1 /* $NetBSD: ip_carp.c,v 1.80 2016/12/12 03:55:57 ozaki-r 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 1.80 2016/12/12 03:55:57 ozaki-r 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; 693 694 CARP_STATINC(CARP_STAT_BADAUTH); 695 sc->sc_if.if_ierrors++; 696 697 switch(af) { 698 699 case AF_INET: 700 ip = mtod(m, struct ip *); 701 CARP_LOG(sc, ("incorrect hash from %s", 702 in_fmtaddr(ip->ip_src))); 703 break; 704 705 case AF_INET6: 706 ip6 = mtod(m, struct ip6_hdr *); 707 CARP_LOG(sc, ("incorrect hash from %s", 708 ip6_sprintf(&ip6->ip6_src))); 709 break; 710 711 default: CARP_LOG(sc, ("incorrect hash")); 712 break; 713 } 714 m_freem(m); 715 return; 716 } 717 718 tmp_counter = ntohl(ch->carp_counter[0]); 719 tmp_counter = tmp_counter<<32; 720 tmp_counter += ntohl(ch->carp_counter[1]); 721 722 /* XXX Replay protection goes here */ 723 724 sc->sc_init_counter = 0; 725 sc->sc_counter = tmp_counter; 726 727 728 sc_tv.tv_sec = sc->sc_advbase; 729 if (carp_suppress_preempt && sc->sc_advskew < 240) 730 sc_tv.tv_usec = 240 * 1000000 / 256; 731 else 732 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256; 733 ch_tv.tv_sec = ch->carp_advbase; 734 ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256; 735 736 switch (sc->sc_state) { 737 case INIT: 738 break; 739 case MASTER: 740 /* 741 * If we receive an advertisement from a backup who's going to 742 * be more frequent than us, go into BACKUP state. 743 */ 744 if (timercmp(&sc_tv, &ch_tv, >) || 745 timercmp(&sc_tv, &ch_tv, ==)) { 746 callout_stop(&sc->sc_ad_tmo); 747 CARP_LOG(sc, ("MASTER -> BACKUP (more frequent advertisement received)")); 748 carp_set_state(sc, BACKUP); 749 carp_setrun(sc, 0); 750 carp_setroute(sc, RTM_DELETE); 751 } 752 break; 753 case BACKUP: 754 /* 755 * If we're pre-empting masters who advertise slower than us, 756 * and this one claims to be slower, treat him as down. 757 */ 758 if (carp_opts[CARPCTL_PREEMPT] && timercmp(&sc_tv, &ch_tv, <)) { 759 CARP_LOG(sc, ("BACKUP -> MASTER (preempting a slower master)")); 760 carp_master_down(sc); 761 break; 762 } 763 764 /* 765 * If the master is going to advertise at such a low frequency 766 * that he's guaranteed to time out, we'd might as well just 767 * treat him as timed out now. 768 */ 769 sc_tv.tv_sec = sc->sc_advbase * 3; 770 if (timercmp(&sc_tv, &ch_tv, <)) { 771 CARP_LOG(sc, ("BACKUP -> MASTER (master timed out)")); 772 carp_master_down(sc); 773 break; 774 } 775 776 /* 777 * Otherwise, we reset the counter and wait for the next 778 * advertisement. 779 */ 780 carp_setrun(sc, af); 781 break; 782 } 783 784 m_freem(m); 785 return; 786 } 787 788 /* 789 * Interface side of the CARP implementation. 790 */ 791 792 /* ARGSUSED */ 793 void 794 carpattach(int n) 795 { 796 if_clone_attach(&carp_cloner); 797 798 carpstat_percpu = percpu_alloc(sizeof(uint64_t) * CARP_NSTATS); 799 } 800 801 static int 802 carp_clone_create(struct if_clone *ifc, int unit) 803 { 804 extern int ifqmaxlen; 805 struct carp_softc *sc; 806 struct ifnet *ifp; 807 808 sc = malloc(sizeof(*sc), M_DEVBUF, M_NOWAIT|M_ZERO); 809 if (!sc) 810 return (ENOMEM); 811 812 sc->sc_suppress = 0; 813 sc->sc_advbase = CARP_DFLTINTV; 814 sc->sc_vhid = -1; /* required setting */ 815 sc->sc_advskew = 0; 816 sc->sc_init_counter = 1; 817 sc->sc_naddrs = sc->sc_naddrs6 = 0; 818 #ifdef INET6 819 sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL; 820 #endif /* INET6 */ 821 822 callout_init(&sc->sc_ad_tmo, 0); 823 callout_init(&sc->sc_md_tmo, 0); 824 callout_init(&sc->sc_md6_tmo, 0); 825 826 callout_setfunc(&sc->sc_ad_tmo, carp_send_ad, sc); 827 callout_setfunc(&sc->sc_md_tmo, carp_master_down, sc); 828 callout_setfunc(&sc->sc_md6_tmo, carp_master_down, sc); 829 830 LIST_INIT(&sc->carp_mc_listhead); 831 ifp = &sc->sc_if; 832 ifp->if_softc = sc; 833 snprintf(ifp->if_xname, sizeof ifp->if_xname, "%s%d", ifc->ifc_name, 834 unit); 835 ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST; 836 ifp->if_ioctl = carp_ioctl; 837 ifp->if_start = carp_start; 838 ifp->if_output = carp_output; 839 ifp->if_type = IFT_CARP; 840 ifp->if_addrlen = ETHER_ADDR_LEN; 841 ifp->if_hdrlen = ETHER_HDR_LEN; 842 ifp->if_mtu = ETHERMTU; 843 IFQ_SET_MAXLEN(&ifp->if_snd, ifqmaxlen); 844 IFQ_SET_READY(&ifp->if_snd); 845 if_attach(ifp); 846 847 if_alloc_sadl(ifp); 848 ifp->if_broadcastaddr = etherbroadcastaddr; 849 carp_set_enaddr(sc); 850 LIST_INIT(&sc->sc_ac.ec_multiaddrs); 851 bpf_attach(ifp, DLT_EN10MB, ETHER_HDR_LEN); 852 #ifdef MBUFTRACE 853 strlcpy(sc->sc_ac.ec_tx_mowner.mo_name, ifp->if_xname, 854 sizeof(sc->sc_ac.ec_tx_mowner.mo_name)); 855 strlcpy(sc->sc_ac.ec_tx_mowner.mo_descr, "tx", 856 sizeof(sc->sc_ac.ec_tx_mowner.mo_descr)); 857 strlcpy(sc->sc_ac.ec_rx_mowner.mo_name, ifp->if_xname, 858 sizeof(sc->sc_ac.ec_rx_mowner.mo_name)); 859 strlcpy(sc->sc_ac.ec_rx_mowner.mo_descr, "rx", 860 sizeof(sc->sc_ac.ec_rx_mowner.mo_descr)); 861 MOWNER_ATTACH(&sc->sc_ac.ec_tx_mowner); 862 MOWNER_ATTACH(&sc->sc_ac.ec_rx_mowner); 863 ifp->if_mowner = &sc->sc_ac.ec_tx_mowner; 864 #endif 865 return (0); 866 } 867 868 static int 869 carp_clone_destroy(struct ifnet *ifp) 870 { 871 struct carp_softc *sc = ifp->if_softc; 872 873 carpdetach(ifp->if_softc); 874 ether_ifdetach(ifp); 875 if_detach(ifp); 876 callout_destroy(&sc->sc_ad_tmo); 877 callout_destroy(&sc->sc_md_tmo); 878 callout_destroy(&sc->sc_md6_tmo); 879 free(ifp->if_softc, M_DEVBUF); 880 881 return (0); 882 } 883 884 static void 885 carpdetach(struct carp_softc *sc) 886 { 887 struct carp_if *cif; 888 int s; 889 890 callout_stop(&sc->sc_ad_tmo); 891 callout_stop(&sc->sc_md_tmo); 892 callout_stop(&sc->sc_md6_tmo); 893 894 if (sc->sc_suppress) 895 carp_suppress_preempt--; 896 sc->sc_suppress = 0; 897 898 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) 899 carp_suppress_preempt--; 900 sc->sc_sendad_errors = 0; 901 902 carp_set_state(sc, INIT); 903 sc->sc_if.if_flags &= ~IFF_UP; 904 carp_setrun(sc, 0); 905 carp_multicast_cleanup(sc); 906 907 KERNEL_LOCK(1, NULL); 908 s = splnet(); 909 if (sc->sc_carpdev != NULL) { 910 /* XXX linkstatehook removal */ 911 cif = (struct carp_if *)sc->sc_carpdev->if_carp; 912 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list); 913 if (!