sys/netinet/ip_carp.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2002 Michael Shalayeff.
   5  * Copyright (c) 2003 Ryan McBride.
   6  * Copyright (c) 2011 Gleb Smirnoff <glebius@FreeBSD.org>
   7  * All rights reserved.
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  *
  18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  19  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  20  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  21  * IN NO EVENT SHALL THE AUTHOR OR HIS RELATIVES BE LIABLE FOR ANY DIRECT,
  22  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  23  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  24  * SERVICES; LOSS OF MIND, USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
  26  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
  27  * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
  28  * THE POSSIBILITY OF SUCH DAMAGE.
  29  */
  30
  31 #include <sys/cdefs.h>
  32 __FBSDID("$FreeBSD$");
  33
  34 #include "opt_bpf.h"
  35 #include "opt_inet.h"
  36 #include "opt_inet6.h"
  37
  38 #include <sys/param.h>
  39 #include <sys/systm.h>
  40 #include <sys/bus.h>
  41 #include <sys/jail.h>
  42 #include <sys/kernel.h>
  43 #include <sys/limits.h>
  44 #include <sys/malloc.h>
  45 #include <sys/mbuf.h>
  46 #include <sys/module.h>
  47 #include <sys/priv.h>
  48 #include <sys/proc.h>
  49 #include <sys/protosw.h>
  50 #include <sys/socket.h>
  51 #include <sys/sockio.h>
  52 #include <sys/sysctl.h>
  53 #include <sys/syslog.h>
  54 #include <sys/taskqueue.h>
  55 #include <sys/counter.h>
  56
  57 #include <net/ethernet.h>
  58 #include <net/if.h>
  59 #include <net/if_var.h>
  60 #include <net/if_dl.h>
  61 #include <net/if_llatbl.h>
  62 #include <net/if_types.h>
  63 #include <net/route.h>
  64 #include <net/vnet.h>
  65
  66 #if defined(INET) || defined(INET6)
  67 #include <netinet/in.h>
  68 #include <netinet/in_var.h>
  69 #include <netinet/ip_carp.h>
  70 #include <netinet/ip.h>
  71 #include <machine/in_cksum.h>
  72 #endif
  73 #ifdef INET
  74 #include <netinet/ip_var.h>
  75 #include <netinet/if_ether.h>
  76 #endif
  77
  78 #ifdef INET6
  79 #include <netinet/icmp6.h>
  80 #include <netinet/ip6.h>
  81 #include <netinet6/in6_var.h>
  82 #include <netinet6/ip6_var.h>
  83 #include <netinet6/scope6_var.h>
  84 #include <netinet6/nd6.h>
  85 #endif
  86
  87 #include <crypto/sha1.h>
  88
  89 static MALLOC_DEFINE(M_CARP, "CARP", "CARP addresses");
  90
  91 struct carp_softc {
  92         struct ifnet            *sc_carpdev;    /* Pointer to parent ifnet. */
  93         struct ifaddr           **sc_ifas;      /* Our ifaddrs. */
  94         struct sockaddr_dl      sc_addr;        /* Our link level address. */
  95         struct callout          sc_ad_tmo;      /* Advertising timeout. */
  96 #ifdef INET
  97         struct callout          sc_md_tmo;      /* Master down timeout. */
  98 #endif
  99 #ifdef INET6
 100         struct callout          sc_md6_tmo;     /* XXX: Master down timeout. */
 101 #endif
 102         struct mtx              sc_mtx;
 103
 104         int                     sc_vhid;
 105         int                     sc_advskew;
 106         int                     sc_advbase;
 107
 108         int                     sc_naddrs;
 109         int                     sc_naddrs6;
 110         int                     sc_ifasiz;
 111         enum { INIT = 0, BACKUP, MASTER }       sc_state;
 112         int                     sc_suppress;
 113         int                     sc_sendad_errors;
 114 #define CARP_SENDAD_MAX_ERRORS  3
 115         int                     sc_sendad_success;
 116 #define CARP_SENDAD_MIN_SUCCESS 3
 117
 118         int                     sc_init_counter;
 119         uint64_t                sc_counter;
 120
 121         /* authentication */
 122 #define CARP_HMAC_PAD   64
 123         unsigned char sc_key[CARP_KEY_LEN];
 124         unsigned char sc_pad[CARP_HMAC_PAD];
 125         SHA1_CTX sc_sha1;
 126
 127         TAILQ_ENTRY(carp_softc) sc_list;        /* On the carp_if list. */
 128         LIST_ENTRY(carp_softc)  sc_next;        /* On the global list. */
 129 };
 130
 131 struct carp_if {
 132 #ifdef INET
 133         int     cif_naddrs;
 134 #endif
 135 #ifdef INET6
 136         int     cif_naddrs6;
 137 #endif
 138         TAILQ_HEAD(, carp_softc) cif_vrs;
 139 #ifdef INET
 140         struct ip_moptions       cif_imo;
 141 #endif
 142 #ifdef INET6
 143         struct ip6_moptions      cif_im6o;
 144 #endif
 145         struct ifnet    *cif_ifp;
 146         struct mtx      cif_mtx;
 147         uint32_t        cif_flags;
 148 #define CIF_PROMISC     0x00000001
 149 };
 150
 151 #define CARP_INET       0
 152 #define CARP_INET6      1
 153 static int proto_reg[] = {-1, -1};
 154
 155 /*
 156  * Brief design of carp(4).
 157  *
 158  * Any carp-capable ifnet may have a list of carp softcs hanging off
 159  * its ifp->if_carp pointer. Each softc represents one unique virtual
 160  * host id, or vhid. The softc has a back pointer to the ifnet. All
 161  * softcs are joined in a global list, which has quite limited use.
 162  *
 163  * Any interface address that takes part in CARP negotiation has a
 164  * pointer to the softc of its vhid, ifa->ifa_carp. That could be either
 165  * AF_INET or AF_INET6 address.
 166  *
 167  * Although, one can get the softc's backpointer to ifnet and traverse
 168  * through its ifp->if_addrhead queue to find all interface addresses
 169  * involved in CARP, we keep a growable array of ifaddr pointers. This
 170  * allows us to avoid grabbing the IF_ADDR_LOCK() in many traversals that
 171  * do calls into the network stack, thus avoiding LORs.
 172  *
 173  * Locking:
 174  *
 175  * Each softc has a lock sc_mtx. It is used to synchronise carp_input_c(),
 176  * callout-driven events and ioctl()s.
 177  *
 178  * To traverse the list of softcs on an ifnet we use CIF_LOCK() or carp_sx.
 179  * To traverse the global list we use the mutex carp_mtx.
 180  *
 181  * Known issues with locking:
 182  *
 183  * - Sending ad, we put the pointer to the softc in an mtag, and no reference
 184  *   counting is done on the softc.
 185  * - On module unload we may race (?) with packet processing thread
 186  *   dereferencing our function pointers.
 187  */
 188
 189 /* Accept incoming CARP packets. */
 190 VNET_DEFINE_STATIC(int, carp_allow) = 1;
 191 #define V_carp_allow    VNET(carp_allow)
 192
 193 /* Set DSCP in outgoing CARP packets. */
 194 VNET_DEFINE_STATIC(int, carp_dscp) = 56;
 195 #define V_carp_dscp     VNET(carp_dscp)
 196
 197 /* Preempt slower nodes. */
 198 VNET_DEFINE_STATIC(int, carp_preempt) = 0;
 199 #define V_carp_preempt  VNET(carp_preempt)
 200
 201 /* Log level. */
 202 VNET_DEFINE_STATIC(int, carp_log) = 1;
 203 #define V_carp_log      VNET(carp_log)
 204
 205 /* Global advskew demotion. */
 206 VNET_DEFINE_STATIC(int, carp_demotion) = 0;
 207 #define V_carp_demotion VNET(carp_demotion)
 208
 209 /* Send error demotion factor. */
 210 VNET_DEFINE_STATIC(int, carp_senderr_adj) = CARP_MAXSKEW;
 211 #define V_carp_senderr_adj      VNET(carp_senderr_adj)
 212
 213 /* Iface down demotion factor. */
 214 VNET_DEFINE_STATIC(int, carp_ifdown_adj) = CARP_MAXSKEW;
 215 #define V_carp_ifdown_adj       VNET(carp_ifdown_adj)
 216
 217 static int carp_allow_sysctl(SYSCTL_HANDLER_ARGS);
 218 static int carp_dscp_sysctl(SYSCTL_HANDLER_ARGS);
 219 static int carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS);
 220
 221 SYSCTL_NODE(_net_inet, IPPROTO_CARP, carp, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 222     "CARP");
 223 SYSCTL_PROC(_net_inet_carp, OID_AUTO, allow,
 224     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
 225     0, 0, carp_allow_sysctl, "I",
 226     "Accept incoming CARP packets");
 227 SYSCTL_PROC(_net_inet_carp, OID_AUTO, dscp,
 228     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
 229     0, 0, carp_dscp_sysctl, "I",
 230     "DSCP value for carp packets");
 231 SYSCTL_INT(_net_inet_carp, OID_AUTO, preempt, CTLFLAG_VNET | CTLFLAG_RW,
 232     &VNET_NAME(carp_preempt), 0, "High-priority backup preemption mode");
 233 SYSCTL_INT(_net_inet_carp, OID_AUTO, log, CTLFLAG_VNET | CTLFLAG_RW,
 234     &VNET_NAME(carp_log), 0, "CARP log level");
 235 SYSCTL_PROC(_net_inet_carp, OID_AUTO, demotion,
 236     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
 237     0, 0, carp_demote_adj_sysctl, "I",
 238     "Adjust demotion factor (skew of advskew)");
 239 SYSCTL_INT(_net_inet_carp, OID_AUTO, senderr_demotion_factor,
 240     CTLFLAG_VNET | CTLFLAG_RW,
 241     &VNET_NAME(carp_senderr_adj), 0, "Send error demotion factor adjustment");
 242 SYSCTL_INT(_net_inet_carp, OID_AUTO, ifdown_demotion_factor,
 243     CTLFLAG_VNET | CTLFLAG_RW,
 244     &VNET_NAME(carp_ifdown_adj), 0,
