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/proc.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/syslog.h>
  47 #include <sys/signalvar.h>
  48 #include <sys/filio.h>
  49 #include <sys/sockio.h>
  50
  51 #include <sys/socket.h>
  52 #include <sys/vnode.h>
  53
  54 #include <machine/stdarg.h>
  55
  56 #include <net/bpf.h>
  57 #include <net/ethernet.h>
  58 #include <net/fddi.h>
  59 #include <net/iso88025.h>
  60 #include <net/if.h>
  61 #include <net/if_clone.h>
  62 #include <net/if_dl.h>
  63 #include <net/if_types.h>
  64 #include <net/route.h>
  65
  66 #ifdef INET
  67 #include <netinet/in.h>
  68 #include <netinet/in_var.h>
  69 #include <netinet/in_systm.h>
  70 #include <netinet/ip.h>
  71 #include <netinet/ip_var.h>
  72 #include <netinet/if_ether.h>
  73 #include <machine/in_cksum.h>
  74 #endif
  75
  76 #ifdef INET6
  77 #include <netinet/icmp6.h>
  78 #include <netinet/ip6.h>
  79 #include <netinet6/ip6_var.h>
  80 #include <netinet6/scope6_var.h>
  81 #include <netinet6/nd6.h>
  82 #endif
  83
  84 #include <crypto/sha1.h>
  85 #include <netinet/ip_carp.h>
  86
  87 #define CARP_IFNAME     "carp"
  88 static MALLOC_DEFINE(M_CARP, "CARP", "CARP interfaces");
  89 SYSCTL_DECL(_net_inet_carp);
  90
  91 struct carp_softc {
  92         struct ifnet            *sc_ifp;        /* Interface clue */
  93         struct ifnet            *sc_carpdev;    /* Pointer to parent interface */
  94         struct in_ifaddr        *sc_ia;         /* primary iface address */
  95         struct ip_moptions       sc_imo;
  96 #ifdef INET6
  97         struct in6_ifaddr       *sc_ia6;        /* primary iface address v6 */
  98         struct ip6_moptions      sc_im6o;
  99 #endif /* INET6 */
 100         TAILQ_ENTRY(carp_softc)  sc_list;
 101
 102         enum { INIT = 0, BACKUP, MASTER }       sc_state;
 103
 104         int                      sc_flags_backup;
 105         int                      sc_suppress;
 106
 107         int                      sc_sendad_errors;
 108 #define CARP_SENDAD_MAX_ERRORS  3
 109         int                      sc_sendad_success;
 110 #define CARP_SENDAD_MIN_SUCCESS 3
 111
 112         int                      sc_vhid;
 113         int                      sc_advskew;
 114         int                      sc_naddrs;
 115         int                      sc_naddrs6;
 116         int                      sc_advbase;    /* seconds */
 117         int                      sc_init_counter;
 118         u_int64_t                sc_counter;
 119
 120         /* authentication */
 121 #define CARP_HMAC_PAD   64
 122         unsigned char sc_key[CARP_KEY_LEN];
 123         unsigned char sc_pad[CARP_HMAC_PAD];
 124         SHA1_CTX sc_sha1;
 125
 126         struct callout           sc_ad_tmo;     /* advertisement timeout */
 127         struct callout           sc_md_tmo;     /* master down timeout */
 128         struct callout           sc_md6_tmo;    /* master down timeout */
 129
 130         LIST_ENTRY(carp_softc)   sc_next;       /* Interface clue */
 131 };
 132 #define SC2IFP(sc)      ((sc)->sc_ifp)
 133
 134 int carp_suppress_preempt = 0;
 135 int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 1, 0, 0 };    /* XXX for now */
 136 SYSCTL_INT(_net_inet_carp, CARPCTL_ALLOW, allow, CTLFLAG_RW,
 137     &carp_opts[CARPCTL_ALLOW], 0, "Accept incoming CARP packets");
 138 SYSCTL_INT(_net_inet_carp, CARPCTL_PREEMPT, preempt, CTLFLAG_RW,
 139     &carp_opts[CARPCTL_PREEMPT], 0, "high-priority backup preemption mode");
 140 SYSCTL_INT(_net_inet_carp, CARPCTL_LOG, log, CTLFLAG_RW,
 141     &carp_opts[CARPCTL_LOG], 0, "log bad carp packets");
 142 SYSCTL_INT(_net_inet_carp, CARPCTL_ARPBALANCE, arpbalance, CTLFLAG_RW,
 143     &carp_opts[CARPCTL_ARPBALANCE], 0, "balance arp responses");
 144 SYSCTL_INT(_net_inet_carp, OID_AUTO, suppress_preempt, CTLFLAG_RD,
 145     &carp_suppress_preempt, 0, "Preemption is suppressed");
 146
 147 struct carpstats carpstats;
 148 SYSCTL_STRUCT(_net_inet_carp, CARPCTL_STATS, stats, CTLFLAG_RW,
 149     &carpstats, carpstats,
 150     "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 151
 152 struct carp_if {
 153         TAILQ_HEAD(, carp_softc) vhif_vrs;
 154         int vhif_nvrs;
 155
 156         struct ifnet    *vhif_ifp;
 157         struct mtx       vhif_mtx;
 158 };
 159
 160 /* Get carp_if from softc. Valid after carp_set_addr{,6}. */
 161 #define SC2CIF(sc)              ((struct carp_if *)(sc)->sc_carpdev->if_carp)
 162
 163 /* lock per carp_if queue */
 164 #define CARP_LOCK_INIT(cif)     mtx_init(&(cif)->vhif_mtx, "carp_if",   \
 165         NULL, MTX_DEF)
 166 #define CARP_LOCK_DESTROY(cif)  mtx_destroy(&(cif)->vhif_mtx)
 167 #define CARP_LOCK_ASSERT(cif)   mtx_assert(&(cif)->vhif_mtx, MA_OWNED)
 168 #define CARP_LOCK(cif)          mtx_lock(&(cif)->vhif_mtx)
 169 #define CARP_UNLOCK(cif)        mtx_unlock(&(cif)->vhif_mtx)
 170
 171 #define CARP_SCLOCK(sc)         mtx_lock(&SC2CIF(sc)->vhif_mtx)
 172 #define CARP_SCUNLOCK(sc)       mtx_unlock(&SC2CIF(sc)->vhif_mtx)
 173 #define CARP_SCLOCK_ASSERT(sc)  mtx_assert(&SC2CIF(sc)->vhif_mtx, MA_OWNED)
 174
 175 #define CARP_LOG(...)   do {                            \
 176         if (carp_opts[CARPCTL_LOG] > 0)                 \
 177                 log(LOG_INFO, __VA_ARGS__);             \
 178 } while (0)
 179
 180 #define CARP_DEBUG(...) do {                            \
 181         if (carp_opts[CARPCTL_LOG] > 1)                 \
 182                 log(LOG_DEBUG, __VA_ARGS__);            \
 183 } while (0)
 184
 185 static void     carp_hmac_prepare(struct carp_softc *);
 186 static void     carp_hmac_generate(struct carp_softc *, u_int32_t *,
 187                     unsigned char *);
 188 static int      carp_hmac_verify(struct carp_softc *, u_int32_t *,
 189                     unsigned char *);
 190 static void     carp_setroute(struct carp_softc *, int);
 191 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 192 static int      carp_clone_create(struct if_clone *, int);
 193 static void     carp_clone_destroy(struct ifnet *);
 194 static void     carpdetach(struct carp_softc *);
 195 static int      carp_prepare_ad(struct mbuf *, struct carp_softc *,
 196                     struct carp_header *);
 197 static void     carp_send_ad_all(void);
 198 static void     carp_send_ad(void *);
 199 static void     carp_send_ad_locked(struct carp_softc *);
 200 static void     carp_send_arp(struct carp_softc *);
 201 static void     carp_master_down(void *);
 202 static void     carp_master_down_locked(struct carp_softc *);
 203 static int      carp_ioctl(struct ifnet *, u_long, caddr_t);
 204 static int      carp_looutput(struct ifnet *, struct mbuf *, struct sockaddr *,
 205                     struct rtentry *);
 206 static void     carp_start(struct ifnet *);
 207 static void     carp_setrun(struct carp_softc *, sa_family_t);
 208 static void     carp_set_state(struct carp_softc *, int);
 209 static int      carp_addrcount(struct carp_if *, struct in_ifaddr *, int);
 210 enum    { CARP_COUNT_MASTER, CARP_COUNT_RUNNING };
 211
 212 static void     carp_multicast_cleanup(struct carp_softc *);
 213 static int      carp_set_addr(struct carp_softc *, struct sockaddr_in *);
 214 static int      carp_del_addr(struct carp_softc *, struct sockaddr_in *);
 215 static void     carp_carpdev_state_locked(struct carp_if *);
 216 static void     carp_sc_state_locked(struct carp_softc *);
 217 #ifdef INET6
 218 static void     carp_send_na(struct carp_softc *);
 219 static int      carp_set_addr6(struct carp_softc *, struct sockaddr_in6 *);
 220 static int      carp_del_addr6(struct carp_softc *, struct sockaddr_in6 *);
 221 #endif
 222
 223 static LIST_HEAD(, carp_softc) carpif_list;
 224 static struct mtx carp_mtx;
 225 IFC_SIMPLE_DECLARE(carp, 0);
 226
 227 static eventhandler_tag if_detach_event_tag;
 228
 229 static __inline u_int16_t
 230 carp_cksum(struct mbuf *m, int len)
 231 {
 232         return (in_cksum(m, len));
 233 }
 234
 235 static void
 236 carp_hmac_prepare(struct carp_softc *sc)
 237 {
 238         u_int8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 239         u_int8_t vhid = sc->sc_vhid & 0xff;
 240         struct ifaddr *ifa;
 241         int i;
 242 #ifdef INET6
 243         struct in6_addr in6;
 244 #endif
