sys/netinet/ip_carp.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2002 Michael Shalayeff.
   5  * Copyright (c) 2003 Ryan McBride.
   6  * Copyright (c) 2011 Gleb Smirnoff <glebius@FreeBSD.org>
   7  * All rights reserved.
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  *
  18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  19  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  20  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  21  * IN NO EVENT SHALL THE AUTHOR OR HIS RELATIVES BE LIABLE FOR ANY DIRECT,
  22  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  23  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  24  * SERVICES; LOSS OF MIND, USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
  26  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
  27  * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
  28  * THE POSSIBILITY OF SUCH DAMAGE.
  29  */
  30
  31 #include <sys/cdefs.h>
  32 __FBSDID("$FreeBSD$");
  33
  34 #include "opt_bpf.h"
  35 #include "opt_inet.h"
  36 #include "opt_inet6.h"
  37
  38 #include <sys/param.h>
  39 #include <sys/systm.h>
  40 #include <sys/bus.h>
  41 #include <sys/jail.h>
  42 #include <sys/kernel.h>
  43 #include <sys/limits.h>
  44 #include <sys/malloc.h>
  45 #include <sys/mbuf.h>
  46 #include <sys/module.h>
  47 #include <sys/priv.h>
  48 #include <sys/proc.h>
  49 #include <sys/protosw.h>
  50 #include <sys/socket.h>
  51 #include <sys/sockio.h>
  52 #include <sys/sysctl.h>
  53 #include <sys/syslog.h>
  54 #include <sys/taskqueue.h>
  55 #include <sys/counter.h>
  56
  57 #include <net/ethernet.h>
  58 #include <net/if.h>
  59 #include <net/if_var.h>
  60 #include <net/if_dl.h>
  61 #include <net/if_llatbl.h>
  62 #include <net/if_types.h>
  63 #include <net/route.h>
  64 #include <net/vnet.h>
  65
  66 #if defined(INET) || defined(INET6)
  67 #include <netinet/in.h>
  68 #include <netinet/in_var.h>
  69 #include <netinet/ip_carp.h>
  70 #include <netinet/ip.h>
  71 #include <machine/in_cksum.h>
  72 #endif
  73 #ifdef INET
  74 #include <netinet/ip_var.h>
  75 #include <netinet/if_ether.h>
  76 #endif
  77
  78 #ifdef INET6
  79 #include <netinet/icmp6.h>
  80 #include <netinet/ip6.h>
  81 #include <netinet6/in6_var.h>
  82 #include <netinet6/ip6_var.h>
  83 #include <netinet6/scope6_var.h>
  84 #include <netinet6/nd6.h>
  85 #endif
  86
  87 #include <crypto/sha1.h>
  88
  89 static MALLOC_DEFINE(M_CARP, "CARP", "CARP addresses");
  90
  91 struct carp_softc {
  92         struct ifnet            *sc_carpdev;    /* Pointer to parent ifnet. */
  93         struct ifaddr           **sc_ifas;      /* Our ifaddrs. */
  94         struct sockaddr_dl      sc_addr;        /* Our link level address. */
  95         struct callout          sc_ad_tmo;      /* Advertising timeout. */
  96 #ifdef INET
  97         struct callout          sc_md_tmo;      /* Master down timeout. */
  98 #endif
  99 #ifdef INET6
 100         struct callout          sc_md6_tmo;     /* XXX: Master down timeout. */
 101 #endif
 102         struct mtx              sc_mtx;
 103
 104         int                     sc_vhid;
 105         int                     sc_advskew;
 106         int                     sc_advbase;
 107
 108         int                     sc_naddrs;
 109         int                     sc_naddrs6;
 110         int                     sc_ifasiz;
 111         enum { INIT = 0, BACKUP, MASTER }       sc_state;
 112         int                     sc_suppress;
 113         int                     sc_sendad_errors;
 114 #define CARP_SENDAD_MAX_ERRORS  3
 115         int                     sc_sendad_success;
 116 #define CARP_SENDAD_MIN_SUCCESS 3
 117
 118         int                     sc_init_counter;
 119         uint64_t                sc_counter;
 120
 121         /* authentication */
 122 #define CARP_HMAC_PAD   64
 123         unsigned char sc_key[CARP_KEY_LEN];
 124         unsigned char sc_pad[CARP_HMAC_PAD];
 125         SHA1_CTX sc_sha1;
 126
 127         TAILQ_ENTRY(carp_softc) sc_list;        /* On the carp_if list. */
 128         LIST_ENTRY(carp_softc)  sc_next;        /* On the global list. */
 129 };
 130
 131 struct carp_if {
 132 #ifdef INET
 133         int     cif_naddrs;
 134 #endif
 135 #ifdef INET6
 136         int     cif_naddrs6;
 137 #endif
 138         TAILQ_HEAD(, carp_softc) cif_vrs;
 139 #ifdef INET
 140         struct ip_moptions       cif_imo;
 141 #endif
 142 #ifdef INET6
 143         struct ip6_moptions      cif_im6o;
 144 #endif
 145         struct ifnet    *cif_ifp;
 146         struct mtx      cif_mtx;
 147         uint32_t        cif_flags;
 148 #define CIF_PROMISC     0x00000001
 149 };
 150
 151 #define CARP_INET       0
 152 #define CARP_INET6      1
 153 static int proto_reg[] = {-1, -1};
 154
 155 /*
 156  * Brief design of carp(4).
 157  *
 158  * Any carp-capable ifnet may have a list of carp softcs hanging off
 159  * its ifp->if_carp pointer. Each softc represents one unique virtual
 160  * host id, or vhid. The softc has a back pointer to the ifnet. All
 161  * softcs are joined in a global list, which has quite limited use.
 162  *
 163  * Any interface address that takes part in CARP negotiation has a
 164  * pointer to the softc of its vhid, ifa->ifa_carp. That could be either
 165  * AF_INET or AF_INET6 address.
 166  *
 167  * Although, one can get the softc's backpointer to ifnet and traverse
 168  * through its ifp->if_addrhead queue to find all interface addresses
 169  * involved in CARP, we keep a growable array of ifaddr pointers. This
 170  * allows us to avoid grabbing the IF_ADDR_LOCK() in many traversals that
 171  * do calls into the network stack, thus avoiding LORs.
 172  *
 173  * Locking:
 174  *
 175  * Each softc has a lock sc_mtx. It is used to synchronise carp_input_c(),
 176  * callout-driven events and ioctl()s.
 177  *
 178  * To traverse the list of softcs on an ifnet we use CIF_LOCK() or carp_sx.
 179  * To traverse the global list we use the mutex carp_mtx.
 180  *
 181  * Known issues with locking:
 182  *
 183  * - Sending ad, we put the pointer to the softc in an mtag, and no reference
 184  *   counting is done on the softc.
 185  * - On module unload we may race (?) with packet processing thread
 186  *   dereferencing our function pointers.
 187  */
 188
 189 /* Accept incoming CARP packets. */
 190 static VNET_DEFINE(int, carp_allow) = 1;
 191 #define V_carp_allow    VNET(carp_allow)
 192
 193 /* Preempt slower nodes. */
 194 static VNET_DEFINE(int, carp_preempt) = 0;
 195 #define V_carp_preempt  VNET(carp_preempt)
 196
 197 /* Log level. */
 198 static VNET_DEFINE(int, carp_log) = 1;
 199 #define V_carp_log      VNET(carp_log)
 200
 201 /* Global advskew demotion. */
 202 static VNET_DEFINE(int, carp_demotion) = 0;
 203 #define V_carp_demotion VNET(carp_demotion)
 204
 205 /* Send error demotion factor. */
 206 static VNET_DEFINE(int, carp_senderr_adj) = CARP_MAXSKEW;
 207 #define V_carp_senderr_adj      VNET(carp_senderr_adj)
 208
 209 /* Iface down demotion factor. */
 210 static VNET_DEFINE(int, carp_ifdown_adj) = CARP_MAXSKEW;
 211 #define V_carp_ifdown_adj       VNET(carp_ifdown_adj)
 212
 213 static int carp_allow_sysctl(SYSCTL_HANDLER_ARGS);
 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_PROC(_net_inet_carp, OID_AUTO, allow,
 218     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW, 0, 0, carp_allow_sysctl, "I",
 219     "Accept incoming CARP packets");
 220 SYSCTL_INT(_net_inet_carp, OID_AUTO, preempt, CTLFLAG_VNET | CTLFLAG_RW,
 221     &VNET_NAME(carp_preempt), 0, "High-priority backup preemption mode");
 222 SYSCTL_INT(_net_inet_carp, OID_AUTO, log, CTLFLAG_VNET | CTLFLAG_RW,
 223     &VNET_NAME(carp_log), 0, "CARP log level");
 224 SYSCTL_PROC(_net_inet_carp, OID_AUTO, demotion,
 225     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW,
 226     0, 0, carp_demote_adj_sysctl, "I",
 227     "Adjust demotion factor (skew of advskew)");
 228 SYSCTL_INT(_net_inet_carp, OID_AUTO, senderr_demotion_factor,
 229     CTLFLAG_VNET | CTLFLAG_RW,
 230     &VNET_NAME(carp_senderr_adj), 0, "Send error demotion factor adjustment");
 231 SYSCTL_INT(_net_inet_carp, OID_AUTO, ifdown_demotion_factor,
 232     CTLFLAG_VNET | CTLFLAG_RW,
 233     &VNET_NAME(carp_ifdown_adj), 0,
 234     "Interface down demotion factor adjustment");
 235
 236 VNET_PCPUSTAT_DEFINE(struct carpstats, carpstats);
 237 VNET_PCPUSTAT_SYSINIT(carpstats);
 238 VNET_PCPUSTAT_SYSUNINIT(carpstats);
 239
 240 #define CARPSTATS_ADD(name, val)        \
 241     counter_u64_add(VNET(carpstats)[offsetof(struct carpstats, name) / \
 242         sizeof(uint64_t)], (val))
 243 #define CARPSTATS_INC(name)             CARPSTATS_ADD(name, 1)
 244
 245 SYSCTL_VNET_PCPUSTAT(_net_inet_carp, OID_AUTO, stats, struct carpstats,
 246     carpstats, "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 247
 248 #define CARP_LOCK_INIT(sc)      mtx_init(&(sc)->sc_mtx, "carp_softc",   \
 249         NULL, MTX_DEF)
 250 #define CARP_LOCK_DESTROY(sc)   mtx_destroy(&(sc)->sc_mtx)
 251 #define CARP_LOCK_ASSERT(sc)    mtx_assert(&(sc)->sc_mtx, MA_OWNED)
 252 #define CARP_LOCK(sc)           mtx_lock(&(sc)->sc_mtx)
 253 #define CARP_UNLOCK(sc)         mtx_unlock(&(sc)->sc_mtx)
 254 #define CIF_LOCK_INIT(cif)      mtx_init(&(cif)->cif_mtx, "carp_if",   \
 255         NULL, MTX_DEF)
 256 #define CIF_LOCK_DESTROY(cif)   mtx_destroy(&(cif)->cif_mtx)
 257 #define CIF_LOCK_ASSERT(cif)    mtx_assert(&(cif)->cif_mtx, MA_OWNED)
 258 #define CIF_LOCK(cif)           mtx_lock(&(cif)->cif_mtx)
 259 #define CIF_UNLOCK(cif)         mtx_unlock(&(cif)->cif_mtx)
 260 #define CIF_FREE(cif)   do {                            \
 261                 CIF_LOCK(cif);                          \
 262                 if (TAILQ_EMPTY(&(cif)->cif_vrs))       \
 263                         carp_free_if(cif);              \
 264                 else                                    \
 265                         CIF_UNLOCK(cif);                \
 266 } while (0)
 267
 268 #define CARP_LOG(...)   do {                            \
 269         if (V_carp_log > 0)                             \
 270                 log(LOG_INFO, "carp: " __VA_ARGS__);    \
 271 } while (0)
 272
 273 #define CARP_DEBUG(...) do {                            \
 274         if (V_carp_log > 1)                             \
 275                 log(LOG_DEBUG, __VA_ARGS__);            \
 276 } while (0)
 277
 278 #define IFNET_FOREACH_IFA(ifp, ifa)                                     \
 279         CK_STAILQ_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         KASSERT(mtx_owned(&ifp->if_carp->cif_mtx) ||                    \
 291             sx_xlocked(&carp_sx), ("cif_vrs not locked"));              \
 292         TAILQ_FOREACH((sc), &(ifp)->if_carp->cif_vrs, sc_list)
 293
 294 #define DEMOTE_ADVSKEW(sc)                                      \
