sys/netinet/ip_carp.c

   1 /*
   2  * Copyright (c) 2002 Michael Shalayeff. All rights reserved.
   3  * Copyright (c) 2003 Ryan McBride. All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  *
  14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  15  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  16  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  17  * IN NO EVENT SHALL THE AUTHOR OR HIS RELATIVES BE LIABLE FOR ANY DIRECT,
  18  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  19  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  20  * SERVICES; LOSS OF MIND, USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
  22  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
  23  * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
  24  * THE POSSIBILITY OF SUCH DAMAGE.
  25  */
  26 /*
  27  * $FreeBSD: src/sys/netinet/ip_carp.c,v 1.48 2007/02/02 09:39:09 glebius Exp $
  28  */
  29
  30 #include "opt_carp.h"
  31 #include "opt_inet.h"
  32 #include "opt_inet6.h"
  33
  34 #include <sys/param.h>
  35 #include <sys/systm.h>
  36 #include <sys/kernel.h>
  37 #include <sys/in_cksum.h>
  38 #include <sys/limits.h>
  39 #include <sys/malloc.h>
  40 #include <sys/mbuf.h>
  41 #include <sys/msgport2.h>
  42 #include <sys/time.h>
  43 #include <sys/proc.h>
  44 #include <sys/priv.h>
  45 #include <sys/sockio.h>
  46 #include <sys/socket.h>
  47 #include <sys/sysctl.h>
  48 #include <sys/syslog.h>
  49 #include <sys/thread.h>
  50
  51 #include <machine/stdarg.h>
  52 #include <crypto/sha1.h>
  53
  54 #include <net/bpf.h>
  55 #include <net/ethernet.h>
  56 #include <net/if.h>
  57 #include <net/if_dl.h>
  58 #include <net/if_types.h>
  59 #include <net/route.h>
  60 #include <net/if_clone.h>
  61 #include <net/if_var.h>
  62 #include <net/ifq_var.h>
  63
  64 #ifdef INET
  65 #include <netinet/in.h>
  66 #include <netinet/in_var.h>
  67 #include <netinet/in_systm.h>
  68 #include <netinet/ip.h>
  69 #include <netinet/ip_var.h>
  70 #include <netinet/if_ether.h>
  71 #endif
  72
  73 #ifdef INET6
  74 #include <netinet/icmp6.h>
  75 #include <netinet/ip6.h>
  76 #include <netinet6/ip6_var.h>
  77 #include <netinet6/scope6_var.h>
  78 #include <netinet6/nd6.h>
  79 #endif
  80
  81 #include <netinet/ip_carp.h>
  82
  83 #define CARP_IFNAME             "carp"
  84 #define CARP_IS_RUNNING(ifp)    \
  85         (((ifp)->if_flags & (IFF_UP | IFF_RUNNING)) == (IFF_UP | IFF_RUNNING))
  86
  87 struct carp_vhaddr {
  88         uint32_t                vha_flags;      /* CARP_VHAF_ */
  89         struct in_ifaddr        *vha_ia;        /* carp address */
  90         struct in_ifaddr        *vha_iaback;    /* backing address */
  91         TAILQ_ENTRY(carp_vhaddr) vha_link;
  92 };
  93 TAILQ_HEAD(carp_vhaddr_list, carp_vhaddr);
  94
  95 struct carp_softc {
  96         struct arpcom            arpcom;
  97         struct ifnet            *sc_carpdev;    /* parent interface */
  98         struct carp_vhaddr_list  sc_vha_list;   /* virtual addr list */
  99
 100         const struct in_ifaddr  *sc_ia;         /* primary iface address v4 */
 101         struct ip_moptions       sc_imo;
 102
 103 #ifdef INET6
 104         struct in6_ifaddr       *sc_ia6;        /* primary iface address v6 */
 105         struct ip6_moptions      sc_im6o;
 106 #endif /* INET6 */
 107         TAILQ_ENTRY(carp_softc)  sc_list;
 108
 109         enum { INIT = 0, BACKUP, MASTER }
 110                                  sc_state;
 111         int                      sc_dead;
 112
 113         int                      sc_suppress;
 114
 115         int                      sc_sendad_errors;
 116 #define CARP_SENDAD_MAX_ERRORS  3
 117         int                      sc_sendad_success;
 118 #define CARP_SENDAD_MIN_SUCCESS 3
 119
 120         int                      sc_vhid;
 121         int                      sc_advskew;
 122         int                      sc_naddrs;     /* actually used IPv4 vha */
 123         int                      sc_naddrs6;
 124         int                      sc_advbase;    /* seconds */
 125         int                      sc_init_counter;
 126         uint64_t                 sc_counter;
 127
 128         /* authentication */
 129 #define CARP_HMAC_PAD   64
 130         unsigned char            sc_key[CARP_KEY_LEN];
 131         unsigned char            sc_pad[CARP_HMAC_PAD];
 132         SHA1_CTX                 sc_sha1;
 133
 134         struct callout           sc_ad_tmo;     /* advertisement timeout */
 135         struct callout           sc_md_tmo;     /* master down timeout */
 136         struct callout           sc_md6_tmo;    /* master down timeout */
 137
 138         LIST_ENTRY(carp_softc)   sc_next;       /* Interface clue */
 139 };
 140
 141 #define sc_if   arpcom.ac_if
 142
 143 struct carp_if {
 144         TAILQ_HEAD(, carp_softc) vhif_vrs;
 145 };
 146
 147 enum    { CARP_COUNT_MASTER, CARP_COUNT_RUNNING };
 148
 149 SYSCTL_DECL(_net_inet_carp);
 150
 151 static int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 1, 0, 0 }; /* XXX for now */
 152 SYSCTL_INT(_net_inet_carp, CARPCTL_ALLOW, allow, CTLFLAG_RW,
 153     &carp_opts[CARPCTL_ALLOW], 0, "Accept incoming CARP packets");
 154 SYSCTL_INT(_net_inet_carp, CARPCTL_PREEMPT, preempt, CTLFLAG_RW,
 155     &carp_opts[CARPCTL_PREEMPT], 0, "high-priority backup preemption mode");
 156 SYSCTL_INT(_net_inet_carp, CARPCTL_LOG, log, CTLFLAG_RW,
 157     &carp_opts[CARPCTL_LOG], 0, "log bad carp packets");
 158 SYSCTL_INT(_net_inet_carp, CARPCTL_ARPBALANCE, arpbalance, CTLFLAG_RW,
 159     &carp_opts[CARPCTL_ARPBALANCE], 0, "balance arp responses");
 160
 161 static int carp_suppress_preempt = 0;
 162 SYSCTL_INT(_net_inet_carp, OID_AUTO, suppress_preempt, CTLFLAG_RD,
 163     &carp_suppress_preempt, 0, "Preemption is suppressed");
 164
 165 static struct carpstats carpstats;
 166 SYSCTL_STRUCT(_net_inet_carp, CARPCTL_STATS, stats, CTLFLAG_RW,
 167     &carpstats, carpstats,
 168     "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 169
 170 #define CARP_LOG(...)   do {                            \
 171         if (carp_opts[CARPCTL_LOG] > 0)                 \
 172                 log(LOG_INFO, __VA_ARGS__);             \
 173 } while (0)
 174
 175 #define CARP_DEBUG(...) do {                            \
 176         if (carp_opts[CARPCTL_LOG] > 1)                 \
 177                 log(LOG_DEBUG, __VA_ARGS__);            \
 178 } while (0)
 179
 180 static struct lwkt_token carp_tok = LWKT_TOKEN_INITIALIZER(carp_token);
 181
 182 static void     carp_hmac_prepare(struct carp_softc *);
 183 static void     carp_hmac_generate(struct carp_softc *, uint32_t *,
 184                     unsigned char *);
 185 static int      carp_hmac_verify(struct carp_softc *, uint32_t *,
 186                     unsigned char *);
 187 static void     carp_setroute(struct carp_softc *, int);
 188 static void     carp_proto_input_c(struct carp_softc *, struct mbuf *,
 189                     struct carp_header *, sa_family_t);
 190 static int      carp_clone_create(struct if_clone *, int, caddr_t);
 191 static int      carp_clone_destroy(struct ifnet *);
 192 static void     carp_detach(struct carp_softc *, int, boolean_t);
 193 static void     carp_prepare_ad(struct carp_softc *, struct carp_header *);
 194 static void     carp_send_ad_all(void);
 195 static void     carp_send_ad_timeout(void *);
 196 static void     carp_send_ad(struct carp_softc *);
 197 static void     carp_send_arp(struct carp_softc *);
 198 static void     carp_master_down_timeout(void *);
 199 static void     carp_master_down(struct carp_softc *);
 200 static void     carp_setrun(struct carp_softc *, sa_family_t);
 201 static void     carp_set_state(struct carp_softc *, int);
 202 static struct ifnet *carp_forus(struct carp_if *, const uint8_t *);
 203
 204 static void     carp_init(void *);
 205 static int      carp_ioctl(struct ifnet *, u_long, caddr_t, struct ucred *);
 206 static int      carp_output(struct ifnet *, struct mbuf *, struct sockaddr *,
 207                     struct rtentry *);
 208 static void     carp_start(struct ifnet *);
 209 static void     carp_serialize(struct ifnet *, enum ifnet_serialize);
 210 static void     carp_deserialize(struct ifnet *, enum ifnet_serialize);
 211 static int      carp_tryserialize(struct ifnet *, enum ifnet_serialize);
 212 #ifdef INVARIANTS
 213 static void     carp_serialize_assert(struct ifnet *, enum ifnet_serialize,
 214                     boolean_t);
 215 #endif
 216
 217 static void     carp_multicast_cleanup(struct carp_softc *);
 218 static void     carp_add_addr(struct carp_softc *, struct ifaddr *);
 219 static void     carp_del_addr(struct carp_softc *, struct ifaddr *);
 220 static void     carp_config_addr(struct carp_softc *, struct ifaddr *);
 221 static void     carp_link_addrs(struct carp_softc *, struct ifnet *,
 222                     struct ifaddr *);
 223 static void     carp_unlink_addrs(struct carp_softc *, struct ifnet *,
 224                     struct ifaddr *);
 225 static void     carp_update_addrs(struct carp_softc *, struct ifaddr *);
 226
 227 static int      carp_get_vhaddr(struct carp_softc *, struct ifdrv *);
 228 static int      carp_config_vhaddr(struct carp_softc *, struct carp_vhaddr *,
 229                     struct in_ifaddr *);
 230 static int      carp_activate_vhaddr(struct carp_softc *, struct carp_vhaddr *,
 231                     struct ifnet *, struct in_ifaddr *, int);
 232 static void     carp_deactivate_vhaddr(struct carp_softc *,
 233                     struct carp_vhaddr *, boolean_t);
 234 static int      carp_addroute_vhaddr(struct carp_softc *, struct carp_vhaddr *);
 235 static void     carp_delroute_vhaddr(struct carp_softc *, struct carp_vhaddr *,
 236                     boolean_t);
 237
 238 static void     carp_sc_state(struct carp_softc *);
 239 #ifdef INET6
 240 static void     carp_send_na(struct carp_softc *);
 241 #ifdef notyet
 242 static int      carp_set_addr6(struct carp_softc *, struct sockaddr_in6 *);
 243 static int      carp_del_addr6(struct carp_softc *, struct sockaddr_in6 *);
 244 #endif
