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 1.5 2008/01/06 16:55:52 swildner 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         sc->sc_flags_backup = 0;
 364         sc->sc_suppress = 0;
 365         sc->sc_advbase = CARP_DFLTINTV;
 366         sc->sc_vhid = -1;       /* required setting */
 367         sc->sc_advskew = 0;
 368         sc->sc_init_counter = 1;
 369         sc->sc_naddrs = sc->sc_naddrs6 = 0; /* M_ZERO? */
 370
 371 #ifdef INET6
 372         sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL;
 373 #endif
 374
 375 /*      sc->sc_imo.imo_membership = kmalloc((sizeof(struct in_multi) * IP_MAX_MEMBERSHIPS), M_CARP,M_WAITOK);*/
 376 /*
 377         sc->sc_imo.imo_max_memberships = IP_MAX_MEMBERSHIPS;
 378         sc->sc_imo.imo_multicast_vif = -1;
 379 */
 380         callout_init(&sc->sc_ad_tmo);
 381         callout_init(&sc->sc_md_tmo);
 382         callout_init(&sc->sc_md6_tmo);
 383
 384         ifp->if_softc = sc;
 385         if_initname(ifp, CARP_IFNAME, unit);
 386         ifp->if_mtu = ETHERMTU;
 387         ifp->if_flags = IFF_LOOPBACK;
 388         ifp->if_ioctl = carp_ioctl;
 389         ifp->if_output = carp_looutput;
 390         ifp->if_start = carp_start;
 391         ifp->if_type = IFT_CARP;
 392         ifp->if_snd.ifq_maxlen = ifqmaxlen;
 393         ifp->if_hdrlen = 0;
 394         if_attach(ifp, NULL);
 395         bpfattach(ifp, DLT_NULL, sizeof(u_int));
 396
 397         crit_enter();
 398         LIST_INSERT_HEAD(&carpif_list, sc, sc_next);
 399         crit_exit();
 400
 401         return (0);
 402 }
 403
 404 static void
 405 carp_clone_destroy(struct ifnet *ifp)
 406 {
 407         struct carp_softc *sc = ifp->if_softc;
 408
 409         if (sc->sc_carpdev)
 410                 CARP_SCLOCK(sc);
 411         carpdetach(sc, 1);      /* Returns unlocked. */
 412
 413         crit_enter();
 414         LIST_REMOVE(sc, sc_next);
 415         crit_exit();
 416         bpfdetach(ifp);
 417         if_detach(ifp);
 418 /*      if_free_type(ifp, IFT_ETHER);*/
 419 /*      kfree(sc->sc_imo.imo_membership, M_CARP); */
 420         kfree(sc, M_CARP);
 421 }
 422
 423 /*
 424  * This function can be called on CARP interface destroy path,
 425  * and in case of the removal of the underlying interface as
 426  * well. We differentiate these two cases. In the latter case
 427  * we do not cleanup our multicast memberships, since they
 428  * are already freed. Also, in the latter case we do not
 429  * release the lock on return, because the function will be
 430  * called once more, for another CARP instance on the same
 431  * interface.
 432  */
 433 static void
 434 carpdetach(struct carp_softc *sc, int unlock)
 435 {
 436         struct carp_if *cif;
 437
 438         callout_stop(&sc->sc_ad_tmo);
 439         callout_stop(&sc->sc_md_tmo);
 440         callout_stop(&sc->sc_md6_tmo);
 441
 442         if (sc->sc_suppress)
 443                 carp_suppress_preempt--;
 444         sc->sc_suppress = 0;
 445
 446         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS)
 447                 carp_suppress_preempt--;
 448         sc->sc_sendad_errors = 0;
 449
 450         carp_set_state(sc, INIT);
 451         SC2IFP(sc)->if_flags &= ~IFF_UP;
 452         carp_setrun(sc, 0);
 453         if (unlock)
 454                 carp_multicast_cleanup(sc);
 455 #ifdef INET6
 456         carp_multicast6_cleanup(sc);
 457 #endif
 458
 459         if (sc->sc_carpdev != NULL) {
 460                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
 461                 CARP_LOCK_ASSERT(cif);
 462                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
 463                 if (!--cif->vhif_nvrs) {
 464                         ifpromisc(sc->sc_carpdev, 0);
 465                         sc->sc_carpdev->if_carp = NULL;
 466                         CARP_LOCK_DESTROY(cif);
 467                         FREE(cif, M_IFADDR);
 468                 } else if (unlock)
 469                         CARP_UNLOCK(cif);
 470                 sc->sc_carpdev = NULL;
 471         }
 472 }
 473
 474 /* Detach an interface from the carp. */
 475 static void
 476 carp_ifdetach(void *arg __unused, struct ifnet *ifp)
 477 {
 478         struct carp_if *cif = (struct carp_if *)ifp->if_carp;
 479         struct carp_softc *sc, *nextsc;
 480
 481         if (cif == NULL)
 482                 return;
 483
 484         /*
 485          * XXX: At the end of for() cycle the lock will be destroyed.
 486          */
 487         CARP_LOCK(cif);
 488         for (sc = TAILQ_FIRST(&cif->vhif_vrs); sc; sc = nextsc) {
 489                 nextsc = TAILQ_NEXT(sc, sc_list);
 490                 carpdetach(sc, 0);
 491         }
 492         CARP_UNLOCK(cif);
 493 }
 494
 495 /*
 496  * process input packet.
 497  * we have rearranged checks order compared to the rfc,
 498  * but it seems more efficient this way or not possible otherwise.
 499  */
 500 void
 501 carp_input(struct mbuf *m, int hlen)
 502 {
 503         struct ip *ip = mtod(m, struct ip *);
 504         struct carp_header *ch;
 505         int iplen, len;
 506
 507         carpstats.carps_ipackets++;
 508
 509         if (!carp_opts[CARPCTL_ALLOW]) {
 510                 m_freem(m);
 511                 return;
 512         }
 513
 514         /* check if received on a valid carp interface */
 515         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 516                 carpstats.carps_badif++;
 517                 CARP_LOG("carp_input: packet received on non-carp "
 518                     "interface: %s\n",
 519                     m->m_pkthdr.rcvif->if_xname);
 520                 m_freem(m);
 521                 return;
 522         }
 523
 524         /* verify that the IP TTL is 255.  */
 525         if (ip->ip_ttl != CARP_DFLTTL) {
 526                 carpstats.carps_badttl++;
 527                 CARP_LOG("carp_input: received ttl %d != 255i on %s\n",
 528                     ip->ip_ttl,
 529                     m->m_pkthdr.rcvif->if_xname);
 530                 m_freem(m);
 531                 return;
 532         }
 533
 534         iplen = ip->ip_hl << 2;
 535
 536         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 537                 carpstats.carps_badlen++;
 538                 CARP_LOG("carp_input: received len %zd < "
 539                     "sizeof(struct carp_header)\n",
 540                     m->m_len - sizeof(struct ip));
 541                 m_freem(m);
 542                 return;
 543         }
 544
 545         if (iplen + sizeof(*ch) < m->m_len) {
 546                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 547                         carpstats.carps_hdrops++;
 548                         CARP_LOG("carp_input: pullup failed\n");
 549                         return;
 550                 }
 551                 ip = mtod(m, struct ip *);
 552         }
 553         ch = (struct carp_header *)((char *)ip + iplen);
 554
 555         /*
 556          * verify that the received packet length is
 557          * equal to the CARP header
 558          */
 559         len = iplen + sizeof(*ch);
 560         if (len > m->m_pkthdr.len) {
 561                 carpstats.carps_badlen++;
 562                 CARP_LOG("carp_input: packet too short %d on %s\n",
 563                     m->m_pkthdr.len,
 564                     m->m_pkthdr.rcvif->if_xname);
 565                 m_freem(m);
 566                 return;
 567         }
 568
 569         if ((m = m_pullup(m, len)) == NULL) {
 570                 carpstats.carps_hdrops++;
 571                 return;
 572         }
 573         ip = mtod(m, struct ip *);
 574         ch = (struct carp_header *)((char *)ip + iplen);
 575
 576         /* verify the CARP checksum */
 577         m->m_data += iplen;
 578         if (carp_cksum(m, len - iplen)) {
