2 * Copyright (c) 2001 McAfee, Inc.
3 * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG
6 * This software was developed for the FreeBSD Project by Jonathan Lemon
7 * and McAfee Research, the Security Research Division of McAfee, Inc. under
8 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
9 * DARPA CHATS research program. [2001 McAfee, Inc.]
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in the
18 * documentation and/or other materials provided with the distribution.
20 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 #include "opt_pcbgroup.h"
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/kernel.h>
44 #include <sys/sysctl.h>
45 #include <sys/limits.h>
47 #include <sys/mutex.h>
48 #include <sys/malloc.h>
50 #include <sys/proc.h> /* for proc0 declaration */
51 #include <sys/random.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #include <sys/syslog.h>
55 #include <sys/ucred.h>
58 #include <crypto/siphash/siphash.h>
63 #include <net/route.h>
66 #include <netinet/in.h>
67 #include <netinet/in_systm.h>
68 #include <netinet/ip.h>
69 #include <netinet/in_var.h>
70 #include <netinet/in_pcb.h>
71 #include <netinet/ip_var.h>
72 #include <netinet/ip_options.h>
74 #include <netinet/ip6.h>
75 #include <netinet/icmp6.h>
76 #include <netinet6/nd6.h>
77 #include <netinet6/ip6_var.h>
78 #include <netinet6/in6_pcb.h>
80 #include <netinet/tcp.h>
81 #include <netinet/tcp_fsm.h>
82 #include <netinet/tcp_seq.h>
83 #include <netinet/tcp_timer.h>
84 #include <netinet/tcp_var.h>
85 #include <netinet/tcp_syncache.h>
87 #include <netinet6/tcp6_var.h>
90 #include <netinet/toecore.h>
94 #include <netipsec/ipsec.h>
96 #include <netipsec/ipsec6.h>
98 #include <netipsec/key.h>
101 #include <machine/in_cksum.h>
103 #include <security/mac/mac_framework.h>
105 static VNET_DEFINE(int, tcp_syncookies) = 1;
106 #define V_tcp_syncookies VNET(tcp_syncookies)
107 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_RW,
108 &VNET_NAME(tcp_syncookies), 0,
109 "Use TCP SYN cookies if the syncache overflows");
111 static VNET_DEFINE(int, tcp_syncookiesonly) = 0;
112 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly)
113 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_RW,
114 &VNET_NAME(tcp_syncookiesonly), 0,
115 "Use only TCP SYN cookies");
118 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
121 static void syncache_drop(struct syncache *, struct syncache_head *);
122 static void syncache_free(struct syncache *);
123 static void syncache_insert(struct syncache *, struct syncache_head *);
124 struct syncache *syncache_lookup(struct in_conninfo *, struct syncache_head **);
125 static int syncache_respond(struct syncache *);
126 static struct socket *syncache_socket(struct syncache *, struct socket *,
128 static int syncache_sysctl_count(SYSCTL_HANDLER_ARGS);
129 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch,
131 static void syncache_timer(void *);
133 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
134 uint8_t *, uintptr_t);
135 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *);
136 static struct syncache
137 *syncookie_lookup(struct in_conninfo *, struct syncache_head *,
138 struct syncache *, struct tcphdr *, struct tcpopt *,
140 static void syncookie_reseed(void *);
142 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
143 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
148 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
149 * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds,
150 * the odds are that the user has given up attempting to connect by then.
152 #define SYNCACHE_MAXREXMTS 3
154 /* Arbitrary values */
155 #define TCP_SYNCACHE_HASHSIZE 512
156 #define TCP_SYNCACHE_BUCKETLIMIT 30
158 static VNET_DEFINE(struct tcp_syncache, tcp_syncache);
159 #define V_tcp_syncache VNET(tcp_syncache)
161 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0,
164 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RDTUN,
165 &VNET_NAME(tcp_syncache.bucket_limit), 0,
166 "Per-bucket hash limit for syncache");
168 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_RDTUN,
169 &VNET_NAME(tcp_syncache.cache_limit), 0,
170 "Overall entry limit for syncache");
172 SYSCTL_VNET_PROC(_net_inet_tcp_syncache, OID_AUTO, count, (CTLTYPE_UINT|CTLFLAG_RD),
173 NULL, 0, &syncache_sysctl_count, "IU",
174 "Current number of entries in syncache");
176 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RDTUN,
177 &VNET_NAME(tcp_syncache.hashsize), 0,
178 "Size of TCP syncache hashtable");
180 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_RW,
181 &VNET_NAME(tcp_syncache.rexmt_limit), 0,
182 "Limit on SYN/ACK retransmissions");
184 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
185 SYSCTL_VNET_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
186 CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
187 "Send reset on socket allocation failure");
189 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
191 #define SYNCACHE_HASH(inc, mask) \
192 ((V_tcp_syncache.hash_secret ^ \
193 (inc)->inc_faddr.s_addr ^ \
194 ((inc)->inc_faddr.s_addr >> 16) ^ \
195 (inc)->inc_fport ^ (inc)->inc_lport) & mask)
197 #define SYNCACHE_HASH6(inc, mask) \
198 ((V_tcp_syncache.hash_secret ^ \
199 (inc)->inc6_faddr.s6_addr32[0] ^ \
200 (inc)->inc6_faddr.s6_addr32[3] ^ \
201 (inc)->inc_fport ^ (inc)->inc_lport) & mask)
203 #define ENDPTS_EQ(a, b) ( \
204 (a)->ie_fport == (b)->ie_fport && \
205 (a)->ie_lport == (b)->ie_lport && \
206 (a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr && \
207 (a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr \
210 #define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0)
212 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx)
213 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx)
214 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED)
217 * Requires the syncache entry to be already removed from the bucket list.
