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         char *r_opt;
 181         unsigned opts;
 182
 183         /* Ugly trick, but I want these zeroed in one go */
 184         struct {
 185                 const struct in_addr null_ip;
 186                 const struct ether_addr null_addr;
 187                 struct in_addr ip;
 188                 struct ifreq ifr;
 189                 char *script_av[3];
 190                 int timeout_ms; /* must be signed */
 191                 unsigned conflicts;
 192                 unsigned nprobes;
 193                 unsigned nclaims;
 194                 int ready;
 195                 int verbose;
 196         } L;
 197 #define null_ip    (L.null_ip   )
 198 #define null_addr  (L.null_addr )
 199 #define ip         (L.ip        )
 200 #define ifr        (L.ifr       )
 201 #define script_av  (L.script_av )
 202 #define timeout_ms (L.timeout_ms)
 203 #define conflicts  (L.conflicts )
 204 #define nprobes    (L.nprobes   )
 205 #define nclaims    (L.nclaims   )
 206 #define ready      (L.ready     )
 207 #define verbose    (L.verbose   )
 208
 209         memset(&L, 0, sizeof(L));
 210
 211 #define FOREGROUND (opts & 1)
 212 #define QUIT       (opts & 2)
 213         // parse commandline: prog [options] ifname script
 214         // exactly 2 args; -v accumulates and implies -f
 215         opt_complementary = "=2:vv:vf";
 216         opts = getopt32(argv, "fqr:v", &r_opt, &verbose);
 217         if (!FOREGROUND) {
 218                 /* Do it early, before all bb_xx_msg calls */
 219                 openlog(applet_name, 0, LOG_DAEMON);
 220                 logmode |= LOGMODE_SYSLOG;
 221         }
 222         if (opts & 4) { // -r n.n.n.n
 223                 if (inet_aton(r_opt, &ip) == 0
 224                  || (ntohl(ip.s_addr) & IN_CLASSB_NET) != LINKLOCAL_ADDR
 225                 ) {
 226                         bb_error_msg_and_die("invalid link address");
 227                 }
 228         }
 229         // On NOMMU reexec early (or else we will rerun things twice)
 230 #if !BB_MMU
 231         if (!FOREGROUND)
 232                 bb_daemonize_or_rexec(DAEMON_CHDIR_ROOT, argv);
 233 #endif
 234         argc -= optind;
 235         argv += optind;
 236
 237         intf = argv[0];
 238         script_av[0] = argv[1];
 239         setenv("interface", intf, 1);
 240
 241         // initialize the interface (modprobe, ifup, etc)
 242         script_av[1] = (char*)"init";
 243         if (run(script_av, NULL))
 244                 return EXIT_FAILURE;
 245
 246         // initialize saddr
 247         // saddr is: { u16 sa_family; u8 sa_data[14]; }
 248         //memset(&saddr, 0, sizeof(saddr));
 249         //TODO: are we leaving sa_family == 0 (AF_UNSPEC)?!
 250         safe_strncpy(saddr.sa_data, intf, sizeof(saddr.sa_data));
 251
 252         // open an ARP socket
 253         xmove_fd(xsocket(AF_PACKET, SOCK_PACKET, htons(ETH_P_ARP)), sock_fd);
 254         // bind to the interface's ARP socket
 255         xbind(sock_fd, &saddr, sizeof(saddr));
 256
 257         // get the interface's ethernet address
 258         //memset(&ifr, 0, sizeof(ifr));
 259         strncpy(ifr.ifr_name, intf, sizeof(ifr.ifr_name));
 260         xioctl(sock_fd, SIOCGIFHWADDR, &ifr);
 261         memcpy(&eth_addr, &ifr.ifr_hwaddr.sa_data, ETH_ALEN);
 262
 263         // start with some stable ip address, either a function of
 264         // the hardware address or else the last address we used.
 265         // we are taking low-order four bytes, as top-order ones
 266         // aren't random enough.
 267         // NOTE: the sequence of addresses we try changes only
 268         // depending on when we detect conflicts.
