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 eth;
  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(...) do { } while (0)
  73
  74
  75 enum {
  76         sock_fd = 3
  77 };
  78
  79 struct globals {
  80         char *intf;
  81         struct sockaddr saddr;
  82 };
  83 #define G (*(struct globals*)&bb_common_bufsiz1)
  84 #define intf  (G.intf )
  85 #define saddr (G.saddr)
  86
  87
  88 /**
  89  * Pick a random link local IP address on 169.254/16, except that
  90  * the first and last 256 addresses are reserved.
  91  */
  92 static void pick(struct in_addr *ip)
  93 {
  94         unsigned tmp;
  95
  96         do {
  97                 tmp = rand() & IN_CLASSB_HOST;
  98         } while (tmp > (IN_CLASSB_HOST - 0x0200));
  99         ip->s_addr = htonl((LINKLOCAL_ADDR + 0x0100) + tmp);
 100 }
 101
 102 /**
 103  * Broadcast an ARP packet.
 104  */
 105 static void arp(int op,
 106         const struct ether_addr *source_eth, struct in_addr source_ip,
 107         const struct ether_addr *target_eth, struct in_addr target_ip)
 108 {
 109         struct arp_packet p;
 110         memset(&p, 0, sizeof(p));
 111
 112         // ether header
 113         p.eth.ether_type = htons(ETHERTYPE_ARP);
 114         memcpy(p.eth.ether_shost, source_eth, ETH_ALEN);
 115         memset(p.eth.ether_dhost, 0xff, ETH_ALEN);
 116
 117         // arp request
 118         p.arp.arp_hrd = htons(ARPHRD_ETHER);
 119         p.arp.arp_pro = htons(ETHERTYPE_IP);
 120         p.arp.arp_hln = ETH_ALEN;
 121         p.arp.arp_pln = 4;
 122         p.arp.arp_op = htons(op);
 123         memcpy(&p.arp.arp_sha, source_eth, ETH_ALEN);
 124         memcpy(&p.arp.arp_spa, &source_ip, sizeof(p.arp.arp_spa));
 125         memcpy(&p.arp.arp_tha, target_eth, ETH_ALEN);
 126         memcpy(&p.arp.arp_tpa, &target_ip, sizeof(p.arp.arp_tpa));
 127
 128         // send it
 129         xsendto(sock_fd, &p, sizeof(p), &saddr, sizeof(saddr));
 130
 131         // Currently all callers ignore errors, that's why returns are
 132         // commented out...
 133         //return 0;
 134 }
 135
 136 /**
 137  * Run a script. argv[2] is already NULL.
 138  */
 139 static int run(char *argv[3], struct in_addr *ip)
 140 {
 141         int status;
 142         char *addr = addr; /* for gcc */
 143         const char *fmt = "%s %s %s" + 3;
 144
 145         VDBG("%s run %s %s\n", intf, argv[0], argv[1]);
 146
 147         if (ip) {
 148                 addr = inet_ntoa(*ip);
 149                 setenv("ip", addr, 1);
 150                 fmt -= 3;
 151         }
 152         bb_info_msg(fmt, argv[1], intf, addr);
 153
 154         status = wait4pid(spawn(argv));
 155         if (status < 0) {
 156                 bb_perror_msg("%s %s %s" + 3, argv[1], intf);
 157                 return -errno;
 158         }
 159         if (status != 0)
 160                 bb_error_msg("script %s %s failed, exitcode=%d", argv[0], argv[1], status);
 161         return status;
 162 }
 163
 164 /**
 165  * Return milliseconds of random delay, up to "secs" seconds.
