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 #define CIF_FREE(cif)   do {                            \
 227                 CIF_LOCK_ASSERT(cif);                   \
 228                 if (TAILQ_EMPTY(&(cif)->cif_vrs))       \
 229                         carp_free_if(cif);              \
 230                 else                                    \
 231                         CIF_UNLOCK(cif);                \
 232 } while (0)
 233
 234 #define CARP_LOG(...)   do {                            \
 235         if (carp_log > 0)                               \
 236                 log(LOG_INFO, "carp: " __VA_ARGS__);    \
 237 } while (0)
 238
 239 #define CARP_DEBUG(...) do {                            \
 240         if (carp_log > 1)                               \
 241                 log(LOG_DEBUG, __VA_ARGS__);            \
 242 } while (0)
 243
 244 #define IFNET_FOREACH_IFA(ifp, ifa)                                     \
 245         IF_ADDR_LOCK_ASSERT(ifp);                                       \
 246         TAILQ_FOREACH((ifa), &(ifp)->if_addrhead, ifa_link)             \
 247                 if ((ifa)->ifa_carp != NULL)
 248
 249 #define CARP_FOREACH_IFA(sc, ifa)                                       \
 250         CARP_LOCK_ASSERT(sc);                                           \
 251         for (int _i = 0;                                                \
 252                 _i < (sc)->sc_naddrs + (sc)->sc_naddrs6 &&              \
 253                 ((ifa) = sc->sc_ifas[_i]) != NULL;                      \
 254                 ++_i)
 255
 256 #define IFNET_FOREACH_CARP(ifp, sc)                                     \
 257         CIF_LOCK_ASSERT(ifp->if_carp);                                  \
 258         TAILQ_FOREACH((sc), &(ifp)->if_carp->cif_vrs, sc_list)
 259
 260 #define DEMOTE_ADVSKEW(sc)                                      \
 261     (((sc)->sc_advskew + carp_demotion > CARP_MAXSKEW) ?        \
 262     CARP_MAXSKEW : ((sc)->sc_advskew + carp_demotion))
 263
 264 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 265 static struct carp_softc
 266                 *carp_alloc(struct ifnet *);
 267 static void     carp_detach_locked(struct ifaddr *);
 268 static void     carp_destroy(struct carp_softc *);
 269 static struct carp_if
 270                 *carp_alloc_if(struct ifnet *);
 271 static void     carp_free_if(struct carp_if *);
 272 static void     carp_set_state(struct carp_softc *, int);
 273 static void     carp_sc_state(struct carp_softc *);
 274 static void     carp_setrun(struct carp_softc *, sa_family_t);
 275 static void     carp_master_down(void *);
 276 static void     carp_master_down_locked(struct carp_softc *);
 277 static void     carp_send_ad(void *);
 278 static void     carp_send_ad_locked(struct carp_softc *);
 279 static void     carp_addroute(struct carp_softc *);
 280 static void     carp_ifa_addroute(struct ifaddr *);
 281 static void     carp_delroute(struct carp_softc *);
 282 static void     carp_ifa_delroute(struct ifaddr *);
 283 static void     carp_send_ad_all(void *, int);
 284 static void     carp_demote_adj(int, char *);
 285
 286 static LIST_HEAD(, carp_softc) carp_list;
 287 static struct mtx carp_mtx;
 288 static struct task carp_sendall_task =
 289     TASK_INITIALIZER(0, carp_send_ad_all, NULL);
 290
 291 static __inline uint16_t
 292 carp_cksum(struct mbuf *m, int len)
 293 {
 294         return (in_cksum(m, len));
 295 }
 296
 297 static void
 298 carp_hmac_prepare(struct carp_softc *sc)
 299 {
 300         uint8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 301         uint8_t vhid = sc->sc_vhid & 0xff;
 302         struct ifaddr *ifa;
 303         int i, found;
 304 #ifdef INET
 305         struct in_addr last, cur, in;
 306 #endif
 307 #ifdef INET6
 308         struct in6_addr last6, cur6, in6;
 309 #endif
 310
 311         CARP_LOCK_ASSERT(sc);
 312
 313         /* Compute ipad from key. */
 314         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 315         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 316         for (i = 0; i < sizeof(sc->sc_pad); i++)
 317                 sc->sc_pad[i] ^= 0x36;
 318
 319         /* Precompute first part of inner hash. */
 320         SHA1Init(&sc->sc_sha1);
 321         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 322         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 323         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 324         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 325 #ifdef INET
 326         cur.s_addr = 0;
 327         do {
 328                 found = 0;
 329                 last = cur;
 330                 cur.s_addr = 0xffffffff;
 331                 CARP_FOREACH_IFA(sc, ifa) {
 332                         in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 333                         if (ifa->ifa_addr->sa_family == AF_INET &&
 334                             ntohl(in.s_addr) > ntohl(last.s_addr) &&
 335                             ntohl(in.s_addr) < ntohl(cur.s_addr)) {
 336                                 cur.s_addr = in.s_addr;
 337                                 found++;
 338                         }
 339                 }
 340                 if (found)
 341                         SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur));
 342         } while (found);
 343 #endif /* INET */
 344 #ifdef INET6
 345         memset(&cur6, 0, sizeof(cur6));
 346         do {
 347                 found = 0;
 348                 last6 = cur6;
 349                 memset(&cur6, 0xff, sizeof(cur6));
 350                 CARP_FOREACH_IFA(sc, ifa) {
 351                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 352                         if (IN6_IS_SCOPE_EMBED(&in6))
 353                                 in6.s6_addr16[1] = 0;
 354                         if (ifa->ifa_addr->sa_family == AF_INET6 &&
 355                             memcmp(&in6, &last6, sizeof(in6)) > 0 &&
 356                             memcmp(&in6, &cur6, sizeof(in6)) < 0) {
 357                                 cur6 = in6;
 358                                 found++;
 359                         }
 360                 }
 361                 if (found)
 362                         SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6));
 363         } while (found);
 364 #endif /* INET6 */
 365
 366         /* convert ipad to opad */
 367         for (i = 0; i < sizeof(sc->sc_pad); i++)
 368                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 369 }
 370
 371 static void
 372 carp_hmac_generate(struct carp_softc *sc, uint32_t counter[2],
 373     unsigned char md[20])
 374 {
 375         SHA1_CTX sha1ctx;
 376
 377         CARP_LOCK_ASSERT(sc);
 378
 379         /* fetch first half of inner hash */
 380         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 381
 382         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 383         SHA1Final(md, &sha1ctx);
 384
 385         /* outer hash */
 386         SHA1Init(&sha1ctx);
 387         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 388         SHA1Update(&sha1ctx, md, 20);
 389         SHA1Final(md, &sha1ctx);
 390 }
 391
 392 static int
 393 carp_hmac_verify(struct carp_softc *sc, uint32_t counter[2],
 394     unsigned char md[20])
 395 {
 396         unsigned char md2[20];
 397
 398         CARP_LOCK_ASSERT(sc);
 399
 400         carp_hmac_generate(sc, counter, md2);
 401
 402         return (bcmp(md, md2, sizeof(md2)));
 403 }
 404
 405 /*
 406  * process input packet.
 407  * we have rearranged checks order compared to the rfc,
 408  * but it seems more efficient this way or not possible otherwise.
