sys/netinet/ip_carp.c

   1 /*-
   2  * Copyright (c) 2002 Michael Shalayeff.
   3  * Copyright (c) 2003 Ryan McBride.
   4  * Copyright (c) 2011 Gleb Smirnoff <glebius@FreeBSD.org>
   5  * All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  17  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  18  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  19  * IN NO EVENT SHALL THE AUTHOR OR HIS RELATIVES BE LIABLE FOR ANY DIRECT,
  20  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  21  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  22  * SERVICES; LOSS OF MIND, USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
  24  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
  25  * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
  26  * THE POSSIBILITY OF SUCH DAMAGE.
  27  */
  28
  29 #include <sys/cdefs.h>
  30 __FBSDID("$FreeBSD$");
  31
  32 #include "opt_bpf.h"
  33 #include "opt_inet.h"
  34 #include "opt_inet6.h"
  35
  36 #include <sys/param.h>
  37 #include <sys/systm.h>
  38 #include <sys/bus.h>
  39 #include <sys/jail.h>
  40 #include <sys/kernel.h>
  41 #include <sys/limits.h>
  42 #include <sys/malloc.h>
  43 #include <sys/mbuf.h>
  44 #include <sys/module.h>
  45 #include <sys/priv.h>
  46 #include <sys/proc.h>
  47 #include <sys/protosw.h>
  48 #include <sys/socket.h>
  49 #include <sys/sockio.h>
  50 #include <sys/sysctl.h>
  51 #include <sys/syslog.h>
  52 #include <sys/taskqueue.h>
  53 #include <sys/counter.h>
  54
  55 #include <net/ethernet.h>
  56 #include <net/fddi.h>
  57 #include <net/if.h>
  58 #include <net/if_var.h>
  59 #include <net/if_dl.h>
  60 #include <net/if_llatbl.h>
  61 #include <net/if_types.h>
  62 #include <net/iso88025.h>
  63 #include <net/route.h>
  64 #include <net/vnet.h>
  65
  66 #if defined(INET) || defined(INET6)
  67 #include <netinet/in.h>
  68 #include <netinet/in_var.h>
  69 #include <netinet/ip_carp.h>
  70 #include <netinet/ip.h>
  71 #include <machine/in_cksum.h>
  72 #endif
  73 #ifdef INET
  74 #include <netinet/ip_var.h>
  75 #include <netinet/if_ether.h>
  76 #endif
  77
  78 #ifdef INET6
  79 #include <netinet/icmp6.h>
  80 #include <netinet/ip6.h>
  81 #include <netinet6/in6_var.h>
  82 #include <netinet6/ip6_var.h>
  83 #include <netinet6/scope6_var.h>
  84 #include <netinet6/nd6.h>
  85 #endif
  86
  87 #include <crypto/sha1.h>
  88
  89 static MALLOC_DEFINE(M_CARP, "CARP", "CARP addresses");
  90
  91 struct carp_softc {
  92         struct ifnet            *sc_carpdev;    /* Pointer to parent ifnet. */
  93         struct ifaddr           **sc_ifas;      /* Our ifaddrs. */
  94         struct sockaddr_dl      sc_addr;        /* Our link level address. */
  95         struct callout          sc_ad_tmo;      /* Advertising timeout. */
  96 #ifdef INET
  97         struct callout          sc_md_tmo;      /* Master down timeout. */
  98 #endif
  99 #ifdef INET6
 100         struct callout          sc_md6_tmo;     /* XXX: Master down timeout. */
 101 #endif
 102         struct mtx              sc_mtx;
 103
 104         int                     sc_vhid;
 105         int                     sc_advskew;
 106         int                     sc_advbase;
 107
 108         int                     sc_naddrs;
 109         int                     sc_naddrs6;
 110         int                     sc_ifasiz;
 111         enum { INIT = 0, BACKUP, MASTER }       sc_state;
 112         int                     sc_suppress;
 113         int                     sc_sendad_errors;
 114 #define CARP_SENDAD_MAX_ERRORS  3
 115         int                     sc_sendad_success;
 116 #define CARP_SENDAD_MIN_SUCCESS 3
 117
 118         int                     sc_init_counter;
 119         uint64_t                sc_counter;
 120
 121         /* authentication */
 122 #define CARP_HMAC_PAD   64
 123         unsigned char sc_key[CARP_KEY_LEN];
 124         unsigned char sc_pad[CARP_HMAC_PAD];
 125         SHA1_CTX sc_sha1;
 126
 127         TAILQ_ENTRY(carp_softc) sc_list;        /* On the carp_if list. */
 128         LIST_ENTRY(carp_softc)  sc_next;        /* On the global list. */
 129 };
 130
 131 struct carp_if {
 132 #ifdef INET
 133         int     cif_naddrs;
 134 #endif
 135 #ifdef INET6
 136         int     cif_naddrs6;
 137 #endif
 138         TAILQ_HEAD(, carp_softc) cif_vrs;
 139 #ifdef INET
 140         struct ip_moptions       cif_imo;
 141 #endif
 142 #ifdef INET6
 143         struct ip6_moptions      cif_im6o;
 144 #endif
 145         struct ifnet    *cif_ifp;
 146         struct mtx      cif_mtx;
 147         uint32_t        cif_flags;
 148 #define CIF_PROMISC     0x00000001
 149 };
 150
 151 #define CARP_INET       0
 152 #define CARP_INET6      1
 153 static int proto_reg[] = {-1, -1};
 154
 155 /*
 156  * Brief design of carp(4).
 157  *
 158  * Any carp-capable ifnet may have a list of carp softcs hanging off
 159  * its ifp->if_carp pointer. Each softc represents one unique virtual
 160  * host id, or vhid. The softc has a back pointer to the ifnet. All
 161  * softcs are joined in a global list, which has quite limited use.
 162  *
 163  * Any interface address that takes part in CARP negotiation has a
 164  * pointer to the softc of its vhid, ifa->ifa_carp. That could be either
 165  * AF_INET or AF_INET6 address.
 166  *
 167  * Although, one can get the softc's backpointer to ifnet and traverse
 168  * through its ifp->if_addrhead queue to find all interface addresses
 169  * involved in CARP, we keep a growable array of ifaddr pointers. This
 170  * allows us to avoid grabbing the IF_ADDR_LOCK() in many traversals that
 171  * do calls into the network stack, thus avoiding LORs.
 172  *
 173  * Locking:
 174  *
 175  * Each softc has a lock sc_mtx. It is used to synchronise carp_input_c(),
 176  * callout-driven events and ioctl()s.
 177  *
 178  * To traverse the list of softcs on an ifnet we use CIF_LOCK(), to
 179  * traverse the global list we use the mutex carp_mtx.
 180  *
 181  * Known issues with locking:
 182  *
 183  * - Sending ad, we put the pointer to the softc in an mtag, and no reference
 184  *   counting is done on the softc.
 185  * - On module unload we may race (?) with packet processing thread
 186  *   dereferencing our function pointers.
 187  */
 188
 189 /* Accept incoming CARP packets. */
 190 static VNET_DEFINE(int, carp_allow) = 1;
 191 #define V_carp_allow    VNET(carp_allow)
 192
 193 /* Preempt slower nodes. */
 194 static VNET_DEFINE(int, carp_preempt) = 0;
 195 #define V_carp_preempt  VNET(carp_preempt)
 196
 197 /* Log level. */
 198 static VNET_DEFINE(int, carp_log) = 1;
 199 #define V_carp_log      VNET(carp_log)
 200
 201 /* Global advskew demotion. */
 202 static VNET_DEFINE(int, carp_demotion) = 0;
 203 #define V_carp_demotion VNET(carp_demotion)
 204
 205 /* Send error demotion factor. */
 206 static VNET_DEFINE(int, carp_senderr_adj) = CARP_MAXSKEW;
 207 #define V_carp_senderr_adj      VNET(carp_senderr_adj)
 208
 209 /* Iface down demotion factor. */
 210 static VNET_DEFINE(int, carp_ifdown_adj) = CARP_MAXSKEW;
 211 #define V_carp_ifdown_adj       VNET(carp_ifdown_adj)
 212
 213 static int carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS);
 214
 215 SYSCTL_NODE(_net_inet, IPPROTO_CARP,    carp,   CTLFLAG_RW, 0,  "CARP");
 216 SYSCTL_INT(_net_inet_carp, OID_AUTO, allow, CTLFLAG_VNET | CTLFLAG_RW,
 217     &VNET_NAME(carp_allow), 0, "Accept incoming CARP packets");
 218 SYSCTL_INT(_net_inet_carp, OID_AUTO, preempt, CTLFLAG_VNET | CTLFLAG_RW,
 219     &VNET_NAME(carp_preempt), 0, "High-priority backup preemption mode");
 220 SYSCTL_INT(_net_inet_carp, OID_AUTO, log, CTLFLAG_VNET | CTLFLAG_RW,
 221     &VNET_NAME(carp_log), 0, "CARP log level");
 222 SYSCTL_PROC(_net_inet_carp, OID_AUTO, demotion,
 223     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW,
 224     0, 0, carp_demote_adj_sysctl, "I",
 225     "Adjust demotion factor (skew of advskew)");
 226 SYSCTL_INT(_net_inet_carp, OID_AUTO, senderr_demotion_factor,
 227     CTLFLAG_VNET | CTLFLAG_RW,
 228     &VNET_NAME(carp_senderr_adj), 0, "Send error demotion factor adjustment");
 229 SYSCTL_INT(_net_inet_carp, OID_AUTO, ifdown_demotion_factor,
 230     CTLFLAG_VNET | CTLFLAG_RW,
 231     &VNET_NAME(carp_ifdown_adj), 0,
 232     "Interface down demotion factor adjustment");
 233
 234 VNET_PCPUSTAT_DEFINE(struct carpstats, carpstats);
 235 VNET_PCPUSTAT_SYSINIT(carpstats);
 236 VNET_PCPUSTAT_SYSUNINIT(carpstats);
 237
 238 #define CARPSTATS_ADD(name, val)        \
 239     counter_u64_add(VNET(carpstats)[offsetof(struct carpstats, name) / \
