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 #endif
  78
  79 #ifdef INET6
  80 #include <netinet/icmp6.h>
  81 #include <netinet/ip6.h>
  82 #include <netinet6/ip6_var.h>
  83 #include <netinet6/scope6_var.h>
  84 #include <netinet6/nd6.h>
  85 #endif
  86
  87 #include <crypto/sha1.h>
  88 #include <netinet/ip_carp.h>
  89
  90 #define CARP_IFNAME     "carp"
  91 static MALLOC_DEFINE(M_CARP, "CARP", "CARP interfaces");
  92 SYSCTL_DECL(_net_inet_carp);
  93
  94 struct carp_softc {
  95         struct ifnet            *sc_ifp;        /* Interface clue */
  96         struct ifnet            *sc_carpdev;    /* Pointer to parent interface */
  97         struct in_ifaddr        *sc_ia;         /* primary iface address */
  98         struct ip_moptions       sc_imo;
  99 #ifdef INET6
 100         struct in6_ifaddr       *sc_ia6;        /* primary iface address v6 */
 101         struct ip6_moptions      sc_im6o;
 102 #endif /* INET6 */
 103         TAILQ_ENTRY(carp_softc)  sc_list;
 104
 105         enum { INIT = 0, BACKUP, MASTER }       sc_state;
 106
 107         int                      sc_flags_backup;
 108         int                      sc_suppress;
 109
 110         int                      sc_sendad_errors;
 111 #define CARP_SENDAD_MAX_ERRORS  3
 112         int                      sc_sendad_success;
 113 #define CARP_SENDAD_MIN_SUCCESS 3
 114
 115         int                      sc_vhid;
 116         int                      sc_advskew;
 117         int                      sc_naddrs;
 118         int                      sc_naddrs6;
 119         int                      sc_advbase;    /* seconds */
 120         int                      sc_init_counter;
 121         u_int64_t                sc_counter;
 122
 123         /* authentication */
 124 #define CARP_HMAC_PAD   64
 125         unsigned char sc_key[CARP_KEY_LEN];
 126         unsigned char sc_pad[CARP_HMAC_PAD];
 127         SHA1_CTX sc_sha1;
 128
 129         struct callout           sc_ad_tmo;     /* advertisement timeout */
 130         struct callout           sc_md_tmo;     /* master down timeout */
 131         struct callout           sc_md6_tmo;    /* master down timeout */
 132
 133         LIST_ENTRY(carp_softc)   sc_next;       /* Interface clue */
 134 };
 135 #define SC2IFP(sc)      ((sc)->sc_ifp)
 136
 137 int carp_suppress_preempt = 0;
 138 int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 1, 0, 0 };    /* XXX for now */
 139 SYSCTL_INT(_net_inet_carp, CARPCTL_ALLOW, allow, CTLFLAG_RW,
 140     &carp_opts[CARPCTL_ALLOW], 0, "Accept incoming CARP packets");
 141 SYSCTL_INT(_net_inet_carp, CARPCTL_PREEMPT, preempt, CTLFLAG_RW,
 142     &carp_opts[CARPCTL_PREEMPT], 0, "high-priority backup preemption mode");
 143 SYSCTL_INT(_net_inet_carp, CARPCTL_LOG, log, CTLFLAG_RW,
 144     &carp_opts[CARPCTL_LOG], 0, "log bad carp packets");
 145 SYSCTL_INT(_net_inet_carp, CARPCTL_ARPBALANCE, arpbalance, CTLFLAG_RW,
 146     &carp_opts[CARPCTL_ARPBALANCE], 0, "balance arp responses");
 147 SYSCTL_INT(_net_inet_carp, OID_AUTO, suppress_preempt, CTLFLAG_RD,
 148     &carp_suppress_preempt, 0, "Preemption is suppressed");
 149
 150 struct carpstats carpstats;
 151 SYSCTL_STRUCT(_net_inet_carp, CARPCTL_STATS, stats, CTLFLAG_RW,
 152     &carpstats, carpstats,
 153     "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 154
 155 struct carp_if {
 156         TAILQ_HEAD(, carp_softc) vhif_vrs;
 157         int vhif_nvrs;
 158
 159         struct ifnet    *vhif_ifp;
 160         struct mtx       vhif_mtx;
 161 };
 162
 163 /* Get carp_if from softc. Valid after carp_set_addr{,6}. */
 164 #define SC2CIF(sc)              ((struct carp_if *)(sc)->sc_carpdev->if_carp)
 165
 166 /* lock per carp_if queue */
 167 #define CARP_LOCK_INIT(cif)     mtx_init(&(cif)->vhif_mtx, "carp_if",   \
 168         NULL, MTX_DEF)
 169 #define CARP_LOCK_DESTROY(cif)  mtx_destroy(&(cif)->vhif_mtx)
 170 #define CARP_LOCK_ASSERT(cif)   mtx_assert(&(cif)->vhif_mtx, MA_OWNED)
 171 #define CARP_LOCK(cif)          mtx_lock(&(cif)->vhif_mtx)
 172 #define CARP_UNLOCK(cif)        mtx_unlock(&(cif)->vhif_mtx)
 173
 174 #define CARP_SCLOCK(sc)         mtx_lock(&SC2CIF(sc)->vhif_mtx)
 175 #define CARP_SCUNLOCK(sc)       mtx_unlock(&SC2CIF(sc)->vhif_mtx)
 176 #define CARP_SCLOCK_ASSERT(sc)  mtx_assert(&SC2CIF(sc)->vhif_mtx, MA_OWNED)
 177
 178 #define CARP_LOG(...)   do {                            \
 179         if (carp_opts[CARPCTL_LOG] > 0)                 \
 180                 log(LOG_INFO, __VA_ARGS__);             \
 181 } while (0)
 182
 183 #define CARP_DEBUG(...) do {                            \
 184         if (carp_opts[CARPCTL_LOG] > 1)                 \
 185                 log(LOG_DEBUG, __VA_ARGS__);            \
 186 } while (0)
 187
 188 static void     carp_hmac_prepare(struct carp_softc *);
 189 static void     carp_hmac_generate(struct carp_softc *, u_int32_t *,
 190                     unsigned char *);
 191 static int      carp_hmac_verify(struct carp_softc *, u_int32_t *,
 192                     unsigned char *);
 193 static void     carp_setroute(struct carp_softc *, int);
 194 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 195 static int      carp_clone_create(struct if_clone *, int, caddr_t);
 196 static void     carp_clone_destroy(struct ifnet *);
 197 static void     carpdetach(struct carp_softc *, int);
 198 static int      carp_prepare_ad(struct mbuf *, struct carp_softc *,
 199                     struct carp_header *);
 200 static void     carp_send_ad_all(void);
 201 static void     carp_send_ad(void *);
 202 static void     carp_send_ad_locked(struct carp_softc *);
 203 static void     carp_send_arp(struct carp_softc *);
 204 static void     carp_master_down(void *);
 205 static void     carp_master_down_locked(struct carp_softc *);
 206 static int      carp_ioctl(struct ifnet *, u_long, caddr_t);
 207 static int      carp_looutput(struct ifnet *, struct mbuf *, struct sockaddr *,
 208                     struct rtentry *);
 209 static void     carp_start(struct ifnet *);
 210 static void     carp_setrun(struct carp_softc *, sa_family_t);
 211 static void     carp_set_state(struct carp_softc *, int);
 212 static int      carp_addrcount(struct carp_if *, struct in_ifaddr *, int);
 213 enum    { CARP_COUNT_MASTER, CARP_COUNT_RUNNING };
 214
 215 static void     carp_multicast_cleanup(struct carp_softc *);
 216 static int      carp_set_addr(struct carp_softc *, struct sockaddr_in *);
 217 static int      carp_del_addr(struct carp_softc *, struct sockaddr_in *);
 218 static void     carp_carpdev_state_locked(struct carp_if *);
 219 static void     carp_sc_state_locked(struct carp_softc *);
 220 #ifdef INET6
 221 static void     carp_send_na(struct carp_softc *);
 222 static int      carp_set_addr6(struct carp_softc *, struct sockaddr_in6 *);
 223 static int      carp_del_addr6(struct carp_softc *, struct sockaddr_in6 *);
 224 static void     carp_multicast6_cleanup(struct carp_softc *);
 225 #endif
 226
 227 static LIST_HEAD(, carp_softc) carpif_list;
 228 static struct mtx carp_mtx;
 229 IFC_SIMPLE_DECLARE(carp, 0);
 230
 231 static eventhandler_tag if_detach_event_tag;
 232
 233 static __inline u_int16_t
 234 carp_cksum(struct mbuf *m, int len)
 235 {
 236         return (in_cksum(m, len));
 237 }
 238
 239 static void
 240 carp_hmac_prepare(struct carp_softc *sc)
 241 {
 242         u_int8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 243         u_int8_t vhid = sc->sc_vhid & 0xff;
 244         struct ifaddr *ifa;
 245         int i, found;
 246 #ifdef INET
 247         struct in_addr last, cur, in;
 248 #endif
 249 #ifdef INET6
 250         struct in6_addr last6, cur6, in6;
 251 #endif
 252
 253         if (sc->sc_carpdev)
 254                 CARP_SCLOCK(sc);
 255
 256         /* XXX: possible race here */
 257
 258         /* compute ipad from key */
 259         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 260         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 261         for (i = 0; i < sizeof(sc->sc_pad); i++)
 262                 sc->sc_pad[i] ^= 0x36;
 263
 264         /* precompute first part of inner hash */
 265         SHA1Init(&sc->sc_sha1);
 266         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 267         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 268         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 269         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 270 #ifdef INET
 271         cur.s_addr = 0;
 272         do {
 273                 found = 0;
 274                 last = cur;
 275                 cur.s_addr = 0xffffffff;
 276                 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 277                         in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 278                         if (ifa->ifa_addr->sa_family == AF_INET &&
 279                             ntohl(in.s_addr) > ntohl(last.s_addr) &&
 280                             ntohl(in.s_addr) < ntohl(cur.s_addr)) {
 281                                 cur.s_addr = in.s_addr;
