2 * Copyright (c) 2002 Michael Shalayeff. All rights reserved.
3 * Copyright (c) 2003 Ryan McBride. All rights reserved.
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
15 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
16 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
17 * IN NO EVENT SHALL THE AUTHOR OR HIS RELATIVES BE LIABLE FOR ANY DIRECT,
18 * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
19 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
20 * SERVICES; LOSS OF MIND, USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
22 * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
23 * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
24 * THE POSSIBILITY OF SUCH DAMAGE.
27 * $FreeBSD: src/sys/netinet/ip_carp.c,v 1.48 2007/02/02 09:39:09 glebius Exp $
32 #include "opt_inet6.h"
34 #include <sys/param.h>
35 #include <sys/systm.h>
36 #include <sys/kernel.h>
37 #include <sys/in_cksum.h>
38 #include <sys/limits.h>
39 #include <sys/malloc.h>
41 #include <sys/msgport2.h>
45 #include <sys/sockio.h>
46 #include <sys/socket.h>
47 #include <sys/sysctl.h>
48 #include <sys/syslog.h>
49 #include <sys/thread.h>
51 #include <machine/stdarg.h>
52 #include <crypto/sha1.h>
55 #include <net/ethernet.h>
57 #include <net/if_dl.h>
58 #include <net/if_types.h>
59 #include <net/route.h>
60 #include <net/if_clone.h>
61 #include <net/if_var.h>
62 #include <net/ifq_var.h>
63 #include <net/netmsg2.h>
64 #include <net/netisr2.h>
67 #include <netinet/in.h>
68 #include <netinet/in_var.h>
69 #include <netinet/in_systm.h>
70 #include <netinet/ip.h>
71 #include <netinet/ip_var.h>
72 #include <netinet/if_ether.h>
76 #include <netinet/icmp6.h>
77 #include <netinet/ip6.h>
78 #include <netinet6/ip6_var.h>
79 #include <netinet6/scope6_var.h>
80 #include <netinet6/nd6.h>
83 #include <netinet/ip_carp.h>
86 * Note about carp's MP safe approach:
88 * Brief: carp_softc (softc), carp_softc_container (scc)
90 * - All configuration operation, e.g. ioctl, add/delete inet addresses
91 * is serialized by netisr0; not by carp's serializer
93 * - Backing interface's if_carp and carp_softc's relationship:
96 * if_carp -->| carp_if |
102 * | scc |-->| softc |
108 * | scc |-->| softc |
112 * - if_carp creation, modification and deletion all happen in netisr0,
113 * as stated previously. Since if_carp is accessed by multiple netisrs,
114 * the modification to if_carp is conducted in the following way:
118 * 1) Duplicate the old carp_if to new carp_if (ncif), and insert the
119 * to-be-added carp_softc to the new carp_if (ncif):
124 * +---------+ +---------+
125 * | carp_if | | carp_if |
126 * +---------+ +---------+
130 * +-----+ | | +-----+
131 * | scc |---->| softc |<----| scc |
132 * +-----+ | | +-----+
136 * +-----+ | | +-----+
137 * | scc |---->| softc |<----| scc |
138 * +-----+ | | +-----+
143 * | softc |<----| scc |
147 * 2) Switch save if_carp into ocif and switch if_carp to ncif:
152 * +---------+ +---------+
153 * | carp_if | | carp_if |
154 * +---------+ +---------+
158 * +-----+ | | +-----+
159 * | scc |---->| softc |<----| scc |
160 * +-----+ | | +-----+
164 * +-----+ | | +-----+
165 * | scc |---->| softc |<----| scc |
166 * +-----+ | | +-----+
171 * | softc |<----| scc |
175 * 3) Run netmsg_service_sync(), which will make sure that
176 * ocif is no longer accessed (all network operations
177 * are happened only in network threads).
178 * 4) Free ocif -- only carp_if and scc are freed.
181 * Removing carp_softc:
183 * 1) Duplicate the old carp_if to new carp_if (ncif); the to-be-deleted
184 * carp_softc will not be duplicated.
189 * +---------+ +---------+
190 * | carp_if | | carp_if |
191 * +---------+ +---------+
195 * +-----+ | | +-----+
196 * | scc |---->| softc |<----| scc |
197 * +-----+ | | +-----+
202 * | scc |---->| softc | |
207 * +-----+ | | +-----+
208 * | scc |---->| softc |<----| scc |
209 * +-----+ | | +-----+
212 * 2) Switch save if_carp into ocif and switch if_carp to ncif:
217 * +---------+ +---------+
218 * | carp_if | | carp_if |
219 * +---------+ +---------+
223 * +-----+ | | +-----+
224 * | scc |---->| softc |<----| scc |
225 * +-----+ | | +-----+
230 * | scc |---->| softc | |
235 * +-----+ | | +-----+
236 * | scc |---->| softc |<----| scc |
237 * +-----+ | | +-----+
240 * 3) Run netmsg_service_sync(), which will make sure that
241 * ocif is no longer accessed (all network operations
242 * are happened only in network threads).
243 * 4) Free ocif -- only carp_if and scc are freed.
245 * - if_carp accessing:
246 * The accessing code should cache the if_carp in a local temporary
247 * variable and accessing the temporary variable along the code path
248 * instead of accessing if_carp later on.
251 #define CARP_IFNAME "carp"
252 #define CARP_IS_RUNNING(ifp) \
253 (((ifp)->if_flags & (IFF_UP | IFF_RUNNING)) == (IFF_UP | IFF_RUNNING))
258 uint32_t vha_flags; /* CARP_VHAF_ */
259 struct in_ifaddr *vha_ia; /* carp address */
260 struct in_ifaddr *vha_iaback; /* backing address */
261 TAILQ_ENTRY(carp_vhaddr) vha_link;
263 TAILQ_HEAD(carp_vhaddr_list, carp_vhaddr);
266 struct netmsg_base base;
267 struct ifnet *nc_carpdev;
268 struct carp_softc *nc_softc;
274 struct arpcom arpcom;
275 struct ifnet *sc_carpdev; /* parent interface */
276 struct carp_vhaddr_list sc_vha_list; /* virtual addr list */
278 const struct in_ifaddr *sc_ia; /* primary iface address v4 */
279 struct ip_moptions sc_imo;
282 struct in6_ifaddr *sc_ia6; /* primary iface address v6 */
283 struct ip6_moptions sc_im6o;
286 enum { INIT = 0, BACKUP, MASTER }
292 int sc_sendad_errors;
293 #define CARP_SENDAD_MAX_ERRORS 3
294 int sc_sendad_success;
295 #define CARP_SENDAD_MIN_SUCCESS 3
299 int sc_naddrs; /* actually used IPv4 vha */
301 int sc_advbase; /* seconds */
306 #define CARP_HMAC_PAD 64
307 unsigned char sc_key[CARP_KEY_LEN];
308 unsigned char sc_pad[CARP_HMAC_PAD];
311 struct callout sc_ad_tmo; /* advertisement timeout */
312 struct netmsg_carp sc_ad_msg; /* adv timeout netmsg */
313 struct callout sc_md_tmo; /* ip4 master down timeout */
314 struct callout sc_md6_tmo; /* ip6 master down timeout */
315 struct netmsg_carp sc_md_msg; /* master down timeout netmsg */
317 LIST_ENTRY(carp_softc) sc_next; /* Interface clue */
320 #define sc_if arpcom.ac_if
322 struct carp_softc_container {
323 TAILQ_ENTRY(carp_softc_container) scc_link;
324 struct carp_softc *scc_softc;
326 TAILQ_HEAD(carp_if, carp_softc_container);
328 SYSCTL_DECL(_net_inet_carp);
330 static int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 1, 0, 0, 1 }; /* XXX for now */
331 SYSCTL_INT(_net_inet_carp, CARPCTL_ALLOW, allow, CTLFLAG_RW,
332 &carp_opts[CARPCTL_ALLOW], 0, "Accept incoming CARP packets");
333 SYSCTL_INT(_net_inet_carp, CARPCTL_PREEMPT, preempt, CTLFLAG_RW,
334 &carp_opts[CARPCTL_PREEMPT], 0, "high-priority backup preemption mode");
335 SYSCTL_INT(_net_inet_carp, CARPCTL_LOG, log, CTLFLAG_RW,
336 &carp_opts[CARPCTL_LOG], 0, "log bad carp packets");
337 SYSCTL_INT(_net_inet_carp, CARPCTL_ARPBALANCE, arpbalance, CTLFLAG_RW,
338 &carp_opts[CARPCTL_ARPBALANCE], 0, "balance arp responses");
339 SYSCTL_INT(_net_inet_carp, CARPCTL_SETROUTE, setroute, CTLFLAG_RW,
340 &carp_opts[CARPCTL_SETROUTE], 0, "set route");
342 static int carp_suppress_preempt = 0;
343 SYSCTL_INT(_net_inet_carp, OID_AUTO, suppress_preempt, CTLFLAG_RD,
344 &carp_suppress_preempt, 0, "Preemption is suppressed");
346 static int carp_prio_ad = 1;
347 SYSCTL_INT(_net_inet_carp, OID_AUTO, prio_ad, CTLFLAG_RD,
348 &carp_prio_ad, 0, "Prioritize advertisement packet");
350 static struct carpstats carpstats;
351 SYSCTL_STRUCT(_net_inet_carp, CARPCTL_STATS, stats, CTLFLAG_RW,
352 &carpstats, carpstats,
353 "CARP statistics (struct carpstats, netinet/ip_carp.h)");
355 #define CARP_LOG(...) do { \
356 if (carp_opts[CARPCTL_LOG] > 0) \
357 log(LOG_INFO, __VA_ARGS__); \
360 #define CARP_DEBUG(...) do { \
361 if (carp_opts[CARPCTL_LOG] > 1) \
362 log(LOG_DEBUG, __VA_ARGS__); \
365 static struct lwkt_token carp_listtok = LWKT_TOKEN_INITIALIZER(carp_list_token);
367 static void carp_hmac_prepare(struct carp_softc *);
368 static void carp_hmac_generate(struct carp_softc *, uint32_t *,
370 static int carp_hmac_verify(struct carp_softc *, uint32_t *,
372 static void carp_setroute(struct carp_softc *, int);
373 static void carp_proto_input_c(struct carp_softc *, struct mbuf *,
374 struct carp_header *, sa_family_t);
375 static int carp_clone_create(struct if_clone *, int, caddr_t);
376 static int carp_clone_destroy(struct ifnet *);