 240         sizeof(uint64_t)], (val))
 241 #define CARPSTATS_INC(name)             CARPSTATS_ADD(name, 1)
 242
 243 SYSCTL_VNET_PCPUSTAT(_net_inet_carp, OID_AUTO, stats, struct carpstats,
 244     carpstats, "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 245
 246 #define CARP_LOCK_INIT(sc)      mtx_init(&(sc)->sc_mtx, "carp_softc",   \
 247         NULL, MTX_DEF)
 248 #define CARP_LOCK_DESTROY(sc)   mtx_destroy(&(sc)->sc_mtx)
 249 #define CARP_LOCK_ASSERT(sc)    mtx_assert(&(sc)->sc_mtx, MA_OWNED)
 250 #define CARP_LOCK(sc)           mtx_lock(&(sc)->sc_mtx)
 251 #define CARP_UNLOCK(sc)         mtx_unlock(&(sc)->sc_mtx)
 252 #define CIF_LOCK_INIT(cif)      mtx_init(&(cif)->cif_mtx, "carp_if",   \
 253         NULL, MTX_DEF)
 254 #define CIF_LOCK_DESTROY(cif)   mtx_destroy(&(cif)->cif_mtx)
 255 #define CIF_LOCK_ASSERT(cif)    mtx_assert(&(cif)->cif_mtx, MA_OWNED)
 256 #define CIF_LOCK(cif)           mtx_lock(&(cif)->cif_mtx)
 257 #define CIF_UNLOCK(cif)         mtx_unlock(&(cif)->cif_mtx)
 258 #define CIF_FREE(cif)   do {                            \
 259                 CIF_LOCK(cif);                          \
 260                 if (TAILQ_EMPTY(&(cif)->cif_vrs))       \
 261                         carp_free_if(cif);              \
 262                 else                                    \
 263                         CIF_UNLOCK(cif);                \
 264 } while (0)
 265
 266 #define CARP_LOG(...)   do {                            \
 267         if (V_carp_log > 0)                             \
 268                 log(LOG_INFO, "carp: " __VA_ARGS__);    \
 269 } while (0)
 270
 271 #define CARP_DEBUG(...) do {                            \
 272         if (V_carp_log > 1)                             \
 273                 log(LOG_DEBUG, __VA_ARGS__);            \
 274 } while (0)
 275
 276 #define IFNET_FOREACH_IFA(ifp, ifa)                                     \
 277         IF_ADDR_LOCK_ASSERT(ifp);                                       \
 278         TAILQ_FOREACH((ifa), &(ifp)->if_addrhead, ifa_link)             \
 279                 if ((ifa)->ifa_carp != NULL)
 280
 281 #define CARP_FOREACH_IFA(sc, ifa)                                       \
 282         CARP_LOCK_ASSERT(sc);                                           \
 283         for (int _i = 0;                                                \
 284                 _i < (sc)->sc_naddrs + (sc)->sc_naddrs6 &&              \
 285                 ((ifa) = sc->sc_ifas[_i]) != NULL;                      \
 286                 ++_i)
 287
 288 #define IFNET_FOREACH_CARP(ifp, sc)                                     \
 289         CIF_LOCK_ASSERT(ifp->if_carp);                                  \
 290         TAILQ_FOREACH((sc), &(ifp)->if_carp->cif_vrs, sc_list)
 291
 292 #define DEMOTE_ADVSKEW(sc)                                      \
 293     (((sc)->sc_advskew + V_carp_demotion > CARP_MAXSKEW) ?      \
 294     CARP_MAXSKEW : ((sc)->sc_advskew + V_carp_demotion))
 295
 296 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 297 static struct carp_softc
 298                 *carp_alloc(struct ifnet *);
 299 static void     carp_destroy(struct carp_softc *);
 300 static struct carp_if
 301                 *carp_alloc_if(struct ifnet *);
 302 static void     carp_free_if(struct carp_if *);
 303 static void     carp_set_state(struct carp_softc *, int, const char* reason);
 304 static void     carp_sc_state(struct carp_softc *);
 305 static void     carp_setrun(struct carp_softc *, sa_family_t);
 306 static void     carp_master_down(void *);
 307 static void     carp_master_down_locked(struct carp_softc *,
 308                     const char* reason);
 309 static void     carp_send_ad(void *);
 310 static void     carp_send_ad_locked(struct carp_softc *);
 311 static void     carp_addroute(struct carp_softc *);
 312 static void     carp_ifa_addroute(struct ifaddr *);
 313 static void     carp_delroute(struct carp_softc *);
 314 static void     carp_ifa_delroute(struct ifaddr *);
 315 static void     carp_send_ad_all(void *, int);
 316 static void     carp_demote_adj(int, char *);
 317
 318 static LIST_HEAD(, carp_softc) carp_list;
 319 static struct mtx carp_mtx;
 320 static struct sx carp_sx;
 321 static struct task carp_sendall_task =
 322     TASK_INITIALIZER(0, carp_send_ad_all, NULL);
 323
 324 static void
 325 carp_hmac_prepare(struct carp_softc *sc)
 326 {
 327         uint8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 328         uint8_t vhid = sc->sc_vhid & 0xff;
 329         struct ifaddr *ifa;
 330         int i, found;
 331 #ifdef INET
 332         struct in_addr last, cur, in;
 333 #endif
 334 #ifdef INET6
 335         struct in6_addr last6, cur6, in6;
 336 #endif
 337
 338         CARP_LOCK_ASSERT(sc);
 339
 340         /* Compute ipad from key. */
 341         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 342         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 343         for (i = 0; i < sizeof(sc->sc_pad); i++)
 344                 sc->sc_pad[i] ^= 0x36;
 345
 346         /* Precompute first part of inner hash. */
 347         SHA1Init(&sc->sc_sha1);
 348         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 349         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 350         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 351         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 352 #ifdef INET
 353         cur.s_addr = 0;
 354         do {
 355                 found = 0;
 356                 last = cur;
 357                 cur.s_addr = 0xffffffff;
 358                 CARP_FOREACH_IFA(sc, ifa) {
 359                         in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 360                         if (ifa->ifa_addr->sa_family == AF_INET &&
 361                             ntohl(in.s_addr) > ntohl(last.s_addr) &&
 362                             ntohl(in.s_addr) < ntohl(cur.s_addr)) {
 363                                 cur.s_addr = in.s_addr;
 364                                 found++;
 365                         }
 366                 }
 367                 if (found)
 368                         SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur));
 369         } while (found);
 370 #endif /* INET */
 371 #ifdef INET6
 372         memset(&cur6, 0, sizeof(cur6));
 373         do {
 374                 found = 0;
 375                 last6 = cur6;
 376                 memset(&cur6, 0xff, sizeof(cur6));
 377                 CARP_FOREACH_IFA(sc, ifa) {
 378                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 379                         if (IN6_IS_SCOPE_EMBED(&in6))
 380                                 in6.s6_addr16[1] = 0;
 381                         if (ifa->ifa_addr->sa_family == AF_INET6 &&
 382                             memcmp(&in6, &last6, sizeof(in6)) > 0 &&
 383                             memcmp(&in6, &cur6, sizeof(in6)) < 0) {
 384                                 cur6 = in6;
 385                                 found++;
 386                         }
 387                 }
 388                 if (found)
 389                         SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6));
 390         } while (found);
 391 #endif /* INET6 */
 392
 393         /* convert ipad to opad */
 394         for (i = 0; i < sizeof(sc->sc_pad); i++)
 395                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 396 }
 397
 398 static void
 399 carp_hmac_generate(struct carp_softc *sc, uint32_t counter[2],
 400     unsigned char md[20])
 401 {
 402         SHA1_CTX sha1ctx;
 403
 404         CARP_LOCK_ASSERT(sc);
 405
 406         /* fetch first half of inner hash */
 407         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 408
 409         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 410         SHA1Final(md, &sha1ctx);
 411
 412         /* outer hash */
 413         SHA1Init(&sha1ctx);
 414         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 415         SHA1Update(&sha1ctx, md, 20);
 416         SHA1Final(md, &sha1ctx);
 417 }
 418
 419 static int
 420 carp_hmac_verify(struct carp_softc *sc, uint32_t counter[2],
 421     unsigned char md[20])
 422 {
 423         unsigned char md2[20];
 424
 425         CARP_LOCK_ASSERT(sc);
 426
 427         carp_hmac_generate(sc, counter, md2);
 428
 429         return (bcmp(md, md2, sizeof(md2)));
 430 }
 431
 432 /*
 433  * process input packet.
 434  * we have rearranged checks order compared to the rfc,
 435  * but it seems more efficient this way or not possible otherwise.