 579                 carpstats.carps_badsum++;
 580                 CARP_LOG("carp_input: checksum failed on %s\n",
 581                     m->m_pkthdr.rcvif->if_xname);
 582                 m_freem(m);
 583                 return;
 584         }
 585         m->m_data -= iplen;
 586
 587         carp_input_c(m, ch, AF_INET);
 588 }
 589
 590 #ifdef INET6
 591 int
 592 carp6_input(struct mbuf **mp, int *offp, int proto)
 593 {
 594         struct mbuf *m = *mp;
 595         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 596         struct carp_header *ch;
 597         u_int len;
 598
 599         carpstats.carps_ipackets6++;
 600
 601         if (!carp_opts[CARPCTL_ALLOW]) {
 602                 m_freem(m);
 603                 return (IPPROTO_DONE);
 604         }
 605
 606         /* check if received on a valid carp interface */
 607         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 608                 carpstats.carps_badif++;
 609                 CARP_LOG("carp6_input: packet received on non-carp "
 610                     "interface: %s\n",
 611                     m->m_pkthdr.rcvif->if_xname);
 612                 m_freem(m);
 613                 return (IPPROTO_DONE);
 614         }
 615
 616         /* verify that the IP TTL is 255 */
 617         if (ip6->ip6_hlim != CARP_DFLTTL) {
 618                 carpstats.carps_badttl++;
 619                 CARP_LOG("carp6_input: received ttl %d != 255 on %s\n",
 620                     ip6->ip6_hlim,
 621                     m->m_pkthdr.rcvif->if_xname);
 622                 m_freem(m);
 623                 return (IPPROTO_DONE);
 624         }
 625
 626         /* verify that we have a complete carp packet */
 627         len = m->m_len;
 628         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 629         if (ch == NULL) {
 630                 carpstats.carps_badlen++;
 631                 CARP_LOG("carp6_input: packet size %u too small\n", len);
 632                 return (IPPROTO_DONE);
 633         }
 634
 635
 636         /* verify the CARP checksum */
 637         m->m_data += *offp;
 638         if (carp_cksum(m, sizeof(*ch))) {
 639                 carpstats.carps_badsum++;
 640                 CARP_LOG("carp6_input: checksum failed, on %s\n",
 641                     m->m_pkthdr.rcvif->if_xname);
 642                 m_freem(m);
 643                 return (IPPROTO_DONE);
 644         }
 645         m->m_data -= *offp;
 646
 647         carp_input_c(m, ch, AF_INET6);
 648         return (IPPROTO_DONE);
 649 }
 650 #endif /* INET6 */
 651
 652 static void
 653 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 654 {
 655         struct ifnet *ifp = m->m_pkthdr.rcvif;
 656         struct carp_softc *sc;
 657         u_int64_t tmp_counter;
 658         struct timeval sc_tv, ch_tv;
 659
 660         /* verify that the VHID is valid on the receiving interface */
 661         CARP_LOCK(ifp->if_carp);
 662         TAILQ_FOREACH(sc, &((struct carp_if *)ifp->if_carp)->vhif_vrs, sc_list)
 663                 if (sc->sc_vhid == ch->carp_vhid)
 664                         break;
 665
 666         if (!sc || !((SC2IFP(sc)->if_flags & IFF_UP) && (SC2IFP(sc)->if_flags & IFF_RUNNING))) {
 667                 carpstats.carps_badvhid++;
 668                 CARP_UNLOCK(ifp->if_carp);
 669                 m_freem(m);
 670                 return;
 671         }
 672
 673         getmicrotime(&SC2IFP(sc)->if_lastchange);
 674         SC2IFP(sc)->if_ipackets++;
 675         SC2IFP(sc)->if_ibytes += m->m_pkthdr.len;
 676
 677         if (SC2IFP(sc)->if_bpf) {
 678                 struct ip *ip = mtod(m, struct ip *);
 679
 680                 /* BPF wants net byte order */
 681                 ip->ip_len = htons(ip->ip_len + (ip->ip_hl << 2));
 682                 ip->ip_off = htons(ip->ip_off);
 683                 bpf_mtap(SC2IFP(sc)->if_bpf, m);
 684         }
 685
 686         /* verify the CARP version. */
 687         if (ch->carp_version != CARP_VERSION) {
 688                 carpstats.carps_badver++;
 689                 SC2IFP(sc)->if_ierrors++;
 690                 CARP_UNLOCK(ifp->if_carp);
 691                 CARP_LOG("%s; invalid version %d\n",
 692                     SC2IFP(sc)->if_xname,
 693                     ch->carp_version);
 694                 m_freem(m);
 695                 return;
 696         }
 697
 698         /* verify the hash */
 699         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 700                 carpstats.carps_badauth++;
 701                 SC2IFP(sc)->if_ierrors++;
 702                 CARP_UNLOCK(ifp->if_carp);
 703                 CARP_LOG("%s: incorrect hash\n", SC2IFP(sc)->if_xname);
 704                 m_freem(m);
 705                 return;
 706         }
 707
 708         tmp_counter = ntohl(ch->carp_counter[0]);
 709         tmp_counter = tmp_counter<<32;
 710         tmp_counter += ntohl(ch->carp_counter[1]);
 711
 712         /* XXX Replay protection goes here */
 713
 714         sc->sc_init_counter = 0;
 715         sc->sc_counter = tmp_counter;
 716
 717         sc_tv.tv_sec = sc->sc_advbase;
 718         if (carp_suppress_preempt && sc->sc_advskew <  240)
 719                 sc_tv.tv_usec = 240 * 1000000 / 256;
 720         else
 721                 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256;
 722         ch_tv.tv_sec = ch->carp_advbase;
 723         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 724
 725         switch (sc->sc_state) {
 726         case INIT:
 727                 break;
 728         case MASTER:
 729                 /*
 730                  * If we receive an advertisement from a master who's going to
 731                  * be more frequent than us, go into BACKUP state.
 732                  */
 733                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 734                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 735                         callout_stop(&sc->sc_ad_tmo);
 736                         CARP_DEBUG("%s: MASTER -> BACKUP "
 737                            "(more frequent advertisement received)\n",
 738                            SC2IFP(sc)->if_xname);
 739                         carp_set_state(sc, BACKUP);
 740                         carp_setrun(sc, 0);
 741                         carp_setroute(sc, RTM_DELETE);
 742                 }
 743                 break;
 744         case BACKUP:
 745                 /*
 746                  * If we're pre-empting masters who advertise slower than us,
 747                  * and this one claims to be slower, treat him as down.
 748                  */
 749                 if (carp_opts[CARPCTL_PREEMPT] &&
 750                     timevalcmp(&sc_tv, &ch_tv, <)) {
 751                         CARP_DEBUG("%s: BACKUP -> MASTER "
 752                             "(preempting a slower master)\n",
 753                             SC2IFP(sc)->if_xname);
 754                         carp_master_down_locked(sc);
 755                         break;
 756                 }
 757
 758                 /*
 759                  *  If the master is going to advertise at such a low frequency
 760                  *  that he's guaranteed to time out, we'd might as well just
 761                  *  treat him as timed out now.
 762                  */
 763                 sc_tv.tv_sec = sc->sc_advbase * 3;
 764                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 765                         CARP_DEBUG("%s: BACKUP -> MASTER "
 766                             "(master timed out)\n",
 767                             SC2IFP(sc)->if_xname);
 768                         carp_master_down_locked(sc);
 769                         break;
 770                 }
 771
 772                 /*
 773                  * Otherwise, we reset the counter and wait for the next
 774                  * advertisement.