220 syncache_free(struct syncache *sc)
224 (void) m_free(sc->sc_ipopts);
228 mac_syncache_destroy(&sc->sc_label);
231 uma_zfree(V_tcp_syncache.zone, sc);
239 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
240 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
241 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
242 V_tcp_syncache.hash_secret = arc4random();
244 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
245 &V_tcp_syncache.hashsize);
246 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
247 &V_tcp_syncache.bucket_limit);
248 if (!powerof2(V_tcp_syncache.hashsize) ||
249 V_tcp_syncache.hashsize == 0) {
250 printf("WARNING: syncache hash size is not a power of 2.\n");
251 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
253 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
256 V_tcp_syncache.cache_limit =
257 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
258 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
259 &V_tcp_syncache.cache_limit);
261 /* Allocate the hash table. */
262 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
263 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
266 V_tcp_syncache.vnet = curvnet;
269 /* Initialize the hash buckets. */
270 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
271 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
272 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
274 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
275 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
276 V_tcp_syncache.hashbase[i].sch_length = 0;
277 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
280 /* Create the syncache entry zone. */
281 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
282 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
283 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
284 V_tcp_syncache.cache_limit);
286 /* Start the SYN cookie reseeder callout. */
287 callout_init(&V_tcp_syncache.secret.reseed, 1);
288 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
289 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
290 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
291 syncookie_reseed, &V_tcp_syncache);
296 syncache_destroy(void)
298 struct syncache_head *sch;
299 struct syncache *sc, *nsc;
302 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */
303 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
305 sch = &V_tcp_syncache.hashbase[i];
306 callout_drain(&sch->sch_timer);
309 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
310 syncache_drop(sc, sch);
312 KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
313 ("%s: sch->sch_bucket not empty", __func__));
314 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
315 __func__, sch->sch_length));
316 mtx_destroy(&sch->sch_mtx);
319 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
320 ("%s: cache_count not 0", __func__));
322 /* Free the allocated global resources. */
323 uma_zdestroy(V_tcp_syncache.zone);
324 free(V_tcp_syncache.hashbase, M_SYNCACHE);
326 callout_drain(&V_tcp_syncache.secret.reseed);
331 syncache_sysctl_count(SYSCTL_HANDLER_ARGS)
335 count = uma_zone_get_cur(V_tcp_syncache.zone);
336 return (sysctl_handle_int(oidp, &count, 0, req));
340 * Inserts a syncache entry into the specified bucket row.
341 * Locks and unlocks the syncache_head autonomously.
344 syncache_insert(struct syncache *sc, struct syncache_head *sch)
346 struct syncache *sc2;
351 * Make sure that we don't overflow the per-bucket limit.
352 * If the bucket is full, toss the oldest element.
354 if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
355 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
356 ("sch->sch_length incorrect"));
357 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
358 syncache_drop(sc2, sch);
359 TCPSTAT_INC(tcps_sc_bucketoverflow);
362 /* Put it into the bucket. */
363 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
367 if (ADDED_BY_TOE(sc)) {
368 struct toedev *tod = sc->sc_tod;
370 tod->tod_syncache_added(tod, sc->sc_todctx);
374 /* Reinitialize the bucket row's timer. */
375 if (sch->sch_length == 1)
376 sch->sch_nextc = ticks + INT_MAX;
377 syncache_timeout(sc, sch, 1);
381 TCPSTAT_INC(tcps_sc_added);
385 * Remove and free entry from syncache bucket row.
386 * Expects locked syncache head.
389 syncache_drop(struct syncache *sc, struct syncache_head *sch)
392 SCH_LOCK_ASSERT(sch);
394 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
398 if (ADDED_BY_TOE(sc)) {
399 struct toedev *tod = sc->sc_tod;
401 tod->tod_syncache_removed(tod, sc->sc_todctx);
409 * Engage/reengage time on bucket row.
412 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
414 sc->sc_rxttime = ticks +
415 TCPTV_RTOBASE * (tcp_syn_backoff[sc->sc_rxmits]);
417 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
418 sch->sch_nextc = sc->sc_rxttime;
420 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
421 syncache_timer, (void *)sch);
426 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
427 * If we have retransmitted an entry the maximum number of times, expire it.
428 * One separate timer for each bucket row.
431 syncache_timer(void *xsch)
433 struct syncache_head *sch = (struct syncache_head *)xsch;
434 struct syncache *sc, *nsc;
438 CURVNET_SET(sch->sch_sc->vnet);
440 /* NB: syncache_head has already been locked by the callout. */
441 SCH_LOCK_ASSERT(sch);
444 * In the following cycle we may remove some entries and/or
445 * advance some timeouts, so re-initialize the bucket timer.
447 sch->sch_nextc = tick + INT_MAX;
449 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
451 * We do not check if the listen socket still exists
452 * and accept the case where the listen socket may be
453 * gone by the time we resend the SYN/ACK. We do
454 * not expect this to happens often. If it does,
455 * then the RST will be sent by the time the remote
456 * host does the SYN/ACK->ACK.
458 if (TSTMP_GT(sc->sc_rxttime, tick)) {
459 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
460 sch->sch_nextc = sc->sc_rxttime;
463 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
464 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
465 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
466 "giving up and removing syncache entry\n",
470 syncache_drop(sc, sch);
471 TCPSTAT_INC(tcps_sc_stale);
474 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
475 log(LOG_DEBUG, "%s; %s: Response timeout, "
476 "retransmitting (%u) SYN|ACK\n",
477 s, __func__, sc->sc_rxmits);
481 (void) syncache_respond(sc);
482 TCPSTAT_INC(tcps_sc_retransmitted);
483 syncache_timeout(sc, sch, 0);
485 if (!TAILQ_EMPTY(&(sch)->sch_bucket))
486 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
487 syncache_timer, (void *)(sch));
492 * Find an entry in the syncache.
493 * Returns always with locked syncache_head plus a matching entry or NULL.
496 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
499 struct syncache_head *sch;
502 if (inc->inc_flags & INC_ISIPV6) {
503 sch = &V_tcp_syncache.hashbase[
504 SYNCACHE_HASH6(inc, V_tcp_syncache.hashmask)];
509 /* Circle through bucket row to find matching entry. */
510 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
511 if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
517 sch = &V_tcp_syncache.hashbase[
518 SYNCACHE_HASH(inc, V_tcp_syncache.hashmask)];
523 /* Circle through bucket row to find matching entry. */
524 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
526 if (sc->sc_inc.inc_flags & INC_ISIPV6)
529 if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
533 SCH_LOCK_ASSERT(*schp);
534 return (NULL); /* always returns with locked sch */
538 * This function is called when we get a RST for a
539 * non-existent connection, so that we can see if the
540 * connection is in the syn cache. If it is, zap it.