 245
 246         if (sc->sc_carpdev)
 247                 CARP_SCLOCK(sc);
 248
 249         /* XXX: possible race here */
 250
 251         /* compute ipad from key */
 252         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 253         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 254         for (i = 0; i < sizeof(sc->sc_pad); i++)
 255                 sc->sc_pad[i] ^= 0x36;
 256
 257         /* precompute first part of inner hash */
 258         SHA1Init(&sc->sc_sha1);
 259         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 260         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 261         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 262         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 263 #ifdef INET
 264         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 265                 if (ifa->ifa_addr->sa_family == AF_INET)
 266                         SHA1Update(&sc->sc_sha1,
 267                             (void *)&ifatoia(ifa)->ia_addr.sin_addr.s_addr,
 268                             sizeof(struct in_addr));
 269         }
 270 #endif /* INET */
 271 #ifdef INET6
 272         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 273                 if (ifa->ifa_addr->sa_family == AF_INET6) {
 274                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 275                         in6_clearscope(&in6);
 276                         SHA1Update(&sc->sc_sha1, (void *)&in6, sizeof(in6));
 277                 }
 278         }
 279 #endif /* INET6 */
 280
 281         /* convert ipad to opad */
 282         for (i = 0; i < sizeof(sc->sc_pad); i++)
 283                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 284
 285         if (sc->sc_carpdev)
 286                 CARP_SCUNLOCK(sc);
 287 }
 288
 289 static void
 290 carp_hmac_generate(struct carp_softc *sc, u_int32_t counter[2],
 291     unsigned char md[20])
 292 {
 293         SHA1_CTX sha1ctx;
 294
 295         /* fetch first half of inner hash */
 296         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 297
 298         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 299         SHA1Final(md, &sha1ctx);
 300
 301         /* outer hash */
 302         SHA1Init(&sha1ctx);
 303         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 304         SHA1Update(&sha1ctx, md, 20);
 305         SHA1Final(md, &sha1ctx);
 306 }
 307
 308 static int
 309 carp_hmac_verify(struct carp_softc *sc, u_int32_t counter[2],
 310     unsigned char md[20])
 311 {
 312         unsigned char md2[20];
 313
 314         CARP_SCLOCK_ASSERT(sc);
 315
 316         carp_hmac_generate(sc, counter, md2);
 317
 318         return (bcmp(md, md2, sizeof(md2)));
 319 }
 320
 321 static void
 322 carp_setroute(struct carp_softc *sc, int cmd)
 323 {
 324         struct ifaddr *ifa;
 325         int s;
 326
 327         if (sc->sc_carpdev)
 328                 CARP_SCLOCK_ASSERT(sc);
 329
 330         s = splnet();
 331         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 332                 if (ifa->ifa_addr->sa_family == AF_INET &&
 333                     sc->sc_carpdev != NULL) {
 334                         int count = carp_addrcount(
 335                             (struct carp_if *)sc->sc_carpdev->if_carp,
 336                             ifatoia(ifa), CARP_COUNT_MASTER);
 337
 338                         if ((cmd == RTM_ADD && count == 1) ||
 339                             (cmd == RTM_DELETE && count == 0))
 340                                 rtinit(ifa, cmd, RTF_UP | RTF_HOST);
 341                 }
 342 #ifdef INET6
 343                 if (ifa->ifa_addr->sa_family == AF_INET6) {
 344                         if (cmd == RTM_ADD)
 345                                 in6_ifaddloop(ifa);
 346                         else
 347                                 in6_ifremloop(ifa);
 348                 }
 349 #endif /* INET6 */
 350         }
 351         splx(s);
 352 }
 353
 354 static int
 355 carp_clone_create(struct if_clone *ifc, int unit)
 356 {
 357
 358         struct carp_softc *sc;
 359         struct ifnet *ifp;
 360
 361         MALLOC(sc, struct carp_softc *, sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
 362         ifp = SC2IFP(sc) = if_alloc(IFT_ETHER);
 363         if (ifp == NULL) {
 364                 FREE(sc, M_CARP);
 365                 return (ENOSPC);
 366         }
 367
 368         sc->sc_flags_backup = 0;
 369         sc->sc_suppress = 0;
 370         sc->sc_advbase = CARP_DFLTINTV;
 371         sc->sc_vhid = -1;       /* required setting */
 372         sc->sc_advskew = 0;
 373         sc->sc_init_counter = 1;
 374         sc->sc_naddrs = sc->sc_naddrs6 = 0; /* M_ZERO? */
 375 #ifdef INET6
 376         sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL;
 377 #endif
 378
 379         callout_init(&sc->sc_ad_tmo, NET_CALLOUT_MPSAFE);
 380         callout_init(&sc->sc_md_tmo, NET_CALLOUT_MPSAFE);
 381         callout_init(&sc->sc_md6_tmo, NET_CALLOUT_MPSAFE);
 382
 383         ifp->if_softc = sc;
 384         if_initname(ifp, CARP_IFNAME, unit);
 385         ifp->if_mtu = ETHERMTU;
 386         ifp->if_flags = IFF_LOOPBACK;
 387         ifp->if_ioctl = carp_ioctl;
 388         ifp->if_output = carp_looutput;
 389         ifp->if_start = carp_start;
 390         ifp->if_type = IFT_CARP;
 391         ifp->if_snd.ifq_maxlen = ifqmaxlen;
 392         ifp->if_hdrlen = 0;
 393         if_attach(ifp);
 394         bpfattach(SC2IFP(sc), DLT_NULL, sizeof(u_int32_t));
 395         mtx_lock(&carp_mtx);
 396         LIST_INSERT_HEAD(&carpif_list, sc, sc_next);
 397         mtx_unlock(&carp_mtx);
 398         return (0);
 399 }
 400
 401 static void
 402 carp_clone_destroy(struct ifnet *ifp)
 403 {
 404         struct carp_softc *sc = ifp->if_softc;
 405
 406         if (sc->sc_carpdev)
 407                 CARP_SCLOCK(sc);
 408         carpdetach(sc);
 409         if (sc->sc_carpdev)
 410                 CARP_SCUNLOCK(sc);
 411
 412         mtx_lock(&carp_mtx);
 413         LIST_REMOVE(sc, sc_next);
 414         mtx_unlock(&carp_mtx);
 415         bpfdetach(ifp);
 416         if_detach(ifp);
 417         if_free_type(ifp, IFT_ETHER);
 418         free(sc, M_CARP);
 419 }
 420
 421 static void
 422 carpdetach(struct carp_softc *sc)
 423 {
 424         struct carp_if *cif;
 425
 426         callout_stop(&sc->sc_ad_tmo);
 427         callout_stop(&sc->sc_md_tmo);
 428         callout_stop(&sc->sc_md6_tmo);
 429
 430         if (sc->sc_suppress)
 431                 carp_suppress_preempt--;
 432         sc->sc_suppress = 0;
 433
 434         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS)
 435                 carp_suppress_preempt--;
 436         sc->sc_sendad_errors = 0;
 437
 438         carp_set_state(sc, INIT);
 439         SC2IFP(sc)->if_flags &= ~IFF_UP;
 440         carp_setrun(sc, 0);
 441         carp_multicast_cleanup(sc);
 442
 443         if (sc->sc_carpdev != NULL) {
 444                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
 445                 CARP_LOCK_ASSERT(cif);
 446                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
 447                 if (!--cif->vhif_nvrs) {
 448                         ifpromisc(sc->sc_carpdev, 0);
 449                         sc->sc_carpdev->if_carp = NULL;
 450                         CARP_LOCK_DESTROY(cif);
 451                         FREE(cif, M_IFADDR);
 452                 }
 453         }
 454         sc->sc_carpdev = NULL;
 455 }
 456
 457 /* Detach an interface from the carp. */
 458 static void
 459 carp_ifdetach(void *arg __unused, struct ifnet *ifp)
 460 {
 461         struct carp_if *cif = (struct carp_if *)ifp->if_carp;
 462         struct carp_softc *sc, *nextsc;
 463
 464         if (cif == NULL)
 465                 return;
 466
 467         /*
 468          * XXX: At the end of for() cycle the lock will be destroyed.
 469          */
 470         CARP_LOCK(cif);
 471         for (sc = TAILQ_FIRST(&cif->vhif_vrs); sc; sc = nextsc) {
 472                 nextsc = TAILQ_NEXT(sc, sc_list);
 473                 carpdetach(sc);
 474         }
 475 }
 476
 477 /*
 478  * process input packet.
 479  * we have rearranged checks order compared to the rfc,
 480  * but it seems more efficient this way or not possible otherwise.
