ip_carp.c revision 1.115
1 /* $NetBSD: ip_carp.c,v 1.115 2021/06/16 00:21:19 riastradh 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 1.115 2021/06/16 00:21:19 riastradh 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 #include <sys/cpu.h> 63 #include <sys/pserialize.h> 64 #include <sys/psref.h> 65 66 #include <net/if.h> 67 #include <net/pfil.h> 68 #include <net/if_types.h> 69 #include <net/if_ether.h> 70 #include <net/route.h> 71 #include <net/netisr.h> 72 #include <net/net_stats.h> 73 #include <netinet/if_inarp.h> 74 #include <netinet/wqinput.h> 75 76 #ifdef INET 77 #include <netinet/in.h> 78 #include <netinet/in_systm.h> 79 #include <netinet/in_var.h> 80 #include <netinet/ip.h> 81 #include <netinet/ip_var.h> 82 83 #include <net/if_dl.h> 84 #endif 85 86 #ifdef INET6 87 #include <netinet/icmp6.h> 88 #include <netinet/ip6.h> 89 #include <netinet6/ip6_var.h> 90 #include <netinet6/nd6.h> 91 #include <netinet6/scope6_var.h> 92 #include <netinet6/in6_var.h> 93 #endif 94 95 #include <net/bpf.h> 96 97 #include <sys/sha1.h> 98 99 #include <netinet/ip_carp.h> 100 101 #include "ioconf.h" 102 103 struct carp_mc_entry { 104 LIST_ENTRY(carp_mc_entry) mc_entries; 105 union { 106 struct ether_multi *mcu_enm; 107 } mc_u; 108 struct sockaddr_storage mc_addr; 109 }; 110 #define mc_enm mc_u.mcu_enm 111 112 struct carp_softc { 113 struct ethercom sc_ac; 114 #define sc_if sc_ac.ec_if 115 #define sc_carpdev sc_ac.ec_if.if_carpdev 116 int ah_cookie; 117 int lh_cookie; 118 struct ip_moptions sc_imo; 119 #ifdef INET6 120 struct ip6_moptions sc_im6o; 121 #endif /* INET6 */ 122 TAILQ_ENTRY(carp_softc) sc_list; 123 124 enum { INIT = 0, BACKUP, MASTER } sc_state; 125 126 int sc_suppress; 127 int sc_bow_out; 128 129 int sc_sendad_errors; 130 #define CARP_SENDAD_MAX_ERRORS 3 131 int sc_sendad_success; 132 #define CARP_SENDAD_MIN_SUCCESS 3 133 134 int sc_vhid; 135 int sc_advskew; 136 int sc_naddrs; 137 int sc_naddrs6; 138 int sc_advbase; /* seconds */ 139 int sc_init_counter; 140 u_int64_t sc_counter; 141 142 /* authentication */ 143 #define CARP_HMAC_PAD 64 144 unsigned char sc_key[CARP_KEY_LEN]; 145 unsigned char sc_pad[CARP_HMAC_PAD]; 146 SHA1_CTX sc_sha1; 147 u_int32_t sc_hashkey[2]; 148 149 struct callout sc_ad_tmo; /* advertisement timeout */ 150 struct callout sc_md_tmo; /* master down timeout */ 151 struct callout sc_md6_tmo; /* master down timeout */ 152 153 LIST_HEAD(__carp_mchead, carp_mc_entry) carp_mc_listhead; 154 }; 155 156 int carp_suppress_preempt = 0; 157 static int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 0, 0 }; /* XXX for now */ 158 159 static percpu_t *carpstat_percpu; 160 161 #define CARP_STATINC(x) _NET_STATINC(carpstat_percpu, x) 162 163 #ifdef MBUFTRACE 164 static struct mowner carp_proto_mowner_rx = MOWNER_INIT("carp", "rx"); 165 static struct mowner carp_proto_mowner_tx = MOWNER_INIT("carp", "tx"); 166 static struct mowner carp_proto6_mowner_rx = MOWNER_INIT("carp6", "rx"); 167 static struct mowner carp_proto6_mowner_tx = MOWNER_INIT("carp6", "tx"); 168 #endif 169 170 struct carp_if { 171 TAILQ_HEAD(, carp_softc) vhif_vrs; 172 int vhif_nvrs; 173 174 struct ifnet *vhif_ifp; 175 }; 176 177 #define CARP_LOG(sc, s) \ 178 if (carp_opts[CARPCTL_LOG]) { \ 179 if (sc) \ 180 log(LOG_INFO, "%s: ", \ 181 (sc)->sc_if.if_xname); \ 182 else \ 183 log(LOG_INFO, "carp: "); \ 184 addlog s; \ 185 addlog("\n"); \ 186 } 187 188 static void carp_hmac_prepare(struct carp_softc *); 189 static void carp_hmac_generate(struct carp_softc *, u_int32_t *, 190 unsigned char *); 191 static int carp_hmac_verify(struct carp_softc *, u_int32_t *, 192 unsigned char *); 193 static void carp_setroute(struct carp_softc *, int); 194 static void carp_proto_input_c(struct mbuf *, struct carp_header *, 195 sa_family_t); 196 static void carpdetach(struct carp_softc *); 197 static void carp_prepare_ad(struct mbuf *, struct carp_softc *, 198 struct carp_header *); 199 static void carp_send_ad_all(void); 200 static void carp_send_ad(void *); 201 static void carp_send_arp(struct carp_softc *); 202 static void carp_master_down(void *); 203 static int carp_ioctl(struct ifnet *, u_long, void *); 204 static void carp_start(struct ifnet *); 205 static void carp_setrun(struct carp_softc *, sa_family_t); 206 static void carp_set_state(struct carp_softc *, int); 207 static int carp_addrcount(struct carp_if *, struct in_ifaddr *, int); 208 enum { CARP_COUNT_MASTER, CARP_COUNT_RUNNING }; 209 210 static void carp_multicast_cleanup(struct carp_softc *); 211 static int carp_set_ifp(struct carp_softc *, struct ifnet *); 212 static void carp_set_enaddr(struct carp_softc *); 213 #if 0 214 static void carp_addr_updated(void *); 215 #endif 216 static u_int32_t carp_hash(struct carp_softc *, u_char *); 217 static int carp_set_addr(struct carp_softc *, struct sockaddr_in *); 218 static int carp_join_multicast(struct carp_softc *); 219 #ifdef INET6 220 static void carp_send_na(struct carp_softc *); 221 static int carp_set_addr6(struct carp_softc *, struct sockaddr_in6 *); 222 static int carp_join_multicast6(struct carp_softc *); 223 #endif 224 static int carp_clone_create(struct if_clone *, int); 225 static int carp_clone_destroy(struct ifnet *); 226 static int carp_ether_addmulti(struct carp_softc *, struct ifreq *); 227 static int carp_ether_delmulti(struct carp_softc *, struct ifreq *); 228 static void carp_ether_purgemulti(struct carp_softc *); 229 static void carp_update_link_state(struct carp_softc *sc); 230 231 static void sysctl_net_inet_carp_setup(struct sysctllog **); 232 233 /* workqueue-based pr_input */ 234 static struct wqinput *carp_wqinput; 235 static void _carp_proto_input(struct mbuf *, int, int); 236 #ifdef INET6 237 static struct wqinput *carp6_wqinput; 238 static void _carp6_proto_input(struct mbuf *, int, int); 239 #endif 240 241 struct if_clone carp_cloner = 242 IF_CLONE_INITIALIZER("carp", carp_clone_create, carp_clone_destroy); 243 244 static __inline u_int16_t 245 carp_cksum(struct mbuf *m, int len) 246 { 247 return (in_cksum(m, len)); 248 } 249 250 #ifdef INET6 251 static __inline u_int16_t 252 carp6_cksum(struct mbuf *m, uint32_t off, uint32_t len) 253 { 254 return (in6_cksum(m, IPPROTO_CARP, off, len)); 255 } 256 #endif 257 258 static void 259 carp_hmac_prepare(struct carp_softc *sc) 260 { 261 u_int8_t carp_version = CARP_VERSION, type = CARP_ADVERTISEMENT; 262 u_int8_t vhid = sc->sc_vhid & 0xff; 263 SHA1_CTX sha1ctx; 264 u_int32_t kmd[5]; 265 struct ifaddr *ifa; 266 int i, found; 267 struct in_addr last, cur, in; 268 #ifdef INET6 269 struct in6_addr last6, cur6, in6; 270 #endif /* INET6 */ 271 272 /* compute ipad from key */ 273 memset(sc->sc_pad, 0, sizeof(sc->sc_pad)); 274 memcpy(sc->sc_pad, sc->sc_key, sizeof(sc->sc_key)); 275 for (i = 0; i < sizeof(sc->sc_pad); i++) 276 sc->sc_pad[i] ^= 0x36; 277 278 /* precompute first part of inner hash */ 279 SHA1Init(&sc->sc_sha1); 280 SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad)); 281 SHA1Update(&sc->sc_sha1, (void *)&carp_version, sizeof(carp_version)); 282 SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type)); 283 284 /* generate a key for the arpbalance hash, before the vhid is hashed */ 285 memcpy(&sha1ctx, &sc->sc_sha1, sizeof(sha1ctx)); 286 SHA1Final((unsigned char *)kmd, &sha1ctx); 287 sc->sc_hashkey[0] = kmd[0] ^ kmd[1]; 288 sc->sc_hashkey[1] = kmd[2] ^ kmd[3]; 289 290 /* the rest of the precomputation */ 291 SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid)); 292 293 /* Hash the addresses from smallest to largest, not interface order */ 294 #ifdef INET 295 cur.s_addr = 0; 296 do { 297 int s; 298 found = 0; 299 last = cur; 300 cur.s_addr = 0xffffffff; 301 s = pserialize_read_enter(); 302 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 303 in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr; 304 if (ifa->ifa_addr->sa_family == AF_INET && 305 ntohl(in.s_addr) > ntohl(last.s_addr) && 306 ntohl(in.s_addr) < ntohl(cur.s_addr)) { 307 cur.s_addr = in.s_addr; 308 found++; 309 } 310 } 311 pserialize_read_exit(s); 312 if (found) 313 SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur)); 314 } while (found); 315 #endif /* INET */ 316 317 #ifdef INET6 318 memset(&cur6, 0x00, sizeof(cur6)); 319 