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, mtag + 1, sizeof(sc));
 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                         goto resched;
 775                 }
 776                 len = sizeof(*ip) + sizeof(ch);
 777                 m->m_pkthdr.len = len;
 778                 m->m_pkthdr.rcvif = NULL;
 779                 m->m_len = len;
 780                 MH_ALIGN(m, m->m_len);
 781                 m->m_flags |= M_MCAST;
 782                 ip = mtod(m, struct ip *);
 783                 ip->ip_v = IPVERSION;
 784                 ip->ip_hl = sizeof(*ip) >> 2;
 785                 ip->ip_tos = IPTOS_LOWDELAY;
 786                 ip->ip_len = len;
 787                 ip->ip_id = ip_newid();
 788                 ip->ip_off = IP_DF;
 789                 ip->ip_ttl = CARP_DFLTTL;
 790                 ip->ip_p = IPPROTO_CARP;
 791                 ip->ip_sum = 0;
 792
 793                 bzero(&sa, sizeof(sa));
 794                 sa.sa_family = AF_INET;
 795                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 796                 if (ifa != NULL) {
 797                         ip->ip_src.s_addr =
 798                             ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 799                         ifa_free(ifa);
 800                 } else
 801                         ip->ip_src.s_addr = 0;
 802                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 803
 804                 ch_ptr = (struct carp_header *)(&ip[1]);
 805                 bcopy(&ch, ch_ptr, sizeof(ch));
 806                 if (carp_prepare_ad(m, sc, ch_ptr))
 807                         goto resched;
 808
 809                 m->m_data += sizeof(*ip);
 810                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip));
 811                 m->m_data -= sizeof(*ip);
 812
 813                 CARPSTATS_INC(carps_opackets);
 814
 815                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT,
 816                     &sc->sc_carpdev->if_carp->cif_imo, NULL)) {
 817                         if (sc->sc_sendad_errors < INT_MAX)
 818                                 sc->sc_sendad_errors++;
 819                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS)
 820                                 carp_demote_adj(carp_senderr_adj, "send error");
 821                         sc->sc_sendad_success = 0;
 822                 } else {
 823                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 824                                 if (++sc->sc_sendad_success >=
 825                                     CARP_SENDAD_MIN_SUCCESS) {
 826                                         carp_demote_adj(-carp_senderr_adj,
 827                                             "send ok");
 828                                         sc->sc_sendad_errors = 0;
 829                                 }
 830                         } else
 831                                 sc->sc_sendad_errors = 0;
 832                 }
 833         }
 834 #endif /* INET */
 835 #ifdef INET6
 836         if (sc->sc_naddrs6) {
 837                 struct ip6_hdr *ip6;
 838
 839                 MGETHDR(m, M_NOWAIT, MT_HEADER);
 840                 if (m == NULL) {
 841                         CARPSTATS_INC(carps_onomem);
 842                         goto resched;
 843                 }
 844                 len = sizeof(*ip6) + sizeof(ch);
 845                 m->m_pkthdr.len = len;
 846                 m->m_pkthdr.rcvif = NULL;
 847                 m->m_len = len;
 848                 MH_ALIGN(m, m->m_len);
 849                 m->m_flags |= M_MCAST;
 850                 ip6 = mtod(m, struct ip6_hdr *);
 851                 bzero(ip6, sizeof(*ip6));
 852                 ip6->ip6_vfc |= IPV6_VERSION;
 853                 ip6->ip6_hlim = CARP_DFLTTL;
 854                 ip6->ip6_nxt = IPPROTO_CARP;
 855                 bzero(&sa, sizeof(sa));
 856
 857                 /* set the source address */
 858                 sa.sa_family = AF_INET6;
 859                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 860                 if (ifa != NULL) {
 861                         bcopy(IFA_IN6(ifa), &ip6->ip6_src,
 862                             sizeof(struct in6_addr));
 863                         ifa_free(ifa);
 864                 } else
 865                         /* This should never happen with IPv6. */
 866                         bzero(&ip6->ip6_src, sizeof(struct in6_addr));
 867
 868                 /* Set the multicast destination. */
 869                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
 870                 ip6->ip6_dst.s6_addr8[15] = 0x12;
 871                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
 872                         m_freem(m);
 873                         CARP_DEBUG("%s: in6_setscope failed\n", __func__);
 874                         goto resched;
 875                 }
 876
 877                 ch_ptr = (struct carp_header *)(&ip6[1]);
 878                 bcopy(&ch, ch_ptr, sizeof(ch));
 879                 if (carp_prepare_ad(m, sc, ch_ptr))
 880                         goto resched;
 881
 882                 m->m_data += sizeof(*ip6);
 883                 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6));
 884                 m->m_data -= sizeof(*ip6);
 885
 886                 CARPSTATS_INC(carps_opackets6);
 887
 888                 if (ip6_output(m, NULL, NULL, 0,
 889                     &sc->sc_carpdev->if_carp->cif_im6o, NULL, NULL)) {
 890                         if (sc->sc_sendad_errors < INT_MAX)
 891                                 sc->sc_sendad_errors++;
 892                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS)
 893                                 carp_demote_adj(carp_senderr_adj,
 894                                     "send6 error");
 895                         sc->sc_sendad_success = 0;
 896                 } else {
 897                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 898                                 if (++sc->sc_sendad_success >=
 899                                     CARP_SENDAD_MIN_SUCCESS) {
 900                                         carp_demote_adj(-carp_senderr_adj,
 901                                             "send6 ok");
 902                                         sc->sc_sendad_errors = 0;
 903                                 }
 904                         } else
 905                                 sc->sc_sendad_errors = 0;
 906                 }
 907         }
 908 #endif /* INET6 */
 909
 910 resched:
 911         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), carp_send_ad, sc);
 912 }
 913
 914 static void
 915 carp_addroute(struct carp_softc *sc)
 916 {
 917         struct ifaddr *ifa;
 918
 919         CARP_FOREACH_IFA(sc, ifa)
 920                 carp_ifa_addroute(ifa);
 921 }
 922
 923 static void
 924 carp_ifa_addroute(struct ifaddr *ifa)
 925 {
 926
 927         switch (ifa->ifa_addr->sa_family) {
 928 #ifdef INET
 929         case AF_INET:
 930                 in_addprefix(ifatoia(ifa), RTF_UP);
 931                 ifa_add_loopback_route(ifa,
 932                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 933                 break;
 934 #endif
 935 #ifdef INET6
 936         case AF_INET6:
 937                 ifa_add_loopback_route(ifa,
 938                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 939                 in6_ifaddloop(ifa);
 940                 break;
 941 #endif
 942         }
 943 }
 944
 945 static void
 946 carp_delroute(struct carp_softc *sc)
 947 {
 948         struct ifaddr *ifa;
 949
 950         CARP_FOREACH_IFA(sc, ifa)
 951                 carp_ifa_delroute(ifa);
 952 }
 953
 954 static void
 955 carp_ifa_delroute(struct ifaddr *ifa)
 956 {
 957
 958         switch (ifa->ifa_addr->sa_family) {
 959 #ifdef INET
 960         case AF_INET:
 961                 ifa_del_loopback_route(ifa,
 962                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 963                 in_scrubprefix(ifatoia(ifa), LLE_STATIC);
 964                 break;
 965 #endif
 966 #ifdef INET6
 967         case AF_INET6:
 968                 ifa_del_loopback_route(ifa,
 969                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 970                 in6_ifremloop(ifa);
 971                 break;
 972 #endif
 973         }
 974 }
 975
 976 #ifdef INET
 977 /*
 978  * Broadcast a gratuitous ARP request containing
 979  * the virtual router MAC address for each IP address
 980  * associated with the virtual router.
