ip_carp.c revision 1.83
1 /* $NetBSD: ip_carp.c,v 1.83 2017/01/16 15:44:47 christos 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 1.83 2017/01/16 15:44:47 christos 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 #include <netinet6/in6_var.h> 98 #endif 99 100 #include <net/bpf.h> 101 102 #include <sys/sha1.h> 103 104 #include <netinet/ip_carp.h> 105 106 #include "ioconf.h" 107 108 struct carp_mc_entry { 109 LIST_ENTRY(carp_mc_entry) mc_entries; 110 union { 111 struct ether_multi *mcu_enm; 112 } mc_u; 113 struct sockaddr_storage mc_addr; 114 }; 115 #define mc_enm mc_u.mcu_enm 116 117 struct carp_softc { 118 struct ethercom sc_ac; 119 #define sc_if sc_ac.ec_if 120 #define sc_carpdev sc_ac.ec_if.if_carpdev 121 int ah_cookie; 122 int lh_cookie; 123 struct ip_moptions sc_imo; 124 #ifdef INET6 125 struct ip6_moptions sc_im6o; 126 #endif /* INET6 */ 127 TAILQ_ENTRY(carp_softc) sc_list; 128 129 enum { INIT = 0, BACKUP, MASTER } sc_state; 130 131 int sc_suppress; 132 int sc_bow_out; 133 134 int sc_sendad_errors; 135 #define CARP_SENDAD_MAX_ERRORS 3 136 int sc_sendad_success; 137 #define CARP_SENDAD_MIN_SUCCESS 3 138 139 int sc_vhid; 140 int sc_advskew; 141 int sc_naddrs; 142 int sc_naddrs6; 143 int sc_advbase; /* seconds */ 144 int sc_init_counter; 145 u_int64_t sc_counter; 146 147 /* authentication */ 148 #define CARP_HMAC_PAD 64 149 unsigned char sc_key[CARP_KEY_LEN]; 150 unsigned char sc_pad[CARP_HMAC_PAD]; 151 SHA1_CTX sc_sha1; 152 u_int32_t sc_hashkey[2]; 153 154 struct callout sc_ad_tmo; /* advertisement timeout */ 155 struct callout sc_md_tmo; /* master down timeout */ 156 struct callout sc_md6_tmo; /* master down timeout */ 157 158 LIST_HEAD(__carp_mchead, carp_mc_entry) carp_mc_listhead; 159 }; 160 161 int carp_suppress_preempt = 0; 162 static int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 0, 0 }; /* XXX for now */ 163 164 static percpu_t *carpstat_percpu; 165 166 #define CARP_STATINC(x) _NET_STATINC(carpstat_percpu, x) 167 168 #ifdef MBUFTRACE 169 static struct mowner carp_proto_mowner_rx = MOWNER_INIT("carp", "rx"); 170 static struct mowner carp_proto_mowner_tx = MOWNER_INIT("carp", "tx"); 171 static struct mowner carp_proto6_mowner_rx = MOWNER_INIT("carp6", "rx"); 172 static struct mowner carp_proto6_mowner_tx = MOWNER_INIT("carp6", "tx"); 173 #endif 174 175 struct carp_if { 176 TAILQ_HEAD(, carp_softc) vhif_vrs; 177 int vhif_nvrs; 178 179 struct ifnet *vhif_ifp; 180 }; 181 182 #define CARP_LOG(sc, s) \ 183 if (carp_opts[CARPCTL_LOG]) { \ 184 if (sc) \ 185 log(LOG_INFO, "%s: ", \ 186 (sc)->sc_if.if_xname); \ 187 else \ 188 log(LOG_INFO, "carp: "); \ 189 addlog s; \ 190 addlog("\n"); \ 191 } 192 193 static void carp_hmac_prepare(struct carp_softc *); 194 static void carp_hmac_generate(struct carp_softc *, u_int32_t *, 195 unsigned char *); 196 static int carp_hmac_verify(struct carp_softc *, u_int32_t *, 197 unsigned char *); 198 static void carp_setroute(struct carp_softc *, int); 199 static void carp_proto_input_c(struct mbuf *, struct carp_header *, 200 sa_family_t); 201 static void carpdetach(struct carp_softc *); 202 static int carp_prepare_ad(struct mbuf *, struct carp_softc *, 203 struct carp_header *); 204 static void carp_send_ad_all(void); 205 static void carp_send_ad(void *); 206 static void carp_send_arp(struct carp_softc *); 207 static void carp_master_down(void *); 208 static int carp_ioctl(struct ifnet *, u_long, void *); 209 static void carp_start(struct ifnet *); 210 static void carp_setrun(struct carp_softc *, sa_family_t); 211 static void carp_set_state(struct carp_softc *, int); 212 static int carp_addrcount(struct carp_if *, struct in_ifaddr *, int); 213 enum { CARP_COUNT_MASTER, CARP_COUNT_RUNNING }; 214 215 static void carp_multicast_cleanup(struct carp_softc *); 216 static int carp_set_ifp(struct carp_softc *, struct ifnet *); 217 static void carp_set_enaddr(struct carp_softc *); 218 #if 0 219 static void carp_addr_updated(void *); 220 #endif 221 static u_int32_t carp_hash(struct carp_softc *, u_char *); 222 static int carp_set_addr(struct carp_softc *, struct sockaddr_in *); 223 static int carp_join_multicast(struct carp_softc *); 224 #ifdef INET6 225 static void carp_send_na(struct carp_softc *); 226 static int carp_set_addr6(struct carp_softc *, struct sockaddr_in6 *); 227 static int carp_join_multicast6(struct carp_softc *); 228 #endif 229 static int carp_clone_create(struct if_clone *, int); 230 static int carp_clone_destroy(struct ifnet *); 231 static int carp_ether_addmulti(struct carp_softc *, struct ifreq *); 232 static int carp_ether_delmulti(struct carp_softc *, struct ifreq *); 233 static void carp_ether_purgemulti(struct carp_softc *); 234 235 static void sysctl_net_inet_carp_setup(struct sysctllog **); 236 237 struct if_clone carp_cloner = 238 IF_CLONE_INITIALIZER("carp", carp_clone_create, carp_clone_destroy); 239 240 static __inline u_int16_t 241 carp_cksum(struct mbuf *m, int len) 242 { 243 return (in_cksum(m, len)); 244 } 245 246 static void 247 carp_hmac_prepare(struct carp_softc *sc) 248 { 249 u_int8_t carp_version = CARP_VERSION, type = CARP_ADVERTISEMENT; 250 u_int8_t vhid = sc->sc_vhid & 0xff; 251 SHA1_CTX sha1ctx; 252 u_int32_t kmd[5]; 253 struct ifaddr *ifa; 254 int i, found; 255 struct in_addr last, cur, in; 256 #ifdef INET6 257 struct in6_addr last6, cur6, in6; 258 #endif /* INET6 */ 259 260 /* compute ipad from key */ 261 memset(sc->sc_pad, 0, sizeof(sc->sc_pad)); 262 memcpy(sc->sc_pad, sc->sc_key, sizeof(sc->sc_key)); 263 for (i = 0; i < sizeof(sc->sc_pad); i++) 264 sc->sc_pad[i] ^= 0x36; 265 266 /* precompute first part of inner hash */ 267 SHA1Init(&sc->sc_sha1); 268 SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad)); 269 SHA1Update(&sc->sc_sha1, (void *)&carp_version, sizeof(carp_version)); 270 SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type)); 271 272 /* generate a key for the arpbalance hash, before the vhid is hashed */ 273 memcpy(&sha1ctx, &sc->sc_sha1, sizeof(sha1ctx)); 274 SHA1Final((unsigned char *)kmd, &sha1ctx); 275 sc->sc_hashkey[0] = kmd[0] ^ kmd[1]; 276 sc->sc_hashkey[1] = kmd[2] ^ kmd[3]; 277 278 /* the rest of the precomputation */ 279 SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid)); 280 281 /* Hash the addresses from smallest to largest, not interface order */ 282 #ifdef INET 283 cur.s_addr = 0; 284 do { 285 found = 0; 286 last = cur; 287 cur.s_addr = 0xffffffff; 288 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 289 in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr; 290 if (ifa->ifa_addr->sa_family == AF_INET && 291 ntohl(in.s_addr) > ntohl(last.s_addr) && 292 ntohl(in.s_addr) < ntohl(cur.s_addr)) { 293 cur.s_addr = in.s_addr; 294 found++; 295 } 296 } 297 if (found) 298 SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur)); 299 } while (found); 300 #endif /* INET */ 301 302 #ifdef INET6 303 memset(&cur6, 0x00, sizeof(cur6)); 304 do { 305 found = 0; 306 last6 = cur6; 307 memset(&cur6, 0xff, sizeof(cur6)); 308 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 309 in6 = ifatoia6(ifa)->ia_addr.sin6_addr; 310 if (IN6_IS_ADDR_LINKLOCAL(&in6)) 311 in6.s6_addr16[1] = 0; 312 if (ifa->ifa_addr->sa_family == AF_INET6 && 313 memcmp(&in6, &last6, sizeof(in6)) > 0 && 314 memcmp(&in6, &cur6, sizeof(in6)) < 0) { 315 cur6 = in6; 316 found++; 317 } 318 } 319 if (found) 320 SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6)); 321 } while (found); 322 #endif /* INET6 */ 323 324 /* convert ipad to opad */ 325 for (i = 0; i < sizeof(sc->sc_pad); i++) 326 sc->sc_pad[i] ^= 0x36 ^ 0x5c; 327 } 328 329 static void 330 carp_hmac_generate(struct carp_softc *sc, u_int32_t counter[2], 331 unsigned char md[20]) 332 { 333 SHA1_CTX sha1ctx; 334 335 /* fetch first half of inner hash */ 336 memcpy(&sha1ctx, &sc->sc_sha1, sizeof(sha1ctx)); 337 338 SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter)); 339 SHA1Final(md, &sha1ctx); 340 341 /* outer hash */ 342 SHA1Init(&sha1ctx); 343 SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad)); 344 SHA1Update(&sha1ctx, md, 20); 345 SHA1Final(md, &sha1ctx); 346 } 347 348 static int 349 carp_hmac_verify(struct carp_softc *sc, u_int32_t counter[2], 350 unsigned char md[20]) 351 { 352 unsigned char md2[20]; 353 354 carp_hmac_generate(sc, counter, md2); 355 356 return (memcmp(md, md2, sizeof(md2))); 357 } 358 359 static void 360 carp_setroute(struct carp_softc *sc, int cmd) 361 { 362 struct ifaddr *ifa; 363 int s; 364 365 KERNEL_LOCK(1, NULL); 366 s = splsoftnet(); 367 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 368 switch (ifa->ifa_addr->sa_family) { 369 case AF_INET: { 370 int count = 0; 371 struct rtentry *rt; 372 int hr_otherif, nr_ourif; 373 374 /* 375 * Avoid screwing with the routes if there are other 376 * carp interfaces which are master and have the same 377 * address. 