sys/netinet/ip_carp.c

   1 /*
   2  * Copyright (c) 2002 Michael Shalayeff. All rights reserved.
   3  * Copyright (c) 2003 Ryan McBride. All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  *
  14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  15  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  16  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  17  * IN NO EVENT SHALL THE AUTHOR OR HIS RELATIVES BE LIABLE FOR ANY DIRECT,
  18  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  19  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  20  * SERVICES; LOSS OF MIND, USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
  22  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
  23  * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
  24  * THE POSSIBILITY OF SUCH DAMAGE.
  25  */
  26 /*
  27  * $FreeBSD: src/sys/netinet/ip_carp.c,v 1.48 2007/02/02 09:39:09 glebius Exp $
  28  * $DragonFly: src/sys/netinet/ip_carp.c,v 1.3 2007/08/27 16:15:42 hasso Exp $
  29  */
  30
  31 #include "opt_carp.h"
  32 /*#include "opt_bpf.h"*/
  33 #include "opt_inet.h"
  34 #include "opt_inet6.h"
  35
  36 #include <sys/types.h>
  37 #include <sys/param.h>
  38 #include <sys/systm.h>
  39 #include <sys/conf.h>
  40 #include <sys/kernel.h>
  41 #include <machine/limits.h>
  42 #include <sys/malloc.h>
  43 #include <sys/mbuf.h>
  44 #include <sys/module.h>
  45 #include <sys/time.h>
  46 #include <sys/proc.h>
  47 #include <sys/sysctl.h>
  48 #include <sys/syslog.h>
  49 #include <sys/signalvar.h>
  50 #include <sys/filio.h>
  51 #include <sys/sockio.h>
  52 #include <sys/in_cksum.h>
  53 #include <sys/socket.h>
  54 #include <sys/vnode.h>
  55
  56 #include <machine/stdarg.h>
  57
  58 #include <net/bpf.h>
  59 #include <net/ethernet.h>
  60 #include <net/if.h>
  61 #include <net/if_dl.h>
  62 #include <net/if_types.h>
  63 #include <net/route.h>
  64
  65 #ifdef INET
  66 #include <netinet/in.h>
  67 #include <netinet/in_var.h>
  68 #include <netinet/in_systm.h>
  69 #include <netinet/ip.h>
  70 #include <netinet/ip_var.h>
  71 #include <netinet/if_ether.h>
  72 #endif
  73
  74 #ifdef INET6
  75 #include <netinet/icmp6.h>
  76 #include <netinet/ip6.h>
  77 #include <netinet6/ip6_var.h>
  78 #include <netinet6/scope6_var.h>
  79 #include <netinet6/nd6.h>
  80 #endif
  81
  82 #include <crypto/sha1.h>
  83 #include <netinet/ip_carp.h>
  84 #include <sys/lock.h>
  85
  86 #define CARP_IFNAME     "carp"
  87 static MALLOC_DEFINE(M_CARP, "CARP", "CARP interfaces");
  88 static MALLOC_DEFINE(M_IFNET, "IFNET", "IFNET CARP?");
  89 SYSCTL_DECL(_net_inet_carp);
  90
  91 struct carp_softc {
  92         struct ifnet            *sc_ifp;        /* Interface clue */
  93         struct ifnet            *sc_carpdev;    /* Pointer to parent interface */
  94         struct in_ifaddr        *sc_ia;         /* primary iface address */
  95         struct ip_moptions       sc_imo;
  96 #ifdef INET6
  97         struct in6_ifaddr       *sc_ia6;        /* primary iface address v6 */
  98         struct ip6_moptions      sc_im6o;
  99 #endif /* INET6 */
 100         TAILQ_ENTRY(carp_softc)  sc_list;
 101
 102         enum { INIT = 0, BACKUP, MASTER }       sc_state;
 103
 104         int                      sc_flags_backup;
 105         int                      sc_suppress;
 106
 107         int                      sc_sendad_errors;
 108 #define CARP_SENDAD_MAX_ERRORS  3
 109         int                      sc_sendad_success;
 110 #define CARP_SENDAD_MIN_SUCCESS 3
 111
 112         int                      sc_vhid;
 113         int                      sc_advskew;
 114         int                      sc_naddrs;
 115         int                      sc_naddrs6;
 116         int                      sc_advbase;    /* seconds */
 117         int                      sc_init_counter;
 118         u_int64_t                sc_counter;
 119
 120         /* authentication */
 121 #define CARP_HMAC_PAD   64
 122         unsigned char sc_key[CARP_KEY_LEN];
 123         unsigned char sc_pad[CARP_HMAC_PAD];
 124         SHA1_CTX sc_sha1;
 125
 126         struct callout           sc_ad_tmo;     /* advertisement timeout */
 127         struct callout           sc_md_tmo;     /* master down timeout */
 128         struct callout           sc_md6_tmo;    /* master down timeout */
 129
 130         LIST_ENTRY(carp_softc)   sc_next;       /* Interface clue */
 131 };
 132 #define SC2IFP(sc)      ((sc)->sc_ifp)
 133
 134 int carp_suppress_preempt = 0;
 135 int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 1, 0, 0 };    /* XXX for now */
 136 SYSCTL_INT(_net_inet_carp, CARPCTL_ALLOW, allow, CTLFLAG_RW,
 137     &carp_opts[CARPCTL_ALLOW], 0, "Accept incoming CARP packets");
 138 SYSCTL_INT(_net_inet_carp, CARPCTL_PREEMPT, preempt, CTLFLAG_RW,
 139     &carp_opts[CARPCTL_PREEMPT], 0, "high-priority backup preemption mode");
 140 SYSCTL_INT(_net_inet_carp, CARPCTL_LOG, log, CTLFLAG_RW,
 141     &carp_opts[CARPCTL_LOG], 0, "log bad carp packets");
 142 SYSCTL_INT(_net_inet_carp, CARPCTL_ARPBALANCE, arpbalance, CTLFLAG_RW,
 143     &carp_opts[CARPCTL_ARPBALANCE], 0, "balance arp responses");
 144 SYSCTL_INT(_net_inet_carp, OID_AUTO, suppress_preempt, CTLFLAG_RD,
 145     &carp_suppress_preempt, 0, "Preemption is suppressed");
 146
 147 struct carpstats carpstats;
 148 SYSCTL_STRUCT(_net_inet_carp, CARPCTL_STATS, stats, CTLFLAG_RW,
 149     &carpstats, carpstats,
 150     "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 151
 152 struct carp_if {
 153         TAILQ_HEAD(, carp_softc) vhif_vrs;
 154         int vhif_nvrs;
 155
 156         struct ifnet    *vhif_ifp;
 157         struct lock     vhif_lock;
 158 };
 159
 160 /* Get carp_if from softc. Valid after carp_set_addr{,6}. */
 161 #define SC2CIF(sc)              ((struct carp_if *)(sc)->sc_carpdev->if_carp)
 162
 163 #define CARP_LOCK_INIT(cif)     lockinit(&(cif)->vhif_lock, "carp_if", 0,  LK_NOWAIT);
 164 #define CARP_LOCK_DESTROY(cif)  ;
 165 #define CARP_LOCK_ASSERT(cif)   ;
 166 #define CARP_LOCK(cif)          lockmgr(&(cif)->vhif_lock, LK_EXCLUSIVE);
 167 #define CARP_UNLOCK(cif)        lockmgr(&(cif)->vhif_lock, LK_RELEASE);
 168
 169 #define CARP_SCLOCK(sc)         lockmgr(&SC2CIF(sc)->vhif_lock, LK_EXCLUSIVE);
 170 #define CARP_SCUNLOCK(sc)       lockmgr(&SC2CIF(sc)->vhif_lock, LK_RELEASE);
 171 #define CARP_SCLOCK_ASSERT(sc)  ;
 172
 173 #define CARP_LOG(...)   do {                            \
 174         if (carp_opts[CARPCTL_LOG] > 0)                 \
 175                 log(LOG_INFO, __VA_ARGS__);             \
 176 } while (0)
 177
 178 #define CARP_DEBUG(...) do {                            \
 179         if (carp_opts[CARPCTL_LOG] > 1)                 \
 180                 log(LOG_DEBUG, __VA_ARGS__);            \
 181 } while (0)
 182
 183 static void     carp_hmac_prepare(struct carp_softc *);
 184 static void     carp_hmac_generate(struct carp_softc *, u_int32_t *,
 185                     unsigned char *);
 186 static int      carp_hmac_verify(struct carp_softc *, u_int32_t *,
 187                     unsigned char *);
 188 static void     carp_setroute(struct carp_softc *, int);
 189 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 190 static int      carp_clone_create(struct if_clone *, int);
 191 static void     carp_clone_destroy(struct ifnet *);
 192 static void     carpdetach(struct carp_softc *, int);
 193 static int      carp_prepare_ad(struct mbuf *, struct carp_softc *,
 194                     struct carp_header *);
 195 static void     carp_send_ad_all(void);
 196 static void     carp_send_ad(void *);
 197 static void     carp_send_ad_locked(struct carp_softc *);
 198 static void     carp_send_arp(struct carp_softc *);
 199 static void     carp_master_down(void *);
 200 static void     carp_master_down_locked(struct carp_softc *);
 201 static int      carp_ioctl(struct ifnet *, u_long, caddr_t, struct ucred *);
 202 static int      carp_looutput(struct ifnet *, struct mbuf *, struct sockaddr *,
 203                     struct rtentry *);
 204 static void     carp_start(struct ifnet *);
 205 static void     carp_setrun(struct carp_softc *, sa_family_t);
 206 static void     carp_set_state(struct carp_softc *, int);
 207 static int      carp_addrcount(struct carp_if *, struct in_ifaddr *, int);
