sys/netinet/ip_carp.c

   1 /*      $FreeBSD$ */
   2
   3 /*
   4  * Copyright (c) 2002 Michael Shalayeff. All rights reserved.
   5  * Copyright (c) 2003 Ryan McBride. All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  17  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  18  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  19  * IN NO EVENT SHALL THE AUTHOR OR HIS RELATIVES BE LIABLE FOR ANY DIRECT,
  20  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  21  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  22  * SERVICES; LOSS OF MIND, USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
  24  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
  25  * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
  26  * THE POSSIBILITY OF SUCH DAMAGE.
  27  */
  28
  29 #include "opt_carp.h"
  30 #include "opt_bpf.h"
  31 #include "opt_inet.h"
  32 #include "opt_inet6.h"
  33
  34 #include <sys/types.h>
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/conf.h>
  38 #include <sys/kernel.h>
  39 #include <sys/limits.h>
  40 #include <sys/malloc.h>
  41 #include <sys/mbuf.h>
  42 #include <sys/module.h>
  43 #include <sys/time.h>
  44 #include <sys/priv.h>
  45 #include <sys/proc.h>
  46 #include <sys/sysctl.h>
  47 #include <sys/syslog.h>
  48 #include <sys/signalvar.h>
  49 #include <sys/filio.h>
  50 #include <sys/sockio.h>
  51
  52 #include <sys/socket.h>
  53 #include <sys/vnode.h>
  54
  55 #include <machine/stdarg.h>
  56
  57 #include <net/bpf.h>
  58 #include <net/ethernet.h>
  59 #include <net/fddi.h>
  60 #include <net/iso88025.h>
  61 #include <net/if.h>
  62 #include <net/if_clone.h>
  63 #include <net/if_dl.h>
  64 #include <net/if_types.h>
  65 #include <net/route.h>
  66
  67 #ifdef INET
  68 #include <netinet/in.h>
  69 #include <netinet/in_var.h>
  70 #include <netinet/in_systm.h>
  71 #include <netinet/ip.h>
  72 #include <netinet/ip_var.h>
  73 #include <netinet/if_ether.h>
  74 #include <machine/in_cksum.h>
  75 #endif
  76
  77 #ifdef INET6
  78 #include <netinet/icmp6.h>
  79 #include <netinet/ip6.h>
  80 #include <netinet6/ip6_var.h>
  81 #include <netinet6/scope6_var.h>
  82 #include <netinet6/nd6.h>
  83 #endif
  84
  85 #include <crypto/sha1.h>
  86 #include <netinet/ip_carp.h>
  87
  88 #define CARP_IFNAME     "carp"
  89 static MALLOC_DEFINE(M_CARP, "CARP", "CARP interfaces");
  90 SYSCTL_DECL(_net_inet_carp);
  91
  92 struct carp_softc {
  93         struct ifnet            *sc_ifp;        /* Interface clue */
  94         struct ifnet            *sc_carpdev;    /* Pointer to parent interface */
  95         struct in_ifaddr        *sc_ia;         /* primary iface address */
  96         struct ip_moptions       sc_imo;
  97 #ifdef INET6
  98         struct in6_ifaddr       *sc_ia6;        /* primary iface address v6 */
  99         struct ip6_moptions      sc_im6o;
 100 #endif /* INET6 */
 101         TAILQ_ENTRY(carp_softc)  sc_list;
 102
 103         enum { INIT = 0, BACKUP, MASTER }       sc_state;
 104
 105         int                      sc_flags_backup;
 106         int                      sc_suppress;
 107
 108         int                      sc_sendad_errors;
 109 #define CARP_SENDAD_MAX_ERRORS  3
 110         int                      sc_sendad_success;
 111 #define CARP_SENDAD_MIN_SUCCESS 3
 112
 113         int                      sc_vhid;
 114         int                      sc_advskew;
 115         int                      sc_naddrs;
 116         int                      sc_naddrs6;
 117         int                      sc_advbase;    /* seconds */
 118         int                      sc_init_counter;
 119         u_int64_t                sc_counter;
 120
 121         /* authentication */
 122 #define CARP_HMAC_PAD   64
 123         unsigned char sc_key[CARP_KEY_LEN];
 124         unsigned char sc_pad[CARP_HMAC_PAD];
 125         SHA1_CTX sc_sha1;
 126
 127         struct callout           sc_ad_tmo;     /* advertisement timeout */
 128         struct callout           sc_md_tmo;     /* master down timeout */
 129         struct callout           sc_md6_tmo;    /* master down timeout */
 130
 131         LIST_ENTRY(carp_softc)   sc_next;       /* Interface clue */
 132 };
 133 #define SC2IFP(sc)      ((sc)->sc_ifp)
 134
 135 int carp_suppress_preempt = 0;
 136 int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 1, 0, 0 };    /* XXX for now */
 137 SYSCTL_INT(_net_inet_carp, CARPCTL_ALLOW, allow, CTLFLAG_RW,
 138     &carp_opts[CARPCTL_ALLOW], 0, "Accept incoming CARP packets");
 139 SYSCTL_INT(_net_inet_carp, CARPCTL_PREEMPT, preempt, CTLFLAG_RW,
 140     &carp_opts[CARPCTL_PREEMPT], 0, "high-priority backup preemption mode");
 141 SYSCTL_INT(_net_inet_carp, CARPCTL_LOG, log, CTLFLAG_RW,
 142     &carp_opts[CARPCTL_LOG], 0, "log bad carp packets");
 143 SYSCTL_INT(_net_inet_carp, CARPCTL_ARPBALANCE, arpbalance, CTLFLAG_RW,
 144     &carp_opts[CARPCTL_ARPBALANCE], 0, "balance arp responses");
 145 SYSCTL_INT(_net_inet_carp, OID_AUTO, suppress_preempt, CTLFLAG_RD,
 146     &carp_suppress_preempt, 0, "Preemption is suppressed");
 147
 148 struct carpstats carpstats;
 149 SYSCTL_STRUCT(_net_inet_carp, CARPCTL_STATS, stats, CTLFLAG_RW,
 150     &carpstats, carpstats,
 151     "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 152
 153 struct carp_if {
 154         TAILQ_HEAD(, carp_softc) vhif_vrs;
 155         int vhif_nvrs;
 156
 157         struct ifnet    *vhif_ifp;
 158         struct mtx       vhif_mtx;
 159 };
 160
 161 /* Get carp_if from softc. Valid after carp_set_addr{,6}. */
 162 #define SC2CIF(sc)              ((struct carp_if *)(sc)->sc_carpdev->if_carp)
 163
 164 /* lock per carp_if queue */
 165 #define CARP_LOCK_INIT(cif)     mtx_init(&(cif)->vhif_mtx, "carp_if",   \
 166         NULL, MTX_DEF)
 167 #define CARP_LOCK_DESTROY(cif)  mtx_destroy(&(cif)->vhif_mtx)
 168 #define CARP_LOCK_ASSERT(cif)   mtx_assert(&(cif)->vhif_mtx, MA_OWNED)
 169 #define CARP_LOCK(cif)          mtx_lock(&(cif)->vhif_mtx)
 170 #define CARP_UNLOCK(cif)        mtx_unlock(&(cif)->vhif_mtx)
 171
 172 #define CARP_SCLOCK(sc)         mtx_lock(&SC2CIF(sc)->vhif_mtx)
 173 #define CARP_SCUNLOCK(sc)       mtx_unlock(&SC2CIF(sc)->vhif_mtx)
 174 #define CARP_SCLOCK_ASSERT(sc)  mtx_assert(&SC2CIF(sc)->vhif_mtx, MA_OWNED)
 175
 176 #define CARP_LOG(...)   do {                            \
 177         if (carp_opts[CARPCTL_LOG] > 0)                 \
 178                 log(LOG_INFO, __VA_ARGS__);             \
 179 } while (0)
 180
 181 #define CARP_DEBUG(...) do {                            \
 182         if (carp_opts[CARPCTL_LOG] > 1)                 \
 183                 log(LOG_DEBUG, __VA_ARGS__);            \
 184 } while (0)
 185
 186 static void     carp_hmac_prepare(struct carp_softc *);
 187 static void     carp_hmac_generate(struct carp_softc *, u_int32_t *,
 188                     unsigned char *);
 189 static int      carp_hmac_verify(struct carp_softc *, u_int32_t *,
 190                     unsigned char *);
 191 static void     carp_setroute(struct carp_softc *, int);
 192 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 193 static int      carp_clone_create(struct if_clone *, int, caddr_t);
 194 static void     carp_clone_destroy(struct ifnet *);
 195 static void     carpdetach(struct carp_softc *, int);
 196 static int      carp_prepare_ad(struct mbuf *, struct carp_softc *,
 197                     struct carp_header *);
 198 static void     carp_send_ad_all(void);
 199 static void     carp_send_ad(void *);
 200 static void     carp_send_ad_locked(struct carp_softc *);
 201 static void     carp_send_arp(struct carp_softc *);
 202 static void     carp_master_down(void *);
 203 static void     carp_master_down_locked(struct carp_softc *);
 204 static int      carp_ioctl(struct ifnet *, u_long, caddr_t);
 205 static int      carp_looutput(struct ifnet *, struct mbuf *, struct sockaddr *,
 206                     struct rtentry *);
 207 static void     carp_start(struct ifnet *);
 208 static void     carp_setrun(struct carp_softc *, sa_family_t);
 209 static void     carp_set_state(struct carp_softc *, int);
 210 static int      carp_addrcount(struct carp_if *, struct in_ifaddr *, int);
 211 enum    { CARP_COUNT_MASTER, CARP_COUNT_RUNNING };
 212
 213 static void     carp_multicast_cleanup(struct carp_softc *);
 214 static int      carp_set_addr(struct carp_softc *, struct sockaddr_in *);
 215 static int      carp_del_addr(struct carp_softc *, struct sockaddr_in *);
 216 static void     carp_carpdev_state_locked(struct carp_if *);
 217 static void     carp_sc_state_locked(struct carp_softc *);
 218 #ifdef INET6
 219 static void     carp_send_na(struct carp_softc *);
 220 static int      carp_set_addr6(struct carp_softc *, struct sockaddr_in6 *);
 221 static int      carp_del_addr6(struct carp_softc *, struct sockaddr_in6 *);
 222 static void     carp_multicast6_cleanup(struct carp_softc *);
 223 #endif
 224
 225 static LIST_HEAD(, carp_softc) carpif_list;
 226 static struct mtx carp_mtx;
 227 IFC_SIMPLE_DECLARE(carp, 0);
 228
 229 static eventhandler_tag if_detach_event_tag;
 230
 231 static __inline u_int16_t
 232 carp_cksum(struct mbuf *m, int len)
 233 {
 234         return (in_cksum(m, len));
 235 }
 236