543 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
546 struct syncache_head *sch;
549 sc = syncache_lookup(inc, &sch); /* returns locked sch */
550 SCH_LOCK_ASSERT(sch);
553 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
554 * See RFC 793 page 65, section SEGMENT ARRIVES.
556 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
557 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
558 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
559 "FIN flag set, segment ignored\n", s, __func__);
560 TCPSTAT_INC(tcps_badrst);
565 * No corresponding connection was found in syncache.
566 * If syncookies are enabled and possibly exclusively
567 * used, or we are under memory pressure, a valid RST
568 * may not find a syncache entry. In that case we're
569 * done and no SYN|ACK retransmissions will happen.
570 * Otherwise the RST was misdirected or spoofed.
573 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
574 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
575 "syncache entry (possibly syncookie only), "
576 "segment ignored\n", s, __func__);
577 TCPSTAT_INC(tcps_badrst);
582 * If the RST bit is set, check the sequence number to see
583 * if this is a valid reset segment.
585 * In all states except SYN-SENT, all reset (RST) segments
586 * are validated by checking their SEQ-fields. A reset is
587 * valid if its sequence number is in the window.
589 * The sequence number in the reset segment is normally an
590 * echo of our outgoing acknowlegement numbers, but some hosts
591 * send a reset with the sequence number at the rightmost edge
592 * of our receive window, and we have to handle this case.
594 if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
595 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
596 syncache_drop(sc, sch);
597 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
598 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, "
599 "connection attempt aborted by remote endpoint\n",
601 TCPSTAT_INC(tcps_sc_reset);
603 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
604 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
605 "IRS %u (+WND %u), segment ignored\n",
606 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd);
607 TCPSTAT_INC(tcps_badrst);
617 syncache_badack(struct in_conninfo *inc)
620 struct syncache_head *sch;
622 sc = syncache_lookup(inc, &sch); /* returns locked sch */
623 SCH_LOCK_ASSERT(sch);
625 syncache_drop(sc, sch);
626 TCPSTAT_INC(tcps_sc_badack);
632 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th)
635 struct syncache_head *sch;
637 sc = syncache_lookup(inc, &sch); /* returns locked sch */
638 SCH_LOCK_ASSERT(sch);
642 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
643 if (ntohl(th->th_seq) != sc->sc_iss)
647 * If we've rertransmitted 3 times and this is our second error,
648 * we remove the entry. Otherwise, we allow it to continue on.
649 * This prevents us from incorrectly nuking an entry during a
650 * spurious network outage.
654 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
655 sc->sc_flags |= SCF_UNREACH;
658 syncache_drop(sc, sch);
659 TCPSTAT_INC(tcps_sc_unreach);
665 * Build a new TCP socket structure from a syncache entry.
667 static struct socket *
668 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
670 struct inpcb *inp = NULL;
676 INP_INFO_WLOCK_ASSERT(&V_tcbinfo);
679 * Ok, create the full blown connection, and set things up
680 * as they would have been set up if we had created the
681 * connection when the SYN arrived. If we can't create
682 * the connection, abort it.
684 so = sonewconn(lso, SS_ISCONNECTED);
687 * Drop the connection; we will either send a RST or
688 * have the peer retransmit its SYN again after its
691 TCPSTAT_INC(tcps_listendrop);
692 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
693 log(LOG_DEBUG, "%s; %s: Socket create failed "
694 "due to limits or memory shortage\n",
701 mac_socketpeer_set_from_mbuf(m, so);
705 inp->inp_inc.inc_fibnum = so->so_fibnum;
707 INP_HASH_WLOCK(&V_tcbinfo);
709 /* Insert new socket into PCB hash list. */
710 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
712 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
713 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
715 inp->inp_vflag &= ~INP_IPV6;
716 inp->inp_vflag |= INP_IPV4;
718 inp->inp_laddr = sc->sc_inc.inc_laddr;
724 * Install in the reservation hash table for now, but don't yet
725 * install a connection group since the full 4-tuple isn't yet
728 inp->inp_lport = sc->sc_inc.inc_lport;
729 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
731 * Undo the assignments above if we failed to
732 * put the PCB on the hash lists.
735 if (sc->sc_inc.inc_flags & INC_ISIPV6)
736 inp->in6p_laddr = in6addr_any;
739 inp->inp_laddr.s_addr = INADDR_ANY;
741 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
742 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
747 INP_HASH_WUNLOCK(&V_tcbinfo);
751 /* Copy old policy into new socket's. */
752 if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp))
753 printf("syncache_socket: could not copy policy\n");
756 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
757 struct inpcb *oinp = sotoinpcb(lso);
758 struct in6_addr laddr6;
759 struct sockaddr_in6 sin6;
761 * Inherit socket options from the listening socket.
762 * Note that in6p_inputopts are not (and should not be)
763 * copied, since it stores previously received options and is
764 * used to detect if each new option is different than the
765 * previous one and hence should be passed to a user.
766 * If we copied in6p_inputopts, a user would not be able to
767 * receive options just after calling the accept system call.
