sys/netinet/ip_carp.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2002 Michael Shalayeff.
   5  * Copyright (c) 2003 Ryan McBride.
   6  * Copyright (c) 2011 Gleb Smirnoff <glebius@FreeBSD.org>
   7  * All rights reserved.
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  *
  18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  19  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  20  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  21  * IN NO EVENT SHALL THE AUTHOR OR HIS RELATIVES BE LIABLE FOR ANY DIRECT,
  22  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  23  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  24  * SERVICES; LOSS OF MIND, USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
  26  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
  27  * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
  28  * THE POSSIBILITY OF SUCH DAMAGE.
  29  */
  30
  31 #include <sys/cdefs.h>
  32 __FBSDID("$FreeBSD$");
  33
  34 #include "opt_bpf.h"
  35 #include "opt_inet.h"
  36 #include "opt_inet6.h"
  37
  38 #include <sys/param.h>
  39 #include <sys/systm.h>
  40 #include <sys/bus.h>
  41 #include <sys/jail.h>
  42 #include <sys/kernel.h>
  43 #include <sys/limits.h>
  44 #include <sys/malloc.h>
  45 #include <sys/mbuf.h>
  46 #include <sys/module.h>
  47 #include <sys/priv.h>
  48 #include <sys/proc.h>
  49 #include <sys/protosw.h>
  50 #include <sys/socket.h>
  51 #include <sys/sockio.h>
  52 #include <sys/sysctl.h>
  53 #include <sys/syslog.h>
  54 #include <sys/taskqueue.h>
  55 #include <sys/counter.h>
  56
  57 #include <net/ethernet.h>
  58 #include <net/if.h>
  59 #include <net/if_var.h>
  60 #include <net/if_dl.h>
  61 #include <net/if_llatbl.h>
  62 #include <net/if_types.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() or carp_sx.
 179  * To traverse the global list we use the mutex carp_mtx.
 180  *
 181  * Known issues with locking:
 182  *
 183  * - Sending ad, we put the pointer to the softc in an mtag, and no reference
 184  *   counting is done on the softc.
 185  * - On module unload we may race (?) with packet processing thread
 186  *   dereferencing our function pointers.
 187  */
 188
 189 /* Accept incoming CARP packets. */
 190 VNET_DEFINE_STATIC(int, carp_allow) = 1;
 191 #define V_carp_allow    VNET(carp_allow)
 192
 193 /* Set DSCP in outgoing CARP packets. */
 194 VNET_DEFINE_STATIC(int, carp_dscp) = 56;
 195 #define V_carp_dscp     VNET(carp_dscp)
 196
 197 /* Preempt slower nodes. */
 198 VNET_DEFINE_STATIC(int, carp_preempt) = 0;
 199 #define V_carp_preempt  VNET(carp_preempt)
 200
 201 /* Log level. */
 202 VNET_DEFINE_STATIC(int, carp_log) = 1;
 203 #define V_carp_log      VNET(carp_log)
 204
 205 /* Global advskew demotion. */
 206 VNET_DEFINE_STATIC(int, carp_demotion) = 0;
 207 #define V_carp_demotion VNET(carp_demotion)
 208
 209 /* Send error demotion factor. */
 210 VNET_DEFINE_STATIC(int, carp_senderr_adj) = CARP_MAXSKEW;
 211 #define V_carp_senderr_adj      VNET(carp_senderr_adj)
 212
 213 /* Iface down demotion factor. */
 214 VNET_DEFINE_STATIC(int, carp_ifdown_adj) = CARP_MAXSKEW;
 215 #define V_carp_ifdown_adj       VNET(carp_ifdown_adj)
 216
 217 static int carp_allow_sysctl(SYSCTL_HANDLER_ARGS);
 218 static int carp_dscp_sysctl(SYSCTL_HANDLER_ARGS);
 219 static int carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS);
 220
 221 SYSCTL_NODE(_net_inet, IPPROTO_CARP,    carp,   CTLFLAG_RW, 0,  "CARP");
 222 SYSCTL_PROC(_net_inet_carp, OID_AUTO, allow,
 223     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW, 0, 0, carp_allow_sysctl, "I",
 224     "Accept incoming CARP packets");
 225 SYSCTL_PROC(_net_inet_carp, OID_AUTO, dscp,
 226     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW, 0, 0, carp_dscp_sysctl, "I",
 227     "DSCP value for carp packets");
 228 SYSCTL_INT(_net_inet_carp, OID_AUTO, preempt, CTLFLAG_VNET | CTLFLAG_RW,
 229     &VNET_NAME(carp_preempt), 0, "High-priority backup preemption mode");
 230 SYSCTL_INT(_net_inet_carp, OID_AUTO, log, CTLFLAG_VNET | CTLFLAG_RW,
 231     &VNET_NAME(carp_log), 0, "CARP log level");
 232 SYSCTL_PROC(_net_inet_carp, OID_AUTO, demotion,
 233     CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW,
 234     0, 0, carp_demote_adj_sysctl, "I",
 235     "Adjust demotion factor (skew of advskew)");
 236 SYSCTL_INT(_net_inet_carp, OID_AUTO, senderr_demotion_factor,
 237     CTLFLAG_VNET | CTLFLAG_RW,
 238     &VNET_NAME(carp_senderr_adj), 0, "Send error demotion factor adjustment");
 239 SYSCTL_INT(_net_inet_carp, OID_AUTO, ifdown_demotion_factor,
 240     CTLFLAG_VNET | CTLFLAG_RW,
 241     &VNET_NAME(carp_ifdown_adj), 0,
 242     "Interface down demotion factor adjustment");
 243
 244 VNET_PCPUSTAT_DEFINE(struct carpstats, carpstats);
 245 VNET_PCPUSTAT_SYSINIT(carpstats);
 246 VNET_PCPUSTAT_SYSUNINIT(carpstats);
 247
 248 #define CARPSTATS_ADD(name, val)        \
 249     counter_u64_add(VNET(carpstats)[offsetof(struct carpstats, name) / \
 250         sizeof(uint64_t)], (val))
 251 #define CARPSTATS_INC(name)             CARPSTATS_ADD(name, 1)
 252
 253 SYSCTL_VNET_PCPUSTAT(_net_inet_carp, OID_AUTO, stats, struct carpstats,
 254     carpstats, "CARP statistics (struct carpstats, netinet/ip_carp.h)");
 255
 256 #define CARP_LOCK_INIT(sc)      mtx_init(&(sc)->sc_mtx, "carp_softc",   \
 257         NULL, MTX_DEF)
 258 #define CARP_LOCK_DESTROY(sc)   mtx_destroy(&(sc)->sc_mtx)
 259 #define CARP_LOCK_ASSERT(sc)    mtx_assert(&(sc)->sc_mtx, MA_OWNED)
 260 #define CARP_LOCK(sc)           mtx_lock(&(sc)->sc_mtx)
 261 #define CARP_UNLOCK(sc)         mtx_unlock(&(sc)->sc_mtx)
 262 #define CIF_LOCK_INIT(cif)      mtx_init(&(cif)->cif_mtx, "carp_if",   \
 263         NULL, MTX_DEF)
 264 #define CIF_LOCK_DESTROY(cif)   mtx_destroy(&(cif)->cif_mtx)
 265 #define CIF_LOCK_ASSERT(cif)    mtx_assert(&(cif)->cif_mtx, MA_OWNED)
 266 #define CIF_LOCK(cif)           mtx_lock(&(cif)->cif_mtx)
 267 #define CIF_UNLOCK(cif)         mtx_unlock(&(cif)->cif_mtx)
 268 #define CIF_FREE(cif)   do {                            \
 269                 CIF_LOCK(cif);                          \
 270                 if (TAILQ_EMPTY(&(cif)->cif_vrs))       \
 271                         carp_free_if(cif);              \
 272                 else                                    \
 273                         CIF_UNLOCK(cif);                \
 274 } while (0)
 275
 276 #define CARP_LOG(...)   do {                            \
 277         if (V_carp_log > 0)                             \
 278                 log(LOG_INFO, "carp: " __VA_ARGS__);    \
 279 } while (0)
 280
 281 #define CARP_DEBUG(...) do {                            \
 282         if (V_carp_log > 1)                             \
 283                 log(LOG_DEBUG, __VA_ARGS__);            \
 284 } while (0)
 285
 286 #define IFNET_FOREACH_IFA(ifp, ifa)                                     \
 287         CK_STAILQ_FOREACH((ifa), &(ifp)->if_addrhead, ifa_link) \
 288                 if ((ifa)->ifa_carp != NULL)
 289
 290 #define CARP_FOREACH_IFA(sc, ifa)                                       \
 291         CARP_LOCK_ASSERT(sc);                                           \
 292         for (int _i = 0;                                                \
 293                 _i < (sc)->sc_naddrs + (sc)->sc_naddrs6 &&              \
 294                 ((ifa) = sc->sc_ifas[_i]) != NULL;                      \
 295                 ++_i)
 296
 297 #define IFNET_FOREACH_CARP(ifp, sc)                                     \
 298         KASSERT(mtx_owned(&ifp->if_carp->cif_mtx) ||                    \
 299             sx_xlocked(&carp_sx), ("cif_vrs not locked"));              \
 300         TAILQ_FOREACH((sc), &(ifp)->if_carp->cif_vrs, sc_list)
 301
 302 #define DEMOTE_ADVSKEW(sc)                                      \
 303     (((sc)->sc_advskew + V_carp_demotion > CARP_MAXSKEW) ?      \
 304     CARP_MAXSKEW : ((sc)->sc_advskew + V_carp_demotion))
 305
 306 static void     carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
 307 static struct carp_softc
 308                 *carp_alloc(struct ifnet *);
 309 static void     carp_destroy(struct carp_softc *);
 310 static struct carp_if
 311                 *carp_alloc_if(struct ifnet *);
 312 static void     carp_free_if(struct carp_if *);
 313 static void     carp_set_state(struct carp_softc *, int, const char* reason);
 314 static void     carp_sc_state(struct carp_softc *);
 315 static void     carp_setrun(struct carp_softc *, sa_family_t);
 316 static void     carp_master_down(void *);
 317 static void     carp_master_down_locked(struct carp_softc *,
 318                     const char* reason);
 319 static void     carp_send_ad(void *);
 320 static void     carp_send_ad_locked(struct carp_softc *);
 321 static void     carp_addroute(struct carp_softc *);
 322 static void     carp_ifa_addroute(struct ifaddr *);
 323 static void     carp_delroute(struct carp_softc *);
 324 static void     carp_ifa_delroute(struct ifaddr *);
 325 static void     carp_send_ad_all(void *, int);
 326 static void     carp_demote_adj(int, char *);
 327
 328 static LIST_HEAD(, carp_softc) carp_list;
 329 static struct mtx carp_mtx;
 330 static struct sx carp_sx;
 331 static struct task carp_sendall_task =
 332     TASK_INITIALIZER(0, carp_send_ad_all, NULL);
 333
 334 static void
 335 carp_hmac_prepare(struct carp_softc *sc)
 336 {
 337         uint8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
 338         uint8_t vhid = sc->sc_vhid & 0xff;
 339         struct ifaddr *ifa;
 340         int i, found;
 341 #ifdef INET
 342         struct in_addr last, cur, in;
 343 #endif
 344 #ifdef INET6
