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>
99 #include <netinet/udp.h>
100 #include <netinet/udp_var.h>
102 #include <netipsec/ipsec_support.h>
104 #include <machine/in_cksum.h>
106 #include <security/mac/mac_framework.h>
108 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1;
109 #define V_tcp_syncookies VNET(tcp_syncookies)
110 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
111 &VNET_NAME(tcp_syncookies), 0,
112 "Use TCP SYN cookies if the syncache overflows");
114 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0;
115 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly)
116 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
117 &VNET_NAME(tcp_syncookiesonly), 0,
118 "Use only TCP SYN cookies");
120 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1;
121 #define V_functions_inherit_listen_socket_stack \
122 VNET(functions_inherit_listen_socket_stack)
123 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack,
124 CTLFLAG_VNET | CTLFLAG_RW,
125 &VNET_NAME(functions_inherit_listen_socket_stack), 0,
126 "Inherit listen socket's stack");
129 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
132 static void syncache_drop(struct syncache *, struct syncache_head *);
133 static void syncache_free(struct syncache *);
134 static void syncache_insert(struct syncache *, struct syncache_head *);
135 static int syncache_respond(struct syncache *, const struct mbuf *, int);
136 static struct socket *syncache_socket(struct syncache *, struct socket *,
138 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch,
140 static void syncache_timer(void *);
142 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
143 uint8_t *, uintptr_t);
144 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *);
145 static struct syncache
146 *syncookie_lookup(struct in_conninfo *, struct syncache_head *,
147 struct syncache *, struct tcphdr *, struct tcpopt *,
148 struct socket *, uint16_t);
149 static void syncache_pause(struct in_conninfo *);
150 static void syncache_unpause(void *);
151 static void syncookie_reseed(void *);
153 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
154 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
155 struct socket *lso, uint16_t port);
159 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
160 * 3 retransmits corresponds to a timeout with default values of
161 * tcp_rexmit_initial * ( 1 +
164 * tcp_backoff[3]) + 3 * tcp_rexmit_slop,
165 * 1000 ms * (1 + 2 + 4 + 8) + 3 * 200 ms = 15600 ms,
166 * the odds are that the user has given up attempting to connect by then.
168 #define SYNCACHE_MAXREXMTS 3
170 /* Arbitrary values */
171 #define TCP_SYNCACHE_HASHSIZE 512
172 #define TCP_SYNCACHE_BUCKETLIMIT 30
174 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
175 #define V_tcp_syncache VNET(tcp_syncache)
177 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache,
178 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
181 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
182 &VNET_NAME(tcp_syncache.bucket_limit), 0,
183 "Per-bucket hash limit for syncache");
185 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
186 &VNET_NAME(tcp_syncache.cache_limit), 0,
187 "Overall entry limit for syncache");
189 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
190 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
192 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
193 &VNET_NAME(tcp_syncache.hashsize), 0,
194 "Size of TCP syncache hashtable");
196 SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET |
197 CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0,
198 "All syncache(4) entries are visible, ignoring UID/GID, jail(2) "
199 "and mac(4) checks");
202 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
207 new = V_tcp_syncache.rexmt_limit;
208 error = sysctl_handle_int(oidp, &new, 0, req);
209 if ((error == 0) && (req->newptr != NULL)) {
210 if (new > TCP_MAXRXTSHIFT)
213 V_tcp_syncache.rexmt_limit = new;
218 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
219 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
220 &VNET_NAME(tcp_syncache.rexmt_limit), 0,
221 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
222 "Limit on SYN/ACK retransmissions");
224 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
225 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
226 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
227 "Send reset on socket allocation failure");
229 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
231 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx)
232 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx)
233 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED)
236 * Requires the syncache entry to be already removed from the bucket list.
