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 #include <sys/counter.h>
  54
  55 #include <net/ethernet.h>
  56 #include <net/fddi.h>
  57 #include <net/if.h>
  58 #include <net/if_var.h>
  59 #include <net/if_dl.h>
  60 #include <net/if_llatbl.h>
  61 #include <net/if_types.h>
  62 #include <net/iso88025.h>
  63 #include <net/route.h>
  64 #include <net/vnet.h>
  65
  66 #if defined(INET) || defined(INET6)
  67 #include <netinet/in.h>
  68 #include <netinet/in_var.h>
  69 #include <netinet/ip_carp.h>
  70 #include <netinet/ip.h>
  71 #include <machine/in_cksum.h>
  72 #endif
  73 #ifdef INET
  74 #include <netinet/ip_var.h>
  75 #include <netinet/if_ether.h>
  76 #endif
  77
  78 #ifdef INET6
  79 #include <netinet/icmp6.h>
  80 #include <netinet/ip6.h>
  81 #include <netinet6/in6_var.h>
  82 #include <netinet6/ip6_var.h>
  83 #include <netinet6/scope6_var.h>
  84 #include <netinet6/nd6.h>
  85 #endif
  86
  87 #include <crypto/sha1.h>
  88
  89 static MALLOC_DEFINE(M_CARP, "CARP", "CARP addresses");
  90
  91 struct carp_softc {
  92         struct ifnet            *sc_carpdev;    /* Pointer to parent ifnet. */
  93         struct ifaddr           **sc_ifas;      /* Our ifaddrs. */
  94         struct sockaddr_dl      sc_addr;        /* Our link level address. */
  95         struct callout          sc_ad_tmo;      /* Advertising timeout. */
  96 #ifdef INET
  97         struct callout          sc_md_tmo;      /* Master down timeout. */
  98 #endif
  99 #ifdef INET6
 100         struct callout          sc_md6_tmo;     /* XXX: Master down timeout. */
 101 #endif
 102         struct mtx              sc_mtx;
 103
 104         int                     sc_vhid;
 105         int                     sc_advskew;
 106         int                     sc_advbase;
 107
 108         int                     sc_naddrs;
 109         int                     sc_naddrs6;
 110         int                     sc_ifasiz;
 111         enum { INIT = 0, BACKUP, MASTER }       sc_state;
 112         int                     sc_suppress;
 113         int                     sc_sendad_errors;
 114 #define CARP_SENDAD_MAX_ERRORS  3
 115         int                     sc_sendad_success;
 116 #define CARP_SENDAD_MIN_SUCCESS 3
 117
 118         int                     sc_init_counter;
 119         uint64_t                sc_counter;
 120
 121         /* authentication */
 122 #define CARP_HMAC_PAD   64
 123         unsigned char sc_key[CARP_KEY_LEN];
 124         unsigned char sc_pad[CARP_HMAC_PAD];
 125         SHA1_CTX sc_sha1;
 126
 127         TAILQ_ENTRY(carp_softc) sc_list;        /* On the carp_if list. */
 128         LIST_ENTRY(carp_softc)  sc_next;        /* On the global list. */
 129 };
 130
 131 struct carp_if {
 132 #ifdef INET
 133         int     cif_naddrs;
 134 #endif
 135 #ifdef INET6
 136         int     cif_naddrs6;
 137 #endif
 138         TAILQ_HEAD(, carp_softc) cif_vrs;
 139 #ifdef INET
 140         struct ip_moptions       cif_imo;
 141 #endif
 142 #ifdef INET6
 143         struct ip6_moptions      cif_im6o;
 144 #endif
 145         struct ifnet    *cif_ifp;
 146         struct mtx      cif_mtx;
 147         uint32_t        cif_flags;
 148 #define CIF_PROMISC     0x00000001
 149 };
 150
 151 #define CARP_INET       0
 152 #define CARP_INET6      1
 153 static int proto_reg[] = {-1, -1};
 154
 155 /*
 156  * Brief design of carp(4).
 157  *
 158  * Any carp-capable ifnet may have a list of carp softcs hanging off
 159  * its ifp->if_carp pointer. Each softc represents one unique virtual
 160  * host id, or vhid. The softc has a back pointer to the ifnet. All
 161  * softcs are joined in a global list, which has quite limited use.
 162  *
 163  * Any interface address that takes part in CARP negotiation has a
 164  * pointer to the softc of its vhid, ifa->ifa_carp. That could be either
 165  * AF_INET or AF_INET6 address.
 166  *
 167  * Although, one can get the softc's backpointer to ifnet and traverse
 168  * through its ifp->if_addrhead queue to find all interface addresses
 169  * involved in CARP, we keep a growable array of ifaddr pointers. This
 170  * allows us to avoid grabbing the IF_ADDR_LOCK() in many traversals that
 171  * do calls into the network stack, thus avoiding LORs.
 172  *
 173  * Locking:
 174  *
 175  * Each softc has a lock sc_mtx. It is used to synchronise carp_input_c(),
 176  * callout-driven events and ioctl()s.
 177  *
 178  * To traverse the list of softcs on an ifnet we use CIF_LOCK(), to
 179  * traverse the global list we use the mutex carp_mtx.
 180  *
 181  * Known issues with locking:
 182  *
 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 /* Accept incoming CARP packets. */
 190 static VNET_DEFINE(int, carp_allow) = 1;
 191 #define V_carp_allow    VNET(carp_allow)
 192
 193 /* Preempt slower nodes. */
 194 static VNET_DEFINE(int, carp_preempt) = 0;
 195 #define V_carp_preempt  VNET(carp_preempt)
 196
 197 /* Log level. */
 198 static VNET_DEFINE(int, carp_log) = 1;
 199 #define V_carp_log      VNET(carp_log)
 200
 201 /* Global advskew demotion. */
 202 static VNET_DEFINE(int, carp_demotion) = 0;
 203 #define V_carp_demotion VNET(carp_demotion)
 204
 205 /* Send error demotion factor. */
 206 static VNET_DEFINE(int, carp_senderr_adj) = CARP_MAXSKEW;
 207 #define V_carp_senderr_adj      VNET(carp_senderr_adj)
 208
 209 /* Iface down demotion factor. */
 210 static VNET_DEFINE(int, carp_ifdown_adj) = CARP_MAXSKEW;
 211 #define V_carp_ifdown_adj       VNET(carp_ifdown_adj)
 212
 213 static int carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS);
 214
 215 SYSCTL_NODE(_net_inet, IPPROTO_CARP,    carp,   CTLFLAG_RW, 0,  "CARP");
 216 SYSCTL_INT(_net_inet_carp, OID_AUTO, allow, CTLFLAG_VNET | CTLFLAG_RW,
 217     &VNET_NAME(carp_allow), 0, "Accept incoming CARP packets");
 218 SYSCTL_INT(_net_inet_carp, OID_AUTO, preempt, CTLFLAG_VNET | CTLFLAG_RW,
 219     &VNET_NAME(carp_preempt), 0, "High-priority backup preemption mode");
 220 SYSCTL_INT(_net_inet_carp, OID_AUTO, log, CTLFLAG_VNET | CTLFLAG_RW,
 221     &VNET_NAME(carp_log), 0, "CARP log level");
 222 SYSCTL_PROC(_net_inet_carp, OID_AUTO, demotion,
 223     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW,
 224     0, 0, carp_demote_adj_sysctl, "I",
 225     "Adjust demotion factor (skew of advskew)");
 226 SYSCTL_INT(_net_inet_carp, OID_AUTO, senderr_demotion_factor,
 227     CTLFLAG_VNET | CTLFLAG_RW,
 228     &VNET_NAME(carp_senderr_adj), 0, "Send error demotion factor adjustment");
 229 SYSCTL_INT(_net_inet_carp, OID_AUTO, ifdown_demotion_factor,
 230     CTLFLAG_VNET | CTLFLAG_RW,
 231     &VNET_NAME(carp_ifdown_adj), 0,
 232     "Interface down demotion factor adjustment");
 233
 234 VNET_PCPUSTAT_DEFINE(struct carpstats, carpstats);
 235 VNET_PCPUSTAT_SYSINIT(carpstats);
 236 VNET_PCPUSTAT_SYSUNINIT(carpstats);
 237
 238 #define CARPSTATS_ADD(name, val)        \
 239     counter_u64_add(VNET(carpstats)[offsetof(struct carpstats, name) / \
 240         sizeof(uint64_t)], (val))
 241 #define CARPSTATS_INC(name)             CARPSTATS_ADD(name, 1)
 242
 243 SYSCTL_VNET_PCPUSTAT(_net_inet_carp, OID_AUTO, stats, struct carpstats,
 244     carpstats, "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 245
 246 #define CARP_LOCK_INIT(sc)      mtx_init(&(sc)->sc_mtx, "carp_softc",   \
 247         NULL, MTX_DEF)
 248 #define CARP_LOCK_DESTROY(sc)   mtx_destroy(&(sc)->sc_mtx)
 249 #define CARP_LOCK_ASSERT(sc)    mtx_assert(&(sc)->sc_mtx, MA_OWNED)
 250 #define CARP_LOCK(sc)           mtx_lock(&(sc)->sc_mtx)
 251 #define CARP_UNLOCK(sc)         mtx_unlock(&(sc)->sc_mtx)
 252 #define CIF_LOCK_INIT(cif)      mtx_init(&(cif)->cif_mtx, "carp_if",   \
 253         NULL, MTX_DEF)
 254 #define CIF_LOCK_DESTROY(cif)   mtx_destroy(&(cif)->cif_mtx)