377 static void carp_detach(struct carp_softc *, boolean_t, boolean_t);
378 static void carp_prepare_ad(struct carp_softc *, struct carp_header *);
379 static void carp_send_ad_all(void);
380 static void carp_send_ad_timeout(void *);
381 static void carp_send_ad(struct carp_softc *);
382 static void carp_send_arp(struct carp_softc *);
383 static void carp_master_down_timeout(void *);
384 static void carp_master_down(struct carp_softc *);
385 static void carp_setrun(struct carp_softc *, sa_family_t);
386 static void carp_set_state(struct carp_softc *, int);
387 static struct ifnet *carp_forus(struct carp_if *, const uint8_t *);
389 static void carp_init(void *);
390 static int carp_ioctl(struct ifnet *, u_long, caddr_t, struct ucred *);
391 static int carp_output(struct ifnet *, struct mbuf *, struct sockaddr *,
393 static void carp_start(struct ifnet *, struct ifaltq_subque *);
395 static void carp_multicast_cleanup(struct carp_softc *);
396 static void carp_add_addr(struct carp_softc *, struct ifaddr *);
397 static void carp_del_addr(struct carp_softc *, struct ifaddr *);
398 static void carp_config_addr(struct carp_softc *, struct ifaddr *);
399 static void carp_link_addrs(struct carp_softc *, struct ifnet *,
401 static void carp_unlink_addrs(struct carp_softc *, struct ifnet *,
403 static void carp_update_addrs(struct carp_softc *, struct ifaddr *);
405 static int carp_config_vhaddr(struct carp_softc *, struct carp_vhaddr *,
407 static int carp_activate_vhaddr(struct carp_softc *, struct carp_vhaddr *,
408 struct ifnet *, struct in_ifaddr *, int);
409 static void carp_deactivate_vhaddr(struct carp_softc *,
410 struct carp_vhaddr *, boolean_t);
411 static int carp_addroute_vhaddr(struct carp_softc *, struct carp_vhaddr *);
412 static void carp_delroute_vhaddr(struct carp_softc *, struct carp_vhaddr *,
416 static void carp_sc_state(struct carp_softc *);
419 static void carp_send_na(struct carp_softc *);
421 static int carp_set_addr6(struct carp_softc *, struct sockaddr_in6 *);
422 static int carp_del_addr6(struct carp_softc *, struct sockaddr_in6 *);
424 static void carp_multicast6_cleanup(struct carp_softc *);
426 static void carp_stop(struct carp_softc *, boolean_t);
427 static void carp_suspend(struct carp_softc *, boolean_t);
428 static void carp_ioctl_stop(struct carp_softc *);
429 static int carp_ioctl_setvh(struct carp_softc *, void *, struct ucred *);
430 static void carp_ioctl_ifcap(struct carp_softc *, int);
431 static int carp_ioctl_getvh(struct carp_softc *, void *, struct ucred *);
432 static int carp_ioctl_getdevname(struct carp_softc *, struct ifdrv *);
433 static int carp_ioctl_getvhaddr(struct carp_softc *, struct ifdrv *);
435 static struct carp_if *carp_if_remove(struct carp_if *, struct carp_softc *);
436 static struct carp_if *carp_if_insert(struct carp_if *, struct carp_softc *);
437 static void carp_if_free(struct carp_if *);
439 static void carp_ifaddr(void *, struct ifnet *, enum ifaddr_event,
441 static void carp_ifdetach(void *, struct ifnet *);
443 static void carp_ifdetach_dispatch(netmsg_t);
444 static void carp_clone_destroy_dispatch(netmsg_t);
445 static void carp_init_dispatch(netmsg_t);
446 static void carp_ioctl_stop_dispatch(netmsg_t);
447 static void carp_ioctl_setvh_dispatch(netmsg_t);
448 static void carp_ioctl_ifcap_dispatch(netmsg_t);
449 static void carp_ioctl_getvh_dispatch(netmsg_t);
450 static void carp_ioctl_getdevname_dispatch(netmsg_t);
451 static void carp_ioctl_getvhaddr_dispatch(netmsg_t);
452 static void carp_send_ad_timeout_dispatch(netmsg_t);
453 static void carp_master_down_timeout_dispatch(netmsg_t);
455 static MALLOC_DEFINE(M_CARP, "CARP", "CARP interfaces");
457 static LIST_HEAD(, carp_softc) carpif_list;
459 static struct if_clone carp_cloner =
460 IF_CLONE_INITIALIZER(CARP_IFNAME, carp_clone_create, carp_clone_destroy,
463 static const uint8_t carp_etheraddr[ETHER_ADDR_LEN] =
464 { 0, 0, 0x5e, 0, 1, 0 };
466 static eventhandler_tag carp_ifdetach_event;
467 static eventhandler_tag carp_ifaddr_event;
470 carp_insert_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha_new)
472 struct carp_vhaddr *vha;
473 u_long new_addr, addr;
475 KKASSERT((vha_new->vha_flags & CARP_VHAF_ONLIST) == 0);
478 * Virtual address list is sorted; smaller one first
480 new_addr = ntohl(vha_new->vha_ia->ia_addr.sin_addr.s_addr);
482 TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
483 addr = ntohl(vha->vha_ia->ia_addr.sin_addr.s_addr);
489 TAILQ_INSERT_TAIL(&sc->sc_vha_list, vha_new, vha_link);
491 TAILQ_INSERT_BEFORE(vha, vha_new, vha_link);
492 vha_new->vha_flags |= CARP_VHAF_ONLIST;
496 carp_remove_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha)
498 KKASSERT(vha->vha_flags & CARP_VHAF_ONLIST);
499 vha->vha_flags &= ~CARP_VHAF_ONLIST;
500 TAILQ_REMOVE(&sc->sc_vha_list, vha, vha_link);
504 carp_hmac_prepare(struct carp_softc *sc)
506 uint8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
507 uint8_t vhid = sc->sc_vhid & 0xff;
510 struct ifaddr_container *ifac;
514 struct carp_vhaddr *vha;
517 /* XXX: possible race here */
519 /* compute ipad from key */
520 bzero(sc->sc_pad, sizeof(sc->sc_pad));
521 bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
522 for (i = 0; i < sizeof(sc->sc_pad); i++)
523 sc->sc_pad[i] ^= 0x36;
525 /* precompute first part of inner hash */
526 SHA1Init(&sc->sc_sha1);
527 SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
528 SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
529 SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
530 SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
532 TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
533 SHA1Update(&sc->sc_sha1,
534 (const uint8_t *)&vha->vha_ia->ia_addr.sin_addr,
535 sizeof(struct in_addr));
539 TAILQ_FOREACH(ifac, &sc->sc_if.if_addrheads[mycpuid], ifa_link) {
540 struct ifaddr *ifa = ifac->ifa;
542 if (ifa->ifa_addr->sa_family == AF_INET6) {
543 in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
544 in6_clearscope(&in6);
545 SHA1Update(&sc->sc_sha1, (void *)&in6, sizeof(in6));
550 /* convert ipad to opad */
551 for (i = 0; i < sizeof(sc->sc_pad); i++)
552 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
556 carp_hmac_generate(struct carp_softc *sc, uint32_t counter[2],
557 unsigned char md[20])
561 /* fetch first half of inner hash */
562 bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
564 SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
565 SHA1Final(md, &sha1ctx);
569 SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
570 SHA1Update(&sha1ctx, md, 20);
571 SHA1Final(md, &sha1ctx);
575 carp_hmac_verify(struct carp_softc *sc, uint32_t counter[2],
576 unsigned char md[20])
578 unsigned char md2[20];
580 carp_hmac_generate(sc, counter, md2);
581 return (bcmp(md, md2, sizeof(md2)));
585 carp_setroute(struct carp_softc *sc, int cmd)
588 struct ifaddr_container *ifac;
590 struct carp_vhaddr *vha;
592 KKASSERT(cmd == RTM_DELETE || cmd == RTM_ADD);
594 TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
595 if (vha->vha_iaback == NULL)
597 if (cmd == RTM_DELETE)
598 carp_delroute_vhaddr(sc, vha, FALSE);
600 carp_addroute_vhaddr(sc, vha);
604 TAILQ_FOREACH(ifac, &sc->sc_if.if_addrheads[mycpuid], ifa_link) {
605 struct ifaddr *ifa = ifac->ifa;
607 if (ifa->ifa_addr->sa_family == AF_INET6) {
618 carp_clone_create(struct if_clone *ifc, int unit, caddr_t param __unused)
620 struct carp_softc *sc;
623 sc = kmalloc(sizeof(*sc), M_CARP, M_WAITOK | M_ZERO);
627 sc->sc_advbase = CARP_DFLTINTV;
628 sc->sc_vhid = -1; /* required setting */
630 sc->sc_init_counter = 1;
634 TAILQ_INIT(&sc->sc_vha_list);
637 sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL;
640 callout_init_mp(&sc->sc_ad_tmo);
641 netmsg_init(&sc->sc_ad_msg.base, NULL, &netisr_adone_rport,
642 MSGF_DROPABLE | MSGF_PRIORITY, carp_send_ad_timeout_dispatch);
643 sc->sc_ad_msg.nc_softc = sc;
645 callout_init_mp(&sc->sc_md_tmo);
646 callout_init_mp(&sc->sc_md6_tmo);
647 netmsg_init(&sc->sc_md_msg.base, NULL, &netisr_adone_rport,
648 MSGF_DROPABLE | MSGF_PRIORITY, carp_master_down_timeout_dispatch);
649 sc->sc_md_msg.nc_softc = sc;
651 if_initname(ifp, CARP_IFNAME, unit);
653 ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
654 ifp->if_init = carp_init;
655 ifp->if_ioctl = carp_ioctl;
656 ifp->if_start = carp_start;
658 ifp->if_capabilities = IFCAP_TXCSUM | IFCAP_TSO;
659 ifp->if_capenable = ifp->if_capabilities;
661 * Leave if_hwassist as it is; if_hwassist will be
662 * setup when this carp interface has parent.
