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"
41 #include "opt_pcbgroup.h"
43 #include <sys/param.h>
44 #include <sys/systm.h>
46 #include <sys/refcount.h>
47 #include <sys/kernel.h>
48 #include <sys/sysctl.h>
49 #include <sys/limits.h>
51 #include <sys/mutex.h>
52 #include <sys/malloc.h>
54 #include <sys/proc.h> /* for proc0 declaration */
55 #include <sys/random.h>
56 #include <sys/socket.h>
57 #include <sys/socketvar.h>
58 #include <sys/syslog.h>
59 #include <sys/ucred.h>
62 #include <crypto/siphash/siphash.h>
67 #include <net/if_var.h>
68 #include <net/route.h>
71 #include <netinet/in.h>
72 #include <netinet/in_kdtrace.h>
73 #include <netinet/in_systm.h>
74 #include <netinet/ip.h>
75 #include <netinet/in_var.h>
76 #include <netinet/in_pcb.h>
77 #include <netinet/ip_var.h>
78 #include <netinet/ip_options.h>
80 #include <netinet/ip6.h>
81 #include <netinet/icmp6.h>
82 #include <netinet6/nd6.h>
83 #include <netinet6/ip6_var.h>
84 #include <netinet6/in6_pcb.h>
86 #include <netinet/tcp.h>
87 #include <netinet/tcp_fastopen.h>
88 #include <netinet/tcp_fsm.h>
89 #include <netinet/tcp_seq.h>
90 #include <netinet/tcp_timer.h>
91 #include <netinet/tcp_var.h>
92 #include <netinet/tcp_syncache.h>
94 #include <netinet6/tcp6_var.h>
97 #include <netinet/toecore.h>
100 #include <netipsec/ipsec_support.h>
102 #include <machine/in_cksum.h>
104 #include <security/mac/mac_framework.h>
106 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1;
107 #define V_tcp_syncookies VNET(tcp_syncookies)
108 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
109 &VNET_NAME(tcp_syncookies), 0,
110 "Use TCP SYN cookies if the syncache overflows");
112 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0;
113 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly)
114 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
115 &VNET_NAME(tcp_syncookiesonly), 0,
116 "Use only TCP SYN cookies");
118 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1;
119 #define V_functions_inherit_listen_socket_stack \
120 VNET(functions_inherit_listen_socket_stack)
121 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack,
122 CTLFLAG_VNET | CTLFLAG_RW,
123 &VNET_NAME(functions_inherit_listen_socket_stack), 0,
124 "Inherit listen socket's stack");
127 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
130 static void syncache_drop(struct syncache *, struct syncache_head *);
131 static void syncache_free(struct syncache *);
132 static void syncache_insert(struct syncache *, struct syncache_head *);
133 static int syncache_respond(struct syncache *, const struct mbuf *, int);
134 static struct socket *syncache_socket(struct syncache *, struct socket *,
136 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch,
138 static void syncache_timer(void *);
140 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
141 uint8_t *, uintptr_t);
142 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *);
143 static struct syncache
144 *syncookie_lookup(struct in_conninfo *, struct syncache_head *,
145 struct syncache *, struct tcphdr *, struct tcpopt *,
147 static void syncache_pause(struct in_conninfo *);
148 static void syncache_unpause(void *);
149 static void syncookie_reseed(void *);
151 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
152 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
157 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
158 * 3 retransmits corresponds to a timeout with default values of
159 * tcp_rexmit_initial * ( 1 +
162 * tcp_backoff[3]) + 3 * tcp_rexmit_slop,
163 * 1000 ms * (1 + 2 + 4 + 8) + 3 * 200 ms = 15600 ms,
164 * the odds are that the user has given up attempting to connect by then.
166 #define SYNCACHE_MAXREXMTS 3
168 /* Arbitrary values */
169 #define TCP_SYNCACHE_HASHSIZE 512
170 #define TCP_SYNCACHE_BUCKETLIMIT 30
172 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
173 #define V_tcp_syncache VNET(tcp_syncache)
175 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0,
178 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
179 &VNET_NAME(tcp_syncache.bucket_limit), 0,
180 "Per-bucket hash limit for syncache");
182 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
183 &VNET_NAME(tcp_syncache.cache_limit), 0,
184 "Overall entry limit for syncache");
186 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
187 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
189 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
190 &VNET_NAME(tcp_syncache.hashsize), 0,
191 "Size of TCP syncache hashtable");
194 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
199 new = V_tcp_syncache.rexmt_limit;
200 error = sysctl_handle_int(oidp, &new, 0, req);
201 if ((error == 0) && (req->newptr != NULL)) {
202 if (new > TCP_MAXRXTSHIFT)
205 V_tcp_syncache.rexmt_limit = new;
210 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
211 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW,
212 &VNET_NAME(tcp_syncache.rexmt_limit), 0,
213 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
214 "Limit on SYN/ACK retransmissions");
216 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
217 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
218 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
219 "Send reset on socket allocation failure");
221 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
223 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx)
224 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx)
225 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED)
228 * Requires the syncache entry to be already removed from the bucket list.
231 syncache_free(struct syncache *sc)
235 (void) m_free(sc->sc_ipopts);
239 mac_syncache_destroy(&sc->sc_label);
242 uma_zfree(V_tcp_syncache.zone, sc);
250 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
251 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
252 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
253 V_tcp_syncache.hash_secret = arc4random();
255 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
256 &V_tcp_syncache.hashsize);
257 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
258 &V_tcp_syncache.bucket_limit);
259 if (!powerof2(V_tcp_syncache.hashsize) ||
260 V_tcp_syncache.hashsize == 0) {
261 printf("WARNING: syncache hash size is not a power of 2.\n");
262 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
264 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
267 V_tcp_syncache.cache_limit =
268 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
269 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
270 &V_tcp_syncache.cache_limit);
272 /* Allocate the hash table. */
273 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
274 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
277 V_tcp_syncache.vnet = curvnet;
280 /* Initialize the hash buckets. */
281 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
282 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
283 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
285 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
286 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
287 V_tcp_syncache.hashbase[i].sch_length = 0;
288 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
289 V_tcp_syncache.hashbase[i].sch_last_overflow =
290 -(SYNCOOKIE_LIFETIME + 1);
293 /* Create the syncache entry zone. */
294 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
295 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
296 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
297 V_tcp_syncache.cache_limit);
299 /* Start the SYN cookie reseeder callout. */
300 callout_init(&V_tcp_syncache.secret.reseed, 1);
301 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
302 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
303 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
304 syncookie_reseed, &V_tcp_syncache);
306 /* Initialize the pause machinery. */
307 mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF);
308 callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx,
310 V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME;
311 V_tcp_syncache.pause_backoff = 0;
312 V_tcp_syncache.paused = false;
317 syncache_destroy(void)
319 struct syncache_head *sch;
320 struct syncache *sc, *nsc;
324 * Stop the re-seed timer before freeing resources. No need to
325 * possibly schedule it another time.
