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/ip6protosw.h>
  82 #include <netinet6/in6_var.h>
  83 #include <netinet6/ip6_var.h>
  84 #include <netinet6/scope6_var.h>
  85 #include <netinet6/nd6.h>
  86 #endif
  87
  88 #include <crypto/sha1.h>
  89
  90 static MALLOC_DEFINE(M_CARP, "CARP", "CARP addresses");
  91
  92 struct carp_softc {
  93         struct ifnet            *sc_carpdev;    /* Pointer to parent ifnet. */
  94         struct ifaddr           **sc_ifas;      /* Our ifaddrs. */
  95         struct sockaddr_dl      sc_addr;        /* Our link level address. */
  96         struct callout          sc_ad_tmo;      /* Advertising timeout. */
  97 #ifdef INET
  98         struct callout          sc_md_tmo;      /* Master down timeout. */
  99 #endif
 100 #ifdef INET6
 101         struct callout          sc_md6_tmo;     /* XXX: Master down timeout. */
 102 #endif
 103         struct mtx              sc_mtx;
 104
 105         int                     sc_vhid;
 106         int                     sc_advskew;
 107         int                     sc_advbase;
 108
 109         int                     sc_naddrs;
 110         int                     sc_naddrs6;
 111         int                     sc_ifasiz;
 112         enum { INIT = 0, BACKUP, MASTER }       sc_state;
 113         int                     sc_suppress;
 114         int                     sc_sendad_errors;
 115 #define CARP_SENDAD_MAX_ERRORS  3
 116         int                     sc_sendad_success;
 117 #define CARP_SENDAD_MIN_SUCCESS 3
 118
 119         int                     sc_init_counter;
 120         uint64_t                sc_counter;
 121
 122         /* authentication */
 123 #define CARP_HMAC_PAD   64
 124         unsigned char sc_key[CARP_KEY_LEN];
 125         unsigned char sc_pad[CARP_HMAC_PAD];
 126         SHA1_CTX sc_sha1;
 127
 128         TAILQ_ENTRY(carp_softc) sc_list;        /* On the carp_if list. */
 129         LIST_ENTRY(carp_softc)  sc_next;        /* On the global list. */
 130 };
 131
 132 struct carp_if {
 133 #ifdef INET
 134         int     cif_naddrs;
 135 #endif
 136 #ifdef INET6
 137         int     cif_naddrs6;
 138 #endif
 139         TAILQ_HEAD(, carp_softc) cif_vrs;
 140 #ifdef INET
 141         struct ip_moptions       cif_imo;
 142 #endif
 143 #ifdef INET6
 144         struct ip6_moptions      cif_im6o;
 145 #endif
 146         struct ifnet    *cif_ifp;
 147         struct mtx      cif_mtx;
 148         uint32_t        cif_flags;
 149 #define CIF_PROMISC     0x00000001
 150 };
 151
 152 #define CARP_INET       0
 153 #define CARP_INET6      1
 154 static int proto_reg[] = {-1, -1};
 155
 156 /*
 157  * Brief design of carp(4).
 158  *
 159  * Any carp-capable ifnet may have a list of carp softcs hanging off
 160  * its ifp->if_carp pointer. Each softc represents one unique virtual
 161  * host id, or vhid. The softc has a back pointer to the ifnet. All
 162  * softcs are joined in a global list, which has quite limited use.
 163  *
 164  * Any interface address that takes part in CARP negotiation has a
 165  * pointer to the softc of its vhid, ifa->ifa_carp. That could be either
 166  * AF_INET or AF_INET6 address.
 167  *
 168  * Although, one can get the softc's backpointer to ifnet and traverse
 169  * through its ifp->if_addrhead queue to find all interface addresses
 170  * involved in CARP, we keep a growable array of ifaddr pointers. This
 171  * allows us to avoid grabbing the IF_ADDR_LOCK() in many traversals that
 172  * do calls into the network stack, thus avoiding LORs.
 173  *
 174  * Locking:
 175  *
 176  * Each softc has a lock sc_mtx. It is used to synchronise carp_input_c(),
 177  * callout-driven events and ioctl()s.
 178  *
 179  * To traverse the list of softcs on an ifnet we use CIF_LOCK(), to
 180  * traverse the global list we use the mutex carp_mtx.
 181  *
 182  * Known issues with locking:
 183  *
 184  * - There is no protection for races between two ioctl() requests,
 185  *   neither SIOCSVH, nor SIOCAIFADDR & SIOCAIFADDR_IN6. I think that all
 186  *   interface ioctl()s should be serialized right in net/if.c.
 187  * - Sending ad, we put the pointer to the softc in an mtag, and no reference
 188  *   counting is done on the softc.
 189  * - On module unload we may race (?) with packet processing thread
 190  *   dereferencing our function pointers.
 191  */
 192
 193 /* Accept incoming CARP packets. */
 194 static VNET_DEFINE(int, carp_allow) = 1;
 195 #define V_carp_allow    VNET(carp_allow)
 196
 197 /* Preempt slower nodes. */
 198 static VNET_DEFINE(int, carp_preempt) = 0;
 199 #define V_carp_preempt  VNET(carp_preempt)
 200
 201 /* Log level. */
 202 static VNET_DEFINE(int, carp_log) = 1;
 203 #define V_carp_log      VNET(carp_log)
 204
 205 /* Global advskew demotion. */
 206 static VNET_DEFINE(int, carp_demotion) = 0;
 207 #define V_carp_demotion VNET(carp_demotion)
 208
 209 /* Send error demotion factor. */
 210 static VNET_DEFINE(int, carp_senderr_adj) = CARP_MAXSKEW;
 211 #define V_carp_senderr_adj      VNET(carp_senderr_adj)
 212
 213 /* Iface down demotion factor. */
 214 static VNET_DEFINE(int, carp_ifdown_adj) = CARP_MAXSKEW;
 215 #define V_carp_ifdown_adj       VNET(carp_ifdown_adj)
 216
 217 static int carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS);
 218
 219 SYSCTL_NODE(_net_inet, IPPROTO_CARP,    carp,   CTLFLAG_RW, 0,  "CARP");
 220 SYSCTL_VNET_INT(_net_inet_carp, OID_AUTO, allow, CTLFLAG_RW,
 221     &VNET_NAME(carp_allow), 0, "Accept incoming CARP packets");
 222 SYSCTL_VNET_INT(_net_inet_carp, OID_AUTO, preempt, CTLFLAG_RW,
 223     &VNET_NAME(carp_preempt), 0, "High-priority backup preemption mode");
 224 SYSCTL_VNET_INT(_net_inet_carp, OID_AUTO, log, CTLFLAG_RW,
 225     &VNET_NAME(carp_log), 0, "CARP log level");
 226 SYSCTL_VNET_PROC(_net_inet_carp, OID_AUTO, demotion, CTLTYPE_INT|CTLFLAG_RW,
 227     0, 0, carp_demote_adj_sysctl, "I",
 228     "Adjust demotion factor (skew of advskew)");
 229 SYSCTL_VNET_INT(_net_inet_carp, OID_AUTO, senderr_demotion_factor, CTLFLAG_RW,
 230     &VNET_NAME(carp_senderr_adj), 0, "Send error demotion factor adjustment");
 231 SYSCTL_VNET_INT(_net_inet_carp, OID_AUTO, ifdown_demotion_factor, CTLFLAG_RW,
 232     &VNET_NAME(carp_ifdown_adj), 0,
 233     "Interface down demotion factor adjustment");
 234
 235 VNET_PCPUSTAT_DEFINE(struct carpstats, carpstats);
 236 VNET_PCPUSTAT_SYSINIT(carpstats);
 237 VNET_PCPUSTAT_SYSUNINIT(carpstats);
 238
 239 #define CARPSTATS_ADD(name, val)        \
 240     counter_u64_add(VNET(carpstats)[offsetof(struct carpstats, name) / \
 241         sizeof(uint64_t)], (val))
 242 #define CARPSTATS_INC(name)             CARPSTATS_ADD(name, 1)
 243
 244 SYSCTL_VNET_PCPUSTAT(_net_inet_carp, OID_AUTO, stats, struct carpstats,
 245     carpstats, "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 246
 247 #define CARP_LOCK_INIT(sc)      mtx_init(&(sc)->sc_mtx, "carp_softc",   \
 248         NULL, MTX_DEF)
 249 #define CARP_LOCK_DESTROY(sc)   mtx_destroy(&(sc)->sc_mtx)
 250 #define CARP_LOCK_ASSERT(sc)    mtx_assert(&(sc)->sc_mtx, MA_OWNED)
 251 #define CARP_LOCK(sc)           mtx_lock(&(sc)->sc_mtx)
 252 #define CARP_UNLOCK(sc)         mtx_unlock(&(sc)->sc_mtx)
 253 #define CIF_LOCK_INIT(cif)      mtx_init(&(cif)->cif_mtx, "carp_if",   \
 254         NULL, MTX_DEF)
 255 #define CIF_LOCK_DESTROY(cif)   mtx_destroy(&(cif)->cif_mtx)
 256 #define CIF_LOCK_ASSERT(cif)    mtx_assert(&(cif)->cif_mtx, MA_OWNED)
 257 #define CIF_LOCK(cif)           mtx_lock(&(cif)->cif_mtx)
 258 #define CIF_UNLOCK(cif)         mtx_unlock(&(cif)->cif_mtx)
 259 #define CIF_FREE(cif)   do {                            \
 260                 CIF_LOCK_ASSERT(cif);                   \
 261                 if (TAILQ_EMPTY(&(cif)->cif_vrs))       \
 262                         carp_free_if(cif);              \
 263                 else                                    \
 264                         CIF_UNLOCK(cif);                \
 265 } while (0)
 266
 267 #define CARP_LOG(...)   do {                            \
 268         if (V_carp_log > 0)                             \
 269                 log(LOG_INFO, "carp: " __VA_ARGS__);    \
 270 } while (0)
 271
 272 #define CARP_DEBUG(...) do {                            \
 273         if (V_carp_log > 1)                             \
 274                 log(LOG_DEBUG, __VA_ARGS__);            \
 275 } while (0)
 276
 277 #define IFNET_FOREACH_IFA(ifp, ifa)                                     \
 278         IF_ADDR_LOCK_ASSERT(ifp);                                       \
 279         TAILQ_FOREACH((ifa), &(ifp)->if_addrhead, ifa_link)             \
 280                 if ((ifa)->ifa_carp != NULL)
 281
 282 #define CARP_FOREACH_IFA(sc, ifa)                                       \
 283         CARP_LOCK_ASSERT(sc);                                           \
 284         for (int _i = 0;                                                \
 285                 _i < (sc)->sc_naddrs + (sc)->sc_naddrs6 &&              \
 286                 ((ifa) = sc->sc_ifas[_i]) != NULL;                      \
 287                 ++_i)
 288
 289 #define IFNET_FOREACH_CARP(ifp, sc)                                     \