769 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
770 if (oinp->in6p_outputopts)
771 inp->in6p_outputopts =
772 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
774 sin6.sin6_family = AF_INET6;
775 sin6.sin6_len = sizeof(sin6);
776 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
777 sin6.sin6_port = sc->sc_inc.inc_fport;
778 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
779 laddr6 = inp->in6p_laddr;
780 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
781 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
782 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
783 thread0.td_ucred, m)) != 0) {
784 inp->in6p_laddr = laddr6;
785 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
786 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
791 INP_HASH_WUNLOCK(&V_tcbinfo);
794 /* Override flowlabel from in6_pcbconnect. */
795 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
796 inp->inp_flow |= sc->sc_flowlabel;
799 #if defined(INET) && defined(INET6)
804 struct in_addr laddr;
805 struct sockaddr_in sin;
807 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
809 if (inp->inp_options == NULL) {
810 inp->inp_options = sc->sc_ipopts;
811 sc->sc_ipopts = NULL;
814 sin.sin_family = AF_INET;
815 sin.sin_len = sizeof(sin);
816 sin.sin_addr = sc->sc_inc.inc_faddr;
817 sin.sin_port = sc->sc_inc.inc_fport;
818 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
819 laddr = inp->inp_laddr;
820 if (inp->inp_laddr.s_addr == INADDR_ANY)
821 inp->inp_laddr = sc->sc_inc.inc_laddr;
822 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
823 thread0.td_ucred, m)) != 0) {
824 inp->inp_laddr = laddr;
825 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
826 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
831 INP_HASH_WUNLOCK(&V_tcbinfo);
836 INP_HASH_WUNLOCK(&V_tcbinfo);
838 tcp_state_change(tp, TCPS_SYN_RECEIVED);
839 tp->iss = sc->sc_iss;
840 tp->irs = sc->sc_irs;
843 tp->snd_wl1 = sc->sc_irs;
844 tp->snd_max = tp->iss + 1;
845 tp->snd_nxt = tp->iss + 1;
846 tp->rcv_up = sc->sc_irs + 1;
847 tp->rcv_wnd = sc->sc_wnd;
848 tp->rcv_adv += tp->rcv_wnd;
849 tp->last_ack_sent = tp->rcv_nxt;
851 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
852 if (sc->sc_flags & SCF_NOOPT)
853 tp->t_flags |= TF_NOOPT;
855 if (sc->sc_flags & SCF_WINSCALE) {
856 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
857 tp->snd_scale = sc->sc_requested_s_scale;
858 tp->request_r_scale = sc->sc_requested_r_scale;
860 if (sc->sc_flags & SCF_TIMESTAMP) {
861 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
862 tp->ts_recent = sc->sc_tsreflect;
863 tp->ts_recent_age = tcp_ts_getticks();
864 tp->ts_offset = sc->sc_tsoff;
867 if (sc->sc_flags & SCF_SIGNATURE)
868 tp->t_flags |= TF_SIGNATURE;
870 if (sc->sc_flags & SCF_SACK)
871 tp->t_flags |= TF_SACK_PERMIT;
874 if (sc->sc_flags & SCF_ECN)
875 tp->t_flags |= TF_ECN_PERMIT;
878 * Set up MSS and get cached values from tcp_hostcache.
879 * This might overwrite some of the defaults we just set.
881 tcp_mss(tp, sc->sc_peer_mss);
884 * If the SYN,ACK was retransmitted, indicate that CWND to be
885 * limited to one segment in cc_conn_init().
886 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
888 if (sc->sc_rxmits > 1)
893 * Allow a TOE driver to install its hooks. Note that we hold the
894 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
895 * new connection before the TOE driver has done its thing.
897 if (ADDED_BY_TOE(sc)) {
898 struct toedev *tod = sc->sc_tod;
900 tod->tod_offload_socket(tod, sc->sc_todctx, so);
904 * Copy and activate timers.
906 tp->t_keepinit = sototcpcb(lso)->t_keepinit;
907 tp->t_keepidle = sototcpcb(lso)->t_keepidle;
908 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
909 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
910 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
914 TCPSTAT_INC(tcps_accepts);
926 * This function gets called when we receive an ACK for a
927 * socket in the LISTEN state. We look up the connection
928 * in the syncache, and if its there, we pull it out of
929 * the cache and turn it into a full-blown connection in
930 * the SYN-RECEIVED state.
933 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
934 struct socket **lsop, struct mbuf *m)
937 struct syncache_head *sch;
942 * Global TCP locks are held because we manipulate the PCB lists
943 * and create a new socket.
945 INP_INFO_WLOCK_ASSERT(&V_tcbinfo);
946 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
947 ("%s: can handle only ACK", __func__));
949 sc = syncache_lookup(inc, &sch); /* returns locked sch */
950 SCH_LOCK_ASSERT(sch);
954 * Test code for syncookies comparing the syncache stored
955 * values with the reconstructed values from the cookie.
958 syncookie_cmp(inc, sch, sc, th, to, *lsop);
963 * There is no syncache entry, so see if this ACK is
964 * a returning syncookie. To do this, first:
965 * A. See if this socket has had a syncache entry dropped in
966 * the past. We don't want to accept a bogus syncookie
967 * if we've never received a SYN.
968 * B. check that the syncookie is valid. If it is, then
969 * cobble up a fake syncache entry, and return.
971 if (!V_tcp_syncookies) {
973 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
974 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
975 "segment rejected (syncookies disabled)\n",
979 bzero(&scs, sizeof(scs));
980 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
983 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
984 log(LOG_DEBUG, "%s; %s: Segment failed "
985 "SYNCOOKIE authentication, segment rejected "
986 "(probably spoofed)\n", s, __func__);
990 /* Pull out the entry to unlock the bucket row. */
991 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
994 if (ADDED_BY_TOE(sc)) {
995 struct toedev *tod = sc->sc_tod;
997 tod->tod_syncache_removed(tod, sc->sc_todctx);
1004 * Segment validation:
1005 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1007 if (th->th_ack != sc->sc_iss + 1) {
1008 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1009 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1010 "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1015 * The SEQ must fall in the window starting at the received
1016 * initial receive sequence number + 1 (the SYN).
1018 if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1019 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1020 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1021 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1022 "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1027 * If timestamps were not negotiated during SYN/ACK they
1028 * must not appear on any segment during this session.
1030 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1031 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1032 log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1033 "segment rejected\n", s, __func__);
1038 * If timestamps were negotiated during SYN/ACK they should
1039 * appear on every segment during this session.
1040 * XXXAO: This is only informal as there have been unverified
1041 * reports of non-compliants stacks.
1043 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1044 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1045 log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1046 "no action\n", s, __func__);
1053 * If timestamps were negotiated the reflected timestamp
1054 * must be equal to what we actually sent in the SYN|ACK.
1056 if ((to->to_flags & TOF_TS) && to->to_tsecr != sc->sc_ts) {
1057 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1058 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, "
1059 "segment rejected\n",
1060 s, __func__, to->to_tsecr, sc->sc_ts);
1064 *lsop = syncache_socket(sc, *lsop, m);
1067 TCPSTAT_INC(tcps_sc_aborted);
1069 TCPSTAT_INC(tcps_sc_completed);
1071 /* how do we find the inp for the new socket? */
1076 if (sc != NULL && sc != &scs)
1085 * Given a LISTEN socket and an inbound SYN request, add
1086 * this to the syn cache, and send back a segment:
1087 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1090 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1091 * Doing so would require that we hold onto the data and deliver it
1092 * to the application. However, if we are the target of a SYN-flood
1093 * DoS attack, an attacker could send data which would eventually
1094 * consume all available buffer space if it were ACKed. By not ACKing
1095 * the data, we avoid this DoS scenario.
