Deleted Added
sdiff udiff text old ( 227459 ) new ( 228571 )
full compact
ip_carp.c (227459) ip_carp.c (228571)
1/*
2 * Copyright (c) 2002 Michael Shalayeff. All rights reserved.
3 * Copyright (c) 2003 Ryan McBride. All rights reserved.
1/*-
2 * Copyright (c) 2002 Michael Shalayeff.
3 * Copyright (c) 2003 Ryan McBride.
4 * Copyright (c) 2011 Gleb Smirnoff <glebius@FreeBSD.org>
5 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
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

--- 8 unchanged lines hidden (view full) ---

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.
25 */
26
27#include <sys/cdefs.h>
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the

--- 8 unchanged lines hidden (view full) ---

22 * SERVICES; LOSS OF MIND, USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
24 * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
25 * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
26 * THE POSSIBILITY OF SUCH DAMAGE.
27 */
28
29#include <sys/cdefs.h>
28__FBSDID("$FreeBSD: head/sys/netinet/ip_carp.c 227459 2011-11-11 22:57:52Z brooks $");
30__FBSDID("$FreeBSD: head/sys/netinet/ip_carp.c 228571 2011-12-16 12:16:56Z glebius $");
29
30#include "opt_bpf.h"
31#include "opt_inet.h"
32#include "opt_inet6.h"
33
31
32#include "opt_bpf.h"
33#include "opt_inet.h"
34#include "opt_inet6.h"
35
34#include <sys/types.h>
35#include <sys/param.h>
36#include <sys/systm.h>
36#include <sys/param.h>
37#include <sys/systm.h>
37#include <sys/conf.h>
38#include <sys/bus.h>
39#include <sys/jail.h>
38#include <sys/kernel.h>
39#include <sys/limits.h>
40#include <sys/malloc.h>
41#include <sys/mbuf.h>
42#include <sys/module.h>
40#include <sys/kernel.h>
41#include <sys/limits.h>
42#include <sys/malloc.h>
43#include <sys/mbuf.h>
44#include <sys/module.h>
43#include <sys/time.h>
44#include <sys/priv.h>
45#include <sys/proc.h>
46#include <sys/protosw.h>
45#include <sys/priv.h>
46#include <sys/proc.h>
47#include <sys/protosw.h>
48#include <sys/socket.h>
49#include <sys/sockio.h>
47#include <sys/sysctl.h>
48#include <sys/syslog.h>
50#include <sys/sysctl.h>
51#include <sys/syslog.h>
49#include <sys/signalvar.h>
50#include <sys/filio.h>
51#include <sys/sockio.h>
52
52
53#include <sys/socket.h>
54#include <sys/vnode.h>
55
56#include <machine/stdarg.h>
57
58#include <net/bpf.h>
59#include <net/ethernet.h>
60#include <net/fddi.h>
53#include <net/ethernet.h>
54#include <net/fddi.h>
61#include <net/iso88025.h>
62#include <net/if.h>
55#include <net/if.h>
63#include <net/if_clone.h>
64#include <net/if_dl.h>
56#include <net/if_dl.h>
57#include <net/if_llatbl.h>
65#include <net/if_types.h>
58#include <net/if_types.h>
59#include <net/iso88025.h>
66#include <net/route.h>
67#include <net/vnet.h>
68
69#if defined(INET) || defined(INET6)
70#include <netinet/in.h>
71#include <netinet/in_var.h>
72#include <netinet/ip_carp.h>
73#include <netinet/ip.h>
60#include <net/route.h>
61#include <net/vnet.h>
62
63#if defined(INET) || defined(INET6)
64#include <netinet/in.h>
65#include <netinet/in_var.h>
66#include <netinet/ip_carp.h>
67#include <netinet/ip.h>
74
75#include <machine/in_cksum.h>
76#endif
68#include <machine/in_cksum.h>
69#endif
77
78#ifdef INET
70#ifdef INET
79#include <netinet/in_systm.h>
80#include <netinet/ip_var.h>
81#include <netinet/if_ether.h>
82#endif
83
84#ifdef INET6
85#include <netinet/icmp6.h>
86#include <netinet/ip6.h>
87#include <netinet6/ip6protosw.h>
71#include <netinet/ip_var.h>
72#include <netinet/if_ether.h>
73#endif
74
75#ifdef INET6
76#include <netinet/icmp6.h>
77#include <netinet/ip6.h>
78#include <netinet6/ip6protosw.h>
79#include <netinet6/in6_var.h>
88#include <netinet6/ip6_var.h>
89#include <netinet6/scope6_var.h>
80#include <netinet6/ip6_var.h>
81#include <netinet6/scope6_var.h>
90#include <netinet6/in6_var.h>
91#include <netinet6/nd6.h>
92#endif
93
94#include <crypto/sha1.h>
95
82#include <netinet6/nd6.h>
83#endif
84
85#include <crypto/sha1.h>
86
96#define CARP_IFNAME "carp"
97static MALLOC_DEFINE(M_CARP, "CARP", "CARP interfaces");
98SYSCTL_DECL(_net_inet_carp);
87static MALLOC_DEFINE(M_CARP, "CARP", "CARP addresses");
99
100struct carp_softc {
88
89struct carp_softc {
101 struct ifnet *sc_ifp; /* Interface clue */
102 struct ifnet *sc_carpdev; /* Pointer to parent interface */
103 struct in_ifaddr *sc_ia; /* primary iface address */
90 struct ifnet *sc_carpdev; /* Pointer to parent ifnet. */
91 struct ifaddr **sc_ifas; /* Our ifaddrs. */
92 struct sockaddr_dl sc_addr; /* Our link level address. */
93 struct callout sc_ad_tmo; /* Advertising timeout. */
104#ifdef INET
94#ifdef INET
105 struct ip_moptions sc_imo;
95 struct callout sc_md_tmo; /* Master down timeout. */
106#endif
107#ifdef INET6
96#endif
97#ifdef INET6
108 struct in6_ifaddr *sc_ia6; /* primary iface address v6 */
109 struct ip6_moptions sc_im6o;
110#endif /* INET6 */
111 TAILQ_ENTRY(carp_softc) sc_list;
98 struct callout sc_md6_tmo; /* XXX: Master down timeout. */
99#endif
100 struct mtx sc_mtx;
112
101
113 enum { INIT = 0, BACKUP, MASTER } sc_state;
102 int sc_vhid;
103 int sc_advskew;
104 int sc_advbase;
114
105
115 int sc_flags_backup;
116 int sc_suppress;
117
118 int sc_sendad_errors;
106 int sc_naddrs;
107 int sc_naddrs6;
108 int sc_ifasiz;
109 enum { INIT = 0, BACKUP, MASTER } sc_state;
110 int sc_suppress;
111 int sc_sendad_errors;
119#define CARP_SENDAD_MAX_ERRORS 3
112#define CARP_SENDAD_MAX_ERRORS 3
120 int sc_sendad_success;
113 int sc_sendad_success;
121#define CARP_SENDAD_MIN_SUCCESS 3
122
114#define CARP_SENDAD_MIN_SUCCESS 3
115
123 int sc_vhid;
124 int sc_advskew;
125 int sc_naddrs;
126 int sc_naddrs6;
127 int sc_advbase; /* seconds */
128 int sc_init_counter;
129 u_int64_t sc_counter;
116 int sc_init_counter;
117 uint64_t sc_counter;
130
131 /* authentication */
118
119 /* authentication */
132#define CARP_HMAC_PAD 64
120#define CARP_HMAC_PAD 64
133 unsigned char sc_key[CARP_KEY_LEN];
134 unsigned char sc_pad[CARP_HMAC_PAD];
135 SHA1_CTX sc_sha1;
136
121 unsigned char sc_key[CARP_KEY_LEN];
122 unsigned char sc_pad[CARP_HMAC_PAD];
123 SHA1_CTX sc_sha1;
124
137 struct callout sc_ad_tmo; /* advertisement timeout */
138 struct callout sc_md_tmo; /* master down timeout */
139 struct callout sc_md6_tmo; /* master down timeout */
140
141 LIST_ENTRY(carp_softc) sc_next; /* Interface clue */
125 TAILQ_ENTRY(carp_softc) sc_list; /* On the carp_if list. */
126 LIST_ENTRY(carp_softc) sc_next; /* On the global list. */
142};
127};
143#define SC2IFP(sc) ((sc)->sc_ifp)
144
128
129struct carp_if {
130#ifdef INET
131 int cif_naddrs;
132#endif
133#ifdef INET6
134 int cif_naddrs6;
135#endif
136 TAILQ_HEAD(, carp_softc) cif_vrs;
137#ifdef INET
138 struct ip_moptions cif_imo;
139#endif
140#ifdef INET6
141 struct ip6_moptions cif_im6o;
142#endif
143 struct ifnet *cif_ifp;
144 struct mtx cif_mtx;
145};
146
147#define CARP_INET 0
148#define CARP_INET6 1
149static int proto_reg[] = {-1, -1};
150
151/*
152 * Brief design of carp(4).
153 *
154 * Any carp-capable ifnet may have a list of carp softcs hanging off
155 * its ifp->if_carp pointer. Each softc represents one unique virtual
156 * host id, or vhid. The softc has a back pointer to the ifnet. All
157 * softcs are joined in a global list, which has quite limited use.
158 *
159 * Any interface address that takes part in CARP negotiation has a
160 * pointer to the softc of its vhid, ifa->ifa_carp. That could be either
161 * AF_INET or AF_INET6 address.
162 *
163 * Although, one can get the softc's backpointer to ifnet and traverse
164 * through its ifp->if_addrhead queue to find all interface addresses
165 * involved in CARP, we keep a growable array of ifaddr pointers. This
166 * allows us to avoid grabbing the IF_ADDR_LOCK() in many traversals that
167 * do calls into the network stack, thus avoiding LORs.
168 *
169 * Locking:
170 *
171 * Each softc has a lock sc_mtx. It is used to synchronise carp_input_c(),
172 * callout-driven events and ioctl()s.
173 *
174 * To traverse the list of softcs on an ifnet we use CIF_LOCK(), to
175 * traverse the global list we use the mutex carp_mtx.
176 *
177 * Known issues with locking:
178 *
179 * - There is no protection for races between two ioctl() requests,
180 * neither SIOCSVH, nor SIOCAIFADDR & SIOCAIFADDR_IN6. I think that all
181 * interface ioctl()s should be serialized right in net/if.c.
182 * - Sending ad, we put the pointer to the softc in an mtag, and no reference
183 * counting is done on the softc.
184 * - On module unload we may race (?) with packet processing thread
185 * dereferencing our function pointers.
186 */
187
145int carp_suppress_preempt = 0;
188int carp_suppress_preempt = 0;
146int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 1, 0, 0 }; /* XXX for now */
189int carp_opts[CARPCTL_MAXID] = { 0, 1, 0, 1, 0, };
147SYSCTL_NODE(_net_inet, IPPROTO_CARP, carp, CTLFLAG_RW, 0, "CARP");
148SYSCTL_INT(_net_inet_carp, CARPCTL_ALLOW, allow, CTLFLAG_RW,
149 &carp_opts[CARPCTL_ALLOW], 0, "Accept incoming CARP packets");
150SYSCTL_INT(_net_inet_carp, CARPCTL_PREEMPT, preempt, CTLFLAG_RW,
151 &carp_opts[CARPCTL_PREEMPT], 0, "high-priority backup preemption mode");
152SYSCTL_INT(_net_inet_carp, CARPCTL_LOG, log, CTLFLAG_RW,
153 &carp_opts[CARPCTL_LOG], 0, "log bad carp packets");
190SYSCTL_NODE(_net_inet, IPPROTO_CARP, carp, CTLFLAG_RW, 0, "CARP");
191SYSCTL_INT(_net_inet_carp, CARPCTL_ALLOW, allow, CTLFLAG_RW,
192 &carp_opts[CARPCTL_ALLOW], 0, "Accept incoming CARP packets");
193SYSCTL_INT(_net_inet_carp, CARPCTL_PREEMPT, preempt, CTLFLAG_RW,
194 &carp_opts[CARPCTL_PREEMPT], 0, "high-priority backup preemption mode");
195SYSCTL_INT(_net_inet_carp, CARPCTL_LOG, log, CTLFLAG_RW,
196 &carp_opts[CARPCTL_LOG], 0, "log bad carp packets");
154SYSCTL_INT(_net_inet_carp, CARPCTL_ARPBALANCE, arpbalance, CTLFLAG_RW,
155 &carp_opts[CARPCTL_ARPBALANCE], 0, "balance arp responses");
156SYSCTL_INT(_net_inet_carp, OID_AUTO, suppress_preempt, CTLFLAG_RD,
157 &carp_suppress_preempt, 0, "Preemption is suppressed");
158
159struct carpstats carpstats;
160SYSCTL_STRUCT(_net_inet_carp, CARPCTL_STATS, stats, CTLFLAG_RW,
161 &carpstats, carpstats,
162 "CARP statistics (struct carpstats, netinet/ip_carp.h)");
163
197SYSCTL_INT(_net_inet_carp, OID_AUTO, suppress_preempt, CTLFLAG_RD,
198 &carp_suppress_preempt, 0, "Preemption is suppressed");
199
200struct carpstats carpstats;
201SYSCTL_STRUCT(_net_inet_carp, CARPCTL_STATS, stats, CTLFLAG_RW,
202 &carpstats, carpstats,
203 "CARP statistics (struct carpstats, netinet/ip_carp.h)");
204
164struct carp_if {
165 TAILQ_HEAD(, carp_softc) vhif_vrs;
166 int vhif_nvrs;
167
168 struct ifnet *vhif_ifp;
169 struct mtx vhif_mtx;
170};
171
172#define CARP_INET 0
173#define CARP_INET6 1
174static int proto_reg[] = {-1, -1};
175
176/* Get carp_if from softc. Valid after carp_set_addr{,6}. */
177#define SC2CIF(sc) ((struct carp_if *)(sc)->sc_carpdev->if_carp)
178
179/* lock per carp_if queue */
180#define CARP_LOCK_INIT(cif) mtx_init(&(cif)->vhif_mtx, "carp_if", \
205#define CARP_LOCK_INIT(sc) mtx_init(&(sc)->sc_mtx, "carp_softc", \
181 NULL, MTX_DEF)
206 NULL, MTX_DEF)
182#define CARP_LOCK_DESTROY(cif) mtx_destroy(&(cif)->vhif_mtx)
183#define CARP_LOCK_ASSERT(cif) mtx_assert(&(cif)->vhif_mtx, MA_OWNED)
184#define CARP_LOCK(cif) mtx_lock(&(cif)->vhif_mtx)
185#define CARP_UNLOCK(cif) mtx_unlock(&(cif)->vhif_mtx)
207#define CARP_LOCK_DESTROY(sc) mtx_destroy(&(sc)->sc_mtx)
208#define CARP_LOCK_ASSERT(sc) mtx_assert(&(sc)->sc_mtx, MA_OWNED)
209#define CARP_LOCK(sc) mtx_lock(&(sc)->sc_mtx)
210#define CARP_UNLOCK(sc) mtx_unlock(&(sc)->sc_mtx)
211#define CIF_LOCK_INIT(cif) mtx_init(&(cif)->cif_mtx, "carp_if", \
212 NULL, MTX_DEF)
213#define CIF_LOCK_DESTROY(cif) mtx_destroy(&(cif)->cif_mtx)
214#define CIF_LOCK_ASSERT(cif) mtx_assert(&(cif)->cif_mtx, MA_OWNED)
215#define CIF_LOCK(cif) mtx_lock(&(cif)->cif_mtx)
216#define CIF_UNLOCK(cif) mtx_unlock(&(cif)->cif_mtx)
186
217
187#define CARP_SCLOCK(sc) mtx_lock(&SC2CIF(sc)->vhif_mtx)
188#define CARP_SCUNLOCK(sc) mtx_unlock(&SC2CIF(sc)->vhif_mtx)
189#define CARP_SCLOCK_ASSERT(sc) mtx_assert(&SC2CIF(sc)->vhif_mtx, MA_OWNED)
190
191#define CARP_LOG(...) do { \
192 if (carp_opts[CARPCTL_LOG] > 0) \
218#define CARP_LOG(...) do { \
219 if (carp_opts[CARPCTL_LOG] > 0) \
193 log(LOG_INFO, __VA_ARGS__); \
220 log(LOG_INFO, "carp: " __VA_ARGS__); \
194} while (0)
195
196#define CARP_DEBUG(...) do { \
197 if (carp_opts[CARPCTL_LOG] > 1) \
198 log(LOG_DEBUG, __VA_ARGS__); \
199} while (0)
200
221} while (0)
222
223#define CARP_DEBUG(...) do { \
224 if (carp_opts[CARPCTL_LOG] > 1) \
225 log(LOG_DEBUG, __VA_ARGS__); \
226} while (0)
227
201static void carp_hmac_prepare(struct carp_softc *);
202static void carp_hmac_generate(struct carp_softc *, u_int32_t *,
203 unsigned char *);
204static int carp_hmac_verify(struct carp_softc *, u_int32_t *,
205 unsigned char *);
206static void carp_setroute(struct carp_softc *, int);
228#define IFNET_FOREACH_IFA(ifp, ifa) \
229 IF_ADDR_LOCK_ASSERT(ifp); \
230 TAILQ_FOREACH((ifa), &(ifp)->if_addrhead, ifa_link) \
231 if ((ifa)->ifa_carp != NULL)
232
233#define CARP_FOREACH_IFA(sc, ifa) \
234 CARP_LOCK_ASSERT(sc); \
235 for (int _i = 0; \
236 _i < (sc)->sc_naddrs + (sc)->sc_naddrs6 && \
237 ((ifa) = sc->sc_ifas[_i]) != NULL; \
238 ++_i)
239
240#define IFNET_FOREACH_CARP(ifp, sc) \
241 CIF_LOCK_ASSERT(ifp->if_carp); \
242 TAILQ_FOREACH((sc), &(ifp)->if_carp->cif_vrs, sc_list)
243
207static void carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
244static void carp_input_c(struct mbuf *, struct carp_header *, sa_family_t);
208static int carp_clone_create(struct if_clone *, int, caddr_t);
209static void carp_clone_destroy(struct ifnet *);
210static void carpdetach(struct carp_softc *, int);
211static int carp_prepare_ad(struct mbuf *, struct carp_softc *,
212 struct carp_header *);
213static void carp_send_ad_all(void);
214static void carp_send_ad(void *);
215static void carp_send_ad_locked(struct carp_softc *);
216#ifdef INET
217static void carp_send_arp(struct carp_softc *);
218#endif
245static struct carp_softc
246 *carp_alloc(struct ifnet *);
247static void carp_destroy(struct carp_softc *);
248static struct carp_if
249 *carp_alloc_if(struct ifnet *);
250static void carp_free_if(struct carp_if *);
251static void carp_set_state(struct carp_softc *, int);
252static void carp_sc_state(struct carp_softc *);
253static void carp_setrun(struct carp_softc *, sa_family_t);
219static void carp_master_down(void *);
220static void carp_master_down_locked(struct carp_softc *);
254static void carp_master_down(void *);
255static void carp_master_down_locked(struct carp_softc *);
221static int carp_ioctl(struct ifnet *, u_long, caddr_t);
222static int carp_looutput(struct ifnet *, struct mbuf *, struct sockaddr *,
223 struct route *);
224static void carp_start(struct ifnet *);
225static void carp_setrun(struct carp_softc *, sa_family_t);
226static void carp_set_state(struct carp_softc *, int);
227#ifdef INET
228static int carp_addrcount(struct carp_if *, struct in_ifaddr *, int);
229#endif
230enum { CARP_COUNT_MASTER, CARP_COUNT_RUNNING };
256static void carp_send_ad(void *);
257static void carp_send_ad_locked(struct carp_softc *);
258static void carp_addroute(struct carp_softc *);
259static void carp_delroute(struct carp_softc *);
231
260
232#ifdef INET
233static void carp_multicast_cleanup(struct carp_softc *, int dofree);
234static int carp_set_addr(struct carp_softc *, struct sockaddr_in *);
235static int carp_del_addr(struct carp_softc *, struct sockaddr_in *);
236#endif
237static void carp_carpdev_state_locked(struct carp_if *);
238static void carp_sc_state_locked(struct carp_softc *);
239#ifdef INET6
240static void carp_send_na(struct carp_softc *);
241static int carp_set_addr6(struct carp_softc *, struct sockaddr_in6 *);
242static int carp_del_addr6(struct carp_softc *, struct sockaddr_in6 *);
243static void carp_multicast6_cleanup(struct carp_softc *, int dofree);
244#endif
245
246static LIST_HEAD(, carp_softc) carpif_list;
261static LIST_HEAD(, carp_softc) carp_list;
247static struct mtx carp_mtx;
