@%s\n",
 327                                                         nprobes, intf, inet_ntoa(ip));
 328                                         arp(fd, &saddr, ARPOP_REQUEST,
 329                                                         &eth_addr, null_ip,
 330                                                         &null_addr, ip);
 331                                         timeout = PROBE_MIN * 1000;
 332                                         timeout += ms_rdelay(PROBE_MAX
 333                                                         - PROBE_MIN);
 334                                 }
 335                                 else {
 336                                         // Switch to announce state.
 337                                         state = ANNOUNCE;
 338                                         nclaims = 0;
 339                                         VDBG("announce/%d %s@%s\n",
 340                                                         nclaims, intf, inet_ntoa(ip));
 341                                         arp(fd, &saddr, ARPOP_REQUEST,
 342                                                         &eth_addr, ip,
 343                                                         &eth_addr, ip);
 344                                         timeout = ANNOUNCE_INTERVAL * 1000;
 345                                 }
 346                                 break;
 347                         case RATE_LIMIT_PROBE:
 348                                 // timeouts in the RATE_LIMIT_PROBE state mean no conflicting ARP packets
 349                                 // have been received, so we can move immediately to the announce state
 350                                 state = ANNOUNCE;
 351                                 nclaims = 0;
 352                                 VDBG("announce/%d %s@%s\n",
 353                                                 nclaims, intf, inet_ntoa(ip));
 354                                 arp(fd, &saddr, ARPOP_REQUEST,
 355                                                 &eth_addr, ip,
 356                                                 &eth_addr, ip);
 357                                 timeout = ANNOUNCE_INTERVAL * 1000;
 358                                 break;
 359                         case ANNOUNCE:
 360                                 // timeouts in the ANNOUNCE state mean no conflicting ARP packets
 361                                 // have been received, so we can progress through the states
 362                                 if (nclaims < ANNOUNCE_NUM) {
 363                                         nclaims++;
 364                                         VDBG("announce/%d %s@%s\n",
 365                                                         nclaims, intf, inet_ntoa(ip));
 366                                         arp(fd, &saddr, ARPOP_REQUEST,
 367                                                         &eth_addr, ip,
 368                                                         &eth_addr, ip);
 369                                         timeout = ANNOUNCE_INTERVAL * 1000;
 370                                 }
 371                                 else {
 372                                         // Switch to monitor state.
 373                                         state = MONITOR;
 374                                         // link is ok to use earlier
 375                                         // FIXME update filters
 376                                         script_av[1] = (char*)"config";
 377                                         run(script_av, intf, &ip);
 378                                         ready = 1;
 379                                         conflicts = 0;
 380                                         timeout = -1; // Never timeout in the monitor state.
 381
 382                                         // NOTE: all other exit paths
 383                                         // should deconfig ...
 384                                         if (QUIT)
 385                                                 return EXIT_SUCCESS;
 386                                 }
 387                                 break;
 388                         case DEFEND:
 389                                 // We won!  No ARP replies, so just go back to monitor.
 390                                 state = MONITOR;
 391                                 timeout = -1;
 392                                 conflicts = 0;
 393                                 break;
 394                         default:
 395                                 // Invalid, should never happen.  Restart the whole protocol.
 396                                 state = PROBE;
 397                                 pick(&ip);
 398                                 timeout = 0;
 399                                 nprobes = 0;
 400                                 nclaims = 0;
 401                                 break;
 402                         } // switch (state)
 403                         break; // case 0 (timeout)
 404                 // packets arriving
 405                 case 1:
 406                         // We need to adjust the timeout in case we didn't receive
 407                         // a conflicting packet.
 408                         if (timeout > 0) {
 409                                 struct timeval tv2;
 410
 411                                 gettimeofday(&tv2, NULL);
 412                                 if (timercmp(&tv1, &tv2, <)) {
 413                                         // Current time is greater than the expected timeout time.
 414                                         // Should never happen.
 415                                         VDBG("missed an expected timeout\n");
 416                                         timeout = 0;
 417                                 } else {
 418                                         VDBG("adjusting timeout\n");
 419                                         timersub(&tv1, &tv2, &tv1);
 420                                         timeout = 1000 * tv1.tv_sec
 421                                                         + tv1.tv_usec / 1000;
 422                                 }
 423                         }
 424
 425                         if ((fds[0].revents & POLLIN) == 0) {
 426                                 if (fds[0].revents & POLLERR) {
 427                                         // FIXME: links routinely go down;
 428                                         // this shouldn't necessarily exit.
