2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2001 McAfee, Inc.
5 * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG
8 * This software was developed for the FreeBSD Project by Jonathan Lemon
9 * and McAfee Research, the Security Research Division of McAfee, Inc. under
10 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
11 * DARPA CHATS research program. [2001 McAfee, Inc.]
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
16 * 1. Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution.
22 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
39 #include "opt_inet6.h"
40 #include "opt_ipsec.h"
42 #include <sys/param.h>
43 #include <sys/systm.h>
45 #include <sys/refcount.h>
46 #include <sys/kernel.h>
47 #include <sys/sysctl.h>
48 #include <sys/limits.h>
50 #include <sys/mutex.h>
51 #include <sys/malloc.h>
53 #include <sys/proc.h> /* for proc0 declaration */
54 #include <sys/random.h>
55 #include <sys/socket.h>
56 #include <sys/socketvar.h>
57 #include <sys/syslog.h>
58 #include <sys/ucred.h>
61 #include <crypto/siphash/siphash.h>
66 #include <net/if_var.h>
67 #include <net/route.h>
70 #include <netinet/in.h>
71 #include <netinet/in_kdtrace.h>
72 #include <netinet/in_systm.h>
73 #include <netinet/ip.h>
74 #include <netinet/in_var.h>
75 #include <netinet/in_pcb.h>
76 #include <netinet/ip_var.h>
77 #include <netinet/ip_options.h>
79 #include <netinet/ip6.h>
80 #include <netinet/icmp6.h>
81 #include <netinet6/nd6.h>
82 #include <netinet6/ip6_var.h>
83 #include <netinet6/in6_pcb.h>
85 #include <netinet/tcp.h>
86 #include <netinet/tcp_fastopen.h>
87 #include <netinet/tcp_fsm.h>
88 #include <netinet/tcp_seq.h>
89 #include <netinet/tcp_timer.h>
90 #include <netinet/tcp_var.h>
91 #include <netinet/tcp_syncache.h>
92 #include <netinet/tcp_ecn.h>
94 #include <netinet/toecore.h>
96 #include <netinet/udp.h>
98 #include <netipsec/ipsec_support.h>
100 #include <machine/in_cksum.h>
102 #include <security/mac/mac_framework.h>
104 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1;
105 #define V_tcp_syncookies VNET(tcp_syncookies)
106 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
107 &VNET_NAME(tcp_syncookies), 0,
108 "Use TCP SYN cookies if the syncache overflows");
110 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0;
111 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly)
112 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
113 &VNET_NAME(tcp_syncookiesonly), 0,
114 "Use only TCP SYN cookies");
116 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1;
117 #define V_functions_inherit_listen_socket_stack \
118 VNET(functions_inherit_listen_socket_stack)
119 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack,
120 CTLFLAG_VNET | CTLFLAG_RW,
121 &VNET_NAME(functions_inherit_listen_socket_stack), 0,
122 "Inherit listen socket's stack");
125 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
128 static void syncache_drop(struct syncache *, struct syncache_head *);
129 static void syncache_free(struct syncache *);
130 static void syncache_insert(struct syncache *, struct syncache_head *);
131 static int syncache_respond(struct syncache *, const struct mbuf *, int);
132 static struct socket *syncache_socket(struct syncache *, struct socket *,
134 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch,
136 static void syncache_timer(void *);
138 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
139 uint8_t *, uintptr_t);
140 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *);
141 static struct syncache
142 *syncookie_lookup(struct in_conninfo *, struct syncache_head *,
143 struct syncache *, struct tcphdr *, struct tcpopt *,
144 struct socket *, uint16_t);
145 static void syncache_pause(struct in_conninfo *);
146 static void syncache_unpause(void *);
147 static void syncookie_reseed(void *);
149 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
150 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
151 struct socket *lso, uint16_t port);
155 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
156 * 3 retransmits corresponds to a timeout with default values of
157 * tcp_rexmit_initial * ( 1 +
160 * tcp_backoff[3]) + 3 * tcp_rexmit_slop,
161 * 1000 ms * (1 + 2 + 4 + 8) + 3 * 200 ms = 15600 ms,
162 * the odds are that the user has given up attempting to connect by then.
164 #define SYNCACHE_MAXREXMTS 3
166 /* Arbitrary values */
167 #define TCP_SYNCACHE_HASHSIZE 512
168 #define TCP_SYNCACHE_BUCKETLIMIT 30
170 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
171 #define V_tcp_syncache VNET(tcp_syncache)
173 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache,
174 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
177 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
178 &VNET_NAME(tcp_syncache.bucket_limit), 0,
179 "Per-bucket hash limit for syncache");
181 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
182 &VNET_NAME(tcp_syncache.cache_limit), 0,
183 "Overall entry limit for syncache");
185 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
186 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
188 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
189 &VNET_NAME(tcp_syncache.hashsize), 0,
190 "Size of TCP syncache hashtable");
192 SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET |
193 CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0,
194 "All syncache(4) entries are visible, ignoring UID/GID, jail(2) "
195 "and mac(4) checks");
198 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
203 new = V_tcp_syncache.rexmt_limit;
204 error = sysctl_handle_int(oidp, &new, 0, req);
205 if ((error == 0) && (req->newptr != NULL)) {
206 if (new > TCP_MAXRXTSHIFT)
209 V_tcp_syncache.rexmt_limit = new;
214 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
215 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
216 &VNET_NAME(tcp_syncache.rexmt_limit), 0,
217 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
218 "Limit on SYN/ACK retransmissions");
220 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
221 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
222 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
223 "Send reset on socket allocation failure");
225 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
227 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx)
228 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx)
229 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED)
232 * Requires the syncache entry to be already removed from the bucket list.
235 syncache_free(struct syncache *sc)
239 (void) m_free(sc->sc_ipopts);
243 mac_syncache_destroy(&sc->sc_label);
246 uma_zfree(V_tcp_syncache.zone, sc);
254 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
255 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
256 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
257 V_tcp_syncache.hash_secret = arc4random();
259 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
260 &V_tcp_syncache.hashsize);
261 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
262 &V_tcp_syncache.bucket_limit);
263 if (!powerof2(V_tcp_syncache.hashsize) ||
264 V_tcp_syncache.hashsize == 0) {
265 printf("WARNING: syncache hash size is not a power of 2.\n");
266 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
268 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
271 V_tcp_syncache.cache_limit =
272 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
273 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
274 &V_tcp_syncache.cache_limit);
276 /* Allocate the hash table. */
277 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
278 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
281 V_tcp_syncache.vnet = curvnet;
284 /* Initialize the hash buckets. */
285 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
286 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
287 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
289 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
290 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
291 V_tcp_syncache.hashbase[i].sch_length = 0;
292 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
293 V_tcp_syncache.hashbase[i].sch_last_overflow =
294 -(SYNCOOKIE_LIFETIME + 1);
297 /* Create the syncache entry zone. */
298 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
299 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
300 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
301 V_tcp_syncache.cache_limit);
303 /* Start the SYN cookie reseeder callout. */
304 callout_init(&V_tcp_syncache.secret.reseed, 1);
305 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
306 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
307 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
308 syncookie_reseed, &V_tcp_syncache);
310 /* Initialize the pause machinery. */
311 mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF);
312 callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx,
314 V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME;
315 V_tcp_syncache.pause_backoff = 0;
316 V_tcp_syncache.paused = false;
321 syncache_destroy(void)
323 struct syncache_head *sch;
324 struct syncache *sc, *nsc;
328 * Stop the re-seed timer before freeing resources. No need to
329 * possibly schedule it another time.