327 callout_drain(&V_tcp_syncache.secret.reseed);
329 /* Stop the SYN cache pause callout. */
330 mtx_lock(&V_tcp_syncache.pause_mtx);
331 if (callout_stop(&V_tcp_syncache.pause_co) == 0) {
332 mtx_unlock(&V_tcp_syncache.pause_mtx);
333 callout_drain(&V_tcp_syncache.pause_co);
335 mtx_unlock(&V_tcp_syncache.pause_mtx);
337 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */
338 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
340 sch = &V_tcp_syncache.hashbase[i];
341 callout_drain(&sch->sch_timer);
344 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
345 syncache_drop(sc, sch);
347 KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
348 ("%s: sch->sch_bucket not empty", __func__));
349 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
350 __func__, sch->sch_length));
351 mtx_destroy(&sch->sch_mtx);
354 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
355 ("%s: cache_count not 0", __func__));
357 /* Free the allocated global resources. */
358 uma_zdestroy(V_tcp_syncache.zone);
359 free(V_tcp_syncache.hashbase, M_SYNCACHE);
360 mtx_destroy(&V_tcp_syncache.pause_mtx);
365 * Inserts a syncache entry into the specified bucket row.
366 * Locks and unlocks the syncache_head autonomously.
369 syncache_insert(struct syncache *sc, struct syncache_head *sch)
371 struct syncache *sc2;
376 * Make sure that we don't overflow the per-bucket limit.
377 * If the bucket is full, toss the oldest element.
379 if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
380 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
381 ("sch->sch_length incorrect"));
382 syncache_pause(&sc->sc_inc);
383 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
384 sch->sch_last_overflow = time_uptime;
385 syncache_drop(sc2, sch);
388 /* Put it into the bucket. */
389 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
393 if (ADDED_BY_TOE(sc)) {
394 struct toedev *tod = sc->sc_tod;
396 tod->tod_syncache_added(tod, sc->sc_todctx);
400 /* Reinitialize the bucket row's timer. */
401 if (sch->sch_length == 1)
402 sch->sch_nextc = ticks + INT_MAX;
403 syncache_timeout(sc, sch, 1);
407 TCPSTATES_INC(TCPS_SYN_RECEIVED);
408 TCPSTAT_INC(tcps_sc_added);
412 * Remove and free entry from syncache bucket row.
413 * Expects locked syncache head.
416 syncache_drop(struct syncache *sc, struct syncache_head *sch)
419 SCH_LOCK_ASSERT(sch);
421 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
422 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
426 if (ADDED_BY_TOE(sc)) {
427 struct toedev *tod = sc->sc_tod;
429 tod->tod_syncache_removed(tod, sc->sc_todctx);
437 * Engage/reengage time on bucket row.
440 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
444 if (sc->sc_rxmits == 0)
445 rexmt = tcp_rexmit_initial;
448 tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
449 tcp_rexmit_min, TCPTV_REXMTMAX);
450 sc->sc_rxttime = ticks + rexmt;
452 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
453 sch->sch_nextc = sc->sc_rxttime;
455 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
456 syncache_timer, (void *)sch);
461 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
462 * If we have retransmitted an entry the maximum number of times, expire it.
463 * One separate timer for each bucket row.
466 syncache_timer(void *xsch)
468 struct syncache_head *sch = (struct syncache_head *)xsch;
469 struct syncache *sc, *nsc;
470 struct epoch_tracker et;
475 CURVNET_SET(sch->sch_sc->vnet);
477 /* NB: syncache_head has already been locked by the callout. */
478 SCH_LOCK_ASSERT(sch);
481 * In the following cycle we may remove some entries and/or
482 * advance some timeouts, so re-initialize the bucket timer.
484 sch->sch_nextc = tick + INT_MAX;
487 * If we have paused processing, unconditionally remove
488 * all syncache entries.
490 mtx_lock(&V_tcp_syncache.pause_mtx);
491 paused = V_tcp_syncache.paused;
492 mtx_unlock(&V_tcp_syncache.pause_mtx);
494 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
496 syncache_drop(sc, sch);
500 * We do not check if the listen socket still exists
501 * and accept the case where the listen socket may be
502 * gone by the time we resend the SYN/ACK. We do
503 * not expect this to happens often. If it does,
504 * then the RST will be sent by the time the remote
505 * host does the SYN/ACK->ACK.
507 if (TSTMP_GT(sc->sc_rxttime, tick)) {
508 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
509 sch->sch_nextc = sc->sc_rxttime;
512 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
513 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
514 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
515 "giving up and removing syncache entry\n",
519 syncache_drop(sc, sch);
520 TCPSTAT_INC(tcps_sc_stale);
523 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
524 log(LOG_DEBUG, "%s; %s: Response timeout, "
525 "retransmitting (%u) SYN|ACK\n",
526 s, __func__, sc->sc_rxmits);
531 syncache_respond(sc, NULL, TH_SYN|TH_ACK);
533 TCPSTAT_INC(tcps_sc_retransmitted);
534 syncache_timeout(sc, sch, 0);
536 if (!TAILQ_EMPTY(&(sch)->sch_bucket))
537 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
538 syncache_timer, (void *)(sch));
543 * Returns true if the system is only using cookies at the moment.
544 * This could be due to a sysadmin decision to only use cookies, or it
545 * could be due to the system detecting an attack.
548 syncache_cookiesonly(void)
551 return (V_tcp_syncookies && (V_tcp_syncache.paused ||
552 V_tcp_syncookiesonly));
556 * Find the hash bucket for the given connection.
558 static struct syncache_head *
559 syncache_hashbucket(struct in_conninfo *inc)
564 * The hash is built on foreign port + local port + foreign address.
565 * We rely on the fact that struct in_conninfo starts with 16 bits
566 * of foreign port, then 16 bits of local port then followed by 128
567 * bits of foreign address. In case of IPv4 address, the first 3
568 * 32-bit words of the address always are zeroes.
570 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
571 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
573 return (&V_tcp_syncache.hashbase[hash]);
577 * Find an entry in the syncache.
578 * Returns always with locked syncache_head plus a matching entry or NULL.
580 static struct syncache *
581 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
584 struct syncache_head *sch;
586 *schp = sch = syncache_hashbucket(inc);
589 /* Circle through bucket row to find matching entry. */
590 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
591 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
592 sizeof(struct in_endpoints)) == 0)
595 return (sc); /* Always returns with locked sch. */
599 * This function is called when we get a RST for a
600 * non-existent connection, so that we can see if the
601 * connection is in the syn cache. If it is, zap it.
602 * If required send a challenge ACK.
605 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m)
608 struct syncache_head *sch;
611 if (syncache_cookiesonly())
613 sc = syncache_lookup(inc, &sch); /* returns locked sch */
614 SCH_LOCK_ASSERT(sch);
617 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
618 * See RFC 793 page 65, section SEGMENT ARRIVES.
620 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
621 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
622 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
623 "FIN flag set, segment ignored\n", s, __func__);
624 TCPSTAT_INC(tcps_badrst);
629 * No corresponding connection was found in syncache.
630 * If syncookies are enabled and possibly exclusively
631 * used, or we are under memory pressure, a valid RST
632 * may not find a syncache entry. In that case we're
633 * done and no SYN|ACK retransmissions will happen.
634 * Otherwise the RST was misdirected or spoofed.
637 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
638 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
639 "syncache entry (possibly syncookie only), "
640 "segment ignored\n", s, __func__);
641 TCPSTAT_INC(tcps_badrst);
646 * If the RST bit is set, check the sequence number to see
647 * if this is a valid reset segment.
650 * In all states except SYN-SENT, all reset (RST) segments
651 * are validated by checking their SEQ-fields. A reset is
652 * valid if its sequence number is in the window.