 409  */
 410 #ifdef INET
 411 void
 412 carp_input(struct mbuf *m, int hlen)
 413 {
 414         struct ip *ip = mtod(m, struct ip *);
 415         struct carp_header *ch;
 416         int iplen, len;
 417
 418         CARPSTATS_INC(carps_ipackets);
 419
 420         if (!carp_allow) {
 421                 m_freem(m);
 422                 return;
 423         }
 424
 425         /* verify that the IP TTL is 255.  */
 426         if (ip->ip_ttl != CARP_DFLTTL) {
 427                 CARPSTATS_INC(carps_badttl);
 428                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 429                     ip->ip_ttl,
 430                     m->m_pkthdr.rcvif->if_xname);
 431                 m_freem(m);
 432                 return;
 433         }
 434
 435         iplen = ip->ip_hl << 2;
 436
 437         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 438                 CARPSTATS_INC(carps_badlen);
 439                 CARP_DEBUG("%s: received len %zd < sizeof(struct carp_header) "
 440                     "on %s\n", __func__, m->m_len - sizeof(struct ip),
 441                     m->m_pkthdr.rcvif->if_xname);
 442                 m_freem(m);
 443                 return;
 444         }
 445
 446         if (iplen + sizeof(*ch) < m->m_len) {
 447                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 448                         CARPSTATS_INC(carps_hdrops);
 449                         CARP_DEBUG("%s: pullup failed\n", __func__);
 450                         return;
 451                 }
 452                 ip = mtod(m, struct ip *);
 453         }
 454         ch = (struct carp_header *)((char *)ip + iplen);
 455
 456         /*
 457          * verify that the received packet length is
 458          * equal to the CARP header
 459          */
 460         len = iplen + sizeof(*ch);
 461         if (len > m->m_pkthdr.len) {
 462                 CARPSTATS_INC(carps_badlen);
 463                 CARP_DEBUG("%s: packet too short %d on %s\n", __func__,
 464                     m->m_pkthdr.len,
 465                     m->m_pkthdr.rcvif->if_xname);
 466                 m_freem(m);
 467                 return;
 468         }
 469
 470         if ((m = m_pullup(m, len)) == NULL) {
 471                 CARPSTATS_INC(carps_hdrops);
 472                 return;
 473         }
 474         ip = mtod(m, struct ip *);
 475         ch = (struct carp_header *)((char *)ip + iplen);
 476
 477         /* verify the CARP checksum */
 478         m->m_data += iplen;
 479         if (carp_cksum(m, len - iplen)) {
 480                 CARPSTATS_INC(carps_badsum);
 481                 CARP_DEBUG("%s: checksum failed on %s\n", __func__,
 482                     m->m_pkthdr.rcvif->if_xname);
 483                 m_freem(m);
 484                 return;
 485         }
 486         m->m_data -= iplen;
 487
 488         carp_input_c(m, ch, AF_INET);
 489 }
 490 #endif
 491
 492 #ifdef INET6
 493 int
 494 carp6_input(struct mbuf **mp, int *offp, int proto)
 495 {
 496         struct mbuf *m = *mp;
 497         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 498         struct carp_header *ch;
 499         u_int len;
 500
 501         CARPSTATS_INC(carps_ipackets6);
 502
 503         if (!carp_allow) {
 504                 m_freem(m);
 505                 return (IPPROTO_DONE);
 506         }
 507
 508         /* check if received on a valid carp interface */
 509         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 510                 CARPSTATS_INC(carps_badif);
 511                 CARP_DEBUG("%s: packet received on non-carp interface: %s\n",
 512                     __func__, m->m_pkthdr.rcvif->if_xname);
 513                 m_freem(m);
 514                 return (IPPROTO_DONE);
 515         }
 516
 517         /* verify that the IP TTL is 255 */
 518         if (ip6->ip6_hlim != CARP_DFLTTL) {
 519                 CARPSTATS_INC(carps_badttl);
 520                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 521                     ip6->ip6_hlim, m->m_pkthdr.rcvif->if_xname);
 522                 m_freem(m);
 523                 return (IPPROTO_DONE);
 524         }
 525
 526         /* verify that we have a complete carp packet */
 527         len = m->m_len;
 528         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 529         if (ch == NULL) {
 530                 CARPSTATS_INC(carps_badlen);
 531                 CARP_DEBUG("%s: packet size %u too small\n", __func__, len);
 532                 return (IPPROTO_DONE);
 533         }
 534
 535
 536         /* verify the CARP checksum */
 537         m->m_data += *offp;
 538         if (carp_cksum(m, sizeof(*ch))) {
 539                 CARPSTATS_INC(carps_badsum);
 540                 CARP_DEBUG("%s: checksum failed, on %s\n", __func__,
 541                     m->m_pkthdr.rcvif->if_xname);
 542                 m_freem(m);
 543                 return (IPPROTO_DONE);
 544         }
 545         m->m_data -= *offp;
 546
 547         carp_input_c(m, ch, AF_INET6);
 548         return (IPPROTO_DONE);
 549 }
 550 #endif /* INET6 */
 551
 552 static void
 553 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 554 {
 555         struct ifnet *ifp = m->m_pkthdr.rcvif;
 556         struct ifaddr *ifa;
 557         struct carp_softc *sc;
 558         uint64_t tmp_counter;
 559         struct timeval sc_tv, ch_tv;
 560
 561         /* verify that the VHID is valid on the receiving interface */
 562         IF_ADDR_RLOCK(ifp);
 563         IFNET_FOREACH_IFA(ifp, ifa)
 564                 if (ifa->ifa_addr->sa_family == af &&
 565                     ifa->ifa_carp->sc_vhid == ch->carp_vhid) {
 566                         ifa_ref(ifa);
 567                         break;
 568                 }
 569         IF_ADDR_RUNLOCK(ifp);
 570
 571         if (ifa == NULL) {
 572                 CARPSTATS_INC(carps_badvhid);
 573                 m_freem(m);
 574                 return;
 575         }
 576
 577         /* verify the CARP version. */
 578         if (ch->carp_version != CARP_VERSION) {
 579                 CARPSTATS_INC(carps_badver);
 580                 CARP_DEBUG("%s: invalid version %d\n", ifp->if_xname,
 581                     ch->carp_version);
 582                 ifa_free(ifa);
 583                 m_freem(m);
 584                 return;
 585         }
 586
 587         sc = ifa->ifa_carp;
 588         CARP_LOCK(sc);
 589         ifa_free(ifa);
 590
 591         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 592                 CARPSTATS_INC(carps_badauth);
 593                 CARP_DEBUG("%s: incorrect hash for VHID %u@%s\n", __func__,
 594                     sc->sc_vhid, ifp->if_xname);
 595                 goto out;
 596         }
 597
 598         tmp_counter = ntohl(ch->carp_counter[0]);
 599         tmp_counter = tmp_counter<<32;
 600         tmp_counter += ntohl(ch->carp_counter[1]);
 601
 602         /* XXX Replay protection goes here */
 603
 604         sc->sc_init_counter = 0;
 605         sc->sc_counter = tmp_counter;
 606
 607         sc_tv.tv_sec = sc->sc_advbase;
 608         sc_tv.tv_usec = DEMOTE_ADVSKEW(sc) * 1000000 / 256;
 609         ch_tv.tv_sec = ch->carp_advbase;
 610         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 611
 612         switch (sc->sc_state) {
 613         case INIT:
 614                 break;
 615         case MASTER:
 616                 /*
 617                  * If we receive an advertisement from a master who's going to
 618                  * be more frequent than us, go into BACKUP state.
 619                  */
 620                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 621                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 622                         callout_stop(&sc->sc_ad_tmo);
 623                         CARP_LOG("VHID %u@%s: MASTER -> BACKUP "
 624                             "(more frequent advertisement received)\n",
 625                             sc->sc_vhid,
 626                             sc->sc_carpdev->if_xname);
 627                         carp_set_state(sc, BACKUP);
 628                         carp_setrun(sc, 0);
 629                         carp_delroute(sc);
 630                 }
 631                 break;
 632         case BACKUP:
 633                 /*
 634                  * If we're pre-empting masters who advertise slower than us,
 635                  * and this one claims to be slower, treat him as down.
 636                  */
 637                 if (carp_preempt && timevalcmp(&sc_tv, &ch_tv, <)) {
 638                         CARP_LOG("VHID %u@%s: BACKUP -> MASTER "
 639                             "(preempting a slower master)\n",
 640                             sc->sc_vhid,
 641                             sc->sc_carpdev->if_xname);
 642                         carp_master_down_locked(sc);
 643                         break;
 644                 }
 645
 646                 /*
 647                  *  If the master is going to advertise at such a low frequency
 648                  *  that he's guaranteed to time out, we'd might as well just
 649                  *  treat him as timed out now.
 650                  */
 651                 sc_tv.tv_sec = sc->sc_advbase * 3;
 652                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 653                         CARP_LOG("VHID %u@%s: BACKUP -> MASTER "
 654                             "(master timed out)\n",
 655                             sc->sc_vhid,
 656                             sc->sc_carpdev->if_xname);
 657                         carp_master_down_locked(sc);
 658                         break;
 659                 }
 660
 661                 /*
 662                  * Otherwise, we reset the counter and wait for the next
 663                  * advertisement.
 664                  */
 665                 carp_setrun(sc, af);
 666                 break;
 667         }
 668
 669 out:
 670         CARP_UNLOCK(sc);
 671         m_freem(m);
 672 }
 673
 674 static int
 675 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 676 {
 677         struct m_tag *mtag;
 678
 679         if (sc->sc_init_counter) {
 680                 /* this could also be seconds since unix epoch */
 681                 sc->sc_counter = arc4random();
 682                 sc->sc_counter = sc->sc_counter << 32;
 683                 sc->sc_counter += arc4random();
 684         } else
 685                 sc->sc_counter++;
 686
 687         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 688         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 689
 690         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 691
 692         /* Tag packet for carp_output */
 693         if ((mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct carp_softc *),
 694             M_NOWAIT)) == NULL) {
 695                 m_freem(m);
 696                 CARPSTATS_INC(carps_onomem);
 697                 return (ENOMEM);
 698         }
 699         bcopy(&sc, (caddr_t)(mtag + 1), sizeof(struct carp_softc *));
 700         m_tag_prepend(m, mtag);
 701
 702         return (0);
 703 }
 704
 705 /*
 706  * To avoid LORs and possible recursions this function shouldn't
 707  * be called directly, but scheduled via taskqueue.