 237 static void
 238 carp_hmac_prepare(struct carp_softc *sc)
 239 {
 240         u_int8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 241         u_int8_t vhid = sc->sc_vhid & 0xff;
 242         struct ifaddr *ifa;
 243         int i;
 244 #ifdef INET6
 245         struct in6_addr in6;
 246 #endif
 247
 248         if (sc->sc_carpdev)
 249                 CARP_SCLOCK(sc);
 250
 251         /* XXX: possible race here */
 252
 253         /* compute ipad from key */
 254         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 255         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 256         for (i = 0; i < sizeof(sc->sc_pad); i++)
 257                 sc->sc_pad[i] ^= 0x36;
 258
 259         /* precompute first part of inner hash */
 260         SHA1Init(&sc->sc_sha1);
 261         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 262         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 263         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 264         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 265 #ifdef INET
 266         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 267                 if (ifa->ifa_addr->sa_family == AF_INET)
 268                         SHA1Update(&sc->sc_sha1,
 269                             (void *)&ifatoia(ifa)->ia_addr.sin_addr.s_addr,
 270                             sizeof(struct in_addr));
 271         }
 272 #endif /* INET */
 273 #ifdef INET6
 274         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 275                 if (ifa->ifa_addr->sa_family == AF_INET6) {
 276                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 277                         in6_clearscope(&in6);
 278                         SHA1Update(&sc->sc_sha1, (void *)&in6, sizeof(in6));
 279                 }
 280         }
 281 #endif /* INET6 */
 282
 283         /* convert ipad to opad */
 284         for (i = 0; i < sizeof(sc->sc_pad); i++)
 285                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 286
 287         if (sc->sc_carpdev)
 288                 CARP_SCUNLOCK(sc);
 289 }
 290
 291 static void
 292 carp_hmac_generate(struct carp_softc *sc, u_int32_t counter[2],
 293     unsigned char md[20])
 294 {
 295         SHA1_CTX sha1ctx;
 296
 297         /* fetch first half of inner hash */
 298         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 299
 300         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 301         SHA1Final(md, &sha1ctx);
 302
 303         /* outer hash */
 304         SHA1Init(&sha1ctx);
 305         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 306         SHA1Update(&sha1ctx, md, 20);
 307         SHA1Final(md, &sha1ctx);
 308 }
 309
 310 static int
 311 carp_hmac_verify(struct carp_softc *sc, u_int32_t counter[2],
 312     unsigned char md[20])
 313 {
 314         unsigned char md2[20];
 315
 316         CARP_SCLOCK_ASSERT(sc);
 317
 318         carp_hmac_generate(sc, counter, md2);
 319
 320         return (bcmp(md, md2, sizeof(md2)));
 321 }
 322
 323 static void
 324 carp_setroute(struct carp_softc *sc, int cmd)
 325 {
 326         struct ifaddr *ifa;
 327         int s;
 328
 329         if (sc->sc_carpdev)
 330                 CARP_SCLOCK_ASSERT(sc);
 331
 332         s = splnet();
 333         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 334                 if (ifa->ifa_addr->sa_family == AF_INET &&
 335                     sc->sc_carpdev != NULL) {
 336                         int count = carp_addrcount(
 337                             (struct carp_if *)sc->sc_carpdev->if_carp,
 338                             ifatoia(ifa), CARP_COUNT_MASTER);
 339
 340                         if ((cmd == RTM_ADD && count == 1) ||
 341                             (cmd == RTM_DELETE && count == 0))
 342                                 rtinit(ifa, cmd, RTF_UP | RTF_HOST);
 343                 }
 344 #ifdef INET6
 345                 if (ifa->ifa_addr->sa_family == AF_INET6) {
 346                         if (cmd == RTM_ADD)
 347                                 in6_ifaddloop(ifa);
 348                         else
 349                                 in6_ifremloop(ifa);
 350                 }
 351 #endif /* INET6 */
 352         }
 353         splx(s);
 354 }
 355
 356 static int
 357 carp_clone_create(struct if_clone *ifc, int unit, caddr_t params)
 358 {
 359
 360         struct carp_softc *sc;
 361         struct ifnet *ifp;
 362
 363         MALLOC(sc, struct carp_softc *, sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
 364         ifp = SC2IFP(sc) = if_alloc(IFT_ETHER);
 365         if (ifp == NULL) {
 366                 FREE(sc, M_CARP);
 367                 return (ENOSPC);
 368         }
 369
 370         sc->sc_flags_backup = 0;
 371         sc->sc_suppress = 0;
 372         sc->sc_advbase = CARP_DFLTINTV;
 373         sc->sc_vhid = -1;       /* required setting */
 374         sc->sc_advskew = 0;
 375         sc->sc_init_counter = 1;
 376         sc->sc_naddrs = sc->sc_naddrs6 = 0; /* M_ZERO? */
 377 #ifdef INET6
 378         sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL;
 379 #endif
 380         sc->sc_imo.imo_membership = (struct in_multi **)malloc(
 381             (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
 382             M_WAITOK);
 383         sc->sc_imo.imo_mfilters = NULL;
 384         sc->sc_imo.imo_max_memberships = IP_MIN_MEMBERSHIPS;
 385         sc->sc_imo.imo_multicast_vif = -1;
 386
 387         callout_init(&sc->sc_ad_tmo, NET_CALLOUT_MPSAFE);
 388         callout_init(&sc->sc_md_tmo, NET_CALLOUT_MPSAFE);
 389         callout_init(&sc->sc_md6_tmo, NET_CALLOUT_MPSAFE);
 390
 391         ifp->if_softc = sc;
 392         if_initname(ifp, CARP_IFNAME, unit);
 393         ifp->if_mtu = ETHERMTU;
 394         ifp->if_flags = IFF_LOOPBACK;
 395         ifp->if_ioctl = carp_ioctl;
 396         ifp->if_output = carp_looutput;
 397         ifp->if_start = carp_start;
 398         ifp->if_type = IFT_CARP;
 399         ifp->if_snd.ifq_maxlen = ifqmaxlen;
 400         ifp->if_hdrlen = 0;
 401         if_attach(ifp);
 402         bpfattach(SC2IFP(sc), DLT_NULL, sizeof(u_int32_t));
 403         mtx_lock(&carp_mtx);
 404         LIST_INSERT_HEAD(&carpif_list, sc, sc_next);
 405         mtx_unlock(&carp_mtx);
 406         return (0);
 407 }
 408
 409 static void
 410 carp_clone_destroy(struct ifnet *ifp)
 411 {
 412         struct carp_softc *sc = ifp->if_softc;
 413
 414         if (sc->sc_carpdev)
 415                 CARP_SCLOCK(sc);
 416         carpdetach(sc, 1);      /* Returns unlocked. */
 417
 418         mtx_lock(&carp_mtx);
 419         LIST_REMOVE(sc, sc_next);
 420         mtx_unlock(&carp_mtx);
 421         bpfdetach(ifp);
 422         if_detach(ifp);
 423         if_free_type(ifp, IFT_ETHER);
 424         free(sc->sc_imo.imo_membership, M_CARP);
 425         free(sc, M_CARP);
 426 }
 427
 428 /*
 429  * This function can be called on CARP interface destroy path,
 430  * and in case of the removal of the underlying interface as
 431  * well. We differentiate these two cases. In the latter case
 432  * we do not cleanup our multicast memberships, since they
 433  * are already freed. Also, in the latter case we do not
 434  * release the lock on return, because the function will be
 435  * called once more, for another CARP instance on the same
 436  * interface.
 437  */
 438 static void
 439 carpdetach(struct carp_softc *sc, int unlock)
 440 {
 441         struct carp_if *cif;
 442
 443         callout_stop(&sc->sc_ad_tmo);
 444         callout_stop(&sc->sc_md_tmo);
 445         callout_stop(&sc->sc_md6_tmo);
 446
 447         if (sc->sc_suppress)
 448                 carp_suppress_preempt--;
 449         sc->sc_suppress = 0;
 450
 451         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS)
 452                 carp_suppress_preempt--;
 453         sc->sc_sendad_errors = 0;
 454
 455         carp_set_state(sc, INIT);
 456         SC2IFP(sc)->if_flags &= ~IFF_UP;
 457         carp_setrun(sc, 0);
 458         if (unlock)
 459                 carp_multicast_cleanup(sc);
 460 #ifdef INET6
 461         carp_multicast6_cleanup(sc);
 462 #endif
 463
 464         if (sc->sc_carpdev != NULL) {
 465                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
 466                 CARP_LOCK_ASSERT(cif);
 467                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
 468                 if (!--cif->vhif_nvrs) {
 469                         ifpromisc(sc->sc_carpdev, 0);
 470                         sc->sc_carpdev->if_carp = NULL;
 471                         CARP_LOCK_DESTROY(cif);
 472                         FREE(cif, M_IFADDR);
 473                 } else if (unlock)
 474                         CARP_UNLOCK(cif);
 475                 sc->sc_carpdev = NULL;
 476         }
 477 }
 478
 479 /* Detach an interface from the carp. */
 480 static void
 481 carp_ifdetach(void *arg __unused, struct ifnet *ifp)
 482 {
 483         struct carp_if *cif = (struct carp_if *)ifp->if_carp;
 484         struct carp_softc *sc, *nextsc;
 485
 486         if (cif == NULL)
 487                 return;
 488
 489         /*
 490          * XXX: At the end of for() cycle the lock will be destroyed.