655 * There are four cases for the acceptability test for an incoming
658 * Segment Receive Test
660 * ------- ------- -------------------------------------------
661 * 0 0 SEG.SEQ = RCV.NXT
662 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
663 * >0 0 not acceptable
664 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
665 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
667 * Note that when receiving a SYN segment in the LISTEN state,
668 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
669 * described in RFC 793, page 66.
671 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
672 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
673 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
674 if (V_tcp_insecure_rst ||
675 th->th_seq == sc->sc_irs + 1) {
676 syncache_drop(sc, sch);
677 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
679 "%s; %s: Our SYN|ACK was rejected, "
680 "connection attempt aborted by remote "
683 TCPSTAT_INC(tcps_sc_reset);
685 TCPSTAT_INC(tcps_badrst);
686 /* Send challenge ACK. */
687 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
688 log(LOG_DEBUG, "%s; %s: RST with invalid "
689 " SEQ %u != NXT %u (+WND %u), "
690 "sending challenge ACK\n",
692 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
693 syncache_respond(sc, m, TH_ACK);
696 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
697 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
698 "NXT %u (+WND %u), segment ignored\n",
700 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
701 TCPSTAT_INC(tcps_badrst);
711 syncache_badack(struct in_conninfo *inc)
714 struct syncache_head *sch;
716 if (syncache_cookiesonly())
718 sc = syncache_lookup(inc, &sch); /* returns locked sch */
719 SCH_LOCK_ASSERT(sch);
721 syncache_drop(sc, sch);
722 TCPSTAT_INC(tcps_sc_badack);
728 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq)
731 struct syncache_head *sch;
733 if (syncache_cookiesonly())
735 sc = syncache_lookup(inc, &sch); /* returns locked sch */
736 SCH_LOCK_ASSERT(sch);
740 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
741 if (ntohl(th_seq) != sc->sc_iss)
745 * If we've rertransmitted 3 times and this is our second error,
746 * we remove the entry. Otherwise, we allow it to continue on.
747 * This prevents us from incorrectly nuking an entry during a
748 * spurious network outage.
752 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
753 sc->sc_flags |= SCF_UNREACH;
756 syncache_drop(sc, sch);
757 TCPSTAT_INC(tcps_sc_unreach);
763 * Build a new TCP socket structure from a syncache entry.
765 * On success return the newly created socket with its underlying inp locked.
767 static struct socket *
768 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
770 struct tcp_function_block *blk;
771 struct inpcb *inp = NULL;
780 * Ok, create the full blown connection, and set things up
781 * as they would have been set up if we had created the
782 * connection when the SYN arrived. If we can't create
783 * the connection, abort it.
785 so = sonewconn(lso, 0);
788 * Drop the connection; we will either send a RST or
789 * have the peer retransmit its SYN again after its
792 TCPSTAT_INC(tcps_listendrop);
793 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
794 log(LOG_DEBUG, "%s; %s: Socket create failed "
795 "due to limits or memory shortage\n",
802 mac_socketpeer_set_from_mbuf(m, so);
806 inp->inp_inc.inc_fibnum = so->so_fibnum;
809 * Exclusive pcbinfo lock is not required in syncache socket case even
810 * if two inpcb locks can be acquired simultaneously:
811 * - the inpcb in LISTEN state,
812 * - the newly created inp.
814 * In this case, an inp cannot be at same time in LISTEN state and
815 * just created by an accept() call.
817 INP_HASH_WLOCK(&V_tcbinfo);
819 /* Insert new socket into PCB hash list. */
820 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
822 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
823 inp->inp_vflag &= ~INP_IPV4;
824 inp->inp_vflag |= INP_IPV6;
825 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
827 inp->inp_vflag &= ~INP_IPV6;
828 inp->inp_vflag |= INP_IPV4;
830 inp->inp_laddr = sc->sc_inc.inc_laddr;
836 * If there's an mbuf and it has a flowid, then let's initialise the
837 * inp with that particular flowid.
839 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
840 inp->inp_flowid = m->m_pkthdr.flowid;
841 inp->inp_flowtype = M_HASHTYPE_GET(m);
843 inp->inp_numa_domain = m->m_pkthdr.numa_domain;
847 inp->inp_lport = sc->sc_inc.inc_lport;
849 if (inp->inp_vflag & INP_IPV6PROTO) {
850 struct inpcb *oinp = sotoinpcb(lso);
853 * Inherit socket options from the listening socket.
854 * Note that in6p_inputopts are not (and should not be)
855 * copied, since it stores previously received options and is
856 * used to detect if each new option is different than the
857 * previous one and hence should be passed to a user.
858 * If we copied in6p_inputopts, a user would not be able to
859 * receive options just after calling the accept system call.
861 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
862 if (oinp->in6p_outputopts)
863 inp->in6p_outputopts =
864 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
867 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
868 struct in6_addr laddr6;
869 struct sockaddr_in6 sin6;
871 sin6.sin6_family = AF_INET6;
872 sin6.sin6_len = sizeof(sin6);
873 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
874 sin6.sin6_port = sc->sc_inc.inc_fport;
875 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
876 laddr6 = inp->in6p_laddr;
877 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
878 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
879 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
880 thread0.td_ucred, m, false)) != 0) {
881 inp->in6p_laddr = laddr6;
882 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
883 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
888 INP_HASH_WUNLOCK(&V_tcbinfo);
891 /* Override flowlabel from in6_pcbconnect. */
892 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
893 inp->inp_flow |= sc->sc_flowlabel;
896 #if defined(INET) && defined(INET6)
901 struct in_addr laddr;
902 struct sockaddr_in sin;
904 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
906 if (inp->inp_options == NULL) {
907 inp->inp_options = sc->sc_ipopts;
908 sc->sc_ipopts = NULL;
911 sin.sin_family = AF_INET;
912 sin.sin_len = sizeof(sin);
913 sin.sin_addr = sc->sc_inc.inc_faddr;
914 sin.sin_port = sc->sc_inc.inc_fport;
915 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
916 laddr = inp->inp_laddr;
917 if (inp->inp_laddr.s_addr == INADDR_ANY)
918 inp->inp_laddr = sc->sc_inc.inc_laddr;
919 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
920 thread0.td_ucred, m, false)) != 0) {
921 inp->inp_laddr = laddr;
922 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
923 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
928 INP_HASH_WUNLOCK(&V_tcbinfo);
933 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
934 /* Copy old policy into new socket's. */
935 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
936 printf("syncache_socket: could not copy policy\n");
938 INP_HASH_WUNLOCK(&V_tcbinfo);
940 tcp_state_change(tp, TCPS_SYN_RECEIVED);
941 tp->iss = sc->sc_iss;
942 tp->irs = sc->sc_irs;
945 blk = sototcpcb(lso)->t_fb;
946 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
948 * Our parents t_fb was not the default,
949 * we need to release our ref on tp->t_fb and
950 * pickup one on the new entry.
952 struct tcp_function_block *rblk;
954 rblk = find_and_ref_tcp_fb(blk);
955 KASSERT(rblk != NULL,
956 ("cannot find blk %p out of syncache?", blk));
957 if (tp->t_fb->tfb_tcp_fb_fini)
958 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
959 refcount_release(&tp->t_fb->tfb_refcnt);
962 * XXXrrs this is quite dangerous, it is possible
963 * for the new function to fail to init. We also
964 * are not asking if the handoff_is_ok though at
965 * the very start thats probalbly ok.
