sys/netinet/ip_carp.c

   1 /*      $FreeBSD$ */
   2
   3 /*
   4  * Copyright (c) 2002 Michael Shalayeff. All rights reserved.
   5  * Copyright (c) 2003 Ryan McBride. All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  17  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  18  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  19  * IN NO EVENT SHALL THE AUTHOR OR HIS RELATIVES BE LIABLE FOR ANY DIRECT,
  20  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  21  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  22  * SERVICES; LOSS OF MIND, USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
  24  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
  25  * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
  26  * THE POSSIBILITY OF SUCH DAMAGE.
  27  */
  28
  29 #include "opt_carp.h"
  30 #include "opt_bpf.h"
  31 #include "opt_inet.h"
  32 #include "opt_inet6.h"
  33
  34 #include <sys/types.h>
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/conf.h>
  38 #include <sys/kernel.h>
  39 #include <sys/limits.h>
  40 #include <sys/malloc.h>
  41 #include <sys/mbuf.h>
  42 #include <sys/module.h>
  43 #include <sys/time.h>
  44 #include <sys/priv.h>
  45 #include <sys/proc.h>
  46 #include <sys/sysctl.h>
  47 #include <sys/syslog.h>
  48 #include <sys/signalvar.h>
  49 #include <sys/filio.h>
  50 #include <sys/sockio.h>
  51
  52 #include <sys/socket.h>
  53 #include <sys/vnode.h>
  54
  55 #include <machine/stdarg.h>
  56
  57 #include <net/bpf.h>
  58 #include <net/ethernet.h>
  59 #include <net/fddi.h>
  60 #include <net/iso88025.h>
  61 #include <net/if.h>
  62 #include <net/if_clone.h>
  63 #include <net/if_dl.h>
  64 #include <net/if_types.h>
  65 #include <net/route.h>
  66
  67 #ifdef INET
  68 #include <netinet/in.h>
  69 #include <netinet/in_var.h>
  70 #include <netinet/in_systm.h>
  71 #include <netinet/ip.h>
  72 #include <netinet/ip_var.h>
  73 #include <netinet/if_ether.h>
  74 #include <machine/in_cksum.h>
  75 #endif
  76
  77 #ifdef INET6
  78 #include <netinet/icmp6.h>
  79 #include <netinet/ip6.h>
  80 #include <netinet6/ip6_var.h>
  81 #include <netinet6/scope6_var.h>
  82 #include <netinet6/nd6.h>
  83 #endif
  84
  85 #include <crypto/sha1.h>
  86 #include <netinet/ip_carp.h>
  87
  88 #define CARP_IFNAME     "carp"
  89 static MALLOC_DEFINE(M_CARP, "CARP", "CARP interfaces");
  90 SYSCTL_DECL(_net_inet_carp);
  91
  92 struct carp_softc {
  93         struct ifnet            *sc_ifp;        /* Interface clue */
  94         struct ifnet            *sc_carpdev;    /* Pointer to parent interface */
  95         struct in_ifaddr        *sc_ia;         /* primary iface address */
  96         struct ip_moptions       sc_imo;
  97 #ifdef INET6
  98         struct in6_ifaddr       *sc_ia6;        /* primary iface address v6 */
  99         struct ip6_moptions      sc_im6o;
 100 #endif /* INET6 */
 101         TAILQ_ENTRY(carp_softc)  sc_list;
 102
 103         enum { INIT = 0, BACKUP, MASTER }       sc_state;
 104
 105         int                      sc_flags_backup;
 106         int                      sc_suppress;
 107
 108         int                      sc_sendad_errors;
 109 #define CARP_SENDAD_MAX_ERRORS  3
 110         int                      sc_sendad_success;
 111 #define CARP_SENDAD_MIN_SUCCESS 3
 112
 113         int                      sc_vhid;
 114         int                      sc_advskew;
 115         int                      sc_naddrs;
 116         int                      sc_naddrs6;
 117         int                      sc_advbase;    /* seconds */
 118         int                      sc_init_counter;
 119         u_int64_t                sc_counter;
 120
 121         /* authentication */
 122 #define CARP_HMAC_PAD   64
 123         unsigned char sc_key[CARP_KEY_LEN];
 124         unsigned char sc_pad[CARP_HMAC_PAD];
 125         SHA1_CTX sc_sha1;
 126
 127         struct callout           sc_ad_tmo;     /* advertisement timeout */
 128         struct callout           sc_md_tmo;     /* master down timeout */
 129         struct callout           sc_md6_tmo;    /* master down timeout */
 130
 131         LIST_ENTRY(carp_softc)   sc_next;       /* Interface clue */
 132 };
 133 #define SC2IFP(sc)      ((sc)->sc_ifp)
 134
 135 int carp_suppress_preempt = 0;
 136 int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 1, 0, 0 };    /* XXX for now */
 137 SYSCTL_INT(_net_inet_carp, CARPCTL_ALLOW, allow, CTLFLAG_RW,
 138     &carp_opts[CARPCTL_ALLOW], 0, "Accept incoming CARP packets");
 139 SYSCTL_INT(_net_inet_carp, CARPCTL_PREEMPT, preempt, CTLFLAG_RW,
 140     &carp_opts[CARPCTL_PREEMPT], 0, "high-priority backup preemption mode");
 141 SYSCTL_INT(_net_inet_carp, CARPCTL_LOG, log, CTLFLAG_RW,
 142     &carp_opts[CARPCTL_LOG], 0, "log bad carp packets");
 143 SYSCTL_INT(_net_inet_carp, CARPCTL_ARPBALANCE, arpbalance, CTLFLAG_RW,
 144     &carp_opts[CARPCTL_ARPBALANCE], 0, "balance arp responses");
 145 SYSCTL_INT(_net_inet_carp, OID_AUTO, suppress_preempt, CTLFLAG_RD,
 146     &carp_suppress_preempt, 0, "Preemption is suppressed");
 147
 148 struct carpstats carpstats;
 149 SYSCTL_STRUCT(_net_inet_carp, CARPCTL_STATS, stats, CTLFLAG_RW,
 150     &carpstats, carpstats,
 151     "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 152
 153 struct carp_if {
 154         TAILQ_HEAD(, carp_softc) vhif_vrs;
 155         int vhif_nvrs;
 156
 157         struct ifnet    *vhif_ifp;
 158         struct mtx       vhif_mtx;
 159 };
 160
 161 /* Get carp_if from softc. Valid after carp_set_addr{,6}. */
 162 #define SC2CIF(sc)              ((struct carp_if *)(sc)->sc_carpdev->if_carp)
 163
 164 /* lock per carp_if queue */
 165 #define CARP_LOCK_INIT(cif)     mtx_init(&(cif)->vhif_mtx, "carp_if",   \
 166         NULL, MTX_DEF)
 167 #define CARP_LOCK_DESTROY(cif)  mtx_destroy(&(cif)->vhif_mtx)
 168 #define CARP_LOCK_ASSERT(cif)   mtx_assert(&(cif)->vhif_mtx, MA_OWNED)
 169 #define CARP_LOCK(cif)          mtx_lock(&(cif)->vhif_mtx)
 170 #define CARP_UNLOCK(cif)        mtx_unlock(&(cif)->vhif_mtx)
 171
 172 #define CARP_SCLOCK(sc)         mtx_lock(&SC2CIF(sc)->vhif_mtx)
 173 #define CARP_SCUNLOCK(sc)       mtx_unlock(&SC2CIF(sc)->vhif_mtx)
 174 #define CARP_SCLOCK_ASSERT(sc)  mtx_assert(&SC2CIF(sc)->vhif_mtx, MA_OWNED)
 175
 176 #define CARP_LOG(...)   do {                            \
 177         if (carp_opts[CARPCTL_LOG] > 0)                 \
 178                 log(LOG_INFO, __VA_ARGS__);             \
 179 } while (0)
 180
 181 #define CARP_DEBUG(...) do {                            \
 182         if (carp_opts[CARPCTL_LOG] > 1)                 \
 183                 log(LOG_DEBUG, __VA_ARGS__);            \
 184 } while (0)
 185
 186 static void     carp_hmac_prepare(struct carp_softc *);
 187 static void     carp_hmac_generate(struct carp_softc *, u_int32_t *,
 188                     unsigned char *);
 189 static int      carp_hmac_verify(struct carp_softc *, u_int32_t *,
 190                     unsigned char *);
 191 static void     carp_setroute(struct carp_softc *, int);
 192 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 193 static int      carp_clone_create(struct if_clone *, int, caddr_t);
 194 static void     carp_clone_destroy(struct ifnet *);
 195 static void     carpdetach(struct carp_softc *, int);
 196 static int      carp_prepare_ad(struct mbuf *, struct carp_softc *,
 197                     struct carp_header *);
 198 static void     carp_send_ad_all(void);
 199 static void     carp_send_ad(void *);
 200 static void     carp_send_ad_locked(struct carp_softc *);
 201 static void     carp_send_arp(struct carp_softc *);
 202 static void     carp_master_down(void *);
 203 static void     carp_master_down_locked(struct carp_softc *);
 204 static int      carp_ioctl(struct ifnet *, u_long, caddr_t);
 205 static int      carp_looutput(struct ifnet *, struct mbuf *, struct sockaddr *,
 206                     struct rtentry *);
 207 static void     carp_start(struct ifnet *);
 208 static void     carp_setrun(struct carp_softc *, sa_family_t);
 209 static void     carp_set_state(struct carp_softc *, int);
 210 static int      carp_addrcount(struct carp_if *, struct in_ifaddr *, int);
 211 enum    { CARP_COUNT_MASTER, CARP_COUNT_RUNNING };
 212
 213 static void     carp_multicast_cleanup(struct carp_softc *);
 214 static int      carp_set_addr(struct carp_softc *, struct sockaddr_in *);
 215 static int      carp_del_addr(struct carp_softc *, struct sockaddr_in *);
 216 static void     carp_carpdev_state_locked(struct carp_if *);
 217 static void     carp_sc_state_locked(struct carp_softc *);
 218 #ifdef INET6
 219 static void     carp_send_na(struct carp_softc *);
 220 static int      carp_set_addr6(struct carp_softc *, struct sockaddr_in6 *);
 221 static int      carp_del_addr6(struct carp_softc *, struct sockaddr_in6 *);
 222 static void     carp_multicast6_cleanup(struct carp_softc *);
 223 #endif
 224
 225 static LIST_HEAD(, carp_softc) carpif_list;
 226 static struct mtx carp_mtx;
 227 IFC_SIMPLE_DECLARE(carp, 0);
 228
 229 static eventhandler_tag if_detach_event_tag;
 230
 231 static __inline u_int16_t
 232 carp_cksum(struct mbuf *m, int len)
 233 {
 234         return (in_cksum(m, len));
