1 /* $FreeBSD$ */ 2 3 /* 4 * Copyright (c) 2002 Michael Shalayeff. All rights reserved. 5 * Copyright (c) 2003 Ryan McBride. 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 "opt_carp.h" 30 #include "opt_bpf.h" 31 #include "opt_inet.h" 32 #include "opt_inet6.h" 33 34 #include <sys/types.h> 35 #include <sys/param.h> 36 #include <sys/systm.h> 37 #include <sys/conf.h> 38 #include <sys/kernel.h> 39 #include <sys/limits.h> 40 #include <sys/malloc.h> 41 #include <sys/mbuf.h> 42 #include <sys/module.h> 43 #include <sys/time.h> 44 #include <sys/proc.h> 45 #include <sys/sysctl.h> 46 #include <sys/syslog.h> 47 #include <sys/signalvar.h> 48 #include <sys/filio.h> 49 #include <sys/sockio.h> 50 51 #include <sys/socket.h> 52 #include <sys/vnode.h> 53 54 #include <machine/stdarg.h> 55 56 #include <net/bpf.h> 57 #include <net/ethernet.h> 58 #include <net/fddi.h> 59 #include <net/iso88025.h> 60 #include <net/if.h> 61 #include <net/if_clone.h> 62 #include <net/if_types.h> 63 #include <net/route.h> 64 65 #ifdef INET 66 #include <netinet/in.h> 67 #include <netinet/in_var.h> 68 #include <netinet/in_systm.h> 69 #include <netinet/ip.h> 70 #include <netinet/ip_var.h> 71 #include <netinet/if_ether.h> 72 #include <machine/in_cksum.h> 73 #endif 74 75 #ifdef INET6 76 #include <netinet/icmp6.h> 77 #include <netinet/ip6.h> 78 #include <netinet6/ip6_var.h> 79 #include <netinet6/nd6.h> 80 #include <net/if_dl.h> 81 #endif 82 83 #include <crypto/sha1.h> 84 #include <netinet/ip_carp.h> 85 86 #define CARP_IFNAME "carp" 87 static MALLOC_DEFINE(M_CARP, "CARP", "CARP interfaces"); 88 SYSCTL_DECL(_net_inet_carp); 89 90 struct carp_softc { 91 struct ifnet *sc_ifp; /* Interface clue */ 92 struct ifnet *sc_carpdev; /* Pointer to parent interface */ 93 struct in_ifaddr *sc_ia; /* primary iface address */ 94 struct ip_moptions sc_imo; 95 #ifdef INET6 96 struct in6_ifaddr *sc_ia6; /* primary iface address v6 */ 97 struct ip6_moptions sc_im6o; 98 #endif /* INET6 */ 99 TAILQ_ENTRY(carp_softc) sc_list; 100 101 enum { INIT = 0, BACKUP, MASTER } sc_state; 102 103 int sc_flags_backup; 104 int sc_suppress; 105 106 int sc_sendad_errors; 107 #define CARP_SENDAD_MAX_ERRORS 3 108 int sc_sendad_success; 109 #define CARP_SENDAD_MIN_SUCCESS 3 110 111 int sc_vhid; 112 int sc_advskew; 113 int sc_naddrs; 114 int sc_naddrs6; 115 int sc_advbase; /* seconds */ 116 int sc_init_counter; 117 u_int64_t sc_counter; 118 119 /* authentication */ 120 #define CARP_HMAC_PAD 64 121 unsigned char sc_key[CARP_KEY_LEN]; 122 unsigned char sc_pad[CARP_HMAC_PAD]; 123 SHA1_CTX sc_sha1; 124 125 struct callout sc_ad_tmo; /* advertisement timeout */ 126 struct callout sc_md_tmo; /* master down timeout */ 127 struct callout sc_md6_tmo; /* master down timeout */ 128 129 LIST_ENTRY(carp_softc) sc_next; /* Interface clue */ 130 }; 131 #define SC2IFP(sc) ((sc)->sc_ifp) 132 133 int carp_suppress_preempt = 0; 134 int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 1, 0, 0 }; /* XXX for now */ 135 SYSCTL_INT(_net_inet_carp, CARPCTL_ALLOW, allow, CTLFLAG_RW, 136 &carp_opts[CARPCTL_ALLOW], 0, "Accept incoming CARP packets"); 137 SYSCTL_INT(_net_inet_carp, CARPCTL_PREEMPT, preempt, CTLFLAG_RW, 138 &carp_opts[CARPCTL_PREEMPT], 0, "high-priority backup preemption mode"); 139 SYSCTL_INT(_net_inet_carp, CARPCTL_LOG, log, CTLFLAG_RW, 140 &carp_opts[CARPCTL_LOG], 0, "log bad carp packets"); 141 SYSCTL_INT(_net_inet_carp, CARPCTL_ARPBALANCE, arpbalance, CTLFLAG_RW, 142 &carp_opts[CARPCTL_ARPBALANCE], 0, "balance arp responses"); 143 SYSCTL_INT(_net_inet_carp, OID_AUTO, suppress_preempt, CTLFLAG_RD, 144 &carp_suppress_preempt, 0, "Preemption is suppressed"); 145 146 struct carpstats carpstats; 147 SYSCTL_STRUCT(_net_inet_carp, CARPCTL_STATS, stats, CTLFLAG_RW, 148 &carpstats, carpstats, 149 "CARP statistics (struct carpstats, netinet/ip_carp.h)"); 150 151 struct carp_if { 152 TAILQ_HEAD(, carp_softc) vhif_vrs; 153 int vhif_nvrs; 154 155 struct ifnet *vhif_ifp; 156 struct mtx vhif_mtx; 157 }; 158 159 /* Get carp_if from softc. Valid after carp_set_addr{,6}. */ 160 #define SC2CIF(sc) ((struct carp_if *)(sc)->sc_carpdev->if_carp) 161 162 /* lock per carp_if queue */ 163 #define CARP_LOCK_INIT(cif) mtx_init(&(cif)->vhif_mtx, "carp_if", \ 164 NULL, MTX_DEF) 165 #define CARP_LOCK_DESTROY(cif) mtx_destroy(&(cif)->vhif_mtx) 166 #define CARP_LOCK_ASSERT(cif) mtx_assert(&(cif)->vhif_mtx, MA_OWNED) 167 #define CARP_LOCK(cif) mtx_lock(&(cif)->vhif_mtx) 168 #define CARP_UNLOCK(cif) mtx_unlock(&(cif)->vhif_mtx) 169 170 #define CARP_SCLOCK(sc) mtx_lock(&SC2CIF(sc)->vhif_mtx) 171 #define CARP_SCUNLOCK(sc) mtx_unlock(&SC2CIF(sc)->vhif_mtx) 172 #define CARP_SCLOCK_ASSERT(sc) mtx_assert(&SC2CIF(sc)->vhif_mtx, MA_OWNED) 173 174 #define CARP_LOG(...) do { \ 175 if (carp_opts[CARPCTL_LOG] > 0) \ 176 log(LOG_INFO, __VA_ARGS__); \ 177 } while (0) 178 179 #define CARP_DEBUG(...) do { \ 180 if (carp_opts[CARPCTL_LOG] > 1) \ 181 log(LOG_DEBUG, __VA_ARGS__); \ 182 } while (0) 183 184 static void carp_hmac_prepare(struct carp_softc *); 185 static void carp_hmac_generate(struct carp_softc *, u_int32_t *, 186 unsigned char *); 187 static int carp_hmac_verify(struct carp_softc *, u_int32_t *, 188 unsigned char *); 189 static void carp_setroute(struct carp_softc *, int); 190 static void carp_input_c(struct mbuf *, struct carp_header *, sa_family_t); 191 static int carp_clone_create(struct if_clone *, int); 192 static void carp_clone_destroy(struct ifnet *); 193 static void carpdetach(struct carp_softc *); 194 static int carp_prepare_ad(struct mbuf *, struct carp_softc *, 195 struct carp_header *); 196 static void carp_send_ad_all(void); 197 static void carp_send_ad(void *); 198 static void carp_send_ad_locked(struct carp_softc *); 199 static void carp_send_arp(struct carp_softc *); 200 static void carp_master_down(void *); 201 static void carp_master_down_locked(struct carp_softc *); 202 static int carp_ioctl(struct ifnet *, u_long, caddr_t); 203 static int carp_looutput(struct ifnet *, struct mbuf *, struct sockaddr *, 204 struct rtentry *); 205 static void carp_start(struct ifnet *); 206 static void carp_setrun(struct carp_softc *, sa_family_t); 207 static void carp_set_state(struct carp_softc *, int); 208 static int carp_addrcount(struct carp_if *, struct in_ifaddr *, int); 209 enum { CARP_COUNT_MASTER, CARP_COUNT_RUNNING }; 210 211 static int carp_set_addr(struct carp_softc *, struct sockaddr_in *); 212 static int carp_del_addr(struct carp_softc *, struct sockaddr_in *); 213 static void carp_carpdev_state_locked(struct carp_if *); 214 static void carp_sc_state_locked(struct carp_softc *); 215 #ifdef INET6 216 static void carp_send_na(struct carp_softc *); 217 static int carp_set_addr6(struct carp_softc *, struct sockaddr_in6 *); 218 static int carp_del_addr6(struct carp_softc *, struct sockaddr_in6 *); 219 #endif 220 221 static LIST_HEAD(, carp_softc) carpif_list; 222 static struct mtx carp_mtx; 223 IFC_SIMPLE_DECLARE(carp, 0); 224 225 static __inline u_int16_t 226 carp_cksum(struct mbuf *m, int len) 227 { 228 return (in_cksum(m, len)); 229 } 230 231 static void 232 carp_hmac_prepare(struct carp_softc *sc) 233 { 234 u_int8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT; 