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/sockio.h>
  45 #include <sys/socket.h>
  46 #include <sys/sysctl.h>
  47 #include <sys/syslog.h>
  48
  49 #include <machine/stdarg.h>
  50 #include <crypto/sha1.h>
  51
  52 #include <net/bpf.h>
  53 #include <net/ethernet.h>
  54 #include <net/if.h>
  55 #include <net/if_dl.h>
  56 #include <net/if_types.h>
  57 #include <net/route.h>
  58 #include <net/if_clone.h>
  59
  60 #ifdef INET
  61 #include <netinet/in.h>
  62 #include <netinet/in_var.h>
  63 #include <netinet/in_systm.h>
  64 #include <netinet/ip.h>
  65 #include <netinet/ip_var.h>
  66 #include <netinet/if_ether.h>
  67 #endif
  68
  69 #ifdef INET6
  70 #include <netinet/icmp6.h>
  71 #include <netinet/ip6.h>
  72 #include <netinet6/ip6_var.h>
  73 #include <netinet6/scope6_var.h>
  74 #include <netinet6/nd6.h>
  75 #endif
  76
  77 #include <netinet/ip_carp.h>
  78
  79 #define CARP_IFNAME     "carp"
  80
  81 struct carp_softc {
  82         struct ifnet             sc_if;
  83         struct ifnet            *sc_ifp;        /* compat shim */
  84         struct ifnet            *sc_carpdev;    /* parent interface */
  85         struct in_ifaddr        *sc_ia;         /* primary iface address */
  86         struct ip_moptions       sc_imo;
  87 #ifdef INET6
  88         struct in6_ifaddr       *sc_ia6;        /* primary iface address v6 */
  89         struct ip6_moptions      sc_im6o;
  90 #endif /* INET6 */
  91         TAILQ_ENTRY(carp_softc)  sc_list;
  92
  93         enum { INIT = 0, BACKUP, MASTER }
  94                                  sc_state;
  95
  96         int                      sc_flags_backup;
  97         int                      sc_suppress;
  98
  99         int                      sc_sendad_errors;
 100 #define CARP_SENDAD_MAX_ERRORS  3
 101         int                      sc_sendad_success;
 102 #define CARP_SENDAD_MIN_SUCCESS 3
 103
 104         int                      sc_vhid;
 105         int                      sc_advskew;
 106         int                      sc_naddrs;
 107         int                      sc_naddrs6;
 108         int                      sc_advbase;    /* seconds */
 109         int                      sc_init_counter;
 110         uint64_t                 sc_counter;
 111
 112         /* authentication */
 113 #define CARP_HMAC_PAD   64
 114         unsigned char            sc_key[CARP_KEY_LEN];
 115         unsigned char            sc_pad[CARP_HMAC_PAD];
 116         SHA1_CTX                 sc_sha1;
 117
 118         struct callout           sc_ad_tmo;     /* advertisement timeout */
 119         struct callout           sc_md_tmo;     /* master down timeout */
 120         struct callout           sc_md6_tmo;    /* master down timeout */
 121
 122         LIST_ENTRY(carp_softc)   sc_next;       /* Interface clue */
 123 };
 124 #define SC2IFP(sc)      ((sc)->sc_ifp)
 125
 126 struct carp_if {
 127         TAILQ_HEAD(, carp_softc) vhif_vrs;
 128         int             vhif_nvrs;
 129
 130         struct ifnet    *vhif_ifp;
 131 };
 132
 133 enum    { CARP_COUNT_MASTER, CARP_COUNT_RUNNING };
 134
 135 SYSCTL_DECL(_net_inet_carp);
 136
 137 static int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 1, 0, 0 }; /* XXX for now */
 138 SYSCTL_INT(_net_inet_carp, CARPCTL_ALLOW, allow, CTLFLAG_RW,
 139     &carp_opts[CARPCTL_ALLOW], 0, "Accept incoming CARP packets");
 140 SYSCTL_INT(_net_inet_carp, CARPCTL_PREEMPT, preempt, CTLFLAG_RW,
 141     &carp_opts[CARPCTL_PREEMPT], 0, "high-priority backup preemption mode");
 142 SYSCTL_INT(_net_inet_carp, CARPCTL_LOG, log, CTLFLAG_RW,
 143     &carp_opts[CARPCTL_LOG], 0, "log bad carp packets");
 144 SYSCTL_INT(_net_inet_carp, CARPCTL_ARPBALANCE, arpbalance, CTLFLAG_RW,
 145     &carp_opts[CARPCTL_ARPBALANCE], 0, "balance arp responses");
 146
 147 static int carp_suppress_preempt = 0;
 148 SYSCTL_INT(_net_inet_carp, OID_AUTO, suppress_preempt, CTLFLAG_RD,
 149     &carp_suppress_preempt, 0, "Preemption is suppressed");
 150
 151 static struct carpstats carpstats;
 152 SYSCTL_STRUCT(_net_inet_carp, CARPCTL_STATS, stats, CTLFLAG_RW,
 153     &carpstats, carpstats,
 154     "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 155
 156 /* Get carp_if from softc. Valid after carp_set_addr{,6}. */
 157 #define SC2CIF(sc)              ((struct carp_if *)(sc)->sc_carpdev->if_carp)
 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                         FREE(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 ||
 605             !((SC2IFP(sc)->if_flags & IFF_UP) &&
 606               (SC2IFP(sc)->if_flags & IFF_RUNNING))) {
 607                 carpstats.carps_badvhid++;
 608                 m_freem(m);
 609                 return;
 610         }
 611
 612         getmicrotime(&SC2IFP(sc)->if_lastchange);
 613         SC2IFP(sc)->if_ipackets++;
 614         SC2IFP(sc)->if_ibytes += m->m_pkthdr.len;
 615
 616         if (SC2IFP(sc)->if_bpf) {
 617                 struct ip *ip = mtod(m, struct ip *);
 618
 619                 /* BPF wants net byte order */
 620                 ip->ip_len = htons(ip->ip_len + (ip->ip_hl << 2));
 621                 ip->ip_off = htons(ip->ip_off);
 622                 bpf_mtap(SC2IFP(sc)->if_bpf, m);
 623         }
 624
 625         /* verify the CARP version. */
 626         if (ch->carp_version != CARP_VERSION) {
 627                 carpstats.carps_badver++;
 628                 SC2IFP(sc)->if_ierrors++;
 629                 CARP_LOG("%s; invalid version %d\n",
 630                     SC2IFP(sc)->if_xname,
 631                     ch->carp_version);
 632                 m_freem(m);
 633                 return;
 634         }
 635
 636         /* verify the hash */
 637         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 638                 carpstats.carps_badauth++;
 639                 SC2IFP(sc)->if_ierrors++;
 640                 CARP_LOG("%s: incorrect hash\n", SC2IFP(sc)->if_xname);
 641                 m_freem(m);
 642                 return;
 643         }
 644
 645         tmp_counter = ntohl(ch->carp_counter[0]);
 646         tmp_counter = tmp_counter<<32;
 647         tmp_counter += ntohl(ch->carp_counter[1]);
 648
 649         /* XXX Replay protection goes here */
 650
 651         sc->sc_init_counter = 0;
 652         sc->sc_counter = tmp_counter;
 653
 654         sc_tv.tv_sec = sc->sc_advbase;
 655         if (carp_suppress_preempt && sc->sc_advskew <  240)
 656                 sc_tv.tv_usec = 240 * 1000000 / 256;
 657         else
 658                 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256;
 659         ch_tv.tv_sec = ch->carp_advbase;
 660         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 661
 662         switch (sc->sc_state) {
 663         case INIT:
 664                 break;
 665
 666         case MASTER:
 667                 /*
 668                  * If we receive an advertisement from a master who's going to
 669                  * be more frequent than us, go into BACKUP state.
