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