2 * Copyright (c) 2001 McAfee, Inc.
3 * Copyright (c) 2006 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.
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"
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>
51 #include <sys/proc.h> /* for proc0 declaration */
52 #include <sys/random.h>
53 #include <sys/socket.h>
54 #include <sys/socketvar.h>
55 #include <sys/syslog.h>
60 #include <net/route.h>
62 #include <netinet/in.h>
63 #include <netinet/in_systm.h>
64 #include <netinet/ip.h>
65 #include <netinet/in_var.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/ip_var.h>
68 #include <netinet/ip_options.h>
70 #include <netinet/ip6.h>
71 #include <netinet/icmp6.h>
72 #include <netinet6/nd6.h>
73 #include <netinet6/ip6_var.h>
74 #include <netinet6/in6_pcb.h>
76 #include <netinet/tcp.h>
77 #include <netinet/tcp_fsm.h>
78 #include <netinet/tcp_seq.h>
79 #include <netinet/tcp_timer.h>
80 #include <netinet/tcp_var.h>
81 #include <netinet/tcp_syncache.h>
82 #include <netinet/tcp_offload.h>
84 #include <netinet6/tcp6_var.h>
88 #include <netipsec/ipsec.h>
90 #include <netipsec/ipsec6.h>
92 #include <netipsec/key.h>
95 #include <machine/in_cksum.h>
97 #include <security/mac/mac_framework.h>
99 static int tcp_syncookies = 1;
100 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_RW,
102 "Use TCP SYN cookies if the syncache overflows");
104 static int tcp_syncookiesonly = 0;
105 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_RW,
106 &tcp_syncookiesonly, 0,
107 "Use only TCP SYN cookies");
109 #define SYNCOOKIE_SECRET_SIZE 8 /* dwords */
110 #define SYNCOOKIE_LIFETIME 16 /* seconds */
113 TAILQ_ENTRY(syncache) sc_hash;
114 struct in_conninfo sc_inc; /* addresses */
115 int sc_rxttime; /* retransmit time */
116 u_int16_t sc_rxmits; /* retransmit counter */
118 u_int32_t sc_tsreflect; /* timestamp to reflect */
119 u_int32_t sc_ts; /* our timestamp to send */
120 u_int32_t sc_tsoff; /* ts offset w/ syncookies */
121 u_int32_t sc_flowlabel; /* IPv6 flowlabel */
122 tcp_seq sc_irs; /* seq from peer */
123 tcp_seq sc_iss; /* our ISS */
124 struct mbuf *sc_ipopts; /* source route */
126 u_int16_t sc_peer_mss; /* peer's MSS */
127 u_int16_t sc_wnd; /* advertised window */
128 u_int8_t sc_ip_ttl; /* IPv4 TTL */
129 u_int8_t sc_ip_tos; /* IPv4 TOS */
130 u_int8_t sc_requested_s_scale:4,
131 sc_requested_r_scale:4;
133 #define SCF_NOOPT 0x01 /* no TCP options */
134 #define SCF_WINSCALE 0x02 /* negotiated window scaling */
135 #define SCF_TIMESTAMP 0x04 /* negotiated timestamps */
136 /* MSS is implicit */
137 #define SCF_UNREACH 0x10 /* icmp unreachable received */
138 #define SCF_SIGNATURE 0x20 /* send MD5 digests */
139 #define SCF_SACK 0x80 /* send SACK option */
140 #ifndef TCP_OFFLOAD_DISABLE
141 struct toe_usrreqs *sc_tu; /* TOE operations */
142 void *sc_toepcb; /* TOE protocol block */
145 struct label *sc_label; /* MAC label reference */
149 #ifdef TCP_OFFLOAD_DISABLE
150 #define TOEPCB_ISSET(sc) (0)
152 #define TOEPCB_ISSET(sc) ((sc)->sc_toepcb != NULL)
156 struct syncache_head {
158 TAILQ_HEAD(sch_head, syncache) sch_bucket;
159 struct callout sch_timer;
163 u_int32_t sch_secbits_odd[SYNCOOKIE_SECRET_SIZE];
164 u_int32_t sch_secbits_even[SYNCOOKIE_SECRET_SIZE];
165 u_int sch_reseed; /* time_uptime, seconds */
168 static void syncache_drop(struct syncache *, struct syncache_head *);
169 static void syncache_free(struct syncache *);
170 static void syncache_insert(struct syncache *, struct syncache_head *);
171 struct syncache *syncache_lookup(struct in_conninfo *, struct syncache_head **);
172 static int syncache_respond(struct syncache *);
173 static struct socket *syncache_socket(struct syncache *, struct socket *,
175 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch,
177 static void syncache_timer(void *);
178 static void syncookie_generate(struct syncache_head *, struct syncache *,
180 static struct syncache
181 *syncookie_lookup(struct in_conninfo *, struct syncache_head *,
182 struct syncache *, struct tcpopt *, struct tcphdr *,
186 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
187 * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds,
188 * the odds are that the user has given up attempting to connect by then.
190 #define SYNCACHE_MAXREXMTS 3
192 /* Arbitrary values */
193 #define TCP_SYNCACHE_HASHSIZE 512
194 #define TCP_SYNCACHE_BUCKETLIMIT 30
196 struct tcp_syncache {
197 struct syncache_head *hashbase;
202 u_int cache_count; /* XXX: unprotected */
207 static struct tcp_syncache tcp_syncache;
209 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, "TCP SYN cache");
211 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RDTUN,
212 &tcp_syncache.bucket_limit, 0, "Per-bucket hash limit for syncache");
214 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_RDTUN,
215 &tcp_syncache.cache_limit, 0, "Overall entry limit for syncache");
217 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_RD,
218 &tcp_syncache.cache_count, 0, "Current number of entries in syncache");
220 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RDTUN,
221 &tcp_syncache.hashsize, 0, "Size of TCP syncache hashtable");
223 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_RW,
224 &tcp_syncache.rexmt_limit, 0, "Limit on SYN/ACK retransmissions");
226 int tcp_sc_rst_sock_fail = 1;
227 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, CTLFLAG_RW,
228 &tcp_sc_rst_sock_fail, 0, "Send reset on socket allocation failure");
230 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
232 #define SYNCACHE_HASH(inc, mask) \
233 ((tcp_syncache.hash_secret ^ \
234 (inc)->inc_faddr.s_addr ^ \
235 ((inc)->inc_faddr.s_addr >> 16) ^ \
236 (inc)->inc_fport ^ (inc)->inc_lport) & mask)
238 #define SYNCACHE_HASH6(inc, mask) \
239 ((tcp_syncache.hash_secret ^ \
240 (inc)->inc6_faddr.s6_addr32[0] ^ \
241 (inc)->inc6_faddr.s6_addr32[3] ^ \
242 (inc)->inc_fport ^ (inc)->inc_lport) & mask)
244 #define ENDPTS_EQ(a, b) ( \
245 (a)->ie_fport == (b)->ie_fport && \
246 (a)->ie_lport == (b)->ie_lport && \
247 (a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr && \
248 (a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr \
251 #define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0)
253 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx)
254 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx)
255 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED)
258 * Requires the syncache entry to be already removed from the bucket list.