 670                  */
 671                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 672                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 673                         callout_stop(&sc->sc_ad_tmo);
 674                         CARP_DEBUG("%s: MASTER -> BACKUP "
 675                            "(more frequent advertisement received)\n",
 676                            SC2IFP(sc)->if_xname);
 677                         carp_set_state(sc, BACKUP);
 678                         carp_setrun(sc, 0);
 679                         carp_setroute(sc, RTM_DELETE);
 680                 }
 681                 break;
 682
 683         case BACKUP:
 684                 /*
 685                  * If we're pre-empting masters who advertise slower than us,
 686                  * and this one claims to be slower, treat him as down.
 687                  */
 688                 if (carp_opts[CARPCTL_PREEMPT] &&
 689                     timevalcmp(&sc_tv, &ch_tv, <)) {
 690                         CARP_DEBUG("%s: BACKUP -> MASTER "
 691                             "(preempting a slower master)\n",
 692                             SC2IFP(sc)->if_xname);
 693                         carp_master_down_locked(sc);
 694                         break;
 695                 }
 696
 697                 /*
 698                  *  If the master is going to advertise at such a low frequency
 699                  *  that he's guaranteed to time out, we'd might as well just
 700                  *  treat him as timed out now.
 701                  */
 702                 sc_tv.tv_sec = sc->sc_advbase * 3;
 703                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 704                         CARP_DEBUG("%s: BACKUP -> MASTER "
 705                             "(master timed out)\n",
 706                             SC2IFP(sc)->if_xname);
 707                         carp_master_down_locked(sc);
 708                         break;
 709                 }
 710
 711                 /*
 712                  * Otherwise, we reset the counter and wait for the next
 713                  * advertisement.
 714                  */
 715                 carp_setrun(sc, af);
 716                 break;
 717         }
 718         m_freem(m);
 719 }
 720
 721 static int
 722 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 723 {
 724         struct m_tag *mtag;
 725         struct ifnet *ifp = SC2IFP(sc);
 726
 727         if (sc->sc_init_counter) {
 728                 /* this could also be seconds since unix epoch */
 729                 sc->sc_counter = karc4random();
 730                 sc->sc_counter = sc->sc_counter << 32;
 731                 sc->sc_counter += karc4random();
 732         } else {
 733                 sc->sc_counter++;
 734         }
 735
 736         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 737         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 738
 739         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 740
 741         /* Tag packet for carp_output */
 742         mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct ifnet *), MB_DONTWAIT);
 743         if (mtag == NULL) {
 744                 m_freem(m);
 745                 SC2IFP(sc)->if_oerrors++;
 746                 return (ENOMEM);
 747         }
 748         bcopy(&ifp, (caddr_t)(mtag + 1), sizeof(struct ifnet *));
 749         m_tag_prepend(m, mtag);
 750
 751         return (0);
 752 }
 753
 754 static void
 755 carp_send_ad_all(void)
 756 {
 757         struct carp_softc *sc;
 758
 759         LIST_FOREACH(sc, &carpif_list, sc_next) {
 760                 if (sc->sc_carpdev == NULL)
 761                         continue;
 762
 763                 if ((SC2IFP(sc)->if_flags & IFF_UP) &&
 764                     (SC2IFP(sc)->if_flags & IFF_RUNNING) &&
 765                     sc->sc_state == MASTER)
 766                         carp_send_ad_locked(sc);
 767         }
 768 }
 769
 770 static void
 771 carp_send_ad(void *v)
 772 {
 773         struct carp_softc *sc = v;
 774
 775         carp_send_ad_locked(sc);
 776 }
 777
 778 static void
 779 carp_send_ad_locked(struct carp_softc *sc)
 780 {
 781         struct carp_header ch;
 782         struct timeval tv;
 783         struct carp_header *ch_ptr;
 784         struct mbuf *m;
 785         int len, advbase, advskew;
 786
 787         /* bow out if we've lost our UPness or RUNNINGuiness */
 788         if (!((SC2IFP(sc)->if_flags & IFF_UP) &&
 789               (SC2IFP(sc)->if_flags & IFF_RUNNING))) {
 790                 advbase = 255;
 791                 advskew = 255;
 792         } else {
 793                 advbase = sc->sc_advbase;
 794                 if (!carp_suppress_preempt || sc->sc_advskew > 240)
 795                         advskew = sc->sc_advskew;
 796                 else
 797                         advskew = 240;
 798                 tv.tv_sec = advbase;
 799                 tv.tv_usec = advskew * 1000000 / 256;
 800         }
 801
 802         ch.carp_version = CARP_VERSION;
 803         ch.carp_type = CARP_ADVERTISEMENT;
 804         ch.carp_vhid = sc->sc_vhid;
 805         ch.carp_advbase = advbase;
 806         ch.carp_advskew = advskew;
 807         ch.carp_authlen = 7;    /* XXX DEFINE */
 808         ch.carp_pad1 = 0;       /* must be zero */
 809         ch.carp_cksum = 0;
 810
 811 #ifdef INET
 812         if (sc->sc_ia) {
 813                 struct ip *ip;
 814
 815                 MGETHDR(m, M_NOWAIT, MT_HEADER);
 816                 if (m == NULL) {
 817                         SC2IFP(sc)->if_oerrors++;
 818                         carpstats.carps_onomem++;
 819                         /* XXX maybe less ? */
 820                         if (advbase != 255 || advskew != 255)
 821                                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
 822                                     carp_send_ad, sc);
 823                         return;
 824                 }
 825                 len = sizeof(*ip) + sizeof(ch);
 826                 m->m_pkthdr.len = len;
 827                 m->m_pkthdr.rcvif = NULL;
 828                 m->m_len = len;
 829                 MH_ALIGN(m, m->m_len);
 830                 m->m_flags |= M_MCAST;
 831                 ip = mtod(m, struct ip *);
 832                 ip->ip_v = IPVERSION;
 833                 ip->ip_hl = sizeof(*ip) >> 2;
 834                 ip->ip_tos = IPTOS_LOWDELAY;
 835                 ip->ip_len = len;
 836                 ip->ip_id = ip_newid();
 837                 ip->ip_off = IP_DF;
 838                 ip->ip_ttl = CARP_DFLTTL;
 839                 ip->ip_p = IPPROTO_CARP;
 840                 ip->ip_sum = 0;
 841                 ip->ip_src.s_addr = sc->sc_ia->ia_addr.sin_addr.s_addr;
 842                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 843
 844                 ch_ptr = (struct carp_header *)(&ip[1]);
 845                 bcopy(&ch, ch_ptr, sizeof(ch));
 846                 if (carp_prepare_ad(m, sc, ch_ptr))
 847                         return;
 848                 ch_ptr->carp_cksum = in_cksum_skip(m, len, sizeof(*ip));
 849
 850                 getmicrotime(&SC2IFP(sc)->if_lastchange);
 851                 SC2IFP(sc)->if_opackets++;
 852                 SC2IFP(sc)->if_obytes += len;
 853                 carpstats.carps_opackets++;
 854
 855                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, NULL)) {
 856                         SC2IFP(sc)->if_oerrors++;
 857                         if (sc->sc_sendad_errors < INT_MAX)
 858                                 sc->sc_sendad_errors++;