do { 320 int s; 321 found = 0; 322 last6 = cur6; 323 memset(&cur6, 0xff, sizeof(cur6)); 324 s = pserialize_read_enter(); 325 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 326 in6 = ifatoia6(ifa)->ia_addr.sin6_addr; 327 if (IN6_IS_ADDR_LINKLOCAL(&in6)) 328 in6.s6_addr16[1] = 0; 329 if (ifa->ifa_addr->sa_family == AF_INET6 && 330 memcmp(&in6, &last6, sizeof(in6)) > 0 && 331 memcmp(&in6, &cur6, sizeof(in6)) < 0) { 332 cur6 = in6; 333 found++; 334 } 335 } 336 pserialize_read_exit(s); 337 if (found) 338 SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6)); 339 } while (found); 340 #endif /* INET6 */ 341 342 /* convert ipad to opad */ 343 for (i = 0; i < sizeof(sc->sc_pad); i++) 344 sc->sc_pad[i] ^= 0x36 ^ 0x5c; 345 } 346 347 static void 348 carp_hmac_generate(struct carp_softc *sc, u_int32_t counter[2], 349 unsigned char md[20]) 350 { 351 SHA1_CTX sha1ctx; 352 353 /* fetch first half of inner hash */ 354 memcpy(&sha1ctx, &sc->sc_sha1, sizeof(sha1ctx)); 355 356 SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter)); 357 SHA1Final(md, &sha1ctx); 358 359 /* outer hash */ 360 SHA1Init(&sha1ctx); 361 SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad)); 362 SHA1Update(&sha1ctx, md, 20); 363 SHA1Final(md, &sha1ctx); 364 } 365 366 static int 367 carp_hmac_verify(struct carp_softc *sc, u_int32_t counter[2], 368 unsigned char md[20]) 369 { 370 unsigned char md2[20]; 371 372 carp_hmac_generate(sc, counter, md2); 373 374 return (memcmp(md, md2, sizeof(md2))); 375 } 376 377 static void 378 carp_setroute(struct carp_softc *sc, int cmd) 379 { 380 struct ifaddr *ifa; 381 int s, bound; 382 383 KERNEL_LOCK(1, NULL); 384 bound = curlwp_bind(); 385 s = pserialize_read_enter(); 386 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 387 struct psref psref; 388 ifa_acquire(ifa, &psref); 389 pserialize_read_exit(s); 390 391 switch (ifa->ifa_addr->sa_family) { 392 case AF_INET: { 393 int count = 0; 394 struct rtentry *rt; 395 int hr_otherif, nr_ourif; 396 397 /* 398 * Avoid screwing with the routes if there are other 399 * carp interfaces which are master and have the same 400 * address. 401 */ 402 if (sc->sc_carpdev != NULL && 403 sc->sc_carpdev->if_carp != NULL) { 404 count = carp_addrcount( 405 (struct carp_if *)sc->sc_carpdev->if_carp, 406 ifatoia(ifa), CARP_COUNT_MASTER); 407 if ((cmd == RTM_ADD && count != 1) || 408 (cmd == RTM_DELETE && count != 0)) 409 goto next; 410 } 411 412 /* Remove the existing host route, if any */ 413 rtrequest(RTM_DELETE, ifa->ifa_addr, 414 ifa->ifa_addr, ifa->ifa_netmask, 415 RTF_HOST, NULL); 416 417 rt = NULL; 418 (void)rtrequest(RTM_GET, ifa->ifa_addr, ifa->ifa_addr, 419 ifa->ifa_netmask, RTF_HOST, &rt); 420 hr_otherif = (rt && rt->rt_ifp != &sc->sc_if && 421 (rt->rt_flags & RTF_CONNECTED)); 422 if (rt != NULL) { 423 rt_unref(rt); 424 rt = NULL; 425 } 426 427 /* Check for a network route on our interface */ 428 429 rt = NULL; 430 (void)rtrequest(RTM_GET, ifa->ifa_addr, ifa->ifa_addr, 431 ifa->ifa_netmask, 0, &rt); 432 nr_ourif = (rt && rt->rt_ifp == &sc->sc_if); 433 434 switch (cmd) { 435 case RTM_ADD: 436 if (hr_otherif) { 437 ifa->ifa_rtrequest = NULL; 438 ifa->ifa_flags &= ~RTF_CONNECTED; 439 440 rtrequest(RTM_ADD, ifa->ifa_addr, 441 ifa->ifa_addr, ifa->ifa_netmask, 442 RTF_UP | RTF_HOST, NULL); 443 } 444 if (!hr_otherif || nr_ourif || !rt) { 445 if (nr_ourif && 446 (rt->rt_flags & RTF_CONNECTED) == 0) 447 rtrequest(RTM_DELETE, 448 ifa->ifa_addr, 449 ifa->ifa_addr, 450 ifa->ifa_netmask, 0, NULL); 451 452 ifa->ifa_rtrequest = arp_rtrequest; 453 ifa->ifa_flags |= RTF_CONNECTED; 454 455 if (rtrequest(RTM_ADD, ifa->ifa_addr, 456 ifa->ifa_addr, ifa->ifa_netmask, 0, 457 NULL) == 0) 458 ifa->ifa_flags |= IFA_ROUTE; 459 } 460 break; 461 case RTM_DELETE: 462 break; 463 default: 464 break; 465 } 466 if (rt != NULL) { 467 rt_unref(rt); 468 rt = NULL; 469 } 470 break; 471 } 472 473 #ifdef INET6 474 case AF_INET6: 475 if (cmd == RTM_ADD) 476 in6_ifaddlocal(ifa); 477 else 478 in6_ifremlocal(ifa); 479 break; 480 #endif /* INET6 */ 481 default: 482 break; 483 } 484 next: 485 s = pserialize_read_enter(); 486 ifa_release(ifa, &psref); 487 } 488 pserialize_read_exit(s); 489 curlwp_bindx(bound); 490 KERNEL_UNLOCK_ONE(NULL); 491 } 492 493 /* 494 * process input packet. 495 * we have rearranged checks order compared to the rfc, 496 * but it seems more efficient this way or not possible otherwise. 497 */ 498 static void 499 _carp_proto_input(struct mbuf *m, int hlen, int proto) 500 { 501 struct ip *ip = mtod(m, struct ip *); 502 struct carp_softc *sc = NULL; 503 struct carp_header *ch; 504 int iplen, len; 505 struct ifnet *rcvif; 506 507 CARP_STATINC(CARP_STAT_IPACKETS); 508 MCLAIM(m, &carp_proto_mowner_rx); 509 510 if (!carp_opts[CARPCTL_ALLOW]) { 511 m_freem(m); 512 return; 513 } 514 515 rcvif = m_get_rcvif_NOMPSAFE(m); 516 /* check if received on a valid carp interface */ 517 if (rcvif->if_type != IFT_CARP) { 518 CARP_STATINC(CARP_STAT_BADIF); 519 CARP_LOG(sc, ("packet received on non-carp interface: %s", 520 rcvif->if_xname)); 521 m_freem(m); 522 return; 523 } 524 525 /* verify that the IP TTL is 255. */ 526 if (ip->ip_ttl != CARP_DFLTTL) { 527 CARP_STATINC(CARP_STAT_BADTTL); 528 CARP_LOG(sc, ("received ttl %d != %d on %s", ip->ip_ttl, 529 CARP_DFLTTL, rcvif->if_xname)); 530 m_freem(m); 531 return; 532 } 533 534 /* 535 * verify that the received packet length is 536 * equal to the CARP header 537 */ 538 iplen = ip->ip_hl << 2; 539 len = iplen + sizeof(*ch); 540 if (len > m->m_pkthdr.len) { 541 CARP_STATINC(CARP_STAT_BADLEN); 542 CARP_LOG(sc, ("packet too short %d on %s", m->m_pkthdr.len, 543 rcvif->if_xname)); 544 m_freem(m); 545 return; 546 } 547 548 if ((m = m_pullup(m, len)) == NULL) { 549 CARP_STATINC(CARP_STAT_HDROPS); 550 return; 551 } 552 ip = mtod(m, struct ip *); 553 ch = (struct carp_header *)((char *)ip + iplen); 554 /* verify the CARP checksum */ 555 m->m_data += iplen; 556 if (carp_cksum(m, len - iplen)) { 557 CARP_STATINC(CARP_STAT_BADSUM); 558 CARP_LOG(sc, ("checksum failed on %s", 559 rcvif->if_xname)); 560 m_freem(m); 561 return; 562 } 563 m->m_data -= iplen; 564 565 carp_proto_input_c(m, ch, AF_INET); 566 } 567 568 void 569 carp_proto_input(struct mbuf *m, int off, int proto) 570 { 571 572 wqinput_input(carp_wqinput, m, 0, 0); 573 } 574 575 #ifdef INET6 576 static void 577 _carp6_proto_input(struct mbuf *m, int off, int proto) 578 { 579 struct carp_softc *sc = NULL; 580 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 581 struct carp_header *ch; 582 u_int len; 583 struct ifnet *rcvif; 584 585 CARP_STATINC(CARP_STAT_IPACKETS6); 586 MCLAIM(m, &carp_proto6_mowner_rx); 587 588 if (!carp_opts[CARPCTL_ALLOW]) { 589 m_freem(m); 590 return; 591 } 592 593 rcvif = m_get_rcvif_NOMPSAFE(m); 594 595 /* check if received on a valid carp interface */ 596 if (rcvif->if_type != IFT_CARP) { 597 CARP_STATINC(CARP_STAT_BADIF); 598 CARP_LOG(sc, ("packet received on non-carp interface: %s", 599 rcvif->if_xname)); 600 m_freem(m); 601 return; 602 } 603 604 /* verify that the IP TTL is 255 */ 605 if (ip6->ip6_hlim != CARP_DFLTTL) { 606 CARP_STATINC(CARP_STAT_BADTTL); 607 CARP_LOG(sc, ("received ttl %d != %d on %s", ip6->ip6_hlim, 608 CARP_DFLTTL, rcvif->if_xname)); 609 m_freem(m); 610 return; 611 } 612 613 /* verify that we have a complete carp packet */ 614 len = m->m_len; 615 M_REGION_GET(ch, struct carp_header *, m, off, sizeof(*ch)); 616 if (ch == NULL) { 617 CARP_STATINC(CARP_STAT_BADLEN); 618 CARP_LOG(sc, ("packet size %u too small", len)); 619 return; 620 } 621 622 /* verify the CARP checksum */ 623 if (carp6_cksum(m, off, sizeof(*ch))) { 624 CARP_STATINC(CARP_STAT_BADSUM); 625 CARP_LOG(sc, ("checksum failed, on %s", rcvif->if_xname)); 626 m_freem(m); 627 return; 628 } 629 630 carp_proto_input_c(m, ch, AF_INET6); 631 return; 632 } 633 634 int 635 carp6_proto_input(struct mbuf **mp, int *offp, int proto) 636 { 637 638 wqinput_input(carp6_wqinput, *mp, *offp, proto); 639 640 return IPPROTO_DONE; 641 } 642 #endif /* INET6 */ 643 644 static void 645 carp_proto_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af) 646 { 647 struct carp_softc *sc; 648 u_int64_t tmp_counter; 649 struct timeval sc_tv, ch_tv; 650 651 TAILQ_FOREACH(sc, &((struct carp_if *) 652 m_get_rcvif_NOMPSAFE(m)->if_carpdev->if_carp)->vhif_vrs, sc_list) 653 if (sc->sc_vhid == ch->carp_vhid) 654 break; 655 656 if (!sc || (sc->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) != 657 (IFF_UP|IFF_RUNNING)) { 658 CARP_STATINC(CARP_STAT_BADVHID); 659 m_freem(m); 660 return; 661 } 662 663 /* 664 * Check if our own advertisement was duplicated 665 * from a non simplex interface. 