networking/zcip.c

   1 /* vi: set sw=4 ts=4: */
   2 /*
   3  * RFC3927 ZeroConf IPv4 Link-Local addressing
   4  * (see <http://www.zeroconf.org/>)
   5  *
   6  * Copyright (C) 2003 by Arthur van Hoff (avh@strangeberry.com)
   7  * Copyright (C) 2004 by David Brownell
   8  *
   9  * Licensed under the GPL v2 or later, see the file LICENSE in this tarball.
  10  */
  11
  12 /*
  13  * ZCIP just manages the 169.254.*.* addresses.  That network is not
  14  * routed at the IP level, though various proxies or bridges can
  15  * certainly be used.  Its naming is built over multicast DNS.
  16  */
  17
  18 //#define DEBUG
  19
  20 // TODO:
  21 // - more real-world usage/testing, especially daemon mode
  22 // - kernel packet filters to reduce scheduling noise
  23 // - avoid silent script failures, especially under load...
  24 // - link status monitoring (restart on link-up; stop on link-down)
  25
  26 #include "busybox.h"
  27 #include <errno.h>
  28 #include <string.h>
  29 #include <syslog.h>
  30 #include <poll.h>
  31 #include <time.h>
  32
  33 #include <sys/wait.h>
  34
  35 #include <netinet/ether.h>
  36 #include <net/ethernet.h>
  37 #include <net/if.h>
  38 #include <net/if_arp.h>
  39
  40 #include <linux/if_packet.h>
  41 #include <linux/sockios.h>
  42
  43
  44 struct arp_packet {
  45         struct ether_header hdr;
  46         struct ether_arp arp;
  47 } ATTRIBUTE_PACKED;
  48
  49 enum {
  50 /* 169.254.0.0 */
  51         LINKLOCAL_ADDR = 0xa9fe0000,
  52
  53 /* protocol timeout parameters, specified in seconds */
  54         PROBE_WAIT = 1,
  55         PROBE_MIN = 1,
  56         PROBE_MAX = 2,
  57         PROBE_NUM = 3,
  58         MAX_CONFLICTS = 10,
  59         RATE_LIMIT_INTERVAL = 60,
  60         ANNOUNCE_WAIT = 2,
  61         ANNOUNCE_NUM = 2,
  62         ANNOUNCE_INTERVAL = 2,
  63         DEFEND_INTERVAL = 10
  64 };
  65
  66 /* States during the configuration process. */
  67 enum {
  68         PROBE = 0,
  69         RATE_LIMIT_PROBE,
  70         ANNOUNCE,
  71         MONITOR,
  72         DEFEND
  73 };
  74
  75 #define VDBG(fmt,args...) \
  76         do { } while (0)
  77
  78 static unsigned long opts;
  79 #define FOREGROUND (opts & 1)
  80 #define QUIT (opts & 2)
  81
  82 /**
  83  * Pick a random link local IP address on 169.254/16, except that
  84  * the first and last 256 addresses are reserved.
  85  */
  86 static void pick(struct in_addr *ip)
  87 {
  88         unsigned tmp;
  89
  90         /* use cheaper math than lrand48() mod N */
  91         do {
  92                 tmp = (lrand48() >> 16) & IN_CLASSB_HOST;
  93         } while (tmp > (IN_CLASSB_HOST - 0x0200));
  94         ip->s_addr = htonl((LINKLOCAL_ADDR + 0x0100) + tmp);
  95 }
  96
  97 /* TODO: we need a flag to direct bb_[p]error_msg output to stderr. */
  98
  99 /**
 100  * Broadcast an ARP packet.
