1 /* $NetBSD: ip_carp.c,v 1.62 2015/08/24 22:21:26 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 #ifdef _KERNEL_OPT 31 #include "opt_inet.h" 32 #include "opt_mbuftrace.h" 33 #endif 34 35 #include <sys/cdefs.h> 36 __KERNEL_RCSID(0, "$NetBSD: ip_carp.c,v 1.62 2015/08/24 22:21:26 pooka Exp $"); 37 38 /* 39 * TODO: 40 * - iface reconfigure 41 * - support for hardware checksum calculations; 42 * 43 */ 44 45 #include <sys/param.h> 46 #include <sys/proc.h> 47 #include <sys/mbuf.h> 48 #include <sys/socket.h> 49 #include <sys/socketvar.h> 50 #include <sys/callout.h> 51 #include <sys/ioctl.h> 52 #include <sys/errno.h> 53 #include <sys/device.h> 54 #include <sys/time.h> 55 #include <sys/kernel.h> 56 #include <sys/kauth.h> 57 #include <sys/sysctl.h> 58 #include <sys/ucred.h> 59 #include <sys/syslog.h> 60 #include <sys/acct.h> 61 #include <sys/cprng.h> 62 63 #include <sys/cpu.h> 64 65 #include <net/if.h> 66 #include <net/pfil.h> 67 #include <net/if_types.h> 68 #include <net/if_ether.h> 69 #include <net/route.h> 70 #include <net/netisr.h> 71 #include <net/net_stats.h> 72 #include <netinet/if_inarp.h> 73 74 #if NFDDI > 0 75 #include <net/if_fddi.h> 76 #endif 77 #if NTOKEN > 0 78 #include <net/if_token.h> 79 #endif 80 81 #ifdef INET 82 #include <netinet/in.h> 83 #include <netinet/in_systm.h> 84 #include <netinet/in_var.h> 85 #include <netinet/ip.h> 86 #include <netinet/ip_var.h> 87 88 #include <net/if_dl.h> 89 #endif 90 91 #ifdef INET6 92 #include <netinet/icmp6.h> 93 #include <netinet/ip6.h> 94 #include <netinet6/ip6_var.h> 95 #include <netinet6/nd6.h> 96 #include <netinet6/scope6_var.h> 97 #endif 98 99 #include <net/bpf.h> 100 101 #include <sys/sha1.h> 102 103 #include <netinet/ip_carp.h> 104 105 #include "ioconf.h" 106 107 struct carp_mc_entry { 108 LIST_ENTRY(carp_mc_entry) mc_entries; 109 union { 110 struct ether_multi *mcu_enm; 111 } mc_u; 112 struct sockaddr_storage mc_addr; 113 }; 114 #define mc_enm mc_u.mcu_enm 115 116 struct carp_softc { 117 struct ethercom sc_ac; 118 #define sc_if sc_ac.ec_if 119 #define sc_carpdev sc_ac.ec_if.if_carpdev 120 int ah_cookie; 121 int lh_cookie; 122 struct ip_moptions sc_imo; 123 #ifdef INET6 124 struct ip6_moptions sc_im6o; 125 #endif /* INET6 */ 126 TAILQ_ENTRY(carp_softc) sc_list; 127 128 enum { INIT = 0, BACKUP, MASTER } sc_state; 129 130 int sc_suppress; 131 int sc_bow_out; 132 133 int sc_sendad_errors; 134 #define CARP_SENDAD_MAX_ERRORS 3 135 int sc_sendad_success; 136 #define CARP_SENDAD_MIN_SUCCESS 3 137 138 int sc_vhid; 139 int sc_advskew; 140 int sc_naddrs; 141 int sc_naddrs6; 142 int sc_advbase; /* seconds */ 143 int sc_init_counter; 144 u_int64_t sc_counter; 145 146 /* authentication */ 147 #define CARP_HMAC_PAD 64 148 unsigned char sc_key[CARP_KEY_LEN]; 149 unsigned char sc_pad[CARP_HMAC_PAD]; 150 SHA1_CTX sc_sha1; 151 u_int32_t sc_hashkey[2]; 152 153 struct callout sc_ad_tmo; /* advertisement timeout */ 154 struct callout sc_md_tmo; /* master down timeout */ 155 struct callout sc_md6_tmo; /* master down timeout */ 156 157 LIST_HEAD(__carp_mchead, carp_mc_entry) carp_mc_listhead; 158 }; 159 160 int carp_suppress_preempt = 0; 161 static int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 0, 0 }; /* XXX for now */ 162 163 static percpu_t *carpstat_percpu; 164 165 #define CARP_STATINC(x) _NET_STATINC(carpstat_percpu, x) 166 167 #ifdef MBUFTRACE 168 static struct mowner carp_proto_mowner_rx = MOWNER_INIT("carp", "rx"); 169 static struct mowner carp_proto_mowner_tx = MOWNER_INIT("carp", "tx"); 170 static struct mowner carp_proto6_mowner_rx = MOWNER_INIT("carp6", "rx"); 171 static struct mowner carp_proto6_mowner_tx = MOWNER_INIT("carp6", "tx"); 172 #endif 173 174 struct carp_if { 175 TAILQ_HEAD(, carp_softc) vhif_vrs; 176 int vhif_nvrs; 177 178 struct ifnet *vhif_ifp; 179 }; 180 181 #define CARP_LOG(sc, s) \ 182 if (carp_opts[CARPCTL_LOG]) { \ 183 if (sc) \ 184 log(LOG_INFO, "%s: ", \ 185 (sc)->sc_if.if_xname); \ 186 else \ 187 log(LOG_INFO, "carp: "); \ 188 addlog s; \ 189 addlog("\n"); \ 190 } 191 192 static void carp_hmac_prepare(struct carp_softc *); 193 static void carp_hmac_generate(struct carp_softc *, u_int32_t *, 194 unsigned char *); 195 static int carp_hmac_verify(struct carp_softc *, u_int32_t *, 196 unsigned char *); 197 static void carp_setroute(struct carp_softc *, int); 198 static void carp_proto_input_c(struct mbuf *, struct carp_header *, 199 sa_family_t); 200 static void carpdetach(struct carp_softc *); 201 static int carp_prepare_ad(struct mbuf *, struct carp_softc *, 202 struct carp_header *); 203 static void carp_send_ad_all(void); 204 static void carp_send_ad(void *); 205 static void carp_send_arp(struct carp_softc *); 206 static void carp_master_down(void *); 207 static int carp_ioctl(struct ifnet *, u_long, void *); 208 static void carp_start(struct ifnet *); 209 static void carp_setrun(struct carp_softc *, sa_family_t); 210 static void carp_set_state(struct carp_softc *, int); 211 static int carp_addrcount(struct carp_if *, struct in_ifaddr *, int); 212 enum { CARP_COUNT_MASTER, CARP_COUNT_RUNNING }; 213 214 static void carp_multicast_cleanup(struct carp_softc *); 215 static int carp_set_ifp(struct carp_softc *, struct ifnet *); 216 static void carp_set_enaddr(struct carp_softc *); 217 #if 0 218 static void carp_addr_updated(void *); 219 #endif 220 static u_int32_t carp_hash(struct carp_softc *, u_char *); 221 static int carp_set_addr(struct carp_softc *, struct sockaddr_in *); 222 static int carp_join_multicast(struct carp_softc *); 223 #ifdef INET6 224 static void carp_send_na(struct carp_softc *); 225 static int carp_set_addr6(struct carp_softc *, struct sockaddr_in6 *); 226 static int carp_join_multicast6(struct carp_softc *); 227 #endif 228 static int carp_clone_create(struct if_clone *, int); 229 static int carp_clone_destroy(struct ifnet *); 230 static int carp_ether_addmulti(struct carp_softc *, struct ifreq *); 231 static int carp_ether_delmulti(struct carp_softc *, struct ifreq *); 232 static void carp_ether_purgemulti(struct carp_softc *); 233 234 static void sysctl_net_inet_carp_setup(struct sysctllog **); 235 236 struct if_clone carp_cloner = 237 IF_CLONE_INITIALIZER("carp", carp_clone_create, carp_clone_destroy); 238 239 static __inline u_int16_t 240 carp_cksum(struct mbuf *m, int len) 241 { 242 return (in_cksum(m, len)); 243 } 244 245 static void 246 carp_hmac_prepare(struct carp_softc *sc) 247 { 248 u_int8_t carp_version = CARP_VERSION, type = CARP_ADVERTISEMENT; 249 u_int8_t vhid = sc->sc_vhid & 0xff; 250 SHA1_CTX sha1ctx; 251 u_int32_t kmd[5]; 252 struct ifaddr *ifa; 253 int i, found; 254 struct in_addr last, cur, in; 255 #ifdef INET6 256 struct in6_addr last6, cur6, in6; 257 #endif /* INET6 */ 258 259 /* compute ipad from key */ 260 memset(sc->sc_pad, 0, sizeof(sc->sc_pad)); 261 memcpy(sc->sc_pad, sc->sc_key, sizeof(sc->sc_key)); 262 for (i = 0; i < sizeof(sc->sc_pad); i++) 263 sc->sc_pad[i] ^= 0x36; 264 265 /* precompute first part of inner hash */ 266 SHA1Init(&sc->sc_sha1); 267 SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad)); 268 SHA1Update(&sc->sc_sha1, (void *)&carp_version, sizeof(carp_version)); 269 SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type)); 270 271 /* generate a key for the arpbalance hash, before the vhid is hashed */ 272 memcpy(&sha1ctx, &sc->sc_sha1, sizeof(sha1ctx)); 273 SHA1Final((unsigned char *)kmd, &sha1ctx); 274 sc->sc_hashkey[0] = kmd[0] ^ kmd[1]; 275 sc->sc_hashkey[1] = kmd[2] ^ kmd[3]; 276 277 /* the rest of the precomputation */ 278 SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid)); 279 280 /* Hash the addresses from smallest to largest, not interface order */ 281 #ifdef INET 282 cur.s_addr = 0; 283 do { 284 found = 0; 285 last = cur; 286 cur.s_addr = 0xffffffff; 287 IFADDR_FOREACH(ifa, &sc->sc_if) { 288 in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr; 289 if (ifa->ifa_addr->sa_family == AF_INET && 290 