--cif->vhif_nvrs) { 914 ifpromisc(sc->sc_carpdev, 0); 915 sc->sc_carpdev->if_carp = NULL; 916 free(cif, M_IFADDR); 917 } 918 } 919 sc->sc_carpdev = NULL; 920 splx(s); 921 KERNEL_UNLOCK_ONE(NULL); 922 } 923 924 /* Detach an interface from the carp. */ 925 void 926 carp_ifdetach(struct ifnet *ifp) 927 { 928 struct carp_softc *sc, *nextsc; 929 struct carp_if *cif = (struct carp_if *)ifp->if_carp; 930 931 for (sc = TAILQ_FIRST(&cif->vhif_vrs); sc; sc = nextsc) { 932 nextsc = TAILQ_NEXT(sc, sc_list); 933 carpdetach(sc); 934 } 935 } 936 937 static int 938 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, 939 struct carp_header *ch) 940 { 941 if (sc->sc_init_counter) { 942 /* this could also be seconds since unix epoch */ 943 sc->sc_counter = cprng_fast64(); 944 } else 945 sc->sc_counter++; 946 947 ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff); 948 ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff); 949 950 carp_hmac_generate(sc, ch->carp_counter, ch->carp_md); 951 952 return (0); 953 } 954 955 static void 956 carp_send_ad_all(void) 957 { 958 struct ifnet *ifp; 959 struct carp_if *cif; 960 struct carp_softc *vh; 961 int s; 962 int bound = curlwp_bind(); 963 964 s = pserialize_read_enter(); 965 IFNET_READER_FOREACH(ifp) { 966 struct psref psref; 967 if (ifp->if_carp == NULL || ifp->if_type == IFT_CARP) 968 continue; 969 970 psref_acquire(&psref, &ifp->if_psref, ifnet_psref_class); 971 pserialize_read_exit(s); 972 973 cif = (struct carp_if *)ifp->if_carp; 974 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 975 if ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 976 (IFF_UP|IFF_RUNNING) && vh->sc_state == MASTER) 977 carp_send_ad(vh); 978 } 979 980 s = pserialize_read_enter(); 981 psref_release(&psref, &ifp->if_psref, ifnet_psref_class); 982 } 983 pserialize_read_exit(s); 984 curlwp_bindx(bound); 985 } 986 987 988 static void 989 carp_send_ad(void *v) 990 { 991 struct carp_header ch; 992 struct timeval tv; 993 struct carp_softc *sc = v; 994 struct carp_header *ch_ptr; 995 struct mbuf *m; 996 int error, len, advbase, advskew, s; 997 struct sockaddr sa; 998 999 KERNEL_LOCK(1, NULL); 1000 s = splsoftnet(); 1001 1002 advbase = advskew = 0; /* Sssssh compiler */ 1003 if (sc->sc_carpdev == NULL) { 1004 sc->sc_if.if_oerrors++; 1005 goto retry_later; 1006 } 1007 1008 /* bow out if we've gone to backup (the carp interface is going down) */ 1009 if (sc->sc_bow_out) { 1010 sc->sc_bow_out = 0; 1011 advbase = 255; 1012 advskew = 255; 1013 } else { 1014 advbase = sc->sc_advbase; 1015 if (!carp_suppress_preempt || sc->sc_advskew > 240) 1016 advskew = sc->sc_advskew; 1017 else 1018 advskew = 240; 1019 tv.tv_sec = advbase; 1020 tv.tv_usec = advskew * 1000000 / 256; 1021 } 1022 1023 ch.carp_version = CARP_VERSION; 1024 ch.carp_type = CARP_ADVERTISEMENT; 1025 ch.carp_vhid = sc->sc_vhid; 1026 ch.carp_advbase = advbase; 1027 ch.carp_advskew = advskew; 1028 ch.carp_authlen = 7; /* XXX DEFINE */ 1029 ch.carp_pad1 = 0; /* must be zero */ 1030 ch.carp_cksum = 0; 1031 1032 1033 #ifdef INET 1034 if (sc->sc_naddrs) { 1035 struct ip *ip; 1036 struct ifaddr *ifa; 1037 int _s; 1038 1039 MGETHDR(m, M_DONTWAIT, MT_HEADER); 1040 if (m == NULL) { 1041 sc->sc_if.if_oerrors++; 1042 CARP_STATINC(CARP_STAT_ONOMEM); 1043 /* XXX maybe less ? */ 1044 goto retry_later; 1045 } 1046 MCLAIM(m, &carp_proto_mowner_tx); 1047 len = sizeof(*ip) + sizeof(ch); 1048 m->m_pkthdr.len = len; 1049 m_reset_rcvif(m); 1050 m->m_len = len; 1051 MH_ALIGN(m, m->m_len); 1052 m->m_flags |= M_MCAST; 1053 ip = mtod(m, struct ip *); 1054 ip->ip_v = IPVERSION; 1055 ip->ip_hl = sizeof(*ip) >> 2; 1056 ip->ip_tos = IPTOS_LOWDELAY; 1057 ip->ip_len = htons(len); 1058 ip->ip_id = 0; /* no need for id, we don't support fragments */ 1059 ip->ip_off = htons(IP_DF); 1060 ip->ip_ttl = CARP_DFLTTL; 1061 ip->ip_p = IPPROTO_CARP; 1062 ip->ip_sum = 0; 1063 1064 memset(&sa, 0, sizeof(sa)); 1065 sa.sa_family = AF_INET; 1066 _s = pserialize_read_enter(); 1067 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev); 1068 if (ifa == NULL) 1069 ip->ip_src.s_addr = 0; 1070 else 1071 ip->ip_src.s_addr = 1072 ifatoia(ifa)->ia_addr.sin_addr.s_addr; 1073 pserialize_read_exit(_s); 1074 ip->ip_dst.s_addr = INADDR_CARP_GROUP; 1075 1076 ch_ptr = (struct carp_header *)(&ip[1]); 1077 memcpy(ch_ptr, &ch, sizeof(ch)); 1078 if (carp_prepare_ad(m, sc, ch_ptr)) 1079 goto retry_later; 1080 1081 m->m_data += sizeof(*ip); 1082 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip)); 1083 m->m_data -= sizeof(*ip); 1084 1085 nanotime(&sc->sc_if.if_lastchange); 1086 sc->sc_if.if_opackets++; 1087 sc->sc_if.if_obytes += len; 1088 CARP_STATINC(CARP_STAT_OPACKETS); 1089 1090 error = ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, 1091 NULL); 1092 if (error) { 1093 if (error == ENOBUFS) 1094 CARP_STATINC(CARP_STAT_ONOMEM); 1095 else 1096 CARP_LOG(sc, ("ip_output failed: %d", error)); 1097 sc->sc_if.if_oerrors++; 1098 if (sc->sc_sendad_errors < INT_MAX) 1099 sc->sc_sendad_errors++; 1100 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) { 1101 carp_suppress_preempt++; 1102 if (carp_suppress_preempt == 1) 1103 carp_send_ad_all(); 1104 } 1105 sc->sc_sendad_success = 0; 1106 } else { 1107 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) { 1108 if (++sc->sc_sendad_success >= 1109 CARP_SENDAD_MIN_SUCCESS) { 1110 carp_suppress_preempt--; 1111 sc->sc_sendad_errors = 0; 1112 } 1113 } else 1114 sc->sc_sendad_errors = 0; 1115 } 1116 } 1117 #endif /* INET */ 1118 #ifdef INET6_notyet 1119 if (sc->sc_naddrs6) { 1120 struct ip6_hdr *ip6; 1121 struct ifaddr *ifa; 1122 int _s; 1123 1124 MGETHDR(m, M_DONTWAIT, MT_HEADER); 1125 if (m == NULL) { 1126 sc->sc_if.if_oerrors++; 1127 CARP_STATINC(CARP_STAT_ONOMEM); 1128 /* XXX maybe less ? */ 1129 goto retry_later; 1130 } 1131 MCLAIM(m, &carp_proto6_mowner_tx); 1132 len = sizeof(*ip6) + sizeof(ch); 1133 m->m_pkthdr.len = len; 1134 m_reset_rcvif(m); 1135 m->m_len = len; 1136 MH_ALIGN(m, m->m_len); 1137 m->m_flags |= M_MCAST; 1138 ip6 = mtod(m, struct ip6_hdr *); 1139 memset(ip6, 0, sizeof(*ip6)); 1140 ip6->ip6_vfc |= IPV6_VERSION; 1141 ip6->ip6_hlim = CARP_DFLTTL; 1142 ip6->ip6_nxt = IPPROTO_CARP; 1143 1144 /* set the source address */ 1145 memset(&sa, 0, sizeof(sa)); 1146 sa.sa_family = AF_INET6; 1147 _s = pserialize_read_enter(); 1148 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev); 1149 if (ifa == NULL) /* This should never happen with IPv6 */ 1150 memset(&ip6->ip6_src, 0, sizeof(struct in6_addr)); 1151 else 1152 bcopy(ifatoia6(ifa)->ia_addr.sin6_addr.s6_addr, 1153 &ip6->ip6_src, sizeof(struct in6_addr)); 1154 pserialize_read_exit(_s); 1155 /* set the multicast destination */ 1156 1157 ip6->ip6_dst.s6_addr16[0] = htons(0xff02); 1158 ip6->ip6_dst.s6_addr8[15] = 0x12; 1159 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) { 1160 sc->sc_if.if_oerrors++; 1161 m_freem(m); 1162 CARP_LOG(sc, ("in6_setscope failed")); 1163 goto retry_later; 1164 } 1165 1166 ch_ptr = (struct carp_header *)(&ip6[1]); 1167 memcpy(ch_ptr, &ch, sizeof(ch)); 1168 if (carp_prepare_ad(m, sc, ch_ptr)) 1169 goto retry_later; 1170 1171 m->m_data += sizeof(*ip6); 1172 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6)); 1173 m->m_data -= sizeof(*ip6); 1174 1175 nanotime(&sc->sc_if.if_lastchange); 1176 sc->sc_if.if_opackets++; 1177 sc->sc_if.if_obytes += len; 1178 CARP_STATINC(CARP_STAT_OPACKETS6); 1179 1180 error = ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL); 1181 if (error) { 1182 if (error == ENOBUFS) 1183 CARP_STATINC(CARP_STAT_ONOMEM); 1184 else 1185 CARP_LOG(sc, ("ip6_output failed: %d", error)); 1186 sc->sc_if.if_oerrors++; 1187 if (sc->sc_sendad_errors < INT_MAX) 1188 sc->sc_sendad_errors++; 1189 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) { 1190 carp_suppress_preempt++; 1191 if (carp_suppress_preempt == 1) 1192 carp_send_ad_all(); 1193 } 1194 sc->sc_sendad_success = 0; 1195 } else { 1196 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) { 1197 if (++sc->sc_sendad_success >= 1198 CARP_SENDAD_MIN_SUCCESS) { 1199 carp_suppress_preempt--; 1200 sc->sc_sendad_errors = 0; 1201 } 1202 } else 1203 sc->sc_sendad_errors = 0; 1204 } 1205 } 1206 #endif /* INET6 */ 1207 1208 retry_later: 1209 splx(s); 1210 KERNEL_UNLOCK_ONE(NULL); 1211 if (advbase != 255 || advskew != 255) 1212 callout_schedule(&sc->sc_ad_tmo, tvtohz(&tv)); 1213 } 1214 1215 /* 1216 * Broadcast a gratuitous ARP request containing 1217 * the virtual router MAC address for each IP address 1218 * associated with the virtual router. 1219 */ 1220 static void 1221 carp_send_arp(struct carp_softc *sc) 1222 { 1223 struct ifaddr *ifa; 1224 int s; 1225 1226 KERNEL_LOCK(1, NULL); 1227 s = splsoftnet(); 1228 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 1229 1230 if (ifa->ifa_addr->sa_family != AF_INET) 1231 continue; 1232 1233 arpannounce(sc->sc_carpdev, ifa, CLLADDR(sc->sc_if.if_sadl)); 1234 } 1235 splx(s); 1236 KERNEL_UNLOCK_ONE(NULL); 1237 } 1238 1239 #ifdef INET6 1240 static void 1241 carp_send_na(struct carp_softc *sc) 1242 { 1243 struct ifaddr *ifa; 1244 struct in6_addr *in6; 1245 static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT; 1246 int s; 1247 1248 KERNEL_LOCK(1, NULL); 1249 s = splsoftnet(); 1250 1251 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 1252 1253 if (ifa->ifa_addr->sa_family != AF_INET6) 1254 continue; 1255 1256 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr; 1257 nd6_na_output(sc->sc_carpdev, &mcast, in6, 1258 ND_NA_FLAG_OVERRIDE, 1, NULL); 1259 } 1260 splx(s); 1261 KERNEL_UNLOCK_ONE(NULL); 1262 } 1263 #endif /* INET6 */ 1264 1265 /* 1266 * Based on bridge_hash() in if_bridge.c 1267 */ 1268 #define mix(a,b,c) \ 1269 do { \ 1270 a -= b; a -= c; a ^= (c >> 13); \ 1271 b -= c; b -= a; b ^= (a << 8); \ 1272 c -= a; c -= b; c ^= (b >> 13); \ 1273 a -= b; a -= c; a ^= (c >> 12); \ 1274 b -= c; b -= a; b ^= (a << 16); \ 1275 c -= a; c -= b; c ^= (b >> 5); \ 1276 a -= b; a -= c; a ^= (c >> 3); \ 1277 b -= c; b -= a; b ^= (a << 10); \ 1278 c -= a; c -= b; c ^= (b >> 15); \ 1279 } while (0) 1280 1281 static u_int32_t 1282 carp_hash(struct carp_softc *sc, u_char *src) 1283 { 1284 u_int32_t a = 0x9e3779b9, b = sc->sc_hashkey[0], c = sc->sc_hashkey[1]; 1285 1286 c += sc->sc_key[3] << 24; 1287 c += sc->sc_key[2] << 16; 1288 c += sc->sc_key[1] << 8; 1289 c += sc->sc_key[0]; 1290 b += src[5] << 8; 1291 b += src[4]; 1292 a += src[3] << 24; 1293 a += src[2] << 16; 1294 a += src[1] << 8; 1295 a += src[0]; 1296 1297 mix(a, b, c); 1298 return (c); 1299 } 1300 1301 static int 1302 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type) 1303 { 1304 struct carp_softc *vh; 1305 struct ifaddr *ifa; 1306 int count = 0; 1307 1308 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1309 if ((type == CARP_COUNT_RUNNING && 1310 (vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1311 (IFF_UP|IFF_RUNNING)) || 1312 (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) { 1313 IFADDR_READER_FOREACH(ifa, &vh->sc_if) { 1314 if (ifa->ifa_addr->sa_family == AF_INET && 1315 ia->ia_addr.sin_addr.s_addr == 1316 ifatoia(ifa)->ia_addr.sin_addr.s_addr) 1317 count++; 1318 } 1319 } 1320 } 1321 return (count); 1322 } 1323 1324 int 1325 carp_iamatch(struct in_ifaddr *ia, u_char *src, 1326 u_int32_t *count, u_int32_t index) 1327 { 1328 struct carp_softc *sc = ia->ia_ifp->if_softc; 1329 1330 if (carp_opts[CARPCTL_ARPBALANCE]) { 1331 /* 1332 * We use the source ip to decide which virtual host should 1333 * handle the request. If we're master of that virtual host, 1334 * then we respond, otherwise, just drop the arp packet on 1335 * the floor. 