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 GPLv2 or later, see file LICENSE in this source tree.
  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 //usage:#define zcip_trivial_usage
  27 //usage:       "[OPTIONS] IFACE SCRIPT"
  28 //usage:#define zcip_full_usage "\n\n"
  29 //usage:       "Manage a ZeroConf IPv4 link-local address\n"
  30 //usage:     "\n        -f              Run in foreground"
  31 //usage:     "\n        -q              Quit after obtaining address"
  32 //usage:     "\n        -r 169.254.x.x  Request this address first"
  33 //usage:     "\n        -l x.x.0.0      Use this range instead of 169.254"
  34 //usage:     "\n        -v              Verbose"
  35 //usage:     "\n"
  36 //usage:     "\n$LOGGING=none           Suppress logging"
  37 //usage:     "\n$LOGGING=syslog         Log to syslog"
  38 //usage:     "\n"
  39 //usage:     "\nWith no -q, runs continuously monitoring for ARP conflicts,"
  40 //usage:     "\nexits only on I/O errors (link down etc)"
  41
  42 #include "libbb.h"
  43 #include <netinet/ether.h>
  44 #include <net/if.h>
  45 #include <net/if_arp.h>
  46 #include <linux/sockios.h>
  47
  48 #include <syslog.h>
  49
  50 /* We don't need more than 32 bits of the counter */
  51 #define MONOTONIC_US() ((unsigned)monotonic_us())
  52
  53 struct arp_packet {
  54         struct ether_header eth;
  55         struct ether_arp arp;
  56 } PACKED;
  57
  58 enum {
  59         /* 169.254.0.0 */
  60         LINKLOCAL_ADDR = 0xa9fe0000,
  61
  62         /* 0-1 seconds before sending 1st probe */
  63         PROBE_WAIT = 1,
  64         /* 1-2 seconds between probes */
  65         PROBE_MIN = 1,
  66         PROBE_MAX = 2,
  67         PROBE_NUM = 3,          /* total probes to send */
  68         ANNOUNCE_INTERVAL = 2,  /* 2 seconds between announces */
  69         ANNOUNCE_NUM = 3,       /* announces to send */
  70         /* if probe/announce sees a conflict, multiply RANDOM(NUM_CONFLICT) by... */
  71         CONFLICT_MULTIPLIER = 2,
  72         /* if we monitor and see a conflict, how long is defend state? */
  73         DEFEND_INTERVAL = 10
  74 };
  75
  76 /* States during the configuration process. */
  77 enum {
  78         PROBE = 0,
  79         ANNOUNCE,
  80         MONITOR,
  81         DEFEND
  82 };
  83
  84 #define VDBG(...) do { } while (0)
  85
  86
  87 enum {
  88         sock_fd = 3
  89 };
  90
  91 struct globals {
  92         struct sockaddr iface_sockaddr;
  93         struct ether_addr eth_addr;
  94         uint32_t localnet_ip;
  95 } FIX_ALIASING;
  96 #define G (*(struct globals*)&bb_common_bufsiz1)
  97 #define INIT_G() do { } while (0)
  98
  99
 100 /**
 101  * Pick a random link local IP address on 169.254/16, except that
 102  * the first and last 256 addresses are reserved.
 103  */
 104 static uint32_t pick_nip(void)
 105 {
 106         unsigned tmp;
 107
 108         do {
 109                 tmp = rand() & IN_CLASSB_HOST;
 110         } while (tmp > (IN_CLASSB_HOST - 0x0200));
 111         return htonl((G.localnet_ip + 0x0100) + tmp);
 112 }
 113
 114 static const char *nip_to_a(uint32_t nip)
 115 {
 116         struct in_addr in;
 117         in.s_addr = nip;
 118         return inet_ntoa(in);
 119 }
 120
 121 /**
 122  * Broadcast an ARP packet.