 245     "Interface down demotion factor adjustment");
 246
 247 VNET_PCPUSTAT_DEFINE(struct carpstats, carpstats);
 248 VNET_PCPUSTAT_SYSINIT(carpstats);
 249 VNET_PCPUSTAT_SYSUNINIT(carpstats);
 250
 251 #define CARPSTATS_ADD(name, val)        \
 252     counter_u64_add(VNET(carpstats)[offsetof(struct carpstats, name) / \
 253         sizeof(uint64_t)], (val))
 254 #define CARPSTATS_INC(name)             CARPSTATS_ADD(name, 1)
 255
 256 SYSCTL_VNET_PCPUSTAT(_net_inet_carp, OID_AUTO, stats, struct carpstats,
 257     carpstats, "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 258
 259 #define CARP_LOCK_INIT(sc)      mtx_init(&(sc)->sc_mtx, "carp_softc",   \
 260         NULL, MTX_DEF)
 261 #define CARP_LOCK_DESTROY(sc)   mtx_destroy(&(sc)->sc_mtx)
 262 #define CARP_LOCK_ASSERT(sc)    mtx_assert(&(sc)->sc_mtx, MA_OWNED)
 263 #define CARP_LOCK(sc)           mtx_lock(&(sc)->sc_mtx)
 264 #define CARP_UNLOCK(sc)         mtx_unlock(&(sc)->sc_mtx)
 265 #define CIF_LOCK_INIT(cif)      mtx_init(&(cif)->cif_mtx, "carp_if",   \
 266         NULL, MTX_DEF)
 267 #define CIF_LOCK_DESTROY(cif)   mtx_destroy(&(cif)->cif_mtx)
 268 #define CIF_LOCK_ASSERT(cif)    mtx_assert(&(cif)->cif_mtx, MA_OWNED)
 269 #define CIF_LOCK(cif)           mtx_lock(&(cif)->cif_mtx)
 270 #define CIF_UNLOCK(cif)         mtx_unlock(&(cif)->cif_mtx)
 271 #define CIF_FREE(cif)   do {                            \
 272                 CIF_LOCK(cif);                          \
 273                 if (TAILQ_EMPTY(&(cif)->cif_vrs))       \
 274                         carp_free_if(cif);              \
 275                 else                                    \
 276                         CIF_UNLOCK(cif);                \
 277 } while (0)
 278
 279 #define CARP_LOG(...)   do {                            \
 280         if (V_carp_log > 0)                             \
 281                 log(LOG_INFO, "carp: " __VA_ARGS__);    \
 282 } while (0)
 283
 284 #define CARP_DEBUG(...) do {                            \
 285         if (V_carp_log > 1)                             \
 286                 log(LOG_DEBUG, __VA_ARGS__);            \
 287 } while (0)
 288
 289 #define IFNET_FOREACH_IFA(ifp, ifa)                                     \
 290         CK_STAILQ_FOREACH((ifa), &(ifp)->if_addrhead, ifa_link) \
 291                 if ((ifa)->ifa_carp != NULL)
 292
 293 #define CARP_FOREACH_IFA(sc, ifa)                                       \
 294         CARP_LOCK_ASSERT(sc);                                           \
 295         for (int _i = 0;                                                \
 296                 _i < (sc)->sc_naddrs + (sc)->sc_naddrs6 &&              \
 297                 ((ifa) = sc->sc_ifas[_i]) != NULL;                      \
 298                 ++_i)
 299
 300 #define IFNET_FOREACH_CARP(ifp, sc)                                     \
 301         KASSERT(mtx_owned(&ifp->if_carp->cif_mtx) ||                    \
 302             sx_xlocked(&carp_sx), ("cif_vrs not locked"));              \
 303         TAILQ_FOREACH((sc), &(ifp)->if_carp->cif_vrs, sc_list)
 304
 305 #define DEMOTE_ADVSKEW(sc)                                      \
 306     (((sc)->sc_advskew + V_carp_demotion > CARP_MAXSKEW) ?      \
 307     CARP_MAXSKEW : ((sc)->sc_advskew + V_carp_demotion))
 308
 309 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 310 static struct carp_softc
 311                 *carp_alloc(struct ifnet *);
 312 static void     carp_destroy(struct carp_softc *);
 313 static struct carp_if
 314                 *carp_alloc_if(struct ifnet *);
 315 static void     carp_free_if(struct carp_if *);
 316 static void     carp_set_state(struct carp_softc *, int, const char* reason);
 317 static void     carp_sc_state(struct carp_softc *);
 318 static void     carp_setrun(struct carp_softc *, sa_family_t);
 319 static void     carp_master_down(void *);
 320 static void     carp_master_down_locked(struct carp_softc *,
 321                     const char* reason);
 322 static void     carp_send_ad(void *);
 323 static void     carp_send_ad_locked(struct carp_softc *);
 324 static void     carp_addroute(struct carp_softc *);
 325 static void     carp_ifa_addroute(struct ifaddr *);
 326 static void     carp_delroute(struct carp_softc *);
 327 static void     carp_ifa_delroute(struct ifaddr *);
 328 static void     carp_send_ad_all(void *, int);
 329 static void     carp_demote_adj(int, char *);
 330
 331 static LIST_HEAD(, carp_softc) carp_list;
 332 static struct mtx carp_mtx;
 333 static struct sx carp_sx;
 334 static struct task carp_sendall_task =
 335     TASK_INITIALIZER(0, carp_send_ad_all, NULL);
 336
 337 static void
 338 carp_hmac_prepare(struct carp_softc *sc)
 339 {
 340         uint8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 341         uint8_t vhid = sc->sc_vhid & 0xff;
 342         struct ifaddr *ifa;
 343         int i, found;
 344 #ifdef INET
 345         struct in_addr last, cur, in;
 346 #endif
 347 #ifdef INET6
 348         struct in6_addr last6, cur6, in6;
 349 #endif
 350
 351         CARP_LOCK_ASSERT(sc);
 352
 353         /* Compute ipad from key. */
 354         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 355         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 356         for (i = 0; i < sizeof(sc->sc_pad); i++)
 357                 sc->sc_pad[i] ^= 0x36;
 358
 359         /* Precompute first part of inner hash. */
 360         SHA1Init(&sc->sc_sha1);
 361         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 362         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 363         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 364         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 365 #ifdef INET
 366         cur.s_addr = 0;
 367         do {
 368                 found = 0;
 369                 last = cur;
 370                 cur.s_addr = 0xffffffff;
 371                 CARP_FOREACH_IFA(sc, ifa) {
 372                         in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 373                         if (ifa->ifa_addr->sa_family == AF_INET &&
 374                             ntohl(in.s_addr) > ntohl(last.s_addr) &&
 375                             ntohl(in.s_addr) < ntohl(cur.s_addr)) {
 376                                 cur.s_addr = in.s_addr;
 377                                 found++;
 378                         }
 379                 }
 380                 if (found)
 381                         SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur));
 382         } while (found);
 383 #endif /* INET */
 384 #ifdef INET6
 385         memset(&cur6, 0, sizeof(cur6));
 386         do {
 387                 found = 0;
 388                 last6 = cur6;
 389                 memset(&cur6, 0xff, sizeof(cur6));
 390                 CARP_FOREACH_IFA(sc, ifa) {
 391                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 392                         if (IN6_IS_SCOPE_EMBED(&in6))
 393                                 in6.s6_addr16[1] = 0;
 394                         if (ifa->ifa_addr->sa_family == AF_INET6 &&
 395                             memcmp(&in6, &last6, sizeof(in6)) > 0 &&
 396                             memcmp(&in6, &cur6, sizeof(in6)) < 0) {
 397                                 cur6 = in6;
 398                                 found++;
 399                         }
 400                 }
 401                 if (found)
 402                         SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6));
 403         } while (found);
 404 #endif /* INET6 */
 405
 406         /* convert ipad to opad */
 407         for (i = 0; i < sizeof(sc->sc_pad); i++)
 408                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 409 }
 410
 411 static void
 412 carp_hmac_generate(struct carp_softc *sc, uint32_t counter[2],
 413     unsigned char md[20])
 414 {
 415         SHA1_CTX sha1ctx;
 416
 417         CARP_LOCK_ASSERT(sc);
 418
 419         /* fetch first half of inner hash */
 420         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 421
 422         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 423         SHA1Final(md, &sha1ctx);
 424
 425         /* outer hash */
 426         SHA1Init(&sha1ctx);
 427         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 428         SHA1Update(&sha1ctx, md, 20);
 429         SHA1Final(md, &sha1ctx);
 430 }
 431
 432 static int
 433 carp_hmac_verify(struct carp_softc *sc, uint32_t counter[2],
 434     unsigned char md[20])
 435 {
 436         unsigned char md2[20];
 437
 438         CARP_LOCK_ASSERT(sc);
 439
 440         carp_hmac_generate(sc, counter, md2);
 441
 442         return (bcmp(md, md2, sizeof(md2)));
 443 }
 444
 445 /*
 446  * process input packet.
 447  * we have rearranged checks order compared to the rfc,
 448  * but it seems more efficient this way or not possible otherwise.