 436  */
 437 #ifdef INET
 438 int
 439 carp_input(struct mbuf **mp, int *offp, int proto)
 440 {
 441         struct mbuf *m = *mp;
 442         struct ip *ip = mtod(m, struct ip *);
 443         struct carp_header *ch;
 444         int iplen, len;
 445
 446         iplen = *offp;
 447         *mp = NULL;
 448
 449         CARPSTATS_INC(carps_ipackets);
 450
 451         if (!V_carp_allow) {
 452                 m_freem(m);
 453                 return (IPPROTO_DONE);
 454         }
 455
 456         /* verify that the IP TTL is 255.  */
 457         if (ip->ip_ttl != CARP_DFLTTL) {
 458                 CARPSTATS_INC(carps_badttl);
 459                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 460                     ip->ip_ttl,
 461                     m->m_pkthdr.rcvif->if_xname);
 462                 m_freem(m);
 463                 return (IPPROTO_DONE);
 464         }
 465
 466         iplen = ip->ip_hl << 2;
 467
 468         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 469                 CARPSTATS_INC(carps_badlen);
 470                 CARP_DEBUG("%s: received len %zd < sizeof(struct carp_header) "
 471                     "on %s\n", __func__, m->m_len - sizeof(struct ip),
 472                     m->m_pkthdr.rcvif->if_xname);
 473                 m_freem(m);
 474                 return (IPPROTO_DONE);
 475         }
 476
 477         if (iplen + sizeof(*ch) < m->m_len) {
 478                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 479                         CARPSTATS_INC(carps_hdrops);
 480                         CARP_DEBUG("%s: pullup failed\n", __func__);
 481                         return (IPPROTO_DONE);
 482                 }
 483                 ip = mtod(m, struct ip *);
 484         }
 485         ch = (struct carp_header *)((char *)ip + iplen);
 486
 487         /*
 488          * verify that the received packet length is
 489          * equal to the CARP header
 490          */
 491         len = iplen + sizeof(*ch);
 492         if (len > m->m_pkthdr.len) {
 493                 CARPSTATS_INC(carps_badlen);
 494                 CARP_DEBUG("%s: packet too short %d on %s\n", __func__,
 495                     m->m_pkthdr.len,
 496                     m->m_pkthdr.rcvif->if_xname);
 497                 m_freem(m);
 498                 return (IPPROTO_DONE);
 499         }
 500
 501         if ((m = m_pullup(m, len)) == NULL) {
 502                 CARPSTATS_INC(carps_hdrops);
 503                 return (IPPROTO_DONE);
 504         }
 505         ip = mtod(m, struct ip *);
 506         ch = (struct carp_header *)((char *)ip + iplen);
 507
 508         /* verify the CARP checksum */
 509         m->m_data += iplen;
 510         if (in_cksum(m, len - iplen)) {
 511                 CARPSTATS_INC(carps_badsum);
 512                 CARP_DEBUG("%s: checksum failed on %s\n", __func__,
 513                     m->m_pkthdr.rcvif->if_xname);
 514                 m_freem(m);
 515                 return (IPPROTO_DONE);
 516         }
 517         m->m_data -= iplen;
 518
 519         carp_input_c(m, ch, AF_INET);
 520         return (IPPROTO_DONE);
 521 }
 522 #endif
 523
 524 #ifdef INET6
 525 int
 526 carp6_input(struct mbuf **mp, int *offp, int proto)
 527 {
 528         struct mbuf *m = *mp;
 529         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 530         struct carp_header *ch;
 531         u_int len;
 532
 533         CARPSTATS_INC(carps_ipackets6);
 534
 535         if (!V_carp_allow) {
 536                 m_freem(m);
 537                 return (IPPROTO_DONE);
 538         }
 539
 540         /* check if received on a valid carp interface */
 541         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 542                 CARPSTATS_INC(carps_badif);
 543                 CARP_DEBUG("%s: packet received on non-carp interface: %s\n",
 544                     __func__, m->m_pkthdr.rcvif->if_xname);
 545                 m_freem(m);
 546                 return (IPPROTO_DONE);
 547         }
 548
 549         /* verify that the IP TTL is 255 */
 550         if (ip6->ip6_hlim != CARP_DFLTTL) {
 551                 CARPSTATS_INC(carps_badttl);
 552                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 553                     ip6->ip6_hlim, m->m_pkthdr.rcvif->if_xname);
 554                 m_freem(m);
 555                 return (IPPROTO_DONE);
 556         }
 557
 558         /* verify that we have a complete carp packet */
 559         len = m->m_len;
 560         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 561         if (ch == NULL) {
 562                 CARPSTATS_INC(carps_badlen);
 563                 CARP_DEBUG("%s: packet size %u too small\n", __func__, len);
 564                 return (IPPROTO_DONE);
 565         }
 566
 567
 568         /* verify the CARP checksum */
 569         m->m_data += *offp;
 570         if (in_cksum(m, sizeof(*ch))) {
 571                 CARPSTATS_INC(carps_badsum);
 572                 CARP_DEBUG("%s: checksum failed, on %s\n", __func__,
 573                     m->m_pkthdr.rcvif->if_xname);
 574                 m_freem(m);
 575                 return (IPPROTO_DONE);
 576         }
 577         m->m_data -= *offp;
 578
 579         carp_input_c(m, ch, AF_INET6);
 580         return (IPPROTO_DONE);
 581 }
 582 #endif /* INET6 */
 583
 584 /*
 585  * This routine should not be necessary at all, but some switches
 586  * (VMWare ESX vswitches) can echo our own packets back at us,
 587  * and we must ignore them or they will cause us to drop out of
 588  * MASTER mode.
 589  *
 590  * We cannot catch all cases of network loops.  Instead, what we
 591  * do here is catch any packet that arrives with a carp header
 592  * with a VHID of 0, that comes from an address that is our own.
 593  * These packets are by definition "from us" (even if they are from
 594  * a misconfigured host that is pretending to be us).
 595  *
 596  * The VHID test is outside this mini-function.
 597  */
 598 static int
 599 carp_source_is_self(struct mbuf *m, struct ifaddr *ifa, sa_family_t af)
 600 {
 601 #ifdef INET
 602         struct ip *ip4;
 603         struct in_addr in4;
 604 #endif
 605 #ifdef INET6
 606         struct ip6_hdr *ip6;
 607         struct in6_addr in6;
 608 #endif
 609
 610         switch (af) {
 611 #ifdef INET
 612         case AF_INET:
 613                 ip4 = mtod(m, struct ip *);
 614                 in4 = ifatoia(ifa)->ia_addr.sin_addr;
 615                 return (in4.s_addr == ip4->ip_src.s_addr);
 616 #endif
 617 #ifdef INET6
 618         case AF_INET6:
 619                 ip6 = mtod(m, struct ip6_hdr *);
 620                 in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 621                 return (memcmp(&in6, &ip6->ip6_src, sizeof(in6)) == 0);
 622 #endif
 623         default:
 624                 break;
 625         }
 626         return (0);
 627 }
 628
 629 static void
 630 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 631 {
 632         struct ifnet *ifp = m->m_pkthdr.rcvif;
 633         struct ifaddr *ifa, *match;
 634         struct carp_softc *sc;
 635         uint64_t tmp_counter;
 636         struct timeval sc_tv, ch_tv;
 637         int error;
 638
 639         /*
 640          * Verify that the VHID is valid on the receiving interface.
 641          *
 642          * There should be just one match.  If there are none
 643          * the VHID is not valid and we drop the packet.  If
 644          * there are multiple VHID matches, take just the first
 645          * one, for compatibility with previous code.  While we're
 646          * scanning, check for obvious loops in the network topology
 647          * (these should never happen, and as noted above, we may
 648          * miss real loops; this is just a double-check).
 649          */
 650         IF_ADDR_RLOCK(ifp);
 651         error = 0;
 652         match = NULL;
 653         IFNET_FOREACH_IFA(ifp, ifa) {
 654                 if (match == NULL && ifa->ifa_carp != NULL &&
 655                     ifa->ifa_addr->sa_family == af &&
 656                     ifa->ifa_carp->sc_vhid == ch->carp_vhid)
 657                         match = ifa;
 658                 if (ch->carp_vhid == 0 && carp_source_is_self(m, ifa, af))
 659                         error = ELOOP;
 660         }
 661         ifa = error ? NULL : match;
 662         if (ifa != NULL)
 663                 ifa_ref(ifa);
 664         IF_ADDR_RUNLOCK(ifp);
 665
 666         if (ifa == NULL) {
 667                 if (error == ELOOP) {
 668                         CARP_DEBUG("dropping looped packet on interface %s\n",
 669                             ifp->if_xname);
 670                         CARPSTATS_INC(carps_badif);     /* ??? */
 671                 } else {
 672                         CARPSTATS_INC(carps_badvhid);
 673                 }
 674                 m_freem(m);
 675                 return;
 676         }
 677
 678         /* verify the CARP version. */
 679         if (ch->carp_version != CARP_VERSION) {
 680                 CARPSTATS_INC(carps_badver);
 681                 CARP_DEBUG("%s: invalid version %d\n", ifp->if_xname,
 682                     ch->carp_version);
 683                 ifa_free(ifa);
 684                 m_freem(m);
 685                 return;
 686         }
 687
 688         sc = ifa->ifa_carp;
 689         CARP_LOCK(sc);
 690         ifa_free(ifa);
 691
 692         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 693                 CARPSTATS_INC(carps_badauth);
 694                 CARP_DEBUG("%s: incorrect hash for VHID %u@%s\n", __func__,
 695                     sc->sc_vhid, ifp->if_xname);
 696                 goto out;
 697         }
 698
 699         tmp_counter = ntohl(ch->carp_counter[0]);
 700         tmp_counter = tmp_counter<<32;
 701         tmp_counter += ntohl(ch->carp_counter[1]);
 702
 703         /* XXX Replay protection goes here */
 704
 705         sc->sc_init_counter = 0;
 706         sc->sc_counter = tmp_counter;
 707
 708         sc_tv.tv_sec = sc->sc_advbase;
 709         sc_tv.tv_usec = DEMOTE_ADVSKEW(sc) * 1000000 / 256;
 710         ch_tv.tv_sec = ch->carp_advbase;
 711         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 712
 713         switch (sc->sc_state) {
 714         case INIT:
 715                 break;
 716         case MASTER:
 717                 /*
 718                  * If we receive an advertisement from a master who's going to
 719                  * be more frequent than us, go into BACKUP state.
 720                  */
 721                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 722                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 723                         callout_stop(&sc->sc_ad_tmo);
 724                         carp_set_state(sc, BACKUP,
 725                             "more frequent advertisement received");
 726                         carp_setrun(sc, 0);
 727                         carp_delroute(sc);
 728                 }
 729                 break;
 730         case BACKUP:
 731                 /*
 732                  * If we're pre-empting masters who advertise slower than us,
 733                  * and this one claims to be slower, treat him as down.
 734                  */
 735                 if (V_carp_preempt && timevalcmp(&sc_tv, &ch_tv, <)) {
 736                         carp_master_down_locked(sc,
 737                             "preempting a slower master");
 738                         break;
 739                 }
 740
 741                 /*
 742                  *  If the master is going to advertise at such a low frequency
 743                  *  that he's guaranteed to time out, we'd might as well just
 744                  *  treat him as timed out now.
 745                  */
 746                 sc_tv.tv_sec = sc->sc_advbase * 3;
 747                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 748                         carp_master_down_locked(sc, "master will time out");
 749                         break;
 750                 }
 751
 752                 /*
 753                  * Otherwise, we reset the counter and wait for the next
 754                  * advertisement.
 755                  */
 756                 carp_setrun(sc, af);
 757                 break;
 758         }
 759
 760 out:
 761         CARP_UNLOCK(sc);
 762         m_freem(m);
 763 }
 764
 765 static int
 766 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 767 {
 768         struct m_tag *mtag;
 769
 770         if (sc->sc_init_counter) {
 771                 /* this could also be seconds since unix epoch */
 772                 sc->sc_counter = arc4random();
 773                 sc->sc_counter = sc->sc_counter << 32;
 774                 sc->sc_counter += arc4random();
 775         } else
 776                 sc->sc_counter++;
 777
 778         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 779         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 780
 781         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 782
 783         /* Tag packet for carp_output */
 784         if ((mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct carp_softc *),
 785             M_NOWAIT)) == NULL) {
 786                 m_freem(m);
 787                 CARPSTATS_INC(carps_onomem);
 788                 return (ENOMEM);
 789         }
 790         bcopy(&sc, mtag + 1, sizeof(sc));
 791         m_tag_prepend(m, mtag);
 792
 793         return (0);
 794 }
 795
 796 /*
 797  * To avoid LORs and possible recursions this function shouldn't
 798  * be called directly, but scheduled via taskqueue.
