sys/netinet/ip_carp.c

   1 /*
   2  * Copyright (c) 2002 Michael Shalayeff. All rights reserved.
   3  * Copyright (c) 2003 Ryan McBride. All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  *
  14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  15  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  16  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  17  * IN NO EVENT SHALL THE AUTHOR OR HIS RELATIVES BE LIABLE FOR ANY DIRECT,
  18  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  19  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  20  * SERVICES; LOSS OF MIND, USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
  22  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
  23  * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
  24  * THE POSSIBILITY OF SUCH DAMAGE.
  25  */
  26 /*
  27  * $FreeBSD: src/sys/netinet/ip_carp.c,v 1.48 2007/02/02 09:39:09 glebius Exp $
  28  * $DragonFly: src/sys/netinet/ip_carp.c,v 1.10 2008/07/27 10:06:57 sephe Exp $
  29  */
  30
  31 #include "opt_carp.h"
  32 #include "opt_inet.h"
  33 #include "opt_inet6.h"
  34
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/kernel.h>
  38 #include <sys/in_cksum.h>
  39 #include <sys/limits.h>
  40 #include <sys/malloc.h>
  41 #include <sys/mbuf.h>
  42 #include <sys/time.h>
  43 #include <sys/proc.h>
  44 #include <sys/priv.h>
  45 #include <sys/sockio.h>
  46 #include <sys/socket.h>
  47 #include <sys/sysctl.h>
  48 #include <sys/syslog.h>
  49
  50 #include <machine/stdarg.h>
  51 #include <crypto/sha1.h>
  52
  53 #include <net/bpf.h>
  54 #include <net/ethernet.h>
  55 #include <net/if.h>
  56 #include <net/if_dl.h>
  57 #include <net/if_types.h>
  58 #include <net/route.h>
  59 #include <net/if_clone.h>
  60
  61 #ifdef INET
  62 #include <netinet/in.h>
  63 #include <netinet/in_var.h>
  64 #include <netinet/in_systm.h>
  65 #include <netinet/ip.h>
  66 #include <netinet/ip_var.h>
  67 #include <netinet/if_ether.h>
  68 #endif
  69
  70 #ifdef INET6
  71 #include <netinet/icmp6.h>
  72 #include <netinet/ip6.h>
  73 #include <netinet6/ip6_var.h>
  74 #include <netinet6/scope6_var.h>
  75 #include <netinet6/nd6.h>
  76 #endif
  77
  78 #include <netinet/ip_carp.h>
  79
  80 #define CARP_IFNAME             "carp"
  81 #define CARP_IS_RUNNING(ifp)    \
  82         (((ifp)->if_flags & (IFF_UP | IFF_RUNNING)) == (IFF_UP | IFF_RUNNING))
  83
  84 struct carp_softc {
  85         struct ifnet             sc_if;
  86         struct ifnet            *sc_ifp;        /* compat shim */
  87         struct ifnet            *sc_carpdev;    /* parent interface */
  88         struct in_ifaddr        *sc_ia;         /* primary iface address */
  89         struct ip_moptions       sc_imo;
  90 #ifdef INET6
  91         struct in6_ifaddr       *sc_ia6;        /* primary iface address v6 */
  92         struct ip6_moptions      sc_im6o;
  93 #endif /* INET6 */
  94         TAILQ_ENTRY(carp_softc)  sc_list;
  95
  96         enum { INIT = 0, BACKUP, MASTER }
  97                                  sc_state;
  98
  99         int                      sc_flags_backup;
 100         int                      sc_suppress;
 101
 102         int                      sc_sendad_errors;
 103 #define CARP_SENDAD_MAX_ERRORS  3
 104         int                      sc_sendad_success;
 105 #define CARP_SENDAD_MIN_SUCCESS 3
 106
 107         int                      sc_vhid;
 108         int                      sc_advskew;
 109         int                      sc_naddrs;
 110         int                      sc_naddrs6;
 111         int                      sc_advbase;    /* seconds */
 112         int                      sc_init_counter;
 113         uint64_t                 sc_counter;
 114
 115         /* authentication */
 116 #define CARP_HMAC_PAD   64
 117         unsigned char            sc_key[CARP_KEY_LEN];
 118         unsigned char            sc_pad[CARP_HMAC_PAD];
 119         SHA1_CTX                 sc_sha1;
 120
 121         struct callout           sc_ad_tmo;     /* advertisement timeout */
 122         struct callout           sc_md_tmo;     /* master down timeout */
 123         struct callout           sc_md6_tmo;    /* master down timeout */
 124
 125         LIST_ENTRY(carp_softc)   sc_next;       /* Interface clue */
 126 };
 127 #define SC2IFP(sc)      ((sc)->sc_ifp)
 128
 129 struct carp_if {
 130         TAILQ_HEAD(, carp_softc) vhif_vrs;
 131         int             vhif_nvrs;
 132
 133         struct ifnet    *vhif_ifp;
 134 };
 135
 136 enum    { CARP_COUNT_MASTER, CARP_COUNT_RUNNING };
 137
 138 SYSCTL_DECL(_net_inet_carp);
 139
 140 static int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 1, 0, 0 }; /* XXX for now */
 141 SYSCTL_INT(_net_inet_carp, CARPCTL_ALLOW, allow, CTLFLAG_RW,
 142     &carp_opts[CARPCTL_ALLOW], 0, "Accept incoming CARP packets");
 143 SYSCTL_INT(_net_inet_carp, CARPCTL_PREEMPT, preempt, CTLFLAG_RW,
 144     &carp_opts[CARPCTL_PREEMPT], 0, "high-priority backup preemption mode");
 145 SYSCTL_INT(_net_inet_carp, CARPCTL_LOG, log, CTLFLAG_RW,
 146     &carp_opts[CARPCTL_LOG], 0, "log bad carp packets");
 147 SYSCTL_INT(_net_inet_carp, CARPCTL_ARPBALANCE, arpbalance, CTLFLAG_RW,
 148     &carp_opts[CARPCTL_ARPBALANCE], 0, "balance arp responses");
 149
 150 static int carp_suppress_preempt = 0;
 151 SYSCTL_INT(_net_inet_carp, OID_AUTO, suppress_preempt, CTLFLAG_RD,
 152     &carp_suppress_preempt, 0, "Preemption is suppressed");
 153
 154 static struct carpstats carpstats;
 155 SYSCTL_STRUCT(_net_inet_carp, CARPCTL_STATS, stats, CTLFLAG_RW,
 156     &carpstats, carpstats,
 157     "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 158
 159 #define CARP_LOG(...)   do {                            \
 160         if (carp_opts[CARPCTL_LOG] > 0)                 \
 161                 log(LOG_INFO, __VA_ARGS__);             \
 162 } while (0)
 163
 164 #define CARP_DEBUG(...) do {                            \
 165         if (carp_opts[CARPCTL_LOG] > 1)                 \
 166                 log(LOG_DEBUG, __VA_ARGS__);            \
 167 } while (0)
 168
 169 static void     carp_hmac_prepare(struct carp_softc *);
 170 static void     carp_hmac_generate(struct carp_softc *, uint32_t *,
 171                     unsigned char *);
 172 static int      carp_hmac_verify(struct carp_softc *, uint32_t *,
 173                     unsigned char *);
 174 static void     carp_setroute(struct carp_softc *, int);
 175 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 176 static int      carp_clone_create(struct if_clone *, int);
 177 static void     carp_clone_destroy(struct ifnet *);
 178 static void     carpdetach(struct carp_softc *, int);
 179 static int      carp_prepare_ad(struct mbuf *, struct carp_softc *,
 180                     struct carp_header *);
 181 static void     carp_send_ad_all(void);
 182 static void     carp_send_ad(void *);
 183 static void     carp_send_ad_locked(struct carp_softc *);
 184 static void     carp_send_arp(struct carp_softc *);
 185 static void     carp_master_down(void *);
 186 static void     carp_master_down_locked(struct carp_softc *);
 187 static int      carp_ioctl(struct ifnet *, u_long, caddr_t, struct ucred *);
 188 static int      carp_looutput(struct ifnet *, struct mbuf *, struct sockaddr *,
 189                     struct rtentry *);
 190 static void     carp_start(struct ifnet *);
 191 static void     carp_setrun(struct carp_softc *, sa_family_t);
 192 static void     carp_set_state(struct carp_softc *, int);
 193 static int      carp_addrcount(struct carp_if *, struct in_ifaddr *, int);
 194
 195 static void     carp_multicast_cleanup(struct carp_softc *);
 196 static int      carp_set_addr(struct carp_softc *, struct sockaddr_in *);
 197 static int      carp_del_addr(struct carp_softc *, struct sockaddr_in *);
 198 static void     carp_carpdev_state_locked(struct carp_if *);
 199 static void     carp_sc_state_locked(struct carp_softc *);
 200 #ifdef INET6
 201 static void     carp_send_na(struct carp_softc *);
 202 static int      carp_set_addr6(struct carp_softc *, struct sockaddr_in6 *);
 203 static int      carp_del_addr6(struct carp_softc *, struct sockaddr_in6 *);
 204 static void     carp_multicast6_cleanup(struct carp_softc *);
 205 #endif
 206
 207 static MALLOC_DEFINE(M_CARP, "CARP", "CARP interfaces");
 208
 209 static LIST_HEAD(, carp_softc) carpif_list;
 210
 211 static struct if_clone carp_cloner =
 212 IF_CLONE_INITIALIZER(CARP_IFNAME, carp_clone_create, carp_clone_destroy,