262static struct mtx carp_mtx;
248IFC_SIMPLE_DECLARE(carp, 0);
249
263
250static eventhandler_tag if_detach_event_tag;
251
252static __inline u_int16_t
264static __inline uint16_t
253carp_cksum(struct mbuf *m, int len)
254{
255 return (in_cksum(m, len));
256}
257
258static void
259carp_hmac_prepare(struct carp_softc *sc)
260{
265carp_cksum(struct mbuf *m, int len)
266{
267 return (in_cksum(m, len));
268}
269
270static void
271carp_hmac_prepare(struct carp_softc *sc)
272{
261 u_int8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
262 u_int8_t vhid = sc->sc_vhid & 0xff;
273 uint8_t version = CARP_VERSION, type = CARP_ADVERTISEMENT;
274 uint8_t vhid = sc->sc_vhid & 0xff;
263 struct ifaddr *ifa;
264 int i, found;
265#ifdef INET
266 struct in_addr last, cur, in;
267#endif
268#ifdef INET6
269 struct in6_addr last6, cur6, in6;
270#endif
271
275 struct ifaddr *ifa;
276 int i, found;
277#ifdef INET
278 struct in_addr last, cur, in;
279#endif
280#ifdef INET6
281 struct in6_addr last6, cur6, in6;
282#endif
283
272 if (sc->sc_carpdev)
273 CARP_SCLOCK(sc);
284 CARP_LOCK_ASSERT(sc);
274
285
275 /* XXX: possible race here */
276
277 /* compute ipad from key */
286 /* Compute ipad from key. */
278 bzero(sc->sc_pad, sizeof(sc->sc_pad));
279 bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
280 for (i = 0; i < sizeof(sc->sc_pad); i++)
281 sc->sc_pad[i] ^= 0x36;
282
287 bzero(sc->sc_pad, sizeof(sc->sc_pad));
288 bcopy(sc->sc_key, sc->sc_pad, sizeof(sc->sc_key));
289 for (i = 0; i < sizeof(sc->sc_pad); i++)
290 sc->sc_pad[i] ^= 0x36;
291
283 /* precompute first part of inner hash */
292 /* Precompute first part of inner hash. */
284 SHA1Init(&sc->sc_sha1);
285 SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
286 SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
287 SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
288 SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
289#ifdef INET
290 cur.s_addr = 0;
291 do {
292 found = 0;
293 last = cur;
294 cur.s_addr = 0xffffffff;
293 SHA1Init(&sc->sc_sha1);
294 SHA1Update(&sc->sc_sha1, sc->sc_pad, sizeof(sc->sc_pad));
295 SHA1Update(&sc->sc_sha1, (void *)&version, sizeof(version));
296 SHA1Update(&sc->sc_sha1, (void *)&type, sizeof(type));
297 SHA1Update(&sc->sc_sha1, (void *)&vhid, sizeof(vhid));
298#ifdef INET
299 cur.s_addr = 0;
300 do {
301 found = 0;
302 last = cur;
303 cur.s_addr = 0xffffffff;
295 IF_ADDR_LOCK(SC2IFP(sc));
296 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
304 CARP_FOREACH_IFA(sc, ifa) {
297 in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr;
298 if (ifa->ifa_addr->sa_family == AF_INET &&
299 ntohl(in.s_addr) > ntohl(last.s_addr) &&
300 ntohl(in.s_addr) < ntohl(cur.s_addr)) {
301 cur.s_addr = in.s_addr;
302 found++;
303 }
304 }
305 in.s_addr = ifatoia(ifa)->ia_addr.sin_addr.s_addr;
306 if (ifa->ifa_addr->sa_family == AF_INET &&
307 ntohl(in.s_addr) > ntohl(last.s_addr) &&
308 ntohl(in.s_addr) < ntohl(cur.s_addr)) {
309 cur.s_addr = in.s_addr;
310 found++;
311 }
312 }
305 IF_ADDR_UNLOCK(SC2IFP(sc));
306 if (found)
307 SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur));
308 } while (found);
309#endif /* INET */
310#ifdef INET6
311 memset(&cur6, 0, sizeof(cur6));
312 do {
313 found = 0;
314 last6 = cur6;
315 memset(&cur6, 0xff, sizeof(cur6));
313 if (found)
314 SHA1Update(&sc->sc_sha1, (void *)&cur, sizeof(cur));
315 } while (found);
316#endif /* INET */
317#ifdef INET6
318 memset(&cur6, 0, sizeof(cur6));
319 do {
320 found = 0;
321 last6 = cur6;
322 memset(&cur6, 0xff, sizeof(cur6));
316 IF_ADDR_LOCK(SC2IFP(sc));
317 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
323 CARP_FOREACH_IFA(sc, ifa) {
318 in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
319 if (IN6_IS_SCOPE_EMBED(&in6))
320 in6.s6_addr16[1] = 0;
321 if (ifa->ifa_addr->sa_family == AF_INET6 &&
322 memcmp(&in6, &last6, sizeof(in6)) > 0 &&
323 memcmp(&in6, &cur6, sizeof(in6)) < 0) {
324 cur6 = in6;
325 found++;
326 }
327 }
324 in6 = ifatoia6(ifa)->ia_addr.sin6_addr;
325 if (IN6_IS_SCOPE_EMBED(&in6))
326 in6.s6_addr16[1] = 0;
327 if (ifa->ifa_addr->sa_family == AF_INET6 &&
328 memcmp(&in6, &last6, sizeof(in6)) > 0 &&
329 memcmp(&in6, &cur6, sizeof(in6)) < 0) {
330 cur6 = in6;
331 found++;
332 }
333 }
328 IF_ADDR_UNLOCK(SC2IFP(sc));
329 if (found)
330 SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6));
331 } while (found);
332#endif /* INET6 */
333
334 /* convert ipad to opad */
335 for (i = 0; i < sizeof(sc->sc_pad); i++)
336 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
334 if (found)
335 SHA1Update(&sc->sc_sha1, (void *)&cur6, sizeof(cur6));
336 } while (found);
337#endif /* INET6 */
338
339 /* convert ipad to opad */
340 for (i = 0; i < sizeof(sc->sc_pad); i++)
341 sc->sc_pad[i] ^= 0x36 ^ 0x5c;
337
338 if (sc->sc_carpdev)
339 CARP_SCUNLOCK(sc);
340}
341
342static void
342}
343
344static void
343carp_hmac_generate(struct carp_softc *sc, u_int32_t counter[2],
345carp_hmac_generate(struct carp_softc *sc, uint32_t counter[2],
344 unsigned char md[20])
345{
346 SHA1_CTX sha1ctx;
347
346 unsigned char md[20])
347{
348 SHA1_CTX sha1ctx;
349
350 CARP_LOCK_ASSERT(sc);
351
348 /* fetch first half of inner hash */
349 bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
350
351 SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
352 SHA1Final(md, &sha1ctx);
353
354 /* outer hash */
355 SHA1Init(&sha1ctx);
356 SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
357 SHA1Update(&sha1ctx, md, 20);
358 SHA1Final(md, &sha1ctx);
359}
360
361static int
352 /* fetch first half of inner hash */
353 bcopy(&sc->sc_sha1, &sha1ctx, sizeof(sha1ctx));
354
355 SHA1Update(&sha1ctx, (void *)counter, sizeof(sc->sc_counter));
356 SHA1Final(md, &sha1ctx);
357
358 /* outer hash */
359 SHA1Init(&sha1ctx);
360 SHA1Update(&sha1ctx, sc->sc_pad, sizeof(sc->sc_pad));
361 SHA1Update(&sha1ctx, md, 20);
362 SHA1Final(md, &sha1ctx);
363}
364
365static int
362carp_hmac_verify(struct carp_softc *sc, u_int32_t counter[2],
366carp_hmac_verify(struct carp_softc *sc, uint32_t counter[2],
363 unsigned char md[20])
364{
365 unsigned char md2[20];
366
367 unsigned char md[20])
368{
369 unsigned char md2[20];
370
367 CARP_SCLOCK_ASSERT(sc);
371 CARP_LOCK_ASSERT(sc);
368
369 carp_hmac_generate(sc, counter, md2);
370
371 return (bcmp(md, md2, sizeof(md2)));
372}
373
372
373 carp_hmac_generate(sc, counter, md2);
374
375 return (bcmp(md, md2, sizeof(md2)));
376}
377
374static void
375carp_setroute(struct carp_softc *sc, int cmd)
376{
377 struct ifaddr *ifa;
378 int s;
379
380 if (sc->sc_carpdev)
381 CARP_SCLOCK_ASSERT(sc);
382
383 s = splnet();
384 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
385#ifdef INET
386 if (ifa->ifa_addr->sa_family == AF_INET &&
387 sc->sc_carpdev != NULL) {
388 int count = carp_addrcount(
389 (struct carp_if *)sc->sc_carpdev->if_carp,
390 ifatoia(ifa), CARP_COUNT_MASTER);
391
392 if ((cmd == RTM_ADD && count == 1) ||
393 (cmd == RTM_DELETE && count == 0))
394 rtinit(ifa, cmd, RTF_UP | RTF_HOST);
395 }
396#endif
397 }
398 splx(s);
399}
400
401static int
402carp_clone_create(struct if_clone *ifc, int unit, caddr_t params)
403{
404
405 struct carp_softc *sc;
406 struct ifnet *ifp;
407
408 sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
409 ifp = SC2IFP(sc) = if_alloc(IFT_ETHER);
410 if (ifp == NULL) {
411 free(sc, M_CARP);
412 return (ENOSPC);
413 }
414
415 sc->sc_flags_backup = 0;
416 sc->sc_suppress = 0;
417 sc->sc_advbase = CARP_DFLTINTV;
418 sc->sc_vhid = -1; /* required setting */
419 sc->sc_advskew = 0;
420 sc->sc_init_counter = 1;
421 sc->sc_naddrs = sc->sc_naddrs6 = 0; /* M_ZERO? */
422#ifdef INET
423 sc->sc_imo.imo_membership = (struct in_multi **)malloc(
424 (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
425 M_WAITOK);
426 sc->sc_imo.imo_mfilters = NULL;
427 sc->sc_imo.imo_max_memberships = IP_MIN_MEMBERSHIPS;
428 sc->sc_imo.imo_multicast_vif = -1;
429#endif
430#ifdef INET6
431 sc->sc_im6o.im6o_membership = (struct in6_multi **)malloc(
432 (sizeof(struct in6_multi *) * IPV6_MIN_MEMBERSHIPS), M_CARP,
433 M_WAITOK);
434 sc->sc_im6o.im6o_mfilters = NULL;
435 sc->sc_im6o.im6o_max_memberships = IPV6_MIN_MEMBERSHIPS;
436 sc->sc_im6o.im6o_multicast_hlim = CARP_DFLTTL;
437#endif
438
439 callout_init(&sc->sc_ad_tmo, CALLOUT_MPSAFE);
440 callout_init(&sc->sc_md_tmo, CALLOUT_MPSAFE);
441 callout_init(&sc->sc_md6_tmo, CALLOUT_MPSAFE);
442
443 ifp->if_softc = sc;
444 if_initname(ifp, CARP_IFNAME, unit);
445 ifp->if_mtu = ETHERMTU;
446 ifp->if_flags = IFF_LOOPBACK;
447 ifp->if_ioctl = carp_ioctl;
448 ifp->if_output = carp_looutput;
449 ifp->if_start = carp_start;
450 ifp->if_type = IFT_CARP;
451 ifp->if_snd.ifq_maxlen = ifqmaxlen;
452 ifp->if_hdrlen = 0;
453 if_attach(ifp);
454 bpfattach(SC2IFP(sc), DLT_NULL, sizeof(u_int32_t));
455 mtx_lock(&carp_mtx);
456 LIST_INSERT_HEAD(&carpif_list, sc, sc_next);
457 mtx_unlock(&carp_mtx);
458 return (0);
459}
460
461static void
462carp_clone_destroy(struct ifnet *ifp)
463{
464 struct carp_softc *sc = ifp->if_softc;
465
466 if (sc->sc_carpdev)
467 CARP_SCLOCK(sc);
468 carpdetach(sc, 1); /* Returns unlocked. */
469
470 mtx_lock(&carp_mtx);
471 LIST_REMOVE(sc, sc_next);
472 mtx_unlock(&carp_mtx);
473 bpfdetach(ifp);
474 if_detach(ifp);
475 if_free(ifp);
476#ifdef INET
477 free(sc->sc_imo.imo_membership, M_CARP);
478#endif
479#ifdef INET6
480 free(sc->sc_im6o.im6o_membership, M_CARP);
481#endif
482 free(sc, M_CARP);
483}
484
485/*
378/*
486 * This function can be called on CARP interface destroy path,
487 * and in case of the removal of the underlying interface as
488 * well. We differentiate these two cases: in case of destruction
489 * of the underlying interface, we do not cleanup our multicast
490 * memberships, since they are already freed. But we purge pointers
491 * to multicast structures, since they are no longer valid, to
492 * avoid panic in future calls to carpdetach(). Also, we do not
493 * release the lock on return, because the function will be
494 * called once more, for another CARP instance on the same
495 * interface.
496 */
497static void
498carpdetach(struct carp_softc *sc, int unlock)
499{
500 struct carp_if *cif;
501
502 callout_stop(&sc->sc_ad_tmo);
503 callout_stop(&sc->sc_md_tmo);
504 callout_stop(&sc->sc_md6_tmo);
505
506 if (sc->sc_suppress)
507 carp_suppress_preempt--;
508 sc->sc_suppress = 0;
509
510 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS)
511 carp_suppress_preempt--;
512 sc->sc_sendad_errors = 0;
513
514 carp_set_state(sc, INIT);
515 SC2IFP(sc)->if_flags &= ~IFF_UP;
516 carp_setrun(sc, 0);
517#ifdef INET
518 carp_multicast_cleanup(sc, unlock);
519#endif
520#ifdef INET6
521 carp_multicast6_cleanup(sc, unlock);
522#endif
523
524 if (sc->sc_carpdev != NULL) {
525 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
526 CARP_LOCK_ASSERT(cif);
527 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
528 if (!--cif->vhif_nvrs) {
529 ifpromisc(sc->sc_carpdev, 0);
530 sc->sc_carpdev->if_carp = NULL;
531 CARP_LOCK_DESTROY(cif);
532 free(cif, M_CARP);
533 } else if (unlock)
534 CARP_UNLOCK(cif);
535 sc->sc_carpdev = NULL;
536 }
537}
538
539/* Detach an interface from the carp. */
540static void
541carp_ifdetach(void *arg __unused, struct ifnet *ifp)
542{
543 struct carp_if *cif = (struct carp_if *)ifp->if_carp;
544 struct carp_softc *sc, *nextsc;
545
546 if (cif == NULL)
547 return;
548
549 /*
550 * XXX: At the end of for() cycle the lock will be destroyed.
551 */
552 CARP_LOCK(cif);
553 for (sc = TAILQ_FIRST(&cif->vhif_vrs); sc; sc = nextsc) {
554 nextsc = TAILQ_NEXT(sc, sc_list);
555 carpdetach(sc, 0);
556 }
557}
558
559/*
560 * process input packet.
561 * we have rearranged checks order compared to the rfc,
562 * but it seems more efficient this way or not possible otherwise.
563 */
564#ifdef INET
565void
566carp_input(struct mbuf *m, int hlen)
567{
568 struct ip *ip = mtod(m, struct ip *);
569 struct carp_header *ch;
570 int iplen, len;
571
572 CARPSTATS_INC(carps_ipackets);
573
574 if (!carp_opts[CARPCTL_ALLOW]) {
575 m_freem(m);
576 return;
577 }
578
379 * process input packet.
380 * we have rearranged checks order compared to the rfc,
381 * but it seems more efficient this way or not possible otherwise.
382 */
383#ifdef INET
384void
385carp_input(struct mbuf *m, int hlen)
386{
387 struct ip *ip = mtod(m, struct ip *);
388 struct carp_header *ch;
389 int iplen, len;
390
391 CARPSTATS_INC(carps_ipackets);
392
393 if (!carp_opts[CARPCTL_ALLOW]) {
394 m_freem(m);
395 return;
396 }
397
579 /* check if received on a valid carp interface */
580 if (m->m_pkthdr.rcvif->if_carp == NULL) {
581 CARPSTATS_INC(carps_badif);
582 CARP_DEBUG("carp_input: packet received on non-carp "
583 "interface: %s\n",
584 m->m_pkthdr.rcvif->if_xname);
585 m_freem(m);
586 return;
587 }
588
589 /* verify that the IP TTL is 255. */
590 if (ip->ip_ttl != CARP_DFLTTL) {
591 CARPSTATS_INC(carps_badttl);
398 /* verify that the IP TTL is 255. */
399 if (ip->ip_ttl != CARP_DFLTTL) {
400 CARPSTATS_INC(carps_badttl);
592 CARP_DEBUG("carp_input: received ttl %d != 255 on %s\n",
401 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
593 ip->ip_ttl,
594 m->m_pkthdr.rcvif->if_xname);
595 m_freem(m);
596 return;
597 }
598
599 iplen = ip->ip_hl << 2;
600
601 if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
602 CARPSTATS_INC(carps_badlen);
402 ip->ip_ttl,
403 m->m_pkthdr.rcvif->if_xname);
404 m_freem(m);
405 return;
406 }
407
408 iplen = ip->ip_hl << 2;
409
410 if (m->m_pkthdr.len < iplen + sizeof(*ch)) {
411 CARPSTATS_INC(carps_badlen);
603 CARP_DEBUG("carp_input: received len %zd < "
604 "sizeof(struct carp_header) on %s\n",
605 m->m_len - sizeof(struct ip),
412 CARP_DEBUG("%s: received len %zd < sizeof(struct carp_header) "
413 "on %s\n", __func__, m->m_len - sizeof(struct ip),
606 m->m_pkthdr.rcvif->if_xname);
607 m_freem(m);
608 return;
609 }
610
611 if (iplen + sizeof(*ch) < m->m_len) {
612 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
613 CARPSTATS_INC(carps_hdrops);
414 m->m_pkthdr.rcvif->if_xname);
415 m_freem(m);
416 return;
417 }
418
419 if (iplen + sizeof(*ch) < m->m_len) {
420 if ((m = m_pullup(m, iplen + sizeof(*ch))) == NULL) {
421 CARPSTATS_INC(carps_hdrops);
614 CARP_DEBUG("carp_input: pullup failed\n");
422 CARP_DEBUG("%s: pullup failed\n", __func__);
615 return;
616 }
617 ip = mtod(m, struct ip *);
618 }
619 ch = (struct carp_header *)((char *)ip + iplen);
620
621 /*
622 * verify that the received packet length is
623 * equal to the CARP header
624 */
625 len = iplen + sizeof(*ch);
626 if (len > m->m_pkthdr.len) {
627 CARPSTATS_INC(carps_badlen);
423 return;
424 }
425 ip = mtod(m, struct ip *);
426 }
427 ch = (struct carp_header *)((char *)ip + iplen);
428
429 /*
430 * verify that the received packet length is
431 * equal to the CARP header
432 */
433 len = iplen + sizeof(*ch);
434 if (len > m->m_pkthdr.len) {
435 CARPSTATS_INC(carps_badlen);
628 CARP_DEBUG("carp_input: packet too short %d on %s\n",
436 CARP_DEBUG("%s: packet too short %d on %s\n", __func__,
629 m->m_pkthdr.len,
630 m->m_pkthdr.rcvif->if_xname);
631 m_freem(m);
632 return;
633 }
634
635 if ((m = m_pullup(m, len)) == NULL) {
636 CARPSTATS_INC(carps_hdrops);
637 return;
638 }
639 ip = mtod(m, struct ip *);
640 ch = (struct carp_header *)((char *)ip + iplen);
641
642 /* verify the CARP checksum */
643 m->m_data += iplen;
644 if (carp_cksum(m, len - iplen)) {
645 CARPSTATS_INC(carps_badsum);
437 m->m_pkthdr.len,
438 m->m_pkthdr.rcvif->if_xname);
439 m_freem(m);
440 return;
441 }
442
443 if ((m = m_pullup(m, len)) == NULL) {
444 CARPSTATS_INC(carps_hdrops);
445 return;
446 }
447 ip = mtod(m, struct ip *);
448 ch = (struct carp_header *)((char *)ip + iplen);
449
450 /* verify the CARP checksum */
451 m->m_data += iplen;
452 if (carp_cksum(m, len - iplen)) {
453 CARPSTATS_INC(carps_badsum);
646 CARP_DEBUG("carp_input: checksum failed on %s\n",
454 CARP_DEBUG("%s: checksum failed on %s\n", __func__,
647 m->m_pkthdr.rcvif->if_xname);
648 m_freem(m);
649 return;
650 }
651 m->m_data -= iplen;
652
653 carp_input_c(m, ch, AF_INET);
654}