665 ifq_set_maxlen(&ifp->if_snd, ifqmaxlen);
666 ifq_set_ready(&ifp->if_snd);
668 ether_ifattach(ifp, carp_etheraddr, NULL);
670 ifp->if_type = IFT_CARP;
671 ifp->if_output = carp_output;
673 lwkt_gettoken(&carp_listtok);
674 LIST_INSERT_HEAD(&carpif_list, sc, sc_next);
675 lwkt_reltoken(&carp_listtok);
681 carp_clone_destroy_dispatch(netmsg_t msg)
683 struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
684 struct carp_softc *sc = cmsg->nc_softc;
687 carp_detach(sc, TRUE, FALSE);
689 callout_stop_sync(&sc->sc_ad_tmo);
690 callout_stop_sync(&sc->sc_md_tmo);
691 callout_stop_sync(&sc->sc_md6_tmo);
694 lwkt_dropmsg(&sc->sc_ad_msg.base.lmsg);
695 lwkt_dropmsg(&sc->sc_md_msg.base.lmsg);
698 lwkt_replymsg(&cmsg->base.lmsg, 0);
702 carp_clone_destroy(struct ifnet *ifp)
704 struct carp_softc *sc = ifp->if_softc;
705 struct netmsg_carp cmsg;
707 bzero(&cmsg, sizeof(cmsg));
708 netmsg_init(&cmsg.base, NULL, &curthread->td_msgport, 0,
709 carp_clone_destroy_dispatch);
712 lwkt_domsg(netisr_cpuport(0), &cmsg.base.lmsg, 0);
714 lwkt_gettoken(&carp_listtok);
715 LIST_REMOVE(sc, sc_next);
716 lwkt_reltoken(&carp_listtok);
721 KASSERT(sc->sc_naddrs == 0, ("certain inet address is still active"));
727 static struct carp_if *
728 carp_if_remove(struct carp_if *ocif, struct carp_softc *sc)
730 struct carp_softc_container *oscc, *scc;
737 TAILQ_FOREACH(oscc, ocif, scc_link) {
740 if (oscc->scc_softc == sc)
744 KASSERT(found, ("%s carp_softc is not on carp_if", __func__));
747 /* Last one is going to be unlinked */
751 cif = kmalloc(sizeof(*cif), M_CARP, M_WAITOK | M_ZERO);
754 TAILQ_FOREACH(oscc, ocif, scc_link) {
755 if (oscc->scc_softc == sc)
758 scc = kmalloc(sizeof(*scc), M_CARP, M_WAITOK | M_ZERO);
759 scc->scc_softc = oscc->scc_softc;
760 TAILQ_INSERT_TAIL(cif, scc, scc_link);
766 static struct carp_if *
767 carp_if_insert(struct carp_if *ocif, struct carp_softc *sc)
769 struct carp_softc_container *oscc;
774 TAILQ_FOREACH(oscc, ocif, scc_link) {
775 if (oscc->scc_softc == sc)
781 if (sc->sc_carpdev != NULL) {
782 KASSERT(onlist, ("%s is not on %s carp list",
783 sc->sc_if.if_xname, sc->sc_carpdev->if_xname));
785 KASSERT(!onlist, ("%s is already on carp list",
786 sc->sc_if.if_xname));
792 struct carp_softc_container *new_scc, *scc;
795 cif = kmalloc(sizeof(*cif), M_CARP, M_WAITOK | M_ZERO);
798 new_scc = kmalloc(sizeof(*new_scc), M_CARP, M_WAITOK | M_ZERO);
799 new_scc->scc_softc = sc;
802 TAILQ_FOREACH(oscc, ocif, scc_link) {
804 oscc->scc_softc->sc_vhid > sc->sc_vhid) {
805 TAILQ_INSERT_TAIL(cif, new_scc,
810 scc = kmalloc(sizeof(*scc), M_CARP,
812 scc->scc_softc = oscc->scc_softc;
813 TAILQ_INSERT_TAIL(cif, scc, scc_link);
817 TAILQ_INSERT_TAIL(cif, new_scc, scc_link);
826 carp_if_free(struct carp_if *cif)
828 struct carp_softc_container *scc;
830 while ((scc = TAILQ_FIRST(cif)) != NULL) {
831 TAILQ_REMOVE(cif, scc, scc_link);
838 carp_detach(struct carp_softc *sc, boolean_t detach, boolean_t del_iaback)
840 carp_suspend(sc, detach);
842 carp_multicast_cleanup(sc);
844 carp_multicast6_cleanup(sc);
847 if (!sc->sc_dead && detach) {
848 struct carp_vhaddr *vha;
850 TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link)
851 carp_deactivate_vhaddr(sc, vha, del_iaback);
852 KKASSERT(sc->sc_naddrs == 0);
855 if (sc->sc_carpdev != NULL) {
856 struct ifnet *ifp = sc->sc_carpdev;
857 struct carp_if *ocif = ifp->if_carp;
859 ifp->if_carp = carp_if_remove(ocif, sc);
860 KASSERT(ifp->if_carp != ocif,
861 ("%s carp_if_remove failed", __func__));
863 sc->sc_carpdev = NULL;
865 sc->arpcom.ac_if.if_hwassist = 0;
868 * Make sure that all protocol threads see the
869 * sc_carpdev and if_carp changes
871 netmsg_service_sync();
873 if (ifp->if_carp == NULL) {
875 * No more carp interfaces using
876 * ifp as the backing interface,
877 * move it out of promiscous mode.
883 * The old carp list could be safely free now,
884 * since no one can access it.
891 carp_ifdetach_dispatch(netmsg_t msg)
893 struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
894 struct ifnet *ifp = cmsg->nc_carpdev;
896 while (ifp->if_carp) {
897 struct carp_softc_container *scc;
899 scc = TAILQ_FIRST((struct carp_if *)(ifp->if_carp));
900 carp_detach(scc->scc_softc, TRUE, TRUE);
902 lwkt_replymsg(&cmsg->base.lmsg, 0);
905 /* Detach an interface from the carp. */
907 carp_ifdetach(void *arg __unused, struct ifnet *ifp)
909 struct netmsg_carp cmsg;
911 ASSERT_IFNET_NOT_SERIALIZED_ALL(ifp);
913 bzero(&cmsg, sizeof(cmsg));
914 netmsg_init(&cmsg.base, NULL, &curthread->td_msgport, 0,
915 carp_ifdetach_dispatch);
916 cmsg.nc_carpdev = ifp;
918 lwkt_domsg(netisr_cpuport(0), &cmsg.base.lmsg, 0);
922 * process input packet.
923 * we have rearranged checks order compared to the rfc,
924 * but it seems more efficient this way or not possible otherwise.
927 carp_proto_input(struct mbuf **mp, int *offp, int proto)
929 struct mbuf *m = *mp;
930 struct ip *ip = mtod(m, struct ip *);
931 struct ifnet *ifp = m->m_pkthdr.rcvif;
932 struct carp_header *ch;
933 struct carp_softc *sc;
939 carpstats.carps_ipackets++;
941 if (!carp_opts[CARPCTL_ALLOW]) {
946 /* Check if received on a valid carp interface */
947 if (ifp->if_type != IFT_CARP) {
948 carpstats.carps_badif++;
949 CARP_LOG("carp_proto_input: packet received on non-carp "
950 "interface: %s\n", ifp->if_xname);
955 if (!CARP_IS_RUNNING(ifp)) {
956 carpstats.carps_badif++;
957 CARP_LOG("carp_proto_input: packet received on stopped carp "
958 "interface: %s\n", ifp->if_xname);
964 if (sc->sc_carpdev == NULL) {
965 carpstats.carps_badif++;
966 CARP_LOG("carp_proto_input: packet received on defunc carp "
967 "interface: %s\n", ifp->if_xname);
972 if (!IN_MULTICAST(ntohl(ip->ip_dst.s_addr))) {
973 carpstats.carps_badif++;
974 CARP_LOG("carp_proto_input: non-mcast packet on "
975 "interface: %s\n", ifp->if_xname);
980 /* Verify that the IP TTL is CARP_DFLTTL. */
981 if (ip->ip_ttl != CARP_DFLTTL) {
982 carpstats.carps_badttl++;
983 CARP_LOG("carp_proto_input: received ttl %d != %d on %s\n",
984 ip->ip_ttl, CARP_DFLTTL, ifp->if_xname);
989 /* Minimal CARP packet size */
990 len = iphlen + sizeof(*ch);
993 * Verify that the received packet length is
994 * not less than the CARP header
996 if (m->m_pkthdr.len < len) {
997 carpstats.carps_badlen++;
998 CARP_LOG("packet too short %d on %s\n", m->m_pkthdr.len,
1004 /* Make sure that CARP header is contiguous */
1005 if (len > m->m_len) {
1006 m = m_pullup(m, len);
1008 carpstats.carps_hdrops++;
1009 CARP_LOG("carp_proto_input: m_pullup failed\n");
1012 ip = mtod(m, struct ip *);
1014 ch = (struct carp_header *)((uint8_t *)ip + iphlen);
1016 /* Verify the CARP checksum */
1017 if (in_cksum_skip(m, len, iphlen)) {
1018 carpstats.carps_badsum++;
1019 CARP_LOG("carp_proto_input: checksum failed on %s\n",
1024 carp_proto_input_c(sc, m, ch, AF_INET);
1026 return(IPPROTO_DONE);
1031 carp6_proto_input(struct mbuf **mp, int *offp, int proto)
1033 struct mbuf *m = *mp;
1034 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1035 struct ifnet *ifp = m->m_pkthdr.rcvif;
1036 struct carp_header *ch;
1037 struct carp_softc *sc;
1040 carpstats.carps_ipackets6++;
1042 if (!carp_opts[CARPCTL_ALLOW]) {
1047 /* check if received on a valid carp interface */
1048 if (ifp->if_type != IFT_CARP) {
1049 carpstats.carps_badif++;
1050 CARP_LOG("carp6_proto_input: packet received on non-carp "
1051 "interface: %s\n", ifp->if_xname);
1056 if (!CARP_IS_RUNNING(ifp)) {
1057 carpstats.carps_badif++;
1058 CARP_LOG("carp_proto_input: packet received on stopped carp "
1059 "interface: %s\n", ifp->if_xname);
1065 if (sc->sc_carpdev == NULL) {
1066 carpstats.carps_badif++;
1067 CARP_LOG("carp6_proto_input: packet received on defunc-carp "
1068 "interface: %s\n", ifp->if_xname);
1073 /* verify that the IP TTL is 255 */
1074 if (ip6->ip6_hlim != CARP_DFLTTL) {
1075 carpstats.carps_badttl++;
1076 CARP_LOG("carp6_proto_input: received ttl %d != 255 on %s\n",
1077 ip6->ip6_hlim, ifp->if_xname);
1082 /* verify that we have a complete carp packet */
1084 IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
1086 carpstats.carps_badlen++;
1087 CARP_LOG("carp6_proto_input: packet size %u too small\n", len);
1091 /* verify the CARP checksum */
1092 if (in_cksum_range(m, 0, *offp, sizeof(*ch))) {
1093 carpstats.carps_badsum++;
1094 CARP_LOG("carp6_proto_input: checksum failed, on %s\n",
1100 carp_proto_input_c(sc, m, ch, AF_INET6);
1102 return (IPPROTO_DONE);
1107 carp_proto_input_c(struct carp_softc *sc, struct mbuf *m,
1108 struct carp_header *ch, sa_family_t af)
1111 uint64_t tmp_counter;
1112 struct timeval sc_tv, ch_tv;
1114 if (sc->sc_vhid != ch->carp_vhid) {
1116 * CARP uses multicast, however, multicast packets
1117 * are tapped to all CARP interfaces on the physical
1118 * interface receiving the CARP packets, so we don't
1119 * update any stats here.
1126 /* verify the CARP version. */
1127 if (ch->carp_version != CARP_VERSION) {
1128 carpstats.carps_badver++;
1129 CARP_LOG("%s; invalid version %d\n", cifp->if_xname,
1135 /* verify the hash */
1136 if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
1137 carpstats.carps_badauth++;
1138 CARP_LOG("%s: incorrect hash\n", cifp->if_xname);
1143 tmp_counter = ntohl(ch->carp_counter[0]);
1144 tmp_counter = tmp_counter<<32;
1145 tmp_counter += ntohl(ch->carp_counter[1]);
1147 /* XXX Replay protection goes here */
1149 sc->sc_init_counter = 0;
1150 sc->sc_counter = tmp_counter;
1152 sc_tv.tv_sec = sc->sc_advbase;
1153 if (carp_suppress_preempt && sc->sc_advskew < 240)
1154 sc_tv.tv_usec = 240 * 1000000 / 256;
1156 sc_tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1157 ch_tv.tv_sec = ch->carp_advbase;
1158 ch_tv.tv_usec = ch->carp_advskew * 1000000 / 256;
1160 switch (sc->sc_state) {
1166 * If we receive an advertisement from a master who's going to
1167 * be more frequent than us, go into BACKUP state.
1169 if (timevalcmp(&sc_tv, &ch_tv, >) ||
1170 timevalcmp(&sc_tv, &ch_tv, ==)) {
1171 callout_stop(&sc->sc_ad_tmo);
1172 CARP_DEBUG("%s: MASTER -> BACKUP "
1173 "(more frequent advertisement received)\n",
1175 carp_set_state(sc, BACKUP);
1177 if (carp_opts[CARPCTL_SETROUTE])
1178 carp_setroute(sc, RTM_DELETE);
1184 * If we're pre-empting masters who advertise slower than us,
1185 * and this one claims to be slower, treat him as down.
1187 if (carp_opts[CARPCTL_PREEMPT] &&
1188 timevalcmp(&sc_tv, &ch_tv, <)) {
1189 CARP_DEBUG("%s: BACKUP -> MASTER "
1190 "(preempting a slower master)\n", cifp->if_xname);
1191 carp_master_down(sc);
1196 * If the master is going to advertise at such a low frequency
1197 * that he's guaranteed to time out, we'd might as well just
1198 * treat him as timed out now.
