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;
 245 #ifdef INET6
 246         struct in6_addr in6;
 247 #endif
 248
 249         if (sc->sc_carpdev)
 250                 CARP_SCLOCK(sc);
 251
 252         /* XXX: possible race here */
 253
 254         /* compute ipad from key */
 255         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 256         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 257         for (i = 0; i < sizeof(sc->sc_pad); i++)
 258                 sc->sc_pad[i] ^= 0x36;
 259
 260         /* precompute first part of inner hash */
 261         SHA1Init(&sc->sc_sha1);
 262         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 263         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 264         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 265         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 266 #ifdef INET
 267         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 268                 if (ifa->ifa_addr->sa_family == AF_INET)
 269                         SHA1Update(&sc->sc_sha1,
 270                             (void *)&ifatoia(ifa)->ia_addr.sin_addr.s_addr,
 271                             sizeof(struct in_addr));
 272         }
 273 #endif /* INET */
 274 #ifdef INET6
 275         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 276                 if (ifa->ifa_addr->sa_family == AF_INET6) {
 277                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 278                         in6_clearscope(&in6);
 279                         SHA1Update(&sc->sc_sha1, (void *)&in6, sizeof(in6));
 280                 }
 281         }
 282 #endif /* INET6 */
 283
 284         /* convert ipad to opad */
 285         for (i = 0; i < sizeof(sc->sc_pad); i++)
 286                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 287
 288         if (sc->sc_carpdev)
 289                 CARP_SCUNLOCK(sc);
 290 }
 291
 292 static void
 293 carp_hmac_generate(struct carp_softc *sc, u_int32_t counter[2],
 294     unsigned char md[20])
 295 {
 296         SHA1_CTX sha1ctx;
 297
 298         /* fetch first half of inner hash */
 299         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 300
 301         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 302         SHA1Final(md, &sha1ctx);
 303
 304         /* outer hash */
 305         SHA1Init(&sha1ctx);
 306         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 307         SHA1Update(&sha1ctx, md, 20);
 308         SHA1Final(md, &sha1ctx);
 309 }
 310
 311 static int
 312 carp_hmac_verify(struct carp_softc *sc, u_int32_t counter[2],
 313     unsigned char md[20])
 314 {
 315         unsigned char md2[20];
 316
 317         CARP_SCLOCK_ASSERT(sc);
 318
 319         carp_hmac_generate(sc, counter, md2);
 320
 321         return (bcmp(md, md2, sizeof(md2)));
 322 }
 323
 324 static void
 325 carp_setroute(struct carp_softc *sc, int cmd)
 326 {
 327         struct ifaddr *ifa;
 328         int s;
 329
 330         if (sc->sc_carpdev)
 331                 CARP_SCLOCK_ASSERT(sc);
 332
 333         s = splnet();
 334         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 335                 if (ifa->ifa_addr->sa_family == AF_INET &&
 336                     sc->sc_carpdev != NULL) {
 337                         int count = carp_addrcount(
 338                             (struct carp_if *)sc->sc_carpdev->if_carp,
 339                             ifatoia(ifa), CARP_COUNT_MASTER);
 340
 341                         if ((cmd == RTM_ADD && count == 1) ||
 342                             (cmd == RTM_DELETE && count == 0))
 343                                 rtinit(ifa, cmd, RTF_UP | RTF_HOST);
 344                 }
 345 #ifdef INET6
 346                 if (ifa->ifa_addr->sa_family == AF_INET6) {
 347                         if (cmd == RTM_ADD)
 348                                 in6_ifaddloop(ifa);
 349                         else
 350                                 in6_ifremloop(ifa);
 351                 }
 352 #endif /* INET6 */
 353         }
 354         splx(s);
 355 }
 356
 357 static int
 358 carp_clone_create(struct if_clone *ifc, int unit, caddr_t params)
 359 {
 360
 361         struct carp_softc *sc;
 362         struct ifnet *ifp;
 363
 364         MALLOC(sc, struct carp_softc *, sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
 365         ifp = SC2IFP(sc) = if_alloc(IFT_ETHER);
 366         if (ifp == NULL) {
 367                 FREE(sc, M_CARP);
 368                 return (ENOSPC);
 369         }
 370
 371         sc->sc_flags_backup = 0;
 372         sc->sc_suppress = 0;
 373         sc->sc_advbase = CARP_DFLTINTV;
 374         sc->sc_vhid = -1;       /* required setting */
 375         sc->sc_advskew = 0;
 376         sc->sc_init_counter = 1;
 377         sc->sc_naddrs = sc->sc_naddrs6 = 0; /* M_ZERO? */
 378 #ifdef INET6
 379         sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL;
 380 #endif
 381         sc->sc_imo.imo_membership = (struct in_multi **)malloc(
 382             (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
 383             M_WAITOK);
 384         sc->sc_imo.imo_mfilters = NULL;
 385         sc->sc_imo.imo_max_memberships = IP_MIN_MEMBERSHIPS;
 386         sc->sc_imo.imo_multicast_vif = -1;
 387
 388         callout_init(&sc->sc_ad_tmo, CALLOUT_MPSAFE);
 389         callout_init(&sc->sc_md_tmo, CALLOUT_MPSAFE);
 390         callout_init(&sc->sc_md6_tmo, CALLOUT_MPSAFE);
 391
 392         ifp->if_softc = sc;
 393         if_initname(ifp, CARP_IFNAME, unit);
 394         ifp->if_mtu = ETHERMTU;
 395         ifp->if_flags = IFF_LOOPBACK;
 396         ifp->if_ioctl = carp_ioctl;
 397         ifp->if_output = carp_looutput;
 398         ifp->if_start = carp_start;
 399         ifp->if_type = IFT_CARP;
 400         ifp->if_snd.ifq_maxlen = ifqmaxlen;
 401         ifp->if_hdrlen = 0;
 402         if_attach(ifp);
 403         bpfattach(SC2IFP(sc), DLT_NULL, sizeof(u_int32_t));
 404         mtx_lock(&carp_mtx);
 405         LIST_INSERT_HEAD(&carpif_list, sc, sc_next);
 406         mtx_unlock(&carp_mtx);
 407         return (0);
 408 }
 409
 410 static void
 411 carp_clone_destroy(struct ifnet *ifp)
 412 {
 413         struct carp_softc *sc = ifp->if_softc;
 414
 415         if (sc->sc_carpdev)
 416                 CARP_SCLOCK(sc);
 417         carpdetach(sc, 1);      /* Returns unlocked. */
 418
 419         mtx_lock(&carp_mtx);
 420         LIST_REMOVE(sc, sc_next);
 421         mtx_unlock(&carp_mtx);
 422         bpfdetach(ifp);
 423         if_detach(ifp);
 424         if_free_type(ifp, IFT_ETHER);
 425         free(sc->sc_imo.imo_membership, M_CARP);
 426         free(sc, M_CARP);
 427 }
 428
 429 /*
 430  * This function can be called on CARP interface destroy path,
 431  * and in case of the removal of the underlying interface as
 432  * well. We differentiate these two cases. In the latter case
 433  * we do not cleanup our multicast memberships, since they
 434  * are already freed. Also, in the latter case we do not
 435  * release the lock on return, because the function will be
 436  * called once more, for another CARP instance on the same
 437  * interface.
 438  */
 439 static void
 440 carpdetach(struct carp_softc *sc, int unlock)
 441 {
 442         struct carp_if *cif;
 443
 444         callout_stop(&sc->sc_ad_tmo);
 445         callout_stop(&sc->sc_md_tmo);
 446         callout_stop(&sc->sc_md6_tmo);
 447
 448         if (sc->sc_suppress)
 449                 carp_suppress_preempt--;
 450         sc->sc_suppress = 0;
 451
 452         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS)
 453                 carp_suppress_preempt--;
 454         sc->sc_sendad_errors = 0;
 455
 456         carp_set_state(sc, INIT);
 457         SC2IFP(sc)->if_flags &= ~IFF_UP;
 458         carp_setrun(sc, 0);
 459         if (unlock)
 460                 carp_multicast_cleanup(sc);
 461 #ifdef INET6
 462         carp_multicast6_cleanup(sc);
 463 #endif
 464
 465         if (sc->sc_carpdev != NULL) {
 466                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
 467                 CARP_LOCK_ASSERT(cif);
 468                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
 469                 if (!--cif->vhif_nvrs) {
 470                         ifpromisc(sc->sc_carpdev, 0);
 471                         sc->sc_carpdev->if_carp = NULL;
 472                         CARP_LOCK_DESTROY(cif);
 473                         FREE(cif, M_IFADDR);
 474                 } else if (unlock)
 475                         CARP_UNLOCK(cif);
 476                 sc->sc_carpdev = NULL;
 477         }
 478 }
 479
 480 /* Detach an interface from the carp. */
 481 static void
 482 carp_ifdetach(void *arg __unused, struct ifnet *ifp)
 483 {
 484         struct carp_if *cif = (struct carp_if *)ifp->if_carp;
 485         struct carp_softc *sc, *nextsc;
 486
 487         if (cif == NULL)
 488                 return;
 489
 490         /*
 491          * XXX: At the end of for() cycle the lock will be destroyed.