 859                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 860                                 carp_suppress_preempt++;
 861                                 if (carp_suppress_preempt == 1) {
 862                                         carp_send_ad_all();
 863                                 }
 864                         }
 865                         sc->sc_sendad_success = 0;
 866                 } else {
 867                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 868                                 if (++sc->sc_sendad_success >=
 869                                     CARP_SENDAD_MIN_SUCCESS) {
 870                                         carp_suppress_preempt--;
 871                                         sc->sc_sendad_errors = 0;
 872                                 }
 873                         } else {
 874                                 sc->sc_sendad_errors = 0;
 875                         }
 876                 }
 877         }
 878 #endif /* INET */
 879 #ifdef INET6
 880         if (sc->sc_ia6) {
 881                 struct ip6_hdr *ip6;
 882
 883                 MGETHDR(m, M_NOWAIT, MT_HEADER);
 884                 if (m == NULL) {
 885                         SC2IFP(sc)->if_oerrors++;
 886                         carpstats.carps_onomem++;
 887                         /* XXX maybe less ? */
 888                         if (advbase != 255 || advskew != 255)
 889                                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
 890                                     carp_send_ad, sc);
 891                         return;
 892                 }
 893                 len = sizeof(*ip6) + sizeof(ch);
 894                 m->m_pkthdr.len = len;
 895                 m->m_pkthdr.rcvif = NULL;
 896                 m->m_len = len;
 897                 MH_ALIGN(m, m->m_len);
 898                 m->m_flags |= M_MCAST;
 899                 ip6 = mtod(m, struct ip6_hdr *);
 900                 bzero(ip6, sizeof(*ip6));
 901                 ip6->ip6_vfc |= IPV6_VERSION;
 902                 ip6->ip6_hlim = CARP_DFLTTL;
 903                 ip6->ip6_nxt = IPPROTO_CARP;
 904                 bcopy(&sc->sc_ia6->ia_addr.sin6_addr, &ip6->ip6_src,
 905                     sizeof(struct in6_addr));
 906                 /* set the multicast destination */
 907
 908                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
 909                 ip6->ip6_dst.s6_addr8[15] = 0x12;
 910                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
 911                         SC2IFP(sc)->if_oerrors++;
 912                         m_freem(m);
 913                         CARP_LOG("%s: in6_setscope failed\n", __func__);
 914                         return;
 915                 }
 916
 917                 ch_ptr = (struct carp_header *)(&ip6[1]);
 918                 bcopy(&ch, ch_ptr, sizeof(ch));
 919                 if (carp_prepare_ad(m, sc, ch_ptr))
 920                         return;
 921                 ch_ptr->carp_cksum = in_cksum_skip(m, len, sizeof(*ip6));
 922
 923                 getmicrotime(&SC2IFP(sc)->if_lastchange);
 924                 SC2IFP(sc)->if_opackets++;
 925                 SC2IFP(sc)->if_obytes += len;
 926                 carpstats.carps_opackets6++;
 927
 928                 if (ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL)) {
 929                         SC2IFP(sc)->if_oerrors++;
 930                         if (sc->sc_sendad_errors < INT_MAX)
 931                                 sc->sc_sendad_errors++;
 932                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 933                                 carp_suppress_preempt++;
 934                                 if (carp_suppress_preempt == 1) {
 935                                         carp_send_ad_all();
 936                                 }
 937                         }
 938                         sc->sc_sendad_success = 0;
 939                 } else {
 940                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 941                                 if (++sc->sc_sendad_success >=
 942                                     CARP_SENDAD_MIN_SUCCESS) {
 943                                         carp_suppress_preempt--;
 944                                         sc->sc_sendad_errors = 0;
 945                                 }
 946                         } else {
 947                                 sc->sc_sendad_errors = 0;
 948                         }
 949                 }
 950         }
 951 #endif /* INET6 */
 952
 953         if (advbase != 255 || advskew != 255)
 954                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
 955                     carp_send_ad, sc);
 956 }
 957
 958 /*
 959  * Broadcast a gratuitous ARP request containing
 960  * the virtual router MAC address for each IP address
 961  * associated with the virtual router.
 962  */
 963 static void
 964 carp_send_arp(struct carp_softc *sc)
 965 {
 966         struct ifaddr_container *ifac;
 967
 968         TAILQ_FOREACH(ifac, &SC2IFP(sc)->if_addrheads[mycpuid], ifa_link) {
 969                 struct ifaddr *ifa = ifac->ifa;
 970
 971                 if (ifa->ifa_addr->sa_family != AF_INET)
 972                         continue;
 973                 arp_ifinit2(sc->sc_carpdev, ifa, IF_LLADDR(sc->sc_ifp));
 974
 975                 DELAY(1000);    /* XXX */
 976         }
 977 }
 978
 979 #ifdef INET6
 980 static void
 981 carp_send_na(struct carp_softc *sc)
 982 {
 983         struct ifaddr_container *ifac;
 984         struct in6_addr *in6;
 985         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
 986
 987         TAILQ_FOREACH(ifac, &SC2IFP(sc)->if_addrheads[mycpuid], ifa_link) {
 988                 struct ifaddr *ifa = ifac->ifa;
 989
 990                 if (ifa->ifa_addr->sa_family != AF_INET6)
 991                         continue;
 992
 993                 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr;
 994                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
 995                     ND_NA_FLAG_OVERRIDE, 1, NULL);
 996                 DELAY(1000);    /* XXX */
 997         }
 998 }
 999 #endif /* INET6 */
1000
1001 static int
1002 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type)
1003 {
1004         struct carp_softc *vh;
1005         int count = 0;
1006
1007         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1008                 if ((type == CARP_COUNT_RUNNING &&
1009                      (SC2IFP(vh)->if_flags & IFF_UP) &&
1010                      (SC2IFP(vh)->if_flags & IFF_RUNNING)) ||
1011                     (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) {
1012                         struct ifaddr_container *ifac;
1013
1014                         TAILQ_FOREACH(ifac, &SC2IFP(vh)->if_addrheads[mycpuid],
1015                                       ifa_link) {
1016                                 struct ifaddr *ifa = ifac->ifa;
1017
1018                                 if (ifa->ifa_addr->sa_family == AF_INET &&
1019                                     ia->ia_addr.sin_addr.s_addr ==
1020                                     ifatoia(ifa)->ia_addr.sin_addr.s_addr)
1021                                         count++;
1022                         }
1023                 }
1024         }
1025         return (count);
1026 }
1027
1028 int
1029 carp_iamatch(void *v, struct in_ifaddr *ia,
1030     struct in_addr *isaddr, uint8_t **enaddr)
1031 {
1032         struct carp_if *cif = v;
1033         struct carp_softc *vh;
1034         int index, count = 0;
1035
1036         if (carp_opts[CARPCTL_ARPBALANCE]) {
1037                 /*
1038                  * XXX proof of concept implementation.
