sys/netinet/ip_carp.c

   1 /*-
   2  * Copyright (c) 2002 Michael Shalayeff.
   3  * Copyright (c) 2003 Ryan McBride.
   4  * Copyright (c) 2011 Gleb Smirnoff <glebius@FreeBSD.org>
   5  * All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  17  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  18  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  19  * IN NO EVENT SHALL THE AUTHOR OR HIS RELATIVES BE LIABLE FOR ANY DIRECT,
  20  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  21  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  22  * SERVICES; LOSS OF MIND, USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
  24  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
  25  * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
  26  * THE POSSIBILITY OF SUCH DAMAGE.
  27  */
  28
  29 #include <sys/cdefs.h>
  30 __FBSDID("$FreeBSD$");
  31
  32 #include "opt_bpf.h"
  33 #include "opt_inet.h"
  34 #include "opt_inet6.h"
  35
  36 #include <sys/param.h>
  37 #include <sys/systm.h>
  38 #include <sys/bus.h>
  39 #include <sys/jail.h>
  40 #include <sys/kernel.h>
  41 #include <sys/limits.h>
  42 #include <sys/malloc.h>
  43 #include <sys/mbuf.h>
  44 #include <sys/module.h>
  45 #include <sys/priv.h>
  46 #include <sys/proc.h>
  47 #include <sys/protosw.h>
  48 #include <sys/socket.h>
  49 #include <sys/sockio.h>
  50 #include <sys/sysctl.h>
  51 #include <sys/syslog.h>
  52 #include <sys/taskqueue.h>
  53
  54 #include <net/ethernet.h>
  55 #include <net/fddi.h>
  56 #include <net/if.h>
  57 #include <net/if_dl.h>
  58 #include <net/if_llatbl.h>
  59 #include <net/if_types.h>
  60 #include <net/iso88025.h>
  61 #include <net/route.h>
  62 #include <net/vnet.h>
  63
  64 #if defined(INET) || defined(INET6)
  65 #include <netinet/in.h>
  66 #include <netinet/in_var.h>
  67 #include <netinet/ip_carp.h>
  68 #include <netinet/ip.h>
  69 #include <machine/in_cksum.h>
  70 #endif
  71 #ifdef INET
  72 #include <netinet/ip_var.h>
  73 #include <netinet/if_ether.h>
  74 #endif
  75
  76 #ifdef INET6
  77 #include <netinet/icmp6.h>
  78 #include <netinet/ip6.h>
  79 #include <netinet6/ip6protosw.h>
  80 #include <netinet6/in6_var.h>
  81 #include <netinet6/ip6_var.h>
  82 #include <netinet6/scope6_var.h>
  83 #include <netinet6/nd6.h>
  84 #endif
  85
  86 #include <crypto/sha1.h>
  87
  88 static MALLOC_DEFINE(M_CARP, "CARP", "CARP addresses");
  89
  90 struct carp_softc {
  91         struct ifnet            *sc_carpdev;    /* Pointer to parent ifnet. */
  92         struct ifaddr           **sc_ifas;      /* Our ifaddrs. */
  93         struct sockaddr_dl      sc_addr;        /* Our link level address. */
  94         struct callout          sc_ad_tmo;      /* Advertising timeout. */
  95 #ifdef INET
  96         struct callout          sc_md_tmo;      /* Master down timeout. */
  97 #endif
  98 #ifdef INET6
  99         struct callout          sc_md6_tmo;     /* XXX: Master down timeout. */
 100 #endif
 101         struct mtx              sc_mtx;
 102
 103         int                     sc_vhid;
 104         int                     sc_advskew;
 105         int                     sc_advbase;
 106
 107         int                     sc_naddrs;
 108         int                     sc_naddrs6;
 109         int                     sc_ifasiz;
 110         enum { INIT = 0, BACKUP, MASTER }       sc_state;
 111         int                     sc_suppress;
 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_init_counter;
 118         uint64_t                sc_counter;
 119
 120         /* authentication */
 121 #define CARP_HMAC_PAD   64
 122         unsigned char sc_key[CARP_KEY_LEN];
 123         unsigned char sc_pad[CARP_HMAC_PAD];
 124         SHA1_CTX sc_sha1;
 125
 126         TAILQ_ENTRY(carp_softc) sc_list;        /* On the carp_if list. */
 127         LIST_ENTRY(carp_softc)  sc_next;        /* On the global list. */
 128 };
 129
 130 struct carp_if {
 131 #ifdef INET
 132         int     cif_naddrs;
 133 #endif
 134 #ifdef INET6
 135         int     cif_naddrs6;
 136 #endif
 137         TAILQ_HEAD(, carp_softc) cif_vrs;
 138 #ifdef INET
 139         struct ip_moptions       cif_imo;
 140 #endif
 141 #ifdef INET6
 142         struct ip6_moptions      cif_im6o;
 143 #endif
 144         struct ifnet    *cif_ifp;
 145         struct mtx      cif_mtx;
 146 };
 147
 148 #define CARP_INET       0
 149 #define CARP_INET6      1
 150 static int proto_reg[] = {-1, -1};
 151
 152 /*
 153  * Brief design of carp(4).
 154  *
 155  * Any carp-capable ifnet may have a list of carp softcs hanging off
 156  * its ifp->if_carp pointer. Each softc represents one unique virtual
 157  * host id, or vhid. The softc has a back pointer to the ifnet. All
 158  * softcs are joined in a global list, which has quite limited use.
 159  *
 160  * Any interface address that takes part in CARP negotiation has a
 161  * pointer to the softc of its vhid, ifa->ifa_carp. That could be either
 162  * AF_INET or AF_INET6 address.
 163  *
 164  * Although, one can get the softc's backpointer to ifnet and traverse
 165  * through its ifp->if_addrhead queue to find all interface addresses
 166  * involved in CARP, we keep a growable array of ifaddr pointers. This
 167  * allows us to avoid grabbing the IF_ADDR_LOCK() in many traversals that
 168  * do calls into the network stack, thus avoiding LORs.
 169  *
 170  * Locking:
 171  *
 172  * Each softc has a lock sc_mtx. It is used to synchronise carp_input_c(),
 173  * callout-driven events and ioctl()s.
 174  *
 175  * To traverse the list of softcs on an ifnet we use CIF_LOCK(), to
 176  * traverse the global list we use the mutex carp_mtx.
 177  *
 178  * Known issues with locking:
 179  *
 180  * - There is no protection for races between two ioctl() requests,
 181  *   neither SIOCSVH, nor SIOCAIFADDR & SIOCAIFADDR_IN6. I think that all
 182  *   interface ioctl()s should be serialized right in net/if.c.
 183  * - Sending ad, we put the pointer to the softc in an mtag, and no reference
 184  *   counting is done on the softc.
 185  * - On module unload we may race (?) with packet processing thread
 186  *   dereferencing our function pointers.