 492          */
 493         CARP_LOCK(cif);
 494         for (sc = TAILQ_FIRST(&cif->vhif_vrs); sc; sc = nextsc) {
 495                 nextsc = TAILQ_NEXT(sc, sc_list);
 496                 carpdetach(sc, 0);
 497         }
 498 }
 499
 500 /*
 501  * process input packet.
 502  * we have rearranged checks order compared to the rfc,
 503  * but it seems more efficient this way or not possible otherwise.
 504  */
 505 void
 506 carp_input(struct mbuf *m, int hlen)
 507 {
 508         struct ip *ip = mtod(m, struct ip *);
 509         struct carp_header *ch;
 510         int iplen, len;
 511
 512         carpstats.carps_ipackets++;
 513
 514         if (!carp_opts[CARPCTL_ALLOW]) {
 515                 m_freem(m);
 516                 return;
 517         }
 518
 519         /* check if received on a valid carp interface */
 520         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 521                 carpstats.carps_badif++;
 522                 CARP_LOG("carp_input: packet received on non-carp "
 523                     "interface: %s\n",
 524                     m->m_pkthdr.rcvif->if_xname);
 525                 m_freem(m);
 526                 return;
 527         }
 528
 529         /* verify that the IP TTL is 255.  */
 530         if (ip->ip_ttl != CARP_DFLTTL) {
 531                 carpstats.carps_badttl++;
 532                 CARP_LOG("carp_input: received ttl %d != 255i on %s\n",
 533                     ip->ip_ttl,
 534                     m->m_pkthdr.rcvif->if_xname);
 535                 m_freem(m);
 536                 return;
 537         }
 538
 539         iplen = ip->ip_hl << 2;
 540
 541         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 542                 carpstats.carps_badlen++;
 543                 CARP_LOG("carp_input: received len %zd < "
 544                     "sizeof(struct carp_header)\n",
 545                     m->m_len - sizeof(struct ip));
 546                 m_freem(m);
 547                 return;
 548         }
 549
 550         if (iplen + sizeof(*ch) < m->m_len) {
 551                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 552                         carpstats.carps_hdrops++;
 553                         CARP_LOG("carp_input: pullup failed\n");
 554                         return;
 555                 }
 556                 ip = mtod(m, struct ip *);
 557         }
 558         ch = (struct carp_header *)((char *)ip + iplen);
 559
 560         /*
 561          * verify that the received packet length is
 562          * equal to the CARP header
 563          */
 564         len = iplen + sizeof(*ch);
 565         if (len > m->m_pkthdr.len) {
 566                 carpstats.carps_badlen++;
 567                 CARP_LOG("carp_input: packet too short %d on %s\n",
 568                     m->m_pkthdr.len,
 569                     m->m_pkthdr.rcvif->if_xname);
 570                 m_freem(m);
 571                 return;
 572         }
 573
 574         if ((m = m_pullup(m, len)) == NULL) {
 575                 carpstats.carps_hdrops++;
 576                 return;
 577         }
 578         ip = mtod(m, struct ip *);
 579         ch = (struct carp_header *)((char *)ip + iplen);
 580
 581         /* verify the CARP checksum */
 582         m->m_data += iplen;
 583         if (carp_cksum(m, len - iplen)) {
 584                 carpstats.carps_badsum++;
 585                 CARP_LOG("carp_input: checksum failed on %s\n",
 586                     m->m_pkthdr.rcvif->if_xname);
 587                 m_freem(m);
 588                 return;
 589         }
 590         m->m_data -= iplen;
 591
 592         carp_input_c(m, ch, AF_INET);
 593 }
 594
 595 #ifdef INET6
 596 int
 597 carp6_input(struct mbuf **mp, int *offp, int proto)
 598 {
 599         struct mbuf *m = *mp;
 600         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 601         struct carp_header *ch;
 602         u_int len;
 603
 604         carpstats.carps_ipackets6++;
 605
 606         if (!carp_opts[CARPCTL_ALLOW]) {
 607                 m_freem(m);
 608                 return (IPPROTO_DONE);
 609         }
 610
 611         /* check if received on a valid carp interface */
 612         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 613                 carpstats.carps_badif++;
 614                 CARP_LOG("carp6_input: packet received on non-carp "
 615                     "interface: %s\n",
 616                     m->m_pkthdr.rcvif->if_xname);
 617                 m_freem(m);
 618                 return (IPPROTO_DONE);
 619         }
 620
 621         /* verify that the IP TTL is 255 */
 622         if (ip6->ip6_hlim != CARP_DFLTTL) {
 623                 carpstats.carps_badttl++;
 624                 CARP_LOG("carp6_input: received ttl %d != 255 on %s\n",
 625                     ip6->ip6_hlim,
 626                     m->m_pkthdr.rcvif->if_xname);
 627                 m_freem(m);
 628                 return (IPPROTO_DONE);
 629         }
 630
 631         /* verify that we have a complete carp packet */
 632         len = m->m_len;
 633         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 634         if (ch == NULL) {
 635                 carpstats.carps_badlen++;
 636                 CARP_LOG("carp6_input: packet size %u too small\n", len);
 637                 return (IPPROTO_DONE);
 638         }
 639
 640
 641         /* verify the CARP checksum */
 642         m->m_data += *offp;
 643         if (carp_cksum(m, sizeof(*ch))) {
 644                 carpstats.carps_badsum++;
 645                 CARP_LOG("carp6_input: checksum failed, on %s\n",
 646                     m->m_pkthdr.rcvif->if_xname);
 647                 m_freem(m);
 648                 return (IPPROTO_DONE);
 649         }
 650         m->m_data -= *offp;
 651
 652         carp_input_c(m, ch, AF_INET6);
 653         return (IPPROTO_DONE);
 654 }
 655 #endif /* INET6 */
 656
 657 static void
 658 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 659 {
 660         struct ifnet *ifp = m->m_pkthdr.rcvif;
 661         struct carp_softc *sc;
 662         u_int64_t tmp_counter;
 663         struct timeval sc_tv, ch_tv;
 664
 665         /* verify that the VHID is valid on the receiving interface */
 666         CARP_LOCK(ifp->if_carp);
 667         TAILQ_FOREACH(sc, &((struct carp_if *)ifp->if_carp)->vhif_vrs, sc_list)
 668                 if (sc->sc_vhid == ch->carp_vhid)
 669                         break;
 670
 671         if (!sc || !((SC2IFP(sc)->if_flags & IFF_UP) &&
 672             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING))) {
 673                 carpstats.carps_badvhid++;
 674                 CARP_UNLOCK(ifp->if_carp);
 675                 m_freem(m);
 676                 return;
 677         }
 678
 679         getmicrotime(&SC2IFP(sc)->if_lastchange);
 680         SC2IFP(sc)->if_ipackets++;
 681         SC2IFP(sc)->if_ibytes += m->m_pkthdr.len;
 682
 683         if (bpf_peers_present(SC2IFP(sc)->if_bpf)) {
 684                 struct ip *ip = mtod(m, struct ip *);
 685                 uint32_t af1 = af;
 686
 687                 /* BPF wants net byte order */
 688                 ip->ip_len = htons(ip->ip_len + (ip->ip_hl << 2));
 689                 ip->ip_off = htons(ip->ip_off);
 690                 bpf_mtap2(SC2IFP(sc)->if_bpf, &af1, sizeof(af1), m);
 691         }
 692
 693         /* verify the CARP version. */
 694         if (ch->carp_version != CARP_VERSION) {
 695                 carpstats.carps_badver++;
 696                 SC2IFP(sc)->if_ierrors++;
 697                 CARP_UNLOCK(ifp->if_carp);
 698                 CARP_LOG("%s; invalid version %d\n",
 699                     SC2IFP(sc)->if_xname,
 700                     ch->carp_version);
 701                 m_freem(m);
 702                 return;
 703         }
 704
 705         /* verify the hash */
 706         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 707                 carpstats.carps_badauth++;
 708                 SC2IFP(sc)->if_ierrors++;
 709                 CARP_UNLOCK(ifp->if_carp);
 710                 CARP_LOG("%s: incorrect hash\n", SC2IFP(sc)->if_xname);
 711                 m_freem(m);
 712                 return;
 713         }
 714
 715         tmp_counter = ntohl(ch->carp_counter[0]);
 716         tmp_counter = tmp_counter<<32;
 717         tmp_counter += ntohl(ch->carp_counter[1]);
 718
 719         /* XXX Replay protection goes here */
 720
 721         sc->sc_init_counter = 0;
 722         sc->sc_counter = tmp_counter;
 723
 724         sc_tv.tv_sec = sc->sc_advbase;
 725         if (carp_suppress_preempt && sc->sc_advskew <  240)
 726                 sc_tv.tv_usec = 240 * 1000000 / 256;
 727         else
 728                 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256;
 729         ch_tv.tv_sec = ch->carp_advbase;
 730         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 731
 732         switch (sc->sc_state) {
 733         case INIT:
 734                 break;
 735         case MASTER:
 736                 /*
 737                  * If we receive an advertisement from a master who's going to
 738                  * be more frequent than us, go into BACKUP state.
