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 #include <net/netmsg2.h>
  64 #include <net/netisr2.h>
  65
  66 #ifdef INET
  67 #include <netinet/in.h>
  68 #include <netinet/in_var.h>
  69 #include <netinet/in_systm.h>
  70 #include <netinet/ip.h>
  71 #include <netinet/ip_var.h>
  72 #include <netinet/if_ether.h>
  73 #endif
  74
  75 #ifdef INET6
  76 #include <netinet/icmp6.h>
  77 #include <netinet/ip6.h>
  78 #include <netinet6/ip6_var.h>
  79 #include <netinet6/scope6_var.h>
  80 #include <netinet6/nd6.h>
  81 #endif
  82
  83 #include <netinet/ip_carp.h>
  84
  85 /*
  86  * Note about carp's MP safe approach:
  87  *
  88  * Brief: carp_softc (softc), carp_softc_container (scc)
  89  *
  90  * - All configuration operation, e.g. ioctl, add/delete inet addresses
  91  *   is serialized by netisr0; not by carp's serializer
  92  *
  93  * - Backing interface's if_carp and carp_softc's relationship:
  94  *
  95  *                +---------+
  96  *     if_carp -->| carp_if |
  97  *                +---------+
  98  *                     |
  99  *                     |
 100  *                     V      +---------+
 101  *                  +-----+   |         |
 102  *                  | scc |-->|  softc  |
 103  *                  +-----+   |         |
 104  *                     |      +---------+
 105  *                     |
 106  *                     V      +---------+
 107  *                  +-----+   |         |
 108  *                  | scc |-->|  softc  |
 109  *                  +-----+   |         |
 110  *                            +---------+
 111  *
 112  * - if_carp creation, modification and deletion all happen in netisr0,
 113  *   as stated previously.  Since if_carp is accessed by multiple netisrs,
 114  *   the modification to if_carp is conducted in the following way:
 115  *
 116  *   Adding carp_softc:
 117  *
 118  *   1) Duplicate the old carp_if to new carp_if (ncif), and insert the
 119  *      to-be-added carp_softc to the new carp_if (ncif):
 120  *
 121  *        if_carp                     ncif
 122  *           |                         |
 123  *           V                         V
 124  *      +---------+               +---------+
 125  *      | carp_if |               | carp_if |
 126  *      +---------+               +---------+
 127  *           |                         |
 128  *           |                         |
 129  *           V        +-------+        V
 130  *        +-----+     |       |     +-----+
 131  *        | scc |---->| softc |<----| scc |
 132  *        +-----+     |       |     +-----+
 133  *           |        +-------+        |
 134  *           |                         |
 135  *           V        +-------+        V
 136  *        +-----+     |       |     +-----+
 137  *        | scc |---->| softc |<----| scc |
 138  *        +-----+     |       |     +-----+
 139  *                    +-------+        |
 140  *                                     |
 141  *                    +-------+        V
 142  *                    |       |     +-----+
 143  *                    | softc |<----| scc |
 144  *                    |       |     +-----+
 145  *                    +-------+
 146  *
 147  *   2) Switch save if_carp into ocif and switch if_carp to ncif:
 148  *
 149  *          ocif                    if_carp
 150  *           |                         |
 151  *           V                         V
 152  *      +---------+               +---------+
 153  *      | carp_if |               | carp_if |
 154  *      +---------+               +---------+
 155  *           |                         |
 156  *           |                         |
 157  *           V        +-------+        V
 158  *        +-----+     |       |     +-----+
 159  *        | scc |---->| softc |<----| scc |
 160  *        +-----+     |       |     +-----+
 161  *           |        +-------+        |
 162  *           |                         |
 163  *           V        +-------+        V
 164  *        +-----+     |       |     +-----+
 165  *        | scc |---->| softc |<----| scc |
 166  *        +-----+     |       |     +-----+
 167  *                    +-------+        |
 168  *                                     |
 169  *                    +-------+        V
 170  *                    |       |     +-----+
 171  *                    | softc |<----| scc |
 172  *                    |       |     +-----+
 173  *                    +-------+
 174  *
 175  *   3) Run netmsg_service_sync(), which will make sure that
 176  *      ocif is no longer accessed (all network operations
 177  *      are happened only in network threads).
 178  *   4) Free ocif -- only carp_if and scc are freed.
 179  *
 180  *
 181  *   Removing carp_softc:
 182  *
 183  *   1) Duplicate the old carp_if to new carp_if (ncif); the to-be-deleted
 184  *      carp_softc will not be duplicated.
 185  *
 186  *        if_carp                     ncif
 187  *           |                         |
 188  *           V                         V
 189  *      +---------+               +---------+
 190  *      | carp_if |               | carp_if |
 191  *      +---------+               +---------+
 192  *           |                         |
 193  *           |                         |
 194  *           V        +-------+        V
 195  *        +-----+     |       |     +-----+
 196  *        | scc |---->| softc |<----| scc |
 197  *        +-----+     |       |     +-----+
 198  *           |        +-------+        |
 199  *           |                         |
 200  *           V        +-------+        |
 201  *        +-----+     |       |        |
 202  *        | scc |---->| softc |        |
 203  *        +-----+     |       |        |
 204  *           |        +-------+        |
 205  *           |                         |
 206  *           V        +-------+        V
 207  *        +-----+     |       |     +-----+
 208  *        | scc |---->| softc |<----| scc |
 209  *        +-----+     |       |     +-----+
 210  *                    +-------+
 211  *
 212  *   2) Switch save if_carp into ocif and switch if_carp to ncif:
 213  *
 214  *          ocif                    if_carp
 215  *           |                         |
 216  *           V                         V
 217  *      +---------+               +---------+
 218  *      | carp_if |               | carp_if |
 219  *      +---------+               +---------+
 220  *           |                         |
 221  *           |                         |
 222  *           V        +-------+        V
 223  *        +-----+     |       |     +-----+
 224  *        | scc |---->| softc |<----| scc |
 225  *        +-----+     |       |     +-----+
 226  *           |        +-------+        |
 227  *           |                         |
 228  *           V        +-------+        |
 229  *        +-----+     |       |        |
 230  *        | scc |---->| softc |        |
 231  *        +-----+     |       |        |
 232  *           |        +-------+        |
 233  *           |                         |
 234  *           V        +-------+        V
 235  *        +-----+     |       |     +-----+
 236  *        | scc |---->| softc |<----| scc |
 237  *        +-----+     |       |     +-----+
 238  *                    +-------+
 239  *
 240  *   3) Run netmsg_service_sync(), which will make sure that
 241  *      ocif is no longer accessed (all network operations
 242  *      are happened only in network threads).
 243  *   4) Free ocif -- only carp_if and scc are freed.
 244  *
 245  * - if_carp accessing:
 246  *   The accessing code should cache the if_carp in a local temporary
 247  *   variable and accessing the temporary variable along the code path
 248  *   instead of accessing if_carp later on.
 249  */
 250
 251 #define CARP_IFNAME             "carp"
 252 #define CARP_IS_RUNNING(ifp)    \
 253         (((ifp)->if_flags & (IFF_UP | IFF_RUNNING)) == (IFF_UP | IFF_RUNNING))
 254
 255 struct carp_softc;
 256
 257 struct carp_vhaddr {
 258         uint32_t                vha_flags;      /* CARP_VHAF_ */
 259         struct in_ifaddr        *vha_ia;        /* carp address */
 260         struct in_ifaddr        *vha_iaback;    /* backing address */
 261         TAILQ_ENTRY(carp_vhaddr) vha_link;
 262 };
 263 TAILQ_HEAD(carp_vhaddr_list, carp_vhaddr);
 264
 265 struct netmsg_carp {
 266         struct netmsg_base      base;
 267         struct ifnet            *nc_carpdev;
 268         struct carp_softc       *nc_softc;
 269         void                    *nc_data;
 270         size_t                  nc_datalen;
 271 };
 272
 273 struct carp_softc {
 274         struct arpcom            arpcom;
 275         struct ifnet            *sc_carpdev;    /* parent interface */
 276         struct carp_vhaddr_list  sc_vha_list;   /* virtual addr list */
 277
 278         const struct in_ifaddr  *sc_ia;         /* primary iface address v4 */
 279         struct ip_moptions       sc_imo;
 280
 281 #ifdef INET6
 282         struct in6_ifaddr       *sc_ia6;        /* primary iface address v6 */
 283         struct ip6_moptions      sc_im6o;
 284 #endif /* INET6 */
 285
 286         enum { INIT = 0, BACKUP, MASTER }
 287                                  sc_state;
 288         boolean_t                sc_dead;
 289
 290         int                      sc_suppress;
 291
 292         int                      sc_sendad_errors;
 293 #define CARP_SENDAD_MAX_ERRORS  3
 294         int                      sc_sendad_success;
 295 #define CARP_SENDAD_MIN_SUCCESS 3
 296
 297         int                      sc_vhid;
 298         int                      sc_advskew;
 299         int                      sc_naddrs;     /* actually used IPv4 vha */
 300         int                      sc_naddrs6;
 301         int                      sc_advbase;    /* seconds */
 302         int                      sc_init_counter;
 303         uint64_t                 sc_counter;
 304
 305         /* authentication */
 306 #define CARP_HMAC_PAD   64
 307         unsigned char            sc_key[CARP_KEY_LEN];
 308         unsigned char            sc_pad[CARP_HMAC_PAD];
 309         SHA1_CTX                 sc_sha1;
 310
 311         struct callout           sc_ad_tmo;     /* advertisement timeout */
 312         struct netmsg_carp       sc_ad_msg;     /* adv timeout netmsg */
 313         struct callout           sc_md_tmo;     /* ip4 master down timeout */
 314         struct callout           sc_md6_tmo;    /* ip6 master down timeout */
 315         struct netmsg_carp       sc_md_msg;     /* master down timeout netmsg */
 316
 317         LIST_ENTRY(carp_softc)   sc_next;       /* Interface clue */
 318 };
 319
 320 #define sc_if   arpcom.ac_if
 321
 322 struct carp_softc_container {
 323         TAILQ_ENTRY(carp_softc_container) scc_link;
 324         struct carp_softc       *scc_softc;
 325 };
 326 TAILQ_HEAD(carp_if, carp_softc_container);
 327
 328 SYSCTL_DECL(_net_inet_carp);
 329
 330 static int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 1, 0, 0 }; /* XXX for now */
 331 SYSCTL_INT(_net_inet_carp, CARPCTL_ALLOW, allow, CTLFLAG_RW,
 332     &carp_opts[CARPCTL_ALLOW], 0, "Accept incoming CARP packets");
 333 SYSCTL_INT(_net_inet_carp, CARPCTL_PREEMPT, preempt, CTLFLAG_RW,
 334     &carp_opts[CARPCTL_PREEMPT], 0, "high-priority backup preemption mode");
 335 SYSCTL_INT(_net_inet_carp, CARPCTL_LOG, log, CTLFLAG_RW,
 336     &carp_opts[CARPCTL_LOG], 0, "log bad carp packets");
 337 SYSCTL_INT(_net_inet_carp, CARPCTL_ARPBALANCE, arpbalance, CTLFLAG_RW,
 338     &carp_opts[CARPCTL_ARPBALANCE], 0, "balance arp responses");
 339
 340 static int carp_suppress_preempt = 0;
 341 SYSCTL_INT(_net_inet_carp, OID_AUTO, suppress_preempt, CTLFLAG_RD,
 342     &carp_suppress_preempt, 0, "Preemption is suppressed");
 343
 344 static struct carpstats carpstats;
 345 SYSCTL_STRUCT(_net_inet_carp, CARPCTL_STATS, stats, CTLFLAG_RW,
 346     &carpstats, carpstats,
 347     "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 348
 349 #define CARP_LOG(...)   do {                            \
 350         if (carp_opts[CARPCTL_LOG] > 0)                 \
 351                 log(LOG_INFO, __VA_ARGS__);             \
 352 } while (0)
 353
 354 #define CARP_DEBUG(...) do {                            \
 355         if (carp_opts[CARPCTL_LOG] > 1)                 \
 356                 log(LOG_DEBUG, __VA_ARGS__);            \
 357 } while (0)
 358
 359 static struct lwkt_token carp_listtok = LWKT_TOKEN_INITIALIZER(carp_list_token);
 360
 361 static void     carp_hmac_prepare(struct carp_softc *);
 362 static void     carp_hmac_generate(struct carp_softc *, uint32_t *,
 363                     unsigned char *);
 364 static int      carp_hmac_verify(struct carp_softc *, uint32_t *,
 365                     unsigned char *);
 366 static void     carp_setroute(struct carp_softc *, int);
 367 static void     carp_proto_input_c(struct carp_softc *, struct mbuf *,
 368                     struct carp_header *, sa_family_t);
 369 static int      carp_clone_create(struct if_clone *, int, caddr_t);
 370 static int      carp_clone_destroy(struct ifnet *);
 371 static void     carp_detach(struct carp_softc *, boolean_t, boolean_t);
 372 static void     carp_prepare_ad(struct carp_softc *, struct carp_header *);
 373 static void     carp_send_ad_all(void);
 374 static void     carp_send_ad_timeout(void *);
 375 static void     carp_send_ad(struct carp_softc *);
 376 static void     carp_send_arp(struct carp_softc *);
 377 static void     carp_master_down_timeout(void *);
 378 static void     carp_master_down(struct carp_softc *);
 379 static void     carp_setrun(struct carp_softc *, sa_family_t);
 380 static void     carp_set_state(struct carp_softc *, int);
 381 static struct ifnet *carp_forus(struct carp_if *, const uint8_t *);
 382
 383 static void     carp_init(void *);
 384 static int      carp_ioctl(struct ifnet *, u_long, caddr_t, struct ucred *);
 385 static int      carp_output(struct ifnet *, struct mbuf *, struct sockaddr *,
 386                     struct rtentry *);
 387 static void     carp_start(struct ifnet *, struct ifaltq_subque *);
 388
 389 static void     carp_multicast_cleanup(struct carp_softc *);
 390 static void     carp_add_addr(struct carp_softc *, struct ifaddr *);
 391 static void     carp_del_addr(struct carp_softc *, struct ifaddr *);
 392 static void     carp_config_addr(struct carp_softc *, struct ifaddr *);
 393 static void     carp_link_addrs(struct carp_softc *, struct ifnet *,
 394                     struct ifaddr *);
 395 static void     carp_unlink_addrs(struct carp_softc *, struct ifnet *,
 396                     struct ifaddr *);
 397 static void     carp_update_addrs(struct carp_softc *, struct ifaddr *);
 398
 399 static int      carp_config_vhaddr(struct carp_softc *, struct carp_vhaddr *,
 400                     struct in_ifaddr *);
 401 static int      carp_activate_vhaddr(struct carp_softc *, struct carp_vhaddr *,
 402                     struct ifnet *, struct in_ifaddr *, int);
 403 static void     carp_deactivate_vhaddr(struct carp_softc *,
 404                     struct carp_vhaddr *, boolean_t);
 405 static int      carp_addroute_vhaddr(struct carp_softc *, struct carp_vhaddr *);
 406 static void     carp_delroute_vhaddr(struct carp_softc *, struct carp_vhaddr *,
 407                     boolean_t);
 408
 409 #ifdef foo
 410 static void     carp_sc_state(struct carp_softc *);
 411 #endif
 412 #ifdef INET6
 413 static void     carp_send_na(struct carp_softc *);
 414 #ifdef notyet
 415 static int      carp_set_addr6(struct carp_softc *, struct sockaddr_in6 *);
 416 static int      carp_del_addr6(struct carp_softc *, struct sockaddr_in6 *);
 417 #endif
 418 static void     carp_multicast6_cleanup(struct carp_softc *);
 419 #endif
 420 static void     carp_stop(struct carp_softc *, boolean_t);
 421 static void     carp_suspend(struct carp_softc *, boolean_t);
 422 static void     carp_ioctl_stop(struct carp_softc *);
 423 static int      carp_ioctl_setvh(struct carp_softc *, void *, struct ucred *);
