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/ip6protosw.h>
  82 #include <netinet6/in6_var.h>
  83 #include <netinet6/ip6_var.h>
  84 #include <netinet6/scope6_var.h>
  85 #include <netinet6/nd6.h>
  86 #endif
  87
  88 #include <crypto/sha1.h>
  89
  90 static MALLOC_DEFINE(M_CARP, "CARP", "CARP addresses");
  91
  92 struct carp_softc {
  93         struct ifnet            *sc_carpdev;    /* Pointer to parent ifnet. */
  94         struct ifaddr           **sc_ifas;      /* Our ifaddrs. */
  95         struct sockaddr_dl      sc_addr;        /* Our link level address. */
  96         struct callout          sc_ad_tmo;      /* Advertising timeout. */
  97 #ifdef INET
  98         struct callout          sc_md_tmo;      /* Master down timeout. */
  99 #endif
 100 #ifdef INET6
 101         struct callout          sc_md6_tmo;     /* XXX: Master down timeout. */
 102 #endif
 103         struct mtx              sc_mtx;
 104
 105         int                     sc_vhid;
 106         int                     sc_advskew;
 107         int                     sc_advbase;
 108
 109         int                     sc_naddrs;
 110         int                     sc_naddrs6;
 111         int                     sc_ifasiz;
 112         enum { INIT = 0, BACKUP, MASTER }       sc_state;
 113         int                     sc_suppress;
 114         int                     sc_sendad_errors;
 115 #define CARP_SENDAD_MAX_ERRORS  3
 116         int                     sc_sendad_success;
 117 #define CARP_SENDAD_MIN_SUCCESS 3
 118
 119         int                     sc_init_counter;
 120         uint64_t                sc_counter;
 121
 122         /* authentication */
 123 #define CARP_HMAC_PAD   64
 124         unsigned char sc_key[CARP_KEY_LEN];
 125         unsigned char sc_pad[CARP_HMAC_PAD];
 126         SHA1_CTX sc_sha1;
 127
 128         TAILQ_ENTRY(carp_softc) sc_list;        /* On the carp_if list. */
 129         LIST_ENTRY(carp_softc)  sc_next;        /* On the global list. */
 130 };
 131
 132 struct carp_if {
 133 #ifdef INET
 134         int     cif_naddrs;
 135 #endif
 136 #ifdef INET6
 137         int     cif_naddrs6;
 138 #endif
 139         TAILQ_HEAD(, carp_softc) cif_vrs;
 140 #ifdef INET
 141         struct ip_moptions       cif_imo;
 142 #endif
 143 #ifdef INET6
 144         struct ip6_moptions      cif_im6o;
 145 #endif
 146         struct ifnet    *cif_ifp;
 147         struct mtx      cif_mtx;
 148 };
 149
 150 #define CARP_INET       0
 151 #define CARP_INET6      1
 152 static int proto_reg[] = {-1, -1};
 153
 154 /*
 155  * Brief design of carp(4).
 156  *
 157  * Any carp-capable ifnet may have a list of carp softcs hanging off
 158  * its ifp->if_carp pointer. Each softc represents one unique virtual
 159  * host id, or vhid. The softc has a back pointer to the ifnet. All
 160  * softcs are joined in a global list, which has quite limited use.
 161  *
 162  * Any interface address that takes part in CARP negotiation has a
 163  * pointer to the softc of its vhid, ifa->ifa_carp. That could be either
 164  * AF_INET or AF_INET6 address.
 165  *
 166  * Although, one can get the softc's backpointer to ifnet and traverse
 167  * through its ifp->if_addrhead queue to find all interface addresses
 168  * involved in CARP, we keep a growable array of ifaddr pointers. This
 169  * allows us to avoid grabbing the IF_ADDR_LOCK() in many traversals that
 170  * do calls into the network stack, thus avoiding LORs.
 171  *
 172  * Locking:
 173  *
 174  * Each softc has a lock sc_mtx. It is used to synchronise carp_input_c(),
 175  * callout-driven events and ioctl()s.
 176  *
 177  * To traverse the list of softcs on an ifnet we use CIF_LOCK(), to
 178  * traverse the global list we use the mutex carp_mtx.
 179  *
 180  * Known issues with locking:
 181  *
 182  * - There is no protection for races between two ioctl() requests,
 183  *   neither SIOCSVH, nor SIOCAIFADDR & SIOCAIFADDR_IN6. I think that all
 184  *   interface ioctl()s should be serialized right in net/if.c.
 185  * - Sending ad, we put the pointer to the softc in an mtag, and no reference
 186  *   counting is done on the softc.
 187  * - On module unload we may race (?) with packet processing thread
 188  *   dereferencing our function pointers.
 189  */
 190
 191 /* Accept incoming CARP packets. */
 192 static VNET_DEFINE(int, carp_allow) = 1;
 193 #define V_carp_allow    VNET(carp_allow)
 194
 195 /* Preempt slower nodes. */
 196 static VNET_DEFINE(int, carp_preempt) = 0;
 197 #define V_carp_preempt  VNET(carp_preempt)
 198
 199 /* Log level. */
 200 static VNET_DEFINE(int, carp_log) = 1;
 201 #define V_carp_log      VNET(carp_log)
 202
 203 /* Global advskew demotion. */
 204 static VNET_DEFINE(int, carp_demotion) = 0;
 205 #define V_carp_demotion VNET(carp_demotion)
 206
 207 /* Send error demotion factor. */
 208 static VNET_DEFINE(int, carp_senderr_adj) = CARP_MAXSKEW;
 209 #define V_carp_senderr_adj      VNET(carp_senderr_adj)
 210
 211 /* Iface down demotion factor. */
 212 static VNET_DEFINE(int, carp_ifdown_adj) = CARP_MAXSKEW;
 213 #define V_carp_ifdown_adj       VNET(carp_ifdown_adj)
 214
 215 static int carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS);
 216
 217 SYSCTL_NODE(_net_inet, IPPROTO_CARP,    carp,   CTLFLAG_RW, 0,  "CARP");
 218 SYSCTL_VNET_INT(_net_inet_carp, OID_AUTO, allow, CTLFLAG_RW,
 219     &VNET_NAME(carp_allow), 0, "Accept incoming CARP packets");
 220 SYSCTL_VNET_INT(_net_inet_carp, OID_AUTO, preempt, CTLFLAG_RW,
 221     &VNET_NAME(carp_preempt), 0, "High-priority backup preemption mode");
 222 SYSCTL_VNET_INT(_net_inet_carp, OID_AUTO, log, CTLFLAG_RW,
 223     &VNET_NAME(carp_log), 0, "CARP log level");
 224 SYSCTL_VNET_PROC(_net_inet_carp, OID_AUTO, demotion, CTLTYPE_INT|CTLFLAG_RW,
 225     0, 0, carp_demote_adj_sysctl, "I",
 226     "Adjust demotion factor (skew of advskew)");
 227 SYSCTL_VNET_INT(_net_inet_carp, OID_AUTO, senderr_demotion_factor, CTLFLAG_RW,
 228     &VNET_NAME(carp_senderr_adj), 0, "Send error demotion factor adjustment");
 229 SYSCTL_VNET_INT(_net_inet_carp, OID_AUTO, ifdown_demotion_factor, CTLFLAG_RW,
 230     &VNET_NAME(carp_ifdown_adj), 0,
 231     "Interface down demotion factor adjustment");
 232
 233 VNET_PCPUSTAT_DEFINE(struct carpstats, carpstats);
 234 VNET_PCPUSTAT_SYSINIT(carpstats);
 235 VNET_PCPUSTAT_SYSUNINIT(carpstats);
 236
 237 #define CARPSTATS_ADD(name, val)        \
 238     counter_u64_add(VNET(carpstats)[offsetof(struct carpstats, name) / \
 239         sizeof(uint64_t)], (val))
 240 #define CARPSTATS_INC(name)             CARPSTATS_ADD(name, 1)
 241
 242 SYSCTL_VNET_PCPUSTAT(_net_inet_carp, OID_AUTO, stats, struct carpstats,
 243     carpstats, "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 244
 245 #define CARP_LOCK_INIT(sc)      mtx_init(&(sc)->sc_mtx, "carp_softc",   \
 246         NULL, MTX_DEF)
 247 #define CARP_LOCK_DESTROY(sc)   mtx_destroy(&(sc)->sc_mtx)
 248 #define CARP_LOCK_ASSERT(sc)    mtx_assert(&(sc)->sc_mtx, MA_OWNED)
 249 #define CARP_LOCK(sc)           mtx_lock(&(sc)->sc_mtx)
 250 #define CARP_UNLOCK(sc)         mtx_unlock(&(sc)->sc_mtx)
 251 #define CIF_LOCK_INIT(cif)      mtx_init(&(cif)->cif_mtx, "carp_if",   \
 252         NULL, MTX_DEF)
 253 #define CIF_LOCK_DESTROY(cif)   mtx_destroy(&(cif)->cif_mtx)
 254 #define CIF_LOCK_ASSERT(cif)    mtx_assert(&(cif)->cif_mtx, MA_OWNED)
 255 #define CIF_LOCK(cif)           mtx_lock(&(cif)->cif_mtx)
 256 #define CIF_UNLOCK(cif)         mtx_unlock(&(cif)->cif_mtx)