ntohl(in.s_addr) > ntohl(last.s_addr) && 291 ntohl(in.s_addr) < ntohl(cur.s_addr)) { 292 cur.s_addr = in.s_addr; 293 found++; 294 } 295 } 296 if (found) 297 SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur)); 298 } while (found); 299 #endif /* INET */ 300 301 #ifdef INET6 302 memset(&cur6, 0x00, sizeof(cur6)); 303 do { 304 found = 0; 305 last6 = cur6; 306 memset(&cur6, 0xff, sizeof(cur6)); 307 IFADDR_FOREACH(ifa, &sc->sc_if) { 308 in6 = ifatoia6(ifa)->ia_addr.sin6_addr; 309 if (IN6_IS_ADDR_LINKLOCAL(&in6)) 310 in6.s6_addr16[1] = 0; 311 if (ifa->ifa_addr->sa_family == AF_INET6 && 312 memcmp(&in6, &last6, sizeof(in6)) > 0 && 313 memcmp(&in6, &cur6, sizeof(in6)) < 0) { 314 cur6 = in6; 315 found++; 316 } 317 } 318 if (found) 319 SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6)); 320 } while (found); 321 #endif /* INET6 */ 322 323 /* convert ipad to opad */ 324 for (i = 0; i < sizeof(sc->sc_pad); i++) 325 sc->sc_pad[i] ^= 0x36 ^ 0x5c; 326 } 327 328 static void 329 carp_hmac_generate(struct carp_softc *sc, u_int32_t counter[2], 330 unsigned char md[20]) 331 { 332 SHA1_CTX sha1ctx; 333 334 /* fetch first half of inner hash */ 335 memcpy(&sha1ctx, &sc->sc_sha1, sizeof(sha1ctx)); 336 337 SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter)); 338 SHA1Final(md, &sha1ctx); 339 340 /* outer hash */ 341 SHA1Init(&sha1ctx); 342 SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad)); 343 SHA1Update(&sha1ctx, md, 20); 344 SHA1Final(md, &sha1ctx); 345 } 346 347 static int 348 carp_hmac_verify(struct carp_softc *sc, u_int32_t counter[2], 349 unsigned char md[20]) 350 { 351 unsigned char md2[20]; 352 353 carp_hmac_generate(sc, counter, md2); 354 355 return (memcmp(md, md2, sizeof(md2))); 356 } 357 358 static void 359 carp_setroute(struct carp_softc *sc, int cmd) 360 { 361 struct ifaddr *ifa; 362 int s; 363 364 KERNEL_LOCK(1, NULL); 365 s = splsoftnet(); 366 IFADDR_FOREACH(ifa, &sc->sc_if) { 367 switch (ifa->ifa_addr->sa_family) { 368 case AF_INET: { 369 int count = 0; 370 struct rtentry *rt; 371 int hr_otherif, nr_ourif; 372 373 /* 374 * Avoid screwing with the routes if there are other 375 * carp interfaces which are master and have the same 376 * address. 377 */ 378 if (sc->sc_carpdev != NULL && 379 sc->sc_carpdev->if_carp != NULL) { 380 count = carp_addrcount( 381 (struct carp_if *)sc->sc_carpdev->if_carp, 382 ifatoia(ifa), CARP_COUNT_MASTER); 383 if ((cmd == RTM_ADD && count != 1) || 384 (cmd == RTM_DELETE && count != 0)) 385 continue; 386 } 387 388 /* Remove the existing host route, if any */ 389 rtrequest(RTM_DELETE, ifa->ifa_addr, 390 ifa->ifa_addr, ifa->ifa_netmask, 391 RTF_HOST, NULL); 392 393 rt = NULL; 394 (void)rtrequest(RTM_GET, ifa->ifa_addr, ifa->ifa_addr, 395 ifa->ifa_netmask, RTF_HOST, &rt); 396 hr_otherif = (rt && rt->rt_ifp != &sc->sc_if && 397 rt->rt_flags & (RTF_CLONING|RTF_CLONED)); 398 if (rt != NULL) { 399 rtfree(rt); 400 rt = NULL; 401 } 402 403 /* Check for a network route on our interface */ 404 405 rt = NULL; 406 (void)rtrequest(RTM_GET, ifa->ifa_addr, ifa->ifa_addr, 407 ifa->ifa_netmask, 0, &rt); 408 nr_ourif = (rt && rt->rt_ifp == &sc->sc_if); 409 410 switch (cmd) { 411 case RTM_ADD: 412 if (hr_otherif) { 413 ifa->ifa_rtrequest = NULL; 414 ifa->ifa_flags &= ~RTF_CLONING; 415 416 rtrequest(RTM_ADD, ifa->ifa_addr, 417 ifa->ifa_addr, ifa->ifa_netmask, 418 RTF_UP | RTF_HOST, NULL); 419 } 420 if (!hr_otherif || nr_ourif || !rt) { 421 if (nr_ourif && !(rt->rt_flags & 422 RTF_CLONING)) 423 rtrequest(RTM_DELETE, 424 ifa->ifa_addr, 425 ifa->ifa_addr, 426 ifa->ifa_netmask, 0, NULL); 427 428 ifa->ifa_rtrequest = arp_rtrequest; 429 ifa->ifa_flags |= RTF_CLONING; 430 431 if (rtrequest(RTM_ADD, ifa->ifa_addr, 432 ifa->ifa_addr, ifa->ifa_netmask, 0, 433 NULL) == 0) 434 ifa->ifa_flags |= IFA_ROUTE; 435 } 436 break; 437 case RTM_DELETE: 438 break; 439 default: 440 break; 441 } 442 if (rt != NULL) { 443 rtfree(rt); 444 rt = NULL; 445 } 446 break; 447 } 448 449 #ifdef INET6 450 case AF_INET6: 451 if (cmd == RTM_ADD) 452 in6_ifaddlocal(ifa); 453 else 454 in6_ifremlocal(ifa); 455 break; 456 #endif /* INET6 */ 457 default: 458 break; 459 } 460 } 461 splx(s); 462 KERNEL_UNLOCK_ONE(NULL); 463 } 464 465 /* 466 * process input packet. 467 * we have rearranged checks order compared to the rfc, 468 * but it seems more efficient this way or not possible otherwise. 469 */ 470 void 471 carp_proto_input(struct mbuf *m, ...) 472 { 473 struct ip *ip = mtod(m, struct ip *); 474 struct carp_softc *sc = NULL; 475 struct carp_header *ch; 476 int iplen, len; 477 va_list ap; 478 479 va_start(ap, m); 480 va_end(ap); 481 482 CARP_STATINC(CARP_STAT_IPACKETS); 483 MCLAIM(m, &carp_proto_mowner_rx); 484 485 if (!carp_opts[CARPCTL_ALLOW]) { 486 m_freem(m); 487 return; 488 } 489 490 /* check if received on a valid carp interface */ 491 if (m->m_pkthdr.rcvif->if_type != IFT_CARP) { 492 CARP_STATINC(CARP_STAT_BADIF); 493 CARP_LOG(sc, ("packet received on non-carp interface: %s", 494 m->m_pkthdr.rcvif->if_xname)); 495 m_freem(m); 496 return; 497 } 498 499 /* verify that the IP TTL is 255. */ 500 if (ip->ip_ttl != CARP_DFLTTL) { 501 CARP_STATINC(CARP_STAT_BADTTL); 502 CARP_LOG(sc, ("received ttl %d != %d on %s", ip->ip_ttl, 503 CARP_DFLTTL, m->m_pkthdr.rcvif->if_xname)); 504 m_freem(m); 505 return; 506 } 507 508 /* 509 * verify that the received packet length is 510 * equal to the CARP header 511 */ 512 iplen = ip->ip_hl << 2; 513 len = iplen + sizeof(*ch); 514 if (len > m->m_pkthdr.len) { 515 CARP_STATINC(CARP_STAT_BADLEN); 516 CARP_LOG(sc, ("packet too short %d on %s", m->m_pkthdr.len, 517 m->m_pkthdr.rcvif->if_xname)); 518 m_freem(m); 519 return; 520 } 521 522 if ((m = m_pullup(m, len)) == NULL) { 523 CARP_STATINC(CARP_STAT_HDROPS); 524 return; 525 } 526 ip = mtod(m, struct ip *); 527 ch = (struct carp_header *)((char *)ip + iplen); 528 /* verify the CARP checksum */ 529 m->m_data += iplen; 530 if (carp_cksum(m, len - iplen)) { 531 CARP_STATINC(CARP_STAT_BADSUM); 532 CARP_LOG(sc, ("checksum failed on %s", 533 m->m_pkthdr.rcvif->if_xname)); 534 m_freem(m); 535 return; 536 } 537 m->m_data -= iplen; 538 539 carp_proto_input_c(m, ch, AF_INET); 540 } 541 542 #ifdef INET6 543 int 544 carp6_proto_input(struct mbuf **mp, int *offp, int proto) 545 { 546 struct mbuf *m = *mp; 547 struct carp_softc *sc = NULL; 548 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 549 struct carp_header *ch; 550 u_int len; 551 552 CARP_STATINC(CARP_STAT_IPACKETS6); 553 MCLAIM(m, &carp_proto6_mowner_rx); 554 555 if (!carp_opts[CARPCTL_ALLOW]) { 556 m_freem(m); 557 return (IPPROTO_DONE); 558 } 559 560 /* check if received on a valid carp interface */ 561 if (m->m_pkthdr.rcvif->if_type != IFT_CARP) { 562 CARP_STATINC(CARP_STAT_BADIF); 563 CARP_LOG(sc, ("packet received on non-carp interface: %s", 564 m->m_pkthdr.rcvif->if_xname)); 565 m_freem(m); 566 return (IPPROTO_DONE); 567 } 568 569 /* verify that the IP TTL is 255 */ 570 if (ip6->ip6_hlim != CARP_DFLTTL) { 571 CARP_STATINC(CARP_STAT_BADTTL); 572 CARP_LOG(sc, ("received ttl %d != %d on %s", ip6->ip6_hlim, 573 CARP_DFLTTL, m->m_pkthdr.rcvif->if_xname)); 574 m_freem(m); 575 return (IPPROTO_DONE); 576 } 577 578 /* verify that we have a complete carp packet */ 579 len = m->m_len; 580 IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch)); 581 if (ch == NULL) { 582 CARP_STATINC(CARP_STAT_BADLEN); 583 CARP_LOG(sc, ("packet size %u too small", len)); 584 return (IPPROTO_DONE); 585 } 586 587 588 /* verify the CARP checksum */ 589 m->m_data += *offp; 590 if (carp_cksum(m, sizeof(*ch))) { 591 CARP_STATINC(CARP_STAT_BADSUM); 592 CARP_LOG(sc, ("checksum failed, on %s", 593 m->m_pkthdr.rcvif->if_xname)); 594 m_freem(m); 595 return (IPPROTO_DONE); 596 } 597 m->m_data -= *offp; 598 599 carp_proto_input_c(m, ch, AF_INET6); 600 return (IPPROTO_DONE); 601 } 602 #endif /* INET6 */ 603 604 static void 605 carp_proto_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af) 606 { 607 struct carp_softc *sc; 608 u_int64_t tmp_counter; 609 struct timeval sc_tv, ch_tv; 610 611 TAILQ_FOREACH(sc, &((struct carp_if *) 612 m->m_pkthdr.rcvif->if_carpdev->if_carp)->vhif_vrs, sc_list) 613 if (sc->sc_vhid == ch->carp_vhid) 614 break; 615 616 if (!sc || (sc->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) != 617 (IFF_UP|IFF_RUNNING)) { 618 CARP_STATINC(CARP_STAT_BADVHID); 619 m_freem(m); 620 return; 621 } 622 623 /* 624 * Check if our own advertisement was duplicated 625 * from a non simplex interface. 