 282                                 found++;
 283                         }
 284                 }
 285                 if (found)
 286                         SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur));
 287         } while (found);
 288 #endif /* INET */
 289 #ifdef INET6
 290         memset(&cur6, 0, sizeof(cur6));
 291         do {
 292                 found = 0;
 293                 last6 = cur6;
 294                 memset(&cur6, 0xff, sizeof(cur6));
 295                 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 296                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 297                         if (IN6_IS_SCOPE_EMBED(&in6))
 298                                 in6.s6_addr16[1] = 0;
 299                         if (ifa->ifa_addr->sa_family == AF_INET6 &&
 300                             memcmp(&in6, &last6, sizeof(in6)) > 0 &&
 301                             memcmp(&in6, &cur6, sizeof(in6)) < 0) {
 302                                 cur6 = in6;
 303                                 found++;
 304                         }
 305                 }
 306                 if (found)
 307                         SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6));
 308         } while (found);
 309 #endif /* INET6 */
 310
 311         /* convert ipad to opad */
 312         for (i = 0; i < sizeof(sc->sc_pad); i++)
 313                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 314
 315         if (sc->sc_carpdev)
 316                 CARP_SCUNLOCK(sc);
 317 }
 318
 319 static void
 320 carp_hmac_generate(struct carp_softc *sc, u_int32_t counter[2],
 321     unsigned char md[20])
 322 {
 323         SHA1_CTX sha1ctx;
 324
 325         /* fetch first half of inner hash */
 326         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 327
 328         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 329         SHA1Final(md, &sha1ctx);
 330
 331         /* outer hash */
 332         SHA1Init(&sha1ctx);
 333         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 334         SHA1Update(&sha1ctx, md, 20);
 335         SHA1Final(md, &sha1ctx);
 336 }
 337
 338 static int
 339 carp_hmac_verify(struct carp_softc *sc, u_int32_t counter[2],
 340     unsigned char md[20])
 341 {
 342         unsigned char md2[20];
 343
 344         CARP_SCLOCK_ASSERT(sc);
 345
 346         carp_hmac_generate(sc, counter, md2);
 347
 348         return (bcmp(md, md2, sizeof(md2)));
 349 }
 350
 351 static void
 352 carp_setroute(struct carp_softc *sc, int cmd)
 353 {
 354         struct ifaddr *ifa;
 355         int s;
 356
 357         if (sc->sc_carpdev)
 358                 CARP_SCLOCK_ASSERT(sc);
 359
 360         s = splnet();
 361         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
 362                 if (ifa->ifa_addr->sa_family == AF_INET &&
 363                     sc->sc_carpdev != NULL) {
 364                         int count = carp_addrcount(
 365                             (struct carp_if *)sc->sc_carpdev->if_carp,
 366                             ifatoia(ifa), CARP_COUNT_MASTER);
 367
 368                         if ((cmd == RTM_ADD && count == 1) ||
 369                             (cmd == RTM_DELETE && count == 0))
 370                                 rtinit(ifa, cmd, RTF_UP | RTF_HOST);
 371                 }
 372 #ifdef INET6
 373                 if (ifa->ifa_addr->sa_family == AF_INET6) {
 374                         if (cmd == RTM_ADD)
 375                                 in6_ifaddloop(ifa);
 376                         else
 377                                 in6_ifremloop(ifa);
 378                 }
 379 #endif /* INET6 */
 380         }
 381         splx(s);
 382 }
 383
 384 static int
 385 carp_clone_create(struct if_clone *ifc, int unit, caddr_t params)
 386 {
 387
 388         struct carp_softc *sc;
 389         struct ifnet *ifp;
 390
 391         MALLOC(sc, struct carp_softc *, sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
 392         ifp = SC2IFP(sc) = if_alloc(IFT_ETHER);
 393         if (ifp == NULL) {
 394                 FREE(sc, M_CARP);
 395                 return (ENOSPC);
 396         }
 397
 398         sc->sc_flags_backup = 0;
 399         sc->sc_suppress = 0;
 400         sc->sc_advbase = CARP_DFLTINTV;
 401         sc->sc_vhid = -1;       /* required setting */
 402         sc->sc_advskew = 0;
 403         sc->sc_init_counter = 1;
 404         sc->sc_naddrs = sc->sc_naddrs6 = 0; /* M_ZERO? */
 405 #ifdef INET6
 406         sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL;
 407 #endif
 408         sc->sc_imo.imo_membership = (struct in_multi **)malloc(
 409             (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
 410             M_WAITOK);
 411         sc->sc_imo.imo_mfilters = NULL;
 412         sc->sc_imo.imo_max_memberships = IP_MIN_MEMBERSHIPS;
 413         sc->sc_imo.imo_multicast_vif = -1;
 414
 415         callout_init(&sc->sc_ad_tmo, CALLOUT_MPSAFE);
 416         callout_init(&sc->sc_md_tmo, CALLOUT_MPSAFE);
 417         callout_init(&sc->sc_md6_tmo, CALLOUT_MPSAFE);
 418
 419         ifp->if_softc = sc;
 420         if_initname(ifp, CARP_IFNAME, unit);
 421         ifp->if_mtu = ETHERMTU;
 422         ifp->if_flags = IFF_LOOPBACK;
 423         ifp->if_ioctl = carp_ioctl;
 424         ifp->if_output = carp_looutput;
 425         ifp->if_start = carp_start;
 426         ifp->if_type = IFT_CARP;
 427         ifp->if_snd.ifq_maxlen = ifqmaxlen;
 428         ifp->if_hdrlen = 0;
 429         if_attach(ifp);
 430         bpfattach(SC2IFP(sc), DLT_NULL, sizeof(u_int32_t));
 431         mtx_lock(&carp_mtx);
 432         LIST_INSERT_HEAD(&carpif_list, sc, sc_next);
 433         mtx_unlock(&carp_mtx);
 434         return (0);
 435 }
 436
 437 static void
 438 carp_clone_destroy(struct ifnet *ifp)
 439 {
 440         struct carp_softc *sc = ifp->if_softc;
 441
 442         if (sc->sc_carpdev)
 443                 CARP_SCLOCK(sc);
 444         carpdetach(sc, 1);      /* Returns unlocked. */
 445
 446         mtx_lock(&carp_mtx);
 447         LIST_REMOVE(sc, sc_next);
 448         mtx_unlock(&carp_mtx);
 449         bpfdetach(ifp);
 450         if_detach(ifp);
 451         if_free_type(ifp, IFT_ETHER);
 452         free(sc->sc_imo.imo_membership, M_CARP);
 453         free(sc, M_CARP);
 454 }
 455
 456 /*
 457  * This function can be called on CARP interface destroy path,
 458  * and in case of the removal of the underlying interface as
 459  * well. We differentiate these two cases. In the latter case
 460  * we do not cleanup our multicast memberships, since they
 461  * are already freed. Also, in the latter case we do not
 462  * release the lock on return, because the function will be
 463  * called once more, for another CARP instance on the same
 464  * interface.
 465  */
 466 static void
 467 carpdetach(struct carp_softc *sc, int unlock)
 468 {
 469         struct carp_if *cif;
 470
 471         callout_stop(&sc->sc_ad_tmo);
 472         callout_stop(&sc->sc_md_tmo);
 473         callout_stop(&sc->sc_md6_tmo);
 474
 475         if (sc->sc_suppress)
 476                 carp_suppress_preempt--;
 477         sc->sc_suppress = 0;
 478
 479         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS)
 480                 carp_suppress_preempt--;
 481         sc->sc_sendad_errors = 0;
 482
 483         carp_set_state(sc, INIT);
 484         SC2IFP(sc)->if_flags &= ~IFF_UP;
 485         carp_setrun(sc, 0);
 486         if (unlock)
 487                 carp_multicast_cleanup(sc);
 488 #ifdef INET6
 489         carp_multicast6_cleanup(sc);
 490 #endif
 491
 492         if (sc->sc_carpdev != NULL) {
 493                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
 494                 CARP_LOCK_ASSERT(cif);
 495                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
 496                 if (!--cif->vhif_nvrs) {
 497                         ifpromisc(sc->sc_carpdev, 0);
 498                         sc->sc_carpdev->if_carp = NULL;
 499                         CARP_LOCK_DESTROY(cif);
 500                         FREE(cif, M_IFADDR);
 501                 } else if (unlock)
 502                         CARP_UNLOCK(cif);
 503                 sc->sc_carpdev = NULL;
 504         }
 505 }
 506
 507 /* Detach an interface from the carp. */
 508 static void
 509 carp_ifdetach(void *arg __unused, struct ifnet *ifp)
 510 {
 511         struct carp_if *cif = (struct carp_if *)ifp->if_carp;
 512         struct carp_softc *sc, *nextsc;
 513
 514         if (cif == NULL)
 515                 return;
 516
 517         /*
 518          * XXX: At the end of for() cycle the lock will be destroyed.
 519          */
 520         CARP_LOCK(cif);
 521         for (sc = TAILQ_FIRST(&cif->vhif_vrs); sc; sc = nextsc) {
 522                 nextsc = TAILQ_NEXT(sc, sc_list);
 523                 carpdetach(sc, 0);
 524         }
 525 }
 526
 527 /*
 528  * process input packet.
 529  * we have rearranged checks order compared to the rfc,
 530  * but it seems more efficient this way or not possible otherwise.
