1 /* $NetBSD: ip_carp.c,v 1.39 2009/09/16 15:23:05 pooka Exp $ */ 2 /* $OpenBSD: ip_carp.c,v 1.113 2005/11/04 08:11:54 mcbride Exp $ */ 3 4 /* 5 * Copyright (c) 2002 Michael Shalayeff. All rights reserved. 6 * Copyright (c) 2003 Ryan McBride. All rights reserved. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 18 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 19 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 20 * IN NO EVENT SHALL THE AUTHOR OR HIS RELATIVES BE LIABLE FOR ANY DIRECT, 21 * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES 22 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR 23 * SERVICES; LOSS OF MIND, USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, 25 * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING 26 * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF 27 * THE POSSIBILITY OF SUCH DAMAGE. 28 */ 29 30 #include <sys/cdefs.h> 31 __KERNEL_RCSID(0, "$NetBSD: ip_carp.c,v 1.39 2009/09/16 15:23:05 pooka Exp $"); 32 33 /* 34 * TODO: 35 * - iface reconfigure 36 * - support for hardware checksum calculations; 37 * 38 */ 39 40 #include <sys/param.h> 41 #include <sys/proc.h> 42 #include <sys/mbuf.h> 43 #include <sys/socket.h> 44 #include <sys/socketvar.h> 45 #include <sys/callout.h> 46 #include <sys/ioctl.h> 47 #include <sys/errno.h> 48 #include <sys/device.h> 49 #include <sys/time.h> 50 #include <sys/kernel.h> 51 #include <sys/kauth.h> 52 #include <sys/sysctl.h> 53 #include <sys/ucred.h> 54 #include <sys/syslog.h> 55 #include <sys/acct.h> 56 57 #include <sys/cpu.h> 58 59 #include <net/if.h> 60 #include <net/pfil.h> 61 #include <net/if_types.h> 62 #include <net/if_ether.h> 63 #include <net/route.h> 64 #include <net/netisr.h> 65 #include <net/net_stats.h> 66 #include <netinet/if_inarp.h> 67 68 #include <machine/stdarg.h> 69 70 #if NFDDI > 0 71 #include <net/if_fddi.h> 72 #endif 73 #if NTOKEN > 0 74 #include <net/if_token.h> 75 #endif 76 77 #ifdef INET 78 #include <netinet/in.h> 79 #include <netinet/in_systm.h> 80 #include <netinet/in_var.h> 81 #include <netinet/ip.h> 82 #include <netinet/ip_var.h> 83 84 #include <net/if_dl.h> 85 #endif 86 87 #ifdef INET6 88 #include <netinet/icmp6.h> 89 #include <netinet/ip6.h> 90 #include <netinet6/ip6_var.h> 91 #include <netinet6/nd6.h> 92 #include <netinet6/scope6_var.h> 93 #endif 94 95 #include "bpfilter.h" 96 #if NBPFILTER > 0 97 #include <net/bpf.h> 98 #endif 99 100 #include <sys/sha1.h> 101 102 #include <netinet/ip_carp.h> 103 104 struct carp_mc_entry { 105 LIST_ENTRY(carp_mc_entry) mc_entries; 106 union { 107 struct ether_multi *mcu_enm; 108 } mc_u; 109 struct sockaddr_storage mc_addr; 110 }; 111 #define mc_enm mc_u.mcu_enm 112 113 struct carp_softc { 114 struct ethercom sc_ac; 115 #define sc_if sc_ac.ec_if 116 #define sc_carpdev sc_ac.ec_if.if_carpdev 117 int ah_cookie; 118 int lh_cookie; 119 struct ip_moptions sc_imo; 120 #ifdef INET6 121 struct ip6_moptions sc_im6o; 122 #endif /* INET6 */ 123 TAILQ_ENTRY(carp_softc) sc_list; 124 125 enum { INIT = 0, BACKUP, MASTER } sc_state; 126 127 int sc_suppress; 128 int sc_bow_out; 129 130 int sc_sendad_errors; 131 #define CARP_SENDAD_MAX_ERRORS 3 132 int sc_sendad_success; 133 #define CARP_SENDAD_MIN_SUCCESS 3 134 135 int sc_vhid; 136 int sc_advskew; 137 int sc_naddrs; 138 int sc_naddrs6; 139 int sc_advbase; /* seconds */ 140 int sc_init_counter; 141 u_int64_t sc_counter; 142 143 /* authentication */ 144 #define CARP_HMAC_PAD 64 145 unsigned char sc_key[CARP_KEY_LEN]; 146 unsigned char sc_pad[CARP_HMAC_PAD]; 147 SHA1_CTX sc_sha1; 148 u_int32_t sc_hashkey[2]; 149 150 struct callout sc_ad_tmo; /* advertisement timeout */ 151 struct callout sc_md_tmo; /* master down timeout */ 152 struct callout sc_md6_tmo; /* master down timeout */ 153 154 LIST_HEAD(__carp_mchead, carp_mc_entry) carp_mc_listhead; 155 }; 156 157 int carp_suppress_preempt = 0; 158 int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 0, 0 }; /* XXX for now */ 159 160 static percpu_t *carpstat_percpu; 161 162 #define CARP_STATINC(x) _NET_STATINC(carpstat_percpu, x) 163 164 struct carp_if { 165 TAILQ_HEAD(, carp_softc) vhif_vrs; 166 int vhif_nvrs; 167 168 struct ifnet *vhif_ifp; 169 }; 170 171 #define CARP_LOG(sc, s) \ 172 if (carp_opts[CARPCTL_LOG]) { \ 173 if (sc) \ 174 log(LOG_INFO, "%s: ", \ 175 (sc)->sc_if.if_xname); \ 176 else \ 177 log(LOG_INFO, "carp: "); \ 178 addlog s; \ 179 addlog("\n"); \ 180 } 181 182 void carp_hmac_prepare(struct carp_softc *); 183 void carp_hmac_generate(struct carp_softc *, u_int32_t *, 184 unsigned char *); 185 int carp_hmac_verify(struct carp_softc *, u_int32_t *, 186 unsigned char *); 187 void carp_setroute(struct carp_softc *, int); 188 void carp_proto_input_c(struct mbuf *, struct carp_header *, sa_family_t); 189 void carpattach(int); 190 void carpdetach(struct carp_softc *); 191 int carp_prepare_ad(struct mbuf *, struct carp_softc *, 192 struct carp_header *); 193 void carp_send_ad_all(void); 194 void carp_send_ad(void *); 195 void carp_send_arp(struct carp_softc *); 196 void carp_master_down(void *); 197 int carp_ioctl(struct ifnet *, u_long, void *); 198 void carp_start(struct ifnet *); 199 void carp_setrun(struct carp_softc *, sa_family_t); 200 void carp_set_state(struct carp_softc *, int); 201 int carp_addrcount(struct carp_if *, struct in_ifaddr *, int); 202 enum { CARP_COUNT_MASTER, CARP_COUNT_RUNNING }; 203 204 void carp_multicast_cleanup(struct carp_softc *); 205 int carp_set_ifp(struct carp_softc *, struct ifnet *); 206 void carp_set_enaddr(struct carp_softc *); 207 void carp_addr_updated(void *); 208 u_int32_t carp_hash(struct carp_softc *, u_char *); 209 int carp_set_addr(struct carp_softc *, struct sockaddr_in *); 210 int carp_join_multicast(struct carp_softc *); 211 #ifdef INET6 212 void carp_send_na(struct carp_softc *); 213 int carp_set_addr6(struct carp_softc *, struct sockaddr_in6 *); 214 int carp_join_multicast6(struct carp_softc *); 215 #endif 216 int carp_clone_create(struct if_clone *, int); 217 int carp_clone_destroy(struct ifnet *); 218 int carp_ether_addmulti(struct carp_softc *, struct ifreq *); 219 int carp_ether_delmulti(struct carp_softc *, struct ifreq *); 220 void carp_ether_purgemulti(struct carp_softc *); 221 222 static void sysctl_net_inet_carp_setup(struct sysctllog **); 223 224 struct if_clone carp_cloner = 225 IF_CLONE_INITIALIZER("carp", carp_clone_create, carp_clone_destroy); 226 227 static __inline u_int16_t 228 carp_cksum(struct mbuf *m, int len) 229 { 230 return (in_cksum(m, len)); 231 } 232 233 void 234 carp_hmac_prepare(struct carp_softc *sc) 235 { 236 u_int8_t carp_version = CARP_VERSION, type = CARP_ADVERTISEMENT; 237 u_int8_t vhid = sc->sc_vhid & 0xff; 238 SHA1_CTX sha1ctx; 239 u_int32_t kmd[5]; 240 struct ifaddr *ifa; 241 int i, found; 242 struct in_addr last, cur, in; 243 #ifdef INET6 244 struct in6_addr last6, cur6, in6; 245 #endif /* INET6 */ 246 247 /* compute ipad from key */ 248 memset(sc->sc_pad, 0, sizeof(sc->sc_pad)); 249 memcpy(sc->sc_pad, sc->sc_key, 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 *)&carp_version, sizeof(carp_version)); 257 SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type)); 258 259 /* generate a key for the arpbalance hash, before the vhid is hashed */ 260 memcpy(&sha1ctx, &sc->sc_sha1, sizeof(sha1ctx)); 261 SHA1Final((unsigned char *)kmd, &sha1ctx); 262 sc->sc_hashkey[0] = kmd[0] ^ kmd[1]; 263 sc->sc_hashkey[1] = kmd[2] ^ kmd[3]; 264 265 /* the rest of the precomputation */ 266 SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid)); 267 268 /* Hash the addresses from smallest to largest, not interface order */ 269 #ifdef INET 270 cur.s_addr = 0; 271 do { 272 found = 0; 273 last = cur; 274 cur.s_addr = 0xffffffff; 275 IFADDR_FOREACH(ifa, &sc->sc_if) { 276 