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
  68 #ifdef INET
  69 #include <netinet/in.h>
  70 #include <netinet/in_var.h>
  71 #include <netinet/in_systm.h>
  72 #include <netinet/ip.h>
  73 #include <netinet/ip_var.h>
  74 #include <netinet/if_ether.h>
  75 #include <machine/in_cksum.h>
  76 #endif
  77
  78 #ifdef INET6
  79 #include <netinet/icmp6.h>
  80 #include <netinet/ip6.h>
  81 #include <netinet6/ip6_var.h>
  82 #include <netinet6/scope6_var.h>
  83 #include <netinet6/nd6.h>
  84 #endif
  85
  86 #include <crypto/sha1.h>
  87 #include <netinet/ip_carp.h>
  88
  89 #define CARP_IFNAME     "carp"
  90 static MALLOC_DEFINE(M_CARP, "CARP", "CARP interfaces");
  91 SYSCTL_DECL(_net_inet_carp);
  92
  93 struct carp_softc {
  94         struct ifnet            *sc_ifp;        /* Interface clue */
  95         struct ifnet            *sc_carpdev;    /* Pointer to parent interface */
  96         struct in_ifaddr        *sc_ia;         /* primary iface address */
  97         struct ip_moptions       sc_imo;
  98 #ifdef INET6
  99         struct in6_ifaddr       *sc_ia6;        /* primary iface address v6 */
 100         struct ip6_moptions      sc_im6o;
 101 #endif /* INET6 */
 102         TAILQ_ENTRY(carp_softc)  sc_list;
 103
 104         enum { INIT = 0, BACKUP, MASTER }       sc_state;
 105
 106         int                      sc_flags_backup;
 107         int                      sc_suppress;
 108
 109         int                      sc_sendad_errors;
 110 #define CARP_SENDAD_MAX_ERRORS  3
 111         int                      sc_sendad_success;
 112 #define CARP_SENDAD_MIN_SUCCESS 3
 113
 114         int                      sc_vhid;
 115         int                      sc_advskew;
 116         int                      sc_naddrs;
 117         int                      sc_naddrs6;
 118         int                      sc_advbase;    /* seconds */
 119         int                      sc_init_counter;
 120         u_int64_t                sc_counter;
 121
 122         /* authentication */
 123 #define CARP_HMAC_PAD   64
 124         unsigned char sc_key[CARP_KEY_LEN];
 125         unsigned char sc_pad[CARP_HMAC_PAD];
 126         SHA1_CTX sc_sha1;
 127
 128         struct callout           sc_ad_tmo;     /* advertisement timeout */
 129         struct callout           sc_md_tmo;     /* master down timeout */
 130         struct callout           sc_md6_tmo;    /* master down timeout */
 131
 132         LIST_ENTRY(carp_softc)   sc_next;       /* Interface clue */
 133 };
 134 #define SC2IFP(sc)      ((sc)->sc_ifp)
 135
 136 int carp_suppress_preempt = 0;
 137 int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 1, 0, 0 };    /* XXX for now */
 138 SYSCTL_INT(_net_inet_carp, CARPCTL_ALLOW, allow, CTLFLAG_RW,
 139     &carp_opts[CARPCTL_ALLOW], 0, "Accept incoming CARP packets");
 140 SYSCTL_INT(_net_inet_carp, CARPCTL_PREEMPT, preempt, CTLFLAG_RW,
 141     &carp_opts[CARPCTL_PREEMPT], 0, "high-priority backup preemption mode");
 142 SYSCTL_INT(_net_inet_carp, CARPCTL_LOG, log, CTLFLAG_RW,
 143     &carp_opts[CARPCTL_LOG], 0, "log bad carp packets");
 144 SYSCTL_INT(_net_inet_carp, CARPCTL_ARPBALANCE, arpbalance, CTLFLAG_RW,
 145     &carp_opts[CARPCTL_ARPBALANCE], 0, "balance arp responses");
 146 SYSCTL_INT(_net_inet_carp, OID_AUTO, suppress_preempt, CTLFLAG_RD,
 147     &carp_suppress_preempt, 0, "Preemption is suppressed");
 148
 149 struct carpstats carpstats;
 150 SYSCTL_STRUCT(_net_inet_carp, CARPCTL_STATS, stats, CTLFLAG_RW,
 151     &carpstats, carpstats,
 152     "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 153
 154 struct carp_if {
 155         TAILQ_HEAD(, carp_softc) vhif_vrs;
 156         int vhif_nvrs;
 157
 158         struct ifnet    *vhif_ifp;
 159         struct mtx       vhif_mtx;
 160 };
 161
 162 /* Get carp_if from softc. Valid after carp_set_addr{,6}. */
 163 #define SC2CIF(sc)              ((struct carp_if *)(sc)->sc_carpdev->if_carp)
 164
 165 /* lock per carp_if queue */
 166 #define CARP_LOCK_INIT(cif)     mtx_init(&(cif)->vhif_mtx, "carp_if",   \
 167         NULL, MTX_DEF)
 168 #define CARP_LOCK_DESTROY(cif)  mtx_destroy(&(cif)->vhif_mtx)
 169 #define CARP_LOCK_ASSERT(cif)   mtx_assert(&(cif)->vhif_mtx, MA_OWNED)
 170 #define CARP_LOCK(cif)          mtx_lock(&(cif)->vhif_mtx)
 171 #define CARP_UNLOCK(cif)        mtx_unlock(&(cif)->vhif_mtx)
 172
 173 #define CARP_SCLOCK(sc)         mtx_lock(&SC2CIF(sc)->vhif_mtx)
 174 #define CARP_SCUNLOCK(sc)       mtx_unlock(&SC2CIF(sc)->vhif_mtx)
 175 #define CARP_SCLOCK_ASSERT(sc)  mtx_assert(&SC2CIF(sc)->vhif_mtx, MA_OWNED)
 176
 177 #define CARP_LOG(...)   do {                            \
 178         if (carp_opts[CARPCTL_LOG] > 0)                 \
 179                 log(LOG_INFO, __VA_ARGS__);             \
 180 } while (0)
 181
 182 #define CARP_DEBUG(...) do {                            \
 183         if (carp_opts[CARPCTL_LOG] > 1)                 \
 184                 log(LOG_DEBUG, __VA_ARGS__);            \
 185 } while (0)
 186
 187 static void     carp_hmac_prepare(struct carp_softc *);
 188 static void     carp_hmac_generate(struct carp_softc *, u_int32_t *,
 189                     unsigned char *);
 190 static int      carp_hmac_verify(struct carp_softc *, u_int32_t *,
 191                     unsigned char *);
 192 static void     carp_setroute(struct carp_softc *, int);
 193 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 194 static int      carp_clone_create(struct if_clone *, int, caddr_t);
 195 static void     carp_clone_destroy(struct ifnet *);
 196 static void     carpdetach(struct carp_softc *, int);
 197 static int      carp_prepare_ad(struct mbuf *, struct carp_softc *,
 198                     struct carp_header *);
 199 static void     carp_send_ad_all(void);
 200 static void     carp_send_ad(void *);
 201 static void     carp_send_ad_locked(struct carp_softc *);
 202 static void     carp_send_arp(struct carp_softc *);
 203 static void     carp_master_down(void *);
 204 static void     carp_master_down_locked(struct carp_softc *);
 205 static int      carp_ioctl(struct ifnet *, u_long, caddr_t);
 206 static int      carp_looutput(struct ifnet *, struct mbuf *, struct sockaddr *,
 207                     struct rtentry *);
 208 static void     carp_start(struct ifnet *);
 209 static void     carp_setrun(struct carp_softc *, sa_family_t);
 210 static void     carp_set_state(struct carp_softc *, int);
 211 static int      carp_addrcount(struct carp_if *, struct in_ifaddr *, int);
 212 enum    { CARP_COUNT_MASTER, CARP_COUNT_RUNNING };
 213
 214 static void     carp_multicast_cleanup(struct carp_softc *);
 215 static int      carp_set_addr(struct carp_softc *, struct sockaddr_in *);
 216 static int      carp_del_addr(struct carp_softc *, struct sockaddr_in *);
 217 static void     carp_carpdev_state_locked(struct carp_if *);
 218 static void     carp_sc_state_locked(struct carp_softc *);
 219 #ifdef INET6
 220 static void     carp_send_na(struct carp_softc *);
 221 static int      carp_set_addr6(struct carp_softc *, struct sockaddr_in6 *);
 222 static int      carp_del_addr6(struct carp_softc *, struct sockaddr_in6 *);
 223 static void     carp_multicast6_cleanup(struct carp_softc *);
 224 #endif
 225
 226 static LIST_HEAD(, carp_softc) carpif_list;
 227 static struct mtx carp_mtx;
 228 IFC_SIMPLE_DECLARE(carp, 0);
 229
 230 static eventhandler_tag if_detach_event_tag;
 231
 232 static __inline u_int16_t
 233 carp_cksum(struct mbuf *m, int len)
 234 {
 235         return (in_cksum(m, len));
 236 }
 237
 238 static void
 239 carp_hmac_prepare(struct carp_softc *sc)
 240 {
 241         u_int8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 242         u_int8_t vhid = sc->sc_vhid & 0xff;
 243         struct ifaddr *ifa;
 244         int i, found;
 245 #ifdef INET
 246         struct in_addr last, cur, in;
 247 #endif
 248 #ifdef INET6
 249         struct in6_addr last6, cur6, in6;
 250 #endif
 251
 252         if (sc->sc_carpdev)
 253                 CARP_SCLOCK(sc);
 254
 255         /* XXX: possible race here */
 256
 257         /* compute ipad from key */
 258         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 259         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 260         for (i = 0; i < sizeof(sc->sc_pad); i++)
 261                 sc->sc_pad[i] ^= 0x36;
 262
 263         /* precompute first part of inner hash */
 264         SHA1Init(&sc->sc_sha1);
 265         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 266         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 267         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 268         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 269 #ifdef INET
 270         cur.s_addr = 0;
 271         do {
 272                 found = 0;
 273                 last = cur;
 274                 cur.s_addr = 0xffffffff;
 275                 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 276                         in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 277                         if (ifa->ifa_addr->sa_family == AF_INET &&
 278                             ntohl(in.s_addr) > ntohl(last.s_addr) &&