626 * XXX If there is no address on our physical interface 627 * there is no way to distinguish our ads from the ones 628 * another carp host might have sent us. 629 */ 630 if ((sc->sc_carpdev->if_flags & IFF_SIMPLEX) == 0) { 631 struct sockaddr sa; 632 struct ifaddr *ifa; 633 634 memset(&sa, 0, sizeof(sa)); 635 sa.sa_family = af; 636 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev); 637 638 if (ifa && af == AF_INET) { 639 struct ip *ip = mtod(m, struct ip *); 640 if (ip->ip_src.s_addr == 641 ifatoia(ifa)->ia_addr.sin_addr.s_addr) { 642 m_freem(m); 643 return; 644 } 645 } 646 #ifdef INET6 647 if (ifa && af == AF_INET6) { 648 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 649 struct in6_addr in6_src, in6_found; 650 651 in6_src = ip6->ip6_src; 652 in6_found = ifatoia6(ifa)->ia_addr.sin6_addr; 653 if (IN6_IS_ADDR_LINKLOCAL(&in6_src)) 654 in6_src.s6_addr16[1] = 0; 655 if (IN6_IS_ADDR_LINKLOCAL(&in6_found)) 656 in6_found.s6_addr16[1] = 0; 657 if (IN6_ARE_ADDR_EQUAL(&in6_src, &in6_found)) { 658 m_freem(m); 659 return; 660 } 661 } 662 #endif /* INET6 */ 663 } 664 665 nanotime(&sc->sc_if.if_lastchange); 666 sc->sc_if.if_ipackets++; 667 sc->sc_if.if_ibytes += m->m_pkthdr.len; 668 669 /* verify the CARP version. */ 670 if (ch->carp_version != CARP_VERSION) { 671 CARP_STATINC(CARP_STAT_BADVER); 672 sc->sc_if.if_ierrors++; 673 CARP_LOG(sc, ("invalid version %d != %d", 674 ch->carp_version, CARP_VERSION)); 675 m_freem(m); 676 return; 677 } 678 679 /* verify the hash */ 680 if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) { 681 CARP_STATINC(CARP_STAT_BADAUTH); 682 sc->sc_if.if_ierrors++; 683 CARP_LOG(sc, ("incorrect hash")); 684 m_freem(m); 685 return; 686 } 687 688 tmp_counter = ntohl(ch->carp_counter[0]); 689 tmp_counter = tmp_counter<<32; 690 tmp_counter += ntohl(ch->carp_counter[1]); 691 692 /* XXX Replay protection goes here */ 693 694 sc->sc_init_counter = 0; 695 sc->sc_counter = tmp_counter; 696 697 698 sc_tv.tv_sec = sc->sc_advbase; 699 if (carp_suppress_preempt && sc->sc_advskew < 240) 700 sc_tv.tv_usec = 240 * 1000000 / 256; 701 else 702 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256; 703 ch_tv.tv_sec = ch->carp_advbase; 704 ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256; 705 706 switch (sc->sc_state) { 707 case INIT: 708 break; 709 case MASTER: 710 /* 711 * If we receive an advertisement from a backup who's going to 712 * be more frequent than us, go into BACKUP state. 713 */ 714 if (timercmp(&sc_tv, &ch_tv, >) || 715 timercmp(&sc_tv, &ch_tv, ==)) { 716 callout_stop(&sc->sc_ad_tmo); 717 CARP_LOG(sc, ("MASTER -> BACKUP (more frequent advertisement received)")); 718 carp_set_state(sc, BACKUP); 719 carp_setrun(sc, 0); 720 carp_setroute(sc, RTM_DELETE); 721 } 722 break; 723 case BACKUP: 724 /* 725 * If we're pre-empting masters who advertise slower than us, 726 * and this one claims to be slower, treat him as down. 727 */ 728 if (carp_opts[CARPCTL_PREEMPT] && timercmp(&sc_tv, &ch_tv, <)) { 729 CARP_LOG(sc, ("BACKUP -> MASTER (preempting a slower master)")); 730 carp_master_down(sc); 731 break; 732 } 733 734 /* 735 * If the master is going to advertise at such a low frequency 736 * that he's guaranteed to time out, we'd might as well just 737 * treat him as timed out now. 738 */ 739 sc_tv.tv_sec = sc->sc_advbase * 3; 740 if (timercmp(&sc_tv, &ch_tv, <)) { 741 CARP_LOG(sc, ("BACKUP -> MASTER (master timed out)")); 742 carp_master_down(sc); 743 break; 744 } 745 746 /* 747 * Otherwise, we reset the counter and wait for the next 748 * advertisement. 749 */ 750 carp_setrun(sc, af); 751 break; 752 } 753 754 m_freem(m); 755 return; 756 } 757 758 /* 759 * Interface side of the CARP implementation. 760 */ 761 762 /* ARGSUSED */ 763 void 764 carpattach(int n) 765 { 766 if_clone_attach(&carp_cloner); 767 768 carpstat_percpu = percpu_alloc(sizeof(uint64_t) * CARP_NSTATS); 769 } 770 771 static int 772 carp_clone_create(struct if_clone *ifc, int unit) 773 { 774 extern int ifqmaxlen; 775 struct carp_softc *sc; 776 struct ifnet *ifp; 777 778 sc = malloc(sizeof(*sc), M_DEVBUF, M_NOWAIT|M_ZERO); 779 if (!sc) 780 return (ENOMEM); 781 782 sc->sc_suppress = 0; 783 sc->sc_advbase = CARP_DFLTINTV; 784 sc->sc_vhid = -1; /* required setting */ 785 sc->sc_advskew = 0; 786 sc->sc_init_counter = 1; 787 sc->sc_naddrs = sc->sc_naddrs6 = 0; 788 #ifdef INET6 789 sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL; 790 #endif /* INET6 */ 791 792 callout_init(&sc->sc_ad_tmo, 0); 793 callout_init(&sc->sc_md_tmo, 0); 794 callout_init(&sc->sc_md6_tmo, 0); 795 796 callout_setfunc(&sc->sc_ad_tmo, carp_send_ad, sc); 797 callout_setfunc(&sc->sc_md_tmo, carp_master_down, sc); 798 callout_setfunc(&sc->sc_md6_tmo, carp_master_down, sc); 799 800 LIST_INIT(&sc->carp_mc_listhead); 801 ifp = &sc->sc_if; 802 ifp->if_softc = sc; 803 snprintf(ifp->if_xname, sizeof ifp->if_xname, "%s%d", ifc->ifc_name, 804 unit); 805 ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST; 806 ifp->if_ioctl = carp_ioctl; 807 ifp->if_start = carp_start; 808 ifp->if_output = carp_output; 809 ifp->if_type = IFT_CARP; 810 ifp->if_addrlen = ETHER_ADDR_LEN; 811 ifp->if_hdrlen = ETHER_HDR_LEN; 812 ifp->if_mtu = ETHERMTU; 813 IFQ_SET_MAXLEN(&ifp->if_snd, ifqmaxlen); 814 IFQ_SET_READY(&ifp->if_snd); 815 if_attach(ifp); 816 817 if_alloc_sadl(ifp); 818 ifp->if_broadcastaddr = etherbroadcastaddr; 819 carp_set_enaddr(sc); 820 LIST_INIT(&sc->sc_ac.ec_multiaddrs); 821 bpf_attach(ifp, DLT_EN10MB, ETHER_HDR_LEN); 822 #ifdef MBUFTRACE 823 strlcpy(sc->sc_ac.ec_tx_mowner.mo_name, ifp->if_xname, 824 sizeof(sc->sc_ac.ec_tx_mowner.mo_name)); 825 strlcpy(sc->sc_ac.ec_tx_mowner.mo_descr, "tx", 826 sizeof(sc->sc_ac.ec_tx_mowner.mo_descr)); 827 strlcpy(sc->sc_ac.ec_rx_mowner.mo_name, ifp->if_xname, 828 sizeof(sc->sc_ac.ec_rx_mowner.mo_name)); 829 strlcpy(sc->sc_ac.ec_rx_mowner.mo_descr, "rx", 830 sizeof(sc->sc_ac.ec_rx_mowner.mo_descr)); 831 MOWNER_ATTACH(&sc->sc_ac.ec_tx_mowner); 832 MOWNER_ATTACH(&sc->sc_ac.ec_rx_mowner); 833 ifp->if_mowner = &sc->sc_ac.ec_tx_mowner; 834 #endif 835 return (0); 836 } 837 838 static int 839 carp_clone_destroy(struct ifnet *ifp) 840 { 841 struct carp_softc *sc = ifp->if_softc; 842 843 carpdetach(ifp->if_softc); 844 ether_ifdetach(ifp); 845 if_detach(ifp); 846 callout_destroy(&sc->sc_ad_tmo); 847 callout_destroy(&sc->sc_md_tmo); 848 callout_destroy(&sc->sc_md6_tmo); 849 free(ifp->if_softc, M_DEVBUF); 850 851 return (0); 852 } 853 854 static void 855 carpdetach(struct carp_softc *sc) 856 { 857 struct carp_if *cif; 858 int s; 859 860 callout_stop(&sc->sc_ad_tmo); 861 callout_stop(&sc->sc_md_tmo); 862 callout_stop(&sc->sc_md6_tmo); 863 864 if (sc->sc_suppress) 865 carp_suppress_preempt--; 866 sc->sc_suppress = 0; 867 868 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) 869 carp_suppress_preempt--; 870 sc->sc_sendad_errors = 0; 871 872 carp_set_state(sc, INIT); 873 sc->sc_if.if_flags &= ~IFF_UP; 874 carp_setrun(sc, 0); 875 carp_multicast_cleanup(sc); 876 877 KERNEL_LOCK(1, NULL); 878 s = splnet(); 879 if (sc->sc_carpdev != NULL) { 880 /* XXX linkstatehook removal */ 881 cif = (struct carp_if *)sc->sc_carpdev->if_carp; 882 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list); 883 if (!