sys/netinet/ip_carp.c

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