 981  */
 982 static void
 983 carp_send_arp(struct carp_softc *sc)
 984 {
 985         struct ifaddr *ifa;
 986
 987         CARP_FOREACH_IFA(sc, ifa)
 988                 if (ifa->ifa_addr->sa_family == AF_INET)
 989                         arp_ifinit2(sc->sc_carpdev, ifa, LLADDR(&sc->sc_addr));
 990 }
 991
 992 int
 993 carp_iamatch(struct ifaddr *ifa, uint8_t **enaddr)
 994 {
 995         struct carp_softc *sc = ifa->ifa_carp;
 996
 997         if (sc->sc_state == MASTER) {
 998                 *enaddr = LLADDR(&sc->sc_addr);
 999                 return (1);
1000         }
1001
1002         return (0);
1003 }
1004 #endif
1005
1006 #ifdef INET6
1007 static void
1008 carp_send_na(struct carp_softc *sc)
1009 {
1010         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1011         struct ifaddr *ifa;
1012         struct in6_addr *in6;
1013
1014         CARP_FOREACH_IFA(sc, ifa) {
1015                 if (ifa->ifa_addr->sa_family != AF_INET6)
1016                         continue;
1017
1018                 in6 = IFA_IN6(ifa);
1019                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1020                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1021                 DELAY(1000);    /* XXX */
1022         }
1023 }
1024
1025 /*
1026  * Returns ifa in case it's a carp address and it is MASTER, or if the address
1027  * matches and is not a carp address.  Returns NULL otherwise.
1028  */
1029 struct ifaddr *
1030 carp_iamatch6(struct ifnet *ifp, struct in6_addr *taddr)
1031 {
1032         struct ifaddr *ifa;
1033
1034         ifa = NULL;
1035         IF_ADDR_RLOCK(ifp);
1036         TAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
1037                 if (ifa->ifa_addr->sa_family != AF_INET6)
1038                         continue;
1039                 if (!IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa)))
1040                         continue;
1041                 if (ifa->ifa_carp && ifa->ifa_carp->sc_state != MASTER)
1042                         ifa = NULL;
1043                 else
1044                         ifa_ref(ifa);
1045                 break;
1046         }
1047         IF_ADDR_RUNLOCK(ifp);
1048
1049         return (ifa);
1050 }
1051
1052 caddr_t
1053 carp_macmatch6(struct ifnet *ifp, struct mbuf *m, const struct in6_addr *taddr)
1054 {
1055         struct ifaddr *ifa;
1056
1057         IF_ADDR_RLOCK(ifp);
1058         IFNET_FOREACH_IFA(ifp, ifa)
1059                 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1060                     IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1061                         struct carp_softc *sc = ifa->ifa_carp;
1062                         struct m_tag *mtag;
1063
1064                         IF_ADDR_RUNLOCK(ifp);
1065
1066                         mtag = m_tag_get(PACKET_TAG_CARP,
1067                             sizeof(struct carp_softc *), M_NOWAIT);
1068                         if (mtag == NULL)
1069                                 /* Better a bit than nothing. */
1070                                 return (LLADDR(&sc->sc_addr));
1071
1072                         bcopy(&sc, mtag + 1, sizeof(sc));
1073                         m_tag_prepend(m, mtag);
1074
1075                         return (LLADDR(&sc->sc_addr));
1076                 }
1077         IF_ADDR_RUNLOCK(ifp);
1078
1079         return (NULL);
1080 }
1081 #endif /* INET6 */
1082
1083 int
1084 carp_forus(struct ifnet *ifp, u_char *dhost)
1085 {
1086         struct carp_softc *sc;
1087         uint8_t *ena = dhost;
1088
1089         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1090                 return (0);
1091
1092         CIF_LOCK(ifp->if_carp);
1093         IFNET_FOREACH_CARP(ifp, sc) {
1094                 CARP_LOCK(sc);
1095                 if (sc->sc_state == MASTER && !bcmp(dhost, LLADDR(&sc->sc_addr),
1096                     ETHER_ADDR_LEN)) {
1097                         CARP_UNLOCK(sc);
1098                         CIF_UNLOCK(ifp->if_carp);
1099                         return (1);
1100                 }
1101                 CARP_UNLOCK(sc);
1102         }
1103         CIF_UNLOCK(ifp->if_carp);
1104
1105         return (0);
1106 }
1107
1108 /* Master down timeout event, executed in callout context. */
1109 static void
1110 carp_master_down(void *v)
1111 {
1112         struct carp_softc *sc = v;
1113
1114         CARP_LOCK_ASSERT(sc);
1115
1116         CURVNET_SET(sc->sc_carpdev->if_vnet);
1117         if (sc->sc_state == BACKUP) {
1118                 CARP_LOG("VHID %u@%s: BACKUP -> MASTER (master down)\n",
1119                     sc->sc_vhid,
1120                     sc->sc_carpdev->if_xname);
1121                 carp_master_down_locked(sc);
1122         }
1123         CURVNET_RESTORE();
1124
1125         CARP_UNLOCK(sc);
1126 }
1127
1128 static void
1129 carp_master_down_locked(struct carp_softc *sc)
1130 {
1131
1132         CARP_LOCK_ASSERT(sc);
1133
1134         switch (sc->sc_state) {
1135         case BACKUP:
1136                 carp_set_state(sc, MASTER);
1137                 carp_send_ad_locked(sc);
1138 #ifdef INET
1139                 carp_send_arp(sc);
1140 #endif
1141 #ifdef INET6
1142                 carp_send_na(sc);
1143 #endif
1144                 carp_setrun(sc, 0);
1145                 carp_addroute(sc);
1146                 break;
1147         case INIT:
1148         case MASTER:
1149 #ifdef INVARIANTS
1150                 panic("carp: VHID %u@%s: master_down event in %s state\n",
1151                     sc->sc_vhid,
1152                     sc->sc_carpdev->if_xname,
1153                     sc->sc_state ? "MASTER" : "INIT");
1154 #endif
1155                 break;
1156         }
1157 }
1158
1159 /*
1160  * When in backup state, af indicates whether to reset the master down timer
1161  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1162  */
1163 static void
1164 carp_setrun(struct carp_softc *sc, sa_family_t af)
1165 {
1166         struct timeval tv;
1167
1168         CARP_LOCK_ASSERT(sc);
1169
1170         if ((sc->sc_carpdev->if_flags & IFF_UP) == 0 ||
1171             sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1172             (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0))
1173                 return;
1174
