sys/netinet/ip_carp.c

   1 /*
   2  * Copyright (c) 2002 Michael Shalayeff. All rights reserved.
   3  * Copyright (c) 2003 Ryan McBride. All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  *
  14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  15  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  16  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  17  * IN NO EVENT SHALL THE AUTHOR OR HIS RELATIVES BE LIABLE FOR ANY DIRECT,
  18  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  19  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  20  * SERVICES; LOSS OF MIND, USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
  22  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
  23  * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
  24  * THE POSSIBILITY OF SUCH DAMAGE.
  25  */
  26 /*
  27  * $FreeBSD: src/sys/netinet/ip_carp.c,v 1.48 2007/02/02 09:39:09 glebius Exp $
  28  * $DragonFly: src/sys/netinet/ip_carp.c,v 1.9 2008/06/08 08:38:05 sephe Exp $
  29  */
  30
  31 #include "opt_carp.h"
  32 /*#include "opt_bpf.h"*/
  33 #include "opt_inet.h"
  34 #include "opt_inet6.h"
  35
  36 #include <sys/types.h>
  37 #include <sys/param.h>
  38 #include <sys/systm.h>
  39 #include <sys/conf.h>
  40 #include <sys/kernel.h>
  41 #include <machine/limits.h>
  42 #include <sys/malloc.h>
  43 #include <sys/mbuf.h>
  44 #include <sys/module.h>
  45 #include <sys/time.h>
  46 #include <sys/proc.h>
  47 #include <sys/sysctl.h>
  48 #include <sys/syslog.h>
  49 #include <sys/signalvar.h>
  50 #include <sys/filio.h>
  51 #include <sys/sockio.h>
  52 #include <sys/in_cksum.h>
  53 #include <sys/socket.h>
  54 #include <sys/vnode.h>
  55
  56 #include <machine/stdarg.h>
  57
  58 #include <net/bpf.h>
  59 #include <net/ethernet.h>
  60 #include <net/if.h>
  61 #include <net/if_dl.h>
  62 #include <net/if_types.h>
  63 #include <net/route.h>
  64 #include <net/if_clone.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 #endif
  74
  75 #ifdef INET6
  76 #include <netinet/icmp6.h>
  77 #include <netinet/ip6.h>
  78 #include <netinet6/ip6_var.h>
  79 #include <netinet6/scope6_var.h>
  80 #include <netinet6/nd6.h>
  81 #endif
  82
  83 #include <crypto/sha1.h>
  84 #include <netinet/ip_carp.h>
  85 #include <sys/lock.h>
  86
  87 #define CARP_IFNAME     "carp"
  88 static MALLOC_DEFINE(M_CARP, "CARP", "CARP interfaces");
  89 static MALLOC_DEFINE(M_IFNET, "IFNET", "IFNET CARP?");
  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 lock     vhif_lock;
 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 #define CARP_LOCK_INIT(cif)     lockinit(&(cif)->vhif_lock, "carp_if", 0,  LK_NOWAIT);
 165 #define CARP_LOCK_DESTROY(cif)  ;
 166 #define CARP_LOCK_ASSERT(cif)   ;
 167 #define CARP_LOCK(cif)          lockmgr(&(cif)->vhif_lock, LK_EXCLUSIVE);
 168 #define CARP_UNLOCK(cif)        lockmgr(&(cif)->vhif_lock, LK_RELEASE);
 169
 170 #define CARP_SCLOCK(sc)         lockmgr(&SC2CIF(sc)->vhif_lock, LK_EXCLUSIVE);
 171 #define CARP_SCUNLOCK(sc)       lockmgr(&SC2CIF(sc)->vhif_lock, LK_RELEASE);
 172 #define CARP_SCLOCK_ASSERT(sc)  ;
 173
 174 #define CARP_LOG(...)   do {                            \
 175         if (carp_opts[CARPCTL_LOG] > 0)                 \
 176                 log(LOG_INFO, __VA_ARGS__);             \
 177 } while (0)
 178
 179 #define CARP_DEBUG(...) do {                            \
 180         if (carp_opts[CARPCTL_LOG] > 1)                 \
 181                 log(LOG_DEBUG, __VA_ARGS__);            \
 182 } while (0)
 183
 184 static void     carp_hmac_prepare(struct carp_softc *);
 185 static void     carp_hmac_generate(struct carp_softc *, u_int32_t *,
 186                     unsigned char *);
 187 static int      carp_hmac_verify(struct carp_softc *, u_int32_t *,
 188                     unsigned char *);
 189 static void     carp_setroute(struct carp_softc *, int);
 190 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 191 static int      carp_clone_create(struct if_clone *, int);
 192 static void     carp_clone_destroy(struct ifnet *);
 193 static void     carpdetach(struct carp_softc *, int);
 194 static int      carp_prepare_ad(struct mbuf *, struct carp_softc *,
 195                     struct carp_header *);
 196 static void     carp_send_ad_all(void);
 197 static void     carp_send_ad(void *);
 198 static void     carp_send_ad_locked(struct carp_softc *);
 199 static void     carp_send_arp(struct carp_softc *);
 200 static void     carp_master_down(void *);
 201 static void     carp_master_down_locked(struct carp_softc *);
 202 static int      carp_ioctl(struct ifnet *, u_long, caddr_t, struct ucred *);
 203 static int      carp_looutput(struct ifnet *, struct mbuf *, struct sockaddr *,
 204                     struct rtentry *);
 205 static void     carp_start(struct ifnet *);
 206 static void     carp_setrun(struct carp_softc *, sa_family_t);
 207 static void     carp_set_state(struct carp_softc *, int);
 208 static int      carp_addrcount(struct carp_if *, struct in_ifaddr *, int);
 209 enum    { CARP_COUNT_MASTER, CARP_COUNT_RUNNING };
 210
 211 static void     carp_multicast_cleanup(struct carp_softc *);
 212 static int      carp_set_addr(struct carp_softc *, struct sockaddr_in *);
 213 static int      carp_del_addr(struct carp_softc *, struct sockaddr_in *);
 214 static void     carp_carpdev_state_locked(struct carp_if *);
 215 static void     carp_sc_state_locked(struct carp_softc *);
 216 #ifdef INET6
 217 static void     carp_send_na(struct carp_softc *);
 218 static int      carp_set_addr6(struct carp_softc *, struct sockaddr_in6 *);
 219 static int      carp_del_addr6(struct carp_softc *, struct sockaddr_in6 *);
 220 static void     carp_multicast6_cleanup(struct carp_softc *);
 221 #endif
 222
 223 static LIST_HEAD(, carp_softc) carpif_list;
 224
 225 struct if_clone carp_cloner = IF_CLONE_INITIALIZER(CARP_IFNAME, carp_clone_create, carp_clone_destroy, 0, IF_MAXUNIT);
 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_container *ifac;
 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(ifac, &SC2IFP(sc)->if_addrheads[mycpuid], ifa_link) {
 265                 struct ifaddr *ifa = ifac->ifa;
 266
 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(ifac, &SC2IFP(sc)->if_addrheads[mycpuid], ifa_link) {
 275                 struct ifaddr *ifa = ifac->ifa;
 276
 277                 if (ifa->ifa_addr->sa_family == AF_INET6) {
 278                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 279                         in6_clearscope(&in6);
 280                         SHA1Update(&sc->sc_sha1, (void *)&in6, sizeof(in6));
 281                 }
 282         }
 283 #endif /* INET6 */
 284
 285         /* convert ipad to opad */
 286         for (i = 0; i < sizeof(sc->sc_pad); i++)
 287                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 288
 289         if (sc->sc_carpdev)
 290                 CARP_SCUNLOCK(sc);
 291 }
 292
 293 static void
 294 carp_hmac_generate(struct carp_softc *sc, u_int32_t counter[2],
 295     unsigned char md[20])
 296 {
 297         SHA1_CTX sha1ctx;
 298
 299         /* fetch first half of inner hash */
 300         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 301
 302         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 303         SHA1Final(md, &sha1ctx);
 304
 305         /* outer hash */
 306         SHA1Init(&sha1ctx);
 307         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 308         SHA1Update(&sha1ctx, md, 20);
 309         SHA1Final(md, &sha1ctx);
 310 }
 311
 312 static int
 313 carp_hmac_verify(struct carp_softc *sc, u_int32_t counter[2],
 314     unsigned char md[20])
 315 {
 316         unsigned char md2[20];
 317
 318         CARP_SCLOCK_ASSERT(sc);
 319
 320         carp_hmac_generate(sc, counter, md2);
 321
 322         return (bcmp(md, md2, sizeof(md2)));
 323 }
 324
 325 static void
 326 carp_setroute(struct carp_softc *sc, int cmd)
 327 {
 328         struct ifaddr_container *ifac;
 329
 330         if (sc->sc_carpdev)
 331                 CARP_SCLOCK_ASSERT(sc);
 332
 333         crit_enter();
 334         TAILQ_FOREACH(ifac, &SC2IFP(sc)->if_addrheads[mycpuid], ifa_link) {
 335                 struct ifaddr *ifa = ifac->ifa;
 336
 337                 if (ifa->ifa_addr->sa_family == AF_INET &&
 338                     sc->sc_carpdev != NULL) {
 339                         int count = carp_addrcount(
 340                             (struct carp_if *)sc->sc_carpdev->if_carp,
 341                             ifatoia(ifa), CARP_COUNT_MASTER);
 342
 343                         if ((cmd == RTM_ADD && count == 1) ||
 344                             (cmd == RTM_DELETE && count == 0))
 345                                 rtinit(ifa, cmd, RTF_UP | RTF_HOST);
 346                 }
 347 #ifdef INET6
 348                 if (ifa->ifa_addr->sa_family == AF_INET6) {
 349                         if (cmd == RTM_ADD)
 350                                 in6_ifaddloop(ifa);
 351                         else
 352                                 in6_ifremloop(ifa);
 353                 }
 354 #endif /* INET6 */
 355         }
 356         crit_exit();
 357
 358 }
 359
 360 static int
 361 carp_clone_create(struct if_clone *ifc, int unit)
 362 {
 363
 364         struct carp_softc *sc;
 365         struct ifnet *ifp;
 366
 367         MALLOC(sc, struct carp_softc *, sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
 368         ifp = SC2IFP(sc) = kmalloc(sizeof(struct ifnet), M_IFNET, M_WAITOK|M_ZERO);
