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 #include <sys/cdefs.h>
  28 __FBSDID("$FreeBSD$");
  29
  30 #include "opt_carp.h"
  31 #include "opt_bpf.h"
  32 #include "opt_inet.h"
  33 #include "opt_inet6.h"
  34
  35 #include <sys/types.h>
  36 #include <sys/param.h>
  37 #include <sys/systm.h>
  38 #include <sys/conf.h>
  39 #include <sys/kernel.h>
  40 #include <sys/limits.h>
  41 #include <sys/malloc.h>
  42 #include <sys/mbuf.h>
  43 #include <sys/module.h>
  44 #include <sys/time.h>
  45 #include <sys/priv.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
  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/fddi.h>
  61 #include <net/iso88025.h>
  62 #include <net/if.h>
  63 #include <net/if_clone.h>
  64 #include <net/if_dl.h>
  65 #include <net/if_types.h>
  66 #include <net/route.h>
  67 #include <net/vnet.h>
  68
  69 #ifdef INET
  70 #include <netinet/in.h>
  71 #include <netinet/in_var.h>
  72 #include <netinet/in_systm.h>
  73 #include <netinet/ip.h>
  74 #include <netinet/ip_var.h>
  75 #include <netinet/if_ether.h>
  76 #include <machine/in_cksum.h>
  77 #endif
  78
  79 #ifdef INET6
  80 #include <netinet/icmp6.h>
  81 #include <netinet/ip6.h>
  82 #include <netinet6/ip6_var.h>
  83 #include <netinet6/scope6_var.h>
  84 #include <netinet6/nd6.h>
  85 #endif
  86
  87 #include <crypto/sha1.h>
  88 #include <netinet/ip_carp.h>
  89
  90 #define CARP_IFNAME     "carp"
  91 static MALLOC_DEFINE(M_CARP, "CARP", "CARP interfaces");
  92 SYSCTL_DECL(_net_inet_carp);
  93
  94 struct carp_softc {
  95         struct ifnet            *sc_ifp;        /* Interface clue */
  96         struct ifnet            *sc_carpdev;    /* Pointer to parent interface */
  97         struct in_ifaddr        *sc_ia;         /* primary iface address */
  98         struct ip_moptions       sc_imo;
  99 #ifdef INET6
 100         struct in6_ifaddr       *sc_ia6;        /* primary iface address v6 */
 101         struct ip6_moptions      sc_im6o;
 102 #endif /* INET6 */
 103         TAILQ_ENTRY(carp_softc)  sc_list;
 104
 105         enum { INIT = 0, BACKUP, MASTER }       sc_state;
 106
 107         int                      sc_flags_backup;
 108         int                      sc_suppress;
 109
 110         int                      sc_sendad_errors;
 111 #define CARP_SENDAD_MAX_ERRORS  3
 112         int                      sc_sendad_success;
 113 #define CARP_SENDAD_MIN_SUCCESS 3
 114
 115         int                      sc_vhid;
 116         int                      sc_advskew;
 117         int                      sc_naddrs;
 118         int                      sc_naddrs6;
 119         int                      sc_advbase;    /* seconds */
 120         int                      sc_init_counter;
 121         u_int64_t                sc_counter;
 122
 123         /* authentication */
 124 #define CARP_HMAC_PAD   64
 125         unsigned char sc_key[CARP_KEY_LEN];
 126         unsigned char sc_pad[CARP_HMAC_PAD];
 127         SHA1_CTX sc_sha1;
 128
 129         struct callout           sc_ad_tmo;     /* advertisement timeout */
 130         struct callout           sc_md_tmo;     /* master down timeout */
 131         struct callout           sc_md6_tmo;    /* master down timeout */
 132
 133         LIST_ENTRY(carp_softc)   sc_next;       /* Interface clue */
 134 };
 135 #define SC2IFP(sc)      ((sc)->sc_ifp)
 136
 137 int carp_suppress_preempt = 0;
 138 int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 1, 0, 0 };    /* XXX for now */
 139 SYSCTL_INT(_net_inet_carp, CARPCTL_ALLOW, allow, CTLFLAG_RW,
 140     &carp_opts[CARPCTL_ALLOW], 0, "Accept incoming CARP packets");
 141 SYSCTL_INT(_net_inet_carp, CARPCTL_PREEMPT, preempt, CTLFLAG_RW,
 142     &carp_opts[CARPCTL_PREEMPT], 0, "high-priority backup preemption mode");
 143 SYSCTL_INT(_net_inet_carp, CARPCTL_LOG, log, CTLFLAG_RW,
 144     &carp_opts[CARPCTL_LOG], 0, "log bad carp packets");
 145 SYSCTL_INT(_net_inet_carp, CARPCTL_ARPBALANCE, arpbalance, CTLFLAG_RW,
 146     &carp_opts[CARPCTL_ARPBALANCE], 0, "balance arp responses");
 147 SYSCTL_INT(_net_inet_carp, OID_AUTO, suppress_preempt, CTLFLAG_RD,
 148     &carp_suppress_preempt, 0, "Preemption is suppressed");
 149
 150 struct carpstats carpstats;
 151 SYSCTL_STRUCT(_net_inet_carp, CARPCTL_STATS, stats, CTLFLAG_RW,
 152     &carpstats, carpstats,
 153     "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 154
 155 struct carp_if {
 156         TAILQ_HEAD(, carp_softc) vhif_vrs;
 157         int vhif_nvrs;
 158
 159         struct ifnet    *vhif_ifp;
 160         struct mtx       vhif_mtx;
 161 };
 162
 163 /* Get carp_if from softc. Valid after carp_set_addr{,6}. */
 164 #define SC2CIF(sc)              ((struct carp_if *)(sc)->sc_carpdev->if_carp)
 165
 166 /* lock per carp_if queue */
 167 #define CARP_LOCK_INIT(cif)     mtx_init(&(cif)->vhif_mtx, "carp_if",   \
 168         NULL, MTX_DEF)
 169 #define CARP_LOCK_DESTROY(cif)  mtx_destroy(&(cif)->vhif_mtx)
 170 #define CARP_LOCK_ASSERT(cif)   mtx_assert(&(cif)->vhif_mtx, MA_OWNED)
 171 #define CARP_LOCK(cif)          mtx_lock(&(cif)->vhif_mtx)
 172 #define CARP_UNLOCK(cif)        mtx_unlock(&(cif)->vhif_mtx)
 173
 174 #define CARP_SCLOCK(sc)         mtx_lock(&SC2CIF(sc)->vhif_mtx)
 175 #define CARP_SCUNLOCK(sc)       mtx_unlock(&SC2CIF(sc)->vhif_mtx)
 176 #define CARP_SCLOCK_ASSERT(sc)  mtx_assert(&SC2CIF(sc)->vhif_mtx, MA_OWNED)
 177
 178 #define CARP_LOG(...)   do {                            \
 179         if (carp_opts[CARPCTL_LOG] > 0)                 \
 180                 log(LOG_INFO, __VA_ARGS__);             \
 181 } while (0)
 182
 183 #define CARP_DEBUG(...) do {                            \
 184         if (carp_opts[CARPCTL_LOG] > 1)                 \
 185                 log(LOG_DEBUG, __VA_ARGS__);            \
 186 } while (0)
 187
 188 static void     carp_hmac_prepare(struct carp_softc *);
 189 static void     carp_hmac_generate(struct carp_softc *, u_int32_t *,
 190                     unsigned char *);
 191 static int      carp_hmac_verify(struct carp_softc *, u_int32_t *,
 192                     unsigned char *);
 193 static void     carp_setroute(struct carp_softc *, int);
 194 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 195 static int      carp_clone_create(struct if_clone *, int, caddr_t);
 196 static void     carp_clone_destroy(struct ifnet *);
 197 static void     carpdetach(struct carp_softc *, int);
 198 static int      carp_prepare_ad(struct mbuf *, struct carp_softc *,
 199                     struct carp_header *);
 200 static void     carp_send_ad_all(void);
 201 static void     carp_send_ad(void *);
 202 static void     carp_send_ad_locked(struct carp_softc *);
 203 static void     carp_send_arp(struct carp_softc *);
 204 static void     carp_master_down(void *);
 205 static void     carp_master_down_locked(struct carp_softc *);
 206 static int      carp_ioctl(struct ifnet *, u_long, caddr_t);
 207 static int      carp_looutput(struct ifnet *, struct mbuf *, struct sockaddr *,
 208                     struct route *);
 209 static void     carp_start(struct ifnet *);
 210 static void     carp_setrun(struct carp_softc *, sa_family_t);
 211 static void     carp_set_state(struct carp_softc *, int);
 212 static int      carp_addrcount(struct carp_if *, struct in_ifaddr *, int);
 213 enum    { CARP_COUNT_MASTER, CARP_COUNT_RUNNING };
 214
 215 static void     carp_multicast_cleanup(struct carp_softc *);
 216 static int      carp_set_addr(struct carp_softc *, struct sockaddr_in *);
 217 static int      carp_del_addr(struct carp_softc *, struct sockaddr_in *);
 218 static void     carp_carpdev_state_locked(struct carp_if *);
 219 static void     carp_sc_state_locked(struct carp_softc *);
 220 #ifdef INET6
 221 static void     carp_send_na(struct carp_softc *);
 222 static int      carp_set_addr6(struct carp_softc *, struct sockaddr_in6 *);
 223 static int      carp_del_addr6(struct carp_softc *, struct sockaddr_in6 *);
 224 static void     carp_multicast6_cleanup(struct carp_softc *);
 225 #endif
 226
 227 static LIST_HEAD(, carp_softc) carpif_list;
 228 static struct mtx carp_mtx;
 229 IFC_SIMPLE_DECLARE(carp, 0);
 230
 231 static eventhandler_tag if_detach_event_tag;
 232
 233 static __inline u_int16_t
 234 carp_cksum(struct mbuf *m, int len)
 235 {
 236         return (in_cksum(m, len));
 237 }
 238
 239 static void
 240 carp_hmac_prepare(struct carp_softc *sc)
 241 {
 242         u_int8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 243         u_int8_t vhid = sc->sc_vhid & 0xff;
 244         struct ifaddr *ifa;
 245         int i, found;
 246 #ifdef INET
 247         struct in_addr last, cur, in;
 248 #endif
 249 #ifdef INET6
 250         struct in6_addr last6, cur6, in6;
 251 #endif
 252
 253         if (sc->sc_carpdev)
 254                 CARP_SCLOCK(sc);
 255
 256         /* XXX: possible race here */
 257
 258         /* compute ipad from key */
 259         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 260         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 261         for (i = 0; i < sizeof(sc->sc_pad); i++)
 262                 sc->sc_pad[i] ^= 0x36;
 263
 264         /* precompute first part of inner hash */
 265         SHA1Init(&sc->sc_sha1);
 266         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 267         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 268         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 269         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 270 #ifdef INET
 271         cur.s_addr = 0;
 272         do {
 273                 found = 0;
 274                 last = cur;