 531  */
 532 void
 533 carp_input(struct mbuf *m, int hlen)
 534 {
 535         struct ip *ip = mtod(m, struct ip *);
 536         struct carp_header *ch;
 537         int iplen, len;
 538
 539         carpstats.carps_ipackets++;
 540
 541         if (!carp_opts[CARPCTL_ALLOW]) {
 542                 m_freem(m);
 543                 return;
 544         }
 545
 546         /* check if received on a valid carp interface */
 547         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 548                 carpstats.carps_badif++;
 549                 CARP_LOG("carp_input: packet received on non-carp "
 550                     "interface: %s\n",
 551                     m->m_pkthdr.rcvif->if_xname);
 552                 m_freem(m);
 553                 return;
 554         }
 555
 556         /* verify that the IP TTL is 255.  */
 557         if (ip->ip_ttl != CARP_DFLTTL) {
 558                 carpstats.carps_badttl++;
 559                 CARP_LOG("carp_input: received ttl %d != 255i on %s\n",
 560                     ip->ip_ttl,
 561                     m->m_pkthdr.rcvif->if_xname);
 562                 m_freem(m);
 563                 return;
 564         }
 565
 566         iplen = ip->ip_hl << 2;
 567
 568         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 569                 carpstats.carps_badlen++;
 570                 CARP_LOG("carp_input: received len %zd < "
 571                     "sizeof(struct carp_header)\n",
 572                     m->m_len - sizeof(struct ip));
 573                 m_freem(m);
 574                 return;
 575         }
 576
 577         if (iplen + sizeof(*ch) < m->m_len) {
 578                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 579                         carpstats.carps_hdrops++;
 580                         CARP_LOG("carp_input: pullup failed\n");
 581                         return;
 582                 }
 583                 ip = mtod(m, struct ip *);
 584         }
 585         ch = (struct carp_header *)((char *)ip + iplen);
 586
 587         /*
 588          * verify that the received packet length is
 589          * equal to the CARP header
 590          */
 591         len = iplen + sizeof(*ch);
 592         if (len > m->m_pkthdr.len) {
 593                 carpstats.carps_badlen++;
 594                 CARP_LOG("carp_input: packet too short %d on %s\n",
 595                     m->m_pkthdr.len,
 596                     m->m_pkthdr.rcvif->if_xname);
 597                 m_freem(m);
 598                 return;
 599         }
 600
 601         if ((m = m_pullup(m, len)) == NULL) {
 602                 carpstats.carps_hdrops++;
 603                 return;
 604         }
 605         ip = mtod(m, struct ip *);
 606         ch = (struct carp_header *)((char *)ip + iplen);
 607
 608         /* verify the CARP checksum */
 609         m->m_data += iplen;
 610         if (carp_cksum(m, len - iplen)) {
 611                 carpstats.carps_badsum++;
 612                 CARP_LOG("carp_input: checksum failed on %s\n",
 613                     m->m_pkthdr.rcvif->if_xname);
 614                 m_freem(m);
 615                 return;
 616         }
 617         m->m_data -= iplen;
 618
 619         carp_input_c(m, ch, AF_INET);
 620 }
 621
 622 #ifdef INET6
 623 int
 624 carp6_input(struct mbuf **mp, int *offp, int proto)
 625 {
 626         struct mbuf *m = *mp;
 627         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 628         struct carp_header *ch;
 629         u_int len;
 630
 631         carpstats.carps_ipackets6++;
 632
 633         if (!carp_opts[CARPCTL_ALLOW]) {
 634                 m_freem(m);
 635                 return (IPPROTO_DONE);
 636         }
 637
 638         /* check if received on a valid carp interface */
 639         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 640                 carpstats.carps_badif++;
 641                 CARP_LOG("carp6_input: packet received on non-carp "
 642                     "interface: %s\n",
 643                     m->m_pkthdr.rcvif->if_xname);
 644                 m_freem(m);
 645                 return (IPPROTO_DONE);
 646         }
 647
 648         /* verify that the IP TTL is 255 */
 649         if (ip6->ip6_hlim != CARP_DFLTTL) {
 650                 carpstats.carps_badttl++;
 651                 CARP_LOG("carp6_input: received ttl %d != 255 on %s\n",
 652                     ip6->ip6_hlim,
 653                     m->m_pkthdr.rcvif->if_xname);
 654                 m_freem(m);
 655                 return (IPPROTO_DONE);
 656         }
 657
 658         /* verify that we have a complete carp packet */
 659         len = m->m_len;
 660         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 661         if (ch == NULL) {
 662                 carpstats.carps_badlen++;
 663                 CARP_LOG("carp6_input: packet size %u too small\n", len);
 664                 return (IPPROTO_DONE);
 665         }
 666
 667
 668         /* verify the CARP checksum */
 669         m->m_data += *offp;
 670         if (carp_cksum(m, sizeof(*ch))) {
 671                 carpstats.carps_badsum++;
 672                 CARP_LOG("carp6_input: checksum failed, on %s\n",
 673                     m->m_pkthdr.rcvif->if_xname);
 674                 m_freem(m);
 675                 return (IPPROTO_DONE);
 676         }
 677         m->m_data -= *offp;
 678
 679         carp_input_c(m, ch, AF_INET6);
 680         return (IPPROTO_DONE);
 681 }
 682 #endif /* INET6 */
 683
 684 static void
 685 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 686 {
 687         struct ifnet *ifp = m->m_pkthdr.rcvif;
 688         struct carp_softc *sc;
 689         u_int64_t tmp_counter;
 690         struct timeval sc_tv, ch_tv;
 691
 692         /* verify that the VHID is valid on the receiving interface */
 693         CARP_LOCK(ifp->if_carp);
 694         TAILQ_FOREACH(sc, &((struct carp_if *)ifp->if_carp)->vhif_vrs, sc_list)
 695                 if (sc->sc_vhid == ch->carp_vhid)
 696                         break;
 697
 698         if (!sc || !((SC2IFP(sc)->if_flags & IFF_UP) &&
 699             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING))) {
 700                 carpstats.carps_badvhid++;
 701                 CARP_UNLOCK(ifp->if_carp);
 702                 m_freem(m);
 703                 return;
 704         }
 705
 706         getmicrotime(&SC2IFP(sc)->if_lastchange);
 707         SC2IFP(sc)->if_ipackets++;
 708         SC2IFP(sc)->if_ibytes += m->m_pkthdr.len;
 709
 710         if (bpf_peers_present(SC2IFP(sc)->if_bpf)) {
 711                 struct ip *ip = mtod(m, struct ip *);
 712                 uint32_t af1 = af;
 713
 714                 /* BPF wants net byte order */
 715                 ip->ip_len = htons(ip->ip_len + (ip->ip_hl << 2));
 716                 ip->ip_off = htons(ip->ip_off);
 717                 bpf_mtap2(SC2IFP(sc)->if_bpf, &af1, sizeof(af1), m);
 718         }
 719
 720         /* verify the CARP version. */
 721         if (ch->carp_version != CARP_VERSION) {
 722                 carpstats.carps_badver++;
 723                 SC2IFP(sc)->if_ierrors++;
 724                 CARP_UNLOCK(ifp->if_carp);
 725                 CARP_LOG("%s; invalid version %d\n",
 726                     SC2IFP(sc)->if_xname,
 727                     ch->carp_version);
 728                 m_freem(m);
 729                 return;
 730         }
 731
 732         /* verify the hash */
 733         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 734                 carpstats.carps_badauth++;
 735                 SC2IFP(sc)->if_ierrors++;
 736                 CARP_UNLOCK(ifp->if_carp);
 737                 CARP_LOG("%s: incorrect hash\n", SC2IFP(sc)->if_xname);
 738                 m_freem(m);
 739                 return;
 740         }
 741
 742         tmp_counter = ntohl(ch->carp_counter[0]);
 743         tmp_counter = tmp_counter<<32;
 744         tmp_counter += ntohl(ch->carp_counter[1]);
 745
 746         /* XXX Replay protection goes here */
 747
 748         sc->sc_init_counter = 0;
 749         sc->sc_counter = tmp_counter;
 750
 751         sc_tv.tv_sec = sc->sc_advbase;
 752         if (carp_suppress_preempt && sc->sc_advskew <  240)
 753                 sc_tv.tv_usec = 240 * 1000000 / 256;
 754         else
 755                 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256;
 756         ch_tv.tv_sec = ch->carp_advbase;
 757         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 758
 759         switch (sc->sc_state) {
 760         case INIT:
 761                 break;
 762         case MASTER:
 763                 /*
 764                  * If we receive an advertisement from a master who's going to
 765                  * be more frequent than us, go into BACKUP state.
 766                  */
 767                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 768                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 769                         callout_stop(&sc->sc_ad_tmo);
 770                         CARP_DEBUG("%s: MASTER -> BACKUP "
 771                            "(more frequent advertisement received)\n",
 772                            SC2IFP(sc)->if_xname);
 773                         carp_set_state(sc, BACKUP);
 774                         carp_setrun(sc, 0);
 775                         carp_setroute(sc, RTM_DELETE);
 776                 }
 777                 break;
 778         case BACKUP:
 779                 /*
 780                  * If we're pre-empting masters who advertise slower than us,
 781                  * and this one claims to be slower, treat him as down.
 782                  */
 783                 if (carp_opts[CARPCTL_PREEMPT] &&
 784                     timevalcmp(&sc_tv, &ch_tv, <)) {
 785                         CARP_DEBUG("%s: BACKUP -> MASTER "
 786                             "(preempting a slower master)\n",
 787                             SC2IFP(sc)->if_xname);
 788                         carp_master_down_locked(sc);
 789                         break;
 790                 }
 791
 792                 /*
 793                  *  If the master is going to advertise at such a low frequency
 794                  *  that he's guaranteed to time out, we'd might as well just
 795                  *  treat him as timed out now.
 796                  */
 797                 sc_tv.tv_sec = sc->sc_advbase * 3;
 798                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 799                         CARP_DEBUG("%s: BACKUP -> MASTER "
 800                             "(master timed out)\n",
 801                             SC2IFP(sc)->if_xname);
 802                         carp_master_down_locked(sc);
 803                         break;
 804                 }
 805
 806                 /*
 807                  * Otherwise, we reset the counter and wait for the next
 808                  * advertisement.