1200 sc_tv.tv_sec = sc->sc_advbase * 3;
1201 if (timevalcmp(&sc_tv, &ch_tv, <)) {
1202 CARP_DEBUG("%s: BACKUP -> MASTER (master timed out)\n",
1204 carp_master_down(sc);
1209 * Otherwise, we reset the counter and wait for the next
1212 carp_setrun(sc, af);
1219 carp_input(void *v, struct mbuf *m)
1221 struct carp_if *cif = v;
1222 struct ether_header *eh;
1223 struct carp_softc_container *scc;
1226 eh = mtod(m, struct ether_header *);
1228 ifp = carp_forus(cif, eh->ether_dhost);
1230 ether_reinput_oncpu(ifp, m, REINPUT_RUNBPF);
1234 if ((m->m_flags & (M_BCAST | M_MCAST)) == 0)
1238 * XXX Should really check the list of multicast addresses
1239 * for each CARP interface _before_ copying.
1241 TAILQ_FOREACH(scc, cif, scc_link) {
1242 struct carp_softc *sc = scc->scc_softc;
1245 if ((sc->sc_if.if_flags & IFF_UP) == 0)
1248 m0 = m_dup(m, M_NOWAIT);
1252 ether_reinput_oncpu(&sc->sc_if, m0, REINPUT_RUNBPF);
1258 carp_prepare_ad(struct carp_softc *sc, struct carp_header *ch)
1260 if (sc->sc_init_counter) {
1261 /* this could also be seconds since unix epoch */
1262 sc->sc_counter = karc4random();
1263 sc->sc_counter = sc->sc_counter << 32;
1264 sc->sc_counter += karc4random();
1269 ch->carp_counter[0] = htonl((sc->sc_counter >> 32) & 0xffffffff);
1270 ch->carp_counter[1] = htonl(sc->sc_counter & 0xffffffff);
1272 carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
1276 carp_send_ad_all(void)
1278 struct carp_softc *sc;
1280 LIST_FOREACH(sc, &carpif_list, sc_next) {
1281 if (sc->sc_carpdev == NULL)
1284 if (CARP_IS_RUNNING(&sc->sc_if) && sc->sc_state == MASTER)
1290 carp_send_ad_timeout(void *xsc)
1292 struct carp_softc *sc = xsc;
1293 struct netmsg_carp *cmsg = &sc->sc_ad_msg;
1295 KASSERT(mycpuid == 0, ("%s not on cpu0 but on cpu%d",
1296 __func__, mycpuid));
1299 if (cmsg->base.lmsg.ms_flags & MSGF_DONE)
1300 lwkt_sendmsg_oncpu(netisr_cpuport(0), &cmsg->base.lmsg);
1305 carp_send_ad_timeout_dispatch(netmsg_t msg)
1307 struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
1308 struct carp_softc *sc = cmsg->nc_softc;
1312 lwkt_replymsg(&cmsg->base.lmsg, 0);
1319 carp_send_ad(struct carp_softc *sc)
1321 struct ifnet *cifp = &sc->sc_if;
1322 struct carp_header ch;
1324 struct carp_header *ch_ptr;
1326 int len, advbase, advskew;
1328 if (!CARP_IS_RUNNING(cifp)) {
1333 advbase = sc->sc_advbase;
1334 if (!carp_suppress_preempt || sc->sc_advskew > 240)
1335 advskew = sc->sc_advskew;
1338 tv.tv_sec = advbase;
1339 tv.tv_usec = advskew * 1000000 / 256;
1342 ch.carp_version = CARP_VERSION;
1343 ch.carp_type = CARP_ADVERTISEMENT;
1344 ch.carp_vhid = sc->sc_vhid;
1345 ch.carp_advbase = advbase;
1346 ch.carp_advskew = advskew;
1347 ch.carp_authlen = 7; /* XXX DEFINE */
1348 ch.carp_pad1 = 0; /* must be zero */
1352 if (sc->sc_ia != NULL) {
1355 MGETHDR(m, M_NOWAIT, MT_HEADER);
1357 IFNET_STAT_INC(cifp, oerrors, 1);
1358 carpstats.carps_onomem++;
1359 /* XXX maybe less ? */
1360 if (advbase != 255 || advskew != 255)
1361 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
1362 carp_send_ad_timeout, sc);
1365 len = sizeof(*ip) + sizeof(ch);
1366 m->m_pkthdr.len = len;
1367 m->m_pkthdr.rcvif = NULL;
1369 MH_ALIGN(m, m->m_len);
1370 m->m_flags |= M_MCAST;
1372 m->m_flags |= M_PRIO;
1373 ip = mtod(m, struct ip *);
1374 ip->ip_v = IPVERSION;
1375 ip->ip_hl = sizeof(*ip) >> 2;
1376 ip->ip_tos = IPTOS_LOWDELAY;
1378 ip->ip_id = ip_newid();
1380 ip->ip_ttl = CARP_DFLTTL;
1381 ip->ip_p = IPPROTO_CARP;
1383 ip->ip_src = sc->sc_ia->ia_addr.sin_addr;
1384 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
1386 ch_ptr = (struct carp_header *)(&ip[1]);
1387 bcopy(&ch, ch_ptr, sizeof(ch));
1388 carp_prepare_ad(sc, ch_ptr);
1389 ch_ptr->carp_cksum = in_cksum_skip(m, len, sizeof(*ip));
1391 getmicrotime(&cifp->if_lastchange);
1392 IFNET_STAT_INC(cifp, opackets, 1);
1393 IFNET_STAT_INC(cifp, obytes, len);
1394 carpstats.carps_opackets++;
1396 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, NULL)) {
1397 IFNET_STAT_INC(cifp, oerrors, 1);
1398 if (sc->sc_sendad_errors < INT_MAX)
1399 sc->sc_sendad_errors++;
1400 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
1401 carp_suppress_preempt++;
1402 if (carp_suppress_preempt == 1) {
1406 sc->sc_sendad_success = 0;
1408 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
1409 if (++sc->sc_sendad_success >=
1410 CARP_SENDAD_MIN_SUCCESS) {
1411 carp_suppress_preempt--;
1412 sc->sc_sendad_errors = 0;
1415 sc->sc_sendad_errors = 0;
1422 struct ip6_hdr *ip6;
1424 MGETHDR(m, M_NOWAIT, MT_HEADER);
1426 IFNET_STAT_INC(cifp, oerrors, 1);
1427 carpstats.carps_onomem++;
1428 /* XXX maybe less ? */
1429 if (advbase != 255 || advskew != 255)
1430 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
1431 carp_send_ad_timeout, sc);
1434 len = sizeof(*ip6) + sizeof(ch);
1435 m->m_pkthdr.len = len;
1436 m->m_pkthdr.rcvif = NULL;
1438 MH_ALIGN(m, m->m_len);
1439 m->m_flags |= M_MCAST;
1440 ip6 = mtod(m, struct ip6_hdr *);
1441 bzero(ip6, sizeof(*ip6));
1442 ip6->ip6_vfc |= IPV6_VERSION;
1443 ip6->ip6_hlim = CARP_DFLTTL;
1444 ip6->ip6_nxt = IPPROTO_CARP;
1445 bcopy(&sc->sc_ia6->ia_addr.sin6_addr, &ip6->ip6_src,
1446 sizeof(struct in6_addr));
1447 /* set the multicast destination */
1449 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
1450 ip6->ip6_dst.s6_addr8[15] = 0x12;
1451 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
1452 IFNET_STAT_INC(cifp, oerrors, 1);
1454 CARP_LOG("%s: in6_setscope failed\n", __func__);
1458 ch_ptr = (struct carp_header *)(&ip6[1]);
1459 bcopy(&ch, ch_ptr, sizeof(ch));
1460 carp_prepare_ad(sc, ch_ptr);
1461 ch_ptr->carp_cksum = in_cksum_skip(m, len, sizeof(*ip6));
1463 getmicrotime(&cifp->if_lastchange);
1464 IFNET_STAT_INC(cifp, opackets, 1);
1465 IFNET_STAT_INC(cifp, obytes, len);
1466 carpstats.carps_opackets6++;
1468 if (ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL)) {
1469 IFNET_STAT_INC(cifp, oerrors, 1);
1470 if (sc->sc_sendad_errors < INT_MAX)
1471 sc->sc_sendad_errors++;
1472 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
1473 carp_suppress_preempt++;
1474 if (carp_suppress_preempt == 1) {
1478 sc->sc_sendad_success = 0;
1480 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
1481 if (++sc->sc_sendad_success >=
1482 CARP_SENDAD_MIN_SUCCESS) {
1483 carp_suppress_preempt--;
1484 sc->sc_sendad_errors = 0;
1487 sc->sc_sendad_errors = 0;
1493 if (advbase != 255 || advskew != 255)
1494 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
1495 carp_send_ad_timeout, sc);
1499 * Broadcast a gratuitous ARP request containing
1500 * the virtual router MAC address for each IP address
1501 * associated with the virtual router.
1504 carp_send_arp(struct carp_softc *sc)
1506 const struct carp_vhaddr *vha;
1508 TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
1509 if (vha->vha_iaback == NULL)
1511 arp_gratuitous(&sc->sc_if, &vha->vha_ia->ia_ifa);
1517 carp_send_na(struct carp_softc *sc)
1519 struct ifaddr_container *ifac;
1520 struct in6_addr *in6;
1521 static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1523 TAILQ_FOREACH(ifac, &sc->sc_if.if_addrheads[mycpuid], ifa_link) {
1524 struct ifaddr *ifa = ifac->ifa;
1526 if (ifa->ifa_addr->sa_family != AF_INET6)
1529 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr;
1530 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1531 ND_NA_FLAG_OVERRIDE, 1, NULL);
1532 DELAY(1000); /* XXX */
1537 static __inline const struct carp_vhaddr *
1538 carp_find_addr(const struct carp_softc *sc, const struct in_addr *addr)
1540 struct carp_vhaddr *vha;
1542 TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
1543 if (vha->vha_iaback == NULL)
1546 if (vha->vha_ia->ia_addr.sin_addr.s_addr == addr->s_addr)
1554 carp_iamatch_balance(const struct carp_if *cif, const struct in_addr *itaddr,
1555 const struct in_addr *isaddr, uint8_t **enaddr)
1557 const struct carp_softc *vh;
1558 int index, count = 0;
1561 * XXX proof of concept implementation.