 345         struct in6_addr last6, cur6, in6;
 346 #endif
 347
 348         CARP_LOCK_ASSERT(sc);
 349
 350         /* Compute ipad from key. */
 351         bzero(sc->sc_pad, sizeof(sc->sc_pad));
 352         bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
 353         for (i = 0; i < sizeof(sc->sc_pad); i++)
 354                 sc->sc_pad[i] ^= 0x36;
 355
 356         /* Precompute first part of inner hash. */
 357         SHA1Init(&sc->sc_sha1);
 358         SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
 359         SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
 360         SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
 361         SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
 362 #ifdef INET
 363         cur.s_addr = 0;
 364         do {
 365                 found = 0;
 366                 last = cur;
 367                 cur.s_addr = 0xffffffff;
 368                 CARP_FOREACH_IFA(sc, ifa) {
 369                         in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 370                         if (ifa->ifa_addr->sa_family == AF_INET &&
 371                             ntohl(in.s_addr) > ntohl(last.s_addr) &&
 372                             ntohl(in.s_addr) < ntohl(cur.s_addr)) {
 373                                 cur.s_addr = in.s_addr;
 374                                 found++;
 375                         }
 376                 }
 377                 if (found)
 378                         SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur));
 379         } while (found);
 380 #endif /* INET */
 381 #ifdef INET6
 382         memset(&cur6, 0, sizeof(cur6));
 383         do {
 384                 found = 0;
 385                 last6 = cur6;
 386                 memset(&cur6, 0xff, sizeof(cur6));
 387                 CARP_FOREACH_IFA(sc, ifa) {
 388                         in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 389                         if (IN6_IS_SCOPE_EMBED(&in6))
 390                                 in6.s6_addr16[1] = 0;
 391                         if (ifa->ifa_addr->sa_family == AF_INET6 &&
 392                             memcmp(&in6, &last6, sizeof(in6)) > 0 &&
 393                             memcmp(&in6, &cur6, sizeof(in6)) < 0) {
 394                                 cur6 = in6;
 395                                 found++;
 396                         }
 397                 }
 398                 if (found)
 399                         SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6));
 400         } while (found);
 401 #endif /* INET6 */
 402
 403         /* convert ipad to opad */
 404         for (i = 0; i < sizeof(sc->sc_pad); i++)
 405                 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
 406 }
 407
 408 static void
 409 carp_hmac_generate(struct carp_softc *sc, uint32_t counter[2],
 410     unsigned char md[20])
 411 {
 412         SHA1_CTX sha1ctx;
 413
 414         CARP_LOCK_ASSERT(sc);
 415
 416         /* fetch first half of inner hash */
 417         bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
 418
 419         SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
 420         SHA1Final(md, &sha1ctx);
 421
 422         /* outer hash */
 423         SHA1Init(&sha1ctx);
 424         SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
 425         SHA1Update(&sha1ctx, md, 20);
 426         SHA1Final(md, &sha1ctx);
 427 }
 428
 429 static int
 430 carp_hmac_verify(struct carp_softc *sc, uint32_t counter[2],
 431     unsigned char md[20])
 432 {
 433         unsigned char md2[20];
 434
 435         CARP_LOCK_ASSERT(sc);
 436
 437         carp_hmac_generate(sc, counter, md2);
 438
 439         return (bcmp(md, md2, sizeof(md2)));
 440 }
 441
 442 /*
 443  * process input packet.
 444  * we have rearranged checks order compared to the rfc,
 445  * but it seems more efficient this way or not possible otherwise.
 446  */
 447 #ifdef INET
 448 int
 449 carp_input(struct mbuf **mp, int *offp, int proto)
 450 {
 451         struct mbuf *m = *mp;
 452         struct ip *ip = mtod(m, struct ip *);
 453         struct carp_header *ch;
 454         int iplen, len;
 455
 456         iplen = *offp;
 457         *mp = NULL;
 458
 459         CARPSTATS_INC(carps_ipackets);
 460
 461         if (!V_carp_allow) {
 462                 m_freem(m);
 463                 return (IPPROTO_DONE);
 464         }
 465
 466         /* verify that the IP TTL is 255.  */
 467         if (ip->ip_ttl != CARP_DFLTTL) {
 468                 CARPSTATS_INC(carps_badttl);
 469                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 470                     ip->ip_ttl,
 471                     m->m_pkthdr.rcvif->if_xname);
 472                 m_freem(m);
 473                 return (IPPROTO_DONE);
 474         }
 475
 476         iplen = ip->ip_hl << 2;
 477
 478         if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
 479                 CARPSTATS_INC(carps_badlen);
 480                 CARP_DEBUG("%s: received len %zd < sizeof(struct carp_header) "
 481                     "on %s\n", __func__, m->m_len - sizeof(struct ip),
 482                     m->m_pkthdr.rcvif->if_xname);
 483                 m_freem(m);
 484                 return (IPPROTO_DONE);
 485         }
 486
 487         if (iplen + sizeof(*ch) < m->m_len) {
 488                 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
 489                         CARPSTATS_INC(carps_hdrops);
 490                         CARP_DEBUG("%s: pullup failed\n", __func__);
 491                         return (IPPROTO_DONE);
 492                 }
 493                 ip = mtod(m, struct ip *);
 494         }
 495         ch = (struct carp_header *)((char *)ip + iplen);
 496
 497         /*
 498          * verify that the received packet length is
 499          * equal to the CARP header
 500          */
 501         len = iplen + sizeof(*ch);
 502         if (len > m->m_pkthdr.len) {
 503                 CARPSTATS_INC(carps_badlen);
 504                 CARP_DEBUG("%s: packet too short %d on %s\n", __func__,
 505                     m->m_pkthdr.len,
 506                     m->m_pkthdr.rcvif->if_xname);
 507                 m_freem(m);
 508                 return (IPPROTO_DONE);
 509         }
 510
 511         if ((m = m_pullup(m, len)) == NULL) {
 512                 CARPSTATS_INC(carps_hdrops);
 513                 return (IPPROTO_DONE);
 514         }
 515         ip = mtod(m, struct ip *);
 516         ch = (struct carp_header *)((char *)ip + iplen);
 517
 518         /* verify the CARP checksum */
 519         m->m_data += iplen;
 520         if (in_cksum(m, len - iplen)) {
 521                 CARPSTATS_INC(carps_badsum);
 522                 CARP_DEBUG("%s: checksum failed on %s\n", __func__,
 523                     m->m_pkthdr.rcvif->if_xname);
 524                 m_freem(m);
 525                 return (IPPROTO_DONE);
 526         }
 527         m->m_data -= iplen;
 528
 529         carp_input_c(m, ch, AF_INET);
 530         return (IPPROTO_DONE);
 531 }
 532 #endif
 533
 534 #ifdef INET6
 535 int
 536 carp6_input(struct mbuf **mp, int *offp, int proto)
 537 {
 538         struct mbuf *m = *mp;
 539         struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
 540         struct carp_header *ch;
 541         u_int len;
 542
 543         CARPSTATS_INC(carps_ipackets6);
 544
 545         if (!V_carp_allow) {
 546                 m_freem(m);
 547                 return (IPPROTO_DONE);
 548         }
 549
 550         /* check if received on a valid carp interface */
 551         if (m->m_pkthdr.rcvif->if_carp == NULL) {
 552                 CARPSTATS_INC(carps_badif);
 553                 CARP_DEBUG("%s: packet received on non-carp interface: %s\n",
 554                     __func__, m->m_pkthdr.rcvif->if_xname);
 555                 m_freem(m);
 556                 return (IPPROTO_DONE);
 557         }
 558
 559         /* verify that the IP TTL is 255 */
 560         if (ip6->ip6_hlim != CARP_DFLTTL) {
 561                 CARPSTATS_INC(carps_badttl);
 562                 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
 563                     ip6->ip6_hlim, m->m_pkthdr.rcvif->if_xname);
 564                 m_freem(m);
 565                 return (IPPROTO_DONE);
 566         }
 567
 568         /* verify that we have a complete carp packet */
 569         if (m->m_len < *offp + sizeof(*ch)) {
 570                 len = m->m_len;
 571                 m = m_pullup(m, *offp + sizeof(*ch));
 572                 if (m == NULL) {
 573                         CARPSTATS_INC(carps_badlen);
 574                         CARP_DEBUG("%s: packet size %u too small\n", __func__, len);
 575                         return (IPPROTO_DONE);
 576                 }
 577         }
 578         ch = (struct carp_header *)(mtod(m, caddr_t) + *offp);
 579
 580
 581         /* verify the CARP checksum */
 582         m->m_data += *offp;
 583         if (in_cksum(m, sizeof(*ch))) {
 584                 CARPSTATS_INC(carps_badsum);
 585                 CARP_DEBUG("%s: checksum failed, on %s\n", __func__,
 586                     m->m_pkthdr.rcvif->if_xname);
 587                 m_freem(m);
 588                 return (IPPROTO_DONE);
 589         }
 590         m->m_data -= *offp;
 591
 592         carp_input_c(m, ch, AF_INET6);
 593         return (IPPROTO_DONE);
 594 }
 595 #endif /* INET6 */
 596
 597 /*
 598  * This routine should not be necessary at all, but some switches
 599  * (VMWare ESX vswitches) can echo our own packets back at us,
 600  * and we must ignore them or they will cause us to drop out of
 601  * MASTER mode.