 257 #define CIF_FREE(cif)   do {                            \
 258                 CIF_LOCK_ASSERT(cif);                   \
 259                 if (TAILQ_EMPTY(&(cif)->cif_vrs))       \
 260                         carp_free_if(cif);              \
 261                 else                                    \
 262                         CIF_UNLOCK(cif);                \
 263 } while (0)
 264
 265 #define CARP_LOG(...)   do {                            \
 266         if (V_carp_log > 0)                             \
 267                 log(LOG_INFO, "carp: " __VA_ARGS__);    \
 268 } while (0)
 269
 270 #define CARP_DEBUG(...) do {                            \
 271         if (V_carp_log > 1)                             \
 272                 log(LOG_DEBUG, __VA_ARGS__);            \
 273 } while (0)
 274
 275 #define IFNET_FOREACH_IFA(ifp, ifa)                                     \
 276         IF_ADDR_LOCK_ASSERT(ifp);                                       \
 277         TAILQ_FOREACH((ifa), &(ifp)->if_addrhead, ifa_link)             \
 278                 if ((ifa)->ifa_carp != NULL)
 279
 280 #define CARP_FOREACH_IFA(sc, ifa)                                       \
 281         CARP_LOCK_ASSERT(sc);                                           \
 282         for (int _i = 0;                                                \
 283                 _i < (sc)->sc_naddrs + (sc)->sc_naddrs6 &&              \
 284                 ((ifa) = sc->sc_ifas[_i]) != NULL;                      \
 285                 ++_i)
 286
 287 #define IFNET_FOREACH_CARP(ifp, sc)                                     \
 288         CIF_LOCK_ASSERT(ifp->if_carp);                                  \
 289         TAILQ_FOREACH((sc), &(ifp)->if_carp->cif_vrs, sc_list)
 290
 291 #define DEMOTE_ADVSKEW(sc)                                      \
 292     (((sc)->sc_advskew + V_carp_demotion > CARP_MAXSKEW) ?      \
 293     CARP_MAXSKEW : ((sc)->sc_advskew + V_carp_demotion))
 294
 295 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 296 static struct carp_softc
 297                 *carp_alloc(struct ifnet *);
 298 static void     carp_detach_locked(struct ifaddr *);
 299 static void     carp_destroy(struct carp_softc *);
 300 static struct carp_if
 301                 *carp_alloc_if(struct ifnet *);
 302 static void     carp_free_if(struct carp_if *);
 303 static void     carp_set_state(struct carp_softc *, int);
 304 static void     carp_sc_state(struct carp_softc *);
 305 static void     carp_setrun(struct carp_softc *, sa_family_t);
 306 static void     carp_master_down(void *);
 307 static void     carp_master_down_locked(struct carp_softc *);
 308 static void     carp_send_ad(void *);
 309 static void     carp_send_ad_locked(struct carp_softc *);
 310 static void     carp_addroute(struct carp_softc *);
 311 static void     carp_ifa_addroute(struct ifaddr *);
 312 static void     carp_delroute(struct carp_softc *);
 313 static void     carp_ifa_delroute(struct ifaddr *);
 314 static void     carp_send_ad_all(void *, int);
 315 static void     carp_demote_adj(int, char *);
 316
 317 static LIST_HEAD(, carp_softc) carp_list;
 318 static struct mtx carp_mtx;
 319 static struct task carp_sendall_task =
 320     TASK_INITIALIZER(0, carp_send_ad_all, NULL);
 321
 322 static void
 323 carp_hmac_prepare(struct carp_softc *sc)
 324 {
 325         uint8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 326         uint8_t vhid = sc->sc_vhid & 0xff;
 327         struct ifaddr *ifa;
 328         int i, found;
 329 #ifdef INET
 330         struct in_addr last, cur, in;
 331 #endif
 332 #ifdef INET6
 333         struct in6_addr last6, cur6, in6;
 334 #endif
 335
 336         CARP_LOCK_ASSERT(sc);
 337
 338         /* Compute ipad from key. */
 339         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 340         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 341         for (i = 0; i < sizeof(sc->sc_pad); i++)
 342                 sc->sc_pad[i] ^= 0x36;
 343
 344         /* Precompute first part of inner hash. */
 345         SHA1Init(&sc->sc_sha1);
 346         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 347         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 348         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 349         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 350 #ifdef INET
 351         cur.s_addr = 0;
 352         do {
 353                 found = 0;
 354                 last = cur;
 355                 cur.s_addr = 0xffffffff;
 356                 CARP_FOREACH_IFA(sc, ifa) {
 357                         in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 358                         if (ifa->ifa_addr->sa_family == AF_INET &&
 359                             ntohl(in.s_addr) > ntohl(last.s_addr) &&
 360                             ntohl(in.s_addr) < ntohl(cur.s_addr)) {
 361                                 cur.s_addr = in.s_addr;
 362                                 found++;
 363                         }
 364                 }
 365                 if (found)
 366                         SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur));
 367         } while (found);
 368 #endif /* INET */
 369 #ifdef INET6
 370         memset(&cur6, 0, sizeof(cur6));
 371         do {
 372                 found = 0;
 373                 last6 = cur6;
 374                 memset(&cur6, 0xff, sizeof(cur6));
 375                 CARP_FOREACH_IFA(sc, ifa) {
 376                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 377                         if (IN6_IS_SCOPE_EMBED(&in6))
 378                                 in6.s6_addr16[1] = 0;
 379                         if (ifa->ifa_addr->sa_family == AF_INET6 &&
 380                             memcmp(&in6, &last6, sizeof(in6)) > 0 &&
 381                             memcmp(&in6, &cur6, sizeof(in6)) < 0) {
 382                                 cur6 = in6;
 383                                 found++;
 384                         }
 385                 }
 386                 if (found)
 387                         SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6));
 388         } while (found);
 389 #endif /* INET6 */
 390
 391         /* convert ipad to opad */
 392         for (i = 0; i < sizeof(sc->sc_pad); i++)
 393                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 394 }
 395
 396 static void
 397 carp_hmac_generate(struct carp_softc *sc, uint32_t counter[2],
 398     unsigned char md[20])
 399 {
 400         SHA1_CTX sha1ctx;
 401
 402         CARP_LOCK_ASSERT(sc);
 403
 404         /* fetch first half of inner hash */
 405         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 406
 407         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 408         SHA1Final(md, &sha1ctx);
 409
 410         /* outer hash */
 411         SHA1Init(&sha1ctx);
 412         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 413         SHA1Update(&sha1ctx, md, 20);
 414         SHA1Final(md, &sha1ctx);
 415 }
 416
 417 static int
 418 carp_hmac_verify(struct carp_softc *sc, uint32_t counter[2],
 419     unsigned char md[20])
 420 {
 421         unsigned char md2[20];
 422
 423         CARP_LOCK_ASSERT(sc);
 424
 425         carp_hmac_generate(sc, counter, md2);
 426
 427         return (bcmp(md, md2, sizeof(md2)));
 428 }
 429
 430 /*
 431  * process input packet.
 432  * we have rearranged checks order compared to the rfc,
 433  * but it seems more efficient this way or not possible otherwise.