 424 static void     carp_ioctl_ifcap(struct carp_softc *, int);
 425 static int      carp_ioctl_getvh(struct carp_softc *, void *, struct ucred *);
 426 static int      carp_ioctl_getdevname(struct carp_softc *, struct ifdrv *);
 427 static int      carp_ioctl_getvhaddr(struct carp_softc *, struct ifdrv *);
 428
 429 static struct carp_if *carp_if_remove(struct carp_if *, struct carp_softc *);
 430 static struct carp_if *carp_if_insert(struct carp_if *, struct carp_softc *);
 431 static void     carp_if_free(struct carp_if *);
 432
 433 static void     carp_ifaddr(void *, struct ifnet *, enum ifaddr_event,
 434                             struct ifaddr *);
 435 static void     carp_ifdetach(void *, struct ifnet *);
 436
 437 static void     carp_ifdetach_dispatch(netmsg_t);
 438 static void     carp_clone_destroy_dispatch(netmsg_t);
 439 static void     carp_init_dispatch(netmsg_t);
 440 static void     carp_ioctl_stop_dispatch(netmsg_t);
 441 static void     carp_ioctl_setvh_dispatch(netmsg_t);
 442 static void     carp_ioctl_ifcap_dispatch(netmsg_t);
 443 static void     carp_ioctl_getvh_dispatch(netmsg_t);
 444 static void     carp_ioctl_getdevname_dispatch(netmsg_t);
 445 static void     carp_ioctl_getvhaddr_dispatch(netmsg_t);
 446 static void     carp_send_ad_timeout_dispatch(netmsg_t);
 447 static void     carp_master_down_timeout_dispatch(netmsg_t);
 448
 449 static MALLOC_DEFINE(M_CARP, "CARP", "CARP interfaces");
 450
 451 static LIST_HEAD(, carp_softc) carpif_list;
 452
 453 static struct if_clone carp_cloner =
 454 IF_CLONE_INITIALIZER(CARP_IFNAME, carp_clone_create, carp_clone_destroy,
 455                      0, IF_MAXUNIT);
 456
 457 static const uint8_t    carp_etheraddr[ETHER_ADDR_LEN] =
 458         { 0, 0, 0x5e, 0, 1, 0 };
 459
 460 static eventhandler_tag carp_ifdetach_event;
 461 static eventhandler_tag carp_ifaddr_event;
 462
 463 static __inline void
 464 carp_insert_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha_new)
 465 {
 466         struct carp_vhaddr *vha;
 467         u_long new_addr, addr;
 468
 469         KKASSERT((vha_new->vha_flags & CARP_VHAF_ONLIST) == 0);
 470
 471         /*
 472          * Virtual address list is sorted; smaller one first
 473          */
 474         new_addr = ntohl(vha_new->vha_ia->ia_addr.sin_addr.s_addr);
 475
 476         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
 477                 addr = ntohl(vha->vha_ia->ia_addr.sin_addr.s_addr);
 478
 479                 if (addr > new_addr)
 480                         break;
 481         }
 482         if (vha == NULL)
 483                 TAILQ_INSERT_TAIL(&sc->sc_vha_list, vha_new, vha_link);
 484         else
 485                 TAILQ_INSERT_BEFORE(vha, vha_new, vha_link);
 486         vha_new->vha_flags |= CARP_VHAF_ONLIST;
 487 }
 488
 489 static __inline void
 490 carp_remove_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha)
 491 {
 492         KKASSERT(vha->vha_flags & CARP_VHAF_ONLIST);
 493         vha->vha_flags &= ~CARP_VHAF_ONLIST;
 494         TAILQ_REMOVE(&sc->sc_vha_list, vha, vha_link);
 495 }
 496
 497 static void
 498 carp_hmac_prepare(struct carp_softc *sc)
 499 {
 500         uint8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 501         uint8_t vhid = sc->sc_vhid & 0xff;
 502         int i;
 503 #ifdef INET6
 504         struct ifaddr_container *ifac;
 505         struct in6_addr in6;
 506 #endif
 507 #ifdef INET
 508         struct carp_vhaddr *vha;
 509 #endif
 510
 511         /* XXX: possible race here */
 512
 513         /* compute ipad from key */
 514         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 515         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 516         for (i = 0; i < sizeof(sc->sc_pad); i++)
 517                 sc->sc_pad[i] ^= 0x36;
 518
 519         /* precompute first part of inner hash */
 520         SHA1Init(&sc->sc_sha1);
 521         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 522         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 523         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 524         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 525 #ifdef INET
 526         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
 527                 SHA1Update(&sc->sc_sha1,
 528                     (const uint8_t *)&vha->vha_ia->ia_addr.sin_addr,
 529                     sizeof(struct in_addr));
 530         }
 531 #endif /* INET */
 532 #ifdef INET6
 533         TAILQ_FOREACH(ifac, &sc->sc_if.if_addrheads[mycpuid], ifa_link) {
 534                 struct ifaddr *ifa = ifac->ifa;
 535
 536                 if (ifa->ifa_addr->sa_family == AF_INET6) {
 537                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 538                         in6_clearscope(&in6);
 539                         SHA1Update(&sc->sc_sha1, (void *)&in6, sizeof(in6));
 540                 }
 541         }
 542 #endif /* INET6 */
 543
 544         /* convert ipad to opad */
 545         for (i = 0; i < sizeof(sc->sc_pad); i++)
 546                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 547 }
 548
 549 static void
 550 carp_hmac_generate(struct carp_softc *sc, uint32_t counter[2],
 551     unsigned char md[20])
 552 {
 553         SHA1_CTX sha1ctx;
 554
 555         /* fetch first half of inner hash */
 556         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 557
 558         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 559         SHA1Final(md, &sha1ctx);
 560
 561         /* outer hash */
 562         SHA1Init(&sha1ctx);
 563         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 564         SHA1Update(&sha1ctx, md, 20);
 565         SHA1Final(md, &sha1ctx);
 566 }
 567
 568 static int
 569 carp_hmac_verify(struct carp_softc *sc, uint32_t counter[2],
 570     unsigned char md[20])
 571 {
 572         unsigned char md2[20];
 573
 574         carp_hmac_generate(sc, counter, md2);
 575         return (bcmp(md, md2, sizeof(md2)));
 576 }
 577
 578 static void
 579 carp_setroute(struct carp_softc *sc, int cmd)
 580 {
 581 #ifdef INET6
 582         struct ifaddr_container *ifac;
 583 #endif
 584         struct carp_vhaddr *vha;
 585
 586         KKASSERT(cmd == RTM_DELETE || cmd == RTM_ADD);
 587
 588         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
 589                 if (vha->vha_iaback == NULL)
 590                         continue;
 591                 if (cmd == RTM_DELETE)
 592                         carp_delroute_vhaddr(sc, vha, FALSE);
 593                 else
 594                         carp_addroute_vhaddr(sc, vha);
 595         }
 596
 597 #ifdef INET6
 598         TAILQ_FOREACH(ifac, &sc->sc_if.if_addrheads[mycpuid], ifa_link) {
 599                 struct ifaddr *ifa = ifac->ifa;
 600
 601                 if (ifa->ifa_addr->sa_family == AF_INET6) {
 602                         if (cmd == RTM_ADD)
 603                                 in6_ifaddloop(ifa);
 604                         else
 605                                 in6_ifremloop(ifa);
 606                 }
 607         }
 608 #endif /* INET6 */
 609 }
 610
 611 static int
 612 carp_clone_create(struct if_clone *ifc, int unit, caddr_t param __unused)
 613 {
 614         struct carp_softc *sc;
 615         struct ifnet *ifp;
 616
 617         sc = kmalloc(sizeof(*sc), M_CARP, M_WAITOK | M_ZERO);
 618         ifp = &sc->sc_if;
 619
 620         sc->sc_suppress = 0;
 621         sc->sc_advbase = CARP_DFLTINTV;
 622         sc->sc_vhid = -1;       /* required setting */
 623         sc->sc_advskew = 0;
 624         sc->sc_init_counter = 1;
 625         sc->sc_naddrs = 0;
 626         sc->sc_naddrs6 = 0;
 627
 628         TAILQ_INIT(&sc->sc_vha_list);
 629
 630 #ifdef INET6
 631         sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL;
 632 #endif
 633
 634         callout_init_mp(&sc->sc_ad_tmo);
 635         netmsg_init(&sc->sc_ad_msg.base, NULL, &netisr_adone_rport,
 636             MSGF_DROPABLE | MSGF_PRIORITY, carp_send_ad_timeout_dispatch);
 637         sc->sc_ad_msg.nc_softc = sc;
 638
 639         callout_init_mp(&sc->sc_md_tmo);
 640         callout_init_mp(&sc->sc_md6_tmo);
 641         netmsg_init(&sc->sc_md_msg.base, NULL, &netisr_adone_rport,
 642             MSGF_DROPABLE | MSGF_PRIORITY, carp_master_down_timeout_dispatch);
 643         sc->sc_md_msg.nc_softc = sc;
 644
 645         if_initname(ifp, CARP_IFNAME, unit);
 646         ifp->if_softc = sc;
 647         ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
 648         ifp->if_init = carp_init;
 649         ifp->if_ioctl = carp_ioctl;
 650         ifp->if_start = carp_start;
 651
 652         ifp->if_capabilities = IFCAP_TXCSUM | IFCAP_TSO;
 653         ifp->if_capenable = ifp->if_capabilities;
 654         /*
 655          * Leave if_hwassist as it is; if_hwassist will be
 656          * setup when this carp interface has parent.
 657          */
 658
 659         ifq_set_maxlen(&ifp->if_snd, ifqmaxlen);
 660         ifq_set_ready(&ifp->if_snd);
 661
 662         ether_ifattach(ifp, carp_etheraddr, NULL);
 663
 664         ifp->if_type = IFT_CARP;
 665         ifp->if_output = carp_output;
 666
 667         lwkt_gettoken(&carp_listtok);
 668         LIST_INSERT_HEAD(&carpif_list, sc, sc_next);
 669         lwkt_reltoken(&carp_listtok);
 670
 671         return (0);
 672 }
 673
 674 static void
 675 carp_clone_destroy_dispatch(netmsg_t msg)
 676 {
 677         struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
 678         struct carp_softc *sc = cmsg->nc_softc;
 679
 680         sc->sc_dead = TRUE;
 681         carp_detach(sc, TRUE, FALSE);
 682
 683         callout_stop_sync(&sc->sc_ad_tmo);
 684         callout_stop_sync(&sc->sc_md_tmo);
 685         callout_stop_sync(&sc->sc_md6_tmo);
 686
 687         crit_enter();
 688         lwkt_dropmsg(&sc->sc_ad_msg.base.lmsg);
 689         lwkt_dropmsg(&sc->sc_md_msg.base.lmsg);
 690         crit_exit();
 691
 692         lwkt_replymsg(&cmsg->base.lmsg, 0);
 693 }
 694
 695 static int
 696 carp_clone_destroy(struct ifnet *ifp)
 697 {
 698         struct carp_softc *sc = ifp->if_softc;
 699         struct netmsg_carp cmsg;
 700
 701         bzero(&cmsg, sizeof(cmsg));
 702         netmsg_init(&cmsg.base, NULL, &curthread->td_msgport, 0,
 703             carp_clone_destroy_dispatch);
 704         cmsg.nc_softc = sc;
 705
 706         lwkt_domsg(netisr_cpuport(0), &cmsg.base.lmsg, 0);
 707
 708         lwkt_gettoken(&carp_listtok);
 709         LIST_REMOVE(sc, sc_next);
 710         lwkt_reltoken(&carp_listtok);
 711
 712         bpfdetach(ifp);
 713         if_detach(ifp);
 714
 715         KASSERT(sc->sc_naddrs == 0, ("certain inet address is still active"));
 716         kfree(sc, M_CARP);
 717
 718         return 0;
 719 }
 720
 721 static struct carp_if *
 722 carp_if_remove(struct carp_if *ocif, struct carp_softc *sc)
 723 {
 724         struct carp_softc_container *oscc, *scc;
 725         struct carp_if *cif;
 726         int count = 0;
 727 #ifdef INVARIANTS
 728         int found = 0;
 729 #endif
 730
 731         TAILQ_FOREACH(oscc, ocif, scc_link) {
 732                 ++count;
 733 #ifdef INVARIANTS
 734                 if (oscc->scc_softc == sc)
 735                         found = 1;
 736 #endif
 737         }
 738         KASSERT(found, ("%s carp_softc is not on carp_if", __func__));
 739
 740         if (count == 1) {
 741                 /* Last one is going to be unlinked */
 742                 return NULL;
 743         }
 744
 745         cif = kmalloc(sizeof(*cif), M_CARP, M_WAITOK | M_ZERO);
 746         TAILQ_INIT(cif);
 747
 748         TAILQ_FOREACH(oscc, ocif, scc_link) {
 749                 if (oscc->scc_softc == sc)
 750                         continue;
 751
 752                 scc = kmalloc(sizeof(*scc), M_CARP, M_WAITOK | M_ZERO);
 753                 scc->scc_softc = oscc->scc_softc;
 754                 TAILQ_INSERT_TAIL(cif, scc, scc_link);
 755         }
 756
 757         return cif;
 758 }
 759
 760 static struct carp_if *
 761 carp_if_insert(struct carp_if *ocif, struct carp_softc *sc)
 762 {
 763         struct carp_softc_container *oscc;
 764         int onlist;
 765
 766         onlist = 0;
 767         if (ocif != NULL) {
 768                 TAILQ_FOREACH(oscc, ocif, scc_link) {
 769                         if (oscc->scc_softc == sc)
 770                                 onlist = 1;
 771                 }
 772         }
 773
 774 #ifdef INVARIANTS
 775         if (sc->sc_carpdev != NULL) {
 776                 KASSERT(onlist, ("%s is not on %s carp list",
 777                     sc->sc_if.if_xname, sc->sc_carpdev->if_xname));
 778         } else {
 779                 KASSERT(!onlist, ("%s is already on carp list",
 780                     sc->sc_if.if_xname));
 781         }
 782 #endif
 783
 784         if (!onlist) {
 785                 struct carp_if *cif;
 786                 struct carp_softc_container *new_scc, *scc;
 787                 int inserted = 0;
 788
 789                 cif = kmalloc(sizeof(*cif), M_CARP, M_WAITOK | M_ZERO);
 790                 TAILQ_INIT(cif);
 791
 792                 new_scc = kmalloc(sizeof(*new_scc), M_CARP, M_WAITOK | M_ZERO);
 793                 new_scc->scc_softc = sc;
 794
 795                 if (ocif != NULL) {
 796                         TAILQ_FOREACH(oscc, ocif, scc_link) {
 797                                 if (!inserted &&
 798                                     oscc->scc_softc->sc_vhid > sc->sc_vhid) {
 799                                         TAILQ_INSERT_TAIL(cif, new_scc,
 800                                             scc_link);
 801                                         inserted = 1;
 802                                 }
 803
 804                                 scc = kmalloc(sizeof(*scc), M_CARP,
 805                                     M_WAITOK | M_ZERO);
 806                                 scc->scc_softc = oscc->scc_softc;
 807                                 TAILQ_INSERT_TAIL(cif, scc, scc_link);
 808                         }
 809                 }
 810                 if (!inserted)
 811                         TAILQ_INSERT_TAIL(cif, new_scc, scc_link);
 812
 813                 return cif;
 814         } else {
 815                 return ocif;
 816         }
 817 }
 818
 819 static void
 820 carp_if_free(struct carp_if *cif)
 821 {
 822         struct carp_softc_container *scc;
 823
 824         while ((scc = TAILQ_FIRST(cif)) != NULL) {
 825                 TAILQ_REMOVE(cif, scc, scc_link);
 826                 kfree(scc, M_CARP);
 827         }
 828         kfree(cif, M_CARP);
 829 }
 830
 831 static void
 832 carp_detach(struct carp_softc *sc, boolean_t detach, boolean_t del_iaback)
 833 {
 834         carp_suspend(sc, detach);
 835
 836         carp_multicast_cleanup(sc);
 837 #ifdef INET6
 838         carp_multicast6_cleanup(sc);
 839 #endif
 840
 841         if (!sc->sc_dead && detach) {
 842                 struct carp_vhaddr *vha;
 843
 844                 TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link)
 845                         carp_deactivate_vhaddr(sc, vha, del_iaback);
 846                 KKASSERT(sc->sc_naddrs == 0);
 847         }
 848
 849         if (sc->sc_carpdev != NULL) {
 850                 struct ifnet *ifp = sc->sc_carpdev;
 851                 struct carp_if *ocif = ifp->if_carp;
 852
 853                 ifp->if_carp = carp_if_remove(ocif, sc);
 854                 KASSERT(ifp->if_carp != ocif,
 855                     ("%s carp_if_remove failed", __func__));
 856
 857                 sc->sc_carpdev = NULL;
 858                 sc->sc_ia = NULL;
 859                 sc->arpcom.ac_if.if_hwassist = 0;
 860
 861                 /*
 862                  * Make sure that all protocol threads see the
 863                  * sc_carpdev and if_carp changes
 864                  */
 865                 netmsg_service_sync();
 866
 867                 if (ifp->if_carp == NULL) {
 868                         /*
 869                          * No more carp interfaces using
 870                          * ifp as the backing interface,
 871                          * move it out of promiscous mode.