 775                  */
 776                 carp_setrun(sc, af);
 777                 break;
 778         }
 779
 780         CARP_UNLOCK(ifp->if_carp);
 781
 782         m_freem(m);
 783         return;
 784 }
 785
 786 static int
 787 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 788 {
 789         struct m_tag *mtag;
 790         struct ifnet *ifp = SC2IFP(sc);
 791
 792         if (sc->sc_init_counter) {
 793                 /* this could also be seconds since unix epoch */
 794                 sc->sc_counter = karc4random();
 795                 sc->sc_counter = sc->sc_counter << 32;
 796                 sc->sc_counter += karc4random();
 797         } else
 798                 sc->sc_counter++;
 799
 800         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 801         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 802
 803         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 804
 805         /* Tag packet for carp_output */
 806         mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct ifnet *), MB_DONTWAIT);
 807         if (mtag == NULL) {
 808                 m_freem(m);
 809                 SC2IFP(sc)->if_oerrors++;
 810                 return (ENOMEM);
 811         }
 812         bcopy(&ifp, (caddr_t)(mtag + 1), sizeof(struct ifnet *));
 813         m_tag_prepend(m, mtag);
 814
 815         return (0);
 816 }
 817
 818 static void
 819 carp_send_ad_all(void)
 820 {
 821         struct carp_softc *sc;
 822
 823         LIST_FOREACH(sc, &carpif_list, sc_next) {
 824                 if (sc->sc_carpdev == NULL)
 825                         continue;
 826                 CARP_SCLOCK(sc);
 827                 if ((SC2IFP(sc)->if_flags & IFF_UP) && (SC2IFP(sc)->if_flags & IFF_RUNNING) &&
 828                      sc->sc_state == MASTER)
 829                         carp_send_ad_locked(sc);
 830                 CARP_SCUNLOCK(sc);
 831         }
 832 }
 833
 834 static void
 835 carp_send_ad(void *v)
 836 {
 837         struct carp_softc *sc = v;
 838
 839         CARP_SCLOCK(sc);
 840         carp_send_ad_locked(sc);
 841         CARP_SCUNLOCK(sc);
 842 }
 843
 844 static void
 845 carp_send_ad_locked(struct carp_softc *sc)
 846 {
 847         struct carp_header ch;
 848         struct timeval tv;
 849         struct carp_header *ch_ptr;
 850         struct mbuf *m;
 851         int len, advbase, advskew;
 852
 853
 854         /* bow out if we've lost our UPness or RUNNINGuiness */
 855         if (!((SC2IFP(sc)->if_flags & IFF_UP) && (SC2IFP(sc)->if_flags & IFF_RUNNING))) {
 856                 advbase = 255;
 857                 advskew = 255;
 858         } else {
 859                 advbase = sc->sc_advbase;
 860                 if (!carp_suppress_preempt || sc->sc_advskew > 240)
 861                         advskew = sc->sc_advskew;
 862                 else
 863                         advskew = 240;
 864                 tv.tv_sec = advbase;
 865                 tv.tv_usec = advskew * 1000000 / 256;
 866         }
 867
 868         ch.carp_version = CARP_VERSION;
 869         ch.carp_type = CARP_ADVERTISEMENT;
 870         ch.carp_vhid = sc->sc_vhid;
 871         ch.carp_advbase = advbase;
 872         ch.carp_advskew = advskew;
 873         ch.carp_authlen = 7;    /* XXX DEFINE */
 874         ch.carp_pad1 = 0;       /* must be zero */
 875         ch.carp_cksum = 0;
 876
 877 #ifdef INET
 878         if (sc->sc_ia) {
 879                 struct ip *ip;
 880
 881                 MGETHDR(m, M_NOWAIT, MT_HEADER);
 882                 if (m == NULL) {
 883                         SC2IFP(sc)->if_oerrors++;
 884                         carpstats.carps_onomem++;
 885                         /* XXX maybe less ? */
 886                         if (advbase != 255 || advskew != 255)
 887                                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
 888                                     carp_send_ad, sc);
 889                         return;
 890                 }
 891                 len = sizeof(*ip) + sizeof(ch);
 892                 m->m_pkthdr.len = len;
 893                 m->m_pkthdr.rcvif = NULL;
 894                 m->m_len = len;
 895                 MH_ALIGN(m, m->m_len);
 896                 m->m_flags |= M_MCAST;
 897                 ip = mtod(m, struct ip *);
 898                 ip->ip_v = IPVERSION;
 899                 ip->ip_hl = sizeof(*ip) >> 2;
 900                 ip->ip_tos = IPTOS_LOWDELAY;
 901                 ip->ip_len = len;
 902                 ip->ip_id = ip_newid();
 903                 ip->ip_off = IP_DF;
 904                 ip->ip_ttl = CARP_DFLTTL;
 905                 ip->ip_p = IPPROTO_CARP;
 906                 ip->ip_sum = 0;
 907                 ip->ip_src.s_addr = sc->sc_ia->ia_addr.sin_addr.s_addr;
 908                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 909
 910                 ch_ptr = (struct carp_header *)(&ip[1]);
 911                 bcopy(&ch, ch_ptr, sizeof(ch));
 912                 if (carp_prepare_ad(m, sc, ch_ptr))
 913                         return;
 914
 915                 m->m_data += sizeof(*ip);
 916                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip));
 917                 m->m_data -= sizeof(*ip);
 918
 919                 getmicrotime(&SC2IFP(sc)->if_lastchange);
 920                 SC2IFP(sc)->if_opackets++;
 921                 SC2IFP(sc)->if_obytes += len;
 922                 carpstats.carps_opackets++;
 923
 924                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, NULL)) {
 925                         SC2IFP(sc)->if_oerrors++;
 926                         if (sc->sc_sendad_errors < INT_MAX)
 927                                 sc->sc_sendad_errors++;
 928                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 929                                 carp_suppress_preempt++;
 930                                 if (carp_suppress_preempt == 1) {
 931                                         CARP_SCUNLOCK(sc);
 932                                         carp_send_ad_all();
 933                                         CARP_SCLOCK(sc);
 934                                 }
 935                         }
 936                         sc->sc_sendad_success = 0;
 937                 } else {
 938                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 939                                 if (++sc->sc_sendad_success >=
 940                                     CARP_SENDAD_MIN_SUCCESS) {
 941                                         carp_suppress_preempt--;
 942                                         sc->sc_sendad_errors = 0;
 943                                 }
 944                         } else
 945                                 sc->sc_sendad_errors = 0;
 946                 }
 947         }
 948 #endif /* INET */
 949 #ifdef INET6
 950         if (sc->sc_ia6) {
 951                 struct ip6_hdr *ip6;
 952
 953                 MGETHDR(m, M_NOWAIT, MT_HEADER);
 954                 if (m == NULL) {
 955                         SC2IFP(sc)->if_oerrors++;
 956                         carpstats.carps_onomem++;
 957                         /* XXX maybe less ? */
 958                         if (advbase != 255 || advskew != 255)
 959                                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
 960                                     carp_send_ad, sc);
 961                         return;
 962                 }
 963                 len = sizeof(*ip6) + sizeof(ch);
 964                 m->m_pkthdr.len = len;
 965                 m->m_pkthdr.rcvif = NULL;
 966                 m->m_len = len;
 967                 MH_ALIGN(m, m->m_len);
 968                 m->m_flags |= M_MCAST;
 969                 ip6 = mtod(m, struct ip6_hdr *);
 970                 bzero(ip6, sizeof(*ip6));
 971                 ip6->ip6_vfc |= IPV6_VERSION;
 972                 ip6->ip6_hlim = CARP_DFLTTL;
 973                 ip6->ip6_nxt = IPPROTO_CARP;
 974                 bcopy(&sc->sc_ia6->ia_addr.sin6_addr, &ip6->ip6_src,
 975                     sizeof(struct in6_addr));
 976                 /* set the multicast destination */
 977
 978                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
 979                 ip6->ip6_dst.s6_addr8[15] = 0x12;
 980                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
 981                         SC2IFP(sc)->if_oerrors++;
 982                         m_freem(m);
 983                         CARP_LOG("%s: in6_setscope failed\n", __func__);
 984                         return;
 985                 }
 986
 987                 ch_ptr = (struct carp_header *)(&ip6[1]);
 988                 bcopy(&ch, ch_ptr, sizeof(ch));
 989                 if (carp_prepare_ad(m, sc, ch_ptr))
 990                         return;
 991
 992                 m->m_data += sizeof(*ip6);
 993                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6));
 994                 m->m_data -= sizeof(*ip6);
 995
 996                 getmicrotime(&SC2IFP(sc)->if_lastchange);
 997                 SC2IFP(sc)->if_opackets++;
 998                 SC2IFP(sc)->if_obytes += len;
 999                 carpstats.carps_opackets6++;
1000
1001                 if (ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL)) {
1002                         SC2IFP(sc)->if_oerrors++;
1003                         if (sc->sc_sendad_errors < INT_MAX)
1004                                 sc->sc_sendad_errors++;
1005                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
1006                                 carp_suppress_preempt++;
1007                                 if (carp_suppress_preempt == 1) {
1008                                         CARP_SCUNLOCK(sc);
1009                                         carp_send_ad_all();
1010                                         CARP_SCLOCK(sc);
1011                                 }
1012                         }
1013                         sc->sc_sendad_success = 0;
1014                 } else {
1015                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
1016                                 if (++sc->sc_sendad_success >=
1017                                     CARP_SENDAD_MIN_SUCCESS) {
1018                                         carp_suppress_preempt--;
1019                                         sc->sc_sendad_errors = 0;
1020                                 }
1021                         } else
1022                                 sc->sc_sendad_errors = 0;
1023                 }
1024         }
1025 #endif /* INET6 */
1026
1027         if (advbase != 255 || advskew != 255)
1028                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
1029                     carp_send_ad, sc);
1030
1031 }
1032
1033 /*
1034  * Broadcast a gratuitous ARP request containing
1035  * the virtual router MAC address for each IP address
1036  * associated with the virtual router.