 187  */
 188
 189 static int carp_allow = 1;              /* Accept incoming CARP packets. */
 190 static int carp_preempt = 0;            /* Preempt slower nodes. */
 191 static int carp_log = 1;                /* Log level. */
 192 static int carp_demotion = 0;           /* Global advskew demotion. */
 193 static int carp_senderr_adj = CARP_MAXSKEW;     /* Send error demotion factor */
 194 static int carp_ifdown_adj = CARP_MAXSKEW;      /* Iface down demotion factor */
 195
 196 SYSCTL_NODE(_net_inet, IPPROTO_CARP,    carp,   CTLFLAG_RW, 0,  "CARP");
 197 SYSCTL_INT(_net_inet_carp, OID_AUTO, allow, CTLFLAG_RW, &carp_allow, 0,
 198     "Accept incoming CARP packets");
 199 SYSCTL_INT(_net_inet_carp, OID_AUTO, preempt, CTLFLAG_RW, &carp_preempt, 0,
 200     "High-priority backup preemption mode");
 201 SYSCTL_INT(_net_inet_carp, OID_AUTO, log, CTLFLAG_RW, &carp_log, 0,
 202     "CARP log level");
 203 SYSCTL_INT(_net_inet_carp, OID_AUTO, demotion, CTLFLAG_RW, &carp_demotion, 0,
 204     "Demotion factor (skew of advskew)");
 205 SYSCTL_INT(_net_inet_carp, OID_AUTO, senderr_demotion_factor, CTLFLAG_RW,
 206     &carp_senderr_adj, 0, "Send error demotion factor adjustment");
 207 SYSCTL_INT(_net_inet_carp, OID_AUTO, ifdown_demotion_factor, CTLFLAG_RW,
 208     &carp_ifdown_adj, 0, "Interface down demotion factor adjustment");
 209
 210 static struct carpstats carpstats;
 211 SYSCTL_STRUCT(_net_inet_carp, OID_AUTO, stats, CTLFLAG_RW, &carpstats,
 212     carpstats, "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 213
 214 #define CARP_LOCK_INIT(sc)      mtx_init(&(sc)->sc_mtx, "carp_softc",   \
 215         NULL, MTX_DEF)
 216 #define CARP_LOCK_DESTROY(sc)   mtx_destroy(&(sc)->sc_mtx)
 217 #define CARP_LOCK_ASSERT(sc)    mtx_assert(&(sc)->sc_mtx, MA_OWNED)
 218 #define CARP_LOCK(sc)           mtx_lock(&(sc)->sc_mtx)
 219 #define CARP_UNLOCK(sc)         mtx_unlock(&(sc)->sc_mtx)
 220 #define CIF_LOCK_INIT(cif)      mtx_init(&(cif)->cif_mtx, "carp_if",   \
 221         NULL, MTX_DEF)
 222 #define CIF_LOCK_DESTROY(cif)   mtx_destroy(&(cif)->cif_mtx)
 223 #define CIF_LOCK_ASSERT(cif)    mtx_assert(&(cif)->cif_mtx, MA_OWNED)
 224 #define CIF_LOCK(cif)           mtx_lock(&(cif)->cif_mtx)
 225 #define CIF_UNLOCK(cif)         mtx_unlock(&(cif)->cif_mtx)
 226
 227 #define CARP_LOG(...)   do {                            \
 228         if (carp_log > 0)                               \
 229                 log(LOG_INFO, "carp: " __VA_ARGS__);    \
 230 } while (0)
 231
 232 #define CARP_DEBUG(...) do {                            \
 233         if (carp_log > 1)                               \
 234                 log(LOG_DEBUG, __VA_ARGS__);            \
 235 } while (0)
 236
 237 #define IFNET_FOREACH_IFA(ifp, ifa)                                     \
 238         IF_ADDR_LOCK_ASSERT(ifp);                                       \
 239         TAILQ_FOREACH((ifa), &(ifp)->if_addrhead, ifa_link)             \
 240                 if ((ifa)->ifa_carp != NULL)
 241
 242 #define CARP_FOREACH_IFA(sc, ifa)                                       \
 243         CARP_LOCK_ASSERT(sc);                                           \
 244         for (int _i = 0;                                                \
 245                 _i < (sc)->sc_naddrs + (sc)->sc_naddrs6 &&              \
 246                 ((ifa) = sc->sc_ifas[_i]) != NULL;                      \
 247                 ++_i)
 248
 249 #define IFNET_FOREACH_CARP(ifp, sc)                                     \
 250         CIF_LOCK_ASSERT(ifp->if_carp);                                  \
 251         TAILQ_FOREACH((sc), &(ifp)->if_carp->cif_vrs, sc_list)
 252
 253 #define DEMOTE_ADVSKEW(sc)                                      \
 254     (((sc)->sc_advskew + carp_demotion > CARP_MAXSKEW) ?        \
 255     CARP_MAXSKEW : ((sc)->sc_advskew + carp_demotion))
 256
 257 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 258 static struct carp_softc
 259                 *carp_alloc(struct ifnet *);
 260 static void     carp_destroy(struct carp_softc *);
 261 static struct carp_if
 262                 *carp_alloc_if(struct ifnet *);
 263 static void     carp_free_if(struct carp_if *);
 264 static void     carp_set_state(struct carp_softc *, int);
 265 static void     carp_sc_state(struct carp_softc *);
 266 static void     carp_setrun(struct carp_softc *, sa_family_t);
 267 static void     carp_master_down(void *);
 268 static void     carp_master_down_locked(struct carp_softc *);
 269 static void     carp_send_ad(void *);
 270 static void     carp_send_ad_locked(struct carp_softc *);
 271 static void     carp_addroute(struct carp_softc *);
 272 static void     carp_ifa_addroute(struct ifaddr *);
 273 static void     carp_delroute(struct carp_softc *);
 274 static void     carp_ifa_delroute(struct ifaddr *);
 275 static void     carp_send_ad_all(void *, int);
 276 static void     carp_demote_adj(int, char *);
 277
 278 static LIST_HEAD(, carp_softc) carp_list;
 279 static struct mtx carp_mtx;
 280 static struct task carp_sendall_task =
 281     TASK_INITIALIZER(0, carp_send_ad_all, NULL);
 282
 283 static __inline uint16_t
 284 carp_cksum(struct mbuf *m, int len)
 285 {
 286         return (in_cksum(m, len));
 287 }
 288
 289 static void
 290 carp_hmac_prepare(struct carp_softc *sc)
 291 {
 292         uint8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 293         uint8_t vhid = sc->sc_vhid & 0xff;
 294         struct ifaddr *ifa;
 295         int i, found;
 296 #ifdef INET
 297         struct in_addr last, cur, in;
 298 #endif
 299 #ifdef INET6
 300         struct in6_addr last6, cur6, in6;
 301 #endif
 302
 303         CARP_LOCK_ASSERT(sc);
 304
 305         /* Compute ipad from key. */
 306         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 307         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 308         for (i = 0; i < sizeof(sc->sc_pad); i++)
 309                 sc->sc_pad[i] ^= 0x36;
 310
 311         /* Precompute first part of inner hash. */
 312         SHA1Init(&sc->sc_sha1);
 313         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 314         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 315         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 316         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 317 #ifdef INET
 318         cur.s_addr = 0;
 319         do {
 320                 found = 0;
 321                 last = cur;
 322                 cur.s_addr = 0xffffffff;
 323                 CARP_FOREACH_IFA(sc, ifa) {
 324                         in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 325                         if (ifa->ifa_addr->sa_family == AF_INET &&
 326                             ntohl(in.s_addr) > ntohl(last.s_addr) &&
 327                             ntohl(in.s_addr) < ntohl(cur.s_addr)) {
 328                                 cur.s_addr = in.s_addr;
 329                                 found++;
 330                         }
 331                 }
 332                 if (found)
 333                         SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur));
 334         } while (found);
 335 #endif /* INET */
 336 #ifdef INET6
 337         memset(&cur6, 0, sizeof(cur6));
 338         do {
 339                 found = 0;
 340                 last6 = cur6;
 341                 memset(&cur6, 0xff, sizeof(cur6));
 342                 CARP_FOREACH_IFA(sc, ifa) {
 343                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 344                         if (IN6_IS_SCOPE_EMBED(&in6))
 345                                 in6.s6_addr16[1] = 0;
 346                         if (ifa->ifa_addr->sa_family == AF_INET6 &&
 347                             memcmp(&in6, &last6, sizeof(in6)) > 0 &&
 348                             memcmp(&in6, &cur6, sizeof(in6)) < 0) {
 349                                 cur6 = in6;
 350                                 found++;
 351                         }
 352                 }
 353                 if (found)
 354                         SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6));
 355         } while (found);
 356 #endif /* INET6 */
 357
 358         /* convert ipad to opad */
 359         for (i = 0; i < sizeof(sc->sc_pad); i++)
 360                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 361 }
 362
 363 static void
 364 carp_hmac_generate(struct carp_softc *sc, uint32_t counter[2],
 365     unsigned char md[20])
 366 {
 367         SHA1_CTX sha1ctx;
 368
 369         CARP_LOCK_ASSERT(sc);
 370
 371         /* fetch first half of inner hash */
 372         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 373
 374         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 375         SHA1Final(md, &sha1ctx);
 376
 377         /* outer hash */
 378         SHA1Init(&sha1ctx);
 379         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 380         SHA1Update(&sha1ctx, md, 20);
 381         SHA1Final(md, &sha1ctx);
 382 }
 383
 384 static int
 385 carp_hmac_verify(struct carp_softc *sc, uint32_t counter[2],
 386     unsigned char md[20])
 387 {
 388         unsigned char md2[20];
 389
 390         CARP_LOCK_ASSERT(sc);
 391
 392         carp_hmac_generate(sc, counter, md2);
 393
 394         return (bcmp(md, md2, sizeof(md2)));
 395 }
 396
 397 /*
 398  * process input packet.
 399  * we have rearranged checks order compared to the rfc,
 400  * but it seems more efficient this way or not possible otherwise.
