ip_carp.c revision 1.70
1 /* $NetBSD: ip_carp.c,v 1.70 2016/06/20 08:08:13 knakahara 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 1.70 2016/06/20 08:08:13 knakahara 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_CONNECTED)); 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_CONNECTED; 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 && 422 (rt->rt_flags & RTF_CONNECTED) == 0) 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_CONNECTED; 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 struct ifnet *rcvif; 479 480 va_start(ap, m); 481 va_end(ap); 482 483 CARP_STATINC(CARP_STAT_IPACKETS); 484 MCLAIM(m, &carp_proto_mowner_rx); 485 486 if (!carp_opts[CARPCTL_ALLOW]) { 487 m_freem(m); 488 return; 489 } 490 491 rcvif = m_get_rcvif_NOMPSAFE(m); 492 /* check if received on a valid carp interface */ 493 if (rcvif->if_type != IFT_CARP) { 494 CARP_STATINC(CARP_STAT_BADIF); 495 CARP_LOG(sc, ("packet received on non-carp interface: %s", 496 rcvif->if_xname)); 497 m_freem(m); 498 return; 499 } 500 501 /* verify that the IP TTL is 255. */ 502 if (ip->ip_ttl != CARP_DFLTTL) { 503 CARP_STATINC(CARP_STAT_BADTTL); 504 CARP_LOG(sc, ("received ttl %d != %d on %s", ip->ip_ttl, 505 CARP_DFLTTL, rcvif->if_xname)); 506 m_freem(m); 507 return; 508 } 509 510 /* 511 * verify that the received packet length is 512 * equal to the CARP header 513 */ 514 iplen = ip->ip_hl << 2; 515 len = iplen + sizeof(*ch); 516 if (len > m->m_pkthdr.len) { 517 CARP_STATINC(CARP_STAT_BADLEN); 518 CARP_LOG(sc, ("packet too short %d on %s", m->m_pkthdr.len, 519 rcvif->if_xname)); 520 m_freem(m); 521 return; 522 } 523 524 if ((m = m_pullup(m, len)) == NULL) { 525 CARP_STATINC(CARP_STAT_HDROPS); 526 return; 527 } 528 ip = mtod(m, struct ip *); 529 ch = (struct carp_header *)((char *)ip + iplen); 530 /* verify the CARP checksum */ 531 m->m_data += iplen; 532 if (carp_cksum(m, len - iplen)) { 533 CARP_STATINC(CARP_STAT_BADSUM); 534 CARP_LOG(sc, ("checksum failed on %s", 535 rcvif->if_xname)); 536 m_freem(m); 537 return; 538 } 539 m->m_data -= iplen; 540 541 carp_proto_input_c(m, ch, AF_INET); 542 } 543 544 #ifdef INET6 545 int 546 carp6_proto_input(struct mbuf **mp, int *offp, int proto) 547 { 548 struct mbuf *m = *mp; 549 struct carp_softc *sc = NULL; 550 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 551 struct carp_header *ch; 552 u_int len; 553 struct ifnet *rcvif; 554 555 CARP_STATINC(CARP_STAT_IPACKETS6); 556 MCLAIM(m, &carp_proto6_mowner_rx); 557 558 if (!carp_opts[CARPCTL_ALLOW]) { 559 m_freem(m); 560 return (IPPROTO_DONE); 561 } 562 563 rcvif = m_get_rcvif_NOMPSAFE(m); 564 565 /* check if received on a valid carp interface */ 566 if (rcvif->if_type != IFT_CARP) { 567 CARP_STATINC(CARP_STAT_BADIF); 568 CARP_LOG(sc, ("packet received on non-carp interface: %s", 569 rcvif->if_xname)); 570 m_freem(m); 571 return (IPPROTO_DONE); 572 } 573 574 /* verify that the IP TTL is 255 */ 575 if (ip6->ip6_hlim != CARP_DFLTTL) { 576 CARP_STATINC(CARP_STAT_BADTTL); 577 CARP_LOG(sc, ("received ttl %d != %d on %s", ip6->ip6_hlim, 578 CARP_DFLTTL, rcvif->if_xname)); 579 m_freem(m); 580 return (IPPROTO_DONE); 581 } 582 583 /* verify that we have a complete carp packet */ 584 len = m->m_len; 585 IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch)); 586 if (ch == NULL) { 587 CARP_STATINC(CARP_STAT_BADLEN); 588 CARP_LOG(sc, ("packet size %u too small", len)); 589 return (IPPROTO_DONE); 590 } 591 592 593 /* verify the CARP checksum */ 594 m->m_data += *offp; 595 if (carp_cksum(m, sizeof(*ch))) { 596 CARP_STATINC(CARP_STAT_BADSUM); 597 CARP_LOG(sc, ("checksum failed, on %s", rcvif->if_xname)); 598 m_freem(m); 599 return (IPPROTO_DONE); 600 } 601 m->m_data -= *offp; 602 603 carp_proto_input_c(m, ch, AF_INET6); 604 return (IPPROTO_DONE); 605 } 606 #endif /* INET6 */ 607 608 static void 609 carp_proto_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af) 610 { 611 struct carp_softc *sc; 612 u_int64_t tmp_counter; 613 struct timeval sc_tv, ch_tv; 614 615 TAILQ_FOREACH(sc, &((struct carp_if *) 616 m_get_rcvif_NOMPSAFE(m)->if_carpdev->if_carp)->vhif_vrs, sc_list) 617 if (sc->sc_vhid == ch->carp_vhid) 618 break; 619 620 if (!sc || (sc->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) != 621 (IFF_UP|IFF_RUNNING)) { 622 CARP_STATINC(CARP_STAT_BADVHID); 623 m_freem(m); 624 return; 625 } 626 627 /* 628 * Check if our own advertisement was duplicated 629 * from a non simplex interface. 630 * XXX If there is no address on our physical interface 631 * there is no way to distinguish our ads from the ones 632 * another carp host might have sent us. 633 */ 634 if ((sc->sc_carpdev->if_flags & IFF_SIMPLEX) == 0) { 635 struct sockaddr sa; 636 struct ifaddr *ifa; 637 638 memset(&sa, 0, sizeof(sa)); 639 sa.sa_family = af; 640 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev); 641 642 if (ifa && af == AF_INET) { 643 struct ip *ip = mtod(m, struct ip *); 644 if (ip->ip_src.s_addr == 645 ifatoia(ifa)->ia_addr.sin_addr.s_addr) { 646 m_freem(m); 647 return; 648 } 649 } 650 #ifdef INET6 651 if (ifa && af == AF_INET6) { 652 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 653 struct in6_addr in6_src, in6_found; 654 655 in6_src = ip6->ip6_src; 656 in6_found = ifatoia6(ifa)->ia_addr.sin6_addr; 657 if (IN6_IS_ADDR_LINKLOCAL(&in6_src)) 658 in6_src.s6_addr16[1] = 0; 659 if (IN6_IS_ADDR_LINKLOCAL(&in6_found)) 660 in6_found.s6_addr16[1] = 0; 661 if (IN6_ARE_ADDR_EQUAL(&in6_src, &in6_found)) { 662 m_freem(m); 663 return; 664 } 665 } 666 #endif /* INET6 */ 667 } 668 669 nanotime(&sc->sc_if.if_lastchange); 670 sc->sc_if.if_ipackets++; 671 sc->sc_if.if_ibytes += m->m_pkthdr.len; 672 673 /* verify the CARP version. */ 674 if (ch->carp_version != CARP_VERSION) { 675 CARP_STATINC(CARP_STAT_BADVER); 676 sc->sc_if.if_ierrors++; 677 CARP_LOG(sc, ("invalid version %d != %d", 678 ch->carp_version, CARP_VERSION)); 679 m_freem(m); 680 return; 681 } 682 683 /* verify the hash */ 684 if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) { 685 CARP_STATINC(CARP_STAT_BADAUTH); 686 sc->sc_if.if_ierrors++; 687 CARP_LOG(sc, ("incorrect hash")); 688 m_freem(m); 689 return; 690 } 691 692 tmp_counter = ntohl(ch->carp_counter[0]); 693 tmp_counter = tmp_counter<<32; 694 tmp_counter += ntohl(ch->carp_counter[1]); 695 696 /* XXX Replay protection goes here */ 697 698 sc->sc_init_counter = 0; 699 sc->sc_counter = tmp_counter; 700 701 702 sc_tv.tv_sec = sc->sc_advbase; 703 if (carp_suppress_preempt && sc->sc_advskew < 240) 704 sc_tv.tv_usec = 240 * 1000000 / 256; 705 else 706 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256; 707 ch_tv.tv_sec = ch->carp_advbase; 708 ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256; 709 710 switch (sc->sc_state) { 711 case INIT: 712 break; 713 case MASTER: 714 /* 715 * If we receive an advertisement from a backup who's going to 716 * be more frequent than us, go into BACKUP state. 