 602  *
 603  * We cannot catch all cases of network loops.  Instead, what we
 604  * do here is catch any packet that arrives with a carp header
 605  * with a VHID of 0, that comes from an address that is our own.
 606  * These packets are by definition "from us" (even if they are from
 607  * a misconfigured host that is pretending to be us).
 608  *
 609  * The VHID test is outside this mini-function.
 610  */
 611 static int
 612 carp_source_is_self(struct mbuf *m, struct ifaddr *ifa, sa_family_t af)
 613 {
 614 #ifdef INET
 615         struct ip *ip4;
 616         struct in_addr in4;
 617 #endif
 618 #ifdef INET6
 619         struct ip6_hdr *ip6;
 620         struct in6_addr in6;
 621 #endif
 622
 623         switch (af) {
 624 #ifdef INET
 625         case AF_INET:
 626                 ip4 = mtod(m, struct ip *);
 627                 in4 = ifatoia(ifa)->ia_addr.sin_addr;
 628                 return (in4.s_addr == ip4->ip_src.s_addr);
 629 #endif
 630 #ifdef INET6
 631         case AF_INET6:
 632                 ip6 = mtod(m, struct ip6_hdr *);
 633                 in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
 634                 return (memcmp(&in6, &ip6->ip6_src, sizeof(in6)) == 0);
 635 #endif
 636         default:
 637                 break;
 638         }
 639         return (0);
 640 }
 641
 642 static void
 643 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
 644 {
 645         struct ifnet *ifp = m->m_pkthdr.rcvif;
 646         struct ifaddr *ifa, *match;
 647         struct carp_softc *sc;
 648         uint64_t tmp_counter;
 649         struct timeval sc_tv, ch_tv;
 650         int error;
 651
 652         /*
 653          * Verify that the VHID is valid on the receiving interface.
 654          *
 655          * There should be just one match.  If there are none
 656          * the VHID is not valid and we drop the packet.  If
 657          * there are multiple VHID matches, take just the first
 658          * one, for compatibility with previous code.  While we're
 659          * scanning, check for obvious loops in the network topology
 660          * (these should never happen, and as noted above, we may
 661          * miss real loops; this is just a double-check).
 662          */
 663         IF_ADDR_RLOCK(ifp);
 664         error = 0;
 665         match = NULL;
 666         IFNET_FOREACH_IFA(ifp, ifa) {
 667                 if (match == NULL && ifa->ifa_carp != NULL &&
 668                     ifa->ifa_addr->sa_family == af &&
 669                     ifa->ifa_carp->sc_vhid == ch->carp_vhid)
 670                         match = ifa;
 671                 if (ch->carp_vhid == 0 && carp_source_is_self(m, ifa, af))
 672                         error = ELOOP;
 673         }
 674         ifa = error ? NULL : match;
 675         if (ifa != NULL)
 676                 ifa_ref(ifa);
 677         IF_ADDR_RUNLOCK(ifp);
 678
 679         if (ifa == NULL) {
 680                 if (error == ELOOP) {
 681                         CARP_DEBUG("dropping looped packet on interface %s\n",
 682                             ifp->if_xname);
 683                         CARPSTATS_INC(carps_badif);     /* ??? */
 684                 } else {
 685                         CARPSTATS_INC(carps_badvhid);
 686                 }
 687                 m_freem(m);
 688                 return;
 689         }
 690
 691         /* verify the CARP version. */
 692         if (ch->carp_version != CARP_VERSION) {
 693                 CARPSTATS_INC(carps_badver);
 694                 CARP_DEBUG("%s: invalid version %d\n", ifp->if_xname,
 695                     ch->carp_version);
 696                 ifa_free(ifa);
 697                 m_freem(m);
 698                 return;
 699         }
 700
 701         sc = ifa->ifa_carp;
 702         CARP_LOCK(sc);
 703         ifa_free(ifa);
 704
 705         if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
 706                 CARPSTATS_INC(carps_badauth);
 707                 CARP_DEBUG("%s: incorrect hash for VHID %u@%s\n", __func__,
 708                     sc->sc_vhid, ifp->if_xname);
 709                 goto out;
 710         }
 711
 712         tmp_counter = ntohl(ch->carp_counter[0]);
 713         tmp_counter = tmp_counter<<32;
 714         tmp_counter += ntohl(ch->carp_counter[1]);
 715
 716         /* XXX Replay protection goes here */
 717
 718         sc->sc_init_counter = 0;
 719         sc->sc_counter = tmp_counter;
 720
 721         sc_tv.tv_sec = sc->sc_advbase;
 722         sc_tv.tv_usec = DEMOTE_ADVSKEW(sc) * 1000000 / 256;
 723         ch_tv.tv_sec = ch->carp_advbase;
 724         ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
 725
 726         switch (sc->sc_state) {
 727         case INIT:
 728                 break;
 729         case MASTER:
 730                 /*
 731                  * If we receive an advertisement from a master who's going to
 732                  * be more frequent than us, go into BACKUP state.
 733                  */
 734                 if (timevalcmp(&sc_tv, &ch_tv, >) ||
 735                     timevalcmp(&sc_tv, &ch_tv, ==)) {
 736                         callout_stop(&sc->sc_ad_tmo);
 737                         carp_set_state(sc, BACKUP,
 738                             "more frequent advertisement received");
 739                         carp_setrun(sc, 0);
 740                         carp_delroute(sc);
 741                 }
 742                 break;
 743         case BACKUP:
 744                 /*
 745                  * If we're pre-empting masters who advertise slower than us,
 746                  * and this one claims to be slower, treat him as down.
 747                  */
 748                 if (V_carp_preempt && timevalcmp(&sc_tv, &ch_tv, <)) {
 749                         carp_master_down_locked(sc,
 750                             "preempting a slower master");
 751                         break;
 752                 }
 753
 754                 /*
 755                  *  If the master is going to advertise at such a low frequency
 756                  *  that he's guaranteed to time out, we'd might as well just
 757                  *  treat him as timed out now.
 758                  */
 759                 sc_tv.tv_sec = sc->sc_advbase * 3;
 760                 if (timevalcmp(&sc_tv, &ch_tv, <)) {
 761                         carp_master_down_locked(sc, "master will time out");
 762                         break;
 763                 }
 764
 765                 /*
 766                  * Otherwise, we reset the counter and wait for the next
 767                  * advertisement.
 768                  */
 769                 carp_setrun(sc, af);
 770                 break;
 771         }
 772
 773 out:
 774         CARP_UNLOCK(sc);
 775         m_freem(m);
 776 }
 777
 778 static int
 779 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
 780 {
 781         struct m_tag *mtag;
 782
 783         if (sc->sc_init_counter) {
 784                 /* this could also be seconds since unix epoch */
 785                 sc->sc_counter = arc4random();
 786                 sc->sc_counter = sc->sc_counter << 32;
 787                 sc->sc_counter += arc4random();
 788         } else
 789                 sc->sc_counter++;
 790
 791         ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
 792         ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
 793
 794         carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
 795
 796         /* Tag packet for carp_output */
 797         if ((mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct carp_softc *),
 798             M_NOWAIT)) == NULL) {
 799                 m_freem(m);
 800                 CARPSTATS_INC(carps_onomem);
 801                 return (ENOMEM);
 802         }
 803         bcopy(&sc, mtag + 1, sizeof(sc));
 804         m_tag_prepend(m, mtag);
 805
 806         return (0);
 807 }
 808
 809 /*
 810  * To avoid LORs and possible recursions this function shouldn't
 811  * be called directly, but scheduled via taskqueue.
 812  */
 813 static void
 814 carp_send_ad_all(void *ctx __unused, int pending __unused)
 815 {
 816         struct carp_softc *sc;
 817
 818         mtx_lock(&carp_mtx);
 819         LIST_FOREACH(sc, &carp_list, sc_next)
 820                 if (sc->sc_state == MASTER) {
 821                         CARP_LOCK(sc);
 822                         CURVNET_SET(sc->sc_carpdev->if_vnet);
 823                         carp_send_ad_locked(sc);
 824                         CURVNET_RESTORE();
 825                         CARP_UNLOCK(sc);
 826                 }
 827         mtx_unlock(&carp_mtx);
 828 }
 829
 830 /* Send a periodic advertisement, executed in callout context. */
 831 static void
 832 carp_send_ad(void *v)
 833 {
 834         struct carp_softc *sc = v;
 835
 836         CARP_LOCK_ASSERT(sc);
 837         CURVNET_SET(sc->sc_carpdev->if_vnet);
 838         carp_send_ad_locked(sc);
 839         CURVNET_RESTORE();
 840         CARP_UNLOCK(sc);
 841 }
 842
 843 static void
 844 carp_send_ad_error(struct carp_softc *sc, int error)
 845 {
 846
 847         if (error) {
 848                 if (sc->sc_sendad_errors < INT_MAX)
 849                         sc->sc_sendad_errors++;
 850                 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
 851                         static const char fmt[] = "send error %d on %s";
 852                         char msg[sizeof(fmt) + IFNAMSIZ];
 853
 854                         sprintf(msg, fmt, error, sc->sc_carpdev->if_xname);
 855                         carp_demote_adj(V_carp_senderr_adj, msg);
 856                 }
 857                 sc->sc_sendad_success = 0;
 858         } else {
 859                 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS &&
 860                     ++sc->sc_sendad_success >= CARP_SENDAD_MIN_SUCCESS) {
 861                         static const char fmt[] = "send ok on %s";
 862                         char msg[sizeof(fmt) + IFNAMSIZ];
 863
 864                         sprintf(msg, fmt, sc->sc_carpdev->if_xname);
 865                         carp_demote_adj(-V_carp_senderr_adj, msg);
 866                         sc->sc_sendad_errors = 0;
 867                 } else
 868                         sc->sc_sendad_errors = 0;
 869         }
 870 }
 871
 872 /*
 873  * Pick the best ifaddr on the given ifp for sending CARP
 874  * advertisements.
 875  *
 876  * "Best" here is defined by ifa_preferred().  This function is much
 877  * much like ifaof_ifpforaddr() except that we just use ifa_preferred().
 878  *
 879  * (This could be simplified to return the actual address, except that
 880  * it has a different format in AF_INET and AF_INET6.)
