sys/netinet/ip_carp.c

   1 /*
   2  * Copyright (c) 2002 Michael Shalayeff. All rights reserved.
   3  * Copyright (c) 2003 Ryan McBride. All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  *
  14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  15  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  16  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  17  * IN NO EVENT SHALL THE AUTHOR OR HIS RELATIVES BE LIABLE FOR ANY DIRECT,
  18  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  19  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  20  * SERVICES; LOSS OF MIND, USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
  22  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
  23  * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
  24  * THE POSSIBILITY OF SUCH DAMAGE.
  25  */
  26
  27 #include <sys/cdefs.h>
  28 __FBSDID("$FreeBSD$");
  29
  30 #include "opt_carp.h"
  31 #include "opt_bpf.h"
  32 #include "opt_inet.h"
  33 #include "opt_inet6.h"
  34
  35 #include <sys/types.h>
  36 #include <sys/param.h>
  37 #include <sys/systm.h>
  38 #include <sys/conf.h>
  39 #include <sys/kernel.h>
  40 #include <sys/limits.h>
  41 #include <sys/malloc.h>
  42 #include <sys/mbuf.h>
  43 #include <sys/module.h>
  44 #include <sys/time.h>
  45 #include <sys/priv.h>
  46 #include <sys/proc.h>
  47 #include <sys/sysctl.h>
  48 #include <sys/syslog.h>
  49 #include <sys/signalvar.h>
  50 #include <sys/filio.h>
  51 #include <sys/sockio.h>
  52
  53 #include <sys/socket.h>
  54 #include <sys/vnode.h>
  55 #include <sys/vimage.h>
  56
  57 #include <machine/stdarg.h>
  58
  59 #include <net/bpf.h>
  60 #include <net/ethernet.h>
  61 #include <net/fddi.h>
  62 #include <net/iso88025.h>
  63 #include <net/if.h>
  64 #include <net/if_clone.h>
  65 #include <net/if_dl.h>
  66 #include <net/if_types.h>
  67 #include <net/route.h>
  68
  69 #ifdef INET
  70 #include <netinet/in.h>
  71 #include <netinet/in_var.h>
  72 #include <netinet/in_systm.h>
  73 #include <netinet/ip.h>
  74 #include <netinet/ip_var.h>
  75 #include <netinet/if_ether.h>
  76 #include <machine/in_cksum.h>
  77 #include <netinet/vinet.h>
  78 #endif
  79
  80 #ifdef INET6
  81 #include <netinet/icmp6.h>
  82 #include <netinet/ip6.h>
  83 #include <netinet6/ip6_var.h>
  84 #include <netinet6/scope6_var.h>
  85 #include <netinet6/nd6.h>
  86 #include <netinet6/vinet6.h>
  87 #endif
  88
  89 #include <crypto/sha1.h>
  90 #include <netinet/ip_carp.h>
  91
  92 #define CARP_IFNAME     "carp"
  93 static MALLOC_DEFINE(M_CARP, "CARP", "CARP interfaces");
  94 SYSCTL_DECL(_net_inet_carp);
  95
  96 struct carp_softc {
  97         struct ifnet            *sc_ifp;        /* Interface clue */
  98         struct ifnet            *sc_carpdev;    /* Pointer to parent interface */
  99         struct in_ifaddr        *sc_ia;         /* primary iface address */
 100         struct ip_moptions       sc_imo;
 101 #ifdef INET6
 102         struct in6_ifaddr       *sc_ia6;        /* primary iface address v6 */
 103         struct ip6_moptions      sc_im6o;
 104 #endif /* INET6 */
 105         TAILQ_ENTRY(carp_softc)  sc_list;
 106
 107         enum { INIT = 0, BACKUP, MASTER }       sc_state;
 108
 109         int                      sc_flags_backup;
 110         int                      sc_suppress;
 111
 112         int                      sc_sendad_errors;
 113 #define CARP_SENDAD_MAX_ERRORS  3
 114         int                      sc_sendad_success;
 115 #define CARP_SENDAD_MIN_SUCCESS 3
 116
 117         int                      sc_vhid;
 118         int                      sc_advskew;
 119         int                      sc_naddrs;
 120         int                      sc_naddrs6;
 121         int                      sc_advbase;    /* seconds */
 122         int                      sc_init_counter;
 123         u_int64_t                sc_counter;
 124
 125         /* authentication */
 126 #define CARP_HMAC_PAD   64
 127         unsigned char sc_key[CARP_KEY_LEN];
 128         unsigned char sc_pad[CARP_HMAC_PAD];
 129         SHA1_CTX sc_sha1;
 130
 131         struct callout           sc_ad_tmo;     /* advertisement timeout */
 132         struct callout           sc_md_tmo;     /* master down timeout */
 133         struct callout           sc_md6_tmo;    /* master down timeout */
 134
 135         LIST_ENTRY(carp_softc)   sc_next;       /* Interface clue */
 136 };
 137 #define SC2IFP(sc)      ((sc)->sc_ifp)
 138
 139 int carp_suppress_preempt = 0;
 140 int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 1, 0, 0 };    /* XXX for now */
 141 SYSCTL_INT(_net_inet_carp, CARPCTL_ALLOW, allow, CTLFLAG_RW,
 142     &carp_opts[CARPCTL_ALLOW], 0, "Accept incoming CARP packets");
 143 SYSCTL_INT(_net_inet_carp, CARPCTL_PREEMPT, preempt, CTLFLAG_RW,
 144     &carp_opts[CARPCTL_PREEMPT], 0, "high-priority backup preemption mode");
 145 SYSCTL_INT(_net_inet_carp, CARPCTL_LOG, log, CTLFLAG_RW,
 146     &carp_opts[CARPCTL_LOG], 0, "log bad carp packets");
 147 SYSCTL_INT(_net_inet_carp, CARPCTL_ARPBALANCE, arpbalance, CTLFLAG_RW,
 148     &carp_opts[CARPCTL_ARPBALANCE], 0, "balance arp responses");
 149 SYSCTL_INT(_net_inet_carp, OID_AUTO, suppress_preempt, CTLFLAG_RD,
 150     &carp_suppress_preempt, 0, "Preemption is suppressed");
 151
 152 struct carpstats carpstats;
 153 SYSCTL_STRUCT(_net_inet_carp, CARPCTL_STATS, stats, CTLFLAG_RW,
 154     &carpstats, carpstats,
 155     "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 156
 157 struct carp_if {
 158         TAILQ_HEAD(, carp_softc) vhif_vrs;
 159         int vhif_nvrs;
 160
 161         struct ifnet    *vhif_ifp;
 162         struct mtx       vhif_mtx;
 163 };
 164
 165 /* Get carp_if from softc. Valid after carp_set_addr{,6}. */
 166 #define SC2CIF(sc)              ((struct carp_if *)(sc)->sc_carpdev->if_carp)
 167
 168 /* lock per carp_if queue */
 169 #define CARP_LOCK_INIT(cif)     mtx_init(&(cif)->vhif_mtx, "carp_if",   \
 170         NULL, MTX_DEF)
 171 #define CARP_LOCK_DESTROY(cif)  mtx_destroy(&(cif)->vhif_mtx)
 172 #define CARP_LOCK_ASSERT(cif)   mtx_assert(&(cif)->vhif_mtx, MA_OWNED)
 173 #define CARP_LOCK(cif)          mtx_lock(&(cif)->vhif_mtx)
 174 #define CARP_UNLOCK(cif)        mtx_unlock(&(cif)->vhif_mtx)
 175
 176 #define CARP_SCLOCK(sc)         mtx_lock(&SC2CIF(sc)->vhif_mtx)
 177 #define CARP_SCUNLOCK(sc)       mtx_unlock(&SC2CIF(sc)->vhif_mtx)
 178 #define CARP_SCLOCK_ASSERT(sc)  mtx_assert(&SC2CIF(sc)->vhif_mtx, MA_OWNED)
 179
 180 #define CARP_LOG(...)   do {                            \
 181         if (carp_opts[CARPCTL_LOG] > 0)                 \
 182                 log(LOG_INFO, __VA_ARGS__);             \
 183 } while (0)
 184
 185 #define CARP_DEBUG(...) do {                            \
 186         if (carp_opts[CARPCTL_LOG] > 1)                 \
 187                 log(LOG_DEBUG, __VA_ARGS__);            \
 188 } while (0)
 189
 190 static void     carp_hmac_prepare(struct carp_softc *);
 191 static void     carp_hmac_generate(struct carp_softc *, u_int32_t *,
 192                     unsigned char *);
 193 static int      carp_hmac_verify(struct carp_softc *, u_int32_t *,
 194                     unsigned char *);
 195 static void     carp_setroute(struct carp_softc *, int);
 196 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 197 static int      carp_clone_create(struct if_clone *, int, caddr_t);
 198 static void     carp_clone_destroy(struct ifnet *);
 199 static void     carpdetach(struct carp_softc *, int);
 200 static int      carp_prepare_ad(struct mbuf *, struct carp_softc *,
 201                     struct carp_header *);
 202 static void     carp_send_ad_all(void);
 203 static void     carp_send_ad(void *);
 204 static void     carp_send_ad_locked(struct carp_softc *);
 205 static void     carp_send_arp(struct carp_softc *);
 206 static void     carp_master_down(void *);
 207 static void     carp_master_down_locked(struct carp_softc *);
 208 static int      carp_ioctl(struct ifnet *, u_long, caddr_t);
 209 static int      carp_looutput(struct ifnet *, struct mbuf *, struct sockaddr *,
 210                     struct route *);
 211 static void     carp_start(struct ifnet *);
 212 static void     carp_setrun(struct carp_softc *, sa_family_t);
 213 static void     carp_set_state(struct carp_softc *, int);
 214 static int      carp_addrcount(struct carp_if *, struct in_ifaddr *, int);
 215 enum    { CARP_COUNT_MASTER, CARP_COUNT_RUNNING };
 216
 217 static void     carp_multicast_cleanup(struct carp_softc *);
 218 static int      carp_set_addr(struct carp_softc *, struct sockaddr_in *);
 219 static int      carp_del_addr(struct carp_softc *, struct sockaddr_in *);
 220 static void     carp_carpdev_state_locked(struct carp_if *);
 221 static void     carp_sc_state_locked(struct carp_softc *);
 222 #ifdef INET6
 223 static void     carp_send_na(struct carp_softc *);
 224 static int      carp_set_addr6(struct carp_softc *, struct sockaddr_in6 *);
 225 static int      carp_del_addr6(struct carp_softc *, struct sockaddr_in6 *);
 226 static void     carp_multicast6_cleanup(struct carp_softc *);
 227 #endif
 228
 229 static LIST_HEAD(, carp_softc) carpif_list;
 230 static struct mtx carp_mtx;
 231 IFC_SIMPLE_DECLARE(carp, 0);
 232
 233 static eventhandler_tag if_detach_event_tag;
 234
 235 static __inline u_int16_t
 236 carp_cksum(struct mbuf *m, int len)
 237 {
 238         return (in_cksum(m, len));
 239 }
 240
 241 static void
 242 carp_hmac_prepare(struct carp_softc *sc)
 243 {
 244         u_int8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 245         u_int8_t vhid = sc->sc_vhid & 0xff;
 246         struct ifaddr *ifa;
 247         int i, found;
 248 #ifdef INET
 249         struct in_addr last, cur, in;
 250 #endif
 251 #ifdef INET6
 252         struct in6_addr last6, cur6, in6;
 253 #endif
 254
