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         // FIXME this part is netinet/if_ether.h "struct ether_arp"
  47         struct arphdr arp;
  48         struct ether_addr source_addr;
  49         struct in_addr source_ip;
  50         struct ether_addr target_addr;
  51         struct in_addr target_ip;
  52 } ATTRIBUTE_PACKED;
  53
  54 enum {
  55 /* 169.254.0.0 */
  56         LINKLOCAL_ADDR = 0xa9fe0000,
  57
  58 /* protocol timeout parameters, specified in seconds */
  59         PROBE_WAIT = 1,
  60         PROBE_MIN = 1,
  61         PROBE_MAX = 2,
  62         PROBE_NUM = 3,
  63         MAX_CONFLICTS = 10,
  64         RATE_LIMIT_INTERVAL = 60,
  65         ANNOUNCE_WAIT = 2,
  66         ANNOUNCE_NUM = 2,
  67         ANNOUNCE_INTERVAL = 2,
  68         DEFEND_INTERVAL = 10
  69 };
  70
  71 static const struct in_addr null_ip = { 0 };
  72 static const struct ether_addr null_addr = { {0, 0, 0, 0, 0, 0} };
  73
  74 static int verbose = 0;
  75
  76 #define DBG(fmt,args...) \
  77         do { } while (0)
  78 #define VDBG    DBG
  79
  80 /**
  81  * Pick a random link local IP address on 169.254/16, except that
  82  * the first and last 256 addresses are reserved.
  83  */
  84 static void pick(struct in_addr *ip)
  85 {
  86         unsigned        tmp;
  87
  88         /* use cheaper math than lrand48() mod N */
  89         do {
  90                 tmp = (lrand48() >> 16) & IN_CLASSB_HOST;
  91         } while (tmp > (IN_CLASSB_HOST - 0x0200));
  92         ip->s_addr = htonl((LINKLOCAL_ADDR + 0x0100) + tmp);
  93 }
  94
  95 /**
  96  * Broadcast an ARP packet.
  97  */
  98 static int arp(int fd, struct sockaddr *saddr, int op,
  99         const struct ether_addr *source_addr, struct in_addr source_ip,
 100         const struct ether_addr *target_addr, struct in_addr target_ip)
 101 {
 102         struct arp_packet p;
 103
 104         // ether header
 105         p.hdr.ether_type = htons(ETHERTYPE_ARP);
 106         memcpy(p.hdr.ether_shost, source_addr, ETH_ALEN);
 107         memset(p.hdr.ether_dhost, 0xff, ETH_ALEN);
 108
 109         // arp request
 110         p.arp.ar_hrd = htons(ARPHRD_ETHER);
 111         p.arp.ar_pro = htons(ETHERTYPE_IP);
 112         p.arp.ar_hln = ETH_ALEN;
 113         p.arp.ar_pln = 4;
 114         p.arp.ar_op = htons(op);
 115         memcpy(&p.source_addr, source_addr, ETH_ALEN);
 116         memcpy(&p.source_ip, &source_ip, sizeof (p.source_ip));
 117         memcpy(&p.target_addr, target_addr, ETH_ALEN);
 118         memcpy(&p.target_ip, &target_ip, sizeof (p.target_ip));
 119
 120         // send it
 121         if (sendto(fd, &p, sizeof (p), 0, saddr, sizeof (*saddr)) < 0) {
 122                 perror("sendto");
 123                 return -errno;
 124         }
 125         return 0;
 126 }
 127
 128 /**
 129  * Run a script.