 481  */
 482 void
 483 carp_input(struct mbuf *m, int hlen)
 484 {
 485         struct ip *ip = mtod(m, struct ip *);
 486         struct carp_header *ch;
 487         int iplen, len;
 488
 489         carpstats.carps_ipackets++;
 490
 491         if (!carp_opts[CARPCTL_ALLOW]) {
 492                 m_freem(m);
 493                 return;
 494         }
 495
 496         /* check if received on a valid carp interface */
 497         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 498                 carpstats.carps_badif++;
 499                 CARP_LOG("carp_input: packet received on non-carp "
 500                     "interface: %s\n",
 501                     m->m_pkthdr.rcvif->if_xname);
 502                 m_freem(m);
 503                 return;
 504         }
 505
 506         /* verify that the IP TTL is 255.  */
 507         if (ip->ip_ttl != CARP_DFLTTL) {
 508                 carpstats.carps_badttl++;
 509                 CARP_LOG("carp_input: received ttl %d != 255i on %s\n",
 510                     ip->ip_ttl,
 511                     m->m_pkthdr.rcvif->if_xname);
 512                 m_freem(m);
 513                 return;
 514         }
 515
 516         iplen = ip->ip_hl << 2;
 517
 518         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 519                 carpstats.carps_badlen++;
 520                 CARP_LOG("carp_input: received len %zd < "
 521                     "sizeof(struct carp_header)\n",
 522                     m->m_len - sizeof(struct ip));
 523                 m_freem(m);
 524                 return;
 525         }
 526
 527         if (iplen + sizeof(*ch) < m->m_len) {
 528                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 529                         carpstats.carps_hdrops++;
 530                         CARP_LOG("carp_input: pullup failed\n");
 531                         return;
 532                 }
 533                 ip = mtod(m, struct ip *);
 534         }
 535         ch = (struct carp_header *)((char *)ip + iplen);
 536
 537         /*
 538          * verify that the received packet length is
 539          * equal to the CARP header
 540          */
 541         len = iplen + sizeof(*ch);
 542         if (len > m->m_pkthdr.len) {
 543                 carpstats.carps_badlen++;
 544                 CARP_LOG("carp_input: packet too short %d on %s\n",
 545                     m->m_pkthdr.len,
 546                     m->m_pkthdr.rcvif->if_xname);
 547                 m_freem(m);
 548                 return;
 549         }
 550
 551         if ((m = m_pullup(m, len)) == NULL) {
 552                 carpstats.carps_hdrops++;
 553                 return;
 554         }
 555         ip = mtod(m, struct ip *);
 556         ch = (struct carp_header *)((char *)ip + iplen);
 557
 558         /* verify the CARP checksum */
 559         m->m_data += iplen;
 560         if (carp_cksum(m, len - iplen)) {
 561                 carpstats.carps_badsum++;
 562                 CARP_LOG("carp_input: checksum failed on %s\n",
 563                     m->m_pkthdr.rcvif->if_xname);
 564                 m_freem(m);
 565                 return;
 566         }
 567         m->m_data -= iplen;
 568
 569         carp_input_c(m, ch, AF_INET);
 570 }
 571
 572 #ifdef INET6
 573 int
 574 carp6_input(struct mbuf **mp, int *offp, int proto)
 575 {
 576         struct mbuf *m = *mp;
 577         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 578         struct carp_header *ch;
 579         u_int len;
 580
 581         carpstats.carps_ipackets6++;
 582
 583         if (!carp_opts[CARPCTL_ALLOW]) {
 584                 m_freem(m);
 585                 return (IPPROTO_DONE);
 586         }
 587
 588         /* check if received on a valid carp interface */
 589         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 590                 carpstats.carps_badif++;
 591                 CARP_LOG("carp6_input: packet received on non-carp "
 592                     "interface: %s\n",
 593                     m->m_pkthdr.rcvif->if_xname);
 594                 m_freem(m);
 595                 return (IPPROTO_DONE);
 596         }
 597
 598         /* verify that the IP TTL is 255 */
 599         if (ip6->ip6_hlim != CARP_DFLTTL) {
 600                 carpstats.carps_badttl++;
 601                 CARP_LOG("carp6_input: received ttl %d != 255 on %s\n",
 602                     ip6->ip6_hlim,
 603                     m->m_pkthdr.rcvif->if_xname);
 604                 m_freem(m);
 605                 return (IPPROTO_DONE);
 606         }
 607
 608         /* verify that we have a complete carp packet */
 609         len = m->m_len;
 610         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 611         if (ch == NULL) {
 612                 carpstats.carps_badlen++;
 613                 CARP_LOG("carp6_input: packet size %u too small\n", len);
 614                 return (IPPROTO_DONE);
 615         }
 616
 617
 618         /* verify the CARP checksum */
 619         m->m_data += *offp;
 620         if (carp_cksum(m, sizeof(*ch))) {
 621                 carpstats.carps_badsum++;
 622                 CARP_LOG("carp6_input: checksum failed, on %s\n",
 623                     m->m_pkthdr.rcvif->if_xname);
 624                 m_freem(m);
 625                 return (IPPROTO_DONE);
 626         }
 627         m->m_data -= *offp;
 628
 629         carp_input_c(m, ch, AF_INET6);
 630         return (IPPROTO_DONE);
 631 }
 632 #endif /* INET6 */
 633
 634 static void
 635 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 636 {
 637         struct ifnet *ifp = m->m_pkthdr.rcvif;
 638         struct carp_softc *sc;
 639         u_int64_t tmp_counter;
 640         struct timeval sc_tv, ch_tv;
 641
 642         /* verify that the VHID is valid on the receiving interface */
 643         CARP_LOCK(ifp->if_carp);
 644         TAILQ_FOREACH(sc, &((struct carp_if *)ifp->if_carp)->vhif_vrs, sc_list)
 645                 if (sc->sc_vhid == ch->carp_vhid)
 646                         break;
 647
 648         if (!sc || !((SC2IFP(sc)->if_flags & IFF_UP) &&
 649             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING))) {
 650                 carpstats.carps_badvhid++;
 651                 CARP_UNLOCK(ifp->if_carp);
 652                 m_freem(m);
 653                 return;
 654         }
 655
 656         getmicrotime(&SC2IFP(sc)->if_lastchange);
 657         SC2IFP(sc)->if_ipackets++;
 658         SC2IFP(sc)->if_ibytes += m->m_pkthdr.len;
 659
 660         if (bpf_peers_present(SC2IFP(sc)->if_bpf)) {
 661                 struct ip *ip = mtod(m, struct ip *);
 662                 uint32_t af1 = af;
 663
 664                 /* BPF wants net byte order */
 665                 ip->ip_len = htons(ip->ip_len + (ip->ip_hl << 2));
 666                 ip->ip_off = htons(ip->ip_off);
 667                 bpf_mtap2(SC2IFP(sc)->if_bpf, &af1, sizeof(af1), m);
 668         }
 669
 670         /* verify the CARP version. */
 671         if (ch->carp_version != CARP_VERSION) {
 672                 carpstats.carps_badver++;
 673                 SC2IFP(sc)->if_ierrors++;
 674                 CARP_UNLOCK(ifp->if_carp);
 675                 CARP_LOG("%s; invalid version %d\n",
 676                     SC2IFP(sc)->if_xname,
 677                     ch->carp_version);
 678                 m_freem(m);
 679                 return;
 680         }
 681
 682         /* verify the hash */
 683         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 684                 carpstats.carps_badauth++;
 685                 SC2IFP(sc)->if_ierrors++;
 686                 CARP_UNLOCK(ifp->if_carp);
 687                 CARP_LOG("%s: incorrect hash\n", SC2IFP(sc)->if_xname);
 688                 m_freem(m);
 689                 return;
 690         }
 691
 692         tmp_counter = ntohl(ch->carp_counter[0]);
 693         tmp_counter = tmp_counter<<32;
 694         tmp_counter += ntohl(ch->carp_counter[1]);
 695
 696         /* XXX Replay protection goes here */
 697
 698         sc->sc_init_counter = 0;
 699         sc->sc_counter = tmp_counter;
 700
 701         sc_tv.tv_sec = sc->sc_advbase;
 702         if (carp_suppress_preempt && sc->sc_advskew <  240)
 703                 sc_tv.tv_usec = 240 * 1000000 / 256;
 704         else
 705                 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256;
 706         ch_tv.tv_sec = ch->carp_advbase;
 707         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 708
 709         switch (sc->sc_state) {
 710         case INIT:
 711                 break;
 712         case MASTER:
 713                 /*
 714                  * If we receive an advertisement from a master who's going to
 715                  * be more frequent than us, go into BACKUP state.
 716                  */
 717                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 718                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 719                         callout_stop(&sc->sc_ad_tmo);
 720                         CARP_DEBUG("%s: MASTER -> BACKUP "
 721                            "(more frequent advertisement received)\n",
 722                            SC2IFP(sc)->if_xname);
 723                         carp_set_state(sc, BACKUP);
 724                         carp_setrun(sc, 0);
 725                         carp_setroute(sc, RTM_DELETE);
 726                 }
 727                 break;
 728         case BACKUP:
 729                 /*
 730                  * If we're pre-empting masters who advertise slower than us,
 731                  * and this one claims to be slower, treat him as down.
 732                  */
 733                 if (carp_opts[CARPCTL_PREEMPT] &&
 734                     timevalcmp(&sc_tv, &ch_tv, <)) {
 735                         CARP_DEBUG("%s: BACKUP -> MASTER "
 736                             "(preempting a slower master)\n",
 737                             SC2IFP(sc)->if_xname);
 738                         carp_master_down_locked(sc);
 739                         break;
 740                 }
 741
 742                 /*
 743                  *  If the master is going to advertise at such a low frequency
 744                  *  that he's guaranteed to time out, we'd might as well just
 745                  *  treat him as timed out now.
 746                  */
 747                 sc_tv.tv_sec = sc->sc_advbase * 3;
 748                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 749                         CARP_DEBUG("%s: BACKUP -> MASTER "
 750                             "(master timed out)\n",
 751                             SC2IFP(sc)->if_xname);
 752                         carp_master_down_locked(sc);
 753                         break;
 754                 }
 755
 756                 /*
 757                  * Otherwise, we reset the counter and wait for the next
 758                  * advertisement.