in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr; 277 if (ifa->ifa_addr->sa_family == AF_INET && 278 ntohl(in.s_addr) > ntohl(last.s_addr) && 279 ntohl(in.s_addr) < ntohl(cur.s_addr)) { 280 cur.s_addr = in.s_addr; 281 found++; 282 } 283 } 284 if (found) 285 SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur)); 286 } while (found); 287 #endif /* INET */ 288 289 #ifdef INET6 290 memset(&cur6, 0x00, sizeof(cur6)); 291 do { 292 found = 0; 293 last6 = cur6; 294 memset(&cur6, 0xff, sizeof(cur6)); 295 IFADDR_FOREACH(ifa, &sc->sc_if) { 296 in6 = ifatoia6(ifa)->ia_addr.sin6_addr; 297 if (IN6_IS_ADDR_LINKLOCAL(&in6)) 298 in6.s6_addr16[1] = 0; 299 if (ifa->ifa_addr->sa_family == AF_INET6 && 300 memcmp(&in6, &last6, sizeof(in6)) > 0 && 301 memcmp(&in6, &cur6, sizeof(in6)) < 0) { 302 cur6 = in6; 303 found++; 304 } 305 } 306 if (found) 307 SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6)); 308 } while (found); 309 #endif /* INET6 */ 310 311 /* convert ipad to opad */ 312 for (i = 0; i < sizeof(sc->sc_pad); i++) 313 sc->sc_pad[i] ^= 0x36 ^ 0x5c; 314 } 315 316 void 317 carp_hmac_generate(struct carp_softc *sc, u_int32_t counter[2], 318 unsigned char md[20]) 319 { 320 SHA1_CTX sha1ctx; 321 322 /* fetch first half of inner hash */ 323 memcpy(&sha1ctx, &sc->sc_sha1, sizeof(sha1ctx)); 324 325 SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter)); 326 SHA1Final(md, &sha1ctx); 327 328 /* outer hash */ 329 SHA1Init(&sha1ctx); 330 SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad)); 331 SHA1Update(&sha1ctx, md, 20); 332 SHA1Final(md, &sha1ctx); 333 } 334 335 int 336 carp_hmac_verify(struct carp_softc *sc, u_int32_t counter[2], 337 unsigned char md[20]) 338 { 339 unsigned char md2[20]; 340 341 carp_hmac_generate(sc, counter, md2); 342 343 return (memcmp(md, md2, sizeof(md2))); 344 } 345 346 void 347 carp_setroute(struct carp_softc *sc, int cmd) 348 { 349 struct ifaddr *ifa; 350 int s; 351 352 s = splsoftnet(); 353 IFADDR_FOREACH(ifa, &sc->sc_if) { 354 switch (ifa->ifa_addr->sa_family) { 355 case AF_INET: { 356 int count = 0; 357 struct rtentry *rt; 358 int hr_otherif, nr_ourif; 359 360 /* 361 * Avoid screwing with the routes if there are other 362 * carp interfaces which are master and have the same 363 * address. 364 */ 365 if (sc->sc_carpdev != NULL && 366 sc->sc_carpdev->if_carp != NULL) { 367 count = carp_addrcount( 368 (struct carp_if *)sc->sc_carpdev->if_carp, 369 ifatoia(ifa), CARP_COUNT_MASTER); 370 if ((cmd == RTM_ADD && count != 1) || 371 (cmd == RTM_DELETE && count != 0)) 372 continue; 373 } 374 375 /* Remove the existing host route, if any */ 376 rtrequest(RTM_DELETE, ifa->ifa_addr, 377 ifa->ifa_addr, ifa->ifa_netmask, 378 RTF_HOST, NULL); 379 380 rt = NULL; 381 (void)rtrequest(RTM_GET, ifa->ifa_addr, ifa->ifa_addr, 382 ifa->ifa_netmask, RTF_HOST, &rt); 383 hr_otherif = (rt && rt->rt_ifp != &sc->sc_if && 384 rt->rt_flags & (RTF_CLONING|RTF_CLONED)); 385 if (rt != NULL) { 386 RTFREE(rt); 387 rt = NULL; 388 } 389 390 /* Check for a network route on our interface */ 391 392 rt = NULL; 393 (void)rtrequest(RTM_GET, ifa->ifa_addr, ifa->ifa_addr, 394 ifa->ifa_netmask, 0, &rt); 395 nr_ourif = (rt && rt->rt_ifp == &sc->sc_if); 396 397 switch (cmd) { 398 case RTM_ADD: 399 if (hr_otherif) { 400 ifa->ifa_rtrequest = NULL; 401 ifa->ifa_flags &= ~RTF_CLONING; 402 403 rtrequest(RTM_ADD, ifa->ifa_addr, 404 ifa->ifa_addr, ifa->ifa_netmask, 405 RTF_UP | RTF_HOST, NULL); 406 } 407 if (!hr_otherif || nr_ourif || !rt) { 408 if (nr_ourif && !(rt->rt_flags & 409 RTF_CLONING)) 410 rtrequest(RTM_DELETE, 411 ifa->ifa_addr, 412 ifa->ifa_addr, 413 ifa->ifa_netmask, 0, NULL); 414 415 ifa->ifa_rtrequest = arp_rtrequest; 416 ifa->ifa_flags |= RTF_CLONING; 417 418 if (rtrequest(RTM_ADD, ifa->ifa_addr, 419 ifa->ifa_addr, ifa->ifa_netmask, 0, 420 NULL) == 0) 421 ifa->ifa_flags |= IFA_ROUTE; 422 } 423 break; 424 case RTM_DELETE: 425 break; 426 default: 427 break; 428 } 429 if (rt != NULL) { 430 RTFREE(rt); 431 rt = NULL; 432 } 433 break; 434 } 435 436 #ifdef INET6 437 case AF_INET6: 438 if (cmd == RTM_ADD) 439 in6_ifaddloop(ifa); 440 else 441 in6_ifremloop(ifa); 442 break; 443 #endif /* INET6 */ 444 default: 445 break; 446 } 447 } 448 splx(s); 449 } 450 451 /* 452 * process input packet. 453 * we have rearranged checks order compared to the rfc, 454 * but it seems more efficient this way or not possible otherwise. 455 */ 456 void 457 carp_proto_input(struct mbuf *m, ...) 458 { 459 struct ip *ip = mtod(m, struct ip *); 460 struct carp_softc *sc = NULL; 461 struct carp_header *ch; 462 int iplen, len, hlen; 463 va_list ap; 464 465 va_start(ap, m); 466 hlen = va_arg(ap, int); 467 va_end(ap); 468 469 CARP_STATINC(CARP_STAT_IPACKETS); 470 471 if (!carp_opts[CARPCTL_ALLOW]) { 472 m_freem(m); 473 return; 474 } 475 476 /* check if received on a valid carp interface */ 477 if (m->m_pkthdr.rcvif->if_type != IFT_CARP) { 478 CARP_STATINC(CARP_STAT_BADIF); 479 CARP_LOG(sc, ("packet received on non-carp interface: %s", 480 m->m_pkthdr.rcvif->if_xname)); 481 m_freem(m); 482 return; 483 } 484 485 /* verify that the IP TTL is 255. */ 486 if (ip->ip_ttl != CARP_DFLTTL) { 487 CARP_STATINC(CARP_STAT_BADTTL); 488 CARP_LOG(sc, ("received ttl %d != %d on %s", ip->ip_ttl, 489 CARP_DFLTTL, m->m_pkthdr.rcvif->if_xname)); 490 m_freem(m); 491 return; 492 } 493 494 /* 495 * verify that the received packet length is 496 * equal to the CARP header 497 */ 498 iplen = ip->ip_hl << 2; 499 len = iplen + sizeof(*ch); 500 if (len > m->m_pkthdr.len) { 501 CARP_STATINC(CARP_STAT_BADLEN); 502 CARP_LOG(sc, ("packet too short %d on %s", m->m_pkthdr.len, 503 m->m_pkthdr.rcvif->if_xname)); 504 m_freem(m); 505 return; 506 } 507 508 if ((m = m_pullup(m, len)) == NULL) { 509 CARP_STATINC(CARP_STAT_HDROPS); 510 return; 511 } 512 ip = mtod(m, struct ip *); 513 ch = (struct carp_header *)((char *)ip + iplen); 514 /* verify the CARP checksum */ 515 m->m_data += iplen; 516 if (carp_cksum(m, len - iplen)) { 517 CARP_STATINC(CARP_STAT_BADSUM); 518 CARP_LOG(sc, ("checksum failed on %s", 519 m->m_pkthdr.rcvif->if_xname)); 520 m_freem(m); 521 return; 522 } 523 m->m_data -= iplen; 524 525 carp_proto_input_c(m, ch, AF_INET); 526 } 527 528 #ifdef INET6 529 int 530 carp6_proto_input(struct mbuf **mp, int *offp, int proto) 531 { 532 struct mbuf *m = *mp; 533 struct carp_softc *sc = NULL; 534 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 535 struct carp_header *ch; 536 u_int len; 537 538 CARP_STATINC(CARP_STAT_IPACKETS6); 539 540 if (!carp_opts[CARPCTL_ALLOW]) { 541 m_freem(m); 542 return (IPPROTO_DONE); 543 } 544 545 /* check if received on a valid carp interface */ 546 if (m->m_pkthdr.rcvif->if_type != IFT_CARP) { 547 CARP_STATINC(CARP_STAT_BADIF); 548 CARP_LOG(sc, ("packet received on non-carp interface: %s", 549 m->m_pkthdr.rcvif->if_xname)); 550 m_freem(m); 551 return (IPPROTO_DONE); 552 } 553 554 /* verify that the IP TTL is 255 */ 555 if (ip6->ip6_hlim != CARP_DFLTTL) { 556 CARP_STATINC(CARP_STAT_BADTTL); 557 CARP_LOG(sc, ("received ttl %d != %d on %s", ip6->ip6_hlim, 558 CARP_DFLTTL, m->m_pkthdr.rcvif->if_xname)); 559 m_freem(m); 560 return (IPPROTO_DONE); 561 } 562 563 /* verify that we have a complete carp packet */ 564 len = m->m_len; 565 IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch)); 566 if (ch == NULL) { 567 CARP_STATINC(CARP_STAT_BADLEN); 568 CARP_LOG(sc, ("packet size %u too small", len)); 569 return (IPPROTO_DONE); 570 } 571 572 573 /* verify the CARP checksum */ 574 m->m_data += *offp; 575 if (carp_cksum(m, sizeof(*ch))) { 576 CARP_STATINC(CARP_STAT_BADSUM); 577 CARP_LOG(sc, ("checksum failed, on %s", 578 m->m_pkthdr.rcvif->if_xname)); 579 m_freem(m); 580 return (IPPROTO_DONE); 581 } 582 m->m_data -= *offp; 583 584 carp_proto_input_c(m, ch, AF_INET6); 585 return (IPPROTO_DONE); 586 } 587 #endif /* INET6 */ 588 589 void 590 carp_proto_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af) 591 { 592 struct carp_softc *sc; 593 u_int64_t tmp_counter; 594 struct timeval sc_tv, ch_tv; 595 596 TAILQ_FOREACH(sc, &((struct carp_if *) 597 m->m_pkthdr.rcvif->if_carpdev->if_carp)->vhif_vrs, sc_list) 598 if (sc->sc_vhid == ch->carp_vhid) 599 break; 600 601 if (!sc || (sc->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) != 602 (IFF_UP|IFF_RUNNING)) { 603 CARP_STATINC(CARP_STAT_BADVHID); 604 m_freem(m); 605 return; 606 } 607 608 /* 609 * Check if our own advertisement was duplicated 610 * from a non simplex interface. 