 449  */
 450 #ifdef INET
 451 int
 452 carp_input(struct mbuf **mp, int *offp, int proto)
 453 {
 454         struct mbuf *m = *mp;
 455         struct ip *ip = mtod(m, struct ip *);
 456         struct carp_header *ch;
 457         int iplen, len;
 458
 459         iplen = *offp;
 460         *mp = NULL;
 461
 462         CARPSTATS_INC(carps_ipackets);
 463
 464         if (!V_carp_allow) {
 465                 m_freem(m);
 466                 return (IPPROTO_DONE);
 467         }
 468
 469         /* verify that the IP TTL is 255.  */
 470         if (ip->ip_ttl != CARP_DFLTTL) {
 471                 CARPSTATS_INC(carps_badttl);
 472                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 473                     ip->ip_ttl,
 474                     m->m_pkthdr.rcvif->if_xname);
 475                 m_freem(m);
 476                 return (IPPROTO_DONE);
 477         }
 478
 479         iplen = ip->ip_hl << 2;
 480
 481         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 482                 CARPSTATS_INC(carps_badlen);
 483                 CARP_DEBUG("%s: received len %zd < sizeof(struct carp_header) "
 484                     "on %s\n", __func__, m->m_len - sizeof(struct ip),
 485                     m->m_pkthdr.rcvif->if_xname);
 486                 m_freem(m);
 487                 return (IPPROTO_DONE);
 488         }
 489
 490         if (iplen + sizeof(*ch) < m->m_len) {
 491                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 492                         CARPSTATS_INC(carps_hdrops);
 493                         CARP_DEBUG("%s: pullup failed\n", __func__);
 494                         return (IPPROTO_DONE);
 495                 }
 496                 ip = mtod(m, struct ip *);
 497         }
 498         ch = (struct carp_header *)((char *)ip + iplen);
 499
 500         /*
 501          * verify that the received packet length is
 502          * equal to the CARP header
 503          */
 504         len = iplen + sizeof(*ch);
 505         if (len > m->m_pkthdr.len) {
 506                 CARPSTATS_INC(carps_badlen);
 507                 CARP_DEBUG("%s: packet too short %d on %s\n", __func__,
 508                     m->m_pkthdr.len,
 509                     m->m_pkthdr.rcvif->if_xname);
 510                 m_freem(m);
 511                 return (IPPROTO_DONE);
 512         }
 513
 514         if ((m = m_pullup(m, len)) == NULL) {
 515                 CARPSTATS_INC(carps_hdrops);
 516                 return (IPPROTO_DONE);
 517         }
 518         ip = mtod(m, struct ip *);
 519         ch = (struct carp_header *)((char *)ip + iplen);
 520
 521         /* verify the CARP checksum */
 522         m->m_data += iplen;
 523         if (in_cksum(m, len - iplen)) {
 524                 CARPSTATS_INC(carps_badsum);
 525                 CARP_DEBUG("%s: checksum failed on %s\n", __func__,
 526                     m->m_pkthdr.rcvif->if_xname);
 527                 m_freem(m);
 528                 return (IPPROTO_DONE);
 529         }
 530         m->m_data -= iplen;
 531
 532         carp_input_c(m, ch, AF_INET);
 533         return (IPPROTO_DONE);
 534 }
 535 #endif
 536
 537 #ifdef INET6
 538 int
 539 carp6_input(struct mbuf **mp, int *offp, int proto)
 540 {
 541         struct mbuf *m = *mp;
 542         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 543         struct carp_header *ch;
 544         u_int len;
 545
 546         CARPSTATS_INC(carps_ipackets6);
 547
 548         if (!V_carp_allow) {
 549                 m_freem(m);
 550                 return (IPPROTO_DONE);
 551         }
 552
 553         /* check if received on a valid carp interface */
 554         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 555                 CARPSTATS_INC(carps_badif);
 556                 CARP_DEBUG("%s: packet received on non-carp interface: %s\n",
 557                     __func__, m->m_pkthdr.rcvif->if_xname);
 558                 m_freem(m);
 559                 return (IPPROTO_DONE);
 560         }
 561
 562         /* verify that the IP TTL is 255 */
 563         if (ip6->ip6_hlim != CARP_DFLTTL) {
 564                 CARPSTATS_INC(carps_badttl);
 565                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 566                     ip6->ip6_hlim, m->m_pkthdr.rcvif->if_xname);
 567                 m_freem(m);
 568                 return (IPPROTO_DONE);
 569         }
 570
 571         /* verify that we have a complete carp packet */
 572         if (m->m_len < *offp + sizeof(*ch)) {
 573                 len = m->m_len;
 574                 m = m_pullup(m, *offp + sizeof(*ch));
 575                 if (m == NULL) {
 576                         CARPSTATS_INC(carps_badlen);
 577                         CARP_DEBUG("%s: packet size %u too small\n", __func__, len);
 578                         return (IPPROTO_DONE);
 579                 }
 580         }
 581         ch = (struct carp_header *)(mtod(m, char *) + *offp);
 582
 583
 584         /* verify the CARP checksum */
 585         m->m_data += *offp;
 586         if (in_cksum(m, sizeof(*ch))) {
 587                 CARPSTATS_INC(carps_badsum);
 588                 CARP_DEBUG("%s: checksum failed, on %s\n", __func__,
 589                     m->m_pkthdr.rcvif->if_xname);
 590                 m_freem(m);
 591                 return (IPPROTO_DONE);
 592         }
 593         m->m_data -= *offp;
 594
 595         carp_input_c(m, ch, AF_INET6);
 596         return (IPPROTO_DONE);
 597 }
 598 #endif /* INET6 */
 599
 600 /*
 601  * This routine should not be necessary at all, but some switches
 602  * (VMWare ESX vswitches) can echo our own packets back at us,
 603  * and we must ignore them or they will cause us to drop out of
 604  * MASTER mode.
 605  *
 606  * We cannot catch all cases of network loops.  Instead, what we
 607  * do here is catch any packet that arrives with a carp header
 608  * with a VHID of 0, that comes from an address that is our own.
 609  * These packets are by definition "from us" (even if they are from
 610  * a misconfigured host that is pretending to be us).
 611  *
 612  * The VHID test is outside this mini-function.
 613  */
 614 static int
 615 carp_source_is_self(struct mbuf *m, struct ifaddr *ifa, sa_family_t af)
 616 {
 617 #ifdef INET
 618         struct ip *ip4;
 619         struct in_addr in4;
 620 #endif
 621 #ifdef INET6
 622         struct ip6_hdr *ip6;
 623         struct in6_addr in6;
 624 #endif
 625
 626         switch (af) {
 627 #ifdef INET
 628         case AF_INET:
 629                 ip4 = mtod(m, struct ip *);
 630                 in4 = ifatoia(ifa)->ia_addr.sin_addr;
 631                 return (in4.s_addr == ip4->ip_src.s_addr);
 632 #endif
 633 #ifdef INET6
 634         case AF_INET6:
 635                 ip6 = mtod(m, struct ip6_hdr *);
 636                 in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 637                 return (memcmp(&in6, &ip6->ip6_src, sizeof(in6)) == 0);
 638 #endif
 639         default:
 640                 break;
 641         }
 642         return (0);
 643 }
 644
 645 static void
 646 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 647 {
 648         struct ifnet *ifp = m->m_pkthdr.rcvif;
 649         struct ifaddr *ifa, *match;
 650         struct carp_softc *sc;
 651         uint64_t tmp_counter;
 652         struct timeval sc_tv, ch_tv;
 653         int error;
 654
 655         NET_EPOCH_ASSERT();
 656
 657         /*
 658          * Verify that the VHID is valid on the receiving interface.
 659          *
 660          * There should be just one match.  If there are none
 661          * the VHID is not valid and we drop the packet.  If
 662          * there are multiple VHID matches, take just the first
 663          * one, for compatibility with previous code.  While we're
 664          * scanning, check for obvious loops in the network topology
 665          * (these should never happen, and as noted above, we may
 666          * miss real loops; this is just a double-check).
 667          */
 668         error = 0;
 669         match = NULL;
 670         IFNET_FOREACH_IFA(ifp, ifa) {
 671                 if (match == NULL && ifa->ifa_carp != NULL &&
 672                     ifa->ifa_addr->sa_family == af &&
 673                     ifa->ifa_carp->sc_vhid == ch->carp_vhid)
 674                         match = ifa;
 675                 if (ch->carp_vhid == 0 && carp_source_is_self(m, ifa, af))
 676                         error = ELOOP;
 677         }
 678         ifa = error ? NULL : match;
 679         if (ifa != NULL)
 680                 ifa_ref(ifa);
 681
 682         if (ifa == NULL) {
 683                 if (error == ELOOP) {
 684                         CARP_DEBUG("dropping looped packet on interface %s\n",
 685                             ifp->if_xname);
 686                         CARPSTATS_INC(carps_badif);     /* ??? */
 687                 } else {
 688                         CARPSTATS_INC(carps_badvhid);
 689                 }
 690                 m_freem(m);
 691                 return;
 692         }
 693
 694         /* verify the CARP version. */
 695         if (ch->carp_version != CARP_VERSION) {
 696                 CARPSTATS_INC(carps_badver);
 697                 CARP_DEBUG("%s: invalid version %d\n", ifp->if_xname,
 698                     ch->carp_version);
 699                 ifa_free(ifa);
 700                 m_freem(m);
 701                 return;
 702         }
 703
 704         sc = ifa->ifa_carp;
 705         CARP_LOCK(sc);
 706         ifa_free(ifa);
 707
 708         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 709                 CARPSTATS_INC(carps_badauth);
 710                 CARP_DEBUG("%s: incorrect hash for VHID %u@%s\n", __func__,
 711                     sc->sc_vhid, ifp->if_xname);
 712                 goto out;
 713         }
 714
 715         tmp_counter = ntohl(ch->carp_counter[0]);
 716         tmp_counter = tmp_counter<<32;
 717         tmp_counter += ntohl(ch->carp_counter[1]);
 718
 719         /* XXX Replay protection goes here */
 720
 721         sc->sc_init_counter = 0;
 722         sc->sc_counter = tmp_counter;
 723
 724         sc_tv.tv_sec = sc->sc_advbase;
 725         sc_tv.tv_usec = DEMOTE_ADVSKEW(sc) * 1000000 / 256;
 726         ch_tv.tv_sec = ch->carp_advbase;
 727         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 728
 729         switch (sc->sc_state) {
 730         case INIT:
 731                 break;
 732         case MASTER:
 733                 /*
 734                  * If we receive an advertisement from a master who's going to
 735                  * be more frequent than us, go into BACKUP state.
 736                  */
 737                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 738                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 739                         callout_stop(&sc->sc_ad_tmo);
 740                         carp_set_state(sc, BACKUP,
 741                             "more frequent advertisement received");
 742                         carp_setrun(sc, 0);
 743                         carp_delroute(sc);
 744                 }
 745                 break;
 746         case BACKUP:
 747                 /*
 748                  * If we're pre-empting masters who advertise slower than us,
 749                  * and this one claims to be slower, treat him as down.
 750                  */
 751                 if (V_carp_preempt && timevalcmp(&sc_tv, &ch_tv, <)) {
 752                         carp_master_down_locked(sc,
 753                             "preempting a slower master");
 754                         break;
 755                 }
 756
 757                 /*
 758                  *  If the master is going to advertise at such a low frequency
 759                  *  that he's guaranteed to time out, we'd might as well just
 760                  *  treat him as timed out now.
 761                  */
 762                 sc_tv.tv_sec = sc->sc_advbase * 3;
 763                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 764                         carp_master_down_locked(sc, "master will time out");
 765                         break;
 766                 }
 767
 768                 /*
 769                  * Otherwise, we reset the counter and wait for the next
 770                  * advertisement.
 771                  */
 772                 carp_setrun(sc, af);
 773                 break;
 774         }
 775
 776 out:
 777         CARP_UNLOCK(sc);
 778         m_freem(m);
 779 }
 780
 781 static int
 782 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 783 {
 784         struct m_tag *mtag;
 785
 786         if (sc->sc_init_counter) {
 787                 /* this could also be seconds since unix epoch */
 788                 sc->sc_counter = arc4random();
 789                 sc->sc_counter = sc->sc_counter << 32;
 790                 sc->sc_counter += arc4random();
 791         } else
 792                 sc->sc_counter++;
 793
 794         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 795         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 796
 797         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 798
 799         /* Tag packet for carp_output */
 800         if ((mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct carp_softc *),
 801             M_NOWAIT)) == NULL) {
 802                 m_freem(m);
 803                 CARPSTATS_INC(carps_onomem);
 804                 return (ENOMEM);
 805         }
 806         bcopy(&sc, mtag + 1, sizeof(sc));
 807         m_tag_prepend(m, mtag);
 808
 809         return (0);
 810 }
 811
 812 /*
 813  * To avoid LORs and possible recursions this function shouldn't
 814  * be called directly, but scheduled via taskqueue.