 809                  */
 810                 carp_setrun(sc, af);
 811                 break;
 812         }
 813
 814         CARP_UNLOCK(ifp->if_carp);
 815
 816         m_freem(m);
 817         return;
 818 }
 819
 820 static int
 821 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 822 {
 823         struct m_tag *mtag;
 824         struct ifnet *ifp = SC2IFP(sc);
 825
 826         if (sc->sc_init_counter) {
 827                 /* this could also be seconds since unix epoch */
 828                 sc->sc_counter = arc4random();
 829                 sc->sc_counter = sc->sc_counter << 32;
 830                 sc->sc_counter += arc4random();
 831         } else
 832                 sc->sc_counter++;
 833
 834         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 835         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 836
 837         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 838
 839         /* Tag packet for carp_output */
 840         mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct ifnet *), M_NOWAIT);
 841         if (mtag == NULL) {
 842                 m_freem(m);
 843                 SC2IFP(sc)->if_oerrors++;
 844                 return (ENOMEM);
 845         }
 846         bcopy(&ifp, (caddr_t)(mtag + 1), sizeof(struct ifnet *));
 847         m_tag_prepend(m, mtag);
 848
 849         return (0);
 850 }
 851
 852 static void
 853 carp_send_ad_all(void)
 854 {
 855         struct carp_softc *sc;
 856
 857         mtx_lock(&carp_mtx);
 858         LIST_FOREACH(sc, &carpif_list, sc_next) {
 859                 if (sc->sc_carpdev == NULL)
 860                         continue;
 861                 CARP_SCLOCK(sc);
 862                 if ((SC2IFP(sc)->if_flags & IFF_UP) &&
 863                     (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING) &&
 864                      sc->sc_state == MASTER)
 865                         carp_send_ad_locked(sc);
 866                 CARP_SCUNLOCK(sc);
 867         }
 868         mtx_unlock(&carp_mtx);
 869 }
 870
 871 static void
 872 carp_send_ad(void *v)
 873 {
 874         struct carp_softc *sc = v;
 875
 876         CARP_SCLOCK(sc);
 877         carp_send_ad_locked(sc);
 878         CARP_SCUNLOCK(sc);
 879 }
 880
 881 static void
 882 carp_send_ad_locked(struct carp_softc *sc)
 883 {
 884         struct carp_header ch;
 885         struct timeval tv;
 886         struct carp_header *ch_ptr;
 887         struct mbuf *m;
 888         int len, advbase, advskew;
 889
 890         CARP_SCLOCK_ASSERT(sc);
 891
 892         /* bow out if we've lost our UPness or RUNNINGuiness */
 893         if (!((SC2IFP(sc)->if_flags & IFF_UP) &&
 894             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING))) {
 895                 advbase = 255;
 896                 advskew = 255;
 897         } else {
 898                 advbase = sc->sc_advbase;
 899                 if (!carp_suppress_preempt || sc->sc_advskew > 240)
 900                         advskew = sc->sc_advskew;
 901                 else
 902                         advskew = 240;
 903                 tv.tv_sec = advbase;
 904                 tv.tv_usec = advskew * 1000000 / 256;
 905         }
 906
 907         ch.carp_version = CARP_VERSION;
 908         ch.carp_type = CARP_ADVERTISEMENT;
 909         ch.carp_vhid = sc->sc_vhid;
 910         ch.carp_advbase = advbase;
 911         ch.carp_advskew = advskew;
 912         ch.carp_authlen = 7;    /* XXX DEFINE */
 913         ch.carp_pad1 = 0;       /* must be zero */
 914         ch.carp_cksum = 0;
 915
 916 #ifdef INET
 917         INIT_VNET_INET(curvnet);
 918         if (sc->sc_ia) {
 919                 struct ip *ip;
 920
 921                 MGETHDR(m, M_DONTWAIT, MT_HEADER);
 922                 if (m == NULL) {
 923                         SC2IFP(sc)->if_oerrors++;
 924                         carpstats.carps_onomem++;
 925                         /* XXX maybe less ? */
 926                         if (advbase != 255 || advskew != 255)
 927                                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
 928                                     carp_send_ad, sc);
 929                         return;
 930                 }
 931                 len = sizeof(*ip) + sizeof(ch);
 932                 m->m_pkthdr.len = len;
 933                 m->m_pkthdr.rcvif = NULL;
 934                 m->m_len = len;
 935                 MH_ALIGN(m, m->m_len);
 936                 m->m_flags |= M_MCAST;
 937                 ip = mtod(m, struct ip *);
 938                 ip->ip_v = IPVERSION;
 939                 ip->ip_hl = sizeof(*ip) >> 2;
 940                 ip->ip_tos = IPTOS_LOWDELAY;
 941                 ip->ip_len = len;
 942                 ip->ip_id = ip_newid();
 943                 ip->ip_off = IP_DF;
 944                 ip->ip_ttl = CARP_DFLTTL;
 945                 ip->ip_p = IPPROTO_CARP;
 946                 ip->ip_sum = 0;
 947                 ip->ip_src.s_addr = sc->sc_ia->ia_addr.sin_addr.s_addr;
 948                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 949
 950                 ch_ptr = (struct carp_header *)(&ip[1]);
 951                 bcopy(&ch, ch_ptr, sizeof(ch));
 952                 if (carp_prepare_ad(m, sc, ch_ptr))
 953                         return;
 954
 955                 m->m_data += sizeof(*ip);
 956                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip));
 957                 m->m_data -= sizeof(*ip);
 958
 959                 getmicrotime(&SC2IFP(sc)->if_lastchange);
 960                 SC2IFP(sc)->if_opackets++;
 961                 SC2IFP(sc)->if_obytes += len;
 962                 carpstats.carps_opackets++;
 963
 964                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, NULL)) {
 965                         SC2IFP(sc)->if_oerrors++;
 966                         if (sc->sc_sendad_errors < INT_MAX)
 967                                 sc->sc_sendad_errors++;
 968                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 969                                 carp_suppress_preempt++;
 970                                 if (carp_suppress_preempt == 1) {
 971                                         CARP_SCUNLOCK(sc);
 972                                         carp_send_ad_all();
 973                                         CARP_SCLOCK(sc);
 974                                 }
 975                         }
 976                         sc->sc_sendad_success = 0;
 977                 } else {
 978                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 979                                 if (++sc->sc_sendad_success >=
 980                                     CARP_SENDAD_MIN_SUCCESS) {
 981                                         carp_suppress_preempt--;
 982                                         sc->sc_sendad_errors = 0;
 983                                 }
 984                         } else
 985                                 sc->sc_sendad_errors = 0;
 986                 }
 987         }
 988 #endif /* INET */
 989 #ifdef INET6
 990         if (sc->sc_ia6) {
 991                 struct ip6_hdr *ip6;
 992
 993                 MGETHDR(m, M_DONTWAIT, MT_HEADER);
 994                 if (m == NULL) {
 995                         SC2IFP(sc)->if_oerrors++;
 996                         carpstats.carps_onomem++;
 997                         /* XXX maybe less ? */
 998                         if (advbase != 255 || advskew != 255)
 999                                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1000                                     carp_send_ad, sc);
1001                         return;
1002                 }
1003                 len = sizeof(*ip6) + sizeof(ch);
1004                 m->m_pkthdr.len = len;
1005                 m->m_pkthdr.rcvif = NULL;
1006                 m->m_len = len;
1007                 MH_ALIGN(m, m->m_len);
1008                 m->m_flags |= M_MCAST;
1009                 ip6 = mtod(m, struct ip6_hdr *);
1010                 bzero(ip6, sizeof(*ip6));
1011                 ip6->ip6_vfc |= IPV6_VERSION;
1012                 ip6->ip6_hlim = CARP_DFLTTL;
1013                 ip6->ip6_nxt = IPPROTO_CARP;
1014                 bcopy(&sc->sc_ia6->ia_addr.sin6_addr, &ip6->ip6_src,
1015                     sizeof(struct in6_addr));
1016                 /* set the multicast destination */
1017
1018                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
1019                 ip6->ip6_dst.s6_addr8[15] = 0x12;
1020                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
1021                         SC2IFP(sc)->if_oerrors++;
1022                         m_freem(m);
1023                         CARP_LOG("%s: in6_setscope failed\n", __func__);
1024                         return;
1025                 }
1026
1027                 ch_ptr = (struct carp_header *)(&ip6[1]);
1028                 bcopy(&ch, ch_ptr, sizeof(ch));
1029                 if (carp_prepare_ad(m, sc, ch_ptr))
1030                         return;
1031
1032                 m->m_data += sizeof(*ip6);
1033                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6));
1034                 m->m_data -= sizeof(*ip6);
1035
1036                 getmicrotime(&SC2IFP(sc)->if_lastchange);
1037                 SC2IFP(sc)->if_opackets++;
1038                 SC2IFP(sc)->if_obytes += len;
1039                 carpstats.carps_opackets6++;
1040
1041                 if (ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL)) {
1042                         SC2IFP(sc)->if_oerrors++;
1043                         if (sc->sc_sendad_errors < INT_MAX)
1044                                 sc->sc_sendad_errors++;
1045                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
1046                                 carp_suppress_preempt++;
1047                                 if (carp_suppress_preempt == 1) {
1048                                         CARP_SCUNLOCK(sc);
1049                                         carp_send_ad_all();
1050                                         CARP_SCLOCK(sc);
1051                                 }
1052                         }
1053                         sc->sc_sendad_success = 0;
1054                 } else {
1055                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
1056                                 if (++sc->sc_sendad_success >=
1057                                     CARP_SENDAD_MIN_SUCCESS) {
1058                                         carp_suppress_preempt--;
1059                                         sc->sc_sendad_errors = 0;
1060                                 }
1061                         } else
1062                                 sc->sc_sendad_errors = 0;
1063                 }
1064         }
1065 #endif /* INET6 */
1066
1067         if (advbase != 255 || advskew != 255)
1068                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1069                     carp_send_ad, sc);
1070
1071 }
1072
1073 /*
1074  * Broadcast a gratuitous ARP request containing
1075  * the virtual router MAC address for each IP address
1076  * associated with the virtual router.