 872                          */
 873                         ifpromisc(ifp, 0);
 874                 }
 875
 876                 /*
 877                  * The old carp list could be safely free now,
 878                  * since no one can access it.
 879                  */
 880                 carp_if_free(ocif);
 881         }
 882 }
 883
 884 static void
 885 carp_ifdetach_dispatch(netmsg_t msg)
 886 {
 887         struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
 888         struct ifnet *ifp = cmsg->nc_carpdev;
 889
 890         while (ifp->if_carp) {
 891                 struct carp_softc_container *scc;
 892
 893                 scc = TAILQ_FIRST((struct carp_if *)(ifp->if_carp));
 894                 carp_detach(scc->scc_softc, TRUE, TRUE);
 895         }
 896         lwkt_replymsg(&cmsg->base.lmsg, 0);
 897 }
 898
 899 /* Detach an interface from the carp. */
 900 static void
 901 carp_ifdetach(void *arg __unused, struct ifnet *ifp)
 902 {
 903         struct netmsg_carp cmsg;
 904
 905         ASSERT_IFNET_NOT_SERIALIZED_ALL(ifp);
 906
 907         bzero(&cmsg, sizeof(cmsg));
 908         netmsg_init(&cmsg.base, NULL, &curthread->td_msgport, 0,
 909             carp_ifdetach_dispatch);
 910         cmsg.nc_carpdev = ifp;
 911
 912         lwkt_domsg(netisr_cpuport(0), &cmsg.base.lmsg, 0);
 913 }
 914
 915 /*
 916  * process input packet.
 917  * we have rearranged checks order compared to the rfc,
 918  * but it seems more efficient this way or not possible otherwise.
 919  */
 920 int
 921 carp_proto_input(struct mbuf **mp, int *offp, int proto)
 922 {
 923         struct mbuf *m = *mp;
 924         struct ip *ip = mtod(m, struct ip *);
 925         struct ifnet *ifp = m->m_pkthdr.rcvif;
 926         struct carp_header *ch;
 927         struct carp_softc *sc;
 928         int len, iphlen;
 929
 930         iphlen = *offp;
 931         *mp = NULL;
 932
 933         carpstats.carps_ipackets++;
 934
 935         if (!carp_opts[CARPCTL_ALLOW]) {
 936                 m_freem(m);
 937                 goto back;
 938         }
 939
 940         /* Check if received on a valid carp interface */
 941         if (ifp->if_type != IFT_CARP) {
 942                 carpstats.carps_badif++;
 943                 CARP_LOG("carp_proto_input: packet received on non-carp "
 944                     "interface: %s\n", ifp->if_xname);
 945                 m_freem(m);
 946                 goto back;
 947         }
 948
 949         if (!CARP_IS_RUNNING(ifp)) {
 950                 carpstats.carps_badif++;
 951                 CARP_LOG("carp_proto_input: packet received on stopped carp "
 952                     "interface: %s\n", ifp->if_xname);
 953                 m_freem(m);
 954                 goto back;
 955         }
 956
 957         sc = ifp->if_softc;
 958         if (sc->sc_carpdev == NULL) {
 959                 carpstats.carps_badif++;
 960                 CARP_LOG("carp_proto_input: packet received on defunc carp "
 961                     "interface: %s\n", ifp->if_xname);
 962                 m_freem(m);
 963                 goto back;
 964         }
 965
 966         if (!IN_MULTICAST(ntohl(ip->ip_dst.s_addr))) {
 967                 carpstats.carps_badif++;
 968                 CARP_LOG("carp_proto_input: non-mcast packet on "
 969                     "interface: %s\n", ifp->if_xname);
 970                 m_freem(m);
 971                 goto back;
 972         }
 973
 974         /* Verify that the IP TTL is CARP_DFLTTL. */
 975         if (ip->ip_ttl != CARP_DFLTTL) {
 976                 carpstats.carps_badttl++;
 977                 CARP_LOG("carp_proto_input: received ttl %d != %d on %s\n",
 978                     ip->ip_ttl, CARP_DFLTTL, ifp->if_xname);
 979                 m_freem(m);
 980                 goto back;
 981         }
 982
 983         /* Minimal CARP packet size */
 984         len = iphlen + sizeof(*ch);
 985
 986         /*
 987          * Verify that the received packet length is
 988          * not less than the CARP header
 989          */
 990         if (m->m_pkthdr.len < len) {
 991                 carpstats.carps_badlen++;
 992                 CARP_LOG("packet too short %d on %s\n", m->m_pkthdr.len,
 993                     ifp->if_xname);
 994                 m_freem(m);
 995                 goto back;
 996         }
 997
 998         /* Make sure that CARP header is contiguous */
 999         if (len > m->m_len) {
1000                 m = m_pullup(m, len);
1001                 if (m == NULL) {
1002                         carpstats.carps_hdrops++;
1003                         CARP_LOG("carp_proto_input: m_pullup failed\n");
1004                         goto back;
1005                 }
1006                 ip = mtod(m, struct ip *);
1007         }
1008         ch = (struct carp_header *)((uint8_t *)ip + iphlen);
1009
1010         /* Verify the CARP checksum */
1011         if (in_cksum_skip(m, len, iphlen)) {
1012                 carpstats.carps_badsum++;
1013                 CARP_LOG("carp_proto_input: checksum failed on %s\n",
1014                     ifp->if_xname);
1015                 m_freem(m);
1016                 goto back;
1017         }
1018         carp_proto_input_c(sc, m, ch, AF_INET);
1019 back:
1020         return(IPPROTO_DONE);
1021 }
1022
1023 #ifdef INET6
1024 int
1025 carp6_proto_input(struct mbuf **mp, int *offp, int proto)
1026 {
1027         struct mbuf *m = *mp;
1028         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1029         struct ifnet *ifp = m->m_pkthdr.rcvif;
1030         struct carp_header *ch;
1031         struct carp_softc *sc;
1032         u_int len;
1033
1034         carpstats.carps_ipackets6++;
1035
1036         if (!carp_opts[CARPCTL_ALLOW]) {
1037                 m_freem(m);
1038                 goto back;
1039         }
1040
1041         /* check if received on a valid carp interface */
1042         if (ifp->if_type != IFT_CARP) {
1043                 carpstats.carps_badif++;
1044                 CARP_LOG("carp6_proto_input: packet received on non-carp "
1045                     "interface: %s\n", ifp->if_xname);
1046                 m_freem(m);
1047                 goto back;
1048         }
1049
1050         if (!CARP_IS_RUNNING(ifp)) {
1051                 carpstats.carps_badif++;
1052                 CARP_LOG("carp_proto_input: packet received on stopped carp "
1053                     "interface: %s\n", ifp->if_xname);
1054                 m_freem(m);
1055                 goto back;
1056         }
1057
1058         sc = ifp->if_softc;
1059         if (sc->sc_carpdev == NULL) {
1060                 carpstats.carps_badif++;
1061                 CARP_LOG("carp6_proto_input: packet received on defunc-carp "
1062                     "interface: %s\n", ifp->if_xname);
1063                 m_freem(m);
1064                 goto back;
1065         }
1066
1067         /* verify that the IP TTL is 255 */
1068         if (ip6->ip6_hlim != CARP_DFLTTL) {
1069                 carpstats.carps_badttl++;
1070                 CARP_LOG("carp6_proto_input: received ttl %d != 255 on %s\n",
1071                     ip6->ip6_hlim, ifp->if_xname);
1072                 m_freem(m);
1073                 goto back;
1074         }
1075
1076         /* verify that we have a complete carp packet */
1077         len = m->m_len;
1078         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
1079         if (ch == NULL) {
1080                 carpstats.carps_badlen++;
1081                 CARP_LOG("carp6_proto_input: packet size %u too small\n", len);
1082                 goto back;
1083         }
1084
1085         /* verify the CARP checksum */
1086         if (in_cksum_range(m, 0, *offp, sizeof(*ch))) {
1087                 carpstats.carps_badsum++;
1088                 CARP_LOG("carp6_proto_input: checksum failed, on %s\n",
1089                     ifp->if_xname);
1090                 m_freem(m);
1091                 goto back;
1092         }
1093
1094         carp_proto_input_c(sc, m, ch, AF_INET6);
1095 back:
1096         return (IPPROTO_DONE);
1097 }
1098 #endif /* INET6 */
1099
1100 static void
1101 carp_proto_input_c(struct carp_softc *sc, struct mbuf *m,
1102     struct carp_header *ch, sa_family_t af)
1103 {
1104         struct ifnet *cifp;
1105         uint64_t tmp_counter;
1106         struct timeval sc_tv, ch_tv;
1107
1108         if (sc->sc_vhid != ch->carp_vhid) {
1109                 /*
1110                  * CARP uses multicast, however, multicast packets
1111                  * are tapped to all CARP interfaces on the physical
1112                  * interface receiving the CARP packets, so we don't
1113                  * update any stats here.
1114                  */
1115                 m_freem(m);
1116                 return;
1117         }
1118         cifp = &sc->sc_if;
1119
1120         /* verify the CARP version. */
1121         if (ch->carp_version != CARP_VERSION) {
1122                 carpstats.carps_badver++;
1123                 CARP_LOG("%s; invalid version %d\n", cifp->if_xname,
1124                          ch->carp_version);
1125                 m_freem(m);
1126                 return;
1127         }
1128
1129         /* verify the hash */
1130         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
1131                 carpstats.carps_badauth++;
1132                 CARP_LOG("%s: incorrect hash\n", cifp->if_xname);
1133                 m_freem(m);
1134                 return;
1135         }
1136
1137         tmp_counter = ntohl(ch->carp_counter[0]);
1138         tmp_counter = tmp_counter<<32;
1139         tmp_counter += ntohl(ch->carp_counter[1]);
1140
1141         /* XXX Replay protection goes here */
1142
1143         sc->sc_init_counter = 0;
1144         sc->sc_counter = tmp_counter;
1145
1146         sc_tv.tv_sec = sc->sc_advbase;
1147         if (carp_suppress_preempt && sc->sc_advskew <  240)
1148                 sc_tv.tv_usec = 240 * 1000000 / 256;
1149         else
1150                 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1151         ch_tv.tv_sec = ch->carp_advbase;
1152         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
1153
1154         switch (sc->sc_state) {
1155         case INIT:
1156                 break;
1157
1158         case MASTER:
1159                 /*
1160                  * If we receive an advertisement from a master who's going to
1161                  * be more frequent than us, go into BACKUP state.
1162                  */
1163                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
1164                     timevalcmp(&sc_tv, &ch_tv, ==)) {
1165                         callout_stop(&sc->sc_ad_tmo);
1166                         CARP_DEBUG("%s: MASTER -> BACKUP "
1167                            "(more frequent advertisement received)\n",
1168                            cifp->if_xname);
1169                         carp_set_state(sc, BACKUP);
1170                         carp_setrun(sc, 0);
1171                         carp_setroute(sc, RTM_DELETE);
1172                 }
1173                 break;
1174
1175         case BACKUP:
1176                 /*
1177                  * If we're pre-empting masters who advertise slower than us,
1178                  * and this one claims to be slower, treat him as down.
1179                  */
1180                 if (carp_opts[CARPCTL_PREEMPT] &&
1181                     timevalcmp(&sc_tv, &ch_tv, <)) {
1182                         CARP_DEBUG("%s: BACKUP -> MASTER "
1183                             "(preempting a slower master)\n", cifp->if_xname);
1184                         carp_master_down(sc);
1185                         break;
1186                 }
1187
1188                 /*
1189                  *  If the master is going to advertise at such a low frequency
1190                  *  that he's guaranteed to time out, we'd might as well just
1191                  *  treat him as timed out now.
1192                  */
1193                 sc_tv.tv_sec = sc->sc_advbase * 3;
1194                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
1195                         CARP_DEBUG("%s: BACKUP -> MASTER (master timed out)\n",
1196                                    cifp->if_xname);
1197                         carp_master_down(sc);
1198                         break;
1199                 }
1200
1201                 /*
1202                  * Otherwise, we reset the counter and wait for the next
1203                  * advertisement.