 255         if (sc->sc_carpdev)
 256                 CARP_SCLOCK(sc);
 257
 258         /* XXX: possible race here */
 259
 260         /* compute ipad from key */
 261         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 262         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 263         for (i = 0; i < sizeof(sc->sc_pad); i++)
 264                 sc->sc_pad[i] ^= 0x36;
 265
 266         /* precompute first part of inner hash */
 267         SHA1Init(&sc->sc_sha1);
 268         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 269         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 270         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 271         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 272 #ifdef INET
 273         cur.s_addr = 0;
 274         do {
 275                 found = 0;
 276                 last = cur;
 277                 cur.s_addr = 0xffffffff;
 278                 IF_ADDR_LOCK(SC2IFP(sc));
 279                 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 280                         in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 281                         if (ifa->ifa_addr->sa_family == AF_INET &&
 282                             ntohl(in.s_addr) > ntohl(last.s_addr) &&
 283                             ntohl(in.s_addr) < ntohl(cur.s_addr)) {
 284                                 cur.s_addr = in.s_addr;
 285                                 found++;
 286                         }
 287                 }
 288                 IF_ADDR_UNLOCK(SC2IFP(sc));
 289                 if (found)
 290                         SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur));
 291         } while (found);
 292 #endif /* INET */
 293 #ifdef INET6
 294         memset(&cur6, 0, sizeof(cur6));
 295         do {
 296                 found = 0;
 297                 last6 = cur6;
 298                 memset(&cur6, 0xff, sizeof(cur6));
 299                 IF_ADDR_LOCK(SC2IFP(sc));
 300                 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 301                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 302                         if (IN6_IS_SCOPE_EMBED(&in6))
 303                                 in6.s6_addr16[1] = 0;
 304                         if (ifa->ifa_addr->sa_family == AF_INET6 &&
 305                             memcmp(&in6, &last6, sizeof(in6)) > 0 &&
 306                             memcmp(&in6, &cur6, sizeof(in6)) < 0) {
 307                                 cur6 = in6;
 308                                 found++;
 309                         }
 310                 }
 311                 IF_ADDR_UNLOCK(SC2IFP(sc));
 312                 if (found)
 313                         SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6));
 314         } while (found);
 315 #endif /* INET6 */
 316
 317         /* convert ipad to opad */
 318         for (i = 0; i < sizeof(sc->sc_pad); i++)
 319                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 320
 321         if (sc->sc_carpdev)
 322                 CARP_SCUNLOCK(sc);
 323 }
 324
 325 static void
 326 carp_hmac_generate(struct carp_softc *sc, u_int32_t counter[2],
 327     unsigned char md[20])
 328 {
 329         SHA1_CTX sha1ctx;
 330
 331         /* fetch first half of inner hash */
 332         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 333
 334         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 335         SHA1Final(md, &sha1ctx);
 336
 337         /* outer hash */
 338         SHA1Init(&sha1ctx);
 339         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 340         SHA1Update(&sha1ctx, md, 20);
 341         SHA1Final(md, &sha1ctx);
 342 }
 343
 344 static int
 345 carp_hmac_verify(struct carp_softc *sc, u_int32_t counter[2],
 346     unsigned char md[20])
 347 {
 348         unsigned char md2[20];
 349
 350         CARP_SCLOCK_ASSERT(sc);
 351
 352         carp_hmac_generate(sc, counter, md2);
 353
 354         return (bcmp(md, md2, sizeof(md2)));
 355 }
 356
 357 static void
 358 carp_setroute(struct carp_softc *sc, int cmd)
 359 {
 360         struct ifaddr *ifa;
 361         int s;
 362
 363         if (sc->sc_carpdev)
 364                 CARP_SCLOCK_ASSERT(sc);
 365
 366         s = splnet();
 367         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 368                 if (ifa->ifa_addr->sa_family == AF_INET &&
 369                     sc->sc_carpdev != NULL) {
 370                         int count = carp_addrcount(
 371                             (struct carp_if *)sc->sc_carpdev->if_carp,
 372                             ifatoia(ifa), CARP_COUNT_MASTER);
 373
 374                         if ((cmd == RTM_ADD && count == 1) ||
 375                             (cmd == RTM_DELETE && count == 0))
 376                                 rtinit(ifa, cmd, RTF_UP | RTF_HOST);
 377                 }
 378         }
 379         splx(s);
 380 }
 381
 382 static int
 383 carp_clone_create(struct if_clone *ifc, int unit, caddr_t params)
 384 {
 385
 386         struct carp_softc *sc;
 387         struct ifnet *ifp;
 388
 389         sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
 390         ifp = SC2IFP(sc) = if_alloc(IFT_ETHER);
 391         if (ifp == NULL) {
 392                 free(sc, M_CARP);
 393                 return (ENOSPC);
 394         }
 395
 396         sc->sc_flags_backup = 0;
 397         sc->sc_suppress = 0;
 398         sc->sc_advbase = CARP_DFLTINTV;
 399         sc->sc_vhid = -1;       /* required setting */
 400         sc->sc_advskew = 0;
 401         sc->sc_init_counter = 1;
 402         sc->sc_naddrs = sc->sc_naddrs6 = 0; /* M_ZERO? */
 403 #ifdef INET6
 404         sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL;
 405 #endif
 406         sc->sc_imo.imo_membership = (struct in_multi **)malloc(
 407             (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
 408             M_WAITOK);
 409         sc->sc_imo.imo_mfilters = NULL;
 410         sc->sc_imo.imo_max_memberships = IP_MIN_MEMBERSHIPS;
 411         sc->sc_imo.imo_multicast_vif = -1;
 412
 413         callout_init(&sc->sc_ad_tmo, CALLOUT_MPSAFE);
 414         callout_init(&sc->sc_md_tmo, CALLOUT_MPSAFE);
 415         callout_init(&sc->sc_md6_tmo, CALLOUT_MPSAFE);
 416
 417         ifp->if_softc = sc;
 418         if_initname(ifp, CARP_IFNAME, unit);
 419         ifp->if_mtu = ETHERMTU;
 420         ifp->if_flags = IFF_LOOPBACK;
 421         ifp->if_ioctl = carp_ioctl;
 422         ifp->if_output = carp_looutput;
 423         ifp->if_start = carp_start;
 424         ifp->if_type = IFT_CARP;
 425         ifp->if_snd.ifq_maxlen = ifqmaxlen;
 426         ifp->if_hdrlen = 0;
 427         if_attach(ifp);
 428         bpfattach(SC2IFP(sc), DLT_NULL, sizeof(u_int32_t));
 429         mtx_lock(&carp_mtx);
 430         LIST_INSERT_HEAD(&carpif_list, sc, sc_next);
 431         mtx_unlock(&carp_mtx);
 432         return (0);
 433 }
 434
 435 static void
 436 carp_clone_destroy(struct ifnet *ifp)
 437 {
 438         struct carp_softc *sc = ifp->if_softc;
 439
 440         if (sc->sc_carpdev)
 441                 CARP_SCLOCK(sc);
 442         carpdetach(sc, 1);      /* Returns unlocked. */
 443
 444         mtx_lock(&carp_mtx);
 445         LIST_REMOVE(sc, sc_next);
 446         mtx_unlock(&carp_mtx);
 447         bpfdetach(ifp);
 448         if_detach(ifp);
 449         if_free_type(ifp, IFT_ETHER);
 450         free(sc->sc_imo.imo_membership, M_CARP);
 451         free(sc, M_CARP);
 452 }
 453
 454 /*
 455  * This function can be called on CARP interface destroy path,
 456  * and in case of the removal of the underlying interface as
 457  * well. We differentiate these two cases. In the latter case
 458  * we do not cleanup our multicast memberships, since they
 459  * are already freed. Also, in the latter case we do not
 460  * release the lock on return, because the function will be
 461  * called once more, for another CARP instance on the same
 462  * interface.
 463  */
 464 static void
 465 carpdetach(struct carp_softc *sc, int unlock)
 466 {
 467         struct carp_if *cif;
 468
 469         callout_stop(&sc->sc_ad_tmo);
 470         callout_stop(&sc->sc_md_tmo);
 471         callout_stop(&sc->sc_md6_tmo);
 472
 473         if (sc->sc_suppress)
 474                 carp_suppress_preempt--;
 475         sc->sc_suppress = 0;
 476
 477         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS)
 478                 carp_suppress_preempt--;
 479         sc->sc_sendad_errors = 0;
 480
 481         carp_set_state(sc, INIT);
 482         SC2IFP(sc)->if_flags &= ~IFF_UP;
 483         carp_setrun(sc, 0);
 484         if (unlock)
 485                 carp_multicast_cleanup(sc);
 486 #ifdef INET6
 487         carp_multicast6_cleanup(sc);
 488 #endif
 489
 490         if (sc->sc_carpdev != NULL) {
 491                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
 492                 CARP_LOCK_ASSERT(cif);
 493                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
 494                 if (!--cif->vhif_nvrs) {
 495                         ifpromisc(sc->sc_carpdev, 0);
 496                         sc->sc_carpdev->if_carp = NULL;
 497                         CARP_LOCK_DESTROY(cif);
 498                         free(cif, M_IFADDR);
 499                 } else if (unlock)
 500                         CARP_UNLOCK(cif);
 501                 sc->sc_carpdev = NULL;
 502         }
 503 }
 504
 505 /* Detach an interface from the carp. */
 506 static void
 507 carp_ifdetach(void *arg __unused, struct ifnet *ifp)
 508 {
 509         struct carp_if *cif = (struct carp_if *)ifp->if_carp;
 510         struct carp_softc *sc, *nextsc;
 511
 512         if (cif == NULL)
 513                 return;
 514
 515         /*
 516          * XXX: At the end of for() cycle the lock will be destroyed.