 290         CIF_LOCK_ASSERT(ifp->if_carp);                                  \
 291         TAILQ_FOREACH((sc), &(ifp)->if_carp->cif_vrs, sc_list)
 292
 293 #define DEMOTE_ADVSKEW(sc)                                      \
 294     (((sc)->sc_advskew + V_carp_demotion > CARP_MAXSKEW) ?      \
 295     CARP_MAXSKEW : ((sc)->sc_advskew + V_carp_demotion))
 296
 297 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 298 static struct carp_softc
 299                 *carp_alloc(struct ifnet *);
 300 static void     carp_detach_locked(struct ifaddr *);
 301 static void     carp_destroy(struct carp_softc *);
 302 static struct carp_if
 303                 *carp_alloc_if(struct ifnet *);
 304 static void     carp_free_if(struct carp_if *);
 305 static void     carp_set_state(struct carp_softc *, int);
 306 static void     carp_sc_state(struct carp_softc *);
 307 static void     carp_setrun(struct carp_softc *, sa_family_t);
 308 static void     carp_master_down(void *);
 309 static void     carp_master_down_locked(struct carp_softc *);
 310 static void     carp_send_ad(void *);
 311 static void     carp_send_ad_locked(struct carp_softc *);
 312 static void     carp_addroute(struct carp_softc *);
 313 static void     carp_ifa_addroute(struct ifaddr *);
 314 static void     carp_delroute(struct carp_softc *);
 315 static void     carp_ifa_delroute(struct ifaddr *);
 316 static void     carp_send_ad_all(void *, int);
 317 static void     carp_demote_adj(int, char *);
 318
 319 static LIST_HEAD(, carp_softc) carp_list;
 320 static struct mtx carp_mtx;
 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 void
 439 carp_input(struct mbuf *m, int hlen)
 440 {
 441         struct ip *ip = mtod(m, struct ip *);
 442         struct carp_header *ch;
 443         int iplen, len;
 444
 445         CARPSTATS_INC(carps_ipackets);
 446
 447         if (!V_carp_allow) {
 448                 m_freem(m);
 449                 return;
 450         }
 451
 452         /* verify that the IP TTL is 255.  */
 453         if (ip->ip_ttl != CARP_DFLTTL) {
 454                 CARPSTATS_INC(carps_badttl);
 455                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 456                     ip->ip_ttl,
 457                     m->m_pkthdr.rcvif->if_xname);
 458                 m_freem(m);
 459                 return;
 460         }
 461
 462         iplen = ip->ip_hl << 2;
 463
 464         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 465                 CARPSTATS_INC(carps_badlen);
 466                 CARP_DEBUG("%s: received len %zd < sizeof(struct carp_header) "
 467                     "on %s\n", __func__, m->m_len - sizeof(struct ip),
 468                     m->m_pkthdr.rcvif->if_xname);
 469                 m_freem(m);
 470                 return;
 471         }
 472
 473         if (iplen + sizeof(*ch) < m->m_len) {
 474                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 475                         CARPSTATS_INC(carps_hdrops);
 476                         CARP_DEBUG("%s: pullup failed\n", __func__);
 477                         return;
 478                 }
 479                 ip = mtod(m, struct ip *);
 480         }
 481         ch = (struct carp_header *)((char *)ip + iplen);
 482
 483         /*
 484          * verify that the received packet length is
 485          * equal to the CARP header
 486          */
 487         len = iplen + sizeof(*ch);
 488         if (len > m->m_pkthdr.len) {
 489                 CARPSTATS_INC(carps_badlen);
 490                 CARP_DEBUG("%s: packet too short %d on %s\n", __func__,
 491                     m->m_pkthdr.len,
 492                     m->m_pkthdr.rcvif->if_xname);
 493                 m_freem(m);
 494                 return;
 495         }
 496
 497         if ((m = m_pullup(m, len)) == NULL) {
 498                 CARPSTATS_INC(carps_hdrops);
 499                 return;
 500         }
 501         ip = mtod(m, struct ip *);
 502         ch = (struct carp_header *)((char *)ip + iplen);
 503
 504         /* verify the CARP checksum */
 505         m->m_data += iplen;
 506         if (in_cksum(m, len - iplen)) {
 507                 CARPSTATS_INC(carps_badsum);
 508                 CARP_DEBUG("%s: checksum failed on %s\n", __func__,
 509                     m->m_pkthdr.rcvif->if_xname);
 510                 m_freem(m);
 511                 return;
 512         }
 513         m->m_data -= iplen;
 514
 515         carp_input_c(m, ch, AF_INET);
 516 }
 517 #endif
 518
 519 #ifdef INET6
 520 int
 521 carp6_input(struct mbuf **mp, int *offp, int proto)
 522 {
 523         struct mbuf *m = *mp;
 524         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 525         struct carp_header *ch;
 526         u_int len;
 527
 528         CARPSTATS_INC(carps_ipackets6);
 529
 530         if (!V_carp_allow) {
 531                 m_freem(m);
 532                 return (IPPROTO_DONE);
 533         }
 534
 535         /* check if received on a valid carp interface */
 536         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 537                 CARPSTATS_INC(carps_badif);
 538                 CARP_DEBUG("%s: packet received on non-carp interface: %s\n",
 539                     __func__, m->m_pkthdr.rcvif->if_xname);
 540                 m_freem(m);
 541                 return (IPPROTO_DONE);
 542         }
 543
 544         /* verify that the IP TTL is 255 */
 545         if (ip6->ip6_hlim != CARP_DFLTTL) {
 546                 CARPSTATS_INC(carps_badttl);
 547                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 548                     ip6->ip6_hlim, m->m_pkthdr.rcvif->if_xname);
 549                 m_freem(m);
 550                 return (IPPROTO_DONE);
 551         }
 552
 553         /* verify that we have a complete carp packet */
 554         len = m->m_len;
 555         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 556         if (ch == NULL) {
 557                 CARPSTATS_INC(carps_badlen);
 558                 CARP_DEBUG("%s: packet size %u too small\n", __func__, len);
 559                 return (IPPROTO_DONE);
 560         }
 561
 562
 563         /* verify the CARP checksum */
 564         m->m_data += *offp;
 565         if (in_cksum(m, sizeof(*ch))) {
 566                 CARPSTATS_INC(carps_badsum);
 567                 CARP_DEBUG("%s: checksum failed, on %s\n", __func__,
 568                     m->m_pkthdr.rcvif->if_xname);
 569                 m_freem(m);
 570                 return (IPPROTO_DONE);
 571         }
 572         m->m_data -= *offp;
 573
 574         carp_input_c(m, ch, AF_INET6);
 575         return (IPPROTO_DONE);
 576 }
 577 #endif /* INET6 */
 578
 579 static void
 580 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 581 {
 582         struct ifnet *ifp = m->m_pkthdr.rcvif;
 583         struct ifaddr *ifa;
 584         struct carp_softc *sc;
 585         uint64_t tmp_counter;
 586         struct timeval sc_tv, ch_tv;
 587
 588         /* verify that the VHID is valid on the receiving interface */
 589         IF_ADDR_RLOCK(ifp);
 590         IFNET_FOREACH_IFA(ifp, ifa)
 591                 if (ifa->ifa_addr->sa_family == af &&
 592                     ifa->ifa_carp->sc_vhid == ch->carp_vhid) {
 593                         ifa_ref(ifa);
 594                         break;
 595                 }
 596         IF_ADDR_RUNLOCK(ifp);
 597
 598         if (ifa == NULL) {
 599                 CARPSTATS_INC(carps_badvhid);
 600                 m_freem(m);
 601                 return;
 602         }
 603
 604         /* verify the CARP version. */
 605         if (ch->carp_version != CARP_VERSION) {
 606                 CARPSTATS_INC(carps_badver);
 607                 CARP_DEBUG("%s: invalid version %d\n", ifp->if_xname,
 608                     ch->carp_version);
 609                 ifa_free(ifa);
 610                 m_freem(m);
 611                 return;
 612         }
 613
 614         sc = ifa->ifa_carp;
 615         CARP_LOCK(sc);
 616         ifa_free(ifa);
 617
 618         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 619                 CARPSTATS_INC(carps_badauth);
 620                 CARP_DEBUG("%s: incorrect hash for VHID %u@%s\n", __func__,
 621                     sc->sc_vhid, ifp->if_xname);
 622                 goto out;
 623         }
 624
 625         tmp_counter = ntohl(ch->carp_counter[0]);
 626         tmp_counter = tmp_counter<<32;
 627         tmp_counter += ntohl(ch->carp_counter[1]);
 628
 629         /* XXX Replay protection goes here */
 630
 631         sc->sc_init_counter = 0;
 632         sc->sc_counter = tmp_counter;
 633
 634         sc_tv.tv_sec = sc->sc_advbase;
 635         sc_tv.tv_usec = DEMOTE_ADVSKEW(sc) * 1000000 / 256;
 636         ch_tv.tv_sec = ch->carp_advbase;
 637         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 638
 639         switch (sc->sc_state) {
 640         case INIT:
 641                 break;
 642         case MASTER:
 643                 /*
 644                  * If we receive an advertisement from a master who's going to
 645                  * be more frequent than us, go into BACKUP state.