 491          */
 492         CARP_LOCK(cif);
 493         for (sc = TAILQ_FIRST(&cif->vhif_vrs); sc; sc = nextsc) {
 494                 nextsc = TAILQ_NEXT(sc, sc_list);
 495                 carpdetach(sc, 0);
 496         }
 497 }
 498
 499 /*
 500  * process input packet.
 501  * we have rearranged checks order compared to the rfc,
 502  * but it seems more efficient this way or not possible otherwise.
 503  */
 504 void
 505 carp_input(struct mbuf *m, int hlen)
 506 {
 507         struct ip *ip = mtod(m, struct ip *);
 508         struct carp_header *ch;
 509         int iplen, len;
 510
 511         carpstats.carps_ipackets++;
 512
 513         if (!carp_opts[CARPCTL_ALLOW]) {
 514                 m_freem(m);
 515                 return;
 516         }
 517
 518         /* check if received on a valid carp interface */
 519         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 520                 carpstats.carps_badif++;
 521                 CARP_LOG("carp_input: packet received on non-carp "
 522                     "interface: %s\n",
 523                     m->m_pkthdr.rcvif->if_xname);
 524                 m_freem(m);
 525                 return;
 526         }
 527
 528         /* verify that the IP TTL is 255.  */
 529         if (ip->ip_ttl != CARP_DFLTTL) {
 530                 carpstats.carps_badttl++;
 531                 CARP_LOG("carp_input: received ttl %d != 255i on %s\n",
 532                     ip->ip_ttl,
 533                     m->m_pkthdr.rcvif->if_xname);
 534                 m_freem(m);
 535                 return;
 536         }
 537
 538         iplen = ip->ip_hl << 2;
 539
 540         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 541                 carpstats.carps_badlen++;
 542                 CARP_LOG("carp_input: received len %zd < "
 543                     "sizeof(struct carp_header)\n",
 544                     m->m_len - sizeof(struct ip));
 545                 m_freem(m);
 546                 return;
 547         }
 548
 549         if (iplen + sizeof(*ch) < m->m_len) {
 550                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 551                         carpstats.carps_hdrops++;
 552                         CARP_LOG("carp_input: pullup failed\n");
 553                         return;
 554                 }
 555                 ip = mtod(m, struct ip *);
 556         }
 557         ch = (struct carp_header *)((char *)ip + iplen);
 558
 559         /*
 560          * verify that the received packet length is
 561          * equal to the CARP header
 562          */
 563         len = iplen + sizeof(*ch);
 564         if (len > m->m_pkthdr.len) {
 565                 carpstats.carps_badlen++;
 566                 CARP_LOG("carp_input: packet too short %d on %s\n",
 567                     m->m_pkthdr.len,
 568                     m->m_pkthdr.rcvif->if_xname);
 569                 m_freem(m);
 570                 return;
 571         }
 572
 573         if ((m = m_pullup(m, len)) == NULL) {
 574                 carpstats.carps_hdrops++;
 575                 return;
 576         }
 577         ip = mtod(m, struct ip *);
 578         ch = (struct carp_header *)((char *)ip + iplen);
 579
 580         /* verify the CARP checksum */
 581         m->m_data += iplen;
 582         if (carp_cksum(m, len - iplen)) {
 583                 carpstats.carps_badsum++;
 584                 CARP_LOG("carp_input: checksum failed on %s\n",
 585                     m->m_pkthdr.rcvif->if_xname);
 586                 m_freem(m);
 587                 return;
 588         }
 589         m->m_data -= iplen;
 590
 591         carp_input_c(m, ch, AF_INET);
 592 }
 593
 594 #ifdef INET6
 595 int
 596 carp6_input(struct mbuf **mp, int *offp, int proto)
 597 {
 598         struct mbuf *m = *mp;
 599         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 600         struct carp_header *ch;
 601         u_int len;
 602
 603         carpstats.carps_ipackets6++;
 604
 605         if (!carp_opts[CARPCTL_ALLOW]) {
 606                 m_freem(m);
 607                 return (IPPROTO_DONE);
 608         }
 609
 610         /* check if received on a valid carp interface */
 611         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 612                 carpstats.carps_badif++;
 613                 CARP_LOG("carp6_input: packet received on non-carp "
 614                     "interface: %s\n",
 615                     m->m_pkthdr.rcvif->if_xname);
 616                 m_freem(m);
 617                 return (IPPROTO_DONE);
 618         }
 619
 620         /* verify that the IP TTL is 255 */
 621         if (ip6->ip6_hlim != CARP_DFLTTL) {
 622                 carpstats.carps_badttl++;
 623                 CARP_LOG("carp6_input: received ttl %d != 255 on %s\n",
 624                     ip6->ip6_hlim,
 625                     m->m_pkthdr.rcvif->if_xname);
 626                 m_freem(m);
 627                 return (IPPROTO_DONE);
 628         }
 629
 630         /* verify that we have a complete carp packet */
 631         len = m->m_len;
 632         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 633         if (ch == NULL) {
 634                 carpstats.carps_badlen++;
 635                 CARP_LOG("carp6_input: packet size %u too small\n", len);
 636                 return (IPPROTO_DONE);
 637         }
 638
 639
 640         /* verify the CARP checksum */
 641         m->m_data += *offp;
 642         if (carp_cksum(m, sizeof(*ch))) {
 643                 carpstats.carps_badsum++;
 644                 CARP_LOG("carp6_input: checksum failed, on %s\n",
 645                     m->m_pkthdr.rcvif->if_xname);
 646                 m_freem(m);
 647                 return (IPPROTO_DONE);
 648         }
 649         m->m_data -= *offp;
 650
 651         carp_input_c(m, ch, AF_INET6);
 652         return (IPPROTO_DONE);
 653 }
 654 #endif /* INET6 */
 655
 656 static void
 657 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 658 {
 659         struct ifnet *ifp = m->m_pkthdr.rcvif;
 660         struct carp_softc *sc;
 661         u_int64_t tmp_counter;
 662         struct timeval sc_tv, ch_tv;
 663
 664         /* verify that the VHID is valid on the receiving interface */
 665         CARP_LOCK(ifp->if_carp);
 666         TAILQ_FOREACH(sc, &((struct carp_if *)ifp->if_carp)->vhif_vrs, sc_list)
 667                 if (sc->sc_vhid == ch->carp_vhid)
 668                         break;
 669
 670         if (!sc || !((SC2IFP(sc)->if_flags & IFF_UP) &&
 671             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING))) {
 672                 carpstats.carps_badvhid++;
 673                 CARP_UNLOCK(ifp->if_carp);
 674                 m_freem(m);
 675                 return;
 676         }
 677
 678         getmicrotime(&SC2IFP(sc)->if_lastchange);
 679         SC2IFP(sc)->if_ipackets++;
 680         SC2IFP(sc)->if_ibytes += m->m_pkthdr.len;
 681
 682         if (bpf_peers_present(SC2IFP(sc)->if_bpf)) {
 683                 struct ip *ip = mtod(m, struct ip *);
 684                 uint32_t af1 = af;
 685
 686                 /* BPF wants net byte order */
 687                 ip->ip_len = htons(ip->ip_len + (ip->ip_hl << 2));
 688                 ip->ip_off = htons(ip->ip_off);
 689                 bpf_mtap2(SC2IFP(sc)->if_bpf, &af1, sizeof(af1), m);
 690         }
 691
 692         /* verify the CARP version. */
 693         if (ch->carp_version != CARP_VERSION) {
 694                 carpstats.carps_badver++;
 695                 SC2IFP(sc)->if_ierrors++;
 696                 CARP_UNLOCK(ifp->if_carp);
 697                 CARP_LOG("%s; invalid version %d\n",
 698                     SC2IFP(sc)->if_xname,
 699                     ch->carp_version);
 700                 m_freem(m);
 701                 return;
 702         }
 703
 704         /* verify the hash */
 705         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 706                 carpstats.carps_badauth++;
 707                 SC2IFP(sc)->if_ierrors++;
 708                 CARP_UNLOCK(ifp->if_carp);
 709                 CARP_LOG("%s: incorrect hash\n", SC2IFP(sc)->if_xname);
 710                 m_freem(m);
 711                 return;
 712         }
 713
 714         tmp_counter = ntohl(ch->carp_counter[0]);
 715         tmp_counter = tmp_counter<<32;
 716         tmp_counter += ntohl(ch->carp_counter[1]);
 717
 718         /* XXX Replay protection goes here */
 719
 720         sc->sc_init_counter = 0;
 721         sc->sc_counter = tmp_counter;
 722
 723         sc_tv.tv_sec = sc->sc_advbase;
 724         if (carp_suppress_preempt && sc->sc_advskew <  240)
 725                 sc_tv.tv_usec = 240 * 1000000 / 256;
 726         else
 727                 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256;
 728         ch_tv.tv_sec = ch->carp_advbase;
 729         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 730
 731         switch (sc->sc_state) {
 732         case INIT:
 733                 break;
 734         case MASTER:
 735                 /*
 736                  * If we receive an advertisement from a master who's going to
 737                  * be more frequent than us, go into BACKUP state.
 738                  */
 739                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 740                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 741                         callout_stop(&sc->sc_ad_tmo);
 742                         CARP_DEBUG("%s: MASTER -> BACKUP "
 743                            "(more frequent advertisement received)\n",
 744                            SC2IFP(sc)->if_xname);
 745                         carp_set_state(sc, BACKUP);
 746                         carp_setrun(sc, 0);
 747                         carp_setroute(sc, RTM_DELETE);
 748                 }
 749                 break;
 750         case BACKUP:
 751                 /*
 752                  * If we're pre-empting masters who advertise slower than us,
 753                  * and this one claims to be slower, treat him as down.