331 callout_drain(&V_tcp_syncache.secret.reseed);
333 /* Stop the SYN cache pause callout. */
334 mtx_lock(&V_tcp_syncache.pause_mtx);
335 if (callout_stop(&V_tcp_syncache.pause_co) == 0) {
336 mtx_unlock(&V_tcp_syncache.pause_mtx);
337 callout_drain(&V_tcp_syncache.pause_co);
339 mtx_unlock(&V_tcp_syncache.pause_mtx);
341 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */
342 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
343 sch = &V_tcp_syncache.hashbase[i];
344 callout_drain(&sch->sch_timer);
347 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
348 syncache_drop(sc, sch);
350 KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
351 ("%s: sch->sch_bucket not empty", __func__));
352 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
353 __func__, sch->sch_length));
354 mtx_destroy(&sch->sch_mtx);
357 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
358 ("%s: cache_count not 0", __func__));
360 /* Free the allocated global resources. */
361 uma_zdestroy(V_tcp_syncache.zone);
362 free(V_tcp_syncache.hashbase, M_SYNCACHE);
363 mtx_destroy(&V_tcp_syncache.pause_mtx);
368 * Inserts a syncache entry into the specified bucket row.
369 * Locks and unlocks the syncache_head autonomously.
372 syncache_insert(struct syncache *sc, struct syncache_head *sch)
374 struct syncache *sc2;
379 * Make sure that we don't overflow the per-bucket limit.
380 * If the bucket is full, toss the oldest element.
382 if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
383 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
384 ("sch->sch_length incorrect"));
385 syncache_pause(&sc->sc_inc);
386 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
387 sch->sch_last_overflow = time_uptime;
388 syncache_drop(sc2, sch);
391 /* Put it into the bucket. */
392 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
396 if (ADDED_BY_TOE(sc)) {
397 struct toedev *tod = sc->sc_tod;
399 tod->tod_syncache_added(tod, sc->sc_todctx);
403 /* Reinitialize the bucket row's timer. */
404 if (sch->sch_length == 1)
405 sch->sch_nextc = ticks + INT_MAX;
406 syncache_timeout(sc, sch, 1);
410 TCPSTATES_INC(TCPS_SYN_RECEIVED);
411 TCPSTAT_INC(tcps_sc_added);
415 * Remove and free entry from syncache bucket row.
416 * Expects locked syncache head.
419 syncache_drop(struct syncache *sc, struct syncache_head *sch)
422 SCH_LOCK_ASSERT(sch);
424 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
425 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
429 if (ADDED_BY_TOE(sc)) {
430 struct toedev *tod = sc->sc_tod;
432 tod->tod_syncache_removed(tod, sc->sc_todctx);
440 * Engage/reengage time on bucket row.
443 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
447 if (sc->sc_rxmits == 0)
448 rexmt = tcp_rexmit_initial;
451 tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
452 tcp_rexmit_min, TCPTV_REXMTMAX);
453 sc->sc_rxttime = ticks + rexmt;
455 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
456 sch->sch_nextc = sc->sc_rxttime;
458 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
459 syncache_timer, (void *)sch);
464 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
465 * If we have retransmitted an entry the maximum number of times, expire it.
466 * One separate timer for each bucket row.
469 syncache_timer(void *xsch)
471 struct syncache_head *sch = (struct syncache_head *)xsch;
472 struct syncache *sc, *nsc;
473 struct epoch_tracker et;
478 CURVNET_SET(sch->sch_sc->vnet);
480 /* NB: syncache_head has already been locked by the callout. */
481 SCH_LOCK_ASSERT(sch);
484 * In the following cycle we may remove some entries and/or
485 * advance some timeouts, so re-initialize the bucket timer.
487 sch->sch_nextc = tick + INT_MAX;
490 * If we have paused processing, unconditionally remove
491 * all syncache entries.
493 mtx_lock(&V_tcp_syncache.pause_mtx);
494 paused = V_tcp_syncache.paused;
495 mtx_unlock(&V_tcp_syncache.pause_mtx);
497 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
499 syncache_drop(sc, sch);
503 * We do not check if the listen socket still exists
504 * and accept the case where the listen socket may be
505 * gone by the time we resend the SYN/ACK. We do
506 * not expect this to happens often. If it does,
507 * then the RST will be sent by the time the remote
508 * host does the SYN/ACK->ACK.
510 if (TSTMP_GT(sc->sc_rxttime, tick)) {
511 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
512 sch->sch_nextc = sc->sc_rxttime;
515 if (sc->sc_rxmits > V_tcp_ecn_maxretries) {
516 sc->sc_flags &= ~SCF_ECN_MASK;
518 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
519 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
520 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
521 "giving up and removing syncache entry\n",
525 syncache_drop(sc, sch);
526 TCPSTAT_INC(tcps_sc_stale);
529 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
530 log(LOG_DEBUG, "%s; %s: Response timeout, "
531 "retransmitting (%u) SYN|ACK\n",
532 s, __func__, sc->sc_rxmits);
537 syncache_respond(sc, NULL, TH_SYN|TH_ACK);
539 TCPSTAT_INC(tcps_sc_retransmitted);
540 syncache_timeout(sc, sch, 0);
542 if (!TAILQ_EMPTY(&(sch)->sch_bucket))
543 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
544 syncache_timer, (void *)(sch));
549 * Returns true if the system is only using cookies at the moment.
550 * This could be due to a sysadmin decision to only use cookies, or it
551 * could be due to the system detecting an attack.
554 syncache_cookiesonly(void)
557 return (V_tcp_syncookies && (V_tcp_syncache.paused ||
558 V_tcp_syncookiesonly));
562 * Find the hash bucket for the given connection.
564 static struct syncache_head *
565 syncache_hashbucket(struct in_conninfo *inc)
570 * The hash is built on foreign port + local port + foreign address.
571 * We rely on the fact that struct in_conninfo starts with 16 bits
572 * of foreign port, then 16 bits of local port then followed by 128
573 * bits of foreign address. In case of IPv4 address, the first 3
574 * 32-bit words of the address always are zeroes.
576 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
577 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
579 return (&V_tcp_syncache.hashbase[hash]);
583 * Find an entry in the syncache.
584 * Returns always with locked syncache_head plus a matching entry or NULL.
586 static struct syncache *
587 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
590 struct syncache_head *sch;
592 *schp = sch = syncache_hashbucket(inc);
595 /* Circle through bucket row to find matching entry. */
596 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
597 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
598 sizeof(struct in_endpoints)) == 0)
601 return (sc); /* Always returns with locked sch. */
605 * This function is called when we get a RST for a
606 * non-existent connection, so that we can see if the
607 * connection is in the syn cache. If it is, zap it.
608 * If required send a challenge ACK.
611 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m,
615 struct syncache_head *sch;
618 if (syncache_cookiesonly())
620 sc = syncache_lookup(inc, &sch); /* returns locked sch */
621 SCH_LOCK_ASSERT(sch);
624 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
625 * See RFC 793 page 65, section SEGMENT ARRIVES.
627 if (tcp_get_flags(th) & (TH_ACK|TH_SYN|TH_FIN)) {
628 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
629 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
630 "FIN flag set, segment ignored\n", s, __func__);
631 TCPSTAT_INC(tcps_badrst);
636 * No corresponding connection was found in syncache.
637 * If syncookies are enabled and possibly exclusively
638 * used, or we are under memory pressure, a valid RST
639 * may not find a syncache entry. In that case we're
640 * done and no SYN|ACK retransmissions will happen.
641 * Otherwise the RST was misdirected or spoofed.
644 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
645 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
646 "syncache entry (possibly syncookie only), "
647 "segment ignored\n", s, __func__);
648 TCPSTAT_INC(tcps_badrst);
652 /* The remote UDP encaps port does not match. */
653 if (sc->sc_port != port) {
654 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
655 log(LOG_DEBUG, "%s; %s: Spurious RST with matching "
656 "syncache entry but non-matching UDP encaps port, "
657 "segment ignored\n", s, __func__);
658 TCPSTAT_INC(tcps_badrst);
663 * If the RST bit is set, check the sequence number to see
664 * if this is a valid reset segment.
667 * In all states except SYN-SENT, all reset (RST) segments
668 * are validated by checking their SEQ-fields. A reset is
669 * valid if its sequence number is in the window.