717 */ 718 if (timercmp(&sc_tv, &ch_tv, >) || 719 timercmp(&sc_tv, &ch_tv, ==)) { 720 callout_stop(&sc->sc_ad_tmo); 721 CARP_LOG(sc, ("MASTER -> BACKUP (more frequent advertisement received)")); 722 carp_set_state(sc, BACKUP); 723 carp_setrun(sc, 0); 724 carp_setroute(sc, RTM_DELETE); 725 } 726 break; 727 case BACKUP: 728 /* 729 * If we're pre-empting masters who advertise slower than us, 730 * and this one claims to be slower, treat him as down. 731 */ 732 if (carp_opts[CARPCTL_PREEMPT] && timercmp(&sc_tv, &ch_tv, <)) { 733 CARP_LOG(sc, ("BACKUP -> MASTER (preempting a slower master)")); 734 carp_master_down(sc); 735 break; 736 } 737 738 /* 739 * If the master is going to advertise at such a low frequency 740 * that he's guaranteed to time out, we'd might as well just 741 * treat him as timed out now. 742 */ 743 sc_tv.tv_sec = sc->sc_advbase * 3; 744 if (timercmp(&sc_tv, &ch_tv, <)) { 745 CARP_LOG(sc, ("BACKUP -> MASTER (master timed out)")); 746 carp_master_down(sc); 747 break; 748 } 749 750 /* 751 * Otherwise, we reset the counter and wait for the next 752 * advertisement. 753 */ 754 carp_setrun(sc, af); 755 break; 756 } 757 758 m_freem(m); 759 return; 760 } 761 762 /* 763 * Interface side of the CARP implementation. 764 */ 765 766 /* ARGSUSED */ 767 void 768 carpattach(int n) 769 { 770 if_clone_attach(&carp_cloner); 771 772 carpstat_percpu = percpu_alloc(sizeof(uint64_t) * CARP_NSTATS); 773 } 774 775 static int 776 carp_clone_create(struct if_clone *ifc, int unit) 777 { 778 extern int ifqmaxlen; 779 struct carp_softc *sc; 780 struct ifnet *ifp; 781 782 sc = malloc(sizeof(*sc), M_DEVBUF, M_NOWAIT|M_ZERO); 783 if (!sc) 784 return (ENOMEM); 785 786 sc->sc_suppress = 0; 787 sc->sc_advbase = CARP_DFLTINTV; 788 sc->sc_vhid = -1; /* required setting */ 789 sc->sc_advskew = 0; 790 sc->sc_init_counter = 1; 791 sc->sc_naddrs = sc->sc_naddrs6 = 0; 792 #ifdef INET6 793 sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL; 794 #endif /* INET6 */ 795 796 callout_init(&sc->sc_ad_tmo, 0); 797 callout_init(&sc->sc_md_tmo, 0); 798 callout_init(&sc->sc_md6_tmo, 0); 799 800 callout_setfunc(&sc->sc_ad_tmo, carp_send_ad, sc); 801 callout_setfunc(&sc->sc_md_tmo, carp_master_down, sc); 802 callout_setfunc(&sc->sc_md6_tmo, carp_master_down, sc); 803 804 LIST_INIT(&sc->carp_mc_listhead); 805 ifp = &sc->sc_if; 806 ifp->if_softc = sc; 807 snprintf(ifp->if_xname, sizeof ifp->if_xname, "%s%d", ifc->ifc_name, 808 unit); 809 ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST; 810 ifp->if_ioctl = carp_ioctl; 811 ifp->if_start = carp_start; 812 ifp->if_output = carp_output; 813 ifp->if_type = IFT_CARP; 814 ifp->if_addrlen = ETHER_ADDR_LEN; 815 ifp->if_hdrlen = ETHER_HDR_LEN; 816 ifp->if_mtu = ETHERMTU; 817 IFQ_SET_MAXLEN(&ifp->if_snd, ifqmaxlen); 818 IFQ_SET_READY(&ifp->if_snd); 819 if_attach(ifp); 820 821 if_alloc_sadl(ifp); 822 ifp->if_broadcastaddr = etherbroadcastaddr; 823 carp_set_enaddr(sc); 824 LIST_INIT(&sc->sc_ac.ec_multiaddrs); 825 bpf_attach(ifp, DLT_EN10MB, ETHER_HDR_LEN); 826 #ifdef MBUFTRACE 827 strlcpy(sc->sc_ac.ec_tx_mowner.mo_name, ifp->if_xname, 828 sizeof(sc->sc_ac.ec_tx_mowner.mo_name)); 829 strlcpy(sc->sc_ac.ec_tx_mowner.mo_descr, "tx", 830 sizeof(sc->sc_ac.ec_tx_mowner.mo_descr)); 831 strlcpy(sc->sc_ac.ec_rx_mowner.mo_name, ifp->if_xname, 832 sizeof(sc->sc_ac.ec_rx_mowner.mo_name)); 833 strlcpy(sc->sc_ac.ec_rx_mowner.mo_descr, "rx", 834 sizeof(sc->sc_ac.ec_rx_mowner.mo_descr)); 835 MOWNER_ATTACH(&sc->sc_ac.ec_tx_mowner); 836 MOWNER_ATTACH(&sc->sc_ac.ec_rx_mowner); 837 ifp->if_mowner = &sc->sc_ac.ec_tx_mowner; 838 #endif 839 return (0); 840 } 841 842 static int 843 carp_clone_destroy(struct ifnet *ifp) 844 { 845 struct carp_softc *sc = ifp->if_softc; 846 847 carpdetach(ifp->if_softc); 848 ether_ifdetach(ifp); 849 if_detach(ifp); 850 callout_destroy(&sc->sc_ad_tmo); 851 callout_destroy(&sc->sc_md_tmo); 852 callout_destroy(&sc->sc_md6_tmo); 853 free(ifp->if_softc, M_DEVBUF); 854 855 return (0); 856 } 857 858 static void 859 carpdetach(struct carp_softc *sc) 860 { 861 struct carp_if *cif; 862 int s; 863 864 callout_stop(&sc->sc_ad_tmo); 865 callout_stop(&sc->sc_md_tmo); 866 callout_stop(&sc->sc_md6_tmo); 867 868 if (sc->sc_suppress) 869 carp_suppress_preempt--; 870 sc->sc_suppress = 0; 871 872 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) 873 carp_suppress_preempt--; 874 sc->sc_sendad_errors = 0; 875 876 carp_set_state(sc, INIT); 877 sc->sc_if.if_flags &= ~IFF_UP; 878 carp_setrun(sc, 0); 879 carp_multicast_cleanup(sc); 880 881 KERNEL_LOCK(1, NULL); 882 s = splnet(); 883 if (sc->sc_carpdev != NULL) { 884 /* XXX linkstatehook removal */ 885 cif = (struct carp_if *)sc->sc_carpdev->if_carp; 886 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list); 887 if (!--cif->vhif_nvrs) { 888 ifpromisc(sc->sc_carpdev, 0); 889 sc->sc_carpdev->if_carp = NULL; 890 free(cif, M_IFADDR); 891 } 892 } 893 sc->sc_carpdev = NULL; 894 splx(s); 895 KERNEL_UNLOCK_ONE(NULL); 896 } 897 898 /* Detach an interface from the carp. */ 899 void 900 carp_ifdetach(struct ifnet *ifp) 901 { 902 struct carp_softc *sc, *nextsc; 903 struct carp_if *cif = (struct carp_if *)ifp->if_carp; 904 905 for (sc = TAILQ_FIRST(&cif->vhif_vrs); sc; sc = nextsc) { 906 nextsc = TAILQ_NEXT(sc, sc_list); 907 carpdetach(sc); 908 } 909 } 910 911 static int 912 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, 913 struct carp_header *ch) 914 { 915 if (sc->sc_init_counter) { 916 /* this could also be seconds since unix epoch */ 917 sc->sc_counter = cprng_fast64(); 918 } else 919 sc->sc_counter++; 920 921 ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff); 922 ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff); 923 924 carp_hmac_generate(sc, ch->carp_counter, ch->carp_md); 925 926 return (0); 927 } 928 929 static void 930 carp_send_ad_all(void) 931 { 932 struct ifnet *ifp; 933 struct carp_if *cif; 934 struct carp_softc *vh; 935 int s; 936 int bound = curlwp_bind(); 937 938 s = pserialize_read_enter(); 939 IFNET_READER_FOREACH(ifp) { 940 struct psref psref; 941 if (ifp->if_carp == NULL || ifp->if_type == IFT_CARP) 942 continue; 943 944 psref_acquire(&psref, &ifp->if_psref, ifnet_psref_class); 945 pserialize_read_exit(s); 946 947 cif = (struct carp_if *)ifp->if_carp; 948 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 949 if ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 950 (IFF_UP|IFF_RUNNING) && vh->sc_state == MASTER) 951 carp_send_ad(vh); 952 } 953 954 s = pserialize_read_enter(); 955 psref_release(&psref, &ifp->if_psref, ifnet_psref_class); 956 } 957 pserialize_read_exit(s); 958 curlwp_bindx(bound); 959 } 960 961 962 static void 963 carp_send_ad(void *v) 964 { 965 struct carp_header ch; 966 struct timeval tv; 967 struct carp_softc *sc = v; 968 struct carp_header *ch_ptr; 969 struct mbuf *m; 970 int error, len, advbase, advskew, s; 971 struct ifaddr *ifa; 972 struct sockaddr sa; 973 974 KERNEL_LOCK(1, NULL); 975 s = splsoftnet(); 976 977 advbase = advskew = 0; /* Sssssh compiler */ 978 if (sc->sc_carpdev == NULL) { 979 sc->sc_if.if_oerrors++; 980 goto retry_later; 981 } 982 983 /* bow out if we've gone to backup (the carp interface is going down) */ 984 if (sc->sc_bow_out) { 985 sc->sc_bow_out = 0; 986 advbase = 255; 987 advskew = 255; 988 } else { 989 advbase = sc->sc_advbase; 990 if (!carp_suppress_preempt || sc->sc_advskew > 240) 991 advskew = sc->sc_advskew; 992 else 993 advskew = 240; 994 tv.tv_sec = advbase; 995 tv.tv_usec = advskew * 1000000 / 256; 996 } 997 998 ch.carp_version = CARP_VERSION; 999 ch.carp_type = CARP_ADVERTISEMENT; 1000 ch.carp_vhid = sc->sc_vhid; 1001 ch.carp_advbase = advbase; 1002 ch.carp_advskew = advskew; 1003 ch.carp_authlen = 7; /* XXX DEFINE */ 1004 ch.carp_pad1 = 0; /* must be zero */ 1005 ch.carp_cksum = 0; 1006 1007 1008 #ifdef INET 1009 if (sc->sc_naddrs) { 1010 struct ip *ip; 1011 1012 MGETHDR(m, M_DONTWAIT, MT_HEADER); 1013 if (m == NULL) { 1014 sc->sc_if.if_oerrors++; 1015 CARP_STATINC(CARP_STAT_ONOMEM); 1016 /* XXX maybe less ? */ 1017 goto retry_later; 1018 } 1019 MCLAIM(m, &carp_proto_mowner_tx); 1020 len = sizeof(*ip) + sizeof(ch); 1021 m->m_pkthdr.len = len; 1022 m_reset_rcvif(m); 1023 m->m_len = len; 1024 MH_ALIGN(m, m->m_len); 1025 m->m_flags |= M_MCAST; 1026 ip = mtod(m, struct ip *); 1027 ip->ip_v = IPVERSION; 1028 ip->ip_hl = sizeof(*ip) >> 2; 1029 ip->ip_tos = IPTOS_LOWDELAY; 1030 ip->ip_len = htons(len); 1031 ip->ip_id = 0; /* no need for id, we don't support fragments */ 1032 ip->ip_off = htons(IP_DF); 1033 ip->ip_ttl = CARP_DFLTTL; 1034 ip->ip_p = IPPROTO_CARP; 1035 ip->ip_sum = 0; 1036 1037 memset(&sa, 0, sizeof(sa)); 1038 sa.sa_family = AF_INET; 1039 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev); 1040 if (ifa == NULL) 1041 ip->ip_src.s_addr = 0; 1042 else 1043 ip->ip_src.s_addr = 1044 ifatoia(ifa)->ia_addr.sin_addr.s_addr; 1045 ip->ip_dst.s_addr = INADDR_CARP_GROUP; 1046 1047 ch_ptr = (struct carp_header *)(&ip[1]); 1048 memcpy(ch_ptr, &ch, sizeof(ch)); 1049 if (carp_prepare_ad(m, sc, ch_ptr)) 1050 goto retry_later; 1051 1052 m->m_data += sizeof(*ip); 1053 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip)); 1054 m->m_data -= sizeof(*ip); 1055 1056 nanotime(&sc->sc_if.if_lastchange); 1057 sc->sc_if.if_opackets++; 1058 sc->sc_if.if_obytes += len; 1059 CARP_STATINC(CARP_STAT_OPACKETS); 1060 1061 error = ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, 1062 NULL); 1063 if (error) { 1064 if (error == ENOBUFS) 1065 CARP_STATINC(CARP_STAT_ONOMEM); 1066 else 1067 CARP_LOG(sc, ("ip_output failed: %d", error)); 1068 sc->sc_if.if_oerrors++; 1069 if (sc->sc_sendad_errors < INT_MAX) 1070 sc->sc_sendad_errors++; 1071 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) { 1072 carp_suppress_preempt++; 1073 if (carp_suppress_preempt == 1) 1074 carp_send_ad_all(); 1075 } 1076 sc->sc_sendad_success = 0; 1077 } else { 1078 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) { 1079 if (++sc->sc_sendad_success >= 1080 CARP_SENDAD_MIN_SUCCESS) { 1081 carp_suppress_preempt--; 1082 sc->sc_sendad_errors = 0; 1083 } 1084 } else 1085 sc->sc_sendad_errors = 0; 1086 } 1087 } 1088 #endif /* INET */ 1089 #ifdef INET6 1090 if (sc->sc_naddrs6) { 1091 struct ip6_hdr *ip6; 1092 1093 MGETHDR(m, M_DONTWAIT, MT_HEADER); 1094 if (m == NULL) { 1095 sc->sc_if.if_oerrors++; 1096 CARP_STATINC(CARP_STAT_ONOMEM); 1097 /* XXX maybe less ? */ 1098 goto retry_later; 1099 } 1100 MCLAIM(m, &carp_proto6_mowner_tx); 1101 len = sizeof(*ip6) + sizeof(ch); 1102 m->m_pkthdr.len = len; 1103 m_reset_rcvif(m); 1104 m->m_len = len; 1105 MH_ALIGN(m, m->m_len); 1106 m->m_flags |= M_MCAST; 1107 ip6 = mtod(m, struct ip6_hdr *); 1108 memset(ip6, 0, sizeof(*ip6)); 1109 ip6->ip6_vfc |= IPV6_VERSION; 1110 ip6->ip6_hlim = CARP_DFLTTL; 1111 ip6->ip6_nxt = IPPROTO_CARP; 1112 1113 /* set the source address */ 1114 memset(&sa, 0, sizeof(sa)); 1115 sa.sa_family = AF_INET6; 1116 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev); 1117 if (ifa == NULL) /* This should never happen with IPv6 */ 1118 memset(&ip6->ip6_src, 0, sizeof(struct in6_addr)); 1119 else 1120 bcopy(ifatoia6(ifa)->ia_addr.sin6_addr.s6_addr, 1121 &ip6->ip6_src, sizeof(struct in6_addr)); 1122 /* set the multicast destination */ 1123 1124 ip6->ip6_dst.s6_addr16[0] = htons(0xff02); 1125 ip6->ip6_dst.s6_addr8[15] = 0x12; 1126 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) { 1127 sc->sc_if.if_oerrors++; 1128 m_freem(m); 1129 CARP_LOG(sc, ("in6_setscope failed")); 1130 goto retry_later; 1131 } 1132 1133 ch_ptr = (struct carp_header *)(&ip6[1]); 1134 memcpy(ch_ptr, &ch, sizeof(ch)); 1135 if (carp_prepare_ad(m, sc, ch_ptr)) 1136 goto retry_later; 1137 1138 m->m_data += sizeof(*ip6); 1139 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6)); 1140 m->m_data -= sizeof(*ip6); 1141 1142 nanotime(&sc->sc_if.if_lastchange); 1143 sc->sc_if.if_opackets++; 1144 sc->sc_if.if_obytes += len; 1145 CARP_STATINC(CARP_STAT_OPACKETS6); 1146 1147 error = ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL); 1148 if (error) { 1149 if (error == ENOBUFS) 1150 CARP_STATINC(CARP_STAT_ONOMEM); 1151 else 1152 CARP_LOG(sc, ("ip6_output failed: %d", error)); 1153 sc->sc_if.if_oerrors++; 1154 if (sc->sc_sendad_errors < INT_MAX) 1155 sc->sc_sendad_errors++; 1156 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) { 1157 carp_suppress_preempt++; 1158 if (carp_suppress_preempt == 1) 1159 carp_send_ad_all(); 1160 } 1161 sc->sc_sendad_success = 0; 1162 } else { 1163 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) { 1164 if (++sc->sc_sendad_success >= 1165 CARP_SENDAD_MIN_SUCCESS) { 1166 carp_suppress_preempt--; 1167 sc->sc_sendad_errors = 0; 1168 } 1169 } else 1170 sc->sc_sendad_errors = 0; 1171 } 1172 } 1173 #endif /* INET6 */ 1174 1175 retry_later: 1176 splx(s); 1177 KERNEL_UNLOCK_ONE(NULL); 1178 if (advbase != 255 || advskew != 255) 1179 callout_schedule(&sc->sc_ad_tmo, tvtohz(&tv)); 1180 } 1181 1182 /* 1183 * Broadcast a gratuitous ARP request containing 1184 * the virtual router MAC address for each IP address 1185 * associated with the virtual router. 