 517          */
 518         CARP_LOCK(cif);
 519         for (sc = TAILQ_FIRST(&cif->vhif_vrs); sc; sc = nextsc) {
 520                 nextsc = TAILQ_NEXT(sc, sc_list);
 521                 carpdetach(sc, 0);
 522         }
 523 }
 524
 525 /*
 526  * process input packet.
 527  * we have rearranged checks order compared to the rfc,
 528  * but it seems more efficient this way or not possible otherwise.
 529  */
 530 void
 531 carp_input(struct mbuf *m, int hlen)
 532 {
 533         struct ip *ip = mtod(m, struct ip *);
 534         struct carp_header *ch;
 535         int iplen, len;
 536
 537         CARPSTATS_INC(carps_ipackets);
 538
 539         if (!carp_opts[CARPCTL_ALLOW]) {
 540                 m_freem(m);
 541                 return;
 542         }
 543
 544         /* check if received on a valid carp interface */
 545         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 546                 CARPSTATS_INC(carps_badif);
 547                 CARP_LOG("carp_input: packet received on non-carp "
 548                     "interface: %s\n",
 549                     m->m_pkthdr.rcvif->if_xname);
 550                 m_freem(m);
 551                 return;
 552         }
 553
 554         /* verify that the IP TTL is 255.  */
 555         if (ip->ip_ttl != CARP_DFLTTL) {
 556                 CARPSTATS_INC(carps_badttl);
 557                 CARP_LOG("carp_input: received ttl %d != 255i on %s\n",
 558                     ip->ip_ttl,
 559                     m->m_pkthdr.rcvif->if_xname);
 560                 m_freem(m);
 561                 return;
 562         }
 563
 564         iplen = ip->ip_hl << 2;
 565
 566         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 567                 CARPSTATS_INC(carps_badlen);
 568                 CARP_LOG("carp_input: received len %zd < "
 569                     "sizeof(struct carp_header)\n",
 570                     m->m_len - sizeof(struct ip));
 571                 m_freem(m);
 572                 return;
 573         }
 574
 575         if (iplen + sizeof(*ch) < m->m_len) {
 576                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 577                         CARPSTATS_INC(carps_hdrops);
 578                         CARP_LOG("carp_input: pullup failed\n");
 579                         return;
 580                 }
 581                 ip = mtod(m, struct ip *);
 582         }
 583         ch = (struct carp_header *)((char *)ip + iplen);
 584
 585         /*
 586          * verify that the received packet length is
 587          * equal to the CARP header
 588          */
 589         len = iplen + sizeof(*ch);
 590         if (len > m->m_pkthdr.len) {
 591                 CARPSTATS_INC(carps_badlen);
 592                 CARP_LOG("carp_input: packet too short %d on %s\n",
 593                     m->m_pkthdr.len,
 594                     m->m_pkthdr.rcvif->if_xname);
 595                 m_freem(m);
 596                 return;
 597         }
 598
 599         if ((m = m_pullup(m, len)) == NULL) {
 600                 CARPSTATS_INC(carps_hdrops);
 601                 return;
 602         }
 603         ip = mtod(m, struct ip *);
 604         ch = (struct carp_header *)((char *)ip + iplen);
 605
 606         /* verify the CARP checksum */
 607         m->m_data += iplen;
 608         if (carp_cksum(m, len - iplen)) {
 609                 CARPSTATS_INC(carps_badsum);
 610                 CARP_LOG("carp_input: checksum failed on %s\n",
 611                     m->m_pkthdr.rcvif->if_xname);
 612                 m_freem(m);
 613                 return;
 614         }
 615         m->m_data -= iplen;
 616
 617         carp_input_c(m, ch, AF_INET);
 618 }
 619
 620 #ifdef INET6
 621 int
 622 carp6_input(struct mbuf **mp, int *offp, int proto)
 623 {
 624         struct mbuf *m = *mp;
 625         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 626         struct carp_header *ch;
 627         u_int len;
 628
 629         CARPSTATS_INC(carps_ipackets6);
 630
 631         if (!carp_opts[CARPCTL_ALLOW]) {
 632                 m_freem(m);
 633                 return (IPPROTO_DONE);
 634         }
 635
 636         /* check if received on a valid carp interface */
 637         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 638                 CARPSTATS_INC(carps_badif);
 639                 CARP_LOG("carp6_input: packet received on non-carp "
 640                     "interface: %s\n",
 641                     m->m_pkthdr.rcvif->if_xname);
 642                 m_freem(m);
 643                 return (IPPROTO_DONE);
 644         }
 645
 646         /* verify that the IP TTL is 255 */
 647         if (ip6->ip6_hlim != CARP_DFLTTL) {
 648                 CARPSTATS_INC(carps_badttl);
 649                 CARP_LOG("carp6_input: received ttl %d != 255 on %s\n",
 650                     ip6->ip6_hlim,
 651                     m->m_pkthdr.rcvif->if_xname);
 652                 m_freem(m);
 653                 return (IPPROTO_DONE);
 654         }
 655
 656         /* verify that we have a complete carp packet */
 657         len = m->m_len;
 658         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 659         if (ch == NULL) {
 660                 CARPSTATS_INC(carps_badlen);
 661                 CARP_LOG("carp6_input: packet size %u too small\n", len);
 662                 return (IPPROTO_DONE);
 663         }
 664
 665
 666         /* verify the CARP checksum */
 667         m->m_data += *offp;
 668         if (carp_cksum(m, sizeof(*ch))) {
 669                 CARPSTATS_INC(carps_badsum);
 670                 CARP_LOG("carp6_input: checksum failed, on %s\n",
 671                     m->m_pkthdr.rcvif->if_xname);
 672                 m_freem(m);
 673                 return (IPPROTO_DONE);
 674         }
 675         m->m_data -= *offp;
 676
 677         carp_input_c(m, ch, AF_INET6);
 678         return (IPPROTO_DONE);
 679 }
 680 #endif /* INET6 */
 681
 682 static void
 683 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 684 {
 685         struct ifnet *ifp = m->m_pkthdr.rcvif;
 686         struct carp_softc *sc;
 687         u_int64_t tmp_counter;
 688         struct timeval sc_tv, ch_tv;
 689
 690         /* verify that the VHID is valid on the receiving interface */
 691         CARP_LOCK(ifp->if_carp);
 692         TAILQ_FOREACH(sc, &((struct carp_if *)ifp->if_carp)->vhif_vrs, sc_list)
 693                 if (sc->sc_vhid == ch->carp_vhid)
 694                         break;
 695
 696         if (!sc || !((SC2IFP(sc)->if_flags & IFF_UP) &&
 697             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING))) {
 698                 CARPSTATS_INC(carps_badvhid);
 699                 CARP_UNLOCK(ifp->if_carp);
 700                 m_freem(m);
 701                 return;
 702         }
 703
 704         getmicrotime(&SC2IFP(sc)->if_lastchange);
 705         SC2IFP(sc)->if_ipackets++;
 706         SC2IFP(sc)->if_ibytes += m->m_pkthdr.len;
 707
 708         if (bpf_peers_present(SC2IFP(sc)->if_bpf)) {
 709                 struct ip *ip = mtod(m, struct ip *);
 710                 uint32_t af1 = af;
 711
 712                 /* BPF wants net byte order */
 713                 ip->ip_len = htons(ip->ip_len + (ip->ip_hl << 2));
 714                 ip->ip_off = htons(ip->ip_off);
 715                 bpf_mtap2(SC2IFP(sc)->if_bpf, &af1, sizeof(af1), m);
 716         }
 717
 718         /* verify the CARP version. */
 719         if (ch->carp_version != CARP_VERSION) {
 720                 CARPSTATS_INC(carps_badver);
 721                 SC2IFP(sc)->if_ierrors++;
 722                 CARP_UNLOCK(ifp->if_carp);
 723                 CARP_LOG("%s; invalid version %d\n",
 724                     SC2IFP(sc)->if_xname,
 725                     ch->carp_version);
 726                 m_freem(m);
 727                 return;
 728         }
 729
 730         /* verify the hash */
 731         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 732                 CARPSTATS_INC(carps_badauth);
 733                 SC2IFP(sc)->if_ierrors++;
 734                 CARP_UNLOCK(ifp->if_carp);
 735                 CARP_LOG("%s: incorrect hash\n", SC2IFP(sc)->if_xname);
 736                 m_freem(m);
 737                 return;
 738         }
 739
 740         tmp_counter = ntohl(ch->carp_counter[0]);
 741         tmp_counter = tmp_counter<<32;
 742         tmp_counter += ntohl(ch->carp_counter[1]);
 743
 744         /* XXX Replay protection goes here */
 745
 746         sc->sc_init_counter = 0;
 747         sc->sc_counter = tmp_counter;
 748
 749         sc_tv.tv_sec = sc->sc_advbase;
 750         if (carp_suppress_preempt && sc->sc_advskew <  240)
 751                 sc_tv.tv_usec = 240 * 1000000 / 256;
 752         else
 753                 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256;
 754         ch_tv.tv_sec = ch->carp_advbase;
 755         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 756
 757         switch (sc->sc_state) {
 758         case INIT:
 759                 break;
 760         case MASTER:
 761                 /*
 762                  * If we receive an advertisement from a master who's going to
 763                  * be more frequent than us, go into BACKUP state.
 764                  */
 765                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 766                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 767                         callout_stop(&sc->sc_ad_tmo);
 768                         CARP_DEBUG("%s: MASTER -> BACKUP "
 769                            "(more frequent advertisement received)\n",
 770                            SC2IFP(sc)->if_xname);
 771                         carp_set_state(sc, BACKUP);
 772                         carp_setrun(sc, 0);
 773                         carp_setroute(sc, RTM_DELETE);
 774                 }
 775                 break;
 776         case BACKUP:
 777                 /*
 778                  * If we're pre-empting masters who advertise slower than us,
 779                  * and this one claims to be slower, treat him as down.