611 * XXX If there is no address on our physical interface 612 * there is no way to distinguish our ads from the ones 613 * another carp host might have sent us. 614 */ 615 if ((sc->sc_carpdev->if_flags & IFF_SIMPLEX) == 0) { 616 struct sockaddr sa; 617 struct ifaddr *ifa; 618 619 memset(&sa, 0, sizeof(sa)); 620 sa.sa_family = af; 621 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev); 622 623 if (ifa && af == AF_INET) { 624 struct ip *ip = mtod(m, struct ip *); 625 if (ip->ip_src.s_addr == 626 ifatoia(ifa)->ia_addr.sin_addr.s_addr) { 627 m_freem(m); 628 return; 629 } 630 } 631 #ifdef INET6 632 if (ifa && af == AF_INET6) { 633 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 634 struct in6_addr in6_src, in6_found; 635 636 in6_src = ip6->ip6_src; 637 in6_found = ifatoia6(ifa)->ia_addr.sin6_addr; 638 if (IN6_IS_ADDR_LINKLOCAL(&in6_src)) 639 in6_src.s6_addr16[1] = 0; 640 if (IN6_IS_ADDR_LINKLOCAL(&in6_found)) 641 in6_found.s6_addr16[1] = 0; 642 if (IN6_ARE_ADDR_EQUAL(&in6_src, &in6_found)) { 643 m_freem(m); 644 return; 645 } 646 } 647 #endif /* INET6 */ 648 } 649 650 nanotime(&sc->sc_if.if_lastchange); 651 sc->sc_if.if_ipackets++; 652 sc->sc_if.if_ibytes += m->m_pkthdr.len; 653 654 /* verify the CARP version. */ 655 if (ch->carp_version != CARP_VERSION) { 656 CARP_STATINC(CARP_STAT_BADVER); 657 sc->sc_if.if_ierrors++; 658 CARP_LOG(sc, ("invalid version %d != %d", 659 ch->carp_version, CARP_VERSION)); 660 m_freem(m); 661 return; 662 } 663 664 /* verify the hash */ 665 if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) { 666 CARP_STATINC(CARP_STAT_BADAUTH); 667 sc->sc_if.if_ierrors++; 668 CARP_LOG(sc, ("incorrect hash")); 669 m_freem(m); 670 return; 671 } 672 673 tmp_counter = ntohl(ch->carp_counter[0]); 674 tmp_counter = tmp_counter<<32; 675 tmp_counter += ntohl(ch->carp_counter[1]); 676 677 /* XXX Replay protection goes here */ 678 679 sc->sc_init_counter = 0; 680 sc->sc_counter = tmp_counter; 681 682 683 sc_tv.tv_sec = sc->sc_advbase; 684 if (carp_suppress_preempt && sc->sc_advskew < 240) 685 sc_tv.tv_usec = 240 * 1000000 / 256; 686 else 687 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256; 688 ch_tv.tv_sec = ch->carp_advbase; 689 ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256; 690 691 switch (sc->sc_state) { 692 case INIT: 693 break; 694 case MASTER: 695 /* 696 * If we receive an advertisement from a backup who's going to 697 * be more frequent than us, go into BACKUP state. 698 */ 699 if (timercmp(&sc_tv, &ch_tv, >) || 700 timercmp(&sc_tv, &ch_tv, ==)) { 701 callout_stop(&sc->sc_ad_tmo); 702 CARP_LOG(sc, ("MASTER -> BACKUP (more frequent advertisement received)")); 703 carp_set_state(sc, BACKUP); 704 carp_setrun(sc, 0); 705 carp_setroute(sc, RTM_DELETE); 706 } 707 break; 708 case BACKUP: 709 /* 710 * If we're pre-empting masters who advertise slower than us, 711 * and this one claims to be slower, treat him as down. 712 */ 713 if (carp_opts[CARPCTL_PREEMPT] && timercmp(&sc_tv, &ch_tv, <)) { 714 CARP_LOG(sc, ("BACKUP -> MASTER (preempting a slower master)")); 715 carp_master_down(sc); 716 break; 717 } 718 719 /* 720 * If the master is going to advertise at such a low frequency 721 * that he's guaranteed to time out, we'd might as well just 722 * treat him as timed out now. 723 */ 724 sc_tv.tv_sec = sc->sc_advbase * 3; 725 if (timercmp(&sc_tv, &ch_tv, <)) { 726 CARP_LOG(sc, ("BACKUP -> MASTER (master timed out)")); 727 carp_master_down(sc); 728 break; 729 } 730 731 /* 732 * Otherwise, we reset the counter and wait for the next 733 * advertisement. 734 */ 735 carp_setrun(sc, af); 736 break; 737 } 738 739 m_freem(m); 740 return; 741 } 742 743 /* 744 * Interface side of the CARP implementation. 745 */ 746 747 /* ARGSUSED */ 748 void 749 carpattach(int n) 750 { 751 if_clone_attach(&carp_cloner); 752 753 carpstat_percpu = percpu_alloc(sizeof(uint64_t) * CARP_NSTATS); 754 } 755 756 int 757 carp_clone_create(struct if_clone *ifc, int unit) 758 { 759 extern int ifqmaxlen; 760 struct carp_softc *sc; 761 struct ifnet *ifp; 762 763 sc = malloc(sizeof(*sc), M_DEVBUF, M_NOWAIT|M_ZERO); 764 if (!sc) 765 return (ENOMEM); 766 767 sc->sc_suppress = 0; 768 sc->sc_advbase = CARP_DFLTINTV; 769 sc->sc_vhid = -1; /* required setting */ 770 sc->sc_advskew = 0; 771 sc->sc_init_counter = 1; 772 sc->sc_naddrs = sc->sc_naddrs6 = 0; 773 #ifdef INET6 774 sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL; 775 #endif /* INET6 */ 776 777 callout_init(&sc->sc_ad_tmo, 0); 778 callout_init(&sc->sc_md_tmo, 0); 779 callout_init(&sc->sc_md6_tmo, 0); 780 781 callout_setfunc(&sc->sc_ad_tmo, carp_send_ad, sc); 782 callout_setfunc(&sc->sc_md_tmo, carp_master_down, sc); 783 callout_setfunc(&sc->sc_md6_tmo, carp_master_down, sc); 784 785 LIST_INIT(&sc->carp_mc_listhead); 786 ifp = &sc->sc_if; 787 ifp->if_softc = sc; 788 snprintf(ifp->if_xname, sizeof ifp->if_xname, "%s%d", ifc->ifc_name, 789 unit); 790 ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST; 791 ifp->if_ioctl = carp_ioctl; 792 ifp->if_start = carp_start; 793 ifp->if_output = carp_output; 794 ifp->if_type = IFT_CARP; 795 ifp->if_addrlen = ETHER_ADDR_LEN; 796 ifp->if_hdrlen = ETHER_HDR_LEN; 797 ifp->if_mtu = ETHERMTU; 798 IFQ_SET_MAXLEN(&ifp->if_snd, ifqmaxlen); 799 IFQ_SET_READY(&ifp->if_snd); 800 if_attach(ifp); 801 802 if_alloc_sadl(ifp); 803 ifp->if_broadcastaddr = etherbroadcastaddr; 804 carp_set_enaddr(sc); 805 LIST_INIT(&sc->sc_ac.ec_multiaddrs); 806 #if NBPFILTER > 0 807 bpfattach(ifp, DLT_EN10MB, ETHER_HDR_LEN); 808 #endif 809 return (0); 810 } 811 812 int 813 carp_clone_destroy(struct ifnet *ifp) 814 { 815 struct carp_softc *sc = ifp->if_softc; 816 817 carpdetach(ifp->if_softc); 818 ether_ifdetach(ifp); 819 if_detach(ifp); 820 callout_destroy(&sc->sc_ad_tmo); 821 callout_destroy(&sc->sc_md_tmo); 822 callout_destroy(&sc->sc_md6_tmo); 823 free(ifp->if_softc, M_DEVBUF); 824 825 return (0); 826 } 827 828 void 829 carpdetach(struct carp_softc *sc) 830 { 831 struct carp_if *cif; 832 int s; 833 834 callout_stop(&sc->sc_ad_tmo); 835 callout_stop(&sc->sc_md_tmo); 836 callout_stop(&sc->sc_md6_tmo); 837 838 if (sc->sc_suppress) 839 carp_suppress_preempt--; 840 sc->sc_suppress = 0; 841 842 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) 843 carp_suppress_preempt--; 844 sc->sc_sendad_errors = 0; 845 846 carp_set_state(sc, INIT); 847 sc->sc_if.if_flags &= ~IFF_UP; 848 carp_setrun(sc, 0); 849 carp_multicast_cleanup(sc); 850 851 s = splnet(); 852 if (sc->sc_carpdev != NULL) { 853 /* XXX linkstatehook removal */ 854 cif = (struct carp_if *)sc->sc_carpdev->if_carp; 855 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list); 856 if (!