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 "libbb.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         xsendto(fd, &p, sizeof(p), saddr, sizeof(*saddr));
 118
 119         // Currently all callers ignore errors, that's why returns are
 120         // commented out...
 121         //return 0;
 122 }
 123
 124 /**
 125  * Run a script. argv[2] is already NULL.
 126  */
 127 static int run(char *argv[3], const char *intf, struct in_addr *ip)
 128 {
 129         int status;
 130
 131         VDBG("%s run %s %s\n", intf, argv[0], argv[1]);
 132
 133         if (ip) {
 134                 char *addr = inet_ntoa(*ip);
 135                 setenv("ip", addr, 1);
 136                 bb_info_msg("%s %s %s", argv[1], intf, addr);
 137         }
 138
 139         status = wait4pid(spawn(argv));
 140         if (status < 0) {
 141                 bb_perror_msg("%s %s", argv[1], intf);
 142                 return -errno;
 143         }
 144         if (status != 0)
 145                 bb_error_msg("script %s %s failed, exitcode=%d", argv[0], argv[1], status);
 146         return status;
 147 }
 148
 149 /**
 150  * Return milliseconds of random delay, up to "secs" seconds.
 151  */
 152 static unsigned ALWAYS_INLINE ms_rdelay(unsigned secs)
 153 {
 154         return lrand48() % (secs * 1000);
 155 }
 156
 157 /**
 158  * main program
 159  */
 160 int zcip_main(int argc, char **argv);
 161 int zcip_main(int argc, char **argv)
 162 {
 163         int state = PROBE;
 164         struct ether_addr eth_addr;
 165         const char *why;
 166         int fd;
 167         char *r_opt;
 168         unsigned opts;
 169
 170         /* Ugly trick, but I want these zeroed in one go */
 171         struct {
 172                 const struct in_addr null_ip;
 173                 const struct ether_addr null_addr;
 174                 struct sockaddr saddr;
 175                 struct in_addr ip;
 176                 struct ifreq ifr;
 177                 char *intf;
 178                 char *script_av[3];
 179                 suseconds_t timeout; // milliseconds
 180                 unsigned conflicts;
 181                 unsigned nprobes;
 182                 unsigned nclaims;
 183                 int ready;
 184                 int verbose;
 185         } L;
 186 #define null_ip   (L.null_ip  )
 187 #define null_addr (L.null_addr)
 188 #define saddr     (L.saddr    )
 189 #define ip        (L.ip       )
 190 #define ifr       (L.ifr      )
 191 #define intf      (L.intf     )
 192 #define script_av (L.script_av)
 193 #define timeout   (L.timeout  )
 194 #define conflicts (L.conflicts)
 195 #define nprobes   (L.nprobes  )
 196 #define nclaims   (L.nclaims  )
 197 #define ready     (L.ready    )
 198 #define verbose   (L.verbose  )
 199
 200         memset(&L, 0, sizeof(L));
 201
 202 #define FOREGROUND (opts & 1)
 203 #define QUIT       (opts & 2)
 204         // parse commandline: prog [options] ifname script
 205         // exactly 2 args; -v accumulates and implies -f
 206         opt_complementary = "=2:vv:vf";
 207         opts = getopt32(argc, argv, "fqr:v", &r_opt, &verbose);
 208         if (!FOREGROUND) {
 209                 /* Do it early, before all bb_xx_msg calls */
 210                 openlog(applet_name, 0, LOG_DAEMON);
 211                 logmode |= LOGMODE_SYSLOG;
 212         }
 213         if (opts & 4) { // -r n.n.n.n
 214                 if (inet_aton(r_opt, &ip) == 0
 215                  || (ntohl(ip.s_addr) & IN_CLASSB_NET) != LINKLOCAL_ADDR
 216                 ) {
 217                         bb_error_msg_and_die("invalid link address");
 218                 }
 219         }
 220         // On NOMMU reexec early (or else we will rerun things twice)
 221 #if !BB_MMU
 222         if (!FOREGROUND)
 223                 bb_daemonize_or_rexec(DAEMON_CHDIR_ROOT, argv);
 224 #endif
 225         argc -= optind;
 226         argv += optind;
 227
 228         intf = argv[0];
 229         script_av[0] = argv[1];
 230         setenv("interface", intf, 1);
 231
 232         // initialize the interface (modprobe, ifup, etc)
 233         script_av[1] = (char*)"init";
 234         if (run(script_av, intf, NULL))
 235                 return EXIT_FAILURE;
 236
 237         // initialize saddr
 238         //memset(&saddr, 0, sizeof(saddr));
 239         safe_strncpy(saddr.sa_data, intf, sizeof(saddr.sa_data));
 240
 241         // open an ARP socket
 242         fd = xsocket(PF_PACKET, SOCK_PACKET, htons(ETH_P_ARP));
 243         // bind to the interface's ARP socket
 244         xbind(fd, &saddr, sizeof(saddr));
 245
 246         // get the interface's ethernet address
 247         //memset(&ifr, 0, sizeof(ifr));
 248         strncpy(ifr.ifr_name, intf, sizeof(ifr.ifr_name));
 249         if (ioctl(fd, SIOCGIFHWADDR, &ifr) < 0) {
 250                 bb_perror_msg_and_die("get ethernet address");
 251         }
 252         memcpy(&eth_addr, &ifr.ifr_hwaddr.sa_data, ETH_ALEN);
 253
 254         // start with some stable ip address, either a function of
 255         // the hardware address or else the last address we used.