1039                  * We use the source ip to decide which virtual host should
1040                  * handle the request. If we're master of that virtual host,
1041                  * then we respond, otherwise, just drop the arp packet on
1042                  * the floor.
1043                  */
1044                 count = carp_addrcount(cif, ia, CARP_COUNT_RUNNING);
1045                 if (count == 0) {
1046                         /* should never reach this */
1047                         return (0);
1048                 }
1049
1050                 /* this should be a hash, like pf_hash() */
1051                 index = ntohl(isaddr->s_addr) % count;
1052                 count = 0;
1053
1054                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1055                         if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1056                             (SC2IFP(vh)->if_flags & IFF_RUNNING)) {
1057                                 struct ifaddr_container *ifac;
1058
1059                                 TAILQ_FOREACH(ifac,
1060                                 &SC2IFP(vh)->if_addrheads[mycpuid], ifa_link) {
1061                                         struct ifaddr *ifa = ifac->ifa;
1062
1063                                         if (ifa->ifa_addr->sa_family ==
1064                                             AF_INET &&
1065                                             ia->ia_addr.sin_addr.s_addr ==
1066                                             ifatoia(ifa)->ia_addr.sin_addr.s_addr) {
1067                                                 if (count == index) {
1068                                                         if (vh->sc_state == MASTER) {
1069                                                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1070                                                                 return (1);
1071                                                         } else {
1072                                                                 return (0);
1073                                                         }
1074                                                 }
1075                                                 count++;
1076                                         }
1077                                 }
1078                         }
1079                 }
1080         } else {
1081                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1082                         if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1083                             (SC2IFP(vh)->if_flags & IFF_RUNNING) &&
1084                             vh->sc_state == MASTER) {
1085                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1086                                 return (1);
1087                         }
1088                 }
1089         }
1090         return(0);
1091 }
1092
1093 #ifdef INET6
1094 struct ifaddr *
1095 carp_iamatch6(void *v, struct in6_addr *taddr)
1096 {
1097         struct carp_if *cif = v;
1098         struct carp_softc *vh;
1099
1100         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1101                 struct ifaddr_container *ifac;
1102
1103                 TAILQ_FOREACH(ifac, &SC2IFP(vh)->if_addrheads[mycpuid],
1104                               ifa_link) {
1105                         struct ifaddr *ifa = ifac->ifa;
1106
1107                         if (IN6_ARE_ADDR_EQUAL(taddr,
1108                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1109                             (SC2IFP(vh)->if_flags & IFF_UP) &&
1110                             (SC2IFP(vh)->if_flags & IFF_RUNNING) &&
1111                             vh->sc_state == MASTER) {
1112                                 return (ifa);
1113                         }
1114                 }
1115         }
1116         return (NULL);
1117 }
1118
1119 void *
1120 carp_macmatch6(void *v, struct mbuf *m, const struct in6_addr *taddr)
1121 {
1122         struct m_tag *mtag;
1123         struct carp_if *cif = v;
1124         struct carp_softc *sc;
1125
1126         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) {
1127                 struct ifaddr_container *ifac;
1128
1129                 TAILQ_FOREACH(ifac, &SC2IFP(sc)->if_addrheads[mycpuid],
1130                               ifa_link) {
1131                         struct ifaddr *ifa = ifac->ifa;
1132
1133                         if (IN6_ARE_ADDR_EQUAL(taddr,
1134                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1135                             (SC2IFP(sc)->if_flags & IFF_UP) &&
1136                             (SC2IFP(sc)->if_flags & IFF_RUNNING)) {
1137                                 struct ifnet *ifp = SC2IFP(sc);
1138                                 mtag = m_tag_get(PACKET_TAG_CARP,
1139                                     sizeof(struct ifnet *), MB_DONTWAIT);
1140                                 if (mtag == NULL) {
1141                                         /* better a bit than nothing */
1142                                         return (IF_LLADDR(sc->sc_ifp));
1143                                 }
1144                                 bcopy(&ifp, (caddr_t)(mtag + 1),
1145                                     sizeof(struct ifnet *));
1146                                 m_tag_prepend(m, mtag);
1147
1148                                 return (IF_LLADDR(sc->sc_ifp));
1149                         }
1150                 }
1151         }
1152         return (NULL);
1153 }
1154 #endif
1155
1156 struct ifnet *
1157 carp_forus(void *v, void *dhost)
1158 {
1159         struct carp_if *cif = v;
1160         struct carp_softc *vh;
1161         uint8_t *ena = dhost;
1162
1163         /**
1164          * XXX: See here for check on MAC adr is not for virtual use
1165          *
1166          **/
1167
1168         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1169                 return (NULL);
1170
1171         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1172                 if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1173                     (SC2IFP(vh)->if_flags & IFF_RUNNING) &&
1174                     vh->sc_state == MASTER &&
1175                     !bcmp(dhost, IF_LLADDR(vh->sc_ifp), ETHER_ADDR_LEN)) {
1176                         return (SC2IFP(vh));
1177                 }
1178         }
1179         return (NULL);
1180 }
1181
1182 static void
1183 carp_master_down(void *v)
1184 {
1185         struct carp_softc *sc = v;
1186
1187         lwkt_serialize_enter(sc->sc_ifp->if_serializer);
1188         carp_master_down_locked(sc);
1189         lwkt_serialize_exit(sc->sc_ifp->if_serializer);
1190 }
1191
1192 static void
1193 carp_master_down_locked(struct carp_softc *sc)
1194 {
1195         switch (sc->sc_state) {
1196         case INIT:
1197                 kprintf("%s: master_down event in INIT state\n",
1198                     SC2IFP(sc)->if_xname);
1199                 break;
1200
1201         case MASTER:
1202                 break;
1203
1204         case BACKUP:
1205                 carp_set_state(sc, MASTER);
1206                 carp_send_ad_locked(sc);
1207                 carp_send_arp(sc);
1208 #ifdef INET6
1209                 carp_send_na(sc);
1210 #endif /* INET6 */
1211                 carp_setrun(sc, 0);
1212                 carp_setroute(sc, RTM_ADD);
1213                 break;
1214         }
1215 }
1216
1217 /*
1218  * When in backup state, af indicates whether to reset the master down timer
1219  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1220  */
1221 static void
1222 carp_setrun(struct carp_softc *sc, sa_family_t af)
1223 {
1224         struct timeval tv;
1225
1226         if (sc->sc_carpdev == NULL) {
1227                 SC2IFP(sc)->if_flags &= ~IFF_RUNNING;
1228                 carp_set_state(sc, INIT);
1229                 return;
1230         }
1231
1232         if (SC2IFP(sc)->if_flags & IFF_UP &&
1233             sc->sc_vhid > 0 && (sc->sc_naddrs || sc->sc_naddrs6)) {
1234                 SC2IFP(sc)->if_flags |= IFF_RUNNING;