239 syncache_free(struct syncache *sc)
243 (void) m_free(sc->sc_ipopts);
247 mac_syncache_destroy(&sc->sc_label);
250 uma_zfree(V_tcp_syncache.zone, sc);
258 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
259 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
260 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
261 V_tcp_syncache.hash_secret = arc4random();
263 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
264 &V_tcp_syncache.hashsize);
265 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
266 &V_tcp_syncache.bucket_limit);
267 if (!powerof2(V_tcp_syncache.hashsize) ||
268 V_tcp_syncache.hashsize == 0) {
269 printf("WARNING: syncache hash size is not a power of 2.\n");
270 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
272 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
275 V_tcp_syncache.cache_limit =
276 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
277 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
278 &V_tcp_syncache.cache_limit);
280 /* Allocate the hash table. */
281 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
282 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
285 V_tcp_syncache.vnet = curvnet;
288 /* Initialize the hash buckets. */
289 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
290 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
291 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
293 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
294 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
295 V_tcp_syncache.hashbase[i].sch_length = 0;
296 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
297 V_tcp_syncache.hashbase[i].sch_last_overflow =
298 -(SYNCOOKIE_LIFETIME + 1);
301 /* Create the syncache entry zone. */
302 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
303 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
304 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
305 V_tcp_syncache.cache_limit);
307 /* Start the SYN cookie reseeder callout. */
308 callout_init(&V_tcp_syncache.secret.reseed, 1);
309 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
310 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
311 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
312 syncookie_reseed, &V_tcp_syncache);
314 /* Initialize the pause machinery. */
315 mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF);
316 callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx,
318 V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME;
319 V_tcp_syncache.pause_backoff = 0;
320 V_tcp_syncache.paused = false;
325 syncache_destroy(void)
327 struct syncache_head *sch;
328 struct syncache *sc, *nsc;
332 * Stop the re-seed timer before freeing resources. No need to
333 * possibly schedule it another time.
335 callout_drain(&V_tcp_syncache.secret.reseed);
337 /* Stop the SYN cache pause callout. */
338 mtx_lock(&V_tcp_syncache.pause_mtx);
339 if (callout_stop(&V_tcp_syncache.pause_co) == 0) {
340 mtx_unlock(&V_tcp_syncache.pause_mtx);
341 callout_drain(&V_tcp_syncache.pause_co);
343 mtx_unlock(&V_tcp_syncache.pause_mtx);
345 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */
346 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
347 sch = &V_tcp_syncache.hashbase[i];
348 callout_drain(&sch->sch_timer);
351 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
352 syncache_drop(sc, sch);
354 KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
355 ("%s: sch->sch_bucket not empty", __func__));
356 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
357 __func__, sch->sch_length));
358 mtx_destroy(&sch->sch_mtx);
361 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
362 ("%s: cache_count not 0", __func__));
364 /* Free the allocated global resources. */
365 uma_zdestroy(V_tcp_syncache.zone);
366 free(V_tcp_syncache.hashbase, M_SYNCACHE);
367 mtx_destroy(&V_tcp_syncache.pause_mtx);
372 * Inserts a syncache entry into the specified bucket row.
373 * Locks and unlocks the syncache_head autonomously.
376 syncache_insert(struct syncache *sc, struct syncache_head *sch)
378 struct syncache *sc2;
383 * Make sure that we don't overflow the per-bucket limit.
384 * If the bucket is full, toss the oldest element.
386 if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
387 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
388 ("sch->sch_length incorrect"));
389 syncache_pause(&sc->sc_inc);
390 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
391 sch->sch_last_overflow = time_uptime;
392 syncache_drop(sc2, sch);
395 /* Put it into the bucket. */
396 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
400 if (ADDED_BY_TOE(sc)) {
401 struct toedev *tod = sc->sc_tod;
403 tod->tod_syncache_added(tod, sc->sc_todctx);
407 /* Reinitialize the bucket row's timer. */
408 if (sch->sch_length == 1)
409 sch->sch_nextc = ticks + INT_MAX;
410 syncache_timeout(sc, sch, 1);
414 TCPSTATES_INC(TCPS_SYN_RECEIVED);
415 TCPSTAT_INC(tcps_sc_added);
419 * Remove and free entry from syncache bucket row.
420 * Expects locked syncache head.
423 syncache_drop(struct syncache *sc, struct syncache_head *sch)
426 SCH_LOCK_ASSERT(sch);
428 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
429 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
433 if (ADDED_BY_TOE(sc)) {
434 struct toedev *tod = sc->sc_tod;
436 tod->tod_syncache_removed(tod, sc->sc_todctx);
444 * Engage/reengage time on bucket row.
447 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
451 if (sc->sc_rxmits == 0)
452 rexmt = tcp_rexmit_initial;
455 tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
456 tcp_rexmit_min, TCPTV_REXMTMAX);
457 sc->sc_rxttime = ticks + rexmt;
459 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
460 sch->sch_nextc = sc->sc_rxttime;
462 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
463 syncache_timer, (void *)sch);
468 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
469 * If we have retransmitted an entry the maximum number of times, expire it.
470 * One separate timer for each bucket row.
473 syncache_timer(void *xsch)
475 struct syncache_head *sch = (struct syncache_head *)xsch;
476 struct syncache *sc, *nsc;
477 struct epoch_tracker et;
482 CURVNET_SET(sch->sch_sc->vnet);
484 /* NB: syncache_head has already been locked by the callout. */
485 SCH_LOCK_ASSERT(sch);
488 * In the following cycle we may remove some entries and/or
489 * advance some timeouts, so re-initialize the bucket timer.