1336 */ 1337 1338 /* Count the elegible carp interfaces with this address */ 1339 if (*count == 0) 1340 *count = carp_addrcount( 1341 (struct carp_if *)ia->ia_ifp->if_carpdev->if_carp, 1342 ia, CARP_COUNT_RUNNING); 1343 1344 /* This should never happen, but... */ 1345 if (*count == 0) 1346 return (0); 1347 1348 if (carp_hash(sc, src) % *count == index - 1 && 1349 sc->sc_state == MASTER) { 1350 return (1); 1351 } 1352 } else { 1353 if (sc->sc_state == MASTER) 1354 return (1); 1355 } 1356 1357 return (0); 1358 } 1359 1360 #ifdef INET6 1361 struct ifaddr * 1362 carp_iamatch6(void *v, struct in6_addr *taddr) 1363 { 1364 struct carp_if *cif = v; 1365 struct carp_softc *vh; 1366 struct ifaddr *ifa; 1367 1368 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1369 IFADDR_READER_FOREACH(ifa, &vh->sc_if) { 1370 if (IN6_ARE_ADDR_EQUAL(taddr, 1371 &ifatoia6(ifa)->ia_addr.sin6_addr) && 1372 ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1373 (IFF_UP|IFF_RUNNING)) && vh->sc_state == MASTER) 1374 return (ifa); 1375 } 1376 } 1377 1378 return (NULL); 1379 } 1380 #endif /* INET6 */ 1381 1382 struct ifnet * 1383 carp_ourether(void *v, struct ether_header *eh, u_char iftype, int src) 1384 { 1385 struct carp_if *cif = (struct carp_if *)v; 1386 struct carp_softc *vh; 1387 u_int8_t *ena; 1388 1389 if (src) 1390 ena = (u_int8_t *)&eh->ether_shost; 1391 else 1392 ena = (u_int8_t *)&eh->ether_dhost; 1393 1394 switch (iftype) { 1395 case IFT_ETHER: 1396 case IFT_FDDI: 1397 if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1) 1398 return (NULL); 1399 break; 1400 case IFT_ISO88025: 1401 if (ena[0] != 3 || ena[1] || ena[4] || ena[5]) 1402 return (NULL); 1403 break; 1404 default: 1405 return (NULL); 1406 break; 1407 } 1408 1409 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) 1410 if ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1411 (IFF_UP|IFF_RUNNING) && vh->sc_state == MASTER && 1412 !memcmp(ena, CLLADDR(vh->sc_if.if_sadl), 1413 ETHER_ADDR_LEN)) { 1414 return (&vh->sc_if); 1415 } 1416 1417 return (NULL); 1418 } 1419 1420 int 1421 carp_input(struct mbuf *m, u_int8_t *shost, u_int8_t *dhost, u_int16_t etype) 1422 { 1423 struct ether_header eh; 1424 struct carp_if *cif = (struct carp_if *)m_get_rcvif_NOMPSAFE(m)->if_carp; 1425 struct ifnet *ifp; 1426 1427 memcpy(&eh.ether_shost, shost, sizeof(eh.ether_shost)); 1428 memcpy(&eh.ether_dhost, dhost, sizeof(eh.ether_dhost)); 1429 eh.ether_type = etype; 1430 1431 if (m->m_flags & (M_BCAST|M_MCAST)) { 1432 struct carp_softc *vh; 1433 struct mbuf *m0; 1434 1435 /* 1436 * XXX Should really check the list of multicast addresses 1437 * for each CARP interface _before_ copying. 1438 */ 1439 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1440 m0 = m_copym(m, 0, M_COPYALL, M_DONTWAIT); 1441 if (m0 == NULL) 1442 continue; 1443 m_set_rcvif(m0, &vh->sc_if); 1444 ether_input(&vh->sc_if, m0); 1445 } 1446 return (1); 1447 } 1448 1449 ifp = carp_ourether(cif, &eh, m_get_rcvif_NOMPSAFE(m)->if_type, 0); 1450 if (ifp == NULL) { 1451 return (1); 1452 } 1453 1454 m_set_rcvif(m, ifp); 1455 1456 bpf_mtap(ifp, m); 1457 ifp->if_ipackets++; 1458 ether_input(ifp, m); 1459 return (0); 1460 } 1461 1462 static void 1463 carp_master_down(void *v) 1464 { 1465 struct carp_softc *sc = v; 1466 1467 switch (sc->sc_state) { 1468 case INIT: 1469 printf("%s: master_down event in INIT state\n", 1470 sc->sc_if.if_xname); 1471 break; 1472 case MASTER: 1473 break; 1474 case BACKUP: 1475 CARP_LOG(sc, ("INIT -> MASTER (preempting)")); 1476 carp_set_state(sc, MASTER); 1477 carp_send_ad(sc); 1478 carp_send_arp(sc); 1479 #ifdef INET6 1480 carp_send_na(sc); 1481 #endif /* INET6 */ 1482 carp_setrun(sc, 0); 1483 carp_setroute(sc, RTM_ADD); 1484 break; 1485 } 1486 } 1487 1488 /* 1489 * When in backup state, af indicates whether to reset the master down timer 1490 * for v4 or v6. If it's set to zero, reset the ones which are already pending. 1491 */ 1492 static void 1493 carp_setrun(struct carp_softc *sc, sa_family_t af) 1494 { 1495 struct timeval tv; 1496 1497 if (sc->sc_carpdev == NULL) { 1498 sc->sc_if.if_flags &= ~IFF_RUNNING; 1499 carp_set_state(sc, INIT); 1500 return; 1501 } 1502 1503 if (sc->sc_if.if_flags & IFF_UP && sc->sc_vhid > 0 && 1504 (sc->sc_naddrs || sc->sc_naddrs6) && !sc->sc_suppress) { 1505 sc->sc_if.if_flags |= IFF_RUNNING; 1506 } else { 1507 sc->sc_if.if_flags &= ~IFF_RUNNING; 1508 carp_setroute(sc, RTM_DELETE); 1509 return; 1510 } 1511 1512 switch (sc->sc_state) { 1513 case INIT: 1514 carp_set_state(sc, BACKUP); 1515 carp_setroute(sc, RTM_DELETE); 1516 carp_setrun(sc, 0); 1517 break; 1518 case BACKUP: 1519 callout_stop(&sc->sc_ad_tmo); 1520 tv.tv_sec = 3 * sc->sc_advbase; 1521 tv.tv_usec = sc->sc_advskew * 1000000 / 256; 1522 switch (af) { 1523 #ifdef INET 1524 case AF_INET: 1525 callout_schedule(&sc->sc_md_tmo, tvtohz(&tv)); 1526 break; 1527 #endif /* INET */ 1528 #ifdef INET6_notyet 1529 case AF_INET6: 1530 callout_schedule(&sc->sc_md6_tmo, tvtohz(&tv)); 1531 break; 1532 #endif /* INET6 */ 1533 default: 1534 if (sc->sc_naddrs) 1535 callout_schedule(&sc->sc_md_tmo, tvtohz(&tv)); 1536 #ifdef INET6_notyet 1537 if (sc->sc_naddrs6) 1538 callout_schedule(&sc->sc_md6_tmo, tvtohz(&tv)); 1539 #endif /* INET6 */ 1540 break; 1541 } 1542 break; 1543 case MASTER: 1544 tv.tv_sec = sc->sc_advbase; 1545 tv.tv_usec = sc->sc_advskew * 1000000 / 256; 1546 callout_schedule(&sc->sc_ad_tmo, tvtohz(&tv)); 1547 break; 1548 } 1549 } 1550 1551 static void 1552 carp_multicast_cleanup(struct carp_softc *sc) 1553 { 1554 struct ip_moptions *imo = &sc->sc_imo; 1555 #ifdef