--- 13 unchanged lines hidden (view full) ---

668 if (!carp_opts[CARPCTL_ALLOW]) {
669 m_freem(m);
670 return (IPPROTO_DONE);
671 }
672
673 /* check if received on a valid carp interface */
674 if (m->m_pkthdr.rcvif->if_carp == NULL) {
675 CARPSTATS_INC(carps_badif);
455 m->m_pkthdr.rcvif->if_xname);
456 m_freem(m);
457 return;
458 }
459 m->m_data -= iplen;
460
461 carp_input_c(m, ch, AF_INET);
462}

--- 13 unchanged lines hidden (view full) ---

476 if (!carp_opts[CARPCTL_ALLOW]) {
477 m_freem(m);
478 return (IPPROTO_DONE);
479 }
480
481 /* check if received on a valid carp interface */
482 if (m->m_pkthdr.rcvif->if_carp == NULL) {
483 CARPSTATS_INC(carps_badif);
676 CARP_DEBUG("carp6_input: packet received on non-carp "
677 "interface: %s\n",
678 m->m_pkthdr.rcvif->if_xname);
484 CARP_DEBUG("%s: packet received on non-carp interface: %s\n",
485 __func__, m->m_pkthdr.rcvif->if_xname);
679 m_freem(m);
680 return (IPPROTO_DONE);
681 }
682
683 /* verify that the IP TTL is 255 */
684 if (ip6->ip6_hlim != CARP_DFLTTL) {
685 CARPSTATS_INC(carps_badttl);
486 m_freem(m);
487 return (IPPROTO_DONE);
488 }
489
490 /* verify that the IP TTL is 255 */
491 if (ip6->ip6_hlim != CARP_DFLTTL) {
492 CARPSTATS_INC(carps_badttl);
686 CARP_DEBUG("carp6_input: received ttl %d != 255 on %s\n",
687 ip6->ip6_hlim,
688 m->m_pkthdr.rcvif->if_xname);
493 CARP_DEBUG("%s: received ttl %d != 255 on %s\n", __func__,
494 ip6->ip6_hlim, m->m_pkthdr.rcvif->if_xname);
689 m_freem(m);
690 return (IPPROTO_DONE);
691 }
692
693 /* verify that we have a complete carp packet */
694 len = m->m_len;
695 IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
696 if (ch == NULL) {
697 CARPSTATS_INC(carps_badlen);
495 m_freem(m);
496 return (IPPROTO_DONE);
497 }
498
499 /* verify that we have a complete carp packet */
500 len = m->m_len;
501 IP6_EXTHDR_GET(ch, struct carp_header *, m, *offp, sizeof(*ch));
502 if (ch == NULL) {
503 CARPSTATS_INC(carps_badlen);
698 CARP_DEBUG("carp6_input: packet size %u too small\n", len);
504 CARP_DEBUG("%s: packet size %u too small\n", __func__, len);
699 return (IPPROTO_DONE);
700 }
701
702
703 /* verify the CARP checksum */
704 m->m_data += *offp;
705 if (carp_cksum(m, sizeof(*ch))) {
706 CARPSTATS_INC(carps_badsum);
505 return (IPPROTO_DONE);
506 }
507
508
509 /* verify the CARP checksum */
510 m->m_data += *offp;
511 if (carp_cksum(m, sizeof(*ch))) {
512 CARPSTATS_INC(carps_badsum);
707 CARP_DEBUG("carp6_input: checksum failed, on %s\n",
513 CARP_DEBUG("%s: checksum failed, on %s\n", __func__,
708 m->m_pkthdr.rcvif->if_xname);
709 m_freem(m);
710 return (IPPROTO_DONE);
711 }
712 m->m_data -= *offp;
713
714 carp_input_c(m, ch, AF_INET6);
715 return (IPPROTO_DONE);
716}
717#endif /* INET6 */
718
719static void
720carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
721{
722 struct ifnet *ifp = m->m_pkthdr.rcvif;
514 m->m_pkthdr.rcvif->if_xname);
515 m_freem(m);
516 return (IPPROTO_DONE);
517 }
518 m->m_data -= *offp;
519
520 carp_input_c(m, ch, AF_INET6);
521 return (IPPROTO_DONE);
522}
523#endif /* INET6 */
524
525static void
526carp_input_c(struct mbuf *m, struct carp_header *ch, sa_family_t af)
527{
528 struct ifnet *ifp = m->m_pkthdr.rcvif;
529 struct ifaddr *ifa;
723 struct carp_softc *sc;
530 struct carp_softc *sc;
724 u_int64_t tmp_counter;
531 uint64_t tmp_counter;
725 struct timeval sc_tv, ch_tv;
726
727 /* verify that the VHID is valid on the receiving interface */
532 struct timeval sc_tv, ch_tv;
533
534 /* verify that the VHID is valid on the receiving interface */
728 CARP_LOCK(ifp->if_carp);
729 TAILQ_FOREACH(sc, &((struct carp_if *)ifp->if_carp)->vhif_vrs, sc_list)
730 if (sc->sc_vhid == ch->carp_vhid)
535 IF_ADDR_LOCK(ifp);
536 IFNET_FOREACH_IFA(ifp, ifa)
537 if (ifa->ifa_addr->sa_family == af &&
538 ifa->ifa_carp->sc_vhid == ch->carp_vhid) {
539 ifa_ref(ifa);
731 break;
540 break;
541 }
542 IF_ADDR_UNLOCK(ifp);
732
543
733 if (!sc || !((SC2IFP(sc)->if_flags & IFF_UP) &&
734 (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING))) {
544 if (ifa == NULL) {
735 CARPSTATS_INC(carps_badvhid);
545 CARPSTATS_INC(carps_badvhid);
736 CARP_UNLOCK(ifp->if_carp);
737 m_freem(m);
738 return;
739 }
740
546 m_freem(m);
547 return;
548 }
549
741 getmicrotime(&SC2IFP(sc)->if_lastchange);
742 SC2IFP(sc)->if_ipackets++;
743 SC2IFP(sc)->if_ibytes += m->m_pkthdr.len;
744
745 if (bpf_peers_present(SC2IFP(sc)->if_bpf)) {
746 uint32_t af1 = af;
747#ifdef INET
748 struct ip *ip = mtod(m, struct ip *);
749
750 /* BPF wants net byte order */
751 if (af == AF_INET) {
752 ip->ip_len = htons(ip->ip_len + (ip->ip_hl << 2));
753 ip->ip_off = htons(ip->ip_off);
754 }
755#endif
756 bpf_mtap2(SC2IFP(sc)->if_bpf, &af1, sizeof(af1), m);
757 }
758
759 /* verify the CARP version. */
760 if (ch->carp_version != CARP_VERSION) {
761 CARPSTATS_INC(carps_badver);
550 /* verify the CARP version. */
551 if (ch->carp_version != CARP_VERSION) {
552 CARPSTATS_INC(carps_badver);
762 SC2IFP(sc)->if_ierrors++;
763 CARP_UNLOCK(ifp->if_carp);
764 CARP_DEBUG("%s; invalid version %d\n",
765 SC2IFP(sc)->if_xname,
553 CARP_DEBUG("%s: invalid version %d\n", ifp->if_xname,
766 ch->carp_version);
554 ch->carp_version);
555 ifa_free(ifa);
767 m_freem(m);
768 return;
769 }
770
556 m_freem(m);
557 return;
558 }
559
771 /* verify the hash */
560 sc = ifa->ifa_carp;
561 CARP_LOCK(sc);
562 ifa_free(ifa);
563
772 if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
773 CARPSTATS_INC(carps_badauth);
564 if (carp_hmac_verify(sc, ch->carp_counter, ch->carp_md)) {
565 CARPSTATS_INC(carps_badauth);
774 SC2IFP(sc)->if_ierrors++;
775 CARP_UNLOCK(ifp->if_carp);
776 CARP_DEBUG("%s: incorrect hash\n", SC2IFP(sc)->if_xname);
777 m_freem(m);
778 return;
566 CARP_DEBUG("%s: incorrect hash for VHID %u@%s\n", __func__,
567 sc->sc_vhid, ifp->if_xname);
568 goto out;
779 }
780
781 tmp_counter = ntohl(ch->carp_counter[0]);
782 tmp_counter = tmp_counter<<32;
783 tmp_counter += ntohl(ch->carp_counter[1]);
784
785 /* XXX Replay protection goes here */
786

--- 14 unchanged lines hidden (view full) ---

801 case MASTER:
802 /*
803 * If we receive an advertisement from a master who's going to
804 * be more frequent than us, go into BACKUP state.
805 */
806 if (timevalcmp(&sc_tv, &ch_tv, >) ||
807 timevalcmp(&sc_tv, &ch_tv, ==)) {
808 callout_stop(&sc->sc_ad_tmo);
569 }
570
571 tmp_counter = ntohl(ch->carp_counter[0]);
572 tmp_counter = tmp_counter<<32;
573 tmp_counter += ntohl(ch->carp_counter[1]);
574
575 /* XXX Replay protection goes here */
576