1186 */ 1187 static void 1188 carp_send_arp(struct carp_softc *sc) 1189 { 1190 struct ifaddr *ifa; 1191 struct in_addr *in; 1192 int s; 1193 1194 KERNEL_LOCK(1, NULL); 1195 s = splsoftnet(); 1196 IFADDR_FOREACH(ifa, &sc->sc_if) { 1197 1198 if (ifa->ifa_addr->sa_family != AF_INET) 1199 continue; 1200 1201 in = &ifatoia(ifa)->ia_addr.sin_addr; 1202 arprequest(sc->sc_carpdev, in, in, CLLADDR(sc->sc_if.if_sadl)); 1203 } 1204 splx(s); 1205 KERNEL_UNLOCK_ONE(NULL); 1206 } 1207 1208 #ifdef INET6 1209 static void 1210 carp_send_na(struct carp_softc *sc) 1211 { 1212 struct ifaddr *ifa; 1213 struct in6_addr *in6; 1214 static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT; 1215 int s; 1216 1217 KERNEL_LOCK(1, NULL); 1218 s = splsoftnet(); 1219 1220 IFADDR_FOREACH(ifa, &sc->sc_if) { 1221 1222 if (ifa->ifa_addr->sa_family != AF_INET6) 1223 continue; 1224 1225 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr; 1226 nd6_na_output(sc->sc_carpdev, &mcast, in6, 1227 ND_NA_FLAG_OVERRIDE, 1, NULL); 1228 } 1229 splx(s); 1230 KERNEL_UNLOCK_ONE(NULL); 1231 } 1232 #endif /* INET6 */ 1233 1234 /* 1235 * Based on bridge_hash() in if_bridge.c 1236 */ 1237 #define mix(a,b,c) \ 1238 do { \ 1239 a -= b; a -= c; a ^= (c >> 13); \ 1240 b -= c; b -= a; b ^= (a << 8); \ 1241 c -= a; c -= b; c ^= (b >> 13); \ 1242 a -= b; a -= c; a ^= (c >> 12); \ 1243 b -= c; b -= a; b ^= (a << 16); \ 1244 c -= a; c -= b; c ^= (b >> 5); \ 1245 a -= b; a -= c; a ^= (c >> 3); \ 1246 b -= c; b -= a; b ^= (a << 10); \ 1247 c -= a; c -= b; c ^= (b >> 15); \ 1248 } while (0) 1249 1250 static u_int32_t 1251 carp_hash(struct carp_softc *sc, u_char *src) 1252 { 1253 u_int32_t a = 0x9e3779b9, b = sc->sc_hashkey[0], c = sc->sc_hashkey[1]; 1254 1255 c += sc->sc_key[3] << 24; 1256 c += sc->sc_key[2] << 16; 1257 c += sc->sc_key[1] << 8; 1258 c += sc->sc_key[0]; 1259 b += src[5] << 8; 1260 b += src[4]; 1261 a += src[3] << 24; 1262 a += src[2] << 16; 1263 a += src[1] << 8; 1264 a += src[0]; 1265 1266 mix(a, b, c); 1267 return (c); 1268 } 1269 1270 static int 1271 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type) 1272 { 1273 struct carp_softc *vh; 1274 struct ifaddr *ifa; 1275 int count = 0; 1276 1277 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1278 if ((type == CARP_COUNT_RUNNING && 1279 (vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1280 (IFF_UP|IFF_RUNNING)) || 1281 (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) { 1282 IFADDR_FOREACH(ifa, &vh->sc_if) { 1283 if (ifa->ifa_addr->sa_family == AF_INET && 1284 ia->ia_addr.sin_addr.s_addr == 1285 ifatoia(ifa)->ia_addr.sin_addr.s_addr) 1286 count++; 1287 } 1288 } 1289 } 1290 return (count); 1291 } 1292 1293 int 1294 carp_iamatch(struct in_ifaddr *ia, u_char *src, 1295 u_int32_t *count, u_int32_t index) 1296 { 1297 struct carp_softc *sc = ia->ia_ifp->if_softc; 1298 1299 if (carp_opts[CARPCTL_ARPBALANCE]) { 1300 /* 1301 * We use the source ip to decide which virtual host should 1302 * handle the request. If we're master of that virtual host, 1303 * then we respond, otherwise, just drop the arp packet on 1304 * the floor. 1305 */ 1306 1307 /* Count the elegible carp interfaces with this address */ 1308 if (*count == 0) 1309 *count = carp_addrcount( 1310 (struct carp_if *)ia->ia_ifp->if_carpdev->if_carp, 1311 ia, CARP_COUNT_RUNNING); 1312 1313 /* This should never happen, but... */ 1314 if (*count == 0) 1315 return (0); 1316 1317 if (carp_hash(sc, src) % *count == index - 1 && 1318 sc->sc_state == MASTER) { 1319 return (1); 1320 } 1321 } else { 1322 if (sc->sc_state == MASTER) 1323 return (1); 1324 } 1325 1326 return (0); 1327 } 1328 1329 #ifdef INET6 1330 struct ifaddr * 1331 carp_iamatch6(void *v, struct in6_addr *taddr) 1332 { 1333 struct carp_if *cif = v; 1334 struct carp_softc *vh; 1335 struct ifaddr *ifa; 1336 1337 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1338 IFADDR_FOREACH(ifa, &vh->sc_if) { 1339 if (IN6_ARE_ADDR_EQUAL(taddr, 1340 &ifatoia6(ifa)->ia_addr.sin6_addr) && 1341 ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1342 (IFF_UP|IFF_RUNNING)) && vh->sc_state == MASTER) 1343 return (ifa); 1344 } 1345 } 1346 1347 return (NULL); 1348 } 1349 #endif /* INET6 */ 1350 1351 struct ifnet * 1352 carp_ourether(void *v, struct ether_header *eh, u_char iftype, int src) 1353 { 1354 struct carp_if *cif = (struct carp_if *)v; 1355 struct carp_softc *vh; 1356 u_int8_t *ena; 1357 1358 if (src) 1359 ena = (u_int8_t *)&eh->ether_shost; 1360 else 1361 ena = (u_int8_t *)&eh->ether_dhost; 1362 1363 switch (iftype) { 1364 case IFT_ETHER: 1365 case IFT_FDDI: 1366 if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1) 1367 return (NULL); 1368 break; 1369 case IFT_ISO88025: 1370 if (ena[0] != 3 || ena[1] || ena[4] || ena[5]) 1371 return (NULL); 1372 break; 1373 default: 1374 return (NULL); 1375 break; 1376 } 1377 1378 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) 1379 if ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1380 (IFF_UP|IFF_RUNNING) && vh->sc_state == MASTER && 1381 !memcmp(ena, CLLADDR(vh->sc_if.if_sadl), 1382 ETHER_ADDR_LEN)) { 1383 return (&vh->sc_if); 1384 } 1385 1386 return (NULL); 1387 } 1388 1389 int 1390 carp_input(struct mbuf *m, u_int8_t *shost, u_int8_t *dhost, u_int16_t etype) 1391 { 1392 struct ether_header eh; 1393 struct carp_if *cif = (struct carp_if *)m_get_rcvif_NOMPSAFE(m)->if_carp; 1394 struct ifnet *ifp; 1395 1396 memcpy(&eh.ether_shost, shost, sizeof(eh.ether_shost)); 1397 memcpy(&eh.ether_dhost, dhost, sizeof(eh.ether_dhost)); 1398 eh.ether_type = etype; 1399 1400 if (m->m_flags & (M_BCAST|M_MCAST)) { 1401 struct carp_softc *vh; 1402 struct mbuf *m0; 1403 1404 /* 1405 * XXX Should really check the list of multicast addresses 1406 * for each CARP interface _before_ copying. 