 759                  */
 760                 carp_setrun(sc, af);
 761                 break;
 762         }
 763
 764         CARP_UNLOCK(ifp->if_carp);
 765
 766         m_freem(m);
 767         return;
 768 }
 769
 770 static int
 771 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 772 {
 773         struct m_tag *mtag;
 774         struct ifnet *ifp = SC2IFP(sc);
 775
 776         if (sc->sc_init_counter) {
 777                 /* this could also be seconds since unix epoch */
 778                 sc->sc_counter = arc4random();
 779                 sc->sc_counter = sc->sc_counter << 32;
 780                 sc->sc_counter += arc4random();
 781         } else
 782                 sc->sc_counter++;
 783
 784         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 785         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 786
 787         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 788
 789         /* Tag packet for carp_output */
 790         mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct ifnet *), M_NOWAIT);
 791         if (mtag == NULL) {
 792                 m_freem(m);
 793                 SC2IFP(sc)->if_oerrors++;
 794                 return (ENOMEM);
 795         }
 796         bcopy(&ifp, (caddr_t)(mtag + 1), sizeof(struct ifnet *));
 797         m_tag_prepend(m, mtag);
 798
 799         return (0);
 800 }
 801
 802 static void
 803 carp_send_ad_all(void)
 804 {
 805         struct carp_softc *sc;
 806
 807         mtx_lock(&carp_mtx);
 808         LIST_FOREACH(sc, &carpif_list, sc_next) {
 809                 if (sc->sc_carpdev == NULL)
 810                         continue;
 811                 CARP_SCLOCK(sc);
 812                 if ((SC2IFP(sc)->if_flags & IFF_UP) &&
 813                     (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING) &&
 814                      sc->sc_state == MASTER)
 815                         carp_send_ad_locked(sc);
 816                 CARP_SCUNLOCK(sc);
 817         }
 818         mtx_unlock(&carp_mtx);
 819 }
 820
 821 static void
 822 carp_send_ad(void *v)
 823 {
 824         struct carp_softc *sc = v;
 825
 826         CARP_SCLOCK(sc);
 827         carp_send_ad_locked(sc);
 828         CARP_SCUNLOCK(sc);
 829 }
 830
 831 static void
 832 carp_send_ad_locked(struct carp_softc *sc)
 833 {
 834         struct carp_header ch;
 835         struct timeval tv;
 836         struct carp_header *ch_ptr;
 837         struct mbuf *m;
 838         int len, advbase, advskew;
 839
 840         CARP_SCLOCK_ASSERT(sc);
 841
 842         /* bow out if we've lost our UPness or RUNNINGuiness */
 843         if (!((SC2IFP(sc)->if_flags & IFF_UP) &&
 844             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING))) {
 845                 advbase = 255;
 846                 advskew = 255;
 847         } else {
 848                 advbase = sc->sc_advbase;
 849                 if (!carp_suppress_preempt || sc->sc_advskew > 240)
 850                         advskew = sc->sc_advskew;
 851                 else
 852                         advskew = 240;
 853                 tv.tv_sec = advbase;
 854                 tv.tv_usec = advskew * 1000000 / 256;
 855         }
 856
 857         ch.carp_version = CARP_VERSION;
 858         ch.carp_type = CARP_ADVERTISEMENT;
 859         ch.carp_vhid = sc->sc_vhid;
 860         ch.carp_advbase = advbase;
 861         ch.carp_advskew = advskew;
 862         ch.carp_authlen = 7;    /* XXX DEFINE */
 863         ch.carp_pad1 = 0;       /* must be zero */
 864         ch.carp_cksum = 0;
 865
 866 #ifdef INET
 867         if (sc->sc_ia) {
 868                 struct ip *ip;
 869
 870                 MGETHDR(m, M_DONTWAIT, MT_HEADER);
 871                 if (m == NULL) {
 872                         SC2IFP(sc)->if_oerrors++;
 873                         carpstats.carps_onomem++;
 874                         /* XXX maybe less ? */
 875                         if (advbase != 255 || advskew != 255)
 876                                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
 877                                     carp_send_ad, sc);
 878                         return;
 879                 }
 880                 len = sizeof(*ip) + sizeof(ch);
 881                 m->m_pkthdr.len = len;
 882                 m->m_pkthdr.rcvif = NULL;
 883                 m->m_len = len;
 884                 MH_ALIGN(m, m->m_len);
 885                 m->m_flags |= M_MCAST;
 886                 ip = mtod(m, struct ip *);
 887                 ip->ip_v = IPVERSION;
 888                 ip->ip_hl = sizeof(*ip) >> 2;
 889                 ip->ip_tos = IPTOS_LOWDELAY;
 890                 ip->ip_len = len;
 891                 ip->ip_id = ip_newid();
 892                 ip->ip_off = IP_DF;
 893                 ip->ip_ttl = CARP_DFLTTL;
 894                 ip->ip_p = IPPROTO_CARP;
 895                 ip->ip_sum = 0;
 896                 ip->ip_src.s_addr = sc->sc_ia->ia_addr.sin_addr.s_addr;
 897                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 898
 899                 ch_ptr = (struct carp_header *)(&ip[1]);
 900                 bcopy(&ch, ch_ptr, sizeof(ch));
 901                 if (carp_prepare_ad(m, sc, ch_ptr))
 902                         return;
 903
 904                 m->m_data += sizeof(*ip);
 905                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip));
 906                 m->m_data -= sizeof(*ip);
 907
 908                 getmicrotime(&SC2IFP(sc)->if_lastchange);
 909                 SC2IFP(sc)->if_opackets++;
 910                 SC2IFP(sc)->if_obytes += len;
 911                 carpstats.carps_opackets++;
 912
 913                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, NULL)) {
 914                         SC2IFP(sc)->if_oerrors++;
 915                         if (sc->sc_sendad_errors < INT_MAX)
 916                                 sc->sc_sendad_errors++;
 917                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 918                                 carp_suppress_preempt++;
 919                                 if (carp_suppress_preempt == 1) {
 920                                         CARP_SCUNLOCK(sc);
 921                                         carp_send_ad_all();
 922                                         CARP_SCLOCK(sc);
 923                                 }
 924                         }
 925                         sc->sc_sendad_success = 0;
 926                 } else {
 927                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 928                                 if (++sc->sc_sendad_success >=
 929                                     CARP_SENDAD_MIN_SUCCESS) {
 930                                         carp_suppress_preempt--;
 931                                         sc->sc_sendad_errors = 0;
 932                                 }
 933                         } else
 934                                 sc->sc_sendad_errors = 0;
 935                 }
 936         }
 937 #endif /* INET */
 938 #ifdef INET6
 939         if (sc->sc_ia6) {
 940                 struct ip6_hdr *ip6;
 941
 942                 MGETHDR(m, M_DONTWAIT, MT_HEADER);
 943                 if (m == NULL) {
 944                         SC2IFP(sc)->if_oerrors++;
 945                         carpstats.carps_onomem++;
 946                         /* XXX maybe less ? */
 947                         if (advbase != 255 || advskew != 255)
 948                                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
 949                                     carp_send_ad, sc);
 950                         return;
 951                 }
 952                 len = sizeof(*ip6) + sizeof(ch);
 953                 m->m_pkthdr.len = len;
 954                 m->m_pkthdr.rcvif = NULL;
 955                 m->m_len = len;
 956                 MH_ALIGN(m, m->m_len);
 957                 m->m_flags |= M_MCAST;
 958                 ip6 = mtod(m, struct ip6_hdr *);
 959                 bzero(ip6, sizeof(*ip6));
 960                 ip6->ip6_vfc |= IPV6_VERSION;
 961                 ip6->ip6_hlim = CARP_DFLTTL;
 962                 ip6->ip6_nxt = IPPROTO_CARP;
 963                 bcopy(&sc->sc_ia6->ia_addr.sin6_addr, &ip6->ip6_src,
 964                     sizeof(struct in6_addr));
 965                 /* set the multicast destination */
 966
 967                 ip6->ip6_dst.s6_addr8[0] = 0xff;
 968                 ip6->ip6_dst.s6_addr8[1] = 0x02;
 969                 ip6->ip6_dst.s6_addr8[15] = 0x12;
 970
 971                 ch_ptr = (struct carp_header *)(&ip6[1]);
 972                 bcopy(&ch, ch_ptr, sizeof(ch));
 973                 if (carp_prepare_ad(m, sc, ch_ptr))
 974                         return;
 975
 976                 m->m_data += sizeof(*ip6);
 977                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6));
 978                 m->m_data -= sizeof(*ip6);
 979
 980                 getmicrotime(&SC2IFP(sc)->if_lastchange);
 981                 SC2IFP(sc)->if_opackets++;
 982                 SC2IFP(sc)->if_obytes += len;
 983                 carpstats.carps_opackets6++;
 984
 985                 if (ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL)) {
 986                         SC2IFP(sc)->if_oerrors++;
 987                         if (sc->sc_sendad_errors < INT_MAX)
 988                                 sc->sc_sendad_errors++;
 989                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 990                                 carp_suppress_preempt++;
 991                                 if (carp_suppress_preempt == 1) {
 992                                         CARP_SCUNLOCK(sc);
 993                                         carp_send_ad_all();
 994                                         CARP_SCLOCK(sc);
 995                                 }
 996                         }
 997                         sc->sc_sendad_success = 0;
 998                 } else {
 999                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
1000                                 if (++sc->sc_sendad_success >=
1001                                     CARP_SENDAD_MIN_SUCCESS) {
1002                                         carp_suppress_preempt--;
1003                                         sc->sc_sendad_errors = 0;
1004                                 }
1005                         } else
1006                                 sc->sc_sendad_errors = 0;
1007                 }
1008         }
1009 #endif /* INET6 */
1010
1011         if (advbase != 255 || advskew != 255)
1012                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1013                     carp_send_ad, sc);
1014
1015 }
1016
1017 /*
1018  * Broadcast a gratuitous ARP request containing
1019  * the virtual router MAC address for each IP address
1020  * associated with the virtual router.
