sys/netinet/ip_carp.c

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