261 syncache_free(struct syncache *sc)
264 (void) m_free(sc->sc_ipopts);
266 mac_syncache_destroy(&sc->sc_label);
269 uma_zfree(tcp_syncache.zone, sc);
277 tcp_syncache.cache_count = 0;
278 tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
279 tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
280 tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
281 tcp_syncache.hash_secret = arc4random();
283 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
284 &tcp_syncache.hashsize);
285 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
286 &tcp_syncache.bucket_limit);
287 if (!powerof2(tcp_syncache.hashsize) || tcp_syncache.hashsize == 0) {
288 printf("WARNING: syncache hash size is not a power of 2.\n");
289 tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
291 tcp_syncache.hashmask = tcp_syncache.hashsize - 1;
294 tcp_syncache.cache_limit =
295 tcp_syncache.hashsize * tcp_syncache.bucket_limit;
296 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
297 &tcp_syncache.cache_limit);
299 /* Allocate the hash table. */
300 MALLOC(tcp_syncache.hashbase, struct syncache_head *,
301 tcp_syncache.hashsize * sizeof(struct syncache_head),
302 M_SYNCACHE, M_WAITOK | M_ZERO);
304 /* Initialize the hash buckets. */
305 for (i = 0; i < tcp_syncache.hashsize; i++) {
306 TAILQ_INIT(&tcp_syncache.hashbase[i].sch_bucket);
307 mtx_init(&tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
309 callout_init_mtx(&tcp_syncache.hashbase[i].sch_timer,
310 &tcp_syncache.hashbase[i].sch_mtx, 0);
311 tcp_syncache.hashbase[i].sch_length = 0;
314 /* Create the syncache entry zone. */
315 tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
316 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
317 uma_zone_set_max(tcp_syncache.zone, tcp_syncache.cache_limit);
321 * Inserts a syncache entry into the specified bucket row.
322 * Locks and unlocks the syncache_head autonomously.
325 syncache_insert(struct syncache *sc, struct syncache_head *sch)
327 struct syncache *sc2;
332 * Make sure that we don't overflow the per-bucket limit.
333 * If the bucket is full, toss the oldest element.
335 if (sch->sch_length >= tcp_syncache.bucket_limit) {
336 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
337 ("sch->sch_length incorrect"));
338 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
339 syncache_drop(sc2, sch);
340 tcpstat.tcps_sc_bucketoverflow++;
343 /* Put it into the bucket. */
344 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
347 /* Reinitialize the bucket row's timer. */
348 if (sch->sch_length == 1)
349 sch->sch_nextc = ticks + INT_MAX;
350 syncache_timeout(sc, sch, 1);
354 tcp_syncache.cache_count++;
355 tcpstat.tcps_sc_added++;
359 * Remove and free entry from syncache bucket row.
360 * Expects locked syncache head.
363 syncache_drop(struct syncache *sc, struct syncache_head *sch)
366 SCH_LOCK_ASSERT(sch);
368 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
371 #ifndef TCP_OFFLOAD_DISABLE
373 sc->sc_tu->tu_syncache_event(TOE_SC_DROP, sc->sc_toepcb);
376 tcp_syncache.cache_count--;
380 * Engage/reengage time on bucket row.
383 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
385 sc->sc_rxttime = ticks +
386 TCPTV_RTOBASE * (tcp_backoff[sc->sc_rxmits]);
388 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
389 sch->sch_nextc = sc->sc_rxttime;
391 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
392 syncache_timer, (void *)sch);
397 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
398 * If we have retransmitted an entry the maximum number of times, expire it.
399 * One separate timer for each bucket row.
402 syncache_timer(void *xsch)
404 struct syncache_head *sch = (struct syncache_head *)xsch;
405 struct syncache *sc, *nsc;
409 /* NB: syncache_head has already been locked by the callout. */
410 SCH_LOCK_ASSERT(sch);
413 * In the following cycle we may remove some entries and/or
414 * advance some timeouts, so re-initialize the bucket timer.
416 sch->sch_nextc = tick + INT_MAX;
418 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
420 * We do not check if the listen socket still exists
421 * and accept the case where the listen socket may be
422 * gone by the time we resend the SYN/ACK. We do
423 * not expect this to happens often. If it does,
424 * then the RST will be sent by the time the remote
425 * host does the SYN/ACK->ACK.
427 if (TSTMP_GT(sc->sc_rxttime, tick)) {
428 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
429 sch->sch_nextc = sc->sc_rxttime;
432 if (sc->sc_rxmits > tcp_syncache.rexmt_limit) {
433 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
434 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
435 "giving up and removing syncache entry\n",
439 syncache_drop(sc, sch);
440 tcpstat.tcps_sc_stale++;
443 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
444 log(LOG_DEBUG, "%s; %s: Response timeout, "
445 "retransmitting (%u) SYN|ACK\n",
446 s, __func__, sc->sc_rxmits);
450 (void) syncache_respond(sc);
451 tcpstat.tcps_sc_retransmitted++;
452 syncache_timeout(sc, sch, 0);
454 if (!TAILQ_EMPTY(&(sch)->sch_bucket))
455 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
456 syncache_timer, (void *)(sch));
460 * Find an entry in the syncache.
461 * Returns always with locked syncache_head plus a matching entry or NULL.