378 */ 379 if (sc->sc_carpdev != NULL && 380 sc->sc_carpdev->if_carp != NULL) { 381 count = carp_addrcount( 382 (struct carp_if *)sc->sc_carpdev->if_carp, 383 ifatoia(ifa), CARP_COUNT_MASTER); 384 if ((cmd == RTM_ADD && count != 1) || 385 (cmd == RTM_DELETE && count != 0)) 386 continue; 387 } 388 389 /* Remove the existing host route, if any */ 390 rtrequest(RTM_DELETE, ifa->ifa_addr, 391 ifa->ifa_addr, ifa->ifa_netmask, 392 RTF_HOST, NULL); 393 394 rt = NULL; 395 (void)rtrequest(RTM_GET, ifa->ifa_addr, ifa->ifa_addr, 396 ifa->ifa_netmask, RTF_HOST, &rt); 397 hr_otherif = (rt && rt->rt_ifp != &sc->sc_if && 398 (rt->rt_flags & RTF_CONNECTED)); 399 if (rt != NULL) { 400 rt_unref(rt); 401 rt = NULL; 402 } 403 404 /* Check for a network route on our interface */ 405 406 rt = NULL; 407 (void)rtrequest(RTM_GET, ifa->ifa_addr, ifa->ifa_addr, 408 ifa->ifa_netmask, 0, &rt); 409 nr_ourif = (rt && rt->rt_ifp == &sc->sc_if); 410 411 switch (cmd) { 412 case RTM_ADD: 413 if (hr_otherif) { 414 ifa->ifa_rtrequest = NULL; 415 ifa->ifa_flags &= ~RTF_CONNECTED; 416 417 rtrequest(RTM_ADD, ifa->ifa_addr, 418 ifa->ifa_addr, ifa->ifa_netmask, 419 RTF_UP | RTF_HOST, NULL); 420 } 421 if (!hr_otherif || nr_ourif || !rt) { 422 if (nr_ourif && 423 (rt->rt_flags & RTF_CONNECTED) == 0) 424 rtrequest(RTM_DELETE, 425 ifa->ifa_addr, 426 ifa->ifa_addr, 427 ifa->ifa_netmask, 0, NULL); 428 429 ifa->ifa_rtrequest = arp_rtrequest; 430 ifa->ifa_flags |= RTF_CONNECTED; 431 432 if (rtrequest(RTM_ADD, ifa->ifa_addr, 433 ifa->ifa_addr, ifa->ifa_netmask, 0, 434 NULL) == 0) 435 ifa->ifa_flags |= IFA_ROUTE; 436 } 437 break; 438 case RTM_DELETE: 439 break; 440 default: 441 break; 442 } 443 if (rt != NULL) { 444 rt_unref(rt); 445 rt = NULL; 446 } 447 break; 448 } 449 450 #ifdef INET6 451 case AF_INET6: 452 if (cmd == RTM_ADD) 453 in6_ifaddlocal(ifa); 454 else 455 in6_ifremlocal(ifa); 456 break; 457 #endif /* INET6 */ 458 default: 459 break; 460 } 461 } 462 splx(s); 463 KERNEL_UNLOCK_ONE(NULL); 464 } 465 466 /* 467 * process input packet. 468 * we have rearranged checks order compared to the rfc, 469 * but it seems more efficient this way or not possible otherwise. 470 */ 471 void 472 carp_proto_input(struct mbuf *m, ...) 473 { 474 struct ip *ip = mtod(m, struct ip *); 475 struct carp_softc *sc = NULL; 476 struct carp_header *ch; 477 int iplen, len; 478 va_list ap; 479 struct ifnet *rcvif; 480 481 va_start(ap, m); 482 va_end(ap); 483 484 CARP_STATINC(CARP_STAT_IPACKETS); 485 MCLAIM(m, &carp_proto_mowner_rx); 486 487 if (!carp_opts[CARPCTL_ALLOW]) { 488 m_freem(m); 489 return; 490 } 491 492 rcvif = m_get_rcvif_NOMPSAFE(m); 493 /* check if received on a valid carp interface */ 494 if (rcvif->if_type != IFT_CARP) { 495 CARP_STATINC(CARP_STAT_BADIF); 496 CARP_LOG(sc, ("packet received on non-carp interface: %s", 497 rcvif->if_xname)); 498 m_freem(m); 499 return; 500 } 501 502 /* verify that the IP TTL is 255. */ 503 if (ip->ip_ttl != CARP_DFLTTL) { 504 CARP_STATINC(CARP_STAT_BADTTL); 505 CARP_LOG(sc, ("received ttl %d != %d on %s", ip->ip_ttl, 506 CARP_DFLTTL, rcvif->if_xname)); 507 m_freem(m); 508 return; 509 } 510 511 /* 512 * verify that the received packet length is 513 * equal to the CARP header 514 */ 515 iplen = ip->ip_hl << 2; 516 len = iplen + sizeof(*ch); 517 if (len > m->m_pkthdr.len) { 518 CARP_STATINC(CARP_STAT_BADLEN); 519 CARP_LOG(sc, ("packet too short %d on %s", m->m_pkthdr.len, 520 rcvif->if_xname)); 521 m_freem(m); 522 return; 523 } 524 525 if ((m = m_pullup(m, len)) == NULL) { 526 CARP_STATINC(CARP_STAT_HDROPS); 527 return; 528 } 529 ip = mtod(m, struct ip *); 530 ch = (struct carp_header *)((char *)ip + iplen); 531 /* verify the CARP checksum */ 532 m->m_data += iplen; 533 if (carp_cksum(m, len - iplen)) { 534 CARP_STATINC(CARP_STAT_BADSUM); 535 CARP_LOG(sc, ("checksum failed on %s", 536 rcvif->if_xname)); 537 m_freem(m); 538 return; 539 } 540 m->m_data -= iplen; 541 542 carp_proto_input_c(m, ch, AF_INET); 543 } 544 545 #ifdef INET6 546 int 547 carp6_proto_input(struct mbuf **mp, int *offp, int proto) 548 { 549 struct mbuf *m = *mp; 550 struct carp_softc *sc = NULL; 551 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 552 struct carp_header *ch; 553 u_int len; 554 struct ifnet *rcvif; 555 556 CARP_STATINC(CARP_STAT_IPACKETS6); 557 MCLAIM(m, &carp_proto6_mowner_rx); 558 559 if (!carp_opts[CARPCTL_ALLOW]) { 560 m_freem(m); 561 return (IPPROTO_DONE); 562 } 563 564 rcvif = m_get_rcvif_NOMPSAFE(m); 565 566 /* check if received on a valid carp interface */ 567 if (rcvif->if_type != IFT_CARP) { 568 CARP_STATINC(CARP_STAT_BADIF); 569 CARP_LOG(sc, ("packet received on non-carp interface: %s", 570 rcvif->if_xname)); 571 m_freem(m); 572 return (IPPROTO_DONE); 573 } 574 575 /* verify that the IP TTL is 255 */ 576 if (ip6->ip6_hlim != CARP_DFLTTL) { 577 CARP_STATINC(CARP_STAT_BADTTL); 578 CARP_LOG(sc, ("received ttl %d != %d on %s", ip6->ip6_hlim, 579 CARP_DFLTTL, rcvif->if_xname)); 580 m_freem(m); 581 return (IPPROTO_DONE); 582 } 583 584 /* verify that we have a complete carp packet */ 585 len = m->m_len; 586 IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch)); 587 if (ch == NULL) { 588 CARP_STATINC(CARP_STAT_BADLEN); 589 CARP_LOG(sc, ("packet size %u too small", len)); 590 return (IPPROTO_DONE); 591 } 592 593 594 /* verify the CARP checksum */ 595 m->m_data += *offp; 596 if (carp_cksum(m, sizeof(*ch))) { 597 CARP_STATINC(CARP_STAT_BADSUM); 598 CARP_LOG(sc, ("checksum failed, on %s", rcvif->if_xname)); 599 m_freem(m); 600 return (IPPROTO_DONE); 601 } 602 m->m_data -= *offp; 603 604 carp_proto_input_c(m, ch, AF_INET6); 605 return (IPPROTO_DONE); 606 } 607 #endif /* INET6 */ 608 609 static void 610 carp_proto_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af) 611 { 612 struct carp_softc *sc; 613 u_int64_t tmp_counter; 614 struct timeval sc_tv, ch_tv; 615 616 TAILQ_FOREACH(sc, &((struct carp_if *) 617 m_get_rcvif_NOMPSAFE(m)->if_carpdev->if_carp)->vhif_vrs, sc_list) 618 if (sc->sc_vhid == ch->carp_vhid) 619 break; 620 621 if (!sc || (sc->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) != 622 (IFF_UP|IFF_RUNNING)) { 623 CARP_STATINC(CARP_STAT_BADVHID); 624 m_freem(m); 625 return; 626 } 627 628 /* 629 * Check if our own advertisement was duplicated 630 * from a non simplex interface. 