 269         {
 270                 uint32_t t;
 271                 memcpy(&t, (char*)&eth_addr + 2, 4);
 272                 srand(t);
 273         }
 274         if (ip.s_addr == 0)
 275                 pick(&ip);
 276
 277         // FIXME cases to handle:
 278         //  - zcip already running!
 279         //  - link already has local address... just defend/update
 280
 281         // daemonize now; don't delay system startup
 282         if (!FOREGROUND) {
 283 #if BB_MMU
 284                 bb_daemonize(DAEMON_CHDIR_ROOT);
 285 #endif
 286                 bb_info_msg("start, interface %s", intf);
 287         }
 288
 289         // run the dynamic address negotiation protocol,
 290         // restarting after address conflicts:
 291         //  - start with some address we want to try
 292         //  - short random delay
 293         //  - arp probes to see if another host else uses it
 294         //  - arp announcements that we're claiming it
 295         //  - use it
 296         //  - defend it, within limits
 297         while (1) {
 298                 struct pollfd fds[1];
 299                 unsigned deadline_us;
 300                 struct arp_packet p;
 301                 int source_ip_conflict;
 302                 int target_ip_conflict;
 303
 304                 fds[0].fd = sock_fd;
 305                 fds[0].events = POLLIN;
 306                 fds[0].revents = 0;
 307
 308                 // poll, being ready to adjust current timeout
 309                 if (!timeout_ms) {
 310                         timeout_ms = random_delay_ms(PROBE_WAIT);
 311                         // FIXME setsockopt(sock_fd, SO_ATTACH_FILTER, ...) to
 312                         // make the kernel filter out all packets except
 313                         // ones we'd care about.
 314                 }
 315                 // set deadline_us to the point in time when we timeout
 316                 deadline_us = MONOTONIC_US() + timeout_ms * 1000;
 317
 318                 VDBG("...wait %d %s nprobes=%u, nclaims=%u\n",
 319                                 timeout_ms, intf, nprobes, nclaims);
 320
 321                 switch (safe_poll(fds, 1, timeout_ms)) {
 322
 323                 default:
 324                         /*bb_perror_msg("poll"); - done in safe_poll */
 325                         return EXIT_FAILURE;
 326
 327                 // timeout
 328                 case 0:
 329                         VDBG("state = %d\n", state);
 330                         switch (state) {
 331                         case PROBE:
 332                                 // timeouts in the PROBE state mean no conflicting ARP packets
 333                                 // have been received, so we can progress through the states
 334                                 if (nprobes < PROBE_NUM) {
 335                                         nprobes++;
 336                                         VDBG("probe/%u %s@%s\n",
 337                                                         nprobes, intf, inet_ntoa(ip));
 338                                         arp(ARPOP_REQUEST,
 339                                                         &eth_addr, null_ip,
 340                                                         &null_addr, ip);
 341                                         timeout_ms = PROBE_MIN * 1000;
 342                                         timeout_ms += random_delay_ms(PROBE_MAX - PROBE_MIN);
 343                                 }
 344                                 else {
 345                                         // Switch to announce state.
 346                                         state = ANNOUNCE;
 347                                         nclaims = 0;
 348                                         VDBG("announce/%u %s@%s\n",
 349                                                         nclaims, intf, inet_ntoa(ip));
 350                                         arp(ARPOP_REQUEST,
 351                                                         &eth_addr, ip,
 352                                                         &eth_addr, ip);
 353                                         timeout_ms = ANNOUNCE_INTERVAL * 1000;
 354                                 }
 355                                 break;
 356                         case RATE_LIMIT_PROBE:
 357                                 // timeouts in the RATE_LIMIT_PROBE state mean no conflicting ARP packets
 358                                 // have been received, so we can move immediately to the announce state
 359                                 state = ANNOUNCE;
 360                                 nclaims = 0;
 361                                 VDBG("announce/%u %s@%s\n",
 362                                                 nclaims, intf, inet_ntoa(ip));
 363                                 arp(ARPOP_REQUEST,
 364                                                 &eth_addr, ip,
 365                                                 &eth_addr, ip);
 366                                 timeout_ms = ANNOUNCE_INTERVAL * 1000;
 367                                 break;
 368                         case ANNOUNCE:
 369                                 // timeouts in the ANNOUNCE state mean no conflicting ARP packets
 370                                 // have been received, so we can progress through the states
 371                                 if (nclaims < ANNOUNCE_NUM) {
 372                                         nclaims++;
 373                                         VDBG("announce/%u %s@%s\n",
 374                                                         nclaims, intf, inet_ntoa(ip));
 375                                         arp(ARPOP_REQUEST,
 376                                                         &eth_addr, ip,
 377                                                         &eth_addr, ip);
 378                                         timeout_ms = ANNOUNCE_INTERVAL * 1000;
 379                                 }
 380                                 else {
 381                                         // Switch to monitor state.