 123  */
 124 static void arp(
 125         /* int op, - always ARPOP_REQUEST */
 126         /* const struct ether_addr *source_eth, - always &G.eth_addr */
 127                                         uint32_t source_nip,
 128         const struct ether_addr *target_eth, uint32_t target_nip)
 129 {
 130         enum { op = ARPOP_REQUEST };
 131 #define source_eth (&G.eth_addr)
 132
 133         struct arp_packet p;
 134         memset(&p, 0, sizeof(p));
 135
 136         // ether header
 137         p.eth.ether_type = htons(ETHERTYPE_ARP);
 138         memcpy(p.eth.ether_shost, source_eth, ETH_ALEN);
 139         memset(p.eth.ether_dhost, 0xff, ETH_ALEN);
 140
 141         // arp request
 142         p.arp.arp_hrd = htons(ARPHRD_ETHER);
 143         p.arp.arp_pro = htons(ETHERTYPE_IP);
 144         p.arp.arp_hln = ETH_ALEN;
 145         p.arp.arp_pln = 4;
 146         p.arp.arp_op = htons(op);
 147         memcpy(&p.arp.arp_sha, source_eth, ETH_ALEN);
 148         memcpy(&p.arp.arp_spa, &source_nip, 4);
 149         memcpy(&p.arp.arp_tha, target_eth, ETH_ALEN);
 150         memcpy(&p.arp.arp_tpa, &target_nip, 4);
 151
 152         // send it
 153         // Even though sock_fd is already bound to G.iface_sockaddr, just send()
 154         // won't work, because "socket is not connected"
 155         // (and connect() won't fix that, "operation not supported").
 156         // Thus we sendto() to G.iface_sockaddr. I wonder which sockaddr
 157         // (from bind() or from sendto()?) kernel actually uses
 158         // to determine iface to emit the packet from...
 159         xsendto(sock_fd, &p, sizeof(p), &G.iface_sockaddr, sizeof(G.iface_sockaddr));
 160 #undef source_eth
 161 }
 162
 163 /**
 164  * Run a script.
 165  * argv[0]:intf argv[1]:script_name argv[2]:junk argv[3]:NULL
 166  */
 167 static int run(char *argv[3], const char *param, uint32_t nip)
 168 {
 169         int status;
 170         const char *addr = addr; /* for gcc */
 171         const char *fmt = "%s %s %s" + 3;
 172
 173         argv[2] = (char*)param;
 174
 175         VDBG("%s run %s %s\n", argv[0], argv[1], argv[2]);
 176
 177         if (nip != 0) {
 178                 addr = nip_to_a(nip);
 179                 xsetenv("ip", addr);
 180                 fmt -= 3;
 181         }
 182         bb_info_msg(fmt, argv[2], argv[0], addr);
 183
 184         status = spawn_and_wait(argv + 1);
 185         if (status < 0) {
 186                 bb_perror_msg("%s %s %s" + 3, argv[2], argv[0]);
 187                 return -errno;
 188         }
 189         if (status != 0)
 190                 bb_error_msg("script %s %s failed, exitcode=%d", argv[1], argv[2], status & 0xff);
 191         return status;
 192 }
 193
 194 /**
 195  * Return milliseconds of random delay, up to "secs" seconds.