 245 static void     carp_multicast6_cleanup(struct carp_softc *);
 246 #endif
 247 static void     carp_stop(struct carp_softc *, int);
 248 static void     carp_suspend(struct carp_softc *, int);
 249
 250 static void     carp_ifaddr(void *, struct ifnet *, enum ifaddr_event,
 251                             struct ifaddr *);
 252 static void     carp_ifdetach(void *, struct ifnet *);
 253
 254 static MALLOC_DEFINE(M_CARP, "CARP", "CARP interfaces");
 255
 256 static LIST_HEAD(, carp_softc) carpif_list;
 257
 258 static struct if_clone carp_cloner =
 259 IF_CLONE_INITIALIZER(CARP_IFNAME, carp_clone_create, carp_clone_destroy,
 260                      0, IF_MAXUNIT);
 261
 262 static uint8_t  carp_etheraddr[ETHER_ADDR_LEN] = { 0, 0, 0x5e, 0, 1, 0 };
 263
 264 static eventhandler_tag carp_ifdetach_event;
 265 static eventhandler_tag carp_ifaddr_event;
 266
 267 static __inline void
 268 carp_insert_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha_new)
 269 {
 270         struct carp_vhaddr *vha;
 271         u_long new_addr, addr;
 272
 273         KKASSERT((vha_new->vha_flags & CARP_VHAF_ONLIST) == 0);
 274
 275         /*
 276          * Virtual address list is sorted; smaller one first
 277          */
 278         new_addr = ntohl(vha_new->vha_ia->ia_addr.sin_addr.s_addr);
 279
 280         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
 281                 addr = ntohl(vha->vha_ia->ia_addr.sin_addr.s_addr);
 282
 283                 if (addr > new_addr)
 284                         break;
 285         }
 286         if (vha == NULL)
 287                 TAILQ_INSERT_TAIL(&sc->sc_vha_list, vha_new, vha_link);
 288         else
 289                 TAILQ_INSERT_BEFORE(vha, vha_new, vha_link);
 290         vha_new->vha_flags |= CARP_VHAF_ONLIST;
 291 }
 292
 293 static __inline void
 294 carp_remove_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha)
 295 {
 296         KKASSERT(vha->vha_flags & CARP_VHAF_ONLIST);
 297         vha->vha_flags &= ~CARP_VHAF_ONLIST;
 298         TAILQ_REMOVE(&sc->sc_vha_list, vha, vha_link);
 299 }
 300
 301 static void
 302 carp_hmac_prepare(struct carp_softc *sc)
 303 {
 304         uint8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 305         uint8_t vhid = sc->sc_vhid & 0xff;
 306         int i;
 307 #ifdef INET6
 308         struct ifaddr_container *ifac;
 309         struct in6_addr in6;
 310 #endif
 311 #ifdef INET
 312         struct carp_vhaddr *vha;
 313 #endif
 314
 315         /* XXX: possible race here */
 316
 317         /* compute ipad from key */
 318         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 319         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 320         for (i = 0; i < sizeof(sc->sc_pad); i++)
 321                 sc->sc_pad[i] ^= 0x36;
 322
 323         /* precompute first part of inner hash */
 324         SHA1Init(&sc->sc_sha1);
 325         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 326         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 327         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 328         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 329 #ifdef INET
 330         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
 331                 SHA1Update(&sc->sc_sha1,
 332                     (const uint8_t *)&vha->vha_ia->ia_addr.sin_addr,
 333                     sizeof(struct in_addr));
 334         }
 335 #endif /* INET */
 336 #ifdef INET6
 337         TAILQ_FOREACH(ifac, &sc->sc_if.if_addrheads[mycpuid], ifa_link) {
 338                 struct ifaddr *ifa = ifac->ifa;
 339
 340                 if (ifa->ifa_addr->sa_family == AF_INET6) {
 341                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 342                         in6_clearscope(&in6);
 343                         SHA1Update(&sc->sc_sha1, (void *)&in6, sizeof(in6));
 344                 }
 345         }
 346 #endif /* INET6 */
 347
 348         /* convert ipad to opad */
 349         for (i = 0; i < sizeof(sc->sc_pad); i++)
 350                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 351 }
 352
 353 static void
 354 carp_hmac_generate(struct carp_softc *sc, uint32_t counter[2],
 355     unsigned char md[20])
 356 {
 357         SHA1_CTX sha1ctx;
 358
 359         /* fetch first half of inner hash */
 360         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 361
 362         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 363         SHA1Final(md, &sha1ctx);
 364
 365         /* outer hash */
 366         SHA1Init(&sha1ctx);
 367         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 368         SHA1Update(&sha1ctx, md, 20);
 369         SHA1Final(md, &sha1ctx);
 370 }
 371
 372 static int
 373 carp_hmac_verify(struct carp_softc *sc, uint32_t counter[2],
 374     unsigned char md[20])
 375 {
 376         unsigned char md2[20];
 377
 378         carp_hmac_generate(sc, counter, md2);
 379         return (bcmp(md, md2, sizeof(md2)));
 380 }
 381
 382 static void
 383 carp_setroute(struct carp_softc *sc, int cmd)
 384 {
 385 #ifdef INET6
 386         struct ifaddr_container *ifac;
 387 #endif
 388         struct carp_vhaddr *vha;
 389
 390         KKASSERT(cmd == RTM_DELETE || cmd == RTM_ADD);
 391
 392         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
 393                 if (vha->vha_iaback == NULL)
 394                         continue;
 395                 if (cmd == RTM_DELETE)
 396                         carp_delroute_vhaddr(sc, vha, FALSE);
 397                 else
 398                         carp_addroute_vhaddr(sc, vha);
 399         }
 400
 401 #ifdef INET6
 402         TAILQ_FOREACH(ifac, &sc->sc_if.if_addrheads[mycpuid], ifa_link) {
 403                 struct ifaddr *ifa = ifac->ifa;
 404
 405                 if (ifa->ifa_addr->sa_family == AF_INET6) {
 406                         if (cmd == RTM_ADD)
 407                                 in6_ifaddloop(ifa);
 408                         else
 409                                 in6_ifremloop(ifa);
 410                 }
 411         }
 412 #endif /* INET6 */
 413 }
 414
 415 static int
 416 carp_clone_create(struct if_clone *ifc, int unit, caddr_t param __unused)
 417 {
 418         struct carp_softc *sc;
 419         struct ifnet *ifp;
 420
 421         sc = kmalloc(sizeof(*sc), M_CARP, M_WAITOK | M_ZERO);
 422         ifp = &sc->sc_if;
 423
 424         sc->sc_suppress = 0;
 425         sc->sc_advbase = CARP_DFLTINTV;
 426         sc->sc_vhid = -1;       /* required setting */
 427         sc->sc_advskew = 0;
 428         sc->sc_init_counter = 1;
 429         sc->sc_naddrs = 0;
 430         sc->sc_naddrs6 = 0;
 431
 432         TAILQ_INIT(&sc->sc_vha_list);
 433
 434 #ifdef INET6
 435         sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL;
 436 #endif
 437
 438         callout_init_mp(&sc->sc_ad_tmo);
 439         callout_init_mp(&sc->sc_md_tmo);
 440         callout_init_mp(&sc->sc_md6_tmo);
 441
 442         if_initname(ifp, CARP_IFNAME, unit);
 443         ifp->if_softc = sc;
 444         ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
 445         ifp->if_init = carp_init;
 446         ifp->if_ioctl = carp_ioctl;
 447         ifp->if_start = carp_start;
 448         ifp->if_serialize = carp_serialize;
 449         ifp->if_deserialize = carp_deserialize;
 450         ifp->if_tryserialize = carp_tryserialize;
 451 #ifdef INVARIANTS
 452         ifp->if_serialize_assert = carp_serialize_assert;
 453 #endif
 454         ifq_set_maxlen(&ifp->if_snd, ifqmaxlen);
 455         ifq_set_ready(&ifp->if_snd);
 456
 457         ether_ifattach(ifp, carp_etheraddr, NULL);
 458
 459         ifp->if_type = IFT_CARP;
 460         ifp->if_output = carp_output;
 461
 462         carp_gettok();
 463         LIST_INSERT_HEAD(&carpif_list, sc, sc_next);
 464         carp_reltok();
 465
 466         return (0);
 467 }
 468
 469 static int
 470 carp_clone_destroy(struct ifnet *ifp)
 471 {
 472         struct carp_softc *sc = ifp->if_softc;
 473
 474         carp_gettok();
 475
 476         sc->sc_dead = 1;
 477         carp_detach(sc, 1, FALSE);
 478         LIST_REMOVE(sc, sc_next);
 479
 480         carp_reltok();
 481
 482         bpfdetach(ifp);
 483         if_detach(ifp);
 484
 485         KASSERT(sc->sc_naddrs == 0, ("certain inet address is still active\n"));
 486         kfree(sc, M_CARP);
 487
 488         return 0;
 489 }
 490
 491 static void
 492 carp_detach(struct carp_softc *sc, int detach, boolean_t del_iaback)
 493 {
 494         struct carp_if *cif;
 495
 496         carp_suspend(sc, detach);
 497
 498         carp_multicast_cleanup(sc);
 499 #ifdef INET6
 500         carp_multicast6_cleanup(sc);
 501 #endif
 502
 503         if (!sc->sc_dead && detach) {
 504                 struct carp_vhaddr *vha;
 505
 506                 TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link)
 507                         carp_deactivate_vhaddr(sc, vha, del_iaback);
 508                 KKASSERT(sc->sc_naddrs == 0);
 509         }
 510
 511         if (sc->sc_carpdev != NULL) {
 512                 cif = sc->sc_carpdev->if_carp;
 513                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
 514                 if (TAILQ_EMPTY(&cif->vhif_vrs)) {
 515                         ifpromisc(sc->sc_carpdev, 0);
 516                         sc->sc_carpdev->if_carp = NULL;
 517                         kfree(cif, M_CARP);
 518                 }
 519                 sc->sc_carpdev = NULL;
 520                 sc->sc_ia = NULL;
 521         }
 522 }
 523
 524 /* Detach an interface from the carp. */
 525 static void
 526 carp_ifdetach(void *arg __unused, struct ifnet *ifp)
 527 {
 528         struct carp_if *cif = ifp->if_carp;
 529         struct carp_softc *sc;
 530
 531         carp_gettok();
 532
 533         while (ifp->if_carp &&
 534                (sc = TAILQ_FIRST(&cif->vhif_vrs)) != NULL)
 535                 carp_detach(sc, 1, TRUE);
 536
 537         carp_reltok();
 538 }
 539
 540 /*
 541  * process input packet.
 542  * we have rearranged checks order compared to the rfc,
 543  * but it seems more efficient this way or not possible otherwise.