 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
 378 #ifdef INET6
 379         sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL;
 380 #endif
 381
 382 /*      sc->sc_imo.imo_membership = kmalloc((sizeof(struct in_multi) * IP_MAX_MEMBERSHIPS), M_CARP,M_WAITOK);*/
 383 /*
 384         sc->sc_imo.imo_max_memberships = IP_MAX_MEMBERSHIPS;
 385         sc->sc_imo.imo_multicast_vif = -1;
 386 */
 387         callout_init(&sc->sc_ad_tmo);
 388         callout_init(&sc->sc_md_tmo);
 389         callout_init(&sc->sc_md6_tmo);
 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, NULL);
 402         bpfattach(ifp, DLT_NULL, sizeof(u_int));
 403
 404         crit_enter();
 405         LIST_INSERT_HEAD(&carpif_list, sc, sc_next);
 406         crit_exit();
 407
 408         return (0);
 409 }
 410
 411 static void
 412 carp_clone_destroy(struct ifnet *ifp)
 413 {
 414         struct carp_softc *sc = ifp->if_softc;
 415
 416         if (sc->sc_carpdev)
 417                 CARP_SCLOCK(sc);
 418         carpdetach(sc, 1);      /* Returns unlocked. */
 419
 420         crit_enter();
 421         LIST_REMOVE(sc, sc_next);
 422         crit_exit();
 423         bpfdetach(ifp);
 424         if_detach(ifp);
 425 /*      if_free_type(ifp, IFT_ETHER);*/
 426 /*      kfree(sc->sc_imo.imo_membership, M_CARP); */
 427         kfree(sc, M_CARP);
 428 }
 429
 430 /*
 431  * This function can be called on CARP interface destroy path,
 432  * and in case of the removal of the underlying interface as
 433  * well. We differentiate these two cases. In the latter case
 434  * we do not cleanup our multicast memberships, since they
 435  * are already freed. Also, in the latter case we do not
 436  * release the lock on return, because the function will be
 437  * called once more, for another CARP instance on the same
 438  * interface.
 439  */
 440 static void
 441 carpdetach(struct carp_softc *sc, int unlock)
 442 {
 443         struct carp_if *cif;
 444
 445         callout_stop(&sc->sc_ad_tmo);
 446         callout_stop(&sc->sc_md_tmo);
 447         callout_stop(&sc->sc_md6_tmo);
 448
 449         if (sc->sc_suppress)
 450                 carp_suppress_preempt--;
 451         sc->sc_suppress = 0;
 452
 453         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS)
 454                 carp_suppress_preempt--;
 455         sc->sc_sendad_errors = 0;
 456
 457         carp_set_state(sc, INIT);
 458         SC2IFP(sc)->if_flags &= ~IFF_UP;
 459         carp_setrun(sc, 0);
 460         if (unlock)
 461                 carp_multicast_cleanup(sc);
 462 #ifdef INET6
 463         carp_multicast6_cleanup(sc);
 464 #endif
 465
 466         if (sc->sc_carpdev != NULL) {
 467                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
 468                 CARP_LOCK_ASSERT(cif);
 469                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
 470                 if (!--cif->vhif_nvrs) {
 471                         ifpromisc(sc->sc_carpdev, 0);
 472                         sc->sc_carpdev->if_carp = NULL;
 473                         CARP_LOCK_DESTROY(cif);
 474                         FREE(cif, M_IFADDR);
 475                 } else if (unlock)
 476                         CARP_UNLOCK(cif);
 477                 sc->sc_carpdev = NULL;
 478         }
 479 }
 480
 481 /* Detach an interface from the carp. */
 482 static void
 483 carp_ifdetach(void *arg __unused, struct ifnet *ifp)
 484 {
 485         struct carp_if *cif = (struct carp_if *)ifp->if_carp;
 486         struct carp_softc *sc, *nextsc;
 487
 488         if (cif == NULL)
 489                 return;
 490
 491         /*
 492          * XXX: At the end of for() cycle the lock will be destroyed.
 493          */
 494         CARP_LOCK(cif);
 495         for (sc = TAILQ_FIRST(&cif->vhif_vrs); sc; sc = nextsc) {
 496                 nextsc = TAILQ_NEXT(sc, sc_list);
 497                 carpdetach(sc, 0);
 498         }
 499         CARP_UNLOCK(cif);
 500 }
 501
 502 /*
 503  * process input packet.
 504  * we have rearranged checks order compared to the rfc,
 505  * but it seems more efficient this way or not possible otherwise.
 506  */
 507 void
 508 carp_input(struct mbuf *m, int hlen)
 509 {
 510         struct ip *ip = mtod(m, struct ip *);
 511         struct carp_header *ch;
 512         int iplen, len;
 513
 514         carpstats.carps_ipackets++;
 515
 516         if (!carp_opts[CARPCTL_ALLOW]) {
 517                 m_freem(m);
 518                 return;
 519         }
 520
 521         /* check if received on a valid carp interface */
 522         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 523                 carpstats.carps_badif++;
 524                 CARP_LOG("carp_input: packet received on non-carp "
 525                     "interface: %s\n",
 526                     m->m_pkthdr.rcvif->if_xname);
 527                 m_freem(m);
 528                 return;
 529         }
 530
 531         /* verify that the IP TTL is 255.  */
 532         if (ip->ip_ttl != CARP_DFLTTL) {
 533                 carpstats.carps_badttl++;
 534                 CARP_LOG("carp_input: received ttl %d != 255i on %s\n",
 535                     ip->ip_ttl,
 536                     m->m_pkthdr.rcvif->if_xname);
 537                 m_freem(m);
 538                 return;
 539         }
 540
 541         iplen = ip->ip_hl << 2;
 542
 543         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 544                 carpstats.carps_badlen++;
 545                 CARP_LOG("carp_input: received len %zd < "
 546                     "sizeof(struct carp_header)\n",
 547                     m->m_len - sizeof(struct ip));
 548                 m_freem(m);
 549                 return;
 550         }
 551
 552         if (iplen + sizeof(*ch) < m->m_len) {
 553                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 554                         carpstats.carps_hdrops++;
 555                         CARP_LOG("carp_input: pullup failed\n");
 556                         return;
 557                 }
 558                 ip = mtod(m, struct ip *);
 559         }
 560         ch = (struct carp_header *)((char *)ip + iplen);
 561
 562         /*
 563          * verify that the received packet length is
 564          * equal to the CARP header
 565          */
 566         len = iplen + sizeof(*ch);
 567         if (len > m->m_pkthdr.len) {
 568                 carpstats.carps_badlen++;
 569                 CARP_LOG("carp_input: packet too short %d on %s\n",
 570                     m->m_pkthdr.len,
 571                     m->m_pkthdr.rcvif->if_xname);
 572                 m_freem(m);
 573                 return;
 574         }
 575
 576         if ((m = m_pullup(m, len)) == NULL) {
 577                 carpstats.carps_hdrops++;
 578                 return;
 579         }
 580         ip = mtod(m, struct ip *);
 581         ch = (struct carp_header *)((char *)ip + iplen);
 582
 583         /* verify the CARP checksum */
 584         m->m_data += iplen;
 585         if (carp_cksum(m, len - iplen)) {
 586                 carpstats.carps_badsum++;
 587                 CARP_LOG("carp_input: checksum failed on %s\n",
 588                     m->m_pkthdr.rcvif->if_xname);
 589                 m_freem(m);
 590                 return;
 591         }
 592         m->m_data -= iplen;
 593
 594         carp_input_c(m, ch, AF_INET);
 595 }
 596
 597 #ifdef INET6
 598 int
 599 carp6_input(struct mbuf **mp, int *offp, int proto)
 600 {
 601         struct mbuf *m = *mp;
 602         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 603         struct carp_header *ch;
 604         u_int len;
 605
 606         carpstats.carps_ipackets6++;
 607
 608         if (!carp_opts[CARPCTL_ALLOW]) {
 609                 m_freem(m);
 610                 return (IPPROTO_DONE);
 611         }
 612
 613         /* check if received on a valid carp interface */
 614         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 615                 carpstats.carps_badif++;
 616                 CARP_LOG("carp6_input: packet received on non-carp "
 617                     "interface: %s\n",
 618                     m->m_pkthdr.rcvif->if_xname);
 619                 m_freem(m);
 620                 return (IPPROTO_DONE);
 621         }
 622
 623         /* verify that the IP TTL is 255 */
 624         if (ip6->ip6_hlim != CARP_DFLTTL) {
 625                 carpstats.carps_badttl++;
 626                 CARP_LOG("carp6_input: received ttl %d != 255 on %s\n",
 627                     ip6->ip6_hlim,
 628                     m->m_pkthdr.rcvif->if_xname);
 629                 m_freem(m);
 630                 return (IPPROTO_DONE);
 631         }
 632
 633         /* verify that we have a complete carp packet */
 634         len = m->m_len;
 635         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 636         if (ch == NULL) {
 637                 carpstats.carps_badlen++;
 638                 CARP_LOG("carp6_input: packet size %u too small\n", len);
 639                 return (IPPROTO_DONE);
 640         }
 641
 642
 643         /* verify the CARP checksum */
 644         m->m_data += *offp;
 645         if (carp_cksum(m, sizeof(*ch))) {
 646                 carpstats.carps_badsum++;
 647                 CARP_LOG("carp6_input: checksum failed, on %s\n",
 648                     m->m_pkthdr.rcvif->if_xname);
 649                 m_freem(m);
 650                 return (IPPROTO_DONE);
 651         }
 652         m->m_data -= *offp;
 653