464 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
467 struct syncache_head *sch;
470 if (inc->inc_isipv6) {
471 sch = &tcp_syncache.hashbase[
472 SYNCACHE_HASH6(inc, tcp_syncache.hashmask)];
477 /* Circle through bucket row to find matching entry. */
478 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
479 if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
485 sch = &tcp_syncache.hashbase[
486 SYNCACHE_HASH(inc, tcp_syncache.hashmask)];
491 /* Circle through bucket row to find matching entry. */
492 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
494 if (sc->sc_inc.inc_isipv6)
497 if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
501 SCH_LOCK_ASSERT(*schp);
502 return (NULL); /* always returns with locked sch */
506 * This function is called when we get a RST for a
507 * non-existent connection, so that we can see if the
508 * connection is in the syn cache. If it is, zap it.
511 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
514 struct syncache_head *sch;
517 sc = syncache_lookup(inc, &sch); /* returns locked sch */
518 SCH_LOCK_ASSERT(sch);
521 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
522 * See RFC 793 page 65, section SEGMENT ARRIVES.
524 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
525 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
526 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
527 "FIN flag set, segment ignored\n", s, __func__);
528 tcpstat.tcps_badrst++;
533 * No corresponding connection was found in syncache.
534 * If syncookies are enabled and possibly exclusively
535 * used, or we are under memory pressure, a valid RST
536 * may not find a syncache entry. In that case we're
537 * done and no SYN|ACK retransmissions will happen.
538 * Otherwise the the RST was misdirected or spoofed.
541 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
542 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
543 "syncache entry (possibly syncookie only), "
544 "segment ignored\n", s, __func__);
545 tcpstat.tcps_badrst++;
550 * If the RST bit is set, check the sequence number to see
551 * if this is a valid reset segment.
553 * In all states except SYN-SENT, all reset (RST) segments
554 * are validated by checking their SEQ-fields. A reset is
555 * valid if its sequence number is in the window.
557 * The sequence number in the reset segment is normally an
558 * echo of our outgoing acknowlegement numbers, but some hosts
559 * send a reset with the sequence number at the rightmost edge
560 * of our receive window, and we have to handle this case.
562 if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
563 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
564 syncache_drop(sc, sch);
565 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
566 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, "
567 "connection attempt aborted by remote endpoint\n",
569 tcpstat.tcps_sc_reset++;
571 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
572 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
573 "IRS %u (+WND %u), segment ignored\n",
574 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd);
575 tcpstat.tcps_badrst++;
585 syncache_badack(struct in_conninfo *inc)
588 struct syncache_head *sch;
590 sc = syncache_lookup(inc, &sch); /* returns locked sch */
591 SCH_LOCK_ASSERT(sch);
593 syncache_drop(sc, sch);
594 tcpstat.tcps_sc_badack++;
600 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th)
603 struct syncache_head *sch;
605 sc = syncache_lookup(inc, &sch); /* returns locked sch */
606 SCH_LOCK_ASSERT(sch);
610 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
611 if (ntohl(th->th_seq) != sc->sc_iss)
615 * If we've rertransmitted 3 times and this is our second error,
616 * we remove the entry. Otherwise, we allow it to continue on.
617 * This prevents us from incorrectly nuking an entry during a
618 * spurious network outage.
622 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
623 sc->sc_flags |= SCF_UNREACH;
626 syncache_drop(sc, sch);
627 tcpstat.tcps_sc_unreach++;
633 * Build a new TCP socket structure from a syncache entry.
635 static struct socket *
636 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
638 struct inpcb *inp = NULL;
643 INP_INFO_WLOCK_ASSERT(&tcbinfo);
646 * Ok, create the full blown connection, and set things up
647 * as they would have been set up if we had created the
648 * connection when the SYN arrived. If we can't create
649 * the connection, abort it.
651 so = sonewconn(lso, SS_ISCONNECTED);
654 * Drop the connection; we will either send a RST or
655 * have the peer retransmit its SYN again after its
658 tcpstat.tcps_listendrop++;
659 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
660 log(LOG_DEBUG, "%s; %s: Socket create failed "
661 "due to limits or memory shortage\n",
669 mac_socketpeer_set_from_mbuf(m, so);
674 inp->inp_inc.inc_fibnum = sc->sc_inc.inc_fibnum;
675 so->so_fibnum = sc->sc_inc.inc_fibnum;
678 /* Insert new socket into PCB hash list. */
679 inp->inp_inc.inc_isipv6 = sc->sc_inc.inc_isipv6;
681 if (sc->sc_inc.inc_isipv6) {
682 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
684 inp->inp_vflag &= ~INP_IPV6;
685 inp->inp_vflag |= INP_IPV4;
687 inp->inp_laddr = sc->sc_inc.inc_laddr;
691 inp->inp_lport = sc->sc_inc.inc_lport;
692 if (in_pcbinshash(inp) != 0) {
694 * Undo the assignments above if we failed to
695 * put the PCB on the hash lists.
698 if (sc->sc_inc.inc_isipv6)
699 inp->in6p_laddr = in6addr_any;
702 inp->inp_laddr.s_addr = INADDR_ANY;
707 /* Copy old policy into new socket's. */
708 if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp))
709 printf("syncache_socket: could not copy policy\n");
712 if (sc->sc_inc.inc_isipv6) {
713 struct inpcb *oinp = sotoinpcb(lso);
714 struct in6_addr laddr6;
715 struct sockaddr_in6 sin6;
717 * Inherit socket options from the listening socket.
718 * Note that in6p_inputopts are not (and should not be)
719 * copied, since it stores previously received options and is
720 * used to detect if each new option is different than the
721 * previous one and hence should be passed to a user.
722 * If we copied in6p_inputopts, a user would not be able to
723 * receive options just after calling the accept system call.