INET6 1556 struct ip6_moptions *im6o = &sc->sc_im6o; 1557 #endif 1558 u_int16_t n = imo->imo_num_memberships; 1559 1560 /* Clean up our own multicast memberships */ 1561 while (n-- > 0) { 1562 if (imo->imo_membership[n] != NULL) { 1563 in_delmulti(imo->imo_membership[n]); 1564 imo->imo_membership[n] = NULL; 1565 } 1566 } 1567 imo->imo_num_memberships = 0; 1568 imo->imo_multicast_if_index = 0; 1569 1570 #ifdef INET6 1571 while (!LIST_EMPTY(&im6o->im6o_memberships)) { 1572 struct in6_multi_mship *imm = 1573 LIST_FIRST(&im6o->im6o_memberships); 1574 1575 LIST_REMOVE(imm, i6mm_chain); 1576 in6_leavegroup(imm); 1577 } 1578 im6o->im6o_multicast_if_index = 0; 1579 #endif 1580 1581 /* And any other multicast memberships */ 1582 carp_ether_purgemulti(sc); 1583 } 1584 1585 static int 1586 carp_set_ifp(struct carp_softc *sc, struct ifnet *ifp) 1587 { 1588 struct carp_if *cif, *ncif = NULL; 1589 struct carp_softc *vr, *after = NULL; 1590 int myself = 0, error = 0; 1591 int s; 1592 1593 if (ifp == sc->sc_carpdev) 1594 return (0); 1595 1596 if (ifp != NULL) { 1597 if ((ifp->if_flags & IFF_MULTICAST) == 0) 1598 return (EADDRNOTAVAIL); 1599 1600 if (ifp->if_type == IFT_CARP) 1601 return (EINVAL); 1602 1603 if (ifp->if_carp == NULL) { 1604 ncif = malloc(sizeof(*cif), M_IFADDR, M_NOWAIT); 1605 if (ncif == NULL) 1606 return (ENOBUFS); 1607 if ((error = ifpromisc(ifp, 1))) { 1608 free(ncif, M_IFADDR); 1609 return (error); 1610 } 1611 1612 ncif->vhif_ifp = ifp; 1613 TAILQ_INIT(&ncif->vhif_vrs); 1614 } else { 1615 cif = (struct carp_if *)ifp->if_carp; 1616 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) 1617 if (vr != sc && vr->sc_vhid == sc->sc_vhid) 1618 return (EINVAL); 1619 } 1620 1621 /* detach from old interface */ 1622 if (sc->sc_carpdev != NULL) 1623 carpdetach(sc); 1624 1625 /* join multicast groups */ 1626 if (sc->sc_naddrs < 0 && 1627 (error = carp_join_multicast(sc)) != 0) { 1628 if (ncif != NULL) 1629 free(ncif, M_IFADDR); 1630 return (error); 1631 } 1632 1633 #ifdef INET6 1634 if (sc->sc_naddrs6 < 0 && 1635 (error = carp_join_multicast6(sc)) != 0) { 1636 if (ncif != NULL) 1637 free(ncif, M_IFADDR); 1638 carp_multicast_cleanup(sc); 1639 return (error); 1640 } 1641 #endif 1642 1643 /* attach carp interface to physical interface */ 1644 if (ncif != NULL) 1645 ifp->if_carp = (void *)ncif; 1646 sc->sc_carpdev = ifp; 1647 sc->sc_if.if_capabilities = ifp->if_capabilities & 1648 (IFCAP_TSOv4 | IFCAP_TSOv6 | 1649 IFCAP_CSUM_IPv4_Tx|IFCAP_CSUM_IPv4_Rx| 1650 IFCAP_CSUM_TCPv4_Tx|IFCAP_CSUM_TCPv4_Rx| 1651 IFCAP_CSUM_UDPv4_Tx|IFCAP_CSUM_UDPv4_Rx| 1652 IFCAP_CSUM_TCPv6_Tx|IFCAP_CSUM_TCPv6_Rx| 1653 IFCAP_CSUM_UDPv6_Tx|IFCAP_CSUM_UDPv6_Rx); 1654 1655 cif = (struct carp_if *)ifp->if_carp; 1656 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) { 1657 if (vr == sc) 1658 myself = 1; 1659 if (vr->sc_vhid < sc->sc_vhid) 1660 after = vr; 1661 } 1662 1663 if (!myself) { 1664 /* We're trying to keep things in order */ 1665 if (after == NULL) { 1666 TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list); 1667 } else { 1668 TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, 1669 sc, sc_list); 1670 } 1671 cif->vhif_nvrs++; 1672 } 1673 if (sc->sc_naddrs || sc->sc_naddrs6) 1674 sc->sc_if.if_flags |= IFF_UP; 1675 carp_set_enaddr(sc); 1676 KERNEL_LOCK(1, NULL); 1677 s = splnet(); 1678 /* XXX linkstatehooks establish */ 1679 carp_carpdev_state(ifp); 1680 splx(s); 1681 KERNEL_UNLOCK_ONE(NULL); 1682 } else { 1683 carpdetach(sc); 1684 sc->sc_if.if_flags &= ~(IFF_UP|IFF_RUNNING); 1685 } 1686 return (0); 1687 } 1688 1689 static void 1690 carp_set_enaddr(struct carp_softc *sc) 1691 { 1692 uint8_t enaddr[ETHER_ADDR_LEN]; 1693 if (sc->sc_carpdev && sc->sc_carpdev->if_type == IFT_ISO88025) { 1694 enaddr[0] = 3; 1695 enaddr[1] = 0; 1696 enaddr[2] = 0x40 >> (sc->sc_vhid - 1); 1697 enaddr[3] = 0x40000 >> (sc->sc_vhid - 1); 1698 enaddr[4] = 0; 1699 enaddr[5] = 0; 1700 } else { 1701 enaddr[0] = 0; 1702 enaddr[1] = 0; 1703 enaddr[2] = 0x5e; 1704 enaddr[3] = 0; 1705 enaddr[4] = 1; 1706 enaddr[5] = sc->sc_vhid; 1707 } 1708 if_set_sadl(&sc->sc_if, enaddr, sizeof(enaddr), false); 1709 } 1710 1711 #if 0 1712 static void 1713 carp_addr_updated(void *v) 1714 { 1715 struct carp_softc *sc = (struct carp_softc *) v; 1716 struct ifaddr *ifa; 1717 int new_naddrs = 0, new_naddrs6 = 0; 1718 1719 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 1720 if (ifa->ifa_addr->sa_family == AF_INET) 1721 new_naddrs++; 1722 else if (ifa->ifa_addr->sa_family == AF_INET6) 1723 new_naddrs6++; 1724 } 1725 1726 /* Handle a callback after SIOCDIFADDR */ 1727 if (new_naddrs < sc->sc_naddrs || new_naddrs6 < sc->sc_naddrs6) { 1728 struct in_addr mc_addr; 1729 1730 sc->sc_naddrs = new_naddrs; 1731 sc->sc_naddrs6 = new_naddrs6; 1732 1733 /* Re-establish multicast membership removed by in_control */ 1734 mc_addr.s_addr = INADDR_CARP_GROUP; 1735 if (!in_multi_group(mc_addr, &sc->sc_if, 0)) { 1736 memset(&sc->sc_imo, 0, sizeof(sc->sc_imo)); 1737 1738 if (sc->sc_carpdev != NULL && sc->sc_naddrs > 0) 1739 carp_join_multicast(sc); 1740 } 1741 1742 if (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) { 1743 sc->sc_if.if_flags &= ~IFF_UP; 1744 carp_set_state(sc, INIT); 1745 } else 1746 carp_hmac_prepare(sc); 1747 } 1748 1749 carp_setrun(sc, 0); 1750 } 1751 #endif 1752 1753 static int 1754 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin) 1755 { 1756 struct ifnet *ifp = sc->sc_carpdev; 1757 struct in_ifaddr *ia, *ia_if; 1758 int error = 0; 1759 int s; 1760 1761 if (sin->sin_addr.s_addr == 0) { 1762 if (!