 544  */
 545 int
 546 carp_proto_input(struct mbuf **mp, int *offp, int proto)
 547 {
 548         struct mbuf *m = *mp;
 549         struct ip *ip = mtod(m, struct ip *);
 550         struct ifnet *ifp = m->m_pkthdr.rcvif;
 551         struct carp_header *ch;
 552         struct carp_softc *sc;
 553         int len, iphlen;
 554
 555         carp_gettok();
 556
 557         iphlen = *offp;
 558         *mp = NULL;
 559
 560         carpstats.carps_ipackets++;
 561
 562         if (!carp_opts[CARPCTL_ALLOW]) {
 563                 m_freem(m);
 564                 goto back;
 565         }
 566
 567         /* Check if received on a valid carp interface */
 568         if (ifp->if_type != IFT_CARP) {
 569                 carpstats.carps_badif++;
 570                 CARP_LOG("carp_proto_input: packet received on non-carp "
 571                     "interface: %s\n", ifp->if_xname);
 572                 m_freem(m);
 573                 goto back;
 574         }
 575
 576         if (!CARP_IS_RUNNING(ifp)) {
 577                 carpstats.carps_badif++;
 578                 CARP_LOG("carp_proto_input: packet received on stopped carp "
 579                     "interface: %s\n", ifp->if_xname);
 580                 m_freem(m);
 581                 goto back;
 582         }
 583
 584         sc = ifp->if_softc;
 585         if (sc->sc_carpdev == NULL) {
 586                 carpstats.carps_badif++;
 587                 CARP_LOG("carp_proto_input: packet received on defunc carp "
 588                     "interface: %s\n", ifp->if_xname);
 589                 m_freem(m);
 590                 goto back;
 591         }
 592
 593         if (!IN_MULTICAST(ntohl(ip->ip_dst.s_addr))) {
 594                 carpstats.carps_badif++;
 595                 CARP_LOG("carp_proto_input: non-mcast packet on "
 596                     "interface: %s\n", ifp->if_xname);
 597                 m_freem(m);
 598                 goto back;
 599         }
 600
 601         /* Verify that the IP TTL is CARP_DFLTTL. */
 602         if (ip->ip_ttl != CARP_DFLTTL) {
 603                 carpstats.carps_badttl++;
 604                 CARP_LOG("carp_proto_input: received ttl %d != %d on %s\n",
 605                     ip->ip_ttl, CARP_DFLTTL, ifp->if_xname);
 606                 m_freem(m);
 607                 goto back;
 608         }
 609
 610         /* Minimal CARP packet size */
 611         len = iphlen + sizeof(*ch);
 612
 613         /*
 614          * Verify that the received packet length is
 615          * not less than the CARP header
 616          */
 617         if (m->m_pkthdr.len < len) {
 618                 carpstats.carps_badlen++;
 619                 CARP_LOG("packet too short %d on %s\n", m->m_pkthdr.len,
 620                     ifp->if_xname);
 621                 m_freem(m);
 622                 goto back;
 623         }
 624
 625         /* Make sure that CARP header is contiguous */
 626         if (len > m->m_len) {
 627                 m = m_pullup(m, len);
 628                 if (m == NULL) {
 629                         carpstats.carps_hdrops++;
 630                         CARP_LOG("carp_proto_input: m_pullup failed\n");
 631                         goto back;
 632                 }
 633                 ip = mtod(m, struct ip *);
 634         }
 635         ch = (struct carp_header *)((uint8_t *)ip + iphlen);
 636
 637         /* Verify the CARP checksum */
 638         if (in_cksum_skip(m, len, iphlen)) {
 639                 carpstats.carps_badsum++;
 640                 CARP_LOG("carp_proto_input: checksum failed on %s\n",
 641                     ifp->if_xname);
 642                 m_freem(m);
 643                 goto back;
 644         }
 645         carp_proto_input_c(sc, m, ch, AF_INET);
 646 back:
 647         carp_reltok();
 648         return(IPPROTO_DONE);
 649 }
 650
 651 #ifdef INET6
 652 int
 653 carp6_proto_input(struct mbuf **mp, int *offp, int proto)
 654 {
 655         struct mbuf *m = *mp;
 656         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 657         struct ifnet *ifp = m->m_pkthdr.rcvif;
 658         struct carp_header *ch;
 659         struct carp_softc *sc;
 660         u_int len;
 661
 662         carp_gettok();
 663
 664         carpstats.carps_ipackets6++;
 665
 666         if (!carp_opts[CARPCTL_ALLOW]) {
 667                 m_freem(m);
 668                 goto back;
 669         }
 670
 671         /* check if received on a valid carp interface */
 672         if (ifp->if_type != IFT_CARP) {
 673                 carpstats.carps_badif++;
 674                 CARP_LOG("carp6_proto_input: packet received on non-carp "
 675                     "interface: %s\n", ifp->if_xname);
 676                 m_freem(m);
 677                 goto back;
 678         }
 679
 680         if (!CARP_IS_RUNNING(ifp)) {
 681                 carpstats.carps_badif++;
 682                 CARP_LOG("carp_proto_input: packet received on stopped carp "
 683                     "interface: %s\n", ifp->if_xname);
 684                 m_freem(m);
 685                 goto back;
 686         }
 687
 688         sc = ifp->if_softc;
 689         if (sc->sc_carpdev == NULL) {
 690                 carpstats.carps_badif++;
 691                 CARP_LOG("carp6_proto_input: packet received on defunc-carp "
 692                     "interface: %s\n", ifp->if_xname);
 693                 m_freem(m);
 694                 goto back;
 695         }
 696
 697         /* verify that the IP TTL is 255 */
 698         if (ip6->ip6_hlim != CARP_DFLTTL) {
 699                 carpstats.carps_badttl++;
 700                 CARP_LOG("carp6_proto_input: received ttl %d != 255 on %s\n",
 701                     ip6->ip6_hlim, ifp->if_xname);
 702                 m_freem(m);
 703                 goto back;
 704         }
 705
 706         /* verify that we have a complete carp packet */
 707         len = m->m_len;
 708         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 709         if (ch == NULL) {
 710                 carpstats.carps_badlen++;
 711                 CARP_LOG("carp6_proto_input: packet size %u too small\n", len);
 712                 goto back;
 713         }
 714
 715         /* verify the CARP checksum */
 716         if (in_cksum_range(m, 0, *offp, sizeof(*ch))) {
 717                 carpstats.carps_badsum++;
 718                 CARP_LOG("carp6_proto_input: checksum failed, on %s\n",
 719                     ifp->if_xname);
 720                 m_freem(m);
 721                 goto back;
 722         }
 723
 724         carp_proto_input_c(sc, m, ch, AF_INET6);
 725 back:
 726         carp_reltok();
 727         return (IPPROTO_DONE);
 728 }
 729 #endif /* INET6 */
 730
 731 static void
 732 carp_proto_input_c(struct carp_softc *sc, struct mbuf *m,
 733     struct carp_header *ch, sa_family_t af)
 734 {
 735         struct ifnet *cifp;
 736         uint64_t tmp_counter;
 737         struct timeval sc_tv, ch_tv;
 738
 739         if (sc->sc_vhid != ch->carp_vhid) {
 740                 /*
 741                  * CARP uses multicast, however, multicast packets
 742                  * are tapped to all CARP interfaces on the physical
 743                  * interface receiving the CARP packets, so we don't
 744                  * update any stats here.
 745                  */
 746                 m_freem(m);
 747                 return;
 748         }
 749         cifp = &sc->sc_if;
 750
 751         /* verify the CARP version. */
 752         if (ch->carp_version != CARP_VERSION) {
 753                 carpstats.carps_badver++;
 754                 CARP_LOG("%s; invalid version %d\n", cifp->if_xname,
 755                          ch->carp_version);
 756                 m_freem(m);
 757                 return;
 758         }
 759
 760         /* verify the hash */
 761         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 762                 carpstats.carps_badauth++;
 763                 CARP_LOG("%s: incorrect hash\n", cifp->if_xname);
 764                 m_freem(m);
 765                 return;
 766         }
 767
 768         tmp_counter = ntohl(ch->carp_counter[0]);
 769         tmp_counter = tmp_counter<<32;
 770         tmp_counter += ntohl(ch->carp_counter[1]);
 771
 772         /* XXX Replay protection goes here */
 773
 774         sc->sc_init_counter = 0;
 775         sc->sc_counter = tmp_counter;
 776
 777         sc_tv.tv_sec = sc->sc_advbase;
 778         if (carp_suppress_preempt && sc->sc_advskew <  240)
 779                 sc_tv.tv_usec = 240 * 1000000 / 256;
 780         else
 781                 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256;
 782         ch_tv.tv_sec = ch->carp_advbase;
 783         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 784
 785         switch (sc->sc_state) {
 786         case INIT:
 787                 break;
 788
 789         case MASTER:
 790                 /*
 791                  * If we receive an advertisement from a master who's going to
 792                  * be more frequent than us, go into BACKUP state.
 793                  */
 794                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 795                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 796                         callout_stop(&sc->sc_ad_tmo);
 797                         CARP_DEBUG("%s: MASTER -> BACKUP "
 798                            "(more frequent advertisement received)\n",
 799                            cifp->if_xname);
 800                         carp_set_state(sc, BACKUP);
 801                         carp_setrun(sc, 0);
 802                         carp_setroute(sc, RTM_DELETE);
 803                 }
 804                 break;
 805
 806         case BACKUP:
 807                 /*
 808                  * If we're pre-empting masters who advertise slower than us,
 809                  * and this one claims to be slower, treat him as down.
 810                  */
 811                 if (carp_opts[CARPCTL_PREEMPT] &&
 812                     timevalcmp(&sc_tv, &ch_tv, <)) {
 813                         CARP_DEBUG("%s: BACKUP -> MASTER "
 814                             "(preempting a slower master)\n", cifp->if_xname);
 815                         carp_master_down(sc);
 816                         break;
 817                 }
 818
 819                 /*
 820                  *  If the master is going to advertise at such a low frequency
 821                  *  that he's guaranteed to time out, we'd might as well just
 822                  *  treat him as timed out now.
 823                  */
 824                 sc_tv.tv_sec = sc->sc_advbase * 3;
 825                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 826                         CARP_DEBUG("%s: BACKUP -> MASTER (master timed out)\n",
 827                                    cifp->if_xname);
 828                         carp_master_down(sc);
 829                         break;
 830                 }
 831
 832                 /*
 833                  * Otherwise, we reset the counter and wait for the next
 834                  * advertisement.
 835                  */
 836                 carp_setrun(sc, af);
 837                 break;
 838         }
 839         m_freem(m);
 840 }
 841
 842 struct mbuf *
 843 carp_input(void *v, struct mbuf *m)
 844 {
 845         struct carp_if *cif = v;
 846         struct ether_header *eh;
 847         struct carp_softc *sc;
 848         struct ifnet *ifp;
 849
 850         ASSERT_LWKT_TOKEN_HELD(&carp_tok);
 851
 852         eh = mtod(m, struct ether_header *);
 853
 854         ifp = carp_forus(cif, eh->ether_dhost);
 855         if (ifp != NULL) {
 856                 ether_reinput_oncpu(ifp, m, REINPUT_RUNBPF);
 857                 return NULL;
 858         }
 859
 860         if ((m->m_flags & (M_BCAST | M_MCAST)) == 0)
 861                 return m;
 862
 863         /*
 864          * XXX Should really check the list of multicast addresses
 865          * for each CARP interface _before_ copying.