 754                  */
 755                 if (carp_opts[CARPCTL_PREEMPT] &&
 756                     timevalcmp(&sc_tv, &ch_tv, <)) {
 757                         CARP_DEBUG("%s: BACKUP -> MASTER "
 758                             "(preempting a slower master)\n",
 759                             SC2IFP(sc)->if_xname);
 760                         carp_master_down_locked(sc);
 761                         break;
 762                 }
 763
 764                 /*
 765                  *  If the master is going to advertise at such a low frequency
 766                  *  that he's guaranteed to time out, we'd might as well just
 767                  *  treat him as timed out now.
 768                  */
 769                 sc_tv.tv_sec = sc->sc_advbase * 3;
 770                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 771                         CARP_DEBUG("%s: BACKUP -> MASTER "
 772                             "(master timed out)\n",
 773                             SC2IFP(sc)->if_xname);
 774                         carp_master_down_locked(sc);
 775                         break;
 776                 }
 777
 778                 /*
 779                  * Otherwise, we reset the counter and wait for the next
 780                  * advertisement.
 781                  */
 782                 carp_setrun(sc, af);
 783                 break;
 784         }
 785
 786         CARP_UNLOCK(ifp->if_carp);
 787
 788         m_freem(m);
 789         return;
 790 }
 791
 792 static int
 793 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 794 {
 795         struct m_tag *mtag;
 796         struct ifnet *ifp = SC2IFP(sc);
 797
 798         if (sc->sc_init_counter) {
 799                 /* this could also be seconds since unix epoch */
 800                 sc->sc_counter = arc4random();
 801                 sc->sc_counter = sc->sc_counter << 32;
 802                 sc->sc_counter += arc4random();
 803         } else
 804                 sc->sc_counter++;
 805
 806         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 807         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 808
 809         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 810
 811         /* Tag packet for carp_output */
 812         mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct ifnet *), M_NOWAIT);
 813         if (mtag == NULL) {
 814                 m_freem(m);
 815                 SC2IFP(sc)->if_oerrors++;
 816                 return (ENOMEM);
 817         }
 818         bcopy(&ifp, (caddr_t)(mtag + 1), sizeof(struct ifnet *));
 819         m_tag_prepend(m, mtag);
 820
 821         return (0);
 822 }
 823
 824 static void
 825 carp_send_ad_all(void)
 826 {
 827         struct carp_softc *sc;
 828
 829         mtx_lock(&carp_mtx);
 830         LIST_FOREACH(sc, &carpif_list, sc_next) {
 831                 if (sc->sc_carpdev == NULL)
 832                         continue;
 833                 CARP_SCLOCK(sc);
 834                 if ((SC2IFP(sc)->if_flags & IFF_UP) &&
 835                     (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING) &&
 836                      sc->sc_state == MASTER)
 837                         carp_send_ad_locked(sc);
 838                 CARP_SCUNLOCK(sc);
 839         }
 840         mtx_unlock(&carp_mtx);
 841 }
 842
 843 static void
 844 carp_send_ad(void *v)
 845 {
 846         struct carp_softc *sc = v;
 847
 848         CARP_SCLOCK(sc);
 849         carp_send_ad_locked(sc);
 850         CARP_SCUNLOCK(sc);
 851 }
 852
 853 static void
 854 carp_send_ad_locked(struct carp_softc *sc)
 855 {
 856         struct carp_header ch;
 857         struct timeval tv;
 858         struct carp_header *ch_ptr;
 859         struct mbuf *m;
 860         int len, advbase, advskew;
 861
 862         CARP_SCLOCK_ASSERT(sc);
 863
 864         /* bow out if we've lost our UPness or RUNNINGuiness */
 865         if (!((SC2IFP(sc)->if_flags & IFF_UP) &&
 866             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING))) {
 867                 advbase = 255;
 868                 advskew = 255;
 869         } else {
 870                 advbase = sc->sc_advbase;
 871                 if (!carp_suppress_preempt || sc->sc_advskew > 240)
 872                         advskew = sc->sc_advskew;
 873                 else
 874                         advskew = 240;
 875                 tv.tv_sec = advbase;
 876                 tv.tv_usec = advskew * 1000000 / 256;
 877         }
 878
 879         ch.carp_version = CARP_VERSION;
 880         ch.carp_type = CARP_ADVERTISEMENT;
 881         ch.carp_vhid = sc->sc_vhid;
 882         ch.carp_advbase = advbase;
 883         ch.carp_advskew = advskew;
 884         ch.carp_authlen = 7;    /* XXX DEFINE */
 885         ch.carp_pad1 = 0;       /* must be zero */
 886         ch.carp_cksum = 0;
 887
 888 #ifdef INET
 889         if (sc->sc_ia) {
 890                 struct ip *ip;
 891
 892                 MGETHDR(m, M_DONTWAIT, MT_HEADER);
 893                 if (m == NULL) {
 894                         SC2IFP(sc)->if_oerrors++;
 895                         carpstats.carps_onomem++;
 896                         /* XXX maybe less ? */
 897                         if (advbase != 255 || advskew != 255)
 898                                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
 899                                     carp_send_ad, sc);
 900                         return;
 901                 }
 902                 len = sizeof(*ip) + sizeof(ch);
 903                 m->m_pkthdr.len = len;
 904                 m->m_pkthdr.rcvif = NULL;
 905                 m->m_len = len;
 906                 MH_ALIGN(m, m->m_len);
 907                 m->m_flags |= M_MCAST;
 908                 ip = mtod(m, struct ip *);
 909                 ip->ip_v = IPVERSION;
 910                 ip->ip_hl = sizeof(*ip) >> 2;
 911                 ip->ip_tos = IPTOS_LOWDELAY;
 912                 ip->ip_len = len;
 913                 ip->ip_id = ip_newid();
 914                 ip->ip_off = IP_DF;
 915                 ip->ip_ttl = CARP_DFLTTL;
 916                 ip->ip_p = IPPROTO_CARP;
 917                 ip->ip_sum = 0;
 918                 ip->ip_src.s_addr = sc->sc_ia->ia_addr.sin_addr.s_addr;
 919                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 920
 921                 ch_ptr = (struct carp_header *)(&ip[1]);
 922                 bcopy(&ch, ch_ptr, sizeof(ch));
 923                 if (carp_prepare_ad(m, sc, ch_ptr))
 924                         return;
 925
 926                 m->m_data += sizeof(*ip);
 927                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip));
 928                 m->m_data -= sizeof(*ip);
 929
 930                 getmicrotime(&SC2IFP(sc)->if_lastchange);
 931                 SC2IFP(sc)->if_opackets++;
 932                 SC2IFP(sc)->if_obytes += len;
 933                 carpstats.carps_opackets++;
 934
 935                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, NULL)) {
 936                         SC2IFP(sc)->if_oerrors++;
 937                         if (sc->sc_sendad_errors < INT_MAX)
 938                                 sc->sc_sendad_errors++;
 939                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 940                                 carp_suppress_preempt++;
 941                                 if (carp_suppress_preempt == 1) {
 942                                         CARP_SCUNLOCK(sc);
 943                                         carp_send_ad_all();
 944                                         CARP_SCLOCK(sc);
 945                                 }
 946                         }
 947                         sc->sc_sendad_success = 0;
 948                 } else {
 949                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 950                                 if (++sc->sc_sendad_success >=
 951                                     CARP_SENDAD_MIN_SUCCESS) {
 952                                         carp_suppress_preempt--;
 953                                         sc->sc_sendad_errors = 0;
 954                                 }
 955                         } else
 956                                 sc->sc_sendad_errors = 0;
 957                 }
 958         }
 959 #endif /* INET */
 960 #ifdef INET6
 961         if (sc->sc_ia6) {
 962                 struct ip6_hdr *ip6;
 963
 964                 MGETHDR(m, M_DONTWAIT, MT_HEADER);
 965                 if (m == NULL) {
 966                         SC2IFP(sc)->if_oerrors++;
 967                         carpstats.carps_onomem++;
 968                         /* XXX maybe less ? */
 969                         if (advbase != 255 || advskew != 255)
 970                                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
 971                                     carp_send_ad, sc);
 972                         return;
 973                 }
 974                 len = sizeof(*ip6) + sizeof(ch);
 975                 m->m_pkthdr.len = len;
 976                 m->m_pkthdr.rcvif = NULL;
 977                 m->m_len = len;
 978                 MH_ALIGN(m, m->m_len);
 979                 m->m_flags |= M_MCAST;
 980                 ip6 = mtod(m, struct ip6_hdr *);
 981                 bzero(ip6, sizeof(*ip6));
 982                 ip6->ip6_vfc |= IPV6_VERSION;
 983                 ip6->ip6_hlim = CARP_DFLTTL;
 984                 ip6->ip6_nxt = IPPROTO_CARP;
 985                 bcopy(&sc->sc_ia6->ia_addr.sin6_addr, &ip6->ip6_src,
 986                     sizeof(struct in6_addr));
 987                 /* set the multicast destination */
 988
 989                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
 990                 ip6->ip6_dst.s6_addr8[15] = 0x12;
 991                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
 992                         SC2IFP(sc)->if_oerrors++;
 993                         m_freem(m);
 994                         CARP_LOG("%s: in6_setscope failed\n", __func__);
 995                         return;
 996                 }
 997
 998                 ch_ptr = (struct carp_header *)(&ip6[1]);
 999                 bcopy(&ch, ch_ptr, sizeof(ch));
1000                 if (carp_prepare_ad(m, sc, ch_ptr))
1001                         return;
1002
1003                 m->m_data += sizeof(*ip6);
1004                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6));
1005                 m->m_data -= sizeof(*ip6);
1006
1007                 getmicrotime(&SC2IFP(sc)->if_lastchange);
1008                 SC2IFP(sc)->if_opackets++;
1009                 SC2IFP(sc)->if_obytes += len;
1010                 carpstats.carps_opackets6++;
1011
1012                 if (ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL)) {
1013                         SC2IFP(sc)->if_oerrors++;
1014                         if (sc->sc_sendad_errors < INT_MAX)
1015                                 sc->sc_sendad_errors++;
1016                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
1017                                 carp_suppress_preempt++;
1018                                 if (carp_suppress_preempt == 1) {
1019                                         CARP_SCUNLOCK(sc);
1020                                         carp_send_ad_all();
1021                                         CARP_SCLOCK(sc);
1022                                 }
1023                         }
1024                         sc->sc_sendad_success = 0;
1025                 } else {
1026                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
1027                                 if (++sc->sc_sendad_success >=
1028                                     CARP_SENDAD_MIN_SUCCESS) {
1029                                         carp_suppress_preempt--;
1030                                         sc->sc_sendad_errors = 0;
1031                                 }
1032                         } else
1033                                 sc->sc_sendad_errors = 0;
1034                 }
1035         }
1036 #endif /* INET6 */
1037
1038         if (advbase != 255 || advskew != 255)
1039                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1040                     carp_send_ad, sc);
1041
1042 }
1043
1044 /*
1045  * Broadcast a gratuitous ARP request containing
1046  * the virtual router MAC address for each IP address
1047  * associated with the virtual router.