235 u_int8_t vhid = sc->sc_vhid & 0xff; 236 struct ifaddr *ifa; 237 int i; 238 #ifdef INET6 239 struct in6_addr in6; 240 #endif 241 242 if (sc->sc_carpdev) 243 CARP_SCLOCK(sc); 244 245 /* XXX: possible race here */ 246 247 /* compute ipad from key */ 248 bzero(sc->sc_pad, sizeof(sc->sc_pad)); 249 bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key)); 250 for (i = 0; i < sizeof(sc->sc_pad); i++) 251 sc->sc_pad[i] ^= 0x36; 252 253 /* precompute first part of inner hash */ 254 SHA1Init(&sc->sc_sha1); 255 SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad)); 256 SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version)); 257 SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type)); 258 SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid)); 259 #ifdef INET 260 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) { 261 if (ifa->ifa_addr->sa_family == AF_INET) 262 SHA1Update(&sc->sc_sha1, 263 (void *)&ifatoia(ifa)->ia_addr.sin_addr.s_addr, 264 sizeof(struct in_addr)); 265 } 266 #endif /* INET */ 267 #ifdef INET6 268 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) { 269 if (ifa->ifa_addr->sa_family == AF_INET6) { 270 in6 = ifatoia6(ifa)->ia_addr.sin6_addr; 271 if (IN6_IS_ADDR_LINKLOCAL(&in6)) 272 in6.s6_addr16[1] = 0; 273 SHA1Update(&sc->sc_sha1, (void *)&in6, sizeof(in6)); 274 } 275 } 276 #endif /* INET6 */ 277 278 /* convert ipad to opad */ 279 for (i = 0; i < sizeof(sc->sc_pad); i++) 280 sc->sc_pad[i] ^= 0x36 ^ 0x5c; 281 282 if (sc->sc_carpdev) 283 CARP_SCUNLOCK(sc); 284 } 285 286 static void 287 carp_hmac_generate(struct carp_softc *sc, u_int32_t counter[2], 288 unsigned char md[20]) 289 { 290 SHA1_CTX sha1ctx; 291 292 /* fetch first half of inner hash */ 293 bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx)); 294 295 SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter)); 296 SHA1Final(md, &sha1ctx); 297 298 /* outer hash */ 299 SHA1Init(&sha1ctx); 300 SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad)); 301 SHA1Update(&sha1ctx, md, 20); 302 SHA1Final(md, &sha1ctx); 303 } 304 305 static int 306 carp_hmac_verify(struct carp_softc *sc, u_int32_t counter[2], 307 unsigned char md[20]) 308 { 309 unsigned char md2[20]; 310 311 CARP_SCLOCK_ASSERT(sc); 312 313 carp_hmac_generate(sc, counter, md2); 314 315 return (bcmp(md, md2, sizeof(md2))); 316 } 317 318 static void 319 carp_setroute(struct carp_softc *sc, int cmd) 320 { 321 struct ifaddr *ifa; 322 int s; 323 324 if (sc->sc_carpdev) 325 CARP_SCLOCK_ASSERT(sc); 326 327 s = splnet(); 328 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) { 329 if (ifa->ifa_addr->sa_family == AF_INET && 330 sc->sc_carpdev != NULL) { 331 int count = carp_addrcount( 332 (struct carp_if *)sc->sc_carpdev->if_carp, 333 ifatoia(ifa), CARP_COUNT_MASTER); 334 335 if ((cmd == RTM_ADD && count == 1) || 336 (cmd == RTM_DELETE && count == 0)) 337 rtinit(ifa, cmd, RTF_UP | RTF_HOST); 338 } 339 #ifdef INET6 340 if (ifa->ifa_addr->sa_family == AF_INET6) { 341 if (cmd == RTM_ADD) 342 in6_ifaddloop(ifa); 343 else 344 in6_ifremloop(ifa); 345 } 346 #endif /* INET6 */ 347 } 348 splx(s); 349 } 350 351 static int 352 carp_clone_create(struct if_clone *ifc, int unit) 353 { 354 355 struct carp_softc *sc; 356 struct ifnet *ifp; 357 358 MALLOC(sc, struct carp_softc *, sizeof(*sc), M_CARP, M_WAITOK|M_ZERO); 359 ifp = SC2IFP(sc) = if_alloc(IFT_ETHER); 360 if (ifp == NULL) { 361 FREE(sc, M_CARP); 362 return (ENOSPC); 363 } 364 365 sc->sc_flags_backup = 0; 366 sc->sc_suppress = 0; 367 sc->sc_advbase = CARP_DFLTINTV; 368 sc->sc_vhid = -1; /* required setting */ 369 sc->sc_advskew = 0; 370 sc->sc_init_counter = 1; 371 sc->sc_naddrs = sc->sc_naddrs6 = 0; /* M_ZERO? */ 372 #ifdef INET6 373 sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL; 374 #endif 375 376 callout_init(&sc->sc_ad_tmo, NET_CALLOUT_MPSAFE); 377 callout_init(&sc->sc_md_tmo, NET_CALLOUT_MPSAFE); 378 callout_init(&sc->sc_md6_tmo, NET_CALLOUT_MPSAFE); 379 380 ifp->if_softc = sc; 381 if_initname(ifp, CARP_IFNAME, unit); 382 ifp->if_mtu = ETHERMTU; 383 ifp->if_flags = 0; 384 ifp->if_ioctl = carp_ioctl; 385 ifp->if_output = carp_looutput; 386 ifp->if_start = carp_start; 387 ifp->if_type = IFT_CARP; 388 ifp->if_snd.ifq_maxlen = ifqmaxlen; 389 ifp->if_hdrlen = 0; 390 if_attach(ifp); 391 bpfattach(SC2IFP(sc), DLT_NULL, sizeof(u_int32_t)); 392 mtx_lock(&carp_mtx); 393 LIST_INSERT_HEAD(&carpif_list, sc, sc_next); 394 mtx_unlock(&carp_mtx); 395 return (0); 396 } 397 398 static void 399 carp_clone_destroy(struct ifnet *ifp) 400 { 401 struct carp_softc *sc = ifp->if_softc; 402 struct carp_if *cif; 403 struct ip_moptions *imo = &sc->sc_imo; 404 #ifdef INET6 405 struct ip6_moptions *im6o = &sc->sc_im6o; 406 #endif 407 408 /* carpdetach(sc); */ 409 410 /* 411 * If an interface is destroyed which is suppressing the preemption, 412 * decrease the global counter, otherwise the host will never get 413 * out of the carp supressing state. 414 */ 415 if (sc->sc_suppress) 416 carp_suppress_preempt--; 417 sc->sc_suppress = 0; 418 419 callout_stop(&sc->sc_ad_tmo); 420 callout_stop(&sc->sc_md_tmo); 421 callout_stop(&sc->sc_md6_tmo); 422 423 if (imo->imo_num_memberships) { 424 in_delmulti(imo->imo_membership[--imo->imo_num_memberships]); 425 imo->imo_multicast_ifp = NULL; 426 } 427 #ifdef INET6 428 while (!LIST_EMPTY(&im6o->im6o_memberships)) { 429 struct in6_multi_mship *imm = 430 LIST_FIRST(&im6o->im6o_memberships); 431 LIST_REMOVE(imm, i6mm_chain); 432 in6_leavegroup(imm); 433 } 434 im6o->im6o_multicast_ifp = NULL; 435 #endif 436 437 /* Remove ourself from parents if_carp queue */ 438 if (sc->sc_carpdev && (cif = sc->sc_carpdev->if_carp)) { 439 CARP_LOCK(cif); 440 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list); 441 if (!--cif->vhif_nvrs) { 442 sc->sc_carpdev->if_carp = NULL; 443 CARP_LOCK_DESTROY(cif); 444 FREE(cif, M_CARP); 445 ifpromisc(sc->sc_carpdev, 0); 446 } else { 447 CARP_UNLOCK(cif); 448 } 449 } 450 451 mtx_lock(&carp_mtx); 452 LIST_REMOVE(sc, sc_next); 453 mtx_unlock(&carp_mtx); 454 bpfdetach(ifp); 455 if_detach(ifp); 456 if_free_type(ifp, IFT_ETHER); 457 free(sc, M_CARP); 458 } 459 460 /* 461 * process input packet. 462 * we have rearranged checks order compared to the rfc, 463 * but it seems more efficient this way or not possible otherwise. 