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 static int carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS);
 196
 197 SYSCTL_NODE(_net_inet, IPPROTO_CARP,    carp,   CTLFLAG_RW, 0,  "CARP");
 198 SYSCTL_INT(_net_inet_carp, OID_AUTO, allow, CTLFLAG_RW, &carp_allow, 0,
 199     "Accept incoming CARP packets");
 200 SYSCTL_INT(_net_inet_carp, OID_AUTO, preempt, CTLFLAG_RW, &carp_preempt, 0,
 201     "High-priority backup preemption mode");
 202 SYSCTL_INT(_net_inet_carp, OID_AUTO, log, CTLFLAG_RW, &carp_log, 0,
 203     "CARP log level");
 204 SYSCTL_PROC(_net_inet_carp, OID_AUTO, demotion, CTLTYPE_INT|CTLFLAG_RW,
 205     0, 0, carp_demote_adj_sysctl, "I",
 206     "Adjust demotion factor (skew of advskew)");
 207 SYSCTL_INT(_net_inet_carp, OID_AUTO, senderr_demotion_factor, CTLFLAG_RW,
 208     &carp_senderr_adj, 0, "Send error demotion factor adjustment");
 209 SYSCTL_INT(_net_inet_carp, OID_AUTO, ifdown_demotion_factor, CTLFLAG_RW,
 210     &carp_ifdown_adj, 0, "Interface down demotion factor adjustment");
 211
 212 static struct carpstats carpstats;
 213 SYSCTL_STRUCT(_net_inet_carp, OID_AUTO, stats, CTLFLAG_RW, &carpstats,
 214     carpstats, "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 215
 216 #define CARP_LOCK_INIT(sc)      mtx_init(&(sc)->sc_mtx, "carp_softc",   \
 217         NULL, MTX_DEF)
 218 #define CARP_LOCK_DESTROY(sc)   mtx_destroy(&(sc)->sc_mtx)
 219 #define CARP_LOCK_ASSERT(sc)    mtx_assert(&(sc)->sc_mtx, MA_OWNED)
 220 #define CARP_LOCK(sc)           mtx_lock(&(sc)->sc_mtx)
 221 #define CARP_UNLOCK(sc)         mtx_unlock(&(sc)->sc_mtx)
 222 #define CIF_LOCK_INIT(cif)      mtx_init(&(cif)->cif_mtx, "carp_if",   \
 223         NULL, MTX_DEF)
 224 #define CIF_LOCK_DESTROY(cif)   mtx_destroy(&(cif)->cif_mtx)
 225 #define CIF_LOCK_ASSERT(cif)    mtx_assert(&(cif)->cif_mtx, MA_OWNED)
 226 #define CIF_LOCK(cif)           mtx_lock(&(cif)->cif_mtx)
 227 #define CIF_UNLOCK(cif)         mtx_unlock(&(cif)->cif_mtx)
 228 #define CIF_FREE(cif)   do {                            \
 229                 CIF_LOCK_ASSERT(cif);                   \
 230                 if (TAILQ_EMPTY(&(cif)->cif_vrs))       \
 231                         carp_free_if(cif);              \
 232                 else                                    \
 233                         CIF_UNLOCK(cif);                \
 234 } while (0)
 235
 236 #define CARP_LOG(...)   do {                            \
 237         if (carp_log > 0)                               \
 238                 log(LOG_INFO, "carp: " __VA_ARGS__);    \
 239 } while (0)
 240
 241 #define CARP_DEBUG(...) do {                            \
 242         if (carp_log > 1)                               \
 243                 log(LOG_DEBUG, __VA_ARGS__);            \
 244 } while (0)
 245
 246 #define IFNET_FOREACH_IFA(ifp, ifa)                                     \
 247         IF_ADDR_LOCK_ASSERT(ifp);                                       \
 248         TAILQ_FOREACH((ifa), &(ifp)->if_addrhead, ifa_link)             \
 249                 if ((ifa)->ifa_carp != NULL)
 250
 251 #define CARP_FOREACH_IFA(sc, ifa)                                       \
 252         CARP_LOCK_ASSERT(sc);                                           \
 253         for (int _i = 0;                                                \
 254                 _i < (sc)->sc_naddrs + (sc)->sc_naddrs6 &&              \
 255                 ((ifa) = sc->sc_ifas[_i]) != NULL;                      \
 256                 ++_i)
 257
 258 #define IFNET_FOREACH_CARP(ifp, sc)                                     \
 259         CIF_LOCK_ASSERT(ifp->if_carp);                                  \
 260         TAILQ_FOREACH((sc), &(ifp)->if_carp->cif_vrs, sc_list)
 261
 262 #define DEMOTE_ADVSKEW(sc)                                      \
 263     (((sc)->sc_advskew + carp_demotion > CARP_MAXSKEW) ?        \
 264     CARP_MAXSKEW : ((sc)->sc_advskew + carp_demotion))
 265
 266 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 267 static struct carp_softc
 268                 *carp_alloc(struct ifnet *);
 269 static void     carp_detach_locked(struct ifaddr *);
 270 static void     carp_destroy(struct carp_softc *);
 271 static struct carp_if
 272                 *carp_alloc_if(struct ifnet *);
 273 static void     carp_free_if(struct carp_if *);
 274 static void     carp_set_state(struct carp_softc *, int);
 275 static void     carp_sc_state(struct carp_softc *);
 276 static void     carp_setrun(struct carp_softc *, sa_family_t);
 277 static void     carp_master_down(void *);
 278 static void     carp_master_down_locked(struct carp_softc *);
 279 static void     carp_send_ad(void *);
 280 static void     carp_send_ad_locked(struct carp_softc *);
 281 static void     carp_addroute(struct carp_softc *);
 282 static void     carp_ifa_addroute(struct ifaddr *);
 283 static void     carp_delroute(struct carp_softc *);
 284 static void     carp_ifa_delroute(struct ifaddr *);
 285 static void     carp_send_ad_all(void *, int);
 286 static void     carp_demote_adj(int, char *);
 287
 288 static LIST_HEAD(, carp_softc) carp_list;
 289 static struct mtx carp_mtx;
 290 static struct task carp_sendall_task =
 291     TASK_INITIALIZER(0, carp_send_ad_all, NULL);
 292
 293 static void
 294 carp_hmac_prepare(struct carp_softc *sc)
 295 {
 296         uint8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 297         uint8_t vhid = sc->sc_vhid & 0xff;
 298         struct ifaddr *ifa;
 299         int i, found;
 300 #ifdef INET
 301         struct in_addr last, cur, in;
 302 #endif
 303 #ifdef INET6
 304         struct in6_addr last6, cur6, in6;
 305 #endif
 306
 307         CARP_LOCK_ASSERT(sc);
 308
 309         /* Compute ipad from key. */
 310         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 311         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 312         for (i = 0; i < sizeof(sc->sc_pad); i++)
 313                 sc->sc_pad[i] ^= 0x36;
 314
 315         /* Precompute first part of inner hash. */
 316         SHA1Init(&sc->sc_sha1);
 317         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 318         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 319         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 320         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 321 #ifdef INET
 322         cur.s_addr = 0;
 323         do {
 324                 found = 0;
 325                 last = cur;
 326                 cur.s_addr = 0xffffffff;
 327                 CARP_FOREACH_IFA(sc, ifa) {
 328                         in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 329                         if (ifa->ifa_addr->sa_family == AF_INET &&
 330                             ntohl(in.s_addr) > ntohl(last.s_addr) &&
 331                             ntohl(in.s_addr) < ntohl(cur.s_addr)) {
 332                                 cur.s_addr = in.s_addr;
 333                                 found++;
 334                         }
 335                 }
 336                 if (found)
 337                         SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur));
 338         } while (found);
 339 #endif /* INET */
 340 #ifdef INET6
 341         memset(&cur6, 0, sizeof(cur6));
 342         do {
 343                 found = 0;
 344                 last6 = cur6;
 345                 memset(&cur6, 0xff, sizeof(cur6));
 346                 CARP_FOREACH_IFA(sc, ifa) {
 347                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 348                         if (IN6_IS_SCOPE_EMBED(&in6))
 349                                 in6.s6_addr16[1] = 0;
 350                         if (ifa->ifa_addr->sa_family == AF_INET6 &&
 351                             memcmp(&in6, &last6, sizeof(in6)) > 0 &&
 352                             memcmp(&in6, &cur6, sizeof(in6)) < 0) {
 353                                 cur6 = in6;
 354                                 found++;
 355                         }
 356                 }
 357                 if (found)
 358                         SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6));
 359         } while (found);
 360 #endif /* INET6 */
 361
 362         /* convert ipad to opad */
 363         for (i = 0; i < sizeof(sc->sc_pad); i++)
 364                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 365 }
 366
 367 static void
 368 carp_hmac_generate(struct carp_softc *sc, uint32_t counter[2],
 369     unsigned char md[20])
 370 {
 371         SHA1_CTX sha1ctx;
 372
 373         CARP_LOCK_ASSERT(sc);
 374
 375         /* fetch first half of inner hash */
 376         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 377
 378         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 379         SHA1Final(md, &sha1ctx);
 380
 381         /* outer hash */
 382         SHA1Init(&sha1ctx);
 383         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 384         SHA1Update(&sha1ctx, md, 20);
 385         SHA1Final(md, &sha1ctx);
 386 }
 387
 388 static int
 389 carp_hmac_verify(struct carp_softc *sc, uint32_t counter[2],
 390     unsigned char md[20])
 391 {
 392         unsigned char md2[20];
 393
 394         CARP_LOCK_ASSERT(sc);
 395
 396         carp_hmac_generate(sc, counter, md2);
 397
 398         return (bcmp(md, md2, sizeof(md2)));
 399 }
 400
 401 /*
 402  * process input packet.