--cif->vhif_nvrs) { 857 ifpromisc(sc->sc_carpdev, 0); 858 sc->sc_carpdev->if_carp = NULL; 859 free(cif, M_IFADDR); 860 } 861 } 862 sc->sc_carpdev = NULL; 863 splx(s); 864 } 865 866 /* Detach an interface from the carp. */ 867 void 868 carp_ifdetach(struct ifnet *ifp) 869 { 870 struct carp_softc *sc, *nextsc; 871 struct carp_if *cif = (struct carp_if *)ifp->if_carp; 872 873 for (sc = TAILQ_FIRST(&cif->vhif_vrs); sc; sc = nextsc) { 874 nextsc = TAILQ_NEXT(sc, sc_list); 875 carpdetach(sc); 876 } 877 } 878 879 int 880 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, 881 struct carp_header *ch) 882 { 883 if (sc->sc_init_counter) { 884 /* this could also be seconds since unix epoch */ 885 sc->sc_counter = arc4random(); 886 sc->sc_counter = sc->sc_counter << 32; 887 sc->sc_counter += arc4random(); 888 } else 889 sc->sc_counter++; 890 891 ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff); 892 ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff); 893 894 carp_hmac_generate(sc, ch->carp_counter, ch->carp_md); 895 896 return (0); 897 } 898 899 void 900 carp_send_ad_all(void) 901 { 902 struct ifnet *ifp; 903 struct carp_if *cif; 904 struct carp_softc *vh; 905 906 TAILQ_FOREACH(ifp, &ifnet, if_list) { 907 if (ifp->if_carp == NULL || ifp->if_type == IFT_CARP) 908 continue; 909 910 cif = (struct carp_if *)ifp->if_carp; 911 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 912 if ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 913 (IFF_UP|IFF_RUNNING) && vh->sc_state == MASTER) 914 carp_send_ad(vh); 915 } 916 } 917 } 918 919 920 void 921 carp_send_ad(void *v) 922 { 923 struct carp_header ch; 924 struct timeval tv; 925 struct carp_softc *sc = v; 926 struct carp_header *ch_ptr; 927 struct mbuf *m; 928 int error, len, advbase, advskew, s; 929 struct ifaddr *ifa; 930 struct sockaddr sa; 931 932 s = splsoftnet(); 933 934 advbase = advskew = 0; /* Sssssh compiler */ 935 if (sc->sc_carpdev == NULL) { 936 sc->sc_if.if_oerrors++; 937 goto retry_later; 938 } 939 940 /* bow out if we've gone to backup (the carp interface is going down) */ 941 if (sc->sc_bow_out) { 942 sc->sc_bow_out = 0; 943 advbase = 255; 944 advskew = 255; 945 } else { 946 advbase = sc->sc_advbase; 947 if (!carp_suppress_preempt || sc->sc_advskew > 240) 948 advskew = sc->sc_advskew; 949 else 950 advskew = 240; 951 tv.tv_sec = advbase; 952 tv.tv_usec = advskew * 1000000 / 256; 953 } 954 955 ch.carp_version = CARP_VERSION; 956 ch.carp_type = CARP_ADVERTISEMENT; 957 ch.carp_vhid = sc->sc_vhid; 958 ch.carp_advbase = advbase; 959 ch.carp_advskew = advskew; 960 ch.carp_authlen = 7; /* XXX DEFINE */ 961 ch.carp_pad1 = 0; /* must be zero */ 962 ch.carp_cksum = 0; 963 964 965 #ifdef INET 966 if (sc->sc_naddrs) { 967 struct ip *ip; 968 969 MGETHDR(m, M_DONTWAIT, MT_HEADER); 970 if (m == NULL) { 971 sc->sc_if.if_oerrors++; 972 CARP_STATINC(CARP_STAT_ONOMEM); 973 /* XXX maybe less ? */ 974 goto retry_later; 975 } 976 len = sizeof(*ip) + sizeof(ch); 977 m->m_pkthdr.len = len; 978 m->m_pkthdr.rcvif = NULL; 979 m->m_len = len; 980 MH_ALIGN(m, m->m_len); 981 m->m_flags |= M_MCAST; 982 ip = mtod(m, struct ip *); 983 ip->ip_v = IPVERSION; 984 ip->ip_hl = sizeof(*ip) >> 2; 985 ip->ip_tos = IPTOS_LOWDELAY; 986 ip->ip_len = htons(len); 987 ip->ip_id = 0; /* no need for id, we don't support fragments */ 988 ip->ip_off = htons(IP_DF); 989 ip->ip_ttl = CARP_DFLTTL; 990 ip->ip_p = IPPROTO_CARP; 991 ip->ip_sum = 0; 992 993 memset(&sa, 0, sizeof(sa)); 994 sa.sa_family = AF_INET; 995 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev); 996 if (ifa == NULL) 997 ip->ip_src.s_addr = 0; 998 else 999 ip->ip_src.s_addr = 1000 ifatoia(ifa)->ia_addr.sin_addr.s_addr; 1001 ip->ip_dst.s_addr = INADDR_CARP_GROUP; 1002 1003 ch_ptr = (struct carp_header *)(&ip[1]); 1004 memcpy(ch_ptr, &ch, sizeof(ch)); 1005 if (carp_prepare_ad(m, sc, ch_ptr)) 1006 goto retry_later; 1007 1008 m->m_data += sizeof(*ip); 1009 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip)); 1010 m->m_data -= sizeof(*ip); 1011 1012 nanotime(&sc->sc_if.if_lastchange); 1013 sc->sc_if.if_opackets++; 1014 sc->sc_if.if_obytes += len; 1015 CARP_STATINC(CARP_STAT_OPACKETS); 1016 1017 error = ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, 1018 NULL); 1019 if (error) { 1020 if (error == ENOBUFS) 1021 CARP_STATINC(CARP_STAT_ONOMEM); 1022 else 1023 CARP_LOG(sc, ("ip_output failed: %d", error)); 1024 sc->sc_if.if_oerrors++; 1025 if (sc->sc_sendad_errors < INT_MAX) 1026 sc->sc_sendad_errors++; 1027 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) { 1028 carp_suppress_preempt++; 1029 if (carp_suppress_preempt == 1) 1030 carp_send_ad_all(); 1031 } 1032 sc->sc_sendad_success = 0; 1033 } else { 1034 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) { 1035 if (++sc->sc_sendad_success >= 1036 CARP_SENDAD_MIN_SUCCESS) { 1037 carp_suppress_preempt--; 1038 sc->sc_sendad_errors = 0; 1039 } 1040 } else 1041 sc->sc_sendad_errors = 0; 1042 } 1043 } 1044 #endif /* INET */ 1045 #ifdef INET6 1046 if (sc->sc_naddrs6) { 1047 struct ip6_hdr *ip6; 1048 1049 MGETHDR(m, M_DONTWAIT, MT_HEADER); 1050 if (m == NULL) { 1051 sc->sc_if.if_oerrors++; 1052 CARP_STATINC(CARP_STAT_ONOMEM); 1053 /* XXX maybe less ? */ 1054 goto retry_later; 1055 } 1056 len = sizeof(*ip6) + sizeof(ch); 1057 m->m_pkthdr.len = len; 1058 m->m_pkthdr.rcvif = NULL; 1059 m->m_len = len; 1060 MH_ALIGN(m, m->m_len); 1061 m->m_flags |= M_MCAST; 1062 ip6 = mtod(m, struct ip6_hdr *); 1063 memset(ip6, 0, sizeof(*ip6)); 1064 ip6->ip6_vfc |= IPV6_VERSION; 1065 ip6->ip6_hlim = CARP_DFLTTL; 1066 ip6->ip6_nxt = IPPROTO_CARP; 1067 1068 /* set the source address */ 1069 memset(&sa, 0, sizeof(sa)); 1070 sa.sa_family = AF_INET6; 1071 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev); 1072 if (ifa == NULL) /* This should never happen with IPv6 */ 1073 memset(&ip6->ip6_src, 0, sizeof(struct in6_addr)); 1074 else 1075 bcopy(ifatoia6(ifa)->ia_addr.sin6_addr.s6_addr, 1076 &ip6->ip6_src, sizeof(struct in6_addr)); 1077 /* set the multicast destination */ 1078 1079 ip6->ip6_dst.s6_addr16[0] = htons(0xff02); 1080 ip6->ip6_dst.s6_addr8[15] = 0x12; 1081 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) { 1082 sc->sc_if.if_oerrors++; 1083 m_freem(m); 1084 CARP_LOG(sc, ("in6_setscope failed")); 1085 goto retry_later; 1086 } 1087 1088 ch_ptr = (struct carp_header *)(&ip6[1]); 1089 memcpy(ch_ptr, &ch, sizeof(ch)); 1090 if (carp_prepare_ad(m, sc, ch_ptr)) 1091 goto retry_later; 1092 1093 m->m_data += sizeof(*ip6); 1094 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6)); 1095 m->m_data -= sizeof(*ip6); 1096 1097 nanotime(&sc->sc_if.if_lastchange); 1098 sc->sc_if.if_opackets++; 1099 sc->sc_if.if_obytes += len; 1100 CARP_STATINC(CARP_STAT_OPACKETS6); 1101 1102 error = ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL); 1103 if (error) { 1104 if (error == ENOBUFS) 1105 CARP_STATINC(CARP_STAT_ONOMEM); 1106 else 1107 CARP_LOG(sc, ("ip6_output failed: %d", error)); 1108 sc->sc_if.if_oerrors++; 1109 if (sc->sc_sendad_errors < INT_MAX) 1110 sc->sc_sendad_errors++; 1111 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) { 1112 carp_suppress_preempt++; 1113 if (carp_suppress_preempt == 1) 1114 carp_send_ad_all(); 1115 } 1116 sc->sc_sendad_success = 0; 1117 } else { 1118 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) { 1119 if (++sc->sc_sendad_success >= 1120 CARP_SENDAD_MIN_SUCCESS) { 1121 carp_suppress_preempt--; 1122 sc->sc_sendad_errors = 0; 1123 } 1124 } else 1125 sc->sc_sendad_errors = 0; 1126 } 1127 } 1128 #endif /* INET6 */ 1129 1130 retry_later: 1131 splx(s); 1132 if (advbase != 255 || advskew != 255) 1133 callout_schedule(&sc->sc_ad_tmo, tvtohz(&tv)); 1134 } 1135 1136 /* 1137 * Broadcast a gratuitous ARP request containing 1138 * the virtual router MAC address for each IP address 1139 * associated with the virtual router. 