491 sch->sch_nextc = tick + INT_MAX;
494 * If we have paused processing, unconditionally remove
495 * all syncache entries.
497 mtx_lock(&V_tcp_syncache.pause_mtx);
498 paused = V_tcp_syncache.paused;
499 mtx_unlock(&V_tcp_syncache.pause_mtx);
501 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
503 syncache_drop(sc, sch);
507 * We do not check if the listen socket still exists
508 * and accept the case where the listen socket may be
509 * gone by the time we resend the SYN/ACK. We do
510 * not expect this to happens often. If it does,
511 * then the RST will be sent by the time the remote
512 * host does the SYN/ACK->ACK.
514 if (TSTMP_GT(sc->sc_rxttime, tick)) {
515 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
516 sch->sch_nextc = sc->sc_rxttime;
519 if (sc->sc_rxmits > V_tcp_ecn_maxretries) {
520 sc->sc_flags &= ~SCF_ECN;
522 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
523 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
524 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
525 "giving up and removing syncache entry\n",
529 syncache_drop(sc, sch);
530 TCPSTAT_INC(tcps_sc_stale);
533 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
534 log(LOG_DEBUG, "%s; %s: Response timeout, "
535 "retransmitting (%u) SYN|ACK\n",
536 s, __func__, sc->sc_rxmits);
541 syncache_respond(sc, NULL, TH_SYN|TH_ACK);
543 TCPSTAT_INC(tcps_sc_retransmitted);
544 syncache_timeout(sc, sch, 0);
546 if (!TAILQ_EMPTY(&(sch)->sch_bucket))
547 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
548 syncache_timer, (void *)(sch));
553 * Returns true if the system is only using cookies at the moment.
554 * This could be due to a sysadmin decision to only use cookies, or it
555 * could be due to the system detecting an attack.
558 syncache_cookiesonly(void)
561 return (V_tcp_syncookies && (V_tcp_syncache.paused ||
562 V_tcp_syncookiesonly));
566 * Find the hash bucket for the given connection.
568 static struct syncache_head *
569 syncache_hashbucket(struct in_conninfo *inc)
574 * The hash is built on foreign port + local port + foreign address.
575 * We rely on the fact that struct in_conninfo starts with 16 bits
576 * of foreign port, then 16 bits of local port then followed by 128
577 * bits of foreign address. In case of IPv4 address, the first 3
578 * 32-bit words of the address always are zeroes.
580 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
581 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
583 return (&V_tcp_syncache.hashbase[hash]);
587 * Find an entry in the syncache.
588 * Returns always with locked syncache_head plus a matching entry or NULL.
590 static struct syncache *
591 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
594 struct syncache_head *sch;
596 *schp = sch = syncache_hashbucket(inc);
599 /* Circle through bucket row to find matching entry. */
600 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
601 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
602 sizeof(struct in_endpoints)) == 0)
605 return (sc); /* Always returns with locked sch. */
609 * This function is called when we get a RST for a
610 * non-existent connection, so that we can see if the
611 * connection is in the syn cache. If it is, zap it.
612 * If required send a challenge ACK.
615 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m,
619 struct syncache_head *sch;
622 if (syncache_cookiesonly())
624 sc = syncache_lookup(inc, &sch); /* returns locked sch */
625 SCH_LOCK_ASSERT(sch);
628 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
629 * See RFC 793 page 65, section SEGMENT ARRIVES.
631 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
632 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
633 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
634 "FIN flag set, segment ignored\n", s, __func__);
635 TCPSTAT_INC(tcps_badrst);
640 * No corresponding connection was found in syncache.
641 * If syncookies are enabled and possibly exclusively
642 * used, or we are under memory pressure, a valid RST
643 * may not find a syncache entry. In that case we're
644 * done and no SYN|ACK retransmissions will happen.
645 * Otherwise the RST was misdirected or spoofed.
648 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
649 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
650 "syncache entry (possibly syncookie only), "
651 "segment ignored\n", s, __func__);
652 TCPSTAT_INC(tcps_badrst);
656 /* The remote UDP encaps port does not match. */
657 if (sc->sc_port != port) {
658 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
659 log(LOG_DEBUG, "%s; %s: Spurious RST with matching "
660 "syncache entry but non-matching UDP encaps port, "
661 "segment ignored\n", s, __func__);
662 TCPSTAT_INC(tcps_badrst);
667 * If the RST bit is set, check the sequence number to see
668 * if this is a valid reset segment.
671 * In all states except SYN-SENT, all reset (RST) segments
672 * are validated by checking their SEQ-fields. A reset is
673 * valid if its sequence number is in the window.
