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_VNET_INT(_net_inet_carp, OID_AUTO, allow, CTLFLAG_RW,
 220     &VNET_NAME(carp_allow), 0, "Accept incoming CARP packets");
 221 SYSCTL_VNET_INT(_net_inet_carp, OID_AUTO, preempt, CTLFLAG_RW,
 222     &VNET_NAME(carp_preempt), 0, "High-priority backup preemption mode");
 223 SYSCTL_VNET_INT(_net_inet_carp, OID_AUTO, log, CTLFLAG_RW,
 224     &VNET_NAME(carp_log), 0, "CARP log level");
 225 SYSCTL_VNET_PROC(_net_inet_carp, OID_AUTO, demotion, CTLTYPE_INT|CTLFLAG_RW,
 226     0, 0, carp_demote_adj_sysctl, "I",
 227     "Adjust demotion factor (skew of advskew)");
 228 SYSCTL_VNET_INT(_net_inet_carp, OID_AUTO, senderr_demotion_factor, CTLFLAG_RW,
 229     &VNET_NAME(carp_senderr_adj), 0, "Send error demotion factor adjustment");
 230 SYSCTL_VNET_INT(_net_inet_carp, OID_AUTO, ifdown_demotion_factor, CTLFLAG_RW,
 231     &VNET_NAME(carp_ifdown_adj), 0,
 232     "Interface down demotion factor adjustment");
 233
 234 VNET_PCPUSTAT_DEFINE(struct carpstats, carpstats);
 235 VNET_PCPUSTAT_SYSINIT(carpstats);
 236 VNET_PCPUSTAT_SYSUNINIT(carpstats);
 237
 238 #define CARPSTATS_ADD(name, val)        \
 239     counter_u64_add(VNET(carpstats)[offsetof(struct carpstats, name) / \
 240         sizeof(uint64_t)], (val))
 241 #define CARPSTATS_INC(name)             CARPSTATS_ADD(name, 1)
 242
 243 SYSCTL_VNET_PCPUSTAT(_net_inet_carp, OID_AUTO, stats, struct carpstats,
 244     carpstats, "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 245
 246 #define CARP_LOCK_INIT(sc)      mtx_init(&(sc)->sc_mtx, "carp_softc",   \
 247         NULL, MTX_DEF)
 248 #define CARP_LOCK_DESTROY(sc)   mtx_destroy(&(sc)->sc_mtx)
 249 #define CARP_LOCK_ASSERT(sc)    mtx_assert(&(sc)->sc_mtx, MA_OWNED)
 250 #define CARP_LOCK(sc)           mtx_lock(&(sc)->sc_mtx)
 251 #define CARP_UNLOCK(sc)         mtx_unlock(&(sc)->sc_mtx)
 252 #define CIF_LOCK_INIT(cif)      mtx_init(&(cif)->cif_mtx, "carp_if",   \
 253         NULL, MTX_DEF)
 254 #define CIF_LOCK_DESTROY(cif)   mtx_destroy(&(cif)->cif_mtx)
 255 #define CIF_LOCK_ASSERT(cif)    mtx_assert(&(cif)->cif_mtx, MA_OWNED)
 256 #define CIF_LOCK(cif)           mtx_lock(&(cif)->cif_mtx)
 257 #define CIF_UNLOCK(cif)         mtx_unlock(&(cif)->cif_mtx)
 258 #define CIF_FREE(cif)   do {                            \
 259                 CIF_LOCK_ASSERT(cif);                   \
 260                 if (TAILQ_EMPTY(&(cif)->cif_vrs))       \
 261                         carp_free_if(cif);              \
 262                 else                                    \
 263                         CIF_UNLOCK(cif);                \
 264 } while (0)
 265
 266 #define CARP_LOG(...)   do {                            \
 267         if (V_carp_log > 0)                             \
 268                 log(LOG_INFO, "carp: " __VA_ARGS__);    \
 269 } while (0)
 270
 271 #define CARP_DEBUG(...) do {                            \
 272         if (V_carp_log > 1)                             \
 273                 log(LOG_DEBUG, __VA_ARGS__);            \
 274 } while (0)
 275
 276 #define IFNET_FOREACH_IFA(ifp, ifa)                                     \
 277         IF_ADDR_LOCK_ASSERT(ifp);                                       \
 278         TAILQ_FOREACH((ifa), &(ifp)->if_addrhead, ifa_link)             \
 279                 if ((ifa)->ifa_carp != NULL)
 280
 281 #define CARP_FOREACH_IFA(sc, ifa)                                       \
 282         CARP_LOCK_ASSERT(sc);                                           \
 283         for (int _i = 0;                                                \
 284                 _i < (sc)->sc_naddrs + (sc)->sc_naddrs6 &&              \
 285                 ((ifa) = sc->sc_ifas[_i]) != NULL;                      \
 286                 ++_i)
 287
 288 #define IFNET_FOREACH_CARP(ifp, sc)                                     \
 289         CIF_LOCK_ASSERT(ifp->if_carp);                                  \
 290         TAILQ_FOREACH((sc), &(ifp)->if_carp->cif_vrs, sc_list)
 291
 292 #define DEMOTE_ADVSKEW(sc)                                      \
 293     (((sc)->sc_advskew + V_carp_demotion > CARP_MAXSKEW) ?      \
 294     CARP_MAXSKEW : ((sc)->sc_advskew + V_carp_demotion))
 295
 296 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 297 static struct carp_softc
 298                 *carp_alloc(struct ifnet *);
 299 static void     carp_detach_locked(struct ifaddr *);
 300 static void     carp_destroy(struct carp_softc *);
 301 static struct carp_if
 302                 *carp_alloc_if(struct ifnet *);
 303 static void     carp_free_if(struct carp_if *);
 304 static void     carp_set_state(struct carp_softc *, int);
 305 static void     carp_sc_state(struct carp_softc *);
 306 static void     carp_setrun(struct carp_softc *, sa_family_t);
 307 static void     carp_master_down(void *);
 308 static void     carp_master_down_locked(struct carp_softc *);
 309 static void     carp_send_ad(void *);
 310 static void     carp_send_ad_locked(struct carp_softc *);
 311 static void     carp_addroute(struct carp_softc *);
 312 static void     carp_ifa_addroute(struct ifaddr *);
 313 static void     carp_delroute(struct carp_softc *);
 314 static void     carp_ifa_delroute(struct ifaddr *);
 315 static void     carp_send_ad_all(void *, int);
 316 static void     carp_demote_adj(int, char *);
 317
 318 static LIST_HEAD(, carp_softc) carp_list;
 319 static struct mtx carp_mtx;
 320 static struct task carp_sendall_task =
 321     TASK_INITIALIZER(0, carp_send_ad_all, NULL);
 322
 323 static void
 324 carp_hmac_prepare(struct carp_softc *sc)
 325 {
 326         uint8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 327         uint8_t vhid = sc->sc_vhid & 0xff;
 328         struct ifaddr *ifa;
 329         int i, found;
 330 #ifdef INET
 331         struct in_addr last, cur, in;
 332 #endif
 333 #ifdef INET6
 334         struct in6_addr last6, cur6, in6;
 335 #endif
 336
 337         CARP_LOCK_ASSERT(sc);
 338
 339         /* Compute ipad from key. */
 340         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 341         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 342         for (i = 0; i < sizeof(sc->sc_pad); i++)
 343                 sc->sc_pad[i] ^= 0x36;
 344
 345         /* Precompute first part of inner hash. */
 346         SHA1Init(&sc->sc_sha1);
 347         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 348         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 349         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 350         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 351 #ifdef INET
 352         cur.s_addr = 0;
 353         do {
 354                 found = 0;
 355                 last = cur;
 356                 cur.s_addr = 0xffffffff;
 357                 CARP_FOREACH_IFA(sc, ifa) {
 358                         in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 359                         if (ifa->ifa_addr->sa_family == AF_INET &&
 360                             ntohl(in.s_addr) > ntohl(last.s_addr) &&
 361                             ntohl(in.s_addr) < ntohl(cur.s_addr)) {
 362                                 cur.s_addr = in.s_addr;
 363                                 found++;
 364                         }
 365                 }
 366                 if (found)
 367                         SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur));
 368         } while (found);
 369 #endif /* INET */
 370 #ifdef INET6
 371         memset(&cur6, 0, sizeof(cur6));
 372         do {
 373                 found = 0;
 374                 last6 = cur6;
 375                 memset(&cur6, 0xff, sizeof(cur6));
 376                 CARP_FOREACH_IFA(sc, ifa) {
 377                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 378                         if (IN6_IS_SCOPE_EMBED(&in6))
 379                                 in6.s6_addr16[1] = 0;
 380                         if (ifa->ifa_addr->sa_family == AF_INET6 &&
 381                             memcmp(&in6, &last6, sizeof(in6)) > 0 &&
 382                             memcmp(&in6, &cur6, sizeof(in6)) < 0) {
 383                                 cur6 = in6;
 384                                 found++;
 385                         }
 386                 }
 387                 if (found)
 388                         SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6));
 389         } while (found);
 390 #endif /* INET6 */
 391
 392         /* convert ipad to opad */
 393         for (i = 0; i < sizeof(sc->sc_pad); i++)
 394                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 395 }
 396
 397 static void
 398 carp_hmac_generate(struct carp_softc *sc, uint32_t counter[2],
 399     unsigned char md[20])
 400 {
 401         SHA1_CTX sha1ctx;
 402
 403         CARP_LOCK_ASSERT(sc);
 404
 405         /* fetch first half of inner hash */
 406         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 407
 408         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 409         SHA1Final(md, &sha1ctx);
 410
 411         /* outer hash */
 412         SHA1Init(&sha1ctx);
 413         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 414         SHA1Update(&sha1ctx, md, 20);
 415         SHA1Final(md, &sha1ctx);
 416 }
 417
 418 static int
 419 carp_hmac_verify(struct carp_softc *sc, uint32_t counter[2],
 420     unsigned char md[20])
 421 {
 422         unsigned char md2[20];
 423
 424         CARP_LOCK_ASSERT(sc);
 425
 426         carp_hmac_generate(sc, counter, md2);
 427
 428         return (bcmp(md, md2, sizeof(md2)));
 429 }
 430
 431 /*
 432  * process input packet.