 654         carp_input_c(m, ch, AF_INET6);
 655         return (IPPROTO_DONE);
 656 }
 657 #endif /* INET6 */
 658
 659 static void
 660 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 661 {
 662         struct ifnet *ifp = m->m_pkthdr.rcvif;
 663         struct carp_softc *sc;
 664         u_int64_t tmp_counter;
 665         struct timeval sc_tv, ch_tv;
 666
 667         /* verify that the VHID is valid on the receiving interface */
 668         CARP_LOCK(ifp->if_carp);
 669         TAILQ_FOREACH(sc, &((struct carp_if *)ifp->if_carp)->vhif_vrs, sc_list)
 670                 if (sc->sc_vhid == ch->carp_vhid)
 671                         break;
 672
 673         if (!sc || !((SC2IFP(sc)->if_flags & IFF_UP) && (SC2IFP(sc)->if_flags & IFF_RUNNING))) {
 674                 carpstats.carps_badvhid++;
 675                 CARP_UNLOCK(ifp->if_carp);
 676                 m_freem(m);
 677                 return;
 678         }
 679
 680         getmicrotime(&SC2IFP(sc)->if_lastchange);
 681         SC2IFP(sc)->if_ipackets++;
 682         SC2IFP(sc)->if_ibytes += m->m_pkthdr.len;
 683
 684         if (SC2IFP(sc)->if_bpf) {
 685                 struct ip *ip = mtod(m, struct ip *);
 686
 687                 /* BPF wants net byte order */
 688                 ip->ip_len = htons(ip->ip_len + (ip->ip_hl << 2));
 689                 ip->ip_off = htons(ip->ip_off);
 690                 bpf_mtap(SC2IFP(sc)->if_bpf, m);
 691         }
 692
 693         /* verify the CARP version. */
 694         if (ch->carp_version != CARP_VERSION) {
 695                 carpstats.carps_badver++;
 696                 SC2IFP(sc)->if_ierrors++;
 697                 CARP_UNLOCK(ifp->if_carp);
 698                 CARP_LOG("%s; invalid version %d\n",
 699                     SC2IFP(sc)->if_xname,
 700                     ch->carp_version);
 701                 m_freem(m);
 702                 return;
 703         }
 704
 705         /* verify the hash */
 706         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 707                 carpstats.carps_badauth++;
 708                 SC2IFP(sc)->if_ierrors++;
 709                 CARP_UNLOCK(ifp->if_carp);
 710                 CARP_LOG("%s: incorrect hash\n", SC2IFP(sc)->if_xname);
 711                 m_freem(m);
 712                 return;
 713         }
 714
 715         tmp_counter = ntohl(ch->carp_counter[0]);
 716         tmp_counter = tmp_counter<<32;
 717         tmp_counter += ntohl(ch->carp_counter[1]);
 718
 719         /* XXX Replay protection goes here */
 720
 721         sc->sc_init_counter = 0;
 722         sc->sc_counter = tmp_counter;
 723
 724         sc_tv.tv_sec = sc->sc_advbase;
 725         if (carp_suppress_preempt && sc->sc_advskew <  240)
 726                 sc_tv.tv_usec = 240 * 1000000 / 256;
 727         else
 728                 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256;
 729         ch_tv.tv_sec = ch->carp_advbase;
 730         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 731
 732         switch (sc->sc_state) {
 733         case INIT:
 734                 break;
 735         case MASTER:
 736                 /*
 737                  * If we receive an advertisement from a master who's going to
 738                  * be more frequent than us, go into BACKUP state.
 739                  */
 740                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 741                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 742                         callout_stop(&sc->sc_ad_tmo);
 743                         CARP_DEBUG("%s: MASTER -> BACKUP "
 744                            "(more frequent advertisement received)\n",
 745                            SC2IFP(sc)->if_xname);
 746                         carp_set_state(sc, BACKUP);
 747                         carp_setrun(sc, 0);
 748                         carp_setroute(sc, RTM_DELETE);
 749                 }
 750                 break;
 751         case BACKUP:
 752                 /*
 753                  * If we're pre-empting masters who advertise slower than us,
 754                  * and this one claims to be slower, treat him as down.
 755                  */
 756                 if (carp_opts[CARPCTL_PREEMPT] &&
 757                     timevalcmp(&sc_tv, &ch_tv, <)) {
 758                         CARP_DEBUG("%s: BACKUP -> MASTER "
 759                             "(preempting a slower master)\n",
 760                             SC2IFP(sc)->if_xname);
 761                         carp_master_down_locked(sc);
 762                         break;
 763                 }
 764
 765                 /*
 766                  *  If the master is going to advertise at such a low frequency
 767                  *  that he's guaranteed to time out, we'd might as well just
 768                  *  treat him as timed out now.
 769                  */
 770                 sc_tv.tv_sec = sc->sc_advbase * 3;
 771                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 772                         CARP_DEBUG("%s: BACKUP -> MASTER "
 773                             "(master timed out)\n",
 774                             SC2IFP(sc)->if_xname);
 775                         carp_master_down_locked(sc);
 776                         break;
 777                 }
 778
 779                 /*
 780                  * Otherwise, we reset the counter and wait for the next
 781                  * advertisement.
 782                  */
 783                 carp_setrun(sc, af);
 784                 break;
 785         }
 786
 787         CARP_UNLOCK(ifp->if_carp);
 788
 789         m_freem(m);
 790         return;
 791 }
 792
 793 static int
 794 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 795 {
 796         struct m_tag *mtag;
 797         struct ifnet *ifp = SC2IFP(sc);
 798
 799         if (sc->sc_init_counter) {
 800                 /* this could also be seconds since unix epoch */
 801                 sc->sc_counter = karc4random();
 802                 sc->sc_counter = sc->sc_counter << 32;
 803                 sc->sc_counter += karc4random();
 804         } else
 805                 sc->sc_counter++;
 806
 807         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 808         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 809
 810         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 811
 812         /* Tag packet for carp_output */
 813         mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct ifnet *), MB_DONTWAIT);
 814         if (mtag == NULL) {
 815                 m_freem(m);
 816                 SC2IFP(sc)->if_oerrors++;
 817                 return (ENOMEM);
 818         }
 819         bcopy(&ifp, (caddr_t)(mtag + 1), sizeof(struct ifnet *));
 820         m_tag_prepend(m, mtag);
 821
 822         return (0);
 823 }
 824
 825 static void
 826 carp_send_ad_all(void)
 827 {
 828         struct carp_softc *sc;
 829
 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) && (SC2IFP(sc)->if_flags & IFF_RUNNING) &&
 835                      sc->sc_state == MASTER)
 836                         carp_send_ad_locked(sc);
 837                 CARP_SCUNLOCK(sc);
 838         }
 839 }
 840
 841 static void
 842 carp_send_ad(void *v)
 843 {
 844         struct carp_softc *sc = v;
 845
 846         CARP_SCLOCK(sc);
 847         carp_send_ad_locked(sc);
 848         CARP_SCUNLOCK(sc);
 849 }
 850
 851 static void
 852 carp_send_ad_locked(struct carp_softc *sc)
 853 {
 854         struct carp_header ch;
 855         struct timeval tv;
 856         struct carp_header *ch_ptr;
 857         struct mbuf *m;
 858         int len, advbase, advskew;
 859
 860
 861         /* bow out if we've lost our UPness or RUNNINGuiness */
 862         if (!((SC2IFP(sc)->if_flags & IFF_UP) && (SC2IFP(sc)->if_flags & IFF_RUNNING))) {
 863                 advbase = 255;
 864                 advskew = 255;
 865         } else {
 866                 advbase = sc->sc_advbase;
 867                 if (!carp_suppress_preempt || sc->sc_advskew > 240)
 868                         advskew = sc->sc_advskew;
 869                 else
 870                         advskew = 240;
 871                 tv.tv_sec = advbase;
 872                 tv.tv_usec = advskew * 1000000 / 256;
 873         }
 874
 875         ch.carp_version = CARP_VERSION;
 876         ch.carp_type = CARP_ADVERTISEMENT;
 877         ch.carp_vhid = sc->sc_vhid;
 878         ch.carp_advbase = advbase;
 879         ch.carp_advskew = advskew;
 880         ch.carp_authlen = 7;    /* XXX DEFINE */
 881         ch.carp_pad1 = 0;       /* must be zero */
 882         ch.carp_cksum = 0;
 883
 884 #ifdef INET
 885         if (sc->sc_ia) {
 886                 struct ip *ip;
 887
 888                 MGETHDR(m, M_NOWAIT, MT_HEADER);
 889                 if (m == NULL) {
 890                         SC2IFP(sc)->if_oerrors++;
 891                         carpstats.carps_onomem++;
 892                         /* XXX maybe less ? */
 893                         if (advbase != 255 || advskew != 255)
 894                                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
 895                                     carp_send_ad, sc);
 896                         return;
 897                 }
 898                 len = sizeof(*ip) + sizeof(ch);
 899                 m->m_pkthdr.len = len;
 900                 m->m_pkthdr.rcvif = NULL;
 901                 m->m_len = len;
 902                 MH_ALIGN(m, m->m_len);
 903                 m->m_flags |= M_MCAST;
 904                 ip = mtod(m, struct ip *);