 208 enum    { CARP_COUNT_MASTER, CARP_COUNT_RUNNING };
 209
 210 static void     carp_multicast_cleanup(struct carp_softc *);
 211 static int      carp_set_addr(struct carp_softc *, struct sockaddr_in *);
 212 static int      carp_del_addr(struct carp_softc *, struct sockaddr_in *);
 213 static void     carp_carpdev_state_locked(struct carp_if *);
 214 static void     carp_sc_state_locked(struct carp_softc *);
 215 #ifdef INET6
 216 static void     carp_send_na(struct carp_softc *);
 217 static int      carp_set_addr6(struct carp_softc *, struct sockaddr_in6 *);
 218 static int      carp_del_addr6(struct carp_softc *, struct sockaddr_in6 *);
 219 static void     carp_multicast6_cleanup(struct carp_softc *);
 220 #endif
 221
 222 static LIST_HEAD(, carp_softc) carpif_list;
 223
 224 struct if_clone carp_cloner = IF_CLONE_INITIALIZER(CARP_IFNAME, carp_clone_create, carp_clone_destroy, 0, IF_MAXUNIT);
 225
 226 static eventhandler_tag if_detach_event_tag;
 227
 228 static __inline u_int16_t
 229 carp_cksum(struct mbuf *m, int len)
 230 {
 231         return (in_cksum(m, len));
 232 }
 233
 234 static void
 235 carp_hmac_prepare(struct carp_softc *sc)
 236 {
 237         u_int8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 238         u_int8_t vhid = sc->sc_vhid & 0xff;
 239         struct ifaddr *ifa;
 240         int i;
 241 #ifdef INET6
 242         struct in6_addr in6;
 243 #endif
 244
 245         if (sc->sc_carpdev)
 246                 CARP_SCLOCK(sc);
 247
 248         /* XXX: possible race here */
 249
 250         /* compute ipad from key */
 251         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 252         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 253         for (i = 0; i < sizeof(sc->sc_pad); i++)
 254                 sc->sc_pad[i] ^= 0x36;
 255
 256         /* precompute first part of inner hash */
 257         SHA1Init(&sc->sc_sha1);
 258         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 259         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 260         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 261         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 262 #ifdef INET
 263         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 264                 if (ifa->ifa_addr->sa_family == AF_INET)
 265                         SHA1Update(&sc->sc_sha1,
 266                             (void *)&ifatoia(ifa)->ia_addr.sin_addr.s_addr,
 267                             sizeof(struct in_addr));
 268         }
 269 #endif /* INET */
 270 #ifdef INET6
 271         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 272                 if (ifa->ifa_addr->sa_family == AF_INET6) {
 273                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 274                         in6_clearscope(&in6);
 275                         SHA1Update(&sc->sc_sha1, (void *)&in6, sizeof(in6));
 276                 }
 277         }
 278 #endif /* INET6 */
 279
 280         /* convert ipad to opad */
 281         for (i = 0; i < sizeof(sc->sc_pad); i++)
 282                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 283
 284         if (sc->sc_carpdev)
 285                 CARP_SCUNLOCK(sc);
 286 }
 287
 288 static void
 289 carp_hmac_generate(struct carp_softc *sc, u_int32_t counter[2],
 290     unsigned char md[20])
 291 {
 292         SHA1_CTX sha1ctx;
 293
 294         /* fetch first half of inner hash */
 295         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 296
 297         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 298         SHA1Final(md, &sha1ctx);
 299
 300         /* outer hash */
 301         SHA1Init(&sha1ctx);
 302         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 303         SHA1Update(&sha1ctx, md, 20);
 304         SHA1Final(md, &sha1ctx);
 305 }
 306
 307 static int
 308 carp_hmac_verify(struct carp_softc *sc, u_int32_t counter[2],
 309     unsigned char md[20])
 310 {
 311         unsigned char md2[20];
 312
 313         CARP_SCLOCK_ASSERT(sc);
 314
 315         carp_hmac_generate(sc, counter, md2);
 316
 317         return (bcmp(md, md2, sizeof(md2)));
 318 }
 319
 320 static void
 321 carp_setroute(struct carp_softc *sc, int cmd)
 322 {
 323         struct ifaddr *ifa;
 324
 325         if (sc->sc_carpdev)
 326                 CARP_SCLOCK_ASSERT(sc);
 327
 328         crit_enter();
 329         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 330                 if (ifa->ifa_addr->sa_family == AF_INET &&
 331                     sc->sc_carpdev != NULL) {
 332                         int count = carp_addrcount(
 333                             (struct carp_if *)sc->sc_carpdev->if_carp,
 334                             ifatoia(ifa), CARP_COUNT_MASTER);
 335
 336                         if ((cmd == RTM_ADD && count == 1) ||
 337                             (cmd == RTM_DELETE && count == 0))
 338                                 rtinit(ifa, cmd, RTF_UP | RTF_HOST);
 339                 }
 340 #ifdef INET6
 341                 if (ifa->ifa_addr->sa_family == AF_INET6) {
 342                         if (cmd == RTM_ADD)
 343                                 in6_ifaddloop(ifa);
 344                         else
 345                                 in6_ifremloop(ifa);
 346                 }
 347 #endif /* INET6 */
 348         }
 349         crit_exit();
 350
 351 }
 352
 353 static int
 354 carp_clone_create(struct if_clone *ifc, int unit)
 355 {
 356
 357         struct carp_softc *sc;
 358         struct ifnet *ifp;
 359
 360         MALLOC(sc, struct carp_softc *, sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
 361         ifp = SC2IFP(sc) = kmalloc(sizeof(struct ifnet), M_IFNET, M_WAITOK|M_ZERO);
 362
 363         if (ifp == NULL) {
 364                 FREE(sc, M_CARP);
 365                 return (ENOSPC);
 366         }
 367
 368         sc->sc_flags_backup = 0;
 369         sc->sc_suppress = 0;
 370         sc->sc_advbase = CARP_DFLTINTV;
 371         sc->sc_vhid = -1;       /* required setting */
 372         sc->sc_advskew = 0;
 373         sc->sc_init_counter = 1;
 374         sc->sc_naddrs = sc->sc_naddrs6 = 0; /* M_ZERO? */
 375
 376 #ifdef INET6
 377         sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL;
 378 #endif
 379
 380 /*      sc->sc_imo.imo_membership = kmalloc((sizeof(struct in_multi) * IP_MAX_MEMBERSHIPS), M_CARP,M_WAITOK);*/
 381 /*
 382         sc->sc_imo.imo_max_memberships = IP_MAX_MEMBERSHIPS;
 383         sc->sc_imo.imo_multicast_vif = -1;
 384 */
 385         callout_init(&sc->sc_ad_tmo);
 386         callout_init(&sc->sc_md_tmo);
 387         callout_init(&sc->sc_md6_tmo);
 388
 389         ifp->if_softc = sc;
 390         if_initname(ifp, CARP_IFNAME, unit);
 391         ifp->if_mtu = ETHERMTU;
 392         ifp->if_flags = IFF_LOOPBACK;
 393         ifp->if_ioctl = carp_ioctl;
 394         ifp->if_output = carp_looutput;
 395         ifp->if_start = carp_start;
 396         ifp->if_type = IFT_CARP;
 397         ifp->if_snd.ifq_maxlen = ifqmaxlen;
 398         ifp->if_hdrlen = 0;
 399         if_attach(ifp, NULL);
 400         bpfattach(ifp, DLT_NULL, sizeof(u_int));
 401
 402         crit_enter();
 403         LIST_INSERT_HEAD(&carpif_list, sc, sc_next);
 404         crit_exit();
 405
 406         return (0);
 407 }
 408
 409 static void
 410 carp_clone_destroy(struct ifnet *ifp)
 411 {
 412         struct carp_softc *sc = ifp->if_softc;
 413
 414         if (sc->sc_carpdev)
 415                 CARP_SCLOCK(sc);
 416         carpdetach(sc, 1);      /* Returns unlocked. */
 417
 418         crit_enter();
 419         LIST_REMOVE(sc, sc_next);
 420         crit_exit();
 421         bpfdetach(ifp);
 422         if_detach(ifp);
 423 /*      if_free_type(ifp, IFT_ETHER);*/
 424 /*      kfree(sc->sc_imo.imo_membership, M_CARP); */
 425         kfree(sc, M_CARP);
 426 }
 427
 428 /*
 429  * This function can be called on CARP interface destroy path,
 430  * and in case of the removal of the underlying interface as
 431  * well. We differentiate these two cases. In the latter case
 432  * we do not cleanup our multicast memberships, since they
 433  * are already freed. Also, in the latter case we do not
 434  * release the lock on return, because the function will be
 435  * called once more, for another CARP instance on the same
 436  * interface.