666 * XXX If there is no address on our physical interface 667 * there is no way to distinguish our ads from the ones 668 * another carp host might have sent us. 669 */ 670 if ((sc->sc_carpdev->if_flags & IFF_SIMPLEX) == 0) { 671 struct sockaddr sa; 672 struct ifaddr *ifa; 673 int s; 674 675 memset(&sa, 0, sizeof(sa)); 676 sa.sa_family = af; 677 s = pserialize_read_enter(); 678 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev); 679 680 if (ifa && af == AF_INET) { 681 struct ip *ip = mtod(m, struct ip *); 682 if (ip->ip_src.s_addr == 683 ifatoia(ifa)->ia_addr.sin_addr.s_addr) { 684 pserialize_read_exit(s); 685 m_freem(m); 686 return; 687 } 688 } 689 #ifdef INET6 690 if (ifa && af == AF_INET6) { 691 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 692 struct in6_addr in6_src, in6_found; 693 694 in6_src = ip6->ip6_src; 695 in6_found = ifatoia6(ifa)->ia_addr.sin6_addr; 696 if (IN6_IS_ADDR_LINKLOCAL(&in6_src)) 697 in6_src.s6_addr16[1] = 0; 698 if (IN6_IS_ADDR_LINKLOCAL(&in6_found)) 699 in6_found.s6_addr16[1] = 0; 700 if (IN6_ARE_ADDR_EQUAL(&in6_src, &in6_found)) { 701 pserialize_read_exit(s); 702 m_freem(m); 703 return; 704 } 705 } 706 #endif /* INET6 */ 707 pserialize_read_exit(s); 708 } 709 710 nanotime(&sc->sc_if.if_lastchange); 711 if_statadd2(&sc->sc_if, if_ipackets, 1, if_ibytes, m->m_pkthdr.len); 712 713 /* verify the CARP version. */ 714 if (ch->carp_version != CARP_VERSION) { 715 CARP_STATINC(CARP_STAT_BADVER); 716 if_statinc(&sc->sc_if, if_ierrors); 717 CARP_LOG(sc, ("invalid version %d != %d", 718 ch->carp_version, CARP_VERSION)); 719 m_freem(m); 720 return; 721 } 722 723 /* verify the hash */ 724 if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) { 725 struct ip *ip; 726 char ipbuf[INET_ADDRSTRLEN]; 727 #ifdef INET6 728 struct ip6_hdr *ip6; 729 char ip6buf[INET6_ADDRSTRLEN]; 730 #endif 731 732 CARP_STATINC(CARP_STAT_BADAUTH); 733 if_statinc(&sc->sc_if, if_ierrors); 734 735 switch(af) { 736 case AF_INET: 737 ip = mtod(m, struct ip *); 738 CARP_LOG(sc, ("incorrect hash from %s", 739 IN_PRINT(ipbuf, &ip->ip_src))); 740 break; 741 742 #ifdef INET6 743 case AF_INET6: 744 ip6 = mtod(m, struct ip6_hdr *); 745 CARP_LOG(sc, ("incorrect hash from %s", 746 IN6_PRINT(ip6buf, &ip6->ip6_src))); 747 break; 748 #endif 749 750 default: CARP_LOG(sc, ("incorrect hash")); 751 break; 752 } 753 m_freem(m); 754 return; 755 } 756 757 tmp_counter = ntohl(ch->carp_counter[0]); 758 tmp_counter = tmp_counter<<32; 759 tmp_counter += ntohl(ch->carp_counter[1]); 760 761 /* XXX Replay protection goes here */ 762 763 sc->sc_init_counter = 0; 764 sc->sc_counter = tmp_counter; 765 766 767 sc_tv.tv_sec = sc->sc_advbase; 768 if (carp_suppress_preempt && sc->sc_advskew < 240) 769 sc_tv.tv_usec = 240 * 1000000 / 256; 770 else 771 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256; 772 ch_tv.tv_sec = ch->carp_advbase; 773 ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256; 774 775 switch (sc->sc_state) { 776 case INIT: 777 break; 778 case MASTER: 779 /* 780 * If we receive an advertisement from a backup who's going to 781 * be more frequent than us, go into BACKUP state. 782 */ 783 if (timercmp(&sc_tv, &ch_tv, >) || 784 timercmp(&sc_tv, &ch_tv, ==)) { 785 callout_stop(&sc->sc_ad_tmo); 786 CARP_LOG(sc, ("MASTER -> BACKUP (more frequent advertisement received)")); 787 carp_set_state(sc, BACKUP); 788 carp_setrun(sc, 0); 789 carp_setroute(sc, RTM_DELETE); 790 } 791 break; 792 case BACKUP: 793 /* 794 * If we're pre-empting masters who advertise slower than us, 795 * and this one claims to be slower, treat him as down. 796 */ 797 if (carp_opts[CARPCTL_PREEMPT] && timercmp(&sc_tv, &ch_tv, <)) { 798 CARP_LOG(sc, ("BACKUP -> MASTER (preempting a slower master)")); 799 carp_master_down(sc); 800 break; 801 } 802 803 /* 804 * If the master is going to advertise at such a low frequency 805 * that he's guaranteed to time out, we'd might as well just 806 * treat him as timed out now. 807 */ 808 sc_tv.tv_sec = sc->sc_advbase * 3; 809 if (timercmp(&sc_tv, &ch_tv, <)) { 810 CARP_LOG(sc, ("BACKUP -> MASTER (master timed out)")); 811 carp_master_down(sc); 812 break; 813 } 814 815 /* 816 * Otherwise, we reset the counter and wait for the next 817 * advertisement. 818 */ 819 carp_setrun(sc, af); 820 break; 821 } 822 823 m_freem(m); 824 return; 825 } 826 827 /* 828 * Interface side of the CARP implementation. 829 */ 830 831 /* ARGSUSED */ 832 void 833 carpattach(int n) 834 { 835 if_clone_attach(&carp_cloner); 836 837 carpstat_percpu = percpu_alloc(sizeof(uint64_t) * CARP_NSTATS); 838 } 839 840 static int 841 carp_clone_create(struct if_clone *ifc, int unit) 842 { 843 extern int ifqmaxlen; 844 struct carp_softc *sc; 845 struct ifnet *ifp; 846 847 sc = malloc(sizeof(*sc), M_DEVBUF, M_NOWAIT|M_ZERO); 848 if (!sc) 849 return (ENOMEM); 850 851 sc->sc_suppress = 0; 852 sc->sc_advbase = CARP_DFLTINTV; 853 sc->sc_vhid = -1; /* required setting */ 854 sc->sc_advskew = 0; 855 sc->sc_init_counter = 1; 856 sc->sc_naddrs = sc->sc_naddrs6 = 0; 857 #ifdef INET6 858 sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL; 859 #endif /* INET6 */ 860 861 callout_init(&sc->sc_ad_tmo, 0); 862 callout_init(&sc->sc_md_tmo, 0); 863 callout_init(&sc->sc_md6_tmo, 0); 864 865 callout_setfunc(&sc->sc_ad_tmo, carp_send_ad, sc); 866 callout_setfunc(&sc->sc_md_tmo, carp_master_down, sc); 867 callout_setfunc(&sc->sc_md6_tmo, carp_master_down, sc); 868 869 LIST_INIT(&sc->carp_mc_listhead); 870 ifp = &sc->sc_if; 871 ifp->if_softc = sc; 872 snprintf(ifp->if_xname, sizeof ifp->if_xname, "%s%d", ifc->ifc_name, 873 unit); 874 ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST; 875 ifp->if_ioctl = carp_ioctl; 876 ifp->if_start = carp_start; 877 IFQ_SET_MAXLEN(&ifp->if_snd, ifqmaxlen); 878 IFQ_SET_READY(&ifp->if_snd); 879 if_initialize(ifp); 880 ether_ifattach(ifp, NULL); 881 /* Overwrite ethernet defaults */ 882 ifp->if_type = IFT_CARP; 883 ifp->if_output = carp_output; 884 ifp->if_link_state = LINK_STATE_DOWN; 885 carp_set_enaddr(sc); 886 if_register(ifp); 887 888 return (0); 889 } 890 891 static int 892 carp_clone_destroy(struct ifnet *ifp) 893 { 894 struct carp_softc *sc = ifp->if_softc; 895 896 carpdetach(ifp->if_softc); 897 ether_ifdetach(ifp); 898 if_detach(ifp); 899 callout_destroy(&sc->sc_ad_tmo); 900 callout_destroy(&sc->sc_md_tmo); 901 callout_destroy(&sc->sc_md6_tmo); 902 free(ifp->if_softc, M_DEVBUF); 903 904 return (0); 905 } 906 907 static void 908 carpdetach(struct carp_softc *sc) 909 { 910 struct carp_if *cif; 911 int s; 912 913 callout_stop(&sc->sc_ad_tmo); 914 callout_stop(&sc->sc_md_tmo); 915 callout_stop(&sc->sc_md6_tmo); 916 917 if (sc->sc_suppress) 918 carp_suppress_preempt--; 919 sc->sc_suppress = 0; 920 921 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) 922 carp_suppress_preempt--; 923 sc->sc_sendad_errors = 0; 924 925 carp_set_state(sc, INIT); 926 sc->sc_if.if_flags &= ~IFF_UP; 927 carp_setrun(sc, 0); 928 carp_multicast_cleanup(sc); 929 930 KERNEL_LOCK(1, NULL); 931 s = splnet(); 932 if (sc->sc_carpdev != NULL) { 933 /* XXX linkstatehook removal */ 934 cif = (struct carp_if *)sc->sc_carpdev->if_carp; 935 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list); 936 if (!