 433  * we have rearranged checks order compared to the rfc,
 434  * but it seems more efficient this way or not possible otherwise.
 435  */
 436 #ifdef INET
 437 int
 438 carp_input(struct mbuf **mp, int *offp, int proto)
 439 {
 440         struct mbuf *m = *mp;
 441         struct ip *ip = mtod(m, struct ip *);
 442         struct carp_header *ch;
 443         int iplen, len;
 444
 445         iplen = *offp;
 446         *mp = NULL;
 447
 448         CARPSTATS_INC(carps_ipackets);
 449
 450         if (!V_carp_allow) {
 451                 m_freem(m);
 452                 return (IPPROTO_DONE);
 453         }
 454
 455         /* verify that the IP TTL is 255.  */
 456         if (ip->ip_ttl != CARP_DFLTTL) {
 457                 CARPSTATS_INC(carps_badttl);
 458                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 459                     ip->ip_ttl,
 460                     m->m_pkthdr.rcvif->if_xname);
 461                 m_freem(m);
 462                 return (IPPROTO_DONE);
 463         }
 464
 465         iplen = ip->ip_hl << 2;
 466
 467         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 468                 CARPSTATS_INC(carps_badlen);
 469                 CARP_DEBUG("%s: received len %zd < sizeof(struct carp_header) "
 470                     "on %s\n", __func__, m->m_len - sizeof(struct ip),
 471                     m->m_pkthdr.rcvif->if_xname);
 472                 m_freem(m);
 473                 return (IPPROTO_DONE);
 474         }
 475
 476         if (iplen + sizeof(*ch) < m->m_len) {
 477                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 478                         CARPSTATS_INC(carps_hdrops);
 479                         CARP_DEBUG("%s: pullup failed\n", __func__);
 480                         return (IPPROTO_DONE);
 481                 }
 482                 ip = mtod(m, struct ip *);
 483         }
 484         ch = (struct carp_header *)((char *)ip + iplen);
 485
 486         /*
 487          * verify that the received packet length is
 488          * equal to the CARP header
 489          */
 490         len = iplen + sizeof(*ch);
 491         if (len > m->m_pkthdr.len) {
 492                 CARPSTATS_INC(carps_badlen);
 493                 CARP_DEBUG("%s: packet too short %d on %s\n", __func__,
 494                     m->m_pkthdr.len,
 495                     m->m_pkthdr.rcvif->if_xname);
 496                 m_freem(m);
 497                 return (IPPROTO_DONE);
 498         }
 499
 500         if ((m = m_pullup(m, len)) == NULL) {
 501                 CARPSTATS_INC(carps_hdrops);
 502                 return (IPPROTO_DONE);
 503         }
 504         ip = mtod(m, struct ip *);
 505         ch = (struct carp_header *)((char *)ip + iplen);
 506
 507         /* verify the CARP checksum */
 508         m->m_data += iplen;
 509         if (in_cksum(m, len - iplen)) {
 510                 CARPSTATS_INC(carps_badsum);
 511                 CARP_DEBUG("%s: checksum failed on %s\n", __func__,
 512                     m->m_pkthdr.rcvif->if_xname);
 513                 m_freem(m);
 514                 return (IPPROTO_DONE);
 515         }
 516         m->m_data -= iplen;
 517
 518         carp_input_c(m, ch, AF_INET);
 519         return (IPPROTO_DONE);
 520 }
 521 #endif
 522
 523 #ifdef INET6
 524 int
 525 carp6_input(struct mbuf **mp, int *offp, int proto)
 526 {
 527         struct mbuf *m = *mp;
 528         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 529         struct carp_header *ch;
 530         u_int len;
 531
 532         CARPSTATS_INC(carps_ipackets6);
 533
 534         if (!V_carp_allow) {
 535                 m_freem(m);
 536                 return (IPPROTO_DONE);
 537         }
 538
 539         /* check if received on a valid carp interface */
 540         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 541                 CARPSTATS_INC(carps_badif);
 542                 CARP_DEBUG("%s: packet received on non-carp interface: %s\n",
 543                     __func__, m->m_pkthdr.rcvif->if_xname);
 544                 m_freem(m);
 545                 return (IPPROTO_DONE);
 546         }
 547
 548         /* verify that the IP TTL is 255 */
 549         if (ip6->ip6_hlim != CARP_DFLTTL) {
 550                 CARPSTATS_INC(carps_badttl);
 551                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 552                     ip6->ip6_hlim, m->m_pkthdr.rcvif->if_xname);
 553                 m_freem(m);
 554                 return (IPPROTO_DONE);
 555         }
 556
 557         /* verify that we have a complete carp packet */
 558         len = m->m_len;
 559         IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
 560         if (ch == NULL) {
 561                 CARPSTATS_INC(carps_badlen);
 562                 CARP_DEBUG("%s: packet size %u too small\n", __func__, len);
 563                 return (IPPROTO_DONE);
 564         }
 565
 566
 567         /* verify the CARP checksum */
 568         m->m_data += *offp;
 569         if (in_cksum(m, sizeof(*ch))) {
 570                 CARPSTATS_INC(carps_badsum);
 571                 CARP_DEBUG("%s: checksum failed, on %s\n", __func__,
 572                     m->m_pkthdr.rcvif->if_xname);
 573                 m_freem(m);
 574                 return (IPPROTO_DONE);
 575         }
 576         m->m_data -= *offp;
 577
 578         carp_input_c(m, ch, AF_INET6);
 579         return (IPPROTO_DONE);
 580 }
 581 #endif /* INET6 */
 582
 583 static void
 584 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 585 {
 586         struct ifnet *ifp = m->m_pkthdr.rcvif;
 587         struct ifaddr *ifa;
 588         struct carp_softc *sc;
 589         uint64_t tmp_counter;
 590         struct timeval sc_tv, ch_tv;
 591
 592         /* verify that the VHID is valid on the receiving interface */
 593         IF_ADDR_RLOCK(ifp);
 594         IFNET_FOREACH_IFA(ifp, ifa)
 595                 if (ifa->ifa_addr->sa_family == af &&
 596                     ifa->ifa_carp->sc_vhid == ch->carp_vhid) {
 597                         ifa_ref(ifa);
 598                         break;
 599                 }
 600         IF_ADDR_RUNLOCK(ifp);
 601
 602         if (ifa == NULL) {
 603                 CARPSTATS_INC(carps_badvhid);
 604                 m_freem(m);
 605                 return;
 606         }
 607
 608         /* verify the CARP version. */
 609         if (ch->carp_version != CARP_VERSION) {
 610                 CARPSTATS_INC(carps_badver);
 611                 CARP_DEBUG("%s: invalid version %d\n", ifp->if_xname,
 612                     ch->carp_version);
 613                 ifa_free(ifa);
 614                 m_freem(m);
 615                 return;
 616         }
 617
 618         sc = ifa->ifa_carp;
 619         CARP_LOCK(sc);
 620         ifa_free(ifa);
 621
 622         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 623                 CARPSTATS_INC(carps_badauth);
 624                 CARP_DEBUG("%s: incorrect hash for VHID %u@%s\n", __func__,
 625                     sc->sc_vhid, ifp->if_xname);
 626                 goto out;
 627         }
 628
 629         tmp_counter = ntohl(ch->carp_counter[0]);
 630         tmp_counter = tmp_counter<<32;
 631         tmp_counter += ntohl(ch->carp_counter[1]);
 632
 633         /* XXX Replay protection goes here */
 634
 635         sc->sc_init_counter = 0;
 636         sc->sc_counter = tmp_counter;
 637
 638         sc_tv.tv_sec = sc->sc_advbase;
 639         sc_tv.tv_usec = DEMOTE_ADVSKEW(sc) * 1000000 / 256;
 640         ch_tv.tv_sec = ch->carp_advbase;
 641         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 642
 643         switch (sc->sc_state) {
 644         case INIT:
 645                 break;
 646         case MASTER:
 647                 /*
 648                  * If we receive an advertisement from a master who's going to
 649                  * be more frequent than us, go into BACKUP state.