 437  */
 438 static void
 439 carpdetach(struct carp_softc *sc, int unlock)
 440 {
 441         struct carp_if *cif;
 442
 443         callout_stop(&sc->sc_ad_tmo);
 444         callout_stop(&sc->sc_md_tmo);
 445         callout_stop(&sc->sc_md6_tmo);
 446
 447         if (sc->sc_suppress)
 448                 carp_suppress_preempt--;
 449         sc->sc_suppress = 0;
 450
 451         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS)
 452                 carp_suppress_preempt--;
 453         sc->sc_sendad_errors = 0;
 454
 455         carp_set_state(sc, INIT);
 456         SC2IFP(sc)->if_flags &= ~IFF_UP;
 457         carp_setrun(sc, 0);
 458         if (unlock)
 459                 carp_multicast_cleanup(sc);
 460 #ifdef INET6
 461         carp_multicast6_cleanup(sc);
 462 #endif
 463
 464         if (sc->sc_carpdev != NULL) {
 465                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
 466                 CARP_LOCK_ASSERT(cif);
 467                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
 468                 if (!--cif->vhif_nvrs) {
 469                         ifpromisc(sc->sc_carpdev, 0);
 470                         sc->sc_carpdev->if_carp = NULL;
 471                         CARP_LOCK_DESTROY(cif);
 472                         FREE(cif, M_IFADDR);
 473                 } else if (unlock)
 474                         CARP_UNLOCK(cif);
 475                 sc->sc_carpdev = NULL;
 476         }
 477 }
 478
 479 /* Detach an interface from the carp. */
 480 static void
 481 carp_ifdetach(void *arg __unused, struct ifnet *ifp)
 482 {
 483         struct carp_if *cif = (struct carp_if *)ifp->if_carp;
 484         struct carp_softc *sc, *nextsc;
 485
 486         if (cif == NULL)
 487                 return;
 488
 489         /*
 490          * XXX: At the end of for() cycle the lock will be destroyed.
 491          */
 492         CARP_LOCK(cif);
 493         for (sc = TAILQ_FIRST(&cif->vhif_vrs); sc; sc = nextsc) {
 494                 nextsc = TAILQ_NEXT(sc, sc_list);
 495                 carpdetach(sc, 0);
 496         }
 497         CARP_UNLOCK(cif);
 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) && (SC2IFP(sc)->if_flags & IFF_RUNNING))) {
 672                 carpstats.carps_badvhid++;
 673                 CARP_UNLOCK(ifp->if_carp);
 674                 m_freem(m);
 675                 return;
 676         }
 677
 678         getmicrotime(&SC2IFP(sc)->if_lastchange);
 679         SC2IFP(sc)->if_ipackets++;
 680         SC2IFP(sc)->if_ibytes += m->m_pkthdr.len;
 681
 682         if (SC2IFP(sc)->if_bpf) {
 683                 struct ip *ip = mtod(m, struct ip *);
 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_mtap(SC2IFP(sc)->if_bpf, 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 = karc4random();
 800                 sc->sc_counter = sc->sc_counter << 32;
 801                 sc->sc_counter += karc4random();
 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         LIST_FOREACH(sc, &carpif_list, sc_next) {
 829                 if (sc->sc_carpdev == NULL)
 830                         continue;
 831                 CARP_SCLOCK(sc);
 832                 if ((SC2IFP(sc)->if_flags & IFF_UP) && (SC2IFP(sc)->if_flags & IFF_RUNNING) &&
 833                      sc->sc_state == MASTER)
 834                         carp_send_ad_locked(sc);
 835                 CARP_SCUNLOCK(sc);
 836         }
 837 }
 838
 839 static void
 840 carp_send_ad(void *v)
 841 {
 842         struct carp_softc *sc = v;
 843
 844         CARP_SCLOCK(sc);
 845         carp_send_ad_locked(sc);
 846         CARP_SCUNLOCK(sc);
 847 }
 848
 849 static void
 850 carp_send_ad_locked(struct carp_softc *sc)
 851 {
 852         struct carp_header ch;
 853         struct timeval tv;
 854         struct carp_header *ch_ptr;
 855         struct mbuf *m;
 856         int len, advbase, advskew;
 857
 858
 859         /* bow out if we've lost our UPness or RUNNINGuiness */
 860         if (!((SC2IFP(sc)->if_flags & IFF_UP) && (SC2IFP(sc)->if_flags & IFF_RUNNING))) {
 861                 advbase = 255;
 862                 advskew = 255;
 863         } else {
 864                 advbase = sc->sc_advbase;
 865                 if (!carp_suppress_preempt || sc->sc_advskew > 240)
 866                         advskew = sc->sc_advskew;
 867                 else
 868                         advskew = 240;
 869                 tv.tv_sec = advbase;
 870                 tv.tv_usec = advskew * 1000000 / 256;
 871         }
 872
 873         ch.carp_version = CARP_VERSION;
 874         ch.carp_type = CARP_ADVERTISEMENT;
 875         ch.carp_vhid = sc->sc_vhid;
 876         ch.carp_advbase = advbase;
 877         ch.carp_advskew = advskew;
 878         ch.carp_authlen = 7;    /* XXX DEFINE */
 879         ch.carp_pad1 = 0;       /* must be zero */
 880         ch.carp_cksum = 0;
 881
 882 #ifdef INET
 883         if (sc->sc_ia) {
 884                 struct ip *ip;
 885
 886                 MGETHDR(m, M_NOWAIT, MT_HEADER);
 887                 if (m == NULL) {
 888                         SC2IFP(sc)->if_oerrors++;
 889                         carpstats.carps_onomem++;
 890                         /* XXX maybe less ? */
 891                         if (advbase != 255 || advskew != 255)
 892                                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
 893                                     carp_send_ad, sc);
 894                         return;
 895                 }
 896                 len = sizeof(*ip) + sizeof(ch);
 897                 m->m_pkthdr.len = len;
 898                 m->m_pkthdr.rcvif = NULL;
 899                 m->m_len = len;
 900                 MH_ALIGN(m, m->m_len);
 901                 m->m_flags |= M_MCAST;
 902                 ip = mtod(m, struct ip *);
 903                 ip->ip_v = IPVERSION;
 904                 ip->ip_hl = sizeof(*ip) >> 2;
 905                 ip->ip_tos = IPTOS_LOWDELAY;
 906                 ip->ip_len = len;
 907                 ip->ip_id = ip_newid();
 908                 ip->ip_off = IP_DF;
 909                 ip->ip_ttl = CARP_DFLTTL;
 910                 ip->ip_p = IPPROTO_CARP;
 911                 ip->ip_sum = 0;
 912                 ip->ip_src.s_addr = sc->sc_ia->ia_addr.sin_addr.s_addr;
 913                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 914
 915                 ch_ptr = (struct carp_header *)(&ip[1]);
 916                 bcopy(&ch, ch_ptr, sizeof(ch));
 917                 if (carp_prepare_ad(m, sc, ch_ptr))
 918                         return;
 919
 920                 m->m_data += sizeof(*ip);
 921                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip));
 922                 m->m_data -= sizeof(*ip);
 923
 924                 getmicrotime(&SC2IFP(sc)->if_lastchange);
 925                 SC2IFP(sc)->if_opackets++;
 926                 SC2IFP(sc)->if_obytes += len;
 927                 carpstats.carps_opackets++;
 928
 929                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, NULL)) {
 930                         SC2IFP(sc)->if_oerrors++;
 931                         if (sc->sc_sendad_errors < INT_MAX)
 932                                 sc->sc_sendad_errors++;
 933                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 934                                 carp_suppress_preempt++;
 935                                 if (carp_suppress_preempt == 1) {
 936                                         CARP_SCUNLOCK(sc);
 937                                         carp_send_ad_all();
 938                                         CARP_SCLOCK(sc);
 939                                 }
 940                         }
 941                         sc->sc_sendad_success = 0;
