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                         carp_send_ad_locked(sc);
 711                         CARP_UNLOCK(sc);
 712                 }
 713         mtx_unlock(&carp_mtx);
 714 }
 715
 716 static void
 717 carp_send_ad(void *v)
 718 {
 719         struct carp_softc *sc = v;
 720
 721         CARP_LOCK_ASSERT(sc);
 722         carp_send_ad_locked(sc);
 723         CARP_UNLOCK(sc);
 724 }
 725
 726 static void
 727 carp_send_ad_locked(struct carp_softc *sc)
 728 {
 729         struct carp_header ch;
 730         struct timeval tv;
 731         struct sockaddr sa;
 732         struct ifaddr *ifa;
 733         struct carp_header *ch_ptr;
 734         struct mbuf *m;
 735         int len, advskew;
 736
 737         CARP_LOCK_ASSERT(sc);
 738
 739         advskew = DEMOTE_ADVSKEW(sc);
 740         tv.tv_sec = sc->sc_advbase;
 741         tv.tv_usec = advskew * 1000000 / 256;
 742
 743         ch.carp_version = CARP_VERSION;
 744         ch.carp_type = CARP_ADVERTISEMENT;
 745         ch.carp_vhid = sc->sc_vhid;
 746         ch.carp_advbase = sc->sc_advbase;
 747         ch.carp_advskew = advskew;
 748         ch.carp_authlen = 7;    /* XXX DEFINE */
 749         ch.carp_pad1 = 0;       /* must be zero */
 750         ch.carp_cksum = 0;
 751
 752         /* XXXGL: OpenBSD picks first ifaddr with needed family. */
 753
 754 #ifdef INET
 755         if (sc->sc_naddrs) {
 756                 struct ip *ip;
 757
 758                 MGETHDR(m, M_NOWAIT, MT_HEADER);
 759                 if (m == NULL) {
 760                         CARPSTATS_INC(carps_onomem);
 761                         /* XXX maybe less ? */
 762                         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
 763                             carp_send_ad, sc);
 764                         return;
 765                 }
 766                 len = sizeof(*ip) + sizeof(ch);
 767                 m->m_pkthdr.len = len;
 768                 m->m_pkthdr.rcvif = NULL;
 769                 m->m_len = len;
 770                 MH_ALIGN(m, m->m_len);
 771                 m->m_flags |= M_MCAST;
 772                 ip = mtod(m, struct ip *);
 773                 ip->ip_v = IPVERSION;
 774                 ip->ip_hl = sizeof(*ip) >> 2;
 775                 ip->ip_tos = IPTOS_LOWDELAY;
 776                 ip->ip_len = len;
 777                 ip->ip_id = ip_newid();
 778                 ip->ip_off = IP_DF;
 779                 ip->ip_ttl = CARP_DFLTTL;
 780                 ip->ip_p = IPPROTO_CARP;
 781                 ip->ip_sum = 0;
 782
 783                 bzero(&sa, sizeof(sa));
 784                 sa.sa_family = AF_INET;
 785                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 786                 if (ifa != NULL) {
 787                         ip->ip_src.s_addr =
 788                             ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 789                         ifa_free(ifa);
 790                 } else
 791                         ip->ip_src.s_addr = 0;
 792                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 793
 794                 ch_ptr = (struct carp_header *)(&ip[1]);
 795                 bcopy(&ch, ch_ptr, sizeof(ch));
 796                 if (carp_prepare_ad(m, sc, ch_ptr))
 797                         return;
 798
 799                 m->m_data += sizeof(*ip);
 800                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip));
 801                 m->m_data -= sizeof(*ip);
 802
 803                 CARPSTATS_INC(carps_opackets);
 804
 805                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT,
 806                     &sc->sc_carpdev->if_carp->cif_imo, NULL)) {
 807                         if (sc->sc_sendad_errors < INT_MAX)
 808                                 sc->sc_sendad_errors++;
 809                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS)
 810                                 carp_demote_adj(carp_senderr_adj, "send error");
 811                         sc->sc_sendad_success = 0;
 812                 } else {
 813                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 814                                 if (++sc->sc_sendad_success >=
 815                                     CARP_SENDAD_MIN_SUCCESS) {
 816                                         carp_demote_adj(-carp_senderr_adj,
 817                                             "send ok");
 818                                         sc->sc_sendad_errors = 0;
 819                                 }
 820                         } else
 821                                 sc->sc_sendad_errors = 0;
 822                 }
 823         }
 824 #endif /* INET */
 825 #ifdef INET6
 826         if (sc->sc_naddrs6) {
 827                 struct ip6_hdr *ip6;
 828
 829                 MGETHDR(m, M_NOWAIT, MT_HEADER);
 830                 if (m == NULL) {
 831                         CARPSTATS_INC(carps_onomem);
 832                         /* XXX maybe less ? */
 833                         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
 834                             carp_send_ad, sc);
 835                         return;
 836                 }
 837                 len = sizeof(*ip6) + sizeof(ch);
 838                 m->m_pkthdr.len = len;
 839                 m->m_pkthdr.rcvif = NULL;
 840                 m->m_len = len;
 841                 MH_ALIGN(m, m->m_len);
 842                 m->m_flags |= M_MCAST;
 843                 ip6 = mtod(m, struct ip6_hdr *);
 844                 bzero(ip6, sizeof(*ip6));
 845                 ip6->ip6_vfc |= IPV6_VERSION;
 846                 ip6->ip6_hlim = CARP_DFLTTL;
 847                 ip6->ip6_nxt = IPPROTO_CARP;
 848                 bzero(&sa, sizeof(sa));
 849
 850                 /* set the source address */
 851                 sa.sa_family = AF_INET6;
 852                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 853                 if (ifa != NULL) {
 854                         bcopy(IFA_IN6(ifa), &ip6->ip6_src,