725 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
726 if (oinp->in6p_outputopts)
727 inp->in6p_outputopts =
728 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
730 sin6.sin6_family = AF_INET6;
731 sin6.sin6_len = sizeof(sin6);
732 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
733 sin6.sin6_port = sc->sc_inc.inc_fport;
734 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
735 laddr6 = inp->in6p_laddr;
736 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
737 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
738 if (in6_pcbconnect(inp, (struct sockaddr *)&sin6,
740 inp->in6p_laddr = laddr6;
743 /* Override flowlabel from in6_pcbconnect. */
744 inp->in6p_flowinfo &= ~IPV6_FLOWLABEL_MASK;
745 inp->in6p_flowinfo |= sc->sc_flowlabel;
749 struct in_addr laddr;
750 struct sockaddr_in sin;
752 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
754 if (inp->inp_options == NULL) {
755 inp->inp_options = sc->sc_ipopts;
756 sc->sc_ipopts = NULL;
759 sin.sin_family = AF_INET;
760 sin.sin_len = sizeof(sin);
761 sin.sin_addr = sc->sc_inc.inc_faddr;
762 sin.sin_port = sc->sc_inc.inc_fport;
763 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
764 laddr = inp->inp_laddr;
765 if (inp->inp_laddr.s_addr == INADDR_ANY)
766 inp->inp_laddr = sc->sc_inc.inc_laddr;
767 if (in_pcbconnect(inp, (struct sockaddr *)&sin,
769 inp->inp_laddr = laddr;
774 tp->t_state = TCPS_SYN_RECEIVED;
775 tp->iss = sc->sc_iss;
776 tp->irs = sc->sc_irs;
779 tp->snd_wl1 = sc->sc_irs;
780 tp->snd_max = tp->iss + 1;
781 tp->snd_nxt = tp->iss + 1;
782 tp->rcv_up = sc->sc_irs + 1;
783 tp->rcv_wnd = sc->sc_wnd;
784 tp->rcv_adv += tp->rcv_wnd;
785 tp->last_ack_sent = tp->rcv_nxt;
787 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
788 if (sc->sc_flags & SCF_NOOPT)
789 tp->t_flags |= TF_NOOPT;
791 if (sc->sc_flags & SCF_WINSCALE) {
792 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
793 tp->snd_scale = sc->sc_requested_s_scale;
794 tp->request_r_scale = sc->sc_requested_r_scale;
796 if (sc->sc_flags & SCF_TIMESTAMP) {
797 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
798 tp->ts_recent = sc->sc_tsreflect;
799 tp->ts_recent_age = ticks;
800 tp->ts_offset = sc->sc_tsoff;
803 if (sc->sc_flags & SCF_SIGNATURE)
804 tp->t_flags |= TF_SIGNATURE;
806 if (sc->sc_flags & SCF_SACK)
807 tp->t_flags |= TF_SACK_PERMIT;
811 * Set up MSS and get cached values from tcp_hostcache.
812 * This might overwrite some of the defaults we just set.
814 tcp_mss(tp, sc->sc_peer_mss);
817 * If the SYN,ACK was retransmitted, reset cwnd to 1 segment.
820 tp->snd_cwnd = tp->t_maxseg;
821 tcp_timer_activate(tp, TT_KEEP, tcp_keepinit);
825 tcpstat.tcps_accepts++;
837 * This function gets called when we receive an ACK for a
838 * socket in the LISTEN state. We look up the connection
839 * in the syncache, and if its there, we pull it out of
840 * the cache and turn it into a full-blown connection in
841 * the SYN-RECEIVED state.
844 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
845 struct socket **lsop, struct mbuf *m)
848 struct syncache_head *sch;
853 * Global TCP locks are held because we manipulate the PCB lists
854 * and create a new socket.
856 INP_INFO_WLOCK_ASSERT(&tcbinfo);
857 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
858 ("%s: can handle only ACK", __func__));
860 sc = syncache_lookup(inc, &sch); /* returns locked sch */
861 SCH_LOCK_ASSERT(sch);
864 * There is no syncache entry, so see if this ACK is
865 * a returning syncookie. To do this, first:
866 * A. See if this socket has had a syncache entry dropped in
867 * the past. We don't want to accept a bogus syncookie
868 * if we've never received a SYN.
869 * B. check that the syncookie is valid. If it is, then
870 * cobble up a fake syncache entry, and return.
872 if (!tcp_syncookies) {
874 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
875 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
876 "segment rejected (syncookies disabled)\n",
880 bzero(&scs, sizeof(scs));
881 sc = syncookie_lookup(inc, sch, &scs, to, th, *lsop);
884 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
885 log(LOG_DEBUG, "%s; %s: Segment failed "
886 "SYNCOOKIE authentication, segment rejected "
887 "(probably spoofed)\n", s, __func__);
891 /* Pull out the entry to unlock the bucket row. */
892 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
894 tcp_syncache.cache_count--;
899 * Segment validation:
900 * ACK must match our initial sequence number + 1 (the SYN|ACK).
902 if (th->th_ack != sc->sc_iss + 1 && !TOEPCB_ISSET(sc)) {
903 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
904 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
905 "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
909 * The SEQ must match the received initial receive sequence
910 * number + 1 (the SYN) because we didn't ACK any data that
911 * may have come with the SYN.
913 if (th->th_seq != sc->sc_irs + 1 && !TOEPCB_ISSET(sc)) {
914 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
915 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
916 "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
920 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
921 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
922 log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
923 "segment rejected\n", s, __func__);
927 * If timestamps were negotiated the reflected timestamp
928 * must be equal to what we actually sent in the SYN|ACK.
930 if ((to->to_flags & TOF_TS) && to->to_tsecr != sc->sc_ts &&
932 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
933 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, "
934 "segment rejected\n",
935 s, __func__, to->to_tsecr, sc->sc_ts);
939 *lsop = syncache_socket(sc, *lsop, m);
942 tcpstat.tcps_sc_aborted++;
944 tcpstat.tcps_sc_completed++;
946 /* how do we find the inp for the new socket? */
951 if (sc != NULL && sc != &scs)
960 * Given a LISTEN socket and an inbound SYN request, add
961 * this to the syn cache, and send back a segment:
962 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
965 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
966 * Doing so would require that we hold onto the data and deliver it
967 * to the application. However, if we are the target of a SYN-flood
968 * DoS attack, an attacker could send data which would eventually
969 * consume all available buffer space if it were ACKed. By not ACKing
970 * the data, we avoid this DoS scenario.
