1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2002 Michael Shalayeff. 5 * Copyright (c) 2003 Ryan McBride. 6 * Copyright (c) 2011 Gleb Smirnoff <glebius@FreeBSD.org> 7 * All rights reserved. 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 1. Redistributions of source code must retain the above copyright 13 * notice, this list of conditions and the following disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in the 16 * documentation and/or other materials provided with the distribution. 17 * 18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 19 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 20 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 21 * IN NO EVENT SHALL THE AUTHOR OR HIS RELATIVES BE LIABLE FOR ANY DIRECT, 22 * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES 23 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR 24 * SERVICES; LOSS OF MIND, USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, 26 * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING 27 * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF 28 * THE POSSIBILITY OF SUCH DAMAGE. 29 */ 30 31 #include <sys/cdefs.h> 32 __FBSDID("$FreeBSD$"); 33 34 #include "opt_bpf.h" 35 #include "opt_inet.h" 36 #include "opt_inet6.h" 37 38 #include <sys/param.h> 39 #include <sys/systm.h> 40 #include <sys/devctl.h> 41 #include <sys/jail.h> 42 #include <sys/kernel.h> 43 #include <sys/limits.h> 44 #include <sys/malloc.h> 45 #include <sys/mbuf.h> 46 #include <sys/module.h> 47 #include <sys/priv.h> 48 #include <sys/proc.h> 49 #include <sys/protosw.h> 50 #include <sys/socket.h> 51 #include <sys/sockio.h> 52 #include <sys/sysctl.h> 53 #include <sys/syslog.h> 54 #include <sys/taskqueue.h> 55 #include <sys/counter.h> 56 57 #include <net/ethernet.h> 58 #include <net/if.h> 59 #include <net/if_var.h> 60 #include <net/if_dl.h> 61 #include <net/if_llatbl.h> 62 #include <net/if_types.h> 63 #include <net/route.h> 64 #include <net/vnet.h> 65 66 #if defined(INET) || defined(INET6) 67 #include <netinet/in.h> 68 #include <netinet/in_var.h> 69 #include <netinet/ip_carp.h> 70 #include <netinet/ip.h> 71 #include <machine/in_cksum.h> 72 #endif 73 #ifdef INET 74 #include <netinet/ip_var.h> 75 #include <netinet/if_ether.h> 76 #endif 77 78 #ifdef INET6 79 #include <netinet/icmp6.h> 80 #include <netinet/ip6.h> 81 #include <netinet6/in6_var.h> 82 #include <netinet6/ip6_var.h> 83 #include <netinet6/scope6_var.h> 84 #include <netinet6/nd6.h> 85 #endif 86 87 #include <crypto/sha1.h> 88 89 static MALLOC_DEFINE(M_CARP, "CARP", "CARP addresses"); 90 91 struct carp_softc { 92 struct ifnet *sc_carpdev; /* Pointer to parent ifnet. */ 93 struct ifaddr **sc_ifas; /* Our ifaddrs. */ 94 struct sockaddr_dl sc_addr; /* Our link level address. */ 95 struct callout sc_ad_tmo; /* Advertising timeout. */ 96 #ifdef INET 97 struct callout sc_md_tmo; /* Master down timeout. */ 98 #endif 99 #ifdef INET6 100 struct callout sc_md6_tmo; /* XXX: Master down timeout. */ 101 #endif 102 struct mtx sc_mtx; 103 104 int sc_vhid; 105 int sc_advskew; 106 int sc_advbase; 107 108 int sc_naddrs; 109 int sc_naddrs6; 110 int sc_ifasiz; 111 enum { INIT = 0, BACKUP, MASTER } sc_state; 112 int sc_suppress; 113 int sc_sendad_errors; 114 #define CARP_SENDAD_MAX_ERRORS 3 115 int sc_sendad_success; 116 #define CARP_SENDAD_MIN_SUCCESS 3 117 118 int sc_init_counter; 119 uint64_t sc_counter; 120 121 /* authentication */ 122 #define CARP_HMAC_PAD 64 123 unsigned char sc_key[CARP_KEY_LEN]; 124 unsigned char sc_pad[CARP_HMAC_PAD]; 125 SHA1_CTX sc_sha1; 126 127 TAILQ_ENTRY(carp_softc) sc_list; /* On the carp_if list. */ 128 LIST_ENTRY(carp_softc) sc_next; /* On the global list. */ 129 }; 130 131 struct carp_if { 132 #ifdef INET 133 int cif_naddrs; 134 #endif 135 #ifdef INET6 136 int cif_naddrs6; 137 #endif 138 TAILQ_HEAD(, carp_softc) cif_vrs; 139 #ifdef INET 140 struct ip_moptions cif_imo; 141 #endif 142 #ifdef INET6 143 struct ip6_moptions cif_im6o; 144 #endif 145 struct ifnet *cif_ifp; 146 struct mtx cif_mtx; 147 uint32_t cif_flags; 148 #define CIF_PROMISC 0x00000001 149 }; 150 151 #define CARP_INET 0 152 #define CARP_INET6 1 153 static int proto_reg[] = {-1, -1}; 154 155 /* 156 * Brief design of carp(4). 157 * 158 * Any carp-capable ifnet may have a list of carp softcs hanging off 159 * its ifp->if_carp pointer. Each softc represents one unique virtual 160 * host id, or vhid. The softc has a back pointer to the ifnet. All 161 * softcs are joined in a global list, which has quite limited use. 162 * 163 * Any interface address that takes part in CARP negotiation has a 164 * pointer to the softc of its vhid, ifa->ifa_carp. That could be either 165 * AF_INET or AF_INET6 address. 166 * 167 * Although, one can get the softc's backpointer to ifnet and traverse 168 * through its ifp->if_addrhead queue to find all interface addresses 169 * involved in CARP, we keep a growable array of ifaddr pointers. This 170 * allows us to avoid grabbing the IF_ADDR_LOCK() in many traversals that 171 * do calls into the network stack, thus avoiding LORs. 172 * 173 * Locking: 174 * 175 * Each softc has a lock sc_mtx. It is used to synchronise carp_input_c(), 176 * callout-driven events and ioctl()s. 177 * 178 * To traverse the list of softcs on an ifnet we use CIF_LOCK() or carp_sx. 179 * To traverse the global list we use the mutex carp_mtx. 180 * 181 * Known issues with locking: 182 * 183 * - Sending ad, we put the pointer to the softc in an mtag, and no reference 184 * counting is done on the softc. 185 * - On module unload we may race (?) with packet processing thread 186 * dereferencing our function pointers. 187 */ 188 189 /* Accept incoming CARP packets. */ 190 VNET_DEFINE_STATIC(int, carp_allow) = 1; 191 #define V_carp_allow VNET(carp_allow) 192 193 /* Set DSCP in outgoing CARP packets. */ 194 VNET_DEFINE_STATIC(int, carp_dscp) = 56; 195 #define V_carp_dscp VNET(carp_dscp) 196 197 /* Preempt slower nodes. */ 198 VNET_DEFINE_STATIC(int, carp_preempt) = 0; 199 #define V_carp_preempt VNET(carp_preempt) 200 201 /* Log level. */ 202 VNET_DEFINE_STATIC(int, carp_log) = 1; 203 #define V_carp_log VNET(carp_log) 204 205 /* Global advskew demotion. */ 206 VNET_DEFINE_STATIC(int, carp_demotion) = 0; 207 #define V_carp_demotion VNET(carp_demotion) 208 209 /* Send error demotion factor. */ 210 VNET_DEFINE_STATIC(int, carp_senderr_adj) = CARP_MAXSKEW; 211 #define V_carp_senderr_adj VNET(carp_senderr_adj) 212 213 /* Iface down demotion factor. */ 214 VNET_DEFINE_STATIC(int, carp_ifdown_adj) = CARP_MAXSKEW; 215 #define V_carp_ifdown_adj VNET(carp_ifdown_adj) 216 217 static int carp_allow_sysctl(SYSCTL_HANDLER_ARGS); 218 static int carp_dscp_sysctl(SYSCTL_HANDLER_ARGS); 219 static int carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS); 220 221 SYSCTL_NODE(_net_inet, IPPROTO_CARP, carp, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 222 "CARP"); 223 SYSCTL_PROC(_net_inet_carp, OID_AUTO, allow, 224 CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, 225 0, 0, carp_allow_sysctl, "I", 226 "Accept incoming CARP packets"); 227 SYSCTL_PROC(_net_inet_carp, OID_AUTO, dscp, 228 CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, 229 0, 0, carp_dscp_sysctl, "I", 230 "DSCP value for carp packets"); 231 SYSCTL_INT(_net_inet_carp, OID_AUTO, preempt, CTLFLAG_VNET | CTLFLAG_RW, 232 &VNET_NAME(carp_preempt), 0, "High-priority backup preemption mode"); 233 SYSCTL_INT(_net_inet_carp, OID_AUTO, log, CTLFLAG_VNET | CTLFLAG_RW, 234 &VNET_NAME(carp_log), 0, "CARP log level"); 235 SYSCTL_PROC(_net_inet_carp, OID_AUTO, demotion, 236 CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, 237 0, 0, carp_demote_adj_sysctl, "I", 238 "Adjust demotion factor (skew of advskew)"); 239 SYSCTL_INT(_net_inet_carp, OID_AUTO, senderr_demotion_factor, 240 CTLFLAG_VNET | CTLFLAG_RW, 241 &VNET_NAME(carp_senderr_adj), 0, "Send error demotion factor adjustment"); 242 SYSCTL_INT(_net_inet_carp, OID_AUTO, ifdown_demotion_factor, 243 CTLFLAG_VNET | CTLFLAG_RW, 244 &VNET_NAME(carp_ifdown_adj), 0, 245 "Interface down demotion factor adjustment"); 246 247 VNET_PCPUSTAT_DEFINE(struct carpstats, carpstats); 248 VNET_PCPUSTAT_SYSINIT(carpstats); 249 VNET_PCPUSTAT_SYSUNINIT(carpstats); 250 251 #define CARPSTATS_ADD(name, val) \ 252 counter_u64_add(VNET(carpstats)[offsetof(struct carpstats, name) / \ 253 sizeof(uint64_t)], (val)) 254 #define CARPSTATS_INC(name) CARPSTATS_ADD(name, 1) 255 256 SYSCTL_VNET_PCPUSTAT(_net_inet_carp, OID_AUTO, stats, struct carpstats, 257 carpstats, "CARP statistics (struct carpstats, netinet/ip_carp.h)"); 258 259 #define CARP_LOCK_INIT(sc) mtx_init(&(sc)->sc_mtx, "carp_softc", \ 260 NULL, MTX_DEF) 261 #define CARP_LOCK_DESTROY(sc) mtx_destroy(&(sc)->sc_mtx) 262 #define CARP_LOCK_ASSERT(sc) mtx_assert(&(sc)->sc_mtx, MA_OWNED) 263 #define CARP_LOCK(sc) mtx_lock(&(sc)->sc_mtx) 264 #define CARP_UNLOCK(sc) mtx_unlock(&(sc)->sc_mtx) 265 #define CIF_LOCK_INIT(cif) mtx_init(&(cif)->cif_mtx, "carp_if", \ 266 NULL, MTX_DEF) 267 #define CIF_LOCK_DESTROY(cif) mtx_destroy(&(cif)->cif_mtx) 268 #define CIF_LOCK_ASSERT(cif) mtx_assert(&(cif)->cif_mtx, MA_OWNED) 269 #define CIF_LOCK(cif) mtx_lock(&(cif)->cif_mtx) 270 #define CIF_UNLOCK(cif) mtx_unlock(&(cif)->cif_mtx) 271 #define CIF_FREE(cif) do { \ 272 CIF_LOCK(cif); \ 273 if (TAILQ_EMPTY(&(cif)->cif_vrs)) \ 274 carp_free_if(cif); \ 275 else \ 276 CIF_UNLOCK(cif); \ 277 } while (0) 278 279 #define CARP_LOG(...) do { \ 280 if (V_carp_log > 0) \ 281 log(LOG_INFO, "carp: " __VA_ARGS__); \ 282 } while (0) 283 284 #define CARP_DEBUG(...) do { \ 285 if (V_carp_log > 1) \ 286 log(LOG_DEBUG, __VA_ARGS__); \ 287 } while (0) 288 289 #define IFNET_FOREACH_IFA(ifp, ifa) \ 290 CK_STAILQ_FOREACH((ifa), &(ifp)->if_addrhead, ifa_link) \ 291 if ((ifa)->ifa_carp != NULL) 292 293 #define CARP_FOREACH_IFA(sc, ifa) \ 294 CARP_LOCK_ASSERT(sc); \ 295 for (int _i = 0; \ 296 _i < (sc)->sc_naddrs + (sc)->sc_naddrs6 && \ 297 ((ifa) = sc->sc_ifas[_i]) != NULL; \ 298 ++_i) 299 300 #define IFNET_FOREACH_CARP(ifp, sc) \ 301 KASSERT(mtx_owned(&ifp->if_carp->cif_mtx) || \ 302 sx_xlocked(&carp_sx), ("cif_vrs not locked")); \ 303 TAILQ_FOREACH((sc), &(ifp)->if_carp->cif_vrs, sc_list) 304 305 #define DEMOTE_ADVSKEW(sc) \ 306 (((sc)->sc_advskew + V_carp_demotion > CARP_MAXSKEW) ? \ 307 CARP_MAXSKEW : \ 308 (((sc)->sc_advskew + V_carp_demotion < 0) ? \ 309 0 : ((sc)->sc_advskew + V_carp_demotion))) 310 311 static void carp_input_c(struct mbuf *, struct carp_header *, sa_family_t); 312 static struct carp_softc 313 *carp_alloc(struct ifnet *); 314 static void carp_destroy(struct carp_softc *); 315 static struct carp_if 316 *carp_alloc_if(struct ifnet *); 317 static void carp_free_if(struct carp_if *); 318 static void carp_set_state(struct carp_softc *, int, const char* reason); 319 static void carp_sc_state(struct carp_softc *); 320 static void carp_setrun(struct carp_softc *, sa_family_t); 321 static void carp_master_down(void *); 322 static void carp_master_down_locked(struct carp_softc *, 323 const char* reason); 324 static void carp_send_ad(void *); 325 static void carp_send_ad_locked(struct carp_softc *); 326 static void carp_addroute(struct carp_softc *); 327 static void carp_ifa_addroute(struct ifaddr *); 328 static void carp_delroute(struct carp_softc *); 329 static void carp_ifa_delroute(struct ifaddr *); 330 static void carp_send_ad_all(void *, int); 331 static void carp_demote_adj(int, char *); 332 333 static LIST_HEAD(, carp_softc) carp_list; 334 static struct mtx carp_mtx; 335 static struct sx carp_sx; 336 static struct task carp_sendall_task = 337 TASK_INITIALIZER(0, carp_send_ad_all, NULL); 338 339 static void 340 carp_hmac_prepare(struct carp_softc *sc) 341 { 342 uint8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT; 343 uint8_t vhid = sc->sc_vhid & 0xff; 344 struct ifaddr *ifa; 345 int i, found; 346 #ifdef INET 347 struct in_addr last, cur, in; 348 #endif 349 #ifdef INET6 350 struct in6_addr last6, cur6, in6; 351 #endif 352 353 CARP_LOCK_ASSERT(sc); 354 355 /* Compute ipad from key. */ 356 bzero(sc->sc_pad, sizeof(sc->sc_pad)); 357 bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key)); 358 for (i = 0; i < sizeof(sc->sc_pad); i++) 359 sc->sc_pad[i] ^= 0x36; 360 361 /* Precompute first part of inner hash. */ 362 SHA1Init(&sc->sc_sha1); 363 SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad)); 364 SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version)); 365 SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type)); 366 SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid)); 367 #ifdef INET 368 cur.s_addr = 0; 369 do { 370 found = 0; 371 last = cur; 372 cur.s_addr = 0xffffffff; 373 CARP_FOREACH_IFA(sc, ifa) { 374 in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr; 375 if (ifa->ifa_addr->sa_family == AF_INET && 376 ntohl(in.s_addr) > ntohl(last.s_addr) && 377 ntohl(in.s_addr) < ntohl(cur.s_addr)) { 378 cur.s_addr = in.s_addr; 379 found++; 380 } 381 } 382 if (found) 383 SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur)); 384 } while (found); 385 #endif /* INET */ 386 #ifdef INET6 387 memset(&cur6, 0, sizeof(cur6)); 388 do { 389 found = 0; 390 last6 = cur6; 391 memset(&cur6, 0xff, sizeof(cur6)); 392 CARP_FOREACH_IFA(sc, ifa) { 393 in6 = ifatoia6(ifa)->ia_addr.sin6_addr; 394 if (IN6_IS_SCOPE_EMBED(&in6)) 395 in6.s6_addr16[1] = 0; 396 if (ifa->ifa_addr->sa_family == AF_INET6 && 397 memcmp(&in6, &last6, sizeof(in6)) > 0 && 398 memcmp(&in6, &cur6, sizeof(in6)) < 0) { 399 cur6 = in6; 400 found++; 401 } 402 } 403 if (found) 404 SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6)); 405 } while (found); 406 #endif /* INET6 */ 407 408 /* convert ipad to opad */ 409 for (i = 0; i < sizeof(sc->sc_pad); i++) 410 sc->sc_pad[i] ^= 0x36 ^ 0x5c; 411 } 412 413 static void 414 carp_hmac_generate(struct carp_softc *sc, uint32_t counter[2], 415 unsigned char md[20]) 416 { 417 SHA1_CTX sha1ctx; 418 419 CARP_LOCK_ASSERT(sc); 420 421 /* fetch first half of inner hash */ 422 bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx)); 423 424 SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter)); 425 SHA1Final(md, &sha1ctx); 426 427 /* outer hash */ 428 SHA1Init(&sha1ctx); 429 SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad)); 430 SHA1Update(&sha1ctx, md, 20); 431 SHA1Final(md, &sha1ctx); 432 } 433 434 static int 435 carp_hmac_verify(struct carp_softc *sc, uint32_t counter[2], 436 unsigned char md[20]) 437 { 438 unsigned char md2[20]; 439 440 CARP_LOCK_ASSERT(sc); 441 442 carp_hmac_generate(sc, counter, md2); 443 444 return (bcmp(md, md2, sizeof(md2))); 445 } 446 447 /* 448 * process input packet. 449 * we have rearranged checks order compared to the rfc, 450 * but it seems more efficient this way or not possible otherwise. 