676 * There are four cases for the acceptability test for an incoming
679 * Segment Receive Test
681 * ------- ------- -------------------------------------------
682 * 0 0 SEG.SEQ = RCV.NXT
683 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
684 * >0 0 not acceptable
685 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
686 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
688 * Note that when receiving a SYN segment in the LISTEN state,
689 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
690 * described in RFC 793, page 66.
692 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
693 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
694 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
695 if (V_tcp_insecure_rst ||
696 th->th_seq == sc->sc_irs + 1) {
697 syncache_drop(sc, sch);
698 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
700 "%s; %s: Our SYN|ACK was rejected, "
701 "connection attempt aborted by remote "
704 TCPSTAT_INC(tcps_sc_reset);
706 TCPSTAT_INC(tcps_badrst);
707 /* Send challenge ACK. */
708 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
709 log(LOG_DEBUG, "%s; %s: RST with invalid "
710 " SEQ %u != NXT %u (+WND %u), "
711 "sending challenge ACK\n",
713 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
714 syncache_respond(sc, m, TH_ACK);
717 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
718 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
719 "NXT %u (+WND %u), segment ignored\n",
721 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
722 TCPSTAT_INC(tcps_badrst);
732 syncache_badack(struct in_conninfo *inc, uint16_t port)
735 struct syncache_head *sch;
737 if (syncache_cookiesonly())
739 sc = syncache_lookup(inc, &sch); /* returns locked sch */
740 SCH_LOCK_ASSERT(sch);
741 if ((sc != NULL) && (sc->sc_port == port)) {
742 syncache_drop(sc, sch);
743 TCPSTAT_INC(tcps_sc_badack);
749 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port)
752 struct syncache_head *sch;
754 if (syncache_cookiesonly())
756 sc = syncache_lookup(inc, &sch); /* returns locked sch */
757 SCH_LOCK_ASSERT(sch);
761 /* If the port != sc_port, then it's a bogus ICMP msg */
762 if (port != sc->sc_port)
765 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
766 if (ntohl(th_seq) != sc->sc_iss)
770 * If we've rertransmitted 3 times and this is our second error,
771 * we remove the entry. Otherwise, we allow it to continue on.
772 * This prevents us from incorrectly nuking an entry during a
773 * spurious network outage.
777 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
778 sc->sc_flags |= SCF_UNREACH;
781 syncache_drop(sc, sch);
782 TCPSTAT_INC(tcps_sc_unreach);
788 * Build a new TCP socket structure from a syncache entry.
790 * On success return the newly created socket with its underlying inp locked.
792 static struct socket *
793 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
795 struct tcp_function_block *blk;
796 struct inpcb *inp = NULL;
805 * Ok, create the full blown connection, and set things up
806 * as they would have been set up if we had created the
807 * connection when the SYN arrived. If we can't create
808 * the connection, abort it.
810 so = sonewconn(lso, 0);
813 * Drop the connection; we will either send a RST or
814 * have the peer retransmit its SYN again after its
817 TCPSTAT_INC(tcps_listendrop);
818 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
819 log(LOG_DEBUG, "%s; %s: Socket create failed "
820 "due to limits or memory shortage\n",
827 mac_socketpeer_set_from_mbuf(m, so);
831 inp->inp_inc.inc_fibnum = so->so_fibnum;
834 * Exclusive pcbinfo lock is not required in syncache socket case even
835 * if two inpcb locks can be acquired simultaneously:
836 * - the inpcb in LISTEN state,
837 * - the newly created inp.
839 * In this case, an inp cannot be at same time in LISTEN state and
840 * just created by an accept() call.
842 INP_HASH_WLOCK(&V_tcbinfo);
844 /* Insert new socket into PCB hash list. */
845 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
847 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
848 inp->inp_vflag &= ~INP_IPV4;
849 inp->inp_vflag |= INP_IPV6;
850 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
852 inp->inp_vflag &= ~INP_IPV6;
853 inp->inp_vflag |= INP_IPV4;
855 inp->inp_ip_ttl = sc->sc_ip_ttl;
856 inp->inp_ip_tos = sc->sc_ip_tos;
857 inp->inp_laddr = sc->sc_inc.inc_laddr;
863 * If there's an mbuf and it has a flowid, then let's initialise the
864 * inp with that particular flowid.
866 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
867 inp->inp_flowid = m->m_pkthdr.flowid;
868 inp->inp_flowtype = M_HASHTYPE_GET(m);
870 inp->inp_numa_domain = m->m_pkthdr.numa_domain;
874 inp->inp_lport = sc->sc_inc.inc_lport;
876 if (inp->inp_vflag & INP_IPV6PROTO) {
877 struct inpcb *oinp = sotoinpcb(lso);
880 * Inherit socket options from the listening socket.