 708  */
 709 static void
 710 carp_send_ad_all(void *ctx __unused, int pending __unused)
 711 {
 712         struct carp_softc *sc;
 713
 714         mtx_lock(&carp_mtx);
 715         LIST_FOREACH(sc, &carp_list, sc_next)
 716                 if (sc->sc_state == MASTER) {
 717                         CARP_LOCK(sc);
 718                         CURVNET_SET(sc->sc_carpdev->if_vnet);
 719                         carp_send_ad_locked(sc);
 720                         CURVNET_RESTORE();
 721                         CARP_UNLOCK(sc);
 722                 }
 723         mtx_unlock(&carp_mtx);
 724 }
 725
 726 /* Send a periodic advertisement, executed in callout context. */
 727 static void
 728 carp_send_ad(void *v)
 729 {
 730         struct carp_softc *sc = v;
 731
 732         CARP_LOCK_ASSERT(sc);
 733         CURVNET_SET(sc->sc_carpdev->if_vnet);
 734         carp_send_ad_locked(sc);
 735         CURVNET_RESTORE();
 736         CARP_UNLOCK(sc);
 737 }
 738
 739 static void
 740 carp_send_ad_locked(struct carp_softc *sc)
 741 {
 742         struct carp_header ch;
 743         struct timeval tv;
 744         struct sockaddr sa;
 745         struct ifaddr *ifa;
 746         struct carp_header *ch_ptr;
 747         struct mbuf *m;
 748         int len, advskew;
 749
 750         CARP_LOCK_ASSERT(sc);
 751
 752         advskew = DEMOTE_ADVSKEW(sc);
 753         tv.tv_sec = sc->sc_advbase;
 754         tv.tv_usec = advskew * 1000000 / 256;
 755
 756         ch.carp_version = CARP_VERSION;
 757         ch.carp_type = CARP_ADVERTISEMENT;
 758         ch.carp_vhid = sc->sc_vhid;
 759         ch.carp_advbase = sc->sc_advbase;
 760         ch.carp_advskew = advskew;
 761         ch.carp_authlen = 7;    /* XXX DEFINE */
 762         ch.carp_pad1 = 0;       /* must be zero */
 763         ch.carp_cksum = 0;
 764
 765         /* XXXGL: OpenBSD picks first ifaddr with needed family. */
 766
 767 #ifdef INET
 768         if (sc->sc_naddrs) {
 769                 struct ip *ip;
 770
 771                 MGETHDR(m, M_NOWAIT, MT_HEADER);
 772                 if (m == NULL) {
 773                         CARPSTATS_INC(carps_onomem);
 774                         /* XXX maybe less ? */
 775                         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
 776                             carp_send_ad, sc);
 777                         return;
 778                 }
 779                 len = sizeof(*ip) + sizeof(ch);
 780                 m->m_pkthdr.len = len;
 781                 m->m_pkthdr.rcvif = NULL;
 782                 m->m_len = len;
 783                 MH_ALIGN(m, m->m_len);
 784                 m->m_flags |= M_MCAST;
 785                 ip = mtod(m, struct ip *);
 786                 ip->ip_v = IPVERSION;
 787                 ip->ip_hl = sizeof(*ip) >> 2;
 788                 ip->ip_tos = IPTOS_LOWDELAY;
 789                 ip->ip_len = len;
 790                 ip->ip_id = ip_newid();
 791                 ip->ip_off = IP_DF;
 792                 ip->ip_ttl = CARP_DFLTTL;
 793                 ip->ip_p = IPPROTO_CARP;
 794                 ip->ip_sum = 0;
 795
 796                 bzero(&sa, sizeof(sa));
 797                 sa.sa_family = AF_INET;
 798                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 799                 if (ifa != NULL) {
 800                         ip->ip_src.s_addr =
 801                             ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 802                         ifa_free(ifa);
 803                 } else
 804                         ip->ip_src.s_addr = 0;
 805                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 806
 807                 ch_ptr = (struct carp_header *)(&ip[1]);
 808                 bcopy(&ch, ch_ptr, sizeof(ch));
 809                 if (carp_prepare_ad(m, sc, ch_ptr))
 810                         return;
 811
 812                 m->m_data += sizeof(*ip);
 813                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip));
 814                 m->m_data -= sizeof(*ip);
 815
 816                 CARPSTATS_INC(carps_opackets);
 817
 818                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT,
 819                     &sc->sc_carpdev->if_carp->cif_imo, NULL)) {
 820                         if (sc->sc_sendad_errors < INT_MAX)
 821                                 sc->sc_sendad_errors++;
 822                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS)
 823                                 carp_demote_adj(carp_senderr_adj, "send error");
 824                         sc->sc_sendad_success = 0;
 825                 } else {
 826                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 827                                 if (++sc->sc_sendad_success >=
 828                                     CARP_SENDAD_MIN_SUCCESS) {
 829                                         carp_demote_adj(-carp_senderr_adj,
 830                                             "send ok");
 831                                         sc->sc_sendad_errors = 0;
 832                                 }
 833                         } else
 834                                 sc->sc_sendad_errors = 0;
 835                 }
 836         }
 837 #endif /* INET */
 838 #ifdef INET6
 839         if (sc->sc_naddrs6) {
 840                 struct ip6_hdr *ip6;
 841
 842                 MGETHDR(m, M_NOWAIT, MT_HEADER);
 843                 if (m == NULL) {
 844                         CARPSTATS_INC(carps_onomem);
 845                         /* XXX maybe less ? */
 846                         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
 847                             carp_send_ad, sc);
 848                         return;
 849                 }
 850                 len = sizeof(*ip6) + sizeof(ch);
 851                 m->m_pkthdr.len = len;
 852                 m->m_pkthdr.rcvif = NULL;
 853                 m->m_len = len;
 854                 MH_ALIGN(m, m->m_len);
 855                 m->m_flags |= M_MCAST;
 856                 ip6 = mtod(m, struct ip6_hdr *);
 857                 bzero(ip6, sizeof(*ip6));
 858                 ip6->ip6_vfc |= IPV6_VERSION;
 859                 ip6->ip6_hlim = CARP_DFLTTL;
 860                 ip6->ip6_nxt = IPPROTO_CARP;
 861                 bzero(&sa, sizeof(sa));
 862
 863                 /* set the source address */
 864                 sa.sa_family = AF_INET6;
 865                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 866                 if (ifa != NULL) {
 867                         bcopy(IFA_IN6(ifa), &ip6->ip6_src,
 868                             sizeof(struct in6_addr));
 869                         ifa_free(ifa);
 870                 } else
 871                         /* This should never happen with IPv6. */
 872                         bzero(&ip6->ip6_src, sizeof(struct in6_addr));
 873
 874                 /* Set the multicast destination. */
 875                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
 876                 ip6->ip6_dst.s6_addr8[15] = 0x12;
 877                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
 878                         m_freem(m);
 879                         CARP_DEBUG("%s: in6_setscope failed\n", __func__);
 880                         return;
 881                 }
 882
 883                 ch_ptr = (struct carp_header *)(&ip6[1]);
 884                 bcopy(&ch, ch_ptr, sizeof(ch));
 885                 if (carp_prepare_ad(m, sc, ch_ptr))
 886                         return;
 887
 888                 m->m_data += sizeof(*ip6);
 889                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6));
 890                 m->m_data -= sizeof(*ip6);
 891
 892                 CARPSTATS_INC(carps_opackets6);
 893
 894                 if (ip6_output(m, NULL, NULL, 0,
 895                     &sc->sc_carpdev->if_carp->cif_im6o, NULL, NULL)) {
 896                         if (sc->sc_sendad_errors < INT_MAX)
 897                                 sc->sc_sendad_errors++;
 898                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS)
 899                                 carp_demote_adj(carp_senderr_adj,
 900                                     "send6 error");
 901                         sc->sc_sendad_success = 0;
 902                 } else {
 903                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 904                                 if (++sc->sc_sendad_success >=
 905                                     CARP_SENDAD_MIN_SUCCESS) {
 906                                         carp_demote_adj(-carp_senderr_adj,
 907                                             "send6 ok");
 908                                         sc->sc_sendad_errors = 0;
 909                                 }
 910                         } else
 911                                 sc->sc_sendad_errors = 0;
 912                 }
 913         }
 914 #endif /* INET6 */
 915
 916         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), carp_send_ad, sc);
 917 }
 918
 919 static void
 920 carp_addroute(struct carp_softc *sc)
 921 {
 922         struct ifaddr *ifa;
 923
 924         CARP_FOREACH_IFA(sc, ifa)
 925                 carp_ifa_addroute(ifa);
 926 }
 927
 928 static void
 929 carp_ifa_addroute(struct ifaddr *ifa)
 930 {
 931
 932         switch (ifa->ifa_addr->sa_family) {
 933 #ifdef INET
 934         case AF_INET:
 935                 in_addprefix(ifatoia(ifa), RTF_UP);
 936                 ifa_add_loopback_route(ifa,
 937                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 938                 break;
 939 #endif
 940 #ifdef INET6
 941         case AF_INET6:
 942                 ifa_add_loopback_route(ifa,
 943                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 944                 in6_ifaddloop(ifa);
 945                 break;
 946 #endif
 947         }
 948 }
 949
 950 static void
 951 carp_delroute(struct carp_softc *sc)
 952 {
 953         struct ifaddr *ifa;
 954
 955         CARP_FOREACH_IFA(sc, ifa)
 956                 carp_ifa_delroute(ifa);
 957 }
 958
 959 static void
 960 carp_ifa_delroute(struct ifaddr *ifa)
 961 {
 962
 963         switch (ifa->ifa_addr->sa_family) {
 964 #ifdef INET
 965         case AF_INET:
 966                 ifa_del_loopback_route(ifa,
 967                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 968                 in_scrubprefix(ifatoia(ifa), LLE_STATIC);
 969                 break;
 970 #endif
 971 #ifdef INET6
 972         case AF_INET6:
 973                 ifa_del_loopback_route(ifa,
 974                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 975                 in6_ifremloop(ifa);
 976                 break;
 977 #endif
 978         }
 979 }
 980
 981 #ifdef INET
 982 /*
 983  * Broadcast a gratuitous ARP request containing
 984  * the virtual router MAC address for each IP address
 985  * associated with the virtual router.