--- 14 unchanged lines hidden (view full) ---

591 case MASTER:
592 /*
593 * If we receive an advertisement from a master who's going to
594 * be more frequent than us, go into BACKUP state.
595 */
596 if (timevalcmp(&sc_tv, &ch_tv, >) ||
597 timevalcmp(&sc_tv, &ch_tv, ==)) {
598 callout_stop(&sc->sc_ad_tmo);
809 CARP_LOG("%s: MASTER -> BACKUP "
810 "(more frequent advertisement received)\n",
811 SC2IFP(sc)->if_xname);
599 CARP_LOG("VHID %u@%s: MASTER -> BACKUP "
600 "(more frequent advertisement received)\n",
601 sc->sc_vhid,
602 sc->sc_carpdev->if_xname);
812 carp_set_state(sc, BACKUP);
813 carp_setrun(sc, 0);
603 carp_set_state(sc, BACKUP);
604 carp_setrun(sc, 0);
814 carp_setroute(sc, RTM_DELETE);
605 carp_delroute(sc);
815 }
816 break;
817 case BACKUP:
818 /*
819 * If we're pre-empting masters who advertise slower than us,
820 * and this one claims to be slower, treat him as down.
821 */
822 if (carp_opts[CARPCTL_PREEMPT] &&
823 timevalcmp(&sc_tv, &ch_tv, <)) {
606 }
607 break;
608 case BACKUP:
609 /*
610 * If we're pre-empting masters who advertise slower than us,
611 * and this one claims to be slower, treat him as down.
612 */
613 if (carp_opts[CARPCTL_PREEMPT] &&
614 timevalcmp(&sc_tv, &ch_tv, <)) {
824 CARP_LOG("%s: BACKUP -> MASTER "
615 CARP_LOG("VHID %u@%s: BACKUP -> MASTER "
825 "(preempting a slower master)\n",
616 "(preempting a slower master)\n",
826 SC2IFP(sc)->if_xname);
617 sc->sc_vhid,
618 sc->sc_carpdev->if_xname);
827 carp_master_down_locked(sc);
828 break;
829 }
830
831 /*
832 * If the master is going to advertise at such a low frequency
833 * that he's guaranteed to time out, we'd might as well just
834 * treat him as timed out now.
835 */
836 sc_tv.tv_sec = sc->sc_advbase * 3;
837 if (timevalcmp(&sc_tv, &ch_tv, <)) {
619 carp_master_down_locked(sc);
620 break;
621 }
622
623 /*
624 * If the master is going to advertise at such a low frequency
625 * that he's guaranteed to time out, we'd might as well just
626 * treat him as timed out now.
627 */
628 sc_tv.tv_sec = sc->sc_advbase * 3;
629 if (timevalcmp(&sc_tv, &ch_tv, <)) {
838 CARP_LOG("%s: BACKUP -> MASTER "
630 CARP_LOG("VHID %u@%s: BACKUP -> MASTER "
839 "(master timed out)\n",
631 "(master timed out)\n",
840 SC2IFP(sc)->if_xname);
632 sc->sc_vhid,
633 sc->sc_carpdev->if_xname);
841 carp_master_down_locked(sc);
842 break;
843 }
844
845 /*
846 * Otherwise, we reset the counter and wait for the next
847 * advertisement.
848 */
849 carp_setrun(sc, af);
850 break;
851 }
852
634 carp_master_down_locked(sc);
635 break;
636 }
637
638 /*
639 * Otherwise, we reset the counter and wait for the next
640 * advertisement.
641 */
642 carp_setrun(sc, af);
643 break;
644 }
645
853 CARP_UNLOCK(ifp->if_carp);
854
646out:
647 CARP_UNLOCK(sc);
855 m_freem(m);
648 m_freem(m);
856 return;
857}
858
859static int
860carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
861{
862 struct m_tag *mtag;
649}
650
651static int
652carp_prepare_ad(struct mbuf *m, struct carp_softc *sc, struct carp_header *ch)
653{
654 struct m_tag *mtag;
863 struct ifnet *ifp = SC2IFP(sc);
864
865 if (sc->sc_init_counter) {
866 /* this could also be seconds since unix epoch */
867 sc->sc_counter = arc4random();
868 sc->sc_counter = sc->sc_counter << 32;
869 sc->sc_counter += arc4random();
870 } else
871 sc->sc_counter++;
872
873 ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
874 ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
875
876 carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
877
878 /* Tag packet for carp_output */
655
656 if (sc->sc_init_counter) {
657 /* this could also be seconds since unix epoch */
658 sc->sc_counter = arc4random();
659 sc->sc_counter = sc->sc_counter << 32;
660 sc->sc_counter += arc4random();
661 } else
662 sc->sc_counter++;
663
664 ch->carp_counter[0] = htonl((sc->sc_counter>>32)&0xffffffff);
665 ch->carp_counter[1] = htonl(sc->sc_counter&0xffffffff);
666
667 carp_hmac_generate(sc, ch->carp_counter, ch->carp_md);
668
669 /* Tag packet for carp_output */
879 mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct ifnet *), M_NOWAIT);
880 if (mtag == NULL) {
670 if ((mtag = m_tag_get(PACKET_TAG_CARP, sizeof(struct carp_softc *),
671 M_NOWAIT)) == NULL) {
881 m_freem(m);
672 m_freem(m);
882 SC2IFP(sc)->if_oerrors++;
673 CARPSTATS_INC(carps_onomem);
883 return (ENOMEM);
884 }
674 return (ENOMEM);
675 }
885 bcopy(&ifp, (caddr_t)(mtag + 1), sizeof(struct ifnet *));
676 bcopy(&sc, (caddr_t)(mtag + 1), sizeof(struct carp_softc *));
886 m_tag_prepend(m, mtag);
887
888 return (0);
889}
890
891static void
677 m_tag_prepend(m, mtag);
678
679 return (0);
680}
681
682static void
892carp_send_ad_all(void)
683carp_send_ad_all(struct carp_softc *badsc)
893{
894 struct carp_softc *sc;
895
684{
685 struct carp_softc *sc;
686
687 /*
688 * Avoid LOR and recursive call to carp_send_ad_locked().
689 */
690 CARP_UNLOCK(badsc);
691
896 mtx_lock(&carp_mtx);
692 mtx_lock(&carp_mtx);
897 LIST_FOREACH(sc, &carpif_list, sc_next) {
898 if (sc->sc_carpdev == NULL)
899 continue;
900 CARP_SCLOCK(sc);
901 if ((SC2IFP(sc)->if_flags & IFF_UP) &&
902 (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING) &&
903 sc->sc_state == MASTER)
693 LIST_FOREACH(sc, &carp_list, sc_next)
694 if (sc != badsc && sc->sc_state == MASTER) {
695 CARP_LOCK(sc);
904 carp_send_ad_locked(sc);
696 carp_send_ad_locked(sc);
905 CARP_SCUNLOCK(sc);
906 }
697 CARP_UNLOCK(sc);
698 }
907 mtx_unlock(&carp_mtx);
699 mtx_unlock(&carp_mtx);
700
701 CARP_LOCK(badsc);
908}
909
910static void
911carp_send_ad(void *v)
912{
913 struct carp_softc *sc = v;
914
702}
703
704static void
705carp_send_ad(void *v)
706{
707 struct carp_softc *sc = v;
708
915 CARP_SCLOCK(sc);
709 CARP_LOCK_ASSERT(sc);
916 carp_send_ad_locked(sc);
710 carp_send_ad_locked(sc);
917 CARP_SCUNLOCK(sc);
711 CARP_UNLOCK(sc);
918}
919
920static void
921carp_send_ad_locked(struct carp_softc *sc)
922{
923 struct carp_header ch;
924 struct timeval tv;
712}
713
714static void
715carp_send_ad_locked(struct carp_softc *sc)
716{
717 struct carp_header ch;
718 struct timeval tv;
719 struct sockaddr sa;
720 struct ifaddr *ifa;
925 struct carp_header *ch_ptr;
926 struct mbuf *m;
721 struct carp_header *ch_ptr;
722 struct mbuf *m;
927 int len, advbase, advskew;
723 int len, advskew;
928
724
929 CARP_SCLOCK_ASSERT(sc);
725 CARP_LOCK_ASSERT(sc);
930
726
931 /* bow out if we've lost our UPness or RUNNINGuiness */
932 if (!((SC2IFP(sc)->if_flags & IFF_UP) &&
933 (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING))) {
934 advbase = 255;
935 advskew = 255;
936 } else {
937 advbase = sc->sc_advbase;
938 if (!carp_suppress_preempt || sc->sc_advskew > 240)
939 advskew = sc->sc_advskew;
940 else
941 advskew = 240;
942 tv.tv_sec = advbase;
943 tv.tv_usec = advskew * 1000000 / 256;
944 }
727 if (!carp_suppress_preempt || sc->sc_advskew > 240)
728 advskew = sc->sc_advskew;
729 else
730 advskew = 240;
731 tv.tv_sec = sc->sc_advbase;
732 tv.tv_usec = advskew * 1000000 / 256;
945
946 ch.carp_version = CARP_VERSION;
947 ch.carp_type = CARP_ADVERTISEMENT;
948 ch.carp_vhid = sc->sc_vhid;
733
734 ch.carp_version = CARP_VERSION;
735 ch.carp_type = CARP_ADVERTISEMENT;
736 ch.carp_vhid = sc->sc_vhid;
949 ch.carp_advbase = advbase;
737 ch.carp_advbase = sc->sc_advbase;
950 ch.carp_advskew = advskew;
951 ch.carp_authlen = 7; /* XXX DEFINE */
952 ch.carp_pad1 = 0; /* must be zero */
953 ch.carp_cksum = 0;
954
738 ch.carp_advskew = advskew;
739 ch.carp_authlen = 7; /* XXX DEFINE */
740 ch.carp_pad1 = 0; /* must be zero */
741 ch.carp_cksum = 0;
742
743 /* XXXGL: OpenBSD picks first ifaddr with needed family. */
744
955#ifdef INET
745#ifdef INET
956 if (sc->sc_ia) {
746 if (sc->sc_naddrs) {
957 struct ip *ip;
958
747 struct ip *ip;
748
959 MGETHDR(m, M_DONTWAIT, MT_HEADER);
749 MGETHDR(m, M_NOWAIT, MT_HEADER);
960 if (m == NULL) {
750 if (m == NULL) {
961 SC2IFP(sc)->if_oerrors++;
962 CARPSTATS_INC(carps_onomem);
963 /* XXX maybe less ? */
751 CARPSTATS_INC(carps_onomem);
752 /* XXX maybe less ? */
964 if (advbase != 255 || advskew != 255)
965 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
966 carp_send_ad, sc);
753 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
754 carp_send_ad, sc);
967 return;
968 }
969 len = sizeof(*ip) + sizeof(ch);
970 m->m_pkthdr.len = len;
971 m->m_pkthdr.rcvif = NULL;
972 m->m_len = len;
973 MH_ALIGN(m, m->m_len);
974 m->m_flags |= M_MCAST;
975 ip = mtod(m, struct ip *);
976 ip->ip_v = IPVERSION;
977 ip->ip_hl = sizeof(*ip) >> 2;
978 ip->ip_tos = IPTOS_LOWDELAY;
979 ip->ip_len = len;
980 ip->ip_id = ip_newid();
981 ip->ip_off = IP_DF;
982 ip->ip_ttl = CARP_DFLTTL;
983 ip->ip_p = IPPROTO_CARP;
984 ip->ip_sum = 0;
755 return;
756 }
757 len = sizeof(*ip) + sizeof(ch);
758 m->m_pkthdr.len = len;
759 m->m_pkthdr.rcvif = NULL;
760 m->m_len = len;
761 MH_ALIGN(m, m->m_len);
762 m->m_flags |= M_MCAST;
763 ip = mtod(m, struct ip *);
764 ip->ip_v = IPVERSION;
765 ip->ip_hl = sizeof(*ip) >> 2;
766 ip->ip_tos = IPTOS_LOWDELAY;
767 ip->ip_len = len;
768 ip->ip_id = ip_newid();
769 ip->ip_off = IP_DF;
770 ip->ip_ttl = CARP_DFLTTL;
771 ip->ip_p = IPPROTO_CARP;
772 ip->ip_sum = 0;
985 ip->ip_src.s_addr = sc->sc_ia->ia_addr.sin_addr.s_addr;
773
774 bzero(&sa, sizeof(sa));
775 sa.sa_family = AF_INET;
776 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
777 if (ifa != NULL) {
778 ip->ip_src.s_addr =
779 ifatoia(ifa)->ia_addr.sin_addr.s_addr;
780 ifa_free(ifa);
781 } else
782 ip->ip_src.s_addr = 0;
986 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
987
988 ch_ptr = (struct carp_header *)(&ip[1]);
989 bcopy(&ch, ch_ptr, sizeof(ch));
990 if (carp_prepare_ad(m, sc, ch_ptr))
991 return;
992
993 m->m_data += sizeof(*ip);
994 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip));
995 m->m_data -= sizeof(*ip);
996
783 ip->ip_dst.s_addr = htonl(INADDR_CARP_GROUP);
784
785 ch_ptr = (struct carp_header *)(&ip[1]);
786 bcopy(&ch, ch_ptr, sizeof(ch));
787 if (carp_prepare_ad(m, sc, ch_ptr))
788 return;
789
790 m->m_data += sizeof(*ip);
791 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip));
792 m->m_data -= sizeof(*ip);
793
997 getmicrotime(&SC2IFP(sc)->if_lastchange);
998 SC2IFP(sc)->if_opackets++;
999 SC2IFP(sc)->if_obytes += len;
1000 CARPSTATS_INC(carps_opackets);
1001
794 CARPSTATS_INC(carps_opackets);
795
1002 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT, &sc->sc_imo, NULL)) {
1003 SC2IFP(sc)->if_oerrors++;
796 if (ip_output(m, NULL, NULL, IP_RAWOUTPUT,
797 &sc->sc_carpdev->if_carp->cif_imo, NULL)) {
1004 if (sc->sc_sendad_errors < INT_MAX)
1005 sc->sc_sendad_errors++;
1006 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
1007 carp_suppress_preempt++;
798 if (sc->sc_sendad_errors < INT_MAX)
799 sc->sc_sendad_errors++;
800 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
801 carp_suppress_preempt++;
1008 if (carp_suppress_preempt == 1) {
1009 CARP_SCUNLOCK(sc);
1010 carp_send_ad_all();
1011 CARP_SCLOCK(sc);
1012 }
802 if (carp_suppress_preempt == 1)
803 carp_send_ad_all(sc);
1013 }
1014 sc->sc_sendad_success = 0;
1015 } else {
1016 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
1017 if (++sc->sc_sendad_success >=
1018 CARP_SENDAD_MIN_SUCCESS) {
1019 carp_suppress_preempt--;
1020 sc->sc_sendad_errors = 0;
1021 }
1022 } else
1023 sc->sc_sendad_errors = 0;
1024 }
1025 }
1026#endif /* INET */
1027#ifdef INET6
804 }
805 sc->sc_sendad_success = 0;
806 } else {
807 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
808 if (++sc->sc_sendad_success >=
809 CARP_SENDAD_MIN_SUCCESS) {
810 carp_suppress_preempt--;
811 sc->sc_sendad_errors = 0;
812 }
813 } else
814 sc->sc_sendad_errors = 0;
815 }
816 }
817#endif /* INET */
818#ifdef INET6
1028 if (sc->sc_ia6) {
819 if (sc->sc_naddrs6) {
1029 struct ip6_hdr *ip6;
1030
820 struct ip6_hdr *ip6;
821
1031 MGETHDR(m, M_DONTWAIT, MT_HEADER);
822 MGETHDR(m, M_NOWAIT, MT_HEADER);
1032 if (m == NULL) {
823 if (m == NULL) {
1033 SC2IFP(sc)->if_oerrors++;
1034 CARPSTATS_INC(carps_onomem);
1035 /* XXX maybe less ? */
824 CARPSTATS_INC(carps_onomem);
825 /* XXX maybe less ? */
1036 if (advbase != 255 || advskew != 255)
1037 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1038 carp_send_ad, sc);
826 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
827 carp_send_ad, sc);
1039 return;
1040 }
1041 len = sizeof(*ip6) + sizeof(ch);
1042 m->m_pkthdr.len = len;
1043 m->m_pkthdr.rcvif = NULL;
1044 m->m_len = len;
1045 MH_ALIGN(m, m->m_len);
1046 m->m_flags |= M_MCAST;
1047 ip6 = mtod(m, struct ip6_hdr *);
1048 bzero(ip6, sizeof(*ip6));
1049 ip6->ip6_vfc |= IPV6_VERSION;
1050 ip6->ip6_hlim = CARP_DFLTTL;
1051 ip6->ip6_nxt = IPPROTO_CARP;
828 return;
829 }
830 len = sizeof(*ip6) + sizeof(ch);
831 m->m_pkthdr.len = len;
832 m->m_pkthdr.rcvif = NULL;
833 m->m_len = len;
834 MH_ALIGN(m, m->m_len);
835 m->m_flags |= M_MCAST;
836 ip6 = mtod(m, struct ip6_hdr *);
837 bzero(ip6, sizeof(*ip6));
838 ip6->ip6_vfc |= IPV6_VERSION;
839 ip6->ip6_hlim = CARP_DFLTTL;
840 ip6->ip6_nxt = IPPROTO_CARP;
1052 bcopy(&sc->sc_ia6->ia_addr.sin6_addr, &ip6->ip6_src,
1053 sizeof(struct in6_addr));
1054 /* set the multicast destination */
841 bzero(&sa, sizeof(sa));
1055
842
843 /* set the source address */
844 sa.sa_family = AF_INET6;
845 ifa = ifaof_ifpforaddr(&sa, sc->sc_carpdev);
846 if (ifa != NULL) {
847 bcopy(IFA_IN6(ifa), &ip6->ip6_src,
848 sizeof(struct in6_addr));
849 ifa_free(ifa);
850 } else
851 /* This should never happen with IPv6. */
852 bzero(&ip6->ip6_src, sizeof(struct in6_addr));
853
854 /* Set the multicast destination. */
1056 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
1057 ip6->ip6_dst.s6_addr8[15] = 0x12;
1058 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
855 ip6->ip6_dst.s6_addr16[0] = htons(0xff02);
856 ip6->ip6_dst.s6_addr8[15] = 0x12;
857 if (in6_setscope(&ip6->ip6_dst, sc->sc_carpdev, NULL) != 0) {
1059 SC2IFP(sc)->if_oerrors++;
1060 m_freem(m);
1061 CARP_DEBUG("%s: in6_setscope failed\n", __func__);
1062 return;
1063 }
1064
1065 ch_ptr = (struct carp_header *)(&ip6[1]);
1066 bcopy(&ch, ch_ptr, sizeof(ch));
1067 if (carp_prepare_ad(m, sc, ch_ptr))
1068 return;
1069
1070 m->m_data += sizeof(*ip6);
1071 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6));
1072 m->m_data -= sizeof(*ip6);
1073
858 m_freem(m);
859 CARP_DEBUG("%s: in6_setscope failed\n", __func__);
860 return;
861 }
862
863 ch_ptr = (struct carp_header *)(&ip6[1]);
864 bcopy(&ch, ch_ptr, sizeof(ch));
865 if (carp_prepare_ad(m, sc, ch_ptr))
866 return;
867
868 m->m_data += sizeof(*ip6);
869 ch_ptr->carp_cksum = carp_cksum(m, len - sizeof(*ip6));
870 m->m_data -= sizeof(*ip6);
871
1074 getmicrotime(&SC2IFP(sc)->if_lastchange);
1075 SC2IFP(sc)->if_opackets++;
1076 SC2IFP(sc)->if_obytes += len;
1077 CARPSTATS_INC(carps_opackets6);
1078
872 CARPSTATS_INC(carps_opackets6);
873
1079 if (ip6_output(m, NULL, NULL, 0, &sc->sc_im6o, NULL, NULL)) {
1080 SC2IFP(sc)->if_oerrors++;
874 if (ip6_output(m, NULL, NULL, 0,
875 &sc->sc_carpdev->if_carp->cif_im6o, NULL, NULL)) {
1081 if (sc->sc_sendad_errors < INT_MAX)
1082 sc->sc_sendad_errors++;
1083 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
1084 carp_suppress_preempt++;
876 if (sc->sc_sendad_errors < INT_MAX)
877 sc->sc_sendad_errors++;
878 if (sc->sc_sendad_errors == CARP_SENDAD_MAX_ERRORS) {
879 carp_suppress_preempt++;
1085 if (carp_suppress_preempt == 1) {
1086 CARP_SCUNLOCK(sc);
1087 carp_send_ad_all();
1088 CARP_SCLOCK(sc);
1089 }
880 if (carp_suppress_preempt == 1)
881 carp_send_ad_all(sc);
1090 }
1091 sc->sc_sendad_success = 0;
1092 } else {
1093 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
1094 if (++sc->sc_sendad_success >=
1095 CARP_SENDAD_MIN_SUCCESS) {
1096 carp_suppress_preempt--;
1097 sc->sc_sendad_errors = 0;
1098 }
1099 } else
1100 sc->sc_sendad_errors = 0;
1101 }
1102 }
1103#endif /* INET6 */
1104
882 }
883 sc->sc_sendad_success = 0;
884 } else {
885 if (sc->sc_sendad_errors >= CARP_SENDAD_MAX_ERRORS) {
886 if (++sc->sc_sendad_success >=
887 CARP_SENDAD_MIN_SUCCESS) {
888 carp_suppress_preempt--;
889 sc->sc_sendad_errors = 0;
890 }
891 } else
892 sc->sc_sendad_errors = 0;
893 }
894 }
895#endif /* INET6 */
896
1105 if (advbase != 255 || advskew != 255)
1106 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1107 carp_send_ad, sc);
897 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv), carp_send_ad, sc);
898}
1108
899
900static void
901carp_addroute(struct carp_softc *sc)
902{
903 struct ifaddr *ifa;
904
905 CARP_FOREACH_IFA(sc, ifa)
906 switch (ifa->ifa_addr->sa_family) {
907#ifdef INET
908 case AF_INET:
909 in_addprefix(ifatoia(ifa), RTF_UP);
910 ifa_add_loopback_route(ifa,
911 (struct sockaddr *)&ifatoia(ifa)->ia_addr);
912 break;
913#endif
914#ifdef INET6
915 case AF_INET6:
916 ifa_add_loopback_route(ifa,
917 (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
918 in6_ifaddloop(ifa);
919 break;
920#endif
921 }
1109}
1110
922}
923
924static void
925carp_delroute(struct carp_softc *sc)
926{
927 struct ifaddr *ifa;
928
929 CARP_FOREACH_IFA(sc, ifa)
930 switch (ifa->ifa_addr->sa_family) {
1111#ifdef INET
931#ifdef INET
932 case AF_INET:
933 ifa_del_loopback_route(ifa,
934 (struct sockaddr *)&ifatoia(ifa)->ia_addr);
935 in_scrubprefix(ifatoia(ifa), LLE_STATIC);
936 break;
937#endif
938#ifdef INET6
939 case AF_INET6:
940 ifa_del_loopback_route(ifa,
941 (struct sockaddr *)&ifatoia6(ifa)->ia_addr);
942 in6_ifremloop(ifa);
943 break;
944#endif
945 }
946}
947
948#ifdef INET
1112/*
1113 * Broadcast a gratuitous ARP request containing
1114 * the virtual router MAC address for each IP address
1115 * associated with the virtual router.
1116 */
1117static void
1118carp_send_arp(struct carp_softc *sc)
1119{
1120 struct ifaddr *ifa;
1121
949/*
950 * Broadcast a gratuitous ARP request containing
951 * the virtual router MAC address for each IP address
952 * associated with the virtual router.