 646                  */
 647                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 648                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 649                         callout_stop(&sc->sc_ad_tmo);
 650                         CARP_LOG("VHID %u@%s: MASTER -> BACKUP "
 651                             "(more frequent advertisement received)\n",
 652                             sc->sc_vhid,
 653                             sc->sc_carpdev->if_xname);
 654                         carp_set_state(sc, BACKUP);
 655                         carp_setrun(sc, 0);
 656                         carp_delroute(sc);
 657                 }
 658                 break;
 659         case BACKUP:
 660                 /*
 661                  * If we're pre-empting masters who advertise slower than us,
 662                  * and this one claims to be slower, treat him as down.
 663                  */
 664                 if (V_carp_preempt && timevalcmp(&sc_tv, &ch_tv, <)) {
 665                         CARP_LOG("VHID %u@%s: BACKUP -> MASTER "
 666                             "(preempting a slower master)\n",
 667                             sc->sc_vhid,
 668                             sc->sc_carpdev->if_xname);
 669                         carp_master_down_locked(sc);
 670                         break;
 671                 }
 672
 673                 /*
 674                  *  If the master is going to advertise at such a low frequency
 675                  *  that he's guaranteed to time out, we'd might as well just
 676                  *  treat him as timed out now.
 677                  */
 678                 sc_tv.tv_sec = sc->sc_advbase * 3;
 679                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 680                         CARP_LOG("VHID %u@%s: BACKUP -> MASTER "
 681                             "(master timed out)\n",
 682                             sc->sc_vhid,
 683                             sc->sc_carpdev->if_xname);
 684                         carp_master_down_locked(sc);
 685                         break;
 686                 }
 687
 688                 /*
 689                  * Otherwise, we reset the counter and wait for the next
 690                  * advertisement.
 691                  */
 692                 carp_setrun(sc, af);
 693                 break;
 694         }
 695
 696 out:
 697         CARP_UNLOCK(sc);
 698         m_freem(m);
 699 }
 700
 701 static int
 702 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 703 {
 704         struct m_tag *mtag;
 705
 706         if (sc->sc_init_counter) {
 707                 /* this could also be seconds since unix epoch */
 708                 sc->sc_counter = arc4random();
 709                 sc->sc_counter = sc->sc_counter << 32;
 710                 sc->sc_counter += arc4random();
 711         } else
 712                 sc->sc_counter++;
 713
 714         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 715         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 716
 717         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 718
 719         /* Tag packet for carp_output */
 720         if ((mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct carp_softc *),
 721             M_NOWAIT)) == NULL) {
 722                 m_freem(m);
 723                 CARPSTATS_INC(carps_onomem);
 724                 return (ENOMEM);
 725         }
 726         bcopy(&sc, mtag + 1, sizeof(sc));
 727         m_tag_prepend(m, mtag);
 728
 729         return (0);
 730 }
 731
 732 /*
 733  * To avoid LORs and possible recursions this function shouldn't
 734  * be called directly, but scheduled via taskqueue.
 735  */
 736 static void
 737 carp_send_ad_all(void *ctx __unused, int pending __unused)
 738 {
 739         struct carp_softc *sc;
 740
 741         mtx_lock(&carp_mtx);
 742         LIST_FOREACH(sc, &carp_list, sc_next)
 743                 if (sc->sc_state == MASTER) {
 744                         CARP_LOCK(sc);
 745                         CURVNET_SET(sc->sc_carpdev->if_vnet);
 746                         carp_send_ad_locked(sc);
 747                         CURVNET_RESTORE();
 748                         CARP_UNLOCK(sc);
 749                 }
 750         mtx_unlock(&carp_mtx);
 751 }
 752
 753 /* Send a periodic advertisement, executed in callout context. */
 754 static void
 755 carp_send_ad(void *v)
 756 {
 757         struct carp_softc *sc = v;
 758
 759         CARP_LOCK_ASSERT(sc);
 760         CURVNET_SET(sc->sc_carpdev->if_vnet);
 761         carp_send_ad_locked(sc);
 762         CURVNET_RESTORE();
 763         CARP_UNLOCK(sc);
 764 }
 765
 766 static void
 767 carp_send_ad_error(struct carp_softc *sc, int error)
 768 {
 769
 770         if (error) {
 771                 if (sc->sc_sendad_errors < INT_MAX)
 772                         sc->sc_sendad_errors++;
 773                 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 774                         static const char fmt[] = "send error %d on %s";
 775                         char msg[sizeof(fmt) + IFNAMSIZ];
 776
 777                         sprintf(msg, fmt, error, sc->sc_carpdev->if_xname);
 778                         carp_demote_adj(V_carp_senderr_adj, msg);
 779                 }
 780                 sc->sc_sendad_success = 0;
 781         } else {
 782                 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS &&
 783                     ++sc->sc_sendad_success >= CARP_SENDAD_MIN_SUCCESS) {
 784                         static const char fmt[] = "send ok on %s";
 785                         char msg[sizeof(fmt) + IFNAMSIZ];
 786
 787                         sprintf(msg, fmt, sc->sc_carpdev->if_xname);
 788                         carp_demote_adj(-V_carp_senderr_adj, msg);
 789                         sc->sc_sendad_errors = 0;
 790                 } else
 791                         sc->sc_sendad_errors = 0;
 792         }
 793 }
 794
 795 static void
 796 carp_send_ad_locked(struct carp_softc *sc)
 797 {
 798         struct carp_header ch;
 799         struct timeval tv;
 800         struct sockaddr sa;
 801         struct ifaddr *ifa;
 802         struct carp_header *ch_ptr;
 803         struct mbuf *m;
 804         int len, advskew;
 805
 806         CARP_LOCK_ASSERT(sc);
 807
 808         advskew = DEMOTE_ADVSKEW(sc);
 809         tv.tv_sec = sc->sc_advbase;
 810         tv.tv_usec = advskew * 1000000 / 256;
 811
 812         ch.carp_version = CARP_VERSION;
 813         ch.carp_type = CARP_ADVERTISEMENT;
 814         ch.carp_vhid = sc->sc_vhid;
 815         ch.carp_advbase = sc->sc_advbase;
 816         ch.carp_advskew = advskew;
 817         ch.carp_authlen = 7;    /* XXX DEFINE */
 818         ch.carp_pad1 = 0;       /* must be zero */
 819         ch.carp_cksum = 0;
 820
 821         /* XXXGL: OpenBSD picks first ifaddr with needed family. */
 822
 823 #ifdef INET
 824         if (sc->sc_naddrs) {
 825                 struct ip *ip;
 826
 827                 m = m_gethdr(M_NOWAIT, MT_DATA);
 828                 if (m == NULL) {
 829                         CARPSTATS_INC(carps_onomem);
 830                         goto resched;
 831                 }
 832                 len = sizeof(*ip) + sizeof(ch);
 833                 m->m_pkthdr.len = len;
 834                 m->m_pkthdr.rcvif = NULL;
 835                 m->m_len = len;
 836                 MH_ALIGN(m, m->m_len);
 837                 m->m_flags |= M_MCAST;
 838                 ip = mtod(m, struct ip *);
 839                 ip->ip_v = IPVERSION;
 840                 ip->ip_hl = sizeof(*ip) >> 2;
 841                 ip->ip_tos = IPTOS_LOWDELAY;