 196  */
 197 static ALWAYS_INLINE unsigned random_delay_ms(unsigned secs)
 198 {
 199         return rand() % (secs * 1000);
 200 }
 201
 202 /**
 203  * main program
 204  */
 205 int zcip_main(int argc, char **argv) MAIN_EXTERNALLY_VISIBLE;
 206 int zcip_main(int argc UNUSED_PARAM, char **argv)
 207 {
 208         int state;
 209         char *r_opt;
 210         const char *l_opt = "169.254.0.0";
 211         unsigned opts;
 212
 213         // ugly trick, but I want these zeroed in one go
 214         struct {
 215                 const struct ether_addr null_addr;
 216                 struct ifreq ifr;
 217                 uint32_t chosen_nip;
 218                 int timeout_ms; /* must be signed */
 219                 unsigned conflicts;
 220                 unsigned nprobes;
 221                 unsigned nclaims;
 222                 int verbose;
 223         } L;
 224 #define null_addr  (L.null_addr )
 225 #define chosen_nip (L.chosen_nip)
 226 #define ifr        (L.ifr       )
 227 #define timeout_ms (L.timeout_ms)
 228 #define conflicts  (L.conflicts )
 229 #define nprobes    (L.nprobes   )
 230 #define nclaims    (L.nclaims   )
 231 #define verbose    (L.verbose   )
 232
 233         memset(&L, 0, sizeof(L));
 234         INIT_G();
 235
 236 #define FOREGROUND (opts & 1)
 237 #define QUIT       (opts & 2)
 238         // parse commandline: prog [options] ifname script
 239         // exactly 2 args; -v accumulates and implies -f
 240         opt_complementary = "=2:vv:vf";
 241         opts = getopt32(argv, "fqr:l:v", &r_opt, &l_opt, &verbose);
 242 #if !BB_MMU
 243         // on NOMMU reexec early (or else we will rerun things twice)
 244         if (!FOREGROUND)
 245                 bb_daemonize_or_rexec(0 /*was: DAEMON_CHDIR_ROOT*/, argv);
 246 #endif
 247         // open an ARP socket
 248         // (need to do it before openlog to prevent openlog from taking
 249         // fd 3 (sock_fd==3))
 250         xmove_fd(xsocket(AF_PACKET, SOCK_PACKET, htons(ETH_P_ARP)), sock_fd);
 251         if (!FOREGROUND) {
 252                 // do it before all bb_xx_msg calls
 253                 openlog(applet_name, 0, LOG_DAEMON);
 254                 logmode |= LOGMODE_SYSLOG;
 255         }
 256         bb_logenv_override();
 257
 258         { // -l n.n.n.n
 259                 struct in_addr net;
 260                 if (inet_aton(l_opt, &net) == 0
 261                  || (net.s_addr & htonl(IN_CLASSB_NET)) != net.s_addr
 262                 ) {
 263                         bb_error_msg_and_die("invalid network address");
 264                 }
 265                 G.localnet_ip = ntohl(net.s_addr);
 266         }
 267         if (opts & 4) { // -r n.n.n.n
 268                 struct in_addr ip;
 269                 if (inet_aton(r_opt, &ip) == 0
 270                  || (ntohl(ip.s_addr) & IN_CLASSB_NET) != G.localnet_ip
 271                 ) {
 272                         bb_error_msg_and_die("invalid link address");
 273                 }
 274                 chosen_nip = ip.s_addr;
 275         }
 276         argv += optind - 1;
 277
 278         /* Now: argv[0]:junk argv[1]:intf argv[2]:script argv[3]:NULL */
 279         /* We need to make space for script argument: */
 280         argv[0] = argv[1];
 281         argv[1] = argv[2];
 282         /* Now: argv[0]:intf argv[1]:script argv[2]:junk argv[3]:NULL */
 283 #define argv_intf (argv[0])
 284
 285         xsetenv("interface", argv_intf);
 286
 287         // initialize the interface (modprobe, ifup, etc)
 288         if (run(argv, "init", 0))
 289                 return EXIT_FAILURE;
 290
 291         // initialize G.iface_sockaddr
 292         // G.iface_sockaddr is: { u16 sa_family; u8 sa_data[14]; }
 293         //memset(&G.iface_sockaddr, 0, sizeof(G.iface_sockaddr));
 294         //TODO: are we leaving sa_family == 0 (AF_UNSPEC)?!