1175         switch (sc->sc_state) {
1176         case INIT:
1177                 CARP_LOG("VHID %u@%s: INIT -> BACKUP\n",
1178                     sc->sc_vhid,
1179                     sc->sc_carpdev->if_xname);
1180                 carp_set_state(sc, BACKUP);
1181                 carp_setrun(sc, 0);
1182                 break;
1183         case BACKUP:
1184                 callout_stop(&sc->sc_ad_tmo);
1185                 tv.tv_sec = 3 * sc->sc_advbase;
1186                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1187                 switch (af) {
1188 #ifdef INET
1189                 case AF_INET:
1190                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1191                             carp_master_down, sc);
1192                         break;
1193 #endif
1194 #ifdef INET6
1195                 case AF_INET6:
1196                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1197                             carp_master_down, sc);
1198                         break;
1199 #endif
1200                 default:
1201 #ifdef INET
1202                         if (sc->sc_naddrs)
1203                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1204                                     carp_master_down, sc);
1205 #endif
1206 #ifdef INET6
1207                         if (sc->sc_naddrs6)
1208                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1209                                     carp_master_down, sc);
1210 #endif
1211                         break;
1212                 }
1213                 break;
1214         case MASTER:
1215                 tv.tv_sec = sc->sc_advbase;
1216                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1217                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1218                     carp_send_ad, sc);
1219                 break;
1220         }
1221 }
1222
1223 /*
1224  * Setup multicast structures.
1225  */
1226 static int
1227 carp_multicast_setup(struct carp_if *cif, sa_family_t sa)
1228 {
1229         struct ifnet *ifp = cif->cif_ifp;
1230         int error = 0;
1231
1232         CIF_LOCK_ASSERT(cif);
1233
1234         switch (sa) {
1235 #ifdef INET
1236         case AF_INET:
1237             {
1238                 struct ip_moptions *imo = &cif->cif_imo;
1239                 struct in_addr addr;
1240
1241                 if (imo->imo_membership)
1242                         return (0);
1243
1244                 imo->imo_membership = (struct in_multi **)malloc(
1245                     (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
1246                     M_NOWAIT);
1247                 if (imo->imo_membership == NULL)
1248                         return (ENOMEM);
1249                 imo->imo_mfilters = NULL;
1250                 imo->imo_max_memberships = IP_MIN_MEMBERSHIPS;
1251                 imo->imo_multicast_vif = -1;
1252
1253                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1254                 if ((error = in_joingroup(ifp, &addr, NULL,
1255                     &imo->imo_membership[0])) != 0) {
1256                         free(imo->imo_membership, M_CARP);
1257                         break;
1258                 }
1259                 imo->imo_num_memberships++;
1260                 imo->imo_multicast_ifp = ifp;
1261                 imo->imo_multicast_ttl = CARP_DFLTTL;
1262                 imo->imo_multicast_loop = 0;
1263                 break;
1264            }
1265 #endif
1266 #ifdef INET6
1267         case AF_INET6:
1268             {
1269                 struct ip6_moptions *im6o = &cif->cif_im6o;
1270                 struct in6_addr in6;
1271                 struct in6_multi *in6m;
1272
1273                 if (im6o->im6o_membership)
1274                         return (0);
1275
1276                 im6o->im6o_membership = (struct in6_multi **)malloc(
1277                     (sizeof(struct in6_multi *) * IPV6_MIN_MEMBERSHIPS), M_CARP,
1278                     M_ZERO | M_NOWAIT);
1279                 if (im6o->im6o_membership == NULL)
1280                         return (ENOMEM);
1281                 im6o->im6o_mfilters = NULL;
1282                 im6o->im6o_max_memberships = IPV6_MIN_MEMBERSHIPS;
1283                 im6o->im6o_multicast_hlim = CARP_DFLTTL;
1284                 im6o->im6o_multicast_ifp = ifp;
1285
1286                 /* Join IPv6 CARP multicast group. */
1287                 bzero(&in6, sizeof(in6));
1288                 in6.s6_addr16[0] = htons(0xff02);
1289                 in6.s6_addr8[15] = 0x12;
1290                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1291                         free(im6o->im6o_membership, M_CARP);
1292                         break;
1293                 }
1294                 in6m = NULL;
1295                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1296                         free(im6o->im6o_membership, M_CARP);
1297                         break;
1298                 }
1299                 im6o->im6o_membership[0] = in6m;
1300                 im6o->im6o_num_memberships++;
1301
1302                 /* Join solicited multicast address. */
1303                 bzero(&in6, sizeof(in6));
1304                 in6.s6_addr16[0] = htons(0xff02);
1305                 in6.s6_addr32[1] = 0;
1306                 in6.s6_addr32[2] = htonl(1);
1307                 in6.s6_addr32[3] = 0;
1308                 in6.s6_addr8[12] = 0xff;
1309                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1310                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1311                         free(im6o->im6o_membership, M_CARP);
1312                         break;
1313                 }
1314                 in6m = NULL;
1315                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1316                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1317                         free(im6o->im6o_membership, M_CARP);
1318                         break;
1319                 }
1320                 im6o->im6o_membership[1] = in6m;
1321                 im6o->im6o_num_memberships++;
1322                 break;
1323             }
1324 #endif
1325         }
1326
1327         return (error);
1328 }
1329
1330 /*
1331  * Free multicast structures.