1048  */
1049 static void
1050 carp_send_arp(struct carp_softc *sc)
1051 {
1052         struct ifaddr *ifa;
1053
1054         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1055
1056                 if (ifa->ifa_addr->sa_family != AF_INET)
1057                         continue;
1058
1059 /*              arprequest(sc->sc_carpdev, &in, &in, IF_LLADDR(sc->sc_ifp)); */
1060                 arp_ifinit2(sc->sc_carpdev, ifa, IF_LLADDR(sc->sc_ifp));
1061
1062                 DELAY(1000);    /* XXX */
1063         }
1064 }
1065
1066 #ifdef INET6
1067 static void
1068 carp_send_na(struct carp_softc *sc)
1069 {
1070         struct ifaddr *ifa;
1071         struct in6_addr *in6;
1072         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1073
1074         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1075
1076                 if (ifa->ifa_addr->sa_family != AF_INET6)
1077                         continue;
1078
1079                 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr;
1080                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1081                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1082                 DELAY(1000);    /* XXX */
1083         }
1084 }
1085 #endif /* INET6 */
1086
1087 static int
1088 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type)
1089 {
1090         struct carp_softc *vh;
1091         struct ifaddr *ifa;
1092         int count = 0;
1093
1094         CARP_LOCK_ASSERT(cif);
1095
1096         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1097                 if ((type == CARP_COUNT_RUNNING &&
1098                     (SC2IFP(vh)->if_flags & IFF_UP) &&
1099                     (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING)) ||
1100                     (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) {
1101                         TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1102                             ifa_list) {
1103                                 if (ifa->ifa_addr->sa_family == AF_INET &&
1104                                     ia->ia_addr.sin_addr.s_addr ==
1105                                     ifatoia(ifa)->ia_addr.sin_addr.s_addr)
1106                                         count++;
1107                         }
1108                 }
1109         }
1110         return (count);
1111 }
1112
1113 int
1114 carp_iamatch(void *v, struct in_ifaddr *ia,
1115     struct in_addr *isaddr, u_int8_t **enaddr)
1116 {
1117         struct carp_if *cif = v;
1118         struct carp_softc *vh;
1119         int index, count = 0;
1120         struct ifaddr *ifa;
1121
1122         CARP_LOCK(cif);
1123
1124         if (carp_opts[CARPCTL_ARPBALANCE]) {
1125                 /*
1126                  * XXX proof of concept implementation.
1127                  * We use the source ip to decide which virtual host should
1128                  * handle the request. If we're master of that virtual host,
1129                  * then we respond, otherwise, just drop the arp packet on
1130                  * the floor.
1131                  */
1132                 count = carp_addrcount(cif, ia, CARP_COUNT_RUNNING);
1133                 if (count == 0) {
1134                         /* should never reach this */
1135                         CARP_UNLOCK(cif);
1136                         return (0);
1137                 }
1138
1139                 /* this should be a hash, like pf_hash() */
1140                 index = ntohl(isaddr->s_addr) % count;
1141                 count = 0;
1142
1143                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1144                         if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1145                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING)) {
1146                                 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1147                                     ifa_list) {
1148                                         if (ifa->ifa_addr->sa_family ==
1149                                             AF_INET &&
1150                                             ia->ia_addr.sin_addr.s_addr ==
1151                                             ifatoia(ifa)->ia_addr.sin_addr.s_addr) {
1152                                                 if (count == index) {
1153                                                         if (vh->sc_state ==
1154                                                             MASTER) {
1155                                                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1156                                                                 CARP_UNLOCK(cif);
1157                                                                 return (1);
1158                                                         } else {
1159                                                                 CARP_UNLOCK(cif);
1160                                                                 return (0);
1161                                                         }
1162                                                 }
1163                                                 count++;
1164                                         }
1165                                 }
1166                         }
1167                 }
1168         } else {
1169                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1170                         if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1171                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1172                             ia->ia_ifp == SC2IFP(vh) &&
1173                             vh->sc_state == MASTER) {
1174                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1175                                 CARP_UNLOCK(cif);
1176                                 return (1);
1177                         }
1178                 }
1179         }
1180         CARP_UNLOCK(cif);
1181         return (0);
1182 }
1183
1184 #ifdef INET6
1185 struct ifaddr *
1186 carp_iamatch6(void *v, struct in6_addr *taddr)
1187 {
1188         struct carp_if *cif = v;
1189         struct carp_softc *vh;
1190         struct ifaddr *ifa;
1191
1192         CARP_LOCK(cif);
1193         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1194                 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist, ifa_list) {
1195                         if (IN6_ARE_ADDR_EQUAL(taddr,
1196                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1197                             (SC2IFP(vh)->if_flags & IFF_UP) &&
1198                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1199                             vh->sc_state == MASTER) {
1200                                 CARP_UNLOCK(cif);
1201                                 return (ifa);
1202                         }
1203                 }
1204         }
1205         CARP_UNLOCK(cif);
1206
1207         return (NULL);
1208 }
1209
1210 void *
1211 carp_macmatch6(void *v, struct mbuf *m, const struct in6_addr *taddr)
1212 {
1213         struct m_tag *mtag;
1214         struct carp_if *cif = v;
1215         struct carp_softc *sc;
1216         struct ifaddr *ifa;
1217
1218         CARP_LOCK(cif);
1219         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) {
1220                 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1221                         if (IN6_ARE_ADDR_EQUAL(taddr,
1222                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1223                             (SC2IFP(sc)->if_flags & IFF_UP) &&
1224                             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING)) {
1225                                 struct ifnet *ifp = SC2IFP(sc);
1226                                 mtag = m_tag_get(PACKET_TAG_CARP,
1227                                     sizeof(struct ifnet *), M_NOWAIT);
1228                                 if (mtag == NULL) {
1229                                         /* better a bit than nothing */
1230                                         CARP_UNLOCK(cif);
1231                                         return (IF_LLADDR(sc->sc_ifp));
1232                                 }
1233                                 bcopy(&ifp, (caddr_t)(mtag + 1),
1234                                     sizeof(struct ifnet *));
1235                                 m_tag_prepend(m, mtag);
1236
1237                                 CARP_UNLOCK(cif);
1238                                 return (IF_LLADDR(sc->sc_ifp));
1239                         }
1240                 }
1241         }
1242         CARP_UNLOCK(cif);
1243
1244         return (NULL);
1245 }
1246 #endif
1247
1248 struct ifnet *
1249 carp_forus(void *v, void *dhost)
1250 {
1251         struct carp_if *cif = v;
1252         struct carp_softc *vh;
1253         u_int8_t *ena = dhost;
1254
1255         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1256                 return (NULL);
1257
1258         CARP_LOCK(cif);
1259         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list)
1260                 if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1261                     (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1262                     vh->sc_state == MASTER &&
1263                     !bcmp(dhost, IF_LLADDR(vh->sc_ifp), ETHER_ADDR_LEN)) {
1264                         CARP_UNLOCK(cif);
1265                         return (SC2IFP(vh));
1266                 }
1267
1268         CARP_UNLOCK(cif);
1269         return (NULL);
1270 }
1271
1272 static void
1273 carp_master_down(void *v)
1274 {
1275         struct carp_softc *sc = v;
1276
1277         CARP_SCLOCK(sc);
1278         carp_master_down_locked(sc);
1279         CARP_SCUNLOCK(sc);
1280 }
1281
1282 static void
1283 carp_master_down_locked(struct carp_softc *sc)
1284 {
1285         if (sc->sc_carpdev)
1286                 CARP_SCLOCK_ASSERT(sc);
1287
1288         switch (sc->sc_state) {
1289         case INIT:
1290                 printf("%s: master_down event in INIT state\n",
1291                     SC2IFP(sc)->if_xname);
1292                 break;
1293         case MASTER:
1294                 break;
1295         case BACKUP:
1296                 carp_set_state(sc, MASTER);
1297                 carp_send_ad_locked(sc);
1298                 carp_send_arp(sc);
1299 #ifdef INET6
1300                 carp_send_na(sc);
1301 #endif /* INET6 */
1302                 carp_setrun(sc, 0);
1303                 carp_setroute(sc, RTM_ADD);
1304                 break;
1305         }
1306 }
1307
1308 /*
1309  * When in backup state, af indicates whether to reset the master down timer
1310  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1311  */
1312 static void
1313 carp_setrun(struct carp_softc *sc, sa_family_t af)
1314 {
1315         struct timeval tv;
1316
1317         if (sc->sc_carpdev == NULL) {
1318                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1319                 carp_set_state(sc, INIT);
1320                 return;
1321         } else
1322                 CARP_SCLOCK_ASSERT(sc);
1323
1324         if (SC2IFP(sc)->if_flags & IFF_UP &&
1325             sc->sc_vhid > 0 && (sc->sc_naddrs || sc->sc_naddrs6))
1326                 SC2IFP(sc)->if_drv_flags |= IFF_DRV_RUNNING;
1327         else {
1328                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1329                 carp_setroute(sc, RTM_DELETE);
1330                 return;
1331         }
1332
1333         switch (sc->sc_state) {
1334         case INIT:
1335                 if (carp_opts[CARPCTL_PREEMPT] && !carp_suppress_preempt) {
1336                         carp_send_ad_locked(sc);
1337                         carp_send_arp(sc);
1338 #ifdef INET6
1339                         carp_send_na(sc);
1340 #endif /* INET6 */
1341                         CARP_DEBUG("%s: INIT -> MASTER (preempting)\n",
1342                             SC2IFP(sc)->if_xname);
1343                         carp_set_state(sc, MASTER);
1344                         carp_setroute(sc, RTM_ADD);
1345                 } else {
1346                         CARP_DEBUG("%s: INIT -> BACKUP\n", SC2IFP(sc)->if_xname);
1347                         carp_set_state(sc, BACKUP);
1348                         carp_setroute(sc, RTM_DELETE);
1349                         carp_setrun(sc, 0);
1350                 }
1351                 break;
1352         case BACKUP:
1353                 callout_stop(&sc->sc_ad_tmo);
1354                 tv.tv_sec = 3 * sc->sc_advbase;