 295         safe_strncpy(G.iface_sockaddr.sa_data, argv_intf, sizeof(G.iface_sockaddr.sa_data));
 296
 297         // bind to the interface's ARP socket
 298         xbind(sock_fd, &G.iface_sockaddr, sizeof(G.iface_sockaddr));
 299
 300         // get the interface's ethernet address
 301         //memset(&ifr, 0, sizeof(ifr));
 302         strncpy_IFNAMSIZ(ifr.ifr_name, argv_intf);
 303         xioctl(sock_fd, SIOCGIFHWADDR, &ifr);
 304         memcpy(&G.eth_addr, &ifr.ifr_hwaddr.sa_data, ETH_ALEN);
 305
 306         // start with some stable ip address, either a function of
 307         // the hardware address or else the last address we used.
 308         // we are taking low-order four bytes, as top-order ones
 309         // aren't random enough.
 310         // NOTE: the sequence of addresses we try changes only
 311         // depending on when we detect conflicts.
 312         {
 313                 uint32_t t;
 314                 move_from_unaligned32(t, ((char *)&G.eth_addr + 2));
 315                 srand(t);
 316         }
 317         if (chosen_nip == 0)
 318                 chosen_nip = pick_nip();
 319
 320         // FIXME cases to handle:
 321         //  - zcip already running!
 322         //  - link already has local address... just defend/update
 323
 324         // daemonize now; don't delay system startup
 325         if (!FOREGROUND) {
 326 #if BB_MMU
 327                 bb_daemonize(0 /*was: DAEMON_CHDIR_ROOT*/);
 328 #endif
 329                 bb_info_msg("start, interface %s", argv_intf);
 330         }
 331
 332         // run the dynamic address negotiation protocol,
 333         // restarting after address conflicts:
 334         //  - start with some address we want to try
 335         //  - short random delay
 336         //  - arp probes to see if another host uses it
 337         //    00:04:e2:64:23:c2 > ff:ff:ff:ff:ff:ff, ARP (0x0806): arp who-has 169.254.194.171 tell 0.0.0.0
 338         //  - arp announcements that we're claiming it
 339         //    00:04:e2:64:23:c2 > ff:ff:ff:ff:ff:ff, ARP (0x0806): arp who-has 169.254.194.171 (00:04:e2:64:23:c2) tell 169.254.194.171
 340         //  - use it
 341         //  - defend it, within limits
 342         // exit if:
 343         // - address is successfully obtained and -q was given:
 344         //   run "<script> config", then exit with exitcode 0
 345         // - poll error (when does this happen?)
 346         // - read error (when does this happen?)
 347         // - sendto error (in arp()) (when does this happen?)
 348         // - revents & POLLERR (link down). run "<script> deconfig" first
 349         state = PROBE;
 350         while (1) {
 351                 struct pollfd fds[1];
 352                 unsigned deadline_us;
 353                 struct arp_packet p;
 354                 int source_ip_conflict;
 355                 int target_ip_conflict;
 356
 357                 fds[0].fd = sock_fd;
 358                 fds[0].events = POLLIN;
 359                 fds[0].revents = 0;
 360
 361                 // poll, being ready to adjust current timeout
 362                 if (!timeout_ms) {
 363                         timeout_ms = random_delay_ms(PROBE_WAIT);
 364                         // FIXME setsockopt(sock_fd, SO_ATTACH_FILTER, ...) to
 365                         // make the kernel filter out all packets except
 366                         // ones we'd care about.
 367                 }
 368                 // set deadline_us to the point in time when we timeout
 369                 deadline_us = MONOTONIC_US() + timeout_ms * 1000;
 370
 371                 VDBG("...wait %d %s nprobes=%u, nclaims=%u\n",
 372                                 timeout_ms, argv_intf, nprobes, nclaims);
 373
 374                 switch (safe_poll(fds, 1, timeout_ms)) {
 375
 376                 default:
 377                         //bb_perror_msg("poll"); - done in safe_poll
 378                         return EXIT_FAILURE;
 379
 380                 // timeout
 381                 case 0:
 382                         VDBG("state = %d\n", state);
 383                         switch (state) {
 384                         case PROBE:
 385                                 // timeouts in the PROBE state mean no conflicting ARP packets
 386                                 // have been received, so we can progress through the states
 387                                 if (nprobes < PROBE_NUM) {
 388                                         nprobes++;
 389                                         VDBG("probe/%u %s@%s\n",
 390                                                         nprobes, argv_intf, nip_to_a(chosen_nip));
 391                                         timeout_ms = PROBE_MIN * 1000;
 392                                         timeout_ms += random_delay_ms(PROBE_MAX - PROBE_MIN);
 393                                         arp(/* ARPOP_REQUEST, */
 394                                                         /* &G.eth_addr, */ 0,
 395                                                         &null_addr, chosen_nip);
 396                                         break;
 397                                 }
 398                                 // Switch to announce state.
