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         int error;
 648
 649         /*
 650          * Verify that the VHID is valid on the receiving interface.
 651          *
 652          * There should be just one match.  If there are none
 653          * the VHID is not valid and we drop the packet.  If
 654          * there are multiple VHID matches, take just the first
 655          * one, for compatibility with previous code.  While we're
 656          * scanning, check for obvious loops in the network topology
 657          * (these should never happen, and as noted above, we may
 658          * miss real loops; this is just a double-check).
 659          */
 660         IF_ADDR_RLOCK(ifp);
 661         error = 0;
 662         match = NULL;
 663         IFNET_FOREACH_IFA(ifp, ifa) {
 664                 if (match == NULL && ifa->ifa_carp != NULL &&
 665                     ifa->ifa_addr->sa_family == af &&
 666                     ifa->ifa_carp->sc_vhid == ch->carp_vhid)
 667                         match = ifa;
 668                 if (ch->carp_vhid == 0 && carp_source_is_self(m, ifa, af))
 669                         error = ELOOP;
 670         }
 671         ifa = error ? NULL : match;
 672         if (ifa != NULL)
 673                 ifa_ref(ifa);
 674         IF_ADDR_RUNLOCK(ifp);
 675
 676         if (ifa == NULL) {
 677                 if (error == ELOOP) {
 678                         CARP_DEBUG("dropping looped packet on interface %s\n",
 679                             ifp->if_xname);
 680                         CARPSTATS_INC(carps_badif);     /* ??? */
 681                 } else {
 682                         CARPSTATS_INC(carps_badvhid);
 683                 }
 684                 m_freem(m);
 685                 return;
 686         }
 687
 688         /* verify the CARP version. */
 689         if (ch->carp_version != CARP_VERSION) {
 690                 CARPSTATS_INC(carps_badver);
 691                 CARP_DEBUG("%s: invalid version %d\n", ifp->if_xname,
 692                     ch->carp_version);
 693                 ifa_free(ifa);
 694                 m_freem(m);
 695                 return;
 696         }
 697
 698         sc = ifa->ifa_carp;
 699         CARP_LOCK(sc);
 700         ifa_free(ifa);
 701
 702         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 703                 CARPSTATS_INC(carps_badauth);
 704                 CARP_DEBUG("%s: incorrect hash for VHID %u@%s\n", __func__,
 705                     sc->sc_vhid, ifp->if_xname);
 706                 goto out;
 707         }
 708
 709         tmp_counter = ntohl(ch->carp_counter[0]);
 710         tmp_counter = tmp_counter<<32;
 711         tmp_counter += ntohl(ch->carp_counter[1]);
 712
 713         /* XXX Replay protection goes here */
 714
 715         sc->sc_init_counter = 0;
 716         sc->sc_counter = tmp_counter;
 717
 718         sc_tv.tv_sec = sc->sc_advbase;
 719         sc_tv.tv_usec = DEMOTE_ADVSKEW(sc) * 1000000 / 256;
 720         ch_tv.tv_sec = ch->carp_advbase;
 721         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 722
 723         switch (sc->sc_state) {
 724         case INIT:
 725                 break;
 726         case MASTER:
 727                 /*
 728                  * If we receive an advertisement from a master who's going to
 729                  * be more frequent than us, go into BACKUP state.
 730                  */
 731                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 732                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 733                         callout_stop(&sc->sc_ad_tmo);
 734                         carp_set_state(sc, BACKUP,
 735                             "more frequent advertisement received");
 736                         carp_setrun(sc, 0);
 737                         carp_delroute(sc);
 738                 }
 739                 break;
 740         case BACKUP:
 741                 /*
 742                  * If we're pre-empting masters who advertise slower than us,
 743                  * and this one claims to be slower, treat him as down.
 744                  */
 745                 if (V_carp_preempt && timevalcmp(&sc_tv, &ch_tv, <)) {
 746                         carp_master_down_locked(sc,
 747                             "preempting a slower master");
 748                         break;
 749                 }
 750
 751                 /*
 752                  *  If the master is going to advertise at such a low frequency
 753                  *  that he's guaranteed to time out, we'd might as well just
 754                  *  treat him as timed out now.
 755                  */
 756                 sc_tv.tv_sec = sc->sc_advbase * 3;
 757                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 758                         carp_master_down_locked(sc, "master will time out");
 759                         break;
 760                 }
 761
 762                 /*
 763                  * Otherwise, we reset the counter and wait for the next
 764                  * advertisement.
 765                  */
 766                 carp_setrun(sc, af);
 767                 break;
 768         }
 769
 770 out:
 771         CARP_UNLOCK(sc);
 772         m_freem(m);
 773 }
 774
 775 static int
 776 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 777 {
 778         struct m_tag *mtag;
 779
 780         if (sc->sc_init_counter) {
 781                 /* this could also be seconds since unix epoch */
 782                 sc->sc_counter = arc4random();
 783                 sc->sc_counter = sc->sc_counter << 32;
 784                 sc->sc_counter += arc4random();
 785         } else
 786                 sc->sc_counter++;
 787
 788         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 789         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 790
 791         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 792
 793         /* Tag packet for carp_output */
 794         if ((mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct carp_softc *),
 795             M_NOWAIT)) == NULL) {
 796                 m_freem(m);
 797                 CARPSTATS_INC(carps_onomem);
 798                 return (ENOMEM);
 799         }
 800         bcopy(&sc, mtag + 1, sizeof(sc));
 801         m_tag_prepend(m, mtag);
 802
 803         return (0);
 804 }
 805
 806 /*
 807  * To avoid LORs and possible recursions this function shouldn't
 808  * be called directly, but scheduled via taskqueue.
 809  */
 810 static void
 811 carp_send_ad_all(void *ctx __unused, int pending __unused)
 812 {
 813         struct carp_softc *sc;
 814
 815         mtx_lock(&carp_mtx);
 816         LIST_FOREACH(sc, &carp_list, sc_next)
 817                 if (sc->sc_state == MASTER) {
 818                         CARP_LOCK(sc);
 819                         CURVNET_SET(sc->sc_carpdev->if_vnet);
 820                         carp_send_ad_locked(sc);
 821                         CURVNET_RESTORE();
 822                         CARP_UNLOCK(sc);
 823                 }
 824         mtx_unlock(&carp_mtx);
 825 }
 826
 827 /* Send a periodic advertisement, executed in callout context. */
 828 static void
 829 carp_send_ad(void *v)
 830 {
 831         struct carp_softc *sc = v;
 832
 833         CARP_LOCK_ASSERT(sc);
 834         CURVNET_SET(sc->sc_carpdev->if_vnet);
 835         carp_send_ad_locked(sc);
 836         CURVNET_RESTORE();
 837         CARP_UNLOCK(sc);
 838 }
 839
 840 static void
 841 carp_send_ad_error(struct carp_softc *sc, int error)
 842 {
 843
 844         if (error) {
 845                 if (sc->sc_sendad_errors < INT_MAX)
 846                         sc->sc_sendad_errors++;
 847                 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 848                         static const char fmt[] = "send error %d on %s";
 849                         char msg[sizeof(fmt) + IFNAMSIZ];
 850
 851                         sprintf(msg, fmt, error, sc->sc_carpdev->if_xname);
 852                         carp_demote_adj(V_carp_senderr_adj, msg);
 853                 }
 854                 sc->sc_sendad_success = 0;
 855         } else {
 856                 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS &&
 857                     ++sc->sc_sendad_success >= CARP_SENDAD_MIN_SUCCESS) {
 858                         static const char fmt[] = "send ok on %s";
 859                         char msg[sizeof(fmt) + IFNAMSIZ];
 860
 861                         sprintf(msg, fmt, sc->sc_carpdev->if_xname);
 862                         carp_demote_adj(-V_carp_senderr_adj, msg);
 863                         sc->sc_sendad_errors = 0;
 864                 } else
 865                         sc->sc_sendad_errors = 0;
 866         }
 867 }
 868
 869 /*
 870  * Pick the best ifaddr on the given ifp for sending CARP
 871  * advertisements.