1332  */
1333 static void
1334 carp_multicast_cleanup(struct carp_if *cif, sa_family_t sa)
1335 {
1336
1337         CIF_LOCK_ASSERT(cif);
1338         switch (sa) {
1339 #ifdef INET
1340         case AF_INET:
1341                 if (cif->cif_naddrs == 0) {
1342                         struct ip_moptions *imo = &cif->cif_imo;
1343
1344                         in_leavegroup(imo->imo_membership[0], NULL);
1345                         KASSERT(imo->imo_mfilters == NULL,
1346                             ("%s: imo_mfilters != NULL", __func__));
1347                         free(imo->imo_membership, M_CARP);
1348                         imo->imo_membership = NULL;
1349
1350                 }
1351                 break;
1352 #endif
1353 #ifdef INET6
1354         case AF_INET6:
1355                 if (cif->cif_naddrs6 == 0) {
1356                         struct ip6_moptions *im6o = &cif->cif_im6o;
1357
1358                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1359                         in6_mc_leave(im6o->im6o_membership[1], NULL);
1360                         KASSERT(im6o->im6o_mfilters == NULL,
1361                             ("%s: im6o_mfilters != NULL", __func__));
1362                         free(im6o->im6o_membership, M_CARP);
1363                         im6o->im6o_membership = NULL;
1364                 }
1365                 break;
1366 #endif
1367         }
1368 }
1369
1370 int
1371 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *sa)
1372 {
1373         struct m_tag *mtag;
1374         struct carp_softc *sc;
1375
1376         if (!sa)
1377                 return (0);
1378
1379         switch (sa->sa_family) {
1380 #ifdef INET
1381         case AF_INET:
1382                 break;
1383 #endif
1384 #ifdef INET6
1385         case AF_INET6:
1386                 break;
1387 #endif
1388         default:
1389                 return (0);
1390         }
1391
1392         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
1393         if (mtag == NULL)
1394                 return (0);
1395
1396         bcopy(mtag + 1, &sc, sizeof(sc));
1397
1398         /* Set the source MAC address to the Virtual Router MAC Address. */
1399         switch (ifp->if_type) {
1400         case IFT_ETHER:
1401         case IFT_BRIDGE:
1402         case IFT_L2VLAN: {
1403                         struct ether_header *eh;
1404
1405                         eh = mtod(m, struct ether_header *);
1406                         eh->ether_shost[0] = 0;
1407                         eh->ether_shost[1] = 0;
1408                         eh->ether_shost[2] = 0x5e;
1409                         eh->ether_shost[3] = 0;
1410                         eh->ether_shost[4] = 1;
1411                         eh->ether_shost[5] = sc->sc_vhid;
1412                 }
1413                 break;
1414         case IFT_FDDI: {
1415                         struct fddi_header *fh;
1416
1417                         fh = mtod(m, struct fddi_header *);
1418                         fh->fddi_shost[0] = 0;
1419                         fh->fddi_shost[1] = 0;
1420                         fh->fddi_shost[2] = 0x5e;
1421                         fh->fddi_shost[3] = 0;
1422                         fh->fddi_shost[4] = 1;
1423                         fh->fddi_shost[5] = sc->sc_vhid;
1424                 }
1425                 break;
1426         case IFT_ISO88025: {
1427                         struct iso88025_header *th;
1428                         th = mtod(m, struct iso88025_header *);
1429                         th->iso88025_shost[0] = 3;
1430                         th->iso88025_shost[1] = 0;
1431                         th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
1432                         th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
1433                         th->iso88025_shost[4] = 0;
1434                         th->iso88025_shost[5] = 0;
1435                 }
1436                 break;
1437         default:
1438                 printf("%s: carp is not supported for the %d interface type\n",
1439                     ifp->if_xname, ifp->if_type);
1440                 return (EOPNOTSUPP);
1441         }
1442
1443         return (0);
1444 }
1445
1446 static struct carp_softc*
1447 carp_alloc(struct ifnet *ifp)
1448 {
1449         struct carp_softc *sc;
1450         struct carp_if *cif;
1451
1452         if ((cif = ifp->if_carp) == NULL) {
1453                 cif = carp_alloc_if(ifp);
1454                 if (cif == NULL)
1455                         return (NULL);
1456         }
1457
1458         sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
1459
1460         sc->sc_advbase = CARP_DFLTINTV;
1461         sc->sc_vhid = -1;       /* required setting */
1462         sc->sc_init_counter = 1;
1463         sc->sc_state = INIT;
1464
1465         sc->sc_ifasiz = sizeof(struct ifaddr *);
1466         sc->sc_ifas = malloc(sc->sc_ifasiz, M_CARP, M_WAITOK|M_ZERO);
1467         sc->sc_carpdev = ifp;
1468
1469         CARP_LOCK_INIT(sc);
1470 #ifdef INET
1471         callout_init_mtx(&sc->sc_md_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1472 #endif
1473 #ifdef INET6
1474         callout_init_mtx(&sc->sc_md6_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1475 #endif
1476         callout_init_mtx(&sc->sc_ad_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1477
1478         CIF_LOCK(cif);
1479         TAILQ_INSERT_TAIL(&cif->cif_vrs, sc, sc_list);
1480         CIF_UNLOCK(cif);
1481
1482         mtx_lock(&carp_mtx);
1483         LIST_INSERT_HEAD(&carp_list, sc, sc_next);
1484         mtx_unlock(&carp_mtx);