967 if (tp->t_fb->tfb_tcp_fb_init) {
968 (*tp->t_fb->tfb_tcp_fb_init)(tp);
971 tp->snd_wl1 = sc->sc_irs;
972 tp->snd_max = tp->iss + 1;
973 tp->snd_nxt = tp->iss + 1;
974 tp->rcv_up = sc->sc_irs + 1;
975 tp->rcv_wnd = sc->sc_wnd;
976 tp->rcv_adv += tp->rcv_wnd;
977 tp->last_ack_sent = tp->rcv_nxt;
979 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
980 if (sc->sc_flags & SCF_NOOPT)
981 tp->t_flags |= TF_NOOPT;
983 if (sc->sc_flags & SCF_WINSCALE) {
984 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
985 tp->snd_scale = sc->sc_requested_s_scale;
986 tp->request_r_scale = sc->sc_requested_r_scale;
988 if (sc->sc_flags & SCF_TIMESTAMP) {
989 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
990 tp->ts_recent = sc->sc_tsreflect;
991 tp->ts_recent_age = tcp_ts_getticks();
992 tp->ts_offset = sc->sc_tsoff;
994 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
995 if (sc->sc_flags & SCF_SIGNATURE)
996 tp->t_flags |= TF_SIGNATURE;
998 if (sc->sc_flags & SCF_SACK)
999 tp->t_flags |= TF_SACK_PERMIT;
1002 if (sc->sc_flags & SCF_ECN)
1003 tp->t_flags2 |= TF2_ECN_PERMIT;
1006 * Set up MSS and get cached values from tcp_hostcache.
1007 * This might overwrite some of the defaults we just set.
1009 tcp_mss(tp, sc->sc_peer_mss);
1012 * If the SYN,ACK was retransmitted, indicate that CWND to be
1013 * limited to one segment in cc_conn_init().
1014 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
1016 if (sc->sc_rxmits > 1)
1021 * Allow a TOE driver to install its hooks. Note that we hold the
1022 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
1023 * new connection before the TOE driver has done its thing.
1025 if (ADDED_BY_TOE(sc)) {
1026 struct toedev *tod = sc->sc_tod;
1028 tod->tod_offload_socket(tod, sc->sc_todctx, so);
1032 * Copy and activate timers.
1034 tp->t_keepinit = sototcpcb(lso)->t_keepinit;
1035 tp->t_keepidle = sototcpcb(lso)->t_keepidle;
1036 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
1037 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1038 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1040 TCPSTAT_INC(tcps_accepts);
1052 * This function gets called when we receive an ACK for a
1053 * socket in the LISTEN state. We look up the connection
1054 * in the syncache, and if its there, we pull it out of
1055 * the cache and turn it into a full-blown connection in
1056 * the SYN-RECEIVED state.
1058 * On syncache_socket() success the newly created socket
1059 * has its underlying inp locked.
1062 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1063 struct socket **lsop, struct mbuf *m)
1065 struct syncache *sc;
1066 struct syncache_head *sch;
1067 struct syncache scs;
1072 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1073 ("%s: can handle only ACK", __func__));
1075 if (syncache_cookiesonly()) {
1077 sch = syncache_hashbucket(inc);
1080 sc = syncache_lookup(inc, &sch); /* returns locked sch */
1082 SCH_LOCK_ASSERT(sch);
1087 * Test code for syncookies comparing the syncache stored
1088 * values with the reconstructed values from the cookie.
1091 syncookie_cmp(inc, sch, sc, th, to, *lsop);
1096 * There is no syncache entry, so see if this ACK is
1097 * a returning syncookie. To do this, first:
1098 * A. Check if syncookies are used in case of syncache
1100 * B. See if this socket has had a syncache entry dropped in
1101 * the recent past. We don't want to accept a bogus
1102 * syncookie if we've never received a SYN or accept it
1104 * C. check that the syncookie is valid. If it is, then
1105 * cobble up a fake syncache entry, and return.
1107 if (locked && !V_tcp_syncookies) {
1109 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1110 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1111 "segment rejected (syncookies disabled)\n",
1115 if (locked && !V_tcp_syncookiesonly &&
1116 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1118 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1119 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1120 "segment rejected (no syncache entry)\n",
1124 bzero(&scs, sizeof(scs));
1125 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
1129 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1130 log(LOG_DEBUG, "%s; %s: Segment failed "
1131 "SYNCOOKIE authentication, segment rejected "
1132 "(probably spoofed)\n", s, __func__);
1135 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1136 /* If received ACK has MD5 signature, check it. */
1137 if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1138 (!TCPMD5_ENABLED() ||
1139 TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1141 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1142 log(LOG_DEBUG, "%s; %s: Segment rejected, "
1143 "MD5 signature doesn't match.\n",
1147 TCPSTAT_INC(tcps_sig_err_sigopt);
1148 return (-1); /* Do not send RST */
1150 #endif /* TCP_SIGNATURE */
1152 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1154 * If listening socket requested TCP digests, check that
1155 * received ACK has signature and it is correct.
1156 * If not, drop the ACK and leave sc entry in th cache,
1157 * because SYN was received with correct signature.
1159 if (sc->sc_flags & SCF_SIGNATURE) {
1160 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1162 TCPSTAT_INC(tcps_sig_err_nosigopt);
1164 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1165 log(LOG_DEBUG, "%s; %s: Segment "
1166 "rejected, MD5 signature wasn't "
1167 "provided.\n", s, __func__);
1170 return (-1); /* Do not send RST */
1172 if (!TCPMD5_ENABLED() ||
1173 TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1174 /* Doesn't match or no SA */
1176 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1177 log(LOG_DEBUG, "%s; %s: Segment "
1178 "rejected, MD5 signature doesn't "
1179 "match.\n", s, __func__);
1182 return (-1); /* Do not send RST */
1185 #endif /* TCP_SIGNATURE */
1188 * RFC 7323 PAWS: If we have a timestamp on this segment and
1189 * it's less than ts_recent, drop it.
1190 * XXXMT: RFC 7323 also requires to send an ACK.
1191 * In tcp_input.c this is only done for TCP segments
1192 * with user data, so be consistent here and just drop
1195 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1196 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1198 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1200 "%s; %s: SEG.TSval %u < TS.Recent %u, "
1201 "segment dropped\n", s, __func__,
1202 to->to_tsval, sc->sc_tsreflect);
1205 return (-1); /* Do not send RST */
1209 * Pull out the entry to unlock the bucket row.
1211 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1212 * tcp_state_change(). The tcpcb is not existent at this
1213 * moment. A new one will be allocated via syncache_socket->
1214 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1215 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1217 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1218 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1221 if (ADDED_BY_TOE(sc)) {
1222 struct toedev *tod = sc->sc_tod;
1224 tod->tod_syncache_removed(tod, sc->sc_todctx);
1231 * Segment validation:
1232 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1234 if (th->th_ack != sc->sc_iss + 1) {
1235 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1236 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1237 "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1242 * The SEQ must fall in the window starting at the received
1243 * initial receive sequence number + 1 (the SYN).
1245 if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1246 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1247 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1248 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1249 "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1254 * If timestamps were not negotiated during SYN/ACK they
1255 * must not appear on any segment during this session.
1257 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1258 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1259 log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1260 "segment rejected\n", s, __func__);
1265 * If timestamps were negotiated during SYN/ACK they should
1266 * appear on every segment during this session.
1267 * XXXAO: This is only informal as there have been unverified
1268 * reports of non-compliants stacks.