 275                 cur.s_addr = 0xffffffff;
 276                 IF_ADDR_LOCK(SC2IFP(sc));
 277                 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 278                         in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 279                         if (ifa->ifa_addr->sa_family == AF_INET &&
 280                             ntohl(in.s_addr) > ntohl(last.s_addr) &&
 281                             ntohl(in.s_addr) < ntohl(cur.s_addr)) {
 282                                 cur.s_addr = in.s_addr;
 283                                 found++;
 284                         }
 285                 }
 286                 IF_ADDR_UNLOCK(SC2IFP(sc));
 287                 if (found)
 288                         SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur));
 289         } while (found);
 290 #endif /* INET */
 291 #ifdef INET6
 292         memset(&cur6, 0, sizeof(cur6));
 293         do {
 294                 found = 0;
 295                 last6 = cur6;
 296                 memset(&cur6, 0xff, sizeof(cur6));
 297                 IF_ADDR_LOCK(SC2IFP(sc));
 298                 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 299                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 300                         if (IN6_IS_SCOPE_EMBED(&in6))
 301                                 in6.s6_addr16[1] = 0;
 302                         if (ifa->ifa_addr->sa_family == AF_INET6 &&
 303                             memcmp(&in6, &last6, sizeof(in6)) > 0 &&
 304                             memcmp(&in6, &cur6, sizeof(in6)) < 0) {
 305                                 cur6 = in6;
 306                                 found++;
 307                         }
 308                 }
 309                 IF_ADDR_UNLOCK(SC2IFP(sc));
 310                 if (found)
 311                         SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6));
 312         } while (found);
 313 #endif /* INET6 */
 314
 315         /* convert ipad to opad */
 316         for (i = 0; i < sizeof(sc->sc_pad); i++)
 317                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 318
 319         if (sc->sc_carpdev)
 320                 CARP_SCUNLOCK(sc);
 321 }
 322
 323 static void
 324 carp_hmac_generate(struct carp_softc *sc, u_int32_t counter[2],
 325     unsigned char md[20])
 326 {
 327         SHA1_CTX sha1ctx;
 328
 329         /* fetch first half of inner hash */
 330         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 331
 332         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 333         SHA1Final(md, &sha1ctx);
 334
 335         /* outer hash */
 336         SHA1Init(&sha1ctx);
 337         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 338         SHA1Update(&sha1ctx, md, 20);
 339         SHA1Final(md, &sha1ctx);
 340 }
 341
 342 static int
 343 carp_hmac_verify(struct carp_softc *sc, u_int32_t counter[2],
 344     unsigned char md[20])
 345 {
 346         unsigned char md2[20];
 347
 348         CARP_SCLOCK_ASSERT(sc);
 349
 350         carp_hmac_generate(sc, counter, md2);
 351
 352         return (bcmp(md, md2, sizeof(md2)));
 353 }
 354
 355 static void
 356 carp_setroute(struct carp_softc *sc, int cmd)
 357 {
 358         struct ifaddr *ifa;
 359         int s;
 360
 361         if (sc->sc_carpdev)
 362                 CARP_SCLOCK_ASSERT(sc);
 363
 364         s = splnet();
 365         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 366                 if (ifa->ifa_addr->sa_family == AF_INET &&
 367                     sc->sc_carpdev != NULL) {
 368                         int count = carp_addrcount(
 369                             (struct carp_if *)sc->sc_carpdev->if_carp,
 370                             ifatoia(ifa), CARP_COUNT_MASTER);
 371
 372                         if ((cmd == RTM_ADD && count == 1) ||
 373                             (cmd == RTM_DELETE && count == 0))
 374                                 rtinit(ifa, cmd, RTF_UP | RTF_HOST);
 375                 }
 376         }
 377         splx(s);
 378 }
 379
 380 static int
 381 carp_clone_create(struct if_clone *ifc, int unit, caddr_t params)
 382 {
 383
 384         struct carp_softc *sc;
 385         struct ifnet *ifp;
 386
 387         sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
 388         ifp = SC2IFP(sc) = if_alloc(IFT_ETHER);
 389         if (ifp == NULL) {
 390                 free(sc, M_CARP);
 391                 return (ENOSPC);
 392         }
 393
 394         sc->sc_flags_backup = 0;
 395         sc->sc_suppress = 0;
 396         sc->sc_advbase = CARP_DFLTINTV;
 397         sc->sc_vhid = -1;       /* required setting */
 398         sc->sc_advskew = 0;
 399         sc->sc_init_counter = 1;
 400         sc->sc_naddrs = sc->sc_naddrs6 = 0; /* M_ZERO? */
 401         sc->sc_imo.imo_membership = (struct in_multi **)malloc(
 402             (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
 403             M_WAITOK);
 404         sc->sc_imo.imo_mfilters = NULL;
 405         sc->sc_imo.imo_max_memberships = IP_MIN_MEMBERSHIPS;
 406         sc->sc_imo.imo_multicast_vif = -1;
 407 #ifdef INET6
 408         sc->sc_im6o.im6o_membership = (struct in6_multi **)malloc(
 409             (sizeof(struct in6_multi *) * IPV6_MIN_MEMBERSHIPS), M_CARP,
 410             M_WAITOK);
 411         sc->sc_im6o.im6o_mfilters = NULL;
 412         sc->sc_im6o.im6o_max_memberships = IPV6_MIN_MEMBERSHIPS;
 413         sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL;
 414 #endif
 415
 416         callout_init(&sc->sc_ad_tmo, CALLOUT_MPSAFE);
 417         callout_init(&sc->sc_md_tmo, CALLOUT_MPSAFE);
 418         callout_init(&sc->sc_md6_tmo, CALLOUT_MPSAFE);
 419
 420         ifp->if_softc = sc;
 421         if_initname(ifp, CARP_IFNAME, unit);
 422         ifp->if_mtu = ETHERMTU;
 423         ifp->if_flags = IFF_LOOPBACK;
 424         ifp->if_ioctl = carp_ioctl;
 425         ifp->if_output = carp_looutput;
 426         ifp->if_start = carp_start;
 427         ifp->if_type = IFT_CARP;
 428         ifp->if_snd.ifq_maxlen = ifqmaxlen;
 429         ifp->if_hdrlen = 0;
 430         if_attach(ifp);
 431         bpfattach(SC2IFP(sc), DLT_NULL, sizeof(u_int32_t));
 432         mtx_lock(&carp_mtx);
 433         LIST_INSERT_HEAD(&carpif_list, sc, sc_next);
 434         mtx_unlock(&carp_mtx);
 435         return (0);
 436 }
 437
 438 static void
 439 carp_clone_destroy(struct ifnet *ifp)
 440 {
 441         struct carp_softc *sc = ifp->if_softc;
 442
 443         if (sc->sc_carpdev)
 444                 CARP_SCLOCK(sc);
 445         carpdetach(sc, 1);      /* Returns unlocked. */
 446
 447         mtx_lock(&carp_mtx);
 448         LIST_REMOVE(sc, sc_next);
 449         mtx_unlock(&carp_mtx);
 450         bpfdetach(ifp);
 451         if_detach(ifp);
 452         if_free_type(ifp, IFT_ETHER);
 453         free(sc->sc_imo.imo_membership, M_CARP);
 454 #ifdef INET6
 455         free(sc->sc_im6o.im6o_membership, M_CARP);
 456 #endif
 457         free(sc, M_CARP);
 458 }
 459
 460 /*
 461  * This function can be called on CARP interface destroy path,
 462  * and in case of the removal of the underlying interface as
 463  * well. We differentiate these two cases. In the latter case
 464  * we do not cleanup our multicast memberships, since they
 465  * are already freed. Also, in the latter case we do not
 466  * release the lock on return, because the function will be
 467  * called once more, for another CARP instance on the same
 468  * interface.
 469  */
 470 static void
 471 carpdetach(struct carp_softc *sc, int unlock)
 472 {
 473         struct carp_if *cif;
 474
 475         callout_stop(&sc->sc_ad_tmo);
 476         callout_stop(&sc->sc_md_tmo);
 477         callout_stop(&sc->sc_md6_tmo);
 478
 479         if (sc->sc_suppress)
 480                 carp_suppress_preempt--;
 481         sc->sc_suppress = 0;
 482
 483         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS)
 484                 carp_suppress_preempt--;
 485         sc->sc_sendad_errors = 0;
 486
 487         carp_set_state(sc, INIT);
 488         SC2IFP(sc)->if_flags &= ~IFF_UP;
 489         carp_setrun(sc, 0);
 490         if (unlock)
 491                 carp_multicast_cleanup(sc);
 492 #ifdef INET6
 493         carp_multicast6_cleanup(sc);
 494 #endif
 495
 496         if (sc->sc_carpdev != NULL) {
 497                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
 498                 CARP_LOCK_ASSERT(cif);
 499                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
 500                 if (!--cif->vhif_nvrs) {
 501                         ifpromisc(sc->sc_carpdev, 0);
 502                         sc->sc_carpdev->if_carp = NULL;
 503                         CARP_LOCK_DESTROY(cif);
 504                         free(cif, M_CARP);
 505                 } else if (unlock)
 506                         CARP_UNLOCK(cif);
 507                 sc->sc_carpdev = NULL;
 508         }
 509 }
 510
 511 /* Detach an interface from the carp. */
 512 static void
 513 carp_ifdetach(void *arg __unused, struct ifnet *ifp)
 514 {
 515         struct carp_if *cif = (struct carp_if *)ifp->if_carp;
 516         struct carp_softc *sc, *nextsc;
 517
 518         if (cif == NULL)
 519                 return;
 520
 521         /*
 522          * XXX: At the end of for() cycle the lock will be destroyed.
 523          */
 524         CARP_LOCK(cif);
 525         for (sc = TAILQ_FIRST(&cif->vhif_vrs); sc; sc = nextsc) {
 526                 nextsc = TAILQ_NEXT(sc, sc_list);
 527                 carpdetach(sc, 0);
 528         }
 529 }
 530
 531 /*
 532  * process input packet.
