1 /* $NetBSD: ip_carp.c,v 1.84 2017/02/02 02:52:10 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 1.84 2017/02/02 02:52:10 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 #include <netinet/wqinput.h> 74 75 #if NFDDI > 0 76 #include <net/if_fddi.h> 77 #endif 78 #if NTOKEN > 0 79 #include <net/if_token.h> 80 #endif 81 82 #ifdef INET 83 #include <netinet/in.h> 84 #include <netinet/in_systm.h> 85 #include <netinet/in_var.h> 86 #include <netinet/ip.h> 87 #include <netinet/ip_var.h> 88 89 #include <net/if_dl.h> 90 #endif 91 92 #ifdef INET6 93 #include <netinet/icmp6.h> 94 #include <netinet/ip6.h> 95 #include <netinet6/ip6_var.h> 96 #include <netinet6/nd6.h> 97 #include <netinet6/scope6_var.h> 98 #include <netinet6/in6_var.h> 99 #endif 100 101 #include <net/bpf.h> 102 103 #include <sys/sha1.h> 104 105 #include <netinet/ip_carp.h> 106 107 #include "ioconf.h" 108 109 struct carp_mc_entry { 110 LIST_ENTRY(carp_mc_entry) mc_entries; 111 union { 112 struct ether_multi *mcu_enm; 113 } mc_u; 114 struct sockaddr_storage mc_addr; 115 }; 116 #define mc_enm mc_u.mcu_enm 117 118 struct carp_softc { 119 struct ethercom sc_ac; 120 #define sc_if sc_ac.ec_if 121 #define sc_carpdev sc_ac.ec_if.if_carpdev 122 int ah_cookie; 123 int lh_cookie; 124 struct ip_moptions sc_imo; 125 #ifdef INET6 126 struct ip6_moptions sc_im6o; 127 #endif /* INET6 */ 128 TAILQ_ENTRY(carp_softc) sc_list; 129 130 enum { INIT = 0, BACKUP, MASTER } sc_state; 131 132 int sc_suppress; 133 int sc_bow_out; 134 135 int sc_sendad_errors; 136 #define CARP_SENDAD_MAX_ERRORS 3 137 int sc_sendad_success; 138 #define CARP_SENDAD_MIN_SUCCESS 3 139 140 int sc_vhid; 141 int sc_advskew; 142 int sc_naddrs; 143 int sc_naddrs6; 144 int sc_advbase; /* seconds */ 145 int sc_init_counter; 146 u_int64_t sc_counter; 147 148 /* authentication */ 149 #define CARP_HMAC_PAD 64 150 unsigned char sc_key[CARP_KEY_LEN]; 151 unsigned char sc_pad[CARP_HMAC_PAD]; 152 SHA1_CTX sc_sha1; 153 u_int32_t sc_hashkey[2]; 154 155 struct callout sc_ad_tmo; /* advertisement timeout */ 156 struct callout sc_md_tmo; /* master down timeout */ 157 struct callout sc_md6_tmo; /* master down timeout */ 158 159 LIST_HEAD(__carp_mchead, carp_mc_entry) carp_mc_listhead; 160 }; 161 162 int carp_suppress_preempt = 0; 163 static int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 0, 0 }; /* XXX for now */ 164 165 static percpu_t *carpstat_percpu; 166 167 #define CARP_STATINC(x) _NET_STATINC(carpstat_percpu, x) 168 169 #ifdef MBUFTRACE 170 static struct mowner carp_proto_mowner_rx = MOWNER_INIT("carp", "rx"); 171 static struct mowner carp_proto_mowner_tx = MOWNER_INIT("carp", "tx"); 172 static struct mowner carp_proto6_mowner_rx = MOWNER_INIT("carp6", "rx"); 173 static struct mowner carp_proto6_mowner_tx = MOWNER_INIT("carp6", "tx"); 174 #endif 175 176 struct carp_if { 177 TAILQ_HEAD(, carp_softc) vhif_vrs; 178 int vhif_nvrs; 179 180 struct ifnet *vhif_ifp; 181 }; 182 183 #define CARP_LOG(sc, s) \ 184 if (carp_opts[CARPCTL_LOG]) { \ 185 if (sc) \ 186 log(LOG_INFO, "%s: ", \ 187 (sc)->sc_if.if_xname); \ 188 else \ 189 log(LOG_INFO, "carp: "); \ 190 addlog s; \ 191 addlog("\n"); \ 192 } 193 194 static void carp_hmac_prepare(struct carp_softc *); 195 static void carp_hmac_generate(struct carp_softc *, u_int32_t *, 196 unsigned char *); 197 static int carp_hmac_verify(struct carp_softc *, u_int32_t *, 198 unsigned char *); 199 static void carp_setroute(struct carp_softc *, int); 200 static void carp_proto_input_c(struct mbuf *, struct carp_header *, 201 sa_family_t); 202 static void carpdetach(struct carp_softc *); 203 static int carp_prepare_ad(struct mbuf *, struct carp_softc *, 204 struct carp_header *); 205 static void carp_send_ad_all(void); 206 static void carp_send_ad(void *); 207 static void carp_send_arp(struct carp_softc *); 208 static void carp_master_down(void *); 209 static int carp_ioctl(struct ifnet *, u_long, void *); 210 static void carp_start(struct ifnet *); 211 static void carp_setrun(struct carp_softc *, sa_family_t); 212 static void carp_set_state(struct carp_softc *, int); 213 static int carp_addrcount(struct carp_if *, struct in_ifaddr *, int); 214 enum { CARP_COUNT_MASTER, CARP_COUNT_RUNNING }; 215 216 static void carp_multicast_cleanup(struct carp_softc *); 217 static int carp_set_ifp(struct carp_softc *, struct ifnet *); 218 static void carp_set_enaddr(struct carp_softc *); 219 #if 0 220 static void carp_addr_updated(void *); 221 #endif 222 static u_int32_t carp_hash(struct carp_softc *, u_char *); 223 static int carp_set_addr(struct carp_softc *, struct sockaddr_in *); 224 static int carp_join_multicast(struct carp_softc *); 225 #ifdef INET6 226 static void carp_send_na(struct carp_softc *); 227 static int carp_set_addr6(struct carp_softc *, struct sockaddr_in6 *); 228 static int carp_join_multicast6(struct carp_softc *); 229 #endif 230 static int carp_clone_create(struct if_clone *, int); 231 static int carp_clone_destroy(struct ifnet *); 232 static int carp_ether_addmulti(struct carp_softc *, struct ifreq *); 233 static int carp_ether_delmulti(struct carp_softc *, struct ifreq *); 234 static void carp_ether_purgemulti(struct carp_softc *); 235 236 static void sysctl_net_inet_carp_setup(struct sysctllog **); 237 238 /* workqueue-based pr_input */ 239 static struct wqinput *carp_wqinput; 240 static void _carp_proto_input(struct mbuf *, int, int); 241 #ifdef INET6 242 static struct wqinput *carp6_wqinput; 243 static void _carp6_proto_input(struct mbuf *, int, int); 244 #endif 245 246 struct if_clone carp_cloner = 247 IF_CLONE_INITIALIZER("carp", carp_clone_create, carp_clone_destroy); 248 249 static __inline u_int16_t 250 carp_cksum(struct mbuf *m, int len) 251 { 252 return (in_cksum(m, len)); 253 } 254 255 static void 256 carp_hmac_prepare(struct carp_softc *sc) 257 { 258 u_int8_t carp_version = CARP_VERSION, type = CARP_ADVERTISEMENT; 259 u_int8_t vhid = sc->sc_vhid & 0xff; 260 SHA1_CTX sha1ctx; 261 u_int32_t kmd[5]; 262 struct ifaddr *ifa; 263 int i, found; 264 struct in_addr last, cur, in; 265 #ifdef INET6 266 struct in6_addr last6, cur6, in6; 267 #endif /* INET6 */ 268 269 /* compute ipad from key */ 270 memset(sc->sc_pad, 0, sizeof(sc->sc_pad)); 271 memcpy(sc->sc_pad, sc->sc_key, sizeof(sc->sc_key)); 272 for (i = 0; i < sizeof(sc->sc_pad); i++) 273 sc->sc_pad[i] ^= 0x36; 274 275 /* precompute first part of inner hash */ 276 SHA1Init(&sc->sc_sha1); 277 SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad)); 278 SHA1Update(&sc->sc_sha1, (void *)&carp_version, sizeof(carp_version)); 279 SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type)); 280 281 /* generate a key for the arpbalance hash, before the vhid is hashed */ 282 memcpy(&sha1ctx, &sc->sc_sha1, sizeof(sha1ctx)); 283 SHA1Final((unsigned char *)kmd, &sha1ctx); 284 sc->sc_hashkey[0] = kmd[0] ^ kmd[1]; 285 sc->sc_hashkey[1] = kmd[2] ^ kmd[3]; 286 287 /* the rest of the precomputation */ 288 SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid)); 289 290 /* Hash the addresses from smallest to largest, not interface order */ 291 #ifdef INET 292 cur.s_addr = 0; 293 do { 294 found = 0; 295 last = cur; 296 cur.s_addr = 0xffffffff; 297 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 298 in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr; 299 if (ifa->ifa_addr->sa_family == AF_INET && 300 ntohl(in.s_addr) > ntohl(last.s_addr) && 301 ntohl(in.s_addr) < ntohl(cur.s_addr)) { 302 cur.s_addr = in.s_addr; 303 found++; 304 } 305 } 306 if (found) 307 SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur)); 308 } while (found); 309 #endif /* INET */ 310 311 #ifdef INET6 312 memset(&cur6, 0x00, sizeof(cur6)); 313 do { 314 found = 0; 315 last6 = cur6; 316 memset(&cur6, 0xff, sizeof(cur6)); 317 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 318 in6 = ifatoia6(ifa)->ia_addr.sin6_addr; 319 if (IN6_IS_ADDR_LINKLOCAL(&in6)) 320 in6.s6_addr16[1] = 0; 321 if (ifa->ifa_addr->sa_family == AF_INET6 && 322 