--cif->vhif_nvrs) { 937 ifpromisc(sc->sc_carpdev, 0); 938 sc->sc_carpdev->if_carp = NULL; 939 free(cif, M_IFADDR); 940 } 941 } 942 sc->sc_carpdev = NULL; 943 splx(s); 944 KERNEL_UNLOCK_ONE(NULL); 945 } 946 947 /* Detach an interface from the carp. */ 948 void 949 carp_ifdetach(struct ifnet *ifp) 950 { 951 struct carp_softc *sc, *nextsc; 952 struct carp_if *cif = (struct carp_if *)ifp->if_carp; 953 954 for (sc = TAILQ_FIRST(&cif->vhif_vrs); sc; sc = nextsc) { 955 nextsc = TAILQ_NEXT(sc, sc_list); 956 carpdetach(sc); 957 } 958 } 959 960 static void 961 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, 962 struct carp_header *ch) 963 { 964 if (sc->sc_init_counter) { 965 /* this could also be seconds since unix epoch */ 966 sc->sc_counter = cprng_fast64(); 967 } else 968 sc->sc_counter++; 969 970 ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff); 971 ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff); 972 973 carp_hmac_generate(sc, ch->carp_counter, ch->carp_md); 974 } 975 976 static void 977 carp_send_ad_all(void) 978 { 979 struct ifnet *ifp; 980 struct carp_if *cif; 981 struct carp_softc *vh; 982 int s; 983 int bound = curlwp_bind(); 984 985 s = pserialize_read_enter(); 986 IFNET_READER_FOREACH(ifp) { 987 struct psref psref; 988 if (ifp->if_carp == NULL || ifp->if_type == IFT_CARP) 989 continue; 990 991 if_acquire(ifp, &psref); 992 pserialize_read_exit(s); 993 994 cif = (struct carp_if *)ifp->if_carp; 995 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 996 if ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 997 (IFF_UP|IFF_RUNNING) && vh->sc_state == MASTER) 998 carp_send_ad(vh); 999 } 1000 1001 s = pserialize_read_enter(); 1002 if_release(ifp, &psref); 1003 } 1004 pserialize_read_exit(s); 1005 curlwp_bindx(bound); 1006 } 1007 1008 1009 static void 1010 carp_send_ad(void *v) 1011 { 1012 struct carp_header ch; 1013 struct timeval tv; 1014 struct carp_softc *sc = v; 1015 struct carp_header *ch_ptr; 1016 struct mbuf *m; 1017 int error, len, advbase, advskew, s; 1018 struct sockaddr sa; 1019 1020 KERNEL_LOCK(1, NULL); 1021 s = splsoftnet(); 1022 1023 advbase = advskew = 0; /* Sssssh compiler */ 1024 if (sc->sc_carpdev == NULL) { 1025 if_statinc(&sc->sc_if, if_oerrors); 1026 goto retry_later; 1027 } 1028 1029 /* bow out if we've gone to backup (the carp interface is going down) */ 1030 if (sc->sc_bow_out) { 1031 sc->sc_bow_out = 0; 1032 advbase = 255; 1033 advskew = 255; 1034 } else { 1035 advbase = sc->sc_advbase; 1036 if (!carp_suppress_preempt || sc->sc_advskew > 240) 1037 advskew = sc->sc_advskew; 1038 else 1039 advskew = 240; 1040 tv.tv_sec = advbase; 1041 tv.tv_usec = advskew * 1000000 / 256; 1042 } 1043 1044 ch.carp_version = CARP_VERSION; 1045 ch.carp_type = CARP_ADVERTISEMENT; 1046 ch.carp_vhid = sc->sc_vhid; 1047 ch.carp_advbase = advbase; 1048 ch.carp_advskew = advskew; 1049 ch.carp_authlen = 7; /* XXX DEFINE */ 1050 ch.carp_pad1 = 0; /* must be zero */ 1051 ch.carp_cksum = 0; 1052 1053 1054 #ifdef INET 1055 if (sc->sc_naddrs) { 1056 struct ip *ip; 1057 struct ifaddr *ifa; 1058 int _s; 1059 1060 MGETHDR(m, M_DONTWAIT, MT_HEADER); 1061 if (m == NULL) { 1062 if_statinc(&sc->sc_if, if_oerrors); 1063 CARP_STATINC(CARP_STAT_ONOMEM); 1064 /* XXX maybe less ? */ 1065 goto retry_later; 1066 } 1067 MCLAIM(m, &carp_proto_mowner_tx); 1068 len = sizeof(*ip) + sizeof(ch); 1069 m->m_pkthdr.len = len; 1070 m_reset_rcvif(m); 1071 m->m_len = len; 1072 m_align(m, m->m_len); 1073 m->m_flags |= M_MCAST; 1074 ip = mtod(m, struct ip *); 1075 ip->ip_v = IPVERSION; 1076 ip->ip_hl = sizeof(*ip) >> 2; 1077 ip->ip_tos = IPTOS_LOWDELAY; 1078 ip->ip_len = htons(len); 1079 ip->ip_id = 0; /* no need for id, we don't support fragments */ 1080 ip->ip_off = htons(IP_DF); 1081 ip->ip_ttl = CARP_DFLTTL; 1082 ip->ip_p = IPPROTO_CARP; 1083 ip->ip_sum = 0; 1084 1085 memset(&sa, 0, sizeof(sa)); 1086 sa.sa_family = AF_INET; 1087 _s = pserialize_read_enter(); 1088 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev); 1089 if (ifa == NULL) 1090 ip->ip_src.s_addr = 0; 1091 else 1092 ip->ip_src.s_addr = 1093 ifatoia(ifa)->ia_addr.sin_addr.s_addr; 1094 pserialize_read_exit(_s); 1095 ip->ip_dst.s_addr = INADDR_CARP_GROUP; 1096 1097 ch_ptr = (struct carp_header *)(&ip[1]); 1098 memcpy(ch_ptr, &ch, sizeof(ch)); 1099 carp_prepare_ad(m, sc, ch_ptr); 1100 1101 m->m_data += sizeof(*ip); 1102 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip)); 1103 m->m_data -= sizeof(*ip); 1104 1105 nanotime(&sc->sc_if.if_lastchange); 1106 if_statadd2(&sc->sc_if, if_opackets, 1, if_obytes, len); 1107 CARP_STATINC(CARP_STAT_OPACKETS); 1108 1109 error = ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, 1110 NULL); 1111 if (error) { 1112 if (error == ENOBUFS) 1113 CARP_STATINC(CARP_STAT_ONOMEM); 1114 else 1115 CARP_LOG(sc, ("ip_output failed: %d", error)); 1116 if_statinc(&sc->sc_if, if_oerrors); 1117 if (sc->sc_sendad_errors < INT_MAX) 1118 sc->sc_sendad_errors++; 1119 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) { 1120 carp_suppress_preempt++; 1121 if (carp_suppress_preempt == 1) 1122 carp_send_ad_all(); 1123 } 1124 sc->sc_sendad_success = 0; 1125 } else { 1126 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) { 1127 if (++sc->sc_sendad_success >= 1128 CARP_SENDAD_MIN_SUCCESS) { 1129 carp_suppress_preempt--; 1130 sc->sc_sendad_errors = 0; 1131 } 1132 } else 1133 sc->sc_sendad_errors = 0; 1134 } 1135 } 1136 #endif /* INET */ 1137 #ifdef INET6 1138 if (sc->sc_naddrs6) { 1139 struct ip6_hdr *ip6; 1140 struct ifaddr *ifa; 1141 int _s; 1142 1143 MGETHDR(m, M_DONTWAIT, MT_HEADER); 1144 if (m == NULL) { 1145 if_statinc(&sc->sc_if, if_oerrors); 1146 CARP_STATINC(CARP_STAT_ONOMEM); 1147 /* XXX maybe less ? */ 1148 goto retry_later; 1149 } 1150 MCLAIM(m, &carp_proto6_mowner_tx); 1151 len = sizeof(*ip6) + sizeof(ch); 1152 m->m_pkthdr.len = len; 1153 m_reset_rcvif(m); 1154 m->m_len = len; 1155 m_align(m, m->m_len); 1156 m->m_flags |= M_MCAST; 1157 ip6 = mtod(m, struct ip6_hdr *); 1158 memset(ip6, 0, sizeof(*ip6)); 1159 ip6->ip6_vfc |= IPV6_VERSION; 1160 ip6->ip6_hlim = CARP_DFLTTL; 1161 ip6->ip6_nxt = IPPROTO_CARP; 1162 1163 /* set the source address */ 1164 memset(&sa, 0, sizeof(sa)); 1165 sa.sa_family = AF_INET6; 1166 _s = pserialize_read_enter(); 1167 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev); 1168 if (ifa == NULL) /* This should never happen with IPv6 */ 1169 memset(&ip6->ip6_src, 0, sizeof(struct in6_addr)); 1170 else 1171 bcopy(ifatoia6(ifa)->ia_addr.sin6_addr.s6_addr, 1172 &ip6->ip6_src, sizeof(struct in6_addr)); 1173 pserialize_read_exit(_s); 1174 /* set the multicast destination */ 1175 1176 ip6->ip6_dst.s6_addr16[0] = htons(0xff02); 1177 ip6->ip6_dst.s6_addr8[15] = 0x12; 1178 if (in6_setscope(&ip6->ip6_dst, &sc->sc_if, NULL) != 0) { 1179 if_statinc(&sc->sc_if, if_oerrors); 1180 m_freem(m); 1181 CARP_LOG(sc, ("in6_setscope failed")); 1182 goto retry_later; 1183 } 1184 1185 ch_ptr = (struct carp_header *)(&ip6[1]); 1186 memcpy(ch_ptr, &ch, sizeof(ch)); 1187 carp_prepare_ad(m, sc, ch_ptr); 1188 1189 ch_ptr->carp_cksum = carp6_cksum(m, sizeof(*ip6), 1190 len - sizeof(*ip6)); 1191 1192 nanotime(&sc->sc_if.if_lastchange); 1193 if_statadd2(&sc->sc_if, if_opackets, 1, if_obytes, len); 1194 CARP_STATINC(CARP_STAT_OPACKETS6); 1195 1196 error = ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL); 1197 if (error) { 1198 if (error == ENOBUFS) 1199 CARP_STATINC(CARP_STAT_ONOMEM); 1200 else 1201 CARP_LOG(sc, ("ip6_output failed: %d", error)); 1202 if_statinc(&sc->sc_if, if_oerrors); 1203 if (sc->sc_sendad_errors < INT_MAX) 1204 sc->sc_sendad_errors++; 1205 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) { 1206 carp_suppress_preempt++; 1207 if (carp_suppress_preempt == 1) 1208 carp_send_ad_all(); 1209 } 1210 sc->sc_sendad_success = 0; 1211 } else { 1212 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) { 1213 if (++sc->sc_sendad_success >= 1214 CARP_SENDAD_MIN_SUCCESS) { 1215 carp_suppress_preempt--; 1216 sc->sc_sendad_errors = 0; 1217 } 1218 } else 1219 sc->sc_sendad_errors = 0; 1220 } 1221 } 1222 #endif /* INET6 */ 1223 1224 retry_later: 1225 splx(s); 1226 KERNEL_UNLOCK_ONE(NULL); 1227 if (advbase != 255 || advskew != 255) 1228 callout_schedule(&sc->sc_ad_tmo, tvtohz(&tv)); 1229 } 1230 1231 /* 1232 * Broadcast a gratuitous ARP request containing 1233 * the virtual router MAC address for each IP address 1234 * associated with the virtual router. 