 881  */
 882 static struct ifaddr *
 883 carp_best_ifa(int af, struct ifnet *ifp)
 884 {
 885         struct ifaddr *ifa, *best;
 886
 887         if (af >= AF_MAX)
 888                 return (NULL);
 889         best = NULL;
 890         IF_ADDR_RLOCK(ifp);
 891         CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
 892                 if (ifa->ifa_addr->sa_family == af &&
 893                     (best == NULL || ifa_preferred(best, ifa)))
 894                         best = ifa;
 895         }
 896         IF_ADDR_RUNLOCK(ifp);
 897         if (best != NULL)
 898                 ifa_ref(best);
 899         return (best);
 900 }
 901
 902 static void
 903 carp_send_ad_locked(struct carp_softc *sc)
 904 {
 905         struct carp_header ch;
 906         struct timeval tv;
 907         struct ifaddr *ifa;
 908         struct carp_header *ch_ptr;
 909         struct mbuf *m;
 910         int len, advskew;
 911
 912         CARP_LOCK_ASSERT(sc);
 913
 914         advskew = DEMOTE_ADVSKEW(sc);
 915         tv.tv_sec = sc->sc_advbase;
 916         tv.tv_usec = advskew * 1000000 / 256;
 917
 918         ch.carp_version = CARP_VERSION;
 919         ch.carp_type = CARP_ADVERTISEMENT;
 920         ch.carp_vhid = sc->sc_vhid;
 921         ch.carp_advbase = sc->sc_advbase;
 922         ch.carp_advskew = advskew;
 923         ch.carp_authlen = 7;    /* XXX DEFINE */
 924         ch.carp_pad1 = 0;       /* must be zero */
 925         ch.carp_cksum = 0;
 926
 927         /* XXXGL: OpenBSD picks first ifaddr with needed family. */
 928
 929 #ifdef INET
 930         if (sc->sc_naddrs) {
 931                 struct ip *ip;
 932
 933                 m = m_gethdr(M_NOWAIT, MT_DATA);
 934                 if (m == NULL) {
 935                         CARPSTATS_INC(carps_onomem);
 936                         goto resched;
 937                 }
 938                 len = sizeof(*ip) + sizeof(ch);
 939                 m->m_pkthdr.len = len;
 940                 m->m_pkthdr.rcvif = NULL;
 941                 m->m_len = len;
 942                 M_ALIGN(m, m->m_len);
 943                 m->m_flags |= M_MCAST;
 944                 ip = mtod(m, struct ip *);
 945                 ip->ip_v = IPVERSION;
 946                 ip->ip_hl = sizeof(*ip) >> 2;
 947                 ip->ip_tos = V_carp_dscp << IPTOS_DSCP_OFFSET;
 948                 ip->ip_len = htons(len);
 949                 ip->ip_off = htons(IP_DF);
 950                 ip->ip_ttl = CARP_DFLTTL;
 951                 ip->ip_p = IPPROTO_CARP;
 952                 ip->ip_sum = 0;
 953                 ip_fillid(ip);
 954
 955                 ifa = carp_best_ifa(AF_INET, sc->sc_carpdev);
 956                 if (ifa != NULL) {
 957                         ip->ip_src.s_addr =
 958                             ifatoia(ifa)->ia_addr.sin_addr.s_addr;
 959                         ifa_free(ifa);
 960                 } else
 961                         ip->ip_src.s_addr = 0;
 962                 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
 963
 964                 ch_ptr = (struct carp_header *)(&ip[1]);
 965                 bcopy(&ch, ch_ptr, sizeof(ch));
 966                 if (carp_prepare_ad(m, sc, ch_ptr))
 967                         goto resched;
 968
 969                 m->m_data += sizeof(*ip);
 970                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip));
 971                 m->m_data -= sizeof(*ip);
 972
 973                 CARPSTATS_INC(carps_opackets);
 974
 975                 carp_send_ad_error(sc, ip_output(m, NULL, NULL, IP_RAWOUTPUT,
 976                     &sc->sc_carpdev->if_carp->cif_imo, NULL));
 977         }
 978 #endif /* INET */
 979 #ifdef INET6
 980         if (sc->sc_naddrs6) {
 981                 struct ip6_hdr *ip6;
 982
 983                 m = m_gethdr(M_NOWAIT, MT_DATA);
 984                 if (m == NULL) {
 985                         CARPSTATS_INC(carps_onomem);
 986                         goto resched;
 987                 }
 988                 len = sizeof(*ip6) + sizeof(ch);
 989                 m->m_pkthdr.len = len;
 990                 m->m_pkthdr.rcvif = NULL;
 991                 m->m_len = len;
 992                 M_ALIGN(m, m->m_len);
 993                 m->m_flags |= M_MCAST;
 994                 ip6 = mtod(m, struct ip6_hdr *);
 995                 bzero(ip6, sizeof(*ip6));
 996                 ip6->ip6_vfc |= IPV6_VERSION;
 997                 /* Traffic class isn't defined in ip6 struct instead
 998                  * it gets offset into flowid field */
 999                 ip6->ip6_flow |= htonl(V_carp_dscp << (IPV6_FLOWLABEL_LEN +
1000                     IPTOS_DSCP_OFFSET));
1001                 ip6->ip6_hlim = CARP_DFLTTL;
1002                 ip6->ip6_nxt = IPPROTO_CARP;
1003
1004                 /* set the source address */
1005                 ifa = carp_best_ifa(AF_INET6, sc->sc_carpdev);
1006                 if (ifa != NULL) {
1007                         bcopy(IFA_IN6(ifa), &ip6->ip6_src,
1008                             sizeof(struct in6_addr));
1009                         ifa_free(ifa);
1010                 } else
1011                         /* This should never happen with IPv6. */
1012                         bzero(&ip6->ip6_src, sizeof(struct in6_addr));
1013
1014                 /* Set the multicast destination. */
1015                 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
1016                 ip6->ip6_dst.s6_addr8[15] = 0x12;
1017                 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
1018                         m_freem(m);
1019                         CARP_DEBUG("%s: in6_setscope failed\n", __func__);
1020                         goto resched;
1021                 }
1022
1023                 ch_ptr = (struct carp_header *)(&ip6[1]);
1024                 bcopy(&ch, ch_ptr, sizeof(ch));
1025                 if (carp_prepare_ad(m, sc, ch_ptr))
1026                         goto resched;
1027
1028                 m->m_data += sizeof(*ip6);
1029                 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip6));
1030                 m->m_data -= sizeof(*ip6);
1031
1032                 CARPSTATS_INC(carps_opackets6);
1033
1034                 carp_send_ad_error(sc, ip6_output(m, NULL, NULL, 0,
1035                     &sc->sc_carpdev->if_carp->cif_im6o, NULL, NULL));
1036         }
1037 #endif /* INET6 */
1038
1039 resched:
1040         callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), carp_send_ad, sc);
1041 }
1042
1043 static void
1044 carp_addroute(struct carp_softc *sc)
1045 {
1046         struct ifaddr *ifa;
1047
1048         CARP_FOREACH_IFA(sc, ifa)
1049                 carp_ifa_addroute(ifa);
1050 }
1051
1052 static void
1053 carp_ifa_addroute(struct ifaddr *ifa)
1054 {
1055
1056         switch (ifa->ifa_addr->sa_family) {
1057 #ifdef INET
1058         case AF_INET:
1059                 in_addprefix(ifatoia(ifa), RTF_UP);
1060                 ifa_add_loopback_route(ifa,
1061                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
1062                 break;
1063 #endif
1064 #ifdef INET6
1065         case AF_INET6:
1066                 ifa_add_loopback_route(ifa,
1067                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
1068                 nd6_add_ifa_lle(ifatoia6(ifa));
1069                 break;
1070 #endif
1071         }
1072 }
1073
1074 static void
1075 carp_delroute(struct carp_softc *sc)
1076 {
1077         struct ifaddr *ifa;
1078
1079         CARP_FOREACH_IFA(sc, ifa)
1080                 carp_ifa_delroute(ifa);
1081 }
1082
1083 static void
1084 carp_ifa_delroute(struct ifaddr *ifa)
1085 {
1086
1087         switch (ifa->ifa_addr->sa_family) {
1088 #ifdef INET
1089         case AF_INET:
1090                 ifa_del_loopback_route(ifa,
1091                     (struct sockaddr *)&ifatoia(ifa)->ia_addr);
1092                 in_scrubprefix(ifatoia(ifa), LLE_STATIC);
1093                 break;
1094 #endif
1095 #ifdef INET6
1096         case AF_INET6:
1097                 ifa_del_loopback_route(ifa,
1098                     (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
1099                 nd6_rem_ifa_lle(ifatoia6(ifa), 1);
1100                 break;
1101 #endif
1102         }
1103 }
1104
1105 int
1106 carp_master(struct ifaddr *ifa)
1107 {
1108         struct carp_softc *sc = ifa->ifa_carp;
1109
1110         return (sc->sc_state == MASTER);
1111 }
1112
1113 #ifdef INET
1114 /*
1115  * Broadcast a gratuitous ARP request containing
1116  * the virtual router MAC address for each IP address
1117  * associated with the virtual router.
1118  */
1119 static void
1120 carp_send_arp(struct carp_softc *sc)
1121 {
1122         struct ifaddr *ifa;
1123         struct in_addr addr;
1124
1125         CARP_FOREACH_IFA(sc, ifa) {
1126                 if (ifa->ifa_addr->sa_family != AF_INET)
1127                         continue;
1128                 addr = ((struct sockaddr_in *)ifa->ifa_addr)->sin_addr;
1129                 arp_announce_ifaddr(sc->sc_carpdev, addr, LLADDR(&sc->sc_addr));
1130         }
1131 }
1132
1133 int
1134 carp_iamatch(struct ifaddr *ifa, uint8_t **enaddr)
1135 {
1136         struct carp_softc *sc = ifa->ifa_carp;
1137
1138         if (sc->sc_state == MASTER) {
1139                 *enaddr = LLADDR(&sc->sc_addr);
1140                 return (1);
1141         }
1142
1143         return (0);
1144 }
1145 #endif
1146
1147 #ifdef INET6
1148 static void
1149 carp_send_na(struct carp_softc *sc)
1150 {
1151         static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1152         struct ifaddr *ifa;
1153         struct in6_addr *in6;
1154
1155         CARP_FOREACH_IFA(sc, ifa) {
1156                 if (ifa->ifa_addr->sa_family != AF_INET6)
1157                         continue;
1158
1159                 in6 = IFA_IN6(ifa);
1160                 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1161                     ND_NA_FLAG_OVERRIDE, 1, NULL);
1162                 DELAY(1000);    /* XXX */
1163         }
1164 }
1165
1166 /*
1167  * Returns ifa in case it's a carp address and it is MASTER, or if the address
1168  * matches and is not a carp address.  Returns NULL otherwise.