672 * There are four cases for the acceptability test for an incoming
675 * Segment Receive Test
677 * ------- ------- -------------------------------------------
678 * 0 0 SEG.SEQ = RCV.NXT
679 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
680 * >0 0 not acceptable
681 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
682 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
684 * Note that when receiving a SYN segment in the LISTEN state,
685 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
686 * described in RFC 793, page 66.
688 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
689 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
690 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
691 if (V_tcp_insecure_rst ||
692 th->th_seq == sc->sc_irs + 1) {
693 syncache_drop(sc, sch);
694 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
696 "%s; %s: Our SYN|ACK was rejected, "
697 "connection attempt aborted by remote "
700 TCPSTAT_INC(tcps_sc_reset);
702 TCPSTAT_INC(tcps_badrst);
703 /* Send challenge ACK. */
704 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
705 log(LOG_DEBUG, "%s; %s: RST with invalid "
706 " SEQ %u != NXT %u (+WND %u), "
707 "sending challenge ACK\n",
709 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
710 syncache_respond(sc, m, TH_ACK);
713 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
714 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
715 "NXT %u (+WND %u), segment ignored\n",
717 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
718 TCPSTAT_INC(tcps_badrst);
728 syncache_badack(struct in_conninfo *inc, uint16_t port)
731 struct syncache_head *sch;
733 if (syncache_cookiesonly())
735 sc = syncache_lookup(inc, &sch); /* returns locked sch */
736 SCH_LOCK_ASSERT(sch);
737 if ((sc != NULL) && (sc->sc_port == port)) {
738 syncache_drop(sc, sch);
739 TCPSTAT_INC(tcps_sc_badack);
745 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port)
748 struct syncache_head *sch;
750 if (syncache_cookiesonly())
752 sc = syncache_lookup(inc, &sch); /* returns locked sch */
753 SCH_LOCK_ASSERT(sch);
757 /* If the port != sc_port, then it's a bogus ICMP msg */
758 if (port != sc->sc_port)
761 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
762 if (ntohl(th_seq) != sc->sc_iss)
766 * If we've rertransmitted 3 times and this is our second error,
767 * we remove the entry. Otherwise, we allow it to continue on.
768 * This prevents us from incorrectly nuking an entry during a
769 * spurious network outage.
773 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
774 sc->sc_flags |= SCF_UNREACH;
777 syncache_drop(sc, sch);
778 TCPSTAT_INC(tcps_sc_unreach);
784 * Build a new TCP socket structure from a syncache entry.
786 * On success return the newly created socket with its underlying inp locked.
788 static struct socket *
789 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
791 struct tcp_function_block *blk;
792 struct inpcb *inp = NULL;
801 * Ok, create the full blown connection, and set things up
802 * as they would have been set up if we had created the
803 * connection when the SYN arrived.
805 if ((so = solisten_clone(lso)) == NULL)
808 mac_socketpeer_set_from_mbuf(m, so);
810 error = in_pcballoc(so, &V_tcbinfo);
816 if ((tp = tcp_newtcpcb(inp)) == NULL) {
822 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
824 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
825 inp->inp_vflag &= ~INP_IPV4;
826 inp->inp_vflag |= INP_IPV6;
827 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
829 inp->inp_vflag &= ~INP_IPV6;
830 inp->inp_vflag |= INP_IPV4;
832 inp->inp_ip_ttl = sc->sc_ip_ttl;
833 inp->inp_ip_tos = sc->sc_ip_tos;
834 inp->inp_laddr = sc->sc_inc.inc_laddr;
840 * If there's an mbuf and it has a flowid, then let's initialise the
841 * inp with that particular flowid.
843 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
844 inp->inp_flowid = m->m_pkthdr.flowid;
845 inp->inp_flowtype = M_HASHTYPE_GET(m);
847 inp->inp_numa_domain = m->m_pkthdr.numa_domain;
851 inp->inp_lport = sc->sc_inc.inc_lport;
853 if (inp->inp_vflag & INP_IPV6PROTO) {
854 struct inpcb *oinp = sotoinpcb(lso);
857 * Inherit socket options from the listening socket.
858 * Note that in6p_inputopts are not (and should not be)
859 * copied, since it stores previously received options and is
860 * used to detect if each new option is different than the
861 * previous one and hence should be passed to a user.
862 * If we copied in6p_inputopts, a user would not be able to
863 * receive options just after calling the accept system call.
865 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
866 if (oinp->in6p_outputopts)
867 inp->in6p_outputopts =
868 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
869 inp->in6p_hops = oinp->in6p_hops;
872 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
873 struct sockaddr_in6 sin6;
875 sin6.sin6_family = AF_INET6;
876 sin6.sin6_len = sizeof(sin6);
877 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
878 sin6.sin6_port = sc->sc_inc.inc_fport;
879 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
880 INP_HASH_WLOCK(&V_tcbinfo);
881 error = in6_pcbconnect(inp, &sin6, thread0.td_ucred, false);
882 INP_HASH_WUNLOCK(&V_tcbinfo);
885 /* Override flowlabel from in6_pcbconnect. */
886 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
887 inp->inp_flow |= sc->sc_flowlabel;
890 #if defined(INET) && defined(INET6)
895 struct sockaddr_in sin;
897 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
899 if (inp->inp_options == NULL) {
900 inp->inp_options = sc->sc_ipopts;
901 sc->sc_ipopts = NULL;
904 sin.sin_family = AF_INET;
905 sin.sin_len = sizeof(sin);
906 sin.sin_addr = sc->sc_inc.inc_faddr;
907 sin.sin_port = sc->sc_inc.inc_fport;
908 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
909 INP_HASH_WLOCK(&V_tcbinfo);
910 error = in_pcbconnect(inp, &sin, thread0.td_ucred, false);
911 INP_HASH_WUNLOCK(&V_tcbinfo);
916 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
917 /* Copy old policy into new socket's. */
918 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
919 printf("syncache_socket: could not copy policy\n");
921 tp->t_state = TCPS_SYN_RECEIVED;
922 tp->iss = sc->sc_iss;
923 tp->irs = sc->sc_irs;
924 tp->t_port = sc->sc_port;
927 blk = sototcpcb(lso)->t_fb;
928 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
930 * Our parents t_fb was not the default,
931 * we need to release our ref on tp->t_fb and
932 * pickup one on the new entry.
934 struct tcp_function_block *rblk;
937 rblk = find_and_ref_tcp_fb(blk);
938 KASSERT(rblk != NULL,
939 ("cannot find blk %p out of syncache?", blk));
941 if (rblk->tfb_tcp_fb_init == NULL ||
942 (*rblk->tfb_tcp_fb_init)(tp, &ptr) == 0) {
943 /* Release the old stack */
944 if (tp->t_fb->tfb_tcp_fb_fini != NULL)
945 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
946 refcount_release(&tp->t_fb->tfb_refcnt);
947 /* Now set in all the pointers */
952 * Initialization failed. Release the reference count on
953 * the looked up default stack.
955 refcount_release(&rblk->tfb_refcnt);
958 tp->snd_wl1 = sc->sc_irs;
959 tp->snd_max = tp->iss + 1;
960 tp->snd_nxt = tp->iss + 1;
961 tp->rcv_up = sc->sc_irs + 1;
962 tp->rcv_wnd = sc->sc_wnd;
963 tp->rcv_adv += tp->rcv_wnd;
964 tp->last_ack_sent = tp->rcv_nxt;
966 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
967 if (sc->sc_flags & SCF_NOOPT)
968 tp->t_flags |= TF_NOOPT;
970 if (sc->sc_flags & SCF_WINSCALE) {
971 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
972 tp->snd_scale = sc->sc_requested_s_scale;
973 tp->request_r_scale = sc->sc_requested_r_scale;
975 if (sc->sc_flags & SCF_TIMESTAMP) {
976 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
977 tp->ts_recent = sc->sc_tsreflect;
978 tp->ts_recent_age = tcp_ts_getticks();
979 tp->ts_offset = sc->sc_tsoff;
981 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
982 if (sc->sc_flags & SCF_SIGNATURE)
983 tp->t_flags |= TF_SIGNATURE;
985 if (sc->sc_flags & SCF_SACK)
986 tp->t_flags |= TF_SACK_PERMIT;
989 tcp_ecn_syncache_socket(tp, sc);
992 * Set up MSS and get cached values from tcp_hostcache.