 235 }
 236
 237 static void
 238 carp_hmac_prepare(struct carp_softc *sc)
 239 {
 240         u_int8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 241         u_int8_t vhid = sc->sc_vhid & 0xff;
 242         struct ifaddr *ifa;
 243         int i;
 244 #ifdef INET6
 245         struct in6_addr in6;
 246 #endif
 247
 248         if (sc->sc_carpdev)
 249                 CARP_SCLOCK(sc);
 250
 251         /* XXX: possible race here */
 252
 253         /* compute ipad from key */
 254         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 255         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 256         for (i = 0; i < sizeof(sc->sc_pad); i++)
 257                 sc->sc_pad[i] ^= 0x36;
 258
 259         /* precompute first part of inner hash */
 260         SHA1Init(&sc->sc_sha1);
 261         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 262         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 263         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 264         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 265 #ifdef INET
 266         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 267                 if (ifa->ifa_addr->sa_family == AF_INET)
 268                         SHA1Update(&sc->sc_sha1,
 269                             (void *)&ifatoia(ifa)->ia_addr.sin_addr.s_addr,
 270                             sizeof(struct in_addr));
 271         }
 272 #endif /* INET */
 273 #ifdef INET6
 274         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 275                 if (ifa->ifa_addr->sa_family == AF_INET6) {
 276                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 277                         in6_clearscope(&in6);
 278                         SHA1Update(&sc->sc_sha1, (void *)&in6, sizeof(in6));
 279                 }
 280         }
 281 #endif /* INET6 */
 282
 283         /* convert ipad to opad */
 284         for (i = 0; i < sizeof(sc->sc_pad); i++)
 285                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 286
 287         if (sc->sc_carpdev)
 288                 CARP_SCUNLOCK(sc);
 289 }
 290
 291 static void
 292 carp_hmac_generate(struct carp_softc *sc, u_int32_t counter[2],
 293     unsigned char md[20])
 294 {
 295         SHA1_CTX sha1ctx;
 296
 297         /* fetch first half of inner hash */
 298         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 299
 300         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 301         SHA1Final(md, &sha1ctx);
 302
 303         /* outer hash */
 304         SHA1Init(&sha1ctx);
 305         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 306         SHA1Update(&sha1ctx, md, 20);
 307         SHA1Final(md, &sha1ctx);
 308 }
 309
 310 static int
 311 carp_hmac_verify(struct carp_softc *sc, u_int32_t counter[2],
 312     unsigned char md[20])
 313 {
 314         unsigned char md2[20];
 315
 316         CARP_SCLOCK_ASSERT(sc);
 317
 318         carp_hmac_generate(sc, counter, md2);
 319
 320         return (bcmp(md, md2, sizeof(md2)));
 321 }
 322
 323 static void
 324 carp_setroute(struct carp_softc *sc, int cmd)
 325 {
 326         struct ifaddr *ifa;
 327         int s;
 328
 329         if (sc->sc_carpdev)
 330                 CARP_SCLOCK_ASSERT(sc);
 331
 332         s = splnet();
 333         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 334                 if (ifa->ifa_addr->sa_family == AF_INET &&
 335                     sc->sc_carpdev != NULL) {
 336                         int count = carp_addrcount(
 337                             (struct carp_if *)sc->sc_carpdev->if_carp,
 338                             ifatoia(ifa), CARP_COUNT_MASTER);
 339
 340                         if ((cmd == RTM_ADD && count == 1) ||
 341                             (cmd == RTM_DELETE && count == 0))
 342                                 rtinit(ifa, cmd, RTF_UP | RTF_HOST);
 343                 }
 344 #ifdef INET6
 345                 if (ifa->ifa_addr->sa_family == AF_INET6) {
 346                         if (cmd == RTM_ADD)
 347                                 in6_ifaddloop(ifa);
 348                         else
 349                                 in6_ifremloop(ifa);
 350                 }
 351 #endif /* INET6 */
 352         }
 353         splx(s);
 354 }
 355
 356 static int
 357 carp_clone_create(struct if_clone *ifc, int unit, caddr_t params)
 358 {
 359
 360         struct carp_softc *sc;
 361         struct ifnet *ifp;
 362
 363         MALLOC(sc, struct carp_softc *, sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
 364         ifp = SC2IFP(sc) = if_alloc(IFT_ETHER);
 365         if (ifp == NULL) {
 366                 FREE(sc, M_CARP);
 367                 return (ENOSPC);
 368         }
 369
 370         sc->sc_flags_backup = 0;
 371         sc->sc_suppress = 0;
 372         sc->sc_advbase = CARP_DFLTINTV;
 373         sc->sc_vhid = -1;       /* required setting */
 374         sc->sc_advskew = 0;
 375         sc->sc_init_counter = 1;
 376         sc->sc_naddrs = sc->sc_naddrs6 = 0; /* M_ZERO? */
 377 #ifdef INET6
 378         sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL;
 379 #endif
 380         sc->sc_imo.imo_membership = (struct in_multi **)malloc(
 381             (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
 382             M_WAITOK);
 383         sc->sc_imo.imo_max_memberships = IP_MIN_MEMBERSHIPS;
 384         sc->sc_imo.imo_multicast_vif = -1;
 385
 386         callout_init(&sc->sc_ad_tmo, NET_CALLOUT_MPSAFE);
 387         callout_init(&sc->sc_md_tmo, NET_CALLOUT_MPSAFE);
 388         callout_init(&sc->sc_md6_tmo, NET_CALLOUT_MPSAFE);
 389
 390         ifp->if_softc = sc;
 391         if_initname(ifp, CARP_IFNAME, unit);
 392         ifp->if_mtu = ETHERMTU;
 393         ifp->if_flags = IFF_LOOPBACK;
 394         ifp->if_ioctl = carp_ioctl;
 395         ifp->if_output = carp_looutput;
 396         ifp->if_start = carp_start;
 397         ifp->if_type = IFT_CARP;
 398         ifp->if_snd.ifq_maxlen = ifqmaxlen;
 399         ifp->if_hdrlen = 0;
 400         if_attach(ifp);
 401         bpfattach(SC2IFP(sc), DLT_NULL, sizeof(u_int32_t));
 402         mtx_lock(&carp_mtx);
 403         LIST_INSERT_HEAD(&carpif_list, sc, sc_next);
 404         mtx_unlock(&carp_mtx);
 405         return (0);
 406 }
 407
 408 static void
 409 carp_clone_destroy(struct ifnet *ifp)
 410 {
 411         struct carp_softc *sc = ifp->if_softc;
 412
 413         if (sc->sc_carpdev)
 414                 CARP_SCLOCK(sc);
 415         carpdetach(sc, 1);      /* Returns unlocked. */
 416
 417         mtx_lock(&carp_mtx);
 418         LIST_REMOVE(sc, sc_next);
 419         mtx_unlock(&carp_mtx);
 420         bpfdetach(ifp);
 421         if_detach(ifp);
 422         if_free_type(ifp, IFT_ETHER);
 423         free(sc->sc_imo.imo_membership, M_CARP);
 424         free(sc, M_CARP);
 425 }
 426
 427 /*
 428  * This function can be called on CARP interface destroy path,
 429  * and in case of the removal of the underlying interface as
 430  * well. We differentiate these two cases. In the latter case
 431  * we do not cleanup our multicast memberships, since they
 432  * are already freed. Also, in the latter case we do not
 433  * release the lock on return, because the function will be
 434  * called once more, for another CARP instance on the same
 435  * interface.
 436  */
 437 static void
 438 carpdetach(struct carp_softc *sc, int unlock)
 439 {
 440         struct carp_if *cif;
 441
 442         callout_stop(&sc->sc_ad_tmo);
 443         callout_stop(&sc->sc_md_tmo);
 444         callout_stop(&sc->sc_md6_tmo);
 445
 446         if (sc->sc_suppress)
 447                 carp_suppress_preempt--;
 448         sc->sc_suppress = 0;
 449
 450         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS)
 451                 carp_suppress_preempt--;
 452         sc->sc_sendad_errors = 0;
 453
 454         carp_set_state(sc, INIT);
 455         SC2IFP(sc)->if_flags &= ~IFF_UP;
 456         carp_setrun(sc, 0);
 457         if (unlock)
 458                 carp_multicast_cleanup(sc);
 459 #ifdef INET6
 460         carp_multicast6_cleanup(sc);
 461 #endif
 462
 463         if (sc->sc_carpdev != NULL) {
 464                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
 465                 CARP_LOCK_ASSERT(cif);
 466                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
 467                 if (!--cif->vhif_nvrs) {
 468                         ifpromisc(sc->sc_carpdev, 0);
 469                         sc->sc_carpdev->if_carp = NULL;
 470                         CARP_LOCK_DESTROY(cif);
 471                         FREE(cif, M_IFADDR);
 472                 } else if (unlock)
 473                         CARP_UNLOCK(cif);
 474                 sc->sc_carpdev = NULL;
 475         }
 476 }
 477
 478 /* Detach an interface from the carp. */
 479 static void
 480 carp_ifdetach(void *arg __unused, struct ifnet *ifp)
 481 {
 482         struct carp_if *cif = (struct carp_if *)ifp->if_carp;
 483         struct carp_softc *sc, *nextsc;
 484
 485         if (cif == NULL)
 486                 return;
 487
 488         /*
 489          * XXX: At the end of for() cycle the lock will be destroyed.