 382                                         state = MONITOR;
 383                                         // link is ok to use earlier
 384                                         // FIXME update filters
 385                                         script_av[1] = (char*)"config";
 386                                         run(script_av, &ip);
 387                                         ready = 1;
 388                                         conflicts = 0;
 389                                         timeout_ms = -1; // Never timeout in the monitor state.
 390
 391                                         // NOTE: all other exit paths
 392                                         // should deconfig ...
 393                                         if (QUIT)
 394                                                 return EXIT_SUCCESS;
 395                                 }
 396                                 break;
 397                         case DEFEND:
 398                                 // We won!  No ARP replies, so just go back to monitor.
 399                                 state = MONITOR;
 400                                 timeout_ms = -1;
 401                                 conflicts = 0;
 402                                 break;
 403                         default:
 404                                 // Invalid, should never happen.  Restart the whole protocol.
 405                                 state = PROBE;
 406                                 pick(&ip);
 407                                 timeout_ms = 0;
 408                                 nprobes = 0;
 409                                 nclaims = 0;
 410                                 break;
 411                         } // switch (state)
 412                         break; // case 0 (timeout)
 413
 414                 // packets arriving, or link went down
 415                 case 1:
 416                         // We need to adjust the timeout in case we didn't receive
 417                         // a conflicting packet.
 418                         if (timeout_ms > 0) {
 419                                 unsigned diff = deadline_us - MONOTONIC_US();
 420                                 if ((int)(diff) < 0) {
 421                                         // Current time is greater than the expected timeout time.
 422                                         // Should never happen.
 423                                         VDBG("missed an expected timeout\n");
 424                                         timeout_ms = 0;
 425                                 } else {
 426                                         VDBG("adjusting timeout\n");
 427                                         timeout_ms = (diff / 1000) | 1; /* never 0 */
 428                                 }
 429                         }
 430
 431                         if ((fds[0].revents & POLLIN) == 0) {
 432                                 if (fds[0].revents & POLLERR) {
 433                                         // FIXME: links routinely go down;
 434                                         // this shouldn't necessarily exit.