 942                 } else {
 943                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 944                                 if (++sc->sc_sendad_success >=
 945                                     CARP_SENDAD_MIN_SUCCESS) {
 946                                         carp_suppress_preempt--;
 947                                         sc->sc_sendad_errors = 0;
 948                                 }
 949                         } else
 950                                 sc->sc_sendad_errors = 0;
 951                 }
 952         }
 953 #endif /* INET */
 954 #ifdef INET6
 955         if (sc->sc_ia6) {
 956                 struct ip6_hdr *ip6;
 957
 958                 MGETHDR(m, M_NOWAIT, MT_HEADER);
 959                 if (m == NULL) {
 960                         SC2IFP(sc)->if_oerrors++;
 961                         carpstats.carps_onomem++;
 962                         /* XXX maybe less ? */
 963                         if (advbase != 255 || advskew != 255)
 964                                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
 965                                     carp_send_ad, sc);
 966                         return;
 967                 }
 968                 len = sizeof(*ip6) + sizeof(ch);
 969                 m->m_pkthdr.len = len;
 970                 m->m_pkthdr.rcvif = NULL;
 971                 m->m_len = len;
 972                 MH_ALIGN(m, m->m_len);
 973                 m->m_flags |= M_MCAST;
 974                 ip6 = mtod(m, struct ip6_hdr *);
 975                 bzero(ip6, sizeof(*ip6));
 976                 ip6->ip6_vfc |= IPV6_VERSION;
 977                 ip6->ip6_hlim = CARP_DFLTTL;
 978                 ip6->ip6_nxt = IPPROTO_CARP;
 979                 bcopy(&sc->sc_ia6->ia_addr.sin6_addr, &ip6->ip6_src,
 980                     sizeof(struct in6_addr));
 981                 /* set the multicast destination */
 982
 983                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
 984                 ip6->ip6_dst.s6_addr8[15] = 0x12;
 985                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
 986                         SC2IFP(sc)->if_oerrors++;
 987                         m_freem(m);
 988                         CARP_LOG("%s: in6_setscope failed\n", __func__);
 989                         return;
 990                 }
 991
 992                 ch_ptr = (struct carp_header *)(&ip6[1]);
 993                 bcopy(&ch, ch_ptr, sizeof(ch));
 994                 if (carp_prepare_ad(m, sc, ch_ptr))
 995                         return;
 996
 997                 m->m_data += sizeof(*ip6);
 998                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6));
 999                 m->m_data -= sizeof(*ip6);
1000
1001                 getmicrotime(&SC2IFP(sc)->if_lastchange);
1002                 SC2IFP(sc)->if_opackets++;
1003                 SC2IFP(sc)->if_obytes += len;
1004                 carpstats.carps_opackets6++;
1005
1006                 if (ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL)) {
1007                         SC2IFP(sc)->if_oerrors++;
1008                         if (sc->sc_sendad_errors < INT_MAX)
1009                                 sc->sc_sendad_errors++;
1010                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
1011                                 carp_suppress_preempt++;
1012                                 if (carp_suppress_preempt == 1) {
1013                                         CARP_SCUNLOCK(sc);
1014                                         carp_send_ad_all();
1015                                         CARP_SCLOCK(sc);
1016                                 }
1017                         }
1018                         sc->sc_sendad_success = 0;
1019                 } else {
1020                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
1021                                 if (++sc->sc_sendad_success >=
1022                                     CARP_SENDAD_MIN_SUCCESS) {
1023                                         carp_suppress_preempt--;
1024                                         sc->sc_sendad_errors = 0;
1025                                 }
1026                         } else
1027                                 sc->sc_sendad_errors = 0;
1028                 }
1029         }
1030 #endif /* INET6 */
1031
1032         if (advbase != 255 || advskew != 255)
1033                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
1034                     carp_send_ad, sc);
1035
1036 }
1037
1038 /*
1039  * Broadcast a gratuitous ARP request containing
1040  * the virtual router MAC address for each IP address
1041  * associated with the virtual router.
1042  */
1043 static void
1044 carp_send_arp(struct carp_softc *sc)
1045 {
1046         struct ifaddr *ifa;
1047
1048         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1049
1050                 if (ifa->ifa_addr->sa_family != AF_INET)
1051                         continue;
1052                 lwkt_serialize_enter(sc->sc_carpdev->if_serializer);
1053                 arp_ifinit2(sc->sc_carpdev, ifa, IF_LLADDR(sc->sc_ifp));
1054                 lwkt_serialize_exit(sc->sc_carpdev->if_serializer);
1055
1056                 DELAY(1000);    /* XXX */
1057         }
1058 }
1059
1060 #ifdef INET6
1061 static void
1062 carp_send_na(struct carp_softc *sc)
1063 {
1064         struct ifaddr *ifa;
1065         struct in6_addr *in6;
1066         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1067
1068         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1069
1070                 if (ifa->ifa_addr->sa_family != AF_INET6)
1071                         continue;
1072
1073                 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr;
1074                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1075                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1076                 DELAY(1000);    /* XXX */
1077         }
1078 }
1079 #endif /* INET6 */
1080
1081 static int
1082 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type)
1083 {
1084         struct carp_softc *vh;
1085         struct ifaddr *ifa;
1086         int count = 0;
1087
1088         CARP_LOCK_ASSERT(cif);
1089
1090         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1091                 if ((type == CARP_COUNT_RUNNING &&
1092                     (SC2IFP(vh)->if_flags & IFF_UP) && (SC2IFP(vh)->if_flags & IFF_RUNNING)) ||
1093                     (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) {
1094                         TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1095                             ifa_list) {
1096                                 if (ifa->ifa_addr->sa_family == AF_INET &&
1097                                     ia->ia_addr.sin_addr.s_addr ==
1098                                     ifatoia(ifa)->ia_addr.sin_addr.s_addr)
1099                                         count++;
1100                         }
1101                 }
1102         }
1103         return (count);
1104 }
1105
1106 int
1107 carp_iamatch(void *v, struct in_ifaddr *ia,
1108     struct in_addr *isaddr, u_int8_t **enaddr)
1109 {
1110         struct carp_if *cif = v;
1111         struct carp_softc *vh;
1112         int index, count = 0;
1113         struct ifaddr *ifa;
1114
1115         CARP_LOCK(cif);
1116
1117         if (carp_opts[CARPCTL_ARPBALANCE]) {
1118                 /*
1119                  * XXX proof of concept implementation.
1120                  * We use the source ip to decide which virtual host should
1121                  * handle the request. If we're master of that virtual host,
1122                  * then we respond, otherwise, just drop the arp packet on
1123                  * the floor.