 490          */
 491         CARP_LOCK(cif);
 492         for (sc = TAILQ_FIRST(&cif->vhif_vrs); sc; sc = nextsc) {
 493                 nextsc = TAILQ_NEXT(sc, sc_list);
 494                 carpdetach(sc, 0);
 495         }
 496 }
 497
 498 /*
 499  * process input packet.
 500  * we have rearranged checks order compared to the rfc,
 501  * but it seems more efficient this way or not possible otherwise.
 502  */
 503 void
 504 carp_input(struct mbuf *m, int hlen)
 505 {
 506         struct ip *ip = mtod(m, struct ip *);
 507         struct carp_header *ch;
 508         int iplen, len;
 509
 510         carpstats.carps_ipackets++;
 511
 512         if (!carp_opts[CARPCTL_ALLOW]) {
 513                 m_freem(m);
 514                 return;
 515         }
 516
 517         /* check if received on a valid carp interface */
 518         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 519                 carpstats.carps_badif++;
 520                 CARP_LOG("carp_input: packet received on non-carp "
 521                     "interface: %s\n",
 522                     m->m_pkthdr.rcvif->if_xname);
 523                 m_freem(m);
 524                 return;
 525         }
 526
 527         /* verify that the IP TTL is 255.  */
 528         if (ip->ip_ttl != CARP_DFLTTL) {
 529                 carpstats.carps_badttl++;
 530                 CARP_LOG("carp_input: received ttl %d != 255i on %s\n",
 531                     ip->ip_ttl,
 532                     m->m_pkthdr.rcvif->if_xname);
 533                 m_freem(m);
 534                 return;
 535         }
 536
 537         iplen = ip->ip_hl << 2;
 538
 539         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 540                 carpstats.carps_badlen++;
 541                 CARP_LOG("carp_input: received len %zd < "
 542                     "sizeof(struct carp_header)\n",
 543                     m->m_len - sizeof(struct ip));
 544                 m_freem(m);
 545                 return;
 546         }
 547
 548         if (iplen + sizeof(*ch) < m->m_len) {
 549                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 550                         carpstats.carps_hdrops++;
 551                         CARP_LOG("carp_input: pullup failed\n");
 552                         return;
 553                 }
 554                 ip = mtod(m, struct ip *);
 555         }
 556         ch = (struct carp_header *)((char *)ip + iplen);
 557
 558         /*
 559          * verify that the received packet length is
 560          * equal to the CARP header
 561          */
 562         len = iplen + sizeof(*ch);
 563         if (len > m->m_pkthdr.len) {
 564                 carpstats.carps_badlen++;
 565                 CARP_LOG("carp_input: packet too short %d on %s\n",
 566                     m->m_pkthdr.len,
 567                     m->m_pkthdr.rcvif->if_xname);
 568                 m_freem(m);
 569                 return;
 570         }
 571
 572         if ((m = m_pullup(m, len)) == NULL) {
 573                 carpstats.carps_hdrops++;
 574                 return;
 575         }
 576         ip = mtod(m, struct ip *);
 577         ch = (struct carp_header *)((char *)ip + iplen);
 578
 579         /* verify the CARP checksum */
 580         m->m_data += iplen;
 581         if (carp_cksum(m, len - iplen)) {
 582                 carpstats.carps_badsum++;
 583                 CARP_LOG("carp_input: checksum failed on %s\n",
 584                     m->m_pkthdr.rcvif->if_xname);
 585                 m_freem(m);
 586                 return;
 587         }
 588         m->m_data -= iplen;
 589
 590         carp_input_c(m, ch, AF_INET);
 591 }
 592
 593 #ifdef INET6
 594 int
 595 carp6_input(struct mbuf **mp, int *offp, int proto)
 596 {
 597         struct mbuf *m = *mp;
 598         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 599         struct carp_header *ch;
 600         u_int len;
 601
 602         carpstats.carps_ipackets6++;
 603
 604         if (!carp_opts[CARPCTL_ALLOW]) {
 605                 m_freem(m);
 606                 return (IPPROTO_DONE);
 607         }
 608
 609         /* check if received on a valid carp interface */
 610         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 611                 carpstats.carps_badif++;
 612                 CARP_LOG("carp6_input: packet received on non-carp "
 613                     "interface: %s\n",
 614                     m->m_pkthdr.rcvif->if_xname);
 615                 m_freem(m);
 616                 return (IPPROTO_DONE);
 617         }
 618
 619         /* verify that the IP TTL is 255 */
 620         if (ip6->ip6_hlim != CARP_DFLTTL) {
 621                 carpstats.carps_badttl++;
 622                 CARP_LOG("carp6_input: received ttl %d != 255 on %s\n",
 623                     ip6->ip6_hlim,
 624                     m->m_pkthdr.rcvif->if_xname);
 625                 m_freem(m);
 626                 return (IPPROTO_DONE);
 627         }
 628
 629         /* verify that we have a complete carp packet */
 630         len = m->m_len;
 631         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 632         if (ch == NULL) {
 633                 carpstats.carps_badlen++;
 634                 CARP_LOG("carp6_input: packet size %u too small\n", len);
 635                 return (IPPROTO_DONE);
 636         }
 637
 638
 639         /* verify the CARP checksum */
 640         m->m_data += *offp;
 641         if (carp_cksum(m, sizeof(*ch))) {
 642                 carpstats.carps_badsum++;
 643                 CARP_LOG("carp6_input: checksum failed, on %s\n",
 644                     m->m_pkthdr.rcvif->if_xname);
 645                 m_freem(m);
 646                 return (IPPROTO_DONE);
 647         }
 648         m->m_data -= *offp;
 649
 650         carp_input_c(m, ch, AF_INET6);
 651         return (IPPROTO_DONE);
 652 }
 653 #endif /* INET6 */
 654
 655 static void
 656 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 657 {
 658         struct ifnet *ifp = m->m_pkthdr.rcvif;
 659         struct carp_softc *sc;
 660         u_int64_t tmp_counter;
 661         struct timeval sc_tv, ch_tv;
 662
 663         /* verify that the VHID is valid on the receiving interface */
 664         CARP_LOCK(ifp->if_carp);
 665         TAILQ_FOREACH(sc, &((struct carp_if *)ifp->if_carp)->vhif_vrs, sc_list)
 666                 if (sc->sc_vhid == ch->carp_vhid)
 667                         break;
 668
 669         if (!sc || !((SC2IFP(sc)->if_flags & IFF_UP) &&
 670             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING))) {
 671                 carpstats.carps_badvhid++;
 672                 CARP_UNLOCK(ifp->if_carp);
 673                 m_freem(m);
 674                 return;
 675         }
 676
 677         getmicrotime(&SC2IFP(sc)->if_lastchange);
 678         SC2IFP(sc)->if_ipackets++;
 679         SC2IFP(sc)->if_ibytes += m->m_pkthdr.len;
 680
 681         if (bpf_peers_present(SC2IFP(sc)->if_bpf)) {
 682                 struct ip *ip = mtod(m, struct ip *);
 683                 uint32_t af1 = af;
 684
 685                 /* BPF wants net byte order */
 686                 ip->ip_len = htons(ip->ip_len + (ip->ip_hl << 2));
 687                 ip->ip_off = htons(ip->ip_off);
 688                 bpf_mtap2(SC2IFP(sc)->if_bpf, &af1, sizeof(af1), m);
 689         }
 690
 691         /* verify the CARP version. */
 692         if (ch->carp_version != CARP_VERSION) {
 693                 carpstats.carps_badver++;
 694                 SC2IFP(sc)->if_ierrors++;
 695                 CARP_UNLOCK(ifp->if_carp);
 696                 CARP_LOG("%s; invalid version %d\n",
 697                     SC2IFP(sc)->if_xname,
 698                     ch->carp_version);
 699                 m_freem(m);
 700                 return;
 701         }
 702
 703         /* verify the hash */
 704         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 705                 carpstats.carps_badauth++;
 706                 SC2IFP(sc)->if_ierrors++;
 707                 CARP_UNLOCK(ifp->if_carp);
 708                 CARP_LOG("%s: incorrect hash\n", SC2IFP(sc)->if_xname);
 709                 m_freem(m);
 710                 return;
 711         }
 712
 713         tmp_counter = ntohl(ch->carp_counter[0]);
 714         tmp_counter = tmp_counter<<32;
 715         tmp_counter += ntohl(ch->carp_counter[1]);
 716
 717         /* XXX Replay protection goes here */
 718
 719         sc->sc_init_counter = 0;
 720         sc->sc_counter = tmp_counter;
 721
 722         sc_tv.tv_sec = sc->sc_advbase;
 723         if (carp_suppress_preempt && sc->sc_advskew <  240)
 724                 sc_tv.tv_usec = 240 * 1000000 / 256;
 725         else
 726                 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256;
 727         ch_tv.tv_sec = ch->carp_advbase;
 728         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 729
 730         switch (sc->sc_state) {
 731         case INIT:
 732                 break;
 733         case MASTER:
 734                 /*
 735                  * If we receive an advertisement from a master who's going to
 736                  * be more frequent than us, go into BACKUP state.
 737                  */
 738                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 739                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 740                         callout_stop(&sc->sc_ad_tmo);
 741                         CARP_DEBUG("%s: MASTER -> BACKUP "
 742                            "(more frequent advertisement received)\n",
 743                            SC2IFP(sc)->if_xname);
 744                         carp_set_state(sc, BACKUP);
 745                         carp_setrun(sc, 0);
 746                         carp_setroute(sc, RTM_DELETE);
 747                 }
 748                 break;
 749         case BACKUP:
 750                 /*
 751                  * If we're pre-empting masters who advertise slower than us,
 752                  * and this one claims to be slower, treat him as down.