1098 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1099 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1104 struct syncache *sc = NULL;
1105 struct syncache_head *sch;
1106 struct mbuf *ipopts = NULL;
1108 int win, sb_hiwat, ip_ttl, ip_tos;
1111 int autoflowlabel = 0;
1114 struct label *maclabel;
1116 struct syncache scs;
1119 INP_INFO_WLOCK_ASSERT(&V_tcbinfo);
1120 INP_WLOCK_ASSERT(inp); /* listen socket */
1121 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1122 ("%s: unexpected tcp flags", __func__));
1125 * Combine all so/tp operations very early to drop the INP lock as
1130 cred = crhold(so->so_cred);
1133 if ((inc->inc_flags & INC_ISIPV6) &&
1134 (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1137 ip_ttl = inp->inp_ip_ttl;
1138 ip_tos = inp->inp_ip_tos;
1139 win = sbspace(&so->so_rcv);
1140 sb_hiwat = so->so_rcv.sb_hiwat;
1141 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1143 /* By the time we drop the lock these should no longer be used. */
1148 if (mac_syncache_init(&maclabel) != 0) {
1150 INP_INFO_WUNLOCK(&V_tcbinfo);
1153 mac_syncache_create(maclabel, inp);
1156 INP_INFO_WUNLOCK(&V_tcbinfo);
1159 * Remember the IP options, if any.
1162 if (!(inc->inc_flags & INC_ISIPV6))
1165 ipopts = (m) ? ip_srcroute(m) : NULL;
1171 * See if we already have an entry for this connection.
1172 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1174 * XXX: should the syncache be re-initialized with the contents
1175 * of the new SYN here (which may have different options?)
1177 * XXX: We do not check the sequence number to see if this is a
1178 * real retransmit or a new connection attempt. The question is
1179 * how to handle such a case; either ignore it as spoofed, or
1180 * drop the current entry and create a new one?
1182 sc = syncache_lookup(inc, &sch); /* returns locked entry */
1183 SCH_LOCK_ASSERT(sch);
1185 TCPSTAT_INC(tcps_sc_dupsyn);
1188 * If we were remembering a previous source route,
1189 * forget it and use the new one we've been given.
1192 (void) m_free(sc->sc_ipopts);
1193 sc->sc_ipopts = ipopts;
1196 * Update timestamp if present.
1198 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1199 sc->sc_tsreflect = to->to_tsval;
1201 sc->sc_flags &= ~SCF_TIMESTAMP;
1204 * Since we have already unconditionally allocated label
1205 * storage, free it up. The syncache entry will already
1206 * have an initialized label we can use.
1208 mac_syncache_destroy(&maclabel);
1210 /* Retransmit SYN|ACK and reset retransmit count. */
1211 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1212 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1213 "resetting timer and retransmitting SYN|ACK\n",
1217 if (syncache_respond(sc) == 0) {
1219 syncache_timeout(sc, sch, 1);
1220 TCPSTAT_INC(tcps_sndacks);
1221 TCPSTAT_INC(tcps_sndtotal);
1227 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1230 * The zone allocator couldn't provide more entries.
1231 * Treat this as if the cache was full; drop the oldest
1232 * entry and insert the new one.
1234 TCPSTAT_INC(tcps_sc_zonefail);
1235 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL)
1236 syncache_drop(sc, sch);
1237 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1239 if (V_tcp_syncookies) {
1240 bzero(&scs, sizeof(scs));
1245 (void) m_free(ipopts);
1252 * Fill in the syncache values.
1255 sc->sc_label = maclabel;
1259 sc->sc_ipopts = ipopts;
1260 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1262 if (!(inc->inc_flags & INC_ISIPV6))
1265 sc->sc_ip_tos = ip_tos;
1266 sc->sc_ip_ttl = ip_ttl;
1270 sc->sc_todctx = todctx;
1272 sc->sc_irs = th->th_seq;
1273 sc->sc_iss = arc4random();
1275 sc->sc_flowlabel = 0;
1278 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1279 * win was derived from socket earlier in the function.
1282 win = imin(win, TCP_MAXWIN);
1285 if (V_tcp_do_rfc1323) {
1287 * A timestamp received in a SYN makes
1288 * it ok to send timestamp requests and replies.
1290 if (to->to_flags & TOF_TS) {
1291 sc->sc_tsreflect = to->to_tsval;
1292 sc->sc_ts = tcp_ts_getticks();
1293 sc->sc_flags |= SCF_TIMESTAMP;
1295 if (to->to_flags & TOF_SCALE) {
1299 * Pick the smallest possible scaling factor that
1300 * will still allow us to scale up to sb_max, aka
1301 * kern.ipc.maxsockbuf.
1303 * We do this because there are broken firewalls that
1304 * will corrupt the window scale option, leading to
1305 * the other endpoint believing that our advertised
1306 * window is unscaled. At scale factors larger than
1307 * 5 the unscaled window will drop below 1500 bytes,
1308 * leading to serious problems when traversing these
1311 * With the default maxsockbuf of 256K, a scale factor
1312 * of 3 will be chosen by this algorithm. Those who
1313 * choose a larger maxsockbuf should watch out
1314 * for the compatiblity problems mentioned above.
1316 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1317 * or <SYN,ACK>) segment itself is never scaled.
1319 while (wscale < TCP_MAX_WINSHIFT &&
1320 (TCP_MAXWIN << wscale) < sb_max)
1322 sc->sc_requested_r_scale = wscale;
1323 sc->sc_requested_s_scale = to->to_wscale;
1324 sc->sc_flags |= SCF_WINSCALE;
1327 #ifdef TCP_SIGNATURE
1329 * If listening socket requested TCP digests, and received SYN
1330 * contains the option, flag this in the syncache so that
1331 * syncache_respond() will do the right thing with the SYN+ACK.