 401  */
 402 #ifdef INET
 403 void
 404 carp_input(struct mbuf *m, int hlen)
 405 {
 406         struct ip *ip = mtod(m, struct ip *);
 407         struct carp_header *ch;
 408         int iplen, len;
 409
 410         CARPSTATS_INC(carps_ipackets);
 411
 412         if (!carp_allow) {
 413                 m_freem(m);
 414                 return;
 415         }
 416
 417         /* verify that the IP TTL is 255.  */
 418         if (ip->ip_ttl != CARP_DFLTTL) {
 419                 CARPSTATS_INC(carps_badttl);
 420                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 421                     ip->ip_ttl,
 422                     m->m_pkthdr.rcvif->if_xname);
 423                 m_freem(m);
 424                 return;
 425         }
 426
 427         iplen = ip->ip_hl << 2;
 428
 429         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 430                 CARPSTATS_INC(carps_badlen);
 431                 CARP_DEBUG("%s: received len %zd < sizeof(struct carp_header) "
 432                     "on %s\n", __func__, m->m_len - sizeof(struct ip),
 433                     m->m_pkthdr.rcvif->if_xname);
 434                 m_freem(m);
 435                 return;
 436         }
 437
 438         if (iplen + sizeof(*ch) < m->m_len) {
 439                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 440                         CARPSTATS_INC(carps_hdrops);
 441                         CARP_DEBUG("%s: pullup failed\n", __func__);
 442                         return;
 443                 }
 444                 ip = mtod(m, struct ip *);
 445         }
 446         ch = (struct carp_header *)((char *)ip + iplen);
 447
 448         /*
 449          * verify that the received packet length is
 450          * equal to the CARP header
 451          */
 452         len = iplen + sizeof(*ch);
 453         if (len > m->m_pkthdr.len) {
 454                 CARPSTATS_INC(carps_badlen);
 455                 CARP_DEBUG("%s: packet too short %d on %s\n", __func__,
 456                     m->m_pkthdr.len,
 457                     m->m_pkthdr.rcvif->if_xname);
 458                 m_freem(m);
 459                 return;
 460         }
 461
 462         if ((m = m_pullup(m, len)) == NULL) {
 463                 CARPSTATS_INC(carps_hdrops);
 464                 return;
 465         }
 466         ip = mtod(m, struct ip *);
 467         ch = (struct carp_header *)((char *)ip + iplen);
 468
 469         /* verify the CARP checksum */
 470         m->m_data += iplen;
 471         if (carp_cksum(m, len - iplen)) {
 472                 CARPSTATS_INC(carps_badsum);
 473                 CARP_DEBUG("%s: checksum failed on %s\n", __func__,
 474                     m->m_pkthdr.rcvif->if_xname);
 475                 m_freem(m);
 476                 return;
 477         }
 478         m->m_data -= iplen;
 479
 480         carp_input_c(m, ch, AF_INET);
 481 }
 482 #endif
 483
 484 #ifdef INET6
 485 int
 486 carp6_input(struct mbuf **mp, int *offp, int proto)
 487 {
 488         struct mbuf *m = *mp;
 489         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 490         struct carp_header *ch;
 491         u_int len;
 492
 493         CARPSTATS_INC(carps_ipackets6);
 494
 495         if (!carp_allow) {
 496                 m_freem(m);
 497                 return (IPPROTO_DONE);
 498         }
 499
 500         /* check if received on a valid carp interface */
 501         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 502                 CARPSTATS_INC(carps_badif);
 503                 CARP_DEBUG("%s: packet received on non-carp interface: %s\n",
 504                     __func__, m->m_pkthdr.rcvif->if_xname);
 505                 m_freem(m);
 506                 return (IPPROTO_DONE);
 507         }
 508
 509         /* verify that the IP TTL is 255 */
 510         if (ip6->ip6_hlim != CARP_DFLTTL) {
 511                 CARPSTATS_INC(carps_badttl);
 512                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 513                     ip6->ip6_hlim, m->m_pkthdr.rcvif->if_xname);
 514                 m_freem(m);
 515                 return (IPPROTO_DONE);
 516         }
 517
 518         /* verify that we have a complete carp packet */
 519         len = m->m_len;
 520         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 521         if (ch == NULL) {
 522                 CARPSTATS_INC(carps_badlen);
 523                 CARP_DEBUG("%s: packet size %u too small\n", __func__, len);
 524                 return (IPPROTO_DONE);
 525         }
 526
 527
 528         /* verify the CARP checksum */
 529         m->m_data += *offp;
 530         if (carp_cksum(m, sizeof(*ch))) {
 531                 CARPSTATS_INC(carps_badsum);
 532                 CARP_DEBUG("%s: checksum failed, on %s\n", __func__,
 533                     m->m_pkthdr.rcvif->if_xname);
 534                 m_freem(m);
 535                 return (IPPROTO_DONE);
 536         }
 537         m->m_data -= *offp;
 538
 539         carp_input_c(m, ch, AF_INET6);
 540         return (IPPROTO_DONE);
 541 }
 542 #endif /* INET6 */
 543
 544 static void
 545 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 546 {
 547         struct ifnet *ifp = m->m_pkthdr.rcvif;
 548         struct ifaddr *ifa;
 549         struct carp_softc *sc;
 550         uint64_t tmp_counter;
 551         struct timeval sc_tv, ch_tv;
 552
 553         /* verify that the VHID is valid on the receiving interface */
 554         IF_ADDR_RLOCK(ifp);
 555         IFNET_FOREACH_IFA(ifp, ifa)
 556                 if (ifa->ifa_addr->sa_family == af &&
 557                     ifa->ifa_carp->sc_vhid == ch->carp_vhid) {
 558                         ifa_ref(ifa);
 559                         break;
 560                 }
 561         IF_ADDR_RUNLOCK(ifp);
 562
 563         if (ifa == NULL) {
 564                 CARPSTATS_INC(carps_badvhid);
 565                 m_freem(m);
 566                 return;
 567         }
 568
 569         /* verify the CARP version. */
 570         if (ch->carp_version != CARP_VERSION) {
 571                 CARPSTATS_INC(carps_badver);
 572                 CARP_DEBUG("%s: invalid version %d\n", ifp->if_xname,
 573                     ch->carp_version);
 574                 ifa_free(ifa);
 575                 m_freem(m);
 576                 return;
 577         }
 578
 579         sc = ifa->ifa_carp;
 580         CARP_LOCK(sc);
 581         ifa_free(ifa);
 582
 583         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 584                 CARPSTATS_INC(carps_badauth);
 585                 CARP_DEBUG("%s: incorrect hash for VHID %u@%s\n", __func__,
 586                     sc->sc_vhid, ifp->if_xname);
 587                 goto out;
 588         }
 589
 590         tmp_counter = ntohl(ch->carp_counter[0]);
 591         tmp_counter = tmp_counter<<32;
 592         tmp_counter += ntohl(ch->carp_counter[1]);
 593
 594         /* XXX Replay protection goes here */
 595
 596         sc->sc_init_counter = 0;
 597         sc->sc_counter = tmp_counter;
 598
 599         sc_tv.tv_sec = sc->sc_advbase;
 600         sc_tv.tv_usec = DEMOTE_ADVSKEW(sc) * 1000000 / 256;
 601         ch_tv.tv_sec = ch->carp_advbase;
 602         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 603
 604         switch (sc->sc_state) {
 605         case INIT:
 606                 break;
 607         case MASTER:
 608                 /*
 609                  * If we receive an advertisement from a master who's going to
 610                  * be more frequent than us, go into BACKUP state.
 611                  */
 612                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 613                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 614                         callout_stop(&sc->sc_ad_tmo);
 615                         CARP_LOG("VHID %u@%s: MASTER -> BACKUP "
 616                             "(more frequent advertisement received)\n",
 617                             sc->sc_vhid,
 618                             sc->sc_carpdev->if_xname);
 619                         carp_set_state(sc, BACKUP);
 620                         carp_setrun(sc, 0);
 621                         carp_delroute(sc);
 622                 }
 623                 break;
 624         case BACKUP:
 625                 /*
 626                  * If we're pre-empting masters who advertise slower than us,
 627                  * and this one claims to be slower, treat him as down.
 628                  */
 629                 if (carp_preempt && timevalcmp(&sc_tv, &ch_tv, <)) {
 630                         CARP_LOG("VHID %u@%s: BACKUP -> MASTER "
 631                             "(preempting a slower master)\n",
 632                             sc->sc_vhid,
 633                             sc->sc_carpdev->if_xname);
 634                         carp_master_down_locked(sc);
 635                         break;
 636                 }
 637
 638                 /*
 639                  *  If the master is going to advertise at such a low frequency
 640                  *  that he's guaranteed to time out, we'd might as well just
 641                  *  treat him as timed out now.
 642                  */
 643                 sc_tv.tv_sec = sc->sc_advbase * 3;
 644                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 645                         CARP_LOG("VHID %u@%s: BACKUP -> MASTER "
 646                             "(master timed out)\n",
 647                             sc->sc_vhid,
 648                             sc->sc_carpdev->if_xname);
 649                         carp_master_down_locked(sc);
 650                         break;
 651                 }
 652
 653                 /*
 654                  * Otherwise, we reset the counter and wait for the next
 655                  * advertisement.
 656                  */
 657                 carp_setrun(sc, af);
 658                 break;
 659         }
 660
 661 out:
 662         CARP_UNLOCK(sc);
 663         m_freem(m);
 664 }
 665
 666 static int
 667 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 668 {
 669         struct m_tag *mtag;
 670
 671         if (sc->sc_init_counter) {
 672                 /* this could also be seconds since unix epoch */
 673                 sc->sc_counter = arc4random();
 674                 sc->sc_counter = sc->sc_counter << 32;
 675                 sc->sc_counter += arc4random();
 676         } else
 677                 sc->sc_counter++;
 678
 679         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 680         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 681
 682         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 683
 684         /* Tag packet for carp_output */
 685         if ((mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct carp_softc *),
 686             M_NOWAIT)) == NULL) {
 687                 m_freem(m);
 688                 CARPSTATS_INC(carps_onomem);
 689                 return (ENOMEM);
 690         }
 691         bcopy(&sc, (caddr_t)(mtag + 1), sizeof(struct carp_softc *));
 692         m_tag_prepend(m, mtag);
 693
 694         return (0);
 695 }
 696
 697 /*
 698  * To avoid LORs and possible recursions this function shouldn't
 699  * be called directly, but scheduled via taskqueue.