 650                  */
 651                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 652                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 653                         callout_stop(&sc->sc_ad_tmo);
 654                         CARP_LOG("VHID %u@%s: MASTER -> BACKUP "
 655                             "(more frequent advertisement received)\n",
 656                             sc->sc_vhid,
 657                             sc->sc_carpdev->if_xname);
 658                         carp_set_state(sc, BACKUP);
 659                         carp_setrun(sc, 0);
 660                         carp_delroute(sc);
 661                 }
 662                 break;
 663         case BACKUP:
 664                 /*
 665                  * If we're pre-empting masters who advertise slower than us,
 666                  * and this one claims to be slower, treat him as down.
 667                  */
 668                 if (V_carp_preempt && timevalcmp(&sc_tv, &ch_tv, <)) {
 669                         CARP_LOG("VHID %u@%s: BACKUP -> MASTER "
 670                             "(preempting a slower master)\n",
 671                             sc->sc_vhid,
 672                             sc->sc_carpdev->if_xname);
 673                         carp_master_down_locked(sc);
 674                         break;
 675                 }
 676
 677                 /*
 678                  *  If the master is going to advertise at such a low frequency
 679                  *  that he's guaranteed to time out, we'd might as well just
 680                  *  treat him as timed out now.
 681                  */
 682                 sc_tv.tv_sec = sc->sc_advbase * 3;
 683                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 684                         CARP_LOG("VHID %u@%s: BACKUP -> MASTER "
 685                             "(master timed out)\n",
 686                             sc->sc_vhid,
 687                             sc->sc_carpdev->if_xname);
 688                         carp_master_down_locked(sc);
 689                         break;
 690                 }
 691
 692                 /*
 693                  * Otherwise, we reset the counter and wait for the next
 694                  * advertisement.
 695                  */
 696                 carp_setrun(sc, af);
 697                 break;
 698         }
 699
 700 out:
 701         CARP_UNLOCK(sc);
 702         m_freem(m);
 703 }
 704
 705 static int
 706 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 707 {
 708         struct m_tag *mtag;
 709
 710         if (sc->sc_init_counter) {
 711                 /* this could also be seconds since unix epoch */
 712                 sc->sc_counter = arc4random();
 713                 sc->sc_counter = sc->sc_counter << 32;
 714                 sc->sc_counter += arc4random();
 715         } else
 716                 sc->sc_counter++;
 717
 718         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 719         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 720
 721         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 722
 723         /* Tag packet for carp_output */
 724         if ((mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct carp_softc *),
 725             M_NOWAIT)) == NULL) {
 726                 m_freem(m);
 727                 CARPSTATS_INC(carps_onomem);
 728                 return (ENOMEM);
 729         }
 730         bcopy(&sc, mtag + 1, sizeof(sc));
 731         m_tag_prepend(m, mtag);
 732
 733         return (0);
 734 }
 735
 736 /*
 737  * To avoid LORs and possible recursions this function shouldn't
 738  * be called directly, but scheduled via taskqueue.
 739  */
 740 static void
 741 carp_send_ad_all(void *ctx __unused, int pending __unused)
 742 {
 743         struct carp_softc *sc;
 744
 745         mtx_lock(&carp_mtx);
 746         LIST_FOREACH(sc, &carp_list, sc_next)
 747                 if (sc->sc_state == MASTER) {
 748                         CARP_LOCK(sc);
 749                         CURVNET_SET(sc->sc_carpdev->if_vnet);
 750                         carp_send_ad_locked(sc);
 751                         CURVNET_RESTORE();
 752                         CARP_UNLOCK(sc);
 753                 }
 754         mtx_unlock(&carp_mtx);
 755 }
 756
 757 /* Send a periodic advertisement, executed in callout context. */
 758 static void
 759 carp_send_ad(void *v)
 760 {
 761         struct carp_softc *sc = v;
 762
 763         CARP_LOCK_ASSERT(sc);
 764         CURVNET_SET(sc->sc_carpdev->if_vnet);
 765         carp_send_ad_locked(sc);
 766         CURVNET_RESTORE();
 767         CARP_UNLOCK(sc);
 768 }
 769
 770 static void
 771 carp_send_ad_error(struct carp_softc *sc, int error)
 772 {
 773
 774         if (error) {
 775                 if (sc->sc_sendad_errors < INT_MAX)
 776                         sc->sc_sendad_errors++;
 777                 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 778                         static const char fmt[] = "send error %d on %s";
 779                         char msg[sizeof(fmt) + IFNAMSIZ];
 780
 781                         sprintf(msg, fmt, error, sc->sc_carpdev->if_xname);
 782                         carp_demote_adj(V_carp_senderr_adj, msg);
 783                 }
 784                 sc->sc_sendad_success = 0;
 785         } else {
 786                 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS &&
 787                     ++sc->sc_sendad_success >= CARP_SENDAD_MIN_SUCCESS) {
 788                         static const char fmt[] = "send ok on %s";
 789                         char msg[sizeof(fmt) + IFNAMSIZ];
 790
 791                         sprintf(msg, fmt, sc->sc_carpdev->if_xname);
 792                         carp_demote_adj(-V_carp_senderr_adj, msg);
 793                         sc->sc_sendad_errors = 0;
 794                 } else
 795                         sc->sc_sendad_errors = 0;
 796         }
 797 }
 798
 799 static void
 800 carp_send_ad_locked(struct carp_softc *sc)
 801 {
 802         struct carp_header ch;
 803         struct timeval tv;
 804         struct sockaddr sa;
 805         struct ifaddr *ifa;
 806         struct carp_header *ch_ptr;
 807         struct mbuf *m;
 808         int len, advskew;
 809
 810         CARP_LOCK_ASSERT(sc);
 811
 812         advskew = DEMOTE_ADVSKEW(sc);
 813         tv.tv_sec = sc->sc_advbase;
 814         tv.tv_usec = advskew * 1000000 / 256;
 815
 816         ch.carp_version = CARP_VERSION;
 817         ch.carp_type = CARP_ADVERTISEMENT;
 818         ch.carp_vhid = sc->sc_vhid;
 819         ch.carp_advbase = sc->sc_advbase;
 820         ch.carp_advskew = advskew;
 821         ch.carp_authlen = 7;    /* XXX DEFINE */
 822         ch.carp_pad1 = 0;       /* must be zero */
 823         ch.carp_cksum = 0;
 824
 825         /* XXXGL: OpenBSD picks first ifaddr with needed family. */
 826
 827 #ifdef INET
 828         if (sc->sc_naddrs) {
 829                 struct ip *ip;
 830
 831                 m = m_gethdr(M_NOWAIT, MT_DATA);
 832                 if (m == NULL) {
 833                         CARPSTATS_INC(carps_onomem);
 834                         goto resched;
 835                 }
 836                 len = sizeof(*ip) + sizeof(ch);
 837                 m->m_pkthdr.len = len;
 838                 m->m_pkthdr.rcvif = NULL;
 839                 m->m_len = len;
 840                 MH_ALIGN(m, m->m_len);
 841                 m->m_flags |= M_MCAST;
 842                 ip = mtod(m, struct ip *);
 843                 ip->ip_v = IPVERSION;
 844                 ip->ip_hl = sizeof(*ip) >> 2;
 845                 ip->ip_tos = IPTOS_LOWDELAY;
 846                 ip->ip_len = htons(len);
 847                 ip->ip_id = ip_newid();
 848                 ip->ip_off = htons(IP_DF);
 849                 ip->ip_ttl = CARP_DFLTTL;
 850                 ip->ip_p = IPPROTO_CARP;
 851                 ip->ip_sum = 0;
 852
 853                 bzero(&sa, sizeof(sa));
 854                 sa.sa_family = AF_INET;
 855                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 856                 if (ifa != NULL) {
 857                         ip->ip_src.s_addr =
 858                             ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 859                         ifa_free(ifa);
 860                 } else
 861                         ip->ip_src.s_addr = 0;
 862                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 863
 864                 ch_ptr = (struct carp_header *)(&ip[1]);
 865                 bcopy(&ch, ch_ptr, sizeof(ch));
 866                 if (carp_prepare_ad(m, sc, ch_ptr))
 867                         goto resched;
 868
 869                 m->m_data += sizeof(*ip);
 870                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip));