993 * This might overwrite some of the defaults we just set.
995 tcp_mss(tp, sc->sc_peer_mss);
998 * If the SYN,ACK was retransmitted, indicate that CWND to be
999 * limited to one segment in cc_conn_init().
1000 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
1002 if (sc->sc_rxmits > 1)
1007 * Allow a TOE driver to install its hooks. Note that we hold the
1008 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
1009 * new connection before the TOE driver has done its thing.
1011 if (ADDED_BY_TOE(sc)) {
1012 struct toedev *tod = sc->sc_tod;
1014 tod->tod_offload_socket(tod, sc->sc_todctx, so);
1018 * Copy and activate timers.
1020 tp->t_maxunacktime = sototcpcb(lso)->t_maxunacktime;
1021 tp->t_keepinit = sototcpcb(lso)->t_keepinit;
1022 tp->t_keepidle = sototcpcb(lso)->t_keepidle;
1023 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
1024 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1025 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1027 TCPSTAT_INC(tcps_accepts);
1028 TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, TCPS_LISTEN);
1030 if (!solisten_enqueue(so, SS_ISCONNECTED))
1031 tp->t_flags |= TF_SONOTCONN;
1037 * Drop the connection; we will either send a RST or have the peer
1038 * retransmit its SYN again after its RTO and try again.
1040 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1041 log(LOG_DEBUG, "%s; %s: Socket create failed "
1042 "due to limits or memory shortage\n",
1046 TCPSTAT_INC(tcps_listendrop);
1053 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1054 log(LOG_DEBUG, "%s; %s: in%s_pcbconnect failed with error %i\n",
1055 s, __func__, (sc->sc_inc.inc_flags & INC_ISIPV6) ? "6" : "",
1059 TCPSTAT_INC(tcps_listendrop);
1064 * This function gets called when we receive an ACK for a
1065 * socket in the LISTEN state. We look up the connection
1066 * in the syncache, and if its there, we pull it out of
1067 * the cache and turn it into a full-blown connection in
1068 * the SYN-RECEIVED state.
1070 * On syncache_socket() success the newly created socket
1071 * has its underlying inp locked.
1074 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1075 struct socket **lsop, struct mbuf *m, uint16_t port)
1077 struct syncache *sc;
1078 struct syncache_head *sch;
1079 struct syncache scs;
1084 KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1085 ("%s: can handle only ACK", __func__));
1087 if (syncache_cookiesonly()) {
1089 sch = syncache_hashbucket(inc);
1092 sc = syncache_lookup(inc, &sch); /* returns locked sch */
1094 SCH_LOCK_ASSERT(sch);
1099 * Test code for syncookies comparing the syncache stored
1100 * values with the reconstructed values from the cookie.
1103 syncookie_cmp(inc, sch, sc, th, to, *lsop, port);
1108 * There is no syncache entry, so see if this ACK is
1109 * a returning syncookie. To do this, first:
1110 * A. Check if syncookies are used in case of syncache
1112 * B. See if this socket has had a syncache entry dropped in
1113 * the recent past. We don't want to accept a bogus
1114 * syncookie if we've never received a SYN or accept it
1116 * C. check that the syncookie is valid. If it is, then
1117 * cobble up a fake syncache entry, and return.
1119 if (locked && !V_tcp_syncookies) {
1121 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1122 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1123 "segment rejected (syncookies disabled)\n",
1127 if (locked && !V_tcp_syncookiesonly &&
1128 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1130 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1131 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1132 "segment rejected (no syncache entry)\n",
1136 bzero(&scs, sizeof(scs));
1137 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port);
1141 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1142 log(LOG_DEBUG, "%s; %s: Segment failed "
1143 "SYNCOOKIE authentication, segment rejected "
1144 "(probably spoofed)\n", s, __func__);
1147 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1148 /* If received ACK has MD5 signature, check it. */
1149 if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1150 (!TCPMD5_ENABLED() ||
1151 TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1153 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1154 log(LOG_DEBUG, "%s; %s: Segment rejected, "
1155 "MD5 signature doesn't match.\n",
1159 TCPSTAT_INC(tcps_sig_err_sigopt);
1160 return (-1); /* Do not send RST */
1162 #endif /* TCP_SIGNATURE */
1163 TCPSTATES_INC(TCPS_SYN_RECEIVED);
1165 if (sc->sc_port != port) {
1169 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1171 * If listening socket requested TCP digests, check that
1172 * received ACK has signature and it is correct.
1173 * If not, drop the ACK and leave sc entry in th cache,
1174 * because SYN was received with correct signature.
1176 if (sc->sc_flags & SCF_SIGNATURE) {
1177 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1179 TCPSTAT_INC(tcps_sig_err_nosigopt);
1181 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1182 log(LOG_DEBUG, "%s; %s: Segment "
1183 "rejected, MD5 signature wasn't "
1184 "provided.\n", s, __func__);
1187 return (-1); /* Do not send RST */
1189 if (!TCPMD5_ENABLED() ||
1190 TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1191 /* Doesn't match or no SA */
1193 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1194 log(LOG_DEBUG, "%s; %s: Segment "
1195 "rejected, MD5 signature doesn't "
1196 "match.\n", s, __func__);
1199 return (-1); /* Do not send RST */
1202 #endif /* TCP_SIGNATURE */
1205 * RFC 7323 PAWS: If we have a timestamp on this segment and
1206 * it's less than ts_recent, drop it.
1207 * XXXMT: RFC 7323 also requires to send an ACK.
1208 * In tcp_input.c this is only done for TCP segments
1209 * with user data, so be consistent here and just drop
1212 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1213 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1215 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1217 "%s; %s: SEG.TSval %u < TS.Recent %u, "
1218 "segment dropped\n", s, __func__,
1219 to->to_tsval, sc->sc_tsreflect);
1222 return (-1); /* Do not send RST */
1226 * If timestamps were not negotiated during SYN/ACK and a
1227 * segment with a timestamp is received, ignore the
1228 * timestamp and process the packet normally.
1229 * See section 3.2 of RFC 7323.
1231 if (!(sc->sc_flags & SCF_TIMESTAMP) &&
1232 (to->to_flags & TOF_TS)) {
1233 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1234 log(LOG_DEBUG, "%s; %s: Timestamp not "
1235 "expected, segment processed normally\n",
1243 * If timestamps were negotiated during SYN/ACK and a
1244 * segment without a timestamp is received, silently drop
1245 * the segment, unless the missing timestamps are tolerated.
1246 * See section 3.2 of RFC 7323.
1248 if ((sc->sc_flags & SCF_TIMESTAMP) &&
1249 !(to->to_flags & TOF_TS)) {
1250 if (V_tcp_tolerate_missing_ts) {
1251 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1253 "%s; %s: Timestamp missing, "
1254 "segment processed normally\n",
1260 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1262 "%s; %s: Timestamp missing, "
1263 "segment silently dropped\n",
1267 return (-1); /* Do not send RST */
1270 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1273 if (ADDED_BY_TOE(sc)) {
1274 struct toedev *tod = sc->sc_tod;
1276 tod->tod_syncache_removed(tod, sc->sc_todctx);
1283 * Segment validation:
1284 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1286 if (th->th_ack != sc->sc_iss + 1) {
1287 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1288 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1289 "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1294 * The SEQ must fall in the window starting at the received
1295 * initial receive sequence number + 1 (the SYN).