 986  */
 987 static void
 988 carp_send_arp(struct carp_softc *sc)
 989 {
 990         struct ifaddr *ifa;
 991
 992         CARP_FOREACH_IFA(sc, ifa)
 993                 if (ifa->ifa_addr->sa_family == AF_INET)
 994                         arp_ifinit2(sc->sc_carpdev, ifa, LLADDR(&sc->sc_addr));
 995 }
 996
 997 int
 998 carp_iamatch(struct ifaddr *ifa, uint8_t **enaddr)
 999 {
1000         struct carp_softc *sc = ifa->ifa_carp;
1001
1002         if (sc->sc_state == MASTER) {
1003                 *enaddr = LLADDR(&sc->sc_addr);
1004                 return (1);
1005         }
1006
1007         return (0);
1008 }
1009 #endif
1010
1011 #ifdef INET6
1012 static void
1013 carp_send_na(struct carp_softc *sc)
1014 {
1015         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1016         struct ifaddr *ifa;
1017         struct in6_addr *in6;
1018
1019         CARP_FOREACH_IFA(sc, ifa) {
1020                 if (ifa->ifa_addr->sa_family != AF_INET6)
1021                         continue;
1022
1023                 in6 = IFA_IN6(ifa);
1024                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1025                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1026                 DELAY(1000);    /* XXX */
1027         }
1028 }
1029
1030 struct ifaddr *
1031 carp_iamatch6(struct ifnet *ifp, struct in6_addr *taddr)
1032 {
1033         struct ifaddr *ifa;
1034
1035         IF_ADDR_RLOCK(ifp);
1036         IFNET_FOREACH_IFA(ifp, ifa)
1037                 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1038                     ifa->ifa_carp->sc_state == MASTER &&
1039                     IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1040                         ifa_ref(ifa);
1041                         IF_ADDR_RUNLOCK(ifp);
1042                         return (ifa);
1043                 }
1044         IF_ADDR_RUNLOCK(ifp);
1045
1046         return (NULL);
1047 }
1048
1049 caddr_t
1050 carp_macmatch6(struct ifnet *ifp, struct mbuf *m, const struct in6_addr *taddr)
1051 {
1052         struct ifaddr *ifa;
1053
1054         IF_ADDR_RLOCK(ifp);
1055         IFNET_FOREACH_IFA(ifp, ifa)
1056                 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1057                     IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1058                         struct carp_softc *sc = ifa->ifa_carp;
1059                         struct m_tag *mtag;
1060
1061                         IF_ADDR_RUNLOCK(ifp);
1062
1063                         mtag = m_tag_get(PACKET_TAG_CARP,
1064                             sizeof(struct ifnet *), M_NOWAIT);
1065                         if (mtag == NULL)
1066                                 /* Better a bit than nothing. */
1067                                 return (LLADDR(&sc->sc_addr));
1068
1069                         bcopy(&ifp, (caddr_t)(mtag + 1),
1070                             sizeof(struct ifnet *));
1071                         m_tag_prepend(m, mtag);
1072
1073                         return (LLADDR(&sc->sc_addr));
1074                 }
1075         IF_ADDR_RUNLOCK(ifp);
1076
1077         return (NULL);
1078 }
1079 #endif /* INET6 */
1080
1081 int
1082 carp_forus(struct ifnet *ifp, u_char *dhost)
1083 {
1084         struct carp_softc *sc;
1085         uint8_t *ena = dhost;
1086
1087         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1088                 return (0);
1089
1090         CIF_LOCK(ifp->if_carp);
1091         IFNET_FOREACH_CARP(ifp, sc) {
1092                 CARP_LOCK(sc);
1093                 if (sc->sc_state == MASTER && !bcmp(dhost, LLADDR(&sc->sc_addr),
1094                     ETHER_ADDR_LEN)) {
1095                         CARP_UNLOCK(sc);
1096                         CIF_UNLOCK(ifp->if_carp);
1097                         return (1);
1098                 }
1099                 CARP_UNLOCK(sc);
1100         }
1101         CIF_UNLOCK(ifp->if_carp);
1102
1103         return (0);
1104 }
1105
1106 /* Master down timeout event, executed in callout context. */
1107 static void
1108 carp_master_down(void *v)
1109 {
1110         struct carp_softc *sc = v;
1111
1112         CARP_LOCK_ASSERT(sc);
1113
1114         CURVNET_SET(sc->sc_carpdev->if_vnet);
1115         if (sc->sc_state == BACKUP) {
1116                 CARP_LOG("VHID %u@%s: BACKUP -> MASTER (master down)\n",
1117                     sc->sc_vhid,
1118                     sc->sc_carpdev->if_xname);
1119                 carp_master_down_locked(sc);
1120         }
1121         CURVNET_RESTORE();
1122
1123         CARP_UNLOCK(sc);
1124 }
1125
1126 static void
1127 carp_master_down_locked(struct carp_softc *sc)
1128 {
1129
1130         CARP_LOCK_ASSERT(sc);
1131
1132         switch (sc->sc_state) {
1133         case BACKUP:
1134                 carp_set_state(sc, MASTER);
1135                 carp_send_ad_locked(sc);
1136 #ifdef INET
1137                 carp_send_arp(sc);
1138 #endif
1139 #ifdef INET6
1140                 carp_send_na(sc);
1141 #endif
1142                 carp_setrun(sc, 0);
1143                 carp_addroute(sc);
1144                 break;
1145         case INIT:
1146         case MASTER:
1147 #ifdef INVARIANTS
1148                 panic("carp: VHID %u@%s: master_down event in %s state\n",
1149                     sc->sc_vhid,
1150                     sc->sc_carpdev->if_xname,
1151                     sc->sc_state ? "MASTER" : "INIT");
1152 #endif
1153                 break;
1154         }
1155 }
1156
1157 /*
1158  * When in backup state, af indicates whether to reset the master down timer
1159  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1160  */
1161 static void
1162 carp_setrun(struct carp_softc *sc, sa_family_t af)
1163 {
1164         struct timeval tv;
1165
1166         CARP_LOCK_ASSERT(sc);
1167
1168         if ((sc->sc_carpdev->if_flags & IFF_UP) == 0 ||
1169             sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1170             (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0))
1171                 return;
1172
1173         switch (sc->sc_state) {
1174         case INIT:
1175                 CARP_LOG("VHID %u@%s: INIT -> BACKUP\n",
1176                     sc->sc_vhid,
1177                     sc->sc_carpdev->if_xname);
1178                 carp_set_state(sc, BACKUP);
1179                 carp_setrun(sc, 0);
1180                 break;
1181         case BACKUP:
1182                 callout_stop(&sc->sc_ad_tmo);
1183                 tv.tv_sec = 3 * sc->sc_advbase;
1184                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1185                 switch (af) {
1186 #ifdef INET
1187                 case AF_INET:
1188                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1189                             carp_master_down, sc);
1190                         break;
1191 #endif
1192 #ifdef INET6
1193                 case AF_INET6:
1194                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1195                             carp_master_down, sc);
1196                         break;
1197 #endif
1198                 default:
1199 #ifdef INET
1200                         if (sc->sc_naddrs)
1201                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1202                                     carp_master_down, sc);
1203 #endif
1204 #ifdef INET6
1205                         if (sc->sc_naddrs6)
1206                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1207                                     carp_master_down, sc);
1208 #endif
1209                         break;
1210                 }
1211                 break;
1212         case MASTER:
1213                 tv.tv_sec = sc->sc_advbase;
1214                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1215                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1216                     carp_send_ad, sc);
1217                 break;
1218         }
1219 }
1220
1221 /*
1222  * Setup multicast structures.