1204                  */
1205                 carp_setrun(sc, af);
1206                 break;
1207         }
1208         m_freem(m);
1209 }
1210
1211 struct mbuf *
1212 carp_input(void *v, struct mbuf *m)
1213 {
1214         struct carp_if *cif = v;
1215         struct ether_header *eh;
1216         struct carp_softc_container *scc;
1217         struct ifnet *ifp;
1218
1219         eh = mtod(m, struct ether_header *);
1220
1221         ifp = carp_forus(cif, eh->ether_dhost);
1222         if (ifp != NULL) {
1223                 ether_reinput_oncpu(ifp, m, REINPUT_RUNBPF);
1224                 return NULL;
1225         }
1226
1227         if ((m->m_flags & (M_BCAST | M_MCAST)) == 0)
1228                 return m;
1229
1230         /*
1231          * XXX Should really check the list of multicast addresses
1232          * for each CARP interface _before_ copying.
1233          */
1234         TAILQ_FOREACH(scc, cif, scc_link) {
1235                 struct carp_softc *sc = scc->scc_softc;
1236                 struct mbuf *m0;
1237
1238                 if ((sc->sc_if.if_flags & IFF_UP) == 0)
1239                         continue;
1240
1241                 m0 = m_dup(m, MB_DONTWAIT);
1242                 if (m0 == NULL)
1243                         continue;
1244
1245                 ether_reinput_oncpu(&sc->sc_if, m0, REINPUT_RUNBPF);
1246         }
1247         return m;
1248 }
1249
1250 static void
1251 carp_prepare_ad(struct carp_softc *sc, struct carp_header *ch)
1252 {
1253         if (sc->sc_init_counter) {
1254                 /* this could also be seconds since unix epoch */
1255                 sc->sc_counter = karc4random();
1256                 sc->sc_counter = sc->sc_counter << 32;
1257                 sc->sc_counter += karc4random();
1258         } else {
1259                 sc->sc_counter++;
1260         }
1261
1262         ch->carp_counter[0] = htonl((sc->sc_counter >> 32) & 0xffffffff);
1263         ch->carp_counter[1] = htonl(sc->sc_counter & 0xffffffff);
1264
1265         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
1266 }
1267
1268 static void
1269 carp_send_ad_all(void)
1270 {
1271         struct carp_softc *sc;
1272
1273         LIST_FOREACH(sc, &carpif_list, sc_next) {
1274                 if (sc->sc_carpdev == NULL)
1275                         continue;
1276
1277                 if (CARP_IS_RUNNING(&sc->sc_if) && sc->sc_state == MASTER)
1278                         carp_send_ad(sc);
1279         }
1280 }
1281
1282 static void
1283 carp_send_ad_timeout(void *xsc)
1284 {
1285         struct carp_softc *sc = xsc;
1286         struct netmsg_carp *cmsg = &sc->sc_ad_msg;
1287
1288         KASSERT(mycpuid == 0, ("%s not on cpu0 but on cpu%d",
1289             __func__, mycpuid));
1290
1291         crit_enter();
1292         if (cmsg->base.lmsg.ms_flags & MSGF_DONE)
1293                 lwkt_sendmsg(netisr_cpuport(0), &cmsg->base.lmsg);
1294         crit_exit();
1295 }
1296
1297 static void
1298 carp_send_ad_timeout_dispatch(netmsg_t msg)
1299 {
1300         struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
1301         struct carp_softc *sc = cmsg->nc_softc;
1302
1303         /* Reply ASAP */
1304         crit_enter();
1305         lwkt_replymsg(&cmsg->base.lmsg, 0);
1306         crit_exit();
1307
1308         carp_send_ad(sc);
1309 }
1310
1311 static void
1312 carp_send_ad(struct carp_softc *sc)
1313 {
1314         struct ifnet *cifp = &sc->sc_if;
1315         struct carp_header ch;
1316         struct timeval tv;
1317         struct carp_header *ch_ptr;
1318         struct mbuf *m;
1319         int len, advbase, advskew;
1320
1321         if (!CARP_IS_RUNNING(cifp)) {
1322                 /* Bow out */
1323                 advbase = 255;
1324                 advskew = 255;
1325         } else {
1326                 advbase = sc->sc_advbase;
1327                 if (!carp_suppress_preempt || sc->sc_advskew > 240)
1328                         advskew = sc->sc_advskew;
1329                 else
1330                         advskew = 240;
1331                 tv.tv_sec = advbase;
1332                 tv.tv_usec = advskew * 1000000 / 256;
1333         }
1334
1335         ch.carp_version = CARP_VERSION;
1336         ch.carp_type = CARP_ADVERTISEMENT;
1337         ch.carp_vhid = sc->sc_vhid;
1338         ch.carp_advbase = advbase;
1339         ch.carp_advskew = advskew;
1340         ch.carp_authlen = 7;    /* XXX DEFINE */
1341         ch.carp_pad1 = 0;       /* must be zero */
1342         ch.carp_cksum = 0;
1343
1344 #ifdef INET
1345         if (sc->sc_ia != NULL) {
1346                 struct ip *ip;
1347
1348                 MGETHDR(m, MB_DONTWAIT, MT_HEADER);
1349                 if (m == NULL) {
1350                         IFNET_STAT_INC(cifp, oerrors, 1);
1351                         carpstats.carps_onomem++;
1352                         /* XXX maybe less ? */
1353                         if (advbase != 255 || advskew != 255)
1354                                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
1355                                     carp_send_ad_timeout, sc);
1356                         return;
1357                 }
1358                 len = sizeof(*ip) + sizeof(ch);
1359                 m->m_pkthdr.len = len;
1360                 m->m_pkthdr.rcvif = NULL;
1361                 m->m_len = len;
1362                 MH_ALIGN(m, m->m_len);
1363                 m->m_flags |= M_MCAST;
1364                 ip = mtod(m, struct ip *);
1365                 ip->ip_v = IPVERSION;
1366                 ip->ip_hl = sizeof(*ip) >> 2;
1367                 ip->ip_tos = IPTOS_LOWDELAY;
1368                 ip->ip_len = len;
1369                 ip->ip_id = ip_newid();
1370                 ip->ip_off = IP_DF;
1371                 ip->ip_ttl = CARP_DFLTTL;
1372                 ip->ip_p = IPPROTO_CARP;
1373                 ip->ip_sum = 0;
1374                 ip->ip_src = sc->sc_ia->ia_addr.sin_addr;
1375                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
1376
1377                 ch_ptr = (struct carp_header *)(&ip[1]);
1378                 bcopy(&ch, ch_ptr, sizeof(ch));
1379                 carp_prepare_ad(sc, ch_ptr);
1380                 ch_ptr->carp_cksum = in_cksum_skip(m, len, sizeof(*ip));
1381
1382                 getmicrotime(&cifp->if_lastchange);
1383                 IFNET_STAT_INC(cifp, opackets, 1);
1384                 IFNET_STAT_INC(cifp, obytes, len);
1385                 carpstats.carps_opackets++;
1386
1387                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, NULL)) {
1388                         IFNET_STAT_INC(cifp, oerrors, 1);
1389                         if (sc->sc_sendad_errors < INT_MAX)
1390                                 sc->sc_sendad_errors++;
1391                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
1392                                 carp_suppress_preempt++;
1393                                 if (carp_suppress_preempt == 1) {
1394                                         carp_send_ad_all();
1395                                 }
1396                         }
1397                         sc->sc_sendad_success = 0;
1398                 } else {
1399                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
1400                                 if (++sc->sc_sendad_success >=
1401                                     CARP_SENDAD_MIN_SUCCESS) {
1402                                         carp_suppress_preempt--;
1403                                         sc->sc_sendad_errors = 0;
1404                                 }
1405                         } else {
1406                                 sc->sc_sendad_errors = 0;
1407                         }
1408                 }
1409         }
1410 #endif /* INET */
1411 #ifdef INET6
1412         if (sc->sc_ia6) {
1413                 struct ip6_hdr *ip6;
1414
1415                 MGETHDR(m, MB_DONTWAIT, MT_HEADER);
1416                 if (m == NULL) {
1417                         IFNET_STAT_INC(cifp, oerrors, 1);
1418                         carpstats.carps_onomem++;
1419                         /* XXX maybe less ? */
1420                         if (advbase != 255 || advskew != 255)
1421                                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
1422                                     carp_send_ad_timeout, sc);
1423                         return;
1424                 }
1425                 len = sizeof(*ip6) + sizeof(ch);
1426                 m->m_pkthdr.len = len;
1427                 m->m_pkthdr.rcvif = NULL;
1428                 m->m_len = len;
1429                 MH_ALIGN(m, m->m_len);
1430                 m->m_flags |= M_MCAST;
1431                 ip6 = mtod(m, struct ip6_hdr *);
1432                 bzero(ip6, sizeof(*ip6));
1433                 ip6->ip6_vfc |= IPV6_VERSION;
1434                 ip6->ip6_hlim = CARP_DFLTTL;
1435                 ip6->ip6_nxt = IPPROTO_CARP;
1436                 bcopy(&sc->sc_ia6->ia_addr.sin6_addr, &ip6->ip6_src,
1437                     sizeof(struct in6_addr));
1438                 /* set the multicast destination */
1439
1440                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
1441                 ip6->ip6_dst.s6_addr8[15] = 0x12;
1442                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
1443                         IFNET_STAT_INC(cifp, oerrors, 1);
1444                         m_freem(m);
1445                         CARP_LOG("%s: in6_setscope failed\n", __func__);
1446                         return;
1447                 }
1448
1449                 ch_ptr = (struct carp_header *)(&ip6[1]);
1450                 bcopy(&ch, ch_ptr, sizeof(ch));
1451                 carp_prepare_ad(sc, ch_ptr);
1452                 ch_ptr->carp_cksum = in_cksum_skip(m, len, sizeof(*ip6));
1453
1454                 getmicrotime(&cifp->if_lastchange);
1455                 IFNET_STAT_INC(cifp, opackets, 1);
1456                 IFNET_STAT_INC(cifp, obytes, len);
1457                 carpstats.carps_opackets6++;
1458
1459                 if (ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL)) {
1460                         IFNET_STAT_INC(cifp, oerrors, 1);
1461                         if (sc->sc_sendad_errors < INT_MAX)
1462                                 sc->sc_sendad_errors++;
1463                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
1464                                 carp_suppress_preempt++;
1465                                 if (carp_suppress_preempt == 1) {
1466                                         carp_send_ad_all();
1467                                 }
1468                         }
1469                         sc->sc_sendad_success = 0;
1470                 } else {
1471                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
1472                                 if (++sc->sc_sendad_success >=
1473                                     CARP_SENDAD_MIN_SUCCESS) {
1474                                         carp_suppress_preempt--;
1475                                         sc->sc_sendad_errors = 0;
1476                                 }
1477                         } else {
1478                                 sc->sc_sendad_errors = 0;
1479                         }
1480                 }
1481         }
1482 #endif /* INET6 */
1483
1484         if (advbase != 255 || advskew != 255)
1485                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
1486                     carp_send_ad_timeout, sc);
1487 }
1488
1489 /*
1490  * Broadcast a gratuitous ARP request containing
1491  * the virtual router MAC address for each IP address
1492  * associated with the virtual router.
1493  */
1494 static void
1495 carp_send_arp(struct carp_softc *sc)
1496 {
1497         const struct carp_vhaddr *vha;
1498
1499         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
1500                 if (vha->vha_iaback == NULL)
1501                         continue;
1502                 arp_gratuitous(&sc->sc_if, &vha->vha_ia->ia_ifa);
1503         }
1504 }
1505
1506 #ifdef INET6
1507 static void
1508 carp_send_na(struct carp_softc *sc)
1509 {
1510         struct ifaddr_container *ifac;
1511         struct in6_addr *in6;
1512         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1513
1514         TAILQ_FOREACH(ifac, &sc->sc_if.if_addrheads[mycpuid], ifa_link) {
1515                 struct ifaddr *ifa = ifac->ifa;
1516
1517                 if (ifa->ifa_addr->sa_family != AF_INET6)
1518                         continue;
1519
1520                 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr;
1521                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1522                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1523                 DELAY(1000);    /* XXX */
1524         }
1525 }
1526 #endif /* INET6 */
1527
1528 static __inline const struct carp_vhaddr *
1529 carp_find_addr(const struct carp_softc *sc, const struct in_addr *addr)
1530 {
1531         struct carp_vhaddr *vha;
1532
1533         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
1534                 if (vha->vha_iaback == NULL)
1535                         continue;
1536
1537                 if (vha->vha_ia->ia_addr.sin_addr.s_addr == addr->s_addr)
1538                         return vha;
1539         }
1540         return NULL;
1541 }
1542
1543 #ifdef notyet
1544 static int
1545 carp_iamatch_balance(const struct carp_if *cif, const struct in_addr *itaddr,
1546                      const struct in_addr *isaddr, uint8_t **enaddr)
1547 {
1548         const struct carp_softc *vh;
1549         int index, count = 0;
1550
1551         /*
1552          * XXX proof of concept implementation.
1553          * We use the source ip to decide which virtual host should
1554          * handle the request. If we're master of that virtual host,
1555          * then we respond, otherwise, just drop the arp packet on
1556          * the floor.