1169  */
1170 struct ifaddr *
1171 carp_iamatch6(struct ifnet *ifp, struct in6_addr *taddr)
1172 {
1173         struct ifaddr *ifa;
1174
1175         ifa = NULL;
1176         IF_ADDR_RLOCK(ifp);
1177         CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
1178                 if (ifa->ifa_addr->sa_family != AF_INET6)
1179                         continue;
1180                 if (!IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa)))
1181                         continue;
1182                 if (ifa->ifa_carp && ifa->ifa_carp->sc_state != MASTER)
1183                         ifa = NULL;
1184                 else
1185                         ifa_ref(ifa);
1186                 break;
1187         }
1188         IF_ADDR_RUNLOCK(ifp);
1189
1190         return (ifa);
1191 }
1192
1193 caddr_t
1194 carp_macmatch6(struct ifnet *ifp, struct mbuf *m, const struct in6_addr *taddr)
1195 {
1196         struct ifaddr *ifa;
1197
1198         IF_ADDR_RLOCK(ifp);
1199         IFNET_FOREACH_IFA(ifp, ifa)
1200                 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1201                     IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1202                         struct carp_softc *sc = ifa->ifa_carp;
1203                         struct m_tag *mtag;
1204
1205                         IF_ADDR_RUNLOCK(ifp);
1206
1207                         mtag = m_tag_get(PACKET_TAG_CARP,
1208                             sizeof(struct carp_softc *), M_NOWAIT);
1209                         if (mtag == NULL)
1210                                 /* Better a bit than nothing. */
1211                                 return (LLADDR(&sc->sc_addr));
1212
1213                         bcopy(&sc, mtag + 1, sizeof(sc));
1214                         m_tag_prepend(m, mtag);
1215
1216                         return (LLADDR(&sc->sc_addr));
1217                 }
1218         IF_ADDR_RUNLOCK(ifp);
1219
1220         return (NULL);
1221 }
1222 #endif /* INET6 */
1223
1224 int
1225 carp_forus(struct ifnet *ifp, u_char *dhost)
1226 {
1227         struct carp_softc *sc;
1228         uint8_t *ena = dhost;
1229
1230         if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1231                 return (0);
1232
1233         CIF_LOCK(ifp->if_carp);
1234         IFNET_FOREACH_CARP(ifp, sc) {
1235                 CARP_LOCK(sc);
1236                 if (sc->sc_state == MASTER && !bcmp(dhost, LLADDR(&sc->sc_addr),
1237                     ETHER_ADDR_LEN)) {
1238                         CARP_UNLOCK(sc);
1239                         CIF_UNLOCK(ifp->if_carp);
1240                         return (1);
1241                 }
1242                 CARP_UNLOCK(sc);
1243         }
1244         CIF_UNLOCK(ifp->if_carp);
1245
1246         return (0);
1247 }
1248
1249 /* Master down timeout event, executed in callout context. */
1250 static void
1251 carp_master_down(void *v)
1252 {
1253         struct carp_softc *sc = v;
1254
1255         CARP_LOCK_ASSERT(sc);
1256
1257         CURVNET_SET(sc->sc_carpdev->if_vnet);
1258         if (sc->sc_state == BACKUP) {
1259                 carp_master_down_locked(sc, "master timed out");
1260         }
1261         CURVNET_RESTORE();
1262
1263         CARP_UNLOCK(sc);
1264 }
1265
1266 static void
1267 carp_master_down_locked(struct carp_softc *sc, const char *reason)
1268 {
1269
1270         CARP_LOCK_ASSERT(sc);
1271
1272         switch (sc->sc_state) {
1273         case BACKUP:
1274                 carp_set_state(sc, MASTER, reason);
1275                 carp_send_ad_locked(sc);
1276 #ifdef INET
1277                 carp_send_arp(sc);
1278 #endif
1279 #ifdef INET6
1280                 carp_send_na(sc);
1281 #endif
1282                 carp_setrun(sc, 0);
1283                 carp_addroute(sc);
1284                 break;
1285         case INIT:
1286         case MASTER:
1287 #ifdef INVARIANTS
1288                 panic("carp: VHID %u@%s: master_down event in %s state\n",
1289                     sc->sc_vhid,
1290                     sc->sc_carpdev->if_xname,
1291                     sc->sc_state ? "MASTER" : "INIT");
1292 #endif
1293                 break;
1294         }
1295 }
1296
1297 /*
1298  * When in backup state, af indicates whether to reset the master down timer
1299  * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1300  */
1301 static void
1302 carp_setrun(struct carp_softc *sc, sa_family_t af)
1303 {
1304         struct timeval tv;
1305
1306         CARP_LOCK_ASSERT(sc);
1307
1308         if ((sc->sc_carpdev->if_flags & IFF_UP) == 0 ||
1309             sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1310             (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) ||
1311             !V_carp_allow)
1312                 return;
1313
1314         switch (sc->sc_state) {
1315         case INIT:
1316                 carp_set_state(sc, BACKUP, "initialization complete");
1317                 carp_setrun(sc, 0);
1318                 break;
1319         case BACKUP:
1320                 callout_stop(&sc->sc_ad_tmo);
1321                 tv.tv_sec = 3 * sc->sc_advbase;
1322                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1323                 switch (af) {
1324 #ifdef INET
1325                 case AF_INET:
1326                         callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1327                             carp_master_down, sc);
1328                         break;
1329 #endif
1330 #ifdef INET6
1331                 case AF_INET6:
1332                         callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1333                             carp_master_down, sc);
1334                         break;
1335 #endif
1336                 default:
1337 #ifdef INET
1338                         if (sc->sc_naddrs)
1339                                 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1340                                     carp_master_down, sc);
1341 #endif
1342 #ifdef INET6
1343                         if (sc->sc_naddrs6)
1344                                 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1345                                     carp_master_down, sc);
1346 #endif
1347                         break;
1348                 }
1349                 break;
1350         case MASTER:
1351                 tv.tv_sec = sc->sc_advbase;
1352                 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1353                 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1354                     carp_send_ad, sc);
1355                 break;
1356         }
1357 }
1358
1359 /*
1360  * Setup multicast structures.
1361  */
1362 static int
1363 carp_multicast_setup(struct carp_if *cif, sa_family_t sa)
1364 {
1365         struct ifnet *ifp = cif->cif_ifp;
1366         int error = 0;
1367
1368         switch (sa) {
1369 #ifdef INET
1370         case AF_INET:
1371             {
1372                 struct ip_moptions *imo = &cif->cif_imo;
1373                 struct in_mfilter *imf;
1374                 struct in_addr addr;
1375
1376                 if (ip_mfilter_first(&imo->imo_head) != NULL)
1377                         return (0);
1378
1379                 imf = ip_mfilter_alloc(M_WAITOK, 0, 0);
1380                 ip_mfilter_init(&imo->imo_head);
1381                 imo->imo_multicast_vif = -1;
1382
1383                 addr.s_addr = htonl(INADDR_CARP_GROUP);
1384                 if ((error = in_joingroup(ifp, &addr, NULL,
1385                     &imf->imf_inm)) != 0) {
1386                         ip_mfilter_free(imf);
1387                         break;
1388                 }
1389
1390                 ip_mfilter_insert(&imo->imo_head, imf);
1391                 imo->imo_multicast_ifp = ifp;
1392                 imo->imo_multicast_ttl = CARP_DFLTTL;
1393                 imo->imo_multicast_loop = 0;
1394                 break;
1395            }
1396 #endif
1397 #ifdef INET6
1398         case AF_INET6:
1399             {
1400                 struct ip6_moptions *im6o = &cif->cif_im6o;
1401                 struct in6_mfilter *im6f[2];
1402                 struct in6_addr in6;
1403
1404                 if (ip6_mfilter_first(&im6o->im6o_head))
1405                         return (0);
1406
1407                 im6f[0] = ip6_mfilter_alloc(M_WAITOK, 0, 0);
1408                 im6f[1] = ip6_mfilter_alloc(M_WAITOK, 0, 0);
1409
1410                 ip6_mfilter_init(&im6o->im6o_head);
1411                 im6o->im6o_multicast_hlim = CARP_DFLTTL;
1412                 im6o->im6o_multicast_ifp = ifp;
1413
1414                 /* Join IPv6 CARP multicast group. */
1415                 bzero(&in6, sizeof(in6));
1416                 in6.s6_addr16[0] = htons(0xff02);
1417                 in6.s6_addr8[15] = 0x12;
1418                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1419                         ip6_mfilter_free(im6f[0]);
1420                         ip6_mfilter_free(im6f[1]);
1421                         break;
1422                 }
1423                 if ((error = in6_joingroup(ifp, &in6, NULL, &im6f[0]->im6f_in6m, 0)) != 0) {
1424                         ip6_mfilter_free(im6f[0]);
1425                         ip6_mfilter_free(im6f[1]);
1426                         break;
1427                 }
1428
1429                 /* Join solicited multicast address. */
1430                 bzero(&in6, sizeof(in6));
1431                 in6.s6_addr16[0] = htons(0xff02);
1432                 in6.s6_addr32[1] = 0;
1433                 in6.s6_addr32[2] = htonl(1);
1434                 in6.s6_addr32[3] = 0;
1435                 in6.s6_addr8[12] = 0xff;
1436
1437                 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1438                         ip6_mfilter_free(im6f[0]);
1439                         ip6_mfilter_free(im6f[1]);
1440                         break;
1441                 }
1442
1443                 if ((error = in6_joingroup(ifp, &in6, NULL, &im6f[1]->im6f_in6m, 0)) != 0) {
1444                         in6_leavegroup(im6f[0]->im6f_in6m, NULL);
1445                         ip6_mfilter_free(im6f[0]);
1446                         ip6_mfilter_free(im6f[1]);
1447                         break;
1448                 }
1449                 ip6_mfilter_insert(&im6o->im6o_head, im6f[0]);
1450                 ip6_mfilter_insert(&im6o->im6o_head, im6f[1]);
1451                 break;
1452             }
1453 #endif
1454         }
1455
1456         return (error);
1457 }
1458
1459 /*
1460  * Free multicast structures.