 213                      0, IF_MAXUNIT);
 214
 215 static eventhandler_tag carp_ifdetach_event;
 216
 217 static void
 218 carp_hmac_prepare(struct carp_softc *sc)
 219 {
 220         uint8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 221         uint8_t vhid = sc->sc_vhid & 0xff;
 222         struct ifaddr_container *ifac;
 223         int i;
 224 #ifdef INET6
 225         struct in6_addr in6;
 226 #endif
 227
 228         /* XXX: possible race here */
 229
 230         /* compute ipad from key */
 231         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 232         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 233         for (i = 0; i < sizeof(sc->sc_pad); i++)
 234                 sc->sc_pad[i] ^= 0x36;
 235
 236         /* precompute first part of inner hash */
 237         SHA1Init(&sc->sc_sha1);
 238         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 239         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 240         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 241         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 242 #ifdef INET
 243         TAILQ_FOREACH(ifac, &SC2IFP(sc)->if_addrheads[mycpuid], ifa_link) {
 244                 struct ifaddr *ifa = ifac->ifa;
 245
 246                 if (ifa->ifa_addr->sa_family == AF_INET)
 247                         SHA1Update(&sc->sc_sha1,
 248                             (void *)&ifatoia(ifa)->ia_addr.sin_addr.s_addr,
 249                             sizeof(struct in_addr));
 250         }
 251 #endif /* INET */
 252 #ifdef INET6
 253         TAILQ_FOREACH(ifac, &SC2IFP(sc)->if_addrheads[mycpuid], ifa_link) {
 254                 struct ifaddr *ifa = ifac->ifa;
 255
 256                 if (ifa->ifa_addr->sa_family == AF_INET6) {
 257                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 258                         in6_clearscope(&in6);
 259                         SHA1Update(&sc->sc_sha1, (void *)&in6, sizeof(in6));
 260                 }
 261         }
 262 #endif /* INET6 */
 263
 264         /* convert ipad to opad */
 265         for (i = 0; i < sizeof(sc->sc_pad); i++)
 266                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 267 }
 268
 269 static void
 270 carp_hmac_generate(struct carp_softc *sc, uint32_t counter[2],
 271     unsigned char md[20])
 272 {
 273         SHA1_CTX sha1ctx;
 274
 275         /* fetch first half of inner hash */
 276         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 277
 278         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 279         SHA1Final(md, &sha1ctx);
 280
 281         /* outer hash */
 282         SHA1Init(&sha1ctx);
 283         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 284         SHA1Update(&sha1ctx, md, 20);
 285         SHA1Final(md, &sha1ctx);
 286 }
 287
 288 static int
 289 carp_hmac_verify(struct carp_softc *sc, uint32_t counter[2],
 290     unsigned char md[20])
 291 {
 292         unsigned char md2[20];
 293
 294         carp_hmac_generate(sc, counter, md2);
 295         return (bcmp(md, md2, sizeof(md2)));
 296 }
 297
 298 static void
 299 carp_setroute(struct carp_softc *sc, int cmd)
 300 {
 301         struct ifaddr_container *ifac;
 302
 303         crit_enter();
 304         TAILQ_FOREACH(ifac, &SC2IFP(sc)->if_addrheads[mycpuid], ifa_link) {
 305                 struct ifaddr *ifa = ifac->ifa;
 306
 307                 if (ifa->ifa_addr->sa_family == AF_INET &&
 308                     sc->sc_carpdev != NULL) {
 309                         int count = carp_addrcount(
 310                             (struct carp_if *)sc->sc_carpdev->if_carp,
 311                             ifatoia(ifa), CARP_COUNT_MASTER);
 312
 313                         if ((cmd == RTM_ADD && count == 1) ||
 314                             (cmd == RTM_DELETE && count == 0))
 315                                 rtinit(ifa, cmd, RTF_UP | RTF_HOST);
 316                 }
 317 #ifdef INET6
 318                 if (ifa->ifa_addr->sa_family == AF_INET6) {
 319                         if (cmd == RTM_ADD)
 320                                 in6_ifaddloop(ifa);
 321                         else
 322                                 in6_ifremloop(ifa);
 323                 }
 324 #endif /* INET6 */
 325         }
 326         crit_exit();
 327 }
 328
 329 static int
 330 carp_clone_create(struct if_clone *ifc, int unit)
 331 {
 332         struct carp_softc *sc;
 333         struct ifnet *ifp;
 334
 335         sc = kmalloc(sizeof(*sc), M_CARP, M_WAITOK | M_ZERO);
 336         ifp = sc->sc_ifp = &sc->sc_if;
 337
 338         sc->sc_flags_backup = 0;
 339         sc->sc_suppress = 0;
 340         sc->sc_advbase = CARP_DFLTINTV;
 341         sc->sc_vhid = -1;       /* required setting */
 342         sc->sc_advskew = 0;
 343         sc->sc_init_counter = 1;
 344         sc->sc_naddrs = sc->sc_naddrs6 = 0;
 345
 346 #ifdef INET6
 347         sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL;
 348 #endif
 349
 350         callout_init(&sc->sc_ad_tmo);
 351         callout_init(&sc->sc_md_tmo);
 352         callout_init(&sc->sc_md6_tmo);
 353
 354         ifp->if_softc = sc;
 355         if_initname(ifp, CARP_IFNAME, unit);
 356         ifp->if_mtu = ETHERMTU;
 357         ifp->if_flags = IFF_LOOPBACK;
 358         ifp->if_ioctl = carp_ioctl;
 359         ifp->if_output = carp_looutput;
 360         ifp->if_start = carp_start;
 361         ifp->if_type = IFT_CARP;
 362         ifp->if_snd.ifq_maxlen = ifqmaxlen;
 363         ifp->if_hdrlen = 0;
 364         if_attach(ifp, NULL);
 365         bpfattach(ifp, DLT_NULL, sizeof(u_int));
 366
 367         crit_enter();
 368         LIST_INSERT_HEAD(&carpif_list, sc, sc_next);
 369         crit_exit();
 370
 371         return (0);
 372 }
 373
 374 static void
 375 carp_clone_destroy(struct ifnet *ifp)
 376 {
 377         struct carp_softc *sc = ifp->if_softc;
 378
 379         carpdetach(sc, 1);
 380
 381         crit_enter();
 382         LIST_REMOVE(sc, sc_next);
 383         crit_exit();
 384         bpfdetach(ifp);
 385         if_detach(ifp);
 386         kfree(sc, M_CARP);
 387 }
 388
 389 /*
 390  * This function can be called on CARP interface destroy path,
 391  * and in case of the removal of the underlying interface as
 392  * well. We differentiate these two cases. In the latter case
 393  * we do not cleanup our multicast memberships, since they
 394  * are already freed.
 395  */
 396 static void
 397 carpdetach(struct carp_softc *sc, int unlock)
 398 {
 399         struct carp_if *cif;
 400
 401         callout_stop(&sc->sc_ad_tmo);
 402         callout_stop(&sc->sc_md_tmo);
 403         callout_stop(&sc->sc_md6_tmo);
 404
 405         if (sc->sc_suppress)
 406                 carp_suppress_preempt--;
 407         sc->sc_suppress = 0;
 408
 409         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS)
 410                 carp_suppress_preempt--;
 411         sc->sc_sendad_errors = 0;
 412
 413         carp_set_state(sc, INIT);
 414         SC2IFP(sc)->if_flags &= ~IFF_UP;
 415         carp_setrun(sc, 0);
 416         if (unlock)
 417                 carp_multicast_cleanup(sc);
 418 #ifdef INET6
 419         carp_multicast6_cleanup(sc);
 420 #endif
 421
 422         if (sc->sc_carpdev != NULL) {
 423                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
 424                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
 425                 if (!--cif->vhif_nvrs) {
 426                         ifpromisc(sc->sc_carpdev, 0);
 427                         sc->sc_carpdev->if_carp = NULL;
 428                         kfree(cif, M_IFADDR);
 429                 }
 430                 sc->sc_carpdev = NULL;
 431         }
 432 }
 433
 434 /* Detach an interface from the carp. */
 435 static void
 436 carp_ifdetach(void *arg __unused, struct ifnet *ifp)
 437 {
 438         struct carp_if *cif = (struct carp_if *)ifp->if_carp;
 439         struct carp_softc *sc, *nextsc;
 440
 441         if (cif == NULL)
 442                 return;
 443
 444         for (sc = TAILQ_FIRST(&cif->vhif_vrs); sc; sc = nextsc) {
 445                 nextsc = TAILQ_NEXT(sc, sc_list);
 446                 carpdetach(sc, 0);
 447         }
 448 }
 449
 450 /*
 451  * process input packet.
 452  * we have rearranged checks order compared to the rfc,
 453  * but it seems more efficient this way or not possible otherwise.
 454  */
 455 void
 456 carp_input(struct mbuf *m, ...)