memcmp(&in6, &last6, sizeof(in6)) > 0 && 323 memcmp(&in6, &cur6, sizeof(in6)) < 0) { 324 cur6 = in6; 325 found++; 326 } 327 } 328 if (found) 329 SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6)); 330 } while (found); 331 #endif /* INET6 */ 332 333 /* convert ipad to opad */ 334 for (i = 0; i < sizeof(sc->sc_pad); i++) 335 sc->sc_pad[i] ^= 0x36 ^ 0x5c; 336 } 337 338 static void 339 carp_hmac_generate(struct carp_softc *sc, u_int32_t counter[2], 340 unsigned char md[20]) 341 { 342 SHA1_CTX sha1ctx; 343 344 /* fetch first half of inner hash */ 345 memcpy(&sha1ctx, &sc->sc_sha1, sizeof(sha1ctx)); 346 347 SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter)); 348 SHA1Final(md, &sha1ctx); 349 350 /* outer hash */ 351 SHA1Init(&sha1ctx); 352 SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad)); 353 SHA1Update(&sha1ctx, md, 20); 354 SHA1Final(md, &sha1ctx); 355 } 356 357 static int 358 carp_hmac_verify(struct carp_softc *sc, u_int32_t counter[2], 359 unsigned char md[20]) 360 { 361 unsigned char md2[20]; 362 363 carp_hmac_generate(sc, counter, md2); 364 365 return (memcmp(md, md2, sizeof(md2))); 366 } 367 368 static void 369 carp_setroute(struct carp_softc *sc, int cmd) 370 { 371 struct ifaddr *ifa; 372 int s; 373 374 KERNEL_LOCK(1, NULL); 375 s = splsoftnet(); 376 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 377 switch (ifa->ifa_addr->sa_family) { 378 case AF_INET: { 379 int count = 0; 380 struct rtentry *rt; 381 int hr_otherif, nr_ourif; 382 383 /* 384 * Avoid screwing with the routes if there are other 385 * carp interfaces which are master and have the same 386 * address. 387 */ 388 if (sc->sc_carpdev != NULL && 389 sc->sc_carpdev->if_carp != NULL) { 390 count = carp_addrcount( 391 (struct carp_if *)sc->sc_carpdev->if_carp, 392 ifatoia(ifa), CARP_COUNT_MASTER); 393 if ((cmd == RTM_ADD && count != 1) || 394 (cmd == RTM_DELETE && count != 0)) 395 continue; 396 } 397 398 /* Remove the existing host route, if any */ 399 rtrequest(RTM_DELETE, ifa->ifa_addr, 400 ifa->ifa_addr, ifa->ifa_netmask, 401 RTF_HOST, NULL); 402 403 rt = NULL; 404 (void)rtrequest(RTM_GET, ifa->ifa_addr, ifa->ifa_addr, 405 ifa->ifa_netmask, RTF_HOST, &rt); 406 hr_otherif = (rt && rt->rt_ifp != &sc->sc_if && 407 (rt->rt_flags & RTF_CONNECTED)); 408 if (rt != NULL) { 409 rt_unref(rt); 410 rt = NULL; 411 } 412 413 /* Check for a network route on our interface */ 414 415 rt = NULL; 416 (void)rtrequest(RTM_GET, ifa->ifa_addr, ifa->ifa_addr, 417 ifa->ifa_netmask, 0, &rt); 418 nr_ourif = (rt && rt->rt_ifp == &sc->sc_if); 419 420 switch (cmd) { 421 case RTM_ADD: 422 if (hr_otherif) { 423 ifa->ifa_rtrequest = NULL; 424 ifa->ifa_flags &= ~RTF_CONNECTED; 425 426 rtrequest(RTM_ADD, ifa->ifa_addr, 427 ifa->ifa_addr, ifa->ifa_netmask, 428 RTF_UP | RTF_HOST, NULL); 429 } 430 if (!hr_otherif || nr_ourif || !rt) { 431 if (nr_ourif && 432 (rt->rt_flags & RTF_CONNECTED) == 0) 433 rtrequest(RTM_DELETE, 434 ifa->ifa_addr, 435 ifa->ifa_addr, 436 ifa->ifa_netmask, 0, NULL); 437 438 ifa->ifa_rtrequest = arp_rtrequest; 439 ifa->ifa_flags |= RTF_CONNECTED; 440 441 if (rtrequest(RTM_ADD, ifa->ifa_addr, 442 ifa->ifa_addr, ifa->ifa_netmask, 0, 443 NULL) == 0) 444 ifa->ifa_flags |= IFA_ROUTE; 445 } 446 break; 447 case RTM_DELETE: 448 break; 449 default: 450 break; 451 } 452 if (rt != NULL) { 453 rt_unref(rt); 454 rt = NULL; 455 } 456 break; 457 } 458 459 #ifdef INET6 460 case AF_INET6: 461 if (cmd == RTM_ADD) 462 in6_ifaddlocal(ifa); 463 else 464 in6_ifremlocal(ifa); 465 break; 466 #endif /* INET6 */ 467 default: 468 break; 469 } 470 } 471 splx(s); 472 KERNEL_UNLOCK_ONE(NULL); 473 } 474 475 /* 476 * process input packet. 477 * we have rearranged checks order compared to the rfc, 478 * but it seems more efficient this way or not possible otherwise. 479 */ 480 static void 481 _carp_proto_input(struct mbuf *m, int hlen, int proto) 482 { 483 struct ip *ip = mtod(m, struct ip *); 484 struct carp_softc *sc = NULL; 485 struct carp_header *ch; 486 int iplen, len; 487 struct ifnet *rcvif; 488 489 CARP_STATINC(CARP_STAT_IPACKETS); 490 MCLAIM(m, &carp_proto_mowner_rx); 491 492 if (!carp_opts[CARPCTL_ALLOW]) { 493 m_freem(m); 494 return; 495 } 496 497 rcvif = m_get_rcvif_NOMPSAFE(m); 498 /* check if received on a valid carp interface */ 499 if (rcvif->if_type != IFT_CARP) { 500 CARP_STATINC(CARP_STAT_BADIF); 501 CARP_LOG(sc, ("packet received on non-carp interface: %s", 502 rcvif->if_xname)); 503 m_freem(m); 504 return; 505 } 506 507 /* verify that the IP TTL is 255. */ 508 if (ip->ip_ttl != CARP_DFLTTL) { 509 CARP_STATINC(CARP_STAT_BADTTL); 510 CARP_LOG(sc, ("received ttl %d != %d on %s", ip->ip_ttl, 511 CARP_DFLTTL, rcvif->if_xname)); 512 m_freem(m); 513 return; 514 } 515 516 /* 517 * verify that the received packet length is 518 * equal to the CARP header 519 */ 520 iplen = ip->ip_hl << 2; 521 len = iplen + sizeof(*ch); 522 if (len > m->m_pkthdr.len) { 523 CARP_STATINC(CARP_STAT_BADLEN); 524 CARP_LOG(sc, ("packet too short %d on %s", m->m_pkthdr.len, 525 rcvif->if_xname)); 526 m_freem(m); 527 return; 528 } 529 530 if ((m = m_pullup(m, len)) == NULL) { 531 CARP_STATINC(CARP_STAT_HDROPS); 532 return; 533 } 534 ip = mtod(m, struct ip *); 535 ch = (struct carp_header *)((char *)ip + iplen); 536 /* verify the CARP checksum */ 537 m->m_data += iplen; 538 if (carp_cksum(m, len - iplen)) { 539 CARP_STATINC(CARP_STAT_BADSUM); 540 CARP_LOG(sc, ("checksum failed on %s", 541 rcvif->if_xname)); 542 m_freem(m); 543 return; 544 } 545 m->m_data -= iplen; 546 547 carp_proto_input_c(m, ch, AF_INET); 548 } 549 550 void 551 carp_proto_input(struct mbuf *m, ...) 552 { 553 554 wqinput_input(carp_wqinput, m, 0, 0); 555 } 556 557 #ifdef INET6 558 static void 559 _carp6_proto_input(struct mbuf *m, int off, int proto) 560 { 561 struct carp_softc *sc = NULL; 562 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 563 struct carp_header *ch; 564 u_int len; 565 struct ifnet *rcvif; 566 567 CARP_STATINC(CARP_STAT_IPACKETS6); 568 MCLAIM(m, &carp_proto6_mowner_rx); 569 570 if (!carp_opts[CARPCTL_ALLOW]) { 571 m_freem(m); 572 return; 573 } 574 575 rcvif = m_get_rcvif_NOMPSAFE(m); 576 577 /* check if received on a valid carp interface */ 578 if (rcvif->if_type != IFT_CARP) { 579 CARP_STATINC(CARP_STAT_BADIF); 580 CARP_LOG(sc, ("packet received on non-carp interface: %s", 581 rcvif->if_xname)); 582 m_freem(m); 583 return; 584 } 585 586 /* verify that the IP TTL is 255 */ 587 if (ip6->ip6_hlim != CARP_DFLTTL) { 588 CARP_STATINC(CARP_STAT_BADTTL); 589 CARP_LOG(sc, ("received ttl %d != %d on %s", ip6->ip6_hlim, 590 CARP_DFLTTL, rcvif->if_xname)); 591 m_freem(m); 592 return; 593 } 594 595 /* verify that we have a complete carp packet */ 596 len = m->m_len; 597 IP6_EXTHDR_GET(ch, struct carp_header *, m, off, sizeof(*ch)); 598 if (ch == NULL) { 599 CARP_STATINC(CARP_STAT_BADLEN); 600 CARP_LOG(sc, ("packet size %u too small", len)); 601 return; 602 } 603 604 605 /* verify the CARP checksum */ 606 m->m_data += off; 607 if (carp_cksum(m, sizeof(*ch))) { 608 CARP_STATINC(CARP_STAT_BADSUM); 609 CARP_LOG(sc, ("checksum failed, on %s", rcvif->if_xname)); 610 m_freem(m); 611 return; 612 } 613 m->m_data -= off; 614 615 carp_proto_input_c(m, ch, AF_INET6); 616 return; 617 } 618 619 int 620 carp6_proto_input(struct mbuf **mp, int *offp, int proto) 621 { 622 623 wqinput_input(carp6_wqinput, *mp, *offp, proto); 624 625 return IPPROTO_DONE; 626 } 627 #endif /* INET6 */ 628 629 static void 630 carp_proto_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af) 631 { 632 struct carp_softc *sc; 633 u_int64_t tmp_counter; 634 struct timeval sc_tv, ch_tv; 635 636 TAILQ_FOREACH(sc, &((struct carp_if *) 637 m_get_rcvif_NOMPSAFE(m)->if_carpdev->if_carp)->vhif_vrs, sc_list) 638 if (sc->sc_vhid == ch->carp_vhid) 639 break; 640 641 if (!sc || (sc->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) != 642 (IFF_UP|IFF_RUNNING)) { 643 CARP_STATINC(CARP_STAT_BADVHID); 644 m_freem(m); 645 return; 646 } 647 648 /* 649 * Check if our own advertisement was duplicated 650 * from a non simplex interface. 