 166  */
 167 static unsigned ALWAYS_INLINE random_delay_ms(unsigned secs)
 168 {
 169         return rand() % (secs * 1000);
 170 }
 171
 172 /**
 173  * main program
 174  */
 175 int zcip_main(int argc, char **argv) MAIN_EXTERNALLY_VISIBLE;
 176 int zcip_main(int argc, char **argv)
 177 {
 178         int state = PROBE;
 179         struct ether_addr eth_addr;
 180         const char *why;
 181         char *r_opt;
 182         unsigned opts;
 183
 184         /* Ugly trick, but I want these zeroed in one go */
 185         struct {
 186                 const struct in_addr null_ip;
 187                 const struct ether_addr null_addr;
 188                 struct in_addr ip;
 189                 struct ifreq ifr;
 190                 char *script_av[3];
 191                 int timeout_ms; /* must be signed */
 192                 unsigned conflicts;
 193                 unsigned nprobes;
 194                 unsigned nclaims;
 195                 int ready;
 196                 int verbose;
 197         } L;
 198 #define null_ip    (L.null_ip   )
 199 #define null_addr  (L.null_addr )
 200 #define ip         (L.ip        )
 201 #define ifr        (L.ifr       )
 202 #define script_av  (L.script_av )
 203 #define timeout_ms (L.timeout_ms)
 204 #define conflicts  (L.conflicts )
 205 #define nprobes    (L.nprobes   )
 206 #define nclaims    (L.nclaims   )
 207 #define ready      (L.ready     )
 208 #define verbose    (L.verbose   )
 209
 210         memset(&L, 0, sizeof(L));
 211
 212 #define FOREGROUND (opts & 1)
 213 #define QUIT       (opts & 2)
 214         // parse commandline: prog [options] ifname script
 215         // exactly 2 args; -v accumulates and implies -f
 216         opt_complementary = "=2:vv:vf";
 217         opts = getopt32(argv, "fqr:v", &r_opt, &verbose);
 218         if (!FOREGROUND) {
 219                 /* Do it early, before all bb_xx_msg calls */
 220                 openlog(applet_name, 0, LOG_DAEMON);
 221                 logmode |= LOGMODE_SYSLOG;
 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         // On NOMMU reexec early (or else we will rerun things twice)
 231 #if !BB_MMU
 232         if (!FOREGROUND)
 233                 bb_daemonize_or_rexec(DAEMON_CHDIR_ROOT, argv);
 234 #endif
 235         argc -= optind;
 236         argv += optind;
 237
 238         intf = argv[0];
 239         script_av[0] = argv[1];
 240         setenv("interface", intf, 1);
 241
 242         // initialize the interface (modprobe, ifup, etc)
 243         script_av[1] = (char*)"init";
 244         if (run(script_av, NULL))
 245                 return EXIT_FAILURE;
 246
 247         // initialize saddr
 248         // saddr is: { u16 sa_family; u8 sa_data[14]; }
 249         //memset(&saddr, 0, sizeof(saddr));
 250         //TODO: are we leaving sa_family == 0 (AF_UNSPEC)?!
 251         safe_strncpy(saddr.sa_data, intf, sizeof(saddr.sa_data));
 252
 253         // open an ARP socket
 254         xmove_fd(xsocket(AF_PACKET, SOCK_PACKET, htons(ETH_P_ARP)), sock_fd);
 255         // bind to the interface's ARP socket
 256         xbind(sock_fd, &saddr, sizeof(saddr));
 257
 258         // get the interface's ethernet address
 259         //memset(&ifr, 0, sizeof(ifr));
 260         strncpy(ifr.ifr_name, intf, sizeof(ifr.ifr_name));
 261         xioctl(sock_fd, SIOCGIFHWADDR, &ifr);
 262         memcpy(&eth_addr, &ifr.ifr_hwaddr.sa_data, ETH_ALEN);
 263
 264         // start with some stable ip address, either a function of
 265         // the hardware address or else the last address we used.
 266         // we are taking low-order four bytes, as top-order ones
 267         // aren't random enough.
 268         // NOTE: the sequence of addresses we try changes only
 269         // depending on when we detect conflicts.
 270         {
 271                 uint32_t t;
 272                 memcpy(&t, (char*)&eth_addr + 2, 4);
 273                 srand(t);
 274         }
 275         if (ip.s_addr == 0)
 276                 pick(&ip);
 277
 278         // FIXME cases to handle:
 279         //  - zcip already running!