1021  */
1022 static void
1023 carp_send_arp(struct carp_softc *sc)
1024 {
1025         struct ifaddr *ifa;
1026
1027         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1028
1029                 if (ifa->ifa_addr->sa_family != AF_INET)
1030                         continue;
1031
1032 /*              arprequest(sc->sc_carpdev, &in, &in, IF_LLADDR(sc->sc_ifp)); */
1033                 arp_ifinit2(sc->sc_carpdev, ifa, IF_LLADDR(sc->sc_ifp));
1034
1035                 DELAY(1000);    /* XXX */
1036         }
1037 }
1038
1039 #ifdef INET6
1040 static void
1041 carp_send_na(struct carp_softc *sc)
1042 {
1043         struct ifaddr *ifa;
1044         struct in6_addr *in6;
1045         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1046
1047         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1048
1049                 if (ifa->ifa_addr->sa_family != AF_INET6)
1050                         continue;
1051
1052                 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr;
1053                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1054                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1055                 DELAY(1000);    /* XXX */
1056         }
1057 }
1058 #endif /* INET6 */
1059
1060 static int
1061 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type)
1062 {
1063         struct carp_softc *vh;
1064         struct ifaddr *ifa;
1065         int count = 0;
1066
1067         CARP_LOCK_ASSERT(cif);
1068
1069         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1070                 if ((type == CARP_COUNT_RUNNING &&
1071                     (SC2IFP(vh)->if_flags & IFF_UP) &&
1072                     (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING)) ||
1073                     (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) {
1074                         TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1075                             ifa_list) {
1076                                 if (ifa->ifa_addr->sa_family == AF_INET &&
1077                                     ia->ia_addr.sin_addr.s_addr ==
1078                                     ifatoia(ifa)->ia_addr.sin_addr.s_addr)
1079                                         count++;
1080                         }
1081                 }
1082         }
1083         return (count);
1084 }
1085
1086 int
1087 carp_iamatch(void *v, struct in_ifaddr *ia,
1088     struct in_addr *isaddr, u_int8_t **enaddr)
1089 {
1090         struct carp_if *cif = v;
1091         struct carp_softc *vh;
1092         int index, count = 0;
1093         struct ifaddr *ifa;
1094
1095         CARP_LOCK(cif);
1096
1097         if (carp_opts[CARPCTL_ARPBALANCE]) {
1098                 /*
1099                  * XXX proof of concept implementation.
1100                  * We use the source ip to decide which virtual host should
1101                  * handle the request. If we're master of that virtual host,
1102                  * then we respond, otherwise, just drop the arp packet on
1103                  * the floor.
1104                  */
1105                 count = carp_addrcount(cif, ia, CARP_COUNT_RUNNING);
1106                 if (count == 0) {
1107                         /* should never reach this */
1108                         CARP_UNLOCK(cif);
1109                         return (0);
1110                 }
1111
1112                 /* this should be a hash, like pf_hash() */
1113                 index = ntohl(isaddr->s_addr) % count;
1114                 count = 0;
1115
1116                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1117                         if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1118                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING)) {
1119                                 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1120                                     ifa_list) {
1121                                         if (ifa->ifa_addr->sa_family ==
1122                                             AF_INET &&
1123                                             ia->ia_addr.sin_addr.s_addr ==
1124                                             ifatoia(ifa)->ia_addr.sin_addr.s_addr) {
1125                                                 if (count == index) {
1126                                                         if (vh->sc_state ==
1127                                                             MASTER) {
1128                                                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1129                                                                 CARP_UNLOCK(cif);
1130                                                                 return (1);
1131                                                         } else {
1132                                                                 CARP_UNLOCK(cif);
1133                                                                 return (0);
1134                                                         }
1135                                                 }
1136                                                 count++;
1137                                         }
1138                                 }
1139                         }
1140                 }
1141         } else {
1142                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1143                         if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1144                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1145                             ia->ia_ifp == SC2IFP(vh) &&
1146                             vh->sc_state == MASTER) {
1147                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1148                                 CARP_UNLOCK(cif);
1149                                 return (1);
1150                         }
1151                 }
1152         }
1153         CARP_UNLOCK(cif);
1154         return (0);
1155 }
1156
1157 #ifdef INET6
1158 struct ifaddr *
1159 carp_iamatch6(void *v, struct in6_addr *taddr)
1160 {
1161         struct carp_if *cif = v;
1162         struct carp_softc *vh;
1163         struct ifaddr *ifa;
1164
1165         CARP_LOCK(cif);
1166         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1167                 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist, ifa_list) {
1168                         if (IN6_ARE_ADDR_EQUAL(taddr,
1169                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1170                             (SC2IFP(vh)->if_flags & IFF_UP) &&
1171                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1172                             vh->sc_state == MASTER) {
1173                                 CARP_UNLOCK(cif);
1174                                 return (ifa);
1175                         }
1176                 }
1177         }
1178         CARP_UNLOCK(cif);
1179
1180         return (NULL);
1181 }
1182
1183 void *
1184 carp_macmatch6(void *v, struct mbuf *m, const struct in6_addr *taddr)
1185 {
1186         struct m_tag *mtag;
1187         struct carp_if *cif = v;
1188         struct carp_softc *sc;
1189         struct ifaddr *ifa;
1190
1191         CARP_LOCK(cif);
1192         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) {
1193                 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1194                         if (IN6_ARE_ADDR_EQUAL(taddr,
1195                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1196                             (SC2IFP(sc)->if_flags & IFF_UP) &&
1197                             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING)) {
1198                                 struct ifnet *ifp = SC2IFP(sc);
1199                                 mtag = m_tag_get(PACKET_TAG_CARP,
1200                                     sizeof(struct ifnet *), M_NOWAIT);
1201                                 if (mtag == NULL) {
1202                                         /* better a bit than nothing */
1203                                         CARP_UNLOCK(cif);
1204                                         return (IF_LLADDR(sc->sc_ifp));
1205                                 }
1206                                 bcopy(&ifp, (caddr_t)(mtag + 1),
1207                                     sizeof(struct ifnet *));
1208                                 m_tag_prepend(m, mtag);
1209
1210                                 CARP_UNLOCK(cif);
1211                                 return (IF_LLADDR(sc->sc_ifp));
1212                         }
1213                 }
1214         }
1215         CARP_UNLOCK(cif);
1216
1217         return (NULL);
1218 }
1219 #endif
1220
1221 struct ifnet *
1222 carp_forus(void *v, void *dhost)
1223 {
1224         struct carp_if *cif = v;
1225         struct carp_softc *vh;
1226         u_int8_t *ena = dhost;
1227
1228         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1229                 return (NULL);
1230
1231         CARP_LOCK(cif);
1232         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list)
1233                 if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1234                     (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1235                     vh->sc_state == MASTER &&
1236                     !bcmp(dhost, IF_LLADDR(vh->sc_ifp), ETHER_ADDR_LEN)) {
1237                         CARP_UNLOCK(cif);
1238                         return (SC2IFP(vh));
1239                 }
1240
1241         CARP_UNLOCK(cif);
1242         return (NULL);
1243 }
1244
1245 static void
1246 carp_master_down(void *v)
1247 {
1248         struct carp_softc *sc = v;
1249
1250         CARP_SCLOCK(sc);
1251         carp_master_down_locked(sc);
1252         CARP_SCUNLOCK(sc);
1253 }
1254
1255 static void
1256 carp_master_down_locked(struct carp_softc *sc)
1257 {
1258         if (sc->sc_carpdev)
1259                 CARP_SCLOCK_ASSERT(sc);
1260
1261         switch (sc->sc_state) {
1262         case INIT:
1263                 printf("%s: master_down event in INIT state\n",
1264                     SC2IFP(sc)->if_xname);
1265                 break;
1266         case MASTER:
1267                 break;
1268         case BACKUP:
1269                 carp_set_state(sc, MASTER);
1270                 carp_send_ad_locked(sc);
1271                 carp_send_arp(sc);
1272 #ifdef INET6
1273                 carp_send_na(sc);
1274 #endif /* INET6 */
1275                 carp_setrun(sc, 0);
1276                 carp_setroute(sc, RTM_ADD);
1277                 break;
1278         }
1279 }
1280
1281 /*
1282  * When in backup state, af indicates whether to reset the master down timer
1283  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1284  */
1285 static void
1286 carp_setrun(struct carp_softc *sc, sa_family_t af)
1287 {
1288         struct timeval tv;
1289
1290         if (sc->sc_carpdev == NULL) {
1291                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1292                 carp_set_state(sc, INIT);
1293                 return;
1294         } else
1295                 CARP_SCLOCK_ASSERT(sc);
1296
1297         if (SC2IFP(sc)->if_flags & IFF_UP &&
1298             sc->sc_vhid > 0 && (sc->sc_naddrs || sc->sc_naddrs6))
1299                 SC2IFP(sc)->if_drv_flags |= IFF_DRV_RUNNING;
1300         else {
1301                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1302                 carp_setroute(sc, RTM_DELETE);
1303                 return;
1304         }
1305
1306         switch (sc->sc_state) {
1307         case INIT:
1308                 if (carp_opts[CARPCTL_PREEMPT] && !carp_suppress_preempt) {
1309                         carp_send_ad_locked(sc);
1310                         carp_send_arp(sc);
1311 #ifdef INET6
1312                         carp_send_na(sc);
1313 #endif /* INET6 */
1314                         CARP_DEBUG("%s: INIT -> MASTER (preempting)\n",
1315                             SC2IFP(sc)->if_xname);
1316                         carp_set_state(sc, MASTER);
1317                         carp_setroute(sc, RTM_ADD);
1318                 } else {
1319                         CARP_DEBUG("%s: INIT -> BACKUP\n", SC2IFP(sc)->if_xname);
1320                         carp_set_state(sc, BACKUP);
1321                         carp_setroute(sc, RTM_DELETE);
1322                         carp_setrun(sc, 0);
1323                 }
