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_dl.h>
  59 #include <net/if_llatbl.h>
  60 #include <net/if_types.h>
  61 #include <net/iso88025.h>
  62 #include <net/route.h>
  63 #include <net/vnet.h>
  64
  65 #if defined(INET) || defined(INET6)
  66 #include <netinet/in.h>
  67 #include <netinet/in_var.h>
  68 #include <netinet/ip_carp.h>
  69 #include <netinet/ip.h>
  70 #include <machine/in_cksum.h>
  71 #endif
  72 #ifdef INET
  73 #include <netinet/ip_var.h>
  74 #include <netinet/if_ether.h>
  75 #endif
  76
  77 #ifdef INET6
  78 #include <netinet/icmp6.h>
  79 #include <netinet/ip6.h>
  80 #include <netinet6/ip6protosw.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 };
 148
 149 #define CARP_INET       0
 150 #define CARP_INET6      1
 151 static int proto_reg[] = {-1, -1};
 152
 153 /*
 154  * Brief design of carp(4).
 155  *
 156  * Any carp-capable ifnet may have a list of carp softcs hanging off
 157  * its ifp->if_carp pointer. Each softc represents one unique virtual
 158  * host id, or vhid. The softc has a back pointer to the ifnet. All
 159  * softcs are joined in a global list, which has quite limited use.
 160  *
 161  * Any interface address that takes part in CARP negotiation has a
 162  * pointer to the softc of its vhid, ifa->ifa_carp. That could be either
 163  * AF_INET or AF_INET6 address.
 164  *
 165  * Although, one can get the softc's backpointer to ifnet and traverse
 166  * through its ifp->if_addrhead queue to find all interface addresses
 167  * involved in CARP, we keep a growable array of ifaddr pointers. This
 168  * allows us to avoid grabbing the IF_ADDR_LOCK() in many traversals that
 169  * do calls into the network stack, thus avoiding LORs.
 170  *
 171  * Locking:
 172  *
 173  * Each softc has a lock sc_mtx. It is used to synchronise carp_input_c(),
 174  * callout-driven events and ioctl()s.
 175  *
 176  * To traverse the list of softcs on an ifnet we use CIF_LOCK(), to
 177  * traverse the global list we use the mutex carp_mtx.
 178  *
 179  * Known issues with locking:
 180  *
 181  * - There is no protection for races between two ioctl() requests,
 182  *   neither SIOCSVH, nor SIOCAIFADDR & SIOCAIFADDR_IN6. I think that all
 183  *   interface ioctl()s should be serialized right in net/if.c.
 184  * - Sending ad, we put the pointer to the softc in an mtag, and no reference
 185  *   counting is done on the softc.
 186  * - On module unload we may race (?) with packet processing thread
 187  *   dereferencing our function pointers.
 188  */
 189
 190 /* Accept incoming CARP packets. */
 191 static VNET_DEFINE(int, carp_allow) = 1;
 192 #define V_carp_allow    VNET(carp_allow)
 193
 194 /* Preempt slower nodes. */
 195 static VNET_DEFINE(int, carp_preempt) = 0;
 196 #define V_carp_preempt  VNET(carp_preempt)
 197
 198 /* Log level. */
 199 static VNET_DEFINE(int, carp_log) = 1;
 200 #define V_carp_log      VNET(carp_log)
 201
 202 /* Global advskew demotion. */
 203 static VNET_DEFINE(int, carp_demotion) = 0;
 204 #define V_carp_demotion VNET(carp_demotion)
 205
 206 /* Send error demotion factor. */
 207 static VNET_DEFINE(int, carp_senderr_adj) = CARP_MAXSKEW;
 208 #define V_carp_senderr_adj      VNET(carp_senderr_adj)
 209
 210 /* Iface down demotion factor. */
 211 static VNET_DEFINE(int, carp_ifdown_adj) = CARP_MAXSKEW;
 212 #define V_carp_ifdown_adj       VNET(carp_ifdown_adj)
 213
 214 static int carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS);
 215
 216 SYSCTL_NODE(_net_inet, IPPROTO_CARP,    carp,   CTLFLAG_RW, 0,  "CARP");
 217 SYSCTL_VNET_INT(_net_inet_carp, OID_AUTO, allow, CTLFLAG_RW,
 218     &VNET_NAME(carp_allow), 0, "Accept incoming CARP packets");
 219 SYSCTL_VNET_INT(_net_inet_carp, OID_AUTO, preempt, CTLFLAG_RW,
 220     &VNET_NAME(carp_preempt), 0, "High-priority backup preemption mode");
 221 SYSCTL_VNET_INT(_net_inet_carp, OID_AUTO, log, CTLFLAG_RW,
 222     &VNET_NAME(carp_log), 0, "CARP log level");
 223 SYSCTL_VNET_PROC(_net_inet_carp, OID_AUTO, demotion, CTLTYPE_INT|CTLFLAG_RW,
 224     0, 0, carp_demote_adj_sysctl, "I",
 225     "Adjust demotion factor (skew of advskew)");
 226 SYSCTL_VNET_INT(_net_inet_carp, OID_AUTO, senderr_demotion_factor, CTLFLAG_RW,
 227     &VNET_NAME(carp_senderr_adj), 0, "Send error demotion factor adjustment");
 228 SYSCTL_VNET_INT(_net_inet_carp, OID_AUTO, ifdown_demotion_factor, CTLFLAG_RW,
 229     &VNET_NAME(carp_ifdown_adj), 0,
 230     "Interface down demotion factor adjustment");
 231
 232 VNET_PCPUSTAT_DEFINE(struct carpstats, carpstats);
 233 VNET_PCPUSTAT_SYSINIT(carpstats);
 234 VNET_PCPUSTAT_SYSUNINIT(carpstats);
 235
 236 #define CARPSTATS_ADD(name, val)        \
 237     counter_u64_add(VNET(carpstats)[offsetof(struct carpstats, name) / \
 238         sizeof(uint64_t)], (val))
 239 #define CARPSTATS_INC(name)             CARPSTATS_ADD(name, 1)
 240
 241 SYSCTL_VNET_PCPUSTAT(_net_inet_carp, OID_AUTO, stats, struct carpstats,
 242     carpstats, "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 243
 244 #define CARP_LOCK_INIT(sc)      mtx_init(&(sc)->sc_mtx, "carp_softc",   \
 245         NULL, MTX_DEF)
 246 #define CARP_LOCK_DESTROY(sc)   mtx_destroy(&(sc)->sc_mtx)
 247 #define CARP_LOCK_ASSERT(sc)    mtx_assert(&(sc)->sc_mtx, MA_OWNED)
 248 #define CARP_LOCK(sc)           mtx_lock(&(sc)->sc_mtx)
 249 #define CARP_UNLOCK(sc)         mtx_unlock(&(sc)->sc_mtx)
 250 #define CIF_LOCK_INIT(cif)      mtx_init(&(cif)->cif_mtx, "carp_if",   \
 251         NULL, MTX_DEF)
 252 #define CIF_LOCK_DESTROY(cif)   mtx_destroy(&(cif)->cif_mtx)
 253 #define CIF_LOCK_ASSERT(cif)    mtx_assert(&(cif)->cif_mtx, MA_OWNED)
 254 #define CIF_LOCK(cif)           mtx_lock(&(cif)->cif_mtx)
 255 #define CIF_UNLOCK(cif)         mtx_unlock(&(cif)->cif_mtx)
 256 #define CIF_FREE(cif)   do {                            \
 257                 CIF_LOCK_ASSERT(cif);                   \
 258                 if (TAILQ_EMPTY(&(cif)->cif_vrs))       \
 259                         carp_free_if(cif);              \
 260                 else                                    \
 261                         CIF_UNLOCK(cif);                \
 262 } while (0)
 263
 264 #define CARP_LOG(...)   do {                            \
 265         if (V_carp_log > 0)                             \
 266                 log(LOG_INFO, "carp: " __VA_ARGS__);    \
 267 } while (0)
 268
 269 #define CARP_DEBUG(...) do {                            \
 270         if (V_carp_log > 1)                             \
 271                 log(LOG_DEBUG, __VA_ARGS__);            \
 272 } while (0)
 273
 274 #define IFNET_FOREACH_IFA(ifp, ifa)                                     \
 275         IF_ADDR_LOCK_ASSERT(ifp);                                       \
 276         TAILQ_FOREACH((ifa), &(ifp)->if_addrhead, ifa_link)             \
 277                 if ((ifa)->ifa_carp != NULL)
 278
 279 #define CARP_FOREACH_IFA(sc, ifa)                                       \
 280         CARP_LOCK_ASSERT(sc);                                           \
 281         for (int _i = 0;                                                \
 282                 _i < (sc)->sc_naddrs + (sc)->sc_naddrs6 &&              \
 283                 ((ifa) = sc->sc_ifas[_i]) != NULL;                      \
 284                 ++_i)
 285
 286 #define IFNET_FOREACH_CARP(ifp, sc)                                     \
 287         CIF_LOCK_ASSERT(ifp->if_carp);                                  \
 288         TAILQ_FOREACH((sc), &(ifp)->if_carp->cif_vrs, sc_list)
 289
 290 #define DEMOTE_ADVSKEW(sc)                                      \
 291     (((sc)->sc_advskew + V_carp_demotion > CARP_MAXSKEW) ?      \
 292     CARP_MAXSKEW : ((sc)->sc_advskew + V_carp_demotion))
 293
 294 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 295 static struct carp_softc
 296                 *carp_alloc(struct ifnet *);
 297 static void     carp_detach_locked(struct ifaddr *);
 298 static void     carp_destroy(struct carp_softc *);
 299 static struct carp_if
 300                 *carp_alloc_if(struct ifnet *);
 301 static void     carp_free_if(struct carp_if *);
 302 static void     carp_set_state(struct carp_softc *, int);