 403  * we have rearranged checks order compared to the rfc,
 404  * but it seems more efficient this way or not possible otherwise.
 405  */
 406 #ifdef INET
 407 void
 408 carp_input(struct mbuf *m, int hlen)
 409 {
 410         struct ip *ip = mtod(m, struct ip *);
 411         struct carp_header *ch;
 412         int iplen, len;
 413
 414         CARPSTATS_INC(carps_ipackets);
 415
 416         if (!carp_allow) {
 417                 m_freem(m);
 418                 return;
 419         }
 420
 421         /* verify that the IP TTL is 255.  */
 422         if (ip->ip_ttl != CARP_DFLTTL) {
 423                 CARPSTATS_INC(carps_badttl);
 424                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 425                     ip->ip_ttl,
 426                     m->m_pkthdr.rcvif->if_xname);
 427                 m_freem(m);
 428                 return;
 429         }
 430
 431         iplen = ip->ip_hl << 2;
 432
 433         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 434                 CARPSTATS_INC(carps_badlen);
 435                 CARP_DEBUG("%s: received len %zd < sizeof(struct carp_header) "
 436                     "on %s\n", __func__, m->m_len - sizeof(struct ip),
 437                     m->m_pkthdr.rcvif->if_xname);
 438                 m_freem(m);
 439                 return;
 440         }
 441
 442         if (iplen + sizeof(*ch) < m->m_len) {
 443                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 444                         CARPSTATS_INC(carps_hdrops);
 445                         CARP_DEBUG("%s: pullup failed\n", __func__);
 446                         return;
 447                 }
 448                 ip = mtod(m, struct ip *);
 449         }
 450         ch = (struct carp_header *)((char *)ip + iplen);
 451
 452         /*
 453          * verify that the received packet length is
 454          * equal to the CARP header
 455          */
 456         len = iplen + sizeof(*ch);
 457         if (len > m->m_pkthdr.len) {
 458                 CARPSTATS_INC(carps_badlen);
 459                 CARP_DEBUG("%s: packet too short %d on %s\n", __func__,
 460                     m->m_pkthdr.len,
 461                     m->m_pkthdr.rcvif->if_xname);
 462                 m_freem(m);
 463                 return;
 464         }
 465
 466         if ((m = m_pullup(m, len)) == NULL) {
 467                 CARPSTATS_INC(carps_hdrops);
 468                 return;
 469         }
 470         ip = mtod(m, struct ip *);
 471         ch = (struct carp_header *)((char *)ip + iplen);
 472
 473         /* verify the CARP checksum */
 474         m->m_data += iplen;
 475         if (in_cksum(m, len - iplen)) {
 476                 CARPSTATS_INC(carps_badsum);
 477                 CARP_DEBUG("%s: checksum failed on %s\n", __func__,
 478                     m->m_pkthdr.rcvif->if_xname);
 479                 m_freem(m);
 480                 return;
 481         }
 482         m->m_data -= iplen;
 483
 484         carp_input_c(m, ch, AF_INET);
 485 }
 486 #endif
 487
 488 #ifdef INET6
 489 int
 490 carp6_input(struct mbuf **mp, int *offp, int proto)
 491 {
 492         struct mbuf *m = *mp;
 493         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 494         struct carp_header *ch;
 495         u_int len;
 496
 497         CARPSTATS_INC(carps_ipackets6);
 498
 499         if (!carp_allow) {
 500                 m_freem(m);
 501                 return (IPPROTO_DONE);
 502         }
 503
 504         /* check if received on a valid carp interface */
 505         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 506                 CARPSTATS_INC(carps_badif);
 507                 CARP_DEBUG("%s: packet received on non-carp interface: %s\n",
 508                     __func__, m->m_pkthdr.rcvif->if_xname);
 509                 m_freem(m);
 510                 return (IPPROTO_DONE);
 511         }
 512
 513         /* verify that the IP TTL is 255 */
 514         if (ip6->ip6_hlim != CARP_DFLTTL) {
 515                 CARPSTATS_INC(carps_badttl);
 516                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 517                     ip6->ip6_hlim, m->m_pkthdr.rcvif->if_xname);
 518                 m_freem(m);
 519                 return (IPPROTO_DONE);
 520         }
 521
 522         /* verify that we have a complete carp packet */
 523         len = m->m_len;
 524         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 525         if (ch == NULL) {
 526                 CARPSTATS_INC(carps_badlen);
 527                 CARP_DEBUG("%s: packet size %u too small\n", __func__, len);
 528                 return (IPPROTO_DONE);
 529         }
 530
 531
 532         /* verify the CARP checksum */
 533         m->m_data += *offp;
 534         if (in_cksum(m, sizeof(*ch))) {
 535                 CARPSTATS_INC(carps_badsum);
 536                 CARP_DEBUG("%s: checksum failed, on %s\n", __func__,
 537                     m->m_pkthdr.rcvif->if_xname);
 538                 m_freem(m);
 539                 return (IPPROTO_DONE);
 540         }
 541         m->m_data -= *offp;
 542
 543         carp_input_c(m, ch, AF_INET6);
 544         return (IPPROTO_DONE);
 545 }
 546 #endif /* INET6 */
 547
 548 static void
 549 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 550 {
 551         struct ifnet *ifp = m->m_pkthdr.rcvif;
 552         struct ifaddr *ifa;
 553         struct carp_softc *sc;
 554         uint64_t tmp_counter;
 555         struct timeval sc_tv, ch_tv;
 556
 557         /* verify that the VHID is valid on the receiving interface */
 558         IF_ADDR_RLOCK(ifp);
 559         IFNET_FOREACH_IFA(ifp, ifa)
 560                 if (ifa->ifa_addr->sa_family == af &&
 561                     ifa->ifa_carp->sc_vhid == ch->carp_vhid) {
 562                         ifa_ref(ifa);
 563                         break;
 564                 }
 565         IF_ADDR_RUNLOCK(ifp);
 566
 567         if (ifa == NULL) {
 568                 CARPSTATS_INC(carps_badvhid);
 569                 m_freem(m);
 570                 return;
 571         }
 572
 573         /* verify the CARP version. */
 574         if (ch->carp_version != CARP_VERSION) {
 575                 CARPSTATS_INC(carps_badver);
 576                 CARP_DEBUG("%s: invalid version %d\n", ifp->if_xname,
 577                     ch->carp_version);
 578                 ifa_free(ifa);
 579                 m_freem(m);
 580                 return;
 581         }
 582
 583         sc = ifa->ifa_carp;
 584         CARP_LOCK(sc);
 585         ifa_free(ifa);
 586
 587         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 588                 CARPSTATS_INC(carps_badauth);
 589                 CARP_DEBUG("%s: incorrect hash for VHID %u@%s\n", __func__,
 590                     sc->sc_vhid, ifp->if_xname);
 591                 goto out;
 592         }
 593
 594         tmp_counter = ntohl(ch->carp_counter[0]);
 595         tmp_counter = tmp_counter<<32;
 596         tmp_counter += ntohl(ch->carp_counter[1]);
 597
 598         /* XXX Replay protection goes here */
 599
 600         sc->sc_init_counter = 0;
 601         sc->sc_counter = tmp_counter;
 602
 603         sc_tv.tv_sec = sc->sc_advbase;
 604         sc_tv.tv_usec = DEMOTE_ADVSKEW(sc) * 1000000 / 256;
 605         ch_tv.tv_sec = ch->carp_advbase;
 606         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 607
 608         switch (sc->sc_state) {
 609         case INIT:
 610                 break;
 611         case MASTER:
 612                 /*
 613                  * If we receive an advertisement from a master who's going to
 614                  * be more frequent than us, go into BACKUP state.
 615                  */
 616                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 617                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 618                         callout_stop(&sc->sc_ad_tmo);
 619                         CARP_LOG("VHID %u@%s: MASTER -> BACKUP "
 620                             "(more frequent advertisement received)\n",
 621                             sc->sc_vhid,
 622                             sc->sc_carpdev->if_xname);
 623                         carp_set_state(sc, BACKUP);
 624                         carp_setrun(sc, 0);
 625                         carp_delroute(sc);
 626                 }
 627                 break;
 628         case BACKUP:
 629                 /*
 630                  * If we're pre-empting masters who advertise slower than us,
 631                  * and this one claims to be slower, treat him as down.