 295     (((sc)->sc_advskew + V_carp_demotion > CARP_MAXSKEW) ?      \
 296     CARP_MAXSKEW : ((sc)->sc_advskew + V_carp_demotion))
 297
 298 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 299 static struct carp_softc
 300                 *carp_alloc(struct ifnet *);
 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, const char* reason);
 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                     const char* reason);
 311 static void     carp_send_ad(void *);
 312 static void     carp_send_ad_locked(struct carp_softc *);
 313 static void     carp_addroute(struct carp_softc *);
 314 static void     carp_ifa_addroute(struct ifaddr *);
 315 static void     carp_delroute(struct carp_softc *);
 316 static void     carp_ifa_delroute(struct ifaddr *);
 317 static void     carp_send_ad_all(void *, int);
 318 static void     carp_demote_adj(int, char *);
 319
 320 static LIST_HEAD(, carp_softc) carp_list;
 321 static struct mtx carp_mtx;
 322 static struct sx carp_sx;
 323 static struct task carp_sendall_task =
 324     TASK_INITIALIZER(0, carp_send_ad_all, NULL);
 325
 326 static void
 327 carp_hmac_prepare(struct carp_softc *sc)
 328 {
 329         uint8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 330         uint8_t vhid = sc->sc_vhid & 0xff;
 331         struct ifaddr *ifa;
 332         int i, found;
 333 #ifdef INET
 334         struct in_addr last, cur, in;
 335 #endif
 336 #ifdef INET6
 337         struct in6_addr last6, cur6, in6;
 338 #endif
 339
 340         CARP_LOCK_ASSERT(sc);
 341
 342         /* Compute ipad from key. */
 343         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 344         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 345         for (i = 0; i < sizeof(sc->sc_pad); i++)
 346                 sc->sc_pad[i] ^= 0x36;
 347
 348         /* Precompute first part of inner hash. */
 349         SHA1Init(&sc->sc_sha1);
 350         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 351         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 352         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 353         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 354 #ifdef INET
 355         cur.s_addr = 0;
 356         do {
 357                 found = 0;
 358                 last = cur;
 359                 cur.s_addr = 0xffffffff;
 360                 CARP_FOREACH_IFA(sc, ifa) {
 361                         in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 362                         if (ifa->ifa_addr->sa_family == AF_INET &&
 363                             ntohl(in.s_addr) > ntohl(last.s_addr) &&
 364                             ntohl(in.s_addr) < ntohl(cur.s_addr)) {
 365                                 cur.s_addr = in.s_addr;
 366                                 found++;
 367                         }
 368                 }
 369                 if (found)
 370                         SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur));
 371         } while (found);
 372 #endif /* INET */
 373 #ifdef INET6
 374         memset(&cur6, 0, sizeof(cur6));
 375         do {
 376                 found = 0;
 377                 last6 = cur6;
 378                 memset(&cur6, 0xff, sizeof(cur6));
 379                 CARP_FOREACH_IFA(sc, ifa) {
 380                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 381                         if (IN6_IS_SCOPE_EMBED(&in6))
 382                                 in6.s6_addr16[1] = 0;
 383                         if (ifa->ifa_addr->sa_family == AF_INET6 &&
 384                             memcmp(&in6, &last6, sizeof(in6)) > 0 &&
 385                             memcmp(&in6, &cur6, sizeof(in6)) < 0) {
 386                                 cur6 = in6;
 387                                 found++;
 388                         }
 389                 }
 390                 if (found)
 391                         SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6));
 392         } while (found);
 393 #endif /* INET6 */
 394
 395         /* convert ipad to opad */
 396         for (i = 0; i < sizeof(sc->sc_pad); i++)
 397                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 398 }
 399
 400 static void
 401 carp_hmac_generate(struct carp_softc *sc, uint32_t counter[2],
 402     unsigned char md[20])
 403 {
 404         SHA1_CTX sha1ctx;
 405
 406         CARP_LOCK_ASSERT(sc);
 407
 408         /* fetch first half of inner hash */
 409         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 410
 411         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 412         SHA1Final(md, &sha1ctx);
 413
 414         /* outer hash */
 415         SHA1Init(&sha1ctx);
 416         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 417         SHA1Update(&sha1ctx, md, 20);
 418         SHA1Final(md, &sha1ctx);
 419 }
 420
 421 static int
 422 carp_hmac_verify(struct carp_softc *sc, uint32_t counter[2],
 423     unsigned char md[20])
 424 {
 425         unsigned char md2[20];
 426
 427         CARP_LOCK_ASSERT(sc);
 428
 429         carp_hmac_generate(sc, counter, md2);
 430
 431         return (bcmp(md, md2, sizeof(md2)));
 432 }
 433
 434 /*
 435  * process input packet.
 436  * we have rearranged checks order compared to the rfc,
 437  * but it seems more efficient this way or not possible otherwise.
 438  */
 439 #ifdef INET
 440 int
 441 carp_input(struct mbuf **mp, int *offp, int proto)
 442 {
 443         struct mbuf *m = *mp;
 444         struct ip *ip = mtod(m, struct ip *);
 445         struct carp_header *ch;
 446         int iplen, len;
 447
 448         iplen = *offp;
 449         *mp = NULL;
 450
 451         CARPSTATS_INC(carps_ipackets);
 452
 453         if (!V_carp_allow) {
 454                 m_freem(m);
 455                 return (IPPROTO_DONE);
 456         }
 457
 458         /* verify that the IP TTL is 255.  */
 459         if (ip->ip_ttl != CARP_DFLTTL) {
 460                 CARPSTATS_INC(carps_badttl);
 461                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 462                     ip->ip_ttl,
 463                     m->m_pkthdr.rcvif->if_xname);
 464                 m_freem(m);
 465                 return (IPPROTO_DONE);
 466         }
 467
 468         iplen = ip->ip_hl << 2;
 469
 470         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 471                 CARPSTATS_INC(carps_badlen);
 472                 CARP_DEBUG("%s: received len %zd < sizeof(struct carp_header) "
 473                     "on %s\n", __func__, m->m_len - sizeof(struct ip),
 474                     m->m_pkthdr.rcvif->if_xname);
 475                 m_freem(m);
 476                 return (IPPROTO_DONE);
 477         }
 478
 479         if (iplen + sizeof(*ch) < m->m_len) {
 480                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 481                         CARPSTATS_INC(carps_hdrops);
 482                         CARP_DEBUG("%s: pullup failed\n", __func__);
 483                         return (IPPROTO_DONE);
 484                 }
 485                 ip = mtod(m, struct ip *);
 486         }
 487         ch = (struct carp_header *)((char *)ip + iplen);
 488
 489         /*
 490          * verify that the received packet length is
 491          * equal to the CARP header
 492          */
 493         len = iplen + sizeof(*ch);
 494         if (len > m->m_pkthdr.len) {
 495                 CARPSTATS_INC(carps_badlen);
 496                 CARP_DEBUG("%s: packet too short %d on %s\n", __func__,
 497                     m->m_pkthdr.len,
 498                     m->m_pkthdr.rcvif->if_xname);
 499                 m_freem(m);
 500                 return (IPPROTO_DONE);
 501         }
 502
 503         if ((m = m_pullup(m, len)) == NULL) {
 504                 CARPSTATS_INC(carps_hdrops);
 505                 return (IPPROTO_DONE);
 506         }
 507         ip = mtod(m, struct ip *);
 508         ch = (struct carp_header *)((char *)ip + iplen);
 509
 510         /* verify the CARP checksum */
 511         m->m_data += iplen;
 512         if (in_cksum(m, len - iplen)) {
 513                 CARPSTATS_INC(carps_badsum);
 514                 CARP_DEBUG("%s: checksum failed on %s\n", __func__,
 515                     m->m_pkthdr.rcvif->if_xname);
 516                 m_freem(m);
 517                 return (IPPROTO_DONE);
 518         }
 519         m->m_data -= iplen;
 520
 521         carp_input_c(m, ch, AF_INET);
 522         return (IPPROTO_DONE);
 523 }
 524 #endif
 525
 526 #ifdef INET6
 527 int
 528 carp6_input(struct mbuf **mp, int *offp, int proto)
 529 {
 530         struct mbuf *m = *mp;
 531         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 532         struct carp_header *ch;
 533         u_int len;
 534
 535         CARPSTATS_INC(carps_ipackets6);
 536
 537         if (!V_carp_allow) {
 538                 m_freem(m);
 539                 return (IPPROTO_DONE);
 540         }
 541
 542         /* check if received on a valid carp interface */
 543         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 544                 CARPSTATS_INC(carps_badif);
 545                 CARP_DEBUG("%s: packet received on non-carp interface: %s\n",
 546                     __func__, m->m_pkthdr.rcvif->if_xname);
 547                 m_freem(m);
 548                 return (IPPROTO_DONE);
 549         }
 550
 551         /* verify that the IP TTL is 255 */
 552         if (ip6->ip6_hlim != CARP_DFLTTL) {
 553                 CARPSTATS_INC(carps_badttl);
 554                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 555                     ip6->ip6_hlim, m->m_pkthdr.rcvif->if_xname);
 556                 m_freem(m);
 557                 return (IPPROTO_DONE);
 558         }
 559
 560         /* verify that we have a complete carp packet */
 561         len = m->m_len;
 562         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 563         if (ch == NULL) {
 564                 CARPSTATS_INC(carps_badlen);
 565                 CARP_DEBUG("%s: packet size %u too small\n", __func__, len);
 566                 return (IPPROTO_DONE);
 567         }
 568
 569
 570         /* verify the CARP checksum */
 571         m->m_data += *offp;
 572         if (in_cksum(m, sizeof(*ch))) {
 573                 CARPSTATS_INC(carps_badsum);
 574                 CARP_DEBUG("%s: checksum failed, on %s\n", __func__,
 575                     m->m_pkthdr.rcvif->if_xname);
 576                 m_freem(m);
 577                 return (IPPROTO_DONE);
 578         }
 579         m->m_data -= *offp;
 580
 581         carp_input_c(m, ch, AF_INET6);
 582         return (IPPROTO_DONE);
 583 }
 584 #endif /* INET6 */
 585
 586 /*
 587  * This routine should not be necessary at all, but some switches
 588  * (VMWare ESX vswitches) can echo our own packets back at us,
 589  * and we must ignore them or they will cause us to drop out of
 590  * MASTER mode.