 700  */
 701 static void
 702 carp_send_ad_all(void *ctx __unused, int pending __unused)
 703 {
 704         struct carp_softc *sc;
 705
 706         mtx_lock(&carp_mtx);
 707         LIST_FOREACH(sc, &carp_list, sc_next)
 708                 if (sc->sc_state == MASTER) {
 709                         CARP_LOCK(sc);
 710                         CURVNET_SET(sc->sc_carpdev->if_vnet);
 711                         carp_send_ad_locked(sc);
 712                         CURVNET_RESTORE();
 713                         CARP_UNLOCK(sc);
 714                 }
 715         mtx_unlock(&carp_mtx);
 716 }
 717
 718 /* Send a periodic advertisement, executed in callout context. */
 719 static void
 720 carp_send_ad(void *v)
 721 {
 722         struct carp_softc *sc = v;
 723
 724         CARP_LOCK_ASSERT(sc);
 725         CURVNET_SET(sc->sc_carpdev->if_vnet);
 726         carp_send_ad_locked(sc);
 727         CURVNET_RESTORE();
 728         CARP_UNLOCK(sc);
 729 }
 730
 731 static void
 732 carp_send_ad_locked(struct carp_softc *sc)
 733 {
 734         struct carp_header ch;
 735         struct timeval tv;
 736         struct sockaddr sa;
 737         struct ifaddr *ifa;
 738         struct carp_header *ch_ptr;
 739         struct mbuf *m;
 740         int len, advskew;
 741
 742         CARP_LOCK_ASSERT(sc);
 743
 744         advskew = DEMOTE_ADVSKEW(sc);
 745         tv.tv_sec = sc->sc_advbase;
 746         tv.tv_usec = advskew * 1000000 / 256;
 747
 748         ch.carp_version = CARP_VERSION;
 749         ch.carp_type = CARP_ADVERTISEMENT;
 750         ch.carp_vhid = sc->sc_vhid;
 751         ch.carp_advbase = sc->sc_advbase;
 752         ch.carp_advskew = advskew;
 753         ch.carp_authlen = 7;    /* XXX DEFINE */
 754         ch.carp_pad1 = 0;       /* must be zero */
 755         ch.carp_cksum = 0;
 756
 757         /* XXXGL: OpenBSD picks first ifaddr with needed family. */
 758
 759 #ifdef INET
 760         if (sc->sc_naddrs) {
 761                 struct ip *ip;
 762
 763                 MGETHDR(m, M_NOWAIT, MT_HEADER);
 764                 if (m == NULL) {
 765                         CARPSTATS_INC(carps_onomem);
 766                         /* XXX maybe less ? */
 767                         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
 768                             carp_send_ad, sc);
 769                         return;
 770                 }
 771                 len = sizeof(*ip) + sizeof(ch);
 772                 m->m_pkthdr.len = len;
 773                 m->m_pkthdr.rcvif = NULL;
 774                 m->m_len = len;
 775                 MH_ALIGN(m, m->m_len);
 776                 m->m_flags |= M_MCAST;
 777                 ip = mtod(m, struct ip *);
 778                 ip->ip_v = IPVERSION;
 779                 ip->ip_hl = sizeof(*ip) >> 2;
 780                 ip->ip_tos = IPTOS_LOWDELAY;
 781                 ip->ip_len = len;
 782                 ip->ip_id = ip_newid();
 783                 ip->ip_off = IP_DF;
 784                 ip->ip_ttl = CARP_DFLTTL;
 785                 ip->ip_p = IPPROTO_CARP;
 786                 ip->ip_sum = 0;
 787
 788                 bzero(&sa, sizeof(sa));
 789                 sa.sa_family = AF_INET;
 790                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 791                 if (ifa != NULL) {
 792                         ip->ip_src.s_addr =
 793                             ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 794                         ifa_free(ifa);
 795                 } else
 796                         ip->ip_src.s_addr = 0;
 797                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 798
 799                 ch_ptr = (struct carp_header *)(&ip[1]);
 800                 bcopy(&ch, ch_ptr, sizeof(ch));
 801                 if (carp_prepare_ad(m, sc, ch_ptr))
 802                         return;
 803
 804                 m->m_data += sizeof(*ip);
 805                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip));
 806                 m->m_data -= sizeof(*ip);
 807
 808                 CARPSTATS_INC(carps_opackets);
 809
 810                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT,
 811                     &sc->sc_carpdev->if_carp->cif_imo, NULL)) {
 812                         if (sc->sc_sendad_errors < INT_MAX)
 813                                 sc->sc_sendad_errors++;
 814                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS)
 815                                 carp_demote_adj(carp_senderr_adj, "send error");
 816                         sc->sc_sendad_success = 0;
 817                 } else {
 818                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 819                                 if (++sc->sc_sendad_success >=
 820                                     CARP_SENDAD_MIN_SUCCESS) {
 821                                         carp_demote_adj(-carp_senderr_adj,
 822                                             "send ok");
 823                                         sc->sc_sendad_errors = 0;
 824                                 }
 825                         } else
 826                                 sc->sc_sendad_errors = 0;
 827                 }
 828         }
 829 #endif /* INET */
 830 #ifdef INET6
 831         if (sc->sc_naddrs6) {
 832                 struct ip6_hdr *ip6;
 833
 834                 MGETHDR(m, M_NOWAIT, MT_HEADER);
 835                 if (m == NULL) {
 836                         CARPSTATS_INC(carps_onomem);
 837                         /* XXX maybe less ? */
 838                         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
 839                             carp_send_ad, sc);
 840                         return;
 841                 }
 842                 len = sizeof(*ip6) + sizeof(ch);
 843                 m->m_pkthdr.len = len;
 844                 m->m_pkthdr.rcvif = NULL;
 845                 m->m_len = len;
 846                 MH_ALIGN(m, m->m_len);
 847                 m->m_flags |= M_MCAST;
 848                 ip6 = mtod(m, struct ip6_hdr *);
 849                 bzero(ip6, sizeof(*ip6));
 850                 ip6->ip6_vfc |= IPV6_VERSION;
 851                 ip6->ip6_hlim = CARP_DFLTTL;
 852                 ip6->ip6_nxt = IPPROTO_CARP;
 853                 bzero(&sa, sizeof(sa));
 854
 855                 /* set the source address */
 856                 sa.sa_family = AF_INET6;
 857                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 858                 if (ifa != NULL) {
 859                         bcopy(IFA_IN6(ifa), &ip6->ip6_src,
 860                             sizeof(struct in6_addr));
 861                         ifa_free(ifa);
 862                 } else
 863                         /* This should never happen with IPv6. */
 864                         bzero(&ip6->ip6_src, sizeof(struct in6_addr));
 865
 866                 /* Set the multicast destination. */
 867                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
 868                 ip6->ip6_dst.s6_addr8[15] = 0x12;
 869                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
 870                         m_freem(m);
 871                         CARP_DEBUG("%s: in6_setscope failed\n", __func__);
 872                         return;
 873                 }
 874
 875                 ch_ptr = (struct carp_header *)(&ip6[1]);
 876                 bcopy(&ch, ch_ptr, sizeof(ch));
 877                 if (carp_prepare_ad(m, sc, ch_ptr))
 878                         return;
 879
 880                 m->m_data += sizeof(*ip6);
 881                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6));
 882                 m->m_data -= sizeof(*ip6);
 883
 884                 CARPSTATS_INC(carps_opackets6);
 885
 886                 if (ip6_output(m, NULL, NULL, 0,
 887                     &sc->sc_carpdev->if_carp->cif_im6o, NULL, NULL)) {
 888                         if (sc->sc_sendad_errors < INT_MAX)
 889                                 sc->sc_sendad_errors++;
 890                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS)
 891                                 carp_demote_adj(carp_senderr_adj,
 892                                     "send6 error");
 893                         sc->sc_sendad_success = 0;
 894                 } else {
 895                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 896                                 if (++sc->sc_sendad_success >=
 897                                     CARP_SENDAD_MIN_SUCCESS) {
 898                                         carp_demote_adj(-carp_senderr_adj,
 899                                             "send6 ok");
 900                                         sc->sc_sendad_errors = 0;
 901                                 }
 902                         } else
 903                                 sc->sc_sendad_errors = 0;
 904                 }
 905         }
 906 #endif /* INET6 */
 907
 908         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), carp_send_ad, sc);
 909 }
 910
 911 static void
 912 carp_addroute(struct carp_softc *sc)
 913 {
 914         struct ifaddr *ifa;
 915
 916         CARP_FOREACH_IFA(sc, ifa)
 917                 carp_ifa_addroute(ifa);
 918 }
 919
 920 static void
 921 carp_ifa_addroute(struct ifaddr *ifa)
 922 {
 923
 924         switch (ifa->ifa_addr->sa_family) {
 925 #ifdef INET
 926         case AF_INET:
 927                 in_addprefix(ifatoia(ifa), RTF_UP);
 928                 ifa_add_loopback_route(ifa,
 929                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 930                 break;
 931 #endif
 932 #ifdef INET6
 933         case AF_INET6:
 934                 ifa_add_loopback_route(ifa,
 935                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 936                 in6_ifaddloop(ifa);
 937                 break;
 938 #endif
 939         }
 940 }
 941
 942 static void
 943 carp_delroute(struct carp_softc *sc)
 944 {
 945         struct ifaddr *ifa;
 946
 947         CARP_FOREACH_IFA(sc, ifa)
 948                 carp_ifa_delroute(ifa);
 949 }
 950
 951 static void
 952 carp_ifa_delroute(struct ifaddr *ifa)
 953 {
 954
 955         switch (ifa->ifa_addr->sa_family) {
 956 #ifdef INET
 957         case AF_INET:
 958                 ifa_del_loopback_route(ifa,
 959                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 960                 in_scrubprefix(ifatoia(ifa), LLE_STATIC);
 961                 break;
 962 #endif
 963 #ifdef INET6
 964         case AF_INET6:
 965                 ifa_del_loopback_route(ifa,
 966                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 967                 in6_ifremloop(ifa);
 968                 break;
 969 #endif
 970         }
 971 }
 972
 973 #ifdef INET
 974 /*
 975  * Broadcast a gratuitous ARP request containing
 976  * the virtual router MAC address for each IP address
 977  * associated with the virtual router.