1270 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1271 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1272 log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1273 "no action\n", s, __func__);
1279 *lsop = syncache_socket(sc, *lsop, m);
1282 TCPSTAT_INC(tcps_sc_aborted);
1284 TCPSTAT_INC(tcps_sc_completed);
1286 /* how do we find the inp for the new socket? */
1291 if (sc != NULL && sc != &scs)
1300 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
1301 uint64_t response_cookie)
1305 unsigned int *pending_counter;
1309 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
1310 *lsop = syncache_socket(sc, *lsop, m);
1311 if (*lsop == NULL) {
1312 TCPSTAT_INC(tcps_sc_aborted);
1313 atomic_subtract_int(pending_counter, 1);
1315 soisconnected(*lsop);
1316 inp = sotoinpcb(*lsop);
1317 tp = intotcpcb(inp);
1318 tp->t_flags |= TF_FASTOPEN;
1319 tp->t_tfo_cookie.server = response_cookie;
1320 tp->snd_max = tp->iss;
1321 tp->snd_nxt = tp->iss;
1322 tp->t_tfo_pending = pending_counter;
1323 TCPSTAT_INC(tcps_sc_completed);
1328 * Given a LISTEN socket and an inbound SYN request, add
1329 * this to the syn cache, and send back a segment:
1330 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1333 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1334 * Doing so would require that we hold onto the data and deliver it
1335 * to the application. However, if we are the target of a SYN-flood
1336 * DoS attack, an attacker could send data which would eventually
1337 * consume all available buffer space if it were ACKed. By not ACKing
1338 * the data, we avoid this DoS scenario.
1340 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1341 * cookie is processed and a new socket is created. In this case, any data
1342 * accompanying the SYN will be queued to the socket by tcp_input() and will
1343 * be ACKed either when the application sends response data or the delayed
1344 * ACK timer expires, whichever comes first.
1347 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1348 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1349 void *todctx, uint8_t iptos)
1353 struct syncache *sc = NULL;
1354 struct syncache_head *sch;
1355 struct mbuf *ipopts = NULL;
1357 int win, ip_ttl, ip_tos;
1361 int autoflowlabel = 0;
1364 struct label *maclabel;
1366 struct syncache scs;
1368 uint64_t tfo_response_cookie;
1369 unsigned int *tfo_pending = NULL;
1370 int tfo_cookie_valid = 0;
1371 int tfo_response_cookie_valid = 0;
1374 INP_WLOCK_ASSERT(inp); /* listen socket */
1375 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1376 ("%s: unexpected tcp flags", __func__));
1379 * Combine all so/tp operations very early to drop the INP lock as
1383 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1385 cred = crhold(so->so_cred);
1388 if ((inc->inc_flags & INC_ISIPV6) &&
1389 (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1392 ip_ttl = inp->inp_ip_ttl;
1393 ip_tos = inp->inp_ip_tos;
1394 win = so->sol_sbrcv_hiwat;
1395 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1397 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
1398 (tp->t_tfo_pending != NULL) &&
1399 (to->to_flags & TOF_FASTOPEN)) {
1401 * Limit the number of pending TFO connections to
1402 * approximately half of the queue limit. This prevents TFO
1403 * SYN floods from starving the service by filling the
1404 * listen queue with bogus TFO connections.
1406 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1407 (so->sol_qlimit / 2)) {
1410 result = tcp_fastopen_check_cookie(inc,
1411 to->to_tfo_cookie, to->to_tfo_len,
1412 &tfo_response_cookie);
1413 tfo_cookie_valid = (result > 0);
1414 tfo_response_cookie_valid = (result >= 0);
1418 * Remember the TFO pending counter as it will have to be
1419 * decremented below if we don't make it to syncache_tfo_expand().
1421 tfo_pending = tp->t_tfo_pending;
1424 /* By the time we drop the lock these should no longer be used. */
1429 if (mac_syncache_init(&maclabel) != 0) {
1433 mac_syncache_create(maclabel, inp);
1435 if (!tfo_cookie_valid)
1439 * Remember the IP options, if any.
1442 if (!(inc->inc_flags & INC_ISIPV6))
1445 ipopts = (m) ? ip_srcroute(m) : NULL;
1450 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1452 * If listening socket requested TCP digests, check that received
1453 * SYN has signature and it is correct. If signature doesn't match
1454 * or TCP_SIGNATURE support isn't enabled, drop the packet.
1456 if (ltflags & TF_SIGNATURE) {
1457 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1458 TCPSTAT_INC(tcps_sig_err_nosigopt);
1461 if (!TCPMD5_ENABLED() ||
1462 TCPMD5_INPUT(m, th, to->to_signature) != 0)
1465 #endif /* TCP_SIGNATURE */
1467 * See if we already have an entry for this connection.
1468 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1470 * XXX: should the syncache be re-initialized with the contents
1471 * of the new SYN here (which may have different options?)
1473 * XXX: We do not check the sequence number to see if this is a
1474 * real retransmit or a new connection attempt. The question is
1475 * how to handle such a case; either ignore it as spoofed, or
1476 * drop the current entry and create a new one?
1478 if (syncache_cookiesonly()) {
1480 sch = syncache_hashbucket(inc);
1483 sc = syncache_lookup(inc, &sch); /* returns locked sch */
1485 SCH_LOCK_ASSERT(sch);
1488 if (tfo_cookie_valid)
1490 TCPSTAT_INC(tcps_sc_dupsyn);
1493 * If we were remembering a previous source route,
1494 * forget it and use the new one we've been given.
1497 (void) m_free(sc->sc_ipopts);
1498 sc->sc_ipopts = ipopts;
1501 * Update timestamp if present.
1503 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1504 sc->sc_tsreflect = to->to_tsval;
1506 sc->sc_flags &= ~SCF_TIMESTAMP;
1509 * Since we have already unconditionally allocated label
1510 * storage, free it up. The syncache entry will already
1511 * have an initialized label we can use.
1513 mac_syncache_destroy(&maclabel);
1515 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1516 /* Retransmit SYN|ACK and reset retransmit count. */
1517 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1518 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1519 "resetting timer and retransmitting SYN|ACK\n",
1523 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1525 syncache_timeout(sc, sch, 1);
1526 TCPSTAT_INC(tcps_sndacks);
1527 TCPSTAT_INC(tcps_sndtotal);
1533 if (tfo_cookie_valid) {
1534 bzero(&scs, sizeof(scs));
1540 * Skip allocating a syncache entry if we are just going to discard
1544 bzero(&scs, sizeof(scs));
1547 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1550 * The zone allocator couldn't provide more entries.
1551 * Treat this as if the cache was full; drop the oldest
1552 * entry and insert the new one.
1554 TCPSTAT_INC(tcps_sc_zonefail);
1555 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1556 sch->sch_last_overflow = time_uptime;
1557 syncache_drop(sc, sch);
1558 syncache_pause(inc);
1560 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1562 if (V_tcp_syncookies) {
1563 bzero(&scs, sizeof(scs));
1567 ("%s: bucket unexpectedly unlocked",
1571 (void) m_free(ipopts);
1578 if (!tfo_cookie_valid && tfo_response_cookie_valid)
1579 sc->sc_tfo_cookie = &tfo_response_cookie;
1582 * Fill in the syncache values.
1585 sc->sc_label = maclabel;
1589 sc->sc_ipopts = ipopts;
1590 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1592 if (!(inc->inc_flags & INC_ISIPV6))
1595 sc->sc_ip_tos = ip_tos;
1596 sc->sc_ip_ttl = ip_ttl;
1600 sc->sc_todctx = todctx;
1602 sc->sc_irs = th->th_seq;
1604 sc->sc_flowlabel = 0;
1607 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1608 * win was derived from socket earlier in the function.