 842                 ip->ip_len = htons(len);
 843                 ip->ip_id = ip_newid();
 844                 ip->ip_off = htons(IP_DF);
 845                 ip->ip_ttl = CARP_DFLTTL;
 846                 ip->ip_p = IPPROTO_CARP;
 847                 ip->ip_sum = 0;
 848
 849                 bzero(&sa, sizeof(sa));
 850                 sa.sa_family = AF_INET;
 851                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 852                 if (ifa != NULL) {
 853                         ip->ip_src.s_addr =
 854                             ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 855                         ifa_free(ifa);
 856                 } else
 857                         ip->ip_src.s_addr = 0;
 858                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 859
 860                 ch_ptr = (struct carp_header *)(&ip[1]);
 861                 bcopy(&ch, ch_ptr, sizeof(ch));
 862                 if (carp_prepare_ad(m, sc, ch_ptr))
 863                         goto resched;
 864
 865                 m->m_data += sizeof(*ip);
 866                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip));
 867                 m->m_data -= sizeof(*ip);
 868
 869                 CARPSTATS_INC(carps_opackets);
 870
 871                 carp_send_ad_error(sc, ip_output(m, NULL, NULL, IP_RAWOUTPUT,
 872                     &sc->sc_carpdev->if_carp->cif_imo, NULL));
 873         }
 874 #endif /* INET */
 875 #ifdef INET6
 876         if (sc->sc_naddrs6) {
 877                 struct ip6_hdr *ip6;
 878
 879                 m = m_gethdr(M_NOWAIT, MT_DATA);
 880                 if (m == NULL) {
 881                         CARPSTATS_INC(carps_onomem);
 882                         goto resched;
 883                 }
 884                 len = sizeof(*ip6) + sizeof(ch);
 885                 m->m_pkthdr.len = len;
 886                 m->m_pkthdr.rcvif = NULL;
 887                 m->m_len = len;
 888                 MH_ALIGN(m, m->m_len);
 889                 m->m_flags |= M_MCAST;
 890                 ip6 = mtod(m, struct ip6_hdr *);
 891                 bzero(ip6, sizeof(*ip6));
 892                 ip6->ip6_vfc |= IPV6_VERSION;
 893                 ip6->ip6_hlim = CARP_DFLTTL;
 894                 ip6->ip6_nxt = IPPROTO_CARP;
 895                 bzero(&sa, sizeof(sa));
 896
 897                 /* set the source address */
 898                 sa.sa_family = AF_INET6;
 899                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 900                 if (ifa != NULL) {
 901                         bcopy(IFA_IN6(ifa), &ip6->ip6_src,
 902                             sizeof(struct in6_addr));
 903                         ifa_free(ifa);
 904                 } else
 905                         /* This should never happen with IPv6. */
 906                         bzero(&ip6->ip6_src, sizeof(struct in6_addr));
 907
 908                 /* Set the multicast destination. */
 909                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
 910                 ip6->ip6_dst.s6_addr8[15] = 0x12;
 911                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
 912                         m_freem(m);
 913                         CARP_DEBUG("%s: in6_setscope failed\n", __func__);
 914                         goto resched;
 915                 }
 916
 917                 ch_ptr = (struct carp_header *)(&ip6[1]);
 918                 bcopy(&ch, ch_ptr, sizeof(ch));
 919                 if (carp_prepare_ad(m, sc, ch_ptr))
 920                         goto resched;
 921
 922                 m->m_data += sizeof(*ip6);
 923                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip6));
 924                 m->m_data -= sizeof(*ip6);
 925
 926                 CARPSTATS_INC(carps_opackets6);
 927
 928                 carp_send_ad_error(sc, ip6_output(m, NULL, NULL, 0,
 929                     &sc->sc_carpdev->if_carp->cif_im6o, NULL, NULL));
 930         }
 931 #endif /* INET6 */
 932
 933 resched:
 934         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), carp_send_ad, sc);
 935 }
 936
 937 static void
 938 carp_addroute(struct carp_softc *sc)
 939 {
 940         struct ifaddr *ifa;
 941
 942         CARP_FOREACH_IFA(sc, ifa)
 943                 carp_ifa_addroute(ifa);
 944 }
 945
 946 static void
 947 carp_ifa_addroute(struct ifaddr *ifa)
 948 {
 949
 950         switch (ifa->ifa_addr->sa_family) {
 951 #ifdef INET
 952         case AF_INET:
 953                 in_addprefix(ifatoia(ifa), RTF_UP);
 954                 ifa_add_loopback_route(ifa,
 955                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 956                 break;
 957 #endif
 958 #ifdef INET6
 959         case AF_INET6:
 960                 ifa_add_loopback_route(ifa,
 961                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 962                 nd6_add_ifa_lle(ifatoia6(ifa));
 963                 break;
 964 #endif
 965         }
 966 }
 967
 968 static void
 969 carp_delroute(struct carp_softc *sc)
 970 {
 971         struct ifaddr *ifa;
 972
 973         CARP_FOREACH_IFA(sc, ifa)
 974                 carp_ifa_delroute(ifa);
 975 }
 976
 977 static void
 978 carp_ifa_delroute(struct ifaddr *ifa)
 979 {
 980
 981         switch (ifa->ifa_addr->sa_family) {
 982 #ifdef INET
 983         case AF_INET:
 984                 ifa_del_loopback_route(ifa,
 985                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 986                 in_scrubprefix(ifatoia(ifa), LLE_STATIC);
 987                 break;
 988 #endif
 989 #ifdef INET6
 990         case AF_INET6:
 991                 ifa_del_loopback_route(ifa,
 992                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 993                 nd6_rem_ifa_lle(ifatoia6(ifa));
 994                 break;
 995 #endif
 996         }
 997 }
 998
 999 int
1000 carp_master(struct ifaddr *ifa)
1001 {
1002         struct carp_softc *sc = ifa->ifa_carp;
1003
1004         return (sc->sc_state == MASTER);
1005 }
1006
1007 #ifdef INET
1008 /*
1009  * Broadcast a gratuitous ARP request containing
1010  * the virtual router MAC address for each IP address
1011  * associated with the virtual router.
1012  */
1013 static void
1014 carp_send_arp(struct carp_softc *sc)
1015 {
1016         struct ifaddr *ifa;
1017
1018         CARP_FOREACH_IFA(sc, ifa)
1019                 if (ifa->ifa_addr->sa_family == AF_INET)
1020                         arp_ifinit2(sc->sc_carpdev, ifa, LLADDR(&sc->sc_addr));
1021 }
1022
1023 int
1024 carp_iamatch(struct ifaddr *ifa, uint8_t **enaddr)
1025 {
1026         struct carp_softc *sc = ifa->ifa_carp;
1027
1028         if (sc->sc_state == MASTER) {
1029                 *enaddr = LLADDR(&sc->sc_addr);
1030                 return (1);
1031         }
1032
1033         return (0);
1034 }
1035 #endif
1036
1037 #ifdef INET6
1038 static void
1039 carp_send_na(struct carp_softc *sc)
1040 {
1041         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1042         struct ifaddr *ifa;
1043         struct in6_addr *in6;
1044
1045         CARP_FOREACH_IFA(sc, ifa) {
1046                 if (ifa->ifa_addr->sa_family != AF_INET6)
1047                         continue;
1048
1049                 in6 = IFA_IN6(ifa);
1050                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1051                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1052                 DELAY(1000);    /* XXX */
1053         }
1054 }
1055
1056 /*
1057  * Returns ifa in case it's a carp address and it is MASTER, or if the address
1058  * matches and is not a carp address.  Returns NULL otherwise.