 435                                         bb_error_msg("iface %s is down", intf);
 436                                         if (ready) {
 437                                                 script_av[1] = (char*)"deconfig";
 438                                                 run(script_av, &ip);
 439                                         }
 440                                         return EXIT_FAILURE;
 441                                 }
 442                                 continue;
 443                         }
 444
 445                         // read ARP packet
 446                         if (safe_read(sock_fd, &p, sizeof(p)) < 0) {
 447                                 bb_perror_msg_and_die(bb_msg_read_error);
 448                         }
 449                         if (p.eth.ether_type != htons(ETHERTYPE_ARP))
 450                                 continue;
 451 #ifdef DEBUG
 452                         {
 453                                 struct ether_addr *sha = (struct ether_addr *) p.arp.arp_sha;
 454                                 struct ether_addr *tha = (struct ether_addr *) p.arp.arp_tha;
 455                                 struct in_addr *spa = (struct in_addr *) p.arp.arp_spa;
 456                                 struct in_addr *tpa = (struct in_addr *) p.arp.arp_tpa;
 457                                 VDBG("%s recv arp type=%d, op=%d,\n",
 458                                         intf, ntohs(p.eth.ether_type),
 459                                         ntohs(p.arp.arp_op));
 460                                 VDBG("\tsource=%s %s\n",
 461                                         ether_ntoa(sha),
 462                                         inet_ntoa(*spa));
 463                                 VDBG("\ttarget=%s %s\n",
 464                                         ether_ntoa(tha),
 465                                         inet_ntoa(*tpa));
 466                         }
 467 #endif
 468                         if (p.arp.arp_op != htons(ARPOP_REQUEST)
 469                          && p.arp.arp_op != htons(ARPOP_REPLY))
 470                                 continue;
 471
 472                         source_ip_conflict = 0;
 473                         target_ip_conflict = 0;
 474
 475                         if (memcmp(p.arp.arp_spa, &ip.s_addr, sizeof(struct in_addr)) == 0
 476                          && memcmp(&p.arp.arp_sha, &eth_addr, ETH_ALEN) != 0
 477                         ) {
 478                                 source_ip_conflict = 1;
 479                         }
 480                         if (p.arp.arp_op == htons(ARPOP_REQUEST)
 481                          && memcmp(p.arp.arp_tpa, &ip.s_addr, sizeof(struct in_addr)) == 0
 482                          && memcmp(&p.arp.arp_tha, &eth_addr, ETH_ALEN) != 0
 483                         ) {
 484                                 target_ip_conflict = 1;
 485                         }
 486
 487                         VDBG("state = %d, source ip conflict = %d, target ip conflict = %d\n",
 488                                 state, source_ip_conflict, target_ip_conflict);
 489                         switch (state) {
 490                         case PROBE:
 491                         case ANNOUNCE:
 492                                 // When probing or announcing, check for source IP conflicts
 493                                 // and other hosts doing ARP probes (target IP conflicts).
 494                                 if (source_ip_conflict || target_ip_conflict) {
 495                                         conflicts++;
 496                                         if (conflicts >= MAX_CONFLICTS) {
 497                                                 VDBG("%s ratelimit\n", intf);
 498                                                 timeout_ms = RATE_LIMIT_INTERVAL * 1000;
 499                                                 state = RATE_LIMIT_PROBE;
 500                                         }
 501
 502                                         // restart the whole protocol
 503                                         pick(&ip);
 504                                         timeout_ms = 0;
 505                                         nprobes = 0;
 506                                         nclaims = 0;
 507                                 }
 508                                 break;
 509                         case MONITOR:
 510                                 // If a conflict, we try to defend with a single ARP probe.
 511                                 if (source_ip_conflict) {
 512                                         VDBG("monitor conflict -- defending\n");
 513                                         state = DEFEND;
 514                                         timeout_ms = DEFEND_INTERVAL * 1000;
 515                                         arp(ARPOP_REQUEST,
 516                                                 &eth_addr, ip,
 517                                                 &eth_addr, ip);
 518                                 }
 519                                 break;
 520                         case DEFEND:
 521                                 // Well, we tried.  Start over (on conflict).
 522                                 if (source_ip_conflict) {
 523                                         state = PROBE;
 524                                         VDBG("defend conflict -- starting over\n");
 525                                         ready = 0;
 526                                         script_av[1] = (char*)"deconfig";
 527                                         run(script_av, &ip);
 528
 529                                         // restart the whole protocol
 530                                         pick(&ip);
 531                                         timeout_ms = 0;
 532                                         nprobes = 0;
 533                                         nclaims = 0;
 534                                 }
 535                                 break;
 536                         default:
 537                                 // Invalid, should never happen.  Restart the whole protocol.
 538                                 VDBG("invalid state -- starting over\n");
 539                                 state = PROBE;
 540                                 pick(&ip);
 541                                 timeout_ms = 0;
 542                                 nprobes = 0;
 543                                 nclaims = 0;
 544                                 break;
 545                         } // switch state
 546                         break; // case 1 (packets arriving)
 547                 } // switch poll
 548         } // while (1)
 549 }