sys/netinet/ip_carp.c

   1 /*-
   2  * Copyright (c) 2002 Michael Shalayeff.
   3  * Copyright (c) 2003 Ryan McBride.
   4  * Copyright (c) 2011 Gleb Smirnoff <glebius@FreeBSD.org>
   5  * All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  17  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  18  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  19  * IN NO EVENT SHALL THE AUTHOR OR HIS RELATIVES BE LIABLE FOR ANY DIRECT,
  20  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  21  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  22  * SERVICES; LOSS OF MIND, USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
  24  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
  25  * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
  26  * THE POSSIBILITY OF SUCH DAMAGE.
  27  */
  28
  29 #include <sys/cdefs.h>
  30 __FBSDID("$FreeBSD$");
  31
  32 #include "opt_bpf.h"
  33 #include "opt_inet.h"
  34 #include "opt_inet6.h"
  35
  36 #include <sys/param.h>
  37 #include <sys/systm.h>
  38 #include <sys/bus.h>
  39 #include <sys/jail.h>
  40 #include <sys/kernel.h>
  41 #include <sys/limits.h>
  42 #include <sys/malloc.h>
  43 #include <sys/mbuf.h>
  44 #include <sys/module.h>
  45 #include <sys/priv.h>
  46 #include <sys/proc.h>
  47 #include <sys/protosw.h>
  48 #include <sys/socket.h>
  49 #include <sys/sockio.h>
  50 #include <sys/sysctl.h>
  51 #include <sys/syslog.h>
  52 #include <sys/taskqueue.h>
  53 #include <sys/counter.h>
  54
  55 #include <net/ethernet.h>
  56 #include <net/fddi.h>
  57 #include <net/if.h>
  58 #include <net/if_var.h>
  59 #include <net/if_dl.h>
  60 #include <net/if_llatbl.h>
  61 #include <net/if_types.h>
  62 #include <net/iso88025.h>
  63 #include <net/route.h>
  64 #include <net/vnet.h>
  65
  66 #if defined(INET) || defined(INET6)
  67 #include <netinet/in.h>
  68 #include <netinet/in_var.h>
  69 #include <netinet/ip_carp.h>
  70 #include <netinet/ip.h>
  71 #include <machine/in_cksum.h>
  72 #endif
  73 #ifdef INET
  74 #include <netinet/ip_var.h>
  75 #include <netinet/if_ether.h>
  76 #endif
  77
  78 #ifdef INET6
  79 #include <netinet/icmp6.h>
  80 #include <netinet/ip6.h>
  81 #include <netinet6/in6_var.h>
  82 #include <netinet6/ip6_var.h>
  83 #include <netinet6/scope6_var.h>
  84 #include <netinet6/nd6.h>
  85 #endif
  86
  87 #include <crypto/sha1.h>
  88
  89 static MALLOC_DEFINE(M_CARP, "CARP", "CARP addresses");
  90
  91 struct carp_softc {
  92         struct ifnet            *sc_carpdev;    /* Pointer to parent ifnet. */
  93         struct ifaddr           **sc_ifas;      /* Our ifaddrs. */
  94         struct sockaddr_dl      sc_addr;        /* Our link level address. */
  95         struct callout          sc_ad_tmo;      /* Advertising timeout. */
  96 #ifdef INET
  97         struct callout          sc_md_tmo;      /* Master down timeout. */
  98 #endif
  99 #ifdef INET6
 100         struct callout          sc_md6_tmo;     /* XXX: Master down timeout. */
 101 #endif
 102         struct mtx              sc_mtx;
 103
 104         int                     sc_vhid;
 105         int                     sc_advskew;
 106         int                     sc_advbase;
 107
 108         int                     sc_naddrs;
 109         int                     sc_naddrs6;
 110         int                     sc_ifasiz;
 111         enum { INIT = 0, BACKUP, MASTER }       sc_state;
 112         int                     sc_suppress;
 113         int                     sc_sendad_errors;
 114 #define CARP_SENDAD_MAX_ERRORS  3
 115         int                     sc_sendad_success;
 116 #define CARP_SENDAD_MIN_SUCCESS 3
 117
 118         int                     sc_init_counter;
 119         uint64_t                sc_counter;
 120
 121         /* authentication */
 122 #define CARP_HMAC_PAD   64
 123         unsigned char sc_key[CARP_KEY_LEN];
 124         unsigned char sc_pad[CARP_HMAC_PAD];
 125         SHA1_CTX sc_sha1;
 126
 127         TAILQ_ENTRY(carp_softc) sc_list;        /* On the carp_if list. */
 128         LIST_ENTRY(carp_softc)  sc_next;        /* On the global list. */
 129 };
 130
 131 struct carp_if {
 132 #ifdef INET
 133         int     cif_naddrs;
 134 #endif
 135 #ifdef INET6
 136         int     cif_naddrs6;
 137 #endif
 138         TAILQ_HEAD(, carp_softc) cif_vrs;
 139 #ifdef INET
 140         struct ip_moptions       cif_imo;
 141 #endif
 142 #ifdef INET6
 143         struct ip6_moptions      cif_im6o;
 144 #endif
 145         struct ifnet    *cif_ifp;
 146         struct mtx      cif_mtx;
 147         uint32_t        cif_flags;
 148 #define CIF_PROMISC     0x00000001
 149 };
 150
 151 #define CARP_INET       0
 152 #define CARP_INET6      1
 153 static int proto_reg[] = {-1, -1};
 154
 155 /*
 156  * Brief design of carp(4).
 157  *
 158  * Any carp-capable ifnet may have a list of carp softcs hanging off
 159  * its ifp->if_carp pointer. Each softc represents one unique virtual
 160  * host id, or vhid. The softc has a back pointer to the ifnet. All
 161  * softcs are joined in a global list, which has quite limited use.
 162  *
 163  * Any interface address that takes part in CARP negotiation has a
 164  * pointer to the softc of its vhid, ifa->ifa_carp. That could be either
 165  * AF_INET or AF_INET6 address.
 166  *
 167  * Although, one can get the softc's backpointer to ifnet and traverse
 168  * through its ifp->if_addrhead queue to find all interface addresses
 169  * involved in CARP, we keep a growable array of ifaddr pointers. This
 170  * allows us to avoid grabbing the IF_ADDR_LOCK() in many traversals that
 171  * do calls into the network stack, thus avoiding LORs.
 172  *
 173  * Locking:
 174  *
 175  * Each softc has a lock sc_mtx. It is used to synchronise carp_input_c(),
 176  * callout-driven events and ioctl()s.
 177  *
 178  * To traverse the list of softcs on an ifnet we use CIF_LOCK(), to
 179  * traverse the global list we use the mutex carp_mtx.
 180  *
 181  * Known issues with locking:
 182  *
 183  * - There is no protection for races between two ioctl() requests,
 184  *   neither SIOCSVH, nor SIOCAIFADDR & SIOCAIFADDR_IN6. I think that all
 185  *   interface ioctl()s should be serialized right in net/if.c.
 186  * - Sending ad, we put the pointer to the softc in an mtag, and no reference
 187  *   counting is done on the softc.
 188  * - On module unload we may race (?) with packet processing thread
 189  *   dereferencing our function pointers.