 255 #define CIF_LOCK_ASSERT(cif)    mtx_assert(&(cif)->cif_mtx, MA_OWNED)
 256 #define CIF_LOCK(cif)           mtx_lock(&(cif)->cif_mtx)
 257 #define CIF_UNLOCK(cif)         mtx_unlock(&(cif)->cif_mtx)
 258 #define CIF_FREE(cif)   do {                            \
 259                 CIF_LOCK(cif);                          \
 260                 if (TAILQ_EMPTY(&(cif)->cif_vrs))       \
 261                         carp_free_if(cif);              \
 262                 else                                    \
 263                         CIF_UNLOCK(cif);                \
 264 } while (0)
 265
 266 #define CARP_LOG(...)   do {                            \
 267         if (V_carp_log > 0)                             \
 268                 log(LOG_INFO, "carp: " __VA_ARGS__);    \
 269 } while (0)
 270
 271 #define CARP_DEBUG(...) do {                            \
 272         if (V_carp_log > 1)                             \
 273                 log(LOG_DEBUG, __VA_ARGS__);            \
 274 } while (0)
 275
 276 #define IFNET_FOREACH_IFA(ifp, ifa)                                     \
 277         IF_ADDR_LOCK_ASSERT(ifp);                                       \
 278         TAILQ_FOREACH((ifa), &(ifp)->if_addrhead, ifa_link)             \
 279                 if ((ifa)->ifa_carp != NULL)
 280
 281 #define CARP_FOREACH_IFA(sc, ifa)                                       \
 282         CARP_LOCK_ASSERT(sc);                                           \
 283         for (int _i = 0;                                                \
 284                 _i < (sc)->sc_naddrs + (sc)->sc_naddrs6 &&              \
 285                 ((ifa) = sc->sc_ifas[_i]) != NULL;                      \
 286                 ++_i)
 287
 288 #define IFNET_FOREACH_CARP(ifp, sc)                                     \
 289         CIF_LOCK_ASSERT(ifp->if_carp);                                  \
 290         TAILQ_FOREACH((sc), &(ifp)->if_carp->cif_vrs, sc_list)
 291
 292 #define DEMOTE_ADVSKEW(sc)                                      \
 293     (((sc)->sc_advskew + V_carp_demotion > CARP_MAXSKEW) ?      \
 294     CARP_MAXSKEW : ((sc)->sc_advskew + V_carp_demotion))
 295
 296 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 297 static struct carp_softc
 298                 *carp_alloc(struct ifnet *);
 299 static void     carp_destroy(struct carp_softc *);
 300 static struct carp_if
 301                 *carp_alloc_if(struct ifnet *);
 302 static void     carp_free_if(struct carp_if *);
 303 static void     carp_set_state(struct carp_softc *, int, const char* reason);
 304 static void     carp_sc_state(struct carp_softc *);
 305 static void     carp_setrun(struct carp_softc *, sa_family_t);
 306 static void     carp_master_down(void *);
 307 static void     carp_master_down_locked(struct carp_softc *,
 308                     const char* reason);
 309 static void     carp_send_ad(void *);
 310 static void     carp_send_ad_locked(struct carp_softc *);
 311 static void     carp_addroute(struct carp_softc *);
 312 static void     carp_ifa_addroute(struct ifaddr *);
 313 static void     carp_delroute(struct carp_softc *);
 314 static void     carp_ifa_delroute(struct ifaddr *);
 315 static void     carp_send_ad_all(void *, int);
 316 static void     carp_demote_adj(int, char *);
 317
 318 static LIST_HEAD(, carp_softc) carp_list;
 319 static struct mtx carp_mtx;
 320 static struct sx carp_sx;
 321 static struct task carp_sendall_task =
 322     TASK_INITIALIZER(0, carp_send_ad_all, NULL);
 323
 324 static void
 325 carp_hmac_prepare(struct carp_softc *sc)
 326 {
 327         uint8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 328         uint8_t vhid = sc->sc_vhid & 0xff;
 329         struct ifaddr *ifa;
 330         int i, found;
 331 #ifdef INET
 332         struct in_addr last, cur, in;
 333 #endif
 334 #ifdef INET6
 335         struct in6_addr last6, cur6, in6;
 336 #endif
 337
 338         CARP_LOCK_ASSERT(sc);
 339
 340         /* Compute ipad from key. */
 341         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 342         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 343         for (i = 0; i < sizeof(sc->sc_pad); i++)
 344                 sc->sc_pad[i] ^= 0x36;
 345
 346         /* Precompute first part of inner hash. */
 347         SHA1Init(&sc->sc_sha1);
 348         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 349         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 350         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 351         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 352 #ifdef INET
 353         cur.s_addr = 0;
 354         do {
 355                 found = 0;
 356                 last = cur;
 357                 cur.s_addr = 0xffffffff;
 358                 CARP_FOREACH_IFA(sc, ifa) {
 359                         in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 360                         if (ifa->ifa_addr->sa_family == AF_INET &&
 361                             ntohl(in.s_addr) > ntohl(last.s_addr) &&
 362                             ntohl(in.s_addr) < ntohl(cur.s_addr)) {
 363                                 cur.s_addr = in.s_addr;
 364                                 found++;
 365                         }
 366                 }
 367                 if (found)
 368                         SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur));
 369         } while (found);
 370 #endif /* INET */
 371 #ifdef INET6
 372         memset(&cur6, 0, sizeof(cur6));
 373         do {
 374                 found = 0;
 375                 last6 = cur6;
 376                 memset(&cur6, 0xff, sizeof(cur6));
 377                 CARP_FOREACH_IFA(sc, ifa) {
 378                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 379                         if (IN6_IS_SCOPE_EMBED(&in6))
 380                                 in6.s6_addr16[1] = 0;
 381                         if (ifa->ifa_addr->sa_family == AF_INET6 &&
 382                             memcmp(&in6, &last6, sizeof(in6)) > 0 &&
 383                             memcmp(&in6, &cur6, sizeof(in6)) < 0) {
 384                                 cur6 = in6;
 385                                 found++;
 386                         }
 387                 }
 388                 if (found)
 389                         SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6));
 390         } while (found);
 391 #endif /* INET6 */
 392
 393         /* convert ipad to opad */
 394         for (i = 0; i < sizeof(sc->sc_pad); i++)
 395                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 396 }
 397
 398 static void
 399 carp_hmac_generate(struct carp_softc *sc, uint32_t counter[2],
 400     unsigned char md[20])
 401 {
 402         SHA1_CTX sha1ctx;
 403
 404         CARP_LOCK_ASSERT(sc);
 405
 406         /* fetch first half of inner hash */
 407         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 408
 409         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 410         SHA1Final(md, &sha1ctx);
 411
 412         /* outer hash */
 413         SHA1Init(&sha1ctx);
 414         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 415         SHA1Update(&sha1ctx, md, 20);
 416         SHA1Final(md, &sha1ctx);
 417 }
 418
 419 static int
 420 carp_hmac_verify(struct carp_softc *sc, uint32_t counter[2],
 421     unsigned char md[20])
 422 {
 423         unsigned char md2[20];
 424
 425         CARP_LOCK_ASSERT(sc);
 426
 427         carp_hmac_generate(sc, counter, md2);
 428
 429         return (bcmp(md, md2, sizeof(md2)));
 430 }
 431
 432 /*
 433  * process input packet.
 434  * we have rearranged checks order compared to the rfc,
 435  * but it seems more efficient this way or not possible otherwise.