 279                             ntohl(in.s_addr) < ntohl(cur.s_addr)) {
 280                                 cur.s_addr = in.s_addr;
 281                                 found++;
 282                         }
 283                 }
 284                 if (found)
 285                         SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur));
 286         } while (found);
 287 #endif /* INET */
 288 #ifdef INET6
 289         memset(&cur6, 0, sizeof(cur6));
 290         do {
 291                 found = 0;
 292                 last6 = cur6;
 293                 memset(&cur6, 0xff, sizeof(cur6));
 294                 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 295                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 296                         if (IN6_IS_SCOPE_EMBED(&in6))
 297                                 in6.s6_addr16[1] = 0;
 298                         if (ifa->ifa_addr->sa_family == AF_INET6 &&
 299                             memcmp(&in6, &last6, sizeof(in6)) > 0 &&
 300                             memcmp(&in6, &cur6, sizeof(in6)) < 0) {
 301                                 cur6 = in6;
 302                                 found++;
 303                         }
 304                 }
 305                 if (found)
 306                         SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6));
 307         } while (found);
 308 #endif /* INET6 */
 309
 310         /* convert ipad to opad */
 311         for (i = 0; i < sizeof(sc->sc_pad); i++)
 312                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 313
 314         if (sc->sc_carpdev)
 315                 CARP_SCUNLOCK(sc);
 316 }
 317
 318 static void
 319 carp_hmac_generate(struct carp_softc *sc, u_int32_t counter[2],
 320     unsigned char md[20])
 321 {
 322         SHA1_CTX sha1ctx;
 323
 324         /* fetch first half of inner hash */
 325         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 326
 327         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 328         SHA1Final(md, &sha1ctx);
 329
 330         /* outer hash */
 331         SHA1Init(&sha1ctx);
 332         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 333         SHA1Update(&sha1ctx, md, 20);
 334         SHA1Final(md, &sha1ctx);
 335 }
 336
 337 static int
 338 carp_hmac_verify(struct carp_softc *sc, u_int32_t counter[2],
 339     unsigned char md[20])
 340 {
 341         unsigned char md2[20];
 342
 343         CARP_SCLOCK_ASSERT(sc);
 344
 345         carp_hmac_generate(sc, counter, md2);
 346
 347         return (bcmp(md, md2, sizeof(md2)));
 348 }
 349
 350 static void
 351 carp_setroute(struct carp_softc *sc, int cmd)
 352 {
 353         struct ifaddr *ifa;
 354         int s;
 355
 356         if (sc->sc_carpdev)
 357                 CARP_SCLOCK_ASSERT(sc);
 358
 359         s = splnet();
 360         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 361                 if (ifa->ifa_addr->sa_family == AF_INET &&
 362                     sc->sc_carpdev != NULL) {
 363                         int count = carp_addrcount(
 364                             (struct carp_if *)sc->sc_carpdev->if_carp,
 365                             ifatoia(ifa), CARP_COUNT_MASTER);
 366
 367                         if ((cmd == RTM_ADD && count == 1) ||
 368                             (cmd == RTM_DELETE && count == 0))
 369                                 rtinit(ifa, cmd, RTF_UP | RTF_HOST);
 370                 }
 371 #ifdef INET6
 372                 if (ifa->ifa_addr->sa_family == AF_INET6) {
 373                         if (cmd == RTM_ADD)
 374                                 in6_ifaddloop(ifa);
 375                         else
 376                                 in6_ifremloop(ifa);
 377                 }
 378 #endif /* INET6 */
 379         }
 380         splx(s);
 381 }
 382
 383 static int
 384 carp_clone_create(struct if_clone *ifc, int unit, caddr_t params)
 385 {
 386
 387         struct carp_softc *sc;
 388         struct ifnet *ifp;
 389
 390         MALLOC(sc, struct carp_softc *, sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
 391         ifp = SC2IFP(sc) = if_alloc(IFT_ETHER);
 392         if (ifp == NULL) {
 393                 FREE(sc, M_CARP);
 394                 return (ENOSPC);
 395         }
 396
 397         sc->sc_flags_backup = 0;
 398         sc->sc_suppress = 0;
 399         sc->sc_advbase = CARP_DFLTINTV;
 400         sc->sc_vhid = -1;       /* required setting */
 401         sc->sc_advskew = 0;
 402         sc->sc_init_counter = 1;
 403         sc->sc_naddrs = sc->sc_naddrs6 = 0; /* M_ZERO? */
 404 #ifdef INET6
 405         sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL;
 406 #endif
 407         sc->sc_imo.imo_membership = (struct in_multi **)malloc(
 408             (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
 409             M_WAITOK);
 410         sc->sc_imo.imo_mfilters = NULL;
 411         sc->sc_imo.imo_max_memberships = IP_MIN_MEMBERSHIPS;
 412         sc->sc_imo.imo_multicast_vif = -1;
 413
 414         callout_init(&sc->sc_ad_tmo, CALLOUT_MPSAFE);
 415         callout_init(&sc->sc_md_tmo, CALLOUT_MPSAFE);
 416         callout_init(&sc->sc_md6_tmo, CALLOUT_MPSAFE);
 417
 418         ifp->if_softc = sc;
 419         if_initname(ifp, CARP_IFNAME, unit);
 420         ifp->if_mtu = ETHERMTU;
 421         ifp->if_flags = IFF_LOOPBACK;
 422         ifp->if_ioctl = carp_ioctl;
 423         ifp->if_output = carp_looutput;
 424         ifp->if_start = carp_start;
 425         ifp->if_type = IFT_CARP;
 426         ifp->if_snd.ifq_maxlen = ifqmaxlen;
 427         ifp->if_hdrlen = 0;
 428         if_attach(ifp);
 429         bpfattach(SC2IFP(sc), DLT_NULL, sizeof(u_int32_t));
 430         mtx_lock(&carp_mtx);
 431         LIST_INSERT_HEAD(&carpif_list, sc, sc_next);
 432         mtx_unlock(&carp_mtx);
 433         return (0);
 434 }
 435
 436 static void
 437 carp_clone_destroy(struct ifnet *ifp)
 438 {
 439         struct carp_softc *sc = ifp->if_softc;
 440
 441         if (sc->sc_carpdev)
 442                 CARP_SCLOCK(sc);
 443         carpdetach(sc, 1);      /* Returns unlocked. */
 444
 445         mtx_lock(&carp_mtx);
 446         LIST_REMOVE(sc, sc_next);
 447         mtx_unlock(&carp_mtx);
 448         bpfdetach(ifp);
 449         if_detach(ifp);
 450         if_free_type(ifp, IFT_ETHER);
 451         free(sc->sc_imo.imo_membership, M_CARP);
 452         free(sc, M_CARP);
 453 }
 454
 455 /*
 456  * This function can be called on CARP interface destroy path,
 457  * and in case of the removal of the underlying interface as
 458  * well. We differentiate these two cases. In the latter case
 459  * we do not cleanup our multicast memberships, since they
 460  * are already freed. Also, in the latter case we do not
 461  * release the lock on return, because the function will be
 462  * called once more, for another CARP instance on the same
 463  * interface.
 464  */
 465 static void
 466 carpdetach(struct carp_softc *sc, int unlock)
 467 {
 468         struct carp_if *cif;
 469
 470         callout_stop(&sc->sc_ad_tmo);
 471         callout_stop(&sc->sc_md_tmo);
 472         callout_stop(&sc->sc_md6_tmo);
 473
 474         if (sc->sc_suppress)
 475                 carp_suppress_preempt--;
 476         sc->sc_suppress = 0;
 477
 478         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS)
 479                 carp_suppress_preempt--;
 480         sc->sc_sendad_errors = 0;
 481
 482         carp_set_state(sc, INIT);
 483         SC2IFP(sc)->if_flags &= ~IFF_UP;
 484         carp_setrun(sc, 0);
 485         if (unlock)
 486                 carp_multicast_cleanup(sc);
 487 #ifdef INET6
 488         carp_multicast6_cleanup(sc);
 489 #endif
 490
 491         if (sc->sc_carpdev != NULL) {
 492                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
 493                 CARP_LOCK_ASSERT(cif);
 494                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
 495                 if (!--cif->vhif_nvrs) {
 496                         ifpromisc(sc->sc_carpdev, 0);
 497                         sc->sc_carpdev->if_carp = NULL;
 498                         CARP_LOCK_DESTROY(cif);
 499                         FREE(cif, M_IFADDR);
 500                 } else if (unlock)
 501                         CARP_UNLOCK(cif);
 502                 sc->sc_carpdev = NULL;
 503         }
 504 }
 505
 506 /* Detach an interface from the carp. */
 507 static void
 508 carp_ifdetach(void *arg __unused, struct ifnet *ifp)
 509 {
 510         struct carp_if *cif = (struct carp_if *)ifp->if_carp;
 511         struct carp_softc *sc, *nextsc;
 512
 513         if (cif == NULL)
 514                 return;
 515
 516         /*
 517          * XXX: At the end of for() cycle the lock will be destroyed.
 518          */
 519         CARP_LOCK(cif);
 520         for (sc = TAILQ_FIRST(&cif->vhif_vrs); sc; sc = nextsc) {
 521                 nextsc = TAILQ_NEXT(sc, sc_list);
 522                 carpdetach(sc, 0);
 523         }
 524 }
 525
 526 /*
 527  * process input packet.
 528  * we have rearranged checks order compared to the rfc,
 529  * but it seems more efficient this way or not possible otherwise.
