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 <syslog.h>
  27 #include <poll.h>
  28 #include <sys/wait.h>
  29 #include <netinet/ether.h>
  30 #include <net/ethernet.h>
  31 #include <net/if.h>
  32 #include <net/if_arp.h>
  33 #include <linux/if_packet.h>
  34 #include <linux/sockios.h>
  35
  36 #include "libbb.h"
  37
  38 /* We don't need more than 32 bits of the counter */
  39 #define MONOTONIC_US() ((unsigned)monotonic_us())
  40
  41 struct arp_packet {
  42         struct ether_header hdr;
  43         struct ether_arp arp;
  44 } ATTRIBUTE_PACKED;
  45
  46 enum {
  47 /* 169.254.0.0 */
  48         LINKLOCAL_ADDR = 0xa9fe0000,
  49
  50 /* protocol timeout parameters, specified in seconds */
  51         PROBE_WAIT = 1,
  52         PROBE_MIN = 1,
  53         PROBE_MAX = 2,
  54         PROBE_NUM = 3,
  55         MAX_CONFLICTS = 10,
  56         RATE_LIMIT_INTERVAL = 60,
  57         ANNOUNCE_WAIT = 2,
  58         ANNOUNCE_NUM = 2,
  59         ANNOUNCE_INTERVAL = 2,
  60         DEFEND_INTERVAL = 10
  61 };
  62
  63 /* States during the configuration process. */
  64 enum {
  65         PROBE = 0,
  66         RATE_LIMIT_PROBE,
  67         ANNOUNCE,
  68         MONITOR,
  69         DEFEND
  70 };
  71
  72 #define VDBG(fmt,args...) \
  73         do { } while (0)
  74
  75 /**
  76  * Pick a random link local IP address on 169.254/16, except that
  77  * the first and last 256 addresses are reserved.
  78  */
  79 static void pick(struct in_addr *ip)
  80 {
  81         unsigned tmp;
  82
  83         do {
  84                 tmp = rand() & IN_CLASSB_HOST;
  85         } while (tmp > (IN_CLASSB_HOST - 0x0200));
  86         ip->s_addr = htonl((LINKLOCAL_ADDR + 0x0100) + tmp);
  87 }
  88
  89 /**
  90  * Broadcast an ARP packet.
  91  */
  92 static void arp(int fd, struct sockaddr *saddr, int op,
  93         const struct ether_addr *source_addr, struct in_addr source_ip,
  94         const struct ether_addr *target_addr, struct in_addr target_ip)
  95 {
  96         struct arp_packet p;
  97         memset(&p, 0, sizeof(p));
  98
  99         // ether header
 100         p.hdr.ether_type = htons(ETHERTYPE_ARP);
 101         memcpy(p.hdr.ether_shost, source_addr, ETH_ALEN);
 102         memset(p.hdr.ether_dhost, 0xff, ETH_ALEN);
 103
 104         // arp request
 105         p.arp.arp_hrd = htons(ARPHRD_ETHER);
 106         p.arp.arp_pro = htons(ETHERTYPE_IP);
 107         p.arp.arp_hln = ETH_ALEN;
 108         p.arp.arp_pln = 4;
 109         p.arp.arp_op = htons(op);
 110         memcpy(&p.arp.arp_sha, source_addr, ETH_ALEN);
 111         memcpy(&p.arp.arp_spa, &source_ip, sizeof(p.arp.arp_spa));
 112         memcpy(&p.arp.arp_tha, target_addr, ETH_ALEN);
 113         memcpy(&p.arp.arp_tpa, &target_ip, sizeof(p.arp.arp_tpa));
 114
 115         // send it
 116         xsendto(fd, &p, sizeof(p), saddr, sizeof(*saddr));
 117
 118         // Currently all callers ignore errors, that's why returns are
 119         // commented out...
 120         //return 0;
 121 }
 122
 123 /**
 124  * Run a script. argv[2] is already NULL.