1407 */ 1408 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1409 m0 = m_copym(m, 0, M_COPYALL, M_DONTWAIT); 1410 if (m0 == NULL) 1411 continue; 1412 m_set_rcvif(m0, &vh->sc_if); 1413 ether_input(&vh->sc_if, m0); 1414 } 1415 return (1); 1416 } 1417 1418 ifp = carp_ourether(cif, &eh, m_get_rcvif_NOMPSAFE(m)->if_type, 0); 1419 if (ifp == NULL) { 1420 return (1); 1421 } 1422 1423 m_set_rcvif(m, ifp); 1424 1425 bpf_mtap(ifp, m); 1426 ifp->if_ipackets++; 1427 ether_input(ifp, m); 1428 return (0); 1429 } 1430 1431 static void 1432 carp_master_down(void *v) 1433 { 1434 struct carp_softc *sc = v; 1435 1436 switch (sc->sc_state) { 1437 case INIT: 1438 printf("%s: master_down event in INIT state\n", 1439 sc->sc_if.if_xname); 1440 break; 1441 case MASTER: 1442 break; 1443 case BACKUP: 1444 CARP_LOG(sc, ("INIT -> MASTER (preempting)")); 1445 carp_set_state(sc, MASTER); 1446 carp_send_ad(sc); 1447 carp_send_arp(sc); 1448 #ifdef INET6 1449 carp_send_na(sc); 1450 #endif /* INET6 */ 1451 carp_setrun(sc, 0); 1452 carp_setroute(sc, RTM_ADD); 1453 break; 1454 } 1455 } 1456 1457 /* 1458 * When in backup state, af indicates whether to reset the master down timer 1459 * for v4 or v6. If it's set to zero, reset the ones which are already pending. 1460 */ 1461 static void 1462 carp_setrun(struct carp_softc *sc, sa_family_t af) 1463 { 1464 struct timeval tv; 1465 1466 if (sc->sc_carpdev == NULL) { 1467 sc->sc_if.if_flags &= ~IFF_RUNNING; 1468 carp_set_state(sc, INIT); 1469 return; 1470 } 1471 1472 if (sc->sc_if.if_flags & IFF_UP && sc->sc_vhid > 0 && 1473 (sc->sc_naddrs || sc->sc_naddrs6) && !sc->sc_suppress) { 1474 sc->sc_if.if_flags |= IFF_RUNNING; 1475 } else { 1476 sc->sc_if.if_flags &= ~IFF_RUNNING; 1477 carp_setroute(sc, RTM_DELETE); 1478 return; 1479 } 1480 1481 switch (sc->sc_state) { 1482 case INIT: 1483 carp_set_state(sc, BACKUP); 1484 carp_setroute(sc, RTM_DELETE); 1485 carp_setrun(sc, 0); 1486 break; 1487 case BACKUP: 1488 callout_stop(&sc->sc_ad_tmo); 1489 tv.tv_sec = 3 * sc->sc_advbase; 1490 tv.tv_usec = sc->sc_advskew * 1000000 / 256; 1491 switch (af) { 1492 #ifdef INET 1493 case AF_INET: 1494 callout_schedule(&sc->sc_md_tmo, tvtohz(&tv)); 1495 break; 1496 #endif /* INET */ 1497 #ifdef INET6 1498 case AF_INET6: 1499 callout_schedule(&sc->sc_md6_tmo, tvtohz(&tv)); 1500 break; 1501 #endif /* INET6 */ 1502 default: 1503 if (sc->sc_naddrs) 1504 callout_schedule(&sc->sc_md_tmo, tvtohz(&tv)); 1505 if (sc->sc_naddrs6) 1506 callout_schedule(&sc->sc_md6_tmo, tvtohz(&tv)); 1507 break; 1508 } 1509 break; 1510 case MASTER: 1511 tv.tv_sec = sc->sc_advbase; 1512 tv.tv_usec = sc->sc_advskew * 1000000 / 256; 1513 callout_schedule(&sc->sc_ad_tmo, tvtohz(&tv)); 1514 break; 1515 } 1516 } 1517 1518 static void 1519 carp_multicast_cleanup(struct carp_softc *sc) 1520 { 1521 struct ip_moptions *imo = &sc->sc_imo; 1522 #ifdef INET6 1523 struct ip6_moptions *im6o = &sc->sc_im6o; 1524 #endif 1525 u_int16_t n = imo->imo_num_memberships; 1526 1527 /* Clean up our own multicast memberships */ 1528 while (n-- > 0) { 1529 if (imo->imo_membership[n] != NULL) { 1530 in_delmulti(imo->imo_membership[n]); 1531 imo->imo_membership[n] = NULL; 1532 } 1533 } 1534 imo->imo_num_memberships = 0; 1535 imo->imo_multicast_ifp = NULL; 1536 1537 #ifdef INET6 1538 while (!LIST_EMPTY(&im6o->im6o_memberships)) { 1539 struct in6_multi_mship *imm = 1540 LIST_FIRST(&im6o->im6o_memberships); 1541 1542 LIST_REMOVE(imm, i6mm_chain); 1543 in6_leavegroup(imm); 1544 } 1545 im6o->im6o_multicast_ifp = NULL; 1546 #endif 1547 1548 /* And any other multicast memberships */ 1549 carp_ether_purgemulti(sc); 1550 } 1551 1552 static int 1553 carp_set_ifp(struct carp_softc *sc, struct ifnet *ifp) 1554 { 1555 struct carp_if *cif, *ncif = NULL; 1556 struct carp_softc *vr, *after = NULL; 1557 int myself = 0, error = 0; 1558 int s; 1559 1560 if (ifp == sc->sc_carpdev) 1561 return (0); 1562 1563 if (ifp != NULL) { 1564 if ((ifp->if_flags & IFF_MULTICAST) == 0) 1565 return (EADDRNOTAVAIL); 1566 1567 if (ifp->if_type == IFT_CARP) 1568 return (EINVAL); 1569 1570 if (ifp->if_carp == NULL) { 1571 ncif = malloc(sizeof(*cif), M_IFADDR, M_NOWAIT); 1572 if (ncif == NULL) 1573 return (ENOBUFS); 1574 if ((error = ifpromisc(ifp, 1))) { 1575 free(ncif, M_IFADDR); 1576 return (error); 1577 } 1578 1579 ncif->vhif_ifp = ifp; 1580 TAILQ_INIT(&ncif->vhif_vrs); 1581 } else { 1582 cif = (struct carp_if *)ifp->if_carp; 1583 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) 1584 if (vr != sc && vr->sc_vhid == sc->sc_vhid) 1585 return (EINVAL); 1586 } 1587 1588 /* detach from old interface */ 1589 if (sc->sc_carpdev != NULL) 1590 carpdetach(sc); 1591 1592 /* join multicast groups */ 1593 if (sc->sc_naddrs < 0 && 1594 (error = carp_join_multicast(sc)) != 0) { 1595 if (ncif != NULL) 1596 free(ncif, M_IFADDR); 1597 return (error); 1598 } 1599 1600 #ifdef INET6 1601 if (sc->sc_naddrs6 < 0 && 1602 (error = carp_join_multicast6(sc)) != 0) { 1603 if (ncif != NULL) 1604 free(ncif, M_IFADDR); 1605 carp_multicast_cleanup(sc); 1606 return (error); 1607 } 1608 #endif 1609 1610 /* attach carp interface to physical interface */ 1611 if (ncif != NULL) 1612 ifp->if_carp = (void *)ncif; 1613 sc->sc_carpdev = ifp; 1614 sc->sc_if.if_capabilities = ifp->if_capabilities & 1615 (IFCAP_TSOv4 | IFCAP_TSOv6 | 1616 IFCAP_CSUM_IPv4_Tx|IFCAP_CSUM_IPv4_Rx| 1617 IFCAP_CSUM_TCPv4_Tx|IFCAP_CSUM_TCPv4_Rx| 1618 IFCAP_CSUM_UDPv4_Tx|IFCAP_CSUM_UDPv4_Rx| 1619 IFCAP_CSUM_TCPv6_Tx|IFCAP_CSUM_TCPv6_Rx| 1620 IFCAP_CSUM_UDPv6_Tx|IFCAP_CSUM_UDPv6_Rx); 1621 1622 cif = (struct carp_if *)ifp->if_carp; 1623 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) { 1624 if (vr == sc) 1625 myself = 1; 1626 if (vr->sc_vhid < sc->sc_vhid) 1627 after = vr; 1628 } 1629 1630 if (!myself) { 1631 /* We're trying to keep things in order */ 1632 if (after == NULL) { 1633 TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list); 1634 } else { 1635 TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, 1636 sc, sc_list); 1637 } 1638 cif->vhif_nvrs++; 1639 } 1640 if (sc->sc_naddrs || sc->sc_naddrs6) 1641 sc->sc_if.if_flags |= IFF_UP; 1642 carp_set_enaddr(sc); 1643 KERNEL_LOCK(1, NULL); 1644 s = splnet(); 1645 /* XXX linkstatehooks establish */ 1646 carp_carpdev_state(ifp); 1647 splx(s); 1648 KERNEL_UNLOCK_ONE(NULL); 1649 } else { 1650 carpdetach(sc); 1651 sc->sc_if.if_flags &= ~(IFF_UP|IFF_RUNNING); 1652 } 1653 return (0); 1654 } 1655 1656 static void 1657 carp_set_enaddr(struct carp_softc *sc) 1658 { 1659 uint8_t enaddr[ETHER_ADDR_LEN]; 1660 if (sc->sc_carpdev && sc->sc_carpdev->if_type == IFT_ISO88025) { 1661 enaddr[0] = 3; 1662 enaddr[1] = 0; 1663 enaddr[2] = 0x40 >> (sc->sc_vhid - 