1611 win = imin(win, TCP_MAXWIN);
1614 if (V_tcp_do_rfc1323) {
1616 * A timestamp received in a SYN makes
1617 * it ok to send timestamp requests and replies.
1619 if (to->to_flags & TOF_TS) {
1620 sc->sc_tsreflect = to->to_tsval;
1621 sc->sc_flags |= SCF_TIMESTAMP;
1622 sc->sc_tsoff = tcp_new_ts_offset(inc);
1624 if (to->to_flags & TOF_SCALE) {
1628 * Pick the smallest possible scaling factor that
1629 * will still allow us to scale up to sb_max, aka
1630 * kern.ipc.maxsockbuf.
1632 * We do this because there are broken firewalls that
1633 * will corrupt the window scale option, leading to
1634 * the other endpoint believing that our advertised
1635 * window is unscaled. At scale factors larger than
1636 * 5 the unscaled window will drop below 1500 bytes,
1637 * leading to serious problems when traversing these
1640 * With the default maxsockbuf of 256K, a scale factor
1641 * of 3 will be chosen by this algorithm. Those who
1642 * choose a larger maxsockbuf should watch out
1643 * for the compatibility problems mentioned above.
1645 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1646 * or <SYN,ACK>) segment itself is never scaled.
1648 while (wscale < TCP_MAX_WINSHIFT &&
1649 (TCP_MAXWIN << wscale) < sb_max)
1651 sc->sc_requested_r_scale = wscale;
1652 sc->sc_requested_s_scale = to->to_wscale;
1653 sc->sc_flags |= SCF_WINSCALE;
1656 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1658 * If listening socket requested TCP digests, flag this in the
1659 * syncache so that syncache_respond() will do the right thing
1662 if (ltflags & TF_SIGNATURE)
1663 sc->sc_flags |= SCF_SIGNATURE;
1664 #endif /* TCP_SIGNATURE */
1665 if (to->to_flags & TOF_SACKPERM)
1666 sc->sc_flags |= SCF_SACK;
1667 if (to->to_flags & TOF_MSS)
1668 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */
1669 if (ltflags & TF_NOOPT)
1670 sc->sc_flags |= SCF_NOOPT;
1671 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1672 sc->sc_flags |= SCF_ECN;
1674 if (V_tcp_syncookies)
1675 sc->sc_iss = syncookie_generate(sch, sc);
1677 sc->sc_iss = arc4random();
1679 if (autoflowlabel) {
1680 if (V_tcp_syncookies)
1681 sc->sc_flowlabel = sc->sc_iss;
1683 sc->sc_flowlabel = ip6_randomflowlabel();
1684 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1690 if (tfo_cookie_valid) {
1691 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie);
1692 /* INP_WUNLOCK(inp) will be performed by the caller */
1697 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1699 * Do a standard 3-way handshake.
1701 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1702 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1704 else if (sc != &scs)
1705 syncache_insert(sc, sch); /* locks and unlocks sch */
1706 TCPSTAT_INC(tcps_sndacks);
1707 TCPSTAT_INC(tcps_sndtotal);
1711 TCPSTAT_INC(tcps_sc_dropped);
1716 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1723 * If tfo_pending is not NULL here, then a TFO SYN that did not
1724 * result in a new socket was processed and the associated pending
1725 * counter has not yet been decremented. All such TFO processing paths
1726 * transit this point.
1728 if (tfo_pending != NULL)
1729 tcp_fastopen_decrement_counter(tfo_pending);
1736 mac_syncache_destroy(&maclabel);
1742 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment,
1743 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1746 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1748 struct ip *ip = NULL;
1750 struct tcphdr *th = NULL;
1751 int optlen, error = 0; /* Make compiler happy */
1752 u_int16_t hlen, tlen, mssopt;
1755 struct ip6_hdr *ip6 = NULL;
1762 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1765 tlen = hlen + sizeof(struct tcphdr);
1767 /* Determine MSS we advertize to other end of connection. */
1768 mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss);
1770 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1771 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1772 ("syncache: mbuf too small"));
1774 /* Create the IP+TCP header from scratch. */
1775 m = m_gethdr(M_NOWAIT, MT_DATA);
1779 mac_syncache_create_mbuf(sc->sc_label, m);
1781 m->m_data += max_linkhdr;
1783 m->m_pkthdr.len = tlen;
1784 m->m_pkthdr.rcvif = NULL;
1787 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1788 ip6 = mtod(m, struct ip6_hdr *);
1789 ip6->ip6_vfc = IPV6_VERSION;
1790 ip6->ip6_nxt = IPPROTO_TCP;
1791 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1792 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1793 ip6->ip6_plen = htons(tlen - hlen);
1794 /* ip6_hlim is set after checksum */
1795 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1796 ip6->ip6_flow |= sc->sc_flowlabel;
1798 th = (struct tcphdr *)(ip6 + 1);
1801 #if defined(INET6) && defined(INET)
1806 ip = mtod(m, struct ip *);
1807 ip->ip_v = IPVERSION;
1808 ip->ip_hl = sizeof(struct ip) >> 2;
1809 ip->ip_len = htons(tlen);
1813 ip->ip_p = IPPROTO_TCP;
1814 ip->ip_src = sc->sc_inc.inc_laddr;
1815 ip->ip_dst = sc->sc_inc.inc_faddr;
1816 ip->ip_ttl = sc->sc_ip_ttl;
1817 ip->ip_tos = sc->sc_ip_tos;
1820 * See if we should do MTU discovery. Route lookups are
1821 * expensive, so we will only unset the DF bit if:
1823 * 1) path_mtu_discovery is disabled
1824 * 2) the SCF_UNREACH flag has been set
1826 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1827 ip->ip_off |= htons(IP_DF);
1829 th = (struct tcphdr *)(ip + 1);
1832 th->th_sport = sc->sc_inc.inc_lport;
1833 th->th_dport = sc->sc_inc.inc_fport;
1836 th->th_seq = htonl(sc->sc_iss);
1838 th->th_seq = htonl(sc->sc_iss + 1);
1839 th->th_ack = htonl(sc->sc_irs + 1);
1840 th->th_off = sizeof(struct tcphdr) >> 2;
1842 th->th_flags = flags;
1843 th->th_win = htons(sc->sc_wnd);
1846 if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) {
1847 th->th_flags |= TH_ECE;
1848 TCPSTAT_INC(tcps_ecn_shs);
1851 /* Tack on the TCP options. */
1852 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1855 if (flags & TH_SYN) {
1857 to.to_flags = TOF_MSS;
1858 if (sc->sc_flags & SCF_WINSCALE) {
1859 to.to_wscale = sc->sc_requested_r_scale;
1860 to.to_flags |= TOF_SCALE;
1862 if (sc->sc_flags & SCF_SACK)
1863 to.to_flags |= TOF_SACKPERM;
1864 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1865 if (sc->sc_flags & SCF_SIGNATURE)
1866 to.to_flags |= TOF_SIGNATURE;
1868 if (sc->sc_tfo_cookie) {
1869 to.to_flags |= TOF_FASTOPEN;
1870 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1871 to.to_tfo_cookie = sc->sc_tfo_cookie;
1872 /* don't send cookie again when retransmitting response */
1873 sc->sc_tfo_cookie = NULL;
1876 if (sc->sc_flags & SCF_TIMESTAMP) {
1877 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1878 to.to_tsecr = sc->sc_tsreflect;
1879 to.to_flags |= TOF_TS;
1881 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1883 /* Adjust headers by option size. */
1884 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1886 m->m_pkthdr.len += optlen;
1888 if (sc->sc_inc.inc_flags & INC_ISIPV6)
1889 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1892 ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1893 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1894 if (sc->sc_flags & SCF_SIGNATURE) {
1895 KASSERT(to.to_flags & TOF_SIGNATURE,
1896 ("tcp_addoptions() didn't set tcp_signature"));
1898 /* NOTE: to.to_signature is inside of mbuf */
1899 if (!TCPMD5_ENABLED() ||
1900 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1909 M_SETFIB(m, sc->sc_inc.inc_fibnum);
1910 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1912 * If we have peer's SYN and it has a flowid, then let's assign it to
1913 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid
1914 * to SYN|ACK due to lack of inp here.