1557          */
1558
1559         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1560                 if (!CARP_IS_RUNNING(&vh->sc_if))
1561                         continue;
1562
1563                 if (carp_find_addr(vh, itaddr) != NULL)
1564                         count++;
1565         }
1566         if (count == 0)
1567                 return 0;
1568
1569         /* this should be a hash, like pf_hash() */
1570         index = ntohl(isaddr->s_addr) % count;
1571         count = 0;
1572
1573         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1574                 if (!CARP_IS_RUNNING(&vh->sc_if))
1575                         continue;
1576
1577                 if (carp_find_addr(vh, itaddr) == NULL)
1578                         continue;
1579
1580                 if (count == index) {
1581                         if (vh->sc_state == MASTER) {
1582                                 *enaddr = IF_LLADDR(&vh->sc_if);
1583                                 return 1;
1584                         } else {
1585                                 return 0;
1586                         }
1587                 }
1588                 count++;
1589         }
1590         return 0;
1591 }
1592 #endif
1593
1594 int
1595 carp_iamatch(const struct in_ifaddr *ia)
1596 {
1597         const struct carp_softc *sc = ia->ia_ifp->if_softc;
1598
1599         KASSERT(&curthread->td_msgport == netisr_cpuport(0),
1600             ("not in netisr0"));
1601
1602 #ifdef notyet
1603         if (carp_opts[CARPCTL_ARPBALANCE])
1604                 return carp_iamatch_balance(cif, itaddr, isaddr, enaddr);
1605 #endif
1606
1607         if (!CARP_IS_RUNNING(&sc->sc_if) || sc->sc_state != MASTER)
1608                 return 0;
1609
1610         return 1;
1611 }
1612
1613 #ifdef INET6
1614 struct ifaddr *
1615 carp_iamatch6(void *v, struct in6_addr *taddr)
1616 {
1617 #ifdef foo
1618         struct carp_if *cif = v;
1619         struct carp_softc *vh;
1620
1621         TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1622                 struct ifaddr_container *ifac;
1623
1624                 TAILQ_FOREACH(ifac, &vh->sc_if.if_addrheads[mycpuid],
1625                               ifa_link) {
1626                         struct ifaddr *ifa = ifac->ifa;
1627
1628                         if (IN6_ARE_ADDR_EQUAL(taddr,
1629                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1630                             CARP_IS_RUNNING(&vh->sc_if) &&
1631                             vh->sc_state == MASTER) {
1632                                 return (ifa);
1633                         }
1634                 }
1635         }
1636 #endif
1637         return (NULL);
1638 }
1639
1640 void *
1641 carp_macmatch6(void *v, struct mbuf *m, const struct in6_addr *taddr)
1642 {
1643 #ifdef foo
1644         struct m_tag *mtag;
1645         struct carp_if *cif = v;
1646         struct carp_softc *sc;
1647
1648         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) {
1649                 struct ifaddr_container *ifac;
1650
1651                 TAILQ_FOREACH(ifac, &sc->sc_if.if_addrheads[mycpuid],
1652                               ifa_link) {
1653                         struct ifaddr *ifa = ifac->ifa;
1654
1655                         if (IN6_ARE_ADDR_EQUAL(taddr,
1656                             &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1657                             CARP_IS_RUNNING(&sc->sc_if)) {
1658                                 struct ifnet *ifp = &sc->sc_if;
1659
1660                                 mtag = m_tag_get(PACKET_TAG_CARP,
1661                                     sizeof(struct ifnet *), MB_DONTWAIT);
1662                                 if (mtag == NULL) {
1663                                         /* better a bit than nothing */
1664                                         return (IF_LLADDR(ifp));
1665                                 }
1666                                 bcopy(&ifp, (caddr_t)(mtag + 1),
1667                                     sizeof(struct ifnet *));
1668                                 m_tag_prepend(m, mtag);
1669
1670                                 return (IF_LLADDR(ifp));
1671                         }
1672                 }
1673         }
1674 #endif
1675         return (NULL);
1676 }
1677 #endif
1678
1679 static struct ifnet *
1680 carp_forus(struct carp_if *cif, const uint8_t *dhost)
1681 {
1682         struct carp_softc_container *scc;
1683
1684         if (memcmp(dhost, carp_etheraddr, ETHER_ADDR_LEN - 1) != 0)
1685                 return NULL;
1686
1687         TAILQ_FOREACH(scc, cif, scc_link) {
1688                 struct carp_softc *sc = scc->scc_softc;
1689                 struct ifnet *ifp = &sc->sc_if;
1690
1691                 if (CARP_IS_RUNNING(ifp) && sc->sc_state == MASTER &&
1692                     !bcmp(dhost, IF_LLADDR(ifp), ETHER_ADDR_LEN))
1693                         return ifp;
1694         }
1695         return NULL;
1696 }
1697
1698 static void
1699 carp_master_down_timeout(void *xsc)
1700 {
1701         struct carp_softc *sc = xsc;
1702         struct netmsg_carp *cmsg = &sc->sc_md_msg;
1703
1704         KASSERT(mycpuid == 0, ("%s not on cpu0 but on cpu%d",
1705             __func__, mycpuid));
1706
1707         crit_enter();
1708         if (cmsg->base.lmsg.ms_flags & MSGF_DONE)
1709                 lwkt_sendmsg(netisr_cpuport(0), &cmsg->base.lmsg);
1710         crit_exit();
1711 }
1712
1713 static void
1714 carp_master_down_timeout_dispatch(netmsg_t msg)
1715 {
1716         struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
1717         struct carp_softc *sc = cmsg->nc_softc;
1718
1719         /* Reply ASAP */
1720         crit_enter();
1721         lwkt_replymsg(&cmsg->base.lmsg, 0);
1722         crit_exit();
1723
1724         CARP_DEBUG("%s: BACKUP -> MASTER (master timed out)\n",
1725                    sc->sc_if.if_xname);
1726         carp_master_down(sc);
1727 }
1728
1729 static void
1730 carp_master_down(struct carp_softc *sc)
1731 {
1732         switch (sc->sc_state) {
1733         case INIT:
1734                 kprintf("%s: master_down event in INIT state\n",
1735                         sc->sc_if.if_xname);
1736                 break;
1737
1738         case MASTER:
1739                 break;
1740
1741         case BACKUP:
1742                 carp_set_state(sc, MASTER);
1743                 carp_send_ad(sc);
1744                 carp_send_arp(sc);
1745 #ifdef INET6
1746                 carp_send_na(sc);
1747 #endif /* INET6 */
1748                 carp_setrun(sc, 0);
1749                 carp_setroute(sc, RTM_ADD);
1750                 break;
1751         }
1752 }
1753
1754 /*
1755  * When in backup state, af indicates whether to reset the master down timer
1756  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1757  */
1758 static void
1759 carp_setrun(struct carp_softc *sc, sa_family_t af)
1760 {
1761         struct ifnet *cifp = &sc->sc_if;
1762         struct timeval tv;
1763
1764         if (sc->sc_carpdev == NULL) {
1765                 carp_set_state(sc, INIT);
1766                 return;
1767         }
1768
1769         if ((cifp->if_flags & IFF_RUNNING) && sc->sc_vhid > 0 &&
1770             (sc->sc_naddrs || sc->sc_naddrs6)) {
1771                 /* Nothing */
1772         } else {
1773                 carp_setroute(sc, RTM_DELETE);
1774                 return;
1775         }
1776
1777         switch (sc->sc_state) {
1778         case INIT:
1779                 if (carp_opts[CARPCTL_PREEMPT] && !carp_suppress_preempt) {
1780                         carp_send_ad(sc);
1781                         carp_send_arp(sc);
1782 #ifdef INET6
1783                         carp_send_na(sc);
1784 #endif /* INET6 */
1785                         CARP_DEBUG("%s: INIT -> MASTER (preempting)\n",
1786                                    cifp->if_xname);
1787                         carp_set_state(sc, MASTER);
1788                         carp_setroute(sc, RTM_ADD);
1789                 } else {
1790                         CARP_DEBUG("%s: INIT -> BACKUP\n", cifp->if_xname);
1791                         carp_set_state(sc, BACKUP);
1792                         carp_setroute(sc, RTM_DELETE);
1793                         carp_setrun(sc, 0);
1794                 }
1795                 break;
1796
1797         case BACKUP:
1798                 callout_stop(&sc->sc_ad_tmo);
1799                 tv.tv_sec = 3 * sc->sc_advbase;
1800                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1801                 switch (af) {
1802 #ifdef INET
1803                 case AF_INET:
1804                         callout_reset(&sc->sc_md_tmo, tvtohz_high(&tv),
1805                             carp_master_down_timeout, sc);
1806                         break;
1807 #endif /* INET */
1808 #ifdef INET6
1809                 case AF_INET6:
1810                         callout_reset(&sc->sc_md6_tmo, tvtohz_high(&tv),
1811                             carp_master_down_timeout, sc);
1812                         break;
1813 #endif /* INET6 */
1814                 default:
1815                         if (sc->sc_naddrs)
1816                                 callout_reset(&sc->sc_md_tmo, tvtohz_high(&tv),
1817                                     carp_master_down_timeout, sc);
1818                         if (sc->sc_naddrs6)
1819                                 callout_reset(&sc->sc_md6_tmo, tvtohz_high(&tv),
1820                                     carp_master_down_timeout, sc);
1821                         break;
1822                 }
1823                 break;
1824
1825         case MASTER:
1826                 tv.tv_sec = sc->sc_advbase;
1827                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1828                 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
1829                     carp_send_ad_timeout, sc);
1830                 break;
1831         }
1832 }
1833
1834 static void
1835 carp_multicast_cleanup(struct carp_softc *sc)
1836 {
1837         struct ip_moptions *imo = &sc->sc_imo;
1838
1839         if (imo->imo_num_memberships == 0)
1840                 return;
1841         KKASSERT(imo->imo_num_memberships == 1);
1842
1843         in_delmulti(imo->imo_membership[0]);
1844         imo->imo_membership[0] = NULL;
1845         imo->imo_num_memberships = 0;
1846         imo->imo_multicast_ifp = NULL;
1847 }
1848
1849 #ifdef INET6
1850 static void
1851 carp_multicast6_cleanup(struct carp_softc *sc)
1852 {
1853         struct ip6_moptions *im6o = &sc->sc_im6o;
1854
1855         while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1856                 struct in6_multi_mship *imm =
1857                     LIST_FIRST(&im6o->im6o_memberships);
1858
1859                 LIST_REMOVE(imm, i6mm_chain);
1860                 in6_leavegroup(imm);
1861         }
1862         im6o->im6o_multicast_ifp = NULL;
1863 }
1864 #endif
1865
1866 static void
1867 carp_ioctl_getvhaddr_dispatch(netmsg_t msg)
1868 {
1869         struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
1870         struct carp_softc *sc = cmsg->nc_softc;
1871         const struct carp_vhaddr *vha;
1872         struct ifcarpvhaddr *carpa, *carpa0;
1873         int count, len, error = 0;
1874
1875         count = 0;
1876         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link)
1877                 ++count;
1878
1879         if (cmsg->nc_datalen == 0) {
1880                 cmsg->nc_datalen = count * sizeof(*carpa);
1881                 goto back;
1882         } else if (count == 0 || cmsg->nc_datalen < sizeof(*carpa)) {
1883                 cmsg->nc_datalen = 0;
1884                 goto back;
1885         }
1886         len = min(cmsg->nc_datalen, sizeof(*carpa) * count);
1887         KKASSERT(len >= sizeof(*carpa));
1888
1889         carpa0 = carpa = kmalloc(len, M_TEMP, M_WAITOK | M_NULLOK | M_ZERO);
1890         if (carpa == NULL) {
1891                 error = ENOMEM;
1892                 goto back;
1893         }
1894
1895         count = 0;
1896         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
1897                 if (len < sizeof(*carpa))
1898                         break;
1899
1900                 carpa->carpa_flags = vha->vha_flags;
1901                 carpa->carpa_addr.sin_family = AF_INET;
1902                 carpa->carpa_addr.sin_addr = vha->vha_ia->ia_addr.sin_addr;
1903
1904                 carpa->carpa_baddr.sin_family = AF_INET;
1905                 if (vha->vha_iaback == NULL) {
1906                         carpa->carpa_baddr.sin_addr.s_addr = INADDR_ANY;
1907                 } else {
1908                         carpa->carpa_baddr.sin_addr =
1909                         vha->vha_iaback->ia_addr.sin_addr;
1910                 }
1911
1912                 ++carpa;
1913                 ++count;
1914                 len -= sizeof(*carpa);
1915         }
1916         cmsg->nc_datalen = sizeof(*carpa) * count;
1917         KKASSERT(cmsg->nc_datalen > 0);
1918
1919         cmsg->nc_data = carpa0;
1920
1921 back:
1922         lwkt_replymsg(&cmsg->base.lmsg, error);
1923 }
1924
1925 static int
1926 carp_ioctl_getvhaddr(struct carp_softc *sc, struct ifdrv *ifd)
1927 {
1928         struct ifnet *ifp = &sc->arpcom.ac_if;
1929         struct netmsg_carp cmsg;
1930         int error;
1931
1932         ASSERT_IFNET_SERIALIZED_ALL(ifp);
1933         ifnet_deserialize_all(ifp);
1934
1935         bzero(&cmsg, sizeof(cmsg));
1936         netmsg_init(&cmsg.base, NULL, &curthread->td_msgport, 0,
1937             carp_ioctl_getvhaddr_dispatch);
1938         cmsg.nc_softc = sc;
1939         cmsg.nc_datalen = ifd->ifd_len;
1940
1941         error = lwkt_domsg(netisr_cpuport(0), &cmsg.base.lmsg, 0);
1942
1943         if (!error) {
1944                 if (cmsg.nc_data != NULL) {
1945                         error = copyout(cmsg.nc_data, ifd->ifd_data,
1946                             cmsg.nc_datalen);
1947                         kfree(cmsg.nc_data, M_TEMP);
1948                 }
1949                 ifd->ifd_len = cmsg.nc_datalen;
1950         } else {
1951                 KASSERT(cmsg.nc_data == NULL,
1952                     ("%s temp vhaddr is alloc upon error", __func__));
1953         }
1954
1955         ifnet_serialize_all(ifp);
1956         return error;
1957 }
1958
1959 static int
1960 carp_config_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha,
1961     struct in_ifaddr *ia_del)
1962 {
1963         struct ifnet *ifp;
1964         struct in_ifaddr *ia_if;
1965         const struct in_ifaddr *ia_vha;
1966         struct in_ifaddr_container *iac;
1967         int own, ia_match_carpdev;
1968
1969         KKASSERT(vha->vha_ia != NULL);
1970         ia_vha = vha->vha_ia;
1971
1972         ia_if = NULL;
1973         own = 0;
1974         ia_match_carpdev = 0;
1975         TAILQ_FOREACH(iac, &in_ifaddrheads[mycpuid], ia_link) {
1976                 struct in_ifaddr *ia = iac->ia;
1977
1978                 if (ia == ia_del)
1979                         continue;
1980
1981                 if (ia->ia_ifp->if_type == IFT_CARP)
1982                         continue;
1983
1984                 if ((ia->ia_ifp->if_flags & IFF_UP) == 0)
1985                         continue;
1986
1987                 /* and, yeah, we need a multicast-capable iface too */
1988                 if ((ia->ia_ifp->if_flags & IFF_MULTICAST) == 0)
1989                         continue;
1990
1991                 if (ia_vha->ia_subnetmask == ia->ia_subnetmask &&
1992                     ia_vha->ia_subnet == ia->ia_subnet) {
1993                         if (ia_vha->ia_addr.sin_addr.s_addr ==
1994                             ia->ia_addr.sin_addr.s_addr)
1995                                 own = 1;
1996                         if (ia_if == NULL) {
1997                                 ia_if = ia;
1998                         } else if (sc->sc_carpdev != NULL &&
1999                             sc->sc_carpdev == ia->ia_ifp) {
2000                                 ia_if = ia;
2001                                 if (ia_if->ia_flags & IFA_ROUTE) {
2002                                         /*
2003                                          * Address with prefix route
2004                                          * is prefered
2005                                          */
2006                                         break;
2007                                 }
2008                                 ia_match_carpdev = 1;
2009                         } else if (!ia_match_carpdev) {
2010                                 if (ia->ia_flags & IFA_ROUTE) {
2011                                         /*
2012                                          * Address with prefix route
2013                                          * is prefered over others.