 591  *
 592  * We cannot catch all cases of network loops.  Instead, what we
 593  * do here is catch any packet that arrives with a carp header
 594  * with a VHID of 0, that comes from an address that is our own.
 595  * These packets are by definition "from us" (even if they are from
 596  * a misconfigured host that is pretending to be us).
 597  *
 598  * The VHID test is outside this mini-function.
 599  */
 600 static int
 601 carp_source_is_self(struct mbuf *m, struct ifaddr *ifa, sa_family_t af)
 602 {
 603 #ifdef INET
 604         struct ip *ip4;
 605         struct in_addr in4;
 606 #endif
 607 #ifdef INET6
 608         struct ip6_hdr *ip6;
 609         struct in6_addr in6;
 610 #endif
 611
 612         switch (af) {
 613 #ifdef INET
 614         case AF_INET:
 615                 ip4 = mtod(m, struct ip *);
 616                 in4 = ifatoia(ifa)->ia_addr.sin_addr;
 617                 return (in4.s_addr == ip4->ip_src.s_addr);
 618 #endif
 619 #ifdef INET6
 620         case AF_INET6:
 621                 ip6 = mtod(m, struct ip6_hdr *);
 622                 in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 623                 return (memcmp(&in6, &ip6->ip6_src, sizeof(in6)) == 0);
 624 #endif
 625         default:
 626                 break;
 627         }
 628         return (0);
 629 }
 630
 631 static void
 632 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 633 {
 634         struct ifnet *ifp = m->m_pkthdr.rcvif;
 635         struct ifaddr *ifa, *match;
 636         struct carp_softc *sc;
 637         uint64_t tmp_counter;
 638         struct timeval sc_tv, ch_tv;
 639         int error;
 640
 641         /*
 642          * Verify that the VHID is valid on the receiving interface.
 643          *
 644          * There should be just one match.  If there are none
 645          * the VHID is not valid and we drop the packet.  If
 646          * there are multiple VHID matches, take just the first
 647          * one, for compatibility with previous code.  While we're
 648          * scanning, check for obvious loops in the network topology
 649          * (these should never happen, and as noted above, we may
 650          * miss real loops; this is just a double-check).
 651          */
 652         IF_ADDR_RLOCK(ifp);
 653         error = 0;
 654         match = NULL;
 655         IFNET_FOREACH_IFA(ifp, ifa) {
 656                 if (match == NULL && ifa->ifa_carp != NULL &&
 657                     ifa->ifa_addr->sa_family == af &&
 658                     ifa->ifa_carp->sc_vhid == ch->carp_vhid)
 659                         match = ifa;
 660                 if (ch->carp_vhid == 0 && carp_source_is_self(m, ifa, af))
 661                         error = ELOOP;
 662         }
 663         ifa = error ? NULL : match;
 664         if (ifa != NULL)
 665                 ifa_ref(ifa);
 666         IF_ADDR_RUNLOCK(ifp);
 667
 668         if (ifa == NULL) {
 669                 if (error == ELOOP) {
 670                         CARP_DEBUG("dropping looped packet on interface %s\n",
 671                             ifp->if_xname);
 672                         CARPSTATS_INC(carps_badif);     /* ??? */
 673                 } else {
 674                         CARPSTATS_INC(carps_badvhid);
 675                 }
 676                 m_freem(m);
 677                 return;
 678         }
 679
 680         /* verify the CARP version. */
 681         if (ch->carp_version != CARP_VERSION) {
 682                 CARPSTATS_INC(carps_badver);
 683                 CARP_DEBUG("%s: invalid version %d\n", ifp->if_xname,
 684                     ch->carp_version);
 685                 ifa_free(ifa);
 686                 m_freem(m);
 687                 return;
 688         }
 689
 690         sc = ifa->ifa_carp;
 691         CARP_LOCK(sc);
 692         ifa_free(ifa);
 693
 694         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 695                 CARPSTATS_INC(carps_badauth);
 696                 CARP_DEBUG("%s: incorrect hash for VHID %u@%s\n", __func__,
 697                     sc->sc_vhid, ifp->if_xname);
 698                 goto out;
 699         }
 700
 701         tmp_counter = ntohl(ch->carp_counter[0]);
 702         tmp_counter = tmp_counter<<32;
 703         tmp_counter += ntohl(ch->carp_counter[1]);
 704
 705         /* XXX Replay protection goes here */
 706
 707         sc->sc_init_counter = 0;
 708         sc->sc_counter = tmp_counter;
 709
 710         sc_tv.tv_sec = sc->sc_advbase;
 711         sc_tv.tv_usec = DEMOTE_ADVSKEW(sc) * 1000000 / 256;
 712         ch_tv.tv_sec = ch->carp_advbase;
 713         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 714
 715         switch (sc->sc_state) {
 716         case INIT:
 717                 break;
 718         case MASTER:
 719                 /*
 720                  * If we receive an advertisement from a master who's going to
 721                  * be more frequent than us, go into BACKUP state.
 722                  */
 723                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 724                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 725                         callout_stop(&sc->sc_ad_tmo);
 726                         carp_set_state(sc, BACKUP,
 727                             "more frequent advertisement received");
 728                         carp_setrun(sc, 0);
 729                         carp_delroute(sc);
 730                 }
 731                 break;
 732         case BACKUP:
 733                 /*
 734                  * If we're pre-empting masters who advertise slower than us,
 735                  * and this one claims to be slower, treat him as down.
 736                  */
 737                 if (V_carp_preempt && timevalcmp(&sc_tv, &ch_tv, <)) {
 738                         carp_master_down_locked(sc,
 739                             "preempting a slower master");
 740                         break;
 741                 }
 742
 743                 /*
 744                  *  If the master is going to advertise at such a low frequency
 745                  *  that he's guaranteed to time out, we'd might as well just
 746                  *  treat him as timed out now.
 747                  */
 748                 sc_tv.tv_sec = sc->sc_advbase * 3;
 749                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 750                         carp_master_down_locked(sc, "master will time out");
 751                         break;
 752                 }
 753
 754                 /*
 755                  * Otherwise, we reset the counter and wait for the next
 756                  * advertisement.
 757                  */
 758                 carp_setrun(sc, af);
 759                 break;
 760         }
 761
 762 out:
 763         CARP_UNLOCK(sc);
 764         m_freem(m);
 765 }
 766
 767 static int
 768 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 769 {
 770         struct m_tag *mtag;
 771
 772         if (sc->sc_init_counter) {
 773                 /* this could also be seconds since unix epoch */
 774                 sc->sc_counter = arc4random();
 775                 sc->sc_counter = sc->sc_counter << 32;
 776                 sc->sc_counter += arc4random();
 777         } else
 778                 sc->sc_counter++;
 779
 780         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 781         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 782
 783         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 784
 785         /* Tag packet for carp_output */
 786         if ((mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct carp_softc *),
 787             M_NOWAIT)) == NULL) {
 788                 m_freem(m);
 789                 CARPSTATS_INC(carps_onomem);
 790                 return (ENOMEM);
 791         }
 792         bcopy(&sc, mtag + 1, sizeof(sc));
 793         m_tag_prepend(m, mtag);
 794
 795         return (0);
 796 }
 797
 798 /*
 799  * To avoid LORs and possible recursions this function shouldn't
 800  * be called directly, but scheduled via taskqueue.