 434  */
 435 #ifdef INET
 436 void
 437 carp_input(struct mbuf *m, int hlen)
 438 {
 439         struct ip *ip = mtod(m, struct ip *);
 440         struct carp_header *ch;
 441         int iplen, len;
 442
 443         CARPSTATS_INC(carps_ipackets);
 444
 445         if (!V_carp_allow) {
 446                 m_freem(m);
 447                 return;
 448         }
 449
 450         /* verify that the IP TTL is 255.  */
 451         if (ip->ip_ttl != CARP_DFLTTL) {
 452                 CARPSTATS_INC(carps_badttl);
 453                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 454                     ip->ip_ttl,
 455                     m->m_pkthdr.rcvif->if_xname);
 456                 m_freem(m);
 457                 return;
 458         }
 459
 460         iplen = ip->ip_hl << 2;
 461
 462         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 463                 CARPSTATS_INC(carps_badlen);
 464                 CARP_DEBUG("%s: received len %zd < sizeof(struct carp_header) "
 465                     "on %s\n", __func__, m->m_len - sizeof(struct ip),
 466                     m->m_pkthdr.rcvif->if_xname);
 467                 m_freem(m);
 468                 return;
 469         }
 470
 471         if (iplen + sizeof(*ch) < m->m_len) {
 472                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 473                         CARPSTATS_INC(carps_hdrops);
 474                         CARP_DEBUG("%s: pullup failed\n", __func__);
 475                         return;
 476                 }
 477                 ip = mtod(m, struct ip *);
 478         }
 479         ch = (struct carp_header *)((char *)ip + iplen);
 480
 481         /*
 482          * verify that the received packet length is
 483          * equal to the CARP header
 484          */
 485         len = iplen + sizeof(*ch);
 486         if (len > m->m_pkthdr.len) {
 487                 CARPSTATS_INC(carps_badlen);
 488                 CARP_DEBUG("%s: packet too short %d on %s\n", __func__,
 489                     m->m_pkthdr.len,
 490                     m->m_pkthdr.rcvif->if_xname);
 491                 m_freem(m);
 492                 return;
 493         }
 494
 495         if ((m = m_pullup(m, len)) == NULL) {
 496                 CARPSTATS_INC(carps_hdrops);
 497                 return;
 498         }
 499         ip = mtod(m, struct ip *);
 500         ch = (struct carp_header *)((char *)ip + iplen);
 501
 502         /* verify the CARP checksum */
 503         m->m_data += iplen;
 504         if (in_cksum(m, len - iplen)) {
 505                 CARPSTATS_INC(carps_badsum);
 506                 CARP_DEBUG("%s: checksum failed on %s\n", __func__,
 507                     m->m_pkthdr.rcvif->if_xname);
 508                 m_freem(m);
 509                 return;
 510         }
 511         m->m_data -= iplen;
 512
 513         carp_input_c(m, ch, AF_INET);
 514 }
 515 #endif
 516
 517 #ifdef INET6
 518 int
 519 carp6_input(struct mbuf **mp, int *offp, int proto)
 520 {
 521         struct mbuf *m = *mp;
 522         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 523         struct carp_header *ch;
 524         u_int len;
 525
 526         CARPSTATS_INC(carps_ipackets6);
 527
 528         if (!V_carp_allow) {
 529                 m_freem(m);
 530                 return (IPPROTO_DONE);
 531         }
 532
 533         /* check if received on a valid carp interface */
 534         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 535                 CARPSTATS_INC(carps_badif);
 536                 CARP_DEBUG("%s: packet received on non-carp interface: %s\n",
 537                     __func__, m->m_pkthdr.rcvif->if_xname);
 538                 m_freem(m);
 539                 return (IPPROTO_DONE);
 540         }
 541
 542         /* verify that the IP TTL is 255 */
 543         if (ip6->ip6_hlim != CARP_DFLTTL) {
 544                 CARPSTATS_INC(carps_badttl);
 545                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 546                     ip6->ip6_hlim, m->m_pkthdr.rcvif->if_xname);
 547                 m_freem(m);
 548                 return (IPPROTO_DONE);
 549         }
 550
 551         /* verify that we have a complete carp packet */
 552         len = m->m_len;
 553         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 554         if (ch == NULL) {
 555                 CARPSTATS_INC(carps_badlen);
 556                 CARP_DEBUG("%s: packet size %u too small\n", __func__, len);
 557                 return (IPPROTO_DONE);
 558         }
 559
 560
 561         /* verify the CARP checksum */
 562         m->m_data += *offp;
 563         if (in_cksum(m, sizeof(*ch))) {
 564                 CARPSTATS_INC(carps_badsum);
 565                 CARP_DEBUG("%s: checksum failed, on %s\n", __func__,
 566                     m->m_pkthdr.rcvif->if_xname);
 567                 m_freem(m);
 568                 return (IPPROTO_DONE);
 569         }
 570         m->m_data -= *offp;
 571
 572         carp_input_c(m, ch, AF_INET6);
 573         return (IPPROTO_DONE);
 574 }
 575 #endif /* INET6 */
 576
 577 static void
 578 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 579 {
 580         struct ifnet *ifp = m->m_pkthdr.rcvif;
 581         struct ifaddr *ifa;
 582         struct carp_softc *sc;
 583         uint64_t tmp_counter;
 584         struct timeval sc_tv, ch_tv;
 585
 586         /* verify that the VHID is valid on the receiving interface */
 587         IF_ADDR_RLOCK(ifp);
 588         IFNET_FOREACH_IFA(ifp, ifa)
 589                 if (ifa->ifa_addr->sa_family == af &&
 590                     ifa->ifa_carp->sc_vhid == ch->carp_vhid) {
 591                         ifa_ref(ifa);
 592                         break;
 593                 }
 594         IF_ADDR_RUNLOCK(ifp);
 595
 596         if (ifa == NULL) {
 597                 CARPSTATS_INC(carps_badvhid);
 598                 m_freem(m);
 599                 return;
 600         }
 601
 602         /* verify the CARP version. */
 603         if (ch->carp_version != CARP_VERSION) {
 604                 CARPSTATS_INC(carps_badver);
 605                 CARP_DEBUG("%s: invalid version %d\n", ifp->if_xname,
 606                     ch->carp_version);
 607                 ifa_free(ifa);
 608                 m_freem(m);
 609                 return;
 610         }
 611
 612         sc = ifa->ifa_carp;
 613         CARP_LOCK(sc);
 614         ifa_free(ifa);
 615
 616         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 617                 CARPSTATS_INC(carps_badauth);
 618                 CARP_DEBUG("%s: incorrect hash for VHID %u@%s\n", __func__,
 619                     sc->sc_vhid, ifp->if_xname);
 620                 goto out;
 621         }
 622
 623         tmp_counter = ntohl(ch->carp_counter[0]);
 624         tmp_counter = tmp_counter<<32;
 625         tmp_counter += ntohl(ch->carp_counter[1]);
 626
 627         /* XXX Replay protection goes here */
 628
 629         sc->sc_init_counter = 0;
 630         sc->sc_counter = tmp_counter;
 631
 632         sc_tv.tv_sec = sc->sc_advbase;
 633         sc_tv.tv_usec = DEMOTE_ADVSKEW(sc) * 1000000 / 256;
 634         ch_tv.tv_sec = ch->carp_advbase;
 635         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 636
 637         switch (sc->sc_state) {
 638         case INIT:
 639                 break;
 640         case MASTER:
 641                 /*
 642                  * If we receive an advertisement from a master who's going to
 643                  * be more frequent than us, go into BACKUP state.
 644                  */
 645                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 646                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 647                         callout_stop(&sc->sc_ad_tmo);
 648                         CARP_LOG("VHID %u@%s: MASTER -> BACKUP "
 649                             "(more frequent advertisement received)\n",
 650                             sc->sc_vhid,
 651                             sc->sc_carpdev->if_xname);
 652                         carp_set_state(sc, BACKUP);
 653                         carp_setrun(sc, 0);
 654                         carp_delroute(sc);
 655                 }
 656                 break;
 657         case BACKUP:
 658                 /*
 659                  * If we're pre-empting masters who advertise slower than us,
 660                  * and this one claims to be slower, treat him as down.
 661                  */
 662                 if (V_carp_preempt && timevalcmp(&sc_tv, &ch_tv, <)) {
 663                         CARP_LOG("VHID %u@%s: BACKUP -> MASTER "
 664                             "(preempting a slower master)\n",
 665                             sc->sc_vhid,
 666                             sc->sc_carpdev->if_xname);
 667                         carp_master_down_locked(sc);
 668                         break;
 669                 }
 670
 671                 /*
 672                  *  If the master is going to advertise at such a low frequency
 673                  *  that he's guaranteed to time out, we'd might as well just
 674                  *  treat him as timed out now.
 675                  */
 676                 sc_tv.tv_sec = sc->sc_advbase * 3;
 677                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 678                         CARP_LOG("VHID %u@%s: BACKUP -> MASTER "
 679                             "(master timed out)\n",
 680                             sc->sc_vhid,
 681                             sc->sc_carpdev->if_xname);
 682                         carp_master_down_locked(sc);
 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_locked(struct carp_softc *sc)
 766 {
 767         struct carp_header ch;
 768         struct timeval tv;
 769         struct sockaddr sa;
 770         struct ifaddr *ifa;
 771         struct carp_header *ch_ptr;
 772         struct mbuf *m;
 773         int len, advskew;
 774
 775         CARP_LOCK_ASSERT(sc);
 776
 777         advskew = DEMOTE_ADVSKEW(sc);
 778         tv.tv_sec = sc->sc_advbase;
 779         tv.tv_usec = advskew * 1000000 / 256;
 780
 781         ch.carp_version = CARP_VERSION;
 782         ch.carp_type = CARP_ADVERTISEMENT;
 783         ch.carp_vhid = sc->sc_vhid;
 784         ch.carp_advbase = sc->sc_advbase;
 785         ch.carp_advskew = advskew;
 786         ch.carp_authlen = 7;    /* XXX DEFINE */
 787         ch.carp_pad1 = 0;       /* must be zero */
 788         ch.carp_cksum = 0;