1485
1486         return (sc);
1487 }
1488
1489 static int
1490 carp_grow_ifas(struct carp_softc *sc)
1491 {
1492         struct ifaddr **new;
1493
1494         CARP_LOCK_ASSERT(sc);
1495
1496         new = malloc(sc->sc_ifasiz * 2, M_CARP, M_NOWAIT|M_ZERO);
1497         if (new == NULL)
1498                 return (ENOMEM);
1499         bcopy(sc->sc_ifas, new, sc->sc_ifasiz);
1500         free(sc->sc_ifas, M_CARP);
1501         sc->sc_ifas = new;
1502         sc->sc_ifasiz *= 2;
1503
1504         return (0);
1505 }
1506
1507 static void
1508 carp_destroy(struct carp_softc *sc)
1509 {
1510         struct ifnet *ifp = sc->sc_carpdev;
1511         struct carp_if *cif = ifp->if_carp;
1512
1513         CIF_LOCK_ASSERT(cif);
1514
1515         TAILQ_REMOVE(&cif->cif_vrs, sc, sc_list);
1516
1517         mtx_lock(&carp_mtx);
1518         LIST_REMOVE(sc, sc_next);
1519         mtx_unlock(&carp_mtx);
1520
1521         CARP_LOCK(sc);
1522         if (sc->sc_suppress)
1523                 carp_demote_adj(-carp_ifdown_adj, "vhid removed");
1524         callout_drain(&sc->sc_ad_tmo);
1525 #ifdef INET
1526         callout_drain(&sc->sc_md_tmo);
1527 #endif
1528 #ifdef INET6
1529         callout_drain(&sc->sc_md6_tmo);
1530 #endif
1531         CARP_LOCK_DESTROY(sc);
1532
1533         free(sc->sc_ifas, M_CARP);
1534         free(sc, M_CARP);
1535 }
1536
1537 static struct carp_if*
1538 carp_alloc_if(struct ifnet *ifp)
1539 {
1540         struct carp_if *cif;
1541
1542         cif = malloc(sizeof(*cif), M_CARP, M_WAITOK|M_ZERO);
1543
1544         if (ifpromisc(ifp, 1) != 0)
1545                 goto cleanup;
1546
1547         CIF_LOCK_INIT(cif);
1548         cif->cif_ifp = ifp;
1549         TAILQ_INIT(&cif->cif_vrs);
1550
1551         IF_ADDR_WLOCK(ifp);
1552         ifp->if_carp = cif;
1553         if_ref(ifp);
1554         IF_ADDR_WUNLOCK(ifp);
1555
1556         return (cif);
1557
1558 cleanup:
1559         free(cif, M_CARP);
1560
1561         return (NULL);
1562 }
1563
1564 static void
1565 carp_free_if(struct carp_if *cif)
1566 {
1567         struct ifnet *ifp = cif->cif_ifp;
1568
1569         CIF_LOCK_ASSERT(cif);
1570         KASSERT(TAILQ_EMPTY(&cif->cif_vrs), ("%s: softc list not empty",
1571             __func__));
1572
1573         IF_ADDR_WLOCK(ifp);
1574         ifp->if_carp = NULL;
1575         if_rele(ifp);
1576         IF_ADDR_WUNLOCK(ifp);
1577
1578         CIF_LOCK_DESTROY(cif);
1579
1580         ifpromisc(ifp, 0);
1581
1582         free(cif, M_CARP);
1583 }
1584
1585 static void
1586 carp_carprcp(struct carpreq *carpr, struct carp_softc *sc, int priv)
1587 {
1588
1589         CARP_LOCK(sc);
1590         carpr->carpr_state = sc->sc_state;
1591         carpr->carpr_vhid = sc->sc_vhid;
1592         carpr->carpr_advbase = sc->sc_advbase;
1593         carpr->carpr_advskew = sc->sc_advskew;
1594         if (priv)
1595                 bcopy(sc->sc_key, carpr->carpr_key, sizeof(carpr->carpr_key));
1596         else
1597                 bzero(carpr->carpr_key, sizeof(carpr->carpr_key));
1598         CARP_UNLOCK(sc);
1599 }
1600
1601 int
1602 carp_ioctl(struct ifreq *ifr, u_long cmd, struct thread *td)
1603 {
1604         struct carpreq carpr;
1605         struct ifnet *ifp;
1606         struct carp_softc *sc = NULL;
1607         int error = 0, locked = 0;
1608
1609         if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1610                 return (error);
1611
1612         ifp = ifunit_ref(ifr->ifr_name);
1613         if (ifp == NULL)
1614                 return (ENXIO);
1615
1616         switch (ifp->if_type) {
1617         case IFT_ETHER:
1618         case IFT_L2VLAN:
1619         case IFT_BRIDGE:
1620         case IFT_FDDI:
1621         case IFT_ISO88025:
1622                 break;
1623         default:
1624                 error = EOPNOTSUPP;
1625                 goto out;
1626         }
1627
1628         if ((ifp->if_flags & IFF_MULTICAST) == 0) {
1629                 error = EADDRNOTAVAIL;
1630                 goto out;
1631         }
1632
1633         switch (cmd) {
1634         case SIOCSVH:
1635                 if ((error = priv_check(td, PRIV_NETINET_CARP)))
1636                         break;
1637                 if (carpr.carpr_vhid <= 0 || carpr.carpr_vhid > CARP_MAXVHID ||
1638                     carpr.carpr_advbase < 0 || carpr.carpr_advskew < 0) {
1639                         error = EINVAL;
1640                         break;
1641                 }
1642
1643                 if (ifp->if_carp) {
1644                         CIF_LOCK(ifp->if_carp);
1645                         IFNET_FOREACH_CARP(ifp, sc)
1646                                 if (sc->sc_vhid == carpr.carpr_vhid)
1647                                         break;
1648                         CIF_UNLOCK(ifp->if_carp);
1649                 }
1650                 if (sc == NULL) {
1651                         sc = carp_alloc(ifp);
1652                         if (sc == NULL) {
1653                                 error = EINVAL; /* XXX: ifpromisc failed */
1654                                 break;
1655                         }
1656
1657                         CARP_LOCK(sc);
1658                         sc->sc_vhid = carpr.carpr_vhid;
1659                         LLADDR(&sc->sc_addr)[0] = 0;
1660                         LLADDR(&sc->sc_addr)[1] = 0;
1661                         LLADDR(&sc->sc_addr)[2] = 0x5e;
1662                         LLADDR(&sc->sc_addr)[3] = 0;
1663                         LLADDR(&sc->sc_addr)[4] = 1;
1664                         LLADDR(&sc->sc_addr)[5] = sc->sc_vhid;