1124                  */
1125                 count = carp_addrcount(cif, ia, CARP_COUNT_RUNNING);
1126                 if (count == 0) {
1127                         /* should never reach this */
1128                         CARP_UNLOCK(cif);
1129                         return (0);
1130                 }
1131
1132                 /* this should be a hash, like pf_hash() */
1133                 index = ntohl(isaddr->s_addr) % count;
1134                 count = 0;
1135
1136                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1137                         if ((SC2IFP(vh)->if_flags & IFF_UP) && (SC2IFP(vh)->if_flags & IFF_RUNNING)) {
1138                                 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1139                                     ifa_list) {
1140                                         if (ifa->ifa_addr->sa_family ==
1141                                             AF_INET &&
1142                                             ia->ia_addr.sin_addr.s_addr ==
1143                                             ifatoia(ifa)->ia_addr.sin_addr.s_addr) {
1144                                                 if (count == index) {
1145                                                         if (vh->sc_state ==
1146                                                             MASTER) {
1147                                                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1148                                                                 CARP_UNLOCK(cif);
1149                                                                 return (1);
1150                                                         } else {
1151                                                                 CARP_UNLOCK(cif);
1152                                                                 return (0);
1153                                                         }
1154                                                 }
1155                                                 count++;
1156                                         }
1157                                 }
1158                         }
1159                 }
1160         } else {
1161                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1162                         if ((SC2IFP(vh)->if_flags & IFF_UP) && (SC2IFP(vh)->if_flags & IFF_RUNNING) &&
1163                             vh->sc_state == MASTER) {
1164                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1165                                 CARP_UNLOCK(cif);
1166                                 return (1);
1167                         }
1168                 }
1169         }
1170         CARP_UNLOCK(cif);
1171         return(0);
1172 }
1173
1174 #ifdef INET6
1175 struct ifaddr *
1176 carp_iamatch6(void *v, struct in6_addr *taddr)
1177 {
1178         struct carp_if *cif = v;
1179         struct carp_softc *vh;
1180         struct ifaddr *ifa;
1181
1182         CARP_LOCK(cif);
1183         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1184                 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist, ifa_list) {
1185                         if (IN6_ARE_ADDR_EQUAL(taddr,
1186                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1187                             (SC2IFP(vh)->if_flags & IFF_UP) && (SC2IFP(vh)->if_flags & IFF_RUNNING) &&
1188                             vh->sc_state == MASTER) {
1189                                 CARP_UNLOCK(cif);
1190                                 return (ifa);
1191                         }
1192                 }
1193         }
1194         CARP_UNLOCK(cif);
1195
1196         return (NULL);
1197 }
1198
1199 void *
1200 carp_macmatch6(void *v, struct mbuf *m, const struct in6_addr *taddr)
1201 {
1202         struct m_tag *mtag;
1203         struct carp_if *cif = v;
1204         struct carp_softc *sc;
1205         struct ifaddr *ifa;
1206
1207         CARP_LOCK(cif);
1208         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) {
1209                 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1210                         if (IN6_ARE_ADDR_EQUAL(taddr,
1211                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1212                             (SC2IFP(sc)->if_flags & IFF_UP) && (SC2IFP(sc)->if_flags & IFF_RUNNING)) {
1213                                 struct ifnet *ifp = SC2IFP(sc);
1214                                 mtag = m_tag_get(PACKET_TAG_CARP,
1215                                     sizeof(struct ifnet *), M_NOWAIT);
1216                                 if (mtag == NULL) {
1217                                         /* better a bit than nothing */
1218                                         CARP_UNLOCK(cif);
1219                                         return (IF_LLADDR(sc->sc_ifp));
1220                                 }
1221                                 bcopy(&ifp, (caddr_t)(mtag + 1),
1222                                     sizeof(struct ifnet *));
1223                                 m_tag_prepend(m, mtag);
1224
1225                                 CARP_UNLOCK(cif);
1226                                 return (IF_LLADDR(sc->sc_ifp));
1227                         }
1228                 }
1229         }
1230         CARP_UNLOCK(cif);
1231
1232         return (NULL);
1233 }
1234 #endif
1235
1236 struct ifnet *
1237 carp_forus(void *v, void *dhost)
1238 {
1239         struct carp_if *cif = v;
1240         struct carp_softc *vh;
1241         u_int8_t *ena = dhost;
1242
1243         /**
1244          * XXX: See here for check on MAC adr is not for virtual use
1245          *
1246          **/
1247
1248         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1249         {
1250                 return (NULL);
1251         }
1252
1253         CARP_LOCK(cif);
1254         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list)
1255                 if ((SC2IFP(vh)->if_flags & IFF_UP) && (SC2IFP(vh)->if_flags & IFF_RUNNING) &&
1256                     vh->sc_state == MASTER &&
1257                     !bcmp(dhost, IF_LLADDR(vh->sc_ifp), ETHER_ADDR_LEN)) {
1258                         CARP_UNLOCK(cif);
1259                         return (SC2IFP(vh));
1260                 }
1261
1262         CARP_UNLOCK(cif);
1263         return (NULL);
1264 }
1265
1266 static void
1267 carp_master_down(void *v)
1268 {
1269         struct carp_softc *sc = v;
1270
1271         lwkt_serialize_enter(sc->sc_ifp->if_serializer);
1272         carp_master_down_locked(sc);
1273         lwkt_serialize_exit(sc->sc_ifp->if_serializer);
1274 }
1275
1276 static void
1277 carp_master_down_locked(struct carp_softc *sc)
1278 {
1279         if (sc->sc_carpdev)
1280                 CARP_SCLOCK_ASSERT(sc);
1281
1282         switch (sc->sc_state) {
1283         case INIT:
1284                 kprintf("%s: master_down event in INIT state\n",
1285                     SC2IFP(sc)->if_xname);
1286                 break;
1287         case MASTER:
1288                 break;
1289         case BACKUP:
1290                 carp_set_state(sc, MASTER);
1291                 carp_send_ad_locked(sc);
1292                 carp_send_arp(sc);
1293 #ifdef INET6
1294                 carp_send_na(sc);
1295 #endif /* INET6 */
1296                 carp_setrun(sc, 0);
1297                 carp_setroute(sc, RTM_ADD);
1298                 break;
1299         }
1300 }
1301
1302 /*
1303  * When in backup state, af indicates whether to reset the master down timer
1304  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1305  */
1306 static void
1307 carp_setrun(struct carp_softc *sc, sa_family_t af)
1308 {
1309         struct timeval tv;
1310
1311         if (sc->sc_carpdev == NULL) {
1312                 SC2IFP(sc)->if_flags &= ~IFF_RUNNING;
1313                 carp_set_state(sc, INIT);
1314                 return;
1315         }
1316
1317         if (SC2IFP(sc)->if_flags & IFF_UP &&
1318             sc->sc_vhid > 0 && (sc->sc_naddrs || sc->sc_naddrs6))
1319                 SC2IFP(sc)->if_flags |= IFF_RUNNING;
1320         else {
1321                 SC2IFP(sc)->if_flags &= ~IFF_RUNNING;
1322                 carp_setroute(sc, RTM_DELETE);
1323                 return;
1324         }
1325
1326         switch (sc->sc_state) {
1327         case INIT:
1328                 if (carp_opts[CARPCTL_PREEMPT] && !carp_suppress_preempt) {
1329                         carp_send_ad_locked(sc);
1330                         carp_send_arp(sc);
1331 #ifdef INET6
1332                         carp_send_na(sc);
1333 #endif /* INET6 */
1334                         CARP_DEBUG("%s: INIT -> MASTER (preempting)\n",
1335                             SC2IFP(sc)->if_xname);
1336                         carp_set_state(sc, MASTER);
1337                         carp_setroute(sc, RTM_ADD);
1338                 } else {
1339                         CARP_DEBUG("%s: INIT -> BACKUP\n", SC2IFP(sc)->if_xname);
1340                         carp_set_state(sc, BACKUP);
1341                         carp_setroute(sc, RTM_DELETE);
1342                         carp_setrun(sc, 0);
1343                 }
1344                 break;
1345         case BACKUP:
1346                 callout_stop(&sc->sc_ad_tmo);
1347                 tv.tv_sec = 3 * sc->sc_advbase;
1348                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1349                 switch (af) {
1350 #ifdef INET
1351                 case AF_INET:
1352                         callout_reset(&sc->sc_md_tmo, tvtohz_high(&tv),
1353                             carp_master_down, sc);
1354                         break;
1355 #endif /* INET */
1356 #ifdef INET6
1357                 case AF_INET6:
1358                         callout_reset(&sc->sc_md6_tmo, tvtohz_high(&tv),
1359                             carp_master_down, sc);
1360                         break;
1361 #endif /* INET6 */
1362                 default:
1363                         if (sc->sc_naddrs)
1364                                 callout_reset(&sc->sc_md_tmo, tvtohz_high(&tv),
1365                                     carp_master_down, sc);
1366                         if (sc->sc_naddrs6)
1367                                 callout_reset(&sc->sc_md6_tmo, tvtohz_high(&tv),
1368                                     carp_master_down, sc);
1369                         break;
1370                 }
1371                 break;
1372         case MASTER:
1373                 tv.tv_sec = sc->sc_advbase;
1374                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1375                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
1376                     carp_send_ad, sc);
1377                 break;
1378         }
1379 }
1380
1381 static void
1382 carp_multicast_cleanup(struct carp_softc *sc)
1383 {
1384         struct ip_moptions *imo = &sc->sc_imo;
1385         u_int16_t n = imo->imo_num_memberships;
1386
1387         /* Clean up our own multicast memberships */
1388         while (n-- > 0) {
1389                 if (imo->imo_membership[n] != NULL) {
1390                         in_delmulti(imo->imo_membership[n]);
1391                         imo->imo_membership[n] = NULL;
1392                 }
1393         }
1394         imo->imo_num_memberships = 0;
1395         imo->imo_multicast_ifp = NULL;
1396 }
1397
1398 #ifdef INET6
1399 static void
1400 carp_multicast6_cleanup(struct carp_softc *sc)