 125  */
 126 static int run(char *argv[3], const char *intf, struct in_addr *ip)
 127 {
 128         int status;
 129
 130         VDBG("%s run %s %s\n", intf, argv[0], argv[1]);
 131
 132         if (ip) {
 133                 char *addr = inet_ntoa(*ip);
 134                 setenv("ip", addr, 1);
 135                 bb_info_msg("%s %s %s", argv[1], intf, addr);
 136         }
 137
 138         status = wait4pid(spawn(argv));
 139         if (status < 0) {
 140                 bb_perror_msg("%s %s", argv[1], intf);
 141                 return -errno;
 142         }
 143         if (status != 0)
 144                 bb_error_msg("script %s %s failed, exitcode=%d", argv[0], argv[1], status);
 145         return status;
 146 }
 147
 148 /**
 149  * Return milliseconds of random delay, up to "secs" seconds.
 150  */
 151 static unsigned ALWAYS_INLINE ms_rdelay(unsigned secs)
 152 {
 153         return rand() % (secs * 1000);
 154 }
 155
 156 /**
 157  * main program
 158  */
 159 int zcip_main(int argc, char **argv);
 160 int zcip_main(int argc, char **argv)
 161 {
 162         int state = PROBE;
 163         struct ether_addr eth_addr;
 164         const char *why;
 165         int fd;
 166         char *r_opt;
 167         unsigned opts;
 168
 169         /* Ugly trick, but I want these zeroed in one go */
 170         struct {
 171                 const struct in_addr null_ip;
 172                 const struct ether_addr null_addr;
 173                 struct sockaddr saddr;
 174                 struct in_addr ip;
 175                 struct ifreq ifr;
 176                 char *intf;
 177                 char *script_av[3];
 178                 int timeout_ms; /* must be signed */
 179                 unsigned conflicts;
 180                 unsigned nprobes;
 181                 unsigned nclaims;
 182                 int ready;
 183                 int verbose;
 184         } L;
 185 #define null_ip    (L.null_ip   )
 186 #define null_addr  (L.null_addr )
 187 #define saddr      (L.saddr     )
 188 #define ip         (L.ip        )
 189 #define ifr        (L.ifr       )
 190 #define intf       (L.intf      )
 191 #define script_av  (L.script_av )
 192 #define timeout_ms (L.timeout_ms)
 193 #define conflicts  (L.conflicts )
 194 #define nprobes    (L.nprobes   )
 195 #define nclaims    (L.nclaims   )
 196 #define ready      (L.ready     )
 197 #define verbose    (L.verbose   )
 198
 199         memset(&L, 0, sizeof(L));
 200
 201         srand(MONOTONIC_US());
 202
 203 #define FOREGROUND (opts & 1)
 204 #define QUIT       (opts & 2)
 205         // parse commandline: prog [options] ifname script
 206         // exactly 2 args; -v accumulates and implies -f
 207         opt_complementary = "=2:vv:vf";
 208         opts = getopt32(argc, argv, "fqr:v", &r_opt, &verbose);
 209         if (!FOREGROUND) {
 210                 /* Do it early, before all bb_xx_msg calls */
 211                 openlog(applet_name, 0, LOG_DAEMON);
 212                 logmode |= LOGMODE_SYSLOG;
 213         }
 214         if (opts & 4) { // -r n.n.n.n
 215                 if (inet_aton(r_opt, &ip) == 0
 216                  || (ntohl(ip.s_addr) & IN_CLASSB_NET) != LINKLOCAL_ADDR
 217                 ) {
 218                         bb_error_msg_and_die("invalid link address");
 219                 }
 220         }
 221         // On NOMMU reexec early (or else we will rerun things twice)
 222 #if !BB_MMU
 223         if (!FOREGROUND)
 224                 bb_daemonize_or_rexec(DAEMON_CHDIR_ROOT, argv);
 225 #endif
 226         argc -= optind;
 227         argv += optind;
 228
 229         intf = argv[0];
 230         script_av[0] = argv[1];
 231         setenv("interface", intf, 1);
 232
 233         // initialize the interface (modprobe, ifup, etc)
 234         script_av[1] = (char*)"init";
 235         if (run(script_av, intf, NULL))
 236                 return EXIT_FAILURE;
 237
 238         // initialize saddr
 239         //memset(&saddr, 0, sizeof(saddr));
 240         safe_strncpy(saddr.sa_data, intf, sizeof(saddr.sa_data));
 241
 242         // open an ARP socket
 243         fd = xsocket(PF_PACKET, SOCK_PACKET, htons(ETH_P_ARP));
 244         // bind to the interface's ARP socket
 245         xbind(fd, &saddr, sizeof(saddr));
 246
 247         // get the interface's ethernet address
 248         //memset(&ifr, 0, sizeof(ifr));
 249         strncpy(ifr.ifr_name, intf, sizeof(ifr.ifr_name));
 250         if (ioctl(fd, SIOCGIFHWADDR, &ifr) < 0) {
 251                 bb_perror_msg_and_die("get ethernet address");
 252         }
 253         memcpy(&eth_addr, &ifr.ifr_hwaddr.sa_data, ETH_ALEN);
 254
 255         // start with some stable ip address, either a function of
 256         // the hardware address or else the last address we used.