 190  */
 191
 192 /* Accept incoming CARP packets. */
 193 static VNET_DEFINE(int, carp_allow) = 1;
 194 #define V_carp_allow    VNET(carp_allow)
 195
 196 /* Preempt slower nodes. */
 197 static VNET_DEFINE(int, carp_preempt) = 0;
 198 #define V_carp_preempt  VNET(carp_preempt)
 199
 200 /* Log level. */
 201 static VNET_DEFINE(int, carp_log) = 1;
 202 #define V_carp_log      VNET(carp_log)
 203
 204 /* Global advskew demotion. */
 205 static VNET_DEFINE(int, carp_demotion) = 0;
 206 #define V_carp_demotion VNET(carp_demotion)
 207
 208 /* Send error demotion factor. */
 209 static VNET_DEFINE(int, carp_senderr_adj) = CARP_MAXSKEW;
 210 #define V_carp_senderr_adj      VNET(carp_senderr_adj)
 211
 212 /* Iface down demotion factor. */
 213 static VNET_DEFINE(int, carp_ifdown_adj) = CARP_MAXSKEW;
 214 #define V_carp_ifdown_adj       VNET(carp_ifdown_adj)
 215
 216 static int carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS);
 217
 218 SYSCTL_NODE(_net_inet, IPPROTO_CARP,    carp,   CTLFLAG_RW, 0,  "CARP");
 219 SYSCTL_INT(_net_inet_carp, OID_AUTO, allow, CTLFLAG_VNET | CTLFLAG_RW,
 220     &VNET_NAME(carp_allow), 0, "Accept incoming CARP packets");
 221 SYSCTL_INT(_net_inet_carp, OID_AUTO, preempt, CTLFLAG_VNET | CTLFLAG_RW,
 222     &VNET_NAME(carp_preempt), 0, "High-priority backup preemption mode");
 223 SYSCTL_INT(_net_inet_carp, OID_AUTO, log, CTLFLAG_VNET | CTLFLAG_RW,
 224     &VNET_NAME(carp_log), 0, "CARP log level");
 225 SYSCTL_PROC(_net_inet_carp, OID_AUTO, demotion,
 226     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW,
 227     0, 0, carp_demote_adj_sysctl, "I",
 228     "Adjust demotion factor (skew of advskew)");
 229 SYSCTL_INT(_net_inet_carp, OID_AUTO, senderr_demotion_factor,
 230     CTLFLAG_VNET | CTLFLAG_RW,
 231     &VNET_NAME(carp_senderr_adj), 0, "Send error demotion factor adjustment");
 232 SYSCTL_INT(_net_inet_carp, OID_AUTO, ifdown_demotion_factor,
 233     CTLFLAG_VNET | CTLFLAG_RW,
 234     &VNET_NAME(carp_ifdown_adj), 0,
 235     "Interface down demotion factor adjustment");
 236
 237 VNET_PCPUSTAT_DEFINE(struct carpstats, carpstats);
 238 VNET_PCPUSTAT_SYSINIT(carpstats);
 239 VNET_PCPUSTAT_SYSUNINIT(carpstats);
 240
 241 #define CARPSTATS_ADD(name, val)        \
 242     counter_u64_add(VNET(carpstats)[offsetof(struct carpstats, name) / \
 243         sizeof(uint64_t)], (val))
 244 #define CARPSTATS_INC(name)             CARPSTATS_ADD(name, 1)
 245
 246 SYSCTL_VNET_PCPUSTAT(_net_inet_carp, OID_AUTO, stats, struct carpstats,
 247     carpstats, "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 248
 249 #define CARP_LOCK_INIT(sc)      mtx_init(&(sc)->sc_mtx, "carp_softc",   \
 250         NULL, MTX_DEF)
 251 #define CARP_LOCK_DESTROY(sc)   mtx_destroy(&(sc)->sc_mtx)
 252 #define CARP_LOCK_ASSERT(sc)    mtx_assert(&(sc)->sc_mtx, MA_OWNED)
 253 #define CARP_LOCK(sc)           mtx_lock(&(sc)->sc_mtx)
 254 #define CARP_UNLOCK(sc)         mtx_unlock(&(sc)->sc_mtx)
 255 #define CIF_LOCK_INIT(cif)      mtx_init(&(cif)->cif_mtx, "carp_if",   \
 256         NULL, MTX_DEF)
 257 #define CIF_LOCK_DESTROY(cif)   mtx_destroy(&(cif)->cif_mtx)
 258 #define CIF_LOCK_ASSERT(cif)    mtx_assert(&(cif)->cif_mtx, MA_OWNED)
 259 #define CIF_LOCK(cif)           mtx_lock(&(cif)->cif_mtx)
 260 #define CIF_UNLOCK(cif)         mtx_unlock(&(cif)->cif_mtx)
 261 #define CIF_FREE(cif)   do {                            \
 262                 CIF_LOCK_ASSERT(cif);                   \
 263                 if (TAILQ_EMPTY(&(cif)->cif_vrs))       \
 264                         carp_free_if(cif);              \
 265                 else                                    \
 266                         CIF_UNLOCK(cif);                \
 267 } while (0)
 268
 269 #define CARP_LOG(...)   do {                            \
 270         if (V_carp_log > 0)                             \
 271                 log(LOG_INFO, "carp: " __VA_ARGS__);    \
 272 } while (0)
 273
 274 #define CARP_DEBUG(...) do {                            \
 275         if (V_carp_log > 1)                             \
 276                 log(LOG_DEBUG, __VA_ARGS__);            \
 277 } while (0)
 278
 279 #define IFNET_FOREACH_IFA(ifp, ifa)                                     \
 280         IF_ADDR_LOCK_ASSERT(ifp);                                       \
 281         TAILQ_FOREACH((ifa), &(ifp)->if_addrhead, ifa_link)             \
 282                 if ((ifa)->ifa_carp != NULL)
 283
 284 #define CARP_FOREACH_IFA(sc, ifa)                                       \
 285         CARP_LOCK_ASSERT(sc);                                           \
 286         for (int _i = 0;                                                \
 287                 _i < (sc)->sc_naddrs + (sc)->sc_naddrs6 &&              \
 288                 ((ifa) = sc->sc_ifas[_i]) != NULL;                      \
 289                 ++_i)
 290
 291 #define IFNET_FOREACH_CARP(ifp, sc)                                     \
 292         CIF_LOCK_ASSERT(ifp->if_carp);                                  \
 293         TAILQ_FOREACH((sc), &(ifp)->if_carp->cif_vrs, sc_list)
 294
 295 #define DEMOTE_ADVSKEW(sc)                                      \
 296     (((sc)->sc_advskew + V_carp_demotion > CARP_MAXSKEW) ?      \
 297     CARP_MAXSKEW : ((sc)->sc_advskew + V_carp_demotion))
 298
 299 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 300 static struct carp_softc
 301                 *carp_alloc(struct ifnet *);
 302 static void     carp_detach_locked(struct ifaddr *);
 303 static void     carp_destroy(struct carp_softc *);
 304 static struct carp_if
 305                 *carp_alloc_if(struct ifnet *);
 306 static void     carp_free_if(struct carp_if *);
 307 static void     carp_set_state(struct carp_softc *, int, const char* reason);
 308 static void     carp_sc_state(struct carp_softc *);
 309 static void     carp_setrun(struct carp_softc *, sa_family_t);
 310 static void     carp_master_down(void *);
 311 static void     carp_master_down_locked(struct carp_softc *,
 312                     const char* reason);
 313 static void     carp_send_ad(void *);
 314 static void     carp_send_ad_locked(struct carp_softc *);
 315 static void     carp_addroute(struct carp_softc *);
 316 static void     carp_ifa_addroute(struct ifaddr *);
 317 static void     carp_delroute(struct carp_softc *);
 318 static void     carp_ifa_delroute(struct ifaddr *);
 319 static void     carp_send_ad_all(void *, int);
 320 static void     carp_demote_adj(int, char *);
 321
 322 static LIST_HEAD(, carp_softc) carp_list;
 323 static struct mtx carp_mtx;
 324 static struct task carp_sendall_task =
 325     TASK_INITIALIZER(0, carp_send_ad_all, NULL);
 326
 327 static void
 328 carp_hmac_prepare(struct carp_softc *sc)
 329 {
 330         uint8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 331         uint8_t vhid = sc->sc_vhid & 0xff;
 332         struct ifaddr *ifa;
 333         int i, found;
 334 #ifdef INET
 335         struct in_addr last, cur, in;
 336 #endif
 337 #ifdef INET6
 338         struct in6_addr last6, cur6, in6;
 339 #endif
 340
 341         CARP_LOCK_ASSERT(sc);
 342
 343         /* Compute ipad from key. */
 344         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 345         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 346         for (i = 0; i < sizeof(sc->sc_pad); i++)
 347                 sc->sc_pad[i] ^= 0x36;
 348
 349         /* Precompute first part of inner hash. */
 350         SHA1Init(&sc->sc_sha1);
 351         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 352         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 353         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 354         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 355 #ifdef INET
 356         cur.s_addr = 0;
 357         do {
 358                 found = 0;
 359                 last = cur;
 360                 cur.s_addr = 0xffffffff;
 361                 CARP_FOREACH_IFA(sc, ifa) {
 362                         in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 363                         if (ifa->ifa_addr->sa_family == AF_INET &&
 364                             ntohl(in.s_addr) > ntohl(last.s_addr) &&
 365                             ntohl(in.s_addr) < ntohl(cur.s_addr)) {
 366                                 cur.s_addr = in.s_addr;
 367                                 found++;
 368                         }
 369                 }
 370                 if (found)
 371                         SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur));
 372         } while (found);
 373 #endif /* INET */
 374 #ifdef INET6
 375         memset(&cur6, 0, sizeof(cur6));
 376         do {
 377                 found = 0;
 378                 last6 = cur6;
 379                 memset(&cur6, 0xff, sizeof(cur6));
 380                 CARP_FOREACH_IFA(sc, ifa) {
 381                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 382                         if (IN6_IS_SCOPE_EMBED(&in6))
 383                                 in6.s6_addr16[1] = 0;
 384                         if (ifa->ifa_addr->sa_family == AF_INET6 &&
 385                             memcmp(&in6, &last6, sizeof(in6)) > 0 &&
 386                             memcmp(&in6, &cur6, sizeof(in6)) < 0) {
 387                                 cur6 = in6;
 388                                 found++;
 389                         }
 390                 }
 391                 if (found)
 392                         SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6));
 393         } while (found);
 394 #endif /* INET6 */
 395
 396         /* convert ipad to opad */
 397         for (i = 0; i < sizeof(sc->sc_pad); i++)
 398                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 399 }
 400
 401 static void
 402 carp_hmac_generate(struct carp_softc *sc, uint32_t counter[2],
 403     unsigned char md[20])
 404 {
 405         SHA1_CTX sha1ctx;
 406
 407         CARP_LOCK_ASSERT(sc);
 408
 409         /* fetch first half of inner hash */
 410         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 411
 412         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 413         SHA1Final(md, &sha1ctx);
 414
 415         /* outer hash */
 416         SHA1Init(&sha1ctx);
 417         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 418         SHA1Update(&sha1ctx, md, 20);
 419         SHA1Final(md, &sha1ctx);
 420 }
 421
 422 static int
 423 carp_hmac_verify(struct carp_softc *sc, uint32_t counter[2],
 424     unsigned char md[20])
 425 {
 426         unsigned char md2[20];
 427
 428         CARP_LOCK_ASSERT(sc);
 429
 430         carp_hmac_generate(sc, counter, md2);
 431
 432         return (bcmp(md, md2, sizeof(md2)));
 433 }
 434
 435 /*
 436  * process input packet.
 437  * we have rearranged checks order compared to the rfc,
 438  * but it seems more efficient this way or not possible otherwise.