 533  * we have rearranged checks order compared to the rfc,
 534  * but it seems more efficient this way or not possible otherwise.
 535  */
 536 void
 537 carp_input(struct mbuf *m, int hlen)
 538 {
 539         struct ip *ip = mtod(m, struct ip *);
 540         struct carp_header *ch;
 541         int iplen, len;
 542
 543         CARPSTATS_INC(carps_ipackets);
 544
 545         if (!carp_opts[CARPCTL_ALLOW]) {
 546                 m_freem(m);
 547                 return;
 548         }
 549
 550         /* check if received on a valid carp interface */
 551         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 552                 CARPSTATS_INC(carps_badif);
 553                 CARP_DEBUG("carp_input: packet received on non-carp "
 554                     "interface: %s\n",
 555                     m->m_pkthdr.rcvif->if_xname);
 556                 m_freem(m);
 557                 return;
 558         }
 559
 560         /* verify that the IP TTL is 255.  */
 561         if (ip->ip_ttl != CARP_DFLTTL) {
 562                 CARPSTATS_INC(carps_badttl);
 563                 CARP_DEBUG("carp_input: received ttl %d != 255 on %s\n",
 564                     ip->ip_ttl,
 565                     m->m_pkthdr.rcvif->if_xname);
 566                 m_freem(m);
 567                 return;
 568         }
 569
 570         iplen = ip->ip_hl << 2;
 571
 572         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 573                 CARPSTATS_INC(carps_badlen);
 574                 CARP_DEBUG("carp_input: received len %zd < "
 575                     "sizeof(struct carp_header) on %s\n",
 576                     m->m_len - sizeof(struct ip),
 577                     m->m_pkthdr.rcvif->if_xname);
 578                 m_freem(m);
 579                 return;
 580         }
 581
 582         if (iplen + sizeof(*ch) < m->m_len) {
 583                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 584                         CARPSTATS_INC(carps_hdrops);
 585                         CARP_DEBUG("carp_input: pullup failed\n");
 586                         return;
 587                 }
 588                 ip = mtod(m, struct ip *);
 589         }
 590         ch = (struct carp_header *)((char *)ip + iplen);
 591
 592         /*
 593          * verify that the received packet length is
 594          * equal to the CARP header
 595          */
 596         len = iplen + sizeof(*ch);
 597         if (len > m->m_pkthdr.len) {
 598                 CARPSTATS_INC(carps_badlen);
 599                 CARP_DEBUG("carp_input: packet too short %d on %s\n",
 600                     m->m_pkthdr.len,
 601                     m->m_pkthdr.rcvif->if_xname);
 602                 m_freem(m);
 603                 return;
 604         }
 605
 606         if ((m = m_pullup(m, len)) == NULL) {
 607                 CARPSTATS_INC(carps_hdrops);
 608                 return;
 609         }
 610         ip = mtod(m, struct ip *);
 611         ch = (struct carp_header *)((char *)ip + iplen);
 612
 613         /* verify the CARP checksum */
 614         m->m_data += iplen;
 615         if (carp_cksum(m, len - iplen)) {
 616                 CARPSTATS_INC(carps_badsum);
 617                 CARP_DEBUG("carp_input: checksum failed on %s\n",
 618                     m->m_pkthdr.rcvif->if_xname);
 619                 m_freem(m);
 620                 return;
 621         }
 622         m->m_data -= iplen;
 623
 624         carp_input_c(m, ch, AF_INET);
 625 }
 626
 627 #ifdef INET6
 628 int
 629 carp6_input(struct mbuf **mp, int *offp, int proto)
 630 {
 631         struct mbuf *m = *mp;
 632         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 633         struct carp_header *ch;
 634         u_int len;
 635
 636         CARPSTATS_INC(carps_ipackets6);
 637
 638         if (!carp_opts[CARPCTL_ALLOW]) {
 639                 m_freem(m);
 640                 return (IPPROTO_DONE);
 641         }
 642
 643         /* check if received on a valid carp interface */
 644         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 645                 CARPSTATS_INC(carps_badif);
 646                 CARP_DEBUG("carp6_input: packet received on non-carp "
 647                     "interface: %s\n",
 648                     m->m_pkthdr.rcvif->if_xname);
 649                 m_freem(m);
 650                 return (IPPROTO_DONE);
 651         }
 652
 653         /* verify that the IP TTL is 255 */
 654         if (ip6->ip6_hlim != CARP_DFLTTL) {
 655                 CARPSTATS_INC(carps_badttl);
 656                 CARP_DEBUG("carp6_input: received ttl %d != 255 on %s\n",
 657                     ip6->ip6_hlim,
 658                     m->m_pkthdr.rcvif->if_xname);
 659                 m_freem(m);
 660                 return (IPPROTO_DONE);
 661         }
 662
 663         /* verify that we have a complete carp packet */
 664         len = m->m_len;
 665         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 666         if (ch == NULL) {
 667                 CARPSTATS_INC(carps_badlen);
 668                 CARP_DEBUG("carp6_input: packet size %u too small\n", len);
 669                 return (IPPROTO_DONE);
 670         }
 671
 672
 673         /* verify the CARP checksum */
 674         m->m_data += *offp;
 675         if (carp_cksum(m, sizeof(*ch))) {
 676                 CARPSTATS_INC(carps_badsum);
 677                 CARP_DEBUG("carp6_input: checksum failed, on %s\n",
 678                     m->m_pkthdr.rcvif->if_xname);
 679                 m_freem(m);
 680                 return (IPPROTO_DONE);
 681         }
 682         m->m_data -= *offp;
 683
 684         carp_input_c(m, ch, AF_INET6);
 685         return (IPPROTO_DONE);
 686 }
 687 #endif /* INET6 */
 688
 689 static void
 690 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 691 {
 692         struct ifnet *ifp = m->m_pkthdr.rcvif;
 693         struct carp_softc *sc;
 694         u_int64_t tmp_counter;
 695         struct timeval sc_tv, ch_tv;
 696
 697         /* verify that the VHID is valid on the receiving interface */
 698         CARP_LOCK(ifp->if_carp);
 699         TAILQ_FOREACH(sc, &((struct carp_if *)ifp->if_carp)->vhif_vrs, sc_list)
 700                 if (sc->sc_vhid == ch->carp_vhid)
 701                         break;
 702
 703         if (!sc || !((SC2IFP(sc)->if_flags & IFF_UP) &&
 704             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING))) {
 705                 CARPSTATS_INC(carps_badvhid);
 706                 CARP_UNLOCK(ifp->if_carp);
 707                 m_freem(m);
 708                 return;
 709         }
 710
 711         getmicrotime(&SC2IFP(sc)->if_lastchange);
 712         SC2IFP(sc)->if_ipackets++;
 713         SC2IFP(sc)->if_ibytes += m->m_pkthdr.len;
 714
 715         if (bpf_peers_present(SC2IFP(sc)->if_bpf)) {
 716                 struct ip *ip = mtod(m, struct ip *);
 717                 uint32_t af1 = af;
 718
 719                 /* BPF wants net byte order */
 720                 ip->ip_len = htons(ip->ip_len + (ip->ip_hl << 2));
 721                 ip->ip_off = htons(ip->ip_off);
 722                 bpf_mtap2(SC2IFP(sc)->if_bpf, &af1, sizeof(af1), m);
 723         }
 724
 725         /* verify the CARP version. */
 726         if (ch->carp_version != CARP_VERSION) {
 727                 CARPSTATS_INC(carps_badver);
 728                 SC2IFP(sc)->if_ierrors++;
 729                 CARP_UNLOCK(ifp->if_carp);
 730                 CARP_DEBUG("%s; invalid version %d\n",
 731                     SC2IFP(sc)->if_xname,
 732                     ch->carp_version);
 733                 m_freem(m);
 734                 return;
 735         }
 736
 737         /* verify the hash */
 738         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 739                 CARPSTATS_INC(carps_badauth);
 740                 SC2IFP(sc)->if_ierrors++;
 741                 CARP_UNLOCK(ifp->if_carp);
 742                 CARP_DEBUG("%s: incorrect hash\n", SC2IFP(sc)->if_xname);
 743                 m_freem(m);
 744                 return;
 745         }
 746
 747         tmp_counter = ntohl(ch->carp_counter[0]);
 748         tmp_counter = tmp_counter<<32;
 749         tmp_counter += ntohl(ch->carp_counter[1]);
 750
 751         /* XXX Replay protection goes here */
 752
 753         sc->sc_init_counter = 0;
 754         sc->sc_counter = tmp_counter;
 755
 756         sc_tv.tv_sec = sc->sc_advbase;
 757         if (carp_suppress_preempt && sc->sc_advskew <  240)
 758                 sc_tv.tv_usec = 240 * 1000000 / 256;
 759         else
 760                 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256;
 761         ch_tv.tv_sec = ch->carp_advbase;
 762         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 763
 764         switch (sc->sc_state) {
 765         case INIT:
 766                 break;
 767         case MASTER:
 768                 /*
 769                  * If we receive an advertisement from a master who's going to
 770                  * be more frequent than us, go into BACKUP state.
 771                  */
 772                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 773                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 774                         callout_stop(&sc->sc_ad_tmo);
 775                         CARP_LOG("%s: MASTER -> BACKUP "
 776                            "(more frequent advertisement received)\n",
 777                            SC2IFP(sc)->if_xname);
 778                         carp_set_state(sc, BACKUP);
 779                         carp_setrun(sc, 0);
 780                         carp_setroute(sc, RTM_DELETE);
 781                 }
 782                 break;
 783         case BACKUP:
 784                 /*
 785                  * If we're pre-empting masters who advertise slower than us,
 786                  * and this one claims to be slower, treat him as down.
 787                  */
 788                 if (carp_opts[CARPCTL_PREEMPT] &&
 789                     timevalcmp(&sc_tv, &ch_tv, <)) {
 790                         CARP_LOG("%s: BACKUP -> MASTER "
 791                             "(preempting a slower master)\n",
 792                             SC2IFP(sc)->if_xname);
 793                         carp_master_down_locked(sc);
 794                         break;
 795                 }
 796
 797                 /*
 798                  *  If the master is going to advertise at such a low frequency
 799                  *  that he's guaranteed to time out, we'd might as well just
 800                  *  treat him as timed out now.