 280         //  - link already has local address... just defend/update
 281
 282         // daemonize now; don't delay system startup
 283         if (!FOREGROUND) {
 284 #if BB_MMU
 285                 bb_daemonize(DAEMON_CHDIR_ROOT);
 286 #endif
 287                 bb_info_msg("start, interface %s", intf);
 288         }
 289
 290         // run the dynamic address negotiation protocol,
 291         // restarting after address conflicts:
 292         //  - start with some address we want to try
 293         //  - short random delay
 294         //  - arp probes to see if another host else uses it
 295         //  - arp announcements that we're claiming it
 296         //  - use it
 297         //  - defend it, within limits
 298         while (1) {
 299                 struct pollfd fds[1];
 300                 unsigned deadline_us;
 301                 struct arp_packet p;
 302                 int source_ip_conflict;
 303                 int target_ip_conflict;
 304
 305                 fds[0].fd = sock_fd;
 306                 fds[0].events = POLLIN;
 307                 fds[0].revents = 0;
 308
 309                 // poll, being ready to adjust current timeout
 310                 if (!timeout_ms) {
 311                         timeout_ms = random_delay_ms(PROBE_WAIT);
 312                         // FIXME setsockopt(sock_fd, SO_ATTACH_FILTER, ...) to
 313                         // make the kernel filter out all packets except
 314                         // ones we'd care about.
 315                 }
 316                 // set deadline_us to the point in time when we timeout
 317                 deadline_us = MONOTONIC_US() + timeout_ms * 1000;
 318
 319                 VDBG("...wait %d %s nprobes=%u, nclaims=%u\n",
 320                                 timeout_ms, intf, nprobes, nclaims);
 321
 322         // FIXME: do we really receive ALL packets here??
 323         // if yes, set up filtering to get ARPs only!!! (see arping)
 324                 switch (safe_poll(fds, 1, timeout_ms)) {
 325
 326                 default:
 327                         /*bb_perror_msg("poll"); - done in safe_poll */
 328                         return EXIT_FAILURE;
 329
 330                 // timeout
 331                 case 0:
 332                         VDBG("state = %d\n", state);
 333                         switch (state) {
 334                         case PROBE:
 335                                 // timeouts in the PROBE state mean no conflicting ARP packets
 336                                 // have been received, so we can progress through the states
 337                                 if (nprobes < PROBE_NUM) {
 338                                         nprobes++;
 339                                         VDBG("probe/%u %s@%s\n",
 340                                                         nprobes, intf, inet_ntoa(ip));
 341                                         arp(ARPOP_REQUEST,
 342                                                         &eth_addr, null_ip,
 343                                                         &null_addr, ip);
 344                                         timeout_ms = PROBE_MIN * 1000;
 345                                         timeout_ms += random_delay_ms(PROBE_MAX - PROBE_MIN);
 346                                 }
 347                                 else {
 348                                         // Switch to announce state.
 349                                         state = ANNOUNCE;
 350                                         nclaims = 0;
 351                                         VDBG("announce/%u %s@%s\n",
 352                                                         nclaims, intf, inet_ntoa(ip));
 353                                         arp(ARPOP_REQUEST,
 354                                                         &eth_addr, ip,
 355                                                         &eth_addr, ip);
 356                                         timeout_ms = ANNOUNCE_INTERVAL * 1000;
 357                                 }
 358                                 break;
 359                         case RATE_LIMIT_PROBE:
 360                                 // timeouts in the RATE_LIMIT_PROBE state mean no conflicting ARP packets
 361                                 // have been received, so we can move immediately to the announce state
 362                                 state = ANNOUNCE;
 363                                 nclaims = 0;
 364                                 VDBG("announce/%u %s@%s\n",
 365                                                 nclaims, intf, inet_ntoa(ip));
 366                                 arp(ARPOP_REQUEST,
 367                                                 &eth_addr, ip,
 368                                                 &eth_addr, ip);
 369                                 timeout_ms = ANNOUNCE_INTERVAL * 1000;
 370                                 break;
 371                         case ANNOUNCE:
 372                                 // timeouts in the ANNOUNCE state mean no conflicting ARP packets
 373                                 // have been received, so we can progress through the states
 374                                 if (nclaims < ANNOUNCE_NUM) {
 375                                         nclaims++;
 376                                         VDBG("announce/%u %s@%s\n",
 377                                                         nclaims, intf, inet_ntoa(ip));
 378                                         arp(ARPOP_REQUEST,
 379                                                         &eth_addr, ip,
 380                                                         &eth_addr, ip);
 381                                         timeout_ms = ANNOUNCE_INTERVAL * 1000;
 382                                 }
 383                                 else {
 384                                         // Switch to monitor state.