1324                 break;
1325         case BACKUP:
1326                 callout_stop(&sc->sc_ad_tmo);
1327                 tv.tv_sec = 3 * sc->sc_advbase;
1328                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1329                 switch (af) {
1330 #ifdef INET
1331                 case AF_INET:
1332                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1333                             carp_master_down, sc);
1334                         break;
1335 #endif /* INET */
1336 #ifdef INET6
1337                 case AF_INET6:
1338                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1339                             carp_master_down, sc);
1340                         break;
1341 #endif /* INET6 */
1342                 default:
1343                         if (sc->sc_naddrs)
1344                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1345                                     carp_master_down, sc);
1346                         if (sc->sc_naddrs6)
1347                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1348                                     carp_master_down, sc);
1349                         break;
1350                 }
1351                 break;
1352         case MASTER:
1353                 tv.tv_sec = sc->sc_advbase;
1354                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1355                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1356                     carp_send_ad, sc);
1357                 break;
1358         }
1359 }
1360
1361 void
1362 carp_multicast_cleanup(struct carp_softc *sc)
1363 {
1364         struct ip_moptions *imo = &sc->sc_imo;
1365 #ifdef INET6
1366         struct ip6_moptions *im6o = &sc->sc_im6o;
1367 #endif
1368         u_int16_t n = imo->imo_num_memberships;
1369
1370         /* Clean up our own multicast memberships */
1371         while (n-- > 0) {
1372                 if (imo->imo_membership[n] != NULL) {
1373                         in_delmulti(imo->imo_membership[n]);
1374                         imo->imo_membership[n] = NULL;
1375                 }
1376         }
1377         imo->imo_num_memberships = 0;
1378         imo->imo_multicast_ifp = NULL;
1379
1380 #ifdef INET6
1381         while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1382                 struct in6_multi_mship *imm =
1383                     LIST_FIRST(&im6o->im6o_memberships);
1384
1385                 LIST_REMOVE(imm, i6mm_chain);
1386                 in6_leavegroup(imm);
1387         }
1388         im6o->im6o_multicast_ifp = NULL;
1389 #endif
1390 }
1391
1392 static int
1393 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1394 {
1395         struct ifnet *ifp;
1396         struct carp_if *cif;
1397         struct in_ifaddr *ia, *ia_if;
1398         struct ip_moptions *imo = &sc->sc_imo;
1399         struct in_addr addr;
1400         u_long iaddr = htonl(sin->sin_addr.s_addr);
1401         int own, error;
1402
1403         if (sin->sin_addr.s_addr == 0) {
1404                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1405                         carp_set_state(sc, INIT);
1406                 if (sc->sc_naddrs)
1407                         SC2IFP(sc)->if_flags |= IFF_UP;
1408                 carp_setrun(sc, 0);
1409                 return (0);
1410         }
1411
1412         /* we have to do it by hands to check we won't match on us */
1413         ia_if = NULL; own = 0;
1414         TAILQ_FOREACH(ia, &in_ifaddrhead, ia_link) {
1415                 /* and, yeah, we need a multicast-capable iface too */
1416                 if (ia->ia_ifp != SC2IFP(sc) &&
1417                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1418                     (iaddr & ia->ia_subnetmask) == ia->ia_subnet) {
1419                         if (!ia_if)
1420                                 ia_if = ia;
1421                         if (sin->sin_addr.s_addr ==
1422                             ia->ia_addr.sin_addr.s_addr)
1423                                 own++;
1424                 }
1425         }
1426
1427         if (!ia_if)
1428                 return (EADDRNOTAVAIL);
1429
1430         ia = ia_if;
1431         ifp = ia->ia_ifp;
1432
1433         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1434             (imo->imo_multicast_ifp && imo->imo_multicast_ifp != ifp))
1435                 return (EADDRNOTAVAIL);
1436
1437         if (imo->imo_num_memberships == 0) {
1438                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1439                 if ((imo->imo_membership[0] = in_addmulti(&addr, ifp)) == NULL)
1440                         return (ENOBUFS);
1441                 imo->imo_num_memberships++;
1442                 imo->imo_multicast_ifp = ifp;
1443                 imo->imo_multicast_ttl = CARP_DFLTTL;
1444                 imo->imo_multicast_loop = 0;
1445         }
1446
1447         if (!ifp->if_carp) {
1448
1449                 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP,
1450                     M_WAITOK|M_ZERO);
1451                 if (!cif) {
1452                         error = ENOBUFS;
1453                         goto cleanup;
1454                 }
1455                 if ((error = ifpromisc(ifp, 1))) {
1456                         FREE(cif, M_CARP);
1457                         goto cleanup;
1458                 }
1459
1460                 CARP_LOCK_INIT(cif);
1461                 CARP_LOCK(cif);
1462                 cif->vhif_ifp = ifp;
1463                 TAILQ_INIT(&cif->vhif_vrs);
1464                 ifp->if_carp = cif;
1465
1466         } else {
1467                 struct carp_softc *vr;
1468
1469                 cif = (struct carp_if *)ifp->if_carp;
1470                 CARP_LOCK(cif);
1471                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1472                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1473                                 CARP_UNLOCK(cif);
1474                                 error = EINVAL;
1475                                 goto cleanup;
1476                         }
1477         }
1478         sc->sc_ia = ia;
1479         sc->sc_carpdev = ifp;
1480
1481         { /* XXX prevent endless loop if already in queue */
1482         struct carp_softc *vr, *after = NULL;
1483         int myself = 0;
1484         cif = (struct carp_if *)ifp->if_carp;
1485
1486         /* XXX: cif should not change, right? So we still hold the lock */
1487         CARP_LOCK_ASSERT(cif);
1488
1489         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1490                 if (vr == sc)
1491                         myself = 1;
1492                 if (vr->sc_vhid < sc->sc_vhid)
1493                         after = vr;
1494         }
1495
1496         if (!myself) {
1497                 /* We're trying to keep things in order */
1498                 if (after == NULL) {
1499                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1500                 } else {
1501                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1502                 }
1503                 cif->vhif_nvrs++;
1504         }
1505         }
1506
1507         sc->sc_naddrs++;
1508         SC2IFP(sc)->if_flags |= IFF_UP;
1509         if (own)
1510                 sc->sc_advskew = 0;
1511         carp_sc_state_locked(sc);
1512         carp_setrun(sc, 0);
1513
1514         CARP_UNLOCK(cif);
1515
1516         return (0);
1517
1518 cleanup:
1519         in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1520         return (error);
1521 }
1522
1523 static int
1524 carp_del_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1525 {
1526         int error = 0;
1527
1528         if (!--sc->sc_naddrs) {
1529                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1530                 struct ip_moptions *imo = &sc->sc_imo;
1531
1532                 CARP_LOCK(cif);
1533                 callout_stop(&sc->sc_ad_tmo);
1534                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1535                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1536                 sc->sc_vhid = -1;
1537                 in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1538                 imo->imo_multicast_ifp = NULL;
1539                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1540                 if (!--cif->vhif_nvrs) {
1541                         sc->sc_carpdev->if_carp = NULL;
1542                         CARP_LOCK_DESTROY(cif);
1543                         FREE(cif, M_IFADDR);
1544                 } else {
1545                         CARP_UNLOCK(cif);
1546                 }
1547         }
1548
1549         return (error);
1550 }
1551
1552 #ifdef INET6
1553 static int
1554 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1555 {
1556         struct ifnet *ifp;
1557         struct carp_if *cif;
1558         struct in6_ifaddr *ia, *ia_if;
1559         struct ip6_moptions *im6o = &sc->sc_im6o;
1560         struct in6_multi_mship *imm;
1561         struct in6_addr in6;
1562         int own, error;
1563
1564         if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
1565                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1566                         carp_set_state(sc, INIT);
1567                 if (sc->sc_naddrs6)
1568                         SC2IFP(sc)->if_flags |= IFF_UP;
1569                 carp_setrun(sc, 0);
1570                 return (0);
1571         }
1572
1573         /* we have to do it by hands to check we won't match on us */
1574         ia_if = NULL; own = 0;
1575         for (ia = in6_ifaddr; ia; ia = ia->ia_next) {
1576                 int i;
1577
1578                 for (i = 0; i < 4; i++) {
1579                         if ((sin6->sin6_addr.s6_addr32[i] &
1580                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]) !=
1581                             (ia->ia_addr.sin6_addr.s6_addr32[i] &
1582                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]))
1583                                 break;
1584                 }
1585                 /* and, yeah, we need a multicast-capable iface too */
1586                 if (ia->ia_ifp != SC2IFP(sc) &&
1587                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1588                     (i == 4)) {
1589                         if (!ia_if)
1590                                 ia_if = ia;
1591                         if (IN6_ARE_ADDR_EQUAL(&sin6->sin6_addr,
1592                             &ia->ia_addr.sin6_addr))
1593                                 own++;
1594                 }
1595         }
1596
1597         if (!ia_if)
1598                 return (EADDRNOTAVAIL);
1599         ia = ia_if;
1600         ifp = ia->ia_ifp;
1601
1602         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1603             (im6o->im6o_multicast_ifp && im6o->im6o_multicast_ifp != ifp))
1604                 return (EADDRNOTAVAIL);
1605
1606         if (!sc->sc_naddrs6) {
1607                 im6o->im6o_multicast_ifp = ifp;
1608
1609                 /* join CARP multicast address */
1610                 bzero(&in6, sizeof(in6));
1611                 in6.s6_addr16[0] = htons(0xff02);
1612                 in6.s6_addr8[15] = 0x12;
1613                 if (in6_setscope(&in6, ifp, NULL) != 0)
1614                         goto cleanup;
1615                 if ((imm = in6_joingroup(ifp, &in6, &error, 0)) == NULL)
1616                         goto cleanup;
1617                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1618
1619                 /* join solicited multicast address */
1620                 bzero(&in6, sizeof(in6));
1621                 in6.s6_addr16[0] = htons(0xff02);
1622                 in6.s6_addr32[1] = 0;
1623                 in6.s6_addr32[2] = htonl(1);
1624                 in6.s6_addr32[3] = sin6->sin6_addr.s6_addr32[3];
1625                 in6.s6_addr8[12] = 0xff;
1626                 if (in6_setscope(&in6, ifp, NULL) != 0)
1627                         goto cleanup;
1628                 if ((imm = in6_joingroup(ifp, &in6, &error, 0)) == NULL)
1629                         goto cleanup;
1630                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1631         }
1632
1633         if (!ifp->if_carp) {
1634                 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP,
1635                     M_WAITOK|M_ZERO);