1665                 } else
1666                         CARP_LOCK(sc);
1667                 locked = 1;
1668                 if (carpr.carpr_advbase > 0) {
1669                         if (carpr.carpr_advbase > 255 ||
1670                             carpr.carpr_advbase < CARP_DFLTINTV) {
1671                                 error = EINVAL;
1672                                 break;
1673                         }
1674                         sc->sc_advbase = carpr.carpr_advbase;
1675                 }
1676                 if (carpr.carpr_advskew > 0) {
1677                         if (carpr.carpr_advskew >= 255) {
1678                                 error = EINVAL;
1679                                 break;
1680                         }
1681                         sc->sc_advskew = carpr.carpr_advskew;
1682                 }
1683                 if (carpr.carpr_key[0] != '\0') {
1684                         bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1685                         carp_hmac_prepare(sc);
1686                 }
1687                 if (sc->sc_state != INIT &&
1688                     carpr.carpr_state != sc->sc_state) {
1689                         switch (carpr.carpr_state) {
1690                         case BACKUP:
1691                                 callout_stop(&sc->sc_ad_tmo);
1692                                 carp_set_state(sc, BACKUP);
1693                                 carp_setrun(sc, 0);
1694                                 carp_delroute(sc);
1695                                 break;
1696                         case MASTER:
1697                                 carp_master_down_locked(sc);
1698                                 break;
1699                         default:
1700                                 break;
1701                         }
1702                 }
1703                 break;
1704
1705         case SIOCGVH:
1706             {
1707                 int priveleged;
1708
1709                 if (carpr.carpr_vhid < 0 || carpr.carpr_vhid > CARP_MAXVHID) {
1710                         error = EINVAL;
1711                         break;
1712                 }
1713                 if (carpr.carpr_count < 1) {
1714                         error = EMSGSIZE;
1715                         break;
1716                 }
1717                 if (ifp->if_carp == NULL) {
1718                         error = ENOENT;
1719                         break;
1720                 }
1721
1722                 priveleged = (priv_check(td, PRIV_NETINET_CARP) == 0);
1723                 if (carpr.carpr_vhid != 0) {
1724                         CIF_LOCK(ifp->if_carp);
1725                         IFNET_FOREACH_CARP(ifp, sc)
1726                                 if (sc->sc_vhid == carpr.carpr_vhid)
1727                                         break;
1728                         CIF_UNLOCK(ifp->if_carp);
1729                         if (sc == NULL) {
1730                                 error = ENOENT;
1731                                 break;
1732                         }
1733                         carp_carprcp(&carpr, sc, priveleged);
1734                         error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1735                 } else  {
1736                         int i, count;
1737
1738                         count = 0;
1739                         CIF_LOCK(ifp->if_carp);
1740                         IFNET_FOREACH_CARP(ifp, sc)
1741                                 count++;
1742
1743                         if (count > carpr.carpr_count) {
1744                                 CIF_UNLOCK(ifp->if_carp);
1745                                 error = EMSGSIZE;
1746                                 break;
1747                         }
1748
1749                         i = 0;
1750                         IFNET_FOREACH_CARP(ifp, sc) {
1751                                 carp_carprcp(&carpr, sc, priveleged);
1752                                 carpr.carpr_count = count;
1753                                 error = copyout(&carpr, ifr->ifr_data +
1754                                     (i * sizeof(carpr)), sizeof(carpr));
1755                                 if (error) {
1756                                         CIF_UNLOCK(ifp->if_carp);
1757                                         break;
1758                                 }
1759                                 i++;
1760                         }
1761                         CIF_UNLOCK(ifp->if_carp);
1762                 }
1763                 break;
1764             }
1765         default:
1766                 error = EINVAL;
1767         }
1768
1769 out:
1770         if (locked)
1771                 CARP_UNLOCK(sc);
1772         if_rele(ifp);
1773
1774         return (error);
1775 }
1776
1777 static int
1778 carp_get_vhid(struct ifaddr *ifa)
1779 {
1780
1781         if (ifa == NULL || ifa->ifa_carp == NULL)
1782                 return (0);
1783
1784         return (ifa->ifa_carp->sc_vhid);
1785 }
1786
1787 int
1788 carp_attach(struct ifaddr *ifa, int vhid)
1789 {
1790         struct ifnet *ifp = ifa->ifa_ifp;
1791         struct carp_if *cif = ifp->if_carp;
1792         struct carp_softc *sc;
1793         int index, error;
1794
1795         if (ifp->if_carp == NULL)
1796                 return (ENOPROTOOPT);
1797
1798         switch (ifa->ifa_addr->sa_family) {
1799 #ifdef INET
1800         case AF_INET:
1801 #endif
1802 #ifdef INET6
1803         case AF_INET6:
1804 #endif
1805                 break;
1806         default:
1807                 return (EPROTOTYPE);
1808         }
1809
1810         CIF_LOCK(cif);