1562 * We use the source ip to decide which virtual host should
1563 * handle the request. If we're master of that virtual host,
1564 * then we respond, otherwise, just drop the arp packet on
1568 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1569 if (!CARP_IS_RUNNING(&vh->sc_if))
1572 if (carp_find_addr(vh, itaddr) != NULL)
1578 /* this should be a hash, like pf_hash() */
1579 index = ntohl(isaddr->s_addr) % count;
1582 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1583 if (!CARP_IS_RUNNING(&vh->sc_if))
1586 if (carp_find_addr(vh, itaddr) == NULL)
1589 if (count == index) {
1590 if (vh->sc_state == MASTER) {
1591 *enaddr = IF_LLADDR(&vh->sc_if);
1604 carp_iamatch(const struct in_ifaddr *ia)
1606 const struct carp_softc *sc = ia->ia_ifp->if_softc;
1611 if (carp_opts[CARPCTL_ARPBALANCE])
1612 return carp_iamatch_balance(cif, itaddr, isaddr, enaddr);
1615 if (!CARP_IS_RUNNING(&sc->sc_if) || sc->sc_state != MASTER)
1623 carp_iamatch6(void *v, struct in6_addr *taddr)
1626 struct carp_if *cif = v;
1627 struct carp_softc *vh;
1629 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1630 struct ifaddr_container *ifac;
1632 TAILQ_FOREACH(ifac, &vh->sc_if.if_addrheads[mycpuid],
1634 struct ifaddr *ifa = ifac->ifa;
1636 if (IN6_ARE_ADDR_EQUAL(taddr,
1637 &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1638 CARP_IS_RUNNING(&vh->sc_if) &&
1639 vh->sc_state == MASTER) {
1649 carp_macmatch6(void *v, struct mbuf *m, const struct in6_addr *taddr)
1653 struct carp_if *cif = v;
1654 struct carp_softc *sc;
1656 TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) {
1657 struct ifaddr_container *ifac;
1659 TAILQ_FOREACH(ifac, &sc->sc_if.if_addrheads[mycpuid],
1661 struct ifaddr *ifa = ifac->ifa;
1663 if (IN6_ARE_ADDR_EQUAL(taddr,
1664 &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1665 CARP_IS_RUNNING(&sc->sc_if)) {
1666 struct ifnet *ifp = &sc->sc_if;
1668 mtag = m_tag_get(PACKET_TAG_CARP,
1669 sizeof(struct ifnet *), M_NOWAIT);
1671 /* better a bit than nothing */
1672 return (IF_LLADDR(ifp));
1674 bcopy(&ifp, (caddr_t)(mtag + 1),
1675 sizeof(struct ifnet *));
1676 m_tag_prepend(m, mtag);
1678 return (IF_LLADDR(ifp));
1687 static struct ifnet *
1688 carp_forus(struct carp_if *cif, const uint8_t *dhost)
1690 struct carp_softc_container *scc;
1692 if (memcmp(dhost, carp_etheraddr, ETHER_ADDR_LEN - 1) != 0)
1695 TAILQ_FOREACH(scc, cif, scc_link) {
1696 struct carp_softc *sc = scc->scc_softc;
1697 struct ifnet *ifp = &sc->sc_if;
1699 if (CARP_IS_RUNNING(ifp) && sc->sc_state == MASTER &&
1700 !bcmp(dhost, IF_LLADDR(ifp), ETHER_ADDR_LEN))
1707 carp_master_down_timeout(void *xsc)
1709 struct carp_softc *sc = xsc;
1710 struct netmsg_carp *cmsg = &sc->sc_md_msg;
1712 KASSERT(mycpuid == 0, ("%s not on cpu0 but on cpu%d",
1713 __func__, mycpuid));
1716 if (cmsg->base.lmsg.ms_flags & MSGF_DONE)
1717 lwkt_sendmsg_oncpu(netisr_cpuport(0), &cmsg->base.lmsg);
1722 carp_master_down_timeout_dispatch(netmsg_t msg)
1724 struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
1725 struct carp_softc *sc = cmsg->nc_softc;
1729 lwkt_replymsg(&cmsg->base.lmsg, 0);
1732 CARP_DEBUG("%s: BACKUP -> MASTER (master timed out)\n",
1733 sc->sc_if.if_xname);
1734 carp_master_down(sc);
1738 carp_master_down(struct carp_softc *sc)
1740 switch (sc->sc_state) {
1742 kprintf("%s: master_down event in INIT state\n",
1743 sc->sc_if.if_xname);
1750 carp_set_state(sc, MASTER);
1757 if (carp_opts[CARPCTL_SETROUTE])
1758 carp_setroute(sc, RTM_ADD);
1764 * When in backup state, af indicates whether to reset the master down timer
1765 * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1768 carp_setrun(struct carp_softc *sc, sa_family_t af)
1770 struct ifnet *cifp = &sc->sc_if;
1773 if (sc->sc_carpdev == NULL) {
1774 carp_set_state(sc, INIT);
1778 if ((cifp->if_flags & IFF_RUNNING) && sc->sc_vhid > 0 &&
1779 (sc->sc_naddrs || sc->sc_naddrs6)) {
1782 if (carp_opts[CARPCTL_SETROUTE])
1783 carp_setroute(sc, RTM_DELETE);
1787 switch (sc->sc_state) {
1789 if (carp_opts[CARPCTL_PREEMPT] && !carp_suppress_preempt) {
1795 CARP_DEBUG("%s: INIT -> MASTER (preempting)\n",
1797 carp_set_state(sc, MASTER);
1798 if (carp_opts[CARPCTL_SETROUTE])
1799 carp_setroute(sc, RTM_ADD);
1801 CARP_DEBUG("%s: INIT -> BACKUP\n", cifp->if_xname);
1802 carp_set_state(sc, BACKUP);
1803 if (carp_opts[CARPCTL_SETROUTE])
1804 carp_setroute(sc, RTM_DELETE);
1810 callout_stop(&sc->sc_ad_tmo);
1811 tv.tv_sec = 3 * sc->sc_advbase;
1812 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1816 callout_reset(&sc->sc_md_tmo, tvtohz_high(&tv),
1817 carp_master_down_timeout, sc);
1822 callout_reset(&sc->sc_md6_tmo, tvtohz_high(&tv),
1823 carp_master_down_timeout, sc);
1828 callout_reset(&sc->sc_md_tmo, tvtohz_high(&tv),
1829 carp_master_down_timeout, sc);
1831 callout_reset(&sc->sc_md6_tmo, tvtohz_high(&tv),
1832 carp_master_down_timeout, sc);
1838 tv.tv_sec = sc->sc_advbase;
1839 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1840 callout_reset(&sc->sc_ad_tmo, tvtohz_high(&tv),
1841 carp_send_ad_timeout, sc);
1847 carp_multicast_cleanup(struct carp_softc *sc)
1849 struct ip_moptions *imo = &sc->sc_imo;
1851 if (imo->imo_num_memberships == 0)
1853 KKASSERT(imo->imo_num_memberships == 1);
1855 in_delmulti(imo->imo_membership[0]);
1856 imo->imo_membership[0] = NULL;
1857 imo->imo_num_memberships = 0;
1858 imo->imo_multicast_ifp = NULL;
1863 carp_multicast6_cleanup(struct carp_softc *sc)
1865 struct ip6_moptions *im6o = &sc->sc_im6o;
1867 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
1868 struct in6_multi_mship *imm =
1869 LIST_FIRST(&im6o->im6o_memberships);
1871 LIST_REMOVE(imm, i6mm_chain);
1872 in6_leavegroup(imm);
1874 im6o->im6o_multicast_ifp = NULL;
1879 carp_ioctl_getvhaddr_dispatch(netmsg_t msg)
1881 struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
1882 struct carp_softc *sc = cmsg->nc_softc;
1883 const struct carp_vhaddr *vha;
1884 struct ifcarpvhaddr *carpa, *carpa0;
1885 int count, len, error = 0;
1888 TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link)
1891 if (cmsg->nc_datalen == 0) {
1892 cmsg->nc_datalen = count * sizeof(*carpa);
1894 } else if (count == 0 || cmsg->nc_datalen < sizeof(*carpa)) {
1895 cmsg->nc_datalen = 0;
1898 len = min(cmsg->nc_datalen, sizeof(*carpa) * count);
1899 KKASSERT(len >= sizeof(*carpa));
1901 carpa0 = carpa = kmalloc(len, M_TEMP, M_WAITOK | M_NULLOK | M_ZERO);
1902 if (carpa == NULL) {
1908 TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
1909 if (len < sizeof(*carpa))
1912 carpa->carpa_flags = vha->vha_flags;
1913 carpa->carpa_addr.sin_family = AF_INET;
1914 carpa->carpa_addr.sin_addr = vha->vha_ia->ia_addr.sin_addr;
1916 carpa->carpa_baddr.sin_family = AF_INET;
1917 if (vha->vha_iaback == NULL) {
1918 carpa->carpa_baddr.sin_addr.s_addr = INADDR_ANY;
1920 carpa->carpa_baddr.sin_addr =
1921 vha->vha_iaback->ia_addr.sin_addr;
1926 len -= sizeof(*carpa);
1928 cmsg->nc_datalen = sizeof(*carpa) * count;
1929 KKASSERT(cmsg->nc_datalen > 0);
1931 cmsg->nc_data = carpa0;
1934 lwkt_replymsg(&cmsg->base.lmsg, error);
1938 carp_ioctl_getvhaddr(struct carp_softc *sc, struct ifdrv *ifd)
1940 struct ifnet *ifp = &sc->arpcom.ac_if;
1941 struct netmsg_carp cmsg;
1944 ASSERT_IFNET_SERIALIZED_ALL(ifp);
1945 ifnet_deserialize_all(ifp);
1947 bzero(&cmsg, sizeof(cmsg));
1948 netmsg_init(&cmsg.base, NULL, &curthread->td_msgport, 0,
1949 carp_ioctl_getvhaddr_dispatch);
1951 cmsg.nc_datalen = ifd->ifd_len;
1953 error = lwkt_domsg(netisr_cpuport(0), &cmsg.base.lmsg, 0);
1956 if (cmsg.nc_data != NULL) {
1957 error = copyout(cmsg.nc_data, ifd->ifd_data,
1959 kfree(cmsg.nc_data, M_TEMP);
1961 ifd->ifd_len = cmsg.nc_datalen;
1963 KASSERT(cmsg.nc_data == NULL,
1964 ("%s temp vhaddr is alloc upon error", __func__));
1967 ifnet_serialize_all(ifp);
1972 carp_config_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha,
1973 struct in_ifaddr *ia_del)
1976 struct in_ifaddr *ia_if;
1977 const struct in_ifaddr *ia_vha;
1978 struct in_ifaddr_container *iac;
1979 int own, ia_match_carpdev;
1981 KKASSERT(vha->vha_ia != NULL);
1982 ia_vha = vha->vha_ia;
1986 ia_match_carpdev = 0;
1987 TAILQ_FOREACH(iac, &in_ifaddrheads[mycpuid], ia_link) {
1988 struct in_ifaddr *ia = iac->ia;
1993 if (ia->ia_ifp->if_type == IFT_CARP)
1996 if ((ia->ia_ifp->if_flags & IFF_UP) == 0)
1999 /* and, yeah, we need a multicast-capable iface too */
2000 if ((ia->ia_ifp->if_flags & IFF_MULTICAST) == 0)
2003 if (ia_vha->ia_subnetmask == ia->ia_subnetmask &&
2004 ia_vha->ia_subnet == ia->ia_subnet) {
2005 if (ia_vha->ia_addr.sin_addr.s_addr ==
2006 ia->ia_addr.sin_addr.s_addr)
2008 if (ia_if == NULL) {
2010 } else if (sc->sc_carpdev != NULL &&
2011 sc->sc_carpdev == ia->ia_ifp) {
2013 if (ia_if->ia_flags & IFA_ROUTE) {
2015 * Address with prefix route
2020 ia_match_carpdev = 1;
2021 } else if (!ia_match_carpdev) {
2022 if (ia->ia_flags & IFA_ROUTE) {
2024 * Address with prefix route
2025 * is prefered over others.