1059  */
1060 struct ifaddr *
1061 carp_iamatch6(struct ifnet *ifp, struct in6_addr *taddr)
1062 {
1063         struct ifaddr *ifa;
1064
1065         ifa = NULL;
1066         IF_ADDR_RLOCK(ifp);
1067         TAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
1068                 if (ifa->ifa_addr->sa_family != AF_INET6)
1069                         continue;
1070                 if (!IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa)))
1071                         continue;
1072                 if (ifa->ifa_carp && ifa->ifa_carp->sc_state != MASTER)
1073                         ifa = NULL;
1074                 else
1075                         ifa_ref(ifa);
1076                 break;
1077         }
1078         IF_ADDR_RUNLOCK(ifp);
1079
1080         return (ifa);
1081 }
1082
1083 caddr_t
1084 carp_macmatch6(struct ifnet *ifp, struct mbuf *m, const struct in6_addr *taddr)
1085 {
1086         struct ifaddr *ifa;
1087
1088         IF_ADDR_RLOCK(ifp);
1089         IFNET_FOREACH_IFA(ifp, ifa)
1090                 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1091                     IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1092                         struct carp_softc *sc = ifa->ifa_carp;
1093                         struct m_tag *mtag;
1094
1095                         IF_ADDR_RUNLOCK(ifp);
1096
1097                         mtag = m_tag_get(PACKET_TAG_CARP,
1098                             sizeof(struct carp_softc *), M_NOWAIT);
1099                         if (mtag == NULL)
1100                                 /* Better a bit than nothing. */
1101                                 return (LLADDR(&sc->sc_addr));
1102
1103                         bcopy(&sc, mtag + 1, sizeof(sc));
1104                         m_tag_prepend(m, mtag);
1105
1106                         return (LLADDR(&sc->sc_addr));
1107                 }
1108         IF_ADDR_RUNLOCK(ifp);
1109
1110         return (NULL);
1111 }
1112 #endif /* INET6 */
1113
1114 int
1115 carp_forus(struct ifnet *ifp, u_char *dhost)
1116 {
1117         struct carp_softc *sc;
1118         uint8_t *ena = dhost;
1119
1120         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1121                 return (0);
1122
1123         CIF_LOCK(ifp->if_carp);
1124         IFNET_FOREACH_CARP(ifp, sc) {
1125                 CARP_LOCK(sc);
1126                 if (sc->sc_state == MASTER && !bcmp(dhost, LLADDR(&sc->sc_addr),
1127                     ETHER_ADDR_LEN)) {
1128                         CARP_UNLOCK(sc);
1129                         CIF_UNLOCK(ifp->if_carp);
1130                         return (1);
1131                 }
1132                 CARP_UNLOCK(sc);
1133         }
1134         CIF_UNLOCK(ifp->if_carp);
1135
1136         return (0);
1137 }
1138
1139 /* Master down timeout event, executed in callout context. */
1140 static void
1141 carp_master_down(void *v)
1142 {
1143         struct carp_softc *sc = v;
1144
1145         CARP_LOCK_ASSERT(sc);
1146
1147         CURVNET_SET(sc->sc_carpdev->if_vnet);
1148         if (sc->sc_state == BACKUP) {
1149                 CARP_LOG("VHID %u@%s: BACKUP -> MASTER (master down)\n",
1150                     sc->sc_vhid,
1151                     sc->sc_carpdev->if_xname);
1152                 carp_master_down_locked(sc);
1153         }
1154         CURVNET_RESTORE();
1155
1156         CARP_UNLOCK(sc);
1157 }
1158
1159 static void
1160 carp_master_down_locked(struct carp_softc *sc)
1161 {
1162
1163         CARP_LOCK_ASSERT(sc);
1164
1165         switch (sc->sc_state) {
1166         case BACKUP:
1167                 carp_set_state(sc, MASTER);
1168                 carp_send_ad_locked(sc);
1169 #ifdef INET
1170                 carp_send_arp(sc);
1171 #endif
1172 #ifdef INET6
1173                 carp_send_na(sc);
1174 #endif
1175                 carp_setrun(sc, 0);
1176                 carp_addroute(sc);
1177                 break;
1178         case INIT:
1179         case MASTER:
1180 #ifdef INVARIANTS
1181                 panic("carp: VHID %u@%s: master_down event in %s state\n",
1182                     sc->sc_vhid,
1183                     sc->sc_carpdev->if_xname,
1184                     sc->sc_state ? "MASTER" : "INIT");
1185 #endif
1186                 break;
1187         }
1188 }
1189
1190 /*
1191  * When in backup state, af indicates whether to reset the master down timer
1192  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1193  */
1194 static void
1195 carp_setrun(struct carp_softc *sc, sa_family_t af)
1196 {
1197         struct timeval tv;
1198
1199         CARP_LOCK_ASSERT(sc);
1200
1201         if ((sc->sc_carpdev->if_flags & IFF_UP) == 0 ||
1202             sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1203             (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0))
1204                 return;
1205
1206         switch (sc->sc_state) {
1207         case INIT:
1208                 CARP_LOG("VHID %u@%s: INIT -> BACKUP\n",
1209                     sc->sc_vhid,
1210                     sc->sc_carpdev->if_xname);
1211                 carp_set_state(sc, BACKUP);
1212                 carp_setrun(sc, 0);
1213                 break;
1214         case BACKUP:
1215                 callout_stop(&sc->sc_ad_tmo);
1216                 tv.tv_sec = 3 * sc->sc_advbase;
1217                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1218                 switch (af) {
1219 #ifdef INET
1220                 case AF_INET:
1221                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1222                             carp_master_down, sc);
1223                         break;
1224 #endif
1225 #ifdef INET6
1226                 case AF_INET6:
1227                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1228                             carp_master_down, sc);
1229                         break;
1230 #endif
1231                 default:
1232 #ifdef INET
1233                         if (sc->sc_naddrs)
1234                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1235                                     carp_master_down, sc);
1236 #endif
1237 #ifdef INET6
1238                         if (sc->sc_naddrs6)
1239                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1240                                     carp_master_down, sc);
1241 #endif
1242                         break;
1243                 }
1244                 break;
1245         case MASTER:
1246                 tv.tv_sec = sc->sc_advbase;
1247                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1248                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1249                     carp_send_ad, sc);
1250                 break;
1251         }
1252 }
1253
1254 /*
1255  * Setup multicast structures.
1256  */
1257 static int
1258 carp_multicast_setup(struct carp_if *cif, sa_family_t sa)
1259 {
1260         struct ifnet *ifp = cif->cif_ifp;
1261         int error = 0;
1262
1263         CIF_LOCK_ASSERT(cif);
1264
1265         switch (sa) {
1266 #ifdef INET
1267         case AF_INET:
1268             {
1269                 struct ip_moptions *imo = &cif->cif_imo;
1270                 struct in_addr addr;
1271
1272                 if (imo->imo_membership)
1273                         return (0);
1274
1275                 imo->imo_membership = (struct in_multi **)malloc(
1276                     (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
1277                     M_NOWAIT);
1278                 if (imo->imo_membership == NULL)
1279                         return (ENOMEM);
1280                 imo->imo_mfilters = NULL;
1281                 imo->imo_max_memberships = IP_MIN_MEMBERSHIPS;
1282                 imo->imo_multicast_vif = -1;
1283
1284                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1285                 if ((error = in_joingroup(ifp, &addr, NULL,
1286                     &imo->imo_membership[0])) != 0) {
1287                         free(imo->imo_membership, M_CARP);
1288                         break;
1289                 }
1290                 imo->imo_num_memberships++;
1291                 imo->imo_multicast_ifp = ifp;
1292                 imo->imo_multicast_ttl = CARP_DFLTTL;
1293                 imo->imo_multicast_loop = 0;
1294                 break;
1295            }
1296 #endif
1297 #ifdef INET6
1298         case AF_INET6:
1299             {
1300                 struct ip6_moptions *im6o = &cif->cif_im6o;
1301                 struct in6_addr in6;
1302                 struct in6_multi *in6m;
1303
1304                 if (im6o->im6o_membership)
1305                         return (0);
1306
1307                 im6o->im6o_membership = (struct in6_multi **)malloc(
1308                     (sizeof(struct in6_multi *) * IPV6_MIN_MEMBERSHIPS), M_CARP,
1309                     M_ZERO | M_NOWAIT);
1310                 if (im6o->im6o_membership == NULL)
1311                         return (ENOMEM);
1312                 im6o->im6o_mfilters = NULL;
1313                 im6o->im6o_max_memberships = IPV6_MIN_MEMBERSHIPS;
1314                 im6o->im6o_multicast_hlim = CARP_DFLTTL;
1315                 im6o->im6o_multicast_ifp = ifp;
1316
1317                 /* Join IPv6 CARP multicast group. */
1318                 bzero(&in6, sizeof(in6));
1319                 in6.s6_addr16[0] = htons(0xff02);
1320                 in6.s6_addr8[15] = 0x12;
1321                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1322                         free(im6o->im6o_membership, M_CARP);
1323                         break;
1324                 }
1325                 in6m = NULL;
1326                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1327                         free(im6o->im6o_membership, M_CARP);
1328                         break;
1329                 }
1330                 im6o->im6o_membership[0] = in6m;
1331                 im6o->im6o_num_memberships++;
1332
1333                 /* Join solicited multicast address. */
1334                 bzero(&in6, sizeof(in6));
1335                 in6.s6_addr16[0] = htons(0xff02);
1336                 in6.s6_addr32[1] = 0;
1337                 in6.s6_addr32[2] = htonl(1);
1338                 in6.s6_addr32[3] = 0;
1339                 in6.s6_addr8[12] = 0xff;
1340                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1341                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1342                         free(im6o->im6o_membership, M_CARP);
1343                         break;
1344                 }
1345                 in6m = NULL;
1346                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1347                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1348                         free(im6o->im6o_membership, M_CARP);
1349                         break;
1350                 }
1351                 im6o->im6o_membership[1] = in6m;
1352                 im6o->im6o_num_memberships++;
1353                 break;
1354             }
1355 #endif
1356         }
1357
1358         return (error);
1359 }
1360
1361 /*
1362  * Free multicast structures.