451 */ 452 #ifdef INET 453 int 454 carp_input(struct mbuf **mp, int *offp, int proto) 455 { 456 struct mbuf *m = *mp; 457 struct ip *ip = mtod(m, struct ip *); 458 struct carp_header *ch; 459 int iplen, len; 460 461 iplen = *offp; 462 *mp = NULL; 463 464 CARPSTATS_INC(carps_ipackets); 465 466 if (!V_carp_allow) { 467 m_freem(m); 468 return (IPPROTO_DONE); 469 } 470 471 /* verify that the IP TTL is 255. */ 472 if (ip->ip_ttl != CARP_DFLTTL) { 473 CARPSTATS_INC(carps_badttl); 474 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__, 475 ip->ip_ttl, 476 m->m_pkthdr.rcvif->if_xname); 477 m_freem(m); 478 return (IPPROTO_DONE); 479 } 480 481 iplen = ip->ip_hl << 2; 482 483 if (m->m_pkthdr.len < iplen + sizeof(*ch)) { 484 CARPSTATS_INC(carps_badlen); 485 CARP_DEBUG("%s: received len %zd < sizeof(struct carp_header) " 486 "on %s\n", __func__, m->m_len - sizeof(struct ip), 487 m->m_pkthdr.rcvif->if_xname); 488 m_freem(m); 489 return (IPPROTO_DONE); 490 } 491 492 if (iplen + sizeof(*ch) < m->m_len) { 493 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) { 494 CARPSTATS_INC(carps_hdrops); 495 CARP_DEBUG("%s: pullup failed\n", __func__); 496 return (IPPROTO_DONE); 497 } 498 ip = mtod(m, struct ip *); 499 } 500 ch = (struct carp_header *)((char *)ip + iplen); 501 502 /* 503 * verify that the received packet length is 504 * equal to the CARP header 505 */ 506 len = iplen + sizeof(*ch); 507 if (len > m->m_pkthdr.len) { 508 CARPSTATS_INC(carps_badlen); 509 CARP_DEBUG("%s: packet too short %d on %s\n", __func__, 510 m->m_pkthdr.len, 511 m->m_pkthdr.rcvif->if_xname); 512 m_freem(m); 513 return (IPPROTO_DONE); 514 } 515 516 if ((m = m_pullup(m, len)) == NULL) { 517 CARPSTATS_INC(carps_hdrops); 518 return (IPPROTO_DONE); 519 } 520 ip = mtod(m, struct ip *); 521 ch = (struct carp_header *)((char *)ip + iplen); 522 523 /* verify the CARP checksum */ 524 m->m_data += iplen; 525 if (in_cksum(m, len - iplen)) { 526 CARPSTATS_INC(carps_badsum); 527 CARP_DEBUG("%s: checksum failed on %s\n", __func__, 528 m->m_pkthdr.rcvif->if_xname); 529 m_freem(m); 530 return (IPPROTO_DONE); 531 } 532 m->m_data -= iplen; 533 534 carp_input_c(m, ch, AF_INET); 535 return (IPPROTO_DONE); 536 } 537 #endif 538 539 #ifdef INET6 540 int 541 carp6_input(struct mbuf **mp, int *offp, int proto) 542 { 543 struct mbuf *m = *mp; 544 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 545 struct carp_header *ch; 546 u_int len; 547 548 CARPSTATS_INC(carps_ipackets6); 549 550 if (!V_carp_allow) { 551 m_freem(m); 552 return (IPPROTO_DONE); 553 } 554 555 /* check if received on a valid carp interface */ 556 if (m->m_pkthdr.rcvif->if_carp == NULL) { 557 CARPSTATS_INC(carps_badif); 558 CARP_DEBUG("%s: packet received on non-carp interface: %s\n", 559 __func__, m->m_pkthdr.rcvif->if_xname); 560 m_freem(m); 561 return (IPPROTO_DONE); 562 } 563 564 /* verify that the IP TTL is 255 */ 565 if (ip6->ip6_hlim != CARP_DFLTTL) { 566 CARPSTATS_INC(carps_badttl); 567 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__, 568 ip6->ip6_hlim, m->m_pkthdr.rcvif->if_xname); 569 m_freem(m); 570 return (IPPROTO_DONE); 571 } 572 573 /* verify that we have a complete carp packet */ 574 if (m->m_len < *offp + sizeof(*ch)) { 575 len = m->m_len; 576 m = m_pullup(m, *offp + sizeof(*ch)); 577 if (m == NULL) { 578 CARPSTATS_INC(carps_badlen); 579 CARP_DEBUG("%s: packet size %u too small\n", __func__, len); 580 return (IPPROTO_DONE); 581 } 582 } 583 ch = (struct carp_header *)(mtod(m, char *) + *offp); 584 585 /* verify the CARP checksum */ 586 m->m_data += *offp; 587 if (in_cksum(m, sizeof(*ch))) { 588 CARPSTATS_INC(carps_badsum); 589 CARP_DEBUG("%s: checksum failed, on %s\n", __func__, 590 m->m_pkthdr.rcvif->if_xname); 591 m_freem(m); 592 return (IPPROTO_DONE); 593 } 594 m->m_data -= *offp; 595 596 carp_input_c(m, ch, AF_INET6); 597 return (IPPROTO_DONE); 598 } 599 #endif /* INET6 */ 600 601 /* 602 * This routine should not be necessary at all, but some switches 603 * (VMWare ESX vswitches) can echo our own packets back at us, 604 * and we must ignore them or they will cause us to drop out of 605 * MASTER mode. 606 * 607 * We cannot catch all cases of network loops. Instead, what we 608 * do here is catch any packet that arrives with a carp header 609 * with a VHID of 0, that comes from an address that is our own. 610 * These packets are by definition "from us" (even if they are from 611 * a misconfigured host that is pretending to be us). 612 * 613 * The VHID test is outside this mini-function. 614 */ 615 static int 616 carp_source_is_self(struct mbuf *m, struct ifaddr *ifa, sa_family_t af) 617 { 618 #ifdef INET 619 struct ip *ip4; 620 struct in_addr in4; 621 #endif 622 #ifdef INET6 623 struct ip6_hdr *ip6; 624 struct in6_addr in6; 625 #endif 626 627 switch (af) { 628 #ifdef INET 629 case AF_INET: 630 ip4 = mtod(m, struct ip *); 631 in4 = ifatoia(ifa)->ia_addr.sin_addr; 632 return (in4.s_addr == ip4->ip_src.s_addr); 633 #endif 634 #ifdef INET6 635 case AF_INET6: 636 ip6 = mtod(m, struct ip6_hdr *); 637 in6 = ifatoia6(ifa)->ia_addr.sin6_addr; 638 return (memcmp(&in6, &ip6->ip6_src, sizeof(in6)) == 0); 639 #endif 640 default: 641 break; 642 } 643 return (0); 644 } 645 646 static void 647 carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af) 648 { 649 struct ifnet *ifp = m->m_pkthdr.rcvif; 650 struct ifaddr *ifa, *match; 651 struct carp_softc *sc; 652 uint64_t tmp_counter; 653 struct timeval sc_tv, ch_tv; 654 int error; 655 656 NET_EPOCH_ASSERT(); 657 658 /* 659 * Verify that the VHID is valid on the receiving interface. 660 * 661 * There should be just one match. If there are none 662 * the VHID is not valid and we drop the packet. If 663 * there are multiple VHID matches, take just the first 664 * one, for compatibility with previous code. While we're 665 * scanning, check for obvious loops in the network topology 666 * (these should never happen, and as noted above, we may 667 * miss real loops; this is just a double-check). 668 */ 669 error = 0; 670 match = NULL; 671 IFNET_FOREACH_IFA(ifp, ifa) { 672 if (match == NULL && ifa->ifa_carp != NULL && 673 ifa->ifa_addr->sa_family == af && 674 ifa->ifa_carp->sc_vhid == ch->carp_vhid) 675 match = ifa; 676 if (ch->carp_vhid == 0 && carp_source_is_self(m, ifa, af)) 677 error = ELOOP; 678 } 679 ifa = error ? NULL : match; 680 if (ifa != NULL) 681 ifa_ref(ifa); 682 683 if (ifa == NULL) { 684 if (error == ELOOP) { 685 CARP_DEBUG("dropping looped packet on interface %s\n", 686 ifp->if_xname); 687 CARPSTATS_INC(carps_badif); /* ??? */ 688 } else { 689 CARPSTATS_INC(carps_badvhid); 690 } 691 m_freem(m); 692 return; 693 } 694 695 /* verify the CARP version. */ 696 if (ch->carp_version != CARP_VERSION) { 697 CARPSTATS_INC(carps_badver); 698 CARP_DEBUG("%s: invalid version %d\n", ifp->if_xname, 699 ch->carp_version); 700 ifa_free(ifa); 701 m_freem(m); 702 return; 703 } 704 705 sc = ifa->ifa_carp; 706 CARP_LOCK(sc); 707 ifa_free(ifa); 708 709 if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) { 710 CARPSTATS_INC(carps_badauth); 711 CARP_DEBUG("%s: incorrect hash for VHID %u@%s\n", __func__, 712 sc->sc_vhid, ifp->if_xname); 713 goto out; 714 } 715 716 tmp_counter = ntohl(ch->carp_counter[0]); 717 tmp_counter = tmp_counter<<32; 718 tmp_counter += ntohl(ch->carp_counter[1]); 719 720 /* XXX Replay protection goes here */ 721 722 sc->sc_init_counter = 0; 723 sc->sc_counter = tmp_counter; 724 725 sc_tv.tv_sec = sc->sc_advbase; 726 sc_tv.tv_usec = DEMOTE_ADVSKEW(sc) * 1000000 / 256; 727 ch_tv.tv_sec = ch->carp_advbase; 728 ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256; 729 730 switch (sc->sc_state) { 731 case INIT: 732 break; 733 case MASTER: 734 /* 735 * If we receive an advertisement from a master who's going to 736 * be more frequent than us, go into BACKUP state. 737 */ 738 if (timevalcmp(&sc_tv, &ch_tv, >) || 739 timevalcmp(&sc_tv, &ch_tv, ==)) { 740 callout_stop(&sc->sc_ad_tmo); 741 carp_set_state(sc, BACKUP, 742 "more frequent advertisement received"); 743 carp_setrun(sc, 0); 744 carp_delroute(sc); 745 } 746 break; 747 case BACKUP: 748 /* 749 * If we're pre-empting masters who advertise slower than us, 750 * and this one claims to be slower, treat him as down. 751 */ 752 if (V_carp_preempt && timevalcmp(&sc_tv, &ch_tv, <)) { 753 carp_master_down_locked(sc, 754 "preempting a slower master"); 755 break; 756 } 757 758 /* 759 * If the master is going to advertise at such a low frequency 760 * that he's guaranteed to time out, we'd might as well just 761 * treat him as timed out now. 762 */ 763 sc_tv.tv_sec = sc->sc_advbase * 3; 764 if (timevalcmp(&sc_tv, &ch_tv, <)) { 765 carp_master_down_locked(sc, "master will time out"); 766 break; 767 } 768 769 /* 770 * Otherwise, we reset the counter and wait for the next 771 * advertisement. 