1037  */
1038 static void
1039 carp_send_arp(struct carp_softc *sc)
1040 {
1041         struct ifaddr *ifa;
1042
1043         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1044
1045                 if (ifa->ifa_addr->sa_family != AF_INET)
1046                         continue;
1047                 lwkt_serialize_enter(sc->sc_carpdev->if_serializer);
1048                 arp_ifinit2(sc->sc_carpdev, ifa, IF_LLADDR(sc->sc_ifp));
1049                 lwkt_serialize_exit(sc->sc_carpdev->if_serializer);
1050
1051                 DELAY(1000);    /* XXX */
1052         }
1053 }
1054
1055 #ifdef INET6
1056 static void
1057 carp_send_na(struct carp_softc *sc)
1058 {
1059         struct ifaddr *ifa;
1060         struct in6_addr *in6;
1061         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1062
1063         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1064
1065                 if (ifa->ifa_addr->sa_family != AF_INET6)
1066                         continue;
1067
1068                 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr;
1069                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1070                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1071                 DELAY(1000);    /* XXX */
1072         }
1073 }
1074 #endif /* INET6 */
1075
1076 static int
1077 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type)
1078 {
1079         struct carp_softc *vh;
1080         struct ifaddr *ifa;
1081         int count = 0;
1082
1083         CARP_LOCK_ASSERT(cif);
1084
1085         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1086                 if ((type == CARP_COUNT_RUNNING &&
1087                     (SC2IFP(vh)->if_flags & IFF_UP) && (SC2IFP(vh)->if_flags & IFF_RUNNING)) ||
1088                     (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) {
1089                         TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1090                             ifa_list) {
1091                                 if (ifa->ifa_addr->sa_family == AF_INET &&
1092                                     ia->ia_addr.sin_addr.s_addr ==
1093                                     ifatoia(ifa)->ia_addr.sin_addr.s_addr)
1094                                         count++;
1095                         }
1096                 }
1097         }
1098         return (count);
1099 }
1100
1101 int
1102 carp_iamatch(void *v, struct in_ifaddr *ia,
1103     struct in_addr *isaddr, u_int8_t **enaddr)
1104 {
1105         struct carp_if *cif = v;
1106         struct carp_softc *vh;
1107         int index, count = 0;
1108         struct ifaddr *ifa;
1109
1110         CARP_LOCK(cif);
1111
1112         if (carp_opts[CARPCTL_ARPBALANCE]) {
1113                 /*
1114                  * XXX proof of concept implementation.
1115                  * We use the source ip to decide which virtual host should
1116                  * handle the request. If we're master of that virtual host,
1117                  * then we respond, otherwise, just drop the arp packet on
1118                  * the floor.
1119                  */
1120                 count = carp_addrcount(cif, ia, CARP_COUNT_RUNNING);
1121                 if (count == 0) {
1122                         /* should never reach this */
1123                         CARP_UNLOCK(cif);
1124                         return (0);
1125                 }
1126
1127                 /* this should be a hash, like pf_hash() */
1128                 index = ntohl(isaddr->s_addr) % count;
1129                 count = 0;
1130
1131                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1132                         if ((SC2IFP(vh)->if_flags & IFF_UP) && (SC2IFP(vh)->if_flags & IFF_RUNNING)) {
1133                                 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1134                                     ifa_list) {
1135                                         if (ifa->ifa_addr->sa_family ==
1136                                             AF_INET &&
1137                                             ia->ia_addr.sin_addr.s_addr ==
1138                                             ifatoia(ifa)->ia_addr.sin_addr.s_addr) {
1139                                                 if (count == index) {
1140                                                         if (vh->sc_state ==
1141                                                             MASTER) {
1142                                                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1143                                                                 CARP_UNLOCK(cif);
1144                                                                 return (1);
1145                                                         } else {
1146                                                                 CARP_UNLOCK(cif);
1147                                                                 return (0);
1148                                                         }
1149                                                 }
1150                                                 count++;
1151                                         }
1152                                 }
1153                         }
1154                 }
1155         } else {
1156                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1157                         if ((SC2IFP(vh)->if_flags & IFF_UP) && (SC2IFP(vh)->if_flags & IFF_RUNNING) &&
1158                             vh->sc_state == MASTER) {
1159                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1160                                 CARP_UNLOCK(cif);
1161                                 return (1);
1162                         }
1163                 }
1164         }
1165         CARP_UNLOCK(cif);
1166         return(0);
1167 }
1168
1169 #ifdef INET6
1170 struct ifaddr *
1171 carp_iamatch6(void *v, struct in6_addr *taddr)
1172 {
1173         struct carp_if *cif = v;
1174         struct carp_softc *vh;
1175         struct ifaddr *ifa;
1176
1177         CARP_LOCK(cif);
1178         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1179                 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist, ifa_list) {
1180                         if (IN6_ARE_ADDR_EQUAL(taddr,
1181                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1182                             (SC2IFP(vh)->if_flags & IFF_UP) && (SC2IFP(vh)->if_flags & IFF_RUNNING) &&
1183                             vh->sc_state == MASTER) {
1184                                 CARP_UNLOCK(cif);
1185                                 return (ifa);
1186                         }
1187                 }
1188         }
1189         CARP_UNLOCK(cif);
1190
1191         return (NULL);
1192 }
1193
1194 void *
1195 carp_macmatch6(void *v, struct mbuf *m, const struct in6_addr *taddr)
1196 {
1197         struct m_tag *mtag;
1198         struct carp_if *cif = v;
1199         struct carp_softc *sc;
1200         struct ifaddr *ifa;
1201
1202         CARP_LOCK(cif);
1203         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) {
1204                 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1205                         if (IN6_ARE_ADDR_EQUAL(taddr,
1206                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1207                             (SC2IFP(sc)->if_flags & IFF_UP) && (SC2IFP(sc)->if_flags & IFF_RUNNING)) {
1208                                 struct ifnet *ifp = SC2IFP(sc);
1209                                 mtag = m_tag_get(PACKET_TAG_CARP,
1210                                     sizeof(struct ifnet *), MB_DONTWAIT);
1211                                 if (mtag == NULL) {
1212                                         /* better a bit than nothing */
1213                                         CARP_UNLOCK(cif);
1214                                         return (IF_LLADDR(sc->sc_ifp));
1215                                 }
1216                                 bcopy(&ifp, (caddr_t)(mtag + 1),
1217                                     sizeof(struct ifnet *));
1218                                 m_tag_prepend(m, mtag);
1219
1220                                 CARP_UNLOCK(cif);
1221                                 return (IF_LLADDR(sc->sc_ifp));
1222                         }
1223                 }
1224         }
1225         CARP_UNLOCK(cif);
1226
1227         return (NULL);
1228 }
1229 #endif
1230
1231 struct ifnet *
1232 carp_forus(void *v, void *dhost)
1233 {
1234         struct carp_if *cif = v;
1235         struct carp_softc *vh;
1236         u_int8_t *ena = dhost;
1237
1238         /**
1239          * XXX: See here for check on MAC adr is not for virtual use
1240          *
1241          **/
1242
1243         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1244         {
1245                 return (NULL);
1246         }
1247
1248         CARP_LOCK(cif);
1249         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list)
1250                 if ((SC2IFP(vh)->if_flags & IFF_UP) && (SC2IFP(vh)->if_flags & IFF_RUNNING) &&
1251                     vh->sc_state == MASTER &&
1252                     !bcmp(dhost, IF_LLADDR(vh->sc_ifp), ETHER_ADDR_LEN)) {
1253                         CARP_UNLOCK(cif);