 457 {
 458         struct ip *ip = mtod(m, struct ip *);
 459         struct carp_header *ch;
 460         int len, iphlen;
 461         __va_list ap;
 462
 463         __va_start(ap, m);
 464         iphlen = __va_arg(ap, int);
 465         __va_end(ap);
 466
 467         carpstats.carps_ipackets++;
 468
 469         if (!carp_opts[CARPCTL_ALLOW]) {
 470                 m_freem(m);
 471                 return;
 472         }
 473
 474         /* Check if received on a valid carp interface */
 475         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 476                 carpstats.carps_badif++;
 477                 CARP_LOG("carp_input: packet received on non-carp "
 478                     "interface: %s\n",
 479                     m->m_pkthdr.rcvif->if_xname);
 480                 m_freem(m);
 481                 return;
 482         }
 483
 484         /* Verify that the IP TTL is CARP_DFLTTL. */
 485         if (ip->ip_ttl != CARP_DFLTTL) {
 486                 carpstats.carps_badttl++;
 487                 CARP_LOG("carp_input: received ttl %d != %d on %s\n",
 488                     ip->ip_ttl, CARP_DFLTTL,
 489                     m->m_pkthdr.rcvif->if_xname);
 490                 m_freem(m);
 491                 return;
 492         }
 493
 494         /* Minimal CARP packet size */
 495         len = iphlen + sizeof(*ch);
 496
 497         /*
 498          * Verify that the received packet length is
 499          * not less than the CARP header
 500          */
 501         if (m->m_pkthdr.len < len) {
 502                 carpstats.carps_badlen++;
 503                 CARP_LOG("packet too short %d on %s\n", m->m_pkthdr.len,
 504                          m->m_pkthdr.rcvif->if_xname);
 505                 m_freem(m);
 506                 return;
 507         }
 508
 509         /* Make sure that CARP header is contiguous */
 510         if (len > m->m_len) {
 511                 m = m_pullup(m, len);
 512                 if (m == NULL) {
 513                         carpstats.carps_hdrops++;
 514                         CARP_LOG("carp_input: m_pullup failed\n");
 515                         return;
 516                 }
 517                 ip = mtod(m, struct ip *);
 518         }
 519         ch = (struct carp_header *)((uint8_t *)ip + iphlen);
 520
 521         /* Verify the CARP checksum */
 522         if (in_cksum_skip(m, len, iphlen)) {
 523                 carpstats.carps_badsum++;
 524                 CARP_LOG("carp_input: checksum failed on %s\n",
 525                     m->m_pkthdr.rcvif->if_xname);
 526                 m_freem(m);
 527                 return;
 528         }
 529         carp_input_c(m, ch, AF_INET);
 530 }
 531
 532 #ifdef INET6
 533 int
 534 carp6_input(struct mbuf **mp, int *offp, int proto)
 535 {
 536         struct mbuf *m = *mp;
 537         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 538         struct carp_header *ch;
 539         u_int len;
 540
 541         carpstats.carps_ipackets6++;
 542
 543         if (!carp_opts[CARPCTL_ALLOW]) {
 544                 m_freem(m);
 545                 return (IPPROTO_DONE);
 546         }
 547
 548         /* check if received on a valid carp interface */
 549         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 550                 carpstats.carps_badif++;
 551                 CARP_LOG("carp6_input: packet received on non-carp "
 552                     "interface: %s\n",
 553                     m->m_pkthdr.rcvif->if_xname);
 554                 m_freem(m);
 555                 return (IPPROTO_DONE);
 556         }
 557
 558         /* verify that the IP TTL is 255 */
 559         if (ip6->ip6_hlim != CARP_DFLTTL) {
 560                 carpstats.carps_badttl++;
 561                 CARP_LOG("carp6_input: received ttl %d != 255 on %s\n",
 562                     ip6->ip6_hlim,
 563                     m->m_pkthdr.rcvif->if_xname);
 564                 m_freem(m);
 565                 return (IPPROTO_DONE);
 566         }
 567
 568         /* verify that we have a complete carp packet */
 569         len = m->m_len;
 570         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 571         if (ch == NULL) {
 572                 carpstats.carps_badlen++;
 573                 CARP_LOG("carp6_input: packet size %u too small\n", len);
 574                 return (IPPROTO_DONE);
 575         }
 576
 577         /* verify the CARP checksum */
 578         if (in_cksum_range(m, 0, *offp, sizeof(*ch))) {
 579                 carpstats.carps_badsum++;
 580                 CARP_LOG("carp6_input: checksum failed, on %s\n",
 581                     m->m_pkthdr.rcvif->if_xname);
 582                 m_freem(m);
 583                 return (IPPROTO_DONE);
 584         }
 585
 586         carp_input_c(m, ch, AF_INET6);
 587         return (IPPROTO_DONE);
 588 }
 589 #endif /* INET6 */
 590
 591 static void
 592 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 593 {
 594         struct ifnet *ifp = m->m_pkthdr.rcvif;
 595         struct carp_softc *sc;
 596         uint64_t tmp_counter;
 597         struct timeval sc_tv, ch_tv;
 598
 599         /* verify that the VHID is valid on the receiving interface */
 600         TAILQ_FOREACH(sc, &((struct carp_if *)ifp->if_carp)->vhif_vrs, sc_list)
 601                 if (sc->sc_vhid == ch->carp_vhid)
 602                         break;
 603
 604         if (!sc || !CARP_IS_RUNNING(SC2IFP(sc))) {
 605                 carpstats.carps_badvhid++;
 606                 m_freem(m);
 607                 return;
 608         }
 609
 610         getmicrotime(&SC2IFP(sc)->if_lastchange);
 611         SC2IFP(sc)->if_ipackets++;
 612         SC2IFP(sc)->if_ibytes += m->m_pkthdr.len;
 613
 614         if (SC2IFP(sc)->if_bpf) {
 615                 struct ip *ip = mtod(m, struct ip *);
 616
 617                 /* BPF wants net byte order */
 618                 ip->ip_len = htons(ip->ip_len + (ip->ip_hl << 2));
 619                 ip->ip_off = htons(ip->ip_off);
 620                 bpf_mtap(SC2IFP(sc)->if_bpf, m);
 621         }
 622
 623         /* verify the CARP version. */
 624         if (ch->carp_version != CARP_VERSION) {
 625                 carpstats.carps_badver++;
 626                 SC2IFP(sc)->if_ierrors++;
 627                 CARP_LOG("%s; invalid version %d\n",
 628                     SC2IFP(sc)->if_xname,
 629                     ch->carp_version);
 630                 m_freem(m);
 631                 return;
 632         }
 633
 634         /* verify the hash */
 635         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 636                 carpstats.carps_badauth++;
 637                 SC2IFP(sc)->if_ierrors++;
 638                 CARP_LOG("%s: incorrect hash\n", SC2IFP(sc)->if_xname);
 639                 m_freem(m);
 640                 return;
 641         }
 642
 643         tmp_counter = ntohl(ch->carp_counter[0]);
 644         tmp_counter = tmp_counter<<32;
 645         tmp_counter += ntohl(ch->carp_counter[1]);
 646
 647         /* XXX Replay protection goes here */
 648
 649         sc->sc_init_counter = 0;
 650         sc->sc_counter = tmp_counter;
 651
 652         sc_tv.tv_sec = sc->sc_advbase;
 653         if (carp_suppress_preempt && sc->sc_advskew <  240)
 654                 sc_tv.tv_usec = 240 * 1000000 / 256;
 655         else
 656                 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256;
 657         ch_tv.tv_sec = ch->carp_advbase;
 658         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 659
 660         switch (sc->sc_state) {
 661         case INIT:
 662                 break;
 663
 664         case MASTER:
 665                 /*
 666                  * If we receive an advertisement from a master who's going to
 667                  * be more frequent than us, go into BACKUP state.
 668                  */
 669                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 670                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 671                         callout_stop(&sc->sc_ad_tmo);
 672                         CARP_DEBUG("%s: MASTER -> BACKUP "
 673                            "(more frequent advertisement received)\n",
 674                            SC2IFP(sc)->if_xname);
 675                         carp_set_state(sc, BACKUP);
 676                         carp_setrun(sc, 0);
 677                         carp_setroute(sc, RTM_DELETE);
 678                 }
 679                 break;
 680
 681         case BACKUP:
 682                 /*
 683                  * If we're pre-empting masters who advertise slower than us,
 684                  * and this one claims to be slower, treat him as down.
 685                  */
 686                 if (carp_opts[CARPCTL_PREEMPT] &&
 687                     timevalcmp(&sc_tv, &ch_tv, <)) {
 688                         CARP_DEBUG("%s: BACKUP -> MASTER "
 689                             "(preempting a slower master)\n",
 690                             SC2IFP(sc)->if_xname);
 691                         carp_master_down_locked(sc);
 692                         break;
 693                 }
 694
 695                 /*
 696                  *  If the master is going to advertise at such a low frequency
 697                  *  that he's guaranteed to time out, we'd might as well just
 698                  *  treat him as timed out now.
 699                  */
 700                 sc_tv.tv_sec = sc->sc_advbase * 3;
 701                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 702                         CARP_DEBUG("%s: BACKUP -> MASTER "
 703                             "(master timed out)\n",
 704                             SC2IFP(sc)->if_xname);
 705                         carp_master_down_locked(sc);
 706                         break;
 707                 }
 708
 709                 /*
 710                  * Otherwise, we reset the counter and wait for the next
 711                  * advertisement.