 753                  */
 754                 if (carp_opts[CARPCTL_PREEMPT] &&
 755                     timevalcmp(&sc_tv, &ch_tv, <)) {
 756                         CARP_DEBUG("%s: BACKUP -> MASTER "
 757                             "(preempting a slower master)\n",
 758                             SC2IFP(sc)->if_xname);
 759                         carp_master_down_locked(sc);
 760                         break;
 761                 }
 762
 763                 /*
 764                  *  If the master is going to advertise at such a low frequency
 765                  *  that he's guaranteed to time out, we'd might as well just
 766                  *  treat him as timed out now.
 767                  */
 768                 sc_tv.tv_sec = sc->sc_advbase * 3;
 769                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 770                         CARP_DEBUG("%s: BACKUP -> MASTER "
 771                             "(master timed out)\n",
 772                             SC2IFP(sc)->if_xname);
 773                         carp_master_down_locked(sc);
 774                         break;
 775                 }
 776
 777                 /*
 778                  * Otherwise, we reset the counter and wait for the next
 779                  * advertisement.
 780                  */
 781                 carp_setrun(sc, af);
 782                 break;
 783         }
 784
 785         CARP_UNLOCK(ifp->if_carp);
 786
 787         m_freem(m);
 788         return;
 789 }
 790
 791 static int
 792 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 793 {
 794         struct m_tag *mtag;
 795         struct ifnet *ifp = SC2IFP(sc);
 796
 797         if (sc->sc_init_counter) {
 798                 /* this could also be seconds since unix epoch */
 799                 sc->sc_counter = arc4random();
 800                 sc->sc_counter = sc->sc_counter << 32;
 801                 sc->sc_counter += arc4random();
 802         } else
 803                 sc->sc_counter++;
 804
 805         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 806         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 807
 808         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 809
 810         /* Tag packet for carp_output */
 811         mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct ifnet *), M_NOWAIT);
 812         if (mtag == NULL) {
 813                 m_freem(m);
 814                 SC2IFP(sc)->if_oerrors++;
 815                 return (ENOMEM);
 816         }
 817         bcopy(&ifp, (caddr_t)(mtag + 1), sizeof(struct ifnet *));
 818         m_tag_prepend(m, mtag);
 819
 820         return (0);
 821 }
 822
 823 static void
 824 carp_send_ad_all(void)
 825 {
 826         struct carp_softc *sc;
 827
 828         mtx_lock(&carp_mtx);
 829         LIST_FOREACH(sc, &carpif_list, sc_next) {
 830                 if (sc->sc_carpdev == NULL)
 831                         continue;
 832                 CARP_SCLOCK(sc);
 833                 if ((SC2IFP(sc)->if_flags & IFF_UP) &&
 834                     (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING) &&
 835                      sc->sc_state == MASTER)
 836                         carp_send_ad_locked(sc);
 837                 CARP_SCUNLOCK(sc);
 838         }
 839         mtx_unlock(&carp_mtx);
 840 }
 841
 842 static void
 843 carp_send_ad(void *v)
 844 {
 845         struct carp_softc *sc = v;
 846
 847         CARP_SCLOCK(sc);
 848         carp_send_ad_locked(sc);
 849         CARP_SCUNLOCK(sc);
 850 }
 851
 852 static void
 853 carp_send_ad_locked(struct carp_softc *sc)
 854 {
 855         struct carp_header ch;
 856         struct timeval tv;
 857         struct carp_header *ch_ptr;
 858         struct mbuf *m;
 859         int len, advbase, advskew;
 860
 861         CARP_SCLOCK_ASSERT(sc);
 862
 863         /* bow out if we've lost our UPness or RUNNINGuiness */
 864         if (!((SC2IFP(sc)->if_flags & IFF_UP) &&
 865             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING))) {
 866                 advbase = 255;
 867                 advskew = 255;
 868         } else {
 869                 advbase = sc->sc_advbase;
 870                 if (!carp_suppress_preempt || sc->sc_advskew > 240)
 871                         advskew = sc->sc_advskew;
 872                 else
 873                         advskew = 240;
 874                 tv.tv_sec = advbase;
 875                 tv.tv_usec = advskew * 1000000 / 256;
 876         }
 877
 878         ch.carp_version = CARP_VERSION;
 879         ch.carp_type = CARP_ADVERTISEMENT;
 880         ch.carp_vhid = sc->sc_vhid;
 881         ch.carp_advbase = advbase;
 882         ch.carp_advskew = advskew;
 883         ch.carp_authlen = 7;    /* XXX DEFINE */
 884         ch.carp_pad1 = 0;       /* must be zero */
 885         ch.carp_cksum = 0;
 886
 887 #ifdef INET
 888         if (sc->sc_ia) {
 889                 struct ip *ip;
 890
 891                 MGETHDR(m, M_DONTWAIT, MT_HEADER);
 892                 if (m == NULL) {
 893                         SC2IFP(sc)->if_oerrors++;
 894                         carpstats.carps_onomem++;
 895                         /* XXX maybe less ? */
 896                         if (advbase != 255 || advskew != 255)
 897                                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
 898                                     carp_send_ad, sc);
 899                         return;
 900                 }
 901                 len = sizeof(*ip) + sizeof(ch);
 902                 m->m_pkthdr.len = len;
 903                 m->m_pkthdr.rcvif = NULL;
 904                 m->m_len = len;
 905                 MH_ALIGN(m, m->m_len);
 906                 m->m_flags |= M_MCAST;
 907                 ip = mtod(m, struct ip *);
 908                 ip->ip_v = IPVERSION;
 909                 ip->ip_hl = sizeof(*ip) >> 2;
 910                 ip->ip_tos = IPTOS_LOWDELAY;
 911                 ip->ip_len = len;
 912                 ip->ip_id = ip_newid();
 913                 ip->ip_off = IP_DF;
 914                 ip->ip_ttl = CARP_DFLTTL;
 915                 ip->ip_p = IPPROTO_CARP;
 916                 ip->ip_sum = 0;
 917                 ip->ip_src.s_addr = sc->sc_ia->ia_addr.sin_addr.s_addr;
 918                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 919
 920                 ch_ptr = (struct carp_header *)(&ip[1]);
 921                 bcopy(&ch, ch_ptr, sizeof(ch));
 922                 if (carp_prepare_ad(m, sc, ch_ptr))
 923                         return;
 924
 925                 m->m_data += sizeof(*ip);
 926                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip));
 927                 m->m_data -= sizeof(*ip);
 928
 929                 getmicrotime(&SC2IFP(sc)->if_lastchange);
 930                 SC2IFP(sc)->if_opackets++;
 931                 SC2IFP(sc)->if_obytes += len;
 932                 carpstats.carps_opackets++;
 933
 934                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, NULL)) {
 935                         SC2IFP(sc)->if_oerrors++;
 936                         if (sc->sc_sendad_errors < INT_MAX)
 937                                 sc->sc_sendad_errors++;
 938                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 939                                 carp_suppress_preempt++;
 940                                 if (carp_suppress_preempt == 1) {
 941                                         CARP_SCUNLOCK(sc);
 942                                         carp_send_ad_all();
 943                                         CARP_SCLOCK(sc);
 944                                 }
 945                         }
 946                         sc->sc_sendad_success = 0;
 947                 } else {
 948                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 949                                 if (++sc->sc_sendad_success >=
 950                                     CARP_SENDAD_MIN_SUCCESS) {
 951                                         carp_suppress_preempt--;
 952                                         sc->sc_sendad_errors = 0;
 953                                 }
 954                         } else
 955                                 sc->sc_sendad_errors = 0;
 956                 }
 957         }
 958 #endif /* INET */
 959 #ifdef INET6
 960         if (sc->sc_ia6) {
 961                 struct ip6_hdr *ip6;
 962
 963                 MGETHDR(m, M_DONTWAIT, MT_HEADER);
 964                 if (m == NULL) {
 965                         SC2IFP(sc)->if_oerrors++;
 966                         carpstats.carps_onomem++;
 967                         /* XXX maybe less ? */
 968                         if (advbase != 255 || advskew != 255)
 969                                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
 970                                     carp_send_ad, sc);
 971                         return;
 972                 }
 973                 len = sizeof(*ip6) + sizeof(ch);
 974                 m->m_pkthdr.len = len;
 975                 m->m_pkthdr.rcvif = NULL;
 976                 m->m_len = len;
 977                 MH_ALIGN(m, m->m_len);
 978                 m->m_flags |= M_MCAST;
 979                 ip6 = mtod(m, struct ip6_hdr *);
 980                 bzero(ip6, sizeof(*ip6));
 981                 ip6->ip6_vfc |= IPV6_VERSION;
 982                 ip6->ip6_hlim = CARP_DFLTTL;
 983                 ip6->ip6_nxt = IPPROTO_CARP;
 984                 bcopy(&sc->sc_ia6->ia_addr.sin6_addr, &ip6->ip6_src,
 985                     sizeof(struct in6_addr));
 986                 /* set the multicast destination */
 987
 988                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
 989                 ip6->ip6_dst.s6_addr8[15] = 0x12;
 990                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
 991                         SC2IFP(sc)->if_oerrors++;
 992                         m_freem(m);
 993                         CARP_LOG("%s: in6_setscope failed\n", __func__);
 994                         return;
 995                 }
 996
 997                 ch_ptr = (struct carp_header *)(&ip6[1]);
 998                 bcopy(&ch, ch_ptr, sizeof(ch));
 999                 if (carp_prepare_ad(m, sc, ch_ptr))
1000                         return;
1001
1002                 m->m_data += sizeof(*ip6);
1003                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6));
1004                 m->m_data -= sizeof(*ip6);
1005
1006                 getmicrotime(&SC2IFP(sc)->if_lastchange);
1007                 SC2IFP(sc)->if_opackets++;
1008                 SC2IFP(sc)->if_obytes += len;
1009                 carpstats.carps_opackets6++;
1010
1011                 if (ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL)) {
1012                         SC2IFP(sc)->if_oerrors++;
1013                         if (sc->sc_sendad_errors < INT_MAX)
1014                                 sc->sc_sendad_errors++;
1015                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
1016                                 carp_suppress_preempt++;
1017                                 if (carp_suppress_preempt == 1) {
1018                                         CARP_SCUNLOCK(sc);
1019                                         carp_send_ad_all();
1020                                         CARP_SCLOCK(sc);
1021                                 }
1022                         }
1023                         sc->sc_sendad_success = 0;
1024                 } else {
1025                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
1026                                 if (++sc->sc_sendad_success >=
1027                                     CARP_SENDAD_MIN_SUCCESS) {
1028                                         carp_suppress_preempt--;
1029                                         sc->sc_sendad_errors = 0;
1030                                 }
1031                         } else
1032                                 sc->sc_sendad_errors = 0;
1033                 }
1034         }
1035 #endif /* INET6 */
1036
1037         if (advbase != 255 || advskew != 255)
1038                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1039                     carp_send_ad, sc);
1040
1041 }
1042
1043 /*
1044  * Broadcast a gratuitous ARP request containing
1045  * the virtual router MAC address for each IP address
1046  * associated with the virtual router.