1235 */ 1236 static void 1237 carp_send_arp(struct carp_softc *sc) 1238 { 1239 struct ifaddr *ifa; 1240 int s, bound; 1241 1242 KERNEL_LOCK(1, NULL); 1243 bound = curlwp_bind(); 1244 s = pserialize_read_enter(); 1245 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 1246 struct psref psref; 1247 1248 if (ifa->ifa_addr->sa_family != AF_INET) 1249 continue; 1250 1251 ifa_acquire(ifa, &psref); 1252 pserialize_read_exit(s); 1253 1254 arpannounce(sc->sc_carpdev, ifa, CLLADDR(sc->sc_if.if_sadl)); 1255 1256 s = pserialize_read_enter(); 1257 ifa_release(ifa, &psref); 1258 } 1259 pserialize_read_exit(s); 1260 curlwp_bindx(bound); 1261 KERNEL_UNLOCK_ONE(NULL); 1262 } 1263 1264 #ifdef INET6 1265 static void 1266 carp_send_na(struct carp_softc *sc) 1267 { 1268 struct ifaddr *ifa; 1269 struct in6_addr *in6; 1270 static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT; 1271 int s, bound; 1272 1273 KERNEL_LOCK(1, NULL); 1274 bound = curlwp_bind(); 1275 s = pserialize_read_enter(); 1276 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 1277 struct psref psref; 1278 1279 if (ifa->ifa_addr->sa_family != AF_INET6) 1280 continue; 1281 1282 ifa_acquire(ifa, &psref); 1283 pserialize_read_exit(s); 1284 1285 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr; 1286 nd6_na_output(sc->sc_carpdev, &mcast, in6, 1287 ND_NA_FLAG_OVERRIDE, 1, NULL); 1288 1289 s = pserialize_read_enter(); 1290 ifa_release(ifa, &psref); 1291 } 1292 pserialize_read_exit(s); 1293 curlwp_bindx(bound); 1294 KERNEL_UNLOCK_ONE(NULL); 1295 } 1296 #endif /* INET6 */ 1297 1298 /* 1299 * Based on bridge_hash() in if_bridge.c 1300 */ 1301 #define mix(a,b,c) \ 1302 do { \ 1303 a -= b; a -= c; a ^= (c >> 13); \ 1304 b -= c; b -= a; b ^= (a << 8); \ 1305 c -= a; c -= b; c ^= (b >> 13); \ 1306 a -= b; a -= c; a ^= (c >> 12); \ 1307 b -= c; b -= a; b ^= (a << 16); \ 1308 c -= a; c -= b; c ^= (b >> 5); \ 1309 a -= b; a -= c; a ^= (c >> 3); \ 1310 b -= c; b -= a; b ^= (a << 10); \ 1311 c -= a; c -= b; c ^= (b >> 15); \ 1312 } while (0) 1313 1314 static u_int32_t 1315 carp_hash(struct carp_softc *sc, u_char *src) 1316 { 1317 u_int32_t a = 0x9e3779b9, b = sc->sc_hashkey[0], c = sc->sc_hashkey[1]; 1318 1319 c += sc->sc_key[3] << 24; 1320 c += sc->sc_key[2] << 16; 1321 c += sc->sc_key[1] << 8; 1322 c += sc->sc_key[0]; 1323 b += src[5] << 8; 1324 b += src[4]; 1325 a += src[3] << 24; 1326 a += src[2] << 16; 1327 a += src[1] << 8; 1328 a += src[0]; 1329 1330 mix(a, b, c); 1331 return (c); 1332 } 1333 1334 static int 1335 carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type) 1336 { 1337 struct carp_softc *vh; 1338 struct ifaddr *ifa; 1339 int count = 0; 1340 1341 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1342 if ((type == CARP_COUNT_RUNNING && 1343 (vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1344 (IFF_UP|IFF_RUNNING)) || 1345 (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) { 1346 int s = pserialize_read_enter(); 1347 IFADDR_READER_FOREACH(ifa, &vh->sc_if) { 1348 if (ifa->ifa_addr->sa_family == AF_INET && 1349 ia->ia_addr.sin_addr.s_addr == 1350 ifatoia(ifa)->ia_addr.sin_addr.s_addr) 1351 count++; 1352 } 1353 pserialize_read_exit(s); 1354 } 1355 } 1356 return (count); 1357 } 1358 1359 int 1360 carp_iamatch(struct in_ifaddr *ia, u_char *src, 1361 u_int32_t *count, u_int32_t index) 1362 { 1363 struct carp_softc *sc = ia->ia_ifp->if_softc; 1364 1365 if (carp_opts[CARPCTL_ARPBALANCE]) { 1366 /* 1367 * We use the source ip to decide which virtual host should 1368 * handle the request. If we're master of that virtual host, 1369 * then we respond, otherwise, just drop the arp packet on 1370 * the floor. 1371 */ 1372 1373 /* Count the elegible carp interfaces with this address */ 1374 if (*count == 0) 1375 *count = carp_addrcount( 1376 (struct carp_if *)ia->ia_ifp->if_carpdev->if_carp, 1377 ia, CARP_COUNT_RUNNING); 1378 1379 /* This should never happen, but... */ 1380 if (*count == 0) 1381 return (0); 1382 1383 if (carp_hash(sc, src) % *count == index - 1 && 1384 sc->sc_state == MASTER) { 1385 return (1); 1386 } 1387 } else { 1388 if (sc->sc_state == MASTER) 1389 return (1); 1390 } 1391 1392 return (0); 1393 } 1394 1395 #ifdef INET6 1396 struct ifaddr * 1397 carp_iamatch6(void *v, struct in6_addr *taddr) 1398 { 1399 struct carp_if *cif = v; 1400 struct carp_softc *vh; 1401 struct ifaddr *ifa; 1402 1403 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1404 int s = pserialize_read_enter(); 1405 IFADDR_READER_FOREACH(ifa, &vh->sc_if) { 1406 if (IN6_ARE_ADDR_EQUAL(taddr, 1407 &ifatoia6(ifa)->ia_addr.sin6_addr) && 1408 ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1409 (IFF_UP|IFF_RUNNING)) && vh->sc_state == MASTER) 1410 return (ifa); 1411 } 1412 pserialize_read_exit(s); 1413 } 1414 1415 return (NULL); 1416 } 1417 #endif /* INET6 */ 1418 1419 struct ifnet * 1420 carp_ourether(void *v, struct ether_header *eh, u_char iftype, int src) 1421 { 1422 struct carp_if *cif = (struct carp_if *)v; 1423 struct carp_softc *vh; 1424 u_int8_t *ena; 1425 1426 if (src) 1427 ena = (u_int8_t *)&eh->ether_shost; 1428 else 1429 ena = (u_int8_t *)&eh->ether_dhost; 1430 1431 switch (iftype) { 1432 case IFT_ETHER: 1433 case IFT_FDDI: 1434 if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1) 1435 return (NULL); 1436 break; 1437 case IFT_ISO88025: 1438 if (ena[0] != 3 || ena[1] || ena[4] || ena[5]) 1439 return (NULL); 1440 break; 1441 default: 1442 return (NULL); 1443 break; 1444 } 1445 1446 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) 1447 if ((vh->sc_if.if_flags & (IFF_UP|IFF_RUNNING)) == 1448 (IFF_UP|IFF_RUNNING) && vh->sc_state == MASTER && 1449 !memcmp(ena, CLLADDR(vh->sc_if.if_sadl), 1450 ETHER_ADDR_LEN)) { 1451 return (&vh->sc_if); 1452 } 1453 1454 return (NULL); 1455 } 1456 1457 int 1458 carp_input(struct mbuf *m, u_int8_t *shost, u_int8_t *dhost, u_int16_t etype) 1459 { 1460 struct ether_header eh; 1461 struct carp_if *cif = (struct carp_if *)m_get_rcvif_NOMPSAFE(m)->if_carp; 1462 struct ifnet *ifp; 1463 1464 memcpy(&eh.ether_shost, shost, sizeof(eh.ether_shost)); 1465 memcpy(&eh.ether_dhost, dhost, sizeof(eh.ether_dhost)); 1466 eh.ether_type = etype; 1467 1468 if (m->m_flags & (M_BCAST|M_MCAST)) { 1469 struct carp_softc *vh; 1470 struct mbuf *m0; 1471 1472 /* 1473 * XXX Should really check the list of multicast addresses 1474 * for each CARP interface _before_ copying. 1475 */ 1476 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) { 1477 m0 = m_copym(m, 0, M_COPYALL, M_DONTWAIT); 1478 if (m0 == NULL) 1479 continue; 1480 m_set_rcvif(m0, &vh->sc_if); 1481 ether_input(&vh->sc_if, m0); 1482 } 1483 return (1); 1484 } 1485 1486 ifp = carp_ourether(cif, &eh, m_get_rcvif_NOMPSAFE(m)->if_type, 0); 1487 if (ifp == NULL) { 1488 return (1); 1489 } 1490 1491 m_set_rcvif(m, ifp); 1492 1493 bpf_mtap(ifp, m, BPF_D_IN); 1494 if_statinc(ifp, if_ipackets); 1495 ether_input(ifp, m); 1496 return (0); 1497 } 1498 1499 static void 1500 carp_master_down(void *v) 1501 { 1502 struct carp_softc *sc = v; 1503 1504 switch (sc->sc_state) { 1505 case INIT: 1506 printf("%s: master_down event in INIT state\n", 1507 sc->sc_if.if_xname); 1508 break; 1509 case MASTER: 1510 break; 1511 case BACKUP: 1512 CARP_LOG(sc, ("INIT -> MASTER (preempting)")); 1513 carp_set_state(sc, MASTER); 1514 carp_send_ad(sc); 1515 carp_send_arp(sc); 1516 #ifdef INET6 1517 carp_send_na(sc); 1518 #endif /* INET6 */ 1519 carp_setrun(sc, 0); 1520 carp_setroute(sc, RTM_ADD); 1521 break; 1522 } 1523 } 1524 1525 /* 1526 * When in backup state, af indicates whether to reset the master down timer 1527 * for v4 or v6. If it's set to zero, reset the ones which are already pending. 