 303 static void     carp_sc_state(struct carp_softc *);
 304 static void     carp_setrun(struct carp_softc *, sa_family_t);
 305 static void     carp_master_down(void *);
 306 static void     carp_master_down_locked(struct carp_softc *);
 307 static void     carp_send_ad(void *);
 308 static void     carp_send_ad_locked(struct carp_softc *);
 309 static void     carp_addroute(struct carp_softc *);
 310 static void     carp_ifa_addroute(struct ifaddr *);
 311 static void     carp_delroute(struct carp_softc *);
 312 static void     carp_ifa_delroute(struct ifaddr *);
 313 static void     carp_send_ad_all(void *, int);
 314 static void     carp_demote_adj(int, char *);
 315
 316 static LIST_HEAD(, carp_softc) carp_list;
 317 static struct mtx carp_mtx;
 318 static struct task carp_sendall_task =
 319     TASK_INITIALIZER(0, carp_send_ad_all, NULL);
 320
 321 static void
 322 carp_hmac_prepare(struct carp_softc *sc)
 323 {
 324         uint8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 325         uint8_t vhid = sc->sc_vhid & 0xff;
 326         struct ifaddr *ifa;
 327         int i, found;
 328 #ifdef INET
 329         struct in_addr last, cur, in;
 330 #endif
 331 #ifdef INET6
 332         struct in6_addr last6, cur6, in6;
 333 #endif
 334
 335         CARP_LOCK_ASSERT(sc);
 336
 337         /* Compute ipad from key. */
 338         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 339         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 340         for (i = 0; i < sizeof(sc->sc_pad); i++)
 341                 sc->sc_pad[i] ^= 0x36;
 342
 343         /* Precompute first part of inner hash. */
 344         SHA1Init(&sc->sc_sha1);
 345         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 346         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 347         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 348         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 349 #ifdef INET
 350         cur.s_addr = 0;
 351         do {
 352                 found = 0;
 353                 last = cur;
 354                 cur.s_addr = 0xffffffff;
 355                 CARP_FOREACH_IFA(sc, ifa) {
 356                         in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 357                         if (ifa->ifa_addr->sa_family == AF_INET &&
 358                             ntohl(in.s_addr) > ntohl(last.s_addr) &&
 359                             ntohl(in.s_addr) < ntohl(cur.s_addr)) {
 360                                 cur.s_addr = in.s_addr;
 361                                 found++;
 362                         }
 363                 }
 364                 if (found)
 365                         SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur));
 366         } while (found);
 367 #endif /* INET */
 368 #ifdef INET6
 369         memset(&cur6, 0, sizeof(cur6));
 370         do {
 371                 found = 0;
 372                 last6 = cur6;
 373                 memset(&cur6, 0xff, sizeof(cur6));
 374                 CARP_FOREACH_IFA(sc, ifa) {
 375                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 376                         if (IN6_IS_SCOPE_EMBED(&in6))
 377                                 in6.s6_addr16[1] = 0;
 378                         if (ifa->ifa_addr->sa_family == AF_INET6 &&
 379                             memcmp(&in6, &last6, sizeof(in6)) > 0 &&
 380                             memcmp(&in6, &cur6, sizeof(in6)) < 0) {
 381                                 cur6 = in6;
 382                                 found++;
 383                         }
 384                 }
 385                 if (found)
 386                         SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6));
 387         } while (found);
 388 #endif /* INET6 */
 389
 390         /* convert ipad to opad */
 391         for (i = 0; i < sizeof(sc->sc_pad); i++)
 392                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 393 }
 394
 395 static void
 396 carp_hmac_generate(struct carp_softc *sc, uint32_t counter[2],
 397     unsigned char md[20])
 398 {
 399         SHA1_CTX sha1ctx;
 400
 401         CARP_LOCK_ASSERT(sc);
 402
 403         /* fetch first half of inner hash */
 404         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 405
 406         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 407         SHA1Final(md, &sha1ctx);
 408
 409         /* outer hash */
 410         SHA1Init(&sha1ctx);
 411         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 412         SHA1Update(&sha1ctx, md, 20);
 413         SHA1Final(md, &sha1ctx);
 414 }
 415
 416 static int
 417 carp_hmac_verify(struct carp_softc *sc, uint32_t counter[2],
 418     unsigned char md[20])
 419 {
 420         unsigned char md2[20];
 421
 422         CARP_LOCK_ASSERT(sc);
 423
 424         carp_hmac_generate(sc, counter, md2);
 425
 426         return (bcmp(md, md2, sizeof(md2)));
 427 }
 428
 429 /*
 430  * process input packet.
 431  * we have rearranged checks order compared to the rfc,
 432  * but it seems more efficient this way or not possible otherwise.
 433  */
 434 #ifdef INET
 435 void
 436 carp_input(struct mbuf *m, int hlen)
 437 {
 438         struct ip *ip = mtod(m, struct ip *);
 439         struct carp_header *ch;
 440         int iplen, len;
 441
 442         CARPSTATS_INC(carps_ipackets);
 443
 444         if (!V_carp_allow) {
 445                 m_freem(m);
 446                 return;
 447         }
 448
 449         /* verify that the IP TTL is 255.  */
 450         if (ip->ip_ttl != CARP_DFLTTL) {
 451                 CARPSTATS_INC(carps_badttl);
 452                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 453                     ip->ip_ttl,
 454                     m->m_pkthdr.rcvif->if_xname);
 455                 m_freem(m);
 456                 return;
 457         }
 458
 459         iplen = ip->ip_hl << 2;
 460
 461         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 462                 CARPSTATS_INC(carps_badlen);
 463                 CARP_DEBUG("%s: received len %zd < sizeof(struct carp_header) "
 464                     "on %s\n", __func__, m->m_len - sizeof(struct ip),
 465                     m->m_pkthdr.rcvif->if_xname);
 466                 m_freem(m);
 467                 return;
 468         }
 469
 470         if (iplen + sizeof(*ch) < m->m_len) {
 471                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 472                         CARPSTATS_INC(carps_hdrops);
 473                         CARP_DEBUG("%s: pullup failed\n", __func__);
 474                         return;
 475                 }
 476                 ip = mtod(m, struct ip *);
 477         }
 478         ch = (struct carp_header *)((char *)ip + iplen);
 479
 480         /*
 481          * verify that the received packet length is
 482          * equal to the CARP header
 483          */
 484         len = iplen + sizeof(*ch);
 485         if (len > m->m_pkthdr.len) {
 486                 CARPSTATS_INC(carps_badlen);
 487                 CARP_DEBUG("%s: packet too short %d on %s\n", __func__,
 488                     m->m_pkthdr.len,
 489                     m->m_pkthdr.rcvif->if_xname);
 490                 m_freem(m);
 491                 return;
 492         }
 493
 494         if ((m = m_pullup(m, len)) == NULL) {
 495                 CARPSTATS_INC(carps_hdrops);
 496                 return;
 497         }
 498         ip = mtod(m, struct ip *);
 499         ch = (struct carp_header *)((char *)ip + iplen);
 500
 501         /* verify the CARP checksum */
 502         m->m_data += iplen;
 503         if (in_cksum(m, len - iplen)) {
 504                 CARPSTATS_INC(carps_badsum);
 505                 CARP_DEBUG("%s: checksum failed on %s\n", __func__,
 506                     m->m_pkthdr.rcvif->if_xname);
 507                 m_freem(m);
 508                 return;
 509         }
 510         m->m_data -= iplen;
 511
 512         carp_input_c(m, ch, AF_INET);
 513 }
 514 #endif
 515
 516 #ifdef INET6
 517 int
 518 carp6_input(struct mbuf **mp, int *offp, int proto)
 519 {
 520         struct mbuf *m = *mp;