464 */ 465 void 466 carp_input(struct mbuf *m, int hlen) 467 { 468 struct ip *ip = mtod(m, struct ip *); 469 struct carp_header *ch; 470 int iplen, len; 471 472 carpstats.carps_ipackets++; 473 474 if (!carp_opts[CARPCTL_ALLOW]) { 475 m_freem(m); 476 return; 477 } 478 479 /* check if received on a valid carp interface */ 480 if (m->m_pkthdr.rcvif->if_carp == NULL) { 481 carpstats.carps_badif++; 482 CARP_LOG("carp_input: packet received on non-carp " 483 "interface: %s\n", 484 m->m_pkthdr.rcvif->if_xname); 485 m_freem(m); 486 return; 487 } 488 489 /* verify that the IP TTL is 255. */ 490 if (ip->ip_ttl != CARP_DFLTTL) { 491 carpstats.carps_badttl++; 492 CARP_LOG("carp_input: received ttl %d != 255i on %s\n", 493 ip->ip_ttl, 494 m->m_pkthdr.rcvif->if_xname); 495 m_freem(m); 496 return; 497 } 498 499 iplen = ip->ip_hl << 2; 500 501 if (m->m_pkthdr.len < iplen + sizeof(*ch)) { 502 carpstats.carps_badlen++; 503 CARP_LOG("carp_input: received len %zd < " 504 "sizeof(struct carp_header)\n", 505 m->m_len - sizeof(struct ip)); 506 m_freem(m); 507 return; 508 } 509 510 if (iplen + sizeof(*ch) < m->m_len) { 511 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) { 512 carpstats.carps_hdrops++; 513 CARP_LOG("carp_input: pullup failed\n"); 514 return; 515 } 516 ip = mtod(m, struct ip *); 517 } 518 ch = (struct carp_header *)((char *)ip + iplen); 519 520 /* 521 * verify that the received packet length is 522 * equal to the CARP header 523 */ 524 len = iplen + sizeof(*ch); 525 if (len > m->m_pkthdr.len) { 526 carpstats.carps_badlen++; 527 CARP_LOG("carp_input: packet too short %d on %s\n", 528 m->m_pkthdr.len, 529 m->m_pkthdr.rcvif->if_xname); 530 m_freem(m); 531 return; 532 } 533 534 if ((m = m_pullup(m, len)) == NULL) { 535 carpstats.carps_hdrops++; 536 return; 537 } 538 ip = mtod(m, struct ip *); 539 ch = (struct carp_header *)((char *)ip + iplen); 540 541 /* verify the CARP checksum */ 542 m->m_data += iplen; 543 if (carp_cksum(m, len - iplen)) { 544 carpstats.carps_badsum++; 545 CARP_LOG("carp_input: checksum failed on %s\n", 546 m->m_pkthdr.rcvif->if_xname); 547 m_freem(m); 548 return; 549 } 550 m->m_data -= iplen; 551 552 carp_input_c(m, ch, AF_INET); 553 } 554 555 #ifdef INET6 556 int 557 carp6_input(struct mbuf **mp, int *offp, int proto) 558 { 559 struct mbuf *m = *mp; 560 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 561 struct carp_header *ch; 562 u_int len; 563 564 carpstats.carps_ipackets6++; 565 566 if (!carp_opts[CARPCTL_ALLOW]) { 567 m_freem(m); 568 return (IPPROTO_DONE); 569 } 570 571 /* check if received on a valid carp interface */ 572 if (m->m_pkthdr.rcvif->if_carp == NULL) { 573 carpstats.carps_badif++; 574 CARP_LOG("carp6_input: packet received on non-carp " 575 "interface: %s\n", 576 m->m_pkthdr.rcvif->if_xname); 577 m_freem(m); 578 return (IPPROTO_DONE); 579 } 580 581 /* verify that the IP TTL is 255 */ 582 if (ip6->ip6_hlim != CARP_DFLTTL) { 583 carpstats.carps_badttl++; 584 CARP_LOG("carp6_input: received ttl %d != 255 on %s\n", 585 ip6->ip6_hlim, 586 m->m_pkthdr.rcvif->if_xname); 587 m_freem(m); 588 return (IPPROTO_DONE); 589 } 590 591 /* verify that we have a complete carp packet */ 592 len = m->m_len; 593 IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch)); 594 if (ch == NULL) { 595 carpstats.carps_badlen++; 596 CARP_LOG("carp6_input: packet size %u too small\n", len); 597 return (IPPROTO_DONE); 598 } 599 600 601 /* verify the CARP checksum */ 602 m->m_data += *offp; 603 if (carp_cksum(m, sizeof(*ch))) { 604 carpstats.carps_badsum++; 605 CARP_LOG("carp6_input: checksum failed, on %s\n", 606 m->m_pkthdr.rcvif->if_xname); 607 m_freem(m); 608 return (IPPROTO_DONE); 609 } 610 m->m_data -= *offp; 611 612 carp_input_c(m, ch, AF_INET6); 613 return (IPPROTO_DONE); 614 } 615 #endif /* INET6 */ 616 617 static void 618 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af) 619 { 620 struct ifnet *ifp = m->m_pkthdr.rcvif; 621 struct carp_softc *sc; 622 u_int64_t tmp_counter; 623 struct timeval sc_tv, ch_tv; 624 625 /* verify that the VHID is valid on the receiving interface */ 626 CARP_LOCK(ifp->if_carp); 627 TAILQ_FOREACH(sc, &((struct carp_if *)ifp->if_carp)->vhif_vrs, sc_list) 628 if (sc->sc_vhid == ch->carp_vhid) 629 break; 630 631 if (!sc || (SC2IFP(sc)->if_flags & (IFF_UP|IFF_RUNNING)) != 632 (IFF_UP|IFF_RUNNING)) { 633 carpstats.carps_badvhid++; 634 CARP_UNLOCK(ifp->if_carp); 635 m_freem(m); 636 return; 637 } 638 639 getmicrotime(&SC2IFP(sc)->if_lastchange); 640 SC2IFP(sc)->if_ipackets++; 641 SC2IFP(sc)->if_ibytes += m->m_pkthdr.len; 642 643 if (SC2IFP(sc)->if_bpf) { 644 struct ip *ip = mtod(m, struct ip *); 645 uint32_t af1 = af; 646 647 /* BPF wants net byte order */ 648 ip->ip_len = htons(ip->ip_len + (ip->ip_hl << 2)); 649 ip->ip_off = htons(ip->ip_off); 650 bpf_mtap2(SC2IFP(sc)->if_bpf, &af1, sizeof(af1), m); 651 } 652 653 /* verify the CARP version. */ 654 if (ch->carp_version != CARP_VERSION) { 655 carpstats.carps_badver++; 656 SC2IFP(sc)->if_ierrors++; 657 CARP_UNLOCK(ifp->if_carp); 658 CARP_LOG("%s; invalid version %d\n", 659 SC2IFP(sc)->if_xname, 660 ch->carp_version); 661 m_freem(m); 662 return; 663 } 664 665 /* verify the hash */ 666 if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) { 667 carpstats.carps_badauth++; 668 SC2IFP(sc)->if_ierrors++; 669 CARP_UNLOCK(ifp->if_carp); 670 CARP_LOG("%s: incorrect hash\n", SC2IFP(sc)->if_xname); 671 m_freem(m); 672 return; 673 } 674 675 tmp_counter = ntohl(ch->carp_counter[0]); 676 tmp_counter = tmp_counter<<32; 677 tmp_counter += ntohl(ch->carp_counter[1]); 678 679 /* XXX Replay protection goes here */ 680 681 sc->sc_init_counter = 0; 682 sc->sc_counter = tmp_counter; 683 684 sc_tv.tv_sec = sc->sc_advbase; 685 if (carp_suppress_preempt && sc->sc_advskew < 240) 686 sc_tv.tv_usec = 240 * 1000000 / 256; 687 else 688 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256; 689 ch_tv.tv_sec = ch->carp_advbase; 690 ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256; 691 692 switch (sc->sc_state) { 693 case INIT: 694 break; 695 case MASTER: 696 /* 697 * If we receive an advertisement from a master who's going to 698 * be more frequent than us, go into BACKUP state. 699 */ 700 if (timevalcmp(&sc_tv, &ch_tv, >) || 701 timevalcmp(&sc_tv, &ch_tv, ==)) { 702 callout_stop(&sc->sc_ad_tmo); 703 CARP_DEBUG("%s: MASTER -> BACKUP " 704 "(more frequent advertisement received)\n", 705 SC2IFP(sc)->if_xname); 706 carp_set_state(sc, BACKUP); 707 carp_setrun(sc, 0); 708 carp_setroute(sc, RTM_DELETE); 709 } 710 break; 711 case BACKUP: 712 /* 713 * If we're pre-empting masters who advertise slower than us, 714 * and this one claims to be slower, treat him as down. 715 */ 716 if (carp_opts[CARPCTL_PREEMPT] && 717 timevalcmp(&sc_tv, &ch_tv, <)) { 718 CARP_DEBUG("%s: BACKUP -> MASTER " 719 "(preempting a slower master)\n", 720 SC2IFP(sc)->if_xname); 721 carp_master_down_locked(sc); 722 break; 723 } 724 725 /* 726 * If the master is going to advertise at such a low frequency 727 * that he's guaranteed to time out, we'd might as well just 728 * treat him as timed out now. 729 */ 730 sc_tv.tv_sec = sc->sc_advbase * 3; 731 if (timevalcmp(&sc_tv, &ch_tv, <)) { 732 CARP_DEBUG("%s: BACKUP -> MASTER " 733 "(master timed out)\n", 734 SC2IFP(sc)->if_xname); 735 carp_master_down_locked(sc); 736 break; 737 } 738 739 /* 740 * Otherwise, we reset the counter and wait for the next 741 * advertisement. 