 799  */
 800 static void
 801 carp_send_ad_all(void *ctx __unused, int pending __unused)
 802 {
 803         struct carp_softc *sc;
 804
 805         mtx_lock(&carp_mtx);
 806         LIST_FOREACH(sc, &carp_list, sc_next)
 807                 if (sc->sc_state == MASTER) {
 808                         CARP_LOCK(sc);
 809                         CURVNET_SET(sc->sc_carpdev->if_vnet);
 810                         carp_send_ad_locked(sc);
 811                         CURVNET_RESTORE();
 812                         CARP_UNLOCK(sc);
 813                 }
 814         mtx_unlock(&carp_mtx);
 815 }
 816
 817 /* Send a periodic advertisement, executed in callout context. */
 818 static void
 819 carp_send_ad(void *v)
 820 {
 821         struct carp_softc *sc = v;
 822
 823         CARP_LOCK_ASSERT(sc);
 824         CURVNET_SET(sc->sc_carpdev->if_vnet);
 825         carp_send_ad_locked(sc);
 826         CURVNET_RESTORE();
 827         CARP_UNLOCK(sc);
 828 }
 829
 830 static void
 831 carp_send_ad_error(struct carp_softc *sc, int error)
 832 {
 833
 834         if (error) {
 835                 if (sc->sc_sendad_errors < INT_MAX)
 836                         sc->sc_sendad_errors++;
 837                 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 838                         static const char fmt[] = "send error %d on %s";
 839                         char msg[sizeof(fmt) + IFNAMSIZ];
 840
 841                         sprintf(msg, fmt, error, sc->sc_carpdev->if_xname);
 842                         carp_demote_adj(V_carp_senderr_adj, msg);
 843                 }
 844                 sc->sc_sendad_success = 0;
 845         } else {
 846                 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS &&
 847                     ++sc->sc_sendad_success >= CARP_SENDAD_MIN_SUCCESS) {
 848                         static const char fmt[] = "send ok on %s";
 849                         char msg[sizeof(fmt) + IFNAMSIZ];
 850
 851                         sprintf(msg, fmt, sc->sc_carpdev->if_xname);
 852                         carp_demote_adj(-V_carp_senderr_adj, msg);
 853                         sc->sc_sendad_errors = 0;
 854                 } else
 855                         sc->sc_sendad_errors = 0;
 856         }
 857 }
 858
 859 /*
 860  * Pick the best ifaddr on the given ifp for sending CARP
 861  * advertisements.
 862  *
 863  * "Best" here is defined by ifa_preferred().  This function is much
 864  * much like ifaof_ifpforaddr() except that we just use ifa_preferred().
 865  *
 866  * (This could be simplified to return the actual address, except that
 867  * it has a different format in AF_INET and AF_INET6.)
 868  */
 869 static struct ifaddr *
 870 carp_best_ifa(int af, struct ifnet *ifp)
 871 {
 872         struct ifaddr *ifa, *best;
 873
 874         if (af >= AF_MAX)
 875                 return (NULL);
 876         best = NULL;
 877         IF_ADDR_RLOCK(ifp);
 878         TAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
 879                 if (ifa->ifa_addr->sa_family == af &&
 880                     (best == NULL || ifa_preferred(best, ifa)))
 881                         best = ifa;
 882         }
 883         IF_ADDR_RUNLOCK(ifp);
 884         if (best != NULL)
 885                 ifa_ref(best);
 886         return (best);
 887 }
 888
 889 static void
 890 carp_send_ad_locked(struct carp_softc *sc)
 891 {
 892         struct carp_header ch;
 893         struct timeval tv;
 894         struct ifaddr *ifa;
 895         struct carp_header *ch_ptr;
 896         struct mbuf *m;
 897         int len, advskew;
 898
 899         CARP_LOCK_ASSERT(sc);
 900
 901         advskew = DEMOTE_ADVSKEW(sc);
 902         tv.tv_sec = sc->sc_advbase;
 903         tv.tv_usec = advskew * 1000000 / 256;
 904
 905         ch.carp_version = CARP_VERSION;
 906         ch.carp_type = CARP_ADVERTISEMENT;
 907         ch.carp_vhid = sc->sc_vhid;
 908         ch.carp_advbase = sc->sc_advbase;
 909         ch.carp_advskew = advskew;
 910         ch.carp_authlen = 7;    /* XXX DEFINE */
 911         ch.carp_pad1 = 0;       /* must be zero */
 912         ch.carp_cksum = 0;
 913
 914         /* XXXGL: OpenBSD picks first ifaddr with needed family. */
 915
 916 #ifdef INET
 917         if (sc->sc_naddrs) {
 918                 struct ip *ip;
 919
 920                 m = m_gethdr(M_NOWAIT, MT_DATA);
 921                 if (m == NULL) {
 922                         CARPSTATS_INC(carps_onomem);
 923                         goto resched;
 924                 }
 925                 len = sizeof(*ip) + sizeof(ch);
 926                 m->m_pkthdr.len = len;
 927                 m->m_pkthdr.rcvif = NULL;
 928                 m->m_len = len;
 929                 M_ALIGN(m, m->m_len);
 930                 m->m_flags |= M_MCAST;
 931                 ip = mtod(m, struct ip *);
 932                 ip->ip_v = IPVERSION;
 933                 ip->ip_hl = sizeof(*ip) >> 2;
 934                 ip->ip_tos = IPTOS_LOWDELAY;
 935                 ip->ip_len = htons(len);
 936                 ip->ip_off = htons(IP_DF);
 937                 ip->ip_ttl = CARP_DFLTTL;
 938                 ip->ip_p = IPPROTO_CARP;
 939                 ip->ip_sum = 0;
 940                 ip_fillid(ip);
 941
 942                 ifa = carp_best_ifa(AF_INET, sc->sc_carpdev);
 943                 if (ifa != NULL) {
 944                         ip->ip_src.s_addr =
 945                             ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 946                         ifa_free(ifa);
 947                 } else
 948                         ip->ip_src.s_addr = 0;
 949                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 950
 951                 ch_ptr = (struct carp_header *)(&ip[1]);
 952                 bcopy(&ch, ch_ptr, sizeof(ch));
 953                 if (carp_prepare_ad(m, sc, ch_ptr))
 954                         goto resched;
 955
 956                 m->m_data += sizeof(*ip);
 957                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip));
 958                 m->m_data -= sizeof(*ip);
 959
 960                 CARPSTATS_INC(carps_opackets);
 961
 962                 carp_send_ad_error(sc, ip_output(m, NULL, NULL, IP_RAWOUTPUT,
 963                     &sc->sc_carpdev->if_carp->cif_imo, NULL));
 964         }
 965 #endif /* INET */
 966 #ifdef INET6
 967         if (sc->sc_naddrs6) {
 968                 struct ip6_hdr *ip6;
 969
 970                 m = m_gethdr(M_NOWAIT, MT_DATA);
 971                 if (m == NULL) {
 972                         CARPSTATS_INC(carps_onomem);
 973                         goto resched;
 974                 }
 975                 len = sizeof(*ip6) + sizeof(ch);
 976                 m->m_pkthdr.len = len;
 977                 m->m_pkthdr.rcvif = NULL;
 978                 m->m_len = len;
 979                 M_ALIGN(m, m->m_len);
 980                 m->m_flags |= M_MCAST;
 981                 ip6 = mtod(m, struct ip6_hdr *);
 982                 bzero(ip6, sizeof(*ip6));
 983                 ip6->ip6_vfc |= IPV6_VERSION;
 984                 ip6->ip6_hlim = CARP_DFLTTL;
 985                 ip6->ip6_nxt = IPPROTO_CARP;
 986
 987                 /* set the source address */
 988                 ifa = carp_best_ifa(AF_INET6, sc->sc_carpdev);
 989                 if (ifa != NULL) {
 990                         bcopy(IFA_IN6(ifa), &ip6->ip6_src,
 991                             sizeof(struct in6_addr));
 992                         ifa_free(ifa);
 993                 } else
 994                         /* This should never happen with IPv6. */
 995                         bzero(&ip6->ip6_src, sizeof(struct in6_addr));
 996
 997                 /* Set the multicast destination. */
 998                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
 999                 ip6->ip6_dst.s6_addr8[15] = 0x12;
1000                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
1001                         m_freem(m);
1002                         CARP_DEBUG("%s: in6_setscope failed\n", __func__);
1003                         goto resched;
1004                 }
1005
1006                 ch_ptr = (struct carp_header *)(&ip6[1]);
1007                 bcopy(&ch, ch_ptr, sizeof(ch));
1008                 if (carp_prepare_ad(m, sc, ch_ptr))
1009                         goto resched;
1010
1011                 m->m_data += sizeof(*ip6);
1012                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip6));
1013                 m->m_data -= sizeof(*ip6);
1014
1015                 CARPSTATS_INC(carps_opackets6);
1016
1017                 carp_send_ad_error(sc, ip6_output(m, NULL, NULL, 0,
1018                     &sc->sc_carpdev->if_carp->cif_im6o, NULL, NULL));
1019         }
1020 #endif /* INET6 */
1021
1022 resched:
1023         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), carp_send_ad, sc);
1024 }
1025
1026 static void
1027 carp_addroute(struct carp_softc *sc)
1028 {
1029         struct ifaddr *ifa;
1030
1031         CARP_FOREACH_IFA(sc, ifa)
1032                 carp_ifa_addroute(ifa);
1033 }
1034
1035 static void
1036 carp_ifa_addroute(struct ifaddr *ifa)
1037 {
1038
1039         switch (ifa->ifa_addr->sa_family) {
1040 #ifdef INET
1041         case AF_INET:
1042                 in_addprefix(ifatoia(ifa), RTF_UP);
1043                 ifa_add_loopback_route(ifa,
1044                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
1045                 break;
1046 #endif
1047 #ifdef INET6
1048         case AF_INET6:
1049                 ifa_add_loopback_route(ifa,
1050                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
1051                 nd6_add_ifa_lle(ifatoia6(ifa));
1052                 break;
1053 #endif
1054         }
1055 }
1056
1057 static void
1058 carp_delroute(struct carp_softc *sc)
1059 {
1060         struct ifaddr *ifa;
1061
1062         CARP_FOREACH_IFA(sc, ifa)
1063                 carp_ifa_delroute(ifa);
1064 }
1065
1066 static void
1067 carp_ifa_delroute(struct ifaddr *ifa)
1068 {
1069
1070         switch (ifa->ifa_addr->sa_family) {
1071 #ifdef INET
1072         case AF_INET:
1073                 ifa_del_loopback_route(ifa,
1074                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
1075                 in_scrubprefix(ifatoia(ifa), LLE_STATIC);
1076                 break;
1077 #endif
1078 #ifdef INET6
1079         case AF_INET6:
1080                 ifa_del_loopback_route(ifa,
1081                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
1082                 nd6_rem_ifa_lle(ifatoia6(ifa), 1);
1083                 break;
1084 #endif
1085         }
1086 }
1087
1088 int
1089 carp_master(struct ifaddr *ifa)
1090 {
1091         struct carp_softc *sc = ifa->ifa_carp;
1092
1093         return (sc->sc_state == MASTER);
1094 }
1095
1096 #ifdef INET
1097 /*
1098  * Broadcast a gratuitous ARP request containing
1099  * the virtual router MAC address for each IP address
1100  * associated with the virtual router.