1528 */ 1529 static void 1530 carp_setrun(struct carp_softc *sc, sa_family_t af) 1531 { 1532 struct timeval tv; 1533 1534 if (sc->sc_carpdev == NULL) { 1535 sc->sc_if.if_flags &= ~IFF_RUNNING; 1536 carp_set_state(sc, INIT); 1537 return; 1538 } 1539 1540 if (sc->sc_if.if_flags & IFF_UP && sc->sc_vhid > 0 && 1541 (sc->sc_naddrs || sc->sc_naddrs6) && !sc->sc_suppress) { 1542 sc->sc_if.if_flags |= IFF_RUNNING; 1543 } else { 1544 sc->sc_if.if_flags &= ~IFF_RUNNING; 1545 carp_setroute(sc, RTM_DELETE); 1546 return; 1547 } 1548 1549 switch (sc->sc_state) { 1550 case INIT: 1551 carp_set_state(sc, BACKUP); 1552 carp_setroute(sc, RTM_DELETE); 1553 carp_setrun(sc, 0); 1554 break; 1555 case BACKUP: 1556 callout_stop(&sc->sc_ad_tmo); 1557 tv.tv_sec = 3 * sc->sc_advbase; 1558 tv.tv_usec = sc->sc_advskew * 1000000 / 256; 1559 switch (af) { 1560 #ifdef INET 1561 case AF_INET: 1562 callout_schedule(&sc->sc_md_tmo, tvtohz(&tv)); 1563 break; 1564 #endif /* INET */ 1565 #ifdef INET6 1566 case AF_INET6: 1567 callout_schedule(&sc->sc_md6_tmo, tvtohz(&tv)); 1568 break; 1569 #endif /* INET6 */ 1570 default: 1571 if (sc->sc_naddrs) 1572 callout_schedule(&sc->sc_md_tmo, tvtohz(&tv)); 1573 #ifdef INET6 1574 if (sc->sc_naddrs6) 1575 callout_schedule(&sc->sc_md6_tmo, tvtohz(&tv)); 1576 #endif /* INET6 */ 1577 break; 1578 } 1579 break; 1580 case MASTER: 1581 tv.tv_sec = sc->sc_advbase; 1582 tv.tv_usec = sc->sc_advskew * 1000000 / 256; 1583 callout_schedule(&sc->sc_ad_tmo, tvtohz(&tv)); 1584 break; 1585 } 1586 } 1587 1588 static void 1589 carp_multicast_cleanup(struct carp_softc *sc) 1590 { 1591 struct ip_moptions *imo = &sc->sc_imo; 1592 #ifdef INET6 1593 struct ip6_moptions *im6o = &sc->sc_im6o; 1594 #endif 1595 u_int16_t n = imo->imo_num_memberships; 1596 1597 /* Clean up our own multicast memberships */ 1598 while (n-- > 0) { 1599 if (imo->imo_membership[n] != NULL) { 1600 in_delmulti(imo->imo_membership[n]); 1601 imo->imo_membership[n] = NULL; 1602 } 1603 } 1604 imo->imo_num_memberships = 0; 1605 imo->imo_multicast_if_index = 0; 1606 1607 #ifdef INET6 1608 while (!LIST_EMPTY(&im6o->im6o_memberships)) { 1609 struct in6_multi_mship *imm = 1610 LIST_FIRST(&im6o->im6o_memberships); 1611 1612 LIST_REMOVE(imm, i6mm_chain); 1613 in6_leavegroup(imm); 1614 } 1615 im6o->im6o_multicast_if_index = 0; 1616 #endif 1617 1618 /* And any other multicast memberships */ 1619 carp_ether_purgemulti(sc); 1620 } 1621 1622 static int 1623 carp_set_ifp(struct carp_softc *sc, struct ifnet *ifp) 1624 { 1625 struct carp_if *cif, *ncif = NULL; 1626 struct carp_softc *vr, *after = NULL; 1627 int myself = 0, error = 0; 1628 int s; 1629 1630 if (ifp == sc->sc_carpdev) 1631 return (0); 1632 1633 if (ifp != NULL) { 1634 if ((ifp->if_flags & IFF_MULTICAST) == 0) 1635 return (EADDRNOTAVAIL); 1636 1637 if (ifp->if_type == IFT_CARP) 1638 return (EINVAL); 1639 1640 if (ifp->if_carp == NULL) { 1641 ncif = malloc(sizeof(*cif), M_IFADDR, M_WAITOK); 1642 if ((error = ifpromisc(ifp, 1))) { 1643 free(ncif, M_IFADDR); 1644 return (error); 1645 } 1646 1647 ncif->vhif_ifp = ifp; 1648 TAILQ_INIT(&ncif->vhif_vrs); 1649 } else { 1650 cif = (struct carp_if *)ifp->if_carp; 1651 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) 1652 if (vr != sc && vr->sc_vhid == sc->sc_vhid) 1653 return (EINVAL); 1654 } 1655 1656 /* detach from old interface */ 1657 if (sc->sc_carpdev != NULL) 1658 carpdetach(sc); 1659 1660 /* join multicast groups */ 1661 if (sc->sc_naddrs < 0 && 1662 (error = carp_join_multicast(sc)) != 0) { 1663 if (ncif != NULL) 1664 free(ncif, M_IFADDR); 1665 return (error); 1666 } 1667 1668 #ifdef INET6 1669 if (sc->sc_naddrs6 < 0 && 1670 (error = carp_join_multicast6(sc)) != 0) { 1671 if (ncif != NULL) 1672 free(ncif, M_IFADDR); 1673 carp_multicast_cleanup(sc); 1674 return (error); 1675 } 1676 #endif 1677 1678 /* attach carp interface to physical interface */ 1679 if (ncif != NULL) 1680 ifp->if_carp = (void *)ncif; 1681 sc->sc_carpdev = ifp; 1682 sc->sc_if.if_capabilities = ifp->if_capabilities & 1683 (IFCAP_TSOv4 | IFCAP_TSOv6 | 1684 IFCAP_CSUM_IPv4_Tx|IFCAP_CSUM_IPv4_Rx| 1685 IFCAP_CSUM_TCPv4_Tx|IFCAP_CSUM_TCPv4_Rx| 1686 IFCAP_CSUM_UDPv4_Tx|IFCAP_CSUM_UDPv4_Rx| 1687 IFCAP_CSUM_TCPv6_Tx|IFCAP_CSUM_TCPv6_Rx| 1688 IFCAP_CSUM_UDPv6_Tx|IFCAP_CSUM_UDPv6_Rx); 1689 1690 cif = (struct carp_if *)ifp->if_carp; 1691 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) { 1692 if (vr == sc) 1693 myself = 1; 1694 if (vr->sc_vhid < sc->sc_vhid) 1695 after = vr; 1696 } 1697 1698 if (!myself) { 1699 /* We're trying to keep things in order */ 1700 if (after == NULL) { 1701 TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list); 1702 } else { 1703 TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, 1704 sc, sc_list); 1705 } 1706 cif->vhif_nvrs++; 1707 } 1708 if (sc->sc_naddrs || sc->sc_naddrs6) 1709 sc->sc_if.if_flags |= IFF_UP; 1710 carp_set_enaddr(sc); 1711 KERNEL_LOCK(1, NULL); 1712 s = splnet(); 1713 /* XXX linkstatehooks establish */ 1714 carp_carpdev_state(ifp); 1715 splx(s); 1716 KERNEL_UNLOCK_ONE(NULL); 1717 } else { 1718 carpdetach(sc); 1719 sc->sc_if.if_flags &= ~(IFF_UP|IFF_RUNNING); 1720 } 1721 return (0); 1722 } 1723 1724 static void 1725 carp_set_enaddr(struct carp_softc *sc) 1726 { 1727 struct ifnet *ifp = &sc->sc_if; 1728 uint8_t enaddr[ETHER_ADDR_LEN]; 1729 1730 if (sc->sc_vhid == -1) { 1731 ifp->if_addrlen = 0; 1732 if_alloc_sadl(ifp); 1733 return; 1734 } 1735 1736 if (sc->sc_carpdev && sc->sc_carpdev->if_type == IFT_ISO88025) { 1737 enaddr[0] = 3; 1738 enaddr[1] = 0; 1739 enaddr[2] = 0x40 >> (sc->sc_vhid - 1); 1740 enaddr[3] = 0x40000 >> (sc->sc_vhid - 1); 1741 enaddr[4] = 0; 1742 enaddr[5] = 0; 1743 } else { 1744 enaddr[0] = 0; 1745 enaddr[1] = 0; 1746 enaddr[2] = 0x5e; 1747 enaddr[3] = 0; 1748 enaddr[4] = 1; 1749 enaddr[5] = sc->sc_vhid; 1750 } 1751 1752 if_set_sadl(ifp, enaddr, sizeof(enaddr), false); 1753 } 1754 1755 #if 0 1756 static void 1757 carp_addr_updated(void *v) 1758 { 1759 struct carp_softc *sc = (struct carp_softc *) v; 1760 struct ifaddr *ifa; 1761 int new_naddrs = 0, new_naddrs6 = 0; 1762 1763 IFADDR_READER_FOREACH(ifa, &sc->sc_if) { 1764 if (ifa->ifa_addr->sa_family == AF_INET) 1765 new_naddrs++; 1766 else if (ifa->ifa_addr->sa_family == AF_INET6) 1767 new_naddrs6++; 1768 } 1769 1770 /* Handle a callback after SIOCDIFADDR */ 1771 if (new_naddrs < sc->sc_naddrs || new_naddrs6 < sc->sc_naddrs6) { 1772 struct in_addr mc_addr; 1773 1774 sc->sc_naddrs = new_naddrs; 1775 sc->sc_naddrs6 = new_naddrs6; 1776 1777 /* Re-establish multicast membership removed by in_control */ 1778 mc_addr.s_addr = INADDR_CARP_GROUP; 1779 if (!in_multi_group(mc_addr, &sc->sc_if, 0)) { 1780 memset(&sc->sc_imo, 0, sizeof(sc->sc_imo)); 1781 1782 if (sc->sc_carpdev != NULL && sc->sc_naddrs > 0) 1783 carp_join_multicast(sc); 1784 } 1785 1786 if (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) { 1787 sc->sc_if.if_flags &= ~IFF_UP; 1788 carp_set_state(sc, INIT); 1789 } else 1790 carp_hmac_prepare(sc); 1791 } 1792 1793 carp_setrun(sc, 0); 1794 } 1795 #endif 1796 1797 static int 1798 carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin) 1799 { 1800 struct ifnet *ifp = sc->sc_carpdev; 1801 struct in_ifaddr *ia, *ia_if; 1802 int error = 0; 1803 int s; 1804 1805 if (sin->sin_addr.s_addr == 0) { 1806 if (!(sc->sc_if.if_flags & IFF_UP)) 1807 carp_set_state(sc, INIT); 1808 if (sc->sc_naddrs) 1809 sc->sc_if.if_flags |= IFF_UP; 1810 carp_setrun(sc, 0); 1811 return (0); 1812 } 1813 1814 /* we have to do this by hand to ensure we don't match on ourselves */ 1815 ia_if = NULL; 1816 s = pserialize_read_enter(); 1817 IN_ADDRLIST_READER_FOREACH(ia) { 1818 /* and, yeah, we need a multicast-capable iface too */ 1819 if (ia->ia_ifp != &sc->sc_if && 1820 ia->ia_ifp->if_type != IFT_CARP && 1821 (ia->ia_ifp->if_flags & IFF_MULTICAST) && 1822 (sin->sin_addr.s_addr & ia->ia_subnetmask) == 1823 ia->ia_subnet) { 1824 if (!ia_if) 1825 ia_if = ia; 1826 } 1827 } 1828 1829 if (ia_if) { 1830 ia = ia_if; 1831 if (ifp) { 1832 if (ifp != ia->ia_ifp) 1833 return (EADDRNOTAVAIL); 1834 } else { 1835 /* FIXME NOMPSAFE */ 1836 ifp = ia->ia_ifp; 1837 } 1838 } 1839 pserialize_read_exit(s); 1840 1841 if ((error = carp_set_ifp(sc, ifp))) 1842 return (error); 1843 1844 if (sc->sc_carpdev == NULL) 1845 return (EADDRNOTAVAIL); 1846 1847 if (sc->sc_naddrs == 0 && (error = carp_join_multicast(sc)) != 0) 1848 return (error); 1849 1850 sc->sc_naddrs++; 1851 if (sc->sc_carpdev != NULL) 1852 sc->sc_if.if_flags |= IFF_UP; 1853 1854 carp_set_state(sc, INIT); 1855 carp_setrun(sc, 0); 1856 1857 /* 1858 * Hook if_addrhooks so that we get a callback after in_ifinit has run, 1859 * to correct any inappropriate routes that it inserted. 1860 */ 1861 if (sc->ah_cookie == 0) { 1862 /* XXX link address hook */ 1863 } 1864 1865 return (0); 1866 } 1867 1868 static int 1869 carp_join_multicast(struct carp_softc *sc) 1870 { 1871 struct ip_moptions *imo = &sc->sc_imo, tmpimo; 1872 struct in_addr addr; 1873 1874 memset(&tmpimo, 0, sizeof(tmpimo)); 1875 addr.s_addr = INADDR_CARP_GROUP; 1876 if ((tmpimo.imo_membership[0] = 1877 in_addmulti(&addr, &sc->sc_if)) == NULL) { 1878 return (ENOBUFS); 1879 } 1880 1881 imo->imo_membership[0] = tmpimo.imo_membership[0]; 1882 imo->imo_num_memberships = 1; 1883 imo->imo_multicast_if_index = sc->sc_if.if_index; 1884 imo->imo_multicast_ttl = CARP_DFLTTL; 1885 imo->imo_multicast_loop = 0; 1886 return (0); 1887 } 1888 1889 1890 #ifdef INET6 1891 static int 1892 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6) 1893 { 1894 struct ifnet *ifp = sc->sc_carpdev; 1895 struct in6_ifaddr *ia, *ia_if; 1896 int error = 0; 1897 int s; 1898 1899 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) { 1900 if (!