 789
 790         /* XXXGL: OpenBSD picks first ifaddr with needed family. */
 791
 792 #ifdef INET
 793         if (sc->sc_naddrs) {
 794                 struct ip *ip;
 795
 796                 m = m_gethdr(M_NOWAIT, MT_DATA);
 797                 if (m == NULL) {
 798                         CARPSTATS_INC(carps_onomem);
 799                         goto resched;
 800                 }
 801                 len = sizeof(*ip) + sizeof(ch);
 802                 m->m_pkthdr.len = len;
 803                 m->m_pkthdr.rcvif = NULL;
 804                 m->m_len = len;
 805                 MH_ALIGN(m, m->m_len);
 806                 m->m_flags |= M_MCAST;
 807                 ip = mtod(m, struct ip *);
 808                 ip->ip_v = IPVERSION;
 809                 ip->ip_hl = sizeof(*ip) >> 2;
 810                 ip->ip_tos = IPTOS_LOWDELAY;
 811                 ip->ip_len = htons(len);
 812                 ip->ip_id = ip_newid();
 813                 ip->ip_off = htons(IP_DF);
 814                 ip->ip_ttl = CARP_DFLTTL;
 815                 ip->ip_p = IPPROTO_CARP;
 816                 ip->ip_sum = 0;
 817
 818                 bzero(&sa, sizeof(sa));
 819                 sa.sa_family = AF_INET;
 820                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 821                 if (ifa != NULL) {
 822                         ip->ip_src.s_addr =
 823                             ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 824                         ifa_free(ifa);
 825                 } else
 826                         ip->ip_src.s_addr = 0;
 827                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 828
 829                 ch_ptr = (struct carp_header *)(&ip[1]);
 830                 bcopy(&ch, ch_ptr, sizeof(ch));
 831                 if (carp_prepare_ad(m, sc, ch_ptr))
 832                         goto resched;
 833
 834                 m->m_data += sizeof(*ip);
 835                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip));
 836                 m->m_data -= sizeof(*ip);
 837
 838                 CARPSTATS_INC(carps_opackets);
 839
 840                 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT,
 841                     &sc->sc_carpdev->if_carp->cif_imo, NULL)) {
 842                         if (sc->sc_sendad_errors < INT_MAX)
 843                                 sc->sc_sendad_errors++;
 844                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS)
 845                                 carp_demote_adj(V_carp_senderr_adj,
 846                                     "send error");
 847                         sc->sc_sendad_success = 0;
 848                 } else {
 849                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 850                                 if (++sc->sc_sendad_success >=
 851                                     CARP_SENDAD_MIN_SUCCESS) {
 852                                         carp_demote_adj(-V_carp_senderr_adj,
 853                                             "send ok");
 854                                         sc->sc_sendad_errors = 0;
 855                                 }
 856                         } else
 857                                 sc->sc_sendad_errors = 0;
 858                 }
 859         }
 860 #endif /* INET */
 861 #ifdef INET6
 862         if (sc->sc_naddrs6) {
 863                 struct ip6_hdr *ip6;
 864
 865                 m = m_gethdr(M_NOWAIT, MT_DATA);
 866                 if (m == NULL) {
 867                         CARPSTATS_INC(carps_onomem);
 868                         goto resched;
 869                 }
 870                 len = sizeof(*ip6) + sizeof(ch);
 871                 m->m_pkthdr.len = len;
 872                 m->m_pkthdr.rcvif = NULL;
 873                 m->m_len = len;
 874                 MH_ALIGN(m, m->m_len);
 875                 m->m_flags |= M_MCAST;
 876                 ip6 = mtod(m, struct ip6_hdr *);
 877                 bzero(ip6, sizeof(*ip6));
 878                 ip6->ip6_vfc |= IPV6_VERSION;
 879                 ip6->ip6_hlim = CARP_DFLTTL;
 880                 ip6->ip6_nxt = IPPROTO_CARP;
 881                 bzero(&sa, sizeof(sa));
 882
 883                 /* set the source address */
 884                 sa.sa_family = AF_INET6;
 885                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 886                 if (ifa != NULL) {
 887                         bcopy(IFA_IN6(ifa), &ip6->ip6_src,
 888                             sizeof(struct in6_addr));
 889                         ifa_free(ifa);
 890                 } else
 891                         /* This should never happen with IPv6. */
 892                         bzero(&ip6->ip6_src, sizeof(struct in6_addr));
 893
 894                 /* Set the multicast destination. */
 895                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
 896                 ip6->ip6_dst.s6_addr8[15] = 0x12;
 897                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
 898                         m_freem(m);
 899                         CARP_DEBUG("%s: in6_setscope failed\n", __func__);
 900                         goto resched;
 901                 }
 902
 903                 ch_ptr = (struct carp_header *)(&ip6[1]);
 904                 bcopy(&ch, ch_ptr, sizeof(ch));
 905                 if (carp_prepare_ad(m, sc, ch_ptr))
 906                         goto resched;
 907
 908                 m->m_data += sizeof(*ip6);
 909                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip6));
 910                 m->m_data -= sizeof(*ip6);
 911
 912                 CARPSTATS_INC(carps_opackets6);
 913
 914                 if (ip6_output(m, NULL, NULL, 0,
 915                     &sc->sc_carpdev->if_carp->cif_im6o, NULL, NULL)) {
 916                         if (sc->sc_sendad_errors < INT_MAX)
 917                                 sc->sc_sendad_errors++;
 918                         if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS)
 919                                 carp_demote_adj(V_carp_senderr_adj,
 920                                     "send6 error");
 921                         sc->sc_sendad_success = 0;
 922                 } else {
 923                         if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
 924                                 if (++sc->sc_sendad_success >=
 925                                     CARP_SENDAD_MIN_SUCCESS) {
 926                                         carp_demote_adj(-V_carp_senderr_adj,
 927                                             "send6 ok");
 928                                         sc->sc_sendad_errors = 0;
 929                                 }
 930                         } else
 931                                 sc->sc_sendad_errors = 0;
 932                 }
 933         }
 934 #endif /* INET6 */
 935
 936 resched:
 937         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), carp_send_ad, sc);
 938 }
 939
 940 static void
 941 carp_addroute(struct carp_softc *sc)
 942 {
 943         struct ifaddr *ifa;
 944
 945         CARP_FOREACH_IFA(sc, ifa)
 946                 carp_ifa_addroute(ifa);
 947 }
 948
 949 static void
 950 carp_ifa_addroute(struct ifaddr *ifa)
 951 {
 952
 953         switch (ifa->ifa_addr->sa_family) {
 954 #ifdef INET
 955         case AF_INET:
 956                 in_addprefix(ifatoia(ifa), RTF_UP);
 957                 ifa_add_loopback_route(ifa,
 958                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 959                 break;
 960 #endif
 961 #ifdef INET6
 962         case AF_INET6:
 963                 ifa_add_loopback_route(ifa,
 964                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 965                 in6_ifaddloop(ifa);
 966                 break;
 967 #endif
 968         }
 969 }
 970
 971 static void
 972 carp_delroute(struct carp_softc *sc)
 973 {
 974         struct ifaddr *ifa;
 975
 976         CARP_FOREACH_IFA(sc, ifa)
 977                 carp_ifa_delroute(ifa);
 978 }
 979
 980 static void
 981 carp_ifa_delroute(struct ifaddr *ifa)
 982 {
 983
 984         switch (ifa->ifa_addr->sa_family) {
 985 #ifdef INET
 986         case AF_INET:
 987                 ifa_del_loopback_route(ifa,
 988                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 989                 in_scrubprefix(ifatoia(ifa), LLE_STATIC);
 990                 break;
 991 #endif
 992 #ifdef INET6
 993         case AF_INET6:
 994                 ifa_del_loopback_route(ifa,
 995                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 996                 in6_ifremloop(ifa);
 997                 break;
 998 #endif
 999         }
1000 }
1001
1002 int
1003 carp_master(struct ifaddr *ifa)
1004 {
1005         struct carp_softc *sc = ifa->ifa_carp;
1006
1007         return (sc->sc_state == MASTER);
1008 }
1009
1010 #ifdef INET
1011 /*
1012  * Broadcast a gratuitous ARP request containing
1013  * the virtual router MAC address for each IP address
1014  * associated with the virtual router.
1015  */
1016 static void
1017 carp_send_arp(struct carp_softc *sc)
1018 {
1019         struct ifaddr *ifa;
1020
1021         CARP_FOREACH_IFA(sc, ifa)
1022                 if (ifa->ifa_addr->sa_family == AF_INET)
1023                         arp_ifinit2(sc->sc_carpdev, ifa, LLADDR(&sc->sc_addr));
1024 }
1025
1026 int
1027 carp_iamatch(struct ifaddr *ifa, uint8_t **enaddr)
1028 {
1029         struct carp_softc *sc = ifa->ifa_carp;
1030
1031         if (sc->sc_state == MASTER) {
1032                 *enaddr = LLADDR(&sc->sc_addr);
1033                 return (1);
1034         }
1035
1036         return (0);
1037 }
1038 #endif
1039
1040 #ifdef INET6
1041 static void
1042 carp_send_na(struct carp_softc *sc)
1043 {
1044         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1045         struct ifaddr *ifa;
1046         struct in6_addr *in6;
1047
1048         CARP_FOREACH_IFA(sc, ifa) {
1049                 if (ifa->ifa_addr->sa_family != AF_INET6)
1050                         continue;
1051
1052                 in6 = IFA_IN6(ifa);
1053                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1054                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1055                 DELAY(1000);    /* XXX */
1056         }
1057 }
1058
1059 /*
1060  * Returns ifa in case it's a carp address and it is MASTER, or if the address
1061  * matches and is not a carp address.  Returns NULL otherwise.