1401 {
1402         struct ip6_moptions *im6o = &sc->sc_im6o;
1403
1404         while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1405                 struct in6_multi_mship *imm =
1406                     LIST_FIRST(&im6o->im6o_memberships);
1407
1408                 LIST_REMOVE(imm, i6mm_chain);
1409                 in6_leavegroup(imm);
1410         }
1411         im6o->im6o_multicast_ifp = NULL;
1412 }
1413 #endif
1414
1415 static int
1416 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1417 {
1418         struct ifnet *ifp;
1419         struct carp_if *cif;
1420         struct in_ifaddr *ia, *ia_if;
1421         struct ip_moptions *imo = &sc->sc_imo;
1422         struct in_addr addr;
1423         u_long iaddr = htonl(sin->sin_addr.s_addr);
1424         int own, error;
1425
1426         if (sin->sin_addr.s_addr == 0)
1427         {
1428                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1429                 {
1430                         carp_set_state(sc, INIT);
1431                 }
1432                 if (sc->sc_naddrs)
1433                 {
1434                         SC2IFP(sc)->if_flags |= IFF_UP;
1435                 }
1436                 carp_setrun(sc, 0);
1437                 return (0);
1438         }
1439         /* we have to do it by hands to check we won't match on us */
1440         ia_if = NULL; own = 0;
1441         TAILQ_FOREACH(ia, &in_ifaddrhead, ia_link) {
1442                 /* and, yeah, we need a multicast-capable iface too */
1443                 if (ia->ia_ifp != SC2IFP(sc) &&
1444                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1445                     (iaddr & ia->ia_subnetmask) == ia->ia_subnet) {
1446                         if (!ia_if)
1447                                 ia_if = ia;
1448                         if (sin->sin_addr.s_addr ==
1449                             ia->ia_addr.sin_addr.s_addr)
1450                                 own++;
1451                 }
1452         }
1453
1454
1455         if (!ia_if)
1456                 return (EADDRNOTAVAIL);
1457
1458         ia = ia_if;
1459         ifp = ia->ia_ifp;
1460
1461         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1462             (imo->imo_multicast_ifp && imo->imo_multicast_ifp != ifp))
1463                 return (EADDRNOTAVAIL);
1464
1465         if (imo->imo_num_memberships == 0) {
1466                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1467                 if ((imo->imo_membership[0] = in_addmulti(&addr, ifp)) == NULL)
1468                         return (ENOBUFS);
1469                 imo->imo_num_memberships++;
1470                 imo->imo_multicast_ifp = ifp;
1471                 imo->imo_multicast_ttl = CARP_DFLTTL;
1472                 imo->imo_multicast_loop = 0;
1473         }
1474
1475         if (!ifp->if_carp) {
1476
1477                 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP,
1478                     M_WAITOK|M_ZERO);
1479                 if (!cif) {
1480                         error = ENOBUFS;
1481                         goto cleanup;
1482                 }
1483                 if ((error = ifpromisc(ifp, 1))) {
1484                         FREE(cif, M_CARP);
1485                         goto cleanup;
1486                 }
1487
1488                 CARP_LOCK_INIT(cif);
1489                 CARP_LOCK(cif);
1490                 cif->vhif_ifp = ifp;
1491                 TAILQ_INIT(&cif->vhif_vrs);
1492                 ifp->if_carp = cif;
1493
1494         } else {
1495                 struct carp_softc *vr;
1496
1497                 cif = (struct carp_if *)ifp->if_carp;
1498                 CARP_LOCK(cif);
1499                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1500                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1501                                 CARP_UNLOCK(cif);
1502                                 error = EINVAL;
1503                                 goto cleanup;
1504                         }
1505         }
1506         sc->sc_ia = ia;
1507         sc->sc_carpdev = ifp;
1508
1509         { /* XXX prevent endless loop if already in queue */
1510         struct carp_softc *vr, *after = NULL;
1511         int myself = 0;
1512         cif = (struct carp_if *)ifp->if_carp;
1513
1514         /* XXX: cif should not change, right? So we still hold the lock */
1515         CARP_LOCK_ASSERT(cif);
1516
1517         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1518                 if (vr == sc)
1519                         myself = 1;
1520                 if (vr->sc_vhid < sc->sc_vhid)
1521                         after = vr;
1522         }
1523
1524         if (!myself) {
1525                 /* We're trying to keep things in order */
1526                 if (after == NULL) {
1527                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1528                 } else {
1529                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1530                 }
1531                 cif->vhif_nvrs++;
1532         }
1533         }
1534
1535         sc->sc_naddrs++;
1536         SC2IFP(sc)->if_flags |= IFF_UP;
1537         if (own)
1538                 sc->sc_advskew = 0;
1539
1540
1541         carp_sc_state_locked(sc);
1542         carp_setrun(sc, 0);
1543
1544         CARP_UNLOCK(cif);
1545
1546         return (0);
1547
1548 cleanup:
1549         in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1550         return (error);
1551
1552 }
1553
1554 static int
1555 carp_del_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1556 {
1557         int error = 0;
1558
1559         if (!--sc->sc_naddrs) {
1560                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1561                 struct ip_moptions *imo = &sc->sc_imo;
1562
1563                 CARP_LOCK(cif);
1564                 callout_stop(&sc->sc_ad_tmo);
1565                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1566                 SC2IFP(sc)->if_flags &= ~IFF_RUNNING;
1567                 sc->sc_vhid = -1;
1568                 in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1569                 imo->imo_multicast_ifp = NULL;
1570                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1571                 if (!--cif->vhif_nvrs) {
1572                         sc->sc_carpdev->if_carp = NULL;
1573                         CARP_LOCK_DESTROY(cif);
1574                         FREE(cif, M_IFADDR);
1575                 } else {
1576                         CARP_UNLOCK(cif);
1577                 }
1578         }
1579
1580         return (error);
1581 }
1582
1583 #ifdef INET6
1584 static int
1585 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1586 {
1587         struct ifnet *ifp;
1588         struct carp_if *cif;
1589         struct in6_ifaddr *ia, *ia_if;
1590         struct ip6_moptions *im6o = &sc->sc_im6o;
1591         struct in6_multi_mship *imm;
1592         struct in6_addr in6;
1593         int own, error;
1594
1595         if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
1596                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1597                         carp_set_state(sc, INIT);
1598                 if (sc->sc_naddrs6)
1599                         SC2IFP(sc)->if_flags |= IFF_UP;
1600                 carp_setrun(sc, 0);
1601                 return (0);
1602         }
1603
1604         /* we have to do it by hands to check we won't match on us */
1605         ia_if = NULL; own = 0;
1606         for (ia = in6_ifaddr; ia; ia = ia->ia_next) {
1607                 int i;
1608
1609                 for (i = 0; i < 4; i++) {
1610                         if ((sin6->sin6_addr.s6_addr32[i] &
1611                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]) !=
1612                             (ia->ia_addr.sin6_addr.s6_addr32[i] &
1613                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]))
1614                                 break;
1615                 }
1616                 /* and, yeah, we need a multicast-capable iface too */
1617                 if (ia->ia_ifp != SC2IFP(sc) &&
1618                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1619                     (i == 4)) {
1620                         if (!ia_if)
1621                                 ia_if = ia;
1622                         if (IN6_ARE_ADDR_EQUAL(&sin6->sin6_addr,
1623                             &ia->ia_addr.sin6_addr))
1624                                 own++;
1625                 }
1626         }
1627
1628         if (!ia_if)
1629                 return (EADDRNOTAVAIL);
1630         ia = ia_if;
1631         ifp = ia->ia_ifp;
1632
1633         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1634             (im6o->im6o_multicast_ifp && im6o->im6o_multicast_ifp != ifp))
1635                 return (EADDRNOTAVAIL);
1636
1637         if (!sc->sc_naddrs6) {
1638                 im6o->im6o_multicast_ifp = ifp;
1639
1640                 /* join CARP multicast address */
1641                 bzero(&in6, sizeof(in6));
1642                 in6.s6_addr16[0] = htons(0xff02);
1643                 in6.s6_addr8[15] = 0x12;
1644                 if (in6_setscope(&in6, ifp, NULL) != 0)
1645                         goto cleanup;
1646                 if ((imm = in6_joingroup(ifp, &in6, &error)) == NULL)
1647                         goto cleanup;
1648                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1649
1650                 /* join solicited multicast address */
1651                 bzero(&in6, sizeof(in6));
1652                 in6.s6_addr16[0] = htons(0xff02);
1653                 in6.s6_addr32[1] = 0;
1654                 in6.s6_addr32[2] = htonl(1);
1655                 in6.s6_addr32[3] = sin6->sin6_addr.s6_addr32[3];
1656                 in6.s6_addr8[12] = 0xff;
1657                 if (in6_setscope(&in6, ifp, NULL) != 0)
1658                         goto cleanup;
1659                 if ((imm = in6_joingroup(ifp, &in6, &error)) == NULL)
1660                         goto cleanup;
1661                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1662         }
1663
1664         if (!ifp->if_carp) {
1665                 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP,
1666                     M_WAITOK|M_ZERO);
1667                 if (!cif) {
1668                         error = ENOBUFS;
1669                         goto cleanup;
1670                 }
1671                 if ((error = ifpromisc(ifp, 1))) {
1672                         FREE(cif, M_CARP);
1673                         goto cleanup;
1674                 }
1675
1676                 CARP_LOCK_INIT(cif);
1677                 CARP_LOCK(cif);
1678                 cif->vhif_ifp = ifp;
1679                 TAILQ_INIT(&cif->vhif_vrs);
1680                 ifp->if_carp = cif;
1681
1682         } else {
1683                 struct carp_softc *vr;
1684
1685                 cif = (struct carp_if *)ifp->if_carp;
1686                 CARP_LOCK(cif);
1687                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1688                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1689                                 CARP_UNLOCK(cif);
1690                                 error = EINVAL;
1691                                 goto cleanup;
1692                         }
1693         }