--cif->vhif_nvrs) { 884 ifpromisc(sc->sc_carpdev, 0); 885 sc->sc_carpdev->if_carp = NULL; 886 free(cif, M_IFADDR); 887 } 888 } 889 sc->sc_carpdev = NULL; 890 splx(s); 891 KERNEL_UNLOCK_ONE(NULL); 892 } 893 894 /* Detach an interface from the carp. */ 895 void 896 carp_ifdetach(struct ifnet *ifp) 897 { 898 struct carp_softc *sc, *nextsc; 899 struct carp_if *cif = (struct carp_if *)ifp->if_carp; 900 901 for (sc = TAILQ_FIRST(&cif->vhif_vrs); sc; sc = nextsc) { 902 nextsc = TAILQ_NEXT(sc, sc_list); 903 carpdetach(sc); 904 } 905 } 906 907 static int 908 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, 909 struct carp_header *ch) 910 { 911 if (sc->sc_init_counter) { 912 /* this could also be seconds since unix epoch */ 913 sc->sc_counter = cprng_fast64(); 914 } else 915 sc->sc_counter++; 916 917 ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff); 918 ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff); 919 920 carp_hmac_generate(sc, ch->carp_counter, ch->carp_md); 921 922 return (0); 923 } 924 925 static void 926 carp_send_ad_all(void) 927 { 928 struct ifnet *ifp; 929 struct carp_if *cif; 930 struct carp_softc *vh; 931 932 IFNET_FOREACH(ifp) { 933 if (ifp->if_carp == NULL || ifp->if_type == IFT_CARP) 934 continue; 935 936 cif = (struct carp_if *)ifp->if_carp; 937 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 938 if ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 939 (IFF_UP|IFF_RUNNING) && vh->sc_state == MASTER) 940 carp_send_ad(vh); 941 } 942 } 943 } 944 945 946 static void 947 carp_send_ad(void *v) 948 { 949 struct carp_header ch; 950 struct timeval tv; 951 struct carp_softc *sc = v; 952 struct carp_header *ch_ptr; 953 struct mbuf *m; 954 int error, len, advbase, advskew, s; 955 struct ifaddr *ifa; 956 struct sockaddr sa; 957 958 KERNEL_LOCK(1, NULL); 959 s = splsoftnet(); 960 961 advbase = advskew = 0; /* Sssssh compiler */ 962 if (sc->sc_carpdev == NULL) { 963 sc->sc_if.if_oerrors++; 964 goto retry_later; 965 } 966 967 /* bow out if we've gone to backup (the carp interface is going down) */ 968 if (sc->sc_bow_out) { 969 sc->sc_bow_out = 0; 970 advbase = 255; 971 advskew = 255; 972 } else { 973 advbase = sc->sc_advbase; 974 if (!carp_suppress_preempt || sc->sc_advskew > 240) 975 advskew = sc->sc_advskew; 976 else 977 advskew = 240; 978 tv.tv_sec = advbase; 979 tv.tv_usec = advskew * 1000000 / 256; 980 } 981 982 ch.carp_version = CARP_VERSION; 983 ch.carp_type = CARP_ADVERTISEMENT; 984 ch.carp_vhid = sc->sc_vhid; 985 ch.carp_advbase = advbase; 986 ch.carp_advskew = advskew; 987 ch.carp_authlen = 7; /* XXX DEFINE */ 988 ch.carp_pad1 = 0; /* must be zero */ 989 ch.carp_cksum = 0; 990 991 992 #ifdef INET 993 if (sc->sc_naddrs) { 994 struct ip *ip; 995 996 MGETHDR(m, M_DONTWAIT, MT_HEADER); 997 if (m == NULL) { 998 sc->sc_if.if_oerrors++; 999 CARP_STATINC(CARP_STAT_ONOMEM); 1000 /* XXX maybe less ? */ 1001 goto retry_later; 1002 } 1003 MCLAIM(m, &carp_proto_mowner_tx); 1004 len = sizeof(*ip) + sizeof(ch); 1005 m->m_pkthdr.len = len; 1006 m->m_pkthdr.rcvif = NULL; 1007 m->m_len = len; 1008 MH_ALIGN(m, m->m_len); 1009 m->m_flags |= M_MCAST; 1010 ip = mtod(m, struct ip *); 1011 ip->ip_v = IPVERSION; 1012 ip->ip_hl = sizeof(*ip) >> 2; 1013 ip->ip_tos = IPTOS_LOWDELAY; 1014 ip->ip_len = htons(len); 1015 ip->ip_id = 0; /* no need for id, we don't support fragments */ 1016 ip->ip_off = htons(IP_DF); 1017 ip->ip_ttl = CARP_DFLTTL; 1018 ip->ip_p = IPPROTO_CARP; 1019 ip->ip_sum = 0; 1020 1021 memset(&sa, 0, sizeof(sa)); 1022 sa.sa_family = AF_INET; 1023 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev); 1024 if (ifa == NULL) 1025 ip->ip_src.s_addr = 0; 1026 else 1027 ip->ip_src.s_addr = 1028 ifatoia(ifa)->ia_addr.sin_addr.s_addr; 1029 ip->ip_dst.s_addr = INADDR_CARP_GROUP; 1030 1031 ch_ptr = (struct carp_header *)(&ip[1]); 1032 memcpy(ch_ptr, &ch, sizeof(ch)); 1033 if (carp_prepare_ad(m, sc, ch_ptr)) 1034 goto retry_later; 1035 1036 m->m_data += sizeof(*ip); 1037 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip)); 1038 m->m_data -= sizeof(*ip); 1039 1040 nanotime(&sc->sc_if.if_lastchange); 1041 sc->sc_if.if_opackets++; 1042 sc->sc_if.if_obytes += len; 1043 CARP_STATINC(CARP_STAT_OPACKETS); 1044 1045 error = ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, 1046 NULL); 1047 if (error) { 1048 if (error == ENOBUFS) 1049 CARP_STATINC(CARP_STAT_ONOMEM); 1050 else 1051 CARP_LOG(sc, ("ip_output failed: %d", error)); 1052 sc->sc_if.if_oerrors++; 1053 if (sc->sc_sendad_errors < INT_MAX) 1054 sc->sc_sendad_errors++; 1055 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) { 1056 carp_suppress_preempt++; 1057 if (carp_suppress_preempt == 1) 1058 carp_send_ad_all(); 1059 } 1060 sc->sc_sendad_success = 0; 1061 } else { 1062 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) { 1063 if (++sc->sc_sendad_success >= 1064 CARP_SENDAD_MIN_SUCCESS) { 1065 carp_suppress_preempt--; 1066 sc->sc_sendad_errors = 0; 1067 } 1068 } else 1069 sc->sc_sendad_errors = 0; 1070 } 1071 } 1072 #endif /* INET */ 1073 #ifdef INET6 1074 if (sc->sc_naddrs6) { 1075 struct ip6_hdr *ip6; 1076 1077 MGETHDR(m, M_DONTWAIT, MT_HEADER); 1078 if (m == NULL) { 1079 sc->sc_if.if_oerrors++; 1080 CARP_STATINC(CARP_STAT_ONOMEM); 1081 /* XXX maybe less ? */ 1082 goto retry_later; 1083 } 1084 MCLAIM(m, &carp_proto6_mowner_tx); 1085 len = sizeof(*ip6) + sizeof(ch); 1086 m->m_pkthdr.len = len; 1087 m->m_pkthdr.rcvif = NULL; 1088 m->m_len = len; 1089 MH_ALIGN(m, m->m_len); 1090 m->m_flags |= M_MCAST; 1091 ip6 = mtod(m, struct ip6_hdr *); 1092 memset(ip6, 0, sizeof(*ip6)); 1093 ip6->ip6_vfc |= IPV6_VERSION; 1094 ip6->ip6_hlim = CARP_DFLTTL; 1095 ip6->ip6_nxt = IPPROTO_CARP; 1096 1097 /* set the source address */ 1098 memset(&sa, 0, sizeof(sa)); 1099 sa.sa_family = AF_INET6; 1100 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev); 1101 if (ifa == NULL) /* This should never happen with IPv6 */ 1102 memset(&ip6->ip6_src, 0, sizeof(struct in6_addr)); 1103 else 1104 bcopy(ifatoia6(ifa)->ia_addr.sin6_addr.s6_addr, 1105 &ip6->ip6_src, sizeof(struct in6_addr)); 1106 /* set the multicast destination */ 1107 1108 ip6->ip6_dst.s6_addr16[0] = htons(0xff02); 1109 ip6->ip6_dst.s6_addr8[15] = 0x12; 1110 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) { 1111 sc->sc_if.if_oerrors++; 1112 m_freem(m); 1113 CARP_LOG(sc, ("in6_setscope failed")); 1114 goto retry_later; 1115 } 1116 1117 ch_ptr = (struct carp_header *)(&ip6[1]); 1118 memcpy(ch_ptr, &ch, sizeof(ch)); 1119 if (carp_prepare_ad(m, sc, ch_ptr)) 1120 goto retry_later; 1121 1122 m->m_data += sizeof(*ip6); 1123 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6)); 1124 m->m_data -= sizeof(*ip6); 1125 1126 nanotime(&sc->sc_if.if_lastchange); 1127 sc->sc_if.if_opackets++; 1128 sc->sc_if.if_obytes += len; 1129 CARP_STATINC(CARP_STAT_OPACKETS6); 1130 1131 error = ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL); 1132 if (error) { 1133 if (error == ENOBUFS) 1134 CARP_STATINC(CARP_STAT_ONOMEM); 1135 else 1136 CARP_LOG(sc, ("ip6_output failed: %d", error)); 1137 sc->sc_if.if_oerrors++; 1138 if (sc->sc_sendad_errors < INT_MAX) 1139 sc->sc_sendad_errors++; 1140 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) { 1141 carp_suppress_preempt++; 1142 if (carp_suppress_preempt == 1) 1143 carp_send_ad_all(); 1144 } 1145 sc->sc_sendad_success = 0; 1146 } else { 1147 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) { 1148 if (++sc->sc_sendad_success >= 1149 CARP_SENDAD_MIN_SUCCESS) { 1150 carp_suppress_preempt--; 1151 sc->sc_sendad_errors = 0; 1152 } 1153 } else 1154 sc->sc_sendad_errors = 0; 1155 } 1156 } 1157 #endif /* INET6 */ 1158 1159 retry_later: 1160 splx(s); 1161 KERNEL_UNLOCK_ONE(NULL); 1162 if (advbase != 255 || advskew != 255) 1163 callout_schedule(&sc->sc_ad_tmo, tvtohz(&tv)); 1164 } 1165 1166 /* 1167 * Broadcast a gratuitous ARP request containing 1168 * the virtual router MAC address for each IP address 1169 * associated with the virtual router. 