 712                  */
 713                 carp_setrun(sc, af);
 714                 break;
 715         }
 716         m_freem(m);
 717 }
 718
 719 static int
 720 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 721 {
 722         struct m_tag *mtag;
 723         struct ifnet *ifp = SC2IFP(sc);
 724
 725         if (sc->sc_init_counter) {
 726                 /* this could also be seconds since unix epoch */
 727                 sc->sc_counter = karc4random();
 728                 sc->sc_counter = sc->sc_counter << 32;
 729                 sc->sc_counter += karc4random();
 730         } else {
 731                 sc->sc_counter++;
 732         }
 733
 734         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 735         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 736
 737         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 738
 739         /* Tag packet for carp_output */
 740         mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct ifnet *), MB_DONTWAIT);
 741         if (mtag == NULL) {
 742                 m_freem(m);
 743                 SC2IFP(sc)->if_oerrors++;
 744                 return (ENOMEM);
 745         }
 746         bcopy(&ifp, (caddr_t)(mtag + 1), sizeof(struct ifnet *));
 747         m_tag_prepend(m, mtag);
 748
 749         return (0);
 750 }
 751
 752 static void
 753 carp_send_ad_all(void)
 754 {
 755         struct carp_softc *sc;
 756
 757         LIST_FOREACH(sc, &carpif_list, sc_next) {
 758                 if (sc->sc_carpdev == NULL)
 759                         continue;
 760
 761                 if (CARP_IS_RUNNING(SC2IFP(sc)) && sc->sc_state == MASTER)
 762                         carp_send_ad_locked(sc);
 763         }
 764 }
 765
 766 static void
 767 carp_send_ad(void *v)
 768 {
 769         struct carp_softc *sc = v;
 770
 771         carp_send_ad_locked(sc);
 772 }
 773
 774 static void
 775 carp_send_ad_locked(struct carp_softc *sc)
 776 {
 777         struct carp_header ch;
 778         struct timeval tv;
 779         struct carp_header *ch_ptr;
 780         struct mbuf *m;
 781         int len, advbase, advskew;
 782
 783         /* bow out if we've lost our UPness or RUNNINGuiness */
 784         if (!CARP_IS_RUNNING(SC2IFP(sc))) {
 785                 advbase = 255;
 786                 advskew = 255;
 787         } else {
 788                 advbase = sc->sc_advbase;
 789                 if (!carp_suppress_preempt || sc->sc_advskew > 240)
 790                         advskew = sc->sc_advskew;
 791                 else
 792                         advskew = 240;
 793                 tv.tv_sec = advbase;
 794                 tv.tv_usec = advskew * 1000000 / 256;
 795         }
 796
 797         ch.carp_version = CARP_VERSION;
 798         ch.carp_type = CARP_ADVERTISEMENT;
 799         ch.carp_vhid = sc->sc_vhid;
 800         ch.carp_advbase = advbase;
 801         ch.carp_advskew = advskew;
 802         ch.carp_authlen = 7;    /* XXX DEFINE */
 803         ch.carp_pad1 = 0;       /* must be zero */
 804         ch.carp_cksum = 0;
 805
 806 #ifdef INET
 807         if (sc->sc_ia) {
 808                 struct ip *ip;
 809
 810                 MGETHDR(m, M_NOWAIT, MT_HEADER);
 811                 if (m == NULL) {
 812                         SC2IFP(sc)->if_oerrors++;
 813                         carpstats.carps_onomem++;
 814                         /* XXX maybe less ? */
 815                         if (advbase != 255 || advskew != 255)
 816                                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
 817                                     carp_send_ad, sc);
 818                         return;
 819                 }
 820                 len = sizeof(*ip) + sizeof(ch);
 821                 m->m_pkthdr.len = len;
 822                 m->m_pkthdr.rcvif = NULL;
 823                 m->m_len = len;
 824                 MH_ALIGN(m, m->m_len);
 825                 m->m_flags |= M_MCAST;
 826                 ip = mtod(m, struct ip *);
 827                 ip->ip_v = IPVERSION;
 828                 ip->ip_hl = sizeof(*ip) >> 2;
 829                 ip->ip_tos = IPTOS_LOWDELAY;
 830                 ip->ip_len = len;
 831                 ip->ip_id = ip_newid();
 832                 ip->ip_off = IP_DF;
 833                 ip->ip_ttl = CARP_DFLTTL;
 834                 ip->ip_p = IPPROTO_CARP;
 835                 ip->ip_sum = 0;
 836                 ip->ip_src.s_addr = sc->sc_ia->ia_addr.sin_addr.s_addr;
 837                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 838
 839                 ch_ptr = (struct carp_header *)(&ip[1]);
 840                 bcopy(&ch, ch_ptr, sizeof(ch));
 841                 if (carp_prepare_ad(m, sc, ch_ptr))
 842                         return;
 843                 ch_ptr->carp_cksum = in_cksum_skip(m, len, sizeof(*ip));
 844
 845                 getmicrotime(&SC2IFP(sc)->if_lastchange);
 846                 SC2IFP(sc)->if_opackets++;
 847                 SC2IFP(sc)->if_obytes += len;
 848                 carpstats.carps_opackets++;
 849
 850                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, NULL)) {
 851                         SC2IFP(sc)->if_oerrors++;
 852                         if (sc->sc_sendad_errors < INT_MAX)
 853                                 sc->sc_sendad_errors++;
 854                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 855                                 carp_suppress_preempt++;
 856                                 if (carp_suppress_preempt == 1) {
 857                                         carp_send_ad_all();
 858                                 }
 859                         }
 860                         sc->sc_sendad_success = 0;
 861                 } else {
 862                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 863                                 if (++sc->sc_sendad_success >=
 864                                     CARP_SENDAD_MIN_SUCCESS) {
 865                                         carp_suppress_preempt--;
 866                                         sc->sc_sendad_errors = 0;
 867                                 }
 868                         } else {
 869                                 sc->sc_sendad_errors = 0;
 870                         }
 871                 }
 872         }
 873 #endif /* INET */
 874 #ifdef INET6
 875         if (sc->sc_ia6) {
 876                 struct ip6_hdr *ip6;
 877
 878                 MGETHDR(m, M_NOWAIT, MT_HEADER);
 879                 if (m == NULL) {
 880                         SC2IFP(sc)->if_oerrors++;
 881                         carpstats.carps_onomem++;
 882                         /* XXX maybe less ? */
 883                         if (advbase != 255 || advskew != 255)
 884                                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
 885                                     carp_send_ad, sc);
 886                         return;
 887                 }
 888                 len = sizeof(*ip6) + sizeof(ch);
 889                 m->m_pkthdr.len = len;
 890                 m->m_pkthdr.rcvif = NULL;
 891                 m->m_len = len;
 892                 MH_ALIGN(m, m->m_len);
 893                 m->m_flags |= M_MCAST;
 894                 ip6 = mtod(m, struct ip6_hdr *);
 895                 bzero(ip6, sizeof(*ip6));
 896                 ip6->ip6_vfc |= IPV6_VERSION;
 897                 ip6->ip6_hlim = CARP_DFLTTL;
 898                 ip6->ip6_nxt = IPPROTO_CARP;
 899                 bcopy(&sc->sc_ia6->ia_addr.sin6_addr, &ip6->ip6_src,
 900                     sizeof(struct in6_addr));
 901                 /* set the multicast destination */
 902
 903                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
 904                 ip6->ip6_dst.s6_addr8[15] = 0x12;
 905                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
 906                         SC2IFP(sc)->if_oerrors++;
 907                         m_freem(m);
 908                         CARP_LOG("%s: in6_setscope failed\n", __func__);
 909                         return;
 910                 }
 911
 912                 ch_ptr = (struct carp_header *)(&ip6[1]);
 913                 bcopy(&ch, ch_ptr, sizeof(ch));
 914                 if (carp_prepare_ad(m, sc, ch_ptr))
 915                         return;
 916                 ch_ptr->carp_cksum = in_cksum_skip(m, len, sizeof(*ip6));
 917
 918                 getmicrotime(&SC2IFP(sc)->if_lastchange);
 919                 SC2IFP(sc)->if_opackets++;
 920                 SC2IFP(sc)->if_obytes += len;
 921                 carpstats.carps_opackets6++;
 922
 923                 if (ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL)) {
 924                         SC2IFP(sc)->if_oerrors++;
 925                         if (sc->sc_sendad_errors < INT_MAX)
 926                                 sc->sc_sendad_errors++;
 927                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 928                                 carp_suppress_preempt++;
 929                                 if (carp_suppress_preempt == 1) {
 930                                         carp_send_ad_all();
 931                                 }
 932                         }
 933                         sc->sc_sendad_success = 0;
 934                 } else {
 935                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 936                                 if (++sc->sc_sendad_success >=
 937                                     CARP_SENDAD_MIN_SUCCESS) {
 938                                         carp_suppress_preempt--;
 939                                         sc->sc_sendad_errors = 0;
 940                                 }
 941                         } else {
 942                                 sc->sc_sendad_errors = 0;
 943                         }
 944                 }
 945         }
 946 #endif /* INET6 */
 947
 948         if (advbase != 255 || advskew != 255)
 949                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
 950                     carp_send_ad, sc);
 951 }
 952
 953 /*
 954  * Broadcast a gratuitous ARP request containing
 955  * the virtual router MAC address for each IP address
 956  * associated with the virtual router.
