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, 0,  "CARP");
 222 SYSCTL_PROC(_net_inet_carp, OID_AUTO, allow,
 223     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW, 0, 0, carp_allow_sysctl, "I",
 224     "Accept incoming CARP packets");
 225 SYSCTL_PROC(_net_inet_carp, OID_AUTO, dscp,
 226     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW, 0, 0, carp_dscp_sysctl, "I",
 227     "DSCP value for carp packets");
 228 SYSCTL_INT(_net_inet_carp, OID_AUTO, preempt, CTLFLAG_VNET | CTLFLAG_RW,
 229     &VNET_NAME(carp_preempt), 0, "High-priority backup preemption mode");
 230 SYSCTL_INT(_net_inet_carp, OID_AUTO, log, CTLFLAG_VNET | CTLFLAG_RW,
 231     &VNET_NAME(carp_log), 0, "CARP log level");
 232 SYSCTL_PROC(_net_inet_carp, OID_AUTO, demotion,
 233     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW,
 234     0, 0, carp_demote_adj_sysctl, "I",
 235     "Adjust demotion factor (skew of advskew)");
 236 SYSCTL_INT(_net_inet_carp, OID_AUTO, senderr_demotion_factor,
 237     CTLFLAG_VNET | CTLFLAG_RW,
 238     &VNET_NAME(carp_senderr_adj), 0, "Send error demotion factor adjustment");
 239 SYSCTL_INT(_net_inet_carp, OID_AUTO, ifdown_demotion_factor,
 240     CTLFLAG_VNET | CTLFLAG_RW,
 241     &VNET_NAME(carp_ifdown_adj), 0,
 242     "Interface down demotion factor adjustment");
 243
 244 VNET_PCPUSTAT_DEFINE(struct carpstats, carpstats);
 245 VNET_PCPUSTAT_SYSINIT(carpstats);
 246 VNET_PCPUSTAT_SYSUNINIT(carpstats);
 247
 248 #define CARPSTATS_ADD(name, val)        \
 249     counter_u64_add(VNET(carpstats)[offsetof(struct carpstats, name) / \
 250         sizeof(uint64_t)], (val))
 251 #define CARPSTATS_INC(name)             CARPSTATS_ADD(name, 1)
 252
 253 SYSCTL_VNET_PCPUSTAT(_net_inet_carp, OID_AUTO, stats, struct carpstats,
 254     carpstats, "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 255
 256 #define CARP_LOCK_INIT(sc)      mtx_init(&(sc)->sc_mtx, "carp_softc",   \
 257         NULL, MTX_DEF)
 258 #define CARP_LOCK_DESTROY(sc)   mtx_destroy(&(sc)->sc_mtx)
 259 #define CARP_LOCK_ASSERT(sc)    mtx_assert(&(sc)->sc_mtx, MA_OWNED)
 260 #define CARP_LOCK(sc)           mtx_lock(&(sc)->sc_mtx)
 261 #define CARP_UNLOCK(sc)         mtx_unlock(&(sc)->sc_mtx)
 262 #define CIF_LOCK_INIT(cif)      mtx_init(&(cif)->cif_mtx, "carp_if",   \
 263         NULL, MTX_DEF)
 264 #define CIF_LOCK_DESTROY(cif)   mtx_destroy(&(cif)->cif_mtx)
 265 #define CIF_LOCK_ASSERT(cif)    mtx_assert(&(cif)->cif_mtx, MA_OWNED)
 266 #define CIF_LOCK(cif)           mtx_lock(&(cif)->cif_mtx)
 267 #define CIF_UNLOCK(cif)         mtx_unlock(&(cif)->cif_mtx)
 268 #define CIF_FREE(cif)   do {                            \
 269                 CIF_LOCK(cif);                          \
 270                 if (TAILQ_EMPTY(&(cif)->cif_vrs))       \
 271                         carp_free_if(cif);              \
 272                 else                                    \
 273                         CIF_UNLOCK(cif);                \
 274 } while (0)
 275
 276 #define CARP_LOG(...)   do {                            \
 277         if (V_carp_log > 0)                             \
 278                 log(LOG_INFO, "carp: " __VA_ARGS__);    \
 279 } while (0)
 280
 281 #define CARP_DEBUG(...) do {                            \
 282         if (V_carp_log > 1)                             \
 283                 log(LOG_DEBUG, __VA_ARGS__);            \
 284 } while (0)
 285
 286 #define IFNET_FOREACH_IFA(ifp, ifa)                                     \
 287         CK_STAILQ_FOREACH((ifa), &(ifp)->if_addrhead, ifa_link) \
 288                 if ((ifa)->ifa_carp != NULL)
 289
 290 #define CARP_FOREACH_IFA(sc, ifa)                                       \
 291         CARP_LOCK_ASSERT(sc);                                           \
 292         for (int _i = 0;                                                \
 293                 _i < (sc)->sc_naddrs + (sc)->sc_naddrs6 &&              \
 294                 ((ifa) = sc->sc_ifas[_i]) != NULL;                      \
 295                 ++_i)
 296
 297 #define IFNET_FOREACH_CARP(ifp, sc)                                     \
 298         KASSERT(mtx_owned(&ifp->if_carp->cif_mtx) ||                    \
 299             sx_xlocked(&carp_sx), ("cif_vrs not locked"));              \
 300         TAILQ_FOREACH((sc), &(ifp)->if_carp->cif_vrs, sc_list)
 301
 302 #define DEMOTE_ADVSKEW(sc)                                      \
 303     (((sc)->sc_advskew + V_carp_demotion > CARP_MAXSKEW) ?      \
 304     CARP_MAXSKEW : ((sc)->sc_advskew + V_carp_demotion))
 305
 306 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 307 static struct carp_softc
 308                 *carp_alloc(struct ifnet *);
 309 static void     carp_destroy(struct carp_softc *);
 310 static struct carp_if
 311                 *carp_alloc_if(struct ifnet *);
 312 static void     carp_free_if(struct carp_if *);
 313 static void     carp_set_state(struct carp_softc *, int, const char* reason);
 314 static void     carp_sc_state(struct carp_softc *);
 315 static void     carp_setrun(struct carp_softc *, sa_family_t);
 316 static void     carp_master_down(void *);
 317 static void     carp_master_down_locked(struct carp_softc *,
 318                     const char* reason);
 319 static void     carp_send_ad(void *);
 320 static void     carp_send_ad_locked(struct carp_softc *);
 321 static void     carp_addroute(struct carp_softc *);
 322 static void     carp_ifa_addroute(struct ifaddr *);
 323 static void     carp_delroute(struct carp_softc *);
 324 static void     carp_ifa_delroute(struct ifaddr *);
 325 static void     carp_send_ad_all(void *, int);
 326 static void     carp_demote_adj(int, char *);
 327
 328 static LIST_HEAD(, carp_softc) carp_list;
 329 static struct mtx carp_mtx;
 330 static struct sx carp_sx;
 331 static struct task carp_sendall_task =
 332     TASK_INITIALIZER(0, carp_send_ad_all, NULL);
 333
 334 static void
 335 carp_hmac_prepare(struct carp_softc *sc)
 336 {
 337         uint8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 338         uint8_t vhid = sc->sc_vhid & 0xff;
 339         struct ifaddr *ifa;
 340         int i, found;
 341 #ifdef INET
 342         struct in_addr last, cur, in;
 343 #endif
 344 #ifdef INET6
 345         struct in6_addr last6, cur6, in6;
 346 #endif
 347
 348         CARP_LOCK_ASSERT(sc);
 349
 350         /* Compute ipad from key. */
 351         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 352         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 353         for (i = 0; i < sizeof(sc->sc_pad); i++)
 354                 sc->sc_pad[i] ^= 0x36;
 355
 356         /* Precompute first part of inner hash. */
 357         SHA1Init(&sc->sc_sha1);
 358         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 359         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 360         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 361         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 362 #ifdef INET
 363         cur.s_addr = 0;
 364         do {
 365                 found = 0;
 366                 last = cur;
 367                 cur.s_addr = 0xffffffff;
 368                 CARP_FOREACH_IFA(sc, ifa) {
 369                         in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 370                         if (ifa->ifa_addr->sa_family == AF_INET &&
 371                             ntohl(in.s_addr) > ntohl(last.s_addr) &&
 372                             ntohl(in.s_addr) < ntohl(cur.s_addr)) {
 373                                 cur.s_addr = in.s_addr;
 374                                 found++;
 375                         }
 376                 }
 377                 if (found)
 378                         SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur));
 379         } while (found);
 380 #endif /* INET */
 381 #ifdef INET6
 382         memset(&cur6, 0, sizeof(cur6));
 383         do {
 384                 found = 0;
 385                 last6 = cur6;
 386                 memset(&cur6, 0xff, sizeof(cur6));
 387                 CARP_FOREACH_IFA(sc, ifa) {
 388                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 389                         if (IN6_IS_SCOPE_EMBED(&in6))
 390                                 in6.s6_addr16[1] = 0;
 391                         if (ifa->ifa_addr->sa_family == AF_INET6 &&
 392                             memcmp(&in6, &last6, sizeof(in6)) > 0 &&
 393                             memcmp(&in6, &cur6, sizeof(in6)) < 0) {
 394                                 cur6 = in6;
 395                                 found++;
 396                         }
 397                 }
 398                 if (found)
 399                         SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6));
 400         } while (found);
 401 #endif /* INET6 */
 402
 403         /* convert ipad to opad */
 404         for (i = 0; i < sizeof(sc->sc_pad); i++)
 405                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 406 }
 407
 408 static void
 409 carp_hmac_generate(struct carp_softc *sc, uint32_t counter[2],
 410     unsigned char md[20])
 411 {
 412         SHA1_CTX sha1ctx;
 413
 414         CARP_LOCK_ASSERT(sc);
 415
 416         /* fetch first half of inner hash */
 417         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 418
 419         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 420         SHA1Final(md, &sha1ctx);
 421
 422         /* outer hash */
 423         SHA1Init(&sha1ctx);
 424         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 425         SHA1Update(&sha1ctx, md, 20);
 426         SHA1Final(md, &sha1ctx);
 427 }
 428
 429 static int
 430 carp_hmac_verify(struct carp_softc *sc, uint32_t counter[2],
 431     unsigned char md[20])
 432 {
 433         unsigned char md2[20];
 434
 435         CARP_LOCK_ASSERT(sc);
 436
 437         carp_hmac_generate(sc, counter, md2);
 438
 439         return (bcmp(md, md2, sizeof(md2)));
 440 }
 441
 442 /*
 443  * process input packet.