1254                         return (SC2IFP(vh));
1255                 }
1256
1257         CARP_UNLOCK(cif);
1258         return (NULL);
1259 }
1260
1261 static void
1262 carp_master_down(void *v)
1263 {
1264         struct carp_softc *sc = v;
1265
1266         lwkt_serialize_enter(sc->sc_ifp->if_serializer);
1267         carp_master_down_locked(sc);
1268         lwkt_serialize_exit(sc->sc_ifp->if_serializer);
1269 }
1270
1271 static void
1272 carp_master_down_locked(struct carp_softc *sc)
1273 {
1274         if (sc->sc_carpdev)
1275                 CARP_SCLOCK_ASSERT(sc);
1276
1277         switch (sc->sc_state) {
1278         case INIT:
1279                 kprintf("%s: master_down event in INIT state\n",
1280                     SC2IFP(sc)->if_xname);
1281                 break;
1282         case MASTER:
1283                 break;
1284         case BACKUP:
1285                 carp_set_state(sc, MASTER);
1286                 carp_send_ad_locked(sc);
1287                 carp_send_arp(sc);
1288 #ifdef INET6
1289                 carp_send_na(sc);
1290 #endif /* INET6 */
1291                 carp_setrun(sc, 0);
1292                 carp_setroute(sc, RTM_ADD);
1293                 break;
1294         }
1295 }
1296
1297 /*
1298  * When in backup state, af indicates whether to reset the master down timer
1299  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1300  */
1301 static void
1302 carp_setrun(struct carp_softc *sc, sa_family_t af)
1303 {
1304         struct timeval tv;
1305
1306         if (sc->sc_carpdev == NULL) {
1307                 SC2IFP(sc)->if_flags &= ~IFF_RUNNING;
1308                 carp_set_state(sc, INIT);
1309                 return;
1310         }
1311
1312         if (SC2IFP(sc)->if_flags & IFF_UP &&
1313             sc->sc_vhid > 0 && (sc->sc_naddrs || sc->sc_naddrs6))
1314                 SC2IFP(sc)->if_flags |= IFF_RUNNING;
1315         else {
1316                 SC2IFP(sc)->if_flags &= ~IFF_RUNNING;
1317                 carp_setroute(sc, RTM_DELETE);
1318                 return;
1319         }
1320
1321         switch (sc->sc_state) {
1322         case INIT:
1323                 if (carp_opts[CARPCTL_PREEMPT] && !carp_suppress_preempt) {
1324                         carp_send_ad_locked(sc);
1325                         carp_send_arp(sc);
1326 #ifdef INET6
1327                         carp_send_na(sc);
1328 #endif /* INET6 */
1329                         CARP_DEBUG("%s: INIT -> MASTER (preempting)\n",
1330                             SC2IFP(sc)->if_xname);
1331                         carp_set_state(sc, MASTER);
1332                         carp_setroute(sc, RTM_ADD);
1333                 } else {
1334                         CARP_DEBUG("%s: INIT -> BACKUP\n", SC2IFP(sc)->if_xname);
1335                         carp_set_state(sc, BACKUP);
1336                         carp_setroute(sc, RTM_DELETE);
1337                         carp_setrun(sc, 0);
1338                 }
1339                 break;
1340         case BACKUP:
1341                 callout_stop(&sc->sc_ad_tmo);
1342                 tv.tv_sec = 3 * sc->sc_advbase;
1343                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1344                 switch (af) {
1345 #ifdef INET
1346                 case AF_INET:
1347                         callout_reset(&sc->sc_md_tmo, tvtohz_high(&tv),
1348                             carp_master_down, sc);
1349                         break;
1350 #endif /* INET */
1351 #ifdef INET6
1352                 case AF_INET6:
1353                         callout_reset(&sc->sc_md6_tmo, tvtohz_high(&tv),
1354                             carp_master_down, sc);
1355                         break;
1356 #endif /* INET6 */
1357                 default:
1358                         if (sc->sc_naddrs)
1359                                 callout_reset(&sc->sc_md_tmo, tvtohz_high(&tv),
1360                                     carp_master_down, sc);
1361                         if (sc->sc_naddrs6)
1362                                 callout_reset(&sc->sc_md6_tmo, tvtohz_high(&tv),
1363                                     carp_master_down, sc);
1364                         break;
1365                 }
1366                 break;
1367         case MASTER:
1368                 tv.tv_sec = sc->sc_advbase;
1369                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1370                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
1371                     carp_send_ad, sc);
1372                 break;
1373         }
1374 }
1375
1376 static void
1377 carp_multicast_cleanup(struct carp_softc *sc)
1378 {
1379         struct ip_moptions *imo = &sc->sc_imo;
1380         u_int16_t n = imo->imo_num_memberships;
1381
1382         /* Clean up our own multicast memberships */
1383         while (n-- > 0) {
1384                 if (imo->imo_membership[n] != NULL) {
1385                         in_delmulti(imo->imo_membership[n]);
1386                         imo->imo_membership[n] = NULL;
1387                 }
1388         }
1389         imo->imo_num_memberships = 0;
1390         imo->imo_multicast_ifp = NULL;
1391 }
1392
1393 #ifdef INET6
1394 static void
1395 carp_multicast6_cleanup(struct carp_softc *sc)
1396 {
1397         struct ip6_moptions *im6o = &sc->sc_im6o;
1398
1399         while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1400                 struct in6_multi_mship *imm =
1401                     LIST_FIRST(&im6o->im6o_memberships);
1402
1403                 LIST_REMOVE(imm, i6mm_chain);
1404                 in6_leavegroup(imm);
1405         }
1406         im6o->im6o_multicast_ifp = NULL;
1407 }
1408 #endif
1409
1410 static int
1411 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1412 {
1413         struct ifnet *ifp;
1414         struct carp_if *cif;
1415         struct in_ifaddr *ia, *ia_if;
1416         struct ip_moptions *imo = &sc->sc_imo;
1417         struct in_addr addr;
1418         u_long iaddr = htonl(sin->sin_addr.s_addr);
1419         int own, error;
1420
1421         if (sin->sin_addr.s_addr == 0)
1422         {
1423                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1424                 {
1425                         carp_set_state(sc, INIT);
1426                 }
1427                 if (sc->sc_naddrs)
1428                 {
1429                         SC2IFP(sc)->if_flags |= IFF_UP;
1430                 }
1431                 carp_setrun(sc, 0);
1432                 return (0);
1433         }
1434         /* we have to do it by hands to check we won't match on us */
1435         ia_if = NULL; own = 0;
1436         TAILQ_FOREACH(ia, &in_ifaddrhead, ia_link) {
1437                 /* and, yeah, we need a multicast-capable iface too */
1438                 if (ia->ia_ifp != SC2IFP(sc) &&
1439                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1440                     (iaddr & ia->ia_subnetmask) == ia->ia_subnet) {
1441                         if (!ia_if)
1442                                 ia_if = ia;
1443                         if (sin->sin_addr.s_addr ==
1444                             ia->ia_addr.sin_addr.s_addr)
1445                                 own++;
1446                 }
1447         }
1448
1449
1450         if (!ia_if)
1451                 return (EADDRNOTAVAIL);
1452
1453         ia = ia_if;
1454         ifp = ia->ia_ifp;
1455
1456         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1457             (imo->imo_multicast_ifp && imo->imo_multicast_ifp != ifp))
1458                 return (EADDRNOTAVAIL);
1459
1460         if (imo->imo_num_memberships == 0) {
1461                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1462                 if ((imo->imo_membership[0] = in_addmulti(&addr, ifp)) == NULL)
1463                         return (ENOBUFS);
1464                 imo->imo_num_memberships++;
1465                 imo->imo_multicast_ifp = ifp;
1466                 imo->imo_multicast_ttl = CARP_DFLTTL;
1467                 imo->imo_multicast_loop = 0;
1468         }
1469
1470         if (!ifp->if_carp) {
1471
1472                 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP,
1473                     M_WAITOK|M_ZERO);
1474                 if ((error = ifpromisc(ifp, 1))) {
1475                         FREE(cif, M_CARP);
1476                         goto cleanup;
1477                 }
1478
1479                 CARP_LOCK_INIT(cif);
1480                 CARP_LOCK(cif);
1481                 cif->vhif_ifp = ifp;
1482                 TAILQ_INIT(&cif->vhif_vrs);
1483                 ifp->if_carp = cif;
1484
1485         } else {
1486                 struct carp_softc *vr;
1487
1488                 cif = (struct carp_if *)ifp->if_carp;
1489                 CARP_LOCK(cif);
1490                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1491                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1492                                 CARP_UNLOCK(cif);
1493                                 error = EINVAL;
1494                                 goto cleanup;
1495                         }
1496         }
1497         sc->sc_ia = ia;
1498         sc->sc_carpdev = ifp;
1499