953 */
954static void
955carp_send_arp(struct carp_softc *sc)
956{
957 struct ifaddr *ifa;
958
1122 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
959 CARP_FOREACH_IFA(sc, ifa)
960 if (ifa->ifa_addr->sa_family == AF_INET)
961 arp_ifinit2(sc->sc_carpdev, ifa, LLADDR(&sc->sc_addr));
962}
1123
963
1124 if (ifa->ifa_addr->sa_family != AF_INET)
1125 continue;
964int
965carp_iamatch(struct ifaddr *ifa, uint8_t **enaddr)
966{
967 struct carp_softc *sc = ifa->ifa_carp;
1126
968
1127/* arprequest(sc->sc_carpdev, &in, &in, IF_LLADDR(sc->sc_ifp)); */
1128 arp_ifinit2(sc->sc_carpdev, ifa, IF_LLADDR(sc->sc_ifp));
1129
1130 DELAY(1000); /* XXX */
969 if (sc->sc_state == MASTER) {
970 *enaddr = LLADDR(&sc->sc_addr);
971 return (1);
1131 }
972 }
973
974 return (0);
1132}
1133#endif
1134
1135#ifdef INET6
1136static void
1137carp_send_na(struct carp_softc *sc)
1138{
975}
976#endif
977
978#ifdef INET6
979static void
980carp_send_na(struct carp_softc *sc)
981{
982 static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1139 struct ifaddr *ifa;
1140 struct in6_addr *in6;
983 struct ifaddr *ifa;
984 struct in6_addr *in6;
1141 static struct in6_addr mcast = IN6ADDR_LINKLOCAL_ALLNODES_INIT;
1142
985
1143 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1144
986 CARP_FOREACH_IFA(sc, ifa) {
1145 if (ifa->ifa_addr->sa_family != AF_INET6)
1146 continue;
1147
987 if (ifa->ifa_addr->sa_family != AF_INET6)
988 continue;
989
1148 in6 = &ifatoia6(ifa)->ia_addr.sin6_addr;
990 in6 = IFA_IN6(ifa);
1149 nd6_na_output(sc->sc_carpdev, &mcast, in6,
1150 ND_NA_FLAG_OVERRIDE, 1, NULL);
1151 DELAY(1000); /* XXX */
1152 }
1153}
991 nd6_na_output(sc->sc_carpdev, &mcast, in6,
992 ND_NA_FLAG_OVERRIDE, 1, NULL);
993 DELAY(1000); /* XXX */
994 }
995}
1154#endif /* INET6 */
1155
996
1156#ifdef INET
1157static int
1158carp_addrcount(struct carp_if *cif, struct in_ifaddr *ia, int type)
1159{
1160 struct carp_softc *vh;
1161 struct ifaddr *ifa;
1162 int count = 0;
1163
1164 CARP_LOCK_ASSERT(cif);
1165
1166 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1167 if ((type == CARP_COUNT_RUNNING &&
1168 (SC2IFP(vh)->if_flags & IFF_UP) &&
1169 (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING)) ||
1170 (type == CARP_COUNT_MASTER && vh->sc_state == MASTER)) {
1171 IF_ADDR_LOCK(SC2IFP(vh));
1172 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1173 ifa_list) {
1174 if (ifa->ifa_addr->sa_family == AF_INET &&
1175 ia->ia_addr.sin_addr.s_addr ==
1176 ifatoia(ifa)->ia_addr.sin_addr.s_addr)
1177 count++;
1178 }
1179 IF_ADDR_UNLOCK(SC2IFP(vh));
1180 }
1181 }
1182 return (count);
1183}
1184
1185int
1186carp_iamatch(struct ifnet *ifp, struct in_ifaddr *ia,
1187 struct in_addr *isaddr, u_int8_t **enaddr)
1188{
1189 struct carp_if *cif;
1190 struct carp_softc *vh;
1191 int index, count = 0;
1192 struct ifaddr *ifa;
1193
1194 cif = ifp->if_carp;
1195 CARP_LOCK(cif);
1196
1197 if (carp_opts[CARPCTL_ARPBALANCE]) {
1198 /*
1199 * XXX proof of concept implementation.
1200 * We use the source ip to decide which virtual host should
1201 * handle the request. If we're master of that virtual host,
1202 * then we respond, otherwise, just drop the arp packet on
1203 * the floor.
1204 */
1205 count = carp_addrcount(cif, ia, CARP_COUNT_RUNNING);
1206 if (count == 0) {
1207 /* should never reach this */
1208 CARP_UNLOCK(cif);
1209 return (0);
1210 }
1211
1212 /* this should be a hash, like pf_hash() */
1213 index = ntohl(isaddr->s_addr) % count;
1214 count = 0;
1215
1216 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1217 if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1218 (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING)) {
1219 IF_ADDR_LOCK(SC2IFP(vh));
1220 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist,
1221 ifa_list) {
1222 if (ifa->ifa_addr->sa_family ==
1223 AF_INET &&
1224 ia->ia_addr.sin_addr.s_addr ==
1225 ifatoia(ifa)->ia_addr.sin_addr.s_addr) {
1226 if (count == index) {
1227 if (vh->sc_state ==
1228 MASTER) {
1229 *enaddr = IF_LLADDR(vh->sc_ifp);
1230 IF_ADDR_UNLOCK(SC2IFP(vh));
1231 CARP_UNLOCK(cif);
1232 return (1);
1233 } else {
1234 IF_ADDR_UNLOCK(SC2IFP(vh));
1235 CARP_UNLOCK(cif);
1236 return (0);
1237 }
1238 }
1239 count++;
1240 }
1241 }
1242 IF_ADDR_UNLOCK(SC2IFP(vh));
1243 }
1244 }
1245 } else {
1246 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1247 if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1248 (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1249 ia->ia_ifp == SC2IFP(vh) &&
1250 vh->sc_state == MASTER) {
1251 *enaddr = IF_LLADDR(vh->sc_ifp);
1252 CARP_UNLOCK(cif);
1253 return (1);
1254 }
1255 }
1256 }
1257 CARP_UNLOCK(cif);
1258 return (0);
1259}
1260#endif
1261
1262#ifdef INET6
1263struct ifaddr *
1264carp_iamatch6(struct ifnet *ifp, struct in6_addr *taddr)
1265{
997struct ifaddr *
998carp_iamatch6(struct ifnet *ifp, struct in6_addr *taddr)
999{
1266 struct carp_if *cif;
1267 struct carp_softc *vh;
1268 struct ifaddr *ifa;
1269
1000 struct ifaddr *ifa;
1001
1270 cif = ifp->if_carp;
1271 CARP_LOCK(cif);
1272 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list) {
1273 IF_ADDR_LOCK(SC2IFP(vh));
1274 TAILQ_FOREACH(ifa, &SC2IFP(vh)->if_addrlist, ifa_list) {
1275 if (IN6_ARE_ADDR_EQUAL(taddr,
1276 &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1277 (SC2IFP(vh)->if_flags & IFF_UP) &&
1278 (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1279 vh->sc_state == MASTER) {
1280 ifa_ref(ifa);
1281 IF_ADDR_UNLOCK(SC2IFP(vh));
1282 CARP_UNLOCK(cif);
1283 return (ifa);
1284 }
1002 IF_ADDR_LOCK(ifp);
1003 IFNET_FOREACH_IFA(ifp, ifa)
1004 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1005 ifa->ifa_carp->sc_state == MASTER &&
1006 IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1007 ifa_ref(ifa);
1008 IF_ADDR_UNLOCK(ifp);
1009 return (ifa);
1285 }
1010 }
1286 IF_ADDR_UNLOCK(SC2IFP(vh));
1287 }
1288 CARP_UNLOCK(cif);
1289
1011 IF_ADDR_UNLOCK(ifp);
1012
1290 return (NULL);
1291}
1292
1293caddr_t
1294carp_macmatch6(struct ifnet *ifp, struct mbuf *m, const struct in6_addr *taddr)
1295{
1013 return (NULL);
1014}
1015
1016caddr_t
1017carp_macmatch6(struct ifnet *ifp, struct mbuf *m, const struct in6_addr *taddr)
1018{
1296 struct m_tag *mtag;
1297 struct carp_if *cif;
1298 struct carp_softc *sc;
1299 struct ifaddr *ifa;
1300
1019 struct ifaddr *ifa;
1020
1301 cif = ifp->if_carp;
1302 CARP_LOCK(cif);
1303 TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list) {
1304 IF_ADDR_LOCK(SC2IFP(sc));
1305 TAILQ_FOREACH(ifa, &SC2IFP(sc)->if_addrlist, ifa_list) {
1306 if (IN6_ARE_ADDR_EQUAL(taddr,
1307 &ifatoia6(ifa)->ia_addr.sin6_addr) &&
1308 (SC2IFP(sc)->if_flags & IFF_UP) &&
1309 (SC2IFP(sc)->if_drv_flags & IFF_DRV_RUNNING)) {
1310 struct ifnet *ifp = SC2IFP(sc);
1311 mtag = m_tag_get(PACKET_TAG_CARP,
1312 sizeof(struct ifnet *), M_NOWAIT);
1313 if (mtag == NULL) {
1314 /* better a bit than nothing */
1315 IF_ADDR_UNLOCK(SC2IFP(sc));
1316 CARP_UNLOCK(cif);
1317 return (IF_LLADDR(sc->sc_ifp));
1318 }
1319 bcopy(&ifp, (caddr_t)(mtag + 1),
1320 sizeof(struct ifnet *));
1321 m_tag_prepend(m, mtag);
1021 IF_ADDR_LOCK(ifp);
1022 IFNET_FOREACH_IFA(ifp, ifa)
1023 if (ifa->ifa_addr->sa_family == AF_INET6 &&
1024 IN6_ARE_ADDR_EQUAL(taddr, IFA_IN6(ifa))) {
1025 struct carp_softc *sc = ifa->ifa_carp;
1026 struct m_tag *mtag;
1322
1027
1323 IF_ADDR_UNLOCK(SC2IFP(sc));
1324 CARP_UNLOCK(cif);
1325 return (IF_LLADDR(sc->sc_ifp));
1326 }
1028 IF_ADDR_UNLOCK(ifp);
1029
1030 mtag = m_tag_get(PACKET_TAG_CARP,
1031 sizeof(struct ifnet *), M_NOWAIT);
1032 if (mtag == NULL)
1033 /* Better a bit than nothing. */
1034 return (LLADDR(&sc->sc_addr));
1035
1036 bcopy(&ifp, (caddr_t)(mtag + 1),
1037 sizeof(struct ifnet *));
1038 m_tag_prepend(m, mtag);
1039
1040 return (LLADDR(&sc->sc_addr));
1327 }
1041 }
1328 IF_ADDR_UNLOCK(SC2IFP(sc));
1329 }
1330 CARP_UNLOCK(cif);
1042 IF_ADDR_UNLOCK(ifp);
1331
1332 return (NULL);
1333}
1043
1044 return (NULL);
1045}
1334#endif
1046#endif /* INET6 */
1335
1047
1336struct ifnet *
1048int
1337carp_forus(struct ifnet *ifp, u_char *dhost)
1338{
1049carp_forus(struct ifnet *ifp, u_char *dhost)
1050{
1339 struct carp_if *cif;
1340 struct carp_softc *vh;
1341 u_int8_t *ena = dhost;
1051 struct carp_softc *sc;
1052 uint8_t *ena = dhost;
1342
1343 if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1053
1054 if (ena[0] || ena[1] || ena[2] != 0x5e || ena[3] || ena[4] != 1)
1344 return (NULL);
1055 return (0);
1345
1056
1346 cif = ifp->if_carp;
1347 CARP_LOCK(cif);
1348 TAILQ_FOREACH(vh, &cif->vhif_vrs, sc_list)
1349 if ((SC2IFP(vh)->if_flags & IFF_UP) &&
1350 (SC2IFP(vh)->if_drv_flags & IFF_DRV_RUNNING) &&
1351 vh->sc_state == MASTER &&
1352 !bcmp(dhost, IF_LLADDR(vh->sc_ifp), ETHER_ADDR_LEN)) {
1353 CARP_UNLOCK(cif);
1354 return (SC2IFP(vh));
1057 CIF_LOCK(ifp->if_carp);
1058 IFNET_FOREACH_CARP(ifp, sc) {
1059 CARP_LOCK(sc);
1060 if (sc->sc_state == MASTER && !bcmp(dhost, LLADDR(&sc->sc_addr),
1061 ETHER_ADDR_LEN)) {
1062 CARP_UNLOCK(sc);
1063 CIF_UNLOCK(ifp->if_carp);
1064 return (1);
1355 }
1065 }
1066 CARP_UNLOCK(sc);
1067 }
1068 CIF_UNLOCK(ifp->if_carp);
1356
1069
1357 CARP_UNLOCK(cif);
1358 return (NULL);
1070 return (0);
1359}
1360
1361static void
1362carp_master_down(void *v)
1363{
1364 struct carp_softc *sc = v;
1365
1071}
1072
1073static void
1074carp_master_down(void *v)
1075{
1076 struct carp_softc *sc = v;
1077
1366 CARP_SCLOCK(sc);
1367 carp_master_down_locked(sc);
1368 CARP_SCUNLOCK(sc);
1078 CARP_LOCK_ASSERT(sc);
1079
1080 if (sc->sc_state == BACKUP) {
1081 CARP_LOG("VHID %u@%s: BACKUP -> MASTER (preempting)\n",
1082 sc->sc_vhid,
1083 sc->sc_carpdev->if_xname);
1084 carp_master_down_locked(sc);
1085 }
1086
1087 CARP_UNLOCK(sc);
1369}
1370
1371static void
1372carp_master_down_locked(struct carp_softc *sc)
1373{
1088}
1089
1090static void
1091carp_master_down_locked(struct carp_softc *sc)
1092{
1374 if (sc->sc_carpdev)
1375 CARP_SCLOCK_ASSERT(sc);
1376
1093
1094 CARP_LOCK_ASSERT(sc);
1095
1377 switch (sc->sc_state) {
1096 switch (sc->sc_state) {
1378 case INIT:
1379 printf("%s: master_down event in INIT state\n",
1380 SC2IFP(sc)->if_xname);
1381 break;
1382 case MASTER:
1383 break;
1384 case BACKUP:
1385 carp_set_state(sc, MASTER);
1386 carp_send_ad_locked(sc);
1387#ifdef INET
1388 carp_send_arp(sc);
1389#endif
1390#ifdef INET6
1391 carp_send_na(sc);
1097 case BACKUP:
1098 carp_set_state(sc, MASTER);
1099 carp_send_ad_locked(sc);
1100#ifdef INET
1101 carp_send_arp(sc);
1102#endif
1103#ifdef INET6
1104 carp_send_na(sc);
1392#endif /* INET6 */
1105#endif
1393 carp_setrun(sc, 0);
1106 carp_setrun(sc, 0);
1394 carp_setroute(sc, RTM_ADD);
1107 carp_addroute(sc);
1395 break;
1108 break;
1109 case INIT:
1110 case MASTER:
1111#ifdef INVARIANTS
1112 panic("carp: VHID %u@%s: master_down event in %s state\n",
1113 sc->sc_vhid,
1114 sc->sc_carpdev->if_xname,
1115 sc->sc_state ? "MASTER" : "INIT");
1116#endif
1117 break;
1396 }
1397}
1398
1399/*
1400 * When in backup state, af indicates whether to reset the master down timer
1401 * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1402 */
1403static void
1404carp_setrun(struct carp_softc *sc, sa_family_t af)
1405{
1406 struct timeval tv;
1407
1118 }
1119}
1120
1121/*
1122 * When in backup state, af indicates whether to reset the master down timer
1123 * for v4 or v6. If it's set to zero, reset the ones which are already pending.
1124 */
1125static void
1126carp_setrun(struct carp_softc *sc, sa_family_t af)
1127{
1128 struct timeval tv;
1129
1408 if (sc->sc_carpdev == NULL) {
1409 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1410 carp_set_state(sc, INIT);
1411 return;
1412 } else
1413 CARP_SCLOCK_ASSERT(sc);
1130 CARP_LOCK_ASSERT(sc);
1414
1131
1415 if (SC2IFP(sc)->if_flags & IFF_UP &&
1416 sc->sc_vhid > 0 && (sc->sc_naddrs || sc->sc_naddrs6) &&
1417 sc->sc_carpdev->if_link_state == LINK_STATE_UP)
1418 SC2IFP(sc)->if_drv_flags |= IFF_DRV_RUNNING;
1419 else {
1420 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1421 carp_setroute(sc, RTM_DELETE);
1132 if ((sc->sc_carpdev->if_flags & IFF_UP) == 0 ||
1133 sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1134 (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0))
1422 return;
1135 return;
1423 }
1424
1425 switch (sc->sc_state) {
1426 case INIT:
1136
1137 switch (sc->sc_state) {
1138 case INIT:
1427 CARP_LOG("%s: INIT -> BACKUP\n", SC2IFP(sc)->if_xname);
1139 CARP_LOG("VHID %u@%s: INIT -> BACKUP\n",
1140 sc->sc_vhid,
1141 sc->sc_carpdev->if_xname);
1428 carp_set_state(sc, BACKUP);
1142 carp_set_state(sc, BACKUP);
1429 carp_setroute(sc, RTM_DELETE);
1430 carp_setrun(sc, 0);
1431 break;
1432 case BACKUP:
1433 callout_stop(&sc->sc_ad_tmo);
1434 tv.tv_sec = 3 * sc->sc_advbase;
1435 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1436 switch (af) {
1437#ifdef INET
1438 case AF_INET:
1439 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1440 carp_master_down, sc);
1441 break;
1143 carp_setrun(sc, 0);
1144 break;
1145 case BACKUP:
1146 callout_stop(&sc->sc_ad_tmo);
1147 tv.tv_sec = 3 * sc->sc_advbase;
1148 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1149 switch (af) {
1150#ifdef INET
1151 case AF_INET:
1152 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1153 carp_master_down, sc);
1154 break;
1442#endif /* INET */
1155#endif
1443#ifdef INET6
1444 case AF_INET6:
1445 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1446 carp_master_down, sc);
1447 break;
1156#ifdef INET6
1157 case AF_INET6:
1158 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1159 carp_master_down, sc);
1160 break;
1448#endif /* INET6 */
1161#endif
1449 default:
1162 default:
1163#ifdef INET
1450 if (sc->sc_naddrs)
1451 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1452 carp_master_down, sc);
1164 if (sc->sc_naddrs)
1165 callout_reset(&sc->sc_md_tmo, tvtohz(&tv),
1166 carp_master_down, sc);
1167#endif
1168#ifdef INET6
1453 if (sc->sc_naddrs6)
1454 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1455 carp_master_down, sc);
1169 if (sc->sc_naddrs6)
1170 callout_reset(&sc->sc_md6_tmo, tvtohz(&tv),
1171 carp_master_down, sc);
1172#endif
1456 break;
1457 }
1458 break;
1459 case MASTER:
1460 tv.tv_sec = sc->sc_advbase;
1461 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1462 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1463 carp_send_ad, sc);
1464 break;
1465 }
1466}
1467
1173 break;
1174 }
1175 break;
1176 case MASTER:
1177 tv.tv_sec = sc->sc_advbase;
1178 tv.tv_usec = sc->sc_advskew * 1000000 / 256;
1179 callout_reset(&sc->sc_ad_tmo, tvtohz(&tv),
1180 carp_send_ad, sc);
1181 break;
1182 }
1183}
1184
1468#ifdef INET
1469static void
1470carp_multicast_cleanup(struct carp_softc *sc, int dofree)
1185/*
1186 * Setup multicast structures.
1187 */
1188static int
1189carp_multicast_setup(struct carp_softc *sc, sa_family_t sa)
1471{
1190{
1472 struct ip_moptions *imo = &sc->sc_imo;
1473 u_int16_t n = imo->imo_num_memberships;
1191 struct ifnet *ifp = sc->sc_carpdev;
1192 struct carp_if *cif = ifp->if_carp;
1193 int error = 0;
1474
1194
1475 /* Clean up our own multicast memberships */
1476 while (n-- > 0) {
1477 if (imo->imo_membership[n] != NULL) {
1478 if (dofree)
1479 in_delmulti(imo->imo_membership[n]);
1480 imo->imo_membership[n] = NULL;
1195 switch (sa) {
1196#ifdef INET
1197 case AF_INET:
1198 {
1199 struct ip_moptions *imo = &cif->cif_imo;
1200 struct in_addr addr;
1201
1202 if (imo->imo_membership)
1203 return (0);
1204
1205 imo->imo_membership = (struct in_multi **)malloc(
1206 (sizeof(struct in_multi *) * IP_MIN_MEMBERSHIPS), M_CARP,
1207 M_WAITOK);
1208 imo->imo_mfilters = NULL;
1209 imo->imo_max_memberships = IP_MIN_MEMBERSHIPS;
1210 imo->imo_multicast_vif = -1;
1211
1212 addr.s_addr = htonl(INADDR_CARP_GROUP);
1213 if ((error = in_joingroup(ifp, &addr, NULL,
1214 &imo->imo_membership[0])) != 0) {
1215 free(imo->imo_membership, M_CARP);
1216 break;
1481 }
1217 }
1218 imo->imo_num_memberships++;
1219 imo->imo_multicast_ifp = ifp;
1220 imo->imo_multicast_ttl = CARP_DFLTTL;
1221 imo->imo_multicast_loop = 0;
1222 break;
1223 }
1224#endif
1225#ifdef INET6
1226 case AF_INET6:
1227 {
1228 struct ip6_moptions *im6o = &cif->cif_im6o;
1229 struct in6_addr in6;
1230 struct in6_multi *in6m;
1231
1232 if (im6o->im6o_membership)
1233 return (0);
1234
1235 im6o->im6o_membership = (struct in6_multi **)malloc(
1236 (sizeof(struct in6_multi *) * IPV6_MIN_MEMBERSHIPS), M_CARP,
1237 M_ZERO|M_WAITOK);
1238 im6o->im6o_mfilters = NULL;
1239 im6o->im6o_max_memberships = IPV6_MIN_MEMBERSHIPS;
1240 im6o->im6o_multicast_hlim = CARP_DFLTTL;
1241 im6o->im6o_multicast_ifp = ifp;
1242
1243 /* Join IPv6 CARP multicast group. */
1244 bzero(&in6, sizeof(in6));
1245 in6.s6_addr16[0] = htons(0xff02);
1246 in6.s6_addr8[15] = 0x12;
1247 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1248 free(im6o->im6o_membership, M_CARP);
1249 break;
1250 }
1251 in6m = NULL;
1252 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1253 free(im6o->im6o_membership, M_CARP);
1254 break;
1255 }
1256 im6o->im6o_membership[0] = in6m;
1257 im6o->im6o_num_memberships++;
1258
1259 /* Join solicited multicast address. */
1260 bzero(&in6, sizeof(in6));
1261 in6.s6_addr16[0] = htons(0xff02);
1262 in6.s6_addr32[1] = 0;
1263 in6.s6_addr32[2] = htonl(1);
1264 in6.s6_addr32[3] = 0;
1265 in6.s6_addr8[12] = 0xff;
1266 if ((error = in6_setscope(&in6, ifp, NULL)) != 0) {
1267 in6_mc_leave(im6o->im6o_membership[0], NULL);
1268 free(im6o->im6o_membership, M_CARP);
1269 break;
1270 }
1271 in6m = NULL;
1272 if ((error = in6_mc_join(ifp, &in6, NULL, &in6m, 0)) != 0) {
1273 in6_mc_leave(im6o->im6o_membership[0], NULL);
1274 free(im6o->im6o_membership, M_CARP);
1275 break;
1276 }
1277 im6o->im6o_membership[1] = in6m;
1278 im6o->im6o_num_memberships++;
1279 break;
1280 }
1281#endif
1482 }
1282 }
1483 KASSERT(imo->imo_mfilters == NULL,
1484 ("%s: imo_mfilters != NULL", __func__));