 855                             sizeof(struct in6_addr));
 856                         ifa_free(ifa);
 857                 } else
 858                         /* This should never happen with IPv6. */
 859                         bzero(&ip6->ip6_src, sizeof(struct in6_addr));
 860
 861                 /* Set the multicast destination. */
 862                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
 863                 ip6->ip6_dst.s6_addr8[15] = 0x12;
 864                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
 865                         m_freem(m);
 866                         CARP_DEBUG("%s: in6_setscope failed\n", __func__);
 867                         return;
 868                 }
 869
 870                 ch_ptr = (struct carp_header *)(&ip6[1]);
 871                 bcopy(&ch, ch_ptr, sizeof(ch));
 872                 if (carp_prepare_ad(m, sc, ch_ptr))
 873                         return;
 874
 875                 m->m_data += sizeof(*ip6);
 876                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6));
 877                 m->m_data -= sizeof(*ip6);
 878
 879                 CARPSTATS_INC(carps_opackets6);
 880
 881                 if (ip6_output(m, NULL, NULL, 0,
 882                     &sc->sc_carpdev->if_carp->cif_im6o, NULL, NULL)) {
 883                         if (sc->sc_sendad_errors < INT_MAX)
 884                                 sc->sc_sendad_errors++;
 885                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS)
 886                                 carp_demote_adj(carp_senderr_adj,
 887                                     "send6 error");
 888                         sc->sc_sendad_success = 0;
 889                 } else {
 890                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 891                                 if (++sc->sc_sendad_success >=
 892                                     CARP_SENDAD_MIN_SUCCESS) {
 893                                         carp_demote_adj(-carp_senderr_adj,
 894                                             "send6 ok");
 895                                         sc->sc_sendad_errors = 0;
 896                                 }
 897                         } else
 898                                 sc->sc_sendad_errors = 0;
 899                 }
 900         }
 901 #endif /* INET6 */
 902
 903         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), carp_send_ad, sc);
 904 }
 905
 906 static void
 907 carp_addroute(struct carp_softc *sc)
 908 {
 909         struct ifaddr *ifa;
 910
 911         CARP_FOREACH_IFA(sc, ifa)
 912                 carp_ifa_addroute(ifa);
 913 }
 914
 915 static void
 916 carp_ifa_addroute(struct ifaddr *ifa)
 917 {
 918
 919         switch (ifa->ifa_addr->sa_family) {
 920 #ifdef INET
 921         case AF_INET:
 922                 in_addprefix(ifatoia(ifa), RTF_UP);
 923                 ifa_add_loopback_route(ifa,
 924                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 925                 break;
 926 #endif
 927 #ifdef INET6
 928         case AF_INET6:
 929                 ifa_add_loopback_route(ifa,
 930                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 931                 in6_ifaddloop(ifa);
 932                 break;
 933 #endif
 934         }
 935 }
 936
 937 static void
 938 carp_delroute(struct carp_softc *sc)
 939 {
 940         struct ifaddr *ifa;
 941
 942         CARP_FOREACH_IFA(sc, ifa)
 943                 carp_ifa_delroute(ifa);
 944 }
 945
 946 static void
 947 carp_ifa_delroute(struct ifaddr *ifa)
 948 {
 949
 950         switch (ifa->ifa_addr->sa_family) {
 951 #ifdef INET
 952         case AF_INET:
 953                 ifa_del_loopback_route(ifa,
 954                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 955                 in_scrubprefix(ifatoia(ifa), LLE_STATIC);
 956                 break;
 957 #endif
 958 #ifdef INET6
 959         case AF_INET6:
 960                 ifa_del_loopback_route(ifa,
 961                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 962                 in6_ifremloop(ifa);
 963                 break;
 964 #endif
 965         }
 966 }
 967
 968 #ifdef INET
 969 /*
 970  * Broadcast a gratuitous ARP request containing
 971  * the virtual router MAC address for each IP address
 972  * associated with the virtual router.
 973  */
 974 static void
 975 carp_send_arp(struct carp_softc *sc)
 976 {
 977         struct ifaddr *ifa;
 978
 979         CARP_FOREACH_IFA(sc, ifa)
 980                 if (ifa->ifa_addr->sa_family == AF_INET)
 981                         arp_ifinit2(sc->sc_carpdev, ifa, LLADDR(&sc->sc_addr));
 982 }
 983
 984 int
 985 carp_iamatch(struct ifaddr *ifa, uint8_t **enaddr)
 986 {
 987         struct carp_softc *sc = ifa->ifa_carp;
 988
 989         if (sc->sc_state == MASTER) {
 990                 *enaddr = LLADDR(&sc->sc_addr);
 991                 return (1);
 992         }
 993
 994         return (0);
 995 }
 996 #endif
 997
 998 #ifdef INET6
 999 static void
1000 carp_send_na(struct carp_softc *sc)
1001 {
1002         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1003         struct ifaddr *ifa;
1004         struct in6_addr *in6;
1005
1006         CARP_FOREACH_IFA(sc, ifa) {
1007                 if (ifa->ifa_addr->sa_family != AF_INET6)
1008                         continue;
1009
1010                 in6 = IFA_IN6(ifa);
1011                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1012                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1013                 DELAY(1000);    /* XXX */
1014         }
1015 }
1016
1017 struct ifaddr *
1018 carp_iamatch6(struct ifnet *ifp, struct in6_addr *taddr)
1019 {
1020         struct ifaddr *ifa;
1021
1022         IF_ADDR_RLOCK(ifp);
1023         IFNET_FOREACH_IFA(ifp, ifa)
1024                 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1025                     ifa->ifa_carp->sc_state == MASTER &&
1026                     IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1027                         ifa_ref(ifa);