1062  */
1063 struct ifaddr *
1064 carp_iamatch6(struct ifnet *ifp, struct in6_addr *taddr)
1065 {
1066         struct ifaddr *ifa;
1067
1068         ifa = NULL;
1069         IF_ADDR_RLOCK(ifp);
1070         TAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
1071                 if (ifa->ifa_addr->sa_family != AF_INET6)
1072                         continue;
1073                 if (!IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa)))
1074                         continue;
1075                 if (ifa->ifa_carp && ifa->ifa_carp->sc_state != MASTER)
1076                         ifa = NULL;
1077                 else
1078                         ifa_ref(ifa);
1079                 break;
1080         }
1081         IF_ADDR_RUNLOCK(ifp);
1082
1083         return (ifa);
1084 }
1085
1086 caddr_t
1087 carp_macmatch6(struct ifnet *ifp, struct mbuf *m, const struct in6_addr *taddr)
1088 {
1089         struct ifaddr *ifa;
1090
1091         IF_ADDR_RLOCK(ifp);
1092         IFNET_FOREACH_IFA(ifp, ifa)
1093                 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1094                     IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1095                         struct carp_softc *sc = ifa->ifa_carp;
1096                         struct m_tag *mtag;
1097
1098                         IF_ADDR_RUNLOCK(ifp);
1099
1100                         mtag = m_tag_get(PACKET_TAG_CARP,
1101                             sizeof(struct carp_softc *), M_NOWAIT);
1102                         if (mtag == NULL)
1103                                 /* Better a bit than nothing. */
1104                                 return (LLADDR(&sc->sc_addr));
1105
1106                         bcopy(&sc, mtag + 1, sizeof(sc));
1107                         m_tag_prepend(m, mtag);
1108
1109                         return (LLADDR(&sc->sc_addr));
1110                 }
1111         IF_ADDR_RUNLOCK(ifp);
1112
1113         return (NULL);
1114 }
1115 #endif /* INET6 */
1116
1117 int
1118 carp_forus(struct ifnet *ifp, u_char *dhost)
1119 {
1120         struct carp_softc *sc;
1121         uint8_t *ena = dhost;
1122
1123         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1124                 return (0);
1125
1126         CIF_LOCK(ifp->if_carp);
1127         IFNET_FOREACH_CARP(ifp, sc) {
1128                 CARP_LOCK(sc);
1129                 if (sc->sc_state == MASTER && !bcmp(dhost, LLADDR(&sc->sc_addr),
1130                     ETHER_ADDR_LEN)) {
1131                         CARP_UNLOCK(sc);
1132                         CIF_UNLOCK(ifp->if_carp);
1133                         return (1);
1134                 }
1135                 CARP_UNLOCK(sc);
1136         }
1137         CIF_UNLOCK(ifp->if_carp);
1138
1139         return (0);
1140 }
1141
1142 /* Master down timeout event, executed in callout context. */
1143 static void
1144 carp_master_down(void *v)
1145 {
1146         struct carp_softc *sc = v;
1147
1148         CARP_LOCK_ASSERT(sc);
1149
1150         CURVNET_SET(sc->sc_carpdev->if_vnet);
1151         if (sc->sc_state == BACKUP) {
1152                 CARP_LOG("VHID %u@%s: BACKUP -> MASTER (master down)\n",
1153                     sc->sc_vhid,
1154                     sc->sc_carpdev->if_xname);
1155                 carp_master_down_locked(sc);
1156         }
1157         CURVNET_RESTORE();
1158
1159         CARP_UNLOCK(sc);
1160 }
1161
1162 static void
1163 carp_master_down_locked(struct carp_softc *sc)
1164 {
1165
1166         CARP_LOCK_ASSERT(sc);
1167
1168         switch (sc->sc_state) {
1169         case BACKUP:
1170                 carp_set_state(sc, MASTER);
1171                 carp_send_ad_locked(sc);
1172 #ifdef INET
1173                 carp_send_arp(sc);
1174 #endif
1175 #ifdef INET6
1176                 carp_send_na(sc);
1177 #endif
1178                 carp_setrun(sc, 0);
1179                 carp_addroute(sc);
1180                 break;
1181         case INIT:
1182         case MASTER:
1183 #ifdef INVARIANTS
1184                 panic("carp: VHID %u@%s: master_down event in %s state\n",
1185                     sc->sc_vhid,
1186                     sc->sc_carpdev->if_xname,
1187                     sc->sc_state ? "MASTER" : "INIT");
1188 #endif
1189                 break;
1190         }
1191 }
1192
1193 /*
1194  * When in backup state, af indicates whether to reset the master down timer
1195  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1196  */
1197 static void
1198 carp_setrun(struct carp_softc *sc, sa_family_t af)
1199 {
1200         struct timeval tv;
1201
1202         CARP_LOCK_ASSERT(sc);
1203
1204         if ((sc->sc_carpdev->if_flags & IFF_UP) == 0 ||
1205             sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1206             (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0))
1207                 return;
1208
1209         switch (sc->sc_state) {
1210         case INIT:
1211                 CARP_LOG("VHID %u@%s: INIT -> BACKUP\n",
1212                     sc->sc_vhid,
1213                     sc->sc_carpdev->if_xname);
1214                 carp_set_state(sc, BACKUP);
1215                 carp_setrun(sc, 0);
1216                 break;
1217         case BACKUP:
1218                 callout_stop(&sc->sc_ad_tmo);
1219                 tv.tv_sec = 3 * sc->sc_advbase;
1220                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1221                 switch (af) {
1222 #ifdef INET
1223                 case AF_INET:
1224                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1225                             carp_master_down, sc);
1226                         break;
1227 #endif
1228 #ifdef INET6
1229                 case AF_INET6:
1230                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1231                             carp_master_down, sc);
1232                         break;
1233 #endif
1234                 default:
1235 #ifdef INET
1236                         if (sc->sc_naddrs)
1237                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1238                                     carp_master_down, sc);
1239 #endif
1240 #ifdef INET6
1241                         if (sc->sc_naddrs6)
1242                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1243                                     carp_master_down, sc);
1244 #endif
1245                         break;
1246                 }
1247                 break;
1248         case MASTER:
1249                 tv.tv_sec = sc->sc_advbase;
1250                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1251                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1252                     carp_send_ad, sc);
1253                 break;
1254         }
1255 }
1256
1257 /*
1258  * Setup multicast structures.
1259  */
1260 static int
1261 carp_multicast_setup(struct carp_if *cif, sa_family_t sa)
1262 {
1263         struct ifnet *ifp = cif->cif_ifp;
1264         int error = 0;
1265
1266         CIF_LOCK_ASSERT(cif);
1267
1268         switch (sa) {
1269 #ifdef INET
1270         case AF_INET:
1271             {
1272                 struct ip_moptions *imo = &cif->cif_imo;
1273                 struct in_addr addr;
1274
1275                 if (imo->imo_membership)
1276                         return (0);
1277
1278                 imo->imo_membership = (struct in_multi **)malloc(
1279                     (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
1280                     M_NOWAIT);
1281                 if (imo->imo_membership == NULL)
1282                         return (ENOMEM);
1283                 imo->imo_mfilters = NULL;
1284                 imo->imo_max_memberships = IP_MIN_MEMBERSHIPS;
1285                 imo->imo_multicast_vif = -1;
1286
1287                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1288                 if ((error = in_joingroup(ifp, &addr, NULL,
1289                     &imo->imo_membership[0])) != 0) {
1290                         free(imo->imo_membership, M_CARP);
1291                         break;
1292                 }
1293                 imo->imo_num_memberships++;
1294                 imo->imo_multicast_ifp = ifp;
1295                 imo->imo_multicast_ttl = CARP_DFLTTL;
1296                 imo->imo_multicast_loop = 0;
1297                 break;
1298            }
1299 #endif
1300 #ifdef INET6
1301         case AF_INET6:
1302             {
1303                 struct ip6_moptions *im6o = &cif->cif_im6o;
1304                 struct in6_addr in6;
1305                 struct in6_multi *in6m;
1306
1307                 if (im6o->im6o_membership)
1308                         return (0);
1309
1310                 im6o->im6o_membership = (struct in6_multi **)malloc(
1311                     (sizeof(struct in6_multi *) * IPV6_MIN_MEMBERSHIPS), M_CARP,
1312                     M_ZERO | M_NOWAIT);
1313                 if (im6o->im6o_membership == NULL)
1314                         return (ENOMEM);
1315                 im6o->im6o_mfilters = NULL;
1316                 im6o->im6o_max_memberships = IPV6_MIN_MEMBERSHIPS;
1317                 im6o->im6o_multicast_hlim = CARP_DFLTTL;
1318                 im6o->im6o_multicast_ifp = ifp;
1319
1320                 /* Join IPv6 CARP multicast group. */
1321                 bzero(&in6, sizeof(in6));
1322                 in6.s6_addr16[0] = htons(0xff02);
1323                 in6.s6_addr8[15] = 0x12;
1324                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1325                         free(im6o->im6o_membership, M_CARP);
1326                         break;
1327                 }
1328                 in6m = NULL;
1329                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1330                         free(im6o->im6o_membership, M_CARP);
1331                         break;
1332                 }
1333                 im6o->im6o_membership[0] = in6m;
1334                 im6o->im6o_num_memberships++;
1335
1336                 /* Join solicited multicast address. */
1337                 bzero(&in6, sizeof(in6));
1338                 in6.s6_addr16[0] = htons(0xff02);
1339                 in6.s6_addr32[1] = 0;
1340                 in6.s6_addr32[2] = htonl(1);
1341                 in6.s6_addr32[3] = 0;
1342                 in6.s6_addr8[12] = 0xff;
1343                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1344                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1345                         free(im6o->im6o_membership, M_CARP);
1346                         break;
1347                 }
1348                 in6m = NULL;
1349                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1350                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1351                         free(im6o->im6o_membership, M_CARP);
1352                         break;
1353                 }
1354                 im6o->im6o_membership[1] = in6m;
1355                 im6o->im6o_num_memberships++;
1356                 break;
1357             }
1358 #endif
1359         }
1360
1361         return (error);
1362 }
1363
1364 /*
1365  * Free multicast structures.