742 */ 743 carp_setrun(sc, af); 744 break; 745 } 746 747 CARP_UNLOCK(ifp->if_carp); 748 749 m_freem(m); 750 return; 751 } 752 753 static void 754 carpdetach(struct carp_softc *sc) 755 { 756 struct ifaddr *ifa; 757 758 callout_stop(&sc->sc_ad_tmo); 759 callout_stop(&sc->sc_md_tmo); 760 callout_stop(&sc->sc_md6_tmo); 761 762 while ((ifa = TAILQ_FIRST(&SC2IFP(sc)->if_addrlist)) != NULL) 763 if (ifa->ifa_addr->sa_family == AF_INET) { 764 struct in_ifaddr *ia = ifatoia(ifa); 765 766 carp_del_addr(sc, &ia->ia_addr); 767 768 /* ripped screaming from in_control(SIOCDIFADDR) */ 769 in_ifscrub(SC2IFP(sc), ia); 770 TAILQ_REMOVE(&SC2IFP(sc)->if_addrlist, ifa, ifa_link); 771 TAILQ_REMOVE(&in_ifaddrhead, ia, ia_link); 772 IFAFREE((&ia->ia_ifa)); 773 } 774 } 775 776 /* Detach an interface from the carp. */ 777 void 778 carp_ifdetach(struct ifnet *ifp) 779 { 780 struct carp_softc *sc; 781 struct carp_if *cif = (struct carp_if *)ifp->if_carp; 782 783 CARP_LOCK(cif); 784 TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) 785 carpdetach(sc); 786 CARP_UNLOCK(cif); 787 } 788 789 static int 790 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch) 791 { 792 struct m_tag *mtag; 793 struct ifnet *ifp = SC2IFP(sc); 794 795 if (sc->sc_init_counter) { 796 /* this could also be seconds since unix epoch */ 797 sc->sc_counter = arc4random(); 798 sc->sc_counter = sc->sc_counter << 32; 799 sc->sc_counter += arc4random(); 800 } else 801 sc->sc_counter++; 802 803 ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff); 804 ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff); 805 806 carp_hmac_generate(sc, ch->carp_counter, ch->carp_md); 807 808 /* Tag packet for carp_output */ 809 mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct ifnet *), M_NOWAIT); 810 if (mtag == NULL) { 811 m_freem(m); 812 SC2IFP(sc)->if_oerrors++; 813 return (ENOMEM); 814 } 815 bcopy(&ifp, (caddr_t)(mtag + 1), sizeof(struct ifnet *)); 816 m_tag_prepend(m, mtag); 817 818 return (0); 819 } 820 821 static void 822 carp_send_ad_all(void) 823 { 824 struct carp_softc *sc; 825 826 mtx_lock(&carp_mtx); 827 LIST_FOREACH(sc, &carpif_list, sc_next) { 828 if (sc->sc_carpdev == NULL) 829 continue; 830 CARP_SCLOCK(sc); 831 if ((SC2IFP(sc)->if_flags & (IFF_UP|IFF_RUNNING)) && 832 sc->sc_state == MASTER) 833 carp_send_ad_locked(sc); 834 CARP_SCUNLOCK(sc); 835 } 836 mtx_unlock(&carp_mtx); 837 } 838 839 static void 840 carp_send_ad(void *v) 841 { 842 struct carp_softc *sc = v; 843 844 CARP_SCLOCK(sc); 845 carp_send_ad_locked(sc); 846 CARP_SCUNLOCK(sc); 847 } 848 849 static void 850 carp_send_ad_locked(struct carp_softc *sc) 851 { 852 struct carp_header ch; 853 struct timeval tv; 854 struct carp_header *ch_ptr; 855 struct mbuf *m; 856 int len, advbase, advskew; 857 858 CARP_SCLOCK_ASSERT(sc); 859 860 /* bow out if we've lost our UPness or RUNNINGuiness */ 861 if ((SC2IFP(sc)->if_flags & 862 (IFF_UP|IFF_RUNNING)) != (IFF_UP|IFF_RUNNING)) { 863 advbase = 255; 864 advskew = 255; 865 } else { 866 advbase = sc->sc_advbase; 867 if (!carp_suppress_preempt || sc->sc_advskew > 240) 868 advskew = sc->sc_advskew; 869 else 870 advskew = 240; 871 tv.tv_sec = advbase; 872 tv.tv_usec = advskew * 1000000 / 256; 873 } 874 875 ch.carp_version = CARP_VERSION; 876 ch.carp_type = CARP_ADVERTISEMENT; 877 ch.carp_vhid = sc->sc_vhid; 878 ch.carp_advbase = advbase; 879 ch.carp_advskew = advskew; 880 ch.carp_authlen = 7; /* XXX DEFINE */ 881 ch.carp_pad1 = 0; /* must be zero */ 882 ch.carp_cksum = 0; 883 884 #ifdef INET 885 if (sc->sc_ia) { 886 struct ip *ip; 887 888 MGETHDR(m, M_DONTWAIT, MT_HEADER); 889 if (m == NULL) { 890 SC2IFP(sc)->if_oerrors++; 891 carpstats.carps_onomem++; 892 /* XXX maybe less ? */ 893 if (advbase != 255 || advskew != 255) 894 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), 895 carp_send_ad, sc); 896 return; 897 } 898 len = sizeof(*ip) + sizeof(ch); 899 m->m_pkthdr.len = len; 900 m->m_pkthdr.rcvif = NULL; 901 m->m_len = len; 902 MH_ALIGN(m, m->m_len); 903 m->m_flags |= M_MCAST; 904 ip = mtod(m, struct ip *); 905 ip->ip_v = IPVERSION; 906 ip->ip_hl = sizeof(*ip) >> 2; 907 ip->ip_tos = IPTOS_LOWDELAY; 908 ip->ip_len = len; 909 ip->ip_id = ip_newid(); 910 ip->ip_off = IP_DF; 911 ip->ip_ttl = CARP_DFLTTL; 912 ip->ip_p = IPPROTO_CARP; 913 ip->ip_sum = 0; 914 ip->ip_src.s_addr = sc->sc_ia->ia_addr.sin_addr.s_addr; 915 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP); 916 917 ch_ptr = (struct carp_header *)(&ip[1]); 918 bcopy(&ch, ch_ptr, sizeof(ch)); 919 if (carp_prepare_ad(m, sc, ch_ptr)) 920 return; 921 922 m->m_data += sizeof(*ip); 923 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip)); 924 m->m_data -= sizeof(*ip); 925 926 getmicrotime(&SC2IFP(sc)->if_lastchange); 927 SC2IFP(sc)->if_opackets++; 928 SC2IFP(sc)->if_obytes += len; 929 carpstats.carps_opackets++; 930 931 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, NULL)) { 932 SC2IFP(sc)->if_oerrors++; 933 if (sc->sc_sendad_errors < INT_MAX) 934 sc->sc_sendad_errors++; 935 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) { 936 carp_suppress_preempt++; 937 if (carp_suppress_preempt == 1) { 938 CARP_SCUNLOCK(sc); 939 carp_send_ad_all(); 940 CARP_SCLOCK(sc); 941 } 942 } 943 sc->sc_sendad_success = 0; 944 } else { 945 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) { 946 if (++sc->sc_sendad_success >= 947 CARP_SENDAD_MIN_SUCCESS) { 948 carp_suppress_preempt--; 949 sc->sc_sendad_errors = 0; 950 } 951 } else 952 sc->sc_sendad_errors = 0; 953 } 954 } 955 #endif /* INET */ 956 #ifdef INET6 957 if (sc->sc_ia6) { 958 struct ip6_hdr *ip6; 959 960 MGETHDR(m, M_DONTWAIT, MT_HEADER); 961 if (m == NULL) { 962 SC2IFP(sc)->if_oerrors++; 963 carpstats.carps_onomem++; 964 /* XXX maybe less ? */ 965 if (advbase != 255 || advskew != 255) 966 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), 967 carp_send_ad, sc); 968 return; 969 } 970 len = sizeof(*ip6) + sizeof(ch); 971 m->m_pkthdr.len = len; 972 m->m_pkthdr.rcvif = NULL; 973 m->m_len = len; 974 MH_ALIGN(m, m->m_len); 975 m->m_flags |= M_MCAST; 976 ip6 = mtod(m, struct ip6_hdr *); 977 bzero(ip6, sizeof(*ip6)); 978 ip6->ip6_vfc |= IPV6_VERSION; 979 ip6->ip6_hlim = CARP_DFLTTL; 980 ip6->ip6_nxt = IPPROTO_CARP; 981 bcopy(&sc->sc_ia6->ia_addr.sin6_addr, &ip6->ip6_src, 982 sizeof(struct in6_addr)); 983 /* set the multicast destination */ 984 985 ip6->ip6_dst.s6_addr8[0] = 0xff; 986 ip6->ip6_dst.s6_addr8[1] = 0x02; 987 ip6->ip6_dst.s6_addr8[15] = 0x12; 988 989 ch_ptr = (struct carp_header *)(&ip6[1]); 990 bcopy(&ch, ch_ptr, sizeof(ch)); 991 if (carp_prepare_ad(m, sc, ch_ptr)) 992 return; 993 994 m->m_data += sizeof(*ip6); 995 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6)); 996 m->m_data -= sizeof(*ip6); 997 998 getmicrotime(&SC2IFP(sc)->if_lastchange); 999 SC2IFP(sc)->if_opackets++; 1000 SC2IFP(sc)->if_obytes += len; 1001 carpstats.carps_opackets6++; 1002 1003 if (ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL)) { 1004 SC2IFP(sc)->if_oerrors++; 1005 if (sc->sc_sendad_errors < INT_MAX) 1006 sc->sc_sendad_errors++; 1007 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) { 1008 carp_suppress_preempt++; 1009 if (carp_suppress_preempt == 1) { 1010 CARP_SCUNLOCK(sc); 1011 carp_send_ad_all(); 1012 CARP_SCLOCK(sc); 1013 } 1014 } 1015 sc->sc_sendad_success = 0; 1016 } else { 1017 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) { 1018 if (++sc->sc_sendad_success >= 1019 CARP_SENDAD_MIN_SUCCESS) { 1020 carp_suppress_preempt--; 1021 sc->sc_sendad_errors = 0; 1022 } 1023 } else 1024 sc->sc_sendad_errors = 0; 1025 } 1026 } 1027 #endif /* INET6 */ 1028 1029 if (advbase != 255 || advskew != 255) 1030 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), 1031 carp_send_ad, sc); 1032 1033 } 1034 1035 /* 1036 * Broadcast a gratuitous ARP request containing 1037 * the virtual router MAC address for each IP address 1038 * associated with the virtual router. 