 866          */
 867         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) {
 868                 struct mbuf *m0;
 869
 870                 if ((sc->sc_if.if_flags & IFF_UP) == 0)
 871                         continue;
 872
 873                 m0 = m_dup(m, MB_DONTWAIT);
 874                 if (m0 == NULL)
 875                         continue;
 876
 877                 ether_reinput_oncpu(&sc->sc_if, m0, REINPUT_RUNBPF);
 878         }
 879         return m;
 880 }
 881
 882 static void
 883 carp_prepare_ad(struct carp_softc *sc, struct carp_header *ch)
 884 {
 885         if (sc->sc_init_counter) {
 886                 /* this could also be seconds since unix epoch */
 887                 sc->sc_counter = karc4random();
 888                 sc->sc_counter = sc->sc_counter << 32;
 889                 sc->sc_counter += karc4random();
 890         } else {
 891                 sc->sc_counter++;
 892         }
 893
 894         ch->carp_counter[0] = htonl((sc->sc_counter >> 32) & 0xffffffff);
 895         ch->carp_counter[1] = htonl(sc->sc_counter & 0xffffffff);
 896
 897         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 898 }
 899
 900 static void
 901 carp_send_ad_all(void)
 902 {
 903         struct carp_softc *sc;
 904
 905         LIST_FOREACH(sc, &carpif_list, sc_next) {
 906                 if (sc->sc_carpdev == NULL)
 907                         continue;
 908
 909                 if (CARP_IS_RUNNING(&sc->sc_if) && sc->sc_state == MASTER)
 910                         carp_send_ad(sc);
 911         }
 912 }
 913
 914 static void
 915 carp_send_ad_timeout(void *xsc)
 916 {
 917         carp_gettok();
 918         carp_send_ad(xsc);
 919         carp_reltok();
 920 }
 921
 922 static void
 923 carp_send_ad(struct carp_softc *sc)
 924 {
 925         struct ifnet *cifp = &sc->sc_if;
 926         struct carp_header ch;
 927         struct timeval tv;
 928         struct carp_header *ch_ptr;
 929         struct mbuf *m;
 930         int len, advbase, advskew;
 931
 932         if (!CARP_IS_RUNNING(cifp)) {
 933                 /* Bow out */
 934                 advbase = 255;
 935                 advskew = 255;
 936         } else {
 937                 advbase = sc->sc_advbase;
 938                 if (!carp_suppress_preempt || sc->sc_advskew > 240)
 939                         advskew = sc->sc_advskew;
 940                 else
 941                         advskew = 240;
 942                 tv.tv_sec = advbase;
 943                 tv.tv_usec = advskew * 1000000 / 256;
 944         }
 945
 946         ch.carp_version = CARP_VERSION;
 947         ch.carp_type = CARP_ADVERTISEMENT;
 948         ch.carp_vhid = sc->sc_vhid;
 949         ch.carp_advbase = advbase;
 950         ch.carp_advskew = advskew;
 951         ch.carp_authlen = 7;    /* XXX DEFINE */
 952         ch.carp_pad1 = 0;       /* must be zero */
 953         ch.carp_cksum = 0;
 954
 955 #ifdef INET
 956         if (sc->sc_ia != NULL) {
 957                 struct ip *ip;
 958
 959                 MGETHDR(m, MB_DONTWAIT, MT_HEADER);
 960                 if (m == NULL) {
 961                         cifp->if_oerrors++;
 962                         carpstats.carps_onomem++;
 963                         /* XXX maybe less ? */
 964                         if (advbase != 255 || advskew != 255)
 965                                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
 966                                     carp_send_ad_timeout, sc);
 967                         return;
 968                 }
 969                 len = sizeof(*ip) + sizeof(ch);
 970                 m->m_pkthdr.len = len;
 971                 m->m_pkthdr.rcvif = NULL;
 972                 m->m_len = len;
 973                 MH_ALIGN(m, m->m_len);
 974                 m->m_flags |= M_MCAST;
 975                 ip = mtod(m, struct ip *);
 976                 ip->ip_v = IPVERSION;
 977                 ip->ip_hl = sizeof(*ip) >> 2;
 978                 ip->ip_tos = IPTOS_LOWDELAY;
 979                 ip->ip_len = len;
 980                 ip->ip_id = ip_newid();
 981                 ip->ip_off = IP_DF;
 982                 ip->ip_ttl = CARP_DFLTTL;
 983                 ip->ip_p = IPPROTO_CARP;
 984                 ip->ip_sum = 0;
 985                 ip->ip_src = sc->sc_ia->ia_addr.sin_addr;
 986                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 987
 988                 ch_ptr = (struct carp_header *)(&ip[1]);
 989                 bcopy(&ch, ch_ptr, sizeof(ch));
 990                 carp_prepare_ad(sc, ch_ptr);
 991                 ch_ptr->carp_cksum = in_cksum_skip(m, len, sizeof(*ip));
 992
 993                 getmicrotime(&cifp->if_lastchange);
 994                 cifp->if_opackets++;
 995                 cifp->if_obytes += len;
 996                 carpstats.carps_opackets++;
 997
 998                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, NULL)) {
 999                         cifp->if_oerrors++;
1000                         if (sc->sc_sendad_errors < INT_MAX)
1001                                 sc->sc_sendad_errors++;
1002                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
1003                                 carp_suppress_preempt++;
1004                                 if (carp_suppress_preempt == 1) {
1005                                         carp_send_ad_all();
1006                                 }
1007                         }
1008                         sc->sc_sendad_success = 0;
1009                 } else {
1010                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
1011                                 if (++sc->sc_sendad_success >=
1012                                     CARP_SENDAD_MIN_SUCCESS) {
1013                                         carp_suppress_preempt--;
1014                                         sc->sc_sendad_errors = 0;
1015                                 }
1016                         } else {
1017                                 sc->sc_sendad_errors = 0;
1018                         }
1019                 }
1020         }
1021 #endif /* INET */
1022 #ifdef INET6
1023         if (sc->sc_ia6) {
1024                 struct ip6_hdr *ip6;
1025
1026                 MGETHDR(m, MB_DONTWAIT, MT_HEADER);
1027                 if (m == NULL) {
1028                         cifp->if_oerrors++;
1029                         carpstats.carps_onomem++;
1030                         /* XXX maybe less ? */
1031                         if (advbase != 255 || advskew != 255)
1032                                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
1033                                     carp_send_ad_timeout, sc);
1034                         return;
1035                 }
1036                 len = sizeof(*ip6) + sizeof(ch);
1037                 m->m_pkthdr.len = len;
1038                 m->m_pkthdr.rcvif = NULL;
1039                 m->m_len = len;
1040                 MH_ALIGN(m, m->m_len);
1041                 m->m_flags |= M_MCAST;
1042                 ip6 = mtod(m, struct ip6_hdr *);
1043                 bzero(ip6, sizeof(*ip6));
1044                 ip6->ip6_vfc |= IPV6_VERSION;
1045                 ip6->ip6_hlim = CARP_DFLTTL;
1046                 ip6->ip6_nxt = IPPROTO_CARP;
1047                 bcopy(&sc->sc_ia6->ia_addr.sin6_addr, &ip6->ip6_src,
1048                     sizeof(struct in6_addr));
1049                 /* set the multicast destination */
1050
1051                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
1052                 ip6->ip6_dst.s6_addr8[15] = 0x12;
1053                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
1054                         cifp->if_oerrors++;
1055                         m_freem(m);
1056                         CARP_LOG("%s: in6_setscope failed\n", __func__);
1057                         return;
1058                 }
1059
1060                 ch_ptr = (struct carp_header *)(&ip6[1]);
1061                 bcopy(&ch, ch_ptr, sizeof(ch));
1062                 carp_prepare_ad(sc, ch_ptr);
1063                 ch_ptr->carp_cksum = in_cksum_skip(m, len, sizeof(*ip6));
1064
1065                 getmicrotime(&cifp->if_lastchange);
1066                 cifp->if_opackets++;
1067                 cifp->if_obytes += len;
1068                 carpstats.carps_opackets6++;
1069
1070                 if (ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL)) {
1071                         cifp->if_oerrors++;
1072                         if (sc->sc_sendad_errors < INT_MAX)
1073                                 sc->sc_sendad_errors++;
1074                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
1075                                 carp_suppress_preempt++;
1076                                 if (carp_suppress_preempt == 1) {
1077                                         carp_send_ad_all();
1078                                 }
1079                         }
1080                         sc->sc_sendad_success = 0;
1081                 } else {
1082                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
1083                                 if (++sc->sc_sendad_success >=
1084                                     CARP_SENDAD_MIN_SUCCESS) {
1085                                         carp_suppress_preempt--;
1086                                         sc->sc_sendad_errors = 0;
1087                                 }
1088                         } else {
1089                                 sc->sc_sendad_errors = 0;
1090                         }
1091                 }
1092         }
1093 #endif /* INET6 */
1094
1095         if (advbase != 255 || advskew != 255)
1096                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
1097                     carp_send_ad_timeout, sc);
1098 }
1099
1100 /*
1101  * Broadcast a gratuitous ARP request containing
1102  * the virtual router MAC address for each IP address
1103  * associated with the virtual router.
1104  */
1105 static void
1106 carp_send_arp(struct carp_softc *sc)
1107 {
1108         const struct carp_vhaddr *vha;
1109
1110         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
1111                 if (vha->vha_iaback == NULL)
1112                         continue;
1113                 arp_gratuitous(&sc->sc_if, &vha->vha_ia->ia_ifa);
1114         }
1115 }
1116
1117 #ifdef INET6
1118 static void
1119 carp_send_na(struct carp_softc *sc)
1120 {
1121         struct ifaddr_container *ifac;
1122         struct in6_addr *in6;
1123         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1124
1125         TAILQ_FOREACH(ifac, &sc->sc_if.if_addrheads[mycpuid], ifa_link) {
1126                 struct ifaddr *ifa = ifac->ifa;
1127
1128                 if (ifa->ifa_addr->sa_family != AF_INET6)
1129                         continue;
1130
1131                 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr;
1132                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1133                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1134                 DELAY(1000);    /* XXX */
1135         }
1136 }
1137 #endif /* INET6 */
1138
1139 static __inline const struct carp_vhaddr *
1140 carp_find_addr(const struct carp_softc *sc, const struct in_addr *addr)
1141 {
1142         struct carp_vhaddr *vha;
1143
1144         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
1145                 if (vha->vha_iaback == NULL)
1146                         continue;
1147
1148                 if (vha->vha_ia->ia_addr.sin_addr.s_addr == addr->s_addr)
1149                         return vha;
1150         }
1151         return NULL;
1152 }
1153
1154 #ifdef notyet
1155 static int
1156 carp_iamatch_balance(const struct carp_if *cif, const struct in_addr *itaddr,
1157                      const struct in_addr *isaddr, uint8_t **enaddr)
1158 {
1159         const struct carp_softc *vh;
1160         int index, count = 0;
1161
1162         /*
1163          * XXX proof of concept implementation.
1164          * We use the source ip to decide which virtual host should
1165          * handle the request. If we're master of that virtual host,
1166          * then we respond, otherwise, just drop the arp packet on
1167          * the floor.
1168          */
1169
1170         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1171                 if (!CARP_IS_RUNNING(&vh->sc_if))
1172                         continue;
1173
1174                 if (carp_find_addr(vh, itaddr) != NULL)
1175                         count++;
1176         }
1177         if (count == 0)
1178                 return 0;
1179
1180         /* this should be a hash, like pf_hash() */
1181         index = ntohl(isaddr->s_addr) % count;
1182         count = 0;
1183
1184         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1185                 if (!CARP_IS_RUNNING(&vh->sc_if))
1186                         continue;
1187
1188                 if (carp_find_addr(vh, itaddr) == NULL)
1189                         continue;
1190
1191                 if (count == index) {
1192                         if (vh->sc_state == MASTER) {
1193                                 *enaddr = IF_LLADDR(&vh->sc_if);
1194                                 return 1;
1195                         } else {
1196                                 return 0;
1197                         }
1198                 }
1199                 count++;
1200         }
1201         return 0;
1202 }
1203 #endif
1204
1205 int
1206 carp_iamatch(const struct in_ifaddr *ia)
1207 {
1208         const struct carp_softc *sc = ia->ia_ifp->if_softc;
1209
1210         ASSERT_LWKT_TOKEN_HELD(&carp_tok);
1211
1212 #ifdef notyet
1213         if (carp_opts[CARPCTL_ARPBALANCE])
1214                 return carp_iamatch_balance(cif, itaddr, isaddr, enaddr);
1215 #endif
1216
1217         if (!CARP_IS_RUNNING(&sc->sc_if) || sc->sc_state != MASTER)
1218                 return 0;
1219
1220         return 1;
1221 }
1222
1223 #ifdef INET6
1224 struct ifaddr *