 739                  */
 740                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 741                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 742                         callout_stop(&sc->sc_ad_tmo);
 743                         CARP_DEBUG("%s: MASTER -> BACKUP "
 744                            "(more frequent advertisement received)\n",
 745                            SC2IFP(sc)->if_xname);
 746                         carp_set_state(sc, BACKUP);
 747                         carp_setrun(sc, 0);
 748                         carp_setroute(sc, RTM_DELETE);
 749                 }
 750                 break;
 751         case BACKUP:
 752                 /*
 753                  * If we're pre-empting masters who advertise slower than us,
 754                  * and this one claims to be slower, treat him as down.
 755                  */
 756                 if (carp_opts[CARPCTL_PREEMPT] &&
 757                     timevalcmp(&sc_tv, &ch_tv, <)) {
 758                         CARP_DEBUG("%s: BACKUP -> MASTER "
 759                             "(preempting a slower master)\n",
 760                             SC2IFP(sc)->if_xname);
 761                         carp_master_down_locked(sc);
 762                         break;
 763                 }
 764
 765                 /*
 766                  *  If the master is going to advertise at such a low frequency
 767                  *  that he's guaranteed to time out, we'd might as well just
 768                  *  treat him as timed out now.
 769                  */
 770                 sc_tv.tv_sec = sc->sc_advbase * 3;
 771                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 772                         CARP_DEBUG("%s: BACKUP -> MASTER "
 773                             "(master timed out)\n",
 774                             SC2IFP(sc)->if_xname);
 775                         carp_master_down_locked(sc);
 776                         break;
 777                 }
 778
 779                 /*
 780                  * Otherwise, we reset the counter and wait for the next
 781                  * advertisement.
 782                  */
 783                 carp_setrun(sc, af);
 784                 break;
 785         }
 786
 787         CARP_UNLOCK(ifp->if_carp);
 788
 789         m_freem(m);
 790         return;
 791 }
 792
 793 static int
 794 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 795 {
 796         struct m_tag *mtag;
 797         struct ifnet *ifp = SC2IFP(sc);
 798
 799         if (sc->sc_init_counter) {
 800                 /* this could also be seconds since unix epoch */
 801                 sc->sc_counter = arc4random();
 802                 sc->sc_counter = sc->sc_counter << 32;
 803                 sc->sc_counter += arc4random();
 804         } else
 805                 sc->sc_counter++;
 806
 807         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 808         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 809
 810         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 811
 812         /* Tag packet for carp_output */
 813         mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct ifnet *), M_NOWAIT);
 814         if (mtag == NULL) {
 815                 m_freem(m);
 816                 SC2IFP(sc)->if_oerrors++;
 817                 return (ENOMEM);
 818         }
 819         bcopy(&ifp, (caddr_t)(mtag + 1), sizeof(struct ifnet *));
 820         m_tag_prepend(m, mtag);
 821
 822         return (0);
 823 }
 824
 825 static void
 826 carp_send_ad_all(void)
 827 {
 828         struct carp_softc *sc;
 829
 830         mtx_lock(&carp_mtx);
 831         LIST_FOREACH(sc, &carpif_list, sc_next) {
 832                 if (sc->sc_carpdev == NULL)
 833                         continue;
 834                 CARP_SCLOCK(sc);
 835                 if ((SC2IFP(sc)->if_flags & IFF_UP) &&
 836                     (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING) &&
 837                      sc->sc_state == MASTER)
 838                         carp_send_ad_locked(sc);
 839                 CARP_SCUNLOCK(sc);
 840         }
 841         mtx_unlock(&carp_mtx);
 842 }
 843
 844 static void
 845 carp_send_ad(void *v)
 846 {
 847         struct carp_softc *sc = v;
 848
 849         CARP_SCLOCK(sc);
 850         carp_send_ad_locked(sc);
 851         CARP_SCUNLOCK(sc);
 852 }
 853
 854 static void
 855 carp_send_ad_locked(struct carp_softc *sc)
 856 {
 857         struct carp_header ch;
 858         struct timeval tv;
 859         struct carp_header *ch_ptr;
 860         struct mbuf *m;
 861         int len, advbase, advskew;
 862
 863         CARP_SCLOCK_ASSERT(sc);
 864
 865         /* bow out if we've lost our UPness or RUNNINGuiness */
 866         if (!((SC2IFP(sc)->if_flags & IFF_UP) &&
 867             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING))) {
 868                 advbase = 255;
 869                 advskew = 255;
 870         } else {
 871                 advbase = sc->sc_advbase;
 872                 if (!carp_suppress_preempt || sc->sc_advskew > 240)
 873                         advskew = sc->sc_advskew;
 874                 else
 875                         advskew = 240;
 876                 tv.tv_sec = advbase;
 877                 tv.tv_usec = advskew * 1000000 / 256;
 878         }
 879
 880         ch.carp_version = CARP_VERSION;
 881         ch.carp_type = CARP_ADVERTISEMENT;
 882         ch.carp_vhid = sc->sc_vhid;
 883         ch.carp_advbase = advbase;
 884         ch.carp_advskew = advskew;
 885         ch.carp_authlen = 7;    /* XXX DEFINE */
 886         ch.carp_pad1 = 0;       /* must be zero */
 887         ch.carp_cksum = 0;
 888
 889 #ifdef INET
 890         if (sc->sc_ia) {
 891                 struct ip *ip;
 892
 893                 MGETHDR(m, M_DONTWAIT, MT_HEADER);
 894                 if (m == NULL) {
 895                         SC2IFP(sc)->if_oerrors++;
 896                         carpstats.carps_onomem++;
 897                         /* XXX maybe less ? */
 898                         if (advbase != 255 || advskew != 255)
 899                                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
 900                                     carp_send_ad, sc);
 901                         return;
 902                 }
 903                 len = sizeof(*ip) + sizeof(ch);
 904                 m->m_pkthdr.len = len;
 905                 m->m_pkthdr.rcvif = NULL;
 906                 m->m_len = len;
 907                 MH_ALIGN(m, m->m_len);
 908                 m->m_flags |= M_MCAST;
 909                 ip = mtod(m, struct ip *);
 910                 ip->ip_v = IPVERSION;
 911                 ip->ip_hl = sizeof(*ip) >> 2;
 912                 ip->ip_tos = IPTOS_LOWDELAY;
 913                 ip->ip_len = len;
 914                 ip->ip_id = ip_newid();
 915                 ip->ip_off = IP_DF;
 916                 ip->ip_ttl = CARP_DFLTTL;
 917                 ip->ip_p = IPPROTO_CARP;
 918                 ip->ip_sum = 0;
 919                 ip->ip_src.s_addr = sc->sc_ia->ia_addr.sin_addr.s_addr;
 920                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 921
 922                 ch_ptr = (struct carp_header *)(&ip[1]);
 923                 bcopy(&ch, ch_ptr, sizeof(ch));
 924                 if (carp_prepare_ad(m, sc, ch_ptr))
 925                         return;
 926
 927                 m->m_data += sizeof(*ip);
 928                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip));
 929                 m->m_data -= sizeof(*ip);
 930
 931                 getmicrotime(&SC2IFP(sc)->if_lastchange);
 932                 SC2IFP(sc)->if_opackets++;
 933                 SC2IFP(sc)->if_obytes += len;
 934                 carpstats.carps_opackets++;
 935
 936                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, NULL)) {
 937                         SC2IFP(sc)->if_oerrors++;
 938                         if (sc->sc_sendad_errors < INT_MAX)
 939                                 sc->sc_sendad_errors++;
 940                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 941                                 carp_suppress_preempt++;