1235         } else {
1236                 SC2IFP(sc)->if_flags &= ~IFF_RUNNING;
1237                 carp_setroute(sc, RTM_DELETE);
1238                 return;
1239         }
1240
1241         switch (sc->sc_state) {
1242         case INIT:
1243                 if (carp_opts[CARPCTL_PREEMPT] && !carp_suppress_preempt) {
1244                         carp_send_ad_locked(sc);
1245                         carp_send_arp(sc);
1246 #ifdef INET6
1247                         carp_send_na(sc);
1248 #endif /* INET6 */
1249                         CARP_DEBUG("%s: INIT -> MASTER (preempting)\n",
1250                                    SC2IFP(sc)->if_xname);
1251                         carp_set_state(sc, MASTER);
1252                         carp_setroute(sc, RTM_ADD);
1253                 } else {
1254                         CARP_DEBUG("%s: INIT -> BACKUP\n",
1255                                    SC2IFP(sc)->if_xname);
1256                         carp_set_state(sc, BACKUP);
1257                         carp_setroute(sc, RTM_DELETE);
1258                         carp_setrun(sc, 0);
1259                 }
1260                 break;
1261
1262         case BACKUP:
1263                 callout_stop(&sc->sc_ad_tmo);
1264                 tv.tv_sec = 3 * sc->sc_advbase;
1265                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1266                 switch (af) {
1267 #ifdef INET
1268                 case AF_INET:
1269                         callout_reset(&sc->sc_md_tmo, tvtohz_high(&tv),
1270                             carp_master_down, sc);
1271                         break;
1272 #endif /* INET */
1273 #ifdef INET6
1274                 case AF_INET6:
1275                         callout_reset(&sc->sc_md6_tmo, tvtohz_high(&tv),
1276                             carp_master_down, sc);
1277                         break;
1278 #endif /* INET6 */
1279                 default:
1280                         if (sc->sc_naddrs)
1281                                 callout_reset(&sc->sc_md_tmo, tvtohz_high(&tv),
1282                                     carp_master_down, sc);
1283                         if (sc->sc_naddrs6)
1284                                 callout_reset(&sc->sc_md6_tmo, tvtohz_high(&tv),
1285                                     carp_master_down, sc);
1286                         break;
1287                 }
1288                 break;
1289
1290         case MASTER:
1291                 tv.tv_sec = sc->sc_advbase;
1292                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1293                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
1294                     carp_send_ad, sc);
1295                 break;
1296         }
1297 }
1298
1299 static void
1300 carp_multicast_cleanup(struct carp_softc *sc)
1301 {
1302         struct ip_moptions *imo = &sc->sc_imo;
1303         uint16_t n = imo->imo_num_memberships;
1304
1305         /* Clean up our own multicast memberships */
1306         while (n-- > 0) {
1307                 if (imo->imo_membership[n] != NULL) {
1308                         in_delmulti(imo->imo_membership[n]);
1309                         imo->imo_membership[n] = NULL;
1310                 }
1311         }
1312         imo->imo_num_memberships = 0;
1313         imo->imo_multicast_ifp = NULL;
1314 }
1315
1316 #ifdef INET6
1317 static void
1318 carp_multicast6_cleanup(struct carp_softc *sc)
1319 {
1320         struct ip6_moptions *im6o = &sc->sc_im6o;
1321
1322         while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1323                 struct in6_multi_mship *imm =
1324                     LIST_FIRST(&im6o->im6o_memberships);
1325
1326                 LIST_REMOVE(imm, i6mm_chain);
1327                 in6_leavegroup(imm);
1328         }
1329         im6o->im6o_multicast_ifp = NULL;
1330 }
1331 #endif
1332
1333 static int
1334 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1335 {
1336         struct ifnet *ifp;
1337         struct carp_if *cif;
1338         struct in_ifaddr *ia, *ia_if;
1339         struct in_ifaddr_container *iac;
1340         struct ip_moptions *imo = &sc->sc_imo;
1341         struct in_addr addr;
1342         u_long iaddr = htonl(sin->sin_addr.s_addr);
1343         int own, error;
1344
1345         if (sin->sin_addr.s_addr == 0) {
1346                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1347                         carp_set_state(sc, INIT);
1348                 if (sc->sc_naddrs)
1349                         SC2IFP(sc)->if_flags |= IFF_UP;
1350                 carp_setrun(sc, 0);
1351                 return (0);
1352         }
1353         /* we have to do it by hands to check we won't match on us */
1354         ia_if = NULL; own = 0;
1355         TAILQ_FOREACH(iac, &in_ifaddrheads[mycpuid], ia_link) {
1356                 ia = iac->ia;
1357
1358                 /* and, yeah, we need a multicast-capable iface too */
1359                 if (ia->ia_ifp != SC2IFP(sc) &&
1360                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1361                     (iaddr & ia->ia_subnetmask) == ia->ia_subnet) {
1362                         if (!ia_if)
1363                                 ia_if = ia;
1364                         if (sin->sin_addr.s_addr ==
1365                             ia->ia_addr.sin_addr.s_addr)
1366                                 own++;
1367                 }
1368         }
1369
1370         if (!ia_if)
1371                 return (EADDRNOTAVAIL);
1372
1373         ia = ia_if;
1374         ifp = ia->ia_ifp;
1375
1376         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1377             (imo->imo_multicast_ifp && imo->imo_multicast_ifp != ifp))
1378                 return (EADDRNOTAVAIL);
1379
1380         if (imo->imo_num_memberships == 0) {
1381                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1382                 if ((imo->imo_membership[0] = in_addmulti(&addr, ifp)) == NULL)
1383                         return (ENOBUFS);
1384                 imo->imo_num_memberships++;
1385                 imo->imo_multicast_ifp = ifp;
1386                 imo->imo_multicast_ttl = CARP_DFLTTL;
1387                 imo->imo_multicast_loop = 0;
1388         }
1389
1390         if (!ifp->if_carp) {
1391                 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP,
1392                     M_WAITOK|M_ZERO);
1393                 if ((error = ifpromisc(ifp, 1))) {
1394                         FREE(cif, M_CARP);
1395                         goto cleanup;
1396                 }
1397
1398                 cif->vhif_ifp = ifp;
1399                 TAILQ_INIT(&cif->vhif_vrs);
1400                 ifp->if_carp = cif;
1401         } else {
1402                 struct carp_softc *vr;
1403
1404                 cif = (struct carp_if *)ifp->if_carp;
1405                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1406                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1407                                 error = EINVAL;
1408                                 goto cleanup;
1409                         }
1410                 }
1411         }
1412         sc->sc_ia = ia;
1413         sc->sc_carpdev = ifp;
1414
1415         { /* XXX prevent endless loop if already in queue */
1416         struct carp_softc *vr, *after = NULL;
1417         int myself = 0;
1418         cif = (struct carp_if *)ifp->if_carp;
1419
1420         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1421                 if (vr == sc)
1422                         myself = 1;
1423                 if (vr->sc_vhid < sc->sc_vhid)
1424                         after = vr;
1425         }
1426
1427         if (!myself) {
1428                 /* We're trying to keep things in order */
1429                 if (after == NULL)
1430                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1431                 else
1432                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1433                 cif->vhif_nvrs++;
1434         }
1435         }
1436
1437         sc->sc_naddrs++;