1332 * XXX: Currently we always record the option by default and will
1333 * attempt to use it in syncache_respond().
1335 if (to->to_flags & TOF_SIGNATURE || ltflags & TF_SIGNATURE)
1336 sc->sc_flags |= SCF_SIGNATURE;
1338 if (to->to_flags & TOF_SACKPERM)
1339 sc->sc_flags |= SCF_SACK;
1340 if (to->to_flags & TOF_MSS)
1341 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */
1342 if (ltflags & TF_NOOPT)
1343 sc->sc_flags |= SCF_NOOPT;
1344 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1345 sc->sc_flags |= SCF_ECN;
1347 if (V_tcp_syncookies)
1348 sc->sc_iss = syncookie_generate(sch, sc);
1350 if (autoflowlabel) {
1351 if (V_tcp_syncookies)
1352 sc->sc_flowlabel = sc->sc_iss;
1354 sc->sc_flowlabel = ip6_randomflowlabel();
1355 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1361 * Do a standard 3-way handshake.
1363 if (syncache_respond(sc) == 0) {
1364 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1366 else if (sc != &scs)
1367 syncache_insert(sc, sch); /* locks and unlocks sch */
1368 TCPSTAT_INC(tcps_sndacks);
1369 TCPSTAT_INC(tcps_sndtotal);
1373 TCPSTAT_INC(tcps_sc_dropped);
1381 mac_syncache_destroy(&maclabel);
1391 syncache_respond(struct syncache *sc)
1393 struct ip *ip = NULL;
1395 struct tcphdr *th = NULL;
1396 int optlen, error = 0; /* Make compiler happy */
1397 u_int16_t hlen, tlen, mssopt;
1400 struct ip6_hdr *ip6 = NULL;
1405 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1408 tlen = hlen + sizeof(struct tcphdr);
1410 /* Determine MSS we advertize to other end of connection. */
1411 mssopt = tcp_mssopt(&sc->sc_inc);
1412 if (sc->sc_peer_mss)
1413 mssopt = max( min(sc->sc_peer_mss, mssopt), V_tcp_minmss);
1415 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1416 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1417 ("syncache: mbuf too small"));
1419 /* Create the IP+TCP header from scratch. */
1420 m = m_gethdr(M_NOWAIT, MT_DATA);
1424 mac_syncache_create_mbuf(sc->sc_label, m);
1426 m->m_data += max_linkhdr;
1428 m->m_pkthdr.len = tlen;
1429 m->m_pkthdr.rcvif = NULL;
1432 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1433 ip6 = mtod(m, struct ip6_hdr *);
1434 ip6->ip6_vfc = IPV6_VERSION;
1435 ip6->ip6_nxt = IPPROTO_TCP;
1436 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1437 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1438 ip6->ip6_plen = htons(tlen - hlen);
1439 /* ip6_hlim is set after checksum */
1440 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1441 ip6->ip6_flow |= sc->sc_flowlabel;
1443 th = (struct tcphdr *)(ip6 + 1);
1446 #if defined(INET6) && defined(INET)
1451 ip = mtod(m, struct ip *);
1452 ip->ip_v = IPVERSION;
1453 ip->ip_hl = sizeof(struct ip) >> 2;
1454 ip->ip_len = htons(tlen);
1458 ip->ip_p = IPPROTO_TCP;
1459 ip->ip_src = sc->sc_inc.inc_laddr;
1460 ip->ip_dst = sc->sc_inc.inc_faddr;
1461 ip->ip_ttl = sc->sc_ip_ttl;
1462 ip->ip_tos = sc->sc_ip_tos;
1465 * See if we should do MTU discovery. Route lookups are
1466 * expensive, so we will only unset the DF bit if:
1468 * 1) path_mtu_discovery is disabled
1469 * 2) the SCF_UNREACH flag has been set
1471 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1472 ip->ip_off |= htons(IP_DF);
1474 th = (struct tcphdr *)(ip + 1);
1477 th->th_sport = sc->sc_inc.inc_lport;
1478 th->th_dport = sc->sc_inc.inc_fport;
1480 th->th_seq = htonl(sc->sc_iss);
1481 th->th_ack = htonl(sc->sc_irs + 1);
1482 th->th_off = sizeof(struct tcphdr) >> 2;
1484 th->th_flags = TH_SYN|TH_ACK;
1485 th->th_win = htons(sc->sc_wnd);
1488 if (sc->sc_flags & SCF_ECN) {
1489 th->th_flags |= TH_ECE;
1490 TCPSTAT_INC(tcps_ecn_shs);
1493 /* Tack on the TCP options. */
1494 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1498 to.to_flags = TOF_MSS;
1499 if (sc->sc_flags & SCF_WINSCALE) {
1500 to.to_wscale = sc->sc_requested_r_scale;
1501 to.to_flags |= TOF_SCALE;
1503 if (sc->sc_flags & SCF_TIMESTAMP) {
1504 /* Virgin timestamp or TCP cookie enhanced one. */
1505 to.to_tsval = sc->sc_ts;
1506 to.to_tsecr = sc->sc_tsreflect;
1507 to.to_flags |= TOF_TS;
1509 if (sc->sc_flags & SCF_SACK)
1510 to.to_flags |= TOF_SACKPERM;
1511 #ifdef TCP_SIGNATURE
1512 if (sc->sc_flags & SCF_SIGNATURE)
1513 to.to_flags |= TOF_SIGNATURE;
1515 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1517 /* Adjust headers by option size. */
1518 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1520 m->m_pkthdr.len += optlen;
1522 #ifdef TCP_SIGNATURE
1523 if (sc->sc_flags & SCF_SIGNATURE)
1524 tcp_signature_compute(m, 0, 0, optlen,
1525 to.to_signature, IPSEC_DIR_OUTBOUND);
1528 if (sc->sc_inc.inc_flags & INC_ISIPV6)
1529 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1532 ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1536 M_SETFIB(m, sc->sc_inc.inc_fibnum);
1537 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1539 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1540 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1541 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1543 ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1545 if (ADDED_BY_TOE(sc)) {
1546 struct toedev *tod = sc->sc_tod;
1548 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1553 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1556 #if defined(INET6) && defined(INET)
1561 m->m_pkthdr.csum_flags = CSUM_TCP;
1562 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1563 htons(tlen + optlen - hlen + IPPROTO_TCP));
1565 if (ADDED_BY_TOE(sc)) {
1566 struct toedev *tod = sc->sc_tod;
1568 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1573 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1580 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1581 * that exceed the capacity of the syncache by avoiding the storage of any
1582 * of the SYNs we receive. Syncookies defend against blind SYN flooding
1583 * attacks where the attacker does not have access to our responses.