1297 if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1298 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1299 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1300 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1301 "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1305 *lsop = syncache_socket(sc, *lsop, m);
1307 if (__predict_false(*lsop == NULL)) {
1308 TCPSTAT_INC(tcps_sc_aborted);
1309 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1311 TCPSTAT_INC(tcps_sc_completed);
1313 /* how do we find the inp for the new socket? */
1319 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1329 static struct socket *
1330 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m,
1331 uint64_t response_cookie)
1335 unsigned int *pending_counter;
1340 pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending;
1341 so = syncache_socket(sc, lso, m);
1343 TCPSTAT_INC(tcps_sc_aborted);
1344 atomic_subtract_int(pending_counter, 1);
1347 inp = sotoinpcb(so);
1348 tp = intotcpcb(inp);
1349 tp->t_flags |= TF_FASTOPEN;
1350 tp->t_tfo_cookie.server = response_cookie;
1351 tp->snd_max = tp->iss;
1352 tp->snd_nxt = tp->iss;
1353 tp->t_tfo_pending = pending_counter;
1354 TCPSTATES_INC(TCPS_SYN_RECEIVED);
1355 TCPSTAT_INC(tcps_sc_completed);
1362 * Given a LISTEN socket and an inbound SYN request, add
1363 * this to the syn cache, and send back a segment:
1364 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1367 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1368 * Doing so would require that we hold onto the data and deliver it
1369 * to the application. However, if we are the target of a SYN-flood
1370 * DoS attack, an attacker could send data which would eventually
1371 * consume all available buffer space if it were ACKed. By not ACKing
1372 * the data, we avoid this DoS scenario.
1374 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1375 * cookie is processed and a new socket is created. In this case, any data
1376 * accompanying the SYN will be queued to the socket by tcp_input() and will
1377 * be ACKed either when the application sends response data or the delayed
1378 * ACK timer expires, whichever comes first.
1381 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1382 struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod,
1383 void *todctx, uint8_t iptos, uint16_t port)
1386 struct socket *rv = NULL;
1387 struct syncache *sc = NULL;
1388 struct syncache_head *sch;
1389 struct mbuf *ipopts = NULL;
1391 int win, ip_ttl, ip_tos;
1394 int autoflowlabel = 0;
1397 struct label *maclabel;
1399 struct syncache scs;
1401 uint64_t tfo_response_cookie;
1402 unsigned int *tfo_pending = NULL;
1403 int tfo_cookie_valid = 0;
1404 int tfo_response_cookie_valid = 0;
1407 INP_RLOCK_ASSERT(inp); /* listen socket */
1408 KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1409 ("%s: unexpected tcp flags", __func__));
1412 * Combine all so/tp operations very early to drop the INP lock as
1415 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1417 cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred);
1420 if (inc->inc_flags & INC_ISIPV6) {
1421 if (inp->inp_flags & IN6P_AUTOFLOWLABEL) {
1424 ip_ttl = in6_selecthlim(inp, NULL);
1425 if ((inp->in6p_outputopts == NULL) ||
1426 (inp->in6p_outputopts->ip6po_tclass == -1)) {
1429 ip_tos = inp->in6p_outputopts->ip6po_tclass;
1433 #if defined(INET6) && defined(INET)
1438 ip_ttl = inp->inp_ip_ttl;
1439 ip_tos = inp->inp_ip_tos;
1442 win = so->sol_sbrcv_hiwat;
1443 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1445 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
1446 (tp->t_tfo_pending != NULL) &&
1447 (to->to_flags & TOF_FASTOPEN)) {
1449 * Limit the number of pending TFO connections to
1450 * approximately half of the queue limit. This prevents TFO
1451 * SYN floods from starving the service by filling the
1452 * listen queue with bogus TFO connections.
1454 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1455 (so->sol_qlimit / 2)) {
1458 result = tcp_fastopen_check_cookie(inc,
1459 to->to_tfo_cookie, to->to_tfo_len,
1460 &tfo_response_cookie);
1461 tfo_cookie_valid = (result > 0);
1462 tfo_response_cookie_valid = (result >= 0);
1466 * Remember the TFO pending counter as it will have to be
1467 * decremented below if we don't make it to syncache_tfo_expand().
1469 tfo_pending = tp->t_tfo_pending;
1473 if (mac_syncache_init(&maclabel) != 0) {
1477 mac_syncache_create(maclabel, inp);
1479 if (!tfo_cookie_valid)
1483 * Remember the IP options, if any.
1486 if (!(inc->inc_flags & INC_ISIPV6))
1489 ipopts = (m) ? ip_srcroute(m) : NULL;
1494 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1496 * When the socket is TCP-MD5 enabled check that,
1497 * - a signed packet is valid
1498 * - a non-signed packet does not have a security association
1500 * If a signed packet fails validation or a non-signed packet has a
1501 * security association, the packet will be dropped.
1503 if (ltflags & TF_SIGNATURE) {
1504 if (to->to_flags & TOF_SIGNATURE) {
1505 if (!TCPMD5_ENABLED() ||
1506 TCPMD5_INPUT(m, th, to->to_signature) != 0)
1509 if (TCPMD5_ENABLED() &&
1510 TCPMD5_INPUT(m, NULL, NULL) != ENOENT)
1513 } else if (to->to_flags & TOF_SIGNATURE)
1515 #endif /* TCP_SIGNATURE */
1517 * See if we already have an entry for this connection.
1518 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1520 * XXX: should the syncache be re-initialized with the contents
1521 * of the new SYN here (which may have different options?)
1523 * XXX: We do not check the sequence number to see if this is a
1524 * real retransmit or a new connection attempt. The question is
1525 * how to handle such a case; either ignore it as spoofed, or
1526 * drop the current entry and create a new one?
1528 if (syncache_cookiesonly()) {
1530 sch = syncache_hashbucket(inc);
1533 sc = syncache_lookup(inc, &sch); /* returns locked sch */
1535 SCH_LOCK_ASSERT(sch);
1538 if (tfo_cookie_valid)
1540 TCPSTAT_INC(tcps_sc_dupsyn);
1543 * If we were remembering a previous source route,
1544 * forget it and use the new one we've been given.
1547 (void) m_free(sc->sc_ipopts);
1548 sc->sc_ipopts = ipopts;
1551 * Update timestamp if present.
1553 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1554 sc->sc_tsreflect = to->to_tsval;
1556 sc->sc_flags &= ~SCF_TIMESTAMP;
1558 * Adjust ECN response if needed, e.g. different
1559 * IP ECN field, or a fallback by the remote host.
1561 if (sc->sc_flags & SCF_ECN_MASK) {
1562 sc->sc_flags &= ~SCF_ECN_MASK;
1563 sc->sc_flags = tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1567 * Since we have already unconditionally allocated label
1568 * storage, free it up. The syncache entry will already
1569 * have an initialized label we can use.
1571 mac_syncache_destroy(&maclabel);
1573 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1574 /* Retransmit SYN|ACK and reset retransmit count. */
1575 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1576 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1577 "resetting timer and retransmitting SYN|ACK\n",
1581 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1583 syncache_timeout(sc, sch, 1);
1584 TCPSTAT_INC(tcps_sndacks);
1585 TCPSTAT_INC(tcps_sndtotal);
1591 if (tfo_cookie_valid) {
1592 bzero(&scs, sizeof(scs));
1598 * Skip allocating a syncache entry if we are just going to discard
1602 bzero(&scs, sizeof(scs));
1605 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1608 * The zone allocator couldn't provide more entries.
1609 * Treat this as if the cache was full; drop the oldest
1610 * entry and insert the new one.
1612 TCPSTAT_INC(tcps_sc_zonefail);
1613 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1614 sch->sch_last_overflow = time_uptime;
1615 syncache_drop(sc, sch);
1616 syncache_pause(inc);
1618 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1620 if (V_tcp_syncookies) {
1621 bzero(&scs, sizeof(scs));
1625 ("%s: bucket unexpectedly unlocked",
1629 (void) m_free(ipopts);
1636 if (!tfo_cookie_valid && tfo_response_cookie_valid)
1637 sc->sc_tfo_cookie = &tfo_response_cookie;
1640 * Fill in the syncache values.
1643 sc->sc_label = maclabel;
1648 sc->sc_ipopts = ipopts;
1649 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1650 sc->sc_ip_tos = ip_tos;
1651 sc->sc_ip_ttl = ip_ttl;
1654 sc->sc_todctx = todctx;
1656 sc->sc_irs = th->th_seq;
1658 sc->sc_flowlabel = 0;
1661 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1662 * win was derived from socket earlier in the function.