1438         SC2IFP(sc)->if_flags |= IFF_UP;
1439         if (own)
1440                 sc->sc_advskew = 0;
1441
1442         carp_sc_state_locked(sc);
1443         carp_setrun(sc, 0);
1444
1445         return (0);
1446
1447 cleanup:
1448         in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1449         return (error);
1450 }
1451
1452 static int
1453 carp_del_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1454 {
1455         int error = 0;
1456
1457         if (!--sc->sc_naddrs) {
1458                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1459                 struct ip_moptions *imo = &sc->sc_imo;
1460
1461                 callout_stop(&sc->sc_ad_tmo);
1462                 SC2IFP(sc)->if_flags &= ~(IFF_UP | IFF_RUNNING);
1463                 sc->sc_vhid = -1;
1464                 in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1465                 imo->imo_multicast_ifp = NULL;
1466                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1467                 if (!--cif->vhif_nvrs) {
1468                         sc->sc_carpdev->if_carp = NULL;
1469                         FREE(cif, M_IFADDR);
1470                 }
1471         }
1472         return (error);
1473 }
1474
1475 #ifdef INET6
1476 static int
1477 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1478 {
1479         struct ifnet *ifp;
1480         struct carp_if *cif;
1481         struct in6_ifaddr *ia, *ia_if;
1482         struct ip6_moptions *im6o = &sc->sc_im6o;
1483         struct in6_multi_mship *imm;
1484         struct in6_addr in6;
1485         int own, error;
1486
1487         if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
1488                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1489                         carp_set_state(sc, INIT);
1490                 if (sc->sc_naddrs6)
1491                         SC2IFP(sc)->if_flags |= IFF_UP;
1492                 carp_setrun(sc, 0);
1493                 return (0);
1494         }
1495
1496         /* we have to do it by hands to check we won't match on us */
1497         ia_if = NULL; own = 0;
1498         for (ia = in6_ifaddr; ia; ia = ia->ia_next) {
1499                 int i;
1500
1501                 for (i = 0; i < 4; i++) {
1502                         if ((sin6->sin6_addr.s6_addr32[i] &
1503                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]) !=
1504                             (ia->ia_addr.sin6_addr.s6_addr32[i] &
1505                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]))
1506                                 break;
1507                 }
1508                 /* and, yeah, we need a multicast-capable iface too */
1509                 if (ia->ia_ifp != SC2IFP(sc) &&
1510                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1511                     (i == 4)) {
1512                         if (!ia_if)
1513                                 ia_if = ia;
1514                         if (IN6_ARE_ADDR_EQUAL(&sin6->sin6_addr,
1515                             &ia->ia_addr.sin6_addr))
1516                                 own++;
1517                 }
1518         }
1519
1520         if (!ia_if)
1521                 return (EADDRNOTAVAIL);
1522         ia = ia_if;
1523         ifp = ia->ia_ifp;
1524
1525         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1526             (im6o->im6o_multicast_ifp && im6o->im6o_multicast_ifp != ifp))
1527                 return (EADDRNOTAVAIL);
1528
1529         if (!sc->sc_naddrs6) {
1530                 im6o->im6o_multicast_ifp = ifp;
1531
1532                 /* join CARP multicast address */
1533                 bzero(&in6, sizeof(in6));
1534                 in6.s6_addr16[0] = htons(0xff02);
1535                 in6.s6_addr8[15] = 0x12;
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                 /* join solicited multicast address */
1543                 bzero(&in6, sizeof(in6));
1544                 in6.s6_addr16[0] = htons(0xff02);
1545                 in6.s6_addr32[1] = 0;
1546                 in6.s6_addr32[2] = htonl(1);
1547                 in6.s6_addr32[3] = sin6->sin6_addr.s6_addr32[3];
1548                 in6.s6_addr8[12] = 0xff;
1549                 if (in6_setscope(&in6, ifp, NULL) != 0)
1550                         goto cleanup;
1551                 if ((imm = in6_joingroup(ifp, &in6, &error)) == NULL)
1552                         goto cleanup;
1553                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1554         }
1555
1556         if (!ifp->if_carp) {
1557                 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP,
1558                     M_WAITOK|M_ZERO);
1559                 if ((error = ifpromisc(ifp, 1))) {
1560                         FREE(cif, M_CARP);
1561                         goto cleanup;
1562                 }
1563
1564                 cif->vhif_ifp = ifp;
1565                 TAILQ_INIT(&cif->vhif_vrs);
1566                 ifp->if_carp = cif;
1567         } else {
1568                 struct carp_softc *vr;
1569
1570                 cif = (struct carp_if *)ifp->if_carp;
1571                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1572                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1573                                 error = EINVAL;
1574                                 goto cleanup;
1575                         }
1576                 }
1577         }
1578         sc->sc_ia6 = ia;
1579         sc->sc_carpdev = ifp;
1580
1581         { /* XXX prevent endless loop if already in queue */
1582         struct carp_softc *vr, *after = NULL;
1583         int myself = 0;
1584         cif = (struct carp_if *)ifp->if_carp;
1585
1586         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1587                 if (vr == sc)
1588                         myself = 1;
1589                 if (vr->sc_vhid < sc->sc_vhid)
1590                         after = vr;
1591         }
1592
1593         if (!myself) {
1594                 /* We're trying to keep things in order */
1595                 if (after == NULL)
1596                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1597                 else
1598                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1599                 cif->vhif_nvrs++;
1600         }
1601         }
1602
1603         sc->sc_naddrs6++;
1604         SC2IFP(sc)->if_flags |= IFF_UP;
1605         if (own)
1606                 sc->sc_advskew = 0;
1607         carp_sc_state_locked(sc);
1608         carp_setrun(sc, 0);
1609
1610         return (0);
1611
1612 cleanup:
1613         /* clean up multicast memberships */
1614         if (!sc->sc_naddrs6) {
1615                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1616                         imm = LIST_FIRST(&im6o->im6o_memberships);
1617                         LIST_REMOVE(imm, i6mm_chain);
1618                         in6_leavegroup(imm);
1619                 }
1620         }
1621         return (error);
1622 }
1623
1624 static int
1625 carp_del_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1626 {
1627         int error = 0;
1628
1629         if (!--sc->sc_naddrs6) {
1630                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1631                 struct ip6_moptions *im6o = &sc->sc_im6o;
1632
1633                 callout_stop(&sc->sc_ad_tmo);
1634                 SC2IFP(sc)->if_flags &= ~(IFF_UP | IFF_RUNNING);
1635                 sc->sc_vhid = -1;
1636                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1637                         struct in6_multi_mship *imm =
1638                             LIST_FIRST(&im6o->im6o_memberships);
1639
1640                         LIST_REMOVE(imm, i6mm_chain);
1641                         in6_leavegroup(imm);
1642                 }
1643                 im6o->im6o_multicast_ifp = NULL;
1644                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1645                 if (!--cif->vhif_nvrs) {
1646                         sc->sc_carpdev->if_carp = NULL;
1647                         FREE(cif, M_IFADDR);