1047  */
1048 static void
1049 carp_send_arp(struct carp_softc *sc)
1050 {
1051         struct ifaddr *ifa;
1052
1053         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1054
1055                 if (ifa->ifa_addr->sa_family != AF_INET)
1056                         continue;
1057
1058 /*              arprequest(sc->sc_carpdev, &in, &in, IF_LLADDR(sc->sc_ifp)); */
1059                 arp_ifinit2(sc->sc_carpdev, ifa, IF_LLADDR(sc->sc_ifp));
1060
1061                 DELAY(1000);    /* XXX */
1062         }
1063 }
1064
1065 #ifdef INET6
1066 static void
1067 carp_send_na(struct carp_softc *sc)
1068 {
1069         struct ifaddr *ifa;
1070         struct in6_addr *in6;
1071         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1072
1073         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1074
1075                 if (ifa->ifa_addr->sa_family != AF_INET6)
1076                         continue;
1077
1078                 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr;
1079                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1080                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1081                 DELAY(1000);    /* XXX */
1082         }
1083 }
1084 #endif /* INET6 */
1085
1086 static int
1087 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type)
1088 {
1089         struct carp_softc *vh;
1090         struct ifaddr *ifa;
1091         int count = 0;
1092
1093         CARP_LOCK_ASSERT(cif);
1094
1095         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1096                 if ((type == CARP_COUNT_RUNNING &&
1097                     (SC2IFP(vh)->if_flags & IFF_UP) &&
1098                     (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING)) ||
1099                     (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) {
1100                         TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1101                             ifa_list) {
1102                                 if (ifa->ifa_addr->sa_family == AF_INET &&
1103                                     ia->ia_addr.sin_addr.s_addr ==
1104                                     ifatoia(ifa)->ia_addr.sin_addr.s_addr)
1105                                         count++;
1106                         }
1107                 }
1108         }
1109         return (count);
1110 }
1111
1112 int
1113 carp_iamatch(void *v, struct in_ifaddr *ia,
1114     struct in_addr *isaddr, u_int8_t **enaddr)
1115 {
1116         struct carp_if *cif = v;
1117         struct carp_softc *vh;
1118         int index, count = 0;
1119         struct ifaddr *ifa;
1120
1121         CARP_LOCK(cif);
1122
1123         if (carp_opts[CARPCTL_ARPBALANCE]) {
1124                 /*
1125                  * XXX proof of concept implementation.
1126                  * We use the source ip to decide which virtual host should
1127                  * handle the request. If we're master of that virtual host,
1128                  * then we respond, otherwise, just drop the arp packet on
1129                  * the floor.
1130                  */
1131                 count = carp_addrcount(cif, ia, CARP_COUNT_RUNNING);
1132                 if (count == 0) {
1133                         /* should never reach this */
1134                         CARP_UNLOCK(cif);
1135                         return (0);
1136                 }
1137
1138                 /* this should be a hash, like pf_hash() */
1139                 index = ntohl(isaddr->s_addr) % count;
1140                 count = 0;
1141
1142                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1143                         if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1144                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING)) {
1145                                 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1146                                     ifa_list) {
1147                                         if (ifa->ifa_addr->sa_family ==
1148                                             AF_INET &&
1149                                             ia->ia_addr.sin_addr.s_addr ==
1150                                             ifatoia(ifa)->ia_addr.sin_addr.s_addr) {
1151                                                 if (count == index) {
1152                                                         if (vh->sc_state ==
1153                                                             MASTER) {
1154                                                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1155                                                                 CARP_UNLOCK(cif);
1156                                                                 return (1);
1157                                                         } else {
1158                                                                 CARP_UNLOCK(cif);
1159                                                                 return (0);
1160                                                         }
1161                                                 }
1162                                                 count++;
1163                                         }
1164                                 }
1165                         }
1166                 }
1167         } else {
1168                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1169                         if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1170                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1171                             ia->ia_ifp == SC2IFP(vh) &&
1172                             vh->sc_state == MASTER) {
1173                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1174                                 CARP_UNLOCK(cif);
1175                                 return (1);
1176                         }
1177                 }
1178         }
1179         CARP_UNLOCK(cif);
1180         return (0);
1181 }
1182
1183 #ifdef INET6
1184 struct ifaddr *
1185 carp_iamatch6(void *v, struct in6_addr *taddr)
1186 {
1187         struct carp_if *cif = v;
1188         struct carp_softc *vh;
1189         struct ifaddr *ifa;
1190
1191         CARP_LOCK(cif);
1192         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1193                 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist, ifa_list) {
1194                         if (IN6_ARE_ADDR_EQUAL(taddr,
1195                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1196                             (SC2IFP(vh)->if_flags & IFF_UP) &&
1197                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1198                             vh->sc_state == MASTER) {
1199                                 CARP_UNLOCK(cif);
1200                                 return (ifa);
1201                         }
1202                 }
1203         }
1204         CARP_UNLOCK(cif);
1205
1206         return (NULL);
1207 }
1208
1209 void *
1210 carp_macmatch6(void *v, struct mbuf *m, const struct in6_addr *taddr)
1211 {
1212         struct m_tag *mtag;
1213         struct carp_if *cif = v;
1214         struct carp_softc *sc;
1215         struct ifaddr *ifa;
1216
1217         CARP_LOCK(cif);
1218         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) {
1219                 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1220                         if (IN6_ARE_ADDR_EQUAL(taddr,
1221                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1222                             (SC2IFP(sc)->if_flags & IFF_UP) &&
1223                             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING)) {
1224                                 struct ifnet *ifp = SC2IFP(sc);
1225                                 mtag = m_tag_get(PACKET_TAG_CARP,
1226                                     sizeof(struct ifnet *), M_NOWAIT);
1227                                 if (mtag == NULL) {
1228                                         /* better a bit than nothing */
1229                                         CARP_UNLOCK(cif);
1230                                         return (IF_LLADDR(sc->sc_ifp));
1231                                 }
1232                                 bcopy(&ifp, (caddr_t)(mtag + 1),
1233                                     sizeof(struct ifnet *));
1234                                 m_tag_prepend(m, mtag);
1235
1236                                 CARP_UNLOCK(cif);
1237                                 return (IF_LLADDR(sc->sc_ifp));
1238                         }
1239                 }
1240         }
1241         CARP_UNLOCK(cif);
1242
1243         return (NULL);
1244 }
1245 #endif
1246
1247 struct ifnet *
1248 carp_forus(void *v, void *dhost)
1249 {
1250         struct carp_if *cif = v;
1251         struct carp_softc *vh;
1252         u_int8_t *ena = dhost;
1253
1254         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1255                 return (NULL);
1256
1257         CARP_LOCK(cif);
1258         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list)
1259                 if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1260                     (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1261                     vh->sc_state == MASTER &&
1262                     !bcmp(dhost, IF_LLADDR(vh->sc_ifp), ETHER_ADDR_LEN)) {
1263                         CARP_UNLOCK(cif);
1264                         return (SC2IFP(vh));
1265                 }
1266
1267         CARP_UNLOCK(cif);
1268         return (NULL);
1269 }
1270
1271 static void
1272 carp_master_down(void *v)
1273 {
1274         struct carp_softc *sc = v;
1275
1276         CARP_SCLOCK(sc);
1277         carp_master_down_locked(sc);
1278         CARP_SCUNLOCK(sc);
1279 }
1280
1281 static void
1282 carp_master_down_locked(struct carp_softc *sc)
1283 {
1284         if (sc->sc_carpdev)
1285                 CARP_SCLOCK_ASSERT(sc);
1286
1287         switch (sc->sc_state) {
1288         case INIT:
1289                 printf("%s: master_down event in INIT state\n",
1290                     SC2IFP(sc)->if_xname);
1291                 break;
1292         case MASTER:
1293                 break;
1294         case BACKUP:
1295                 carp_set_state(sc, MASTER);
1296                 carp_send_ad_locked(sc);
1297                 carp_send_arp(sc);
1298 #ifdef INET6
1299                 carp_send_na(sc);
1300 #endif /* INET6 */
1301                 carp_setrun(sc, 0);
1302                 carp_setroute(sc, RTM_ADD);
1303                 break;
1304         }
1305 }
1306
1307 /*
1308  * When in backup state, af indicates whether to reset the master down timer
1309  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1310  */
1311 static void
1312 carp_setrun(struct carp_softc *sc, sa_family_t af)
1313 {
1314         struct timeval tv;
1315
1316         if (sc->sc_carpdev == NULL) {
1317                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1318                 carp_set_state(sc, INIT);
1319                 return;
1320         } else
1321                 CARP_SCLOCK_ASSERT(sc);
1322
1323         if (SC2IFP(sc)->if_flags & IFF_UP &&
1324             sc->sc_vhid > 0 && (sc->sc_naddrs || sc->sc_naddrs6))
1325                 SC2IFP(sc)->if_drv_flags |= IFF_DRV_RUNNING;
1326         else {
1327                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1328                 carp_setroute(sc, RTM_DELETE);
1329                 return;
1330         }
1331
1332         switch (sc->sc_state) {
1333         case INIT:
1334                 if (carp_opts[CARPCTL_PREEMPT] && !carp_suppress_preempt) {
1335                         carp_send_ad_locked(sc);
1336                         carp_send_arp(sc);
1337 #ifdef INET6
1338                         carp_send_na(sc);
1339 #endif /* INET6 */
1340                         CARP_DEBUG("%s: INIT -> MASTER (preempting)\n",
1341                             SC2IFP(sc)->if_xname);
1342                         carp_set_state(sc, MASTER);
1343                         carp_setroute(sc, RTM_ADD);
1344                 } else {
1345                         CARP_DEBUG("%s: INIT -> BACKUP\n", SC2IFP(sc)->if_xname);
1346                         carp_set_state(sc, BACKUP);
1347                         carp_setroute(sc, RTM_DELETE);
1348                         carp_setrun(sc, 0);
1349                 }
1350                 break;
1351         case BACKUP:
1352                 callout_stop(&sc->sc_ad_tmo);
1353                 tv.tv_sec = 3 * sc->sc_advbase;
1354                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1355                 switch (af) {
1356 #ifdef INET
1357                 case AF_INET:
1358                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1359                             carp_master_down, sc);
1360                         break;
1361 #endif /* INET */
1362 #ifdef INET6
1363                 case AF_INET6:
1364                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1365                             carp_master_down, sc);
1366                         break;
1367 #endif /* INET6 */
1368                 default:
1369                         if (sc->sc_naddrs)
1370                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1371                                     carp_master_down, sc);
1372                         if (sc->sc_naddrs6)
1373                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1374                                     carp_master_down, sc);
1375                         break;
1376                 }
1377                 break;
1378         case MASTER:
1379                 tv.tv_sec = sc->sc_advbase;
1380                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1381                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1382                     carp_send_ad, sc);
1383                 break;
1384         }
1385 }
1386
1387 static void
1388 carp_multicast_cleanup(struct carp_softc *sc)
1389 {
1390         struct ip_moptions *imo = &sc->sc_imo;
1391         u_int16_t n = imo->imo_num_memberships;
1392
1393         /* Clean up our own multicast memberships */
1394         while (n-- > 0) {
1395                 if (imo->imo_membership[n] != NULL) {
1396                         in_delmulti(imo->imo_membership[n]);
1397                         imo->imo_membership[n] = NULL;
1398                 }
1399         }
1400         imo->imo_num_memberships = 0;
1401         imo->imo_multicast_ifp = NULL;
1402 }
1403
1404 #ifdef INET6
1405 static void
1406 carp_multicast6_cleanup(struct carp_softc *sc)
1407 {
1408         struct ip6_moptions *im6o = &sc->sc_im6o;
1409
1410         while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1411                 struct in6_multi_mship *imm =
1412                     LIST_FIRST(&im6o->im6o_memberships);
1413
1414                 LIST_REMOVE(imm, i6mm_chain);
1415                 in6_leavegroup(imm);
1416         }
1417         im6o->im6o_multicast_ifp = NULL;
1418 }
1419 #endif
1420
1421 static int
1422 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1423 {
1424         struct ifnet *ifp;
1425         struct carp_if *cif;
1426         struct in_ifaddr *ia, *ia_if;
1427         struct ip_moptions *imo = &sc->sc_imo;
1428         struct in_addr addr;
1429         u_long iaddr = htonl(sin->sin_addr.s_addr);
1430         int own, error;
1431
1432         if (sin->sin_addr.s_addr == 0) {
1433                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1434                         carp_set_state(sc, INIT);
1435                 if (sc->sc_naddrs)
1436                         SC2IFP(sc)->if_flags |= IFF_UP;
1437                 carp_setrun(sc, 0);
1438                 return (0);
1439         }
1440
1441         /* we have to do it by hands to check we won't match on us */
1442         ia_if = NULL; own = 0;
1443         TAILQ_FOREACH(ia, &in_ifaddrhead, ia_link) {
1444                 /* and, yeah, we need a multicast-capable iface too */
1445                 if (ia->ia_ifp != SC2IFP(sc) &&
1446                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1447                     (iaddr & ia->ia_subnetmask) == ia->ia_subnet) {
1448                         if (!ia_if)
1449                                 ia_if = ia;
1450                         if (sin->sin_addr.s_addr ==
1451                             ia->ia_addr.sin_addr.s_addr)
1452                                 own++;
1453                 }
1454         }
1455
1456         if (!ia_if)
1457                 return (EADDRNOTAVAIL);
1458
1459         ia = ia_if;
1460         ifp = ia->ia_ifp;
1461
1462         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1463             (imo->imo_multicast_ifp && imo->imo_multicast_ifp != ifp))
1464                 return (EADDRNOTAVAIL);
1465
1466         if (imo->imo_num_memberships == 0) {
1467                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1468                 if ((imo->imo_membership[0] = in_addmulti(&addr, ifp)) == NULL)
1469                         return (ENOBUFS);
1470                 imo->imo_num_memberships++;
1471                 imo->imo_multicast_ifp = ifp;
1472                 imo->imo_multicast_ttl = CARP_DFLTTL;
1473                 imo->imo_multicast_loop = 0;
1474         }
1475
1476         if (!ifp->if_carp) {
1477
1478                 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP,
1479                     M_WAITOK|M_ZERO);
1480                 if (!cif) {
1481                         error = ENOBUFS;
1482                         goto cleanup;
1483                 }
1484                 if ((error = ifpromisc(ifp, 1))) {
1485                         FREE(cif, M_CARP);
1486                         goto cleanup;
1487                 }
1488
1489                 CARP_LOCK_INIT(cif);
1490                 CARP_LOCK(cif);
1491                 cif->vhif_ifp = ifp;
1492                 TAILQ_INIT(&cif->vhif_vrs);
1493                 ifp->if_carp = cif;
1494
1495         } else {
1496                 struct carp_softc *vr;
1497
1498                 cif = (struct carp_if *)ifp->if_carp;
1499                 CARP_LOCK(cif);
1500                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1501                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1502                                 CARP_UNLOCK(cif);
1503                                 error = EINVAL;
1504                                 goto cleanup;
1505                         }
1506         }
1507         sc->sc_ia = ia;
1508         sc->sc_carpdev = ifp;
1509
1510         { /* XXX prevent endless loop if already in queue */
1511         struct carp_softc *vr, *after = NULL;
1512         int myself = 0;
1513         cif = (struct carp_if *)ifp->if_carp;
1514
1515         /* XXX: cif should not change, right? So we still hold the lock */
1516         CARP_LOCK_ASSERT(cif);
1517
1518         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1519                 if (vr == sc)
1520                         myself = 1;
1521                 if (vr->sc_vhid < sc->sc_vhid)
1522                         after = vr;
1523         }
1524
1525         if (!myself) {
1526                 /* We're trying to keep things in order */
1527                 if (after == NULL) {
1528                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1529                 } else {
1530                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1531                 }
1532                 cif->vhif_nvrs++;
1533         }
1534         }
1535
1536         sc->sc_naddrs++;
1537         SC2IFP(sc)->if_flags |= IFF_UP;
1538         if (own)
1539                 sc->sc_advskew = 0;
1540         carp_sc_state_locked(sc);
1541         carp_setrun(sc, 0);
1542
1543         CARP_UNLOCK(cif);
1544
1545         return (0);
1546
1547 cleanup:
1548         in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1549         return (error);
1550 }
1551
1552 static int
1553 carp_del_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1554 {
1555         int error = 0;
1556
1557         if (!--sc->sc_naddrs) {
1558                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1559                 struct ip_moptions *imo = &sc->sc_imo;
1560
1561                 CARP_LOCK(cif);
1562                 callout_stop(&sc->sc_ad_tmo);
1563                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1564                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1565                 sc->sc_vhid = -1;
1566                 in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1567                 imo->imo_multicast_ifp = NULL;
1568                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1569                 if (!--cif->vhif_nvrs) {
1570                         sc->sc_carpdev->if_carp = NULL;
1571                         CARP_LOCK_DESTROY(cif);
1572                         FREE(cif, M_IFADDR);
1573                 } else {
1574                         CARP_UNLOCK(cif);
1575                 }
1576         }
1577
1578         return (error);
1579 }
1580
1581 #ifdef INET6
1582 static int
1583 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1584 {
1585         struct ifnet *ifp;
1586         struct carp_if *cif;
1587         struct in6_ifaddr *ia, *ia_if;
1588         struct ip6_moptions *im6o = &sc->sc_im6o;
1589         struct in6_multi_mship *imm;
1590         struct in6_addr in6;
1591         int own, error;
1592
1593         if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
1594                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1595                         carp_set_state(sc, INIT);
1596                 if (sc->sc_naddrs6)
1597                         SC2IFP(sc)->if_flags |= IFF_UP;
1598                 carp_setrun(sc, 0);
1599                 return (0);
1600         }
1601
1602         /* we have to do it by hands to check we won't match on us */
1603         ia_if = NULL; own = 0;
1604         for (ia = in6_ifaddr; ia; ia = ia->ia_next) {
1605                 int i;
1606
1607                 for (i = 0; i < 4; i++) {
1608                         if ((sin6->sin6_addr.s6_addr32[i] &
1609                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]) !=
1610                             (ia->ia_addr.sin6_addr.s6_addr32[i] &
1611                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]))
1612                                 break;
1613                 }
1614                 /* and, yeah, we need a multicast-capable iface too */
1615                 if (ia->ia_ifp != SC2IFP(sc) &&
1616                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1617                     (i == 4)) {
1618                         if (!ia_if)
1619                                 ia_if = ia;
1620                         if (IN6_ARE_ADDR_EQUAL(&sin6->sin6_addr,
1621                             &ia->ia_addr.sin6_addr))
1622                                 own++;
1623                 }
1624         }
1625
1626         if (!ia_if)
1627                 return (EADDRNOTAVAIL);
1628         ia = ia_if;
1629         ifp = ia->ia_ifp;
1630
1631         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1632             (im6o->im6o_multicast_ifp && im6o->im6o_multicast_ifp != ifp))
1633                 return (EADDRNOTAVAIL);
1634
1635         if (!sc->sc_naddrs6) {
1636                 im6o->im6o_multicast_ifp = ifp;
1637
1638                 /* join CARP multicast address */
1639                 bzero(&in6, sizeof(in6));
1640                 in6.s6_addr16[0] = htons(0xff02);
1641                 in6.s6_addr8[15] = 0x12;
1642                 if (in6_setscope(&in6, ifp, NULL) != 0)
1643                         goto cleanup;
1644                 if ((imm = in6_joingroup(ifp, &in6, &error, 0)) == NULL)
1645                         goto cleanup;
1646                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1647
1648                 /* join solicited multicast address */
1649                 bzero(&in6, sizeof(in6));
1650                 in6.s6_addr16[0] = htons(0xff02);
1651                 in6.s6_addr32[1] = 0;
1652                 in6.s6_addr32[2] = htonl(1);
1653                 in6.s6_addr32[3] = sin6->sin6_addr.s6_addr32[3];
1654                 in6.s6_addr8[12] = 0xff;
1655                 if (in6_setscope(&in6, ifp, NULL) != 0)
1656                         goto cleanup;
1657                 if ((imm = in6_joingroup(ifp, &in6, &error, 0)) == NULL)
1658                         goto cleanup;
1659                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1660         }
1661
1662         if (!ifp->if_carp) {
1663                 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP,
1664                     M_WAITOK|M_ZERO);
1665                 if (!cif) {