1); 1664 enaddr[3] = 0x40000 >> (sc->sc_vhid - 1); 1665 enaddr[4] = 0; 1666 enaddr[5] = 0; 1667 } else { 1668 enaddr[0] = 0; 1669 enaddr[1] = 0; 1670 enaddr[2] = 0x5e; 1671 enaddr[3] = 0; 1672 enaddr[4] = 1; 1673 enaddr[5] = sc->sc_vhid; 1674 } 1675 if_set_sadl(&sc->sc_if, enaddr, sizeof(enaddr), false); 1676 } 1677 1678 #if 0 1679 static void 1680 carp_addr_updated(void *v) 1681 { 1682 struct carp_softc *sc = (struct carp_softc *) v; 1683 struct ifaddr *ifa; 1684 int new_naddrs = 0, new_naddrs6 = 0; 1685 1686 IFADDR_FOREACH(ifa, &sc->sc_if) { 1687 if (ifa->ifa_addr->sa_family == AF_INET) 1688 new_naddrs++; 1689 else if (ifa->ifa_addr->sa_family == AF_INET6) 1690 new_naddrs6++; 1691 } 1692 1693 /* Handle a callback after SIOCDIFADDR */ 1694 if (new_naddrs < sc->sc_naddrs || new_naddrs6 < sc->sc_naddrs6) { 1695 struct in_addr mc_addr; 1696 1697 sc->sc_naddrs = new_naddrs; 1698 sc->sc_naddrs6 = new_naddrs6; 1699 1700 /* Re-establish multicast membership removed by in_control */ 1701 mc_addr.s_addr = INADDR_CARP_GROUP; 1702 if (!in_multi_group(mc_addr, &sc->sc_if, 0)) { 1703 memset(&sc->sc_imo, 0, sizeof(sc->sc_imo)); 1704 1705 if (sc->sc_carpdev != NULL && sc->sc_naddrs > 0) 1706 carp_join_multicast(sc); 1707 } 1708 1709 if (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) { 1710 sc->sc_if.if_flags &= ~IFF_UP; 1711 carp_set_state(sc, INIT); 1712 } else 1713 carp_hmac_prepare(sc); 1714 } 1715 1716 carp_setrun(sc, 0); 1717 } 1718 #endif 1719 1720 static int 1721 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin) 1722 { 1723 struct ifnet *ifp = sc->sc_carpdev; 1724 struct in_ifaddr *ia, *ia_if; 1725 int error = 0; 1726 1727 if (sin->sin_addr.s_addr == 0) { 1728 if (!(sc->sc_if.if_flags & IFF_UP)) 1729 carp_set_state(sc, INIT); 1730 if (sc->sc_naddrs) 1731 sc->sc_if.if_flags |= IFF_UP; 1732 carp_setrun(sc, 0); 1733 return (0); 1734 } 1735 1736 /* we have to do this by hand to ensure we don't match on ourselves */ 1737 ia_if = NULL; 1738 for (ia = TAILQ_FIRST(&in_ifaddrhead); ia; 1739 ia = TAILQ_NEXT(ia, ia_list)) { 1740 1741 /* and, yeah, we need a multicast-capable iface too */ 1742 if (ia->ia_ifp != &sc->sc_if && 1743 ia->ia_ifp->if_type != IFT_CARP && 1744 (ia->ia_ifp->if_flags & IFF_MULTICAST) && 1745 (sin->sin_addr.s_addr & ia->ia_subnetmask) == 1746 ia->ia_subnet) { 1747 if (!ia_if) 1748 ia_if = ia; 1749 } 1750 } 1751 1752 if (ia_if) { 1753 ia = ia_if; 1754 if (ifp) { 1755 if (ifp != ia->ia_ifp) 1756 return (EADDRNOTAVAIL); 1757 } else { 1758 ifp = ia->ia_ifp; 1759 } 1760 } 1761 1762 if ((error = carp_set_ifp(sc, ifp))) 1763 return (error); 1764 1765 if (sc->sc_carpdev == NULL) 1766 return (EADDRNOTAVAIL); 1767 1768 if (sc->sc_naddrs == 0 && (error = carp_join_multicast(sc)) != 0) 1769 return (error); 1770 1771 sc->sc_naddrs++; 1772 if (sc->sc_carpdev != NULL) 1773 sc->sc_if.if_flags |= IFF_UP; 1774 1775 carp_set_state(sc, INIT); 1776 carp_setrun(sc, 0); 1777 1778 /* 1779 * Hook if_addrhooks so that we get a callback after in_ifinit has run, 1780 * to correct any inappropriate routes that it inserted. 1781 */ 1782 if (sc->ah_cookie == 0) { 1783 /* XXX link address hook */ 1784 } 1785 1786 return (0); 1787 } 1788 1789 static int 1790 carp_join_multicast(struct carp_softc *sc) 1791 { 1792 struct ip_moptions *imo = &sc->sc_imo, tmpimo; 1793 struct in_addr addr; 1794 1795 memset(&tmpimo, 0, sizeof(tmpimo)); 1796 addr.s_addr = INADDR_CARP_GROUP; 1797 if ((tmpimo.imo_membership[0] = 1798 in_addmulti(&addr, &sc->sc_if)) == NULL) { 1799 return (ENOBUFS); 1800 } 1801 1802 imo->imo_membership[0] = tmpimo.imo_membership[0]; 1803 imo->imo_num_memberships = 1; 1804 imo->imo_multicast_ifp = &sc->sc_if; 1805 imo->imo_multicast_ttl = CARP_DFLTTL; 1806 imo->imo_multicast_loop = 0; 1807 return (0); 1808 } 1809 1810 1811 #ifdef INET6 1812 static int 1813 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6) 1814 { 1815 struct ifnet *ifp = sc->sc_carpdev; 1816 struct in6_ifaddr *ia, *ia_if; 1817 int error = 0; 1818 1819 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) { 1820 if (!(sc->sc_if.if_flags & IFF_UP)) 1821 carp_set_state(sc, INIT); 1822 if (sc->sc_naddrs6) 1823 sc->sc_if.if_flags |= IFF_UP; 1824 carp_setrun(sc, 0); 1825 return (0); 1826 } 1827 1828 /* we have to do this by hand to ensure we don't match on ourselves */ 1829 ia_if = NULL; 1830 for (ia = in6_ifaddr; ia; ia = ia->ia_next) { 1831 int i; 1832 1833 for (i = 0; i < 4; i++) { 1834 if ((sin6->sin6_addr.s6_addr32[i] & 1835 ia->ia_prefixmask.sin6_addr.s6_addr32[i]) != 1836 (ia->ia_addr.sin6_addr.s6_addr32[i] & 1837 ia->ia_prefixmask.sin6_addr.s6_addr32[i])) 1838 break; 1839 } 1840 /* and, yeah, we need a multicast-capable iface too */ 1841 if (ia->ia_ifp != &sc->sc_if && 1842 ia->ia_ifp->if_type != IFT_CARP && 1843 (ia->ia_ifp->if_flags & IFF_MULTICAST) && 1844 (i == 4)) { 1845 if (!ia_if) 1846 ia_if = ia; 1847 } 1848 } 1849 1850 if (ia_if) { 1851 ia = ia_if; 1852 if (sc->sc_carpdev) { 1853 if (sc->sc_carpdev != ia->ia_ifp) 1854 return (EADDRNOTAVAIL); 1855 } else { 1856 ifp = ia->ia_ifp; 1857 } 1858 } 1859 1860 if ((error = carp_set_ifp(sc, ifp))) 1861 return (error); 1862 1863 if (sc->sc_carpdev == NULL) 1864 return (EADDRNOTAVAIL); 1865 1866 if (sc->sc_naddrs6 == 0 && (error = carp_join_multicast6(sc)) != 0) 1867 return (error); 1868 1869 sc->sc_naddrs6++; 1870 if (sc->sc_carpdev != NULL) 1871 sc->sc_if.if_flags |= IFF_UP; 1872 carp_set_state(sc, INIT); 1873 carp_setrun(sc, 0); 1874 1875 return (0); 1876 } 1877 1878 static int 1879 carp_join_multicast6(struct carp_softc *sc) 1880 { 1881 struct in6_multi_mship *imm, *imm2; 1882 struct ip6_moptions *im6o = &sc->sc_im6o; 1883 struct sockaddr_in6 addr6; 1884 int error; 1885 1886 /* Join IPv6 CARP multicast group */ 1887 memset(&addr6, 0, sizeof(addr6)); 1888 addr6.sin6_family = AF_INET6; 1889 addr6.sin6_len = sizeof(addr6); 1890 addr6.sin6_addr.s6_addr16[0] = htons(0xff02); 1891 addr6.sin6_addr.s6_addr16[1] = htons(sc->sc_if.if_index); 1892 addr6.sin6_addr.s6_addr8[15] = 0x12; 1893 if ((imm = in6_joingroup(&sc->sc_if, 1894 &addr6.sin6_addr, &error, 0)) == NULL) { 1895 return (error); 1896 } 1897 /* join solicited multicast address */ 1898 memset(&addr6.sin6_addr, 0, sizeof(addr6.sin6_addr)); 1899 addr6.sin6_addr.s6_addr16[0] = htons(0xff02); 1900 addr6.sin6_addr.s6_addr16[1] = htons(sc->sc_if.if_index); 1901 addr6.sin6_addr.s6_addr32[1] = 0; 1902 addr6.sin6_addr.s6_addr32[2] = htonl(1); 1903 addr6.sin6_addr.s6_addr32[3] = 0; 1904 