 632                  */
 633                 if (carp_preempt && timevalcmp(&sc_tv, &ch_tv, <)) {
 634                         CARP_LOG("VHID %u@%s: BACKUP -> MASTER "
 635                             "(preempting a slower master)\n",
 636                             sc->sc_vhid,
 637                             sc->sc_carpdev->if_xname);
 638                         carp_master_down_locked(sc);
 639                         break;
 640                 }
 641
 642                 /*
 643                  *  If the master is going to advertise at such a low frequency
 644                  *  that he's guaranteed to time out, we'd might as well just
 645                  *  treat him as timed out now.
 646                  */
 647                 sc_tv.tv_sec = sc->sc_advbase * 3;
 648                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 649                         CARP_LOG("VHID %u@%s: BACKUP -> MASTER "
 650                             "(master timed out)\n",
 651                             sc->sc_vhid,
 652                             sc->sc_carpdev->if_xname);
 653                         carp_master_down_locked(sc);
 654                         break;
 655                 }
 656
 657                 /*
 658                  * Otherwise, we reset the counter and wait for the next
 659                  * advertisement.
 660                  */
 661                 carp_setrun(sc, af);
 662                 break;
 663         }
 664
 665 out:
 666         CARP_UNLOCK(sc);
 667         m_freem(m);
 668 }
 669
 670 static int
 671 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 672 {
 673         struct m_tag *mtag;
 674
 675         if (sc->sc_init_counter) {
 676                 /* this could also be seconds since unix epoch */
 677                 sc->sc_counter = arc4random();
 678                 sc->sc_counter = sc->sc_counter << 32;
 679                 sc->sc_counter += arc4random();
 680         } else
 681                 sc->sc_counter++;
 682
 683         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 684         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 685
 686         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 687
 688         /* Tag packet for carp_output */
 689         if ((mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct carp_softc *),
 690             M_NOWAIT)) == NULL) {
 691                 m_freem(m);
 692                 CARPSTATS_INC(carps_onomem);
 693                 return (ENOMEM);
 694         }
 695         bcopy(&sc, mtag + 1, sizeof(sc));
 696         m_tag_prepend(m, mtag);
 697
 698         return (0);
 699 }
 700
 701 /*
 702  * To avoid LORs and possible recursions this function shouldn't
 703  * be called directly, but scheduled via taskqueue.
 704  */
 705 static void
 706 carp_send_ad_all(void *ctx __unused, int pending __unused)
 707 {
 708         struct carp_softc *sc;
 709
 710         mtx_lock(&carp_mtx);
 711         LIST_FOREACH(sc, &carp_list, sc_next)
 712                 if (sc->sc_state == MASTER) {
 713                         CARP_LOCK(sc);
 714                         CURVNET_SET(sc->sc_carpdev->if_vnet);
 715                         carp_send_ad_locked(sc);
 716                         CURVNET_RESTORE();
 717                         CARP_UNLOCK(sc);
 718                 }
 719         mtx_unlock(&carp_mtx);
 720 }
 721
 722 /* Send a periodic advertisement, executed in callout context. */
 723 static void
 724 carp_send_ad(void *v)
 725 {
 726         struct carp_softc *sc = v;
 727
 728         CARP_LOCK_ASSERT(sc);
 729         CURVNET_SET(sc->sc_carpdev->if_vnet);
 730         carp_send_ad_locked(sc);
 731         CURVNET_RESTORE();
 732         CARP_UNLOCK(sc);
 733 }
 734
 735 static void
 736 carp_send_ad_locked(struct carp_softc *sc)
 737 {
 738         struct carp_header ch;
 739         struct timeval tv;
 740         struct sockaddr sa;
 741         struct ifaddr *ifa;
 742         struct carp_header *ch_ptr;
 743         struct mbuf *m;
 744         int len, advskew;
 745
 746         CARP_LOCK_ASSERT(sc);
 747
 748         advskew = DEMOTE_ADVSKEW(sc);
 749         tv.tv_sec = sc->sc_advbase;
 750         tv.tv_usec = advskew * 1000000 / 256;
 751
 752         ch.carp_version = CARP_VERSION;
 753         ch.carp_type = CARP_ADVERTISEMENT;
 754         ch.carp_vhid = sc->sc_vhid;
 755         ch.carp_advbase = sc->sc_advbase;
 756         ch.carp_advskew = advskew;
 757         ch.carp_authlen = 7;    /* XXX DEFINE */
 758         ch.carp_pad1 = 0;       /* must be zero */
 759         ch.carp_cksum = 0;
 760
 761         /* XXXGL: OpenBSD picks first ifaddr with needed family. */
 762
 763 #ifdef INET
 764         if (sc->sc_naddrs) {
 765                 struct ip *ip;
 766
 767                 m = m_gethdr(M_NOWAIT, MT_DATA);
 768                 if (m == NULL) {
 769                         CARPSTATS_INC(carps_onomem);
 770                         goto resched;
 771                 }
 772                 len = sizeof(*ip) + sizeof(ch);
 773                 m->m_pkthdr.len = len;
 774                 m->m_pkthdr.rcvif = NULL;
 775                 m->m_len = len;
 776                 MH_ALIGN(m, m->m_len);
 777                 m->m_flags |= M_MCAST;
 778                 ip = mtod(m, struct ip *);
 779                 ip->ip_v = IPVERSION;
 780                 ip->ip_hl = sizeof(*ip) >> 2;
 781                 ip->ip_tos = IPTOS_LOWDELAY;
 782                 ip->ip_len = htons(len);
 783                 ip->ip_id = ip_newid();
 784                 ip->ip_off = htons(IP_DF);
 785                 ip->ip_ttl = CARP_DFLTTL;
 786                 ip->ip_p = IPPROTO_CARP;
 787                 ip->ip_sum = 0;
 788
 789                 bzero(&sa, sizeof(sa));
 790                 sa.sa_family = AF_INET;
 791                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 792                 if (ifa != NULL) {
 793                         ip->ip_src.s_addr =
 794                             ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 795                         ifa_free(ifa);
 796                 } else
 797                         ip->ip_src.s_addr = 0;
 798                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 799
 800                 ch_ptr = (struct carp_header *)(&ip[1]);
 801                 bcopy(&ch, ch_ptr, sizeof(ch));
 802                 if (carp_prepare_ad(m, sc, ch_ptr))
 803                         goto resched;
 804
 805                 m->m_data += sizeof(*ip);
 806                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip));
 807                 m->m_data -= sizeof(*ip);
 808
 809                 CARPSTATS_INC(carps_opackets);
 810
 811                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT,
 812                     &sc->sc_carpdev->if_carp->cif_imo, NULL)) {
 813                         if (sc->sc_sendad_errors < INT_MAX)
 814                                 sc->sc_sendad_errors++;
 815                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS)
 816                                 carp_demote_adj(carp_senderr_adj, "send error");
 817                         sc->sc_sendad_success = 0;
 818                 } else {
 819                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 820                                 if (++sc->sc_sendad_success >=
 821                                     CARP_SENDAD_MIN_SUCCESS) {
 822                                         carp_demote_adj(-carp_senderr_adj,
 823                                             "send ok");
 824                                         sc->sc_sendad_errors = 0;
 825                                 }
 826                         } else
 827                                 sc->sc_sendad_errors = 0;
 828                 }
 829         }
 830 #endif /* INET */
 831 #ifdef INET6
 832         if (sc->sc_naddrs6) {
 833                 struct ip6_hdr *ip6;
 834
 835                 m = m_gethdr(M_NOWAIT, MT_DATA);
 836                 if (m == NULL) {
 837                         CARPSTATS_INC(carps_onomem);
 838                         goto resched;
 839                 }
 840                 len = sizeof(*ip6) + sizeof(ch);
 841                 m->m_pkthdr.len = len;
 842                 m->m_pkthdr.rcvif = NULL;
 843                 m->m_len = len;
 844                 MH_ALIGN(m, m->m_len);
 845                 m->m_flags |= M_MCAST;
 846                 ip6 = mtod(m, struct ip6_hdr *);
 847                 bzero(ip6, sizeof(*ip6));
 848                 ip6->ip6_vfc |= IPV6_VERSION;
 849                 ip6->ip6_hlim = CARP_DFLTTL;
 850                 ip6->ip6_nxt = IPPROTO_CARP;
 851                 bzero(&sa, sizeof(sa));
 852
 853                 /* set the source address */
 854                 sa.sa_family = AF_INET6;
 855                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 856                 if (ifa != NULL) {
 857                         bcopy(IFA_IN6(ifa), &ip6->ip6_src,
 858                             sizeof(struct in6_addr));
 859                         ifa_free(ifa);
 860                 } else
 861                         /* This should never happen with IPv6. */
 862                         bzero(&ip6->ip6_src, sizeof(struct in6_addr));
 863
 864                 /* Set the multicast destination. */
 865                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
 866                 ip6->ip6_dst.s6_addr8[15] = 0x12;
 867                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
 868                         m_freem(m);
 869                         CARP_DEBUG("%s: in6_setscope failed\n", __func__);
 870                         goto resched;
 871                 }
 872
 873                 ch_ptr = (struct carp_header *)(&ip6[1]);
 874                 bcopy(&ch, ch_ptr, sizeof(ch));
 875                 if (carp_prepare_ad(m, sc, ch_ptr))
 876                         goto resched;
 877
 878                 m->m_data += sizeof(*ip6);
 879                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip6));
 880                 m->m_data -= sizeof(*ip6);
 881
 882                 CARPSTATS_INC(carps_opackets6);
 883
 884                 if (ip6_output(m, NULL, NULL, 0,
 885                     &sc->sc_carpdev->if_carp->cif_im6o, NULL, NULL)) {
 886                         if (sc->sc_sendad_errors < INT_MAX)
 887                                 sc->sc_sendad_errors++;
 888                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS)
 889                                 carp_demote_adj(carp_senderr_adj,
 890                                     "send6 error");
 891                         sc->sc_sendad_success = 0;
 892                 } else {
 893                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 894                                 if (++sc->sc_sendad_success >=