1170 */ 1171 static void 1172 carp_send_arp(struct carp_softc *sc) 1173 { 1174 struct ifaddr *ifa; 1175 struct in_addr *in; 1176 int s; 1177 1178 KERNEL_LOCK(1, NULL); 1179 s = splsoftnet(); 1180 IFADDR_FOREACH(ifa, &sc->sc_if) { 1181 1182 if (ifa->ifa_addr->sa_family != AF_INET) 1183 continue; 1184 1185 in = &ifatoia(ifa)->ia_addr.sin_addr; 1186 arprequest(sc->sc_carpdev, in, in, CLLADDR(sc->sc_if.if_sadl)); 1187 } 1188 splx(s); 1189 KERNEL_UNLOCK_ONE(NULL); 1190 } 1191 1192 #ifdef INET6 1193 static void 1194 carp_send_na(struct carp_softc *sc) 1195 { 1196 struct ifaddr *ifa; 1197 struct in6_addr *in6; 1198 static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT; 1199 int s; 1200 1201 KERNEL_LOCK(1, NULL); 1202 s = splsoftnet(); 1203 1204 IFADDR_FOREACH(ifa, &sc->sc_if) { 1205 1206 if (ifa->ifa_addr->sa_family != AF_INET6) 1207 continue; 1208 1209 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr; 1210 nd6_na_output(sc->sc_carpdev, &mcast, in6, 1211 ND_NA_FLAG_OVERRIDE, 1, NULL); 1212 } 1213 splx(s); 1214 KERNEL_UNLOCK_ONE(NULL); 1215 } 1216 #endif /* INET6 */ 1217 1218 /* 1219 * Based on bridge_hash() in if_bridge.c 1220 */ 1221 #define mix(a,b,c) \ 1222 do { \ 1223 a -= b; a -= c; a ^= (c >> 13); \ 1224 b -= c; b -= a; b ^= (a << 8); \ 1225 c -= a; c -= b; c ^= (b >> 13); \ 1226 a -= b; a -= c; a ^= (c >> 12); \ 1227 b -= c; b -= a; b ^= (a << 16); \ 1228 c -= a; c -= b; c ^= (b >> 5); \ 1229 a -= b; a -= c; a ^= (c >> 3); \ 1230 b -= c; b -= a; b ^= (a << 10); \ 1231 c -= a; c -= b; c ^= (b >> 15); \ 1232 } while (0) 1233 1234 static u_int32_t 1235 carp_hash(struct carp_softc *sc, u_char *src) 1236 { 1237 u_int32_t a = 0x9e3779b9, b = sc->sc_hashkey[0], c = sc->sc_hashkey[1]; 1238 1239 c += sc->sc_key[3] << 24; 1240 c += sc->sc_key[2] << 16; 1241 c += sc->sc_key[1] << 8; 1242 c += sc->sc_key[0]; 1243 b += src[5] << 8; 1244 b += src[4]; 1245 a += src[3] << 24; 1246 a += src[2] << 16; 1247 a += src[1] << 8; 1248 a += src[0]; 1249 1250 mix(a, b, c); 1251 return (c); 1252 } 1253 1254 static int 1255 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type) 1256 { 1257 struct carp_softc *vh; 1258 struct ifaddr *ifa; 1259 int count = 0; 1260 1261 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1262 if ((type == CARP_COUNT_RUNNING && 1263 (vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1264 (IFF_UP|IFF_RUNNING)) || 1265 (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) { 1266 IFADDR_FOREACH(ifa, &vh->sc_if) { 1267 if (ifa->ifa_addr->sa_family == AF_INET && 1268 ia->ia_addr.sin_addr.s_addr == 1269 ifatoia(ifa)->ia_addr.sin_addr.s_addr) 1270 count++; 1271 } 1272 } 1273 } 1274 return (count); 1275 } 1276 1277 int 1278 carp_iamatch(struct in_ifaddr *ia, u_char *src, 1279 u_int32_t *count, u_int32_t index) 1280 { 1281 struct carp_softc *sc = ia->ia_ifp->if_softc; 1282 1283 if (carp_opts[CARPCTL_ARPBALANCE]) { 1284 /* 1285 * We use the source ip to decide which virtual host should 1286 * handle the request. If we're master of that virtual host, 1287 * then we respond, otherwise, just drop the arp packet on 1288 * the floor. 1289 */ 1290 1291 /* Count the elegible carp interfaces with this address */ 1292 if (*count == 0) 1293 *count = carp_addrcount( 1294 (struct carp_if *)ia->ia_ifp->if_carpdev->if_carp, 1295 ia, CARP_COUNT_RUNNING); 1296 1297 /* This should never happen, but... */ 1298 if (*count == 0) 1299 return (0); 1300 1301 if (carp_hash(sc, src) % *count == index - 1 && 1302 sc->sc_state == MASTER) { 1303 return (1); 1304 } 1305 } else { 1306 if (sc->sc_state == MASTER) 1307 return (1); 1308 } 1309 1310 return (0); 1311 } 1312 1313 #ifdef INET6 1314 struct ifaddr * 1315 carp_iamatch6(void *v, struct in6_addr *taddr) 1316 { 1317 struct carp_if *cif = v; 1318 struct carp_softc *vh; 1319 struct ifaddr *ifa; 1320 1321 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1322 IFADDR_FOREACH(ifa, &vh->sc_if) { 1323 if (IN6_ARE_ADDR_EQUAL(taddr, 1324 &ifatoia6(ifa)->ia_addr.sin6_addr) && 1325 ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1326 (IFF_UP|IFF_RUNNING)) && vh->sc_state == MASTER) 1327 return (ifa); 1328 } 1329 } 1330 1331 return (NULL); 1332 } 1333 #endif /* INET6 */ 1334 1335 struct ifnet * 1336 carp_ourether(void *v, struct ether_header *eh, u_char iftype, int src) 1337 { 1338 struct carp_if *cif = (struct carp_if *)v; 1339 struct carp_softc *vh; 1340 u_int8_t *ena; 1341 1342 if (src) 1343 ena = (u_int8_t *)&eh->ether_shost; 1344 else 1345 ena = (u_int8_t *)&eh->ether_dhost; 1346 1347 switch (iftype) { 1348 case IFT_ETHER: 1349 case IFT_FDDI: 1350 if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1) 1351 return (NULL); 1352 break; 1353 case IFT_ISO88025: 1354 if (ena[0] != 3 || ena[1] || ena[4] || ena[5]) 1355 return (NULL); 1356 break; 1357 default: 1358 return (NULL); 1359 break; 1360 } 1361 1362 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) 1363 if ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1364 (IFF_UP|IFF_RUNNING) && vh->sc_state == MASTER && 1365 !memcmp(ena, CLLADDR(vh->sc_if.if_sadl), 1366 ETHER_ADDR_LEN)) { 1367 return (&vh->sc_if); 1368 } 1369 1370 return (NULL); 1371 } 1372 1373 int 1374 carp_input(struct mbuf *m, u_int8_t *shost, u_int8_t *dhost, u_int16_t etype) 1375 { 1376 struct ether_header eh; 1377 struct carp_if *cif = (struct carp_if *)m->m_pkthdr.rcvif->if_carp; 1378 struct ifnet *ifp; 1379 1380 memcpy(&eh.ether_shost, shost, sizeof(eh.ether_shost)); 1381 memcpy(&eh.ether_dhost, dhost, sizeof(eh.ether_dhost)); 1382 eh.ether_type = etype; 1383 1384 if (m->m_flags & (M_BCAST|M_MCAST)) { 1385 struct carp_softc *vh; 1386 struct mbuf *m0; 1387 1388 /* 1389 * XXX Should really check the list of multicast addresses 1390 * for each CARP interface _before_ copying. 1391 */ 1392 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1393 m0 = m_copym(m, 0, M_COPYALL, M_DONTWAIT); 1394 if (m0 == NULL) 1395 continue; 1396 m0->m_pkthdr.rcvif = &vh->sc_if; 1397 ether_input(&vh->sc_if, m0); 1398 } 1399 return (1); 1400 } 1401 1402 ifp = carp_ourether(cif, &eh, m->m_pkthdr.rcvif->if_type, 0); 1403 if (ifp == NULL) { 1404 return (1); 1405 } 1406 1407 m->m_pkthdr.rcvif = ifp; 1408 1409 bpf_mtap(ifp, m); 1410 ifp->if_ipackets++; 1411 ether_input(ifp, m); 1412 return (0); 1413 } 1414 1415 static void 1416 carp_master_down(void *v) 1417 { 1418 struct carp_softc *sc = v; 1419 1420 switch (sc->sc_state) { 1421 case INIT: 1422 printf("%s: master_down event in INIT state\n", 1423 sc->sc_if.if_xname); 1424 break; 1425 case MASTER: 1426 break; 1427 case BACKUP: 1428 CARP_LOG(sc, ("INIT -> MASTER (preempting)")); 1429 carp_set_state(sc, MASTER); 1430 carp_send_ad(sc); 1431 carp_send_arp(sc); 1432 #ifdef INET6 1433 carp_send_na(sc); 1434 #endif /* INET6 */ 1435 carp_setrun(sc, 0); 1436 carp_setroute(sc, RTM_ADD); 1437 break; 1438 } 1439 } 1440 1441 /* 1442 * When in backup state, af indicates whether to reset the master down timer 1443 * for v4 or v6. If it's set to zero, reset the ones which are already pending. 