 957  */
 958 static void
 959 carp_send_arp(struct carp_softc *sc)
 960 {
 961         struct ifaddr_container *ifac;
 962
 963         TAILQ_FOREACH(ifac, &SC2IFP(sc)->if_addrheads[mycpuid], ifa_link) {
 964                 struct ifaddr *ifa = ifac->ifa;
 965
 966                 if (ifa->ifa_addr->sa_family != AF_INET)
 967                         continue;
 968                 arp_ifinit2(sc->sc_carpdev, ifa, IF_LLADDR(sc->sc_ifp));
 969
 970                 DELAY(1000);    /* XXX */
 971         }
 972 }
 973
 974 #ifdef INET6
 975 static void
 976 carp_send_na(struct carp_softc *sc)
 977 {
 978         struct ifaddr_container *ifac;
 979         struct in6_addr *in6;
 980         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
 981
 982         TAILQ_FOREACH(ifac, &SC2IFP(sc)->if_addrheads[mycpuid], ifa_link) {
 983                 struct ifaddr *ifa = ifac->ifa;
 984
 985                 if (ifa->ifa_addr->sa_family != AF_INET6)
 986                         continue;
 987
 988                 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr;
 989                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
 990                     ND_NA_FLAG_OVERRIDE, 1, NULL);
 991                 DELAY(1000);    /* XXX */
 992         }
 993 }
 994 #endif /* INET6 */
 995
 996 static int
 997 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type)
 998 {
 999         struct carp_softc *vh;
1000         int count = 0;
1001
1002         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1003                 if ((type == CARP_COUNT_RUNNING &&
1004                      CARP_IS_RUNNING(SC2IFP(vh))) ||
1005                     (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) {
1006                         struct ifaddr_container *ifac;
1007
1008                         TAILQ_FOREACH(ifac, &SC2IFP(vh)->if_addrheads[mycpuid],
1009                                       ifa_link) {
1010                                 struct ifaddr *ifa = ifac->ifa;
1011
1012                                 if (ifa->ifa_addr->sa_family == AF_INET &&
1013                                     ia->ia_addr.sin_addr.s_addr ==
1014                                     ifatoia(ifa)->ia_addr.sin_addr.s_addr)
1015                                         count++;
1016                         }
1017                 }
1018         }
1019         return (count);
1020 }
1021
1022 int
1023 carp_iamatch(void *v, struct in_ifaddr *ia,
1024     struct in_addr *isaddr, uint8_t **enaddr)
1025 {
1026         struct carp_if *cif = v;
1027         struct carp_softc *vh;
1028         int index, count = 0;
1029
1030         if (carp_opts[CARPCTL_ARPBALANCE]) {
1031                 /*
1032                  * XXX proof of concept implementation.
1033                  * We use the source ip to decide which virtual host should
1034                  * handle the request. If we're master of that virtual host,
1035                  * then we respond, otherwise, just drop the arp packet on
1036                  * the floor.
1037                  */
1038                 count = carp_addrcount(cif, ia, CARP_COUNT_RUNNING);
1039                 if (count == 0) {
1040                         /* should never reach this */
1041                         return (0);
1042                 }
1043
1044                 /* this should be a hash, like pf_hash() */
1045                 index = ntohl(isaddr->s_addr) % count;
1046                 count = 0;
1047
1048                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1049                         if (CARP_IS_RUNNING(SC2IFP(vh))) {
1050                                 struct ifaddr_container *ifac;
1051
1052                                 TAILQ_FOREACH(ifac,
1053                                 &SC2IFP(vh)->if_addrheads[mycpuid], ifa_link) {
1054                                         struct ifaddr *ifa = ifac->ifa;
1055
1056                                         if (ifa->ifa_addr->sa_family ==
1057                                             AF_INET &&
1058                                             ia->ia_addr.sin_addr.s_addr ==
1059                                             ifatoia(ifa)->ia_addr.sin_addr.s_addr) {
1060                                                 if (count == index) {
1061                                                         if (vh->sc_state == MASTER) {
1062                                                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1063                                                                 return (1);
1064                                                         } else {
1065                                                                 return (0);
1066                                                         }
1067                                                 }
1068                                                 count++;
1069                                         }
1070                                 }
1071                         }
1072                 }
1073         } else {
1074                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1075                         if (CARP_IS_RUNNING(SC2IFP(vh)) &&
1076                             vh->sc_state == MASTER) {
1077                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1078                                 return (1);
1079                         }
1080                 }
1081         }
1082         return(0);
1083 }
1084
1085 #ifdef INET6
1086 struct ifaddr *
1087 carp_iamatch6(void *v, struct in6_addr *taddr)
1088 {
1089         struct carp_if *cif = v;
1090         struct carp_softc *vh;
1091
1092         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1093                 struct ifaddr_container *ifac;
1094
1095                 TAILQ_FOREACH(ifac, &SC2IFP(vh)->if_addrheads[mycpuid],
1096                               ifa_link) {
1097                         struct ifaddr *ifa = ifac->ifa;
1098
1099                         if (IN6_ARE_ADDR_EQUAL(taddr,
1100                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1101                             CARP_IS_RUNNING(SC2IFP(vh)) &&
1102                             vh->sc_state == MASTER) {
1103                                 return (ifa);
1104                         }
1105                 }
1106         }
1107         return (NULL);
1108 }
1109
1110 void *
1111 carp_macmatch6(void *v, struct mbuf *m, const struct in6_addr *taddr)
1112 {
1113         struct m_tag *mtag;
1114         struct carp_if *cif = v;
1115         struct carp_softc *sc;
1116
1117         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) {
1118                 struct ifaddr_container *ifac;
1119
1120                 TAILQ_FOREACH(ifac, &SC2IFP(sc)->if_addrheads[mycpuid],
1121                               ifa_link) {
1122                         struct ifaddr *ifa = ifac->ifa;
1123
1124                         if (IN6_ARE_ADDR_EQUAL(taddr,
1125                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1126                             CARP_IS_RUNNING(SC2IFP(sc))) {
1127                                 struct ifnet *ifp = SC2IFP(sc);
1128                                 mtag = m_tag_get(PACKET_TAG_CARP,
1129                                     sizeof(struct ifnet *), MB_DONTWAIT);
1130                                 if (mtag == NULL) {
1131                                         /* better a bit than nothing */
1132                                         return (IF_LLADDR(sc->sc_ifp));
1133                                 }
1134                                 bcopy(&ifp, (caddr_t)(mtag + 1),
1135                                     sizeof(struct ifnet *));
1136                                 m_tag_prepend(m, mtag);
1137
1138                                 return (IF_LLADDR(sc->sc_ifp));
1139                         }
1140                 }
1141         }
1142         return (NULL);
1143 }
1144 #endif
1145
1146 struct ifnet *
1147 carp_forus(void *v, void *dhost)
1148 {
1149         struct carp_if *cif = v;
1150         struct carp_softc *vh;
1151         uint8_t *ena = dhost;
1152
1153         /**
1154          * XXX: See here for check on MAC adr is not for virtual use
1155          *
1156          **/
1157
1158         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1159                 return (NULL);
1160
1161         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1162                 if (CARP_IS_RUNNING(SC2IFP(vh)) && vh->sc_state == MASTER &&
1163                     !bcmp(dhost, IF_LLADDR(vh->sc_ifp), ETHER_ADDR_LEN)) {
1164                         return (SC2IFP(vh));
1165                 }
1166         }
1167         return (NULL);
1168 }
1169
1170 static void
1171 carp_master_down(void *v)
1172 {
1173         struct carp_softc *sc = v;
1174
1175         lwkt_serialize_enter(sc->sc_ifp->if_serializer);
1176         carp_master_down_locked(sc);
1177         lwkt_serialize_exit(sc->sc_ifp->if_serializer);
1178 }
1179
1180 static void
1181 carp_master_down_locked(struct carp_softc *sc)
1182 {
1183         switch (sc->sc_state) {
1184         case INIT:
1185                 kprintf("%s: master_down event in INIT state\n",
1186                     SC2IFP(sc)->if_xname);
1187                 break;
1188
1189         case MASTER:
1190                 break;
1191
1192         case BACKUP:
1193                 carp_set_state(sc, MASTER);
1194                 carp_send_ad_locked(sc);
1195                 carp_send_arp(sc);
1196 #ifdef INET6
1197                 carp_send_na(sc);
1198 #endif /* INET6 */
1199                 carp_setrun(sc, 0);
1200                 carp_setroute(sc, RTM_ADD);
1201                 break;
1202         }
1203 }
1204
1205 /*
1206  * When in backup state, af indicates whether to reset the master down timer
1207  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1208  */
1209 static void
1210 carp_setrun(struct carp_softc *sc, sa_family_t af)
1211 {
1212         struct timeval tv;
1213
1214         if (sc->sc_carpdev == NULL) {
1215                 SC2IFP(sc)->if_flags &= ~IFF_RUNNING;
1216                 carp_set_state(sc, INIT);
1217                 return;
1218         }
1219
1220         if (SC2IFP(sc)->if_flags & IFF_UP &&
1221             sc->sc_vhid > 0 && (sc->sc_naddrs || sc->sc_naddrs6)) {
1222                 SC2IFP(sc)->if_flags |= IFF_RUNNING;
1223         } else {
1224                 SC2IFP(sc)->if_flags &= ~IFF_RUNNING;
1225                 carp_setroute(sc, RTM_DELETE);
1226                 return;
1227         }
1228
1229         switch (sc->sc_state) {
1230         case INIT:
1231                 if (carp_opts[CARPCTL_PREEMPT] && !carp_suppress_preempt) {
1232                         carp_send_ad_locked(sc);
1233                         carp_send_arp(sc);
1234 #ifdef INET6
1235                         carp_send_na(sc);
1236 #endif /* INET6 */
1237                         CARP_DEBUG("%s: INIT -> MASTER (preempting)\n",
1238                                    SC2IFP(sc)->if_xname);
1239                         carp_set_state(sc, MASTER);
1240                         carp_setroute(sc, RTM_ADD);
1241                 } else {
1242                         CARP_DEBUG("%s: INIT -> BACKUP\n",
1243                                    SC2IFP(sc)->if_xname);
1244                         carp_set_state(sc, BACKUP);
1245                         carp_setroute(sc, RTM_DELETE);
1246                         carp_setrun(sc, 0);
1247                 }
1248                 break;
1249
1250         case BACKUP:
1251                 callout_stop(&sc->sc_ad_tmo);