 101  */
 102 static void arp(int fd, struct sockaddr *saddr, int op,
 103         const struct ether_addr *source_addr, struct in_addr source_ip,
 104         const struct ether_addr *target_addr, struct in_addr target_ip)
 105 {
 106         struct arp_packet p;
 107         memset(&p, 0, sizeof(p));
 108
 109         // ether header
 110         p.hdr.ether_type = htons(ETHERTYPE_ARP);
 111         memcpy(p.hdr.ether_shost, source_addr, ETH_ALEN);
 112         memset(p.hdr.ether_dhost, 0xff, ETH_ALEN);
 113
 114         // arp request
 115         p.arp.arp_hrd = htons(ARPHRD_ETHER);
 116         p.arp.arp_pro = htons(ETHERTYPE_IP);
 117         p.arp.arp_hln = ETH_ALEN;
 118         p.arp.arp_pln = 4;
 119         p.arp.arp_op = htons(op);
 120         memcpy(&p.arp.arp_sha, source_addr, ETH_ALEN);
 121         memcpy(&p.arp.arp_spa, &source_ip, sizeof (p.arp.arp_spa));
 122         memcpy(&p.arp.arp_tha, target_addr, ETH_ALEN);
 123         memcpy(&p.arp.arp_tpa, &target_ip, sizeof (p.arp.arp_tpa));
 124
 125         // send it
 126         if (sendto(fd, &p, sizeof (p), 0, saddr, sizeof (*saddr)) < 0) {
 127                 bb_perror_msg("sendto");
 128                 //return -errno;
 129         }
 130         // Currently all callers ignore errors, that's why returns are
 131         // commented out...
 132         //return 0;
 133 }
 134
 135 /**
 136  * Run a script.
 137  */
 138 static int run(char *script, char *arg, char *intf, struct in_addr *ip)
 139 {
 140         int pid, status;
 141         char *why;
 142
 143         if(1) { //always true: if (script != NULL)
 144                 VDBG("%s run %s %s\n", intf, script, arg);
 145                 if (ip != NULL) {
 146                         char *addr = inet_ntoa(*ip);
 147                         setenv("ip", addr, 1);
 148                         bb_info_msg("%s %s %s", arg, intf, addr);
 149                 }
 150
 151                 pid = vfork();
 152                 if (pid < 0) {                  // error
 153                         why = "vfork";
 154                         goto bad;
 155                 } else if (pid == 0) {          // child
 156                         execl(script, script, arg, NULL);
 157                         bb_perror_msg("execl");
 158                         _exit(EXIT_FAILURE);
 159                 }
 160
 161                 if (waitpid(pid, &status, 0) <= 0) {
 162                         why = "waitpid";
 163                         goto bad;
 164                 }
 165                 if (WEXITSTATUS(status) != 0) {
 166                         bb_error_msg("script %s failed, exit=%d",
 167                                 script, WEXITSTATUS(status));
 168                         return -errno;
 169                 }
 170         }
 171         return 0;
 172 bad:
 173         status = -errno;
 174         bb_perror_msg("%s %s, %s", arg, intf, why);
 175         return status;
 176 }
 177
 178
 179 /**
 180  * Return milliseconds of random delay, up to "secs" seconds.
 181  */
 182 static unsigned ATTRIBUTE_ALWAYS_INLINE ms_rdelay(unsigned secs)
 183 {
 184         return lrand48() % (secs * 1000);
 185 }
 186
 187 /**
 188  * main program
 189  */
 190
 191 /* Used to be auto variables on main() stack, but
 192  * most of them were zero-inited. Moving them to bss
 193  * is more space-efficient.
 194  */
 195 static  const struct in_addr null_ip; // = { 0 };
 196 static  const struct ether_addr null_addr; // = { {0, 0, 0, 0, 0, 0} };
 197
 198 static  struct sockaddr saddr; // memset(0);
 199 static  struct in_addr ip; // = { 0 };
 200 static  struct ifreq ifr; //memset(0);
 201
 202 static  char *intf; // = NULL;
 203 static  char *script; // = NULL;
 204 static  suseconds_t timeout; // = 0;    // milliseconds
 205 static  unsigned conflicts; // = 0;
 206 static  unsigned nprobes; // = 0;
 207 static  unsigned nclaims; // = 0;
 208 static  int ready; // = 0;
 209 static  int verbose; // = 0;
 210 static  int state = PROBE;
 211
 212 int zcip_main(int argc, char *argv[])