 872  *
 873  * "Best" here is defined by ifa_preferred().  This function is much
 874  * much like ifaof_ifpforaddr() except that we just use ifa_preferred().
 875  *
 876  * (This could be simplified to return the actual address, except that
 877  * it has a different format in AF_INET and AF_INET6.)
 878  */
 879 static struct ifaddr *
 880 carp_best_ifa(int af, struct ifnet *ifp)
 881 {
 882         struct ifaddr *ifa, *best;
 883
 884         if (af >= AF_MAX)
 885                 return (NULL);
 886         best = NULL;
 887         IF_ADDR_RLOCK(ifp);
 888         CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
 889                 if (ifa->ifa_addr->sa_family == af &&
 890                     (best == NULL || ifa_preferred(best, ifa)))
 891                         best = ifa;
 892         }
 893         IF_ADDR_RUNLOCK(ifp);
 894         if (best != NULL)
 895                 ifa_ref(best);
 896         return (best);
 897 }
 898
 899 static void
 900 carp_send_ad_locked(struct carp_softc *sc)
 901 {
 902         struct carp_header ch;
 903         struct timeval tv;
 904         struct ifaddr *ifa;
 905         struct carp_header *ch_ptr;
 906         struct mbuf *m;
 907         int len, advskew;
 908
 909         CARP_LOCK_ASSERT(sc);
 910
 911         advskew = DEMOTE_ADVSKEW(sc);
 912         tv.tv_sec = sc->sc_advbase;
 913         tv.tv_usec = advskew * 1000000 / 256;
 914
 915         ch.carp_version = CARP_VERSION;
 916         ch.carp_type = CARP_ADVERTISEMENT;
 917         ch.carp_vhid = sc->sc_vhid;
 918         ch.carp_advbase = sc->sc_advbase;
 919         ch.carp_advskew = advskew;
 920         ch.carp_authlen = 7;    /* XXX DEFINE */
 921         ch.carp_pad1 = 0;       /* must be zero */
 922         ch.carp_cksum = 0;
 923
 924         /* XXXGL: OpenBSD picks first ifaddr with needed family. */
 925
 926 #ifdef INET
 927         if (sc->sc_naddrs) {
 928                 struct ip *ip;
 929
 930                 m = m_gethdr(M_NOWAIT, MT_DATA);
 931                 if (m == NULL) {
 932                         CARPSTATS_INC(carps_onomem);
 933                         goto resched;
 934                 }
 935                 len = sizeof(*ip) + sizeof(ch);
 936                 m->m_pkthdr.len = len;
 937                 m->m_pkthdr.rcvif = NULL;
 938                 m->m_len = len;
 939                 M_ALIGN(m, m->m_len);
 940                 m->m_flags |= M_MCAST;
 941                 ip = mtod(m, struct ip *);
 942                 ip->ip_v = IPVERSION;
 943                 ip->ip_hl = sizeof(*ip) >> 2;
 944                 ip->ip_tos = V_carp_dscp << IPTOS_DSCP_OFFSET;
 945                 ip->ip_len = htons(len);
 946                 ip->ip_off = htons(IP_DF);
 947                 ip->ip_ttl = CARP_DFLTTL;
 948                 ip->ip_p = IPPROTO_CARP;
 949                 ip->ip_sum = 0;
 950                 ip_fillid(ip);
 951
 952                 ifa = carp_best_ifa(AF_INET, sc->sc_carpdev);
 953                 if (ifa != NULL) {
 954                         ip->ip_src.s_addr =
 955                             ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 956                         ifa_free(ifa);
 957                 } else
 958                         ip->ip_src.s_addr = 0;
 959                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 960
 961                 ch_ptr = (struct carp_header *)(&ip[1]);
 962                 bcopy(&ch, ch_ptr, sizeof(ch));
 963                 if (carp_prepare_ad(m, sc, ch_ptr))
 964                         goto resched;
 965
 966                 m->m_data += sizeof(*ip);
 967                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip));
 968                 m->m_data -= sizeof(*ip);
 969
 970                 CARPSTATS_INC(carps_opackets);
 971
 972                 carp_send_ad_error(sc, ip_output(m, NULL, NULL, IP_RAWOUTPUT,
 973                     &sc->sc_carpdev->if_carp->cif_imo, NULL));
 974         }
 975 #endif /* INET */
 976 #ifdef INET6
 977         if (sc->sc_naddrs6) {
 978                 struct ip6_hdr *ip6;
 979
 980                 m = m_gethdr(M_NOWAIT, MT_DATA);
 981                 if (m == NULL) {
 982                         CARPSTATS_INC(carps_onomem);
 983                         goto resched;
 984                 }
 985                 len = sizeof(*ip6) + sizeof(ch);
 986                 m->m_pkthdr.len = len;
 987                 m->m_pkthdr.rcvif = NULL;
 988                 m->m_len = len;
 989                 M_ALIGN(m, m->m_len);
 990                 m->m_flags |= M_MCAST;
 991                 ip6 = mtod(m, struct ip6_hdr *);
 992                 bzero(ip6, sizeof(*ip6));
 993                 ip6->ip6_vfc |= IPV6_VERSION;
 994                 /* Traffic class isn't defined in ip6 struct instead
 995                  * it gets offset into flowid field */
 996                 ip6->ip6_flow |= htonl(V_carp_dscp << (IPV6_FLOWLABEL_LEN +
 997                     IPTOS_DSCP_OFFSET));
 998                 ip6->ip6_hlim = CARP_DFLTTL;
 999                 ip6->ip6_nxt = IPPROTO_CARP;
1000
1001                 /* set the source address */
1002                 ifa = carp_best_ifa(AF_INET6, sc->sc_carpdev);
1003                 if (ifa != NULL) {
1004                         bcopy(IFA_IN6(ifa), &ip6->ip6_src,
1005                             sizeof(struct in6_addr));
1006                         ifa_free(ifa);
1007                 } else
1008                         /* This should never happen with IPv6. */
1009                         bzero(&ip6->ip6_src, sizeof(struct in6_addr));
1010
1011                 /* Set the multicast destination. */
1012                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
1013                 ip6->ip6_dst.s6_addr8[15] = 0x12;
1014                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
1015                         m_freem(m);
1016                         CARP_DEBUG("%s: in6_setscope failed\n", __func__);
1017                         goto resched;
1018                 }
1019
1020                 ch_ptr = (struct carp_header *)(&ip6[1]);
1021                 bcopy(&ch, ch_ptr, sizeof(ch));
1022                 if (carp_prepare_ad(m, sc, ch_ptr))
1023                         goto resched;
1024
1025                 m->m_data += sizeof(*ip6);
1026                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip6));
1027                 m->m_data -= sizeof(*ip6);
1028
1029                 CARPSTATS_INC(carps_opackets6);
1030
1031                 carp_send_ad_error(sc, ip6_output(m, NULL, NULL, 0,
1032                     &sc->sc_carpdev->if_carp->cif_im6o, NULL, NULL));
1033         }
1034 #endif /* INET6 */
1035
1036 resched:
1037         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), carp_send_ad, sc);
1038 }
1039
1040 static void
1041 carp_addroute(struct carp_softc *sc)
1042 {
1043         struct ifaddr *ifa;
1044
1045         CARP_FOREACH_IFA(sc, ifa)
1046                 carp_ifa_addroute(ifa);