 444  * we have rearranged checks order compared to the rfc,
 445  * but it seems more efficient this way or not possible otherwise.
 446  */
 447 #ifdef INET
 448 int
 449 carp_input(struct mbuf **mp, int *offp, int proto)
 450 {
 451         struct mbuf *m = *mp;
 452         struct ip *ip = mtod(m, struct ip *);
 453         struct carp_header *ch;
 454         int iplen, len;
 455
 456         iplen = *offp;
 457         *mp = NULL;
 458
 459         CARPSTATS_INC(carps_ipackets);
 460
 461         if (!V_carp_allow) {
 462                 m_freem(m);
 463                 return (IPPROTO_DONE);
 464         }
 465
 466         /* verify that the IP TTL is 255.  */
 467         if (ip->ip_ttl != CARP_DFLTTL) {
 468                 CARPSTATS_INC(carps_badttl);
 469                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 470                     ip->ip_ttl,
 471                     m->m_pkthdr.rcvif->if_xname);
 472                 m_freem(m);
 473                 return (IPPROTO_DONE);
 474         }
 475
 476         iplen = ip->ip_hl << 2;
 477
 478         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 479                 CARPSTATS_INC(carps_badlen);
 480                 CARP_DEBUG("%s: received len %zd < sizeof(struct carp_header) "
 481                     "on %s\n", __func__, m->m_len - sizeof(struct ip),
 482                     m->m_pkthdr.rcvif->if_xname);
 483                 m_freem(m);
 484                 return (IPPROTO_DONE);
 485         }
 486
 487         if (iplen + sizeof(*ch) < m->m_len) {
 488                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 489                         CARPSTATS_INC(carps_hdrops);
 490                         CARP_DEBUG("%s: pullup failed\n", __func__);
 491                         return (IPPROTO_DONE);
 492                 }
 493                 ip = mtod(m, struct ip *);
 494         }
 495         ch = (struct carp_header *)((char *)ip + iplen);
 496
 497         /*
 498          * verify that the received packet length is
 499          * equal to the CARP header
 500          */
 501         len = iplen + sizeof(*ch);
 502         if (len > m->m_pkthdr.len) {
 503                 CARPSTATS_INC(carps_badlen);
 504                 CARP_DEBUG("%s: packet too short %d on %s\n", __func__,
 505                     m->m_pkthdr.len,
 506                     m->m_pkthdr.rcvif->if_xname);
 507                 m_freem(m);
 508                 return (IPPROTO_DONE);
 509         }
 510
 511         if ((m = m_pullup(m, len)) == NULL) {
 512                 CARPSTATS_INC(carps_hdrops);
 513                 return (IPPROTO_DONE);
 514         }
 515         ip = mtod(m, struct ip *);
 516         ch = (struct carp_header *)((char *)ip + iplen);
 517
 518         /* verify the CARP checksum */
 519         m->m_data += iplen;
 520         if (in_cksum(m, len - iplen)) {
 521                 CARPSTATS_INC(carps_badsum);
 522                 CARP_DEBUG("%s: checksum failed on %s\n", __func__,
 523                     m->m_pkthdr.rcvif->if_xname);
 524                 m_freem(m);
 525                 return (IPPROTO_DONE);
 526         }
 527         m->m_data -= iplen;
 528
 529         carp_input_c(m, ch, AF_INET);
 530         return (IPPROTO_DONE);
 531 }
 532 #endif
 533
 534 #ifdef INET6
 535 int
 536 carp6_input(struct mbuf **mp, int *offp, int proto)
 537 {
 538         struct mbuf *m = *mp;
 539         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 540         struct carp_header *ch;
 541         u_int len;
 542
 543         CARPSTATS_INC(carps_ipackets6);
 544
 545         if (!V_carp_allow) {
 546                 m_freem(m);
 547                 return (IPPROTO_DONE);
 548         }
 549
 550         /* check if received on a valid carp interface */
 551         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 552                 CARPSTATS_INC(carps_badif);
 553                 CARP_DEBUG("%s: packet received on non-carp interface: %s\n",
 554                     __func__, m->m_pkthdr.rcvif->if_xname);
 555                 m_freem(m);
 556                 return (IPPROTO_DONE);
 557         }
 558
 559         /* verify that the IP TTL is 255 */
 560         if (ip6->ip6_hlim != CARP_DFLTTL) {
 561                 CARPSTATS_INC(carps_badttl);
 562                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 563                     ip6->ip6_hlim, m->m_pkthdr.rcvif->if_xname);
 564                 m_freem(m);
 565                 return (IPPROTO_DONE);
 566         }
 567
 568         /* verify that we have a complete carp packet */
 569         len = m->m_len;
 570         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 571         if (ch == NULL) {
 572                 CARPSTATS_INC(carps_badlen);
 573                 CARP_DEBUG("%s: packet size %u too small\n", __func__, len);
 574                 return (IPPROTO_DONE);
 575         }
 576
 577
 578         /* verify the CARP checksum */
 579         m->m_data += *offp;
 580         if (in_cksum(m, sizeof(*ch))) {
 581                 CARPSTATS_INC(carps_badsum);
 582                 CARP_DEBUG("%s: checksum failed, on %s\n", __func__,
 583                     m->m_pkthdr.rcvif->if_xname);
 584                 m_freem(m);
 585                 return (IPPROTO_DONE);
 586         }
 587         m->m_data -= *offp;
 588
 589         carp_input_c(m, ch, AF_INET6);
 590         return (IPPROTO_DONE);
 591 }
 592 #endif /* INET6 */
 593
 594 /*
 595  * This routine should not be necessary at all, but some switches
 596  * (VMWare ESX vswitches) can echo our own packets back at us,
 597  * and we must ignore them or they will cause us to drop out of
 598  * MASTER mode.
 599  *
 600  * We cannot catch all cases of network loops.  Instead, what we
 601  * do here is catch any packet that arrives with a carp header
 602  * with a VHID of 0, that comes from an address that is our own.
 603  * These packets are by definition "from us" (even if they are from
 604  * a misconfigured host that is pretending to be us).
 605  *
 606  * The VHID test is outside this mini-function.
 607  */
 608 static int
 609 carp_source_is_self(struct mbuf *m, struct ifaddr *ifa, sa_family_t af)
 610 {
 611 #ifdef INET
 612         struct ip *ip4;
 613         struct in_addr in4;
 614 #endif
 615 #ifdef INET6
 616         struct ip6_hdr *ip6;
 617         struct in6_addr in6;
 618 #endif
 619
 620         switch (af) {
 621 #ifdef INET
 622         case AF_INET:
 623                 ip4 = mtod(m, struct ip *);
 624                 in4 = ifatoia(ifa)->ia_addr.sin_addr;
 625                 return (in4.s_addr == ip4->ip_src.s_addr);
 626 #endif
 627 #ifdef INET6
 628         case AF_INET6:
 629                 ip6 = mtod(m, struct ip6_hdr *);
 630                 in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 631                 return (memcmp(&in6, &ip6->ip6_src, sizeof(in6)) == 0);
 632 #endif
 633         default:
 634                 break;
 635         }
 636         return (0);
 637 }
 638
 639 static void
 640 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 641 {
 642         struct ifnet *ifp = m->m_pkthdr.rcvif;
 643         struct ifaddr *ifa, *match;
 644         struct carp_softc *sc;
 645         uint64_t tmp_counter;
 646         struct timeval sc_tv, ch_tv;
 647         struct epoch_tracker et;
 648         int error;
 649
 650         /*
 651          * Verify that the VHID is valid on the receiving interface.
 652          *
 653          * There should be just one match.  If there are none
 654          * the VHID is not valid and we drop the packet.  If
 655          * there are multiple VHID matches, take just the first
 656          * one, for compatibility with previous code.  While we're
 657          * scanning, check for obvious loops in the network topology
 658          * (these should never happen, and as noted above, we may
 659          * miss real loops; this is just a double-check).