1500         { /* XXX prevent endless loop if already in queue */
1501         struct carp_softc *vr, *after = NULL;
1502         int myself = 0;
1503         cif = (struct carp_if *)ifp->if_carp;
1504
1505         /* XXX: cif should not change, right? So we still hold the lock */
1506         CARP_LOCK_ASSERT(cif);
1507
1508         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1509                 if (vr == sc)
1510                         myself = 1;
1511                 if (vr->sc_vhid < sc->sc_vhid)
1512                         after = vr;
1513         }
1514
1515         if (!myself) {
1516                 /* We're trying to keep things in order */
1517                 if (after == NULL) {
1518                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1519                 } else {
1520                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1521                 }
1522                 cif->vhif_nvrs++;
1523         }
1524         }
1525
1526         sc->sc_naddrs++;
1527         SC2IFP(sc)->if_flags |= IFF_UP;
1528         if (own)
1529                 sc->sc_advskew = 0;
1530
1531
1532         carp_sc_state_locked(sc);
1533         carp_setrun(sc, 0);
1534
1535         CARP_UNLOCK(cif);
1536
1537         return (0);
1538
1539 cleanup:
1540         in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1541         return (error);
1542
1543 }
1544
1545 static int
1546 carp_del_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1547 {
1548         int error = 0;
1549
1550         if (!--sc->sc_naddrs) {
1551                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1552                 struct ip_moptions *imo = &sc->sc_imo;
1553
1554                 CARP_LOCK(cif);
1555                 callout_stop(&sc->sc_ad_tmo);
1556                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1557                 SC2IFP(sc)->if_flags &= ~IFF_RUNNING;
1558                 sc->sc_vhid = -1;
1559                 in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1560                 imo->imo_multicast_ifp = NULL;
1561                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1562                 if (!--cif->vhif_nvrs) {
1563                         sc->sc_carpdev->if_carp = NULL;
1564                         CARP_LOCK_DESTROY(cif);
1565                         FREE(cif, M_IFADDR);
1566                 } else {
1567                         CARP_UNLOCK(cif);
1568                 }
1569         }
1570
1571         return (error);
1572 }
1573
1574 #ifdef INET6
1575 static int
1576 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1577 {
1578         struct ifnet *ifp;
1579         struct carp_if *cif;
1580         struct in6_ifaddr *ia, *ia_if;
1581         struct ip6_moptions *im6o = &sc->sc_im6o;
1582         struct in6_multi_mship *imm;
1583         struct in6_addr in6;
1584         int own, error;
1585
1586         if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
1587                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1588                         carp_set_state(sc, INIT);
1589                 if (sc->sc_naddrs6)
1590                         SC2IFP(sc)->if_flags |= IFF_UP;
1591                 carp_setrun(sc, 0);
1592                 return (0);
1593         }
1594
1595         /* we have to do it by hands to check we won't match on us */
1596         ia_if = NULL; own = 0;
1597         for (ia = in6_ifaddr; ia; ia = ia->ia_next) {
1598                 int i;
1599
1600                 for (i = 0; i < 4; i++) {
1601                         if ((sin6->sin6_addr.s6_addr32[i] &
1602                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]) !=
1603                             (ia->ia_addr.sin6_addr.s6_addr32[i] &
1604                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]))
1605                                 break;
1606                 }
1607                 /* and, yeah, we need a multicast-capable iface too */
1608                 if (ia->ia_ifp != SC2IFP(sc) &&
1609                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1610                     (i == 4)) {
1611                         if (!ia_if)
1612                                 ia_if = ia;
1613                         if (IN6_ARE_ADDR_EQUAL(&sin6->sin6_addr,
1614                             &ia->ia_addr.sin6_addr))
1615                                 own++;
1616                 }
1617         }
1618
1619         if (!ia_if)
1620                 return (EADDRNOTAVAIL);
1621         ia = ia_if;
1622         ifp = ia->ia_ifp;
1623
1624         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1625             (im6o->im6o_multicast_ifp && im6o->im6o_multicast_ifp != ifp))
1626                 return (EADDRNOTAVAIL);
1627
1628         if (!sc->sc_naddrs6) {
1629                 im6o->im6o_multicast_ifp = ifp;
1630
1631                 /* join CARP multicast address */
1632                 bzero(&in6, sizeof(in6));
1633                 in6.s6_addr16[0] = htons(0xff02);
1634                 in6.s6_addr8[15] = 0x12;
1635                 if (in6_setscope(&in6, ifp, NULL) != 0)
1636                         goto cleanup;
1637                 if ((imm = in6_joingroup(ifp, &in6, &error)) == NULL)
1638                         goto cleanup;
1639                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1640
1641                 /* join solicited multicast address */
1642                 bzero(&in6, sizeof(in6));
1643                 in6.s6_addr16[0] = htons(0xff02);
1644                 in6.s6_addr32[1] = 0;
1645                 in6.s6_addr32[2] = htonl(1);
1646                 in6.s6_addr32[3] = sin6->sin6_addr.s6_addr32[3];
1647                 in6.s6_addr8[12] = 0xff;
1648                 if (in6_setscope(&in6, ifp, NULL) != 0)
1649                         goto cleanup;
1650                 if ((imm = in6_joingroup(ifp, &in6, &error)) == NULL)
1651                         goto cleanup;
1652                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1653         }
1654
1655         if (!ifp->if_carp) {
1656                 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP,
1657                     M_WAITOK|M_ZERO);
1658                 if ((error = ifpromisc(ifp, 1))) {
1659                         FREE(cif, M_CARP);
1660                         goto cleanup;
1661                 }
1662
1663                 CARP_LOCK_INIT(cif);
1664                 CARP_LOCK(cif);
1665                 cif->vhif_ifp = ifp;
1666                 TAILQ_INIT(&cif->vhif_vrs);
1667                 ifp->if_carp = cif;
1668
1669         } else {
1670                 struct carp_softc *vr;
1671
1672                 cif = (struct carp_if *)ifp->if_carp;
1673                 CARP_LOCK(cif);
1674                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1675                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1676                                 CARP_UNLOCK(cif);
1677                                 error = EINVAL;
1678                                 goto cleanup;
1679                         }
1680         }
1681         sc->sc_ia6 = ia;
1682         sc->sc_carpdev = ifp;
1683
1684         { /* XXX prevent endless loop if already in queue */
1685         struct carp_softc *vr, *after = NULL;
1686         int myself = 0;
1687         cif = (struct carp_if *)ifp->if_carp;
1688         CARP_LOCK_ASSERT(cif);
1689
1690         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1691                 if (vr == sc)
1692                         myself = 1;
1693                 if (vr->sc_vhid < sc->sc_vhid)
1694                         after = vr;
1695         }
1696
1697         if (!myself) {
1698                 /* We're trying to keep things in order */
1699                 if (after == NULL) {
1700                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1701                 } else {
1702                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1703                 }
1704                 cif->vhif_nvrs++;
1705         }
1706         }
1707
1708         sc->sc_naddrs6++;
1709         SC2IFP(sc)->if_flags |= IFF_UP;
1710         if (own)
1711                 sc->sc_advskew = 0;
1712         carp_sc_state_locked(sc);
1713         carp_setrun(sc, 0);
1714
1715         CARP_UNLOCK(cif);
1716
1717         return (0);
1718
1719 cleanup:
1720         /* clean up multicast memberships */
1721         if (!sc->sc_naddrs6) {
1722                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1723                         imm = LIST_FIRST(&im6o->im6o_memberships);
1724                         LIST_REMOVE(imm, i6mm_chain);
1725                         in6_leavegroup(imm);
1726                 }
1727         }
1728         return (error);
1729 }
1730
1731 static int
1732 carp_del_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1733 {
1734         int error = 0;
1735
1736         if (!--sc->sc_naddrs6) {
1737                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1738                 struct ip6_moptions *im6o = &sc->sc_im6o;
1739
1740                 CARP_LOCK(cif);
1741                 callout_stop(&sc->sc_ad_tmo);
1742                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1743                 SC2IFP(sc)->if_flags &= ~IFF_RUNNING;