651 * XXX If there is no address on our physical interface 652 * there is no way to distinguish our ads from the ones 653 * another carp host might have sent us. 654 */ 655 if ((sc->sc_carpdev->if_flags & IFF_SIMPLEX) == 0) { 656 struct sockaddr sa; 657 struct ifaddr *ifa; 658 int s; 659 660 memset(&sa, 0, sizeof(sa)); 661 sa.sa_family = af; 662 s = pserialize_read_enter(); 663 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev); 664 665 if (ifa && af == AF_INET) { 666 struct ip *ip = mtod(m, struct ip *); 667 if (ip->ip_src.s_addr == 668 ifatoia(ifa)->ia_addr.sin_addr.s_addr) { 669 pserialize_read_exit(s); 670 m_freem(m); 671 return; 672 } 673 } 674 #ifdef INET6 675 if (ifa && af == AF_INET6) { 676 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 677 struct in6_addr in6_src, in6_found; 678 679 in6_src = ip6->ip6_src; 680 in6_found = ifatoia6(ifa)->ia_addr.sin6_addr; 681 if (IN6_IS_ADDR_LINKLOCAL(&in6_src)) 682 in6_src.s6_addr16[1] = 0; 683 if (IN6_IS_ADDR_LINKLOCAL(&in6_found)) 684 in6_found.s6_addr16[1] = 0; 685 if (IN6_ARE_ADDR_EQUAL(&in6_src, &in6_found)) { 686 pserialize_read_exit(s); 687 m_freem(m); 688 return; 689 } 690 } 691 #endif /* INET6 */ 692 pserialize_read_exit(s); 693 } 694 695 nanotime(&sc->sc_if.if_lastchange); 696 sc->sc_if.if_ipackets++; 697 sc->sc_if.if_ibytes += m->m_pkthdr.len; 698 699 /* verify the CARP version. */ 700 if (ch->carp_version != CARP_VERSION) { 701 CARP_STATINC(CARP_STAT_BADVER); 702 sc->sc_if.if_ierrors++; 703 CARP_LOG(sc, ("invalid version %d != %d", 704 ch->carp_version, CARP_VERSION)); 705 m_freem(m); 706 return; 707 } 708 709 /* verify the hash */ 710 if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) { 711 struct ip *ip; 712 struct ip6_hdr *ip6; 713 char ip6buf[INET6_ADDRSTRLEN]; 714 char ipbuf[INET_ADDRSTRLEN]; 715 716 CARP_STATINC(CARP_STAT_BADAUTH); 717 sc->sc_if.if_ierrors++; 718 719 switch(af) { 720 case AF_INET: 721 ip = mtod(m, struct ip *); 722 CARP_LOG(sc, ("incorrect hash from %s", 723 in_fmtaddr(ipbuf, ip->ip_src))); 724 break; 725 726 case AF_INET6: 727 ip6 = mtod(m, struct ip6_hdr *); 728 CARP_LOG(sc, ("incorrect hash from %s", 729 IN6_PRINT(ip6buf, &ip6->ip6_src))); 730 break; 731 732 default: CARP_LOG(sc, ("incorrect hash")); 733 break; 734 } 735 m_freem(m); 736 return; 737 } 738 739 tmp_counter = ntohl(ch->carp_counter[0]); 740 tmp_counter = tmp_counter<<32; 741 tmp_counter += ntohl(ch->carp_counter[1]); 742 743 /* XXX Replay protection goes here */ 744 745 sc->sc_init_counter = 0; 746 sc->sc_counter = tmp_counter; 747 748 749 sc_tv.tv_sec = sc->sc_advbase; 750 if (carp_suppress_preempt && sc->sc_advskew < 240) 751 sc_tv.tv_usec = 240 * 1000000 / 256; 752 else 753 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256; 754 ch_tv.tv_sec = ch->carp_advbase; 755 ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256; 756 757 switch (sc->sc_state) { 758 case INIT: 759 break; 760 case MASTER: 761 /* 762 * If we receive an advertisement from a backup who's going to 763 * be more frequent than us, go into BACKUP state. 764 */ 765 if (timercmp(&sc_tv, &ch_tv, >) || 766 timercmp(&sc_tv, &ch_tv, ==)) { 767 callout_stop(&sc->sc_ad_tmo); 768 CARP_LOG(sc, ("MASTER -> BACKUP (more frequent advertisement received)")); 769 carp_set_state(sc, BACKUP); 770 carp_setrun(sc, 0); 771 carp_setroute(sc, RTM_DELETE); 772 } 773 break; 774 case BACKUP: 775 /* 776 * If we're pre-empting masters who advertise slower than us, 777 * and this one claims to be slower, treat him as down. 778 */ 779 if (carp_opts[CARPCTL_PREEMPT] && timercmp(&sc_tv, &ch_tv, <)) { 780 CARP_LOG(sc, ("BACKUP -> MASTER (preempting a slower master)")); 781 carp_master_down(sc); 782 break; 783 } 784 785 /* 786 * If the master is going to advertise at such a low frequency 787 * that he's guaranteed to time out, we'd might as well just 788 * treat him as timed out now. 789 */ 790 sc_tv.tv_sec = sc->sc_advbase * 3; 791 if (timercmp(&sc_tv, &ch_tv, <)) { 792 CARP_LOG(sc, ("BACKUP -> MASTER (master timed out)")); 793 carp_master_down(sc); 794 break; 795 } 796 797 /* 798 * Otherwise, we reset the counter and wait for the next 799 * advertisement. 800 */ 801 carp_setrun(sc, af); 802 break; 803 } 804 805 m_freem(m); 806 return; 807 } 808 809 /* 810 * Interface side of the CARP implementation. 811 */ 812 813 /* ARGSUSED */ 814 void 815 carpattach(int n) 816 { 817 if_clone_attach(&carp_cloner); 818 819 carpstat_percpu = percpu_alloc(sizeof(uint64_t) * CARP_NSTATS); 820 } 821 822 static int 823 carp_clone_create(struct if_clone *ifc, int unit) 824 { 825 extern int ifqmaxlen; 826 struct carp_softc *sc; 827 struct ifnet *ifp; 828 829 sc = malloc(sizeof(*sc), M_DEVBUF, M_NOWAIT|M_ZERO); 830 if (!sc) 831 return (ENOMEM); 832 833 sc->sc_suppress = 0; 834 sc->sc_advbase = CARP_DFLTINTV; 835 sc->sc_vhid = -1; /* required setting */ 836 sc->sc_advskew = 0; 837 sc->sc_init_counter = 1; 838 sc->sc_naddrs = sc->sc_naddrs6 = 0; 839 #ifdef INET6 840 sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL; 841 #endif /* INET6 */ 842 843 callout_init(&sc->sc_ad_tmo, 0); 844 callout_init(&sc->sc_md_tmo, 0); 845 callout_init(&sc->sc_md6_tmo, 0); 846 847 callout_setfunc(&sc->sc_ad_tmo, carp_send_ad, sc); 848 callout_setfunc(&sc->sc_md_tmo, carp_master_down, sc); 849 callout_setfunc(&sc->sc_md6_tmo, carp_master_down, sc); 850 851 LIST_INIT(&sc->carp_mc_listhead); 852 ifp = &sc->sc_if; 853 ifp->if_softc = sc; 854 snprintf(ifp->if_xname, sizeof ifp->if_xname, "%s%d", ifc->ifc_name, 855 unit); 856 ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST; 857 ifp->if_ioctl = carp_ioctl; 858 ifp->if_start = carp_start; 859 IFQ_SET_MAXLEN(&ifp->if_snd, ifqmaxlen); 860 IFQ_SET_READY(&ifp->if_snd); 861 if_initialize(ifp); 862 ether_ifattach(ifp, NULL); 863 carp_set_enaddr(sc); 864 /* Overwrite ethernet defaults */ 865 ifp->if_type = IFT_CARP; 866 ifp->if_output = carp_output; 867 ifp->if_extflags &= ~IFEF_OUTPUT_MPSAFE; 868 if_register(ifp); 869 870 return (0); 871 } 872 873 static int 874 carp_clone_destroy(struct ifnet *ifp) 875 { 876 struct carp_softc *sc = ifp->if_softc; 877 878 carpdetach(ifp->if_softc); 879 ether_ifdetach(ifp); 880 if_detach(ifp); 881 callout_destroy(&sc->sc_ad_tmo); 882 callout_destroy(&sc->sc_md_tmo); 883 callout_destroy(&sc->sc_md6_tmo); 884 free(ifp->if_softc, M_DEVBUF); 885 886 return (0); 887 } 888 889 static void 890 carpdetach(struct carp_softc *sc) 891 { 892 struct carp_if *cif; 893 int s; 894 895 callout_stop(&sc->sc_ad_tmo); 896 callout_stop(&sc->sc_md_tmo); 897 callout_stop(&sc->sc_md6_tmo); 898 899 if (sc->sc_suppress) 900 carp_suppress_preempt--; 901 sc->sc_suppress = 0; 902 903 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) 904 carp_suppress_preempt--; 905 sc->sc_sendad_errors = 0; 906 907 carp_set_state(sc, INIT); 908 sc->sc_if.if_flags &= ~IFF_UP; 909 carp_setrun(sc, 0); 910 carp_multicast_cleanup(sc); 911 912 KERNEL_LOCK(1, NULL); 913 s = splnet(); 914 if (sc->sc_carpdev != NULL) { 915 /* XXX linkstatehook removal */ 916 cif = (struct carp_if *)sc->sc_carpdev->if_carp; 917 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list); 918 if (!