 871                 m->m_data -= sizeof(*ip);
 872
 873                 CARPSTATS_INC(carps_opackets);
 874
 875                 carp_send_ad_error(sc, ip_output(m, NULL, NULL, IP_RAWOUTPUT,
 876                     &sc->sc_carpdev->if_carp->cif_imo, NULL));
 877         }
 878 #endif /* INET */
 879 #ifdef INET6
 880         if (sc->sc_naddrs6) {
 881                 struct ip6_hdr *ip6;
 882
 883                 m = m_gethdr(M_NOWAIT, MT_DATA);
 884                 if (m == NULL) {
 885                         CARPSTATS_INC(carps_onomem);
 886                         goto resched;
 887                 }
 888                 len = sizeof(*ip6) + sizeof(ch);
 889                 m->m_pkthdr.len = len;
 890                 m->m_pkthdr.rcvif = NULL;
 891                 m->m_len = len;
 892                 MH_ALIGN(m, m->m_len);
 893                 m->m_flags |= M_MCAST;
 894                 ip6 = mtod(m, struct ip6_hdr *);
 895                 bzero(ip6, sizeof(*ip6));
 896                 ip6->ip6_vfc |= IPV6_VERSION;
 897                 ip6->ip6_hlim = CARP_DFLTTL;
 898                 ip6->ip6_nxt = IPPROTO_CARP;
 899                 bzero(&sa, sizeof(sa));
 900
 901                 /* set the source address */
 902                 sa.sa_family = AF_INET6;
 903                 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
 904                 if (ifa != NULL) {
 905                         bcopy(IFA_IN6(ifa), &ip6->ip6_src,
 906                             sizeof(struct in6_addr));
 907                         ifa_free(ifa);
 908                 } else
 909                         /* This should never happen with IPv6. */
 910                         bzero(&ip6->ip6_src, sizeof(struct in6_addr));
 911
 912                 /* Set the multicast destination. */
 913                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
 914                 ip6->ip6_dst.s6_addr8[15] = 0x12;
 915                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
 916                         m_freem(m);
 917                         CARP_DEBUG("%s: in6_setscope failed\n", __func__);
 918                         goto resched;
 919                 }
 920
 921                 ch_ptr = (struct carp_header *)(&ip6[1]);
 922                 bcopy(&ch, ch_ptr, sizeof(ch));
 923                 if (carp_prepare_ad(m, sc, ch_ptr))
 924                         goto resched;
 925
 926                 m->m_data += sizeof(*ip6);
 927                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip6));
 928                 m->m_data -= sizeof(*ip6);
 929
 930                 CARPSTATS_INC(carps_opackets6);
 931
 932                 carp_send_ad_error(sc, ip6_output(m, NULL, NULL, 0,
 933                     &sc->sc_carpdev->if_carp->cif_im6o, NULL, NULL));
 934         }
 935 #endif /* INET6 */
 936
 937 resched:
 938         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), carp_send_ad, sc);
 939 }
 940
 941 static void
 942 carp_addroute(struct carp_softc *sc)
 943 {
 944         struct ifaddr *ifa;
 945
 946         CARP_FOREACH_IFA(sc, ifa)
 947                 carp_ifa_addroute(ifa);
 948 }
 949
 950 static void
 951 carp_ifa_addroute(struct ifaddr *ifa)
 952 {
 953
 954         switch (ifa->ifa_addr->sa_family) {
 955 #ifdef INET
 956         case AF_INET:
 957                 in_addprefix(ifatoia(ifa), RTF_UP);
 958                 ifa_add_loopback_route(ifa,
 959                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 960                 break;
 961 #endif
 962 #ifdef INET6
 963         case AF_INET6:
 964                 ifa_add_loopback_route(ifa,
 965                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 966                 nd6_add_ifa_lle(ifatoia6(ifa));
 967                 break;
 968 #endif
 969         }
 970 }
 971
 972 static void
 973 carp_delroute(struct carp_softc *sc)
 974 {
 975         struct ifaddr *ifa;
 976
 977         CARP_FOREACH_IFA(sc, ifa)
 978                 carp_ifa_delroute(ifa);
 979 }
 980
 981 static void
 982 carp_ifa_delroute(struct ifaddr *ifa)
 983 {
 984
 985         switch (ifa->ifa_addr->sa_family) {
 986 #ifdef INET
 987         case AF_INET:
 988                 ifa_del_loopback_route(ifa,
 989                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
 990                 in_scrubprefix(ifatoia(ifa), LLE_STATIC);
 991                 break;
 992 #endif
 993 #ifdef INET6
 994         case AF_INET6:
 995                 ifa_del_loopback_route(ifa,
 996                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
 997                 nd6_rem_ifa_lle(ifatoia6(ifa));
 998                 break;
 999 #endif
1000         }
1001 }
1002
1003 int
1004 carp_master(struct ifaddr *ifa)
1005 {
1006         struct carp_softc *sc = ifa->ifa_carp;
1007
1008         return (sc->sc_state == MASTER);
1009 }
1010
1011 #ifdef INET
1012 /*
1013  * Broadcast a gratuitous ARP request containing
1014  * the virtual router MAC address for each IP address
1015  * associated with the virtual router.
1016  */
1017 static void
1018 carp_send_arp(struct carp_softc *sc)
1019 {
1020         struct ifaddr *ifa;
1021
1022         CARP_FOREACH_IFA(sc, ifa)
1023                 if (ifa->ifa_addr->sa_family == AF_INET)
1024                         arp_ifinit2(sc->sc_carpdev, ifa, LLADDR(&sc->sc_addr));
1025 }
1026
1027 int
1028 carp_iamatch(struct ifaddr *ifa, uint8_t **enaddr)
1029 {
1030         struct carp_softc *sc = ifa->ifa_carp;
1031
1032         if (sc->sc_state == MASTER) {
1033                 *enaddr = LLADDR(&sc->sc_addr);
1034                 return (1);
1035         }
1036
1037         return (0);
1038 }
1039 #endif
1040
1041 #ifdef INET6
1042 static void
1043 carp_send_na(struct carp_softc *sc)
1044 {
1045         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1046         struct ifaddr *ifa;
1047         struct in6_addr *in6;
1048
1049         CARP_FOREACH_IFA(sc, ifa) {
1050                 if (ifa->ifa_addr->sa_family != AF_INET6)
1051                         continue;
1052
1053                 in6 = IFA_IN6(ifa);
1054                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1055                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1056                 DELAY(1000);    /* XXX */
1057         }
1058 }
1059
1060 /*
1061  * Returns ifa in case it's a carp address and it is MASTER, or if the address
1062  * matches and is not a carp address.  Returns NULL otherwise.
1063  */
1064 struct ifaddr *
1065 carp_iamatch6(struct ifnet *ifp, struct in6_addr *taddr)
1066 {
1067         struct ifaddr *ifa;
1068
1069         ifa = NULL;
1070         IF_ADDR_RLOCK(ifp);
1071         TAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
1072                 if (ifa->ifa_addr->sa_family != AF_INET6)
1073                         continue;
1074                 if (!IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa)))
1075                         continue;
1076                 if (ifa->ifa_carp && ifa->ifa_carp->sc_state != MASTER)
1077                         ifa = NULL;
1078                 else
1079                         ifa_ref(ifa);
1080                 break;
1081         }
1082         IF_ADDR_RUNLOCK(ifp);
1083
1084         return (ifa);
1085 }
1086
1087 caddr_t
1088 carp_macmatch6(struct ifnet *ifp, struct mbuf *m, const struct in6_addr *taddr)
1089 {
1090         struct ifaddr *ifa;
1091
1092         IF_ADDR_RLOCK(ifp);
1093         IFNET_FOREACH_IFA(ifp, ifa)
1094                 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1095                     IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1096                         struct carp_softc *sc = ifa->ifa_carp;
1097                         struct m_tag *mtag;
1098
1099                         IF_ADDR_RUNLOCK(ifp);
1100
1101                         mtag = m_tag_get(PACKET_TAG_CARP,
1102                             sizeof(struct carp_softc *), M_NOWAIT);
1103                         if (mtag == NULL)
1104                                 /* Better a bit than nothing. */
1105                                 return (LLADDR(&sc->sc_addr));
1106
1107                         bcopy(&sc, mtag + 1, sizeof(sc));
1108                         m_tag_prepend(m, mtag);
1109
1110                         return (LLADDR(&sc->sc_addr));
1111                 }
1112         IF_ADDR_RUNLOCK(ifp);
1113
1114         return (NULL);
1115 }
1116 #endif /* INET6 */
1117
1118 int
1119 carp_forus(struct ifnet *ifp, u_char *dhost)
1120 {
1121         struct carp_softc *sc;
1122         uint8_t *ena = dhost;
1123
1124         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1125                 return (0);
1126
1127         CIF_LOCK(ifp->if_carp);
1128         IFNET_FOREACH_CARP(ifp, sc) {