1225 carp_iamatch6(void *v, struct in6_addr *taddr)
1226 {
1227         struct carp_if *cif = v;
1228         struct carp_softc *vh;
1229
1230         ASSERT_LWKT_TOKEN_HELD(&carp_tok);
1231
1232         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1233                 struct ifaddr_container *ifac;
1234
1235                 TAILQ_FOREACH(ifac, &vh->sc_if.if_addrheads[mycpuid],
1236                               ifa_link) {
1237                         struct ifaddr *ifa = ifac->ifa;
1238
1239                         if (IN6_ARE_ADDR_EQUAL(taddr,
1240                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1241                             CARP_IS_RUNNING(&vh->sc_if) &&
1242                             vh->sc_state == MASTER) {
1243                                 return (ifa);
1244                         }
1245                 }
1246         }
1247         return (NULL);
1248 }
1249
1250 void *
1251 carp_macmatch6(void *v, struct mbuf *m, const struct in6_addr *taddr)
1252 {
1253         struct m_tag *mtag;
1254         struct carp_if *cif = v;
1255         struct carp_softc *sc;
1256
1257         ASSERT_LWKT_TOKEN_HELD(&carp_tok);
1258
1259         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) {
1260                 struct ifaddr_container *ifac;
1261
1262                 TAILQ_FOREACH(ifac, &sc->sc_if.if_addrheads[mycpuid],
1263                               ifa_link) {
1264                         struct ifaddr *ifa = ifac->ifa;
1265
1266                         if (IN6_ARE_ADDR_EQUAL(taddr,
1267                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1268                             CARP_IS_RUNNING(&sc->sc_if)) {
1269                                 struct ifnet *ifp = &sc->sc_if;
1270
1271                                 mtag = m_tag_get(PACKET_TAG_CARP,
1272                                     sizeof(struct ifnet *), MB_DONTWAIT);
1273                                 if (mtag == NULL) {
1274                                         /* better a bit than nothing */
1275                                         return (IF_LLADDR(ifp));
1276                                 }
1277                                 bcopy(&ifp, (caddr_t)(mtag + 1),
1278                                     sizeof(struct ifnet *));
1279                                 m_tag_prepend(m, mtag);
1280
1281                                 return (IF_LLADDR(ifp));
1282                         }
1283                 }
1284         }
1285         return (NULL);
1286 }
1287 #endif
1288
1289 static struct ifnet *
1290 carp_forus(struct carp_if *cif, const uint8_t *dhost)
1291 {
1292         struct carp_softc *sc;
1293
1294         ASSERT_LWKT_TOKEN_HELD(&carp_tok);
1295
1296         if (memcmp(dhost, carp_etheraddr, ETHER_ADDR_LEN - 1) != 0)
1297                 return NULL;
1298
1299         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) {
1300                 struct ifnet *ifp = &sc->sc_if;
1301
1302                 if (CARP_IS_RUNNING(ifp) && sc->sc_state == MASTER &&
1303                     !bcmp(dhost, IF_LLADDR(ifp), ETHER_ADDR_LEN))
1304                         return ifp;
1305         }
1306         return NULL;
1307 }
1308
1309 static void
1310 carp_master_down_timeout(void *xsc)
1311 {
1312         struct carp_softc *sc = xsc;
1313
1314         CARP_DEBUG("%s: BACKUP -> MASTER (master timed out)\n",
1315                    sc->sc_if.if_xname);
1316         carp_gettok();
1317         carp_master_down(sc);
1318         carp_reltok();
1319 }
1320
1321 static void
1322 carp_master_down(struct carp_softc *sc)
1323 {
1324         switch (sc->sc_state) {
1325         case INIT:
1326                 kprintf("%s: master_down event in INIT state\n",
1327                         sc->sc_if.if_xname);
1328                 break;
1329
1330         case MASTER:
1331                 break;
1332
1333         case BACKUP:
1334                 carp_set_state(sc, MASTER);
1335                 carp_send_ad(sc);
1336                 carp_send_arp(sc);
1337 #ifdef INET6
1338                 carp_send_na(sc);
1339 #endif /* INET6 */
1340                 carp_setrun(sc, 0);
1341                 carp_setroute(sc, RTM_ADD);
1342                 break;
1343         }
1344 }
1345
1346 /*
1347  * When in backup state, af indicates whether to reset the master down timer
1348  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1349  */
1350 static void
1351 carp_setrun(struct carp_softc *sc, sa_family_t af)
1352 {
1353         struct ifnet *cifp = &sc->sc_if;
1354         struct timeval tv;
1355
1356         if (sc->sc_carpdev == NULL) {
1357                 carp_set_state(sc, INIT);
1358                 return;
1359         }
1360
1361         if ((cifp->if_flags & IFF_RUNNING) && sc->sc_vhid > 0 &&
1362             (sc->sc_naddrs || sc->sc_naddrs6)) {
1363                 /* Nothing */
1364         } else {
1365                 carp_setroute(sc, RTM_DELETE);
1366                 return;
1367         }
1368
1369         switch (sc->sc_state) {
1370         case INIT:
1371                 if (carp_opts[CARPCTL_PREEMPT] && !carp_suppress_preempt) {
1372                         carp_send_ad(sc);
1373                         carp_send_arp(sc);
1374 #ifdef INET6
1375                         carp_send_na(sc);
1376 #endif /* INET6 */
1377                         CARP_DEBUG("%s: INIT -> MASTER (preempting)\n",
1378                                    cifp->if_xname);
1379                         carp_set_state(sc, MASTER);
1380                         carp_setroute(sc, RTM_ADD);
1381                 } else {
1382                         CARP_DEBUG("%s: INIT -> BACKUP\n", cifp->if_xname);
1383                         carp_set_state(sc, BACKUP);
1384                         carp_setroute(sc, RTM_DELETE);
1385                         carp_setrun(sc, 0);
1386                 }
1387                 break;
1388
1389         case BACKUP:
1390                 callout_stop(&sc->sc_ad_tmo);
1391                 tv.tv_sec = 3 * sc->sc_advbase;
1392                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1393                 switch (af) {
1394 #ifdef INET
1395                 case AF_INET:
1396                         callout_reset(&sc->sc_md_tmo, tvtohz_high(&tv),
1397                             carp_master_down_timeout, sc);
1398                         break;
1399 #endif /* INET */
1400 #ifdef INET6
1401                 case AF_INET6:
1402                         callout_reset(&sc->sc_md6_tmo, tvtohz_high(&tv),
1403                             carp_master_down_timeout, sc);
1404                         break;
1405 #endif /* INET6 */
1406                 default:
1407                         if (sc->sc_naddrs)
1408                                 callout_reset(&sc->sc_md_tmo, tvtohz_high(&tv),
1409                                     carp_master_down_timeout, sc);
1410                         if (sc->sc_naddrs6)
1411                                 callout_reset(&sc->sc_md6_tmo, tvtohz_high(&tv),
1412                                     carp_master_down_timeout, sc);
1413                         break;
1414                 }
1415                 break;
1416
1417         case MASTER:
1418                 tv.tv_sec = sc->sc_advbase;
1419                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1420                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
1421                     carp_send_ad_timeout, sc);
1422                 break;
1423         }
1424 }
1425
1426 static void
1427 carp_multicast_cleanup(struct carp_softc *sc)
1428 {
1429         struct ip_moptions *imo = &sc->sc_imo;
1430
1431         if (imo->imo_num_memberships == 0)
1432                 return;
1433         KKASSERT(imo->imo_num_memberships == 1);
1434
1435         in_delmulti(imo->imo_membership[0]);
1436         imo->imo_membership[0] = NULL;
1437         imo->imo_num_memberships = 0;
1438         imo->imo_multicast_ifp = NULL;
1439 }
1440
1441 #ifdef INET6
1442 static void
1443 carp_multicast6_cleanup(struct carp_softc *sc)
1444 {
1445         struct ip6_moptions *im6o = &sc->sc_im6o;
1446
1447         while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1448                 struct in6_multi_mship *imm =
1449                     LIST_FIRST(&im6o->im6o_memberships);
1450
1451                 LIST_REMOVE(imm, i6mm_chain);
1452                 in6_leavegroup(imm);
1453         }
1454         im6o->im6o_multicast_ifp = NULL;
1455 }
1456 #endif
1457
1458 static int
1459 carp_get_vhaddr(struct carp_softc *sc, struct ifdrv *ifd)
1460 {
1461         const struct carp_vhaddr *vha;
1462         struct ifcarpvhaddr *carpa, *carpa0;
1463         int count, len, error;
1464
1465         count = 0;
1466         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link)
1467                 ++count;
1468
1469         if (ifd->ifd_len == 0) {
1470                 ifd->ifd_len = count * sizeof(*carpa);
1471                 return 0;
1472         } else if (count == 0 || ifd->ifd_len < sizeof(*carpa)) {
1473                 ifd->ifd_len = 0;
1474                 return 0;
1475         }
1476         len = min(ifd->ifd_len, sizeof(*carpa) * count);
1477         KKASSERT(len >= sizeof(*carpa));
1478
1479         carpa0 = carpa = kmalloc(len, M_TEMP, M_WAITOK | M_NULLOK | M_ZERO);
1480         if (carpa == NULL)
1481                 return ENOMEM;
1482
1483         count = 0;
1484         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
1485                 if (len < sizeof(*carpa))
1486                         break;
1487
1488                 carpa->carpa_flags = vha->vha_flags;
1489                 carpa->carpa_addr.sin_family = AF_INET;
1490                 carpa->carpa_addr.sin_addr = vha->vha_ia->ia_addr.sin_addr;
1491
1492                 carpa->carpa_baddr.sin_family = AF_INET;
1493                 if (vha->vha_iaback == NULL) {
1494                         carpa->carpa_baddr.sin_addr.s_addr = INADDR_ANY;
1495                 } else {
1496                         carpa->carpa_baddr.sin_addr =
1497                         vha->vha_iaback->ia_addr.sin_addr;
1498                 }
1499
1500                 ++carpa;
1501                 ++count;
1502                 len -= sizeof(*carpa);
1503         }
1504         ifd->ifd_len = sizeof(*carpa) * count;
1505         KKASSERT(ifd->ifd_len > 0);
1506
1507         error = copyout(carpa0, ifd->ifd_data, ifd->ifd_len);
1508         kfree(carpa0, M_TEMP);
1509         return error;
1510 }
1511
1512 static int
1513 carp_config_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha,
1514     struct in_ifaddr *ia_del)
1515 {
1516         struct ifnet *ifp;
1517         struct in_ifaddr *ia_if;
1518         struct in_ifaddr_container *iac;
1519         const struct sockaddr_in *sin;
1520         u_long iaddr;
1521         int own;
1522
1523         KKASSERT(vha->vha_ia != NULL);
1524
1525         sin = &vha->vha_ia->ia_addr;
1526         iaddr = ntohl(sin->sin_addr.s_addr);
1527
1528         ia_if = NULL;
1529         own = 0;
1530         TAILQ_FOREACH(iac, &in_ifaddrheads[mycpuid], ia_link) {
1531                 struct in_ifaddr *ia = iac->ia;
1532
1533                 if (ia == ia_del)
1534                         continue;
1535
1536                 if (ia->ia_ifp->if_type == IFT_CARP)
1537                         continue;
1538
1539                 if ((ia->ia_ifp->if_flags & IFF_UP) == 0)
1540                         continue;
1541
1542                 /* and, yeah, we need a multicast-capable iface too */
1543                 if ((ia->ia_ifp->if_flags & IFF_MULTICAST) == 0)
1544                         continue;
1545
1546                 if ((iaddr & ia->ia_subnetmask) == ia->ia_subnet) {
1547                         if (sin->sin_addr.s_addr ==
1548                             ia->ia_addr.sin_addr.s_addr)
1549                                 own = 1;
1550                         if (ia_if == NULL)
1551                                 ia_if = ia;
1552                         else if (sc->sc_carpdev != NULL &&
1553                                  sc->sc_carpdev == ia->ia_ifp)
1554                                 ia_if = ia;
1555                 }
1556         }
1557
1558         carp_deactivate_vhaddr(sc, vha, FALSE);
1559         if (!ia_if)
1560                 return ENOENT;
1561
1562         ifp = ia_if->ia_ifp;
1563
1564         /* XXX Don't allow parent iface to be changed */
1565         if (sc->sc_carpdev != NULL && sc->sc_carpdev != ifp)
1566                 return EEXIST;
1567
1568         return carp_activate_vhaddr(sc, vha, ifp, ia_if, own);
1569 }
1570
1571 static void
1572 carp_add_addr(struct carp_softc *sc, struct ifaddr *carp_ifa)
1573 {
1574         struct carp_vhaddr *vha_new;
1575         struct in_ifaddr *carp_ia;
1576 #ifdef INVARIANTS
1577         struct carp_vhaddr *vha;
1578 #endif
1579
1580         KKASSERT(carp_ifa->ifa_addr->sa_family == AF_INET);
1581         carp_ia = ifatoia(carp_ifa);
1582
1583 #ifdef INVARIANTS
1584         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link)
1585                 KKASSERT(vha->vha_ia != NULL && vha->vha_ia != carp_ia);
1586 #endif
1587
1588         vha_new = kmalloc(sizeof(*vha_new), M_CARP, M_WAITOK | M_ZERO);
1589         vha_new->vha_ia = carp_ia;
1590         carp_insert_vhaddr(sc, vha_new);
1591
1592         if (carp_config_vhaddr(sc, vha_new, NULL) != 0) {
1593                 /*
1594                  * If the above configuration fails, it may only mean
1595                  * that the new address is problematic.  However, the
1596                  * carp(4) interface may already have several working
1597                  * addresses.  Since the expected behaviour of
1598                  * SIOC[AS]IFADDR is to put the NIC into working state,
1599                  * we try starting the state machine manually here with
1600                  * the hope that the carp(4)'s previously working
1601                  * addresses still could be brought up.