631 * XXX If there is no address on our physical interface 632 * there is no way to distinguish our ads from the ones 633 * another carp host might have sent us. 634 */ 635 if ((sc->sc_carpdev->if_flags & IFF_SIMPLEX) == 0) { 636 struct sockaddr sa; 637 struct ifaddr *ifa; 638 int s; 639 640 memset(&sa, 0, sizeof(sa)); 641 sa.sa_family = af; 642 s = pserialize_read_enter(); 643 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev); 644 645 if (ifa && af == AF_INET) { 646 struct ip *ip = mtod(m, struct ip *); 647 if (ip->ip_src.s_addr == 648 ifatoia(ifa)->ia_addr.sin_addr.s_addr) { 649 pserialize_read_exit(s); 650 m_freem(m); 651 return; 652 } 653 } 654 #ifdef INET6 655 if (ifa && af == AF_INET6) { 656 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 657 struct in6_addr in6_src, in6_found; 658 659 in6_src = ip6->ip6_src; 660 in6_found = ifatoia6(ifa)->ia_addr.sin6_addr; 661 if (IN6_IS_ADDR_LINKLOCAL(&in6_src)) 662 in6_src.s6_addr16[1] = 0; 663 if (IN6_IS_ADDR_LINKLOCAL(&in6_found)) 664 in6_found.s6_addr16[1] = 0; 665 if (IN6_ARE_ADDR_EQUAL(&in6_src, &in6_found)) { 666 pserialize_read_exit(s); 667 m_freem(m); 668 return; 669 } 670 } 671 #endif /* INET6 */ 672 pserialize_read_exit(s); 673 } 674 675 nanotime(&sc->sc_if.if_lastchange); 676 sc->sc_if.if_ipackets++; 677 sc->sc_if.if_ibytes += m->m_pkthdr.len; 678 679 /* verify the CARP version. */ 680 if (ch->carp_version != CARP_VERSION) { 681 CARP_STATINC(CARP_STAT_BADVER); 682 sc->sc_if.if_ierrors++; 683 CARP_LOG(sc, ("invalid version %d != %d", 684 ch->carp_version, CARP_VERSION)); 685 m_freem(m); 686 return; 687 } 688 689 /* verify the hash */ 690 if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) { 691 struct ip *ip; 692 struct ip6_hdr *ip6; 693 char ip6buf[INET6_ADDRSTRLEN]; 694 char ipbuf[INET_ADDRSTRLEN]; 695 696 CARP_STATINC(CARP_STAT_BADAUTH); 697 sc->sc_if.if_ierrors++; 698 699 switch(af) { 700 case AF_INET: 701 ip = mtod(m, struct ip *); 702 CARP_LOG(sc, ("incorrect hash from %s", 703 in_fmtaddr(ipbuf, ip->ip_src))); 704 break; 705 706 case AF_INET6: 707 ip6 = mtod(m, struct ip6_hdr *); 708 CARP_LOG(sc, ("incorrect hash from %s", 709 IN6_PRINT(ip6buf, &ip6->ip6_src))); 710 break; 711 712 default: CARP_LOG(sc, ("incorrect hash")); 713 break; 714 } 715 m_freem(m); 716 return; 717 } 718 719 tmp_counter = ntohl(ch->carp_counter[0]); 720 tmp_counter = tmp_counter<<32; 721 tmp_counter += ntohl(ch->carp_counter[1]); 722 723 /* XXX Replay protection goes here */ 724 725 sc->sc_init_counter = 0; 726 sc->sc_counter = tmp_counter; 727 728 729 sc_tv.tv_sec = sc->sc_advbase; 730 if (carp_suppress_preempt && sc->sc_advskew < 240) 731 sc_tv.tv_usec = 240 * 1000000 / 256; 732 else 733 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256; 734 ch_tv.tv_sec = ch->carp_advbase; 735 ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256; 736 737 switch (sc->sc_state) { 738 case INIT: 739 break; 740 case MASTER: 741 /* 742 * If we receive an advertisement from a backup who's going to 743 * be more frequent than us, go into BACKUP state. 744 */ 745 if (timercmp(&sc_tv, &ch_tv, >) || 746 timercmp(&sc_tv, &ch_tv, ==)) { 747 callout_stop(&sc->sc_ad_tmo); 748 CARP_LOG(sc, ("MASTER -> BACKUP (more frequent advertisement received)")); 749 carp_set_state(sc, BACKUP); 750 carp_setrun(sc, 0); 751 carp_setroute(sc, RTM_DELETE); 752 } 753 break; 754 case BACKUP: 755 /* 756 * If we're pre-empting masters who advertise slower than us, 757 * and this one claims to be slower, treat him as down. 758 */ 759 if (carp_opts[CARPCTL_PREEMPT] && timercmp(&sc_tv, &ch_tv, <)) { 760 CARP_LOG(sc, ("BACKUP -> MASTER (preempting a slower master)")); 761 carp_master_down(sc); 762 break; 763 } 764 765 /* 766 * If the master is going to advertise at such a low frequency 767 * that he's guaranteed to time out, we'd might as well just 768 * treat him as timed out now. 769 */ 770 sc_tv.tv_sec = sc->sc_advbase * 3; 771 if (timercmp(&sc_tv, &ch_tv, <)) { 772 CARP_LOG(sc, ("BACKUP -> MASTER (master timed out)")); 773 carp_master_down(sc); 774 break; 775 } 776 777 /* 778 * Otherwise, we reset the counter and wait for the next 779 * advertisement. 780 */ 781 carp_setrun(sc, af); 782 break; 783 } 784 785 m_freem(m); 786 return; 787 } 788 789 /* 790 * Interface side of the CARP implementation. 791 */ 792 793 /* ARGSUSED */ 794 void 795 carpattach(int n) 796 { 797 if_clone_attach(&carp_cloner); 798 799 carpstat_percpu = percpu_alloc(sizeof(uint64_t) * CARP_NSTATS); 800 } 801 802 static int 803 carp_clone_create(struct if_clone *ifc, int unit) 804 { 805 extern int ifqmaxlen; 806 struct carp_softc *sc; 807 struct ifnet *ifp; 808 809 sc = malloc(sizeof(*sc), M_DEVBUF, M_NOWAIT|M_ZERO); 810 if (!sc) 811 return (ENOMEM); 812 813 sc->sc_suppress = 0; 814 sc->sc_advbase = CARP_DFLTINTV; 815 sc->sc_vhid = -1; /* required setting */ 816 sc->sc_advskew = 0; 817 sc->sc_init_counter = 1; 818 sc->sc_naddrs = sc->sc_naddrs6 = 0; 819 #ifdef INET6 820 sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL; 821 #endif /* INET6 */ 822 823 callout_init(&sc->sc_ad_tmo, 0); 824 callout_init(&sc->sc_md_tmo, 0); 825 callout_init(&sc->sc_md6_tmo, 0); 826 827 callout_setfunc(&sc->sc_ad_tmo, carp_send_ad, sc); 828 callout_setfunc(&sc->sc_md_tmo, carp_master_down, sc); 829 callout_setfunc(&sc->sc_md6_tmo, carp_master_down, sc); 830 831 LIST_INIT(&sc->carp_mc_listhead); 832 ifp = &sc->sc_if; 833 ifp->if_softc = sc; 834 snprintf(ifp->if_xname, sizeof ifp->if_xname, "%s%d", ifc->ifc_name, 835 unit); 836 ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST; 837 ifp->if_ioctl = carp_ioctl; 838 ifp->if_start = carp_start; 839 IFQ_SET_MAXLEN(&ifp->if_snd, ifqmaxlen); 840 IFQ_SET_READY(&ifp->if_snd); 841 if_initialize(ifp); 842 ether_ifattach(ifp, NULL); 843 carp_set_enaddr(sc); 844 /* Overwrite ethernet defaults */ 845 ifp->if_type = IFT_CARP; 846 ifp->if_output = carp_output; 847 ifp->if_extflags &= ~IFEF_OUTPUT_MPSAFE; 848 if_register(ifp); 849 850 return (0); 851 } 852 853 static int 854 carp_clone_destroy(struct ifnet *ifp) 855 { 856 struct carp_softc *sc = ifp->if_softc; 857 858 carpdetach(ifp->if_softc); 859 ether_ifdetach(ifp); 860 if_detach(ifp); 861 callout_destroy(&sc->sc_ad_tmo); 862 callout_destroy(&sc->sc_md_tmo); 863 callout_destroy(&sc->sc_md6_tmo); 864 free(ifp->if_softc, M_DEVBUF); 865 866 return (0); 867 } 868 869 static void 870 carpdetach(struct carp_softc *sc) 871 { 872 struct carp_if *cif; 873 int s; 874 875 callout_stop(&sc->sc_ad_tmo); 876 callout_stop(&sc->sc_md_tmo); 877 callout_stop(&sc->sc_md6_tmo); 878 879 if (sc->sc_suppress) 880 carp_suppress_preempt--; 881 sc->sc_suppress = 0; 882 883 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) 884 carp_suppress_preempt--; 885 sc->sc_sendad_errors = 0; 886 887 carp_set_state(sc, INIT); 888 sc->sc_if.if_flags &= ~IFF_UP; 889 carp_setrun(sc, 0); 890 carp_multicast_cleanup(sc); 891 892 KERNEL_LOCK(1, NULL); 893 s = splnet(); 894 if (sc->sc_carpdev != NULL) { 895 /* XXX linkstatehook removal */ 896 cif = (struct carp_if *)sc->sc_carpdev->if_carp; 897 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list); 898 if (!--cif->vhif_nvrs) { 899 ifpromisc(sc->sc_carpdev, 0); 900 sc->sc_carpdev->if_carp = NULL; 901 free(cif, M_IFADDR); 902 } 903 } 904 sc->sc_carpdev = NULL; 905 splx(s); 906 KERNEL_UNLOCK_ONE(NULL); 907 } 908 909 /* Detach an interface from the carp. */ 910 void 911 carp_ifdetach(struct ifnet *ifp) 912 { 913 struct carp_softc *sc, *nextsc; 914 struct carp_if *cif = (struct carp_if *)ifp->if_carp; 915 916 for (sc = TAILQ_FIRST(&cif->vhif_vrs); sc; sc = nextsc) { 917 nextsc = TAILQ_NEXT(sc, sc_list); 918 carpdetach(sc); 919 } 920 } 921 922 static int 923 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, 924 struct carp_header *ch) 925 { 926 if (sc->sc_init_counter) { 927 /* this could also be seconds since unix epoch */ 928 sc->sc_counter = cprng_fast64(); 929 } else 930 sc->sc_counter++; 931 932 ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff); 933 ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff); 934 935 carp_hmac_generate(sc, ch->carp_counter, ch->carp_md); 936 937 return (0); 938 } 939 940 static void 941 carp_send_ad_all(void) 942 { 943 struct ifnet *ifp; 944 struct carp_if *cif; 945 struct carp_softc *vh; 946 int s; 947 int bound = curlwp_bind(); 948 949 s = pserialize_read_enter(); 950 IFNET_READER_FOREACH(ifp) { 951 struct psref psref; 952 if (ifp->if_carp == NULL || ifp->if_type == IFT_CARP) 953 continue; 954 955 psref_acquire(&psref, &ifp->if_psref, ifnet_psref_class); 956 pserialize_read_exit(s); 957 958 cif = (struct carp_if *)ifp->if_carp; 959 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 960 if ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 961 (IFF_UP|IFF_RUNNING) && vh->sc_state == MASTER) 962 carp_send_ad(vh); 963 } 964 965 s = pserialize_read_enter(); 966 psref_release(&psref, &ifp->if_psref, ifnet_psref_class); 967 } 968 pserialize_read_exit(s); 969 curlwp_bindx(bound); 970 } 971 972 973 static void 974 carp_send_ad(void *v) 975 { 976 struct carp_header ch; 977 struct timeval tv; 978 struct carp_softc *sc = v; 979 struct carp_header *ch_ptr; 980 struct mbuf *m; 981 int error, len, advbase, advskew, s; 982 struct sockaddr sa; 983 984 KERNEL_LOCK(1, NULL); 985 s = splsoftnet(); 986 987 advbase = advskew = 0; /* Sssssh compiler */ 988 if (sc->sc_carpdev == NULL) { 989 sc->sc_if.if_oerrors++; 990 goto retry_later; 991 } 992 993 /* bow out if we've gone to backup (the carp interface is going down) */ 994 if (sc->sc_bow_out) { 995 sc->sc_bow_out = 0; 996 advbase = 255; 997 advskew = 255; 998 } else { 999 advbase = sc->sc_advbase; 1000 if (!carp_suppress_preempt || sc->sc_advskew > 240) 1001 advskew = sc->sc_advskew; 1002 else 1003 advskew = 240; 1004 tv.tv_sec = advbase; 1005 tv.tv_usec = advskew * 1000000 / 256; 1006 } 1007 1008 ch.carp_version = CARP_VERSION; 1009 ch.carp_type = CARP_ADVERTISEMENT; 1010 ch.carp_vhid = sc->sc_vhid; 1011 ch.carp_advbase = advbase; 1012 ch.carp_advskew = advskew; 1013 ch.carp_authlen = 7; /* XXX DEFINE */ 1014 ch.carp_pad1 = 0; /* must be zero */ 1015 ch.carp_cksum = 0; 1016 1017 1018 #ifdef INET 1019 if (sc->sc_naddrs) { 1020 struct ip *ip; 1021 struct ifaddr *ifa; 1022 int _s; 1023 1024 MGETHDR(m, M_DONTWAIT, MT_HEADER); 1025 if (m == NULL) { 1026 sc->sc_if.if_oerrors++; 1027 CARP_STATINC(CARP_STAT_ONOMEM); 1028 /* XXX maybe less ? */ 1029 goto retry_later; 1030 } 1031 MCLAIM(m, &carp_proto_mowner_tx); 1032 len = sizeof(*ip) + sizeof(ch); 1033 m->m_pkthdr.len = len; 1034 m_reset_rcvif(m); 1035 m->m_len = len; 1036 MH_ALIGN(m, m->m_len); 1037 m->m_flags |= M_MCAST; 1038 ip = mtod(m, struct ip *); 1039 ip->ip_v = IPVERSION; 1040 ip->ip_hl = sizeof(*ip) >> 2; 1041 ip->ip_tos = IPTOS_LOWDELAY; 1042 ip->ip_len = htons(len); 1043 ip->ip_id = 0; /* no need for id, we don't support fragments */ 1044 ip->ip_off = htons(IP_DF); 1045 ip->ip_ttl = CARP_DFLTTL; 1046 ip->ip_p = IPPROTO_CARP; 1047 ip->ip_sum = 0; 1048 1049 memset(&sa, 0, sizeof(sa)); 1050 sa.sa_family = AF_INET; 1051 _s = pserialize_read_enter(); 1052 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev); 1053 if (ifa == NULL) 1054 ip->ip_src.s_addr = 0; 1055 else 1056 ip->ip_src.s_addr = 1057 ifatoia(ifa)->ia_addr.sin_addr.s_addr; 1058 pserialize_read_exit(_s); 1059 ip->ip_dst.s_addr = INADDR_CARP_GROUP; 1060 1061 ch_ptr = (struct carp_header *)(&ip[1]); 1062 memcpy(ch_ptr, &ch, sizeof(ch)); 1063 if (carp_prepare_ad(m, sc, ch_ptr)) 1064 goto retry_later; 1065 1066 m->m_data += sizeof(*ip); 1067 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip)); 1068 m->m_data -= sizeof(*ip); 1069 1070 nanotime(&sc->sc_if.if_lastchange); 1071 sc->sc_if.if_opackets++; 1072 sc->sc_if.if_obytes += len; 1073 CARP_STATINC(CARP_STAT_OPACKETS); 1074 1075 error = ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, 1076 NULL); 1077 if (error) { 1078 if (error == ENOBUFS) 1079 CARP_STATINC(CARP_STAT_ONOMEM); 1080 else 1081 CARP_LOG(sc, ("ip_output failed: %d", error)); 1082 sc->sc_if.if_oerrors++; 1083 if (sc->sc_sendad_errors < INT_MAX) 1084 sc->sc_sendad_errors++; 1085 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) { 1086 carp_suppress_preempt++; 1087 if (carp_suppress_preempt == 1) 1088 carp_send_ad_all(); 1089 } 1090 sc->sc_sendad_success = 0; 1091 } else { 1092 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) { 1093 if (++sc->sc_sendad_success >= 1094 CARP_SENDAD_MIN_SUCCESS) { 1095 carp_suppress_preempt--; 1096 sc->sc_sendad_errors = 0; 1097 } 1098 } else 1099 sc->sc_sendad_errors = 0; 1100 } 1101 } 1102 #endif /* INET */ 1103 #ifdef INET6_notyet 1104 if (sc->sc_naddrs6) { 1105 struct ip6_hdr *ip6; 1106 struct ifaddr *ifa; 1107 int _s; 1108 1109 MGETHDR(m, M_DONTWAIT, MT_HEADER); 1110 if (m == NULL) { 1111 sc->sc_if.if_oerrors++; 1112 CARP_STATINC(CARP_STAT_ONOMEM); 1113 /* XXX maybe less ? */ 1114 goto retry_later; 1115 } 1116 MCLAIM(m, &carp_proto6_mowner_tx); 1117 len = sizeof(*ip6) + sizeof(ch); 1118 m->m_pkthdr.len = len; 1119 m_reset_rcvif(m); 1120 m->m_len = len; 1121 MH_ALIGN(m, m->m_len); 1122 m->m_flags |= M_MCAST; 1123 ip6 = mtod(m, struct ip6_hdr *); 1124 memset(ip6, 0, sizeof(*ip6)); 1125 ip6->ip6_vfc |= IPV6_VERSION; 1126 ip6->ip6_hlim = CARP_DFLTTL; 1127 ip6->ip6_nxt = IPPROTO_CARP; 1128 1129 /* set the source address */ 1130 memset(&sa, 0, sizeof(sa)); 1131 sa.sa_family = AF_INET6; 1132 _s = pserialize_read_enter(); 1133 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev); 1134 if (ifa == NULL) /* This should never happen with IPv6 */ 1135 memset(&ip6->ip6_src, 0, sizeof(struct in6_addr)); 1136 else 1137 bcopy(ifatoia6(ifa)->ia_addr.sin6_addr.s6_addr, 1138 &ip6->ip6_src, sizeof(struct in6_addr)); 1139 pserialize_read_exit(_s); 1140 /* set the multicast destination */ 1141 1142 ip6->ip6_dst.s6_addr16[0] = htons(0xff02); 1143 ip6->ip6_dst.s6_addr8[15] = 0x12; 1144 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) { 1145 sc->sc_if.if_oerrors++; 1146 m_freem(m); 1147 CARP_LOG(sc, ("in6_setscope failed")); 1148 goto retry_later; 1149 } 1150 1151 ch_ptr = (struct carp_header *)(&ip6[1]); 1152 memcpy(ch_ptr, &ch, sizeof(ch)); 1153 if (carp_prepare_ad(m, sc, ch_ptr)) 1154 goto retry_later; 1155 1156 m->m_data += sizeof(*ip6); 1157 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6)); 1158 m->m_data -= sizeof(*ip6); 1159 1160 nanotime(&sc->sc_if.if_lastchange); 1161 sc->sc_if.if_opackets++; 1162 sc->sc_if.if_obytes += len; 1163 CARP_STATINC(CARP_STAT_OPACKETS6); 1164 1165 error = ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL); 1166 if (error) { 1167 if (error == ENOBUFS) 1168 CARP_STATINC(CARP_STAT_ONOMEM); 1169 else 1170 CARP_LOG(sc, ("ip6_output failed: %d", error)); 1171 sc->sc_if.if_oerrors++; 1172 if (sc->sc_sendad_errors < INT_MAX) 1173 sc->sc_sendad_errors++; 1174 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) { 1175 carp_suppress_preempt++; 1176 if (carp_suppress_preempt == 1) 1177 carp_send_ad_all(); 1178 } 1179 sc->sc_sendad_success = 0; 1180 } else { 1181 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) { 1182 if (++sc->sc_sendad_success >= 1183 CARP_SENDAD_MIN_SUCCESS) { 1184 carp_suppress_preempt--; 1185 sc->sc_sendad_errors = 0; 1186 } 1187 } else 1188 sc->sc_sendad_errors = 0; 1189 } 1190 } 1191 #endif /* INET6 */ 1192 1193 retry_later: 1194 splx(s); 1195 KERNEL_UNLOCK_ONE(NULL); 1196 if (advbase != 255 || advskew != 255) 1197 callout_schedule(&sc->sc_ad_tmo, tvtohz(&tv)); 1198 } 1199 1200 /* 1201 * Broadcast a gratuitous ARP request containing 1202 * the virtual router MAC address for each IP address 1203 * associated with the virtual router. 