 521         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 522         struct carp_header *ch;
 523         u_int len;
 524
 525         CARPSTATS_INC(carps_ipackets6);
 526
 527         if (!V_carp_allow) {
 528                 m_freem(m);
 529                 return (IPPROTO_DONE);
 530         }
 531
 532         /* check if received on a valid carp interface */
 533         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 534                 CARPSTATS_INC(carps_badif);
 535                 CARP_DEBUG("%s: packet received on non-carp interface: %s\n",
 536                     __func__, m->m_pkthdr.rcvif->if_xname);
 537                 m_freem(m);
 538                 return (IPPROTO_DONE);
 539         }
 540
 541         /* verify that the IP TTL is 255 */
 542         if (ip6->ip6_hlim != CARP_DFLTTL) {
 543                 CARPSTATS_INC(carps_badttl);
 544                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 545                     ip6->ip6_hlim, m->m_pkthdr.rcvif->if_xname);
 546                 m_freem(m);
 547                 return (IPPROTO_DONE);
 548         }
 549
 550         /* verify that we have a complete carp packet */
 551         len = m->m_len;
 552         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 553         if (ch == NULL) {
 554                 CARPSTATS_INC(carps_badlen);
 555                 CARP_DEBUG("%s: packet size %u too small\n", __func__, len);
 556                 return (IPPROTO_DONE);
 557         }
 558
 559
 560         /* verify the CARP checksum */
 561         m->m_data += *offp;
 562         if (in_cksum(m, sizeof(*ch))) {
 563                 CARPSTATS_INC(carps_badsum);
 564                 CARP_DEBUG("%s: checksum failed, on %s\n", __func__,
 565                     m->m_pkthdr.rcvif->if_xname);
 566                 m_freem(m);
 567                 return (IPPROTO_DONE);
 568         }
 569         m->m_data -= *offp;
 570
 571         carp_input_c(m, ch, AF_INET6);
 572         return (IPPROTO_DONE);
 573 }
 574 #endif /* INET6 */
 575
 576 static void
 577 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 578 {
 579         struct ifnet *ifp = m->m_pkthdr.rcvif;
 580         struct ifaddr *ifa;
 581         struct carp_softc *sc;
 582         uint64_t tmp_counter;
 583         struct timeval sc_tv, ch_tv;
 584
 585         /* verify that the VHID is valid on the receiving interface */
 586         IF_ADDR_RLOCK(ifp);
 587         IFNET_FOREACH_IFA(ifp, ifa)
 588                 if (ifa->ifa_addr->sa_family == af &&
 589                     ifa->ifa_carp->sc_vhid == ch->carp_vhid) {
 590                         ifa_ref(ifa);
 591                         break;
 592                 }
 593         IF_ADDR_RUNLOCK(ifp);
 594
 595         if (ifa == NULL) {
 596                 CARPSTATS_INC(carps_badvhid);
 597                 m_freem(m);
 598                 return;
 599         }
 600
 601         /* verify the CARP version. */
 602         if (ch->carp_version != CARP_VERSION) {
 603                 CARPSTATS_INC(carps_badver);
 604                 CARP_DEBUG("%s: invalid version %d\n", ifp->if_xname,
 605                     ch->carp_version);
 606                 ifa_free(ifa);
 607                 m_freem(m);
 608                 return;
 609         }
 610
 611         sc = ifa->ifa_carp;
 612         CARP_LOCK(sc);
 613         ifa_free(ifa);
 614
 615         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 616                 CARPSTATS_INC(carps_badauth);
 617                 CARP_DEBUG("%s: incorrect hash for VHID %u@%s\n", __func__,
 618                     sc->sc_vhid, ifp->if_xname);
 619                 goto out;
 620         }
 621
 622         tmp_counter = ntohl(ch->carp_counter[0]);
 623         tmp_counter = tmp_counter<<32;
 624         tmp_counter += ntohl(ch->carp_counter[1]);
 625
 626         /* XXX Replay protection goes here */
 627
 628         sc->sc_init_counter = 0;
 629         sc->sc_counter = tmp_counter;
 630
 631         sc_tv.tv_sec = sc->sc_advbase;
 632         sc_tv.tv_usec = DEMOTE_ADVSKEW(sc) * 1000000 / 256;
 633         ch_tv.tv_sec = ch->carp_advbase;
 634         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 635
 636         switch (sc->sc_state) {
 637         case INIT:
 638                 break;
 639         case MASTER:
 640                 /*
 641                  * If we receive an advertisement from a master who's going to
 642                  * be more frequent than us, go into BACKUP state.
 643                  */
 644                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 645                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 646                         callout_stop(&sc->sc_ad_tmo);
 647                         CARP_LOG("VHID %u@%s: MASTER -> BACKUP "
 648                             "(more frequent advertisement received)\n",
 649                             sc->sc_vhid,
 650                             sc->sc_carpdev->if_xname);
 651                         carp_set_state(sc, BACKUP);
 652                         carp_setrun(sc, 0);
 653                         carp_delroute(sc);
 654                 }
 655                 break;
 656         case BACKUP:
 657                 /*
 658                  * If we're pre-empting masters who advertise slower than us,
 659                  * and this one claims to be slower, treat him as down.
 660                  */
 661                 if (V_carp_preempt && timevalcmp(&sc_tv, &ch_tv, <)) {
 662                         CARP_LOG("VHID %u@%s: BACKUP -> MASTER "
 663                             "(preempting a slower master)\n",
 664                             sc->sc_vhid,
 665                             sc->sc_carpdev->if_xname);
 666                         carp_master_down_locked(sc);
 667                         break;
 668                 }
 669
 670                 /*
 671                  *  If the master is going to advertise at such a low frequency
 672                  *  that he's guaranteed to time out, we'd might as well just
 673                  *  treat him as timed out now.
 674                  */
 675                 sc_tv.tv_sec = sc->sc_advbase * 3;
 676                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 677                         CARP_LOG("VHID %u@%s: BACKUP -> MASTER "
 678                             "(master timed out)\n",
 679                             sc->sc_vhid,
 680                             sc->sc_carpdev->if_xname);
 681                         carp_master_down_locked(sc);
 682                         break;
 683                 }
 684
 685                 /*
 686                  * Otherwise, we reset the counter and wait for the next
 687                  * advertisement.
 688                  */
 689                 carp_setrun(sc, af);
 690                 break;
 691         }
 692
 693 out:
 694         CARP_UNLOCK(sc);
 695         m_freem(m);
 696 }
 697
 698 static int
 699 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 700 {
 701         struct m_tag *mtag;
 702
 703         if (sc->sc_init_counter) {
 704                 /* this could also be seconds since unix epoch */
 705                 sc->sc_counter = arc4random();
 706                 sc->sc_counter = sc->sc_counter << 32;
 707                 sc->sc_counter += arc4random();
 708         } else
 709                 sc->sc_counter++;
 710
 711         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 712         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 713
 714         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 715
 716         /* Tag packet for carp_output */
 717         if ((mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct carp_softc *),
 718             M_NOWAIT)) == NULL) {
 719                 m_freem(m);
 720                 CARPSTATS_INC(carps_onomem);
 721                 return (ENOMEM);
 722         }
 723         bcopy(&sc, mtag + 1, sizeof(sc));
 724         m_tag_prepend(m, mtag);
 725
 726         return (0);
 727 }
 728
 729 /*
 730  * To avoid LORs and possible recursions this function shouldn't
 731  * be called directly, but scheduled via taskqueue.