1039 */ 1040 static void 1041 carp_send_arp(struct carp_softc *sc) 1042 { 1043 struct ifaddr *ifa; 1044 1045 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) { 1046 1047 if (ifa->ifa_addr->sa_family != AF_INET) 1048 continue; 1049 1050 /* arprequest(sc->sc_carpdev, &in, &in, IFP2ENADDR(sc->sc_ifp)); */ 1051 arp_ifinit2(sc->sc_carpdev, ifa, IFP2ENADDR(sc->sc_ifp)); 1052 1053 DELAY(1000); /* XXX */ 1054 } 1055 } 1056 1057 #ifdef INET6 1058 static void 1059 carp_send_na(struct carp_softc *sc) 1060 { 1061 struct ifaddr *ifa; 1062 struct in6_addr *in6; 1063 static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT; 1064 1065 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) { 1066 1067 if (ifa->ifa_addr->sa_family != AF_INET6) 1068 continue; 1069 1070 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr; 1071 nd6_na_output(sc->sc_carpdev, &mcast, in6, 1072 ND_NA_FLAG_OVERRIDE, 1, NULL); 1073 DELAY(1000); /* XXX */ 1074 } 1075 } 1076 #endif /* INET6 */ 1077 1078 static int 1079 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type) 1080 { 1081 struct carp_softc *vh; 1082 struct ifaddr *ifa; 1083 int count = 0; 1084 1085 CARP_LOCK_ASSERT(cif); 1086 1087 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1088 if ((type == CARP_COUNT_RUNNING && 1089 (SC2IFP(vh)->if_flags & (IFF_UP|IFF_RUNNING)) == 1090 (IFF_UP|IFF_RUNNING)) || 1091 (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) { 1092 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist, 1093 ifa_list) { 1094 if (ifa->ifa_addr->sa_family == AF_INET && 1095 ia->ia_addr.sin_addr.s_addr == 1096 ifatoia(ifa)->ia_addr.sin_addr.s_addr) 1097 count++; 1098 } 1099 } 1100 } 1101 return (count); 1102 } 1103 1104 int 1105 carp_iamatch(void *v, struct in_ifaddr *ia, 1106 struct in_addr *isaddr, u_int8_t **enaddr) 1107 { 1108 struct carp_if *cif = v; 1109 struct carp_softc *vh; 1110 int index, count = 0; 1111 struct ifaddr *ifa; 1112 1113 CARP_LOCK(cif); 1114 1115 if (carp_opts[CARPCTL_ARPBALANCE]) { 1116 /* 1117 * XXX proof of concept implementation. 1118 * We use the source ip to decide which virtual host should 1119 * handle the request. If we're master of that virtual host, 1120 * then we respond, otherwise, just drop the arp packet on 1121 * the floor. 1122 */ 1123 count = carp_addrcount(cif, ia, CARP_COUNT_RUNNING); 1124 if (count == 0) { 1125 /* should never reach this */ 1126 CARP_UNLOCK(cif); 1127 return (0); 1128 } 1129 1130 /* this should be a hash, like pf_hash() */ 1131 index = isaddr->s_addr % count; 1132 count = 0; 1133 1134 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1135 if ((SC2IFP(vh)->if_flags & (IFF_UP|IFF_RUNNING)) == 1136 (IFF_UP|IFF_RUNNING)) { 1137 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist, 1138 ifa_list) { 1139 if (ifa->ifa_addr->sa_family == 1140 AF_INET && 1141 ia->ia_addr.sin_addr.s_addr == 1142 ifatoia(ifa)->ia_addr.sin_addr.s_addr) { 1143 if (count == index) { 1144 if (vh->sc_state == 1145 MASTER) { 1146 *enaddr = IFP2ENADDR(vh->sc_ifp); 1147 CARP_UNLOCK(cif); 1148 return (1); 1149 } else { 1150 CARP_UNLOCK(cif); 1151 return (0); 1152 } 1153 } 1154 count++; 1155 } 1156 } 1157 } 1158 } 1159 } else { 1160 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1161 if ((SC2IFP(vh)->if_flags & (IFF_UP|IFF_RUNNING)) == 1162 (IFF_UP|IFF_RUNNING) && ia->ia_ifp == 1163 SC2IFP(vh)) { 1164 *enaddr = IFP2ENADDR(vh->sc_ifp); 1165 CARP_UNLOCK(cif); 1166 return (1); 1167 } 1168 } 1169 } 1170 CARP_UNLOCK(cif); 1171 return (0); 1172 } 1173 1174 #ifdef INET6 1175 struct ifaddr * 1176 carp_iamatch6(void *v, struct in6_addr *taddr) 1177 { 1178 struct carp_if *cif = v; 1179 struct carp_softc *vh; 1180 struct ifaddr *ifa; 1181 1182 CARP_LOCK(cif); 1183 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1184 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist, ifa_list) { 1185 if (IN6_ARE_ADDR_EQUAL(taddr, 1186 &ifatoia6(ifa)->ia_addr.sin6_addr) && 1187 ((SC2IFP(vh)->if_flags & 1188 (IFF_UP|IFF_RUNNING)) == (IFF_UP|IFF_RUNNING))) { 1189 CARP_UNLOCK(cif); 1190 return (ifa); 1191 } 1192 } 1193 } 1194 CARP_UNLOCK(cif); 1195 1196 return (NULL); 1197 } 1198 1199 void * 1200 carp_macmatch6(void *v, struct mbuf *m, const struct in6_addr *taddr) 1201 { 1202 struct m_tag *mtag; 1203 struct carp_if *cif = v; 1204 struct carp_softc *sc; 1205 struct ifaddr *ifa; 1206 1207 CARP_LOCK(cif); 1208 TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) { 1209 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) { 1210 if (IN6_ARE_ADDR_EQUAL(taddr, 1211 &ifatoia6(ifa)->ia_addr.sin6_addr) && 1212 ((SC2IFP(sc)->if_flags & 1213 (IFF_UP|IFF_RUNNING)) == (IFF_UP|IFF_RUNNING))) { 1214 struct ifnet *ifp = SC2IFP(sc); 1215 mtag = m_tag_get(PACKET_TAG_CARP, 1216 sizeof(struct ifnet *), M_NOWAIT); 1217 if (mtag == NULL) { 1218 /* better a bit than nothing */ 1219 CARP_UNLOCK(cif); 1220 return (IFP2ENADDR(sc->sc_ifp)); 1221 } 1222 bcopy(&ifp, (caddr_t)(mtag + 1), 1223 sizeof(struct ifnet *)); 1224 m_tag_prepend(m, mtag); 1225 1226 CARP_UNLOCK(cif); 1227 return (IFP2ENADDR(sc->sc_ifp)); 1228 } 1229 } 1230 } 1231 CARP_UNLOCK(cif); 1232 1233 return (NULL); 1234 } 1235 #endif 1236 1237 struct ifnet * 1238 carp_forus(void *v, void *dhost) 1239 { 1240 struct carp_if *cif = v; 1241 struct carp_softc *vh; 1242 u_int8_t *ena = dhost; 1243 1244 if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1) 1245 return (NULL); 1246 1247 CARP_LOCK(cif); 1248 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) 1249 if ((SC2IFP(vh)->if_flags & (IFF_UP|IFF_RUNNING)) == 1250 (IFF_UP|IFF_RUNNING) && vh->sc_state == MASTER && 1251 !bcmp(dhost, IFP2ENADDR(vh->sc_ifp), ETHER_ADDR_LEN)) { 1252 CARP_UNLOCK(cif); 1253 return (SC2IFP(vh)); 1254 } 1255 1256 CARP_UNLOCK(cif); 1257 return (NULL); 1258 } 1259 1260 static void 1261 carp_master_down(void *v) 1262 { 1263 struct carp_softc *sc = v; 1264 1265 CARP_SCLOCK(sc); 1266 carp_master_down_locked(sc); 1267 CARP_SCUNLOCK(sc); 1268 } 1269 1270 static void 1271 carp_master_down_locked(struct carp_softc *sc) 1272 { 1273 if (sc->sc_carpdev) 1274 CARP_SCLOCK_ASSERT(sc); 1275 1276 switch (sc->sc_state) { 1277 case INIT: 1278 printf("%s: master_down event in INIT state\n", 1279 SC2IFP(sc)->if_xname); 1280 break; 1281 case MASTER: 1282 break; 1283 case BACKUP: 1284 carp_set_state(sc, MASTER); 1285 carp_send_ad_locked(sc); 1286 carp_send_arp(sc); 1287 #ifdef INET6 1288 carp_send_na(sc); 1289 #endif /* INET6 */ 1290 carp_setrun(sc, 0); 1291 carp_setroute(sc, RTM_ADD); 1292 break; 1293 } 1294 } 1295 1296 /* 1297 * When in backup state, af indicates whether to reset the master down timer 1298 * for v4 or v6. If it's set to zero, reset the ones which are already pending. 