1602                  */
1603                 carp_hmac_prepare(sc);
1604                 carp_set_state(sc, INIT);
1605                 carp_setrun(sc, 0);
1606         }
1607 }
1608
1609 static void
1610 carp_del_addr(struct carp_softc *sc, struct ifaddr *carp_ifa)
1611 {
1612         struct carp_vhaddr *vha;
1613         struct in_ifaddr *carp_ia;
1614
1615         KKASSERT(carp_ifa->ifa_addr->sa_family == AF_INET);
1616         carp_ia = ifatoia(carp_ifa);
1617
1618         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
1619                 KKASSERT(vha->vha_ia != NULL);
1620                 if (vha->vha_ia == carp_ia)
1621                         break;
1622         }
1623         KASSERT(vha != NULL, ("no corresponding vhaddr %p\n", carp_ifa));
1624
1625         /*
1626          * Remove the vhaddr from the list before deactivating
1627          * the vhaddr, so that the HMAC could be correctly
1628          * updated in carp_deactivate_vhaddr()
1629          */
1630         carp_remove_vhaddr(sc, vha);
1631
1632         carp_deactivate_vhaddr(sc, vha, FALSE);
1633         kfree(vha, M_CARP);
1634 }
1635
1636 static void
1637 carp_config_addr(struct carp_softc *sc, struct ifaddr *carp_ifa)
1638 {
1639         struct carp_vhaddr *vha;
1640         struct in_ifaddr *carp_ia;
1641
1642         KKASSERT(carp_ifa->ifa_addr->sa_family == AF_INET);
1643         carp_ia = ifatoia(carp_ifa);
1644
1645         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
1646                 KKASSERT(vha->vha_ia != NULL);
1647                 if (vha->vha_ia == carp_ia)
1648                         break;
1649         }
1650         KASSERT(vha != NULL, ("no corresponding vhaddr %p\n", carp_ifa));
1651
1652         /* Remove then reinsert, to keep the vhaddr list sorted */
1653         carp_remove_vhaddr(sc, vha);
1654         carp_insert_vhaddr(sc, vha);
1655
1656         if (carp_config_vhaddr(sc, vha, NULL) != 0) {
1657                 /* See the comment in carp_add_addr() */
1658                 carp_hmac_prepare(sc);
1659                 carp_set_state(sc, INIT);
1660                 carp_setrun(sc, 0);
1661         }
1662 }
1663
1664 #ifdef notyet
1665
1666 #ifdef INET6
1667 static int
1668 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1669 {
1670         struct ifnet *ifp;
1671         struct carp_if *cif;
1672         struct in6_ifaddr *ia, *ia_if;
1673         struct ip6_moptions *im6o = &sc->sc_im6o;
1674         struct in6_multi_mship *imm;
1675         struct in6_addr in6;
1676         int own, error;
1677
1678         if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
1679                 carp_setrun(sc, 0);
1680                 return (0);
1681         }
1682
1683         /* we have to do it by hands to check we won't match on us */
1684         ia_if = NULL; own = 0;
1685         for (ia = in6_ifaddr; ia; ia = ia->ia_next) {
1686                 int i;
1687
1688                 for (i = 0; i < 4; i++) {
1689                         if ((sin6->sin6_addr.s6_addr32[i] &
1690                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]) !=
1691                             (ia->ia_addr.sin6_addr.s6_addr32[i] &
1692                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]))
1693                                 break;
1694                 }
1695                 /* and, yeah, we need a multicast-capable iface too */
1696                 if (ia->ia_ifp != &sc->sc_if &&
1697                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1698                     (i == 4)) {
1699                         if (!ia_if)
1700                                 ia_if = ia;
1701                         if (IN6_ARE_ADDR_EQUAL(&sin6->sin6_addr,
1702                             &ia->ia_addr.sin6_addr))
1703                                 own++;
1704                 }
1705         }
1706
1707         if (!ia_if)
1708                 return (EADDRNOTAVAIL);
1709         ia = ia_if;
1710         ifp = ia->ia_ifp;
1711
1712         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1713             (im6o->im6o_multicast_ifp && im6o->im6o_multicast_ifp != ifp))
1714                 return (EADDRNOTAVAIL);
1715
1716         if (!sc->sc_naddrs6) {
1717                 im6o->im6o_multicast_ifp = ifp;
1718
1719                 /* join CARP multicast address */
1720                 bzero(&in6, sizeof(in6));
1721                 in6.s6_addr16[0] = htons(0xff02);
1722                 in6.s6_addr8[15] = 0x12;
1723                 if (in6_setscope(&in6, ifp, NULL) != 0)
1724                         goto cleanup;
1725                 if ((imm = in6_joingroup(ifp, &in6, &error)) == NULL)
1726                         goto cleanup;
1727                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1728
1729                 /* join solicited multicast address */
1730                 bzero(&in6, sizeof(in6));
1731                 in6.s6_addr16[0] = htons(0xff02);
1732                 in6.s6_addr32[1] = 0;
1733                 in6.s6_addr32[2] = htonl(1);
1734                 in6.s6_addr32[3] = sin6->sin6_addr.s6_addr32[3];
1735                 in6.s6_addr8[12] = 0xff;
1736                 if (in6_setscope(&in6, ifp, NULL) != 0)
1737                         goto cleanup;
1738                 if ((imm = in6_joingroup(ifp, &in6, &error)) == NULL)
1739                         goto cleanup;
1740                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
1741         }
1742
1743         if (!ifp->if_carp) {
1744                 cif = kmalloc(sizeof(*cif), M_CARP, M_WAITOK | M_ZERO);
1745
1746                 if ((error = ifpromisc(ifp, 1))) {
1747                         kfree(cif, M_CARP);
1748                         goto cleanup;
1749                 }
1750
1751                 TAILQ_INIT(&cif->vhif_vrs);
1752                 ifp->if_carp = cif;
1753         } else {
1754                 struct carp_softc *vr;
1755
1756                 cif = ifp->if_carp;
1757                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1758                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1759                                 error = EINVAL;
1760                                 goto cleanup;
1761                         }
1762                 }
1763         }
1764         sc->sc_ia6 = ia;
1765         sc->sc_carpdev = ifp;
1766
1767         { /* XXX prevent endless loop if already in queue */
1768         struct carp_softc *vr, *after = NULL;
1769         int myself = 0;
1770         cif = ifp->if_carp;
1771
1772         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1773                 if (vr == sc)
1774                         myself = 1;
1775                 if (vr->sc_vhid < sc->sc_vhid)
1776                         after = vr;
1777         }
1778
1779         if (!myself) {
1780                 /* We're trying to keep things in order */
1781                 if (after == NULL)
1782                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1783                 else
1784                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1785         }
1786         }
1787
1788         sc->sc_naddrs6++;
1789         if (own)
1790                 sc->sc_advskew = 0;
1791         carp_sc_state(sc);
1792         carp_setrun(sc, 0);
1793
1794         return (0);
1795
1796 cleanup:
1797         /* clean up multicast memberships */
1798         if (!sc->sc_naddrs6) {
1799                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1800                         imm = LIST_FIRST(&im6o->im6o_memberships);
1801                         LIST_REMOVE(imm, i6mm_chain);
1802                         in6_leavegroup(imm);
1803                 }
1804         }
1805         return (error);
1806 }
1807
1808 static int
1809 carp_del_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1810 {
1811         int error = 0;
1812
1813         if (!--sc->sc_naddrs6) {
1814                 struct carp_if *cif = sc->sc_carpdev->if_carp;
1815                 struct ip6_moptions *im6o = &sc->sc_im6o;
1816
1817                 callout_stop(&sc->sc_ad_tmo);
1818                 sc->sc_vhid = -1;
1819                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1820                         struct in6_multi_mship *imm =
1821                             LIST_FIRST(&im6o->im6o_memberships);
1822
1823                         LIST_REMOVE(imm, i6mm_chain);
1824                         in6_leavegroup(imm);
1825                 }
1826                 im6o->im6o_multicast_ifp = NULL;
1827                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1828                 if (TAILQ_EMPTY(&cif->vhif_vrs)) {
1829                         sc->sc_carpdev->if_carp = NULL;
1830                         kfree(cif, M_IFADDR);
1831                 }
1832         }
1833         return (error);
1834 }
1835 #endif /* INET6 */
1836
1837 #endif
1838
1839 static int
1840 carp_ioctl(struct ifnet *ifp, u_long cmd, caddr_t addr, struct ucred *cr)
1841 {
1842         struct carp_softc *sc = ifp->if_softc, *vr;
1843         struct carpreq carpr;
1844         struct ifaddr *ifa;
1845         struct ifreq *ifr;
1846         struct ifaliasreq *ifra;
1847         struct ifdrv *ifd;
1848         char devname[IFNAMSIZ];
1849         int error = 0;
1850
1851         carp_gettok();
1852
1853         ifa = (struct ifaddr *)addr;
1854         ifra = (struct ifaliasreq *)addr;
1855         ifr = (struct ifreq *)addr;
1856         ifd = (struct ifdrv *)addr;
1857
1858         switch (cmd) {
1859         case SIOCSIFFLAGS:
1860                 if (ifp->if_flags & IFF_UP) {
1861                         if ((ifp->if_flags & IFF_RUNNING) == 0)
1862                                 carp_init(sc);
1863                 } else if (ifp->if_flags & IFF_RUNNING) {
1864                         carp_stop(sc, 0);
1865                 }
1866                 break;
1867
1868         case SIOCSVH:
1869                 error = priv_check_cred(cr, PRIV_ROOT, NULL_CRED_OKAY);
1870                 if (error)
1871                         break;
1872                 error = copyin(ifr->ifr_data, &carpr, sizeof(carpr));
1873                 if (error)
1874                         break;
1875
1876                 error = 1;
1877                 if ((ifp->if_flags & IFF_RUNNING) &&
1878                     sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) {
1879                         switch (carpr.carpr_state) {
1880                         case BACKUP:
1881                                 callout_stop(&sc->sc_ad_tmo);
1882                                 carp_set_state(sc, BACKUP);
1883                                 carp_setrun(sc, 0);
1884                                 carp_setroute(sc, RTM_DELETE);
1885                                 break;
1886
1887                         case MASTER:
1888                                 carp_master_down(sc);
1889                                 break;
1890
1891                         default:
1892                                 break;
1893                         }
1894                 }
1895                 if (carpr.carpr_vhid > 0) {
1896                         if (carpr.carpr_vhid > 255) {
1897                                 error = EINVAL;
1898                                 break;
1899                         }
1900                         if (sc->sc_carpdev) {
1901                                 struct carp_if *cif = sc->sc_carpdev->if_carp;
1902
1903                                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1904                                         if (vr != sc &&
1905                                             vr->sc_vhid == carpr.carpr_vhid) {
1906                                                 carp_reltok();
1907                                                 return EEXIST;
1908                                         }
1909                                 }
1910                         }
1911                         sc->sc_vhid = carpr.carpr_vhid;
1912
1913                         IF_LLADDR(ifp)[5] = sc->sc_vhid;
1914                         bcopy(IF_LLADDR(ifp), sc->arpcom.ac_enaddr,
1915                             ETHER_ADDR_LEN);
1916
1917                         error--;
1918                 }
1919                 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) {
1920                         if (carpr.carpr_advskew >= 255) {
1921                                 error = EINVAL;
1922                                 break;
1923                         }
1924                         if (carpr.carpr_advbase > 255) {
1925                                 error = EINVAL;
1926                                 break;
1927                         }
1928                         sc->sc_advbase = carpr.carpr_advbase;
1929                         sc->sc_advskew = carpr.carpr_advskew;
1930                         error--;
1931                 }
1932                 bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1933                 if (error > 0) {
1934                         error = EINVAL;
1935                 } else {
1936                         error = 0;
1937                         carp_setrun(sc, 0);
1938                 }
1939                 break;
1940
1941         case SIOCGVH:
1942                 bzero(&carpr, sizeof(carpr));
1943                 carpr.carpr_state = sc->sc_state;
1944                 carpr.carpr_vhid = sc->sc_vhid;
1945                 carpr.carpr_advbase = sc->sc_advbase;
1946                 carpr.carpr_advskew = sc->sc_advskew;
1947                 error = priv_check_cred(cr, PRIV_ROOT, NULL_CRED_OKAY);
1948                 if (error == 0) {
1949                         bcopy(sc->sc_key, carpr.carpr_key,
1950                               sizeof(carpr.carpr_key));
1951                 }
1952
1953                 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1954                 break;
1955
1956         case SIOCGDRVSPEC:
1957                 switch (ifd->ifd_cmd) {
1958                 case CARPGDEVNAME:
1959                         if (ifd->ifd_len != sizeof(devname))
1960                                 error = EINVAL;
1961                         break;
1962
1963                 case CARPGVHADDR:
1964                         break;
1965
1966                 default:
1967                         error = EINVAL;
1968                         break;
1969                 }
1970                 if (error)
1971                         break;
1972
1973                 switch (ifd->ifd_cmd) {
1974                 case CARPGVHADDR:
1975                         error = carp_get_vhaddr(sc, ifd);
1976                         break;
1977
1978                 case CARPGDEVNAME:
1979                         bzero(devname, sizeof(devname));
1980                         if (sc->sc_carpdev != NULL) {
1981                                 strlcpy(devname, sc->sc_carpdev->if_xname,
1982                                         sizeof(devname));
1983                         }
1984                         error = copyout(devname, ifd->ifd_data,
1985                                         sizeof(devname));
1986                         break;
1987                 }
1988                 break;
1989
1990         default:
1991                 error = ether_ioctl(ifp, cmd, addr);
1992                 break;
1993         }
1994         carp_hmac_prepare(sc);
1995
1996         carp_reltok();
1997         return error;
1998 }
1999
2000 static void
2001 carp_init(void *xsc)
2002 {
2003         struct carp_softc *sc = xsc;
2004
2005         carp_gettok();
2006
2007         sc->sc_if.if_flags |= IFF_RUNNING;
2008         carp_set_state(sc, INIT);
2009         carp_setrun(sc, 0);
2010
2011         carp_reltok();
2012 }
2013
2014 static int
2015 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *dst,
2016     struct rtentry *rt)
2017 {
2018         struct carp_softc *sc = ifp->if_softc;
2019         int error = 0;
2020
2021         carp_gettok();
2022         if (sc->sc_carpdev) {
2023                 /*
2024                  * NOTE:
2025                  * CARP's ifp is passed to backing device's
2026                  * if_output method.