 439  */
 440 #ifdef INET
 441 int
 442 carp_input(struct mbuf **mp, int *offp, int proto)
 443 {
 444         struct mbuf *m = *mp;
 445         struct ip *ip = mtod(m, struct ip *);
 446         struct carp_header *ch;
 447         int iplen, len;
 448
 449         iplen = *offp;
 450         *mp = NULL;
 451
 452         CARPSTATS_INC(carps_ipackets);
 453
 454         if (!V_carp_allow) {
 455                 m_freem(m);
 456                 return (IPPROTO_DONE);
 457         }
 458
 459         /* verify that the IP TTL is 255.  */
 460         if (ip->ip_ttl != CARP_DFLTTL) {
 461                 CARPSTATS_INC(carps_badttl);
 462                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 463                     ip->ip_ttl,
 464                     m->m_pkthdr.rcvif->if_xname);
 465                 m_freem(m);
 466                 return (IPPROTO_DONE);
 467         }
 468
 469         iplen = ip->ip_hl << 2;
 470
 471         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 472                 CARPSTATS_INC(carps_badlen);
 473                 CARP_DEBUG("%s: received len %zd < sizeof(struct carp_header) "
 474                     "on %s\n", __func__, m->m_len - sizeof(struct ip),
 475                     m->m_pkthdr.rcvif->if_xname);
 476                 m_freem(m);
 477                 return (IPPROTO_DONE);
 478         }
 479
 480         if (iplen + sizeof(*ch) < m->m_len) {
 481                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 482                         CARPSTATS_INC(carps_hdrops);
 483                         CARP_DEBUG("%s: pullup failed\n", __func__);
 484                         return (IPPROTO_DONE);
 485                 }
 486                 ip = mtod(m, struct ip *);
 487         }
 488         ch = (struct carp_header *)((char *)ip + iplen);
 489
 490         /*
 491          * verify that the received packet length is
 492          * equal to the CARP header
 493          */
 494         len = iplen + sizeof(*ch);
 495         if (len > m->m_pkthdr.len) {
 496                 CARPSTATS_INC(carps_badlen);
 497                 CARP_DEBUG("%s: packet too short %d on %s\n", __func__,
 498                     m->m_pkthdr.len,
 499                     m->m_pkthdr.rcvif->if_xname);
 500                 m_freem(m);
 501                 return (IPPROTO_DONE);
 502         }
 503
 504         if ((m = m_pullup(m, len)) == NULL) {
 505                 CARPSTATS_INC(carps_hdrops);
 506                 return (IPPROTO_DONE);
 507         }
 508         ip = mtod(m, struct ip *);
 509         ch = (struct carp_header *)((char *)ip + iplen);
 510
 511         /* verify the CARP checksum */
 512         m->m_data += iplen;
 513         if (in_cksum(m, len - iplen)) {
 514                 CARPSTATS_INC(carps_badsum);
 515                 CARP_DEBUG("%s: checksum failed on %s\n", __func__,
 516                     m->m_pkthdr.rcvif->if_xname);
 517                 m_freem(m);
 518                 return (IPPROTO_DONE);
 519         }
 520         m->m_data -= iplen;
 521
 522         carp_input_c(m, ch, AF_INET);
 523         return (IPPROTO_DONE);
 524 }
 525 #endif
 526
 527 #ifdef INET6
 528 int
 529 carp6_input(struct mbuf **mp, int *offp, int proto)
 530 {
 531         struct mbuf *m = *mp;
 532         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 533         struct carp_header *ch;
 534         u_int len;
 535
 536         CARPSTATS_INC(carps_ipackets6);
 537
 538         if (!V_carp_allow) {
 539                 m_freem(m);
 540                 return (IPPROTO_DONE);
 541         }
 542
 543         /* check if received on a valid carp interface */
 544         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 545                 CARPSTATS_INC(carps_badif);
 546                 CARP_DEBUG("%s: packet received on non-carp interface: %s\n",
 547                     __func__, m->m_pkthdr.rcvif->if_xname);
 548                 m_freem(m);
 549                 return (IPPROTO_DONE);
 550         }
 551
 552         /* verify that the IP TTL is 255 */
 553         if (ip6->ip6_hlim != CARP_DFLTTL) {
 554                 CARPSTATS_INC(carps_badttl);
 555                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 556                     ip6->ip6_hlim, m->m_pkthdr.rcvif->if_xname);
 557                 m_freem(m);
 558                 return (IPPROTO_DONE);
 559         }
 560
 561         /* verify that we have a complete carp packet */
 562         len = m->m_len;
 563         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 564         if (ch == NULL) {
 565                 CARPSTATS_INC(carps_badlen);
 566                 CARP_DEBUG("%s: packet size %u too small\n", __func__, len);
 567                 return (IPPROTO_DONE);
 568         }
 569
 570
 571         /* verify the CARP checksum */
 572         m->m_data += *offp;
 573         if (in_cksum(m, sizeof(*ch))) {
 574                 CARPSTATS_INC(carps_badsum);
 575                 CARP_DEBUG("%s: checksum failed, on %s\n", __func__,
 576                     m->m_pkthdr.rcvif->if_xname);
 577                 m_freem(m);
 578                 return (IPPROTO_DONE);
 579         }
 580         m->m_data -= *offp;
 581
 582         carp_input_c(m, ch, AF_INET6);
 583         return (IPPROTO_DONE);
 584 }
 585 #endif /* INET6 */
 586
 587 static void
 588 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 589 {
 590         struct ifnet *ifp = m->m_pkthdr.rcvif;
 591         struct ifaddr *ifa;
 592         struct carp_softc *sc;
 593         uint64_t tmp_counter;
 594         struct timeval sc_tv, ch_tv;
 595
 596         /* verify that the VHID is valid on the receiving interface */
 597         IF_ADDR_RLOCK(ifp);
 598         IFNET_FOREACH_IFA(ifp, ifa)
 599                 if (ifa->ifa_addr->sa_family == af &&
 600                     ifa->ifa_carp->sc_vhid == ch->carp_vhid) {
 601                         ifa_ref(ifa);
 602                         break;
 603                 }
 604         IF_ADDR_RUNLOCK(ifp);
 605
 606         if (ifa == NULL) {
 607                 CARPSTATS_INC(carps_badvhid);
 608                 m_freem(m);
 609                 return;
 610         }
 611
 612         /* verify the CARP version. */
 613         if (ch->carp_version != CARP_VERSION) {
 614                 CARPSTATS_INC(carps_badver);
 615                 CARP_DEBUG("%s: invalid version %d\n", ifp->if_xname,
 616                     ch->carp_version);
 617                 ifa_free(ifa);
 618                 m_freem(m);
 619                 return;
 620         }
 621
 622         sc = ifa->ifa_carp;
 623         CARP_LOCK(sc);
 624         ifa_free(ifa);
 625
 626         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 627                 CARPSTATS_INC(carps_badauth);
 628                 CARP_DEBUG("%s: incorrect hash for VHID %u@%s\n", __func__,
 629                     sc->sc_vhid, ifp->if_xname);
 630                 goto out;
 631         }
 632
 633         tmp_counter = ntohl(ch->carp_counter[0]);
 634         tmp_counter = tmp_counter<<32;
 635         tmp_counter += ntohl(ch->carp_counter[1]);
 636
 637         /* XXX Replay protection goes here */
 638
 639         sc->sc_init_counter = 0;
 640         sc->sc_counter = tmp_counter;
 641
 642         sc_tv.tv_sec = sc->sc_advbase;
 643         sc_tv.tv_usec = DEMOTE_ADVSKEW(sc) * 1000000 / 256;
 644         ch_tv.tv_sec = ch->carp_advbase;
 645         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 646
 647         switch (sc->sc_state) {
 648         case INIT:
 649                 break;
 650         case MASTER:
 651                 /*
 652                  * If we receive an advertisement from a master who's going to
 653                  * be more frequent than us, go into BACKUP state.
 654                  */
 655                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 656                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 657                         callout_stop(&sc->sc_ad_tmo);
 658                         carp_set_state(sc, BACKUP,
 659                             "more frequent advertisement received");
 660                         carp_setrun(sc, 0);
 661                         carp_delroute(sc);
 662                 }
 663                 break;
 664         case BACKUP:
 665                 /*
 666                  * If we're pre-empting masters who advertise slower than us,
 667                  * and this one claims to be slower, treat him as down.
 668                  */
 669                 if (V_carp_preempt && timevalcmp(&sc_tv, &ch_tv, <)) {
 670                         carp_master_down_locked(sc,
 671                             "preempting a slower master");
 672                         break;
 673                 }
 674
 675                 /*
 676                  *  If the master is going to advertise at such a low frequency
 677                  *  that he's guaranteed to time out, we'd might as well just
 678                  *  treat him as timed out now.
 679                  */
 680                 sc_tv.tv_sec = sc->sc_advbase * 3;
 681                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 682                         carp_master_down_locked(sc, "master will time out");
 683                         break;
 684                 }
 685
 686                 /*
 687                  * Otherwise, we reset the counter and wait for the next
 688                  * advertisement.
 689                  */
 690                 carp_setrun(sc, af);
 691                 break;
 692         }
 693
 694 out:
 695         CARP_UNLOCK(sc);
 696         m_freem(m);
 697 }
 698
 699 static int
 700 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 701 {
 702         struct m_tag *mtag;
 703
 704         if (sc->sc_init_counter) {
 705                 /* this could also be seconds since unix epoch */
 706                 sc->sc_counter = arc4random();
 707                 sc->sc_counter = sc->sc_counter << 32;
 708                 sc->sc_counter += arc4random();
 709         } else
 710                 sc->sc_counter++;
 711
 712         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 713         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 714
 715         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 716
 717         /* Tag packet for carp_output */
 718         if ((mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct carp_softc *),
 719             M_NOWAIT)) == NULL) {
 720                 m_freem(m);
 721                 CARPSTATS_INC(carps_onomem);
 722                 return (ENOMEM);
 723         }
 724         bcopy(&sc, mtag + 1, sizeof(sc));
 725         m_tag_prepend(m, mtag);
 726
 727         return (0);
 728 }
 729
 730 /*
 731  * To avoid LORs and possible recursions this function shouldn't
 732  * be called directly, but scheduled via taskqueue.