 978  */
 979 static void
 980 carp_send_arp(struct carp_softc *sc)
 981 {
 982         struct ifaddr *ifa;
 983
 984         CARP_FOREACH_IFA(sc, ifa)
 985                 if (ifa->ifa_addr->sa_family == AF_INET)
 986                         arp_ifinit2(sc->sc_carpdev, ifa, LLADDR(&sc->sc_addr));
 987 }
 988
 989 int
 990 carp_iamatch(struct ifaddr *ifa, uint8_t **enaddr)
 991 {
 992         struct carp_softc *sc = ifa->ifa_carp;
 993
 994         if (sc->sc_state == MASTER) {
 995                 *enaddr = LLADDR(&sc->sc_addr);
 996                 return (1);
 997         }
 998
 999         return (0);
1000 }
1001 #endif
1002
1003 #ifdef INET6
1004 static void
1005 carp_send_na(struct carp_softc *sc)
1006 {
1007         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1008         struct ifaddr *ifa;
1009         struct in6_addr *in6;
1010
1011         CARP_FOREACH_IFA(sc, ifa) {
1012                 if (ifa->ifa_addr->sa_family != AF_INET6)
1013                         continue;
1014
1015                 in6 = IFA_IN6(ifa);
1016                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1017                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1018                 DELAY(1000);    /* XXX */
1019         }
1020 }
1021
1022 struct ifaddr *
1023 carp_iamatch6(struct ifnet *ifp, struct in6_addr *taddr)
1024 {
1025         struct ifaddr *ifa;
1026
1027         IF_ADDR_RLOCK(ifp);
1028         IFNET_FOREACH_IFA(ifp, ifa)
1029                 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1030                     ifa->ifa_carp->sc_state == MASTER &&
1031                     IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1032                         ifa_ref(ifa);
1033                         IF_ADDR_RUNLOCK(ifp);
1034                         return (ifa);
1035                 }
1036         IF_ADDR_RUNLOCK(ifp);
1037
1038         return (NULL);
1039 }
1040
1041 caddr_t
1042 carp_macmatch6(struct ifnet *ifp, struct mbuf *m, const struct in6_addr *taddr)
1043 {
1044         struct ifaddr *ifa;
1045
1046         IF_ADDR_RLOCK(ifp);
1047         IFNET_FOREACH_IFA(ifp, ifa)
1048                 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1049                     IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1050                         struct carp_softc *sc = ifa->ifa_carp;
1051                         struct m_tag *mtag;
1052
1053                         IF_ADDR_RUNLOCK(ifp);
1054
1055                         mtag = m_tag_get(PACKET_TAG_CARP,
1056                             sizeof(struct ifnet *), M_NOWAIT);
1057                         if (mtag == NULL)
1058                                 /* Better a bit than nothing. */
1059                                 return (LLADDR(&sc->sc_addr));
1060
1061                         bcopy(&ifp, (caddr_t)(mtag + 1),
1062                             sizeof(struct ifnet *));
1063                         m_tag_prepend(m, mtag);
1064
1065                         return (LLADDR(&sc->sc_addr));
1066                 }
1067         IF_ADDR_RUNLOCK(ifp);
1068
1069         return (NULL);
1070 }
1071 #endif /* INET6 */
1072
1073 int
1074 carp_forus(struct ifnet *ifp, u_char *dhost)
1075 {
1076         struct carp_softc *sc;
1077         uint8_t *ena = dhost;
1078
1079         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1080                 return (0);
1081
1082         CIF_LOCK(ifp->if_carp);
1083         IFNET_FOREACH_CARP(ifp, sc) {
1084                 CARP_LOCK(sc);
1085                 if (sc->sc_state == MASTER && !bcmp(dhost, LLADDR(&sc->sc_addr),
1086                     ETHER_ADDR_LEN)) {
1087                         CARP_UNLOCK(sc);
1088                         CIF_UNLOCK(ifp->if_carp);
1089                         return (1);
1090                 }
1091                 CARP_UNLOCK(sc);
1092         }
1093         CIF_UNLOCK(ifp->if_carp);
1094
1095         return (0);
1096 }
1097
1098 /* Master down timeout event, executed in callout context. */
1099 static void
1100 carp_master_down(void *v)
1101 {
1102         struct carp_softc *sc = v;
1103
1104         CARP_LOCK_ASSERT(sc);
1105
1106         CURVNET_SET(sc->sc_carpdev->if_vnet);
1107         if (sc->sc_state == BACKUP) {
1108                 CARP_LOG("VHID %u@%s: BACKUP -> MASTER (master down)\n",
1109                     sc->sc_vhid,
1110                     sc->sc_carpdev->if_xname);
1111                 carp_master_down_locked(sc);
1112         }
1113         CURVNET_RESTORE();
1114
1115         CARP_UNLOCK(sc);
1116 }
1117
1118 static void
1119 carp_master_down_locked(struct carp_softc *sc)
1120 {
1121
1122         CARP_LOCK_ASSERT(sc);
1123
1124         switch (sc->sc_state) {
1125         case BACKUP:
1126                 carp_set_state(sc, MASTER);
1127                 carp_send_ad_locked(sc);
1128 #ifdef INET
1129                 carp_send_arp(sc);
1130 #endif
1131 #ifdef INET6
1132                 carp_send_na(sc);
1133 #endif
1134                 carp_setrun(sc, 0);
1135                 carp_addroute(sc);
1136                 break;
1137         case INIT:
1138         case MASTER:
1139 #ifdef INVARIANTS
1140                 panic("carp: VHID %u@%s: master_down event in %s state\n",
1141                     sc->sc_vhid,
1142                     sc->sc_carpdev->if_xname,
1143                     sc->sc_state ? "MASTER" : "INIT");
1144 #endif
1145                 break;
1146         }
1147 }
1148
1149 /*
1150  * When in backup state, af indicates whether to reset the master down timer
1151  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1152  */
1153 static void
1154 carp_setrun(struct carp_softc *sc, sa_family_t af)
1155 {
1156         struct timeval tv;
1157
1158         CARP_LOCK_ASSERT(sc);
1159
1160         if ((sc->sc_carpdev->if_flags & IFF_UP) == 0 ||
1161             sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1162             (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0))
1163                 return;
1164
1165         switch (sc->sc_state) {
1166         case INIT:
1167                 CARP_LOG("VHID %u@%s: INIT -> BACKUP\n",
1168                     sc->sc_vhid,
1169                     sc->sc_carpdev->if_xname);
1170                 carp_set_state(sc, BACKUP);
1171                 carp_setrun(sc, 0);
1172                 break;
1173         case BACKUP:
1174                 callout_stop(&sc->sc_ad_tmo);
1175                 tv.tv_sec = 3 * sc->sc_advbase;
1176                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1177                 switch (af) {
1178 #ifdef INET
1179                 case AF_INET:
1180                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1181                             carp_master_down, sc);
1182                         break;
1183 #endif
1184 #ifdef INET6
1185                 case AF_INET6:
1186                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1187                             carp_master_down, sc);
1188                         break;
1189 #endif
1190                 default:
1191 #ifdef INET
1192                         if (sc->sc_naddrs)
1193                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1194                                     carp_master_down, sc);
1195 #endif
1196 #ifdef INET6
1197                         if (sc->sc_naddrs6)
1198                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1199                                     carp_master_down, sc);
1200 #endif
1201                         break;
1202                 }
1203                 break;
1204         case MASTER:
1205                 tv.tv_sec = sc->sc_advbase;
1206                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1207                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1208                     carp_send_ad, sc);
1209                 break;
1210         }
1211 }
1212
1213 /*
1214  * Setup multicast structures.
1215  */
1216 static int
1217 carp_multicast_setup(struct carp_softc *sc, sa_family_t sa)
1218 {
1219         struct ifnet *ifp = sc->sc_carpdev;
1220         struct carp_if *cif = ifp->if_carp;
1221         int error = 0;
1222
1223         switch (sa) {
1224 #ifdef INET
1225         case AF_INET:
1226             {
1227                 struct ip_moptions *imo = &cif->cif_imo;
1228                 struct in_addr addr;
1229
1230                 if (imo->imo_membership)
1231                         return (0);
1232
1233                 imo->imo_membership = (struct in_multi **)malloc(
1234                     (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
1235                     M_WAITOK);
1236                 imo->imo_mfilters = NULL;
1237                 imo->imo_max_memberships = IP_MIN_MEMBERSHIPS;
1238                 imo->imo_multicast_vif = -1;
1239
1240                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1241                 if ((error = in_joingroup(ifp, &addr, NULL,
1242                     &imo->imo_membership[0])) != 0) {
1243                         free(imo->imo_membership, M_CARP);
1244                         break;
1245                 }
1246                 imo->imo_num_memberships++;
1247                 imo->imo_multicast_ifp = ifp;
1248                 imo->imo_multicast_ttl = CARP_DFLTTL;
1249                 imo->imo_multicast_loop = 0;
1250                 break;
1251            }
1252 #endif
1253 #ifdef INET6
1254         case AF_INET6:
1255             {
1256                 struct ip6_moptions *im6o = &cif->cif_im6o;
1257                 struct in6_addr in6;
1258                 struct in6_multi *in6m;
1259
1260                 if (im6o->im6o_membership)
1261                         return (0);
1262
1263                 im6o->im6o_membership = (struct in6_multi **)malloc(
1264                     (sizeof(struct in6_multi *) * IPV6_MIN_MEMBERSHIPS), M_CARP,
1265                     M_ZERO|M_WAITOK);
1266                 im6o->im6o_mfilters = NULL;
1267                 im6o->im6o_max_memberships = IPV6_MIN_MEMBERSHIPS;
1268                 im6o->im6o_multicast_hlim = CARP_DFLTTL;
1269                 im6o->im6o_multicast_ifp = ifp;
1270
1271                 /* Join IPv6 CARP multicast group. */
1272                 bzero(&in6, sizeof(in6));
1273                 in6.s6_addr16[0] = htons(0xff02);
1274                 in6.s6_addr8[15] = 0x12;
1275                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1276                         free(im6o->im6o_membership, M_CARP);
1277                         break;
1278                 }
1279                 in6m = NULL;
1280                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1281                         free(im6o->im6o_membership, M_CARP);
1282                         break;
1283                 }
1284                 im6o->im6o_membership[0] = in6m;
1285                 im6o->im6o_num_memberships++;
1286
1287                 /* Join solicited multicast address. */
1288                 bzero(&in6, sizeof(in6));
1289                 in6.s6_addr16[0] = htons(0xff02);
1290                 in6.s6_addr32[1] = 0;
1291                 in6.s6_addr32[2] = htonl(1);
1292                 in6.s6_addr32[3] = 0;
1293                 in6.s6_addr8[12] = 0xff;
1294                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1295                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1296                         free(im6o->im6o_membership, M_CARP);
1297                         break;
1298                 }
1299                 in6m = NULL;
1300                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1301                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1302                         free(im6o->im6o_membership, M_CARP);
1303                         break;
1304                 }
1305                 im6o->im6o_membership[1] = in6m;
1306                 im6o->im6o_num_memberships++;
1307                 break;
1308             }
1309 #endif
1310         }
1311
1312         return (error);
1313 }
1314
1315 /*
1316  * Free multicast structures.