 895                                     CARP_SENDAD_MIN_SUCCESS) {
 896                                         carp_demote_adj(-carp_senderr_adj,
 897                                             "send6 ok");
 898                                         sc->sc_sendad_errors = 0;
 899                                 }
 900                         } else
 901                                 sc->sc_sendad_errors = 0;
 902                 }
 903         }
 904 #endif /* INET6 */
 905
 906 resched:
 907         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), carp_send_ad, sc);
 908 }
 909
 910 static void
 911 carp_addroute(struct carp_softc *sc)
 912 {
 913         struct ifaddr *ifa;
 914
 915         CARP_FOREACH_IFA(sc, ifa)
 916                 carp_ifa_addroute(ifa);
 917 }
 918
 919 static void
 920 carp_ifa_addroute(struct ifaddr *ifa)
 921 {
 922
 923         switch (ifa->ifa_addr->sa_family) {
 924 #ifdef INET
 925         case AF_INET:
 926                 in_addprefix(ifatoia(ifa), RTF_UP);
 927                 ifa_add_loopback_route(ifa,
 928                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 929                 break;
 930 #endif
 931 #ifdef INET6
 932         case AF_INET6:
 933                 ifa_add_loopback_route(ifa,
 934                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 935                 in6_ifaddloop(ifa);
 936                 break;
 937 #endif
 938         }
 939 }
 940
 941 static void
 942 carp_delroute(struct carp_softc *sc)
 943 {
 944         struct ifaddr *ifa;
 945
 946         CARP_FOREACH_IFA(sc, ifa)
 947                 carp_ifa_delroute(ifa);
 948 }
 949
 950 static void
 951 carp_ifa_delroute(struct ifaddr *ifa)
 952 {
 953
 954         switch (ifa->ifa_addr->sa_family) {
 955 #ifdef INET
 956         case AF_INET:
 957                 ifa_del_loopback_route(ifa,
 958                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 959                 in_scrubprefix(ifatoia(ifa), LLE_STATIC);
 960                 break;
 961 #endif
 962 #ifdef INET6
 963         case AF_INET6:
 964                 ifa_del_loopback_route(ifa,
 965                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 966                 in6_ifremloop(ifa);
 967                 break;
 968 #endif
 969         }
 970 }
 971
 972 int
 973 carp_master(struct ifaddr *ifa)
 974 {
 975         struct carp_softc *sc = ifa->ifa_carp;
 976
 977         return (sc->sc_state == MASTER);
 978 }
 979
 980 #ifdef INET
 981 /*
 982  * Broadcast a gratuitous ARP request containing
 983  * the virtual router MAC address for each IP address
 984  * associated with the virtual router.
 985  */
 986 static void
 987 carp_send_arp(struct carp_softc *sc)
 988 {
 989         struct ifaddr *ifa;
 990
 991         CARP_FOREACH_IFA(sc, ifa)
 992                 if (ifa->ifa_addr->sa_family == AF_INET)
 993                         arp_ifinit2(sc->sc_carpdev, ifa, LLADDR(&sc->sc_addr));
 994 }
 995
 996 int
 997 carp_iamatch(struct ifaddr *ifa, uint8_t **enaddr)
 998 {
 999         struct carp_softc *sc = ifa->ifa_carp;
1000
1001         if (sc->sc_state == MASTER) {
1002                 *enaddr = LLADDR(&sc->sc_addr);
1003                 return (1);
1004         }
1005
1006         return (0);
1007 }
1008 #endif
1009
1010 #ifdef INET6
1011 static void
1012 carp_send_na(struct carp_softc *sc)
1013 {
1014         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1015         struct ifaddr *ifa;
1016         struct in6_addr *in6;
1017
1018         CARP_FOREACH_IFA(sc, ifa) {
1019                 if (ifa->ifa_addr->sa_family != AF_INET6)
1020                         continue;
1021
1022                 in6 = IFA_IN6(ifa);
1023                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1024                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1025                 DELAY(1000);    /* XXX */
1026         }
1027 }
1028
1029 /*
1030  * Returns ifa in case it's a carp address and it is MASTER, or if the address
1031  * matches and is not a carp address.  Returns NULL otherwise.
1032  */
1033 struct ifaddr *
1034 carp_iamatch6(struct ifnet *ifp, struct in6_addr *taddr)
1035 {
1036         struct ifaddr *ifa;
1037
1038         ifa = NULL;
1039         IF_ADDR_RLOCK(ifp);
1040         TAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
1041                 if (ifa->ifa_addr->sa_family != AF_INET6)
1042                         continue;
1043                 if (!IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa)))
1044                         continue;
1045                 if (ifa->ifa_carp && ifa->ifa_carp->sc_state != MASTER)
1046                         ifa = NULL;
1047                 else
1048                         ifa_ref(ifa);
1049                 break;
1050         }
1051         IF_ADDR_RUNLOCK(ifp);
1052
1053         return (ifa);
1054 }
1055
1056 caddr_t
1057 carp_macmatch6(struct ifnet *ifp, struct mbuf *m, const struct in6_addr *taddr)
1058 {
1059         struct ifaddr *ifa;
1060
1061         IF_ADDR_RLOCK(ifp);
1062         IFNET_FOREACH_IFA(ifp, ifa)
1063                 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1064                     IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1065                         struct carp_softc *sc = ifa->ifa_carp;
1066                         struct m_tag *mtag;
1067
1068                         IF_ADDR_RUNLOCK(ifp);
1069
1070                         mtag = m_tag_get(PACKET_TAG_CARP,
1071                             sizeof(struct carp_softc *), M_NOWAIT);
1072                         if (mtag == NULL)
1073                                 /* Better a bit than nothing. */
1074                                 return (LLADDR(&sc->sc_addr));
1075
1076                         bcopy(&sc, mtag + 1, sizeof(sc));
1077                         m_tag_prepend(m, mtag);
1078
1079                         return (LLADDR(&sc->sc_addr));
1080                 }
1081         IF_ADDR_RUNLOCK(ifp);
1082
1083         return (NULL);
1084 }
1085 #endif /* INET6 */
1086
1087 int
1088 carp_forus(struct ifnet *ifp, u_char *dhost)
1089 {
1090         struct carp_softc *sc;
1091         uint8_t *ena = dhost;
1092
1093         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1094                 return (0);
1095
1096         CIF_LOCK(ifp->if_carp);
1097         IFNET_FOREACH_CARP(ifp, sc) {
1098                 CARP_LOCK(sc);
1099                 if (sc->sc_state == MASTER && !bcmp(dhost, LLADDR(&sc->sc_addr),
1100                     ETHER_ADDR_LEN)) {
1101                         CARP_UNLOCK(sc);
1102                         CIF_UNLOCK(ifp->if_carp);
1103                         return (1);
1104                 }
1105                 CARP_UNLOCK(sc);
1106         }
1107         CIF_UNLOCK(ifp->if_carp);
1108
1109         return (0);
1110 }
1111
1112 /* Master down timeout event, executed in callout context. */
1113 static void
1114 carp_master_down(void *v)
1115 {
1116         struct carp_softc *sc = v;
1117
1118         CARP_LOCK_ASSERT(sc);
1119
1120         CURVNET_SET(sc->sc_carpdev->if_vnet);
1121         if (sc->sc_state == BACKUP) {
1122                 CARP_LOG("VHID %u@%s: BACKUP -> MASTER (master down)\n",
1123                     sc->sc_vhid,
1124                     sc->sc_carpdev->if_xname);
1125                 carp_master_down_locked(sc);
1126         }
1127         CURVNET_RESTORE();
1128
1129         CARP_UNLOCK(sc);
1130 }
1131
1132 static void
1133 carp_master_down_locked(struct carp_softc *sc)
1134 {
1135
1136         CARP_LOCK_ASSERT(sc);
1137
1138         switch (sc->sc_state) {
1139         case BACKUP:
1140                 carp_set_state(sc, MASTER);
1141                 carp_send_ad_locked(sc);
1142 #ifdef INET
1143                 carp_send_arp(sc);
1144 #endif
1145 #ifdef INET6
1146                 carp_send_na(sc);
1147 #endif
1148                 carp_setrun(sc, 0);
1149                 carp_addroute(sc);
1150                 break;
1151         case INIT:
1152         case MASTER:
1153 #ifdef INVARIANTS
1154                 panic("carp: VHID %u@%s: master_down event in %s state\n",
1155                     sc->sc_vhid,
1156                     sc->sc_carpdev->if_xname,
1157                     sc->sc_state ? "MASTER" : "INIT");
1158 #endif
1159                 break;
1160         }
1161 }
1162
1163 /*
1164  * When in backup state, af indicates whether to reset the master down timer
1165  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1166  */
1167 static void
1168 carp_setrun(struct carp_softc *sc, sa_family_t af)
1169 {
1170         struct timeval tv;
1171
1172         CARP_LOCK_ASSERT(sc);
1173
1174         if ((sc->sc_carpdev->if_flags & IFF_UP) == 0 ||
1175             sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1176             (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0))
1177                 return;
1178
1179         switch (sc->sc_state) {
1180         case INIT:
1181                 CARP_LOG("VHID %u@%s: INIT -> BACKUP\n",
1182                     sc->sc_vhid,
1183                     sc->sc_carpdev->if_xname);
1184                 carp_set_state(sc, BACKUP);
1185                 carp_setrun(sc, 0);
1186                 break;
1187         case BACKUP:
1188                 callout_stop(&sc->sc_ad_tmo);
1189                 tv.tv_sec = 3 * sc->sc_advbase;
1190                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1191                 switch (af) {
1192 #ifdef INET
1193                 case AF_INET:
1194                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1195                             carp_master_down, sc);
1196                         break;
1197 #endif
1198 #ifdef INET6
1199                 case AF_INET6:
1200                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1201                             carp_master_down, sc);
1202                         break;
1203 #endif
1204                 default:
1205 #ifdef INET
1206                         if (sc->sc_naddrs)
1207                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1208                                     carp_master_down, sc);
1209 #endif
1210 #ifdef INET6
1211                         if (sc->sc_naddrs6)
1212                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1213                                     carp_master_down, sc);
1214 #endif
1215                         break;
1216                 }
1217                 break;
1218         case MASTER:
1219                 tv.tv_sec = sc->sc_advbase;
1220                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1221                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1222                     carp_send_ad, sc);
1223                 break;
1224         }
1225 }
1226
1227 /*
1228  * Setup multicast structures.