1047 }
1048
1049 static void
1050 carp_ifa_addroute(struct ifaddr *ifa)
1051 {
1052
1053         switch (ifa->ifa_addr->sa_family) {
1054 #ifdef INET
1055         case AF_INET:
1056                 in_addprefix(ifatoia(ifa), RTF_UP);
1057                 ifa_add_loopback_route(ifa,
1058                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
1059                 break;
1060 #endif
1061 #ifdef INET6
1062         case AF_INET6:
1063                 ifa_add_loopback_route(ifa,
1064                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
1065                 nd6_add_ifa_lle(ifatoia6(ifa));
1066                 break;
1067 #endif
1068         }
1069 }
1070
1071 static void
1072 carp_delroute(struct carp_softc *sc)
1073 {
1074         struct ifaddr *ifa;
1075
1076         CARP_FOREACH_IFA(sc, ifa)
1077                 carp_ifa_delroute(ifa);
1078 }
1079
1080 static void
1081 carp_ifa_delroute(struct ifaddr *ifa)
1082 {
1083
1084         switch (ifa->ifa_addr->sa_family) {
1085 #ifdef INET
1086         case AF_INET:
1087                 ifa_del_loopback_route(ifa,
1088                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
1089                 in_scrubprefix(ifatoia(ifa), LLE_STATIC);
1090                 break;
1091 #endif
1092 #ifdef INET6
1093         case AF_INET6:
1094                 ifa_del_loopback_route(ifa,
1095                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
1096                 nd6_rem_ifa_lle(ifatoia6(ifa), 1);
1097                 break;
1098 #endif
1099         }
1100 }
1101
1102 int
1103 carp_master(struct ifaddr *ifa)
1104 {
1105         struct carp_softc *sc = ifa->ifa_carp;
1106
1107         return (sc->sc_state == MASTER);
1108 }
1109
1110 #ifdef INET
1111 /*
1112  * Broadcast a gratuitous ARP request containing
1113  * the virtual router MAC address for each IP address
1114  * associated with the virtual router.
1115  */
1116 static void
1117 carp_send_arp(struct carp_softc *sc)
1118 {
1119         struct ifaddr *ifa;
1120         struct in_addr addr;
1121
1122         CARP_FOREACH_IFA(sc, ifa) {
1123                 if (ifa->ifa_addr->sa_family != AF_INET)
1124                         continue;
1125                 addr = ((struct sockaddr_in *)ifa->ifa_addr)->sin_addr;
1126                 arp_announce_ifaddr(sc->sc_carpdev, addr, LLADDR(&sc->sc_addr));
1127         }
1128 }
1129
1130 int
1131 carp_iamatch(struct ifaddr *ifa, uint8_t **enaddr)
1132 {
1133         struct carp_softc *sc = ifa->ifa_carp;
1134
1135         if (sc->sc_state == MASTER) {
1136                 *enaddr = LLADDR(&sc->sc_addr);
1137                 return (1);
1138         }
1139
1140         return (0);
1141 }
1142 #endif
1143
1144 #ifdef INET6
1145 static void
1146 carp_send_na(struct carp_softc *sc)
1147 {
1148         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1149         struct ifaddr *ifa;
1150         struct in6_addr *in6;
1151
1152         CARP_FOREACH_IFA(sc, ifa) {
1153                 if (ifa->ifa_addr->sa_family != AF_INET6)
1154                         continue;
1155
1156                 in6 = IFA_IN6(ifa);
1157                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1158                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1159                 DELAY(1000);    /* XXX */
1160         }
1161 }
1162
1163 /*
1164  * Returns ifa in case it's a carp address and it is MASTER, or if the address
1165  * matches and is not a carp address.  Returns NULL otherwise.
1166  */
1167 struct ifaddr *
1168 carp_iamatch6(struct ifnet *ifp, struct in6_addr *taddr)
1169 {
1170         struct ifaddr *ifa;
1171
1172         ifa = NULL;
1173         IF_ADDR_RLOCK(ifp);
1174         CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
1175                 if (ifa->ifa_addr->sa_family != AF_INET6)
1176                         continue;
1177                 if (!IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa)))
1178                         continue;
1179                 if (ifa->ifa_carp && ifa->ifa_carp->sc_state != MASTER)
1180                         ifa = NULL;
1181                 else
1182                         ifa_ref(ifa);
1183                 break;
1184         }
1185         IF_ADDR_RUNLOCK(ifp);
1186
1187         return (ifa);
1188 }
1189
1190 caddr_t
1191 carp_macmatch6(struct ifnet *ifp, struct mbuf *m, const struct in6_addr *taddr)
1192 {
1193         struct ifaddr *ifa;
1194
1195         IF_ADDR_RLOCK(ifp);
1196         IFNET_FOREACH_IFA(ifp, ifa)
1197                 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1198                     IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1199                         struct carp_softc *sc = ifa->ifa_carp;
1200                         struct m_tag *mtag;
1201
1202                         IF_ADDR_RUNLOCK(ifp);
1203
1204                         mtag = m_tag_get(PACKET_TAG_CARP,
1205                             sizeof(struct carp_softc *), M_NOWAIT);
1206                         if (mtag == NULL)
1207                                 /* Better a bit than nothing. */
1208                                 return (LLADDR(&sc->sc_addr));
1209
1210                         bcopy(&sc, mtag + 1, sizeof(sc));
1211                         m_tag_prepend(m, mtag);
1212
1213                         return (LLADDR(&sc->sc_addr));
1214                 }
1215         IF_ADDR_RUNLOCK(ifp);
1216
1217         return (NULL);
1218 }
1219 #endif /* INET6 */
1220
1221 int
1222 carp_forus(struct ifnet *ifp, u_char *dhost)
1223 {
1224         struct carp_softc *sc;
1225         uint8_t *ena = dhost;
1226
1227         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1228                 return (0);
1229
1230         CIF_LOCK(ifp->if_carp);
1231         IFNET_FOREACH_CARP(ifp, sc) {
1232                 CARP_LOCK(sc);
1233                 if (sc->sc_state == MASTER && !bcmp(dhost, LLADDR(&sc->sc_addr),
1234                     ETHER_ADDR_LEN)) {
1235                         CARP_UNLOCK(sc);
1236                         CIF_UNLOCK(ifp->if_carp);
1237                         return (1);
1238                 }
1239                 CARP_UNLOCK(sc);
1240         }
1241         CIF_UNLOCK(ifp->if_carp);
1242
1243         return (0);
1244 }
1245
1246 /* Master down timeout event, executed in callout context. */
1247 static void
1248 carp_master_down(void *v)
1249 {
1250         struct carp_softc *sc = v;
1251
1252         CARP_LOCK_ASSERT(sc);
1253
1254         CURVNET_SET(sc->sc_carpdev->if_vnet);
1255         if (sc->sc_state == BACKUP) {
1256                 carp_master_down_locked(sc, "master timed out");
1257         }
1258         CURVNET_RESTORE();
1259
1260         CARP_UNLOCK(sc);
1261 }
1262
1263 static void
1264 carp_master_down_locked(struct carp_softc *sc, const char *reason)
1265 {
1266
1267         CARP_LOCK_ASSERT(sc);
1268
1269         switch (sc->sc_state) {
1270         case BACKUP:
1271                 carp_set_state(sc, MASTER, reason);
1272                 carp_send_ad_locked(sc);
1273 #ifdef INET
1274                 carp_send_arp(sc);
1275 #endif
1276 #ifdef INET6
1277                 carp_send_na(sc);
1278 #endif
1279                 carp_setrun(sc, 0);
1280                 carp_addroute(sc);
1281                 break;
1282         case INIT:
1283         case MASTER:
1284 #ifdef INVARIANTS
1285                 panic("carp: VHID %u@%s: master_down event in %s state\n",
1286                     sc->sc_vhid,
1287                     sc->sc_carpdev->if_xname,
1288                     sc->sc_state ? "MASTER" : "INIT");
1289 #endif
1290                 break;
1291         }
1292 }
1293
1294 /*
1295  * When in backup state, af indicates whether to reset the master down timer
1296  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1297  */
1298 static void
1299 carp_setrun(struct carp_softc *sc, sa_family_t af)
1300 {
1301         struct timeval tv;
1302
1303         CARP_LOCK_ASSERT(sc);
1304
1305         if ((sc->sc_carpdev->if_flags & IFF_UP) == 0 ||
1306             sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1307             (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) ||
1308             !V_carp_allow)
1309                 return;
1310
1311         switch (sc->sc_state) {
1312         case INIT:
1313                 carp_set_state(sc, BACKUP, "initialization complete");
1314                 carp_setrun(sc, 0);
1315                 break;
1316         case BACKUP:
1317                 callout_stop(&sc->sc_ad_tmo);
1318                 tv.tv_sec = 3 * sc->sc_advbase;
1319                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1320                 switch (af) {
1321 #ifdef INET
1322                 case AF_INET:
1323                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1324                             carp_master_down, sc);
1325                         break;
1326 #endif
1327 #ifdef INET6
1328                 case AF_INET6:
1329                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1330                             carp_master_down, sc);
1331                         break;
1332 #endif
1333                 default:
1334 #ifdef INET
1335                         if (sc->sc_naddrs)
1336                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1337                                     carp_master_down, sc);
1338 #endif
1339 #ifdef INET6
1340                         if (sc->sc_naddrs6)
1341                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1342                                     carp_master_down, sc);
1343 #endif
1344                         break;
1345                 }
1346                 break;
1347         case MASTER:
1348                 tv.tv_sec = sc->sc_advbase;
1349                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1350                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1351                     carp_send_ad, sc);
1352                 break;
1353         }
1354 }
1355
1356 /*
1357  * Setup multicast structures.
1358  */
1359 static int
1360 carp_multicast_setup(struct carp_if *cif, sa_family_t sa)
1361 {
1362         struct ifnet *ifp = cif->cif_ifp;
1363         int error = 0;
1364
1365         switch (sa) {
1366 #ifdef INET
1367         case AF_INET:
1368             {
1369                 struct ip_moptions *imo = &cif->cif_imo;
1370                 struct in_mfilter *imf;
1371                 struct in_addr addr;
1372
1373                 if (ip_mfilter_first(&imo->imo_head) != NULL)
1374                         return (0);
1375
1376                 imf = ip_mfilter_alloc(M_WAITOK, 0, 0);
1377                 ip_mfilter_init(&imo->imo_head);
1378                 imo->imo_multicast_vif = -1;
1379
1380                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1381                 if ((error = in_joingroup(ifp, &addr, NULL,
1382                     &imf->imf_inm)) != 0) {
1383                         ip_mfilter_free(imf);
1384                         break;
1385                 }
1386
1387                 ip_mfilter_insert(&imo->imo_head, imf);
1388                 imo->imo_multicast_ifp = ifp;
1389                 imo->imo_multicast_ttl = CARP_DFLTTL;
1390                 imo->imo_multicast_loop = 0;
1391                 break;
1392            }
1393 #endif
1394 #ifdef INET6
1395         case AF_INET6:
1396             {
1397                 struct ip6_moptions *im6o = &cif->cif_im6o;
1398                 struct in6_mfilter *im6f[2];
1399                 struct in6_addr in6;
1400
1401                 if (ip6_mfilter_first(&im6o->im6o_head))
1402                         return (0);
1403
1404                 im6f[0] = ip6_mfilter_alloc(M_WAITOK, 0, 0);
1405                 im6f[1] = ip6_mfilter_alloc(M_WAITOK, 0, 0);
1406
1407                 ip6_mfilter_init(&im6o->im6o_head);
1408                 im6o->im6o_multicast_hlim = CARP_DFLTTL;
1409                 im6o->im6o_multicast_ifp = ifp;
1410
1411                 /* Join IPv6 CARP multicast group. */
1412                 bzero(&in6, sizeof(in6));
1413                 in6.s6_addr16[0] = htons(0xff02);
1414                 in6.s6_addr8[15] = 0x12;
1415                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1416                         ip6_mfilter_free(im6f[0]);
1417                         ip6_mfilter_free(im6f[1]);
1418                         break;
1419                 }
1420                 if ((error = in6_joingroup(ifp, &in6, NULL, &im6f[0]->im6f_in6m, 0)) != 0) {
1421                         ip6_mfilter_free(im6f[0]);
1422                         ip6_mfilter_free(im6f[1]);
1423                         break;
1424                 }
1425
1426                 /* Join solicited multicast address. */
1427                 bzero(&in6, sizeof(in6));
1428                 in6.s6_addr16[0] = htons(0xff02);
1429                 in6.s6_addr32[1] = 0;
1430                 in6.s6_addr32[2] = htonl(1);
1431                 in6.s6_addr32[3] = 0;
1432                 in6.s6_addr8[12] = 0xff;
1433
1434                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1435                         ip6_mfilter_free(im6f[0]);
1436                         ip6_mfilter_free(im6f[1]);
1437                         break;
1438                 }
1439
1440                 if ((error = in6_joingroup(ifp, &in6, NULL, &im6f[1]->im6f_in6m, 0)) != 0) {
1441                         in6_leavegroup(im6f[0]->im6f_in6m, NULL);
1442                         ip6_mfilter_free(im6f[0]);
1443                         ip6_mfilter_free(im6f[1]);
1444                         break;
1445                 }
1446                 ip6_mfilter_insert(&im6o->im6o_head, im6f[0]);
1447                 ip6_mfilter_insert(&im6o->im6o_head, im6f[1]);
1448                 break;
1449             }
1450 #endif
1451         }
1452
1453         return (error);
1454 }
1455
1456 /*
1457  * Free multicast structures.