1666                         error = ENOBUFS;
1667                         goto cleanup;
1668                 }
1669                 if ((error = ifpromisc(ifp, 1))) {
1670                         FREE(cif, M_CARP);
1671                         goto cleanup;
1672                 }
1673
1674                 CARP_LOCK_INIT(cif);
1675                 CARP_LOCK(cif);
1676                 cif->vhif_ifp = ifp;
1677                 TAILQ_INIT(&cif->vhif_vrs);
1678                 ifp->if_carp = cif;
1679
1680         } else {
1681                 struct carp_softc *vr;
1682
1683                 cif = (struct carp_if *)ifp->if_carp;
1684                 CARP_LOCK(cif);
1685                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1686                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1687                                 CARP_UNLOCK(cif);
1688                                 error = EINVAL;
1689                                 goto cleanup;
1690                         }
1691         }
1692         sc->sc_ia6 = ia;
1693         sc->sc_carpdev = ifp;
1694
1695         { /* XXX prevent endless loop if already in queue */
1696         struct carp_softc *vr, *after = NULL;
1697         int myself = 0;
1698         cif = (struct carp_if *)ifp->if_carp;
1699         CARP_LOCK_ASSERT(cif);
1700
1701         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1702                 if (vr == sc)
1703                         myself = 1;
1704                 if (vr->sc_vhid < sc->sc_vhid)
1705                         after = vr;
1706         }
1707
1708         if (!myself) {
1709                 /* We're trying to keep things in order */
1710                 if (after == NULL) {
1711                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1712                 } else {
1713                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1714                 }
1715                 cif->vhif_nvrs++;
1716         }
1717         }
1718
1719         sc->sc_naddrs6++;
1720         SC2IFP(sc)->if_flags |= IFF_UP;
1721         if (own)
1722                 sc->sc_advskew = 0;
1723         carp_sc_state_locked(sc);
1724         carp_setrun(sc, 0);
1725
1726         CARP_UNLOCK(cif);
1727
1728         return (0);
1729
1730 cleanup:
1731         /* clean up multicast memberships */
1732         if (!sc->sc_naddrs6) {
1733                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1734                         imm = LIST_FIRST(&im6o->im6o_memberships);
1735                         LIST_REMOVE(imm, i6mm_chain);
1736                         in6_leavegroup(imm);
1737                 }
1738         }
1739         return (error);
1740 }
1741
1742 static int
1743 carp_del_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1744 {
1745         int error = 0;
1746
1747         if (!--sc->sc_naddrs6) {
1748                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1749                 struct ip6_moptions *im6o = &sc->sc_im6o;
1750
1751                 CARP_LOCK(cif);
1752                 callout_stop(&sc->sc_ad_tmo);
1753                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1754                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1755                 sc->sc_vhid = -1;
1756                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1757                         struct in6_multi_mship *imm =
1758                             LIST_FIRST(&im6o->im6o_memberships);
1759
1760                         LIST_REMOVE(imm, i6mm_chain);
1761                         in6_leavegroup(imm);
1762                 }
1763                 im6o->im6o_multicast_ifp = NULL;
1764                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1765                 if (!--cif->vhif_nvrs) {
1766                         CARP_LOCK_DESTROY(cif);
1767                         sc->sc_carpdev->if_carp = NULL;
1768                         FREE(cif, M_IFADDR);
1769                 } else
1770                         CARP_UNLOCK(cif);
1771         }
1772
1773         return (error);
1774 }
1775 #endif /* INET6 */
1776
1777 static int
1778 carp_ioctl(struct ifnet *ifp, u_long cmd, caddr_t addr)
1779 {
1780         struct carp_softc *sc = ifp->if_softc, *vr;
1781         struct carpreq carpr;
1782         struct ifaddr *ifa;
1783         struct ifreq *ifr;
1784         struct ifaliasreq *ifra;
1785         int locked = 0, error = 0;
1786
1787         ifa = (struct ifaddr *)addr;
1788         ifra = (struct ifaliasreq *)addr;
1789         ifr = (struct ifreq *)addr;
1790
1791         switch (cmd) {
1792         case SIOCSIFADDR:
1793                 switch (ifa->ifa_addr->sa_family) {
1794 #ifdef INET
1795                 case AF_INET:
1796                         SC2IFP(sc)->if_flags |= IFF_UP;
1797                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1798                             sizeof(struct sockaddr));
1799                         error = carp_set_addr(sc, satosin(ifa->ifa_addr));
1800                         break;
1801 #endif /* INET */
1802 #ifdef INET6
1803                 case AF_INET6:
1804                         SC2IFP(sc)->if_flags |= IFF_UP;
1805                         error = carp_set_addr6(sc, satosin6(ifa->ifa_addr));
1806                         break;
1807 #endif /* INET6 */
1808                 default:
1809                         error = EAFNOSUPPORT;
1810                         break;
1811                 }
1812                 break;
1813
1814         case SIOCAIFADDR:
1815                 switch (ifa->ifa_addr->sa_family) {
1816 #ifdef INET
1817                 case AF_INET:
1818                         SC2IFP(sc)->if_flags |= IFF_UP;
1819                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1820                             sizeof(struct sockaddr));
1821                         error = carp_set_addr(sc, satosin(&ifra->ifra_addr));
1822                         break;
1823 #endif /* INET */
1824 #ifdef INET6
1825                 case AF_INET6:
1826                         SC2IFP(sc)->if_flags |= IFF_UP;
1827                         error = carp_set_addr6(sc, satosin6(&ifra->ifra_addr));
1828                         break;
1829 #endif /* INET6 */
1830                 default:
1831                         error = EAFNOSUPPORT;
1832                         break;
1833                 }
1834                 break;
1835
1836         case SIOCDIFADDR:
1837                 switch (ifa->ifa_addr->sa_family) {
1838 #ifdef INET
1839                 case AF_INET:
1840                         error = carp_del_addr(sc, satosin(&ifra->ifra_addr));
1841                         break;
1842 #endif /* INET */
1843 #ifdef INET6
1844                 case AF_INET6:
1845                         error = carp_del_addr6(sc, satosin6(&ifra->ifra_addr));
1846                         break;
1847 #endif /* INET6 */
1848                 default:
1849                         error = EAFNOSUPPORT;
1850                         break;
1851                 }
1852                 break;
1853
1854         case SIOCSIFFLAGS:
1855                 if (sc->sc_carpdev) {
1856                         locked = 1;
1857                         CARP_SCLOCK(sc);
1858                 }
1859                 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) {
1860                         callout_stop(&sc->sc_ad_tmo);
1861                         callout_stop(&sc->sc_md_tmo);
1862                         callout_stop(&sc->sc_md6_tmo);
1863                         if (sc->sc_state == MASTER)
1864                                 carp_send_ad_locked(sc);
1865                         carp_set_state(sc, INIT);
1866                         carp_setrun(sc, 0);
1867                 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) {
1868                         SC2IFP(sc)->if_flags |= IFF_UP;
1869                         carp_setrun(sc, 0);
1870                 }
1871                 break;
1872
1873         case SIOCSVH:
1874                 error = priv_check(curthread, PRIV_NETINET_CARP);
1875                 if (error)
1876                         break;
1877                 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1878                         break;
1879                 error = 1;
1880                 if (sc->sc_carpdev) {
1881                         locked = 1;
1882                         CARP_SCLOCK(sc);
1883                 }
1884                 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) {
1885                         switch (carpr.carpr_state) {
1886                         case BACKUP:
1887                                 callout_stop(&sc->sc_ad_tmo);
1888                                 carp_set_state(sc, BACKUP);
1889                                 carp_setrun(sc, 0);
1890                                 carp_setroute(sc, RTM_DELETE);
1891                                 break;
1892                         case MASTER:
1893                                 carp_master_down_locked(sc);
1894                                 break;
1895                         default:
1896                                 break;
1897                         }
1898                 }
1899                 if (carpr.carpr_vhid > 0) {
1900                         if (carpr.carpr_vhid > 255) {
1901                                 error = EINVAL;
1902                                 break;
1903                         }
1904                         if (sc->sc_carpdev) {
1905                                 struct carp_if *cif;
1906                                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1907                                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1908                                         if (vr != sc &&
1909                                             vr->sc_vhid == carpr.carpr_vhid)
1910                                                 return EEXIST;
1911                         }
1912                         sc->sc_vhid = carpr.carpr_vhid;
1913                         IF_LLADDR(sc->sc_ifp)[0] = 0;
1914                         IF_LLADDR(sc->sc_ifp)[1] = 0;
1915                         IF_LLADDR(sc->sc_ifp)[2] = 0x5e;
1916                         IF_LLADDR(sc->sc_ifp)[3] = 0;
1917                         IF_LLADDR(sc->sc_ifp)[4] = 1;
1918                         IF_LLADDR(sc->sc_ifp)[5] = sc->sc_vhid;
1919                         error--;
1920                 }
1921                 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) {
1922                         if (carpr.carpr_advskew >= 255) {
1923                                 error = EINVAL;
1924                                 break;
1925                         }
1926                         if (carpr.carpr_advbase > 255) {
1927                                 error = EINVAL;
1928                                 break;
1929                         }
1930                         sc->sc_advbase = carpr.carpr_advbase;
1931                         sc->sc_advskew = carpr.carpr_advskew;
1932                         error--;
1933                 }
1934                 bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1935                 if (error > 0)
1936                         error = EINVAL;
1937                 else {
1938                         error = 0;
1939                         carp_setrun(sc, 0);
1940                 }
1941                 break;
1942
1943         case SIOCGVH:
1944                 /* XXX: lockless read */
1945                 bzero(&carpr, sizeof(carpr));
1946                 carpr.carpr_state = sc->sc_state;
1947                 carpr.carpr_vhid = sc->sc_vhid;
1948                 carpr.carpr_advbase = sc->sc_advbase;
1949                 carpr.carpr_advskew = sc->sc_advskew;
1950                 error = priv_check(curthread, PRIV_NETINET_CARP);
1951                 if (error == 0)
1952                         bcopy(sc->sc_key, carpr.carpr_key,
1953                             sizeof(carpr.carpr_key));
1954                 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1955                 break;
1956
1957         default:
1958                 error = EINVAL;
1959         }
1960
1961         if (locked)
1962                 CARP_SCUNLOCK(sc);
1963
1964         carp_hmac_prepare(sc);
1965
1966         return (error);
1967 }
1968
1969 /*
1970  * XXX: this is looutput. We should eventually use it from there.