 905                 ip->ip_v = IPVERSION;
 906                 ip->ip_hl = sizeof(*ip) >> 2;
 907                 ip->ip_tos = IPTOS_LOWDELAY;
 908                 ip->ip_len = len;
 909                 ip->ip_id = ip_newid();
 910                 ip->ip_off = IP_DF;
 911                 ip->ip_ttl = CARP_DFLTTL;
 912                 ip->ip_p = IPPROTO_CARP;
 913                 ip->ip_sum = 0;
 914                 ip->ip_src.s_addr = sc->sc_ia->ia_addr.sin_addr.s_addr;
 915                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 916
 917                 ch_ptr = (struct carp_header *)(&ip[1]);
 918                 bcopy(&ch, ch_ptr, sizeof(ch));
 919                 if (carp_prepare_ad(m, sc, ch_ptr))
 920                         return;
 921
 922                 m->m_data += sizeof(*ip);
 923                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip));
 924                 m->m_data -= sizeof(*ip);
 925
 926                 getmicrotime(&SC2IFP(sc)->if_lastchange);
 927                 SC2IFP(sc)->if_opackets++;
 928                 SC2IFP(sc)->if_obytes += len;
 929                 carpstats.carps_opackets++;
 930
 931                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, NULL)) {
 932                         SC2IFP(sc)->if_oerrors++;
 933                         if (sc->sc_sendad_errors < INT_MAX)
 934                                 sc->sc_sendad_errors++;
 935                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 936                                 carp_suppress_preempt++;
 937                                 if (carp_suppress_preempt == 1) {
 938                                         CARP_SCUNLOCK(sc);
 939                                         carp_send_ad_all();
 940                                         CARP_SCLOCK(sc);
 941                                 }
 942                         }
 943                         sc->sc_sendad_success = 0;
 944                 } else {
 945                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 946                                 if (++sc->sc_sendad_success >=
 947                                     CARP_SENDAD_MIN_SUCCESS) {
 948                                         carp_suppress_preempt--;
 949                                         sc->sc_sendad_errors = 0;
 950                                 }
 951                         } else
 952                                 sc->sc_sendad_errors = 0;
 953                 }
 954         }
 955 #endif /* INET */
 956 #ifdef INET6
 957         if (sc->sc_ia6) {
 958                 struct ip6_hdr *ip6;
 959
 960                 MGETHDR(m, M_NOWAIT, MT_HEADER);
 961                 if (m == NULL) {
 962                         SC2IFP(sc)->if_oerrors++;
 963                         carpstats.carps_onomem++;
 964                         /* XXX maybe less ? */
 965                         if (advbase != 255 || advskew != 255)
 966                                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
 967                                     carp_send_ad, sc);
 968                         return;
 969                 }
 970                 len = sizeof(*ip6) + sizeof(ch);
 971                 m->m_pkthdr.len = len;
 972                 m->m_pkthdr.rcvif = NULL;
 973                 m->m_len = len;
 974                 MH_ALIGN(m, m->m_len);
 975                 m->m_flags |= M_MCAST;
 976                 ip6 = mtod(m, struct ip6_hdr *);
 977                 bzero(ip6, sizeof(*ip6));
 978                 ip6->ip6_vfc |= IPV6_VERSION;
 979                 ip6->ip6_hlim = CARP_DFLTTL;
 980                 ip6->ip6_nxt = IPPROTO_CARP;
 981                 bcopy(&sc->sc_ia6->ia_addr.sin6_addr, &ip6->ip6_src,
 982                     sizeof(struct in6_addr));
 983                 /* set the multicast destination */
 984
 985                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
 986                 ip6->ip6_dst.s6_addr8[15] = 0x12;
 987                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
 988                         SC2IFP(sc)->if_oerrors++;
 989                         m_freem(m);
 990                         CARP_LOG("%s: in6_setscope failed\n", __func__);
 991                         return;
 992                 }
 993
 994                 ch_ptr = (struct carp_header *)(&ip6[1]);
 995                 bcopy(&ch, ch_ptr, sizeof(ch));
 996                 if (carp_prepare_ad(m, sc, ch_ptr))
 997                         return;
 998
 999                 m->m_data += sizeof(*ip6);
1000                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6));
1001                 m->m_data -= sizeof(*ip6);
1002
1003                 getmicrotime(&SC2IFP(sc)->if_lastchange);
1004                 SC2IFP(sc)->if_opackets++;
1005                 SC2IFP(sc)->if_obytes += len;
1006                 carpstats.carps_opackets6++;
1007
1008                 if (ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL)) {
1009                         SC2IFP(sc)->if_oerrors++;
1010                         if (sc->sc_sendad_errors < INT_MAX)
1011                                 sc->sc_sendad_errors++;
1012                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
1013                                 carp_suppress_preempt++;
1014                                 if (carp_suppress_preempt == 1) {
1015                                         CARP_SCUNLOCK(sc);
1016                                         carp_send_ad_all();
1017                                         CARP_SCLOCK(sc);
1018                                 }
1019                         }
1020                         sc->sc_sendad_success = 0;
1021                 } else {
1022                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
1023                                 if (++sc->sc_sendad_success >=
1024                                     CARP_SENDAD_MIN_SUCCESS) {
1025                                         carp_suppress_preempt--;
1026                                         sc->sc_sendad_errors = 0;
1027                                 }
1028                         } else
1029                                 sc->sc_sendad_errors = 0;
1030                 }
1031         }
1032 #endif /* INET6 */
1033
1034         if (advbase != 255 || advskew != 255)
1035                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
1036                     carp_send_ad, sc);
1037
1038 }
1039
1040 /*
1041  * Broadcast a gratuitous ARP request containing
1042  * the virtual router MAC address for each IP address
1043  * associated with the virtual router.
1044  */
1045 static void
1046 carp_send_arp(struct carp_softc *sc)
1047 {
1048         struct ifaddr_container *ifac;
1049
1050         TAILQ_FOREACH(ifac, &SC2IFP(sc)->if_addrheads[mycpuid], ifa_link) {
1051                 struct ifaddr *ifa = ifac->ifa;
1052
1053                 if (ifa->ifa_addr->sa_family != AF_INET)
1054                         continue;
1055                 arp_ifinit2(sc->sc_carpdev, ifa, IF_LLADDR(sc->sc_ifp));
1056
1057                 DELAY(1000);    /* XXX */
1058         }
1059 }
1060
1061 #ifdef INET6
1062 static void
1063 carp_send_na(struct carp_softc *sc)
1064 {
1065         struct ifaddr_container *ifac;
1066         struct in6_addr *in6;
1067         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1068
1069         TAILQ_FOREACH(ifac, &SC2IFP(sc)->if_addrheads[mycpuid], ifa_link) {
1070                 struct ifaddr *ifa = ifac->ifa;
1071
1072                 if (ifa->ifa_addr->sa_family != AF_INET6)
1073                         continue;
1074
1075                 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr;
1076                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1077                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1078                 DELAY(1000);    /* XXX */
1079         }
1080 }
1081 #endif /* INET6 */
1082
1083 static int
1084 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type)
1085 {
1086         struct carp_softc *vh;
1087         int count = 0;
1088
1089         CARP_LOCK_ASSERT(cif);
1090
1091         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1092                 if ((type == CARP_COUNT_RUNNING &&
1093                     (SC2IFP(vh)->if_flags & IFF_UP) && (SC2IFP(vh)->if_flags & IFF_RUNNING)) ||
1094                     (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) {
1095                         struct ifaddr_container *ifac;
1096
1097                         TAILQ_FOREACH(ifac, &SC2IFP(vh)->if_addrheads[mycpuid],
1098                                       ifa_link) {
1099                                 struct ifaddr *ifa = ifac->ifa;
1100
1101                                 if (ifa->ifa_addr->sa_family == AF_INET &&
1102                                     ia->ia_addr.sin_addr.s_addr ==
1103                                     ifatoia(ifa)->ia_addr.sin_addr.s_addr)
1104                                         count++;
1105                         }
1106                 }
1107         }
1108         return (count);
1109 }
1110
1111 int
1112 carp_iamatch(void *v, struct in_ifaddr *ia,
1113     struct in_addr *isaddr, u_int8_t **enaddr)
1114 {
1115         struct carp_if *cif = v;
1116         struct carp_softc *vh;
1117         int index, count = 0;
1118
1119         CARP_LOCK(cif);
1120
1121         if (carp_opts[CARPCTL_ARPBALANCE]) {
1122                 /*
1123                  * XXX proof of concept implementation.