 733  */
 734 static void
 735 carp_send_ad_all(void *ctx __unused, int pending __unused)
 736 {
 737         struct carp_softc *sc;
 738
 739         mtx_lock(&carp_mtx);
 740         LIST_FOREACH(sc, &carp_list, sc_next)
 741                 if (sc->sc_state == MASTER) {
 742                         CARP_LOCK(sc);
 743                         CURVNET_SET(sc->sc_carpdev->if_vnet);
 744                         carp_send_ad_locked(sc);
 745                         CURVNET_RESTORE();
 746                         CARP_UNLOCK(sc);
 747                 }
 748         mtx_unlock(&carp_mtx);
 749 }
 750
 751 /* Send a periodic advertisement, executed in callout context. */
 752 static void
 753 carp_send_ad(void *v)
 754 {
 755         struct carp_softc *sc = v;
 756
 757         CARP_LOCK_ASSERT(sc);
 758         CURVNET_SET(sc->sc_carpdev->if_vnet);
 759         carp_send_ad_locked(sc);
 760         CURVNET_RESTORE();
 761         CARP_UNLOCK(sc);
 762 }
 763
 764 static void
 765 carp_send_ad_error(struct carp_softc *sc, int error)
 766 {
 767
 768         if (error) {
 769                 if (sc->sc_sendad_errors < INT_MAX)
 770                         sc->sc_sendad_errors++;
 771                 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 772                         static const char fmt[] = "send error %d on %s";
 773                         char msg[sizeof(fmt) + IFNAMSIZ];
 774
 775                         sprintf(msg, fmt, error, sc->sc_carpdev->if_xname);
 776                         carp_demote_adj(V_carp_senderr_adj, msg);
 777                 }
 778                 sc->sc_sendad_success = 0;
 779         } else {
 780                 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS &&
 781                     ++sc->sc_sendad_success >= CARP_SENDAD_MIN_SUCCESS) {
 782                         static const char fmt[] = "send ok on %s";
 783                         char msg[sizeof(fmt) + IFNAMSIZ];
 784
 785                         sprintf(msg, fmt, sc->sc_carpdev->if_xname);
 786                         carp_demote_adj(-V_carp_senderr_adj, msg);
 787                         sc->sc_sendad_errors = 0;
 788                 } else
 789                         sc->sc_sendad_errors = 0;
 790         }
 791 }
 792
 793 static void
 794 carp_send_ad_locked(struct carp_softc *sc)
 795 {
 796         struct carp_header ch;
 797         struct timeval tv;
 798         struct sockaddr sa;
 799         struct ifaddr *ifa;
 800         struct carp_header *ch_ptr;
 801         struct mbuf *m;
 802         int len, advskew;
 803
 804         CARP_LOCK_ASSERT(sc);
 805
 806         advskew = DEMOTE_ADVSKEW(sc);
 807         tv.tv_sec = sc->sc_advbase;
 808         tv.tv_usec = advskew * 1000000 / 256;
 809
 810         ch.carp_version = CARP_VERSION;
 811         ch.carp_type = CARP_ADVERTISEMENT;
 812         ch.carp_vhid = sc->sc_vhid;
 813         ch.carp_advbase = sc->sc_advbase;
 814         ch.carp_advskew = advskew;
 815         ch.carp_authlen = 7;    /* XXX DEFINE */
 816         ch.carp_pad1 = 0;       /* must be zero */
 817         ch.carp_cksum = 0;
 818
 819         /* XXXGL: OpenBSD picks first ifaddr with needed family. */
 820
 821 #ifdef INET
 822         if (sc->sc_naddrs) {
 823                 struct ip *ip;
 824
 825                 m = m_gethdr(M_NOWAIT, MT_DATA);
 826                 if (m == NULL) {
 827                         CARPSTATS_INC(carps_onomem);
 828                         goto resched;
 829                 }
 830                 len = sizeof(*ip) + sizeof(ch);
 831                 m->m_pkthdr.len = len;
 832                 m->m_pkthdr.rcvif = NULL;
 833                 m->m_len = len;
 834                 M_ALIGN(m, m->m_len);
 835                 m->m_flags |= M_MCAST;
 836                 ip = mtod(m, struct ip *);
 837                 ip->ip_v = IPVERSION;
 838                 ip->ip_hl = sizeof(*ip) >> 2;
 839                 ip->ip_tos = IPTOS_LOWDELAY;
 840                 ip->ip_len = htons(len);
 841                 ip->ip_off = htons(IP_DF);
 842                 ip->ip_ttl = CARP_DFLTTL;
 843                 ip->ip_p = IPPROTO_CARP;
 844                 ip->ip_sum = 0;
 845                 ip_fillid(ip);
 846
 847                 bzero(&sa, sizeof(sa));
 848                 sa.sa_family = AF_INET;
 849                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 850                 if (ifa != NULL) {
 851                         ip->ip_src.s_addr =
 852                             ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 853                         ifa_free(ifa);
 854                 } else
 855                         ip->ip_src.s_addr = 0;
 856                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 857
 858                 ch_ptr = (struct carp_header *)(&ip[1]);
 859                 bcopy(&ch, ch_ptr, sizeof(ch));
 860                 if (carp_prepare_ad(m, sc, ch_ptr))
 861                         goto resched;
 862
 863                 m->m_data += sizeof(*ip);
 864                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip));
 865                 m->m_data -= sizeof(*ip);
 866
 867                 CARPSTATS_INC(carps_opackets);
 868
 869                 carp_send_ad_error(sc, ip_output(m, NULL, NULL, IP_RAWOUTPUT,
 870                     &sc->sc_carpdev->if_carp->cif_imo, NULL));
 871         }
 872 #endif /* INET */
 873 #ifdef INET6
 874         if (sc->sc_naddrs6) {
 875                 struct ip6_hdr *ip6;
 876
 877                 m = m_gethdr(M_NOWAIT, MT_DATA);
 878                 if (m == NULL) {
 879                         CARPSTATS_INC(carps_onomem);
 880                         goto resched;
 881                 }
 882                 len = sizeof(*ip6) + sizeof(ch);
 883                 m->m_pkthdr.len = len;
 884                 m->m_pkthdr.rcvif = NULL;
 885                 m->m_len = len;
 886                 M_ALIGN(m, m->m_len);
 887                 m->m_flags |= M_MCAST;
 888                 ip6 = mtod(m, struct ip6_hdr *);
 889                 bzero(ip6, sizeof(*ip6));
 890                 ip6->ip6_vfc |= IPV6_VERSION;
 891                 ip6->ip6_hlim = CARP_DFLTTL;
 892                 ip6->ip6_nxt = IPPROTO_CARP;
 893                 bzero(&sa, sizeof(sa));
 894
 895                 /* set the source address */
 896                 sa.sa_family = AF_INET6;
 897                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 898                 if (ifa != NULL) {
 899                         bcopy(IFA_IN6(ifa), &ip6->ip6_src,
 900                             sizeof(struct in6_addr));
 901                         ifa_free(ifa);
 902                 } else
 903                         /* This should never happen with IPv6. */
 904                         bzero(&ip6->ip6_src, sizeof(struct in6_addr));
 905
 906                 /* Set the multicast destination. */
 907                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
 908                 ip6->ip6_dst.s6_addr8[15] = 0x12;
 909                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
 910                         m_freem(m);
 911                         CARP_DEBUG("%s: in6_setscope failed\n", __func__);
 912                         goto resched;
 913                 }
 914
 915                 ch_ptr = (struct carp_header *)(&ip6[1]);
 916                 bcopy(&ch, ch_ptr, sizeof(ch));
 917                 if (carp_prepare_ad(m, sc, ch_ptr))
 918                         goto resched;
 919
 920                 m->m_data += sizeof(*ip6);
 921                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip6));
 922                 m->m_data -= sizeof(*ip6);
 923
 924                 CARPSTATS_INC(carps_opackets6);
 925
 926                 carp_send_ad_error(sc, ip6_output(m, NULL, NULL, 0,
 927                     &sc->sc_carpdev->if_carp->cif_im6o, NULL, NULL));
 928         }
 929 #endif /* INET6 */
 930
 931 resched:
 932         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), carp_send_ad, sc);
 933 }
 934
 935 static void
 936 carp_addroute(struct carp_softc *sc)
 937 {
 938         struct ifaddr *ifa;
 939
 940         CARP_FOREACH_IFA(sc, ifa)
 941                 carp_ifa_addroute(ifa);
 942 }
 943
 944 static void
 945 carp_ifa_addroute(struct ifaddr *ifa)
 946 {
 947
 948         switch (ifa->ifa_addr->sa_family) {
 949 #ifdef INET
 950         case AF_INET:
 951                 in_addprefix(ifatoia(ifa), RTF_UP);
 952                 ifa_add_loopback_route(ifa,
 953                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 954                 break;
 955 #endif
 956 #ifdef INET6
 957         case AF_INET6:
 958                 ifa_add_loopback_route(ifa,
 959                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 960                 nd6_add_ifa_lle(ifatoia6(ifa));
 961                 break;
 962 #endif
 963         }
 964 }
 965
 966 static void
 967 carp_delroute(struct carp_softc *sc)
 968 {
 969         struct ifaddr *ifa;
 970
 971         CARP_FOREACH_IFA(sc, ifa)
 972                 carp_ifa_delroute(ifa);
 973 }
 974
 975 static void
 976 carp_ifa_delroute(struct ifaddr *ifa)
 977 {
 978
 979         switch (ifa->ifa_addr->sa_family) {
 980 #ifdef INET
 981         case AF_INET:
 982                 ifa_del_loopback_route(ifa,
 983                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 984                 in_scrubprefix(ifatoia(ifa), LLE_STATIC);
 985                 break;
 986 #endif
 987 #ifdef INET6
 988         case AF_INET6:
 989                 ifa_del_loopback_route(ifa,
 990                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 991                 nd6_rem_ifa_lle(ifatoia6(ifa));
 992                 break;
 993 #endif
 994         }
 995 }
 996
 997 int
 998 carp_master(struct ifaddr *ifa)
 999 {
1000         struct carp_softc *sc = ifa->ifa_carp;
1001
1002         return (sc->sc_state == MASTER);
1003 }
1004
1005 #ifdef INET
1006 /*
1007  * Broadcast a gratuitous ARP request containing
1008  * the virtual router MAC address for each IP address
1009  * associated with the virtual router.