 801  */
 802 static void
 803 carp_send_ad_all(void *ctx __unused, int pending __unused)
 804 {
 805         struct carp_softc *sc;
 806
 807         mtx_lock(&carp_mtx);
 808         LIST_FOREACH(sc, &carp_list, sc_next)
 809                 if (sc->sc_state == MASTER) {
 810                         CARP_LOCK(sc);
 811                         CURVNET_SET(sc->sc_carpdev->if_vnet);
 812                         carp_send_ad_locked(sc);
 813                         CURVNET_RESTORE();
 814                         CARP_UNLOCK(sc);
 815                 }
 816         mtx_unlock(&carp_mtx);
 817 }
 818
 819 /* Send a periodic advertisement, executed in callout context. */
 820 static void
 821 carp_send_ad(void *v)
 822 {
 823         struct carp_softc *sc = v;
 824
 825         CARP_LOCK_ASSERT(sc);
 826         CURVNET_SET(sc->sc_carpdev->if_vnet);
 827         carp_send_ad_locked(sc);
 828         CURVNET_RESTORE();
 829         CARP_UNLOCK(sc);
 830 }
 831
 832 static void
 833 carp_send_ad_error(struct carp_softc *sc, int error)
 834 {
 835
 836         if (error) {
 837                 if (sc->sc_sendad_errors < INT_MAX)
 838                         sc->sc_sendad_errors++;
 839                 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 840                         static const char fmt[] = "send error %d on %s";
 841                         char msg[sizeof(fmt) + IFNAMSIZ];
 842
 843                         sprintf(msg, fmt, error, sc->sc_carpdev->if_xname);
 844                         carp_demote_adj(V_carp_senderr_adj, msg);
 845                 }
 846                 sc->sc_sendad_success = 0;
 847         } else {
 848                 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS &&
 849                     ++sc->sc_sendad_success >= CARP_SENDAD_MIN_SUCCESS) {
 850                         static const char fmt[] = "send ok on %s";
 851                         char msg[sizeof(fmt) + IFNAMSIZ];
 852
 853                         sprintf(msg, fmt, sc->sc_carpdev->if_xname);
 854                         carp_demote_adj(-V_carp_senderr_adj, msg);
 855                         sc->sc_sendad_errors = 0;
 856                 } else
 857                         sc->sc_sendad_errors = 0;
 858         }
 859 }
 860
 861 /*
 862  * Pick the best ifaddr on the given ifp for sending CARP
 863  * advertisements.
 864  *
 865  * "Best" here is defined by ifa_preferred().  This function is much
 866  * much like ifaof_ifpforaddr() except that we just use ifa_preferred().
 867  *
 868  * (This could be simplified to return the actual address, except that
 869  * it has a different format in AF_INET and AF_INET6.)
 870  */
 871 static struct ifaddr *
 872 carp_best_ifa(int af, struct ifnet *ifp)
 873 {
 874         struct ifaddr *ifa, *best;
 875
 876         if (af >= AF_MAX)
 877                 return (NULL);
 878         best = NULL;
 879         IF_ADDR_RLOCK(ifp);
 880         CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
 881                 if (ifa->ifa_addr->sa_family == af &&
 882                     (best == NULL || ifa_preferred(best, ifa)))
 883                         best = ifa;
 884         }
 885         IF_ADDR_RUNLOCK(ifp);
 886         if (best != NULL)
 887                 ifa_ref(best);
 888         return (best);
 889 }
 890
 891 static void
 892 carp_send_ad_locked(struct carp_softc *sc)
 893 {
 894         struct carp_header ch;
 895         struct timeval tv;
 896         struct ifaddr *ifa;
 897         struct carp_header *ch_ptr;
 898         struct mbuf *m;
 899         int len, advskew;
 900
 901         CARP_LOCK_ASSERT(sc);
 902
 903         advskew = DEMOTE_ADVSKEW(sc);
 904         tv.tv_sec = sc->sc_advbase;
 905         tv.tv_usec = advskew * 1000000 / 256;
 906
 907         ch.carp_version = CARP_VERSION;
 908         ch.carp_type = CARP_ADVERTISEMENT;
 909         ch.carp_vhid = sc->sc_vhid;
 910         ch.carp_advbase = sc->sc_advbase;
 911         ch.carp_advskew = advskew;
 912         ch.carp_authlen = 7;    /* XXX DEFINE */
 913         ch.carp_pad1 = 0;       /* must be zero */
 914         ch.carp_cksum = 0;
 915
 916         /* XXXGL: OpenBSD picks first ifaddr with needed family. */
 917
 918 #ifdef INET
 919         if (sc->sc_naddrs) {
 920                 struct ip *ip;
 921
 922                 m = m_gethdr(M_NOWAIT, MT_DATA);
 923                 if (m == NULL) {
 924                         CARPSTATS_INC(carps_onomem);
 925                         goto resched;
 926                 }
 927                 len = sizeof(*ip) + sizeof(ch);
 928                 m->m_pkthdr.len = len;
 929                 m->m_pkthdr.rcvif = NULL;
 930                 m->m_len = len;
 931                 M_ALIGN(m, m->m_len);
 932                 m->m_flags |= M_MCAST;
 933                 ip = mtod(m, struct ip *);
 934                 ip->ip_v = IPVERSION;
 935                 ip->ip_hl = sizeof(*ip) >> 2;
 936                 ip->ip_tos = IPTOS_LOWDELAY;
 937                 ip->ip_len = htons(len);
 938                 ip->ip_off = htons(IP_DF);
 939                 ip->ip_ttl = CARP_DFLTTL;
 940                 ip->ip_p = IPPROTO_CARP;
 941                 ip->ip_sum = 0;
 942                 ip_fillid(ip);
 943
 944                 ifa = carp_best_ifa(AF_INET, sc->sc_carpdev);
 945                 if (ifa != NULL) {
 946                         ip->ip_src.s_addr =
 947                             ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 948                         ifa_free(ifa);
 949                 } else
 950                         ip->ip_src.s_addr = 0;
 951                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 952
 953                 ch_ptr = (struct carp_header *)(&ip[1]);
 954                 bcopy(&ch, ch_ptr, sizeof(ch));
 955                 if (carp_prepare_ad(m, sc, ch_ptr))
 956                         goto resched;
 957
 958                 m->m_data += sizeof(*ip);
 959                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip));
 960                 m->m_data -= sizeof(*ip);
 961
 962                 CARPSTATS_INC(carps_opackets);
 963
 964                 carp_send_ad_error(sc, ip_output(m, NULL, NULL, IP_RAWOUTPUT,
 965                     &sc->sc_carpdev->if_carp->cif_imo, NULL));
 966         }
 967 #endif /* INET */
 968 #ifdef INET6
 969         if (sc->sc_naddrs6) {
 970                 struct ip6_hdr *ip6;
 971
 972                 m = m_gethdr(M_NOWAIT, MT_DATA);
 973                 if (m == NULL) {
 974                         CARPSTATS_INC(carps_onomem);
 975                         goto resched;
 976                 }
 977                 len = sizeof(*ip6) + sizeof(ch);
 978                 m->m_pkthdr.len = len;
 979                 m->m_pkthdr.rcvif = NULL;
 980                 m->m_len = len;
 981                 M_ALIGN(m, m->m_len);
 982                 m->m_flags |= M_MCAST;
 983                 ip6 = mtod(m, struct ip6_hdr *);
 984                 bzero(ip6, sizeof(*ip6));
 985                 ip6->ip6_vfc |= IPV6_VERSION;
 986                 ip6->ip6_hlim = CARP_DFLTTL;
 987                 ip6->ip6_nxt = IPPROTO_CARP;
 988
 989                 /* set the source address */
 990                 ifa = carp_best_ifa(AF_INET6, sc->sc_carpdev);
 991                 if (ifa != NULL) {
 992                         bcopy(IFA_IN6(ifa), &ip6->ip6_src,
 993                             sizeof(struct in6_addr));
 994                         ifa_free(ifa);
 995                 } else
 996                         /* This should never happen with IPv6. */
 997                         bzero(&ip6->ip6_src, sizeof(struct in6_addr));
 998
 999                 /* Set the multicast destination. */
1000                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
1001                 ip6->ip6_dst.s6_addr8[15] = 0x12;
1002                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
1003                         m_freem(m);
1004                         CARP_DEBUG("%s: in6_setscope failed\n", __func__);
1005                         goto resched;
1006                 }
1007
1008                 ch_ptr = (struct carp_header *)(&ip6[1]);
1009                 bcopy(&ch, ch_ptr, sizeof(ch));
1010                 if (carp_prepare_ad(m, sc, ch_ptr))
1011                         goto resched;
1012
1013                 m->m_data += sizeof(*ip6);
1014                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip6));
1015                 m->m_data -= sizeof(*ip6);
1016
1017                 CARPSTATS_INC(carps_opackets6);
1018
1019                 carp_send_ad_error(sc, ip6_output(m, NULL, NULL, 0,
1020                     &sc->sc_carpdev->if_carp->cif_im6o, NULL, NULL));
1021         }
1022 #endif /* INET6 */
1023
1024 resched:
1025         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), carp_send_ad, sc);
1026 }
1027
1028 static void
1029 carp_addroute(struct carp_softc *sc)
1030 {
1031         struct ifaddr *ifa;
1032
1033         CARP_FOREACH_IFA(sc, ifa)
1034                 carp_ifa_addroute(ifa);
1035 }
1036
1037 static void
1038 carp_ifa_addroute(struct ifaddr *ifa)
1039 {
1040
1041         switch (ifa->ifa_addr->sa_family) {
1042 #ifdef INET
1043         case AF_INET:
1044                 in_addprefix(ifatoia(ifa), RTF_UP);
1045                 ifa_add_loopback_route(ifa,
1046                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
1047                 break;
1048 #endif
1049 #ifdef INET6
1050         case AF_INET6:
1051                 ifa_add_loopback_route(ifa,
1052                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
1053                 nd6_add_ifa_lle(ifatoia6(ifa));
1054                 break;
1055 #endif
1056         }
1057 }
1058
1059 static void
1060 carp_delroute(struct carp_softc *sc)
1061 {
1062         struct ifaddr *ifa;
1063
1064         CARP_FOREACH_IFA(sc, ifa)
1065                 carp_ifa_delroute(ifa);
1066 }
1067
1068 static void
1069 carp_ifa_delroute(struct ifaddr *ifa)
1070 {
1071
1072         switch (ifa->ifa_addr->sa_family) {
1073 #ifdef INET
1074         case AF_INET:
1075                 ifa_del_loopback_route(ifa,
1076                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
1077                 in_scrubprefix(ifatoia(ifa), LLE_STATIC);
1078                 break;
1079 #endif
1080 #ifdef INET6
1081         case AF_INET6:
1082                 ifa_del_loopback_route(ifa,
1083                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
1084                 nd6_rem_ifa_lle(ifatoia6(ifa), 1);
1085                 break;
1086 #endif
1087         }
1088 }
1089
1090 int
1091 carp_master(struct ifaddr *ifa)
1092 {
1093         struct carp_softc *sc = ifa->ifa_carp;
1094
1095         return (sc->sc_state == MASTER);
1096 }
1097
1098 #ifdef INET
1099 /*
1100  * Broadcast a gratuitous ARP request containing
1101  * the virtual router MAC address for each IP address
1102  * associated with the virtual router.