1355                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1356                 switch (af) {
1357 #ifdef INET
1358                 case AF_INET:
1359                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1360                             carp_master_down, sc);
1361                         break;
1362 #endif /* INET */
1363 #ifdef INET6
1364                 case AF_INET6:
1365                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1366                             carp_master_down, sc);
1367                         break;
1368 #endif /* INET6 */
1369                 default:
1370                         if (sc->sc_naddrs)
1371                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1372                                     carp_master_down, sc);
1373                         if (sc->sc_naddrs6)
1374                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1375                                     carp_master_down, sc);
1376                         break;
1377                 }
1378                 break;
1379         case MASTER:
1380                 tv.tv_sec = sc->sc_advbase;
1381                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1382                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1383                     carp_send_ad, sc);
1384                 break;
1385         }
1386 }
1387
1388 static void
1389 carp_multicast_cleanup(struct carp_softc *sc)
1390 {
1391         struct ip_moptions *imo = &sc->sc_imo;
1392         u_int16_t n = imo->imo_num_memberships;
1393
1394         /* Clean up our own multicast memberships */
1395         while (n-- > 0) {
1396                 if (imo->imo_membership[n] != NULL) {
1397                         in_delmulti(imo->imo_membership[n]);
1398                         imo->imo_membership[n] = NULL;
1399                 }
1400         }
1401         KASSERT(imo->imo_mfilters == NULL,
1402            ("%s: imo_mfilters != NULL", __func__));
1403         imo->imo_num_memberships = 0;
1404         imo->imo_multicast_ifp = NULL;
1405 }
1406
1407 #ifdef INET6
1408 static void
1409 carp_multicast6_cleanup(struct carp_softc *sc)
1410 {
1411         struct ip6_moptions *im6o = &sc->sc_im6o;
1412
1413         while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1414                 struct in6_multi_mship *imm =
1415                     LIST_FIRST(&im6o->im6o_memberships);
1416
1417                 LIST_REMOVE(imm, i6mm_chain);
1418                 in6_leavegroup(imm);
1419         }
1420         im6o->im6o_multicast_ifp = NULL;
1421 }
1422 #endif
1423
1424 static int
1425 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1426 {
1427         struct ifnet *ifp;
1428         struct carp_if *cif;
1429         struct in_ifaddr *ia, *ia_if;
1430         struct ip_moptions *imo = &sc->sc_imo;
1431         struct in_addr addr;
1432         u_long iaddr = htonl(sin->sin_addr.s_addr);
1433         int own, error;
1434
1435         if (sin->sin_addr.s_addr == 0) {
1436                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1437                         carp_set_state(sc, INIT);
1438                 if (sc->sc_naddrs)
1439                         SC2IFP(sc)->if_flags |= IFF_UP;
1440                 carp_setrun(sc, 0);
1441                 return (0);
1442         }
1443
1444         /* we have to do it by hands to check we won't match on us */
1445         ia_if = NULL; own = 0;
1446         TAILQ_FOREACH(ia, &in_ifaddrhead, ia_link) {
1447                 /* and, yeah, we need a multicast-capable iface too */
1448                 if (ia->ia_ifp != SC2IFP(sc) &&
1449                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1450                     (iaddr & ia->ia_subnetmask) == ia->ia_subnet) {
1451                         if (!ia_if)
1452                                 ia_if = ia;
1453                         if (sin->sin_addr.s_addr ==
1454                             ia->ia_addr.sin_addr.s_addr)
1455                                 own++;
1456                 }
1457         }
1458
1459         if (!ia_if)
1460                 return (EADDRNOTAVAIL);
1461
1462         ia = ia_if;
1463         ifp = ia->ia_ifp;
1464
1465         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1466             (imo->imo_multicast_ifp && imo->imo_multicast_ifp != ifp))
1467                 return (EADDRNOTAVAIL);
1468
1469         if (imo->imo_num_memberships == 0) {
1470                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1471                 if ((imo->imo_membership[0] = in_addmulti(&addr, ifp)) == NULL)
1472                         return (ENOBUFS);
1473                 imo->imo_num_memberships++;
1474                 imo->imo_multicast_ifp = ifp;
1475                 imo->imo_multicast_ttl = CARP_DFLTTL;
1476                 imo->imo_multicast_loop = 0;
1477         }
1478
1479         if (!ifp->if_carp) {
1480
1481                 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP,
1482                     M_WAITOK|M_ZERO);
1483                 if (!cif) {
1484                         error = ENOBUFS;
1485                         goto cleanup;
1486                 }
1487                 if ((error = ifpromisc(ifp, 1))) {
1488                         FREE(cif, M_CARP);
1489                         goto cleanup;
1490                 }
1491
1492                 CARP_LOCK_INIT(cif);
1493                 CARP_LOCK(cif);
1494                 cif->vhif_ifp = ifp;
1495                 TAILQ_INIT(&cif->vhif_vrs);
1496                 ifp->if_carp = cif;
1497
1498         } else {
1499                 struct carp_softc *vr;
1500
1501                 cif = (struct carp_if *)ifp->if_carp;
1502                 CARP_LOCK(cif);
1503                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1504                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1505                                 CARP_UNLOCK(cif);
1506                                 error = EINVAL;
1507                                 goto cleanup;
1508                         }
1509         }
1510         sc->sc_ia = ia;
1511         sc->sc_carpdev = ifp;
1512
1513         { /* XXX prevent endless loop if already in queue */
1514         struct carp_softc *vr, *after = NULL;
1515         int myself = 0;
1516         cif = (struct carp_if *)ifp->if_carp;
1517
1518         /* XXX: cif should not change, right? So we still hold the lock */
1519         CARP_LOCK_ASSERT(cif);
1520
1521         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1522                 if (vr == sc)
1523                         myself = 1;
1524                 if (vr->sc_vhid < sc->sc_vhid)
1525                         after = vr;
1526         }
1527
1528         if (!myself) {
1529                 /* We're trying to keep things in order */
1530                 if (after == NULL) {
1531                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1532                 } else {
1533                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1534                 }
1535                 cif->vhif_nvrs++;
1536         }
1537         }
1538
1539         sc->sc_naddrs++;
1540         SC2IFP(sc)->if_flags |= IFF_UP;
1541         if (own)
1542                 sc->sc_advskew = 0;
1543         carp_sc_state_locked(sc);
1544         carp_setrun(sc, 0);
1545
1546         CARP_UNLOCK(cif);
1547
1548         return (0);
1549
1550 cleanup:
1551         in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1552         return (error);
1553 }
1554
1555 static int
1556 carp_del_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1557 {
1558         int error = 0;
1559
1560         if (!--sc->sc_naddrs) {
1561                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1562                 struct ip_moptions *imo = &sc->sc_imo;
1563
1564                 CARP_LOCK(cif);
1565                 callout_stop(&sc->sc_ad_tmo);
1566                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1567                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1568                 sc->sc_vhid = -1;
1569                 in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1570                 imo->imo_multicast_ifp = NULL;
1571                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1572                 if (!--cif->vhif_nvrs) {
1573                         sc->sc_carpdev->if_carp = NULL;
1574                         CARP_LOCK_DESTROY(cif);
1575                         FREE(cif, M_IFADDR);
1576                 } else {
1577                         CARP_UNLOCK(cif);
1578                 }
1579         }
1580
1581         return (error);
1582 }
1583
1584 #ifdef INET6
1585 static int
1586 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1587 {
1588         struct ifnet *ifp;
1589         struct carp_if *cif;
1590         struct in6_ifaddr *ia, *ia_if;
1591         struct ip6_moptions *im6o = &sc->sc_im6o;
1592         struct in6_multi_mship *imm;
1593         struct in6_addr in6;
1594         int own, error;
1595
1596         if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
1597                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1598                         carp_set_state(sc, INIT);
1599                 if (sc->sc_naddrs6)
1600                         SC2IFP(sc)->if_flags |= IFF_UP;
1601                 carp_setrun(sc, 0);
1602                 return (0);
1603         }
1604
1605         /* we have to do it by hands to check we won't match on us */
1606         ia_if = NULL; own = 0;
1607         for (ia = in6_ifaddr; ia; ia = ia->ia_next) {
1608                 int i;
1609
1610                 for (i = 0; i < 4; i++) {
1611                         if ((sin6->sin6_addr.s6_addr32[i] &
1612                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]) !=
1613                             (ia->ia_addr.sin6_addr.s6_addr32[i] &
1614                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]))
1615                                 break;
1616                 }
1617                 /* and, yeah, we need a multicast-capable iface too */
1618                 if (ia->ia_ifp != SC2IFP(sc) &&
1619                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1620                     (i == 4)) {
1621                         if (!ia_if)
1622                                 ia_if = ia;
1623                         if (IN6_ARE_ADDR_EQUAL(&sin6->sin6_addr,
1624                             &ia->ia_addr.sin6_addr))
1625                                 own++;
1626                 }
1627         }
1628
1629         if (!ia_if)
1630                 return (EADDRNOTAVAIL);
1631         ia = ia_if;
1632         ifp = ia->ia_ifp;
1633
1634         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1635             (im6o->im6o_multicast_ifp && im6o->im6o_multicast_ifp != ifp))
1636                 return (EADDRNOTAVAIL);
1637
1638         if (!sc->sc_naddrs6) {
1639                 im6o->im6o_multicast_ifp = ifp;
1640
1641                 /* join CARP multicast address */
1642                 bzero(&in6, sizeof(in6));
1643                 in6.s6_addr16[0] = htons(0xff02);
1644                 in6.s6_addr8[15] = 0x12;
1645                 if (in6_setscope(&in6, ifp, NULL) != 0)
1646                         goto cleanup;
1647                 if ((imm = in6_joingroup(ifp, &in6, &error, 0)) == NULL)
1648                         goto cleanup;
1649                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1650
1651                 /* join solicited multicast address */
1652                 bzero(&in6, sizeof(in6));
1653                 in6.s6_addr16[0] = htons(0xff02);
1654                 in6.s6_addr32[1] = 0;
1655                 in6.s6_addr32[2] = htonl(1);
1656                 in6.s6_addr32[3] = sin6->sin6_addr.s6_addr32[3];
1657                 in6.s6_addr8[12] = 0xff;
1658                 if (in6_setscope(&in6, ifp, NULL) != 0)
1659                         goto cleanup;
1660                 if ((imm = in6_joingroup(ifp, &in6, &error, 0)) == NULL)
1661                         goto cleanup;
1662                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1663         }
1664
1665         if (!ifp->if_carp) {
1666                 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP,
1667                     M_WAITOK|M_ZERO);
1668                 if (!cif) {
1669                         error = ENOBUFS;
1670                         goto cleanup;
1671                 }