1317  */
1318 static void
1319 carp_multicast_cleanup(struct carp_softc *sc, sa_family_t sa)
1320 {
1321         struct ifnet *ifp = sc->sc_carpdev;
1322         struct carp_if *cif = ifp->if_carp;
1323
1324         switch (sa) {
1325 #ifdef INET
1326         case AF_INET:
1327                 if (sc->sc_naddrs == 0) {
1328                         struct ip_moptions *imo = &cif->cif_imo;
1329
1330                         in_leavegroup(imo->imo_membership[0], NULL);
1331                         KASSERT(imo->imo_mfilters == NULL,
1332                             ("%s: imo_mfilters != NULL", __func__));
1333                         free(imo->imo_membership, M_CARP);
1334                         imo->imo_membership = NULL;
1335
1336                 }
1337                 break;
1338 #endif
1339 #ifdef INET6
1340         case AF_INET6:
1341                 if (sc->sc_naddrs6 == 0) {
1342                         struct ip6_moptions *im6o = &cif->cif_im6o;
1343
1344                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1345                         in6_mc_leave(im6o->im6o_membership[1], NULL);
1346                         KASSERT(im6o->im6o_mfilters == NULL,
1347                             ("%s: im6o_mfilters != NULL", __func__));
1348                         free(im6o->im6o_membership, M_CARP);
1349                         im6o->im6o_membership = NULL;
1350                 }
1351                 break;
1352 #endif
1353         }
1354 }
1355
1356 int
1357 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *sa)
1358 {
1359         struct m_tag *mtag;
1360         struct carp_softc *sc;
1361
1362         if (!sa)
1363                 return (0);
1364
1365         switch (sa->sa_family) {
1366 #ifdef INET
1367         case AF_INET:
1368                 break;
1369 #endif
1370 #ifdef INET6
1371         case AF_INET6:
1372                 break;
1373 #endif
1374         default:
1375                 return (0);
1376         }
1377
1378         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
1379         if (mtag == NULL)
1380                 return (0);
1381
1382         bcopy(mtag + 1, &sc, sizeof(struct carp_softc *));
1383
1384         /* Set the source MAC address to the Virtual Router MAC Address. */
1385         switch (ifp->if_type) {
1386         case IFT_ETHER:
1387         case IFT_L2VLAN: {
1388                         struct ether_header *eh;
1389
1390                         eh = mtod(m, struct ether_header *);
1391                         eh->ether_shost[0] = 0;
1392                         eh->ether_shost[1] = 0;
1393                         eh->ether_shost[2] = 0x5e;
1394                         eh->ether_shost[3] = 0;
1395                         eh->ether_shost[4] = 1;
1396                         eh->ether_shost[5] = sc->sc_vhid;
1397                 }
1398                 break;
1399         case IFT_FDDI: {
1400                         struct fddi_header *fh;
1401
1402                         fh = mtod(m, struct fddi_header *);
1403                         fh->fddi_shost[0] = 0;
1404                         fh->fddi_shost[1] = 0;
1405                         fh->fddi_shost[2] = 0x5e;
1406                         fh->fddi_shost[3] = 0;
1407                         fh->fddi_shost[4] = 1;
1408                         fh->fddi_shost[5] = sc->sc_vhid;
1409                 }
1410                 break;
1411         case IFT_ISO88025: {
1412                         struct iso88025_header *th;
1413                         th = mtod(m, struct iso88025_header *);
1414                         th->iso88025_shost[0] = 3;
1415                         th->iso88025_shost[1] = 0;
1416                         th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
1417                         th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
1418                         th->iso88025_shost[4] = 0;
1419                         th->iso88025_shost[5] = 0;
1420                 }
1421                 break;
1422         default:
1423                 printf("%s: carp is not supported for the %d interface type\n",
1424                     ifp->if_xname, ifp->if_type);
1425                 return (EOPNOTSUPP);
1426         }
1427
1428         return (0);
1429 }
1430
1431 static struct carp_softc*
1432 carp_alloc(struct ifnet *ifp)
1433 {
1434         struct carp_softc *sc;
1435         struct carp_if *cif;
1436
1437         if ((cif = ifp->if_carp) == NULL) {
1438                 cif = carp_alloc_if(ifp);
1439                 if (cif == NULL)
1440                         return (NULL);
1441         }
1442
1443         sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
1444
1445         sc->sc_advbase = CARP_DFLTINTV;
1446         sc->sc_vhid = -1;       /* required setting */
1447         sc->sc_init_counter = 1;
1448         sc->sc_state = INIT;
1449
1450         sc->sc_ifasiz = sizeof(struct ifaddr *);
1451         sc->sc_ifas = malloc(sc->sc_ifasiz, M_CARP, M_WAITOK|M_ZERO);
1452         sc->sc_carpdev = ifp;
1453
1454         CARP_LOCK_INIT(sc);
1455 #ifdef INET
1456         callout_init_mtx(&sc->sc_md_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1457 #endif
1458 #ifdef INET6
1459         callout_init_mtx(&sc->sc_md6_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1460 #endif
1461         callout_init_mtx(&sc->sc_ad_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1462
1463         CIF_LOCK(cif);
1464         TAILQ_INSERT_TAIL(&cif->cif_vrs, sc, sc_list);
1465         CIF_UNLOCK(cif);
1466
1467         mtx_lock(&carp_mtx);
1468         LIST_INSERT_HEAD(&carp_list, sc, sc_next);
1469         mtx_unlock(&carp_mtx);
1470
1471         return (sc);
1472 }
1473
1474 static int
1475 carp_grow_ifas(struct carp_softc *sc)
1476 {
1477         struct ifaddr **new;
1478
1479         CARP_LOCK_ASSERT(sc);
1480
1481         new = malloc(sc->sc_ifasiz * 2, M_CARP, M_NOWAIT|M_ZERO);
1482         if (new == NULL)
1483                 return (ENOMEM);
1484         bcopy(sc->sc_ifas, new, sc->sc_ifasiz);
1485         free(sc->sc_ifas, M_CARP);
1486         sc->sc_ifas = new;
1487         sc->sc_ifasiz *= 2;
1488
1489         return (0);
1490 }
1491
1492 static void
1493 carp_destroy(struct carp_softc *sc)
1494 {
1495         struct ifnet *ifp = sc->sc_carpdev;
1496         struct carp_if *cif = ifp->if_carp;
1497
1498         CIF_LOCK(cif);
1499         TAILQ_REMOVE(&cif->cif_vrs, sc, sc_list);
1500         if (TAILQ_EMPTY(&cif->cif_vrs))
1501                 carp_free_if(cif);
1502         else
1503                 CIF_UNLOCK(cif);
1504
1505         mtx_lock(&carp_mtx);
1506         LIST_REMOVE(sc, sc_next);
1507         mtx_unlock(&carp_mtx);
1508
1509         CARP_LOCK(sc);
1510         if (sc->sc_suppress)
1511                 carp_demote_adj(-carp_ifdown_adj, "vhid removed");
1512         callout_drain(&sc->sc_ad_tmo);
1513 #ifdef INET
1514         callout_drain(&sc->sc_md_tmo);
1515 #endif
1516 #ifdef INET6
1517         callout_drain(&sc->sc_md6_tmo);
1518 #endif
1519         CARP_LOCK_DESTROY(sc);
1520
1521         free(sc->sc_ifas, M_CARP);
1522         free(sc, M_CARP);
1523 }
1524
1525 static struct carp_if*
1526 carp_alloc_if(struct ifnet *ifp)
1527 {
1528         struct carp_if *cif;
1529
1530         cif = malloc(sizeof(*cif), M_CARP, M_WAITOK|M_ZERO);
1531
1532         if (ifpromisc(ifp, 1) != 0)
1533                 goto cleanup;
1534
1535         CIF_LOCK_INIT(cif);
1536         cif->cif_ifp = ifp;
1537         TAILQ_INIT(&cif->cif_vrs);
1538
1539         IF_ADDR_WLOCK(ifp);
1540         ifp->if_carp = cif;
1541         if_ref(ifp);
1542         IF_ADDR_WUNLOCK(ifp);
1543
1544         return (cif);
1545
1546 cleanup:
1547         free(cif, M_CARP);
1548
1549         return (NULL);
1550 }
1551
1552 static void
1553 carp_free_if(struct carp_if *cif)
1554 {
1555         struct ifnet *ifp = cif->cif_ifp;
1556
1557         CIF_LOCK_ASSERT(cif);
1558         KASSERT(TAILQ_EMPTY(&cif->cif_vrs), ("%s: softc list not empty",
1559             __func__));
1560
1561         IF_ADDR_WLOCK(ifp);
1562         ifp->if_carp = NULL;
1563         if_rele(ifp);
1564         IF_ADDR_WUNLOCK(ifp);
1565
1566         CIF_LOCK_DESTROY(cif);
1567
1568         ifpromisc(ifp, 0);
1569
1570         free(cif, M_CARP);
1571 }
1572
1573 static void
1574 carp_carprcp(struct carpreq *carpr, struct carp_softc *sc, int priv)
1575 {
1576
1577         CARP_LOCK(sc);
1578         carpr->carpr_state = sc->sc_state;
1579         carpr->carpr_vhid = sc->sc_vhid;
1580         carpr->carpr_advbase = sc->sc_advbase;
1581         carpr->carpr_advskew = sc->sc_advskew;
1582         if (priv)
1583                 bcopy(sc->sc_key, carpr->carpr_key, sizeof(carpr->carpr_key));
1584         else
1585                 bzero(carpr->carpr_key, sizeof(carpr->carpr_key));
1586         CARP_UNLOCK(sc);
1587 }
1588
1589 int
1590 carp_ioctl(struct ifreq *ifr, u_long cmd, struct thread *td)
1591 {
1592         struct carpreq carpr;
1593         struct ifnet *ifp;
1594         struct carp_softc *sc = NULL;
1595         int error = 0, locked = 0;
1596