1252                 tv.tv_sec = 3 * sc->sc_advbase;
1253                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1254                 switch (af) {
1255 #ifdef INET
1256                 case AF_INET:
1257                         callout_reset(&sc->sc_md_tmo, tvtohz_high(&tv),
1258                             carp_master_down, sc);
1259                         break;
1260 #endif /* INET */
1261 #ifdef INET6
1262                 case AF_INET6:
1263                         callout_reset(&sc->sc_md6_tmo, tvtohz_high(&tv),
1264                             carp_master_down, sc);
1265                         break;
1266 #endif /* INET6 */
1267                 default:
1268                         if (sc->sc_naddrs)
1269                                 callout_reset(&sc->sc_md_tmo, tvtohz_high(&tv),
1270                                     carp_master_down, sc);
1271                         if (sc->sc_naddrs6)
1272                                 callout_reset(&sc->sc_md6_tmo, tvtohz_high(&tv),
1273                                     carp_master_down, sc);
1274                         break;
1275                 }
1276                 break;
1277
1278         case MASTER:
1279                 tv.tv_sec = sc->sc_advbase;
1280                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1281                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
1282                     carp_send_ad, sc);
1283                 break;
1284         }
1285 }
1286
1287 static void
1288 carp_multicast_cleanup(struct carp_softc *sc)
1289 {
1290         struct ip_moptions *imo = &sc->sc_imo;
1291         uint16_t n = imo->imo_num_memberships;
1292
1293         /* Clean up our own multicast memberships */
1294         while (n-- > 0) {
1295                 if (imo->imo_membership[n] != NULL) {
1296                         in_delmulti(imo->imo_membership[n]);
1297                         imo->imo_membership[n] = NULL;
1298                 }
1299         }
1300         imo->imo_num_memberships = 0;
1301         imo->imo_multicast_ifp = NULL;
1302 }
1303
1304 #ifdef INET6
1305 static void
1306 carp_multicast6_cleanup(struct carp_softc *sc)
1307 {
1308         struct ip6_moptions *im6o = &sc->sc_im6o;
1309
1310         while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1311                 struct in6_multi_mship *imm =
1312                     LIST_FIRST(&im6o->im6o_memberships);
1313
1314                 LIST_REMOVE(imm, i6mm_chain);
1315                 in6_leavegroup(imm);
1316         }
1317         im6o->im6o_multicast_ifp = NULL;
1318 }
1319 #endif
1320
1321 static int
1322 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1323 {
1324         struct ifnet *ifp;
1325         struct carp_if *cif;
1326         struct in_ifaddr *ia, *ia_if;
1327         struct in_ifaddr_container *iac;
1328         struct ip_moptions *imo = &sc->sc_imo;
1329         struct in_addr addr;
1330         u_long iaddr = htonl(sin->sin_addr.s_addr);
1331         int own, error;
1332
1333         if (sin->sin_addr.s_addr == 0) {
1334                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1335                         carp_set_state(sc, INIT);
1336                 if (sc->sc_naddrs)
1337                         SC2IFP(sc)->if_flags |= IFF_UP;
1338                 carp_setrun(sc, 0);
1339                 return (0);
1340         }
1341         /* we have to do it by hands to check we won't match on us */
1342         ia_if = NULL; own = 0;
1343         TAILQ_FOREACH(iac, &in_ifaddrheads[mycpuid], ia_link) {
1344                 ia = iac->ia;
1345
1346                 /* and, yeah, we need a multicast-capable iface too */
1347                 if (ia->ia_ifp != SC2IFP(sc) &&
1348                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1349                     (iaddr & ia->ia_subnetmask) == ia->ia_subnet) {
1350                         if (!ia_if)
1351                                 ia_if = ia;
1352                         if (sin->sin_addr.s_addr ==
1353                             ia->ia_addr.sin_addr.s_addr)
1354                                 own++;
1355                 }
1356         }
1357
1358         if (!ia_if)
1359                 return (EADDRNOTAVAIL);
1360
1361         ia = ia_if;
1362         ifp = ia->ia_ifp;
1363
1364         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1365             (imo->imo_multicast_ifp && imo->imo_multicast_ifp != ifp))
1366                 return (EADDRNOTAVAIL);
1367
1368         if (imo->imo_num_memberships == 0) {
1369                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1370                 if ((imo->imo_membership[0] = in_addmulti(&addr, ifp)) == NULL)
1371                         return (ENOBUFS);
1372                 imo->imo_num_memberships++;
1373                 imo->imo_multicast_ifp = ifp;
1374                 imo->imo_multicast_ttl = CARP_DFLTTL;
1375                 imo->imo_multicast_loop = 0;
1376         }
1377
1378         if (!ifp->if_carp) {
1379                 cif = kmalloc(sizeof(*cif), M_CARP, M_WAITOK | M_ZERO);
1380                 if ((error = ifpromisc(ifp, 1))) {
1381                         kfree(cif, M_CARP);
1382                         goto cleanup;
1383                 }
1384
1385                 cif->vhif_ifp = ifp;
1386                 TAILQ_INIT(&cif->vhif_vrs);
1387                 ifp->if_carp = cif;
1388         } else {
1389                 struct carp_softc *vr;
1390
1391                 cif = (struct carp_if *)ifp->if_carp;
1392                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1393                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1394                                 error = EINVAL;
1395                                 goto cleanup;
1396                         }
1397                 }
1398         }
1399         sc->sc_ia = ia;
1400         sc->sc_carpdev = ifp;
1401
1402         { /* XXX prevent endless loop if already in queue */
1403         struct carp_softc *vr, *after = NULL;
1404         int myself = 0;
1405         cif = (struct carp_if *)ifp->if_carp;
1406
1407         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1408                 if (vr == sc)
1409                         myself = 1;
1410                 if (vr->sc_vhid < sc->sc_vhid)
1411                         after = vr;
1412         }
1413
1414         if (!myself) {
1415                 /* We're trying to keep things in order */
1416                 if (after == NULL)
1417                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1418                 else
1419                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1420                 cif->vhif_nvrs++;
1421         }
1422         }
1423
1424         sc->sc_naddrs++;
1425         SC2IFP(sc)->if_flags |= IFF_UP;
1426         if (own)
1427                 sc->sc_advskew = 0;
1428
1429         carp_sc_state_locked(sc);
1430         carp_setrun(sc, 0);
1431
1432         return (0);
1433
1434 cleanup:
1435         in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1436         return (error);
1437 }
1438
1439 static int
1440 carp_del_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1441 {
1442         int error = 0;
1443
1444         if (!--sc->sc_naddrs) {
1445                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1446                 struct ip_moptions *imo = &sc->sc_imo;
1447
1448                 callout_stop(&sc->sc_ad_tmo);
1449                 SC2IFP(sc)->if_flags &= ~(IFF_UP | IFF_RUNNING);
1450                 sc->sc_vhid = -1;
1451                 in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1452                 imo->imo_multicast_ifp = NULL;
1453                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1454                 if (!--cif->vhif_nvrs) {
1455                         sc->sc_carpdev->if_carp = NULL;
1456                         kfree(cif, M_IFADDR);
1457                 }
1458         }
1459         return (error);
1460 }
1461
1462 #ifdef INET6
1463 static int
1464 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1465 {
1466         struct ifnet *ifp;
1467         struct carp_if *cif;
1468         struct in6_ifaddr *ia, *ia_if;
1469         struct ip6_moptions *im6o = &sc->sc_im6o;
1470         struct in6_multi_mship *imm;
1471         struct in6_addr in6;
1472         int own, error;
1473
1474         if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
1475                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1476                         carp_set_state(sc, INIT);
1477                 if (sc->sc_naddrs6)
1478                         SC2IFP(sc)->if_flags |= IFF_UP;
1479                 carp_setrun(sc, 0);
1480                 return (0);
1481         }
1482
1483         /* we have to do it by hands to check we won't match on us */
1484         ia_if = NULL; own = 0;
1485         for (ia = in6_ifaddr; ia; ia = ia->ia_next) {
1486                 int i;
1487
1488                 for (i = 0; i < 4; i++) {
1489                         if ((sin6->sin6_addr.s6_addr32[i] &
1490                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]) !=
1491                             (ia->ia_addr.sin6_addr.s6_addr32[i] &
1492                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]))
1493                                 break;
1494                 }
1495                 /* and, yeah, we need a multicast-capable iface too */
1496                 if (ia->ia_ifp != SC2IFP(sc) &&
1497                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1498                     (i == 4)) {
1499                         if (!ia_if)
1500                                 ia_if = ia;
1501                         if (IN6_ARE_ADDR_EQUAL(&sin6->sin6_addr,
1502                             &ia->ia_addr.sin6_addr))
1503                                 own++;
1504                 }
1505         }
1506
1507         if (!ia_if)
1508                 return (EADDRNOTAVAIL);
1509         ia = ia_if;
1510         ifp = ia->ia_ifp;
1511
1512         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1513             (im6o->im6o_multicast_ifp && im6o->im6o_multicast_ifp != ifp))
1514                 return (EADDRNOTAVAIL);
1515
1516         if (!sc->sc_naddrs6) {
1517                 im6o->im6o_multicast_ifp = ifp;
1518
1519                 /* join CARP multicast address */
1520                 bzero(&in6, sizeof(in6));
1521                 in6.s6_addr16[0] = htons(0xff02);
1522                 in6.s6_addr8[15] = 0x12;
1523                 if (in6_setscope(&in6, ifp, NULL) != 0)
1524                         goto cleanup;
1525                 if ((imm = in6_joingroup(ifp, &in6, &error)) == NULL)
1526                         goto cleanup;
1527                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1528
1529                 /* join solicited multicast address */
1530                 bzero(&in6, sizeof(in6));
1531                 in6.s6_addr16[0] = htons(0xff02);
1532                 in6.s6_addr32[1] = 0;
1533                 in6.s6_addr32[2] = htonl(1);
1534                 in6.s6_addr32[3] = sin6->sin6_addr.s6_addr32[3];
1535                 in6.s6_addr8[12] = 0xff;
1536                 if (in6_setscope(&in6, ifp, NULL) != 0)
1537                         goto cleanup;
1538                 if ((imm = in6_joingroup(ifp, &in6, &error)) == NULL)
1539                         goto cleanup;
1540                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1541         }
1542
1543         if (!ifp->if_carp) {