973 _syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
974 struct inpcb *inp, struct socket **lsop, struct mbuf *m,
975 struct toe_usrreqs *tu, void *toepcb)
979 struct syncache *sc = NULL;
980 struct syncache_head *sch;
981 struct mbuf *ipopts = NULL;
983 int win, sb_hiwat, ip_ttl, ip_tos, noopt;
986 int autoflowlabel = 0;
989 struct label *maclabel;
993 INP_INFO_WLOCK_ASSERT(&tcbinfo);
994 INP_WLOCK_ASSERT(inp); /* listen socket */
995 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
996 ("%s: unexpected tcp flags", __func__));
999 * Combine all so/tp operations very early to drop the INP lock as
1006 if (inc->inc_isipv6 &&
1007 (inp->in6p_flags & IN6P_AUTOFLOWLABEL))
1010 ip_ttl = inp->inp_ip_ttl;
1011 ip_tos = inp->inp_ip_tos;
1012 win = sbspace(&so->so_rcv);
1013 sb_hiwat = so->so_rcv.sb_hiwat;
1014 noopt = (tp->t_flags & TF_NOOPT);
1020 if (mac_syncache_init(&maclabel) != 0) {
1022 INP_INFO_WUNLOCK(&tcbinfo);
1025 mac_syncache_create(maclabel, inp);
1028 INP_INFO_WUNLOCK(&tcbinfo);
1031 * Remember the IP options, if any.
1034 if (!inc->inc_isipv6)
1036 ipopts = (m) ? ip_srcroute(m) : NULL;
1039 * See if we already have an entry for this connection.
1040 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1042 * XXX: should the syncache be re-initialized with the contents
1043 * of the new SYN here (which may have different options?)
1045 * XXX: We do not check the sequence number to see if this is a
1046 * real retransmit or a new connection attempt. The question is
1047 * how to handle such a case; either ignore it as spoofed, or
1048 * drop the current entry and create a new one?
1050 sc = syncache_lookup(inc, &sch); /* returns locked entry */
1051 SCH_LOCK_ASSERT(sch);
1053 #ifndef TCP_OFFLOAD_DISABLE
1055 sc->sc_tu->tu_syncache_event(TOE_SC_ENTRY_PRESENT,
1058 tcpstat.tcps_sc_dupsyn++;
1061 * If we were remembering a previous source route,
1062 * forget it and use the new one we've been given.
1065 (void) m_free(sc->sc_ipopts);
1066 sc->sc_ipopts = ipopts;
1069 * Update timestamp if present.
1071 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1072 sc->sc_tsreflect = to->to_tsval;
1074 sc->sc_flags &= ~SCF_TIMESTAMP;
1077 * Since we have already unconditionally allocated label
1078 * storage, free it up. The syncache entry will already
1079 * have an initialized label we can use.
1081 mac_syncache_destroy(&maclabel);
1082 KASSERT(sc->sc_label != NULL,
1083 ("%s: label not initialized", __func__));
1085 /* Retransmit SYN|ACK and reset retransmit count. */
1086 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1087 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1088 "resetting timer and retransmitting SYN|ACK\n",
1092 if (!TOEPCB_ISSET(sc) && syncache_respond(sc) == 0) {
1094 syncache_timeout(sc, sch, 1);
1095 tcpstat.tcps_sndacks++;
1096 tcpstat.tcps_sndtotal++;
1102 sc = uma_zalloc(tcp_syncache.zone, M_NOWAIT | M_ZERO);
1105 * The zone allocator couldn't provide more entries.
1106 * Treat this as if the cache was full; drop the oldest
1107 * entry and insert the new one.
1109 tcpstat.tcps_sc_zonefail++;
1110 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL)
1111 syncache_drop(sc, sch);
1112 sc = uma_zalloc(tcp_syncache.zone, M_NOWAIT | M_ZERO);
1114 if (tcp_syncookies) {
1115 bzero(&scs, sizeof(scs));
1120 (void) m_free(ipopts);
1127 * Fill in the syncache values.
1130 sc->sc_label = maclabel;
1132 sc->sc_ipopts = ipopts;
1133 sc->sc_inc.inc_fibnum = inp->inp_inc.inc_fibnum;
1134 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1136 if (!inc->inc_isipv6)
1139 sc->sc_ip_tos = ip_tos;
1140 sc->sc_ip_ttl = ip_ttl;
1142 #ifndef TCP_OFFLOAD_DISABLE
1144 sc->sc_toepcb = toepcb;
1146 sc->sc_irs = th->th_seq;
1147 sc->sc_iss = arc4random();
1149 sc->sc_flowlabel = 0;
1152 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1153 * win was derived from socket earlier in the function.
1156 win = imin(win, TCP_MAXWIN);
1159 if (tcp_do_rfc1323) {
1161 * A timestamp received in a SYN makes
1162 * it ok to send timestamp requests and replies.
1164 if (to->to_flags & TOF_TS) {
1165 sc->sc_tsreflect = to->to_tsval;
1167 sc->sc_flags |= SCF_TIMESTAMP;
1169 if (to->to_flags & TOF_SCALE) {
1173 * Pick the smallest possible scaling factor that
1174 * will still allow us to scale up to sb_max, aka
1175 * kern.ipc.maxsockbuf.
1177 * We do this because there are broken firewalls that
1178 * will corrupt the window scale option, leading to
1179 * the other endpoint believing that our advertised
1180 * window is unscaled. At scale factors larger than
1181 * 5 the unscaled window will drop below 1500 bytes,
1182 * leading to serious problems when traversing these
1185 * With the default maxsockbuf of 256K, a scale factor
1186 * of 3 will be chosen by this algorithm. Those who
1187 * choose a larger maxsockbuf should watch out
1188 * for the compatiblity problems mentioned above.
1190 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1191 * or <SYN,ACK>) segment itself is never scaled.
1193 while (wscale < TCP_MAX_WINSHIFT &&
1194 (TCP_MAXWIN << wscale) < sb_max)
1196 sc->sc_requested_r_scale = wscale;
1197 sc->sc_requested_s_scale = to->to_wscale;
1198 sc->sc_flags |= SCF_WINSCALE;
1201 #ifdef TCP_SIGNATURE
1203 * If listening socket requested TCP digests, and received SYN
1204 * contains the option, flag this in the syncache so that
1205 * syncache_respond() will do the right thing with the SYN+ACK.
1206 * XXX: Currently we always record the option by default and will
1207 * attempt to use it in syncache_respond().