1485 imo->imo_num_memberships = 0;
1486 imo->imo_multicast_ifp = NULL;
1283
1284 return (error);
1487}
1285}
1488#endif
1489
1286
1490#ifdef INET6
1287/*
1288 * Free multicast structures.
1289 */
1491static void
1290static void
1492carp_multicast6_cleanup(struct carp_softc *sc, int dofree)
1291carp_multicast_cleanup(struct carp_softc *sc, sa_family_t sa)
1493{
1292{
1494 struct ip6_moptions *im6o = &sc->sc_im6o;
1495 u_int16_t n = im6o->im6o_num_memberships;
1293 struct ifnet *ifp = sc->sc_carpdev;
1294 struct carp_if *cif = ifp->if_carp;
1496
1295
1497 while (n-- > 0) {
1498 if (im6o->im6o_membership[n] != NULL) {
1499 if (dofree)
1500 in6_mc_leave(im6o->im6o_membership[n], NULL);
1501 im6o->im6o_membership[n] = NULL;
1296 switch (sa) {
1297#ifdef INET
1298 case AF_INET:
1299 if (sc->sc_naddrs == 0) {
1300 struct ip_moptions *imo = &cif->cif_imo;
1301
1302 in_leavegroup(imo->imo_membership[0], NULL);
1303 KASSERT(imo->imo_mfilters == NULL,
1304 ("%s: imo_mfilters != NULL", __func__));
1305 free(imo->imo_membership, M_CARP);
1306 imo->imo_membership = NULL;
1307
1502 }
1308 }
1309 break;
1310#endif
1311#ifdef INET6
1312 case AF_INET6:
1313 if (sc->sc_naddrs6 == 0) {
1314 struct ip6_moptions *im6o = &cif->cif_im6o;
1315
1316 in6_mc_leave(im6o->im6o_membership[0], NULL);
1317 in6_mc_leave(im6o->im6o_membership[1], NULL);
1318 KASSERT(im6o->im6o_mfilters == NULL,
1319 ("%s: im6o_mfilters != NULL", __func__));
1320 free(im6o->im6o_membership, M_CARP);
1321 im6o->im6o_membership = NULL;
1322 }
1323 break;
1324#endif
1503 }
1325 }
1504 KASSERT(im6o->im6o_mfilters == NULL,
1505 ("%s: im6o_mfilters != NULL", __func__));
1506 im6o->im6o_num_memberships = 0;
1507 im6o->im6o_multicast_ifp = NULL;
1508}
1326}
1509#endif
1510
1327
1511#ifdef INET
1512static int
1513carp_set_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1328int
1329carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *sa)
1514{
1330{
1515 struct ifnet *ifp;
1516 struct carp_if *cif;
1517 struct in_ifaddr *ia, *ia_if;
1518 struct ip_moptions *imo = &sc->sc_imo;
1519 struct in_addr addr;
1520 u_long iaddr = htonl(sin->sin_addr.s_addr);
1521 int own, error;
1331 struct m_tag *mtag;
1332 struct carp_softc *sc;
1522
1333
1523 if (sin->sin_addr.s_addr == 0) {
1524 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1525 carp_set_state(sc, INIT);
1526 if (sc->sc_naddrs)
1527 SC2IFP(sc)->if_flags |= IFF_UP;
1528 if (sc->sc_carpdev)
1529 CARP_SCLOCK(sc);
1530 carp_setrun(sc, 0);
1531 if (sc->sc_carpdev)
1532 CARP_SCUNLOCK(sc);
1334 if (!sa)
1533 return (0);
1335 return (0);
1534 }
1535
1336
1536 /* we have to do it by hands to check we won't match on us */
1537 ia_if = NULL; own = 0;
1538 IN_IFADDR_RLOCK();
1539 TAILQ_FOREACH(ia, &V_in_ifaddrhead, ia_link) {
1540 /* and, yeah, we need a multicast-capable iface too */
1541 if (ia->ia_ifp != SC2IFP(sc) &&
1542 (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1543 (iaddr & ia->ia_subnetmask) == ia->ia_subnet) {
1544 if (!ia_if)
1545 ia_if = ia;
1546 if (sin->sin_addr.s_addr ==
1547 ia->ia_addr.sin_addr.s_addr)
1548 own++;
1549 }
1337 switch (sa->sa_family) {
1338#ifdef INET
1339 case AF_INET:
1340 break;
1341#endif
1342#ifdef INET6
1343 case AF_INET6:
1344 break;
1345#endif
1346 default:
1347 return (0);
1550 }
1551
1348 }
1349
1552 if (!ia_if) {
1553 IN_IFADDR_RUNLOCK();
1554 return (EADDRNOTAVAIL);
1555 }
1350 mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
1351 if (mtag == NULL)
1352 return (0);
1556
1353
1557 ia = ia_if;
1558 ifa_ref(&ia->ia_ifa);
1559 IN_IFADDR_RUNLOCK();
1354 bcopy(mtag + 1, &sc, sizeof(struct carp_softc *));
1560
1355
1561 ifp = ia->ia_ifp;
1356 /* Set the source MAC address to the Virtual Router MAC Address. */
1357 switch (ifp->if_type) {
1358 case IFT_ETHER:
1359 case IFT_L2VLAN: {
1360 struct ether_header *eh;
1562
1361
1563 if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1564 (imo->imo_multicast_ifp && imo->imo_multicast_ifp != ifp)) {
1565 ifa_free(&ia->ia_ifa);
1566 return (EADDRNOTAVAIL);
1567 }
1568
1569 if (imo->imo_num_memberships == 0) {
1570 addr.s_addr = htonl(INADDR_CARP_GROUP);
1571 if ((imo->imo_membership[0] = in_addmulti(&addr, ifp)) ==
1572 NULL) {
1573 ifa_free(&ia->ia_ifa);
1574 return (ENOBUFS);
1362 eh = mtod(m, struct ether_header *);
1363 eh->ether_shost[0] = 0;
1364 eh->ether_shost[1] = 0;
1365 eh->ether_shost[2] = 0x5e;
1366 eh->ether_shost[3] = 0;
1367 eh->ether_shost[4] = 1;
1368 eh->ether_shost[5] = sc->sc_vhid;
1575 }
1369 }
1576 imo->imo_num_memberships++;
1577 imo->imo_multicast_ifp = ifp;
1578 imo->imo_multicast_ttl = CARP_DFLTTL;
1579 imo->imo_multicast_loop = 0;
1580 }
1370 break;
1371 case IFT_FDDI: {
1372 struct fddi_header *fh;
1581
1373
1582 if (!ifp->if_carp) {
1583
1584 cif = malloc(sizeof(*cif), M_CARP,
1585 M_WAITOK|M_ZERO);
1586 if (!cif) {
1587 error = ENOBUFS;
1588 goto cleanup;
1374 fh = mtod(m, struct fddi_header *);
1375 fh->fddi_shost[0] = 0;
1376 fh->fddi_shost[1] = 0;
1377 fh->fddi_shost[2] = 0x5e;
1378 fh->fddi_shost[3] = 0;
1379 fh->fddi_shost[4] = 1;
1380 fh->fddi_shost[5] = sc->sc_vhid;
1589 }
1381 }
1590 if ((error = ifpromisc(ifp, 1))) {
1591 free(cif, M_CARP);
1592 goto cleanup;
1382 break;
1383 case IFT_ISO88025: {
1384 struct iso88025_header *th;
1385 th = mtod(m, struct iso88025_header *);
1386 th->iso88025_shost[0] = 3;
1387 th->iso88025_shost[1] = 0;
1388 th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
1389 th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
1390 th->iso88025_shost[4] = 0;
1391 th->iso88025_shost[5] = 0;
1593 }
1392 }
1594
1595 CARP_LOCK_INIT(cif);
1596 CARP_LOCK(cif);
1597 cif->vhif_ifp = ifp;
1598 TAILQ_INIT(&cif->vhif_vrs);
1599 ifp->if_carp = cif;
1600
1601 } else {
1602 struct carp_softc *vr;
1603
1604 cif = (struct carp_if *)ifp->if_carp;
1605 CARP_LOCK(cif);
1606 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1607 if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1608 CARP_UNLOCK(cif);
1609 error = EEXIST;
1610 goto cleanup;
1611 }
1393 break;
1394 default:
1395 printf("%s: carp is not supported for the %d interface type\n",
1396 ifp->if_xname, ifp->if_type);
1397 return (EOPNOTSUPP);
1612 }
1398 }
1613 sc->sc_ia = ia;
1614 sc->sc_carpdev = ifp;
1615
1399
1616 { /* XXX prevent endless loop if already in queue */
1617 struct carp_softc *vr, *after = NULL;
1618 int myself = 0;
1619 cif = (struct carp_if *)ifp->if_carp;
1400 return (0);
1401}
1620
1402
1621 /* XXX: cif should not change, right? So we still hold the lock */
1622 CARP_LOCK_ASSERT(cif);
1403static struct carp_softc*
1404carp_alloc(struct ifnet *ifp)
1405{
1406 struct carp_softc *sc;
1407 struct carp_if *cif;
1623
1408
1624 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1625 if (vr == sc)
1626 myself = 1;
1627 if (vr->sc_vhid < sc->sc_vhid)
1628 after = vr;
1409 if ((cif = ifp->if_carp) == NULL) {
1410 cif = carp_alloc_if(ifp);
1411 if (cif == NULL)
1412 return (NULL);
1629 }
1630
1413 }
1414
1631 if (!myself) {
1632 /* We're trying to keep things in order */
1633 if (after == NULL) {
1634 TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1635 } else {
1636 TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1637 }
1638 cif->vhif_nvrs++;
1639 }
1640 }
1415 sc = malloc(sizeof(*sc), M_CARP, M_WAITOK|M_ZERO);
1641
1416
1642 sc->sc_naddrs++;
1643 SC2IFP(sc)->if_flags |= IFF_UP;
1644 if (own)
1645 sc->sc_advskew = 0;
1646 carp_sc_state_locked(sc);
1647 carp_setrun(sc, 0);
1417 sc->sc_advbase = CARP_DFLTINTV;
1418 sc->sc_vhid = -1; /* required setting */
1419 sc->sc_init_counter = 1;
1420 sc->sc_state = INIT;
1648
1421
1649 CARP_UNLOCK(cif);
1650 ifa_free(&ia->ia_ifa); /* XXXRW: should hold reference for softc. */
1422 sc->sc_ifasiz = sizeof(struct ifaddr *);
1423 sc->sc_ifas = malloc(sc->sc_ifasiz, M_CARP, M_WAITOK|M_ZERO);
1424 sc->sc_carpdev = ifp;
1651
1425
1652 return (0);
1426 CARP_LOCK_INIT(sc);
1427#ifdef INET
1428 callout_init_mtx(&sc->sc_md_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1429#endif
1430#ifdef INET6
1431 callout_init_mtx(&sc->sc_md6_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1432#endif
1433 callout_init_mtx(&sc->sc_ad_tmo, &sc->sc_mtx, CALLOUT_RETURNUNLOCKED);
1653
1434
1654cleanup:
1655 in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1656 ifa_free(&ia->ia_ifa);
1657 return (error);
1435 CIF_LOCK(cif);
1436 TAILQ_INSERT_TAIL(&cif->cif_vrs, sc, sc_list);
1437 CIF_UNLOCK(cif);
1438
1439 mtx_lock(&carp_mtx);
1440 LIST_INSERT_HEAD(&carp_list, sc, sc_next);
1441 mtx_unlock(&carp_mtx);
1442
1443 return (sc);
1658}
1659
1660static int
1444}
1445
1446static int
1661carp_del_addr(struct carp_softc *sc, struct sockaddr_in *sin)
1447carp_grow_ifas(struct carp_softc *sc)
1662{
1448{
1663 int error = 0;
1449 struct ifaddr **new;
1664
1450
1665 if (!--sc->sc_naddrs) {
1666 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1667 struct ip_moptions *imo = &sc->sc_imo;
1451 CARP_LOCK_ASSERT(sc);
1668
1452
1669 CARP_LOCK(cif);
1670 callout_stop(&sc->sc_ad_tmo);
1671 SC2IFP(sc)->if_flags &= ~IFF_UP;
1672 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1673 sc->sc_vhid = -1;
1674 in_delmulti(imo->imo_membership[--imo->imo_num_memberships]);
1675 imo->imo_multicast_ifp = NULL;
1676 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1677 if (!--cif->vhif_nvrs) {
1678 sc->sc_carpdev->if_carp = NULL;
1679 CARP_LOCK_DESTROY(cif);
1680 free(cif, M_CARP);
1681 } else {
1682 CARP_UNLOCK(cif);
1683 }
1684 }
1453 new = malloc(sc->sc_ifasiz * 2, M_CARP, M_NOWAIT|M_ZERO);
1454 if (new == NULL)
1455 return (ENOMEM);
1456 bcopy(sc->sc_ifas, new, sc->sc_ifasiz);
1457 free(sc->sc_ifas, M_CARP);
1458 sc->sc_ifas = new;
1459 sc->sc_ifasiz *= 2;
1685
1460
1686 return (error);
1461 return (0);
1687}
1462}
1688#endif
1689
1463
1690#ifdef INET6
1691static int
1692carp_set_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1464static void
1465carp_destroy(struct carp_softc *sc)
1693{
1466{
1694 struct ifnet *ifp;
1695 struct carp_if *cif;
1696 struct in6_ifaddr *ia, *ia_if;
1697 struct ip6_moptions *im6o = &sc->sc_im6o;
1698 struct in6_addr in6;
1699 int own, error;
1467 struct ifnet *ifp = sc->sc_carpdev;
1468 struct carp_if *cif = ifp->if_carp;
1700
1469
1701 error = 0;
1470 CIF_LOCK(cif);
1471 TAILQ_REMOVE(&cif->cif_vrs, sc, sc_list);
1472 if (TAILQ_EMPTY(&cif->cif_vrs))
1473 carp_free_if(cif);
1474 else
1475 CIF_UNLOCK(cif);
1702
1476
1703 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
1704 if (!(SC2IFP(sc)->if_flags & IFF_UP))
1705 carp_set_state(sc, INIT);
1706 if (sc->sc_naddrs6)
1707 SC2IFP(sc)->if_flags |= IFF_UP;
1708 if (sc->sc_carpdev)
1709 CARP_SCLOCK(sc);
1710 carp_setrun(sc, 0);
1711 if (sc->sc_carpdev)
1712 CARP_SCUNLOCK(sc);
1713 return (0);
1714 }
1477 mtx_lock(&carp_mtx);
1478 LIST_REMOVE(sc, sc_next);
1479 mtx_unlock(&carp_mtx);
1715
1480
1716 /* we have to do it by hands to check we won't match on us */
1717 ia_if = NULL; own = 0;
1718 IN6_IFADDR_RLOCK();
1719 TAILQ_FOREACH(ia, &V_in6_ifaddrhead, ia_link) {
1720 int i;
1481 CARP_LOCK(sc);
1482 callout_drain(&sc->sc_ad_tmo);
1483#ifdef INET
1484 callout_drain(&sc->sc_md_tmo);
1485#endif
1486#ifdef INET6
1487 callout_drain(&sc->sc_md6_tmo);
1488#endif
1489 CARP_LOCK_DESTROY(sc);
1721
1490
1722 for (i = 0; i < 4; i++) {
1723 if ((sin6->sin6_addr.s6_addr32[i] &
1724 ia->ia_prefixmask.sin6_addr.s6_addr32[i]) !=
1725 (ia->ia_addr.sin6_addr.s6_addr32[i] &
1726 ia->ia_prefixmask.sin6_addr.s6_addr32[i]))
1727 break;
1728 }
1729 /* and, yeah, we need a multicast-capable iface too */
1730 if (ia->ia_ifp != SC2IFP(sc) &&
1731 (ia->ia_ifp->if_flags & IFF_MULTICAST) &&
1732 (i == 4)) {
1733 if (!ia_if)
1734 ia_if = ia;
1735 if (IN6_ARE_ADDR_EQUAL(&sin6->sin6_addr,
1736 &ia->ia_addr.sin6_addr))
1737 own++;
1738 }
1739 }
1491 free(sc->sc_ifas, M_CARP);
1492 free(sc, M_CARP);
1493}
1740
1494
1741 if (!ia_if) {
1742 IN6_IFADDR_RUNLOCK();
1743 return (EADDRNOTAVAIL);
1744 }
1745 ia = ia_if;
1746 ifa_ref(&ia->ia_ifa);
1747 IN6_IFADDR_RUNLOCK();
1748 ifp = ia->ia_ifp;
1495static struct carp_if*
1496carp_alloc_if(struct ifnet *ifp)
1497{
1498 struct carp_if *cif;
1749
1499
1750 if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0 ||
1751 (im6o->im6o_multicast_ifp && im6o->im6o_multicast_ifp != ifp)) {
1752 ifa_free(&ia->ia_ifa);
1753 return (EADDRNOTAVAIL);
1754 }
1500 cif = malloc(sizeof(*cif), M_CARP, M_WAITOK|M_ZERO);
1755
1501
1756 if (!sc->sc_naddrs6) {
1757 struct in6_multi *in6m;
1502 if (ifpromisc(ifp, 1) != 0)
1503 goto cleanup;
1758
1504
1759 im6o->im6o_multicast_ifp = ifp;
1505 CIF_LOCK_INIT(cif);
1506 cif->cif_ifp = ifp;
1507 TAILQ_INIT(&cif->cif_vrs);
1760
1508
1761 /* join CARP multicast address */
1762 bzero(&in6, sizeof(in6));
1763 in6.s6_addr16[0] = htons(0xff02);
1764 in6.s6_addr8[15] = 0x12;
1765 if (in6_setscope(&in6, ifp, NULL) != 0)
1766 goto cleanup;
1767 in6m = NULL;
1768 error = in6_mc_join(ifp, &in6, NULL, &in6m, 0);
1769 if (error)
1770 goto cleanup;
1771 im6o->im6o_membership[0] = in6m;
1772 im6o->im6o_num_memberships++;
1509 IF_ADDR_LOCK(ifp);
1510 ifp->if_carp = cif;
1511 if_ref(ifp);
1512 IF_ADDR_UNLOCK(ifp);
1773
1513
1774 /* join solicited multicast address */
1775 bzero(&in6, sizeof(in6));
1776 in6.s6_addr16[0] = htons(0xff02);
1777 in6.s6_addr32[1] = 0;
1778 in6.s6_addr32[2] = htonl(1);
1779 in6.s6_addr32[3] = sin6->sin6_addr.s6_addr32[3];
1780 in6.s6_addr8[12] = 0xff;
1781 if (in6_setscope(&in6, ifp, NULL) != 0)
1782 goto cleanup;
1783 in6m = NULL;
1784 error = in6_mc_join(ifp, &in6, NULL, &in6m, 0);
1785 if (error)
1786 goto cleanup;
1787 im6o->im6o_membership[1] = in6m;
1788 im6o->im6o_num_memberships++;
1789 }
1514 return (cif);
1790
1515
1791 if (!ifp->if_carp) {
1792 cif = malloc(sizeof(*cif), M_CARP,
1793 M_WAITOK|M_ZERO);
1794 if (!cif) {
1795 error = ENOBUFS;
1796 goto cleanup;
1797 }
1798 if ((error = ifpromisc(ifp, 1))) {
1799 free(cif, M_CARP);
1800 goto cleanup;
1801 }
1516cleanup:
1517 free(cif, M_CARP);
1802
1518
1803 CARP_LOCK_INIT(cif);
1804 CARP_LOCK(cif);
1805 cif->vhif_ifp = ifp;
1806 TAILQ_INIT(&cif->vhif_vrs);
1807 ifp->if_carp = cif;
1519 return (NULL);
1520}
1808
1521
1809 } else {
1810 struct carp_softc *vr;
1522static void
1523carp_free_if(struct carp_if *cif)
1524{
1525 struct ifnet *ifp = cif->cif_ifp;
1811
1526
1812 cif = (struct carp_if *)ifp->if_carp;
1813 CARP_LOCK(cif);
1814 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
1815 if (vr != sc && vr->sc_vhid == sc->sc_vhid) {
1816 CARP_UNLOCK(cif);
1817 error = EINVAL;
1818 goto cleanup;
1819 }
1820 }
1821 sc->sc_ia6 = ia;
1822 sc->sc_carpdev = ifp;
1527 CIF_LOCK_ASSERT(cif);
1528 KASSERT(TAILQ_EMPTY(&cif->cif_vrs), ("%s: softc list not empty",
1529 __func__));
1823
1530
1824 { /* XXX prevent endless loop if already in queue */
1825 struct carp_softc *vr, *after = NULL;
1826 int myself = 0;
1827 cif = (struct carp_if *)ifp->if_carp;
1828 CARP_LOCK_ASSERT(cif);
1531 IF_ADDR_LOCK(ifp);
1532 ifp->if_carp = NULL;
1533 if_rele(ifp);
1534 IF_ADDR_UNLOCK(ifp);
1829
1535
1830 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list) {
1831 if (vr == sc)
1832 myself = 1;
1833 if (vr->sc_vhid < sc->sc_vhid)
1834 after = vr;
1835 }
1536 CIF_LOCK_DESTROY(cif);
1836
1537
1837 if (!myself) {
1838 /* We're trying to keep things in order */
1839 if (after == NULL) {
1840 TAILQ_INSERT_TAIL(&cif->vhif_vrs, sc, sc_list);
1841 } else {
1842 TAILQ_INSERT_AFTER(&cif->vhif_vrs, after, sc, sc_list);
1843 }
1844 cif->vhif_nvrs++;
1845 }
1846 }
1538 ifpromisc(ifp, 0);
1847
1539
1848 sc->sc_naddrs6++;
1849 SC2IFP(sc)->if_flags |= IFF_UP;
1850 if (own)
1851 sc->sc_advskew = 0;
1852 carp_sc_state_locked(sc);
1853 carp_setrun(sc, 0);
1854
1855 CARP_UNLOCK(cif);
1856 ifa_free(&ia->ia_ifa); /* XXXRW: should hold reference for softc. */
1857
1858 return (0);
1859
1860cleanup:
1861 if (!sc->sc_naddrs6)
1862 carp_multicast6_cleanup(sc, 1);
1863 ifa_free(&ia->ia_ifa);
1864 return (error);
1540 free(cif, M_CARP);
1865}
1866
1541}
1542
1867static int
1868carp_del_addr6(struct carp_softc *sc, struct sockaddr_in6 *sin6)
1543static void
1544carp_carprcp(struct carpreq *carpr, struct carp_softc *sc, int priv)
1869{
1545{
1870 int error = 0;
1871
1546
1872 if (!--sc->sc_naddrs6) {
1873 struct carp_if *cif = (struct carp_if *)sc->sc_carpdev->if_carp;
1874
1875 CARP_LOCK(cif);
1876 callout_stop(&sc->sc_ad_tmo);
1877 SC2IFP(sc)->if_flags &= ~IFF_UP;
1878 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
1879 sc->sc_vhid = -1;
1880 carp_multicast6_cleanup(sc, 1);
1881 TAILQ_REMOVE(&cif->vhif_vrs, sc, sc_list);
1882 if (!--cif->vhif_nvrs) {
1883 CARP_LOCK_DESTROY(cif);
1884 sc->sc_carpdev->if_carp = NULL;
1885 free(cif, M_CARP);
1886 } else
1887 CARP_UNLOCK(cif);
1888 }
1889
1890 return (error);
1547 CARP_LOCK(sc);
1548 carpr->carpr_state = sc->sc_state;
1549 carpr->carpr_vhid = sc->sc_vhid;
1550 carpr->carpr_advbase = sc->sc_advbase;
1551 carpr->carpr_advskew = sc->sc_advskew;
1552 if (priv)
1553 bcopy(sc->sc_key, carpr->carpr_key, sizeof(carpr->carpr_key));
1554 else
1555 bzero(carpr->carpr_key, sizeof(carpr->carpr_key));
1556 CARP_UNLOCK(sc);
1891}
1557}
1892#endif /* INET6 */
1893
1558
1894static int
1895carp_ioctl(struct ifnet *ifp, u_long cmd, caddr_t addr)
1559int
1560carp_ioctl(struct ifreq *ifr, u_long cmd, struct thread *td)
1896{
1561{
1897 struct carp_softc *sc = ifp->if_softc, *vr;
1898 struct carpreq carpr;
1562 struct carpreq carpr;
1899 struct ifaddr *ifa;
1900 struct ifreq *ifr;
1901 struct ifaliasreq *ifra;
1902 int locked = 0, error = 0;
1563 struct ifnet *ifp;
1564 struct carp_softc *sc = NULL;
1565 int error = 0, locked = 0;
1903
1566
1904 ifa = (struct ifaddr *)addr;
1905 ifra = (struct ifaliasreq *)addr;
1906 ifr = (struct ifreq *)addr;
1567 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1568 return (error);
1907
1569
1908 switch (cmd) {
1909 case SIOCSIFADDR:
1910 switch (ifa->ifa_addr->sa_family) {
1911#ifdef INET
1912 case AF_INET:
1913 SC2IFP(sc)->if_flags |= IFF_UP;
1914 bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1915 sizeof(struct sockaddr));
1916 error = carp_set_addr(sc, satosin(ifa->ifa_addr));
1917 break;
1918#endif /* INET */
1919#ifdef INET6
1920 case AF_INET6:
1921 SC2IFP(sc)->if_flags |= IFF_UP;
1922 error = carp_set_addr6(sc, satosin6(ifa->ifa_addr));
1923 break;
1924#endif /* INET6 */
1925 default:
1926 error = EAFNOSUPPORT;
1927 break;
1928 }
1570 ifp = ifunit_ref(ifr->ifr_name);
1571 if (ifp == NULL)
1572 return (ENXIO);
1573
1574 switch (ifp->if_type) {
1575 case IFT_ETHER:
1576 case IFT_L2VLAN:
1577 case IFT_FDDI:
1578 case IFT_ISO88025:
1929 break;
1579 break;
1580 default:
1581 error = EOPNOTSUPP;
1582 goto out;
1583 }
1930
1584
1931 case SIOCAIFADDR:
1932 switch (ifa->ifa_addr->sa_family) {
1933#ifdef INET
1934 case AF_INET:
1935 SC2IFP(sc)->if_flags |= IFF_UP;
1936 bcopy(ifa->ifa_addr, ifa->ifa_dstaddr,
1937 sizeof(struct sockaddr));
1938 error = carp_set_addr(sc, satosin(&ifra->ifra_addr));
1585 if ((ifp->if_flags & IFF_MULTICAST) == 0) {
1586 error = EADDRNOTAVAIL;
1587 goto out;
1588 }
1589
1590 switch (cmd) {
1591 case SIOCSVH:
1592 if ((error = priv_check(td, PRIV_NETINET_CARP)))
1939 break;
1593 break;
1940#endif /* INET */
1941#ifdef INET6
1942 case AF_INET6:
1943 SC2IFP(sc)->if_flags |= IFF_UP;
1944 error = carp_set_addr6(sc, satosin6(&ifra->ifra_addr));
1594 if (carpr.carpr_vhid <= 0 || carpr.carpr_vhid > CARP_MAXVHID ||
1595 carpr.carpr_advbase < 0 || carpr.carpr_advskew < 0) {
1596 error = EINVAL;
1945 break;
1597 break;
1946#endif /* INET6 */
1947 default:
1948 error = EAFNOSUPPORT;
1949 break;
1950 }
1598 }
1951 break;
1952
1599
1953 case SIOCDIFADDR:
1954 switch (ifa->ifa_addr->sa_family) {
1955#ifdef INET
1956 case AF_INET:
1957 error = carp_del_addr(sc, satosin(&ifra->ifra_addr));
1958 break;
1959#endif /* INET */
1960#ifdef INET6
1961 case AF_INET6:
1962 error = carp_del_addr6(sc, satosin6(&ifra->ifra_addr));
1963 break;
1964#endif /* INET6 */
1965 default:
1966 error = EAFNOSUPPORT;
1967 break;
1600 if (ifp->if_carp) {
1601 CIF_LOCK(ifp->if_carp);
1602 IFNET_FOREACH_CARP(ifp, sc)
1603 if (sc->sc_vhid == carpr.carpr_vhid)
1604 break;
1605 CIF_UNLOCK(ifp->if_carp);
1968 }
1606 }
1969 break;
1607 if (sc == NULL) {
1608 sc = carp_alloc(ifp);
1609 if (sc == NULL) {
1610 error = EINVAL; /* XXX: ifpromisc failed */
1611 break;
1612 }
1970
1613
1971 case SIOCSIFFLAGS:
1972 if (sc->sc_carpdev) {
1973 locked = 1;
1974 CARP_SCLOCK(sc);
1614 CARP_LOCK(sc);
1615 sc->sc_vhid = carpr.carpr_vhid;
1616 LLADDR(&sc->sc_addr)[0] = 0;
1617 LLADDR(&sc->sc_addr)[1] = 0;
1618 LLADDR(&sc->sc_addr)[2] = 0x5e;
1619 LLADDR(&sc->sc_addr)[3] = 0;
1620 LLADDR(&sc->sc_addr)[4] = 1;
1621 LLADDR(&sc->sc_addr)[5] = sc->sc_vhid;
1622 } else
1623 CARP_LOCK(sc);
1624 locked = 1;
1625 if (carpr.carpr_advbase > 0) {
1626 if (carpr.carpr_advbase > 255 ||
1627 carpr.carpr_advbase < CARP_DFLTINTV) {
1628 error = EINVAL;
1629 break;
1630 }
1631 sc->sc_advbase = carpr.carpr_advbase;
1975 }
1632 }
1976 if (sc->sc_state != INIT && !(ifr->ifr_flags & IFF_UP)) {
1977 callout_stop(&sc->sc_ad_tmo);
1978 callout_stop(&sc->sc_md_tmo);
1979 callout_stop(&sc->sc_md6_tmo);
1980 if (sc->sc_state == MASTER)
1981 carp_send_ad_locked(sc);
1982 carp_set_state(sc, INIT);
1983 carp_setrun(sc, 0);
1984 } else if (sc->sc_state == INIT && (ifr->ifr_flags & IFF_UP)) {
1985 SC2IFP(sc)->if_flags |= IFF_UP;
1986 carp_setrun(sc, 0);
1633 if (carpr.carpr_advskew > 0) {
1634 if (carpr.carpr_advskew >= 255) {
1635 error = EINVAL;
1636 break;
1637 }
1638 sc->sc_advskew = carpr.carpr_advskew;
1987 }
1639 }
1988 break;
1989
1990 case SIOCSVH:
1991 error = priv_check(curthread, PRIV_NETINET_CARP);
1992 if (error)
1993 break;
1994 if ((error = copyin(ifr->ifr_data, &carpr, sizeof carpr)))
1995 break;
1996 error = 1;
1997 if (sc->sc_carpdev) {
1998 locked = 1;
1999 CARP_SCLOCK(sc);
1640 if (carpr.carpr_key[0] != '\0') {
1641 bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
1642 carp_hmac_prepare(sc);
2000 }
1643 }
2001 if (sc->sc_state != INIT && carpr.carpr_state != sc->sc_state) {
1644 if (sc->sc_state != INIT &&
1645 carpr.carpr_state != sc->sc_state) {
2002 switch (carpr.carpr_state) {
2003 case BACKUP:
2004 callout_stop(&sc->sc_ad_tmo);
2005 carp_set_state(sc, BACKUP);
2006 carp_setrun(sc, 0);
1646 switch (carpr.carpr_state) {
1647 case BACKUP:
1648 callout_stop(&sc->sc_ad_tmo);
1649 carp_set_state(sc, BACKUP);
1650 carp_setrun(sc, 0);
2007 carp_setroute(sc, RTM_DELETE);
1651 carp_delroute(sc);
2008 break;
2009 case MASTER:
2010 carp_master_down_locked(sc);
2011 break;
2012 default:
2013 break;
2014 }
2015 }
1652 break;
1653 case MASTER:
1654 carp_master_down_locked(sc);
1655 break;
1656 default:
1657 break;
1658 }
1659 }
2016 if (carpr.carpr_vhid > 0) {
2017 if (carpr.carpr_vhid > 255) {
2018 error = EINVAL;
2019 break;
2020 }
2021 if (sc->sc_carpdev) {
2022 struct carp_if *cif;
2023 cif = (struct carp_if *)sc->sc_carpdev->if_carp;
2024 TAILQ_FOREACH(vr, &cif->vhif_vrs, sc_list)
2025 if (vr != sc &&
2026 vr->sc_vhid == carpr.carpr_vhid) {
2027 error = EEXIST;
2028 break;
2029 }
2030 if (error == EEXIST)
2031 break;
2032 }
2033 sc->sc_vhid = carpr.carpr_vhid;
2034 IF_LLADDR(sc->sc_ifp)[0] = 0;
2035 IF_LLADDR(sc->sc_ifp)[1] = 0;
2036 IF_LLADDR(sc->sc_ifp)[2] = 0x5e;
2037 IF_LLADDR(sc->sc_ifp)[3] = 0;
2038 IF_LLADDR(sc->sc_ifp)[4] = 1;
2039 IF_LLADDR(sc->sc_ifp)[5] = sc->sc_vhid;
2040 error--;
1660 break;
1661
1662 case SIOCGVH:
1663 {
1664 int priveleged;
1665
1666 if (carpr.carpr_vhid < 0 || carpr.carpr_vhid > CARP_MAXVHID) {
1667 error = EINVAL;
1668 break;
2041 }
1669 }
2042 if (carpr.carpr_advbase > 0 || carpr.carpr_advskew > 0) {
2043 if (carpr.carpr_advskew >= 255) {
2044 error = EINVAL;
1670 if (carpr.carpr_count < 1) {
1671 error = EMSGSIZE;
1672 break;
1673 }
1674 if (ifp->if_carp == NULL) {
1675 error = ENOENT;
1676 break;
1677 }
1678
1679 priveleged = (priv_check(td, PRIV_NETINET_CARP) == 0);
1680 if (carpr.carpr_vhid != 0) {
1681 CIF_LOCK(ifp->if_carp);
1682 IFNET_FOREACH_CARP(ifp, sc)
1683 if (sc->sc_vhid == carpr.carpr_vhid)
1684 break;
1685 CIF_UNLOCK(ifp->if_carp);
1686 if (sc == NULL) {
1687 error = ENOENT;
2045 break;
2046 }
1688 break;
1689 }
2047 if (carpr.carpr_advbase > 255) {
2048 error = EINVAL;
1690 carp_carprcp(&carpr, sc, priveleged);
1691 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
1692 } else {
1693 int i, count;
1694
1695 count = 0;
1696 CIF_LOCK(ifp->if_carp);
1697 IFNET_FOREACH_CARP(ifp, sc)
1698 count++;
1699
1700 if (count > carpr.carpr_count) {
1701 CIF_UNLOCK(ifp->if_carp);
1702 error = EMSGSIZE;
2049 break;
2050 }
1703 break;
1704 }
2051 sc->sc_advbase = carpr.carpr_advbase;
2052 sc->sc_advskew = carpr.carpr_advskew;
2053 error--;
1705
1706 i = 0;
1707 IFNET_FOREACH_CARP(ifp, sc) {
1708 carp_carprcp(&carpr, sc, priveleged);
1709 carpr.carpr_count = count;
1710 error = copyout(&carpr, ifr->ifr_data +
1711 (i * sizeof(carpr)), sizeof(carpr));
1712 if (error) {
1713 CIF_UNLOCK(ifp->if_carp);
1714 break;
1715 }
1716 i++;
1717 }
1718 CIF_UNLOCK(ifp->if_carp);
2054 }
1719 }
2055 bcopy(carpr.carpr_key, sc->sc_key, sizeof(sc->sc_key));
2056 if (error > 0)
2057 error = EINVAL;
2058 else {
2059 error = 0;
2060 carp_setrun(sc, 0);
2061 }
2062 break;
1720 break;
2063
2064 case SIOCGVH:
2065 /* XXX: lockless read */
2066 bzero(&carpr, sizeof(carpr));
2067 carpr.carpr_state = sc->sc_state;
2068 carpr.carpr_vhid = sc->sc_vhid;
2069 carpr.carpr_advbase = sc->sc_advbase;
2070 carpr.carpr_advskew = sc->sc_advskew;
2071 error = priv_check(curthread, PRIV_NETINET_CARP);
2072 if (error == 0)
2073 bcopy(sc->sc_key, carpr.carpr_key,
2074 sizeof(carpr.carpr_key));
2075 error = copyout(&carpr, ifr->ifr_data, sizeof(carpr));
2076 break;
2077
1721 }
2078 default:
2079 error = EINVAL;
2080 }
2081
1722 default:
1723 error = EINVAL;
1724 }
1725
1726out:
2082 if (locked)
1727 if (locked)
2083 CARP_SCUNLOCK(sc);
1728 CARP_UNLOCK(sc);
1729 if_rele(ifp);
2084
1730
2085 carp_hmac_prepare(sc);
2086
2087 return (error);
2088}
2089
1731 return (error);
1732}
1733
2090/*
2091 * XXX: this is looutput. We should eventually use it from there.
2092 */
2093static int
1734static int
2094carp_looutput(struct ifnet *ifp, struct mbuf *m, struct sockaddr *dst,
2095 struct route *ro)
1735carp_get_vhid(struct ifaddr *ifa)
2096{
1736{
2097 u_int32_t af;
2098 struct rtentry *rt = NULL;
2099
1737
2100 M_ASSERTPKTHDR(m); /* check if we have the packet header */
1738 if (ifa == NULL || ifa->ifa_carp == NULL)
1739 return (0);
2101
1740
2102 if (ro != NULL)
2103 rt = ro->ro_rt;
2104 if (rt && rt->rt_flags & (RTF_REJECT|RTF_BLACKHOLE)) {
2105 m_freem(m);
2106 return (rt->rt_flags & RTF_BLACKHOLE ? 0 :
2107 rt->rt_flags & RTF_HOST ? EHOSTUNREACH : ENETUNREACH);
2108 }
2109
2110 ifp->if_opackets++;
2111 ifp->if_obytes += m->m_pkthdr.len;
2112
2113 /* BPF writes need to be handled specially. */
2114 if (dst->sa_family == AF_UNSPEC) {
2115 bcopy(dst->sa_data, &af, sizeof(af));
2116 dst->sa_family = af;
2117 }
2118
2119#if 1 /* XXX */
2120 switch (dst->sa_family) {
2121 case AF_INET:
2122 case AF_INET6:
2123 case AF_IPX:
2124 case AF_APPLETALK:
2125 break;
2126 default:
2127 printf("carp_looutput: af=%d unexpected\n", dst->sa_family);
2128 m_freem(m);
2129 return (EAFNOSUPPORT);
2130 }
2131#endif
2132 return(if_simloop(ifp, m, dst->sa_family, 0));
1741 return (ifa->ifa_carp->sc_vhid);
2133}
2134
1742}
1743
2135/*
2136 * Start output on carp interface. This function should never be called.
2137 */
2138static void
2139carp_start(struct ifnet *ifp)
2140{
2141#ifdef DEBUG
2142 printf("%s: start called\n", ifp->if_xname);
2143#endif
2144}
2145
2146int
1744int
2147carp_output(struct ifnet *ifp, struct mbuf *m, struct sockaddr *sa,
2148 struct rtentry *rt)
1745carp_attach(struct ifaddr *ifa, int vhid)
2149{
1746{
2150 struct m_tag *mtag;
1747 struct ifnet *ifp = ifa->ifa_ifp;
2151 struct carp_softc *sc;
1748 struct carp_softc *sc;
2152 struct ifnet *carp_ifp;
1749 int index, error;
2153
1750
2154 if (!sa)
2155 return (0);
1751 if (ifp->if_carp == NULL)
1752 return (ENOPROTOOPT);
2156
1753
2157 switch (sa->sa_family) {
1754 switch (ifa->ifa_addr->sa_family) {
2158#ifdef INET
2159 case AF_INET:
1755#ifdef INET
1756 case AF_INET:
2160 break;
2161#endif /* INET */
1757#endif
2162#ifdef INET6
2163 case AF_INET6:
1758#ifdef INET6
1759 case AF_INET6:
1760#endif
2164 break;
1761 break;
2165#endif /* INET6 */
2166 default:
1762 default:
2167 return (0);
1763 return (EPROTOTYPE);
2168 }
2169
1764 }
1765
2170 mtag = m_tag_find(m, PACKET_TAG_CARP, NULL);
2171 if (mtag == NULL)
2172 return (0);
1766 CIF_LOCK(ifp->if_carp);
1767 IFNET_FOREACH_CARP(ifp, sc)
1768 if (sc->sc_vhid == vhid)
1769 break;
1770 CIF_UNLOCK(ifp->if_carp);
1771 if (sc == NULL)
1772 return (ENOENT);
2173
1773
2174 bcopy(mtag + 1, &carp_ifp, sizeof(struct ifnet *));
2175 sc = carp_ifp->if_softc;
1774 if (ifa->ifa_carp) {
1775 if (ifa->ifa_carp->sc_vhid != vhid)
1776 carp_detach(ifa);
1777 else
1778 return (0);
1779 }
2176
1780
2177 /* Set the source MAC address to Virtual Router MAC Address */
2178 switch (ifp->if_type) {
2179 case IFT_ETHER:
2180 case IFT_L2VLAN: {
2181 struct ether_header *eh;
1781 error = carp_multicast_setup(sc, ifa->ifa_addr->sa_family);
1782 if (error)
1783 return (error);
2182
1784
2183 eh = mtod(m, struct ether_header *);
2184 eh->ether_shost[0] = 0;
2185 eh->ether_shost[1] = 0;
2186 eh->ether_shost[2] = 0x5e;
2187 eh->ether_shost[3] = 0;
2188 eh->ether_shost[4] = 1;
2189 eh->ether_shost[5] = sc->sc_vhid;
1785 CARP_LOCK(sc);
1786 index = sc->sc_naddrs + sc->sc_naddrs6 + 1;
1787 if (index > sc->sc_ifasiz / sizeof(struct ifaddr *))
1788 if ((error = carp_grow_ifas(sc)) != 0) {
1789 carp_multicast_cleanup(sc,
1790 ifa->ifa_addr->sa_family);
1791 CARP_UNLOCK(sc);
1792 return (error);
2190 }
1793 }
2191 break;
2192 case IFT_FDDI: {
2193 struct fddi_header *fh;
2194
1794
2195 fh = mtod(m, struct fddi_header *);
2196 fh->fddi_shost[0] = 0;
2197 fh->fddi_shost[1] = 0;
2198 fh->fddi_shost[2] = 0x5e;
2199 fh->fddi_shost[3] = 0;
2200 fh->fddi_shost[4] = 1;
2201 fh->fddi_shost[5] = sc->sc_vhid;
2202 }
1795 switch (ifa->ifa_addr->sa_family) {
1796#ifdef INET
1797 case AF_INET:
1798 sc->sc_naddrs++;
2203 break;
1799 break;
2204 case IFT_ISO88025: {
2205 struct iso88025_header *th;
2206 th = mtod(m, struct iso88025_header *);
2207 th->iso88025_shost[0] = 3;
2208 th->iso88025_shost[1] = 0;
2209 th->iso88025_shost[2] = 0x40 >> (sc->sc_vhid - 1);
2210 th->iso88025_shost[3] = 0x40000 >> (sc->sc_vhid - 1);
2211 th->iso88025_shost[4] = 0;
2212 th->iso88025_shost[5] = 0;
2213 }
1800#endif
1801#ifdef INET6
1802 case AF_INET6:
1803 sc->sc_naddrs6++;
2214 break;
1804 break;
2215 default:
2216 printf("%s: carp is not supported for this interface type\n",
2217 ifp->if_xname);
2218 return (EOPNOTSUPP);
1805#endif
2219 }
2220
1806 }
1807
1808 ifa_ref(ifa);
1809 sc->sc_ifas[index - 1] = ifa;
1810 ifa->ifa_carp = sc;
1811
1812 carp_hmac_prepare(sc);
1813 carp_sc_state(sc);
1814
1815 CARP_UNLOCK(sc);
1816
2221 return (0);
2222}
2223
1817 return (0);
1818}
1819
2224static void
2225carp_set_state(struct carp_softc *sc, int state)
1820void
1821carp_detach(struct ifaddr *ifa)
2226{
1822{
2227 int link_state;
1823 struct carp_softc *sc = ifa->ifa_carp;
1824 int i, index;
2228
1825
2229 if (sc->sc_carpdev)
2230 CARP_SCLOCK_ASSERT(sc);
1826 KASSERT(sc != NULL, ("%s: %p not attached", __func__, ifa));
2231
1827
2232 if (sc->sc_state == state)
2233 return;
1828 CARP_LOCK(sc);
2234
1829
2235 sc->sc_state = state;
2236 switch (state) {
2237 case BACKUP:
2238 link_state = LINK_STATE_DOWN;
1830 /* Shift array. */
1831 index = sc->sc_naddrs + sc->sc_naddrs6;
1832 for (i = 0; i < index; i++)
1833 if (sc->sc_ifas[i] == ifa)
1834 break;
1835 KASSERT(i < index, ("%s: %p no backref", __func__, ifa));
1836 for (; i < index - 1; i++)
1837 sc->sc_ifas[i] = sc->sc_ifas[i+1];
1838 sc->sc_ifas[index - 1] = NULL;
1839
1840 switch (ifa->ifa_addr->sa_family) {
1841#ifdef INET
1842 case AF_INET:
1843 sc->sc_naddrs--;
2239 break;
1844 break;
2240 case MASTER:
2241 link_state = LINK_STATE_UP;
1845#endif
1846#ifdef INET6
1847 case AF_INET6:
1848 sc->sc_naddrs6--;
2242 break;
1849 break;
2243 default:
2244 link_state = LINK_STATE_UNKNOWN;
2245 break;
1850#endif
2246 }
1851 }
2247 if_link_state_change(SC2IFP(sc), link_state);
1852
1853 carp_multicast_cleanup(sc, ifa->ifa_addr->sa_family);
1854
1855 ifa->ifa_carp = NULL;
1856 ifa_free(ifa);
1857
1858 carp_hmac_prepare(sc);
1859 carp_sc_state(sc);
1860
1861 if (sc->sc_naddrs == 0 && sc->sc_naddrs6 == 0) {
1862 CARP_UNLOCK(sc);
1863 carp_destroy(sc);
1864 } else
1865 CARP_UNLOCK(sc);
2248}
2249
1866}
1867
2250void
2251carp_carpdev_state(struct ifnet *ifp)
1868static void
1869carp_set_state(struct carp_softc *sc, int state)
2252{
1870{
2253 struct carp_if *cif;
2254
1871
2255 cif = ifp->if_carp;
2256 CARP_LOCK(cif);
2257 carp_carpdev_state_locked(cif);
2258 CARP_UNLOCK(cif);
1872 CARP_LOCK_ASSERT(sc);
1873
1874 if (sc->sc_state != state) {
1875 const char *carp_states[] = { CARP_STATES };
1876 char subsys[IFNAMSIZ+5];
1877
1878 sc->sc_state = state;
1879
1880 snprintf(subsys, IFNAMSIZ+5, "%u@%s", sc->sc_vhid,
1881 sc->sc_carpdev->if_xname);
1882 devctl_notify("CARP", subsys, carp_states[state], NULL);
1883 }
2259}
2260
2261static void
1884}
1885
1886static void
2262carp_carpdev_state_locked(struct carp_if *cif)
1887carp_linkstate(struct ifnet *ifp)
2263{
2264 struct carp_softc *sc;
2265
1888{
1889 struct carp_softc *sc;
1890
2266 TAILQ_FOREACH(sc, &cif->vhif_vrs, sc_list)
2267 carp_sc_state_locked(sc);
1891 CIF_LOCK(ifp->if_carp);
1892 IFNET_FOREACH_CARP(ifp, sc) {
1893 CARP_LOCK(sc);
1894 carp_sc_state(sc);
1895 CARP_UNLOCK(sc);
1896 }
1897 CIF_UNLOCK(ifp->if_carp);
2268}
2269
2270static void
1898}
1899
1900static void
2271carp_sc_state_locked(struct carp_softc *sc)
1901carp_sc_state(struct carp_softc *sc)
2272{
1902{
2273 CARP_SCLOCK_ASSERT(sc);
2274
1903
1904 CARP_LOCK_ASSERT(sc);
1905
2275 if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
2276 !(sc->sc_carpdev->if_flags & IFF_UP)) {
1906 if (sc->sc_carpdev->if_link_state != LINK_STATE_UP ||
1907 !(sc->sc_carpdev->if_flags & IFF_UP)) {
2277 sc->sc_flags_backup = SC2IFP(sc)->if_flags;
2278 SC2IFP(sc)->if_flags &= ~IFF_UP;
2279 SC2IFP(sc)->if_drv_flags &= ~IFF_DRV_RUNNING;
2280 callout_stop(&sc->sc_ad_tmo);
1908 callout_stop(&sc->sc_ad_tmo);
1909#ifdef INET
2281 callout_stop(&sc->sc_md_tmo);
1910 callout_stop(&sc->sc_md_tmo);
1911#endif
1912#ifdef INET6
2282 callout_stop(&sc->sc_md6_tmo);
1913 callout_stop(&sc->sc_md6_tmo);
1914#endif
2283 carp_set_state(sc, INIT);
2284 carp_setrun(sc, 0);
2285 if (!sc->sc_suppress) {
2286 carp_suppress_preempt++;
1915 carp_set_state(sc, INIT);
1916 carp_setrun(sc, 0);
1917 if (!sc->sc_suppress) {
1918 carp_suppress_preempt++;
2287 if (carp_suppress_preempt == 1) {
2288 CARP_SCUNLOCK(sc);
2289 carp_send_ad_all();
2290 CARP_SCLOCK(sc);
2291 }
1919 if (carp_suppress_preempt == 1)
1920 carp_send_ad_all(sc);
2292 }
2293 sc->sc_suppress = 1;
2294 } else {
1921 }
1922 sc->sc_suppress = 1;
1923 } else {
2295 SC2IFP(sc)->if_flags |= sc->sc_flags_backup;
2296 carp_set_state(sc, INIT);
2297 carp_setrun(sc, 0);
2298 if (sc->sc_suppress)
2299 carp_suppress_preempt--;
2300 sc->sc_suppress = 0;
2301 }
1924 carp_set_state(sc, INIT);
1925 carp_setrun(sc, 0);
1926 if (sc->sc_suppress)
1927 carp_suppress_preempt--;
1928 sc->sc_suppress = 0;
1929 }
2302
2303 return;
2304}
2305
1930}
1931
1932
2306#ifdef INET
2307extern struct domain inetdomain;
2308static struct protosw in_carp_protosw = {
2309 .pr_type = SOCK_RAW,
2310 .pr_domain = &inetdomain,
2311 .pr_protocol = IPPROTO_CARP,
2312 .pr_flags = PR_ATOMIC|PR_ADDR,
2313 .pr_input = carp_input,