1140 */ 1141 void 1142 carp_send_arp(struct carp_softc *sc) 1143 { 1144 struct ifaddr *ifa; 1145 struct in_addr *in; 1146 int s = splsoftnet(); 1147 1148 IFADDR_FOREACH(ifa, &sc->sc_if) { 1149 1150 if (ifa->ifa_addr->sa_family != AF_INET) 1151 continue; 1152 1153 in = &ifatoia(ifa)->ia_addr.sin_addr; 1154 arprequest(sc->sc_carpdev, in, in, CLLADDR(sc->sc_if.if_sadl)); 1155 DELAY(1000); /* XXX */ 1156 } 1157 splx(s); 1158 } 1159 1160 #ifdef INET6 1161 void 1162 carp_send_na(struct carp_softc *sc) 1163 { 1164 struct ifaddr *ifa; 1165 struct in6_addr *in6; 1166 static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT; 1167 int s = splsoftnet(); 1168 1169 IFADDR_FOREACH(ifa, &sc->sc_if) { 1170 1171 if (ifa->ifa_addr->sa_family != AF_INET6) 1172 continue; 1173 1174 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr; 1175 nd6_na_output(sc->sc_carpdev, &mcast, in6, 1176 ND_NA_FLAG_OVERRIDE, 1, NULL); 1177 DELAY(1000); /* XXX */ 1178 } 1179 splx(s); 1180 } 1181 #endif /* INET6 */ 1182 1183 /* 1184 * Based on bridge_hash() in if_bridge.c 1185 */ 1186 #define mix(a,b,c) \ 1187 do { \ 1188 a -= b; a -= c; a ^= (c >> 13); \ 1189 b -= c; b -= a; b ^= (a << 8); \ 1190 c -= a; c -= b; c ^= (b >> 13); \ 1191 a -= b; a -= c; a ^= (c >> 12); \ 1192 b -= c; b -= a; b ^= (a << 16); \ 1193 c -= a; c -= b; c ^= (b >> 5); \ 1194 a -= b; a -= c; a ^= (c >> 3); \ 1195 b -= c; b -= a; b ^= (a << 10); \ 1196 c -= a; c -= b; c ^= (b >> 15); \ 1197 } while (0) 1198 1199 u_int32_t 1200 carp_hash(struct carp_softc *sc, u_char *src) 1201 { 1202 u_int32_t a = 0x9e3779b9, b = sc->sc_hashkey[0], c = sc->sc_hashkey[1]; 1203 1204 c += sc->sc_key[3] << 24; 1205 c += sc->sc_key[2] << 16; 1206 c += sc->sc_key[1] << 8; 1207 c += sc->sc_key[0]; 1208 b += src[5] << 8; 1209 b += src[4]; 1210 a += src[3] << 24; 1211 a += src[2] << 16; 1212 a += src[1] << 8; 1213 a += src[0]; 1214 1215 mix(a, b, c); 1216 return (c); 1217 } 1218 1219 int 1220 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type) 1221 { 1222 struct carp_softc *vh; 1223 struct ifaddr *ifa; 1224 int count = 0; 1225 1226 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1227 if ((type == CARP_COUNT_RUNNING && 1228 (vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1229 (IFF_UP|IFF_RUNNING)) || 1230 (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) { 1231 IFADDR_FOREACH(ifa, &vh->sc_if) { 1232 if (ifa->ifa_addr->sa_family == AF_INET && 1233 ia->ia_addr.sin_addr.s_addr == 1234 ifatoia(ifa)->ia_addr.sin_addr.s_addr) 1235 count++; 1236 } 1237 } 1238 } 1239 return (count); 1240 } 1241 1242 int 1243 carp_iamatch(struct in_ifaddr *ia, u_char *src, 1244 u_int32_t *count, u_int32_t index) 1245 { 1246 struct carp_softc *sc = ia->ia_ifp->if_softc; 1247 1248 if (carp_opts[CARPCTL_ARPBALANCE]) { 1249 /* 1250 * We use the source ip to decide which virtual host should 1251 * handle the request. If we're master of that virtual host, 1252 * then we respond, otherwise, just drop the arp packet on 1253 * the floor. 1254 */ 1255 1256 /* Count the elegible carp interfaces with this address */ 1257 if (*count == 0) 1258 *count = carp_addrcount( 1259 (struct carp_if *)ia->ia_ifp->if_carpdev->if_carp, 1260 ia, CARP_COUNT_RUNNING); 1261 1262 /* This should never happen, but... */ 1263 if (*count == 0) 1264 return (0); 1265 1266 if (carp_hash(sc, src) % *count == index - 1 && 1267 sc->sc_state == MASTER) { 1268 return (1); 1269 } 1270 } else { 1271 if (sc->sc_state == MASTER) 1272 return (1); 1273 } 1274 1275 return (0); 1276 } 1277 1278 #ifdef INET6 1279 struct ifaddr * 1280 carp_iamatch6(void *v, struct in6_addr *taddr) 1281 { 1282 struct carp_if *cif = v; 1283 struct carp_softc *vh; 1284 struct ifaddr *ifa; 1285 1286 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1287 IFADDR_FOREACH(ifa, &vh->sc_if) { 1288 if (IN6_ARE_ADDR_EQUAL(taddr, 1289 &ifatoia6(ifa)->ia_addr.sin6_addr) && 1290 ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1291 (IFF_UP|IFF_RUNNING)) && vh->sc_state == MASTER) 1292 return (ifa); 1293 } 1294 } 1295 1296 return (NULL); 1297 } 1298 #endif /* INET6 */ 1299 1300 struct ifnet * 1301 carp_ourether(void *v, struct ether_header *eh, u_char iftype, int src) 1302 { 1303 struct carp_if *cif = (struct carp_if *)v; 1304 struct carp_softc *vh; 1305 u_int8_t *ena; 1306 1307 if (src) 1308 ena = (u_int8_t *)&eh->ether_shost; 1309 else 1310 ena = (u_int8_t *)&eh->ether_dhost; 1311 1312 switch (iftype) { 1313 case IFT_ETHER: 1314 case IFT_FDDI: 1315 if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1) 1316 return (NULL); 1317 break; 1318 case IFT_ISO88025: 1319 if (ena[0] != 3 || ena[1] || ena[4] || ena[5]) 1320 return (NULL); 1321 break; 1322 default: 1323 return (NULL); 1324 break; 1325 } 1326 1327 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) 1328 if ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1329 (IFF_UP|IFF_RUNNING) && vh->sc_state == MASTER && 1330 !memcmp(ena, CLLADDR(vh->sc_if.if_sadl), 1331 ETHER_ADDR_LEN)) { 1332 return (&vh->sc_if); 1333 } 1334 1335 return (NULL); 1336 } 1337 1338 int 1339 carp_input(struct mbuf *m, u_int8_t *shost, u_int8_t *dhost, u_int16_t etype) 1340 { 1341 struct ether_header eh; 1342 struct carp_if *cif = (struct carp_if *)m->m_pkthdr.rcvif->if_carp; 1343 struct ifnet *ifp; 1344 1345 memcpy(&eh.ether_shost, shost, sizeof(eh.ether_shost)); 1346 memcpy(&eh.ether_dhost, dhost, sizeof(eh.ether_dhost)); 1347 eh.ether_type = etype; 1348 1349 if (m->m_flags & (M_BCAST|M_MCAST)) { 1350 struct carp_softc *vh; 1351 struct mbuf *m0; 1352 1353 /* 1354 * XXX Should really check the list of multicast addresses 1355 * for each CARP interface _before_ copying. 1356 */ 1357 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1358 m0 = m_copym(m, 0, M_COPYALL, M_DONTWAIT); 1359 if (m0 == NULL) 1360 continue; 1361 m0->m_pkthdr.rcvif = &vh->sc_if; 1362 ether_input(&vh->sc_if, m0); 1363 } 1364 return (1); 1365 } 1366 1367 ifp = carp_ourether(cif, &eh, m->m_pkthdr.rcvif->if_type, 0); 1368 if (ifp == NULL) { 1369 return (1); 1370 } 1371 1372 m->m_pkthdr.rcvif = ifp; 1373 1374 #if NBPFILTER > 0 1375 if (ifp->if_bpf) 1376 bpf_mtap(ifp->if_bpf, m); 1377 #endif 1378 ifp->if_ipackets++; 1379 ether_input(ifp, m); 1380 return (0); 1381 } 1382 1383 void 1384 carp_master_down(void *v) 1385 { 1386 struct carp_softc *sc = v; 1387 1388 switch (sc->sc_state) { 1389 case INIT: 1390 printf("%s: master_down event in INIT state\n", 1391 sc->sc_if.if_xname); 1392 break; 1393 case MASTER: 1394 break; 1395 case BACKUP: 1396 CARP_LOG(sc, ("INIT -> MASTER (preempting)")); 1397 carp_set_state(sc, MASTER); 1398 carp_send_ad(sc); 1399 carp_send_arp(sc); 1400 #ifdef INET6 1401 carp_send_na(sc); 1402 #endif /* INET6 */ 1403 carp_setrun(sc, 0); 1404 carp_setroute(sc, RTM_ADD); 1405 break; 1406 } 1407 } 1408 1409 /* 1410 * When in backup state, af indicates whether to reset the master down timer 1411 * for v4 or v6. If it's set to zero, reset the ones which are already pending. 