1028                         IF_ADDR_RUNLOCK(ifp);
1029                         return (ifa);
1030                 }
1031         IF_ADDR_RUNLOCK(ifp);
1032
1033         return (NULL);
1034 }
1035
1036 caddr_t
1037 carp_macmatch6(struct ifnet *ifp, struct mbuf *m, const struct in6_addr *taddr)
1038 {
1039         struct ifaddr *ifa;
1040
1041         IF_ADDR_RLOCK(ifp);
1042         IFNET_FOREACH_IFA(ifp, ifa)
1043                 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1044                     IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1045                         struct carp_softc *sc = ifa->ifa_carp;
1046                         struct m_tag *mtag;
1047
1048                         IF_ADDR_RUNLOCK(ifp);
1049
1050                         mtag = m_tag_get(PACKET_TAG_CARP,
1051                             sizeof(struct ifnet *), M_NOWAIT);
1052                         if (mtag == NULL)
1053                                 /* Better a bit than nothing. */
1054                                 return (LLADDR(&sc->sc_addr));
1055
1056                         bcopy(&ifp, (caddr_t)(mtag + 1),
1057                             sizeof(struct ifnet *));
1058                         m_tag_prepend(m, mtag);
1059
1060                         return (LLADDR(&sc->sc_addr));
1061                 }
1062         IF_ADDR_RUNLOCK(ifp);
1063
1064         return (NULL);
1065 }
1066 #endif /* INET6 */
1067
1068 int
1069 carp_forus(struct ifnet *ifp, u_char *dhost)
1070 {
1071         struct carp_softc *sc;
1072         uint8_t *ena = dhost;
1073
1074         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1075                 return (0);
1076
1077         CIF_LOCK(ifp->if_carp);
1078         IFNET_FOREACH_CARP(ifp, sc) {
1079                 CARP_LOCK(sc);
1080                 if (sc->sc_state == MASTER && !bcmp(dhost, LLADDR(&sc->sc_addr),
1081                     ETHER_ADDR_LEN)) {
1082                         CARP_UNLOCK(sc);
1083                         CIF_UNLOCK(ifp->if_carp);
1084                         return (1);
1085                 }
1086                 CARP_UNLOCK(sc);
1087         }
1088         CIF_UNLOCK(ifp->if_carp);
1089
1090         return (0);
1091 }
1092
1093 static void
1094 carp_master_down(void *v)
1095 {
1096         struct carp_softc *sc = v;
1097
1098         CARP_LOCK_ASSERT(sc);
1099
1100         if (sc->sc_state == BACKUP) {
1101                 CARP_LOG("VHID %u@%s: BACKUP -> MASTER (master down)\n",
1102                     sc->sc_vhid,
1103                     sc->sc_carpdev->if_xname);
1104                 carp_master_down_locked(sc);
1105         }
1106
1107         CARP_UNLOCK(sc);
1108 }
1109
1110 static void
1111 carp_master_down_locked(struct carp_softc *sc)
1112 {
1113
1114         CARP_LOCK_ASSERT(sc);
1115
1116         switch (sc->sc_state) {
1117         case BACKUP:
1118                 carp_set_state(sc, MASTER);
1119                 carp_send_ad_locked(sc);
1120 #ifdef INET
1121                 carp_send_arp(sc);
1122 #endif
1123 #ifdef INET6
1124                 carp_send_na(sc);
1125 #endif
1126                 carp_setrun(sc, 0);
1127                 carp_addroute(sc);
1128                 break;
1129         case INIT:
1130         case MASTER:
1131 #ifdef INVARIANTS
1132                 panic("carp: VHID %u@%s: master_down event in %s state\n",
1133                     sc->sc_vhid,
1134                     sc->sc_carpdev->if_xname,
1135                     sc->sc_state ? "MASTER" : "INIT");
1136 #endif
1137                 break;
1138         }
1139 }
1140
1141 /*
1142  * When in backup state, af indicates whether to reset the master down timer
1143  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1144  */
1145 static void
1146 carp_setrun(struct carp_softc *sc, sa_family_t af)
1147 {
1148         struct timeval tv;
1149
1150         CARP_LOCK_ASSERT(sc);
1151
1152         if ((sc->sc_carpdev->if_flags & IFF_UP) == 0 ||
1153             sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1154             (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0))
1155                 return;
1156
1157         switch (sc->sc_state) {
1158         case INIT:
1159                 CARP_LOG("VHID %u@%s: INIT -> BACKUP\n",
1160                     sc->sc_vhid,
1161                     sc->sc_carpdev->if_xname);
1162                 carp_set_state(sc, BACKUP);
1163                 carp_setrun(sc, 0);
1164                 break;
1165         case BACKUP:
1166                 callout_stop(&sc->sc_ad_tmo);
1167                 tv.tv_sec = 3 * sc->sc_advbase;
1168                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1169                 switch (af) {
1170 #ifdef INET
1171                 case AF_INET:
1172                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1173                             carp_master_down, sc);
1174                         break;
1175 #endif
1176 #ifdef INET6
1177                 case AF_INET6:
1178                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1179                             carp_master_down, sc);
1180                         break;
1181 #endif
1182                 default:
1183 #ifdef INET
1184                         if (sc->sc_naddrs)
1185                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1186                                     carp_master_down, sc);
1187 #endif
1188 #ifdef INET6
1189                         if (sc->sc_naddrs6)
1190                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1191                                     carp_master_down, sc);
1192 #endif
1193                         break;
1194                 }
1195                 break;
1196         case MASTER:
1197                 tv.tv_sec = sc->sc_advbase;
1198                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1199                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1200                     carp_send_ad, sc);
1201                 break;
1202         }
1203 }
1204
1205 /*
1206  * Setup multicast structures.