1204 */ 1205 static void 1206 carp_send_arp(struct carp_softc *sc) 1207 { 1208 struct ifaddr *ifa; 1209 int s; 1210 1211 KERNEL_LOCK(1, NULL); 1212 s = splsoftnet(); 1213 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 1214 1215 if (ifa->ifa_addr->sa_family != AF_INET) 1216 continue; 1217 1218 arpannounce(sc->sc_carpdev, ifa, CLLADDR(sc->sc_if.if_sadl)); 1219 } 1220 splx(s); 1221 KERNEL_UNLOCK_ONE(NULL); 1222 } 1223 1224 #ifdef INET6 1225 static void 1226 carp_send_na(struct carp_softc *sc) 1227 { 1228 struct ifaddr *ifa; 1229 struct in6_addr *in6; 1230 static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT; 1231 int s; 1232 1233 KERNEL_LOCK(1, NULL); 1234 s = splsoftnet(); 1235 1236 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 1237 1238 if (ifa->ifa_addr->sa_family != AF_INET6) 1239 continue; 1240 1241 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr; 1242 nd6_na_output(sc->sc_carpdev, &mcast, in6, 1243 ND_NA_FLAG_OVERRIDE, 1, NULL); 1244 } 1245 splx(s); 1246 KERNEL_UNLOCK_ONE(NULL); 1247 } 1248 #endif /* INET6 */ 1249 1250 /* 1251 * Based on bridge_hash() in if_bridge.c 1252 */ 1253 #define mix(a,b,c) \ 1254 do { \ 1255 a -= b; a -= c; a ^= (c >> 13); \ 1256 b -= c; b -= a; b ^= (a << 8); \ 1257 c -= a; c -= b; c ^= (b >> 13); \ 1258 a -= b; a -= c; a ^= (c >> 12); \ 1259 b -= c; b -= a; b ^= (a << 16); \ 1260 c -= a; c -= b; c ^= (b >> 5); \ 1261 a -= b; a -= c; a ^= (c >> 3); \ 1262 b -= c; b -= a; b ^= (a << 10); \ 1263 c -= a; c -= b; c ^= (b >> 15); \ 1264 } while (0) 1265 1266 static u_int32_t 1267 carp_hash(struct carp_softc *sc, u_char *src) 1268 { 1269 u_int32_t a = 0x9e3779b9, b = sc->sc_hashkey[0], c = sc->sc_hashkey[1]; 1270 1271 c += sc->sc_key[3] << 24; 1272 c += sc->sc_key[2] << 16; 1273 c += sc->sc_key[1] << 8; 1274 c += sc->sc_key[0]; 1275 b += src[5] << 8; 1276 b += src[4]; 1277 a += src[3] << 24; 1278 a += src[2] << 16; 1279 a += src[1] << 8; 1280 a += src[0]; 1281 1282 mix(a, b, c); 1283 return (c); 1284 } 1285 1286 static int 1287 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type) 1288 { 1289 struct carp_softc *vh; 1290 struct ifaddr *ifa; 1291 int count = 0; 1292 1293 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1294 if ((type == CARP_COUNT_RUNNING && 1295 (vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1296 (IFF_UP|IFF_RUNNING)) || 1297 (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) { 1298 IFADDR_READER_FOREACH(ifa, &vh->sc_if) { 1299 if (ifa->ifa_addr->sa_family == AF_INET && 1300 ia->ia_addr.sin_addr.s_addr == 1301 ifatoia(ifa)->ia_addr.sin_addr.s_addr) 1302 count++; 1303 } 1304 } 1305 } 1306 return (count); 1307 } 1308 1309 int 1310 carp_iamatch(struct in_ifaddr *ia, u_char *src, 1311 u_int32_t *count, u_int32_t index) 1312 { 1313 struct carp_softc *sc = ia->ia_ifp->if_softc; 1314 1315 if (carp_opts[CARPCTL_ARPBALANCE]) { 1316 /* 1317 * We use the source ip to decide which virtual host should 1318 * handle the request. If we're master of that virtual host, 1319 * then we respond, otherwise, just drop the arp packet on 1320 * the floor. 1321 */ 1322 1323 /* Count the elegible carp interfaces with this address */ 1324 if (*count == 0) 1325 *count = carp_addrcount( 1326 (struct carp_if *)ia->ia_ifp->if_carpdev->if_carp, 1327 ia, CARP_COUNT_RUNNING); 1328 1329 /* This should never happen, but... */ 1330 if (*count == 0) 1331 return (0); 1332 1333 if (carp_hash(sc, src) % *count == index - 1 && 1334 sc->sc_state == MASTER) { 1335 return (1); 1336 } 1337 } else { 1338 if (sc->sc_state == MASTER) 1339 return (1); 1340 } 1341 1342 return (0); 1343 } 1344 1345 #ifdef INET6 1346 struct ifaddr * 1347 carp_iamatch6(void *v, struct in6_addr *taddr) 1348 { 1349 struct carp_if *cif = v; 1350 struct carp_softc *vh; 1351 struct ifaddr *ifa; 1352 1353 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1354 IFADDR_READER_FOREACH(ifa, &vh->sc_if) { 1355 if (IN6_ARE_ADDR_EQUAL(taddr, 1356 &ifatoia6(ifa)->ia_addr.sin6_addr) && 1357 ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1358 (IFF_UP|IFF_RUNNING)) && vh->sc_state == MASTER) 1359 return (ifa); 1360 } 1361 } 1362 1363 return (NULL); 1364 } 1365 #endif /* INET6 */ 1366 1367 struct ifnet * 1368 carp_ourether(void *v, struct ether_header *eh, u_char iftype, int src) 1369 { 1370 struct carp_if *cif = (struct carp_if *)v; 1371 struct carp_softc *vh; 1372 u_int8_t *ena; 1373 1374 if (src) 1375 ena = (u_int8_t *)&eh->ether_shost; 1376 else 1377 ena = (u_int8_t *)&eh->ether_dhost; 1378 1379 switch (iftype) { 1380 case IFT_ETHER: 1381 case IFT_FDDI: 1382 if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1) 1383 return (NULL); 1384 break; 1385 case IFT_ISO88025: 1386 if (ena[0] != 3 || ena[1] || ena[4] || ena[5]) 1387 return (NULL); 1388 break; 1389 default: 1390 return (NULL); 1391 break; 1392 } 1393 1394 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) 1395 if ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1396 (IFF_UP|IFF_RUNNING) && vh->sc_state == MASTER && 1397 !memcmp(ena, CLLADDR(vh->sc_if.if_sadl), 1398 ETHER_ADDR_LEN)) { 1399 return (&vh->sc_if); 1400 } 1401 1402 return (NULL); 1403 } 1404 1405 int 1406 carp_input(struct mbuf *m, u_int8_t *shost, u_int8_t *dhost, u_int16_t etype) 1407 { 1408 struct ether_header eh; 1409 struct carp_if *cif = (struct carp_if *)m_get_rcvif_NOMPSAFE(m)->if_carp; 1410 struct ifnet *ifp; 1411 1412 memcpy(&eh.ether_shost, shost, sizeof(eh.ether_shost)); 1413 memcpy(&eh.ether_dhost, dhost, sizeof(eh.ether_dhost)); 1414 eh.ether_type = etype; 1415 1416 if (m->m_flags & (M_BCAST|M_MCAST)) { 1417 struct carp_softc *vh; 1418 struct mbuf *m0; 1419 1420 /* 1421 * XXX Should really check the list of multicast addresses 1422 * for each CARP interface _before_ copying. 1423 */ 1424 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1425 m0 = m_copym(m, 0, M_COPYALL, M_DONTWAIT); 1426 if (m0 == NULL) 1427 continue; 1428 m_set_rcvif(m0, &vh->sc_if); 1429 ether_input(&vh->sc_if, m0); 1430 } 1431 return (1); 1432 } 1433 1434 ifp = carp_ourether(cif, &eh, m_get_rcvif_NOMPSAFE(m)->if_type, 0); 1435 if (ifp == NULL) { 1436 return (1); 1437 } 1438 1439 m_set_rcvif(m, ifp); 1440 1441 bpf_mtap(ifp, m); 1442 ifp->if_ipackets++; 1443 ether_input(ifp, m); 1444 return (0); 1445 } 1446 1447 static void 1448 carp_master_down(void *v) 1449 { 1450 struct carp_softc *sc = v; 1451 1452 switch (sc->sc_state) { 1453 case INIT: 1454 printf("%s: master_down event in INIT state\n", 1455 sc->sc_if.if_xname); 1456 break; 1457 case MASTER: 1458 break; 1459 case BACKUP: 1460 CARP_LOG(sc, ("INIT -> MASTER (preempting)")); 1461 carp_set_state(sc, MASTER); 1462 carp_send_ad(sc); 1463 carp_send_arp(sc); 1464 #ifdef INET6 1465 carp_send_na(sc); 1466 #endif /* INET6 */ 1467 carp_setrun(sc, 0); 1468 carp_setroute(sc, RTM_ADD); 1469 break; 1470 } 1471 } 1472 1473 /* 1474 * When in backup state, af indicates whether to reset the master down timer 1475 * for v4 or v6. If it's set to zero, reset the ones which are already pending. 1476 */ 1477 static void 1478 carp_setrun(struct carp_softc *sc, sa_family_t af) 1479 { 1480 struct timeval tv; 1481 1482 if (sc->sc_carpdev == NULL) { 1483 sc->sc_if.if_flags &= ~IFF_RUNNING; 1484 carp_set_state(sc, INIT); 1485 return; 1486 } 1487 1488 if (sc->sc_if.if_flags & IFF_UP && sc->sc_vhid > 0 && 1489 (sc->sc_naddrs || sc->sc_naddrs6) && !sc->sc_suppress) { 1490 sc->sc_if.if_flags |= IFF_RUNNING; 1491 } else { 1492 sc->sc_if.if_flags &= ~IFF_RUNNING; 1493 carp_setroute(sc, RTM_DELETE); 1494 return; 1495 } 1496 1497 switch (sc->sc_state) { 1498 case INIT: 1499 carp_set_state(sc, BACKUP); 1500 carp_setroute(sc, RTM_DELETE); 1501 carp_setrun(sc, 0); 1502 break; 1503 case BACKUP: 1504 callout_stop(&sc->sc_ad_tmo); 1505 tv.tv_sec = 3 * sc->sc_advbase; 1506 tv.tv_usec = sc->sc_advskew * 1000000 / 256; 1507 switch (af) { 1508 #ifdef INET 1509 case AF_INET: 1510 callout_schedule(&sc->sc_md_tmo, tvtohz(&tv)); 1511 break; 1512 #endif /* INET */ 1513 #ifdef INET6_notyet 1514 case AF_INET6: 1515 callout_schedule(&sc->sc_md6_tmo, tvtohz(&tv)); 1516 break; 1517 #endif /* INET6 */ 1518 default: 1519 if (sc->sc_naddrs) 1520 callout_schedule(&sc->sc_md_tmo, tvtohz(&tv)); 1521 #ifdef INET6_notyet 1522 if (sc->sc_naddrs6) 1523 callout_schedule(&sc->sc_md6_tmo, tvtohz(&tv)); 1524 #endif /* INET6 */ 1525 break; 1526 } 1527 break; 1528 case MASTER: 1529 tv.tv_sec = sc->sc_advbase; 1530 tv.tv_usec = sc->sc_advskew * 1000000 / 256; 1531 callout_schedule(&sc->sc_ad_tmo, tvtohz(&tv)); 1532 break; 1533 } 1534 } 1535 1536 static void 1537 carp_multicast_cleanup(struct carp_softc *sc) 1538 { 1539 struct ip_moptions *imo = &sc->sc_imo; 1540 #ifdef INET6 1541 struct ip6_moptions *im6o = &sc->sc_im6o; 1542 #endif 1543 u_int16_t n = imo->imo_num_memberships; 1544 1545 /* Clean up our own multicast memberships */ 1546 while (n-- > 0) { 1547 if (imo->imo_membership[n] != NULL) { 1548 in_delmulti(imo->imo_membership[n]); 1549 imo->imo_membership[n] = NULL; 1550 } 1551 } 1552 imo->imo_num_memberships = 0; 1553 imo->imo_multicast_if_index = 0; 1554 1555 #ifdef INET6 1556 while (!LIST_EMPTY(&im6o->im6o_memberships)) { 1557 struct in6_multi_mship *imm = 1558 LIST_FIRST(&im6o->im6o_memberships); 1559 1560 LIST_REMOVE(imm, i6mm_chain); 1561 in6_leavegroup(imm); 1562 } 1563 im6o->im6o_multicast_if_index = 0; 1564 #endif 1565 1566 /* And any other multicast memberships */ 1567 carp_ether_purgemulti(sc); 1568 } 1569 1570 static int 1571 carp_set_ifp(struct carp_softc *sc, struct ifnet *ifp) 1572 { 1573 struct carp_if *cif, *ncif = NULL; 1574 struct carp_softc *vr, *after = NULL; 1575 int myself = 0, error = 0; 1576 int s; 1577 1578 if (ifp == sc->sc_carpdev) 1579 return (0); 1580 1581 if (ifp != NULL) { 1582 if ((ifp->if_flags & IFF_MULTICAST) == 0) 1583 return (EADDRNOTAVAIL); 1584 1585 if (ifp->if_type == IFT_CARP) 1586 return (EINVAL); 1587 1588 if (ifp->if_carp == NULL) { 1589 ncif = malloc(sizeof(*cif), M_IFADDR, M_NOWAIT); 1590 if (ncif == NULL) 1591 return (ENOBUFS); 1592 if ((error = ifpromisc(ifp, 1))) { 1593 free(ncif, M_IFADDR); 1594 return (error); 1595 } 1596 1597 ncif->vhif_ifp = ifp; 1598 TAILQ_INIT(&ncif->vhif_vrs); 1599 } else { 1600 cif = (struct carp_if *)ifp->if_carp; 1601 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) 1602 if (vr != sc && vr->sc_vhid == sc->sc_vhid) 1603 return (EINVAL); 1604 } 1605 1606 /* detach from old interface */ 1607 if (sc->sc_carpdev != NULL) 1608 carpdetach(sc); 1609 1610 /* join multicast groups */ 1611 if (sc->sc_naddrs < 0 && 1612 (error = carp_join_multicast(sc)) != 0) { 1613 if (ncif != NULL) 1614 free(ncif, M_IFADDR); 1615 return (error); 1616 } 1617 1618 #ifdef INET6 1619 if (sc->sc_naddrs6 < 0 && 1620 (error = carp_join_multicast6(sc)) != 0) { 1621 if (ncif != NULL) 1622 free(ncif, M_IFADDR); 1623 carp_multicast_cleanup(sc); 1624 return (error); 1625 } 1626 #endif 1627 1628 /* attach carp interface to physical interface */ 1629 if (ncif != NULL) 1630 ifp->if_carp = (void *)ncif; 1631 sc->sc_carpdev = ifp; 1632 sc->sc_if.if_capabilities = ifp->if_capabilities & 1633 (IFCAP_TSOv4 | IFCAP_TSOv6 | 1634 IFCAP_CSUM_IPv4_Tx|IFCAP_CSUM_IPv4_Rx| 1635 IFCAP_CSUM_TCPv4_Tx|IFCAP_CSUM_TCPv4_Rx| 1636 IFCAP_CSUM_UDPv4_Tx|IFCAP_CSUM_UDPv4_Rx| 1637 IFCAP_CSUM_TCPv6_Tx|IFCAP_CSUM_TCPv6_Rx| 1638 IFCAP_CSUM_UDPv6_Tx|IFCAP_CSUM_UDPv6_Rx); 1639 1640 cif = (struct carp_if *)ifp->if_carp; 1641 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) { 1642 if (vr == sc) 1643 myself = 1; 1644 if (vr->sc_vhid < sc->sc_vhid) 1645 after = vr; 1646 } 1647 1648 if (!myself) { 1649 /* We're trying to keep things in order */ 1650 if (after == NULL) { 1651 TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list); 1652 } else { 1653 TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, 1654 sc, sc_list); 1655 } 1656 cif->vhif_nvrs++; 1657 } 1658 if (sc->sc_naddrs || sc->sc_naddrs6) 1659 sc->sc_if.if_flags |= IFF_UP; 1660 carp_set_enaddr(sc); 1661 KERNEL_LOCK(1, NULL); 1662 s = splnet(); 1663 /* XXX linkstatehooks establish */ 1664 carp_carpdev_state(ifp); 1665 splx(s); 1666 KERNEL_UNLOCK_ONE(NULL); 1667 } else { 1668 carpdetach(sc); 1669 sc->sc_if.if_flags &= ~(IFF_UP|IFF_RUNNING); 1670 } 1671 return (0); 1672 } 1673 1674 static void 1675 carp_set_enaddr(struct carp_softc *sc) 1676 { 1677 uint8_t enaddr[ETHER_ADDR_LEN]; 1678 if (sc->sc_carpdev && sc->sc_carpdev->if_type == IFT_ISO88025) { 1679 enaddr[0] = 3; 1680 enaddr[1] = 0; 1681 enaddr[2] = 0x40 >> (sc->sc_vhid - 1); 1682 enaddr[3] = 0x40000 >> (sc->sc_vhid - 1); 1683 enaddr[4] = 0; 1684 enaddr[5] = 0; 1685 } else { 1686 enaddr[0] = 0; 1687 enaddr[1] = 0; 1688 enaddr[2] = 0x5e; 1689 enaddr[3] = 0; 1690 enaddr[4] = 1; 1691 enaddr[5] = sc->sc_vhid; 1692 } 1693 if_set_sadl(&sc->sc_if, enaddr, sizeof(enaddr), false); 1694 } 1695 1696 #if 0 1697 static void 1698 carp_addr_updated(void *v) 1699 { 1700 struct carp_softc *sc = (struct carp_softc *) v; 1701 struct ifaddr *ifa; 1702 int new_naddrs = 0, new_naddrs6 = 0; 1703 1704 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 1705 if (ifa->ifa_addr->sa_family == AF_INET) 1706 new_naddrs++; 1707 else if (ifa->ifa_addr->sa_family == AF_INET6) 1708 new_naddrs6++; 1709 } 1710 1711 /* Handle a callback after SIOCDIFADDR */ 1712 if (new_naddrs < sc->sc_naddrs || new_naddrs6 < sc->sc_naddrs6) { 1713 struct in_addr mc_addr; 1714 1715 sc->sc_naddrs = new_naddrs; 1716 sc->sc_naddrs6 = new_naddrs6; 1717 1718 /* Re-establish multicast membership removed by in_control */ 1719 mc_addr.s_addr = INADDR_CARP_GROUP; 1720 if (!in_multi_group(mc_addr, &sc->sc_if, 0)) { 1721 memset(&sc->sc_imo, 0, sizeof(sc->sc_imo)); 1722 1723 if (sc->sc_carpdev != NULL && sc->sc_naddrs > 0) 1724 carp_join_multicast(sc); 1725 } 1726 1727 if (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) { 1728 sc->sc_if.if_flags &= ~IFF_UP; 1729 carp_set_state(sc, INIT); 1730 } else 1731 carp_hmac_prepare(sc); 1732 } 1733 1734 carp_setrun(sc, 0); 1735 } 1736 #endif 1737 1738 static int 1739 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin) 1740 { 1741 struct ifnet *ifp = sc->sc_carpdev; 1742 struct in_ifaddr *ia, *ia_if; 1743 int error = 0; 1744 int s; 1745 1746 if (sin->sin_addr.s_addr == 0) { 1747 if (!(sc->sc_if.if_flags & IFF_UP)) 1748 carp_set_state(sc, INIT); 1749 if (sc->sc_naddrs) 1750 sc->sc_if.if_flags |= IFF_UP; 1751 carp_setrun(sc, 0); 1752 return (0); 1753 } 1754 1755 /* we have to do this by hand to ensure we don't match on ourselves */ 1756 ia_if = NULL; 1757 s = pserialize_read_enter(); 1758 IN_ADDRLIST_READER_FOREACH(ia) { 1759 /* and, yeah, we need a multicast-capable iface too */ 1760 if (ia->ia_ifp != &sc->sc_if && 1761 ia->ia_ifp->if_type != IFT_CARP && 1762 (ia->ia_ifp->if_flags & IFF_MULTICAST) && 1763 (sin->sin_addr.s_addr & ia->ia_subnetmask) == 1764 ia->ia_subnet) { 1765 if (!ia_if) 1766 ia_if = ia; 1767 } 1768 } 1769 1770 if (ia_if) { 1771 ia = ia_if; 1772 if (ifp) { 1773 if (ifp != ia->ia_ifp) 1774 return (EADDRNOTAVAIL); 1775 } else { 1776 /* FIXME NOMPSAFE */ 1777 ifp = ia->ia_ifp; 1778 } 1779 } 1780 pserialize_read_exit(s); 1781 1782 if ((error = carp_set_ifp(sc, ifp))) 1783 return (error); 1784 1785 if (sc->sc_carpdev == NULL) 1786 return (EADDRNOTAVAIL); 1787 1788 if (sc->sc_naddrs == 0 && (error = carp_join_multicast(sc)) != 0) 1789 return (error); 1790 1791 sc->sc_naddrs++; 1792 if (sc->sc_carpdev != NULL) 1793 sc->sc_if.if_flags |= IFF_UP; 1794 1795 carp_set_state(sc, INIT); 1796 carp_setrun(sc, 0); 1797 1798 /* 1799 * Hook if_addrhooks so that we get a callback after in_ifinit has run, 1800 * to correct any inappropriate routes that it inserted. 1801 */ 1802 if (sc->ah_cookie == 0) { 1803 /* XXX link address hook */ 1804 } 1805 1806 return (0); 1807 } 1808 1809 static int 1810 carp_join_multicast(struct carp_softc *sc) 1811 { 1812 struct ip_moptions *imo = &sc->sc_imo, tmpimo; 1813 struct in_addr addr; 1814 1815 memset(&tmpimo, 0, sizeof(tmpimo)); 1816 addr.s_addr = INADDR_CARP_GROUP; 1817 if ((tmpimo.imo_membership[0] = 1818 in_addmulti(&addr, &sc->sc_if)) == NULL) { 1819 return (ENOBUFS); 1820 } 1821 1822 imo->imo_membership[0] = tmpimo.imo_membership[0]; 1823 imo->imo_num_memberships = 1; 1824 imo->imo_multicast_if_index = sc->sc_if.if_index; 1825 imo->imo_multicast_ttl = CARP_DFLTTL; 1826 imo->imo_multicast_loop = 0; 1827 return (0); 1828 } 1829 1830 1831 #ifdef INET6 1832 static int 1833 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6) 1834 { 1835 struct ifnet *ifp = sc->sc_carpdev; 1836 struct in6_ifaddr *ia, *ia_if; 1837 int error = 0; 1838 int s; 1839 1840 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) { 1841 if (!