 130  */
 131 static int run(char *script, char *arg, char *intf, struct in_addr *ip)
 132 {
 133         int pid, status;
 134         char *why;
 135
 136         if (script != NULL) {
 137                 VDBG("%s run %s %s\n", intf, script, arg);
 138                 if (ip != NULL) {
 139                         char *addr = inet_ntoa(*ip);
 140                         setenv("ip", addr, 1);
 141                         syslog(LOG_INFO, "%s %s %s", arg, intf, addr);
 142                 }
 143
 144                 pid = vfork();
 145                 if (pid < 0) {                  // error
 146                         why = "vfork";
 147                         goto bad;
 148                 } else if (pid == 0) {          // child
 149                         execl(script, script, arg, NULL);
 150                         perror("execl");
 151                         _exit(EXIT_FAILURE);
 152                 }
 153
 154                 if (waitpid(pid, &status, 0) <= 0) {
 155                         why = "waitpid";
 156                         goto bad;
 157                 }
 158                 if (WEXITSTATUS(status) != 0) {
 159                         bb_error_msg("script %s failed, exit=%d\n",
 160                                         script, WEXITSTATUS(status));
 161                         return -errno;
 162                 }
 163         }
 164         return 0;
 165 bad:
 166         status = -errno;
 167         syslog(LOG_ERR, "%s %s, %s error: %s",
 168                 arg, intf, why, strerror(errno));
 169         return status;
 170 }
 171
 172
 173 /**
 174  * Return milliseconds of random delay, up to "secs" seconds.
 175  */
 176 static inline unsigned ms_rdelay(unsigned secs)
 177 {
 178         return lrand48() % (secs * 1000);
 179 }
 180
 181 /**
 182  * main program
 183  */
 184
 185 int zcip_main(int argc, char *argv[])
 186 {
 187         char *intf = NULL;
 188         char *script = NULL;
 189         int quit = 0;
 190         int foreground = 0;
 191
 192         char *why;
 193         struct sockaddr saddr;
 194         struct ether_addr addr;
 195         struct in_addr ip = { 0 };
 196         int fd;
 197         int ready = 0;
 198         suseconds_t timeout = 0;        // milliseconds
 199         time_t defend = 0;
 200         unsigned conflicts = 0;
 201         unsigned nprobes = 0;
 202         unsigned nclaims = 0;
 203         int t;
 204
 205         // parse commandline: prog [options] ifname script
 206         while ((t = getopt(argc, argv, "fqr:v")) != EOF) {
 207                 switch (t) {
 208                 case 'f':
 209                         foreground = 1;
 210                         continue;
 211                 case 'q':
 212                         quit = 1;
 213                         continue;
 214                 case 'r':
 215                         if (inet_aton(optarg, &ip) == 0
 216                                         || (ntohl(ip.s_addr) & IN_CLASSB_NET)
 217                                                 != LINKLOCAL_ADDR) {
 218                                 bb_error_msg_and_die("invalid link address");
 219                         }
 220                         continue;
 221                 case 'v':
 222                         verbose++;
 223                         foreground = 1;
 224                         continue;
 225                 default:
 226                         bb_error_msg_and_die("bad option");
 227                 }
 228         }
 229         if (optind < argc - 1) {
 230                 intf = argv[optind++];
 231                 setenv("interface", intf, 1);
 232                 script = argv[optind++];
 233         }
 234         if (optind != argc || !intf)
 235                 bb_show_usage();
 236         openlog(bb_applet_name, 0, LOG_DAEMON);
 237
 238         // initialize the interface (modprobe, ifup, etc)
 239         if (run(script, "init", intf, NULL) < 0)
 240                 return EXIT_FAILURE;
 241
 242         // initialize saddr
 243         memset(&saddr, 0, sizeof (saddr));
 244         safe_strncpy(saddr.sa_data, intf, sizeof (saddr.sa_data));
 245
 246         // open an ARP socket
 247         if ((fd = socket(PF_PACKET, SOCK_PACKET, htons(ETH_P_ARP))) < 0) {
 248                 why = "open";
 249 fail:
 250                 foreground = 1;
 251                 goto bad;
 252         }
 253         // bind to the interface's ARP socket
 254         if (bind(fd, &saddr, sizeof (saddr)) < 0) {
 255                 why = "bind";
 256                 goto fail;
 257         } else {
 258                 struct ifreq ifr;
 259                 unsigned short seed[3];
 260
 261                 // get the interface's ethernet address
 262                 memset(&ifr, 0, sizeof (ifr));
 263                 strncpy(ifr.ifr_name, intf, sizeof (ifr.ifr_name));
 264                 if (ioctl(fd, SIOCGIFHWADDR, &ifr) < 0) {
 265                         why = "get ethernet address";
 266                         goto fail;
 267                 }
 268                 memcpy(&addr, &ifr.ifr_hwaddr.sa_data, ETH_ALEN);
 269
 270                 // start with some stable ip address, either a function of
 271                 // the hardware address or else the last address we used.