1129                 CARP_LOCK(sc);
1130                 if (sc->sc_state == MASTER && !bcmp(dhost, LLADDR(&sc->sc_addr),
1131                     ETHER_ADDR_LEN)) {
1132                         CARP_UNLOCK(sc);
1133                         CIF_UNLOCK(ifp->if_carp);
1134                         return (1);
1135                 }
1136                 CARP_UNLOCK(sc);
1137         }
1138         CIF_UNLOCK(ifp->if_carp);
1139
1140         return (0);
1141 }
1142
1143 /* Master down timeout event, executed in callout context. */
1144 static void
1145 carp_master_down(void *v)
1146 {
1147         struct carp_softc *sc = v;
1148
1149         CARP_LOCK_ASSERT(sc);
1150
1151         CURVNET_SET(sc->sc_carpdev->if_vnet);
1152         if (sc->sc_state == BACKUP) {
1153                 CARP_LOG("VHID %u@%s: BACKUP -> MASTER (master down)\n",
1154                     sc->sc_vhid,
1155                     sc->sc_carpdev->if_xname);
1156                 carp_master_down_locked(sc);
1157         }
1158         CURVNET_RESTORE();
1159
1160         CARP_UNLOCK(sc);
1161 }
1162
1163 static void
1164 carp_master_down_locked(struct carp_softc *sc)
1165 {
1166
1167         CARP_LOCK_ASSERT(sc);
1168
1169         switch (sc->sc_state) {
1170         case BACKUP:
1171                 carp_set_state(sc, MASTER);
1172                 carp_send_ad_locked(sc);
1173 #ifdef INET
1174                 carp_send_arp(sc);
1175 #endif
1176 #ifdef INET6
1177                 carp_send_na(sc);
1178 #endif
1179                 carp_setrun(sc, 0);
1180                 carp_addroute(sc);
1181                 break;
1182         case INIT:
1183         case MASTER:
1184 #ifdef INVARIANTS
1185                 panic("carp: VHID %u@%s: master_down event in %s state\n",
1186                     sc->sc_vhid,
1187                     sc->sc_carpdev->if_xname,
1188                     sc->sc_state ? "MASTER" : "INIT");
1189 #endif
1190                 break;
1191         }
1192 }
1193
1194 /*
1195  * When in backup state, af indicates whether to reset the master down timer
1196  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1197  */
1198 static void
1199 carp_setrun(struct carp_softc *sc, sa_family_t af)
1200 {
1201         struct timeval tv;
1202
1203         CARP_LOCK_ASSERT(sc);
1204
1205         if ((sc->sc_carpdev->if_flags & IFF_UP) == 0 ||
1206             sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1207             (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0))
1208                 return;
1209
1210         switch (sc->sc_state) {
1211         case INIT:
1212                 CARP_LOG("VHID %u@%s: INIT -> BACKUP\n",
1213                     sc->sc_vhid,
1214                     sc->sc_carpdev->if_xname);
1215                 carp_set_state(sc, BACKUP);
1216                 carp_setrun(sc, 0);
1217                 break;
1218         case BACKUP:
1219                 callout_stop(&sc->sc_ad_tmo);
1220                 tv.tv_sec = 3 * sc->sc_advbase;
1221                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1222                 switch (af) {
1223 #ifdef INET
1224                 case AF_INET:
1225                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1226                             carp_master_down, sc);
1227                         break;
1228 #endif
1229 #ifdef INET6
1230                 case AF_INET6:
1231                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1232                             carp_master_down, sc);
1233                         break;
1234 #endif
1235                 default:
1236 #ifdef INET
1237                         if (sc->sc_naddrs)
1238                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1239                                     carp_master_down, sc);
1240 #endif
1241 #ifdef INET6
1242                         if (sc->sc_naddrs6)
1243                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1244                                     carp_master_down, sc);
1245 #endif
1246                         break;
1247                 }
1248                 break;
1249         case MASTER:
1250                 tv.tv_sec = sc->sc_advbase;
1251                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1252                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1253                     carp_send_ad, sc);
1254                 break;
1255         }
1256 }
1257
1258 /*
1259  * Setup multicast structures.
1260  */
1261 static int
1262 carp_multicast_setup(struct carp_if *cif, sa_family_t sa)
1263 {
1264         struct ifnet *ifp = cif->cif_ifp;
1265         int error = 0;
1266
1267         CIF_LOCK_ASSERT(cif);
1268
1269         switch (sa) {
1270 #ifdef INET
1271         case AF_INET:
1272             {
1273                 struct ip_moptions *imo = &cif->cif_imo;
1274                 struct in_addr addr;
1275
1276                 if (imo->imo_membership)
1277                         return (0);
1278
1279                 imo->imo_membership = (struct in_multi **)malloc(
1280                     (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
1281                     M_NOWAIT);
1282                 if (imo->imo_membership == NULL)
1283                         return (ENOMEM);
1284                 imo->imo_mfilters = NULL;
1285                 imo->imo_max_memberships = IP_MIN_MEMBERSHIPS;
1286                 imo->imo_multicast_vif = -1;
1287
1288                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1289                 if ((error = in_joingroup(ifp, &addr, NULL,
1290                     &imo->imo_membership[0])) != 0) {
1291                         free(imo->imo_membership, M_CARP);
1292                         break;
1293                 }
1294                 imo->imo_num_memberships++;
1295                 imo->imo_multicast_ifp = ifp;
1296                 imo->imo_multicast_ttl = CARP_DFLTTL;
1297                 imo->imo_multicast_loop = 0;
1298                 break;
1299            }
1300 #endif
1301 #ifdef INET6
1302         case AF_INET6:
1303             {
1304                 struct ip6_moptions *im6o = &cif->cif_im6o;
1305                 struct in6_addr in6;
1306                 struct in6_multi *in6m;
1307
1308                 if (im6o->im6o_membership)
1309                         return (0);
1310
1311                 im6o->im6o_membership = (struct in6_multi **)malloc(
1312                     (sizeof(struct in6_multi *) * IPV6_MIN_MEMBERSHIPS), M_CARP,
1313                     M_ZERO | M_NOWAIT);
1314                 if (im6o->im6o_membership == NULL)
1315                         return (ENOMEM);
1316                 im6o->im6o_mfilters = NULL;
1317                 im6o->im6o_max_memberships = IPV6_MIN_MEMBERSHIPS;
1318                 im6o->im6o_multicast_hlim = CARP_DFLTTL;
1319                 im6o->im6o_multicast_ifp = ifp;
1320
1321                 /* Join IPv6 CARP multicast group. */
1322                 bzero(&in6, sizeof(in6));
1323                 in6.s6_addr16[0] = htons(0xff02);
1324                 in6.s6_addr8[15] = 0x12;
1325                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1326                         free(im6o->im6o_membership, M_CARP);
1327                         break;
1328                 }
1329                 in6m = NULL;
1330                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1331                         free(im6o->im6o_membership, M_CARP);
1332                         break;
1333                 }
1334                 im6o->im6o_membership[0] = in6m;
1335                 im6o->im6o_num_memberships++;
1336
1337                 /* Join solicited multicast address. */
1338                 bzero(&in6, sizeof(in6));
1339                 in6.s6_addr16[0] = htons(0xff02);
1340                 in6.s6_addr32[1] = 0;
1341                 in6.s6_addr32[2] = htonl(1);
1342                 in6.s6_addr32[3] = 0;
1343                 in6.s6_addr8[12] = 0xff;
1344                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1345                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1346                         free(im6o->im6o_membership, M_CARP);
1347                         break;
1348                 }
1349                 in6m = NULL;
1350                 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1351                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1352                         free(im6o->im6o_membership, M_CARP);
1353                         break;
1354                 }
1355                 im6o->im6o_membership[1] = in6m;
1356                 im6o->im6o_num_memberships++;
1357                 break;
1358             }
1359 #endif
1360         }
1361
1362         return (error);
1363 }
1364
1365 /*
1366  * Free multicast structures.