 801                  */
 802                 sc_tv.tv_sec = sc->sc_advbase * 3;
 803                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 804                         CARP_LOG("%s: BACKUP -> MASTER "
 805                             "(master timed out)\n",
 806                             SC2IFP(sc)->if_xname);
 807                         carp_master_down_locked(sc);
 808                         break;
 809                 }
 810
 811                 /*
 812                  * Otherwise, we reset the counter and wait for the next
 813                  * advertisement.
 814                  */
 815                 carp_setrun(sc, af);
 816                 break;
 817         }
 818
 819         CARP_UNLOCK(ifp->if_carp);
 820
 821         m_freem(m);
 822         return;
 823 }
 824
 825 static int
 826 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 827 {
 828         struct m_tag *mtag;
 829         struct ifnet *ifp = SC2IFP(sc);
 830
 831         if (sc->sc_init_counter) {
 832                 /* this could also be seconds since unix epoch */
 833                 sc->sc_counter = arc4random();
 834                 sc->sc_counter = sc->sc_counter << 32;
 835                 sc->sc_counter += arc4random();
 836         } else
 837                 sc->sc_counter++;
 838
 839         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 840         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 841
 842         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 843
 844         /* Tag packet for carp_output */
 845         mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct ifnet *), M_NOWAIT);
 846         if (mtag == NULL) {
 847                 m_freem(m);
 848                 SC2IFP(sc)->if_oerrors++;
 849                 return (ENOMEM);
 850         }
 851         bcopy(&ifp, (caddr_t)(mtag + 1), sizeof(struct ifnet *));
 852         m_tag_prepend(m, mtag);
 853
 854         return (0);
 855 }
 856
 857 static void
 858 carp_send_ad_all(void)
 859 {
 860         struct carp_softc *sc;
 861
 862         mtx_lock(&carp_mtx);
 863         LIST_FOREACH(sc, &carpif_list, sc_next) {
 864                 if (sc->sc_carpdev == NULL)
 865                         continue;
 866                 CARP_SCLOCK(sc);
 867                 if ((SC2IFP(sc)->if_flags & IFF_UP) &&
 868                     (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING) &&
 869                      sc->sc_state == MASTER)
 870                         carp_send_ad_locked(sc);
 871                 CARP_SCUNLOCK(sc);
 872         }
 873         mtx_unlock(&carp_mtx);
 874 }
 875
 876 static void
 877 carp_send_ad(void *v)
 878 {
 879         struct carp_softc *sc = v;
 880
 881         CARP_SCLOCK(sc);
 882         carp_send_ad_locked(sc);
 883         CARP_SCUNLOCK(sc);
 884 }
 885
 886 static void
 887 carp_send_ad_locked(struct carp_softc *sc)
 888 {
 889         struct carp_header ch;
 890         struct timeval tv;
 891         struct carp_header *ch_ptr;
 892         struct mbuf *m;
 893         int len, advbase, advskew;
 894
 895         CARP_SCLOCK_ASSERT(sc);
 896
 897         /* bow out if we've lost our UPness or RUNNINGuiness */
 898         if (!((SC2IFP(sc)->if_flags & IFF_UP) &&
 899             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING))) {
 900                 advbase = 255;
 901                 advskew = 255;
 902         } else {
 903                 advbase = sc->sc_advbase;
 904                 if (!carp_suppress_preempt || sc->sc_advskew > 240)
 905                         advskew = sc->sc_advskew;
 906                 else
 907                         advskew = 240;
 908                 tv.tv_sec = advbase;
 909                 tv.tv_usec = advskew * 1000000 / 256;
 910         }
 911
 912         ch.carp_version = CARP_VERSION;
 913         ch.carp_type = CARP_ADVERTISEMENT;
 914         ch.carp_vhid = sc->sc_vhid;
 915         ch.carp_advbase = advbase;
 916         ch.carp_advskew = advskew;
 917         ch.carp_authlen = 7;    /* XXX DEFINE */
 918         ch.carp_pad1 = 0;       /* must be zero */
 919         ch.carp_cksum = 0;
 920
 921 #ifdef INET
 922         if (sc->sc_ia) {
 923                 struct ip *ip;
 924
 925                 MGETHDR(m, M_DONTWAIT, MT_HEADER);
 926                 if (m == NULL) {
 927                         SC2IFP(sc)->if_oerrors++;
 928                         CARPSTATS_INC(carps_onomem);
 929                         /* XXX maybe less ? */
 930                         if (advbase != 255 || advskew != 255)
 931                                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
 932                                     carp_send_ad, sc);
 933                         return;
 934                 }
 935                 len = sizeof(*ip) + sizeof(ch);
 936                 m->m_pkthdr.len = len;
 937                 m->m_pkthdr.rcvif = NULL;
 938                 m->m_len = len;
 939                 MH_ALIGN(m, m->m_len);
 940                 m->m_flags |= M_MCAST;
 941                 ip = mtod(m, struct ip *);
 942                 ip->ip_v = IPVERSION;
 943                 ip->ip_hl = sizeof(*ip) >> 2;
 944                 ip->ip_tos = IPTOS_LOWDELAY;
 945                 ip->ip_len = len;
 946                 ip->ip_id = ip_newid();
 947                 ip->ip_off = IP_DF;
 948                 ip->ip_ttl = CARP_DFLTTL;
 949                 ip->ip_p = IPPROTO_CARP;
 950                 ip->ip_sum = 0;
 951                 ip->ip_src.s_addr = sc->sc_ia->ia_addr.sin_addr.s_addr;
 952                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 953
 954                 ch_ptr = (struct carp_header *)(&ip[1]);
 955                 bcopy(&ch, ch_ptr, sizeof(ch));
 956                 if (carp_prepare_ad(m, sc, ch_ptr))
 957                         return;
 958
 959                 m->m_data += sizeof(*ip);
 960                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip));
 961                 m->m_data -= sizeof(*ip);
 962
 963                 getmicrotime(&SC2IFP(sc)->if_lastchange);
 964                 SC2IFP(sc)->if_opackets++;
 965                 SC2IFP(sc)->if_obytes += len;
 966                 CARPSTATS_INC(carps_opackets);
 967
 968                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, NULL)) {
 969                         SC2IFP(sc)->if_oerrors++;
 970                         if (sc->sc_sendad_errors < INT_MAX)
 971                                 sc->sc_sendad_errors++;
 972                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 973                                 carp_suppress_preempt++;
 974                                 if (carp_suppress_preempt == 1) {
 975                                         CARP_SCUNLOCK(sc);
 976                                         carp_send_ad_all();
 977                                         CARP_SCLOCK(sc);
 978                                 }
 979                         }
 980                         sc->sc_sendad_success = 0;
 981                 } else {
 982                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 983                                 if (++sc->sc_sendad_success >=
 984                                     CARP_SENDAD_MIN_SUCCESS) {
 985                                         carp_suppress_preempt--;
 986                                         sc->sc_sendad_errors = 0;
 987                                 }
 988                         } else
 989                                 sc->sc_sendad_errors = 0;
 990                 }
 991         }
 992 #endif /* INET */
 993 #ifdef INET6
 994         if (sc->sc_ia6) {
 995                 struct ip6_hdr *ip6;
 996
 997                 MGETHDR(m, M_DONTWAIT, MT_HEADER);
 998                 if (m == NULL) {
 999                         SC2IFP(sc)->if_oerrors++;
1000                         CARPSTATS_INC(carps_onomem);
1001                         /* XXX maybe less ? */
1002                         if (advbase != 255 || advskew != 255)
1003                                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1004                                     carp_send_ad, sc);
1005                         return;
1006                 }
1007                 len = sizeof(*ip6) + sizeof(ch);
1008                 m->m_pkthdr.len = len;
1009                 m->m_pkthdr.rcvif = NULL;
1010                 m->m_len = len;
1011                 MH_ALIGN(m, m->m_len);
1012                 m->m_flags |= M_MCAST;
1013                 ip6 = mtod(m, struct ip6_hdr *);
1014                 bzero(ip6, sizeof(*ip6));
1015                 ip6->ip6_vfc |= IPV6_VERSION;
1016                 ip6->ip6_hlim = CARP_DFLTTL;
1017                 ip6->ip6_nxt = IPPROTO_CARP;
1018                 bcopy(&sc->sc_ia6->ia_addr.sin6_addr, &ip6->ip6_src,
1019                     sizeof(struct in6_addr));
1020                 /* set the multicast destination */
1021
1022                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
1023                 ip6->ip6_dst.s6_addr8[15] = 0x12;
1024                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
1025                         SC2IFP(sc)->if_oerrors++;
1026                         m_freem(m);
1027                         CARP_DEBUG("%s: in6_setscope failed\n", __func__);
1028                         return;
1029                 }
1030
1031                 ch_ptr = (struct carp_header *)(&ip6[1]);
1032                 bcopy(&ch, ch_ptr, sizeof(ch));
1033                 if (carp_prepare_ad(m, sc, ch_ptr))
1034                         return;
1035
1036                 m->m_data += sizeof(*ip6);
1037                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6));
1038                 m->m_data -= sizeof(*ip6);
1039
1040                 getmicrotime(&SC2IFP(sc)->if_lastchange);
1041                 SC2IFP(sc)->if_opackets++;
1042                 SC2IFP(sc)->if_obytes += len;
1043                 CARPSTATS_INC(carps_opackets6);
1044
1045                 if (ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL)) {
1046                         SC2IFP(sc)->if_oerrors++;
1047                         if (sc->sc_sendad_errors < INT_MAX)
1048                                 sc->sc_sendad_errors++;
1049                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
1050                                 carp_suppress_preempt++;
1051                                 if (carp_suppress_preempt == 1) {
1052                                         CARP_SCUNLOCK(sc);
1053                                         carp_send_ad_all();
1054                                         CARP_SCLOCK(sc);
1055                                 }
1056                         }
1057                         sc->sc_sendad_success = 0;
1058                 } else {
1059                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
1060                                 if (++sc->sc_sendad_success >=
1061                                     CARP_SENDAD_MIN_SUCCESS) {
1062                                         carp_suppress_preempt--;
1063                                         sc->sc_sendad_errors = 0;
1064                                 }
1065                         } else
1066                                 sc->sc_sendad_errors = 0;
1067                 }
1068         }
1069 #endif /* INET6 */
1070
1071         if (advbase != 255 || advskew != 255)
1072                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1073                     carp_send_ad, sc);
1074
1075 }
1076
1077 /*
1078  * Broadcast a gratuitous ARP request containing
1079  * the virtual router MAC address for each IP address
1080  * associated with the virtual router.