1694         sc->sc_ia6 = ia;
1695         sc->sc_carpdev = ifp;
1696
1697         { /* XXX prevent endless loop if already in queue */
1698         struct carp_softc *vr, *after = NULL;
1699         int myself = 0;
1700         cif = (struct carp_if *)ifp->if_carp;
1701         CARP_LOCK_ASSERT(cif);
1702
1703         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1704                 if (vr == sc)
1705                         myself = 1;
1706                 if (vr->sc_vhid < sc->sc_vhid)
1707                         after = vr;
1708         }
1709
1710         if (!myself) {
1711                 /* We're trying to keep things in order */
1712                 if (after == NULL) {
1713                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1714                 } else {
1715                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1716                 }
1717                 cif->vhif_nvrs++;
1718         }
1719         }
1720
1721         sc->sc_naddrs6++;
1722         SC2IFP(sc)->if_flags |= IFF_UP;
1723         if (own)
1724                 sc->sc_advskew = 0;
1725         carp_sc_state_locked(sc);
1726         carp_setrun(sc, 0);
1727
1728         CARP_UNLOCK(cif);
1729
1730         return (0);
1731
1732 cleanup:
1733         /* clean up multicast memberships */
1734         if (!sc->sc_naddrs6) {
1735                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1736                         imm = LIST_FIRST(&im6o->im6o_memberships);
1737                         LIST_REMOVE(imm, i6mm_chain);
1738                         in6_leavegroup(imm);
1739                 }
1740         }
1741         return (error);
1742 }
1743
1744 static int
1745 carp_del_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1746 {
1747         int error = 0;
1748
1749         if (!--sc->sc_naddrs6) {
1750                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1751                 struct ip6_moptions *im6o = &sc->sc_im6o;
1752
1753                 CARP_LOCK(cif);
1754                 callout_stop(&sc->sc_ad_tmo);
1755                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1756                 SC2IFP(sc)->if_flags &= ~IFF_RUNNING;
1757                 sc->sc_vhid = -1;
1758                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1759                         struct in6_multi_mship *imm =
1760                             LIST_FIRST(&im6o->im6o_memberships);
1761
1762                         LIST_REMOVE(imm, i6mm_chain);
1763                         in6_leavegroup(imm);
1764                 }
1765                 im6o->im6o_multicast_ifp = NULL;
1766                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1767                 if (!--cif->vhif_nvrs) {
1768                         CARP_LOCK_DESTROY(cif);
1769                         sc->sc_carpdev->if_carp = NULL;
1770                         FREE(cif, M_IFADDR);
1771                 } else
1772                         CARP_UNLOCK(cif);
1773         }
1774
1775         return (error);
1776 }
1777 #endif /* INET6 */
1778
1779 static int
1780 carp_ioctl(struct ifnet *ifp, u_long cmd, caddr_t addr, struct ucred *creds)
1781 {
1782         struct carp_softc *sc = ifp->if_softc, *vr;
1783         struct carpreq carpr;
1784         struct ifaddr *ifa;
1785         struct ifreq *ifr;
1786         struct ifaliasreq *ifra;
1787         int locked = 0, error = 0;
1788
1789         ifa = (struct ifaddr *)addr;
1790         ifra = (struct ifaliasreq *)addr;
1791         ifr = (struct ifreq *)addr;
1792
1793
1794         switch (cmd) {
1795         case SIOCSIFADDR:
1796                 switch (ifa->ifa_addr->sa_family) {
1797 #ifdef INET
1798                 case AF_INET:
1799                         SC2IFP(sc)->if_flags |= IFF_UP;
1800                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1801                             sizeof(struct sockaddr));
1802                         error = carp_set_addr(sc, satosin(ifa->ifa_addr));
1803                         break;
1804 #endif /* INET */
1805 #ifdef INET6
1806                 case AF_INET6:
1807                         SC2IFP(sc)->if_flags |= IFF_UP;
1808                         error = carp_set_addr6(sc, satosin6(ifa->ifa_addr));
1809                         break;
1810 #endif /* INET6 */
1811                 default:
1812                         error = EAFNOSUPPORT;
1813                         break;
1814                 }
1815                 break;
1816
1817         case SIOCAIFADDR:
1818                 switch (ifa->ifa_addr->sa_family) {
1819 #ifdef INET
1820                 case AF_INET:
1821                         SC2IFP(sc)->if_flags |= IFF_UP;
1822                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1823                             sizeof(struct sockaddr));
1824                         error = carp_set_addr(sc, satosin(&ifra->ifra_addr));
1825                         break;
1826 #endif /* INET */
1827 #ifdef INET6
1828                 case AF_INET6:
1829                         SC2IFP(sc)->if_flags |= IFF_UP;
1830                         error = carp_set_addr6(sc, satosin6(&ifra->ifra_addr));
1831                         break;
1832 #endif /* INET6 */
1833                 default:
1834                         error = EAFNOSUPPORT;
1835                         break;
1836                 }
1837                 break;
1838
1839         case SIOCDIFADDR:
1840                 switch (ifa->ifa_addr->sa_family) {
1841 #ifdef INET
1842                 case AF_INET:
1843                         error = carp_del_addr(sc, satosin(&ifra->ifra_addr));
1844                         break;
1845 #endif /* INET */
1846 #ifdef INET6
1847                 case AF_INET6:
1848                         error = carp_del_addr6(sc, satosin6(&ifra->ifra_addr));
1849                         break;
1850 #endif /* INET6 */
1851                 default:
1852                         error = EAFNOSUPPORT;
1853                         break;
1854                 }
1855                 break;
1856
1857         case SIOCSIFFLAGS:
1858                 if (sc->sc_carpdev) {
1859                         locked = 1;
1860                         CARP_SCLOCK(sc);
1861                 }
1862                 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) {
1863                         callout_stop(&sc->sc_ad_tmo);
1864                         callout_stop(&sc->sc_md_tmo);
1865                         callout_stop(&sc->sc_md6_tmo);
1866                         if (sc->sc_state == MASTER)
1867                                 carp_send_ad_locked(sc);
1868                         carp_set_state(sc, INIT);
1869                         carp_setrun(sc, 0);
1870                 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) {
1871                         SC2IFP(sc)->if_flags |= IFF_UP;
1872                         carp_setrun(sc, 0);
1873                 }
1874                 break;
1875
1876         case SIOCSVH:
1877                 error = suser(curthread);
1878                 if (error)
1879                         break;
1880                 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1881                         break;
1882                 error = 1;
1883                 if (sc->sc_carpdev) {
1884                         locked = 1;
1885                         CARP_SCLOCK(sc);
1886                 }
1887                 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) {
1888                         switch (carpr.carpr_state) {
1889                         case BACKUP:
1890                                 callout_stop(&sc->sc_ad_tmo);
1891                                 carp_set_state(sc, BACKUP);
1892                                 carp_setrun(sc, 0);
1893                                 carp_setroute(sc, RTM_DELETE);
1894                                 break;
1895                         case MASTER:
1896                                 carp_master_down_locked(sc);
1897                                 break;
1898                         default:
1899                                 break;
1900                         }
1901                 }
1902                 if (carpr.carpr_vhid > 0) {
1903                         if (carpr.carpr_vhid > 255) {
1904                                 error = EINVAL;
1905                                 break;
1906                         }
1907                         if (sc->sc_carpdev) {
1908                                 struct carp_if *cif;
1909                                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1910                                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1911                                         if (vr != sc &&
1912                                             vr->sc_vhid == carpr.carpr_vhid)
1913                                                 return EEXIST;
1914                         }
1915                         sc->sc_vhid = carpr.carpr_vhid;
1916                         IF_LLADDR(sc->sc_ifp)[0] = 0;
1917                         IF_LLADDR(sc->sc_ifp)[1] = 0;
1918                         IF_LLADDR(sc->sc_ifp)[2] = 0x5e;
1919                         IF_LLADDR(sc->sc_ifp)[3] = 0;
1920                         IF_LLADDR(sc->sc_ifp)[4] = 1;
1921                         IF_LLADDR(sc->sc_ifp)[5] = sc->sc_vhid;
1922                         error--;
1923                 }
1924                 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) {
1925                         if (carpr.carpr_advskew >= 255) {
1926                                 error = EINVAL;
1927                                 break;
1928                         }
1929                         if (carpr.carpr_advbase > 255) {
1930                                 error = EINVAL;
1931                                 break;
1932                         }
1933                         sc->sc_advbase = carpr.carpr_advbase;
1934                         sc->sc_advskew = carpr.carpr_advskew;
1935                         error--;
1936                 }
1937                 bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1938                 if (error > 0)
1939                         error = EINVAL;
1940                 else {
1941                         error = 0;
1942                         carp_setrun(sc, 0);
1943                 }
1944                 break;
1945
1946         case SIOCGVH:
1947                 /* XXX: lockless read */
1948                 bzero(&carpr, sizeof(carpr));
1949                 carpr.carpr_state = sc->sc_state;
1950                 carpr.carpr_vhid = sc->sc_vhid;
1951                 carpr.carpr_advbase = sc->sc_advbase;
1952                 carpr.carpr_advskew = sc->sc_advskew;
1953                 error = suser(curthread);
1954                 if (error == 0)
1955                         bcopy(sc->sc_key, carpr.carpr_key,
1956                             sizeof(carpr.carpr_key));
1957                 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1958                 break;
1959
1960         default:
1961                 error = EINVAL;
1962         }
1963
1964         if (locked)
1965                 CARP_SCUNLOCK(sc);
1966
1967         carp_hmac_prepare(sc);
1968
1969         return (error);
1970 }
1971
1972 /*
1973  * XXX: this is looutput. We should eventually use it from there.