1744                 sc->sc_vhid = -1;
1745                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1746                         struct in6_multi_mship *imm =
1747                             LIST_FIRST(&im6o->im6o_memberships);
1748
1749                         LIST_REMOVE(imm, i6mm_chain);
1750                         in6_leavegroup(imm);
1751                 }
1752                 im6o->im6o_multicast_ifp = NULL;
1753                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1754                 if (!--cif->vhif_nvrs) {
1755                         CARP_LOCK_DESTROY(cif);
1756                         sc->sc_carpdev->if_carp = NULL;
1757                         FREE(cif, M_IFADDR);
1758                 } else
1759                         CARP_UNLOCK(cif);
1760         }
1761
1762         return (error);
1763 }
1764 #endif /* INET6 */
1765
1766 static int
1767 carp_ioctl(struct ifnet *ifp, u_long cmd, caddr_t addr, struct ucred *creds)
1768 {
1769         struct carp_softc *sc = ifp->if_softc, *vr;
1770         struct carpreq carpr;
1771         struct ifaddr *ifa;
1772         struct ifreq *ifr;
1773         struct ifaliasreq *ifra;
1774         int locked = 0, error = 0;
1775
1776         ifa = (struct ifaddr *)addr;
1777         ifra = (struct ifaliasreq *)addr;
1778         ifr = (struct ifreq *)addr;
1779
1780
1781         switch (cmd) {
1782         case SIOCSIFADDR:
1783                 switch (ifa->ifa_addr->sa_family) {
1784 #ifdef INET
1785                 case AF_INET:
1786                         SC2IFP(sc)->if_flags |= IFF_UP;
1787                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1788                             sizeof(struct sockaddr));
1789                         error = carp_set_addr(sc, satosin(ifa->ifa_addr));
1790                         break;
1791 #endif /* INET */
1792 #ifdef INET6
1793                 case AF_INET6:
1794                         SC2IFP(sc)->if_flags |= IFF_UP;
1795                         error = carp_set_addr6(sc, satosin6(ifa->ifa_addr));
1796                         break;
1797 #endif /* INET6 */
1798                 default:
1799                         error = EAFNOSUPPORT;
1800                         break;
1801                 }
1802                 break;
1803
1804         case SIOCAIFADDR:
1805                 switch (ifa->ifa_addr->sa_family) {
1806 #ifdef INET
1807                 case AF_INET:
1808                         SC2IFP(sc)->if_flags |= IFF_UP;
1809                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1810                             sizeof(struct sockaddr));
1811                         error = carp_set_addr(sc, satosin(&ifra->ifra_addr));
1812                         break;
1813 #endif /* INET */
1814 #ifdef INET6
1815                 case AF_INET6:
1816                         SC2IFP(sc)->if_flags |= IFF_UP;
1817                         error = carp_set_addr6(sc, satosin6(&ifra->ifra_addr));
1818                         break;
1819 #endif /* INET6 */
1820                 default:
1821                         error = EAFNOSUPPORT;
1822                         break;
1823                 }
1824                 break;
1825
1826         case SIOCDIFADDR:
1827                 switch (ifa->ifa_addr->sa_family) {
1828 #ifdef INET
1829                 case AF_INET:
1830                         error = carp_del_addr(sc, satosin(&ifra->ifra_addr));
1831                         break;
1832 #endif /* INET */
1833 #ifdef INET6
1834                 case AF_INET6:
1835                         error = carp_del_addr6(sc, satosin6(&ifra->ifra_addr));
1836                         break;
1837 #endif /* INET6 */
1838                 default:
1839                         error = EAFNOSUPPORT;
1840                         break;
1841                 }
1842                 break;
1843
1844         case SIOCSIFFLAGS:
1845                 if (sc->sc_carpdev) {
1846                         locked = 1;
1847                         CARP_SCLOCK(sc);
1848                 }
1849                 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) {
1850                         callout_stop(&sc->sc_ad_tmo);
1851                         callout_stop(&sc->sc_md_tmo);
1852                         callout_stop(&sc->sc_md6_tmo);
1853                         if (sc->sc_state == MASTER)
1854                                 carp_send_ad_locked(sc);
1855                         carp_set_state(sc, INIT);
1856                         carp_setrun(sc, 0);
1857                 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) {
1858                         SC2IFP(sc)->if_flags |= IFF_UP;
1859                         carp_setrun(sc, 0);
1860                 }
1861                 break;
1862
1863         case SIOCSVH:
1864                 error = suser(curthread);
1865                 if (error)
1866                         break;
1867                 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1868                         break;
1869                 error = 1;
1870                 if (sc->sc_carpdev) {
1871                         locked = 1;
1872                         CARP_SCLOCK(sc);
1873                 }
1874                 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) {
1875                         switch (carpr.carpr_state) {
1876                         case BACKUP:
1877                                 callout_stop(&sc->sc_ad_tmo);
1878                                 carp_set_state(sc, BACKUP);
1879                                 carp_setrun(sc, 0);
1880                                 carp_setroute(sc, RTM_DELETE);
1881                                 break;
1882                         case MASTER:
1883                                 carp_master_down_locked(sc);
1884                                 break;
1885                         default:
1886                                 break;
1887                         }
1888                 }
1889                 if (carpr.carpr_vhid > 0) {
1890                         if (carpr.carpr_vhid > 255) {
1891                                 error = EINVAL;
1892                                 break;
1893                         }
1894                         if (sc->sc_carpdev) {
1895                                 struct carp_if *cif;
1896                                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1897                                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1898                                         if (vr != sc &&
1899                                             vr->sc_vhid == carpr.carpr_vhid)
1900                                                 return EEXIST;
1901                         }
1902                         sc->sc_vhid = carpr.carpr_vhid;
1903                         IF_LLADDR(sc->sc_ifp)[0] = 0;
1904                         IF_LLADDR(sc->sc_ifp)[1] = 0;
1905                         IF_LLADDR(sc->sc_ifp)[2] = 0x5e;
1906                         IF_LLADDR(sc->sc_ifp)[3] = 0;
1907                         IF_LLADDR(sc->sc_ifp)[4] = 1;
1908                         IF_LLADDR(sc->sc_ifp)[5] = sc->sc_vhid;
1909                         error--;
1910                 }
1911                 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) {
1912                         if (carpr.carpr_advskew >= 255) {
1913                                 error = EINVAL;
1914                                 break;
1915                         }
1916                         if (carpr.carpr_advbase > 255) {
1917                                 error = EINVAL;
1918                                 break;
1919                         }
1920                         sc->sc_advbase = carpr.carpr_advbase;
1921                         sc->sc_advskew = carpr.carpr_advskew;
1922                         error--;
1923                 }
1924                 bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1925                 if (error > 0)
1926                         error = EINVAL;
1927                 else {
1928                         error = 0;
1929                         carp_setrun(sc, 0);
1930                 }
1931                 break;
1932
1933         case SIOCGVH:
1934                 /* XXX: lockless read */
1935                 bzero(&carpr, sizeof(carpr));
1936                 carpr.carpr_state = sc->sc_state;
1937                 carpr.carpr_vhid = sc->sc_vhid;
1938                 carpr.carpr_advbase = sc->sc_advbase;
1939                 carpr.carpr_advskew = sc->sc_advskew;
1940                 error = suser(curthread);
1941                 if (error == 0)
1942                         bcopy(sc->sc_key, carpr.carpr_key,
1943                             sizeof(carpr.carpr_key));
1944                 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1945                 break;
1946
1947         default:
1948                 error = EINVAL;
1949         }
1950
1951         if (locked)
1952                 CARP_SCUNLOCK(sc);
1953
1954         carp_hmac_prepare(sc);
1955
1956         return (error);
1957 }
1958
1959 /*
1960  * XXX: this is looutput. We should eventually use it from there.