 256         // NOTE: the sequence of addresses we try changes only
 257         // depending on when we detect conflicts.
 258         // (SVID 3 bogon: who says that "short" is always 16 bits?)
 259         seed48( (unsigned short*)&ifr.ifr_hwaddr.sa_data );
 260         if (ip.s_addr == 0)
 261                 pick(&ip);
 262
 263         // FIXME cases to handle:
 264         //  - zcip already running!
 265         //  - link already has local address... just defend/update
 266
 267         // daemonize now; don't delay system startup
 268         if (!FOREGROUND) {
 269 #if BB_MMU
 270                 bb_daemonize(DAEMON_CHDIR_ROOT);
 271 #endif
 272                 bb_info_msg("start, interface %s", intf);
 273         }
 274
 275         // run the dynamic address negotiation protocol,
 276         // restarting after address conflicts:
 277         //  - start with some address we want to try
 278         //  - short random delay
 279         //  - arp probes to see if another host else uses it
 280         //  - arp announcements that we're claiming it
 281         //  - use it
 282         //  - defend it, within limits
 283         while (1) {
 284                 struct pollfd fds[1];
 285                 struct timeval tv1;
 286                 struct arp_packet p;
 287
 288                 int source_ip_conflict = 0;
 289                 int target_ip_conflict = 0;
 290
 291                 fds[0].fd = fd;
 292                 fds[0].events = POLLIN;
 293                 fds[0].revents = 0;
 294
 295                 // poll, being ready to adjust current timeout
 296                 if (!timeout) {
 297                         timeout = ms_rdelay(PROBE_WAIT);
 298                         // FIXME setsockopt(fd, SO_ATTACH_FILTER, ...) to
 299                         // make the kernel filter out all packets except
 300                         // ones we'd care about.
 301                 }
 302                 // set tv1 to the point in time when we timeout
 303                 gettimeofday(&tv1, NULL);
 304                 tv1.tv_usec += (timeout % 1000) * 1000;
 305                 while (tv1.tv_usec > 1000000) {
 306                         tv1.tv_usec -= 1000000;
 307                         tv1.tv_sec++;
 308                 }
 309                 tv1.tv_sec += timeout / 1000;
 310
 311                 VDBG("...wait %ld %s nprobes=%d, nclaims=%d\n",
 312                                 timeout, intf, nprobes, nclaims);
 313                 switch (poll(fds, 1, timeout)) {
 314
 315                 // timeout
 316                 case 0:
 317                         VDBG("state = %d\n", state);
 318                         switch (state) {
 319                         case PROBE:
 320                                 // timeouts in the PROBE state mean no conflicting ARP packets
 321                                 // have been received, so we can progress through the states
 322                                 if (nprobes < PROBE_NUM) {
 323                                         nprobes++;
 324                                         VDBG("probe/%d %s@%s\n",
 325                                                         nprobes, intf, inet_ntoa(ip));
 326                                         arp(fd, &saddr, ARPOP_REQUEST,
 327                                                         &eth_addr, null_ip,
 328                                                         &null_addr, ip);
 329                                         timeout = PROBE_MIN * 1000;
 330                                         timeout += ms_rdelay(PROBE_MAX
 331                                                         - PROBE_MIN);
 332                                 }
 333                                 else {
 334                                         // Switch to announce state.