 942                                 if (carp_suppress_preempt == 1) {
 943                                         CARP_SCUNLOCK(sc);
 944                                         carp_send_ad_all();
 945                                         CARP_SCLOCK(sc);
 946                                 }
 947                         }
 948                         sc->sc_sendad_success = 0;
 949                 } else {
 950                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 951                                 if (++sc->sc_sendad_success >=
 952                                     CARP_SENDAD_MIN_SUCCESS) {
 953                                         carp_suppress_preempt--;
 954                                         sc->sc_sendad_errors = 0;
 955                                 }
 956                         } else
 957                                 sc->sc_sendad_errors = 0;
 958                 }
 959         }
 960 #endif /* INET */
 961 #ifdef INET6
 962         if (sc->sc_ia6) {
 963                 struct ip6_hdr *ip6;
 964
 965                 MGETHDR(m, M_DONTWAIT, MT_HEADER);
 966                 if (m == NULL) {
 967                         SC2IFP(sc)->if_oerrors++;
 968                         carpstats.carps_onomem++;
 969                         /* XXX maybe less ? */
 970                         if (advbase != 255 || advskew != 255)
 971                                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
 972                                     carp_send_ad, sc);
 973                         return;
 974                 }
 975                 len = sizeof(*ip6) + sizeof(ch);
 976                 m->m_pkthdr.len = len;
 977                 m->m_pkthdr.rcvif = NULL;
 978                 m->m_len = len;
 979                 MH_ALIGN(m, m->m_len);
 980                 m->m_flags |= M_MCAST;
 981                 ip6 = mtod(m, struct ip6_hdr *);
 982                 bzero(ip6, sizeof(*ip6));
 983                 ip6->ip6_vfc |= IPV6_VERSION;
 984                 ip6->ip6_hlim = CARP_DFLTTL;
 985                 ip6->ip6_nxt = IPPROTO_CARP;
 986                 bcopy(&sc->sc_ia6->ia_addr.sin6_addr, &ip6->ip6_src,
 987                     sizeof(struct in6_addr));
 988                 /* set the multicast destination */
 989
 990                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
 991                 ip6->ip6_dst.s6_addr8[15] = 0x12;
 992                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
 993                         SC2IFP(sc)->if_oerrors++;
 994                         m_freem(m);
 995                         CARP_LOG("%s: in6_setscope failed\n", __func__);
 996                         return;
 997                 }
 998
 999                 ch_ptr = (struct carp_header *)(&ip6[1]);
1000                 bcopy(&ch, ch_ptr, sizeof(ch));
1001                 if (carp_prepare_ad(m, sc, ch_ptr))
1002                         return;
1003
1004                 m->m_data += sizeof(*ip6);
1005                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6));
1006                 m->m_data -= sizeof(*ip6);
1007
1008                 getmicrotime(&SC2IFP(sc)->if_lastchange);
1009                 SC2IFP(sc)->if_opackets++;
1010                 SC2IFP(sc)->if_obytes += len;
1011                 carpstats.carps_opackets6++;
1012
1013                 if (ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL)) {
1014                         SC2IFP(sc)->if_oerrors++;
1015                         if (sc->sc_sendad_errors < INT_MAX)
1016                                 sc->sc_sendad_errors++;
1017                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
1018                                 carp_suppress_preempt++;
1019                                 if (carp_suppress_preempt == 1) {
1020                                         CARP_SCUNLOCK(sc);
1021                                         carp_send_ad_all();
1022                                         CARP_SCLOCK(sc);
1023                                 }
1024                         }
1025                         sc->sc_sendad_success = 0;
1026                 } else {
1027                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
1028                                 if (++sc->sc_sendad_success >=
1029                                     CARP_SENDAD_MIN_SUCCESS) {
1030                                         carp_suppress_preempt--;
1031                                         sc->sc_sendad_errors = 0;
1032                                 }
1033                         } else
1034                                 sc->sc_sendad_errors = 0;
1035                 }
1036         }
1037 #endif /* INET6 */
1038
1039         if (advbase != 255 || advskew != 255)
1040                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1041                     carp_send_ad, sc);
1042
1043 }
1044
1045 /*
1046  * Broadcast a gratuitous ARP request containing
1047  * the virtual router MAC address for each IP address
1048  * associated with the virtual router.
1049  */
1050 static void
1051 carp_send_arp(struct carp_softc *sc)
1052 {
1053         struct ifaddr *ifa;
1054
1055         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1056
1057                 if (ifa->ifa_addr->sa_family != AF_INET)
1058                         continue;
1059
1060 /*              arprequest(sc->sc_carpdev, &in, &in, IF_LLADDR(sc->sc_ifp)); */
1061                 arp_ifinit2(sc->sc_carpdev, ifa, IF_LLADDR(sc->sc_ifp));
1062
1063                 DELAY(1000);    /* XXX */
1064         }
1065 }
1066
1067 #ifdef INET6
1068 static void
1069 carp_send_na(struct carp_softc *sc)
1070 {
1071         struct ifaddr *ifa;
1072         struct in6_addr *in6;
1073         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1074
1075         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1076
1077                 if (ifa->ifa_addr->sa_family != AF_INET6)
1078                         continue;
1079
1080                 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr;
1081                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1082                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1083                 DELAY(1000);    /* XXX */
1084         }
1085 }
1086 #endif /* INET6 */
1087
1088 static int
1089 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type)
1090 {
1091         struct carp_softc *vh;
1092         struct ifaddr *ifa;
1093         int count = 0;
1094
1095         CARP_LOCK_ASSERT(cif);
1096
1097         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1098                 if ((type == CARP_COUNT_RUNNING &&
1099                     (SC2IFP(vh)->if_flags & IFF_UP) &&
1100                     (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING)) ||
1101                     (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) {
1102                         TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1103                             ifa_list) {
1104                                 if (ifa->ifa_addr->sa_family == AF_INET &&
1105                                     ia->ia_addr.sin_addr.s_addr ==
1106                                     ifatoia(ifa)->ia_addr.sin_addr.s_addr)
1107                                         count++;
1108                         }
1109                 }
1110         }
1111         return (count);
1112 }
1113
1114 int
1115 carp_iamatch(void *v, struct in_ifaddr *ia,
1116     struct in_addr *isaddr, u_int8_t **enaddr)
1117 {
1118         struct carp_if *cif = v;
1119         struct carp_softc *vh;
1120         int index, count = 0;
1121         struct ifaddr *ifa;
1122
1123         CARP_LOCK(cif);
1124
1125         if (carp_opts[CARPCTL_ARPBALANCE]) {
1126                 /*
1127                  * XXX proof of concept implementation.
1128                  * We use the source ip to decide which virtual host should
1129                  * handle the request. If we're master of that virtual host,
1130                  * then we respond, otherwise, just drop the arp packet on
1131                  * the floor.