2014                                          */
2015                                         ia_if = ia;
2016                                 }
2017                         }
2018                 }
2019         }
2020
2021         carp_deactivate_vhaddr(sc, vha, FALSE);
2022         if (!ia_if)
2023                 return ENOENT;
2024
2025         ifp = ia_if->ia_ifp;
2026
2027         /* XXX Don't allow parent iface to be changed */
2028         if (sc->sc_carpdev != NULL && sc->sc_carpdev != ifp)
2029                 return EEXIST;
2030
2031         return carp_activate_vhaddr(sc, vha, ifp, ia_if, own);
2032 }
2033
2034 static void
2035 carp_add_addr(struct carp_softc *sc, struct ifaddr *carp_ifa)
2036 {
2037         struct carp_vhaddr *vha_new;
2038         struct in_ifaddr *carp_ia;
2039 #ifdef INVARIANTS
2040         struct carp_vhaddr *vha;
2041 #endif
2042
2043         KKASSERT(carp_ifa->ifa_addr->sa_family == AF_INET);
2044         carp_ia = ifatoia(carp_ifa);
2045
2046 #ifdef INVARIANTS
2047         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link)
2048                 KKASSERT(vha->vha_ia != NULL && vha->vha_ia != carp_ia);
2049 #endif
2050
2051         vha_new = kmalloc(sizeof(*vha_new), M_CARP, M_WAITOK | M_ZERO);
2052         vha_new->vha_ia = carp_ia;
2053         carp_insert_vhaddr(sc, vha_new);
2054
2055         if (carp_config_vhaddr(sc, vha_new, NULL) != 0) {
2056                 /*
2057                  * If the above configuration fails, it may only mean
2058                  * that the new address is problematic.  However, the
2059                  * carp(4) interface may already have several working
2060                  * addresses.  Since the expected behaviour of
2061                  * SIOC[AS]IFADDR is to put the NIC into working state,
2062                  * we try starting the state machine manually here with
2063                  * the hope that the carp(4)'s previously working
2064                  * addresses still could be brought up.
2065                  */
2066                 carp_hmac_prepare(sc);
2067                 carp_set_state(sc, INIT);
2068                 carp_setrun(sc, 0);
2069         }
2070 }
2071
2072 static void
2073 carp_del_addr(struct carp_softc *sc, struct ifaddr *carp_ifa)
2074 {
2075         struct carp_vhaddr *vha;
2076         struct in_ifaddr *carp_ia;
2077
2078         KKASSERT(carp_ifa->ifa_addr->sa_family == AF_INET);
2079         carp_ia = ifatoia(carp_ifa);
2080
2081         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
2082                 KKASSERT(vha->vha_ia != NULL);
2083                 if (vha->vha_ia == carp_ia)
2084                         break;
2085         }
2086         KASSERT(vha != NULL, ("no corresponding vhaddr %p", carp_ifa));
2087
2088         /*
2089          * Remove the vhaddr from the list before deactivating
2090          * the vhaddr, so that the HMAC could be correctly
2091          * updated in carp_deactivate_vhaddr()
2092          */
2093         carp_remove_vhaddr(sc, vha);
2094
2095         carp_deactivate_vhaddr(sc, vha, FALSE);
2096         kfree(vha, M_CARP);
2097 }
2098
2099 static void
2100 carp_config_addr(struct carp_softc *sc, struct ifaddr *carp_ifa)
2101 {
2102         struct carp_vhaddr *vha;
2103         struct in_ifaddr *carp_ia;
2104
2105         KKASSERT(carp_ifa->ifa_addr->sa_family == AF_INET);
2106         carp_ia = ifatoia(carp_ifa);
2107
2108         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
2109                 KKASSERT(vha->vha_ia != NULL);
2110                 if (vha->vha_ia == carp_ia)
2111                         break;
2112         }
2113         KASSERT(vha != NULL, ("no corresponding vhaddr %p", carp_ifa));
2114
2115         /* Remove then reinsert, to keep the vhaddr list sorted */
2116         carp_remove_vhaddr(sc, vha);
2117         carp_insert_vhaddr(sc, vha);
2118
2119         if (carp_config_vhaddr(sc, vha, NULL) != 0) {
2120                 /* See the comment in carp_add_addr() */
2121                 carp_hmac_prepare(sc);
2122                 carp_set_state(sc, INIT);
2123                 carp_setrun(sc, 0);
2124         }
2125 }
2126
2127 #ifdef notyet
2128
2129 #ifdef INET6
2130 static int
2131 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
2132 {
2133         struct ifnet *ifp;
2134         struct carp_if *cif;
2135         struct in6_ifaddr *ia, *ia_if;
2136         struct ip6_moptions *im6o = &sc->sc_im6o;
2137         struct in6_multi_mship *imm;
2138         struct in6_addr in6;
2139         int own, error;
2140
2141         if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
2142                 carp_setrun(sc, 0);
2143                 return (0);
2144         }
2145
2146         /* we have to do it by hands to check we won't match on us */
2147         ia_if = NULL; own = 0;
2148         for (ia = in6_ifaddr; ia; ia = ia->ia_next) {
2149                 int i;
2150
2151                 for (i = 0; i < 4; i++) {
2152                         if ((sin6->sin6_addr.s6_addr32[i] &
2153                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]) !=
2154                             (ia->ia_addr.sin6_addr.s6_addr32[i] &
2155                             ia->ia_prefixmask.sin6_addr.s6_addr32[i]))
2156                                 break;
2157                 }
2158                 /* and, yeah, we need a multicast-capable iface too */
2159                 if (ia->ia_ifp != &sc->sc_if &&
2160                     (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
2161                     (i == 4)) {
2162                         if (!ia_if)
2163                                 ia_if = ia;
2164                         if (IN6_ARE_ADDR_EQUAL(&sin6->sin6_addr,
2165                             &ia->ia_addr.sin6_addr))
2166                                 own++;
2167                 }
2168         }
2169
2170         if (!ia_if)
2171                 return (EADDRNOTAVAIL);
2172         ia = ia_if;
2173         ifp = ia->ia_ifp;
2174
2175         if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
2176             (im6o->im6o_multicast_ifp && im6o->im6o_multicast_ifp != ifp))
2177                 return (EADDRNOTAVAIL);
2178
2179         if (!sc->sc_naddrs6) {
2180                 im6o->im6o_multicast_ifp = ifp;
2181
2182                 /* join CARP multicast address */
2183                 bzero(&in6, sizeof(in6));
2184                 in6.s6_addr16[0] = htons(0xff02);
2185                 in6.s6_addr8[15] = 0x12;
2186                 if (in6_setscope(&in6, ifp, NULL) != 0)
2187                         goto cleanup;
2188                 if ((imm = in6_joingroup(ifp, &in6, &error)) == NULL)
2189                         goto cleanup;
2190                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
2191
2192                 /* join solicited multicast address */
2193                 bzero(&in6, sizeof(in6));
2194                 in6.s6_addr16[0] = htons(0xff02);
2195                 in6.s6_addr32[1] = 0;
2196                 in6.s6_addr32[2] = htonl(1);
2197                 in6.s6_addr32[3] = sin6->sin6_addr.s6_addr32[3];
2198                 in6.s6_addr8[12] = 0xff;
2199                 if (in6_setscope(&in6, ifp, NULL) != 0)
2200                         goto cleanup;
2201                 if ((imm = in6_joingroup(ifp, &in6, &error)) == NULL)
2202                         goto cleanup;
2203                 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
2204         }
2205
2206 #ifdef foo
2207         if (!ifp->if_carp) {
2208                 cif = kmalloc(sizeof(*cif), M_CARP, M_WAITOK | M_ZERO);
2209
2210                 if ((error = ifpromisc(ifp, 1))) {
2211                         kfree(cif, M_CARP);
2212                         goto cleanup;
2213                 }
2214
2215                 TAILQ_INIT(&cif->vhif_vrs);
2216                 ifp->if_carp = cif;
2217         } else {
2218                 struct carp_softc *vr;
2219
2220                 cif = ifp->if_carp;
2221                 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
2222                         if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
2223                                 error = EINVAL;
2224                                 goto cleanup;
2225                         }
2226                 }
2227         }
2228 #endif
2229         sc->sc_ia6 = ia;
2230         sc->sc_carpdev = ifp;
2231
2232 #ifdef foo
2233         { /* XXX prevent endless loop if already in queue */
2234         struct carp_softc *vr, *after = NULL;
2235         int myself = 0;
2236         cif = ifp->if_carp;
2237
2238         TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
2239                 if (vr == sc)
2240                         myself = 1;
2241                 if (vr->sc_vhid < sc->sc_vhid)
2242                         after = vr;
2243         }
2244
2245         if (!myself) {
2246                 /* We're trying to keep things in order */
2247                 if (after == NULL)
2248                         TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
2249                 else
2250                         TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
2251         }
2252         }
2253 #endif
2254
2255         sc->sc_naddrs6++;
2256         if (own)
2257                 sc->sc_advskew = 0;
2258         carp_sc_state(sc);
2259         carp_setrun(sc, 0);
2260
2261         return (0);
2262
2263 cleanup:
2264         /* clean up multicast memberships */
2265         if (!sc->sc_naddrs6) {
2266                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
2267                         imm = LIST_FIRST(&im6o->im6o_memberships);
2268                         LIST_REMOVE(imm, i6mm_chain);
2269                         in6_leavegroup(imm);
2270                 }
2271         }
2272         return (error);
2273 }
2274
2275 static int
2276 carp_del_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
2277 {
2278         int error = 0;
2279
2280         if (!--sc->sc_naddrs6) {
2281                 struct carp_if *cif = sc->sc_carpdev->if_carp;
2282                 struct ip6_moptions *im6o = &sc->sc_im6o;
2283
2284                 callout_stop(&sc->sc_ad_tmo);
2285                 sc->sc_vhid = -1;
2286                 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
2287                         struct in6_multi_mship *imm =
2288                             LIST_FIRST(&im6o->im6o_memberships);
2289
2290                         LIST_REMOVE(imm, i6mm_chain);
2291                         in6_leavegroup(imm);
2292                 }
2293                 im6o->im6o_multicast_ifp = NULL;
2294 #ifdef foo
2295                 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
2296                 if (TAILQ_EMPTY(&cif->vhif_vrs)) {
2297                         sc->sc_carpdev->if_carp = NULL;
2298                         kfree(cif, M_IFADDR);
2299                 }
2300 #endif
2301         }
2302         return (error);
2303 }
2304 #endif /* INET6 */
2305
2306 #endif
2307
2308 static int
2309 carp_ioctl(struct ifnet *ifp, u_long cmd, caddr_t addr, struct ucred *cr)
2310 {
2311         struct carp_softc *sc = ifp->if_softc;
2312         struct ifreq *ifr = (struct ifreq *)addr;
2313         struct ifdrv *ifd = (struct ifdrv *)addr;
2314         int error = 0;
2315
2316         ASSERT_IFNET_SERIALIZED_ALL(ifp);
2317
2318         switch (cmd) {
2319         case SIOCSIFFLAGS:
2320                 if (ifp->if_flags & IFF_UP) {
2321                         if ((ifp->if_flags & IFF_RUNNING) == 0)
2322                                 carp_init(sc);
2323                 } else if (ifp->if_flags & IFF_RUNNING) {
2324                         carp_ioctl_stop(sc);
2325                 }
2326                 break;
2327
2328         case SIOCSIFCAP:
2329                 carp_ioctl_ifcap(sc, ifr->ifr_reqcap);
2330                 break;
2331
2332         case SIOCSVH:
2333                 error = carp_ioctl_setvh(sc, ifr->ifr_data, cr);
2334                 break;
2335
2336         case SIOCGVH:
2337                 error = carp_ioctl_getvh(sc, ifr->ifr_data, cr);
2338                 break;
2339
2340         case SIOCGDRVSPEC:
2341                 switch (ifd->ifd_cmd) {
2342                 case CARPGDEVNAME:
2343                         error = carp_ioctl_getdevname(sc, ifd);
2344                         break;
2345
2346                 case CARPGVHADDR:
2347                         error = carp_ioctl_getvhaddr(sc, ifd);
2348                         break;
2349
2350                 default:
2351                         error = EINVAL;
2352                         break;
2353                 }
2354                 break;
2355
2356         default:
2357                 error = ether_ioctl(ifp, cmd, addr);
2358                 break;
2359         }
2360
2361         return error;
2362 }
2363
2364 static void
2365 carp_ioctl_stop_dispatch(netmsg_t msg)
2366 {
2367         struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
2368         struct carp_softc *sc = cmsg->nc_softc;
2369
2370         carp_stop(sc, FALSE);
2371         lwkt_replymsg(&cmsg->base.lmsg, 0);
2372 }
2373
2374 static void
2375 carp_ioctl_stop(struct carp_softc *sc)
2376 {
2377         struct ifnet *ifp = &sc->arpcom.ac_if;
2378         struct netmsg_carp cmsg;
2379
2380         ASSERT_IFNET_SERIALIZED_ALL(ifp);
2381
2382         ifnet_deserialize_all(ifp);
2383
2384         bzero(&cmsg, sizeof(cmsg));
2385         netmsg_init(&cmsg.base, NULL, &curthread->td_msgport, 0,
2386             carp_ioctl_stop_dispatch);
2387         cmsg.nc_softc = sc;
2388
2389         lwkt_domsg(netisr_cpuport(0), &cmsg.base.lmsg, 0);
2390
2391         ifnet_serialize_all(ifp);
2392 }
2393
2394 static void
2395 carp_ioctl_setvh_dispatch(netmsg_t msg)
2396 {
2397         struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
2398         struct carp_softc *sc = cmsg->nc_softc;
2399         struct ifnet *ifp = &sc->arpcom.ac_if;
2400         const struct carpreq *carpr = cmsg->nc_data;
2401         int error;
2402
2403         error = 1;
2404         if ((ifp->if_flags & IFF_RUNNING) &&
2405             sc->sc_state != INIT && carpr->carpr_state != sc->sc_state) {
2406                 switch (carpr->carpr_state) {
2407                 case BACKUP:
2408                         callout_stop(&sc->sc_ad_tmo);
2409                         carp_set_state(sc, BACKUP);
2410                         carp_setrun(sc, 0);
2411                         carp_setroute(sc, RTM_DELETE);
2412                         break;
2413
2414                 case MASTER:
2415                         carp_master_down(sc);
2416                         break;
2417
2418                 default:
2419                         break;
2420                 }
2421         }
2422         if (carpr->carpr_vhid > 0) {
2423                 if (carpr->carpr_vhid > 255) {
2424                         error = EINVAL;
2425                         goto back;
2426                 }
2427                 if (sc->sc_carpdev) {
2428                         struct carp_if *cif = sc->sc_carpdev->if_carp;
2429                         struct carp_softc_container *scc;
2430
2431                         TAILQ_FOREACH(scc, cif, scc_link) {
2432                                 struct carp_softc *vr = scc->scc_softc;
2433
2434                                 if (vr != sc &&
2435                                     vr->sc_vhid == carpr->carpr_vhid) {
2436                                         error = EEXIST;
2437                                         goto back;
2438                                 }
2439                         }
2440                 }
2441                 sc->sc_vhid = carpr->carpr_vhid;