 436  */
 437 #ifdef INET
 438 int
 439 carp_input(struct mbuf **mp, int *offp, int proto)
 440 {
 441         struct mbuf *m = *mp;
 442         struct ip *ip = mtod(m, struct ip *);
 443         struct carp_header *ch;
 444         int iplen, len;
 445
 446         iplen = *offp;
 447         *mp = NULL;
 448
 449         CARPSTATS_INC(carps_ipackets);
 450
 451         if (!V_carp_allow) {
 452                 m_freem(m);
 453                 return (IPPROTO_DONE);
 454         }
 455
 456         /* verify that the IP TTL is 255.  */
 457         if (ip->ip_ttl != CARP_DFLTTL) {
 458                 CARPSTATS_INC(carps_badttl);
 459                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 460                     ip->ip_ttl,
 461                     m->m_pkthdr.rcvif->if_xname);
 462                 m_freem(m);
 463                 return (IPPROTO_DONE);
 464         }
 465
 466         iplen = ip->ip_hl << 2;
 467
 468         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 469                 CARPSTATS_INC(carps_badlen);
 470                 CARP_DEBUG("%s: received len %zd < sizeof(struct carp_header) "
 471                     "on %s\n", __func__, m->m_len - sizeof(struct ip),
 472                     m->m_pkthdr.rcvif->if_xname);
 473                 m_freem(m);
 474                 return (IPPROTO_DONE);
 475         }
 476
 477         if (iplen + sizeof(*ch) < m->m_len) {
 478                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 479                         CARPSTATS_INC(carps_hdrops);
 480                         CARP_DEBUG("%s: pullup failed\n", __func__);
 481                         return (IPPROTO_DONE);
 482                 }
 483                 ip = mtod(m, struct ip *);
 484         }
 485         ch = (struct carp_header *)((char *)ip + iplen);
 486
 487         /*
 488          * verify that the received packet length is
 489          * equal to the CARP header
 490          */
 491         len = iplen + sizeof(*ch);
 492         if (len > m->m_pkthdr.len) {
 493                 CARPSTATS_INC(carps_badlen);
 494                 CARP_DEBUG("%s: packet too short %d on %s\n", __func__,
 495                     m->m_pkthdr.len,
 496                     m->m_pkthdr.rcvif->if_xname);
 497                 m_freem(m);
 498                 return (IPPROTO_DONE);
 499         }
 500
 501         if ((m = m_pullup(m, len)) == NULL) {
 502                 CARPSTATS_INC(carps_hdrops);
 503                 return (IPPROTO_DONE);
 504         }
 505         ip = mtod(m, struct ip *);
 506         ch = (struct carp_header *)((char *)ip + iplen);
 507
 508         /* verify the CARP checksum */
 509         m->m_data += iplen;
 510         if (in_cksum(m, len - iplen)) {
 511                 CARPSTATS_INC(carps_badsum);
 512                 CARP_DEBUG("%s: checksum failed on %s\n", __func__,
 513                     m->m_pkthdr.rcvif->if_xname);
 514                 m_freem(m);
 515                 return (IPPROTO_DONE);
 516         }
 517         m->m_data -= iplen;
 518
 519         carp_input_c(m, ch, AF_INET);
 520         return (IPPROTO_DONE);
 521 }
 522 #endif
 523
 524 #ifdef INET6
 525 int
 526 carp6_input(struct mbuf **mp, int *offp, int proto)
 527 {
 528         struct mbuf *m = *mp;
 529         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 530         struct carp_header *ch;
 531         u_int len;
 532
 533         CARPSTATS_INC(carps_ipackets6);
 534
 535         if (!V_carp_allow) {
 536                 m_freem(m);
 537                 return (IPPROTO_DONE);
 538         }
 539
 540         /* check if received on a valid carp interface */
 541         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 542                 CARPSTATS_INC(carps_badif);
 543                 CARP_DEBUG("%s: packet received on non-carp interface: %s\n",
 544                     __func__, m->m_pkthdr.rcvif->if_xname);
 545                 m_freem(m);
 546                 return (IPPROTO_DONE);
 547         }
 548
 549         /* verify that the IP TTL is 255 */
 550         if (ip6->ip6_hlim != CARP_DFLTTL) {
 551                 CARPSTATS_INC(carps_badttl);
 552                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 553                     ip6->ip6_hlim, m->m_pkthdr.rcvif->if_xname);
 554                 m_freem(m);
 555                 return (IPPROTO_DONE);
 556         }
 557
 558         /* verify that we have a complete carp packet */
 559         len = m->m_len;
 560         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 561         if (ch == NULL) {
 562                 CARPSTATS_INC(carps_badlen);
 563                 CARP_DEBUG("%s: packet size %u too small\n", __func__, len);
 564                 return (IPPROTO_DONE);
 565         }
 566
 567
 568         /* verify the CARP checksum */
 569         m->m_data += *offp;
 570         if (in_cksum(m, sizeof(*ch))) {
 571                 CARPSTATS_INC(carps_badsum);
 572                 CARP_DEBUG("%s: checksum failed, on %s\n", __func__,
 573                     m->m_pkthdr.rcvif->if_xname);
 574                 m_freem(m);
 575                 return (IPPROTO_DONE);
 576         }
 577         m->m_data -= *offp;
 578
 579         carp_input_c(m, ch, AF_INET6);
 580         return (IPPROTO_DONE);
 581 }
 582 #endif /* INET6 */
 583
 584 static void
 585 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 586 {
 587         struct ifnet *ifp = m->m_pkthdr.rcvif;
 588         struct ifaddr *ifa;
 589         struct carp_softc *sc;
 590         uint64_t tmp_counter;
 591         struct timeval sc_tv, ch_tv;
 592
 593         /* verify that the VHID is valid on the receiving interface */
 594         IF_ADDR_RLOCK(ifp);
 595         IFNET_FOREACH_IFA(ifp, ifa)
 596                 if (ifa->ifa_addr->sa_family == af &&
 597                     ifa->ifa_carp->sc_vhid == ch->carp_vhid) {
 598                         ifa_ref(ifa);
 599                         break;
 600                 }
 601         IF_ADDR_RUNLOCK(ifp);
 602
 603         if (ifa == NULL) {
 604                 CARPSTATS_INC(carps_badvhid);
 605                 m_freem(m);
 606                 return;
 607         }
 608
 609         /* verify the CARP version. */
 610         if (ch->carp_version != CARP_VERSION) {
 611                 CARPSTATS_INC(carps_badver);
 612                 CARP_DEBUG("%s: invalid version %d\n", ifp->if_xname,
 613                     ch->carp_version);
 614                 ifa_free(ifa);
 615                 m_freem(m);
 616                 return;
 617         }
 618
 619         sc = ifa->ifa_carp;
 620         CARP_LOCK(sc);
 621         ifa_free(ifa);
 622
 623         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 624                 CARPSTATS_INC(carps_badauth);
 625                 CARP_DEBUG("%s: incorrect hash for VHID %u@%s\n", __func__,
 626                     sc->sc_vhid, ifp->if_xname);
 627                 goto out;
 628         }
 629
 630         tmp_counter = ntohl(ch->carp_counter[0]);
 631         tmp_counter = tmp_counter<<32;
 632         tmp_counter += ntohl(ch->carp_counter[1]);
 633
 634         /* XXX Replay protection goes here */
 635
 636         sc->sc_init_counter = 0;
 637         sc->sc_counter = tmp_counter;
 638
 639         sc_tv.tv_sec = sc->sc_advbase;
 640         sc_tv.tv_usec = DEMOTE_ADVSKEW(sc) * 1000000 / 256;
 641         ch_tv.tv_sec = ch->carp_advbase;
 642         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 643
 644         switch (sc->sc_state) {
 645         case INIT:
 646                 break;
 647         case MASTER:
 648                 /*
 649                  * If we receive an advertisement from a master who's going to
 650                  * be more frequent than us, go into BACKUP state.
 651                  */
 652                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 653                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 654                         callout_stop(&sc->sc_ad_tmo);
 655                         carp_set_state(sc, BACKUP,
 656                             "more frequent advertisement received");
 657                         carp_setrun(sc, 0);
 658                         carp_delroute(sc);
 659                 }
 660                 break;
 661         case BACKUP:
 662                 /*
 663                  * If we're pre-empting masters who advertise slower than us,
 664                  * and this one claims to be slower, treat him as down.
 665                  */
 666                 if (V_carp_preempt && timevalcmp(&sc_tv, &ch_tv, <)) {
 667                         carp_master_down_locked(sc,
 668                             "preempting a slower master");
 669                         break;
 670                 }
 671
 672                 /*
 673                  *  If the master is going to advertise at such a low frequency
 674                  *  that he's guaranteed to time out, we'd might as well just
 675                  *  treat him as timed out now.
 676                  */
 677                 sc_tv.tv_sec = sc->sc_advbase * 3;
 678                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 679                         carp_master_down_locked(sc, "master will time out");
 680                         break;
 681                 }
 682
 683                 /*
 684                  * Otherwise, we reset the counter and wait for the next
 685                  * advertisement.
 686                  */
 687                 carp_setrun(sc, af);
 688                 break;
 689         }
 690
 691 out:
 692         CARP_UNLOCK(sc);
 693         m_freem(m);
 694 }
 695
 696 static int
 697 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 698 {
 699         struct m_tag *mtag;
 700
 701         if (sc->sc_init_counter) {
 702                 /* this could also be seconds since unix epoch */
 703                 sc->sc_counter = arc4random();
 704                 sc->sc_counter = sc->sc_counter << 32;
 705                 sc->sc_counter += arc4random();
 706         } else
 707                 sc->sc_counter++;
 708
 709         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 710         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 711
 712         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 713
 714         /* Tag packet for carp_output */
 715         if ((mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct carp_softc *),
 716             M_NOWAIT)) == NULL) {
 717                 m_freem(m);
 718                 CARPSTATS_INC(carps_onomem);
 719                 return (ENOMEM);
 720         }
 721         bcopy(&sc, mtag + 1, sizeof(sc));
 722         m_tag_prepend(m, mtag);
 723
 724         return (0);
 725 }
 726
 727 /*
 728  * To avoid LORs and possible recursions this function shouldn't
 729  * be called directly, but scheduled via taskqueue.