 335                                         state = ANNOUNCE;
 336                                         nclaims = 0;
 337                                         VDBG("announce/%d %s@%s\n",
 338                                                         nclaims, intf, inet_ntoa(ip));
 339                                         arp(fd, &saddr, ARPOP_REQUEST,
 340                                                         &eth_addr, ip,
 341                                                         &eth_addr, ip);
 342                                         timeout = ANNOUNCE_INTERVAL * 1000;
 343                                 }
 344                                 break;
 345                         case RATE_LIMIT_PROBE:
 346                                 // timeouts in the RATE_LIMIT_PROBE state mean no conflicting ARP packets
 347                                 // have been received, so we can move immediately to the announce state
 348                                 state = ANNOUNCE;
 349                                 nclaims = 0;
 350                                 VDBG("announce/%d %s@%s\n",
 351                                                 nclaims, intf, inet_ntoa(ip));
 352                                 arp(fd, &saddr, ARPOP_REQUEST,
 353                                                 &eth_addr, ip,
 354                                                 &eth_addr, ip);
 355                                 timeout = ANNOUNCE_INTERVAL * 1000;
 356                                 break;
 357                         case ANNOUNCE:
 358                                 // timeouts in the ANNOUNCE state mean no conflicting ARP packets
 359                                 // have been received, so we can progress through the states
 360                                 if (nclaims < ANNOUNCE_NUM) {
 361                                         nclaims++;
 362                                         VDBG("announce/%d %s@%s\n",
 363                                                         nclaims, intf, inet_ntoa(ip));
 364                                         arp(fd, &saddr, ARPOP_REQUEST,
 365                                                         &eth_addr, ip,
 366                                                         &eth_addr, ip);
 367                                         timeout = ANNOUNCE_INTERVAL * 1000;
 368                                 }
 369                                 else {
 370                                         // Switch to monitor state.
 371                                         state = MONITOR;
 372                                         // link is ok to use earlier
 373                                         // FIXME update filters
 374                                         script_av[1] = (char*)"config";
 375                                         run(script_av, intf, &ip);
 376                                         ready = 1;
 377                                         conflicts = 0;
 378                                         timeout = -1; // Never timeout in the monitor state.
 379
 380                                         // NOTE: all other exit paths
 381                                         // should deconfig ...
 382                                         if (QUIT)
 383                                                 return EXIT_SUCCESS;
 384                                 }
 385                                 break;
 386                         case DEFEND:
 387                                 // We won!  No ARP replies, so just go back to monitor.
 388                                 state = MONITOR;
 389                                 timeout = -1;
 390                                 conflicts = 0;
 391                                 break;
 392                         default:
 393                                 // Invalid, should never happen.  Restart the whole protocol.
 394                                 state = PROBE;
 395                                 pick(&ip);
 396                                 timeout = 0;
 397                                 nprobes = 0;
 398                                 nclaims = 0;
 399                                 break;
 400                         } // switch (state)
 401                         break; // case 0 (timeout)
 402                 // packets arriving
 403                 case 1:
 404                         // We need to adjust the timeout in case we didn't receive
 405                         // a conflicting packet.
 406                         if (timeout > 0) {
 407                                 struct timeval tv2;
 408
 409                                 gettimeofday(&tv2, NULL);
 410                                 if (timercmp(&tv1, &tv2, <)) {
 411                                         // Current time is greater than the expected timeout time.
 412                                         // Should never happen.
 413                                         VDBG("missed an expected timeout\n");
 414                                         timeout = 0;
 415                                 } else {
 416                                         VDBG("adjusting timeout\n");
 417                                         timersub(&tv1, &tv2, &tv1);
 418                                         timeout = 1000 * tv1.tv_sec
 419                                                         + tv1.tv_usec / 1000;
 420                                 }
 421                         }
 422
 423                         if ((fds[0].revents & POLLIN) == 0) {
 424                                 if (fds[0].revents & POLLERR) {
 425                                         // FIXME: links routinely go down;
 426                                         // this shouldn't necessarily exit.