 213 {
 214         struct ether_addr eth_addr;
 215         char *why;
 216         int fd;
 217
 218         // parse commandline: prog [options] ifname script
 219         char *r_opt;
 220         bb_opt_complementally = "vv:vf"; // -v accumulates and implies -f
 221         opts = bb_getopt_ulflags(argc, argv, "fqr:v", &r_opt, &verbose);
 222         if (!FOREGROUND) {
 223                 /* Do it early, before all bb_xx_msg calls */
 224                 logmode = LOGMODE_SYSLOG;
 225                 openlog(bb_applet_name, 0, LOG_DAEMON);
 226         }
 227         if (opts & 4) { // -r n.n.n.n
 228                 if (inet_aton(r_opt, &ip) == 0
 229                 || (ntohl(ip.s_addr) & IN_CLASSB_NET) != LINKLOCAL_ADDR) {
 230                         bb_error_msg_and_die("invalid link address");
 231                 }
 232         }
 233         argc -= optind;
 234         argv += optind;
 235         if (argc != 2)
 236                 bb_show_usage();
 237         intf = argv[0];
 238         script = argv[1];
 239         setenv("interface", intf, 1);
 240
 241         // initialize the interface (modprobe, ifup, etc)
 242         if (run(script, "init", intf, NULL) < 0)
 243                 return EXIT_FAILURE;
 244
 245         // initialize saddr
 246         //memset(&saddr, 0, sizeof (saddr));
 247         safe_strncpy(saddr.sa_data, intf, sizeof (saddr.sa_data));
 248
 249         // open an ARP socket
 250         fd = xsocket(PF_PACKET, SOCK_PACKET, htons(ETH_P_ARP));
 251         // bind to the interface's ARP socket
 252         xbind(fd, &saddr, sizeof (saddr));
 253
 254         // get the interface's ethernet address
 255         //memset(&ifr, 0, sizeof (ifr));
 256         strncpy(ifr.ifr_name, intf, sizeof (ifr.ifr_name));
 257         if (ioctl(fd, SIOCGIFHWADDR, &ifr) < 0) {
 258                 bb_perror_msg_and_die("get ethernet address");
 259         }
 260         memcpy(&eth_addr, &ifr.ifr_hwaddr.sa_data, ETH_ALEN);
 261
 262         // start with some stable ip address, either a function of
 263         // the hardware address or else the last address we used.
 264         // NOTE: the sequence of addresses we try changes only
 265         // depending on when we detect conflicts.
 266         // (SVID 3 bogon: who says that "short" is always 16 bits?)
 267         seed48( (unsigned short*)&ifr.ifr_hwaddr.sa_data );
 268         if (ip.s_addr == 0)
 269                 pick(&ip);
 270
 271         // FIXME cases to handle:
 272         //  - zcip already running!
 273         //  - link already has local address... just defend/update
 274
 275         // daemonize now; don't delay system startup
 276         if (!FOREGROUND) {
 277                 setsid();
 278                 xdaemon(0, 0);
 279                 bb_info_msg("start, interface %s", intf);
 280         }
 281
 282         // run the dynamic address negotiation protocol,
 283         // restarting after address conflicts:
 284         //  - start with some address we want to try
 285         //  - short random delay
 286         //  - arp probes to see if another host else uses it
 287         //  - arp announcements that we're claiming it
 288         //  - use it
 289         //  - defend it, within limits
 290         while (1) {
 291                 struct pollfd fds[1];
 292                 struct timeval tv1;
 293                 struct arp_packet p;
 294
 295                 int source_ip_conflict = 0;
 296                 int target_ip_conflict = 0;
 297
 298                 fds[0].fd = fd;
 299                 fds[0].events = POLLIN;
 300                 fds[0].revents = 0;
 301
 302                 // poll, being ready to adjust current timeout
 303                 if (!timeout) {
 304                         timeout = ms_rdelay(PROBE_WAIT);
 305                         // FIXME setsockopt(fd, SO_ATTACH_FILTER, ...) to
 306                         // make the kernel filter out all packets except
 307                         // ones we'd care about.