1103  */
1104 static void
1105 carp_send_arp(struct carp_softc *sc)
1106 {
1107         struct ifaddr *ifa;
1108         struct in_addr addr;
1109
1110         CARP_FOREACH_IFA(sc, ifa) {
1111                 if (ifa->ifa_addr->sa_family != AF_INET)
1112                         continue;
1113                 addr = ((struct sockaddr_in *)ifa->ifa_addr)->sin_addr;
1114                 arp_announce_ifaddr(sc->sc_carpdev, addr, LLADDR(&sc->sc_addr));
1115         }
1116 }
1117
1118 int
1119 carp_iamatch(struct ifaddr *ifa, uint8_t **enaddr)
1120 {
1121         struct carp_softc *sc = ifa->ifa_carp;
1122
1123         if (sc->sc_state == MASTER) {
1124                 *enaddr = LLADDR(&sc->sc_addr);
1125                 return (1);
1126         }
1127
1128         return (0);
1129 }
1130 #endif
1131
1132 #ifdef INET6
1133 static void
1134 carp_send_na(struct carp_softc *sc)
1135 {
1136         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1137         struct ifaddr *ifa;
1138         struct in6_addr *in6;
1139
1140         CARP_FOREACH_IFA(sc, ifa) {
1141                 if (ifa->ifa_addr->sa_family != AF_INET6)
1142                         continue;
1143
1144                 in6 = IFA_IN6(ifa);
1145                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1146                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1147                 DELAY(1000);    /* XXX */
1148         }
1149 }
1150
1151 /*
1152  * Returns ifa in case it's a carp address and it is MASTER, or if the address
1153  * matches and is not a carp address.  Returns NULL otherwise.
1154  */
1155 struct ifaddr *
1156 carp_iamatch6(struct ifnet *ifp, struct in6_addr *taddr)
1157 {
1158         struct ifaddr *ifa;
1159
1160         ifa = NULL;
1161         IF_ADDR_RLOCK(ifp);
1162         CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
1163                 if (ifa->ifa_addr->sa_family != AF_INET6)
1164                         continue;
1165                 if (!IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa)))
1166                         continue;
1167                 if (ifa->ifa_carp && ifa->ifa_carp->sc_state != MASTER)
1168                         ifa = NULL;
1169                 else
1170                         ifa_ref(ifa);
1171                 break;
1172         }
1173         IF_ADDR_RUNLOCK(ifp);
1174
1175         return (ifa);
1176 }
1177
1178 caddr_t
1179 carp_macmatch6(struct ifnet *ifp, struct mbuf *m, const struct in6_addr *taddr)
1180 {
1181         struct ifaddr *ifa;
1182
1183         IF_ADDR_RLOCK(ifp);
1184         IFNET_FOREACH_IFA(ifp, ifa)
1185                 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1186                     IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1187                         struct carp_softc *sc = ifa->ifa_carp;
1188                         struct m_tag *mtag;
1189
1190                         IF_ADDR_RUNLOCK(ifp);
1191
1192                         mtag = m_tag_get(PACKET_TAG_CARP,
1193                             sizeof(struct carp_softc *), M_NOWAIT);
1194                         if (mtag == NULL)
1195                                 /* Better a bit than nothing. */
1196                                 return (LLADDR(&sc->sc_addr));
1197
1198                         bcopy(&sc, mtag + 1, sizeof(sc));
1199                         m_tag_prepend(m, mtag);
1200
1201                         return (LLADDR(&sc->sc_addr));
1202                 }
1203         IF_ADDR_RUNLOCK(ifp);
1204
1205         return (NULL);
1206 }
1207 #endif /* INET6 */
1208
1209 int
1210 carp_forus(struct ifnet *ifp, u_char *dhost)
1211 {
1212         struct carp_softc *sc;
1213         uint8_t *ena = dhost;
1214
1215         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1216                 return (0);
1217
1218         CIF_LOCK(ifp->if_carp);
1219         IFNET_FOREACH_CARP(ifp, sc) {
1220                 CARP_LOCK(sc);
1221                 if (sc->sc_state == MASTER && !bcmp(dhost, LLADDR(&sc->sc_addr),
1222                     ETHER_ADDR_LEN)) {
1223                         CARP_UNLOCK(sc);
1224                         CIF_UNLOCK(ifp->if_carp);
1225                         return (1);
1226                 }
1227                 CARP_UNLOCK(sc);
1228         }
1229         CIF_UNLOCK(ifp->if_carp);
1230
1231         return (0);
1232 }
1233
1234 /* Master down timeout event, executed in callout context. */
1235 static void
1236 carp_master_down(void *v)
1237 {
1238         struct carp_softc *sc = v;
1239
1240         CARP_LOCK_ASSERT(sc);
1241
1242         CURVNET_SET(sc->sc_carpdev->if_vnet);
1243         if (sc->sc_state == BACKUP) {
1244                 carp_master_down_locked(sc, "master timed out");
1245         }
1246         CURVNET_RESTORE();
1247
1248         CARP_UNLOCK(sc);
1249 }
1250
1251 static void
1252 carp_master_down_locked(struct carp_softc *sc, const char *reason)
1253 {
1254
1255         CARP_LOCK_ASSERT(sc);
1256
1257         switch (sc->sc_state) {
1258         case BACKUP:
1259                 carp_set_state(sc, MASTER, reason);
1260                 carp_send_ad_locked(sc);
1261 #ifdef INET
1262                 carp_send_arp(sc);
1263 #endif
1264 #ifdef INET6
1265                 carp_send_na(sc);
1266 #endif
1267                 carp_setrun(sc, 0);
1268                 carp_addroute(sc);
1269                 break;
1270         case INIT:
1271         case MASTER:
1272 #ifdef INVARIANTS
1273                 panic("carp: VHID %u@%s: master_down event in %s state\n",
1274                     sc->sc_vhid,
1275                     sc->sc_carpdev->if_xname,
1276                     sc->sc_state ? "MASTER" : "INIT");
1277 #endif
1278                 break;
1279         }
1280 }
1281
1282 /*
1283  * When in backup state, af indicates whether to reset the master down timer
1284  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1285  */
1286 static void
1287 carp_setrun(struct carp_softc *sc, sa_family_t af)
1288 {
1289         struct timeval tv;
1290
1291         CARP_LOCK_ASSERT(sc);
1292
1293         if ((sc->sc_carpdev->if_flags & IFF_UP) == 0 ||
1294             sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1295             (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) ||
1296             !V_carp_allow)
1297                 return;
1298
1299         switch (sc->sc_state) {
1300         case INIT:
1301                 carp_set_state(sc, BACKUP, "initialization complete");
1302                 carp_setrun(sc, 0);
1303                 break;
1304         case BACKUP:
1305                 callout_stop(&sc->sc_ad_tmo);
1306                 tv.tv_sec = 3 * sc->sc_advbase;
1307                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1308                 switch (af) {
1309 #ifdef INET
1310                 case AF_INET:
1311                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1312                             carp_master_down, sc);
1313                         break;
1314 #endif
1315 #ifdef INET6
1316                 case AF_INET6:
1317                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1318                             carp_master_down, sc);
1319                         break;
1320 #endif
1321                 default:
1322 #ifdef INET
1323                         if (sc->sc_naddrs)
1324                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1325                                     carp_master_down, sc);
1326 #endif
1327 #ifdef INET6
1328                         if (sc->sc_naddrs6)
1329                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1330                                     carp_master_down, sc);
1331 #endif
1332                         break;
1333                 }
1334                 break;
1335         case MASTER:
1336                 tv.tv_sec = sc->sc_advbase;
1337                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1338                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1339                     carp_send_ad, sc);
1340                 break;
1341         }
1342 }
1343
1344 /*
1345  * Setup multicast structures.
1346  */
1347 static int
1348 carp_multicast_setup(struct carp_if *cif, sa_family_t sa)
1349 {
1350         struct ifnet *ifp = cif->cif_ifp;
1351         int error = 0;
1352
1353         switch (sa) {
1354 #ifdef INET
1355         case AF_INET:
1356             {
1357                 struct ip_moptions *imo = &cif->cif_imo;
1358                 struct in_addr addr;
1359
1360                 if (imo->imo_membership)
1361                         return (0);
1362
1363                 imo->imo_membership = (struct in_multi **)malloc(
1364                     (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
1365                     M_WAITOK);
1366                 imo->imo_mfilters = NULL;
1367                 imo->imo_max_memberships = IP_MIN_MEMBERSHIPS;
1368                 imo->imo_multicast_vif = -1;
1369
1370                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1371                 if ((error = in_joingroup(ifp, &addr, NULL,
1372                     &imo->imo_membership[0])) != 0) {
1373                         free(imo->imo_membership, M_CARP);
1374                         break;
1375                 }
1376                 imo->imo_num_memberships++;
1377                 imo->imo_multicast_ifp = ifp;
1378                 imo->imo_multicast_ttl = CARP_DFLTTL;
1379                 imo->imo_multicast_loop = 0;
1380                 break;
1381            }
1382 #endif
1383 #ifdef INET6
1384         case AF_INET6:
1385             {
1386                 struct ip6_moptions *im6o = &cif->cif_im6o;
1387                 struct in6_addr in6;
1388                 struct in6_multi *in6m;
1389
1390                 if (im6o->im6o_membership)
1391                         return (0);
1392
1393                 im6o->im6o_membership = (struct in6_multi **)malloc(
1394                     (sizeof(struct in6_multi *) * IPV6_MIN_MEMBERSHIPS), M_CARP,
1395                     M_ZERO | M_WAITOK);
1396                 im6o->im6o_mfilters = NULL;
1397                 im6o->im6o_max_memberships = IPV6_MIN_MEMBERSHIPS;
1398                 im6o->im6o_multicast_hlim = CARP_DFLTTL;
1399                 im6o->im6o_multicast_ifp = ifp;
1400
1401                 /* Join IPv6 CARP multicast group. */
1402                 bzero(&in6, sizeof(in6));
1403                 in6.s6_addr16[0] = htons(0xff02);
1404                 in6.s6_addr8[15] = 0x12;
1405                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1406                         free(im6o->im6o_membership, M_CARP);
1407                         break;
1408                 }
1409                 in6m = NULL;
1410                 if ((error = in6_joingroup(ifp, &in6, NULL, &in6m, 0)) != 0) {
1411                         free(im6o->im6o_membership, M_CARP);
1412                         break;
1413                 }
1414                 im6o->im6o_membership[0] = in6m;
1415                 im6o->im6o_num_memberships++;
1416
1417                 /* Join solicited multicast address. */
1418                 bzero(&in6, sizeof(in6));
1419                 in6.s6_addr16[0] = htons(0xff02);
1420                 in6.s6_addr32[1] = 0;
1421                 in6.s6_addr32[2] = htonl(1);
1422                 in6.s6_addr32[3] = 0;
1423                 in6.s6_addr8[12] = 0xff;
1424                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1425                         in6_leavegroup(im6o->im6o_membership[0], NULL);
1426                         free(im6o->im6o_membership, M_CARP);
1427                         break;
1428                 }
1429                 in6m = NULL;
1430                 if ((error = in6_joingroup(ifp, &in6, NULL, &in6m, 0)) != 0) {
1431                         in6_leavegroup(im6o->im6o_membership[0], NULL);
1432                         free(im6o->im6o_membership, M_CARP);
1433                         break;
1434                 }
1435                 im6o->im6o_membership[1] = in6m;
1436                 im6o->im6o_num_memberships++;
1437                 break;
1438             }
1439 #endif
1440         }
1441
1442         return (error);
1443 }
1444
1445 /*
1446  * Free multicast structures.