(sc->sc_if.if_flags & IFF_UP)) 1842 carp_set_state(sc, INIT); 1843 if (sc->sc_naddrs6) 1844 sc->sc_if.if_flags |= IFF_UP; 1845 carp_setrun(sc, 0); 1846 return (0); 1847 } 1848 1849 /* we have to do this by hand to ensure we don't match on ourselves */ 1850 ia_if = NULL; 1851 s = pserialize_read_enter(); 1852 IN6_ADDRLIST_READER_FOREACH(ia) { 1853 int i; 1854 1855 for (i = 0; i < 4; i++) { 1856 if ((sin6->sin6_addr.s6_addr32[i] & 1857 ia->ia_prefixmask.sin6_addr.s6_addr32[i]) != 1858 (ia->ia_addr.sin6_addr.s6_addr32[i] & 1859 ia->ia_prefixmask.sin6_addr.s6_addr32[i])) 1860 break; 1861 } 1862 /* and, yeah, we need a multicast-capable iface too */ 1863 if (ia->ia_ifp != &sc->sc_if && 1864 ia->ia_ifp->if_type != IFT_CARP && 1865 (ia->ia_ifp->if_flags & IFF_MULTICAST) && 1866 (i == 4)) { 1867 if (!ia_if) 1868 ia_if = ia; 1869 } 1870 } 1871 pserialize_read_exit(s); 1872 1873 if (ia_if) { 1874 ia = ia_if; 1875 if (sc->sc_carpdev) { 1876 if (sc->sc_carpdev != ia->ia_ifp) 1877 return (EADDRNOTAVAIL); 1878 } else { 1879 ifp = ia->ia_ifp; 1880 } 1881 } 1882 1883 if ((error = carp_set_ifp(sc, ifp))) 1884 return (error); 1885 1886 if (sc->sc_carpdev == NULL) 1887 return (EADDRNOTAVAIL); 1888 1889 if (sc->sc_naddrs6 == 0 && (error = carp_join_multicast6(sc)) != 0) 1890 return (error); 1891 1892 sc->sc_naddrs6++; 1893 if (sc->sc_carpdev != NULL) 1894 sc->sc_if.if_flags |= IFF_UP; 1895 carp_set_state(sc, INIT); 1896 carp_setrun(sc, 0); 1897 1898 return (0); 1899 } 1900 1901 static int 1902 carp_join_multicast6(struct carp_softc *sc) 1903 { 1904 struct in6_multi_mship *imm, *imm2; 1905 struct ip6_moptions *im6o = &sc->sc_im6o; 1906 struct sockaddr_in6 addr6; 1907 int error; 1908 1909 /* Join IPv6 CARP multicast group */ 1910 memset(&addr6, 0, sizeof(addr6)); 1911 addr6.sin6_family = AF_INET6; 1912 addr6.sin6_len = sizeof(addr6); 1913 addr6.sin6_addr.s6_addr16[0] = htons(0xff02); 1914 addr6.sin6_addr.s6_addr16[1] = htons(sc->sc_if.if_index); 1915 addr6.sin6_addr.s6_addr8[15] = 0x12; 1916 if ((imm = in6_joingroup(&sc->sc_if, 1917 &addr6.sin6_addr, &error, 0)) == NULL) { 1918 return (error); 1919 } 1920 /* join solicited multicast address */ 1921 memset(&addr6.sin6_addr, 0, sizeof(addr6.sin6_addr)); 1922 addr6.sin6_addr.s6_addr16[0] = htons(0xff02); 1923 addr6.sin6_addr.s6_addr16[1] = htons(sc->sc_if.if_index); 1924 addr6.sin6_addr.s6_addr32[1] = 0; 1925 addr6.sin6_addr.s6_addr32[2] = htonl(1); 1926 addr6.sin6_addr.s6_addr32[3] = 0; 1927 addr6.sin6_addr.s6_addr8[12] = 0xff; 1928 if ((imm2 = in6_joingroup(&sc->sc_if, 1929 &addr6.sin6_addr, &error, 0)) == NULL) { 1930 in6_leavegroup(imm); 1931 return (error); 1932 } 1933 1934 /* apply v6 multicast membership */ 1935 im6o->im6o_multicast_if_index = sc->sc_if.if_index; 1936 if (imm) 1937 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, 1938 i6mm_chain); 1939 if (imm2) 1940 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm2, 1941 i6mm_chain); 1942 1943 return (0); 1944 } 1945 1946 #endif /* INET6 */ 1947 1948 static int 1949 carp_ioctl(struct ifnet *ifp, u_long cmd, void *data) 1950 { 1951 struct lwp *l = curlwp; /* XXX */ 1952 struct carp_softc *sc = ifp->if_softc, *vr; 1953 struct carpreq carpr; 1954 struct ifaddr *ifa; 1955 struct ifreq *ifr; 1956 struct ifnet *cdev = NULL; 1957 int error = 0; 1958 1959 ifa = (struct ifaddr *)data; 1960 ifr = (struct ifreq *)data; 1961 1962 switch (cmd) { 1963 case SIOCINITIFADDR: 1964 switch (ifa->ifa_addr->sa_family) { 1965 #ifdef INET 1966 case AF_INET: 1967 sc->sc_if.if_flags |= IFF_UP; 1968 memcpy(ifa->ifa_dstaddr, ifa->ifa_addr, 1969 sizeof(struct sockaddr)); 1970 error = carp_set_addr(sc, satosin(ifa->ifa_addr)); 1971 break; 1972 #endif /* INET */ 1973 #ifdef INET6 1974 case AF_INET6: 1975 sc->sc_if.if_flags|= IFF_UP; 1976 error = carp_set_addr6(sc, satosin6(ifa->ifa_addr)); 1977 break; 1978 #endif /* INET6 */ 1979 default: 1980 error = EAFNOSUPPORT; 1981 break; 1982 } 1983 break; 1984 1985 case SIOCSIFFLAGS: 1986 if ((error = ifioctl_common(ifp, cmd, data)) != 0) 1987 break; 1988 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) { 1989 callout_stop(&sc->sc_ad_tmo); 1990 callout_stop(&sc->sc_md_tmo); 1991 callout_stop(&sc->sc_md6_tmo); 1992 if (sc->sc_state == MASTER) { 1993 /* we need the interface up to bow out */ 1994 sc->sc_if.if_flags |= IFF_UP; 1995 sc->sc_bow_out = 1; 1996 carp_send_ad(sc); 1997 } 1998 sc->sc_if.if_flags &= ~IFF_UP; 1999 carp_set_state(sc, INIT); 2000 carp_setrun(sc, 0); 2001 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) { 2002 sc->sc_if.if_flags |= IFF_UP; 2003 carp_setrun(sc, 0); 2004 } 2005 break; 2006 2007 case SIOCSVH: 2008 if (l == NULL) 2009 break; 2010 if ((error = kauth_authorize_network(l->l_cred, 2011 KAUTH_NETWORK_INTERFACE, 2012 KAUTH_REQ_NETWORK_INTERFACE_SETPRIV, ifp, (void *)cmd, 2013 NULL)) != 0) 2014 break; 2015 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr))) 2016 break; 2017 error = 1; 2018 if (carpr.carpr_carpdev[0] != '\0' && 2019 (cdev = ifunit(carpr.carpr_carpdev)) == NULL) 2020 return (EINVAL); 2021 if ((error = carp_set_ifp(sc, cdev))) 2022 return (error); 2023 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) { 2024 switch (carpr.carpr_state) { 2025 case BACKUP: 2026 callout_stop(&sc->sc_ad_tmo); 2027 carp_set_state(sc, BACKUP); 2028 carp_setrun(sc, 0); 2029 carp_setroute(sc, RTM_DELETE); 2030 break; 2031 case MASTER: 2032 carp_master_down(sc); 2033 break; 2034 default: 2035 break; 2036 } 2037 } 2038 if (carpr.carpr_vhid > 0) { 2039 if (carpr.carpr_vhid > 255) { 2040 error = EINVAL; 2041 break; 2042 } 2043 if (sc->sc_carpdev) { 2044 struct carp_if *cif; 2045 cif = (struct carp_if *)sc->sc_carpdev->if_carp; 2046 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) 2047 if (vr != sc && 2048 vr->sc_vhid == carpr.carpr_vhid) 2049 return (EINVAL); 2050 } 2051 sc->sc_vhid = carpr.carpr_vhid; 2052 carp_set_enaddr(sc); 2053 carp_set_state(sc, INIT); 2054 error--; 2055 } 2056 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) { 2057 if (carpr.carpr_advskew > 254) { 2058 error = EINVAL; 2059 break; 2060 } 2061 if (carpr.carpr_advbase > 255) { 2062 error = EINVAL; 2063 break; 2064 } 2065 sc->sc_advbase = carpr.carpr_advbase; 2066 sc->sc_advskew = carpr.carpr_advskew; 2067 error--; 2068 } 2069 memcpy(sc->sc_key, carpr.carpr_key, sizeof(sc->sc_key)); 2070 if (error > 0) 2071 error = EINVAL; 2072 else { 2073 error = 0; 2074 carp_setrun(sc, 0); 2075 } 2076 break; 2077 2078 case SIOCGVH: 2079 memset(&carpr, 0, sizeof(carpr)); 2080 if (sc->sc_carpdev != NULL) 2081 strlcpy(carpr.carpr_carpdev, sc->sc_carpdev->if_xname, 2082 IFNAMSIZ); 2083 carpr.carpr_state = sc->sc_state; 2084 carpr.carpr_vhid = sc->sc_vhid; 2085 carpr.carpr_advbase = sc->sc_advbase; 2086 carpr.carpr_advskew = sc->sc_advskew; 2087 2088 if ((l != NULL) && (error = kauth_authorize_network(l->l_cred, 2089 KAUTH_NETWORK_INTERFACE, 2090 KAUTH_REQ_NETWORK_INTERFACE_SETPRIV, ifp, (void *)cmd, 2091 NULL)) == 0) 2092 memcpy(carpr.carpr_key, sc->sc_key, 2093 sizeof(carpr.carpr_key)); 2094 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr)); 2095 break; 2096 2097 case SIOCADDMULTI: 2098 error = carp_ether_addmulti(sc, ifr); 2099 break; 2100 2101 case SIOCDELMULTI: 2102 error = carp_ether_delmulti(sc, ifr); 2103 break; 2104 2105 case SIOCSIFCAP: 2106 if ((error = ifioctl_common(ifp, cmd, data)) == ENETRESET) 2107 error = 0; 2108 break; 2109 2110 default: 2111 error = ether_ioctl(ifp, cmd, data); 2112 } 2113 2114 carp_hmac_prepare(sc); 2115 return (error); 2116 } 2117 2118 2119 /* 2120 * Start output on carp interface. This function should never be called. 