1209 if (to->to_flags & TOF_SIGNATURE)
1210 sc->sc_flags |= SCF_SIGNATURE;
1212 if (to->to_flags & TOF_SACKPERM)
1213 sc->sc_flags |= SCF_SACK;
1214 if (to->to_flags & TOF_MSS)
1215 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */
1217 sc->sc_flags |= SCF_NOOPT;
1219 if (tcp_syncookies) {
1220 syncookie_generate(sch, sc, &flowtmp);
1223 sc->sc_flowlabel = flowtmp;
1229 (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK);
1235 * Do a standard 3-way handshake.
1237 if (TOEPCB_ISSET(sc) || syncache_respond(sc) == 0) {
1238 if (tcp_syncookies && tcp_syncookiesonly && sc != &scs)
1240 else if (sc != &scs)
1241 syncache_insert(sc, sch); /* locks and unlocks sch */
1242 tcpstat.tcps_sndacks++;
1243 tcpstat.tcps_sndtotal++;
1247 tcpstat.tcps_sc_dropped++;
1253 mac_syncache_destroy(&maclabel);
1264 syncache_respond(struct syncache *sc)
1266 struct ip *ip = NULL;
1270 u_int16_t hlen, tlen, mssopt;
1273 struct ip6_hdr *ip6 = NULL;
1278 (sc->sc_inc.inc_isipv6) ? sizeof(struct ip6_hdr) :
1281 tlen = hlen + sizeof(struct tcphdr);
1283 /* Determine MSS we advertize to other end of connection. */
1284 mssopt = tcp_mssopt(&sc->sc_inc);
1285 if (sc->sc_peer_mss)
1286 mssopt = max( min(sc->sc_peer_mss, mssopt), tcp_minmss);
1288 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1289 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1290 ("syncache: mbuf too small"));
1292 /* Create the IP+TCP header from scratch. */
1293 m = m_gethdr(M_DONTWAIT, MT_DATA);
1297 mac_syncache_create_mbuf(sc->sc_label, m);
1299 m->m_data += max_linkhdr;
1301 m->m_pkthdr.len = tlen;
1302 m->m_pkthdr.rcvif = NULL;
1305 if (sc->sc_inc.inc_isipv6) {
1306 ip6 = mtod(m, struct ip6_hdr *);
1307 ip6->ip6_vfc = IPV6_VERSION;
1308 ip6->ip6_nxt = IPPROTO_TCP;
1309 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1310 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1311 ip6->ip6_plen = htons(tlen - hlen);
1312 /* ip6_hlim is set after checksum */
1313 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1314 ip6->ip6_flow |= sc->sc_flowlabel;
1316 th = (struct tcphdr *)(ip6 + 1);
1320 ip = mtod(m, struct ip *);
1321 ip->ip_v = IPVERSION;
1322 ip->ip_hl = sizeof(struct ip) >> 2;
1327 ip->ip_p = IPPROTO_TCP;
1328 ip->ip_src = sc->sc_inc.inc_laddr;
1329 ip->ip_dst = sc->sc_inc.inc_faddr;
1330 ip->ip_ttl = sc->sc_ip_ttl;
1331 ip->ip_tos = sc->sc_ip_tos;
1334 * See if we should do MTU discovery. Route lookups are
1335 * expensive, so we will only unset the DF bit if:
1337 * 1) path_mtu_discovery is disabled
1338 * 2) the SCF_UNREACH flag has been set
1340 if (path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1341 ip->ip_off |= IP_DF;
1343 th = (struct tcphdr *)(ip + 1);
1345 th->th_sport = sc->sc_inc.inc_lport;
1346 th->th_dport = sc->sc_inc.inc_fport;
1348 th->th_seq = htonl(sc->sc_iss);
1349 th->th_ack = htonl(sc->sc_irs + 1);
1350 th->th_off = sizeof(struct tcphdr) >> 2;
1352 th->th_flags = TH_SYN|TH_ACK;
1353 th->th_win = htons(sc->sc_wnd);
1356 /* Tack on the TCP options. */
1357 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1361 to.to_flags = TOF_MSS;
1362 if (sc->sc_flags & SCF_WINSCALE) {
1363 to.to_wscale = sc->sc_requested_r_scale;
1364 to.to_flags |= TOF_SCALE;
1366 if (sc->sc_flags & SCF_TIMESTAMP) {
1367 /* Virgin timestamp or TCP cookie enhanced one. */
1368 to.to_tsval = sc->sc_ts;
1369 to.to_tsecr = sc->sc_tsreflect;
1370 to.to_flags |= TOF_TS;
1372 if (sc->sc_flags & SCF_SACK)
1373 to.to_flags |= TOF_SACKPERM;
1374 #ifdef TCP_SIGNATURE
1375 if (sc->sc_flags & SCF_SIGNATURE)
1376 to.to_flags |= TOF_SIGNATURE;
1378 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1380 /* Adjust headers by option size. */
1381 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1383 m->m_pkthdr.len += optlen;
1385 #ifdef TCP_SIGNATURE
1386 if (sc->sc_flags & SCF_SIGNATURE)
1387 tcp_signature_compute(m, sizeof(struct ip), 0, optlen,
1388 to.to_signature, IPSEC_DIR_OUTBOUND);
1391 if (sc->sc_inc.inc_isipv6)
1392 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1395 ip->ip_len += optlen;
1400 if (sc->sc_inc.inc_isipv6) {
1402 th->th_sum = in6_cksum(m, IPPROTO_TCP, hlen,
1403 tlen + optlen - hlen);
1404 ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1405 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1409 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1410 htons(tlen + optlen - hlen + IPPROTO_TCP));
1411 m->m_pkthdr.csum_flags = CSUM_TCP;
1412 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1413 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1419 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1420 struct inpcb *inp, struct socket **lsop, struct mbuf *m)
1422 _syncache_add(inc, to, th, inp, lsop, m, NULL, NULL);
1426 syncache_offload_add(struct in_conninfo *inc, struct tcpopt *to,
1427 struct tcphdr *th, struct inpcb *inp, struct socket **lsop,
1428 struct toe_usrreqs *tu, void *toepcb)
1431 INP_INFO_WLOCK(&tcbinfo);
1433 _syncache_add(inc, to, th, inp, lsop, NULL, tu, toepcb);
1437 * The purpose of SYN cookies is to avoid keeping track of all SYN's we
1438 * receive and to be able to handle SYN floods from bogus source addresses
1439 * (where we will never receive any reply). SYN floods try to exhaust all
1440 * our memory and available slots in the SYN cache table to cause a denial
1441 * of service to legitimate users of the local host.