2442
2443                 IF_LLADDR(ifp)[5] = sc->sc_vhid;
2444                 bcopy(IF_LLADDR(ifp), sc->arpcom.ac_enaddr,
2445                     ETHER_ADDR_LEN);
2446
2447                 error--;
2448         }
2449         if (carpr->carpr_advbase > 0 || carpr->carpr_advskew > 0) {
2450                 if (carpr->carpr_advskew >= 255) {
2451                         error = EINVAL;
2452                         goto back;
2453                 }
2454                 if (carpr->carpr_advbase > 255) {
2455                         error = EINVAL;
2456                         goto back;
2457                 }
2458                 sc->sc_advbase = carpr->carpr_advbase;
2459                 sc->sc_advskew = carpr->carpr_advskew;
2460                 error--;
2461         }
2462         bcopy(carpr->carpr_key, sc->sc_key, sizeof(sc->sc_key));
2463         if (error > 0) {
2464                 error = EINVAL;
2465         } else {
2466                 error = 0;
2467                 carp_setrun(sc, 0);
2468         }
2469 back:
2470         carp_hmac_prepare(sc);
2471
2472         lwkt_replymsg(&cmsg->base.lmsg, error);
2473 }
2474
2475 static int
2476 carp_ioctl_setvh(struct carp_softc *sc, void *udata, struct ucred *cr)
2477 {
2478         struct ifnet *ifp = &sc->arpcom.ac_if;
2479         struct netmsg_carp cmsg;
2480         struct carpreq carpr;
2481         int error;
2482
2483         ASSERT_IFNET_SERIALIZED_ALL(ifp);
2484         ifnet_deserialize_all(ifp);
2485
2486         error = priv_check_cred(cr, PRIV_ROOT, NULL_CRED_OKAY);
2487         if (error)
2488                 goto back;
2489
2490         error = copyin(udata, &carpr, sizeof(carpr));
2491         if (error)
2492                 goto back;
2493
2494         bzero(&cmsg, sizeof(cmsg));
2495         netmsg_init(&cmsg.base, NULL, &curthread->td_msgport, 0,
2496             carp_ioctl_setvh_dispatch);
2497         cmsg.nc_softc = sc;
2498         cmsg.nc_data = &carpr;
2499
2500         error = lwkt_domsg(netisr_cpuport(0), &cmsg.base.lmsg, 0);
2501
2502 back:
2503         ifnet_serialize_all(ifp);
2504         return error;
2505 }
2506
2507 static void
2508 carp_ioctl_ifcap_dispatch(netmsg_t msg)
2509 {
2510         struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
2511         struct carp_softc *sc = cmsg->nc_softc;
2512         struct ifnet *ifp = &sc->arpcom.ac_if;
2513         int reqcap = *((const int *)(cmsg->nc_data));
2514         int mask;
2515
2516         mask = reqcap ^ ifp->if_capenable;
2517         if (mask & IFCAP_TXCSUM) {
2518                 ifp->if_capenable ^= IFCAP_TXCSUM;
2519                 if ((ifp->if_capenable & IFCAP_TXCSUM) &&
2520                     sc->sc_carpdev != NULL) {
2521                         ifp->if_hwassist |=
2522                             (sc->sc_carpdev->if_hwassist &
2523                              (CSUM_IP | CSUM_UDP | CSUM_TCP));
2524                 } else {
2525                         ifp->if_hwassist &= ~(CSUM_IP | CSUM_UDP | CSUM_TCP);
2526                 }
2527         }
2528         if (mask & IFCAP_TSO) {
2529                 ifp->if_capenable ^= IFCAP_TSO;
2530                 if ((ifp->if_capenable & IFCAP_TSO) &&
2531                     sc->sc_carpdev != NULL) {
2532                         ifp->if_hwassist |=
2533                             (sc->sc_carpdev->if_hwassist & CSUM_TSO);
2534                 } else {
2535                         ifp->if_hwassist &= ~CSUM_TSO;
2536                 }
2537         }
2538
2539         lwkt_replymsg(&cmsg->base.lmsg, 0);
2540 }
2541
2542 static void
2543 carp_ioctl_ifcap(struct carp_softc *sc, int reqcap)
2544 {
2545         struct ifnet *ifp = &sc->arpcom.ac_if;
2546         struct netmsg_carp cmsg;
2547
2548         ASSERT_IFNET_SERIALIZED_ALL(ifp);
2549         ifnet_deserialize_all(ifp);
2550
2551         bzero(&cmsg, sizeof(cmsg));
2552         netmsg_init(&cmsg.base, NULL, &curthread->td_msgport, 0,
2553             carp_ioctl_ifcap_dispatch);
2554         cmsg.nc_softc = sc;
2555         cmsg.nc_data = &reqcap;
2556
2557         lwkt_domsg(netisr_cpuport(0), &cmsg.base.lmsg, 0);
2558
2559         ifnet_serialize_all(ifp);
2560 }
2561
2562 static void
2563 carp_ioctl_getvh_dispatch(netmsg_t msg)
2564 {
2565         struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
2566         struct carp_softc *sc = cmsg->nc_softc;
2567         struct carpreq *carpr = cmsg->nc_data;
2568
2569         carpr->carpr_state = sc->sc_state;
2570         carpr->carpr_vhid = sc->sc_vhid;
2571         carpr->carpr_advbase = sc->sc_advbase;
2572         carpr->carpr_advskew = sc->sc_advskew;
2573         bcopy(sc->sc_key, carpr->carpr_key, sizeof(carpr->carpr_key));
2574
2575         lwkt_replymsg(&cmsg->base.lmsg, 0);
2576 }
2577
2578 static int
2579 carp_ioctl_getvh(struct carp_softc *sc, void *udata, struct ucred *cr)
2580 {
2581         struct ifnet *ifp = &sc->arpcom.ac_if;
2582         struct netmsg_carp cmsg;
2583         struct carpreq carpr;
2584         int error;
2585
2586         ASSERT_IFNET_SERIALIZED_ALL(ifp);
2587         ifnet_deserialize_all(ifp);
2588
2589         bzero(&cmsg, sizeof(cmsg));
2590         netmsg_init(&cmsg.base, NULL, &curthread->td_msgport, 0,
2591             carp_ioctl_getvh_dispatch);
2592         cmsg.nc_softc = sc;
2593         cmsg.nc_data = &carpr;
2594
2595         lwkt_domsg(netisr_cpuport(0), &cmsg.base.lmsg, 0);
2596
2597         error = priv_check_cred(cr, PRIV_ROOT, NULL_CRED_OKAY);
2598         if (error)
2599                 bzero(carpr.carpr_key, sizeof(carpr.carpr_key));
2600
2601         error = copyout(&carpr, udata, sizeof(carpr));
2602
2603         ifnet_serialize_all(ifp);
2604         return error;
2605 }
2606
2607 static void
2608 carp_ioctl_getdevname_dispatch(netmsg_t msg)
2609 {
2610         struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
2611         struct carp_softc *sc = cmsg->nc_softc;
2612         char *devname = cmsg->nc_data;
2613
2614         bzero(devname, IFNAMSIZ);
2615         if (sc->sc_carpdev != NULL)
2616                 strlcpy(devname, sc->sc_carpdev->if_xname, IFNAMSIZ);
2617
2618         lwkt_replymsg(&cmsg->base.lmsg, 0);
2619 }
2620
2621 static int
2622 carp_ioctl_getdevname(struct carp_softc *sc, struct ifdrv *ifd)
2623 {
2624         struct ifnet *ifp = &sc->arpcom.ac_if;
2625         struct netmsg_carp cmsg;
2626         char devname[IFNAMSIZ];
2627         int error;
2628
2629         ASSERT_IFNET_SERIALIZED_ALL(ifp);
2630
2631         if (ifd->ifd_len != sizeof(devname))
2632                 return EINVAL;
2633
2634         ifnet_deserialize_all(ifp);
2635
2636         bzero(&cmsg, sizeof(cmsg));
2637         netmsg_init(&cmsg.base, NULL, &curthread->td_msgport, 0,
2638             carp_ioctl_getdevname_dispatch);
2639         cmsg.nc_softc = sc;
2640         cmsg.nc_data = devname;
2641
2642         lwkt_domsg(netisr_cpuport(0), &cmsg.base.lmsg, 0);
2643
2644         error = copyout(devname, ifd->ifd_data, sizeof(devname));
2645
2646         ifnet_serialize_all(ifp);
2647         return error;
2648 }
2649
2650 static void
2651 carp_init_dispatch(netmsg_t msg)
2652 {
2653         struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
2654         struct carp_softc *sc = cmsg->nc_softc;
2655
2656         sc->sc_if.if_flags |= IFF_RUNNING;
2657         carp_hmac_prepare(sc);
2658         carp_set_state(sc, INIT);
2659         carp_setrun(sc, 0);
2660
2661         lwkt_replymsg(&cmsg->base.lmsg, 0);
2662 }
2663
2664 static void
2665 carp_init(void *xsc)
2666 {
2667         struct carp_softc *sc = xsc;
2668         struct ifnet *ifp = &sc->arpcom.ac_if;
2669         struct netmsg_carp cmsg;
2670
2671         ASSERT_IFNET_SERIALIZED_ALL(ifp);
2672
2673         ifnet_deserialize_all(ifp);
2674
2675         bzero(&cmsg, sizeof(cmsg));
2676         netmsg_init(&cmsg.base, NULL, &curthread->td_msgport, 0,
2677             carp_init_dispatch);
2678         cmsg.nc_softc = sc;
2679
2680         lwkt_domsg(netisr_cpuport(0), &cmsg.base.lmsg, 0);
2681
2682         ifnet_serialize_all(ifp);
2683 }
2684
2685 static int
2686 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *dst,
2687     struct rtentry *rt)
2688 {
2689         struct carp_softc *sc = ifp->if_softc;
2690         struct ifnet *carpdev;
2691         int error = 0;
2692
2693         carpdev = sc->sc_carpdev;
2694         if (carpdev != NULL) {
2695                 if (m->m_flags & M_MCAST)
2696                         IFNET_STAT_INC(ifp, omcasts, 1);
2697                 IFNET_STAT_INC(ifp, obytes, m->m_pkthdr.len + ETHER_HDR_LEN);
2698                 IFNET_STAT_INC(ifp, opackets, 1);
2699
2700                 /*
2701                  * NOTE:
2702                  * CARP's ifp is passed to backing device's
2703                  * if_output method.
2704                  */
2705                 carpdev->if_output(ifp, m, dst, rt);
2706         } else {
2707                 IFNET_STAT_INC(ifp, oerrors, 1);
2708                 m_freem(m);
2709                 error = ENETUNREACH;
2710         }
2711         return error;
2712 }
2713
2714 /*
2715  * Start output on carp interface. This function should never be called.
2716  */
2717 static void
2718 carp_start(struct ifnet *ifp, struct ifaltq_subque *ifsq __unused)
2719 {
2720         panic("%s: start called", ifp->if_xname);
2721 }
2722
2723 static void
2724 carp_set_state(struct carp_softc *sc, int state)
2725 {
2726         struct ifnet *cifp = &sc->sc_if;
2727
2728         if (sc->sc_state == state)
2729                 return;
2730         sc->sc_state = state;
2731
2732         switch (sc->sc_state) {
2733         case BACKUP:
2734                 cifp->if_link_state = LINK_STATE_DOWN;
2735                 break;
2736
2737         case MASTER:
2738                 cifp->if_link_state = LINK_STATE_UP;
2739                 break;
2740
2741         default:
2742                 cifp->if_link_state = LINK_STATE_UNKNOWN;
2743                 break;
2744         }
2745         rt_ifmsg(cifp);
2746 }
2747
2748 void
2749 carp_group_demote_adj(struct ifnet *ifp, int adj)
2750 {
2751         struct ifg_list *ifgl;
2752         int *dm;
2753
2754         TAILQ_FOREACH(ifgl, &ifp->if_groups, ifgl_next) {
2755                 if (!strcmp(ifgl->ifgl_group->ifg_group, IFG_ALL))
2756                         continue;
2757                 dm = &ifgl->ifgl_group->ifg_carp_demoted;
2758
2759                 if (*dm + adj >= 0)
2760                         *dm += adj;
2761                 else
2762                         *dm = 0;
2763
2764                 if (adj > 0 && *dm == 1)
2765                         carp_send_ad_all();
2766                 CARP_LOG("%s demoted group %s to %d", ifp->if_xname,
2767                     ifgl->ifgl_group->ifg_group, *dm);
2768         }
2769 }
2770
2771 #ifdef foo
2772 void
2773 carp_carpdev_state(void *v)
2774 {
2775         struct carp_if *cif = v;
2776         struct carp_softc *sc;
2777
2778         TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list)
2779                 carp_sc_state(sc);
2780 }
2781
2782 static void
2783 carp_sc_state(struct carp_softc *sc)
2784 {
2785         if (!(sc->sc_carpdev->if_flags & IFF_UP)) {
2786                 callout_stop(&sc->sc_ad_tmo);
2787                 callout_stop(&sc->sc_md_tmo);
2788                 callout_stop(&sc->sc_md6_tmo);
2789                 carp_set_state(sc, INIT);
2790                 carp_setrun(sc, 0);
2791                 if (!sc->sc_suppress) {
2792                         carp_suppress_preempt++;
2793                         if (carp_suppress_preempt == 1)
2794                                 carp_send_ad_all();
2795                 }
2796                 sc->sc_suppress = 1;
2797         } else {
2798                 carp_set_state(sc, INIT);
2799                 carp_setrun(sc, 0);
2800                 if (sc->sc_suppress)
2801                         carp_suppress_preempt--;
2802                 sc->sc_suppress = 0;
2803         }
2804 }
2805 #endif
2806
2807 static void
2808 carp_stop(struct carp_softc *sc, boolean_t detach)
2809 {
2810         sc->sc_if.if_flags &= ~IFF_RUNNING;
2811
2812         callout_stop(&sc->sc_ad_tmo);
2813         callout_stop(&sc->sc_md_tmo);
2814         callout_stop(&sc->sc_md6_tmo);
2815
2816         if (!detach && sc->sc_state == MASTER)
2817                 carp_send_ad(sc);
2818
2819         if (sc->sc_suppress)
2820                 carp_suppress_preempt--;
2821         sc->sc_suppress = 0;
2822
2823         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS)
2824                 carp_suppress_preempt--;
2825         sc->sc_sendad_errors = 0;
2826         sc->sc_sendad_success = 0;
2827
2828         carp_set_state(sc, INIT);
2829         carp_setrun(sc, 0);
2830 }
2831
2832 static void
2833 carp_suspend(struct carp_softc *sc, boolean_t detach)
2834 {
2835         struct ifnet *cifp = &sc->sc_if;
2836
2837         carp_stop(sc, detach);
2838
2839         /* Retain the running state, if we are not dead yet */
2840         if (!sc->sc_dead && (cifp->if_flags & IFF_UP))
2841                 cifp->if_flags |= IFF_RUNNING;
2842 }
2843
2844 static int
2845 carp_activate_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha,
2846     struct ifnet *ifp, struct in_ifaddr *ia_if, int own)
2847 {
2848         struct ip_moptions *imo = &sc->sc_imo;
2849         struct carp_if *ocif = ifp->if_carp;
2850         int error;
2851
2852         KKASSERT(vha->vha_ia != NULL);
2853
2854         KASSERT(ia_if != NULL, ("NULL backing address"));
2855         KASSERT(vha->vha_iaback == NULL, ("%p is already activated", vha));
2856         KASSERT((vha->vha_flags & CARP_VHAF_OWNER) == 0,
2857                 ("inactive vhaddr %p is the address owner", vha));
2858
2859         KASSERT(sc->sc_carpdev == NULL || sc->sc_carpdev == ifp,
2860                 ("%s is already on %s", sc->sc_if.if_xname,
2861                  sc->sc_carpdev->if_xname));
2862
2863         if (ocif == NULL) {
2864                 KASSERT(sc->sc_carpdev == NULL,
2865                         ("%s is already on %s", sc->sc_if.if_xname,
2866                          sc->sc_carpdev->if_xname));
2867
2868                 error = ifpromisc(ifp, 1);
2869                 if (error)
2870                         return error;
2871         } else {
2872                 struct carp_softc_container *scc;
2873
2874                 TAILQ_FOREACH(scc, ocif, scc_link) {
2875                         struct carp_softc *vr = scc->scc_softc;
2876
2877                         if (vr != sc && vr->sc_vhid == sc->sc_vhid)
2878                                 return EINVAL;
2879                 }
2880         }
2881
2882         ifp->if_carp = carp_if_insert(ocif, sc);
2883         KASSERT(ifp->if_carp != NULL, ("%s carp_if_insert failed", __func__));
2884
2885         sc->sc_ia = ia_if;
2886         sc->sc_carpdev = ifp;
2887         sc->arpcom.ac_if.if_hwassist = 0;
2888         if (sc->arpcom.ac_if.if_capenable & IFCAP_TXCSUM) {
2889                 sc->arpcom.ac_if.if_hwassist |=
2890                     (ifp->if_hwassist & (CSUM_IP | CSUM_UDP | CSUM_TCP));
2891         }
2892         if (sc->arpcom.ac_if.if_capenable & IFCAP_TSO)
2893                 sc->arpcom.ac_if.if_hwassist |= (ifp->if_hwassist & CSUM_TSO);
2894
2895         /*
2896          * Make sure that all protocol threads see the sc_carpdev and
2897          * if_carp changes
2898          */
2899         netmsg_service_sync();
2900
2901         if (ocif != NULL && ifp->if_carp != ocif) {
2902                 /*
2903                  * The old carp list could be safely free now,
2904                  * since no one can access it.