881 * Note that in6p_inputopts are not (and should not be)
882 * copied, since it stores previously received options and is
883 * used to detect if each new option is different than the
884 * previous one and hence should be passed to a user.
885 * If we copied in6p_inputopts, a user would not be able to
886 * receive options just after calling the accept system call.
888 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
889 if (oinp->in6p_outputopts)
890 inp->in6p_outputopts =
891 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
892 inp->in6p_hops = oinp->in6p_hops;
895 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
896 struct in6_addr laddr6;
897 struct sockaddr_in6 sin6;
899 sin6.sin6_family = AF_INET6;
900 sin6.sin6_len = sizeof(sin6);
901 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
902 sin6.sin6_port = sc->sc_inc.inc_fport;
903 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
904 laddr6 = inp->in6p_laddr;
905 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
906 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
907 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
908 thread0.td_ucred, m, false)) != 0) {
909 inp->in6p_laddr = laddr6;
910 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
911 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
916 INP_HASH_WUNLOCK(&V_tcbinfo);
919 /* Override flowlabel from in6_pcbconnect. */
920 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
921 inp->inp_flow |= sc->sc_flowlabel;
924 #if defined(INET) && defined(INET6)
929 struct in_addr laddr;
930 struct sockaddr_in sin;
932 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
934 if (inp->inp_options == NULL) {
935 inp->inp_options = sc->sc_ipopts;
936 sc->sc_ipopts = NULL;
939 sin.sin_family = AF_INET;
940 sin.sin_len = sizeof(sin);
941 sin.sin_addr = sc->sc_inc.inc_faddr;
942 sin.sin_port = sc->sc_inc.inc_fport;
943 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
944 laddr = inp->inp_laddr;
945 if (inp->inp_laddr.s_addr == INADDR_ANY)
946 inp->inp_laddr = sc->sc_inc.inc_laddr;
947 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
948 thread0.td_ucred, m, false)) != 0) {
949 inp->inp_laddr = laddr;
950 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
951 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
956 INP_HASH_WUNLOCK(&V_tcbinfo);
961 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
962 /* Copy old policy into new socket's. */
963 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
964 printf("syncache_socket: could not copy policy\n");
966 INP_HASH_WUNLOCK(&V_tcbinfo);
968 tcp_state_change(tp, TCPS_SYN_RECEIVED);
969 tp->iss = sc->sc_iss;
970 tp->irs = sc->sc_irs;
971 tp->t_port = sc->sc_port;
974 blk = sototcpcb(lso)->t_fb;
975 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
977 * Our parents t_fb was not the default,
978 * we need to release our ref on tp->t_fb and
979 * pickup one on the new entry.
981 struct tcp_function_block *rblk;
983 rblk = find_and_ref_tcp_fb(blk);
984 KASSERT(rblk != NULL,
985 ("cannot find blk %p out of syncache?", blk));
986 if (tp->t_fb->tfb_tcp_fb_fini)
987 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
988 refcount_release(&tp->t_fb->tfb_refcnt);
991 * XXXrrs this is quite dangerous, it is possible
992 * for the new function to fail to init. We also
993 * are not asking if the handoff_is_ok though at
994 * the very start thats probalbly ok.
996 if (tp->t_fb->tfb_tcp_fb_init) {
997 (*tp->t_fb->tfb_tcp_fb_init)(tp);
1000 tp->snd_wl1 = sc->sc_irs;
1001 tp->snd_max = tp->iss + 1;
1002 tp->snd_nxt = tp->iss + 1;
1003 tp->rcv_up = sc->sc_irs + 1;
1004 tp->rcv_wnd = sc->sc_wnd;
1005 tp->rcv_adv += tp->rcv_wnd;
1006 tp->last_ack_sent = tp->rcv_nxt;
1008 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
1009 if (sc->sc_flags & SCF_NOOPT)
1010 tp->t_flags |= TF_NOOPT;
1012 if (sc->sc_flags & SCF_WINSCALE) {
1013 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
1014 tp->snd_scale = sc->sc_requested_s_scale;
1015 tp->request_r_scale = sc->sc_requested_r_scale;
1017 if (sc->sc_flags & SCF_TIMESTAMP) {
1018 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
1019 tp->ts_recent = sc->sc_tsreflect;
1020 tp->ts_recent_age = tcp_ts_getticks();
1021 tp->ts_offset = sc->sc_tsoff;
1023 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1024 if (sc->sc_flags & SCF_SIGNATURE)
1025 tp->t_flags |= TF_SIGNATURE;
1027 if (sc->sc_flags & SCF_SACK)
1028 tp->t_flags |= TF_SACK_PERMIT;
1031 if (sc->sc_flags & SCF_ECN)
1032 tp->t_flags2 |= TF2_ECN_PERMIT;
1035 * Set up MSS and get cached values from tcp_hostcache.