 385                                         state = MONITOR;
 386                                         // link is ok to use earlier
 387                                         // FIXME update filters
 388                                         script_av[1] = (char*)"config";
 389                                         run(script_av, &ip);
 390                                         ready = 1;
 391                                         conflicts = 0;
 392                                         timeout_ms = -1; // Never timeout in the monitor state.
 393
 394                                         // NOTE: all other exit paths
 395                                         // should deconfig ...
 396                                         if (QUIT)
 397                                                 return EXIT_SUCCESS;
 398                                 }
 399                                 break;
 400                         case DEFEND:
 401                                 // We won!  No ARP replies, so just go back to monitor.
 402                                 state = MONITOR;
 403                                 timeout_ms = -1;
 404                                 conflicts = 0;
 405                                 break;
 406                         default:
 407                                 // Invalid, should never happen.  Restart the whole protocol.
 408                                 state = PROBE;
 409                                 pick(&ip);
 410                                 timeout_ms = 0;
 411                                 nprobes = 0;
 412                                 nclaims = 0;
 413                                 break;
 414                         } // switch (state)
 415                         break; // case 0 (timeout)
 416
 417                 // packets arriving, or link went down
 418                 case 1:
 419                         // We need to adjust the timeout in case we didn't receive
 420                         // a conflicting packet.
 421                         if (timeout_ms > 0) {
 422                                 unsigned diff = deadline_us - MONOTONIC_US();
 423                                 if ((int)(diff) < 0) {
 424                                         // Current time is greater than the expected timeout time.
 425                                         // Should never happen.
 426                                         VDBG("missed an expected timeout\n");
 427                                         timeout_ms = 0;
 428                                 } else {
 429                                         VDBG("adjusting timeout\n");
 430                                         timeout_ms = (diff / 1000) | 1; /* never 0 */
 431                                 }
 432                         }
 433
 434                         if ((fds[0].revents & POLLIN) == 0) {
 435                                 if (fds[0].revents & POLLERR) {
 436                                         // FIXME: links routinely go down;
 437                                         // this shouldn't necessarily exit.
 438                                         bb_error_msg("iface %s is down", intf);
 439                                         if (ready) {
 440                                                 script_av[1] = (char*)"deconfig";
 441                                                 run(script_av, &ip);
 442                                         }
 443                                         return EXIT_FAILURE;
 444                                 }
 445                                 continue;
 446                         }
 447
 448                         // read ARP packet
 449                         if (safe_read(sock_fd, &p, sizeof(p)) < 0) {
 450                                 why = "recv";
 451                                 goto bad;
 452                         }
 453                         if (p.eth.ether_type != htons(ETHERTYPE_ARP))
 454                                 continue;
 455 #ifdef DEBUG
 456                         {
 457                                 struct ether_addr *sha = (struct ether_addr *) p.arp.arp_sha;
 458                                 struct ether_addr *tha = (struct ether_addr *) p.arp.arp_tha;
 459                                 struct in_addr *spa = (struct in_addr *) p.arp.arp_spa;
 460                                 struct in_addr *tpa = (struct in_addr *) p.arp.arp_tpa;
 461                                 VDBG("%s recv arp type=%d, op=%d,\n",
 462                                         intf, ntohs(p.eth.ether_type),
 463                                         ntohs(p.arp.arp_op));
 464                                 VDBG("\tsource=%s %s\n",
 465                                         ether_ntoa(sha),
 466                                         inet_ntoa(*spa));
 467                                 VDBG("\ttarget=%s %s\n",
 468                                         ether_ntoa(tha),