1101  */
1102 static void
1103 carp_send_arp(struct carp_softc *sc)
1104 {
1105         struct ifaddr *ifa;
1106         struct in_addr addr;
1107
1108         CARP_FOREACH_IFA(sc, ifa) {
1109                 if (ifa->ifa_addr->sa_family != AF_INET)
1110                         continue;
1111                 addr = ((struct sockaddr_in *)ifa->ifa_addr)->sin_addr;
1112                 arp_announce_ifaddr(sc->sc_carpdev, addr, LLADDR(&sc->sc_addr));
1113         }
1114 }
1115
1116 int
1117 carp_iamatch(struct ifaddr *ifa, uint8_t **enaddr)
1118 {
1119         struct carp_softc *sc = ifa->ifa_carp;
1120
1121         if (sc->sc_state == MASTER) {
1122                 *enaddr = LLADDR(&sc->sc_addr);
1123                 return (1);
1124         }
1125
1126         return (0);
1127 }
1128 #endif
1129
1130 #ifdef INET6
1131 static void
1132 carp_send_na(struct carp_softc *sc)
1133 {
1134         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1135         struct ifaddr *ifa;
1136         struct in6_addr *in6;
1137
1138         CARP_FOREACH_IFA(sc, ifa) {
1139                 if (ifa->ifa_addr->sa_family != AF_INET6)
1140                         continue;
1141
1142                 in6 = IFA_IN6(ifa);
1143                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1144                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1145                 DELAY(1000);    /* XXX */
1146         }
1147 }
1148
1149 /*
1150  * Returns ifa in case it's a carp address and it is MASTER, or if the address
1151  * matches and is not a carp address.  Returns NULL otherwise.
1152  */
1153 struct ifaddr *
1154 carp_iamatch6(struct ifnet *ifp, struct in6_addr *taddr)
1155 {
1156         struct ifaddr *ifa;
1157
1158         ifa = NULL;
1159         IF_ADDR_RLOCK(ifp);
1160         TAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
1161                 if (ifa->ifa_addr->sa_family != AF_INET6)
1162                         continue;
1163                 if (!IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa)))
1164                         continue;
1165                 if (ifa->ifa_carp && ifa->ifa_carp->sc_state != MASTER)
1166                         ifa = NULL;
1167                 else
1168                         ifa_ref(ifa);
1169                 break;
1170         }
1171         IF_ADDR_RUNLOCK(ifp);
1172
1173         return (ifa);
1174 }
1175
1176 caddr_t
1177 carp_macmatch6(struct ifnet *ifp, struct mbuf *m, const struct in6_addr *taddr)
1178 {
1179         struct ifaddr *ifa;
1180
1181         IF_ADDR_RLOCK(ifp);
1182         IFNET_FOREACH_IFA(ifp, ifa)
1183                 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1184                     IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1185                         struct carp_softc *sc = ifa->ifa_carp;
1186                         struct m_tag *mtag;
1187
1188                         IF_ADDR_RUNLOCK(ifp);
1189
1190                         mtag = m_tag_get(PACKET_TAG_CARP,
1191                             sizeof(struct carp_softc *), M_NOWAIT);
1192                         if (mtag == NULL)
1193                                 /* Better a bit than nothing. */
1194                                 return (LLADDR(&sc->sc_addr));
1195
1196                         bcopy(&sc, mtag + 1, sizeof(sc));
1197                         m_tag_prepend(m, mtag);
1198
1199                         return (LLADDR(&sc->sc_addr));
1200                 }
1201         IF_ADDR_RUNLOCK(ifp);
1202
1203         return (NULL);
1204 }
1205 #endif /* INET6 */
1206
1207 int
1208 carp_forus(struct ifnet *ifp, u_char *dhost)
1209 {
1210         struct carp_softc *sc;
1211         uint8_t *ena = dhost;
1212
1213         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1214                 return (0);
1215
1216         CIF_LOCK(ifp->if_carp);
1217         IFNET_FOREACH_CARP(ifp, sc) {
1218                 CARP_LOCK(sc);
1219                 if (sc->sc_state == MASTER && !bcmp(dhost, LLADDR(&sc->sc_addr),
1220                     ETHER_ADDR_LEN)) {
1221                         CARP_UNLOCK(sc);
1222                         CIF_UNLOCK(ifp->if_carp);
1223                         return (1);
1224                 }
1225                 CARP_UNLOCK(sc);
1226         }
1227         CIF_UNLOCK(ifp->if_carp);
1228
1229         return (0);
1230 }
1231
1232 /* Master down timeout event, executed in callout context. */
1233 static void
1234 carp_master_down(void *v)
1235 {
1236         struct carp_softc *sc = v;
1237
1238         CARP_LOCK_ASSERT(sc);
1239
1240         CURVNET_SET(sc->sc_carpdev->if_vnet);
1241         if (sc->sc_state == BACKUP) {
1242                 carp_master_down_locked(sc, "master timed out");
1243         }
1244         CURVNET_RESTORE();
1245
1246         CARP_UNLOCK(sc);
1247 }
1248
1249 static void
1250 carp_master_down_locked(struct carp_softc *sc, const char *reason)
1251 {
1252
1253         CARP_LOCK_ASSERT(sc);
1254
1255         switch (sc->sc_state) {
1256         case BACKUP:
1257                 carp_set_state(sc, MASTER, reason);
1258                 carp_send_ad_locked(sc);
1259 #ifdef INET
1260                 carp_send_arp(sc);
1261 #endif
1262 #ifdef INET6
1263                 carp_send_na(sc);
1264 #endif
1265                 carp_setrun(sc, 0);
1266                 carp_addroute(sc);
1267                 break;
1268         case INIT:
1269         case MASTER:
1270 #ifdef INVARIANTS
1271                 panic("carp: VHID %u@%s: master_down event in %s state\n",
1272                     sc->sc_vhid,
1273                     sc->sc_carpdev->if_xname,
1274                     sc->sc_state ? "MASTER" : "INIT");
1275 #endif
1276                 break;
1277         }
1278 }
1279
1280 /*
1281  * When in backup state, af indicates whether to reset the master down timer
1282  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1283  */
1284 static void
1285 carp_setrun(struct carp_softc *sc, sa_family_t af)
1286 {
1287         struct timeval tv;
1288
1289         CARP_LOCK_ASSERT(sc);
1290
1291         if ((sc->sc_carpdev->if_flags & IFF_UP) == 0 ||
1292             sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1293             (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0))
1294                 return;
1295
1296         switch (sc->sc_state) {
1297         case INIT:
1298                 carp_set_state(sc, BACKUP, "initialization complete");
1299                 carp_setrun(sc, 0);
1300                 break;
1301         case BACKUP:
1302                 callout_stop(&sc->sc_ad_tmo);
1303                 tv.tv_sec = 3 * sc->sc_advbase;
1304                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1305                 switch (af) {
1306 #ifdef INET
1307                 case AF_INET:
1308                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1309                             carp_master_down, sc);
1310                         break;
1311 #endif
1312 #ifdef INET6
1313                 case AF_INET6:
1314                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1315                             carp_master_down, sc);
1316                         break;
1317 #endif
1318                 default:
1319 #ifdef INET
1320                         if (sc->sc_naddrs)
1321                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1322                                     carp_master_down, sc);
1323 #endif
1324 #ifdef INET6
1325                         if (sc->sc_naddrs6)
1326                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1327                                     carp_master_down, sc);
1328 #endif
1329                         break;
1330                 }
1331                 break;
1332         case MASTER:
1333                 tv.tv_sec = sc->sc_advbase;
1334                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1335                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1336                     carp_send_ad, sc);
1337                 break;
1338         }
1339 }
1340
1341 /*
1342  * Setup multicast structures.