1124                  * We use the source ip to decide which virtual host should
1125                  * handle the request. If we're master of that virtual host,
1126                  * then we respond, otherwise, just drop the arp packet on
1127                  * the floor.
1128                  */
1129                 count = carp_addrcount(cif, ia, CARP_COUNT_RUNNING);
1130                 if (count == 0) {
1131                         /* should never reach this */
1132                         CARP_UNLOCK(cif);
1133                         return (0);
1134                 }
1135
1136                 /* this should be a hash, like pf_hash() */
1137                 index = ntohl(isaddr->s_addr) % count;
1138                 count = 0;
1139
1140                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1141                         if ((SC2IFP(vh)->if_flags & IFF_UP) && (SC2IFP(vh)->if_flags & IFF_RUNNING)) {
1142                                 struct ifaddr_container *ifac;
1143
1144                                 TAILQ_FOREACH(ifac, &SC2IFP(vh)->if_addrheads[mycpuid], ifa_link) {
1145                                         struct ifaddr *ifa = ifac->ifa;
1146
1147                                         if (ifa->ifa_addr->sa_family ==
1148                                             AF_INET &&
1149                                             ia->ia_addr.sin_addr.s_addr ==
1150                                             ifatoia(ifa)->ia_addr.sin_addr.s_addr) {
1151                                                 if (count == index) {
1152                                                         if (vh->sc_state ==
1153                                                             MASTER) {
1154                                                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1155                                                                 CARP_UNLOCK(cif);
1156                                                                 return (1);
1157                                                         } else {
1158                                                                 CARP_UNLOCK(cif);
1159                                                                 return (0);
1160                                                         }
1161                                                 }
1162                                                 count++;
1163                                         }
1164                                 }
1165                         }
1166                 }
1167         } else {
1168                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1169                         if ((SC2IFP(vh)->if_flags & IFF_UP) && (SC2IFP(vh)->if_flags & IFF_RUNNING) &&
1170                             vh->sc_state == MASTER) {
1171                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1172                                 CARP_UNLOCK(cif);
1173                                 return (1);
1174                         }
1175                 }
1176         }
1177         CARP_UNLOCK(cif);
1178         return(0);
1179 }
1180
1181 #ifdef INET6
1182 struct ifaddr *
1183 carp_iamatch6(void *v, struct in6_addr *taddr)
1184 {
1185         struct carp_if *cif = v;
1186         struct carp_softc *vh;
1187
1188         CARP_LOCK(cif);
1189         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1190                 struct ifaddr_container *ifac;
1191
1192                 TAILQ_FOREACH(ifac, &SC2IFP(vh)->if_addrheads[mycpuid], ifa_link) {
1193                         struct ifaddr *ifa = ifac->ifa;
1194
1195                         if (IN6_ARE_ADDR_EQUAL(taddr,
1196                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1197                             (SC2IFP(vh)->if_flags & IFF_UP) && (SC2IFP(vh)->if_flags & IFF_RUNNING) &&
1198                             vh->sc_state == MASTER) {
1199                                 CARP_UNLOCK(cif);
1200                                 return (ifa);
1201                         }
1202                 }
1203         }
1204         CARP_UNLOCK(cif);
1205
1206         return (NULL);
1207 }
1208
1209 void *
1210 carp_macmatch6(void *v, struct mbuf *m, const struct in6_addr *taddr)
1211 {
1212         struct m_tag *mtag;
1213         struct carp_if *cif = v;
1214         struct carp_softc *sc;
1215
1216         CARP_LOCK(cif);
1217         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) {
1218                 struct ifaddr_container *ifac;
1219
1220                 TAILQ_FOREACH(ifac, &SC2IFP(sc)->if_addrheads[mycpuid], ifa_link) {
1221                         struct ifaddr *ifa = ifac->ifa;
1222
1223                         if (IN6_ARE_ADDR_EQUAL(taddr,
1224                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1225                             (SC2IFP(sc)->if_flags & IFF_UP) && (SC2IFP(sc)->if_flags & IFF_RUNNING)) {
1226                                 struct ifnet *ifp = SC2IFP(sc);
1227                                 mtag = m_tag_get(PACKET_TAG_CARP,
1228                                     sizeof(struct ifnet *), MB_DONTWAIT);
1229                                 if (mtag == NULL) {
1230                                         /* better a bit than nothing */
1231                                         CARP_UNLOCK(cif);
1232                                         return (IF_LLADDR(sc->sc_ifp));
1233                                 }
1234                                 bcopy(&ifp, (caddr_t)(mtag + 1),
1235                                     sizeof(struct ifnet *));
1236                                 m_tag_prepend(m, mtag);
1237
1238                                 CARP_UNLOCK(cif);
1239                                 return (IF_LLADDR(sc->sc_ifp));
1240                         }
1241                 }
1242         }
1243         CARP_UNLOCK(cif);
1244
1245         return (NULL);
1246 }
1247 #endif
1248
1249 struct ifnet *
1250 carp_forus(void *v, void *dhost)
1251 {
1252         struct carp_if *cif = v;
1253         struct carp_softc *vh;
1254         u_int8_t *ena = dhost;
1255
1256         /**
1257          * XXX: See here for check on MAC adr is not for virtual use
1258          *
1259          **/
1260
1261         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1262         {
1263                 return (NULL);
1264         }
1265
1266         CARP_LOCK(cif);
1267         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list)
1268                 if ((SC2IFP(vh)->if_flags & IFF_UP) && (SC2IFP(vh)->if_flags & IFF_RUNNING) &&
1269                     vh->sc_state == MASTER &&
1270                     !bcmp(dhost, IF_LLADDR(vh->sc_ifp), ETHER_ADDR_LEN)) {
1271                         CARP_UNLOCK(cif);
1272                         return (SC2IFP(vh));
1273                 }
1274
1275         CARP_UNLOCK(cif);
1276         return (NULL);
1277 }
1278
1279 static void
1280 carp_master_down(void *v)
1281 {
1282         struct carp_softc *sc = v;
1283
1284         lwkt_serialize_enter(sc->sc_ifp->if_serializer);
1285         carp_master_down_locked(sc);
1286         lwkt_serialize_exit(sc->sc_ifp->if_serializer);
1287 }
1288
1289 static void
1290 carp_master_down_locked(struct carp_softc *sc)
1291 {
1292         if (sc->sc_carpdev)
1293                 CARP_SCLOCK_ASSERT(sc);
1294
1295         switch (sc->sc_state) {
1296         case INIT:
1297                 kprintf("%s: master_down event in INIT state\n",
1298                     SC2IFP(sc)->if_xname);
1299                 break;
1300         case MASTER:
1301                 break;
1302         case BACKUP:
1303                 carp_set_state(sc, MASTER);
1304                 carp_send_ad_locked(sc);
1305                 carp_send_arp(sc);
1306 #ifdef INET6
1307                 carp_send_na(sc);
1308 #endif /* INET6 */
1309                 carp_setrun(sc, 0);
1310                 carp_setroute(sc, RTM_ADD);
1311                 break;
1312         }
1313 }
1314
1315 /*
1316  * When in backup state, af indicates whether to reset the master down timer
1317  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1318  */
1319 static void
1320 carp_setrun(struct carp_softc *sc, sa_family_t af)
1321 {
1322         struct timeval tv;
1323
1324         if (sc->sc_carpdev == NULL) {
1325                 SC2IFP(sc)->if_flags &= ~IFF_RUNNING;
1326                 carp_set_state(sc, INIT);
1327                 return;
1328         }
1329
1330         if (SC2IFP(sc)->if_flags & IFF_UP &&
1331             sc->sc_vhid > 0 && (sc->sc_naddrs || sc->sc_naddrs6))
1332                 SC2IFP(sc)->if_flags |= IFF_RUNNING;
1333         else {
1334                 SC2IFP(sc)->if_flags &= ~IFF_RUNNING;
1335                 carp_setroute(sc, RTM_DELETE);
1336                 return;
1337         }
1338
1339         switch (sc->sc_state) {
1340         case INIT:
1341                 if (carp_opts[CARPCTL_PREEMPT] && !carp_suppress_preempt) {
1342                         carp_send_ad_locked(sc);
1343                         carp_send_arp(sc);
1344 #ifdef INET6
1345                         carp_send_na(sc);
1346 #endif /* INET6 */
1347                         CARP_DEBUG("%s: INIT -> MASTER (preempting)\n",
1348                             SC2IFP(sc)->if_xname);
1349                         carp_set_state(sc, MASTER);
1350                         carp_setroute(sc, RTM_ADD);
1351                 } else {
1352                         CARP_DEBUG("%s: INIT -> BACKUP\n", SC2IFP(sc)->if_xname);
1353                         carp_set_state(sc, BACKUP);
1354                         carp_setroute(sc, RTM_DELETE);
1355                         carp_setrun(sc, 0);
1356                 }
1357                 break;
1358         case BACKUP:
1359                 callout_stop(&sc->sc_ad_tmo);
1360                 tv.tv_sec = 3 * sc->sc_advbase;
1361                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1362                 switch (af) {
1363 #ifdef INET
1364                 case AF_INET:
1365                         callout_reset(&sc->sc_md_tmo, tvtohz_high(&tv),
1366                             carp_master_down, sc);
1367                         break;
1368 #endif /* INET */
1369 #ifdef INET6
1370                 case AF_INET6:
1371                         callout_reset(&sc->sc_md6_tmo, tvtohz_high(&tv),
1372                             carp_master_down, sc);
1373                         break;
1374 #endif /* INET6 */
1375                 default:
1376                         if (sc->sc_naddrs)
1377                                 callout_reset(&sc->sc_md_tmo, tvtohz_high(&tv),
1378                                     carp_master_down, sc);
1379                         if (sc->sc_naddrs6)
1380                                 callout_reset(&sc->sc_md6_tmo, tvtohz_high(&tv),
1381                                     carp_master_down, sc);
1382                         break;
1383                 }
1384                 break;
1385         case MASTER:
1386                 tv.tv_sec = sc->sc_advbase;
1387                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1388                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
1389                     carp_send_ad, sc);
1390                 break;
1391         }
1392 }
1393
1394 static void
1395 carp_multicast_cleanup(struct carp_softc *sc)
1396 {
1397         struct ip_moptions *imo = &sc->sc_imo;
1398         u_int16_t n = imo->imo_num_memberships;
1399