772 */ 773 carp_setrun(sc, af); 774 break; 775 } 776 777 out: 778 CARP_UNLOCK(sc); 779 m_freem(m); 780 } 781 782 static int 783 carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch) 784 { 785 struct m_tag *mtag; 786 787 if (sc->sc_init_counter) { 788 /* this could also be seconds since unix epoch */ 789 sc->sc_counter = arc4random(); 790 sc->sc_counter = sc->sc_counter << 32; 791 sc->sc_counter += arc4random(); 792 } else 793 sc->sc_counter++; 794 795 ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff); 796 ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff); 797 798 carp_hmac_generate(sc, ch->carp_counter, ch->carp_md); 799 800 /* Tag packet for carp_output */ 801 if ((mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct carp_softc *), 802 M_NOWAIT)) == NULL) { 803 m_freem(m); 804 CARPSTATS_INC(carps_onomem); 805 return (ENOMEM); 806 } 807 bcopy(&sc, mtag + 1, sizeof(sc)); 808 m_tag_prepend(m, mtag); 809 810 return (0); 811 } 812 813 /* 814 * To avoid LORs and possible recursions this function shouldn't 815 * be called directly, but scheduled via taskqueue. 816 */ 817 static void 818 carp_send_ad_all(void *ctx __unused, int pending __unused) 819 { 820 struct carp_softc *sc; 821 struct epoch_tracker et; 822 823 NET_EPOCH_ENTER(et); 824 mtx_lock(&carp_mtx); 825 LIST_FOREACH(sc, &carp_list, sc_next) 826 if (sc->sc_state == MASTER) { 827 CARP_LOCK(sc); 828 CURVNET_SET(sc->sc_carpdev->if_vnet); 829 carp_send_ad_locked(sc); 830 CURVNET_RESTORE(); 831 CARP_UNLOCK(sc); 832 } 833 mtx_unlock(&carp_mtx); 834 NET_EPOCH_EXIT(et); 835 } 836 837 /* Send a periodic advertisement, executed in callout context. */ 838 static void 839 carp_send_ad(void *v) 840 { 841 struct carp_softc *sc = v; 842 struct epoch_tracker et; 843 844 NET_EPOCH_ENTER(et); 845 CARP_LOCK_ASSERT(sc); 846 CURVNET_SET(sc->sc_carpdev->if_vnet); 847 carp_send_ad_locked(sc); 848 CURVNET_RESTORE(); 849 CARP_UNLOCK(sc); 850 NET_EPOCH_EXIT(et); 851 } 852 853 static void 854 carp_send_ad_error(struct carp_softc *sc, int error) 855 { 856 857 if (error) { 858 if (sc->sc_sendad_errors < INT_MAX) 859 sc->sc_sendad_errors++; 860 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) { 861 static const char fmt[] = "send error %d on %s"; 862 char msg[sizeof(fmt) + IFNAMSIZ]; 863 864 sprintf(msg, fmt, error, sc->sc_carpdev->if_xname); 865 carp_demote_adj(V_carp_senderr_adj, msg); 866 } 867 sc->sc_sendad_success = 0; 868 } else { 869 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS && 870 ++sc->sc_sendad_success >= CARP_SENDAD_MIN_SUCCESS) { 871 static const char fmt[] = "send ok on %s"; 872 char msg[sizeof(fmt) + IFNAMSIZ]; 873 874 sprintf(msg, fmt, sc->sc_carpdev->if_xname); 875 carp_demote_adj(-V_carp_senderr_adj, msg); 876 sc->sc_sendad_errors = 0; 877 } else 878 sc->sc_sendad_errors = 0; 879 } 880 } 881 882 /* 883 * Pick the best ifaddr on the given ifp for sending CARP 884 * advertisements. 885 * 886 * "Best" here is defined by ifa_preferred(). This function is much 887 * much like ifaof_ifpforaddr() except that we just use ifa_preferred(). 888 * 889 * (This could be simplified to return the actual address, except that 890 * it has a different format in AF_INET and AF_INET6.) 891 */ 892 static struct ifaddr * 893 carp_best_ifa(int af, struct ifnet *ifp) 894 { 895 struct ifaddr *ifa, *best; 896 897 NET_EPOCH_ASSERT(); 898 899 if (af >= AF_MAX) 900 return (NULL); 901 best = NULL; 902 CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) { 903 if (ifa->ifa_addr->sa_family == af && 904 (best == NULL || ifa_preferred(best, ifa))) 905 best = ifa; 906 } 907 if (best != NULL) 908 ifa_ref(best); 909 return (best); 910 } 911 912 static void 913 carp_send_ad_locked(struct carp_softc *sc) 914 { 915 struct carp_header ch; 916 struct timeval tv; 917 struct ifaddr *ifa; 918 struct carp_header *ch_ptr; 919 struct mbuf *m; 920 int len, advskew; 921 922 NET_EPOCH_ASSERT(); 923 CARP_LOCK_ASSERT(sc); 924 925 advskew = DEMOTE_ADVSKEW(sc); 926 tv.tv_sec = sc->sc_advbase; 927 tv.tv_usec = advskew * 1000000 / 256; 928 929 ch.carp_version = CARP_VERSION; 930 ch.carp_type = CARP_ADVERTISEMENT; 931 ch.carp_vhid = sc->sc_vhid; 932 ch.carp_advbase = sc->sc_advbase; 933 ch.carp_advskew = advskew; 934 ch.carp_authlen = 7; /* XXX DEFINE */ 935 ch.carp_pad1 = 0; /* must be zero */ 936 ch.carp_cksum = 0; 937 938 /* XXXGL: OpenBSD picks first ifaddr with needed family. */ 939 940 #ifdef INET 941 if (sc->sc_naddrs) { 942 struct ip *ip; 943 944 m = m_gethdr(M_NOWAIT, MT_DATA); 945 if (m == NULL) { 946 CARPSTATS_INC(carps_onomem); 947 goto resched; 948 } 949 len = sizeof(*ip) + sizeof(ch); 950 m->m_pkthdr.len = len; 951 m->m_pkthdr.rcvif = NULL; 952 m->m_len = len; 953 M_ALIGN(m, m->m_len); 954 m->m_flags |= M_MCAST; 955 ip = mtod(m, struct ip *); 956 ip->ip_v = IPVERSION; 957 ip->ip_hl = sizeof(*ip) >> 2; 958 ip->ip_tos = V_carp_dscp << IPTOS_DSCP_OFFSET; 959 ip->ip_len = htons(len); 960 ip->ip_off = htons(IP_DF); 961 ip->ip_ttl = CARP_DFLTTL; 962 ip->ip_p = IPPROTO_CARP; 963 ip->ip_sum = 0; 964 ip_fillid(ip); 965 966 ifa = carp_best_ifa(AF_INET, sc->sc_carpdev); 967 if (ifa != NULL) { 968 ip->ip_src.s_addr = 969 ifatoia(ifa)->ia_addr.sin_addr.s_addr; 970 ifa_free(ifa); 971 } else 972 ip->ip_src.s_addr = 0; 973 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP); 974 975 ch_ptr = (struct carp_header *)(&ip[1]); 976 bcopy(&ch, ch_ptr, sizeof(ch)); 977 if (carp_prepare_ad(m, sc, ch_ptr)) 978 goto resched; 979 980 m->m_data += sizeof(*ip); 981 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip)); 982 m->m_data -= sizeof(*ip); 983 984 CARPSTATS_INC(carps_opackets); 985 986 carp_send_ad_error(sc, ip_output(m, NULL, NULL, IP_RAWOUTPUT, 987 &sc->sc_carpdev->if_carp->cif_imo, NULL)); 988 } 989 #endif /* INET */ 990 #ifdef INET6 991 if (sc->sc_naddrs6) { 992 struct ip6_hdr *ip6; 993 994 m = m_gethdr(M_NOWAIT, MT_DATA); 995 if (m == NULL) { 996 CARPSTATS_INC(carps_onomem); 997 goto resched; 998 } 999 len = sizeof(*ip6) + sizeof(ch); 1000 m->m_pkthdr.len = len; 1001 m->m_pkthdr.rcvif = NULL; 1002 m->m_len = len; 1003 M_ALIGN(m, m->m_len); 1004 m->m_flags |= M_MCAST; 1005 ip6 = mtod(m, struct ip6_hdr *); 1006 bzero(ip6, sizeof(*ip6)); 1007 ip6->ip6_vfc |= IPV6_VERSION; 1008 /* Traffic class isn't defined in ip6 struct instead 1009 * it gets offset into flowid field */ 1010 ip6->ip6_flow |= htonl(V_carp_dscp << (IPV6_FLOWLABEL_LEN + 1011 IPTOS_DSCP_OFFSET)); 1012 ip6->ip6_hlim = CARP_DFLTTL; 1013 ip6->ip6_nxt = IPPROTO_CARP; 1014 1015 /* set the source address */ 1016 ifa = carp_best_ifa(AF_INET6, sc->sc_carpdev); 1017 if (ifa != NULL) { 1018 bcopy(IFA_IN6(ifa), &ip6->ip6_src, 1019 sizeof(struct in6_addr)); 1020 ifa_free(ifa); 1021 } else 1022 /* This should never happen with IPv6. */ 1023 bzero(&ip6->ip6_src, sizeof(struct in6_addr)); 1024 1025 /* Set the multicast destination. */ 1026 ip6->ip6_dst.s6_addr16[0] = htons(0xff02); 1027 ip6->ip6_dst.s6_addr8[15] = 0x12; 1028 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) { 1029 m_freem(m); 1030 CARP_DEBUG("%s: in6_setscope failed\n", __func__); 1031 goto resched; 1032 } 1033 1034 ch_ptr = (struct carp_header *)(&ip6[1]); 1035 bcopy(&ch, ch_ptr, sizeof(ch)); 1036 if (carp_prepare_ad(m, sc, ch_ptr)) 1037 goto resched; 1038 1039 m->m_data += sizeof(*ip6); 1040 ch_ptr->carp_cksum = in_cksum(m, len - sizeof(*ip6)); 1041 m->m_data -= sizeof(*ip6); 1042 1043 CARPSTATS_INC(carps_opackets6); 1044 1045 carp_send_ad_error(sc, ip6_output(m, NULL, NULL, 0, 1046 &sc->sc_carpdev->if_carp->cif_im6o, NULL, NULL)); 1047 } 1048 #endif /* INET6 */ 1049 1050 resched: 1051 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), carp_send_ad, sc); 1052 } 1053 1054 static void 1055 carp_addroute(struct carp_softc *sc) 1056 { 1057 struct ifaddr *ifa; 1058 1059 CARP_FOREACH_IFA(sc, ifa) 1060 carp_ifa_addroute(ifa); 1061 } 1062 1063 static void 1064 carp_ifa_addroute(struct ifaddr *ifa) 1065 { 1066 1067 switch (ifa->ifa_addr->sa_family) { 1068 #ifdef INET 1069 case AF_INET: 1070 in_addprefix(ifatoia(ifa)); 1071 ifa_add_loopback_route(ifa, 1072 (struct sockaddr *)&ifatoia(ifa)->ia_addr); 1073 break; 1074 #endif 1075 #ifdef INET6 1076 case AF_INET6: 1077 ifa_add_loopback_route(ifa, 1078 (struct sockaddr *)&ifatoia6(ifa)->ia_addr); 1079 nd6_add_ifa_lle(ifatoia6(ifa)); 1080 break; 1081 #endif 1082 } 1083 } 1084 1085 static void 1086 carp_delroute(struct carp_softc *sc) 1087 { 1088 struct ifaddr *ifa; 1089 1090 CARP_FOREACH_IFA(sc, ifa) 1091 carp_ifa_delroute(ifa); 1092 } 1093 1094 static void 1095 carp_ifa_delroute(struct ifaddr *ifa) 1096 { 1097 1098 switch (ifa->ifa_addr->sa_family) { 1099 #ifdef INET 1100 case AF_INET: 1101 ifa_del_loopback_route(ifa, 1102 (struct sockaddr *)&ifatoia(ifa)->ia_addr); 1103 in_scrubprefix(ifatoia(ifa), LLE_STATIC); 1104 break; 1105 #endif 1106 #ifdef INET6 1107 case AF_INET6: 1108 ifa_del_loopback_route(ifa, 1109 (struct sockaddr *)&ifatoia6(ifa)->ia_addr); 1110 nd6_rem_ifa_lle(ifatoia6(ifa), 1); 1111 break; 1112 #endif 1113 } 1114 } 1115 1116 int 1117 carp_master(struct ifaddr *ifa) 1118 { 1119 struct carp_softc *sc = ifa->ifa_carp; 1120 1121 return (sc->sc_state == MASTER); 1122 } 1123 1124 #ifdef INET 1125 /* 1126 * Broadcast a gratuitous ARP request containing 1127 * the virtual router MAC address for each IP address 1128 * associated with the virtual router. 1129 */ 1130 static void 1131 carp_send_arp(struct carp_softc *sc) 1132 { 1133 struct ifaddr *ifa; 1134 struct in_addr addr; 1135 1136 NET_EPOCH_ASSERT(); 1137 1138 CARP_FOREACH_IFA(sc, ifa) { 1139 if (ifa->ifa_addr->sa_family != AF_INET) 1140 continue; 1141 addr = ((struct sockaddr_in *)ifa->ifa_addr)->sin_addr; 1142 arp_announce_ifaddr(sc->sc_carpdev, addr, LLADDR(&sc->sc_addr)); 1143 } 1144 } 1145 1146 int 1147 carp_iamatch(struct ifaddr *ifa, uint8_t **enaddr) 1148 { 1149 struct carp_softc *sc = ifa->ifa_carp; 1150 1151 if (sc->sc_state == MASTER) { 1152 *enaddr = LLADDR(&sc->sc_addr); 1153 return (1); 1154 } 1155 1156 return (0); 1157 } 1158 #endif 1159 1160 #ifdef INET6 1161 static void 1162 carp_send_na(struct carp_softc *sc) 1163 { 1164 static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT; 1165 struct ifaddr *ifa; 1166 struct in6_addr *in6; 1167 1168 CARP_FOREACH_IFA(sc, ifa) { 1169 if (ifa->ifa_addr->sa_family != AF_INET6) 1170 continue; 1171 1172 in6 = IFA_IN6(ifa); 1173 nd6_na_output(sc->sc_carpdev, &mcast, in6, 1174 ND_NA_FLAG_OVERRIDE, 1, NULL); 1175 DELAY(1000); /* XXX */ 1176 } 1177 } 1178 1179 /* 1180 * Returns ifa in case it's a carp address and it is MASTER, or if the address 1181 * matches and is not a carp address. Returns NULL otherwise. 1182 */ 1183 struct ifaddr * 1184 carp_iamatch6(struct ifnet *ifp, struct in6_addr *taddr) 1185 { 1186 struct ifaddr *ifa; 1187 1188 NET_EPOCH_ASSERT(); 1189 1190 ifa = NULL; 1191 CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) { 1192 if (ifa->ifa_addr->sa_family != AF_INET6) 1193 continue; 1194 if (!IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) 1195 continue; 1196 if (ifa->ifa_carp && ifa->ifa_carp->sc_state != MASTER) 1197 ifa = NULL; 1198 else 1199 ifa_ref(ifa); 1200 break; 1201 } 1202 1203 return (ifa); 1204 } 1205 1206 char * 1207 carp_macmatch6(struct ifnet *ifp, struct mbuf *m, const struct in6_addr *taddr) 1208 { 1209 struct ifaddr *ifa; 1210 1211 NET_EPOCH_ASSERT(); 1212 1213 IFNET_FOREACH_IFA(ifp, ifa) 1214 if (ifa->ifa_addr->sa_family == AF_INET6 && 1215 IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) { 1216 struct carp_softc *sc = ifa->ifa_carp; 1217 struct m_tag *mtag; 1218 1219 mtag = m_tag_get(PACKET_TAG_CARP, 1220 sizeof(struct carp_softc *), M_NOWAIT); 1221 if (mtag == NULL) 1222 /* Better a bit than nothing. */ 1223 return (LLADDR(&sc->sc_addr)); 1224 1225 bcopy(&sc, mtag + 1, sizeof(sc)); 1226 m_tag_prepend(m, mtag); 1227 1228 return (LLADDR(&sc->sc_addr)); 1229 } 1230 1231 return (NULL); 1232 } 1233 #endif /* INET6 */ 1234 1235 int 1236 carp_forus(struct ifnet *ifp, u_char *dhost) 1237 { 1238 struct carp_softc *sc; 1239 uint8_t *ena = dhost; 1240 1241 if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1) 1242 return (0); 1243 1244 CIF_LOCK(ifp->if_carp); 1245 IFNET_FOREACH_CARP(ifp, sc) { 1246 /* 1247 * CARP_LOCK() is not here, since would protect nothing, but 1248 * cause deadlock with if_bridge, calling this under its lock. 1249 */ 1250 if (sc->sc_state == MASTER && !bcmp(dhost, LLADDR(&sc->sc_addr), 1251 ETHER_ADDR_LEN)) { 1252 CIF_UNLOCK(ifp->if_carp); 1253 return (1); 1254 } 1255 } 1256 CIF_UNLOCK(ifp->if_carp); 1257 1258 return (0); 1259 } 1260 1261 /* Master down timeout event, executed in callout context. */ 1262 static void 1263 carp_master_down(void *v) 1264 { 1265 struct carp_softc *sc = v; 1266 struct epoch_tracker et; 1267 1268 NET_EPOCH_ENTER(et); 1269 CARP_LOCK_ASSERT(sc); 1270 1271 CURVNET_SET(sc->sc_carpdev->if_vnet); 1272 if (sc->sc_state == BACKUP) { 1273 carp_master_down_locked(sc, "master timed out"); 1274 } 1275 CURVNET_RESTORE(); 1276 1277 CARP_UNLOCK(sc); 1278 NET_EPOCH_EXIT(et); 1279 } 1280 1281 static void 1282 carp_master_down_locked(struct carp_softc *sc, const char *reason) 1283 { 1284 1285 NET_EPOCH_ASSERT(); 1286 CARP_LOCK_ASSERT(sc); 1287 1288 switch (sc->sc_state) { 1289 case BACKUP: 1290 carp_set_state(sc, MASTER, reason); 1291 carp_send_ad_locked(sc); 1292 #ifdef INET 1293 carp_send_arp(sc); 1294 #endif 1295 #ifdef INET6 1296 carp_send_na(sc); 1297 #endif 1298 carp_setrun(sc, 0); 1299 carp_addroute(sc); 1300 break; 1301 case INIT: 1302 case MASTER: 1303 #ifdef INVARIANTS 1304 panic("carp: VHID %u@%s: master_down event in %s state\n", 1305 sc->sc_vhid, 1306 sc->sc_carpdev->if_xname, 1307 sc->sc_state ? "MASTER" : "INIT"); 1308 #endif 1309 break; 1310 } 1311 } 1312 1313 /* 1314 * When in backup state, af indicates whether to reset the master down timer 1315 * for v4 or v6. If it's set to zero, reset the ones which are already pending. 1316 */ 1317 static void 1318 carp_setrun(struct carp_softc *sc, sa_family_t af) 1319 { 1320 struct timeval tv; 1321 1322 CARP_LOCK_ASSERT(sc); 1323 1324 if ((sc->sc_carpdev->if_flags & IFF_UP) == 0 || 1325 sc->sc_carpdev->if_link_state != LINK_STATE_UP || 1326 (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) || 1327 !V_carp_allow) 1328 return; 1329 1330 switch (sc->sc_state) { 1331 case INIT: 1332 carp_set_state(sc, BACKUP, "initialization complete"); 1333 carp_setrun(sc, 0); 1334 break; 1335 case BACKUP: 1336 callout_stop(&sc->sc_ad_tmo); 1337 tv.tv_sec = 3 * sc->sc_advbase; 1338 tv.tv_usec = sc->sc_advskew * 1000000 / 256; 1339 switch (af) { 1340 #ifdef INET 1341 case AF_INET: 1342 callout_reset(&sc->sc_md_tmo, tvtohz(&tv), 1343 carp_master_down, sc); 1344 break; 1345 #endif 1346 #ifdef INET6 1347 case AF_INET6: 1348 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv), 1349 carp_master_down, sc); 1350 break; 1351 #endif 1352 default: 1353 #ifdef INET 1354 if (sc->sc_naddrs) 1355 callout_reset(&sc->sc_md_tmo, tvtohz(&tv), 1356 carp_master_down, sc); 1357 #endif 1358 #ifdef INET6 1359 if (sc->sc_naddrs6) 1360 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv), 1361 carp_master_down, sc); 1362 #endif 1363 break; 1364 } 1365 break; 1366 case MASTER: 1367 tv.tv_sec = sc->sc_advbase; 1368 tv.tv_usec = sc->sc_advskew * 1000000 / 256; 1369 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), 1370 carp_send_ad, sc); 1371 break; 1372 } 1373 } 1374 1375 /* 1376 * Setup multicast structures. 