 815  */
 816 static void
 817 carp_send_ad_all(void *ctx __unused, int pending __unused)
 818 {
 819         struct carp_softc *sc;
 820         struct epoch_tracker et;
 821
 822         NET_EPOCH_ENTER(et);
 823         mtx_lock(&carp_mtx);
 824         LIST_FOREACH(sc, &carp_list, sc_next)
 825                 if (sc->sc_state == MASTER) {
 826                         CARP_LOCK(sc);
 827                         CURVNET_SET(sc->sc_carpdev->if_vnet);
 828                         carp_send_ad_locked(sc);
 829                         CURVNET_RESTORE();
 830                         CARP_UNLOCK(sc);
 831                 }
 832         mtx_unlock(&carp_mtx);
 833         NET_EPOCH_EXIT(et);
 834 }
 835
 836 /* Send a periodic advertisement, executed in callout context. */
 837 static void
 838 carp_send_ad(void *v)
 839 {
 840         struct carp_softc *sc = v;
 841         struct epoch_tracker et;
 842
 843         NET_EPOCH_ENTER(et);
 844         CARP_LOCK_ASSERT(sc);
 845         CURVNET_SET(sc->sc_carpdev->if_vnet);
 846         carp_send_ad_locked(sc);
 847         CURVNET_RESTORE();
 848         CARP_UNLOCK(sc);
 849         NET_EPOCH_EXIT(et);
 850 }
 851
 852 static void
 853 carp_send_ad_error(struct carp_softc *sc, int error)
 854 {
 855
 856         if (error) {
 857                 if (sc->sc_sendad_errors < INT_MAX)
 858                         sc->sc_sendad_errors++;
 859                 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 860                         static const char fmt[] = "send error %d on %s";
 861                         char msg[sizeof(fmt) + IFNAMSIZ];
 862
 863                         sprintf(msg, fmt, error, sc->sc_carpdev->if_xname);
 864                         carp_demote_adj(V_carp_senderr_adj, msg);
 865                 }
 866                 sc->sc_sendad_success = 0;
 867         } else {
 868                 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS &&
 869                     ++sc->sc_sendad_success >= CARP_SENDAD_MIN_SUCCESS) {
 870                         static const char fmt[] = "send ok on %s";
 871                         char msg[sizeof(fmt) + IFNAMSIZ];
 872
 873                         sprintf(msg, fmt, sc->sc_carpdev->if_xname);
 874                         carp_demote_adj(-V_carp_senderr_adj, msg);
 875                         sc->sc_sendad_errors = 0;
 876                 } else
 877                         sc->sc_sendad_errors = 0;
 878         }
 879 }
 880
 881 /*
 882  * Pick the best ifaddr on the given ifp for sending CARP
 883  * advertisements.
 884  *
 885  * "Best" here is defined by ifa_preferred().  This function is much
 886  * much like ifaof_ifpforaddr() except that we just use ifa_preferred().
 887  *
 888  * (This could be simplified to return the actual address, except that
 889  * it has a different format in AF_INET and AF_INET6.)
 890  */
 891 static struct ifaddr *
 892 carp_best_ifa(int af, struct ifnet *ifp)
 893 {
 894         struct ifaddr *ifa, *best;
 895
 896         NET_EPOCH_ASSERT();
 897
 898         if (af >= AF_MAX)
 899                 return (NULL);
 900         best = NULL;
 901         CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
 902                 if (ifa->ifa_addr->sa_family == af &&
 903                     (best == NULL || ifa_preferred(best, ifa)))
 904                         best = ifa;
 905         }
 906         if (best != NULL)
 907                 ifa_ref(best);
 908         return (best);
 909 }
 910
 911 static void
 912 carp_send_ad_locked(struct carp_softc *sc)
 913 {
 914         struct carp_header ch;
 915         struct timeval tv;
 916         struct ifaddr *ifa;
 917         struct carp_header *ch_ptr;
 918         struct mbuf *m;
 919         int len, advskew;
 920
 921         NET_EPOCH_ASSERT();
 922         CARP_LOCK_ASSERT(sc);
 923
 924         advskew = DEMOTE_ADVSKEW(sc);
 925         tv.tv_sec = sc->sc_advbase;
 926         tv.tv_usec = advskew * 1000000 / 256;
 927
 928         ch.carp_version = CARP_VERSION;
 929         ch.carp_type = CARP_ADVERTISEMENT;
 930         ch.carp_vhid = sc->sc_vhid;
 931         ch.carp_advbase = sc->sc_advbase;
 932         ch.carp_advskew = advskew;
 933         ch.carp_authlen = 7;    /* XXX DEFINE */
 934         ch.carp_pad1 = 0;       /* must be zero */
 935         ch.carp_cksum = 0;
 936
 937         /* XXXGL: OpenBSD picks first ifaddr with needed family. */
 938
 939 #ifdef INET
 940         if (sc->sc_naddrs) {
 941                 struct ip *ip;
 942
 943                 m = m_gethdr(M_NOWAIT, MT_DATA);
 944                 if (m == NULL) {
 945                         CARPSTATS_INC(carps_onomem);
 946                         goto resched;
 947                 }
 948                 len = sizeof(*ip) + sizeof(ch);
 949                 m->m_pkthdr.len = len;
 950                 m->m_pkthdr.rcvif = NULL;
 951                 m->m_len = len;
 952                 M_ALIGN(m, m->m_len);
 953                 m->m_flags |= M_MCAST;
 954                 ip = mtod(m, struct ip *);
 955                 ip->ip_v = IPVERSION;
 956                 ip->ip_hl = sizeof(*ip) >> 2;
 957                 ip->ip_tos = V_carp_dscp << IPTOS_DSCP_OFFSET;
 958                 ip->ip_len = htons(len);
 959                 ip->ip_off = htons(IP_DF);
 960                 ip->ip_ttl = CARP_DFLTTL;
 961                 ip->ip_p = IPPROTO_CARP;
 962                 ip->ip_sum = 0;
 963                 ip_fillid(ip);
 964
 965                 ifa = carp_best_ifa(AF_INET, sc->sc_carpdev);
 966                 if (ifa != NULL) {
 967                         ip->ip_src.s_addr =
 968                             ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 969                         ifa_free(ifa);
 970                 } else
 971                         ip->ip_src.s_addr = 0;
 972                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 973
 974                 ch_ptr = (struct carp_header *)(&ip[1]);
 975                 bcopy(&ch, ch_ptr, sizeof(ch));
 976                 if (carp_prepare_ad(m, sc, ch_ptr))
 977                         goto resched;
 978
 979                 m->m_data += sizeof(*ip);
 980                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip));
 981                 m->m_data -= sizeof(*ip);
 982
 983                 CARPSTATS_INC(carps_opackets);
 984
 985                 carp_send_ad_error(sc, ip_output(m, NULL, NULL, IP_RAWOUTPUT,
 986                     &sc->sc_carpdev->if_carp->cif_imo, NULL));
 987         }
 988 #endif /* INET */
 989 #ifdef INET6
 990         if (sc->sc_naddrs6) {
 991                 struct ip6_hdr *ip6;
 992
 993                 m = m_gethdr(M_NOWAIT, MT_DATA);
 994                 if (m == NULL) {
 995                         CARPSTATS_INC(carps_onomem);
 996                         goto resched;
 997                 }
 998                 len = sizeof(*ip6) + sizeof(ch);
 999                 m->m_pkthdr.len = len;
1000                 m->m_pkthdr.rcvif = NULL;
1001                 m->m_len = len;
1002                 M_ALIGN(m, m->m_len);
1003                 m->m_flags |= M_MCAST;
1004                 ip6 = mtod(m, struct ip6_hdr *);
1005                 bzero(ip6, sizeof(*ip6));
1006                 ip6->ip6_vfc |= IPV6_VERSION;
1007                 /* Traffic class isn't defined in ip6 struct instead
1008                  * it gets offset into flowid field */
1009                 ip6->ip6_flow |= htonl(V_carp_dscp << (IPV6_FLOWLABEL_LEN +
1010                     IPTOS_DSCP_OFFSET));
1011                 ip6->ip6_hlim = CARP_DFLTTL;
1012                 ip6->ip6_nxt = IPPROTO_CARP;
1013
1014                 /* set the source address */
1015                 ifa = carp_best_ifa(AF_INET6, sc->sc_carpdev);
1016                 if (ifa != NULL) {
1017                         bcopy(IFA_IN6(ifa), &ip6->ip6_src,
1018                             sizeof(struct in6_addr));
1019                         ifa_free(ifa);
1020                 } else
1021                         /* This should never happen with IPv6. */
1022                         bzero(&ip6->ip6_src, sizeof(struct in6_addr));
1023
1024                 /* Set the multicast destination. */
1025                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
1026                 ip6->ip6_dst.s6_addr8[15] = 0x12;
1027                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
1028                         m_freem(m);
1029                         CARP_DEBUG("%s: in6_setscope failed\n", __func__);
1030                         goto resched;
1031                 }
1032
1033                 ch_ptr = (struct carp_header *)(&ip6[1]);
1034                 bcopy(&ch, ch_ptr, sizeof(ch));
1035                 if (carp_prepare_ad(m, sc, ch_ptr))
1036                         goto resched;
1037
1038                 m->m_data += sizeof(*ip6);
1039                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip6));
1040                 m->m_data -= sizeof(*ip6);
1041
1042                 CARPSTATS_INC(carps_opackets6);
1043
1044                 carp_send_ad_error(sc, ip6_output(m, NULL, NULL, 0,
1045                     &sc->sc_carpdev->if_carp->cif_im6o, NULL, NULL));
1046         }
1047 #endif /* INET6 */
1048
1049 resched:
1050         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), carp_send_ad, sc);
1051 }
1052
1053 static void
1054 carp_addroute(struct carp_softc *sc)
1055 {
1056         struct ifaddr *ifa;
1057
1058         CARP_FOREACH_IFA(sc, ifa)
1059                 carp_ifa_addroute(ifa);
1060 }
1061
1062 static void
1063 carp_ifa_addroute(struct ifaddr *ifa)
1064 {
1065
1066         switch (ifa->ifa_addr->sa_family) {
1067 #ifdef INET
1068         case AF_INET:
1069                 in_addprefix(ifatoia(ifa), RTF_UP);
1070                 ifa_add_loopback_route(ifa,
1071                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
1072                 break;
1073 #endif
1074 #ifdef INET6
1075         case AF_INET6:
1076                 ifa_add_loopback_route(ifa,
1077                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
1078                 nd6_add_ifa_lle(ifatoia6(ifa));
1079                 break;
1080 #endif
1081         }
1082 }
1083
1084 static void
1085 carp_delroute(struct carp_softc *sc)
1086 {
1087         struct ifaddr *ifa;
1088
1089         CARP_FOREACH_IFA(sc, ifa)
1090                 carp_ifa_delroute(ifa);
1091 }
1092
1093 static void
1094 carp_ifa_delroute(struct ifaddr *ifa)
1095 {
1096
1097         switch (ifa->ifa_addr->sa_family) {
1098 #ifdef INET
1099         case AF_INET:
1100                 ifa_del_loopback_route(ifa,
1101                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
1102                 in_scrubprefix(ifatoia(ifa), LLE_STATIC);
1103                 break;
1104 #endif
1105 #ifdef INET6
1106         case AF_INET6:
1107                 ifa_del_loopback_route(ifa,
1108                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
1109                 nd6_rem_ifa_lle(ifatoia6(ifa), 1);
1110                 break;
1111 #endif
1112         }
1113 }
1114
1115 int
1116 carp_master(struct ifaddr *ifa)
1117 {
1118         struct carp_softc *sc = ifa->ifa_carp;
1119
1120         return (sc->sc_state == MASTER);
1121 }
1122
1123 #ifdef INET
1124 /*
1125  * Broadcast a gratuitous ARP request containing
1126  * the virtual router MAC address for each IP address
1127  * associated with the virtual router.