 732  */
 733 static void
 734 carp_send_ad_all(void *ctx __unused, int pending __unused)
 735 {
 736         struct carp_softc *sc;
 737
 738         mtx_lock(&carp_mtx);
 739         LIST_FOREACH(sc, &carp_list, sc_next)
 740                 if (sc->sc_state == MASTER) {
 741                         CARP_LOCK(sc);
 742                         CURVNET_SET(sc->sc_carpdev->if_vnet);
 743                         carp_send_ad_locked(sc);
 744                         CURVNET_RESTORE();
 745                         CARP_UNLOCK(sc);
 746                 }
 747         mtx_unlock(&carp_mtx);
 748 }
 749
 750 /* Send a periodic advertisement, executed in callout context. */
 751 static void
 752 carp_send_ad(void *v)
 753 {
 754         struct carp_softc *sc = v;
 755
 756         CARP_LOCK_ASSERT(sc);
 757         CURVNET_SET(sc->sc_carpdev->if_vnet);
 758         carp_send_ad_locked(sc);
 759         CURVNET_RESTORE();
 760         CARP_UNLOCK(sc);
 761 }
 762
 763 static void
 764 carp_send_ad_error(struct carp_softc *sc, int error)
 765 {
 766
 767         if (error) {
 768                 if (sc->sc_sendad_errors < INT_MAX)
 769                         sc->sc_sendad_errors++;
 770                 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 771                         static const char fmt[] = "send error %d on %s";
 772                         char msg[sizeof(fmt) + IFNAMSIZ];
 773
 774                         sprintf(msg, fmt, error, sc->sc_carpdev->if_xname);
 775                         carp_demote_adj(V_carp_senderr_adj, msg);
 776                 }
 777                 sc->sc_sendad_success = 0;
 778         } else {
 779                 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS &&
 780                     ++sc->sc_sendad_success >= CARP_SENDAD_MIN_SUCCESS) {
 781                         static const char fmt[] = "send ok on %s";
 782                         char msg[sizeof(fmt) + IFNAMSIZ];
 783
 784                         sprintf(msg, fmt, sc->sc_carpdev->if_xname);
 785                         carp_demote_adj(-V_carp_senderr_adj, msg);
 786                         sc->sc_sendad_errors = 0;
 787                 } else
 788                         sc->sc_sendad_errors = 0;
 789         }
 790 }
 791
 792 static void
 793 carp_send_ad_locked(struct carp_softc *sc)
 794 {
 795         struct carp_header ch;
 796         struct timeval tv;
 797         struct sockaddr sa;
 798         struct ifaddr *ifa;
 799         struct carp_header *ch_ptr;
 800         struct mbuf *m;
 801         int len, advskew;
 802
 803         CARP_LOCK_ASSERT(sc);
 804
 805         advskew = DEMOTE_ADVSKEW(sc);
 806         tv.tv_sec = sc->sc_advbase;
 807         tv.tv_usec = advskew * 1000000 / 256;
 808
 809         ch.carp_version = CARP_VERSION;
 810         ch.carp_type = CARP_ADVERTISEMENT;
 811         ch.carp_vhid = sc->sc_vhid;
 812         ch.carp_advbase = sc->sc_advbase;
 813         ch.carp_advskew = advskew;
 814         ch.carp_authlen = 7;    /* XXX DEFINE */
 815         ch.carp_pad1 = 0;       /* must be zero */
 816         ch.carp_cksum = 0;
 817
 818         /* XXXGL: OpenBSD picks first ifaddr with needed family. */
 819
 820 #ifdef INET
 821         if (sc->sc_naddrs) {
 822                 struct ip *ip;
 823
 824                 m = m_gethdr(M_NOWAIT, MT_DATA);
 825                 if (m == NULL) {
 826                         CARPSTATS_INC(carps_onomem);
 827                         goto resched;
 828                 }
 829                 len = sizeof(*ip) + sizeof(ch);
 830                 m->m_pkthdr.len = len;
 831                 m->m_pkthdr.rcvif = NULL;
 832                 m->m_len = len;
 833                 MH_ALIGN(m, m->m_len);
 834                 m->m_flags |= M_MCAST;
 835                 ip = mtod(m, struct ip *);
 836                 ip->ip_v = IPVERSION;
 837                 ip->ip_hl = sizeof(*ip) >> 2;
 838                 ip->ip_tos = IPTOS_LOWDELAY;
 839                 ip->ip_len = htons(len);
 840                 ip->ip_id = ip_newid();
 841                 ip->ip_off = htons(IP_DF);
 842                 ip->ip_ttl = CARP_DFLTTL;
 843                 ip->ip_p = IPPROTO_CARP;
 844                 ip->ip_sum = 0;
 845
 846                 bzero(&sa, sizeof(sa));
 847                 sa.sa_family = AF_INET;
 848                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 849                 if (ifa != NULL) {
 850                         ip->ip_src.s_addr =
 851                             ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 852                         ifa_free(ifa);
 853                 } else
 854                         ip->ip_src.s_addr = 0;
 855                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 856
 857                 ch_ptr = (struct carp_header *)(&ip[1]);
 858                 bcopy(&ch, ch_ptr, sizeof(ch));
 859                 if (carp_prepare_ad(m, sc, ch_ptr))
 860                         goto resched;
 861
 862                 m->m_data += sizeof(*ip);
 863                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip));
 864                 m->m_data -= sizeof(*ip);
 865
 866                 CARPSTATS_INC(carps_opackets);
 867
 868                 carp_send_ad_error(sc, ip_output(m, NULL, NULL, IP_RAWOUTPUT,
 869                     &sc->sc_carpdev->if_carp->cif_imo, NULL));
 870         }
 871 #endif /* INET */
 872 #ifdef INET6
 873         if (sc->sc_naddrs6) {
 874                 struct ip6_hdr *ip6;
 875
 876                 m = m_gethdr(M_NOWAIT, MT_DATA);
 877                 if (m == NULL) {
 878                         CARPSTATS_INC(carps_onomem);
 879                         goto resched;
 880                 }
 881                 len = sizeof(*ip6) + sizeof(ch);
 882                 m->m_pkthdr.len = len;
 883                 m->m_pkthdr.rcvif = NULL;
 884                 m->m_len = len;
 885                 MH_ALIGN(m, m->m_len);
 886                 m->m_flags |= M_MCAST;
 887                 ip6 = mtod(m, struct ip6_hdr *);
 888                 bzero(ip6, sizeof(*ip6));
 889                 ip6->ip6_vfc |= IPV6_VERSION;
 890                 ip6->ip6_hlim = CARP_DFLTTL;
 891                 ip6->ip6_nxt = IPPROTO_CARP;
 892                 bzero(&sa, sizeof(sa));
 893
 894                 /* set the source address */
 895                 sa.sa_family = AF_INET6;
 896                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 897                 if (ifa != NULL) {
 898                         bcopy(IFA_IN6(ifa), &ip6->ip6_src,
 899                             sizeof(struct in6_addr));
 900                         ifa_free(ifa);
 901                 } else
 902                         /* This should never happen with IPv6. */
 903                         bzero(&ip6->ip6_src, sizeof(struct in6_addr));
 904
 905                 /* Set the multicast destination. */
 906                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
 907                 ip6->ip6_dst.s6_addr8[15] = 0x12;
 908                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
 909                         m_freem(m);
 910                         CARP_DEBUG("%s: in6_setscope failed\n", __func__);
 911                         goto resched;
 912                 }
 913
 914                 ch_ptr = (struct carp_header *)(&ip6[1]);
 915                 bcopy(&ch, ch_ptr, sizeof(ch));
 916                 if (carp_prepare_ad(m, sc, ch_ptr))
 917                         goto resched;
 918
 919                 m->m_data += sizeof(*ip6);
 920                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip6));
 921                 m->m_data -= sizeof(*ip6);
 922
 923                 CARPSTATS_INC(carps_opackets6);
 924
 925                 carp_send_ad_error(sc, ip6_output(m, NULL, NULL, 0,
 926                     &sc->sc_carpdev->if_carp->cif_im6o, NULL, NULL));
 927         }
 928 #endif /* INET6 */
 929
 930 resched:
 931         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), carp_send_ad, sc);
 932 }
 933
 934 static void
 935 carp_addroute(struct carp_softc *sc)
 936 {
 937         struct ifaddr *ifa;
 938
 939         CARP_FOREACH_IFA(sc, ifa)
 940                 carp_ifa_addroute(ifa);
 941 }
 942
 943 static void
 944 carp_ifa_addroute(struct ifaddr *ifa)
 945 {
 946
 947         switch (ifa->ifa_addr->sa_family) {
 948 #ifdef INET
 949         case AF_INET:
 950                 in_addprefix(ifatoia(ifa), RTF_UP);
 951                 ifa_add_loopback_route(ifa,
 952                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 953                 break;
 954 #endif
 955 #ifdef INET6
 956         case AF_INET6:
 957                 ifa_add_loopback_route(ifa,
 958                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 959                 in6_ifaddloop(ifa);
 960                 break;
 961 #endif
 962         }
 963 }
 964
 965 static void
 966 carp_delroute(struct carp_softc *sc)
 967 {
 968         struct ifaddr *ifa;
 969
 970         CARP_FOREACH_IFA(sc, ifa)
 971                 carp_ifa_delroute(ifa);
 972 }
 973
 974 static void
 975 carp_ifa_delroute(struct ifaddr *ifa)
 976 {
 977
 978         switch (ifa->ifa_addr->sa_family) {
 979 #ifdef INET
 980         case AF_INET:
 981                 ifa_del_loopback_route(ifa,
 982                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 983                 in_scrubprefix(ifatoia(ifa), LLE_STATIC);
 984                 break;
 985 #endif
 986 #ifdef INET6
 987         case AF_INET6:
 988                 ifa_del_loopback_route(ifa,
 989                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 990                 in6_ifremloop(ifa);
 991                 break;
 992 #endif
 993         }
 994 }
 995
 996 int
 997 carp_master(struct ifaddr *ifa)
 998 {
 999         struct carp_softc *sc = ifa->ifa_carp;
1000
1001         return (sc->sc_state == MASTER);
1002 }
1003
1004 #ifdef INET
1005 /*
1006  * Broadcast a gratuitous ARP request containing
1007  * the virtual router MAC address for each IP address
1008  * associated with the virtual router.