(sc->sc_if.if_flags & IFF_UP)) 1763 carp_set_state(sc, INIT); 1764 if (sc->sc_naddrs) 1765 sc->sc_if.if_flags |= IFF_UP; 1766 carp_setrun(sc, 0); 1767 return (0); 1768 } 1769 1770 /* we have to do this by hand to ensure we don't match on ourselves */ 1771 ia_if = NULL; 1772 s = pserialize_read_enter(); 1773 IN_ADDRLIST_READER_FOREACH(ia) { 1774 /* and, yeah, we need a multicast-capable iface too */ 1775 if (ia->ia_ifp != &sc->sc_if && 1776 ia->ia_ifp->if_type != IFT_CARP && 1777 (ia->ia_ifp->if_flags & IFF_MULTICAST) && 1778 (sin->sin_addr.s_addr & ia->ia_subnetmask) == 1779 ia->ia_subnet) { 1780 if (!ia_if) 1781 ia_if = ia; 1782 } 1783 } 1784 1785 if (ia_if) { 1786 ia = ia_if; 1787 if (ifp) { 1788 if (ifp != ia->ia_ifp) 1789 return (EADDRNOTAVAIL); 1790 } else { 1791 /* FIXME NOMPSAFE */ 1792 ifp = ia->ia_ifp; 1793 } 1794 } 1795 pserialize_read_exit(s); 1796 1797 if ((error = carp_set_ifp(sc, ifp))) 1798 return (error); 1799 1800 if (sc->sc_carpdev == NULL) 1801 return (EADDRNOTAVAIL); 1802 1803 if (sc->sc_naddrs == 0 && (error = carp_join_multicast(sc)) != 0) 1804 return (error); 1805 1806 sc->sc_naddrs++; 1807 if (sc->sc_carpdev != NULL) 1808 sc->sc_if.if_flags |= IFF_UP; 1809 1810 carp_set_state(sc, INIT); 1811 carp_setrun(sc, 0); 1812 1813 /* 1814 * Hook if_addrhooks so that we get a callback after in_ifinit has run, 1815 * to correct any inappropriate routes that it inserted. 1816 */ 1817 if (sc->ah_cookie == 0) { 1818 /* XXX link address hook */ 1819 } 1820 1821 return (0); 1822 } 1823 1824 static int 1825 carp_join_multicast(struct carp_softc *sc) 1826 { 1827 struct ip_moptions *imo = &sc->sc_imo, tmpimo; 1828 struct in_addr addr; 1829 1830 memset(&tmpimo, 0, sizeof(tmpimo)); 1831 addr.s_addr = INADDR_CARP_GROUP; 1832 if ((tmpimo.imo_membership[0] = 1833 in_addmulti(&addr, &sc->sc_if)) == NULL) { 1834 return (ENOBUFS); 1835 } 1836 1837 imo->imo_membership[0] = tmpimo.imo_membership[0]; 1838 imo->imo_num_memberships = 1; 1839 imo->imo_multicast_if_index = sc->sc_if.if_index; 1840 imo->imo_multicast_ttl = CARP_DFLTTL; 1841 imo->imo_multicast_loop = 0; 1842 return (0); 1843 } 1844 1845 1846 #ifdef INET6 1847 static int 1848 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6) 1849 { 1850 struct ifnet *ifp = sc->sc_carpdev; 1851 struct in6_ifaddr *ia, *ia_if; 1852 int error = 0; 1853 int s; 1854 1855 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) { 1856 if (!(sc->sc_if.if_flags & IFF_UP)) 1857 carp_set_state(sc, INIT); 1858 if (sc->sc_naddrs6) 1859 sc->sc_if.if_flags |= IFF_UP; 1860 carp_setrun(sc, 0); 1861 return (0); 1862 } 1863 1864 /* we have to do this by hand to ensure we don't match on ourselves */ 1865 ia_if = NULL; 1866 s = pserialize_read_enter(); 1867 IN6_ADDRLIST_READER_FOREACH(ia) { 1868 int i; 1869 1870 for (i = 0; i < 4; i++) { 1871 if ((sin6->sin6_addr.s6_addr32[i] & 1872 ia->ia_prefixmask.sin6_addr.s6_addr32[i]) != 1873 (ia->ia_addr.sin6_addr.s6_addr32[i] & 1874 ia->ia_prefixmask.sin6_addr.s6_addr32[i])) 1875 break; 1876 } 1877 /* and, yeah, we need a multicast-capable iface too */ 1878 if (ia->ia_ifp != &sc->sc_if && 1879 ia->ia_ifp->if_type != IFT_CARP && 1880 (ia->ia_ifp->if_flags & IFF_MULTICAST) && 1881 (i == 4)) { 1882 if (!ia_if) 1883 ia_if = ia; 1884 } 1885 } 1886 pserialize_read_exit(s); 1887 1888 if (ia_if) { 1889 ia = ia_if; 1890 if (sc->sc_carpdev) { 1891 if (sc->sc_carpdev != ia->ia_ifp) 1892 return (EADDRNOTAVAIL); 1893 } else { 1894 ifp = ia->ia_ifp; 1895 } 1896 } 1897 1898 if ((error = carp_set_ifp(sc, ifp))) 1899 return (error); 1900 1901 if (sc->sc_carpdev == NULL) 1902 return (EADDRNOTAVAIL); 1903 1904 if (sc->sc_naddrs6 == 0 && (error = carp_join_multicast6(sc)) != 0) 1905 return (error); 1906 1907 sc->sc_naddrs6++; 1908 if (sc->sc_carpdev != NULL) 1909 sc->sc_if.if_flags |= IFF_UP; 1910 carp_set_state(sc, INIT); 1911 carp_setrun(sc, 0); 1912 1913 return (0); 1914 } 1915 1916 static int 1917 carp_join_multicast6(struct carp_softc *sc) 1918 { 1919 struct in6_multi_mship *imm, *imm2; 1920 struct ip6_moptions *im6o = &sc->sc_im6o; 1921 struct sockaddr_in6 addr6; 1922 int error; 1923 1924 /* Join IPv6 CARP multicast group */ 1925 memset(&addr6, 0, sizeof(addr6)); 1926 addr6.sin6_family = AF_INET6; 1927 addr6.sin6_len = sizeof(addr6); 1928 addr6.sin6_addr.s6_addr16[0] = htons(0xff02); 1929 addr6.sin6_addr.s6_addr16[1] = htons(sc->sc_if.if_index); 1930 addr6.sin6_addr.s6_addr8[15] = 0x12; 1931 if ((imm = in6_joingroup(&sc->sc_if, 1932 &addr6.sin6_addr, &error, 0)) == NULL) { 1933 return (error); 1934 } 1935 /* join solicited multicast address */ 1936 memset(&addr6.sin6_addr, 0, sizeof(addr6.sin6_addr)); 1937 addr6.sin6_addr.s6_addr16[0] = htons(0xff02); 1938 addr6.sin6_addr.s6_addr16[1] = htons(sc->sc_if.if_index); 1939 addr6.sin6_addr.s6_addr32[1] = 0; 1940 addr6.sin6_addr.s6_addr32[2] = htonl(1); 1941 addr6.sin6_addr.s6_addr32[3] = 0; 1942 addr6.sin6_addr.s6_addr8[12] = 0xff; 1943 if ((imm2 = in6_joingroup(&sc->sc_if, 1944 &addr6.sin6_addr, &error, 0)) == NULL) { 1945 in6_leavegroup(imm); 1946 return (error); 1947 } 1948 1949 /* apply v6 multicast membership */ 1950 im6o->im6o_multicast_if_index = sc->sc_if.if_index; 1951 if (imm) 1952 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, 1953 i6mm_chain); 1954 if (imm2) 1955 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm2, 1956 i6mm_chain); 1957 1958 return (0); 1959 } 1960 1961 #endif /* INET6 */ 1962 1963 static int 1964 carp_ioctl(struct ifnet *ifp, u_long cmd, void *data) 1965 { 1966 struct lwp *l = curlwp; /* XXX */ 1967 struct carp_softc *sc = ifp->if_softc, *vr; 1968 struct carpreq carpr; 1969 struct ifaddr *ifa; 1970 struct ifreq *ifr; 1971 struct ifnet *cdev = NULL; 1972 int error = 0; 1973 1974 ifa = (struct ifaddr *)data; 1975 ifr = (struct ifreq *)data; 1976 1977 switch (cmd) { 1978 case SIOCINITIFADDR: 1979 switch (ifa->ifa_addr->sa_family) { 1980 #ifdef INET 1981 case AF_INET: 1982 sc->sc_if.if_flags |= IFF_UP; 1983 memcpy(ifa->ifa_dstaddr, ifa->ifa_addr, 1984 sizeof(struct sockaddr)); 1985 error = carp_set_addr(sc, satosin(ifa->ifa_addr)); 1986 break; 1987 #endif /* INET */ 1988 #ifdef