 730  */
 731 static void
 732 carp_send_ad_all(void *ctx __unused, int pending __unused)
 733 {
 734         struct carp_softc *sc;
 735
 736         mtx_lock(&carp_mtx);
 737         LIST_FOREACH(sc, &carp_list, sc_next)
 738                 if (sc->sc_state == MASTER) {
 739                         CARP_LOCK(sc);
 740                         CURVNET_SET(sc->sc_carpdev->if_vnet);
 741                         carp_send_ad_locked(sc);
 742                         CURVNET_RESTORE();
 743                         CARP_UNLOCK(sc);
 744                 }
 745         mtx_unlock(&carp_mtx);
 746 }
 747
 748 /* Send a periodic advertisement, executed in callout context. */
 749 static void
 750 carp_send_ad(void *v)
 751 {
 752         struct carp_softc *sc = v;
 753
 754         CARP_LOCK_ASSERT(sc);
 755         CURVNET_SET(sc->sc_carpdev->if_vnet);
 756         carp_send_ad_locked(sc);
 757         CURVNET_RESTORE();
 758         CARP_UNLOCK(sc);
 759 }
 760
 761 static void
 762 carp_send_ad_error(struct carp_softc *sc, int error)
 763 {
 764
 765         if (error) {
 766                 if (sc->sc_sendad_errors < INT_MAX)
 767                         sc->sc_sendad_errors++;
 768                 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 769                         static const char fmt[] = "send error %d on %s";
 770                         char msg[sizeof(fmt) + IFNAMSIZ];
 771
 772                         sprintf(msg, fmt, error, sc->sc_carpdev->if_xname);
 773                         carp_demote_adj(V_carp_senderr_adj, msg);
 774                 }
 775                 sc->sc_sendad_success = 0;
 776         } else {
 777                 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS &&
 778                     ++sc->sc_sendad_success >= CARP_SENDAD_MIN_SUCCESS) {
 779                         static const char fmt[] = "send ok on %s";
 780                         char msg[sizeof(fmt) + IFNAMSIZ];
 781
 782                         sprintf(msg, fmt, sc->sc_carpdev->if_xname);
 783                         carp_demote_adj(-V_carp_senderr_adj, msg);
 784                         sc->sc_sendad_errors = 0;
 785                 } else
 786                         sc->sc_sendad_errors = 0;
 787         }
 788 }
 789
 790 static void
 791 carp_send_ad_locked(struct carp_softc *sc)
 792 {
 793         struct carp_header ch;
 794         struct timeval tv;
 795         struct sockaddr sa;
 796         struct ifaddr *ifa;
 797         struct carp_header *ch_ptr;
 798         struct mbuf *m;
 799         int len, advskew;
 800
 801         CARP_LOCK_ASSERT(sc);
 802
 803         advskew = DEMOTE_ADVSKEW(sc);
 804         tv.tv_sec = sc->sc_advbase;
 805         tv.tv_usec = advskew * 1000000 / 256;
 806
 807         ch.carp_version = CARP_VERSION;
 808         ch.carp_type = CARP_ADVERTISEMENT;
 809         ch.carp_vhid = sc->sc_vhid;
 810         ch.carp_advbase = sc->sc_advbase;
 811         ch.carp_advskew = advskew;
 812         ch.carp_authlen = 7;    /* XXX DEFINE */
 813         ch.carp_pad1 = 0;       /* must be zero */
 814         ch.carp_cksum = 0;
 815
 816         /* XXXGL: OpenBSD picks first ifaddr with needed family. */
 817
 818 #ifdef INET
 819         if (sc->sc_naddrs) {
 820                 struct ip *ip;
 821
 822                 m = m_gethdr(M_NOWAIT, MT_DATA);
 823                 if (m == NULL) {
 824                         CARPSTATS_INC(carps_onomem);
 825                         goto resched;
 826                 }
 827                 len = sizeof(*ip) + sizeof(ch);
 828                 m->m_pkthdr.len = len;
 829                 m->m_pkthdr.rcvif = NULL;
 830                 m->m_len = len;
 831                 M_ALIGN(m, m->m_len);
 832                 m->m_flags |= M_MCAST;
 833                 ip = mtod(m, struct ip *);
 834                 ip->ip_v = IPVERSION;
 835                 ip->ip_hl = sizeof(*ip) >> 2;
 836                 ip->ip_tos = IPTOS_LOWDELAY;
 837                 ip->ip_len = htons(len);
 838                 ip->ip_off = htons(IP_DF);
 839                 ip->ip_ttl = CARP_DFLTTL;
 840                 ip->ip_p = IPPROTO_CARP;
 841                 ip->ip_sum = 0;
 842                 ip_fillid(ip);
 843
 844                 bzero(&sa, sizeof(sa));
 845                 sa.sa_family = AF_INET;
 846                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 847                 if (ifa != NULL) {
 848                         ip->ip_src.s_addr =
 849                             ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 850                         ifa_free(ifa);
 851                 } else
 852                         ip->ip_src.s_addr = 0;
 853                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 854
 855                 ch_ptr = (struct carp_header *)(&ip[1]);
 856                 bcopy(&ch, ch_ptr, sizeof(ch));
 857                 if (carp_prepare_ad(m, sc, ch_ptr))
 858                         goto resched;
 859
 860                 m->m_data += sizeof(*ip);
 861                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip));
 862                 m->m_data -= sizeof(*ip);
 863
 864                 CARPSTATS_INC(carps_opackets);
 865
 866                 carp_send_ad_error(sc, ip_output(m, NULL, NULL, IP_RAWOUTPUT,
 867                     &sc->sc_carpdev->if_carp->cif_imo, NULL));
 868         }
 869 #endif /* INET */
 870 #ifdef INET6
 871         if (sc->sc_naddrs6) {
 872                 struct ip6_hdr *ip6;
 873
 874                 m = m_gethdr(M_NOWAIT, MT_DATA);
 875                 if (m == NULL) {
 876                         CARPSTATS_INC(carps_onomem);
 877                         goto resched;
 878                 }
 879                 len = sizeof(*ip6) + sizeof(ch);
 880                 m->m_pkthdr.len = len;
 881                 m->m_pkthdr.rcvif = NULL;
 882                 m->m_len = len;
 883                 M_ALIGN(m, m->m_len);
 884                 m->m_flags |= M_MCAST;
 885                 ip6 = mtod(m, struct ip6_hdr *);
 886                 bzero(ip6, sizeof(*ip6));
 887                 ip6->ip6_vfc |= IPV6_VERSION;
 888                 ip6->ip6_hlim = CARP_DFLTTL;
 889                 ip6->ip6_nxt = IPPROTO_CARP;
 890                 bzero(&sa, sizeof(sa));
 891
 892                 /* set the source address */
 893                 sa.sa_family = AF_INET6;
 894                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 895                 if (ifa != NULL) {
 896                         bcopy(IFA_IN6(ifa), &ip6->ip6_src,
 897                             sizeof(struct in6_addr));
 898                         ifa_free(ifa);
 899                 } else
 900                         /* This should never happen with IPv6. */
 901                         bzero(&ip6->ip6_src, sizeof(struct in6_addr));
 902
 903                 /* Set the multicast destination. */
 904                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
 905                 ip6->ip6_dst.s6_addr8[15] = 0x12;
 906                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
 907                         m_freem(m);
 908                         CARP_DEBUG("%s: in6_setscope failed\n", __func__);
 909                         goto resched;
 910                 }
 911
 912                 ch_ptr = (struct carp_header *)(&ip6[1]);
 913                 bcopy(&ch, ch_ptr, sizeof(ch));
 914                 if (carp_prepare_ad(m, sc, ch_ptr))
 915                         goto resched;
 916
 917                 m->m_data += sizeof(*ip6);
 918                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip6));
 919                 m->m_data -= sizeof(*ip6);
 920
 921                 CARPSTATS_INC(carps_opackets6);
 922
 923                 carp_send_ad_error(sc, ip6_output(m, NULL, NULL, 0,
 924                     &sc->sc_carpdev->if_carp->cif_im6o, NULL, NULL));
 925         }
 926 #endif /* INET6 */
 927
 928 resched:
 929         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), carp_send_ad, sc);
 930 }
 931
 932 static void
 933 carp_addroute(struct carp_softc *sc)
 934 {
 935         struct ifaddr *ifa;
 936
 937         CARP_FOREACH_IFA(sc, ifa)
 938                 carp_ifa_addroute(ifa);
 939 }
 940
 941 static void
 942 carp_ifa_addroute(struct ifaddr *ifa)
 943 {
 944
 945         switch (ifa->ifa_addr->sa_family) {
 946 #ifdef INET
 947         case AF_INET:
 948                 in_addprefix(ifatoia(ifa), RTF_UP);
 949                 ifa_add_loopback_route(ifa,
 950                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 951                 break;
 952 #endif
 953 #ifdef INET6
 954         case AF_INET6:
 955                 ifa_add_loopback_route(ifa,
 956                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 957                 nd6_add_ifa_lle(ifatoia6(ifa));
 958                 break;
 959 #endif
 960         }
 961 }
 962
 963 static void
 964 carp_delroute(struct carp_softc *sc)
 965 {
 966         struct ifaddr *ifa;
 967
 968         CARP_FOREACH_IFA(sc, ifa)
 969                 carp_ifa_delroute(ifa);
 970 }
 971
 972 static void
 973 carp_ifa_delroute(struct ifaddr *ifa)
 974 {
 975
 976         switch (ifa->ifa_addr->sa_family) {
 977 #ifdef INET
 978         case AF_INET:
 979                 ifa_del_loopback_route(ifa,
 980                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 981                 in_scrubprefix(ifatoia(ifa), LLE_STATIC);
 982                 break;
 983 #endif
 984 #ifdef INET6
 985         case AF_INET6:
 986                 ifa_del_loopback_route(ifa,
 987                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 988                 nd6_rem_ifa_lle(ifatoia6(ifa), 1);
 989                 break;
 990 #endif
 991         }
 992 }
 993
 994 int
 995 carp_master(struct ifaddr *ifa)
 996 {
 997         struct carp_softc *sc = ifa->ifa_carp;
 998
 999         return (sc->sc_state == MASTER);
1000 }
1001
1002 #ifdef INET
1003 /*
1004  * Broadcast a gratuitous ARP request containing
1005  * the virtual router MAC address for each IP address
1006  * associated with the virtual router.
1007  */
1008 static void
1009 carp_send_arp(struct carp_softc *sc)
1010 {
1011         struct ifaddr *ifa;
1012
1013         CARP_FOREACH_IFA(sc, ifa)
1014                 if (ifa->ifa_addr->sa_family == AF_INET)
1015                         arp_ifinit2(sc->sc_carpdev, ifa, LLADDR(&sc->sc_addr));
1016 }
1017
1018 int
1019 carp_iamatch(struct ifaddr *ifa, uint8_t **enaddr)
1020 {
1021         struct carp_softc *sc = ifa->ifa_carp;
1022
1023         if (sc->sc_state == MASTER) {
1024                 *enaddr = LLADDR(&sc->sc_addr);
1025                 return (1);
1026         }
1027
1028         return (0);
1029 }
1030 #endif
1031
1032 #ifdef INET6
1033 static void
1034 carp_send_na(struct carp_softc *sc)
1035 {
1036         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1037         struct ifaddr *ifa;
1038         struct in6_addr *in6;
1039
1040         CARP_FOREACH_IFA(sc, ifa) {
1041                 if (ifa->ifa_addr->sa_family != AF_INET6)
1042                         continue;
1043
1044                 in6 = IFA_IN6(ifa);
1045                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1046                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1047                 DELAY(1000);    /* XXX */
1048         }
1049 }
1050
1051 /*
1052  * Returns ifa in case it's a carp address and it is MASTER, or if the address
1053  * matches and is not a carp address.  Returns NULL otherwise.