1971  */
1972 static int
1973 carp_looutput(struct ifnet *ifp, struct mbuf *m, struct sockaddr *dst,
1974     struct rtentry *rt)
1975 {
1976         u_int32_t af;
1977
1978         M_ASSERTPKTHDR(m); /* check if we have the packet header */
1979
1980         if (rt && rt->rt_flags & (RTF_REJECT|RTF_BLACKHOLE)) {
1981                 m_freem(m);
1982                 return (rt->rt_flags & RTF_BLACKHOLE ? 0 :
1983                         rt->rt_flags & RTF_HOST ? EHOSTUNREACH : ENETUNREACH);
1984         }
1985
1986         ifp->if_opackets++;
1987         ifp->if_obytes += m->m_pkthdr.len;
1988
1989         /* BPF writes need to be handled specially. */
1990         if (dst->sa_family == AF_UNSPEC) {
1991                 bcopy(dst->sa_data, &af, sizeof(af));
1992                 dst->sa_family = af;
1993         }
1994
1995 #if 1   /* XXX */
1996         switch (dst->sa_family) {
1997         case AF_INET:
1998         case AF_INET6:
1999         case AF_IPX:
2000         case AF_APPLETALK:
2001                 break;
2002         default:
2003                 printf("carp_looutput: af=%d unexpected\n", dst->sa_family);
2004                 m_freem(m);
2005                 return (EAFNOSUPPORT);
2006         }
2007 #endif
2008         return(if_simloop(ifp, m, dst->sa_family, 0));
2009 }
2010
2011 /*
2012  * Start output on carp interface. This function should never be called.
2013  */
2014 static void
2015 carp_start(struct ifnet *ifp)
2016 {
2017 #ifdef DEBUG
2018         printf("%s: start called\n", ifp->if_xname);
2019 #endif
2020 }
2021
2022 int
2023 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *sa,
2024     struct rtentry *rt)
2025 {
2026         struct m_tag *mtag;
2027         struct carp_softc *sc;
2028         struct ifnet *carp_ifp;
2029
2030         if (!sa)
2031                 return (0);
2032
2033         switch (sa->sa_family) {
2034 #ifdef INET
2035         case AF_INET:
2036                 break;
2037 #endif /* INET */
2038 #ifdef INET6
2039         case AF_INET6:
2040                 break;
2041 #endif /* INET6 */
2042         default:
2043                 return (0);
2044         }
2045
2046         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
2047         if (mtag == NULL)
2048                 return (0);
2049
2050         bcopy(mtag + 1, &carp_ifp, sizeof(struct ifnet *));
2051         sc = carp_ifp->if_softc;
2052
2053         /* Set the source MAC address to Virtual Router MAC Address */
2054         switch (ifp->if_type) {
2055         case IFT_ETHER:
2056         case IFT_L2VLAN: {
2057                         struct ether_header *eh;
2058
2059                         eh = mtod(m, struct ether_header *);
2060                         eh->ether_shost[0] = 0;
2061                         eh->ether_shost[1] = 0;
2062                         eh->ether_shost[2] = 0x5e;
2063                         eh->ether_shost[3] = 0;
2064                         eh->ether_shost[4] = 1;
2065                         eh->ether_shost[5] = sc->sc_vhid;
2066                 }
2067                 break;
2068         case IFT_FDDI: {
2069                         struct fddi_header *fh;
2070
2071                         fh = mtod(m, struct fddi_header *);
2072                         fh->fddi_shost[0] = 0;
2073                         fh->fddi_shost[1] = 0;
2074                         fh->fddi_shost[2] = 0x5e;
2075                         fh->fddi_shost[3] = 0;
2076                         fh->fddi_shost[4] = 1;
2077                         fh->fddi_shost[5] = sc->sc_vhid;
2078                 }
2079                 break;
2080         case IFT_ISO88025: {
2081                         struct iso88025_header *th;
2082                         th = mtod(m, struct iso88025_header *);
2083                         th->iso88025_shost[0] = 3;
2084                         th->iso88025_shost[1] = 0;
2085                         th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
2086                         th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
2087                         th->iso88025_shost[4] = 0;
2088                         th->iso88025_shost[5] = 0;
2089                 }
2090                 break;
2091         default:
2092                 printf("%s: carp is not supported for this interface type\n",
2093                     ifp->if_xname);
2094                 return (EOPNOTSUPP);
2095         }
2096
2097         return (0);
2098 }
2099
2100 static void
2101 carp_set_state(struct carp_softc *sc, int state)
2102 {
2103
2104         if (sc->sc_carpdev)
2105                 CARP_SCLOCK_ASSERT(sc);
2106
2107         if (sc->sc_state == state)
2108                 return;
2109
2110         sc->sc_state = state;
2111         switch (state) {
2112         case BACKUP:
2113                 SC2IFP(sc)->if_link_state = LINK_STATE_DOWN;
2114                 break;
2115         case MASTER:
2116                 SC2IFP(sc)->if_link_state = LINK_STATE_UP;
2117                 break;
2118         default:
2119                 SC2IFP(sc)->if_link_state = LINK_STATE_UNKNOWN;
2120                 break;
2121         }
2122         rt_ifmsg(SC2IFP(sc));
2123 }
2124
2125 void
2126 carp_carpdev_state(void *v)
2127 {
2128         struct carp_if *cif = v;
2129
2130         CARP_LOCK(cif);
2131         carp_carpdev_state_locked(cif);
2132         CARP_UNLOCK(cif);
2133 }
2134
2135 static void
2136 carp_carpdev_state_locked(struct carp_if *cif)
2137 {
2138         struct carp_softc *sc;
2139
2140         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list)
2141                 carp_sc_state_locked(sc);
2142 }
2143
2144 static void
2145 carp_sc_state_locked(struct carp_softc *sc)
2146 {
2147         CARP_SCLOCK_ASSERT(sc);
2148
2149         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
2150             !(sc->sc_carpdev->if_flags & IFF_UP)) {
2151                 sc->sc_flags_backup = SC2IFP(sc)->if_flags;
2152                 SC2IFP(sc)->if_flags &= ~IFF_UP;
2153                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
2154                 callout_stop(&sc->sc_ad_tmo);
2155                 callout_stop(&sc->sc_md_tmo);
2156                 callout_stop(&sc->sc_md6_tmo);
2157                 carp_set_state(sc, INIT);
2158                 carp_setrun(sc, 0);
2159                 if (!sc->sc_suppress) {
2160                         carp_suppress_preempt++;
2161                         if (carp_suppress_preempt == 1) {
2162                                 CARP_SCUNLOCK(sc);
2163                                 carp_send_ad_all();
2164                                 CARP_SCLOCK(sc);
2165                         }
2166                 }
2167                 sc->sc_suppress = 1;
2168         } else {
2169                 SC2IFP(sc)->if_flags |= sc->sc_flags_backup;
2170                 carp_set_state(sc, INIT);
2171                 carp_setrun(sc, 0);
2172                 if (sc->sc_suppress)
2173                         carp_suppress_preempt--;
2174                 sc->sc_suppress = 0;
2175         }
2176
2177         return;
2178 }
2179
2180 static int
2181 carp_modevent(module_t mod, int type, void *data)
2182 {
2183         switch (type) {
2184         case MOD_LOAD:
2185                 if_detach_event_tag = EVENTHANDLER_REGISTER(ifnet_departure_event,
2186                     carp_ifdetach, NULL, EVENTHANDLER_PRI_ANY);
2187                 if (if_detach_event_tag == NULL)
2188                         return (ENOMEM);
2189                 mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2190                 LIST_INIT(&carpif_list);
2191                 if_clone_attach(&carp_cloner);
2192                 break;
2193
2194         case MOD_UNLOAD:
2195                 EVENTHANDLER_DEREGISTER(ifnet_departure_event, if_detach_event_tag);
2196                 if_clone_detach(&carp_cloner);
2197                 mtx_destroy(&carp_mtx);
2198                 break;
2199
2200         default:
2201                 return (EINVAL);
2202         }
2203
2204         return (0);
2205 }
2206
2207 static moduledata_t carp_mod = {
2208         "carp",
2209         carp_modevent,
2210         0
2211 };
2212
2213 DECLARE_MODULE(carp, carp_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);