1916 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1917 m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1918 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1921 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1922 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1923 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1925 ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1927 if (ADDED_BY_TOE(sc)) {
1928 struct toedev *tod = sc->sc_tod;
1930 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1935 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
1936 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1939 #if defined(INET6) && defined(INET)
1944 m->m_pkthdr.csum_flags = CSUM_TCP;
1945 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1946 htons(tlen + optlen - hlen + IPPROTO_TCP));
1948 if (ADDED_BY_TOE(sc)) {
1949 struct toedev *tod = sc->sc_tod;
1951 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1956 TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
1957 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1964 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1965 * that exceed the capacity of the syncache by avoiding the storage of any
1966 * of the SYNs we receive. Syncookies defend against blind SYN flooding
1967 * attacks where the attacker does not have access to our responses.
1969 * Syncookies encode and include all necessary information about the
1970 * connection setup within the SYN|ACK that we send back. That way we
1971 * can avoid keeping any local state until the ACK to our SYN|ACK returns
1972 * (if ever). Normally the syncache and syncookies are running in parallel
1973 * with the latter taking over when the former is exhausted. When matching
1974 * syncache entry is found the syncookie is ignored.
1976 * The only reliable information persisting the 3WHS is our initial sequence
1977 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient
1978 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1979 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK
1980 * returns and signifies a legitimate connection if it matches the ACK.
1982 * The available space of 32 bits to store the hash and to encode the SYN
1983 * option information is very tight and we should have at least 24 bits for
1984 * the MAC to keep the number of guesses by blind spoofing reasonably high.
1986 * SYN option information we have to encode to fully restore a connection:
1987 * MSS: is imporant to chose an optimal segment size to avoid IP level
1988 * fragmentation along the path. The common MSS values can be encoded
1989 * in a 3-bit table. Uncommon values are captured by the next lower value
1990 * in the table leading to a slight increase in packetization overhead.
1991 * WSCALE: is necessary to allow large windows to be used for high delay-
1992 * bandwidth product links. Not scaling the window when it was initially
1993 * negotiated is bad for performance as lack of scaling further decreases
1994 * the apparent available send window. We only need to encode the WSCALE
1995 * we received from the remote end. Our end can be recalculated at any
1996 * time. The common WSCALE values can be encoded in a 3-bit table.
1997 * Uncommon values are captured by the next lower value in the table
1998 * making us under-estimate the available window size halving our
1999 * theoretically possible maximum throughput for that connection.
2000 * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2001 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2002 * that are included in all segments on a connection. We enable them when
2005 * Security of syncookies and attack vectors:
2007 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2008 * together with the gloabl secret to make it unique per connection attempt.
2009 * Thus any change of any of those parameters results in a different MAC output
2010 * in an unpredictable way unless a collision is encountered. 24 bits of the
2011 * MAC are embedded into the ISS.
2013 * To prevent replay attacks two rotating global secrets are updated with a
2014 * new random value every 15 seconds. The life-time of a syncookie is thus
2017 * Vector 1: Attacking the secret. This requires finding a weakness in the
2018 * MAC itself or the way it is used here. The attacker can do a chosen plain
2019 * text attack by varying and testing the all parameters under his control.
2020 * The strength depends on the size and randomness of the secret, and the
2021 * cryptographic security of the MAC function. Due to the constant updating
2022 * of the secret the attacker has at most 29.999 seconds to find the secret
2023 * and launch spoofed connections. After that he has to start all over again.
2025 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC
2026 * size an average of 4,823 attempts are required for a 50% chance of success
2027 * to spoof a single syncookie (birthday collision paradox). However the
2028 * attacker is blind and doesn't know if one of his attempts succeeded unless
2029 * he has a side channel to interfere success from. A single connection setup
2030 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2031 * This many attempts are required for each one blind spoofed connection. For
2032 * every additional spoofed connection he has to launch another N attempts.
2033 * Thus for a sustained rate 100 spoofed connections per second approximately
2034 * 1,800,000 packets per second would have to be sent.
2036 * NB: The MAC function should be fast so that it doesn't become a CPU
2037 * exhaustion attack vector itself.
2040 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2041 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2042 * http://cr.yp.to/syncookies.html (overview)
2043 * http://cr.yp.to/syncookies/archive (details)
2046 * Schematic construction of a syncookie enabled Initial Sequence Number:
2048 * 12345678901234567890123456789012
2049 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2051 * x 24 MAC (truncated)
2052 * W 3 Send Window Scale index
2054 * S 1 SACK permitted
2055 * P 1 Odd/even secret
2059 * Distribution and probability of certain MSS values. Those in between are
2060 * rounded down to the next lower one.
2061 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2062 * .2% .3% 5% 7% 7% 20% 15% 45%
2064 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2067 * Distribution and probability of certain WSCALE values. We have to map the
2068 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2069 * bits based on prevalence of certain values. Where we don't have an exact
2070 * match for are rounded down to the next lower one letting us under-estimate
2071 * the true available window. At the moment this would happen only for the
2072 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2073 * and window size). The absence of the WSCALE option (no scaling in either
2074 * direction) is encoded with index zero.
2075 * [WSCALE values histograms, Allman, 2012]
2076 * X 10 10 35 5 6 14 10% by host
2077 * X 11 4 5 5 18 49 3% by connections
2079 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2082 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed
2083 * and good cryptographic properties.
2086 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2087 uint8_t *secbits, uintptr_t secmod)
2090 uint32_t siphash[2];
2092 SipHash24_Init(&ctx);
2093 SipHash_SetKey(&ctx, secbits);
2094 switch (inc->inc_flags & INC_ISIPV6) {
2097 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2098 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2103 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2104 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2108 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2109 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2110 SipHash_Update(&ctx, &irs, sizeof(irs));
2111 SipHash_Update(&ctx, &flags, sizeof(flags));
2112 SipHash_Update(&ctx, &secmod, sizeof(secmod));
2113 SipHash_Final((u_int8_t *)&siphash, &ctx);
2115 return (siphash[0] ^ siphash[1]);
2119 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2121 u_int i, secbit, wscale;
2124 union syncookie cookie;
2128 /* Map our computed MSS into the 3-bit index. */
2129 for (i = nitems(tcp_sc_msstab) - 1;
2130 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2133 cookie.flags.mss_idx = i;
2136 * Map the send window scale into the 3-bit index but only if
2137 * the wscale option was received.