1229  */
1230 static int
1231 carp_multicast_setup(struct carp_if *cif, sa_family_t sa)
1232 {
1233         struct ifnet *ifp = cif->cif_ifp;
1234         int error = 0;
1235
1236         CIF_LOCK_ASSERT(cif);
1237
1238         switch (sa) {
1239 #ifdef INET
1240         case AF_INET:
1241             {
1242                 struct ip_moptions *imo = &cif->cif_imo;
1243                 struct in_addr addr;
1244
1245                 if (imo->imo_membership)
1246                         return (0);
1247
1248                 imo->imo_membership = (struct in_multi **)malloc(
1249                     (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
1250                     M_NOWAIT);
1251                 if (imo->imo_membership == NULL)
1252                         return (ENOMEM);
1253                 imo->imo_mfilters = NULL;
1254                 imo->imo_max_memberships = IP_MIN_MEMBERSHIPS;
1255                 imo->imo_multicast_vif = -1;
1256
1257                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1258                 if ((error = in_joingroup(ifp, &addr, NULL,
1259                     &imo->imo_membership[0])) != 0) {
1260                         free(imo->imo_membership, M_CARP);
1261                         break;
1262                 }
1263                 imo->imo_num_memberships++;
1264                 imo->imo_multicast_ifp = ifp;
1265                 imo->imo_multicast_ttl = CARP_DFLTTL;
1266                 imo->imo_multicast_loop = 0;
1267                 break;
1268            }
1269 #endif
1270 #ifdef INET6
1271         case AF_INET6:
1272             {
1273                 struct ip6_moptions *im6o = &cif->cif_im6o;
1274                 struct in6_addr in6;
1275                 struct in6_multi *in6m;
1276
1277                 if (im6o->im6o_membership)
1278                         return (0);
1279
1280                 im6o->im6o_membership = (struct in6_multi **)malloc(
1281                     (sizeof(struct in6_multi *) * IPV6_MIN_MEMBERSHIPS), M_CARP,
1282                     M_ZERO | M_NOWAIT);
1283                 if (im6o->im6o_membership == NULL)
1284                         return (ENOMEM);
1285                 im6o->im6o_mfilters = NULL;
1286                 im6o->im6o_max_memberships = IPV6_MIN_MEMBERSHIPS;
1287                 im6o->im6o_multicast_hlim = CARP_DFLTTL;
1288                 im6o->im6o_multicast_ifp = ifp;
1289
1290                 /* Join IPv6 CARP multicast group. */
1291                 bzero(&in6, sizeof(in6));
1292                 in6.s6_addr16[0] = htons(0xff02);
1293                 in6.s6_addr8[15] = 0x12;
1294                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1295                         free(im6o->im6o_membership, M_CARP);
1296                         break;
1297                 }
1298                 in6m = NULL;
1299                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1300                         free(im6o->im6o_membership, M_CARP);
1301                         break;
1302                 }
1303                 im6o->im6o_membership[0] = in6m;
1304                 im6o->im6o_num_memberships++;
1305
1306                 /* Join solicited multicast address. */
1307                 bzero(&in6, sizeof(in6));
1308                 in6.s6_addr16[0] = htons(0xff02);
1309                 in6.s6_addr32[1] = 0;
1310                 in6.s6_addr32[2] = htonl(1);
1311                 in6.s6_addr32[3] = 0;
1312                 in6.s6_addr8[12] = 0xff;
1313                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1314                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1315                         free(im6o->im6o_membership, M_CARP);
1316                         break;
1317                 }
1318                 in6m = NULL;
1319                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1320                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1321                         free(im6o->im6o_membership, M_CARP);
1322                         break;
1323                 }
1324                 im6o->im6o_membership[1] = in6m;
1325                 im6o->im6o_num_memberships++;
1326                 break;
1327             }
1328 #endif
1329         }
1330
1331         return (error);
1332 }
1333
1334 /*
1335  * Free multicast structures.