1299 */ 1300 static void 1301 carp_setrun(struct carp_softc *sc, sa_family_t af) 1302 { 1303 struct timeval tv; 1304 1305 if (sc->sc_carpdev) 1306 CARP_SCLOCK_ASSERT(sc); 1307 1308 if (SC2IFP(sc)->if_flags & IFF_UP && 1309 sc->sc_vhid > 0 && (sc->sc_naddrs || sc->sc_naddrs6)) 1310 SC2IFP(sc)->if_flags |= IFF_RUNNING; 1311 else { 1312 SC2IFP(sc)->if_flags &= ~IFF_RUNNING; 1313 carp_setroute(sc, RTM_DELETE); 1314 return; 1315 } 1316 1317 switch (sc->sc_state) { 1318 case INIT: 1319 if (carp_opts[CARPCTL_PREEMPT] && !carp_suppress_preempt) { 1320 carp_send_ad_locked(sc); 1321 carp_send_arp(sc); 1322 #ifdef INET6 1323 carp_send_na(sc); 1324 #endif /* INET6 */ 1325 CARP_DEBUG("%s: INIT -> MASTER (preempting)\n", 1326 SC2IFP(sc)->if_xname); 1327 carp_set_state(sc, MASTER); 1328 carp_setroute(sc, RTM_ADD); 1329 } else { 1330 CARP_DEBUG("%s: INIT -> BACKUP\n", SC2IFP(sc)->if_xname); 1331 carp_set_state(sc, BACKUP); 1332 carp_setroute(sc, RTM_DELETE); 1333 carp_setrun(sc, 0); 1334 } 1335 break; 1336 case BACKUP: 1337 callout_stop(&sc->sc_ad_tmo); 1338 tv.tv_sec = 3 * sc->sc_advbase; 1339 tv.tv_usec = sc->sc_advskew * 1000000 / 256; 1340 switch (af) { 1341 #ifdef INET 1342 case AF_INET: 1343 callout_reset(&sc->sc_md_tmo, tvtohz(&tv), 1344 carp_master_down, sc); 1345 break; 1346 #endif /* INET */ 1347 #ifdef INET6 1348 case AF_INET6: 1349 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv), 1350 carp_master_down, sc); 1351 break; 1352 #endif /* INET6 */ 1353 default: 1354 if (sc->sc_naddrs) 1355 callout_reset(&sc->sc_md_tmo, tvtohz(&tv), 1356 carp_master_down, sc); 1357 if (sc->sc_naddrs6) 1358 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv), 1359 carp_master_down, sc); 1360 break; 1361 } 1362 break; 1363 case MASTER: 1364 tv.tv_sec = sc->sc_advbase; 1365 tv.tv_usec = sc->sc_advskew * 1000000 / 256; 1366 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), 1367 carp_send_ad, sc); 1368 break; 1369 } 1370 } 1371 1372 static int 1373 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin) 1374 { 1375 struct ifnet *ifp; 1376 struct carp_if *cif; 1377 struct in_ifaddr *ia, *ia_if; 1378 struct ip_moptions *imo = &sc->sc_imo; 1379 struct in_addr addr; 1380 u_long iaddr = htonl(sin->sin_addr.s_addr); 1381 int own, error; 1382 1383 if (sin->sin_addr.s_addr == 0) { 1384 if (!(SC2IFP(sc)->if_flags & IFF_UP)) 1385 carp_set_state(sc, INIT); 1386 if (sc->sc_naddrs) 1387 SC2IFP(sc)->if_flags |= IFF_UP; 1388 carp_setrun(sc, 0); 1389 return (0); 1390 } 1391 1392 /* we have to do it by hands to check we won't match on us */ 1393 ia_if = NULL; own = 0; 1394 TAILQ_FOREACH(ia, &in_ifaddrhead, ia_link) { 1395 /* and, yeah, we need a multicast-capable iface too */ 1396 if (ia->ia_ifp != SC2IFP(sc) && 1397 (ia->ia_ifp->if_flags & IFF_MULTICAST) && 1398 (iaddr & ia->ia_subnetmask) == ia->ia_subnet) { 1399 if (!ia_if) 1400 ia_if = ia; 1401 if (sin->sin_addr.s_addr == 1402 ia->ia_addr.sin_addr.s_addr) 1403 own++; 1404 } 1405 } 1406 1407 if (!ia_if) 1408 return (EADDRNOTAVAIL); 1409 1410 ia = ia_if; 1411 ifp = ia->ia_ifp; 1412 1413 if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 || 1414 (imo->imo_multicast_ifp && imo->imo_multicast_ifp != ifp)) 1415 return (EADDRNOTAVAIL); 1416 1417 if (imo->imo_num_memberships == 0) { 1418 addr.s_addr = htonl(INADDR_CARP_GROUP); 1419 if ((imo->imo_membership[0] = in_addmulti(&addr, ifp)) == NULL) 1420 return (ENOBUFS); 1421 imo->imo_num_memberships++; 1422 imo->imo_multicast_ifp = ifp; 1423 imo->imo_multicast_ttl = CARP_DFLTTL; 1424 imo->imo_multicast_loop = 0; 1425 } 1426 1427 if (!ifp->if_carp) { 1428 1429 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP, 1430 M_WAITOK|M_ZERO); 1431 if (!cif) { 1432 error = ENOBUFS; 1433 goto cleanup; 1434 } 1435 if ((error = ifpromisc(ifp, 1))) { 1436 FREE(cif, M_CARP); 1437 goto cleanup; 1438 } 1439 1440 CARP_LOCK_INIT(cif); 1441 CARP_LOCK(cif); 1442 cif->vhif_ifp = ifp; 1443 TAILQ_INIT(&cif->vhif_vrs); 1444 ifp->if_carp = cif; 1445 1446 } else { 1447 struct carp_softc *vr; 1448 1449 cif = (struct carp_if *)ifp->if_carp; 1450 CARP_LOCK(cif); 1451 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) 1452 if (vr != sc && vr->sc_vhid == sc->sc_vhid) { 1453 CARP_UNLOCK(cif); 1454 error = EINVAL; 1455 goto cleanup; 1456 } 1457 } 1458 sc->sc_ia = ia; 1459 sc->sc_carpdev = ifp; 1460 1461 { /* XXX prevent endless loop if already in queue */ 1462 struct carp_softc *vr, *after = NULL; 1463 int myself = 0; 1464 cif = (struct carp_if *)ifp->if_carp; 1465 1466 /* XXX: cif should not change, right? So we still hold the lock */ 1467 CARP_LOCK_ASSERT(cif); 1468 1469 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) { 1470 if (vr == sc) 1471 myself = 1; 1472 if (vr->sc_vhid < sc->sc_vhid) 1473 after = vr; 1474 } 1475 1476 if (!myself) { 1477 /* We're trying to keep things in order */ 1478 if (after == NULL) { 1479 TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list); 1480 } else { 1481 TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list); 1482 } 1483 cif->vhif_nvrs++; 1484 } 1485 } 1486 1487 sc->sc_naddrs++; 1488 SC2IFP(sc)->if_flags |= IFF_UP; 1489 if (own) 1490 sc->sc_advskew = 0; 1491 carp_sc_state_locked(sc); 1492 carp_setrun(sc, 0); 1493 1494 CARP_UNLOCK(cif); 1495 1496 return (0); 1497 1498 cleanup: 1499 in_delmulti(imo->imo_membership[--imo->imo_num_memberships]); 1500 return (error); 1501 } 1502 1503 static int 1504 carp_del_addr(struct carp_softc *sc, struct sockaddr_in *sin) 1505 { 1506 int error = 0; 1507 1508 if (!--sc->sc_naddrs) { 1509 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp; 1510 struct ip_moptions *imo = &sc->sc_imo; 1511 1512 CARP_LOCK(cif); 1513 callout_stop(&sc->sc_ad_tmo); 1514 SC2IFP(sc)->if_flags &= ~(IFF_UP|IFF_RUNNING); 1515 sc->sc_vhid = -1; 1516 in_delmulti(imo->imo_membership[--imo->imo_num_memberships]); 1517 imo->imo_multicast_ifp = NULL; 1518 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list); 1519 if (!--cif->vhif_nvrs) { 1520 sc->sc_carpdev->if_carp = NULL; 1521 CARP_LOCK_DESTROY(cif); 1522 FREE(cif, M_IFADDR); 1523 } else { 1524 CARP_UNLOCK(cif); 1525 } 1526 } 1527 1528 return (error); 1529 } 1530 1531 #ifdef INET6 1532 static int 1533 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6) 1534 { 1535 struct ifnet *ifp; 1536 struct carp_if *cif; 1537 struct in6_ifaddr *ia, *ia_if; 1538 struct ip6_moptions *im6o = &sc->sc_im6o; 1539 struct in6_multi_mship *imm; 1540 struct sockaddr_in6 addr; 1541 int own, error; 1542 1543 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) { 1544 if (!(SC2IFP(sc)->if_flags & IFF_UP)) 1545 carp_set_state(sc, INIT); 1546 if (sc->sc_naddrs6) 1547 SC2IFP(sc)->if_flags |= IFF_UP; 1548 carp_setrun(sc, 0); 1549 return (0); 1550 } 1551 1552 /* we have to do it by hands to check we won't match on us */ 1553 ia_if = NULL; own = 0; 1554 for (ia = in6_ifaddr; ia; ia = ia->ia_next) { 1555 int i; 1556 1557 for (i = 0; i < 4; i++) { 1558 if ((sin6->sin6_addr.s6_addr32[i] & 1559 ia->ia_prefixmask.sin6_addr.s6_addr32[i]) != 1560 (ia->ia_addr.sin6_addr.s6_addr32[i] & 1561 ia->ia_prefixmask.sin6_addr.s6_addr32[i])) 1562 break; 1563 } 1564 /* and, yeah, we need a multicast-capable iface too */ 1565 if (ia->ia_ifp != SC2IFP(sc) && 1566 (ia->ia_ifp->if_flags & IFF_MULTICAST) && 1567 (i == 4)) { 1568 if (!ia_if) 1569 ia_if = ia; 1570 if (IN6_ARE_ADDR_EQUAL(&sin6->sin6_addr, 1571 &ia->ia_addr.sin6_addr)) 1572 own++; 1573 } 1574 } 1575 1576 if (!ia_if) 1577 return (EADDRNOTAVAIL); 1578 ia = ia_if; 1579 ifp = ia->ia_ifp; 1580 1581 if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 || 1582 (im6o->im6o_multicast_ifp && im6o->im6o_multicast_ifp != ifp)) 