1400         /* Clean up our own multicast memberships */
1401         while (n-- > 0) {
1402                 if (imo->imo_membership[n] != NULL) {
1403                         in_delmulti(imo->imo_membership[n]);
1404                         imo->imo_membership[n] = NULL;
1405                 }
1406         }
1407         imo->imo_num_memberships = 0;
1408         imo->imo_multicast_ifp = NULL;
1409 }
1410
1411 #ifdef INET6
1412 static void
1413 carp_multicast6_cleanup(struct carp_softc *sc)
1414 {
1415         struct ip6_moptions *im6o = &sc->sc_im6o;
1416
1417         while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1418                 struct in6_multi_mship *imm =
1419                     LIST_FIRST(&im6o->im6o_memberships);
1420
1421                 LIST_REMOVE(imm, i6mm_chain);
1422                 in6_leavegroup(imm);
1423         }
1424         im6o->im6o_multicast_ifp = NULL;
1425 }
1426 #endif
1427
1428 static int
1429 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1430 {
1431         struct ifnet *ifp;
1432         struct carp_if *cif;
1433         struct in_ifaddr *ia, *ia_if;
1434         struct in_ifaddr_container *iac;
1435         struct ip_moptions *imo = &sc->sc_imo;
1436         struct in_addr addr;
1437         u_long iaddr = htonl(sin->sin_addr.s_addr);
1438         int own, error;
1439
1440         if (sin->sin_addr.s_addr == 0)
1441         {
1442                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1443                 {
1444                         carp_set_state(sc, INIT);
1445                 }
1446                 if (sc->sc_naddrs)
1447                 {
1448                         SC2IFP(sc)->if_flags |= IFF_UP;
1449                 }
1450                 carp_setrun(sc, 0);
1451                 return (0);
1452         }
1453         /* we have to do it by hands to check we won't match on us */
1454         ia_if = NULL; own = 0;
1455         TAILQ_FOREACH(iac, &in_ifaddrheads[mycpuid], ia_link) {
1456                 ia = iac->ia;
1457
1458                 /* and, yeah, we need a multicast-capable iface too */
1459                 if (ia->ia_ifp != SC2IFP(sc) &&
1460                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1461                     (iaddr & ia->ia_subnetmask) == ia->ia_subnet) {
1462                         if (!ia_if)
1463                                 ia_if = ia;
1464                         if (sin->sin_addr.s_addr ==
1465                             ia->ia_addr.sin_addr.s_addr)
1466                                 own++;
1467                 }
1468         }
1469
1470
1471         if (!ia_if)
1472                 return (EADDRNOTAVAIL);
1473
1474         ia = ia_if;
1475         ifp = ia->ia_ifp;
1476
1477         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1478             (imo->imo_multicast_ifp && imo->imo_multicast_ifp != ifp))
1479                 return (EADDRNOTAVAIL);
1480
1481         if (imo->imo_num_memberships == 0) {
1482                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1483                 if ((imo->imo_membership[0] = in_addmulti(&addr, ifp)) == NULL)
1484                         return (ENOBUFS);
1485                 imo->imo_num_memberships++;
1486                 imo->imo_multicast_ifp = ifp;
1487                 imo->imo_multicast_ttl = CARP_DFLTTL;
1488                 imo->imo_multicast_loop = 0;
1489         }
1490
1491         if (!ifp->if_carp) {
1492
1493                 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP,
1494                     M_WAITOK|M_ZERO);
1495                 if ((error = ifpromisc(ifp, 1))) {
1496                         FREE(cif, M_CARP);
1497                         goto cleanup;
1498                 }
1499
1500                 CARP_LOCK_INIT(cif);
1501                 CARP_LOCK(cif);
1502                 cif->vhif_ifp = ifp;
1503                 TAILQ_INIT(&cif->vhif_vrs);
1504                 ifp->if_carp = cif;
1505
1506         } else {
1507                 struct carp_softc *vr;
1508
1509                 cif = (struct carp_if *)ifp->if_carp;
1510                 CARP_LOCK(cif);
1511                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1512                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1513                                 CARP_UNLOCK(cif);
1514                                 error = EINVAL;
1515                                 goto cleanup;
1516                         }
1517         }
1518         sc->sc_ia = ia;
1519         sc->sc_carpdev = ifp;
1520
1521         { /* XXX prevent endless loop if already in queue */
1522         struct carp_softc *vr, *after = NULL;
1523         int myself = 0;
1524         cif = (struct carp_if *)ifp->if_carp;
1525
1526         /* XXX: cif should not change, right? So we still hold the lock */
1527         CARP_LOCK_ASSERT(cif);
1528
1529         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1530                 if (vr == sc)
1531                         myself = 1;
1532                 if (vr->sc_vhid < sc->sc_vhid)
1533                         after = vr;
1534         }
1535
1536         if (!myself) {
1537                 /* We're trying to keep things in order */
1538                 if (after == NULL) {
1539                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1540                 } else {
1541                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1542                 }
1543                 cif->vhif_nvrs++;
1544         }
1545         }
1546
1547         sc->sc_naddrs++;
1548         SC2IFP(sc)->if_flags |= IFF_UP;
1549         if (own)
1550                 sc->sc_advskew = 0;
1551
1552
1553         carp_sc_state_locked(sc);
1554         carp_setrun(sc, 0);
1555
1556         CARP_UNLOCK(cif);
1557
1558         return (0);
1559
1560 cleanup:
1561         in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1562         return (error);
1563
1564 }
1565
1566 static int
1567 carp_del_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1568 {
1569         int error = 0;
1570
1571         if (!--sc->sc_naddrs) {
1572                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1573                 struct ip_moptions *imo = &sc->sc_imo;
1574
1575                 CARP_LOCK(cif);
1576                 callout_stop(&sc->sc_ad_tmo);
1577                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1578                 SC2IFP(sc)->if_flags &= ~IFF_RUNNING;
1579                 sc->sc_vhid = -1;
1580                 in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1581                 imo->imo_multicast_ifp = NULL;
1582                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1583                 if (!--cif->vhif_nvrs) {
1584                         sc->sc_carpdev->if_carp = NULL;
1585                         CARP_LOCK_DESTROY(cif);
1586                         FREE(cif, M_IFADDR);
1587                 } else {
1588                         CARP_UNLOCK(cif);
1589                 }
1590         }
1591
1592         return (error);
1593 }
1594
1595 #ifdef INET6
1596 static int
1597 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1598 {
1599         struct ifnet *ifp;
1600         struct carp_if *cif;
1601         struct in6_ifaddr *ia, *ia_if;
1602         struct ip6_moptions *im6o = &sc->sc_im6o;
1603         struct in6_multi_mship *imm;
1604         struct in6_addr in6;
1605         int own, error;
1606
1607         if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
1608                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1609                         carp_set_state(sc, INIT);
1610                 if (sc->sc_naddrs6)
1611                         SC2IFP(sc)->if_flags |= IFF_UP;
1612                 carp_setrun(sc, 0);
1613                 return (0);
1614         }
1615
1616         /* we have to do it by hands to check we won't match on us */
1617         ia_if = NULL; own = 0;
1618         for (ia = in6_ifaddr; ia; ia = ia->ia_next) {
1619                 int i;
1620
1621                 for (i = 0; i < 4; i++) {
1622                         if ((sin6->sin6_addr.s6_addr32[i] &
1623                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]) !=
1624                             (ia->ia_addr.sin6_addr.s6_addr32[i] &
1625                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]))
1626                                 break;
1627                 }
1628                 /* and, yeah, we need a multicast-capable iface too */
1629                 if (ia->ia_ifp != SC2IFP(sc) &&
1630                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1631                     (i == 4)) {
1632                         if (!ia_if)
1633                                 ia_if = ia;
1634                         if (IN6_ARE_ADDR_EQUAL(&sin6->sin6_addr,
1635                             &ia->ia_addr.sin6_addr))
1636                                 own++;
1637                 }
1638         }
1639
1640         if (!ia_if)
1641                 return (EADDRNOTAVAIL);
1642         ia = ia_if;
1643         ifp = ia->ia_ifp;
1644
1645         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1646             (im6o->im6o_multicast_ifp && im6o->im6o_multicast_ifp != ifp))
1647                 return (EADDRNOTAVAIL);
1648
1649         if (!sc->sc_naddrs6) {
1650                 im6o->im6o_multicast_ifp = ifp;
1651
1652                 /* join CARP multicast address */
1653                 bzero(&in6, sizeof(in6));
1654                 in6.s6_addr16[0] = htons(0xff02);
1655                 in6.s6_addr8[15] = 0x12;
1656                 if (in6_setscope(&in6, ifp, NULL) != 0)
1657                         goto cleanup;
1658                 if ((imm = in6_joingroup(ifp, &in6, &error)) == NULL)
1659                         goto cleanup;
1660                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1661
1662                 /* join solicited multicast address */
1663                 bzero(&in6, sizeof(in6));
1664                 in6.s6_addr16[0] = htons(0xff02);
1665                 in6.s6_addr32[1] = 0;
1666                 in6.s6_addr32[2] = htonl(1);
1667                 in6.s6_addr32[3] = sin6->sin6_addr.s6_addr32[3];
1668                 in6.s6_addr8[12] = 0xff;
1669                 if (in6_setscope(&in6, ifp, NULL) != 0)
1670                         goto cleanup;
1671                 if ((imm = in6_joingroup(ifp, &in6, &error)) == NULL)
1672                         goto cleanup;
1673                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1674         }
1675
1676         if (!ifp->if_carp) {
1677                 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP,
1678                     M_WAITOK|M_ZERO);
1679                 if ((error = ifpromisc(ifp, 1))) {
1680                         FREE(cif, M_CARP);
1681                         goto cleanup;
1682                 }
1683
1684                 CARP_LOCK_INIT(cif);
1685                 CARP_LOCK(cif);
1686                 cif->vhif_ifp = ifp;
1687                 TAILQ_INIT(&cif->vhif_vrs);
1688                 ifp->if_carp = cif;
1689
1690         } else {
1691                 struct carp_softc *vr;
1692
1693                 cif = (struct carp_if *)ifp->if_carp;
1694                 CARP_LOCK(cif);
1695                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1696                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1697                                 CARP_UNLOCK(cif);