1336  */
1337 static void
1338 carp_multicast_cleanup(struct carp_if *cif, sa_family_t sa)
1339 {
1340
1341         CIF_LOCK_ASSERT(cif);
1342         switch (sa) {
1343 #ifdef INET
1344         case AF_INET:
1345                 if (cif->cif_naddrs == 0) {
1346                         struct ip_moptions *imo = &cif->cif_imo;
1347
1348                         in_leavegroup(imo->imo_membership[0], NULL);
1349                         KASSERT(imo->imo_mfilters == NULL,
1350                             ("%s: imo_mfilters != NULL", __func__));
1351                         free(imo->imo_membership, M_CARP);
1352                         imo->imo_membership = NULL;
1353
1354                 }
1355                 break;
1356 #endif
1357 #ifdef INET6
1358         case AF_INET6:
1359                 if (cif->cif_naddrs6 == 0) {
1360                         struct ip6_moptions *im6o = &cif->cif_im6o;
1361
1362                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1363                         in6_mc_leave(im6o->im6o_membership[1], NULL);
1364                         KASSERT(im6o->im6o_mfilters == NULL,
1365                             ("%s: im6o_mfilters != NULL", __func__));
1366                         free(im6o->im6o_membership, M_CARP);
1367                         im6o->im6o_membership = NULL;
1368                 }
1369                 break;
1370 #endif
1371         }
1372 }
1373
1374 int
1375 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa)
1376 {
1377         struct m_tag *mtag;
1378         struct carp_softc *sc;
1379
1380         if (!sa)
1381                 return (0);
1382
1383         switch (sa->sa_family) {
1384 #ifdef INET
1385         case AF_INET:
1386                 break;
1387 #endif
1388 #ifdef INET6
1389         case AF_INET6:
1390                 break;
1391 #endif
1392         default:
1393                 return (0);
1394         }
1395
1396         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
1397         if (mtag == NULL)
1398                 return (0);
1399
1400         bcopy(mtag + 1, &sc, sizeof(sc));
1401
1402         /* Set the source MAC address to the Virtual Router MAC Address. */
1403         switch (ifp->if_type) {
1404         case IFT_ETHER:
1405         case IFT_BRIDGE:
1406         case IFT_L2VLAN: {
1407                         struct ether_header *eh;
1408
1409                         eh = mtod(m, struct ether_header *);
1410                         eh->ether_shost[0] = 0;
1411                         eh->ether_shost[1] = 0;
1412                         eh->ether_shost[2] = 0x5e;
1413                         eh->ether_shost[3] = 0;
1414                         eh->ether_shost[4] = 1;
1415                         eh->ether_shost[5] = sc->sc_vhid;
1416                 }
1417                 break;
1418         case IFT_FDDI: {
1419                         struct fddi_header *fh;
1420
1421                         fh = mtod(m, struct fddi_header *);
1422                         fh->fddi_shost[0] = 0;
1423                         fh->fddi_shost[1] = 0;
1424                         fh->fddi_shost[2] = 0x5e;
1425                         fh->fddi_shost[3] = 0;
1426                         fh->fddi_shost[4] = 1;
1427                         fh->fddi_shost[5] = sc->sc_vhid;
1428                 }
1429                 break;
1430         case IFT_ISO88025: {
1431                         struct iso88025_header *th;
1432                         th = mtod(m, struct iso88025_header *);
1433                         th->iso88025_shost[0] = 3;
1434                         th->iso88025_shost[1] = 0;
1435                         th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
1436                         th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
1437                         th->iso88025_shost[4] = 0;
1438                         th->iso88025_shost[5] = 0;
1439                 }
1440                 break;
1441         default:
1442                 printf("%s: carp is not supported for the %d interface type\n",
1443                     ifp->if_xname, ifp->if_type);
1444                 return (EOPNOTSUPP);
1445         }
1446
1447         return (0);
1448 }
1449
1450 static struct carp_softc*
1451 carp_alloc(struct ifnet *ifp)
1452 {
1453         struct carp_softc *sc;
1454         struct carp_if *cif;
1455
1456         if ((cif = ifp->if_carp) == NULL) {
1457                 cif = carp_alloc_if(ifp);
1458                 if (cif == NULL)
1459                         return (NULL);
1460         }
1461
1462         sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
1463
1464         sc->sc_advbase = CARP_DFLTINTV;
1465         sc->sc_vhid = -1;       /* required setting */
1466         sc->sc_init_counter = 1;
1467         sc->sc_state = INIT;
1468
1469         sc->sc_ifasiz = sizeof(struct ifaddr *);
1470         sc->sc_ifas = malloc(sc->sc_ifasiz, M_CARP, M_WAITOK|M_ZERO);
1471         sc->sc_carpdev = ifp;
1472
1473         CARP_LOCK_INIT(sc);
1474 #ifdef INET
1475         callout_init_mtx(&sc->sc_md_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1476 #endif
1477 #ifdef INET6
1478         callout_init_mtx(&sc->sc_md6_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1479 #endif
1480         callout_init_mtx(&sc->sc_ad_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1481
1482         CIF_LOCK(cif);
1483         TAILQ_INSERT_TAIL(&cif->cif_vrs, sc, sc_list);
1484         CIF_UNLOCK(cif);
1485
1486         mtx_lock(&carp_mtx);
1487         LIST_INSERT_HEAD(&carp_list, sc, sc_next);
1488         mtx_unlock(&carp_mtx);
1489
1490         return (sc);
1491 }
1492
1493 static int
1494 carp_grow_ifas(struct carp_softc *sc)
1495 {
1496         struct ifaddr **new;
1497
1498         CARP_LOCK_ASSERT(sc);
1499
1500         new = malloc(sc->sc_ifasiz * 2, M_CARP, M_NOWAIT|M_ZERO);
1501         if (new == NULL)
1502                 return (ENOMEM);
1503         bcopy(sc->sc_ifas, new, sc->sc_ifasiz);
1504         free(sc->sc_ifas, M_CARP);
1505         sc->sc_ifas = new;
1506         sc->sc_ifasiz *= 2;
1507
1508         return (0);
1509 }
1510
1511 static void
1512 carp_destroy(struct carp_softc *sc)
1513 {
1514         struct ifnet *ifp = sc->sc_carpdev;
1515         struct carp_if *cif = ifp->if_carp;
1516
1517         CIF_LOCK_ASSERT(cif);
1518
1519         TAILQ_REMOVE(&cif->cif_vrs, sc, sc_list);
1520
1521         mtx_lock(&carp_mtx);
1522         LIST_REMOVE(sc, sc_next);
1523         mtx_unlock(&carp_mtx);
1524
1525         CARP_LOCK(sc);
1526         if (sc->sc_suppress)
1527                 carp_demote_adj(-carp_ifdown_adj, "vhid removed");
1528         callout_drain(&sc->sc_ad_tmo);
1529 #ifdef INET
1530         callout_drain(&sc->sc_md_tmo);
1531 #endif
1532 #ifdef INET6
1533         callout_drain(&sc->sc_md6_tmo);
1534 #endif
1535         CARP_LOCK_DESTROY(sc);
1536
1537         free(sc->sc_ifas, M_CARP);
1538         free(sc, M_CARP);
1539 }
1540
1541 static struct carp_if*
1542 carp_alloc_if(struct ifnet *ifp)
1543 {
1544         struct carp_if *cif;
1545
1546         cif = malloc(sizeof(*cif), M_CARP, M_WAITOK|M_ZERO);
1547
1548         if (ifpromisc(ifp, 1) != 0)
1549                 goto cleanup;
1550
1551         CIF_LOCK_INIT(cif);
1552         cif->cif_ifp = ifp;
1553         TAILQ_INIT(&cif->cif_vrs);
1554
1555         IF_ADDR_WLOCK(ifp);
1556         ifp->if_carp = cif;
1557         if_ref(ifp);
1558         IF_ADDR_WUNLOCK(ifp);
1559
1560         return (cif);
1561
1562 cleanup:
1563         free(cif, M_CARP);
1564
1565         return (NULL);
1566 }
1567
1568 static void
1569 carp_free_if(struct carp_if *cif)
1570 {
1571         struct ifnet *ifp = cif->cif_ifp;
1572
1573         CIF_LOCK_ASSERT(cif);
1574         KASSERT(TAILQ_EMPTY(&cif->cif_vrs), ("%s: softc list not empty",
1575             __func__));
1576
1577         IF_ADDR_WLOCK(ifp);
1578         ifp->if_carp = NULL;
1579         if_rele(ifp);
1580         IF_ADDR_WUNLOCK(ifp);
1581
1582         CIF_LOCK_DESTROY(cif);
1583
1584         ifpromisc(ifp, 0);
1585
1586         free(cif, M_CARP);
1587 }
1588
1589 static void
1590 carp_carprcp(struct carpreq *carpr, struct carp_softc *sc, int priv)
1591 {
1592
1593         CARP_LOCK(sc);
1594         carpr->carpr_state = sc->sc_state;
1595         carpr->carpr_vhid = sc->sc_vhid;
1596         carpr->carpr_advbase = sc->sc_advbase;
1597         carpr->carpr_advskew = sc->sc_advskew;
1598         if (priv)
1599                 bcopy(sc->sc_key, carpr->carpr_key, sizeof(carpr->carpr_key));
1600         else
1601                 bzero(carpr->carpr_key, sizeof(carpr->carpr_key));
1602         CARP_UNLOCK(sc);
1603 }
1604
1605 int
1606 carp_ioctl(struct ifreq *ifr, u_long cmd, struct thread *td)
1607 {
1608         struct carpreq carpr;
1609         struct ifnet *ifp;
1610         struct carp_softc *sc = NULL;
1611         int error = 0, locked = 0;
1612
1613         if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1614                 return (error);
1615
1616         ifp = ifunit_ref(ifr->ifr_name);
1617         if (ifp == NULL)
1618                 return (ENXIO);
1619
1620         switch (ifp->if_type) {
1621         case IFT_ETHER:
1622         case IFT_L2VLAN:
1623         case IFT_BRIDGE:
1624         case IFT_FDDI:
1625         case IFT_ISO88025:
1626                 break;
1627         default:
1628                 error = EOPNOTSUPP;
1629                 goto out;
1630         }
1631