2139 if (sc->sc_flags & SCF_WINSCALE) {
2140 wscale = sc->sc_requested_s_scale;
2141 for (i = nitems(tcp_sc_wstab) - 1;
2142 tcp_sc_wstab[i] > wscale && i > 0;
2145 cookie.flags.wscale_idx = i;
2148 /* Can we do SACK? */
2149 if (sc->sc_flags & SCF_SACK)
2150 cookie.flags.sack_ok = 1;
2152 /* Which of the two secrets to use. */
2153 secbit = V_tcp_syncache.secret.oddeven & 0x1;
2154 cookie.flags.odd_even = secbit;
2156 secbits = V_tcp_syncache.secret.key[secbit];
2157 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2161 * Put the flags into the hash and XOR them to get better ISS number
2162 * variance. This doesn't enhance the cryptographic strength and is
2163 * done to prevent the 8 cookie bits from showing up directly on the
2167 iss |= cookie.cookie ^ (hash >> 24);
2169 TCPSTAT_INC(tcps_sc_sendcookie);
2173 static struct syncache *
2174 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2175 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2181 int wnd, wscale = 0;
2182 union syncookie cookie;
2185 * Pull information out of SYN-ACK/ACK and revert sequence number
2188 ack = th->th_ack - 1;
2189 seq = th->th_seq - 1;
2192 * Unpack the flags containing enough information to restore the
2195 cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2197 /* Which of the two secrets to use. */
2198 secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2200 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2202 /* The recomputed hash matches the ACK if this was a genuine cookie. */
2203 if ((ack & ~0xff) != (hash & ~0xff))
2206 /* Fill in the syncache values. */
2208 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2209 sc->sc_ipopts = NULL;
2214 switch (inc->inc_flags & INC_ISIPV6) {
2217 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2218 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2223 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2224 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
2229 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2231 /* We can simply recompute receive window scale we sent earlier. */
2232 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2235 /* Only use wscale if it was enabled in the orignal SYN. */
2236 if (cookie.flags.wscale_idx > 0) {
2237 sc->sc_requested_r_scale = wscale;
2238 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2239 sc->sc_flags |= SCF_WINSCALE;
2242 wnd = lso->sol_sbrcv_hiwat;
2244 wnd = imin(wnd, TCP_MAXWIN);
2247 if (cookie.flags.sack_ok)
2248 sc->sc_flags |= SCF_SACK;
2250 if (to->to_flags & TOF_TS) {
2251 sc->sc_flags |= SCF_TIMESTAMP;
2252 sc->sc_tsreflect = to->to_tsval;
2253 sc->sc_tsoff = tcp_new_ts_offset(inc);
2256 if (to->to_flags & TOF_SIGNATURE)
2257 sc->sc_flags |= SCF_SIGNATURE;
2261 TCPSTAT_INC(tcps_sc_recvcookie);
2267 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2268 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2271 struct syncache scs, *scx;
2274 bzero(&scs, sizeof(scs));
2275 scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
2277 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2281 if (sc->sc_peer_mss != scx->sc_peer_mss)
2282 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2283 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2285 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2286 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2287 s, __func__, sc->sc_requested_r_scale,
2288 scx->sc_requested_r_scale);
2290 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2291 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2292 s, __func__, sc->sc_requested_s_scale,
2293 scx->sc_requested_s_scale);
2295 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2296 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2303 #endif /* INVARIANTS */
2306 syncookie_reseed(void *arg)
2308 struct tcp_syncache *sc = arg;
2313 * Reseeding the secret doesn't have to be protected by a lock.
2314 * It only must be ensured that the new random values are visible
2315 * to all CPUs in a SMP environment. The atomic with release
2316 * semantics ensures that.
2318 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2319 secbits = sc->secret.key[secbit];
2320 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2321 atomic_add_rel_int(&sc->secret.oddeven, 1);
2323 /* Reschedule ourself. */
2324 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2328 * We have overflowed a bucket. Let's pause dealing with the syncache.
2329 * This function will increment the bucketoverflow statistics appropriately
2330 * (once per pause when pausing is enabled; otherwise, once per overflow).
2333 syncache_pause(struct in_conninfo *inc)
2339 * 2. Add sysctl read here so we don't get the benefit of this
2340 * change without the new sysctl.
2344 * Try an unlocked read. If we already know that another thread
2345 * has activated the feature, there is no need to proceed.
2347 if (V_tcp_syncache.paused)
2350 /* Are cookied enabled? If not, we can't pause. */
2351 if (!V_tcp_syncookies) {
2352 TCPSTAT_INC(tcps_sc_bucketoverflow);
2357 * We may be the first thread to find an overflow. Get the lock
2358 * and evaluate if we need to take action.
2360 mtx_lock(&V_tcp_syncache.pause_mtx);
2361 if (V_tcp_syncache.paused) {
2362 mtx_unlock(&V_tcp_syncache.pause_mtx);
2366 /* Activate protection. */
2367 V_tcp_syncache.paused = true;
2368 TCPSTAT_INC(tcps_sc_bucketoverflow);
2371 * Determine the last backoff time. If we are seeing a re-newed
2372 * attack within that same time after last reactivating the syncache,
2373 * consider it an extension of the same attack.
2375 delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2376 if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2377 if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2379 V_tcp_syncache.pause_backoff++;
2382 delta = TCP_SYNCACHE_PAUSE_TIME;
2383 V_tcp_syncache.pause_backoff = 0;
2386 /* Log a warning, including IP addresses, if able. */
2388 s = tcp_log_addrs(inc, NULL, NULL, NULL);
2390 s = (const char *)NULL;
2391 log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2392 "the next %lld seconds%s%s%s\n", (long long)delta,
2393 (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2394 (s != NULL) ? ")" : "");
2395 free(__DECONST(void *, s), M_TCPLOG);
2397 /* Use the calculated delta to set a new pause time. */
2398 V_tcp_syncache.pause_until = time_uptime + delta;
2399 callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2401 mtx_unlock(&V_tcp_syncache.pause_mtx);
2404 /* Evaluate whether we need to unpause. */
2406 syncache_unpause(void *arg)
2408 struct tcp_syncache *sc;
2412 mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2413 callout_deactivate(&sc->pause_co);
2416 * Check to make sure we are not running early. If the pause
2417 * time has expired, then deactivate the protection.
2419 if ((delta = sc->pause_until - time_uptime) > 0)
2420 callout_schedule(&sc->pause_co, delta * hz);
2426 * Exports the syncache entries to userland so that netstat can display
2427 * them alongside the other sockets. This function is intended to be
2428 * called only from tcp_pcblist.
2430 * Due to concurrency on an active system, the number of pcbs exported
2431 * may have no relation to max_pcbs. max_pcbs merely indicates the
2432 * amount of space the caller allocated for this function to use.
2435 syncache_pcblist(struct sysctl_req *req)
2438 struct syncache *sc;
2439 struct syncache_head *sch;
2442 bzero(&xt, sizeof(xt));
2443 xt.xt_len = sizeof(xt);
2444 xt.t_state = TCPS_SYN_RECEIVED;
2445 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2446 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2447 xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2448 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2450 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2451 sch = &V_tcp_syncache.hashbase[i];
2453 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2454 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2456 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2457 xt.xt_inp.inp_vflag = INP_IPV6;
2459 xt.xt_inp.inp_vflag = INP_IPV4;
2460 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2461 sizeof (struct in_conninfo));
2462 error = SYSCTL_OUT(req, &xt, sizeof xt);