 308                 }
 309                 // set tv1 to the point in time when we timeout
 310                 gettimeofday(&tv1, NULL);
 311                 tv1.tv_usec += (timeout % 1000) * 1000;
 312                 while (tv1.tv_usec > 1000000) {
 313                         tv1.tv_usec -= 1000000;
 314                         tv1.tv_sec++;
 315                 }
 316                 tv1.tv_sec += timeout / 1000;
 317
 318                 VDBG("...wait %ld %s nprobes=%d, nclaims=%d\n",
 319                                 timeout, intf, nprobes, nclaims);
 320                 switch (poll(fds, 1, timeout)) {
 321
 322                 // timeout
 323                 case 0:
 324                         VDBG("state = %d\n", state);
 325                         switch (state) {
 326                         case PROBE:
 327                                 // timeouts in the PROBE state mean no conflicting ARP packets
 328                                 // have been received, so we can progress through the states
 329                                 if (nprobes < PROBE_NUM) {
 330                                         nprobes++;
 331                                         VDBG("probe/%d %s@%s\n",
 332                                                         nprobes, intf, inet_ntoa(ip));
 333                                         arp(fd, &saddr, ARPOP_REQUEST,
 334                                                         &eth_addr, null_ip,
 335                                                         &null_addr, ip);
 336                                         timeout = PROBE_MIN * 1000;
 337                                         timeout += ms_rdelay(PROBE_MAX
 338                                                         - PROBE_MIN);
 339                                 }
 340                                 else {
 341                                         // Switch to announce state.
 342                                         state = ANNOUNCE;
 343                                         nclaims = 0;
 344                                         VDBG("announce/%d %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 = ANNOUNCE_INTERVAL * 1000;
 350                                 }
 351                                 break;
 352                         case RATE_LIMIT_PROBE:
 353                                 // timeouts in the RATE_LIMIT_PROBE state mean no conflicting ARP packets
 354                                 // have been received, so we can move immediately to the announce state
 355                                 state = ANNOUNCE;
 356                                 nclaims = 0;
 357                                 VDBG("announce/%d %s@%s\n",
 358                                                 nclaims, intf, inet_ntoa(ip));
 359                                 arp(fd, &saddr, ARPOP_REQUEST,
 360                                                 &eth_addr, ip,
 361                                                 &eth_addr, ip);
 362                                 timeout = ANNOUNCE_INTERVAL * 1000;
 363                                 break;
 364                         case ANNOUNCE:
 365                                 // timeouts in the ANNOUNCE state mean no conflicting ARP packets
 366                                 // have been received, so we can progress through the states
 367                                 if (nclaims < ANNOUNCE_NUM) {
 368                                         nclaims++;
 369                                         VDBG("announce/%d %s@%s\n",
 370                                                         nclaims, intf, inet_ntoa(ip));
 371                                         arp(fd, &saddr, ARPOP_REQUEST,
 372                                                         &eth_addr, ip,
 373                                                         &eth_addr, ip);
 374                                         timeout = ANNOUNCE_INTERVAL * 1000;
 375                                 }
 376                                 else {
 377                                         // Switch to monitor state.
 378                                         state = MONITOR;
 379                                         // link is ok to use earlier
 380                                         // FIXME update filters
 381                                         run(script, "config", intf, &ip);
 382                                         ready = 1;
 383                                         conflicts = 0;
 384                                         timeout = -1; // Never timeout in the monitor state.
 385
 386                                         // NOTE: all other exit paths
 387                                         // should deconfig ...
 388                                         if (QUIT)
 389                                                 return EXIT_SUCCESS;
 390                                 }
 391                                 break;
 392                         case DEFEND:
 393                                 // We won!  No ARP replies, so just go back to monitor.
 394                                 state = MONITOR;
 395                                 timeout = -1;
 396                                 conflicts = 0;
 397                                 break;
 398                         default:
 399                                 // Invalid, should never happen.  Restart the whole protocol.
 400                                 state = PROBE;
 401                                 pick(&ip);
 402                                 timeout = 0;
 403                                 nprobes = 0;
 404                                 nclaims = 0;
 405                                 break;
 406                         } // switch (state)
 407                         break; // case 0 (timeout)
 408                 // packets arriving
 409                 case 1:
 410                         // We need to adjust the timeout in case we didn't receive
 411                         // a conflicting packet.
 412                         if (timeout > 0) {
 413                                 struct timeval tv2;
 414
 415                                 gettimeofday(&tv2, NULL);
 416                                 if (timercmp(&tv1, &tv2, <)) {
 417                                         // Current time is greater than the expected timeout time.
 418                                         // Should never happen.
 419                                         VDBG("missed an expected timeout\n");
 420                                         timeout = 0;
 421                                 } else {
 422                                         VDBG("adjusting timeout\n");
 423                                         timersub(&tv1, &tv2, &tv1);
 424                                         timeout = 1000 * tv1.tv_sec
 425                                                         + tv1.tv_usec / 1000;
 426                                 }
 427                         }
 428
 429                         if ((fds[0].revents & POLLIN) == 0) {
 430                                 if (fds[0].revents & POLLERR) {
 431                                         // FIXME: links routinely go down;
 432                                         // this shouldn't necessarily exit.