1223  */
1224 static int
1225 carp_multicast_setup(struct carp_if *cif, sa_family_t sa)
1226 {
1227         struct ifnet *ifp = cif->cif_ifp;
1228         int error = 0;
1229
1230         CIF_LOCK_ASSERT(cif);
1231
1232         switch (sa) {
1233 #ifdef INET
1234         case AF_INET:
1235             {
1236                 struct ip_moptions *imo = &cif->cif_imo;
1237                 struct in_addr addr;
1238
1239                 if (imo->imo_membership)
1240                         return (0);
1241
1242                 imo->imo_membership = (struct in_multi **)malloc(
1243                     (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
1244                     M_NOWAIT);
1245                 if (imo->imo_membership == NULL)
1246                         return (ENOMEM);
1247                 imo->imo_mfilters = NULL;
1248                 imo->imo_max_memberships = IP_MIN_MEMBERSHIPS;
1249                 imo->imo_multicast_vif = -1;
1250
1251                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1252                 if ((error = in_joingroup(ifp, &addr, NULL,
1253                     &imo->imo_membership[0])) != 0) {
1254                         free(imo->imo_membership, M_CARP);
1255                         break;
1256                 }
1257                 imo->imo_num_memberships++;
1258                 imo->imo_multicast_ifp = ifp;
1259                 imo->imo_multicast_ttl = CARP_DFLTTL;
1260                 imo->imo_multicast_loop = 0;
1261                 break;
1262            }
1263 #endif
1264 #ifdef INET6
1265         case AF_INET6:
1266             {
1267                 struct ip6_moptions *im6o = &cif->cif_im6o;
1268                 struct in6_addr in6;
1269                 struct in6_multi *in6m;
1270
1271                 if (im6o->im6o_membership)
1272                         return (0);
1273
1274                 im6o->im6o_membership = (struct in6_multi **)malloc(
1275                     (sizeof(struct in6_multi *) * IPV6_MIN_MEMBERSHIPS), M_CARP,
1276                     M_ZERO | M_NOWAIT);
1277                 if (im6o->im6o_membership == NULL)
1278                         return (ENOMEM);
1279                 im6o->im6o_mfilters = NULL;
1280                 im6o->im6o_max_memberships = IPV6_MIN_MEMBERSHIPS;
1281                 im6o->im6o_multicast_hlim = CARP_DFLTTL;
1282                 im6o->im6o_multicast_ifp = ifp;
1283
1284                 /* Join IPv6 CARP multicast group. */
1285                 bzero(&in6, sizeof(in6));
1286                 in6.s6_addr16[0] = htons(0xff02);
1287                 in6.s6_addr8[15] = 0x12;
1288                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1289                         free(im6o->im6o_membership, M_CARP);
1290                         break;
1291                 }
1292                 in6m = NULL;
1293                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1294                         free(im6o->im6o_membership, M_CARP);
1295                         break;
1296                 }
1297                 im6o->im6o_membership[0] = in6m;
1298                 im6o->im6o_num_memberships++;
1299
1300                 /* Join solicited multicast address. */
1301                 bzero(&in6, sizeof(in6));
1302                 in6.s6_addr16[0] = htons(0xff02);
1303                 in6.s6_addr32[1] = 0;
1304                 in6.s6_addr32[2] = htonl(1);
1305                 in6.s6_addr32[3] = 0;
1306                 in6.s6_addr8[12] = 0xff;
1307                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1308                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1309                         free(im6o->im6o_membership, M_CARP);
1310                         break;
1311                 }
1312                 in6m = NULL;
1313                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1314                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1315                         free(im6o->im6o_membership, M_CARP);
1316                         break;
1317                 }
1318                 im6o->im6o_membership[1] = in6m;
1319                 im6o->im6o_num_memberships++;
1320                 break;
1321             }
1322 #endif
1323         }
1324
1325         return (error);
1326 }
1327
1328 /*
1329  * Free multicast structures.
1330  */
1331 static void
1332 carp_multicast_cleanup(struct carp_if *cif, sa_family_t sa)
1333 {
1334
1335         CIF_LOCK_ASSERT(cif);
1336         switch (sa) {
1337 #ifdef INET
1338         case AF_INET:
1339                 if (cif->cif_naddrs == 0) {
1340                         struct ip_moptions *imo = &cif->cif_imo;
1341
1342                         in_leavegroup(imo->imo_membership[0], NULL);
1343                         KASSERT(imo->imo_mfilters == NULL,
1344                             ("%s: imo_mfilters != NULL", __func__));
1345                         free(imo->imo_membership, M_CARP);
1346                         imo->imo_membership = NULL;
1347
1348                 }
1349                 break;
1350 #endif
1351 #ifdef INET6
1352         case AF_INET6:
1353                 if (cif->cif_naddrs6 == 0) {
1354                         struct ip6_moptions *im6o = &cif->cif_im6o;
1355
1356                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1357                         in6_mc_leave(im6o->im6o_membership[1], NULL);
1358                         KASSERT(im6o->im6o_mfilters == NULL,
1359                             ("%s: im6o_mfilters != NULL", __func__));
1360                         free(im6o->im6o_membership, M_CARP);
1361                         im6o->im6o_membership = NULL;
1362                 }
1363                 break;
1364 #endif
1365         }
1366 }
1367
1368 int
1369 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *sa)
1370 {
1371         struct m_tag *mtag;
1372         struct carp_softc *sc;
1373
1374         if (!sa)
1375                 return (0);
1376
1377         switch (sa->sa_family) {
1378 #ifdef INET
1379         case AF_INET:
1380                 break;
1381 #endif
1382 #ifdef INET6
1383         case AF_INET6:
1384                 break;
1385 #endif
1386         default:
1387                 return (0);
1388         }
1389
1390         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
1391         if (mtag == NULL)
1392                 return (0);
1393
1394         bcopy(mtag + 1, &sc, sizeof(struct carp_softc *));
1395
1396         /* Set the source MAC address to the Virtual Router MAC Address. */
1397         switch (ifp->if_type) {
1398         case IFT_ETHER:
1399         case IFT_BRIDGE:
1400         case IFT_L2VLAN: {
1401                         struct ether_header *eh;
1402
1403                         eh = mtod(m, struct ether_header *);
1404                         eh->ether_shost[0] = 0;
1405                         eh->ether_shost[1] = 0;
1406                         eh->ether_shost[2] = 0x5e;
1407                         eh->ether_shost[3] = 0;
1408                         eh->ether_shost[4] = 1;
1409                         eh->ether_shost[5] = sc->sc_vhid;
1410                 }
1411                 break;
1412         case IFT_FDDI: {
1413                         struct fddi_header *fh;
1414
1415                         fh = mtod(m, struct fddi_header *);
1416                         fh->fddi_shost[0] = 0;
1417                         fh->fddi_shost[1] = 0;
1418                         fh->fddi_shost[2] = 0x5e;
1419                         fh->fddi_shost[3] = 0;
1420                         fh->fddi_shost[4] = 1;
1421                         fh->fddi_shost[5] = sc->sc_vhid;
1422                 }
1423                 break;
1424         case IFT_ISO88025: {
1425                         struct iso88025_header *th;
1426                         th = mtod(m, struct iso88025_header *);
1427                         th->iso88025_shost[0] = 3;
1428                         th->iso88025_shost[1] = 0;
1429                         th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