1698                                 error = EINVAL;
1699                                 goto cleanup;
1700                         }
1701         }
1702         sc->sc_ia6 = ia;
1703         sc->sc_carpdev = ifp;
1704
1705         { /* XXX prevent endless loop if already in queue */
1706         struct carp_softc *vr, *after = NULL;
1707         int myself = 0;
1708         cif = (struct carp_if *)ifp->if_carp;
1709         CARP_LOCK_ASSERT(cif);
1710
1711         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1712                 if (vr == sc)
1713                         myself = 1;
1714                 if (vr->sc_vhid < sc->sc_vhid)
1715                         after = vr;
1716         }
1717
1718         if (!myself) {
1719                 /* We're trying to keep things in order */
1720                 if (after == NULL) {
1721                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1722                 } else {
1723                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1724                 }
1725                 cif->vhif_nvrs++;
1726         }
1727         }
1728
1729         sc->sc_naddrs6++;
1730         SC2IFP(sc)->if_flags |= IFF_UP;
1731         if (own)
1732                 sc->sc_advskew = 0;
1733         carp_sc_state_locked(sc);
1734         carp_setrun(sc, 0);
1735
1736         CARP_UNLOCK(cif);
1737
1738         return (0);
1739
1740 cleanup:
1741         /* clean up multicast memberships */
1742         if (!sc->sc_naddrs6) {
1743                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1744                         imm = LIST_FIRST(&im6o->im6o_memberships);
1745                         LIST_REMOVE(imm, i6mm_chain);
1746                         in6_leavegroup(imm);
1747                 }
1748         }
1749         return (error);
1750 }
1751
1752 static int
1753 carp_del_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1754 {
1755         int error = 0;
1756
1757         if (!--sc->sc_naddrs6) {
1758                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1759                 struct ip6_moptions *im6o = &sc->sc_im6o;
1760
1761                 CARP_LOCK(cif);
1762                 callout_stop(&sc->sc_ad_tmo);
1763                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1764                 SC2IFP(sc)->if_flags &= ~IFF_RUNNING;
1765                 sc->sc_vhid = -1;
1766                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1767                         struct in6_multi_mship *imm =
1768                             LIST_FIRST(&im6o->im6o_memberships);
1769
1770                         LIST_REMOVE(imm, i6mm_chain);
1771                         in6_leavegroup(imm);
1772                 }
1773                 im6o->im6o_multicast_ifp = NULL;
1774                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1775                 if (!--cif->vhif_nvrs) {
1776                         CARP_LOCK_DESTROY(cif);
1777                         sc->sc_carpdev->if_carp = NULL;
1778                         FREE(cif, M_IFADDR);
1779                 } else
1780                         CARP_UNLOCK(cif);
1781         }
1782
1783         return (error);
1784 }
1785 #endif /* INET6 */
1786
1787 static int
1788 carp_ioctl(struct ifnet *ifp, u_long cmd, caddr_t addr, struct ucred *creds)
1789 {
1790         struct carp_softc *sc = ifp->if_softc, *vr;
1791         struct carpreq carpr;
1792         struct ifaddr *ifa;
1793         struct ifreq *ifr;
1794         struct ifaliasreq *ifra;
1795         int locked = 0, error = 0;
1796
1797         ifa = (struct ifaddr *)addr;
1798         ifra = (struct ifaliasreq *)addr;
1799         ifr = (struct ifreq *)addr;
1800
1801
1802         switch (cmd) {
1803         case SIOCSIFADDR:
1804                 switch (ifa->ifa_addr->sa_family) {
1805 #ifdef INET
1806                 case AF_INET:
1807                         SC2IFP(sc)->if_flags |= IFF_UP;
1808                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1809                             sizeof(struct sockaddr));
1810                         error = carp_set_addr(sc, satosin(ifa->ifa_addr));
1811                         break;
1812 #endif /* INET */
1813 #ifdef INET6
1814                 case AF_INET6:
1815                         SC2IFP(sc)->if_flags |= IFF_UP;
1816                         error = carp_set_addr6(sc, satosin6(ifa->ifa_addr));
1817                         break;
1818 #endif /* INET6 */
1819                 default:
1820                         error = EAFNOSUPPORT;
1821                         break;
1822                 }
1823                 break;
1824
1825         case SIOCAIFADDR:
1826                 switch (ifa->ifa_addr->sa_family) {
1827 #ifdef INET
1828                 case AF_INET:
1829                         SC2IFP(sc)->if_flags |= IFF_UP;
1830                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1831                             sizeof(struct sockaddr));
1832                         error = carp_set_addr(sc, satosin(&ifra->ifra_addr));
1833                         break;
1834 #endif /* INET */
1835 #ifdef INET6
1836                 case AF_INET6:
1837                         SC2IFP(sc)->if_flags |= IFF_UP;
1838                         error = carp_set_addr6(sc, satosin6(&ifra->ifra_addr));
1839                         break;
1840 #endif /* INET6 */
1841                 default:
1842                         error = EAFNOSUPPORT;
1843                         break;
1844                 }
1845                 break;
1846
1847         case SIOCDIFADDR:
1848                 switch (ifa->ifa_addr->sa_family) {
1849 #ifdef INET
1850                 case AF_INET:
1851                         error = carp_del_addr(sc, satosin(&ifra->ifra_addr));
1852                         break;
1853 #endif /* INET */
1854 #ifdef INET6
1855                 case AF_INET6:
1856                         error = carp_del_addr6(sc, satosin6(&ifra->ifra_addr));
1857                         break;
1858 #endif /* INET6 */
1859                 default:
1860                         error = EAFNOSUPPORT;
1861                         break;
1862                 }
1863                 break;
1864
1865         case SIOCSIFFLAGS:
1866                 if (sc->sc_carpdev) {
1867                         locked = 1;
1868                         CARP_SCLOCK(sc);
1869                 }
1870                 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) {
1871                         callout_stop(&sc->sc_ad_tmo);
1872                         callout_stop(&sc->sc_md_tmo);
1873                         callout_stop(&sc->sc_md6_tmo);
1874                         if (sc->sc_state == MASTER)
1875                                 carp_send_ad_locked(sc);
1876                         carp_set_state(sc, INIT);
1877                         carp_setrun(sc, 0);
1878                 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) {
1879                         SC2IFP(sc)->if_flags |= IFF_UP;
1880                         carp_setrun(sc, 0);
1881                 }
1882                 break;
1883
1884         case SIOCSVH:
1885                 error = suser(curthread);
1886                 if (error)
1887                         break;
1888                 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1889                         break;
1890                 error = 1;
1891                 if (sc->sc_carpdev) {
1892                         locked = 1;
1893                         CARP_SCLOCK(sc);
1894                 }
1895                 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) {
1896                         switch (carpr.carpr_state) {
1897                         case BACKUP:
1898                                 callout_stop(&sc->sc_ad_tmo);
1899                                 carp_set_state(sc, BACKUP);
1900                                 carp_setrun(sc, 0);
1901                                 carp_setroute(sc, RTM_DELETE);
1902                                 break;
1903                         case MASTER:
1904                                 carp_master_down_locked(sc);
1905                                 break;
1906                         default:
1907                                 break;
1908                         }
1909                 }
1910                 if (carpr.carpr_vhid > 0) {
1911                         if (carpr.carpr_vhid > 255) {
1912                                 error = EINVAL;
1913                                 break;
1914                         }
1915                         if (sc->sc_carpdev) {
1916                                 struct carp_if *cif;
1917                                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1918                                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1919                                         if (vr != sc &&
1920                                             vr->sc_vhid == carpr.carpr_vhid)
1921                                                 return EEXIST;
1922                         }
1923                         sc->sc_vhid = carpr.carpr_vhid;
1924                         IF_LLADDR(sc->sc_ifp)[0] = 0;
1925                         IF_LLADDR(sc->sc_ifp)[1] = 0;
1926                         IF_LLADDR(sc->sc_ifp)[2] = 0x5e;
1927                         IF_LLADDR(sc->sc_ifp)[3] = 0;
1928                         IF_LLADDR(sc->sc_ifp)[4] = 1;
1929                         IF_LLADDR(sc->sc_ifp)[5] = sc->sc_vhid;
1930                         error--;
1931                 }
1932                 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) {
1933                         if (carpr.carpr_advskew >= 255) {
1934                                 error = EINVAL;
1935                                 break;
1936                         }
1937                         if (carpr.carpr_advbase > 255) {
1938                                 error = EINVAL;
1939                                 break;
1940                         }
1941                         sc->sc_advbase = carpr.carpr_advbase;
1942                         sc->sc_advskew = carpr.carpr_advskew;
1943                         error--;
1944                 }
1945                 bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1946                 if (error > 0)
1947                         error = EINVAL;
1948                 else {
1949                         error = 0;
1950                         carp_setrun(sc, 0);
1951                 }
1952                 break;
1953
1954         case SIOCGVH:
1955                 /* XXX: lockless read */
1956                 bzero(&carpr, sizeof(carpr));
1957                 carpr.carpr_state = sc->sc_state;
1958                 carpr.carpr_vhid = sc->sc_vhid;
1959                 carpr.carpr_advbase = sc->sc_advbase;
1960                 carpr.carpr_advskew = sc->sc_advskew;
1961                 error = suser(curthread);
1962                 if (error == 0)
1963                         bcopy(sc->sc_key, carpr.carpr_key,
1964                             sizeof(carpr.carpr_key));
1965                 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1966                 break;
1967
1968         default:
1969                 error = EINVAL;
1970         }
1971
1972         if (locked)
1973                 CARP_SCUNLOCK(sc);
1974
1975         carp_hmac_prepare(sc);
1976
1977         return (error);
1978 }
1979
1980 /*
1981  * XXX: this is looutput. We should eventually use it from there.