1036 * This might overwrite some of the defaults we just set.
1038 tcp_mss(tp, sc->sc_peer_mss);
1041 * If the SYN,ACK was retransmitted, indicate that CWND to be
1042 * limited to one segment in cc_conn_init().
1043 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
1045 if (sc->sc_rxmits > 1)
1050 * Allow a TOE driver to install its hooks. Note that we hold the
1051 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
1052 * new connection before the TOE driver has done its thing.
1054 if (ADDED_BY_TOE(sc)) {
1055 struct toedev *tod = sc->sc_tod;
1057 tod->tod_offload_socket(tod, sc->sc_todctx, so);
1061 * Copy and activate timers.
1063 tp->t_keepinit = sototcpcb(lso)->t_keepinit;
1064 tp->t_keepidle = sototcpcb(lso)->t_keepidle;
1065 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
1066 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1067 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1069 TCPSTAT_INC(tcps_accepts);
1081 * This function gets called when we receive an ACK for a
1082 * socket in the LISTEN state. We look up the connection
1083 * in the syncache, and if its there, we pull it out of
1084 * the cache and turn it into a full-blown connection in
1085 * the SYN-RECEIVED state.
1087 * On syncache_socket() success the newly created socket
1088 * has its underlying inp locked.
1091 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1092 struct socket **lsop, struct mbuf *m, uint16_t port)
1094 struct syncache *sc;
1095 struct syncache_head *sch;
1096 struct syncache scs;
1101 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1102 ("%s: can handle only ACK", __func__));
1104 if (syncache_cookiesonly()) {
1106 sch = syncache_hashbucket(inc);
1109 sc = syncache_lookup(inc, &sch); /* returns locked sch */
1111 SCH_LOCK_ASSERT(sch);
1116 * Test code for syncookies comparing the syncache stored
1117 * values with the reconstructed values from the cookie.
1120 syncookie_cmp(inc, sch, sc, th, to, *lsop, port);
1125 * There is no syncache entry, so see if this ACK is
1126 * a returning syncookie. To do this, first:
1127 * A. Check if syncookies are used in case of syncache
1129 * B. See if this socket has had a syncache entry dropped in
1130 * the recent past. We don't want to accept a bogus
1131 * syncookie if we've never received a SYN or accept it
1133 * C. check that the syncookie is valid. If it is, then
1134 * cobble up a fake syncache entry, and return.
1136 if (locked && !V_tcp_syncookies) {
1138 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1139 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1140 "segment rejected (syncookies disabled)\n",
1144 if (locked && !V_tcp_syncookiesonly &&
1145 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1147 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1148 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1149 "segment rejected (no syncache entry)\n",
1153 bzero(&scs, sizeof(scs));
1154 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port);
1158 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1159 log(LOG_DEBUG, "%s; %s: Segment failed "
1160 "SYNCOOKIE authentication, segment rejected "
1161 "(probably spoofed)\n", s, __func__);
1164 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1165 /* If received ACK has MD5 signature, check it. */
1166 if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1167 (!TCPMD5_ENABLED() ||
1168 TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1170 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1171 log(LOG_DEBUG, "%s; %s: Segment rejected, "
1172 "MD5 signature doesn't match.\n",
1176 TCPSTAT_INC(tcps_sig_err_sigopt);
1177 return (-1); /* Do not send RST */
1179 #endif /* TCP_SIGNATURE */
1181 if (sc->sc_port != port) {
1185 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1187 * If listening socket requested TCP digests, check that
1188 * received ACK has signature and it is correct.
1189 * If not, drop the ACK and leave sc entry in th cache,
1190 * because SYN was received with correct signature.
1192 if (sc->sc_flags & SCF_SIGNATURE) {
1193 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1195 TCPSTAT_INC(tcps_sig_err_nosigopt);
1197 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1198 log(LOG_DEBUG, "%s; %s: Segment "
1199 "rejected, MD5 signature wasn't "
1200 "provided.\n", s, __func__);
1203 return (-1); /* Do not send RST */
1205 if (!TCPMD5_ENABLED() ||
1206 TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1207 /* Doesn't match or no SA */
1209 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1210 log(LOG_DEBUG, "%s; %s: Segment "
1211 "rejected, MD5 signature doesn't "
1212 "match.\n", s, __func__);
1215 return (-1); /* Do not send RST */
1218 #endif /* TCP_SIGNATURE */
1221 * RFC 7323 PAWS: If we have a timestamp on this segment and
1222 * it's less than ts_recent, drop it.