 272                 // NOTE: the sequence of addresses we try changes only
 273                 // depending on when we detect conflicts.
 274                 memcpy(seed, &ifr.ifr_hwaddr.sa_data, ETH_ALEN);
 275                 seed48(seed);
 276                 if (ip.s_addr == 0)
 277                         pick(&ip);
 278         }
 279
 280         // FIXME cases to handle:
 281         //  - zcip already running!
 282         //  - link already has local address... just defend/update
 283
 284         // daemonize now; don't delay system startup
 285         if (!foreground) {
 286                 if (daemon(0, verbose) < 0) {
 287                         why = "daemon";
 288                         goto bad;
 289                 }
 290                 syslog(LOG_INFO, "start, interface %s", intf);
 291         }
 292
 293         // run the dynamic address negotiation protocol,
 294         // restarting after address conflicts:
 295         //  - start with some address we want to try
 296         //  - short random delay
 297         //  - arp probes to see if another host else uses it
 298         //  - arp announcements that we're claiming it
 299         //  - use it
 300         //  - defend it, within limits
 301         while (1) {
 302                 struct pollfd fds[1];
 303                 struct timeval tv1;
 304                 struct arp_packet p;
 305
 306                 fds[0].fd = fd;
 307                 fds[0].events = POLLIN;
 308                 fds[0].revents = 0;
 309
 310                 // poll, being ready to adjust current timeout
 311                 if (!timeout) {
 312                         timeout = ms_rdelay(PROBE_WAIT);
 313                         // FIXME setsockopt(fd, SO_ATTACH_FILTER, ...) to
 314                         // make the kernel filter out all packets except
 315                         // ones we'd care about.
 316                 }
 317                 gettimeofday(&tv1, NULL);
 318                 tv1.tv_usec += (timeout % 1000) * 1000;
 319                 while (tv1.tv_usec > 1000000) {
 320                         tv1.tv_usec -= 1000000;
 321                         tv1.tv_sec++;
 322                 }
 323                 tv1.tv_sec += timeout / 1000;
 324
 325                 VDBG("...wait %ld %s nprobes=%d, nclaims=%d\n",
 326                                 timeout, intf, nprobes, nclaims);
 327                 switch (poll(fds, 1, timeout)) {
 328
 329                 // timeouts trigger protocol transitions
 330                 case 0:
 331                         // probes
 332                         if (nprobes < PROBE_NUM) {
 333                                 nprobes++;
 334                                 VDBG("probe/%d %s@%s\n",
 335                                                 nprobes, intf, inet_ntoa(ip));
 336                                 (void)arp(fd, &saddr, ARPOP_REQUEST,
 337                                                 &addr, null_ip,
 338                                                 &null_addr, ip);
 339                                 if (nprobes < PROBE_NUM) {
 340                                         timeout = PROBE_MIN * 1000;
 341                                         timeout += ms_rdelay(PROBE_MAX
 342                                                         - PROBE_MIN);
 343                                 } else
 344                                         timeout = ANNOUNCE_WAIT * 1000;
 345                         }
 346                         // then announcements
 347                         else if (nclaims < ANNOUNCE_NUM) {
 348                                 nclaims++;
 349                                 VDBG("announce/%d %s@%s\n",
 350                                                 nclaims, intf, inet_ntoa(ip));
 351                                 (void)arp(fd, &saddr, ARPOP_REQUEST,
 352                                                 &addr, ip,
 353                                                 &addr, ip);
 354                                 if (nclaims < ANNOUNCE_NUM) {
 355                                         timeout = ANNOUNCE_INTERVAL * 1000;
 356                                 } else {
 357                                         // link is ok to use earlier
 358                                         run(script, "config", intf, &ip);
 359                                         ready = 1;
 360                                         conflicts = 0;
 361                                         timeout = -1;
 362
 363                                         // NOTE:  all other exit paths
 364                                         // should deconfig ...