 660          */
 661         NET_EPOCH_ENTER(et);
 662         error = 0;
 663         match = NULL;
 664         IFNET_FOREACH_IFA(ifp, ifa) {
 665                 if (match == NULL && ifa->ifa_carp != NULL &&
 666                     ifa->ifa_addr->sa_family == af &&
 667                     ifa->ifa_carp->sc_vhid == ch->carp_vhid)
 668                         match = ifa;
 669                 if (ch->carp_vhid == 0 && carp_source_is_self(m, ifa, af))
 670                         error = ELOOP;
 671         }
 672         ifa = error ? NULL : match;
 673         if (ifa != NULL)
 674                 ifa_ref(ifa);
 675         NET_EPOCH_EXIT(et);
 676
 677         if (ifa == NULL) {
 678                 if (error == ELOOP) {
 679                         CARP_DEBUG("dropping looped packet on interface %s\n",
 680                             ifp->if_xname);
 681                         CARPSTATS_INC(carps_badif);     /* ??? */
 682                 } else {
 683                         CARPSTATS_INC(carps_badvhid);
 684                 }
 685                 m_freem(m);
 686                 return;
 687         }
 688
 689         /* verify the CARP version. */
 690         if (ch->carp_version != CARP_VERSION) {
 691                 CARPSTATS_INC(carps_badver);
 692                 CARP_DEBUG("%s: invalid version %d\n", ifp->if_xname,
 693                     ch->carp_version);
 694                 ifa_free(ifa);
 695                 m_freem(m);
 696                 return;
 697         }
 698
 699         sc = ifa->ifa_carp;
 700         CARP_LOCK(sc);
 701         ifa_free(ifa);
 702
 703         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 704                 CARPSTATS_INC(carps_badauth);
 705                 CARP_DEBUG("%s: incorrect hash for VHID %u@%s\n", __func__,
 706                     sc->sc_vhid, ifp->if_xname);
 707                 goto out;
 708         }
 709
 710         tmp_counter = ntohl(ch->carp_counter[0]);
 711         tmp_counter = tmp_counter<<32;
 712         tmp_counter += ntohl(ch->carp_counter[1]);
 713
 714         /* XXX Replay protection goes here */
 715
 716         sc->sc_init_counter = 0;
 717         sc->sc_counter = tmp_counter;
 718
 719         sc_tv.tv_sec = sc->sc_advbase;
 720         sc_tv.tv_usec = DEMOTE_ADVSKEW(sc) * 1000000 / 256;
 721         ch_tv.tv_sec = ch->carp_advbase;
 722         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 723
 724         switch (sc->sc_state) {
 725         case INIT:
 726                 break;
 727         case MASTER:
 728                 /*
 729                  * If we receive an advertisement from a master who's going to
 730                  * be more frequent than us, go into BACKUP state.
 731                  */
 732                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 733                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 734                         callout_stop(&sc->sc_ad_tmo);
 735                         carp_set_state(sc, BACKUP,
 736                             "more frequent advertisement received");
 737                         carp_setrun(sc, 0);
 738                         carp_delroute(sc);
 739                 }
 740                 break;
 741         case BACKUP:
 742                 /*
 743                  * If we're pre-empting masters who advertise slower than us,
 744                  * and this one claims to be slower, treat him as down.
 745                  */
 746                 if (V_carp_preempt && timevalcmp(&sc_tv, &ch_tv, <)) {
 747                         carp_master_down_locked(sc,
 748                             "preempting a slower master");
 749                         break;
 750                 }
 751
 752                 /*
 753                  *  If the master is going to advertise at such a low frequency
 754                  *  that he's guaranteed to time out, we'd might as well just
 755                  *  treat him as timed out now.
 756                  */
 757                 sc_tv.tv_sec = sc->sc_advbase * 3;
 758                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 759                         carp_master_down_locked(sc, "master will time out");
 760                         break;
 761                 }
 762
 763                 /*
 764                  * Otherwise, we reset the counter and wait for the next
 765                  * advertisement.
 766                  */
 767                 carp_setrun(sc, af);
 768                 break;
 769         }
 770
 771 out:
 772         CARP_UNLOCK(sc);
 773         m_freem(m);
 774 }
 775
 776 static int
 777 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 778 {
 779         struct m_tag *mtag;
 780
 781         if (sc->sc_init_counter) {
 782                 /* this could also be seconds since unix epoch */
 783                 sc->sc_counter = arc4random();
 784                 sc->sc_counter = sc->sc_counter << 32;
 785                 sc->sc_counter += arc4random();
 786         } else
 787                 sc->sc_counter++;
 788
 789         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 790         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 791
 792         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 793
 794         /* Tag packet for carp_output */
 795         if ((mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct carp_softc *),
 796             M_NOWAIT)) == NULL) {
 797                 m_freem(m);
 798                 CARPSTATS_INC(carps_onomem);
 799                 return (ENOMEM);
 800         }
 801         bcopy(&sc, mtag + 1, sizeof(sc));
 802         m_tag_prepend(m, mtag);
 803
 804         return (0);
 805 }
 806
 807 /*
 808  * To avoid LORs and possible recursions this function shouldn't
 809  * be called directly, but scheduled via taskqueue.
 810  */
 811 static void
 812 carp_send_ad_all(void *ctx __unused, int pending __unused)
 813 {
 814         struct carp_softc *sc;
 815
 816         mtx_lock(&carp_mtx);
 817         LIST_FOREACH(sc, &carp_list, sc_next)
 818                 if (sc->sc_state == MASTER) {
 819                         CARP_LOCK(sc);
 820                         CURVNET_SET(sc->sc_carpdev->if_vnet);
 821                         carp_send_ad_locked(sc);
 822                         CURVNET_RESTORE();
 823                         CARP_UNLOCK(sc);
 824                 }
 825         mtx_unlock(&carp_mtx);
 826 }
 827
 828 /* Send a periodic advertisement, executed in callout context. */
 829 static void
 830 carp_send_ad(void *v)
 831 {
 832         struct carp_softc *sc = v;
 833
 834         CARP_LOCK_ASSERT(sc);
 835         CURVNET_SET(sc->sc_carpdev->if_vnet);
 836         carp_send_ad_locked(sc);
 837         CURVNET_RESTORE();
 838         CARP_UNLOCK(sc);
 839 }
 840
 841 static void
 842 carp_send_ad_error(struct carp_softc *sc, int error)
 843 {
 844
 845         if (error) {
 846                 if (sc->sc_sendad_errors < INT_MAX)
 847                         sc->sc_sendad_errors++;
 848                 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 849                         static const char fmt[] = "send error %d on %s";
 850                         char msg[sizeof(fmt) + IFNAMSIZ];
 851
 852                         sprintf(msg, fmt, error, sc->sc_carpdev->if_xname);
 853                         carp_demote_adj(V_carp_senderr_adj, msg);
 854                 }
 855                 sc->sc_sendad_success = 0;
 856         } else {
 857                 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS &&
 858                     ++sc->sc_sendad_success >= CARP_SENDAD_MIN_SUCCESS) {
 859                         static const char fmt[] = "send ok on %s";
 860                         char msg[sizeof(fmt) + IFNAMSIZ];
 861
 862                         sprintf(msg, fmt, sc->sc_carpdev->if_xname);
 863                         carp_demote_adj(-V_carp_senderr_adj, msg);
 864                         sc->sc_sendad_errors = 0;
 865                 } else
 866                         sc->sc_sendad_errors = 0;
 867         }
 868 }
 869
 870 /*
 871  * Pick the best ifaddr on the given ifp for sending CARP
 872  * advertisements.
 873  *
 874  * "Best" here is defined by ifa_preferred().  This function is much
 875  * much like ifaof_ifpforaddr() except that we just use ifa_preferred().
 876  *
 877  * (This could be simplified to return the actual address, except that
 878  * it has a different format in AF_INET and AF_INET6.)