1207  */
1208 static int
1209 carp_multicast_setup(struct carp_softc *sc, sa_family_t sa)
1210 {
1211         struct ifnet *ifp = sc->sc_carpdev;
1212         struct carp_if *cif = ifp->if_carp;
1213         int error = 0;
1214
1215         switch (sa) {
1216 #ifdef INET
1217         case AF_INET:
1218             {
1219                 struct ip_moptions *imo = &cif->cif_imo;
1220                 struct in_addr addr;
1221
1222                 if (imo->imo_membership)
1223                         return (0);
1224
1225                 imo->imo_membership = (struct in_multi **)malloc(
1226                     (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
1227                     M_WAITOK);
1228                 imo->imo_mfilters = NULL;
1229                 imo->imo_max_memberships = IP_MIN_MEMBERSHIPS;
1230                 imo->imo_multicast_vif = -1;
1231
1232                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1233                 if ((error = in_joingroup(ifp, &addr, NULL,
1234                     &imo->imo_membership[0])) != 0) {
1235                         free(imo->imo_membership, M_CARP);
1236                         break;
1237                 }
1238                 imo->imo_num_memberships++;
1239                 imo->imo_multicast_ifp = ifp;
1240                 imo->imo_multicast_ttl = CARP_DFLTTL;
1241                 imo->imo_multicast_loop = 0;
1242                 break;
1243            }
1244 #endif
1245 #ifdef INET6
1246         case AF_INET6:
1247             {
1248                 struct ip6_moptions *im6o = &cif->cif_im6o;
1249                 struct in6_addr in6;
1250                 struct in6_multi *in6m;
1251
1252                 if (im6o->im6o_membership)
1253                         return (0);
1254
1255                 im6o->im6o_membership = (struct in6_multi **)malloc(
1256                     (sizeof(struct in6_multi *) * IPV6_MIN_MEMBERSHIPS), M_CARP,
1257                     M_ZERO|M_WAITOK);
1258                 im6o->im6o_mfilters = NULL;
1259                 im6o->im6o_max_memberships = IPV6_MIN_MEMBERSHIPS;
1260                 im6o->im6o_multicast_hlim = CARP_DFLTTL;
1261                 im6o->im6o_multicast_ifp = ifp;
1262
1263                 /* Join IPv6 CARP multicast group. */
1264                 bzero(&in6, sizeof(in6));
1265                 in6.s6_addr16[0] = htons(0xff02);
1266                 in6.s6_addr8[15] = 0x12;
1267                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1268                         free(im6o->im6o_membership, M_CARP);
1269                         break;
1270                 }
1271                 in6m = NULL;
1272                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1273                         free(im6o->im6o_membership, M_CARP);
1274                         break;
1275                 }
1276                 im6o->im6o_membership[0] = in6m;
1277                 im6o->im6o_num_memberships++;
1278
1279                 /* Join solicited multicast address. */
1280                 bzero(&in6, sizeof(in6));
1281                 in6.s6_addr16[0] = htons(0xff02);
1282                 in6.s6_addr32[1] = 0;
1283                 in6.s6_addr32[2] = htonl(1);
1284                 in6.s6_addr32[3] = 0;
1285                 in6.s6_addr8[12] = 0xff;
1286                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1287                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1288                         free(im6o->im6o_membership, M_CARP);
1289                         break;
1290                 }
1291                 in6m = NULL;
1292                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1293                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1294                         free(im6o->im6o_membership, M_CARP);
1295                         break;
1296                 }
1297                 im6o->im6o_membership[1] = in6m;
1298                 im6o->im6o_num_memberships++;
1299                 break;
1300             }
1301 #endif
1302         }
1303
1304         return (error);
1305 }
1306
1307 /*
1308  * Free multicast structures.
1309  */
1310 static void
1311 carp_multicast_cleanup(struct carp_softc *sc, sa_family_t sa)
1312 {
1313         struct ifnet *ifp = sc->sc_carpdev;
1314         struct carp_if *cif = ifp->if_carp;
1315
1316         switch (sa) {
1317 #ifdef INET
1318         case AF_INET:
1319                 if (sc->sc_naddrs == 0) {
1320                         struct ip_moptions *imo = &cif->cif_imo;
1321
1322                         in_leavegroup(imo->imo_membership[0], NULL);
1323                         KASSERT(imo->imo_mfilters == NULL,
1324                             ("%s: imo_mfilters != NULL", __func__));
1325                         free(imo->imo_membership, M_CARP);
1326                         imo->imo_membership = NULL;
1327
1328                 }
1329                 break;
1330 #endif
1331 #ifdef INET6
1332         case AF_INET6:
1333                 if (sc->sc_naddrs6 == 0) {
1334                         struct ip6_moptions *im6o = &cif->cif_im6o;
1335
1336                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1337                         in6_mc_leave(im6o->im6o_membership[1], NULL);
1338                         KASSERT(im6o->im6o_mfilters == NULL,
1339                             ("%s: im6o_mfilters != NULL", __func__));
1340                         free(im6o->im6o_membership, M_CARP);
1341                         im6o->im6o_membership = NULL;
1342                 }
1343                 break;
1344 #endif
1345         }
1346 }
1347
1348 int
1349 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *sa)
1350 {
1351         struct m_tag *mtag;
1352         struct carp_softc *sc;
1353
1354         if (!sa)
1355                 return (0);
1356
1357         switch (sa->sa_family) {
1358 #ifdef INET
1359         case AF_INET:
1360                 break;
1361 #endif
1362 #ifdef INET6
1363         case AF_INET6:
1364                 break;
1365 #endif
1366         default:
1367                 return (0);
1368         }
1369
1370         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
1371         if (mtag == NULL)
1372                 return (0);
1373