2033 carp_deactivate_vhaddr(sc, vha, FALSE);
2037 ifp = ia_if->ia_ifp;
2039 /* XXX Don't allow parent iface to be changed */
2040 if (sc->sc_carpdev != NULL && sc->sc_carpdev != ifp)
2043 return carp_activate_vhaddr(sc, vha, ifp, ia_if, own);
2047 carp_add_addr(struct carp_softc *sc, struct ifaddr *carp_ifa)
2049 struct carp_vhaddr *vha_new;
2050 struct in_ifaddr *carp_ia;
2052 struct carp_vhaddr *vha;
2055 KKASSERT(carp_ifa->ifa_addr->sa_family == AF_INET);
2056 carp_ia = ifatoia(carp_ifa);
2059 TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link)
2060 KKASSERT(vha->vha_ia != NULL && vha->vha_ia != carp_ia);
2063 vha_new = kmalloc(sizeof(*vha_new), M_CARP, M_WAITOK | M_ZERO);
2064 vha_new->vha_ia = carp_ia;
2065 carp_insert_vhaddr(sc, vha_new);
2067 if (carp_config_vhaddr(sc, vha_new, NULL) != 0) {
2069 * If the above configuration fails, it may only mean
2070 * that the new address is problematic. However, the
2071 * carp(4) interface may already have several working
2072 * addresses. Since the expected behaviour of
2073 * SIOC[AS]IFADDR is to put the NIC into working state,
2074 * we try starting the state machine manually here with
2075 * the hope that the carp(4)'s previously working
2076 * addresses still could be brought up.
2078 carp_hmac_prepare(sc);
2079 carp_set_state(sc, INIT);
2085 carp_del_addr(struct carp_softc *sc, struct ifaddr *carp_ifa)
2087 struct carp_vhaddr *vha;
2088 struct in_ifaddr *carp_ia;
2090 KKASSERT(carp_ifa->ifa_addr->sa_family == AF_INET);
2091 carp_ia = ifatoia(carp_ifa);
2093 TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
2094 KKASSERT(vha->vha_ia != NULL);
2095 if (vha->vha_ia == carp_ia)
2098 KASSERT(vha != NULL, ("no corresponding vhaddr %p", carp_ifa));
2101 * Remove the vhaddr from the list before deactivating
2102 * the vhaddr, so that the HMAC could be correctly
2103 * updated in carp_deactivate_vhaddr()
2105 carp_remove_vhaddr(sc, vha);
2107 carp_deactivate_vhaddr(sc, vha, FALSE);
2112 carp_config_addr(struct carp_softc *sc, struct ifaddr *carp_ifa)
2114 struct carp_vhaddr *vha;
2115 struct in_ifaddr *carp_ia;
2117 KKASSERT(carp_ifa->ifa_addr->sa_family == AF_INET);
2118 carp_ia = ifatoia(carp_ifa);
2120 TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
2121 KKASSERT(vha->vha_ia != NULL);
2122 if (vha->vha_ia == carp_ia)
2125 KASSERT(vha != NULL, ("no corresponding vhaddr %p", carp_ifa));
2127 /* Remove then reinsert, to keep the vhaddr list sorted */
2128 carp_remove_vhaddr(sc, vha);
2129 carp_insert_vhaddr(sc, vha);
2131 if (carp_config_vhaddr(sc, vha, NULL) != 0) {
2132 /* See the comment in carp_add_addr() */
2133 carp_hmac_prepare(sc);
2134 carp_set_state(sc, INIT);
2143 carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
2146 struct carp_if *cif;
2147 struct in6_ifaddr *ia, *ia_if;
2148 struct ip6_moptions *im6o = &sc->sc_im6o;
2149 struct in6_multi_mship *imm;
2150 struct in6_addr in6;
2153 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
2158 /* we have to do it by hands to check we won't match on us */
2159 ia_if = NULL; own = 0;
2160 for (ia = in6_ifaddr; ia; ia = ia->ia_next) {
2163 for (i = 0; i < 4; i++) {
2164 if ((sin6->sin6_addr.s6_addr32[i] &
2165 ia->ia_prefixmask.sin6_addr.s6_addr32[i]) !=
2166 (ia->ia_addr.sin6_addr.s6_addr32[i] &
2167 ia->ia_prefixmask.sin6_addr.s6_addr32[i]))
2170 /* and, yeah, we need a multicast-capable iface too */
2171 if (ia->ia_ifp != &sc->sc_if &&
2172 (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
2176 if (IN6_ARE_ADDR_EQUAL(&sin6->sin6_addr,
2177 &ia->ia_addr.sin6_addr))
2183 return (EADDRNOTAVAIL);
2187 if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
2188 (im6o->im6o_multicast_ifp && im6o->im6o_multicast_ifp != ifp))
2189 return (EADDRNOTAVAIL);
2191 if (!sc->sc_naddrs6) {
2192 im6o->im6o_multicast_ifp = ifp;
2194 /* join CARP multicast address */
2195 bzero(&in6, sizeof(in6));
2196 in6.s6_addr16[0] = htons(0xff02);
2197 in6.s6_addr8[15] = 0x12;
2198 if (in6_setscope(&in6, ifp, NULL) != 0)
2200 if ((imm = in6_joingroup(ifp, &in6, &error)) == NULL)
2202 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
2204 /* join solicited multicast address */
2205 bzero(&in6, sizeof(in6));
2206 in6.s6_addr16[0] = htons(0xff02);
2207 in6.s6_addr32[1] = 0;
2208 in6.s6_addr32[2] = htonl(1);
2209 in6.s6_addr32[3] = sin6->sin6_addr.s6_addr32[3];
2210 in6.s6_addr8[12] = 0xff;
2211 if (in6_setscope(&in6, ifp, NULL) != 0)
2213 if ((imm = in6_joingroup(ifp, &in6, &error)) == NULL)
2215 LIST_INSERT_HEAD(&im6o->im6o_memberships, imm, i6mm_chain);
2219 if (!ifp->if_carp) {
2220 cif = kmalloc(sizeof(*cif), M_CARP, M_WAITOK | M_ZERO);
2222 if ((error = ifpromisc(ifp, 1))) {
2227 TAILQ_INIT(&cif->vhif_vrs);
2230 struct carp_softc *vr;
2233 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
2234 if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
2242 sc->sc_carpdev = ifp;
2245 { /* XXX prevent endless loop if already in queue */
2246 struct carp_softc *vr, *after = NULL;
2250 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
2253 if (vr->sc_vhid < sc->sc_vhid)
2258 /* We're trying to keep things in order */
2260 TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
2262 TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
2276 /* clean up multicast memberships */
2277 if (!sc->sc_naddrs6) {
2278 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
2279 imm = LIST_FIRST(&im6o->im6o_memberships);
2280 LIST_REMOVE(imm, i6mm_chain);
2281 in6_leavegroup(imm);
2288 carp_del_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
2292 if (!--sc->sc_naddrs6) {
2293 struct carp_if *cif = sc->sc_carpdev->if_carp;
2294 struct ip6_moptions *im6o = &sc->sc_im6o;
2296 callout_stop(&sc->sc_ad_tmo);
2298 while (!LIST_EMPTY(&im6o->im6o_memberships)) {
2299 struct in6_multi_mship *imm =
2300 LIST_FIRST(&im6o->im6o_memberships);
2302 LIST_REMOVE(imm, i6mm_chain);
2303 in6_leavegroup(imm);
2305 im6o->im6o_multicast_ifp = NULL;
2307 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
2308 if (TAILQ_EMPTY(&cif->vhif_vrs)) {
2309 sc->sc_carpdev->if_carp = NULL;
2310 kfree(cif, M_IFADDR);
2321 carp_ioctl(struct ifnet *ifp, u_long cmd, caddr_t addr, struct ucred *cr)
2323 struct carp_softc *sc = ifp->if_softc;
2324 struct ifreq *ifr = (struct ifreq *)addr;
2325 struct ifdrv *ifd = (struct ifdrv *)addr;
2328 ASSERT_IFNET_SERIALIZED_ALL(ifp);
2332 if (ifp->if_flags & IFF_UP) {
2333 if ((ifp->if_flags & IFF_RUNNING) == 0)
2335 } else if (ifp->if_flags & IFF_RUNNING) {
2336 carp_ioctl_stop(sc);
2341 carp_ioctl_ifcap(sc, ifr->ifr_reqcap);
2345 error = carp_ioctl_setvh(sc, ifr->ifr_data, cr);
2349 error = carp_ioctl_getvh(sc, ifr->ifr_data, cr);
2353 switch (ifd->ifd_cmd) {
2355 error = carp_ioctl_getdevname(sc, ifd);
2359 error = carp_ioctl_getvhaddr(sc, ifd);
2369 error = ether_ioctl(ifp, cmd, addr);
2377 carp_ioctl_stop_dispatch(netmsg_t msg)
2379 struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
2380 struct carp_softc *sc = cmsg->nc_softc;
2382 carp_stop(sc, FALSE);
2383 lwkt_replymsg(&cmsg->base.lmsg, 0);
2387 carp_ioctl_stop(struct carp_softc *sc)
2389 struct ifnet *ifp = &sc->arpcom.ac_if;
2390 struct netmsg_carp cmsg;
2392 ASSERT_IFNET_SERIALIZED_ALL(ifp);
2394 ifnet_deserialize_all(ifp);
2396 bzero(&cmsg, sizeof(cmsg));
2397 netmsg_init(&cmsg.base, NULL, &curthread->td_msgport, 0,
2398 carp_ioctl_stop_dispatch);
2401 lwkt_domsg(netisr_cpuport(0), &cmsg.base.lmsg, 0);
2403 ifnet_serialize_all(ifp);
2407 carp_ioctl_setvh_dispatch(netmsg_t msg)
2409 struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
2410 struct carp_softc *sc = cmsg->nc_softc;
2411 struct ifnet *ifp = &sc->arpcom.ac_if;
2412 const struct carpreq *carpr = cmsg->nc_data;
2416 if ((ifp->if_flags & IFF_RUNNING) &&
2417 sc->sc_state != INIT && carpr->carpr_state != sc->sc_state) {
2418 switch (carpr->carpr_state) {
2420 callout_stop(&sc->sc_ad_tmo);
2421 carp_set_state(sc, BACKUP);
2423 if (carp_opts[CARPCTL_SETROUTE])
2424 carp_setroute(sc, RTM_DELETE);
2428 carp_master_down(sc);
2435 if (carpr->carpr_vhid > 0) {
2436 if (carpr->carpr_vhid > 255) {
2440 if (sc->sc_carpdev) {
2441 struct carp_if *cif = sc->sc_carpdev->if_carp;
2442 struct carp_softc_container *scc;
2444 TAILQ_FOREACH(scc, cif, scc_link) {
2445 struct carp_softc *vr = scc->scc_softc;
2448 vr->sc_vhid == carpr->carpr_vhid) {
2454 sc->sc_vhid = carpr->carpr_vhid;
2456 IF_LLADDR(ifp)[5] = sc->sc_vhid;
2457 bcopy(IF_LLADDR(ifp), sc->arpcom.ac_enaddr,
2462 if (carpr->carpr_advbase > 0 || carpr->carpr_advskew > 0) {
2463 if (carpr->carpr_advskew >= 255) {
2467 if (carpr->carpr_advbase > 255) {
2471 sc->sc_advbase = carpr->carpr_advbase;
2472 sc->sc_advskew = carpr->carpr_advskew;
2475 bcopy(carpr->carpr_key, sc->sc_key, sizeof(sc->sc_key));
2483 carp_hmac_prepare(sc);
2485 lwkt_replymsg(&cmsg->base.lmsg, error);
2489 carp_ioctl_setvh(struct carp_softc *sc, void *udata, struct ucred *cr)
2491 struct ifnet *ifp = &sc->arpcom.ac_if;
2492 struct netmsg_carp cmsg;