1412 */ 1413 void 1414 carp_setrun(struct carp_softc *sc, sa_family_t af) 1415 { 1416 struct timeval tv; 1417 1418 if (sc->sc_carpdev == NULL) { 1419 sc->sc_if.if_flags &= ~IFF_RUNNING; 1420 carp_set_state(sc, INIT); 1421 return; 1422 } 1423 1424 if (sc->sc_if.if_flags & IFF_UP && sc->sc_vhid > 0 && 1425 (sc->sc_naddrs || sc->sc_naddrs6) && !sc->sc_suppress) { 1426 sc->sc_if.if_flags |= IFF_RUNNING; 1427 } else { 1428 sc->sc_if.if_flags &= ~IFF_RUNNING; 1429 carp_setroute(sc, RTM_DELETE); 1430 return; 1431 } 1432 1433 switch (sc->sc_state) { 1434 case INIT: 1435 carp_set_state(sc, BACKUP); 1436 carp_setroute(sc, RTM_DELETE); 1437 carp_setrun(sc, 0); 1438 break; 1439 case BACKUP: 1440 callout_stop(&sc->sc_ad_tmo); 1441 tv.tv_sec = 3 * sc->sc_advbase; 1442 tv.tv_usec = sc->sc_advskew * 1000000 / 256; 1443 switch (af) { 1444 #ifdef INET 1445 case AF_INET: 1446 callout_schedule(&sc->sc_md_tmo, tvtohz(&tv)); 1447 break; 1448 #endif /* INET */ 1449 #ifdef INET6 1450 case AF_INET6: 1451 callout_schedule(&sc->sc_md6_tmo, tvtohz(&tv)); 1452 break; 1453 #endif /* INET6 */ 1454 default: 1455 if (sc->sc_naddrs) 1456 callout_schedule(&sc->sc_md_tmo, tvtohz(&tv)); 1457 if (sc->sc_naddrs6) 1458 callout_schedule(&sc->sc_md6_tmo, tvtohz(&tv)); 1459 break; 1460 } 1461 break; 1462 case MASTER: 1463 tv.tv_sec = sc->sc_advbase; 1464 tv.tv_usec = sc->sc_advskew * 1000000 / 256; 1465 callout_schedule(&sc->sc_ad_tmo, tvtohz(&tv)); 1466 break; 1467 } 1468 } 1469 1470 void 1471 carp_multicast_cleanup(struct carp_softc *sc) 1472 { 1473 struct ip_moptions *imo = &sc->sc_imo; 1474 #ifdef INET6 1475 struct ip6_moptions *im6o = &sc->sc_im6o; 1476 #endif 1477 u_int16_t n = imo->imo_num_memberships; 1478 1479 /* Clean up our own multicast memberships */ 1480 while (n-- > 0) { 1481 if (imo->imo_membership[n] != NULL) { 1482 in_delmulti(imo->imo_membership[n]); 1483 imo->imo_membership[n] = NULL; 1484 } 1485 } 1486 imo->imo_num_memberships = 0; 1487 imo->imo_multicast_ifp = NULL; 1488 1489 #ifdef INET6 1490 while (!LIST_EMPTY(&im6o->im6o_memberships)) { 1491 struct in6_multi_mship *imm = 1492 LIST_FIRST(&im6o->im6o_memberships); 1493 1494 LIST_REMOVE(imm, i6mm_chain); 1495 in6_leavegroup(imm); 1496 } 1497 im6o->im6o_multicast_ifp = NULL; 1498 #endif 1499 1500 /* And any other multicast memberships */ 1501 carp_ether_purgemulti(sc); 1502 } 1503 1504 int 1505 carp_set_ifp(struct carp_softc *sc, struct ifnet *ifp) 1506 { 1507 struct carp_if *cif, *ncif = NULL; 1508 struct carp_softc *vr, *after = NULL; 1509 int myself = 0, error = 0; 1510 int s; 1511 1512 if (ifp == sc->sc_carpdev) 1513 return (0); 1514 1515 if (ifp != NULL) { 1516 if ((ifp->if_flags & IFF_MULTICAST) == 0) 1517 return (EADDRNOTAVAIL); 1518 1519 if (ifp->if_type == IFT_CARP) 1520 return (EINVAL); 1521 1522 if (ifp->if_carp == NULL) { 1523 ncif = malloc(sizeof(*cif), M_IFADDR, M_NOWAIT); 1524 if (ncif == NULL) 1525 return (ENOBUFS); 1526 if ((error = ifpromisc(ifp, 1))) { 1527 free(ncif, M_IFADDR); 1528 return (error); 1529 } 1530 1531 ncif->vhif_ifp = ifp; 1532 TAILQ_INIT(&ncif->vhif_vrs); 1533 } else { 1534 cif = (struct carp_if *)ifp->if_carp; 1535 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) 1536 if (vr != sc && vr->sc_vhid == sc->sc_vhid) 1537 return (EINVAL); 1538 } 1539 1540 /* detach from old interface */ 1541 if (sc->sc_carpdev != NULL) 1542 carpdetach(sc); 1543 1544 /* join multicast groups */ 1545 if (sc->sc_naddrs < 0 && 1546 (error = carp_join_multicast(sc)) != 0) { 1547 if (ncif != NULL) 1548 free(ncif, M_IFADDR); 1549 return (error); 1550 } 1551 1552 #ifdef INET6 1553 if (sc->sc_naddrs6 < 0 && 1554 (error = carp_join_multicast6(sc)) != 0) { 1555 if (ncif != NULL) 1556 free(ncif, M_IFADDR); 1557 carp_multicast_cleanup(sc); 1558 return (error); 1559 } 1560 #endif 1561 1562 /* attach carp interface to physical interface */ 1563 if (ncif != NULL) 1564 ifp->if_carp = (void *)ncif; 1565 sc->sc_carpdev = ifp; 1566 cif = (struct carp_if *)ifp->if_carp; 1567 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) { 1568 if (vr == sc) 1569 myself = 1; 1570 if (vr->sc_vhid < sc->sc_vhid) 1571 after = vr; 1572 } 1573 1574 if (!myself) { 1575 /* We're trying to keep things in order */ 1576 if (after == NULL) { 1577 TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list); 1578 } else { 1579 TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, 1580 sc, sc_list); 1581 } 1582 cif->vhif_nvrs++; 1583 } 1584 if (sc->sc_naddrs || sc->sc_naddrs6) 1585 sc->sc_if.if_flags |= IFF_UP; 1586 carp_set_enaddr(sc); 1587 s = splnet(); 1588 /* XXX linkstatehooks establish */ 1589 carp_carpdev_state(ifp); 1590 splx(s); 1591 } else { 1592 carpdetach(sc); 1593 sc->sc_if.if_flags &= ~(IFF_UP|IFF_RUNNING); 1594 } 1595 return (0); 1596 } 1597 1598 void 1599 carp_set_enaddr(struct carp_softc *sc) 1600 { 1601 uint8_t enaddr[ETHER_ADDR_LEN]; 1602 if (sc->sc_carpdev && sc->sc_carpdev->if_type == IFT_ISO88025) { 1603 enaddr[0] = 3; 1604 enaddr[1] = 0; 1605 enaddr[2] = 0x40 >> (sc->sc_vhid - 1); 1606 enaddr[3] = 0x40000 >> (sc->sc_vhid - 1); 1607 enaddr[4] = 0; 1608 enaddr[5] = 0; 1609 } else { 1610 enaddr[0] = 0; 1611 enaddr[1] = 0; 1612 enaddr[2] = 0x5e; 1613 enaddr[3] = 0; 1614 enaddr[4] = 1; 1615 enaddr[5] = sc->sc_vhid; 1616 } 1617 if_set_sadl(&sc->sc_if, enaddr, sizeof(enaddr), false); 1618 } 1619 1620 void 1621 carp_addr_updated(void *v) 1622 { 1623 struct carp_softc *sc = (struct carp_softc *) v; 1624 struct ifaddr *ifa; 1625 int new_naddrs = 0, new_naddrs6 = 0; 1626 1627 IFADDR_FOREACH(ifa, &sc->sc_if) { 1628 if (ifa->ifa_addr->sa_family == AF_INET) 1629 new_naddrs++; 1630 else if (ifa->ifa_addr->sa_family == AF_INET6) 1631 new_naddrs6++; 1632 } 1633 1634 /* Handle a callback after SIOCDIFADDR */ 1635 if (new_naddrs < sc->sc_naddrs || new_naddrs6 < sc->sc_naddrs6) { 1636 struct in_addr mc_addr; 1637 struct in_multi *inm; 1638 1639 sc->sc_naddrs = new_naddrs; 1640 sc->sc_naddrs6 = new_naddrs6; 1641 1642 /* Re-establish multicast membership removed by in_control */ 1643 mc_addr.s_addr = INADDR_CARP_GROUP; 1644 IN_LOOKUP_MULTI(mc_addr, &sc->sc_if, inm); 1645 if (inm == NULL) { 1646 memset(&sc->sc_imo, 0, sizeof(sc->sc_imo)); 1647 1648 if (sc->sc_carpdev != NULL && sc->sc_naddrs > 0) 1649 carp_join_multicast(sc); 1650 } 1651 1652 if (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) { 1653 sc->sc_if.if_flags &= ~IFF_UP; 1654 carp_set_state(sc, INIT); 1655 } else 1656 carp_hmac_prepare(sc); 1657 } 1658 1659 carp_setrun(sc, 0); 1660 } 1661 1662 int 1663 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin) 1664 { 1665 struct ifnet *ifp = sc->sc_carpdev; 1666 struct in_ifaddr *ia, *ia_if; 1667 int error = 0; 1668 1669 if (sin->sin_addr.s_addr == 0) { 1670 if (!(sc->sc_if.if_flags & IFF_UP)) 1671 carp_set_state(sc, INIT); 1672 if (sc->sc_naddrs) 1673 sc->sc_if.if_flags |= IFF_UP; 1674 carp_setrun(sc, 0); 1675 return (0); 1676 } 1677 1678 /* we have to do this by hand to ensure we don't match on ourselves */ 1679 ia_if = NULL; 1680 for (ia = TAILQ_FIRST(&in_ifaddrhead); ia; 1681 ia = TAILQ_NEXT(ia, ia_list)) { 1682 1683 /* and, yeah, we need a multicast-capable iface too */ 1684 if (ia->ia_ifp != &sc->sc_if && 1685 ia->ia_ifp->if_type != IFT_CARP && 1686 (ia->ia_ifp->if_flags & IFF_MULTICAST) && 1687 (sin->sin_addr.s_addr & ia->ia_subnetmask) == 1688 ia->ia_subnet) { 1689 if (!ia_if) 1690 ia_if = ia; 1691 } 1692 } 1693 1694 if (ia_if) { 1695 ia = ia_if; 1696 if (ifp) { 1697 if (ifp != ia->ia_ifp) 1698 return (EADDRNOTAVAIL); 1699 } else { 1700 ifp = ia->ia_ifp; 1701 } 1702 } 1703 1704 if ((error = carp_set_ifp(sc, ifp))) 1705 return (error); 1706 1707 if (sc->sc_carpdev == NULL) 1708 return (EADDRNOTAVAIL); 1709 1710 if (sc->sc_naddrs == 0 && (error = carp_join_multicast(sc)) != 0) 1711 return (error); 1712 1713 sc->sc_naddrs++; 1714 if (sc->sc_carpdev != NULL) 1715 sc->sc_if.if_flags |= IFF_UP; 1716 1717 carp_set_state(sc, INIT); 1718 carp_setrun(sc, 0); 1719 1720 /* 1721 * Hook if_addrhooks so that we get a callback after in_ifinit has run, 1722 * to correct any inappropriate routes that it inserted. 1723 */ 1724 if (sc->ah_cookie == 0) { 1725 /* XXX link address hook */ 1726 } 1727 1728 return (0); 1729 } 1730 1731 int 1732 carp_join_multicast(struct carp_softc *sc) 1733 { 1734 struct ip_moptions *imo = &sc->sc_imo, tmpimo; 1735 struct in_addr addr; 1736 1737 memset(&tmpimo, 0, sizeof(tmpimo)); 1738 addr.s_addr = INADDR_CARP_GROUP; 1739 if ((tmpimo.imo_membership[0] = 1740 in_addmulti(&addr, &sc->sc_if)) == NULL) { 1741 return (ENOBUFS); 1742 } 1743 1744 imo->imo_membership[0] = tmpimo.imo_membership[0]; 1745 imo->imo_num_memberships = 1; 1746 imo->imo_multicast_ifp = &sc->sc_if; 1747 imo->imo_multicast_ttl = CARP_DFLTTL; 1748 imo->imo_multicast_loop = 0; 1749 return (0); 1750 } 1751 1752 1753 #ifdef INET6 1754 int 1755 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6) 1756 { 1757 struct ifnet *ifp = sc->sc_carpdev; 1758 struct in6_ifaddr *ia, *ia_if; 1759 int error = 0; 1760 1761 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) { 1762 if (!