1447  */
1448 static void
1449 carp_multicast_cleanup(struct carp_if *cif, sa_family_t sa)
1450 {
1451
1452         sx_assert(&carp_sx, SA_XLOCKED);
1453
1454         switch (sa) {
1455 #ifdef INET
1456         case AF_INET:
1457                 if (cif->cif_naddrs == 0) {
1458                         struct ip_moptions *imo = &cif->cif_imo;
1459
1460                         in_leavegroup(imo->imo_membership[0], NULL);
1461                         KASSERT(imo->imo_mfilters == NULL,
1462                             ("%s: imo_mfilters != NULL", __func__));
1463                         free(imo->imo_membership, M_CARP);
1464                         imo->imo_membership = NULL;
1465
1466                 }
1467                 break;
1468 #endif
1469 #ifdef INET6
1470         case AF_INET6:
1471                 if (cif->cif_naddrs6 == 0) {
1472                         struct ip6_moptions *im6o = &cif->cif_im6o;
1473
1474                         in6_leavegroup(im6o->im6o_membership[0], NULL);
1475                         in6_leavegroup(im6o->im6o_membership[1], NULL);
1476                         KASSERT(im6o->im6o_mfilters == NULL,
1477                             ("%s: im6o_mfilters != NULL", __func__));
1478                         free(im6o->im6o_membership, M_CARP);
1479                         im6o->im6o_membership = NULL;
1480                 }
1481                 break;
1482 #endif
1483         }
1484 }
1485
1486 int
1487 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa)
1488 {
1489         struct m_tag *mtag;
1490         struct carp_softc *sc;
1491
1492         if (!sa)
1493                 return (0);
1494
1495         switch (sa->sa_family) {
1496 #ifdef INET
1497         case AF_INET:
1498                 break;
1499 #endif
1500 #ifdef INET6
1501         case AF_INET6:
1502                 break;
1503 #endif
1504         default:
1505                 return (0);
1506         }
1507
1508         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
1509         if (mtag == NULL)
1510                 return (0);
1511
1512         bcopy(mtag + 1, &sc, sizeof(sc));
1513
1514         /* Set the source MAC address to the Virtual Router MAC Address. */
1515         switch (ifp->if_type) {
1516         case IFT_ETHER:
1517         case IFT_BRIDGE:
1518         case IFT_L2VLAN: {
1519                         struct ether_header *eh;
1520
1521                         eh = mtod(m, struct ether_header *);
1522                         eh->ether_shost[0] = 0;
1523                         eh->ether_shost[1] = 0;
1524                         eh->ether_shost[2] = 0x5e;
1525                         eh->ether_shost[3] = 0;
1526                         eh->ether_shost[4] = 1;
1527                         eh->ether_shost[5] = sc->sc_vhid;
1528                 }
1529                 break;
1530         default:
1531                 printf("%s: carp is not supported for the %d interface type\n",
1532                     ifp->if_xname, ifp->if_type);
1533                 return (EOPNOTSUPP);
1534         }
1535
1536         return (0);
1537 }
1538
1539 static struct carp_softc*
1540 carp_alloc(struct ifnet *ifp)
1541 {
1542         struct carp_softc *sc;
1543         struct carp_if *cif;
1544
1545         sx_assert(&carp_sx, SA_XLOCKED);
1546
1547         if ((cif = ifp->if_carp) == NULL)
1548                 cif = carp_alloc_if(ifp);
1549
1550         sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
1551
1552         sc->sc_advbase = CARP_DFLTINTV;
1553         sc->sc_vhid = -1;       /* required setting */
1554         sc->sc_init_counter = 1;
1555         sc->sc_state = INIT;
1556
1557         sc->sc_ifasiz = sizeof(struct ifaddr *);
1558         sc->sc_ifas = malloc(sc->sc_ifasiz, M_CARP, M_WAITOK|M_ZERO);
1559         sc->sc_carpdev = ifp;
1560
1561         CARP_LOCK_INIT(sc);
1562 #ifdef INET
1563         callout_init_mtx(&sc->sc_md_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1564 #endif
1565 #ifdef INET6
1566         callout_init_mtx(&sc->sc_md6_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1567 #endif
1568         callout_init_mtx(&sc->sc_ad_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1569
1570         CIF_LOCK(cif);
1571         TAILQ_INSERT_TAIL(&cif->cif_vrs, sc, sc_list);
1572         CIF_UNLOCK(cif);
1573
1574         mtx_lock(&carp_mtx);
1575         LIST_INSERT_HEAD(&carp_list, sc, sc_next);
1576         mtx_unlock(&carp_mtx);
1577
1578         return (sc);
1579 }
1580
1581 static void
1582 carp_grow_ifas(struct carp_softc *sc)
1583 {
1584         struct ifaddr **new;
1585
1586         new = malloc(sc->sc_ifasiz * 2, M_CARP, M_WAITOK | M_ZERO);
1587         CARP_LOCK(sc);
1588         bcopy(sc->sc_ifas, new, sc->sc_ifasiz);
1589         free(sc->sc_ifas, M_CARP);
1590         sc->sc_ifas = new;
1591         sc->sc_ifasiz *= 2;
1592         CARP_UNLOCK(sc);
1593 }
1594
1595 static void
1596 carp_destroy(struct carp_softc *sc)
1597 {
1598         struct ifnet *ifp = sc->sc_carpdev;
1599         struct carp_if *cif = ifp->if_carp;
1600
1601         sx_assert(&carp_sx, SA_XLOCKED);
1602
1603         if (sc->sc_suppress)
1604                 carp_demote_adj(-V_carp_ifdown_adj, "vhid removed");
1605         CARP_UNLOCK(sc);
1606
1607         CIF_LOCK(cif);
1608         TAILQ_REMOVE(&cif->cif_vrs, sc, sc_list);
1609         CIF_UNLOCK(cif);
1610
1611         mtx_lock(&carp_mtx);
1612         LIST_REMOVE(sc, sc_next);
1613         mtx_unlock(&carp_mtx);
1614
1615         callout_drain(&sc->sc_ad_tmo);
1616 #ifdef INET
1617         callout_drain(&sc->sc_md_tmo);
1618 #endif
1619 #ifdef INET6
1620         callout_drain(&sc->sc_md6_tmo);
1621 #endif
1622         CARP_LOCK_DESTROY(sc);
1623
1624         free(sc->sc_ifas, M_CARP);
1625         free(sc, M_CARP);
1626 }
1627
1628 static struct carp_if*
1629 carp_alloc_if(struct ifnet *ifp)
1630 {
1631         struct carp_if *cif;
1632         int error;
1633
1634         cif = malloc(sizeof(*cif), M_CARP, M_WAITOK|M_ZERO);
1635
1636         if ((error = ifpromisc(ifp, 1)) != 0)
1637                 printf("%s: ifpromisc(%s) failed: %d\n",
1638                     __func__, ifp->if_xname, error);
1639         else
1640                 cif->cif_flags |= CIF_PROMISC;
1641
1642         CIF_LOCK_INIT(cif);
1643         cif->cif_ifp = ifp;
1644         TAILQ_INIT(&cif->cif_vrs);
1645
1646         IF_ADDR_WLOCK(ifp);
1647         ifp->if_carp = cif;
1648         if_ref(ifp);
1649         IF_ADDR_WUNLOCK(ifp);
1650
1651         return (cif);
1652 }
1653
1654 static void
1655 carp_free_if(struct carp_if *cif)
1656 {
1657         struct ifnet *ifp = cif->cif_ifp;
1658
1659         CIF_LOCK_ASSERT(cif);
1660         KASSERT(TAILQ_EMPTY(&cif->cif_vrs), ("%s: softc list not empty",
1661             __func__));
1662
1663         IF_ADDR_WLOCK(ifp);
1664         ifp->if_carp = NULL;
1665         IF_ADDR_WUNLOCK(ifp);
1666
1667         CIF_LOCK_DESTROY(cif);
1668
1669         if (cif->cif_flags & CIF_PROMISC)
1670                 ifpromisc(ifp, 0);
1671         if_rele(ifp);
1672
1673         free(cif, M_CARP);
1674 }
1675
1676 static void
1677 carp_carprcp(struct carpreq *carpr, struct carp_softc *sc, int priv)
1678 {
1679
1680         CARP_LOCK(sc);
1681         carpr->carpr_state = sc->sc_state;
1682         carpr->carpr_vhid = sc->sc_vhid;
1683         carpr->carpr_advbase = sc->sc_advbase;
1684         carpr->carpr_advskew = sc->sc_advskew;
1685         if (priv)
1686                 bcopy(sc->sc_key, carpr->carpr_key, sizeof(carpr->carpr_key));
1687         else
1688                 bzero(carpr->carpr_key, sizeof(carpr->carpr_key));
1689         CARP_UNLOCK(sc);
1690 }
1691
1692 int
1693 carp_ioctl(struct ifreq *ifr, u_long cmd, struct thread *td)
1694 {
1695         struct carpreq carpr;
1696         struct ifnet *ifp;
1697         struct carp_softc *sc = NULL;
1698         int error = 0, locked = 0;
1699
1700         if ((error = copyin(ifr_data_get_ptr(ifr), &carpr, sizeof carpr)))
1701                 return (error);
1702
1703         ifp = ifunit_ref(ifr->ifr_name);
1704         if (ifp == NULL)
1705                 return (ENXIO);
1706
1707         switch (ifp->if_type) {
1708         case IFT_ETHER:
1709         case IFT_L2VLAN:
1710         case IFT_BRIDGE:
1711                 break;
1712         default:
1713                 error = EOPNOTSUPP;
1714                 goto out;
1715         }
1716
1717         if ((ifp->if_flags & IFF_MULTICAST) == 0) {
1718                 error = EADDRNOTAVAIL;
1719                 goto out;
1720         }
1721
1722         sx_xlock(&carp_sx);
1723         switch (cmd) {
1724         case SIOCSVH:
1725                 if ((error = priv_check(td, PRIV_NETINET_CARP)))
1726                         break;
1727                 if (carpr.carpr_vhid <= 0 || carpr.carpr_vhid > CARP_MAXVHID ||
1728                     carpr.carpr_advbase < 0 || carpr.carpr_advskew < 0) {
1729                         error = EINVAL;
1730                         break;
1731                 }
1732
1733                 if (ifp->if_carp) {
1734                         IFNET_FOREACH_CARP(ifp, sc)
1735                                 if (sc->sc_vhid == carpr.carpr_vhid)