addr6.sin6_addr.s6_addr8[12] = 0xff; 1905 if ((imm2 = in6_joingroup(&sc->sc_if, 1906 &addr6.sin6_addr, &error, 0)) == NULL) { 1907 in6_leavegroup(imm); 1908 return (error); 1909 } 1910 1911 /* apply v6 multicast membership */ 1912 im6o->im6o_multicast_ifp = &sc->sc_if; 1913 if (imm) 1914 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, 1915 i6mm_chain); 1916 if (imm2) 1917 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm2, 1918 i6mm_chain); 1919 1920 return (0); 1921 } 1922 1923 #endif /* INET6 */ 1924 1925 static int 1926 carp_ioctl(struct ifnet *ifp, u_long cmd, void *data) 1927 { 1928 struct lwp *l = curlwp; /* XXX */ 1929 struct carp_softc *sc = ifp->if_softc, *vr; 1930 struct carpreq carpr; 1931 struct ifaddr *ifa; 1932 struct ifreq *ifr; 1933 struct ifnet *cdev = NULL; 1934 int error = 0; 1935 1936 ifa = (struct ifaddr *)data; 1937 ifr = (struct ifreq *)data; 1938 1939 switch (cmd) { 1940 case SIOCINITIFADDR: 1941 switch (ifa->ifa_addr->sa_family) { 1942 #ifdef INET 1943 case AF_INET: 1944 sc->sc_if.if_flags |= IFF_UP; 1945 memcpy(ifa->ifa_dstaddr, ifa->ifa_addr, 1946 sizeof(struct sockaddr)); 1947 error = carp_set_addr(sc, satosin(ifa->ifa_addr)); 1948 break; 1949 #endif /* INET */ 1950 #ifdef INET6 1951 case AF_INET6: 1952 sc->sc_if.if_flags|= IFF_UP; 1953 error = carp_set_addr6(sc, satosin6(ifa->ifa_addr)); 1954 break; 1955 #endif /* INET6 */ 1956 default: 1957 error = EAFNOSUPPORT; 1958 break; 1959 } 1960 break; 1961 1962 case SIOCSIFFLAGS: 1963 if ((error = ifioctl_common(ifp, cmd, data)) != 0) 1964 break; 1965 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) { 1966 callout_stop(&sc->sc_ad_tmo); 1967 callout_stop(&sc->sc_md_tmo); 1968 callout_stop(&sc->sc_md6_tmo); 1969 if (sc->sc_state == MASTER) { 1970 /* we need the interface up to bow out */ 1971 sc->sc_if.if_flags |= IFF_UP; 1972 sc->sc_bow_out = 1; 1973 carp_send_ad(sc); 1974 } 1975 sc->sc_if.if_flags &= ~IFF_UP; 1976 carp_set_state(sc, INIT); 1977 carp_setrun(sc, 0); 1978 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) { 1979 sc->sc_if.if_flags |= IFF_UP; 1980 carp_setrun(sc, 0); 1981 } 1982 break; 1983 1984 case SIOCSVH: 1985 if (l == NULL) 1986 break; 1987 if ((error = kauth_authorize_network(l->l_cred, 1988 KAUTH_NETWORK_INTERFACE, 1989 KAUTH_REQ_NETWORK_INTERFACE_SETPRIV, ifp, (void *)cmd, 1990 NULL)) != 0) 1991 break; 1992 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr))) 1993 break; 1994 error = 1; 1995 if (carpr.carpr_carpdev[0] != '\0' && 1996 (cdev = ifunit(carpr.carpr_carpdev)) == NULL) 1997 return (EINVAL); 1998 if ((error = carp_set_ifp(sc, cdev))) 1999 return (error); 2000 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) { 2001 switch (carpr.carpr_state) { 2002 case BACKUP: 2003 callout_stop(&sc->sc_ad_tmo); 2004 carp_set_state(sc, BACKUP); 2005 carp_setrun(sc, 0); 2006 carp_setroute(sc, RTM_DELETE); 2007 break; 2008 case MASTER: 2009 carp_master_down(sc); 2010 break; 2011 default: 2012 break; 2013 } 2014 } 2015 if (carpr.carpr_vhid > 0) { 2016 if (carpr.carpr_vhid > 255) { 2017 error = EINVAL; 2018 break; 2019 } 2020 if (sc->sc_carpdev) { 2021 struct carp_if *cif; 2022 cif = (struct carp_if *)sc->sc_carpdev->if_carp; 2023 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) 2024 if (vr != sc && 2025 vr->sc_vhid == carpr.carpr_vhid) 2026 return (EINVAL); 2027 } 2028 sc->sc_vhid = carpr.carpr_vhid; 2029 carp_set_enaddr(sc); 2030 carp_set_state(sc, INIT); 2031 error--; 2032 } 2033 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) { 2034 if (carpr.carpr_advskew > 254) { 2035 error = EINVAL; 2036 break; 2037 } 2038 if (carpr.carpr_advbase > 255) { 2039 error = EINVAL; 2040 break; 2041 } 2042 sc->sc_advbase = carpr.carpr_advbase; 2043 sc->sc_advskew = carpr.carpr_advskew; 2044 error--; 2045 } 2046 memcpy(sc->sc_key, carpr.carpr_key, sizeof(sc->sc_key)); 2047 if (error > 0) 2048 error = EINVAL; 2049 else { 2050 error = 0; 2051 carp_setrun(sc, 0); 2052 } 2053 break; 2054 2055 case SIOCGVH: 2056 memset(&carpr, 0, sizeof(carpr)); 2057 if (sc->sc_carpdev != NULL) 2058 strlcpy(carpr.carpr_carpdev, sc->sc_carpdev->if_xname, 2059 IFNAMSIZ); 2060 carpr.carpr_state = sc->sc_state; 2061 carpr.carpr_vhid = sc->sc_vhid; 2062 carpr.carpr_advbase = sc->sc_advbase; 2063 carpr.carpr_advskew = sc->sc_advskew; 2064 2065 if ((l != NULL) && (error = kauth_authorize_network(l->l_cred, 2066 KAUTH_NETWORK_INTERFACE, 2067 KAUTH_REQ_NETWORK_INTERFACE_SETPRIV, ifp, (void *)cmd, 2068 NULL)) == 0) 2069 memcpy(carpr.carpr_key, sc->sc_key, 2070 sizeof(carpr.carpr_key)); 2071 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr)); 2072 break; 2073 2074 case SIOCADDMULTI: 2075 error = carp_ether_addmulti(sc, ifr); 2076 break; 2077 2078 case SIOCDELMULTI: 2079 error = carp_ether_delmulti(sc, ifr); 2080 break; 2081 2082 case SIOCSIFCAP: 2083 if ((error = ifioctl_common(ifp, cmd, data)) == ENETRESET) 2084 error = 0; 2085 break; 2086 2087 default: 2088 error = ether_ioctl(ifp, cmd, data); 2089 } 2090 2091 carp_hmac_prepare(sc); 2092 return (error); 2093 } 2094 2095 2096 /* 2097 * Start output on carp interface. This function should never be called. 2098 */ 2099 static void 2100 carp_start(struct ifnet *ifp) 2101 { 2102 #ifdef DEBUG 2103 printf("%s: start called\n", ifp->if_xname); 2104 #endif 2105 } 2106 2107 int 2108 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa, 2109 const struct rtentry *rt) 2110 { 2111 struct carp_softc *sc = ((struct carp_softc *)ifp->if_softc); 2112 KASSERT(KERNEL_LOCKED_P()); 2113 2114 if (sc->sc_carpdev != NULL && sc->sc_state == MASTER) { 2115 return if_output_lock(sc->sc_carpdev, ifp, m, sa, rt); 2116 } else { 2117 m_freem(m); 2118 return (ENETUNREACH); 2119 } 2120 } 2121 2122 static void 2123 carp_set_state(struct carp_softc *sc, int state) 2124 { 2125 static const char *carp_states[] = { CARP_STATES }; 2126 if (sc->sc_state == state) 2127 return; 2128 2129 CARP_LOG(sc, ("state transition from: %s -> to: %s", carp_states[sc->sc_state], carp_states[state])); 2130 2131 sc->sc_state = state; 2132 switch (state) { 2133 case BACKUP: 2134 sc->sc_if.if_link_state = LINK_STATE_DOWN; 2135 break; 2136 case MASTER: 2137 sc->sc_if.if_link_state = LINK_STATE_UP; 2138 break; 2139 default: 2140 sc->sc_if.if_link_state = LINK_STATE_UNKNOWN; 2141 break; 2142 } 2143 rt_ifmsg(&sc->sc_if); 2144 } 2145 2146 void 2147 carp_carpdev_state(void *v) 2148 { 2149 struct carp_if *cif; 2150 struct carp_softc *sc; 2151 struct ifnet *ifp = v; 2152 2153 if (ifp->if_type == IFT_CARP) 2154 return; 2155 2156 cif = (struct carp_if *)ifp->if_carp; 2157 2158 TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) { 2159 int suppressed = sc->sc_suppress; 2160 2161 if (sc->sc_carpdev->if_link_state == LINK_STATE_DOWN || 2162 !