1010  */
1011 static void
1012 carp_send_arp(struct carp_softc *sc)
1013 {
1014         struct ifaddr *ifa;
1015
1016         CARP_FOREACH_IFA(sc, ifa)
1017                 if (ifa->ifa_addr->sa_family == AF_INET)
1018                         arp_ifinit2(sc->sc_carpdev, ifa, LLADDR(&sc->sc_addr));
1019 }
1020
1021 int
1022 carp_iamatch(struct ifaddr *ifa, uint8_t **enaddr)
1023 {
1024         struct carp_softc *sc = ifa->ifa_carp;
1025
1026         if (sc->sc_state == MASTER) {
1027                 *enaddr = LLADDR(&sc->sc_addr);
1028                 return (1);
1029         }
1030
1031         return (0);
1032 }
1033 #endif
1034
1035 #ifdef INET6
1036 static void
1037 carp_send_na(struct carp_softc *sc)
1038 {
1039         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1040         struct ifaddr *ifa;
1041         struct in6_addr *in6;
1042
1043         CARP_FOREACH_IFA(sc, ifa) {
1044                 if (ifa->ifa_addr->sa_family != AF_INET6)
1045                         continue;
1046
1047                 in6 = IFA_IN6(ifa);
1048                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1049                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1050                 DELAY(1000);    /* XXX */
1051         }
1052 }
1053
1054 /*
1055  * Returns ifa in case it's a carp address and it is MASTER, or if the address
1056  * matches and is not a carp address.  Returns NULL otherwise.
1057  */
1058 struct ifaddr *
1059 carp_iamatch6(struct ifnet *ifp, struct in6_addr *taddr)
1060 {
1061         struct ifaddr *ifa;
1062
1063         ifa = NULL;
1064         IF_ADDR_RLOCK(ifp);
1065         TAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
1066                 if (ifa->ifa_addr->sa_family != AF_INET6)
1067                         continue;
1068                 if (!IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa)))
1069                         continue;
1070                 if (ifa->ifa_carp && ifa->ifa_carp->sc_state != MASTER)
1071                         ifa = NULL;
1072                 else
1073                         ifa_ref(ifa);
1074                 break;
1075         }
1076         IF_ADDR_RUNLOCK(ifp);
1077
1078         return (ifa);
1079 }
1080
1081 caddr_t
1082 carp_macmatch6(struct ifnet *ifp, struct mbuf *m, const struct in6_addr *taddr)
1083 {
1084         struct ifaddr *ifa;
1085
1086         IF_ADDR_RLOCK(ifp);
1087         IFNET_FOREACH_IFA(ifp, ifa)
1088                 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1089                     IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1090                         struct carp_softc *sc = ifa->ifa_carp;
1091                         struct m_tag *mtag;
1092
1093                         IF_ADDR_RUNLOCK(ifp);
1094
1095                         mtag = m_tag_get(PACKET_TAG_CARP,
1096                             sizeof(struct carp_softc *), M_NOWAIT);
1097                         if (mtag == NULL)
1098                                 /* Better a bit than nothing. */
1099                                 return (LLADDR(&sc->sc_addr));
1100
1101                         bcopy(&sc, mtag + 1, sizeof(sc));
1102                         m_tag_prepend(m, mtag);
1103
1104                         return (LLADDR(&sc->sc_addr));
1105                 }
1106         IF_ADDR_RUNLOCK(ifp);
1107
1108         return (NULL);
1109 }
1110 #endif /* INET6 */
1111
1112 int
1113 carp_forus(struct ifnet *ifp, u_char *dhost)
1114 {
1115         struct carp_softc *sc;
1116         uint8_t *ena = dhost;
1117
1118         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1119                 return (0);
1120
1121         CIF_LOCK(ifp->if_carp);
1122         IFNET_FOREACH_CARP(ifp, sc) {
1123                 CARP_LOCK(sc);
1124                 if (sc->sc_state == MASTER && !bcmp(dhost, LLADDR(&sc->sc_addr),
1125                     ETHER_ADDR_LEN)) {
1126                         CARP_UNLOCK(sc);
1127                         CIF_UNLOCK(ifp->if_carp);
1128                         return (1);
1129                 }
1130                 CARP_UNLOCK(sc);
1131         }
1132         CIF_UNLOCK(ifp->if_carp);
1133
1134         return (0);
1135 }
1136
1137 /* Master down timeout event, executed in callout context. */
1138 static void
1139 carp_master_down(void *v)
1140 {
1141         struct carp_softc *sc = v;
1142
1143         CARP_LOCK_ASSERT(sc);
1144
1145         CURVNET_SET(sc->sc_carpdev->if_vnet);
1146         if (sc->sc_state == BACKUP) {
1147                 carp_master_down_locked(sc, "master timed out");
1148         }
1149         CURVNET_RESTORE();
1150
1151         CARP_UNLOCK(sc);
1152 }
1153
1154 static void
1155 carp_master_down_locked(struct carp_softc *sc, const char *reason)
1156 {
1157
1158         CARP_LOCK_ASSERT(sc);
1159
1160         switch (sc->sc_state) {
1161         case BACKUP:
1162                 carp_set_state(sc, MASTER, reason);
1163                 carp_send_ad_locked(sc);
1164 #ifdef INET
1165                 carp_send_arp(sc);
1166 #endif
1167 #ifdef INET6
1168                 carp_send_na(sc);
1169 #endif
1170                 carp_setrun(sc, 0);
1171                 carp_addroute(sc);
1172                 break;
1173         case INIT:
1174         case MASTER:
1175 #ifdef INVARIANTS
1176                 panic("carp: VHID %u@%s: master_down event in %s state\n",
1177                     sc->sc_vhid,
1178                     sc->sc_carpdev->if_xname,
1179                     sc->sc_state ? "MASTER" : "INIT");
1180 #endif
1181                 break;
1182         }
1183 }
1184
1185 /*
1186  * When in backup state, af indicates whether to reset the master down timer
1187  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1188  */
1189 static void
1190 carp_setrun(struct carp_softc *sc, sa_family_t af)
1191 {
1192         struct timeval tv;
1193
1194         CARP_LOCK_ASSERT(sc);
1195
1196         if ((sc->sc_carpdev->if_flags & IFF_UP) == 0 ||
1197             sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1198             (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0))
1199                 return;
1200
1201         switch (sc->sc_state) {
1202         case INIT:
1203                 carp_set_state(sc, BACKUP, "initialization complete");
1204                 carp_setrun(sc, 0);
1205                 break;
1206         case BACKUP:
1207                 callout_stop(&sc->sc_ad_tmo);
1208                 tv.tv_sec = 3 * sc->sc_advbase;
1209                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1210                 switch (af) {
1211 #ifdef INET
1212                 case AF_INET:
1213                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1214                             carp_master_down, sc);
1215                         break;
1216 #endif
1217 #ifdef INET6
1218                 case AF_INET6:
1219                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1220                             carp_master_down, sc);
1221                         break;
1222 #endif
1223                 default:
1224 #ifdef INET
1225                         if (sc->sc_naddrs)
1226                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1227                                     carp_master_down, sc);
1228 #endif
1229 #ifdef INET6
1230                         if (sc->sc_naddrs6)
1231                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1232                                     carp_master_down, sc);
1233 #endif
1234                         break;
1235                 }
1236                 break;
1237         case MASTER:
1238                 tv.tv_sec = sc->sc_advbase;
1239                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1240                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1241                     carp_send_ad, sc);
1242                 break;
1243         }
1244 }
1245
1246 /*
1247  * Setup multicast structures.
1248  */
1249 static int
1250 carp_multicast_setup(struct carp_if *cif, sa_family_t sa)
1251 {
1252         struct ifnet *ifp = cif->cif_ifp;
1253         int error = 0;
1254
1255         CIF_LOCK_ASSERT(cif);
1256
1257         switch (sa) {
1258 #ifdef INET
1259         case AF_INET:
1260             {
1261                 struct ip_moptions *imo = &cif->cif_imo;
1262                 struct in_addr addr;
1263
1264                 if (imo->imo_membership)
1265                         return (0);
1266
1267                 imo->imo_membership = (struct in_multi **)malloc(
1268                     (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
1269                     M_NOWAIT);
1270                 if (imo->imo_membership == NULL)
1271                         return (ENOMEM);
1272                 imo->imo_mfilters = NULL;
1273                 imo->imo_max_memberships = IP_MIN_MEMBERSHIPS;
1274                 imo->imo_multicast_vif = -1;
1275
1276                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1277                 if ((error = in_joingroup(ifp, &addr, NULL,
1278                     &imo->imo_membership[0])) != 0) {
1279                         free(imo->imo_membership, M_CARP);
1280                         break;
1281                 }
1282                 imo->imo_num_memberships++;
1283                 imo->imo_multicast_ifp = ifp;
1284                 imo->imo_multicast_ttl = CARP_DFLTTL;
1285                 imo->imo_multicast_loop = 0;
1286                 break;
1287            }
1288 #endif
1289 #ifdef INET6
1290         case AF_INET6:
1291             {
1292                 struct ip6_moptions *im6o = &cif->cif_im6o;
1293                 struct in6_addr in6;
1294                 struct in6_multi *in6m;
1295
1296                 if (im6o->im6o_membership)
1297                         return (0);
1298
1299                 im6o->im6o_membership = (struct in6_multi **)malloc(
1300                     (sizeof(struct in6_multi *) * IPV6_MIN_MEMBERSHIPS), M_CARP,
1301                     M_ZERO | M_NOWAIT);
1302                 if (im6o->im6o_membership == NULL)
1303                         return (ENOMEM);
1304                 im6o->im6o_mfilters = NULL;
1305                 im6o->im6o_max_memberships = IPV6_MIN_MEMBERSHIPS;
1306                 im6o->im6o_multicast_hlim = CARP_DFLTTL;
1307                 im6o->im6o_multicast_ifp = ifp;
1308
1309                 /* Join IPv6 CARP multicast group. */
1310                 bzero(&in6, sizeof(in6));
1311                 in6.s6_addr16[0] = htons(0xff02);
1312                 in6.s6_addr8[15] = 0x12;
1313                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1314                         free(im6o->im6o_membership, M_CARP);
1315                         break;
1316                 }
1317                 in6m = NULL;
1318                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1319                         free(im6o->im6o_membership, M_CARP);
1320                         break;
1321                 }
1322                 im6o->im6o_membership[0] = in6m;
1323                 im6o->im6o_num_memberships++;
1324
1325                 /* Join solicited multicast address. */
1326                 bzero(&in6, sizeof(in6));
1327                 in6.s6_addr16[0] = htons(0xff02);
1328                 in6.s6_addr32[1] = 0;
1329                 in6.s6_addr32[2] = htonl(1);
1330                 in6.s6_addr32[3] = 0;
1331                 in6.s6_addr8[12] = 0xff;
1332                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1333                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1334                         free(im6o->im6o_membership, M_CARP);
1335                         break;
1336                 }
1337                 in6m = NULL;
1338                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1339                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1340                         free(im6o->im6o_membership, M_CARP);
1341                         break;
1342                 }
1343                 im6o->im6o_membership[1] = in6m;
1344                 im6o->im6o_num_memberships++;
1345                 break;
1346             }
1347 #endif
1348         }
1349
1350         return (error);
1351 }
1352
1353 /*
1354  * Free multicast structures.