1665 win = imin(win, TCP_MAXWIN);
1668 if (V_tcp_do_rfc1323 &&
1669 !(ltflags & TF_NOOPT)) {
1671 * A timestamp received in a SYN makes
1672 * it ok to send timestamp requests and replies.
1674 if ((to->to_flags & TOF_TS) && (V_tcp_do_rfc1323 != 2)) {
1675 sc->sc_tsreflect = to->to_tsval;
1676 sc->sc_flags |= SCF_TIMESTAMP;
1677 sc->sc_tsoff = tcp_new_ts_offset(inc);
1679 if ((to->to_flags & TOF_SCALE) && (V_tcp_do_rfc1323 != 3)) {
1683 * Pick the smallest possible scaling factor that
1684 * will still allow us to scale up to sb_max, aka
1685 * kern.ipc.maxsockbuf.
1687 * We do this because there are broken firewalls that
1688 * will corrupt the window scale option, leading to
1689 * the other endpoint believing that our advertised
1690 * window is unscaled. At scale factors larger than
1691 * 5 the unscaled window will drop below 1500 bytes,
1692 * leading to serious problems when traversing these
1695 * With the default maxsockbuf of 256K, a scale factor
1696 * of 3 will be chosen by this algorithm. Those who
1697 * choose a larger maxsockbuf should watch out
1698 * for the compatibility problems mentioned above.
1700 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1701 * or <SYN,ACK>) segment itself is never scaled.
1703 while (wscale < TCP_MAX_WINSHIFT &&
1704 (TCP_MAXWIN << wscale) < sb_max)
1706 sc->sc_requested_r_scale = wscale;
1707 sc->sc_requested_s_scale = to->to_wscale;
1708 sc->sc_flags |= SCF_WINSCALE;
1711 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1713 * If incoming packet has an MD5 signature, flag this in the
1714 * syncache so that syncache_respond() will do the right thing
1717 if (to->to_flags & TOF_SIGNATURE)
1718 sc->sc_flags |= SCF_SIGNATURE;
1719 #endif /* TCP_SIGNATURE */
1720 if (to->to_flags & TOF_SACKPERM)
1721 sc->sc_flags |= SCF_SACK;
1722 if (to->to_flags & TOF_MSS)
1723 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */
1724 if (ltflags & TF_NOOPT)
1725 sc->sc_flags |= SCF_NOOPT;
1728 sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1730 if (V_tcp_syncookies)
1731 sc->sc_iss = syncookie_generate(sch, sc);
1733 sc->sc_iss = arc4random();
1735 if (autoflowlabel) {
1736 if (V_tcp_syncookies)
1737 sc->sc_flowlabel = sc->sc_iss;
1739 sc->sc_flowlabel = ip6_randomflowlabel();
1740 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1746 if (tfo_cookie_valid) {
1747 rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie);
1748 /* INP_RUNLOCK(inp) will be performed by the caller */
1752 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1754 * Do a standard 3-way handshake.
1756 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1757 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1759 else if (sc != &scs)
1760 syncache_insert(sc, sch); /* locks and unlocks sch */
1761 TCPSTAT_INC(tcps_sndacks);
1762 TCPSTAT_INC(tcps_sndtotal);
1766 TCPSTAT_INC(tcps_sc_dropped);
1771 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1776 * If tfo_pending is not NULL here, then a TFO SYN that did not
1777 * result in a new socket was processed and the associated pending
1778 * counter has not yet been decremented. All such TFO processing paths
1779 * transit this point.
1781 if (tfo_pending != NULL)
1782 tcp_fastopen_decrement_counter(tfo_pending);
1789 mac_syncache_destroy(&maclabel);
1795 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment,
1796 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1799 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1801 struct ip *ip = NULL;
1803 struct tcphdr *th = NULL;
1804 struct udphdr *udp = NULL;
1805 int optlen, error = 0; /* Make compiler happy */
1806 u_int16_t hlen, tlen, mssopt, ulen;
1809 struct ip6_hdr *ip6 = NULL;
1816 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1819 tlen = hlen + sizeof(struct tcphdr);
1821 tlen += sizeof(struct udphdr);
1823 /* Determine MSS we advertize to other end of connection. */
1824 mssopt = tcp_mssopt(&sc->sc_inc);
1826 mssopt -= V_tcp_udp_tunneling_overhead;
1827 mssopt = max(mssopt, V_tcp_minmss);
1829 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1830 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1831 ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + "
1832 "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port,
1833 max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN));
1835 /* Create the IP+TCP header from scratch. */
1836 m = m_gethdr(M_NOWAIT, MT_DATA);
1840 mac_syncache_create_mbuf(sc->sc_label, m);
1842 m->m_data += max_linkhdr;
1844 m->m_pkthdr.len = tlen;
1845 m->m_pkthdr.rcvif = NULL;
1848 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1849 ip6 = mtod(m, struct ip6_hdr *);
1850 ip6->ip6_vfc = IPV6_VERSION;
1851 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1852 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1853 ip6->ip6_plen = htons(tlen - hlen);
1854 /* ip6_hlim is set after checksum */
1855 /* Zero out traffic class and flow label. */
1856 ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1857 ip6->ip6_flow |= sc->sc_flowlabel;
1858 if (sc->sc_port != 0) {
1859 ip6->ip6_nxt = IPPROTO_UDP;
1860 udp = (struct udphdr *)(ip6 + 1);
1861 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1862 udp->uh_dport = sc->sc_port;
1863 ulen = (tlen - sizeof(struct ip6_hdr));
1864 th = (struct tcphdr *)(udp + 1);
1866 ip6->ip6_nxt = IPPROTO_TCP;
1867 th = (struct tcphdr *)(ip6 + 1);
1869 ip6->ip6_flow |= htonl(sc->sc_ip_tos << 20);
1872 #if defined(INET6) && defined(INET)
1877 ip = mtod(m, struct ip *);
1878 ip->ip_v = IPVERSION;
1879 ip->ip_hl = sizeof(struct ip) >> 2;
1880 ip->ip_len = htons(tlen);
1884 ip->ip_src = sc->sc_inc.inc_laddr;
1885 ip->ip_dst = sc->sc_inc.inc_faddr;
1886 ip->ip_ttl = sc->sc_ip_ttl;
1887 ip->ip_tos = sc->sc_ip_tos;
1890 * See if we should do MTU discovery. Route lookups are
1891 * expensive, so we will only unset the DF bit if:
1893 * 1) path_mtu_discovery is disabled
1894 * 2) the SCF_UNREACH flag has been set
1896 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1897 ip->ip_off |= htons(IP_DF);
1898 if (sc->sc_port == 0) {
1899 ip->ip_p = IPPROTO_TCP;
1900 th = (struct tcphdr *)(ip + 1);
1902 ip->ip_p = IPPROTO_UDP;
1903 udp = (struct udphdr *)(ip + 1);
1904 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1905 udp->uh_dport = sc->sc_port;
1906 ulen = (tlen - sizeof(struct ip));
1907 th = (struct tcphdr *)(udp + 1);
1911 th->th_sport = sc->sc_inc.inc_lport;
1912 th->th_dport = sc->sc_inc.inc_fport;
1915 th->th_seq = htonl(sc->sc_iss);
1917 th->th_seq = htonl(sc->sc_iss + 1);
1918 th->th_ack = htonl(sc->sc_irs + 1);
1919 th->th_off = sizeof(struct tcphdr) >> 2;
1920 th->th_win = htons(sc->sc_wnd);
1923 flags = tcp_ecn_syncache_respond(flags, sc);
1924 tcp_set_flags(th, flags);
1926 /* Tack on the TCP options. */
1927 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1930 if (flags & TH_SYN) {
1932 to.to_flags = TOF_MSS;
1933 if (sc->sc_flags & SCF_WINSCALE) {
1934 to.to_wscale = sc->sc_requested_r_scale;
1935 to.to_flags |= TOF_SCALE;
1937 if (sc->sc_flags & SCF_SACK)
1938 to.to_flags |= TOF_SACKPERM;
1939 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1940 if (sc->sc_flags & SCF_SIGNATURE)
1941 to.to_flags |= TOF_SIGNATURE;
1943 if (sc->sc_tfo_cookie) {
1944 to.to_flags |= TOF_FASTOPEN;
1945 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1946 to.to_tfo_cookie = sc->sc_tfo_cookie;
1947 /* don't send cookie again when retransmitting response */
1948 sc->sc_tfo_cookie = NULL;
1951 if (sc->sc_flags & SCF_TIMESTAMP) {
1952 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1953 to.to_tsecr = sc->sc_tsreflect;
1954 to.to_flags |= TOF_TS;
1956 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1958 /* Adjust headers by option size. */
1959 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1961 m->m_pkthdr.len += optlen;
1963 if (sc->sc_inc.inc_flags & INC_ISIPV6)
1964 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1967 ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1968 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1969 if (sc->sc_flags & SCF_SIGNATURE) {
1970 KASSERT(to.to_flags & TOF_SIGNATURE,
1971 ("tcp_addoptions() didn't set tcp_signature"));
1973 /* NOTE: to.to_signature is inside of mbuf */
1974 if (!TCPMD5_ENABLED() ||
1975 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1986 udp->uh_ulen = htons(ulen);
1988 M_SETFIB(m, sc->sc_inc.inc_fibnum);
1990 * If we have peer's SYN and it has a flowid, then let's assign it to
1991 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid
1992 * to SYN|ACK due to lack of inp here.