 365                                         if (quit)
 366                                                 return EXIT_SUCCESS;
 367                                         // FIXME update filters
 368                                 }
 369                         }
 370                         break;
 371
 372                 // packets arriving
 373                 case 1:
 374                         // maybe adjust timeout
 375                         if (timeout > 0) {
 376                                 struct timeval tv2;
 377
 378                                 gettimeofday(&tv2, NULL);
 379                                 if (timercmp(&tv1, &tv2, <)) {
 380                                         timeout = 0;
 381                                 } else {
 382                                         timersub(&tv1, &tv2, &tv1);
 383                                         timeout = 1000 * tv1.tv_sec
 384                                                         + tv1.tv_usec / 1000;
 385                                 }
 386                         }
 387                         if ((fds[0].revents & POLLIN) == 0) {
 388                                 if (fds[0].revents & POLLERR) {
 389                                         // FIXME: links routinely go down;
 390                                         // this shouldn't necessarily exit.
 391                                         bb_error_msg("%s: poll error\n", intf);
 392                                         if (ready) {
 393                                                 run(script, "deconfig",
 394                                                                 intf, &ip);
 395                                         }
 396                                         return EXIT_FAILURE;
 397                                 }
 398                                 continue;
 399                         }
 400                         // read ARP packet
 401                         if (recv(fd, &p, sizeof (p), 0) < 0) {
 402                                 why = "recv";
 403                                 goto bad;
 404                         }
 405                         if (p.hdr.ether_type != htons(ETHERTYPE_ARP))
 406                                 continue;
 407
 408                         VDBG("%s recv arp type=%d, op=%d,\n",
 409                                         intf, ntohs(p.hdr.ether_type),
 410                                         ntohs(p.arp.ar_op));
 411                         VDBG("\tsource=%s %s\n",
 412                                         ether_ntoa(&p.source_addr),
 413                                         inet_ntoa(p.source_ip));
 414                         VDBG("\ttarget=%s %s\n",
 415                                         ether_ntoa(&p.target_addr),
 416                                         inet_ntoa(p.target_ip));
 417                         if (p.arp.ar_op != htons(ARPOP_REQUEST)
 418                                         && p.arp.ar_op != htons(ARPOP_REPLY))
 419                                 continue;
 420
 421                         // some cases are always conflicts
 422                         if ((p.source_ip.s_addr == ip.s_addr)
 423                                         && (memcmp(&addr, &p.source_addr,
 424                                                         ETH_ALEN) != 0)) {
 425 collision:
 426                                 VDBG("%s ARP conflict from %s\n", intf,
 427                                                 ether_ntoa(&p.source_addr));
 428                                 if (ready) {
 429                                         time_t now = time(0);
 430
 431                                         if ((defend + DEFEND_INTERVAL)
 432                                                         < now) {
 433                                                 defend = now;
 434                                                 (void)arp(fd, &saddr,
 435                                                                 ARPOP_REQUEST,
 436                                                                 &addr, ip,
 437                                                                 &addr, ip);
 438                                                 VDBG("%s defend\n", intf);
 439                                                 timeout = -1;
 440                                                 continue;
 441                                         }
 442                                         defend = now;
 443                                         ready = 0;
 444                                         run(script, "deconfig", intf, &ip);
 445                                         // FIXME rm filters: setsockopt(fd,
 446                                         // SO_DETACH_FILTER, ...)
 447                                 }
 448                                 conflicts++;
 449                                 if (conflicts >= MAX_CONFLICTS) {
 450                                         VDBG("%s ratelimit\n", intf);
 451                                         sleep(RATE_LIMIT_INTERVAL);
 452                                 }
 453                                 // restart the whole protocol
 454                                 pick(&ip);
 455                                 timeout = 0;
 456                                 nprobes = 0;
 457                                 nclaims = 0;
 458                         }
 459                         // two hosts probing one address is a collision too
 460                         else if (p.target_ip.s_addr == ip.s_addr
 461                                         && nclaims == 0
 462                                         && p.arp.ar_op == htons(ARPOP_REQUEST)
 463                                         && memcmp(&addr, &p.target_addr,
 464                                                         ETH_ALEN) != 0) {
 465                                 goto collision;
 466                         }
 467                         break;
 468
 469                 default:
 470                         why = "poll";
 471                         goto bad;
 472                 }
 473         }
 474 bad:
 475         if (foreground)
 476                 perror(why);
 477         else
 478                 syslog(LOG_ERR, "%s %s, %s error: %s",
 479                         bb_applet_name, intf, why, strerror(errno));
 480         return EXIT_FAILURE;
 481 }