 530  */
 531 void
 532 carp_input(struct mbuf *m, int hlen)
 533 {
 534         struct ip *ip = mtod(m, struct ip *);
 535         struct carp_header *ch;
 536         int iplen, len;
 537
 538         carpstats.carps_ipackets++;
 539
 540         if (!carp_opts[CARPCTL_ALLOW]) {
 541                 m_freem(m);
 542                 return;
 543         }
 544
 545         /* check if received on a valid carp interface */
 546         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 547                 carpstats.carps_badif++;
 548                 CARP_LOG("carp_input: packet received on non-carp "
 549                     "interface: %s\n",
 550                     m->m_pkthdr.rcvif->if_xname);
 551                 m_freem(m);
 552                 return;
 553         }
 554
 555         /* verify that the IP TTL is 255.  */
 556         if (ip->ip_ttl != CARP_DFLTTL) {
 557                 carpstats.carps_badttl++;
 558                 CARP_LOG("carp_input: received ttl %d != 255i on %s\n",
 559                     ip->ip_ttl,
 560                     m->m_pkthdr.rcvif->if_xname);
 561                 m_freem(m);
 562                 return;
 563         }
 564
 565         iplen = ip->ip_hl << 2;
 566
 567         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 568                 carpstats.carps_badlen++;
 569                 CARP_LOG("carp_input: received len %zd < "
 570                     "sizeof(struct carp_header)\n",
 571                     m->m_len - sizeof(struct ip));
 572                 m_freem(m);
 573                 return;
 574         }
 575
 576         if (iplen + sizeof(*ch) < m->m_len) {
 577                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 578                         carpstats.carps_hdrops++;
 579                         CARP_LOG("carp_input: pullup failed\n");
 580                         return;
 581                 }
 582                 ip = mtod(m, struct ip *);
 583         }
 584         ch = (struct carp_header *)((char *)ip + iplen);
 585
 586         /*
 587          * verify that the received packet length is
 588          * equal to the CARP header
 589          */
 590         len = iplen + sizeof(*ch);
 591         if (len > m->m_pkthdr.len) {
 592                 carpstats.carps_badlen++;
 593                 CARP_LOG("carp_input: packet too short %d on %s\n",
 594                     m->m_pkthdr.len,
 595                     m->m_pkthdr.rcvif->if_xname);
 596                 m_freem(m);
 597                 return;
 598         }
 599
 600         if ((m = m_pullup(m, len)) == NULL) {
 601                 carpstats.carps_hdrops++;
 602                 return;
 603         }
 604         ip = mtod(m, struct ip *);
 605         ch = (struct carp_header *)((char *)ip + iplen);
 606
 607         /* verify the CARP checksum */
 608         m->m_data += iplen;
 609         if (carp_cksum(m, len - iplen)) {
 610                 carpstats.carps_badsum++;
 611                 CARP_LOG("carp_input: checksum failed on %s\n",
 612                     m->m_pkthdr.rcvif->if_xname);
 613                 m_freem(m);
 614                 return;
 615         }
 616         m->m_data -= iplen;
 617
 618         carp_input_c(m, ch, AF_INET);
 619 }
 620
 621 #ifdef INET6
 622 int
 623 carp6_input(struct mbuf **mp, int *offp, int proto)
 624 {
 625         struct mbuf *m = *mp;
 626         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 627         struct carp_header *ch;
 628         u_int len;
 629
 630         carpstats.carps_ipackets6++;
 631
 632         if (!carp_opts[CARPCTL_ALLOW]) {
 633                 m_freem(m);
 634                 return (IPPROTO_DONE);
 635         }
 636
 637         /* check if received on a valid carp interface */
 638         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 639                 carpstats.carps_badif++;
 640                 CARP_LOG("carp6_input: packet received on non-carp "
 641                     "interface: %s\n",
 642                     m->m_pkthdr.rcvif->if_xname);
 643                 m_freem(m);
 644                 return (IPPROTO_DONE);
 645         }
 646
 647         /* verify that the IP TTL is 255 */
 648         if (ip6->ip6_hlim != CARP_DFLTTL) {
 649                 carpstats.carps_badttl++;
 650                 CARP_LOG("carp6_input: received ttl %d != 255 on %s\n",
 651                     ip6->ip6_hlim,
 652                     m->m_pkthdr.rcvif->if_xname);
 653                 m_freem(m);
 654                 return (IPPROTO_DONE);
 655         }
 656
 657         /* verify that we have a complete carp packet */
 658         len = m->m_len;
 659         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 660         if (ch == NULL) {
 661                 carpstats.carps_badlen++;
 662                 CARP_LOG("carp6_input: packet size %u too small\n", len);
 663                 return (IPPROTO_DONE);
 664         }
 665
 666
 667         /* verify the CARP checksum */
 668         m->m_data += *offp;
 669         if (carp_cksum(m, sizeof(*ch))) {
 670                 carpstats.carps_badsum++;
 671                 CARP_LOG("carp6_input: checksum failed, on %s\n",
 672                     m->m_pkthdr.rcvif->if_xname);
 673                 m_freem(m);
 674                 return (IPPROTO_DONE);
 675         }
 676         m->m_data -= *offp;
 677
 678         carp_input_c(m, ch, AF_INET6);
 679         return (IPPROTO_DONE);
 680 }
 681 #endif /* INET6 */
 682
 683 static void
 684 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 685 {
 686         struct ifnet *ifp = m->m_pkthdr.rcvif;
 687         struct carp_softc *sc;
 688         u_int64_t tmp_counter;
 689         struct timeval sc_tv, ch_tv;
 690
 691         /* verify that the VHID is valid on the receiving interface */
 692         CARP_LOCK(ifp->if_carp);
 693         TAILQ_FOREACH(sc, &((struct carp_if *)ifp->if_carp)->vhif_vrs, sc_list)
 694                 if (sc->sc_vhid == ch->carp_vhid)
 695                         break;
 696
 697         if (!sc || !((SC2IFP(sc)->if_flags & IFF_UP) &&
 698             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING))) {
 699                 carpstats.carps_badvhid++;
 700                 CARP_UNLOCK(ifp->if_carp);
 701                 m_freem(m);
 702                 return;
 703         }
 704
 705         getmicrotime(&SC2IFP(sc)->if_lastchange);
 706         SC2IFP(sc)->if_ipackets++;
 707         SC2IFP(sc)->if_ibytes += m->m_pkthdr.len;
 708
 709         if (bpf_peers_present(SC2IFP(sc)->if_bpf)) {
 710                 struct ip *ip = mtod(m, struct ip *);
 711                 uint32_t af1 = af;
 712
 713                 /* BPF wants net byte order */
 714                 ip->ip_len = htons(ip->ip_len + (ip->ip_hl << 2));
 715                 ip->ip_off = htons(ip->ip_off);
 716                 bpf_mtap2(SC2IFP(sc)->if_bpf, &af1, sizeof(af1), m);
 717         }
 718
 719         /* verify the CARP version. */
 720         if (ch->carp_version != CARP_VERSION) {
 721                 carpstats.carps_badver++;
 722                 SC2IFP(sc)->if_ierrors++;
 723                 CARP_UNLOCK(ifp->if_carp);
 724                 CARP_LOG("%s; invalid version %d\n",
 725                     SC2IFP(sc)->if_xname,
 726                     ch->carp_version);
 727                 m_freem(m);
 728                 return;
 729         }
 730
 731         /* verify the hash */
 732         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 733                 carpstats.carps_badauth++;
 734                 SC2IFP(sc)->if_ierrors++;
 735                 CARP_UNLOCK(ifp->if_carp);
 736                 CARP_LOG("%s: incorrect hash\n", SC2IFP(sc)->if_xname);
 737                 m_freem(m);
 738                 return;
 739         }
 740
 741         tmp_counter = ntohl(ch->carp_counter[0]);
 742         tmp_counter = tmp_counter<<32;
 743         tmp_counter += ntohl(ch->carp_counter[1]);
 744
 745         /* XXX Replay protection goes here */
 746
 747         sc->sc_init_counter = 0;
 748         sc->sc_counter = tmp_counter;
 749
 750         sc_tv.tv_sec = sc->sc_advbase;
 751         if (carp_suppress_preempt && sc->sc_advskew <  240)
 752                 sc_tv.tv_usec = 240 * 1000000 / 256;
 753         else
 754                 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256;
 755         ch_tv.tv_sec = ch->carp_advbase;
 756         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 757
 758         switch (sc->sc_state) {
 759         case INIT:
 760                 break;
 761         case MASTER:
 762                 /*
 763                  * If we receive an advertisement from a master who's going to
 764                  * be more frequent than us, go into BACKUP state.
 765                  */
 766                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 767                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 768                         callout_stop(&sc->sc_ad_tmo);
 769                         CARP_DEBUG("%s: MASTER -> BACKUP "
 770                            "(more frequent advertisement received)\n",
 771                            SC2IFP(sc)->if_xname);
 772                         carp_set_state(sc, BACKUP);
 773                         carp_setrun(sc, 0);
 774                         carp_setroute(sc, RTM_DELETE);
 775                 }
 776                 break;
 777         case BACKUP:
 778                 /*
 779                  * If we're pre-empting masters who advertise slower than us,
 780                  * and this one claims to be slower, treat him as down.
 781                  */
 782                 if (carp_opts[CARPCTL_PREEMPT] &&
 783                     timevalcmp(&sc_tv, &ch_tv, <)) {
 784                         CARP_DEBUG("%s: BACKUP -> MASTER "
 785                             "(preempting a slower master)\n",
 786                             SC2IFP(sc)->if_xname);
 787                         carp_master_down_locked(sc);
 788                         break;
 789                 }
 790
 791                 /*
 792                  *  If the master is going to advertise at such a low frequency
 793                  *  that he's guaranteed to time out, we'd might as well just
 794                  *  treat him as timed out now.
 795                  */
 796                 sc_tv.tv_sec = sc->sc_advbase * 3;
 797                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 798                         CARP_DEBUG("%s: BACKUP -> MASTER "
 799                             "(master timed out)\n",
 800                             SC2IFP(sc)->if_xname);
 801                         carp_master_down_locked(sc);
 802                         break;
 803                 }
 804
 805                 /*
 806                  * Otherwise, we reset the counter and wait for the next
 807                  * advertisement.