--- 16 unchanged lines hidden (view full) ---

2330 .pr_usrreqs = &rip6_usrreqs
2331};
2332#endif
2333
2334static void
2335carp_mod_cleanup(void)
2336{
2337
1933#ifdef INET
1934extern struct domain inetdomain;
1935static struct protosw in_carp_protosw = {
1936 .pr_type = SOCK_RAW,
1937 .pr_domain = &inetdomain,
1938 .pr_protocol = IPPROTO_CARP,
1939 .pr_flags = PR_ATOMIC|PR_ADDR,
1940 .pr_input = carp_input,

--- 16 unchanged lines hidden (view full) ---

1957 .pr_usrreqs = &rip6_usrreqs
1958};
1959#endif
1960
1961static void
1962carp_mod_cleanup(void)
1963{
1964
2338 if (if_detach_event_tag == NULL)
2339 return;
2340 EVENTHANDLER_DEREGISTER(ifnet_departure_event, if_detach_event_tag);
2341 if_clone_detach(&carp_cloner);
2342#ifdef INET
2343 if (proto_reg[CARP_INET] == 0) {
2344 (void)ipproto_unregister(IPPROTO_CARP);
2345 pf_proto_unregister(PF_INET, IPPROTO_CARP, SOCK_RAW);
2346 proto_reg[CARP_INET] = -1;
2347 }
2348 carp_iamatch_p = NULL;
2349#endif
2350#ifdef INET6
2351 if (proto_reg[CARP_INET6] == 0) {
2352 (void)ip6proto_unregister(IPPROTO_CARP);
2353 pf_proto_unregister(PF_INET6, IPPROTO_CARP, SOCK_RAW);
2354 proto_reg[CARP_INET6] = -1;
2355 }
2356 carp_iamatch6_p = NULL;
2357 carp_macmatch6_p = NULL;
2358#endif
1965#ifdef INET
1966 if (proto_reg[CARP_INET] == 0) {
1967 (void)ipproto_unregister(IPPROTO_CARP);
1968 pf_proto_unregister(PF_INET, IPPROTO_CARP, SOCK_RAW);
1969 proto_reg[CARP_INET] = -1;
1970 }
1971 carp_iamatch_p = NULL;
1972#endif
1973#ifdef INET6
1974 if (proto_reg[CARP_INET6] == 0) {
1975 (void)ip6proto_unregister(IPPROTO_CARP);
1976 pf_proto_unregister(PF_INET6, IPPROTO_CARP, SOCK_RAW);
1977 proto_reg[CARP_INET6] = -1;
1978 }
1979 carp_iamatch6_p = NULL;
1980 carp_macmatch6_p = NULL;
1981#endif
1982 carp_ioctl_p = NULL;
1983 carp_attach_p = NULL;
1984 carp_detach_p = NULL;
1985 carp_get_vhid_p = NULL;
2359 carp_linkstate_p = NULL;
2360 carp_forus_p = NULL;
2361 carp_output_p = NULL;
2362 mtx_destroy(&carp_mtx);
2363}
2364
2365static int
2366carp_mod_load(void)
2367{
2368 int err;
2369
1986 carp_linkstate_p = NULL;
1987 carp_forus_p = NULL;
1988 carp_output_p = NULL;
1989 mtx_destroy(&carp_mtx);
1990}
1991
1992static int
1993carp_mod_load(void)
1994{
1995 int err;
1996
2370 if_detach_event_tag = EVENTHANDLER_REGISTER(ifnet_departure_event,
2371 carp_ifdetach, NULL, EVENTHANDLER_PRI_ANY);
2372 if (if_detach_event_tag == NULL)
2373 return (ENOMEM);
2374 mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
1997 mtx_init(&carp_mtx, "carp_mtx", NULL, MTX_DEF);
2375 LIST_INIT(&carpif_list);
2376 if_clone_attach(&carp_cloner);
2377 carp_linkstate_p = carp_carpdev_state;
1998 LIST_INIT(&carp_list);
1999 carp_get_vhid_p = carp_get_vhid;
2378 carp_forus_p = carp_forus;
2379 carp_output_p = carp_output;
2000 carp_forus_p = carp_forus;
2001 carp_output_p = carp_output;
2002 carp_linkstate_p = carp_linkstate;
2003 carp_ioctl_p = carp_ioctl;
2004 carp_attach_p = carp_attach;
2005 carp_detach_p = carp_detach;
2380#ifdef INET6
2381 carp_iamatch6_p = carp_iamatch6;
2382 carp_macmatch6_p = carp_macmatch6;
2383 proto_reg[CARP_INET6] = pf_proto_register(PF_INET6,
2384 (struct protosw *)&in6_carp_protosw);
2006#ifdef INET6
2007 carp_iamatch6_p = carp_iamatch6;
2008 carp_macmatch6_p = carp_macmatch6;
2009 proto_reg[CARP_INET6] = pf_proto_register(PF_INET6,
2010 (struct protosw *)&in6_carp_protosw);
2385 if (proto_reg[CARP_INET6] != 0) {
2011 if (proto_reg[CARP_INET6]) {
2386 printf("carp: error %d attaching to PF_INET6\n",
2387 proto_reg[CARP_INET6]);
2388 carp_mod_cleanup();
2389 return (proto_reg[CARP_INET6]);
2390 }
2391 err = ip6proto_register(IPPROTO_CARP);
2392 if (err) {
2393 printf("carp: error %d registering with INET6\n", err);
2394 carp_mod_cleanup();
2395 return (err);
2396 }
2397#endif
2398#ifdef INET
2399 carp_iamatch_p = carp_iamatch;
2400 proto_reg[CARP_INET] = pf_proto_register(PF_INET, &in_carp_protosw);
2012 printf("carp: error %d attaching to PF_INET6\n",
2013 proto_reg[CARP_INET6]);
2014 carp_mod_cleanup();
2015 return (proto_reg[CARP_INET6]);
2016 }
2017 err = ip6proto_register(IPPROTO_CARP);
2018 if (err) {
2019 printf("carp: error %d registering with INET6\n", err);
2020 carp_mod_cleanup();
2021 return (err);
2022 }
2023#endif
2024#ifdef INET
2025 carp_iamatch_p = carp_iamatch;
2026 proto_reg[CARP_INET] = pf_proto_register(PF_INET, &in_carp_protosw);
2401 if (proto_reg[CARP_INET] != 0) {
2027 if (proto_reg[CARP_INET]) {
2402 printf("carp: error %d attaching to PF_INET\n",
2403 proto_reg[CARP_INET]);
2404 carp_mod_cleanup();
2405 return (proto_reg[CARP_INET]);
2406 }
2407 err = ipproto_register(IPPROTO_CARP);
2408 if (err) {
2409 printf("carp: error %d registering with INET\n", err);
2410 carp_mod_cleanup();
2411 return (err);
2412 }
2413#endif
2028 printf("carp: error %d attaching to PF_INET\n",
2029 proto_reg[CARP_INET]);
2030 carp_mod_cleanup();
2031 return (proto_reg[CARP_INET]);
2032 }
2033 err = ipproto_register(IPPROTO_CARP);
2034 if (err) {
2035 printf("carp: error %d registering with INET\n", err);
2036 carp_mod_cleanup();
2037 return (err);
2038 }
2039#endif
2414 return 0;
2040 return (0);
2415}
2416
2417static int
2418carp_modevent(module_t mod, int type, void *data)
2419{
2420 switch (type) {
2421 case MOD_LOAD:
2422 return carp_mod_load();
2423 /* NOTREACHED */
2424 case MOD_UNLOAD:
2041}
2042
2043static int
2044carp_modevent(module_t mod, int type, void *data)
2045{
2046 switch (type) {
2047 case MOD_LOAD:
2048 return carp_mod_load();
2049 /* NOTREACHED */
2050 case MOD_UNLOAD:
2425 /*
2426 * XXX: For now, disallow module unloading by default due to
2427 * a race condition where a thread may dereference one of the
2428 * function pointer hooks after the module has been
2429 * unloaded, during processing of a packet, causing a panic.
2430 */
2431#ifdef CARPMOD_CAN_UNLOAD
2432 carp_mod_cleanup();
2433#else
2434 return (EBUSY);
2435#endif
2051 mtx_lock(&carp_mtx);
2052 if (LIST_EMPTY(&carp_list))
2053 carp_mod_cleanup();
2054 else {
2055 mtx_unlock(&carp_mtx);
2056 return (EBUSY);
2057 }
2436 break;
2437
2438 default:
2439 return (EINVAL);
2440 }
2441
2442 return (0);
2443}
2444
2445static moduledata_t carp_mod = {
2446 "carp",
2447 carp_modevent,
2448 0
2449};
2450
2451DECLARE_MODULE(carp, carp_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
2058 break;
2059
2060 default:
2061 return (EINVAL);
2062 }
2063
2064 return (0);
2065}
2066
2067static moduledata_t carp_mod = {
2068 "carp",
2069 carp_modevent,
2070 0
2071};
2072
2073DECLARE_MODULE(carp, carp_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);