1461  */
1462 static void
1463 carp_multicast_cleanup(struct carp_if *cif, sa_family_t sa)
1464 {
1465 #ifdef INET
1466         struct ip_moptions *imo = &cif->cif_imo;
1467         struct in_mfilter *imf;
1468 #endif
1469 #ifdef INET6
1470         struct ip6_moptions *im6o = &cif->cif_im6o;
1471         struct in6_mfilter *im6f;
1472 #endif
1473         sx_assert(&carp_sx, SA_XLOCKED);
1474
1475         switch (sa) {
1476 #ifdef INET
1477         case AF_INET:
1478                 if (cif->cif_naddrs != 0)
1479                         break;
1480
1481                 while ((imf = ip_mfilter_first(&imo->imo_head)) != NULL) {
1482                         ip_mfilter_remove(&imo->imo_head, imf);
1483                         in_leavegroup(imf->imf_inm, NULL);
1484                         ip_mfilter_free(imf);
1485                 }
1486                 break;
1487 #endif
1488 #ifdef INET6
1489         case AF_INET6:
1490                 if (cif->cif_naddrs6 != 0)
1491                         break;
1492
1493                 while ((im6f = ip6_mfilter_first(&im6o->im6o_head)) != NULL) {
1494                         ip6_mfilter_remove(&im6o->im6o_head, im6f);
1495                         in6_leavegroup(im6f->im6f_in6m, NULL);
1496                         ip6_mfilter_free(im6f);
1497                 }
1498                 break;
1499 #endif
1500         }
1501 }
1502
1503 int
1504 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa)
1505 {
1506         struct m_tag *mtag;
1507         struct carp_softc *sc;
1508
1509         if (!sa)
1510                 return (0);
1511
1512         switch (sa->sa_family) {
1513 #ifdef INET
1514         case AF_INET:
1515                 break;
1516 #endif
1517 #ifdef INET6
1518         case AF_INET6:
1519                 break;
1520 #endif
1521         default:
1522                 return (0);
1523         }
1524
1525         mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
1526         if (mtag == NULL)
1527                 return (0);
1528
1529         bcopy(mtag + 1, &sc, sizeof(sc));
1530
1531         /* Set the source MAC address to the Virtual Router MAC Address. */
1532         switch (ifp->if_type) {
1533         case IFT_ETHER:
1534         case IFT_BRIDGE:
1535         case IFT_L2VLAN: {
1536                         struct ether_header *eh;
1537
1538                         eh = mtod(m, struct ether_header *);
1539                         eh->ether_shost[0] = 0;
1540                         eh->ether_shost[1] = 0;
1541                         eh->ether_shost[2] = 0x5e;
1542                         eh->ether_shost[3] = 0;
1543                         eh->ether_shost[4] = 1;
1544                         eh->ether_shost[5] = sc->sc_vhid;
1545                 }
1546                 break;
1547         default:
1548                 printf("%s: carp is not supported for the %d interface type\n",
1549                     ifp->if_xname, ifp->if_type);
1550                 return (EOPNOTSUPP);
1551         }
1552
1553         return (0);
1554 }
1555
1556 static struct carp_softc*
1557 carp_alloc(struct ifnet *ifp)
1558 {
1559         struct carp_softc *sc;
1560         struct carp_if *cif;
1561
1562         sx_assert(&carp_sx, SA_XLOCKED);
1563
1564         if ((cif = ifp->if_carp) == NULL)
1565                 cif = carp_alloc_if(ifp);
1566
1567         sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
1568
1569         sc->sc_advbase = CARP_DFLTINTV;
1570         sc->sc_vhid = -1;       /* required setting */
1571         sc->sc_init_counter = 1;
1572         sc->sc_state = INIT;
1573
1574         sc->sc_ifasiz = sizeof(struct ifaddr *);
1575         sc->sc_ifas = malloc(sc->sc_ifasiz, M_CARP, M_WAITOK|M_ZERO);
1576         sc->sc_carpdev = ifp;
1577
1578         CARP_LOCK_INIT(sc);
1579 #ifdef INET
1580         callout_init_mtx(&sc->sc_md_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1581 #endif
1582 #ifdef INET6
1583         callout_init_mtx(&sc->sc_md6_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1584 #endif
1585         callout_init_mtx(&sc->sc_ad_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1586
1587         CIF_LOCK(cif);
1588         TAILQ_INSERT_TAIL(&cif->cif_vrs, sc, sc_list);
1589         CIF_UNLOCK(cif);
1590
1591         mtx_lock(&carp_mtx);
1592         LIST_INSERT_HEAD(&carp_list, sc, sc_next);
1593         mtx_unlock(&carp_mtx);
1594
1595         return (sc);
1596 }
1597
1598 static void
1599 carp_grow_ifas(struct carp_softc *sc)
1600 {
1601         struct ifaddr **new;
1602
1603         new = malloc(sc->sc_ifasiz * 2, M_CARP, M_WAITOK | M_ZERO);
1604         CARP_LOCK(sc);
1605         bcopy(sc->sc_ifas, new, sc->sc_ifasiz);
1606         free(sc->sc_ifas, M_CARP);
1607         sc->sc_ifas = new;
1608         sc->sc_ifasiz *= 2;
1609         CARP_UNLOCK(sc);
1610 }
1611
1612 static void
1613 carp_destroy(struct carp_softc *sc)
1614 {
1615         struct ifnet *ifp = sc->sc_carpdev;
1616         struct carp_if *cif = ifp->if_carp;
1617
1618         sx_assert(&carp_sx, SA_XLOCKED);
1619
1620         if (sc->sc_suppress)
1621                 carp_demote_adj(-V_carp_ifdown_adj, "vhid removed");
1622         CARP_UNLOCK(sc);
1623
1624         CIF_LOCK(cif);
1625         TAILQ_REMOVE(&cif->cif_vrs, sc, sc_list);
1626         CIF_UNLOCK(cif);
1627
1628         mtx_lock(&carp_mtx);
1629         LIST_REMOVE(sc, sc_next);
1630         mtx_unlock(&carp_mtx);
1631
1632         callout_drain(&sc->sc_ad_tmo);
1633 #ifdef INET
1634         callout_drain(&sc->sc_md_tmo);
1635 #endif
1636 #ifdef INET6
1637         callout_drain(&sc->sc_md6_tmo);
1638 #endif
1639         CARP_LOCK_DESTROY(sc);
1640
1641         free(sc->sc_ifas, M_CARP);
1642         free(sc, M_CARP);
1643 }
1644
1645 static struct carp_if*
1646 carp_alloc_if(struct ifnet *ifp)
1647 {
1648         struct carp_if *cif;
1649         int error;
1650
1651         cif = malloc(sizeof(*cif), M_CARP, M_WAITOK|M_ZERO);
1652
1653         if ((error = ifpromisc(ifp, 1)) != 0)
1654                 printf("%s: ifpromisc(%s) failed: %d\n",
1655                     __func__, ifp->if_xname, error);
1656         else
1657                 cif->cif_flags |= CIF_PROMISC;
1658
1659         CIF_LOCK_INIT(cif);
1660         cif->cif_ifp = ifp;
1661         TAILQ_INIT(&cif->cif_vrs);
1662
1663         IF_ADDR_WLOCK(ifp);
1664         ifp->if_carp = cif;
1665         if_ref(ifp);
1666         IF_ADDR_WUNLOCK(ifp);
1667
1668         return (cif);
1669 }
1670
1671 static void
1672 carp_free_if(struct carp_if *cif)
1673 {
1674         struct ifnet *ifp = cif->cif_ifp;
1675
1676         CIF_LOCK_ASSERT(cif);
1677         KASSERT(TAILQ_EMPTY(&cif->cif_vrs), ("%s: softc list not empty",
1678             __func__));
1679
1680         IF_ADDR_WLOCK(ifp);
1681         ifp->if_carp = NULL;
1682         IF_ADDR_WUNLOCK(ifp);
1683
1684         CIF_LOCK_DESTROY(cif);
1685
1686         if (cif->cif_flags & CIF_PROMISC)
1687                 ifpromisc(ifp, 0);
1688         if_rele(ifp);
1689
1690         free(cif, M_CARP);
1691 }
1692
1693 static void
1694 carp_carprcp(struct carpreq *carpr, struct carp_softc *sc, int priv)
1695 {
1696
1697         CARP_LOCK(sc);
1698         carpr->carpr_state = sc->sc_state;
1699         carpr->carpr_vhid = sc->sc_vhid;
1700         carpr->carpr_advbase = sc->sc_advbase;
1701         carpr->carpr_advskew = sc->sc_advskew;
1702         if (priv)
1703                 bcopy(sc->sc_key, carpr->carpr_key, sizeof(carpr->carpr_key));
1704         else
1705                 bzero(carpr->carpr_key, sizeof(carpr->carpr_key));
1706         CARP_UNLOCK(sc);
1707 }
1708
1709 int
1710 carp_ioctl(struct ifreq *ifr, u_long cmd, struct thread *td)
1711 {
1712         struct carpreq carpr;
1713         struct ifnet *ifp;
1714         struct carp_softc *sc = NULL;
1715         int error = 0, locked = 0;
1716
1717         if ((error = copyin(ifr_data_get_ptr(ifr), &carpr, sizeof carpr)))
1718                 return (error);
1719
1720         ifp = ifunit_ref(ifr->ifr_name);
1721         if (ifp == NULL)
1722                 return (ENXIO);
1723
1724         switch (ifp->if_type) {
1725         case IFT_ETHER:
1726         case IFT_L2VLAN:
1727         case IFT_BRIDGE:
1728                 break;
1729         default:
1730                 error = EOPNOTSUPP;
1731                 goto out;
1732         }
1733
1734         if ((ifp->if_flags & IFF_MULTICAST) == 0) {
1735                 error = EADDRNOTAVAIL;
1736                 goto out;
1737         }
1738
1739         sx_xlock(&carp_sx);
1740         switch (cmd) {
1741         case SIOCSVH:
1742                 if ((error = priv_check(td, PRIV_NETINET_CARP)))
1743                         break;
1744                 if (carpr.carpr_vhid <= 0 || carpr.carpr_vhid > CARP_MAXVHID ||
1745                     carpr.carpr_advbase < 0 || carpr.carpr_advskew < 0) {
1746                         error = EINVAL;
1747                         break;
1748                 }
1749
1750                 if (ifp->if_carp) {
1751                         IFNET_FOREACH_CARP(ifp, sc)
1752                                 if (sc->sc_vhid == carpr.carpr_vhid)
1753                                         break;
1754                 }
1755                 if (sc == NULL) {
1756                         sc = carp_alloc(ifp);
1757                         CARP_LOCK(sc);
1758                         sc->sc_vhid = carpr.carpr_vhid;
1759                         LLADDR(&sc->sc_addr)[0] = 0;
1760                         LLADDR(&sc->sc_addr)[1] = 0;