1994 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1995 m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1996 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1999 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
2001 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
2002 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2003 udp->uh_sum = in6_cksum_pseudo(ip6, ulen,
2005 th->th_sum = htons(0);
2007 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
2008 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2009 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
2012 ip6->ip6_hlim = sc->sc_ip_ttl;
2014 if (ADDED_BY_TOE(sc)) {
2015 struct toedev *tod = sc->sc_tod;
2017 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2022 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
2023 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
2026 #if defined(INET6) && defined(INET)
2032 m->m_pkthdr.csum_flags = CSUM_UDP;
2033 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2034 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
2035 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
2036 th->th_sum = htons(0);
2038 m->m_pkthdr.csum_flags = CSUM_TCP;
2039 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2040 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
2041 htons(tlen + optlen - hlen + IPPROTO_TCP));
2044 if (ADDED_BY_TOE(sc)) {
2045 struct toedev *tod = sc->sc_tod;
2047 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2052 TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
2053 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
2060 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
2061 * that exceed the capacity of the syncache by avoiding the storage of any
2062 * of the SYNs we receive. Syncookies defend against blind SYN flooding
2063 * attacks where the attacker does not have access to our responses.
2065 * Syncookies encode and include all necessary information about the
2066 * connection setup within the SYN|ACK that we send back. That way we
2067 * can avoid keeping any local state until the ACK to our SYN|ACK returns
2068 * (if ever). Normally the syncache and syncookies are running in parallel
2069 * with the latter taking over when the former is exhausted. When matching
2070 * syncache entry is found the syncookie is ignored.
2072 * The only reliable information persisting the 3WHS is our initial sequence
2073 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient
2074 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
2075 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK
2076 * returns and signifies a legitimate connection if it matches the ACK.
2078 * The available space of 32 bits to store the hash and to encode the SYN
2079 * option information is very tight and we should have at least 24 bits for
2080 * the MAC to keep the number of guesses by blind spoofing reasonably high.
2082 * SYN option information we have to encode to fully restore a connection:
2083 * MSS: is imporant to chose an optimal segment size to avoid IP level
2084 * fragmentation along the path. The common MSS values can be encoded
2085 * in a 3-bit table. Uncommon values are captured by the next lower value
2086 * in the table leading to a slight increase in packetization overhead.
2087 * WSCALE: is necessary to allow large windows to be used for high delay-
2088 * bandwidth product links. Not scaling the window when it was initially
2089 * negotiated is bad for performance as lack of scaling further decreases
2090 * the apparent available send window. We only need to encode the WSCALE
2091 * we received from the remote end. Our end can be recalculated at any
2092 * time. The common WSCALE values can be encoded in a 3-bit table.
2093 * Uncommon values are captured by the next lower value in the table
2094 * making us under-estimate the available window size halving our
2095 * theoretically possible maximum throughput for that connection.
2096 * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2097 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2098 * that are included in all segments on a connection. We enable them when
2101 * Security of syncookies and attack vectors:
2103 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2104 * together with the gloabl secret to make it unique per connection attempt.
2105 * Thus any change of any of those parameters results in a different MAC output
2106 * in an unpredictable way unless a collision is encountered. 24 bits of the
2107 * MAC are embedded into the ISS.
2109 * To prevent replay attacks two rotating global secrets are updated with a
2110 * new random value every 15 seconds. The life-time of a syncookie is thus
2113 * Vector 1: Attacking the secret. This requires finding a weakness in the
2114 * MAC itself or the way it is used here. The attacker can do a chosen plain
2115 * text attack by varying and testing the all parameters under his control.
2116 * The strength depends on the size and randomness of the secret, and the
2117 * cryptographic security of the MAC function. Due to the constant updating
2118 * of the secret the attacker has at most 29.999 seconds to find the secret
2119 * and launch spoofed connections. After that he has to start all over again.
2121 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC
2122 * size an average of 4,823 attempts are required for a 50% chance of success
2123 * to spoof a single syncookie (birthday collision paradox). However the
2124 * attacker is blind and doesn't know if one of his attempts succeeded unless
2125 * he has a side channel to interfere success from. A single connection setup
2126 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2127 * This many attempts are required for each one blind spoofed connection. For
2128 * every additional spoofed connection he has to launch another N attempts.
2129 * Thus for a sustained rate 100 spoofed connections per second approximately
2130 * 1,800,000 packets per second would have to be sent.
2132 * NB: The MAC function should be fast so that it doesn't become a CPU
2133 * exhaustion attack vector itself.
2136 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2137 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2138 * http://cr.yp.to/syncookies.html (overview)
2139 * http://cr.yp.to/syncookies/archive (details)
2142 * Schematic construction of a syncookie enabled Initial Sequence Number:
2144 * 12345678901234567890123456789012
2145 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2147 * x 24 MAC (truncated)
2148 * W 3 Send Window Scale index
2150 * S 1 SACK permitted
2151 * P 1 Odd/even secret
2155 * Distribution and probability of certain MSS values. Those in between are
2156 * rounded down to the next lower one.
2157 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2158 * .2% .3% 5% 7% 7% 20% 15% 45%
2160 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2163 * Distribution and probability of certain WSCALE values. We have to map the
2164 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2165 * bits based on prevalence of certain values. Where we don't have an exact
2166 * match for are rounded down to the next lower one letting us under-estimate
2167 * the true available window. At the moment this would happen only for the
2168 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2169 * and window size). The absence of the WSCALE option (no scaling in either
2170 * direction) is encoded with index zero.
2171 * [WSCALE values histograms, Allman, 2012]
2172 * X 10 10 35 5 6 14 10% by host
2173 * X 11 4 5 5 18 49 3% by connections
2175 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2178 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed
2179 * and good cryptographic properties.
2182 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2183 uint8_t *secbits, uintptr_t secmod)
2186 uint32_t siphash[2];
2188 SipHash24_Init(&ctx);
2189 SipHash_SetKey(&ctx, secbits);
2190 switch (inc->inc_flags & INC_ISIPV6) {
2193 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2194 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2199 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2200 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2204 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2205 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2206 SipHash_Update(&ctx, &irs, sizeof(irs));
2207 SipHash_Update(&ctx, &flags, sizeof(flags));
2208 SipHash_Update(&ctx, &secmod, sizeof(secmod));
2209 SipHash_Final((u_int8_t *)&siphash, &ctx);
2211 return (siphash[0] ^ siphash[1]);
2215 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2217 u_int i, secbit, wscale;
2220 union syncookie cookie;
2224 /* Map our computed MSS into the 3-bit index. */
2225 for (i = nitems(tcp_sc_msstab) - 1;
2226 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2229 cookie.flags.mss_idx = i;
2232 * Map the send window scale into the 3-bit index but only if
2233 * the wscale option was received.