2121 */ 2122 static void 2123 carp_start(struct ifnet *ifp) 2124 { 2125 #ifdef DEBUG 2126 printf("%s: start called\n", ifp->if_xname); 2127 #endif 2128 } 2129 2130 int 2131 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa, 2132 const struct rtentry *rt) 2133 { 2134 struct carp_softc *sc = ((struct carp_softc *)ifp->if_softc); 2135 KASSERT(KERNEL_LOCKED_P()); 2136 2137 if (sc->sc_carpdev != NULL && sc->sc_state == MASTER) { 2138 return if_output_lock(sc->sc_carpdev, ifp, m, sa, rt); 2139 } else { 2140 m_freem(m); 2141 return (ENETUNREACH); 2142 } 2143 } 2144 2145 static void 2146 carp_set_state(struct carp_softc *sc, int state) 2147 { 2148 static const char *carp_states[] = { CARP_STATES }; 2149 if (sc->sc_state == state) 2150 return; 2151 2152 CARP_LOG(sc, ("state transition from: %s -> to: %s", carp_states[sc->sc_state], carp_states[state])); 2153 2154 sc->sc_state = state; 2155 switch (state) { 2156 case BACKUP: 2157 sc->sc_if.if_link_state = LINK_STATE_DOWN; 2158 break; 2159 case MASTER: 2160 sc->sc_if.if_link_state = LINK_STATE_UP; 2161 break; 2162 default: 2163 sc->sc_if.if_link_state = LINK_STATE_UNKNOWN; 2164 break; 2165 } 2166 rt_ifmsg(&sc->sc_if); 2167 } 2168 2169 void 2170 carp_carpdev_state(void *v) 2171 { 2172 struct carp_if *cif; 2173 struct carp_softc *sc; 2174 struct ifnet *ifp = v; 2175 2176 if (ifp->if_type == IFT_CARP) 2177 return; 2178 2179 cif = (struct carp_if *)ifp->if_carp; 2180 2181 TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) { 2182 int suppressed = sc->sc_suppress; 2183 2184 if (sc->sc_carpdev->if_link_state == LINK_STATE_DOWN || 2185 !(sc->sc_carpdev->if_flags & IFF_UP)) { 2186 sc->sc_if.if_flags &= ~IFF_RUNNING; 2187 callout_stop(&sc->sc_ad_tmo); 2188 callout_stop(&sc->sc_md_tmo); 2189 callout_stop(&sc->sc_md6_tmo); 2190 carp_set_state(sc, INIT); 2191 sc->sc_suppress = 1; 2192 carp_setrun(sc, 0); 2193 if (!suppressed) { 2194 carp_suppress_preempt++; 2195 if (carp_suppress_preempt == 1) 2196 carp_send_ad_all(); 2197 } 2198 } else { 2199 carp_set_state(sc, INIT); 2200 sc->sc_suppress = 0; 2201 carp_setrun(sc, 0); 2202 if (suppressed) 2203 carp_suppress_preempt--; 2204 } 2205 } 2206 } 2207 2208 static int 2209 carp_ether_addmulti(struct carp_softc *sc, struct ifreq *ifr) 2210 { 2211 const struct sockaddr *sa = ifreq_getaddr(SIOCADDMULTI, ifr); 2212 struct ifnet *ifp; 2213 struct carp_mc_entry *mc; 2214 u_int8_t addrlo[ETHER_ADDR_LEN], addrhi[ETHER_ADDR_LEN]; 2215 int error; 2216 2217 ifp = sc->sc_carpdev; 2218 if (ifp == NULL) 2219 return (EINVAL); 2220 2221 error = ether_addmulti(sa, &sc->sc_ac); 2222 if (error != ENETRESET) 2223 return (error); 2224 2225 /* 2226 * This is new multicast address. We have to tell parent 2227 * about it. Also, remember this multicast address so that 2228 * we can delete them on unconfigure. 2229 */ 2230 mc = malloc(sizeof(struct carp_mc_entry), M_DEVBUF, M_NOWAIT); 2231 if (mc == NULL) { 2232 error = ENOMEM; 2233 goto alloc_failed; 2234 } 2235 2236 /* 2237 * As ether_addmulti() returns ENETRESET, following two 2238 * statement shouldn't fail. 2239 */ 2240 (void)ether_multiaddr(sa, addrlo, addrhi); 2241 ETHER_LOOKUP_MULTI(addrlo, addrhi, &sc->sc_ac, mc->mc_enm); 2242 memcpy(&mc->mc_addr, sa, sa->sa_len); 2243 LIST_INSERT_HEAD(&sc->carp_mc_listhead, mc, mc_entries); 2244 2245 error = if_mcast_op(ifp, SIOCADDMULTI, sa); 2246 if (error != 0) 2247 goto ioctl_failed; 2248 2249 return (error); 2250 2251 ioctl_failed: 2252 LIST_REMOVE(mc, mc_entries); 2253 free(mc, M_DEVBUF); 2254 alloc_failed: 2255 (void)ether_delmulti(sa, &sc->sc_ac); 2256 2257 return (error); 2258 } 2259 2260 static int 2261 carp_ether_delmulti(struct carp_softc *sc, struct ifreq *ifr) 2262 { 2263 const struct sockaddr *sa = ifreq_getaddr(SIOCDELMULTI, ifr); 2264 struct ifnet *ifp; 2265 struct ether_multi *enm; 2266 struct carp_mc_entry *mc; 2267 u_int8_t addrlo[ETHER_ADDR_LEN], addrhi[ETHER_ADDR_LEN]; 2268 int error; 2269 2270 ifp = sc->sc_carpdev; 2271 if (ifp == NULL) 2272 return (EINVAL); 2273 2274 /* 2275 * Find a key to lookup carp_mc_entry. We have to do this 2276 * before calling ether_delmulti for obvious reason. 2277 */ 2278 if ((error = ether_multiaddr(sa, addrlo, addrhi)) != 0) 2279 return (error); 2280 ETHER_LOOKUP_MULTI(addrlo, addrhi, &sc->sc_ac, enm); 2281 if (enm == NULL) 2282 return (EINVAL); 2283 2284 LIST_FOREACH(mc, &sc->carp_mc_listhead, mc_entries) 2285 if (mc->mc_enm == enm) 2286 break; 2287 2288 /* We won't delete entries we didn't add */ 2289 if (mc == NULL) 2290 return (EINVAL); 2291 2292 error = ether_delmulti(sa, &sc->sc_ac); 2293 if (error != ENETRESET) 2294 return (error); 2295 2296 /* We no longer use this multicast address. Tell parent so. */ 2297 error = if_mcast_op(ifp, SIOCDELMULTI, sa); 2298 if (error == 0) { 2299 /* And forget about this address. */ 2300 LIST_REMOVE(mc, mc_entries); 2301 free(mc, M_DEVBUF); 2302 } else 2303 (void)ether_addmulti(sa, &sc->sc_ac); 2304 return (error); 2305 } 2306 2307 /* 2308 * Delete any multicast address we have asked to add from parent 2309 * interface. Called when the carp is being unconfigured. 2310 */ 2311 static void 2312 carp_ether_purgemulti(struct carp_softc *sc) 2313 { 2314 struct ifnet *ifp = sc->sc_carpdev; /* Parent. */ 2315 struct carp_mc_entry *mc; 2316 2317 if (ifp == NULL) 2318 return; 2319 2320 while ((mc = LIST_FIRST(&sc->carp_mc_listhead)) != NULL) { 2321 (void)if_mcast_op(ifp, SIOCDELMULTI, sstosa(&mc->mc_addr)); 2322 LIST_REMOVE(mc, mc_entries); 2323 free(mc, M_DEVBUF); 2324 } 2325 } 2326 2327 static int 2328 sysctl_net_inet_carp_stats(SYSCTLFN_ARGS) 2329 { 2330 2331 return (NETSTAT_SYSCTL(carpstat_percpu, CARP_NSTATS)); 2332 } 2333 2334 void 2335 carp_init(void) 2336 { 2337 2338 sysctl_net_inet_carp_setup(NULL); 2339 #ifdef MBUFTRACE 2340 MOWNER_ATTACH(&carp_proto_mowner_rx); 2341 MOWNER_ATTACH(&carp_proto_mowner_tx); 2342 MOWNER_ATTACH(&carp_proto6_mowner_rx); 2343 MOWNER_ATTACH(&carp_proto6_mowner_tx); 2344 #endif 2345 } 2346 2347 static void 2348 sysctl_net_inet_carp_setup(struct sysctllog **clog) 2349 { 2350 2351 sysctl_createv(clog, 0, NULL, NULL, 2352 CTLFLAG_PERMANENT, 2353 CTLTYPE_NODE, "inet", NULL, 2354 NULL, 0, NULL, 0, 2355 CTL_NET, PF_INET, CTL_EOL); 2356 sysctl_createv(clog, 0, NULL, NULL, 2357 CTLFLAG_PERMANENT, 2358 CTLTYPE_NODE, "carp", 2359 SYSCTL_DESCR("CARP related settings"), 2360 NULL, 0, NULL, 0, 2361 CTL_NET, PF_INET, IPPROTO_CARP, CTL_EOL); 2362 2363 sysctl_createv(clog, 0, NULL, NULL, 2364 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2365 CTLTYPE_INT, "preempt", 2366 SYSCTL_DESCR("Enable CARP Preempt"), 2367 NULL, 0, &carp_opts[CARPCTL_PREEMPT], 0, 2368 CTL_NET, PF_INET, IPPROTO_CARP, 2369 CTL_CREATE, CTL_EOL); 2370 sysctl_createv(clog, 0, NULL, NULL, 2371 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2372 CTLTYPE_INT, "arpbalance", 2373 SYSCTL_DESCR("Enable ARP balancing"), 2374 NULL, 0, &carp_opts[CARPCTL_ARPBALANCE], 0, 2375 CTL_NET, PF_INET, IPPROTO_CARP, 2376 CTL_CREATE, CTL_EOL); 2377 sysctl_createv(clog, 0, NULL, NULL, 2378 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2379 CTLTYPE_INT, "allow", 2380 SYSCTL_DESCR("Enable CARP"), 2381 NULL, 0, &carp_opts[CARPCTL_ALLOW], 0, 2382 CTL_NET, PF_INET, IPPROTO_CARP, 2383 CTL_CREATE, CTL_EOL); 2384 sysctl_createv(clog, 0, NULL, NULL, 2385 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2386 CTLTYPE_INT, "log", 2387 SYSCTL_DESCR("CARP logging"), 2388 NULL, 0, &carp_opts[CARPCTL_LOG], 0, 2389 CTL_NET, PF_INET, IPPROTO_CARP, 2390 CTL_CREATE, CTL_EOL); 2391 sysctl_createv(clog, 0, NULL, NULL, 2392 CTLFLAG_PERMANENT, 2393 CTLTYPE_STRUCT, "stats", 2394 SYSCTL_DESCR("CARP statistics"), 2395 sysctl_net_inet_carp_stats, 0, NULL, 0, 2396 CTL_NET, PF_INET, IPPROTO_CARP, CARPCTL_STATS, 2397 CTL_EOL); 2398 } 2399
Indexes created Sat Sep 27 14:09:57 GMT 2025