1443 * The idea of SYN cookies is to encode and include all necessary information
1444 * about the connection setup state within the SYN-ACK we send back and thus
1445 * to get along without keeping any local state until the ACK to the SYN-ACK
1446 * arrives (if ever). Everything we need to know should be available from
1447 * the information we encoded in the SYN-ACK.
1449 * More information about the theory behind SYN cookies and its first
1450 * discussion and specification can be found at:
1451 * http://cr.yp.to/syncookies.html (overview)
1452 * http://cr.yp.to/syncookies/archive (gory details)
1454 * This implementation extends the orginal idea and first implementation
1455 * of FreeBSD by using not only the initial sequence number field to store
1456 * information but also the timestamp field if present. This way we can
1457 * keep track of the entire state we need to know to recreate the session in
1458 * its original form. Almost all TCP speakers implement RFC1323 timestamps
1459 * these days. For those that do not we still have to live with the known
1460 * shortcomings of the ISN only SYN cookies.
1464 * Initial sequence number we send:
1465 * 31|................................|0
1466 * DDDDDDDDDDDDDDDDDDDDDDDDDMMMRRRP
1467 * D = MD5 Digest (first dword)
1469 * R = Rotation of secret
1470 * P = Odd or Even secret
1472 * The MD5 Digest is computed with over following parameters:
1473 * a) randomly rotated secret
1474 * b) struct in_conninfo containing the remote/local ip/port (IPv4&IPv6)
1475 * c) the received initial sequence number from remote host
1476 * d) the rotation offset and odd/even bit
1478 * Timestamp we send:
1479 * 31|................................|0
1480 * DDDDDDDDDDDDDDDDDDDDDDSSSSRRRRA5
1481 * D = MD5 Digest (third dword) (only as filler)
1482 * S = Requested send window scale
1483 * R = Requested receive window scale
1485 * 5 = TCP-MD5 enabled (not implemented yet)
1486 * XORed with MD5 Digest (forth dword)
1488 * The timestamp isn't cryptographically secure and doesn't need to be.
1489 * The double use of the MD5 digest dwords ties it to a specific remote/
1490 * local host/port, remote initial sequence number and our local time
1491 * limited secret. A received timestamp is reverted (XORed) and then
1492 * the contained MD5 dword is compared to the computed one to ensure the
1493 * timestamp belongs to the SYN-ACK we sent. The other parameters may
1494 * have been tampered with but this isn't different from supplying bogus
1495 * values in the SYN in the first place.
1497 * Some problems with SYN cookies remain however:
1498 * Consider the problem of a recreated (and retransmitted) cookie. If the
1499 * original SYN was accepted, the connection is established. The second
1500 * SYN is inflight, and if it arrives with an ISN that falls within the
1501 * receive window, the connection is killed.
1504 * A heuristic to determine when to accept syn cookies is not necessary.
1505 * An ACK flood would cause the syncookie verification to be attempted,
1506 * but a SYN flood causes syncookies to be generated. Both are of equal
1507 * cost, so there's no point in trying to optimize the ACK flood case.
1508 * Also, if you don't process certain ACKs for some reason, then all someone
1509 * would have to do is launch a SYN and ACK flood at the same time, which
1510 * would stop cookie verification and defeat the entire purpose of syncookies.
1512 static int tcp_sc_msstab[] = { 0, 256, 468, 536, 996, 1452, 1460, 8960 };
1515 syncookie_generate(struct syncache_head *sch, struct syncache *sc,
1516 u_int32_t *flowlabel)
1519 u_int32_t md5_buffer[MD5_DIGEST_LENGTH / sizeof(u_int32_t)];
1522 u_int off, pmss, mss;
1525 SCH_LOCK_ASSERT(sch);
1527 /* Which of the two secrets to use. */
1528 secbits = sch->sch_oddeven ?
1529 sch->sch_secbits_odd : sch->sch_secbits_even;
1531 /* Reseed secret if too old. */
1532 if (sch->sch_reseed < time_uptime) {
1533 sch->sch_oddeven = sch->sch_oddeven ? 0 : 1; /* toggle */
1534 secbits = sch->sch_oddeven ?
1535 sch->sch_secbits_odd : sch->sch_secbits_even;
1536 for (i = 0; i < SYNCOOKIE_SECRET_SIZE; i++)
1537 secbits[i] = arc4random();
1538 sch->sch_reseed = time_uptime + SYNCOOKIE_LIFETIME;
1541 /* Secret rotation offset. */
1542 off = sc->sc_iss & 0x7; /* iss was randomized before */
1544 /* Maximum segment size calculation. */
1545 pmss = max( min(sc->sc_peer_mss, tcp_mssopt(&sc->sc_inc)), tcp_minmss);
1546 for (mss = sizeof(tcp_sc_msstab) / sizeof(int) - 1; mss > 0; mss--)
1547 if (tcp_sc_msstab[mss] <= pmss)
1550 /* Fold parameters and MD5 digest into the ISN we will send. */
1551 data = sch->sch_oddeven;/* odd or even secret, 1 bit */
1552 data |= off << 1; /* secret offset, derived from iss, 3 bits */
1553 data |= mss << 4; /* mss, 3 bits */
1556 MD5Update(&ctx, ((u_int8_t *)secbits) + off,
1557 SYNCOOKIE_SECRET_SIZE * sizeof(*secbits) - off);
1558 MD5Update(&ctx, secbits, off);
1559 MD5Update(&ctx, &sc->sc_inc, sizeof(sc->sc_inc));
1560 MD5Update(&ctx, &sc->sc_irs, sizeof(sc->sc_irs));
1561 MD5Update(&ctx, &data, sizeof(data));
1562 MD5Final((u_int8_t *)&md5_buffer, &ctx);
1564 data |= (md5_buffer[0] << 7);
1568 *flowlabel = md5_buffer[1] & IPV6_FLOWLABEL_MASK;
1571 /* Additional parameters are stored in the timestamp if present. */
1572 if (sc->sc_flags & SCF_TIMESTAMP) {
1573 data = ((sc->sc_flags & SCF_SIGNATURE) ? 1 : 0); /* TCP-MD5, 1 bit */
1574 data |= ((sc->sc_flags & SCF_SACK) ? 1 : 0) << 1; /* SACK, 1 bit */
1575 data |= sc->sc_requested_s_scale << 2; /* SWIN scale, 4 bits */
1576 data |= sc->sc_requested_r_scale << 6; /* RWIN scale, 4 bits */
1577 data |= md5_buffer[2] << 10; /* more digest bits */
1578 data ^= md5_buffer[3];
1580 sc->sc_tsoff = data - ticks; /* after XOR */
1583 tcpstat.tcps_sc_sendcookie++;
1587 static struct syncache *
1588 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
1589 struct syncache *sc, struct tcpopt *to, struct tcphdr *th,
1593 u_int32_t md5_buffer[MD5_DIGEST_LENGTH / sizeof(u_int32_t)];
1597 int off, mss, wnd, flags;
1599 SCH_LOCK_ASSERT(sch);
1602 * Pull information out of SYN-ACK/ACK and
1603 * revert sequence number advances.