1374         bcopy(mtag + 1, &sc, sizeof(struct carp_softc *));
1375
1376         /* Set the source MAC address to the Virtual Router MAC Address. */
1377         switch (ifp->if_type) {
1378         case IFT_ETHER:
1379         case IFT_L2VLAN: {
1380                         struct ether_header *eh;
1381
1382                         eh = mtod(m, struct ether_header *);
1383                         eh->ether_shost[0] = 0;
1384                         eh->ether_shost[1] = 0;
1385                         eh->ether_shost[2] = 0x5e;
1386                         eh->ether_shost[3] = 0;
1387                         eh->ether_shost[4] = 1;
1388                         eh->ether_shost[5] = sc->sc_vhid;
1389                 }
1390                 break;
1391         case IFT_FDDI: {
1392                         struct fddi_header *fh;
1393
1394                         fh = mtod(m, struct fddi_header *);
1395                         fh->fddi_shost[0] = 0;
1396                         fh->fddi_shost[1] = 0;
1397                         fh->fddi_shost[2] = 0x5e;
1398                         fh->fddi_shost[3] = 0;
1399                         fh->fddi_shost[4] = 1;
1400                         fh->fddi_shost[5] = sc->sc_vhid;
1401                 }
1402                 break;
1403         case IFT_ISO88025: {
1404                         struct iso88025_header *th;
1405                         th = mtod(m, struct iso88025_header *);
1406                         th->iso88025_shost[0] = 3;
1407                         th->iso88025_shost[1] = 0;
1408                         th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
1409                         th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
1410                         th->iso88025_shost[4] = 0;
1411                         th->iso88025_shost[5] = 0;
1412                 }
1413                 break;
1414         default:
1415                 printf("%s: carp is not supported for the %d interface type\n",
1416                     ifp->if_xname, ifp->if_type);
1417                 return (EOPNOTSUPP);
1418         }
1419
1420         return (0);
1421 }
1422
1423 static struct carp_softc*
1424 carp_alloc(struct ifnet *ifp)
1425 {
1426         struct carp_softc *sc;
1427         struct carp_if *cif;
1428
1429         if ((cif = ifp->if_carp) == NULL) {
1430                 cif = carp_alloc_if(ifp);
1431                 if (cif == NULL)
1432                         return (NULL);
1433         }
1434
1435         sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
1436
1437         sc->sc_advbase = CARP_DFLTINTV;
1438         sc->sc_vhid = -1;       /* required setting */
1439         sc->sc_init_counter = 1;
1440         sc->sc_state = INIT;
1441
1442         sc->sc_ifasiz = sizeof(struct ifaddr *);
1443         sc->sc_ifas = malloc(sc->sc_ifasiz, M_CARP, M_WAITOK|M_ZERO);
1444         sc->sc_carpdev = ifp;
1445
1446         CARP_LOCK_INIT(sc);
1447 #ifdef INET
1448         callout_init_mtx(&sc->sc_md_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1449 #endif
1450 #ifdef INET6
1451         callout_init_mtx(&sc->sc_md6_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1452 #endif
1453         callout_init_mtx(&sc->sc_ad_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1454
1455         CIF_LOCK(cif);
1456         TAILQ_INSERT_TAIL(&cif->cif_vrs, sc, sc_list);
1457         CIF_UNLOCK(cif);
1458
1459         mtx_lock(&carp_mtx);
1460         LIST_INSERT_HEAD(&carp_list, sc, sc_next);
1461         mtx_unlock(&carp_mtx);
1462
1463         return (sc);
1464 }
1465
1466 static int
1467 carp_grow_ifas(struct carp_softc *sc)
1468 {
1469         struct ifaddr **new;
1470
1471         CARP_LOCK_ASSERT(sc);
1472
1473         new = malloc(sc->sc_ifasiz * 2, M_CARP, M_NOWAIT|M_ZERO);
1474         if (new == NULL)
1475                 return (ENOMEM);
1476         bcopy(sc->sc_ifas, new, sc->sc_ifasiz);
1477         free(sc->sc_ifas, M_CARP);
1478         sc->sc_ifas = new;
1479         sc->sc_ifasiz *= 2;
1480
1481         return (0);
1482 }
1483
1484 static void
1485 carp_destroy(struct carp_softc *sc)
1486 {
1487         struct ifnet *ifp = sc->sc_carpdev;
1488         struct carp_if *cif = ifp->if_carp;
1489
1490         CIF_LOCK(cif);
1491         TAILQ_REMOVE(&cif->cif_vrs, sc, sc_list);
1492         if (TAILQ_EMPTY(&cif->cif_vrs))
1493                 carp_free_if(cif);
1494         else
1495                 CIF_UNLOCK(cif);
1496
1497         mtx_lock(&carp_mtx);
1498         LIST_REMOVE(sc, sc_next);
1499         mtx_unlock(&carp_mtx);
1500
1501         CARP_LOCK(sc);
1502         if (sc->sc_suppress)
1503                 carp_demote_adj(-carp_ifdown_adj, "vhid removed");
1504         callout_drain(&sc->sc_ad_tmo);
1505 #ifdef INET
1506         callout_drain(&sc->sc_md_tmo);
1507 #endif
1508 #ifdef INET6
1509         callout_drain(&sc->sc_md6_tmo);
1510 #endif
1511         CARP_LOCK_DESTROY(sc);
1512
1513         free(sc->sc_ifas, M_CARP);
1514         free(sc, M_CARP);
1515 }
1516
1517 static struct carp_if*
1518 carp_alloc_if(struct ifnet *ifp)
1519 {
1520         struct carp_if *cif;
1521
1522         cif = malloc(sizeof(*cif), M_CARP, M_WAITOK|M_ZERO);
1523
1524         if (ifpromisc(ifp, 1) != 0)
1525                 goto cleanup;
1526
1527         CIF_LOCK_INIT(cif);
1528         cif->cif_ifp = ifp;
1529         TAILQ_INIT(&cif->cif_vrs);
1530
1531         IF_ADDR_WLOCK(ifp);
1532         ifp->if_carp = cif;
1533         if_ref(ifp);
1534         IF_ADDR_WUNLOCK(ifp);
1535
1536         return (cif);
1537
1538 cleanup:
1539         free(cif, M_CARP);
1540
1541         return (NULL);
1542 }
1543
1544 static void
1545 carp_free_if(struct carp_if *cif)
1546 {
1547         struct ifnet *ifp = cif->cif_ifp;
1548
1549         CIF_LOCK_ASSERT(cif);
1550         KASSERT(TAILQ_EMPTY(&cif->cif_vrs), ("%s: softc list not empty",
1551             __func__));
1552
1553         IF_ADDR_WLOCK(ifp);
1554         ifp->if_carp = NULL;
1555         if_rele(ifp);
1556         IF_ADDR_WUNLOCK(ifp);
1557
1558         CIF_LOCK_DESTROY(cif);
1559
1560         ifpromisc(ifp, 0);
1561
1562         free(cif, M_CARP);
1563 }
1564
1565 static void
1566 carp_carprcp(struct carpreq *carpr, struct carp_softc *sc, int priv)
1567 {
1568
1569         CARP_LOCK(sc);