 780                  */
 781                 if (carp_opts[CARPCTL_PREEMPT] &&
 782                     timevalcmp(&sc_tv, &ch_tv, <)) {
 783                         CARP_DEBUG("%s: BACKUP -> MASTER "
 784                             "(preempting a slower master)\n",
 785                             SC2IFP(sc)->if_xname);
 786                         carp_master_down_locked(sc);
 787                         break;
 788                 }
 789
 790                 /*
 791                  *  If the master is going to advertise at such a low frequency
 792                  *  that he's guaranteed to time out, we'd might as well just
 793                  *  treat him as timed out now.
 794                  */
 795                 sc_tv.tv_sec = sc->sc_advbase * 3;
 796                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 797                         CARP_DEBUG("%s: BACKUP -> MASTER "
 798                             "(master timed out)\n",
 799                             SC2IFP(sc)->if_xname);
 800                         carp_master_down_locked(sc);
 801                         break;
 802                 }
 803
 804                 /*
 805                  * Otherwise, we reset the counter and wait for the next
 806                  * advertisement.
 807                  */
 808                 carp_setrun(sc, af);
 809                 break;
 810         }
 811
 812         CARP_UNLOCK(ifp->if_carp);
 813
 814         m_freem(m);
 815         return;
 816 }
 817
 818 static int
 819 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 820 {
 821         struct m_tag *mtag;
 822         struct ifnet *ifp = SC2IFP(sc);
 823
 824         if (sc->sc_init_counter) {
 825                 /* this could also be seconds since unix epoch */
 826                 sc->sc_counter = arc4random();
 827                 sc->sc_counter = sc->sc_counter << 32;
 828                 sc->sc_counter += arc4random();
 829         } else
 830                 sc->sc_counter++;
 831
 832         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 833         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 834
 835         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 836
 837         /* Tag packet for carp_output */
 838         mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct ifnet *), M_NOWAIT);
 839         if (mtag == NULL) {
 840                 m_freem(m);
 841                 SC2IFP(sc)->if_oerrors++;
 842                 return (ENOMEM);
 843         }
 844         bcopy(&ifp, (caddr_t)(mtag + 1), sizeof(struct ifnet *));
 845         m_tag_prepend(m, mtag);
 846
 847         return (0);
 848 }
 849
 850 static void
 851 carp_send_ad_all(void)
 852 {
 853         struct carp_softc *sc;
 854
 855         mtx_lock(&carp_mtx);
 856         LIST_FOREACH(sc, &carpif_list, sc_next) {
 857                 if (sc->sc_carpdev == NULL)
 858                         continue;
 859                 CARP_SCLOCK(sc);
 860                 if ((SC2IFP(sc)->if_flags & IFF_UP) &&
 861                     (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING) &&
 862                      sc->sc_state == MASTER)
 863                         carp_send_ad_locked(sc);
 864                 CARP_SCUNLOCK(sc);
 865         }
 866         mtx_unlock(&carp_mtx);
 867 }
 868
 869 static void
 870 carp_send_ad(void *v)
 871 {
 872         struct carp_softc *sc = v;
 873
 874         CARP_SCLOCK(sc);
 875         carp_send_ad_locked(sc);
 876         CARP_SCUNLOCK(sc);
 877 }
 878
 879 static void
 880 carp_send_ad_locked(struct carp_softc *sc)
 881 {
 882         struct carp_header ch;
 883         struct timeval tv;
 884         struct carp_header *ch_ptr;
 885         struct mbuf *m;
 886         int len, advbase, advskew;
 887
 888         CARP_SCLOCK_ASSERT(sc);
 889
 890         /* bow out if we've lost our UPness or RUNNINGuiness */
 891         if (!((SC2IFP(sc)->if_flags & IFF_UP) &&
 892             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING))) {
 893                 advbase = 255;
 894                 advskew = 255;
 895         } else {
 896                 advbase = sc->sc_advbase;
 897                 if (!carp_suppress_preempt || sc->sc_advskew > 240)
 898                         advskew = sc->sc_advskew;
 899                 else
 900                         advskew = 240;
 901                 tv.tv_sec = advbase;
 902                 tv.tv_usec = advskew * 1000000 / 256;
 903         }
 904
 905         ch.carp_version = CARP_VERSION;
 906         ch.carp_type = CARP_ADVERTISEMENT;
 907         ch.carp_vhid = sc->sc_vhid;
 908         ch.carp_advbase = advbase;
 909         ch.carp_advskew = advskew;
 910         ch.carp_authlen = 7;    /* XXX DEFINE */
 911         ch.carp_pad1 = 0;       /* must be zero */
 912         ch.carp_cksum = 0;
 913
 914 #ifdef INET
 915         INIT_VNET_INET(curvnet);
 916         if (sc->sc_ia) {
 917                 struct ip *ip;
 918
 919                 MGETHDR(m, M_DONTWAIT, MT_HEADER);
 920                 if (m == NULL) {
 921                         SC2IFP(sc)->if_oerrors++;
 922                         CARPSTATS_INC(carps_onomem);
 923                         /* XXX maybe less ? */
 924                         if (advbase != 255 || advskew != 255)
 925                                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
 926                                     carp_send_ad, sc);
 927                         return;
 928                 }
 929                 len = sizeof(*ip) + sizeof(ch);
 930                 m->m_pkthdr.len = len;
 931                 m->m_pkthdr.rcvif = NULL;
 932                 m->m_len = len;
 933                 MH_ALIGN(m, m->m_len);
 934                 m->m_flags |= M_MCAST;
 935                 ip = mtod(m, struct ip *);
 936                 ip->ip_v = IPVERSION;
 937                 ip->ip_hl = sizeof(*ip) >> 2;
 938                 ip->ip_tos = IPTOS_LOWDELAY;
 939                 ip->ip_len = len;
 940                 ip->ip_id = ip_newid();
 941                 ip->ip_off = IP_DF;
 942                 ip->ip_ttl = CARP_DFLTTL;
 943                 ip->ip_p = IPPROTO_CARP;
 944                 ip->ip_sum = 0;
 945                 ip->ip_src.s_addr = sc->sc_ia->ia_addr.sin_addr.s_addr;
 946                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 947
 948                 ch_ptr = (struct carp_header *)(&ip[1]);
 949                 bcopy(&ch, ch_ptr, sizeof(ch));
 950                 if (carp_prepare_ad(m, sc, ch_ptr))
 951                         return;
 952
 953                 m->m_data += sizeof(*ip);
 954                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip));
 955                 m->m_data -= sizeof(*ip);
 956
 957                 getmicrotime(&SC2IFP(sc)->if_lastchange);
 958                 SC2IFP(sc)->if_opackets++;
 959                 SC2IFP(sc)->if_obytes += len;
 960                 CARPSTATS_INC(carps_opackets);
 961
 962                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, NULL)) {
 963                         SC2IFP(sc)->if_oerrors++;
 964                         if (sc->sc_sendad_errors < INT_MAX)
 965                                 sc->sc_sendad_errors++;
 966                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 967                                 carp_suppress_preempt++;
 968                                 if (carp_suppress_preempt == 1) {
 969                                         CARP_SCUNLOCK(sc);
 970                                         carp_send_ad_all();
 971                                         CARP_SCLOCK(sc);
 972                                 }
 973                         }
 974                         sc->sc_sendad_success = 0;
 975                 } else {
 976                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 977                                 if (++sc->sc_sendad_success >=
 978                                     CARP_SENDAD_MIN_SUCCESS) {
 979                                         carp_suppress_preempt--;
 980                                         sc->sc_sendad_errors = 0;
 981                                 }
 982                         } else
 983                                 sc->sc_sendad_errors = 0;
 984                 }
 985         }
 986 #endif /* INET */
 987 #ifdef INET6
 988         if (sc->sc_ia6) {
 989                 struct ip6_hdr *ip6;
 990
 991                 MGETHDR(m, M_DONTWAIT, MT_HEADER);
 992                 if (m == NULL) {
 993                         SC2IFP(sc)->if_oerrors++;
 994                         CARPSTATS_INC(carps_onomem);
 995                         /* XXX maybe less ? */
 996                         if (advbase != 255 || advskew != 255)
 997                                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
 998                                     carp_send_ad, sc);
 999                         return;
1000                 }
1001                 len = sizeof(*ip6) + sizeof(ch);
1002                 m->m_pkthdr.len = len;
1003                 m->m_pkthdr.rcvif = NULL;
1004                 m->m_len = len;
1005                 MH_ALIGN(m, m->m_len);
1006                 m->m_flags |= M_MCAST;
1007                 ip6 = mtod(m, struct ip6_hdr *);
1008                 bzero(ip6, sizeof(*ip6));
1009                 ip6->ip6_vfc |= IPV6_VERSION;
1010                 ip6->ip6_hlim = CARP_DFLTTL;
1011                 ip6->ip6_nxt = IPPROTO_CARP;
1012                 bcopy(&sc->sc_ia6->ia_addr.sin6_addr, &ip6->ip6_src,
1013                     sizeof(struct in6_addr));
1014                 /* set the multicast destination */
1015
1016                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
1017                 ip6->ip6_dst.s6_addr8[15] = 0x12;
1018                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
1019                         SC2IFP(sc)->if_oerrors++;
1020                         m_freem(m);
1021                         CARP_LOG("%s: in6_setscope failed\n", __func__);
1022                         return;
1023                 }
1024
1025                 ch_ptr = (struct carp_header *)(&ip6[1]);
1026                 bcopy(&ch, ch_ptr, sizeof(ch));
1027                 if (carp_prepare_ad(m, sc, ch_ptr))
1028                         return;
1029
1030                 m->m_data += sizeof(*ip6);
1031                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6));
1032                 m->m_data -= sizeof(*ip6);
1033
1034                 getmicrotime(&SC2IFP(sc)->if_lastchange);
1035                 SC2IFP(sc)->if_opackets++;
1036                 SC2IFP(sc)->if_obytes += len;
1037                 CARPSTATS_INC(carps_opackets6);
1038
1039                 if (ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL)) {
1040                         SC2IFP(sc)->if_oerrors++;
1041                         if (sc->sc_sendad_errors < INT_MAX)
1042                                 sc->sc_sendad_errors++;
1043                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
1044                                 carp_suppress_preempt++;
1045                                 if (carp_suppress_preempt == 1) {
1046                                         CARP_SCUNLOCK(sc);
1047                                         carp_send_ad_all();
1048                                         CARP_SCLOCK(sc);
1049                                 }
1050                         }
1051                         sc->sc_sendad_success = 0;
1052                 } else {
1053                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
1054                                 if (++sc->sc_sendad_success >=
1055                                     CARP_SENDAD_MIN_SUCCESS) {
1056                                         carp_suppress_preempt--;
1057                                         sc->sc_sendad_errors = 0;
1058                                 }
1059                         } else
1060                                 sc->sc_sendad_errors = 0;
1061                 }
1062         }
1063 #endif /* INET6 */
1064
1065         if (advbase != 255 || advskew != 255)
1066                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1067                     carp_send_ad, sc);
1068
1069 }
1070
1071 /*
1072  * Broadcast a gratuitous ARP request containing
1073  * the virtual router MAC address for each IP address
1074  * associated with the virtual router.