1458  */
1459 static void
1460 carp_multicast_cleanup(struct carp_if *cif, sa_family_t sa)
1461 {
1462 #ifdef INET
1463         struct ip_moptions *imo = &cif->cif_imo;
1464         struct in_mfilter *imf;
1465 #endif
1466 #ifdef INET6
1467         struct ip6_moptions *im6o = &cif->cif_im6o;
1468         struct in6_mfilter *im6f;
1469 #endif
1470         sx_assert(&carp_sx, SA_XLOCKED);
1471
1472         switch (sa) {
1473 #ifdef INET
1474         case AF_INET:
1475                 if (cif->cif_naddrs != 0)
1476                         break;
1477
1478                 while ((imf = ip_mfilter_first(&imo->imo_head)) != NULL) {
1479                         ip_mfilter_remove(&imo->imo_head, imf);
1480                         in_leavegroup(imf->imf_inm, NULL);
1481                         ip_mfilter_free(imf);
1482                 }
1483                 break;
1484 #endif
1485 #ifdef INET6
1486         case AF_INET6:
1487                 if (cif->cif_naddrs6 != 0)
1488                         break;
1489
1490                 while ((im6f = ip6_mfilter_first(&im6o->im6o_head)) != NULL) {
1491                         ip6_mfilter_remove(&im6o->im6o_head, im6f);
1492                         in6_leavegroup(im6f->im6f_in6m, NULL);
1493                         ip6_mfilter_free(im6f);
1494                 }
1495                 break;
1496 #endif
1497         }
1498 }
1499
1500 int
1501 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa)
1502 {
1503         struct m_tag *mtag;
1504         struct carp_softc *sc;
1505
1506         if (!sa)
1507                 return (0);
1508
1509         switch (sa->sa_family) {
1510 #ifdef INET
1511         case AF_INET:
1512                 break;
1513 #endif
1514 #ifdef INET6
1515         case AF_INET6:
1516                 break;
1517 #endif
1518         default:
1519                 return (0);
1520         }
1521
1522         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
1523         if (mtag == NULL)
1524                 return (0);
1525
1526         bcopy(mtag + 1, &sc, sizeof(sc));
1527
1528         /* Set the source MAC address to the Virtual Router MAC Address. */
1529         switch (ifp->if_type) {
1530         case IFT_ETHER:
1531         case IFT_BRIDGE:
1532         case IFT_L2VLAN: {
1533                         struct ether_header *eh;
1534
1535                         eh = mtod(m, struct ether_header *);
1536                         eh->ether_shost[0] = 0;
1537                         eh->ether_shost[1] = 0;
1538                         eh->ether_shost[2] = 0x5e;
1539                         eh->ether_shost[3] = 0;
1540                         eh->ether_shost[4] = 1;
1541                         eh->ether_shost[5] = sc->sc_vhid;
1542                 }
1543                 break;
1544         default:
1545                 printf("%s: carp is not supported for the %d interface type\n",
1546                     ifp->if_xname, ifp->if_type);
1547                 return (EOPNOTSUPP);
1548         }
1549
1550         return (0);
1551 }
1552
1553 static struct carp_softc*
1554 carp_alloc(struct ifnet *ifp)
1555 {
1556         struct carp_softc *sc;
1557         struct carp_if *cif;
1558
1559         sx_assert(&carp_sx, SA_XLOCKED);
1560
1561         if ((cif = ifp->if_carp) == NULL)
1562                 cif = carp_alloc_if(ifp);
1563
1564         sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
1565
1566         sc->sc_advbase = CARP_DFLTINTV;
1567         sc->sc_vhid = -1;       /* required setting */
1568         sc->sc_init_counter = 1;
1569         sc->sc_state = INIT;
1570
1571         sc->sc_ifasiz = sizeof(struct ifaddr *);
1572         sc->sc_ifas = malloc(sc->sc_ifasiz, M_CARP, M_WAITOK|M_ZERO);
1573         sc->sc_carpdev = ifp;
1574
1575         CARP_LOCK_INIT(sc);
1576 #ifdef INET
1577         callout_init_mtx(&sc->sc_md_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1578 #endif
1579 #ifdef INET6
1580         callout_init_mtx(&sc->sc_md6_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1581 #endif
1582         callout_init_mtx(&sc->sc_ad_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1583
1584         CIF_LOCK(cif);
1585         TAILQ_INSERT_TAIL(&cif->cif_vrs, sc, sc_list);
1586         CIF_UNLOCK(cif);
1587
1588         mtx_lock(&carp_mtx);
1589         LIST_INSERT_HEAD(&carp_list, sc, sc_next);
1590         mtx_unlock(&carp_mtx);
1591
1592         return (sc);
1593 }
1594
1595 static void
1596 carp_grow_ifas(struct carp_softc *sc)
1597 {
1598         struct ifaddr **new;
1599
1600         new = malloc(sc->sc_ifasiz * 2, M_CARP, M_WAITOK | M_ZERO);
1601         CARP_LOCK(sc);
1602         bcopy(sc->sc_ifas, new, sc->sc_ifasiz);
1603         free(sc->sc_ifas, M_CARP);
1604         sc->sc_ifas = new;
1605         sc->sc_ifasiz *= 2;
1606         CARP_UNLOCK(sc);
1607 }
1608
1609 static void
1610 carp_destroy(struct carp_softc *sc)
1611 {
1612         struct ifnet *ifp = sc->sc_carpdev;
1613         struct carp_if *cif = ifp->if_carp;
1614
1615         sx_assert(&carp_sx, SA_XLOCKED);
1616
1617         if (sc->sc_suppress)
1618                 carp_demote_adj(-V_carp_ifdown_adj, "vhid removed");
1619         CARP_UNLOCK(sc);
1620
1621         CIF_LOCK(cif);
1622         TAILQ_REMOVE(&cif->cif_vrs, sc, sc_list);
1623         CIF_UNLOCK(cif);
1624
1625         mtx_lock(&carp_mtx);
1626         LIST_REMOVE(sc, sc_next);
1627         mtx_unlock(&carp_mtx);
1628
1629         callout_drain(&sc->sc_ad_tmo);
1630 #ifdef INET
1631         callout_drain(&sc->sc_md_tmo);
1632 #endif
1633 #ifdef INET6
1634         callout_drain(&sc->sc_md6_tmo);
1635 #endif
1636         CARP_LOCK_DESTROY(sc);
1637
1638         free(sc->sc_ifas, M_CARP);
1639         free(sc, M_CARP);
1640 }
1641
1642 static struct carp_if*
1643 carp_alloc_if(struct ifnet *ifp)
1644 {
1645         struct carp_if *cif;
1646         int error;
1647
1648         cif = malloc(sizeof(*cif), M_CARP, M_WAITOK|M_ZERO);
1649
1650         if ((error = ifpromisc(ifp, 1)) != 0)
1651                 printf("%s: ifpromisc(%s) failed: %d\n",
1652                     __func__, ifp->if_xname, error);
1653         else
1654                 cif->cif_flags |= CIF_PROMISC;
1655
1656         CIF_LOCK_INIT(cif);
1657         cif->cif_ifp = ifp;
1658         TAILQ_INIT(&cif->cif_vrs);
1659
1660         IF_ADDR_WLOCK(ifp);
1661         ifp->if_carp = cif;
1662         if_ref(ifp);
1663         IF_ADDR_WUNLOCK(ifp);
1664
1665         return (cif);
1666 }
1667
1668 static void
1669 carp_free_if(struct carp_if *cif)
1670 {
1671         struct ifnet *ifp = cif->cif_ifp;
1672
1673         CIF_LOCK_ASSERT(cif);
1674         KASSERT(TAILQ_EMPTY(&cif->cif_vrs), ("%s: softc list not empty",
1675             __func__));
1676
1677         IF_ADDR_WLOCK(ifp);
1678         ifp->if_carp = NULL;
1679         IF_ADDR_WUNLOCK(ifp);
1680
1681         CIF_LOCK_DESTROY(cif);
1682
1683         if (cif->cif_flags & CIF_PROMISC)
1684                 ifpromisc(ifp, 0);
1685         if_rele(ifp);
1686
1687         free(cif, M_CARP);
1688 }
1689
1690 static void
1691 carp_carprcp(struct carpreq *carpr, struct carp_softc *sc, int priv)
1692 {
1693