 257         // NOTE: the sequence of addresses we try changes only
 258         // depending on when we detect conflicts.
 259         // (SVID 3 bogon: who says that "short" is always 16 bits?)
 260         seed48( (unsigned short*)&ifr.ifr_hwaddr.sa_data );
 261         if (ip.s_addr == 0)
 262                 pick(&ip);
 263
 264         // FIXME cases to handle:
 265         //  - zcip already running!
 266         //  - link already has local address... just defend/update
 267
 268         // daemonize now; don't delay system startup
 269         if (!FOREGROUND) {
 270 #if BB_MMU
 271                 bb_daemonize(DAEMON_CHDIR_ROOT);
 272 #endif
 273                 bb_info_msg("start, interface %s", intf);
 274         }
 275
 276         // run the dynamic address negotiation protocol,
 277         // restarting after address conflicts:
 278         //  - start with some address we want to try
 279         //  - short random delay
 280         //  - arp probes to see if another host else uses it
 281         //  - arp announcements that we're claiming it
 282         //  - use it
 283         //  - defend it, within limits
 284         while (1) {
 285                 struct pollfd fds[1];
 286                 unsigned deadline_us;
 287                 struct arp_packet p;
 288
 289                 int source_ip_conflict = 0;
 290                 int target_ip_conflict = 0;
 291
 292                 fds[0].fd = fd;
 293                 fds[0].events = POLLIN;
 294                 fds[0].revents = 0;
 295
 296                 // poll, being ready to adjust current timeout
 297                 if (!timeout_ms) {
 298                         timeout_ms = ms_rdelay(PROBE_WAIT);
 299                         // FIXME setsockopt(fd, SO_ATTACH_FILTER, ...) to
 300                         // make the kernel filter out all packets except
 301                         // ones we'd care about.
 302                 }
 303                 // set deadline_us to the point in time when we timeout
 304                 deadline_us = MONOTONIC_US() + timeout_ms * 1000;
 305
 306                 VDBG("...wait %d %s nprobes=%u, nclaims=%u\n",
 307                                 timeout_ms, intf, nprobes, nclaims);
 308                 switch (poll(fds, 1, timeout_ms)) {
 309
 310                 // timeout
 311                 case 0:
 312                         VDBG("state = %d\n", state);
 313                         switch (state) {
 314                         case PROBE:
 315                                 // timeouts in the PROBE state mean no conflicting ARP packets
 316                                 // have been received, so we can progress through the states
 317                                 if (nprobes < PROBE_NUM) {
 318                                         nprobes++;
 319                                         VDBG("probe/%u %s@%s\n",
 320                                                         nprobes, intf, inet_ntoa(ip));
 321                                         arp(fd, &saddr, ARPOP_REQUEST,
 322                                                         &eth_addr, null_ip,
 323                                                         &null_addr, ip);
 324                                         timeout_ms = PROBE_MIN * 1000;
 325                                         timeout_ms += ms_rdelay(PROBE_MAX - PROBE_MIN);
 326                                 }
 327                                 else {
 328                                         // Switch to announce state.