1128  */
1129 static void
1130 carp_send_arp(struct carp_softc *sc)
1131 {
1132         struct ifaddr *ifa;
1133         struct in_addr addr;
1134
1135         NET_EPOCH_ASSERT();
1136
1137         CARP_FOREACH_IFA(sc, ifa) {
1138                 if (ifa->ifa_addr->sa_family != AF_INET)
1139                         continue;
1140                 addr = ((struct sockaddr_in *)ifa->ifa_addr)->sin_addr;
1141                 arp_announce_ifaddr(sc->sc_carpdev, addr, LLADDR(&sc->sc_addr));
1142         }
1143 }
1144
1145 int
1146 carp_iamatch(struct ifaddr *ifa, uint8_t **enaddr)
1147 {
1148         struct carp_softc *sc = ifa->ifa_carp;
1149
1150         if (sc->sc_state == MASTER) {
1151                 *enaddr = LLADDR(&sc->sc_addr);
1152                 return (1);
1153         }
1154
1155         return (0);
1156 }
1157 #endif
1158
1159 #ifdef INET6
1160 static void
1161 carp_send_na(struct carp_softc *sc)
1162 {
1163         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1164         struct ifaddr *ifa;
1165         struct in6_addr *in6;
1166
1167         CARP_FOREACH_IFA(sc, ifa) {
1168                 if (ifa->ifa_addr->sa_family != AF_INET6)
1169                         continue;
1170
1171                 in6 = IFA_IN6(ifa);
1172                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1173                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1174                 DELAY(1000);    /* XXX */
1175         }
1176 }
1177
1178 /*
1179  * Returns ifa in case it's a carp address and it is MASTER, or if the address
1180  * matches and is not a carp address.  Returns NULL otherwise.
1181  */
1182 struct ifaddr *
1183 carp_iamatch6(struct ifnet *ifp, struct in6_addr *taddr)
1184 {
1185         struct ifaddr *ifa;
1186
1187         NET_EPOCH_ASSERT();
1188
1189         ifa = NULL;
1190         CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
1191                 if (ifa->ifa_addr->sa_family != AF_INET6)
1192                         continue;
1193                 if (!IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa)))
1194                         continue;
1195                 if (ifa->ifa_carp && ifa->ifa_carp->sc_state != MASTER)
1196                         ifa = NULL;
1197                 else
1198                         ifa_ref(ifa);
1199                 break;
1200         }
1201
1202         return (ifa);
1203 }
1204
1205 char *
1206 carp_macmatch6(struct ifnet *ifp, struct mbuf *m, const struct in6_addr *taddr)
1207 {
1208         struct ifaddr *ifa;
1209
1210         NET_EPOCH_ASSERT();
1211
1212         IFNET_FOREACH_IFA(ifp, ifa)
1213                 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1214                     IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1215                         struct carp_softc *sc = ifa->ifa_carp;
1216                         struct m_tag *mtag;
1217
1218                         mtag = m_tag_get(PACKET_TAG_CARP,
1219                             sizeof(struct carp_softc *), M_NOWAIT);
1220                         if (mtag == NULL)
1221                                 /* Better a bit than nothing. */
1222                                 return (LLADDR(&sc->sc_addr));
1223
1224                         bcopy(&sc, mtag + 1, sizeof(sc));
1225                         m_tag_prepend(m, mtag);
1226
1227                         return (LLADDR(&sc->sc_addr));
1228                 }
1229
1230         return (NULL);
1231 }
1232 #endif /* INET6 */
1233
1234 int
1235 carp_forus(struct ifnet *ifp, u_char *dhost)
1236 {
1237         struct carp_softc *sc;
1238         uint8_t *ena = dhost;
1239
1240         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1241                 return (0);
1242
1243         CIF_LOCK(ifp->if_carp);
1244         IFNET_FOREACH_CARP(ifp, sc) {
1245                 /*
1246                  * CARP_LOCK() is not here, since would protect nothing, but
1247                  * cause deadlock with if_bridge, calling this under its lock.
1248                  */
1249                 if (sc->sc_state == MASTER && !bcmp(dhost, LLADDR(&sc->sc_addr),
1250                     ETHER_ADDR_LEN)) {
1251                         CIF_UNLOCK(ifp->if_carp);
1252                         return (1);
1253                 }
1254         }
1255         CIF_UNLOCK(ifp->if_carp);
1256
1257         return (0);
1258 }
1259
1260 /* Master down timeout event, executed in callout context. */
1261 static void
1262 carp_master_down(void *v)
1263 {
1264         struct carp_softc *sc = v;
1265         struct epoch_tracker et;
1266
1267         NET_EPOCH_ENTER(et);
1268         CARP_LOCK_ASSERT(sc);
1269
1270         CURVNET_SET(sc->sc_carpdev->if_vnet);
1271         if (sc->sc_state == BACKUP) {
1272                 carp_master_down_locked(sc, "master timed out");
1273         }
1274         CURVNET_RESTORE();
1275
1276         CARP_UNLOCK(sc);
1277         NET_EPOCH_EXIT(et);
1278 }
1279
1280 static void
1281 carp_master_down_locked(struct carp_softc *sc, const char *reason)
1282 {
1283
1284         NET_EPOCH_ASSERT();
1285         CARP_LOCK_ASSERT(sc);
1286
1287         switch (sc->sc_state) {
1288         case BACKUP:
1289                 carp_set_state(sc, MASTER, reason);
1290                 carp_send_ad_locked(sc);
1291 #ifdef INET
1292                 carp_send_arp(sc);
1293 #endif
1294 #ifdef INET6
1295                 carp_send_na(sc);
1296 #endif
1297                 carp_setrun(sc, 0);
1298                 carp_addroute(sc);
1299                 break;
1300         case INIT:
1301         case MASTER:
1302 #ifdef INVARIANTS
1303                 panic("carp: VHID %u@%s: master_down event in %s state\n",
1304                     sc->sc_vhid,
1305                     sc->sc_carpdev->if_xname,
1306                     sc->sc_state ? "MASTER" : "INIT");
1307 #endif
1308                 break;
1309         }
1310 }
1311
1312 /*
1313  * When in backup state, af indicates whether to reset the master down timer
1314  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1315  */
1316 static void
1317 carp_setrun(struct carp_softc *sc, sa_family_t af)
1318 {
1319         struct timeval tv;
1320
1321         CARP_LOCK_ASSERT(sc);
1322
1323         if ((sc->sc_carpdev->if_flags & IFF_UP) == 0 ||
1324             sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1325             (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) ||
1326             !V_carp_allow)
1327                 return;
1328
1329         switch (sc->sc_state) {
1330         case INIT:
1331                 carp_set_state(sc, BACKUP, "initialization complete");
1332                 carp_setrun(sc, 0);
1333                 break;
1334         case BACKUP:
1335                 callout_stop(&sc->sc_ad_tmo);
1336                 tv.tv_sec = 3 * sc->sc_advbase;
1337                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1338                 switch (af) {
1339 #ifdef INET
1340                 case AF_INET:
1341                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1342                             carp_master_down, sc);
1343                         break;
1344 #endif
1345 #ifdef INET6
1346                 case AF_INET6:
1347                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1348                             carp_master_down, sc);
1349                         break;
1350 #endif
1351                 default:
1352 #ifdef INET
1353                         if (sc->sc_naddrs)
1354                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1355                                     carp_master_down, sc);
1356 #endif
1357 #ifdef INET6
1358                         if (sc->sc_naddrs6)
1359                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1360                                     carp_master_down, sc);
1361 #endif
1362                         break;
1363                 }
1364                 break;
1365         case MASTER:
1366                 tv.tv_sec = sc->sc_advbase;
1367                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1368                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1369                     carp_send_ad, sc);
1370                 break;
1371         }
1372 }
1373
1374 /*
1375  * Setup multicast structures.