2027                  */
2028                 sc->sc_carpdev->if_output(ifp, m, dst, rt);
2029         } else {
2030                 m_freem(m);
2031                 error = ENETUNREACH;
2032         }
2033         carp_reltok();
2034
2035         return error;
2036 }
2037
2038 /*
2039  * Start output on carp interface. This function should never be called.
2040  */
2041 static void
2042 carp_start(struct ifnet *ifp)
2043 {
2044         panic("%s: start called\n", ifp->if_xname);
2045 }
2046
2047 static void
2048 carp_serialize(struct ifnet *ifp __unused,
2049     enum ifnet_serialize slz __unused)
2050 {
2051 }
2052
2053 static void
2054 carp_deserialize(struct ifnet *ifp __unused,
2055     enum ifnet_serialize slz __unused)
2056 {
2057 }
2058
2059 static int
2060 carp_tryserialize(struct ifnet *ifp __unused,
2061     enum ifnet_serialize slz __unused)
2062 {
2063         return 1;
2064 }
2065
2066 #ifdef INVARIANTS
2067
2068 static void
2069 carp_serialize_assert(struct ifnet *ifp __unused,
2070     enum ifnet_serialize slz __unused, boolean_t serialized __unused)
2071 {
2072 }
2073
2074 #endif  /* INVARIANTS */
2075
2076 static void
2077 carp_set_state(struct carp_softc *sc, int state)
2078 {
2079         struct ifnet *cifp = &sc->sc_if;
2080
2081         if (sc->sc_state == state)
2082                 return;
2083         sc->sc_state = state;
2084
2085         switch (sc->sc_state) {
2086         case BACKUP:
2087                 cifp->if_link_state = LINK_STATE_DOWN;
2088                 break;
2089
2090         case MASTER:
2091                 cifp->if_link_state = LINK_STATE_UP;
2092                 break;
2093
2094         default:
2095                 cifp->if_link_state = LINK_STATE_UNKNOWN;
2096                 break;
2097         }
2098         rt_ifmsg(cifp);
2099 }
2100
2101 void
2102 carp_group_demote_adj(struct ifnet *ifp, int adj)
2103 {
2104         struct ifg_list *ifgl;
2105         int *dm;
2106
2107         carp_gettok();
2108
2109         TAILQ_FOREACH(ifgl, &ifp->if_groups, ifgl_next) {
2110                 if (!strcmp(ifgl->ifgl_group->ifg_group, IFG_ALL))
2111                         continue;
2112                 dm = &ifgl->ifgl_group->ifg_carp_demoted;
2113
2114                 if (*dm + adj >= 0)
2115                         *dm += adj;
2116                 else
2117                         *dm = 0;
2118
2119                 if (adj > 0 && *dm == 1)
2120                         carp_send_ad_all();
2121                 CARP_LOG("%s demoted group %s to %d", ifp->if_xname,
2122                     ifgl->ifgl_group->ifg_group, *dm);
2123         }
2124
2125         carp_reltok();
2126 }
2127
2128 void
2129 carp_carpdev_state(void *v)
2130 {
2131         struct carp_if *cif = v;
2132         struct carp_softc *sc;
2133
2134         carp_gettok();
2135
2136         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list)
2137                 carp_sc_state(sc);
2138
2139         carp_reltok();
2140 }
2141
2142 static void
2143 carp_sc_state(struct carp_softc *sc)
2144 {
2145         if (!(sc->sc_carpdev->if_flags & IFF_UP)) {
2146                 callout_stop(&sc->sc_ad_tmo);
2147                 callout_stop(&sc->sc_md_tmo);
2148                 callout_stop(&sc->sc_md6_tmo);
2149                 carp_set_state(sc, INIT);
2150                 carp_setrun(sc, 0);
2151                 if (!sc->sc_suppress) {
2152                         carp_suppress_preempt++;
2153                         if (carp_suppress_preempt == 1)
2154                                 carp_send_ad_all();
2155                 }
2156                 sc->sc_suppress = 1;
2157         } else {
2158                 carp_set_state(sc, INIT);
2159                 carp_setrun(sc, 0);
2160                 if (sc->sc_suppress)
2161                         carp_suppress_preempt--;
2162                 sc->sc_suppress = 0;
2163         }
2164 }
2165
2166 static void
2167 carp_stop(struct carp_softc *sc, int detach)
2168 {
2169         sc->sc_if.if_flags &= ~IFF_RUNNING;
2170
2171         callout_stop(&sc->sc_ad_tmo);
2172         callout_stop(&sc->sc_md_tmo);
2173         callout_stop(&sc->sc_md6_tmo);
2174
2175         if (!detach && sc->sc_state == MASTER)
2176                 carp_send_ad(sc);
2177
2178         if (sc->sc_suppress)
2179                 carp_suppress_preempt--;
2180         sc->sc_suppress = 0;
2181
2182         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS)
2183                 carp_suppress_preempt--;
2184         sc->sc_sendad_errors = 0;
2185         sc->sc_sendad_success = 0;
2186
2187         carp_set_state(sc, INIT);
2188         carp_setrun(sc, 0);
2189 }
2190
2191 static void
2192 carp_suspend(struct carp_softc *sc, int detach)
2193 {
2194         struct ifnet *cifp = &sc->sc_if;
2195
2196         carp_stop(sc, detach);
2197
2198         /* Retain the running state, if we are not dead yet */
2199         if (!sc->sc_dead && (cifp->if_flags & IFF_UP))
2200                 cifp->if_flags |= IFF_RUNNING;
2201 }
2202
2203 static int
2204 carp_activate_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha,
2205     struct ifnet *ifp, struct in_ifaddr *ia_if, int own)
2206 {
2207         struct ip_moptions *imo = &sc->sc_imo;
2208         struct carp_if *cif;
2209         struct carp_softc *vr, *after = NULL;
2210         int onlist, error;
2211 #ifdef INVARIANTS
2212         int assert_onlist;
2213 #endif
2214
2215         KKASSERT(vha->vha_ia != NULL);
2216
2217         KASSERT(ia_if != NULL, ("NULL backing address\n"));
2218         KASSERT(vha->vha_iaback == NULL, ("%p is already activated\n", vha));
2219         KASSERT((vha->vha_flags & CARP_VHAF_OWNER) == 0,
2220                 ("inactive vhaddr %p is the address owner\n", vha));
2221
2222         KASSERT(sc->sc_carpdev == NULL || sc->sc_carpdev == ifp,
2223                 ("%s is already on %s\n", sc->sc_if.if_xname,
2224                  sc->sc_carpdev->if_xname));
2225
2226         if (!ifp->if_carp) {
2227                 KASSERT(sc->sc_carpdev == NULL,
2228                         ("%s is already on %s\n", sc->sc_if.if_xname,
2229                          sc->sc_carpdev->if_xname));
2230
2231                 cif = kmalloc(sizeof(*cif), M_CARP, M_WAITOK | M_ZERO);
2232
2233                 error = ifpromisc(ifp, 1);
2234                 if (error) {
2235                         kfree(cif, M_CARP);
2236                         return error;
2237                 }
2238
2239                 TAILQ_INIT(&cif->vhif_vrs);
2240                 ifp->if_carp = cif;
2241         } else {
2242                 cif = ifp->if_carp;
2243                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
2244                         if (vr != sc && vr->sc_vhid == sc->sc_vhid)
2245                                 return EINVAL;
2246                 }
2247         }
2248
2249 #ifdef INVARIANTS
2250         if (sc->sc_carpdev != NULL)
2251                 assert_onlist = 1;
2252         else
2253                 assert_onlist = 0;
2254 #endif
2255         sc->sc_ia = ia_if;
2256         sc->sc_carpdev = ifp;
2257
2258         cif = ifp->if_carp;
2259         onlist = 0;
2260         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
2261                 if (vr == sc)
2262                         onlist = 1;
2263                 if (vr->sc_vhid < sc->sc_vhid)
2264                         after = vr;
2265         }
2266
2267 #ifdef INVARIANTS
2268         if (assert_onlist) {
2269                 KASSERT(onlist, ("%s is not on %s carp list\n",
2270                         sc->sc_if.if_xname, ifp->if_xname));
2271         } else {
2272                 KASSERT(!onlist, ("%s is already on %s carp list\n",
2273                         sc->sc_if.if_xname, ifp->if_xname));
2274         }
2275 #endif
2276
2277         if (!onlist) {
2278                 /* We're trying to keep things in order */
2279                 if (after == NULL)
2280                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
2281                 else
2282                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
2283         }
2284
2285         vha->vha_iaback = ia_if;
2286         sc->sc_naddrs++;
2287
2288         if (own) {
2289                 vha->vha_flags |= CARP_VHAF_OWNER;
2290
2291                 /* XXX save user configured advskew? */
2292                 sc->sc_advskew = 0;
2293         }
2294
2295         carp_addroute_vhaddr(sc, vha);
2296
2297         /*
2298          * Join the multicast group only after the backing interface
2299          * has been hooked with the CARP interface.
2300          */
2301         KASSERT(imo->imo_multicast_ifp == NULL ||
2302                 imo->imo_multicast_ifp == &sc->sc_if,
2303                 ("%s didn't leave mcast group on %s\n",
2304                  sc->sc_if.if_xname, imo->imo_multicast_ifp->if_xname));
2305
2306         if (imo->imo_num_memberships == 0) {
2307                 struct in_addr addr;
2308
2309                 addr.s_addr = htonl(INADDR_CARP_GROUP);
2310                 imo->imo_membership[0] = in_addmulti(&addr, &sc->sc_if);
2311                 if (imo->imo_membership[0] == NULL) {
2312                         carp_deactivate_vhaddr(sc, vha, FALSE);
2313                         return ENOBUFS;
2314                 }
2315
2316                 imo->imo_num_memberships++;
2317                 imo->imo_multicast_ifp = &sc->sc_if;
2318                 imo->imo_multicast_ttl = CARP_DFLTTL;
2319                 imo->imo_multicast_loop = 0;
2320         }
2321
2322         carp_hmac_prepare(sc);
2323         carp_set_state(sc, INIT);
2324         carp_setrun(sc, 0);
2325         return 0;
2326 }
2327
2328 static void
2329 carp_deactivate_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha,
2330     boolean_t del_iaback)
2331 {
2332         KKASSERT(vha->vha_ia != NULL);
2333
2334         carp_hmac_prepare(sc);
2335
2336         if (vha->vha_iaback == NULL) {
2337                 KASSERT((vha->vha_flags & CARP_VHAF_OWNER) == 0,
2338                         ("inactive vhaddr %p is the address owner\n", vha));
2339                 return;
2340         }
2341
2342         vha->vha_flags &= ~CARP_VHAF_OWNER;
2343         carp_delroute_vhaddr(sc, vha, del_iaback);
2344
2345         KKASSERT(sc->sc_naddrs > 0);
2346         vha->vha_iaback = NULL;
2347         sc->sc_naddrs--;
2348         if (!sc->sc_naddrs) {
2349                 if (sc->sc_naddrs6) {
2350                         carp_multicast_cleanup(sc);
2351                         sc->sc_ia = NULL;
2352                 } else {
2353                         carp_detach(sc, 0, del_iaback);
2354                 }
2355         }
2356 }
2357
2358 static void
2359 carp_link_addrs(struct carp_softc *sc, struct ifnet *ifp, struct ifaddr *ifa_if)
2360 {
2361         struct carp_vhaddr *vha;
2362         struct in_ifaddr *ia_if;
2363
2364         KKASSERT(ifa_if->ifa_addr->sa_family == AF_INET);
2365         ia_if = ifatoia(ifa_if);
2366
2367         /*
2368          * Test each inactive vhaddr against the newly added address.