1982  */
1983 static int
1984 carp_looutput(struct ifnet *ifp, struct mbuf *m, struct sockaddr *dst,
1985     struct rtentry *rt)
1986 {
1987         u_int32_t af;
1988
1989         M_ASSERTPKTHDR(m); /* check if we have the packet header */
1990
1991         if (rt && rt->rt_flags & (RTF_REJECT|RTF_BLACKHOLE)) {
1992                 m_freem(m);
1993                 return (rt->rt_flags & RTF_BLACKHOLE ? 0 :
1994                         rt->rt_flags & RTF_HOST ? EHOSTUNREACH : ENETUNREACH);
1995         }
1996
1997         ifp->if_opackets++;
1998         ifp->if_obytes += m->m_pkthdr.len;
1999
2000         /* BPF writes need to be handled specially. */
2001         if (dst->sa_family == AF_UNSPEC) {
2002                 bcopy(dst->sa_data, &af, sizeof(af));
2003                 dst->sa_family = af;
2004         }
2005
2006 #if 1   /* XXX */
2007         switch (dst->sa_family) {
2008         case AF_INET:
2009         case AF_INET6:
2010         case AF_IPX:
2011         case AF_APPLETALK:
2012                 break;
2013         default:
2014                 m_freem(m);
2015                 return (EAFNOSUPPORT);
2016         }
2017 #endif
2018         return(if_simloop(ifp, m, dst->sa_family, 0));
2019 }
2020
2021 /*
2022  * Start output on carp interface. This function should never be called.
2023  */
2024 static void
2025 carp_start(struct ifnet *ifp)
2026 {
2027 #ifdef DEBUG
2028         kprintf("%s: start called\n", ifp->if_xname);
2029 #endif
2030 }
2031
2032 int
2033 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *sa,
2034     struct rtentry *rt)
2035 {
2036         struct m_tag *mtag;
2037         struct carp_softc *sc;
2038         struct ifnet *carp_ifp;
2039
2040         if (!sa)
2041                 return (0);
2042
2043         switch (sa->sa_family) {
2044 #ifdef INET
2045         case AF_INET:
2046                 break;
2047 #endif /* INET */
2048 #ifdef INET6
2049         case AF_INET6:
2050                 break;
2051 #endif /* INET6 */
2052         default:
2053                 return (0);
2054         }
2055
2056         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
2057         if (mtag == NULL)
2058                 return (0);
2059
2060         bcopy(mtag + 1, &carp_ifp, sizeof(struct ifnet *));
2061         sc = carp_ifp->if_softc;
2062
2063         /* Set the source MAC address to Virtual Router MAC Address */
2064         switch (ifp->if_type) {
2065         case IFT_ETHER:
2066         case IFT_L2VLAN: {
2067                         struct ether_header *eh;
2068
2069                         eh = mtod(m, struct ether_header *);
2070                         eh->ether_shost[0] = 0;
2071                         eh->ether_shost[1] = 0;
2072                         eh->ether_shost[2] = 0x5e;
2073                         eh->ether_shost[3] = 0;
2074                         eh->ether_shost[4] = 1;
2075                         eh->ether_shost[5] = sc->sc_vhid;
2076                 }
2077                 break;
2078         default:
2079                 kprintf("%s: carp is not supported for this interface type\n",
2080                     ifp->if_xname);
2081                 return (EOPNOTSUPP);
2082         }
2083
2084         return (0);
2085
2086 }
2087
2088 static void
2089 carp_set_state(struct carp_softc *sc, int state)
2090 {
2091
2092         if (sc->sc_carpdev)
2093                 CARP_SCLOCK_ASSERT(sc);
2094
2095         if (sc->sc_state == state)
2096                 return;
2097
2098         sc->sc_state = state;
2099         switch (state) {
2100         case BACKUP:
2101                 SC2IFP(sc)->if_link_state = LINK_STATE_DOWN;
2102                 break;
2103         case MASTER:
2104                 SC2IFP(sc)->if_link_state = LINK_STATE_UP;
2105                 break;
2106         default:
2107                 SC2IFP(sc)->if_link_state = LINK_STATE_UNKNOWN;
2108                 break;
2109         }
2110         rt_ifmsg(SC2IFP(sc));
2111 }
2112
2113 void
2114 carp_carpdev_state(void *v)
2115 {
2116         struct carp_if *cif = v;
2117
2118         CARP_LOCK(cif);
2119         carp_carpdev_state_locked(cif);
2120         CARP_UNLOCK(cif);
2121 }
2122
2123 static void
2124 carp_carpdev_state_locked(struct carp_if *cif)
2125 {
2126         struct carp_softc *sc;
2127
2128         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list)
2129                 carp_sc_state_locked(sc);
2130 }
2131
2132 static void
2133 carp_sc_state_locked(struct carp_softc *sc)
2134 {
2135         CARP_SCLOCK_ASSERT(sc);
2136
2137         if ( !(sc->sc_carpdev->if_flags & IFF_UP)) {
2138                 sc->sc_flags_backup = SC2IFP(sc)->if_flags;
2139                 SC2IFP(sc)->if_flags &= ~IFF_UP;
2140                 SC2IFP(sc)->if_flags &= ~IFF_RUNNING;
2141                 callout_stop(&sc->sc_ad_tmo);
2142                 callout_stop(&sc->sc_md_tmo);
2143                 callout_stop(&sc->sc_md6_tmo);
2144                 carp_set_state(sc, INIT);
2145                 carp_setrun(sc, 0);
2146                 if (!sc->sc_suppress) {
2147                         carp_suppress_preempt++;
2148                         if (carp_suppress_preempt == 1) {
2149                                 CARP_SCUNLOCK(sc);
2150                                 carp_send_ad_all();
2151                                 CARP_SCLOCK(sc);
2152                         }
2153                 }
2154                 sc->sc_suppress = 1;
2155         } else {
2156                 SC2IFP(sc)->if_flags |= sc->sc_flags_backup;
2157                 carp_set_state(sc, INIT);
2158                 carp_setrun(sc, 0);
2159                 if (sc->sc_suppress)
2160                         carp_suppress_preempt--;
2161                 sc->sc_suppress = 0;
2162         }
2163
2164         return;
2165 }
2166
2167 static int
2168 carp_modevent(module_t mod, int type, void *data)
2169 {
2170         switch (type) {
2171         case MOD_LOAD:
2172                 if_detach_event_tag = EVENTHANDLER_REGISTER(ifnet_departure_event,
2173                     carp_ifdetach, NULL, EVENTHANDLER_PRI_ANY);
2174                 if (if_detach_event_tag == NULL)
2175                         return (ENOMEM);
2176
2177                 LIST_INIT(&carpif_list);
2178                 if_clone_attach(&carp_cloner);
2179                 break;
2180
2181         case MOD_UNLOAD:
2182                 EVENTHANDLER_DEREGISTER(ifnet_departure_event, if_detach_event_tag);
2183                 if_clone_detach(&carp_cloner);
2184                 break;
2185
2186         default:
2187                 return (EINVAL);
2188         }
2189
2190         return (0);
2191 }
2192
2193 static moduledata_t carp_mod = {
2194         "carp",
2195         carp_modevent,
2196         0
2197 };
2198
2199 DECLARE_MODULE(carp, carp_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);