1077  */
1078 static void
1079 carp_send_arp(struct carp_softc *sc)
1080 {
1081         struct ifaddr *ifa;
1082
1083         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1084
1085                 if (ifa->ifa_addr->sa_family != AF_INET)
1086                         continue;
1087
1088 /*              arprequest(sc->sc_carpdev, &in, &in, IF_LLADDR(sc->sc_ifp)); */
1089                 arp_ifinit2(sc->sc_carpdev, ifa, IF_LLADDR(sc->sc_ifp));
1090
1091                 DELAY(1000);    /* XXX */
1092         }
1093 }
1094
1095 #ifdef INET6
1096 static void
1097 carp_send_na(struct carp_softc *sc)
1098 {
1099         struct ifaddr *ifa;
1100         struct in6_addr *in6;
1101         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1102
1103         TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1104
1105                 if (ifa->ifa_addr->sa_family != AF_INET6)
1106                         continue;
1107
1108                 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr;
1109                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1110                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1111                 DELAY(1000);    /* XXX */
1112         }
1113 }
1114 #endif /* INET6 */
1115
1116 static int
1117 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type)
1118 {
1119         struct carp_softc *vh;
1120         struct ifaddr *ifa;
1121         int count = 0;
1122
1123         CARP_LOCK_ASSERT(cif);
1124
1125         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1126                 if ((type == CARP_COUNT_RUNNING &&
1127                     (SC2IFP(vh)->if_flags & IFF_UP) &&
1128                     (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING)) ||
1129                     (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) {
1130                         TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1131                             ifa_list) {
1132                                 if (ifa->ifa_addr->sa_family == AF_INET &&
1133                                     ia->ia_addr.sin_addr.s_addr ==
1134                                     ifatoia(ifa)->ia_addr.sin_addr.s_addr)
1135                                         count++;
1136                         }
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                                 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1176                                     ifa_list) {
1177                                         if (ifa->ifa_addr->sa_family ==
1178                                             AF_INET &&
1179                                             ia->ia_addr.sin_addr.s_addr ==
1180                                             ifatoia(ifa)->ia_addr.sin_addr.s_addr) {
1181                                                 if (count == index) {
1182                                                         if (vh->sc_state ==
1183                                                             MASTER) {
1184                                                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1185                                                                 CARP_UNLOCK(cif);
1186                                                                 return (1);
1187                                                         } else {
1188                                                                 CARP_UNLOCK(cif);
1189                                                                 return (0);
1190                                                         }
1191                                                 }
1192                                                 count++;
1193                                         }
1194                                 }
1195                         }
1196                 }
1197         } else {
1198                 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1199                         if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1200                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1201                             ia->ia_ifp == SC2IFP(vh) &&
1202                             vh->sc_state == MASTER) {
1203                                 *enaddr = IF_LLADDR(vh->sc_ifp);
1204                                 CARP_UNLOCK(cif);
1205                                 return (1);
1206                         }
1207                 }
1208         }
1209         CARP_UNLOCK(cif);
1210         return (0);
1211 }
1212
1213 #ifdef INET6
1214 struct ifaddr *
1215 carp_iamatch6(void *v, struct in6_addr *taddr)
1216 {
1217         struct carp_if *cif = v;
1218         struct carp_softc *vh;
1219         struct ifaddr *ifa;
1220
1221         CARP_LOCK(cif);
1222         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1223                 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist, ifa_list) {
1224                         if (IN6_ARE_ADDR_EQUAL(taddr,
1225                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1226                             (SC2IFP(vh)->if_flags & IFF_UP) &&
1227                             (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1228                             vh->sc_state == MASTER) {
1229                                 CARP_UNLOCK(cif);
1230                                 return (ifa);
1231                         }
1232                 }
1233         }
1234         CARP_UNLOCK(cif);
1235
1236         return (NULL);
1237 }
1238
1239 void *
1240 carp_macmatch6(void *v, struct mbuf *m, const struct in6_addr *taddr)
1241 {
1242         struct m_tag *mtag;
1243         struct carp_if *cif = v;
1244         struct carp_softc *sc;
1245         struct ifaddr *ifa;
1246
1247         CARP_LOCK(cif);
1248         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) {
1249                 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1250                         if (IN6_ARE_ADDR_EQUAL(taddr,
1251                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1252                             (SC2IFP(sc)->if_flags & IFF_UP) &&
1253                             (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING)) {
1254                                 struct ifnet *ifp = SC2IFP(sc);
1255                                 mtag = m_tag_get(PACKET_TAG_CARP,
1256                                     sizeof(struct ifnet *), M_NOWAIT);
1257                                 if (mtag == NULL) {
1258                                         /* better a bit than nothing */
1259                                         CARP_UNLOCK(cif);
1260                                         return (IF_LLADDR(sc->sc_ifp));
1261                                 }
1262                                 bcopy(&ifp, (caddr_t)(mtag + 1),
1263                                     sizeof(struct ifnet *));
1264                                 m_tag_prepend(m, mtag);
1265
1266                                 CARP_UNLOCK(cif);
1267                                 return (IF_LLADDR(sc->sc_ifp));
1268                         }
1269                 }
1270         }
1271         CARP_UNLOCK(cif);
1272
1273         return (NULL);
1274 }
1275 #endif
1276
1277 struct ifnet *
1278 carp_forus(void *v, void *dhost)
1279 {
1280         struct carp_if *cif = v;
1281         struct carp_softc *vh;
1282         u_int8_t *ena = dhost;
1283
1284         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1285                 return (NULL);
1286
1287         CARP_LOCK(cif);
1288         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list)
1289                 if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1290                     (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1291                     vh->sc_state == MASTER &&
1292                     !bcmp(dhost, IF_LLADDR(vh->sc_ifp), ETHER_ADDR_LEN)) {
1293                         CARP_UNLOCK(cif);
1294                         return (SC2IFP(vh));
1295                 }
1296
1297         CARP_UNLOCK(cif);
1298         return (NULL);
1299 }
1300
1301 static void
1302 carp_master_down(void *v)
1303 {
1304         struct carp_softc *sc = v;
1305
1306         CARP_SCLOCK(sc);
1307         carp_master_down_locked(sc);
1308         CARP_SCUNLOCK(sc);
1309 }
1310
1311 static void
1312 carp_master_down_locked(struct carp_softc *sc)
1313 {
1314         if (sc->sc_carpdev)
1315                 CARP_SCLOCK_ASSERT(sc);
1316
1317         switch (sc->sc_state) {
1318         case INIT:
1319                 printf("%s: master_down event in INIT state\n",
1320                     SC2IFP(sc)->if_xname);
1321                 break;
1322         case MASTER:
1323                 break;
1324         case BACKUP:
1325                 carp_set_state(sc, MASTER);
1326                 carp_send_ad_locked(sc);
1327                 carp_send_arp(sc);
1328 #ifdef INET6
1329                 carp_send_na(sc);
1330 #endif /* INET6 */
1331                 carp_setrun(sc, 0);
1332                 carp_setroute(sc, RTM_ADD);
1333                 break;
1334         }
1335 }
1336
1337 /*
1338  * When in backup state, af indicates whether to reset the master down timer
1339  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1340  */
1341 static void
1342 carp_setrun(struct carp_softc *sc, sa_family_t af)
1343 {
1344         struct timeval tv;
1345
1346         if (sc->sc_carpdev == NULL) {
1347                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1348                 carp_set_state(sc, INIT);
1349                 return;
1350         } else
1351                 CARP_SCLOCK_ASSERT(sc);
1352
1353         if (SC2IFP(sc)->if_flags & IFF_UP &&
1354             sc->sc_vhid > 0 && (sc->sc_naddrs || sc->sc_naddrs6))
1355                 SC2IFP(sc)->if_drv_flags |= IFF_DRV_RUNNING;
1356         else {
1357                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1358                 carp_setroute(sc, RTM_DELETE);
1359                 return;
1360         }
1361
1362         switch (sc->sc_state) {
1363         case INIT:
1364                 if (carp_opts[CARPCTL_PREEMPT] && !carp_suppress_preempt) {
1365                         carp_send_ad_locked(sc);
1366                         carp_send_arp(sc);
1367 #ifdef INET6
1368                         carp_send_na(sc);
1369 #endif /* INET6 */
1370                         CARP_DEBUG("%s: INIT -> MASTER (preempting)\n",
1371                             SC2IFP(sc)->if_xname);
1372                         carp_set_state(sc, MASTER);
1373                         carp_setroute(sc, RTM_ADD);
1374                 } else {
1375                         CARP_DEBUG("%s: INIT -> BACKUP\n", SC2IFP(sc)->if_xname);
1376                         carp_set_state(sc, BACKUP);
1377                         carp_setroute(sc, RTM_DELETE);
1378                         carp_setrun(sc, 0);
1379                 }
1380                 break;
1381         case BACKUP:
1382                 callout_stop(&sc->sc_ad_tmo);
1383                 tv.tv_sec = 3 * sc->sc_advbase;
1384                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1385                 switch (af) {
1386 #ifdef INET
1387                 case AF_INET:
1388                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1389                             carp_master_down, sc);
1390                         break;
1391 #endif /* INET */
1392 #ifdef INET6
1393                 case AF_INET6:
1394                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1395                             carp_master_down, sc);
1396                         break;
1397 #endif /* INET6 */
1398                 default:
1399                         if (sc->sc_naddrs)
1400                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1401                                     carp_master_down, sc);
1402                         if (sc->sc_naddrs6)
1403                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1404                                     carp_master_down, sc);
1405                         break;
1406                 }
1407                 break;
1408         case MASTER:
1409                 tv.tv_sec = sc->sc_advbase;
1410                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1411                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1412                     carp_send_ad, sc);
1413                 break;
1414         }
1415 }
1416
1417 static void
1418 carp_multicast_cleanup(struct carp_softc *sc)
1419 {
1420         struct ip_moptions *imo = &sc->sc_imo;
1421         u_int16_t n = imo->imo_num_memberships;
1422
1423         /* Clean up our own multicast memberships */
1424         while (n-- > 0) {
1425                 if (imo->imo_membership[n] != NULL) {
1426                         in_delmulti(imo->imo_membership[n]);
1427                         imo->imo_membership[n] = NULL;
1428                 }
1429         }
1430         KASSERT(imo->imo_mfilters == NULL,
1431            ("%s: imo_mfilters != NULL", __func__));
1432         imo->imo_num_memberships = 0;
1433         imo->imo_multicast_ifp = NULL;
1434 }
1435
1436 #ifdef INET6
1437 static void
1438 carp_multicast6_cleanup(struct carp_softc *sc)
1439 {
1440         struct ip6_moptions *im6o = &sc->sc_im6o;
1441
1442         while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1443                 struct in6_multi_mship *imm =
1444                     LIST_FIRST(&im6o->im6o_memberships);
1445
1446                 LIST_REMOVE(imm, i6mm_chain);
1447                 in6_leavegroup(imm);
1448         }
1449         im6o->im6o_multicast_ifp = NULL;
1450 }
1451 #endif
1452
1453 static int
1454 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1455 {
1456         INIT_VNET_INET(curvnet);
1457         struct ifnet *ifp;
1458         struct carp_if *cif;
1459         struct in_ifaddr *ia, *ia_if;
1460         struct ip_moptions *imo = &sc->sc_imo;
1461         struct in_addr addr;
1462         u_long iaddr = htonl(sin->sin_addr.s_addr);
1463         int own, error;
1464
1465         if (sin->sin_addr.s_addr == 0) {
1466                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1467                         carp_set_state(sc, INIT);
1468                 if (sc->sc_naddrs)
1469                         SC2IFP(sc)->if_flags |= IFF_UP;
1470                 if (sc->sc_carpdev)
1471                         CARP_SCLOCK(sc);
1472                 carp_setrun(sc, 0);
1473                 if (sc->sc_carpdev)
1474                         CARP_SCUNLOCK(sc);
1475                 return (0);
1476         }
1477
1478         /* we have to do it by hands to check we won't match on us */
1479         ia_if = NULL; own = 0;
1480         TAILQ_FOREACH(ia, &V_in_ifaddrhead, ia_link) {
1481                 /* and, yeah, we need a multicast-capable iface too */
1482                 if (ia->ia_ifp != SC2IFP(sc) &&
1483                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1484                     (iaddr & ia->ia_subnetmask) == ia->ia_subnet) {
1485                         if (!ia_if)
1486                                 ia_if = ia;
1487                         if (sin->sin_addr.s_addr ==
1488                             ia->ia_addr.sin_addr.s_addr)
1489                                 own++;
1490                 }
1491         }
1492
1493         if (!ia_if)
1494                 return (EADDRNOTAVAIL);
1495
1496         ia = ia_if;
1497         ifp = ia->ia_ifp;
1498
1499         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1500             (imo->imo_multicast_ifp && imo->imo_multicast_ifp != ifp))
1501                 return (EADDRNOTAVAIL);
1502
1503         if (imo->imo_num_memberships == 0) {
1504                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1505                 if ((imo->imo_membership[0] = in_addmulti(&addr, ifp)) == NULL)
1506                         return (ENOBUFS);
1507                 imo->imo_num_memberships++;
1508                 imo->imo_multicast_ifp = ifp;
1509                 imo->imo_multicast_ttl = CARP_DFLTTL;
1510                 imo->imo_multicast_loop = 0;
1511         }
1512
1513         if (!ifp->if_carp) {
1514
1515                 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP,
1516                     M_WAITOK|M_ZERO);
1517                 if (!cif) {
1518                         error = ENOBUFS;
1519                         goto cleanup;
1520                 }
1521                 if ((error = ifpromisc(ifp, 1))) {
1522                         FREE(cif, M_CARP);
1523                         goto cleanup;
1524                 }
1525
1526                 CARP_LOCK_INIT(cif);
1527                 CARP_LOCK(cif);
1528                 cif->vhif_ifp = ifp;
1529                 TAILQ_INIT(&cif->vhif_vrs);
1530                 ifp->if_carp = cif;
1531
1532         } else {
1533                 struct carp_softc *vr;
1534
1535                 cif = (struct carp_if *)ifp->if_carp;
1536                 CARP_LOCK(cif);
1537                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1538                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1539                                 CARP_UNLOCK(cif);
1540                                 error = EEXIST;
1541                                 goto cleanup;
1542                         }
1543         }
1544         sc->sc_ia = ia;
1545         sc->sc_carpdev = ifp;
1546
1547         { /* XXX prevent endless loop if already in queue */
1548         struct carp_softc *vr, *after = NULL;
1549         int myself = 0;
1550         cif = (struct carp_if *)ifp->if_carp;
1551
1552         /* XXX: cif should not change, right? So we still hold the lock */
1553         CARP_LOCK_ASSERT(cif);
1554
1555         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1556                 if (vr == sc)
1557                         myself = 1;
1558                 if (vr->sc_vhid < sc->sc_vhid)
1559                         after = vr;
1560         }
1561
1562         if (!myself) {
1563                 /* We're trying to keep things in order */
1564                 if (after == NULL) {
1565                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1566                 } else {
1567                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1568                 }
1569                 cif->vhif_nvrs++;
1570         }
1571         }
1572
1573         sc->sc_naddrs++;
1574         SC2IFP(sc)->if_flags |= IFF_UP;
1575         if (own)
1576                 sc->sc_advskew = 0;
1577         carp_sc_state_locked(sc);
1578         carp_setrun(sc, 0);
1579
1580         CARP_UNLOCK(cif);
1581
1582         return (0);
1583
1584 cleanup:
1585         in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1586         return (error);
1587 }
1588
1589 static int
1590 carp_del_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1591 {
1592         int error = 0;
1593
1594         if (!--sc->sc_naddrs) {
1595                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1596                 struct ip_moptions *imo = &sc->sc_imo;
1597
1598                 CARP_LOCK(cif);
1599                 callout_stop(&sc->sc_ad_tmo);
1600                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1601                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1602                 sc->sc_vhid = -1;
1603                 in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1604                 imo->imo_multicast_ifp = NULL;
1605                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1606                 if (!--cif->vhif_nvrs) {
1607                         sc->sc_carpdev->if_carp = NULL;
1608                         CARP_LOCK_DESTROY(cif);
1609                         FREE(cif, M_IFADDR);
1610                 } else {
1611                         CARP_UNLOCK(cif);
1612                 }
1613         }
1614
1615         return (error);
1616 }
1617
1618 #ifdef INET6
1619 static int
1620 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1621 {
1622         INIT_VNET_INET6(curvnet);
1623         struct ifnet *ifp;
1624         struct carp_if *cif;
1625         struct in6_ifaddr *ia, *ia_if;
1626         struct ip6_moptions *im6o = &sc->sc_im6o;
1627         struct in6_multi_mship *imm;
1628         struct in6_addr in6;
1629         int own, error;
1630
1631         if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
1632                 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1633                         carp_set_state(sc, INIT);
1634                 if (sc->sc_naddrs6)
1635                         SC2IFP(sc)->if_flags |= IFF_UP;
1636                 if (sc->sc_carpdev)
1637                         CARP_SCLOCK(sc);
1638                 carp_setrun(sc, 0);
1639                 if (sc->sc_carpdev)
1640                         CARP_SCUNLOCK(sc);
1641                 return (0);
1642         }
1643
1644         /* we have to do it by hands to check we won't match on us */
1645         ia_if = NULL; own = 0;
1646         for (ia = V_in6_ifaddr; ia; ia = ia->ia_next) {
1647                 int i;
1648
1649                 for (i = 0; i < 4; i++) {
1650                         if ((sin6->sin6_addr.s6_addr32[i] &
1651                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]) !=
1652                             (ia->ia_addr.sin6_addr.s6_addr32[i] &
1653                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]))
1654                                 break;
1655                 }
1656                 /* and, yeah, we need a multicast-capable iface too */
1657                 if (ia->ia_ifp != SC2IFP(sc) &&
1658                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1659                     (i == 4)) {
1660                         if (!ia_if)
1661                                 ia_if = ia;
1662                         if (IN6_ARE_ADDR_EQUAL(&sin6->sin6_addr,
1663                             &ia->ia_addr.sin6_addr))
1664                                 own++;
1665                 }
1666         }
1667
1668         if (!ia_if)
1669                 return (EADDRNOTAVAIL);
1670         ia = ia_if;
1671         ifp = ia->ia_ifp;
1672
1673         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1674             (im6o->im6o_multicast_ifp && im6o->im6o_multicast_ifp != ifp))
1675                 return (EADDRNOTAVAIL);
1676
1677         if (!sc->sc_naddrs6) {
1678                 im6o->im6o_multicast_ifp = ifp;
1679
1680                 /* join CARP multicast address */
1681                 bzero(&in6, sizeof(in6));
1682                 in6.s6_addr16[0] = htons(0xff02);
1683                 in6.s6_addr8[15] = 0x12;
1684                 if (in6_setscope(&in6, ifp, NULL) != 0)
1685                         goto cleanup;
1686                 if ((imm = in6_joingroup(ifp, &in6, &error, 0)) == NULL)
1687                         goto cleanup;
1688                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1689
1690                 /* join solicited multicast address */
1691                 bzero(&in6, sizeof(in6));
1692                 in6.s6_addr16[0] = htons(0xff02);
1693                 in6.s6_addr32[1] = 0;
1694                 in6.s6_addr32[2] = htonl(1);
1695                 in6.s6_addr32[3] = sin6->sin6_addr.s6_addr32[3];
1696                 in6.s6_addr8[12] = 0xff;
1697                 if (in6_setscope(&in6, ifp, NULL) != 0)
1698                         goto cleanup;
1699                 if ((imm = in6_joingroup(ifp, &in6, &error, 0)) == NULL)
1700                         goto cleanup;
1701                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1702         }
1703
1704         if (!ifp->if_carp) {
1705                 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP,
1706                     M_WAITOK|M_ZERO);
1707                 if (!cif) {
1708                         error = ENOBUFS;
1709                         goto cleanup;
1710                 }
1711                 if ((error = ifpromisc(ifp, 1))) {
1712                         FREE(cif, M_CARP);
1713                         goto cleanup;
1714                 }
1715
1716                 CARP_LOCK_INIT(cif);
1717                 CARP_LOCK(cif);
1718                 cif->vhif_ifp = ifp;
1719                 TAILQ_INIT(&cif->vhif_vrs);