1583 return (EADDRNOTAVAIL); 1584 1585 if (!sc->sc_naddrs6) { 1586 im6o->im6o_multicast_ifp = ifp; 1587 1588 /* join CARP multicast address */ 1589 bzero(&addr, sizeof(addr)); 1590 addr.sin6_family = AF_INET6; 1591 addr.sin6_len = sizeof(addr); 1592 addr.sin6_addr.s6_addr16[0] = htons(0xff02); 1593 addr.sin6_addr.s6_addr16[1] = htons(ifp->if_index); 1594 addr.sin6_addr.s6_addr8[15] = 0x12; 1595 if ((imm = in6_joingroup(ifp, &addr.sin6_addr, &error)) == NULL) 1596 goto cleanup; 1597 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain); 1598 1599 /* join solicited multicast address */ 1600 bzero(&addr.sin6_addr, sizeof(addr.sin6_addr)); 1601 addr.sin6_addr.s6_addr16[0] = htons(0xff02); 1602 addr.sin6_addr.s6_addr16[1] = htons(ifp->if_index); 1603 addr.sin6_addr.s6_addr32[1] = 0; 1604 addr.sin6_addr.s6_addr32[2] = htonl(1); 1605 addr.sin6_addr.s6_addr32[3] = sin6->sin6_addr.s6_addr32[3]; 1606 addr.sin6_addr.s6_addr8[12] = 0xff; 1607 if ((imm = in6_joingroup(ifp, &addr.sin6_addr, &error)) == NULL) 1608 goto cleanup; 1609 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain); 1610 } 1611 1612 if (!ifp->if_carp) { 1613 MALLOC(cif, struct carp_if *, sizeof(*cif), M_CARP, 1614 M_WAITOK|M_ZERO); 1615 if (!cif) { 1616 error = ENOBUFS; 1617 goto cleanup; 1618 } 1619 if ((error = ifpromisc(ifp, 1))) { 1620 FREE(cif, M_CARP); 1621 goto cleanup; 1622 } 1623 1624 CARP_LOCK_INIT(cif); 1625 CARP_LOCK(cif); 1626 cif->vhif_ifp = ifp; 1627 TAILQ_INIT(&cif->vhif_vrs); 1628 ifp->if_carp = cif; 1629 1630 } else { 1631 struct carp_softc *vr; 1632 1633 cif = (struct carp_if *)ifp->if_carp; 1634 CARP_LOCK(cif); 1635 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) 1636 if (vr != sc && vr->sc_vhid == sc->sc_vhid) { 1637 CARP_UNLOCK(cif); 1638 error = EINVAL; 1639 goto cleanup; 1640 } 1641 } 1642 sc->sc_ia6 = ia; 1643 sc->sc_carpdev = ifp; 1644 1645 { /* XXX prevent endless loop if already in queue */ 1646 struct carp_softc *vr, *after = NULL; 1647 int myself = 0; 1648 cif = (struct carp_if *)ifp->if_carp; 1649 CARP_LOCK_ASSERT(cif); 1650 1651 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) { 1652 if (vr == sc) 1653 myself = 1; 1654 if (vr->sc_vhid < sc->sc_vhid) 1655 after = vr; 1656 } 1657 1658 if (!myself) { 1659 /* We're trying to keep things in order */ 1660 if (after == NULL) { 1661 TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list); 1662 } else { 1663 TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list); 1664 } 1665 cif->vhif_nvrs++; 1666 } 1667 } 1668 1669 sc->sc_naddrs6++; 1670 SC2IFP(sc)->if_flags |= IFF_UP; 1671 if (own) 1672 sc->sc_advskew = 0; 1673 carp_sc_state_locked(sc); 1674 carp_setrun(sc, 0); 1675 1676 CARP_UNLOCK(cif); 1677 1678 return (0); 1679 1680 cleanup: 1681 /* clean up multicast memberships */ 1682 if (!sc->sc_naddrs6) { 1683 while (!LIST_EMPTY(&im6o->im6o_memberships)) { 1684 imm = LIST_FIRST(&im6o->im6o_memberships); 1685 LIST_REMOVE(imm, i6mm_chain); 1686 in6_leavegroup(imm); 1687 } 1688 } 1689 return (error); 1690 } 1691 1692 static int 1693 carp_del_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6) 1694 { 1695 int error = 0; 1696 1697 if (!--sc->sc_naddrs6) { 1698 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp; 1699 struct ip6_moptions *im6o = &sc->sc_im6o; 1700 1701 CARP_LOCK(cif); 1702 callout_stop(&sc->sc_ad_tmo); 1703 SC2IFP(sc)->if_flags &= ~(IFF_UP|IFF_RUNNING); 1704 sc->sc_vhid = -1; 1705 while (!LIST_EMPTY(&im6o->im6o_memberships)) { 1706 struct in6_multi_mship *imm = 1707 LIST_FIRST(&im6o->im6o_memberships); 1708 1709 LIST_REMOVE(imm, i6mm_chain); 1710 in6_leavegroup(imm); 1711 } 1712 im6o->im6o_multicast_ifp = NULL; 1713 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list); 1714 if (!--cif->vhif_nvrs) { 1715 CARP_LOCK_DESTROY(cif); 1716 sc->sc_carpdev->if_carp = NULL; 1717 FREE(cif, M_IFADDR); 1718 } else 1719 CARP_UNLOCK(cif); 1720 } 1721 1722 return (error); 1723 } 1724 #endif /* INET6 */ 1725 1726 static int 1727 carp_ioctl(struct ifnet *ifp, u_long cmd, caddr_t addr) 1728 { 1729 struct carp_softc *sc = ifp->if_softc, *vr; 1730 struct carpreq carpr; 1731 struct ifaddr *ifa; 1732 struct ifreq *ifr; 1733 struct ifaliasreq *ifra; 1734 int locked = 0, error = 0; 1735 1736 ifa = (struct ifaddr *)addr; 1737 ifra = (struct ifaliasreq *)addr; 1738 ifr = (struct ifreq *)addr; 1739 1740 switch (cmd) { 1741 case SIOCSIFADDR: 1742 switch (ifa->ifa_addr->sa_family) { 1743 #ifdef INET 1744 case AF_INET: 1745 SC2IFP(sc)->if_flags |= IFF_UP; 1746 bcopy(ifa->ifa_addr, ifa->ifa_dstaddr, 1747 sizeof(struct sockaddr)); 1748 error = carp_set_addr(sc, satosin(ifa->ifa_addr)); 1749 break; 1750 #endif /* INET */ 1751 #ifdef INET6 1752 case AF_INET6: 1753 SC2IFP(sc)->if_flags |= IFF_UP; 1754 error = carp_set_addr6(sc, satosin6(ifa->ifa_addr)); 1755 break; 1756 #endif /* INET6 */ 1757 default: 1758 error = EAFNOSUPPORT; 1759 break; 1760 } 1761 break; 1762 1763 case SIOCAIFADDR: 1764 switch (ifa->ifa_addr->sa_family) { 1765 #ifdef INET 1766 case AF_INET: 1767 SC2IFP(sc)->if_flags |= IFF_UP; 1768 bcopy(ifa->ifa_addr, ifa->ifa_dstaddr, 1769 sizeof(struct sockaddr)); 1770 error = carp_set_addr(sc, satosin(&ifra->ifra_addr)); 1771 break; 1772 #endif /* INET */ 1773 #ifdef INET6 1774 case AF_INET6: 1775 SC2IFP(sc)->if_flags |= IFF_UP; 1776 error = carp_set_addr6(sc, satosin6(&ifra->ifra_addr)); 1777 break; 1778 #endif /* INET6 */ 1779 default: 1780 error = EAFNOSUPPORT; 1781 break; 1782 } 1783 break; 1784 1785 case SIOCDIFADDR: 1786 switch (ifa->ifa_addr->sa_family) { 1787 #ifdef INET 1788 case AF_INET: 1789 error = carp_del_addr(sc, satosin(&ifra->ifra_addr)); 1790 break; 1791 #endif /* INET */ 1792 #ifdef INET6 1793 case AF_INET6: 1794 error = carp_del_addr6(sc, satosin6(&ifra->ifra_addr)); 1795 break; 1796 #endif /* INET6 */ 1797 default: 1798 error = EAFNOSUPPORT; 1799 break; 1800 } 1801 break; 1802 1803 case SIOCSIFFLAGS: 1804 if (sc->sc_carpdev) { 1805 locked = 1; 1806 CARP_SCLOCK(sc); 1807 } 1808 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) { 1809 callout_stop(&sc->sc_ad_tmo); 1810 callout_stop(&sc->sc_md_tmo); 1811 callout_stop(&sc->sc_md6_tmo); 1812 if (sc->sc_state == MASTER) 1813 carp_send_ad_locked(sc); 1814 carp_set_state(sc, INIT); 1815 carp_setrun(sc, 0); 1816 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) { 1817 SC2IFP(sc)->if_flags |= IFF_UP; 1818 carp_setrun(sc, 0); 1819 } 1820 break; 1821 1822 case SIOCSVH: 1823 if ((error = suser(curthread)) != 0) 1824 break; 1825 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr))) 1826 break; 1827 error = 1; 1828 if (sc->sc_carpdev) { 1829 locked = 1; 1830 CARP_SCLOCK(sc); 1831 } 1832 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) { 1833 switch (carpr.carpr_state) { 1834 case BACKUP: 1835 callout_stop(&sc->sc_ad_tmo); 1836 carp_set_state(sc, BACKUP); 1837 carp_setrun(sc, 0); 1838 carp_setroute(sc, RTM_DELETE); 1839 break; 1840 case MASTER: 1841 carp_master_down_locked(sc); 1842 break; 1843 default: 1844 break; 1845 } 1846 } 1847 if (carpr.carpr_vhid > 0) { 1848 if (carpr.carpr_vhid > 255) { 1849 error = EINVAL; 1850 break; 1851 } 1852 if (sc->sc_carpdev) { 1853 struct carp_if *cif; 1854 cif = (struct carp_if *)sc->sc_carpdev->if_carp; 1855 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) 1856 if (vr != sc && 1857 vr->sc_vhid == carpr.carpr_vhid) 1858 return EEXIST; 1859 } 1860 sc->sc_vhid = carpr.carpr_vhid; 1861 IFP2ENADDR(sc->sc_ifp)[0] = 0; 1862 IFP2ENADDR(sc->sc_ifp)[1] = 0; 1863 IFP2ENADDR(sc->sc_ifp)[2] = 0x5e; 1864 IFP2ENADDR(sc->sc_ifp)[3] = 0; 1865 IFP2ENADDR(sc->sc_ifp)[4] = 1; 1866 IFP2ENADDR(sc->sc_ifp)[5] = sc->sc_vhid; 1867 error--; 1868 } 1869 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) { 1870 if (carpr.carpr_advskew >= 255) { 1871 error = EINVAL; 1872 break; 1873 } 1874 if (carpr.carpr_advbase > 255) { 1875 error = EINVAL; 1876 break; 1877 } 1878 sc->sc_advbase = carpr.carpr_advbase; 1879 sc->sc_advskew = carpr.carpr_advskew; 1880 error--; 1881 } 1882 bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key)); 1883 if (error > 0) 1884 error = EINVAL; 1885 else { 1886 error = 0; 1887 carp_setrun(sc, 0); 1888 } 1889 break; 1890 1891 case SIOCGVH: 1892 /* XXX: lockless read */ 1893 bzero(&carpr, sizeof(carpr)); 1894 carpr.carpr_state = sc->sc_state; 1895 carpr.carpr_vhid = sc->sc_vhid; 1896 carpr.carpr_advbase = sc->sc_advbase; 1897 carpr.carpr_advskew = sc->sc_advskew; 1898 if (suser(curthread) == 0) 1899 bcopy(sc->sc_key, carpr.carpr_key, 1900 sizeof(carpr.carpr_key)); 1901 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr)); 1902 break; 1903 1904 default: 1905 error = EINVAL; 1906 } 1907 1908 if (locked) 1909 CARP_SCUNLOCK(sc); 1910 1911 carp_hmac_prepare(sc); 1912 1913 return (error); 1914 } 1915 1916 /* 1917 * XXX: this is looutput. We should eventually use it from there. 1918 */ 1919 static int 1920 carp_looutput(struct ifnet *ifp, struct mbuf *m, struct sockaddr *dst, 1921 struct rtentry *rt) 1922 { 1923 M_ASSERTPKTHDR(m); /* check if we have the packet header */ 1924 1925 if (rt && rt->rt_flags & (RTF_REJECT|RTF_BLACKHOLE)) { 1926 m_freem(m); 1927 return (rt->rt_flags & RTF_BLACKHOLE ? 0 : 1928 rt->rt_flags & RTF_HOST ? EHOSTUNREACH : ENETUNREACH); 1929 } 1930 1931 ifp->if_opackets++; 1932 ifp->if_obytes += m->m_pkthdr.len; 1933 #if 1 /* XXX */ 1934 switch (dst->sa_family) { 1935 case AF_INET: 1936 case AF_INET6: 1937 case AF_IPX: 1938 case AF_APPLETALK: 1939 break; 1940 default: 1941 printf("carp_looutput: af=%d unexpected\n", dst->sa_family); 1942 m_freem(m); 1943 return (EAFNOSUPPORT); 1944 } 1945 #endif 1946 return(if_simloop(ifp, m, dst->sa_family, 0)); 1947 } 1948 1949 /* 1950 * Start output on carp interface. This function should never be called. 1951 */ 1952 static void 1953 carp_start(struct ifnet *ifp) 1954 { 1955 #ifdef DEBUG 1956 printf("%s: start called\n", ifp->if_xname); 1957 #endif 1958 } 1959 1960 int 1961 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *sa, 1962 struct rtentry *rt) 1963 { 1964 struct m_tag *mtag; 1965 struct carp_softc *sc; 1966 struct ifnet *carp_ifp; 1967 1968 if (!sa) 1969 return (0); 1970 1971 switch (sa->sa_family) { 1972 #ifdef INET 1973 case AF_INET: 1974 break; 1975 #endif /* INET */ 1976 #ifdef INET6 1977 case AF_INET6: 1978 break; 1979 #endif /* INET6 */ 1980 default: 1981 return (0); 1982 } 1983 1984 mtag = m_tag_find(m, PACKET_TAG_CARP, NULL); 1985 if (mtag == NULL) 1986 return (0); 1987 1988 bcopy(mtag + 1, &carp_ifp, sizeof(struct ifnet *)); 1989 sc = carp_ifp->if_softc; 1990 1991 /* Set the source MAC address to Virtual Router MAC Address */ 1992 switch (ifp->if_type) { 1993 case IFT_ETHER: 1994 case IFT_L2VLAN: { 1995 struct ether_header *eh; 1996 1997 eh = mtod(m, struct ether_header *); 1998 eh->ether_shost[0] = 0; 1999 eh->ether_shost[1] = 0; 2000 eh->ether_shost[2] = 0x5e; 2001 eh->ether_shost[3] = 0; 2002 eh->ether_shost[4] = 1; 2003 eh->ether_shost[5] = sc->sc_vhid; 2004 } 2005 break; 2006 case IFT_FDDI: { 2007 struct fddi_header *fh; 2008 2009 fh = mtod(m, struct fddi_header *); 2010 fh->fddi_shost[0] = 0; 2011 fh->fddi_shost[1] = 0; 2012 fh->fddi_shost[2] = 0x5e; 2013 fh->fddi_shost[3] = 0; 2014 fh->fddi_shost[4] = 1; 2015 fh->fddi_shost[5] = sc->sc_vhid; 2016 } 2017 break; 2018 case IFT_ISO88025: { 2019 struct iso88025_header *th; 2020 th = mtod(m, struct iso88025_header *); 2021 th->iso88025_shost[0] = 3; 2022 th->iso88025_shost[1] = 0; 2023 th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1); 2024 th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1); 2025 th->iso88025_shost[4] = 0; 2026 th->iso88025_shost[5] = 0; 2027 } 2028 break; 2029 default: 2030 printf("%s: carp is not supported for this interface type\n", 2031 ifp->if_xname); 2032 return (EOPNOTSUPP); 2033 } 2034 2035 return (0); 2036 } 2037 2038 static void 2039 carp_set_state(struct carp_softc *sc, int state) 2040 { 2041 2042 if (sc->sc_carpdev) 2043 CARP_SCLOCK_ASSERT(sc); 2044 2045 if (sc->sc_state == state) 2046 return; 2047 2048 sc->sc_state = state; 2049 switch (state) { 2050 case BACKUP: 2051 SC2IFP(sc)->if_link_state = LINK_STATE_DOWN; 2052 break; 2053 case MASTER: 2054 SC2IFP(sc)->if_link_state = LINK_STATE_UP; 2055 break; 2056 default: 2057 SC2IFP(sc)->if_link_state = LINK_STATE_UNKNOWN; 2058 break; 2059 } 2060 rt_ifmsg(SC2IFP(sc)); 2061 } 2062 2063 void 2064 carp_carpdev_state(void *v) 2065 { 2066 struct carp_if *cif = v; 2067 2068 CARP_LOCK(cif); 2069 carp_carpdev_state_locked(cif); 2070 CARP_UNLOCK(cif); 2071 } 2072 2073 static void 2074 carp_carpdev_state_locked(struct carp_if *cif) 2075 { 2076 struct carp_softc *sc; 2077 2078 TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) 2079 carp_sc_state_locked(sc); 2080 } 2081 2082 static void 2083 carp_sc_state_locked(struct carp_softc *sc) 2084 { 2085 CARP_SCLOCK_ASSERT(sc); 2086 2087 if (sc->sc_carpdev->if_link_state != LINK_STATE_UP || 2088 !(sc->sc_carpdev->if_flags & IFF_UP)) { 2089 sc->sc_flags_backup = SC2IFP(sc)->if_flags; 2090 SC2IFP(sc)->if_flags &= ~(IFF_UP|IFF_RUNNING); 2091 callout_stop(&sc->sc_ad_tmo); 2092 callout_stop(&sc->sc_md_tmo); 2093 callout_stop(&sc->sc_md6_tmo); 2094 carp_set_state(sc, INIT); 2095 carp_setrun(sc, 0); 2096 if (!sc->sc_suppress) { 2097 carp_suppress_preempt++; 2098 if (carp_suppress_preempt == 1) { 2099 CARP_SCUNLOCK(sc); 2100 carp_send_ad_all(); 2101 CARP_SCLOCK(sc); 2102 } 2103 } 2104 sc->sc_suppress = 1; 2105 } else { 2106 SC2IFP(sc)->if_flags |= sc->sc_flags_backup; 2107 carp_set_state(sc, INIT); 2108 carp_setrun(sc, 0); 2109 if (sc->sc_suppress) 2110 carp_suppress_preempt--; 2111 sc->sc_suppress = 0; 2112 } 2113 2114 return; 2115 } 2116 2117 static int 2118 carp_modevent(module_t mod, int type, void *data) 2119 { 2120 int error = 0; 2121 2122 switch (type) { 2123 case MOD_LOAD: 2124 mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF); 2125 LIST_INIT(&carpif_list); 2126 if_clone_attach(&carp_cloner); 2127 break; 2128 2129 case MOD_UNLOAD: 2130 if_clone_detach(&carp_cloner); 2131 while (!LIST_EMPTY(&carpif_list)) 2132 carp_clone_destroy(SC2IFP(LIST_FIRST(&carpif_list))); 2133 mtx_destroy(&carp_mtx); 2134 break; 2135 2136 default: 2137 error = EINVAL; 2138 break; 2139 } 2140 2141 return error; 2142 } 2143 2144 static moduledata_t carp_mod = { 2145 "carp", 2146 carp_modevent, 2147 0 2148 }; 2149 2150 DECLARE_MODULE(carp, carp_mod, SI_SUB_PSEUDO, SI_ORDER_ANY); 2151