1366  */
1367 static void
1368 carp_multicast_cleanup(struct carp_if *cif, sa_family_t sa)
1369 {
1370
1371         CIF_LOCK_ASSERT(cif);
1372         switch (sa) {
1373 #ifdef INET
1374         case AF_INET:
1375                 if (cif->cif_naddrs == 0) {
1376                         struct ip_moptions *imo = &cif->cif_imo;
1377
1378                         in_leavegroup(imo->imo_membership[0], NULL);
1379                         KASSERT(imo->imo_mfilters == NULL,
1380                             ("%s: imo_mfilters != NULL", __func__));
1381                         free(imo->imo_membership, M_CARP);
1382                         imo->imo_membership = NULL;
1383
1384                 }
1385                 break;
1386 #endif
1387 #ifdef INET6
1388         case AF_INET6:
1389                 if (cif->cif_naddrs6 == 0) {
1390                         struct ip6_moptions *im6o = &cif->cif_im6o;
1391
1392                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1393                         in6_mc_leave(im6o->im6o_membership[1], NULL);
1394                         KASSERT(im6o->im6o_mfilters == NULL,
1395                             ("%s: im6o_mfilters != NULL", __func__));
1396                         free(im6o->im6o_membership, M_CARP);
1397                         im6o->im6o_membership = NULL;
1398                 }
1399                 break;
1400 #endif
1401         }
1402 }
1403
1404 int
1405 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa)
1406 {
1407         struct m_tag *mtag;
1408         struct carp_softc *sc;
1409
1410         if (!sa)
1411                 return (0);
1412
1413         switch (sa->sa_family) {
1414 #ifdef INET
1415         case AF_INET:
1416                 break;
1417 #endif
1418 #ifdef INET6
1419         case AF_INET6:
1420                 break;
1421 #endif
1422         default:
1423                 return (0);
1424         }
1425
1426         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
1427         if (mtag == NULL)
1428                 return (0);
1429
1430         bcopy(mtag + 1, &sc, sizeof(sc));
1431
1432         /* Set the source MAC address to the Virtual Router MAC Address. */
1433         switch (ifp->if_type) {
1434         case IFT_ETHER:
1435         case IFT_BRIDGE:
1436         case IFT_L2VLAN: {
1437                         struct ether_header *eh;
1438
1439                         eh = mtod(m, struct ether_header *);
1440                         eh->ether_shost[0] = 0;
1441                         eh->ether_shost[1] = 0;
1442                         eh->ether_shost[2] = 0x5e;
1443                         eh->ether_shost[3] = 0;
1444                         eh->ether_shost[4] = 1;
1445                         eh->ether_shost[5] = sc->sc_vhid;
1446                 }
1447                 break;
1448         case IFT_FDDI: {
1449                         struct fddi_header *fh;
1450
1451                         fh = mtod(m, struct fddi_header *);
1452                         fh->fddi_shost[0] = 0;
1453                         fh->fddi_shost[1] = 0;
1454                         fh->fddi_shost[2] = 0x5e;
1455                         fh->fddi_shost[3] = 0;
1456                         fh->fddi_shost[4] = 1;
1457                         fh->fddi_shost[5] = sc->sc_vhid;
1458                 }
1459                 break;
1460         case IFT_ISO88025: {
1461                         struct iso88025_header *th;
1462                         th = mtod(m, struct iso88025_header *);
1463                         th->iso88025_shost[0] = 3;
1464                         th->iso88025_shost[1] = 0;
1465                         th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
1466                         th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
1467                         th->iso88025_shost[4] = 0;
1468                         th->iso88025_shost[5] = 0;
1469                 }
1470                 break;
1471         default:
1472                 printf("%s: carp is not supported for the %d interface type\n",
1473                     ifp->if_xname, ifp->if_type);
1474                 return (EOPNOTSUPP);
1475         }
1476
1477         return (0);
1478 }
1479
1480 static struct carp_softc*
1481 carp_alloc(struct ifnet *ifp)
1482 {
1483         struct carp_softc *sc;
1484         struct carp_if *cif;
1485
1486         if ((cif = ifp->if_carp) == NULL) {
1487                 cif = carp_alloc_if(ifp);
1488                 if (cif == NULL)
1489                         return (NULL);
1490         }
1491
1492         sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
1493
1494         sc->sc_advbase = CARP_DFLTINTV;
1495         sc->sc_vhid = -1;       /* required setting */
1496         sc->sc_init_counter = 1;
1497         sc->sc_state = INIT;
1498
1499         sc->sc_ifasiz = sizeof(struct ifaddr *);
1500         sc->sc_ifas = malloc(sc->sc_ifasiz, M_CARP, M_WAITOK|M_ZERO);
1501         sc->sc_carpdev = ifp;
1502
1503         CARP_LOCK_INIT(sc);
1504 #ifdef INET
1505         callout_init_mtx(&sc->sc_md_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1506 #endif
1507 #ifdef INET6
1508         callout_init_mtx(&sc->sc_md6_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1509 #endif
1510         callout_init_mtx(&sc->sc_ad_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1511
1512         CIF_LOCK(cif);
1513         TAILQ_INSERT_TAIL(&cif->cif_vrs, sc, sc_list);
1514         CIF_UNLOCK(cif);
1515
1516         mtx_lock(&carp_mtx);
1517         LIST_INSERT_HEAD(&carp_list, sc, sc_next);
1518         mtx_unlock(&carp_mtx);
1519
1520         return (sc);
1521 }
1522
1523 static int
1524 carp_grow_ifas(struct carp_softc *sc)
1525 {
1526         struct ifaddr **new;
1527
1528         CARP_LOCK_ASSERT(sc);
1529
1530         new = malloc(sc->sc_ifasiz * 2, M_CARP, M_NOWAIT|M_ZERO);
1531         if (new == NULL)
1532                 return (ENOMEM);
1533         bcopy(sc->sc_ifas, new, sc->sc_ifasiz);
1534         free(sc->sc_ifas, M_CARP);
1535         sc->sc_ifas = new;
1536         sc->sc_ifasiz *= 2;
1537
1538         return (0);
1539 }
1540
1541 static void
1542 carp_destroy(struct carp_softc *sc)
1543 {
1544         struct ifnet *ifp = sc->sc_carpdev;
1545         struct carp_if *cif = ifp->if_carp;
1546
1547         CIF_LOCK_ASSERT(cif);
1548
1549         TAILQ_REMOVE(&cif->cif_vrs, sc, sc_list);
1550
1551         mtx_lock(&carp_mtx);
1552         LIST_REMOVE(sc, sc_next);
1553         mtx_unlock(&carp_mtx);
1554
1555         CARP_LOCK(sc);
1556         if (sc->sc_suppress)
1557                 carp_demote_adj(-V_carp_ifdown_adj, "vhid removed");
1558         callout_drain(&sc->sc_ad_tmo);
1559 #ifdef INET
1560         callout_drain(&sc->sc_md_tmo);
1561 #endif
1562 #ifdef INET6
1563         callout_drain(&sc->sc_md6_tmo);
1564 #endif
1565         CARP_LOCK_DESTROY(sc);
1566
1567         free(sc->sc_ifas, M_CARP);
1568         free(sc, M_CARP);
1569 }
1570
1571 static struct carp_if*
1572 carp_alloc_if(struct ifnet *ifp)
1573 {
1574         struct carp_if *cif;
1575
1576         cif = malloc(sizeof(*cif), M_CARP, M_WAITOK|M_ZERO);
1577
1578         if (ifpromisc(ifp, 1) != 0)
1579                 goto cleanup;
1580
1581         CIF_LOCK_INIT(cif);
1582         cif->cif_ifp = ifp;
1583         TAILQ_INIT(&cif->cif_vrs);
1584
1585         IF_ADDR_WLOCK(ifp);
1586         ifp->if_carp = cif;
1587         if_ref(ifp);
1588         IF_ADDR_WUNLOCK(ifp);
1589
1590         return (cif);
1591
1592 cleanup:
1593         free(cif, M_CARP);
1594
1595         return (NULL);
1596 }
1597
1598 static void
1599 carp_free_if(struct carp_if *cif)
1600 {
1601         struct ifnet *ifp = cif->cif_ifp;
1602
1603         CIF_LOCK_ASSERT(cif);
1604         KASSERT(TAILQ_EMPTY(&cif->cif_vrs), ("%s: softc list not empty",
1605             __func__));
1606
1607         IF_ADDR_WLOCK(ifp);
1608         ifp->if_carp = NULL;
1609         IF_ADDR_WUNLOCK(ifp);
1610
1611         CIF_LOCK_DESTROY(cif);
1612
1613         ifpromisc(ifp, 0);
1614         if_rele(ifp);
1615
1616         free(cif, M_CARP);
1617 }
1618
1619 static void
1620 carp_carprcp(struct carpreq *carpr, struct carp_softc *sc, int priv)
1621 {
1622
1623         CARP_LOCK(sc);
1624         carpr->carpr_state = sc->sc_state;
1625         carpr->carpr_vhid = sc->sc_vhid;
1626         carpr->carpr_advbase = sc->sc_advbase;
1627         carpr->carpr_advskew = sc->sc_advskew;
1628         if (priv)
1629                 bcopy(sc->sc_key, carpr->carpr_key, sizeof(carpr->carpr_key));
1630         else
1631                 bzero(carpr->carpr_key, sizeof(carpr->carpr_key));
1632         CARP_UNLOCK(sc);
1633 }
1634
1635 int
1636 carp_ioctl(struct ifreq *ifr, u_long cmd, struct thread *td)
1637 {
1638         struct carpreq carpr;
1639         struct ifnet *ifp;
1640         struct carp_softc *sc = NULL;
1641         int error = 0, locked = 0;
1642
1643         if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1644                 return (error);
1645
1646         ifp = ifunit_ref(ifr->ifr_name);
1647         if (ifp == NULL)
1648                 return (ENXIO);
1649
1650         switch (ifp->if_type) {
1651         case IFT_ETHER:
1652         case IFT_L2VLAN:
1653         case IFT_BRIDGE:
1654         case IFT_FDDI:
1655         case IFT_ISO88025:
1656                 break;
1657         default:
1658                 error = EOPNOTSUPP;
1659                 goto out;
1660         }
1661