1367  */
1368 static void
1369 carp_multicast_cleanup(struct carp_if *cif, sa_family_t sa)
1370 {
1371
1372         CIF_LOCK_ASSERT(cif);
1373         switch (sa) {
1374 #ifdef INET
1375         case AF_INET:
1376                 if (cif->cif_naddrs == 0) {
1377                         struct ip_moptions *imo = &cif->cif_imo;
1378
1379                         in_leavegroup(imo->imo_membership[0], NULL);
1380                         KASSERT(imo->imo_mfilters == NULL,
1381                             ("%s: imo_mfilters != NULL", __func__));
1382                         free(imo->imo_membership, M_CARP);
1383                         imo->imo_membership = NULL;
1384
1385                 }
1386                 break;
1387 #endif
1388 #ifdef INET6
1389         case AF_INET6:
1390                 if (cif->cif_naddrs6 == 0) {
1391                         struct ip6_moptions *im6o = &cif->cif_im6o;
1392
1393                         in6_mc_leave(im6o->im6o_membership[0], NULL);
1394                         in6_mc_leave(im6o->im6o_membership[1], NULL);
1395                         KASSERT(im6o->im6o_mfilters == NULL,
1396                             ("%s: im6o_mfilters != NULL", __func__));
1397                         free(im6o->im6o_membership, M_CARP);
1398                         im6o->im6o_membership = NULL;
1399                 }
1400                 break;
1401 #endif
1402         }
1403 }
1404
1405 int
1406 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa)
1407 {
1408         struct m_tag *mtag;
1409         struct carp_softc *sc;
1410
1411         if (!sa)
1412                 return (0);
1413
1414         switch (sa->sa_family) {
1415 #ifdef INET
1416         case AF_INET:
1417                 break;
1418 #endif
1419 #ifdef INET6
1420         case AF_INET6:
1421                 break;
1422 #endif
1423         default:
1424                 return (0);
1425         }
1426
1427         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
1428         if (mtag == NULL)
1429                 return (0);
1430
1431         bcopy(mtag + 1, &sc, sizeof(sc));
1432
1433         /* Set the source MAC address to the Virtual Router MAC Address. */
1434         switch (ifp->if_type) {
1435         case IFT_ETHER:
1436         case IFT_BRIDGE:
1437         case IFT_L2VLAN: {
1438                         struct ether_header *eh;
1439
1440                         eh = mtod(m, struct ether_header *);
1441                         eh->ether_shost[0] = 0;
1442                         eh->ether_shost[1] = 0;
1443                         eh->ether_shost[2] = 0x5e;
1444                         eh->ether_shost[3] = 0;
1445                         eh->ether_shost[4] = 1;
1446                         eh->ether_shost[5] = sc->sc_vhid;
1447                 }
1448                 break;
1449         case IFT_FDDI: {
1450                         struct fddi_header *fh;
1451
1452                         fh = mtod(m, struct fddi_header *);
1453                         fh->fddi_shost[0] = 0;
1454                         fh->fddi_shost[1] = 0;
1455                         fh->fddi_shost[2] = 0x5e;
1456                         fh->fddi_shost[3] = 0;
1457                         fh->fddi_shost[4] = 1;
1458                         fh->fddi_shost[5] = sc->sc_vhid;
1459                 }
1460                 break;
1461         case IFT_ISO88025: {
1462                         struct iso88025_header *th;
1463                         th = mtod(m, struct iso88025_header *);
1464                         th->iso88025_shost[0] = 3;
1465                         th->iso88025_shost[1] = 0;
1466                         th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
1467                         th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
1468                         th->iso88025_shost[4] = 0;
1469                         th->iso88025_shost[5] = 0;
1470                 }
1471                 break;
1472         default:
1473                 printf("%s: carp is not supported for the %d interface type\n",
1474                     ifp->if_xname, ifp->if_type);
1475                 return (EOPNOTSUPP);
1476         }
1477
1478         return (0);
1479 }
1480
1481 static struct carp_softc*
1482 carp_alloc(struct ifnet *ifp)
1483 {
1484         struct carp_softc *sc;
1485         struct carp_if *cif;
1486
1487         if ((cif = ifp->if_carp) == NULL)
1488                 cif = carp_alloc_if(ifp);
1489
1490         sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
1491
1492         sc->sc_advbase = CARP_DFLTINTV;
1493         sc->sc_vhid = -1;       /* required setting */
1494         sc->sc_init_counter = 1;
1495         sc->sc_state = INIT;
1496
1497         sc->sc_ifasiz = sizeof(struct ifaddr *);
1498         sc->sc_ifas = malloc(sc->sc_ifasiz, M_CARP, M_WAITOK|M_ZERO);
1499         sc->sc_carpdev = ifp;
1500
1501         CARP_LOCK_INIT(sc);
1502 #ifdef INET
1503         callout_init_mtx(&sc->sc_md_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1504 #endif
1505 #ifdef INET6
1506         callout_init_mtx(&sc->sc_md6_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1507 #endif
1508         callout_init_mtx(&sc->sc_ad_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1509
1510         CIF_LOCK(cif);
1511         TAILQ_INSERT_TAIL(&cif->cif_vrs, sc, sc_list);
1512         CIF_UNLOCK(cif);
1513
1514         mtx_lock(&carp_mtx);
1515         LIST_INSERT_HEAD(&carp_list, sc, sc_next);
1516         mtx_unlock(&carp_mtx);
1517
1518         return (sc);
1519 }
1520
1521 static int
1522 carp_grow_ifas(struct carp_softc *sc)
1523 {
1524         struct ifaddr **new;
1525
1526         CARP_LOCK_ASSERT(sc);
1527
1528         new = malloc(sc->sc_ifasiz * 2, M_CARP, M_NOWAIT|M_ZERO);
1529         if (new == NULL)
1530                 return (ENOMEM);
1531         bcopy(sc->sc_ifas, new, sc->sc_ifasiz);
1532         free(sc->sc_ifas, M_CARP);
1533         sc->sc_ifas = new;
1534         sc->sc_ifasiz *= 2;
1535
1536         return (0);
1537 }
1538
1539 static void
1540 carp_destroy(struct carp_softc *sc)
1541 {
1542         struct ifnet *ifp = sc->sc_carpdev;
1543         struct carp_if *cif = ifp->if_carp;
1544
1545         CIF_LOCK_ASSERT(cif);
1546
1547         TAILQ_REMOVE(&cif->cif_vrs, sc, sc_list);
1548
1549         mtx_lock(&carp_mtx);
1550         LIST_REMOVE(sc, sc_next);
1551         mtx_unlock(&carp_mtx);
1552
1553         CARP_LOCK(sc);
1554         if (sc->sc_suppress)
1555                 carp_demote_adj(-V_carp_ifdown_adj, "vhid removed");
1556         callout_drain(&sc->sc_ad_tmo);
1557 #ifdef INET
1558         callout_drain(&sc->sc_md_tmo);
1559 #endif
1560 #ifdef INET6
1561         callout_drain(&sc->sc_md6_tmo);
1562 #endif
1563         CARP_LOCK_DESTROY(sc);
1564
1565         free(sc->sc_ifas, M_CARP);
1566         free(sc, M_CARP);
1567 }
1568
1569 static struct carp_if*
1570 carp_alloc_if(struct ifnet *ifp)
1571 {
1572         struct carp_if *cif;
1573         int error;
1574
1575         cif = malloc(sizeof(*cif), M_CARP, M_WAITOK|M_ZERO);
1576
1577         if ((error = ifpromisc(ifp, 1)) != 0)
1578                 printf("%s: ifpromisc(%s) failed: %d\n",
1579                     __func__, ifp->if_xname, error);
1580         else
1581                 cif->cif_flags |= CIF_PROMISC;
1582
1583         CIF_LOCK_INIT(cif);
1584         cif->cif_ifp = ifp;
1585         TAILQ_INIT(&cif->cif_vrs);
1586
1587         IF_ADDR_WLOCK(ifp);
1588         ifp->if_carp = cif;
1589         if_ref(ifp);
1590         IF_ADDR_WUNLOCK(ifp);
1591
1592         return (cif);
1593 }
1594
1595 static void
1596 carp_free_if(struct carp_if *cif)
1597 {
1598         struct ifnet *ifp = cif->cif_ifp;
1599
1600         CIF_LOCK_ASSERT(cif);
1601         KASSERT(TAILQ_EMPTY(&cif->cif_vrs), ("%s: softc list not empty",
1602             __func__));
1603
1604         IF_ADDR_WLOCK(ifp);
1605         ifp->if_carp = NULL;
1606         IF_ADDR_WUNLOCK(ifp);
1607
1608         CIF_LOCK_DESTROY(cif);
1609
1610         if (cif->cif_flags & CIF_PROMISC)
1611                 ifpromisc(ifp, 0);
1612         if_rele(ifp);
1613
1614         free(cif, M_CARP);
1615 }
1616
1617 static void
1618 carp_carprcp(struct carpreq *carpr, struct carp_softc *sc, int priv)
1619 {
1620
1621         CARP_LOCK(sc);
1622         carpr->carpr_state = sc->sc_state;
1623         carpr->carpr_vhid = sc->sc_vhid;
1624         carpr->carpr_advbase = sc->sc_advbase;
1625         carpr->carpr_advskew = sc->sc_advskew;
1626         if (priv)
1627                 bcopy(sc->sc_key, carpr->carpr_key, sizeof(carpr->carpr_key));
1628         else
1629                 bzero(carpr->carpr_key, sizeof(carpr->carpr_key));
1630         CARP_UNLOCK(sc);
1631 }
1632
1633 int
1634 carp_ioctl(struct ifreq *ifr, u_long cmd, struct thread *td)
1635 {
1636         struct carpreq carpr;
1637         struct ifnet *ifp;
1638         struct carp_softc *sc = NULL;
1639         int error = 0, locked = 0;
1640
1641         if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1642                 return (error);
1643
1644         ifp = ifunit_ref(ifr->ifr_name);
1645         if (ifp == NULL)
1646                 return (ENXIO);
1647
1648         switch (ifp->if_type) {
1649         case IFT_ETHER:
1650         case IFT_L2VLAN:
1651         case IFT_BRIDGE:
1652         case IFT_FDDI:
1653         case IFT_ISO88025:
1654                 break;