 879  */
 880 static struct ifaddr *
 881 carp_best_ifa(int af, struct ifnet *ifp)
 882 {
 883         struct epoch_tracker et;
 884         struct ifaddr *ifa, *best;
 885
 886         if (af >= AF_MAX)
 887                 return (NULL);
 888         best = NULL;
 889         NET_EPOCH_ENTER(et);
 890         CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
 891                 if (ifa->ifa_addr->sa_family == af &&
 892                     (best == NULL || ifa_preferred(best, ifa)))
 893                         best = ifa;
 894         }
 895         NET_EPOCH_EXIT(et);
 896         if (best != NULL)
 897                 ifa_ref(best);
 898         return (best);
 899 }
 900
 901 static void
 902 carp_send_ad_locked(struct carp_softc *sc)
 903 {
 904         struct carp_header ch;
 905         struct timeval tv;
 906         struct ifaddr *ifa;
 907         struct carp_header *ch_ptr;
 908         struct mbuf *m;
 909         int len, advskew;
 910
 911         CARP_LOCK_ASSERT(sc);
 912
 913         advskew = DEMOTE_ADVSKEW(sc);
 914         tv.tv_sec = sc->sc_advbase;
 915         tv.tv_usec = advskew * 1000000 / 256;
 916
 917         ch.carp_version = CARP_VERSION;
 918         ch.carp_type = CARP_ADVERTISEMENT;
 919         ch.carp_vhid = sc->sc_vhid;
 920         ch.carp_advbase = sc->sc_advbase;
 921         ch.carp_advskew = advskew;
 922         ch.carp_authlen = 7;    /* XXX DEFINE */
 923         ch.carp_pad1 = 0;       /* must be zero */
 924         ch.carp_cksum = 0;
 925
 926         /* XXXGL: OpenBSD picks first ifaddr with needed family. */
 927
 928 #ifdef INET
 929         if (sc->sc_naddrs) {
 930                 struct ip *ip;
 931
 932                 m = m_gethdr(M_NOWAIT, MT_DATA);
 933                 if (m == NULL) {
 934                         CARPSTATS_INC(carps_onomem);
 935                         goto resched;
 936                 }
 937                 len = sizeof(*ip) + sizeof(ch);
 938                 m->m_pkthdr.len = len;
 939                 m->m_pkthdr.rcvif = NULL;
 940                 m->m_len = len;
 941                 M_ALIGN(m, m->m_len);
 942                 m->m_flags |= M_MCAST;
 943                 ip = mtod(m, struct ip *);
 944                 ip->ip_v = IPVERSION;
 945                 ip->ip_hl = sizeof(*ip) >> 2;
 946                 ip->ip_tos = V_carp_dscp << IPTOS_DSCP_OFFSET;
 947                 ip->ip_len = htons(len);
 948                 ip->ip_off = htons(IP_DF);
 949                 ip->ip_ttl = CARP_DFLTTL;
 950                 ip->ip_p = IPPROTO_CARP;
 951                 ip->ip_sum = 0;
 952                 ip_fillid(ip);
 953
 954                 ifa = carp_best_ifa(AF_INET, sc->sc_carpdev);
 955                 if (ifa != NULL) {
 956                         ip->ip_src.s_addr =
 957                             ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 958                         ifa_free(ifa);
 959                 } else
 960                         ip->ip_src.s_addr = 0;
 961                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 962
 963                 ch_ptr = (struct carp_header *)(&ip[1]);
 964                 bcopy(&ch, ch_ptr, sizeof(ch));
 965                 if (carp_prepare_ad(m, sc, ch_ptr))
 966                         goto resched;
 967
 968                 m->m_data += sizeof(*ip);
 969                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip));
 970                 m->m_data -= sizeof(*ip);
 971
 972                 CARPSTATS_INC(carps_opackets);
 973
 974                 carp_send_ad_error(sc, ip_output(m, NULL, NULL, IP_RAWOUTPUT,
 975                     &sc->sc_carpdev->if_carp->cif_imo, NULL));
 976         }
 977 #endif /* INET */
 978 #ifdef INET6
 979         if (sc->sc_naddrs6) {
 980                 struct ip6_hdr *ip6;
 981
 982                 m = m_gethdr(M_NOWAIT, MT_DATA);
 983                 if (m == NULL) {
 984                         CARPSTATS_INC(carps_onomem);
 985                         goto resched;
 986                 }
 987                 len = sizeof(*ip6) + sizeof(ch);
 988                 m->m_pkthdr.len = len;
 989                 m->m_pkthdr.rcvif = NULL;
 990                 m->m_len = len;
 991                 M_ALIGN(m, m->m_len);
 992                 m->m_flags |= M_MCAST;
 993                 ip6 = mtod(m, struct ip6_hdr *);
 994                 bzero(ip6, sizeof(*ip6));
 995                 ip6->ip6_vfc |= IPV6_VERSION;
 996                 /* Traffic class isn't defined in ip6 struct instead
 997                  * it gets offset into flowid field */
 998                 ip6->ip6_flow |= htonl(V_carp_dscp << (IPV6_FLOWLABEL_LEN +
 999                     IPTOS_DSCP_OFFSET));
1000                 ip6->ip6_hlim = CARP_DFLTTL;
1001                 ip6->ip6_nxt = IPPROTO_CARP;
1002
1003                 /* set the source address */
1004                 ifa = carp_best_ifa(AF_INET6, sc->sc_carpdev);
1005                 if (ifa != NULL) {
1006                         bcopy(IFA_IN6(ifa), &ip6->ip6_src,
1007                             sizeof(struct in6_addr));
1008                         ifa_free(ifa);
1009                 } else
1010                         /* This should never happen with IPv6. */
1011                         bzero(&ip6->ip6_src, sizeof(struct in6_addr));
1012
1013                 /* Set the multicast destination. */
1014                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
1015                 ip6->ip6_dst.s6_addr8[15] = 0x12;
1016                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
1017                         m_freem(m);
1018                         CARP_DEBUG("%s: in6_setscope failed\n", __func__);
1019                         goto resched;
1020                 }
1021
1022                 ch_ptr = (struct carp_header *)(&ip6[1]);
1023                 bcopy(&ch, ch_ptr, sizeof(ch));
1024                 if (carp_prepare_ad(m, sc, ch_ptr))
1025                         goto resched;
1026
1027                 m->m_data += sizeof(*ip6);
1028                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip6));
1029                 m->m_data -= sizeof(*ip6);
1030
1031                 CARPSTATS_INC(carps_opackets6);
1032
1033                 carp_send_ad_error(sc, ip6_output(m, NULL, NULL, 0,
1034                     &sc->sc_carpdev->if_carp->cif_im6o, NULL, NULL));
1035         }
1036 #endif /* INET6 */
1037
1038 resched:
1039         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), carp_send_ad, sc);
1040 }
1041
1042 static void
1043 carp_addroute(struct carp_softc *sc)
1044 {
1045         struct ifaddr *ifa;
1046
1047         CARP_FOREACH_IFA(sc, ifa)
1048                 carp_ifa_addroute(ifa);
1049 }
1050
1051 static void
1052 carp_ifa_addroute(struct ifaddr *ifa)
1053 {
1054
1055         switch (ifa->ifa_addr->sa_family) {
1056 #ifdef INET
1057         case AF_INET:
1058                 in_addprefix(ifatoia(ifa), RTF_UP);
1059                 ifa_add_loopback_route(ifa,
1060                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
1061                 break;
1062 #endif
1063 #ifdef INET6
1064         case AF_INET6:
1065                 ifa_add_loopback_route(ifa,
1066                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
1067                 nd6_add_ifa_lle(ifatoia6(ifa));
1068                 break;
1069 #endif
1070         }
1071 }
1072
1073 static void
1074 carp_delroute(struct carp_softc *sc)
1075 {
1076         struct ifaddr *ifa;
1077
1078         CARP_FOREACH_IFA(sc, ifa)
1079                 carp_ifa_delroute(ifa);
1080 }
1081
1082 static void
1083 carp_ifa_delroute(struct ifaddr *ifa)
1084 {
1085
1086         switch (ifa->ifa_addr->sa_family) {
1087 #ifdef INET
1088         case AF_INET:
1089                 ifa_del_loopback_route(ifa,
1090                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
1091                 in_scrubprefix(ifatoia(ifa), LLE_STATIC);
1092                 break;
1093 #endif
1094 #ifdef INET6
1095         case AF_INET6:
1096                 ifa_del_loopback_route(ifa,
1097                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
1098                 nd6_rem_ifa_lle(ifatoia6(ifa), 1);
1099                 break;
1100 #endif
1101         }
1102 }
1103
1104 int
1105 carp_master(struct ifaddr *ifa)
1106 {
1107         struct carp_softc *sc = ifa->ifa_carp;
1108
1109         return (sc->sc_state == MASTER);
1110 }
1111
1112 #ifdef INET
1113 /*
1114  * Broadcast a gratuitous ARP request containing
1115  * the virtual router MAC address for each IP address
1116  * associated with the virtual router.
1117  */
1118 static void
1119 carp_send_arp(struct carp_softc *sc)
1120 {
1121         struct ifaddr *ifa;
1122         struct in_addr addr;
1123
1124         CARP_FOREACH_IFA(sc, ifa) {
1125                 if (ifa->ifa_addr->sa_family != AF_INET)
1126                         continue;
1127                 addr = ((struct sockaddr_in *)ifa->ifa_addr)->sin_addr;
1128                 arp_announce_ifaddr(sc->sc_carpdev, addr, LLADDR(&sc->sc_addr));
1129         }
1130 }
1131
1132 int
1133 carp_iamatch(struct ifaddr *ifa, uint8_t **enaddr)
1134 {
1135         struct carp_softc *sc = ifa->ifa_carp;
1136
1137         if (sc->sc_state == MASTER) {
1138                 *enaddr = LLADDR(&sc->sc_addr);
1139                 return (1);
1140         }
1141
1142         return (0);
1143 }
1144 #endif
1145
1146 #ifdef INET6
1147 static void
1148 carp_send_na(struct carp_softc *sc)
1149 {
1150         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1151         struct ifaddr *ifa;
1152         struct in6_addr *in6;
1153
1154         CARP_FOREACH_IFA(sc, ifa) {
1155                 if (ifa->ifa_addr->sa_family != AF_INET6)
1156                         continue;
1157
1158                 in6 = IFA_IN6(ifa);
1159                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1160                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1161                 DELAY(1000);    /* XXX */
1162         }
1163 }
1164
1165 /*
1166  * Returns ifa in case it's a carp address and it is MASTER, or if the address
1167  * matches and is not a carp address.  Returns NULL otherwise.