 427                                         bb_error_msg("%s: poll error", intf);
 428                                         if (ready) {
 429                                                 script_av[1] = (char*)"deconfig";
 430                                                 run(script_av, intf, &ip);
 431                                         }
 432                                         return EXIT_FAILURE;
 433                                 }
 434                                 continue;
 435                         }
 436
 437                         // read ARP packet
 438                         if (recv(fd, &p, sizeof(p), 0) < 0) {
 439                                 why = "recv";
 440                                 goto bad;
 441                         }
 442                         if (p.hdr.ether_type != htons(ETHERTYPE_ARP))
 443                                 continue;
 444
 445 #ifdef DEBUG
 446                         {
 447                                 struct ether_addr * sha = (struct ether_addr *) p.arp.arp_sha;
 448                                 struct ether_addr * tha = (struct ether_addr *) p.arp.arp_tha;
 449                                 struct in_addr * spa = (struct in_addr *) p.arp.arp_spa;
 450                                 struct in_addr * tpa = (struct in_addr *) p.arp.arp_tpa;
 451                                 VDBG("%s recv arp type=%d, op=%d,\n",
 452                                         intf, ntohs(p.hdr.ether_type),
 453                                         ntohs(p.arp.arp_op));
 454                                 VDBG("\tsource=%s %s\n",
 455                                         ether_ntoa(sha),
 456                                         inet_ntoa(*spa));
 457                                 VDBG("\ttarget=%s %s\n",
 458                                         ether_ntoa(tha),
 459                                         inet_ntoa(*tpa));
 460                         }
 461 #endif
 462                         if (p.arp.arp_op != htons(ARPOP_REQUEST)
 463                                         && p.arp.arp_op != htons(ARPOP_REPLY))
 464                                 continue;
 465
 466                         if (memcmp(p.arp.arp_spa, &ip.s_addr, sizeof(struct in_addr)) == 0 &&
 467                                 memcmp(&eth_addr, &p.arp.arp_sha, ETH_ALEN) != 0) {
 468                                 source_ip_conflict = 1;
 469                         }
 470                         if (memcmp(p.arp.arp_tpa, &ip.s_addr, sizeof(struct in_addr)) == 0 &&
 471                                 p.arp.arp_op == htons(ARPOP_REQUEST) &&
 472                                 memcmp(&eth_addr, &p.arp.arp_tha, ETH_ALEN) != 0) {
 473                                 target_ip_conflict = 1;
 474                         }
 475
 476                         VDBG("state = %d, source ip conflict = %d, target ip conflict = %d\n",
 477                                 state, source_ip_conflict, target_ip_conflict);
 478                         switch (state) {
 479                         case PROBE:
 480                         case ANNOUNCE:
 481                                 // When probing or announcing, check for source IP conflicts
 482                                 // and other hosts doing ARP probes (target IP conflicts).
 483                                 if (source_ip_conflict || target_ip_conflict) {
 484                                         conflicts++;
 485                                         if (conflicts >= MAX_CONFLICTS) {
 486                                                 VDBG("%s ratelimit\n", intf);
 487                                                 timeout = RATE_LIMIT_INTERVAL * 1000;
 488                                                 state = RATE_LIMIT_PROBE;
 489                                         }
 490
 491                                         // restart the whole protocol
 492                                         pick(&ip);
 493                                         timeout = 0;
 494                                         nprobes = 0;
 495                                         nclaims = 0;
 496                                 }
 497                                 break;
 498                         case MONITOR:
 499                                 // If a conflict, we try to defend with a single ARP probe.
 500                                 if (source_ip_conflict) {
 501                                         VDBG("monitor conflict -- defending\n");
 502                                         state = DEFEND;
 503                                         timeout = DEFEND_INTERVAL * 1000;
 504                                         arp(fd, &saddr,
 505                                                         ARPOP_REQUEST,
 506                                                         &eth_addr, ip,
 507                                                         &eth_addr, ip);
 508                                 }
 509                                 break;
 510                         case DEFEND:
 511                                 // Well, we tried.  Start over (on conflict).
 512                                 if (source_ip_conflict) {
 513                                         state = PROBE;
 514                                         VDBG("defend conflict -- starting over\n");
 515                                         ready = 0;
 516                                         script_av[1] = (char*)"deconfig";
 517                                         run(script_av, intf, &ip);
 518
 519                                         // restart the whole protocol
 520                                         pick(&ip);
 521                                         timeout = 0;
 522                                         nprobes = 0;
 523                                         nclaims = 0;
 524                                 }
 525                                 break;
 526                         default:
 527                                 // Invalid, should never happen.  Restart the whole protocol.
 528                                 VDBG("invalid state -- starting over\n");
 529                                 state = PROBE;
 530                                 pick(&ip);
 531                                 timeout = 0;
 532                                 nprobes = 0;
 533                                 nclaims = 0;
 534                                 break;
 535                         } // switch state
 536
 537                         break; // case 1 (packets arriving)
 538                 default:
 539                         why = "poll";
 540                         goto bad;
 541                 } // switch poll
 542         }
 543  bad:
 544         bb_perror_msg("%s, %s", intf, why);
 545         return EXIT_FAILURE;
 546 }