1736                                         break;
1737                 }
1738                 if (sc == NULL) {
1739                         sc = carp_alloc(ifp);
1740                         CARP_LOCK(sc);
1741                         sc->sc_vhid = carpr.carpr_vhid;
1742                         LLADDR(&sc->sc_addr)[0] = 0;
1743                         LLADDR(&sc->sc_addr)[1] = 0;
1744                         LLADDR(&sc->sc_addr)[2] = 0x5e;
1745                         LLADDR(&sc->sc_addr)[3] = 0;
1746                         LLADDR(&sc->sc_addr)[4] = 1;
1747                         LLADDR(&sc->sc_addr)[5] = sc->sc_vhid;
1748                 } else
1749                         CARP_LOCK(sc);
1750                 locked = 1;
1751                 if (carpr.carpr_advbase > 0) {
1752                         if (carpr.carpr_advbase > 255 ||
1753                             carpr.carpr_advbase < CARP_DFLTINTV) {
1754                                 error = EINVAL;
1755                                 break;
1756                         }
1757                         sc->sc_advbase = carpr.carpr_advbase;
1758                 }
1759                 if (carpr.carpr_advskew >= 255) {
1760                         error = EINVAL;
1761                         break;
1762                 }
1763                 sc->sc_advskew = carpr.carpr_advskew;
1764                 if (carpr.carpr_key[0] != '\0') {
1765                         bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1766                         carp_hmac_prepare(sc);
1767                 }
1768                 if (sc->sc_state != INIT &&
1769                     carpr.carpr_state != sc->sc_state) {
1770                         switch (carpr.carpr_state) {
1771                         case BACKUP:
1772                                 callout_stop(&sc->sc_ad_tmo);
1773                                 carp_set_state(sc, BACKUP,
1774                                     "user requested via ifconfig");
1775                                 carp_setrun(sc, 0);
1776                                 carp_delroute(sc);
1777                                 break;
1778                         case MASTER:
1779                                 carp_master_down_locked(sc,
1780                                     "user requested via ifconfig");
1781                                 break;
1782                         default:
1783                                 break;
1784                         }
1785                 }
1786                 break;
1787
1788         case SIOCGVH:
1789             {
1790                 int priveleged;
1791
1792                 if (carpr.carpr_vhid < 0 || carpr.carpr_vhid > CARP_MAXVHID) {
1793                         error = EINVAL;
1794                         break;
1795                 }
1796                 if (carpr.carpr_count < 1) {
1797                         error = EMSGSIZE;
1798                         break;
1799                 }
1800                 if (ifp->if_carp == NULL) {
1801                         error = ENOENT;
1802                         break;
1803                 }
1804
1805                 priveleged = (priv_check(td, PRIV_NETINET_CARP) == 0);
1806                 if (carpr.carpr_vhid != 0) {
1807                         IFNET_FOREACH_CARP(ifp, sc)
1808                                 if (sc->sc_vhid == carpr.carpr_vhid)
1809                                         break;
1810                         if (sc == NULL) {
1811                                 error = ENOENT;
1812                                 break;
1813                         }
1814                         carp_carprcp(&carpr, sc, priveleged);
1815                         error = copyout(&carpr, ifr_data_get_ptr(ifr),
1816                             sizeof(carpr));
1817                 } else  {
1818                         int i, count;
1819
1820                         count = 0;
1821                         IFNET_FOREACH_CARP(ifp, sc)
1822                                 count++;
1823
1824                         if (count > carpr.carpr_count) {
1825                                 CIF_UNLOCK(ifp->if_carp);
1826                                 error = EMSGSIZE;
1827                                 break;
1828                         }
1829
1830                         i = 0;
1831                         IFNET_FOREACH_CARP(ifp, sc) {
1832                                 carp_carprcp(&carpr, sc, priveleged);
1833                                 carpr.carpr_count = count;
1834                                 error = copyout(&carpr,
1835                                     (caddr_t)ifr_data_get_ptr(ifr) +
1836                                     (i * sizeof(carpr)), sizeof(carpr));
1837                                 if (error) {
1838                                         CIF_UNLOCK(ifp->if_carp);
1839                                         break;
1840                                 }
1841                                 i++;
1842                         }
1843                 }
1844                 break;
1845             }
1846         default:
1847                 error = EINVAL;
1848         }
1849         sx_xunlock(&carp_sx);
1850
1851 out:
1852         if (locked)
1853                 CARP_UNLOCK(sc);
1854         if_rele(ifp);
1855
1856         return (error);
1857 }
1858
1859 static int
1860 carp_get_vhid(struct ifaddr *ifa)
1861 {
1862
1863         if (ifa == NULL || ifa->ifa_carp == NULL)
1864                 return (0);
1865
1866         return (ifa->ifa_carp->sc_vhid);
1867 }
1868
1869 int
1870 carp_attach(struct ifaddr *ifa, int vhid)
1871 {
1872         struct ifnet *ifp = ifa->ifa_ifp;
1873         struct carp_if *cif = ifp->if_carp;
1874         struct carp_softc *sc;
1875         int index, error;
1876
1877         KASSERT(ifa->ifa_carp == NULL, ("%s: ifa %p attached", __func__, ifa));
1878
1879         switch (ifa->ifa_addr->sa_family) {
1880 #ifdef INET
1881         case AF_INET:
1882 #endif
1883 #ifdef INET6
1884         case AF_INET6:
1885 #endif
1886                 break;
1887         default:
1888                 return (EPROTOTYPE);
1889         }
1890
1891         sx_xlock(&carp_sx);
1892         if (ifp->if_carp == NULL) {
1893                 sx_xunlock(&carp_sx);
1894                 return (ENOPROTOOPT);
1895         }
1896
1897         IFNET_FOREACH_CARP(ifp, sc)
1898                 if (sc->sc_vhid == vhid)
1899                         break;
1900         if (sc == NULL) {
1901                 sx_xunlock(&carp_sx);
1902                 return (ENOENT);
1903         }
1904
1905         error = carp_multicast_setup(cif, ifa->ifa_addr->sa_family);
1906         if (error) {
1907                 CIF_FREE(cif);
1908                 sx_xunlock(&carp_sx);
1909                 return (error);
1910         }
1911
1912         index = sc->sc_naddrs + sc->sc_naddrs6 + 1;
1913         if (index > sc->sc_ifasiz / sizeof(struct ifaddr *))
1914                 carp_grow_ifas(sc);
1915
1916         switch (ifa->ifa_addr->sa_family) {
1917 #ifdef INET
1918         case AF_INET:
1919                 cif->cif_naddrs++;
1920                 sc->sc_naddrs++;
1921                 break;
1922 #endif
1923 #ifdef INET6
1924         case AF_INET6:
1925                 cif->cif_naddrs6++;
1926                 sc->sc_naddrs6++;
1927                 break;
1928 #endif
1929         }
1930
1931         ifa_ref(ifa);
1932
1933         CARP_LOCK(sc);
1934         sc->sc_ifas[index - 1] = ifa;
1935         ifa->ifa_carp = sc;
1936         carp_hmac_prepare(sc);
1937         carp_sc_state(sc);
1938         CARP_UNLOCK(sc);
1939
1940         sx_xunlock(&carp_sx);
1941
1942         return (0);
1943 }
1944
1945 void
1946 carp_detach(struct ifaddr *ifa, bool keep_cif)
1947 {
1948         struct ifnet *ifp = ifa->ifa_ifp;
1949         struct carp_if *cif = ifp->if_carp;
1950         struct carp_softc *sc = ifa->ifa_carp;
1951         int i, index;
1952
1953         KASSERT(sc != NULL, ("%s: %p not attached", __func__, ifa));
1954
1955         sx_xlock(&carp_sx);
1956
1957         CARP_LOCK(sc);
1958         /* Shift array. */
1959         index = sc->sc_naddrs + sc->sc_naddrs6;
1960         for (i = 0; i < index; i++)
1961                 if (sc->sc_ifas[i] == ifa)
1962                         break;
1963         KASSERT(i < index, ("%s: %p no backref", __func__, ifa));
1964         for (; i < index - 1; i++)
1965                 sc->sc_ifas[i] = sc->sc_ifas[i+1];
1966         sc->sc_ifas[index - 1] = NULL;
1967
1968         switch (ifa->ifa_addr->sa_family) {
1969 #ifdef INET
1970         case AF_INET:
1971                 cif->cif_naddrs--;
1972                 sc->sc_naddrs--;
1973                 break;
1974 #endif
1975 #ifdef INET6
1976         case AF_INET6:
1977                 cif->cif_naddrs6--;
1978                 sc->sc_naddrs6--;
1979                 break;
1980 #endif
1981         }
1982
1983         carp_ifa_delroute(ifa);
1984         carp_multicast_cleanup(cif, ifa->ifa_addr->sa_family);
1985
1986         ifa->ifa_carp = NULL;
1987         ifa_free(ifa);
1988
1989         carp_hmac_prepare(sc);
1990         carp_sc_state(sc);
1991