1054  */
1055 struct ifaddr *
1056 carp_iamatch6(struct ifnet *ifp, struct in6_addr *taddr)
1057 {
1058         struct ifaddr *ifa;
1059
1060         ifa = NULL;
1061         IF_ADDR_RLOCK(ifp);
1062         TAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
1063                 if (ifa->ifa_addr->sa_family != AF_INET6)
1064                         continue;
1065                 if (!IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa)))
1066                         continue;
1067                 if (ifa->ifa_carp && ifa->ifa_carp->sc_state != MASTER)
1068                         ifa = NULL;
1069                 else
1070                         ifa_ref(ifa);
1071                 break;
1072         }
1073         IF_ADDR_RUNLOCK(ifp);
1074
1075         return (ifa);
1076 }
1077
1078 caddr_t
1079 carp_macmatch6(struct ifnet *ifp, struct mbuf *m, const struct in6_addr *taddr)
1080 {
1081         struct ifaddr *ifa;
1082
1083         IF_ADDR_RLOCK(ifp);
1084         IFNET_FOREACH_IFA(ifp, ifa)
1085                 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1086                     IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1087                         struct carp_softc *sc = ifa->ifa_carp;
1088                         struct m_tag *mtag;
1089
1090                         IF_ADDR_RUNLOCK(ifp);
1091
1092                         mtag = m_tag_get(PACKET_TAG_CARP,
1093                             sizeof(struct carp_softc *), M_NOWAIT);
1094                         if (mtag == NULL)
1095                                 /* Better a bit than nothing. */
1096                                 return (LLADDR(&sc->sc_addr));
1097
1098                         bcopy(&sc, mtag + 1, sizeof(sc));
1099                         m_tag_prepend(m, mtag);
1100
1101                         return (LLADDR(&sc->sc_addr));
1102                 }
1103         IF_ADDR_RUNLOCK(ifp);
1104
1105         return (NULL);
1106 }
1107 #endif /* INET6 */
1108
1109 int
1110 carp_forus(struct ifnet *ifp, u_char *dhost)
1111 {
1112         struct carp_softc *sc;
1113         uint8_t *ena = dhost;
1114
1115         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1116                 return (0);
1117
1118         CIF_LOCK(ifp->if_carp);
1119         IFNET_FOREACH_CARP(ifp, sc) {
1120                 CARP_LOCK(sc);
1121                 if (sc->sc_state == MASTER && !bcmp(dhost, LLADDR(&sc->sc_addr),
1122                     ETHER_ADDR_LEN)) {
1123                         CARP_UNLOCK(sc);
1124                         CIF_UNLOCK(ifp->if_carp);
1125                         return (1);
1126                 }
1127                 CARP_UNLOCK(sc);
1128         }
1129         CIF_UNLOCK(ifp->if_carp);
1130
1131         return (0);
1132 }
1133
1134 /* Master down timeout event, executed in callout context. */
1135 static void
1136 carp_master_down(void *v)
1137 {
1138         struct carp_softc *sc = v;
1139
1140         CARP_LOCK_ASSERT(sc);
1141
1142         CURVNET_SET(sc->sc_carpdev->if_vnet);
1143         if (sc->sc_state == BACKUP) {
1144                 carp_master_down_locked(sc, "master timed out");
1145         }
1146         CURVNET_RESTORE();
1147
1148         CARP_UNLOCK(sc);
1149 }
1150
1151 static void
1152 carp_master_down_locked(struct carp_softc *sc, const char *reason)
1153 {
1154
1155         CARP_LOCK_ASSERT(sc);
1156
1157         switch (sc->sc_state) {
1158         case BACKUP:
1159                 carp_set_state(sc, MASTER, reason);
1160                 carp_send_ad_locked(sc);
1161 #ifdef INET
1162                 carp_send_arp(sc);
1163 #endif
1164 #ifdef INET6
1165                 carp_send_na(sc);
1166 #endif
1167                 carp_setrun(sc, 0);
1168                 carp_addroute(sc);
1169                 break;
1170         case INIT:
1171         case MASTER:
1172 #ifdef INVARIANTS
1173                 panic("carp: VHID %u@%s: master_down event in %s state\n",
1174                     sc->sc_vhid,
1175                     sc->sc_carpdev->if_xname,
1176                     sc->sc_state ? "MASTER" : "INIT");
1177 #endif
1178                 break;
1179         }
1180 }
1181
1182 /*
1183  * When in backup state, af indicates whether to reset the master down timer
1184  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1185  */
1186 static void
1187 carp_setrun(struct carp_softc *sc, sa_family_t af)
1188 {
1189         struct timeval tv;
1190
1191         CARP_LOCK_ASSERT(sc);
1192
1193         if ((sc->sc_carpdev->if_flags & IFF_UP) == 0 ||
1194             sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1195             (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0))
1196                 return;
1197
1198         switch (sc->sc_state) {
1199         case INIT:
1200                 carp_set_state(sc, BACKUP, "initialization complete");
1201                 carp_setrun(sc, 0);
1202                 break;
1203         case BACKUP:
1204                 callout_stop(&sc->sc_ad_tmo);
1205                 tv.tv_sec = 3 * sc->sc_advbase;
1206                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1207                 switch (af) {
1208 #ifdef INET
1209                 case AF_INET:
1210                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1211                             carp_master_down, sc);
1212                         break;
1213 #endif
1214 #ifdef INET6
1215                 case AF_INET6:
1216                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1217                             carp_master_down, sc);
1218                         break;
1219 #endif
1220                 default:
1221 #ifdef INET
1222                         if (sc->sc_naddrs)
1223                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1224                                     carp_master_down, sc);
1225 #endif
1226 #ifdef INET6
1227                         if (sc->sc_naddrs6)
1228                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1229                                     carp_master_down, sc);
1230 #endif
1231                         break;
1232                 }
1233                 break;
1234         case MASTER:
1235                 tv.tv_sec = sc->sc_advbase;
1236                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1237                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1238                     carp_send_ad, sc);
1239                 break;
1240         }
1241 }
1242
1243 /*
1244  * Setup multicast structures.
1245  */
1246 static int
1247 carp_multicast_setup(struct carp_if *cif, sa_family_t sa)
1248 {
1249         struct ifnet *ifp = cif->cif_ifp;
1250         int error = 0;
1251
1252         switch (sa) {
1253 #ifdef INET
1254         case AF_INET:
1255             {
1256                 struct ip_moptions *imo = &cif->cif_imo;
1257                 struct in_addr addr;
1258
1259                 if (imo->imo_membership)
1260                         return (0);
1261
1262                 imo->imo_membership = (struct in_multi **)malloc(
1263                     (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
1264                     M_WAITOK);
1265                 imo->imo_mfilters = NULL;
1266                 imo->imo_max_memberships = IP_MIN_MEMBERSHIPS;
1267                 imo->imo_multicast_vif = -1;
1268
1269                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1270                 if ((error = in_joingroup(ifp, &addr, NULL,
1271                     &imo->imo_membership[0])) != 0) {
1272                         free(imo->imo_membership, M_CARP);
1273                         break;
1274                 }
1275                 imo->imo_num_memberships++;
1276                 imo->imo_multicast_ifp = ifp;
1277                 imo->imo_multicast_ttl = CARP_DFLTTL;
1278                 imo->imo_multicast_loop = 0;
1279                 break;
1280            }
1281 #endif
1282 #ifdef INET6
1283         case AF_INET6:
1284             {
1285                 struct ip6_moptions *im6o = &cif->cif_im6o;
1286                 struct in6_addr in6;
1287                 struct in6_multi *in6m;
1288
1289                 if (im6o->im6o_membership)
1290                         return (0);
1291
1292                 im6o->im6o_membership = (struct in6_multi **)malloc(
1293                     (sizeof(struct in6_multi *) * IPV6_MIN_MEMBERSHIPS), M_CARP,
1294                     M_ZERO | M_WAITOK);
1295                 im6o->im6o_mfilters = NULL;
1296                 im6o->im6o_max_memberships = IPV6_MIN_MEMBERSHIPS;
1297                 im6o->im6o_multicast_hlim = CARP_DFLTTL;
1298                 im6o->im6o_multicast_ifp = ifp;
1299
1300                 /* Join IPv6 CARP multicast group. */
1301                 bzero(&in6, sizeof(in6));
1302                 in6.s6_addr16[0] = htons(0xff02);
1303                 in6.s6_addr8[15] = 0x12;
1304                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1305                         free(im6o->im6o_membership, M_CARP);
1306                         break;
1307                 }
1308                 in6m = NULL;
1309                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1310                         free(im6o->im6o_membership, M_CARP);
1311                         break;
1312                 }
1313                 im6o->im6o_membership[0] = in6m;
1314                 im6o->im6o_num_memberships++;
1315
1316                 /* Join solicited multicast address. */
1317                 bzero(&in6, sizeof(in6));
1318                 in6.s6_addr16[0] = htons(0xff02);
1319                 in6.s6_addr32[1] = 0;
1320                 in6.s6_addr32[2] = htonl(1);
1321                 in6.s6_addr32[3] = 0;
1322                 in6.s6_addr8[12] = 0xff;
1323                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1324                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1325                         free(im6o->im6o_membership, M_CARP);
1326                         break;
1327                 }
1328                 in6m = NULL;
1329                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1330                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1331                         free(im6o->im6o_membership, M_CARP);
1332                         break;
1333                 }
1334                 im6o->im6o_membership[1] = in6m;
1335                 im6o->im6o_num_memberships++;
1336                 break;
1337             }
1338 #endif
1339         }
1340
1341         return (error);
1342 }
1343
1344 /*
1345  * Free multicast structures.