1081  */
1082 static void
1083 carp_send_arp(struct carp_softc *sc)
1084 {
1085         struct ifaddr *ifa;
1086
1087         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1088
1089                 if (ifa->ifa_addr->sa_family != AF_INET)
1090                         continue;
1091
1092 /*              arprequest(sc->sc_carpdev, &in, &in, IF_LLADDR(sc->sc_ifp)); */
1093                 arp_ifinit2(sc->sc_carpdev, ifa, IF_LLADDR(sc->sc_ifp));
1094
1095                 DELAY(1000);    /* XXX */
1096         }
1097 }
1098
1099 #ifdef INET6
1100 static void
1101 carp_send_na(struct carp_softc *sc)
1102 {
1103         struct ifaddr *ifa;
1104         struct in6_addr *in6;
1105         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1106
1107         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1108
1109                 if (ifa->ifa_addr->sa_family != AF_INET6)
1110                         continue;
1111
1112                 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr;
1113                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1114                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1115                 DELAY(1000);    /* XXX */
1116         }
1117 }
1118 #endif /* INET6 */
1119
1120 static int
1121 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type)
1122 {
1123         struct carp_softc *vh;
1124         struct ifaddr *ifa;
1125         int count = 0;
1126
1127         CARP_LOCK_ASSERT(cif);
1128
1129         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1130                 if ((type == CARP_COUNT_RUNNING &&
1131                     (SC2IFP(vh)->if_flags & IFF_UP) &&
1132                     (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING)) ||
1133                     (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) {
1134                         IF_ADDR_LOCK(SC2IFP(vh));
1135                         TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1136                             ifa_list) {
1137                                 if (ifa->ifa_addr->sa_family == AF_INET &&
1138                                     ia->ia_addr.sin_addr.s_addr ==
1139                                     ifatoia(ifa)->ia_addr.sin_addr.s_addr)
1140                                         count++;
1141                         }
1142                         IF_ADDR_UNLOCK(SC2IFP(vh));
1143                 }
1144         }
1145         return (count);
1146 }
1147
1148 int
1149 carp_iamatch(void *v, struct in_ifaddr *ia,
1150     struct in_addr *isaddr, u_int8_t **enaddr)
1151 {
1152         struct carp_if *cif = v;
1153         struct carp_softc *vh;
1154         int index, count = 0;
1155         struct ifaddr *ifa;
1156
1157         CARP_LOCK(cif);
1158
1159         if (carp_opts[CARPCTL_ARPBALANCE]) {
1160                 /*
1161                  * XXX proof of concept implementation.
1162                  * We use the source ip to decide which virtual host should
1163                  * handle the request. If we're master of that virtual host,
1164                  * then we respond, otherwise, just drop the arp packet on
1165                  * the floor.
1166                  */
1167                 count = carp_addrcount(cif, ia, CARP_COUNT_RUNNING);
1168                 if (count == 0) {
1169                         /* should never reach this */
1170                         CARP_UNLOCK(cif);
1171                         return (0);
1172                 }
1173
1174                 /* this should be a hash, like pf_hash() */
1175                 index = ntohl(isaddr->s_addr) % count;
1176                 count = 0;
1177
1178                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1179                         if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1180                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING)) {
1181                                 IF_ADDR_LOCK(SC2IFP(vh));
1182                                 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1183                                     ifa_list) {
1184                                         if (ifa->ifa_addr->sa_family ==
1185                                             AF_INET &&
1186                                             ia->ia_addr.sin_addr.s_addr ==
1187                                             ifatoia(ifa)->ia_addr.sin_addr.s_addr) {
1188                                                 if (count == index) {
1189                                                         if (vh->sc_state ==
1190                                                             MASTER) {
1191                                                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1192                                                                 IF_ADDR_UNLOCK(SC2IFP(vh));
1193                                                                 CARP_UNLOCK(cif);
1194                                                                 return (1);
1195                                                         } else {
1196                                                                 IF_ADDR_UNLOCK(SC2IFP(vh));
1197                                                                 CARP_UNLOCK(cif);
1198                                                                 return (0);
1199                                                         }
1200                                                 }
1201                                                 count++;
1202                                         }
1203                                 }
1204                                 IF_ADDR_UNLOCK(SC2IFP(vh));
1205                         }
1206                 }
1207         } else {
1208                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1209                         if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1210                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1211                             ia->ia_ifp == SC2IFP(vh) &&
1212                             vh->sc_state == MASTER) {
1213                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1214                                 CARP_UNLOCK(cif);
1215                                 return (1);
1216                         }
1217                 }
1218         }
1219         CARP_UNLOCK(cif);
1220         return (0);
1221 }
1222
1223 #ifdef INET6
1224 struct ifaddr *
1225 carp_iamatch6(void *v, struct in6_addr *taddr)
1226 {
1227         struct carp_if *cif = v;
1228         struct carp_softc *vh;
1229         struct ifaddr *ifa;
1230
1231         CARP_LOCK(cif);
1232         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1233                 IF_ADDR_LOCK(SC2IFP(vh));
1234                 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist, ifa_list) {
1235                         if (IN6_ARE_ADDR_EQUAL(taddr,
1236                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1237                             (SC2IFP(vh)->if_flags & IFF_UP) &&
1238                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1239                             vh->sc_state == MASTER) {
1240                                 ifa_ref(ifa);
1241                                 IF_ADDR_UNLOCK(SC2IFP(vh));
1242                                 CARP_UNLOCK(cif);
1243                                 return (ifa);
1244                         }
1245                 }
1246                 IF_ADDR_UNLOCK(SC2IFP(vh));
1247         }
1248         CARP_UNLOCK(cif);
1249
1250         return (NULL);
1251 }
1252
1253 void *
1254 carp_macmatch6(void *v, struct mbuf *m, const struct in6_addr *taddr)
1255 {
1256         struct m_tag *mtag;
1257         struct carp_if *cif = v;
1258         struct carp_softc *sc;
1259         struct ifaddr *ifa;
1260
1261         CARP_LOCK(cif);
1262         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) {
1263                 IF_ADDR_LOCK(SC2IFP(sc));
1264                 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1265                         if (IN6_ARE_ADDR_EQUAL(taddr,
1266                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1267                             (SC2IFP(sc)->if_flags & IFF_UP) &&
1268                             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING)) {
1269                                 struct ifnet *ifp = SC2IFP(sc);
1270                                 mtag = m_tag_get(PACKET_TAG_CARP,
1271                                     sizeof(struct ifnet *), M_NOWAIT);
1272                                 if (mtag == NULL) {
1273                                         /* better a bit than nothing */
1274                                         IF_ADDR_UNLOCK(SC2IFP(sc));
1275                                         CARP_UNLOCK(cif);
1276                                         return (IF_LLADDR(sc->sc_ifp));
1277                                 }
1278                                 bcopy(&ifp, (caddr_t)(mtag + 1),
1279                                     sizeof(struct ifnet *));
1280                                 m_tag_prepend(m, mtag);
1281
1282                                 IF_ADDR_UNLOCK(SC2IFP(sc));
1283                                 CARP_UNLOCK(cif);
1284                                 return (IF_LLADDR(sc->sc_ifp));
1285                         }
1286                 }
1287                 IF_ADDR_UNLOCK(SC2IFP(sc));
1288         }
1289         CARP_UNLOCK(cif);
1290
1291         return (NULL);
1292 }
1293 #endif
1294
1295 struct ifnet *
1296 carp_forus(void *v, void *dhost)
1297 {
1298         struct carp_if *cif = v;
1299         struct carp_softc *vh;
1300         u_int8_t *ena = dhost;
1301
1302         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1303                 return (NULL);
1304
1305         CARP_LOCK(cif);
1306         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list)
1307                 if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1308                     (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1309                     vh->sc_state == MASTER &&
1310                     !bcmp(dhost, IF_LLADDR(vh->sc_ifp), ETHER_ADDR_LEN)) {
1311                         CARP_UNLOCK(cif);
1312                         return (SC2IFP(vh));
1313                 }
1314
1315         CARP_UNLOCK(cif);
1316         return (NULL);
1317 }
1318
1319 static void
1320 carp_master_down(void *v)
1321 {
1322         struct carp_softc *sc = v;
1323
1324         CARP_SCLOCK(sc);
1325         carp_master_down_locked(sc);
1326         CARP_SCUNLOCK(sc);
1327 }
1328
1329 static void
1330 carp_master_down_locked(struct carp_softc *sc)
1331 {
1332         if (sc->sc_carpdev)
1333                 CARP_SCLOCK_ASSERT(sc);
1334
1335         switch (sc->sc_state) {
1336         case INIT:
1337                 printf("%s: master_down event in INIT state\n",
1338                     SC2IFP(sc)->if_xname);
1339                 break;
1340         case MASTER:
1341                 break;
1342         case BACKUP:
1343                 carp_set_state(sc, MASTER);
1344                 carp_send_ad_locked(sc);
1345                 carp_send_arp(sc);
1346 #ifdef INET6
1347                 carp_send_na(sc);
1348 #endif /* INET6 */
1349                 carp_setrun(sc, 0);
1350                 carp_setroute(sc, RTM_ADD);
1351                 break;
1352         }
1353 }
1354
1355 /*
1356  * When in backup state, af indicates whether to reset the master down timer
1357  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1358  */
1359 static void
1360 carp_setrun(struct carp_softc *sc, sa_family_t af)
1361 {
1362         struct timeval tv;
1363
1364         if (sc->sc_carpdev == NULL) {
1365                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1366                 carp_set_state(sc, INIT);
1367                 return;
1368         } else
1369                 CARP_SCLOCK_ASSERT(sc);
1370
1371         if (SC2IFP(sc)->if_flags & IFF_UP &&
1372             sc->sc_vhid > 0 && (sc->sc_naddrs || sc->sc_naddrs6))
1373                 SC2IFP(sc)->if_drv_flags |= IFF_DRV_RUNNING;
1374         else {
1375                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1376                 carp_setroute(sc, RTM_DELETE);
1377                 return;
1378         }
1379
1380         switch (sc->sc_state) {
1381         case INIT:
1382                 if (carp_opts[CARPCTL_PREEMPT] && !carp_suppress_preempt) {
1383                         carp_send_ad_locked(sc);
1384                         carp_send_arp(sc);
1385 #ifdef INET6
1386                         carp_send_na(sc);
1387 #endif /* INET6 */
1388                         CARP_LOG("%s: INIT -> MASTER (preempting)\n",
1389                             SC2IFP(sc)->if_xname);
1390                         carp_set_state(sc, MASTER);
1391                         carp_setroute(sc, RTM_ADD);
1392                 } else {
1393                         CARP_LOG("%s: INIT -> BACKUP\n", SC2IFP(sc)->if_xname);
1394                         carp_set_state(sc, BACKUP);