1648                 }
1649         }
1650         return (error);
1651 }
1652 #endif /* INET6 */
1653
1654 static int
1655 carp_ioctl(struct ifnet *ifp, u_long cmd, caddr_t addr, struct ucred *cr)
1656 {
1657         struct carp_softc *sc = ifp->if_softc, *vr;
1658         struct carpreq carpr;
1659         struct ifaddr *ifa;
1660         struct ifreq *ifr;
1661         struct ifaliasreq *ifra;
1662         int error = 0;
1663
1664         ifa = (struct ifaddr *)addr;
1665         ifra = (struct ifaliasreq *)addr;
1666         ifr = (struct ifreq *)addr;
1667
1668         switch (cmd) {
1669         case SIOCSIFADDR:
1670                 switch (ifa->ifa_addr->sa_family) {
1671 #ifdef INET
1672                 case AF_INET:
1673                         SC2IFP(sc)->if_flags |= IFF_UP;
1674                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1675                             sizeof(struct sockaddr));
1676                         error = carp_set_addr(sc, satosin(ifa->ifa_addr));
1677                         break;
1678 #endif /* INET */
1679 #ifdef INET6
1680                 case AF_INET6:
1681                         SC2IFP(sc)->if_flags |= IFF_UP;
1682                         error = carp_set_addr6(sc, satosin6(ifa->ifa_addr));
1683                         break;
1684 #endif /* INET6 */
1685                 default:
1686                         error = EAFNOSUPPORT;
1687                         break;
1688                 }
1689                 break;
1690
1691         case SIOCAIFADDR:
1692                 switch (ifa->ifa_addr->sa_family) {
1693 #ifdef INET
1694                 case AF_INET:
1695                         SC2IFP(sc)->if_flags |= IFF_UP;
1696                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1697                             sizeof(struct sockaddr));
1698                         error = carp_set_addr(sc, satosin(&ifra->ifra_addr));
1699                         break;
1700 #endif /* INET */
1701 #ifdef INET6
1702                 case AF_INET6:
1703                         SC2IFP(sc)->if_flags |= IFF_UP;
1704                         error = carp_set_addr6(sc, satosin6(&ifra->ifra_addr));
1705                         break;
1706 #endif /* INET6 */
1707                 default:
1708                         error = EAFNOSUPPORT;
1709                         break;
1710                 }
1711                 break;
1712
1713         case SIOCDIFADDR:
1714                 switch (ifa->ifa_addr->sa_family) {
1715 #ifdef INET
1716                 case AF_INET:
1717                         error = carp_del_addr(sc, satosin(&ifra->ifra_addr));
1718                         break;
1719 #endif /* INET */
1720 #ifdef INET6
1721                 case AF_INET6:
1722                         error = carp_del_addr6(sc, satosin6(&ifra->ifra_addr));
1723                         break;
1724 #endif /* INET6 */
1725                 default:
1726                         error = EAFNOSUPPORT;
1727                         break;
1728                 }
1729                 break;
1730
1731         case SIOCSIFFLAGS:
1732                 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) {
1733                         callout_stop(&sc->sc_ad_tmo);
1734                         callout_stop(&sc->sc_md_tmo);
1735                         callout_stop(&sc->sc_md6_tmo);
1736                         if (sc->sc_state == MASTER)
1737                                 carp_send_ad_locked(sc);
1738                         carp_set_state(sc, INIT);
1739                         carp_setrun(sc, 0);
1740                 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) {
1741                         SC2IFP(sc)->if_flags |= IFF_UP;
1742                         carp_setrun(sc, 0);
1743                 }
1744                 break;
1745
1746         case SIOCSVH:
1747                 error = suser_cred(cr, NULL_CRED_OKAY);
1748                 if (error)
1749                         break;
1750                 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1751                         break;
1752                 error = 1;
1753                 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) {
1754                         switch (carpr.carpr_state) {
1755                         case BACKUP:
1756                                 callout_stop(&sc->sc_ad_tmo);
1757                                 carp_set_state(sc, BACKUP);
1758                                 carp_setrun(sc, 0);
1759                                 carp_setroute(sc, RTM_DELETE);
1760                                 break;
1761
1762                         case MASTER:
1763                                 carp_master_down_locked(sc);
1764                                 break;
1765
1766                         default:
1767                                 break;
1768                         }
1769                 }
1770                 if (carpr.carpr_vhid > 0) {
1771                         if (carpr.carpr_vhid > 255) {
1772                                 error = EINVAL;
1773                                 break;
1774                         }
1775                         if (sc->sc_carpdev) {
1776                                 struct carp_if *cif;
1777                                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1778                                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1779                                         if (vr != sc &&
1780                                             vr->sc_vhid == carpr.carpr_vhid)
1781                                                 return EEXIST;
1782                                 }
1783                         }
1784                         sc->sc_vhid = carpr.carpr_vhid;
1785                         IF_LLADDR(sc->sc_ifp)[0] = 0;
1786                         IF_LLADDR(sc->sc_ifp)[1] = 0;
1787                         IF_LLADDR(sc->sc_ifp)[2] = 0x5e;
1788                         IF_LLADDR(sc->sc_ifp)[3] = 0;
1789                         IF_LLADDR(sc->sc_ifp)[4] = 1;
1790                         IF_LLADDR(sc->sc_ifp)[5] = sc->sc_vhid;
1791                         error--;
1792                 }
1793                 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) {
1794                         if (carpr.carpr_advskew >= 255) {
1795                                 error = EINVAL;
1796                                 break;
1797                         }
1798                         if (carpr.carpr_advbase > 255) {
1799                                 error = EINVAL;
1800                                 break;
1801                         }
1802                         sc->sc_advbase = carpr.carpr_advbase;
1803                         sc->sc_advskew = carpr.carpr_advskew;
1804                         error--;
1805                 }
1806                 bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1807                 if (error > 0) {
1808                         error = EINVAL;
1809                 } else {
1810                         error = 0;
1811                         carp_setrun(sc, 0);
1812                 }
1813                 break;
1814
1815         case SIOCGVH:
1816                 bzero(&carpr, sizeof(carpr));
1817                 carpr.carpr_state = sc->sc_state;
1818                 carpr.carpr_vhid = sc->sc_vhid;
1819                 carpr.carpr_advbase = sc->sc_advbase;
1820                 carpr.carpr_advskew = sc->sc_advskew;
1821                 error = suser_cred(cr, NULL_CRED_OKAY);
1822                 if (error == 0) {
1823                         bcopy(sc->sc_key, carpr.carpr_key,
1824                             sizeof(carpr.carpr_key));
1825                 }
1826                 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1827                 break;
1828
1829         default:
1830                 error = EINVAL;
1831         }
1832         carp_hmac_prepare(sc);
1833         return (error);
1834 }
1835
1836 /*
1837  * XXX: this is looutput. We should eventually use it from there.