1672                 if ((error = ifpromisc(ifp, 1))) {
1673                         FREE(cif, M_CARP);
1674                         goto cleanup;
1675                 }
1676
1677                 CARP_LOCK_INIT(cif);
1678                 CARP_LOCK(cif);
1679                 cif->vhif_ifp = ifp;
1680                 TAILQ_INIT(&cif->vhif_vrs);
1681                 ifp->if_carp = cif;
1682
1683         } else {
1684                 struct carp_softc *vr;
1685
1686                 cif = (struct carp_if *)ifp->if_carp;
1687                 CARP_LOCK(cif);
1688                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1689                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1690                                 CARP_UNLOCK(cif);
1691                                 error = EINVAL;
1692                                 goto cleanup;
1693                         }
1694         }
1695         sc->sc_ia6 = ia;
1696         sc->sc_carpdev = ifp;
1697
1698         { /* XXX prevent endless loop if already in queue */
1699         struct carp_softc *vr, *after = NULL;
1700         int myself = 0;
1701         cif = (struct carp_if *)ifp->if_carp;
1702         CARP_LOCK_ASSERT(cif);
1703
1704         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1705                 if (vr == sc)
1706                         myself = 1;
1707                 if (vr->sc_vhid < sc->sc_vhid)
1708                         after = vr;
1709         }
1710
1711         if (!myself) {
1712                 /* We're trying to keep things in order */
1713                 if (after == NULL) {
1714                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1715                 } else {
1716                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1717                 }
1718                 cif->vhif_nvrs++;
1719         }
1720         }
1721
1722         sc->sc_naddrs6++;
1723         SC2IFP(sc)->if_flags |= IFF_UP;
1724         if (own)
1725                 sc->sc_advskew = 0;
1726         carp_sc_state_locked(sc);
1727         carp_setrun(sc, 0);
1728
1729         CARP_UNLOCK(cif);
1730
1731         return (0);
1732
1733 cleanup:
1734         /* clean up multicast memberships */
1735         if (!sc->sc_naddrs6) {
1736                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1737                         imm = LIST_FIRST(&im6o->im6o_memberships);
1738                         LIST_REMOVE(imm, i6mm_chain);
1739                         in6_leavegroup(imm);
1740                 }
1741         }
1742         return (error);
1743 }
1744
1745 static int
1746 carp_del_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1747 {
1748         int error = 0;
1749
1750         if (!--sc->sc_naddrs6) {
1751                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1752                 struct ip6_moptions *im6o = &sc->sc_im6o;
1753
1754                 CARP_LOCK(cif);
1755                 callout_stop(&sc->sc_ad_tmo);
1756                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1757                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1758                 sc->sc_vhid = -1;
1759                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1760                         struct in6_multi_mship *imm =
1761                             LIST_FIRST(&im6o->im6o_memberships);
1762
1763                         LIST_REMOVE(imm, i6mm_chain);
1764                         in6_leavegroup(imm);
1765                 }
1766                 im6o->im6o_multicast_ifp = NULL;
1767                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1768                 if (!--cif->vhif_nvrs) {
1769                         CARP_LOCK_DESTROY(cif);
1770                         sc->sc_carpdev->if_carp = NULL;
1771                         FREE(cif, M_IFADDR);
1772                 } else
1773                         CARP_UNLOCK(cif);
1774         }
1775
1776         return (error);
1777 }
1778 #endif /* INET6 */
1779
1780 static int
1781 carp_ioctl(struct ifnet *ifp, u_long cmd, caddr_t addr)
1782 {
1783         struct carp_softc *sc = ifp->if_softc, *vr;
1784         struct carpreq carpr;
1785         struct ifaddr *ifa;
1786         struct ifreq *ifr;
1787         struct ifaliasreq *ifra;
1788         int locked = 0, error = 0;
1789
1790         ifa = (struct ifaddr *)addr;
1791         ifra = (struct ifaliasreq *)addr;
1792         ifr = (struct ifreq *)addr;
1793
1794         switch (cmd) {
1795         case SIOCSIFADDR:
1796                 switch (ifa->ifa_addr->sa_family) {
1797 #ifdef INET
1798                 case AF_INET:
1799                         SC2IFP(sc)->if_flags |= IFF_UP;
1800                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1801                             sizeof(struct sockaddr));
1802                         error = carp_set_addr(sc, satosin(ifa->ifa_addr));
1803                         break;
1804 #endif /* INET */
1805 #ifdef INET6
1806                 case AF_INET6:
1807                         SC2IFP(sc)->if_flags |= IFF_UP;
1808                         error = carp_set_addr6(sc, satosin6(ifa->ifa_addr));
1809                         break;
1810 #endif /* INET6 */
1811                 default:
1812                         error = EAFNOSUPPORT;
1813                         break;
1814                 }
1815                 break;
1816
1817         case SIOCAIFADDR:
1818                 switch (ifa->ifa_addr->sa_family) {
1819 #ifdef INET
1820                 case AF_INET:
1821                         SC2IFP(sc)->if_flags |= IFF_UP;
1822                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1823                             sizeof(struct sockaddr));
1824                         error = carp_set_addr(sc, satosin(&ifra->ifra_addr));
1825                         break;
1826 #endif /* INET */
1827 #ifdef INET6
1828                 case AF_INET6:
1829                         SC2IFP(sc)->if_flags |= IFF_UP;
1830                         error = carp_set_addr6(sc, satosin6(&ifra->ifra_addr));
1831                         break;
1832 #endif /* INET6 */
1833                 default:
1834                         error = EAFNOSUPPORT;
1835                         break;
1836                 }
1837                 break;
1838
1839         case SIOCDIFADDR:
1840                 switch (ifa->ifa_addr->sa_family) {
1841 #ifdef INET
1842                 case AF_INET:
1843                         error = carp_del_addr(sc, satosin(&ifra->ifra_addr));
1844                         break;
1845 #endif /* INET */
1846 #ifdef INET6
1847                 case AF_INET6:
1848                         error = carp_del_addr6(sc, satosin6(&ifra->ifra_addr));
1849                         break;
1850 #endif /* INET6 */
1851                 default:
1852                         error = EAFNOSUPPORT;
1853                         break;
1854                 }
1855                 break;
1856
1857         case SIOCSIFFLAGS:
1858                 if (sc->sc_carpdev) {
1859                         locked = 1;
1860                         CARP_SCLOCK(sc);
1861                 }
1862                 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) {
1863                         callout_stop(&sc->sc_ad_tmo);
1864                         callout_stop(&sc->sc_md_tmo);
1865                         callout_stop(&sc->sc_md6_tmo);
1866                         if (sc->sc_state == MASTER)
1867                                 carp_send_ad_locked(sc);
1868                         carp_set_state(sc, INIT);
1869                         carp_setrun(sc, 0);
1870                 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) {
1871                         SC2IFP(sc)->if_flags |= IFF_UP;
1872                         carp_setrun(sc, 0);
1873                 }
1874                 break;
1875
1876         case SIOCSVH:
1877                 error = priv_check(curthread, PRIV_NETINET_CARP);
1878                 if (error)
1879                         break;
1880                 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1881                         break;
1882                 error = 1;
1883                 if (sc->sc_carpdev) {
1884                         locked = 1;
1885                         CARP_SCLOCK(sc);
1886                 }
1887                 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) {
1888                         switch (carpr.carpr_state) {
1889                         case BACKUP:
1890                                 callout_stop(&sc->sc_ad_tmo);
1891                                 carp_set_state(sc, BACKUP);
1892                                 carp_setrun(sc, 0);
1893                                 carp_setroute(sc, RTM_DELETE);
1894                                 break;
1895                         case MASTER:
1896                                 carp_master_down_locked(sc);
1897                                 break;
1898                         default:
1899                                 break;
1900                         }
1901                 }
1902                 if (carpr.carpr_vhid > 0) {
1903                         if (carpr.carpr_vhid > 255) {
1904                                 error = EINVAL;
1905                                 break;
1906                         }
1907                         if (sc->sc_carpdev) {
1908                                 struct carp_if *cif;
1909                                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1910                                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1911                                         if (vr != sc &&
1912                                             vr->sc_vhid == carpr.carpr_vhid) {
1913                                                 error = EEXIST;
1914                                                 break;
1915                                         }
1916                                 if (error == EEXIST)
1917                                         break;
1918                         }
1919                         sc->sc_vhid = carpr.carpr_vhid;
1920                         IF_LLADDR(sc->sc_ifp)[0] = 0;
1921                         IF_LLADDR(sc->sc_ifp)[1] = 0;
1922                         IF_LLADDR(sc->sc_ifp)[2] = 0x5e;
1923                         IF_LLADDR(sc->sc_ifp)[3] = 0;
1924                         IF_LLADDR(sc->sc_ifp)[4] = 1;
1925                         IF_LLADDR(sc->sc_ifp)[5] = sc->sc_vhid;
1926                         error--;
1927                 }
1928                 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) {
1929                         if (carpr.carpr_advskew >= 255) {
1930                                 error = EINVAL;
1931                                 break;
1932                         }
1933                         if (carpr.carpr_advbase > 255) {
1934                                 error = EINVAL;
1935                                 break;
1936                         }
1937                         sc->sc_advbase = carpr.carpr_advbase;
1938                         sc->sc_advskew = carpr.carpr_advskew;
1939                         error--;
1940                 }
1941                 bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1942                 if (error > 0)
1943                         error = EINVAL;
1944                 else {
1945                         error = 0;
1946                         carp_setrun(sc, 0);
1947                 }
1948                 break;
1949
1950         case SIOCGVH:
1951                 /* XXX: lockless read */
1952                 bzero(&carpr, sizeof(carpr));
1953                 carpr.carpr_state = sc->sc_state;
1954                 carpr.carpr_vhid = sc->sc_vhid;
1955                 carpr.carpr_advbase = sc->sc_advbase;
1956                 carpr.carpr_advskew = sc->sc_advskew;
1957                 error = priv_check(curthread, PRIV_NETINET_CARP);
1958                 if (error == 0)
1959                         bcopy(sc->sc_key, carpr.carpr_key,
1960                             sizeof(carpr.carpr_key));
1961                 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1962                 break;
1963
1964         default:
1965                 error = EINVAL;
1966         }
1967
1968         if (locked)
1969                 CARP_SCUNLOCK(sc);
1970
1971         carp_hmac_prepare(sc);
1972
1973         return (error);
1974 }
1975
1976 /*
1977  * XXX: this is looutput. We should eventually use it from there.