1009  */
1010 static void
1011 carp_send_arp(struct carp_softc *sc)
1012 {
1013         struct ifaddr *ifa;
1014
1015         CARP_FOREACH_IFA(sc, ifa)
1016                 if (ifa->ifa_addr->sa_family == AF_INET)
1017                         arp_ifinit2(sc->sc_carpdev, ifa, LLADDR(&sc->sc_addr));
1018 }
1019
1020 int
1021 carp_iamatch(struct ifaddr *ifa, uint8_t **enaddr)
1022 {
1023         struct carp_softc *sc = ifa->ifa_carp;
1024
1025         if (sc->sc_state == MASTER) {
1026                 *enaddr = LLADDR(&sc->sc_addr);
1027                 return (1);
1028         }
1029
1030         return (0);
1031 }
1032 #endif
1033
1034 #ifdef INET6
1035 static void
1036 carp_send_na(struct carp_softc *sc)
1037 {
1038         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1039         struct ifaddr *ifa;
1040         struct in6_addr *in6;
1041
1042         CARP_FOREACH_IFA(sc, ifa) {
1043                 if (ifa->ifa_addr->sa_family != AF_INET6)
1044                         continue;
1045
1046                 in6 = IFA_IN6(ifa);
1047                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1048                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1049                 DELAY(1000);    /* XXX */
1050         }
1051 }
1052
1053 /*
1054  * Returns ifa in case it's a carp address and it is MASTER, or if the address
1055  * matches and is not a carp address.  Returns NULL otherwise.
1056  */
1057 struct ifaddr *
1058 carp_iamatch6(struct ifnet *ifp, struct in6_addr *taddr)
1059 {
1060         struct ifaddr *ifa;
1061
1062         ifa = NULL;
1063         IF_ADDR_RLOCK(ifp);
1064         TAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
1065                 if (ifa->ifa_addr->sa_family != AF_INET6)
1066                         continue;
1067                 if (!IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa)))
1068                         continue;
1069                 if (ifa->ifa_carp && ifa->ifa_carp->sc_state != MASTER)
1070                         ifa = NULL;
1071                 else
1072                         ifa_ref(ifa);
1073                 break;
1074         }
1075         IF_ADDR_RUNLOCK(ifp);
1076
1077         return (ifa);
1078 }
1079
1080 caddr_t
1081 carp_macmatch6(struct ifnet *ifp, struct mbuf *m, const struct in6_addr *taddr)
1082 {
1083         struct ifaddr *ifa;
1084
1085         IF_ADDR_RLOCK(ifp);
1086         IFNET_FOREACH_IFA(ifp, ifa)
1087                 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1088                     IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1089                         struct carp_softc *sc = ifa->ifa_carp;
1090                         struct m_tag *mtag;
1091
1092                         IF_ADDR_RUNLOCK(ifp);
1093
1094                         mtag = m_tag_get(PACKET_TAG_CARP,
1095                             sizeof(struct carp_softc *), M_NOWAIT);
1096                         if (mtag == NULL)
1097                                 /* Better a bit than nothing. */
1098                                 return (LLADDR(&sc->sc_addr));
1099
1100                         bcopy(&sc, mtag + 1, sizeof(sc));
1101                         m_tag_prepend(m, mtag);
1102
1103                         return (LLADDR(&sc->sc_addr));
1104                 }
1105         IF_ADDR_RUNLOCK(ifp);
1106
1107         return (NULL);
1108 }
1109 #endif /* INET6 */
1110
1111 int
1112 carp_forus(struct ifnet *ifp, u_char *dhost)
1113 {
1114         struct carp_softc *sc;
1115         uint8_t *ena = dhost;
1116
1117         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1118                 return (0);
1119
1120         CIF_LOCK(ifp->if_carp);
1121         IFNET_FOREACH_CARP(ifp, sc) {
1122                 CARP_LOCK(sc);
1123                 if (sc->sc_state == MASTER && !bcmp(dhost, LLADDR(&sc->sc_addr),
1124                     ETHER_ADDR_LEN)) {
1125                         CARP_UNLOCK(sc);
1126                         CIF_UNLOCK(ifp->if_carp);
1127                         return (1);
1128                 }
1129                 CARP_UNLOCK(sc);
1130         }
1131         CIF_UNLOCK(ifp->if_carp);
1132
1133         return (0);
1134 }
1135
1136 /* Master down timeout event, executed in callout context. */
1137 static void
1138 carp_master_down(void *v)
1139 {
1140         struct carp_softc *sc = v;
1141
1142         CARP_LOCK_ASSERT(sc);
1143
1144         CURVNET_SET(sc->sc_carpdev->if_vnet);
1145         if (sc->sc_state == BACKUP) {
1146                 CARP_LOG("VHID %u@%s: BACKUP -> MASTER (master down)\n",
1147                     sc->sc_vhid,
1148                     sc->sc_carpdev->if_xname);
1149                 carp_master_down_locked(sc);
1150         }
1151         CURVNET_RESTORE();
1152
1153         CARP_UNLOCK(sc);
1154 }
1155
1156 static void
1157 carp_master_down_locked(struct carp_softc *sc)
1158 {
1159
1160         CARP_LOCK_ASSERT(sc);
1161
1162         switch (sc->sc_state) {
1163         case BACKUP:
1164                 carp_set_state(sc, MASTER);
1165                 carp_send_ad_locked(sc);
1166 #ifdef INET
1167                 carp_send_arp(sc);
1168 #endif
1169 #ifdef INET6
1170                 carp_send_na(sc);
1171 #endif
1172                 carp_setrun(sc, 0);
1173                 carp_addroute(sc);
1174                 break;
1175         case INIT:
1176         case MASTER:
1177 #ifdef INVARIANTS
1178                 panic("carp: VHID %u@%s: master_down event in %s state\n",
1179                     sc->sc_vhid,
1180                     sc->sc_carpdev->if_xname,
1181                     sc->sc_state ? "MASTER" : "INIT");
1182 #endif
1183                 break;
1184         }
1185 }
1186
1187 /*
1188  * When in backup state, af indicates whether to reset the master down timer
1189  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1190  */
1191 static void
1192 carp_setrun(struct carp_softc *sc, sa_family_t af)
1193 {
1194         struct timeval tv;
1195
1196         CARP_LOCK_ASSERT(sc);
1197
1198         if ((sc->sc_carpdev->if_flags & IFF_UP) == 0 ||
1199             sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1200             (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0))
1201                 return;
1202
1203         switch (sc->sc_state) {
1204         case INIT:
1205                 CARP_LOG("VHID %u@%s: INIT -> BACKUP\n",
1206                     sc->sc_vhid,
1207                     sc->sc_carpdev->if_xname);
1208                 carp_set_state(sc, BACKUP);
1209                 carp_setrun(sc, 0);
1210                 break;
1211         case BACKUP:
1212                 callout_stop(&sc->sc_ad_tmo);
1213                 tv.tv_sec = 3 * sc->sc_advbase;
1214                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1215                 switch (af) {
1216 #ifdef INET
1217                 case AF_INET:
1218                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1219                             carp_master_down, sc);
1220                         break;
1221 #endif
1222 #ifdef INET6
1223                 case AF_INET6:
1224                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1225                             carp_master_down, sc);
1226                         break;
1227 #endif
1228                 default:
1229 #ifdef INET
1230                         if (sc->sc_naddrs)
1231                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1232                                     carp_master_down, sc);
1233 #endif
1234 #ifdef INET6
1235                         if (sc->sc_naddrs6)
1236                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1237                                     carp_master_down, sc);
1238 #endif
1239                         break;
1240                 }
1241                 break;
1242         case MASTER:
1243                 tv.tv_sec = sc->sc_advbase;
1244                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1245                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1246                     carp_send_ad, sc);
1247                 break;
1248         }
1249 }
1250
1251 /*
1252  * Setup multicast structures.