1132                  */
1133                 count = carp_addrcount(cif, ia, CARP_COUNT_RUNNING);
1134                 if (count == 0) {
1135                         /* should never reach this */
1136                         CARP_UNLOCK(cif);
1137                         return (0);
1138                 }
1139
1140                 /* this should be a hash, like pf_hash() */
1141                 index = ntohl(isaddr->s_addr) % count;
1142                 count = 0;
1143
1144                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1145                         if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1146                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING)) {
1147                                 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1148                                     ifa_list) {
1149                                         if (ifa->ifa_addr->sa_family ==
1150                                             AF_INET &&
1151                                             ia->ia_addr.sin_addr.s_addr ==
1152                                             ifatoia(ifa)->ia_addr.sin_addr.s_addr) {
1153                                                 if (count == index) {
1154                                                         if (vh->sc_state ==
1155                                                             MASTER) {
1156                                                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1157                                                                 CARP_UNLOCK(cif);
1158                                                                 return (1);
1159                                                         } else {
1160                                                                 CARP_UNLOCK(cif);
1161                                                                 return (0);
1162                                                         }
1163                                                 }
1164                                                 count++;
1165                                         }
1166                                 }
1167                         }
1168                 }
1169         } else {
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                             ia->ia_ifp == SC2IFP(vh) &&
1174                             vh->sc_state == MASTER) {
1175                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1176                                 CARP_UNLOCK(cif);
1177                                 return (1);
1178                         }
1179                 }
1180         }
1181         CARP_UNLOCK(cif);
1182         return (0);
1183 }
1184
1185 #ifdef INET6
1186 struct ifaddr *
1187 carp_iamatch6(void *v, struct in6_addr *taddr)
1188 {
1189         struct carp_if *cif = v;
1190         struct carp_softc *vh;
1191         struct ifaddr *ifa;
1192
1193         CARP_LOCK(cif);
1194         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1195                 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist, ifa_list) {
1196                         if (IN6_ARE_ADDR_EQUAL(taddr,
1197                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1198                             (SC2IFP(vh)->if_flags & IFF_UP) &&
1199                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1200                             vh->sc_state == MASTER) {
1201                                 CARP_UNLOCK(cif);
1202                                 return (ifa);
1203                         }
1204                 }
1205         }
1206         CARP_UNLOCK(cif);
1207
1208         return (NULL);
1209 }
1210
1211 void *
1212 carp_macmatch6(void *v, struct mbuf *m, const struct in6_addr *taddr)
1213 {
1214         struct m_tag *mtag;
1215         struct carp_if *cif = v;
1216         struct carp_softc *sc;
1217         struct ifaddr *ifa;
1218
1219         CARP_LOCK(cif);
1220         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) {
1221                 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1222                         if (IN6_ARE_ADDR_EQUAL(taddr,
1223                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1224                             (SC2IFP(sc)->if_flags & IFF_UP) &&
1225                             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING)) {
1226                                 struct ifnet *ifp = SC2IFP(sc);
1227                                 mtag = m_tag_get(PACKET_TAG_CARP,
1228                                     sizeof(struct ifnet *), M_NOWAIT);
1229                                 if (mtag == NULL) {
1230                                         /* better a bit than nothing */
1231                                         CARP_UNLOCK(cif);
1232                                         return (IF_LLADDR(sc->sc_ifp));
1233                                 }
1234                                 bcopy(&ifp, (caddr_t)(mtag + 1),
1235                                     sizeof(struct ifnet *));
1236                                 m_tag_prepend(m, mtag);
1237
1238                                 CARP_UNLOCK(cif);
1239                                 return (IF_LLADDR(sc->sc_ifp));
1240                         }
1241                 }
1242         }
1243         CARP_UNLOCK(cif);
1244
1245         return (NULL);
1246 }
1247 #endif
1248
1249 struct ifnet *
1250 carp_forus(void *v, void *dhost)
1251 {
1252         struct carp_if *cif = v;
1253         struct carp_softc *vh;
1254         u_int8_t *ena = dhost;
1255
1256         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1257                 return (NULL);
1258
1259         CARP_LOCK(cif);
1260         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list)
1261                 if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1262                     (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1263                     vh->sc_state == MASTER &&
1264                     !bcmp(dhost, IF_LLADDR(vh->sc_ifp), ETHER_ADDR_LEN)) {
1265                         CARP_UNLOCK(cif);
1266                         return (SC2IFP(vh));
1267                 }
1268
1269         CARP_UNLOCK(cif);
1270         return (NULL);
1271 }
1272
1273 static void
1274 carp_master_down(void *v)
1275 {
1276         struct carp_softc *sc = v;
1277
1278         CARP_SCLOCK(sc);
1279         carp_master_down_locked(sc);
1280         CARP_SCUNLOCK(sc);
1281 }
1282
1283 static void
1284 carp_master_down_locked(struct carp_softc *sc)
1285 {
1286         if (sc->sc_carpdev)
1287                 CARP_SCLOCK_ASSERT(sc);
1288
1289         switch (sc->sc_state) {
1290         case INIT:
1291                 printf("%s: master_down event in INIT state\n",
1292                     SC2IFP(sc)->if_xname);
1293                 break;
1294         case MASTER:
1295                 break;
1296         case BACKUP:
1297                 carp_set_state(sc, MASTER);
1298                 carp_send_ad_locked(sc);
1299                 carp_send_arp(sc);
1300 #ifdef INET6
1301                 carp_send_na(sc);
1302 #endif /* INET6 */
1303                 carp_setrun(sc, 0);
1304                 carp_setroute(sc, RTM_ADD);
1305                 break;
1306         }
1307 }
1308
1309 /*
1310  * When in backup state, af indicates whether to reset the master down timer
1311  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1312  */
1313 static void
1314 carp_setrun(struct carp_softc *sc, sa_family_t af)
1315 {
1316         struct timeval tv;
1317
1318         if (sc->sc_carpdev == NULL) {
1319                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1320                 carp_set_state(sc, INIT);
1321                 return;
1322         } else
1323                 CARP_SCLOCK_ASSERT(sc);
1324
1325         if (SC2IFP(sc)->if_flags & IFF_UP &&
1326             sc->sc_vhid > 0 && (sc->sc_naddrs || sc->sc_naddrs6))
1327                 SC2IFP(sc)->if_drv_flags |= IFF_DRV_RUNNING;
1328         else {
1329                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1330                 carp_setroute(sc, RTM_DELETE);
1331                 return;
1332         }
1333
1334         switch (sc->sc_state) {
1335         case INIT:
1336                 if (carp_opts[CARPCTL_PREEMPT] && !carp_suppress_preempt) {
1337                         carp_send_ad_locked(sc);
1338                         carp_send_arp(sc);
1339 #ifdef INET6
1340                         carp_send_na(sc);
1341 #endif /* INET6 */
1342                         CARP_DEBUG("%s: INIT -> MASTER (preempting)\n",
1343                             SC2IFP(sc)->if_xname);
1344                         carp_set_state(sc, MASTER);
1345                         carp_setroute(sc, RTM_ADD);
1346                 } else {
1347                         CARP_DEBUG("%s: INIT -> BACKUP\n", SC2IFP(sc)->if_xname);
1348                         carp_set_state(sc, BACKUP);
1349                         carp_setroute(sc, RTM_DELETE);
1350                         carp_setrun(sc, 0);
1351                 }
1352                 break;
1353         case BACKUP:
1354                 callout_stop(&sc->sc_ad_tmo);
1355                 tv.tv_sec = 3 * sc->sc_advbase;
1356                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1357                 switch (af) {
1358 #ifdef INET
1359                 case AF_INET:
1360                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1361                             carp_master_down, sc);
1362                         break;
1363 #endif /* INET */
1364 #ifdef INET6
1365                 case AF_INET6:
1366                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1367                             carp_master_down, sc);
1368                         break;
1369 #endif /* INET6 */
1370                 default:
1371                         if (sc->sc_naddrs)
1372                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1373                                     carp_master_down, sc);
1374                         if (sc->sc_naddrs6)
1375                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1376                                     carp_master_down, sc);
1377                         break;
1378                 }
1379                 break;
1380         case MASTER:
1381                 tv.tv_sec = sc->sc_advbase;
1382                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1383                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1384                     carp_send_ad, sc);
1385                 break;
1386         }
1387 }
1388
1389 static void
1390 carp_multicast_cleanup(struct carp_softc *sc)
1391 {
1392         struct ip_moptions *imo = &sc->sc_imo;
1393         u_int16_t n = imo->imo_num_memberships;
1394
1395         /* Clean up our own multicast memberships */
1396         while (n-- > 0) {
1397                 if (imo->imo_membership[n] != NULL) {
1398                         in_delmulti(imo->imo_membership[n]);
1399                         imo->imo_membership[n] = NULL;
1400                 }
1401         }
1402         KASSERT(imo->imo_mfilters == NULL,
1403            ("%s: imo_mfilters != NULL", __func__));
1404         imo->imo_num_memberships = 0;
1405         imo->imo_multicast_ifp = NULL;
1406 }
1407
1408 #ifdef INET6
1409 static void
1410 carp_multicast6_cleanup(struct carp_softc *sc)