1377 */ 1378 static int 1379 carp_multicast_setup(struct carp_if *cif, sa_family_t sa) 1380 { 1381 struct ifnet *ifp = cif->cif_ifp; 1382 int error = 0; 1383 1384 switch (sa) { 1385 #ifdef INET 1386 case AF_INET: 1387 { 1388 struct ip_moptions *imo = &cif->cif_imo; 1389 struct in_mfilter *imf; 1390 struct in_addr addr; 1391 1392 if (ip_mfilter_first(&imo->imo_head) != NULL) 1393 return (0); 1394 1395 imf = ip_mfilter_alloc(M_WAITOK, 0, 0); 1396 ip_mfilter_init(&imo->imo_head); 1397 imo->imo_multicast_vif = -1; 1398 1399 addr.s_addr = htonl(INADDR_CARP_GROUP); 1400 if ((error = in_joingroup(ifp, &addr, NULL, 1401 &imf->imf_inm)) != 0) { 1402 ip_mfilter_free(imf); 1403 break; 1404 } 1405 1406 ip_mfilter_insert(&imo->imo_head, imf); 1407 imo->imo_multicast_ifp = ifp; 1408 imo->imo_multicast_ttl = CARP_DFLTTL; 1409 imo->imo_multicast_loop = 0; 1410 break; 1411 } 1412 #endif 1413 #ifdef INET6 1414 case AF_INET6: 1415 { 1416 struct ip6_moptions *im6o = &cif->cif_im6o; 1417 struct in6_mfilter *im6f[2]; 1418 struct in6_addr in6; 1419 1420 if (ip6_mfilter_first(&im6o->im6o_head)) 1421 return (0); 1422 1423 im6f[0] = ip6_mfilter_alloc(M_WAITOK, 0, 0); 1424 im6f[1] = ip6_mfilter_alloc(M_WAITOK, 0, 0); 1425 1426 ip6_mfilter_init(&im6o->im6o_head); 1427 im6o->im6o_multicast_hlim = CARP_DFLTTL; 1428 im6o->im6o_multicast_ifp = ifp; 1429 1430 /* Join IPv6 CARP multicast group. */ 1431 bzero(&in6, sizeof(in6)); 1432 in6.s6_addr16[0] = htons(0xff02); 1433 in6.s6_addr8[15] = 0x12; 1434 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) { 1435 ip6_mfilter_free(im6f[0]); 1436 ip6_mfilter_free(im6f[1]); 1437 break; 1438 } 1439 if ((error = in6_joingroup(ifp, &in6, NULL, &im6f[0]->im6f_in6m, 0)) != 0) { 1440 ip6_mfilter_free(im6f[0]); 1441 ip6_mfilter_free(im6f[1]); 1442 break; 1443 } 1444 1445 /* Join solicited multicast address. */ 1446 bzero(&in6, sizeof(in6)); 1447 in6.s6_addr16[0] = htons(0xff02); 1448 in6.s6_addr32[1] = 0; 1449 in6.s6_addr32[2] = htonl(1); 1450 in6.s6_addr32[3] = 0; 1451 in6.s6_addr8[12] = 0xff; 1452 1453 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) { 1454 ip6_mfilter_free(im6f[0]); 1455 ip6_mfilter_free(im6f[1]); 1456 break; 1457 } 1458 1459 if ((error = in6_joingroup(ifp, &in6, NULL, &im6f[1]->im6f_in6m, 0)) != 0) { 1460 in6_leavegroup(im6f[0]->im6f_in6m, NULL); 1461 ip6_mfilter_free(im6f[0]); 1462 ip6_mfilter_free(im6f[1]); 1463 break; 1464 } 1465 ip6_mfilter_insert(&im6o->im6o_head, im6f[0]); 1466 ip6_mfilter_insert(&im6o->im6o_head, im6f[1]); 1467 break; 1468 } 1469 #endif 1470 } 1471 1472 return (error); 1473 } 1474 1475 /* 1476 * Free multicast structures. 1477 */ 1478 static void 1479 carp_multicast_cleanup(struct carp_if *cif, sa_family_t sa) 1480 { 1481 #ifdef INET 1482 struct ip_moptions *imo = &cif->cif_imo; 1483 struct in_mfilter *imf; 1484 #endif 1485 #ifdef INET6 1486 struct ip6_moptions *im6o = &cif->cif_im6o; 1487 struct in6_mfilter *im6f; 1488 #endif 1489 sx_assert(&carp_sx, SA_XLOCKED); 1490 1491 switch (sa) { 1492 #ifdef INET 1493 case AF_INET: 1494 if (cif->cif_naddrs != 0) 1495 break; 1496 1497 while ((imf = ip_mfilter_first(&imo->imo_head)) != NULL) { 1498 ip_mfilter_remove(&imo->imo_head, imf); 1499 in_leavegroup(imf->imf_inm, NULL); 1500 ip_mfilter_free(imf); 1501 } 1502 break; 1503 #endif 1504 #ifdef INET6 1505 case AF_INET6: 1506 if (cif->cif_naddrs6 != 0) 1507 break; 1508 1509 while ((im6f = ip6_mfilter_first(&im6o->im6o_head)) != NULL) { 1510 ip6_mfilter_remove(&im6o->im6o_head, im6f); 1511 in6_leavegroup(im6f->im6f_in6m, NULL); 1512 ip6_mfilter_free(im6f); 1513 } 1514 break; 1515 #endif 1516 } 1517 } 1518 1519 int 1520 carp_output(struct ifnet *ifp, struct mbuf *m, const struct sockaddr *sa) 1521 { 1522 struct m_tag *mtag; 1523 struct carp_softc *sc; 1524 1525 if (!sa) 1526 return (0); 1527 1528 switch (sa->sa_family) { 1529 #ifdef INET 1530 case AF_INET: 1531 break; 1532 #endif 1533 #ifdef INET6 1534 case AF_INET6: 1535 break; 1536 #endif 1537 default: 1538 return (0); 1539 } 1540 1541 mtag = m_tag_find(m, PACKET_TAG_CARP, NULL); 1542 if (mtag == NULL) 1543 return (0); 1544 1545 bcopy(mtag + 1, &sc, sizeof(sc)); 1546 1547 /* Set the source MAC address to the Virtual Router MAC Address. */ 1548 switch (ifp->if_type) { 1549 case IFT_ETHER: 1550 case IFT_BRIDGE: 1551 case IFT_L2VLAN: { 1552 struct ether_header *eh; 1553 1554 eh = mtod(m, struct ether_header *); 1555 eh->ether_shost[0] = 0; 1556 eh->ether_shost[1] = 0; 1557 eh->ether_shost[2] = 0x5e; 1558 eh->ether_shost[3] = 0; 1559 eh->ether_shost[4] = 1; 1560 eh->ether_shost[5] = sc->sc_vhid; 1561 } 1562 break; 1563 default: 1564 printf("%s: carp is not supported for the %d interface type\n", 1565 ifp->if_xname, ifp->if_type); 1566 return (EOPNOTSUPP); 1567 } 1568 1569 return (0); 1570 } 1571 1572 static struct carp_softc* 1573 carp_alloc(struct ifnet *ifp) 1574 { 1575 struct carp_softc *sc; 1576 struct carp_if *cif; 1577 1578 sx_assert(&carp_sx, SA_XLOCKED); 1579 1580 if ((cif = ifp->if_carp) == NULL) 1581 cif = carp_alloc_if(ifp); 1582 1583 sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO); 1584 1585 sc->sc_advbase = CARP_DFLTINTV; 1586 sc->sc_vhid = -1; /* required setting */ 1587 sc->sc_init_counter = 1; 1588 sc->sc_state = INIT; 1589 1590 sc->sc_ifasiz = sizeof(struct ifaddr *); 1591 sc->sc_ifas = malloc(sc->sc_ifasiz, M_CARP, M_WAITOK|M_ZERO); 1592 sc->sc_carpdev = ifp; 1593 1594 CARP_LOCK_INIT(sc); 1595 #ifdef INET 1596 callout_init_mtx(&sc->sc_md_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED); 1597 #endif 1598 #ifdef INET6 1599 callout_init_mtx(&sc->sc_md6_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED); 1600 #endif 1601 callout_init_mtx(&sc->sc_ad_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED); 1602 1603 CIF_LOCK(cif); 1604 TAILQ_INSERT_TAIL(&cif->cif_vrs, sc, sc_list); 1605 CIF_UNLOCK(cif); 1606 1607 mtx_lock(&carp_mtx); 1608 LIST_INSERT_HEAD(&carp_list, sc, sc_next); 1609 mtx_unlock(&carp_mtx); 1610 1611 return (sc); 1612 } 1613 1614 static void 1615 carp_grow_ifas(struct carp_softc *sc) 1616 { 1617 struct ifaddr **new; 1618 1619 new = malloc(sc->sc_ifasiz * 2, M_CARP, M_WAITOK | M_ZERO); 1620 CARP_LOCK(sc); 1621 bcopy(sc->sc_ifas, new, sc->sc_ifasiz); 1622 free(sc->sc_ifas, M_CARP); 1623 sc->sc_ifas = new; 1624 sc->sc_ifasiz *= 2; 1625 CARP_UNLOCK(sc); 1626 } 1627 1628 static void 1629 carp_destroy(struct carp_softc *sc) 1630 { 1631 struct ifnet *ifp = sc->sc_carpdev; 1632 struct carp_if *cif = ifp->if_carp; 1633 1634 sx_assert(&carp_sx, SA_XLOCKED); 1635 1636 if (sc->sc_suppress) 1637 carp_demote_adj(-V_carp_ifdown_adj, "vhid removed"); 1638 CARP_UNLOCK(sc); 1639 1640 CIF_LOCK(cif); 1641 TAILQ_REMOVE(&cif->cif_vrs, sc, sc_list); 1642 CIF_UNLOCK(cif); 1643 1644 mtx_lock(&carp_mtx); 1645 LIST_REMOVE(sc, sc_next); 1646 mtx_unlock(&carp_mtx); 1647 1648 callout_drain(&sc->sc_ad_tmo); 1649 #ifdef INET 1650 callout_drain(&sc->sc_md_tmo); 1651 #endif 1652 #ifdef INET6 1653 callout_drain(&sc->sc_md6_tmo); 1654 #endif 1655 CARP_LOCK_DESTROY(sc); 1656 1657 free(sc->sc_ifas, M_CARP); 1658 free(sc, M_CARP); 1659 } 1660 1661 static struct carp_if* 1662 carp_alloc_if(struct ifnet *ifp) 1663 { 1664 struct carp_if *cif; 1665 int error; 1666 1667 cif = malloc(sizeof(*cif), M_CARP, M_WAITOK|M_ZERO); 1668 1669 if ((error = ifpromisc(ifp, 1)) != 0) 1670 printf("%s: ifpromisc(%s) failed: %d\n", 1671 __func__, ifp->if_xname, error); 1672 else 1673 cif->cif_flags |= CIF_PROMISC; 1674 1675 CIF_LOCK_INIT(cif); 1676 cif->cif_ifp = ifp; 1677 TAILQ_INIT(&cif->cif_vrs); 1678 1679 IF_ADDR_WLOCK(ifp); 1680 ifp->if_carp = cif; 1681 if_ref(ifp); 1682 IF_ADDR_WUNLOCK(ifp); 1683 1684 return (cif); 1685 } 1686 1687 static void 1688 carp_free_if(struct carp_if *cif) 1689 { 1690 struct ifnet *ifp = cif->cif_ifp; 1691 1692 CIF_LOCK_ASSERT(cif); 1693 KASSERT(TAILQ_EMPTY(&cif->cif_vrs), ("%s: softc list not empty", 1694 __func__)); 1695 1696 IF_ADDR_WLOCK(ifp); 1697 ifp->if_carp = NULL; 1698 IF_ADDR_WUNLOCK(ifp); 1699 1700 CIF_LOCK_DESTROY(cif); 1701 1702 if (cif->cif_flags & CIF_PROMISC) 1703 ifpromisc(ifp, 0); 1704 if_rele(ifp); 1705 1706 free(cif, M_CARP); 1707 } 1708 1709 static void 1710 carp_carprcp(struct carpreq *carpr, struct carp_softc *sc, int priv) 1711 { 1712 1713 CARP_LOCK(sc); 1714 carpr->carpr_state = sc->sc_state; 1715 carpr->carpr_vhid = sc->sc_vhid; 1716 carpr->carpr_advbase = sc->sc_advbase; 1717 carpr->carpr_advskew = sc->sc_advskew; 1718 if (priv) 1719 bcopy(sc->sc_key, carpr->carpr_key, sizeof(carpr->carpr_key)); 1720 else 1721 bzero(carpr->carpr_key, sizeof(carpr->carpr_key)); 1722 CARP_UNLOCK(sc); 1723 } 1724 1725 int 1726 carp_ioctl(struct ifreq *ifr, u_long cmd, struct thread *td) 1727 { 1728 struct carpreq carpr; 1729 struct ifnet *ifp; 1730 struct carp_softc *sc = NULL; 1731 int error = 0, locked = 0; 1732 1733 if ((error = copyin(ifr_data_get_ptr(ifr), &carpr, sizeof carpr))) 1734 return (error); 1735 1736 ifp = ifunit_ref(ifr->ifr_name); 1737 if (ifp == NULL) 1738 return (ENXIO); 1739 1740 switch (ifp->if_type) { 1741 case IFT_ETHER: 1742 case IFT_L2VLAN: 1743 case IFT_BRIDGE: 1744 break; 1745 default: 1746 error = EOPNOTSUPP; 1747 goto out; 1748 } 1749 1750 if ((ifp->if_flags & IFF_MULTICAST) == 0) { 1751 error = EADDRNOTAVAIL; 1752 goto out; 1753 } 1754 1755 sx_xlock(&carp_sx); 1756 switch (cmd) { 1757 case SIOCSVH: 1758 if ((error = priv_check(td, PRIV_NETINET_CARP))) 1759 break; 1760 if (carpr.carpr_vhid <= 0 || carpr.carpr_vhid > CARP_MAXVHID || 1761 carpr.carpr_advbase < 0 || carpr.carpr_advskew < 0) { 1762 error = EINVAL; 1763 break; 1764 } 1765 1766 if (ifp->if_carp) { 1767 IFNET_FOREACH_CARP(ifp, sc) 1768 if (sc->sc_vhid == carpr.carpr_vhid) 1769 break; 1770 } 1771 if (sc == NULL) { 1772 sc = carp_alloc(ifp); 1773 CARP_LOCK(sc); 1774 sc->sc_vhid = carpr.carpr_vhid; 1775 LLADDR(&sc->sc_addr)[0] = 0; 1776 LLADDR(&sc->sc_addr)[1] = 0; 1777 LLADDR(&sc->sc_addr)[2] = 0x5e; 1778 LLADDR(&sc->sc_addr)[3] = 0; 1779 LLADDR(&sc->sc_addr)[4] = 1; 1780 LLADDR(&sc->sc_addr)[5] = sc->sc_vhid; 1781 } else 1782 CARP_LOCK(sc); 1783 locked = 1; 1784 if (carpr.carpr_advbase > 0) { 1785 if (carpr.carpr_advbase > 255 || 1786 carpr.carpr_advbase < CARP_DFLTINTV) { 1787 error = EINVAL; 1788 break; 1789 } 1790 sc->sc_advbase = carpr.carpr_advbase; 1791 } 1792 if (carpr.carpr_advskew >= 255) { 1793 error = EINVAL; 1794 break; 1795 } 1796 sc->sc_advskew = carpr.carpr_advskew; 1797 if (carpr.carpr_key[0] != '\0') { 1798 bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key)); 1799 carp_hmac_prepare(sc); 1800 } 1801 if (sc->sc_state != INIT && 1802 carpr.carpr_state != sc->sc_state) { 1803 switch (carpr.carpr_state) { 1804 case BACKUP: 1805 callout_stop(&sc->sc_ad_tmo); 1806 carp_set_state(sc, BACKUP, 1807 "user requested via ifconfig"); 1808 carp_setrun(sc, 0); 1809 carp_delroute(sc); 1810 break; 1811 case MASTER: 1812 carp_master_down_locked(sc, 1813 "user requested via ifconfig"); 1814 break; 1815 default: 1816 break; 1817 } 1818 } 1819 break; 1820 1821 case SIOCGVH: 1822 { 1823 int priveleged; 1824 1825 if (carpr.carpr_vhid < 0 || carpr.carpr_vhid > CARP_MAXVHID) { 1826 error = EINVAL; 1827 break; 1828 } 1829 if (carpr.carpr_count < 1) { 1830 error = EMSGSIZE; 1831 break; 1832 } 1833 if (ifp->if_carp == NULL) { 1834 error = ENOENT; 1835 break; 1836 } 1837 1838 priveleged = (priv_check(td, PRIV_NETINET_CARP) == 0); 1839 if (carpr.carpr_vhid != 0) { 1840 IFNET_FOREACH_CARP(ifp, sc) 1841 if (sc->sc_vhid == carpr.carpr_vhid) 1842 break; 1843 if (sc == NULL) { 1844 error = ENOENT; 1845 break; 1846 } 1847 carp_carprcp(&carpr, sc, priveleged); 1848 error = copyout(&carpr, ifr_data_get_ptr(ifr), 1849 sizeof(carpr)); 1850 } else { 1851 int i, count; 1852 1853 count = 0; 1854 IFNET_FOREACH_CARP(ifp, sc) 1855 count++; 1856 1857 if (count > carpr.carpr_count) { 1858 CIF_UNLOCK(ifp->if_carp); 1859 error = EMSGSIZE; 1860 break; 1861 } 1862 1863 i = 0; 1864 IFNET_FOREACH_CARP(ifp, sc) { 1865 carp_carprcp(&carpr, sc, priveleged); 1866 carpr.carpr_count = count; 1867 error = copyout(&carpr, 1868 (char *)ifr_data_get_ptr(ifr) + 1869 (i * sizeof(carpr)), sizeof(carpr)); 1870 if (error) { 1871 CIF_UNLOCK(ifp->if_carp); 1872 break; 1873 } 1874 i++; 1875 } 1876 } 1877 break; 1878 } 1879 default: 1880 error = EINVAL; 1881 } 1882 sx_xunlock(&carp_sx); 1883 1884 out: 1885 if (locked) 1886 CARP_UNLOCK(sc); 1887 if_rele(ifp); 1888 1889 return (error); 1890 } 1891 1892 static int 1893 carp_get_vhid(struct ifaddr *ifa) 1894 { 1895 1896 if (ifa == NULL || ifa->ifa_carp == NULL) 1897 return (0); 1898 1899 return (ifa->ifa_carp->sc_vhid); 1900 } 1901 1902 int 1903 carp_attach(struct ifaddr *ifa, int vhid) 1904 { 1905 struct ifnet *ifp = ifa->ifa_ifp; 1906 struct carp_if *cif = ifp->if_carp; 1907 struct carp_softc *sc; 1908 int index, error; 1909 1910 KASSERT(ifa->ifa_carp == NULL, ("%s: ifa %p attached", __func__, ifa)); 1911 1912 switch (ifa->ifa_addr->sa_family) { 1913 #ifdef INET 1914 case AF_INET: 1915 #endif 1916 #ifdef INET6 1917 case AF_INET6: 1918 #endif 1919 break; 1920 default: 1921 return (EPROTOTYPE); 1922 } 1923 1924 sx_xlock(&carp_sx); 1925 if (ifp->if_carp == NULL) { 1926 sx_xunlock(&carp_sx); 1927 return (ENOPROTOOPT); 1928 } 1929 1930 IFNET_FOREACH_CARP(ifp, sc) 1931 if (sc->sc_vhid == vhid) 1932 break; 1933 if (sc == NULL) { 1934 sx_xunlock(&carp_sx); 1935 return (ENOENT); 1936 } 1937 1938 error = carp_multicast_setup(cif, ifa->ifa_addr->sa_family); 1939 if (error) { 1940 CIF_FREE(cif); 1941 sx_xunlock(&carp_sx); 1942 return (error); 1943 } 1944 1945 index = sc->sc_naddrs + sc->sc_naddrs6 + 1; 1946 if (index > sc->sc_ifasiz / sizeof(struct ifaddr *)) 1947 carp_grow_ifas(sc); 1948 1949 switch (ifa->ifa_addr->sa_family) { 1950 #ifdef INET 1951 case AF_INET: 1952 cif->cif_naddrs++; 1953 sc->sc_naddrs++; 1954 break; 1955 #endif 1956 #ifdef INET6 1957 case AF_INET6: 1958 cif->cif_naddrs6++; 1959 sc->sc_naddrs6++; 1960 break; 1961 #endif 1962 } 1963 1964 ifa_ref(ifa); 1965 1966 CARP_LOCK(sc); 1967 sc->sc_ifas[index - 1] = ifa; 1968 ifa->ifa_carp = sc; 1969 carp_hmac_prepare(sc); 1970 carp_sc_state(sc); 1971 CARP_UNLOCK(sc); 1972 1973 sx_xunlock(&carp_sx); 1974 1975 return (0); 1976 } 1977 1978 void 1979 carp_detach(struct ifaddr *ifa, bool keep_cif) 1980 { 1981 struct ifnet *ifp = ifa->ifa_ifp; 1982 struct carp_if *cif = ifp->if_carp; 1983 struct carp_softc *sc = ifa->ifa_carp; 1984 int i, index; 1985 1986 KASSERT(sc != NULL, ("%s: %p not attached", __func__, ifa)); 1987 1988 sx_xlock(&carp_sx); 1989 1990 CARP_LOCK(sc); 1991 /* Shift array. */ 1992 index = sc->sc_naddrs + sc->sc_naddrs6; 1993 for (i = 0; i < index; i++) 1994 if (sc->sc_ifas[i] == ifa) 1995 break; 1996 KASSERT(i < index, ("%s: %p no backref", __func__, ifa)); 1997 for (; i < index - 1; i++) 1998 sc->sc_ifas[i] = sc->sc_ifas[i+1]; 1999 sc->sc_ifas[index - 1] = NULL; 2000 2001 switch (ifa->ifa_addr->sa_family) { 2002 #ifdef INET 2003 case AF_INET: 2004 cif->cif_naddrs--; 