1075  */
1076 static void
1077 carp_send_arp(struct carp_softc *sc)
1078 {
1079         struct ifaddr *ifa;
1080
1081         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1082
1083                 if (ifa->ifa_addr->sa_family != AF_INET)
1084                         continue;
1085
1086 /*              arprequest(sc->sc_carpdev, &in, &in, IF_LLADDR(sc->sc_ifp)); */
1087                 arp_ifinit2(sc->sc_carpdev, ifa, IF_LLADDR(sc->sc_ifp));
1088
1089                 DELAY(1000);    /* XXX */
1090         }
1091 }
1092
1093 #ifdef INET6
1094 static void
1095 carp_send_na(struct carp_softc *sc)
1096 {
1097         struct ifaddr *ifa;
1098         struct in6_addr *in6;
1099         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1100
1101         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1102
1103                 if (ifa->ifa_addr->sa_family != AF_INET6)
1104                         continue;
1105
1106                 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr;
1107                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1108                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1109                 DELAY(1000);    /* XXX */
1110         }
1111 }
1112 #endif /* INET6 */
1113
1114 static int
1115 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type)
1116 {
1117         struct carp_softc *vh;
1118         struct ifaddr *ifa;
1119         int count = 0;
1120
1121         CARP_LOCK_ASSERT(cif);
1122
1123         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1124                 if ((type == CARP_COUNT_RUNNING &&
1125                     (SC2IFP(vh)->if_flags & IFF_UP) &&
1126                     (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING)) ||
1127                     (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) {
1128                         IF_ADDR_LOCK(SC2IFP(vh));
1129                         TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1130                             ifa_list) {
1131                                 if (ifa->ifa_addr->sa_family == AF_INET &&
1132                                     ia->ia_addr.sin_addr.s_addr ==
1133                                     ifatoia(ifa)->ia_addr.sin_addr.s_addr)
1134                                         count++;
1135                         }
1136                         IF_ADDR_UNLOCK(SC2IFP(vh));
1137                 }
1138         }
1139         return (count);
1140 }
1141
1142 int
1143 carp_iamatch(void *v, struct in_ifaddr *ia,
1144     struct in_addr *isaddr, u_int8_t **enaddr)
1145 {
1146         struct carp_if *cif = v;
1147         struct carp_softc *vh;
1148         int index, count = 0;
1149         struct ifaddr *ifa;
1150
1151         CARP_LOCK(cif);
1152
1153         if (carp_opts[CARPCTL_ARPBALANCE]) {
1154                 /*
1155                  * XXX proof of concept implementation.
1156                  * We use the source ip to decide which virtual host should
1157                  * handle the request. If we're master of that virtual host,
1158                  * then we respond, otherwise, just drop the arp packet on
1159                  * the floor.
1160                  */
1161                 count = carp_addrcount(cif, ia, CARP_COUNT_RUNNING);
1162                 if (count == 0) {
1163                         /* should never reach this */
1164                         CARP_UNLOCK(cif);
1165                         return (0);
1166                 }
1167
1168                 /* this should be a hash, like pf_hash() */
1169                 index = ntohl(isaddr->s_addr) % count;
1170                 count = 0;
1171
1172                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1173                         if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1174                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING)) {
1175                                 IF_ADDR_LOCK(SC2IFP(vh));
1176                                 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1177                                     ifa_list) {
1178                                         if (ifa->ifa_addr->sa_family ==
1179                                             AF_INET &&
1180                                             ia->ia_addr.sin_addr.s_addr ==
1181                                             ifatoia(ifa)->ia_addr.sin_addr.s_addr) {
1182                                                 if (count == index) {
1183                                                         if (vh->sc_state ==
1184                                                             MASTER) {
1185                                                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1186                                                                 IF_ADDR_UNLOCK(SC2IFP(vh));
1187                                                                 CARP_UNLOCK(cif);
1188                                                                 return (1);
1189                                                         } else {
1190                                                                 IF_ADDR_UNLOCK(SC2IFP(vh));
1191                                                                 CARP_UNLOCK(cif);
1192                                                                 return (0);
1193                                                         }
1194                                                 }
1195                                                 count++;
1196                                         }
1197                                 }
1198                                 IF_ADDR_UNLOCK(SC2IFP(vh));
1199                         }
1200                 }
1201         } else {
1202                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1203                         if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1204                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1205                             ia->ia_ifp == SC2IFP(vh) &&
1206                             vh->sc_state == MASTER) {
1207                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1208                                 CARP_UNLOCK(cif);
1209                                 return (1);
1210                         }
1211                 }
1212         }
1213         CARP_UNLOCK(cif);
1214         return (0);
1215 }
1216
1217 #ifdef INET6
1218 struct ifaddr *
1219 carp_iamatch6(void *v, struct in6_addr *taddr)
1220 {
1221         struct carp_if *cif = v;
1222         struct carp_softc *vh;
1223         struct ifaddr *ifa;
1224
1225         CARP_LOCK(cif);
1226         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1227                 IF_ADDR_LOCK(SC2IFP(vh));
1228                 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist, ifa_list) {
1229                         if (IN6_ARE_ADDR_EQUAL(taddr,
1230                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1231                             (SC2IFP(vh)->if_flags & IFF_UP) &&
1232                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1233                             vh->sc_state == MASTER) {
1234                                 IF_ADDR_UNLOCK(SC2IFP(vh));
1235                                 CARP_UNLOCK(cif);
1236                                 return (ifa);
1237                         }
1238                 }
1239                 IF_ADDR_UNLOCK(SC2IFP(vh));
1240         }
1241         CARP_UNLOCK(cif);
1242
1243         return (NULL);
1244 }
1245
1246 void *
1247 carp_macmatch6(void *v, struct mbuf *m, const struct in6_addr *taddr)
1248 {
1249         struct m_tag *mtag;
1250         struct carp_if *cif = v;
1251         struct carp_softc *sc;
1252         struct ifaddr *ifa;
1253
1254         CARP_LOCK(cif);
1255         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) {
1256                 IF_ADDR_LOCK(SC2IFP(sc));
1257                 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1258                         if (IN6_ARE_ADDR_EQUAL(taddr,
1259                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1260                             (SC2IFP(sc)->if_flags & IFF_UP) &&
1261                             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING)) {
1262                                 struct ifnet *ifp = SC2IFP(sc);
1263                                 mtag = m_tag_get(PACKET_TAG_CARP,
1264                                     sizeof(struct ifnet *), M_NOWAIT);
1265                                 if (mtag == NULL) {
1266                                         /* better a bit than nothing */
1267                                         IF_ADDR_UNLOCK(SC2IFP(sc));
1268                                         CARP_UNLOCK(cif);
1269                                         return (IF_LLADDR(sc->sc_ifp));
1270                                 }
1271                                 bcopy(&ifp, (caddr_t)(mtag + 1),
1272                                     sizeof(struct ifnet *));
1273                                 m_tag_prepend(m, mtag);
1274
1275                                 IF_ADDR_UNLOCK(SC2IFP(sc));
1276                                 CARP_UNLOCK(cif);
1277                                 return (IF_LLADDR(sc->sc_ifp));
1278                         }
1279                 }
1280                 IF_ADDR_UNLOCK(SC2IFP(sc));
1281         }
1282         CARP_UNLOCK(cif);
1283
1284         return (NULL);
1285 }
1286 #endif
1287
1288 struct ifnet *
1289 carp_forus(void *v, void *dhost)
1290 {
1291         struct carp_if *cif = v;
1292         struct carp_softc *vh;
1293         u_int8_t *ena = dhost;
1294
1295         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1296                 return (NULL);
1297
1298         CARP_LOCK(cif);
1299         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list)
1300                 if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1301                     (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1302                     vh->sc_state == MASTER &&
1303                     !bcmp(dhost, IF_LLADDR(vh->sc_ifp), ETHER_ADDR_LEN)) {
1304                         CARP_UNLOCK(cif);
1305                         return (SC2IFP(vh));
1306                 }
1307
1308         CARP_UNLOCK(cif);
1309         return (NULL);
1310 }
1311
1312 static void
1313 carp_master_down(void *v)
1314 {
1315         struct carp_softc *sc = v;
1316
1317         CARP_SCLOCK(sc);
1318         carp_master_down_locked(sc);
1319         CARP_SCUNLOCK(sc);
1320 }
1321
1322 static void
1323 carp_master_down_locked(struct carp_softc *sc)
1324 {
1325         if (sc->sc_carpdev)
1326                 CARP_SCLOCK_ASSERT(sc);
1327
1328         switch (sc->sc_state) {
1329         case INIT:
1330                 printf("%s: master_down event in INIT state\n",
1331                     SC2IFP(sc)->if_xname);
1332                 break;
1333         case MASTER:
1334                 break;
1335         case BACKUP:
1336                 carp_set_state(sc, MASTER);
1337                 carp_send_ad_locked(sc);
1338                 carp_send_arp(sc);
1339 #ifdef INET6
1340                 carp_send_na(sc);
1341 #endif /* INET6 */
1342                 carp_setrun(sc, 0);
1343                 carp_setroute(sc, RTM_ADD);
1344                 break;
1345         }
1346 }
1347
1348 /*
1349  * When in backup state, af indicates whether to reset the master down timer
1350  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1351  */
1352 static void
1353 carp_setrun(struct carp_softc *sc, sa_family_t af)
1354 {
1355         struct timeval tv;
1356
1357         if (sc->sc_carpdev == NULL) {
1358                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1359                 carp_set_state(sc, INIT);
1360                 return;
1361         } else
1362                 CARP_SCLOCK_ASSERT(sc);
1363
1364         if (SC2IFP(sc)->if_flags & IFF_UP &&
1365             sc->sc_vhid > 0 && (sc->sc_naddrs || sc->sc_naddrs6))
1366                 SC2IFP(sc)->if_drv_flags |= IFF_DRV_RUNNING;
1367         else {
1368                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1369                 carp_setroute(sc, RTM_DELETE);
1370                 return;
1371         }
1372
1373         switch (sc->sc_state) {
1374         case INIT:
1375                 if (carp_opts[CARPCTL_PREEMPT] && !carp_suppress_preempt) {
1376                         carp_send_ad_locked(sc);
1377                         carp_send_arp(sc);
1378 #ifdef INET6
1379                         carp_send_na(sc);
1380 #endif /* INET6 */
1381                         CARP_DEBUG("%s: INIT -> MASTER (preempting)\n",
1382                             SC2IFP(sc)->if_xname);
1383                         carp_set_state(sc, MASTER);
1384                         carp_setroute(sc, RTM_ADD);
1385                 } else {