1411 {
1412         struct ip6_moptions *im6o = &sc->sc_im6o;
1413
1414         while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1415                 struct in6_multi_mship *imm =
1416                     LIST_FIRST(&im6o->im6o_memberships);
1417
1418                 LIST_REMOVE(imm, i6mm_chain);
1419                 in6_leavegroup(imm);
1420         }
1421         im6o->im6o_multicast_ifp = NULL;
1422 }
1423 #endif
1424
1425 static int
1426 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1427 {
1428         struct ifnet *ifp;
1429         struct carp_if *cif;
1430         struct in_ifaddr *ia, *ia_if;
1431         struct ip_moptions *imo = &sc->sc_imo;
1432         struct in_addr addr;
1433         u_long iaddr = htonl(sin->sin_addr.s_addr);
1434         int own, error;
1435
1436         if (sin->sin_addr.s_addr == 0) {
1437                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1438                         carp_set_state(sc, INIT);
1439                 if (sc->sc_naddrs)
1440                         SC2IFP(sc)->if_flags |= IFF_UP;
1441                 carp_setrun(sc, 0);
1442                 return (0);
1443         }
1444
1445         /* we have to do it by hands to check we won't match on us */
1446         ia_if = NULL; own = 0;
1447         TAILQ_FOREACH(ia, &in_ifaddrhead, ia_link) {
1448                 /* and, yeah, we need a multicast-capable iface too */
1449                 if (ia->ia_ifp != SC2IFP(sc) &&
1450                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1451                     (iaddr & ia->ia_subnetmask) == ia->ia_subnet) {
1452                         if (!ia_if)
1453                                 ia_if = ia;
1454                         if (sin->sin_addr.s_addr ==
1455                             ia->ia_addr.sin_addr.s_addr)
1456                                 own++;
1457                 }
1458         }
1459
1460         if (!ia_if)
1461                 return (EADDRNOTAVAIL);
1462
1463         ia = ia_if;
1464         ifp = ia->ia_ifp;
1465
1466         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1467             (imo->imo_multicast_ifp && imo->imo_multicast_ifp != ifp))
1468                 return (EADDRNOTAVAIL);
1469
1470         if (imo->imo_num_memberships == 0) {
1471                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1472                 if ((imo->imo_membership[0] = in_addmulti(&addr, ifp)) == NULL)
1473                         return (ENOBUFS);
1474                 imo->imo_num_memberships++;
1475                 imo->imo_multicast_ifp = ifp;
1476                 imo->imo_multicast_ttl = CARP_DFLTTL;
1477                 imo->imo_multicast_loop = 0;
1478         }
1479
1480         if (!ifp->if_carp) {
1481
1482                 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP,
1483                     M_WAITOK|M_ZERO);
1484                 if (!cif) {
1485                         error = ENOBUFS;
1486                         goto cleanup;
1487                 }
1488                 if ((error = ifpromisc(ifp, 1))) {
1489                         FREE(cif, M_CARP);
1490                         goto cleanup;
1491                 }
1492
1493                 CARP_LOCK_INIT(cif);
1494                 CARP_LOCK(cif);
1495                 cif->vhif_ifp = ifp;
1496                 TAILQ_INIT(&cif->vhif_vrs);
1497                 ifp->if_carp = cif;
1498
1499         } else {
1500                 struct carp_softc *vr;
1501
1502                 cif = (struct carp_if *)ifp->if_carp;
1503                 CARP_LOCK(cif);
1504                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1505                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1506                                 CARP_UNLOCK(cif);
1507                                 error = EINVAL;
1508                                 goto cleanup;
1509                         }
1510         }
1511         sc->sc_ia = ia;
1512         sc->sc_carpdev = ifp;
1513
1514         { /* XXX prevent endless loop if already in queue */
1515         struct carp_softc *vr, *after = NULL;
1516         int myself = 0;
1517         cif = (struct carp_if *)ifp->if_carp;
1518
1519         /* XXX: cif should not change, right? So we still hold the lock */
1520         CARP_LOCK_ASSERT(cif);
1521
1522         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1523                 if (vr == sc)
1524                         myself = 1;
1525                 if (vr->sc_vhid < sc->sc_vhid)
1526                         after = vr;
1527         }
1528
1529         if (!myself) {
1530                 /* We're trying to keep things in order */
1531                 if (after == NULL) {
1532                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1533                 } else {
1534                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1535                 }
1536                 cif->vhif_nvrs++;
1537         }
1538         }
1539
1540         sc->sc_naddrs++;
1541         SC2IFP(sc)->if_flags |= IFF_UP;
1542         if (own)
1543                 sc->sc_advskew = 0;
1544         carp_sc_state_locked(sc);
1545         carp_setrun(sc, 0);
1546
1547         CARP_UNLOCK(cif);
1548
1549         return (0);
1550
1551 cleanup:
1552         in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1553         return (error);
1554 }
1555
1556 static int
1557 carp_del_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1558 {
1559         int error = 0;
1560
1561         if (!--sc->sc_naddrs) {
1562                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1563                 struct ip_moptions *imo = &sc->sc_imo;
1564
1565                 CARP_LOCK(cif);
1566                 callout_stop(&sc->sc_ad_tmo);
1567                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1568                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1569                 sc->sc_vhid = -1;
1570                 in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1571                 imo->imo_multicast_ifp = NULL;
1572                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1573                 if (!--cif->vhif_nvrs) {
1574                         sc->sc_carpdev->if_carp = NULL;
1575                         CARP_LOCK_DESTROY(cif);
1576                         FREE(cif, M_IFADDR);
1577                 } else {
1578                         CARP_UNLOCK(cif);
1579                 }
1580         }
1581
1582         return (error);
1583 }
1584
1585 #ifdef INET6
1586 static int
1587 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1588 {
1589         struct ifnet *ifp;
1590         struct carp_if *cif;
1591         struct in6_ifaddr *ia, *ia_if;
1592         struct ip6_moptions *im6o = &sc->sc_im6o;
1593         struct in6_multi_mship *imm;
1594         struct in6_addr in6;
1595         int own, error;
1596
1597         if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
1598                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1599                         carp_set_state(sc, INIT);
1600                 if (sc->sc_naddrs6)
1601                         SC2IFP(sc)->if_flags |= IFF_UP;
1602                 carp_setrun(sc, 0);
1603                 return (0);
1604         }
1605
1606         /* we have to do it by hands to check we won't match on us */
1607         ia_if = NULL; own = 0;
1608         for (ia = in6_ifaddr; ia; ia = ia->ia_next) {
1609                 int i;
1610
1611                 for (i = 0; i < 4; i++) {
1612                         if ((sin6->sin6_addr.s6_addr32[i] &
1613                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]) !=
1614                             (ia->ia_addr.sin6_addr.s6_addr32[i] &
1615                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]))
1616                                 break;
1617                 }
1618                 /* and, yeah, we need a multicast-capable iface too */
1619                 if (ia->ia_ifp != SC2IFP(sc) &&
1620                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1621                     (i == 4)) {
1622                         if (!ia_if)
1623                                 ia_if = ia;
1624                         if (IN6_ARE_ADDR_EQUAL(&sin6->sin6_addr,
1625                             &ia->ia_addr.sin6_addr))
1626                                 own++;
1627                 }
1628         }
1629
1630         if (!ia_if)
1631                 return (EADDRNOTAVAIL);
1632         ia = ia_if;
1633         ifp = ia->ia_ifp;
1634
1635         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1636             (im6o->im6o_multicast_ifp && im6o->im6o_multicast_ifp != ifp))
1637                 return (EADDRNOTAVAIL);
1638
1639         if (!sc->sc_naddrs6) {
1640                 im6o->im6o_multicast_ifp = ifp;
1641
1642                 /* join CARP multicast address */
1643                 bzero(&in6, sizeof(in6));
1644                 in6.s6_addr16[0] = htons(0xff02);
1645                 in6.s6_addr8[15] = 0x12;
1646                 if (in6_setscope(&in6, ifp, NULL) != 0)
1647                         goto cleanup;
1648                 if ((imm = in6_joingroup(ifp, &in6, &error, 0)) == NULL)
1649                         goto cleanup;
1650                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1651
1652                 /* join solicited multicast address */
1653                 bzero(&in6, sizeof(in6));
1654                 in6.s6_addr16[0] = htons(0xff02);
1655                 in6.s6_addr32[1] = 0;
1656                 in6.s6_addr32[2] = htonl(1);
1657                 in6.s6_addr32[3] = sin6->sin6_addr.s6_addr32[3];
1658                 in6.s6_addr8[12] = 0xff;
1659                 if (in6_setscope(&in6, ifp, NULL) != 0)
1660                         goto cleanup;
1661                 if ((imm = in6_joingroup(ifp, &in6, &error, 0)) == NULL)
1662                         goto cleanup;
1663                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1664         }
1665
1666         if (!ifp->if_carp) {
1667                 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP,
1668                     M_WAITOK|M_ZERO);
1669                 if (!cif) {
1670                         error = ENOBUFS;
1671                         goto cleanup;
1672                 }
1673                 if ((error = ifpromisc(ifp, 1))) {
1674                         FREE(cif, M_CARP);
1675                         goto cleanup;
1676                 }
1677
1678                 CARP_LOCK_INIT(cif);
1679                 CARP_LOCK(cif);
1680                 cif->vhif_ifp = ifp;
1681                 TAILQ_INIT(&cif->vhif_vrs);
1682                 ifp->if_carp = cif;
1683
1684         } else {
1685                 struct carp_softc *vr;
1686
1687                 cif = (struct carp_if *)ifp->if_carp;
1688                 CARP_LOCK(cif);
1689                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1690                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1691                                 CARP_UNLOCK(cif);
1692                                 error = EINVAL;
1693                                 goto cleanup;
1694                         }
1695         }
1696         sc->sc_ia6 = ia;
1697         sc->sc_carpdev = ifp;
1698
1699         { /* XXX prevent endless loop if already in queue */
1700         struct carp_softc *vr, *after = NULL;