1694         CARP_LOCK(sc);
1695         carpr->carpr_state = sc->sc_state;
1696         carpr->carpr_vhid = sc->sc_vhid;
1697         carpr->carpr_advbase = sc->sc_advbase;
1698         carpr->carpr_advskew = sc->sc_advskew;
1699         if (priv)
1700                 bcopy(sc->sc_key, carpr->carpr_key, sizeof(carpr->carpr_key));
1701         else
1702                 bzero(carpr->carpr_key, sizeof(carpr->carpr_key));
1703         CARP_UNLOCK(sc);
1704 }
1705
1706 int
1707 carp_ioctl(struct ifreq *ifr, u_long cmd, struct thread *td)
1708 {
1709         struct carpreq carpr;
1710         struct ifnet *ifp;
1711         struct carp_softc *sc = NULL;
1712         int error = 0, locked = 0;
1713
1714         if ((error = copyin(ifr_data_get_ptr(ifr), &carpr, sizeof carpr)))
1715                 return (error);
1716
1717         ifp = ifunit_ref(ifr->ifr_name);
1718         if (ifp == NULL)
1719                 return (ENXIO);
1720
1721         switch (ifp->if_type) {
1722         case IFT_ETHER:
1723         case IFT_L2VLAN:
1724         case IFT_BRIDGE:
1725                 break;
1726         default:
1727                 error = EOPNOTSUPP;
1728                 goto out;
1729         }
1730
1731         if ((ifp->if_flags & IFF_MULTICAST) == 0) {
1732                 error = EADDRNOTAVAIL;
1733                 goto out;
1734         }
1735
1736         sx_xlock(&carp_sx);
1737         switch (cmd) {
1738         case SIOCSVH:
1739                 if ((error = priv_check(td, PRIV_NETINET_CARP)))
1740                         break;
1741                 if (carpr.carpr_vhid <= 0 || carpr.carpr_vhid > CARP_MAXVHID ||
1742                     carpr.carpr_advbase < 0 || carpr.carpr_advskew < 0) {
1743                         error = EINVAL;
1744                         break;
1745                 }
1746
1747                 if (ifp->if_carp) {
1748                         IFNET_FOREACH_CARP(ifp, sc)
1749                                 if (sc->sc_vhid == carpr.carpr_vhid)
1750                                         break;
1751                 }
1752                 if (sc == NULL) {
1753                         sc = carp_alloc(ifp);
1754                         CARP_LOCK(sc);
1755                         sc->sc_vhid = carpr.carpr_vhid;
1756                         LLADDR(&sc->sc_addr)[0] = 0;
1757                         LLADDR(&sc->sc_addr)[1] = 0;
1758                         LLADDR(&sc->sc_addr)[2] = 0x5e;
1759                         LLADDR(&sc->sc_addr)[3] = 0;
1760                         LLADDR(&sc->sc_addr)[4] = 1;
1761                         LLADDR(&sc->sc_addr)[5] = sc->sc_vhid;
1762                 } else
1763                         CARP_LOCK(sc);
1764                 locked = 1;
1765                 if (carpr.carpr_advbase > 0) {
1766                         if (carpr.carpr_advbase > 255 ||
1767                             carpr.carpr_advbase < CARP_DFLTINTV) {
1768                                 error = EINVAL;
1769                                 break;
1770                         }
1771                         sc->sc_advbase = carpr.carpr_advbase;
1772                 }
1773                 if (carpr.carpr_advskew >= 255) {
1774                         error = EINVAL;
1775                         break;
1776                 }
1777                 sc->sc_advskew = carpr.carpr_advskew;
1778                 if (carpr.carpr_key[0] != '\0') {
1779                         bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1780                         carp_hmac_prepare(sc);
1781                 }
1782                 if (sc->sc_state != INIT &&
1783                     carpr.carpr_state != sc->sc_state) {
1784                         switch (carpr.carpr_state) {
1785                         case BACKUP:
1786                                 callout_stop(&sc->sc_ad_tmo);
1787                                 carp_set_state(sc, BACKUP,
1788                                     "user requested via ifconfig");
1789                                 carp_setrun(sc, 0);
1790                                 carp_delroute(sc);
1791                                 break;
1792                         case MASTER:
1793                                 carp_master_down_locked(sc,
1794                                     "user requested via ifconfig");
1795                                 break;
1796                         default:
1797                                 break;
1798                         }
1799                 }
1800                 break;
1801
1802         case SIOCGVH:
1803             {
1804                 int priveleged;
1805
1806                 if (carpr.carpr_vhid < 0 || carpr.carpr_vhid > CARP_MAXVHID) {
1807                         error = EINVAL;
1808                         break;
1809                 }
1810                 if (carpr.carpr_count < 1) {
1811                         error = EMSGSIZE;
1812                         break;
1813                 }
1814                 if (ifp->if_carp == NULL) {
1815                         error = ENOENT;
1816                         break;
1817                 }
1818
1819                 priveleged = (priv_check(td, PRIV_NETINET_CARP) == 0);
1820                 if (carpr.carpr_vhid != 0) {
1821                         IFNET_FOREACH_CARP(ifp, sc)
1822                                 if (sc->sc_vhid == carpr.carpr_vhid)
1823                                         break;
1824                         if (sc == NULL) {
1825                                 error = ENOENT;
1826                                 break;
1827                         }
1828                         carp_carprcp(&carpr, sc, priveleged);
1829                         error = copyout(&carpr, ifr_data_get_ptr(ifr),
1830                             sizeof(carpr));
1831                 } else  {
1832                         int i, count;
1833
1834                         count = 0;
1835                         IFNET_FOREACH_CARP(ifp, sc)
1836                                 count++;
1837
1838                         if (count > carpr.carpr_count) {
1839                                 CIF_UNLOCK(ifp->if_carp);
1840                                 error = EMSGSIZE;
1841                                 break;
1842                         }
1843
1844                         i = 0;
1845                         IFNET_FOREACH_CARP(ifp, sc) {
1846                                 carp_carprcp(&carpr, sc, priveleged);
1847                                 carpr.carpr_count = count;
1848                                 error = copyout(&carpr,
1849                                     (caddr_t)ifr_data_get_ptr(ifr) +
1850                                     (i * sizeof(carpr)), sizeof(carpr));
1851                                 if (error) {
1852                                         CIF_UNLOCK(ifp->if_carp);
1853                                         break;
1854                                 }
1855                                 i++;
1856                         }
1857                 }
1858                 break;
1859             }
1860         default:
1861                 error = EINVAL;
1862         }
1863         sx_xunlock(&carp_sx);