1720                 ifp->if_carp = cif;
1721
1722         } else {
1723                 struct carp_softc *vr;
1724
1725                 cif = (struct carp_if *)ifp->if_carp;
1726                 CARP_LOCK(cif);
1727                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1728                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1729                                 CARP_UNLOCK(cif);
1730                                 error = EINVAL;
1731                                 goto cleanup;
1732                         }
1733         }
1734         sc->sc_ia6 = ia;
1735         sc->sc_carpdev = ifp;
1736
1737         { /* XXX prevent endless loop if already in queue */
1738         struct carp_softc *vr, *after = NULL;
1739         int myself = 0;
1740         cif = (struct carp_if *)ifp->if_carp;
1741         CARP_LOCK_ASSERT(cif);
1742
1743         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1744                 if (vr == sc)
1745                         myself = 1;
1746                 if (vr->sc_vhid < sc->sc_vhid)
1747                         after = vr;
1748         }
1749
1750         if (!myself) {
1751                 /* We're trying to keep things in order */
1752                 if (after == NULL) {
1753                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1754                 } else {
1755                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1756                 }
1757                 cif->vhif_nvrs++;
1758         }
1759         }
1760
1761         sc->sc_naddrs6++;
1762         SC2IFP(sc)->if_flags |= IFF_UP;
1763         if (own)
1764                 sc->sc_advskew = 0;
1765         carp_sc_state_locked(sc);
1766         carp_setrun(sc, 0);
1767
1768         CARP_UNLOCK(cif);
1769
1770         return (0);
1771
1772 cleanup:
1773         /* clean up multicast memberships */
1774         if (!sc->sc_naddrs6) {
1775                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1776                         imm = LIST_FIRST(&im6o->im6o_memberships);
1777                         LIST_REMOVE(imm, i6mm_chain);
1778                         in6_leavegroup(imm);
1779                 }
1780         }
1781         return (error);
1782 }
1783
1784 static int
1785 carp_del_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1786 {
1787         int error = 0;
1788
1789         if (!--sc->sc_naddrs6) {
1790                 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1791                 struct ip6_moptions *im6o = &sc->sc_im6o;
1792
1793                 CARP_LOCK(cif);
1794                 callout_stop(&sc->sc_ad_tmo);
1795                 SC2IFP(sc)->if_flags &= ~IFF_UP;
1796                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1797                 sc->sc_vhid = -1;
1798                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1799                         struct in6_multi_mship *imm =
1800                             LIST_FIRST(&im6o->im6o_memberships);
1801
1802                         LIST_REMOVE(imm, i6mm_chain);
1803                         in6_leavegroup(imm);
1804                 }
1805                 im6o->im6o_multicast_ifp = NULL;
1806                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1807                 if (!--cif->vhif_nvrs) {
1808                         CARP_LOCK_DESTROY(cif);
1809                         sc->sc_carpdev->if_carp = NULL;
1810                         FREE(cif, M_IFADDR);
1811                 } else
1812                         CARP_UNLOCK(cif);
1813         }
1814
1815         return (error);
1816 }
1817 #endif /* INET6 */
1818
1819 static int
1820 carp_ioctl(struct ifnet *ifp, u_long cmd, caddr_t addr)
1821 {
1822         struct carp_softc *sc = ifp->if_softc, *vr;
1823         struct carpreq carpr;
1824         struct ifaddr *ifa;
1825         struct ifreq *ifr;
1826         struct ifaliasreq *ifra;
1827         int locked = 0, error = 0;
1828
1829         ifa = (struct ifaddr *)addr;
1830         ifra = (struct ifaliasreq *)addr;
1831         ifr = (struct ifreq *)addr;
1832
1833         switch (cmd) {
1834         case SIOCSIFADDR:
1835                 switch (ifa->ifa_addr->sa_family) {
1836 #ifdef INET
1837                 case AF_INET:
1838                         SC2IFP(sc)->if_flags |= IFF_UP;
1839                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1840                             sizeof(struct sockaddr));
1841                         error = carp_set_addr(sc, satosin(ifa->ifa_addr));
1842                         break;
1843 #endif /* INET */
1844 #ifdef INET6
1845                 case AF_INET6:
1846                         SC2IFP(sc)->if_flags |= IFF_UP;
1847                         error = carp_set_addr6(sc, satosin6(ifa->ifa_addr));
1848                         break;
1849 #endif /* INET6 */
1850                 default:
1851                         error = EAFNOSUPPORT;
1852                         break;
1853                 }
1854                 break;
1855
1856         case SIOCAIFADDR:
1857                 switch (ifa->ifa_addr->sa_family) {
1858 #ifdef INET
1859                 case AF_INET:
1860                         SC2IFP(sc)->if_flags |= IFF_UP;
1861                         bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1862                             sizeof(struct sockaddr));
1863                         error = carp_set_addr(sc, satosin(&ifra->ifra_addr));
1864                         break;
1865 #endif /* INET */
1866 #ifdef INET6
1867                 case AF_INET6:
1868                         SC2IFP(sc)->if_flags |= IFF_UP;
1869                         error = carp_set_addr6(sc, satosin6(&ifra->ifra_addr));
1870                         break;
1871 #endif /* INET6 */
1872                 default:
1873                         error = EAFNOSUPPORT;
1874                         break;
1875                 }
1876                 break;
1877
1878         case SIOCDIFADDR:
1879                 switch (ifa->ifa_addr->sa_family) {
1880 #ifdef INET
1881                 case AF_INET:
1882                         error = carp_del_addr(sc, satosin(&ifra->ifra_addr));
1883                         break;
1884 #endif /* INET */
1885 #ifdef INET6
1886                 case AF_INET6:
1887                         error = carp_del_addr6(sc, satosin6(&ifra->ifra_addr));
1888                         break;
1889 #endif /* INET6 */
1890                 default:
1891                         error = EAFNOSUPPORT;
1892                         break;
1893                 }
1894                 break;
1895
1896         case SIOCSIFFLAGS:
1897                 if (sc->sc_carpdev) {
1898                         locked = 1;
1899                         CARP_SCLOCK(sc);
1900                 }
1901                 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) {
1902                         callout_stop(&sc->sc_ad_tmo);
1903                         callout_stop(&sc->sc_md_tmo);
1904                         callout_stop(&sc->sc_md6_tmo);
1905                         if (sc->sc_state == MASTER)
1906                                 carp_send_ad_locked(sc);
1907                         carp_set_state(sc, INIT);
1908                         carp_setrun(sc, 0);
1909                 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) {
1910                         SC2IFP(sc)->if_flags |= IFF_UP;
1911                         carp_setrun(sc, 0);
1912                 }
1913                 break;
1914
1915         case SIOCSVH:
1916                 error = priv_check(curthread, PRIV_NETINET_CARP);
1917                 if (error)
1918                         break;
1919                 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1920                         break;
1921                 error = 1;
1922                 if (sc->sc_carpdev) {
1923                         locked = 1;
1924                         CARP_SCLOCK(sc);
1925                 }
1926                 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) {
1927                         switch (carpr.carpr_state) {
1928                         case BACKUP:
1929                                 callout_stop(&sc->sc_ad_tmo);
1930                                 carp_set_state(sc, BACKUP);
1931                                 carp_setrun(sc, 0);
1932                                 carp_setroute(sc, RTM_DELETE);
1933                                 break;
1934                         case MASTER:
1935                                 carp_master_down_locked(sc);
1936                                 break;
1937                         default:
1938                                 break;
1939                         }
1940                 }
1941                 if (carpr.carpr_vhid > 0) {
1942                         if (carpr.carpr_vhid > 255) {
1943                                 error = EINVAL;
1944                                 break;
1945                         }
1946                         if (sc->sc_carpdev) {
1947                                 struct carp_if *cif;
1948                                 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1949                                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1950                                         if (vr != sc &&
1951                                             vr->sc_vhid == carpr.carpr_vhid) {
1952                                                 error = EEXIST;
1953                                                 break;
1954                                         }
1955                                 if (error == EEXIST)
1956                                         break;
1957                         }
1958                         sc->sc_vhid = carpr.carpr_vhid;
1959                         IF_LLADDR(sc->sc_ifp)[0] = 0;
1960                         IF_LLADDR(sc->sc_ifp)[1] = 0;
1961                         IF_LLADDR(sc->sc_ifp)[2] = 0x5e;
1962                         IF_LLADDR(sc->sc_ifp)[3] = 0;
1963                         IF_LLADDR(sc->sc_ifp)[4] = 1;
1964                         IF_LLADDR(sc->sc_ifp)[5] = sc->sc_vhid;
1965                         error--;
1966                 }
1967                 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) {
1968                         if (carpr.carpr_advskew >= 255) {
1969                                 error = EINVAL;
1970                                 break;
1971                         }
1972                         if (carpr.carpr_advbase > 255) {
1973                                 error = EINVAL;
1974                                 break;
1975                         }
1976                         sc->sc_advbase = carpr.carpr_advbase;
1977                         sc->sc_advskew = carpr.carpr_advskew;
1978                         error--;
1979                 }
1980                 bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1981                 if (error > 0)
1982                         error = EINVAL;
1983                 else {
1984                         error = 0;
1985                         carp_setrun(sc, 0);
1986                 }
1987                 break;
1988
1989         case SIOCGVH:
1990                 /* XXX: lockless read */
1991                 bzero(&carpr, sizeof(carpr));
1992                 carpr.carpr_state = sc->sc_state;
1993                 carpr.carpr_vhid = sc->sc_vhid;
1994                 carpr.carpr_advbase = sc->sc_advbase;
1995                 carpr.carpr_advskew = sc->sc_advskew;
1996                 error = priv_check(curthread, PRIV_NETINET_CARP);
1997                 if (error == 0)
1998                         bcopy(sc->sc_key, carpr.carpr_key,
1999                             sizeof(carpr.carpr_key));
2000                 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
2001                 break;
2002
2003         default:
2004                 error = EINVAL;
2005         }
2006
2007         if (locked)
2008                 CARP_SCUNLOCK(sc);
2009
2010         carp_hmac_prepare(sc);
2011
2012         return (error);
2013 }
2014
2015 /*
2016  * XXX: this is looutput. We should eventually use it from there.