INET6 1989 case AF_INET6: 1990 sc->sc_if.if_flags|= IFF_UP; 1991 error = carp_set_addr6(sc, satosin6(ifa->ifa_addr)); 1992 break; 1993 #endif /* INET6 */ 1994 default: 1995 error = EAFNOSUPPORT; 1996 break; 1997 } 1998 break; 1999 2000 case SIOCSIFFLAGS: 2001 if ((error = ifioctl_common(ifp, cmd, data)) != 0) 2002 break; 2003 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) { 2004 callout_stop(&sc->sc_ad_tmo); 2005 callout_stop(&sc->sc_md_tmo); 2006 callout_stop(&sc->sc_md6_tmo); 2007 if (sc->sc_state == MASTER) { 2008 /* we need the interface up to bow out */ 2009 sc->sc_if.if_flags |= IFF_UP; 2010 sc->sc_bow_out = 1; 2011 carp_send_ad(sc); 2012 } 2013 sc->sc_if.if_flags &= ~IFF_UP; 2014 carp_set_state(sc, INIT); 2015 carp_setrun(sc, 0); 2016 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) { 2017 sc->sc_if.if_flags |= IFF_UP; 2018 carp_setrun(sc, 0); 2019 } 2020 break; 2021 2022 case SIOCSVH: 2023 if (l == NULL) 2024 break; 2025 if ((error = kauth_authorize_network(l->l_cred, 2026 KAUTH_NETWORK_INTERFACE, 2027 KAUTH_REQ_NETWORK_INTERFACE_SETPRIV, ifp, (void *)cmd, 2028 NULL)) != 0) 2029 break; 2030 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr))) 2031 break; 2032 error = 1; 2033 if (carpr.carpr_carpdev[0] != '\0' && 2034 (cdev = ifunit(carpr.carpr_carpdev)) == NULL) 2035 return (EINVAL); 2036 if ((error = carp_set_ifp(sc, cdev))) 2037 return (error); 2038 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) { 2039 switch (carpr.carpr_state) { 2040 case BACKUP: 2041 callout_stop(&sc->sc_ad_tmo); 2042 carp_set_state(sc, BACKUP); 2043 carp_setrun(sc, 0); 2044 carp_setroute(sc, RTM_DELETE); 2045 break; 2046 case MASTER: 2047 carp_master_down(sc); 2048 break; 2049 default: 2050 break; 2051 } 2052 } 2053 if (carpr.carpr_vhid > 0) { 2054 if (carpr.carpr_vhid > 255) { 2055 error = EINVAL; 2056 break; 2057 } 2058 if (sc->sc_carpdev) { 2059 struct carp_if *cif; 2060 cif = (struct carp_if *)sc->sc_carpdev->if_carp; 2061 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) 2062 if (vr != sc && 2063 vr->sc_vhid == carpr.carpr_vhid) 2064 return (EINVAL); 2065 } 2066 sc->sc_vhid = carpr.carpr_vhid; 2067 carp_set_enaddr(sc); 2068 carp_set_state(sc, INIT); 2069 error--; 2070 } 2071 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) { 2072 if (carpr.carpr_advskew > 254) { 2073 error = EINVAL; 2074 break; 2075 } 2076 if (carpr.carpr_advbase > 255) { 2077 error = EINVAL; 2078 break; 2079 } 2080 sc->sc_advbase = carpr.carpr_advbase; 2081 sc->sc_advskew = carpr.carpr_advskew; 2082 error--; 2083 } 2084 memcpy(sc->sc_key, carpr.carpr_key, sizeof(sc->sc_key)); 2085 if (error > 0) 2086 error = EINVAL; 2087 else { 2088 error = 0; 2089 carp_setrun(sc, 0); 2090 } 2091 break; 2092 2093 case SIOCGVH: 2094 memset(&carpr, 0, sizeof(carpr)); 2095 if (sc->sc_carpdev != NULL) 2096 strlcpy(carpr.carpr_carpdev, sc->sc_carpdev->if_xname, 2097 IFNAMSIZ); 2098 carpr.carpr_state = sc->sc_state; 2099 carpr.carpr_vhid = sc->sc_vhid; 2100 carpr.carpr_advbase = sc->sc_advbase; 2101 carpr.carpr_advskew = sc->sc_advskew; 2102 2103 if ((l != NULL) && (error = kauth_authorize_network(l->l_cred, 2104 KAUTH_NETWORK_INTERFACE, 2105 KAUTH_REQ_NETWORK_INTERFACE_SETPRIV, ifp, (void *)cmd, 2106 NULL)) == 0) 2107 memcpy(carpr.carpr_key, sc->sc_key, 2108 sizeof(carpr.carpr_key)); 2109 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr)); 2110 break; 2111 2112 case SIOCADDMULTI: 2113 error = carp_ether_addmulti(sc, ifr); 2114 break; 2115 2116 case SIOCDELMULTI: 2117 error = carp_ether_delmulti(sc, ifr); 2118 break; 2119 2120 case SIOCSIFCAP: 2121 if ((error = ifioctl_common(ifp, cmd, data)) == ENETRESET) 2122 error = 0; 2123 break; 2124 2125 default: 2126 error = ether_ioctl(ifp, cmd, data); 2127 } 2128 2129 carp_hmac_prepare(sc); 2130 return (error); 2131 } 2132 2133 2134 /* 2135 * Start output on carp interface. This function should never be called. 2136 */ 2137 static void 2138 carp_start(struct ifnet *ifp) 2139 { 2140 #ifdef DEBUG 2141 printf("%s: start called\n", ifp->if_xname); 2142 #endif 2143 } 2144 2145 int 2146 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa, 2147 const struct rtentry *rt) 2148 { 2149 struct carp_softc *sc = ((struct carp_softc *)ifp->if_softc); 2150 KASSERT(KERNEL_LOCKED_P()); 2151 2152 if (sc->sc_carpdev != NULL && sc->sc_state == MASTER) { 2153 return if_output_lock(sc->sc_carpdev, ifp, m, sa, rt); 2154 } else { 2155 m_freem(m); 2156 return (ENETUNREACH); 2157 } 2158 } 2159 2160 static void 2161 carp_set_state(struct carp_softc *sc, int state) 2162 { 2163 static const char *carp_states[] = { CARP_STATES }; 2164 if (sc->sc_state == state) 2165 return; 2166 2167 CARP_LOG(sc, ("state transition from: %s -> to: %s", carp_states[sc->sc_state], carp_states[state])); 2168 2169 sc->sc_state = state; 2170 switch (state) { 2171 case BACKUP: 2172 sc->sc_if.if_link_state = LINK_STATE_DOWN; 2173 break; 2174 case MASTER: 2175 sc->sc_if.if_link_state = LINK_STATE_UP; 2176 break; 2177 default: 2178 sc->sc_if.if_link_state = LINK_STATE_UNKNOWN; 2179 break; 2180 } 2181 rt_ifmsg(&sc->sc_if); 2182 } 2183 2184 void 2185 carp_carpdev_state(void *v) 2186 { 2187 struct carp_if *cif; 2188 struct carp_softc *sc; 2189 struct ifnet *ifp = v; 2190 2191 if (ifp->if_type == IFT_CARP) 2192 return; 2193 2194 cif = (struct carp_if *)ifp->if_carp; 2195 2196 TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) { 2197 int suppressed = sc->sc_suppress; 2198 2199 if (sc->sc_carpdev->if_link_state == LINK_STATE_DOWN || 2200 !