1395                         carp_setroute(sc, RTM_DELETE);
1396                         carp_setrun(sc, 0);
1397                 }
1398                 break;
1399         case BACKUP:
1400                 callout_stop(&sc->sc_ad_tmo);
1401                 tv.tv_sec = 3 * sc->sc_advbase;
1402                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1403                 switch (af) {
1404 #ifdef INET
1405                 case AF_INET:
1406                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1407                             carp_master_down, sc);
1408                         break;
1409 #endif /* INET */
1410 #ifdef INET6
1411                 case AF_INET6:
1412                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1413                             carp_master_down, sc);
1414                         break;
1415 #endif /* INET6 */
1416                 default:
1417                         if (sc->sc_naddrs)
1418                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1419                                     carp_master_down, sc);
1420                         if (sc->sc_naddrs6)
1421                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1422                                     carp_master_down, sc);
1423                         break;
1424                 }
1425                 break;
1426         case MASTER:
1427                 tv.tv_sec = sc->sc_advbase;
1428                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1429                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1430                     carp_send_ad, sc);
1431                 break;
1432         }
1433 }
1434
1435 static void
1436 carp_multicast_cleanup(struct carp_softc *sc)
1437 {
1438         struct ip_moptions *imo = &sc->sc_imo;
1439         u_int16_t n = imo->imo_num_memberships;
1440
1441         /* Clean up our own multicast memberships */
1442         while (n-- > 0) {
1443                 if (imo->imo_membership[n] != NULL) {
1444                         in_delmulti(imo->imo_membership[n]);
1445                         imo->imo_membership[n] = NULL;
1446                 }
1447         }
1448         KASSERT(imo->imo_mfilters == NULL,
1449            ("%s: imo_mfilters != NULL", __func__));
1450         imo->imo_num_memberships = 0;
1451         imo->imo_multicast_ifp = NULL;
1452 }
1453
1454 #ifdef INET6
1455 static void
1456 carp_multicast6_cleanup(struct carp_softc *sc)
1457 {
1458         struct ip6_moptions *im6o = &sc->sc_im6o;
1459         u_int16_t n = im6o->im6o_num_memberships;
1460
1461         while (n-- > 0) {
1462                 if (im6o->im6o_membership[n] != NULL) {
1463                         in6_mc_leave(im6o->im6o_membership[n], NULL);
1464                         im6o->im6o_membership[n] = NULL;
1465                 }
1466         }
1467         KASSERT(im6o->im6o_mfilters == NULL,
1468            ("%s: im6o_mfilters != NULL", __func__));
1469         im6o->im6o_num_memberships = 0;
1470         im6o->im6o_multicast_ifp = NULL;
1471 }
1472 #endif
1473
1474 static int
1475 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1476 {
1477         struct ifnet *ifp;
1478         struct carp_if *cif;
1479         struct in_ifaddr *ia, *ia_if;
1480         struct ip_moptions *imo = &sc->sc_imo;
1481         struct in_addr addr;
1482         u_long iaddr = htonl(sin->sin_addr.s_addr);
1483         int own, error;
1484
1485         if (sin->sin_addr.s_addr == 0) {
1486                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1487                         carp_set_state(sc, INIT);
1488                 if (sc->sc_naddrs)
1489                         SC2IFP(sc)->if_flags |= IFF_UP;
1490                 if (sc->sc_carpdev)
1491                         CARP_SCLOCK(sc);
1492                 carp_setrun(sc, 0);
1493                 if (sc->sc_carpdev)
1494                         CARP_SCUNLOCK(sc);
1495                 return (0);
1496         }
1497
1498         /* we have to do it by hands to check we won't match on us */
1499         ia_if = NULL; own = 0;
1500         IN_IFADDR_RLOCK();
1501         TAILQ_FOREACH(ia, &V_in_ifaddrhead, ia_link) {
1502                 /* and, yeah, we need a multicast-capable iface too */
1503                 if (ia->ia_ifp != SC2IFP(sc) &&
1504                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1505                     (iaddr & ia->ia_subnetmask) == ia->ia_subnet) {
1506                         if (!ia_if)
1507                                 ia_if = ia;
1508                         if (sin->sin_addr.s_addr ==
1509                             ia->ia_addr.sin_addr.s_addr)
1510                                 own++;
1511                 }
1512         }
1513
1514         if (!ia_if) {
1515                 IN_IFADDR_RUNLOCK();
1516                 return (EADDRNOTAVAIL);
1517         }
1518
1519         ia = ia_if;
1520         ifa_ref(&ia->ia_ifa);
1521         IN_IFADDR_RUNLOCK();
1522
1523         ifp = ia->ia_ifp;
1524
1525         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1526             (imo->imo_multicast_ifp && imo->imo_multicast_ifp != ifp)) {
1527                 ifa_free(&ia->ia_ifa);
1528                 return (EADDRNOTAVAIL);
1529         }
1530
1531         if (imo->imo_num_memberships == 0) {
1532                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1533                 if ((imo->imo_membership[0] = in_addmulti(&addr, ifp)) ==
1534                     NULL) {
1535                         ifa_free(&ia->ia_ifa);
1536                         return (ENOBUFS);
1537                 }
1538                 imo->imo_num_memberships++;
1539                 imo->imo_multicast_ifp = ifp;
1540                 imo->imo_multicast_ttl = CARP_DFLTTL;
1541                 imo->imo_multicast_loop = 0;
1542         }
1543
1544         if (!ifp->if_carp) {
1545
1546                 cif = malloc(sizeof(*cif), M_CARP,
1547                     M_WAITOK|M_ZERO);
1548                 if (!cif) {
1549                         error = ENOBUFS;
1550                         goto cleanup;
1551                 }
1552                 if ((error = ifpromisc(ifp, 1))) {
1553                         free(cif, M_CARP);
1554                         goto cleanup;
1555                 }
1556
1557                 CARP_LOCK_INIT(cif);
1558                 CARP_LOCK(cif);
1559                 cif->vhif_ifp = ifp;
1560                 TAILQ_INIT(&cif->vhif_vrs);
1561                 ifp->if_carp = cif;
1562
1563         } else {
1564                 struct carp_softc *vr;
1565
1566                 cif = (struct carp_if *)ifp->if_carp;
1567                 CARP_LOCK(cif);
1568                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1569                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1570                                 CARP_UNLOCK(cif);
1571                                 error = EEXIST;
1572                                 goto cleanup;
1573                         }
1574         }
1575         sc->sc_ia = ia;
1576         sc->sc_carpdev = ifp;
1577
1578         { /* XXX prevent endless loop if already in queue */
1579         struct carp_softc *vr, *after = NULL;
1580         int myself = 0;
1581         cif = (struct carp_if *)ifp->if_carp;
1582
1583         /* XXX: cif should not change, right? So we still hold the lock */
1584         CARP_LOCK_ASSERT(cif);
1585
1586         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1587                 if (vr == sc)
1588                         myself = 1;
1589                 if (vr->sc_vhid < sc->sc_vhid)
1590                         after = vr;
1591         }
1592
1593         if (!myself) {
1594                 /* We're trying to keep things in order */
1595                 if (after == NULL) {
1596                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1597                 } else {
1598                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1599                 }
1600                 cif->vhif_nvrs++;
1601         }
1602         }
1603
1604         sc->sc_naddrs++;
1605         SC2IFP(sc)->if_flags |= IFF_UP;
1606         if (own)
1607                 sc->sc_advskew = 0;
1608         carp_sc_state_locked(sc);
1609         carp_setrun(sc, 0);
1610
1611         CARP_UNLOCK(cif);
1612         ifa_free(&ia->ia_ifa);  /* XXXRW: should hold reference for softc. */
1613
1614         return (0);
1615
1616 cleanup:
1617         in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1618         ifa_free(&ia->ia_ifa);
1619         return (error);
1620 }
1621
1622 static int
1623 carp_del_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1624 {
1625         int error = 0;
1626
1627         if (!--sc->sc_naddrs) {
1628                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1629                 struct ip_moptions *imo = &sc->sc_imo;
1630
1631                 CARP_LOCK(cif);
1632                 callout_stop(&sc->sc_ad_tmo);
1633                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1634                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1635                 sc->sc_vhid = -1;
1636                 in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1637                 imo->imo_multicast_ifp = NULL;
1638                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1639                 if (!--cif->vhif_nvrs) {
1640                         sc->sc_carpdev->if_carp = NULL;
1641                         CARP_LOCK_DESTROY(cif);
1642                         free(cif, M_CARP);
1643                 } else {
1644                         CARP_UNLOCK(cif);
1645                 }
1646         }
1647
1648         return (error);
1649 }
1650
1651 #ifdef INET6
1652 static int
1653 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1654 {
1655         struct ifnet *ifp;
1656         struct carp_if *cif;
1657         struct in6_ifaddr *ia, *ia_if;
1658         struct ip6_moptions *im6o = &sc->sc_im6o;
1659         struct in6_addr in6;
1660         int own, error;
1661
1662         error = 0;
1663
1664         if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
1665                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1666                         carp_set_state(sc, INIT);
1667                 if (sc->sc_naddrs6)
1668                         SC2IFP(sc)->if_flags |= IFF_UP;
1669                 if (sc->sc_carpdev)
1670                         CARP_SCLOCK(sc);
1671                 carp_setrun(sc, 0);
1672                 if (sc->sc_carpdev)
1673                         CARP_SCUNLOCK(sc);
1674                 return (0);
1675         }
1676
1677         /* we have to do it by hands to check we won't match on us */
1678         ia_if = NULL; own = 0;
1679         IN6_IFADDR_RLOCK();
1680         TAILQ_FOREACH(ia, &V_in6_ifaddrhead, ia_link) {
1681                 int i;
1682
1683                 for (i = 0; i < 4; i++) {
1684                         if ((sin6->sin6_addr.s6_addr32[i] &
1685                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]) !=
1686                             (ia->ia_addr.sin6_addr.s6_addr32[i] &
1687                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]))
1688                                 break;
1689                 }
1690                 /* and, yeah, we need a multicast-capable iface too */
1691                 if (ia->ia_ifp != SC2IFP(sc) &&
1692                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1693                     (i == 4)) {
1694                         if (!ia_if)
1695                                 ia_if = ia;
1696                         if (IN6_ARE_ADDR_EQUAL(&sin6->sin6_addr,
1697                             &ia->ia_addr.sin6_addr))
1698                                 own++;
1699                 }
1700         }
1701
1702         if (!ia_if) {
1703                 IN6_IFADDR_RUNLOCK();
1704                 return (EADDRNOTAVAIL);
1705         }
1706         ia = ia_if;
1707         ifa_ref(&ia->ia_ifa);
1708         IN6_IFADDR_RUNLOCK();
1709         ifp = ia->ia_ifp;
1710
1711         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1712             (im6o->im6o_multicast_ifp && im6o->im6o_multicast_ifp != ifp)) {
1713                 ifa_free(&ia->ia_ifa);
1714                 return (EADDRNOTAVAIL);
1715         }
1716
1717         if (!sc->sc_naddrs6) {
1718                 struct in6_multi *in6m;
1719
1720                 im6o->im6o_multicast_ifp = ifp;
1721
1722                 /* join CARP multicast address */
1723                 bzero(&in6, sizeof(in6));
1724                 in6.s6_addr16[0] = htons(0xff02);
1725                 in6.s6_addr8[15] = 0x12;
1726                 if (in6_setscope(&in6, ifp, NULL) != 0)
1727                         goto cleanup;
1728                 in6m = NULL;
1729                 error = in6_mc_join(ifp, &in6, NULL, &in6m, 0);
1730                 if (error)
1731                         goto cleanup;
1732                 im6o->im6o_membership[0] = in6m;