1761                         LLADDR(&sc->sc_addr)[2] = 0x5e;
1762                         LLADDR(&sc->sc_addr)[3] = 0;
1763                         LLADDR(&sc->sc_addr)[4] = 1;
1764                         LLADDR(&sc->sc_addr)[5] = sc->sc_vhid;
1765                 } else
1766                         CARP_LOCK(sc);
1767                 locked = 1;
1768                 if (carpr.carpr_advbase > 0) {
1769                         if (carpr.carpr_advbase > 255 ||
1770                             carpr.carpr_advbase < CARP_DFLTINTV) {
1771                                 error = EINVAL;
1772                                 break;
1773                         }
1774                         sc->sc_advbase = carpr.carpr_advbase;
1775                 }
1776                 if (carpr.carpr_advskew >= 255) {
1777                         error = EINVAL;
1778                         break;
1779                 }
1780                 sc->sc_advskew = carpr.carpr_advskew;
1781                 if (carpr.carpr_key[0] != '\0') {
1782                         bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1783                         carp_hmac_prepare(sc);
1784                 }
1785                 if (sc->sc_state != INIT &&
1786                     carpr.carpr_state != sc->sc_state) {
1787                         switch (carpr.carpr_state) {
1788                         case BACKUP:
1789                                 callout_stop(&sc->sc_ad_tmo);
1790                                 carp_set_state(sc, BACKUP,
1791                                     "user requested via ifconfig");
1792                                 carp_setrun(sc, 0);
1793                                 carp_delroute(sc);
1794                                 break;
1795                         case MASTER:
1796                                 carp_master_down_locked(sc,
1797                                     "user requested via ifconfig");
1798                                 break;
1799                         default:
1800                                 break;
1801                         }
1802                 }
1803                 break;
1804
1805         case SIOCGVH:
1806             {
1807                 int priveleged;
1808
1809                 if (carpr.carpr_vhid < 0 || carpr.carpr_vhid > CARP_MAXVHID) {
1810                         error = EINVAL;
1811                         break;
1812                 }
1813                 if (carpr.carpr_count < 1) {
1814                         error = EMSGSIZE;
1815                         break;
1816                 }
1817                 if (ifp->if_carp == NULL) {
1818                         error = ENOENT;
1819                         break;
1820                 }
1821
1822                 priveleged = (priv_check(td, PRIV_NETINET_CARP) == 0);
1823                 if (carpr.carpr_vhid != 0) {
1824                         IFNET_FOREACH_CARP(ifp, sc)
1825                                 if (sc->sc_vhid == carpr.carpr_vhid)
1826                                         break;
1827                         if (sc == NULL) {
1828                                 error = ENOENT;
1829                                 break;
1830                         }
1831                         carp_carprcp(&carpr, sc, priveleged);
1832                         error = copyout(&carpr, ifr_data_get_ptr(ifr),
1833                             sizeof(carpr));
1834                 } else  {
1835                         int i, count;
1836
1837                         count = 0;
1838                         IFNET_FOREACH_CARP(ifp, sc)
1839                                 count++;
1840
1841                         if (count > carpr.carpr_count) {
1842                                 CIF_UNLOCK(ifp->if_carp);
1843                                 error = EMSGSIZE;
1844                                 break;
1845                         }
1846
1847                         i = 0;
1848                         IFNET_FOREACH_CARP(ifp, sc) {
1849                                 carp_carprcp(&carpr, sc, priveleged);
1850                                 carpr.carpr_count = count;
1851                                 error = copyout(&carpr,
1852                                     (caddr_t)ifr_data_get_ptr(ifr) +
1853                                     (i * sizeof(carpr)), sizeof(carpr));
1854                                 if (error) {
1855                                         CIF_UNLOCK(ifp->if_carp);
1856                                         break;
1857                                 }
1858                                 i++;
1859                         }
1860                 }
1861                 break;
1862             }
1863         default:
1864                 error = EINVAL;
1865         }
1866         sx_xunlock(&carp_sx);
1867
1868 out:
1869         if (locked)
1870                 CARP_UNLOCK(sc);
1871         if_rele(ifp);
1872
1873         return (error);
1874 }
1875
1876 static int
1877 carp_get_vhid(struct ifaddr *ifa)
1878 {
1879
1880         if (ifa == NULL || ifa->ifa_carp == NULL)
1881                 return (0);
1882
1883         return (ifa->ifa_carp->sc_vhid);
1884 }
1885
1886 int
1887 carp_attach(struct ifaddr *ifa, int vhid)
1888 {
1889         struct ifnet *ifp = ifa->ifa_ifp;
1890         struct carp_if *cif = ifp->if_carp;
1891         struct carp_softc *sc;
1892         int index, error;
1893
1894         KASSERT(ifa->ifa_carp == NULL, ("%s: ifa %p attached", __func__, ifa));
1895
1896         switch (ifa->ifa_addr->sa_family) {
1897 #ifdef INET
1898         case AF_INET:
1899 #endif
1900 #ifdef INET6
1901         case AF_INET6:
1902 #endif
1903                 break;
1904         default:
1905                 return (EPROTOTYPE);
1906         }
1907
1908         sx_xlock(&carp_sx);
1909         if (ifp->if_carp == NULL) {
1910                 sx_xunlock(&carp_sx);
1911                 return (ENOPROTOOPT);
1912         }
1913
1914         IFNET_FOREACH_CARP(ifp, sc)
1915                 if (sc->sc_vhid == vhid)
1916                         break;
1917         if (sc == NULL) {
1918                 sx_xunlock(&carp_sx);
1919                 return (ENOENT);
1920         }
1921
1922         error = carp_multicast_setup(cif, ifa->ifa_addr->sa_family);
1923         if (error) {
1924                 CIF_FREE(cif);
1925                 sx_xunlock(&carp_sx);
1926                 return (error);
1927         }
1928
1929         index = sc->sc_naddrs + sc->sc_naddrs6 + 1;
1930         if (index > sc->sc_ifasiz / sizeof(struct ifaddr *))
1931                 carp_grow_ifas(sc);
1932
1933         switch (ifa->ifa_addr->sa_family) {
1934 #ifdef INET
1935         case AF_INET:
1936                 cif->cif_naddrs++;
1937                 sc->sc_naddrs++;
1938                 break;
1939 #endif
1940 #ifdef INET6
1941         case AF_INET6:
1942                 cif->cif_naddrs6++;
1943                 sc->sc_naddrs6++;
1944                 break;
1945 #endif
1946         }
1947
1948         ifa_ref(ifa);
1949
1950         CARP_LOCK(sc);
1951         sc->sc_ifas[index - 1] = ifa;
1952         ifa->ifa_carp = sc;
1953         carp_hmac_prepare(sc);
1954         carp_sc_state(sc);
1955         CARP_UNLOCK(sc);
1956
1957         sx_xunlock(&carp_sx);
1958
1959         return (0);
1960 }
1961
1962 void
1963 carp_detach(struct ifaddr *ifa, bool keep_cif)
1964 {
1965         struct ifnet *ifp = ifa->ifa_ifp;
1966         struct carp_if *cif = ifp->if_carp;
1967         struct carp_softc *sc = ifa->ifa_carp;
1968         int i, index;
1969
1970         KASSERT(sc != NULL, ("%s: %p not attached", __func__, ifa));
1971
1972         sx_xlock(&carp_sx);
1973
1974         CARP_LOCK(sc);
1975         /* Shift array. */
1976         index = sc->sc_naddrs + sc->sc_naddrs6;
1977         for (i = 0; i < index; i++)
1978                 if (sc->sc_ifas[i] == ifa)
1979                         break;
1980         KASSERT(i < index, ("%s: %p no backref", __func__, ifa));
1981         for (; i < index - 1; i++)
1982                 sc->sc_ifas[i] = sc->sc_ifas[i+1];
1983         sc->sc_ifas[index - 1] = NULL;
1984
1985         switch (ifa->ifa_addr->sa_family) {
1986 #ifdef INET
1987         case AF_INET:
1988                 cif->cif_naddrs--;
1989                 sc->sc_naddrs--;
1990                 break;
1991 #endif
1992 #ifdef INET6
1993         case AF_INET6:
1994                 cif->cif_naddrs6--;
1995                 sc->sc_naddrs6--;
1996                 break;
1997 #endif
1998         }
1999
2000         carp_ifa_delroute(ifa);
2001         carp_multicast_cleanup(cif, ifa->ifa_addr->sa_family);
2002
2003         ifa->ifa_carp = NULL;
2004         ifa_free(ifa);
2005
2006         carp_hmac_prepare(sc);
2007         carp_sc_state(sc);
2008
2009         if (!keep_cif && sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0)
2010                 carp_destroy(sc);
2011         else
2012                 CARP_UNLOCK(sc);
2013
2014         if (!keep_cif)
2015                 CIF_FREE(cif);
2016
2017         sx_xunlock(&carp_sx);
2018 }
2019
2020 static void
2021 carp_set_state(struct carp_softc *sc, int state, const char *reason)
2022 {
2023
2024         CARP_LOCK_ASSERT(sc);
2025
2026         if (sc->sc_state != state) {
2027                 const char *carp_states[] = { CARP_STATES };
2028                 char subsys[IFNAMSIZ+5];
2029
2030                 snprintf(subsys, IFNAMSIZ+5, "%u@%s", sc->sc_vhid,
2031                     sc->sc_carpdev->if_xname);
2032
2033                 CARP_LOG("%s: %s -> %s (%s)\n", subsys,