1346  */
1347 static void
1348 carp_multicast_cleanup(struct carp_if *cif, sa_family_t sa)
1349 {
1350
1351         sx_assert(&carp_sx, SA_XLOCKED);
1352
1353         switch (sa) {
1354 #ifdef INET
1355         case AF_INET:
1356                 if (cif->cif_naddrs == 0) {
1357                         struct ip_moptions *imo = &cif->cif_imo;
1358
1359                         in_leavegroup(imo->imo_membership[0], NULL);
1360                         KASSERT(imo->imo_mfilters == NULL,
1361                             ("%s: imo_mfilters != NULL", __func__));
1362                         free(imo->imo_membership, M_CARP);
1363                         imo->imo_membership = NULL;
1364
1365                 }
1366                 break;
1367 #endif
1368 #ifdef INET6
1369         case AF_INET6:
1370                 if (cif->cif_naddrs6 == 0) {
1371                         struct ip6_moptions *im6o = &cif->cif_im6o;
1372
1373                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1374                         in6_mc_leave(im6o->im6o_membership[1], NULL);
1375                         KASSERT(im6o->im6o_mfilters == NULL,
1376                             ("%s: im6o_mfilters != NULL", __func__));
1377                         free(im6o->im6o_membership, M_CARP);
1378                         im6o->im6o_membership = NULL;
1379                 }
1380                 break;
1381 #endif
1382         }
1383 }
1384
1385 int
1386 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa)
1387 {
1388         struct m_tag *mtag;
1389         struct carp_softc *sc;
1390
1391         if (!sa)
1392                 return (0);
1393
1394         switch (sa->sa_family) {
1395 #ifdef INET
1396         case AF_INET:
1397                 break;
1398 #endif
1399 #ifdef INET6
1400         case AF_INET6:
1401                 break;
1402 #endif
1403         default:
1404                 return (0);
1405         }
1406
1407         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
1408         if (mtag == NULL)
1409                 return (0);
1410
1411         bcopy(mtag + 1, &sc, sizeof(sc));
1412
1413         /* Set the source MAC address to the Virtual Router MAC Address. */
1414         switch (ifp->if_type) {
1415         case IFT_ETHER:
1416         case IFT_BRIDGE:
1417         case IFT_L2VLAN: {
1418                         struct ether_header *eh;
1419
1420                         eh = mtod(m, struct ether_header *);
1421                         eh->ether_shost[0] = 0;
1422                         eh->ether_shost[1] = 0;
1423                         eh->ether_shost[2] = 0x5e;
1424                         eh->ether_shost[3] = 0;
1425                         eh->ether_shost[4] = 1;
1426                         eh->ether_shost[5] = sc->sc_vhid;
1427                 }
1428                 break;
1429         case IFT_FDDI: {
1430                         struct fddi_header *fh;
1431
1432                         fh = mtod(m, struct fddi_header *);
1433                         fh->fddi_shost[0] = 0;
1434                         fh->fddi_shost[1] = 0;
1435                         fh->fddi_shost[2] = 0x5e;
1436                         fh->fddi_shost[3] = 0;
1437                         fh->fddi_shost[4] = 1;
1438                         fh->fddi_shost[5] = sc->sc_vhid;
1439                 }
1440                 break;
1441         case IFT_ISO88025: {
1442                         struct iso88025_header *th;
1443                         th = mtod(m, struct iso88025_header *);
1444                         th->iso88025_shost[0] = 3;
1445                         th->iso88025_shost[1] = 0;
1446                         th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
1447                         th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
1448                         th->iso88025_shost[4] = 0;
1449                         th->iso88025_shost[5] = 0;
1450                 }
1451                 break;
1452         default:
1453                 printf("%s: carp is not supported for the %d interface type\n",
1454                     ifp->if_xname, ifp->if_type);
1455                 return (EOPNOTSUPP);
1456         }
1457
1458         return (0);
1459 }
1460
1461 static struct carp_softc*
1462 carp_alloc(struct ifnet *ifp)
1463 {
1464         struct carp_softc *sc;
1465         struct carp_if *cif;
1466
1467         if ((cif = ifp->if_carp) == NULL)
1468                 cif = carp_alloc_if(ifp);
1469
1470         sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
1471
1472         sc->sc_advbase = CARP_DFLTINTV;
1473         sc->sc_vhid = -1;       /* required setting */
1474         sc->sc_init_counter = 1;
1475         sc->sc_state = INIT;
1476
1477         sc->sc_ifasiz = sizeof(struct ifaddr *);
1478         sc->sc_ifas = malloc(sc->sc_ifasiz, M_CARP, M_WAITOK|M_ZERO);
1479         sc->sc_carpdev = ifp;
1480
1481         CARP_LOCK_INIT(sc);
1482 #ifdef INET
1483         callout_init_mtx(&sc->sc_md_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1484 #endif
1485 #ifdef INET6
1486         callout_init_mtx(&sc->sc_md6_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1487 #endif
1488         callout_init_mtx(&sc->sc_ad_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1489
1490         CIF_LOCK(cif);
1491         TAILQ_INSERT_TAIL(&cif->cif_vrs, sc, sc_list);
1492         CIF_UNLOCK(cif);
1493
1494         mtx_lock(&carp_mtx);
1495         LIST_INSERT_HEAD(&carp_list, sc, sc_next);
1496         mtx_unlock(&carp_mtx);
1497
1498         return (sc);
1499 }
1500
1501 static void
1502 carp_grow_ifas(struct carp_softc *sc)
1503 {
1504         struct ifaddr **new;
1505
1506         new = malloc(sc->sc_ifasiz * 2, M_CARP, M_WAITOK | M_ZERO);
1507         CARP_LOCK(sc);
1508         bcopy(sc->sc_ifas, new, sc->sc_ifasiz);
1509         free(sc->sc_ifas, M_CARP);
1510         sc->sc_ifas = new;
1511         sc->sc_ifasiz *= 2;
1512         CARP_UNLOCK(sc);
1513 }
1514
1515 static void
1516 carp_destroy(struct carp_softc *sc)
1517 {
1518         struct ifnet *ifp = sc->sc_carpdev;
1519         struct carp_if *cif = ifp->if_carp;
1520
1521         sx_assert(&carp_sx, SA_XLOCKED);
1522
1523         if (sc->sc_suppress)
1524                 carp_demote_adj(-V_carp_ifdown_adj, "vhid removed");
1525         CARP_UNLOCK(sc);
1526
1527         CIF_LOCK(cif);
1528         TAILQ_REMOVE(&cif->cif_vrs, sc, sc_list);
1529         CIF_UNLOCK(cif);
1530
1531         mtx_lock(&carp_mtx);
1532         LIST_REMOVE(sc, sc_next);
1533         mtx_unlock(&carp_mtx);
1534
1535         callout_drain(&sc->sc_ad_tmo);
1536 #ifdef INET
1537         callout_drain(&sc->sc_md_tmo);
1538 #endif
1539 #ifdef INET6
1540         callout_drain(&sc->sc_md6_tmo);
1541 #endif
1542         CARP_LOCK_DESTROY(sc);
1543
1544         free(sc->sc_ifas, M_CARP);
1545         free(sc, M_CARP);
1546 }
1547
1548 static struct carp_if*
1549 carp_alloc_if(struct ifnet *ifp)
1550 {
1551         struct carp_if *cif;
1552         int error;
1553
1554         cif = malloc(sizeof(*cif), M_CARP, M_WAITOK|M_ZERO);
1555
1556         if ((error = ifpromisc(ifp, 1)) != 0)
1557                 printf("%s: ifpromisc(%s) failed: %d\n",
1558                     __func__, ifp->if_xname, error);
1559         else
1560                 cif->cif_flags |= CIF_PROMISC;
1561
1562         CIF_LOCK_INIT(cif);
1563         cif->cif_ifp = ifp;
1564         TAILQ_INIT(&cif->cif_vrs);
1565
1566         IF_ADDR_WLOCK(ifp);
1567         ifp->if_carp = cif;
1568         if_ref(ifp);
1569         IF_ADDR_WUNLOCK(ifp);
1570
1571         return (cif);
1572 }
1573
1574 static void
1575 carp_free_if(struct carp_if *cif)
1576 {
1577         struct ifnet *ifp = cif->cif_ifp;
1578
1579         CIF_LOCK_ASSERT(cif);
1580         KASSERT(TAILQ_EMPTY(&cif->cif_vrs), ("%s: softc list not empty",
1581             __func__));
1582
1583         IF_ADDR_WLOCK(ifp);
1584         ifp->if_carp = NULL;
1585         IF_ADDR_WUNLOCK(ifp);
1586
1587         CIF_LOCK_DESTROY(cif);
1588
1589         if (cif->cif_flags & CIF_PROMISC)
1590                 ifpromisc(ifp, 0);
1591         if_rele(ifp);
1592
1593         free(cif, M_CARP);
1594 }
1595
1596 static void
1597 carp_carprcp(struct carpreq *carpr, struct carp_softc *sc, int priv)
1598 {
1599
1600         CARP_LOCK(sc);
1601         carpr->carpr_state = sc->sc_state;
1602         carpr->carpr_vhid = sc->sc_vhid;
1603         carpr->carpr_advbase = sc->sc_advbase;
1604         carpr->carpr_advskew = sc->sc_advskew;
1605         if (priv)
1606                 bcopy(sc->sc_key, carpr->carpr_key, sizeof(carpr->carpr_key));
1607         else
1608                 bzero(carpr->carpr_key, sizeof(carpr->carpr_key));
1609         CARP_UNLOCK(sc);
1610 }
1611
1612 int
1613 carp_ioctl(struct ifreq *ifr, u_long cmd, struct thread *td)
1614 {
1615         struct carpreq carpr;
1616         struct ifnet *ifp;
1617         struct carp_softc *sc = NULL;
1618         int error = 0, locked = 0;
1619
1620         if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1621                 return (error);
1622
1623         ifp = ifunit_ref(ifr->ifr_name);
1624         if (ifp == NULL)
1625                 return (ENXIO);
1626
1627         switch (ifp->if_type) {
1628         case IFT_ETHER:
1629         case IFT_L2VLAN:
1630         case IFT_BRIDGE:
1631         case IFT_FDDI:
1632         case IFT_ISO88025:
1633                 break;
1634         default:
1635                 error = EOPNOTSUPP;
1636                 goto out;
1637         }
1638
1639         if ((ifp->if_flags & IFF_MULTICAST) == 0) {
1640                 error = EADDRNOTAVAIL;
1641                 goto out;
1642         }
1643
1644         sx_xlock(&carp_sx);
1645         switch (cmd) {
1646         case SIOCSVH:
1647                 if ((error = priv_check(td, PRIV_NETINET_CARP)))
1648                         break;
1649                 if (carpr.carpr_vhid <= 0 || carpr.carpr_vhid > CARP_MAXVHID ||
1650                     carpr.carpr_advbase < 0 || carpr.carpr_advskew < 0) {
1651                         error = EINVAL;
1652                         break;
1653                 }
1654
1655                 if (ifp->if_carp) {
1656                         CIF_LOCK(ifp->if_carp);
1657                         IFNET_FOREACH_CARP(ifp, sc)
1658                                 if (sc->sc_vhid == carpr.carpr_vhid)
1659                                         break;
1660                         CIF_UNLOCK(ifp->if_carp);
1661                 }
1662                 if (sc == NULL) {
1663                         sc = carp_alloc(ifp);
1664                         CARP_LOCK(sc);
1665                         sc->sc_vhid = carpr.carpr_vhid;
1666                         LLADDR(&sc->sc_addr)[0] = 0;
1667                         LLADDR(&sc->sc_addr)[1] = 0;
1668                         LLADDR(&sc->sc_addr)[2] = 0x5e;
1669                         LLADDR(&sc->sc_addr)[3] = 0;
1670                         LLADDR(&sc->sc_addr)[4] = 1;
1671                         LLADDR(&sc->sc_addr)[5] = sc->sc_vhid;
1672                 } else
1673                         CARP_LOCK(sc);
1674                 locked = 1;
1675                 if (carpr.carpr_advbase > 0) {
1676                         if (carpr.carpr_advbase > 255 ||
1677                             carpr.carpr_advbase < CARP_DFLTINTV) {
1678                                 error = EINVAL;
1679                                 break;
1680                         }
1681                         sc->sc_advbase = carpr.carpr_advbase;
1682                 }
1683                 if (carpr.carpr_advskew >= 255) {
1684                         error = EINVAL;
1685                         break;
1686                 }
1687                 sc->sc_advskew = carpr.carpr_advskew;
1688                 if (carpr.carpr_key[0] != '\0') {
1689                         bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1690                         carp_hmac_prepare(sc);
1691                 }
1692                 if (sc->sc_state != INIT &&
1693                     carpr.carpr_state != sc->sc_state) {
1694                         switch (carpr.carpr_state) {
1695                         case BACKUP:
1696                                 callout_stop(&sc->sc_ad_tmo);
1697                                 carp_set_state(sc, BACKUP,
1698                                     "user requested via ifconfig");
1699                                 carp_setrun(sc, 0);
1700                                 carp_delroute(sc);
1701                                 break;
1702                         case MASTER:
1703                                 carp_master_down_locked(sc,
1704                                     "user requested via ifconfig");
1705                                 break;
1706                         default:
1707                                 break;
1708                         }
1709                 }
1710                 break;
1711
1712         case SIOCGVH:
1713             {
1714                 int priveleged;
1715
1716                 if (carpr.carpr_vhid < 0 || carpr.carpr_vhid > CARP_MAXVHID) {
1717                         error = EINVAL;
1718                         break;
1719                 }
1720                 if (carpr.carpr_count < 1) {
1721                         error = EMSGSIZE;
1722                         break;
1723                 }
1724                 if (ifp->if_carp == NULL) {
1725                         error = ENOENT;
1726                         break;
1727                 }
1728
1729                 priveleged = (priv_check(td, PRIV_NETINET_CARP) == 0);
1730                 if (carpr.carpr_vhid != 0) {
1731                         CIF_LOCK(ifp->if_carp);
1732                         IFNET_FOREACH_CARP(ifp, sc)
1733                                 if (sc->sc_vhid == carpr.carpr_vhid)
1734                                         break;
1735                         CIF_UNLOCK(ifp->if_carp);
1736                         if (sc == NULL) {
1737                                 error = ENOENT;
1738                                 break;
1739                         }
1740                         carp_carprcp(&carpr, sc, priveleged);
1741                         error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1742                 } else  {
1743                         int i, count;
1744
1745                         count = 0;
1746                         CIF_LOCK(ifp->if_carp);
1747                         IFNET_FOREACH_CARP(ifp, sc)
1748                                 count++;
1749
1750                         if (count > carpr.carpr_count) {
1751                                 CIF_UNLOCK(ifp->if_carp);
1752                                 error = EMSGSIZE;
1753                                 break;
1754                         }
1755
1756                         i = 0;
1757                         IFNET_FOREACH_CARP(ifp, sc) {
1758                                 carp_carprcp(&carpr, sc, priveleged);
1759                                 carpr.carpr_count = count;
1760                                 error = copyout(&carpr, ifr->ifr_data +
1761                                     (i * sizeof(carpr)), sizeof(carpr));
1762                                 if (error) {
1763                                         CIF_UNLOCK(ifp->if_carp);
1764                                         break;
1765                                 }
1766                                 i++;
1767                         }
1768                         CIF_UNLOCK(ifp->if_carp);
1769                 }
1770                 break;
1771             }
1772         default:
1773                 error = EINVAL;
1774         }
1775         sx_xunlock(&carp_sx);
1776
1777 out:
1778         if (locked)
1779                 CARP_UNLOCK(sc);
1780         if_rele(ifp);
1781
1782         return (error);
1783 }
1784
1785 static int
1786 carp_get_vhid(struct ifaddr *ifa)
1787 {
1788
1789         if (ifa == NULL || ifa->ifa_carp == NULL)
1790                 return (0);
1791
1792         return (ifa->ifa_carp->sc_vhid);
1793 }
1794
1795 int
1796 carp_attach(struct ifaddr *ifa, int vhid)
1797 {
1798         struct ifnet *ifp = ifa->ifa_ifp;
1799         struct carp_if *cif = ifp->if_carp;
1800         struct carp_softc *sc;
1801         int index, error;
1802
1803         KASSERT(ifa->ifa_carp == NULL, ("%s: ifa %p attached", __func__, ifa));
1804
1805         switch (ifa->ifa_addr->sa_family) {
1806 #ifdef INET
1807         case AF_INET:
1808 #endif
1809 #ifdef INET6
1810         case AF_INET6:
1811 #endif
1812                 break;
1813         default:
1814                 return (EPROTOTYPE);
1815         }
1816
1817         sx_xlock(&carp_sx);
1818         if (ifp->if_carp == NULL) {
1819                 sx_xunlock(&carp_sx);
1820                 return (ENOPROTOOPT);
1821         }
1822
1823         CIF_LOCK(cif);
1824         IFNET_FOREACH_CARP(ifp, sc)
1825                 if (sc->sc_vhid == vhid)
1826                         break;
1827         CIF_UNLOCK(cif);
1828         if (sc == NULL) {
1829                 sx_xunlock(&carp_sx);
1830                 return (ENOENT);
1831         }
1832
1833         error = carp_multicast_setup(cif, ifa->ifa_addr->sa_family);
1834         if (error) {
1835                 CIF_FREE(cif);
1836                 sx_xunlock(&carp_sx);
1837                 return (error);
1838         }
1839
1840         index = sc->sc_naddrs + sc->sc_naddrs6 + 1;
1841         if (index > sc->sc_ifasiz / sizeof(struct ifaddr *))
1842                 carp_grow_ifas(sc);
1843
1844         switch (ifa->ifa_addr->sa_family) {
1845 #ifdef INET
1846         case AF_INET:
1847                 cif->cif_naddrs++;
1848                 sc->sc_naddrs++;
1849                 break;
1850 #endif
1851 #ifdef INET6
1852         case AF_INET6:
1853                 cif->cif_naddrs6++;
1854                 sc->sc_naddrs6++;
1855                 break;
1856 #endif
1857         }
1858
1859         ifa_ref(ifa);
1860
1861         CARP_LOCK(sc);
1862         sc->sc_ifas[index - 1] = ifa;
1863         ifa->ifa_carp = sc;
1864         carp_hmac_prepare(sc);
1865         carp_sc_state(sc);
1866         CARP_UNLOCK(sc);
1867
1868         sx_xunlock(&carp_sx);
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         struct carp_softc *sc = ifa->ifa_carp;
1879         int i, index;
1880
1881         KASSERT(sc != NULL, ("%s: %p not attached", __func__, ifa));
1882
1883         sx_xlock(&carp_sx);
1884
1885         CARP_LOCK(sc);
1886         /* Shift array. */
1887         index = sc->sc_naddrs + sc->sc_naddrs6;