2017  */
2018 static int
2019 carp_looutput(struct ifnet *ifp, struct mbuf *m, struct sockaddr *dst,
2020     struct rtentry *rt)
2021 {
2022         u_int32_t af;
2023
2024         M_ASSERTPKTHDR(m); /* check if we have the packet header */
2025
2026         if (rt && rt->rt_flags & (RTF_REJECT|RTF_BLACKHOLE)) {
2027                 m_freem(m);
2028                 return (rt->rt_flags & RTF_BLACKHOLE ? 0 :
2029                         rt->rt_flags & RTF_HOST ? EHOSTUNREACH : ENETUNREACH);
2030         }
2031
2032         ifp->if_opackets++;
2033         ifp->if_obytes += m->m_pkthdr.len;
2034
2035         /* BPF writes need to be handled specially. */
2036         if (dst->sa_family == AF_UNSPEC) {
2037                 bcopy(dst->sa_data, &af, sizeof(af));
2038                 dst->sa_family = af;
2039         }
2040
2041 #if 1   /* XXX */
2042         switch (dst->sa_family) {
2043         case AF_INET:
2044         case AF_INET6:
2045         case AF_IPX:
2046         case AF_APPLETALK:
2047                 break;
2048         default:
2049                 printf("carp_looutput: af=%d unexpected\n", dst->sa_family);
2050                 m_freem(m);
2051                 return (EAFNOSUPPORT);
2052         }
2053 #endif
2054         return(if_simloop(ifp, m, dst->sa_family, 0));
2055 }
2056
2057 /*
2058  * Start output on carp interface. This function should never be called.
2059  */
2060 static void
2061 carp_start(struct ifnet *ifp)
2062 {
2063 #ifdef DEBUG
2064         printf("%s: start called\n", ifp->if_xname);
2065 #endif
2066 }
2067
2068 int
2069 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *sa,
2070     struct rtentry *rt)
2071 {
2072         struct m_tag *mtag;
2073         struct carp_softc *sc;
2074         struct ifnet *carp_ifp;
2075
2076         if (!sa)
2077                 return (0);
2078
2079         switch (sa->sa_family) {
2080 #ifdef INET
2081         case AF_INET:
2082                 break;
2083 #endif /* INET */
2084 #ifdef INET6
2085         case AF_INET6:
2086                 break;
2087 #endif /* INET6 */
2088         default:
2089                 return (0);
2090         }
2091
2092         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
2093         if (mtag == NULL)
2094                 return (0);
2095
2096         bcopy(mtag + 1, &carp_ifp, sizeof(struct ifnet *));
2097         sc = carp_ifp->if_softc;
2098
2099         /* Set the source MAC address to Virtual Router MAC Address */
2100         switch (ifp->if_type) {
2101         case IFT_ETHER:
2102         case IFT_L2VLAN: {
2103                         struct ether_header *eh;
2104
2105                         eh = mtod(m, struct ether_header *);
2106                         eh->ether_shost[0] = 0;
2107                         eh->ether_shost[1] = 0;
2108                         eh->ether_shost[2] = 0x5e;
2109                         eh->ether_shost[3] = 0;
2110                         eh->ether_shost[4] = 1;
2111                         eh->ether_shost[5] = sc->sc_vhid;
2112                 }
2113                 break;
2114         case IFT_FDDI: {
2115                         struct fddi_header *fh;
2116
2117                         fh = mtod(m, struct fddi_header *);
2118                         fh->fddi_shost[0] = 0;
2119                         fh->fddi_shost[1] = 0;
2120                         fh->fddi_shost[2] = 0x5e;
2121                         fh->fddi_shost[3] = 0;
2122                         fh->fddi_shost[4] = 1;
2123                         fh->fddi_shost[5] = sc->sc_vhid;
2124                 }
2125                 break;
2126         case IFT_ISO88025: {
2127                         struct iso88025_header *th;
2128                         th = mtod(m, struct iso88025_header *);
2129                         th->iso88025_shost[0] = 3;
2130                         th->iso88025_shost[1] = 0;
2131                         th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
2132                         th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
2133                         th->iso88025_shost[4] = 0;
2134                         th->iso88025_shost[5] = 0;
2135                 }
2136                 break;
2137         default:
2138                 printf("%s: carp is not supported for this interface type\n",
2139                     ifp->if_xname);
2140                 return (EOPNOTSUPP);
2141         }
2142
2143         return (0);
2144 }
2145
2146 static void
2147 carp_set_state(struct carp_softc *sc, int state)
2148 {
2149
2150         if (sc->sc_carpdev)
2151                 CARP_SCLOCK_ASSERT(sc);
2152
2153         if (sc->sc_state == state)
2154                 return;
2155
2156         sc->sc_state = state;
2157         switch (state) {
2158         case BACKUP:
2159                 SC2IFP(sc)->if_link_state = LINK_STATE_DOWN;
2160                 break;
2161         case MASTER:
2162                 SC2IFP(sc)->if_link_state = LINK_STATE_UP;
2163                 break;
2164         default:
2165                 SC2IFP(sc)->if_link_state = LINK_STATE_UNKNOWN;
2166                 break;
2167         }
2168         rt_ifmsg(SC2IFP(sc));
2169 }
2170
2171 void
2172 carp_carpdev_state(void *v)
2173 {
2174         struct carp_if *cif = v;
2175
2176         CARP_LOCK(cif);
2177         carp_carpdev_state_locked(cif);
2178         CARP_UNLOCK(cif);
2179 }
2180
2181 static void
2182 carp_carpdev_state_locked(struct carp_if *cif)
2183 {
2184         struct carp_softc *sc;
2185
2186         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list)
2187                 carp_sc_state_locked(sc);
2188 }
2189
2190 static void
2191 carp_sc_state_locked(struct carp_softc *sc)
2192 {
2193         CARP_SCLOCK_ASSERT(sc);
2194
2195         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
2196             !(sc->sc_carpdev->if_flags & IFF_UP)) {
2197                 sc->sc_flags_backup = SC2IFP(sc)->if_flags;
2198                 SC2IFP(sc)->if_flags &= ~IFF_UP;
2199                 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
2200                 callout_stop(&sc->sc_ad_tmo);
2201                 callout_stop(&sc->sc_md_tmo);
2202                 callout_stop(&sc->sc_md6_tmo);
2203                 carp_set_state(sc, INIT);
2204                 carp_setrun(sc, 0);
2205                 if (!sc->sc_suppress) {
2206                         carp_suppress_preempt++;
2207                         if (carp_suppress_preempt == 1) {
2208                                 CARP_SCUNLOCK(sc);
2209                                 carp_send_ad_all();
2210                                 CARP_SCLOCK(sc);
2211                         }
2212                 }
2213                 sc->sc_suppress = 1;
2214         } else {
2215                 SC2IFP(sc)->if_flags |= sc->sc_flags_backup;
2216                 carp_set_state(sc, INIT);
2217                 carp_setrun(sc, 0);
2218                 if (sc->sc_suppress)
2219                         carp_suppress_preempt--;
2220                 sc->sc_suppress = 0;
2221         }
2222
2223         return;
2224 }
2225
2226 static int
2227 carp_modevent(module_t mod, int type, void *data)
2228 {
2229         switch (type) {
2230         case MOD_LOAD:
2231                 if_detach_event_tag = EVENTHANDLER_REGISTER(ifnet_departure_event,
2232                     carp_ifdetach, NULL, EVENTHANDLER_PRI_ANY);
2233                 if (if_detach_event_tag == NULL)
2234                         return (ENOMEM);
2235                 mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2236                 LIST_INIT(&carpif_list);
2237                 if_clone_attach(&carp_cloner);
2238                 break;
2239
2240         case MOD_UNLOAD:
2241                 EVENTHANDLER_DEREGISTER(ifnet_departure_event, if_detach_event_tag);
2242                 if_clone_detach(&carp_cloner);
2243                 mtx_destroy(&carp_mtx);
2244                 break;
2245
2246         default:
2247                 return (EINVAL);
2248         }
2249
2250         return (0);
2251 }
2252
2253 static moduledata_t carp_mod = {
2254         "carp",
2255         carp_modevent,
2256         0
2257 };
2258
2259 DECLARE_MODULE(carp, carp_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);