 329                                         state = ANNOUNCE;
 330                                         nclaims = 0;
 331                                         VDBG("announce/%u %s@%s\n",
 332                                                         nclaims, intf, inet_ntoa(ip));
 333                                         arp(fd, &saddr, ARPOP_REQUEST,
 334                                                         &eth_addr, ip,
 335                                                         &eth_addr, ip);
 336                                         timeout_ms = ANNOUNCE_INTERVAL * 1000;
 337                                 }
 338                                 break;
 339                         case RATE_LIMIT_PROBE:
 340                                 // timeouts in the RATE_LIMIT_PROBE state mean no conflicting ARP packets
 341                                 // have been received, so we can move immediately to the announce state
 342                                 state = ANNOUNCE;
 343                                 nclaims = 0;
 344                                 VDBG("announce/%u %s@%s\n",
 345                                                 nclaims, intf, inet_ntoa(ip));
 346                                 arp(fd, &saddr, ARPOP_REQUEST,
 347                                                 &eth_addr, ip,
 348                                                 &eth_addr, ip);
 349                                 timeout_ms = ANNOUNCE_INTERVAL * 1000;
 350                                 break;
 351                         case ANNOUNCE:
 352                                 // timeouts in the ANNOUNCE state mean no conflicting ARP packets
 353                                 // have been received, so we can progress through the states
 354                                 if (nclaims < ANNOUNCE_NUM) {
 355                                         nclaims++;
 356                                         VDBG("announce/%u %s@%s\n",
 357                                                         nclaims, intf, inet_ntoa(ip));
 358                                         arp(fd, &saddr, ARPOP_REQUEST,
 359                                                         &eth_addr, ip,
 360                                                         &eth_addr, ip);
 361                                         timeout_ms = ANNOUNCE_INTERVAL * 1000;
 362                                 }
 363                                 else {
 364                                         // Switch to monitor state.
 365                                         state = MONITOR;
 366                                         // link is ok to use earlier
 367                                         // FIXME update filters
 368                                         script_av[1] = (char*)"config";
 369                                         run(script_av, intf, &ip);
 370                                         ready = 1;
 371                                         conflicts = 0;
 372                                         timeout_ms = -1; // Never timeout in the monitor state.
 373
 374                                         // NOTE: all other exit paths
 375                                         // should deconfig ...
 376                                         if (QUIT)
 377                                                 return EXIT_SUCCESS;
 378                                 }
 379                                 break;
 380                         case DEFEND:
 381                                 // We won!  No ARP replies, so just go back to monitor.
 382                                 state = MONITOR;
 383                                 timeout_ms = -1;
 384                                 conflicts = 0;
 385                                 break;
 386                         default:
 387                                 // Invalid, should never happen.  Restart the whole protocol.
 388                                 state = PROBE;
 389                                 pick(&ip);
 390                                 timeout_ms = 0;
 391                                 nprobes = 0;
 392                                 nclaims = 0;
 393                                 break;
 394                         } // switch (state)
 395                         break; // case 0 (timeout)
 396                 // packets arriving
 397                 case 1:
 398                         // We need to adjust the timeout in case we didn't receive
 399                         // a conflicting packet.
 400                         if (timeout_ms > 0) {
 401                                 unsigned diff = deadline_us - MONOTONIC_US();
 402                                 if ((int)(diff) < 0) {
 403                                         // Current time is greater than the expected timeout time.
 404                                         // Should never happen.
 405                                         VDBG("missed an expected timeout\n");
 406                                         timeout_ms = 0;
 407                                 } else {
 408                                         VDBG("adjusting timeout\n");
 409                                         timeout_ms = diff / 1000;
 410                                         if (!timeout_ms) timeout_ms = 1;
 411                                 }
 412                         }
 413
 414                         if ((fds[0].revents & POLLIN) == 0) {
 415                                 if (fds[0].revents & POLLERR) {
 416                                         // FIXME: links routinely go down;
 417                                         // this shouldn't necessarily exit.