1888         for (i = 0; i < index; i++)
1889                 if (sc->sc_ifas[i] == ifa)
1890                         break;
1891         KASSERT(i < index, ("%s: %p no backref", __func__, ifa));
1892         for (; i < index - 1; i++)
1893                 sc->sc_ifas[i] = sc->sc_ifas[i+1];
1894         sc->sc_ifas[index - 1] = NULL;
1895
1896         switch (ifa->ifa_addr->sa_family) {
1897 #ifdef INET
1898         case AF_INET:
1899                 cif->cif_naddrs--;
1900                 sc->sc_naddrs--;
1901                 break;
1902 #endif
1903 #ifdef INET6
1904         case AF_INET6:
1905                 cif->cif_naddrs6--;
1906                 sc->sc_naddrs6--;
1907                 break;
1908 #endif
1909         }
1910
1911         carp_ifa_delroute(ifa);
1912         carp_multicast_cleanup(cif, ifa->ifa_addr->sa_family);
1913
1914         ifa->ifa_carp = NULL;
1915         ifa_free(ifa);
1916
1917         carp_hmac_prepare(sc);
1918         carp_sc_state(sc);
1919
1920         if (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0)
1921                 carp_destroy(sc);
1922         else
1923                 CARP_UNLOCK(sc);
1924
1925         CIF_FREE(cif);
1926
1927         sx_xunlock(&carp_sx);
1928 }
1929
1930 static void
1931 carp_set_state(struct carp_softc *sc, int state, const char *reason)
1932 {
1933
1934         CARP_LOCK_ASSERT(sc);
1935
1936         if (sc->sc_state != state) {
1937                 const char *carp_states[] = { CARP_STATES };
1938                 char subsys[IFNAMSIZ+5];
1939
1940                 snprintf(subsys, IFNAMSIZ+5, "%u@%s", sc->sc_vhid,
1941                     sc->sc_carpdev->if_xname);
1942
1943                 CARP_LOG("%s: %s -> %s (%s)\n", subsys,
1944                     carp_states[sc->sc_state], carp_states[state], reason);
1945
1946                 sc->sc_state = state;
1947
1948                 devctl_notify("CARP", subsys, carp_states[state], NULL);
1949         }
1950 }
1951
1952 static void
1953 carp_linkstate(struct ifnet *ifp)
1954 {
1955         struct carp_softc *sc;
1956
1957         CIF_LOCK(ifp->if_carp);
1958         IFNET_FOREACH_CARP(ifp, sc) {
1959                 CARP_LOCK(sc);
1960                 carp_sc_state(sc);
1961                 CARP_UNLOCK(sc);
1962         }
1963         CIF_UNLOCK(ifp->if_carp);
1964 }
1965
1966 static void
1967 carp_sc_state(struct carp_softc *sc)
1968 {
1969
1970         CARP_LOCK_ASSERT(sc);
1971
1972         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1973             !(sc->sc_carpdev->if_flags & IFF_UP)) {
1974                 callout_stop(&sc->sc_ad_tmo);
1975 #ifdef INET
1976                 callout_stop(&sc->sc_md_tmo);
1977 #endif
1978 #ifdef INET6
1979                 callout_stop(&sc->sc_md6_tmo);
1980 #endif
1981                 carp_set_state(sc, INIT, "hardware interface down");
1982                 carp_setrun(sc, 0);
1983                 if (!sc->sc_suppress)
1984                         carp_demote_adj(V_carp_ifdown_adj, "interface down");
1985                 sc->sc_suppress = 1;
1986         } else {
1987                 carp_set_state(sc, INIT, "hardware interface up");
1988                 carp_setrun(sc, 0);
1989                 if (sc->sc_suppress)
1990                         carp_demote_adj(-V_carp_ifdown_adj, "interface up");
1991                 sc->sc_suppress = 0;
1992         }
1993 }
1994
1995 static void
1996 carp_demote_adj(int adj, char *reason)
1997 {
1998         atomic_add_int(&V_carp_demotion, adj);
1999         CARP_LOG("demoted by %d to %d (%s)\n", adj, V_carp_demotion, reason);
2000         taskqueue_enqueue(taskqueue_swi, &carp_sendall_task);
2001 }
2002
2003 static int
2004 carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS)
2005 {
2006         int new, error;
2007
2008         new = V_carp_demotion;
2009         error = sysctl_handle_int(oidp, &new, 0, req);
2010         if (error || !req->newptr)
2011                 return (error);
2012
2013         carp_demote_adj(new, "sysctl");
2014
2015         return (0);
2016 }
2017
2018 #ifdef INET
2019 extern  struct domain inetdomain;
2020 static struct protosw in_carp_protosw = {
2021         .pr_type =              SOCK_RAW,
2022         .pr_domain =            &inetdomain,
2023         .pr_protocol =          IPPROTO_CARP,
2024         .pr_flags =             PR_ATOMIC|PR_ADDR,
2025         .pr_input =             carp_input,
2026         .pr_output =            rip_output,
2027         .pr_ctloutput =         rip_ctloutput,
2028         .pr_usrreqs =           &rip_usrreqs
2029 };
2030 #endif
2031
2032 #ifdef INET6
2033 extern  struct domain inet6domain;
2034 static struct protosw in6_carp_protosw = {
2035         .pr_type =              SOCK_RAW,
2036         .pr_domain =            &inet6domain,
2037         .pr_protocol =          IPPROTO_CARP,
2038         .pr_flags =             PR_ATOMIC|PR_ADDR,
2039         .pr_input =             carp6_input,
2040         .pr_output =            rip6_output,
2041         .pr_ctloutput =         rip6_ctloutput,
2042         .pr_usrreqs =           &rip6_usrreqs
2043 };
2044 #endif
2045
2046 static void
2047 carp_mod_cleanup(void)
2048 {
2049
2050 #ifdef INET
2051         if (proto_reg[CARP_INET] == 0) {
2052                 (void)ipproto_unregister(IPPROTO_CARP);
2053                 pf_proto_unregister(PF_INET, IPPROTO_CARP, SOCK_RAW);
2054                 proto_reg[CARP_INET] = -1;
2055         }
2056         carp_iamatch_p = NULL;
2057 #endif
2058 #ifdef INET6
2059         if (proto_reg[CARP_INET6] == 0) {
2060                 (void)ip6proto_unregister(IPPROTO_CARP);
2061                 pf_proto_unregister(PF_INET6, IPPROTO_CARP, SOCK_RAW);
2062                 proto_reg[CARP_INET6] = -1;
2063         }
2064         carp_iamatch6_p = NULL;
2065         carp_macmatch6_p = NULL;
2066 #endif
2067         carp_ioctl_p = NULL;
2068         carp_attach_p = NULL;
2069         carp_detach_p = NULL;
2070         carp_get_vhid_p = NULL;
2071         carp_linkstate_p = NULL;
2072         carp_forus_p = NULL;
2073         carp_output_p = NULL;
2074         carp_demote_adj_p = NULL;
2075         carp_master_p = NULL;
2076         mtx_unlock(&carp_mtx);
2077         taskqueue_drain(taskqueue_swi, &carp_sendall_task);
2078         mtx_destroy(&carp_mtx);
2079         sx_destroy(&carp_sx);
2080 }
2081
2082 static int
2083 carp_mod_load(void)
2084 {
2085         int err;
2086
2087         mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2088         sx_init(&carp_sx, "carp_sx");
2089         LIST_INIT(&carp_list);
2090         carp_get_vhid_p = carp_get_vhid;
2091         carp_forus_p = carp_forus;
2092         carp_output_p = carp_output;
2093         carp_linkstate_p = carp_linkstate;
2094         carp_ioctl_p = carp_ioctl;
2095         carp_attach_p = carp_attach;
2096         carp_detach_p = carp_detach;
2097         carp_demote_adj_p = carp_demote_adj;
2098         carp_master_p = carp_master;
2099 #ifdef INET6
2100         carp_iamatch6_p = carp_iamatch6;
2101         carp_macmatch6_p = carp_macmatch6;
2102         proto_reg[CARP_INET6] = pf_proto_register(PF_INET6,
2103             (struct protosw *)&in6_carp_protosw);
2104         if (proto_reg[CARP_INET6]) {
2105                 printf("carp: error %d attaching to PF_INET6\n",
2106                     proto_reg[CARP_INET6]);
2107                 carp_mod_cleanup();
2108                 return (proto_reg[CARP_INET6]);
2109         }
2110         err = ip6proto_register(IPPROTO_CARP);
2111         if (err) {
2112                 printf("carp: error %d registering with INET6\n", err);
2113                 carp_mod_cleanup();
2114                 return (err);
2115         }
2116 #endif
2117 #ifdef INET
2118         carp_iamatch_p = carp_iamatch;
2119         proto_reg[CARP_INET] = pf_proto_register(PF_INET, &in_carp_protosw);
2120         if (proto_reg[CARP_INET]) {
2121                 printf("carp: error %d attaching to PF_INET\n",
2122                     proto_reg[CARP_INET]);
2123                 carp_mod_cleanup();
2124                 return (proto_reg[CARP_INET]);
2125         }
2126         err = ipproto_register(IPPROTO_CARP);
2127         if (err) {
2128                 printf("carp: error %d registering with INET\n", err);
2129                 carp_mod_cleanup();
2130                 return (err);
2131         }
2132 #endif
2133         return (0);
2134 }
2135
2136 static int
2137 carp_modevent(module_t mod, int type, void *data)
2138 {
2139         switch (type) {
2140         case MOD_LOAD:
2141                 return carp_mod_load();
2142                 /* NOTREACHED */
2143         case MOD_UNLOAD:
2144                 mtx_lock(&carp_mtx);
2145                 if (LIST_EMPTY(&carp_list))
2146                         carp_mod_cleanup();
2147                 else {
2148                         mtx_unlock(&carp_mtx);
2149                         return (EBUSY);
2150                 }
2151                 break;
2152
2153         default:
2154                 return (EINVAL);
2155         }
2156
2157         return (0);
2158 }
2159
2160 static moduledata_t carp_mod = {
2161         "carp",
2162         carp_modevent,
2163         0
2164 };
2165
2166 DECLARE_MODULE(carp, carp_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);