 429                                         bb_error_msg("%s: poll error", intf);
 430                                         if (ready) {
 431                                                 script_av[1] = (char*)"deconfig";
 432                                                 run(script_av, intf, &ip);
 433                                         }
 434                                         return EXIT_FAILURE;
 435                                 }
 436                                 continue;
 437                         }
 438
 439                         // read ARP packet
 440                         if (recv(fd, &p, sizeof(p), 0) < 0) {
 441                                 why = "recv";
 442                                 goto bad;
 443                         }
 444                         if (p.hdr.ether_type != htons(ETHERTYPE_ARP))
 445                                 continue;
 446
 447 #ifdef DEBUG
 448                         {
 449                                 struct ether_addr * sha = (struct ether_addr *) p.arp.arp_sha;
 450                                 struct ether_addr * tha = (struct ether_addr *) p.arp.arp_tha;
 451                                 struct in_addr * spa = (struct in_addr *) p.arp.arp_spa;
 452                                 struct in_addr * tpa = (struct in_addr *) p.arp.arp_tpa;
 453                                 VDBG("%s recv arp type=%d, op=%d,\n",
 454                                         intf, ntohs(p.hdr.ether_type),
 455                                         ntohs(p.arp.arp_op));
 456                                 VDBG("\tsource=%s %s\n",
 457                                         ether_ntoa(sha),
 458                                         inet_ntoa(*spa));
 459                                 VDBG("\ttarget=%s %s\n",
 460                                         ether_ntoa(tha),
 461                                         inet_ntoa(*tpa));
 462                         }
 463 #endif
 464                         if (p.arp.arp_op != htons(ARPOP_REQUEST)
 465                                         && p.arp.arp_op != htons(ARPOP_REPLY))
 466                                 continue;
 467
 468                         if (memcmp(p.arp.arp_spa, &ip.s_addr, sizeof(struct in_addr)) == 0 &&
 469                                 memcmp(&eth_addr, &p.arp.arp_sha, ETH_ALEN) != 0) {
 470                                 source_ip_conflict = 1;
 471                         }
 472                         if (memcmp(p.arp.arp_tpa, &ip.s_addr, sizeof(struct in_addr)) == 0 &&
 473                                 p.arp.arp_op == htons(ARPOP_REQUEST) &&
 474                                 memcmp(&eth_addr, &p.arp.arp_tha, ETH_ALEN) != 0) {
 475                                 target_ip_conflict = 1;
 476                         }
 477
 478                         VDBG("state = %d, source ip conflict = %d, target ip conflict = %d\n",
 479                                 state, source_ip_conflict, target_ip_conflict);
 480                         switch (state) {
 481                         case PROBE:
 482                         case ANNOUNCE:
 483                                 // When probing or announcing, check for source IP conflicts
 484                                 // and other hosts doing ARP probes (target IP conflicts).
 485                                 if (source_ip_conflict || target_ip_conflict) {
 486                                         conflicts++;
 487                                         if (conflicts >= MAX_CONFLICTS) {
 488                                                 VDBG("%s ratelimit\n", intf);
 489                                                 timeout = RATE_LIMIT_INTERVAL * 1000;
 490                                                 state = RATE_LIMIT_PROBE;
 491                                         }
 492
 493                                         // restart the whole protocol
 494                                         pick(&ip);
 495                                         timeout = 0;
 496                                         nprobes = 0;
 497                                         nclaims = 0;
 498                                 }
 499                                 break;
 500                         case MONITOR:
 501                                 // If a conflict, we try to defend with a single ARP probe.
 502                                 if (source_ip_conflict) {
 503                                         VDBG("monitor conflict -- defending\n");
 504                                         state = DEFEND;
 505                                         timeout = DEFEND_INTERVAL * 1000;
 506                                         arp(fd, &saddr,
 507                                                         ARPOP_REQUEST,
 508                                                         &eth_addr, ip,
 509                                                         &eth_addr, ip);
 510                                 }
 511                                 break;
 512                         case DEFEND:
 513                                 // Well, we tried.  Start over (on conflict).
 514                                 if (source_ip_conflict) {
 515                                         state = PROBE;
 516                                         VDBG("defend conflict -- starting over\n");
 517                                         ready = 0;
 518                                         script_av[1] = (char*)"deconfig";
 519                                         run(script_av, 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 }