 418                                         bb_error_msg("%s: poll error", intf);
 419                                         if (ready) {
 420                                                 script_av[1] = (char*)"deconfig";
 421                                                 run(script_av, intf, &ip);
 422                                         }
 423                                         return EXIT_FAILURE;
 424                                 }
 425                                 continue;
 426                         }
 427
 428                         // read ARP packet
 429                         if (recv(fd, &p, sizeof(p), 0) < 0) {
 430                                 why = "recv";
 431                                 goto bad;
 432                         }
 433                         if (p.hdr.ether_type != htons(ETHERTYPE_ARP))
 434                                 continue;
 435
 436 #ifdef DEBUG
 437                         {
 438                                 struct ether_addr * sha = (struct ether_addr *) p.arp.arp_sha;
 439                                 struct ether_addr * tha = (struct ether_addr *) p.arp.arp_tha;
 440                                 struct in_addr * spa = (struct in_addr *) p.arp.arp_spa;
 441                                 struct in_addr * tpa = (struct in_addr *) p.arp.arp_tpa;
 442                                 VDBG("%s recv arp type=%d, op=%d,\n",
 443                                         intf, ntohs(p.hdr.ether_type),
 444                                         ntohs(p.arp.arp_op));
 445                                 VDBG("\tsource=%s %s\n",
 446                                         ether_ntoa(sha),
 447                                         inet_ntoa(*spa));
 448                                 VDBG("\ttarget=%s %s\n",
 449                                         ether_ntoa(tha),
 450                                         inet_ntoa(*tpa));
 451                         }
 452 #endif
 453                         if (p.arp.arp_op != htons(ARPOP_REQUEST)
 454                                         && p.arp.arp_op != htons(ARPOP_REPLY))
 455                                 continue;
 456
 457                         if (memcmp(p.arp.arp_spa, &ip.s_addr, sizeof(struct in_addr)) == 0 &&
 458                                 memcmp(&eth_addr, &p.arp.arp_sha, ETH_ALEN) != 0) {
 459                                 source_ip_conflict = 1;
 460                         }
 461                         if (memcmp(p.arp.arp_tpa, &ip.s_addr, sizeof(struct in_addr)) == 0 &&
 462                                 p.arp.arp_op == htons(ARPOP_REQUEST) &&
 463                                 memcmp(&eth_addr, &p.arp.arp_tha, ETH_ALEN) != 0) {
 464                                 target_ip_conflict = 1;
 465                         }
 466
 467                         VDBG("state = %d, source ip conflict = %d, target ip conflict = %d\n",
 468                                 state, source_ip_conflict, target_ip_conflict);
 469                         switch (state) {
 470                         case PROBE:
 471                         case ANNOUNCE:
 472                                 // When probing or announcing, check for source IP conflicts
 473                                 // and other hosts doing ARP probes (target IP conflicts).
 474                                 if (source_ip_conflict || target_ip_conflict) {
 475                                         conflicts++;
 476                                         if (conflicts >= MAX_CONFLICTS) {
 477                                                 VDBG("%s ratelimit\n", intf);
 478                                                 timeout_ms = RATE_LIMIT_INTERVAL * 1000;
 479                                                 state = RATE_LIMIT_PROBE;
 480                                         }
 481
 482                                         // restart the whole protocol
 483                                         pick(&ip);
 484                                         timeout_ms = 0;
 485                                         nprobes = 0;
 486                                         nclaims = 0;
 487                                 }
 488                                 break;
 489                         case MONITOR:
 490                                 // If a conflict, we try to defend with a single ARP probe.
 491                                 if (source_ip_conflict) {
 492                                         VDBG("monitor conflict -- defending\n");
 493                                         state = DEFEND;
 494                                         timeout_ms = DEFEND_INTERVAL * 1000;
 495                                         arp(fd, &saddr,
 496                                                         ARPOP_REQUEST,
 497                                                         &eth_addr, ip,
 498                                                         &eth_addr, ip);
 499                                 }
 500                                 break;
 501                         case DEFEND:
 502                                 // Well, we tried.  Start over (on conflict).
 503                                 if (source_ip_conflict) {
 504                                         state = PROBE;
 505                                         VDBG("defend conflict -- starting over\n");
 506                                         ready = 0;
 507                                         script_av[1] = (char*)"deconfig";
 508                                         run(script_av, intf, &ip);
 509
 510                                         // restart the whole protocol
 511                                         pick(&ip);
 512                                         timeout_ms = 0;
 513                                         nprobes = 0;
 514                                         nclaims = 0;
 515                                 }
 516                                 break;
 517                         default:
 518                                 // Invalid, should never happen.  Restart the whole protocol.
 519                                 VDBG("invalid state -- starting over\n");
 520                                 state = PROBE;
 521                                 pick(&ip);
 522                                 timeout_ms = 0;
 523                                 nprobes = 0;
 524                                 nclaims = 0;
 525                                 break;
 526                         } // switch state
 527
 528                         break; // case 1 (packets arriving)
 529                 default:
 530                         why = "poll";
 531                         goto bad;
 532                 } // switch poll
 533         }
 534  bad:
 535         bb_perror_msg("%s, %s", intf, why);
 536         return EXIT_FAILURE;
 537 }