1444 */ 1445 static void 1446 carp_setrun(struct carp_softc *sc, sa_family_t af) 1447 { 1448 struct timeval tv; 1449 1450 if (sc->sc_carpdev == NULL) { 1451 sc->sc_if.if_flags &= ~IFF_RUNNING; 1452 carp_set_state(sc, INIT); 1453 return; 1454 } 1455 1456 if (sc->sc_if.if_flags & IFF_UP && sc->sc_vhid > 0 && 1457 (sc->sc_naddrs || sc->sc_naddrs6) && !sc->sc_suppress) { 1458 sc->sc_if.if_flags |= IFF_RUNNING; 1459 } else { 1460 sc->sc_if.if_flags &= ~IFF_RUNNING; 1461 carp_setroute(sc, RTM_DELETE); 1462 return; 1463 } 1464 1465 switch (sc->sc_state) { 1466 case INIT: 1467 carp_set_state(sc, BACKUP); 1468 carp_setroute(sc, RTM_DELETE); 1469 carp_setrun(sc, 0); 1470 break; 1471 case BACKUP: 1472 callout_stop(&sc->sc_ad_tmo); 1473 tv.tv_sec = 3 * sc->sc_advbase; 1474 tv.tv_usec = sc->sc_advskew * 1000000 / 256; 1475 switch (af) { 1476 #ifdef INET 1477 case AF_INET: 1478 callout_schedule(&sc->sc_md_tmo, tvtohz(&tv)); 1479 break; 1480 #endif /* INET */ 1481 #ifdef INET6 1482 case AF_INET6: 1483 callout_schedule(&sc->sc_md6_tmo, tvtohz(&tv)); 1484 break; 1485 #endif /* INET6 */ 1486 default: 1487 if (sc->sc_naddrs) 1488 callout_schedule(&sc->sc_md_tmo, tvtohz(&tv)); 1489 if (sc->sc_naddrs6) 1490 callout_schedule(&sc->sc_md6_tmo, tvtohz(&tv)); 1491 break; 1492 } 1493 break; 1494 case MASTER: 1495 tv.tv_sec = sc->sc_advbase; 1496 tv.tv_usec = sc->sc_advskew * 1000000 / 256; 1497 callout_schedule(&sc->sc_ad_tmo, tvtohz(&tv)); 1498 break; 1499 } 1500 } 1501 1502 static void 1503 carp_multicast_cleanup(struct carp_softc *sc) 1504 { 1505 struct ip_moptions *imo = &sc->sc_imo; 1506 #ifdef INET6 1507 struct ip6_moptions *im6o = &sc->sc_im6o; 1508 #endif 1509 u_int16_t n = imo->imo_num_memberships; 1510 1511 /* Clean up our own multicast memberships */ 1512 while (n-- > 0) { 1513 if (imo->imo_membership[n] != NULL) { 1514 in_delmulti(imo->imo_membership[n]); 1515 imo->imo_membership[n] = NULL; 1516 } 1517 } 1518 imo->imo_num_memberships = 0; 1519 imo->imo_multicast_ifp = NULL; 1520 1521 #ifdef INET6 1522 while (!LIST_EMPTY(&im6o->im6o_memberships)) { 1523 struct in6_multi_mship *imm = 1524 LIST_FIRST(&im6o->im6o_memberships); 1525 1526 LIST_REMOVE(imm, i6mm_chain); 1527 in6_leavegroup(imm); 1528 } 1529 im6o->im6o_multicast_ifp = NULL; 1530 #endif 1531 1532 /* And any other multicast memberships */ 1533 carp_ether_purgemulti(sc); 1534 } 1535 1536 static int 1537 carp_set_ifp(struct carp_softc *sc, struct ifnet *ifp) 1538 { 1539 struct carp_if *cif, *ncif = NULL; 1540 struct carp_softc *vr, *after = NULL; 1541 int myself = 0, error = 0; 1542 int s; 1543 1544 if (ifp == sc->sc_carpdev) 1545 return (0); 1546 1547 if (ifp != NULL) { 1548 if ((ifp->if_flags & IFF_MULTICAST) == 0) 1549 return (EADDRNOTAVAIL); 1550 1551 if (ifp->if_type == IFT_CARP) 1552 return (EINVAL); 1553 1554 if (ifp->if_carp == NULL) { 1555 ncif = malloc(sizeof(*cif), M_IFADDR, M_NOWAIT); 1556 if (ncif == NULL) 1557 return (ENOBUFS); 1558 if ((error = ifpromisc(ifp, 1))) { 1559 free(ncif, M_IFADDR); 1560 return (error); 1561 } 1562 1563 ncif->vhif_ifp = ifp; 1564 TAILQ_INIT(&ncif->vhif_vrs); 1565 } else { 1566 cif = (struct carp_if *)ifp->if_carp; 1567 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) 1568 if (vr != sc && vr->sc_vhid == sc->sc_vhid) 1569 return (EINVAL); 1570 } 1571 1572 /* detach from old interface */ 1573 if (sc->sc_carpdev != NULL) 1574 carpdetach(sc); 1575 1576 /* join multicast groups */ 1577 if (sc->sc_naddrs < 0 && 1578 (error = carp_join_multicast(sc)) != 0) { 1579 if (ncif != NULL) 1580 free(ncif, M_IFADDR); 1581 return (error); 1582 } 1583 1584 #ifdef INET6 1585 if (sc->sc_naddrs6 < 0 && 1586 (error = carp_join_multicast6(sc)) != 0) { 1587 if (ncif != NULL) 1588 free(ncif, M_IFADDR); 1589 carp_multicast_cleanup(sc); 1590 return (error); 1591 } 1592 #endif 1593 1594 /* attach carp interface to physical interface */ 1595 if (ncif != NULL) 1596 ifp->if_carp = (void *)ncif; 1597 sc->sc_carpdev = ifp; 1598 sc->sc_if.if_capabilities = ifp->if_capabilities & 1599 (IFCAP_TSOv4 | IFCAP_TSOv6 | 1600 IFCAP_CSUM_IPv4_Tx|IFCAP_CSUM_IPv4_Rx| 1601 IFCAP_CSUM_TCPv4_Tx|IFCAP_CSUM_TCPv4_Rx| 1602 IFCAP_CSUM_UDPv4_Tx|IFCAP_CSUM_UDPv4_Rx| 1603 IFCAP_CSUM_TCPv6_Tx|IFCAP_CSUM_TCPv6_Rx| 1604 IFCAP_CSUM_UDPv6_Tx|IFCAP_CSUM_UDPv6_Rx); 1605 1606 cif = (struct carp_if *)ifp->if_carp; 1607 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) { 1608 if (vr == sc) 1609 myself = 1; 1610 if (vr->sc_vhid < sc->sc_vhid) 1611 after = vr; 1612 } 1613 1614 if (!myself) { 1615 /* We're trying to keep things in order */ 1616 if (after == NULL) { 1617 TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list); 1618 } else { 1619 TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, 1620 sc, sc_list); 1621 } 1622 cif->vhif_nvrs++; 1623 } 1624 if (sc->sc_naddrs || sc->sc_naddrs6) 1625 sc->sc_if.if_flags |= IFF_UP; 1626 carp_set_enaddr(sc); 1627 KERNEL_LOCK(1, NULL); 1628 s = splnet(); 1629 /* XXX linkstatehooks establish */ 1630 carp_carpdev_state(ifp); 1631 splx(s); 1632 KERNEL_UNLOCK_ONE(NULL); 1633 } else { 1634 carpdetach(sc); 1635 sc->sc_if.if_flags &= ~(IFF_UP|IFF_RUNNING); 1636 } 1637 return (0); 1638 } 1639 1640 static void 1641 carp_set_enaddr(struct carp_softc *sc) 1642 { 1643 uint8_t enaddr[ETHER_ADDR_LEN]; 1644 if (sc->sc_carpdev && sc->sc_carpdev->if_type == IFT_ISO88025) { 1645 enaddr[0] = 3; 1646 enaddr[1] = 0; 1647 enaddr[2] = 0x40 >> (sc->sc_vhid - 1); 1648 enaddr[3] = 0x40000 >> (sc->sc_vhid - 1); 1649 enaddr[4] = 0; 1650 enaddr[5] = 0; 1651 } else { 1652 enaddr[0] = 0; 1653 enaddr[1] = 0; 1654 enaddr[2] = 0x5e; 1655 enaddr[3] = 0; 1656 enaddr[4] = 1; 1657 enaddr[5] = sc->sc_vhid; 1658 } 1659 if_set_sadl(&sc->sc_if, enaddr, sizeof(enaddr), false); 1660 } 1661 1662 #if 0 1663 static void 1664 carp_addr_updated(void *v) 1665 { 1666 struct carp_softc *sc = (struct carp_softc *) v; 1667 struct ifaddr *ifa; 1668 int new_naddrs = 0, new_naddrs6 = 0; 1669 1670 IFADDR_FOREACH(ifa, &sc->sc_if) { 1671 if (ifa->ifa_addr->sa_family == AF_INET) 1672 new_naddrs++; 1673 else if (ifa->ifa_addr->sa_family == AF_INET6) 1674 new_naddrs6++; 1675 } 1676 1677 /* Handle a callback after SIOCDIFADDR */ 1678 if (new_naddrs < sc->sc_naddrs || new_naddrs6 < sc->sc_naddrs6) { 1679 struct in_addr mc_addr; 1680 1681 sc->sc_naddrs = new_naddrs; 1682 sc->sc_naddrs6 = new_naddrs6; 1683 1684 /* Re-establish multicast membership removed by in_control */ 1685 mc_addr.s_addr = INADDR_CARP_GROUP; 1686 if (!in_multi_group(mc_addr, &sc->sc_if, 0)) { 1687 memset(&sc->sc_imo, 0, sizeof(sc->sc_imo)); 1688 1689 if (sc->sc_carpdev != NULL && sc->sc_naddrs > 0) 1690 carp_join_multicast(sc); 1691 } 1692 1693 if (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) { 1694 sc->sc_if.if_flags &= ~IFF_UP; 1695 carp_set_state(sc, INIT); 1696 } else 1697 carp_hmac_prepare(sc); 1698 } 1699 1700 carp_setrun(sc, 0); 1701 } 1702 #endif 1703 1704 static int 1705 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin) 1706 { 1707 struct ifnet *ifp = sc->sc_carpdev; 1708 struct in_ifaddr *ia, *ia_if; 1709 int error = 0; 1710 1711 if (sin->sin_addr.s_addr == 0) { 1712 if (!(sc->sc_if.if_flags & IFF_UP)) 1713 carp_set_state(sc, INIT); 1714 if (sc->sc_naddrs) 1715 sc->sc_if.if_flags |= IFF_UP; 1716 carp_setrun(sc, 0); 1717 return (0); 1718 } 1719 1720 /* we have to do this by hand to ensure we don't match on ourselves */ 1721 ia_if = NULL; 1722 for (ia = TAILQ_FIRST(&in_ifaddrhead); ia; 1723 ia = TAILQ_NEXT(ia, ia_list)) { 1724 1725 /* and, yeah, we need a multicast-capable iface too */ 1726 if (ia->ia_ifp != &sc->sc_if && 1727 ia->ia_ifp->if_type != IFT_CARP && 1728 (ia->ia_ifp->if_flags & IFF_MULTICAST) && 1729 (sin->sin_addr.s_addr & ia->ia_subnetmask) == 1730 ia->ia_subnet) { 1731 if (!ia_if) 1732 ia_if = ia; 1733 } 1734 } 1735 1736 if (ia_if) { 1737 ia = ia_if; 1738 if (ifp) { 1739 if (ifp != ia->ia_ifp) 1740 return (EADDRNOTAVAIL); 1741 } else { 1742 ifp = ia->ia_ifp; 1743 } 1744 } 1745 1746 if ((error = carp_set_ifp(sc, ifp))) 1747 return (error); 1748 1749 if (sc->sc_carpdev == NULL) 1750 return (EADDRNOTAVAIL); 1751 1752 if (sc->sc_naddrs == 0 && (error = carp_join_multicast(sc)) != 0) 1753 return (error); 1754 1755 sc->sc_naddrs++; 1756 if (sc->sc_carpdev != NULL) 1757 sc->sc_if.if_flags |= IFF_UP; 1758 1759 carp_set_state(sc, INIT); 1760 carp_setrun(sc, 0); 1761 1762 /* 1763 * Hook if_addrhooks so that we get a callback after in_ifinit has run, 1764 * to correct any inappropriate routes that it inserted. 