1168  */
1169 struct ifaddr *
1170 carp_iamatch6(struct ifnet *ifp, struct in6_addr *taddr)
1171 {
1172         struct epoch_tracker et;
1173         struct ifaddr *ifa;
1174
1175         ifa = NULL;
1176         NET_EPOCH_ENTER(et);
1177         CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
1178                 if (ifa->ifa_addr->sa_family != AF_INET6)
1179                         continue;
1180                 if (!IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa)))
1181                         continue;
1182                 if (ifa->ifa_carp && ifa->ifa_carp->sc_state != MASTER)
1183                         ifa = NULL;
1184                 else
1185                         ifa_ref(ifa);
1186                 break;
1187         }
1188         NET_EPOCH_EXIT(et);
1189
1190         return (ifa);
1191 }
1192
1193 caddr_t
1194 carp_macmatch6(struct ifnet *ifp, struct mbuf *m, const struct in6_addr *taddr)
1195 {
1196         struct epoch_tracker et;
1197         struct ifaddr *ifa;
1198
1199         NET_EPOCH_ENTER(et);
1200         IFNET_FOREACH_IFA(ifp, ifa)
1201                 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1202                     IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1203                         struct carp_softc *sc = ifa->ifa_carp;
1204                         struct m_tag *mtag;
1205
1206                         NET_EPOCH_EXIT(et);
1207
1208                         mtag = m_tag_get(PACKET_TAG_CARP,
1209                             sizeof(struct carp_softc *), M_NOWAIT);
1210                         if (mtag == NULL)
1211                                 /* Better a bit than nothing. */
1212                                 return (LLADDR(&sc->sc_addr));
1213
1214                         bcopy(&sc, mtag + 1, sizeof(sc));
1215                         m_tag_prepend(m, mtag);
1216
1217                         return (LLADDR(&sc->sc_addr));
1218                 }
1219         NET_EPOCH_EXIT(et);
1220
1221         return (NULL);
1222 }
1223 #endif /* INET6 */
1224
1225 int
1226 carp_forus(struct ifnet *ifp, u_char *dhost)
1227 {
1228         struct carp_softc *sc;
1229         uint8_t *ena = dhost;
1230
1231         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1232                 return (0);
1233
1234         CIF_LOCK(ifp->if_carp);
1235         IFNET_FOREACH_CARP(ifp, sc) {
1236                 CARP_LOCK(sc);
1237                 if (sc->sc_state == MASTER && !bcmp(dhost, LLADDR(&sc->sc_addr),
1238                     ETHER_ADDR_LEN)) {
1239                         CARP_UNLOCK(sc);
1240                         CIF_UNLOCK(ifp->if_carp);
1241                         return (1);
1242                 }
1243                 CARP_UNLOCK(sc);
1244         }
1245         CIF_UNLOCK(ifp->if_carp);
1246
1247         return (0);
1248 }
1249
1250 /* Master down timeout event, executed in callout context. */
1251 static void
1252 carp_master_down(void *v)
1253 {
1254         struct carp_softc *sc = v;
1255
1256         CARP_LOCK_ASSERT(sc);
1257
1258         CURVNET_SET(sc->sc_carpdev->if_vnet);
1259         if (sc->sc_state == BACKUP) {
1260                 carp_master_down_locked(sc, "master timed out");
1261         }
1262         CURVNET_RESTORE();
1263
1264         CARP_UNLOCK(sc);
1265 }
1266
1267 static void
1268 carp_master_down_locked(struct carp_softc *sc, const char *reason)
1269 {
1270
1271         CARP_LOCK_ASSERT(sc);
1272
1273         switch (sc->sc_state) {
1274         case BACKUP:
1275                 carp_set_state(sc, MASTER, reason);
1276                 carp_send_ad_locked(sc);
1277 #ifdef INET
1278                 carp_send_arp(sc);
1279 #endif
1280 #ifdef INET6
1281                 carp_send_na(sc);
1282 #endif
1283                 carp_setrun(sc, 0);
1284                 carp_addroute(sc);
1285                 break;
1286         case INIT:
1287         case MASTER:
1288 #ifdef INVARIANTS
1289                 panic("carp: VHID %u@%s: master_down event in %s state\n",
1290                     sc->sc_vhid,
1291                     sc->sc_carpdev->if_xname,
1292                     sc->sc_state ? "MASTER" : "INIT");
1293 #endif
1294                 break;
1295         }
1296 }
1297
1298 /*
1299  * When in backup state, af indicates whether to reset the master down timer
1300  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1301  */
1302 static void
1303 carp_setrun(struct carp_softc *sc, sa_family_t af)
1304 {
1305         struct timeval tv;
1306
1307         CARP_LOCK_ASSERT(sc);
1308
1309         if ((sc->sc_carpdev->if_flags & IFF_UP) == 0 ||
1310             sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1311             (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) ||
1312             !V_carp_allow)
1313                 return;
1314
1315         switch (sc->sc_state) {
1316         case INIT:
1317                 carp_set_state(sc, BACKUP, "initialization complete");
1318                 carp_setrun(sc, 0);
1319                 break;
1320         case BACKUP:
1321                 callout_stop(&sc->sc_ad_tmo);
1322                 tv.tv_sec = 3 * sc->sc_advbase;
1323                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1324                 switch (af) {
1325 #ifdef INET
1326                 case AF_INET:
1327                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1328                             carp_master_down, sc);
1329                         break;
1330 #endif
1331 #ifdef INET6
1332                 case AF_INET6:
1333                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1334                             carp_master_down, sc);
1335                         break;
1336 #endif
1337                 default:
1338 #ifdef INET
1339                         if (sc->sc_naddrs)
1340                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1341                                     carp_master_down, sc);
1342 #endif
1343 #ifdef INET6
1344                         if (sc->sc_naddrs6)
1345                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1346                                     carp_master_down, sc);
1347 #endif
1348                         break;
1349                 }
1350                 break;
1351         case MASTER:
1352                 tv.tv_sec = sc->sc_advbase;
1353                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1354                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1355                     carp_send_ad, sc);
1356                 break;
1357         }
1358 }
1359
1360 /*
1361  * Setup multicast structures.
1362  */
1363 static int
1364 carp_multicast_setup(struct carp_if *cif, sa_family_t sa)
1365 {
1366         struct ifnet *ifp = cif->cif_ifp;
1367         int error = 0;
1368
1369         switch (sa) {
1370 #ifdef INET
1371         case AF_INET:
1372             {
1373                 struct ip_moptions *imo = &cif->cif_imo;
1374                 struct in_mfilter *imf;
1375                 struct in_addr addr;
1376
1377                 if (ip_mfilter_first(&imo->imo_head) != NULL)
1378                         return (0);
1379
1380                 imf = ip_mfilter_alloc(M_WAITOK, 0, 0);
1381                 ip_mfilter_init(&imo->imo_head);
1382                 imo->imo_multicast_vif = -1;
1383
1384                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1385                 if ((error = in_joingroup(ifp, &addr, NULL,
1386                     &imf->imf_inm)) != 0) {
1387                         ip_mfilter_free(imf);
1388                         break;
1389                 }
1390
1391                 ip_mfilter_insert(&imo->imo_head, imf);
1392                 imo->imo_multicast_ifp = ifp;
1393                 imo->imo_multicast_ttl = CARP_DFLTTL;
1394                 imo->imo_multicast_loop = 0;
1395                 break;
1396            }
1397 #endif
1398 #ifdef INET6
1399         case AF_INET6:
1400             {
1401                 struct ip6_moptions *im6o = &cif->cif_im6o;
1402                 struct in6_mfilter *im6f[2];
1403                 struct in6_addr in6;
1404
1405                 if (ip6_mfilter_first(&im6o->im6o_head))
1406                         return (0);
1407
1408                 im6f[0] = ip6_mfilter_alloc(M_WAITOK, 0, 0);
1409                 im6f[1] = ip6_mfilter_alloc(M_WAITOK, 0, 0);
1410
1411                 ip6_mfilter_init(&im6o->im6o_head);
1412                 im6o->im6o_multicast_hlim = CARP_DFLTTL;
1413                 im6o->im6o_multicast_ifp = ifp;
1414
1415                 /* Join IPv6 CARP multicast group. */
1416                 bzero(&in6, sizeof(in6));
1417                 in6.s6_addr16[0] = htons(0xff02);
1418                 in6.s6_addr8[15] = 0x12;
1419                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1420                         ip6_mfilter_free(im6f[0]);
1421                         ip6_mfilter_free(im6f[1]);
1422                         break;
1423                 }
1424                 if ((error = in6_joingroup(ifp, &in6, NULL, &im6f[0]->im6f_in6m, 0)) != 0) {
1425                         ip6_mfilter_free(im6f[0]);
1426                         ip6_mfilter_free(im6f[1]);
1427                         break;
1428                 }
1429
1430                 /* Join solicited multicast address. */
1431                 bzero(&in6, sizeof(in6));
1432                 in6.s6_addr16[0] = htons(0xff02);
1433                 in6.s6_addr32[1] = 0;
1434                 in6.s6_addr32[2] = htonl(1);
1435                 in6.s6_addr32[3] = 0;
1436                 in6.s6_addr8[12] = 0xff;
1437
1438                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1439                         ip6_mfilter_free(im6f[0]);
1440                         ip6_mfilter_free(im6f[1]);
1441                         break;
1442                 }
1443
1444                 if ((error = in6_joingroup(ifp, &in6, NULL, &im6f[1]->im6f_in6m, 0)) != 0) {
1445                         in6_leavegroup(im6f[0]->im6f_in6m, NULL);
1446                         ip6_mfilter_free(im6f[0]);
1447                         ip6_mfilter_free(im6f[1]);
1448                         break;
1449                 }
1450                 ip6_mfilter_insert(&im6o->im6o_head, im6f[0]);
1451                 ip6_mfilter_insert(&im6o->im6o_head, im6f[1]);
1452                 break;
1453             }
1454 #endif
1455         }
1456
1457         return (error);
1458 }
1459
1460 /*
1461  * Free multicast structures.