2493 struct carpreq carpr;
2496 ASSERT_IFNET_SERIALIZED_ALL(ifp);
2497 ifnet_deserialize_all(ifp);
2499 error = priv_check_cred(cr, PRIV_ROOT, NULL_CRED_OKAY);
2503 error = copyin(udata, &carpr, sizeof(carpr));
2507 bzero(&cmsg, sizeof(cmsg));
2508 netmsg_init(&cmsg.base, NULL, &curthread->td_msgport, 0,
2509 carp_ioctl_setvh_dispatch);
2511 cmsg.nc_data = &carpr;
2513 error = lwkt_domsg(netisr_cpuport(0), &cmsg.base.lmsg, 0);
2516 ifnet_serialize_all(ifp);
2521 carp_ioctl_ifcap_dispatch(netmsg_t msg)
2523 struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
2524 struct carp_softc *sc = cmsg->nc_softc;
2525 struct ifnet *ifp = &sc->arpcom.ac_if;
2526 int reqcap = *((const int *)(cmsg->nc_data));
2529 mask = reqcap ^ ifp->if_capenable;
2530 if (mask & IFCAP_TXCSUM) {
2531 ifp->if_capenable ^= IFCAP_TXCSUM;
2532 if ((ifp->if_capenable & IFCAP_TXCSUM) &&
2533 sc->sc_carpdev != NULL) {
2535 (sc->sc_carpdev->if_hwassist &
2536 (CSUM_IP | CSUM_UDP | CSUM_TCP));
2538 ifp->if_hwassist &= ~(CSUM_IP | CSUM_UDP | CSUM_TCP);
2541 if (mask & IFCAP_TSO) {
2542 ifp->if_capenable ^= IFCAP_TSO;
2543 if ((ifp->if_capenable & IFCAP_TSO) &&
2544 sc->sc_carpdev != NULL) {
2546 (sc->sc_carpdev->if_hwassist & CSUM_TSO);
2548 ifp->if_hwassist &= ~CSUM_TSO;
2552 lwkt_replymsg(&cmsg->base.lmsg, 0);
2556 carp_ioctl_ifcap(struct carp_softc *sc, int reqcap)
2558 struct ifnet *ifp = &sc->arpcom.ac_if;
2559 struct netmsg_carp cmsg;
2561 ASSERT_IFNET_SERIALIZED_ALL(ifp);
2562 ifnet_deserialize_all(ifp);
2564 bzero(&cmsg, sizeof(cmsg));
2565 netmsg_init(&cmsg.base, NULL, &curthread->td_msgport, 0,
2566 carp_ioctl_ifcap_dispatch);
2568 cmsg.nc_data = &reqcap;
2570 lwkt_domsg(netisr_cpuport(0), &cmsg.base.lmsg, 0);
2572 ifnet_serialize_all(ifp);
2576 carp_ioctl_getvh_dispatch(netmsg_t msg)
2578 struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
2579 struct carp_softc *sc = cmsg->nc_softc;
2580 struct carpreq *carpr = cmsg->nc_data;
2582 carpr->carpr_state = sc->sc_state;
2583 carpr->carpr_vhid = sc->sc_vhid;
2584 carpr->carpr_advbase = sc->sc_advbase;
2585 carpr->carpr_advskew = sc->sc_advskew;
2586 bcopy(sc->sc_key, carpr->carpr_key, sizeof(carpr->carpr_key));
2588 lwkt_replymsg(&cmsg->base.lmsg, 0);
2592 carp_ioctl_getvh(struct carp_softc *sc, void *udata, struct ucred *cr)
2594 struct ifnet *ifp = &sc->arpcom.ac_if;
2595 struct netmsg_carp cmsg;
2596 struct carpreq carpr;
2599 ASSERT_IFNET_SERIALIZED_ALL(ifp);
2600 ifnet_deserialize_all(ifp);
2602 bzero(&cmsg, sizeof(cmsg));
2603 netmsg_init(&cmsg.base, NULL, &curthread->td_msgport, 0,
2604 carp_ioctl_getvh_dispatch);
2606 cmsg.nc_data = &carpr;
2608 lwkt_domsg(netisr_cpuport(0), &cmsg.base.lmsg, 0);
2610 error = priv_check_cred(cr, PRIV_ROOT, NULL_CRED_OKAY);
2612 bzero(carpr.carpr_key, sizeof(carpr.carpr_key));
2614 error = copyout(&carpr, udata, sizeof(carpr));
2616 ifnet_serialize_all(ifp);
2621 carp_ioctl_getdevname_dispatch(netmsg_t msg)
2623 struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
2624 struct carp_softc *sc = cmsg->nc_softc;
2625 char *devname = cmsg->nc_data;
2627 bzero(devname, IFNAMSIZ);
2628 if (sc->sc_carpdev != NULL)
2629 strlcpy(devname, sc->sc_carpdev->if_xname, IFNAMSIZ);
2631 lwkt_replymsg(&cmsg->base.lmsg, 0);
2635 carp_ioctl_getdevname(struct carp_softc *sc, struct ifdrv *ifd)
2637 struct ifnet *ifp = &sc->arpcom.ac_if;
2638 struct netmsg_carp cmsg;
2639 char devname[IFNAMSIZ];
2642 ASSERT_IFNET_SERIALIZED_ALL(ifp);
2644 if (ifd->ifd_len != sizeof(devname))
2647 ifnet_deserialize_all(ifp);
2649 bzero(&cmsg, sizeof(cmsg));
2650 netmsg_init(&cmsg.base, NULL, &curthread->td_msgport, 0,
2651 carp_ioctl_getdevname_dispatch);
2653 cmsg.nc_data = devname;
2655 lwkt_domsg(netisr_cpuport(0), &cmsg.base.lmsg, 0);
2657 error = copyout(devname, ifd->ifd_data, sizeof(devname));
2659 ifnet_serialize_all(ifp);
2664 carp_init_dispatch(netmsg_t msg)
2666 struct netmsg_carp *cmsg = (struct netmsg_carp *)msg;
2667 struct carp_softc *sc = cmsg->nc_softc;
2669 sc->sc_if.if_flags |= IFF_RUNNING;
2670 carp_hmac_prepare(sc);
2671 carp_set_state(sc, INIT);
2674 lwkt_replymsg(&cmsg->base.lmsg, 0);
2678 carp_init(void *xsc)
2680 struct carp_softc *sc = xsc;
2681 struct ifnet *ifp = &sc->arpcom.ac_if;
2682 struct netmsg_carp cmsg;
2684 ASSERT_IFNET_SERIALIZED_ALL(ifp);
2686 ifnet_deserialize_all(ifp);
2688 bzero(&cmsg, sizeof(cmsg));
2689 netmsg_init(&cmsg.base, NULL, &curthread->td_msgport, 0,
2690 carp_init_dispatch);
2693 lwkt_domsg(netisr_cpuport(0), &cmsg.base.lmsg, 0);
2695 ifnet_serialize_all(ifp);
2699 carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *dst,
2702 struct carp_softc *sc = ifp->if_softc;
2703 struct ifnet *carpdev;
2706 carpdev = sc->sc_carpdev;
2707 if (carpdev != NULL) {
2708 if (m->m_flags & M_MCAST)
2709 IFNET_STAT_INC(ifp, omcasts, 1);
2710 IFNET_STAT_INC(ifp, obytes, m->m_pkthdr.len + ETHER_HDR_LEN);
2711 IFNET_STAT_INC(ifp, opackets, 1);
2715 * CARP's ifp is passed to backing device's
2718 carpdev->if_output(ifp, m, dst, rt);
2720 IFNET_STAT_INC(ifp, oerrors, 1);
2722 error = ENETUNREACH;
2728 * Start output on carp interface. This function should never be called.
2731 carp_start(struct ifnet *ifp, struct ifaltq_subque *ifsq __unused)
2733 panic("%s: start called", ifp->if_xname);
2737 carp_set_state(struct carp_softc *sc, int state)
2739 struct ifnet *cifp = &sc->sc_if;
2741 if (sc->sc_state == state)
2743 sc->sc_state = state;
2745 switch (sc->sc_state) {
2747 cifp->if_link_state = LINK_STATE_DOWN;
2751 cifp->if_link_state = LINK_STATE_UP;
2755 cifp->if_link_state = LINK_STATE_UNKNOWN;
2762 carp_group_demote_adj(struct ifnet *ifp, int adj)
2764 struct ifg_list *ifgl;
2767 TAILQ_FOREACH(ifgl, &ifp->if_groups, ifgl_next) {
2768 if (!strcmp(ifgl->ifgl_group->ifg_group, IFG_ALL))
2770 dm = &ifgl->ifgl_group->ifg_carp_demoted;
2777 if (adj > 0 && *dm == 1)
2779 CARP_LOG("%s demoted group %s to %d", ifp->if_xname,
2780 ifgl->ifgl_group->ifg_group, *dm);
2786 carp_carpdev_state(void *v)
2788 struct carp_if *cif = v;
2789 struct carp_softc *sc;
2791 TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list)
2796 carp_sc_state(struct carp_softc *sc)
2798 if (!(sc->sc_carpdev->if_flags & IFF_UP)) {
2799 callout_stop(&sc->sc_ad_tmo);
2800 callout_stop(&sc->sc_md_tmo);
2801 callout_stop(&sc->sc_md6_tmo);
2802 carp_set_state(sc, INIT);
2804 if (!sc->sc_suppress) {
2805 carp_suppress_preempt++;
2806 if (carp_suppress_preempt == 1)
2809 sc->sc_suppress = 1;
2811 carp_set_state(sc, INIT);
2813 if (sc->sc_suppress)
2814 carp_suppress_preempt--;
2815 sc->sc_suppress = 0;
2821 carp_stop(struct carp_softc *sc, boolean_t detach)
2823 sc->sc_if.if_flags &= ~IFF_RUNNING;
2825 callout_stop(&sc->sc_ad_tmo);
2826 callout_stop(&sc->sc_md_tmo);
2827 callout_stop(&sc->sc_md6_tmo);
2829 if (!detach && sc->sc_state == MASTER)
2832 if (sc->sc_suppress)
2833 carp_suppress_preempt--;
2834 sc->sc_suppress = 0;
2836 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS)
2837 carp_suppress_preempt--;
2838 sc->sc_sendad_errors = 0;
2839 sc->sc_sendad_success = 0;
2841 carp_set_state(sc, INIT);
2846 carp_suspend(struct carp_softc *sc, boolean_t detach)
2848 struct ifnet *cifp = &sc->sc_if;
2850 carp_stop(sc, detach);
2852 /* Retain the running state, if we are not dead yet */
2853 if (!sc->sc_dead && (cifp->if_flags & IFF_UP))
2854 cifp->if_flags |= IFF_RUNNING;
2858 carp_activate_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha,
2859 struct ifnet *ifp, struct in_ifaddr *ia_if, int own)
2861 struct ip_moptions *imo = &sc->sc_imo;
2862 struct carp_if *ocif = ifp->if_carp;
2865 KKASSERT(vha->vha_ia != NULL);
2867 KASSERT(ia_if != NULL, ("NULL backing address"));
2868 KASSERT(vha->vha_iaback == NULL, ("%p is already activated", vha));
2869 KASSERT((vha->vha_flags & CARP_VHAF_OWNER) == 0,
2870 ("inactive vhaddr %p is the address owner", vha));
2872 KASSERT(sc->sc_carpdev == NULL || sc->sc_carpdev == ifp,
2873 ("%s is already on %s", sc->sc_if.if_xname,
2874 sc->sc_carpdev->if_xname));
2877 KASSERT(sc->sc_carpdev == NULL,
2878 ("%s is already on %s", sc->sc_if.if_xname,
2879 sc->sc_carpdev->if_xname));
2881 error = ifpromisc(ifp, 1);
2885 struct carp_softc_container *scc;
2887 TAILQ_FOREACH(scc, ocif, scc_link) {
2888 struct carp_softc *vr = scc->scc_softc;
2890 if (vr != sc && vr->sc_vhid == sc->sc_vhid)
2895 ifp->if_carp = carp_if_insert(ocif, sc);
2896 KASSERT(ifp->if_carp != NULL, ("%s carp_if_insert failed", __func__));
2899 sc->sc_carpdev = ifp;
2900 sc->arpcom.ac_if.if_hwassist = 0;
2901 if (sc->arpcom.ac_if.if_capenable & IFCAP_TXCSUM) {
2902 sc->arpcom.ac_if.if_hwassist |=
2903 (ifp->if_hwassist & (CSUM_IP | CSUM_UDP | CSUM_TCP));
2905 if (sc->arpcom.ac_if.if_capenable & IFCAP_TSO)
2906 sc->arpcom.ac_if.if_hwassist |= (ifp->if_hwassist & CSUM_TSO);
2909 * Make sure that all protocol threads see the sc_carpdev and
2912 netmsg_service_sync();
2914 if (ocif != NULL && ifp->if_carp != ocif) {
2916 * The old carp list could be safely free now,
2917 * since no one can access it.