1864
1865 out:
1866         if (locked)
1867                 CARP_UNLOCK(sc);
1868         if_rele(ifp);
1869
1870         return (error);
1871 }
1872
1873 static int
1874 carp_get_vhid(struct ifaddr *ifa)
1875 {
1876
1877         if (ifa == NULL || ifa->ifa_carp == NULL)
1878                 return (0);
1879
1880         return (ifa->ifa_carp->sc_vhid);
1881 }
1882
1883 int
1884 carp_attach(struct ifaddr *ifa, int vhid)
1885 {
1886         struct ifnet *ifp = ifa->ifa_ifp;
1887         struct carp_if *cif = ifp->if_carp;
1888         struct carp_softc *sc;
1889         int index, error;
1890
1891         KASSERT(ifa->ifa_carp == NULL, ("%s: ifa %p attached", __func__, ifa));
1892
1893         switch (ifa->ifa_addr->sa_family) {
1894 #ifdef INET
1895         case AF_INET:
1896 #endif
1897 #ifdef INET6
1898         case AF_INET6:
1899 #endif
1900                 break;
1901         default:
1902                 return (EPROTOTYPE);
1903         }
1904
1905         sx_xlock(&carp_sx);
1906         if (ifp->if_carp == NULL) {
1907                 sx_xunlock(&carp_sx);
1908                 return (ENOPROTOOPT);
1909         }
1910
1911         IFNET_FOREACH_CARP(ifp, sc)
1912                 if (sc->sc_vhid == vhid)
1913                         break;
1914         if (sc == NULL) {
1915                 sx_xunlock(&carp_sx);
1916                 return (ENOENT);
1917         }
1918
1919         error = carp_multicast_setup(cif, ifa->ifa_addr->sa_family);
1920         if (error) {
1921                 CIF_FREE(cif);
1922                 sx_xunlock(&carp_sx);
1923                 return (error);
1924         }
1925
1926         index = sc->sc_naddrs + sc->sc_naddrs6 + 1;
1927         if (index > sc->sc_ifasiz / sizeof(struct ifaddr *))
1928                 carp_grow_ifas(sc);
1929
1930         switch (ifa->ifa_addr->sa_family) {
1931 #ifdef INET
1932         case AF_INET:
1933                 cif->cif_naddrs++;
1934                 sc->sc_naddrs++;
1935                 break;
1936 #endif
1937 #ifdef INET6
1938         case AF_INET6:
1939                 cif->cif_naddrs6++;
1940                 sc->sc_naddrs6++;
1941                 break;
1942 #endif
1943         }
1944
1945         ifa_ref(ifa);
1946
1947         CARP_LOCK(sc);
1948         sc->sc_ifas[index - 1] = ifa;
1949         ifa->ifa_carp = sc;
1950         carp_hmac_prepare(sc);
1951         carp_sc_state(sc);
1952         CARP_UNLOCK(sc);
1953
1954         sx_xunlock(&carp_sx);
1955
1956         return (0);
1957 }
1958
1959 void
1960 carp_detach(struct ifaddr *ifa, bool keep_cif)
1961 {
1962         struct ifnet *ifp = ifa->ifa_ifp;
1963         struct carp_if *cif = ifp->if_carp;
1964         struct carp_softc *sc = ifa->ifa_carp;
1965         int i, index;
1966
1967         KASSERT(sc != NULL, ("%s: %p not attached", __func__, ifa));
1968
1969         sx_xlock(&carp_sx);
1970
1971         CARP_LOCK(sc);
1972         /* Shift array. */
1973         index = sc->sc_naddrs + sc->sc_naddrs6;
1974         for (i = 0; i < index; i++)
1975                 if (sc->sc_ifas[i] == ifa)
1976                         break;
1977         KASSERT(i < index, ("%s: %p no backref", __func__, ifa));
1978         for (; i < index - 1; i++)
1979                 sc->sc_ifas[i] = sc->sc_ifas[i+1];
1980         sc->sc_ifas[index - 1] = NULL;
1981
1982         switch (ifa->ifa_addr->sa_family) {
1983 #ifdef INET
1984         case AF_INET:
1985                 cif->cif_naddrs--;
1986                 sc->sc_naddrs--;
1987                 break;
1988 #endif
1989 #ifdef INET6
1990         case AF_INET6:
1991                 cif->cif_naddrs6--;
1992                 sc->sc_naddrs6--;
1993                 break;
1994 #endif
1995         }
1996
1997         carp_ifa_delroute(ifa);
1998         carp_multicast_cleanup(cif, ifa->ifa_addr->sa_family);
1999
2000         ifa->ifa_carp = NULL;
2001         ifa_free(ifa);
2002
2003         carp_hmac_prepare(sc);
2004         carp_sc_state(sc);
2005
2006         if (!keep_cif && sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0)
2007                 carp_destroy(sc);
2008         else
2009                 CARP_UNLOCK(sc);
2010
2011         if (!keep_cif)
2012                 CIF_FREE(cif);
2013
2014         sx_xunlock(&carp_sx);
2015 }
2016
2017 static void
2018 carp_set_state(struct carp_softc *sc, int state, const char *reason)
2019 {
2020
2021         CARP_LOCK_ASSERT(sc);
2022
2023         if (sc->sc_state != state) {
2024                 const char *carp_states[] = { CARP_STATES };
2025                 char subsys[IFNAMSIZ+5];
2026
2027                 snprintf(subsys, IFNAMSIZ+5, "%u@%s", sc->sc_vhid,
2028                     sc->sc_carpdev->if_xname);
2029
2030                 CARP_LOG("%s: %s -> %s (%s)\n", subsys,
2031                     carp_states[sc->sc_state], carp_states[state], reason);
2032
2033                 sc->sc_state = state;
2034
2035                 devctl_notify("CARP", subsys, carp_states[state], NULL);
2036         }
2037 }
2038
2039 static void
2040 carp_linkstate(struct ifnet *ifp)
2041 {
2042         struct carp_softc *sc;
2043
2044         CIF_LOCK(ifp->if_carp);
2045         IFNET_FOREACH_CARP(ifp, sc) {
2046                 CARP_LOCK(sc);
2047                 carp_sc_state(sc);
2048                 CARP_UNLOCK(sc);
2049         }
2050         CIF_UNLOCK(ifp->if_carp);
2051 }
2052
2053 static void
2054 carp_sc_state(struct carp_softc *sc)
2055 {
2056
2057         CARP_LOCK_ASSERT(sc);
2058
2059         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
2060             !(sc->sc_carpdev->if_flags & IFF_UP) ||
2061             !V_carp_allow) {
2062                 callout_stop(&sc->sc_ad_tmo);
2063 #ifdef INET
2064                 callout_stop(&sc->sc_md_tmo);
2065 #endif
2066 #ifdef INET6
2067                 callout_stop(&sc->sc_md6_tmo);
2068 #endif
2069                 carp_set_state(sc, INIT, "hardware interface down");
2070                 carp_setrun(sc, 0);
2071                 if (!sc->sc_suppress)
2072                         carp_demote_adj(V_carp_ifdown_adj, "interface down");
2073                 sc->sc_suppress = 1;
2074         } else {
2075                 carp_set_state(sc, INIT, "hardware interface up");
2076                 carp_setrun(sc, 0);
2077                 if (sc->sc_suppress)
2078                         carp_demote_adj(-V_carp_ifdown_adj, "interface up");
2079                 sc->sc_suppress = 0;
2080         }
2081 }
2082
2083 static void
2084 carp_demote_adj(int adj, char *reason)
2085 {
2086         atomic_add_int(&V_carp_demotion, adj);
2087         CARP_LOG("demoted by %d to %d (%s)\n", adj, V_carp_demotion, reason);
2088         taskqueue_enqueue(taskqueue_swi, &carp_sendall_task);
2089 }
2090
2091 static int
2092 carp_allow_sysctl(SYSCTL_HANDLER_ARGS)
2093 {
2094         int new, error;
2095         struct carp_softc *sc;
2096
2097         new = V_carp_allow;
2098         error = sysctl_handle_int(oidp, &new, 0, req);
2099         if (error || !req->newptr)
2100                 return (error);
2101
2102         if (V_carp_allow != new) {
2103                 V_carp_allow = new;
2104