1811         IFNET_FOREACH_CARP(ifp, sc)
1812                 if (sc->sc_vhid == vhid)
1813                         break;
1814         if (sc == NULL) {
1815                 CIF_UNLOCK(cif);
1816                 return (ENOENT);
1817         }
1818
1819         if (ifa->ifa_carp) {
1820                 if (ifa->ifa_carp->sc_vhid != vhid)
1821                         carp_detach_locked(ifa);
1822                 else {
1823                         CIF_UNLOCK(cif);
1824                         return (0);
1825                 }
1826         }
1827
1828         error = carp_multicast_setup(cif, ifa->ifa_addr->sa_family);
1829         if (error) {
1830                 CIF_FREE(cif);
1831                 return (error);
1832         }
1833
1834         CARP_LOCK(sc);
1835         index = sc->sc_naddrs + sc->sc_naddrs6 + 1;
1836         if (index > sc->sc_ifasiz / sizeof(struct ifaddr *))
1837                 if ((error = carp_grow_ifas(sc)) != 0) {
1838                         carp_multicast_cleanup(cif,
1839                             ifa->ifa_addr->sa_family);
1840                         CARP_UNLOCK(sc);
1841                         CIF_FREE(cif);
1842                         return (error);
1843                 }
1844
1845         switch (ifa->ifa_addr->sa_family) {
1846 #ifdef INET
1847         case AF_INET:
1848                 cif->cif_naddrs++;
1849                 sc->sc_naddrs++;
1850                 break;
1851 #endif
1852 #ifdef INET6
1853         case AF_INET6:
1854                 cif->cif_naddrs6++;
1855                 sc->sc_naddrs6++;
1856                 break;
1857 #endif
1858         }
1859
1860         ifa_ref(ifa);
1861         sc->sc_ifas[index - 1] = ifa;
1862         ifa->ifa_carp = sc;
1863
1864         carp_hmac_prepare(sc);
1865         carp_sc_state(sc);
1866
1867         CARP_UNLOCK(sc);
1868         CIF_UNLOCK(cif);
1869
1870         return (0);
1871 }
1872
1873 void
1874 carp_detach(struct ifaddr *ifa)
1875 {
1876         struct ifnet *ifp = ifa->ifa_ifp;
1877         struct carp_if *cif = ifp->if_carp;
1878
1879         CIF_LOCK(cif);
1880         carp_detach_locked(ifa);
1881         CIF_FREE(cif);
1882 }
1883
1884 static void
1885 carp_detach_locked(struct ifaddr *ifa)
1886 {
1887         struct ifnet *ifp = ifa->ifa_ifp;
1888         struct carp_if *cif = ifp->if_carp;
1889         struct carp_softc *sc = ifa->ifa_carp;
1890         int i, index;
1891
1892         KASSERT(sc != NULL, ("%s: %p not attached", __func__, ifa));
1893
1894         CIF_LOCK_ASSERT(cif);
1895         CARP_LOCK(sc);
1896
1897         /* Shift array. */
1898         index = sc->sc_naddrs + sc->sc_naddrs6;
1899         for (i = 0; i < index; i++)
1900                 if (sc->sc_ifas[i] == ifa)
1901                         break;
1902         KASSERT(i < index, ("%s: %p no backref", __func__, ifa));
1903         for (; i < index - 1; i++)
1904                 sc->sc_ifas[i] = sc->sc_ifas[i+1];
1905         sc->sc_ifas[index - 1] = NULL;
1906
1907         switch (ifa->ifa_addr->sa_family) {
1908 #ifdef INET
1909         case AF_INET:
1910                 cif->cif_naddrs--;
1911                 sc->sc_naddrs--;
1912                 break;
1913 #endif
1914 #ifdef INET6
1915         case AF_INET6:
1916                 cif->cif_naddrs6--;
1917                 sc->sc_naddrs6--;
1918                 break;
1919 #endif
1920         }
1921
1922         carp_ifa_delroute(ifa);
1923         carp_multicast_cleanup(cif, ifa->ifa_addr->sa_family);
1924
1925         ifa->ifa_carp = NULL;
1926         ifa_free(ifa);
1927
1928         carp_hmac_prepare(sc);
1929         carp_sc_state(sc);
1930
1931         if (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) {
1932                 CARP_UNLOCK(sc);
1933                 carp_destroy(sc);
1934         } else
1935                 CARP_UNLOCK(sc);
1936 }
1937
1938 static void
1939 carp_set_state(struct carp_softc *sc, int state)
1940 {
1941
1942         CARP_LOCK_ASSERT(sc);
1943
1944         if (sc->sc_state != state) {
1945                 const char *carp_states[] = { CARP_STATES };
1946                 char subsys[IFNAMSIZ+5];
1947
1948                 sc->sc_state = state;
1949
1950                 snprintf(subsys, IFNAMSIZ+5, "%u@%s", sc->sc_vhid,
1951                     sc->sc_carpdev->if_xname);
1952                 devctl_notify("CARP", subsys, carp_states[state], NULL);
1953         }
1954 }
1955
1956 static void
1957 carp_linkstate(struct ifnet *ifp)
1958 {
1959         struct carp_softc *sc;
1960
1961         CIF_LOCK(ifp->if_carp);
1962         IFNET_FOREACH_CARP(ifp, sc) {
1963                 CARP_LOCK(sc);
1964                 carp_sc_state(sc);
1965                 CARP_UNLOCK(sc);
1966         }
1967         CIF_UNLOCK(ifp->if_carp);
1968 }
1969
1970 static void
1971 carp_sc_state(struct carp_softc *sc)
1972 {
1973
1974         CARP_LOCK_ASSERT(sc);
1975
1976         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1977             !(sc->sc_carpdev->if_flags & IFF_UP)) {
1978                 callout_stop(&sc->sc_ad_tmo);
1979 #ifdef INET
1980                 callout_stop(&sc->sc_md_tmo);
1981 #endif
1982 #ifdef INET6
1983                 callout_stop(&sc->sc_md6_tmo);
1984 #endif
1985                 carp_set_state(sc, INIT);
1986                 carp_setrun(sc, 0);
1987                 if (!sc->sc_suppress)
1988                         carp_demote_adj(carp_ifdown_adj, "interface down");
1989                 sc->sc_suppress = 1;
1990         } else {
1991                 carp_set_state(sc, INIT);
1992                 carp_setrun(sc, 0);