1585 * Syncookies encode and include all necessary information about the
1586 * connection setup within the SYN|ACK that we send back. That way we
1587 * can avoid keeping any local state until the ACK to our SYN|ACK returns
1588 * (if ever). Normally the syncache and syncookies are running in parallel
1589 * with the latter taking over when the former is exhausted. When matching
1590 * syncache entry is found the syncookie is ignored.
1592 * The only reliable information persisting the 3WHS is our inital sequence
1593 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient
1594 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1595 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK
1596 * returns and signifies a legitimate connection if it matches the ACK.
1598 * The available space of 32 bits to store the hash and to encode the SYN
1599 * option information is very tight and we should have at least 24 bits for
1600 * the MAC to keep the number of guesses by blind spoofing reasonably high.
1602 * SYN option information we have to encode to fully restore a connection:
1603 * MSS: is imporant to chose an optimal segment size to avoid IP level
1604 * fragmentation along the path. The common MSS values can be encoded
1605 * in a 3-bit table. Uncommon values are captured by the next lower value
1606 * in the table leading to a slight increase in packetization overhead.
1607 * WSCALE: is necessary to allow large windows to be used for high delay-
1608 * bandwidth product links. Not scaling the window when it was initially
1609 * negotiated is bad for performance as lack of scaling further decreases
1610 * the apparent available send window. We only need to encode the WSCALE
1611 * we received from the remote end. Our end can be recalculated at any
1612 * time. The common WSCALE values can be encoded in a 3-bit table.
1613 * Uncommon values are captured by the next lower value in the table
1614 * making us under-estimate the available window size halving our
1615 * theoretically possible maximum throughput for that connection.
1616 * SACK: Greatly assists in packet loss recovery and requires 1 bit.
1617 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
1618 * that are included in all segments on a connection. We enable them when
1621 * Security of syncookies and attack vectors:
1623 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
1624 * together with the gloabl secret to make it unique per connection attempt.
1625 * Thus any change of any of those parameters results in a different MAC output
1626 * in an unpredictable way unless a collision is encountered. 24 bits of the
1627 * MAC are embedded into the ISS.
1629 * To prevent replay attacks two rotating global secrets are updated with a
1630 * new random value every 15 seconds. The life-time of a syncookie is thus
1633 * Vector 1: Attacking the secret. This requires finding a weakness in the
1634 * MAC itself or the way it is used here. The attacker can do a chosen plain
1635 * text attack by varying and testing the all parameters under his control.
1636 * The strength depends on the size and randomness of the secret, and the
1637 * cryptographic security of the MAC function. Due to the constant updating
1638 * of the secret the attacker has at most 29.999 seconds to find the secret
1639 * and launch spoofed connections. After that he has to start all over again.
1641 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC
1642 * size an average of 4,823 attempts are required for a 50% chance of success
1643 * to spoof a single syncookie (birthday collision paradox). However the
1644 * attacker is blind and doesn't know if one of his attempts succeeded unless
1645 * he has a side channel to interfere success from. A single connection setup
1646 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
1647 * This many attempts are required for each one blind spoofed connection. For
1648 * every additional spoofed connection he has to launch another N attempts.
1649 * Thus for a sustained rate 100 spoofed connections per second approximately
1650 * 1,800,000 packets per second would have to be sent.
1652 * NB: The MAC function should be fast so that it doesn't become a CPU
1653 * exhaustion attack vector itself.
1656 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations
1657 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
1658 * http://cr.yp.to/syncookies.html (overview)
1659 * http://cr.yp.to/syncookies/archive (details)
1662 * Schematic construction of a syncookie enabled Initial Sequence Number:
1664 * 12345678901234567890123456789012
1665 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
1667 * x 24 MAC (truncated)
1668 * W 3 Send Window Scale index
1670 * S 1 SACK permitted
1671 * P 1 Odd/even secret
1675 * Distribution and probability of certain MSS values. Those in between are
1676 * rounded down to the next lower one.
1677 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
1678 * .2% .3% 5% 7% 7% 20% 15% 45%
1680 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
1683 * Distribution and probability of certain WSCALE values. We have to map the
1684 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
1685 * bits based on prevalence of certain values. Where we don't have an exact
1686 * match for are rounded down to the next lower one letting us under-estimate
1687 * the true available window. At the moment this would happen only for the
1688 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
1689 * and window size). The absence of the WSCALE option (no scaling in either
1690 * direction) is encoded with index zero.
1691 * [WSCALE values histograms, Allman, 2012]
1692 * X 10 10 35 5 6 14 10% by host
1693 * X 11 4 5 5 18 49 3% by connections
1695 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
1698 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed
1699 * and good cryptographic properties.
1702 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
1703 uint8_t *secbits, uintptr_t secmod)
1706 uint32_t siphash[2];
1708 SipHash24_Init(&ctx);
1709 SipHash_SetKey(&ctx, secbits);
1710 switch (inc->inc_flags & INC_ISIPV6) {
1713 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
1714 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
1719 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
1720 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
1724 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
1725 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
1726 SipHash_Update(&ctx, &flags, sizeof(flags));
1727 SipHash_Update(&ctx, &secmod, sizeof(secmod));
1728 SipHash_Final((u_int8_t *)&siphash, &ctx);
1730 return (siphash[0] ^ siphash[1]);
1734 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
1736 u_int i, mss, secbit, wscale;
1739 union syncookie cookie;
1741 SCH_LOCK_ASSERT(sch);
1745 /* Map our computed MSS into the 3-bit index. */
1746 mss = min(tcp_mssopt(&sc->sc_inc), max(sc->sc_peer_mss, V_tcp_minmss));
1747 for (i = sizeof(tcp_sc_msstab) / sizeof(*tcp_sc_msstab) - 1;
1748 tcp_sc_msstab[i] > mss && i > 0;
1751 cookie.flags.mss_idx = i;
1754 * Map the send window scale into the 3-bit index but only if
1755 * the wscale option was received.