 808                  */
 809                 carp_setrun(sc, af);
 810                 break;
 811         }
 812
 813         CARP_UNLOCK(ifp->if_carp);
 814
 815         m_freem(m);
 816         return;
 817 }
 818
 819 static int
 820 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 821 {
 822         struct m_tag *mtag;
 823         struct ifnet *ifp = SC2IFP(sc);
 824
 825         if (sc->sc_init_counter) {
 826                 /* this could also be seconds since unix epoch */
 827                 sc->sc_counter = arc4random();
 828                 sc->sc_counter = sc->sc_counter << 32;
 829                 sc->sc_counter += arc4random();
 830         } else
 831                 sc->sc_counter++;
 832
 833         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 834         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 835
 836         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 837
 838         /* Tag packet for carp_output */
 839         mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct ifnet *), M_NOWAIT);
 840         if (mtag == NULL) {
 841                 m_freem(m);
 842                 SC2IFP(sc)->if_oerrors++;
 843                 return (ENOMEM);
 844         }
 845         bcopy(&ifp, (caddr_t)(mtag + 1), sizeof(struct ifnet *));
 846         m_tag_prepend(m, mtag);
 847
 848         return (0);
 849 }
 850
 851 static void
 852 carp_send_ad_all(void)
 853 {
 854         struct carp_softc *sc;
 855
 856         mtx_lock(&carp_mtx);
 857         LIST_FOREACH(sc, &carpif_list, sc_next) {
 858                 if (sc->sc_carpdev == NULL)
 859                         continue;
 860                 CARP_SCLOCK(sc);
 861                 if ((SC2IFP(sc)->if_flags & IFF_UP) &&
 862                     (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING) &&
 863                      sc->sc_state == MASTER)
 864                         carp_send_ad_locked(sc);
 865                 CARP_SCUNLOCK(sc);
 866         }
 867         mtx_unlock(&carp_mtx);
 868 }
 869
 870 static void
 871 carp_send_ad(void *v)
 872 {
 873         struct carp_softc *sc = v;
 874
 875         CARP_SCLOCK(sc);
 876         carp_send_ad_locked(sc);
 877         CARP_SCUNLOCK(sc);
 878 }
 879
 880 static void
 881 carp_send_ad_locked(struct carp_softc *sc)
 882 {
 883         struct carp_header ch;
 884         struct timeval tv;
 885         struct carp_header *ch_ptr;
 886         struct mbuf *m;
 887         int len, advbase, advskew;
 888
 889         CARP_SCLOCK_ASSERT(sc);
 890
 891         /* bow out if we've lost our UPness or RUNNINGuiness */
 892         if (!((SC2IFP(sc)->if_flags & IFF_UP) &&
 893             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING))) {
 894                 advbase = 255;
 895                 advskew = 255;
 896         } else {
 897                 advbase = sc->sc_advbase;
 898                 if (!carp_suppress_preempt || sc->sc_advskew > 240)
 899                         advskew = sc->sc_advskew;
 900                 else
 901                         advskew = 240;
 902                 tv.tv_sec = advbase;
 903                 tv.tv_usec = advskew * 1000000 / 256;
 904         }
 905
 906         ch.carp_version = CARP_VERSION;
 907         ch.carp_type = CARP_ADVERTISEMENT;
 908         ch.carp_vhid = sc->sc_vhid;
 909         ch.carp_advbase = advbase;
 910         ch.carp_advskew = advskew;
 911         ch.carp_authlen = 7;    /* XXX DEFINE */
 912         ch.carp_pad1 = 0;       /* must be zero */
 913         ch.carp_cksum = 0;
 914
 915 #ifdef INET
 916         if (sc->sc_ia) {
 917                 struct ip *ip;
 918
 919                 MGETHDR(m, M_DONTWAIT, MT_HEADER);
 920                 if (m == NULL) {
 921                         SC2IFP(sc)->if_oerrors++;
 922                         carpstats.carps_onomem++;
 923                         /* XXX maybe less ? */
 924                         if (advbase != 255 || advskew != 255)
 925                                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
 926                                     carp_send_ad, sc);
 927                         return;
 928                 }
 929                 len = sizeof(*ip) + sizeof(ch);
 930                 m->m_pkthdr.len = len;
 931                 m->m_pkthdr.rcvif = NULL;
 932                 m->m_len = len;
 933                 MH_ALIGN(m, m->m_len);
 934                 m->m_flags |= M_MCAST;
 935                 ip = mtod(m, struct ip *);
 936                 ip->ip_v = IPVERSION;
 937                 ip->ip_hl = sizeof(*ip) >> 2;
 938                 ip->ip_tos = IPTOS_LOWDELAY;
 939                 ip->ip_len = len;
 940                 ip->ip_id = ip_newid();
 941                 ip->ip_off = IP_DF;
 942                 ip->ip_ttl = CARP_DFLTTL;
 943                 ip->ip_p = IPPROTO_CARP;
 944                 ip->ip_sum = 0;
 945                 ip->ip_src.s_addr = sc->sc_ia->ia_addr.sin_addr.s_addr;
 946                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 947
 948                 ch_ptr = (struct carp_header *)(&ip[1]);
 949                 bcopy(&ch, ch_ptr, sizeof(ch));
 950                 if (carp_prepare_ad(m, sc, ch_ptr))
 951                         return;
 952
 953                 m->m_data += sizeof(*ip);
 954                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip));
 955                 m->m_data -= sizeof(*ip);
 956
 957                 getmicrotime(&SC2IFP(sc)->if_lastchange);
 958                 SC2IFP(sc)->if_opackets++;
 959                 SC2IFP(sc)->if_obytes += len;
 960                 carpstats.carps_opackets++;
 961
 962                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, NULL)) {
 963                         SC2IFP(sc)->if_oerrors++;
 964                         if (sc->sc_sendad_errors < INT_MAX)
 965                                 sc->sc_sendad_errors++;
 966                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 967                                 carp_suppress_preempt++;
 968                                 if (carp_suppress_preempt == 1) {
 969                                         CARP_SCUNLOCK(sc);
 970                                         carp_send_ad_all();
 971                                         CARP_SCLOCK(sc);
 972                                 }
 973                         }
 974                         sc->sc_sendad_success = 0;
 975                 } else {
 976                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 977                                 if (++sc->sc_sendad_success >=
 978                                     CARP_SENDAD_MIN_SUCCESS) {
 979                                         carp_suppress_preempt--;
 980                                         sc->sc_sendad_errors = 0;
 981                                 }
 982                         } else
 983                                 sc->sc_sendad_errors = 0;
 984                 }
 985         }
 986 #endif /* INET */
 987 #ifdef INET6
 988         if (sc->sc_ia6) {
 989                 struct ip6_hdr *ip6;
 990
 991                 MGETHDR(m, M_DONTWAIT, MT_HEADER);
 992                 if (m == NULL) {
 993                         SC2IFP(sc)->if_oerrors++;
 994                         carpstats.carps_onomem++;
 995                         /* XXX maybe less ? */
 996                         if (advbase != 255 || advskew != 255)
 997                                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
 998                                     carp_send_ad, sc);
 999                         return;
1000                 }
1001                 len = sizeof(*ip6) + sizeof(ch);
1002                 m->m_pkthdr.len = len;
1003                 m->m_pkthdr.rcvif = NULL;
1004                 m->m_len = len;
1005                 MH_ALIGN(m, m->m_len);
1006                 m->m_flags |= M_MCAST;
1007                 ip6 = mtod(m, struct ip6_hdr *);
1008                 bzero(ip6, sizeof(*ip6));
1009                 ip6->ip6_vfc |= IPV6_VERSION;
1010                 ip6->ip6_hlim = CARP_DFLTTL;
1011                 ip6->ip6_nxt = IPPROTO_CARP;
1012                 bcopy(&sc->sc_ia6->ia_addr.sin6_addr, &ip6->ip6_src,
1013                     sizeof(struct in6_addr));
1014                 /* set the multicast destination */
1015
1016                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
1017                 ip6->ip6_dst.s6_addr8[15] = 0x12;
1018                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
1019                         SC2IFP(sc)->if_oerrors++;
1020                         m_freem(m);
1021                         CARP_LOG("%s: in6_setscope failed\n", __func__);
1022                         return;
1023                 }
1024
1025                 ch_ptr = (struct carp_header *)(&ip6[1]);
1026                 bcopy(&ch, ch_ptr, sizeof(ch));
1027                 if (carp_prepare_ad(m, sc, ch_ptr))
1028                         return;
1029
1030                 m->m_data += sizeof(*ip6);
1031                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6));
1032                 m->m_data -= sizeof(*ip6);
1033
1034                 getmicrotime(&SC2IFP(sc)->if_lastchange);
1035                 SC2IFP(sc)->if_opackets++;
1036                 SC2IFP(sc)->if_obytes += len;
1037                 carpstats.carps_opackets6++;
1038
1039                 if (ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL)) {
1040                         SC2IFP(sc)->if_oerrors++;
1041                         if (sc->sc_sendad_errors < INT_MAX)
1042                                 sc->sc_sendad_errors++;
1043                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
1044                                 carp_suppress_preempt++;
1045                                 if (carp_suppress_preempt == 1) {
1046                                         CARP_SCUNLOCK(sc);
1047                                         carp_send_ad_all();
1048                                         CARP_SCLOCK(sc);
1049                                 }
1050                         }
1051                         sc->sc_sendad_success = 0;
1052                 } else {
1053                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
1054                                 if (++sc->sc_sendad_success >=
1055                                     CARP_SENDAD_MIN_SUCCESS) {
1056                                         carp_suppress_preempt--;
1057                                         sc->sc_sendad_errors = 0;
1058                                 }
1059                         } else
1060                                 sc->sc_sendad_errors = 0;
1061                 }
1062         }
1063 #endif /* INET6 */
1064
1065         if (advbase != 255 || advskew != 255)
1066                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1067                     carp_send_ad, sc);
1068
1069 }
1070
1071 /*
1072  * Broadcast a gratuitous ARP request containing
1073  * the virtual router MAC address for each IP address
1074  * associated with the virtual router.