(sc->sc_if.if_flags & IFF_UP)) 1901 carp_set_state(sc, INIT); 1902 if (sc->sc_naddrs6) 1903 sc->sc_if.if_flags |= IFF_UP; 1904 carp_setrun(sc, 0); 1905 return (0); 1906 } 1907 1908 /* we have to do this by hand to ensure we don't match on ourselves */ 1909 ia_if = NULL; 1910 s = pserialize_read_enter(); 1911 IN6_ADDRLIST_READER_FOREACH(ia) { 1912 int i; 1913 1914 for (i = 0; i < 4; i++) { 1915 if ((sin6->sin6_addr.s6_addr32[i] & 1916 ia->ia_prefixmask.sin6_addr.s6_addr32[i]) != 1917 (ia->ia_addr.sin6_addr.s6_addr32[i] & 1918 ia->ia_prefixmask.sin6_addr.s6_addr32[i])) 1919 break; 1920 } 1921 /* and, yeah, we need a multicast-capable iface too */ 1922 if (ia->ia_ifp != &sc->sc_if && 1923 ia->ia_ifp->if_type != IFT_CARP && 1924 (ia->ia_ifp->if_flags & IFF_MULTICAST) && 1925 (i == 4)) { 1926 if (!ia_if) 1927 ia_if = ia; 1928 } 1929 } 1930 pserialize_read_exit(s); 1931 1932 if (ia_if) { 1933 ia = ia_if; 1934 if (sc->sc_carpdev) { 1935 if (sc->sc_carpdev != ia->ia_ifp) 1936 return (EADDRNOTAVAIL); 1937 } else { 1938 ifp = ia->ia_ifp; 1939 } 1940 } 1941 1942 if ((error = carp_set_ifp(sc, ifp))) 1943 return (error); 1944 1945 if (sc->sc_carpdev == NULL) 1946 return (EADDRNOTAVAIL); 1947 1948 if (sc->sc_naddrs6 == 0 && (error = carp_join_multicast6(sc)) != 0) 1949 return (error); 1950 1951 sc->sc_naddrs6++; 1952 if (sc->sc_carpdev != NULL) 1953 sc->sc_if.if_flags |= IFF_UP; 1954 carp_set_state(sc, INIT); 1955 carp_setrun(sc, 0); 1956 1957 return (0); 1958 } 1959 1960 static int 1961 carp_join_multicast6(struct carp_softc *sc) 1962 { 1963 struct in6_multi_mship *imm, *imm2; 1964 struct ip6_moptions *im6o = &sc->sc_im6o; 1965 struct sockaddr_in6 addr6; 1966 int error; 1967 1968 /* Join IPv6 CARP multicast group */ 1969 memset(&addr6, 0, sizeof(addr6)); 1970 addr6.sin6_family = AF_INET6; 1971 addr6.sin6_len = sizeof(addr6); 1972 addr6.sin6_addr.s6_addr16[0] = htons(0xff02); 1973 addr6.sin6_addr.s6_addr16[1] = htons(sc->sc_if.if_index); 1974 addr6.sin6_addr.s6_addr8[15] = 0x12; 1975 if ((imm = in6_joingroup(&sc->sc_if, 1976 &addr6.sin6_addr, &error, 0)) == NULL) { 1977 return (error); 1978 } 1979 /* join solicited multicast address */ 1980 memset(&addr6.sin6_addr, 0, sizeof(addr6.sin6_addr)); 1981 addr6.sin6_addr.s6_addr16[0] = htons(0xff02); 1982 addr6.sin6_addr.s6_addr16[1] = htons(sc->sc_if.if_index); 1983 addr6.sin6_addr.s6_addr32[1] = 0; 1984 addr6.sin6_addr.s6_addr32[2] = htonl(1); 1985 addr6.sin6_addr.s6_addr32[3] = 0; 1986 addr6.sin6_addr.s6_addr8[12] = 0xff; 1987 if ((imm2 = in6_joingroup(&sc->sc_if, 1988 &addr6.sin6_addr, &error, 0)) == NULL) { 1989 in6_leavegroup(imm); 1990 return (error); 1991 } 1992 1993 /* apply v6 multicast membership */ 1994 im6o->im6o_multicast_if_index = sc->sc_if.if_index; 1995 if (imm) 1996 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, 1997 i6mm_chain); 1998 if (imm2) 1999 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm2, 2000 i6mm_chain); 2001 2002 return (0); 2003 } 2004 2005 #endif /* INET6 */ 2006 2007 static int 2008 carp_ioctl(struct ifnet *ifp, u_long cmd, void *data) 2009 { 2010 struct lwp *l = curlwp; /* XXX */ 2011 struct carp_softc *sc = ifp->if_softc, *vr; 2012 struct carpreq carpr; 2013 struct ifaddr *ifa; 2014 struct ifreq *ifr; 2015 struct ifnet *cdev = NULL; 2016 int error = 0; 2017 2018 ifa = (struct ifaddr *)data; 2019 ifr = (struct ifreq *)data; 2020 2021 switch (cmd) { 2022 case SIOCINITIFADDR: 2023 switch (ifa->ifa_addr->sa_family) { 2024 #ifdef INET 2025 case AF_INET: 2026 sc->sc_if.if_flags |= IFF_UP; 2027 memcpy(ifa->ifa_dstaddr, ifa->ifa_addr, 2028 sizeof(struct sockaddr)); 2029 error = carp_set_addr(sc, satosin(ifa->ifa_addr)); 2030 break; 2031 #endif /* INET */ 2032 #ifdef INET6 2033 case AF_INET6: 2034 sc->sc_if.if_flags|= IFF_UP; 2035 error = carp_set_addr6(sc, satosin6(ifa->ifa_addr)); 2036 break; 2037 #endif /* INET6 */ 2038 default: 2039 error = EAFNOSUPPORT; 2040 break; 2041 } 2042 break; 2043 2044 case SIOCSIFFLAGS: 2045 if ((error = ifioctl_common(ifp, cmd, data)) != 0) 2046 break; 2047 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) { 2048 callout_stop(&sc->sc_ad_tmo); 2049 callout_stop(&sc->sc_md_tmo); 2050 callout_stop(&sc->sc_md6_tmo); 2051 if (sc->sc_state == MASTER) { 2052 /* we need the interface up to bow out */ 2053 sc->sc_if.if_flags |= IFF_UP; 2054 sc->sc_bow_out = 1; 2055 carp_send_ad(sc); 2056 } 2057 sc->sc_if.if_flags &= ~IFF_UP; 2058 carp_set_state(sc, INIT); 2059 carp_setrun(sc, 0); 2060 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) { 2061 sc->sc_if.if_flags |= IFF_UP; 2062 carp_setrun(sc, 0); 2063 } 2064 carp_update_link_state(sc); 2065 break; 2066 2067 case SIOCSVH: 2068 if (l == NULL) 2069 break; 2070 if ((error = kauth_authorize_network(l->l_cred, 2071 KAUTH_NETWORK_INTERFACE, 2072 KAUTH_REQ_NETWORK_INTERFACE_SETPRIV, ifp, (void *)cmd, 2073 NULL)) != 0) 2074 break; 2075 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr))) 2076 break; 2077 error = 1; 2078 if (carpr.carpr_carpdev[0] != '\0' && 2079 (cdev = ifunit(carpr.carpr_carpdev)) == NULL) 2080 return (EINVAL); 2081 if ((error = carp_set_ifp(sc, cdev))) 2082 return (error); 2083 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) { 2084 switch (carpr.carpr_state) { 2085 case BACKUP: 2086 callout_stop(&sc->sc_ad_tmo); 2087 carp_set_state(sc, BACKUP); 2088 carp_setrun(sc, 0); 2089 carp_setroute(sc, RTM_DELETE); 2090 break; 2091 case MASTER: 2092 carp_master_down(sc); 2093 break; 2094 default: 2095 break; 2096 } 2097 } 2098 if (carpr.carpr_vhid > 0) { 2099 if (carpr.carpr_vhid > 255) { 2100 error = EINVAL; 2101 break; 2102 } 2103 if (sc->sc_carpdev) { 2104 struct carp_if *cif; 2105 cif = (struct carp_if *)sc->sc_carpdev->if_carp; 2106 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) 2107 if (vr != sc && 2108 vr->sc_vhid == carpr.carpr_vhid) 2109 return (EINVAL); 2110 } 2111 sc->sc_vhid = carpr.carpr_vhid; 2112 carp_set_enaddr(sc); 2113 carp_set_state(sc, INIT); 2114 error--; 2115 } 2116 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) { 2117 if (carpr.carpr_advskew > 254) { 2118 error = EINVAL; 2119 break; 2120 } 2121 if (carpr.carpr_advbase > 255) { 2122 error = EINVAL; 2123 break; 2124 } 2125 sc->sc_advbase = carpr.carpr_advbase; 2126 sc->sc_advskew = carpr.carpr_advskew; 2127 error--; 2128 } 2129 memcpy(sc->sc_key, carpr.carpr_key, sizeof(sc->sc_key)); 2130 if (error > 0) 2131 error = EINVAL; 2132 else { 2133 error = 0; 2134 carp_setrun(sc, 0); 2135 } 2136 break; 2137 2138 case SIOCGVH: 2139 memset(&carpr, 0, sizeof(carpr)); 2140 if (sc->sc_carpdev != NULL) 2141 strlcpy(carpr.carpr_carpdev, sc->sc_carpdev->if_xname, 2142 IFNAMSIZ); 2143 carpr.carpr_state = sc->sc_state; 2144 carpr.carpr_vhid = sc->sc_vhid; 2145 carpr.carpr_advbase = sc->sc_advbase; 2146 carpr.carpr_advskew = sc->sc_advskew; 2147 2148 if ((l != NULL) && (error = kauth_authorize_network(l->l_cred, 2149 KAUTH_NETWORK_INTERFACE, 2150 KAUTH_REQ_NETWORK_INTERFACE_SETPRIV, ifp, (void *)cmd, 2151 NULL)) == 0) 2152 memcpy(carpr.carpr_key, sc->sc_key, 2153 sizeof(carpr.carpr_key)); 2154 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr)); 2155 break; 2156 2157 case SIOCADDMULTI: 2158 error = carp_ether_addmulti(sc, ifr); 2159 break; 2160 2161 case SIOCDELMULTI: 2162 error = carp_ether_delmulti(sc, ifr); 2163 break; 2164 2165 case SIOCSIFCAP: 2166 if ((error = ifioctl_common(ifp, cmd, data)) == ENETRESET) 2167 error = 0; 2168 break; 2169 2170 default: 2171 error = ether_ioctl(ifp, cmd, data); 2172 } 2173 2174 carp_hmac_prepare(sc); 2175 return (error); 2176 } 2177 2178 2179 /* 2180 * Start output on carp interface. This function should never be called. 