1597         if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1598                 return (error);
1599
1600         ifp = ifunit_ref(ifr->ifr_name);
1601         if (ifp == NULL)
1602                 return (ENXIO);
1603
1604         switch (ifp->if_type) {
1605         case IFT_ETHER:
1606         case IFT_L2VLAN:
1607         case IFT_FDDI:
1608         case IFT_ISO88025:
1609                 break;
1610         default:
1611                 error = EOPNOTSUPP;
1612                 goto out;
1613         }
1614
1615         if ((ifp->if_flags & IFF_MULTICAST) == 0) {
1616                 error = EADDRNOTAVAIL;
1617                 goto out;
1618         }
1619
1620         switch (cmd) {
1621         case SIOCSVH:
1622                 if ((error = priv_check(td, PRIV_NETINET_CARP)))
1623                         break;
1624                 if (carpr.carpr_vhid <= 0 || carpr.carpr_vhid > CARP_MAXVHID ||
1625                     carpr.carpr_advbase < 0 || carpr.carpr_advskew < 0) {
1626                         error = EINVAL;
1627                         break;
1628                 }
1629
1630                 if (ifp->if_carp) {
1631                         CIF_LOCK(ifp->if_carp);
1632                         IFNET_FOREACH_CARP(ifp, sc)
1633                                 if (sc->sc_vhid == carpr.carpr_vhid)
1634                                         break;
1635                         CIF_UNLOCK(ifp->if_carp);
1636                 }
1637                 if (sc == NULL) {
1638                         sc = carp_alloc(ifp);
1639                         if (sc == NULL) {
1640                                 error = EINVAL; /* XXX: ifpromisc failed */
1641                                 break;
1642                         }
1643
1644                         CARP_LOCK(sc);
1645                         sc->sc_vhid = carpr.carpr_vhid;
1646                         LLADDR(&sc->sc_addr)[0] = 0;
1647                         LLADDR(&sc->sc_addr)[1] = 0;
1648                         LLADDR(&sc->sc_addr)[2] = 0x5e;
1649                         LLADDR(&sc->sc_addr)[3] = 0;
1650                         LLADDR(&sc->sc_addr)[4] = 1;
1651                         LLADDR(&sc->sc_addr)[5] = sc->sc_vhid;
1652                 } else
1653                         CARP_LOCK(sc);
1654                 locked = 1;
1655                 if (carpr.carpr_advbase > 0) {
1656                         if (carpr.carpr_advbase > 255 ||
1657                             carpr.carpr_advbase < CARP_DFLTINTV) {
1658                                 error = EINVAL;
1659                                 break;
1660                         }
1661                         sc->sc_advbase = carpr.carpr_advbase;
1662                 }
1663                 if (carpr.carpr_advskew > 0) {
1664                         if (carpr.carpr_advskew >= 255) {
1665                                 error = EINVAL;
1666                                 break;
1667                         }
1668                         sc->sc_advskew = carpr.carpr_advskew;
1669                 }
1670                 if (carpr.carpr_key[0] != '\0') {
1671                         bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1672                         carp_hmac_prepare(sc);
1673                 }
1674                 if (sc->sc_state != INIT &&
1675                     carpr.carpr_state != sc->sc_state) {
1676                         switch (carpr.carpr_state) {
1677                         case BACKUP:
1678                                 callout_stop(&sc->sc_ad_tmo);
1679                                 carp_set_state(sc, BACKUP);
1680                                 carp_setrun(sc, 0);
1681                                 carp_delroute(sc);
1682                                 break;
1683                         case MASTER:
1684                                 carp_master_down_locked(sc);
1685                                 break;
1686                         default:
1687                                 break;
1688                         }
1689                 }
1690                 break;
1691
1692         case SIOCGVH:
1693             {
1694                 int priveleged;
1695
1696                 if (carpr.carpr_vhid < 0 || carpr.carpr_vhid > CARP_MAXVHID) {
1697                         error = EINVAL;
1698                         break;
1699                 }
1700                 if (carpr.carpr_count < 1) {
1701                         error = EMSGSIZE;
1702                         break;
1703                 }
1704                 if (ifp->if_carp == NULL) {
1705                         error = ENOENT;
1706                         break;
1707                 }
1708
1709                 priveleged = (priv_check(td, PRIV_NETINET_CARP) == 0);
1710                 if (carpr.carpr_vhid != 0) {
1711                         CIF_LOCK(ifp->if_carp);
1712                         IFNET_FOREACH_CARP(ifp, sc)
1713                                 if (sc->sc_vhid == carpr.carpr_vhid)
1714                                         break;
1715                         CIF_UNLOCK(ifp->if_carp);
1716                         if (sc == NULL) {
1717                                 error = ENOENT;
1718                                 break;
1719                         }
1720                         carp_carprcp(&carpr, sc, priveleged);
1721                         error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1722                 } else  {
1723                         int i, count;
1724
1725                         count = 0;
1726                         CIF_LOCK(ifp->if_carp);
1727                         IFNET_FOREACH_CARP(ifp, sc)
1728                                 count++;
1729
1730                         if (count > carpr.carpr_count) {
1731                                 CIF_UNLOCK(ifp->if_carp);
1732                                 error = EMSGSIZE;
1733                                 break;
1734                         }
1735
1736                         i = 0;
1737                         IFNET_FOREACH_CARP(ifp, sc) {
1738                                 carp_carprcp(&carpr, sc, priveleged);
1739                                 carpr.carpr_count = count;
1740                                 error = copyout(&carpr, ifr->ifr_data +
1741                                     (i * sizeof(carpr)), sizeof(carpr));
1742                                 if (error) {
1743                                         CIF_UNLOCK(ifp->if_carp);
1744                                         break;
1745                                 }
1746                                 i++;
1747                         }
1748                         CIF_UNLOCK(ifp->if_carp);
1749                 }
1750                 break;
1751             }
1752         default:
1753                 error = EINVAL;
1754         }
1755
1756 out:
1757         if (locked)
1758                 CARP_UNLOCK(sc);
1759         if_rele(ifp);
1760
1761         return (error);
1762 }
1763
1764 static int
1765 carp_get_vhid(struct ifaddr *ifa)
1766 {
1767
1768         if (ifa == NULL || ifa->ifa_carp == NULL)
1769                 return (0);
1770
1771         return (ifa->ifa_carp->sc_vhid);
1772 }
1773
1774 int
1775 carp_attach(struct ifaddr *ifa, int vhid)
1776 {
1777         struct ifnet *ifp = ifa->ifa_ifp;
1778         struct carp_softc *sc;
1779         int index, error;
1780
1781         if (ifp->if_carp == NULL)
1782                 return (ENOPROTOOPT);
1783
1784         switch (ifa->ifa_addr->sa_family) {
1785 #ifdef INET
1786         case AF_INET:
1787 #endif
1788 #ifdef INET6
1789         case AF_INET6:
1790 #endif
1791                 break;
1792         default:
1793                 return (EPROTOTYPE);
1794         }
1795
1796         CIF_LOCK(ifp->if_carp);
1797         IFNET_FOREACH_CARP(ifp, sc)
1798                 if (sc->sc_vhid == vhid)
1799                         break;
1800         CIF_UNLOCK(ifp->if_carp);
1801         if (sc == NULL)
1802                 return (ENOENT);
1803
1804         if (ifa->ifa_carp) {
1805                 if (ifa->ifa_carp->sc_vhid != vhid)
1806                         carp_detach(ifa);
1807                 else
1808                         return (0);
1809         }
1810
1811         error = carp_multicast_setup(sc, ifa->ifa_addr->sa_family);
1812         if (error)
1813                 return (error);
1814
1815         CARP_LOCK(sc);
1816         index = sc->sc_naddrs + sc->sc_naddrs6 + 1;
1817         if (index > sc->sc_ifasiz / sizeof(struct ifaddr *))
1818                 if ((error = carp_grow_ifas(sc)) != 0) {
1819                         carp_multicast_cleanup(sc,
1820                             ifa->ifa_addr->sa_family);
1821                         CARP_UNLOCK(sc);
1822                         return (error);
1823                 }
1824
1825         switch (ifa->ifa_addr->sa_family) {
1826 #ifdef INET
1827         case AF_INET:
1828                 sc->sc_naddrs++;
1829                 break;
1830 #endif
1831 #ifdef INET6
1832         case AF_INET6:
1833                 sc->sc_naddrs6++;