1733                 im6o->im6o_num_memberships++;
1734
1735                 /* join solicited multicast address */
1736                 bzero(&in6, sizeof(in6));
1737                 in6.s6_addr16[0] = htons(0xff02);
1738                 in6.s6_addr32[1] = 0;
1739                 in6.s6_addr32[2] = htonl(1);
1740                 in6.s6_addr32[3] = sin6->sin6_addr.s6_addr32[3];
1741                 in6.s6_addr8[12] = 0xff;
1742                 if (in6_setscope(&in6, ifp, NULL) != 0)
1743                         goto cleanup;
1744                 in6m = NULL;
1745                 error = in6_mc_join(ifp, &in6, NULL, &in6m, 0);
1746                 if (error)
1747                         goto cleanup;
1748                 im6o->im6o_membership[1] = in6m;
1749                 im6o->im6o_num_memberships++;
1750         }
1751
1752         if (!ifp->if_carp) {
1753                 cif = malloc(sizeof(*cif), M_CARP,
1754                     M_WAITOK|M_ZERO);
1755                 if (!cif) {
1756                         error = ENOBUFS;
1757                         goto cleanup;
1758                 }
1759                 if ((error = ifpromisc(ifp, 1))) {
1760                         free(cif, M_CARP);
1761                         goto cleanup;
1762                 }
1763
1764                 CARP_LOCK_INIT(cif);
1765                 CARP_LOCK(cif);
1766                 cif->vhif_ifp = ifp;
1767                 TAILQ_INIT(&cif->vhif_vrs);
1768                 ifp->if_carp = cif;
1769
1770         } else {
1771                 struct carp_softc *vr;
1772
1773                 cif = (struct carp_if *)ifp->if_carp;
1774                 CARP_LOCK(cif);
1775                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1776                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1777                                 CARP_UNLOCK(cif);
1778                                 error = EINVAL;
1779                                 goto cleanup;
1780                         }
1781         }
1782         sc->sc_ia6 = ia;
1783         sc->sc_carpdev = ifp;
1784
1785         { /* XXX prevent endless loop if already in queue */
1786         struct carp_softc *vr, *after = NULL;
1787         int myself = 0;
1788         cif = (struct carp_if *)ifp->if_carp;
1789         CARP_LOCK_ASSERT(cif);
1790
1791         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1792                 if (vr == sc)
1793                         myself = 1;
1794                 if (vr->sc_vhid < sc->sc_vhid)
1795                         after = vr;
1796         }
1797
1798         if (!myself) {
1799                 /* We're trying to keep things in order */
1800                 if (after == NULL) {
1801                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1802                 } else {
1803                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1804                 }
1805                 cif->vhif_nvrs++;
1806         }
1807         }
1808
1809         sc->sc_naddrs6++;
1810         SC2IFP(sc)->if_flags |= IFF_UP;
1811         if (own)
1812                 sc->sc_advskew = 0;
1813         carp_sc_state_locked(sc);
1814         carp_setrun(sc, 0);
1815
1816         CARP_UNLOCK(cif);
1817         ifa_free(&ia->ia_ifa);  /* XXXRW: should hold reference for softc. */
1818
1819         return (0);
1820
1821 cleanup:
1822         if (!sc->sc_naddrs6)
1823                 carp_multicast6_cleanup(sc);
1824         ifa_free(&ia->ia_ifa);
1825         return (error);
1826 }
1827
1828 static int
1829 carp_del_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1830 {
1831         int error = 0;
1832
1833         if (!--sc->sc_naddrs6) {
1834                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1835
1836                 CARP_LOCK(cif);
1837                 callout_stop(&sc->sc_ad_tmo);
1838                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1839                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1840                 sc->sc_vhid = -1;
1841                 carp_multicast6_cleanup(sc);
1842                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1843                 if (!--cif->vhif_nvrs) {
1844                         CARP_LOCK_DESTROY(cif);
1845                         sc->sc_carpdev->if_carp = NULL;
1846                         free(cif, M_CARP);
1847                 } else
1848                         CARP_UNLOCK(cif);
1849         }
1850
1851         return (error);
1852 }
1853 #endif /* INET6 */
1854
1855 static int
1856 carp_ioctl(struct ifnet *ifp, u_long cmd, caddr_t addr)
1857 {
1858         struct carp_softc *sc = ifp->if_softc, *vr;
1859         struct carpreq carpr;
1860         struct ifaddr *ifa;
1861         struct ifreq *ifr;
1862         struct ifaliasreq *ifra;
1863         int locked = 0, error = 0;
1864
1865         ifa = (struct ifaddr *)addr;
1866         ifra = (struct ifaliasreq *)addr;
1867         ifr = (struct ifreq *)addr;
1868
1869         switch (cmd) {
1870         case SIOCSIFADDR:
1871                 switch (ifa->ifa_addr->sa_family) {
1872 #ifdef INET
1873                 case AF_INET:
1874                         SC2IFP(sc)->if_flags |= IFF_UP;
1875                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1876                             sizeof(struct sockaddr));
1877                         error = carp_set_addr(sc, satosin(ifa->ifa_addr));
1878                         break;
1879 #endif /* INET */
1880 #ifdef INET6
1881                 case AF_INET6:
1882                         SC2IFP(sc)->if_flags |= IFF_UP;
1883                         error = carp_set_addr6(sc, satosin6(ifa->ifa_addr));
1884                         break;
1885 #endif /* INET6 */
1886                 default:
1887                         error = EAFNOSUPPORT;
1888                         break;
1889                 }
1890                 break;
1891
1892         case SIOCAIFADDR:
1893                 switch (ifa->ifa_addr->sa_family) {
1894 #ifdef INET
1895                 case AF_INET:
1896                         SC2IFP(sc)->if_flags |= IFF_UP;
1897                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1898                             sizeof(struct sockaddr));
1899                         error = carp_set_addr(sc, satosin(&ifra->ifra_addr));
1900                         break;
1901 #endif /* INET */
1902 #ifdef INET6
1903                 case AF_INET6:
1904                         SC2IFP(sc)->if_flags |= IFF_UP;
1905                         error = carp_set_addr6(sc, satosin6(&ifra->ifra_addr));
1906                         break;
1907 #endif /* INET6 */
1908                 default:
1909                         error = EAFNOSUPPORT;
1910                         break;
1911                 }
1912                 break;
1913
1914         case SIOCDIFADDR:
1915                 switch (ifa->ifa_addr->sa_family) {
1916 #ifdef INET
1917                 case AF_INET:
1918                         error = carp_del_addr(sc, satosin(&ifra->ifra_addr));
1919                         break;
1920 #endif /* INET */
1921 #ifdef INET6
1922                 case AF_INET6:
1923                         error = carp_del_addr6(sc, satosin6(&ifra->ifra_addr));
1924                         break;
1925 #endif /* INET6 */
1926                 default:
1927                         error = EAFNOSUPPORT;
1928                         break;
1929                 }
1930                 break;
1931
1932         case SIOCSIFFLAGS:
1933                 if (sc->sc_carpdev) {
1934                         locked = 1;
1935                         CARP_SCLOCK(sc);
1936                 }
1937                 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) {
1938                         callout_stop(&sc->sc_ad_tmo);
1939                         callout_stop(&sc->sc_md_tmo);
1940                         callout_stop(&sc->sc_md6_tmo);
1941                         if (sc->sc_state == MASTER)
1942                                 carp_send_ad_locked(sc);
1943                         carp_set_state(sc, INIT);
1944                         carp_setrun(sc, 0);
1945                 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) {
1946                         SC2IFP(sc)->if_flags |= IFF_UP;
1947                         carp_setrun(sc, 0);
1948                 }
1949                 break;
1950
1951         case SIOCSVH:
1952                 error = priv_check(curthread, PRIV_NETINET_CARP);
1953                 if (error)
1954                         break;
1955                 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1956                         break;
1957                 error = 1;
1958                 if (sc->sc_carpdev) {
1959                         locked = 1;
1960                         CARP_SCLOCK(sc);
1961                 }
1962                 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) {
1963                         switch (carpr.carpr_state) {
1964                         case BACKUP:
1965                                 callout_stop(&sc->sc_ad_tmo);
1966                                 carp_set_state(sc, BACKUP);
1967                                 carp_setrun(sc, 0);
1968                                 carp_setroute(sc, RTM_DELETE);
1969                                 break;
1970                         case MASTER:
1971                                 carp_master_down_locked(sc);
1972                                 break;
1973                         default:
1974                                 break;
1975                         }
1976                 }
1977                 if (carpr.carpr_vhid > 0) {
1978                         if (carpr.carpr_vhid > 255) {
1979                                 error = EINVAL;
1980                                 break;
1981                         }
1982                         if (sc->sc_carpdev) {
1983                                 struct carp_if *cif;
1984                                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1985                                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1986                                         if (vr != sc &&
1987                                             vr->sc_vhid == carpr.carpr_vhid) {
1988                                                 error = EEXIST;
1989                                                 break;
1990                                         }
1991                                 if (error == EEXIST)
1992                                         break;
1993                         }
1994                         sc->sc_vhid = carpr.carpr_vhid;
1995                         IF_LLADDR(sc->sc_ifp)[0] = 0;
1996                         IF_LLADDR(sc->sc_ifp)[1] = 0;
1997                         IF_LLADDR(sc->sc_ifp)[2] = 0x5e;
1998                         IF_LLADDR(sc->sc_ifp)[3] = 0;
1999                         IF_LLADDR(sc->sc_ifp)[4] = 1;
2000                         IF_LLADDR(sc->sc_ifp)[5] = sc->sc_vhid;
2001                         error--;
2002                 }
2003                 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) {
2004                         if (carpr.carpr_advskew >= 255) {
2005                                 error = EINVAL;
2006                                 break;
2007                         }
2008                         if (carpr.carpr_advbase > 255) {
2009                                 error = EINVAL;
2010                                 break;
2011                         }
2012                         sc->sc_advbase = carpr.carpr_advbase;
2013                         sc->sc_advskew = carpr.carpr_advskew;
2014                         error--;
2015                 }
2016                 bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
2017                 if (error > 0)
2018                         error = EINVAL;
2019                 else {
2020                         error = 0;
2021                         carp_setrun(sc, 0);
2022                 }
2023                 break;
2024
2025         case SIOCGVH:
2026                 /* XXX: lockless read */
2027                 bzero(&carpr, sizeof(carpr));
2028                 carpr.carpr_state = sc->sc_state;
2029                 carpr.carpr_vhid = sc->sc_vhid;
2030                 carpr.carpr_advbase = sc->sc_advbase;
2031                 carpr.carpr_advskew = sc->sc_advskew;
2032                 error = priv_check(curthread, PRIV_NETINET_CARP);
2033                 if (error == 0)
2034                         bcopy(sc->sc_key, carpr.carpr_key,
2035                             sizeof(carpr.carpr_key));
2036                 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
2037                 break;
2038
2039         default:
2040                 error = EINVAL;
2041         }
2042
2043         if (locked)
2044                 CARP_SCUNLOCK(sc);
2045
2046         carp_hmac_prepare(sc);
2047
2048         return (error);
2049 }
2050
2051 /*
2052  * XXX: this is looutput. We should eventually use it from there.