1253  */
1254 static int
1255 carp_multicast_setup(struct carp_if *cif, sa_family_t sa)
1256 {
1257         struct ifnet *ifp = cif->cif_ifp;
1258         int error = 0;
1259
1260         CIF_LOCK_ASSERT(cif);
1261
1262         switch (sa) {
1263 #ifdef INET
1264         case AF_INET:
1265             {
1266                 struct ip_moptions *imo = &cif->cif_imo;
1267                 struct in_addr addr;
1268
1269                 if (imo->imo_membership)
1270                         return (0);
1271
1272                 imo->imo_membership = (struct in_multi **)malloc(
1273                     (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
1274                     M_NOWAIT);
1275                 if (imo->imo_membership == NULL)
1276                         return (ENOMEM);
1277                 imo->imo_mfilters = NULL;
1278                 imo->imo_max_memberships = IP_MIN_MEMBERSHIPS;
1279                 imo->imo_multicast_vif = -1;
1280
1281                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1282                 if ((error = in_joingroup(ifp, &addr, NULL,
1283                     &imo->imo_membership[0])) != 0) {
1284                         free(imo->imo_membership, M_CARP);
1285                         break;
1286                 }
1287                 imo->imo_num_memberships++;
1288                 imo->imo_multicast_ifp = ifp;
1289                 imo->imo_multicast_ttl = CARP_DFLTTL;
1290                 imo->imo_multicast_loop = 0;
1291                 break;
1292            }
1293 #endif
1294 #ifdef INET6
1295         case AF_INET6:
1296             {
1297                 struct ip6_moptions *im6o = &cif->cif_im6o;
1298                 struct in6_addr in6;
1299                 struct in6_multi *in6m;
1300
1301                 if (im6o->im6o_membership)
1302                         return (0);
1303
1304                 im6o->im6o_membership = (struct in6_multi **)malloc(
1305                     (sizeof(struct in6_multi *) * IPV6_MIN_MEMBERSHIPS), M_CARP,
1306                     M_ZERO | M_NOWAIT);
1307                 if (im6o->im6o_membership == NULL)
1308                         return (ENOMEM);
1309                 im6o->im6o_mfilters = NULL;
1310                 im6o->im6o_max_memberships = IPV6_MIN_MEMBERSHIPS;
1311                 im6o->im6o_multicast_hlim = CARP_DFLTTL;
1312                 im6o->im6o_multicast_ifp = ifp;
1313
1314                 /* Join IPv6 CARP multicast group. */
1315                 bzero(&in6, sizeof(in6));
1316                 in6.s6_addr16[0] = htons(0xff02);
1317                 in6.s6_addr8[15] = 0x12;
1318                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1319                         free(im6o->im6o_membership, M_CARP);
1320                         break;
1321                 }
1322                 in6m = NULL;
1323                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1324                         free(im6o->im6o_membership, M_CARP);
1325                         break;
1326                 }
1327                 im6o->im6o_membership[0] = in6m;
1328                 im6o->im6o_num_memberships++;
1329
1330                 /* Join solicited multicast address. */
1331                 bzero(&in6, sizeof(in6));
1332                 in6.s6_addr16[0] = htons(0xff02);
1333                 in6.s6_addr32[1] = 0;
1334                 in6.s6_addr32[2] = htonl(1);
1335                 in6.s6_addr32[3] = 0;
1336                 in6.s6_addr8[12] = 0xff;
1337                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1338                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1339                         free(im6o->im6o_membership, M_CARP);
1340                         break;
1341                 }
1342                 in6m = NULL;
1343                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1344                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1345                         free(im6o->im6o_membership, M_CARP);
1346                         break;
1347                 }
1348                 im6o->im6o_membership[1] = in6m;
1349                 im6o->im6o_num_memberships++;
1350                 break;
1351             }
1352 #endif
1353         }
1354
1355         return (error);
1356 }
1357
1358 /*
1359  * Free multicast structures.
1360  */
1361 static void
1362 carp_multicast_cleanup(struct carp_if *cif, sa_family_t sa)
1363 {
1364
1365         CIF_LOCK_ASSERT(cif);
1366         switch (sa) {
1367 #ifdef INET
1368         case AF_INET:
1369                 if (cif->cif_naddrs == 0) {
1370                         struct ip_moptions *imo = &cif->cif_imo;
1371
1372                         in_leavegroup(imo->imo_membership[0], NULL);
1373                         KASSERT(imo->imo_mfilters == NULL,
1374                             ("%s: imo_mfilters != NULL", __func__));
1375                         free(imo->imo_membership, M_CARP);
1376                         imo->imo_membership = NULL;
1377
1378                 }
1379                 break;
1380 #endif
1381 #ifdef INET6
1382         case AF_INET6:
1383                 if (cif->cif_naddrs6 == 0) {
1384                         struct ip6_moptions *im6o = &cif->cif_im6o;
1385
1386                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1387                         in6_mc_leave(im6o->im6o_membership[1], NULL);
1388                         KASSERT(im6o->im6o_mfilters == NULL,
1389                             ("%s: im6o_mfilters != NULL", __func__));
1390                         free(im6o->im6o_membership, M_CARP);
1391                         im6o->im6o_membership = NULL;
1392                 }
1393                 break;
1394 #endif
1395         }
1396 }
1397
1398 int
1399 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa)
1400 {
1401         struct m_tag *mtag;
1402         struct carp_softc *sc;
1403
1404         if (!sa)
1405                 return (0);
1406
1407         switch (sa->sa_family) {
1408 #ifdef INET
1409         case AF_INET:
1410                 break;
1411 #endif
1412 #ifdef INET6
1413         case AF_INET6:
1414                 break;
1415 #endif
1416         default:
1417                 return (0);
1418         }
1419
1420         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
1421         if (mtag == NULL)
1422                 return (0);
1423
1424         bcopy(mtag + 1, &sc, sizeof(sc));
1425
1426         /* Set the source MAC address to the Virtual Router MAC Address. */
1427         switch (ifp->if_type) {
1428         case IFT_ETHER:
1429         case IFT_BRIDGE:
1430         case IFT_L2VLAN: {
1431                         struct ether_header *eh;
1432
1433                         eh = mtod(m, struct ether_header *);
1434                         eh->ether_shost[0] = 0;
1435                         eh->ether_shost[1] = 0;
1436                         eh->ether_shost[2] = 0x5e;
1437                         eh->ether_shost[3] = 0;
1438                         eh->ether_shost[4] = 1;
1439                         eh->ether_shost[5] = sc->sc_vhid;
1440                 }
1441                 break;
1442         case IFT_FDDI: {
1443                         struct fddi_header *fh;
1444
1445                         fh = mtod(m, struct fddi_header *);
1446                         fh->fddi_shost[0] = 0;
1447                         fh->fddi_shost[1] = 0;
1448                         fh->fddi_shost[2] = 0x5e;
1449                         fh->fddi_shost[3] = 0;
1450                         fh->fddi_shost[4] = 1;
1451                         fh->fddi_shost[5] = sc->sc_vhid;
1452                 }
1453                 break;
1454         case IFT_ISO88025: {
1455                         struct iso88025_header *th;
1456                         th = mtod(m, struct iso88025_header *);
1457                         th->iso88025_shost[0] = 3;
1458                         th->iso88025_shost[1] = 0;
1459                         th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
1460                         th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
1461                         th->iso88025_shost[4] = 0;
1462                         th->iso88025_shost[5] = 0;
1463                 }
1464                 break;
1465         default:
1466                 printf("%s: carp is not supported for the %d interface type\n",
1467                     ifp->if_xname, ifp->if_type);
1468                 return (EOPNOTSUPP);
1469         }
1470
1471         return (0);
1472 }
1473
1474 static struct carp_softc*
1475 carp_alloc(struct ifnet *ifp)
1476 {
1477         struct carp_softc *sc;
1478         struct carp_if *cif;
1479
1480         if ((cif = ifp->if_carp) == NULL) {
1481                 cif = carp_alloc_if(ifp);
1482                 if (cif == NULL)
1483                         return (NULL);
1484         }
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
1570         cif = malloc(sizeof(*cif), M_CARP, M_WAITOK|M_ZERO);
1571
1572         if (ifpromisc(ifp, 1) != 0)
1573                 goto cleanup;
1574
1575         CIF_LOCK_INIT(cif);
1576         cif->cif_ifp = ifp;
1577         TAILQ_INIT(&cif->cif_vrs);
1578
1579         IF_ADDR_WLOCK(ifp);