1430                         th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
1431                         th->iso88025_shost[4] = 0;
1432                         th->iso88025_shost[5] = 0;
1433                 }
1434                 break;
1435         default:
1436                 printf("%s: carp is not supported for the %d interface type\n",
1437                     ifp->if_xname, ifp->if_type);
1438                 return (EOPNOTSUPP);
1439         }
1440
1441         return (0);
1442 }
1443
1444 static struct carp_softc*
1445 carp_alloc(struct ifnet *ifp)
1446 {
1447         struct carp_softc *sc;
1448         struct carp_if *cif;
1449
1450         if ((cif = ifp->if_carp) == NULL) {
1451                 cif = carp_alloc_if(ifp);
1452                 if (cif == NULL)
1453                         return (NULL);
1454         }
1455
1456         sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
1457
1458         sc->sc_advbase = CARP_DFLTINTV;
1459         sc->sc_vhid = -1;       /* required setting */
1460         sc->sc_init_counter = 1;
1461         sc->sc_state = INIT;
1462
1463         sc->sc_ifasiz = sizeof(struct ifaddr *);
1464         sc->sc_ifas = malloc(sc->sc_ifasiz, M_CARP, M_WAITOK|M_ZERO);
1465         sc->sc_carpdev = ifp;
1466
1467         CARP_LOCK_INIT(sc);
1468 #ifdef INET
1469         callout_init_mtx(&sc->sc_md_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1470 #endif
1471 #ifdef INET6
1472         callout_init_mtx(&sc->sc_md6_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1473 #endif
1474         callout_init_mtx(&sc->sc_ad_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1475
1476         CIF_LOCK(cif);
1477         TAILQ_INSERT_TAIL(&cif->cif_vrs, sc, sc_list);
1478         CIF_UNLOCK(cif);
1479
1480         mtx_lock(&carp_mtx);
1481         LIST_INSERT_HEAD(&carp_list, sc, sc_next);
1482         mtx_unlock(&carp_mtx);
1483
1484         return (sc);
1485 }
1486
1487 static int
1488 carp_grow_ifas(struct carp_softc *sc)
1489 {
1490         struct ifaddr **new;
1491
1492         CARP_LOCK_ASSERT(sc);
1493
1494         new = malloc(sc->sc_ifasiz * 2, M_CARP, M_NOWAIT|M_ZERO);
1495         if (new == NULL)
1496                 return (ENOMEM);
1497         bcopy(sc->sc_ifas, new, sc->sc_ifasiz);
1498         free(sc->sc_ifas, M_CARP);
1499         sc->sc_ifas = new;
1500         sc->sc_ifasiz *= 2;
1501
1502         return (0);
1503 }
1504
1505 static void
1506 carp_destroy(struct carp_softc *sc)
1507 {
1508         struct ifnet *ifp = sc->sc_carpdev;
1509         struct carp_if *cif = ifp->if_carp;
1510
1511         CIF_LOCK_ASSERT(cif);
1512
1513         TAILQ_REMOVE(&cif->cif_vrs, sc, sc_list);
1514
1515         mtx_lock(&carp_mtx);
1516         LIST_REMOVE(sc, sc_next);
1517         mtx_unlock(&carp_mtx);
1518
1519         CARP_LOCK(sc);
1520         if (sc->sc_suppress)
1521                 carp_demote_adj(-carp_ifdown_adj, "vhid removed");
1522         callout_drain(&sc->sc_ad_tmo);
1523 #ifdef INET
1524         callout_drain(&sc->sc_md_tmo);
1525 #endif
1526 #ifdef INET6
1527         callout_drain(&sc->sc_md6_tmo);
1528 #endif
1529         CARP_LOCK_DESTROY(sc);
1530
1531         free(sc->sc_ifas, M_CARP);
1532         free(sc, M_CARP);
1533 }
1534
1535 static struct carp_if*
1536 carp_alloc_if(struct ifnet *ifp)
1537 {
1538         struct carp_if *cif;
1539
1540         cif = malloc(sizeof(*cif), M_CARP, M_WAITOK|M_ZERO);
1541
1542         if (ifpromisc(ifp, 1) != 0)
1543                 goto cleanup;
1544
1545         CIF_LOCK_INIT(cif);
1546         cif->cif_ifp = ifp;
1547         TAILQ_INIT(&cif->cif_vrs);
1548
1549         IF_ADDR_WLOCK(ifp);
1550         ifp->if_carp = cif;
1551         if_ref(ifp);
1552         IF_ADDR_WUNLOCK(ifp);
1553
1554         return (cif);
1555
1556 cleanup:
1557         free(cif, M_CARP);
1558
1559         return (NULL);
1560 }
1561
1562 static void
1563 carp_free_if(struct carp_if *cif)
1564 {
1565         struct ifnet *ifp = cif->cif_ifp;
1566
1567         CIF_LOCK_ASSERT(cif);
1568         KASSERT(TAILQ_EMPTY(&cif->cif_vrs), ("%s: softc list not empty",
1569             __func__));
1570
1571         IF_ADDR_WLOCK(ifp);
1572         ifp->if_carp = NULL;
1573         if_rele(ifp);
1574         IF_ADDR_WUNLOCK(ifp);
1575
1576         CIF_LOCK_DESTROY(cif);
1577
1578         ifpromisc(ifp, 0);
1579
1580         free(cif, M_CARP);
1581 }
1582
1583 static void
1584 carp_carprcp(struct carpreq *carpr, struct carp_softc *sc, int priv)
1585 {
1586
1587         CARP_LOCK(sc);
1588         carpr->carpr_state = sc->sc_state;
1589         carpr->carpr_vhid = sc->sc_vhid;
1590         carpr->carpr_advbase = sc->sc_advbase;
1591         carpr->carpr_advskew = sc->sc_advskew;
1592         if (priv)
1593                 bcopy(sc->sc_key, carpr->carpr_key, sizeof(carpr->carpr_key));
1594         else
1595                 bzero(carpr->carpr_key, sizeof(carpr->carpr_key));
1596         CARP_UNLOCK(sc);
1597 }
1598
1599 int
1600 carp_ioctl(struct ifreq *ifr, u_long cmd, struct thread *td)
1601 {
1602         struct carpreq carpr;
1603         struct ifnet *ifp;
1604         struct carp_softc *sc = NULL;
1605         int error = 0, locked = 0;
1606
1607         if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1608                 return (error);
1609
1610         ifp = ifunit_ref(ifr->ifr_name);
1611         if (ifp == NULL)
1612                 return (ENXIO);
1613
1614         switch (ifp->if_type) {
1615         case IFT_ETHER:
1616         case IFT_L2VLAN:
1617         case IFT_BRIDGE:
1618         case IFT_FDDI:
1619         case IFT_ISO88025:
1620                 break;
1621         default:
1622                 error = EOPNOTSUPP;
1623                 goto out;
1624         }
1625
1626         if ((ifp->if_flags & IFF_MULTICAST) == 0) {
1627                 error = EADDRNOTAVAIL;
1628                 goto out;
1629         }
1630
1631         switch (cmd) {
1632         case SIOCSVH:
1633                 if ((error = priv_check(td, PRIV_NETINET_CARP)))
1634                         break;
1635                 if (carpr.carpr_vhid <= 0 || carpr.carpr_vhid > CARP_MAXVHID ||
1636                     carpr.carpr_advbase < 0 || carpr.carpr_advskew < 0) {
1637                         error = EINVAL;
1638                         break;
1639                 }
1640
1641                 if (ifp->if_carp) {
1642                         CIF_LOCK(ifp->if_carp);
1643                         IFNET_FOREACH_CARP(ifp, sc)
1644                                 if (sc->sc_vhid == carpr.carpr_vhid)
1645                                         break;
1646                         CIF_UNLOCK(ifp->if_carp);
1647                 }
1648                 if (sc == NULL) {
1649                         sc = carp_alloc(ifp);
1650                         if (sc == NULL) {
1651                                 error = EINVAL; /* XXX: ifpromisc failed */
1652                                 break;
1653                         }
1654
1655                         CARP_LOCK(sc);
1656                         sc->sc_vhid = carpr.carpr_vhid;
1657                         LLADDR(&sc->sc_addr)[0] = 0;
1658                         LLADDR(&sc->sc_addr)[1] = 0;
1659                         LLADDR(&sc->sc_addr)[2] = 0x5e;
1660                         LLADDR(&sc->sc_addr)[3] = 0;
1661                         LLADDR(&sc->sc_addr)[4] = 1;
1662                         LLADDR(&sc->sc_addr)[5] = sc->sc_vhid;
1663                 } else
1664                         CARP_LOCK(sc);
1665                 locked = 1;
1666                 if (carpr.carpr_advbase > 0) {
1667                         if (carpr.carpr_advbase > 255 ||
1668                             carpr.carpr_advbase < CARP_DFLTINTV) {
1669                                 error = EINVAL;
1670                                 break;
1671                         }
1672                         sc->sc_advbase = carpr.carpr_advbase;
1673                 }
1674                 if (carpr.carpr_advskew > 0) {
1675                         if (carpr.carpr_advskew >= 255) {