1376  */
1377 static int
1378 carp_multicast_setup(struct carp_if *cif, sa_family_t sa)
1379 {
1380         struct ifnet *ifp = cif->cif_ifp;
1381         int error = 0;
1382
1383         switch (sa) {
1384 #ifdef INET
1385         case AF_INET:
1386             {
1387                 struct ip_moptions *imo = &cif->cif_imo;
1388                 struct in_mfilter *imf;
1389                 struct in_addr addr;
1390
1391                 if (ip_mfilter_first(&imo->imo_head) != NULL)
1392                         return (0);
1393
1394                 imf = ip_mfilter_alloc(M_WAITOK, 0, 0);
1395                 ip_mfilter_init(&imo->imo_head);
1396                 imo->imo_multicast_vif = -1;
1397
1398                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1399                 if ((error = in_joingroup(ifp, &addr, NULL,
1400                     &imf->imf_inm)) != 0) {
1401                         ip_mfilter_free(imf);
1402                         break;
1403                 }
1404
1405                 ip_mfilter_insert(&imo->imo_head, imf);
1406                 imo->imo_multicast_ifp = ifp;
1407                 imo->imo_multicast_ttl = CARP_DFLTTL;
1408                 imo->imo_multicast_loop = 0;
1409                 break;
1410            }
1411 #endif
1412 #ifdef INET6
1413         case AF_INET6:
1414             {
1415                 struct ip6_moptions *im6o = &cif->cif_im6o;
1416                 struct in6_mfilter *im6f[2];
1417                 struct in6_addr in6;
1418
1419                 if (ip6_mfilter_first(&im6o->im6o_head))
1420                         return (0);
1421
1422                 im6f[0] = ip6_mfilter_alloc(M_WAITOK, 0, 0);
1423                 im6f[1] = ip6_mfilter_alloc(M_WAITOK, 0, 0);
1424
1425                 ip6_mfilter_init(&im6o->im6o_head);
1426                 im6o->im6o_multicast_hlim = CARP_DFLTTL;
1427                 im6o->im6o_multicast_ifp = ifp;
1428
1429                 /* Join IPv6 CARP multicast group. */
1430                 bzero(&in6, sizeof(in6));
1431                 in6.s6_addr16[0] = htons(0xff02);
1432                 in6.s6_addr8[15] = 0x12;
1433                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1434                         ip6_mfilter_free(im6f[0]);
1435                         ip6_mfilter_free(im6f[1]);
1436                         break;
1437                 }
1438                 if ((error = in6_joingroup(ifp, &in6, NULL, &im6f[0]->im6f_in6m, 0)) != 0) {
1439                         ip6_mfilter_free(im6f[0]);
1440                         ip6_mfilter_free(im6f[1]);
1441                         break;
1442                 }
1443
1444                 /* Join solicited multicast address. */
1445                 bzero(&in6, sizeof(in6));
1446                 in6.s6_addr16[0] = htons(0xff02);
1447                 in6.s6_addr32[1] = 0;
1448                 in6.s6_addr32[2] = htonl(1);
1449                 in6.s6_addr32[3] = 0;
1450                 in6.s6_addr8[12] = 0xff;
1451
1452                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1453                         ip6_mfilter_free(im6f[0]);
1454                         ip6_mfilter_free(im6f[1]);
1455                         break;
1456                 }
1457
1458                 if ((error = in6_joingroup(ifp, &in6, NULL, &im6f[1]->im6f_in6m, 0)) != 0) {
1459                         in6_leavegroup(im6f[0]->im6f_in6m, NULL);
1460                         ip6_mfilter_free(im6f[0]);
1461                         ip6_mfilter_free(im6f[1]);
1462                         break;
1463                 }
1464                 ip6_mfilter_insert(&im6o->im6o_head, im6f[0]);
1465                 ip6_mfilter_insert(&im6o->im6o_head, im6f[1]);
1466                 break;
1467             }
1468 #endif
1469         }
1470
1471         return (error);
1472 }
1473
1474 /*
1475  * Free multicast structures.
1476  */
1477 static void
1478 carp_multicast_cleanup(struct carp_if *cif, sa_family_t sa)
1479 {
1480 #ifdef INET
1481         struct ip_moptions *imo = &cif->cif_imo;
1482         struct in_mfilter *imf;
1483 #endif
1484 #ifdef INET6
1485         struct ip6_moptions *im6o = &cif->cif_im6o;
1486         struct in6_mfilter *im6f;
1487 #endif
1488         sx_assert(&carp_sx, SA_XLOCKED);
1489
1490         switch (sa) {
1491 #ifdef INET
1492         case AF_INET:
1493                 if (cif->cif_naddrs != 0)
1494                         break;
1495
1496                 while ((imf = ip_mfilter_first(&imo->imo_head)) != NULL) {
1497                         ip_mfilter_remove(&imo->imo_head, imf);
1498                         in_leavegroup(imf->imf_inm, NULL);
1499                         ip_mfilter_free(imf);
1500                 }
1501                 break;
1502 #endif
1503 #ifdef INET6
1504         case AF_INET6:
1505                 if (cif->cif_naddrs6 != 0)
1506                         break;
1507
1508                 while ((im6f = ip6_mfilter_first(&im6o->im6o_head)) != NULL) {
1509                         ip6_mfilter_remove(&im6o->im6o_head, im6f);
1510                         in6_leavegroup(im6f->im6f_in6m, NULL);
1511                         ip6_mfilter_free(im6f);
1512                 }
1513                 break;
1514 #endif
1515         }
1516 }
1517
1518 int
1519 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa)
1520 {
1521         struct m_tag *mtag;
1522         struct carp_softc *sc;
1523
1524         if (!sa)
1525                 return (0);
1526
1527         switch (sa->sa_family) {
1528 #ifdef INET
1529         case AF_INET:
1530                 break;
1531 #endif
1532 #ifdef INET6
1533         case AF_INET6:
1534                 break;
1535 #endif
1536         default:
1537                 return (0);
1538         }
1539
1540         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
1541         if (mtag == NULL)
1542                 return (0);
1543
1544         bcopy(mtag + 1, &sc, sizeof(sc));
1545
1546         /* Set the source MAC address to the Virtual Router MAC Address. */
1547         switch (ifp->if_type) {
1548         case IFT_ETHER:
1549         case IFT_BRIDGE:
1550         case IFT_L2VLAN: {
1551                         struct ether_header *eh;
1552
1553                         eh = mtod(m, struct ether_header *);
1554                         eh->ether_shost[0] = 0;
1555                         eh->ether_shost[1] = 0;
1556                         eh->ether_shost[2] = 0x5e;
1557                         eh->ether_shost[3] = 0;
1558                         eh->ether_shost[4] = 1;
1559                         eh->ether_shost[5] = sc->sc_vhid;
1560                 }
1561                 break;
1562         default:
1563                 printf("%s: carp is not supported for the %d interface type\n",
1564                     ifp->if_xname, ifp->if_type);
1565                 return (EOPNOTSUPP);
1566         }
1567
1568         return (0);
1569 }
1570
1571 static struct carp_softc*
1572 carp_alloc(struct ifnet *ifp)
1573 {
1574         struct carp_softc *sc;
1575         struct carp_if *cif;
1576
1577         sx_assert(&carp_sx, SA_XLOCKED);
1578
1579         if ((cif = ifp->if_carp) == NULL)
1580                 cif = carp_alloc_if(ifp);
1581
1582         sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
1583
1584         sc->sc_advbase = CARP_DFLTINTV;
1585         sc->sc_vhid = -1;       /* required setting */
1586         sc->sc_init_counter = 1;
1587         sc->sc_state = INIT;
1588
1589         sc->sc_ifasiz = sizeof(struct ifaddr *);
1590         sc->sc_ifas = malloc(sc->sc_ifasiz, M_CARP, M_WAITOK|M_ZERO);
1591         sc->sc_carpdev = ifp;
1592
1593         CARP_LOCK_INIT(sc);
1594 #ifdef INET
1595         callout_init_mtx(&sc->sc_md_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1596 #endif
1597 #ifdef INET6
1598         callout_init_mtx(&sc->sc_md6_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1599 #endif
1600         callout_init_mtx(&sc->sc_ad_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1601
1602         CIF_LOCK(cif);
1603         TAILQ_INSERT_TAIL(&cif->cif_vrs, sc, sc_list);
1604         CIF_UNLOCK(cif);
1605
1606         mtx_lock(&carp_mtx);
1607         LIST_INSERT_HEAD(&carp_list, sc, sc_next);
1608         mtx_unlock(&carp_mtx);
1609
1610         return (sc);
1611 }
1612
1613 static void
1614 carp_grow_ifas(struct carp_softc *sc)
1615 {
1616         struct ifaddr **new;
1617
1618         new = malloc(sc->sc_ifasiz * 2, M_CARP, M_WAITOK | M_ZERO);
1619         CARP_LOCK(sc);
1620         bcopy(sc->sc_ifas, new, sc->sc_ifasiz);
1621         free(sc->sc_ifas, M_CARP);
1622         sc->sc_ifas = new;
1623         sc->sc_ifasiz *= 2;
1624         CARP_UNLOCK(sc);
1625 }
1626
1627 static void
1628 carp_destroy(struct carp_softc *sc)
1629 {
1630         struct ifnet *ifp = sc->sc_carpdev;
1631         struct carp_if *cif = ifp->if_carp;
1632
1633         sx_assert(&carp_sx, SA_XLOCKED);
1634
1635         if (sc->sc_suppress)
1636                 carp_demote_adj(-V_carp_ifdown_adj, "vhid removed");
1637         CARP_UNLOCK(sc);
1638
1639         CIF_LOCK(cif);
1640         TAILQ_REMOVE(&cif->cif_vrs, sc, sc_list);
1641         CIF_UNLOCK(cif);
1642
1643         mtx_lock(&carp_mtx);
1644         LIST_REMOVE(sc, sc_next);
1645         mtx_unlock(&carp_mtx);
1646
1647         callout_drain(&sc->sc_ad_tmo);
1648 #ifdef INET
1649         callout_drain(&sc->sc_md_tmo);
1650 #endif
1651 #ifdef INET6
1652         callout_drain(&sc->sc_md6_tmo);
1653 #endif
1654         CARP_LOCK_DESTROY(sc);
1655
1656         free(sc->sc_ifas, M_CARP);
1657         free(sc, M_CARP);
1658 }
1659
1660 static struct carp_if*
1661 carp_alloc_if(struct ifnet *ifp)
1662 {
1663         struct carp_if *cif;
1664         int error;
1665
1666         cif = malloc(sizeof(*cif), M_CARP, M_WAITOK|M_ZERO);
1667
1668         if ((error = ifpromisc(ifp, 1)) != 0)
1669                 printf("%s: ifpromisc(%s) failed: %d\n",
1670                     __func__, ifp->if_xname, error);
1671         else
1672                 cif->cif_flags |= CIF_PROMISC;
1673
1674         CIF_LOCK_INIT(cif);
1675         cif->cif_ifp = ifp;
1676         TAILQ_INIT(&cif->cif_vrs);
1677
1678         IF_ADDR_WLOCK(ifp);
1679         ifp->if_carp = cif;
1680         if_ref(ifp);
1681         IF_ADDR_WUNLOCK(ifp);
1682
1683         return (cif);
1684 }
1685
1686 static void
1687 carp_free_if(struct carp_if *cif)
1688 {
1689         struct ifnet *ifp = cif->cif_ifp;
1690
1691         CIF_LOCK_ASSERT(cif);
1692         KASSERT(TAILQ_EMPTY(&cif->cif_vrs), ("%s: softc list not empty",
1693             __func__));
1694