1386                         CARP_DEBUG("%s: INIT -> BACKUP\n", SC2IFP(sc)->if_xname);
1387                         carp_set_state(sc, BACKUP);
1388                         carp_setroute(sc, RTM_DELETE);
1389                         carp_setrun(sc, 0);
1390                 }
1391                 break;
1392         case BACKUP:
1393                 callout_stop(&sc->sc_ad_tmo);
1394                 tv.tv_sec = 3 * sc->sc_advbase;
1395                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1396                 switch (af) {
1397 #ifdef INET
1398                 case AF_INET:
1399                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1400                             carp_master_down, sc);
1401                         break;
1402 #endif /* INET */
1403 #ifdef INET6
1404                 case AF_INET6:
1405                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1406                             carp_master_down, sc);
1407                         break;
1408 #endif /* INET6 */
1409                 default:
1410                         if (sc->sc_naddrs)
1411                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1412                                     carp_master_down, sc);
1413                         if (sc->sc_naddrs6)
1414                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1415                                     carp_master_down, sc);
1416                         break;
1417                 }
1418                 break;
1419         case MASTER:
1420                 tv.tv_sec = sc->sc_advbase;
1421                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1422                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1423                     carp_send_ad, sc);
1424                 break;
1425         }
1426 }
1427
1428 static void
1429 carp_multicast_cleanup(struct carp_softc *sc)
1430 {
1431         struct ip_moptions *imo = &sc->sc_imo;
1432         u_int16_t n = imo->imo_num_memberships;
1433
1434         /* Clean up our own multicast memberships */
1435         while (n-- > 0) {
1436                 if (imo->imo_membership[n] != NULL) {
1437                         in_delmulti(imo->imo_membership[n]);
1438                         imo->imo_membership[n] = NULL;
1439                 }
1440         }
1441         KASSERT(imo->imo_mfilters == NULL,
1442            ("%s: imo_mfilters != NULL", __func__));
1443         imo->imo_num_memberships = 0;
1444         imo->imo_multicast_ifp = NULL;
1445 }
1446
1447 #ifdef INET6
1448 static void
1449 carp_multicast6_cleanup(struct carp_softc *sc)
1450 {
1451         struct ip6_moptions *im6o = &sc->sc_im6o;
1452
1453         while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1454                 struct in6_multi_mship *imm =
1455                     LIST_FIRST(&im6o->im6o_memberships);
1456
1457                 LIST_REMOVE(imm, i6mm_chain);
1458                 in6_leavegroup(imm);
1459         }
1460         im6o->im6o_multicast_ifp = NULL;
1461 }
1462 #endif
1463
1464 static int
1465 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1466 {
1467         INIT_VNET_INET(curvnet);
1468         struct ifnet *ifp;
1469         struct carp_if *cif;
1470         struct in_ifaddr *ia, *ia_if;
1471         struct ip_moptions *imo = &sc->sc_imo;
1472         struct in_addr addr;
1473         u_long iaddr = htonl(sin->sin_addr.s_addr);
1474         int own, error;
1475
1476         if (sin->sin_addr.s_addr == 0) {
1477                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1478                         carp_set_state(sc, INIT);
1479                 if (sc->sc_naddrs)
1480                         SC2IFP(sc)->if_flags |= IFF_UP;
1481                 if (sc->sc_carpdev)
1482                         CARP_SCLOCK(sc);
1483                 carp_setrun(sc, 0);
1484                 if (sc->sc_carpdev)
1485                         CARP_SCUNLOCK(sc);
1486                 return (0);
1487         }
1488
1489         /* we have to do it by hands to check we won't match on us */
1490         ia_if = NULL; own = 0;
1491         TAILQ_FOREACH(ia, &V_in_ifaddrhead, ia_link) {
1492                 /* and, yeah, we need a multicast-capable iface too */
1493                 if (ia->ia_ifp != SC2IFP(sc) &&
1494                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1495                     (iaddr & ia->ia_subnetmask) == ia->ia_subnet) {
1496                         if (!ia_if)
1497                                 ia_if = ia;
1498                         if (sin->sin_addr.s_addr ==
1499                             ia->ia_addr.sin_addr.s_addr)
1500                                 own++;
1501                 }
1502         }
1503
1504         if (!ia_if)
1505                 return (EADDRNOTAVAIL);
1506
1507         ia = ia_if;
1508         ifp = ia->ia_ifp;
1509
1510         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1511             (imo->imo_multicast_ifp && imo->imo_multicast_ifp != ifp))
1512                 return (EADDRNOTAVAIL);
1513
1514         if (imo->imo_num_memberships == 0) {
1515                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1516                 if ((imo->imo_membership[0] = in_addmulti(&addr, ifp)) == NULL)
1517                         return (ENOBUFS);
1518                 imo->imo_num_memberships++;
1519                 imo->imo_multicast_ifp = ifp;
1520                 imo->imo_multicast_ttl = CARP_DFLTTL;
1521                 imo->imo_multicast_loop = 0;
1522         }
1523
1524         if (!ifp->if_carp) {
1525
1526                 cif = malloc(sizeof(*cif), M_CARP,
1527                     M_WAITOK|M_ZERO);
1528                 if (!cif) {
1529                         error = ENOBUFS;
1530                         goto cleanup;
1531                 }
1532                 if ((error = ifpromisc(ifp, 1))) {
1533                         free(cif, M_CARP);
1534                         goto cleanup;
1535                 }
1536
1537                 CARP_LOCK_INIT(cif);
1538                 CARP_LOCK(cif);
1539                 cif->vhif_ifp = ifp;
1540                 TAILQ_INIT(&cif->vhif_vrs);
1541                 ifp->if_carp = cif;
1542
1543         } else {
1544                 struct carp_softc *vr;
1545
1546                 cif = (struct carp_if *)ifp->if_carp;
1547                 CARP_LOCK(cif);
1548                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1549                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1550                                 CARP_UNLOCK(cif);
1551                                 error = EEXIST;
1552                                 goto cleanup;
1553                         }
1554         }
1555         sc->sc_ia = ia;
1556         sc->sc_carpdev = ifp;
1557
1558         { /* XXX prevent endless loop if already in queue */
1559         struct carp_softc *vr, *after = NULL;
1560         int myself = 0;
1561         cif = (struct carp_if *)ifp->if_carp;
1562
1563         /* XXX: cif should not change, right? So we still hold the lock */
1564         CARP_LOCK_ASSERT(cif);
1565
1566         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1567                 if (vr == sc)
1568                         myself = 1;
1569                 if (vr->sc_vhid < sc->sc_vhid)
1570                         after = vr;
1571         }
1572
1573         if (!myself) {
1574                 /* We're trying to keep things in order */
1575                 if (after == NULL) {
1576                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1577                 } else {
1578                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1579                 }
1580                 cif->vhif_nvrs++;
1581         }
1582         }
1583
1584         sc->sc_naddrs++;
1585         SC2IFP(sc)->if_flags |= IFF_UP;
1586         if (own)
1587                 sc->sc_advskew = 0;
1588         carp_sc_state_locked(sc);
1589         carp_setrun(sc, 0);
1590
1591         CARP_UNLOCK(cif);
1592
1593         return (0);
1594
1595 cleanup:
1596         in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1597         return (error);
1598 }
1599
1600 static int
1601 carp_del_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1602 {
1603         int error = 0;
1604
1605         if (!--sc->sc_naddrs) {
1606                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1607                 struct ip_moptions *imo = &sc->sc_imo;
1608
1609                 CARP_LOCK(cif);
1610                 callout_stop(&sc->sc_ad_tmo);
1611                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1612                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1613                 sc->sc_vhid = -1;
1614                 in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1615                 imo->imo_multicast_ifp = NULL;
1616                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1617                 if (!--cif->vhif_nvrs) {
1618                         sc->sc_carpdev->if_carp = NULL;
1619                         CARP_LOCK_DESTROY(cif);
1620                         free(cif, M_IFADDR);
1621                 } else {
1622                         CARP_UNLOCK(cif);
1623                 }
1624         }
1625
1626         return (error);
1627 }
1628
1629 #ifdef INET6
1630 static int
1631 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1632 {
1633         INIT_VNET_INET6(curvnet);
1634         struct ifnet *ifp;
1635         struct carp_if *cif;
1636         struct in6_ifaddr *ia, *ia_if;
1637         struct ip6_moptions *im6o = &sc->sc_im6o;
1638         struct in6_multi_mship *imm;
1639         struct in6_addr in6;
1640         int own, error;
1641
1642         if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
1643                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1644                         carp_set_state(sc, INIT);
1645                 if (sc->sc_naddrs6)
1646                         SC2IFP(sc)->if_flags |= IFF_UP;
1647                 if (sc->sc_carpdev)
1648                         CARP_SCLOCK(sc);
1649                 carp_setrun(sc, 0);
1650                 if (sc->sc_carpdev)
1651                         CARP_SCUNLOCK(sc);
1652                 return (0);
1653         }
1654
1655         /* we have to do it by hands to check we won't match on us */
1656         ia_if = NULL; own = 0;
1657         for (ia = V_in6_ifaddr; ia; ia = ia->ia_next) {
1658                 int i;
1659
1660                 for (i = 0; i < 4; i++) {
1661                         if ((sin6->sin6_addr.s6_addr32[i] &
1662                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]) !=
1663                             (ia->ia_addr.sin6_addr.s6_addr32[i] &
1664                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]))
1665                                 break;
1666                 }
1667                 /* and, yeah, we need a multicast-capable iface too */
1668                 if (ia->ia_ifp != SC2IFP(sc) &&
1669                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1670                     (i == 4)) {
1671                         if (!ia_if)
1672                                 ia_if = ia;
1673                         if (IN6_ARE_ADDR_EQUAL(&sin6->sin6_addr,
1674                             &ia->ia_addr.sin6_addr))
1675                                 own++;
1676                 }
1677         }
1678
1679         if (!ia_if)
1680                 return (EADDRNOTAVAIL);
1681         ia = ia_if;
1682         ifp = ia->ia_ifp;
1683
1684         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1685             (im6o->im6o_multicast_ifp && im6o->im6o_multicast_ifp != ifp))
1686                 return (EADDRNOTAVAIL);
1687
1688         if (!sc->sc_naddrs6) {
1689                 im6o->im6o_multicast_ifp = ifp;
1690
1691                 /* join CARP multicast address */
1692                 bzero(&in6, sizeof(in6));
1693                 in6.s6_addr16[0] = htons(0xff02);
1694                 in6.s6_addr8[15] = 0x12;
1695                 if (in6_setscope(&in6, ifp, NULL) != 0)
1696                         goto cleanup;
1697                 if ((imm = in6_joingroup(ifp, &in6, &error, 0)) == NULL)
1698                         goto cleanup;
1699                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1700
1701                 /* join solicited multicast address */
1702                 bzero(&in6, sizeof(in6));
1703                 in6.s6_addr16[0] = htons(0xff02);
1704                 in6.s6_addr32[1] = 0;
1705                 in6.s6_addr32[2] = htonl(1);
1706                 in6.s6_addr32[3] = sin6->sin6_addr.s6_addr32[3];
1707                 in6.s6_addr8[12] = 0xff;
1708                 if (in6_setscope(&in6, ifp, NULL) != 0)
1709                         goto cleanup;
1710                 if ((imm = in6_joingroup(ifp, &in6, &error, 0)) == NULL)