1636                 if (!cif) {
1637                         error = ENOBUFS;
1638                         goto cleanup;
1639                 }
1640                 if ((error = ifpromisc(ifp, 1))) {
1641                         FREE(cif, M_CARP);
1642                         goto cleanup;
1643                 }
1644
1645                 CARP_LOCK_INIT(cif);
1646                 CARP_LOCK(cif);
1647                 cif->vhif_ifp = ifp;
1648                 TAILQ_INIT(&cif->vhif_vrs);
1649                 ifp->if_carp = cif;
1650
1651         } else {
1652                 struct carp_softc *vr;
1653
1654                 cif = (struct carp_if *)ifp->if_carp;
1655                 CARP_LOCK(cif);
1656                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1657                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1658                                 CARP_UNLOCK(cif);
1659                                 error = EINVAL;
1660                                 goto cleanup;
1661                         }
1662         }
1663         sc->sc_ia6 = ia;
1664         sc->sc_carpdev = ifp;
1665
1666         { /* XXX prevent endless loop if already in queue */
1667         struct carp_softc *vr, *after = NULL;
1668         int myself = 0;
1669         cif = (struct carp_if *)ifp->if_carp;
1670         CARP_LOCK_ASSERT(cif);
1671
1672         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1673                 if (vr == sc)
1674                         myself = 1;
1675                 if (vr->sc_vhid < sc->sc_vhid)
1676                         after = vr;
1677         }
1678
1679         if (!myself) {
1680                 /* We're trying to keep things in order */
1681                 if (after == NULL) {
1682                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1683                 } else {
1684                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1685                 }
1686                 cif->vhif_nvrs++;
1687         }
1688         }
1689
1690         sc->sc_naddrs6++;
1691         SC2IFP(sc)->if_flags |= IFF_UP;
1692         if (own)
1693                 sc->sc_advskew = 0;
1694         carp_sc_state_locked(sc);
1695         carp_setrun(sc, 0);
1696
1697         CARP_UNLOCK(cif);
1698
1699         return (0);
1700
1701 cleanup:
1702         /* clean up multicast memberships */
1703         if (!sc->sc_naddrs6) {
1704                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1705                         imm = LIST_FIRST(&im6o->im6o_memberships);
1706                         LIST_REMOVE(imm, i6mm_chain);
1707                         in6_leavegroup(imm);
1708                 }
1709         }
1710         return (error);
1711 }
1712
1713 static int
1714 carp_del_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1715 {
1716         int error = 0;
1717
1718         if (!--sc->sc_naddrs6) {
1719                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1720                 struct ip6_moptions *im6o = &sc->sc_im6o;
1721
1722                 CARP_LOCK(cif);
1723                 callout_stop(&sc->sc_ad_tmo);
1724                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1725                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1726                 sc->sc_vhid = -1;
1727                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1728                         struct in6_multi_mship *imm =
1729                             LIST_FIRST(&im6o->im6o_memberships);
1730
1731                         LIST_REMOVE(imm, i6mm_chain);
1732                         in6_leavegroup(imm);
1733                 }
1734                 im6o->im6o_multicast_ifp = NULL;
1735                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1736                 if (!--cif->vhif_nvrs) {
1737                         CARP_LOCK_DESTROY(cif);
1738                         sc->sc_carpdev->if_carp = NULL;
1739                         FREE(cif, M_IFADDR);
1740                 } else
1741                         CARP_UNLOCK(cif);
1742         }
1743
1744         return (error);
1745 }
1746 #endif /* INET6 */
1747
1748 static int
1749 carp_ioctl(struct ifnet *ifp, u_long cmd, caddr_t addr)
1750 {
1751         struct carp_softc *sc = ifp->if_softc, *vr;
1752         struct carpreq carpr;
1753         struct ifaddr *ifa;
1754         struct ifreq *ifr;
1755         struct ifaliasreq *ifra;
1756         int locked = 0, error = 0;
1757
1758         ifa = (struct ifaddr *)addr;
1759         ifra = (struct ifaliasreq *)addr;
1760         ifr = (struct ifreq *)addr;
1761
1762         switch (cmd) {
1763         case SIOCSIFADDR:
1764                 switch (ifa->ifa_addr->sa_family) {
1765 #ifdef INET
1766                 case AF_INET:
1767                         SC2IFP(sc)->if_flags |= IFF_UP;
1768                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1769                             sizeof(struct sockaddr));
1770                         error = carp_set_addr(sc, satosin(ifa->ifa_addr));
1771                         break;
1772 #endif /* INET */
1773 #ifdef INET6
1774                 case AF_INET6:
1775                         SC2IFP(sc)->if_flags |= IFF_UP;
1776                         error = carp_set_addr6(sc, satosin6(ifa->ifa_addr));
1777                         break;
1778 #endif /* INET6 */
1779                 default:
1780                         error = EAFNOSUPPORT;
1781                         break;
1782                 }
1783                 break;
1784
1785         case SIOCAIFADDR:
1786                 switch (ifa->ifa_addr->sa_family) {
1787 #ifdef INET
1788                 case AF_INET:
1789                         SC2IFP(sc)->if_flags |= IFF_UP;
1790                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1791                             sizeof(struct sockaddr));
1792                         error = carp_set_addr(sc, satosin(&ifra->ifra_addr));
1793                         break;
1794 #endif /* INET */
1795 #ifdef INET6
1796                 case AF_INET6:
1797                         SC2IFP(sc)->if_flags |= IFF_UP;
1798                         error = carp_set_addr6(sc, satosin6(&ifra->ifra_addr));
1799                         break;
1800 #endif /* INET6 */
1801                 default:
1802                         error = EAFNOSUPPORT;
1803                         break;
1804                 }
1805                 break;
1806
1807         case SIOCDIFADDR:
1808                 switch (ifa->ifa_addr->sa_family) {
1809 #ifdef INET
1810                 case AF_INET:
1811                         error = carp_del_addr(sc, satosin(&ifra->ifra_addr));
1812                         break;
1813 #endif /* INET */
1814 #ifdef INET6
1815                 case AF_INET6:
1816                         error = carp_del_addr6(sc, satosin6(&ifra->ifra_addr));
1817                         break;
1818 #endif /* INET6 */
1819                 default:
1820                         error = EAFNOSUPPORT;
1821                         break;
1822                 }
1823                 break;
1824
1825         case SIOCSIFFLAGS:
1826                 if (sc->sc_carpdev) {
1827                         locked = 1;
1828                         CARP_SCLOCK(sc);
1829                 }
1830                 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) {
1831                         callout_stop(&sc->sc_ad_tmo);
1832                         callout_stop(&sc->sc_md_tmo);
1833                         callout_stop(&sc->sc_md6_tmo);
1834                         if (sc->sc_state == MASTER)
1835                                 carp_send_ad_locked(sc);
1836                         carp_set_state(sc, INIT);
1837                         carp_setrun(sc, 0);
1838                 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) {
1839                         SC2IFP(sc)->if_flags |= IFF_UP;
1840                         carp_setrun(sc, 0);
1841                 }
1842                 break;
1843
1844         case SIOCSVH:
1845                 if ((error = suser(curthread)) != 0)
1846                         break;
1847                 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1848                         break;
1849                 error = 1;
1850                 if (sc->sc_carpdev) {
1851                         locked = 1;
1852                         CARP_SCLOCK(sc);
1853                 }
1854                 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) {
1855                         switch (carpr.carpr_state) {
1856                         case BACKUP:
1857                                 callout_stop(&sc->sc_ad_tmo);
1858                                 carp_set_state(sc, BACKUP);
1859                                 carp_setrun(sc, 0);
1860                                 carp_setroute(sc, RTM_DELETE);
1861                                 break;
1862                         case MASTER:
1863                                 carp_master_down_locked(sc);
1864                                 break;
1865                         default:
1866                                 break;
1867                         }
1868                 }
1869                 if (carpr.carpr_vhid > 0) {
1870                         if (carpr.carpr_vhid > 255) {
1871                                 error = EINVAL;
1872                                 break;
1873                         }
1874                         if (sc->sc_carpdev) {
1875                                 struct carp_if *cif;
1876                                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1877                                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1878                                         if (vr != sc &&
1879                                             vr->sc_vhid == carpr.carpr_vhid)
1880                                                 return EEXIST;
1881                         }
1882                         sc->sc_vhid = carpr.carpr_vhid;
1883                         IF_LLADDR(sc->sc_ifp)[0] = 0;
1884                         IF_LLADDR(sc->sc_ifp)[1] = 0;
1885                         IF_LLADDR(sc->sc_ifp)[2] = 0x5e;
1886                         IF_LLADDR(sc->sc_ifp)[3] = 0;
1887                         IF_LLADDR(sc->sc_ifp)[4] = 1;
1888                         IF_LLADDR(sc->sc_ifp)[5] = sc->sc_vhid;
1889                         error--;
1890                 }
1891                 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) {
1892                         if (carpr.carpr_advskew >= 255) {
1893                                 error = EINVAL;
1894                                 break;
1895                         }
1896                         if (carpr.carpr_advbase > 255) {
1897                                 error = EINVAL;
1898                                 break;
1899                         }
1900                         sc->sc_advbase = carpr.carpr_advbase;
1901                         sc->sc_advskew = carpr.carpr_advskew;
1902                         error--;
1903                 }
1904                 bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1905                 if (error > 0)
1906                         error = EINVAL;
1907                 else {
1908                         error = 0;
1909                         carp_setrun(sc, 0);
1910                 }
1911                 break;
1912
1913         case SIOCGVH:
1914                 /* XXX: lockless read */
1915                 bzero(&carpr, sizeof(carpr));
1916                 carpr.carpr_state = sc->sc_state;
1917                 carpr.carpr_vhid = sc->sc_vhid;
1918                 carpr.carpr_advbase = sc->sc_advbase;
1919                 carpr.carpr_advskew = sc->sc_advskew;
1920                 if (suser(curthread) == 0)
1921                         bcopy(sc->sc_key, carpr.carpr_key,
1922                             sizeof(carpr.carpr_key));
1923                 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1924                 break;
1925
1926         default:
1927                 error = EINVAL;
1928         }
1929
1930         if (locked)
1931                 CARP_SCUNLOCK(sc);
1932
1933         carp_hmac_prepare(sc);
1934
1935         return (error);
1936 }
1937
1938 /*
1939  * XXX: this is looutput. We should eventually use it from there.