1223 * XXXMT: RFC 7323 also requires to send an ACK.
1224 * In tcp_input.c this is only done for TCP segments
1225 * with user data, so be consistent here and just drop
1228 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1229 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1231 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1233 "%s; %s: SEG.TSval %u < TS.Recent %u, "
1234 "segment dropped\n", s, __func__,
1235 to->to_tsval, sc->sc_tsreflect);
1238 return (-1); /* Do not send RST */
1242 * If timestamps were not negotiated during SYN/ACK and a
1243 * segment with a timestamp is received, ignore the
1244 * timestamp and process the packet normally.
1245 * See section 3.2 of RFC 7323.
1247 if (!(sc->sc_flags & SCF_TIMESTAMP) &&
1248 (to->to_flags & TOF_TS)) {
1249 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1250 log(LOG_DEBUG, "%s; %s: Timestamp not "
1251 "expected, segment processed normally\n",
1259 * If timestamps were negotiated during SYN/ACK and a
1260 * segment without a timestamp is received, silently drop
1261 * the segment, unless the missing timestamps are tolerated.
1262 * See section 3.2 of RFC 7323.
1264 if ((sc->sc_flags & SCF_TIMESTAMP) &&
1265 !(to->to_flags & TOF_TS)) {
1266 if (V_tcp_tolerate_missing_ts) {
1267 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1269 "%s; %s: Timestamp missing, "
1270 "segment processed normally\n",
1276 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1278 "%s; %s: Timestamp missing, "
1279 "segment silently dropped\n",
1283 return (-1); /* Do not send RST */
1288 * Pull out the entry to unlock the bucket row.
1290 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1291 * tcp_state_change(). The tcpcb is not existent at this
1292 * moment. A new one will be allocated via syncache_socket->
1293 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1294 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1296 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1297 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1300 if (ADDED_BY_TOE(sc)) {
1301 struct toedev *tod = sc->sc_tod;
1303 tod->tod_syncache_removed(tod, sc->sc_todctx);
1310 * Segment validation:
1311 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1313 if (th->th_ack != sc->sc_iss + 1) {
1314 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1315 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1316 "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1321 * The SEQ must fall in the window starting at the received
1322 * initial receive sequence number + 1 (the SYN).
1324 if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1325 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1326 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1327 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1328 "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1332 *lsop = syncache_socket(sc, *lsop, m);
1335 TCPSTAT_INC(tcps_sc_aborted);
1337 TCPSTAT_INC(tcps_sc_completed);
1339 /* how do we find the inp for the new socket? */
1344 if (sc != NULL && sc != &scs)
1352 static struct socket *
1353 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m,
1354 uint64_t response_cookie)
1358 unsigned int *pending_counter;
1363 pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending;
1364 so = syncache_socket(sc, lso, m);
1366 TCPSTAT_INC(tcps_sc_aborted);
1367 atomic_subtract_int(pending_counter, 1);
1370 inp = sotoinpcb(so);
1371 tp = intotcpcb(inp);
1372 tp->t_flags |= TF_FASTOPEN;
1373 tp->t_tfo_cookie.server = response_cookie;
1374 tp->snd_max = tp->iss;
1375 tp->snd_nxt = tp->iss;
1376 tp->t_tfo_pending = pending_counter;
1377 TCPSTAT_INC(tcps_sc_completed);
1384 * Given a LISTEN socket and an inbound SYN request, add
1385 * this to the syn cache, and send back a segment:
1386 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1389 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1390 * Doing so would require that we hold onto the data and deliver it
1391 * to the application. However, if we are the target of a SYN-flood
1392 * DoS attack, an attacker could send data which would eventually
1393 * consume all available buffer space if it were ACKed. By not ACKing
1394 * the data, we avoid this DoS scenario.
1396 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1397 * cookie is processed and a new socket is created. In this case, any data
1398 * accompanying the SYN will be queued to the socket by tcp_input() and will
1399 * be ACKed either when the application sends response data or the delayed
1400 * ACK timer expires, whichever comes first.