2105                 mtx_lock(&carp_mtx);
2106                 LIST_FOREACH(sc, &carp_list, sc_next) {
2107                         CARP_LOCK(sc);
2108                         if (curvnet == sc->sc_carpdev->if_vnet)
2109                                 carp_sc_state(sc);
2110                         CARP_UNLOCK(sc);
2111                 }
2112                 mtx_unlock(&carp_mtx);
2113         }
2114
2115         return (0);
2116 }
2117
2118 static int
2119 carp_dscp_sysctl(SYSCTL_HANDLER_ARGS)
2120 {
2121         int new, error;
2122
2123         new = V_carp_dscp;
2124         error = sysctl_handle_int(oidp, &new, 0, req);
2125         if (error || !req->newptr)
2126                 return (error);
2127
2128         if (new < 0 || new > 63)
2129                 return (EINVAL);
2130
2131         V_carp_dscp = new;
2132
2133         return (0);
2134 }
2135
2136 static int
2137 carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS)
2138 {
2139         int new, error;
2140
2141         new = V_carp_demotion;
2142         error = sysctl_handle_int(oidp, &new, 0, req);
2143         if (error || !req->newptr)
2144                 return (error);
2145
2146         carp_demote_adj(new, "sysctl");
2147
2148         return (0);
2149 }
2150
2151 #ifdef INET
2152 extern  struct domain inetdomain;
2153 static struct protosw in_carp_protosw = {
2154         .pr_type =              SOCK_RAW,
2155         .pr_domain =            &inetdomain,
2156         .pr_protocol =          IPPROTO_CARP,
2157         .pr_flags =             PR_ATOMIC|PR_ADDR,
2158         .pr_input =             carp_input,
2159         .pr_output =            rip_output,
2160         .pr_ctloutput =         rip_ctloutput,
2161         .pr_usrreqs =           &rip_usrreqs
2162 };
2163 #endif
2164
2165 #ifdef INET6
2166 extern  struct domain inet6domain;
2167 static struct protosw in6_carp_protosw = {
2168         .pr_type =              SOCK_RAW,
2169         .pr_domain =            &inet6domain,
2170         .pr_protocol =          IPPROTO_CARP,
2171         .pr_flags =             PR_ATOMIC|PR_ADDR,
2172         .pr_input =             carp6_input,
2173         .pr_output =            rip6_output,
2174         .pr_ctloutput =         rip6_ctloutput,
2175         .pr_usrreqs =           &rip6_usrreqs
2176 };
2177 #endif
2178
2179 #ifdef VIMAGE
2180 #if defined(__i386__)
2181 /*
2182  * XXX This is a hack to work around an absolute relocation outside
2183  * set_vnet by one (on the stop symbol) for carpstats.  Add a dummy variable
2184  * to the end of the file in the hope that the linker will just keep the
2185  * order (as it seems to do at the moment).  It is understood to be fragile.
2186  * See PR 230857 for a longer discussion of the problem and the referenced
2187  * review for possible alternate solutions.  Each is a hack; we just need
2188  * the least intrusive one for the next release.
2189  */
2190 VNET_DEFINE(char, carp_zzz) = 0xde;
2191 #endif
2192 #endif
2193
2194 static void
2195 carp_mod_cleanup(void)
2196 {
2197
2198 #ifdef INET
2199         if (proto_reg[CARP_INET] == 0) {
2200                 (void)ipproto_unregister(IPPROTO_CARP);
2201                 pf_proto_unregister(PF_INET, IPPROTO_CARP, SOCK_RAW);
2202                 proto_reg[CARP_INET] = -1;
2203         }
2204         carp_iamatch_p = NULL;
2205 #endif
2206 #ifdef INET6
2207         if (proto_reg[CARP_INET6] == 0) {
2208                 (void)ip6proto_unregister(IPPROTO_CARP);
2209                 pf_proto_unregister(PF_INET6, IPPROTO_CARP, SOCK_RAW);
2210                 proto_reg[CARP_INET6] = -1;
2211         }
2212         carp_iamatch6_p = NULL;
2213         carp_macmatch6_p = NULL;
2214 #endif
2215         carp_ioctl_p = NULL;
2216         carp_attach_p = NULL;
2217         carp_detach_p = NULL;
2218         carp_get_vhid_p = NULL;
2219         carp_linkstate_p = NULL;
2220         carp_forus_p = NULL;
2221         carp_output_p = NULL;
2222         carp_demote_adj_p = NULL;
2223         carp_master_p = NULL;
2224         mtx_unlock(&carp_mtx);
2225         taskqueue_drain(taskqueue_swi, &carp_sendall_task);
2226         mtx_destroy(&carp_mtx);
2227         sx_destroy(&carp_sx);
2228 }
2229
2230 static int
2231 carp_mod_load(void)
2232 {
2233         int err;
2234
2235         mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2236         sx_init(&carp_sx, "carp_sx");
2237         LIST_INIT(&carp_list);
2238         carp_get_vhid_p = carp_get_vhid;
2239         carp_forus_p = carp_forus;
2240         carp_output_p = carp_output;
2241         carp_linkstate_p = carp_linkstate;
2242         carp_ioctl_p = carp_ioctl;
2243         carp_attach_p = carp_attach;
2244         carp_detach_p = carp_detach;
2245         carp_demote_adj_p = carp_demote_adj;
2246         carp_master_p = carp_master;
2247 #ifdef INET6
2248         carp_iamatch6_p = carp_iamatch6;
2249         carp_macmatch6_p = carp_macmatch6;
2250         proto_reg[CARP_INET6] = pf_proto_register(PF_INET6,
2251             (struct protosw *)&in6_carp_protosw);
2252         if (proto_reg[CARP_INET6]) {
2253                 printf("carp: error %d attaching to PF_INET6\n",
2254                     proto_reg[CARP_INET6]);
2255                 carp_mod_cleanup();
2256                 return (proto_reg[CARP_INET6]);
2257         }
2258         err = ip6proto_register(IPPROTO_CARP);
2259         if (err) {
2260                 printf("carp: error %d registering with INET6\n", err);
2261                 carp_mod_cleanup();
2262                 return (err);
2263         }
2264 #endif
2265 #ifdef INET
2266         carp_iamatch_p = carp_iamatch;
2267         proto_reg[CARP_INET] = pf_proto_register(PF_INET, &in_carp_protosw);
2268         if (proto_reg[CARP_INET]) {
2269                 printf("carp: error %d attaching to PF_INET\n",
2270                     proto_reg[CARP_INET]);
2271                 carp_mod_cleanup();
2272                 return (proto_reg[CARP_INET]);
2273         }
2274         err = ipproto_register(IPPROTO_CARP);
2275         if (err) {
2276                 printf("carp: error %d registering with INET\n", err);
2277                 carp_mod_cleanup();
2278                 return (err);
2279         }
2280 #endif
2281         return (0);
2282 }
2283
2284 static int
2285 carp_modevent(module_t mod, int type, void *data)
2286 {
2287         switch (type) {
2288         case MOD_LOAD:
2289                 return carp_mod_load();
2290                 /* NOTREACHED */
2291         case MOD_UNLOAD:
2292                 mtx_lock(&carp_mtx);
2293                 if (LIST_EMPTY(&carp_list))
2294                         carp_mod_cleanup();
2295                 else {
2296                         mtx_unlock(&carp_mtx);
2297                         return (EBUSY);
2298                 }
2299                 break;
2300
2301         default:
2302                 return (EINVAL);
2303         }
2304
2305         return (0);
2306 }
2307
2308 static moduledata_t carp_mod = {
2309         "carp",
2310         carp_modevent,
2311         0
2312 };
2313
2314 DECLARE_MODULE(carp, carp_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);