1662         if ((ifp->if_flags & IFF_MULTICAST) == 0) {
1663                 error = EADDRNOTAVAIL;
1664                 goto out;
1665         }
1666
1667         switch (cmd) {
1668         case SIOCSVH:
1669                 if ((error = priv_check(td, PRIV_NETINET_CARP)))
1670                         break;
1671                 if (carpr.carpr_vhid <= 0 || carpr.carpr_vhid > CARP_MAXVHID ||
1672                     carpr.carpr_advbase < 0 || carpr.carpr_advskew < 0) {
1673                         error = EINVAL;
1674                         break;
1675                 }
1676
1677                 if (ifp->if_carp) {
1678                         CIF_LOCK(ifp->if_carp);
1679                         IFNET_FOREACH_CARP(ifp, sc)
1680                                 if (sc->sc_vhid == carpr.carpr_vhid)
1681                                         break;
1682                         CIF_UNLOCK(ifp->if_carp);
1683                 }
1684                 if (sc == NULL) {
1685                         sc = carp_alloc(ifp);
1686                         if (sc == NULL) {
1687                                 error = EINVAL; /* XXX: ifpromisc failed */
1688                                 break;
1689                         }
1690
1691                         CARP_LOCK(sc);
1692                         sc->sc_vhid = carpr.carpr_vhid;
1693                         LLADDR(&sc->sc_addr)[0] = 0;
1694                         LLADDR(&sc->sc_addr)[1] = 0;
1695                         LLADDR(&sc->sc_addr)[2] = 0x5e;
1696                         LLADDR(&sc->sc_addr)[3] = 0;
1697                         LLADDR(&sc->sc_addr)[4] = 1;
1698                         LLADDR(&sc->sc_addr)[5] = sc->sc_vhid;
1699                 } else
1700                         CARP_LOCK(sc);
1701                 locked = 1;
1702                 if (carpr.carpr_advbase > 0) {
1703                         if (carpr.carpr_advbase > 255 ||
1704                             carpr.carpr_advbase < CARP_DFLTINTV) {
1705                                 error = EINVAL;
1706                                 break;
1707                         }
1708                         sc->sc_advbase = carpr.carpr_advbase;
1709                 }
1710                 if (carpr.carpr_advskew > 0) {
1711                         if (carpr.carpr_advskew >= 255) {
1712                                 error = EINVAL;
1713                                 break;
1714                         }
1715                         sc->sc_advskew = carpr.carpr_advskew;
1716                 }
1717                 if (carpr.carpr_key[0] != '\0') {
1718                         bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1719                         carp_hmac_prepare(sc);
1720                 }
1721                 if (sc->sc_state != INIT &&
1722                     carpr.carpr_state != sc->sc_state) {
1723                         switch (carpr.carpr_state) {
1724                         case BACKUP:
1725                                 callout_stop(&sc->sc_ad_tmo);
1726                                 carp_set_state(sc, BACKUP);
1727                                 carp_setrun(sc, 0);
1728                                 carp_delroute(sc);
1729                                 break;
1730                         case MASTER:
1731                                 carp_master_down_locked(sc);
1732                                 break;
1733                         default:
1734                                 break;
1735                         }
1736                 }
1737                 break;
1738
1739         case SIOCGVH:
1740             {
1741                 int priveleged;
1742
1743                 if (carpr.carpr_vhid < 0 || carpr.carpr_vhid > CARP_MAXVHID) {
1744                         error = EINVAL;
1745                         break;
1746                 }
1747                 if (carpr.carpr_count < 1) {
1748                         error = EMSGSIZE;
1749                         break;
1750                 }
1751                 if (ifp->if_carp == NULL) {
1752                         error = ENOENT;
1753                         break;
1754                 }
1755
1756                 priveleged = (priv_check(td, PRIV_NETINET_CARP) == 0);
1757                 if (carpr.carpr_vhid != 0) {
1758                         CIF_LOCK(ifp->if_carp);
1759                         IFNET_FOREACH_CARP(ifp, sc)
1760                                 if (sc->sc_vhid == carpr.carpr_vhid)
1761                                         break;
1762                         CIF_UNLOCK(ifp->if_carp);
1763                         if (sc == NULL) {
1764                                 error = ENOENT;
1765                                 break;
1766                         }
1767                         carp_carprcp(&carpr, sc, priveleged);
1768                         error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1769                 } else  {
1770                         int i, count;
1771
1772                         count = 0;
1773                         CIF_LOCK(ifp->if_carp);
1774                         IFNET_FOREACH_CARP(ifp, sc)
1775                                 count++;
1776
1777                         if (count > carpr.carpr_count) {
1778                                 CIF_UNLOCK(ifp->if_carp);
1779                                 error = EMSGSIZE;
1780                                 break;
1781                         }
1782
1783                         i = 0;
1784                         IFNET_FOREACH_CARP(ifp, sc) {
1785                                 carp_carprcp(&carpr, sc, priveleged);
1786                                 carpr.carpr_count = count;
1787                                 error = copyout(&carpr, ifr->ifr_data +
1788                                     (i * sizeof(carpr)), sizeof(carpr));
1789                                 if (error) {
1790                                         CIF_UNLOCK(ifp->if_carp);
1791                                         break;
1792                                 }
1793                                 i++;
1794                         }
1795                         CIF_UNLOCK(ifp->if_carp);
1796                 }
1797                 break;
1798             }
1799         default:
1800                 error = EINVAL;
1801         }
1802
1803 out:
1804         if (locked)
1805                 CARP_UNLOCK(sc);
1806         if_rele(ifp);
1807
1808         return (error);
1809 }
1810
1811 static int
1812 carp_get_vhid(struct ifaddr *ifa)
1813 {
1814
1815         if (ifa == NULL || ifa->ifa_carp == NULL)
1816                 return (0);
1817
1818         return (ifa->ifa_carp->sc_vhid);
1819 }
1820
1821 int
1822 carp_attach(struct ifaddr *ifa, int vhid)
1823 {
1824         struct ifnet *ifp = ifa->ifa_ifp;
1825         struct carp_if *cif = ifp->if_carp;
1826         struct carp_softc *sc;
1827         int index, error;
1828
1829         if (ifp->if_carp == NULL)
1830                 return (ENOPROTOOPT);
1831
1832         switch (ifa->ifa_addr->sa_family) {
1833 #ifdef INET
1834         case AF_INET:
1835 #endif
1836 #ifdef INET6
1837         case AF_INET6:
1838 #endif
1839                 break;
1840         default:
1841                 return (EPROTOTYPE);
1842         }
1843
1844         CIF_LOCK(cif);
1845         IFNET_FOREACH_CARP(ifp, sc)
1846                 if (sc->sc_vhid == vhid)
1847                         break;
1848         if (sc == NULL) {
1849                 CIF_UNLOCK(cif);
1850                 return (ENOENT);
1851         }
1852
1853         if (ifa->ifa_carp) {
1854                 if (ifa->ifa_carp->sc_vhid != vhid)
1855                         carp_detach_locked(ifa);
1856                 else {
1857                         CIF_UNLOCK(cif);
1858                         return (0);
1859                 }
1860         }
1861
1862         error = carp_multicast_setup(cif, ifa->ifa_addr->sa_family);
1863         if (error) {
1864                 CIF_FREE(cif);
1865                 return (error);
1866         }
1867
1868         CARP_LOCK(sc);
1869         index = sc->sc_naddrs + sc->sc_naddrs6 + 1;
1870         if (index > sc->sc_ifasiz / sizeof(struct ifaddr *))
1871                 if ((error = carp_grow_ifas(sc)) != 0) {
1872                         carp_multicast_cleanup(cif,
1873                             ifa->ifa_addr->sa_family);
1874                         CARP_UNLOCK(sc);
1875                         CIF_FREE(cif);
1876                         return (error);
1877                 }
1878
1879         switch (ifa->ifa_addr->sa_family) {
1880 #ifdef INET
1881         case AF_INET:
1882                 cif->cif_naddrs++;
1883                 sc->sc_naddrs++;
1884                 break;
1885 #endif
1886 #ifdef INET6
1887         case AF_INET6:
1888                 cif->cif_naddrs6++;
1889                 sc->sc_naddrs6++;
1890                 break;
1891 #endif
1892         }
1893
1894         ifa_ref(ifa);
1895         sc->sc_ifas[index - 1] = ifa;
1896         ifa->ifa_carp = sc;
1897
1898         carp_hmac_prepare(sc);
1899         carp_sc_state(sc);
1900
1901         CARP_UNLOCK(sc);
1902         CIF_UNLOCK(cif);
1903
1904         return (0);
1905 }
1906
1907 void
1908 carp_detach(struct ifaddr *ifa)
1909 {
1910         struct ifnet *ifp = ifa->ifa_ifp;
1911         struct carp_if *cif = ifp->if_carp;
1912
1913         CIF_LOCK(cif);
1914         carp_detach_locked(ifa);
1915         CIF_FREE(cif);
1916 }
1917