1363  */
1364 static void
1365 carp_multicast_cleanup(struct carp_if *cif, sa_family_t sa)
1366 {
1367
1368         CIF_LOCK_ASSERT(cif);
1369         switch (sa) {
1370 #ifdef INET
1371         case AF_INET:
1372                 if (cif->cif_naddrs == 0) {
1373                         struct ip_moptions *imo = &cif->cif_imo;
1374
1375                         in_leavegroup(imo->imo_membership[0], NULL);
1376                         KASSERT(imo->imo_mfilters == NULL,
1377                             ("%s: imo_mfilters != NULL", __func__));
1378                         free(imo->imo_membership, M_CARP);
1379                         imo->imo_membership = NULL;
1380
1381                 }
1382                 break;
1383 #endif
1384 #ifdef INET6
1385         case AF_INET6:
1386                 if (cif->cif_naddrs6 == 0) {
1387                         struct ip6_moptions *im6o = &cif->cif_im6o;
1388
1389                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1390                         in6_mc_leave(im6o->im6o_membership[1], NULL);
1391                         KASSERT(im6o->im6o_mfilters == NULL,
1392                             ("%s: im6o_mfilters != NULL", __func__));
1393                         free(im6o->im6o_membership, M_CARP);
1394                         im6o->im6o_membership = NULL;
1395                 }
1396                 break;
1397 #endif
1398         }
1399 }
1400
1401 int
1402 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa)
1403 {
1404         struct m_tag *mtag;
1405         struct carp_softc *sc;
1406
1407         if (!sa)
1408                 return (0);
1409
1410         switch (sa->sa_family) {
1411 #ifdef INET
1412         case AF_INET:
1413                 break;
1414 #endif
1415 #ifdef INET6
1416         case AF_INET6:
1417                 break;
1418 #endif
1419         default:
1420                 return (0);
1421         }
1422
1423         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
1424         if (mtag == NULL)
1425                 return (0);
1426
1427         bcopy(mtag + 1, &sc, sizeof(sc));
1428
1429         /* Set the source MAC address to the Virtual Router MAC Address. */
1430         switch (ifp->if_type) {
1431         case IFT_ETHER:
1432         case IFT_BRIDGE:
1433         case IFT_L2VLAN: {
1434                         struct ether_header *eh;
1435
1436                         eh = mtod(m, struct ether_header *);
1437                         eh->ether_shost[0] = 0;
1438                         eh->ether_shost[1] = 0;
1439                         eh->ether_shost[2] = 0x5e;
1440                         eh->ether_shost[3] = 0;
1441                         eh->ether_shost[4] = 1;
1442                         eh->ether_shost[5] = sc->sc_vhid;
1443                 }
1444                 break;
1445         case IFT_FDDI: {
1446                         struct fddi_header *fh;
1447
1448                         fh = mtod(m, struct fddi_header *);
1449                         fh->fddi_shost[0] = 0;
1450                         fh->fddi_shost[1] = 0;
1451                         fh->fddi_shost[2] = 0x5e;
1452                         fh->fddi_shost[3] = 0;
1453                         fh->fddi_shost[4] = 1;
1454                         fh->fddi_shost[5] = sc->sc_vhid;
1455                 }
1456                 break;
1457         case IFT_ISO88025: {
1458                         struct iso88025_header *th;
1459                         th = mtod(m, struct iso88025_header *);
1460                         th->iso88025_shost[0] = 3;
1461                         th->iso88025_shost[1] = 0;
1462                         th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
1463                         th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
1464                         th->iso88025_shost[4] = 0;
1465                         th->iso88025_shost[5] = 0;
1466                 }
1467                 break;
1468         default:
1469                 printf("%s: carp is not supported for the %d interface type\n",
1470                     ifp->if_xname, ifp->if_type);
1471                 return (EOPNOTSUPP);
1472         }
1473
1474         return (0);
1475 }
1476
1477 static struct carp_softc*
1478 carp_alloc(struct ifnet *ifp)
1479 {
1480         struct carp_softc *sc;
1481         struct carp_if *cif;
1482
1483         if ((cif = ifp->if_carp) == NULL)
1484                 cif = carp_alloc_if(ifp);
1485
1486         sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
1487
1488         sc->sc_advbase = CARP_DFLTINTV;
1489         sc->sc_vhid = -1;       /* required setting */
1490         sc->sc_init_counter = 1;
1491         sc->sc_state = INIT;
1492
1493         sc->sc_ifasiz = sizeof(struct ifaddr *);
1494         sc->sc_ifas = malloc(sc->sc_ifasiz, M_CARP, M_WAITOK|M_ZERO);
1495         sc->sc_carpdev = ifp;
1496
1497         CARP_LOCK_INIT(sc);
1498 #ifdef INET
1499         callout_init_mtx(&sc->sc_md_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1500 #endif
1501 #ifdef INET6
1502         callout_init_mtx(&sc->sc_md6_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1503 #endif
1504         callout_init_mtx(&sc->sc_ad_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1505
1506         CIF_LOCK(cif);
1507         TAILQ_INSERT_TAIL(&cif->cif_vrs, sc, sc_list);
1508         CIF_UNLOCK(cif);
1509
1510         mtx_lock(&carp_mtx);
1511         LIST_INSERT_HEAD(&carp_list, sc, sc_next);
1512         mtx_unlock(&carp_mtx);
1513
1514         return (sc);
1515 }
1516
1517 static int
1518 carp_grow_ifas(struct carp_softc *sc)
1519 {
1520         struct ifaddr **new;
1521
1522         CARP_LOCK_ASSERT(sc);
1523
1524         new = malloc(sc->sc_ifasiz * 2, M_CARP, M_NOWAIT|M_ZERO);
1525         if (new == NULL)
1526                 return (ENOMEM);
1527         bcopy(sc->sc_ifas, new, sc->sc_ifasiz);
1528         free(sc->sc_ifas, M_CARP);
1529         sc->sc_ifas = new;
1530         sc->sc_ifasiz *= 2;
1531
1532         return (0);
1533 }
1534
1535 static void
1536 carp_destroy(struct carp_softc *sc)
1537 {
1538         struct ifnet *ifp = sc->sc_carpdev;
1539         struct carp_if *cif = ifp->if_carp;
1540
1541         CIF_LOCK_ASSERT(cif);
1542
1543         TAILQ_REMOVE(&cif->cif_vrs, sc, sc_list);
1544
1545         mtx_lock(&carp_mtx);
1546         LIST_REMOVE(sc, sc_next);
1547         mtx_unlock(&carp_mtx);
1548
1549         CARP_LOCK(sc);
1550         if (sc->sc_suppress)
1551                 carp_demote_adj(-V_carp_ifdown_adj, "vhid removed");
1552         callout_drain(&sc->sc_ad_tmo);
1553 #ifdef INET
1554         callout_drain(&sc->sc_md_tmo);
1555 #endif
1556 #ifdef INET6
1557         callout_drain(&sc->sc_md6_tmo);
1558 #endif
1559         CARP_LOCK_DESTROY(sc);
1560
1561         free(sc->sc_ifas, M_CARP);
1562         free(sc, M_CARP);
1563 }
1564
1565 static struct carp_if*
1566 carp_alloc_if(struct ifnet *ifp)
1567 {
1568         struct carp_if *cif;
1569         int error;
1570
1571         cif = malloc(sizeof(*cif), M_CARP, M_WAITOK|M_ZERO);
1572
1573         if ((error = ifpromisc(ifp, 1)) != 0)
1574                 printf("%s: ifpromisc(%s) failed: %d\n",
1575                     __func__, ifp->if_xname, error);
1576         else
1577                 cif->cif_flags |= CIF_PROMISC;
1578
1579         CIF_LOCK_INIT(cif);
1580         cif->cif_ifp = ifp;
1581         TAILQ_INIT(&cif->cif_vrs);
1582
1583         IF_ADDR_WLOCK(ifp);
1584         ifp->if_carp = cif;
1585         if_ref(ifp);
1586         IF_ADDR_WUNLOCK(ifp);
1587
1588         return (cif);
1589 }
1590
1591 static void
1592 carp_free_if(struct carp_if *cif)
1593 {
1594         struct ifnet *ifp = cif->cif_ifp;
1595
1596         CIF_LOCK_ASSERT(cif);
1597         KASSERT(TAILQ_EMPTY(&cif->cif_vrs), ("%s: softc list not empty",
1598             __func__));
1599
1600         IF_ADDR_WLOCK(ifp);
1601         ifp->if_carp = NULL;
1602         IF_ADDR_WUNLOCK(ifp);
1603
1604         CIF_LOCK_DESTROY(cif);
1605
1606         if (cif->cif_flags & CIF_PROMISC)
1607                 ifpromisc(ifp, 0);
1608         if_rele(ifp);
1609
1610         free(cif, M_CARP);
1611 }
1612
1613 static void
1614 carp_carprcp(struct carpreq *carpr, struct carp_softc *sc, int priv)
1615 {
1616
1617         CARP_LOCK(sc);
1618         carpr->carpr_state = sc->sc_state;
1619         carpr->carpr_vhid = sc->sc_vhid;
1620         carpr->carpr_advbase = sc->sc_advbase;
1621         carpr->carpr_advskew = sc->sc_advskew;
1622         if (priv)
1623                 bcopy(sc->sc_key, carpr->carpr_key, sizeof(carpr->carpr_key));
1624         else
1625                 bzero(carpr->carpr_key, sizeof(carpr->carpr_key));
1626         CARP_UNLOCK(sc);
1627 }
1628
1629 int
1630 carp_ioctl(struct ifreq *ifr, u_long cmd, struct thread *td)
1631 {
1632         struct carpreq carpr;
1633         struct ifnet *ifp;
1634         struct carp_softc *sc = NULL;
1635         int error = 0, locked = 0;
1636
1637         if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1638                 return (error);
1639
1640         ifp = ifunit_ref(ifr->ifr_name);
1641         if (ifp == NULL)
1642                 return (ENXIO);
1643
1644         switch (ifp->if_type) {
1645         case IFT_ETHER:
1646         case IFT_L2VLAN:
1647         case IFT_BRIDGE:
1648         case IFT_FDDI:
1649         case IFT_ISO88025:
1650                 break;
1651         default:
1652                 error = EOPNOTSUPP;
1653                 goto out;
1654         }
1655
1656         if ((ifp->if_flags & IFF_MULTICAST) == 0) {
1657                 error = EADDRNOTAVAIL;
1658                 goto out;
1659         }
1660
1661         switch (cmd) {
1662         case SIOCSVH:
1663                 if ((error = priv_check(td, PRIV_NETINET_CARP)))
1664                         break;
1665                 if (carpr.carpr_vhid <= 0 || carpr.carpr_vhid > CARP_MAXVHID ||
1666                     carpr.carpr_advbase < 0 || carpr.carpr_advskew < 0) {
1667                         error = EINVAL;
1668                         break;
1669                 }
1670
1671                 if (ifp->if_carp) {
1672                         CIF_LOCK(ifp->if_carp);
1673                         IFNET_FOREACH_CARP(ifp, sc)
1674                                 if (sc->sc_vhid == carpr.carpr_vhid)
1675                                         break;
1676                         CIF_UNLOCK(ifp->if_carp);
1677                 }
1678                 if (sc == NULL) {
1679                         sc = carp_alloc(ifp);
1680                         CARP_LOCK(sc);
1681                         sc->sc_vhid = carpr.carpr_vhid;
1682                         LLADDR(&sc->sc_addr)[0] = 0;
1683                         LLADDR(&sc->sc_addr)[1] = 0;
1684                         LLADDR(&sc->sc_addr)[2] = 0x5e;
1685                         LLADDR(&sc->sc_addr)[3] = 0;
1686                         LLADDR(&sc->sc_addr)[4] = 1;
1687                         LLADDR(&sc->sc_addr)[5] = sc->sc_vhid;
1688                 } else
1689                         CARP_LOCK(sc);
1690                 locked = 1;
1691                 if (carpr.carpr_advbase > 0) {
1692                         if (carpr.carpr_advbase > 255 ||
1693                             carpr.carpr_advbase < CARP_DFLTINTV) {
1694                                 error = EINVAL;
1695                                 break;
1696                         }
1697                         sc->sc_advbase = carpr.carpr_advbase;
1698                 }
1699                 if (carpr.carpr_advskew > 0) {
1700                         if (carpr.carpr_advskew >= 255) {
1701                                 error = EINVAL;
1702                                 break;
1703                         }
1704                         sc->sc_advskew = carpr.carpr_advskew;
1705                 }
1706                 if (carpr.carpr_key[0] != '\0') {
1707                         bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1708                         carp_hmac_prepare(sc);
1709                 }
1710                 if (sc->sc_state != INIT &&
1711                     carpr.carpr_state != sc->sc_state) {
1712                         switch (carpr.carpr_state) {
1713                         case BACKUP:
1714                                 callout_stop(&sc->sc_ad_tmo);
1715                                 carp_set_state(sc, BACKUP);
1716                                 carp_setrun(sc, 0);
1717                                 carp_delroute(sc);
1718                                 break;
1719                         case MASTER:
1720                                 carp_master_down_locked(sc);
1721                                 break;
1722                         default:
1723                                 break;
1724                         }
1725                 }
1726                 break;
1727
1728         case SIOCGVH:
1729             {
1730                 int priveleged;
1731
1732                 if (carpr.carpr_vhid < 0 || carpr.carpr_vhid > CARP_MAXVHID) {
1733                         error = EINVAL;
1734                         break;
1735                 }
1736                 if (carpr.carpr_count < 1) {
1737                         error = EMSGSIZE;
1738                         break;
1739                 }
1740                 if (ifp->if_carp == NULL) {
1741                         error = ENOENT;
1742                         break;
1743                 }
1744
1745                 priveleged = (priv_check(td, PRIV_NETINET_CARP) == 0);
1746                 if (carpr.carpr_vhid != 0) {
1747                         CIF_LOCK(ifp->if_carp);
1748                         IFNET_FOREACH_CARP(ifp, sc)
1749                                 if (sc->sc_vhid == carpr.carpr_vhid)
1750                                         break;
1751                         CIF_UNLOCK(ifp->if_carp);
1752                         if (sc == NULL) {
1753                                 error = ENOENT;
1754                                 break;
1755                         }
1756                         carp_carprcp(&carpr, sc, priveleged);
1757                         error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1758                 } else  {
1759                         int i, count;
1760
1761                         count = 0;
1762                         CIF_LOCK(ifp->if_carp);
1763                         IFNET_FOREACH_CARP(ifp, sc)
1764                                 count++;
1765
1766                         if (count > carpr.carpr_count) {
1767                                 CIF_UNLOCK(ifp->if_carp);
1768                                 error = EMSGSIZE;
1769                                 break;
1770                         }
1771
1772                         i = 0;
1773                         IFNET_FOREACH_CARP(ifp, sc) {
1774                                 carp_carprcp(&carpr, sc, priveleged);
1775                                 carpr.carpr_count = count;
1776                                 error = copyout(&carpr, ifr->ifr_data +
1777                                     (i * sizeof(carpr)), sizeof(carpr));
1778                                 if (error) {
1779                                         CIF_UNLOCK(ifp->if_carp);
1780                                         break;
1781                                 }
1782                                 i++;
1783                         }
1784                         CIF_UNLOCK(ifp->if_carp);
1785                 }
1786                 break;
1787             }
1788         default:
1789                 error = EINVAL;
1790         }
1791
1792 out:
1793         if (locked)
1794                 CARP_UNLOCK(sc);
1795         if_rele(ifp);
1796
1797         return (error);
1798 }
1799
1800 static int
1801 carp_get_vhid(struct ifaddr *ifa)
1802 {
1803
1804         if (ifa == NULL || ifa->ifa_carp == NULL)
1805                 return (0);
1806
1807         return (ifa->ifa_carp->sc_vhid);
1808 }
1809
1810 int
1811 carp_attach(struct ifaddr *ifa, int vhid)
1812 {
1813         struct ifnet *ifp = ifa->ifa_ifp;
1814         struct carp_if *cif = ifp->if_carp;
1815         struct carp_softc *sc;
1816         int index, error;
1817
1818         if (ifp->if_carp == NULL)
1819                 return (ENOPROTOOPT);
1820
1821         switch (ifa->ifa_addr->sa_family) {
1822 #ifdef INET
1823         case AF_INET:
1824 #endif
1825 #ifdef INET6
1826         case AF_INET6:
1827 #endif
1828                 break;
1829         default:
1830                 return (EPROTOTYPE);
1831         }
1832
1833         CIF_LOCK(cif);
1834         IFNET_FOREACH_CARP(ifp, sc)
1835                 if (sc->sc_vhid == vhid)
1836                         break;
1837         if (sc == NULL) {
1838                 CIF_UNLOCK(cif);
1839                 return (ENOENT);
1840         }
1841
1842         if (ifa->ifa_carp) {
1843                 if (ifa->ifa_carp->sc_vhid != vhid)
1844                         carp_detach_locked(ifa);
1845                 else {
1846                         CIF_UNLOCK(cif);
1847                         return (0);
1848                 }
1849         }
1850
1851         error = carp_multicast_setup(cif, ifa->ifa_addr->sa_family);
1852         if (error) {
1853                 CIF_FREE(cif);
1854                 return (error);
1855         }
1856
1857         CARP_LOCK(sc);
1858         index = sc->sc_naddrs + sc->sc_naddrs6 + 1;
1859         if (index > sc->sc_ifasiz / sizeof(struct ifaddr *))
1860                 if ((error = carp_grow_ifas(sc)) != 0) {
1861                         carp_multicast_cleanup(cif,
1862                             ifa->ifa_addr->sa_family);
1863                         CARP_UNLOCK(sc);
1864                         CIF_FREE(cif);
1865                         return (error);
1866                 }
1867
1868         switch (ifa->ifa_addr->sa_family) {
1869 #ifdef INET
1870         case AF_INET:
1871                 cif->cif_naddrs++;
1872                 sc->sc_naddrs++;
1873                 break;
1874 #endif
1875 #ifdef INET6
1876         case AF_INET6:
1877                 cif->cif_naddrs6++;
1878                 sc->sc_naddrs6++;
1879                 break;
1880 #endif
1881         }
1882
1883         ifa_ref(ifa);
1884         sc->sc_ifas[index - 1] = ifa;
1885         ifa->ifa_carp = sc;
1886
1887         carp_hmac_prepare(sc);
1888         carp_sc_state(sc);
1889
1890         CARP_UNLOCK(sc);
1891         CIF_UNLOCK(cif);
1892
1893         return (0);
1894 }
1895
1896 void
1897 carp_detach(struct ifaddr *ifa)
1898 {
1899         struct ifnet *ifp = ifa->ifa_ifp;
1900         struct carp_if *cif = ifp->if_carp;
1901
1902         CIF_LOCK(cif);
1903         carp_detach_locked(ifa);
1904         CIF_FREE(cif);
1905 }
1906
1907 static void
1908 carp_detach_locked(struct ifaddr *ifa)