1580         ifp->if_carp = cif;
1581         if_ref(ifp);
1582         IF_ADDR_WUNLOCK(ifp);
1583
1584         return (cif);
1585
1586 cleanup:
1587         free(cif, M_CARP);
1588
1589         return (NULL);
1590 }
1591
1592 static void
1593 carp_free_if(struct carp_if *cif)
1594 {
1595         struct ifnet *ifp = cif->cif_ifp;
1596
1597         CIF_LOCK_ASSERT(cif);
1598         KASSERT(TAILQ_EMPTY(&cif->cif_vrs), ("%s: softc list not empty",
1599             __func__));
1600
1601         IF_ADDR_WLOCK(ifp);
1602         ifp->if_carp = NULL;
1603         IF_ADDR_WUNLOCK(ifp);
1604
1605         CIF_LOCK_DESTROY(cif);
1606
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                         if (sc == NULL) {
1681                                 error = EINVAL; /* XXX: ifpromisc failed */
1682                                 break;
1683                         }
1684
1685                         CARP_LOCK(sc);
1686                         sc->sc_vhid = carpr.carpr_vhid;
1687                         LLADDR(&sc->sc_addr)[0] = 0;
1688                         LLADDR(&sc->sc_addr)[1] = 0;
1689                         LLADDR(&sc->sc_addr)[2] = 0x5e;
1690                         LLADDR(&sc->sc_addr)[3] = 0;
1691                         LLADDR(&sc->sc_addr)[4] = 1;
1692                         LLADDR(&sc->sc_addr)[5] = sc->sc_vhid;
1693                 } else
1694                         CARP_LOCK(sc);
1695                 locked = 1;
1696                 if (carpr.carpr_advbase > 0) {
1697                         if (carpr.carpr_advbase > 255 ||
1698                             carpr.carpr_advbase < CARP_DFLTINTV) {
1699                                 error = EINVAL;
1700                                 break;
1701                         }
1702                         sc->sc_advbase = carpr.carpr_advbase;
1703                 }
1704                 if (carpr.carpr_advskew >= 255) {
1705                         error = EINVAL;
1706                         break;
1707                 }
1708                 sc->sc_advskew = carpr.carpr_advskew;
1709                 if (carpr.carpr_key[0] != '\0') {
1710                         bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1711                         carp_hmac_prepare(sc);
1712                 }
1713                 if (sc->sc_state != INIT &&
1714                     carpr.carpr_state != sc->sc_state) {
1715                         switch (carpr.carpr_state) {
1716                         case BACKUP:
1717                                 callout_stop(&sc->sc_ad_tmo);
1718                                 carp_set_state(sc, BACKUP);
1719                                 carp_setrun(sc, 0);
1720                                 carp_delroute(sc);
1721                                 break;
1722                         case MASTER:
1723                                 carp_master_down_locked(sc);
1724                                 break;
1725                         default:
1726                                 break;
1727                         }
1728                 }
1729                 break;
1730
1731         case SIOCGVH:
1732             {
1733                 int priveleged;
1734
1735                 if (carpr.carpr_vhid < 0 || carpr.carpr_vhid > CARP_MAXVHID) {
1736                         error = EINVAL;
1737                         break;
1738                 }
1739                 if (carpr.carpr_count < 1) {
1740                         error = EMSGSIZE;
1741                         break;
1742                 }
1743                 if (ifp->if_carp == NULL) {
1744                         error = ENOENT;
1745                         break;
1746                 }
1747
1748                 priveleged = (priv_check(td, PRIV_NETINET_CARP) == 0);
1749                 if (carpr.carpr_vhid != 0) {
1750                         CIF_LOCK(ifp->if_carp);
1751                         IFNET_FOREACH_CARP(ifp, sc)
1752                                 if (sc->sc_vhid == carpr.carpr_vhid)
1753                                         break;
1754                         CIF_UNLOCK(ifp->if_carp);
1755                         if (sc == NULL) {
1756                                 error = ENOENT;
1757                                 break;
1758                         }
1759                         carp_carprcp(&carpr, sc, priveleged);
1760                         error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1761                 } else  {
1762                         int i, count;
1763
1764                         count = 0;
1765                         CIF_LOCK(ifp->if_carp);
1766                         IFNET_FOREACH_CARP(ifp, sc)
1767                                 count++;
1768
1769                         if (count > carpr.carpr_count) {
1770                                 CIF_UNLOCK(ifp->if_carp);
1771                                 error = EMSGSIZE;
1772                                 break;
1773                         }
1774
1775                         i = 0;
1776                         IFNET_FOREACH_CARP(ifp, sc) {
1777                                 carp_carprcp(&carpr, sc, priveleged);
1778                                 carpr.carpr_count = count;
1779                                 error = copyout(&carpr, ifr->ifr_data +
1780                                     (i * sizeof(carpr)), sizeof(carpr));
1781                                 if (error) {
1782                                         CIF_UNLOCK(ifp->if_carp);
1783                                         break;
1784                                 }
1785                                 i++;
1786                         }
1787                         CIF_UNLOCK(ifp->if_carp);
1788                 }
1789                 break;
1790             }
1791         default:
1792                 error = EINVAL;
1793         }
1794
1795 out:
1796         if (locked)
1797                 CARP_UNLOCK(sc);
1798         if_rele(ifp);
1799
1800         return (error);
1801 }
1802
1803 static int
1804 carp_get_vhid(struct ifaddr *ifa)
1805 {
1806
1807         if (ifa == NULL || ifa->ifa_carp == NULL)
1808                 return (0);
1809
1810         return (ifa->ifa_carp->sc_vhid);
1811 }
1812
1813 int
1814 carp_attach(struct ifaddr *ifa, int vhid)
1815 {
1816         struct ifnet *ifp = ifa->ifa_ifp;
1817         struct carp_if *cif = ifp->if_carp;
1818         struct carp_softc *sc;
1819         int index, error;
1820
1821         if (ifp->if_carp == NULL)
1822                 return (ENOPROTOOPT);
1823
1824         switch (ifa->ifa_addr->sa_family) {
1825 #ifdef INET
1826         case AF_INET:
1827 #endif
1828 #ifdef INET6
1829         case AF_INET6:
1830 #endif
1831                 break;
1832         default:
1833                 return (EPROTOTYPE);
1834         }
1835
1836         CIF_LOCK(cif);
1837         IFNET_FOREACH_CARP(ifp, sc)
1838                 if (sc->sc_vhid == vhid)
1839                         break;
1840         if (sc == NULL) {
1841                 CIF_UNLOCK(cif);
1842                 return (ENOENT);
1843         }
1844
1845         if (ifa->ifa_carp) {
1846                 if (ifa->ifa_carp->sc_vhid != vhid)
1847                         carp_detach_locked(ifa);
1848                 else {
1849                         CIF_UNLOCK(cif);
1850                         return (0);
1851                 }
1852         }
1853
1854         error = carp_multicast_setup(cif, ifa->ifa_addr->sa_family);
1855         if (error) {
1856                 CIF_FREE(cif);
1857                 return (error);
1858         }
1859
1860         CARP_LOCK(sc);
1861         index = sc->sc_naddrs + sc->sc_naddrs6 + 1;
1862         if (index > sc->sc_ifasiz / sizeof(struct ifaddr *))
1863                 if ((error = carp_grow_ifas(sc)) != 0) {
1864                         carp_multicast_cleanup(cif,
1865                             ifa->ifa_addr->sa_family);
1866                         CARP_UNLOCK(sc);
1867                         CIF_FREE(cif);
1868                         return (error);
1869                 }
1870
1871         switch (ifa->ifa_addr->sa_family) {
1872 #ifdef INET
1873         case AF_INET:
1874                 cif->cif_naddrs++;
1875                 sc->sc_naddrs++;
1876                 break;
1877 #endif
1878 #ifdef INET6
1879         case AF_INET6:
1880                 cif->cif_naddrs6++;
1881                 sc->sc_naddrs6++;
1882                 break;
1883 #endif
1884         }
1885
1886         ifa_ref(ifa);
1887         sc->sc_ifas[index - 1] = ifa;
1888         ifa->ifa_carp = sc;
1889
1890         carp_hmac_prepare(sc);
1891         carp_sc_state(sc);
1892
1893         CARP_UNLOCK(sc);
1894         CIF_UNLOCK(cif);
1895
1896         return (0);
1897 }
1898
1899 void
1900 carp_detach(struct ifaddr *ifa)
1901 {
1902         struct ifnet *ifp = ifa->ifa_ifp;
1903         struct carp_if *cif = ifp->if_carp;
1904
1905         CIF_LOCK(cif);
1906         carp_detach_locked(ifa);
1907         CIF_FREE(cif);
1908 }
1909
1910 static void
1911 carp_detach_locked(struct ifaddr *ifa)
1912 {
1913         struct ifnet *ifp = ifa->ifa_ifp;
1914         struct carp_if *cif = ifp->if_carp;
1915         struct carp_softc *sc = ifa->ifa_carp;
1916         int i, index;