1075  */
1076 static void
1077 carp_send_arp(struct carp_softc *sc)
1078 {
1079         struct ifaddr *ifa;
1080
1081         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1082
1083                 if (ifa->ifa_addr->sa_family != AF_INET)
1084                         continue;
1085
1086 /*              arprequest(sc->sc_carpdev, &in, &in, IF_LLADDR(sc->sc_ifp)); */
1087                 arp_ifinit2(sc->sc_carpdev, ifa, IF_LLADDR(sc->sc_ifp));
1088
1089                 DELAY(1000);    /* XXX */
1090         }
1091 }
1092
1093 #ifdef INET6
1094 static void
1095 carp_send_na(struct carp_softc *sc)
1096 {
1097         struct ifaddr *ifa;
1098         struct in6_addr *in6;
1099         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1100
1101         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1102
1103                 if (ifa->ifa_addr->sa_family != AF_INET6)
1104                         continue;
1105
1106                 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr;
1107                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1108                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1109                 DELAY(1000);    /* XXX */
1110         }
1111 }
1112 #endif /* INET6 */
1113
1114 static int
1115 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type)
1116 {
1117         struct carp_softc *vh;
1118         struct ifaddr *ifa;
1119         int count = 0;
1120
1121         CARP_LOCK_ASSERT(cif);
1122
1123         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1124                 if ((type == CARP_COUNT_RUNNING &&
1125                     (SC2IFP(vh)->if_flags & IFF_UP) &&
1126                     (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING)) ||
1127                     (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) {
1128                         TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1129                             ifa_list) {
1130                                 if (ifa->ifa_addr->sa_family == AF_INET &&
1131                                     ia->ia_addr.sin_addr.s_addr ==
1132                                     ifatoia(ifa)->ia_addr.sin_addr.s_addr)
1133                                         count++;
1134                         }
1135                 }
1136         }
1137         return (count);
1138 }
1139
1140 int
1141 carp_iamatch(void *v, struct in_ifaddr *ia,
1142     struct in_addr *isaddr, u_int8_t **enaddr)
1143 {
1144         struct carp_if *cif = v;
1145         struct carp_softc *vh;
1146         int index, count = 0;
1147         struct ifaddr *ifa;
1148
1149         CARP_LOCK(cif);
1150
1151         if (carp_opts[CARPCTL_ARPBALANCE]) {
1152                 /*
1153                  * XXX proof of concept implementation.
1154                  * We use the source ip to decide which virtual host should
1155                  * handle the request. If we're master of that virtual host,
1156                  * then we respond, otherwise, just drop the arp packet on
1157                  * the floor.
1158                  */
1159                 count = carp_addrcount(cif, ia, CARP_COUNT_RUNNING);
1160                 if (count == 0) {
1161                         /* should never reach this */
1162                         CARP_UNLOCK(cif);
1163                         return (0);
1164                 }
1165
1166                 /* this should be a hash, like pf_hash() */
1167                 index = ntohl(isaddr->s_addr) % count;
1168                 count = 0;
1169
1170                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1171                         if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1172                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING)) {
1173                                 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1174                                     ifa_list) {
1175                                         if (ifa->ifa_addr->sa_family ==
1176                                             AF_INET &&
1177                                             ia->ia_addr.sin_addr.s_addr ==
1178                                             ifatoia(ifa)->ia_addr.sin_addr.s_addr) {
1179                                                 if (count == index) {
1180                                                         if (vh->sc_state ==
1181                                                             MASTER) {
1182                                                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1183                                                                 CARP_UNLOCK(cif);
1184                                                                 return (1);
1185                                                         } else {
1186                                                                 CARP_UNLOCK(cif);
1187                                                                 return (0);
1188                                                         }
1189                                                 }
1190                                                 count++;
1191                                         }
1192                                 }
1193                         }
1194                 }
1195         } else {
1196                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1197                         if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1198                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1199                             ia->ia_ifp == SC2IFP(vh) &&
1200                             vh->sc_state == MASTER) {
1201                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1202                                 CARP_UNLOCK(cif);
1203                                 return (1);
1204                         }
1205                 }
1206         }
1207         CARP_UNLOCK(cif);
1208         return (0);
1209 }
1210
1211 #ifdef INET6
1212 struct ifaddr *
1213 carp_iamatch6(void *v, struct in6_addr *taddr)
1214 {
1215         struct carp_if *cif = v;
1216         struct carp_softc *vh;
1217         struct ifaddr *ifa;
1218
1219         CARP_LOCK(cif);
1220         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1221                 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist, ifa_list) {
1222                         if (IN6_ARE_ADDR_EQUAL(taddr,
1223                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1224                             (SC2IFP(vh)->if_flags & IFF_UP) &&
1225                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1226                             vh->sc_state == MASTER) {
1227                                 CARP_UNLOCK(cif);
1228                                 return (ifa);
1229                         }
1230                 }
1231         }
1232         CARP_UNLOCK(cif);
1233
1234         return (NULL);
1235 }
1236
1237 void *
1238 carp_macmatch6(void *v, struct mbuf *m, const struct in6_addr *taddr)
1239 {
1240         struct m_tag *mtag;
1241         struct carp_if *cif = v;
1242         struct carp_softc *sc;
1243         struct ifaddr *ifa;
1244
1245         CARP_LOCK(cif);
1246         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) {
1247                 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1248                         if (IN6_ARE_ADDR_EQUAL(taddr,
1249                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1250                             (SC2IFP(sc)->if_flags & IFF_UP) &&
1251                             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING)) {
1252                                 struct ifnet *ifp = SC2IFP(sc);
1253                                 mtag = m_tag_get(PACKET_TAG_CARP,
1254                                     sizeof(struct ifnet *), M_NOWAIT);
1255                                 if (mtag == NULL) {
1256                                         /* better a bit than nothing */
1257                                         CARP_UNLOCK(cif);
1258                                         return (IF_LLADDR(sc->sc_ifp));
1259                                 }
1260                                 bcopy(&ifp, (caddr_t)(mtag + 1),
1261                                     sizeof(struct ifnet *));
1262                                 m_tag_prepend(m, mtag);
1263
1264                                 CARP_UNLOCK(cif);
1265                                 return (IF_LLADDR(sc->sc_ifp));
1266                         }
1267                 }
1268         }
1269         CARP_UNLOCK(cif);
1270
1271         return (NULL);
1272 }
1273 #endif
1274
1275 struct ifnet *
1276 carp_forus(void *v, void *dhost)
1277 {
1278         struct carp_if *cif = v;
1279         struct carp_softc *vh;
1280         u_int8_t *ena = dhost;
1281
1282         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1283                 return (NULL);
1284
1285         CARP_LOCK(cif);
1286         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list)
1287                 if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1288                     (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1289                     vh->sc_state == MASTER &&
1290                     !bcmp(dhost, IF_LLADDR(vh->sc_ifp), ETHER_ADDR_LEN)) {
1291                         CARP_UNLOCK(cif);
1292                         return (SC2IFP(vh));
1293                 }
1294
1295         CARP_UNLOCK(cif);
1296         return (NULL);
1297 }
1298
1299 static void
1300 carp_master_down(void *v)
1301 {
1302         struct carp_softc *sc = v;
1303
1304         CARP_SCLOCK(sc);
1305         carp_master_down_locked(sc);
1306         CARP_SCUNLOCK(sc);
1307 }
1308
1309 static void
1310 carp_master_down_locked(struct carp_softc *sc)
1311 {
1312         if (sc->sc_carpdev)
1313                 CARP_SCLOCK_ASSERT(sc);
1314
1315         switch (sc->sc_state) {
1316         case INIT:
1317                 printf("%s: master_down event in INIT state\n",
1318                     SC2IFP(sc)->if_xname);
1319                 break;
1320         case MASTER:
1321                 break;
1322         case BACKUP:
1323                 carp_set_state(sc, MASTER);
1324                 carp_send_ad_locked(sc);
1325                 carp_send_arp(sc);
1326 #ifdef INET6
1327                 carp_send_na(sc);
1328 #endif /* INET6 */
1329                 carp_setrun(sc, 0);
1330                 carp_setroute(sc, RTM_ADD);
1331                 break;
1332         }
1333 }
1334
1335 /*
1336  * When in backup state, af indicates whether to reset the master down timer
1337  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1338  */
1339 static void
1340 carp_setrun(struct carp_softc *sc, sa_family_t af)
1341 {
1342         struct timeval tv;
1343
1344         if (sc->sc_carpdev == NULL) {
1345                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1346                 carp_set_state(sc, INIT);
1347                 return;
1348         } else
1349                 CARP_SCLOCK_ASSERT(sc);
1350
1351         if (SC2IFP(sc)->if_flags & IFF_UP &&
1352             sc->sc_vhid > 0 && (sc->sc_naddrs || sc->sc_naddrs6))
1353                 SC2IFP(sc)->if_drv_flags |= IFF_DRV_RUNNING;
1354         else {
1355                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1356                 carp_setroute(sc, RTM_DELETE);
1357                 return;
1358         }
1359
1360         switch (sc->sc_state) {
1361         case INIT:
1362                 if (carp_opts[CARPCTL_PREEMPT] && !carp_suppress_preempt) {
1363                         carp_send_ad_locked(sc);
1364                         carp_send_arp(sc);
1365 #ifdef INET6
1366                         carp_send_na(sc);
1367 #endif /* INET6 */
1368                         CARP_DEBUG("%s: INIT -> MASTER (preempting)\n",
1369                             SC2IFP(sc)->if_xname);
1370                         carp_set_state(sc, MASTER);
1371                         carp_setroute(sc, RTM_ADD);
1372                 } else {
1373                         CARP_DEBUG("%s: INIT -> BACKUP\n", SC2IFP(sc)->if_xname);
1374                         carp_set_state(sc, BACKUP);
1375                         carp_setroute(sc, RTM_DELETE);
1376                         carp_setrun(sc, 0);
1377                 }
1378                 break;
1379         case BACKUP:
1380                 callout_stop(&sc->sc_ad_tmo);
1381                 tv.tv_sec = 3 * sc->sc_advbase;
1382                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1383                 switch (af) {
1384 #ifdef INET