--cif->vhif_nvrs) { 919 ifpromisc(sc->sc_carpdev, 0); 920 sc->sc_carpdev->if_carp = NULL; 921 free(cif, M_IFADDR); 922 } 923 } 924 sc->sc_carpdev = NULL; 925 splx(s); 926 KERNEL_UNLOCK_ONE(NULL); 927 } 928 929 /* Detach an interface from the carp. */ 930 void 931 carp_ifdetach(struct ifnet *ifp) 932 { 933 struct carp_softc *sc, *nextsc; 934 struct carp_if *cif = (struct carp_if *)ifp->if_carp; 935 936 for (sc = TAILQ_FIRST(&cif->vhif_vrs); sc; sc = nextsc) { 937 nextsc = TAILQ_NEXT(sc, sc_list); 938 carpdetach(sc); 939 } 940 } 941 942 static int 943 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, 944 struct carp_header *ch) 945 { 946 if (sc->sc_init_counter) { 947 /* this could also be seconds since unix epoch */ 948 sc->sc_counter = cprng_fast64(); 949 } else 950 sc->sc_counter++; 951 952 ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff); 953 ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff); 954 955 carp_hmac_generate(sc, ch->carp_counter, ch->carp_md); 956 957 return (0); 958 } 959 960 static void 961 carp_send_ad_all(void) 962 { 963 struct ifnet *ifp; 964 struct carp_if *cif; 965 struct carp_softc *vh; 966 int s; 967 int bound = curlwp_bind(); 968 969 s = pserialize_read_enter(); 970 IFNET_READER_FOREACH(ifp) { 971 struct psref psref; 972 if (ifp->if_carp == NULL || ifp->if_type == IFT_CARP) 973 continue; 974 975 psref_acquire(&psref, &ifp->if_psref, ifnet_psref_class); 976 pserialize_read_exit(s); 977 978 cif = (struct carp_if *)ifp->if_carp; 979 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 980 if ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 981 (IFF_UP|IFF_RUNNING) && vh->sc_state == MASTER) 982 carp_send_ad(vh); 983 } 984 985 s = pserialize_read_enter(); 986 psref_release(&psref, &ifp->if_psref, ifnet_psref_class); 987 } 988 pserialize_read_exit(s); 989 curlwp_bindx(bound); 990 } 991 992 993 static void 994 carp_send_ad(void *v) 995 { 996 struct carp_header ch; 997 struct timeval tv; 998 struct carp_softc *sc = v; 999 struct carp_header *ch_ptr; 1000 struct mbuf *m; 1001 int error, len, advbase, advskew, s; 1002 struct sockaddr sa; 1003 1004 KERNEL_LOCK(1, NULL); 1005 s = splsoftnet(); 1006 1007 advbase = advskew = 0; /* Sssssh compiler */ 1008 if (sc->sc_carpdev == NULL) { 1009 sc->sc_if.if_oerrors++; 1010 goto retry_later; 1011 } 1012 1013 /* bow out if we've gone to backup (the carp interface is going down) */ 1014 if (sc->sc_bow_out) { 1015 sc->sc_bow_out = 0; 1016 advbase = 255; 1017 advskew = 255; 1018 } else { 1019 advbase = sc->sc_advbase; 1020 if (!carp_suppress_preempt || sc->sc_advskew > 240) 1021 advskew = sc->sc_advskew; 1022 else 1023 advskew = 240; 1024 tv.tv_sec = advbase; 1025 tv.tv_usec = advskew * 1000000 / 256; 1026 } 1027 1028 ch.carp_version = CARP_VERSION; 1029 ch.carp_type = CARP_ADVERTISEMENT; 1030 ch.carp_vhid = sc->sc_vhid; 1031 ch.carp_advbase = advbase; 1032 ch.carp_advskew = advskew; 1033 ch.carp_authlen = 7; /* XXX DEFINE */ 1034 ch.carp_pad1 = 0; /* must be zero */ 1035 ch.carp_cksum = 0; 1036 1037 1038 #ifdef INET 1039 if (sc->sc_naddrs) { 1040 struct ip *ip; 1041 struct ifaddr *ifa; 1042 int _s; 1043 1044 MGETHDR(m, M_DONTWAIT, MT_HEADER); 1045 if (m == NULL) { 1046 sc->sc_if.if_oerrors++; 1047 CARP_STATINC(CARP_STAT_ONOMEM); 1048 /* XXX maybe less ? */ 1049 goto retry_later; 1050 } 1051 MCLAIM(m, &carp_proto_mowner_tx); 1052 len = sizeof(*ip) + sizeof(ch); 1053 m->m_pkthdr.len = len; 1054 m_reset_rcvif(m); 1055 m->m_len = len; 1056 MH_ALIGN(m, m->m_len); 1057 m->m_flags |= M_MCAST; 1058 ip = mtod(m, struct ip *); 1059 ip->ip_v = IPVERSION; 1060 ip->ip_hl = sizeof(*ip) >> 2; 1061 ip->ip_tos = IPTOS_LOWDELAY; 1062 ip->ip_len = htons(len); 1063 ip->ip_id = 0; /* no need for id, we don't support fragments */ 1064 ip->ip_off = htons(IP_DF); 1065 ip->ip_ttl = CARP_DFLTTL; 1066 ip->ip_p = IPPROTO_CARP; 1067 ip->ip_sum = 0; 1068 1069 memset(&sa, 0, sizeof(sa)); 1070 sa.sa_family = AF_INET; 1071 _s = pserialize_read_enter(); 1072 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev); 1073 if (ifa == NULL) 1074 ip->ip_src.s_addr = 0; 1075 else 1076 ip->ip_src.s_addr = 1077 ifatoia(ifa)->ia_addr.sin_addr.s_addr; 1078 pserialize_read_exit(_s); 1079 ip->ip_dst.s_addr = INADDR_CARP_GROUP; 1080 1081 ch_ptr = (struct carp_header *)(&ip[1]); 1082 memcpy(ch_ptr, &ch, sizeof(ch)); 1083 if (carp_prepare_ad(m, sc, ch_ptr)) 1084 goto retry_later; 1085 1086 m->m_data += sizeof(*ip); 1087 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip)); 1088 m->m_data -= sizeof(*ip); 1089 1090 nanotime(&sc->sc_if.if_lastchange); 1091 sc->sc_if.if_opackets++; 1092 sc->sc_if.if_obytes += len; 1093 CARP_STATINC(CARP_STAT_OPACKETS); 1094 1095 error = ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, 1096 NULL); 1097 if (error) { 1098 if (error == ENOBUFS) 1099 CARP_STATINC(CARP_STAT_ONOMEM); 1100 else 1101 CARP_LOG(sc, ("ip_output failed: %d", error)); 1102 sc->sc_if.if_oerrors++; 1103 if (sc->sc_sendad_errors < INT_MAX) 1104 sc->sc_sendad_errors++; 1105 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) { 1106 carp_suppress_preempt++; 1107 if (carp_suppress_preempt == 1) 1108 carp_send_ad_all(); 1109 } 1110 sc->sc_sendad_success = 0; 1111 } else { 1112 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) { 1113 if (++sc->sc_sendad_success >= 1114 CARP_SENDAD_MIN_SUCCESS) { 1115 carp_suppress_preempt--; 1116 sc->sc_sendad_errors = 0; 1117 } 1118 } else 1119 sc->sc_sendad_errors = 0; 1120 } 1121 } 1122 #endif /* INET */ 1123 #ifdef INET6_notyet 1124 if (sc->sc_naddrs6) { 1125 struct ip6_hdr *ip6; 1126 struct ifaddr *ifa; 1127 int _s; 1128 1129 MGETHDR(m, M_DONTWAIT, MT_HEADER); 1130 if (m == NULL) { 1131 sc->sc_if.if_oerrors++; 1132 CARP_STATINC(CARP_STAT_ONOMEM); 1133 /* XXX maybe less ? */ 1134 goto retry_later; 1135 } 1136 MCLAIM(m, &carp_proto6_mowner_tx); 1137 len = sizeof(*ip6) + sizeof(ch); 1138 m->m_pkthdr.len = len; 1139 m_reset_rcvif(m); 1140 m->m_len = len; 1141 MH_ALIGN(m, m->m_len); 1142 m->m_flags |= M_MCAST; 1143 ip6 = mtod(m, struct ip6_hdr *); 1144 memset(ip6, 0, sizeof(*ip6)); 1145 ip6->ip6_vfc |= IPV6_VERSION; 1146 ip6->ip6_hlim = CARP_DFLTTL; 1147 ip6->ip6_nxt = IPPROTO_CARP; 1148 1149 /* set the source address */ 1150 memset(&sa, 0, sizeof(sa)); 1151 sa.sa_family = AF_INET6; 1152 _s = pserialize_read_enter(); 1153 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev); 1154 if (ifa == NULL) /* This should never happen with IPv6 */ 1155 memset(&ip6->ip6_src, 0, sizeof(struct in6_addr)); 1156 else 1157 bcopy(ifatoia6(ifa)->ia_addr.sin6_addr.s6_addr, 1158 &ip6->ip6_src, sizeof(struct in6_addr)); 1159 pserialize_read_exit(_s); 1160 /* set the multicast destination */ 1161 1162 ip6->ip6_dst.s6_addr16[0] = htons(0xff02); 1163 ip6->ip6_dst.s6_addr8[15] = 0x12; 1164 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) { 1165 sc->sc_if.if_oerrors++; 1166 m_freem(m); 1167 CARP_LOG(sc, ("in6_setscope failed")); 1168 goto retry_later; 1169 } 1170 1171 ch_ptr = (struct carp_header *)(&ip6[1]); 1172 memcpy(ch_ptr, &ch, sizeof(ch)); 1173 if (carp_prepare_ad(m, sc, ch_ptr)) 1174 goto retry_later; 1175 1176 m->m_data += sizeof(*ip6); 1177 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6)); 1178 m->m_data -= sizeof(*ip6); 1179 1180 nanotime(&sc->sc_if.if_lastchange); 1181 sc->sc_if.if_opackets++; 1182 sc->sc_if.if_obytes += len; 1183 CARP_STATINC(CARP_STAT_OPACKETS6); 1184 1185 error = ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL); 1186 if (error) { 1187 if (error == ENOBUFS) 1188 CARP_STATINC(CARP_STAT_ONOMEM); 1189 else 1190 CARP_LOG(sc, ("ip6_output failed: %d", error)); 1191 sc->sc_if.if_oerrors++; 1192 if (sc->sc_sendad_errors < INT_MAX) 1193 sc->sc_sendad_errors++; 1194 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) { 1195 carp_suppress_preempt++; 1196 if (carp_suppress_preempt == 1) 1197 carp_send_ad_all(); 1198 } 1199 sc->sc_sendad_success = 0; 1200 } else { 1201 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) { 1202 if (++sc->sc_sendad_success >= 1203 CARP_SENDAD_MIN_SUCCESS) { 1204 carp_suppress_preempt--; 1205 sc->sc_sendad_errors = 0; 1206 } 1207 } else 1208 sc->sc_sendad_errors = 0; 1209 } 1210 } 1211 #endif /* INET6 */ 1212 1213 retry_later: 1214 splx(s); 1215 KERNEL_UNLOCK_ONE(NULL); 1216 if (advbase != 255 || advskew != 255) 1217 callout_schedule(&sc->sc_ad_tmo, tvtohz(&tv)); 1218 } 1219 1220 /* 1221 * Broadcast a gratuitous ARP request containing 1222 * the virtual router MAC address for each IP address 1223 * associated with the virtual router. 