2235 if (sc->sc_flags & SCF_WINSCALE) {
2236 wscale = sc->sc_requested_s_scale;
2237 for (i = nitems(tcp_sc_wstab) - 1;
2238 tcp_sc_wstab[i] > wscale && i > 0;
2241 cookie.flags.wscale_idx = i;
2244 /* Can we do SACK? */
2245 if (sc->sc_flags & SCF_SACK)
2246 cookie.flags.sack_ok = 1;
2248 /* Which of the two secrets to use. */
2249 secbit = V_tcp_syncache.secret.oddeven & 0x1;
2250 cookie.flags.odd_even = secbit;
2252 secbits = V_tcp_syncache.secret.key[secbit];
2253 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2257 * Put the flags into the hash and XOR them to get better ISS number
2258 * variance. This doesn't enhance the cryptographic strength and is
2259 * done to prevent the 8 cookie bits from showing up directly on the
2263 iss |= cookie.cookie ^ (hash >> 24);
2265 TCPSTAT_INC(tcps_sc_sendcookie);
2269 static struct syncache *
2270 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2271 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2272 struct socket *lso, uint16_t port)
2277 int wnd, wscale = 0;
2278 union syncookie cookie;
2281 * Pull information out of SYN-ACK/ACK and revert sequence number
2284 ack = th->th_ack - 1;
2285 seq = th->th_seq - 1;
2288 * Unpack the flags containing enough information to restore the
2291 cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2293 /* Which of the two secrets to use. */
2294 secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2296 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2298 /* The recomputed hash matches the ACK if this was a genuine cookie. */
2299 if ((ack & ~0xff) != (hash & ~0xff))
2302 /* Fill in the syncache values. */
2304 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2305 sc->sc_ipopts = NULL;
2310 switch (inc->inc_flags & INC_ISIPV6) {
2313 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2314 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2319 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2321 htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2326 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2328 /* We can simply recompute receive window scale we sent earlier. */
2329 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2332 /* Only use wscale if it was enabled in the orignal SYN. */
2333 if (cookie.flags.wscale_idx > 0) {
2334 sc->sc_requested_r_scale = wscale;
2335 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2336 sc->sc_flags |= SCF_WINSCALE;
2339 wnd = lso->sol_sbrcv_hiwat;
2341 wnd = imin(wnd, TCP_MAXWIN);
2344 if (cookie.flags.sack_ok)
2345 sc->sc_flags |= SCF_SACK;
2347 if (to->to_flags & TOF_TS) {
2348 sc->sc_flags |= SCF_TIMESTAMP;
2349 sc->sc_tsreflect = to->to_tsval;
2350 sc->sc_tsoff = tcp_new_ts_offset(inc);
2353 if (to->to_flags & TOF_SIGNATURE)
2354 sc->sc_flags |= SCF_SIGNATURE;
2360 TCPSTAT_INC(tcps_sc_recvcookie);
2366 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2367 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2368 struct socket *lso, uint16_t port)
2370 struct syncache scs, *scx;
2373 bzero(&scs, sizeof(scs));
2374 scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port);
2376 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2380 if (sc->sc_peer_mss != scx->sc_peer_mss)
2381 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2382 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2384 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2385 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2386 s, __func__, sc->sc_requested_r_scale,
2387 scx->sc_requested_r_scale);
2389 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2390 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2391 s, __func__, sc->sc_requested_s_scale,
2392 scx->sc_requested_s_scale);
2394 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2395 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2402 #endif /* INVARIANTS */
2405 syncookie_reseed(void *arg)
2407 struct tcp_syncache *sc = arg;
2412 * Reseeding the secret doesn't have to be protected by a lock.
2413 * It only must be ensured that the new random values are visible
2414 * to all CPUs in a SMP environment. The atomic with release
2415 * semantics ensures that.
2417 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2418 secbits = sc->secret.key[secbit];
2419 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2420 atomic_add_rel_int(&sc->secret.oddeven, 1);
2422 /* Reschedule ourself. */
2423 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2427 * We have overflowed a bucket. Let's pause dealing with the syncache.
2428 * This function will increment the bucketoverflow statistics appropriately
2429 * (once per pause when pausing is enabled; otherwise, once per overflow).
2432 syncache_pause(struct in_conninfo *inc)
2438 * 2. Add sysctl read here so we don't get the benefit of this
2439 * change without the new sysctl.
2443 * Try an unlocked read. If we already know that another thread
2444 * has activated the feature, there is no need to proceed.
2446 if (V_tcp_syncache.paused)
2449 /* Are cookied enabled? If not, we can't pause. */
2450 if (!V_tcp_syncookies) {
2451 TCPSTAT_INC(tcps_sc_bucketoverflow);
2456 * We may be the first thread to find an overflow. Get the lock
2457 * and evaluate if we need to take action.
2459 mtx_lock(&V_tcp_syncache.pause_mtx);
2460 if (V_tcp_syncache.paused) {
2461 mtx_unlock(&V_tcp_syncache.pause_mtx);
2465 /* Activate protection. */
2466 V_tcp_syncache.paused = true;
2467 TCPSTAT_INC(tcps_sc_bucketoverflow);
2470 * Determine the last backoff time. If we are seeing a re-newed
2471 * attack within that same time after last reactivating the syncache,
2472 * consider it an extension of the same attack.
2474 delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2475 if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2476 if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2478 V_tcp_syncache.pause_backoff++;
2481 delta = TCP_SYNCACHE_PAUSE_TIME;
2482 V_tcp_syncache.pause_backoff = 0;
2485 /* Log a warning, including IP addresses, if able. */
2487 s = tcp_log_addrs(inc, NULL, NULL, NULL);
2489 s = (const char *)NULL;
2490 log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2491 "the next %lld seconds%s%s%s\n", (long long)delta,
2492 (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2493 (s != NULL) ? ")" : "");
2494 free(__DECONST(void *, s), M_TCPLOG);
2496 /* Use the calculated delta to set a new pause time. */
2497 V_tcp_syncache.pause_until = time_uptime + delta;
2498 callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2500 mtx_unlock(&V_tcp_syncache.pause_mtx);
2503 /* Evaluate whether we need to unpause. */
2505 syncache_unpause(void *arg)
2507 struct tcp_syncache *sc;
2511 mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2512 callout_deactivate(&sc->pause_co);
2515 * Check to make sure we are not running early. If the pause
2516 * time has expired, then deactivate the protection.
2518 if ((delta = sc->pause_until - time_uptime) > 0)
2519 callout_schedule(&sc->pause_co, delta * hz);
2525 * Exports the syncache entries to userland so that netstat can display
2526 * them alongside the other sockets. This function is intended to be
2527 * called only from tcp_pcblist.
2529 * Due to concurrency on an active system, the number of pcbs exported
2530 * may have no relation to max_pcbs. max_pcbs merely indicates the
2531 * amount of space the caller allocated for this function to use.
2534 syncache_pcblist(struct sysctl_req *req)
2537 struct syncache *sc;
2538 struct syncache_head *sch;
2541 bzero(&xt, sizeof(xt));
2542 xt.xt_len = sizeof(xt);
2543 xt.t_state = TCPS_SYN_RECEIVED;
2544 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2545 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2546 xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2547 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2549 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2550 sch = &V_tcp_syncache.hashbase[i];
2552 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2553 if (sc->sc_cred != NULL &&
2554 cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2556 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2557 xt.xt_inp.inp_vflag = INP_IPV6;
2559 xt.xt_inp.inp_vflag = INP_IPV4;
2560 xt.xt_encaps_port = sc->sc_port;
2561 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2562 sizeof (struct in_conninfo));
2563 error = SYSCTL_OUT(req, &xt, sizeof xt);