1544                 cif = kmalloc(sizeof(*cif), M_CARP, M_WAITOK | M_ZERO);
1545                 if ((error = ifpromisc(ifp, 1))) {
1546                         kfree(cif, M_CARP);
1547                         goto cleanup;
1548                 }
1549
1550                 cif->vhif_ifp = ifp;
1551                 TAILQ_INIT(&cif->vhif_vrs);
1552                 ifp->if_carp = cif;
1553         } else {
1554                 struct carp_softc *vr;
1555
1556                 cif = (struct carp_if *)ifp->if_carp;
1557                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1558                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1559                                 error = EINVAL;
1560                                 goto cleanup;
1561                         }
1562                 }
1563         }
1564         sc->sc_ia6 = ia;
1565         sc->sc_carpdev = ifp;
1566
1567         { /* XXX prevent endless loop if already in queue */
1568         struct carp_softc *vr, *after = NULL;
1569         int myself = 0;
1570         cif = (struct carp_if *)ifp->if_carp;
1571
1572         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1573                 if (vr == sc)
1574                         myself = 1;
1575                 if (vr->sc_vhid < sc->sc_vhid)
1576                         after = vr;
1577         }
1578
1579         if (!myself) {
1580                 /* We're trying to keep things in order */
1581                 if (after == NULL)
1582                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1583                 else
1584                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1585                 cif->vhif_nvrs++;
1586         }
1587         }
1588
1589         sc->sc_naddrs6++;
1590         SC2IFP(sc)->if_flags |= IFF_UP;
1591         if (own)
1592                 sc->sc_advskew = 0;
1593         carp_sc_state_locked(sc);
1594         carp_setrun(sc, 0);
1595
1596         return (0);
1597
1598 cleanup:
1599         /* clean up multicast memberships */
1600         if (!sc->sc_naddrs6) {
1601                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1602                         imm = LIST_FIRST(&im6o->im6o_memberships);
1603                         LIST_REMOVE(imm, i6mm_chain);
1604                         in6_leavegroup(imm);
1605                 }
1606         }
1607         return (error);
1608 }
1609
1610 static int
1611 carp_del_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1612 {
1613         int error = 0;
1614
1615         if (!--sc->sc_naddrs6) {
1616                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1617                 struct ip6_moptions *im6o = &sc->sc_im6o;
1618
1619                 callout_stop(&sc->sc_ad_tmo);
1620                 SC2IFP(sc)->if_flags &= ~(IFF_UP | IFF_RUNNING);
1621                 sc->sc_vhid = -1;
1622                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1623                         struct in6_multi_mship *imm =
1624                             LIST_FIRST(&im6o->im6o_memberships);
1625
1626                         LIST_REMOVE(imm, i6mm_chain);
1627                         in6_leavegroup(imm);
1628                 }
1629                 im6o->im6o_multicast_ifp = NULL;
1630                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1631                 if (!--cif->vhif_nvrs) {
1632                         sc->sc_carpdev->if_carp = NULL;
1633                         kfree(cif, M_IFADDR);
1634                 }
1635         }
1636         return (error);
1637 }
1638 #endif /* INET6 */
1639
1640 static int
1641 carp_ioctl(struct ifnet *ifp, u_long cmd, caddr_t addr, struct ucred *cr)
1642 {
1643         struct carp_softc *sc = ifp->if_softc, *vr;
1644         struct carpreq carpr;
1645         struct ifaddr *ifa;
1646         struct ifreq *ifr;
1647         struct ifaliasreq *ifra;
1648         int error = 0;
1649
1650         ifa = (struct ifaddr *)addr;
1651         ifra = (struct ifaliasreq *)addr;
1652         ifr = (struct ifreq *)addr;
1653
1654         switch (cmd) {
1655         case SIOCSIFADDR:
1656                 switch (ifa->ifa_addr->sa_family) {
1657 #ifdef INET
1658                 case AF_INET:
1659                         SC2IFP(sc)->if_flags |= IFF_UP;
1660                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1661                             sizeof(struct sockaddr));
1662                         error = carp_set_addr(sc, satosin(ifa->ifa_addr));
1663                         break;
1664 #endif /* INET */
1665 #ifdef INET6
1666                 case AF_INET6:
1667                         SC2IFP(sc)->if_flags |= IFF_UP;
1668                         error = carp_set_addr6(sc, satosin6(ifa->ifa_addr));
1669                         break;
1670 #endif /* INET6 */
1671                 default:
1672                         error = EAFNOSUPPORT;
1673                         break;
1674                 }
1675                 break;
1676
1677         case SIOCAIFADDR:
1678                 switch (ifa->ifa_addr->sa_family) {
1679 #ifdef INET
1680                 case AF_INET:
1681                         SC2IFP(sc)->if_flags |= IFF_UP;
1682                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1683                             sizeof(struct sockaddr));
1684                         error = carp_set_addr(sc, satosin(&ifra->ifra_addr));
1685                         break;
1686 #endif /* INET */
1687 #ifdef INET6
1688                 case AF_INET6:
1689                         SC2IFP(sc)->if_flags |= IFF_UP;
1690                         error = carp_set_addr6(sc, satosin6(&ifra->ifra_addr));
1691                         break;
1692 #endif /* INET6 */
1693                 default:
1694                         error = EAFNOSUPPORT;
1695                         break;
1696                 }
1697                 break;
1698
1699         case SIOCDIFADDR:
1700                 switch (ifa->ifa_addr->sa_family) {
1701 #ifdef INET
1702                 case AF_INET:
1703                         error = carp_del_addr(sc, satosin(&ifra->ifra_addr));
1704                         break;
1705 #endif /* INET */
1706 #ifdef INET6
1707                 case AF_INET6:
1708                         error = carp_del_addr6(sc, satosin6(&ifra->ifra_addr));
1709                         break;
1710 #endif /* INET6 */
1711                 default:
1712                         error = EAFNOSUPPORT;
1713                         break;
1714                 }
1715                 break;
1716
1717         case SIOCSIFFLAGS:
1718                 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) {
1719                         callout_stop(&sc->sc_ad_tmo);
1720                         callout_stop(&sc->sc_md_tmo);
1721                         callout_stop(&sc->sc_md6_tmo);
1722                         if (sc->sc_state == MASTER)
1723                                 carp_send_ad_locked(sc);
1724                         carp_set_state(sc, INIT);
1725                         carp_setrun(sc, 0);
1726                 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) {
1727                         SC2IFP(sc)->if_flags |= IFF_UP;
1728                         carp_setrun(sc, 0);
1729                 }
1730                 break;
1731
1732         case SIOCSVH:
1733                 error = priv_check_cred(cr, PRIV_ROOT, NULL_CRED_OKAY);
1734                 if (error)
1735                         break;
1736                 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1737                         break;
1738                 error = 1;
1739                 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) {
1740                         switch (carpr.carpr_state) {
1741                         case BACKUP:
1742                                 callout_stop(&sc->sc_ad_tmo);
1743                                 carp_set_state(sc, BACKUP);
1744                                 carp_setrun(sc, 0);
1745                                 carp_setroute(sc, RTM_DELETE);
1746                                 break;
1747
1748                         case MASTER:
1749                                 carp_master_down_locked(sc);
1750                                 break;
1751
1752                         default:
1753                                 break;
1754                         }
1755                 }
1756                 if (carpr.carpr_vhid > 0) {
1757                         if (carpr.carpr_vhid > 255) {
1758                                 error = EINVAL;
1759                                 break;
1760                         }
1761                         if (sc->sc_carpdev) {
1762                                 struct carp_if *cif;
1763                                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1764                                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1765                                         if (vr != sc &&
1766                                             vr->sc_vhid == carpr.carpr_vhid)
1767                                                 return EEXIST;
1768                                 }
1769                         }
1770                         sc->sc_vhid = carpr.carpr_vhid;
1771                         IF_LLADDR(sc->sc_ifp)[0] = 0;
1772                         IF_LLADDR(sc->sc_ifp)[1] = 0;
1773                         IF_LLADDR(sc->sc_ifp)[2] = 0x5e;
1774                         IF_LLADDR(sc->sc_ifp)[3] = 0;
1775                         IF_LLADDR(sc->sc_ifp)[4] = 1;
1776                         IF_LLADDR(sc->sc_ifp)[5] = sc->sc_vhid;
1777                         error--;
1778                 }
1779                 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) {
1780                         if (carpr.carpr_advskew >= 255) {
1781                                 error = EINVAL;
1782                                 break;
1783                         }
1784                         if (carpr.carpr_advbase > 255) {
1785                                 error = EINVAL;
1786                                 break;
1787                         }
1788                         sc->sc_advbase = carpr.carpr_advbase;
1789                         sc->sc_advskew = carpr.carpr_advskew;
1790                         error--;
1791                 }
1792                 bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1793                 if (error > 0) {
1794                         error = EINVAL;
1795                 } else {
1796                         error = 0;
1797                         carp_setrun(sc, 0);
1798                 }
1799                 break;
1800
1801         case SIOCGVH:
1802                 bzero(&carpr, sizeof(carpr));
1803                 carpr.carpr_state = sc->sc_state;
1804                 carpr.carpr_vhid = sc->sc_vhid;
1805                 carpr.carpr_advbase = sc->sc_advbase;
1806                 carpr.carpr_advskew = sc->sc_advskew;
1807                 error = priv_check_cred(cr, PRIV_ROOT, NULL_CRED_OKAY);
1808                 if (error == 0) {
1809                         bcopy(sc->sc_key, carpr.carpr_key,
1810                             sizeof(carpr.carpr_key));
1811                 }
1812                 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1813                 break;
1814
1815         default:
1816                 error = EINVAL;
1817         }
1818         carp_hmac_prepare(sc);
1819         return (error);
1820 }
1821
1822 /*
1823  * XXX: this is looutput. We should eventually use it from there.