1224 */ 1225 static void 1226 carp_send_arp(struct carp_softc *sc) 1227 { 1228 struct ifaddr *ifa; 1229 int s; 1230 1231 KERNEL_LOCK(1, NULL); 1232 s = splsoftnet(); 1233 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 1234 1235 if (ifa->ifa_addr->sa_family != AF_INET) 1236 continue; 1237 1238 arpannounce(sc->sc_carpdev, ifa, CLLADDR(sc->sc_if.if_sadl)); 1239 } 1240 splx(s); 1241 KERNEL_UNLOCK_ONE(NULL); 1242 } 1243 1244 #ifdef INET6 1245 static void 1246 carp_send_na(struct carp_softc *sc) 1247 { 1248 struct ifaddr *ifa; 1249 struct in6_addr *in6; 1250 static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT; 1251 int s; 1252 1253 KERNEL_LOCK(1, NULL); 1254 s = splsoftnet(); 1255 1256 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 1257 1258 if (ifa->ifa_addr->sa_family != AF_INET6) 1259 continue; 1260 1261 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr; 1262 nd6_na_output(sc->sc_carpdev, &mcast, in6, 1263 ND_NA_FLAG_OVERRIDE, 1, NULL); 1264 } 1265 splx(s); 1266 KERNEL_UNLOCK_ONE(NULL); 1267 } 1268 #endif /* INET6 */ 1269 1270 /* 1271 * Based on bridge_hash() in if_bridge.c 1272 */ 1273 #define mix(a,b,c) \ 1274 do { \ 1275 a -= b; a -= c; a ^= (c >> 13); \ 1276 b -= c; b -= a; b ^= (a << 8); \ 1277 c -= a; c -= b; c ^= (b >> 13); \ 1278 a -= b; a -= c; a ^= (c >> 12); \ 1279 b -= c; b -= a; b ^= (a << 16); \ 1280 c -= a; c -= b; c ^= (b >> 5); \ 1281 a -= b; a -= c; a ^= (c >> 3); \ 1282 b -= c; b -= a; b ^= (a << 10); \ 1283 c -= a; c -= b; c ^= (b >> 15); \ 1284 } while (0) 1285 1286 static u_int32_t 1287 carp_hash(struct carp_softc *sc, u_char *src) 1288 { 1289 u_int32_t a = 0x9e3779b9, b = sc->sc_hashkey[0], c = sc->sc_hashkey[1]; 1290 1291 c += sc->sc_key[3] << 24; 1292 c += sc->sc_key[2] << 16; 1293 c += sc->sc_key[1] << 8; 1294 c += sc->sc_key[0]; 1295 b += src[5] << 8; 1296 b += src[4]; 1297 a += src[3] << 24; 1298 a += src[2] << 16; 1299 a += src[1] << 8; 1300 a += src[0]; 1301 1302 mix(a, b, c); 1303 return (c); 1304 } 1305 1306 static int 1307 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type) 1308 { 1309 struct carp_softc *vh; 1310 struct ifaddr *ifa; 1311 int count = 0; 1312 1313 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1314 if ((type == CARP_COUNT_RUNNING && 1315 (vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1316 (IFF_UP|IFF_RUNNING)) || 1317 (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) { 1318 IFADDR_READER_FOREACH(ifa, &vh->sc_if) { 1319 if (ifa->ifa_addr->sa_family == AF_INET && 1320 ia->ia_addr.sin_addr.s_addr == 1321 ifatoia(ifa)->ia_addr.sin_addr.s_addr) 1322 count++; 1323 } 1324 } 1325 } 1326 return (count); 1327 } 1328 1329 int 1330 carp_iamatch(struct in_ifaddr *ia, u_char *src, 1331 u_int32_t *count, u_int32_t index) 1332 { 1333 struct carp_softc *sc = ia->ia_ifp->if_softc; 1334 1335 if (carp_opts[CARPCTL_ARPBALANCE]) { 1336 /* 1337 * We use the source ip to decide which virtual host should 1338 * handle the request. If we're master of that virtual host, 1339 * then we respond, otherwise, just drop the arp packet on 1340 * the floor. 1341 */ 1342 1343 /* Count the elegible carp interfaces with this address */ 1344 if (*count == 0) 1345 *count = carp_addrcount( 1346 (struct carp_if *)ia->ia_ifp->if_carpdev->if_carp, 1347 ia, CARP_COUNT_RUNNING); 1348 1349 /* This should never happen, but... */ 1350 if (*count == 0) 1351 return (0); 1352 1353 if (carp_hash(sc, src) % *count == index - 1 && 1354 sc->sc_state == MASTER) { 1355 return (1); 1356 } 1357 } else { 1358 if (sc->sc_state == MASTER) 1359 return (1); 1360 } 1361 1362 return (0); 1363 } 1364 1365 #ifdef INET6 1366 struct ifaddr * 1367 carp_iamatch6(void *v, struct in6_addr *taddr) 1368 { 1369 struct carp_if *cif = v; 1370 struct carp_softc *vh; 1371 struct ifaddr *ifa; 1372 1373 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1374 IFADDR_READER_FOREACH(ifa, &vh->sc_if) { 1375 if (IN6_ARE_ADDR_EQUAL(taddr, 1376 &ifatoia6(ifa)->ia_addr.sin6_addr) && 1377 ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1378 (IFF_UP|IFF_RUNNING)) && vh->sc_state == MASTER) 1379 return (ifa); 1380 } 1381 } 1382 1383 return (NULL); 1384 } 1385 #endif /* INET6 */ 1386 1387 struct ifnet * 1388 carp_ourether(void *v, struct ether_header *eh, u_char iftype, int src) 1389 { 1390 struct carp_if *cif = (struct carp_if *)v; 1391 struct carp_softc *vh; 1392 u_int8_t *ena; 1393 1394 if (src) 1395 ena = (u_int8_t *)&eh->ether_shost; 1396 else 1397 ena = (u_int8_t *)&eh->ether_dhost; 1398 1399 switch (iftype) { 1400 case IFT_ETHER: 1401 case IFT_FDDI: 1402 if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1) 1403 return (NULL); 1404 break; 1405 case IFT_ISO88025: 1406 if (ena[0] != 3 || ena[1] || ena[4] || ena[5]) 1407 return (NULL); 1408 break; 1409 default: 1410 return (NULL); 1411 break; 1412 } 1413 1414 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) 1415 if ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1416 (IFF_UP|IFF_RUNNING) && vh->sc_state == MASTER && 1417 !memcmp(ena, CLLADDR(vh->sc_if.if_sadl), 1418 ETHER_ADDR_LEN)) { 1419 return (&vh->sc_if); 1420 } 1421 1422 return (NULL); 1423 } 1424 1425 int 1426 carp_input(struct mbuf *m, u_int8_t *shost, u_int8_t *dhost, u_int16_t etype) 1427 { 1428 struct ether_header eh; 1429 struct carp_if *cif = (struct carp_if *)m_get_rcvif_NOMPSAFE(m)->if_carp; 1430 struct ifnet *ifp; 1431 1432 memcpy(&eh.ether_shost, shost, sizeof(eh.ether_shost)); 1433 memcpy(&eh.ether_dhost, dhost, sizeof(eh.ether_dhost)); 1434 eh.ether_type = etype; 1435 1436 if (m->m_flags & (M_BCAST|M_MCAST)) { 1437 struct carp_softc *vh; 1438 struct mbuf *m0; 1439 1440 /* 1441 * XXX Should really check the list of multicast addresses 1442 * for each CARP interface _before_ copying. 1443 */ 1444 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1445 m0 = m_copym(m, 0, M_COPYALL, M_DONTWAIT); 1446 if (m0 == NULL) 1447 continue; 1448 m_set_rcvif(m0, &vh->sc_if); 1449 ether_input(&vh->sc_if, m0); 1450 } 1451 return (1); 1452 } 1453 1454 ifp = carp_ourether(cif, &eh, m_get_rcvif_NOMPSAFE(m)->if_type, 0); 1455 if (ifp == NULL) { 1456 return (1); 1457 } 1458 1459 m_set_rcvif(m, ifp); 1460 1461 bpf_mtap(ifp, m); 1462 ifp->if_ipackets++; 1463 ether_input(ifp, m); 1464 return (0); 1465 } 1466 1467 static void 1468 carp_master_down(void *v) 1469 { 1470 struct carp_softc *sc = v; 1471 1472 switch (sc->sc_state) { 1473 case INIT: 1474 printf("%s: master_down event in INIT state\n", 1475 sc->sc_if.if_xname); 1476 break; 1477 case MASTER: 1478 break; 1479 case BACKUP: 1480 CARP_LOG(sc, ("INIT -> MASTER (preempting)")); 1481 carp_set_state(sc, MASTER); 1482 carp_send_ad(sc); 1483 carp_send_arp(sc); 1484 #ifdef INET6 1485 carp_send_na(sc); 1486 #endif /* INET6 */ 1487 carp_setrun(sc, 0); 1488 carp_setroute(sc, RTM_ADD); 1489 break; 1490 } 1491 } 1492 1493 /* 1494 * When in backup state, af indicates whether to reset the master down timer 1495 * for v4 or v6. If it's set to zero, reset the ones which are already pending. 1496 */ 1497 static void 1498 carp_setrun(struct carp_softc *sc, sa_family_t af) 1499 { 1500 struct timeval tv; 1501 1502 if (sc->sc_carpdev == NULL) { 1503 sc->sc_if.if_flags &= ~IFF_RUNNING; 1504 carp_set_state(sc, INIT); 1505 return; 1506 } 1507 1508 if (sc->sc_if.if_flags & IFF_UP && sc->sc_vhid > 0 && 1509 (sc->sc_naddrs || sc->sc_naddrs6) && !sc->sc_suppress) { 1510 sc->sc_if.if_flags |= IFF_RUNNING; 1511 } else { 1512 sc->sc_if.if_flags &= ~IFF_RUNNING; 1513 carp_setroute(sc, RTM_DELETE); 1514 return; 1515 } 1516 1517 switch (sc->sc_state) { 1518 case INIT: 1519 carp_set_state(sc, BACKUP); 1520 carp_setroute(sc, RTM_DELETE); 1521 carp_setrun(sc, 0); 1522 break; 1523 case BACKUP: 1524 callout_stop(&sc->sc_ad_tmo); 1525 tv.tv_sec = 3 * sc->sc_advbase; 1526 tv.tv_usec = sc->sc_advskew * 1000000 / 256; 1527 switch (af) { 1528 #ifdef INET 1529 case AF_INET: 1530 callout_schedule(&sc->sc_md_tmo, tvtohz(&tv)); 1531 break; 1532 #endif /* INET */ 1533 #ifdef INET6_notyet 1534 case AF_INET6: 1535 callout_schedule(&sc->sc_md6_tmo, tvtohz(&tv)); 1536 break; 1537 #endif /* INET6 */ 1538 default: 1539 if (sc->sc_naddrs) 1540 callout_schedule(&sc->sc_md_tmo, tvtohz(&tv)); 1541 #ifdef INET6_notyet 1542 if (sc->sc_naddrs6) 1543 