2053  */
2054 static int
2055 carp_looutput(struct ifnet *ifp, struct mbuf *m, struct sockaddr *dst,
2056     struct route *ro)
2057 {
2058         u_int32_t af;
2059         struct rtentry *rt = NULL;
2060
2061         M_ASSERTPKTHDR(m); /* check if we have the packet header */
2062
2063         if (ro != NULL)
2064                 rt = ro->ro_rt;
2065         if (rt && rt->rt_flags & (RTF_REJECT|RTF_BLACKHOLE)) {
2066                 m_freem(m);
2067                 return (rt->rt_flags & RTF_BLACKHOLE ? 0 :
2068                         rt->rt_flags & RTF_HOST ? EHOSTUNREACH : ENETUNREACH);
2069         }
2070
2071         ifp->if_opackets++;
2072         ifp->if_obytes += m->m_pkthdr.len;
2073
2074         /* BPF writes need to be handled specially. */
2075         if (dst->sa_family == AF_UNSPEC) {
2076                 bcopy(dst->sa_data, &af, sizeof(af));
2077                 dst->sa_family = af;
2078         }
2079
2080 #if 1   /* XXX */
2081         switch (dst->sa_family) {
2082         case AF_INET:
2083         case AF_INET6:
2084         case AF_IPX:
2085         case AF_APPLETALK:
2086                 break;
2087         default:
2088                 printf("carp_looutput: af=%d unexpected\n", dst->sa_family);
2089                 m_freem(m);
2090                 return (EAFNOSUPPORT);
2091         }
2092 #endif
2093         return(if_simloop(ifp, m, dst->sa_family, 0));
2094 }
2095
2096 /*
2097  * Start output on carp interface. This function should never be called.
2098  */
2099 static void
2100 carp_start(struct ifnet *ifp)
2101 {
2102 #ifdef DEBUG
2103         printf("%s: start called\n", ifp->if_xname);
2104 #endif
2105 }
2106
2107 int
2108 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *sa,
2109     struct rtentry *rt)
2110 {
2111         struct m_tag *mtag;
2112         struct carp_softc *sc;
2113         struct ifnet *carp_ifp;
2114
2115         if (!sa)
2116                 return (0);
2117
2118         switch (sa->sa_family) {
2119 #ifdef INET
2120         case AF_INET:
2121                 break;
2122 #endif /* INET */
2123 #ifdef INET6
2124         case AF_INET6:
2125                 break;
2126 #endif /* INET6 */
2127         default:
2128                 return (0);
2129         }
2130
2131         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
2132         if (mtag == NULL)
2133                 return (0);
2134
2135         bcopy(mtag + 1, &carp_ifp, sizeof(struct ifnet *));
2136         sc = carp_ifp->if_softc;
2137
2138         /* Set the source MAC address to Virtual Router MAC Address */
2139         switch (ifp->if_type) {
2140         case IFT_ETHER:
2141         case IFT_L2VLAN: {
2142                         struct ether_header *eh;
2143
2144                         eh = mtod(m, struct ether_header *);
2145                         eh->ether_shost[0] = 0;
2146                         eh->ether_shost[1] = 0;
2147                         eh->ether_shost[2] = 0x5e;
2148                         eh->ether_shost[3] = 0;
2149                         eh->ether_shost[4] = 1;
2150                         eh->ether_shost[5] = sc->sc_vhid;
2151                 }
2152                 break;
2153         case IFT_FDDI: {
2154                         struct fddi_header *fh;
2155
2156                         fh = mtod(m, struct fddi_header *);
2157                         fh->fddi_shost[0] = 0;
2158                         fh->fddi_shost[1] = 0;
2159                         fh->fddi_shost[2] = 0x5e;
2160                         fh->fddi_shost[3] = 0;
2161                         fh->fddi_shost[4] = 1;
2162                         fh->fddi_shost[5] = sc->sc_vhid;
2163                 }
2164                 break;
2165         case IFT_ISO88025: {
2166                         struct iso88025_header *th;
2167                         th = mtod(m, struct iso88025_header *);
2168                         th->iso88025_shost[0] = 3;
2169                         th->iso88025_shost[1] = 0;
2170                         th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
2171                         th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
2172                         th->iso88025_shost[4] = 0;
2173                         th->iso88025_shost[5] = 0;
2174                 }
2175                 break;
2176         default:
2177                 printf("%s: carp is not supported for this interface type\n",
2178                     ifp->if_xname);
2179                 return (EOPNOTSUPP);
2180         }
2181
2182         return (0);
2183 }
2184
2185 static void
2186 carp_set_state(struct carp_softc *sc, int state)
2187 {
2188         int link_state;
2189
2190         if (sc->sc_carpdev)
2191                 CARP_SCLOCK_ASSERT(sc);
2192
2193         if (sc->sc_state == state)
2194                 return;
2195
2196         sc->sc_state = state;
2197         switch (state) {
2198         case BACKUP:
2199                 link_state = LINK_STATE_DOWN;
2200                 break;
2201         case MASTER:
2202                 link_state = LINK_STATE_UP;
2203                 break;
2204         default:
2205                 link_state = LINK_STATE_UNKNOWN;
2206                 break;
2207         }
2208         if_link_state_change(SC2IFP(sc), link_state);
2209 }
2210
2211 void
2212 carp_carpdev_state(void *v)
2213 {
2214         struct carp_if *cif = v;
2215
2216         CARP_LOCK(cif);
2217         carp_carpdev_state_locked(cif);
2218         CARP_UNLOCK(cif);
2219 }
2220
2221 static void
2222 carp_carpdev_state_locked(struct carp_if *cif)
2223 {
2224         struct carp_softc *sc;
2225
2226         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list)
2227                 carp_sc_state_locked(sc);
2228 }
2229
2230 static void
2231 carp_sc_state_locked(struct carp_softc *sc)
2232 {
2233         CARP_SCLOCK_ASSERT(sc);
2234
2235         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
2236             !(sc->sc_carpdev->if_flags & IFF_UP)) {
2237                 sc->sc_flags_backup = SC2IFP(sc)->if_flags;
2238                 SC2IFP(sc)->if_flags &= ~IFF_UP;
2239                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
2240                 callout_stop(&sc->sc_ad_tmo);
2241                 callout_stop(&sc->sc_md_tmo);
2242                 callout_stop(&sc->sc_md6_tmo);
2243                 carp_set_state(sc, INIT);
2244                 carp_setrun(sc, 0);
2245                 if (!sc->sc_suppress) {
2246                         carp_suppress_preempt++;
2247                         if (carp_suppress_preempt == 1) {
2248                                 CARP_SCUNLOCK(sc);
2249                                 carp_send_ad_all();
2250                                 CARP_SCLOCK(sc);
2251                         }
2252                 }
2253                 sc->sc_suppress = 1;
2254         } else {
2255                 SC2IFP(sc)->if_flags |= sc->sc_flags_backup;
2256                 carp_set_state(sc, INIT);
2257                 carp_setrun(sc, 0);
2258                 if (sc->sc_suppress)
2259                         carp_suppress_preempt--;
2260                 sc->sc_suppress = 0;
2261         }
2262
2263         return;
2264 }
2265
2266 static int
2267 carp_modevent(module_t mod, int type, void *data)
2268 {
2269         switch (type) {
2270         case MOD_LOAD:
2271                 if_detach_event_tag = EVENTHANDLER_REGISTER(ifnet_departure_event,
2272                     carp_ifdetach, NULL, EVENTHANDLER_PRI_ANY);
2273                 if (if_detach_event_tag == NULL)
2274                         return (ENOMEM);
2275                 mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2276                 LIST_INIT(&carpif_list);
2277                 if_clone_attach(&carp_cloner);
2278                 break;
2279
2280         case MOD_UNLOAD:
2281                 EVENTHANDLER_DEREGISTER(ifnet_departure_event, if_detach_event_tag);
2282                 if_clone_detach(&carp_cloner);
2283                 mtx_destroy(&carp_mtx);
2284                 break;
2285
2286         default:
2287                 return (EINVAL);
2288         }
2289
2290         return (0);
2291 }
2292
2293 static moduledata_t carp_mod = {
2294         "carp",
2295         carp_modevent,
2296         0
2297 };
2298
2299 DECLARE_MODULE(carp, carp_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);