1909 {
1910         struct ifnet *ifp = ifa->ifa_ifp;
1911         struct carp_if *cif = ifp->if_carp;
1912         struct carp_softc *sc = ifa->ifa_carp;
1913         int i, index;
1914
1915         KASSERT(sc != NULL, ("%s: %p not attached", __func__, ifa));
1916
1917         CIF_LOCK_ASSERT(cif);
1918         CARP_LOCK(sc);
1919
1920         /* Shift array. */
1921         index = sc->sc_naddrs + sc->sc_naddrs6;
1922         for (i = 0; i < index; i++)
1923                 if (sc->sc_ifas[i] == ifa)
1924                         break;
1925         KASSERT(i < index, ("%s: %p no backref", __func__, ifa));
1926         for (; i < index - 1; i++)
1927                 sc->sc_ifas[i] = sc->sc_ifas[i+1];
1928         sc->sc_ifas[index - 1] = NULL;
1929
1930         switch (ifa->ifa_addr->sa_family) {
1931 #ifdef INET
1932         case AF_INET:
1933                 cif->cif_naddrs--;
1934                 sc->sc_naddrs--;
1935                 break;
1936 #endif
1937 #ifdef INET6
1938         case AF_INET6:
1939                 cif->cif_naddrs6--;
1940                 sc->sc_naddrs6--;
1941                 break;
1942 #endif
1943         }
1944
1945         carp_ifa_delroute(ifa);
1946         carp_multicast_cleanup(cif, ifa->ifa_addr->sa_family);
1947
1948         ifa->ifa_carp = NULL;
1949         ifa_free(ifa);
1950
1951         carp_hmac_prepare(sc);
1952         carp_sc_state(sc);
1953
1954         if (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) {
1955                 CARP_UNLOCK(sc);
1956                 carp_destroy(sc);
1957         } else
1958                 CARP_UNLOCK(sc);
1959 }
1960
1961 static void
1962 carp_set_state(struct carp_softc *sc, int state)
1963 {
1964
1965         CARP_LOCK_ASSERT(sc);
1966
1967         if (sc->sc_state != state) {
1968                 const char *carp_states[] = { CARP_STATES };
1969                 char subsys[IFNAMSIZ+5];
1970
1971                 sc->sc_state = state;
1972
1973                 snprintf(subsys, IFNAMSIZ+5, "%u@%s", sc->sc_vhid,
1974                     sc->sc_carpdev->if_xname);
1975                 devctl_notify("CARP", subsys, carp_states[state], NULL);
1976         }
1977 }
1978
1979 static void
1980 carp_linkstate(struct ifnet *ifp)
1981 {
1982         struct carp_softc *sc;
1983
1984         CIF_LOCK(ifp->if_carp);
1985         IFNET_FOREACH_CARP(ifp, sc) {
1986                 CARP_LOCK(sc);
1987                 carp_sc_state(sc);
1988                 CARP_UNLOCK(sc);
1989         }
1990         CIF_UNLOCK(ifp->if_carp);
1991 }
1992
1993 static void
1994 carp_sc_state(struct carp_softc *sc)
1995 {
1996
1997         CARP_LOCK_ASSERT(sc);
1998
1999         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
2000             !(sc->sc_carpdev->if_flags & IFF_UP)) {
2001                 callout_stop(&sc->sc_ad_tmo);
2002 #ifdef INET
2003                 callout_stop(&sc->sc_md_tmo);
2004 #endif
2005 #ifdef INET6
2006                 callout_stop(&sc->sc_md6_tmo);
2007 #endif
2008                 carp_set_state(sc, INIT);
2009                 carp_setrun(sc, 0);
2010                 if (!sc->sc_suppress)
2011                         carp_demote_adj(V_carp_ifdown_adj, "interface down");
2012                 sc->sc_suppress = 1;
2013         } else {
2014                 carp_set_state(sc, INIT);
2015                 carp_setrun(sc, 0);
2016                 if (sc->sc_suppress)
2017                         carp_demote_adj(-V_carp_ifdown_adj, "interface up");
2018                 sc->sc_suppress = 0;
2019         }
2020 }
2021
2022 static void
2023 carp_demote_adj(int adj, char *reason)
2024 {
2025         atomic_add_int(&V_carp_demotion, adj);
2026         CARP_LOG("demoted by %d to %d (%s)\n", adj, V_carp_demotion, reason);
2027         taskqueue_enqueue(taskqueue_swi, &carp_sendall_task);
2028 }
2029
2030 static int
2031 carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS)
2032 {
2033         int new, error;
2034
2035         new = V_carp_demotion;
2036         error = sysctl_handle_int(oidp, &new, 0, req);
2037         if (error || !req->newptr)
2038                 return (error);
2039
2040         carp_demote_adj(new, "sysctl");
2041
2042         return (0);
2043 }
2044
2045 #ifdef INET
2046 extern  struct domain inetdomain;
2047 static struct protosw in_carp_protosw = {
2048         .pr_type =              SOCK_RAW,
2049         .pr_domain =            &inetdomain,
2050         .pr_protocol =          IPPROTO_CARP,
2051         .pr_flags =             PR_ATOMIC|PR_ADDR,
2052         .pr_input =             carp_input,
2053         .pr_output =            (pr_output_t *)rip_output,
2054         .pr_ctloutput =         rip_ctloutput,
2055         .pr_usrreqs =           &rip_usrreqs
2056 };
2057 #endif
2058
2059 #ifdef INET6
2060 extern  struct domain inet6domain;
2061 static struct ip6protosw in6_carp_protosw = {
2062         .pr_type =              SOCK_RAW,
2063         .pr_domain =            &inet6domain,
2064         .pr_protocol =          IPPROTO_CARP,
2065         .pr_flags =             PR_ATOMIC|PR_ADDR,
2066         .pr_input =             carp6_input,
2067         .pr_output =            rip6_output,
2068         .pr_ctloutput =         rip6_ctloutput,
2069         .pr_usrreqs =           &rip6_usrreqs
2070 };
2071 #endif
2072
2073 static void
2074 carp_mod_cleanup(void)
2075 {
2076
2077 #ifdef INET
2078         if (proto_reg[CARP_INET] == 0) {
2079                 (void)ipproto_unregister(IPPROTO_CARP);
2080                 pf_proto_unregister(PF_INET, IPPROTO_CARP, SOCK_RAW);
2081                 proto_reg[CARP_INET] = -1;
2082         }
2083         carp_iamatch_p = NULL;
2084 #endif
2085 #ifdef INET6
2086         if (proto_reg[CARP_INET6] == 0) {
2087                 (void)ip6proto_unregister(IPPROTO_CARP);
2088                 pf_proto_unregister(PF_INET6, IPPROTO_CARP, SOCK_RAW);
2089                 proto_reg[CARP_INET6] = -1;
2090         }
2091         carp_iamatch6_p = NULL;
2092         carp_macmatch6_p = NULL;
2093 #endif
2094         carp_ioctl_p = NULL;
2095         carp_attach_p = NULL;
2096         carp_detach_p = NULL;
2097         carp_get_vhid_p = NULL;
2098         carp_linkstate_p = NULL;
2099         carp_forus_p = NULL;
2100         carp_output_p = NULL;
2101         carp_demote_adj_p = NULL;
2102         carp_master_p = NULL;
2103         mtx_unlock(&carp_mtx);
2104         taskqueue_drain(taskqueue_swi, &carp_sendall_task);
2105         mtx_destroy(&carp_mtx);
2106 }
2107
2108 static int
2109 carp_mod_load(void)
2110 {
2111         int err;
2112
2113         mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2114         LIST_INIT(&carp_list);
2115         carp_get_vhid_p = carp_get_vhid;
2116         carp_forus_p = carp_forus;
2117         carp_output_p = carp_output;
2118         carp_linkstate_p = carp_linkstate;
2119         carp_ioctl_p = carp_ioctl;
2120         carp_attach_p = carp_attach;
2121         carp_detach_p = carp_detach;
2122         carp_demote_adj_p = carp_demote_adj;
2123         carp_master_p = carp_master;
2124 #ifdef INET6
2125         carp_iamatch6_p = carp_iamatch6;
2126         carp_macmatch6_p = carp_macmatch6;
2127         proto_reg[CARP_INET6] = pf_proto_register(PF_INET6,
2128             (struct protosw *)&in6_carp_protosw);
2129         if (proto_reg[CARP_INET6]) {
2130                 printf("carp: error %d attaching to PF_INET6\n",
2131                     proto_reg[CARP_INET6]);
2132                 carp_mod_cleanup();
2133                 return (proto_reg[CARP_INET6]);
2134         }
2135         err = ip6proto_register(IPPROTO_CARP);
2136         if (err) {
2137                 printf("carp: error %d registering with INET6\n", err);
2138                 carp_mod_cleanup();
2139                 return (err);
2140         }
2141 #endif
2142 #ifdef INET
2143         carp_iamatch_p = carp_iamatch;
2144         proto_reg[CARP_INET] = pf_proto_register(PF_INET, &in_carp_protosw);
2145         if (proto_reg[CARP_INET]) {
2146                 printf("carp: error %d attaching to PF_INET\n",
2147                     proto_reg[CARP_INET]);
2148                 carp_mod_cleanup();
2149                 return (proto_reg[CARP_INET]);
2150         }
2151         err = ipproto_register(IPPROTO_CARP);
2152         if (err) {
2153                 printf("carp: error %d registering with INET\n", err);
2154                 carp_mod_cleanup();
2155                 return (err);
2156         }
2157 #endif
2158         return (0);
2159 }
2160
2161 static int
2162 carp_modevent(module_t mod, int type, void *data)
2163 {
2164         switch (type) {
2165         case MOD_LOAD:
2166                 return carp_mod_load();
2167                 /* NOTREACHED */
2168         case MOD_UNLOAD:
2169                 mtx_lock(&carp_mtx);
2170                 if (LIST_EMPTY(&carp_list))
2171                         carp_mod_cleanup();
2172                 else {
2173                         mtx_unlock(&carp_mtx);
2174                         return (EBUSY);
2175                 }
2176                 break;
2177
2178         default:
2179                 return (EINVAL);
2180         }
2181
2182         return (0);
2183 }
2184
2185 static moduledata_t carp_mod = {
2186         "carp",
2187         carp_modevent,
2188         0
2189 };
2190
2191 DECLARE_MODULE(carp, carp_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);