 469                                         inet_ntoa(*tpa));
 470                         }
 471 #endif
 472                         if (p.arp.arp_op != htons(ARPOP_REQUEST)
 473                          && p.arp.arp_op != htons(ARPOP_REPLY))
 474                                 continue;
 475
 476                         source_ip_conflict = 0;
 477                         target_ip_conflict = 0;
 478
 479                         if (memcmp(p.arp.arp_spa, &ip.s_addr, sizeof(struct in_addr)) == 0
 480                          && memcmp(&p.arp.arp_sha, &eth_addr, ETH_ALEN) != 0
 481                         ) {
 482                                 source_ip_conflict = 1;
 483                         }
 484                         if (p.arp.arp_op == htons(ARPOP_REQUEST)
 485                          && memcmp(p.arp.arp_tpa, &ip.s_addr, sizeof(struct in_addr)) == 0
 486                          && memcmp(&p.arp.arp_tha, &eth_addr, ETH_ALEN) != 0
 487                         ) {
 488                                 target_ip_conflict = 1;
 489                         }
 490
 491                         VDBG("state = %d, source ip conflict = %d, target ip conflict = %d\n",
 492                                 state, source_ip_conflict, target_ip_conflict);
 493                         switch (state) {
 494                         case PROBE:
 495                         case ANNOUNCE:
 496                                 // When probing or announcing, check for source IP conflicts
 497                                 // and other hosts doing ARP probes (target IP conflicts).
 498                                 if (source_ip_conflict || target_ip_conflict) {
 499                                         conflicts++;
 500                                         if (conflicts >= MAX_CONFLICTS) {
 501                                                 VDBG("%s ratelimit\n", intf);
 502                                                 timeout_ms = RATE_LIMIT_INTERVAL * 1000;
 503                                                 state = RATE_LIMIT_PROBE;
 504                                         }
 505
 506                                         // restart the whole protocol
 507                                         pick(&ip);
 508                                         timeout_ms = 0;
 509                                         nprobes = 0;
 510                                         nclaims = 0;
 511                                 }
 512                                 break;
 513                         case MONITOR:
 514                                 // If a conflict, we try to defend with a single ARP probe.
 515                                 if (source_ip_conflict) {
 516                                         VDBG("monitor conflict -- defending\n");
 517                                         state = DEFEND;
 518                                         timeout_ms = DEFEND_INTERVAL * 1000;
 519                                         arp(ARPOP_REQUEST,
 520                                                 &eth_addr, ip,
 521                                                 &eth_addr, ip);
 522                                 }
 523                                 break;
 524                         case DEFEND:
 525                                 // Well, we tried.  Start over (on conflict).
 526                                 if (source_ip_conflict) {
 527                                         state = PROBE;
 528                                         VDBG("defend conflict -- starting over\n");
 529                                         ready = 0;
 530                                         script_av[1] = (char*)"deconfig";
 531                                         run(script_av, &ip);
 532
 533                                         // restart the whole protocol
 534                                         pick(&ip);
 535                                         timeout_ms = 0;
 536                                         nprobes = 0;
 537                                         nclaims = 0;
 538                                 }
 539                                 break;
 540                         default:
 541                                 // Invalid, should never happen.  Restart the whole protocol.
 542                                 VDBG("invalid state -- starting over\n");
 543                                 state = PROBE;
 544                                 pick(&ip);
 545                                 timeout_ms = 0;
 546                                 nprobes = 0;
 547                                 nclaims = 0;
 548                                 break;
 549                         } // switch state
 550                         break; // case 1 (packets arriving)
 551                 } // switch poll
 552         } // while (1)
 553  bad:
 554         bb_perror_msg("%s: %s", intf, why);
 555         return EXIT_FAILURE;
 556 }