1605 ack = th->th_ack - 1;
1606 seq = th->th_seq - 1;
1607 off = (ack >> 1) & 0x7;
1608 mss = (ack >> 4) & 0x7;
1611 /* Which of the two secrets to use. */
1612 secbits = (flags & 0x1) ? sch->sch_secbits_odd : sch->sch_secbits_even;
1615 * The secret wasn't updated for the lifetime of a syncookie,
1616 * so this SYN-ACK/ACK is either too old (replay) or totally bogus.
1618 if (sch->sch_reseed < time_uptime) {
1622 /* Recompute the digest so we can compare it. */
1624 MD5Update(&ctx, ((u_int8_t *)secbits) + off,
1625 SYNCOOKIE_SECRET_SIZE * sizeof(*secbits) - off);
1626 MD5Update(&ctx, secbits, off);
1627 MD5Update(&ctx, inc, sizeof(*inc));
1628 MD5Update(&ctx, &seq, sizeof(seq));
1629 MD5Update(&ctx, &flags, sizeof(flags));
1630 MD5Final((u_int8_t *)&md5_buffer, &ctx);
1632 /* Does the digest part of or ACK'ed ISS match? */
1633 if ((ack & (~0x7f)) != (md5_buffer[0] << 7))
1636 /* Does the digest part of our reflected timestamp match? */
1637 if (to->to_flags & TOF_TS) {
1638 data = md5_buffer[3] ^ to->to_tsecr;
1639 if ((data & (~0x3ff)) != (md5_buffer[2] << 10))
1643 /* Fill in the syncache values. */
1644 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1645 sc->sc_ipopts = NULL;
1651 if (inc->inc_isipv6) {
1652 if (sotoinpcb(so)->in6p_flags & IN6P_AUTOFLOWLABEL)
1653 sc->sc_flowlabel = md5_buffer[1] & IPV6_FLOWLABEL_MASK;
1657 sc->sc_ip_ttl = sotoinpcb(so)->inp_ip_ttl;
1658 sc->sc_ip_tos = sotoinpcb(so)->inp_ip_tos;
1661 /* Additional parameters that were encoded in the timestamp. */
1663 sc->sc_flags |= SCF_TIMESTAMP;
1664 sc->sc_tsreflect = to->to_tsval;
1665 sc->sc_ts = to->to_tsecr;
1666 sc->sc_tsoff = to->to_tsecr - ticks;
1667 sc->sc_flags |= (data & 0x1) ? SCF_SIGNATURE : 0;
1668 sc->sc_flags |= ((data >> 1) & 0x1) ? SCF_SACK : 0;
1669 sc->sc_requested_s_scale = min((data >> 2) & 0xf,
1671 sc->sc_requested_r_scale = min((data >> 6) & 0xf,
1673 if (sc->sc_requested_s_scale || sc->sc_requested_r_scale)
1674 sc->sc_flags |= SCF_WINSCALE;
1676 sc->sc_flags |= SCF_NOOPT;
1678 wnd = sbspace(&so->so_rcv);
1680 wnd = imin(wnd, TCP_MAXWIN);
1684 sc->sc_peer_mss = tcp_sc_msstab[mss];
1686 tcpstat.tcps_sc_recvcookie++;
1691 * Returns the current number of syncache entries. This number
1692 * will probably change before you get around to calling
1697 syncache_pcbcount(void)
1699 struct syncache_head *sch;
1702 for (count = 0, i = 0; i < tcp_syncache.hashsize; i++) {
1703 /* No need to lock for a read. */
1704 sch = &tcp_syncache.hashbase[i];
1705 count += sch->sch_length;
1711 * Exports the syncache entries to userland so that netstat can display
1712 * them alongside the other sockets. This function is intended to be
1713 * called only from tcp_pcblist.
1715 * Due to concurrency on an active system, the number of pcbs exported
1716 * may have no relation to max_pcbs. max_pcbs merely indicates the
1717 * amount of space the caller allocated for this function to use.
1720 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
1723 struct syncache *sc;
1724 struct syncache_head *sch;
1725 int count, error, i;
1727 for (count = 0, error = 0, i = 0; i < tcp_syncache.hashsize; i++) {
1728 sch = &tcp_syncache.hashbase[i];
1730 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
1731 if (count >= max_pcbs) {
1735 bzero(&xt, sizeof(xt));
1736 xt.xt_len = sizeof(xt);
1737 if (sc->sc_inc.inc_isipv6)
1738 xt.xt_inp.inp_vflag = INP_IPV6;
1740 xt.xt_inp.inp_vflag = INP_IPV4;
1741 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo));
1742 xt.xt_tp.t_inpcb = &xt.xt_inp;
1743 xt.xt_tp.t_state = TCPS_SYN_RECEIVED;
1744 xt.xt_socket.xso_protocol = IPPROTO_TCP;
1745 xt.xt_socket.xso_len = sizeof (struct xsocket);
1746 xt.xt_socket.so_type = SOCK_STREAM;
1747 xt.xt_socket.so_state = SS_ISCONNECTING;
1748 error = SYSCTL_OUT(req, &xt, sizeof xt);
1758 *pcbs_exported = count;