1765 */ 1766 if (sc->ah_cookie == 0) { 1767 /* XXX link address hook */ 1768 } 1769 1770 return (0); 1771 } 1772 1773 static int 1774 carp_join_multicast(struct carp_softc *sc) 1775 { 1776 struct ip_moptions *imo = &sc->sc_imo, tmpimo; 1777 struct in_addr addr; 1778 1779 memset(&tmpimo, 0, sizeof(tmpimo)); 1780 addr.s_addr = INADDR_CARP_GROUP; 1781 if ((tmpimo.imo_membership[0] = 1782 in_addmulti(&addr, &sc->sc_if)) == NULL) { 1783 return (ENOBUFS); 1784 } 1785 1786 imo->imo_membership[0] = tmpimo.imo_membership[0]; 1787 imo->imo_num_memberships = 1; 1788 imo->imo_multicast_ifp = &sc->sc_if; 1789 imo->imo_multicast_ttl = CARP_DFLTTL; 1790 imo->imo_multicast_loop = 0; 1791 return (0); 1792 } 1793 1794 1795 #ifdef INET6 1796 static int 1797 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6) 1798 { 1799 struct ifnet *ifp = sc->sc_carpdev; 1800 struct in6_ifaddr *ia, *ia_if; 1801 int error = 0; 1802 1803 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) { 1804 if (!(sc->sc_if.if_flags & IFF_UP)) 1805 carp_set_state(sc, INIT); 1806 if (sc->sc_naddrs6) 1807 sc->sc_if.if_flags |= IFF_UP; 1808 carp_setrun(sc, 0); 1809 return (0); 1810 } 1811 1812 /* we have to do this by hand to ensure we don't match on ourselves */ 1813 ia_if = NULL; 1814 for (ia = in6_ifaddr; ia; ia = ia->ia_next) { 1815 int i; 1816 1817 for (i = 0; i < 4; i++) { 1818 if ((sin6->sin6_addr.s6_addr32[i] & 1819 ia->ia_prefixmask.sin6_addr.s6_addr32[i]) != 1820 (ia->ia_addr.sin6_addr.s6_addr32[i] & 1821 ia->ia_prefixmask.sin6_addr.s6_addr32[i])) 1822 break; 1823 } 1824 /* and, yeah, we need a multicast-capable iface too */ 1825 if (ia->ia_ifp != &sc->sc_if && 1826 ia->ia_ifp->if_type != IFT_CARP && 1827 (ia->ia_ifp->if_flags & IFF_MULTICAST) && 1828 (i == 4)) { 1829 if (!ia_if) 1830 ia_if = ia; 1831 } 1832 } 1833 1834 if (ia_if) { 1835 ia = ia_if; 1836 if (sc->sc_carpdev) { 1837 if (sc->sc_carpdev != ia->ia_ifp) 1838 return (EADDRNOTAVAIL); 1839 } else { 1840 ifp = ia->ia_ifp; 1841 } 1842 } 1843 1844 if ((error = carp_set_ifp(sc, ifp))) 1845 return (error); 1846 1847 if (sc->sc_carpdev == NULL) 1848 return (EADDRNOTAVAIL); 1849 1850 if (sc->sc_naddrs6 == 0 && (error = carp_join_multicast6(sc)) != 0) 1851 return (error); 1852 1853 sc->sc_naddrs6++; 1854 if (sc->sc_carpdev != NULL) 1855 sc->sc_if.if_flags |= IFF_UP; 1856 carp_set_state(sc, INIT); 1857 carp_setrun(sc, 0); 1858 1859 return (0); 1860 } 1861 1862 static int 1863 carp_join_multicast6(struct carp_softc *sc) 1864 { 1865 struct in6_multi_mship *imm, *imm2; 1866 struct ip6_moptions *im6o = &sc->sc_im6o; 1867 struct sockaddr_in6 addr6; 1868 int error; 1869 1870 /* Join IPv6 CARP multicast group */ 1871 memset(&addr6, 0, sizeof(addr6)); 1872 addr6.sin6_family = AF_INET6; 1873 addr6.sin6_len = sizeof(addr6); 1874 addr6.sin6_addr.s6_addr16[0] = htons(0xff02); 1875 addr6.sin6_addr.s6_addr16[1] = htons(sc->sc_if.if_index); 1876 addr6.sin6_addr.s6_addr8[15] = 0x12; 1877 if ((imm = in6_joingroup(&sc->sc_if, 1878 &addr6.sin6_addr, &error, 0)) == NULL) { 1879 return (error); 1880 } 1881 /* join solicited multicast address */ 1882 memset(&addr6.sin6_addr, 0, sizeof(addr6.sin6_addr)); 1883 addr6.sin6_addr.s6_addr16[0] = htons(0xff02); 1884 addr6.sin6_addr.s6_addr16[1] = htons(sc->sc_if.if_index); 1885 addr6.sin6_addr.s6_addr32[1] = 0; 1886 addr6.sin6_addr.s6_addr32[2] = htonl(1); 1887 addr6.sin6_addr.s6_addr32[3] = 0; 1888 addr6.sin6_addr.s6_addr8[12] = 0xff; 1889 if ((imm2 = in6_joingroup(&sc->sc_if, 1890 &addr6.sin6_addr, &error, 0)) == NULL) { 1891 in6_leavegroup(imm); 1892 return (error); 1893 } 1894 1895 /* apply v6 multicast membership */ 1896 im6o->im6o_multicast_ifp = &sc->sc_if; 1897 if (imm) 1898 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, 1899 i6mm_chain); 1900 if (imm2) 1901 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm2, 1902 i6mm_chain); 1903 1904 return (0); 1905 } 1906 1907 #endif /* INET6 */ 1908 1909 static int 1910 carp_ioctl(struct ifnet *ifp, u_long cmd, void *data) 1911 { 1912 struct lwp *l = curlwp; /* XXX */ 1913 struct carp_softc *sc = ifp->if_softc, *vr; 1914 struct carpreq carpr; 1915 struct ifaddr *ifa; 1916 struct ifreq *ifr; 1917 struct ifnet *cdev = NULL; 1918 int error = 0; 1919 1920 ifa = (struct ifaddr *)data; 1921 ifr = (struct ifreq *)data; 1922 1923 switch (cmd) { 1924 case SIOCINITIFADDR: 1925 switch (ifa->ifa_addr->sa_family) { 1926 #ifdef INET 1927 case AF_INET: 1928 sc->sc_if.if_flags |= IFF_UP; 1929 memcpy(ifa->ifa_dstaddr, ifa->ifa_addr, 1930 sizeof(struct sockaddr)); 1931 error = carp_set_addr(sc, satosin(ifa->ifa_addr)); 1932 break; 1933 #endif /* INET */ 1934 #ifdef INET6 1935 case AF_INET6: 1936 sc->sc_if.if_flags|= IFF_UP; 1937 error = carp_set_addr6(sc, satosin6(ifa->ifa_addr)); 1938 break; 1939 #endif /* INET6 */ 1940 default: 1941 error = EAFNOSUPPORT; 1942 break; 1943 } 1944 break; 1945 1946 case SIOCSIFFLAGS: 1947 if ((error = ifioctl_common(ifp, cmd, data)) != 0) 1948 break; 1949 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) { 1950 callout_stop(&sc->sc_ad_tmo); 1951 callout_stop(&sc->sc_md_tmo); 1952 callout_stop(&sc->sc_md6_tmo); 1953 if (sc->sc_state == MASTER) { 1954 /* we need the interface up to bow out */ 1955 sc->sc_if.if_flags |= IFF_UP; 1956 sc->sc_bow_out = 1; 1957 carp_send_ad(sc); 1958 } 1959 sc->sc_if.if_flags &= ~IFF_UP; 1960 carp_set_state(sc, INIT); 1961 carp_setrun(sc, 0); 1962 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) { 1963 sc->sc_if.if_flags |= IFF_UP; 1964 carp_setrun(sc, 0); 1965 } 1966 break; 1967 1968 case SIOCSVH: 1969 if (l == NULL) 1970 break; 1971 if ((error = kauth_authorize_network(l->l_cred, 1972 KAUTH_NETWORK_INTERFACE, 1973 KAUTH_REQ_NETWORK_INTERFACE_SETPRIV, ifp, (void *)cmd, 1974 NULL)) != 0) 1975 break; 1976 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr))) 1977 break; 1978 error = 1; 1979 if (carpr.carpr_carpdev[0] != '\0' && 1980 (cdev = ifunit(carpr.carpr_carpdev)) == NULL) 1981 return (EINVAL); 1982 if ((error = carp_set_ifp(sc, cdev))) 1983 return (error); 1984 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) { 1985 switch (carpr.carpr_state) { 1986 case BACKUP: 1987 callout_stop(&sc->sc_ad_tmo); 1988 carp_set_state(sc, BACKUP); 1989 carp_setrun(sc, 0); 1990 carp_setroute(sc, RTM_DELETE); 1991 break; 1992 case MASTER: 1993 carp_master_down(sc); 1994 break; 1995 default: 1996 break; 1997 } 1998 } 1999 if (carpr.carpr_vhid > 0) { 2000 if (carpr.carpr_vhid > 255) { 2001 error = EINVAL; 2002 break; 2003 } 2004 if (sc->sc_carpdev) { 2005 struct carp_if *cif; 2006 cif = (struct carp_if *)sc->sc_carpdev->if_carp; 2007 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) 2008 if (vr != sc && 2009 vr->sc_vhid == carpr.carpr_vhid) 2010 return (EINVAL); 2011 } 2012 sc->sc_vhid = carpr.carpr_vhid; 2013 carp_set_enaddr(sc); 2014 carp_set_state(sc, INIT); 2015 error--; 2016 } 2017 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) { 2018 if (carpr.carpr_advskew > 254) { 2019 error = EINVAL; 2020 break; 2021 } 2022 if (carpr.carpr_advbase > 255) { 2023 error = EINVAL; 2024 break; 2025 } 2026 sc->sc_advbase = carpr.carpr_advbase; 2027 sc->sc_advskew = carpr.carpr_advskew; 2028 error--; 2029 } 2030 memcpy(sc->sc_key, carpr.carpr_key, sizeof(sc->sc_key)); 2031 if (error > 0) 2032 error = EINVAL; 2033 else { 2034 error = 0; 2035 carp_setrun(sc, 0); 2036 } 2037 break; 2038 2039 case SIOCGVH: 2040 memset(&carpr, 0, sizeof(carpr)); 2041 if (sc->sc_carpdev != NULL) 2042 strlcpy(carpr.carpr_carpdev, sc->sc_carpdev->if_xname, 2043 IFNAMSIZ); 2044 carpr.carpr_state = sc->sc_state; 2045 carpr.carpr_vhid = sc->sc_vhid; 2046 carpr.carpr_advbase = sc->sc_advbase; 2047 carpr.carpr_advskew = sc->sc_advskew; 2048 2049 if ((l != NULL) && (error = kauth_authorize_network(l->l_cred, 2050 KAUTH_NETWORK_INTERFACE, 2051 KAUTH_REQ_NETWORK_INTERFACE_SETPRIV, ifp, (void *)cmd, 2052 NULL)) == 0) 2053 memcpy(carpr.carpr_key, sc->sc_key, 2054 sizeof(carpr.carpr_key)); 2055 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr)); 2056 break; 2057 2058 case SIOCADDMULTI: 2059 error = carp_ether_addmulti(sc, ifr); 2060 break; 2061 2062 case SIOCDELMULTI: 2063 error = carp_ether_delmulti(sc, ifr); 2064 break; 2065 2066 case SIOCSIFCAP: 2067 if ((error = ifioctl_common(ifp, cmd, data)) == ENETRESET) 2068 error = 0; 2069 break; 2070 2071 default: 2072 error = ether_ioctl(ifp, cmd, data); 2073 } 2074 2075 carp_hmac_prepare(sc); 2076 return (error); 2077 } 2078 2079 2080 /* 2081 * Start output on carp interface. This function should never be called. 