2181 */ 2182 static void 2183 carp_start(struct ifnet *ifp) 2184 { 2185 #ifdef DEBUG 2186 printf("%s: start called\n", ifp->if_xname); 2187 #endif 2188 } 2189 2190 int 2191 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa, 2192 const struct rtentry *rt) 2193 { 2194 struct carp_softc *sc = ((struct carp_softc *)ifp->if_softc); 2195 KASSERT(KERNEL_LOCKED_P()); 2196 2197 if (sc->sc_carpdev != NULL && sc->sc_state == MASTER) { 2198 return if_output_lock(sc->sc_carpdev, ifp, m, sa, rt); 2199 } else { 2200 m_freem(m); 2201 return (ENETUNREACH); 2202 } 2203 } 2204 2205 static void 2206 carp_set_state(struct carp_softc *sc, int state) 2207 { 2208 static const char *carp_states[] = { CARP_STATES }; 2209 2210 if (sc->sc_state == state) 2211 return; 2212 2213 CARP_LOG(sc, ("state transition from: %s -> to: %s", carp_states[sc->sc_state], carp_states[state])); 2214 2215 sc->sc_state = state; 2216 carp_update_link_state(sc); 2217 } 2218 2219 static void 2220 carp_update_link_state(struct carp_softc *sc) 2221 { 2222 int link_state; 2223 2224 switch (sc->sc_state) { 2225 case BACKUP: 2226 link_state = LINK_STATE_DOWN; 2227 break; 2228 case MASTER: 2229 link_state = LINK_STATE_UP; 2230 break; 2231 default: 2232 /* Not useable, so down makes perfect sense. */ 2233 link_state = LINK_STATE_DOWN; 2234 break; 2235 } 2236 if_link_state_change(&sc->sc_if, link_state); 2237 } 2238 2239 void 2240 carp_carpdev_state(void *v) 2241 { 2242 struct carp_if *cif; 2243 struct carp_softc *sc; 2244 struct ifnet *ifp = v; 2245 2246 if (ifp->if_type == IFT_CARP) 2247 return; 2248 2249 cif = (struct carp_if *)ifp->if_carp; 2250 2251 TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) { 2252 int suppressed = sc->sc_suppress; 2253 2254 if (sc->sc_carpdev->if_link_state == LINK_STATE_DOWN || 2255 !(sc->sc_carpdev->if_flags & IFF_UP)) { 2256 sc->sc_if.if_flags &= ~IFF_RUNNING; 2257 callout_stop(&sc->sc_ad_tmo); 2258 callout_stop(&sc->sc_md_tmo); 2259 callout_stop(&sc->sc_md6_tmo); 2260 carp_set_state(sc, INIT); 2261 sc->sc_suppress = 1; 2262 carp_setrun(sc, 0); 2263 if (!suppressed) { 2264 carp_suppress_preempt++; 2265 if (carp_suppress_preempt == 1) 2266 carp_send_ad_all(); 2267 } 2268 } else { 2269 carp_set_state(sc, INIT); 2270 sc->sc_suppress = 0; 2271 carp_setrun(sc, 0); 2272 if (suppressed) 2273 carp_suppress_preempt--; 2274 } 2275 } 2276 } 2277 2278 static int 2279 carp_ether_addmulti(struct carp_softc *sc, struct ifreq *ifr) 2280 { 2281 const struct sockaddr *sa = ifreq_getaddr(SIOCADDMULTI, ifr); 2282 struct ifnet *ifp; 2283 struct carp_mc_entry *mc; 2284 u_int8_t addrlo[ETHER_ADDR_LEN], addrhi[ETHER_ADDR_LEN]; 2285 int error; 2286 2287 ifp = sc->sc_carpdev; 2288 if (ifp == NULL) 2289 return (EINVAL); 2290 2291 error = ether_addmulti(sa, &sc->sc_ac); 2292 if (error != ENETRESET) 2293 return (error); 2294 2295 /* 2296 * This is new multicast address. We have to tell parent 2297 * about it. Also, remember this multicast address so that 2298 * we can delete them on unconfigure. 2299 */ 2300 mc = malloc(sizeof(struct carp_mc_entry), M_DEVBUF, M_NOWAIT); 2301 if (mc == NULL) { 2302 error = ENOMEM; 2303 goto alloc_failed; 2304 } 2305 2306 /* 2307 * As ether_addmulti() returns ENETRESET, following two 2308 * statement shouldn't fail. 2309 */ 2310 (void)ether_multiaddr(sa, addrlo, addrhi); 2311 2312 ETHER_LOCK(&sc->sc_ac); 2313 mc->mc_enm = ether_lookup_multi(addrlo, addrhi, &sc->sc_ac); 2314 ETHER_UNLOCK(&sc->sc_ac); 2315 2316 memcpy(&mc->mc_addr, sa, sa->sa_len); 2317 LIST_INSERT_HEAD(&sc->carp_mc_listhead, mc, mc_entries); 2318 2319 error = if_mcast_op(ifp, SIOCADDMULTI, sa); 2320 if (error != 0) 2321 goto ioctl_failed; 2322 2323 return (error); 2324 2325 ioctl_failed: 2326 LIST_REMOVE(mc, mc_entries); 2327 free(mc, M_DEVBUF); 2328 alloc_failed: 2329 (void)ether_delmulti(sa, &sc->sc_ac); 2330 2331 return (error); 2332 } 2333 2334 static int 2335 carp_ether_delmulti(struct carp_softc *sc, struct ifreq *ifr) 2336 { 2337 const struct sockaddr *sa = ifreq_getaddr(SIOCDELMULTI, ifr); 2338 struct ifnet *ifp; 2339 struct ether_multi *enm; 2340 struct carp_mc_entry *mc; 2341 u_int8_t addrlo[ETHER_ADDR_LEN], addrhi[ETHER_ADDR_LEN]; 2342 int error; 2343 2344 ifp = sc->sc_carpdev; 2345 if (ifp == NULL) 2346 return (EINVAL); 2347 2348 /* 2349 * Find a key to lookup carp_mc_entry. We have to do this 2350 * before calling ether_delmulti for obvious reason. 2351 */ 2352 if ((error = ether_multiaddr(sa, addrlo, addrhi)) != 0) 2353 return (error); 2354 2355 ETHER_LOCK(&sc->sc_ac); 2356 enm = ether_lookup_multi(addrlo, addrhi, &sc->sc_ac); 2357 ETHER_UNLOCK(&sc->sc_ac); 2358 if (enm == NULL) 2359 return (EINVAL); 2360 2361 LIST_FOREACH(mc, &sc->carp_mc_listhead, mc_entries) 2362 if (mc->mc_enm == enm) 2363 break; 2364 2365 /* We won't delete entries we didn't add */ 2366 if (mc == NULL) 2367 return (EINVAL); 2368 2369 error = ether_delmulti(sa, &sc->sc_ac); 2370 if (error != ENETRESET) 2371 return (error); 2372 2373 /* We no longer use this multicast address. Tell parent so. */ 2374 error = if_mcast_op(ifp, SIOCDELMULTI, sa); 2375 if (error == 0) { 2376 /* And forget about this address. */ 2377 LIST_REMOVE(mc, mc_entries); 2378 free(mc, M_DEVBUF); 2379 } else 2380 (void)ether_addmulti(sa, &sc->sc_ac); 2381 return (error); 2382 } 2383 2384 /* 2385 * Delete any multicast address we have asked to add from parent 2386 * interface. Called when the carp is being unconfigured. 2387 */ 2388 static void 2389 carp_ether_purgemulti(struct carp_softc *sc) 2390 { 2391 struct ifnet *ifp = sc->sc_carpdev; /* Parent. */ 2392 struct carp_mc_entry *mc; 2393 2394 if (ifp == NULL) 2395 return; 2396 2397 while ((mc = LIST_FIRST(&sc->carp_mc_listhead)) != NULL) { 2398 (void)if_mcast_op(ifp, SIOCDELMULTI, sstosa(&mc->mc_addr)); 2399 LIST_REMOVE(mc, mc_entries); 2400 free(mc, M_DEVBUF); 2401 } 2402 } 2403 2404 static int 2405 sysctl_net_inet_carp_stats(SYSCTLFN_ARGS) 2406 { 2407 2408 return (NETSTAT_SYSCTL(carpstat_percpu, CARP_NSTATS)); 2409 } 2410 2411 void 2412 carp_init(void) 2413 { 2414 2415 sysctl_net_inet_carp_setup(NULL); 2416 #ifdef MBUFTRACE 2417 MOWNER_ATTACH(&carp_proto_mowner_rx); 2418 MOWNER_ATTACH(&carp_proto_mowner_tx); 2419 MOWNER_ATTACH(&carp_proto6_mowner_rx); 2420 MOWNER_ATTACH(&carp_proto6_mowner_tx); 2421 #endif 2422 2423 carp_wqinput = wqinput_create("carp", _carp_proto_input); 2424 #ifdef INET6 2425 carp6_wqinput = wqinput_create("carp6", _carp6_proto_input); 2426 #endif 2427 } 2428 2429 static void 2430 sysctl_net_inet_carp_setup(struct sysctllog **clog) 2431 { 2432 2433 sysctl_createv(clog, 0, NULL, NULL, 2434 CTLFLAG_PERMANENT, 2435 CTLTYPE_NODE, "inet", NULL, 2436 NULL, 0, NULL, 0, 2437 CTL_NET, PF_INET, CTL_EOL); 2438 sysctl_createv(clog, 0, NULL, NULL, 2439 CTLFLAG_PERMANENT, 2440 CTLTYPE_NODE, "carp", 2441 SYSCTL_DESCR("CARP related settings"), 2442 NULL, 0, NULL, 0, 2443 CTL_NET, PF_INET, IPPROTO_CARP, CTL_EOL); 2444 2445 sysctl_createv(clog, 0, NULL, NULL, 2446 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2447 CTLTYPE_INT, "preempt", 2448 SYSCTL_DESCR("Enable CARP Preempt"), 2449 NULL, 0, &carp_opts[CARPCTL_PREEMPT], 0, 2450 CTL_NET, PF_INET, IPPROTO_CARP, 2451 CTL_CREATE, CTL_EOL); 2452 sysctl_createv(clog, 0, NULL, NULL, 2453 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2454 CTLTYPE_INT, "arpbalance", 2455 SYSCTL_DESCR("Enable ARP balancing"), 2456 NULL, 0, &carp_opts[CARPCTL_ARPBALANCE], 0, 2457 CTL_NET, PF_INET, IPPROTO_CARP, 2458 CTL_CREATE, CTL_EOL); 2459 sysctl_createv(clog, 0, NULL, NULL, 2460 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2461 CTLTYPE_INT, "allow", 2462 SYSCTL_DESCR("Enable CARP"), 2463 NULL, 0, &carp_opts[CARPCTL_ALLOW], 0, 2464 CTL_NET, PF_INET, IPPROTO_CARP, 2465 CTL_CREATE, CTL_EOL); 2466 sysctl_createv(clog, 0, NULL, NULL, 2467 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 2468 CTLTYPE_INT, "log", 2469 SYSCTL_DESCR("CARP logging"), 2470 NULL, 0, &carp_opts[CARPCTL_LOG], 0, 2471 CTL_NET, PF_INET, IPPROTO_CARP, 2472 CTL_CREATE, CTL_EOL); 2473 sysctl_createv(clog, 0, NULL, NULL, 2474 CTLFLAG_PERMANENT, 2475 CTLTYPE_STRUCT, "stats", 2476 SYSCTL_DESCR("CARP statistics"), 2477 sysctl_net_inet_carp_stats, 0, NULL, 0, 2478 CTL_NET, PF_INET, IPPROTO_CARP, CARPCTL_STATS, 2479 CTL_EOL); 2480 } 2481
Indexes created Sat Sep 27 14:09:57 GMT 2025