1992         if (!keep_cif && sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0)
1993                 carp_destroy(sc);
1994         else
1995                 CARP_UNLOCK(sc);
1996
1997         if (!keep_cif)
1998                 CIF_FREE(cif);
1999
2000         sx_xunlock(&carp_sx);
2001 }
2002
2003 static void
2004 carp_set_state(struct carp_softc *sc, int state, const char *reason)
2005 {
2006
2007         CARP_LOCK_ASSERT(sc);
2008
2009         if (sc->sc_state != state) {
2010                 const char *carp_states[] = { CARP_STATES };
2011                 char subsys[IFNAMSIZ+5];
2012
2013                 snprintf(subsys, IFNAMSIZ+5, "%u@%s", sc->sc_vhid,
2014                     sc->sc_carpdev->if_xname);
2015
2016                 CARP_LOG("%s: %s -> %s (%s)\n", subsys,
2017                     carp_states[sc->sc_state], carp_states[state], reason);
2018
2019                 sc->sc_state = state;
2020
2021                 devctl_notify("CARP", subsys, carp_states[state], NULL);
2022         }
2023 }
2024
2025 static void
2026 carp_linkstate(struct ifnet *ifp)
2027 {
2028         struct carp_softc *sc;
2029
2030         CIF_LOCK(ifp->if_carp);
2031         IFNET_FOREACH_CARP(ifp, sc) {
2032                 CARP_LOCK(sc);
2033                 carp_sc_state(sc);
2034                 CARP_UNLOCK(sc);
2035         }
2036         CIF_UNLOCK(ifp->if_carp);
2037 }
2038
2039 static void
2040 carp_sc_state(struct carp_softc *sc)
2041 {
2042
2043         CARP_LOCK_ASSERT(sc);
2044
2045         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
2046             !(sc->sc_carpdev->if_flags & IFF_UP) ||
2047             !V_carp_allow) {
2048                 callout_stop(&sc->sc_ad_tmo);
2049 #ifdef INET
2050                 callout_stop(&sc->sc_md_tmo);
2051 #endif
2052 #ifdef INET6
2053                 callout_stop(&sc->sc_md6_tmo);
2054 #endif
2055                 carp_set_state(sc, INIT, "hardware interface down");
2056                 carp_setrun(sc, 0);
2057                 if (!sc->sc_suppress)
2058                         carp_demote_adj(V_carp_ifdown_adj, "interface down");
2059                 sc->sc_suppress = 1;
2060         } else {
2061                 carp_set_state(sc, INIT, "hardware interface up");
2062                 carp_setrun(sc, 0);
2063                 if (sc->sc_suppress)
2064                         carp_demote_adj(-V_carp_ifdown_adj, "interface up");
2065                 sc->sc_suppress = 0;
2066         }
2067 }
2068
2069 static void
2070 carp_demote_adj(int adj, char *reason)
2071 {
2072         atomic_add_int(&V_carp_demotion, adj);
2073         CARP_LOG("demoted by %d to %d (%s)\n", adj, V_carp_demotion, reason);
2074         taskqueue_enqueue(taskqueue_swi, &carp_sendall_task);
2075 }
2076
2077 static int
2078 carp_allow_sysctl(SYSCTL_HANDLER_ARGS)
2079 {
2080         int new, error;
2081         struct carp_softc *sc;
2082
2083         new = V_carp_allow;
2084         error = sysctl_handle_int(oidp, &new, 0, req);
2085         if (error || !req->newptr)
2086                 return (error);
2087
2088         if (V_carp_allow != new) {
2089                 V_carp_allow = new;
2090
2091                 mtx_lock(&carp_mtx);
2092                 LIST_FOREACH(sc, &carp_list, sc_next) {
2093                         CARP_LOCK(sc);
2094                         if (curvnet == sc->sc_carpdev->if_vnet)
2095                                 carp_sc_state(sc);
2096                         CARP_UNLOCK(sc);
2097                 }
2098                 mtx_unlock(&carp_mtx);
2099         }
2100
2101         return (0);
2102 }
2103
2104 static int
2105 carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS)
2106 {
2107         int new, error;
2108
2109         new = V_carp_demotion;
2110         error = sysctl_handle_int(oidp, &new, 0, req);
2111         if (error || !req->newptr)
2112                 return (error);
2113
2114         carp_demote_adj(new, "sysctl");
2115
2116         return (0);
2117 }
2118
2119 #ifdef INET
2120 extern  struct domain inetdomain;
2121 static struct protosw in_carp_protosw = {
2122         .pr_type =              SOCK_RAW,
2123         .pr_domain =            &inetdomain,
2124         .pr_protocol =          IPPROTO_CARP,
2125         .pr_flags =             PR_ATOMIC|PR_ADDR,
2126         .pr_input =             carp_input,
2127         .pr_output =            rip_output,
2128         .pr_ctloutput =         rip_ctloutput,
2129         .pr_usrreqs =           &rip_usrreqs
2130 };
2131 #endif
2132
2133 #ifdef INET6
2134 extern  struct domain inet6domain;
2135 static struct protosw in6_carp_protosw = {
2136         .pr_type =              SOCK_RAW,
2137         .pr_domain =            &inet6domain,
2138         .pr_protocol =          IPPROTO_CARP,
2139         .pr_flags =             PR_ATOMIC|PR_ADDR,
2140         .pr_input =             carp6_input,
2141         .pr_output =            rip6_output,
2142         .pr_ctloutput =         rip6_ctloutput,
2143         .pr_usrreqs =           &rip6_usrreqs
2144 };
2145 #endif
2146
2147 static void
2148 carp_mod_cleanup(void)
2149 {
2150
2151 #ifdef INET
2152         if (proto_reg[CARP_INET] == 0) {
2153                 (void)ipproto_unregister(IPPROTO_CARP);
2154                 pf_proto_unregister(PF_INET, IPPROTO_CARP, SOCK_RAW);
2155                 proto_reg[CARP_INET] = -1;
2156         }
2157         carp_iamatch_p = NULL;
2158 #endif
2159 #ifdef INET6
2160         if (proto_reg[CARP_INET6] == 0) {
2161                 (void)ip6proto_unregister(IPPROTO_CARP);
2162                 pf_proto_unregister(PF_INET6, IPPROTO_CARP, SOCK_RAW);
2163                 proto_reg[CARP_INET6] = -1;
2164         }
2165         carp_iamatch6_p = NULL;
2166         carp_macmatch6_p = NULL;
2167 #endif
2168         carp_ioctl_p = NULL;
2169         carp_attach_p = NULL;
2170         carp_detach_p = NULL;
2171         carp_get_vhid_p = NULL;
2172         carp_linkstate_p = NULL;
2173         carp_forus_p = NULL;
2174         carp_output_p = NULL;
2175         carp_demote_adj_p = NULL;
2176         carp_master_p = NULL;
2177         mtx_unlock(&carp_mtx);
2178         taskqueue_drain(taskqueue_swi, &carp_sendall_task);
2179         mtx_destroy(&carp_mtx);
2180         sx_destroy(&carp_sx);
2181 }
2182
2183 static int
2184 carp_mod_load(void)
2185 {
2186         int err;
2187
2188         mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2189         sx_init(&carp_sx, "carp_sx");
2190         LIST_INIT(&carp_list);
2191         carp_get_vhid_p = carp_get_vhid;
2192         carp_forus_p = carp_forus;
2193         carp_output_p = carp_output;
2194         carp_linkstate_p = carp_linkstate;
2195         carp_ioctl_p = carp_ioctl;
2196         carp_attach_p = carp_attach;
2197         carp_detach_p = carp_detach;
2198         carp_demote_adj_p = carp_demote_adj;
2199         carp_master_p = carp_master;
2200 #ifdef INET6
2201         carp_iamatch6_p = carp_iamatch6;
2202         carp_macmatch6_p = carp_macmatch6;
2203         proto_reg[CARP_INET6] = pf_proto_register(PF_INET6,
2204             (struct protosw *)&in6_carp_protosw);
2205         if (proto_reg[CARP_INET6]) {
2206                 printf("carp: error %d attaching to PF_INET6\n",
2207                     proto_reg[CARP_INET6]);
2208                 carp_mod_cleanup();
2209                 return (proto_reg[CARP_INET6]);
2210         }
2211         err = ip6proto_register(IPPROTO_CARP);
2212         if (err) {
2213                 printf("carp: error %d registering with INET6\n", err);
2214                 carp_mod_cleanup();
2215                 return (err);
2216         }
2217 #endif
2218 #ifdef INET
2219         carp_iamatch_p = carp_iamatch;
2220         proto_reg[CARP_INET] = pf_proto_register(PF_INET, &in_carp_protosw);
2221         if (proto_reg[CARP_INET]) {
2222                 printf("carp: error %d attaching to PF_INET\n",
2223                     proto_reg[CARP_INET]);
2224                 carp_mod_cleanup();
2225                 return (proto_reg[CARP_INET]);
2226         }
2227         err = ipproto_register(IPPROTO_CARP);
2228         if (err) {
2229                 printf("carp: error %d registering with INET\n", err);
2230                 carp_mod_cleanup();
2231                 return (err);
2232         }
2233 #endif
2234         return (0);
2235 }
2236
2237 static int
2238 carp_modevent(module_t mod, int type, void *data)
2239 {
2240         switch (type) {
2241         case MOD_LOAD:
2242                 return carp_mod_load();
2243                 /* NOTREACHED */
2244         case MOD_UNLOAD:
2245                 mtx_lock(&carp_mtx);
2246                 if (LIST_EMPTY(&carp_list))
2247                         carp_mod_cleanup();
2248                 else {
2249                         mtx_unlock(&carp_mtx);
2250                         return (EBUSY);
2251                 }
2252                 break;
2253
2254         default:
2255                 return (EINVAL);
2256         }
2257
2258         return (0);
2259 }
2260
2261 static moduledata_t carp_mod = {
2262         "carp",
2263         carp_modevent,
2264         0
2265 };
2266
2267 DECLARE_MODULE(carp, carp_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);