1695         IF_ADDR_WLOCK(ifp);
1696         ifp->if_carp = NULL;
1697         IF_ADDR_WUNLOCK(ifp);
1698
1699         CIF_LOCK_DESTROY(cif);
1700
1701         if (cif->cif_flags & CIF_PROMISC)
1702                 ifpromisc(ifp, 0);
1703         if_rele(ifp);
1704
1705         free(cif, M_CARP);
1706 }
1707
1708 static void
1709 carp_carprcp(struct carpreq *carpr, struct carp_softc *sc, int priv)
1710 {
1711
1712         CARP_LOCK(sc);
1713         carpr->carpr_state = sc->sc_state;
1714         carpr->carpr_vhid = sc->sc_vhid;
1715         carpr->carpr_advbase = sc->sc_advbase;
1716         carpr->carpr_advskew = sc->sc_advskew;
1717         if (priv)
1718                 bcopy(sc->sc_key, carpr->carpr_key, sizeof(carpr->carpr_key));
1719         else
1720                 bzero(carpr->carpr_key, sizeof(carpr->carpr_key));
1721         CARP_UNLOCK(sc);
1722 }
1723
1724 int
1725 carp_ioctl(struct ifreq *ifr, u_long cmd, struct thread *td)
1726 {
1727         struct carpreq carpr;
1728         struct ifnet *ifp;
1729         struct carp_softc *sc = NULL;
1730         int error = 0, locked = 0;
1731
1732         if ((error = copyin(ifr_data_get_ptr(ifr), &carpr, sizeof carpr)))
1733                 return (error);
1734
1735         ifp = ifunit_ref(ifr->ifr_name);
1736         if (ifp == NULL)
1737                 return (ENXIO);
1738
1739         switch (ifp->if_type) {
1740         case IFT_ETHER:
1741         case IFT_L2VLAN:
1742         case IFT_BRIDGE:
1743                 break;
1744         default:
1745                 error = EOPNOTSUPP;
1746                 goto out;
1747         }
1748
1749         if ((ifp->if_flags & IFF_MULTICAST) == 0) {
1750                 error = EADDRNOTAVAIL;
1751                 goto out;
1752         }
1753
1754         sx_xlock(&carp_sx);
1755         switch (cmd) {
1756         case SIOCSVH:
1757                 if ((error = priv_check(td, PRIV_NETINET_CARP)))
1758                         break;
1759                 if (carpr.carpr_vhid <= 0 || carpr.carpr_vhid > CARP_MAXVHID ||
1760                     carpr.carpr_advbase < 0 || carpr.carpr_advskew < 0) {
1761                         error = EINVAL;
1762                         break;
1763                 }
1764
1765                 if (ifp->if_carp) {
1766                         IFNET_FOREACH_CARP(ifp, sc)
1767                                 if (sc->sc_vhid == carpr.carpr_vhid)
1768                                         break;
1769                 }
1770                 if (sc == NULL) {
1771                         sc = carp_alloc(ifp);
1772                         CARP_LOCK(sc);
1773                         sc->sc_vhid = carpr.carpr_vhid;
1774                         LLADDR(&sc->sc_addr)[0] = 0;
1775                         LLADDR(&sc->sc_addr)[1] = 0;
1776                         LLADDR(&sc->sc_addr)[2] = 0x5e;
1777                         LLADDR(&sc->sc_addr)[3] = 0;
1778                         LLADDR(&sc->sc_addr)[4] = 1;
1779                         LLADDR(&sc->sc_addr)[5] = sc->sc_vhid;
1780                 } else
1781                         CARP_LOCK(sc);
1782                 locked = 1;
1783                 if (carpr.carpr_advbase > 0) {
1784                         if (carpr.carpr_advbase > 255 ||
1785                             carpr.carpr_advbase < CARP_DFLTINTV) {
1786                                 error = EINVAL;
1787                                 break;
1788                         }
1789                         sc->sc_advbase = carpr.carpr_advbase;
1790                 }
1791                 if (carpr.carpr_advskew >= 255) {
1792                         error = EINVAL;
1793                         break;
1794                 }
1795                 sc->sc_advskew = carpr.carpr_advskew;
1796                 if (carpr.carpr_key[0] != '\0') {
1797                         bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1798                         carp_hmac_prepare(sc);
1799                 }
1800                 if (sc->sc_state != INIT &&
1801                     carpr.carpr_state != sc->sc_state) {
1802                         switch (carpr.carpr_state) {
1803                         case BACKUP:
1804                                 callout_stop(&sc->sc_ad_tmo);
1805                                 carp_set_state(sc, BACKUP,
1806                                     "user requested via ifconfig");
1807                                 carp_setrun(sc, 0);
1808                                 carp_delroute(sc);
1809                                 break;
1810                         case MASTER:
1811                                 carp_master_down_locked(sc,
1812                                     "user requested via ifconfig");
1813                                 break;
1814                         default:
1815                                 break;
1816                         }
1817                 }
1818                 break;
1819
1820         case SIOCGVH:
1821             {
1822                 int priveleged;
1823
1824                 if (carpr.carpr_vhid < 0 || carpr.carpr_vhid > CARP_MAXVHID) {
1825                         error = EINVAL;
1826                         break;
1827                 }
1828                 if (carpr.carpr_count < 1) {
1829                         error = EMSGSIZE;
1830                         break;
1831                 }
1832                 if (ifp->if_carp == NULL) {
1833                         error = ENOENT;
1834                         break;
1835                 }
1836
1837                 priveleged = (priv_check(td, PRIV_NETINET_CARP) == 0);
1838                 if (carpr.carpr_vhid != 0) {
1839                         IFNET_FOREACH_CARP(ifp, sc)
1840                                 if (sc->sc_vhid == carpr.carpr_vhid)
1841                                         break;
1842                         if (sc == NULL) {
1843                                 error = ENOENT;
1844                                 break;
1845                         }
1846                         carp_carprcp(&carpr, sc, priveleged);
1847                         error = copyout(&carpr, ifr_data_get_ptr(ifr),
1848                             sizeof(carpr));
1849                 } else  {
1850                         int i, count;
1851
1852                         count = 0;
1853                         IFNET_FOREACH_CARP(ifp, sc)
1854                                 count++;
1855
1856                         if (count > carpr.carpr_count) {
1857                                 CIF_UNLOCK(ifp->if_carp);
1858                                 error = EMSGSIZE;
1859                                 break;
1860                         }
1861
1862                         i = 0;
1863                         IFNET_FOREACH_CARP(ifp, sc) {
1864                                 carp_carprcp(&carpr, sc, priveleged);
1865                                 carpr.carpr_count = count;
1866                                 error = copyout(&carpr,
1867                                     (char *)ifr_data_get_ptr(ifr) +
1868                                     (i * sizeof(carpr)), sizeof(carpr));
1869                                 if (error) {
1870                                         CIF_UNLOCK(ifp->if_carp);
1871                                         break;
1872                                 }
1873                                 i++;
1874                         }
1875                 }
1876                 break;
1877             }
1878         default:
1879                 error = EINVAL;
1880         }
1881         sx_xunlock(&carp_sx);
1882
1883 out:
1884         if (locked)
1885                 CARP_UNLOCK(sc);
1886         if_rele(ifp);
1887
1888         return (error);
1889 }
1890
1891 static int
1892 carp_get_vhid(struct ifaddr *ifa)
1893 {
1894
1895         if (ifa == NULL || ifa->ifa_carp == NULL)
1896                 return (0);
1897
1898         return (ifa->ifa_carp->sc_vhid);
1899 }
1900
1901 int
1902 carp_attach(struct ifaddr *ifa, int vhid)
1903 {
1904         struct ifnet *ifp = ifa->ifa_ifp;
1905         struct carp_if *cif = ifp->if_carp;
1906         struct carp_softc *sc;
1907         int index, error;
1908
1909         KASSERT(ifa->ifa_carp == NULL, ("%s: ifa %p attached", __func__, ifa));
1910
1911         switch (ifa->ifa_addr->sa_family) {
1912 #ifdef INET
1913         case AF_INET:
1914 #endif
1915 #ifdef INET6
1916         case AF_INET6:
1917 #endif
1918                 break;
1919         default:
1920                 return (EPROTOTYPE);
1921         }
1922
1923         sx_xlock(&carp_sx);
1924         if (ifp->if_carp == NULL) {
1925                 sx_xunlock(&carp_sx);
1926                 return (ENOPROTOOPT);
1927         }
1928
1929         IFNET_FOREACH_CARP(ifp, sc)
1930                 if (sc->sc_vhid == vhid)
1931                         break;
1932         if (sc == NULL) {
1933                 sx_xunlock(&carp_sx);
1934                 return (ENOENT);
1935         }
1936
1937         error = carp_multicast_setup(cif, ifa->ifa_addr->sa_family);
1938         if (error) {
1939                 CIF_FREE(cif);
1940                 sx_xunlock(&carp_sx);
1941                 return (error);
1942         }
1943
1944         index = sc->sc_naddrs + sc->sc_naddrs6 + 1;
1945         if (index > sc->sc_ifasiz / sizeof(struct ifaddr *))
1946                 carp_grow_ifas(sc);
1947
1948         switch (ifa->ifa_addr->sa_family) {
1949 #ifdef INET
1950         case AF_INET:
1951                 cif->cif_naddrs++;
1952                 sc->sc_naddrs++;
1953                 break;
1954 #endif
1955 #ifdef INET6
1956         case AF_INET6:
1957                 cif->cif_naddrs6++;
1958                 sc->sc_naddrs6++;
1959                 break;
1960 #endif
1961         }
1962
1963         ifa_ref(ifa);
1964
1965         CARP_LOCK(sc);
1966         sc->sc_ifas[index - 1] = ifa;
1967         ifa->ifa_carp = sc;
1968         carp_hmac_prepare(sc);
1969         carp_sc_state(sc);
1970         CARP_UNLOCK(sc);
1971
1972         sx_xunlock(&carp_sx);
1973
1974         return (0);
1975 }
1976
1977 void
1978 carp_detach(struct ifaddr *ifa, bool keep_cif)
1979 {
1980         struct ifnet *ifp = ifa->ifa_ifp;
1981         struct carp_if *cif = ifp->if_carp;
1982         struct carp_softc *sc = ifa->ifa_carp;
1983         int i, index;
1984
1985         KASSERT(sc != NULL, ("%s: %p not attached", __func__, ifa));
1986
1987         sx_xlock(&carp_sx);
1988
1989         CARP_LOCK(sc);
1990         /* Shift array. */
1991         index = sc->sc_naddrs + sc->sc_naddrs6;
1992         for (i = 0; i < index; i++)