1993                 if (sc->sc_suppress)
1994                         carp_demote_adj(-carp_ifdown_adj, "interface up");
1995                 sc->sc_suppress = 0;
1996         }
1997 }
1998
1999 static void
2000 carp_demote_adj(int adj, char *reason)
2001 {
2002         carp_demotion += adj;
2003         CARP_LOG("demoted by %d to %d (%s)\n", adj, carp_demotion, reason);
2004         taskqueue_enqueue(taskqueue_swi, &carp_sendall_task);
2005 }
2006
2007 #ifdef INET
2008 extern  struct domain inetdomain;
2009 static struct protosw in_carp_protosw = {
2010         .pr_type =              SOCK_RAW,
2011         .pr_domain =            &inetdomain,
2012         .pr_protocol =          IPPROTO_CARP,
2013         .pr_flags =             PR_ATOMIC|PR_ADDR,
2014         .pr_input =             carp_input,
2015         .pr_output =            (pr_output_t *)rip_output,
2016         .pr_ctloutput =         rip_ctloutput,
2017         .pr_usrreqs =           &rip_usrreqs
2018 };
2019 #endif
2020
2021 #ifdef INET6
2022 extern  struct domain inet6domain;
2023 static struct ip6protosw in6_carp_protosw = {
2024         .pr_type =              SOCK_RAW,
2025         .pr_domain =            &inet6domain,
2026         .pr_protocol =          IPPROTO_CARP,
2027         .pr_flags =             PR_ATOMIC|PR_ADDR,
2028         .pr_input =             carp6_input,
2029         .pr_output =            rip6_output,
2030         .pr_ctloutput =         rip6_ctloutput,
2031         .pr_usrreqs =           &rip6_usrreqs
2032 };
2033 #endif
2034
2035 static void
2036 carp_mod_cleanup(void)
2037 {
2038
2039 #ifdef INET
2040         if (proto_reg[CARP_INET] == 0) {
2041                 (void)ipproto_unregister(IPPROTO_CARP);
2042                 pf_proto_unregister(PF_INET, IPPROTO_CARP, SOCK_RAW);
2043                 proto_reg[CARP_INET] = -1;
2044         }
2045         carp_iamatch_p = NULL;
2046 #endif
2047 #ifdef INET6
2048         if (proto_reg[CARP_INET6] == 0) {
2049                 (void)ip6proto_unregister(IPPROTO_CARP);
2050                 pf_proto_unregister(PF_INET6, IPPROTO_CARP, SOCK_RAW);
2051                 proto_reg[CARP_INET6] = -1;
2052         }
2053         carp_iamatch6_p = NULL;
2054         carp_macmatch6_p = NULL;
2055 #endif
2056         carp_ioctl_p = NULL;
2057         carp_attach_p = NULL;
2058         carp_detach_p = NULL;
2059         carp_get_vhid_p = NULL;
2060         carp_linkstate_p = NULL;
2061         carp_forus_p = NULL;
2062         carp_output_p = NULL;
2063         carp_demote_adj_p = NULL;
2064         mtx_unlock(&carp_mtx);
2065         taskqueue_drain(taskqueue_swi, &carp_sendall_task);
2066         mtx_destroy(&carp_mtx);
2067 }
2068
2069 static int
2070 carp_mod_load(void)
2071 {
2072         int err;
2073
2074         mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2075         LIST_INIT(&carp_list);
2076         carp_get_vhid_p = carp_get_vhid;
2077         carp_forus_p = carp_forus;
2078         carp_output_p = carp_output;
2079         carp_linkstate_p = carp_linkstate;
2080         carp_ioctl_p = carp_ioctl;
2081         carp_attach_p = carp_attach;
2082         carp_detach_p = carp_detach;
2083         carp_demote_adj_p = carp_demote_adj;
2084 #ifdef INET6
2085         carp_iamatch6_p = carp_iamatch6;
2086         carp_macmatch6_p = carp_macmatch6;
2087         proto_reg[CARP_INET6] = pf_proto_register(PF_INET6,
2088             (struct protosw *)&in6_carp_protosw);
2089         if (proto_reg[CARP_INET6]) {
2090                 printf("carp: error %d attaching to PF_INET6\n",
2091                     proto_reg[CARP_INET6]);
2092                 carp_mod_cleanup();
2093                 return (proto_reg[CARP_INET6]);
2094         }
2095         err = ip6proto_register(IPPROTO_CARP);
2096         if (err) {
2097                 printf("carp: error %d registering with INET6\n", err);
2098                 carp_mod_cleanup();
2099                 return (err);
2100         }
2101 #endif
2102 #ifdef INET
2103         carp_iamatch_p = carp_iamatch;
2104         proto_reg[CARP_INET] = pf_proto_register(PF_INET, &in_carp_protosw);
2105         if (proto_reg[CARP_INET]) {
2106                 printf("carp: error %d attaching to PF_INET\n",
2107                     proto_reg[CARP_INET]);
2108                 carp_mod_cleanup();
2109                 return (proto_reg[CARP_INET]);
2110         }
2111         err = ipproto_register(IPPROTO_CARP);
2112         if (err) {
2113                 printf("carp: error %d registering with INET\n", err);
2114                 carp_mod_cleanup();
2115                 return (err);
2116         }
2117 #endif
2118         return (0);
2119 }
2120
2121 static int
2122 carp_modevent(module_t mod, int type, void *data)
2123 {
2124         switch (type) {
2125         case MOD_LOAD:
2126                 return carp_mod_load();
2127                 /* NOTREACHED */
2128         case MOD_UNLOAD:
2129                 mtx_lock(&carp_mtx);
2130                 if (LIST_EMPTY(&carp_list))
2131                         carp_mod_cleanup();
2132                 else {
2133                         mtx_unlock(&carp_mtx);
2134                         return (EBUSY);
2135                 }
2136                 break;
2137
2138         default:
2139                 return (EINVAL);
2140         }
2141
2142         return (0);
2143 }
2144
2145 static moduledata_t carp_mod = {
2146         "carp",
2147         carp_modevent,
2148         0
2149 };
2150
2151 DECLARE_MODULE(carp, carp_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);