1757 if (sc->sc_flags & SCF_WINSCALE) {
1758 wscale = sc->sc_requested_s_scale;
1759 for (i = sizeof(tcp_sc_wstab) / sizeof(*tcp_sc_wstab) - 1;
1760 tcp_sc_wstab[i] > wscale && i > 0;
1763 cookie.flags.wscale_idx = i;
1766 /* Can we do SACK? */
1767 if (sc->sc_flags & SCF_SACK)
1768 cookie.flags.sack_ok = 1;
1770 /* Which of the two secrets to use. */
1771 secbit = sch->sch_sc->secret.oddeven & 0x1;
1772 cookie.flags.odd_even = secbit;
1774 secbits = sch->sch_sc->secret.key[secbit];
1775 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
1779 * Put the flags into the hash and XOR them to get better ISS number
1780 * variance. This doesn't enhance the cryptographic strength and is
1781 * done to prevent the 8 cookie bits from showing up directly on the
1785 iss |= cookie.cookie ^ (hash >> 24);
1787 /* Randomize the timestamp. */
1788 if (sc->sc_flags & SCF_TIMESTAMP) {
1789 sc->sc_ts = arc4random();
1790 sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks();
1793 TCPSTAT_INC(tcps_sc_sendcookie);
1797 static struct syncache *
1798 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
1799 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
1805 int wnd, wscale = 0;
1806 union syncookie cookie;
1808 SCH_LOCK_ASSERT(sch);
1811 * Pull information out of SYN-ACK/ACK and revert sequence number
1814 ack = th->th_ack - 1;
1815 seq = th->th_seq - 1;
1818 * Unpack the flags containing enough information to restore the
1821 cookie.cookie = (ack & 0xff) ^ (ack >> 24);
1823 /* Which of the two secrets to use. */
1824 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
1826 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
1828 /* The recomputed hash matches the ACK if this was a genuine cookie. */
1829 if ((ack & ~0xff) != (hash & ~0xff))
1832 /* Fill in the syncache values. */
1834 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1835 sc->sc_ipopts = NULL;
1840 switch (inc->inc_flags & INC_ISIPV6) {
1843 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
1844 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
1849 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
1850 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
1855 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
1857 /* We can simply recompute receive window scale we sent earlier. */
1858 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
1861 /* Only use wscale if it was enabled in the orignal SYN. */
1862 if (cookie.flags.wscale_idx > 0) {
1863 sc->sc_requested_r_scale = wscale;
1864 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
1865 sc->sc_flags |= SCF_WINSCALE;
1868 wnd = sbspace(&lso->so_rcv);
1870 wnd = imin(wnd, TCP_MAXWIN);
1873 if (cookie.flags.sack_ok)
1874 sc->sc_flags |= SCF_SACK;
1876 if (to->to_flags & TOF_TS) {
1877 sc->sc_flags |= SCF_TIMESTAMP;
1878 sc->sc_tsreflect = to->to_tsval;
1879 sc->sc_ts = to->to_tsecr;
1880 sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks();
1883 if (to->to_flags & TOF_SIGNATURE)
1884 sc->sc_flags |= SCF_SIGNATURE;
1888 TCPSTAT_INC(tcps_sc_recvcookie);
1894 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
1895 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
1898 struct syncache scs, *scx;
1901 bzero(&scs, sizeof(scs));
1902 scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
1904 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
1908 if (sc->sc_peer_mss != scx->sc_peer_mss)
1909 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
1910 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
1912 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
1913 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
1914 s, __func__, sc->sc_requested_r_scale,
1915 scx->sc_requested_r_scale);
1917 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
1918 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
1919 s, __func__, sc->sc_requested_s_scale,
1920 scx->sc_requested_s_scale);
1922 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
1923 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
1930 #endif /* INVARIANTS */
1933 syncookie_reseed(void *arg)
1935 struct tcp_syncache *sc = arg;
1940 * Reseeding the secret doesn't have to be protected by a lock.
1941 * It only must be ensured that the new random values are visible
1942 * to all CPUs in a SMP environment. The atomic with release
1943 * semantics ensures that.
1945 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
1946 secbits = sc->secret.key[secbit];
1947 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
1948 atomic_add_rel_int(&sc->secret.oddeven, 1);
1950 /* Reschedule ourself. */
1951 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
1955 * Returns the current number of syncache entries. This number
1956 * will probably change before you get around to calling
1960 syncache_pcbcount(void)
1962 struct syncache_head *sch;
1965 for (count = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
1966 /* No need to lock for a read. */
1967 sch = &V_tcp_syncache.hashbase[i];
1968 count += sch->sch_length;
1974 * Exports the syncache entries to userland so that netstat can display
1975 * them alongside the other sockets. This function is intended to be
1976 * called only from tcp_pcblist.
1978 * Due to concurrency on an active system, the number of pcbs exported
1979 * may have no relation to max_pcbs. max_pcbs merely indicates the
1980 * amount of space the caller allocated for this function to use.
1983 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
1986 struct syncache *sc;
1987 struct syncache_head *sch;
1988 int count, error, i;
1990 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
1991 sch = &V_tcp_syncache.hashbase[i];
1993 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
1994 if (count >= max_pcbs) {
1998 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2000 bzero(&xt, sizeof(xt));
2001 xt.xt_len = sizeof(xt);
2002 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2003 xt.xt_inp.inp_vflag = INP_IPV6;
2005 xt.xt_inp.inp_vflag = INP_IPV4;
2006 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo));
2007 xt.xt_tp.t_inpcb = &xt.xt_inp;
2008 xt.xt_tp.t_state = TCPS_SYN_RECEIVED;
2009 xt.xt_socket.xso_protocol = IPPROTO_TCP;
2010 xt.xt_socket.xso_len = sizeof (struct xsocket);
2011 xt.xt_socket.so_type = SOCK_STREAM;
2012 xt.xt_socket.so_state = SS_ISCONNECTING;
2013 error = SYSCTL_OUT(req, &xt, sizeof xt);
2023 *pcbs_exported = count;