(sc->sc_if.if_flags & IFF_UP)) 1763 carp_set_state(sc, INIT); 1764 if (sc->sc_naddrs6) 1765 sc->sc_if.if_flags |= IFF_UP; 1766 carp_setrun(sc, 0); 1767 return (0); 1768 } 1769 1770 /* we have to do this by hand to ensure we don't match on ourselves */ 1771 ia_if = NULL; 1772 for (ia = in6_ifaddr; ia; ia = ia->ia_next) { 1773 int i; 1774 1775 for (i = 0; i < 4; i++) { 1776 if ((sin6->sin6_addr.s6_addr32[i] & 1777 ia->ia_prefixmask.sin6_addr.s6_addr32[i]) != 1778 (ia->ia_addr.sin6_addr.s6_addr32[i] & 1779 ia->ia_prefixmask.sin6_addr.s6_addr32[i])) 1780 break; 1781 } 1782 /* and, yeah, we need a multicast-capable iface too */ 1783 if (ia->ia_ifp != &sc->sc_if && 1784 ia->ia_ifp->if_type != IFT_CARP && 1785 (ia->ia_ifp->if_flags & IFF_MULTICAST) && 1786 (i == 4)) { 1787 if (!ia_if) 1788 ia_if = ia; 1789 } 1790 } 1791 1792 if (ia_if) { 1793 ia = ia_if; 1794 if (sc->sc_carpdev) { 1795 if (sc->sc_carpdev != ia->ia_ifp) 1796 return (EADDRNOTAVAIL); 1797 } else { 1798 ifp = ia->ia_ifp; 1799 } 1800 } 1801 1802 if ((error = carp_set_ifp(sc, ifp))) 1803 return (error); 1804 1805 if (sc->sc_carpdev == NULL) 1806 return (EADDRNOTAVAIL); 1807 1808 if (sc->sc_naddrs6 == 0 && (error = carp_join_multicast6(sc)) != 0) 1809 return (error); 1810 1811 sc->sc_naddrs6++; 1812 if (sc->sc_carpdev != NULL) 1813 sc->sc_if.if_flags |= IFF_UP; 1814 carp_set_state(sc, INIT); 1815 carp_setrun(sc, 0); 1816 1817 return (0); 1818 } 1819 1820 int 1821 carp_join_multicast6(struct carp_softc *sc) 1822 { 1823 struct in6_multi_mship *imm, *imm2; 1824 struct ip6_moptions *im6o = &sc->sc_im6o; 1825 struct sockaddr_in6 addr6; 1826 int error; 1827 1828 /* Join IPv6 CARP multicast group */ 1829 memset(&addr6, 0, sizeof(addr6)); 1830 addr6.sin6_family = AF_INET6; 1831 addr6.sin6_len = sizeof(addr6); 1832 addr6.sin6_addr.s6_addr16[0] = htons(0xff02); 1833 addr6.sin6_addr.s6_addr16[1] = htons(sc->sc_if.if_index); 1834 addr6.sin6_addr.s6_addr8[15] = 0x12; 1835 if ((imm = in6_joingroup(&sc->sc_if, 1836 &addr6.sin6_addr, &error, 0)) == NULL) { 1837 return (error); 1838 } 1839 /* join solicited multicast address */ 1840 memset(&addr6.sin6_addr, 0, sizeof(addr6.sin6_addr)); 1841 addr6.sin6_addr.s6_addr16[0] = htons(0xff02); 1842 addr6.sin6_addr.s6_addr16[1] = htons(sc->sc_if.if_index); 1843 addr6.sin6_addr.s6_addr32[1] = 0; 1844 addr6.sin6_addr.s6_addr32[2] = htonl(1); 1845 addr6.sin6_addr.s6_addr32[3] = 0; 1846 addr6.sin6_addr.s6_addr8[12] = 0xff; 1847 if ((imm2 = in6_joingroup(&sc->sc_if, 1848 &addr6.sin6_addr, &error, 0)) == NULL) { 1849 in6_leavegroup(imm); 1850 return (error); 1851 } 1852 1853 /* apply v6 multicast membership */ 1854 im6o->im6o_multicast_ifp = &sc->sc_if; 1855 if (imm) 1856 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, 1857 i6mm_chain); 1858 if (imm2) 1859 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm2, 1860 i6mm_chain); 1861 1862 return (0); 1863 } 1864 1865 #endif /* INET6 */ 1866 1867 int 1868 carp_ioctl(struct ifnet *ifp, u_long cmd, void *data) 1869 { 1870 struct lwp *l = curlwp; /* XXX */ 1871 struct carp_softc *sc = ifp->if_softc, *vr; 1872 struct carpreq carpr; 1873 struct ifaddr *ifa; 1874 struct ifreq *ifr; 1875 struct ifnet *cdev = NULL; 1876 int error = 0; 1877 1878 ifa = (struct ifaddr *)data; 1879 ifr = (struct ifreq *)data; 1880 1881 switch (cmd) { 1882 case SIOCINITIFADDR: 1883 switch (ifa->ifa_addr->sa_family) { 1884 #ifdef INET 1885 case AF_INET: 1886 sc->sc_if.if_flags |= IFF_UP; 1887 memcpy(ifa->ifa_dstaddr, ifa->ifa_addr, 1888 sizeof(struct sockaddr)); 1889 error = carp_set_addr(sc, satosin(ifa->ifa_addr)); 1890 break; 1891 #endif /* INET */ 1892 #ifdef INET6 1893 case AF_INET6: 1894 sc->sc_if.if_flags|= IFF_UP; 1895 error = carp_set_addr6(sc, satosin6(ifa->ifa_addr)); 1896 break; 1897 #endif /* INET6 */ 1898 default: 1899 error = EAFNOSUPPORT; 1900 break; 1901 } 1902 break; 1903 1904 case SIOCSIFFLAGS: 1905 if ((error = ifioctl_common(ifp, cmd, data)) != 0) 1906 break; 1907 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) { 1908 callout_stop(&sc->sc_ad_tmo); 1909 callout_stop(&sc->sc_md_tmo); 1910 callout_stop(&sc->sc_md6_tmo); 1911 if (sc->sc_state == MASTER) { 1912 /* we need the interface up to bow out */ 1913 sc->sc_if.if_flags |= IFF_UP; 1914 sc->sc_bow_out = 1; 1915 carp_send_ad(sc); 1916 } 1917 sc->sc_if.if_flags &= ~IFF_UP; 1918 carp_set_state(sc, INIT); 1919 carp_setrun(sc, 0); 1920 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) { 1921 sc->sc_if.if_flags |= IFF_UP; 1922 carp_setrun(sc, 0); 1923 } 1924 break; 1925 1926 case SIOCSVH: 1927 if (l == NULL) 1928 break; 1929 if ((error = kauth_authorize_network(l->l_cred, 1930 KAUTH_NETWORK_INTERFACE, 1931 KAUTH_REQ_NETWORK_INTERFACE_SETPRIV, ifp, (void *)cmd, 1932 NULL)) != 0) 1933 break; 1934 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr))) 1935 break; 1936 error = 1; 1937 if (carpr.carpr_carpdev[0] != '\0' && 1938 (cdev = ifunit(carpr.carpr_carpdev)) == NULL) 1939 return (EINVAL); 1940 if ((error = carp_set_ifp(sc, cdev))) 1941 return (error); 1942 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) { 1943 switch (carpr.carpr_state) { 1944 case BACKUP: 1945 callout_stop(&sc->sc_ad_tmo); 1946 carp_set_state(sc, BACKUP); 1947 carp_setrun(sc, 0); 1948 carp_setroute(sc, RTM_DELETE); 1949 break; 1950 case MASTER: 1951 carp_master_down(sc); 1952 break; 1953 default: 1954 break; 1955 } 1956 } 1957 if (carpr.carpr_vhid > 0) { 1958 if (carpr.carpr_vhid > 255) { 1959 error = EINVAL; 1960 break; 1961 } 1962 if (sc->sc_carpdev) { 1963 struct carp_if *cif; 1964 cif = (struct carp_if *)sc->sc_carpdev->if_carp; 1965 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) 1966 if (vr != sc && 1967 vr->sc_vhid == carpr.carpr_vhid) 1968 return (EINVAL); 1969 } 1970 sc->sc_vhid = carpr.carpr_vhid; 1971 carp_set_enaddr(sc); 1972 carp_set_state(sc, INIT); 1973 error--; 1974 } 1975 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) { 1976 if (carpr.carpr_advskew > 254) { 1977 error = EINVAL; 1978 break; 1979 } 1980 if (carpr.carpr_advbase > 255) { 1981 error = EINVAL; 1982 break; 1983 } 1984 sc->sc_advbase = carpr.carpr_advbase; 1985 sc->sc_advskew = carpr.carpr_advskew; 1986 error--; 1987 } 1988 memcpy(sc->sc_key, carpr.carpr_key, sizeof(sc->sc_key)); 1989 if (error > 0) 1990 error = EINVAL; 1991 else { 1992 error = 0; 1993 carp_setrun(sc, 0); 1994 } 1995 break; 1996 1997 case SIOCGVH: 1998 memset(&carpr, 0, sizeof(carpr)); 1999 if (sc->sc_carpdev != NULL) 2000 strlcpy(carpr.carpr_carpdev, sc->sc_carpdev->if_xname, 2001 IFNAMSIZ); 2002 carpr.carpr_state = sc->sc_state; 2003 carpr.carpr_vhid = sc->sc_vhid; 2004 carpr.carpr_advbase = sc->sc_advbase; 2005 carpr.carpr_advskew = sc->sc_advskew; 2006 2007 if ((l != NULL) && (error = kauth_authorize_network(l->l_cred, 2008 KAUTH_NETWORK_INTERFACE, 2009 KAUTH_REQ_NETWORK_INTERFACE_SETPRIV, ifp, (void *)cmd, 2010 NULL)) == 0) 2011 memcpy(carpr.carpr_key, sc->sc_key, 2012 sizeof(carpr.carpr_key)); 2013 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr)); 2014 break; 2015 2016 case SIOCADDMULTI: 2017 error = carp_ether_addmulti(sc, ifr); 2018 break; 2019 2020 case SIOCDELMULTI: 2021 error = carp_ether_delmulti(sc, ifr); 2022 break; 2023 2024 default: 2025 error = ether_ioctl(ifp, cmd, data); 2026 } 2027 2028 carp_hmac_prepare(sc); 2029 return (error); 2030 } 2031 2032 2033 /* 2034 * Start output on carp interface. This function should never be called. 