2922 vha->vha_iaback = ia_if;
2926 vha->vha_flags |= CARP_VHAF_OWNER;
2928 /* XXX save user configured advskew? */
2932 carp_addroute_vhaddr(sc, vha);
2935 * Join the multicast group only after the backing interface
2936 * has been hooked with the CARP interface.
2938 KASSERT(imo->imo_multicast_ifp == NULL ||
2939 imo->imo_multicast_ifp == &sc->sc_if,
2940 ("%s didn't leave mcast group on %s",
2941 sc->sc_if.if_xname, imo->imo_multicast_ifp->if_xname));
2943 if (imo->imo_num_memberships == 0) {
2944 struct in_addr addr;
2946 addr.s_addr = htonl(INADDR_CARP_GROUP);
2947 imo->imo_membership[0] = in_addmulti(&addr, &sc->sc_if);
2948 if (imo->imo_membership[0] == NULL) {
2949 carp_deactivate_vhaddr(sc, vha, FALSE);
2953 imo->imo_num_memberships++;
2954 imo->imo_multicast_ifp = &sc->sc_if;
2955 imo->imo_multicast_ttl = CARP_DFLTTL;
2956 imo->imo_multicast_loop = 0;
2959 carp_hmac_prepare(sc);
2960 carp_set_state(sc, INIT);
2966 carp_deactivate_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha,
2967 boolean_t del_iaback)
2969 KKASSERT(vha->vha_ia != NULL);
2971 carp_hmac_prepare(sc);
2973 if (vha->vha_iaback == NULL) {
2974 KASSERT((vha->vha_flags & CARP_VHAF_OWNER) == 0,
2975 ("inactive vhaddr %p is the address owner", vha));
2979 vha->vha_flags &= ~CARP_VHAF_OWNER;
2980 carp_delroute_vhaddr(sc, vha, del_iaback);
2982 KKASSERT(sc->sc_naddrs > 0);
2983 vha->vha_iaback = NULL;
2985 if (!sc->sc_naddrs) {
2986 if (sc->sc_naddrs6) {
2987 carp_multicast_cleanup(sc);
2990 carp_detach(sc, FALSE, del_iaback);
2996 carp_link_addrs(struct carp_softc *sc, struct ifnet *ifp, struct ifaddr *ifa_if)
2998 struct carp_vhaddr *vha;
2999 struct in_ifaddr *ia_if;
3001 KKASSERT(ifa_if->ifa_addr->sa_family == AF_INET);
3002 ia_if = ifatoia(ifa_if);
3005 * Test each inactive vhaddr against the newly added address.
3006 * If the newly added address could be the backing address,
3007 * then activate the matching vhaddr.
3009 TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
3010 const struct in_ifaddr *ia;
3013 if (vha->vha_iaback != NULL)
3017 if (ia->ia_subnetmask != ia_if->ia_subnetmask ||
3018 ia->ia_subnet != ia_if->ia_subnet)
3022 if (ia->ia_addr.sin_addr.s_addr ==
3023 ia_if->ia_addr.sin_addr.s_addr)
3026 carp_activate_vhaddr(sc, vha, ifp, ia_if, own);
3031 carp_unlink_addrs(struct carp_softc *sc, struct ifnet *ifp,
3032 struct ifaddr *ifa_if)
3034 struct carp_vhaddr *vha;
3035 struct in_ifaddr *ia_if;
3037 KKASSERT(ifa_if->ifa_addr->sa_family == AF_INET);
3038 ia_if = ifatoia(ifa_if);
3041 * Ad src address is deleted; set it to NULL.
3042 * Following loop will try pick up a new ad src address
3043 * if one of the vhaddr could retain its backing address.
3045 if (sc->sc_ia == ia_if)
3049 * Test each active vhaddr against the deleted address.
3050 * If the deleted address is vhaddr address's backing
3051 * address, then deactivate the vhaddr.
3053 TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link) {
3054 if (vha->vha_iaback == NULL)
3057 if (vha->vha_iaback == ia_if)
3058 carp_deactivate_vhaddr(sc, vha, TRUE);
3059 else if (sc->sc_ia == NULL)
3060 sc->sc_ia = vha->vha_iaback;
3065 carp_update_addrs(struct carp_softc *sc, struct ifaddr *ifa_del)
3067 struct carp_vhaddr *vha;
3069 KKASSERT(sc->sc_carpdev == NULL);
3071 TAILQ_FOREACH(vha, &sc->sc_vha_list, vha_link)
3072 carp_config_vhaddr(sc, vha, ifatoia(ifa_del));
3076 carp_ifaddr(void *arg __unused, struct ifnet *ifp,
3077 enum ifaddr_event event, struct ifaddr *ifa)
3079 struct carp_softc *sc;
3081 if (ifa->ifa_addr->sa_family != AF_INET)
3086 if (ifp->if_type == IFT_CARP) {
3088 * Address is changed on carp(4) interface
3091 case IFADDR_EVENT_ADD:
3092 carp_add_addr(ifp->if_softc, ifa);
3095 case IFADDR_EVENT_CHANGE:
3096 carp_config_addr(ifp->if_softc, ifa);
3099 case IFADDR_EVENT_DELETE:
3100 carp_del_addr(ifp->if_softc, ifa);
3107 * Address is changed on non-carp(4) interface
3109 if ((ifp->if_flags & IFF_MULTICAST) == 0)
3112 LIST_FOREACH(sc, &carpif_list, sc_next) {
3113 if (sc->sc_carpdev != NULL && sc->sc_carpdev != ifp) {
3114 /* Not the parent iface; skip */
3119 case IFADDR_EVENT_ADD:
3120 carp_link_addrs(sc, ifp, ifa);
3123 case IFADDR_EVENT_DELETE:
3124 if (sc->sc_carpdev != NULL) {
3125 carp_unlink_addrs(sc, ifp, ifa);
3126 if (sc->sc_carpdev == NULL) {
3128 * We no longer have the parent
3129 * interface, however, certain
3130 * virtual addresses, which are
3131 * not used because they can't
3132 * match the previous parent
3133 * interface's addresses, may now
3134 * match different interface's
3137 carp_update_addrs(sc, ifa);
3141 * The carp(4) interface didn't have a
3142 * parent iface, so it is not possible
3143 * that it will contain any address to
3149 case IFADDR_EVENT_CHANGE:
3150 if (sc->sc_carpdev == NULL) {
3152 * The carp(4) interface didn't have a
3153 * parent iface, so it is not possible
3154 * that it will contain any address to
3157 carp_link_addrs(sc, ifp, ifa);
3160 * First try breaking tie with the old
3161 * address. Then see whether we could
3162 * link certain vhaddr to the new address.
3163 * If that fails, i.e. carpdev is NULL,
3164 * we try a global update.
3166 * NOTE: The above order is critical.
3168 carp_unlink_addrs(sc, ifp, ifa);
3169 carp_link_addrs(sc, ifp, ifa);
3170 if (sc->sc_carpdev == NULL) {
3172 * See the comment in the above
3173 * IFADDR_EVENT_DELETE block.
3175 carp_update_addrs(sc, NULL);
3184 carp_proto_ctlinput(netmsg_t msg)
3186 int cmd = msg->ctlinput.nm_cmd;
3187 struct sockaddr *sa = msg->ctlinput.nm_arg;
3188 struct in_ifaddr_container *iac;
3190 /* We only process PRC_IFDOWN and PRC_IFUP commands */
3191 if (cmd != PRC_IFDOWN && cmd != PRC_IFUP)
3194 TAILQ_FOREACH(iac, &in_ifaddrheads[mycpuid], ia_link) {
3195 struct in_ifaddr *ia = iac->ia;
3196 struct ifnet *ifp = ia->ia_ifp;
3198 if (ifp->if_type == IFT_CARP)
3201 if (ia->ia_ifa.ifa_addr == sa) {
3202 if (cmd == PRC_IFDOWN) {
3203 carp_ifaddr(NULL, ifp, IFADDR_EVENT_DELETE,
3205 } else if (cmd == PRC_IFUP) {
3206 carp_ifaddr(NULL, ifp, IFADDR_EVENT_ADD,
3213 lwkt_replymsg(&msg->lmsg, 0);
3217 carp_parent(struct ifnet *cifp)
3219 struct carp_softc *sc;
3221 KKASSERT(cifp->if_type == IFT_CARP);
3222 sc = cifp->if_softc;
3224 return sc->sc_carpdev;
3227 #define rtinitflags(x) \
3228 (((x)->ia_ifp->if_flags & (IFF_LOOPBACK | IFF_POINTOPOINT)) \
3232 carp_addroute_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha)
3234 struct in_ifaddr *ia, *iaback;
3236 if (sc->sc_state != MASTER)
3240 KKASSERT(ia != NULL);
3242 iaback = vha->vha_iaback;
3243 KKASSERT(iaback != NULL);
3245 return rtchange(&iaback->ia_ifa, &ia->ia_ifa);
3249 carp_delroute_vhaddr(struct carp_softc *sc, struct carp_vhaddr *vha,
3250 boolean_t del_iaback)
3252 struct in_ifaddr *ia, *iaback;
3255 KKASSERT(ia != NULL);
3257 iaback = vha->vha_iaback;
3258 KKASSERT(iaback != NULL);
3260 if (!del_iaback && (iaback->ia_ifp->if_flags & IFF_UP)) {
3261 rtchange(&ia->ia_ifa, &iaback->ia_ifa);
3265 rtinit(&ia->ia_ifa, RTM_DELETE, rtinitflags(ia));
3266 in_ifadown_force(&ia->ia_ifa, 1);
3267 ia->ia_flags &= ~IFA_ROUTE;
3271 carp_modevent(module_t mod, int type, void *data)
3275 LIST_INIT(&carpif_list);
3276 carp_ifdetach_event =
3277 EVENTHANDLER_REGISTER(ifnet_detach_event, carp_ifdetach, NULL,
3278 EVENTHANDLER_PRI_ANY);
3280 EVENTHANDLER_REGISTER(ifaddr_event, carp_ifaddr, NULL,
3281 EVENTHANDLER_PRI_FIRST);
3282 if_clone_attach(&carp_cloner);
3286 EVENTHANDLER_DEREGISTER(ifnet_detach_event,
3287 carp_ifdetach_event);
3288 EVENTHANDLER_DEREGISTER(ifaddr_event,
3290 if_clone_detach(&carp_cloner);
3299 static moduledata_t carp_mod = {
3304 DECLARE_MODULE(carp, carp_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);