1355  */
1356 static void
1357 carp_multicast_cleanup(struct carp_if *cif, sa_family_t sa)
1358 {
1359
1360         CIF_LOCK_ASSERT(cif);
1361         switch (sa) {
1362 #ifdef INET
1363         case AF_INET:
1364                 if (cif->cif_naddrs == 0) {
1365                         struct ip_moptions *imo = &cif->cif_imo;
1366
1367                         in_leavegroup(imo->imo_membership[0], NULL);
1368                         KASSERT(imo->imo_mfilters == NULL,
1369                             ("%s: imo_mfilters != NULL", __func__));
1370                         free(imo->imo_membership, M_CARP);
1371                         imo->imo_membership = NULL;
1372
1373                 }
1374                 break;
1375 #endif
1376 #ifdef INET6
1377         case AF_INET6:
1378                 if (cif->cif_naddrs6 == 0) {
1379                         struct ip6_moptions *im6o = &cif->cif_im6o;
1380
1381                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1382                         in6_mc_leave(im6o->im6o_membership[1], NULL);
1383                         KASSERT(im6o->im6o_mfilters == NULL,
1384                             ("%s: im6o_mfilters != NULL", __func__));
1385                         free(im6o->im6o_membership, M_CARP);
1386                         im6o->im6o_membership = NULL;
1387                 }
1388                 break;
1389 #endif
1390         }
1391 }
1392
1393 int
1394 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa)
1395 {
1396         struct m_tag *mtag;
1397         struct carp_softc *sc;
1398
1399         if (!sa)
1400                 return (0);
1401
1402         switch (sa->sa_family) {
1403 #ifdef INET
1404         case AF_INET:
1405                 break;
1406 #endif
1407 #ifdef INET6
1408         case AF_INET6:
1409                 break;
1410 #endif
1411         default:
1412                 return (0);
1413         }
1414
1415         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
1416         if (mtag == NULL)
1417                 return (0);
1418
1419         bcopy(mtag + 1, &sc, sizeof(sc));
1420
1421         /* Set the source MAC address to the Virtual Router MAC Address. */
1422         switch (ifp->if_type) {
1423         case IFT_ETHER:
1424         case IFT_BRIDGE:
1425         case IFT_L2VLAN: {
1426                         struct ether_header *eh;
1427
1428                         eh = mtod(m, struct ether_header *);
1429                         eh->ether_shost[0] = 0;
1430                         eh->ether_shost[1] = 0;
1431                         eh->ether_shost[2] = 0x5e;
1432                         eh->ether_shost[3] = 0;
1433                         eh->ether_shost[4] = 1;
1434                         eh->ether_shost[5] = sc->sc_vhid;
1435                 }
1436                 break;
1437         case IFT_FDDI: {
1438                         struct fddi_header *fh;
1439
1440                         fh = mtod(m, struct fddi_header *);
1441                         fh->fddi_shost[0] = 0;
1442                         fh->fddi_shost[1] = 0;
1443                         fh->fddi_shost[2] = 0x5e;
1444                         fh->fddi_shost[3] = 0;
1445                         fh->fddi_shost[4] = 1;
1446                         fh->fddi_shost[5] = sc->sc_vhid;
1447                 }
1448                 break;
1449         case IFT_ISO88025: {
1450                         struct iso88025_header *th;
1451                         th = mtod(m, struct iso88025_header *);
1452                         th->iso88025_shost[0] = 3;
1453                         th->iso88025_shost[1] = 0;
1454                         th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
1455                         th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
1456                         th->iso88025_shost[4] = 0;
1457                         th->iso88025_shost[5] = 0;
1458                 }
1459                 break;
1460         default:
1461                 printf("%s: carp is not supported for the %d interface type\n",
1462                     ifp->if_xname, ifp->if_type);
1463                 return (EOPNOTSUPP);
1464         }
1465
1466         return (0);
1467 }
1468
1469 static struct carp_softc*
1470 carp_alloc(struct ifnet *ifp)
1471 {
1472         struct carp_softc *sc;
1473         struct carp_if *cif;
1474
1475         if ((cif = ifp->if_carp) == NULL)
1476                 cif = carp_alloc_if(ifp);
1477
1478         sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
1479
1480         sc->sc_advbase = CARP_DFLTINTV;
1481         sc->sc_vhid = -1;       /* required setting */
1482         sc->sc_init_counter = 1;
1483         sc->sc_state = INIT;
1484
1485         sc->sc_ifasiz = sizeof(struct ifaddr *);
1486         sc->sc_ifas = malloc(sc->sc_ifasiz, M_CARP, M_WAITOK|M_ZERO);
1487         sc->sc_carpdev = ifp;
1488
1489         CARP_LOCK_INIT(sc);
1490 #ifdef INET
1491         callout_init_mtx(&sc->sc_md_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1492 #endif
1493 #ifdef INET6
1494         callout_init_mtx(&sc->sc_md6_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1495 #endif
1496         callout_init_mtx(&sc->sc_ad_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1497
1498         CIF_LOCK(cif);
1499         TAILQ_INSERT_TAIL(&cif->cif_vrs, sc, sc_list);
1500         CIF_UNLOCK(cif);
1501
1502         mtx_lock(&carp_mtx);
1503         LIST_INSERT_HEAD(&carp_list, sc, sc_next);
1504         mtx_unlock(&carp_mtx);
1505
1506         return (sc);
1507 }
1508
1509 static int
1510 carp_grow_ifas(struct carp_softc *sc)
1511 {
1512         struct ifaddr **new;
1513
1514         CARP_LOCK_ASSERT(sc);
1515
1516         new = malloc(sc->sc_ifasiz * 2, M_CARP, M_NOWAIT|M_ZERO);
1517         if (new == NULL)
1518                 return (ENOMEM);
1519         bcopy(sc->sc_ifas, new, sc->sc_ifasiz);
1520         free(sc->sc_ifas, M_CARP);
1521         sc->sc_ifas = new;
1522         sc->sc_ifasiz *= 2;
1523
1524         return (0);
1525 }
1526
1527 static void
1528 carp_destroy(struct carp_softc *sc)
1529 {
1530         struct ifnet *ifp = sc->sc_carpdev;
1531         struct carp_if *cif = ifp->if_carp;
1532
1533         CIF_LOCK_ASSERT(cif);
1534
1535         TAILQ_REMOVE(&cif->cif_vrs, sc, sc_list);
1536
1537         mtx_lock(&carp_mtx);
1538         LIST_REMOVE(sc, sc_next);
1539         mtx_unlock(&carp_mtx);
1540
1541         CARP_LOCK(sc);
1542         if (sc->sc_suppress)
1543                 carp_demote_adj(-V_carp_ifdown_adj, "vhid removed");
1544         callout_drain(&sc->sc_ad_tmo);
1545 #ifdef INET
1546         callout_drain(&sc->sc_md_tmo);
1547 #endif
1548 #ifdef INET6
1549         callout_drain(&sc->sc_md6_tmo);
1550 #endif
1551         CARP_LOCK_DESTROY(sc);
1552
1553         free(sc->sc_ifas, M_CARP);
1554         free(sc, M_CARP);
1555 }
1556
1557 static struct carp_if*
1558 carp_alloc_if(struct ifnet *ifp)
1559 {
1560         struct carp_if *cif;
1561         int error;
1562
1563         cif = malloc(sizeof(*cif), M_CARP, M_WAITOK|M_ZERO);
1564
1565         if ((error = ifpromisc(ifp, 1)) != 0)
1566                 printf("%s: ifpromisc(%s) failed: %d\n",
1567                     __func__, ifp->if_xname, error);
1568         else
1569                 cif->cif_flags |= CIF_PROMISC;
1570
1571         CIF_LOCK_INIT(cif);
1572         cif->cif_ifp = ifp;
1573         TAILQ_INIT(&cif->cif_vrs);
1574
1575         IF_ADDR_WLOCK(ifp);
1576         ifp->if_carp = cif;
1577         if_ref(ifp);
1578         IF_ADDR_WUNLOCK(ifp);
1579
1580         return (cif);
1581 }
1582
1583 static void
1584 carp_free_if(struct carp_if *cif)
1585 {
1586         struct ifnet *ifp = cif->cif_ifp;
1587
1588         CIF_LOCK_ASSERT(cif);
1589         KASSERT(TAILQ_EMPTY(&cif->cif_vrs), ("%s: softc list not empty",
1590             __func__));
1591
1592         IF_ADDR_WLOCK(ifp);
1593         ifp->if_carp = NULL;
1594         IF_ADDR_WUNLOCK(ifp);
1595
1596         CIF_LOCK_DESTROY(cif);
1597
1598         if (cif->cif_flags & CIF_PROMISC)
1599                 ifpromisc(ifp, 0);
1600         if_rele(ifp);
1601
1602         free(cif, M_CARP);
1603 }
1604
1605 static void
1606 carp_carprcp(struct carpreq *carpr, struct carp_softc *sc, int priv)
1607 {
1608
1609         CARP_LOCK(sc);
1610         carpr->carpr_state = sc->sc_state;
1611         carpr->carpr_vhid = sc->sc_vhid;
1612         carpr->carpr_advbase = sc->sc_advbase;
1613         carpr->carpr_advskew = sc->sc_advskew;
1614         if (priv)
1615                 bcopy(sc->sc_key, carpr->carpr_key, sizeof(carpr->carpr_key));
1616         else
1617                 bzero(carpr->carpr_key, sizeof(carpr->carpr_key));
1618         CARP_UNLOCK(sc);
1619 }
1620
1621 int
1622 carp_ioctl(struct ifreq *ifr, u_long cmd, struct thread *td)
1623 {
1624         struct carpreq carpr;
1625         struct ifnet *ifp;
1626         struct carp_softc *sc = NULL;
1627         int error = 0, locked = 0;
1628
1629         if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1630                 return (error);
1631
1632         ifp = ifunit_ref(ifr->ifr_name);
1633         if (ifp == NULL)
1634                 return (ENXIO);
1635
1636         switch (ifp->if_type) {
1637         case IFT_ETHER:
1638         case IFT_L2VLAN:
1639         case IFT_BRIDGE:
1640         case IFT_FDDI:
1641         case IFT_ISO88025:
1642                 break;
1643         default:
1644                 error = EOPNOTSUPP;
1645                 goto out;