1676                                 error = EINVAL;
1677                                 break;
1678                         }
1679                         sc->sc_advskew = carpr.carpr_advskew;
1680                 }
1681                 if (carpr.carpr_key[0] != '\0') {
1682                         bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1683                         carp_hmac_prepare(sc);
1684                 }
1685                 if (sc->sc_state != INIT &&
1686                     carpr.carpr_state != sc->sc_state) {
1687                         switch (carpr.carpr_state) {
1688                         case BACKUP:
1689                                 callout_stop(&sc->sc_ad_tmo);
1690                                 carp_set_state(sc, BACKUP);
1691                                 carp_setrun(sc, 0);
1692                                 carp_delroute(sc);
1693                                 break;
1694                         case MASTER:
1695                                 carp_master_down_locked(sc);
1696                                 break;
1697                         default:
1698                                 break;
1699                         }
1700                 }
1701                 break;
1702
1703         case SIOCGVH:
1704             {
1705                 int priveleged;
1706
1707                 if (carpr.carpr_vhid < 0 || carpr.carpr_vhid > CARP_MAXVHID) {
1708                         error = EINVAL;
1709                         break;
1710                 }
1711                 if (carpr.carpr_count < 1) {
1712                         error = EMSGSIZE;
1713                         break;
1714                 }
1715                 if (ifp->if_carp == NULL) {
1716                         error = ENOENT;
1717                         break;
1718                 }
1719
1720                 priveleged = (priv_check(td, PRIV_NETINET_CARP) == 0);
1721                 if (carpr.carpr_vhid != 0) {
1722                         CIF_LOCK(ifp->if_carp);
1723                         IFNET_FOREACH_CARP(ifp, sc)
1724                                 if (sc->sc_vhid == carpr.carpr_vhid)
1725                                         break;
1726                         CIF_UNLOCK(ifp->if_carp);
1727                         if (sc == NULL) {
1728                                 error = ENOENT;
1729                                 break;
1730                         }
1731                         carp_carprcp(&carpr, sc, priveleged);
1732                         error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1733                 } else  {
1734                         int i, count;
1735
1736                         count = 0;
1737                         CIF_LOCK(ifp->if_carp);
1738                         IFNET_FOREACH_CARP(ifp, sc)
1739                                 count++;
1740
1741                         if (count > carpr.carpr_count) {
1742                                 CIF_UNLOCK(ifp->if_carp);
1743                                 error = EMSGSIZE;
1744                                 break;
1745                         }
1746
1747                         i = 0;
1748                         IFNET_FOREACH_CARP(ifp, sc) {
1749                                 carp_carprcp(&carpr, sc, priveleged);
1750                                 carpr.carpr_count = count;
1751                                 error = copyout(&carpr, ifr->ifr_data +
1752                                     (i * sizeof(carpr)), sizeof(carpr));
1753                                 if (error) {
1754                                         CIF_UNLOCK(ifp->if_carp);
1755                                         break;
1756                                 }
1757                                 i++;
1758                         }
1759                         CIF_UNLOCK(ifp->if_carp);
1760                 }
1761                 break;
1762             }
1763         default:
1764                 error = EINVAL;
1765         }
1766
1767 out:
1768         if (locked)
1769                 CARP_UNLOCK(sc);
1770         if_rele(ifp);
1771
1772         return (error);
1773 }
1774
1775 static int
1776 carp_get_vhid(struct ifaddr *ifa)
1777 {
1778
1779         if (ifa == NULL || ifa->ifa_carp == NULL)
1780                 return (0);
1781
1782         return (ifa->ifa_carp->sc_vhid);
1783 }
1784
1785 int
1786 carp_attach(struct ifaddr *ifa, int vhid)
1787 {
1788         struct ifnet *ifp = ifa->ifa_ifp;
1789         struct carp_if *cif = ifp->if_carp;
1790         struct carp_softc *sc;
1791         int index, error;
1792
1793         if (ifp->if_carp == NULL)
1794                 return (ENOPROTOOPT);
1795
1796         switch (ifa->ifa_addr->sa_family) {
1797 #ifdef INET
1798         case AF_INET:
1799 #endif
1800 #ifdef INET6
1801         case AF_INET6:
1802 #endif
1803                 break;
1804         default:
1805                 return (EPROTOTYPE);
1806         }
1807
1808         CIF_LOCK(cif);
1809         IFNET_FOREACH_CARP(ifp, sc)
1810                 if (sc->sc_vhid == vhid)
1811                         break;
1812         if (sc == NULL) {
1813                 CIF_UNLOCK(cif);
1814                 return (ENOENT);
1815         }
1816
1817         if (ifa->ifa_carp) {
1818                 if (ifa->ifa_carp->sc_vhid != vhid)
1819                         carp_detach_locked(ifa);
1820                 else {
1821                         CIF_UNLOCK(cif);
1822                         return (0);
1823                 }
1824         }
1825
1826         error = carp_multicast_setup(cif, ifa->ifa_addr->sa_family);
1827         if (error) {
1828                 CIF_FREE(cif);
1829                 return (error);
1830         }
1831
1832         CARP_LOCK(sc);
1833         index = sc->sc_naddrs + sc->sc_naddrs6 + 1;
1834         if (index > sc->sc_ifasiz / sizeof(struct ifaddr *))
1835                 if ((error = carp_grow_ifas(sc)) != 0) {
1836                         carp_multicast_cleanup(cif,
1837                             ifa->ifa_addr->sa_family);
1838                         CARP_UNLOCK(sc);
1839                         CIF_FREE(cif);
1840                         return (error);
1841                 }
1842
1843         switch (ifa->ifa_addr->sa_family) {
1844 #ifdef INET
1845         case AF_INET:
1846                 cif->cif_naddrs++;
1847                 sc->sc_naddrs++;
1848                 break;
1849 #endif
1850 #ifdef INET6
1851         case AF_INET6:
1852                 cif->cif_naddrs6++;
1853                 sc->sc_naddrs6++;
1854                 break;
1855 #endif
1856         }
1857
1858         ifa_ref(ifa);
1859         sc->sc_ifas[index - 1] = ifa;
1860         ifa->ifa_carp = sc;
1861
1862         carp_hmac_prepare(sc);
1863         carp_sc_state(sc);
1864
1865         CARP_UNLOCK(sc);
1866         CIF_UNLOCK(cif);
1867
1868         return (0);
1869 }
1870
1871 void
1872 carp_detach(struct ifaddr *ifa)
1873 {
1874         struct ifnet *ifp = ifa->ifa_ifp;
1875         struct carp_if *cif = ifp->if_carp;
1876
1877         CIF_LOCK(cif);
1878         carp_detach_locked(ifa);
1879         CIF_FREE(cif);
1880 }
1881
1882 static void
1883 carp_detach_locked(struct ifaddr *ifa)
1884 {
1885         struct ifnet *ifp = ifa->ifa_ifp;
1886         struct carp_if *cif = ifp->if_carp;
1887         struct carp_softc *sc = ifa->ifa_carp;
1888         int i, index;
1889
1890         KASSERT(sc != NULL, ("%s: %p not attached", __func__, ifa));
1891
1892         CIF_LOCK_ASSERT(cif);
1893         CARP_LOCK(sc);
1894
1895         /* Shift array. */
1896         index = sc->sc_naddrs + sc->sc_naddrs6;
1897         for (i = 0; i < index; i++)
1898                 if (sc->sc_ifas[i] == ifa)
1899                         break;
1900         KASSERT(i < index, ("%s: %p no backref", __func__, ifa));
1901         for (; i < index - 1; i++)
1902                 sc->sc_ifas[i] = sc->sc_ifas[i+1];
1903         sc->sc_ifas[index - 1] = NULL;
1904
1905         switch (ifa->ifa_addr->sa_family) {
1906 #ifdef INET
1907         case AF_INET:
1908                 cif->cif_naddrs--;
1909                 sc->sc_naddrs--;
1910                 break;
1911 #endif
1912 #ifdef INET6
1913         case AF_INET6:
1914                 cif->cif_naddrs6--;
1915                 sc->sc_naddrs6--;
1916                 break;
1917 #endif
1918         }
1919
1920         carp_ifa_delroute(ifa);
1921         carp_multicast_cleanup(cif, ifa->ifa_addr->sa_family);
1922
1923         ifa->ifa_carp = NULL;
1924         ifa_free(ifa);
1925
1926         carp_hmac_prepare(sc);
1927         carp_sc_state(sc);
1928
1929         if (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) {
1930                 CARP_UNLOCK(sc);
1931                 carp_destroy(sc);
1932         } else
1933                 CARP_UNLOCK(sc);
1934 }
1935
1936 static void
1937 carp_set_state(struct carp_softc *sc, int state)
1938 {
1939
1940         CARP_LOCK_ASSERT(sc);
1941
1942         if (sc->sc_state != state) {
1943                 const char *carp_states[] = { CARP_STATES };
1944                 char subsys[IFNAMSIZ+5];
1945
1946                 sc->sc_state = state;
1947
1948                 snprintf(subsys, IFNAMSIZ+5, "%u@%s", sc->sc_vhid,
1949                     sc->sc_carpdev->if_xname);
1950                 devctl_notify("CARP", subsys, carp_states[state], NULL);
1951         }
1952 }
1953
1954 static void
1955 carp_linkstate(struct ifnet *ifp)
1956 {
1957         struct carp_softc *sc;
1958
1959         CIF_LOCK(ifp->if_carp);
1960         IFNET_FOREACH_CARP(ifp, sc) {
1961                 CARP_LOCK(sc);
1962                 carp_sc_state(sc);
1963                 CARP_UNLOCK(sc);
1964         }
1965         CIF_UNLOCK(ifp->if_carp);
1966 }
1967
1968 static void
1969 carp_sc_state(struct carp_softc *sc)
1970 {
1971
1972         CARP_LOCK_ASSERT(sc);
1973
1974         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1975             !(sc->sc_carpdev->if_flags & IFF_UP)) {
1976                 callout_stop(&sc->sc_ad_tmo);
1977 #ifdef INET
1978                 callout_stop(&sc->sc_md_tmo);
1979 #endif
1980 #ifdef INET6
1981                 callout_stop(&sc->sc_md6_tmo);
1982 #endif
1983                 carp_set_state(sc, INIT);
1984                 carp_setrun(sc, 0);
1985                 if (!sc->sc_suppress)
1986                         carp_demote_adj(carp_ifdown_adj, "interface down");
1987                 sc->sc_suppress = 1;
1988         } else {
1989                 carp_set_state(sc, INIT);
1990                 carp_setrun(sc, 0);
1991                 if (sc->sc_suppress)
1992                         carp_demote_adj(-carp_ifdown_adj, "interface up");
1993                 sc->sc_suppress = 0;
1994         }
1995 }
1996
1997 static void
1998 carp_demote_adj(int adj, char *reason)
1999 {
2000         carp_demotion += adj;
2001         CARP_LOG("demoted by %d to %d (%s)\n", adj, carp_demotion, reason);
2002         taskqueue_enqueue(taskqueue_swi, &carp_sendall_task);
2003 }
2004
2005 #ifdef INET
2006 extern  struct domain inetdomain;
2007 static struct protosw in_carp_protosw = {
2008         .pr_type =              SOCK_RAW,
2009         .pr_domain =            &inetdomain,
2010         .pr_protocol =          IPPROTO_CARP,
2011         .pr_flags =             PR_ATOMIC|PR_ADDR,
2012         .pr_input =             carp_input,
2013         .pr_output =            (pr_output_t *)rip_output,
2014         .pr_ctloutput =         rip_ctloutput,
2015         .pr_usrreqs =           &rip_usrreqs
2016 };
2017 #endif
2018
2019 #ifdef INET6
2020 extern  struct domain inet6domain;
2021 static struct ip6protosw in6_carp_protosw = {
2022         .pr_type =              SOCK_RAW,
2023         .pr_domain =            &inet6domain,
2024         .pr_protocol =          IPPROTO_CARP,
2025         .pr_flags =             PR_ATOMIC|PR_ADDR,
2026         .pr_input =             carp6_input,
2027         .pr_output =            rip6_output,
2028         .pr_ctloutput =         rip6_ctloutput,
2029         .pr_usrreqs =           &rip6_usrreqs
2030 };
2031 #endif
2032
2033 static void
2034 carp_mod_cleanup(void)
2035 {
2036
2037 #ifdef INET
2038         if (proto_reg[CARP_INET] == 0) {
2039                 (void)ipproto_unregister(IPPROTO_CARP);
2040                 pf_proto_unregister(PF_INET, IPPROTO_CARP, SOCK_RAW);
2041                 proto_reg[CARP_INET] = -1;
2042         }
2043         carp_iamatch_p = NULL;
2044 #endif
2045 #ifdef INET6
2046         if (proto_reg[CARP_INET6] == 0) {
2047                 (void)ip6proto_unregister(IPPROTO_CARP);
2048                 pf_proto_unregister(PF_INET6, IPPROTO_CARP, SOCK_RAW);
2049                 proto_reg[CARP_INET6] = -1;
2050         }
2051         carp_iamatch6_p = NULL;
2052         carp_macmatch6_p = NULL;
2053 #endif
2054         carp_ioctl_p = NULL;
2055         carp_attach_p = NULL;
2056         carp_detach_p = NULL;
2057         carp_get_vhid_p = NULL;
2058         carp_linkstate_p = NULL;
2059         carp_forus_p = NULL;
2060         carp_output_p = NULL;
2061         carp_demote_adj_p = NULL;
2062         mtx_unlock(&carp_mtx);
2063         taskqueue_drain(taskqueue_swi, &carp_sendall_task);
2064         mtx_destroy(&carp_mtx);
2065 }
2066
2067 static int
2068 carp_mod_load(void)
2069 {
2070         int err;
2071
2072         mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2073         LIST_INIT(&carp_list);
2074         carp_get_vhid_p = carp_get_vhid;
2075         carp_forus_p = carp_forus;
2076         carp_output_p = carp_output;
2077         carp_linkstate_p = carp_linkstate;
2078         carp_ioctl_p = carp_ioctl;
2079         carp_attach_p = carp_attach;
2080         carp_detach_p = carp_detach;
2081         carp_demote_adj_p = carp_demote_adj;
2082 #ifdef INET6
2083         carp_iamatch6_p = carp_iamatch6;
2084         carp_macmatch6_p = carp_macmatch6;
2085         proto_reg[CARP_INET6] = pf_proto_register(PF_INET6,
2086             (struct protosw *)&in6_carp_protosw);
2087         if (proto_reg[CARP_INET6]) {
2088                 printf("carp: error %d attaching to PF_INET6\n",
2089                     proto_reg[CARP_INET6]);
2090                 carp_mod_cleanup();
2091                 return (proto_reg[CARP_INET6]);
2092         }
2093         err = ip6proto_register(IPPROTO_CARP);
2094         if (err) {
2095                 printf("carp: error %d registering with INET6\n", err);
2096                 carp_mod_cleanup();
2097                 return (err);
2098         }
2099 #endif
2100 #ifdef INET
2101         carp_iamatch_p = carp_iamatch;
2102         proto_reg[CARP_INET] = pf_proto_register(PF_INET, &in_carp_protosw);
2103         if (proto_reg[CARP_INET]) {
2104                 printf("carp: error %d attaching to PF_INET\n",
2105                     proto_reg[CARP_INET]);
2106                 carp_mod_cleanup();
2107                 return (proto_reg[CARP_INET]);
2108         }
2109         err = ipproto_register(IPPROTO_CARP);
2110         if (err) {
2111                 printf("carp: error %d registering with INET\n", err);
2112                 carp_mod_cleanup();
2113                 return (err);
2114         }
2115 #endif
2116         return (0);
2117 }
2118
2119 static int
2120 carp_modevent(module_t mod, int type, void *data)
2121 {
2122         switch (type) {
2123         case MOD_LOAD:
2124                 return carp_mod_load();
2125                 /* NOTREACHED */
2126         case MOD_UNLOAD:
2127                 mtx_lock(&carp_mtx);
2128                 if (LIST_EMPTY(&carp_list))
2129                         carp_mod_cleanup();
2130                 else {
2131                         mtx_unlock(&carp_mtx);
2132                         return (EBUSY);
2133                 }
2134                 break;
2135
2136         default:
2137                 return (EINVAL);
2138         }
2139
2140         return (0);
2141 }
2142
2143 static moduledata_t carp_mod = {
2144         "carp",
2145         carp_modevent,
2146         0
2147 };
2148
2149 DECLARE_MODULE(carp, carp_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);