1655         default:
1656                 error = EOPNOTSUPP;
1657                 goto out;
1658         }
1659
1660         if ((ifp->if_flags & IFF_MULTICAST) == 0) {
1661                 error = EADDRNOTAVAIL;
1662                 goto out;
1663         }
1664
1665         switch (cmd) {
1666         case SIOCSVH:
1667                 if ((error = priv_check(td, PRIV_NETINET_CARP)))
1668                         break;
1669                 if (carpr.carpr_vhid <= 0 || carpr.carpr_vhid > CARP_MAXVHID ||
1670                     carpr.carpr_advbase < 0 || carpr.carpr_advskew < 0) {
1671                         error = EINVAL;
1672                         break;
1673                 }
1674
1675                 if (ifp->if_carp) {
1676                         CIF_LOCK(ifp->if_carp);
1677                         IFNET_FOREACH_CARP(ifp, sc)
1678                                 if (sc->sc_vhid == carpr.carpr_vhid)
1679                                         break;
1680                         CIF_UNLOCK(ifp->if_carp);
1681                 }
1682                 if (sc == NULL) {
1683                         sc = carp_alloc(ifp);
1684                         CARP_LOCK(sc);
1685                         sc->sc_vhid = carpr.carpr_vhid;
1686                         LLADDR(&sc->sc_addr)[0] = 0;
1687                         LLADDR(&sc->sc_addr)[1] = 0;
1688                         LLADDR(&sc->sc_addr)[2] = 0x5e;
1689                         LLADDR(&sc->sc_addr)[3] = 0;
1690                         LLADDR(&sc->sc_addr)[4] = 1;
1691                         LLADDR(&sc->sc_addr)[5] = sc->sc_vhid;
1692                 } else
1693                         CARP_LOCK(sc);
1694                 locked = 1;
1695                 if (carpr.carpr_advbase > 0) {
1696                         if (carpr.carpr_advbase > 255 ||
1697                             carpr.carpr_advbase < CARP_DFLTINTV) {
1698                                 error = EINVAL;
1699                                 break;
1700                         }
1701                         sc->sc_advbase = carpr.carpr_advbase;
1702                 }
1703                 if (carpr.carpr_advskew > 0) {
1704                         if (carpr.carpr_advskew >= 255) {
1705                                 error = EINVAL;
1706                                 break;
1707                         }
1708                         sc->sc_advskew = carpr.carpr_advskew;
1709                 }
1710                 if (carpr.carpr_key[0] != '\0') {
1711                         bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1712                         carp_hmac_prepare(sc);
1713                 }
1714                 if (sc->sc_state != INIT &&
1715                     carpr.carpr_state != sc->sc_state) {
1716                         switch (carpr.carpr_state) {
1717                         case BACKUP:
1718                                 callout_stop(&sc->sc_ad_tmo);
1719                                 carp_set_state(sc, BACKUP);
1720                                 carp_setrun(sc, 0);
1721                                 carp_delroute(sc);
1722                                 break;
1723                         case MASTER:
1724                                 carp_master_down_locked(sc);
1725                                 break;
1726                         default:
1727                                 break;
1728                         }
1729                 }
1730                 break;
1731
1732         case SIOCGVH:
1733             {
1734                 int priveleged;
1735
1736                 if (carpr.carpr_vhid < 0 || carpr.carpr_vhid > CARP_MAXVHID) {
1737                         error = EINVAL;
1738                         break;
1739                 }
1740                 if (carpr.carpr_count < 1) {
1741                         error = EMSGSIZE;
1742                         break;
1743                 }
1744                 if (ifp->if_carp == NULL) {
1745                         error = ENOENT;
1746                         break;
1747                 }
1748
1749                 priveleged = (priv_check(td, PRIV_NETINET_CARP) == 0);
1750                 if (carpr.carpr_vhid != 0) {
1751                         CIF_LOCK(ifp->if_carp);
1752                         IFNET_FOREACH_CARP(ifp, sc)
1753                                 if (sc->sc_vhid == carpr.carpr_vhid)
1754                                         break;
1755                         CIF_UNLOCK(ifp->if_carp);
1756                         if (sc == NULL) {
1757                                 error = ENOENT;
1758                                 break;
1759                         }
1760                         carp_carprcp(&carpr, sc, priveleged);
1761                         error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1762                 } else  {
1763                         int i, count;
1764
1765                         count = 0;
1766                         CIF_LOCK(ifp->if_carp);
1767                         IFNET_FOREACH_CARP(ifp, sc)
1768                                 count++;
1769
1770                         if (count > carpr.carpr_count) {
1771                                 CIF_UNLOCK(ifp->if_carp);
1772                                 error = EMSGSIZE;
1773                                 break;
1774                         }
1775
1776                         i = 0;
1777                         IFNET_FOREACH_CARP(ifp, sc) {
1778                                 carp_carprcp(&carpr, sc, priveleged);
1779                                 carpr.carpr_count = count;
1780                                 error = copyout(&carpr, ifr->ifr_data +
1781                                     (i * sizeof(carpr)), sizeof(carpr));
1782                                 if (error) {
1783                                         CIF_UNLOCK(ifp->if_carp);
1784                                         break;
1785                                 }
1786                                 i++;
1787                         }
1788                         CIF_UNLOCK(ifp->if_carp);
1789                 }
1790                 break;
1791             }
1792         default:
1793                 error = EINVAL;
1794         }
1795
1796 out:
1797         if (locked)
1798                 CARP_UNLOCK(sc);
1799         if_rele(ifp);
1800
1801         return (error);
1802 }
1803
1804 static int
1805 carp_get_vhid(struct ifaddr *ifa)
1806 {
1807
1808         if (ifa == NULL || ifa->ifa_carp == NULL)
1809                 return (0);
1810
1811         return (ifa->ifa_carp->sc_vhid);
1812 }
1813
1814 int
1815 carp_attach(struct ifaddr *ifa, int vhid)
1816 {
1817         struct ifnet *ifp = ifa->ifa_ifp;
1818         struct carp_if *cif = ifp->if_carp;
1819         struct carp_softc *sc;
1820         int index, error;
1821
1822         if (ifp->if_carp == NULL)
1823                 return (ENOPROTOOPT);
1824
1825         switch (ifa->ifa_addr->sa_family) {
1826 #ifdef INET
1827         case AF_INET:
1828 #endif
1829 #ifdef INET6
1830         case AF_INET6:
1831 #endif
1832                 break;
1833         default:
1834                 return (EPROTOTYPE);
1835         }
1836
1837         CIF_LOCK(cif);
1838         IFNET_FOREACH_CARP(ifp, sc)
1839                 if (sc->sc_vhid == vhid)
1840                         break;
1841         if (sc == NULL) {
1842                 CIF_UNLOCK(cif);
1843                 return (ENOENT);
1844         }
1845
1846         if (ifa->ifa_carp) {
1847                 if (ifa->ifa_carp->sc_vhid != vhid)
1848                         carp_detach_locked(ifa);
1849                 else {
1850                         CIF_UNLOCK(cif);
1851                         return (0);
1852                 }
1853         }
1854
1855         error = carp_multicast_setup(cif, ifa->ifa_addr->sa_family);
1856         if (error) {
1857                 CIF_FREE(cif);
1858                 return (error);
1859         }
1860
1861         CARP_LOCK(sc);
1862         index = sc->sc_naddrs + sc->sc_naddrs6 + 1;
1863         if (index > sc->sc_ifasiz / sizeof(struct ifaddr *))
1864                 if ((error = carp_grow_ifas(sc)) != 0) {
1865                         carp_multicast_cleanup(cif,
1866                             ifa->ifa_addr->sa_family);
1867                         CARP_UNLOCK(sc);
1868                         CIF_FREE(cif);
1869                         return (error);
1870                 }
1871
1872         switch (ifa->ifa_addr->sa_family) {
1873 #ifdef INET
1874         case AF_INET:
1875                 cif->cif_naddrs++;
1876                 sc->sc_naddrs++;
1877                 break;
1878 #endif
1879 #ifdef INET6
1880         case AF_INET6:
1881                 cif->cif_naddrs6++;
1882                 sc->sc_naddrs6++;
1883                 break;
1884 #endif
1885         }
1886
1887         ifa_ref(ifa);
1888         sc->sc_ifas[index - 1] = ifa;
1889         ifa->ifa_carp = sc;
1890
1891         carp_hmac_prepare(sc);
1892         carp_sc_state(sc);
1893
1894         CARP_UNLOCK(sc);
1895         CIF_UNLOCK(cif);
1896
1897         return (0);
1898 }
1899
1900 void
1901 carp_detach(struct ifaddr *ifa)
1902 {
1903         struct ifnet *ifp = ifa->ifa_ifp;
1904         struct carp_if *cif = ifp->if_carp;
1905
1906         CIF_LOCK(cif);
1907         carp_detach_locked(ifa);
1908         CIF_FREE(cif);
1909 }
1910
1911 static void