1632         if ((ifp->if_flags & IFF_MULTICAST) == 0) {
1633                 error = EADDRNOTAVAIL;
1634                 goto out;
1635         }
1636
1637         switch (cmd) {
1638         case SIOCSVH:
1639                 if ((error = priv_check(td, PRIV_NETINET_CARP)))
1640                         break;
1641                 if (carpr.carpr_vhid <= 0 || carpr.carpr_vhid > CARP_MAXVHID ||
1642                     carpr.carpr_advbase < 0 || carpr.carpr_advskew < 0) {
1643                         error = EINVAL;
1644                         break;
1645                 }
1646
1647                 if (ifp->if_carp) {
1648                         CIF_LOCK(ifp->if_carp);
1649                         IFNET_FOREACH_CARP(ifp, sc)
1650                                 if (sc->sc_vhid == carpr.carpr_vhid)
1651                                         break;
1652                         CIF_UNLOCK(ifp->if_carp);
1653                 }
1654                 if (sc == NULL) {
1655                         sc = carp_alloc(ifp);
1656                         if (sc == NULL) {
1657                                 error = EINVAL; /* XXX: ifpromisc failed */
1658                                 break;
1659                         }
1660
1661                         CARP_LOCK(sc);
1662                         sc->sc_vhid = carpr.carpr_vhid;
1663                         LLADDR(&sc->sc_addr)[0] = 0;
1664                         LLADDR(&sc->sc_addr)[1] = 0;
1665                         LLADDR(&sc->sc_addr)[2] = 0x5e;
1666                         LLADDR(&sc->sc_addr)[3] = 0;
1667                         LLADDR(&sc->sc_addr)[4] = 1;
1668                         LLADDR(&sc->sc_addr)[5] = sc->sc_vhid;
1669                 } else
1670                         CARP_LOCK(sc);
1671                 locked = 1;
1672                 if (carpr.carpr_advbase > 0) {
1673                         if (carpr.carpr_advbase > 255 ||
1674                             carpr.carpr_advbase < CARP_DFLTINTV) {
1675                                 error = EINVAL;
1676                                 break;
1677                         }
1678                         sc->sc_advbase = carpr.carpr_advbase;
1679                 }
1680                 if (carpr.carpr_advskew > 0) {
1681                         if (carpr.carpr_advskew >= 255) {
1682                                 error = EINVAL;
1683                                 break;
1684                         }
1685                         sc->sc_advskew = carpr.carpr_advskew;
1686                 }
1687                 if (carpr.carpr_key[0] != '\0') {
1688                         bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1689                         carp_hmac_prepare(sc);
1690                 }
1691                 if (sc->sc_state != INIT &&
1692                     carpr.carpr_state != sc->sc_state) {
1693                         switch (carpr.carpr_state) {
1694                         case BACKUP:
1695                                 callout_stop(&sc->sc_ad_tmo);
1696                                 carp_set_state(sc, BACKUP);
1697                                 carp_setrun(sc, 0);
1698                                 carp_delroute(sc);
1699                                 break;
1700                         case MASTER:
1701                                 carp_master_down_locked(sc);
1702                                 break;
1703                         default:
1704                                 break;
1705                         }
1706                 }
1707                 break;
1708
1709         case SIOCGVH:
1710             {
1711                 int priveleged;
1712
1713                 if (carpr.carpr_vhid < 0 || carpr.carpr_vhid > CARP_MAXVHID) {
1714                         error = EINVAL;
1715                         break;
1716                 }
1717                 if (carpr.carpr_count < 1) {
1718                         error = EMSGSIZE;
1719                         break;
1720                 }
1721                 if (ifp->if_carp == NULL) {
1722                         error = ENOENT;
1723                         break;
1724                 }
1725
1726                 priveleged = (priv_check(td, PRIV_NETINET_CARP) == 0);
1727                 if (carpr.carpr_vhid != 0) {
1728                         CIF_LOCK(ifp->if_carp);
1729                         IFNET_FOREACH_CARP(ifp, sc)
1730                                 if (sc->sc_vhid == carpr.carpr_vhid)
1731                                         break;
1732                         CIF_UNLOCK(ifp->if_carp);
1733                         if (sc == NULL) {
1734                                 error = ENOENT;
1735                                 break;
1736                         }
1737                         carp_carprcp(&carpr, sc, priveleged);
1738                         error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1739                 } else  {
1740                         int i, count;
1741
1742                         count = 0;
1743                         CIF_LOCK(ifp->if_carp);
1744                         IFNET_FOREACH_CARP(ifp, sc)
1745                                 count++;
1746
1747                         if (count > carpr.carpr_count) {
1748                                 CIF_UNLOCK(ifp->if_carp);
1749                                 error = EMSGSIZE;
1750                                 break;
1751                         }
1752
1753                         i = 0;
1754                         IFNET_FOREACH_CARP(ifp, sc) {
1755                                 carp_carprcp(&carpr, sc, priveleged);
1756                                 carpr.carpr_count = count;
1757                                 error = copyout(&carpr, ifr->ifr_data +
1758                                     (i * sizeof(carpr)), sizeof(carpr));
1759                                 if (error) {
1760                                         CIF_UNLOCK(ifp->if_carp);
1761                                         break;
1762                                 }
1763                                 i++;
1764                         }
1765                         CIF_UNLOCK(ifp->if_carp);
1766                 }
1767                 break;
1768             }
1769         default:
1770                 error = EINVAL;
1771         }
1772
1773 out:
1774         if (locked)
1775                 CARP_UNLOCK(sc);
1776         if_rele(ifp);
1777
1778         return (error);
1779 }
1780
1781 static int
1782 carp_get_vhid(struct ifaddr *ifa)
1783 {
1784
1785         if (ifa == NULL || ifa->ifa_carp == NULL)
1786                 return (0);
1787
1788         return (ifa->ifa_carp->sc_vhid);
1789 }
1790
1791 int
1792 carp_attach(struct ifaddr *ifa, int vhid)
1793 {
1794         struct ifnet *ifp = ifa->ifa_ifp;
1795         struct carp_if *cif = ifp->if_carp;
1796         struct carp_softc *sc;
1797         int index, error;
1798
1799         if (ifp->if_carp == NULL)
1800                 return (ENOPROTOOPT);
1801
1802         switch (ifa->ifa_addr->sa_family) {
1803 #ifdef INET
1804         case AF_INET:
1805 #endif
1806 #ifdef INET6
1807         case AF_INET6:
1808 #endif
1809                 break;
1810         default:
1811                 return (EPROTOTYPE);
1812         }
1813
1814         CIF_LOCK(cif);
1815         IFNET_FOREACH_CARP(ifp, sc)
1816                 if (sc->sc_vhid == vhid)
1817                         break;
1818         if (sc == NULL) {
1819                 CIF_UNLOCK(cif);
1820                 return (ENOENT);
1821         }
1822
1823         if (ifa->ifa_carp) {
1824                 if (ifa->ifa_carp->sc_vhid != vhid)
1825                         carp_detach_locked(ifa);
1826                 else {
1827                         CIF_UNLOCK(cif);
1828                         return (0);
1829                 }
1830         }
1831
1832         error = carp_multicast_setup(cif, ifa->ifa_addr->sa_family);
1833         if (error) {
1834                 CIF_FREE(cif);
1835                 return (error);
1836         }
1837
1838         CARP_LOCK(sc);
1839         index = sc->sc_naddrs + sc->sc_naddrs6 + 1;
1840         if (index > sc->sc_ifasiz / sizeof(struct ifaddr *))
1841                 if ((error = carp_grow_ifas(sc)) != 0) {
1842                         carp_multicast_cleanup(cif,
1843                             ifa->ifa_addr->sa_family);
1844                         CARP_UNLOCK(sc);
1845                         CIF_FREE(cif);
1846                         return (error);
1847                 }
1848
1849         switch (ifa->ifa_addr->sa_family) {
1850 #ifdef INET
1851         case AF_INET:
1852                 cif->cif_naddrs++;
1853                 sc->sc_naddrs++;
1854                 break;
1855 #endif
1856 #ifdef INET6
1857         case AF_INET6:
1858                 cif->cif_naddrs6++;
1859                 sc->sc_naddrs6++;
1860                 break;
1861 #endif
1862         }
1863
1864         ifa_ref(ifa);
1865         sc->sc_ifas[index - 1] = ifa;
1866         ifa->ifa_carp = sc;
1867
1868         carp_hmac_prepare(sc);
1869         carp_sc_state(sc);
1870
1871         CARP_UNLOCK(sc);
1872         CIF_UNLOCK(cif);
1873
1874         return (0);
1875 }
1876
1877 void
1878 carp_detach(struct ifaddr *ifa)
1879 {
1880         struct ifnet *ifp = ifa->ifa_ifp;
1881         struct carp_if *cif = ifp->if_carp;
1882
1883         CIF_LOCK(cif);
1884         carp_detach_locked(ifa);
1885         CIF_FREE(cif);
1886 }