1917
1918         KASSERT(sc != NULL, ("%s: %p not attached", __func__, ifa));
1919
1920         CIF_LOCK_ASSERT(cif);
1921         CARP_LOCK(sc);
1922
1923         /* Shift array. */
1924         index = sc->sc_naddrs + sc->sc_naddrs6;
1925         for (i = 0; i < index; i++)
1926                 if (sc->sc_ifas[i] == ifa)
1927                         break;
1928         KASSERT(i < index, ("%s: %p no backref", __func__, ifa));
1929         for (; i < index - 1; i++)
1930                 sc->sc_ifas[i] = sc->sc_ifas[i+1];
1931         sc->sc_ifas[index - 1] = NULL;
1932
1933         switch (ifa->ifa_addr->sa_family) {
1934 #ifdef INET
1935         case AF_INET:
1936                 cif->cif_naddrs--;
1937                 sc->sc_naddrs--;
1938                 break;
1939 #endif
1940 #ifdef INET6
1941         case AF_INET6:
1942                 cif->cif_naddrs6--;
1943                 sc->sc_naddrs6--;
1944                 break;
1945 #endif
1946         }
1947
1948         carp_ifa_delroute(ifa);
1949         carp_multicast_cleanup(cif, ifa->ifa_addr->sa_family);
1950
1951         ifa->ifa_carp = NULL;
1952         ifa_free(ifa);
1953
1954         carp_hmac_prepare(sc);
1955         carp_sc_state(sc);
1956
1957         if (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) {
1958                 CARP_UNLOCK(sc);
1959                 carp_destroy(sc);
1960         } else
1961                 CARP_UNLOCK(sc);
1962 }
1963
1964 static void
1965 carp_set_state(struct carp_softc *sc, int state)
1966 {
1967
1968         CARP_LOCK_ASSERT(sc);
1969
1970         if (sc->sc_state != state) {
1971                 const char *carp_states[] = { CARP_STATES };
1972                 char subsys[IFNAMSIZ+5];
1973
1974                 sc->sc_state = state;
1975
1976                 snprintf(subsys, IFNAMSIZ+5, "%u@%s", sc->sc_vhid,
1977                     sc->sc_carpdev->if_xname);
1978                 devctl_notify("CARP", subsys, carp_states[state], NULL);
1979         }
1980 }
1981
1982 static void
1983 carp_linkstate(struct ifnet *ifp)
1984 {
1985         struct carp_softc *sc;
1986
1987         CIF_LOCK(ifp->if_carp);
1988         IFNET_FOREACH_CARP(ifp, sc) {
1989                 CARP_LOCK(sc);
1990                 carp_sc_state(sc);
1991                 CARP_UNLOCK(sc);
1992         }
1993         CIF_UNLOCK(ifp->if_carp);
1994 }
1995
1996 static void
1997 carp_sc_state(struct carp_softc *sc)
1998 {
1999
2000         CARP_LOCK_ASSERT(sc);
2001
2002         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
2003             !(sc->sc_carpdev->if_flags & IFF_UP)) {
2004                 callout_stop(&sc->sc_ad_tmo);
2005 #ifdef INET
2006                 callout_stop(&sc->sc_md_tmo);
2007 #endif
2008 #ifdef INET6
2009                 callout_stop(&sc->sc_md6_tmo);
2010 #endif
2011                 carp_set_state(sc, INIT);
2012                 carp_setrun(sc, 0);
2013                 if (!sc->sc_suppress)
2014                         carp_demote_adj(V_carp_ifdown_adj, "interface down");
2015                 sc->sc_suppress = 1;
2016         } else {
2017                 carp_set_state(sc, INIT);
2018                 carp_setrun(sc, 0);
2019                 if (sc->sc_suppress)
2020                         carp_demote_adj(-V_carp_ifdown_adj, "interface up");
2021                 sc->sc_suppress = 0;
2022         }
2023 }
2024
2025 static void
2026 carp_demote_adj(int adj, char *reason)
2027 {
2028         atomic_add_int(&V_carp_demotion, adj);
2029         CARP_LOG("demoted by %d to %d (%s)\n", adj, V_carp_demotion, reason);
2030         taskqueue_enqueue(taskqueue_swi, &carp_sendall_task);
2031 }
2032
2033 static int
2034 carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS)
2035 {
2036         int new, error;
2037
2038         new = V_carp_demotion;
2039         error = sysctl_handle_int(oidp, &new, 0, req);
2040         if (error || !req->newptr)
2041                 return (error);
2042
2043         carp_demote_adj(new, "sysctl");
2044
2045         return (0);
2046 }
2047
2048 #ifdef INET
2049 extern  struct domain inetdomain;
2050 static struct protosw in_carp_protosw = {
2051         .pr_type =              SOCK_RAW,
2052         .pr_domain =            &inetdomain,
2053         .pr_protocol =          IPPROTO_CARP,
2054         .pr_flags =             PR_ATOMIC|PR_ADDR,
2055         .pr_input =             carp_input,
2056         .pr_output =            (pr_output_t *)rip_output,
2057         .pr_ctloutput =         rip_ctloutput,
2058         .pr_usrreqs =           &rip_usrreqs
2059 };
2060 #endif
2061
2062 #ifdef INET6
2063 extern  struct domain inet6domain;
2064 static struct ip6protosw in6_carp_protosw = {
2065         .pr_type =              SOCK_RAW,
2066         .pr_domain =            &inet6domain,
2067         .pr_protocol =          IPPROTO_CARP,
2068         .pr_flags =             PR_ATOMIC|PR_ADDR,
2069         .pr_input =             carp6_input,
2070         .pr_output =            rip6_output,
2071         .pr_ctloutput =         rip6_ctloutput,
2072         .pr_usrreqs =           &rip6_usrreqs
2073 };
2074 #endif
2075
2076 static void
2077 carp_mod_cleanup(void)
2078 {
2079
2080 #ifdef INET
2081         if (proto_reg[CARP_INET] == 0) {
2082                 (void)ipproto_unregister(IPPROTO_CARP);
2083                 pf_proto_unregister(PF_INET, IPPROTO_CARP, SOCK_RAW);
2084                 proto_reg[CARP_INET] = -1;
2085         }
2086         carp_iamatch_p = NULL;
2087 #endif
2088 #ifdef INET6
2089         if (proto_reg[CARP_INET6] == 0) {
2090                 (void)ip6proto_unregister(IPPROTO_CARP);
2091                 pf_proto_unregister(PF_INET6, IPPROTO_CARP, SOCK_RAW);
2092                 proto_reg[CARP_INET6] = -1;
2093         }
2094         carp_iamatch6_p = NULL;
2095         carp_macmatch6_p = NULL;
2096 #endif
2097         carp_ioctl_p = NULL;
2098         carp_attach_p = NULL;
2099         carp_detach_p = NULL;
2100         carp_get_vhid_p = NULL;
2101         carp_linkstate_p = NULL;
2102         carp_forus_p = NULL;
2103         carp_output_p = NULL;
2104         carp_demote_adj_p = NULL;
2105         carp_master_p = NULL;
2106         mtx_unlock(&carp_mtx);
2107         taskqueue_drain(taskqueue_swi, &carp_sendall_task);
2108         mtx_destroy(&carp_mtx);
2109 }
2110
2111 static int
2112 carp_mod_load(void)
2113 {
2114         int err;
2115
2116         mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2117         LIST_INIT(&carp_list);
2118         carp_get_vhid_p = carp_get_vhid;
2119         carp_forus_p = carp_forus;
2120         carp_output_p = carp_output;
2121         carp_linkstate_p = carp_linkstate;
2122         carp_ioctl_p = carp_ioctl;
2123         carp_attach_p = carp_attach;
2124         carp_detach_p = carp_detach;
2125         carp_demote_adj_p = carp_demote_adj;
2126         carp_master_p = carp_master;
2127 #ifdef INET6
2128         carp_iamatch6_p = carp_iamatch6;
2129         carp_macmatch6_p = carp_macmatch6;
2130         proto_reg[CARP_INET6] = pf_proto_register(PF_INET6,
2131             (struct protosw *)&in6_carp_protosw);
2132         if (proto_reg[CARP_INET6]) {
2133                 printf("carp: error %d attaching to PF_INET6\n",
2134                     proto_reg[CARP_INET6]);
2135                 carp_mod_cleanup();
2136                 return (proto_reg[CARP_INET6]);
2137         }
2138         err = ip6proto_register(IPPROTO_CARP);
2139         if (err) {
2140                 printf("carp: error %d registering with INET6\n", err);
2141                 carp_mod_cleanup();
2142                 return (err);
2143         }
2144 #endif
2145 #ifdef INET
2146         carp_iamatch_p = carp_iamatch;
2147         proto_reg[CARP_INET] = pf_proto_register(PF_INET, &in_carp_protosw);
2148         if (proto_reg[CARP_INET]) {
2149                 printf("carp: error %d attaching to PF_INET\n",
2150                     proto_reg[CARP_INET]);
2151                 carp_mod_cleanup();
2152                 return (proto_reg[CARP_INET]);
2153         }
2154         err = ipproto_register(IPPROTO_CARP);
2155         if (err) {
2156                 printf("carp: error %d registering with INET\n", err);
2157                 carp_mod_cleanup();
2158                 return (err);
2159         }
2160 #endif
2161         return (0);
2162 }
2163
2164 static int
2165 carp_modevent(module_t mod, int type, void *data)
2166 {
2167         switch (type) {
2168         case MOD_LOAD:
2169                 return carp_mod_load();
2170                 /* NOTREACHED */
2171         case MOD_UNLOAD:
2172                 mtx_lock(&carp_mtx);
2173                 if (LIST_EMPTY(&carp_list))
2174                         carp_mod_cleanup();
2175                 else {
2176                         mtx_unlock(&carp_mtx);
2177                         return (EBUSY);
2178                 }
2179                 break;
2180
2181         default:
2182                 return (EINVAL);
2183         }
2184
2185         return (0);
2186 }
2187
2188 static moduledata_t carp_mod = {
2189         "carp",
2190         carp_modevent,
2191         0
2192 };
2193
2194 DECLARE_MODULE(carp, carp_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);