1978  */
1979 static int
1980 carp_looutput(struct ifnet *ifp, struct mbuf *m, struct sockaddr *dst,
1981     struct rtentry *rt)
1982 {
1983         u_int32_t af;
1984
1985         M_ASSERTPKTHDR(m); /* check if we have the packet header */
1986
1987         if (rt && rt->rt_flags & (RTF_REJECT|RTF_BLACKHOLE)) {
1988                 m_freem(m);
1989                 return (rt->rt_flags & RTF_BLACKHOLE ? 0 :
1990                         rt->rt_flags & RTF_HOST ? EHOSTUNREACH : ENETUNREACH);
1991         }
1992
1993         ifp->if_opackets++;
1994         ifp->if_obytes += m->m_pkthdr.len;
1995
1996         /* BPF writes need to be handled specially. */
1997         if (dst->sa_family == AF_UNSPEC) {
1998                 bcopy(dst->sa_data, &af, sizeof(af));
1999                 dst->sa_family = af;
2000         }
2001
2002 #if 1   /* XXX */
2003         switch (dst->sa_family) {
2004         case AF_INET:
2005         case AF_INET6:
2006         case AF_IPX:
2007         case AF_APPLETALK:
2008                 break;
2009         default:
2010                 printf("carp_looutput: af=%d unexpected\n", dst->sa_family);
2011                 m_freem(m);
2012                 return (EAFNOSUPPORT);
2013         }
2014 #endif
2015         return(if_simloop(ifp, m, dst->sa_family, 0));
2016 }
2017
2018 /*
2019  * Start output on carp interface. This function should never be called.
2020  */
2021 static void
2022 carp_start(struct ifnet *ifp)
2023 {
2024 #ifdef DEBUG
2025         printf("%s: start called\n", ifp->if_xname);
2026 #endif
2027 }
2028
2029 int
2030 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *sa,
2031     struct rtentry *rt)
2032 {
2033         struct m_tag *mtag;
2034         struct carp_softc *sc;
2035         struct ifnet *carp_ifp;
2036
2037         if (!sa)
2038                 return (0);
2039
2040         switch (sa->sa_family) {
2041 #ifdef INET
2042         case AF_INET:
2043                 break;
2044 #endif /* INET */
2045 #ifdef INET6
2046         case AF_INET6:
2047                 break;
2048 #endif /* INET6 */
2049         default:
2050                 return (0);
2051         }
2052
2053         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
2054         if (mtag == NULL)
2055                 return (0);
2056
2057         bcopy(mtag + 1, &carp_ifp, sizeof(struct ifnet *));
2058         sc = carp_ifp->if_softc;
2059
2060         /* Set the source MAC address to Virtual Router MAC Address */
2061         switch (ifp->if_type) {
2062         case IFT_ETHER:
2063         case IFT_L2VLAN: {
2064                         struct ether_header *eh;
2065
2066                         eh = mtod(m, struct ether_header *);
2067                         eh->ether_shost[0] = 0;
2068                         eh->ether_shost[1] = 0;
2069                         eh->ether_shost[2] = 0x5e;
2070                         eh->ether_shost[3] = 0;
2071                         eh->ether_shost[4] = 1;
2072                         eh->ether_shost[5] = sc->sc_vhid;
2073                 }
2074                 break;
2075         case IFT_FDDI: {
2076                         struct fddi_header *fh;
2077
2078                         fh = mtod(m, struct fddi_header *);
2079                         fh->fddi_shost[0] = 0;
2080                         fh->fddi_shost[1] = 0;
2081                         fh->fddi_shost[2] = 0x5e;
2082                         fh->fddi_shost[3] = 0;
2083                         fh->fddi_shost[4] = 1;
2084                         fh->fddi_shost[5] = sc->sc_vhid;
2085                 }
2086                 break;
2087         case IFT_ISO88025: {
2088                         struct iso88025_header *th;
2089                         th = mtod(m, struct iso88025_header *);
2090                         th->iso88025_shost[0] = 3;
2091                         th->iso88025_shost[1] = 0;
2092                         th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
2093                         th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
2094                         th->iso88025_shost[4] = 0;
2095                         th->iso88025_shost[5] = 0;
2096                 }
2097                 break;
2098         default:
2099                 printf("%s: carp is not supported for this interface type\n",
2100                     ifp->if_xname);
2101                 return (EOPNOTSUPP);
2102         }
2103
2104         return (0);
2105 }
2106
2107 static void
2108 carp_set_state(struct carp_softc *sc, int state)
2109 {
2110
2111         if (sc->sc_carpdev)
2112                 CARP_SCLOCK_ASSERT(sc);
2113
2114         if (sc->sc_state == state)
2115                 return;
2116
2117         sc->sc_state = state;
2118         switch (state) {
2119         case BACKUP:
2120                 SC2IFP(sc)->if_link_state = LINK_STATE_DOWN;
2121                 break;
2122         case MASTER:
2123                 SC2IFP(sc)->if_link_state = LINK_STATE_UP;
2124                 break;
2125         default:
2126                 SC2IFP(sc)->if_link_state = LINK_STATE_UNKNOWN;
2127                 break;
2128         }
2129         rt_ifmsg(SC2IFP(sc));
2130 }
2131
2132 void
2133 carp_carpdev_state(void *v)
2134 {
2135         struct carp_if *cif = v;
2136
2137         CARP_LOCK(cif);
2138         carp_carpdev_state_locked(cif);
2139         CARP_UNLOCK(cif);
2140 }
2141
2142 static void
2143 carp_carpdev_state_locked(struct carp_if *cif)
2144 {
2145         struct carp_softc *sc;
2146
2147         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list)
2148                 carp_sc_state_locked(sc);
2149 }
2150
2151 static void
2152 carp_sc_state_locked(struct carp_softc *sc)
2153 {
2154         CARP_SCLOCK_ASSERT(sc);
2155
2156         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
2157             !(sc->sc_carpdev->if_flags & IFF_UP)) {
2158                 sc->sc_flags_backup = SC2IFP(sc)->if_flags;
2159                 SC2IFP(sc)->if_flags &= ~IFF_UP;
2160                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
2161                 callout_stop(&sc->sc_ad_tmo);
2162                 callout_stop(&sc->sc_md_tmo);
2163                 callout_stop(&sc->sc_md6_tmo);
2164                 carp_set_state(sc, INIT);
2165                 carp_setrun(sc, 0);
2166                 if (!sc->sc_suppress) {
2167                         carp_suppress_preempt++;
2168                         if (carp_suppress_preempt == 1) {
2169                                 CARP_SCUNLOCK(sc);
2170                                 carp_send_ad_all();
2171                                 CARP_SCLOCK(sc);
2172                         }
2173                 }
2174                 sc->sc_suppress = 1;
2175         } else {
2176                 SC2IFP(sc)->if_flags |= sc->sc_flags_backup;
2177                 carp_set_state(sc, INIT);
2178                 carp_setrun(sc, 0);
2179                 if (sc->sc_suppress)
2180                         carp_suppress_preempt--;
2181                 sc->sc_suppress = 0;
2182         }
2183
2184         return;
2185 }
2186
2187 static int
2188 carp_modevent(module_t mod, int type, void *data)
2189 {
2190         switch (type) {
2191         case MOD_LOAD:
2192                 if_detach_event_tag = EVENTHANDLER_REGISTER(ifnet_departure_event,
2193                     carp_ifdetach, NULL, EVENTHANDLER_PRI_ANY);
2194                 if (if_detach_event_tag == NULL)
2195                         return (ENOMEM);
2196                 mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2197                 LIST_INIT(&carpif_list);
2198                 if_clone_attach(&carp_cloner);
2199                 break;
2200
2201         case MOD_UNLOAD:
2202                 EVENTHANDLER_DEREGISTER(ifnet_departure_event, if_detach_event_tag);
2203                 if_clone_detach(&carp_cloner);
2204                 mtx_destroy(&carp_mtx);
2205                 break;
2206
2207         default:
2208                 return (EINVAL);
2209         }
2210
2211         return (0);
2212 }
2213
2214 static moduledata_t carp_mod = {
2215         "carp",
2216         carp_modevent,
2217         0
2218 };
2219
2220 DECLARE_MODULE(carp, carp_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);