2034                     carp_states[sc->sc_state], carp_states[state], reason);
2035
2036                 sc->sc_state = state;
2037
2038                 devctl_notify("CARP", subsys, carp_states[state], NULL);
2039         }
2040 }
2041
2042 static void
2043 carp_linkstate(struct ifnet *ifp)
2044 {
2045         struct carp_softc *sc;
2046
2047         CIF_LOCK(ifp->if_carp);
2048         IFNET_FOREACH_CARP(ifp, sc) {
2049                 CARP_LOCK(sc);
2050                 carp_sc_state(sc);
2051                 CARP_UNLOCK(sc);
2052         }
2053         CIF_UNLOCK(ifp->if_carp);
2054 }
2055
2056 static void
2057 carp_sc_state(struct carp_softc *sc)
2058 {
2059
2060         CARP_LOCK_ASSERT(sc);
2061
2062         if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
2063             !(sc->sc_carpdev->if_flags & IFF_UP) ||
2064             !V_carp_allow) {
2065                 callout_stop(&sc->sc_ad_tmo);
2066 #ifdef INET
2067                 callout_stop(&sc->sc_md_tmo);
2068 #endif
2069 #ifdef INET6
2070                 callout_stop(&sc->sc_md6_tmo);
2071 #endif
2072                 carp_set_state(sc, INIT, "hardware interface down");
2073                 carp_setrun(sc, 0);
2074                 if (!sc->sc_suppress)
2075                         carp_demote_adj(V_carp_ifdown_adj, "interface down");
2076                 sc->sc_suppress = 1;
2077         } else {
2078                 carp_set_state(sc, INIT, "hardware interface up");
2079                 carp_setrun(sc, 0);
2080                 if (sc->sc_suppress)
2081                         carp_demote_adj(-V_carp_ifdown_adj, "interface up");
2082                 sc->sc_suppress = 0;
2083         }
2084 }
2085
2086 static void
2087 carp_demote_adj(int adj, char *reason)
2088 {
2089         atomic_add_int(&V_carp_demotion, adj);
2090         CARP_LOG("demoted by %d to %d (%s)\n", adj, V_carp_demotion, reason);
2091         taskqueue_enqueue(taskqueue_swi, &carp_sendall_task);
2092 }
2093
2094 static int
2095 carp_allow_sysctl(SYSCTL_HANDLER_ARGS)
2096 {
2097         int new, error;
2098         struct carp_softc *sc;
2099
2100         new = V_carp_allow;
2101         error = sysctl_handle_int(oidp, &new, 0, req);
2102         if (error || !req->newptr)
2103                 return (error);
2104
2105         if (V_carp_allow != new) {
2106                 V_carp_allow = new;
2107
2108                 mtx_lock(&carp_mtx);
2109                 LIST_FOREACH(sc, &carp_list, sc_next) {
2110                         CARP_LOCK(sc);
2111                         if (curvnet == sc->sc_carpdev->if_vnet)
2112                                 carp_sc_state(sc);
2113                         CARP_UNLOCK(sc);
2114                 }
2115                 mtx_unlock(&carp_mtx);
2116         }
2117
2118         return (0);
2119 }
2120
2121 static int
2122 carp_dscp_sysctl(SYSCTL_HANDLER_ARGS)
2123 {
2124         int new, error;
2125
2126         new = V_carp_dscp;
2127         error = sysctl_handle_int(oidp, &new, 0, req);
2128         if (error || !req->newptr)
2129                 return (error);
2130
2131         if (new < 0 || new > 63)
2132                 return (EINVAL);
2133
2134         V_carp_dscp = new;
2135
2136         return (0);
2137 }
2138
2139 static int
2140 carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS)
2141 {
2142         int new, error;
2143
2144         new = V_carp_demotion;
2145         error = sysctl_handle_int(oidp, &new, 0, req);
2146         if (error || !req->newptr)
2147                 return (error);
2148
2149         carp_demote_adj(new, "sysctl");
2150
2151         return (0);
2152 }
2153
2154 #ifdef INET
2155 extern  struct domain inetdomain;
2156 static struct protosw in_carp_protosw = {
2157         .pr_type =              SOCK_RAW,
2158         .pr_domain =            &inetdomain,
2159         .pr_protocol =          IPPROTO_CARP,
2160         .pr_flags =             PR_ATOMIC|PR_ADDR,
2161         .pr_input =             carp_input,
2162         .pr_output =            rip_output,
2163         .pr_ctloutput =         rip_ctloutput,
2164         .pr_usrreqs =           &rip_usrreqs
2165 };
2166 #endif
2167
2168 #ifdef INET6
2169 extern  struct domain inet6domain;
2170 static struct protosw in6_carp_protosw = {
2171         .pr_type =              SOCK_RAW,
2172         .pr_domain =            &inet6domain,
2173         .pr_protocol =          IPPROTO_CARP,
2174         .pr_flags =             PR_ATOMIC|PR_ADDR,
2175         .pr_input =             carp6_input,
2176         .pr_output =            rip6_output,
2177         .pr_ctloutput =         rip6_ctloutput,
2178         .pr_usrreqs =           &rip6_usrreqs
2179 };
2180 #endif
2181
2182 #ifdef VIMAGE
2183 #if defined(__i386__)
2184 /*
2185  * XXX This is a hack to work around an absolute relocation outside
2186  * set_vnet by one (on the stop symbol) for carpstats.  Add a dummy variable
2187  * to the end of the file in the hope that the linker will just keep the
2188  * order (as it seems to do at the moment).  It is understood to be fragile.
2189  * See PR 230857 for a longer discussion of the problem and the referenced
2190  * review for possible alternate solutions.  Each is a hack; we just need
2191  * the least intrusive one for the next release.
2192  */
2193 VNET_DEFINE(char, carp_zzz) = 0xde;
2194 #endif
2195 #endif
2196
2197 static void
2198 carp_mod_cleanup(void)
2199 {
2200
2201 #ifdef INET
2202         if (proto_reg[CARP_INET] == 0) {
2203                 (void)ipproto_unregister(IPPROTO_CARP);
2204                 pf_proto_unregister(PF_INET, IPPROTO_CARP, SOCK_RAW);
2205                 proto_reg[CARP_INET] = -1;
2206         }
2207         carp_iamatch_p = NULL;
2208 #endif
2209 #ifdef INET6
2210         if (proto_reg[CARP_INET6] == 0) {
2211                 (void)ip6proto_unregister(IPPROTO_CARP);
2212                 pf_proto_unregister(PF_INET6, IPPROTO_CARP, SOCK_RAW);
2213                 proto_reg[CARP_INET6] = -1;
2214         }
2215         carp_iamatch6_p = NULL;
2216         carp_macmatch6_p = NULL;
2217 #endif
2218         carp_ioctl_p = NULL;
2219         carp_attach_p = NULL;
2220         carp_detach_p = NULL;
2221         carp_get_vhid_p = NULL;
2222         carp_linkstate_p = NULL;
2223         carp_forus_p = NULL;
2224         carp_output_p = NULL;
2225         carp_demote_adj_p = NULL;
2226         carp_master_p = NULL;
2227         mtx_unlock(&carp_mtx);
2228         taskqueue_drain(taskqueue_swi, &carp_sendall_task);
2229         mtx_destroy(&carp_mtx);
2230         sx_destroy(&carp_sx);
2231 }
2232
2233 static int
2234 carp_mod_load(void)
2235 {
2236         int err;
2237
2238         mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2239         sx_init(&carp_sx, "carp_sx");
2240         LIST_INIT(&carp_list);
2241         carp_get_vhid_p = carp_get_vhid;
2242         carp_forus_p = carp_forus;
2243         carp_output_p = carp_output;
2244         carp_linkstate_p = carp_linkstate;
2245         carp_ioctl_p = carp_ioctl;
2246         carp_attach_p = carp_attach;
2247         carp_detach_p = carp_detach;
2248         carp_demote_adj_p = carp_demote_adj;
2249         carp_master_p = carp_master;
2250 #ifdef INET6
2251         carp_iamatch6_p = carp_iamatch6;
2252         carp_macmatch6_p = carp_macmatch6;
2253         proto_reg[CARP_INET6] = pf_proto_register(PF_INET6,
2254             (struct protosw *)&in6_carp_protosw);
2255         if (proto_reg[CARP_INET6]) {
2256                 printf("carp: error %d attaching to PF_INET6\n",
2257                     proto_reg[CARP_INET6]);
2258                 carp_mod_cleanup();
2259                 return (proto_reg[CARP_INET6]);
2260         }
2261         err = ip6proto_register(IPPROTO_CARP);
2262         if (err) {
2263                 printf("carp: error %d registering with INET6\n", err);
2264                 carp_mod_cleanup();
2265                 return (err);
2266         }
2267 #endif
2268 #ifdef INET
2269         carp_iamatch_p = carp_iamatch;
2270         proto_reg[CARP_INET] = pf_proto_register(PF_INET, &in_carp_protosw);
2271         if (proto_reg[CARP_INET]) {
2272                 printf("carp: error %d attaching to PF_INET\n",
2273                     proto_reg[CARP_INET]);
2274                 carp_mod_cleanup();
2275                 return (proto_reg[CARP_INET]);
2276         }
2277         err = ipproto_register(IPPROTO_CARP);
2278         if (err) {
2279                 printf("carp: error %d registering with INET\n", err);
2280                 carp_mod_cleanup();
2281                 return (err);
2282         }
2283 #endif
2284         return (0);
2285 }
2286
2287 static int
2288 carp_modevent(module_t mod, int type, void *data)
2289 {
2290         switch (type) {
2291         case MOD_LOAD:
2292                 return carp_mod_load();
2293                 /* NOTREACHED */
2294         case MOD_UNLOAD:
2295                 mtx_lock(&carp_mtx);
2296                 if (LIST_EMPTY(&carp_list))
2297                         carp_mod_cleanup();
2298                 else {
2299                         mtx_unlock(&carp_mtx);
2300                         return (EBUSY);
2301                 }
2302                 break;
2303
2304         default:
2305                 return (EINVAL);
2306         }
2307
2308         return (0);
2309 }
2310
2311 static moduledata_t carp_mod = {
2312         "carp",
2313         carp_modevent,
2314         0
2315 };
2316
2317 DECLARE_MODULE(carp, carp_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);