1701         int myself = 0;
1702         cif = (struct carp_if *)ifp->if_carp;
1703         CARP_LOCK_ASSERT(cif);
1704
1705         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1706                 if (vr == sc)
1707                         myself = 1;
1708                 if (vr->sc_vhid < sc->sc_vhid)
1709                         after = vr;
1710         }
1711
1712         if (!myself) {
1713                 /* We're trying to keep things in order */
1714                 if (after == NULL) {
1715                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1716                 } else {
1717                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1718                 }
1719                 cif->vhif_nvrs++;
1720         }
1721         }
1722
1723         sc->sc_naddrs6++;
1724         SC2IFP(sc)->if_flags |= IFF_UP;
1725         if (own)
1726                 sc->sc_advskew = 0;
1727         carp_sc_state_locked(sc);
1728         carp_setrun(sc, 0);
1729
1730         CARP_UNLOCK(cif);
1731
1732         return (0);
1733
1734 cleanup:
1735         /* clean up multicast memberships */
1736         if (!sc->sc_naddrs6) {
1737                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1738                         imm = LIST_FIRST(&im6o->im6o_memberships);
1739                         LIST_REMOVE(imm, i6mm_chain);
1740                         in6_leavegroup(imm);
1741                 }
1742         }
1743         return (error);
1744 }
1745
1746 static int
1747 carp_del_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1748 {
1749         int error = 0;
1750
1751         if (!--sc->sc_naddrs6) {
1752                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1753                 struct ip6_moptions *im6o = &sc->sc_im6o;
1754
1755                 CARP_LOCK(cif);
1756                 callout_stop(&sc->sc_ad_tmo);
1757                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1758                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1759                 sc->sc_vhid = -1;
1760                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1761                         struct in6_multi_mship *imm =
1762                             LIST_FIRST(&im6o->im6o_memberships);
1763
1764                         LIST_REMOVE(imm, i6mm_chain);
1765                         in6_leavegroup(imm);
1766                 }
1767                 im6o->im6o_multicast_ifp = NULL;
1768                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1769                 if (!--cif->vhif_nvrs) {
1770                         CARP_LOCK_DESTROY(cif);
1771                         sc->sc_carpdev->if_carp = NULL;
1772                         FREE(cif, M_IFADDR);
1773                 } else
1774                         CARP_UNLOCK(cif);
1775         }
1776
1777         return (error);
1778 }
1779 #endif /* INET6 */
1780
1781 static int
1782 carp_ioctl(struct ifnet *ifp, u_long cmd, caddr_t addr)
1783 {
1784         struct carp_softc *sc = ifp->if_softc, *vr;
1785         struct carpreq carpr;
1786         struct ifaddr *ifa;
1787         struct ifreq *ifr;
1788         struct ifaliasreq *ifra;
1789         int locked = 0, error = 0;
1790
1791         ifa = (struct ifaddr *)addr;
1792         ifra = (struct ifaliasreq *)addr;
1793         ifr = (struct ifreq *)addr;
1794
1795         switch (cmd) {
1796         case SIOCSIFADDR:
1797                 switch (ifa->ifa_addr->sa_family) {
1798 #ifdef INET
1799                 case AF_INET:
1800                         SC2IFP(sc)->if_flags |= IFF_UP;
1801                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1802                             sizeof(struct sockaddr));
1803                         error = carp_set_addr(sc, satosin(ifa->ifa_addr));
1804                         break;
1805 #endif /* INET */
1806 #ifdef INET6
1807                 case AF_INET6:
1808                         SC2IFP(sc)->if_flags |= IFF_UP;
1809                         error = carp_set_addr6(sc, satosin6(ifa->ifa_addr));
1810                         break;
1811 #endif /* INET6 */
1812                 default:
1813                         error = EAFNOSUPPORT;
1814                         break;
1815                 }
1816                 break;
1817
1818         case SIOCAIFADDR:
1819                 switch (ifa->ifa_addr->sa_family) {
1820 #ifdef INET
1821                 case AF_INET:
1822                         SC2IFP(sc)->if_flags |= IFF_UP;
1823                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1824                             sizeof(struct sockaddr));
1825                         error = carp_set_addr(sc, satosin(&ifra->ifra_addr));
1826                         break;
1827 #endif /* INET */
1828 #ifdef INET6
1829                 case AF_INET6:
1830                         SC2IFP(sc)->if_flags |= IFF_UP;
1831                         error = carp_set_addr6(sc, satosin6(&ifra->ifra_addr));
1832                         break;
1833 #endif /* INET6 */
1834                 default:
1835                         error = EAFNOSUPPORT;
1836                         break;
1837                 }
1838                 break;
1839
1840         case SIOCDIFADDR:
1841                 switch (ifa->ifa_addr->sa_family) {
1842 #ifdef INET
1843                 case AF_INET:
1844                         error = carp_del_addr(sc, satosin(&ifra->ifra_addr));
1845                         break;
1846 #endif /* INET */
1847 #ifdef INET6
1848                 case AF_INET6:
1849                         error = carp_del_addr6(sc, satosin6(&ifra->ifra_addr));
1850                         break;
1851 #endif /* INET6 */
1852                 default:
1853                         error = EAFNOSUPPORT;
1854                         break;
1855                 }
1856                 break;
1857
1858         case SIOCSIFFLAGS:
1859                 if (sc->sc_carpdev) {
1860                         locked = 1;
1861                         CARP_SCLOCK(sc);
1862                 }
1863                 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) {
1864                         callout_stop(&sc->sc_ad_tmo);
1865                         callout_stop(&sc->sc_md_tmo);
1866                         callout_stop(&sc->sc_md6_tmo);
1867                         if (sc->sc_state == MASTER)
1868                                 carp_send_ad_locked(sc);
1869                         carp_set_state(sc, INIT);
1870                         carp_setrun(sc, 0);
1871                 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) {
1872                         SC2IFP(sc)->if_flags |= IFF_UP;
1873                         carp_setrun(sc, 0);
1874                 }
1875                 break;
1876
1877         case SIOCSVH:
1878                 error = priv_check(curthread, PRIV_NETINET_CARP);
1879                 if (error)
1880                         break;
1881                 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1882                         break;
1883                 error = 1;
1884                 if (sc->sc_carpdev) {
1885                         locked = 1;
1886                         CARP_SCLOCK(sc);
1887                 }
1888                 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) {
1889                         switch (carpr.carpr_state) {
1890                         case BACKUP:
1891                                 callout_stop(&sc->sc_ad_tmo);
1892                                 carp_set_state(sc, BACKUP);
1893                                 carp_setrun(sc, 0);
1894                                 carp_setroute(sc, RTM_DELETE);
1895                                 break;
1896                         case MASTER:
1897                                 carp_master_down_locked(sc);
1898                                 break;
1899                         default:
1900                                 break;
1901                         }
1902                 }
1903                 if (carpr.carpr_vhid > 0) {
1904                         if (carpr.carpr_vhid > 255) {
1905                                 error = EINVAL;
1906                                 break;
1907                         }
1908                         if (sc->sc_carpdev) {
1909                                 struct carp_if *cif;
1910                                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1911                                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1912                                         if (vr != sc &&
1913                                             vr->sc_vhid == carpr.carpr_vhid) {
1914                                                 error = EEXIST;
1915                                                 break;
1916                                         }
1917                                 if (error == EEXIST)
1918                                         break;
1919                         }
1920                         sc->sc_vhid = carpr.carpr_vhid;
1921                         IF_LLADDR(sc->sc_ifp)[0] = 0;
1922                         IF_LLADDR(sc->sc_ifp)[1] = 0;
1923                         IF_LLADDR(sc->sc_ifp)[2] = 0x5e;
1924                         IF_LLADDR(sc->sc_ifp)[3] = 0;
1925                         IF_LLADDR(sc->sc_ifp)[4] = 1;
1926                         IF_LLADDR(sc->sc_ifp)[5] = sc->sc_vhid;
1927                         error--;
1928                 }
1929                 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) {
1930                         if (carpr.carpr_advskew >= 255) {
1931                                 error = EINVAL;
1932                                 break;
1933                         }
1934                         if (carpr.carpr_advbase > 255) {
1935                                 error = EINVAL;
1936                                 break;
1937                         }
1938                         sc->sc_advbase = carpr.carpr_advbase;
1939                         sc->sc_advskew = carpr.carpr_advskew;
1940                         error--;
1941                 }
1942                 bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1943                 if (error > 0)
1944                         error = EINVAL;
1945                 else {
1946                         error = 0;
1947                         carp_setrun(sc, 0);
1948                 }
1949                 break;
1950
1951         case SIOCGVH:
1952                 /* XXX: lockless read */
1953                 bzero(&carpr, sizeof(carpr));
1954                 carpr.carpr_state = sc->sc_state;
1955                 carpr.carpr_vhid = sc->sc_vhid;
1956                 carpr.carpr_advbase = sc->sc_advbase;
1957                 carpr.carpr_advskew = sc->sc_advskew;
1958                 error = priv_check(curthread, PRIV_NETINET_CARP);
1959                 if (error == 0)
1960                         bcopy(sc->sc_key, carpr.carpr_key,
1961                             sizeof(carpr.carpr_key));
1962                 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1963                 break;
1964
1965         default:
1966                 error = EINVAL;
1967         }
1968
1969         if (locked)
1970                 CARP_SCUNLOCK(sc);
1971
1972         carp_hmac_prepare(sc);
1973
1974         return (error);
1975 }
1976
1977 /*
1978  * XXX: this is looutput. We should eventually use it from there.