1570         carpr->carpr_state = sc->sc_state;
1571         carpr->carpr_vhid = sc->sc_vhid;
1572         carpr->carpr_advbase = sc->sc_advbase;
1573         carpr->carpr_advskew = sc->sc_advskew;
1574         if (priv)
1575                 bcopy(sc->sc_key, carpr->carpr_key, sizeof(carpr->carpr_key));
1576         else
1577                 bzero(carpr->carpr_key, sizeof(carpr->carpr_key));
1578         CARP_UNLOCK(sc);
1579 }
1580
1581 int
1582 carp_ioctl(struct ifreq *ifr, u_long cmd, struct thread *td)
1583 {
1584         struct carpreq carpr;
1585         struct ifnet *ifp;
1586         struct carp_softc *sc = NULL;
1587         int error = 0, locked = 0;
1588
1589         if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1590                 return (error);
1591
1592         ifp = ifunit_ref(ifr->ifr_name);
1593         if (ifp == NULL)
1594                 return (ENXIO);
1595
1596         switch (ifp->if_type) {
1597         case IFT_ETHER:
1598         case IFT_L2VLAN:
1599         case IFT_FDDI:
1600         case IFT_ISO88025:
1601                 break;
1602         default:
1603                 error = EOPNOTSUPP;
1604                 goto out;
1605         }
1606
1607         if ((ifp->if_flags & IFF_MULTICAST) == 0) {
1608                 error = EADDRNOTAVAIL;
1609                 goto out;
1610         }
1611
1612         switch (cmd) {
1613         case SIOCSVH:
1614                 if ((error = priv_check(td, PRIV_NETINET_CARP)))
1615                         break;
1616                 if (carpr.carpr_vhid <= 0 || carpr.carpr_vhid > CARP_MAXVHID ||
1617                     carpr.carpr_advbase < 0 || carpr.carpr_advskew < 0) {
1618                         error = EINVAL;
1619                         break;
1620                 }
1621
1622                 if (ifp->if_carp) {
1623                         CIF_LOCK(ifp->if_carp);
1624                         IFNET_FOREACH_CARP(ifp, sc)
1625                                 if (sc->sc_vhid == carpr.carpr_vhid)
1626                                         break;
1627                         CIF_UNLOCK(ifp->if_carp);
1628                 }
1629                 if (sc == NULL) {
1630                         sc = carp_alloc(ifp);
1631                         if (sc == NULL) {
1632                                 error = EINVAL; /* XXX: ifpromisc failed */
1633                                 break;
1634                         }
1635
1636                         CARP_LOCK(sc);
1637                         sc->sc_vhid = carpr.carpr_vhid;
1638                         LLADDR(&sc->sc_addr)[0] = 0;
1639                         LLADDR(&sc->sc_addr)[1] = 0;
1640                         LLADDR(&sc->sc_addr)[2] = 0x5e;
1641                         LLADDR(&sc->sc_addr)[3] = 0;
1642                         LLADDR(&sc->sc_addr)[4] = 1;
1643                         LLADDR(&sc->sc_addr)[5] = sc->sc_vhid;
1644                 } else
1645                         CARP_LOCK(sc);
1646                 locked = 1;
1647                 if (carpr.carpr_advbase > 0) {
1648                         if (carpr.carpr_advbase > 255 ||
1649                             carpr.carpr_advbase < CARP_DFLTINTV) {
1650                                 error = EINVAL;
1651                                 break;
1652                         }
1653                         sc->sc_advbase = carpr.carpr_advbase;
1654                 }
1655                 if (carpr.carpr_advskew > 0) {
1656                         if (carpr.carpr_advskew >= 255) {
1657                                 error = EINVAL;
1658                                 break;
1659                         }
1660                         sc->sc_advskew = carpr.carpr_advskew;
1661                 }
1662                 if (carpr.carpr_key[0] != '\0') {
1663                         bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1664                         carp_hmac_prepare(sc);
1665                 }
1666                 if (sc->sc_state != INIT &&
1667                     carpr.carpr_state != sc->sc_state) {
1668                         switch (carpr.carpr_state) {
1669                         case BACKUP:
1670                                 callout_stop(&sc->sc_ad_tmo);
1671                                 carp_set_state(sc, BACKUP);
1672                                 carp_setrun(sc, 0);
1673                                 carp_delroute(sc);
1674                                 break;
1675                         case MASTER:
1676                                 carp_master_down_locked(sc);
1677                                 break;
1678                         default:
1679                                 break;
1680                         }
1681                 }
1682                 break;
1683
1684         case SIOCGVH:
1685             {
1686                 int priveleged;
1687
1688                 if (carpr.carpr_vhid < 0 || carpr.carpr_vhid > CARP_MAXVHID) {
1689                         error = EINVAL;
1690                         break;
1691                 }
1692                 if (carpr.carpr_count < 1) {
1693                         error = EMSGSIZE;
1694                         break;
1695                 }
1696                 if (ifp->if_carp == NULL) {
1697                         error = ENOENT;
1698                         break;
1699                 }
1700
1701                 priveleged = (priv_check(td, PRIV_NETINET_CARP) == 0);
1702                 if (carpr.carpr_vhid != 0) {
1703                         CIF_LOCK(ifp->if_carp);
1704                         IFNET_FOREACH_CARP(ifp, sc)
1705                                 if (sc->sc_vhid == carpr.carpr_vhid)
1706                                         break;
1707                         CIF_UNLOCK(ifp->if_carp);
1708                         if (sc == NULL) {
1709                                 error = ENOENT;
1710                                 break;
1711                         }
1712                         carp_carprcp(&carpr, sc, priveleged);
1713                         error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1714                 } else  {
1715                         int i, count;
1716
1717                         count = 0;
1718                         CIF_LOCK(ifp->if_carp);