2905                  */
2906                 carp_if_free(ocif);
2907         }
2908
2909         vha->vha_iaback = ia_if;
2910         sc->sc_naddrs++;
2911
2912         if (own) {
2913                 vha->vha_flags |= CARP_VHAF_OWNER;
2914
2915                 /* XXX save user configured advskew? */
2916                 sc->sc_advskew = 0;
2917         }
2918
2919         carp_addroute_vhaddr(sc, vha);
2920
2921         /*
2922          * Join the multicast group only after the backing interface
2923          * has been hooked with the CARP interface.
2924          */
2925         KASSERT(imo->imo_multicast_ifp == NULL ||
2926                 imo->imo_multicast_ifp == &sc->sc_if,
2927                 ("%s didn't leave mcast group on %s",
2928                  sc->sc_if.if_xname, imo->imo_multicast_ifp->if_xname));
2929
2930         if (imo->imo_num_memberships == 0) {
2931                 struct in_addr addr;
2932
2933                 addr.s_addr = htonl(INADDR_CARP_GROUP);
2934                 imo->imo_membership[0] = in_addmulti(&addr, &sc->sc_if);
2935                 if (imo->imo_membership[0] == NULL) {
2936                         carp_deactivate_vhaddr(sc, vha, FALSE);
2937                         return ENOBUFS;
2938                 }
2939
2940                 imo->imo_num_memberships++;
2941                 imo->imo_multicast_ifp = &sc->sc_if;
2942                 imo->imo_multicast_ttl = CARP_DFLTTL;
2943                 imo->imo_multicast_loop = 0;
2944         }
2945
2946         carp_hmac_prepare(sc);
2947         carp_set_state(sc, INIT);
2948         carp_setrun(sc, 0);
2949         return 0;
2950 }
2951
2952 static void
2953 carp_deactivate_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha,
2954     boolean_t del_iaback)
2955 {
2956         KKASSERT(vha->vha_ia != NULL);
2957
2958         carp_hmac_prepare(sc);
2959
2960         if (vha->vha_iaback == NULL) {
2961                 KASSERT((vha->vha_flags & CARP_VHAF_OWNER) == 0,
2962                         ("inactive vhaddr %p is the address owner", vha));
2963                 return;
2964         }
2965
2966         vha->vha_flags &= ~CARP_VHAF_OWNER;
2967         carp_delroute_vhaddr(sc, vha, del_iaback);
2968
2969         KKASSERT(sc->sc_naddrs > 0);
2970         vha->vha_iaback = NULL;
2971         sc->sc_naddrs--;
2972         if (!sc->sc_naddrs) {
2973                 if (sc->sc_naddrs6) {
2974                         carp_multicast_cleanup(sc);
2975                         sc->sc_ia = NULL;
2976                 } else {
2977                         carp_detach(sc, FALSE, del_iaback);
2978                 }
2979         }
2980 }
2981
2982 static void
2983 carp_link_addrs(struct carp_softc *sc, struct ifnet *ifp, struct ifaddr *ifa_if)
2984 {
2985         struct carp_vhaddr *vha;
2986         struct in_ifaddr *ia_if;
2987
2988         KKASSERT(ifa_if->ifa_addr->sa_family == AF_INET);
2989         ia_if = ifatoia(ifa_if);
2990
2991         /*
2992          * Test each inactive vhaddr against the newly added address.
2993          * If the newly added address could be the backing address,
2994          * then activate the matching vhaddr.
2995          */
2996         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
2997                 const struct in_ifaddr *ia;
2998                 int own;
2999
3000                 if (vha->vha_iaback != NULL)
3001                         continue;
3002
3003                 ia = vha->vha_ia;
3004                 if (ia->ia_subnetmask != ia_if->ia_subnetmask ||
3005                     ia->ia_subnet != ia_if->ia_subnet)
3006                         continue;
3007
3008                 own = 0;
3009                 if (ia->ia_addr.sin_addr.s_addr ==
3010                     ia_if->ia_addr.sin_addr.s_addr)
3011                         own = 1;
3012
3013                 carp_activate_vhaddr(sc, vha, ifp, ia_if, own);
3014         }
3015 }
3016
3017 static void
3018 carp_unlink_addrs(struct carp_softc *sc, struct ifnet *ifp,
3019                   struct ifaddr *ifa_if)
3020 {
3021         struct carp_vhaddr *vha;
3022         struct in_ifaddr *ia_if;
3023
3024         KKASSERT(ifa_if->ifa_addr->sa_family == AF_INET);
3025         ia_if = ifatoia(ifa_if);
3026
3027         /*
3028          * Ad src address is deleted; set it to NULL.
3029          * Following loop will try pick up a new ad src address
3030          * if one of the vhaddr could retain its backing address.
3031          */
3032         if (sc->sc_ia == ia_if)
3033                 sc->sc_ia = NULL;
3034
3035         /*
3036          * Test each active vhaddr against the deleted address.
3037          * If the deleted address is vhaddr address's backing
3038          * address, then deactivate the vhaddr.
3039          */
3040         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
3041                 if (vha->vha_iaback == NULL)
3042                         continue;
3043
3044                 if (vha->vha_iaback == ia_if)
3045                         carp_deactivate_vhaddr(sc, vha, TRUE);
3046                 else if (sc->sc_ia == NULL)
3047                         sc->sc_ia = vha->vha_iaback;
3048         }
3049 }
3050
3051 static void
3052 carp_update_addrs(struct carp_softc *sc, struct ifaddr *ifa_del)
3053 {
3054         struct carp_vhaddr *vha;
3055
3056         KKASSERT(sc->sc_carpdev == NULL);
3057
3058         TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link)
3059                 carp_config_vhaddr(sc, vha, ifatoia(ifa_del));
3060 }
3061
3062 static void
3063 carp_ifaddr(void *arg __unused, struct ifnet *ifp,
3064             enum ifaddr_event event, struct ifaddr *ifa)
3065 {
3066         struct carp_softc *sc;
3067
3068         if (ifa->ifa_addr->sa_family != AF_INET)
3069                 return;
3070
3071         KASSERT(&curthread->td_msgport == netisr_cpuport(0),
3072             ("not in netisr0"));
3073
3074         if (ifp->if_type == IFT_CARP) {
3075                 /*
3076                  * Address is changed on carp(4) interface
3077                  */
3078                 switch (event) {
3079                 case IFADDR_EVENT_ADD:
3080                         carp_add_addr(ifp->if_softc, ifa);
3081                         break;
3082
3083                 case IFADDR_EVENT_CHANGE:
3084                         carp_config_addr(ifp->if_softc, ifa);
3085                         break;
3086
3087                 case IFADDR_EVENT_DELETE:
3088                         carp_del_addr(ifp->if_softc, ifa);
3089                         break;
3090                 }
3091                 return;
3092         }
3093
3094         /*
3095          * Address is changed on non-carp(4) interface
3096          */
3097         if ((ifp->if_flags & IFF_MULTICAST) == 0)
3098                 return;
3099
3100         LIST_FOREACH(sc, &carpif_list, sc_next) {
3101                 if (sc->sc_carpdev != NULL && sc->sc_carpdev != ifp) {
3102                         /* Not the parent iface; skip */
3103                         continue;
3104                 }
3105
3106                 switch (event) {
3107                 case IFADDR_EVENT_ADD:
3108                         carp_link_addrs(sc, ifp, ifa);
3109                         break;
3110
3111                 case IFADDR_EVENT_DELETE:
3112                         if (sc->sc_carpdev != NULL) {
3113                                 carp_unlink_addrs(sc, ifp, ifa);
3114                                 if (sc->sc_carpdev == NULL) {
3115                                         /*
3116                                          * We no longer have the parent
3117                                          * interface, however, certain
3118                                          * virtual addresses, which are
3119                                          * not used because they can't
3120                                          * match the previous parent
3121                                          * interface's addresses, may now
3122                                          * match different interface's
3123                                          * addresses.
3124                                          */
3125                                         carp_update_addrs(sc, ifa);
3126                                 }
3127                         } else {
3128                                 /*
3129                                  * The carp(4) interface didn't have a
3130                                  * parent iface, so it is not possible
3131                                  * that it will contain any address to
3132                                  * be unlinked.
3133                                  */
3134                         }
3135                         break;
3136
3137                 case IFADDR_EVENT_CHANGE:
3138                         if (sc->sc_carpdev == NULL) {
3139                                 /*
3140                                  * The carp(4) interface didn't have a
3141                                  * parent iface, so it is not possible
3142                                  * that it will contain any address to
3143                                  * be updated.
3144                                  */
3145                                 carp_link_addrs(sc, ifp, ifa);
3146                         } else {
3147                                 /*
3148                                  * First try breaking tie with the old
3149                                  * address.  Then see whether we could
3150                                  * link certain vhaddr to the new address.
3151                                  * If that fails, i.e. carpdev is NULL,
3152                                  * we try a global update.
3153                                  *
3154                                  * NOTE: The above order is critical.
3155                                  */
3156                                 carp_unlink_addrs(sc, ifp, ifa);
3157                                 carp_link_addrs(sc, ifp, ifa);
3158                                 if (sc->sc_carpdev == NULL) {
3159                                         /*
3160                                          * See the comment in the above
3161                                          * IFADDR_EVENT_DELETE block.
3162                                          */
3163                                         carp_update_addrs(sc, NULL);
3164                                 }
3165                         }
3166                         break;
3167                 }
3168         }
3169 }
3170
3171 void
3172 carp_proto_ctlinput(netmsg_t msg)
3173 {
3174         int cmd = msg->ctlinput.nm_cmd;
3175         struct sockaddr *sa = msg->ctlinput.nm_arg;
3176         struct in_ifaddr_container *iac;
3177
3178         TAILQ_FOREACH(iac, &in_ifaddrheads[mycpuid], ia_link) {
3179                 struct in_ifaddr *ia = iac->ia;
3180                 struct ifnet *ifp = ia->ia_ifp;
3181
3182                 if (ifp->if_type == IFT_CARP)
3183                         continue;
3184
3185                 if (ia->ia_ifa.ifa_addr == sa) {
3186                         if (cmd == PRC_IFDOWN) {
3187                                 carp_ifaddr(NULL, ifp, IFADDR_EVENT_DELETE,
3188                                     &ia->ia_ifa);
3189                         } else if (cmd == PRC_IFUP) {
3190                                 carp_ifaddr(NULL, ifp, IFADDR_EVENT_ADD,
3191                                     &ia->ia_ifa);
3192                         }
3193                         break;
3194                 }
3195         }
3196
3197         lwkt_replymsg(&msg->lmsg, 0);
3198 }
3199
3200 struct ifnet *
3201 carp_parent(struct ifnet *cifp)
3202 {
3203         struct carp_softc *sc;
3204
3205         KKASSERT(cifp->if_type == IFT_CARP);
3206         sc = cifp->if_softc;
3207
3208         return sc->sc_carpdev;
3209 }
3210
3211 #define rtinitflags(x) \
3212         (((x)->ia_ifp->if_flags & (IFF_LOOPBACK | IFF_POINTOPOINT)) \
3213                  ? RTF_HOST : 0)
3214
3215 static int
3216 carp_addroute_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha)
3217 {
3218         struct in_ifaddr *ia, *iaback;
3219
3220         if (sc->sc_state != MASTER)
3221                 return 0;
3222
3223         ia = vha->vha_ia;
3224         KKASSERT(ia != NULL);
3225
3226         iaback = vha->vha_iaback;
3227         KKASSERT(iaback != NULL);
3228
3229         return rtchange(&iaback->ia_ifa, &ia->ia_ifa);
3230 }
3231
3232 static void
3233 carp_delroute_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha,
3234     boolean_t del_iaback)
3235 {
3236         struct in_ifaddr *ia, *iaback;
3237
3238         ia = vha->vha_ia;
3239         KKASSERT(ia != NULL);
3240
3241         iaback = vha->vha_iaback;
3242         KKASSERT(iaback != NULL);
3243
3244         if (!del_iaback && (iaback->ia_ifp->if_flags & IFF_UP)) {
3245                 rtchange(&ia->ia_ifa, &iaback->ia_ifa);
3246                 return;
3247         }
3248
3249         rtinit(&ia->ia_ifa, RTM_DELETE, rtinitflags(ia));
3250         in_ifadown_force(&ia->ia_ifa, 1);
3251         ia->ia_flags &= ~IFA_ROUTE;
3252 }
3253
3254 static int
3255 carp_modevent(module_t mod, int type, void *data)
3256 {
3257         switch (type) {
3258         case MOD_LOAD:
3259                 LIST_INIT(&carpif_list);
3260                 carp_ifdetach_event =
3261                 EVENTHANDLER_REGISTER(ifnet_detach_event, carp_ifdetach, NULL,
3262                                       EVENTHANDLER_PRI_ANY);
3263                 carp_ifaddr_event =
3264                 EVENTHANDLER_REGISTER(ifaddr_event, carp_ifaddr, NULL,
3265                                       EVENTHANDLER_PRI_FIRST);
3266                 if_clone_attach(&carp_cloner);
3267                 break;
3268
3269         case MOD_UNLOAD:
3270                 EVENTHANDLER_DEREGISTER(ifnet_detach_event,
3271                                         carp_ifdetach_event);
3272                 EVENTHANDLER_DEREGISTER(ifaddr_event,
3273                                         carp_ifaddr_event);
3274                 if_clone_detach(&carp_cloner);
3275                 break;
3276
3277         default:
3278                 return (EINVAL);
3279         }
3280         return (0);
3281 }
3282
3283 static moduledata_t carp_mod = {
3284         "carp",
3285         carp_modevent,
3286         0
3287 };
3288 DECLARE_MODULE(carp, carp_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);