1918 static void
1919 carp_detach_locked(struct ifaddr *ifa)
1920 {
1921         struct ifnet *ifp = ifa->ifa_ifp;
1922         struct carp_if *cif = ifp->if_carp;
1923         struct carp_softc *sc = ifa->ifa_carp;
1924         int i, index;
1925
1926         KASSERT(sc != NULL, ("%s: %p not attached", __func__, ifa));
1927
1928         CIF_LOCK_ASSERT(cif);
1929         CARP_LOCK(sc);
1930
1931         /* Shift array. */
1932         index = sc->sc_naddrs + sc->sc_naddrs6;
1933         for (i = 0; i < index; i++)
1934                 if (sc->sc_ifas[i] == ifa)
1935                         break;
1936         KASSERT(i < index, ("%s: %p no backref", __func__, ifa));
1937         for (; i < index - 1; i++)
1938                 sc->sc_ifas[i] = sc->sc_ifas[i+1];
1939         sc->sc_ifas[index - 1] = NULL;
1940
1941         switch (ifa->ifa_addr->sa_family) {
1942 #ifdef INET
1943         case AF_INET:
1944                 cif->cif_naddrs--;
1945                 sc->sc_naddrs--;
1946                 break;
1947 #endif
1948 #ifdef INET6
1949         case AF_INET6:
1950                 cif->cif_naddrs6--;
1951                 sc->sc_naddrs6--;
1952                 break;
1953 #endif
1954         }
1955
1956         carp_ifa_delroute(ifa);
1957         carp_multicast_cleanup(cif, ifa->ifa_addr->sa_family);
1958
1959         ifa->ifa_carp = NULL;
1960         ifa_free(ifa);
1961
1962         carp_hmac_prepare(sc);
1963         carp_sc_state(sc);
1964
1965         if (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) {
1966                 CARP_UNLOCK(sc);
1967                 carp_destroy(sc);
1968         } else
1969                 CARP_UNLOCK(sc);
1970 }
1971
1972 static void
1973 carp_set_state(struct carp_softc *sc, int state)
1974 {
1975
1976         CARP_LOCK_ASSERT(sc);
1977
1978         if (sc->sc_state != state) {
1979                 const char *carp_states[] = { CARP_STATES };
1980                 char subsys[IFNAMSIZ+5];
1981
1982                 sc->sc_state = state;
1983
1984                 snprintf(subsys, IFNAMSIZ+5, "%u@%s", sc->sc_vhid,
1985                     sc->sc_carpdev->if_xname);
1986                 devctl_notify("CARP", subsys, carp_states[state], NULL);
1987         }
1988 }
1989
1990 static void
1991 carp_linkstate(struct ifnet *ifp)
1992 {
1993         struct carp_softc *sc;
1994
1995         CIF_LOCK(ifp->if_carp);
1996         IFNET_FOREACH_CARP(ifp, sc) {
1997                 CARP_LOCK(sc);
1998                 carp_sc_state(sc);
1999                 CARP_UNLOCK(sc);
2000         }
2001         CIF_UNLOCK(ifp->if_carp);
2002 }
2003
2004 static void
2005 carp_sc_state(struct carp_softc *sc)
2006 {
2007
2008         CARP_LOCK_ASSERT(sc);
2009
2010         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
2011             !(sc->sc_carpdev->if_flags & IFF_UP)) {
2012                 callout_stop(&sc->sc_ad_tmo);
2013 #ifdef INET
2014                 callout_stop(&sc->sc_md_tmo);
2015 #endif
2016 #ifdef INET6
2017                 callout_stop(&sc->sc_md6_tmo);
2018 #endif
2019                 carp_set_state(sc, INIT);
2020                 carp_setrun(sc, 0);
2021                 if (!sc->sc_suppress)
2022                         carp_demote_adj(V_carp_ifdown_adj, "interface down");
2023                 sc->sc_suppress = 1;
2024         } else {
2025                 carp_set_state(sc, INIT);
2026                 carp_setrun(sc, 0);
2027                 if (sc->sc_suppress)
2028                         carp_demote_adj(-V_carp_ifdown_adj, "interface up");
2029                 sc->sc_suppress = 0;
2030         }
2031 }
2032
2033 static void
2034 carp_demote_adj(int adj, char *reason)
2035 {
2036         atomic_add_int(&V_carp_demotion, adj);
2037         CARP_LOG("demoted by %d to %d (%s)\n", adj, V_carp_demotion, reason);
2038         taskqueue_enqueue(taskqueue_swi, &carp_sendall_task);
2039 }
2040
2041 static int
2042 carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS)
2043 {
2044         int new, error;
2045
2046         new = V_carp_demotion;
2047         error = sysctl_handle_int(oidp, &new, 0, req);
2048         if (error || !req->newptr)
2049                 return (error);
2050
2051         carp_demote_adj(new, "sysctl");
2052
2053         return (0);
2054 }
2055
2056 #ifdef INET
2057 extern  struct domain inetdomain;
2058 static struct protosw in_carp_protosw = {
2059         .pr_type =              SOCK_RAW,
2060         .pr_domain =            &inetdomain,
2061         .pr_protocol =          IPPROTO_CARP,
2062         .pr_flags =             PR_ATOMIC|PR_ADDR,
2063         .pr_input =             carp_input,
2064         .pr_output =            (pr_output_t *)rip_output,
2065         .pr_ctloutput =         rip_ctloutput,
2066         .pr_usrreqs =           &rip_usrreqs
2067 };
2068 #endif
2069
2070 #ifdef INET6
2071 extern  struct domain inet6domain;
2072 static struct ip6protosw in6_carp_protosw = {
2073         .pr_type =              SOCK_RAW,
2074         .pr_domain =            &inet6domain,
2075         .pr_protocol =          IPPROTO_CARP,
2076         .pr_flags =             PR_ATOMIC|PR_ADDR,
2077         .pr_input =             carp6_input,
2078         .pr_output =            rip6_output,
2079         .pr_ctloutput =         rip6_ctloutput,
2080         .pr_usrreqs =           &rip6_usrreqs
2081 };
2082 #endif
2083
2084 static void
2085 carp_mod_cleanup(void)
2086 {
2087
2088 #ifdef INET
2089         if (proto_reg[CARP_INET] == 0) {
2090                 (void)ipproto_unregister(IPPROTO_CARP);
2091                 pf_proto_unregister(PF_INET, IPPROTO_CARP, SOCK_RAW);
2092                 proto_reg[CARP_INET] = -1;
2093         }
2094         carp_iamatch_p = NULL;
2095 #endif
2096 #ifdef INET6
2097         if (proto_reg[CARP_INET6] == 0) {
2098                 (void)ip6proto_unregister(IPPROTO_CARP);
2099                 pf_proto_unregister(PF_INET6, IPPROTO_CARP, SOCK_RAW);
2100                 proto_reg[CARP_INET6] = -1;
2101         }
2102         carp_iamatch6_p = NULL;
2103         carp_macmatch6_p = NULL;
2104 #endif
2105         carp_ioctl_p = NULL;
2106         carp_attach_p = NULL;
2107         carp_detach_p = NULL;
2108         carp_get_vhid_p = NULL;
2109         carp_linkstate_p = NULL;
2110         carp_forus_p = NULL;
2111         carp_output_p = NULL;
2112         carp_demote_adj_p = NULL;
2113         carp_master_p = NULL;
2114         mtx_unlock(&carp_mtx);
2115         taskqueue_drain(taskqueue_swi, &carp_sendall_task);
2116         mtx_destroy(&carp_mtx);
2117 }
2118
2119 static int
2120 carp_mod_load(void)
2121 {
2122         int err;
2123
2124         mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2125         LIST_INIT(&carp_list);
2126         carp_get_vhid_p = carp_get_vhid;
2127         carp_forus_p = carp_forus;
2128         carp_output_p = carp_output;
2129         carp_linkstate_p = carp_linkstate;
2130         carp_ioctl_p = carp_ioctl;
2131         carp_attach_p = carp_attach;
2132         carp_detach_p = carp_detach;
2133         carp_demote_adj_p = carp_demote_adj;
2134         carp_master_p = carp_master;
2135 #ifdef INET6
2136         carp_iamatch6_p = carp_iamatch6;
2137         carp_macmatch6_p = carp_macmatch6;
2138         proto_reg[CARP_INET6] = pf_proto_register(PF_INET6,
2139             (struct protosw *)&in6_carp_protosw);
2140         if (proto_reg[CARP_INET6]) {
2141                 printf("carp: error %d attaching to PF_INET6\n",
2142                     proto_reg[CARP_INET6]);
2143                 carp_mod_cleanup();
2144                 return (proto_reg[CARP_INET6]);
2145         }
2146         err = ip6proto_register(IPPROTO_CARP);
2147         if (err) {
2148                 printf("carp: error %d registering with INET6\n", err);
2149                 carp_mod_cleanup();
2150                 return (err);
2151         }
2152 #endif
2153 #ifdef INET
2154         carp_iamatch_p = carp_iamatch;
2155         proto_reg[CARP_INET] = pf_proto_register(PF_INET, &in_carp_protosw);
2156         if (proto_reg[CARP_INET]) {
2157                 printf("carp: error %d attaching to PF_INET\n",
2158                     proto_reg[CARP_INET]);
2159                 carp_mod_cleanup();
2160                 return (proto_reg[CARP_INET]);
2161         }
2162         err = ipproto_register(IPPROTO_CARP);
2163         if (err) {
2164                 printf("carp: error %d registering with INET\n", err);
2165                 carp_mod_cleanup();
2166                 return (err);
2167         }
2168 #endif
2169         return (0);
2170 }
2171
2172 static int
2173 carp_modevent(module_t mod, int type, void *data)
2174 {
2175         switch (type) {
2176         case MOD_LOAD:
2177                 return carp_mod_load();
2178                 /* NOTREACHED */
2179         case MOD_UNLOAD:
2180                 mtx_lock(&carp_mtx);
2181                 if (LIST_EMPTY(&carp_list))
2182                         carp_mod_cleanup();
2183                 else {
2184                         mtx_unlock(&carp_mtx);
2185                         return (EBUSY);
2186                 }
2187                 break;
2188
2189         default:
2190                 return (EINVAL);
2191         }
2192
2193         return (0);
2194 }
2195
2196 static moduledata_t carp_mod = {
2197         "carp",
2198         carp_modevent,
2199         0
2200 };
2201
2202 DECLARE_MODULE(carp, carp_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);