1912 carp_detach_locked(struct ifaddr *ifa)
1913 {
1914         struct ifnet *ifp = ifa->ifa_ifp;
1915         struct carp_if *cif = ifp->if_carp;
1916         struct carp_softc *sc = ifa->ifa_carp;
1917         int i, index;
1918
1919         KASSERT(sc != NULL, ("%s: %p not attached", __func__, ifa));
1920
1921         CIF_LOCK_ASSERT(cif);
1922         CARP_LOCK(sc);
1923
1924         /* Shift array. */
1925         index = sc->sc_naddrs + sc->sc_naddrs6;
1926         for (i = 0; i < index; i++)
1927                 if (sc->sc_ifas[i] == ifa)
1928                         break;
1929         KASSERT(i < index, ("%s: %p no backref", __func__, ifa));
1930         for (; i < index - 1; i++)
1931                 sc->sc_ifas[i] = sc->sc_ifas[i+1];
1932         sc->sc_ifas[index - 1] = NULL;
1933
1934         switch (ifa->ifa_addr->sa_family) {
1935 #ifdef INET
1936         case AF_INET:
1937                 cif->cif_naddrs--;
1938                 sc->sc_naddrs--;
1939                 break;
1940 #endif
1941 #ifdef INET6
1942         case AF_INET6:
1943                 cif->cif_naddrs6--;
1944                 sc->sc_naddrs6--;
1945                 break;
1946 #endif
1947         }
1948
1949         carp_ifa_delroute(ifa);
1950         carp_multicast_cleanup(cif, ifa->ifa_addr->sa_family);
1951
1952         ifa->ifa_carp = NULL;
1953         ifa_free(ifa);
1954
1955         carp_hmac_prepare(sc);
1956         carp_sc_state(sc);
1957
1958         if (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) {
1959                 CARP_UNLOCK(sc);
1960                 carp_destroy(sc);
1961         } else
1962                 CARP_UNLOCK(sc);
1963 }
1964
1965 static void
1966 carp_set_state(struct carp_softc *sc, int state)
1967 {
1968
1969         CARP_LOCK_ASSERT(sc);
1970
1971         if (sc->sc_state != state) {
1972                 const char *carp_states[] = { CARP_STATES };
1973                 char subsys[IFNAMSIZ+5];
1974
1975                 sc->sc_state = state;
1976
1977                 snprintf(subsys, IFNAMSIZ+5, "%u@%s", sc->sc_vhid,
1978                     sc->sc_carpdev->if_xname);
1979                 devctl_notify("CARP", subsys, carp_states[state], NULL);
1980         }
1981 }
1982
1983 static void
1984 carp_linkstate(struct ifnet *ifp)
1985 {
1986         struct carp_softc *sc;
1987
1988         CIF_LOCK(ifp->if_carp);
1989         IFNET_FOREACH_CARP(ifp, sc) {
1990                 CARP_LOCK(sc);
1991                 carp_sc_state(sc);
1992                 CARP_UNLOCK(sc);
1993         }
1994         CIF_UNLOCK(ifp->if_carp);
1995 }
1996
1997 static void
1998 carp_sc_state(struct carp_softc *sc)
1999 {
2000
2001         CARP_LOCK_ASSERT(sc);
2002
2003         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
2004             !(sc->sc_carpdev->if_flags & IFF_UP)) {
2005                 callout_stop(&sc->sc_ad_tmo);
2006 #ifdef INET
2007                 callout_stop(&sc->sc_md_tmo);
2008 #endif
2009 #ifdef INET6
2010                 callout_stop(&sc->sc_md6_tmo);
2011 #endif
2012                 carp_set_state(sc, INIT);
2013                 carp_setrun(sc, 0);
2014                 if (!sc->sc_suppress)
2015                         carp_demote_adj(V_carp_ifdown_adj, "interface down");
2016                 sc->sc_suppress = 1;
2017         } else {
2018                 carp_set_state(sc, INIT);
2019                 carp_setrun(sc, 0);
2020                 if (sc->sc_suppress)
2021                         carp_demote_adj(-V_carp_ifdown_adj, "interface up");
2022                 sc->sc_suppress = 0;
2023         }
2024 }
2025
2026 static void
2027 carp_demote_adj(int adj, char *reason)
2028 {
2029         atomic_add_int(&V_carp_demotion, adj);
2030         CARP_LOG("demoted by %d to %d (%s)\n", adj, V_carp_demotion, reason);
2031         taskqueue_enqueue(taskqueue_swi, &carp_sendall_task);
2032 }
2033
2034 static int
2035 carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS)
2036 {
2037         int new, error;
2038
2039         new = V_carp_demotion;
2040         error = sysctl_handle_int(oidp, &new, 0, req);
2041         if (error || !req->newptr)
2042                 return (error);
2043
2044         carp_demote_adj(new, "sysctl");
2045
2046         return (0);
2047 }
2048
2049 #ifdef INET
2050 extern  struct domain inetdomain;
2051 static struct protosw in_carp_protosw = {
2052         .pr_type =              SOCK_RAW,
2053         .pr_domain =            &inetdomain,
2054         .pr_protocol =          IPPROTO_CARP,
2055         .pr_flags =             PR_ATOMIC|PR_ADDR,
2056         .pr_input =             carp_input,
2057         .pr_output =            rip_output,
2058         .pr_ctloutput =         rip_ctloutput,
2059         .pr_usrreqs =           &rip_usrreqs
2060 };
2061 #endif
2062
2063 #ifdef INET6
2064 extern  struct domain inet6domain;
2065 static struct protosw in6_carp_protosw = {
2066         .pr_type =              SOCK_RAW,
2067         .pr_domain =            &inet6domain,
2068         .pr_protocol =          IPPROTO_CARP,
2069         .pr_flags =             PR_ATOMIC|PR_ADDR,
2070         .pr_input =             carp6_input,
2071         .pr_output =            rip6_output,
2072         .pr_ctloutput =         rip6_ctloutput,
2073         .pr_usrreqs =           &rip6_usrreqs
2074 };
2075 #endif
2076
2077 static void
2078 carp_mod_cleanup(void)
2079 {
2080
2081 #ifdef INET
2082         if (proto_reg[CARP_INET] == 0) {
2083                 (void)ipproto_unregister(IPPROTO_CARP);
2084                 pf_proto_unregister(PF_INET, IPPROTO_CARP, SOCK_RAW);
2085                 proto_reg[CARP_INET] = -1;
2086         }
2087         carp_iamatch_p = NULL;
2088 #endif
2089 #ifdef INET6
2090         if (proto_reg[CARP_INET6] == 0) {
2091                 (void)ip6proto_unregister(IPPROTO_CARP);
2092                 pf_proto_unregister(PF_INET6, IPPROTO_CARP, SOCK_RAW);
2093                 proto_reg[CARP_INET6] = -1;
2094         }
2095         carp_iamatch6_p = NULL;
2096         carp_macmatch6_p = NULL;
2097 #endif
2098         carp_ioctl_p = NULL;
2099         carp_attach_p = NULL;
2100         carp_detach_p = NULL;
2101         carp_get_vhid_p = NULL;
2102         carp_linkstate_p = NULL;
2103         carp_forus_p = NULL;
2104         carp_output_p = NULL;
2105         carp_demote_adj_p = NULL;
2106         carp_master_p = NULL;
2107         mtx_unlock(&carp_mtx);
2108         taskqueue_drain(taskqueue_swi, &carp_sendall_task);
2109         mtx_destroy(&carp_mtx);
2110 }
2111
2112 static int
2113 carp_mod_load(void)
2114 {
2115         int err;
2116
2117         mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2118         LIST_INIT(&carp_list);
2119         carp_get_vhid_p = carp_get_vhid;
2120         carp_forus_p = carp_forus;
2121         carp_output_p = carp_output;
2122         carp_linkstate_p = carp_linkstate;
2123         carp_ioctl_p = carp_ioctl;
2124         carp_attach_p = carp_attach;
2125         carp_detach_p = carp_detach;
2126         carp_demote_adj_p = carp_demote_adj;
2127         carp_master_p = carp_master;
2128 #ifdef INET6
2129         carp_iamatch6_p = carp_iamatch6;
2130         carp_macmatch6_p = carp_macmatch6;
2131         proto_reg[CARP_INET6] = pf_proto_register(PF_INET6,
2132             (struct protosw *)&in6_carp_protosw);
2133         if (proto_reg[CARP_INET6]) {
2134                 printf("carp: error %d attaching to PF_INET6\n",
2135                     proto_reg[CARP_INET6]);
2136                 carp_mod_cleanup();
2137                 return (proto_reg[CARP_INET6]);
2138         }
2139         err = ip6proto_register(IPPROTO_CARP);
2140         if (err) {
2141                 printf("carp: error %d registering with INET6\n", err);
2142                 carp_mod_cleanup();
2143                 return (err);
2144         }
2145 #endif
2146 #ifdef INET
2147         carp_iamatch_p = carp_iamatch;
2148         proto_reg[CARP_INET] = pf_proto_register(PF_INET, &in_carp_protosw);
2149         if (proto_reg[CARP_INET]) {
2150                 printf("carp: error %d attaching to PF_INET\n",
2151                     proto_reg[CARP_INET]);
2152                 carp_mod_cleanup();
2153                 return (proto_reg[CARP_INET]);
2154         }
2155         err = ipproto_register(IPPROTO_CARP);
2156         if (err) {
2157                 printf("carp: error %d registering with INET\n", err);
2158                 carp_mod_cleanup();
2159                 return (err);
2160         }
2161 #endif
2162         return (0);
2163 }
2164
2165 static int
2166 carp_modevent(module_t mod, int type, void *data)
2167 {
2168         switch (type) {
2169         case MOD_LOAD:
2170                 return carp_mod_load();
2171                 /* NOTREACHED */
2172         case MOD_UNLOAD:
2173                 mtx_lock(&carp_mtx);
2174                 if (LIST_EMPTY(&carp_list))
2175                         carp_mod_cleanup();
2176                 else {
2177                         mtx_unlock(&carp_mtx);
2178                         return (EBUSY);
2179                 }
2180                 break;
2181
2182         default:
2183                 return (EINVAL);
2184         }
2185
2186         return (0);
2187 }
2188
2189 static moduledata_t carp_mod = {
2190         "carp",
2191         carp_modevent,
2192         0
2193 };
2194
2195 DECLARE_MODULE(carp, carp_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);