1887
1888 static void
1889 carp_detach_locked(struct ifaddr *ifa)
1890 {
1891         struct ifnet *ifp = ifa->ifa_ifp;
1892         struct carp_if *cif = ifp->if_carp;
1893         struct carp_softc *sc = ifa->ifa_carp;
1894         int i, index;
1895
1896         KASSERT(sc != NULL, ("%s: %p not attached", __func__, ifa));
1897
1898         CIF_LOCK_ASSERT(cif);
1899         CARP_LOCK(sc);
1900
1901         /* Shift array. */
1902         index = sc->sc_naddrs + sc->sc_naddrs6;
1903         for (i = 0; i < index; i++)
1904                 if (sc->sc_ifas[i] == ifa)
1905                         break;
1906         KASSERT(i < index, ("%s: %p no backref", __func__, ifa));
1907         for (; i < index - 1; i++)
1908                 sc->sc_ifas[i] = sc->sc_ifas[i+1];
1909         sc->sc_ifas[index - 1] = NULL;
1910
1911         switch (ifa->ifa_addr->sa_family) {
1912 #ifdef INET
1913         case AF_INET:
1914                 cif->cif_naddrs--;
1915                 sc->sc_naddrs--;
1916                 break;
1917 #endif
1918 #ifdef INET6
1919         case AF_INET6:
1920                 cif->cif_naddrs6--;
1921                 sc->sc_naddrs6--;
1922                 break;
1923 #endif
1924         }
1925
1926         carp_ifa_delroute(ifa);
1927         carp_multicast_cleanup(cif, ifa->ifa_addr->sa_family);
1928
1929         ifa->ifa_carp = NULL;
1930         ifa_free(ifa);
1931
1932         carp_hmac_prepare(sc);
1933         carp_sc_state(sc);
1934
1935         if (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) {
1936                 CARP_UNLOCK(sc);
1937                 carp_destroy(sc);
1938         } else
1939                 CARP_UNLOCK(sc);
1940 }
1941
1942 static void
1943 carp_set_state(struct carp_softc *sc, int state)
1944 {
1945
1946         CARP_LOCK_ASSERT(sc);
1947
1948         if (sc->sc_state != state) {
1949                 const char *carp_states[] = { CARP_STATES };
1950                 char subsys[IFNAMSIZ+5];
1951
1952                 sc->sc_state = state;
1953
1954                 snprintf(subsys, IFNAMSIZ+5, "%u@%s", sc->sc_vhid,
1955                     sc->sc_carpdev->if_xname);
1956                 devctl_notify("CARP", subsys, carp_states[state], NULL);
1957         }
1958 }
1959
1960 static void
1961 carp_linkstate(struct ifnet *ifp)
1962 {
1963         struct carp_softc *sc;
1964
1965         CIF_LOCK(ifp->if_carp);
1966         IFNET_FOREACH_CARP(ifp, sc) {
1967                 CARP_LOCK(sc);
1968                 carp_sc_state(sc);
1969                 CARP_UNLOCK(sc);
1970         }
1971         CIF_UNLOCK(ifp->if_carp);
1972 }
1973
1974 static void
1975 carp_sc_state(struct carp_softc *sc)
1976 {
1977
1978         CARP_LOCK_ASSERT(sc);
1979
1980         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1981             !(sc->sc_carpdev->if_flags & IFF_UP)) {
1982                 callout_stop(&sc->sc_ad_tmo);
1983 #ifdef INET
1984                 callout_stop(&sc->sc_md_tmo);
1985 #endif
1986 #ifdef INET6
1987                 callout_stop(&sc->sc_md6_tmo);
1988 #endif
1989                 carp_set_state(sc, INIT);
1990                 carp_setrun(sc, 0);
1991                 if (!sc->sc_suppress)
1992                         carp_demote_adj(carp_ifdown_adj, "interface down");
1993                 sc->sc_suppress = 1;
1994         } else {
1995                 carp_set_state(sc, INIT);
1996                 carp_setrun(sc, 0);
1997                 if (sc->sc_suppress)
1998                         carp_demote_adj(-carp_ifdown_adj, "interface up");
1999                 sc->sc_suppress = 0;
2000         }
2001 }
2002
2003 static void
2004 carp_demote_adj(int adj, char *reason)
2005 {
2006         atomic_add_int(&carp_demotion, adj);
2007         CARP_LOG("demoted by %d to %d (%s)\n", adj, carp_demotion, reason);
2008         taskqueue_enqueue(taskqueue_swi, &carp_sendall_task);
2009 }
2010
2011 static int
2012 carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS)
2013 {
2014         int new, error;
2015
2016         new = carp_demotion;
2017         error = sysctl_handle_int(oidp, &new, 0, req);
2018         if (error || !req->newptr)
2019                 return (error);
2020
2021         carp_demote_adj(new, "sysctl");
2022
2023         return (0);
2024 }
2025
2026 #ifdef INET
2027 extern  struct domain inetdomain;
2028 static struct protosw in_carp_protosw = {
2029         .pr_type =              SOCK_RAW,
2030         .pr_domain =            &inetdomain,
2031         .pr_protocol =          IPPROTO_CARP,
2032         .pr_flags =             PR_ATOMIC|PR_ADDR,
2033         .pr_input =             carp_input,
2034         .pr_output =            (pr_output_t *)rip_output,
2035         .pr_ctloutput =         rip_ctloutput,
2036         .pr_usrreqs =           &rip_usrreqs
2037 };
2038 #endif
2039
2040 #ifdef INET6
2041 extern  struct domain inet6domain;
2042 static struct ip6protosw in6_carp_protosw = {
2043         .pr_type =              SOCK_RAW,
2044         .pr_domain =            &inet6domain,
2045         .pr_protocol =          IPPROTO_CARP,
2046         .pr_flags =             PR_ATOMIC|PR_ADDR,
2047         .pr_input =             carp6_input,
2048         .pr_output =            rip6_output,
2049         .pr_ctloutput =         rip6_ctloutput,
2050         .pr_usrreqs =           &rip6_usrreqs
2051 };
2052 #endif
2053
2054 static void
2055 carp_mod_cleanup(void)
2056 {
2057
2058 #ifdef INET
2059         if (proto_reg[CARP_INET] == 0) {
2060                 (void)ipproto_unregister(IPPROTO_CARP);
2061                 pf_proto_unregister(PF_INET, IPPROTO_CARP, SOCK_RAW);
2062                 proto_reg[CARP_INET] = -1;
2063         }
2064         carp_iamatch_p = NULL;
2065 #endif
2066 #ifdef INET6
2067         if (proto_reg[CARP_INET6] == 0) {
2068                 (void)ip6proto_unregister(IPPROTO_CARP);
2069                 pf_proto_unregister(PF_INET6, IPPROTO_CARP, SOCK_RAW);
2070                 proto_reg[CARP_INET6] = -1;
2071         }
2072         carp_iamatch6_p = NULL;
2073         carp_macmatch6_p = NULL;
2074 #endif
2075         carp_ioctl_p = NULL;
2076         carp_attach_p = NULL;
2077         carp_detach_p = NULL;
2078         carp_get_vhid_p = NULL;
2079         carp_linkstate_p = NULL;
2080         carp_forus_p = NULL;
2081         carp_output_p = NULL;
2082         carp_demote_adj_p = NULL;
2083         carp_master_p = NULL;
2084         mtx_unlock(&carp_mtx);
2085         taskqueue_drain(taskqueue_swi, &carp_sendall_task);
2086         mtx_destroy(&carp_mtx);
2087 }
2088
2089 static int
2090 carp_mod_load(void)
2091 {
2092         int err;
2093
2094         mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2095         LIST_INIT(&carp_list);
2096         carp_get_vhid_p = carp_get_vhid;
2097         carp_forus_p = carp_forus;
2098         carp_output_p = carp_output;
2099         carp_linkstate_p = carp_linkstate;
2100         carp_ioctl_p = carp_ioctl;
2101         carp_attach_p = carp_attach;
2102         carp_detach_p = carp_detach;
2103         carp_demote_adj_p = carp_demote_adj;
2104         carp_master_p = carp_master;
2105 #ifdef INET6
2106         carp_iamatch6_p = carp_iamatch6;
2107         carp_macmatch6_p = carp_macmatch6;
2108         proto_reg[CARP_INET6] = pf_proto_register(PF_INET6,
2109             (struct protosw *)&in6_carp_protosw);
2110         if (proto_reg[CARP_INET6]) {
2111                 printf("carp: error %d attaching to PF_INET6\n",
2112                     proto_reg[CARP_INET6]);
2113                 carp_mod_cleanup();
2114                 return (proto_reg[CARP_INET6]);
2115         }
2116         err = ip6proto_register(IPPROTO_CARP);
2117         if (err) {
2118                 printf("carp: error %d registering with INET6\n", err);
2119                 carp_mod_cleanup();
2120                 return (err);
2121         }
2122 #endif
2123 #ifdef INET
2124         carp_iamatch_p = carp_iamatch;
2125         proto_reg[CARP_INET] = pf_proto_register(PF_INET, &in_carp_protosw);
2126         if (proto_reg[CARP_INET]) {
2127                 printf("carp: error %d attaching to PF_INET\n",
2128                     proto_reg[CARP_INET]);
2129                 carp_mod_cleanup();
2130                 return (proto_reg[CARP_INET]);
2131         }
2132         err = ipproto_register(IPPROTO_CARP);
2133         if (err) {
2134                 printf("carp: error %d registering with INET\n", err);
2135                 carp_mod_cleanup();
2136                 return (err);
2137         }
2138 #endif
2139         return (0);
2140 }
2141
2142 static int
2143 carp_modevent(module_t mod, int type, void *data)
2144 {
2145         switch (type) {
2146         case MOD_LOAD:
2147                 return carp_mod_load();
2148                 /* NOTREACHED */
2149         case MOD_UNLOAD:
2150                 mtx_lock(&carp_mtx);
2151                 if (LIST_EMPTY(&carp_list))
2152                         carp_mod_cleanup();
2153                 else {
2154                         mtx_unlock(&carp_mtx);
2155                         return (EBUSY);
2156                 }
2157                 break;
2158
2159         default:
2160                 return (EINVAL);
2161         }
2162
2163         return (0);
2164 }
2165
2166 static moduledata_t carp_mod = {
2167         "carp",
2168         carp_modevent,
2169         0
2170 };
2171
2172 DECLARE_MODULE(carp, carp_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);