 433                                         bb_error_msg("%s: poll error", intf);
 434                                         if (ready) {
 435                                                 run(script, "deconfig",
 436                                                                 intf, &ip);
 437                                         }
 438                                         return EXIT_FAILURE;
 439                                 }
 440                                 continue;
 441                         }
 442
 443                         // read ARP packet
 444                         if (recv(fd, &p, sizeof (p), 0) < 0) {
 445                                 why = "recv";
 446                                 goto bad;
 447                         }
 448                         if (p.hdr.ether_type != htons(ETHERTYPE_ARP))
 449                                 continue;
 450
 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.hdr.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                         if (memcmp(p.arp.arp_spa, &ip.s_addr, sizeof(struct in_addr)) == 0 &&
 473                                 memcmp(&eth_addr, &p.arp.arp_sha, ETH_ALEN) != 0) {
 474                                 source_ip_conflict = 1;
 475                         }
 476                         if (memcmp(p.arp.arp_tpa, &ip.s_addr, sizeof(struct in_addr)) == 0 &&
 477                                 p.arp.arp_op == htons(ARPOP_REQUEST) &&
 478                                 memcmp(&eth_addr, &p.arp.arp_tha, ETH_ALEN) != 0) {
 479                                 target_ip_conflict = 1;
 480                         }
 481
 482                         VDBG("state = %d, source ip conflict = %d, target ip conflict = %d\n",
 483                                 state, source_ip_conflict, target_ip_conflict);
 484                         switch (state) {
 485                         case PROBE:
 486                         case ANNOUNCE:
 487                                 // When probing or announcing, check for source IP conflicts
 488                                 // and other hosts doing ARP probes (target IP conflicts).
 489                                 if (source_ip_conflict || target_ip_conflict) {
 490                                         conflicts++;
 491                                         if (conflicts >= MAX_CONFLICTS) {
 492                                                 VDBG("%s ratelimit\n", intf);
 493                                                 timeout = RATE_LIMIT_INTERVAL * 1000;
 494                                                 state = RATE_LIMIT_PROBE;
 495                                         }
 496
 497                                         // restart the whole protocol
 498                                         pick(&ip);
 499                                         timeout = 0;
 500                                         nprobes = 0;
 501                                         nclaims = 0;
 502                                 }
 503                                 break;
 504                         case MONITOR:
 505                                 // If a conflict, we try to defend with a single ARP probe.
 506                                 if (source_ip_conflict) {
 507                                         VDBG("monitor conflict -- defending\n");
 508                                         state = DEFEND;
 509                                         timeout = DEFEND_INTERVAL * 1000;
 510                                         arp(fd, &saddr,
 511                                                         ARPOP_REQUEST,
 512                                                         &eth_addr, ip,
 513                                                         &eth_addr, ip);
 514                                 }
 515                                 break;
 516                         case DEFEND:
 517                                 // Well, we tried.  Start over (on conflict).
 518                                 if (source_ip_conflict) {
 519                                         state = PROBE;
 520                                         VDBG("defend conflict -- starting over\n");
 521                                         ready = 0;
 522                                         run(script, "deconfig", intf, &ip);
 523
 524                                         // restart the whole protocol
 525                                         pick(&ip);
 526                                         timeout = 0;
 527                                         nprobes = 0;
 528                                         nclaims = 0;
 529                                 }
 530                                 break;
 531                         default:
 532                                 // Invalid, should never happen.  Restart the whole protocol.
 533                                 VDBG("invalid state -- starting over\n");
 534                                 state = PROBE;
 535                                 pick(&ip);
 536                                 timeout = 0;
 537                                 nprobes = 0;
 538                                 nclaims = 0;
 539                                 break;
 540                         } // switch state
 541
 542                         break; // case 1 (packets arriving)
 543                 default:
 544                         why = "poll";
 545                         goto bad;
 546                 } // switch poll
 547         }
 548 bad:
 549         bb_perror_msg("%s, %s", intf, why);
 550         return EXIT_FAILURE;
 551 }