1979  */
1980 static int
1981 carp_looutput(struct ifnet *ifp, struct mbuf *m, struct sockaddr *dst,
1982     struct rtentry *rt)
1983 {
1984         u_int32_t af;
1985
1986         M_ASSERTPKTHDR(m); /* check if we have the packet header */
1987
1988         if (rt && rt->rt_flags & (RTF_REJECT|RTF_BLACKHOLE)) {
1989                 m_freem(m);
1990                 return (rt->rt_flags & RTF_BLACKHOLE ? 0 :
1991                         rt->rt_flags & RTF_HOST ? EHOSTUNREACH : ENETUNREACH);
1992         }
1993
1994         ifp->if_opackets++;
1995         ifp->if_obytes += m->m_pkthdr.len;
1996
1997         /* BPF writes need to be handled specially. */
1998         if (dst->sa_family == AF_UNSPEC) {
1999                 bcopy(dst->sa_data, &af, sizeof(af));
2000                 dst->sa_family = af;
2001         }
2002
2003 #if 1   /* XXX */
2004         switch (dst->sa_family) {
2005         case AF_INET:
2006         case AF_INET6:
2007         case AF_IPX:
2008         case AF_APPLETALK:
2009                 break;
2010         default:
2011                 printf("carp_looutput: af=%d unexpected\n", dst->sa_family);
2012                 m_freem(m);
2013                 return (EAFNOSUPPORT);
2014         }
2015 #endif
2016         return(if_simloop(ifp, m, dst->sa_family, 0));
2017 }
2018
2019 /*
2020  * Start output on carp interface. This function should never be called.
2021  */
2022 static void
2023 carp_start(struct ifnet *ifp)
2024 {
2025 #ifdef DEBUG
2026         printf("%s: start called\n", ifp->if_xname);
2027 #endif
2028 }
2029
2030 int
2031 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *sa,
2032     struct rtentry *rt)
2033 {
2034         struct m_tag *mtag;
2035         struct carp_softc *sc;
2036         struct ifnet *carp_ifp;
2037
2038         if (!sa)
2039                 return (0);
2040
2041         switch (sa->sa_family) {
2042 #ifdef INET
2043         case AF_INET:
2044                 break;
2045 #endif /* INET */
2046 #ifdef INET6
2047         case AF_INET6:
2048                 break;
2049 #endif /* INET6 */
2050         default:
2051                 return (0);
2052         }
2053
2054         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
2055         if (mtag == NULL)
2056                 return (0);
2057
2058         bcopy(mtag + 1, &carp_ifp, sizeof(struct ifnet *));
2059         sc = carp_ifp->if_softc;
2060
2061         /* Set the source MAC address to Virtual Router MAC Address */
2062         switch (ifp->if_type) {
2063         case IFT_ETHER:
2064         case IFT_L2VLAN: {
2065                         struct ether_header *eh;
2066
2067                         eh = mtod(m, struct ether_header *);
2068                         eh->ether_shost[0] = 0;
2069                         eh->ether_shost[1] = 0;
2070                         eh->ether_shost[2] = 0x5e;
2071                         eh->ether_shost[3] = 0;
2072                         eh->ether_shost[4] = 1;
2073                         eh->ether_shost[5] = sc->sc_vhid;
2074                 }
2075                 break;
2076         case IFT_FDDI: {
2077                         struct fddi_header *fh;
2078
2079                         fh = mtod(m, struct fddi_header *);
2080                         fh->fddi_shost[0] = 0;
2081                         fh->fddi_shost[1] = 0;
2082                         fh->fddi_shost[2] = 0x5e;
2083                         fh->fddi_shost[3] = 0;
2084                         fh->fddi_shost[4] = 1;
2085                         fh->fddi_shost[5] = sc->sc_vhid;
2086                 }
2087                 break;
2088         case IFT_ISO88025: {
2089                         struct iso88025_header *th;
2090                         th = mtod(m, struct iso88025_header *);
2091                         th->iso88025_shost[0] = 3;
2092                         th->iso88025_shost[1] = 0;
2093                         th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
2094                         th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
2095                         th->iso88025_shost[4] = 0;
2096                         th->iso88025_shost[5] = 0;
2097                 }
2098                 break;
2099         default:
2100                 printf("%s: carp is not supported for this interface type\n",
2101                     ifp->if_xname);
2102                 return (EOPNOTSUPP);
2103         }
2104
2105         return (0);
2106 }
2107
2108 static void
2109 carp_set_state(struct carp_softc *sc, int state)
2110 {
2111
2112         if (sc->sc_carpdev)
2113                 CARP_SCLOCK_ASSERT(sc);
2114
2115         if (sc->sc_state == state)
2116                 return;
2117
2118         sc->sc_state = state;
2119         switch (state) {
2120         case BACKUP:
2121                 SC2IFP(sc)->if_link_state = LINK_STATE_DOWN;
2122                 break;
2123         case MASTER:
2124                 SC2IFP(sc)->if_link_state = LINK_STATE_UP;
2125                 break;
2126         default:
2127                 SC2IFP(sc)->if_link_state = LINK_STATE_UNKNOWN;
2128                 break;
2129         }
2130         rt_ifmsg(SC2IFP(sc));
2131 }
2132
2133 void
2134 carp_carpdev_state(void *v)
2135 {
2136         struct carp_if *cif = v;
2137
2138         CARP_LOCK(cif);
2139         carp_carpdev_state_locked(cif);
2140         CARP_UNLOCK(cif);
2141 }
2142
2143 static void
2144 carp_carpdev_state_locked(struct carp_if *cif)
2145 {
2146         struct carp_softc *sc;
2147
2148         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list)
2149                 carp_sc_state_locked(sc);
2150 }
2151
2152 static void
2153 carp_sc_state_locked(struct carp_softc *sc)
2154 {
2155         CARP_SCLOCK_ASSERT(sc);
2156
2157         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
2158             !(sc->sc_carpdev->if_flags & IFF_UP)) {
2159                 sc->sc_flags_backup = SC2IFP(sc)->if_flags;
2160                 SC2IFP(sc)->if_flags &= ~IFF_UP;
2161                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
2162                 callout_stop(&sc->sc_ad_tmo);
2163                 callout_stop(&sc->sc_md_tmo);
2164                 callout_stop(&sc->sc_md6_tmo);
2165                 carp_set_state(sc, INIT);
2166                 carp_setrun(sc, 0);
2167                 if (!sc->sc_suppress) {
2168                         carp_suppress_preempt++;
2169                         if (carp_suppress_preempt == 1) {
2170                                 CARP_SCUNLOCK(sc);
2171                                 carp_send_ad_all();
2172                                 CARP_SCLOCK(sc);
2173                         }
2174                 }
2175                 sc->sc_suppress = 1;
2176         } else {
2177                 SC2IFP(sc)->if_flags |= sc->sc_flags_backup;
2178                 carp_set_state(sc, INIT);
2179                 carp_setrun(sc, 0);
2180                 if (sc->sc_suppress)
2181                         carp_suppress_preempt--;
2182                 sc->sc_suppress = 0;
2183         }
2184
2185         return;
2186 }
2187
2188 static int
2189 carp_modevent(module_t mod, int type, void *data)
2190 {
2191         switch (type) {
2192         case MOD_LOAD:
2193                 if_detach_event_tag = EVENTHANDLER_REGISTER(ifnet_departure_event,
2194                     carp_ifdetach, NULL, EVENTHANDLER_PRI_ANY);
2195                 if (if_detach_event_tag == NULL)
2196                         return (ENOMEM);
2197                 mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2198                 LIST_INIT(&carpif_list);
2199                 if_clone_attach(&carp_cloner);
2200                 break;
2201
2202         case MOD_UNLOAD:
2203                 EVENTHANDLER_DEREGISTER(ifnet_departure_event, if_detach_event_tag);
2204                 if_clone_detach(&carp_cloner);
2205                 mtx_destroy(&carp_mtx);
2206                 break;
2207
2208         default:
2209                 return (EINVAL);
2210         }
2211
2212         return (0);
2213 }
2214
2215 static moduledata_t carp_mod = {
2216         "carp",
2217         carp_modevent,
2218         0
2219 };
2220
2221 DECLARE_MODULE(carp, carp_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);