1711                         goto cleanup;
1712                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1713         }
1714
1715         if (!ifp->if_carp) {
1716                 cif = malloc(sizeof(*cif), M_CARP,
1717                     M_WAITOK|M_ZERO);
1718                 if (!cif) {
1719                         error = ENOBUFS;
1720                         goto cleanup;
1721                 }
1722                 if ((error = ifpromisc(ifp, 1))) {
1723                         free(cif, M_CARP);
1724                         goto cleanup;
1725                 }
1726
1727                 CARP_LOCK_INIT(cif);
1728                 CARP_LOCK(cif);
1729                 cif->vhif_ifp = ifp;
1730                 TAILQ_INIT(&cif->vhif_vrs);
1731                 ifp->if_carp = cif;
1732
1733         } else {
1734                 struct carp_softc *vr;
1735
1736                 cif = (struct carp_if *)ifp->if_carp;
1737                 CARP_LOCK(cif);
1738                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1739                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1740                                 CARP_UNLOCK(cif);
1741                                 error = EINVAL;
1742                                 goto cleanup;
1743                         }
1744         }
1745         sc->sc_ia6 = ia;
1746         sc->sc_carpdev = ifp;
1747
1748         { /* XXX prevent endless loop if already in queue */
1749         struct carp_softc *vr, *after = NULL;
1750         int myself = 0;
1751         cif = (struct carp_if *)ifp->if_carp;
1752         CARP_LOCK_ASSERT(cif);
1753
1754         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1755                 if (vr == sc)
1756                         myself = 1;
1757                 if (vr->sc_vhid < sc->sc_vhid)
1758                         after = vr;
1759         }
1760
1761         if (!myself) {
1762                 /* We're trying to keep things in order */
1763                 if (after == NULL) {
1764                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1765                 } else {
1766                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1767                 }
1768                 cif->vhif_nvrs++;
1769         }
1770         }
1771
1772         sc->sc_naddrs6++;
1773         SC2IFP(sc)->if_flags |= IFF_UP;
1774         if (own)
1775                 sc->sc_advskew = 0;
1776         carp_sc_state_locked(sc);
1777         carp_setrun(sc, 0);
1778
1779         CARP_UNLOCK(cif);
1780
1781         return (0);
1782
1783 cleanup:
1784         /* clean up multicast memberships */
1785         if (!sc->sc_naddrs6) {
1786                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1787                         imm = LIST_FIRST(&im6o->im6o_memberships);
1788                         LIST_REMOVE(imm, i6mm_chain);
1789                         in6_leavegroup(imm);
1790                 }
1791         }
1792         return (error);
1793 }
1794
1795 static int
1796 carp_del_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1797 {
1798         int error = 0;
1799
1800         if (!--sc->sc_naddrs6) {
1801                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1802                 struct ip6_moptions *im6o = &sc->sc_im6o;
1803
1804                 CARP_LOCK(cif);
1805                 callout_stop(&sc->sc_ad_tmo);
1806                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1807                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1808                 sc->sc_vhid = -1;
1809                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1810                         struct in6_multi_mship *imm =
1811                             LIST_FIRST(&im6o->im6o_memberships);
1812
1813                         LIST_REMOVE(imm, i6mm_chain);
1814                         in6_leavegroup(imm);
1815                 }
1816                 im6o->im6o_multicast_ifp = NULL;
1817                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1818                 if (!--cif->vhif_nvrs) {
1819                         CARP_LOCK_DESTROY(cif);
1820                         sc->sc_carpdev->if_carp = NULL;
1821                         free(cif, M_IFADDR);
1822                 } else
1823                         CARP_UNLOCK(cif);
1824         }
1825
1826         return (error);
1827 }
1828 #endif /* INET6 */
1829
1830 static int
1831 carp_ioctl(struct ifnet *ifp, u_long cmd, caddr_t addr)
1832 {
1833         struct carp_softc *sc = ifp->if_softc, *vr;
1834         struct carpreq carpr;
1835         struct ifaddr *ifa;
1836         struct ifreq *ifr;
1837         struct ifaliasreq *ifra;
1838         int locked = 0, error = 0;
1839
1840         ifa = (struct ifaddr *)addr;
1841         ifra = (struct ifaliasreq *)addr;
1842         ifr = (struct ifreq *)addr;
1843
1844         switch (cmd) {
1845         case SIOCSIFADDR:
1846                 switch (ifa->ifa_addr->sa_family) {
1847 #ifdef INET
1848                 case AF_INET:
1849                         SC2IFP(sc)->if_flags |= IFF_UP;
1850                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1851                             sizeof(struct sockaddr));
1852                         error = carp_set_addr(sc, satosin(ifa->ifa_addr));
1853                         break;
1854 #endif /* INET */
1855 #ifdef INET6
1856                 case AF_INET6:
1857                         SC2IFP(sc)->if_flags |= IFF_UP;
1858                         error = carp_set_addr6(sc, satosin6(ifa->ifa_addr));
1859                         break;
1860 #endif /* INET6 */
1861                 default:
1862                         error = EAFNOSUPPORT;
1863                         break;
1864                 }
1865                 break;
1866
1867         case SIOCAIFADDR:
1868                 switch (ifa->ifa_addr->sa_family) {
1869 #ifdef INET
1870                 case AF_INET:
1871                         SC2IFP(sc)->if_flags |= IFF_UP;
1872                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1873                             sizeof(struct sockaddr));
1874                         error = carp_set_addr(sc, satosin(&ifra->ifra_addr));
1875                         break;
1876 #endif /* INET */
1877 #ifdef INET6
1878                 case AF_INET6:
1879                         SC2IFP(sc)->if_flags |= IFF_UP;
1880                         error = carp_set_addr6(sc, satosin6(&ifra->ifra_addr));
1881                         break;
1882 #endif /* INET6 */
1883                 default:
1884                         error = EAFNOSUPPORT;
1885                         break;
1886                 }
1887                 break;
1888
1889         case SIOCDIFADDR:
1890                 switch (ifa->ifa_addr->sa_family) {
1891 #ifdef INET
1892                 case AF_INET:
1893                         error = carp_del_addr(sc, satosin(&ifra->ifra_addr));
1894                         break;
1895 #endif /* INET */
1896 #ifdef INET6
1897                 case AF_INET6:
1898                         error = carp_del_addr6(sc, satosin6(&ifra->ifra_addr));
1899                         break;
1900 #endif /* INET6 */
1901                 default:
1902                         error = EAFNOSUPPORT;
1903                         break;
1904                 }
1905                 break;
1906
1907         case SIOCSIFFLAGS:
1908                 if (sc->sc_carpdev) {
1909                         locked = 1;
1910                         CARP_SCLOCK(sc);
1911                 }
1912                 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) {
1913                         callout_stop(&sc->sc_ad_tmo);
1914                         callout_stop(&sc->sc_md_tmo);
1915                         callout_stop(&sc->sc_md6_tmo);
1916                         if (sc->sc_state == MASTER)
1917                                 carp_send_ad_locked(sc);
1918                         carp_set_state(sc, INIT);
1919                         carp_setrun(sc, 0);
1920                 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) {
1921                         SC2IFP(sc)->if_flags |= IFF_UP;
1922                         carp_setrun(sc, 0);
1923                 }
1924                 break;
1925
1926         case SIOCSVH:
1927                 error = priv_check(curthread, PRIV_NETINET_CARP);
1928                 if (error)
1929                         break;
1930                 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1931                         break;
1932                 error = 1;
1933                 if (sc->sc_carpdev) {
1934                         locked = 1;
1935                         CARP_SCLOCK(sc);
1936                 }
1937                 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) {
1938                         switch (carpr.carpr_state) {
1939                         case BACKUP:
1940                                 callout_stop(&sc->sc_ad_tmo);
1941                                 carp_set_state(sc, BACKUP);
1942                                 carp_setrun(sc, 0);
1943                                 carp_setroute(sc, RTM_DELETE);
1944                                 break;
1945                         case MASTER:
1946                                 carp_master_down_locked(sc);
1947                                 break;
1948                         default:
1949                                 break;
1950                         }
1951                 }
1952                 if (carpr.carpr_vhid > 0) {
1953                         if (carpr.carpr_vhid > 255) {
1954                                 error = EINVAL;
1955                                 break;
1956                         }
1957                         if (sc->sc_carpdev) {
1958                                 struct carp_if *cif;
1959                                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1960                                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1961                                         if (vr != sc &&
1962                                             vr->sc_vhid == carpr.carpr_vhid) {
1963                                                 error = EEXIST;
1964                                                 break;
1965                                         }
1966                                 if (error == EEXIST)
1967                                         break;
1968                         }
1969                         sc->sc_vhid = carpr.carpr_vhid;
1970                         IF_LLADDR(sc->sc_ifp)[0] = 0;
1971                         IF_LLADDR(sc->sc_ifp)[1] = 0;
1972                         IF_LLADDR(sc->sc_ifp)[2] = 0x5e;
1973                         IF_LLADDR(sc->sc_ifp)[3] = 0;
1974                         IF_LLADDR(sc->sc_ifp)[4] = 1;
1975                         IF_LLADDR(sc->sc_ifp)[5] = sc->sc_vhid;
1976                         error--;
1977                 }
1978                 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) {
1979                         if (carpr.carpr_advskew >= 255) {
1980                                 error = EINVAL;
1981                                 break;
1982                         }
1983                         if (carpr.carpr_advbase > 255) {
1984                                 error = EINVAL;
1985                                 break;
1986                         }
1987                         sc->sc_advbase = carpr.carpr_advbase;
1988                         sc->sc_advskew = carpr.carpr_advskew;
1989                         error--;
1990                 }
1991                 bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1992                 if (error > 0)
1993                         error = EINVAL;
1994                 else {
1995                         error = 0;
1996                         carp_setrun(sc, 0);
1997                 }
1998                 break;
1999
2000         case SIOCGVH:
2001                 /* XXX: lockless read */
2002                 bzero(&carpr, sizeof(carpr));
2003                 carpr.carpr_state = sc->sc_state;
2004                 carpr.carpr_vhid = sc->sc_vhid;
2005                 carpr.carpr_advbase = sc->sc_advbase;
2006                 carpr.carpr_advskew = sc->sc_advskew;
2007                 error = priv_check(curthread, PRIV_NETINET_CARP);
2008                 if (error == 0)
2009                         bcopy(sc->sc_key, carpr.carpr_key,
2010                             sizeof(carpr.carpr_key));
2011                 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
2012                 break;
2013
2014         default:
2015                 error = EINVAL;
2016         }
2017
2018         if (locked)
2019                 CARP_SCUNLOCK(sc);
2020
2021         carp_hmac_prepare(sc);
2022
2023         return (error);
2024 }
2025
2026 /*
2027  * XXX: this is looutput. We should eventually use it from there.