1385                 case AF_INET:
1386                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1387                             carp_master_down, sc);
1388                         break;
1389 #endif /* INET */
1390 #ifdef INET6
1391                 case AF_INET6:
1392                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1393                             carp_master_down, sc);
1394                         break;
1395 #endif /* INET6 */
1396                 default:
1397                         if (sc->sc_naddrs)
1398                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1399                                     carp_master_down, sc);
1400                         if (sc->sc_naddrs6)
1401                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1402                                     carp_master_down, sc);
1403                         break;
1404                 }
1405                 break;
1406         case MASTER:
1407                 tv.tv_sec = sc->sc_advbase;
1408                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1409                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1410                     carp_send_ad, sc);
1411                 break;
1412         }
1413 }
1414
1415 static void
1416 carp_multicast_cleanup(struct carp_softc *sc)
1417 {
1418         struct ip_moptions *imo = &sc->sc_imo;
1419         u_int16_t n = imo->imo_num_memberships;
1420
1421         /* Clean up our own multicast memberships */
1422         while (n-- > 0) {
1423                 if (imo->imo_membership[n] != NULL) {
1424                         in_delmulti(imo->imo_membership[n]);
1425                         imo->imo_membership[n] = NULL;
1426                 }
1427         }
1428         KASSERT(imo->imo_mfilters == NULL,
1429            ("%s: imo_mfilters != NULL", __func__));
1430         imo->imo_num_memberships = 0;
1431         imo->imo_multicast_ifp = NULL;
1432 }
1433
1434 #ifdef INET6
1435 static void
1436 carp_multicast6_cleanup(struct carp_softc *sc)
1437 {
1438         struct ip6_moptions *im6o = &sc->sc_im6o;
1439
1440         while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1441                 struct in6_multi_mship *imm =
1442                     LIST_FIRST(&im6o->im6o_memberships);
1443
1444                 LIST_REMOVE(imm, i6mm_chain);
1445                 in6_leavegroup(imm);
1446         }
1447         im6o->im6o_multicast_ifp = NULL;
1448 }
1449 #endif
1450
1451 static int
1452 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1453 {
1454         struct ifnet *ifp;
1455         struct carp_if *cif;
1456         struct in_ifaddr *ia, *ia_if;
1457         struct ip_moptions *imo = &sc->sc_imo;
1458         struct in_addr addr;
1459         u_long iaddr = htonl(sin->sin_addr.s_addr);
1460         int own, error;
1461
1462         if (sin->sin_addr.s_addr == 0) {
1463                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1464                         carp_set_state(sc, INIT);
1465                 if (sc->sc_naddrs)
1466                         SC2IFP(sc)->if_flags |= IFF_UP;
1467                 if (sc->sc_carpdev)
1468                         CARP_SCLOCK(sc);
1469                 carp_setrun(sc, 0);
1470                 if (sc->sc_carpdev)
1471                         CARP_SCUNLOCK(sc);
1472                 return (0);
1473         }
1474
1475         /* we have to do it by hands to check we won't match on us */
1476         ia_if = NULL; own = 0;
1477         TAILQ_FOREACH(ia, &in_ifaddrhead, ia_link) {
1478                 /* and, yeah, we need a multicast-capable iface too */
1479                 if (ia->ia_ifp != SC2IFP(sc) &&
1480                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1481                     (iaddr & ia->ia_subnetmask) == ia->ia_subnet) {
1482                         if (!ia_if)
1483                                 ia_if = ia;
1484                         if (sin->sin_addr.s_addr ==
1485                             ia->ia_addr.sin_addr.s_addr)
1486                                 own++;
1487                 }
1488         }
1489
1490         if (!ia_if)
1491                 return (EADDRNOTAVAIL);
1492
1493         ia = ia_if;
1494         ifp = ia->ia_ifp;
1495
1496         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1497             (imo->imo_multicast_ifp && imo->imo_multicast_ifp != ifp))
1498                 return (EADDRNOTAVAIL);
1499
1500         if (imo->imo_num_memberships == 0) {
1501                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1502                 if ((imo->imo_membership[0] = in_addmulti(&addr, ifp)) == NULL)
1503                         return (ENOBUFS);
1504                 imo->imo_num_memberships++;
1505                 imo->imo_multicast_ifp = ifp;
1506                 imo->imo_multicast_ttl = CARP_DFLTTL;
1507                 imo->imo_multicast_loop = 0;
1508         }
1509
1510         if (!ifp->if_carp) {
1511
1512                 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP,
1513                     M_WAITOK|M_ZERO);
1514                 if (!cif) {
1515                         error = ENOBUFS;
1516                         goto cleanup;
1517                 }
1518                 if ((error = ifpromisc(ifp, 1))) {
1519                         FREE(cif, M_CARP);
1520                         goto cleanup;
1521                 }
1522
1523                 CARP_LOCK_INIT(cif);
1524                 CARP_LOCK(cif);
1525                 cif->vhif_ifp = ifp;
1526                 TAILQ_INIT(&cif->vhif_vrs);
1527                 ifp->if_carp = cif;
1528
1529         } else {
1530                 struct carp_softc *vr;
1531
1532                 cif = (struct carp_if *)ifp->if_carp;
1533                 CARP_LOCK(cif);
1534                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1535                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1536                                 CARP_UNLOCK(cif);
1537                                 error = EINVAL;
1538                                 goto cleanup;
1539                         }
1540         }
1541         sc->sc_ia = ia;
1542         sc->sc_carpdev = ifp;
1543
1544         { /* XXX prevent endless loop if already in queue */
1545         struct carp_softc *vr, *after = NULL;
1546         int myself = 0;
1547         cif = (struct carp_if *)ifp->if_carp;
1548
1549         /* XXX: cif should not change, right? So we still hold the lock */
1550         CARP_LOCK_ASSERT(cif);
1551
1552         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1553                 if (vr == sc)
1554                         myself = 1;
1555                 if (vr->sc_vhid < sc->sc_vhid)
1556                         after = vr;
1557         }
1558
1559         if (!myself) {
1560                 /* We're trying to keep things in order */
1561                 if (after == NULL) {
1562                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1563                 } else {
1564                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1565                 }
1566                 cif->vhif_nvrs++;
1567         }
1568         }
1569
1570         sc->sc_naddrs++;
1571         SC2IFP(sc)->if_flags |= IFF_UP;
1572         if (own)
1573                 sc->sc_advskew = 0;
1574         carp_sc_state_locked(sc);
1575         carp_setrun(sc, 0);
1576
1577         CARP_UNLOCK(cif);
1578
1579         return (0);
1580
1581 cleanup:
1582         in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1583         return (error);
1584 }
1585
1586 static int
1587 carp_del_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1588 {
1589         int error = 0;
1590
1591         if (!--sc->sc_naddrs) {
1592                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1593                 struct ip_moptions *imo = &sc->sc_imo;
1594
1595                 CARP_LOCK(cif);
1596                 callout_stop(&sc->sc_ad_tmo);
1597                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1598                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1599                 sc->sc_vhid = -1;
1600                 in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1601                 imo->imo_multicast_ifp = NULL;
1602                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1603                 if (!--cif->vhif_nvrs) {
1604                         sc->sc_carpdev->if_carp = NULL;
1605                         CARP_LOCK_DESTROY(cif);
1606                         FREE(cif, M_IFADDR);
1607                 } else {
1608                         CARP_UNLOCK(cif);
1609                 }
1610         }
1611
1612         return (error);
1613 }
1614
1615 #ifdef INET6
1616 static int
1617 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1618 {
1619         struct ifnet *ifp;
1620         struct carp_if *cif;
1621         struct in6_ifaddr *ia, *ia_if;
1622         struct ip6_moptions *im6o = &sc->sc_im6o;
1623         struct in6_multi_mship *imm;
1624         struct in6_addr in6;
1625         int own, error;
1626
1627         if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
1628                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1629                         carp_set_state(sc, INIT);
1630                 if (sc->sc_naddrs6)
1631                         SC2IFP(sc)->if_flags |= IFF_UP;
1632                 if (sc->sc_carpdev)
1633                         CARP_SCLOCK(sc);
1634                 carp_setrun(sc, 0);
1635                 if (sc->sc_carpdev)
1636                         CARP_SCUNLOCK(sc);
1637                 return (0);
1638         }
1639
1640         /* we have to do it by hands to check we won't match on us */
1641         ia_if = NULL; own = 0;
1642         for (ia = in6_ifaddr; ia; ia = ia->ia_next) {
1643                 int i;
1644
1645                 for (i = 0; i < 4; i++) {
1646                         if ((sin6->sin6_addr.s6_addr32[i] &
1647                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]) !=
1648                             (ia->ia_addr.sin6_addr.s6_addr32[i] &
1649                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]))
1650                                 break;
1651                 }
1652                 /* and, yeah, we need a multicast-capable iface too */
1653                 if (ia->ia_ifp != SC2IFP(sc) &&
1654                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1655                     (i == 4)) {
1656                         if (!ia_if)
1657                                 ia_if = ia;
1658                         if (IN6_ARE_ADDR_EQUAL(&sin6->sin6_addr,
1659                             &ia->ia_addr.sin6_addr))
1660                                 own++;
1661                 }
1662         }
1663
1664         if (!ia_if)
1665                 return (EADDRNOTAVAIL);
1666         ia = ia_if;
1667         ifp = ia->ia_ifp;
1668
1669         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1670             (im6o->im6o_multicast_ifp && im6o->im6o_multicast_ifp != ifp))
1671                 return (EADDRNOTAVAIL);
1672
1673         if (!sc->sc_naddrs6) {
1674                 im6o->im6o_multicast_ifp = ifp;
1675
1676                 /* join CARP multicast address */
1677                 bzero(&in6, sizeof(in6));
1678                 in6.s6_addr16[0] = htons(0xff02);
1679                 in6.s6_addr8[15] = 0x12;
1680                 if (in6_setscope(&in6, ifp, NULL) != 0)
1681                         goto cleanup;
1682                 if ((imm = in6_joingroup(ifp, &in6, &error, 0)) == NULL)
1683                         goto cleanup;
1684                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1685
1686                 /* join solicited multicast address */
1687                 bzero(&in6, sizeof(in6));
1688                 in6.s6_addr16[0] = htons(0xff02);
1689                 in6.s6_addr32[1] = 0;
1690                 in6.s6_addr32[2] = htonl(1);
1691                 in6.s6_addr32[3] = sin6->sin6_addr.s6_addr32[3];
1692                 in6.s6_addr8[12] = 0xff;
1693                 if (in6_setscope(&in6, ifp, NULL) != 0)
1694                         goto cleanup;
1695                 if ((imm = in6_joingroup(ifp, &in6, &error, 0)) == NULL)
1696                         goto cleanup;
1697                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1698         }
1699
1700         if (!ifp->if_carp) {
1701                 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP,
1702                     M_WAITOK|M_ZERO);
1703                 if (!cif) {