callout_schedule(&sc->sc_md6_tmo, tvtohz(&tv)); 1544 #endif /* INET6 */ 1545 break; 1546 } 1547 break; 1548 case MASTER: 1549 tv.tv_sec = sc->sc_advbase; 1550 tv.tv_usec = sc->sc_advskew * 1000000 / 256; 1551 callout_schedule(&sc->sc_ad_tmo, tvtohz(&tv)); 1552 break; 1553 } 1554 } 1555 1556 static void 1557 carp_multicast_cleanup(struct carp_softc *sc) 1558 { 1559 struct ip_moptions *imo = &sc->sc_imo; 1560 #ifdef INET6 1561 struct ip6_moptions *im6o = &sc->sc_im6o; 1562 #endif 1563 u_int16_t n = imo->imo_num_memberships; 1564 1565 /* Clean up our own multicast memberships */ 1566 while (n-- > 0) { 1567 if (imo->imo_membership[n] != NULL) { 1568 in_delmulti(imo->imo_membership[n]); 1569 imo->imo_membership[n] = NULL; 1570 } 1571 } 1572 imo->imo_num_memberships = 0; 1573 imo->imo_multicast_if_index = 0; 1574 1575 #ifdef INET6 1576 while (!LIST_EMPTY(&im6o->im6o_memberships)) { 1577 struct in6_multi_mship *imm = 1578 LIST_FIRST(&im6o->im6o_memberships); 1579 1580 LIST_REMOVE(imm, i6mm_chain); 1581 in6_leavegroup(imm); 1582 } 1583 im6o->im6o_multicast_if_index = 0; 1584 #endif 1585 1586 /* And any other multicast memberships */ 1587 carp_ether_purgemulti(sc); 1588 } 1589 1590 static int 1591 carp_set_ifp(struct carp_softc *sc, struct ifnet *ifp) 1592 { 1593 struct carp_if *cif, *ncif = NULL; 1594 struct carp_softc *vr, *after = NULL; 1595 int myself = 0, error = 0; 1596 int s; 1597 1598 if (ifp == sc->sc_carpdev) 1599 return (0); 1600 1601 if (ifp != NULL) { 1602 if ((ifp->if_flags & IFF_MULTICAST) == 0) 1603 return (EADDRNOTAVAIL); 1604 1605 if (ifp->if_type == IFT_CARP) 1606 return (EINVAL); 1607 1608 if (ifp->if_carp == NULL) { 1609 ncif = malloc(sizeof(*cif), M_IFADDR, M_NOWAIT); 1610 if (ncif == NULL) 1611 return (ENOBUFS); 1612 if ((error = ifpromisc(ifp, 1))) { 1613 free(ncif, M_IFADDR); 1614 return (error); 1615 } 1616 1617 ncif->vhif_ifp = ifp; 1618 TAILQ_INIT(&ncif->vhif_vrs); 1619 } else { 1620 cif = (struct carp_if *)ifp->if_carp; 1621 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) 1622 if (vr != sc && vr->sc_vhid == sc->sc_vhid) 1623 return (EINVAL); 1624 } 1625 1626 /* detach from old interface */ 1627 if (sc->sc_carpdev != NULL) 1628 carpdetach(sc); 1629 1630 /* join multicast groups */ 1631 if (sc->sc_naddrs < 0 && 1632 (error = carp_join_multicast(sc)) != 0) { 1633 if (ncif != NULL) 1634 free(ncif, M_IFADDR); 1635 return (error); 1636 } 1637 1638 #ifdef INET6 1639 if (sc->sc_naddrs6 < 0 && 1640 (error = carp_join_multicast6(sc)) != 0) { 1641 if (ncif != NULL) 1642 free(ncif, M_IFADDR); 1643 carp_multicast_cleanup(sc); 1644 return (error); 1645 } 1646 #endif 1647 1648 /* attach carp interface to physical interface */ 1649 if (ncif != NULL) 1650 ifp->if_carp = (void *)ncif; 1651 sc->sc_carpdev = ifp; 1652 sc->sc_if.if_capabilities = ifp->if_capabilities & 1653 (IFCAP_TSOv4 | IFCAP_TSOv6 | 1654 IFCAP_CSUM_IPv4_Tx|IFCAP_CSUM_IPv4_Rx| 1655 IFCAP_CSUM_TCPv4_Tx|IFCAP_CSUM_TCPv4_Rx| 1656 IFCAP_CSUM_UDPv4_Tx|IFCAP_CSUM_UDPv4_Rx| 1657 IFCAP_CSUM_TCPv6_Tx|IFCAP_CSUM_TCPv6_Rx| 1658 IFCAP_CSUM_UDPv6_Tx|IFCAP_CSUM_UDPv6_Rx); 1659 1660 cif = (struct carp_if *)ifp->if_carp; 1661 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) { 1662 if (vr == sc) 1663 myself = 1; 1664 if (vr->sc_vhid < sc->sc_vhid) 1665 after = vr; 1666 } 1667 1668 if (!myself) { 1669 /* We're trying to keep things in order */ 1670 if (after == NULL) { 1671 TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list); 1672 } else { 1673 TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, 1674 sc, sc_list); 1675 } 1676 cif->vhif_nvrs++; 1677 } 1678 if (sc->sc_naddrs || sc->sc_naddrs6) 1679 sc->sc_if.if_flags |= IFF_UP; 1680 carp_set_enaddr(sc); 1681 KERNEL_LOCK(1, NULL); 1682 s = splnet(); 1683 /* XXX linkstatehooks establish */ 1684 carp_carpdev_state(ifp); 1685 splx(s); 1686 KERNEL_UNLOCK_ONE(NULL); 1687 } else { 1688 carpdetach(sc); 1689 sc->sc_if.if_flags &= ~(IFF_UP|IFF_RUNNING); 1690 } 1691 return (0); 1692 } 1693 1694 static void 1695 carp_set_enaddr(struct carp_softc *sc) 1696 { 1697 uint8_t enaddr[ETHER_ADDR_LEN]; 1698 if (sc->sc_carpdev && sc->sc_carpdev->if_type == IFT_ISO88025) { 1699 enaddr[0] = 3; 1700 enaddr[1] = 0; 1701 enaddr[2] = 0x40 >> (sc->sc_vhid - 1); 1702 enaddr[3] = 0x40000 >> (sc->sc_vhid - 1); 1703 enaddr[4] = 0; 1704 enaddr[5] = 0; 1705 } else { 1706 enaddr[0] = 0; 1707 enaddr[1] = 0; 1708 enaddr[2] = 0x5e; 1709 enaddr[3] = 0; 1710 enaddr[4] = 1; 1711 enaddr[5] = sc->sc_vhid; 1712 } 1713 if_set_sadl(&sc->sc_if, enaddr, sizeof(enaddr), false); 1714 } 1715 1716 #if 0 1717 static void 1718 carp_addr_updated(void *v) 1719 { 1720 struct carp_softc *sc = (struct carp_softc *) v; 1721 struct ifaddr *ifa; 1722 int new_naddrs = 0, new_naddrs6 = 0; 1723 1724 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 1725 if (ifa->ifa_addr->sa_family == AF_INET) 1726 new_naddrs++; 1727 else if (ifa->ifa_addr->sa_family == AF_INET6) 1728 new_naddrs6++; 1729 } 1730 1731 /* Handle a callback after SIOCDIFADDR */ 1732 if (new_naddrs < sc->sc_naddrs || new_naddrs6 < sc->sc_naddrs6) { 1733 struct in_addr mc_addr; 1734 1735 sc->sc_naddrs = new_naddrs; 1736 sc->sc_naddrs6 = new_naddrs6; 1737 1738 /* Re-establish multicast membership removed by in_control */ 1739 mc_addr.s_addr = INADDR_CARP_GROUP; 1740 if (!in_multi_group(mc_addr, &sc->sc_if, 0)) { 1741 memset(&sc->sc_imo, 0, sizeof(sc->sc_imo)); 1742 1743 if (sc->sc_carpdev != NULL && sc->sc_naddrs > 0) 1744 carp_join_multicast(sc); 1745 } 1746 1747 if (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) { 1748 sc->sc_if.if_flags &= ~IFF_UP; 1749 carp_set_state(sc, INIT); 1750 } else 1751 carp_hmac_prepare(sc); 1752 } 1753 1754 carp_setrun(sc, 0); 1755 } 1756 #endif 1757 1758 static int 1759 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin) 1760 { 1761 struct ifnet *ifp = sc->sc_carpdev; 1762 struct in_ifaddr *ia, *ia_if; 1763 int error = 0; 1764 int s; 1765 1766 if (sin->sin_addr.s_addr == 0) { 1767 if (!(sc->sc_if.if_flags & IFF_UP)) 1768 carp_set_state(sc, INIT); 1769 if (sc->sc_naddrs) 1770 sc->sc_if.if_flags |= IFF_UP; 1771 carp_setrun(sc, 0); 1772 return (0); 1773 } 1774 1775 /* we have to do this by hand to ensure we don't match on ourselves */ 1776 ia_if = NULL; 1777 s = pserialize_read_enter(); 1778 IN_ADDRLIST_READER_FOREACH(ia) { 1779 /* and, yeah, we need a multicast-capable iface too */ 1780 if (ia->ia_ifp != &sc->sc_if && 1781 ia->ia_ifp->if_type != IFT_CARP && 1782 (ia->ia_ifp->if_flags & IFF_MULTICAST) && 1783 (sin->sin_addr.s_addr & ia->ia_subnetmask) == 1784 ia->ia_subnet) { 1785 if (!ia_if) 1786 ia_if = ia; 1787 } 1788 } 1789 1790 if (ia_if) { 1791 ia = ia_if; 1792 if (ifp) { 1793 if (ifp != ia->ia_ifp) 1794 return (EADDRNOTAVAIL); 1795 } else { 1796 /* FIXME NOMPSAFE */ 1797 ifp = ia->ia_ifp; 1798 } 1799 } 1800 pserialize_read_exit(s); 1801 1802 if ((error = carp_set_ifp(sc, ifp))) 1803 return (error); 1804 1805 if (sc->sc_carpdev == NULL) 1806 return (EADDRNOTAVAIL); 1807 1808 if (sc->sc_naddrs == 0 && (error = carp_join_multicast(sc)) != 0) 1809 return (error); 1810 1811 sc->sc_naddrs++; 1812 if (sc->sc_carpdev != NULL) 1813 sc->sc_if.if_flags |= IFF_UP; 1814 1815 carp_set_state(sc, INIT); 1816 carp_setrun(sc, 0); 1817 1818 /* 1819 * Hook if_addrhooks so that we get a callback after in_ifinit has run, 1820 * to correct any inappropriate routes that it inserted. 1821 */ 1822 if (sc->ah_cookie == 0) { 1823 /* XXX link address hook */ 1824 } 1825 1826 return (0); 1827 } 1828 1829 static int 1830 carp_join_multicast(struct carp_softc *sc) 1831 { 1832 struct ip_moptions *imo = &sc->sc_imo, tmpimo; 1833 struct in_addr addr; 1834 1835 memset(&tmpimo, 0, sizeof(tmpimo)); 1836 addr.s_addr = INADDR_CARP_GROUP; 1837 if ((tmpimo.imo_membership[0] = 1838 in_addmulti(&addr, &sc->sc_if)) == NULL) { 1839 return (ENOBUFS); 1840 } 1841 1842 imo->imo_membership[0] = tmpimo.imo_membership[0]; 1843 imo->imo_num_memberships = 1; 1844 imo->imo_multicast_if_index = sc->sc_if.if_index; 1845 imo->imo_multicast_ttl = CARP_DFLTTL; 1846 imo->imo_multicast_loop = 0; 1847 return (0); 1848 } 1849 1850 1851 #ifdef INET6 1852 static int 1853 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6) 1854 { 1855 struct ifnet *ifp = sc->sc_carpdev; 1856 struct in6_ifaddr *ia, *ia_if; 1857 int error = 0; 1858 int s; 1859 1860 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) { 1861 if (!