2082 */ 2083 static void 2084 carp_start(struct ifnet *ifp) 2085 { 2086 #ifdef DEBUG 2087 printf("%s: start called\n", ifp->if_xname); 2088 #endif 2089 } 2090 2091 int 2092 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa, 2093 struct rtentry *rt) 2094 { 2095 struct carp_softc *sc = ((struct carp_softc *)ifp->if_softc); 2096 KASSERT(KERNEL_LOCKED_P()); 2097 2098 if (sc->sc_carpdev != NULL && sc->sc_state == MASTER) { 2099 return (sc->sc_carpdev->if_output(ifp, m, sa, rt)); 2100 } else { 2101 m_freem(m); 2102 return (ENETUNREACH); 2103 } 2104 } 2105 2106 static void 2107 carp_set_state(struct carp_softc *sc, int state) 2108 { 2109 static const char *carp_states[] = { CARP_STATES }; 2110 if (sc->sc_state == state) 2111 return; 2112 2113 CARP_LOG(sc, ("state transition from: %s -> to: %s", carp_states[sc->sc_state], carp_states[state])); 2114 2115 sc->sc_state = state; 2116 switch (state) { 2117 case BACKUP: 2118 sc->sc_if.if_link_state = LINK_STATE_DOWN; 2119 break; 2120 case MASTER: 2121 sc->sc_if.if_link_state = LINK_STATE_UP; 2122 break; 2123 default: 2124 sc->sc_if.if_link_state = LINK_STATE_UNKNOWN; 2125 break; 2126 } 2127 rt_ifmsg(&sc->sc_if); 2128 } 2129 2130 void 2131 carp_carpdev_state(void *v) 2132 { 2133 struct carp_if *cif; 2134 struct carp_softc *sc; 2135 struct ifnet *ifp = v; 2136 2137 if (ifp->if_type == IFT_CARP) 2138 return; 2139 2140 cif = (struct carp_if *)ifp->if_carp; 2141 2142 TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) { 2143 int suppressed = sc->sc_suppress; 2144 2145 if (sc->sc_carpdev->if_link_state == LINK_STATE_DOWN || 2146 !(sc->sc_carpdev->if_flags & IFF_UP)) { 2147 sc->sc_if.if_flags &= ~IFF_RUNNING; 2148 callout_stop(&sc->sc_ad_tmo); 2149 callout_stop(&sc->sc_md_tmo); 2150 callout_stop(&sc->sc_md6_tmo); 2151 carp_set_state(sc, INIT); 2152 sc->sc_suppress = 1; 2153 carp_setrun(sc, 0); 2154 if (!suppressed) { 2155 carp_suppress_preempt++; 2156 if (carp_suppress_preempt == 1) 2157 carp_send_ad_all(); 2158 } 2159 } else { 2160 carp_set_state(sc, INIT); 2161 sc->sc_suppress = 0; 2162 carp_setrun(sc, 0); 2163 if (suppressed) 2164 carp_suppress_preempt--; 2165 } 2166 } 2167 } 2168 2169 static int 2170 carp_ether_addmulti(struct carp_softc *sc, struct ifreq *ifr) 2171 { 2172 const struct sockaddr *sa = ifreq_getaddr(SIOCADDMULTI, ifr); 2173 struct ifnet *ifp; 2174 struct carp_mc_entry *mc; 2175 u_int8_t addrlo[ETHER_ADDR_LEN], addrhi[ETHER_ADDR_LEN]; 2176 int error; 2177 2178 ifp = sc->sc_carpdev; 2179 if (ifp == NULL) 2180 return (EINVAL); 2181 2182 error = ether_addmulti(sa, &sc->sc_ac); 2183 if (error != ENETRESET) 2184 return (error); 2185 2186 /* 2187 * This is new multicast address. We have to tell parent 2188 * about it. Also, remember this multicast address so that 2189 * we can delete them on unconfigure. 2190 */ 2191 mc = malloc(sizeof(struct carp_mc_entry), M_DEVBUF, M_NOWAIT); 2192 if (mc == NULL) { 2193 error = ENOMEM; 2194 goto alloc_failed; 2195 } 2196 2197 /* 2198 * As ether_addmulti() returns ENETRESET, following two 2199 * statement shouldn't fail. 2200 */ 2201 (void)ether_multiaddr(sa, addrlo, addrhi); 2202 ETHER_LOOKUP_MULTI(addrlo, addrhi, &sc->sc_ac, mc->mc_enm); 2203 memcpy(&mc->mc_addr, sa, sa->sa_len); 2204 LIST_INSERT_HEAD(&sc->carp_mc_listhead, mc, mc_entries); 2205 2206 error = if_mcast_op(ifp, SIOCADDMULTI, sa); 2207 if (error != 0) 2208 goto ioctl_failed; 2209 2210 return (error); 2211 2212 ioctl_failed: 2213 LIST_REMOVE(mc, mc_entries); 2214 free(mc, M_DEVBUF); 2215 alloc_failed: 2216 (void)ether_delmulti(sa, &sc->sc_ac); 2217 2218 return (error); 2219 } 2220 2221 static int 2222 carp_ether_delmulti(struct carp_softc *sc, struct ifreq *ifr) 2223 { 2224 const struct sockaddr *sa = ifreq_getaddr(SIOCDELMULTI, ifr); 2225 struct ifnet *ifp; 2226 struct ether_multi *enm; 2227 struct carp_mc_entry *mc; 2228 u_int8_t addrlo[ETHER_ADDR_LEN], addrhi[ETHER_ADDR_LEN]; 2229 int error; 2230 2231 ifp = sc->sc_carpdev; 2232 if (ifp == NULL) 2233 return (EINVAL); 2234 2235 /* 2236 * Find a key to lookup carp_mc_entry. We have to do this 2237 * before calling ether_delmulti for obvious reason. 2238 */ 2239 if ((error = ether_multiaddr(sa, addrlo, addrhi)) != 0) 2240 return (error); 2241 ETHER_LOOKUP_MULTI(addrlo, addrhi, &sc->sc_ac, enm); 2242 if (enm == NULL) 2243 return (EINVAL); 2244 2245 LIST_FOREACH(mc, &sc->carp_mc_listhead, mc_entries) 2246 if (mc->mc_enm == enm) 2247 break; 2248 2249 /* We won't delete entries we didn't add */ 2250 if (mc == NULL) 2251 return (EINVAL); 2252 2253 error = ether_delmulti(sa, &sc->sc_ac); 2254 if (error != ENETRESET) 2255 return (error); 2256 2257 /* We no longer use this multicast address. Tell parent so. */ 2258 error = if_mcast_op(ifp, SIOCDELMULTI, sa); 2259 if (error == 0) { 2260 /* And forget about this address. */ 2261 LIST_REMOVE(mc, mc_entries); 2262 free(mc, M_DEVBUF); 2263 } else 2264 (void)ether_addmulti(sa, &sc->sc_ac); 2265 return (error); 2266 } 2267 2268 /* 2269 * Delete any multicast address we have asked to add from parent 2270 * interface. Called when the carp is being unconfigured. 2271 */ 2272 static void 2273 carp_ether_purgemulti(struct carp_softc *sc) 2274 { 2275 struct ifnet *ifp = sc->sc_carpdev; /* Parent. */ 2276 struct carp_mc_entry *mc; 2277 2278 if (ifp == NULL) 2279 return; 2280 2281 while ((mc = LIST_FIRST(&sc->carp_mc_listhead)) != NULL) { 2282 (void)if_mcast_op(ifp, SIOCDELMULTI, sstosa(&mc->mc_addr)); 2283 LIST_REMOVE(mc, mc_entries); 2284 free(mc, M_DEVBUF); 2285 } 2286 } 2287 2288 static int 2289 sysctl_net_inet_carp_stats(SYSCTLFN_ARGS) 2290 { 2291 2292 return (NETSTAT_SYSCTL(carpstat_percpu, CARP_NSTATS)); 2293 } 2294 2295 void 2296 carp_init(void) 2297 { 2298 2299 sysctl_net_inet_carp_setup(NULL); 2300 #ifdef MBUFTRACE 2301 MOWNER_ATTACH(&carp_proto_mowner_rx); 2302 MOWNER_ATTACH(&carp_proto_mowner_tx); 2303 MOWNER_ATTACH(&carp_proto6_mowner_rx); 2304 MOWNER_ATTACH(&carp_proto6_mowner_tx); 2305 #endif 2306 } 2307 2308 static void 2309 sysctl_net_inet_carp_setup(struct sysctllog **clog) 2310 { 2311 2312 sysctl_createv(clog, 0, NULL, NULL, 2313 CTLFLAG_PERMANENT, 2314 CTLTYPE_NODE, "inet", NULL, 2315 NULL, 0, NULL, 0, 2316 CTL_NET, PF_INET, CTL_EOL); 2317 sysctl_createv(clog, 0, NULL, NULL, 2318 CTLFLAG_PERMANENT, 2319 CTLTYPE_NODE, "carp", 2320 SYSCTL_DESCR("CARP related settings"), 2321 NULL, 0, NULL, 0, 2322 CTL_NET, PF_INET, IPPROTO_CARP, CTL_EOL); 2323 2324 sysctl_createv(clog, 0, NULL, NULL, 2325 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2326 CTLTYPE_INT, "preempt", 2327 SYSCTL_DESCR("Enable CARP Preempt"), 2328 NULL, 0, &carp_opts[CARPCTL_PREEMPT], 0, 2329 CTL_NET, PF_INET, IPPROTO_CARP, 2330 CTL_CREATE, CTL_EOL); 2331 sysctl_createv(clog, 0, NULL, NULL, 2332 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2333 CTLTYPE_INT, "arpbalance", 2334 SYSCTL_DESCR("Enable ARP balancing"), 2335 NULL, 0, &carp_opts[CARPCTL_ARPBALANCE], 0, 2336 CTL_NET, PF_INET, IPPROTO_CARP, 2337 CTL_CREATE, CTL_EOL); 2338 sysctl_createv(clog, 0, NULL, NULL, 2339 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2340 CTLTYPE_INT, "allow", 2341 SYSCTL_DESCR("Enable CARP"), 2342 NULL, 0, &carp_opts[CARPCTL_ALLOW], 0, 2343 CTL_NET, PF_INET, IPPROTO_CARP, 2344 CTL_CREATE, CTL_EOL); 2345 sysctl_createv(clog, 0, NULL, NULL, 2346 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2347 CTLTYPE_INT, "log", 2348 SYSCTL_DESCR("CARP logging"), 2349 NULL, 0, &carp_opts[CARPCTL_LOG], 0, 2350 CTL_NET, PF_INET, IPPROTO_CARP, 2351 CTL_CREATE, CTL_EOL); 2352 sysctl_createv(clog, 0, NULL, NULL, 2353 CTLFLAG_PERMANENT, 2354 CTLTYPE_STRUCT, "stats", 2355 SYSCTL_DESCR("CARP statistics"), 2356 sysctl_net_inet_carp_stats, 0, NULL, 0, 2357 CTL_NET, PF_INET, IPPROTO_CARP, CARPCTL_STATS, 2358 CTL_EOL); 2359 } 2360