1646         }
1647
1648         if ((ifp->if_flags & IFF_MULTICAST) == 0) {
1649                 error = EADDRNOTAVAIL;
1650                 goto out;
1651         }
1652
1653         switch (cmd) {
1654         case SIOCSVH:
1655                 if ((error = priv_check(td, PRIV_NETINET_CARP)))
1656                         break;
1657                 if (carpr.carpr_vhid <= 0 || carpr.carpr_vhid > CARP_MAXVHID ||
1658                     carpr.carpr_advbase < 0 || carpr.carpr_advskew < 0) {
1659                         error = EINVAL;
1660                         break;
1661                 }
1662
1663                 if (ifp->if_carp) {
1664                         CIF_LOCK(ifp->if_carp);
1665                         IFNET_FOREACH_CARP(ifp, sc)
1666                                 if (sc->sc_vhid == carpr.carpr_vhid)
1667                                         break;
1668                         CIF_UNLOCK(ifp->if_carp);
1669                 }
1670                 if (sc == NULL) {
1671                         sc = carp_alloc(ifp);
1672                         CARP_LOCK(sc);
1673                         sc->sc_vhid = carpr.carpr_vhid;
1674                         LLADDR(&sc->sc_addr)[0] = 0;
1675                         LLADDR(&sc->sc_addr)[1] = 0;
1676                         LLADDR(&sc->sc_addr)[2] = 0x5e;
1677                         LLADDR(&sc->sc_addr)[3] = 0;
1678                         LLADDR(&sc->sc_addr)[4] = 1;
1679                         LLADDR(&sc->sc_addr)[5] = sc->sc_vhid;
1680                 } else
1681                         CARP_LOCK(sc);
1682                 locked = 1;
1683                 if (carpr.carpr_advbase > 0) {
1684                         if (carpr.carpr_advbase > 255 ||
1685                             carpr.carpr_advbase < CARP_DFLTINTV) {
1686                                 error = EINVAL;
1687                                 break;
1688                         }
1689                         sc->sc_advbase = carpr.carpr_advbase;
1690                 }
1691                 if (carpr.carpr_advskew >= 255) {
1692                         error = EINVAL;
1693                         break;
1694                 }
1695                 sc->sc_advskew = carpr.carpr_advskew;
1696                 if (carpr.carpr_key[0] != '\0') {
1697                         bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1698                         carp_hmac_prepare(sc);
1699                 }
1700                 if (sc->sc_state != INIT &&
1701                     carpr.carpr_state != sc->sc_state) {
1702                         switch (carpr.carpr_state) {
1703                         case BACKUP:
1704                                 callout_stop(&sc->sc_ad_tmo);
1705                                 carp_set_state(sc, BACKUP,
1706                                     "user requested via ifconfig");
1707                                 carp_setrun(sc, 0);
1708                                 carp_delroute(sc);
1709                                 break;
1710                         case MASTER:
1711                                 carp_master_down_locked(sc,
1712                                     "user requested via ifconfig");
1713                                 break;
1714                         default:
1715                                 break;
1716                         }
1717                 }
1718                 break;
1719
1720         case SIOCGVH:
1721             {
1722                 int priveleged;
1723
1724                 if (carpr.carpr_vhid < 0 || carpr.carpr_vhid > CARP_MAXVHID) {
1725                         error = EINVAL;
1726                         break;
1727                 }
1728                 if (carpr.carpr_count < 1) {
1729                         error = EMSGSIZE;
1730                         break;
1731                 }
1732                 if (ifp->if_carp == NULL) {
1733                         error = ENOENT;
1734                         break;
1735                 }
1736
1737                 priveleged = (priv_check(td, PRIV_NETINET_CARP) == 0);
1738                 if (carpr.carpr_vhid != 0) {
1739                         CIF_LOCK(ifp->if_carp);
1740                         IFNET_FOREACH_CARP(ifp, sc)
1741                                 if (sc->sc_vhid == carpr.carpr_vhid)
1742                                         break;
1743                         CIF_UNLOCK(ifp->if_carp);
1744                         if (sc == NULL) {
1745                                 error = ENOENT;
1746                                 break;
1747                         }
1748                         carp_carprcp(&carpr, sc, priveleged);
1749                         error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1750                 } else  {
1751                         int i, count;
1752
1753                         count = 0;
1754                         CIF_LOCK(ifp->if_carp);
1755                         IFNET_FOREACH_CARP(ifp, sc)
1756                                 count++;
1757
1758                         if (count > carpr.carpr_count) {
1759                                 CIF_UNLOCK(ifp->if_carp);
1760                                 error = EMSGSIZE;
1761                                 break;
1762                         }
1763
1764                         i = 0;
1765                         IFNET_FOREACH_CARP(ifp, sc) {
1766                                 carp_carprcp(&carpr, sc, priveleged);
1767                                 carpr.carpr_count = count;
1768                                 error = copyout(&carpr, ifr->ifr_data +
1769                                     (i * sizeof(carpr)), sizeof(carpr));
1770                                 if (error) {
1771                                         CIF_UNLOCK(ifp->if_carp);
1772                                         break;
1773                                 }
1774                                 i++;
1775                         }
1776                         CIF_UNLOCK(ifp->if_carp);
1777                 }
1778                 break;
1779             }
1780         default:
1781                 error = EINVAL;
1782         }
1783
1784 out:
1785         if (locked)
1786                 CARP_UNLOCK(sc);
1787         if_rele(ifp);
1788
1789         return (error);
1790 }
1791
1792 static int
1793 carp_get_vhid(struct ifaddr *ifa)
1794 {
1795
1796         if (ifa == NULL || ifa->ifa_carp == NULL)
1797                 return (0);
1798
1799         return (ifa->ifa_carp->sc_vhid);
1800 }
1801
1802 int
1803 carp_attach(struct ifaddr *ifa, int vhid)
1804 {
1805         struct ifnet *ifp = ifa->ifa_ifp;
1806         struct carp_if *cif = ifp->if_carp;
1807         struct carp_softc *sc;
1808         int index, error;
1809
1810         if (ifp->if_carp == NULL)
1811                 return (ENOPROTOOPT);
1812
1813         switch (ifa->ifa_addr->sa_family) {
1814 #ifdef INET
1815         case AF_INET:
1816 #endif
1817 #ifdef INET6
1818         case AF_INET6:
1819 #endif
1820                 break;
1821         default:
1822                 return (EPROTOTYPE);
1823         }
1824
1825         CIF_LOCK(cif);
1826         IFNET_FOREACH_CARP(ifp, sc)
1827                 if (sc->sc_vhid == vhid)
1828                         break;
1829         if (sc == NULL) {
1830                 CIF_UNLOCK(cif);
1831                 return (ENOENT);
1832         }
1833
1834         if (ifa->ifa_carp) {
1835                 if (ifa->ifa_carp->sc_vhid != vhid)
1836                         carp_detach_locked(ifa);
1837                 else {
1838                         CIF_UNLOCK(cif);
1839                         return (0);
1840                 }
1841         }
1842
1843         error = carp_multicast_setup(cif, ifa->ifa_addr->sa_family);
1844         if (error) {
1845                 CIF_FREE(cif);
1846                 return (error);
1847         }
1848
1849         CARP_LOCK(sc);
1850         index = sc->sc_naddrs + sc->sc_naddrs6 + 1;
1851         if (index > sc->sc_ifasiz / sizeof(struct ifaddr *))
1852                 if ((error = carp_grow_ifas(sc)) != 0) {
1853                         carp_multicast_cleanup(cif,
1854                             ifa->ifa_addr->sa_family);
1855                         CARP_UNLOCK(sc);
1856                         CIF_FREE(cif);
1857                         return (error);
1858                 }
1859
1860         switch (ifa->ifa_addr->sa_family) {
1861 #ifdef INET
1862         case AF_INET:
1863                 cif->cif_naddrs++;
1864                 sc->sc_naddrs++;
1865                 break;
1866 #endif
1867 #ifdef INET6
1868         case AF_INET6:
1869                 cif->cif_naddrs6++;
1870                 sc->sc_naddrs6++;
1871                 break;
1872 #endif
1873         }
1874
1875         ifa_ref(ifa);
1876         sc->sc_ifas[index - 1] = ifa;
1877         ifa->ifa_carp = sc;
1878
1879         carp_hmac_prepare(sc);
1880         carp_sc_state(sc);
1881
1882         CARP_UNLOCK(sc);
1883         CIF_UNLOCK(cif);
1884
1885         return (0);
1886 }
1887
1888 void
1889 carp_detach(struct ifaddr *ifa)
1890 {
1891         struct ifnet *ifp = ifa->ifa_ifp;
1892         struct carp_if *cif = ifp->if_carp;
1893
1894         CIF_LOCK(cif);
1895         carp_detach_locked(ifa);
1896         CIF_FREE(cif);
1897 }
1898
1899 static void
1900 carp_detach_locked(struct ifaddr *ifa)
1901 {
1902         struct ifnet *ifp = ifa->ifa_ifp;
1903         struct carp_if *cif = ifp->if_carp;