(sc->sc_if.if_flags & IFF_UP)) 1862 carp_set_state(sc, INIT); 1863 if (sc->sc_naddrs6) 1864 sc->sc_if.if_flags |= IFF_UP; 1865 carp_setrun(sc, 0); 1866 return (0); 1867 } 1868 1869 /* we have to do this by hand to ensure we don't match on ourselves */ 1870 ia_if = NULL; 1871 s = pserialize_read_enter(); 1872 IN6_ADDRLIST_READER_FOREACH(ia) { 1873 int i; 1874 1875 for (i = 0; i < 4; i++) { 1876 if ((sin6->sin6_addr.s6_addr32[i] & 1877 ia->ia_prefixmask.sin6_addr.s6_addr32[i]) != 1878 (ia->ia_addr.sin6_addr.s6_addr32[i] & 1879 ia->ia_prefixmask.sin6_addr.s6_addr32[i])) 1880 break; 1881 } 1882 /* and, yeah, we need a multicast-capable iface too */ 1883 if (ia->ia_ifp != &sc->sc_if && 1884 ia->ia_ifp->if_type != IFT_CARP && 1885 (ia->ia_ifp->if_flags & IFF_MULTICAST) && 1886 (i == 4)) { 1887 if (!ia_if) 1888 ia_if = ia; 1889 } 1890 } 1891 pserialize_read_exit(s); 1892 1893 if (ia_if) { 1894 ia = ia_if; 1895 if (sc->sc_carpdev) { 1896 if (sc->sc_carpdev != ia->ia_ifp) 1897 return (EADDRNOTAVAIL); 1898 } else { 1899 ifp = ia->ia_ifp; 1900 } 1901 } 1902 1903 if ((error = carp_set_ifp(sc, ifp))) 1904 return (error); 1905 1906 if (sc->sc_carpdev == NULL) 1907 return (EADDRNOTAVAIL); 1908 1909 if (sc->sc_naddrs6 == 0 && (error = carp_join_multicast6(sc)) != 0) 1910 return (error); 1911 1912 sc->sc_naddrs6++; 1913 if (sc->sc_carpdev != NULL) 1914 sc->sc_if.if_flags |= IFF_UP; 1915 carp_set_state(sc, INIT); 1916 carp_setrun(sc, 0); 1917 1918 return (0); 1919 } 1920 1921 static int 1922 carp_join_multicast6(struct carp_softc *sc) 1923 { 1924 struct in6_multi_mship *imm, *imm2; 1925 struct ip6_moptions *im6o = &sc->sc_im6o; 1926 struct sockaddr_in6 addr6; 1927 int error; 1928 1929 /* Join IPv6 CARP multicast group */ 1930 memset(&addr6, 0, sizeof(addr6)); 1931 addr6.sin6_family = AF_INET6; 1932 addr6.sin6_len = sizeof(addr6); 1933 addr6.sin6_addr.s6_addr16[0] = htons(0xff02); 1934 addr6.sin6_addr.s6_addr16[1] = htons(sc->sc_if.if_index); 1935 addr6.sin6_addr.s6_addr8[15] = 0x12; 1936 if ((imm = in6_joingroup(&sc->sc_if, 1937 &addr6.sin6_addr, &error, 0)) == NULL) { 1938 return (error); 1939 } 1940 /* join solicited multicast address */ 1941 memset(&addr6.sin6_addr, 0, sizeof(addr6.sin6_addr)); 1942 addr6.sin6_addr.s6_addr16[0] = htons(0xff02); 1943 addr6.sin6_addr.s6_addr16[1] = htons(sc->sc_if.if_index); 1944 addr6.sin6_addr.s6_addr32[1] = 0; 1945 addr6.sin6_addr.s6_addr32[2] = htonl(1); 1946 addr6.sin6_addr.s6_addr32[3] = 0; 1947 addr6.sin6_addr.s6_addr8[12] = 0xff; 1948 if ((imm2 = in6_joingroup(&sc->sc_if, 1949 &addr6.sin6_addr, &error, 0)) == NULL) { 1950 in6_leavegroup(imm); 1951 return (error); 1952 } 1953 1954 /* apply v6 multicast membership */ 1955 im6o->im6o_multicast_if_index = sc->sc_if.if_index; 1956 if (imm) 1957 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, 1958 i6mm_chain); 1959 if (imm2) 1960 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm2, 1961 i6mm_chain); 1962 1963 return (0); 1964 } 1965 1966 #endif /* INET6 */ 1967 1968 static int 1969 carp_ioctl(struct ifnet *ifp, u_long cmd, void *data) 1970 { 1971 struct lwp *l = curlwp; /* XXX */ 1972 struct carp_softc *sc = ifp->if_softc, *vr; 1973 struct carpreq carpr; 1974 struct ifaddr *ifa; 1975 struct ifreq *ifr; 1976 struct ifnet *cdev = NULL; 1977 int error = 0; 1978 1979 ifa = (struct ifaddr *)data; 1980 ifr = (struct ifreq *)data; 1981 1982 switch (cmd) { 1983 case SIOCINITIFADDR: 1984 switch (ifa->ifa_addr->sa_family) { 1985 #ifdef INET 1986 case AF_INET: 1987 sc->sc_if.if_flags |= IFF_UP; 1988 memcpy(ifa->ifa_dstaddr, ifa->ifa_addr, 1989 sizeof(struct sockaddr)); 1990 error = carp_set_addr(sc, satosin(ifa->ifa_addr)); 1991 break; 1992 #endif /* INET */ 1993 #ifdef INET6 1994 case AF_INET6: 1995 sc->sc_if.if_flags|= IFF_UP; 1996 error = carp_set_addr6(sc, satosin6(ifa->ifa_addr)); 1997 break; 1998 #endif /* INET6 */ 1999 default: 2000 error = EAFNOSUPPORT; 2001 break; 2002 } 2003 break; 2004 2005 case SIOCSIFFLAGS: 2006 if ((error = ifioctl_common(ifp, cmd, data)) != 0) 2007 break; 2008 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) { 2009 callout_stop(&sc->sc_ad_tmo); 2010 callout_stop(&sc->sc_md_tmo); 2011 callout_stop(&sc->sc_md6_tmo); 2012 if (sc->sc_state == MASTER) { 2013 /* we need the interface up to bow out */ 2014 sc->sc_if.if_flags |= IFF_UP; 2015 sc->sc_bow_out = 1; 2016 carp_send_ad(sc); 2017 } 2018 sc->sc_if.if_flags &= ~IFF_UP; 2019 carp_set_state(sc, INIT); 2020 carp_setrun(sc, 0); 2021 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) { 2022 sc->sc_if.if_flags |= IFF_UP; 2023 carp_setrun(sc, 0); 2024 } 2025 break; 2026 2027 case SIOCSVH: 2028 if (l == NULL) 2029 break; 2030 if ((error = kauth_authorize_network(l->l_cred, 2031 KAUTH_NETWORK_INTERFACE, 2032 KAUTH_REQ_NETWORK_INTERFACE_SETPRIV, ifp, (void *)cmd, 2033 NULL)) != 0) 2034 break; 2035 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr))) 2036 break; 2037 error = 1; 2038 if (carpr.carpr_carpdev[0] != '\0' && 2039 (cdev = ifunit(carpr.carpr_carpdev)) == NULL) 2040 return (EINVAL); 2041 if ((error = carp_set_ifp(sc, cdev))) 2042 return (error); 2043 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) { 2044 switch (carpr.carpr_state) { 2045 case BACKUP: 2046 callout_stop(&sc->sc_ad_tmo); 2047 carp_set_state(sc, BACKUP); 2048 carp_setrun(sc, 0); 2049 carp_setroute(sc, RTM_DELETE); 2050 break; 2051 case MASTER: 2052 carp_master_down(sc); 2053 break; 2054 default: 2055 break; 2056 } 2057 } 2058 if (carpr.carpr_vhid > 0) { 2059 if (carpr.carpr_vhid > 255) { 2060 error = EINVAL; 2061 break; 2062 } 2063 if (sc->sc_carpdev) { 2064 struct carp_if *cif; 2065 cif = (struct carp_if *)sc->sc_carpdev->if_carp; 2066 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) 2067 if (vr != sc && 2068 vr->sc_vhid == carpr.carpr_vhid) 2069 return (EINVAL); 2070 } 2071 sc->sc_vhid = carpr.carpr_vhid; 2072 carp_set_enaddr(sc); 2073 carp_set_state(sc, INIT); 2074 error--; 2075 } 2076 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) { 2077 if (carpr.carpr_advskew > 254) { 2078 error = EINVAL; 2079 break; 2080 } 2081 if (carpr.carpr_advbase > 255) { 2082 error = EINVAL; 2083 break; 2084 } 2085 sc->sc_advbase = carpr.carpr_advbase; 2086 sc->sc_advskew = carpr.carpr_advskew; 2087 error--; 2088 } 2089 memcpy(sc->sc_key, carpr.carpr_key, sizeof(sc->sc_key)); 2090 if (error > 0) 2091 error = EINVAL; 2092 else { 2093 error = 0; 2094 carp_setrun(sc, 0); 2095 } 2096 break; 2097 2098 case SIOCGVH: 2099 memset(&carpr, 0, sizeof(carpr)); 2100 if (sc->sc_carpdev != NULL) 2101 strlcpy(carpr.carpr_carpdev, sc->sc_carpdev->if_xname, 2102 IFNAMSIZ); 2103 carpr.carpr_state = sc->sc_state; 2104 carpr.carpr_vhid = sc->sc_vhid; 2105 carpr.carpr_advbase = sc->sc_advbase; 2106 carpr.carpr_advskew = sc->sc_advskew; 2107 2108 if ((l != NULL) && (error = kauth_authorize_network(l->l_cred, 2109 KAUTH_NETWORK_INTERFACE, 2110 KAUTH_REQ_NETWORK_INTERFACE_SETPRIV, ifp, (void *)cmd, 2111 NULL)) == 0) 2112 memcpy(carpr.carpr_key, sc->sc_key, 2113 sizeof(carpr.carpr_key)); 2114 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr)); 2115 break; 2116 2117 case SIOCADDMULTI: 2118 error = carp_ether_addmulti(sc, ifr); 2119 break; 2120 2121 case SIOCDELMULTI: 2122 error = carp_ether_delmulti(sc, ifr); 2123 break; 2124 2125 case SIOCSIFCAP: 2126 if ((error = ifioctl_common(ifp, cmd, data)) == ENETRESET) 2127 error = 0; 2128 break; 2129 2130 default: 2131 error = ether_ioctl(ifp, cmd, data); 2132 } 2133 2134 carp_hmac_prepare(sc); 2135 return (error); 2136 } 2137 2138 2139 /* 2140 * Start output on carp interface. This function should never be called. 