1462  */
1463 static void
1464 carp_multicast_cleanup(struct carp_if *cif, sa_family_t sa)
1465 {
1466 #ifdef INET
1467         struct ip_moptions *imo = &cif->cif_imo;
1468         struct in_mfilter *imf;
1469 #endif
1470 #ifdef INET6
1471         struct ip6_moptions *im6o = &cif->cif_im6o;
1472         struct in6_mfilter *im6f;
1473 #endif
1474         sx_assert(&carp_sx, SA_XLOCKED);
1475
1476         switch (sa) {
1477 #ifdef INET
1478         case AF_INET:
1479                 if (cif->cif_naddrs != 0)
1480                         break;
1481
1482                 while ((imf = ip_mfilter_first(&imo->imo_head)) != NULL) {
1483                         ip_mfilter_remove(&imo->imo_head, imf);
1484                         in_leavegroup(imf->imf_inm, NULL);
1485                         ip_mfilter_free(imf);
1486                 }
1487                 break;
1488 #endif
1489 #ifdef INET6
1490         case AF_INET6:
1491                 if (cif->cif_naddrs6 != 0)
1492                         break;
1493
1494                 while ((im6f = ip6_mfilter_first(&im6o->im6o_head)) != NULL) {
1495                         ip6_mfilter_remove(&im6o->im6o_head, im6f);
1496                         in6_leavegroup(im6f->im6f_in6m, NULL);
1497                         ip6_mfilter_free(im6f);
1498                 }
1499                 break;
1500 #endif
1501         }
1502 }
1503
1504 int
1505 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa)
1506 {
1507         struct m_tag *mtag;
1508         struct carp_softc *sc;
1509
1510         if (!sa)
1511                 return (0);
1512
1513         switch (sa->sa_family) {
1514 #ifdef INET
1515         case AF_INET:
1516                 break;
1517 #endif
1518 #ifdef INET6
1519         case AF_INET6:
1520                 break;
1521 #endif
1522         default:
1523                 return (0);
1524         }
1525
1526         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
1527         if (mtag == NULL)
1528                 return (0);
1529
1530         bcopy(mtag + 1, &sc, sizeof(sc));
1531
1532         /* Set the source MAC address to the Virtual Router MAC Address. */
1533         switch (ifp->if_type) {
1534         case IFT_ETHER:
1535         case IFT_BRIDGE:
1536         case IFT_L2VLAN: {
1537                         struct ether_header *eh;
1538
1539                         eh = mtod(m, struct ether_header *);
1540                         eh->ether_shost[0] = 0;
1541                         eh->ether_shost[1] = 0;
1542                         eh->ether_shost[2] = 0x5e;
1543                         eh->ether_shost[3] = 0;
1544                         eh->ether_shost[4] = 1;
1545                         eh->ether_shost[5] = sc->sc_vhid;
1546                 }
1547                 break;
1548         default:
1549                 printf("%s: carp is not supported for the %d interface type\n",
1550                     ifp->if_xname, ifp->if_type);
1551                 return (EOPNOTSUPP);
1552         }
1553
1554         return (0);
1555 }
1556
1557 static struct carp_softc*
1558 carp_alloc(struct ifnet *ifp)
1559 {
1560         struct carp_softc *sc;
1561         struct carp_if *cif;
1562
1563         sx_assert(&carp_sx, SA_XLOCKED);
1564
1565         if ((cif = ifp->if_carp) == NULL)
1566                 cif = carp_alloc_if(ifp);
1567
1568         sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
1569
1570         sc->sc_advbase = CARP_DFLTINTV;
1571         sc->sc_vhid = -1;       /* required setting */
1572         sc->sc_init_counter = 1;
1573         sc->sc_state = INIT;
1574
1575         sc->sc_ifasiz = sizeof(struct ifaddr *);
1576         sc->sc_ifas = malloc(sc->sc_ifasiz, M_CARP, M_WAITOK|M_ZERO);
1577         sc->sc_carpdev = ifp;
1578
1579         CARP_LOCK_INIT(sc);
1580 #ifdef INET
1581         callout_init_mtx(&sc->sc_md_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1582 #endif
1583 #ifdef INET6
1584         callout_init_mtx(&sc->sc_md6_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1585 #endif
1586         callout_init_mtx(&sc->sc_ad_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1587
1588         CIF_LOCK(cif);
1589         TAILQ_INSERT_TAIL(&cif->cif_vrs, sc, sc_list);
1590         CIF_UNLOCK(cif);
1591
1592         mtx_lock(&carp_mtx);
1593         LIST_INSERT_HEAD(&carp_list, sc, sc_next);
1594         mtx_unlock(&carp_mtx);
1595
1596         return (sc);
1597 }
1598
1599 static void
1600 carp_grow_ifas(struct carp_softc *sc)
1601 {
1602         struct ifaddr **new;
1603
1604         new = malloc(sc->sc_ifasiz * 2, M_CARP, M_WAITOK | M_ZERO);
1605         CARP_LOCK(sc);
1606         bcopy(sc->sc_ifas, new, sc->sc_ifasiz);
1607         free(sc->sc_ifas, M_CARP);
1608         sc->sc_ifas = new;
1609         sc->sc_ifasiz *= 2;
1610         CARP_UNLOCK(sc);
1611 }
1612
1613 static void
1614 carp_destroy(struct carp_softc *sc)
1615 {
1616         struct ifnet *ifp = sc->sc_carpdev;
1617         struct carp_if *cif = ifp->if_carp;
1618
1619         sx_assert(&carp_sx, SA_XLOCKED);
1620
1621         if (sc->sc_suppress)
1622                 carp_demote_adj(-V_carp_ifdown_adj, "vhid removed");
1623         CARP_UNLOCK(sc);
1624
1625         CIF_LOCK(cif);
1626         TAILQ_REMOVE(&cif->cif_vrs, sc, sc_list);
1627         CIF_UNLOCK(cif);
1628
1629         mtx_lock(&carp_mtx);
1630         LIST_REMOVE(sc, sc_next);
1631         mtx_unlock(&carp_mtx);
1632
1633         callout_drain(&sc->sc_ad_tmo);
1634 #ifdef INET
1635         callout_drain(&sc->sc_md_tmo);
1636 #endif
1637 #ifdef INET6
1638         callout_drain(&sc->sc_md6_tmo);
1639 #endif
1640         CARP_LOCK_DESTROY(sc);
1641
1642         free(sc->sc_ifas, M_CARP);
1643         free(sc, M_CARP);
1644 }
1645
1646 static struct carp_if*
1647 carp_alloc_if(struct ifnet *ifp)
1648 {
1649         struct carp_if *cif;
1650         int error;
1651
1652         cif = malloc(sizeof(*cif), M_CARP, M_WAITOK|M_ZERO);
1653
1654         if ((error = ifpromisc(ifp, 1)) != 0)
1655                 printf("%s: ifpromisc(%s) failed: %d\n",
1656                     __func__, ifp->if_xname, error);
1657         else
1658                 cif->cif_flags |= CIF_PROMISC;
1659
1660         CIF_LOCK_INIT(cif);
1661         cif->cif_ifp = ifp;
1662         TAILQ_INIT(&cif->cif_vrs);
1663
1664         IF_ADDR_WLOCK(ifp);
1665         ifp->if_carp = cif;
1666         if_ref(ifp);
1667         IF_ADDR_WUNLOCK(ifp);
1668
1669         return (cif);
1670 }
1671
1672 static void
1673 carp_free_if(struct carp_if *cif)
1674 {
1675         struct ifnet *ifp = cif->cif_ifp;
1676
1677         CIF_LOCK_ASSERT(cif);
1678         KASSERT(TAILQ_EMPTY(&cif->cif_vrs), ("%s: softc list not empty",
1679             __func__));
1680
1681         IF_ADDR_WLOCK(ifp);
1682         ifp->if_carp = NULL;
1683         IF_ADDR_WUNLOCK(ifp);
1684
1685         CIF_LOCK_DESTROY(cif);
1686
1687         if (cif->cif_flags & CIF_PROMISC)
1688                 ifpromisc(ifp, 0);
1689         if_rele(ifp);
1690
1691         free(cif, M_CARP);
1692 }
1693
1694 static void
1695 carp_carprcp(struct carpreq *carpr, struct carp_softc *sc, int priv)
1696 {
1697
1698         CARP_LOCK(sc);
1699         carpr->carpr_state = sc->sc_state;
1700         carpr->carpr_vhid = sc->sc_vhid;
1701         carpr->carpr_advbase = sc->sc_advbase;
1702         carpr->carpr_advskew = sc->sc_advskew;
1703         if (priv)
1704                 bcopy(sc->sc_key, carpr->carpr_key, sizeof(carpr->carpr_key));
1705         else
1706                 bzero(carpr->carpr_key, sizeof(carpr->carpr_key));
1707         CARP_UNLOCK(sc);
1708 }
1709
1710 int
1711 carp_ioctl(struct ifreq *ifr, u_long cmd, struct thread *td)
1712 {
1713         struct carpreq carpr;
1714         struct ifnet *ifp;
1715         struct carp_softc *sc = NULL;
1716         int error = 0, locked = 0;
1717
1718         if ((error = copyin(ifr_data_get_ptr(ifr), &carpr, sizeof carpr)))
1719                 return (error);
1720
1721         ifp = ifunit_ref(ifr->ifr_name);
1722         if (ifp == NULL)
1723                 return (ENXIO);
1724
1725         switch (ifp->if_type) {
1726         case IFT_ETHER:
1727         case IFT_L2VLAN:
1728         case IFT_BRIDGE:
1729                 break;
1730         default:
1731                 error = EOPNOTSUPP;
1732                 goto out;
1733         }
1734
1735         if ((ifp->if_flags & IFF_MULTICAST) == 0) {
1736                 error = EADDRNOTAVAIL;
1737                 goto out;
1738         }
1739
1740         sx_xlock(&carp_sx);
1741         switch (cmd) {
1742         case SIOCSVH:
1743                 if ((error = priv_check(td, PRIV_NETINET_CARP)))
1744                         break;
1745                 if (carpr.carpr_vhid <= 0 || carpr.carpr_vhid > CARP_MAXVHID ||
1746                     carpr.carpr_advbase < 0 || carpr.carpr_advskew < 0) {
1747                         error = EINVAL;
1748                         break;
1749                 }
1750
1751                 if (ifp->if_carp) {
1752                         IFNET_FOREACH_CARP(ifp, sc)
1753                                 if (sc->sc_vhid == carpr.carpr_vhid)
1754                                         break;
1755                 }
1756                 if (sc == NULL) {
1757                         sc = carp_alloc(ifp);