1838  */
1839 static int
1840 carp_looutput(struct ifnet *ifp, struct mbuf *m, struct sockaddr *dst,
1841     struct rtentry *rt)
1842 {
1843         uint32_t af;
1844
1845         M_ASSERTPKTHDR(m); /* check if we have the packet header */
1846
1847         if (rt && rt->rt_flags & (RTF_REJECT|RTF_BLACKHOLE)) {
1848                 m_freem(m);
1849                 return (rt->rt_flags & RTF_BLACKHOLE ? 0 :
1850                         rt->rt_flags & RTF_HOST ? EHOSTUNREACH : ENETUNREACH);
1851         }
1852
1853         ifp->if_opackets++;
1854         ifp->if_obytes += m->m_pkthdr.len;
1855
1856         /* BPF writes need to be handled specially. */
1857         if (dst->sa_family == AF_UNSPEC) {
1858                 bcopy(dst->sa_data, &af, sizeof(af));
1859                 dst->sa_family = af;
1860         }
1861
1862 #if 1   /* XXX */
1863         switch (dst->sa_family) {
1864         case AF_INET:
1865         case AF_INET6:
1866         case AF_IPX:
1867         case AF_APPLETALK:
1868                 break;
1869
1870         default:
1871                 m_freem(m);
1872                 return (EAFNOSUPPORT);
1873         }
1874 #endif
1875         return (if_simloop(ifp, m, dst->sa_family, 0));
1876 }
1877
1878 /*
1879  * Start output on carp interface. This function should never be called.
1880  */
1881 static void
1882 carp_start(struct ifnet *ifp)
1883 {
1884 #ifdef DEBUG
1885         kprintf("%s: start called\n", ifp->if_xname);
1886 #endif
1887 }
1888
1889 int
1890 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *sa,
1891     struct rtentry *rt)
1892 {
1893         struct m_tag *mtag;
1894         struct carp_softc *sc;
1895         struct ifnet *carp_ifp;
1896         struct ether_header *eh;
1897
1898         if (!sa)
1899                 return (0);
1900
1901         switch (sa->sa_family) {
1902 #ifdef INET
1903         case AF_INET:
1904                 break;
1905 #endif /* INET */
1906 #ifdef INET6
1907         case AF_INET6:
1908                 break;
1909 #endif /* INET6 */
1910         default:
1911                 return (0);
1912         }
1913
1914         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
1915         if (mtag == NULL)
1916                 return (0);
1917
1918         bcopy(mtag + 1, &carp_ifp, sizeof(struct ifnet *));
1919         sc = carp_ifp->if_softc;
1920
1921         /* Set the source MAC address to Virtual Router MAC Address */
1922         switch (ifp->if_type) {
1923         case IFT_ETHER:
1924         case IFT_L2VLAN:
1925                 eh = mtod(m, struct ether_header *);
1926                 eh->ether_shost[0] = 0;
1927                 eh->ether_shost[1] = 0;
1928                 eh->ether_shost[2] = 0x5e;
1929                 eh->ether_shost[3] = 0;
1930                 eh->ether_shost[4] = 1;
1931                 eh->ether_shost[5] = sc->sc_vhid;
1932                 break;
1933
1934         default:
1935                 if_printf(ifp, "carp is not supported for this "
1936                           "interface type\n");
1937                 return (EOPNOTSUPP);
1938         }
1939         return (0);
1940 }
1941
1942 static void
1943 carp_set_state(struct carp_softc *sc, int state)
1944 {
1945         if (sc->sc_state == state)
1946                 return;
1947
1948         sc->sc_state = state;
1949         switch (state) {
1950         case BACKUP:
1951                 SC2IFP(sc)->if_link_state = LINK_STATE_DOWN;
1952                 break;
1953
1954         case MASTER:
1955                 SC2IFP(sc)->if_link_state = LINK_STATE_UP;
1956                 break;
1957
1958         default:
1959                 SC2IFP(sc)->if_link_state = LINK_STATE_UNKNOWN;
1960                 break;
1961         }
1962         rt_ifmsg(SC2IFP(sc));
1963 }
1964
1965 void
1966 carp_carpdev_state(void *v)
1967 {
1968         struct carp_if *cif = v;
1969
1970         carp_carpdev_state_locked(cif);
1971 }
1972
1973 static void
1974 carp_carpdev_state_locked(struct carp_if *cif)
1975 {
1976         struct carp_softc *sc;
1977
1978         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list)
1979                 carp_sc_state_locked(sc);
1980 }
1981
1982 static void
1983 carp_sc_state_locked(struct carp_softc *sc)
1984 {
1985         if (!(sc->sc_carpdev->if_flags & IFF_UP)) {
1986                 sc->sc_flags_backup = SC2IFP(sc)->if_flags;
1987                 SC2IFP(sc)->if_flags &= ~(IFF_UP | IFF_RUNNING);
1988                 callout_stop(&sc->sc_ad_tmo);
1989                 callout_stop(&sc->sc_md_tmo);
1990                 callout_stop(&sc->sc_md6_tmo);
1991                 carp_set_state(sc, INIT);
1992                 carp_setrun(sc, 0);
1993                 if (!sc->sc_suppress) {
1994                         carp_suppress_preempt++;
1995                         if (carp_suppress_preempt == 1)
1996                                 carp_send_ad_all();
1997                 }
1998                 sc->sc_suppress = 1;
1999         } else {
2000                 SC2IFP(sc)->if_flags |= sc->sc_flags_backup;
2001                 carp_set_state(sc, INIT);
2002                 carp_setrun(sc, 0);
2003                 if (sc->sc_suppress)
2004                         carp_suppress_preempt--;
2005                 sc->sc_suppress = 0;
2006         }
2007 }
2008
2009 static int
2010 carp_modevent(module_t mod, int type, void *data)
2011 {
2012         switch (type) {
2013         case MOD_LOAD:
2014                 LIST_INIT(&carpif_list);
2015                 carp_ifdetach_event =
2016                 EVENTHANDLER_REGISTER(ifnet_detach_event, carp_ifdetach, NULL,
2017                                       EVENTHANDLER_PRI_ANY);
2018                 if_clone_attach(&carp_cloner);
2019                 break;
2020
2021         case MOD_UNLOAD:
2022                 EVENTHANDLER_DEREGISTER(ifnet_detach_event,
2023                                         carp_ifdetach_event);
2024                 if_clone_detach(&carp_cloner);
2025                 break;
2026
2027         default:
2028                 return (EINVAL);
2029         }
2030         return (0);
2031 }
2032
2033 static moduledata_t carp_mod = {
2034         "carp",
2035         carp_modevent,
2036         0
2037 };
2038 DECLARE_MODULE(carp, carp_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);