1704                         error = ENOBUFS;
1705                         goto cleanup;
1706                 }
1707                 if ((error = ifpromisc(ifp, 1))) {
1708                         FREE(cif, M_CARP);
1709                         goto cleanup;
1710                 }
1711
1712                 CARP_LOCK_INIT(cif);
1713                 CARP_LOCK(cif);
1714                 cif->vhif_ifp = ifp;
1715                 TAILQ_INIT(&cif->vhif_vrs);
1716                 ifp->if_carp = cif;
1717
1718         } else {
1719                 struct carp_softc *vr;
1720
1721                 cif = (struct carp_if *)ifp->if_carp;
1722                 CARP_LOCK(cif);
1723                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1724                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1725                                 CARP_UNLOCK(cif);
1726                                 error = EINVAL;
1727                                 goto cleanup;
1728                         }
1729         }
1730         sc->sc_ia6 = ia;
1731         sc->sc_carpdev = ifp;
1732
1733         { /* XXX prevent endless loop if already in queue */
1734         struct carp_softc *vr, *after = NULL;
1735         int myself = 0;
1736         cif = (struct carp_if *)ifp->if_carp;
1737         CARP_LOCK_ASSERT(cif);
1738
1739         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1740                 if (vr == sc)
1741                         myself = 1;
1742                 if (vr->sc_vhid < sc->sc_vhid)
1743                         after = vr;
1744         }
1745
1746         if (!myself) {
1747                 /* We're trying to keep things in order */
1748                 if (after == NULL) {
1749                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1750                 } else {
1751                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1752                 }
1753                 cif->vhif_nvrs++;
1754         }
1755         }
1756
1757         sc->sc_naddrs6++;
1758         SC2IFP(sc)->if_flags |= IFF_UP;
1759         if (own)
1760                 sc->sc_advskew = 0;
1761         carp_sc_state_locked(sc);
1762         carp_setrun(sc, 0);
1763
1764         CARP_UNLOCK(cif);
1765
1766         return (0);
1767
1768 cleanup:
1769         /* clean up multicast memberships */
1770         if (!sc->sc_naddrs6) {
1771                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1772                         imm = LIST_FIRST(&im6o->im6o_memberships);
1773                         LIST_REMOVE(imm, i6mm_chain);
1774                         in6_leavegroup(imm);
1775                 }
1776         }
1777         return (error);
1778 }
1779
1780 static int
1781 carp_del_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1782 {
1783         int error = 0;
1784
1785         if (!--sc->sc_naddrs6) {
1786                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1787                 struct ip6_moptions *im6o = &sc->sc_im6o;
1788
1789                 CARP_LOCK(cif);
1790                 callout_stop(&sc->sc_ad_tmo);
1791                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1792                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1793                 sc->sc_vhid = -1;
1794                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1795                         struct in6_multi_mship *imm =
1796                             LIST_FIRST(&im6o->im6o_memberships);
1797
1798                         LIST_REMOVE(imm, i6mm_chain);
1799                         in6_leavegroup(imm);
1800                 }
1801                 im6o->im6o_multicast_ifp = NULL;
1802                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1803                 if (!--cif->vhif_nvrs) {
1804                         CARP_LOCK_DESTROY(cif);
1805                         sc->sc_carpdev->if_carp = NULL;
1806                         FREE(cif, M_IFADDR);
1807                 } else
1808                         CARP_UNLOCK(cif);
1809         }
1810
1811         return (error);
1812 }
1813 #endif /* INET6 */
1814
1815 static int
1816 carp_ioctl(struct ifnet *ifp, u_long cmd, caddr_t addr)
1817 {
1818         struct carp_softc *sc = ifp->if_softc, *vr;
1819         struct carpreq carpr;
1820         struct ifaddr *ifa;
1821         struct ifreq *ifr;
1822         struct ifaliasreq *ifra;
1823         int locked = 0, error = 0;
1824
1825         ifa = (struct ifaddr *)addr;
1826         ifra = (struct ifaliasreq *)addr;
1827         ifr = (struct ifreq *)addr;
1828
1829         switch (cmd) {
1830         case SIOCSIFADDR:
1831                 switch (ifa->ifa_addr->sa_family) {
1832 #ifdef INET
1833                 case AF_INET:
1834                         SC2IFP(sc)->if_flags |= IFF_UP;
1835                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1836                             sizeof(struct sockaddr));
1837                         error = carp_set_addr(sc, satosin(ifa->ifa_addr));
1838                         break;
1839 #endif /* INET */
1840 #ifdef INET6
1841                 case AF_INET6:
1842                         SC2IFP(sc)->if_flags |= IFF_UP;
1843                         error = carp_set_addr6(sc, satosin6(ifa->ifa_addr));
1844                         break;
1845 #endif /* INET6 */
1846                 default:
1847                         error = EAFNOSUPPORT;
1848                         break;
1849                 }
1850                 break;
1851
1852         case SIOCAIFADDR:
1853                 switch (ifa->ifa_addr->sa_family) {
1854 #ifdef INET
1855                 case AF_INET:
1856                         SC2IFP(sc)->if_flags |= IFF_UP;
1857                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1858                             sizeof(struct sockaddr));
1859                         error = carp_set_addr(sc, satosin(&ifra->ifra_addr));
1860                         break;
1861 #endif /* INET */
1862 #ifdef INET6
1863                 case AF_INET6:
1864                         SC2IFP(sc)->if_flags |= IFF_UP;
1865                         error = carp_set_addr6(sc, satosin6(&ifra->ifra_addr));
1866                         break;
1867 #endif /* INET6 */
1868                 default:
1869                         error = EAFNOSUPPORT;
1870                         break;
1871                 }
1872                 break;
1873
1874         case SIOCDIFADDR:
1875                 switch (ifa->ifa_addr->sa_family) {
1876 #ifdef INET
1877                 case AF_INET:
1878                         error = carp_del_addr(sc, satosin(&ifra->ifra_addr));
1879                         break;
1880 #endif /* INET */
1881 #ifdef INET6
1882                 case AF_INET6:
1883                         error = carp_del_addr6(sc, satosin6(&ifra->ifra_addr));
1884                         break;
1885 #endif /* INET6 */
1886                 default:
1887                         error = EAFNOSUPPORT;
1888                         break;
1889                 }
1890                 break;
1891
1892         case SIOCSIFFLAGS:
1893                 if (sc->sc_carpdev) {
1894                         locked = 1;
1895                         CARP_SCLOCK(sc);
1896                 }
1897                 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) {
1898                         callout_stop(&sc->sc_ad_tmo);
1899                         callout_stop(&sc->sc_md_tmo);
1900                         callout_stop(&sc->sc_md6_tmo);
1901                         if (sc->sc_state == MASTER)
1902                                 carp_send_ad_locked(sc);
1903                         carp_set_state(sc, INIT);
1904                         carp_setrun(sc, 0);
1905                 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) {
1906                         SC2IFP(sc)->if_flags |= IFF_UP;
1907                         carp_setrun(sc, 0);
1908                 }
1909                 break;
1910
1911         case SIOCSVH:
1912                 error = priv_check(curthread, PRIV_NETINET_CARP);
1913                 if (error)
1914                         break;
1915                 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1916                         break;
1917                 error = 1;
1918                 if (sc->sc_carpdev) {
1919                         locked = 1;
1920                         CARP_SCLOCK(sc);
1921                 }
1922                 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) {
1923                         switch (carpr.carpr_state) {
1924                         case BACKUP:
1925                                 callout_stop(&sc->sc_ad_tmo);
1926                                 carp_set_state(sc, BACKUP);
1927                                 carp_setrun(sc, 0);
1928                                 carp_setroute(sc, RTM_DELETE);
1929                                 break;
1930                         case MASTER:
1931                                 carp_master_down_locked(sc);
1932                                 break;
1933                         default:
1934                                 break;
1935                         }
1936                 }
1937                 if (carpr.carpr_vhid > 0) {
1938                         if (carpr.carpr_vhid > 255) {
1939                                 error = EINVAL;
1940                                 break;
1941                         }
1942                         if (sc->sc_carpdev) {
1943                                 struct carp_if *cif;
1944                                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1945                                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1946                                         if (vr != sc &&
1947                                             vr->sc_vhid == carpr.carpr_vhid) {
1948                                                 error = EEXIST;
1949                                                 break;
1950                                         }
1951                                 if (error == EEXIST)
1952                                         break;
1953                         }
1954                         sc->sc_vhid = carpr.carpr_vhid;
1955                         IF_LLADDR(sc->sc_ifp)[0] = 0;
1956                         IF_LLADDR(sc->sc_ifp)[1] = 0;
1957                         IF_LLADDR(sc->sc_ifp)[2] = 0x5e;
1958                         IF_LLADDR(sc->sc_ifp)[3] = 0;
1959                         IF_LLADDR(sc->sc_ifp)[4] = 1;
1960                         IF_LLADDR(sc->sc_ifp)[5] = sc->sc_vhid;
1961                         error--;
1962                 }
1963                 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) {
1964                         if (carpr.carpr_advskew >= 255) {
1965                                 error = EINVAL;
1966                                 break;
1967                         }
1968                         if (carpr.carpr_advbase > 255) {
1969                                 error = EINVAL;
1970                                 break;
1971                         }
1972                         sc->sc_advbase = carpr.carpr_advbase;
1973                         sc->sc_advskew = carpr.carpr_advskew;
1974                         error--;
1975                 }
1976                 bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1977                 if (error > 0)
1978                         error = EINVAL;
1979                 else {
1980                         error = 0;
1981                         carp_setrun(sc, 0);
1982                 }
1983                 break;
1984
1985         case SIOCGVH:
1986                 /* XXX: lockless read */
1987                 bzero(&carpr, sizeof(carpr));
1988                 carpr.carpr_state = sc->sc_state;
1989                 carpr.carpr_vhid = sc->sc_vhid;
1990                 carpr.carpr_advbase = sc->sc_advbase;
1991                 carpr.carpr_advskew = sc->sc_advskew;
1992                 error = priv_check(curthread, PRIV_NETINET_CARP);
1993                 if (error == 0)
1994                         bcopy(sc->sc_key, carpr.carpr_key,
1995                             sizeof(carpr.carpr_key));
1996                 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1997                 break;
1998
1999         default:
2000                 error = EINVAL;
2001         }
2002
2003         if (locked)
2004                 CARP_SCUNLOCK(sc);
2005
2006         carp_hmac_prepare(sc);
2007
2008         return (error);
2009 }
2010
2011 /*
2012  * XXX: this is looutput. We should eventually use it from there.