1343  */
1344 static int
1345 carp_multicast_setup(struct carp_if *cif, sa_family_t sa)
1346 {
1347         struct ifnet *ifp = cif->cif_ifp;
1348         int error = 0;
1349
1350         switch (sa) {
1351 #ifdef INET
1352         case AF_INET:
1353             {
1354                 struct ip_moptions *imo = &cif->cif_imo;
1355                 struct in_addr addr;
1356
1357                 if (imo->imo_membership)
1358                         return (0);
1359
1360                 imo->imo_membership = (struct in_multi **)malloc(
1361                     (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
1362                     M_WAITOK);
1363                 imo->imo_mfilters = NULL;
1364                 imo->imo_max_memberships = IP_MIN_MEMBERSHIPS;
1365                 imo->imo_multicast_vif = -1;
1366
1367                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1368                 if ((error = in_joingroup(ifp, &addr, NULL,
1369                     &imo->imo_membership[0])) != 0) {
1370                         free(imo->imo_membership, M_CARP);
1371                         break;
1372                 }
1373                 imo->imo_num_memberships++;
1374                 imo->imo_multicast_ifp = ifp;
1375                 imo->imo_multicast_ttl = CARP_DFLTTL;
1376                 imo->imo_multicast_loop = 0;
1377                 break;
1378            }
1379 #endif
1380 #ifdef INET6
1381         case AF_INET6:
1382             {
1383                 struct ip6_moptions *im6o = &cif->cif_im6o;
1384                 struct in6_addr in6;
1385                 struct in6_multi *in6m;
1386
1387                 if (im6o->im6o_membership)
1388                         return (0);
1389
1390                 im6o->im6o_membership = (struct in6_multi **)malloc(
1391                     (sizeof(struct in6_multi *) * IPV6_MIN_MEMBERSHIPS), M_CARP,
1392                     M_ZERO | M_WAITOK);
1393                 im6o->im6o_mfilters = NULL;
1394                 im6o->im6o_max_memberships = IPV6_MIN_MEMBERSHIPS;
1395                 im6o->im6o_multicast_hlim = CARP_DFLTTL;
1396                 im6o->im6o_multicast_ifp = ifp;
1397
1398                 /* Join IPv6 CARP multicast group. */
1399                 bzero(&in6, sizeof(in6));
1400                 in6.s6_addr16[0] = htons(0xff02);
1401                 in6.s6_addr8[15] = 0x12;
1402                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1403                         free(im6o->im6o_membership, M_CARP);
1404                         break;
1405                 }
1406                 in6m = NULL;
1407                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1408                         free(im6o->im6o_membership, M_CARP);
1409                         break;
1410                 }
1411                 im6o->im6o_membership[0] = in6m;
1412                 im6o->im6o_num_memberships++;
1413
1414                 /* Join solicited multicast address. */
1415                 bzero(&in6, sizeof(in6));
1416                 in6.s6_addr16[0] = htons(0xff02);
1417                 in6.s6_addr32[1] = 0;
1418                 in6.s6_addr32[2] = htonl(1);
1419                 in6.s6_addr32[3] = 0;
1420                 in6.s6_addr8[12] = 0xff;
1421                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1422                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1423                         free(im6o->im6o_membership, M_CARP);
1424                         break;
1425                 }
1426                 in6m = NULL;
1427                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1428                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1429                         free(im6o->im6o_membership, M_CARP);
1430                         break;
1431                 }
1432                 im6o->im6o_membership[1] = in6m;
1433                 im6o->im6o_num_memberships++;
1434                 break;
1435             }
1436 #endif
1437         }
1438
1439         return (error);
1440 }
1441
1442 /*
1443  * Free multicast structures.
1444  */
1445 static void
1446 carp_multicast_cleanup(struct carp_if *cif, sa_family_t sa)
1447 {
1448
1449         sx_assert(&carp_sx, SA_XLOCKED);
1450
1451         switch (sa) {
1452 #ifdef INET
1453         case AF_INET:
1454                 if (cif->cif_naddrs == 0) {
1455                         struct ip_moptions *imo = &cif->cif_imo;
1456
1457                         in_leavegroup(imo->imo_membership[0], NULL);
1458                         KASSERT(imo->imo_mfilters == NULL,
1459                             ("%s: imo_mfilters != NULL", __func__));
1460                         free(imo->imo_membership, M_CARP);
1461                         imo->imo_membership = NULL;
1462
1463                 }
1464                 break;
1465 #endif
1466 #ifdef INET6
1467         case AF_INET6:
1468                 if (cif->cif_naddrs6 == 0) {
1469                         struct ip6_moptions *im6o = &cif->cif_im6o;
1470
1471                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1472                         in6_mc_leave(im6o->im6o_membership[1], NULL);
1473                         KASSERT(im6o->im6o_mfilters == NULL,
1474                             ("%s: im6o_mfilters != NULL", __func__));
1475                         free(im6o->im6o_membership, M_CARP);
1476                         im6o->im6o_membership = NULL;
1477                 }
1478                 break;
1479 #endif
1480         }
1481 }
1482
1483 int
1484 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa)
1485 {
1486         struct m_tag *mtag;
1487         struct carp_softc *sc;
1488
1489         if (!sa)
1490                 return (0);
1491
1492         switch (sa->sa_family) {
1493 #ifdef INET
1494         case AF_INET:
1495                 break;
1496 #endif
1497 #ifdef INET6
1498         case AF_INET6:
1499                 break;
1500 #endif
1501         default:
1502                 return (0);
1503         }
1504
1505         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
1506         if (mtag == NULL)
1507                 return (0);
1508
1509         bcopy(mtag + 1, &sc, sizeof(sc));
1510
1511         /* Set the source MAC address to the Virtual Router MAC Address. */
1512         switch (ifp->if_type) {
1513         case IFT_ETHER:
1514         case IFT_BRIDGE:
1515         case IFT_L2VLAN: {
1516                         struct ether_header *eh;
1517
1518                         eh = mtod(m, struct ether_header *);
1519                         eh->ether_shost[0] = 0;
1520                         eh->ether_shost[1] = 0;
1521                         eh->ether_shost[2] = 0x5e;
1522                         eh->ether_shost[3] = 0;
1523                         eh->ether_shost[4] = 1;
1524                         eh->ether_shost[5] = sc->sc_vhid;
1525                 }
1526                 break;
1527         case IFT_FDDI: {
1528                         struct fddi_header *fh;
1529
1530                         fh = mtod(m, struct fddi_header *);
1531                         fh->fddi_shost[0] = 0;
1532                         fh->fddi_shost[1] = 0;
1533                         fh->fddi_shost[2] = 0x5e;
1534                         fh->fddi_shost[3] = 0;
1535                         fh->fddi_shost[4] = 1;
1536                         fh->fddi_shost[5] = sc->sc_vhid;
1537                 }
1538                 break;
1539         case IFT_ISO88025: {
1540                         struct iso88025_header *th;
1541                         th = mtod(m, struct iso88025_header *);
1542                         th->iso88025_shost[0] = 3;
1543                         th->iso88025_shost[1] = 0;
1544                         th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
1545                         th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
1546                         th->iso88025_shost[4] = 0;
1547                         th->iso88025_shost[5] = 0;
1548                 }
1549                 break;
1550         default:
1551                 printf("%s: carp is not supported for the %d interface type\n",
1552                     ifp->if_xname, ifp->if_type);
1553                 return (EOPNOTSUPP);
1554         }
1555
1556         return (0);
1557 }
1558
1559 static struct carp_softc*
1560 carp_alloc(struct ifnet *ifp)
1561 {
1562         struct carp_softc *sc;
1563         struct carp_if *cif;
1564
1565         if ((cif = ifp->if_carp) == NULL)
1566                 cif = carp_alloc_if(ifp);
1567
1568         sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
1569
1570         sc->sc_advbase = CARP_DFLTINTV;
1571         sc->sc_vhid = -1;       /* required setting */
1572         sc->sc_init_counter = 1;
1573         sc->sc_state = INIT;
1574
1575         sc->sc_ifasiz = sizeof(struct ifaddr *);
1576         sc->sc_ifas = malloc(sc->sc_ifasiz, M_CARP, M_WAITOK|M_ZERO);
1577         sc->sc_carpdev = ifp;
1578
1579         CARP_LOCK_INIT(sc);
1580 #ifdef INET
1581         callout_init_mtx(&sc->sc_md_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1582 #endif
1583 #ifdef INET6
1584         callout_init_mtx(&sc->sc_md6_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1585 #endif
1586         callout_init_mtx(&sc->sc_ad_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1587
1588         CIF_LOCK(cif);
1589         TAILQ_INSERT_TAIL(&cif->cif_vrs, sc, sc_list);
1590         CIF_UNLOCK(cif);
1591
1592         mtx_lock(&carp_mtx);
1593         LIST_INSERT_HEAD(&carp_list, sc, sc_next);
1594         mtx_unlock(&carp_mtx);
1595
1596         return (sc);
1597 }
1598
1599 static void
1600 carp_grow_ifas(struct carp_softc *sc)
1601 {
1602         struct ifaddr **new;
1603
1604         new = malloc(sc->sc_ifasiz * 2, M_CARP, M_WAITOK | M_ZERO);
1605         CARP_LOCK(sc);
1606         bcopy(sc->sc_ifas, new, sc->sc_ifasiz);
1607         free(sc->sc_ifas, M_CARP);
1608         sc->sc_ifas = new;
1609         sc->sc_ifasiz *= 2;
1610         CARP_UNLOCK(sc);
1611 }
1612
1613 static void
1614 carp_destroy(struct carp_softc *sc)
1615 {
1616         struct ifnet *ifp = sc->sc_carpdev;
1617         struct carp_if *cif = ifp->if_carp;
1618
1619         sx_assert(&carp_sx, SA_XLOCKED);
1620
1621         if (sc->sc_suppress)
1622                 carp_demote_adj(-V_carp_ifdown_adj, "vhid removed");
1623         CARP_UNLOCK(sc);
1624
1625         CIF_LOCK(cif);
1626         TAILQ_REMOVE(&cif->cif_vrs, sc, sc_list);
1627         CIF_UNLOCK(cif);
1628
1629         mtx_lock(&carp_mtx);
1630         LIST_REMOVE(sc, sc_next);
1631         mtx_unlock(&carp_mtx);
1632
1633         callout_drain(&sc->sc_ad_tmo);
1634 #ifdef INET