2028  */
2029 static int
2030 carp_looutput(struct ifnet *ifp, struct mbuf *m, struct sockaddr *dst,
2031     struct route *ro)
2032 {
2033         u_int32_t af;
2034         struct rtentry *rt = NULL;
2035
2036         M_ASSERTPKTHDR(m); /* check if we have the packet header */
2037
2038         if (ro != NULL)
2039                 rt = ro->ro_rt;
2040         if (rt && rt->rt_flags & (RTF_REJECT|RTF_BLACKHOLE)) {
2041                 m_freem(m);
2042                 return (rt->rt_flags & RTF_BLACKHOLE ? 0 :
2043                         rt->rt_flags & RTF_HOST ? EHOSTUNREACH : ENETUNREACH);
2044         }
2045
2046         ifp->if_opackets++;
2047         ifp->if_obytes += m->m_pkthdr.len;
2048
2049         /* BPF writes need to be handled specially. */
2050         if (dst->sa_family == AF_UNSPEC) {
2051                 bcopy(dst->sa_data, &af, sizeof(af));
2052                 dst->sa_family = af;
2053         }
2054
2055 #if 1   /* XXX */
2056         switch (dst->sa_family) {
2057         case AF_INET:
2058         case AF_INET6:
2059         case AF_IPX:
2060         case AF_APPLETALK:
2061                 break;
2062         default:
2063                 printf("carp_looutput: af=%d unexpected\n", dst->sa_family);
2064                 m_freem(m);
2065                 return (EAFNOSUPPORT);
2066         }
2067 #endif
2068         return(if_simloop(ifp, m, dst->sa_family, 0));
2069 }
2070
2071 /*
2072  * Start output on carp interface. This function should never be called.
2073  */
2074 static void
2075 carp_start(struct ifnet *ifp)
2076 {
2077 #ifdef DEBUG
2078         printf("%s: start called\n", ifp->if_xname);
2079 #endif
2080 }
2081
2082 int
2083 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *sa,
2084     struct rtentry *rt)
2085 {
2086         struct m_tag *mtag;
2087         struct carp_softc *sc;
2088         struct ifnet *carp_ifp;
2089
2090         if (!sa)
2091                 return (0);
2092
2093         switch (sa->sa_family) {
2094 #ifdef INET
2095         case AF_INET:
2096                 break;
2097 #endif /* INET */
2098 #ifdef INET6
2099         case AF_INET6:
2100                 break;
2101 #endif /* INET6 */
2102         default:
2103                 return (0);
2104         }
2105
2106         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
2107         if (mtag == NULL)
2108                 return (0);
2109
2110         bcopy(mtag + 1, &carp_ifp, sizeof(struct ifnet *));
2111         sc = carp_ifp->if_softc;
2112
2113         /* Set the source MAC address to Virtual Router MAC Address */
2114         switch (ifp->if_type) {
2115         case IFT_ETHER:
2116         case IFT_L2VLAN: {
2117                         struct ether_header *eh;
2118
2119                         eh = mtod(m, struct ether_header *);
2120                         eh->ether_shost[0] = 0;
2121                         eh->ether_shost[1] = 0;
2122                         eh->ether_shost[2] = 0x5e;
2123                         eh->ether_shost[3] = 0;
2124                         eh->ether_shost[4] = 1;
2125                         eh->ether_shost[5] = sc->sc_vhid;
2126                 }
2127                 break;
2128         case IFT_FDDI: {
2129                         struct fddi_header *fh;
2130
2131                         fh = mtod(m, struct fddi_header *);
2132                         fh->fddi_shost[0] = 0;
2133                         fh->fddi_shost[1] = 0;
2134                         fh->fddi_shost[2] = 0x5e;
2135                         fh->fddi_shost[3] = 0;
2136                         fh->fddi_shost[4] = 1;
2137                         fh->fddi_shost[5] = sc->sc_vhid;
2138                 }
2139                 break;
2140         case IFT_ISO88025: {
2141                         struct iso88025_header *th;
2142                         th = mtod(m, struct iso88025_header *);
2143                         th->iso88025_shost[0] = 3;
2144                         th->iso88025_shost[1] = 0;
2145                         th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
2146                         th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
2147                         th->iso88025_shost[4] = 0;
2148                         th->iso88025_shost[5] = 0;
2149                 }
2150                 break;
2151         default:
2152                 printf("%s: carp is not supported for this interface type\n",
2153                     ifp->if_xname);
2154                 return (EOPNOTSUPP);
2155         }
2156
2157         return (0);
2158 }
2159
2160 static void
2161 carp_set_state(struct carp_softc *sc, int state)
2162 {
2163         int link_state;
2164
2165         if (sc->sc_carpdev)
2166                 CARP_SCLOCK_ASSERT(sc);
2167
2168         if (sc->sc_state == state)
2169                 return;
2170
2171         sc->sc_state = state;
2172         switch (state) {
2173         case BACKUP:
2174                 link_state = LINK_STATE_DOWN;
2175                 break;
2176         case MASTER:
2177                 link_state = LINK_STATE_UP;
2178                 break;
2179         default:
2180                 link_state = LINK_STATE_UNKNOWN;
2181                 break;
2182         }
2183         if_link_state_change(SC2IFP(sc), link_state);
2184 }
2185
2186 void
2187 carp_carpdev_state(void *v)
2188 {
2189         struct carp_if *cif = v;
2190
2191         CARP_LOCK(cif);
2192         carp_carpdev_state_locked(cif);
2193         CARP_UNLOCK(cif);
2194 }
2195
2196 static void
2197 carp_carpdev_state_locked(struct carp_if *cif)
2198 {
2199         struct carp_softc *sc;
2200
2201         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list)
2202                 carp_sc_state_locked(sc);
2203 }
2204
2205 static void
2206 carp_sc_state_locked(struct carp_softc *sc)
2207 {
2208         CARP_SCLOCK_ASSERT(sc);
2209
2210         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
2211             !(sc->sc_carpdev->if_flags & IFF_UP)) {
2212                 sc->sc_flags_backup = SC2IFP(sc)->if_flags;
2213                 SC2IFP(sc)->if_flags &= ~IFF_UP;
2214                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
2215                 callout_stop(&sc->sc_ad_tmo);
2216                 callout_stop(&sc->sc_md_tmo);
2217                 callout_stop(&sc->sc_md6_tmo);
2218                 carp_set_state(sc, INIT);
2219                 carp_setrun(sc, 0);
2220                 if (!sc->sc_suppress) {
2221                         carp_suppress_preempt++;
2222                         if (carp_suppress_preempt == 1) {
2223                                 CARP_SCUNLOCK(sc);
2224                                 carp_send_ad_all();
2225                                 CARP_SCLOCK(sc);
2226                         }
2227                 }
2228                 sc->sc_suppress = 1;
2229         } else {
2230                 SC2IFP(sc)->if_flags |= sc->sc_flags_backup;
2231                 carp_set_state(sc, INIT);
2232                 carp_setrun(sc, 0);
2233                 if (sc->sc_suppress)
2234                         carp_suppress_preempt--;
2235                 sc->sc_suppress = 0;
2236         }
2237
2238         return;
2239 }
2240
2241 static int
2242 carp_modevent(module_t mod, int type, void *data)
2243 {
2244         switch (type) {
2245         case MOD_LOAD:
2246                 if_detach_event_tag = EVENTHANDLER_REGISTER(ifnet_departure_event,
2247                     carp_ifdetach, NULL, EVENTHANDLER_PRI_ANY);
2248                 if (if_detach_event_tag == NULL)
2249                         return (ENOMEM);
2250                 mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2251                 LIST_INIT(&carpif_list);
2252                 if_clone_attach(&carp_cloner);
2253                 break;
2254
2255         case MOD_UNLOAD:
2256                 EVENTHANDLER_DEREGISTER(ifnet_departure_event, if_detach_event_tag);
2257                 if_clone_detach(&carp_cloner);
2258                 mtx_destroy(&carp_mtx);
2259                 break;
2260
2261         default:
2262                 return (EINVAL);
2263         }
2264
2265         return (0);
2266 }
2267
2268 static moduledata_t carp_mod = {
2269         "carp",
2270         carp_modevent,
2271         0
2272 };
2273
2274 DECLARE_MODULE(carp, carp_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);