 399                                 nclaims = 0;
 400                                 state = ANNOUNCE;
 401                                 goto send_announce;
 402                         case ANNOUNCE:
 403                                 // timeouts in the ANNOUNCE state mean no conflicting ARP packets
 404                                 // have been received, so we can progress through the states
 405                                 if (nclaims < ANNOUNCE_NUM) {
 406  send_announce:
 407                                         nclaims++;
 408                                         VDBG("announce/%u %s@%s\n",
 409                                                         nclaims, argv_intf, nip_to_a(chosen_nip));
 410                                         timeout_ms = ANNOUNCE_INTERVAL * 1000;
 411                                         arp(/* ARPOP_REQUEST, */
 412                                                         /* &G.eth_addr, */ chosen_nip,
 413                                                         &G.eth_addr, chosen_nip);
 414                                         break;
 415                                 }
 416                                 // Switch to monitor state.
 417                                 // FIXME update filters
 418                                 run(argv, "config", chosen_nip);
 419                                 // NOTE: all other exit paths should deconfig...
 420                                 if (QUIT)
 421                                         return EXIT_SUCCESS;
 422                                 conflicts = 0;
 423                                 timeout_ms = -1; // Never timeout in the monitor state.
 424                                 state = MONITOR;
 425                                 break;
 426                         case DEFEND:
 427                                 // Defend period ended with no ARP replies - we won.
 428                                 conflicts = 0;
 429                                 timeout_ms = -1;
 430                                 state = MONITOR;
 431                                 break;
 432                         } // switch (state)
 433                         break; // case 0 (timeout)
 434
 435                 // packets arriving, or link went down
 436                 case 1:
 437                         // We need to adjust the timeout in case we didn't receive
 438                         // a conflicting packet.
 439                         if (timeout_ms > 0) {
 440                                 unsigned diff = deadline_us - MONOTONIC_US();
 441                                 if ((int)(diff) < 0) {
 442                                         // Current time is greater than the expected timeout time.
 443                                         diff = 0;
 444                                 }
 445                                 VDBG("adjusting timeout\n");
 446                                 timeout_ms = (diff / 1000) | 1; /* never 0 */
 447                         }
 448
 449                         if ((fds[0].revents & POLLIN) == 0) {
 450                                 if (fds[0].revents & POLLERR) {
 451                                         // FIXME: links routinely go down;
 452                                         // this shouldn't necessarily exit.