1993                 if (sc->sc_ifas[i] == ifa)
1994                         break;
1995         KASSERT(i < index, ("%s: %p no backref", __func__, ifa));
1996         for (; i < index - 1; i++)
1997                 sc->sc_ifas[i] = sc->sc_ifas[i+1];
1998         sc->sc_ifas[index - 1] = NULL;
1999
2000         switch (ifa->ifa_addr->sa_family) {
2001 #ifdef INET
2002         case AF_INET:
2003                 cif->cif_naddrs--;
2004                 sc->sc_naddrs--;
2005                 break;
2006 #endif
2007 #ifdef INET6
2008         case AF_INET6:
2009                 cif->cif_naddrs6--;
2010                 sc->sc_naddrs6--;
2011                 break;
2012 #endif
2013         }
2014
2015         carp_ifa_delroute(ifa);
2016         carp_multicast_cleanup(cif, ifa->ifa_addr->sa_family);
2017
2018         ifa->ifa_carp = NULL;
2019         ifa_free(ifa);
2020
2021         carp_hmac_prepare(sc);
2022         carp_sc_state(sc);
2023
2024         if (!keep_cif && sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0)
2025                 carp_destroy(sc);
2026         else
2027                 CARP_UNLOCK(sc);
2028
2029         if (!keep_cif)
2030                 CIF_FREE(cif);
2031
2032         sx_xunlock(&carp_sx);
2033 }
2034
2035 static void
2036 carp_set_state(struct carp_softc *sc, int state, const char *reason)
2037 {
2038
2039         CARP_LOCK_ASSERT(sc);
2040
2041         if (sc->sc_state != state) {
2042                 const char *carp_states[] = { CARP_STATES };
2043                 char subsys[IFNAMSIZ+5];
2044
2045                 snprintf(subsys, IFNAMSIZ+5, "%u@%s", sc->sc_vhid,
2046                     sc->sc_carpdev->if_xname);
2047
2048                 CARP_LOG("%s: %s -> %s (%s)\n", subsys,
2049                     carp_states[sc->sc_state], carp_states[state], reason);
2050
2051                 sc->sc_state = state;
2052
2053                 devctl_notify("CARP", subsys, carp_states[state], NULL);
2054         }
2055 }
2056
2057 static void
2058 carp_linkstate(struct ifnet *ifp)
2059 {
2060         struct carp_softc *sc;
2061
2062         CIF_LOCK(ifp->if_carp);
2063         IFNET_FOREACH_CARP(ifp, sc) {
2064                 CARP_LOCK(sc);
2065                 carp_sc_state(sc);
2066                 CARP_UNLOCK(sc);
2067         }
2068         CIF_UNLOCK(ifp->if_carp);
2069 }
2070
2071 static void
2072 carp_sc_state(struct carp_softc *sc)
2073 {
2074
2075         CARP_LOCK_ASSERT(sc);
2076
2077         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
2078             !(sc->sc_carpdev->if_flags & IFF_UP) ||
2079             !V_carp_allow) {
2080                 callout_stop(&sc->sc_ad_tmo);
2081 #ifdef INET
2082                 callout_stop(&sc->sc_md_tmo);
2083 #endif
2084 #ifdef INET6
2085                 callout_stop(&sc->sc_md6_tmo);
2086 #endif
2087                 carp_set_state(sc, INIT, "hardware interface down");
2088                 carp_setrun(sc, 0);
2089                 if (!sc->sc_suppress)
2090                         carp_demote_adj(V_carp_ifdown_adj, "interface down");
2091                 sc->sc_suppress = 1;
2092         } else {
2093                 carp_set_state(sc, INIT, "hardware interface up");
2094                 carp_setrun(sc, 0);
2095                 if (sc->sc_suppress)
2096                         carp_demote_adj(-V_carp_ifdown_adj, "interface up");
2097                 sc->sc_suppress = 0;
2098         }
2099 }
2100
2101 static void
2102 carp_demote_adj(int adj, char *reason)
2103 {
2104         atomic_add_int(&V_carp_demotion, adj);
2105         CARP_LOG("demoted by %d to %d (%s)\n", adj, V_carp_demotion, reason);
2106         taskqueue_enqueue(taskqueue_swi, &carp_sendall_task);
2107 }
2108
2109 static int
2110 carp_allow_sysctl(SYSCTL_HANDLER_ARGS)
2111 {
2112         int new, error;
2113         struct carp_softc *sc;
2114
2115         new = V_carp_allow;
2116         error = sysctl_handle_int(oidp, &new, 0, req);
2117         if (error || !req->newptr)
2118                 return (error);
2119
2120         if (V_carp_allow != new) {
2121                 V_carp_allow = new;
2122
2123                 mtx_lock(&carp_mtx);
2124                 LIST_FOREACH(sc, &carp_list, sc_next) {
2125                         CARP_LOCK(sc);
2126                         if (curvnet == sc->sc_carpdev->if_vnet)
2127                                 carp_sc_state(sc);
2128                         CARP_UNLOCK(sc);
2129                 }
2130                 mtx_unlock(&carp_mtx);
2131         }
2132
2133         return (0);
2134 }
2135
2136 static int
2137 carp_dscp_sysctl(SYSCTL_HANDLER_ARGS)
2138 {
2139         int new, error;
2140
2141         new = V_carp_dscp;
2142         error = sysctl_handle_int(oidp, &new, 0, req);
2143         if (error || !req->newptr)
2144                 return (error);
2145
2146         if (new < 0 || new > 63)
2147                 return (EINVAL);
2148
2149         V_carp_dscp = new;
2150
2151         return (0);
2152 }
2153
2154 static int
2155 carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS)
2156 {
2157         int new, error;
2158
2159         new = V_carp_demotion;
2160         error = sysctl_handle_int(oidp, &new, 0, req);
2161         if (error || !req->newptr)
2162                 return (error);
2163
2164         carp_demote_adj(new, "sysctl");
2165
2166         return (0);
2167 }
2168
2169 #ifdef INET
2170 extern  struct domain inetdomain;
2171 static struct protosw in_carp_protosw = {
2172         .pr_type =              SOCK_RAW,
2173         .pr_domain =            &inetdomain,
2174         .pr_protocol =          IPPROTO_CARP,
2175         .pr_flags =             PR_ATOMIC|PR_ADDR,
2176         .pr_input =             carp_input,
2177         .pr_output =            rip_output,
2178         .pr_ctloutput =         rip_ctloutput,
2179         .pr_usrreqs =           &rip_usrreqs
2180 };
2181 #endif
2182
2183 #ifdef INET6
2184 extern  struct domain inet6domain;
2185 static struct protosw in6_carp_protosw = {
2186         .pr_type =              SOCK_RAW,
2187         .pr_domain =            &inet6domain,
2188         .pr_protocol =          IPPROTO_CARP,
2189         .pr_flags =             PR_ATOMIC|PR_ADDR,
2190         .pr_input =             carp6_input,
2191         .pr_output =            rip6_output,
2192         .pr_ctloutput =         rip6_ctloutput,
2193         .pr_usrreqs =           &rip6_usrreqs
2194 };
2195 #endif
2196
2197 static void
2198 carp_mod_cleanup(void)
2199 {
2200
2201 #ifdef INET
2202         if (proto_reg[CARP_INET] == 0) {
2203                 (void)ipproto_unregister(IPPROTO_CARP);
2204                 pf_proto_unregister(PF_INET, IPPROTO_CARP, SOCK_RAW);
2205                 proto_reg[CARP_INET] = -1;
2206         }
2207         carp_iamatch_p = NULL;
2208 #endif
2209 #ifdef INET6
2210         if (proto_reg[CARP_INET6] == 0) {
2211                 (void)ip6proto_unregister(IPPROTO_CARP);
2212                 pf_proto_unregister(PF_INET6, IPPROTO_CARP, SOCK_RAW);
2213                 proto_reg[CARP_INET6] = -1;
2214         }
2215         carp_iamatch6_p = NULL;
2216         carp_macmatch6_p = NULL;
2217 #endif
2218         carp_ioctl_p = NULL;
2219         carp_attach_p = NULL;
2220         carp_detach_p = NULL;
2221         carp_get_vhid_p = NULL;
2222         carp_linkstate_p = NULL;
2223         carp_forus_p = NULL;
2224         carp_output_p = NULL;
2225         carp_demote_adj_p = NULL;
2226         carp_master_p = NULL;
2227         mtx_unlock(&carp_mtx);
2228         taskqueue_drain(taskqueue_swi, &carp_sendall_task);
2229         mtx_destroy(&carp_mtx);
2230         sx_destroy(&carp_sx);
2231 }
2232
2233 static int
2234 carp_mod_load(void)
2235 {
2236         int err;
2237
2238         mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2239         sx_init(&carp_sx, "carp_sx");
2240         LIST_INIT(&carp_list);
2241         carp_get_vhid_p = carp_get_vhid;
2242         carp_forus_p = carp_forus;
2243         carp_output_p = carp_output;
2244         carp_linkstate_p = carp_linkstate;
2245         carp_ioctl_p = carp_ioctl;
2246         carp_attach_p = carp_attach;
2247         carp_detach_p = carp_detach;
2248         carp_demote_adj_p = carp_demote_adj;
2249         carp_master_p = carp_master;
2250 #ifdef INET6
2251         carp_iamatch6_p = carp_iamatch6;
2252         carp_macmatch6_p = carp_macmatch6;
2253         proto_reg[CARP_INET6] = pf_proto_register(PF_INET6,
2254             (struct protosw *)&in6_carp_protosw);
2255         if (proto_reg[CARP_INET6]) {
2256                 printf("carp: error %d attaching to PF_INET6\n",
2257                     proto_reg[CARP_INET6]);
2258                 carp_mod_cleanup();
2259                 return (proto_reg[CARP_INET6]);
2260         }
2261         err = ip6proto_register(IPPROTO_CARP);
2262         if (err) {
2263                 printf("carp: error %d registering with INET6\n", err);
2264                 carp_mod_cleanup();
2265                 return (err);
2266         }
2267 #endif
2268 #ifdef INET
2269         carp_iamatch_p = carp_iamatch;
2270         proto_reg[CARP_INET] = pf_proto_register(PF_INET, &in_carp_protosw);
2271         if (proto_reg[CARP_INET]) {
2272                 printf("carp: error %d attaching to PF_INET\n",
2273                     proto_reg[CARP_INET]);
2274                 carp_mod_cleanup();
2275                 return (proto_reg[CARP_INET]);
2276         }
2277         err = ipproto_register(IPPROTO_CARP);
2278         if (err) {
2279                 printf("carp: error %d registering with INET\n", err);
2280                 carp_mod_cleanup();
2281                 return (err);
2282         }
2283 #endif
2284         return (0);
2285 }
2286
2287 static int
2288 carp_modevent(module_t mod, int type, void *data)
2289 {
2290         switch (type) {
2291         case MOD_LOAD:
2292                 return carp_mod_load();
2293                 /* NOTREACHED */
2294         case MOD_UNLOAD:
2295                 mtx_lock(&carp_mtx);
2296                 if (LIST_EMPTY(&carp_list))
2297                         carp_mod_cleanup();
2298                 else {
2299                         mtx_unlock(&carp_mtx);
2300                         return (EBUSY);
2301                 }
2302                 break;
2303
2304         default:
2305                 return (EINVAL);
2306         }
2307
2308         return (0);
2309 }
2310
2311 static moduledata_t carp_mod = {
2312         "carp",
2313         carp_modevent,
2314         0
2315 };
2316
2317 DECLARE_MODULE(carp, carp_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);