2035 */ 2036 void 2037 carp_start(struct ifnet *ifp) 2038 { 2039 #ifdef DEBUG 2040 printf("%s: start called\n", ifp->if_xname); 2041 #endif 2042 } 2043 2044 int 2045 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa, 2046 struct rtentry *rt) 2047 { 2048 struct carp_softc *sc = ((struct carp_softc *)ifp->if_softc); 2049 2050 if (sc->sc_carpdev != NULL && sc->sc_state == MASTER) { 2051 return (sc->sc_carpdev->if_output(ifp, m, sa, rt)); 2052 } else { 2053 m_freem(m); 2054 return (ENETUNREACH); 2055 } 2056 } 2057 2058 void 2059 carp_set_state(struct carp_softc *sc, int state) 2060 { 2061 static const char *carp_states[] = { CARP_STATES }; 2062 if (sc->sc_state == state) 2063 return; 2064 2065 CARP_LOG(sc, ("state transition from: %s -> to: %s", carp_states[sc->sc_state], carp_states[state])); 2066 2067 sc->sc_state = state; 2068 switch (state) { 2069 case BACKUP: 2070 sc->sc_if.if_link_state = LINK_STATE_DOWN; 2071 break; 2072 case MASTER: 2073 sc->sc_if.if_link_state = LINK_STATE_UP; 2074 break; 2075 default: 2076 sc->sc_if.if_link_state = LINK_STATE_UNKNOWN; 2077 break; 2078 } 2079 rt_ifmsg(&sc->sc_if); 2080 } 2081 2082 void 2083 carp_carpdev_state(void *v) 2084 { 2085 struct carp_if *cif; 2086 struct carp_softc *sc; 2087 struct ifnet *ifp = v; 2088 2089 if (ifp->if_type == IFT_CARP) 2090 return; 2091 2092 cif = (struct carp_if *)ifp->if_carp; 2093 2094 TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) { 2095 int suppressed = sc->sc_suppress; 2096 2097 if (sc->sc_carpdev->if_link_state == LINK_STATE_DOWN || 2098 !(sc->sc_carpdev->if_flags & IFF_UP)) { 2099 sc->sc_if.if_flags &= ~IFF_RUNNING; 2100 callout_stop(&sc->sc_ad_tmo); 2101 callout_stop(&sc->sc_md_tmo); 2102 callout_stop(&sc->sc_md6_tmo); 2103 carp_set_state(sc, INIT); 2104 sc->sc_suppress = 1; 2105 carp_setrun(sc, 0); 2106 if (!suppressed) { 2107 carp_suppress_preempt++; 2108 if (carp_suppress_preempt == 1) 2109 carp_send_ad_all(); 2110 } 2111 } else { 2112 carp_set_state(sc, INIT); 2113 sc->sc_suppress = 0; 2114 carp_setrun(sc, 0); 2115 if (suppressed) 2116 carp_suppress_preempt--; 2117 } 2118 } 2119 } 2120 2121 int 2122 carp_ether_addmulti(struct carp_softc *sc, struct ifreq *ifr) 2123 { 2124 const struct sockaddr *sa = ifreq_getaddr(SIOCADDMULTI, ifr); 2125 struct ifnet *ifp; 2126 struct carp_mc_entry *mc; 2127 u_int8_t addrlo[ETHER_ADDR_LEN], addrhi[ETHER_ADDR_LEN]; 2128 int error; 2129 2130 ifp = sc->sc_carpdev; 2131 if (ifp == NULL) 2132 return (EINVAL); 2133 2134 error = ether_addmulti(sa, &sc->sc_ac); 2135 if (error != ENETRESET) 2136 return (error); 2137 2138 /* 2139 * This is new multicast address. We have to tell parent 2140 * about it. Also, remember this multicast address so that 2141 * we can delete them on unconfigure. 2142 */ 2143 mc = malloc(sizeof(struct carp_mc_entry), M_DEVBUF, M_NOWAIT); 2144 if (mc == NULL) { 2145 error = ENOMEM; 2146 goto alloc_failed; 2147 } 2148 2149 /* 2150 * As ether_addmulti() returns ENETRESET, following two 2151 * statement shouldn't fail. 2152 */ 2153 (void)ether_multiaddr(sa, addrlo, addrhi); 2154 ETHER_LOOKUP_MULTI(addrlo, addrhi, &sc->sc_ac, mc->mc_enm); 2155 memcpy(&mc->mc_addr, sa, sa->sa_len); 2156 LIST_INSERT_HEAD(&sc->carp_mc_listhead, mc, mc_entries); 2157 2158 error = (*ifp->if_ioctl)(ifp, SIOCADDMULTI, ifr); 2159 if (error != 0) 2160 goto ioctl_failed; 2161 2162 return (error); 2163 2164 ioctl_failed: 2165 LIST_REMOVE(mc, mc_entries); 2166 free(mc, M_DEVBUF); 2167 alloc_failed: 2168 (void)ether_delmulti(sa, &sc->sc_ac); 2169 2170 return (error); 2171 } 2172 2173 int 2174 carp_ether_delmulti(struct carp_softc *sc, struct ifreq *ifr) 2175 { 2176 const struct sockaddr *sa = ifreq_getaddr(SIOCDELMULTI, ifr); 2177 struct ifnet *ifp; 2178 struct ether_multi *enm; 2179 struct carp_mc_entry *mc; 2180 u_int8_t addrlo[ETHER_ADDR_LEN], addrhi[ETHER_ADDR_LEN]; 2181 int error; 2182 2183 ifp = sc->sc_carpdev; 2184 if (ifp == NULL) 2185 return (EINVAL); 2186 2187 /* 2188 * Find a key to lookup carp_mc_entry. We have to do this 2189 * before calling ether_delmulti for obvious reason. 2190 */ 2191 if ((error = ether_multiaddr(sa, addrlo, addrhi)) != 0) 2192 return (error); 2193 ETHER_LOOKUP_MULTI(addrlo, addrhi, &sc->sc_ac, enm); 2194 if (enm == NULL) 2195 return (EINVAL); 2196 2197 LIST_FOREACH(mc, &sc->carp_mc_listhead, mc_entries) 2198 if (mc->mc_enm == enm) 2199 break; 2200 2201 /* We won't delete entries we didn't add */ 2202 if (mc == NULL) 2203 return (EINVAL); 2204 2205 error = ether_delmulti(sa, &sc->sc_ac); 2206 if (error != ENETRESET) 2207 return (error); 2208 2209 /* We no longer use this multicast address. Tell parent so. */ 2210 error = (*ifp->if_ioctl)(ifp, SIOCDELMULTI, ifr); 2211 if (error == 0) { 2212 /* And forget about this address. */ 2213 LIST_REMOVE(mc, mc_entries); 2214 free(mc, M_DEVBUF); 2215 } else 2216 (void)ether_addmulti(sa, &sc->sc_ac); 2217 return (error); 2218 } 2219 2220 /* 2221 * Delete any multicast address we have asked to add from parent 2222 * interface. Called when the carp is being unconfigured. 2223 */ 2224 void 2225 carp_ether_purgemulti(struct carp_softc *sc) 2226 { 2227 struct ifnet *ifp = sc->sc_carpdev; /* Parent. */ 2228 struct carp_mc_entry *mc; 2229 union { 2230 struct ifreq ifreq; 2231 struct { 2232 char ifr_name[IFNAMSIZ]; 2233 struct sockaddr_storage ifr_ss; 2234 } ifreq_storage; 2235 } u; 2236 struct ifreq *ifr = &u.ifreq; 2237 2238 if (ifp == NULL) 2239 return; 2240 2241 memcpy(ifr->ifr_name, ifp->if_xname, IFNAMSIZ); 2242 while ((mc = LIST_FIRST(&sc->carp_mc_listhead)) != NULL) { 2243 memcpy(&ifr->ifr_addr, &mc->mc_addr, mc->mc_addr.ss_len); 2244 (void)(*ifp->if_ioctl)(ifp, SIOCDELMULTI, ifr); 2245 LIST_REMOVE(mc, mc_entries); 2246 free(mc, M_DEVBUF); 2247 } 2248 } 2249 2250 static int 2251 sysctl_net_inet_carp_stats(SYSCTLFN_ARGS) 2252 { 2253 2254 return (NETSTAT_SYSCTL(carpstat_percpu, CARP_NSTATS)); 2255 } 2256 2257 void 2258 carp_init(void) 2259 { 2260 2261 sysctl_net_inet_carp_setup(NULL); 2262 } 2263 2264 static void 2265 sysctl_net_inet_carp_setup(struct sysctllog **clog) 2266 { 2267 2268 sysctl_createv(clog, 0, NULL, NULL, 2269 CTLFLAG_PERMANENT, 2270 CTLTYPE_NODE, "net", NULL, 2271 NULL, 0, NULL, 0, 2272 CTL_NET, CTL_EOL); 2273 sysctl_createv(clog, 0, NULL, NULL, 2274 CTLFLAG_PERMANENT, 2275 CTLTYPE_NODE, "inet", NULL, 2276 NULL, 0, NULL, 0, 2277 CTL_NET, PF_INET, CTL_EOL); 2278 sysctl_createv(clog, 0, NULL, NULL, 2279 CTLFLAG_PERMANENT, 2280 CTLTYPE_NODE, "carp", 2281 SYSCTL_DESCR("CARP related settings"), 2282 NULL, 0, NULL, 0, 2283 CTL_NET, PF_INET, IPPROTO_CARP, CTL_EOL); 2284 2285 sysctl_createv(clog, 0, NULL, NULL, 2286 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2287 CTLTYPE_INT, "preempt", 2288 SYSCTL_DESCR("Enable CARP Preempt"), 2289 NULL, 0, &carp_opts[CARPCTL_PREEMPT], 0, 2290 CTL_NET, PF_INET, IPPROTO_CARP, 2291 CTL_CREATE, CTL_EOL); 2292 sysctl_createv(clog, 0, NULL, NULL, 2293 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2294 CTLTYPE_INT, "arpbalance", 2295 SYSCTL_DESCR("Enable ARP balancing"), 2296 NULL, 0, &carp_opts[CARPCTL_ARPBALANCE], 0, 2297 CTL_NET, PF_INET, IPPROTO_CARP, 2298 CTL_CREATE, CTL_EOL); 2299 sysctl_createv(clog, 0, NULL, NULL, 2300 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2301 CTLTYPE_INT, "allow", 2302 SYSCTL_DESCR("Enable CARP"), 2303 NULL, 0, &carp_opts[CARPCTL_ALLOW], 0, 2304 CTL_NET, PF_INET, IPPROTO_CARP, 2305 CTL_CREATE, CTL_EOL); 2306 sysctl_createv(clog, 0, NULL, NULL, 2307 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2308 CTLTYPE_INT, "log", 2309 SYSCTL_DESCR("CARP logging"), 2310 NULL, 0, &carp_opts[CARPCTL_LOG], 0, 2311 CTL_NET, PF_INET, IPPROTO_CARP, 2312 CTL_CREATE, CTL_EOL); 2313 sysctl_createv(clog, 0, NULL, NULL, 2314 CTLFLAG_PERMANENT, 2315 CTLTYPE_STRUCT, "stats", 2316 SYSCTL_DESCR("CARP statistics"), 2317 sysctl_net_inet_carp_stats, 0, NULL, 0, 2318 CTL_NET, PF_INET, IPPROTO_CARP, CARPCTL_STATS, 2319 CTL_EOL); 2320 } 2321