1974  */
1975 static int
1976 carp_looutput(struct ifnet *ifp, struct mbuf *m, struct sockaddr *dst,
1977     struct rtentry *rt)
1978 {
1979         u_int32_t af;
1980
1981         M_ASSERTPKTHDR(m); /* check if we have the packet header */
1982
1983         if (rt && rt->rt_flags & (RTF_REJECT|RTF_BLACKHOLE)) {
1984                 m_freem(m);
1985                 return (rt->rt_flags & RTF_BLACKHOLE ? 0 :
1986                         rt->rt_flags & RTF_HOST ? EHOSTUNREACH : ENETUNREACH);
1987         }
1988
1989         ifp->if_opackets++;
1990         ifp->if_obytes += m->m_pkthdr.len;
1991
1992         /* BPF writes need to be handled specially. */
1993         if (dst->sa_family == AF_UNSPEC) {
1994                 bcopy(dst->sa_data, &af, sizeof(af));
1995                 dst->sa_family = af;
1996         }
1997
1998 #if 1   /* XXX */
1999         switch (dst->sa_family) {
2000         case AF_INET:
2001         case AF_INET6:
2002         case AF_IPX:
2003         case AF_APPLETALK:
2004                 break;
2005         default:
2006                 m_freem(m);
2007                 return (EAFNOSUPPORT);
2008         }
2009 #endif
2010         return(if_simloop(ifp, m, dst->sa_family, 0));
2011 }
2012
2013 /*
2014  * Start output on carp interface. This function should never be called.
2015  */
2016 static void
2017 carp_start(struct ifnet *ifp)
2018 {
2019 #ifdef DEBUG
2020         kprintf("%s: start called\n", ifp->if_xname);
2021 #endif
2022 }
2023
2024 int
2025 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *sa,
2026     struct rtentry *rt)
2027 {
2028         struct m_tag *mtag;
2029         struct carp_softc *sc;
2030         struct ifnet *carp_ifp;
2031
2032         if (!sa)
2033                 return (0);
2034
2035         switch (sa->sa_family) {
2036 #ifdef INET
2037         case AF_INET:
2038                 break;
2039 #endif /* INET */
2040 #ifdef INET6
2041         case AF_INET6:
2042                 break;
2043 #endif /* INET6 */
2044         default:
2045                 return (0);
2046         }
2047
2048         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
2049         if (mtag == NULL)
2050                 return (0);
2051
2052         bcopy(mtag + 1, &carp_ifp, sizeof(struct ifnet *));
2053         sc = carp_ifp->if_softc;
2054
2055         /* Set the source MAC address to Virtual Router MAC Address */
2056         switch (ifp->if_type) {
2057         case IFT_ETHER:
2058         case IFT_L2VLAN: {
2059                         struct ether_header *eh;
2060
2061                         eh = mtod(m, struct ether_header *);
2062                         eh->ether_shost[0] = 0;
2063                         eh->ether_shost[1] = 0;
2064                         eh->ether_shost[2] = 0x5e;
2065                         eh->ether_shost[3] = 0;
2066                         eh->ether_shost[4] = 1;
2067                         eh->ether_shost[5] = sc->sc_vhid;
2068                 }
2069                 break;
2070         default:
2071                 kprintf("%s: carp is not supported for this interface type\n",
2072                     ifp->if_xname);
2073                 return (EOPNOTSUPP);
2074         }
2075
2076         return (0);
2077
2078 }
2079
2080 static void
2081 carp_set_state(struct carp_softc *sc, int state)
2082 {
2083
2084         if (sc->sc_carpdev)
2085                 CARP_SCLOCK_ASSERT(sc);
2086
2087         if (sc->sc_state == state)
2088                 return;
2089
2090         sc->sc_state = state;
2091         switch (state) {
2092         case BACKUP:
2093                 SC2IFP(sc)->if_link_state = LINK_STATE_DOWN;
2094                 break;
2095         case MASTER:
2096                 SC2IFP(sc)->if_link_state = LINK_STATE_UP;
2097                 break;
2098         default:
2099                 SC2IFP(sc)->if_link_state = LINK_STATE_UNKNOWN;
2100                 break;
2101         }
2102         rt_ifmsg(SC2IFP(sc));
2103 }
2104
2105 void
2106 carp_carpdev_state(void *v)
2107 {
2108         struct carp_if *cif = v;
2109
2110         CARP_LOCK(cif);
2111         carp_carpdev_state_locked(cif);
2112         CARP_UNLOCK(cif);
2113 }
2114
2115 static void
2116 carp_carpdev_state_locked(struct carp_if *cif)
2117 {
2118         struct carp_softc *sc;
2119
2120         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list)
2121                 carp_sc_state_locked(sc);
2122 }
2123
2124 static void
2125 carp_sc_state_locked(struct carp_softc *sc)
2126 {
2127         CARP_SCLOCK_ASSERT(sc);
2128
2129         if ( !(sc->sc_carpdev->if_flags & IFF_UP)) {
2130                 sc->sc_flags_backup = SC2IFP(sc)->if_flags;
2131                 SC2IFP(sc)->if_flags &= ~IFF_UP;
2132                 SC2IFP(sc)->if_flags &= ~IFF_RUNNING;
2133                 callout_stop(&sc->sc_ad_tmo);
2134                 callout_stop(&sc->sc_md_tmo);
2135                 callout_stop(&sc->sc_md6_tmo);
2136                 carp_set_state(sc, INIT);
2137                 carp_setrun(sc, 0);
2138                 if (!sc->sc_suppress) {
2139                         carp_suppress_preempt++;
2140                         if (carp_suppress_preempt == 1) {
2141                                 CARP_SCUNLOCK(sc);
2142                                 carp_send_ad_all();
2143                                 CARP_SCLOCK(sc);
2144                         }
2145                 }
2146                 sc->sc_suppress = 1;
2147         } else {
2148                 SC2IFP(sc)->if_flags |= sc->sc_flags_backup;
2149                 carp_set_state(sc, INIT);
2150                 carp_setrun(sc, 0);
2151                 if (sc->sc_suppress)
2152                         carp_suppress_preempt--;
2153                 sc->sc_suppress = 0;
2154         }
2155
2156         return;
2157 }
2158
2159 static int
2160 carp_modevent(module_t mod, int type, void *data)
2161 {
2162         switch (type) {
2163         case MOD_LOAD:
2164                 if_detach_event_tag = EVENTHANDLER_REGISTER(ifnet_departure_event,
2165                     carp_ifdetach, NULL, EVENTHANDLER_PRI_ANY);
2166                 if (if_detach_event_tag == NULL)
2167                         return (ENOMEM);
2168
2169                 LIST_INIT(&carpif_list);
2170                 if_clone_attach(&carp_cloner);
2171                 break;
2172
2173         case MOD_UNLOAD:
2174                 EVENTHANDLER_DEREGISTER(ifnet_departure_event, if_detach_event_tag);
2175                 if_clone_detach(&carp_cloner);
2176                 break;
2177
2178         default:
2179                 return (EINVAL);
2180         }
2181
2182         return (0);
2183 }
2184
2185 static moduledata_t carp_mod = {
2186         "carp",
2187         carp_modevent,
2188         0
2189 };
2190
2191 DECLARE_MODULE(carp, carp_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);