(sc->sc_carpdev->if_flags & IFF_UP)) { 2163 sc->sc_if.if_flags &= ~IFF_RUNNING; 2164 callout_stop(&sc->sc_ad_tmo); 2165 callout_stop(&sc->sc_md_tmo); 2166 callout_stop(&sc->sc_md6_tmo); 2167 carp_set_state(sc, INIT); 2168 sc->sc_suppress = 1; 2169 carp_setrun(sc, 0); 2170 if (!suppressed) { 2171 carp_suppress_preempt++; 2172 if (carp_suppress_preempt == 1) 2173 carp_send_ad_all(); 2174 } 2175 } else { 2176 carp_set_state(sc, INIT); 2177 sc->sc_suppress = 0; 2178 carp_setrun(sc, 0); 2179 if (suppressed) 2180 carp_suppress_preempt--; 2181 } 2182 } 2183 } 2184 2185 static int 2186 carp_ether_addmulti(struct carp_softc *sc, struct ifreq *ifr) 2187 { 2188 const struct sockaddr *sa = ifreq_getaddr(SIOCADDMULTI, ifr); 2189 struct ifnet *ifp; 2190 struct carp_mc_entry *mc; 2191 u_int8_t addrlo[ETHER_ADDR_LEN], addrhi[ETHER_ADDR_LEN]; 2192 int error; 2193 2194 ifp = sc->sc_carpdev; 2195 if (ifp == NULL) 2196 return (EINVAL); 2197 2198 error = ether_addmulti(sa, &sc->sc_ac); 2199 if (error != ENETRESET) 2200 return (error); 2201 2202 /* 2203 * This is new multicast address. We have to tell parent 2204 * about it. Also, remember this multicast address so that 2205 * we can delete them on unconfigure. 2206 */ 2207 mc = malloc(sizeof(struct carp_mc_entry), M_DEVBUF, M_NOWAIT); 2208 if (mc == NULL) { 2209 error = ENOMEM; 2210 goto alloc_failed; 2211 } 2212 2213 /* 2214 * As ether_addmulti() returns ENETRESET, following two 2215 * statement shouldn't fail. 2216 */ 2217 (void)ether_multiaddr(sa, addrlo, addrhi); 2218 ETHER_LOOKUP_MULTI(addrlo, addrhi, &sc->sc_ac, mc->mc_enm); 2219 memcpy(&mc->mc_addr, sa, sa->sa_len); 2220 LIST_INSERT_HEAD(&sc->carp_mc_listhead, mc, mc_entries); 2221 2222 error = if_mcast_op(ifp, SIOCADDMULTI, sa); 2223 if (error != 0) 2224 goto ioctl_failed; 2225 2226 return (error); 2227 2228 ioctl_failed: 2229 LIST_REMOVE(mc, mc_entries); 2230 free(mc, M_DEVBUF); 2231 alloc_failed: 2232 (void)ether_delmulti(sa, &sc->sc_ac); 2233 2234 return (error); 2235 } 2236 2237 static int 2238 carp_ether_delmulti(struct carp_softc *sc, struct ifreq *ifr) 2239 { 2240 const struct sockaddr *sa = ifreq_getaddr(SIOCDELMULTI, ifr); 2241 struct ifnet *ifp; 2242 struct ether_multi *enm; 2243 struct carp_mc_entry *mc; 2244 u_int8_t addrlo[ETHER_ADDR_LEN], addrhi[ETHER_ADDR_LEN]; 2245 int error; 2246 2247 ifp = sc->sc_carpdev; 2248 if (ifp == NULL) 2249 return (EINVAL); 2250 2251 /* 2252 * Find a key to lookup carp_mc_entry. We have to do this 2253 * before calling ether_delmulti for obvious reason. 2254 */ 2255 if ((error = ether_multiaddr(sa, addrlo, addrhi)) != 0) 2256 return (error); 2257 ETHER_LOOKUP_MULTI(addrlo, addrhi, &sc->sc_ac, enm); 2258 if (enm == NULL) 2259 return (EINVAL); 2260 2261 LIST_FOREACH(mc, &sc->carp_mc_listhead, mc_entries) 2262 if (mc->mc_enm == enm) 2263 break; 2264 2265 /* We won't delete entries we didn't add */ 2266 if (mc == NULL) 2267 return (EINVAL); 2268 2269 error = ether_delmulti(sa, &sc->sc_ac); 2270 if (error != ENETRESET) 2271 return (error); 2272 2273 /* We no longer use this multicast address. Tell parent so. */ 2274 error = if_mcast_op(ifp, SIOCDELMULTI, sa); 2275 if (error == 0) { 2276 /* And forget about this address. */ 2277 LIST_REMOVE(mc, mc_entries); 2278 free(mc, M_DEVBUF); 2279 } else 2280 (void)ether_addmulti(sa, &sc->sc_ac); 2281 return (error); 2282 } 2283 2284 /* 2285 * Delete any multicast address we have asked to add from parent 2286 * interface. Called when the carp is being unconfigured. 2287 */ 2288 static void 2289 carp_ether_purgemulti(struct carp_softc *sc) 2290 { 2291 struct ifnet *ifp = sc->sc_carpdev; /* Parent. */ 2292 struct carp_mc_entry *mc; 2293 2294 if (ifp == NULL) 2295 return; 2296 2297 while ((mc = LIST_FIRST(&sc->carp_mc_listhead)) != NULL) { 2298 (void)if_mcast_op(ifp, SIOCDELMULTI, sstosa(&mc->mc_addr)); 2299 LIST_REMOVE(mc, mc_entries); 2300 free(mc, M_DEVBUF); 2301 } 2302 } 2303 2304 static int 2305 sysctl_net_inet_carp_stats(SYSCTLFN_ARGS) 2306 { 2307 2308 return (NETSTAT_SYSCTL(carpstat_percpu, CARP_NSTATS)); 2309 } 2310 2311 void 2312 carp_init(void) 2313 { 2314 2315 sysctl_net_inet_carp_setup(NULL); 2316 #ifdef MBUFTRACE 2317 MOWNER_ATTACH(&carp_proto_mowner_rx); 2318 MOWNER_ATTACH(&carp_proto_mowner_tx); 2319 MOWNER_ATTACH(&carp_proto6_mowner_rx); 2320 MOWNER_ATTACH(&carp_proto6_mowner_tx); 2321 #endif 2322 } 2323 2324 static void 2325 sysctl_net_inet_carp_setup(struct sysctllog **clog) 2326 { 2327 2328 sysctl_createv(clog, 0, NULL, NULL, 2329 CTLFLAG_PERMANENT, 2330 CTLTYPE_NODE, "inet", NULL, 2331 NULL, 0, NULL, 0, 2332 CTL_NET, PF_INET, CTL_EOL); 2333 sysctl_createv(clog, 0, NULL, NULL, 2334 CTLFLAG_PERMANENT, 2335 CTLTYPE_NODE, "carp", 2336 SYSCTL_DESCR("CARP related settings"), 2337 NULL, 0, NULL, 0, 2338 CTL_NET, PF_INET, IPPROTO_CARP, CTL_EOL); 2339 2340 sysctl_createv(clog, 0, NULL, NULL, 2341 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2342 CTLTYPE_INT, "preempt", 2343 SYSCTL_DESCR("Enable CARP Preempt"), 2344 NULL, 0, &carp_opts[CARPCTL_PREEMPT], 0, 2345 CTL_NET, PF_INET, IPPROTO_CARP, 2346 CTL_CREATE, CTL_EOL); 2347 sysctl_createv(clog, 0, NULL, NULL, 2348 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2349 CTLTYPE_INT, "arpbalance", 2350 SYSCTL_DESCR("Enable ARP balancing"), 2351 NULL, 0, &carp_opts[CARPCTL_ARPBALANCE], 0, 2352 CTL_NET, PF_INET, IPPROTO_CARP, 2353 CTL_CREATE, CTL_EOL); 2354 sysctl_createv(clog, 0, NULL, NULL, 2355 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2356 CTLTYPE_INT, "allow", 2357 SYSCTL_DESCR("Enable CARP"), 2358 NULL, 0, &carp_opts[CARPCTL_ALLOW], 0, 2359 CTL_NET, PF_INET, IPPROTO_CARP, 2360 CTL_CREATE, CTL_EOL); 2361 sysctl_createv(clog, 0, NULL, NULL, 2362 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2363 CTLTYPE_INT, "log", 2364 SYSCTL_DESCR("CARP logging"), 2365 NULL, 0, &carp_opts[CARPCTL_LOG], 0, 2366 CTL_NET, PF_INET, IPPROTO_CARP, 2367 CTL_CREATE, CTL_EOL); 2368 sysctl_createv(clog, 0, NULL, NULL, 2369 CTLFLAG_PERMANENT, 2370 CTLTYPE_STRUCT, "stats", 2371 SYSCTL_DESCR("CARP statistics"), 2372 sysctl_net_inet_carp_stats, 0, NULL, 0, 2373 CTL_NET, PF_INET, IPPROTO_CARP, CARPCTL_STATS, 2374 CTL_EOL); 2375 } 2376
Indexes created Wed Sep 24 15:09:51 GMT 2025