2369          * If the newly added address could be the backing address,
2370          * then activate the matching vhaddr.
2371          */
2372         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
2373                 const struct in_ifaddr *ia;
2374                 u_long iaddr;
2375                 int own;
2376
2377                 if (vha->vha_iaback != NULL)
2378                         continue;
2379
2380                 ia = vha->vha_ia;
2381                 iaddr = ntohl(ia->ia_addr.sin_addr.s_addr);
2382
2383                 if ((iaddr & ia_if->ia_subnetmask) != ia_if->ia_subnet)
2384                         continue;
2385
2386                 own = 0;
2387                 if (ia->ia_addr.sin_addr.s_addr ==
2388                     ia_if->ia_addr.sin_addr.s_addr)
2389                         own = 1;
2390
2391                 carp_activate_vhaddr(sc, vha, ifp, ia_if, own);
2392         }
2393 }
2394
2395 static void
2396 carp_unlink_addrs(struct carp_softc *sc, struct ifnet *ifp,
2397                   struct ifaddr *ifa_if)
2398 {
2399         struct carp_vhaddr *vha;
2400         struct in_ifaddr *ia_if;
2401
2402         KKASSERT(ifa_if->ifa_addr->sa_family == AF_INET);
2403         ia_if = ifatoia(ifa_if);
2404
2405         /*
2406          * Ad src address is deleted; set it to NULL.
2407          * Following loop will try pick up a new ad src address
2408          * if one of the vhaddr could retain its backing address.
2409          */
2410         if (sc->sc_ia == ia_if)
2411                 sc->sc_ia = NULL;
2412
2413         /*
2414          * Test each active vhaddr against the deleted address.
2415          * If the deleted address is vhaddr address's backing
2416          * address, then deactivate the vhaddr.
2417          */
2418         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
2419                 if (vha->vha_iaback == NULL)
2420                         continue;
2421
2422                 if (vha->vha_iaback == ia_if)
2423                         carp_deactivate_vhaddr(sc, vha, TRUE);
2424                 else if (sc->sc_ia == NULL)
2425                         sc->sc_ia = vha->vha_iaback;
2426         }
2427 }
2428
2429 static void
2430 carp_update_addrs(struct carp_softc *sc, struct ifaddr *ifa_del)
2431 {
2432         struct carp_vhaddr *vha;
2433
2434         KKASSERT(sc->sc_carpdev == NULL);
2435
2436         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link)
2437                 carp_config_vhaddr(sc, vha, ifatoia(ifa_del));
2438 }
2439
2440 static void
2441 carp_ifaddr(void *arg __unused, struct ifnet *ifp,
2442             enum ifaddr_event event, struct ifaddr *ifa)
2443 {
2444         struct carp_softc *sc;
2445
2446         carp_gettok();
2447
2448         if (ifa->ifa_addr->sa_family != AF_INET)
2449                 goto back;
2450
2451         if (ifp->if_type == IFT_CARP) {
2452                 /*
2453                  * Address is changed on carp(4) interface
2454                  */
2455                 switch (event) {
2456                 case IFADDR_EVENT_ADD:
2457                         carp_add_addr(ifp->if_softc, ifa);
2458                         break;
2459
2460                 case IFADDR_EVENT_CHANGE:
2461                         carp_config_addr(ifp->if_softc, ifa);
2462                         break;
2463
2464                 case IFADDR_EVENT_DELETE:
2465                         carp_del_addr(ifp->if_softc, ifa);
2466                         break;
2467                 }
2468                 goto back;
2469         }
2470
2471         /*
2472          * Address is changed on non-carp(4) interface
2473          */
2474         if ((ifp->if_flags & IFF_MULTICAST) == 0)
2475                 goto back;
2476
2477         LIST_FOREACH(sc, &carpif_list, sc_next) {
2478                 if (sc->sc_carpdev != NULL && sc->sc_carpdev != ifp) {
2479                         /* Not the parent iface; skip */
2480                         continue;
2481                 }
2482
2483                 switch (event) {
2484                 case IFADDR_EVENT_ADD:
2485                         carp_link_addrs(sc, ifp, ifa);
2486                         break;
2487
2488                 case IFADDR_EVENT_DELETE:
2489                         if (sc->sc_carpdev != NULL) {
2490                                 carp_unlink_addrs(sc, ifp, ifa);
2491                                 if (sc->sc_carpdev == NULL) {
2492                                         /*
2493                                          * We no longer have the parent
2494                                          * interface, however, certain
2495                                          * virtual addresses, which are
2496                                          * not used because they can't
2497                                          * match the previous parent
2498                                          * interface's addresses, may now
2499                                          * match different interface's
2500                                          * addresses.
2501                                          */
2502                                         carp_update_addrs(sc, ifa);
2503                                 }
2504                         } else {
2505                                 /*
2506                                  * The carp(4) interface didn't have a
2507                                  * parent iface, so it is not possible
2508                                  * that it will contain any address to
2509                                  * be unlinked.
2510                                  */
2511                         }
2512                         break;
2513
2514                 case IFADDR_EVENT_CHANGE:
2515                         if (sc->sc_carpdev == NULL) {
2516                                 /*
2517                                  * The carp(4) interface didn't have a
2518                                  * parent iface, so it is not possible
2519                                  * that it will contain any address to
2520                                  * be updated.
2521                                  */
2522                                 carp_link_addrs(sc, ifp, ifa);
2523                         } else {
2524                                 /*
2525                                  * First try breaking tie with the old
2526                                  * address.  Then see whether we could
2527                                  * link certain vhaddr to the new address.
2528                                  * If that fails, i.e. carpdev is NULL,
2529                                  * we try a global update.
2530                                  *
2531                                  * NOTE: The above order is critical.
2532                                  */
2533                                 carp_unlink_addrs(sc, ifp, ifa);
2534                                 carp_link_addrs(sc, ifp, ifa);
2535                                 if (sc->sc_carpdev == NULL) {
2536                                         /*
2537                                          * See the comment in the above
2538                                          * IFADDR_EVENT_DELETE block.
2539                                          */
2540                                         carp_update_addrs(sc, NULL);
2541                                 }
2542                         }
2543                         break;
2544                 }
2545         }
2546
2547 back:
2548         carp_reltok();
2549 }
2550
2551 void
2552 carp_proto_ctlinput(netmsg_t msg)
2553 {
2554         int cmd = msg->ctlinput.nm_cmd;
2555         struct sockaddr *sa = msg->ctlinput.nm_arg;
2556         struct in_ifaddr_container *iac;
2557
2558         carp_gettok();
2559
2560         TAILQ_FOREACH(iac, &in_ifaddrheads[mycpuid], ia_link) {
2561                 struct in_ifaddr *ia = iac->ia;
2562                 struct ifnet *ifp = ia->ia_ifp;
2563
2564                 if (ifp->if_type == IFT_CARP)
2565                         continue;
2566
2567                 if (ia->ia_ifa.ifa_addr == sa) {
2568                         if (cmd == PRC_IFDOWN) {
2569                                 carp_ifaddr(NULL, ifp, IFADDR_EVENT_DELETE,
2570                                     &ia->ia_ifa);
2571                         } else if (cmd == PRC_IFUP) {
2572                                 carp_ifaddr(NULL, ifp, IFADDR_EVENT_ADD,
2573                                     &ia->ia_ifa);
2574                         }
2575                         break;
2576                 }
2577         }
2578
2579         carp_reltok();
2580         lwkt_replymsg(&msg->lmsg, 0);
2581 }
2582
2583 void
2584 carp_gettok(void)
2585 {
2586         lwkt_gettoken(&carp_tok);
2587 }
2588
2589 void
2590 carp_reltok(void)
2591 {
2592         lwkt_reltoken(&carp_tok);
2593 }
2594
2595 struct ifnet *
2596 carp_parent(struct ifnet *cifp)
2597 {
2598         struct carp_softc *sc;
2599
2600         ASSERT_LWKT_TOKEN_HELD(&carp_tok);
2601
2602         KKASSERT(cifp->if_type == IFT_CARP);
2603         sc = cifp->if_softc;
2604
2605         return sc->sc_carpdev;
2606 }
2607
2608 #define rtinitflags(x) \
2609         (((x)->ia_ifp->if_flags & (IFF_LOOPBACK | IFF_POINTOPOINT)) \
2610                  ? RTF_HOST : 0)
2611
2612 static int
2613 carp_addroute_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha)
2614 {
2615         struct in_ifaddr *ia, *iaback;
2616         int error;
2617
2618         if (sc->sc_state != MASTER)
2619                 return 0;
2620
2621         ia = vha->vha_ia;
2622         KKASSERT(ia != NULL);
2623
2624         iaback = vha->vha_iaback;
2625         KKASSERT(iaback != NULL);
2626
2627         rtinit(&iaback->ia_ifa, RTM_DELETE, rtinitflags(iaback));
2628         in_ifadown(&iaback->ia_ifa, 1);
2629         iaback->ia_flags &= ~IFA_ROUTE;
2630
2631         error = rtinit(&ia->ia_ifa, RTM_ADD, rtinitflags(ia) | RTF_UP);
2632         if (!error)
2633                 ia->ia_flags |= IFA_ROUTE;
2634         return error;
2635 }
2636
2637 static void
2638 carp_delroute_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha,
2639     boolean_t del_iaback)
2640 {
2641         struct in_ifaddr *ia, *iaback;
2642
2643         ia = vha->vha_ia;
2644         KKASSERT(ia != NULL);
2645
2646         iaback = vha->vha_iaback;
2647         KKASSERT(iaback != NULL);
2648
2649         rtinit(&ia->ia_ifa, RTM_DELETE, rtinitflags(ia));
2650         in_ifadown(&ia->ia_ifa, 1);
2651         ia->ia_flags &= ~IFA_ROUTE;
2652
2653         if (!del_iaback && (iaback->ia_ifp->if_flags & IFF_UP)) {
2654                 int error;
2655
2656                 error = rtinit(&iaback->ia_ifa, RTM_ADD,
2657                     rtinitflags(iaback) | RTF_UP);
2658                 if (!error)
2659                         iaback->ia_flags |= IFA_ROUTE;
2660         }
2661 }
2662
2663 static int
2664 carp_modevent(module_t mod, int type, void *data)
2665 {
2666         switch (type) {
2667         case MOD_LOAD:
2668                 LIST_INIT(&carpif_list);
2669                 carp_ifdetach_event =
2670                 EVENTHANDLER_REGISTER(ifnet_detach_event, carp_ifdetach, NULL,
2671                                       EVENTHANDLER_PRI_ANY);
2672                 carp_ifaddr_event =
2673                 EVENTHANDLER_REGISTER(ifaddr_event, carp_ifaddr, NULL,
2674                                       EVENTHANDLER_PRI_FIRST);
2675                 if_clone_attach(&carp_cloner);
2676                 break;
2677
2678         case MOD_UNLOAD:
2679                 EVENTHANDLER_DEREGISTER(ifnet_detach_event,
2680                                         carp_ifdetach_event);
2681                 EVENTHANDLER_DEREGISTER(ifaddr_event,
2682                                         carp_ifaddr_event);
2683                 if_clone_detach(&carp_cloner);
2684                 break;
2685
2686         default:
2687                 return (EINVAL);
2688         }
2689         return (0);
2690 }
2691
2692 static moduledata_t carp_mod = {
2693         "carp",
2694         carp_modevent,
2695         0
2696 };
2697 DECLARE_MODULE(carp, carp_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);