1635         callout_drain(&sc->sc_md_tmo);
1636 #endif
1637 #ifdef INET6
1638         callout_drain(&sc->sc_md6_tmo);
1639 #endif
1640         CARP_LOCK_DESTROY(sc);
1641
1642         free(sc->sc_ifas, M_CARP);
1643         free(sc, M_CARP);
1644 }
1645
1646 static struct carp_if*
1647 carp_alloc_if(struct ifnet *ifp)
1648 {
1649         struct carp_if *cif;
1650         int error;
1651
1652         cif = malloc(sizeof(*cif), M_CARP, M_WAITOK|M_ZERO);
1653
1654         if ((error = ifpromisc(ifp, 1)) != 0)
1655                 printf("%s: ifpromisc(%s) failed: %d\n",
1656                     __func__, ifp->if_xname, error);
1657         else
1658                 cif->cif_flags |= CIF_PROMISC;
1659
1660         CIF_LOCK_INIT(cif);
1661         cif->cif_ifp = ifp;
1662         TAILQ_INIT(&cif->cif_vrs);
1663
1664         IF_ADDR_WLOCK(ifp);
1665         ifp->if_carp = cif;
1666         if_ref(ifp);
1667         IF_ADDR_WUNLOCK(ifp);
1668
1669         return (cif);
1670 }
1671
1672 static void
1673 carp_free_if(struct carp_if *cif)
1674 {
1675         struct ifnet *ifp = cif->cif_ifp;
1676
1677         CIF_LOCK_ASSERT(cif);
1678         KASSERT(TAILQ_EMPTY(&cif->cif_vrs), ("%s: softc list not empty",
1679             __func__));
1680
1681         IF_ADDR_WLOCK(ifp);
1682         ifp->if_carp = NULL;
1683         IF_ADDR_WUNLOCK(ifp);
1684
1685         CIF_LOCK_DESTROY(cif);
1686
1687         if (cif->cif_flags & CIF_PROMISC)
1688                 ifpromisc(ifp, 0);
1689         if_rele(ifp);
1690
1691         free(cif, M_CARP);
1692 }
1693
1694 static void
1695 carp_carprcp(struct carpreq *carpr, struct carp_softc *sc, int priv)
1696 {
1697
1698         CARP_LOCK(sc);
1699         carpr->carpr_state = sc->sc_state;
1700         carpr->carpr_vhid = sc->sc_vhid;
1701         carpr->carpr_advbase = sc->sc_advbase;
1702         carpr->carpr_advskew = sc->sc_advskew;
1703         if (priv)
1704                 bcopy(sc->sc_key, carpr->carpr_key, sizeof(carpr->carpr_key));
1705         else
1706                 bzero(carpr->carpr_key, sizeof(carpr->carpr_key));
1707         CARP_UNLOCK(sc);
1708 }
1709
1710 int
1711 carp_ioctl(struct ifreq *ifr, u_long cmd, struct thread *td)
1712 {
1713         struct carpreq carpr;
1714         struct ifnet *ifp;
1715         struct carp_softc *sc = NULL;
1716         int error = 0, locked = 0;
1717
1718         if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1719                 return (error);
1720
1721         ifp = ifunit_ref(ifr->ifr_name);
1722         if (ifp == NULL)
1723                 return (ENXIO);
1724
1725         switch (ifp->if_type) {
1726         case IFT_ETHER:
1727         case IFT_L2VLAN:
1728         case IFT_BRIDGE:
1729         case IFT_FDDI:
1730         case IFT_ISO88025:
1731                 break;
1732         default:
1733                 error = EOPNOTSUPP;
1734                 goto out;
1735         }
1736
1737         if ((ifp->if_flags & IFF_MULTICAST) == 0) {
1738                 error = EADDRNOTAVAIL;
1739                 goto out;
1740         }
1741
1742         sx_xlock(&carp_sx);
1743         switch (cmd) {
1744         case SIOCSVH:
1745                 if ((error = priv_check(td, PRIV_NETINET_CARP)))
1746                         break;
1747                 if (carpr.carpr_vhid <= 0 || carpr.carpr_vhid > CARP_MAXVHID ||
1748                     carpr.carpr_advbase < 0 || carpr.carpr_advskew < 0) {
1749                         error = EINVAL;
1750                         break;
1751                 }
1752
1753                 if (ifp->if_carp) {
1754                         CIF_LOCK(ifp->if_carp);
1755                         IFNET_FOREACH_CARP(ifp, sc)
1756                                 if (sc->sc_vhid == carpr.carpr_vhid)
1757                                         break;
1758                         CIF_UNLOCK(ifp->if_carp);
1759                 }
1760                 if (sc == NULL) {
1761                         sc = carp_alloc(ifp);
1762                         CARP_LOCK(sc);
1763                         sc->sc_vhid = carpr.carpr_vhid;
1764                         LLADDR(&sc->sc_addr)[0] = 0;
1765                         LLADDR(&sc->sc_addr)[1] = 0;
1766                         LLADDR(&sc->sc_addr)[2] = 0x5e;
1767                         LLADDR(&sc->sc_addr)[3] = 0;
1768                         LLADDR(&sc->sc_addr)[4] = 1;
1769                         LLADDR(&sc->sc_addr)[5] = sc->sc_vhid;
1770                 } else
1771                         CARP_LOCK(sc);
1772                 locked = 1;
1773                 if (carpr.carpr_advbase > 0) {
1774                         if (carpr.carpr_advbase > 255 ||
1775                             carpr.carpr_advbase < CARP_DFLTINTV) {
1776                                 error = EINVAL;
1777                                 break;
1778                         }
1779                         sc->sc_advbase = carpr.carpr_advbase;
1780                 }
1781                 if (carpr.carpr_advskew >= 255) {
1782                         error = EINVAL;
1783                         break;
1784                 }
1785                 sc->sc_advskew = carpr.carpr_advskew;
1786                 if (carpr.carpr_key[0] != '\0') {
1787                         bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1788                         carp_hmac_prepare(sc);
1789                 }
1790                 if (sc->sc_state != INIT &&
1791                     carpr.carpr_state != sc->sc_state) {
1792                         switch (carpr.carpr_state) {
1793                         case BACKUP:
1794                                 callout_stop(&sc->sc_ad_tmo);
1795                                 carp_set_state(sc, BACKUP,
1796                                     "user requested via ifconfig");
1797                                 carp_setrun(sc, 0);
1798                                 carp_delroute(sc);
1799                                 break;
1800                         case MASTER:
1801                                 carp_master_down_locked(sc,
1802                                     "user requested via ifconfig");
1803                                 break;
1804                         default:
1805                                 break;
1806                         }
1807                 }
1808                 break;
1809
1810         case SIOCGVH:
1811             {
1812                 int priveleged;
1813
1814                 if (carpr.carpr_vhid < 0 || carpr.carpr_vhid > CARP_MAXVHID) {
1815                         error = EINVAL;
1816                         break;
1817                 }
1818                 if (carpr.carpr_count < 1) {
1819                         error = EMSGSIZE;
1820                         break;
1821                 }
1822                 if (ifp->if_carp == NULL) {
1823                         error = ENOENT;
1824                         break;
1825                 }
1826
1827                 priveleged = (priv_check(td, PRIV_NETINET_CARP) == 0);
1828                 if (carpr.carpr_vhid != 0) {
1829                         CIF_LOCK(ifp->if_carp);
1830                         IFNET_FOREACH_CARP(ifp, sc)
1831                                 if (sc->sc_vhid == carpr.carpr_vhid)
1832                                         break;
1833                         CIF_UNLOCK(ifp->if_carp);
1834                         if (sc == NULL) {
1835                                 error = ENOENT;
1836                                 break;
1837                         }
1838                         carp_carprcp(&carpr, sc, priveleged);
1839                         error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1840                 } else  {
1841                         int i, count;
1842
1843                         count = 0;
1844                         CIF_LOCK(ifp->if_carp);
1845                         IFNET_FOREACH_CARP(ifp, sc)
1846                                 count++;
1847
1848                         if (count > carpr.carpr_count) {
1849                                 CIF_UNLOCK(ifp->if_carp);
1850                                 error = EMSGSIZE;
1851                                 break;
1852                         }
1853
1854                         i = 0;
1855                         IFNET_FOREACH_CARP(ifp, sc) {
1856                                 carp_carprcp(&carpr, sc, priveleged);
1857                                 carpr.carpr_count = count;
1858                                 error = copyout(&carpr, ifr->ifr_data +
1859                                     (i * sizeof(carpr)), sizeof(carpr));
1860                                 if (error) {
1861                                         CIF_UNLOCK(ifp->if_carp);
1862                                         break;
1863                                 }
1864                                 i++;
1865                         }
1866                         CIF_UNLOCK(ifp->if_carp);
1867                 }
1868                 break;
1869             }
1870         default:
1871                 error = EINVAL;
1872         }
1873         sx_xunlock(&carp_sx);
1874
1875 out:
1876         if (locked)
1877                 CARP_UNLOCK(sc);
1878         if_rele(ifp);
1879
1880         return (error);
1881 }
1882
1883 static int
1884 carp_get_vhid(struct ifaddr *ifa)
1885 {
1886
1887         if (ifa == NULL || ifa->ifa_carp == NULL)
1888                 return (0);
1889
1890         return (ifa->ifa_carp->sc_vhid);
1891 }
1892
1893 int
1894 carp_attach(struct ifaddr *ifa, int vhid)
1895 {
1896         struct ifnet *ifp = ifa->ifa_ifp;
1897         struct carp_if *cif = ifp->if_carp;
1898         struct carp_softc *sc;
1899         int index, error;
1900
1901         KASSERT(ifa->ifa_carp == NULL, ("%s: ifa %p attached", __func__, ifa));
1902
1903         switch (ifa->ifa_addr->sa_family) {
1904 #ifdef INET
1905         case AF_INET:
1906 #endif
1907 #ifdef INET6
1908         case AF_INET6:
1909 #endif
1910                 break;
1911         default:
1912                 return (EPROTOTYPE);
1913         }
1914
1915         sx_xlock(&carp_sx);
1916         if (ifp->if_carp == NULL) {
1917                 sx_xunlock(&carp_sx);
1918                 return (ENOPROTOOPT);
1919         }
1920
1921         CIF_LOCK(cif);
1922         IFNET_FOREACH_CARP(ifp, sc)
1923                 if (sc->sc_vhid == vhid)
1924                         break;
1925         CIF_UNLOCK(cif);
1926         if (sc == NULL) {
1927                 sx_xunlock(&carp_sx);
1928                 return (ENOENT);
1929         }
1930
1931         error = carp_multicast_setup(cif, ifa->ifa_addr->sa_family);
1932         if (error) {
1933                 CIF_FREE(cif);
1934                 sx_xunlock(&carp_sx);