 453                                         bb_error_msg("iface %s is down", argv_intf);
 454                                         if (state >= MONITOR) {
 455                                                 /* only if we are in MONITOR or DEFEND */
 456                                                 run(argv, "deconfig", chosen_nip);
 457                                         }
 458                                         return EXIT_FAILURE;
 459                                 }
 460                                 continue;
 461                         }
 462
 463                         // read ARP packet
 464                         if (safe_read(sock_fd, &p, sizeof(p)) < 0) {
 465                                 bb_perror_msg_and_die(bb_msg_read_error);
 466                         }
 467
 468                         if (p.eth.ether_type != htons(ETHERTYPE_ARP))
 469                                 continue;
 470                         if (p.arp.arp_op != htons(ARPOP_REQUEST)
 471                          && p.arp.arp_op != htons(ARPOP_REPLY)
 472                         ) {
 473                                 continue;
 474                         }
 475 #ifdef DEBUG
 476                         {
 477                                 struct ether_addr *sha = (struct ether_addr *) p.arp.arp_sha;
 478                                 struct ether_addr *tha = (struct ether_addr *) p.arp.arp_tha;
 479                                 struct in_addr *spa = (struct in_addr *) p.arp.arp_spa;
 480                                 struct in_addr *tpa = (struct in_addr *) p.arp.arp_tpa;
 481                                 VDBG("%s recv arp type=%d, op=%d,\n",
 482                                         argv_intf, ntohs(p.eth.ether_type),
 483                                         ntohs(p.arp.arp_op));
 484                                 VDBG("\tsource=%s %s\n",
 485                                         ether_ntoa(sha),
 486                                         inet_ntoa(*spa));
 487                                 VDBG("\ttarget=%s %s\n",
 488                                         ether_ntoa(tha),
 489                                         inet_ntoa(*tpa));
 490                         }
 491 #endif
 492                         source_ip_conflict = 0;
 493                         target_ip_conflict = 0;
 494
 495                         if (memcmp(&p.arp.arp_sha, &G.eth_addr, ETH_ALEN) != 0) {
 496                                 if (memcmp(p.arp.arp_spa, &chosen_nip, 4) == 0) {
 497                                         /* A probe or reply with source_ip == chosen ip */
 498                                         source_ip_conflict = 1;
 499                                 }
 500                                 if (p.arp.arp_op == htons(ARPOP_REQUEST)
 501                                  && memcmp(p.arp.arp_spa, &const_int_0, 4) == 0
 502                                  && memcmp(p.arp.arp_tpa, &chosen_nip, 4) == 0
 503                                 ) {
 504                                         /* A probe with source_ip == 0.0.0.0, target_ip == chosen ip:
 505                                          * another host trying to claim this ip!
 506                                          */
 507                                         target_ip_conflict = 1;
 508                                 }
 509                         }
 510
 511                         VDBG("state = %d, source ip conflict = %d, target ip conflict = %d\n",
 512                                 state, source_ip_conflict, target_ip_conflict);
 513                         switch (state) {
 514                         case PROBE:
 515                         case ANNOUNCE:
 516                                 // When probing or announcing, check for source IP conflicts
 517                                 // and other hosts doing ARP probes (target IP conflicts).
 518                                 if (source_ip_conflict || target_ip_conflict) {
 519                                         conflicts++;
 520                                         timeout_ms = PROBE_MIN * 1000
 521                                                 + CONFLICT_MULTIPLIER * random_delay_ms(conflicts);
 522                                         chosen_nip = pick_nip();
 523                                         nprobes = 0;
 524                                         nclaims = 0;
 525                                         state = PROBE;
 526                                 }
 527                                 break;
 528                         case MONITOR:
 529                                 // If a conflict, we try to defend with a single ARP probe.
 530                                 if (source_ip_conflict) {
 531                                         VDBG("monitor conflict -- defending\n");
 532                                         timeout_ms = DEFEND_INTERVAL * 1000;
 533                                         state = DEFEND;
 534                                         arp(/* ARPOP_REQUEST, */
 535                                                 /* &G.eth_addr, */ chosen_nip,
 536                                                 &G.eth_addr, chosen_nip);
 537                                 }
 538                                 break;
 539                         case DEFEND:
 540                                 // Well, we tried.  Start over (on conflict).
 541                                 if (source_ip_conflict) {
 542                                         VDBG("defend conflict -- starting over\n");
 543                                         run(argv, "deconfig", chosen_nip);
 544
 545                                         // restart the whole protocol
 546                                         timeout_ms = 0;
 547                                         chosen_nip = pick_nip();
 548                                         nprobes = 0;
 549                                         nclaims = 0;
 550                                         state = PROBE;
 551                                 }
 552                                 break;
 553                         } // switch state
 554                         break; // case 1 (packets arriving)
 555                 } // switch poll
 556         } // while (1)
 557 #undef argv_intf
 558 }