(sc->sc_carpdev->if_flags & IFF_UP)) { 2201 sc->sc_if.if_flags &= ~IFF_RUNNING; 2202 callout_stop(&sc->sc_ad_tmo); 2203 callout_stop(&sc->sc_md_tmo); 2204 callout_stop(&sc->sc_md6_tmo); 2205 carp_set_state(sc, INIT); 2206 sc->sc_suppress = 1; 2207 carp_setrun(sc, 0); 2208 if (!suppressed) { 2209 carp_suppress_preempt++; 2210 if (carp_suppress_preempt == 1) 2211 carp_send_ad_all(); 2212 } 2213 } else { 2214 carp_set_state(sc, INIT); 2215 sc->sc_suppress = 0; 2216 carp_setrun(sc, 0); 2217 if (suppressed) 2218 carp_suppress_preempt--; 2219 } 2220 } 2221 } 2222 2223 static int 2224 carp_ether_addmulti(struct carp_softc *sc, struct ifreq *ifr) 2225 { 2226 const struct sockaddr *sa = ifreq_getaddr(SIOCADDMULTI, ifr); 2227 struct ifnet *ifp; 2228 struct carp_mc_entry *mc; 2229 u_int8_t addrlo[ETHER_ADDR_LEN], addrhi[ETHER_ADDR_LEN]; 2230 int error; 2231 2232 ifp = sc->sc_carpdev; 2233 if (ifp == NULL) 2234 return (EINVAL); 2235 2236 error = ether_addmulti(sa, &sc->sc_ac); 2237 if (error != ENETRESET) 2238 return (error); 2239 2240 /* 2241 * This is new multicast address. We have to tell parent 2242 * about it. Also, remember this multicast address so that 2243 * we can delete them on unconfigure. 2244 */ 2245 mc = malloc(sizeof(struct carp_mc_entry), M_DEVBUF, M_NOWAIT); 2246 if (mc == NULL) { 2247 error = ENOMEM; 2248 goto alloc_failed; 2249 } 2250 2251 /* 2252 * As ether_addmulti() returns ENETRESET, following two 2253 * statement shouldn't fail. 2254 */ 2255 (void)ether_multiaddr(sa, addrlo, addrhi); 2256 ETHER_LOOKUP_MULTI(addrlo, addrhi, &sc->sc_ac, mc->mc_enm); 2257 memcpy(&mc->mc_addr, sa, sa->sa_len); 2258 LIST_INSERT_HEAD(&sc->carp_mc_listhead, mc, mc_entries); 2259 2260 error = if_mcast_op(ifp, SIOCADDMULTI, sa); 2261 if (error != 0) 2262 goto ioctl_failed; 2263 2264 return (error); 2265 2266 ioctl_failed: 2267 LIST_REMOVE(mc, mc_entries); 2268 free(mc, M_DEVBUF); 2269 alloc_failed: 2270 (void)ether_delmulti(sa, &sc->sc_ac); 2271 2272 return (error); 2273 } 2274 2275 static int 2276 carp_ether_delmulti(struct carp_softc *sc, struct ifreq *ifr) 2277 { 2278 const struct sockaddr *sa = ifreq_getaddr(SIOCDELMULTI, ifr); 2279 struct ifnet *ifp; 2280 struct ether_multi *enm; 2281 struct carp_mc_entry *mc; 2282 u_int8_t addrlo[ETHER_ADDR_LEN], addrhi[ETHER_ADDR_LEN]; 2283 int error; 2284 2285 ifp = sc->sc_carpdev; 2286 if (ifp == NULL) 2287 return (EINVAL); 2288 2289 /* 2290 * Find a key to lookup carp_mc_entry. We have to do this 2291 * before calling ether_delmulti for obvious reason. 2292 */ 2293 if ((error = ether_multiaddr(sa, addrlo, addrhi)) != 0) 2294 return (error); 2295 ETHER_LOOKUP_MULTI(addrlo, addrhi, &sc->sc_ac, enm); 2296 if (enm == NULL) 2297 return (EINVAL); 2298 2299 LIST_FOREACH(mc, &sc->carp_mc_listhead, mc_entries) 2300 if (mc->mc_enm == enm) 2301 break; 2302 2303 /* We won't delete entries we didn't add */ 2304 if (mc == NULL) 2305 return (EINVAL); 2306 2307 error = ether_delmulti(sa, &sc->sc_ac); 2308 if (error != ENETRESET) 2309 return (error); 2310 2311 /* We no longer use this multicast address. Tell parent so. */ 2312 error = if_mcast_op(ifp, SIOCDELMULTI, sa); 2313 if (error == 0) { 2314 /* And forget about this address. */ 2315 LIST_REMOVE(mc, mc_entries); 2316 free(mc, M_DEVBUF); 2317 } else 2318 (void)ether_addmulti(sa, &sc->sc_ac); 2319 return (error); 2320 } 2321 2322 /* 2323 * Delete any multicast address we have asked to add from parent 2324 * interface. Called when the carp is being unconfigured. 2325 */ 2326 static void 2327 carp_ether_purgemulti(struct carp_softc *sc) 2328 { 2329 struct ifnet *ifp = sc->sc_carpdev; /* Parent. */ 2330 struct carp_mc_entry *mc; 2331 2332 if (ifp == NULL) 2333 return; 2334 2335 while ((mc = LIST_FIRST(&sc->carp_mc_listhead)) != NULL) { 2336 (void)if_mcast_op(ifp, SIOCDELMULTI, sstosa(&mc->mc_addr)); 2337 LIST_REMOVE(mc, mc_entries); 2338 free(mc, M_DEVBUF); 2339 } 2340 } 2341 2342 static int 2343 sysctl_net_inet_carp_stats(SYSCTLFN_ARGS) 2344 { 2345 2346 return (NETSTAT_SYSCTL(carpstat_percpu, CARP_NSTATS)); 2347 } 2348 2349 void 2350 carp_init(void) 2351 { 2352 2353 sysctl_net_inet_carp_setup(NULL); 2354 #ifdef MBUFTRACE 2355 MOWNER_ATTACH(&carp_proto_mowner_rx); 2356 MOWNER_ATTACH(&carp_proto_mowner_tx); 2357 MOWNER_ATTACH(&carp_proto6_mowner_rx); 2358 MOWNER_ATTACH(&carp_proto6_mowner_tx); 2359 #endif 2360 } 2361 2362 static void 2363 sysctl_net_inet_carp_setup(struct sysctllog **clog) 2364 { 2365 2366 sysctl_createv(clog, 0, NULL, NULL, 2367 CTLFLAG_PERMANENT, 2368 CTLTYPE_NODE, "inet", NULL, 2369 NULL, 0, NULL, 0, 2370 CTL_NET, PF_INET, CTL_EOL); 2371 sysctl_createv(clog, 0, NULL, NULL, 2372 CTLFLAG_PERMANENT, 2373 CTLTYPE_NODE, "carp", 2374 SYSCTL_DESCR("CARP related settings"), 2375 NULL, 0, NULL, 0, 2376 CTL_NET, PF_INET, IPPROTO_CARP, CTL_EOL); 2377 2378 sysctl_createv(clog, 0, NULL, NULL, 2379 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2380 CTLTYPE_INT, "preempt", 2381 SYSCTL_DESCR("Enable CARP Preempt"), 2382 NULL, 0, &carp_opts[CARPCTL_PREEMPT], 0, 2383 CTL_NET, PF_INET, IPPROTO_CARP, 2384 CTL_CREATE, CTL_EOL); 2385 sysctl_createv(clog, 0, NULL, NULL, 2386 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2387 CTLTYPE_INT, "arpbalance", 2388 SYSCTL_DESCR("Enable ARP balancing"), 2389 NULL, 0, &carp_opts[CARPCTL_ARPBALANCE], 0, 2390 CTL_NET, PF_INET, IPPROTO_CARP, 2391 CTL_CREATE, CTL_EOL); 2392 sysctl_createv(clog, 0, NULL, NULL, 2393 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2394 CTLTYPE_INT, "allow", 2395 SYSCTL_DESCR("Enable CARP"), 2396 NULL, 0, &carp_opts[CARPCTL_ALLOW], 0, 2397 CTL_NET, PF_INET, IPPROTO_CARP, 2398 CTL_CREATE, CTL_EOL); 2399 sysctl_createv(clog, 0, NULL, NULL, 2400 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2401 CTLTYPE_INT, "log", 2402 SYSCTL_DESCR("CARP logging"), 2403 NULL, 0, &carp_opts[CARPCTL_LOG], 0, 2404 CTL_NET, PF_INET, IPPROTO_CARP, 2405 CTL_CREATE, CTL_EOL); 2406 sysctl_createv(clog, 0, NULL, NULL, 2407 CTLFLAG_PERMANENT, 2408 CTLTYPE_STRUCT, "stats", 2409 SYSCTL_DESCR("CARP statistics"), 2410 sysctl_net_inet_carp_stats, 0, NULL, 0, 2411 CTL_NET, PF_INET, IPPROTO_CARP, CARPCTL_STATS, 2412 CTL_EOL); 2413 } 2414