1935                 return (error);
1936         }
1937
1938         index = sc->sc_naddrs + sc->sc_naddrs6 + 1;
1939         if (index > sc->sc_ifasiz / sizeof(struct ifaddr *))
1940                 carp_grow_ifas(sc);
1941
1942         switch (ifa->ifa_addr->sa_family) {
1943 #ifdef INET
1944         case AF_INET:
1945                 cif->cif_naddrs++;
1946                 sc->sc_naddrs++;
1947                 break;
1948 #endif
1949 #ifdef INET6
1950         case AF_INET6:
1951                 cif->cif_naddrs6++;
1952                 sc->sc_naddrs6++;
1953                 break;
1954 #endif
1955         }
1956
1957         ifa_ref(ifa);
1958
1959         CARP_LOCK(sc);
1960         sc->sc_ifas[index - 1] = ifa;
1961         ifa->ifa_carp = sc;
1962         carp_hmac_prepare(sc);
1963         carp_sc_state(sc);
1964         CARP_UNLOCK(sc);
1965
1966         sx_xunlock(&carp_sx);
1967
1968         return (0);
1969 }
1970
1971 void
1972 carp_detach(struct ifaddr *ifa)
1973 {
1974         struct ifnet *ifp = ifa->ifa_ifp;
1975         struct carp_if *cif = ifp->if_carp;
1976         struct carp_softc *sc = ifa->ifa_carp;
1977         int i, index;
1978
1979         KASSERT(sc != NULL, ("%s: %p not attached", __func__, ifa));
1980
1981         sx_xlock(&carp_sx);
1982
1983         CARP_LOCK(sc);
1984         /* Shift array. */
1985         index = sc->sc_naddrs + sc->sc_naddrs6;
1986         for (i = 0; i < index; i++)
1987                 if (sc->sc_ifas[i] == ifa)
1988                         break;
1989         KASSERT(i < index, ("%s: %p no backref", __func__, ifa));
1990         for (; i < index - 1; i++)
1991                 sc->sc_ifas[i] = sc->sc_ifas[i+1];
1992         sc->sc_ifas[index - 1] = NULL;
1993
1994         switch (ifa->ifa_addr->sa_family) {
1995 #ifdef INET
1996         case AF_INET:
1997                 cif->cif_naddrs--;
1998                 sc->sc_naddrs--;
1999                 break;
2000 #endif
2001 #ifdef INET6
2002         case AF_INET6:
2003                 cif->cif_naddrs6--;
2004                 sc->sc_naddrs6--;
2005                 break;
2006 #endif
2007         }
2008
2009         carp_ifa_delroute(ifa);
2010         carp_multicast_cleanup(cif, ifa->ifa_addr->sa_family);
2011
2012         ifa->ifa_carp = NULL;
2013         ifa_free(ifa);
2014
2015         carp_hmac_prepare(sc);
2016         carp_sc_state(sc);
2017
2018         if (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0)
2019                 carp_destroy(sc);
2020         else
2021                 CARP_UNLOCK(sc);
2022
2023         CIF_FREE(cif);
2024
2025         sx_xunlock(&carp_sx);
2026 }
2027
2028 static void
2029 carp_set_state(struct carp_softc *sc, int state, const char *reason)
2030 {
2031
2032         CARP_LOCK_ASSERT(sc);
2033
2034         if (sc->sc_state != state) {
2035                 const char *carp_states[] = { CARP_STATES };
2036                 char subsys[IFNAMSIZ+5];
2037
2038                 snprintf(subsys, IFNAMSIZ+5, "%u@%s", sc->sc_vhid,
2039                     sc->sc_carpdev->if_xname);
2040
2041                 CARP_LOG("%s: %s -> %s (%s)\n", subsys,
2042                     carp_states[sc->sc_state], carp_states[state], reason);
2043
2044                 sc->sc_state = state;
2045
2046                 devctl_notify("CARP", subsys, carp_states[state], NULL);
2047         }
2048 }
2049
2050 static void
2051 carp_linkstate(struct ifnet *ifp)
2052 {
2053         struct carp_softc *sc;
2054
2055         CIF_LOCK(ifp->if_carp);
2056         IFNET_FOREACH_CARP(ifp, sc) {
2057                 CARP_LOCK(sc);
2058                 carp_sc_state(sc);
2059                 CARP_UNLOCK(sc);
2060         }
2061         CIF_UNLOCK(ifp->if_carp);
2062 }
2063
2064 static void
2065 carp_sc_state(struct carp_softc *sc)
2066 {
2067
2068         CARP_LOCK_ASSERT(sc);
2069
2070         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
2071             !(sc->sc_carpdev->if_flags & IFF_UP)) {
2072                 callout_stop(&sc->sc_ad_tmo);
2073 #ifdef INET
2074                 callout_stop(&sc->sc_md_tmo);
2075 #endif
2076 #ifdef INET6
2077                 callout_stop(&sc->sc_md6_tmo);
2078 #endif
2079                 carp_set_state(sc, INIT, "hardware interface down");
2080                 carp_setrun(sc, 0);
2081                 if (!sc->sc_suppress)
2082                         carp_demote_adj(V_carp_ifdown_adj, "interface down");
2083                 sc->sc_suppress = 1;
2084         } else {
2085                 carp_set_state(sc, INIT, "hardware interface up");
2086                 carp_setrun(sc, 0);
2087                 if (sc->sc_suppress)
2088                         carp_demote_adj(-V_carp_ifdown_adj, "interface up");
2089                 sc->sc_suppress = 0;
2090         }
2091 }
2092
2093 static void
2094 carp_demote_adj(int adj, char *reason)
2095 {
2096         atomic_add_int(&V_carp_demotion, adj);
2097         CARP_LOG("demoted by %d to %d (%s)\n", adj, V_carp_demotion, reason);
2098         taskqueue_enqueue(taskqueue_swi, &carp_sendall_task);
2099 }
2100
2101 static int
2102 carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS)
2103 {
2104         int new, error;
2105
2106         new = V_carp_demotion;
2107         error = sysctl_handle_int(oidp, &new, 0, req);
2108         if (error || !req->newptr)
2109                 return (error);
2110
2111         carp_demote_adj(new, "sysctl");
2112
2113         return (0);
2114 }
2115
2116 #ifdef INET
2117 extern  struct domain inetdomain;
2118 static struct protosw in_carp_protosw = {
2119         .pr_type =              SOCK_RAW,
2120         .pr_domain =            &inetdomain,
2121         .pr_protocol =          IPPROTO_CARP,
2122         .pr_flags =             PR_ATOMIC|PR_ADDR,
2123         .pr_input =             carp_input,
2124         .pr_output =            rip_output,
2125         .pr_ctloutput =         rip_ctloutput,
2126         .pr_usrreqs =           &rip_usrreqs
2127 };
2128 #endif
2129
2130 #ifdef INET6
2131 extern  struct domain inet6domain;
2132 static struct protosw in6_carp_protosw = {
2133         .pr_type =              SOCK_RAW,
2134         .pr_domain =            &inet6domain,
2135         .pr_protocol =          IPPROTO_CARP,
2136         .pr_flags =             PR_ATOMIC|PR_ADDR,
2137         .pr_input =             carp6_input,
2138         .pr_output =            rip6_output,
2139         .pr_ctloutput =         rip6_ctloutput,
2140         .pr_usrreqs =           &rip6_usrreqs
2141 };
2142 #endif
2143
2144 static void
2145 carp_mod_cleanup(void)
2146 {
2147
2148 #ifdef INET
2149         if (proto_reg[CARP_INET] == 0) {
2150                 (void)ipproto_unregister(IPPROTO_CARP);
2151                 pf_proto_unregister(PF_INET, IPPROTO_CARP, SOCK_RAW);
2152                 proto_reg[CARP_INET] = -1;
2153         }
2154         carp_iamatch_p = NULL;
2155 #endif
2156 #ifdef INET6
2157         if (proto_reg[CARP_INET6] == 0) {
2158                 (void)ip6proto_unregister(IPPROTO_CARP);
2159                 pf_proto_unregister(PF_INET6, IPPROTO_CARP, SOCK_RAW);
2160                 proto_reg[CARP_INET6] = -1;
2161         }
2162         carp_iamatch6_p = NULL;
2163         carp_macmatch6_p = NULL;
2164 #endif
2165         carp_ioctl_p = NULL;
2166         carp_attach_p = NULL;
2167         carp_detach_p = NULL;
2168         carp_get_vhid_p = NULL;
2169         carp_linkstate_p = NULL;
2170         carp_forus_p = NULL;
2171         carp_output_p = NULL;
2172         carp_demote_adj_p = NULL;
2173         carp_master_p = NULL;
2174         mtx_unlock(&carp_mtx);
2175         taskqueue_drain(taskqueue_swi, &carp_sendall_task);
2176         mtx_destroy(&carp_mtx);
2177         sx_destroy(&carp_sx);
2178 }
2179
2180 static int
2181 carp_mod_load(void)
2182 {
2183         int err;
2184
2185         mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2186         sx_init(&carp_sx, "carp_sx");
2187         LIST_INIT(&carp_list);
2188         carp_get_vhid_p = carp_get_vhid;
2189         carp_forus_p = carp_forus;
2190         carp_output_p = carp_output;
2191         carp_linkstate_p = carp_linkstate;
2192         carp_ioctl_p = carp_ioctl;
2193         carp_attach_p = carp_attach;
2194         carp_detach_p = carp_detach;
2195         carp_demote_adj_p = carp_demote_adj;
2196         carp_master_p = carp_master;
2197 #ifdef INET6
2198         carp_iamatch6_p = carp_iamatch6;
2199         carp_macmatch6_p = carp_macmatch6;
2200         proto_reg[CARP_INET6] = pf_proto_register(PF_INET6,
2201             (struct protosw *)&in6_carp_protosw);
2202         if (proto_reg[CARP_INET6]) {
2203                 printf("carp: error %d attaching to PF_INET6\n",
2204                     proto_reg[CARP_INET6]);
2205                 carp_mod_cleanup();
2206                 return (proto_reg[CARP_INET6]);
2207         }
2208         err = ip6proto_register(IPPROTO_CARP);
2209         if (err) {
2210                 printf("carp: error %d registering with INET6\n", err);
2211                 carp_mod_cleanup();
2212                 return (err);
2213         }
2214 #endif
2215 #ifdef INET
2216         carp_iamatch_p = carp_iamatch;
2217         proto_reg[CARP_INET] = pf_proto_register(PF_INET, &in_carp_protosw);
2218         if (proto_reg[CARP_INET]) {
2219                 printf("carp: error %d attaching to PF_INET\n",
2220                     proto_reg[CARP_INET]);
2221                 carp_mod_cleanup();
2222                 return (proto_reg[CARP_INET]);
2223         }
2224         err = ipproto_register(IPPROTO_CARP);
2225         if (err) {
2226                 printf("carp: error %d registering with INET\n", err);
2227                 carp_mod_cleanup();
2228                 return (err);
2229         }
2230 #endif
2231         return (0);
2232 }
2233
2234 static int
2235 carp_modevent(module_t mod, int type, void *data)
2236 {
2237         switch (type) {
2238         case MOD_LOAD:
2239                 return carp_mod_load();
2240                 /* NOTREACHED */
2241         case MOD_UNLOAD:
2242                 mtx_lock(&carp_mtx);
2243                 if (LIST_EMPTY(&carp_list))
2244                         carp_mod_cleanup();
2245                 else {
2246                         mtx_unlock(&carp_mtx);
2247                         return (EBUSY);
2248                 }
2249                 break;
2250
2251         default:
2252                 return (EINVAL);
2253         }
2254
2255         return (0);
2256 }
2257
2258 static moduledata_t carp_mod = {
2259         "carp",
2260         carp_modevent,
2261         0
2262 };
2263
2264 DECLARE_MODULE(carp, carp_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);