1758                         CARP_LOCK(sc);
1759                         sc->sc_vhid = carpr.carpr_vhid;
1760                         LLADDR(&sc->sc_addr)[0] = 0;
1761                         LLADDR(&sc->sc_addr)[1] = 0;
1762                         LLADDR(&sc->sc_addr)[2] = 0x5e;
1763                         LLADDR(&sc->sc_addr)[3] = 0;
1764                         LLADDR(&sc->sc_addr)[4] = 1;
1765                         LLADDR(&sc->sc_addr)[5] = sc->sc_vhid;
1766                 } else
1767                         CARP_LOCK(sc);
1768                 locked = 1;
1769                 if (carpr.carpr_advbase > 0) {
1770                         if (carpr.carpr_advbase > 255 ||
1771                             carpr.carpr_advbase < CARP_DFLTINTV) {
1772                                 error = EINVAL;
1773                                 break;
1774                         }
1775                         sc->sc_advbase = carpr.carpr_advbase;
1776                 }
1777                 if (carpr.carpr_advskew >= 255) {
1778                         error = EINVAL;
1779                         break;
1780                 }
1781                 sc->sc_advskew = carpr.carpr_advskew;
1782                 if (carpr.carpr_key[0] != '\0') {
1783                         bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1784                         carp_hmac_prepare(sc);
1785                 }
1786                 if (sc->sc_state != INIT &&
1787                     carpr.carpr_state != sc->sc_state) {
1788                         switch (carpr.carpr_state) {
1789                         case BACKUP:
1790                                 callout_stop(&sc->sc_ad_tmo);
1791                                 carp_set_state(sc, BACKUP,
1792                                     "user requested via ifconfig");
1793                                 carp_setrun(sc, 0);
1794                                 carp_delroute(sc);
1795                                 break;
1796                         case MASTER:
1797                                 carp_master_down_locked(sc,
1798                                     "user requested via ifconfig");
1799                                 break;
1800                         default:
1801                                 break;
1802                         }
1803                 }
1804                 break;
1805
1806         case SIOCGVH:
1807             {
1808                 int priveleged;
1809
1810                 if (carpr.carpr_vhid < 0 || carpr.carpr_vhid > CARP_MAXVHID) {
1811                         error = EINVAL;
1812                         break;
1813                 }
1814                 if (carpr.carpr_count < 1) {
1815                         error = EMSGSIZE;
1816                         break;
1817                 }
1818                 if (ifp->if_carp == NULL) {
1819                         error = ENOENT;
1820                         break;
1821                 }
1822
1823                 priveleged = (priv_check(td, PRIV_NETINET_CARP) == 0);
1824                 if (carpr.carpr_vhid != 0) {
1825                         IFNET_FOREACH_CARP(ifp, sc)
1826                                 if (sc->sc_vhid == carpr.carpr_vhid)
1827                                         break;
1828                         if (sc == NULL) {
1829                                 error = ENOENT;
1830                                 break;
1831                         }
1832                         carp_carprcp(&carpr, sc, priveleged);
1833                         error = copyout(&carpr, ifr_data_get_ptr(ifr),
1834                             sizeof(carpr));
1835                 } else  {
1836                         int i, count;
1837
1838                         count = 0;
1839                         IFNET_FOREACH_CARP(ifp, sc)
1840                                 count++;
1841
1842                         if (count > carpr.carpr_count) {
1843                                 CIF_UNLOCK(ifp->if_carp);
1844                                 error = EMSGSIZE;
1845                                 break;
1846                         }
1847
1848                         i = 0;
1849                         IFNET_FOREACH_CARP(ifp, sc) {
1850                                 carp_carprcp(&carpr, sc, priveleged);
1851                                 carpr.carpr_count = count;
1852                                 error = copyout(&carpr,
1853                                     (caddr_t)ifr_data_get_ptr(ifr) +
1854                                     (i * sizeof(carpr)), sizeof(carpr));
1855                                 if (error) {
1856                                         CIF_UNLOCK(ifp->if_carp);
1857                                         break;
1858                                 }
1859                                 i++;
1860                         }
1861                 }
1862                 break;
1863             }
1864         default:
1865                 error = EINVAL;
1866         }
1867         sx_xunlock(&carp_sx);
1868
1869 out:
1870         if (locked)
1871                 CARP_UNLOCK(sc);
1872         if_rele(ifp);
1873
1874         return (error);
1875 }
1876
1877 static int
1878 carp_get_vhid(struct ifaddr *ifa)
1879 {
1880
1881         if (ifa == NULL || ifa->ifa_carp == NULL)
1882                 return (0);
1883
1884         return (ifa->ifa_carp->sc_vhid);
1885 }
1886
1887 int
1888 carp_attach(struct ifaddr *ifa, int vhid)
1889 {
1890         struct ifnet *ifp = ifa->ifa_ifp;
1891         struct carp_if *cif = ifp->if_carp;
1892         struct carp_softc *sc;
1893         int index, error;
1894
1895         KASSERT(ifa->ifa_carp == NULL, ("%s: ifa %p attached", __func__, ifa));
1896
1897         switch (ifa->ifa_addr->sa_family) {
1898 #ifdef INET
1899         case AF_INET:
1900 #endif
1901 #ifdef INET6
1902         case AF_INET6:
1903 #endif
1904                 break;
1905         default:
1906                 return (EPROTOTYPE);
1907         }
1908
1909         sx_xlock(&carp_sx);
1910         if (ifp->if_carp == NULL) {
1911                 sx_xunlock(&carp_sx);
1912                 return (ENOPROTOOPT);
1913         }
1914
1915         IFNET_FOREACH_CARP(ifp, sc)
1916                 if (sc->sc_vhid == vhid)
1917                         break;
1918         if (sc == NULL) {
1919                 sx_xunlock(&carp_sx);
1920                 return (ENOENT);
1921         }
1922
1923         error = carp_multicast_setup(cif, ifa->ifa_addr->sa_family);
1924         if (error) {
1925                 CIF_FREE(cif);
1926                 sx_xunlock(&carp_sx);
1927                 return (error);
1928         }
1929
1930         index = sc->sc_naddrs + sc->sc_naddrs6 + 1;
1931         if (index > sc->sc_ifasiz / sizeof(struct ifaddr *))
1932                 carp_grow_ifas(sc);
1933
1934         switch (ifa->ifa_addr->sa_family) {
1935 #ifdef INET
1936         case AF_INET:
1937                 cif->cif_naddrs++;
1938                 sc->sc_naddrs++;
1939                 break;
1940 #endif
1941 #ifdef INET6
1942         case AF_INET6:
1943                 cif->cif_naddrs6++;
1944                 sc->sc_naddrs6++;
1945                 break;
1946 #endif
1947         }
1948
1949         ifa_ref(ifa);
1950
1951         CARP_LOCK(sc);
1952         sc->sc_ifas[index - 1] = ifa;
1953         ifa->ifa_carp = sc;
1954         carp_hmac_prepare(sc);
1955         carp_sc_state(sc);
1956         CARP_UNLOCK(sc);
1957
1958         sx_xunlock(&carp_sx);
1959
1960         return (0);
1961 }
1962
1963 void
1964 carp_detach(struct ifaddr *ifa, bool keep_cif)
1965 {
1966         struct ifnet *ifp = ifa->ifa_ifp;
1967         struct carp_if *cif = ifp->if_carp;
1968         struct carp_softc *sc = ifa->ifa_carp;
1969         int i, index;
1970
1971         KASSERT(sc != NULL, ("%s: %p not attached", __func__, ifa));
1972
1973         sx_xlock(&carp_sx);
1974
1975         CARP_LOCK(sc);
1976         /* Shift array. */
1977         index = sc->sc_naddrs + sc->sc_naddrs6;
1978         for (i = 0; i < index; i++)
1979                 if (sc->sc_ifas[i] == ifa)
1980                         break;
1981         KASSERT(i < index, ("%s: %p no backref", __func__, ifa));
1982         for (; i < index - 1; i++)
1983                 sc->sc_ifas[i] = sc->sc_ifas[i+1];
1984         sc->sc_ifas[index - 1] = NULL;
1985
1986         switch (ifa->ifa_addr->sa_family) {
1987 #ifdef INET
1988         case AF_INET:
1989                 cif->cif_naddrs--;
1990                 sc->sc_naddrs--;
1991                 break;
1992 #endif
1993 #ifdef INET6
1994         case AF_INET6:
1995                 cif->cif_naddrs6--;
1996                 sc->sc_naddrs6--;
1997                 break;
1998 #endif
1999         }
2000
2001         carp_ifa_delroute(ifa);
2002         carp_multicast_cleanup(cif, ifa->ifa_addr->sa_family);
2003
2004         ifa->ifa_carp = NULL;
2005         ifa_free(ifa);
2006
2007         carp_hmac_prepare(sc);
2008         carp_sc_state(sc);
2009
2010         if (!keep_cif && sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0)
2011                 carp_destroy(sc);
2012         else
2013                 CARP_UNLOCK(sc);
2014
2015         if (!keep_cif)
2016                 CIF_FREE(cif);
2017
2018         sx_xunlock(&carp_sx);
2019 }
2020
2021 static void