1719                         IFNET_FOREACH_CARP(ifp, sc)
1720                                 count++;
1721
1722                         if (count > carpr.carpr_count) {
1723                                 CIF_UNLOCK(ifp->if_carp);
1724                                 error = EMSGSIZE;
1725                                 break;
1726                         }
1727
1728                         i = 0;
1729                         IFNET_FOREACH_CARP(ifp, sc) {
1730                                 carp_carprcp(&carpr, sc, priveleged);
1731                                 carpr.carpr_count = count;
1732                                 error = copyout(&carpr, ifr->ifr_data +
1733                                     (i * sizeof(carpr)), sizeof(carpr));
1734                                 if (error) {
1735                                         CIF_UNLOCK(ifp->if_carp);
1736                                         break;
1737                                 }
1738                                 i++;
1739                         }
1740                         CIF_UNLOCK(ifp->if_carp);
1741                 }
1742                 break;
1743             }
1744         default:
1745                 error = EINVAL;
1746         }
1747
1748 out:
1749         if (locked)
1750                 CARP_UNLOCK(sc);
1751         if_rele(ifp);
1752
1753         return (error);
1754 }
1755
1756 static int
1757 carp_get_vhid(struct ifaddr *ifa)
1758 {
1759
1760         if (ifa == NULL || ifa->ifa_carp == NULL)
1761                 return (0);
1762
1763         return (ifa->ifa_carp->sc_vhid);
1764 }
1765
1766 int
1767 carp_attach(struct ifaddr *ifa, int vhid)
1768 {
1769         struct ifnet *ifp = ifa->ifa_ifp;
1770         struct carp_softc *sc;
1771         int index, error;
1772
1773         if (ifp->if_carp == NULL)
1774                 return (ENOPROTOOPT);
1775
1776         switch (ifa->ifa_addr->sa_family) {
1777 #ifdef INET
1778         case AF_INET:
1779 #endif
1780 #ifdef INET6
1781         case AF_INET6:
1782 #endif
1783                 break;
1784         default:
1785                 return (EPROTOTYPE);
1786         }
1787
1788         CIF_LOCK(ifp->if_carp);
1789         IFNET_FOREACH_CARP(ifp, sc)
1790                 if (sc->sc_vhid == vhid)
1791                         break;
1792         CIF_UNLOCK(ifp->if_carp);
1793         if (sc == NULL)
1794                 return (ENOENT);
1795
1796         if (ifa->ifa_carp) {
1797                 if (ifa->ifa_carp->sc_vhid != vhid)
1798                         carp_detach(ifa);
1799                 else
1800                         return (0);
1801         }
1802
1803         error = carp_multicast_setup(sc, ifa->ifa_addr->sa_family);
1804         if (error)
1805                 return (error);
1806
1807         CARP_LOCK(sc);
1808         index = sc->sc_naddrs + sc->sc_naddrs6 + 1;
1809         if (index > sc->sc_ifasiz / sizeof(struct ifaddr *))
1810                 if ((error = carp_grow_ifas(sc)) != 0) {
1811                         carp_multicast_cleanup(sc,
1812                             ifa->ifa_addr->sa_family);
1813                         CARP_UNLOCK(sc);
1814                         return (error);
1815                 }
1816
1817         switch (ifa->ifa_addr->sa_family) {
1818 #ifdef INET
1819         case AF_INET:
1820                 sc->sc_naddrs++;
1821                 break;
1822 #endif
1823 #ifdef INET6
1824         case AF_INET6:
1825                 sc->sc_naddrs6++;
1826                 break;
1827 #endif
1828         }
1829
1830         ifa_ref(ifa);
1831         sc->sc_ifas[index - 1] = ifa;
1832         ifa->ifa_carp = sc;
1833
1834         carp_hmac_prepare(sc);
1835         carp_sc_state(sc);
1836
1837         CARP_UNLOCK(sc);
1838
1839         return (0);
1840 }
1841
1842 void
1843 carp_detach(struct ifaddr *ifa)
1844 {
1845         struct carp_softc *sc = ifa->ifa_carp;
1846         int i, index;
1847
1848         KASSERT(sc != NULL, ("%s: %p not attached", __func__, ifa));
1849
1850         CARP_LOCK(sc);
1851
1852         /* Shift array. */
1853         index = sc->sc_naddrs + sc->sc_naddrs6;
1854         for (i = 0; i < index; i++)
1855                 if (sc->sc_ifas[i] == ifa)
1856                         break;
1857         KASSERT(i < index, ("%s: %p no backref", __func__, ifa));
1858         for (; i < index - 1; i++)
1859                 sc->sc_ifas[i] = sc->sc_ifas[i+1];
1860         sc->sc_ifas[index - 1] = NULL;
1861
1862         switch (ifa->ifa_addr->sa_family) {
1863 #ifdef INET
1864         case AF_INET:
1865                 sc->sc_naddrs--;
1866                 break;
1867 #endif
1868 #ifdef INET6
1869         case AF_INET6:
1870                 sc->sc_naddrs6--;
1871                 break;
1872 #endif
1873         }
1874
1875         carp_ifa_delroute(ifa);
1876         carp_multicast_cleanup(sc, ifa->ifa_addr->sa_family);
1877
1878         ifa->ifa_carp = NULL;
1879         ifa_free(ifa);
1880
1881         carp_hmac_prepare(sc);
1882         carp_sc_state(sc);
1883
1884         if (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) {
1885                 CARP_UNLOCK(sc);
1886                 carp_destroy(sc);
1887         } else
1888                 CARP_UNLOCK(sc);
1889 }
1890
1891 static void
1892 carp_set_state(struct carp_softc *sc, int state)
1893 {
1894
1895         CARP_LOCK_ASSERT(sc);
1896
1897         if (sc->sc_state != state) {
1898                 const char *carp_states[] = { CARP_STATES };
1899                 char subsys[IFNAMSIZ+5];
1900
1901                 sc->sc_state = state;
1902
1903                 snprintf(subsys, IFNAMSIZ+5, "%u@%s", sc->sc_vhid,
1904                     sc->sc_carpdev->if_xname);
1905                 devctl_notify("CARP", subsys, carp_states[state], NULL);
1906         }
1907 }
1908
1909 static void
1910 carp_linkstate(struct ifnet *ifp)
1911 {
1912         struct carp_softc *sc;
1913
1914         CIF_LOCK(ifp->if_carp);
1915         IFNET_FOREACH_CARP(ifp, sc) {
1916                 CARP_LOCK(sc);
1917                 carp_sc_state(sc);
1918                 CARP_UNLOCK(sc);
1919         }
1920         CIF_UNLOCK(ifp->if_carp);
1921 }
1922
1923 static void
1924 carp_sc_state(struct carp_softc *sc)
1925 {
1926
1927         CARP_LOCK_ASSERT(sc);