1824  */
1825 static int
1826 carp_looutput(struct ifnet *ifp, struct mbuf *m, struct sockaddr *dst,
1827     struct rtentry *rt)
1828 {
1829         uint32_t af;
1830
1831         M_ASSERTPKTHDR(m); /* check if we have the packet header */
1832
1833         if (rt && rt->rt_flags & (RTF_REJECT|RTF_BLACKHOLE)) {
1834                 m_freem(m);
1835                 return (rt->rt_flags & RTF_BLACKHOLE ? 0 :
1836                         rt->rt_flags & RTF_HOST ? EHOSTUNREACH : ENETUNREACH);
1837         }
1838
1839         ifp->if_opackets++;
1840         ifp->if_obytes += m->m_pkthdr.len;
1841
1842         /* BPF writes need to be handled specially. */
1843         if (dst->sa_family == AF_UNSPEC) {
1844                 bcopy(dst->sa_data, &af, sizeof(af));
1845                 dst->sa_family = af;
1846         }
1847
1848 #if 1   /* XXX */
1849         switch (dst->sa_family) {
1850         case AF_INET:
1851         case AF_INET6:
1852         case AF_IPX:
1853         case AF_APPLETALK:
1854                 break;
1855
1856         default:
1857                 m_freem(m);
1858                 return (EAFNOSUPPORT);
1859         }
1860 #endif
1861         return (if_simloop(ifp, m, dst->sa_family, 0));
1862 }
1863
1864 /*
1865  * Start output on carp interface. This function should never be called.
1866  */
1867 static void
1868 carp_start(struct ifnet *ifp)
1869 {
1870 #ifdef DEBUG
1871         kprintf("%s: start called\n", ifp->if_xname);
1872 #endif
1873 }
1874
1875 int
1876 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *sa,
1877     struct rtentry *rt)
1878 {
1879         struct m_tag *mtag;
1880         struct carp_softc *sc;
1881         struct ifnet *carp_ifp;
1882         struct ether_header *eh;
1883
1884         if (!sa)
1885                 return (0);
1886
1887         switch (sa->sa_family) {
1888 #ifdef INET
1889         case AF_INET:
1890                 break;
1891 #endif /* INET */
1892 #ifdef INET6
1893         case AF_INET6:
1894                 break;
1895 #endif /* INET6 */
1896         default:
1897                 return (0);
1898         }
1899
1900         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
1901         if (mtag == NULL)
1902                 return (0);
1903
1904         bcopy(mtag + 1, &carp_ifp, sizeof(struct ifnet *));
1905         sc = carp_ifp->if_softc;
1906
1907         /* Set the source MAC address to Virtual Router MAC Address */
1908         switch (ifp->if_type) {
1909         case IFT_ETHER:
1910         case IFT_L2VLAN:
1911                 eh = mtod(m, struct ether_header *);
1912                 eh->ether_shost[0] = 0;
1913                 eh->ether_shost[1] = 0;
1914                 eh->ether_shost[2] = 0x5e;
1915                 eh->ether_shost[3] = 0;
1916                 eh->ether_shost[4] = 1;
1917                 eh->ether_shost[5] = sc->sc_vhid;
1918                 break;
1919
1920         default:
1921                 if_printf(ifp, "carp is not supported for this "
1922                           "interface type\n");
1923                 return (EOPNOTSUPP);
1924         }
1925         return (0);
1926 }
1927
1928 static void
1929 carp_set_state(struct carp_softc *sc, int state)
1930 {
1931         if (sc->sc_state == state)
1932                 return;
1933
1934         sc->sc_state = state;
1935         switch (state) {
1936         case BACKUP:
1937                 SC2IFP(sc)->if_link_state = LINK_STATE_DOWN;
1938                 break;
1939
1940         case MASTER:
1941                 SC2IFP(sc)->if_link_state = LINK_STATE_UP;
1942                 break;
1943
1944         default:
1945                 SC2IFP(sc)->if_link_state = LINK_STATE_UNKNOWN;
1946                 break;
1947         }
1948         rt_ifmsg(SC2IFP(sc));
1949 }
1950
1951 void
1952 carp_carpdev_state(void *v)
1953 {
1954         struct carp_if *cif = v;
1955
1956         carp_carpdev_state_locked(cif);
1957 }
1958
1959 static void
1960 carp_carpdev_state_locked(struct carp_if *cif)
1961 {
1962         struct carp_softc *sc;
1963
1964         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list)
1965                 carp_sc_state_locked(sc);
1966 }
1967
1968 static void
1969 carp_sc_state_locked(struct carp_softc *sc)
1970 {
1971         if (!(sc->sc_carpdev->if_flags & IFF_UP)) {
1972                 sc->sc_flags_backup = SC2IFP(sc)->if_flags;
1973                 SC2IFP(sc)->if_flags &= ~(IFF_UP | IFF_RUNNING);
1974                 callout_stop(&sc->sc_ad_tmo);
1975                 callout_stop(&sc->sc_md_tmo);
1976                 callout_stop(&sc->sc_md6_tmo);
1977                 carp_set_state(sc, INIT);
1978                 carp_setrun(sc, 0);
1979                 if (!sc->sc_suppress) {
1980                         carp_suppress_preempt++;
1981                         if (carp_suppress_preempt == 1)
1982                                 carp_send_ad_all();
1983                 }
1984                 sc->sc_suppress = 1;
1985         } else {
1986                 SC2IFP(sc)->if_flags |= sc->sc_flags_backup;
1987                 carp_set_state(sc, INIT);
1988                 carp_setrun(sc, 0);
1989                 if (sc->sc_suppress)
1990                         carp_suppress_preempt--;
1991                 sc->sc_suppress = 0;
1992         }
1993 }
1994
1995 static int
1996 carp_modevent(module_t mod, int type, void *data)
1997 {
1998         switch (type) {
1999         case MOD_LOAD:
2000                 LIST_INIT(&carpif_list);
2001                 carp_ifdetach_event =
2002                 EVENTHANDLER_REGISTER(ifnet_detach_event, carp_ifdetach, NULL,
2003                                       EVENTHANDLER_PRI_ANY);
2004                 if_clone_attach(&carp_cloner);
2005                 break;
2006
2007         case MOD_UNLOAD:
2008                 EVENTHANDLER_DEREGISTER(ifnet_detach_event,
2009                                         carp_ifdetach_event);
2010                 if_clone_detach(&carp_cloner);
2011                 break;
2012
2013         default:
2014                 return (EINVAL);
2015         }
2016         return (0);
2017 }
2018
2019 static moduledata_t carp_mod = {
2020         "carp",
2021         carp_modevent,
2022         0
2023 };
2024 DECLARE_MODULE(carp, carp_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);