1940  */
1941 static int
1942 carp_looutput(struct ifnet *ifp, struct mbuf *m, struct sockaddr *dst,
1943     struct rtentry *rt)
1944 {
1945         u_int32_t af;
1946
1947         M_ASSERTPKTHDR(m); /* check if we have the packet header */
1948
1949         if (rt && rt->rt_flags & (RTF_REJECT|RTF_BLACKHOLE)) {
1950                 m_freem(m);
1951                 return (rt->rt_flags & RTF_BLACKHOLE ? 0 :
1952                         rt->rt_flags & RTF_HOST ? EHOSTUNREACH : ENETUNREACH);
1953         }
1954
1955         ifp->if_opackets++;
1956         ifp->if_obytes += m->m_pkthdr.len;
1957
1958         /* BPF writes need to be handled specially. */
1959         if (dst->sa_family == AF_UNSPEC) {
1960                 bcopy(dst->sa_data, &af, sizeof(af));
1961                 dst->sa_family = af;
1962         }
1963
1964 #if 1   /* XXX */
1965         switch (dst->sa_family) {
1966         case AF_INET:
1967         case AF_INET6:
1968         case AF_IPX:
1969         case AF_APPLETALK:
1970                 break;
1971         default:
1972                 printf("carp_looutput: af=%d unexpected\n", dst->sa_family);
1973                 m_freem(m);
1974                 return (EAFNOSUPPORT);
1975         }
1976 #endif
1977         return(if_simloop(ifp, m, dst->sa_family, 0));
1978 }
1979
1980 /*
1981  * Start output on carp interface. This function should never be called.
1982  */
1983 static void
1984 carp_start(struct ifnet *ifp)
1985 {
1986 #ifdef DEBUG
1987         printf("%s: start called\n", ifp->if_xname);
1988 #endif
1989 }
1990
1991 int
1992 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *sa,
1993     struct rtentry *rt)
1994 {
1995         struct m_tag *mtag;
1996         struct carp_softc *sc;
1997         struct ifnet *carp_ifp;
1998
1999         if (!sa)
2000                 return (0);
2001
2002         switch (sa->sa_family) {
2003 #ifdef INET
2004         case AF_INET:
2005                 break;
2006 #endif /* INET */
2007 #ifdef INET6
2008         case AF_INET6:
2009                 break;
2010 #endif /* INET6 */
2011         default:
2012                 return (0);
2013         }
2014
2015         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
2016         if (mtag == NULL)
2017                 return (0);
2018
2019         bcopy(mtag + 1, &carp_ifp, sizeof(struct ifnet *));
2020         sc = carp_ifp->if_softc;
2021
2022         /* Set the source MAC address to Virtual Router MAC Address */
2023         switch (ifp->if_type) {
2024         case IFT_ETHER:
2025         case IFT_L2VLAN: {
2026                         struct ether_header *eh;
2027
2028                         eh = mtod(m, struct ether_header *);
2029                         eh->ether_shost[0] = 0;
2030                         eh->ether_shost[1] = 0;
2031                         eh->ether_shost[2] = 0x5e;
2032                         eh->ether_shost[3] = 0;
2033                         eh->ether_shost[4] = 1;
2034                         eh->ether_shost[5] = sc->sc_vhid;
2035                 }
2036                 break;
2037         case IFT_FDDI: {
2038                         struct fddi_header *fh;
2039
2040                         fh = mtod(m, struct fddi_header *);
2041                         fh->fddi_shost[0] = 0;
2042                         fh->fddi_shost[1] = 0;
2043                         fh->fddi_shost[2] = 0x5e;
2044                         fh->fddi_shost[3] = 0;
2045                         fh->fddi_shost[4] = 1;
2046                         fh->fddi_shost[5] = sc->sc_vhid;
2047                 }
2048                 break;
2049         case IFT_ISO88025: {
2050                         struct iso88025_header *th;
2051                         th = mtod(m, struct iso88025_header *);
2052                         th->iso88025_shost[0] = 3;
2053                         th->iso88025_shost[1] = 0;
2054                         th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
2055                         th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
2056                         th->iso88025_shost[4] = 0;
2057                         th->iso88025_shost[5] = 0;
2058                 }
2059                 break;
2060         default:
2061                 printf("%s: carp is not supported for this interface type\n",
2062                     ifp->if_xname);
2063                 return (EOPNOTSUPP);
2064         }
2065
2066         return (0);
2067 }
2068
2069 static void
2070 carp_set_state(struct carp_softc *sc, int state)
2071 {
2072
2073         if (sc->sc_carpdev)
2074                 CARP_SCLOCK_ASSERT(sc);
2075
2076         if (sc->sc_state == state)
2077                 return;
2078
2079         sc->sc_state = state;
2080         switch (state) {
2081         case BACKUP:
2082                 SC2IFP(sc)->if_link_state = LINK_STATE_DOWN;
2083                 break;
2084         case MASTER:
2085                 SC2IFP(sc)->if_link_state = LINK_STATE_UP;
2086                 break;
2087         default:
2088                 SC2IFP(sc)->if_link_state = LINK_STATE_UNKNOWN;
2089                 break;
2090         }
2091         rt_ifmsg(SC2IFP(sc));
2092 }
2093
2094 void
2095 carp_carpdev_state(void *v)
2096 {
2097         struct carp_if *cif = v;
2098
2099         CARP_LOCK(cif);
2100         carp_carpdev_state_locked(cif);
2101         CARP_UNLOCK(cif);
2102 }
2103
2104 static void
2105 carp_carpdev_state_locked(struct carp_if *cif)
2106 {
2107         struct carp_softc *sc;
2108
2109         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list)
2110                 carp_sc_state_locked(sc);
2111 }
2112
2113 static void
2114 carp_sc_state_locked(struct carp_softc *sc)
2115 {
2116         CARP_SCLOCK_ASSERT(sc);
2117
2118         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
2119             !(sc->sc_carpdev->if_flags & IFF_UP)) {
2120                 sc->sc_flags_backup = SC2IFP(sc)->if_flags;
2121                 SC2IFP(sc)->if_flags &= ~IFF_UP;
2122                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
2123                 callout_stop(&sc->sc_ad_tmo);
2124                 callout_stop(&sc->sc_md_tmo);
2125                 callout_stop(&sc->sc_md6_tmo);
2126                 carp_set_state(sc, INIT);
2127                 carp_setrun(sc, 0);
2128                 if (!sc->sc_suppress) {
2129                         carp_suppress_preempt++;
2130                         if (carp_suppress_preempt == 1) {
2131                                 CARP_SCUNLOCK(sc);
2132                                 carp_send_ad_all();
2133                                 CARP_SCLOCK(sc);
2134                         }
2135                 }
2136                 sc->sc_suppress = 1;
2137         } else {
2138                 SC2IFP(sc)->if_flags |= sc->sc_flags_backup;
2139                 carp_set_state(sc, INIT);
2140                 carp_setrun(sc, 0);
2141                 if (sc->sc_suppress)
2142                         carp_suppress_preempt--;
2143                 sc->sc_suppress = 0;
2144         }
2145
2146         return;
2147 }
2148
2149 static int
2150 carp_modevent(module_t mod, int type, void *data)
2151 {
2152         switch (type) {
2153         case MOD_LOAD:
2154                 if_detach_event_tag = EVENTHANDLER_REGISTER(ifnet_departure_event,
2155                     carp_ifdetach, NULL, EVENTHANDLER_PRI_ANY);
2156                 if (if_detach_event_tag == NULL)
2157                         return (ENOMEM);
2158                 mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2159                 LIST_INIT(&carpif_list);
2160                 if_clone_attach(&carp_cloner);
2161                 break;
2162
2163         case MOD_UNLOAD:
2164                 EVENTHANDLER_DEREGISTER(ifnet_departure_event, if_detach_event_tag);
2165                 if_clone_detach(&carp_cloner);
2166                 mtx_destroy(&carp_mtx);
2167                 break;
2168
2169         default:
2170                 return (EINVAL);
2171         }
2172
2173         return (0);
2174 }
2175
2176 static moduledata_t carp_mod = {
2177         "carp",
2178         carp_modevent,
2179         0
2180 };
2181
2182 DECLARE_MODULE(carp, carp_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);