1834                 break;
1835 #endif
1836         }
1837
1838         ifa_ref(ifa);
1839         sc->sc_ifas[index - 1] = ifa;
1840         ifa->ifa_carp = sc;
1841
1842         carp_hmac_prepare(sc);
1843         carp_sc_state(sc);
1844
1845         CARP_UNLOCK(sc);
1846
1847         return (0);
1848 }
1849
1850 void
1851 carp_detach(struct ifaddr *ifa)
1852 {
1853         struct carp_softc *sc = ifa->ifa_carp;
1854         int i, index;
1855
1856         KASSERT(sc != NULL, ("%s: %p not attached", __func__, ifa));
1857
1858         CARP_LOCK(sc);
1859
1860         /* Shift array. */
1861         index = sc->sc_naddrs + sc->sc_naddrs6;
1862         for (i = 0; i < index; i++)
1863                 if (sc->sc_ifas[i] == ifa)
1864                         break;
1865         KASSERT(i < index, ("%s: %p no backref", __func__, ifa));
1866         for (; i < index - 1; i++)
1867                 sc->sc_ifas[i] = sc->sc_ifas[i+1];
1868         sc->sc_ifas[index - 1] = NULL;
1869
1870         switch (ifa->ifa_addr->sa_family) {
1871 #ifdef INET
1872         case AF_INET:
1873                 sc->sc_naddrs--;
1874                 break;
1875 #endif
1876 #ifdef INET6
1877         case AF_INET6:
1878                 sc->sc_naddrs6--;
1879                 break;
1880 #endif
1881         }
1882
1883         carp_ifa_delroute(ifa);
1884         carp_multicast_cleanup(sc, ifa->ifa_addr->sa_family);
1885
1886         ifa->ifa_carp = NULL;
1887         ifa_free(ifa);
1888
1889         carp_hmac_prepare(sc);
1890         carp_sc_state(sc);
1891
1892         if (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) {
1893                 CARP_UNLOCK(sc);
1894                 carp_destroy(sc);
1895         } else
1896                 CARP_UNLOCK(sc);
1897 }
1898
1899 static void
1900 carp_set_state(struct carp_softc *sc, int state)
1901 {
1902
1903         CARP_LOCK_ASSERT(sc);
1904
1905         if (sc->sc_state != state) {
1906                 const char *carp_states[] = { CARP_STATES };
1907                 char subsys[IFNAMSIZ+5];
1908
1909                 sc->sc_state = state;
1910
1911                 snprintf(subsys, IFNAMSIZ+5, "%u@%s", sc->sc_vhid,
1912                     sc->sc_carpdev->if_xname);
1913                 devctl_notify("CARP", subsys, carp_states[state], NULL);
1914         }
1915 }
1916
1917 static void
1918 carp_linkstate(struct ifnet *ifp)
1919 {
1920         struct carp_softc *sc;
1921
1922         CIF_LOCK(ifp->if_carp);
1923         IFNET_FOREACH_CARP(ifp, sc) {
1924                 CARP_LOCK(sc);
1925                 carp_sc_state(sc);
1926                 CARP_UNLOCK(sc);
1927         }
1928         CIF_UNLOCK(ifp->if_carp);
1929 }
1930
1931 static void
1932 carp_sc_state(struct carp_softc *sc)
1933 {
1934
1935         CARP_LOCK_ASSERT(sc);
1936
1937         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1938             !(sc->sc_carpdev->if_flags & IFF_UP)) {
1939                 callout_stop(&sc->sc_ad_tmo);
1940 #ifdef INET
1941                 callout_stop(&sc->sc_md_tmo);
1942 #endif
1943 #ifdef INET6
1944                 callout_stop(&sc->sc_md6_tmo);
1945 #endif
1946                 carp_set_state(sc, INIT);
1947                 carp_setrun(sc, 0);
1948                 if (!sc->sc_suppress)
1949                         carp_demote_adj(carp_ifdown_adj, "interface down");
1950                 sc->sc_suppress = 1;
1951         } else {
1952                 carp_set_state(sc, INIT);
1953                 carp_setrun(sc, 0);
1954                 if (sc->sc_suppress)
1955                         carp_demote_adj(-carp_ifdown_adj, "interface up");
1956                 sc->sc_suppress = 0;
1957         }
1958 }
1959
1960 static void
1961 carp_demote_adj(int adj, char *reason)
1962 {
1963         carp_demotion += adj;
1964         CARP_LOG("demoted by %d to %d (%s)\n", adj, carp_demotion, reason);
1965         taskqueue_enqueue(taskqueue_swi, &carp_sendall_task);
1966 }
1967
1968 #ifdef INET
1969 extern  struct domain inetdomain;
1970 static struct protosw in_carp_protosw = {
1971         .pr_type =              SOCK_RAW,
1972         .pr_domain =            &inetdomain,
1973         .pr_protocol =          IPPROTO_CARP,
1974         .pr_flags =             PR_ATOMIC|PR_ADDR,
1975         .pr_input =             carp_input,
1976         .pr_output =            (pr_output_t *)rip_output,
1977         .pr_ctloutput =         rip_ctloutput,
1978         .pr_usrreqs =           &rip_usrreqs
1979 };
1980 #endif
1981
1982 #ifdef INET6
1983 extern  struct domain inet6domain;
1984 static struct ip6protosw in6_carp_protosw = {
1985         .pr_type =              SOCK_RAW,
1986         .pr_domain =            &inet6domain,
1987         .pr_protocol =          IPPROTO_CARP,
1988         .pr_flags =             PR_ATOMIC|PR_ADDR,
1989         .pr_input =             carp6_input,
1990         .pr_output =            rip6_output,
1991         .pr_ctloutput =         rip6_ctloutput,
1992         .pr_usrreqs =           &rip6_usrreqs
1993 };
1994 #endif
1995
1996 static void
1997 carp_mod_cleanup(void)
1998 {
1999
2000 #ifdef INET
2001         if (proto_reg[CARP_INET] == 0) {
2002                 (void)ipproto_unregister(IPPROTO_CARP);
2003                 pf_proto_unregister(PF_INET, IPPROTO_CARP, SOCK_RAW);
2004                 proto_reg[CARP_INET] = -1;
2005         }
2006         carp_iamatch_p = NULL;
2007 #endif
2008 #ifdef INET6
2009         if (proto_reg[CARP_INET6] == 0) {
2010                 (void)ip6proto_unregister(IPPROTO_CARP);
2011                 pf_proto_unregister(PF_INET6, IPPROTO_CARP, SOCK_RAW);
2012                 proto_reg[CARP_INET6] = -1;
2013         }
2014         carp_iamatch6_p = NULL;
2015         carp_macmatch6_p = NULL;
2016 #endif
2017         carp_ioctl_p = NULL;
2018         carp_attach_p = NULL;
2019         carp_detach_p = NULL;
2020         carp_get_vhid_p = NULL;
2021         carp_linkstate_p = NULL;
2022         carp_forus_p = NULL;
2023         carp_output_p = NULL;
2024         carp_demote_adj_p = NULL;
2025         mtx_unlock(&carp_mtx);
2026         taskqueue_drain(taskqueue_swi, &carp_sendall_task);
2027         mtx_destroy(&carp_mtx);
2028 }
2029
2030 static int
2031 carp_mod_load(void)
2032 {
2033         int err;
2034
2035         mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2036         LIST_INIT(&carp_list);
2037         carp_get_vhid_p = carp_get_vhid;
2038         carp_forus_p = carp_forus;
2039         carp_output_p = carp_output;
2040         carp_linkstate_p = carp_linkstate;
2041         carp_ioctl_p = carp_ioctl;
2042         carp_attach_p = carp_attach;
2043         carp_detach_p = carp_detach;
2044         carp_demote_adj_p = carp_demote_adj;
2045 #ifdef INET6
2046         carp_iamatch6_p = carp_iamatch6;
2047         carp_macmatch6_p = carp_macmatch6;
2048         proto_reg[CARP_INET6] = pf_proto_register(PF_INET6,
2049             (struct protosw *)&in6_carp_protosw);
2050         if (proto_reg[CARP_INET6]) {
2051                 printf("carp: error %d attaching to PF_INET6\n",
2052                     proto_reg[CARP_INET6]);
2053                 carp_mod_cleanup();
2054                 return (proto_reg[CARP_INET6]);
2055         }
2056         err = ip6proto_register(IPPROTO_CARP);
2057         if (err) {
2058                 printf("carp: error %d registering with INET6\n", err);
2059                 carp_mod_cleanup();
2060                 return (err);
2061         }
2062 #endif
2063 #ifdef INET
2064         carp_iamatch_p = carp_iamatch;
2065         proto_reg[CARP_INET] = pf_proto_register(PF_INET, &in_carp_protosw);
2066         if (proto_reg[CARP_INET]) {
2067                 printf("carp: error %d attaching to PF_INET\n",
2068                     proto_reg[CARP_INET]);
2069                 carp_mod_cleanup();
2070                 return (proto_reg[CARP_INET]);
2071         }
2072         err = ipproto_register(IPPROTO_CARP);
2073         if (err) {
2074                 printf("carp: error %d registering with INET\n", err);
2075                 carp_mod_cleanup();
2076                 return (err);
2077         }
2078 #endif
2079         return (0);
2080 }
2081
2082 static int
2083 carp_modevent(module_t mod, int type, void *data)
2084 {
2085         switch (type) {
2086         case MOD_LOAD:
2087                 return carp_mod_load();
2088                 /* NOTREACHED */
2089         case MOD_UNLOAD:
2090                 mtx_lock(&carp_mtx);
2091                 if (LIST_EMPTY(&carp_list))
2092                         carp_mod_cleanup();
2093                 else {
2094                         mtx_unlock(&carp_mtx);
2095                         return (EBUSY);
2096                 }
2097                 break;
2098
2099         default:
2100                 return (EINVAL);
2101         }
2102
2103         return (0);
2104 }
2105
2106 static moduledata_t carp_mod = {
2107         "carp",
2108         carp_modevent,
2109         0
2110 };
2111
2112 DECLARE_MODULE(carp, carp_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);