2005 sc->sc_naddrs--; 2006 break; 2007 #endif 2008 #ifdef INET6 2009 case AF_INET6: 2010 cif->cif_naddrs6--; 2011 sc->sc_naddrs6--; 2012 break; 2013 #endif 2014 } 2015 2016 carp_ifa_delroute(ifa); 2017 carp_multicast_cleanup(cif, ifa->ifa_addr->sa_family); 2018 2019 ifa->ifa_carp = NULL; 2020 ifa_free(ifa); 2021 2022 carp_hmac_prepare(sc); 2023 carp_sc_state(sc); 2024 2025 if (!keep_cif && sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) 2026 carp_destroy(sc); 2027 else 2028 CARP_UNLOCK(sc); 2029 2030 if (!keep_cif) 2031 CIF_FREE(cif); 2032 2033 sx_xunlock(&carp_sx); 2034 } 2035 2036 static void 2037 carp_set_state(struct carp_softc *sc, int state, const char *reason) 2038 { 2039 2040 CARP_LOCK_ASSERT(sc); 2041 2042 if (sc->sc_state != state) { 2043 const char *carp_states[] = { CARP_STATES }; 2044 char subsys[IFNAMSIZ+5]; 2045 2046 snprintf(subsys, IFNAMSIZ+5, "%u@%s", sc->sc_vhid, 2047 sc->sc_carpdev->if_xname); 2048 2049 CARP_LOG("%s: %s -> %s (%s)\n", subsys, 2050 carp_states[sc->sc_state], carp_states[state], reason); 2051 2052 sc->sc_state = state; 2053 2054 devctl_notify("CARP", subsys, carp_states[state], NULL); 2055 } 2056 } 2057 2058 static void 2059 carp_linkstate(struct ifnet *ifp) 2060 { 2061 struct carp_softc *sc; 2062 2063 CIF_LOCK(ifp->if_carp); 2064 IFNET_FOREACH_CARP(ifp, sc) { 2065 CARP_LOCK(sc); 2066 carp_sc_state(sc); 2067 CARP_UNLOCK(sc); 2068 } 2069 CIF_UNLOCK(ifp->if_carp); 2070 } 2071 2072 static void 2073 carp_sc_state(struct carp_softc *sc) 2074 { 2075 2076 CARP_LOCK_ASSERT(sc); 2077 2078 if (sc->sc_carpdev->if_link_state != LINK_STATE_UP || 2079 !(sc->sc_carpdev->if_flags & IFF_UP) || 2080 !V_carp_allow) { 2081 callout_stop(&sc->sc_ad_tmo); 2082 #ifdef INET 2083 callout_stop(&sc->sc_md_tmo); 2084 #endif 2085 #ifdef INET6 2086 callout_stop(&sc->sc_md6_tmo); 2087 #endif 2088 carp_set_state(sc, INIT, "hardware interface down"); 2089 carp_setrun(sc, 0); 2090 if (!sc->sc_suppress) 2091 carp_demote_adj(V_carp_ifdown_adj, "interface down"); 2092 sc->sc_suppress = 1; 2093 } else { 2094 carp_set_state(sc, INIT, "hardware interface up"); 2095 carp_setrun(sc, 0); 2096 if (sc->sc_suppress) 2097 carp_demote_adj(-V_carp_ifdown_adj, "interface up"); 2098 sc->sc_suppress = 0; 2099 } 2100 } 2101 2102 static void 2103 carp_demote_adj(int adj, char *reason) 2104 { 2105 atomic_add_int(&V_carp_demotion, adj); 2106 CARP_LOG("demoted by %d to %d (%s)\n", adj, V_carp_demotion, reason); 2107 taskqueue_enqueue(taskqueue_swi, &carp_sendall_task); 2108 } 2109 2110 static int 2111 carp_allow_sysctl(SYSCTL_HANDLER_ARGS) 2112 { 2113 int new, error; 2114 struct carp_softc *sc; 2115 2116 new = V_carp_allow; 2117 error = sysctl_handle_int(oidp, &new, 0, req); 2118 if (error || !req->newptr) 2119 return (error); 2120 2121 if (V_carp_allow != new) { 2122 V_carp_allow = new; 2123 2124 mtx_lock(&carp_mtx); 2125 LIST_FOREACH(sc, &carp_list, sc_next) { 2126 CARP_LOCK(sc); 2127 if (curvnet == sc->sc_carpdev->if_vnet) 2128 carp_sc_state(sc); 2129 CARP_UNLOCK(sc); 2130 } 2131 mtx_unlock(&carp_mtx); 2132 } 2133 2134 return (0); 2135 } 2136 2137 static int 2138 carp_dscp_sysctl(SYSCTL_HANDLER_ARGS) 2139 { 2140 int new, error; 2141 2142 new = V_carp_dscp; 2143 error = sysctl_handle_int(oidp, &new, 0, req); 2144 if (error || !req->newptr) 2145 return (error); 2146 2147 if (new < 0 || new > 63) 2148 return (EINVAL); 2149 2150 V_carp_dscp = new; 2151 2152 return (0); 2153 } 2154 2155 static int 2156 carp_demote_adj_sysctl(SYSCTL_HANDLER_ARGS) 2157 { 2158 int new, error; 2159 2160 new = V_carp_demotion; 2161 error = sysctl_handle_int(oidp, &new, 0, req); 2162 if (error || !req->newptr) 2163 return (error); 2164 2165 carp_demote_adj(new, "sysctl"); 2166 2167 return (0); 2168 } 2169 2170 #ifdef INET 2171 extern struct domain inetdomain; 2172 static struct protosw in_carp_protosw = { 2173 .pr_type = SOCK_RAW, 2174 .pr_domain = &inetdomain, 2175 .pr_protocol = IPPROTO_CARP, 2176 .pr_flags = PR_ATOMIC|PR_ADDR, 2177 .pr_input = carp_input, 2178 .pr_output = rip_output, 2179 .pr_ctloutput = rip_ctloutput, 2180 .pr_usrreqs = &rip_usrreqs 2181 }; 2182 #endif 2183 2184 #ifdef INET6 2185 extern struct domain inet6domain; 2186 static struct protosw in6_carp_protosw = { 2187 .pr_type = SOCK_RAW, 2188 .pr_domain = &inet6domain, 2189 .pr_protocol = IPPROTO_CARP, 2190 .pr_flags = PR_ATOMIC|PR_ADDR, 2191 .pr_input = carp6_input, 2192 .pr_output = rip6_output, 2193 .pr_ctloutput = rip6_ctloutput, 2194 .pr_usrreqs = &rip6_usrreqs 2195 }; 2196 #endif 2197 2198 static void 2199 carp_mod_cleanup(void) 2200 { 2201 2202 #ifdef INET 2203 if (proto_reg[CARP_INET] == 0) { 2204 (void)ipproto_unregister(IPPROTO_CARP); 2205 pf_proto_unregister(PF_INET, IPPROTO_CARP, SOCK_RAW); 2206 proto_reg[CARP_INET] = -1; 2207 } 2208 carp_iamatch_p = NULL; 2209 #endif 2210 #ifdef INET6 2211 if (proto_reg[CARP_INET6] == 0) { 2212 (void)ip6proto_unregister(IPPROTO_CARP); 2213 pf_proto_unregister(PF_INET6, IPPROTO_CARP, SOCK_RAW); 2214 proto_reg[CARP_INET6] = -1; 2215 } 2216 carp_iamatch6_p = NULL; 2217 carp_macmatch6_p = NULL; 2218 #endif 2219 carp_ioctl_p = NULL; 2220 carp_attach_p = NULL; 2221 carp_detach_p = NULL; 2222 carp_get_vhid_p = NULL; 2223 carp_linkstate_p = NULL; 2224 carp_forus_p = NULL; 2225 carp_output_p = NULL; 2226 carp_demote_adj_p = NULL; 2227 carp_master_p = NULL; 2228 mtx_unlock(&carp_mtx); 2229 taskqueue_drain(taskqueue_swi, &carp_sendall_task); 2230 mtx_destroy(&carp_mtx); 2231 sx_destroy(&carp_sx); 2232 } 2233 2234 static int 2235 carp_mod_load(void) 2236 { 2237 int err; 2238 2239 mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF); 2240 sx_init(&carp_sx, "carp_sx"); 2241 LIST_INIT(&carp_list); 2242 carp_get_vhid_p = carp_get_vhid; 2243 carp_forus_p = carp_forus; 2244 carp_output_p = carp_output; 2245 carp_linkstate_p = carp_linkstate; 2246 carp_ioctl_p = carp_ioctl; 2247 carp_attach_p = carp_attach; 2248 carp_detach_p = carp_detach; 2249 carp_demote_adj_p = carp_demote_adj; 2250 carp_master_p = carp_master; 2251 #ifdef INET6 2252 carp_iamatch6_p = carp_iamatch6; 2253 carp_macmatch6_p = carp_macmatch6; 2254 proto_reg[CARP_INET6] = pf_proto_register(PF_INET6, 2255 (struct protosw *)&in6_carp_protosw); 2256 if (proto_reg[CARP_INET6]) { 2257 printf("carp: error %d attaching to PF_INET6\n", 2258 proto_reg[CARP_INET6]); 2259 carp_mod_cleanup(); 2260 return (proto_reg[CARP_INET6]); 2261 } 2262 err = ip6proto_register(IPPROTO_CARP); 2263 if (err) { 2264 printf("carp: error %d registering with INET6\n", err); 2265 carp_mod_cleanup(); 2266 return (err); 2267 } 2268 #endif 2269 #ifdef INET 2270 carp_iamatch_p = carp_iamatch; 2271 proto_reg[CARP_INET] = pf_proto_register(PF_INET, &in_carp_protosw); 2272 if (proto_reg[CARP_INET]) { 2273 printf("carp: error %d attaching to PF_INET\n", 2274 proto_reg[CARP_INET]); 2275 carp_mod_cleanup(); 2276 return (proto_reg[CARP_INET]); 2277 } 2278 err = ipproto_register(IPPROTO_CARP); 2279 if (err) { 2280 printf("carp: error %d registering with INET\n", err); 2281 carp_mod_cleanup(); 2282 return (err); 2283 } 2284 #endif 2285 return (0); 2286 } 2287 2288 static int 2289 carp_modevent(module_t mod, int type, void *data) 2290 { 2291 switch (type) { 2292 case MOD_LOAD: 2293 return carp_mod_load(); 2294 /* NOTREACHED */ 2295 case MOD_UNLOAD: 2296 mtx_lock(&carp_mtx); 2297 if (LIST_EMPTY(&carp_list)) 2298 carp_mod_cleanup(); 2299 else { 2300 mtx_unlock(&carp_mtx); 2301 return (EBUSY); 2302 } 2303 break; 2304 2305 default: 2306 return (EINVAL); 2307 } 2308 2309 return (0); 2310 } 2311 2312 static moduledata_t carp_mod = { 2313 "carp", 2314 carp_modevent, 2315 0 2316 }; 2317 2318 DECLARE_MODULE(carp, carp_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY); 2319