1961  */
1962 static int
1963 carp_looutput(struct ifnet *ifp, struct mbuf *m, struct sockaddr *dst,
1964     struct rtentry *rt)
1965 {
1966         u_int32_t af;
1967
1968         M_ASSERTPKTHDR(m); /* check if we have the packet header */
1969
1970         if (rt && rt->rt_flags & (RTF_REJECT|RTF_BLACKHOLE)) {
1971                 m_freem(m);
1972                 return (rt->rt_flags & RTF_BLACKHOLE ? 0 :
1973                         rt->rt_flags & RTF_HOST ? EHOSTUNREACH : ENETUNREACH);
1974         }
1975
1976         ifp->if_opackets++;
1977         ifp->if_obytes += m->m_pkthdr.len;
1978
1979         /* BPF writes need to be handled specially. */
1980         if (dst->sa_family == AF_UNSPEC) {
1981                 bcopy(dst->sa_data, &af, sizeof(af));
1982                 dst->sa_family = af;
1983         }
1984
1985 #if 1   /* XXX */
1986         switch (dst->sa_family) {
1987         case AF_INET:
1988         case AF_INET6:
1989         case AF_IPX:
1990         case AF_APPLETALK:
1991                 break;
1992         default:
1993                 m_freem(m);
1994                 return (EAFNOSUPPORT);
1995         }
1996 #endif
1997         return(if_simloop(ifp, m, dst->sa_family, 0));
1998 }
1999
2000 /*
2001  * Start output on carp interface. This function should never be called.
2002  */
2003 static void
2004 carp_start(struct ifnet *ifp)
2005 {
2006 #ifdef DEBUG
2007         kprintf("%s: start called\n", ifp->if_xname);
2008 #endif
2009 }
2010
2011 int
2012 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *sa,
2013     struct rtentry *rt)
2014 {
2015         struct m_tag *mtag;
2016         struct carp_softc *sc;
2017         struct ifnet *carp_ifp;
2018
2019         if (!sa)
2020                 return (0);
2021
2022         switch (sa->sa_family) {
2023 #ifdef INET
2024         case AF_INET:
2025                 break;
2026 #endif /* INET */
2027 #ifdef INET6
2028         case AF_INET6:
2029                 break;
2030 #endif /* INET6 */
2031         default:
2032                 return (0);
2033         }
2034
2035         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
2036         if (mtag == NULL)
2037                 return (0);
2038
2039         bcopy(mtag + 1, &carp_ifp, sizeof(struct ifnet *));
2040         sc = carp_ifp->if_softc;
2041
2042         /* Set the source MAC address to Virtual Router MAC Address */
2043         switch (ifp->if_type) {
2044         case IFT_ETHER:
2045         case IFT_L2VLAN: {
2046                         struct ether_header *eh;
2047
2048                         eh = mtod(m, struct ether_header *);
2049                         eh->ether_shost[0] = 0;
2050                         eh->ether_shost[1] = 0;
2051                         eh->ether_shost[2] = 0x5e;
2052                         eh->ether_shost[3] = 0;
2053                         eh->ether_shost[4] = 1;
2054                         eh->ether_shost[5] = sc->sc_vhid;
2055                 }
2056                 break;
2057         default:
2058                 kprintf("%s: carp is not supported for this interface type\n",
2059                     ifp->if_xname);
2060                 return (EOPNOTSUPP);
2061         }
2062
2063         return (0);
2064
2065 }
2066
2067 static void
2068 carp_set_state(struct carp_softc *sc, int state)
2069 {
2070
2071         if (sc->sc_carpdev)
2072                 CARP_SCLOCK_ASSERT(sc);
2073
2074         if (sc->sc_state == state)
2075                 return;
2076
2077         sc->sc_state = state;
2078         switch (state) {
2079         case BACKUP:
2080                 SC2IFP(sc)->if_link_state = LINK_STATE_DOWN;
2081                 break;
2082         case MASTER:
2083                 SC2IFP(sc)->if_link_state = LINK_STATE_UP;
2084                 break;
2085         default:
2086                 SC2IFP(sc)->if_link_state = LINK_STATE_UNKNOWN;
2087                 break;
2088         }
2089         rt_ifmsg(SC2IFP(sc));
2090 }
2091
2092 void
2093 carp_carpdev_state(void *v)
2094 {
2095         struct carp_if *cif = v;
2096
2097         CARP_LOCK(cif);
2098         carp_carpdev_state_locked(cif);
2099         CARP_UNLOCK(cif);
2100 }
2101
2102 static void
2103 carp_carpdev_state_locked(struct carp_if *cif)
2104 {
2105         struct carp_softc *sc;
2106
2107         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list)
2108                 carp_sc_state_locked(sc);
2109 }
2110
2111 static void
2112 carp_sc_state_locked(struct carp_softc *sc)
2113 {
2114         CARP_SCLOCK_ASSERT(sc);
2115
2116         if ( !(sc->sc_carpdev->if_flags & IFF_UP)) {
2117                 sc->sc_flags_backup = SC2IFP(sc)->if_flags;
2118                 SC2IFP(sc)->if_flags &= ~IFF_UP;
2119                 SC2IFP(sc)->if_flags &= ~IFF_RUNNING;
2120                 callout_stop(&sc->sc_ad_tmo);
2121                 callout_stop(&sc->sc_md_tmo);
2122                 callout_stop(&sc->sc_md6_tmo);
2123                 carp_set_state(sc, INIT);
2124                 carp_setrun(sc, 0);
2125                 if (!sc->sc_suppress) {
2126                         carp_suppress_preempt++;
2127                         if (carp_suppress_preempt == 1) {
2128                                 CARP_SCUNLOCK(sc);
2129                                 carp_send_ad_all();
2130                                 CARP_SCLOCK(sc);
2131                         }
2132                 }
2133                 sc->sc_suppress = 1;
2134         } else {
2135                 SC2IFP(sc)->if_flags |= sc->sc_flags_backup;
2136                 carp_set_state(sc, INIT);
2137                 carp_setrun(sc, 0);
2138                 if (sc->sc_suppress)
2139                         carp_suppress_preempt--;
2140                 sc->sc_suppress = 0;
2141         }
2142
2143         return;
2144 }
2145
2146 static int
2147 carp_modevent(module_t mod, int type, void *data)
2148 {
2149         switch (type) {
2150         case MOD_LOAD:
2151                 if_detach_event_tag = EVENTHANDLER_REGISTER(ifnet_departure_event,
2152                     carp_ifdetach, NULL, EVENTHANDLER_PRI_ANY);
2153                 if (if_detach_event_tag == NULL)
2154                         return (ENOMEM);
2155
2156                 LIST_INIT(&carpif_list);
2157                 if_clone_attach(&carp_cloner);
2158                 break;
2159
2160         case MOD_UNLOAD:
2161                 EVENTHANDLER_DEREGISTER(ifnet_departure_event, if_detach_event_tag);
2162                 if_clone_detach(&carp_cloner);
2163                 break;
2164
2165         default:
2166                 return (EINVAL);
2167         }
2168
2169         return (0);
2170 }
2171
2172 static moduledata_t carp_mod = {
2173         "carp",
2174         carp_modevent,
2175         0
2176 };
2177
2178 DECLARE_MODULE(carp, carp_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);