2141 */ 2142 static void 2143 carp_start(struct ifnet *ifp) 2144 { 2145 #ifdef DEBUG 2146 printf("%s: start called\n", ifp->if_xname); 2147 #endif 2148 } 2149 2150 int 2151 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa, 2152 const struct rtentry *rt) 2153 { 2154 struct carp_softc *sc = ((struct carp_softc *)ifp->if_softc); 2155 KASSERT(KERNEL_LOCKED_P()); 2156 2157 if (sc->sc_carpdev != NULL && sc->sc_state == MASTER) { 2158 return if_output_lock(sc->sc_carpdev, ifp, m, sa, rt); 2159 } else { 2160 m_freem(m); 2161 return (ENETUNREACH); 2162 } 2163 } 2164 2165 static void 2166 carp_set_state(struct carp_softc *sc, int state) 2167 { 2168 static const char *carp_states[] = { CARP_STATES }; 2169 if (sc->sc_state == state) 2170 return; 2171 2172 CARP_LOG(sc, ("state transition from: %s -> to: %s", carp_states[sc->sc_state], carp_states[state])); 2173 2174 sc->sc_state = state; 2175 switch (state) { 2176 case BACKUP: 2177 sc->sc_if.if_link_state = LINK_STATE_DOWN; 2178 break; 2179 case MASTER: 2180 sc->sc_if.if_link_state = LINK_STATE_UP; 2181 break; 2182 default: 2183 sc->sc_if.if_link_state = LINK_STATE_UNKNOWN; 2184 break; 2185 } 2186 rt_ifmsg(&sc->sc_if); 2187 } 2188 2189 void 2190 carp_carpdev_state(void *v) 2191 { 2192 struct carp_if *cif; 2193 struct carp_softc *sc; 2194 struct ifnet *ifp = v; 2195 2196 if (ifp->if_type == IFT_CARP) 2197 return; 2198 2199 cif = (struct carp_if *)ifp->if_carp; 2200 2201 TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) { 2202 int suppressed = sc->sc_suppress; 2203 2204 if (sc->sc_carpdev->if_link_state == LINK_STATE_DOWN || 2205 !(sc->sc_carpdev->if_flags & IFF_UP)) { 2206 sc->sc_if.if_flags &= ~IFF_RUNNING; 2207 callout_stop(&sc->sc_ad_tmo); 2208 callout_stop(&sc->sc_md_tmo); 2209 callout_stop(&sc->sc_md6_tmo); 2210 carp_set_state(sc, INIT); 2211 sc->sc_suppress = 1; 2212 carp_setrun(sc, 0); 2213 if (!suppressed) { 2214 carp_suppress_preempt++; 2215 if (carp_suppress_preempt == 1) 2216 carp_send_ad_all(); 2217 } 2218 } else { 2219 carp_set_state(sc, INIT); 2220 sc->sc_suppress = 0; 2221 carp_setrun(sc, 0); 2222 if (suppressed) 2223 carp_suppress_preempt--; 2224 } 2225 } 2226 } 2227 2228 static int 2229 carp_ether_addmulti(struct carp_softc *sc, struct ifreq *ifr) 2230 { 2231 const struct sockaddr *sa = ifreq_getaddr(SIOCADDMULTI, ifr); 2232 struct ifnet *ifp; 2233 struct carp_mc_entry *mc; 2234 u_int8_t addrlo[ETHER_ADDR_LEN], addrhi[ETHER_ADDR_LEN]; 2235 int error; 2236 2237 ifp = sc->sc_carpdev; 2238 if (ifp == NULL) 2239 return (EINVAL); 2240 2241 error = ether_addmulti(sa, &sc->sc_ac); 2242 if (error != ENETRESET) 2243 return (error); 2244 2245 /* 2246 * This is new multicast address. We have to tell parent 2247 * about it. Also, remember this multicast address so that 2248 * we can delete them on unconfigure. 2249 */ 2250 mc = malloc(sizeof(struct carp_mc_entry), M_DEVBUF, M_NOWAIT); 2251 if (mc == NULL) { 2252 error = ENOMEM; 2253 goto alloc_failed; 2254 } 2255 2256 /* 2257 * As ether_addmulti() returns ENETRESET, following two 2258 * statement shouldn't fail. 2259 */ 2260 (void)ether_multiaddr(sa, addrlo, addrhi); 2261 ETHER_LOOKUP_MULTI(addrlo, addrhi, &sc->sc_ac, mc->mc_enm); 2262 memcpy(&mc->mc_addr, sa, sa->sa_len); 2263 LIST_INSERT_HEAD(&sc->carp_mc_listhead, mc, mc_entries); 2264 2265 error = if_mcast_op(ifp, SIOCADDMULTI, sa); 2266 if (error != 0) 2267 goto ioctl_failed; 2268 2269 return (error); 2270 2271 ioctl_failed: 2272 LIST_REMOVE(mc, mc_entries); 2273 free(mc, M_DEVBUF); 2274 alloc_failed: 2275 (void)ether_delmulti(sa, &sc->sc_ac); 2276 2277 return (error); 2278 } 2279 2280 static int 2281 carp_ether_delmulti(struct carp_softc *sc, struct ifreq *ifr) 2282 { 2283 const struct sockaddr *sa = ifreq_getaddr(SIOCDELMULTI, ifr); 2284 struct ifnet *ifp; 2285 struct ether_multi *enm; 2286 struct carp_mc_entry *mc; 2287 u_int8_t addrlo[ETHER_ADDR_LEN], addrhi[ETHER_ADDR_LEN]; 2288 int error; 2289 2290 ifp = sc->sc_carpdev; 2291 if (ifp == NULL) 2292 return (EINVAL); 2293 2294 /* 2295 * Find a key to lookup carp_mc_entry. We have to do this 2296 * before calling ether_delmulti for obvious reason. 2297 */ 2298 if ((error = ether_multiaddr(sa, addrlo, addrhi)) != 0) 2299 return (error); 2300 ETHER_LOOKUP_MULTI(addrlo, addrhi, &sc->sc_ac, enm); 2301 if (enm == NULL) 2302 return (EINVAL); 2303 2304 LIST_FOREACH(mc, &sc->carp_mc_listhead, mc_entries) 2305 if (mc->mc_enm == enm) 2306 break; 2307 2308 /* We won't delete entries we didn't add */ 2309 if (mc == NULL) 2310 return (EINVAL); 2311 2312 error = ether_delmulti(sa, &sc->sc_ac); 2313 if (error != ENETRESET) 2314 return (error); 2315 2316 /* We no longer use this multicast address. Tell parent so. */ 2317 error = if_mcast_op(ifp, SIOCDELMULTI, sa); 2318 if (error == 0) { 2319 /* And forget about this address. */ 2320 LIST_REMOVE(mc, mc_entries); 2321 free(mc, M_DEVBUF); 2322 } else 2323 (void)ether_addmulti(sa, &sc->sc_ac); 2324 return (error); 2325 } 2326 2327 /* 2328 * Delete any multicast address we have asked to add from parent 2329 * interface. Called when the carp is being unconfigured. 2330 */ 2331 static void 2332 carp_ether_purgemulti(struct carp_softc *sc) 2333 { 2334 struct ifnet *ifp = sc->sc_carpdev; /* Parent. */ 2335 struct carp_mc_entry *mc; 2336 2337 if (ifp == NULL) 2338 return; 2339 2340 while ((mc = LIST_FIRST(&sc->carp_mc_listhead)) != NULL) { 2341 (void)if_mcast_op(ifp, SIOCDELMULTI, sstosa(&mc->mc_addr)); 2342 LIST_REMOVE(mc, mc_entries); 2343 free(mc, M_DEVBUF); 2344 } 2345 } 2346 2347 static int 2348 sysctl_net_inet_carp_stats(SYSCTLFN_ARGS) 2349 { 2350 2351 return (NETSTAT_SYSCTL(carpstat_percpu, CARP_NSTATS)); 2352 } 2353 2354 void 2355 carp_init(void) 2356 { 2357 2358 sysctl_net_inet_carp_setup(NULL); 2359 #ifdef MBUFTRACE 2360 MOWNER_ATTACH(&carp_proto_mowner_rx); 2361 MOWNER_ATTACH(&carp_proto_mowner_tx); 2362 MOWNER_ATTACH(&carp_proto6_mowner_rx); 2363 MOWNER_ATTACH(&carp_proto6_mowner_tx); 2364 #endif 2365 2366 carp_wqinput = wqinput_create("carp", _carp_proto_input); 2367 #ifdef INET6 2368 carp6_wqinput = wqinput_create("carp6", _carp6_proto_input); 2369 #endif 2370 } 2371 2372 static void 2373 sysctl_net_inet_carp_setup(struct sysctllog **clog) 2374 { 2375 2376 sysctl_createv(clog, 0, NULL, NULL, 2377 CTLFLAG_PERMANENT, 2378 CTLTYPE_NODE, "inet", NULL, 2379 NULL, 0, NULL, 0, 2380 CTL_NET, PF_INET, CTL_EOL); 2381 sysctl_createv(clog, 0, NULL, NULL, 2382 CTLFLAG_PERMANENT, 2383 CTLTYPE_NODE, "carp", 2384 SYSCTL_DESCR("CARP related settings"), 2385 NULL, 0, NULL, 0, 2386 CTL_NET, PF_INET, IPPROTO_CARP, CTL_EOL); 2387 2388 sysctl_createv(clog, 0, NULL, NULL, 2389 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2390 CTLTYPE_INT, "preempt", 2391 SYSCTL_DESCR("Enable CARP Preempt"), 2392 NULL, 0, &carp_opts[CARPCTL_PREEMPT], 0, 2393 CTL_NET, PF_INET, IPPROTO_CARP, 2394 CTL_CREATE, CTL_EOL); 2395 sysctl_createv(clog, 0, NULL, NULL, 2396 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2397 CTLTYPE_INT, "arpbalance", 2398 SYSCTL_DESCR("Enable ARP balancing"), 2399 NULL, 0, &carp_opts[CARPCTL_ARPBALANCE], 0, 2400 CTL_NET, PF_INET, IPPROTO_CARP, 2401 CTL_CREATE, CTL_EOL); 2402 sysctl_createv(clog, 0, NULL, NULL, 2403 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2404 CTLTYPE_INT, "allow", 2405 SYSCTL_DESCR("Enable CARP"), 2406 NULL, 0, &carp_opts[CARPCTL_ALLOW], 0, 2407 CTL_NET, PF_INET, IPPROTO_CARP, 2408 CTL_CREATE, CTL_EOL); 2409 sysctl_createv(clog, 0, NULL, NULL, 2410 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2411 CTLTYPE_INT, "log", 2412 SYSCTL_DESCR("CARP logging"), 2413 NULL, 0, &carp_opts[CARPCTL_LOG], 0, 2414 CTL_NET, PF_INET, IPPROTO_CARP, 2415 CTL_CREATE, CTL_EOL); 2416 sysctl_createv(clog, 0, NULL, NULL, 2417 CTLFLAG_PERMANENT, 2418 CTLTYPE_STRUCT, "stats", 2419 SYSCTL_DESCR("CARP statistics"), 2420 sysctl_net_inet_carp_stats, 0, NULL, 0, 2421 CTL_NET, PF_INET, IPPROTO_CARP, CARPCTL_STATS, 2422 CTL_EOL); 2423 } 2424