1904         struct carp_softc *sc = ifa->ifa_carp;
1905         int i, index;
1906
1907         KASSERT(sc != NULL, ("%s: %p not attached", __func__, ifa));
1908
1909         CIF_LOCK_ASSERT(cif);
1910         CARP_LOCK(sc);
1911
1912         /* Shift array. */
1913         index = sc->sc_naddrs + sc->sc_naddrs6;
1914         for (i = 0; i < index; i++)
1915                 if (sc->sc_ifas[i] == ifa)
1916                         break;
1917         KASSERT(i < index, ("%s: %p no backref", __func__, ifa));
1918         for (; i < index - 1; i++)
1919                 sc->sc_ifas[i] = sc->sc_ifas[i+1];
1920         sc->sc_ifas[index - 1] = NULL;
1921
1922         switch (ifa->ifa_addr->sa_family) {
1923 #ifdef INET
1924         case AF_INET:
1925                 cif->cif_naddrs--;
1926                 sc->sc_naddrs--;
1927                 break;
1928 #endif
1929 #ifdef INET6
1930         case AF_INET6:
1931                 cif->cif_naddrs6--;
1932                 sc->sc_naddrs6--;
1933                 break;
1934 #endif
1935         }
1936
1937         carp_ifa_delroute(ifa);
1938         carp_multicast_cleanup(cif, ifa->ifa_addr->sa_family);
1939
1940         ifa->ifa_carp = NULL;
1941         ifa_free(ifa);
1942
1943         carp_hmac_prepare(sc);
1944         carp_sc_state(sc);
1945
1946         if (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) {
1947                 CARP_UNLOCK(sc);
1948                 carp_destroy(sc);
1949         } else
1950                 CARP_UNLOCK(sc);
1951 }
1952
1953 static void
1954 carp_set_state(struct carp_softc *sc, int state, const char *reason)
1955 {
1956
1957         CARP_LOCK_ASSERT(sc);
1958
1959         if (sc->sc_state != state) {
1960                 const char *carp_states[] = { CARP_STATES };
1961                 char subsys[IFNAMSIZ+5];
1962
1963                 snprintf(subsys, IFNAMSIZ+5, "%u@%s", sc->sc_vhid,
1964                     sc->sc_carpdev->if_xname);
1965
1966                 CARP_LOG("%s: %s -> %s (%s)\n", subsys,
1967                     carp_states[sc->sc_state], carp_states[state], reason);
1968
1969                 sc->sc_state = state;
1970
1971                 devctl_notify("CARP", subsys, carp_states[state], NULL);
1972         }
1973 }
1974
1975 static void
1976 carp_linkstate(struct ifnet *ifp)
1977 {
1978         struct carp_softc *sc;
1979
1980         CIF_LOCK(ifp->if_carp);
1981         IFNET_FOREACH_CARP(ifp, sc) {
1982                 CARP_LOCK(sc);
1983                 carp_sc_state(sc);
1984                 CARP_UNLOCK(sc);
1985         }
1986         CIF_UNLOCK(ifp->if_carp);
1987 }
1988
1989 static void
1990 carp_sc_state(struct carp_softc *sc)
1991 {
1992
1993         CARP_LOCK_ASSERT(sc);
1994
1995         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1996             !(sc->sc_carpdev->if_flags & IFF_UP)) {
1997                 callout_stop(&sc->sc_ad_tmo);
1998 #ifdef INET
1999                 callout_stop(&sc->sc_md_tmo);
2000 #endif
2001 #ifdef INET6
2002                 callout_stop(&sc->sc_md6_tmo);
2003 #endif
2004                 carp_set_state(sc, INIT, "hardware interface down");
2005                 carp_setrun(sc, 0);
2006                 if (!sc->sc_suppress)
2007                         carp_demote_adj(V_carp_ifdown_adj, "interface down");
2008                 sc->sc_suppress = 1;
2009         } else {
2010                 carp_set_state(sc, INIT, "hardware interface up");
2011                 carp_setrun(sc, 0);
2012                 if (sc->sc_suppress)
2013                         carp_demote_adj(-V_carp_ifdown_adj, "interface up");
2014                 sc->sc_suppress = 0;
2015         }
2016 }
2017
2018 static void
2019 carp_demote_adj(int adj, char *reason)
2020 {
2021         atomic_add_int(&V_carp_demotion, adj);
2022         CARP_LOG("demoted by %d to %d (%s)\n", adj, V_carp_demotion, reason);
2023         taskqueue_enqueue(taskqueue_swi, &carp_sendall_task);
2024 }
2025
2026 static int
2027 carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS)
2028 {
2029         int new, error;
2030
2031         new = V_carp_demotion;
2032         error = sysctl_handle_int(oidp, &new, 0, req);
2033         if (error || !req->newptr)
2034                 return (error);
2035
2036         carp_demote_adj(new, "sysctl");
2037
2038         return (0);
2039 }
2040
2041 #ifdef INET
2042 extern  struct domain inetdomain;
2043 static struct protosw in_carp_protosw = {
2044         .pr_type =              SOCK_RAW,
2045         .pr_domain =            &inetdomain,
2046         .pr_protocol =          IPPROTO_CARP,
2047         .pr_flags =             PR_ATOMIC|PR_ADDR,
2048         .pr_input =             carp_input,
2049         .pr_output =            rip_output,
2050         .pr_ctloutput =         rip_ctloutput,
2051         .pr_usrreqs =           &rip_usrreqs
2052 };
2053 #endif
2054
2055 #ifdef INET6
2056 extern  struct domain inet6domain;
2057 static struct protosw in6_carp_protosw = {
2058         .pr_type =              SOCK_RAW,
2059         .pr_domain =            &inet6domain,
2060         .pr_protocol =          IPPROTO_CARP,
2061         .pr_flags =             PR_ATOMIC|PR_ADDR,
2062         .pr_input =             carp6_input,
2063         .pr_output =            rip6_output,
2064         .pr_ctloutput =         rip6_ctloutput,
2065         .pr_usrreqs =           &rip6_usrreqs
2066 };
2067 #endif
2068
2069 static void
2070 carp_mod_cleanup(void)
2071 {
2072
2073 #ifdef INET
2074         if (proto_reg[CARP_INET] == 0) {
2075                 (void)ipproto_unregister(IPPROTO_CARP);
2076                 pf_proto_unregister(PF_INET, IPPROTO_CARP, SOCK_RAW);
2077                 proto_reg[CARP_INET] = -1;
2078         }
2079         carp_iamatch_p = NULL;
2080 #endif
2081 #ifdef INET6
2082         if (proto_reg[CARP_INET6] == 0) {
2083                 (void)ip6proto_unregister(IPPROTO_CARP);
2084                 pf_proto_unregister(PF_INET6, IPPROTO_CARP, SOCK_RAW);
2085                 proto_reg[CARP_INET6] = -1;
2086         }
2087         carp_iamatch6_p = NULL;
2088         carp_macmatch6_p = NULL;
2089 #endif
2090         carp_ioctl_p = NULL;
2091         carp_attach_p = NULL;
2092         carp_detach_p = NULL;
2093         carp_get_vhid_p = NULL;
2094         carp_linkstate_p = NULL;
2095         carp_forus_p = NULL;
2096         carp_output_p = NULL;
2097         carp_demote_adj_p = NULL;
2098         carp_master_p = NULL;
2099         mtx_unlock(&carp_mtx);
2100         taskqueue_drain(taskqueue_swi, &carp_sendall_task);
2101         mtx_destroy(&carp_mtx);
2102 }
2103
2104 static int
2105 carp_mod_load(void)
2106 {
2107         int err;
2108
2109         mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2110         LIST_INIT(&carp_list);
2111         carp_get_vhid_p = carp_get_vhid;
2112         carp_forus_p = carp_forus;
2113         carp_output_p = carp_output;
2114         carp_linkstate_p = carp_linkstate;
2115         carp_ioctl_p = carp_ioctl;
2116         carp_attach_p = carp_attach;
2117         carp_detach_p = carp_detach;
2118         carp_demote_adj_p = carp_demote_adj;
2119         carp_master_p = carp_master;
2120 #ifdef INET6
2121         carp_iamatch6_p = carp_iamatch6;
2122         carp_macmatch6_p = carp_macmatch6;
2123         proto_reg[CARP_INET6] = pf_proto_register(PF_INET6,
2124             (struct protosw *)&in6_carp_protosw);
2125         if (proto_reg[CARP_INET6]) {
2126                 printf("carp: error %d attaching to PF_INET6\n",
2127                     proto_reg[CARP_INET6]);
2128                 carp_mod_cleanup();
2129                 return (proto_reg[CARP_INET6]);
2130         }
2131         err = ip6proto_register(IPPROTO_CARP);
2132         if (err) {
2133                 printf("carp: error %d registering with INET6\n", err);
2134                 carp_mod_cleanup();
2135                 return (err);
2136         }
2137 #endif
2138 #ifdef INET
2139         carp_iamatch_p = carp_iamatch;
2140         proto_reg[CARP_INET] = pf_proto_register(PF_INET, &in_carp_protosw);
2141         if (proto_reg[CARP_INET]) {
2142                 printf("carp: error %d attaching to PF_INET\n",
2143                     proto_reg[CARP_INET]);
2144                 carp_mod_cleanup();
2145                 return (proto_reg[CARP_INET]);
2146         }
2147         err = ipproto_register(IPPROTO_CARP);
2148         if (err) {
2149                 printf("carp: error %d registering with INET\n", err);
2150                 carp_mod_cleanup();
2151                 return (err);
2152         }
2153 #endif
2154         return (0);
2155 }
2156
2157 static int
2158 carp_modevent(module_t mod, int type, void *data)
2159 {
2160         switch (type) {
2161         case MOD_LOAD:
2162                 return carp_mod_load();
2163                 /* NOTREACHED */
2164         case MOD_UNLOAD:
2165                 mtx_lock(&carp_mtx);
2166                 if (LIST_EMPTY(&carp_list))
2167                         carp_mod_cleanup();
2168                 else {
2169                         mtx_unlock(&carp_mtx);
2170                         return (EBUSY);
2171                 }
2172                 break;
2173
2174         default:
2175                 return (EINVAL);
2176         }
2177
2178         return (0);
2179 }
2180
2181 static moduledata_t carp_mod = {
2182         "carp",
2183         carp_modevent,
2184         0
2185 };
2186
2187 DECLARE_MODULE(carp, carp_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);