2013  */
2014 static int
2015 carp_looutput(struct ifnet *ifp, struct mbuf *m, struct sockaddr *dst,
2016     struct rtentry *rt)
2017 {
2018         u_int32_t af;
2019
2020         M_ASSERTPKTHDR(m); /* check if we have the packet header */
2021
2022         if (rt && rt->rt_flags & (RTF_REJECT|RTF_BLACKHOLE)) {
2023                 m_freem(m);
2024                 return (rt->rt_flags & RTF_BLACKHOLE ? 0 :
2025                         rt->rt_flags & RTF_HOST ? EHOSTUNREACH : ENETUNREACH);
2026         }
2027
2028         ifp->if_opackets++;
2029         ifp->if_obytes += m->m_pkthdr.len;
2030
2031         /* BPF writes need to be handled specially. */
2032         if (dst->sa_family == AF_UNSPEC) {
2033                 bcopy(dst->sa_data, &af, sizeof(af));
2034                 dst->sa_family = af;
2035         }
2036
2037 #if 1   /* XXX */
2038         switch (dst->sa_family) {
2039         case AF_INET:
2040         case AF_INET6:
2041         case AF_IPX:
2042         case AF_APPLETALK:
2043                 break;
2044         default:
2045                 printf("carp_looutput: af=%d unexpected\n", dst->sa_family);
2046                 m_freem(m);
2047                 return (EAFNOSUPPORT);
2048         }
2049 #endif
2050         return(if_simloop(ifp, m, dst->sa_family, 0));
2051 }
2052
2053 /*
2054  * Start output on carp interface. This function should never be called.
2055  */
2056 static void
2057 carp_start(struct ifnet *ifp)
2058 {
2059 #ifdef DEBUG
2060         printf("%s: start called\n", ifp->if_xname);
2061 #endif
2062 }
2063
2064 int
2065 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *sa,
2066     struct rtentry *rt)
2067 {
2068         struct m_tag *mtag;
2069         struct carp_softc *sc;
2070         struct ifnet *carp_ifp;
2071
2072         if (!sa)
2073                 return (0);
2074
2075         switch (sa->sa_family) {
2076 #ifdef INET
2077         case AF_INET:
2078                 break;
2079 #endif /* INET */
2080 #ifdef INET6
2081         case AF_INET6:
2082                 break;
2083 #endif /* INET6 */
2084         default:
2085                 return (0);
2086         }
2087
2088         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
2089         if (mtag == NULL)
2090                 return (0);
2091
2092         bcopy(mtag + 1, &carp_ifp, sizeof(struct ifnet *));
2093         sc = carp_ifp->if_softc;
2094
2095         /* Set the source MAC address to Virtual Router MAC Address */
2096         switch (ifp->if_type) {
2097         case IFT_ETHER:
2098         case IFT_L2VLAN: {
2099                         struct ether_header *eh;
2100
2101                         eh = mtod(m, struct ether_header *);
2102                         eh->ether_shost[0] = 0;
2103                         eh->ether_shost[1] = 0;
2104                         eh->ether_shost[2] = 0x5e;
2105                         eh->ether_shost[3] = 0;
2106                         eh->ether_shost[4] = 1;
2107                         eh->ether_shost[5] = sc->sc_vhid;
2108                 }
2109                 break;
2110         case IFT_FDDI: {
2111                         struct fddi_header *fh;
2112
2113                         fh = mtod(m, struct fddi_header *);
2114                         fh->fddi_shost[0] = 0;
2115                         fh->fddi_shost[1] = 0;
2116                         fh->fddi_shost[2] = 0x5e;
2117                         fh->fddi_shost[3] = 0;
2118                         fh->fddi_shost[4] = 1;
2119                         fh->fddi_shost[5] = sc->sc_vhid;
2120                 }
2121                 break;
2122         case IFT_ISO88025: {
2123                         struct iso88025_header *th;
2124                         th = mtod(m, struct iso88025_header *);
2125                         th->iso88025_shost[0] = 3;
2126                         th->iso88025_shost[1] = 0;
2127                         th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
2128                         th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
2129                         th->iso88025_shost[4] = 0;
2130                         th->iso88025_shost[5] = 0;
2131                 }
2132                 break;
2133         default:
2134                 printf("%s: carp is not supported for this interface type\n",
2135                     ifp->if_xname);
2136                 return (EOPNOTSUPP);
2137         }
2138
2139         return (0);
2140 }
2141
2142 static void
2143 carp_set_state(struct carp_softc *sc, int state)
2144 {
2145
2146         if (sc->sc_carpdev)
2147                 CARP_SCLOCK_ASSERT(sc);
2148
2149         if (sc->sc_state == state)
2150                 return;
2151
2152         sc->sc_state = state;
2153         switch (state) {
2154         case BACKUP:
2155                 SC2IFP(sc)->if_link_state = LINK_STATE_DOWN;
2156                 break;
2157         case MASTER:
2158                 SC2IFP(sc)->if_link_state = LINK_STATE_UP;
2159                 break;
2160         default:
2161                 SC2IFP(sc)->if_link_state = LINK_STATE_UNKNOWN;
2162                 break;
2163         }
2164         rt_ifmsg(SC2IFP(sc));
2165 }
2166
2167 void
2168 carp_carpdev_state(void *v)
2169 {
2170         struct carp_if *cif = v;
2171
2172         CARP_LOCK(cif);
2173         carp_carpdev_state_locked(cif);
2174         CARP_UNLOCK(cif);
2175 }
2176
2177 static void
2178 carp_carpdev_state_locked(struct carp_if *cif)
2179 {
2180         struct carp_softc *sc;
2181
2182         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list)
2183                 carp_sc_state_locked(sc);
2184 }
2185
2186 static void
2187 carp_sc_state_locked(struct carp_softc *sc)
2188 {
2189         CARP_SCLOCK_ASSERT(sc);
2190
2191         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
2192             !(sc->sc_carpdev->if_flags & IFF_UP)) {
2193                 sc->sc_flags_backup = SC2IFP(sc)->if_flags;
2194                 SC2IFP(sc)->if_flags &= ~IFF_UP;
2195                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
2196                 callout_stop(&sc->sc_ad_tmo);
2197                 callout_stop(&sc->sc_md_tmo);
2198                 callout_stop(&sc->sc_md6_tmo);
2199                 carp_set_state(sc, INIT);
2200                 carp_setrun(sc, 0);
2201                 if (!sc->sc_suppress) {
2202                         carp_suppress_preempt++;
2203                         if (carp_suppress_preempt == 1) {
2204                                 CARP_SCUNLOCK(sc);
2205                                 carp_send_ad_all();
2206                                 CARP_SCLOCK(sc);
2207                         }
2208                 }
2209                 sc->sc_suppress = 1;
2210         } else {
2211                 SC2IFP(sc)->if_flags |= sc->sc_flags_backup;
2212                 carp_set_state(sc, INIT);
2213                 carp_setrun(sc, 0);
2214                 if (sc->sc_suppress)
2215                         carp_suppress_preempt--;
2216                 sc->sc_suppress = 0;
2217         }
2218
2219         return;
2220 }
2221
2222 static int
2223 carp_modevent(module_t mod, int type, void *data)
2224 {
2225         switch (type) {
2226         case MOD_LOAD:
2227                 if_detach_event_tag = EVENTHANDLER_REGISTER(ifnet_departure_event,
2228                     carp_ifdetach, NULL, EVENTHANDLER_PRI_ANY);
2229                 if (if_detach_event_tag == NULL)
2230                         return (ENOMEM);
2231                 mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2232                 LIST_INIT(&carpif_list);
2233                 if_clone_attach(&carp_cloner);
2234                 break;
2235
2236         case MOD_UNLOAD:
2237                 EVENTHANDLER_DEREGISTER(ifnet_departure_event, if_detach_event_tag);
2238                 if_clone_detach(&carp_cloner);
2239                 mtx_destroy(&carp_mtx);
2240                 break;
2241
2242         default:
2243                 return (EINVAL);
2244         }
2245
2246         return (0);
2247 }
2248
2249 static moduledata_t carp_mod = {
2250         "carp",
2251         carp_modevent,
2252         0
2253 };
2254
2255 DECLARE_MODULE(carp, carp_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);