1928
1929         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1930             !(sc->sc_carpdev->if_flags & IFF_UP)) {
1931                 callout_stop(&sc->sc_ad_tmo);
1932 #ifdef INET
1933                 callout_stop(&sc->sc_md_tmo);
1934 #endif
1935 #ifdef INET6
1936                 callout_stop(&sc->sc_md6_tmo);
1937 #endif
1938                 carp_set_state(sc, INIT);
1939                 carp_setrun(sc, 0);
1940                 if (!sc->sc_suppress)
1941                         carp_demote_adj(carp_ifdown_adj, "interface down");
1942                 sc->sc_suppress = 1;
1943         } else {
1944                 carp_set_state(sc, INIT);
1945                 carp_setrun(sc, 0);
1946                 if (sc->sc_suppress)
1947                         carp_demote_adj(-carp_ifdown_adj, "interface up");
1948                 sc->sc_suppress = 0;
1949         }
1950 }
1951
1952 static void
1953 carp_demote_adj(int adj, char *reason)
1954 {
1955         carp_demotion += adj;
1956         CARP_LOG("demoted by %d to %d (%s)\n", adj, carp_demotion, reason);
1957         taskqueue_enqueue(taskqueue_swi, &carp_sendall_task);
1958 }
1959
1960 #ifdef INET
1961 extern  struct domain inetdomain;
1962 static struct protosw in_carp_protosw = {
1963         .pr_type =              SOCK_RAW,
1964         .pr_domain =            &inetdomain,
1965         .pr_protocol =          IPPROTO_CARP,
1966         .pr_flags =             PR_ATOMIC|PR_ADDR,
1967         .pr_input =             carp_input,
1968         .pr_output =            (pr_output_t *)rip_output,
1969         .pr_ctloutput =         rip_ctloutput,
1970         .pr_usrreqs =           &rip_usrreqs
1971 };
1972 #endif
1973
1974 #ifdef INET6
1975 extern  struct domain inet6domain;
1976 static struct ip6protosw in6_carp_protosw = {
1977         .pr_type =              SOCK_RAW,
1978         .pr_domain =            &inet6domain,
1979         .pr_protocol =          IPPROTO_CARP,
1980         .pr_flags =             PR_ATOMIC|PR_ADDR,
1981         .pr_input =             carp6_input,
1982         .pr_output =            rip6_output,
1983         .pr_ctloutput =         rip6_ctloutput,
1984         .pr_usrreqs =           &rip6_usrreqs
1985 };
1986 #endif
1987
1988 static void
1989 carp_mod_cleanup(void)
1990 {
1991
1992 #ifdef INET
1993         if (proto_reg[CARP_INET] == 0) {
1994                 (void)ipproto_unregister(IPPROTO_CARP);
1995                 pf_proto_unregister(PF_INET, IPPROTO_CARP, SOCK_RAW);
1996                 proto_reg[CARP_INET] = -1;
1997         }
1998         carp_iamatch_p = NULL;
1999 #endif
2000 #ifdef INET6
2001         if (proto_reg[CARP_INET6] == 0) {
2002                 (void)ip6proto_unregister(IPPROTO_CARP);
2003                 pf_proto_unregister(PF_INET6, IPPROTO_CARP, SOCK_RAW);
2004                 proto_reg[CARP_INET6] = -1;
2005         }
2006         carp_iamatch6_p = NULL;
2007         carp_macmatch6_p = NULL;
2008 #endif
2009         carp_ioctl_p = NULL;
2010         carp_attach_p = NULL;
2011         carp_detach_p = NULL;
2012         carp_get_vhid_p = NULL;
2013         carp_linkstate_p = NULL;
2014         carp_forus_p = NULL;
2015         carp_output_p = NULL;
2016         carp_demote_adj_p = NULL;
2017         mtx_unlock(&carp_mtx);
2018         taskqueue_drain(taskqueue_swi, &carp_sendall_task);
2019         mtx_destroy(&carp_mtx);
2020 }
2021
2022 static int
2023 carp_mod_load(void)
2024 {
2025         int err;
2026
2027         mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2028         LIST_INIT(&carp_list);
2029         carp_get_vhid_p = carp_get_vhid;
2030         carp_forus_p = carp_forus;
2031         carp_output_p = carp_output;
2032         carp_linkstate_p = carp_linkstate;
2033         carp_ioctl_p = carp_ioctl;
2034         carp_attach_p = carp_attach;
2035         carp_detach_p = carp_detach;
2036         carp_demote_adj_p = carp_demote_adj;
2037 #ifdef INET6
2038         carp_iamatch6_p = carp_iamatch6;
2039         carp_macmatch6_p = carp_macmatch6;
2040         proto_reg[CARP_INET6] = pf_proto_register(PF_INET6,
2041             (struct protosw *)&in6_carp_protosw);
2042         if (proto_reg[CARP_INET6]) {
2043                 printf("carp: error %d attaching to PF_INET6\n",
2044                     proto_reg[CARP_INET6]);
2045                 carp_mod_cleanup();
2046                 return (proto_reg[CARP_INET6]);
2047         }
2048         err = ip6proto_register(IPPROTO_CARP);
2049         if (err) {
2050                 printf("carp: error %d registering with INET6\n", err);
2051                 carp_mod_cleanup();
2052                 return (err);
2053         }
2054 #endif
2055 #ifdef INET
2056         carp_iamatch_p = carp_iamatch;
2057         proto_reg[CARP_INET] = pf_proto_register(PF_INET, &in_carp_protosw);
2058         if (proto_reg[CARP_INET]) {
2059                 printf("carp: error %d attaching to PF_INET\n",
2060                     proto_reg[CARP_INET]);
2061                 carp_mod_cleanup();
2062                 return (proto_reg[CARP_INET]);
2063         }
2064         err = ipproto_register(IPPROTO_CARP);
2065         if (err) {
2066                 printf("carp: error %d registering with INET\n", err);
2067                 carp_mod_cleanup();
2068                 return (err);
2069         }
2070 #endif
2071         return (0);
2072 }
2073
2074 static int
2075 carp_modevent(module_t mod, int type, void *data)
2076 {
2077         switch (type) {
2078         case MOD_LOAD:
2079                 return carp_mod_load();
2080                 /* NOTREACHED */
2081         case MOD_UNLOAD:
2082                 mtx_lock(&carp_mtx);
2083                 if (LIST_EMPTY(&carp_list))
2084                         carp_mod_cleanup();
2085                 else {
2086                         mtx_unlock(&carp_mtx);
2087                         return (EBUSY);
2088                 }
2089                 break;
2090
2091         default:
2092                 return (EINVAL);
2093         }
2094
2095         return (0);
2096 }
2097
2098 static moduledata_t carp_mod = {
2099         "carp",
2100         carp_modevent,
2101         0
2102 };
2103
2104 DECLARE_MODULE(carp, carp_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);