2022 carp_set_state(struct carp_softc *sc, int state, const char *reason)
2023 {
2024
2025         CARP_LOCK_ASSERT(sc);
2026
2027         if (sc->sc_state != state) {
2028                 const char *carp_states[] = { CARP_STATES };
2029                 char subsys[IFNAMSIZ+5];
2030
2031                 snprintf(subsys, IFNAMSIZ+5, "%u@%s", sc->sc_vhid,
2032                     sc->sc_carpdev->if_xname);
2033
2034                 CARP_LOG("%s: %s -> %s (%s)\n", subsys,
2035                     carp_states[sc->sc_state], carp_states[state], reason);
2036
2037                 sc->sc_state = state;
2038
2039                 devctl_notify("CARP", subsys, carp_states[state], NULL);
2040         }
2041 }
2042
2043 static void
2044 carp_linkstate(struct ifnet *ifp)
2045 {
2046         struct carp_softc *sc;
2047
2048         CIF_LOCK(ifp->if_carp);
2049         IFNET_FOREACH_CARP(ifp, sc) {
2050                 CARP_LOCK(sc);
2051                 carp_sc_state(sc);
2052                 CARP_UNLOCK(sc);
2053         }
2054         CIF_UNLOCK(ifp->if_carp);
2055 }
2056
2057 static void
2058 carp_sc_state(struct carp_softc *sc)
2059 {
2060
2061         CARP_LOCK_ASSERT(sc);
2062
2063         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
2064             !(sc->sc_carpdev->if_flags & IFF_UP) ||
2065             !V_carp_allow) {
2066                 callout_stop(&sc->sc_ad_tmo);
2067 #ifdef INET
2068                 callout_stop(&sc->sc_md_tmo);
2069 #endif
2070 #ifdef INET6
2071                 callout_stop(&sc->sc_md6_tmo);
2072 #endif
2073                 carp_set_state(sc, INIT, "hardware interface down");
2074                 carp_setrun(sc, 0);
2075                 if (!sc->sc_suppress)
2076                         carp_demote_adj(V_carp_ifdown_adj, "interface down");
2077                 sc->sc_suppress = 1;
2078         } else {
2079                 carp_set_state(sc, INIT, "hardware interface up");
2080                 carp_setrun(sc, 0);
2081                 if (sc->sc_suppress)
2082                         carp_demote_adj(-V_carp_ifdown_adj, "interface up");
2083                 sc->sc_suppress = 0;
2084         }
2085 }
2086
2087 static void
2088 carp_demote_adj(int adj, char *reason)
2089 {
2090         atomic_add_int(&V_carp_demotion, adj);
2091         CARP_LOG("demoted by %d to %d (%s)\n", adj, V_carp_demotion, reason);
2092         taskqueue_enqueue(taskqueue_swi, &carp_sendall_task);
2093 }
2094
2095 static int
2096 carp_allow_sysctl(SYSCTL_HANDLER_ARGS)
2097 {
2098         int new, error;
2099         struct carp_softc *sc;
2100
2101         new = V_carp_allow;
2102         error = sysctl_handle_int(oidp, &new, 0, req);
2103         if (error || !req->newptr)
2104                 return (error);
2105
2106         if (V_carp_allow != new) {
2107                 V_carp_allow = new;
2108
2109                 mtx_lock(&carp_mtx);
2110                 LIST_FOREACH(sc, &carp_list, sc_next) {
2111                         CARP_LOCK(sc);
2112                         if (curvnet == sc->sc_carpdev->if_vnet)
2113                                 carp_sc_state(sc);
2114                         CARP_UNLOCK(sc);
2115                 }
2116                 mtx_unlock(&carp_mtx);
2117         }
2118
2119         return (0);
2120 }
2121
2122 static int
2123 carp_dscp_sysctl(SYSCTL_HANDLER_ARGS)
2124 {
2125         int new, error;
2126
2127         new = V_carp_dscp;
2128         error = sysctl_handle_int(oidp, &new, 0, req);
2129         if (error || !req->newptr)
2130                 return (error);
2131
2132         if (new < 0 || new > 63)
2133                 return (EINVAL);
2134
2135         V_carp_dscp = new;
2136
2137         return (0);
2138 }
2139
2140 static int
2141 carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS)
2142 {
2143         int new, error;
2144
2145         new = V_carp_demotion;
2146         error = sysctl_handle_int(oidp, &new, 0, req);
2147         if (error || !req->newptr)
2148                 return (error);
2149
2150         carp_demote_adj(new, "sysctl");
2151
2152         return (0);
2153 }
2154
2155 #ifdef INET
2156 extern  struct domain inetdomain;
2157 static struct protosw in_carp_protosw = {
2158         .pr_type =              SOCK_RAW,
2159         .pr_domain =            &inetdomain,
2160         .pr_protocol =          IPPROTO_CARP,
2161         .pr_flags =             PR_ATOMIC|PR_ADDR,
2162         .pr_input =             carp_input,
2163         .pr_output =            rip_output,
2164         .pr_ctloutput =         rip_ctloutput,
2165         .pr_usrreqs =           &rip_usrreqs
2166 };
2167 #endif
2168
2169 #ifdef INET6
2170 extern  struct domain inet6domain;
2171 static struct protosw in6_carp_protosw = {
2172         .pr_type =              SOCK_RAW,
2173         .pr_domain =            &inet6domain,
2174         .pr_protocol =          IPPROTO_CARP,
2175         .pr_flags =             PR_ATOMIC|PR_ADDR,
2176         .pr_input =             carp6_input,
2177         .pr_output =            rip6_output,
2178         .pr_ctloutput =         rip6_ctloutput,
2179         .pr_usrreqs =           &rip6_usrreqs
2180 };
2181 #endif
2182
2183 static void
2184 carp_mod_cleanup(void)
2185 {
2186
2187 #ifdef INET
2188         if (proto_reg[CARP_INET] == 0) {
2189                 (void)ipproto_unregister(IPPROTO_CARP);
2190                 pf_proto_unregister(PF_INET, IPPROTO_CARP, SOCK_RAW);
2191                 proto_reg[CARP_INET] = -1;
2192         }
2193         carp_iamatch_p = NULL;
2194 #endif
2195 #ifdef INET6
2196         if (proto_reg[CARP_INET6] == 0) {
2197                 (void)ip6proto_unregister(IPPROTO_CARP);
2198                 pf_proto_unregister(PF_INET6, IPPROTO_CARP, SOCK_RAW);
2199                 proto_reg[CARP_INET6] = -1;
2200         }
2201         carp_iamatch6_p = NULL;
2202         carp_macmatch6_p = NULL;
2203 #endif
2204         carp_ioctl_p = NULL;
2205         carp_attach_p = NULL;
2206         carp_detach_p = NULL;
2207         carp_get_vhid_p = NULL;
2208         carp_linkstate_p = NULL;
2209         carp_forus_p = NULL;
2210         carp_output_p = NULL;
2211         carp_demote_adj_p = NULL;
2212         carp_master_p = NULL;
2213         mtx_unlock(&carp_mtx);
2214         taskqueue_drain(taskqueue_swi, &carp_sendall_task);
2215         mtx_destroy(&carp_mtx);
2216         sx_destroy(&carp_sx);
2217 }
2218
2219 static int
2220 carp_mod_load(void)
2221 {
2222         int err;
2223
2224         mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2225         sx_init(&carp_sx, "carp_sx");
2226         LIST_INIT(&carp_list);
2227         carp_get_vhid_p = carp_get_vhid;
2228         carp_forus_p = carp_forus;
2229         carp_output_p = carp_output;
2230         carp_linkstate_p = carp_linkstate;
2231         carp_ioctl_p = carp_ioctl;
2232         carp_attach_p = carp_attach;
2233         carp_detach_p = carp_detach;
2234         carp_demote_adj_p = carp_demote_adj;
2235         carp_master_p = carp_master;
2236 #ifdef INET6
2237         carp_iamatch6_p = carp_iamatch6;
2238         carp_macmatch6_p = carp_macmatch6;
2239         proto_reg[CARP_INET6] = pf_proto_register(PF_INET6,
2240             (struct protosw *)&in6_carp_protosw);
2241         if (proto_reg[CARP_INET6]) {
2242                 printf("carp: error %d attaching to PF_INET6\n",
2243                     proto_reg[CARP_INET6]);
2244                 carp_mod_cleanup();
2245                 return (proto_reg[CARP_INET6]);
2246         }
2247         err = ip6proto_register(IPPROTO_CARP);
2248         if (err) {
2249                 printf("carp: error %d registering with INET6\n", err);
2250                 carp_mod_cleanup();
2251                 return (err);
2252         }
2253 #endif
2254 #ifdef INET
2255         carp_iamatch_p = carp_iamatch;
2256         proto_reg[CARP_INET] = pf_proto_register(PF_INET, &in_carp_protosw);
2257         if (proto_reg[CARP_INET]) {
2258                 printf("carp: error %d attaching to PF_INET\n",
2259                     proto_reg[CARP_INET]);
2260                 carp_mod_cleanup();
2261                 return (proto_reg[CARP_INET]);
2262         }
2263         err = ipproto_register(IPPROTO_CARP);
2264         if (err) {
2265                 printf("carp: error %d registering with INET\n", err);
2266                 carp_mod_cleanup();
2267                 return (err);
2268         }
2269 #endif
2270         return (0);
2271 }
2272
2273 static int
2274 carp_modevent(module_t mod, int type, void *data)
2275 {
2276         switch (type) {
2277         case MOD_LOAD:
2278                 return carp_mod_load();
2279                 /* NOTREACHED */
2280         case MOD_UNLOAD:
2281                 mtx_lock(&carp_mtx);
2282                 if (LIST_EMPTY(&carp_list))
2283                         carp_mod_cleanup();
2284                 else {
2285                         mtx_unlock(&carp_mtx);
2286                         return (EBUSY);
2287                 }
2288                 break;
2289
2290         default:
2291                 return (EINVAL);
2292         }
2293
2294         return (0);
2295 }
2296
2297 static moduledata_t carp_mod = {
2298         "carp",
2299         carp_modevent,
2300         0
2301 };
2302
2303 DECLARE_MODULE(carp, carp_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);