1403 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1404 struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod,
1405 void *todctx, uint8_t iptos, uint16_t port)
1408 struct socket *rv = NULL;
1409 struct syncache *sc = NULL;
1410 struct syncache_head *sch;
1411 struct mbuf *ipopts = NULL;
1413 int win, ip_ttl, ip_tos;
1416 int autoflowlabel = 0;
1419 struct label *maclabel;
1421 struct syncache scs;
1423 uint64_t tfo_response_cookie;
1424 unsigned int *tfo_pending = NULL;
1425 int tfo_cookie_valid = 0;
1426 int tfo_response_cookie_valid = 0;
1429 INP_RLOCK_ASSERT(inp); /* listen socket */
1430 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1431 ("%s: unexpected tcp flags", __func__));
1434 * Combine all so/tp operations very early to drop the INP lock as
1437 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1439 cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred);
1442 if (inc->inc_flags & INC_ISIPV6) {
1443 if (inp->inp_flags & IN6P_AUTOFLOWLABEL) {
1446 ip_ttl = in6_selecthlim(inp, NULL);
1447 if ((inp->in6p_outputopts == NULL) ||
1448 (inp->in6p_outputopts->ip6po_tclass == -1)) {
1451 ip_tos = inp->in6p_outputopts->ip6po_tclass;
1455 #if defined(INET6) && defined(INET)
1460 ip_ttl = inp->inp_ip_ttl;
1461 ip_tos = inp->inp_ip_tos;
1464 win = so->sol_sbrcv_hiwat;
1465 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1467 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
1468 (tp->t_tfo_pending != NULL) &&
1469 (to->to_flags & TOF_FASTOPEN)) {
1471 * Limit the number of pending TFO connections to
1472 * approximately half of the queue limit. This prevents TFO
1473 * SYN floods from starving the service by filling the
1474 * listen queue with bogus TFO connections.
1476 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1477 (so->sol_qlimit / 2)) {
1480 result = tcp_fastopen_check_cookie(inc,
1481 to->to_tfo_cookie, to->to_tfo_len,
1482 &tfo_response_cookie);
1483 tfo_cookie_valid = (result > 0);
1484 tfo_response_cookie_valid = (result >= 0);
1488 * Remember the TFO pending counter as it will have to be
1489 * decremented below if we don't make it to syncache_tfo_expand().
1491 tfo_pending = tp->t_tfo_pending;
1495 if (mac_syncache_init(&maclabel) != 0) {
1499 mac_syncache_create(maclabel, inp);
1501 if (!tfo_cookie_valid)
1505 * Remember the IP options, if any.
1508 if (!(inc->inc_flags & INC_ISIPV6))
1511 ipopts = (m) ? ip_srcroute(m) : NULL;
1516 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1518 * If listening socket requested TCP digests, check that received
1519 * SYN has signature and it is correct. If signature doesn't match
1520 * or TCP_SIGNATURE support isn't enabled, drop the packet.
1522 if (ltflags & TF_SIGNATURE) {
1523 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1524 TCPSTAT_INC(tcps_sig_err_nosigopt);
1527 if (!TCPMD5_ENABLED() ||
1528 TCPMD5_INPUT(m, th, to->to_signature) != 0)
1531 #endif /* TCP_SIGNATURE */
1533 * See if we already have an entry for this connection.
1534 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1536 * XXX: should the syncache be re-initialized with the contents
1537 * of the new SYN here (which may have different options?)
1539 * XXX: We do not check the sequence number to see if this is a
1540 * real retransmit or a new connection attempt. The question is
1541 * how to handle such a case; either ignore it as spoofed, or
1542 * drop the current entry and create a new one?
1544 if (syncache_cookiesonly()) {
1546 sch = syncache_hashbucket(inc);
1549 sc = syncache_lookup(inc, &sch); /* returns locked sch */
1551 SCH_LOCK_ASSERT(sch);
1554 if (tfo_cookie_valid)
1556 TCPSTAT_INC(tcps_sc_dupsyn);
1559 * If we were remembering a previous source route,
1560 * forget it and use the new one we've been given.
1563 (void) m_free(sc->sc_ipopts);
1564 sc->sc_ipopts = ipopts;
1567 * Update timestamp if present.
1569 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1570 sc->sc_tsreflect = to->to_tsval;
1572 sc->sc_flags &= ~SCF_TIMESTAMP;
1574 * Disable ECN if needed.
1576 if ((sc->sc_flags & SCF_ECN) &&
1577 ((th->th_flags & (TH_ECE|TH_CWR)) != (TH_ECE|TH_CWR))) {
1578 sc->sc_flags &= ~SCF_ECN;
1582 * Since we have already unconditionally allocated label
1583 * storage, free it up. The syncache entry will already
1584 * have an initialized label we can use.
1586 mac_syncache_destroy(&maclabel);
1588 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1589 /* Retransmit SYN|ACK and reset retransmit count. */
1590 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1591 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1592 "resetting timer and retransmitting SYN|ACK\n",
1596 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1598 syncache_timeout(sc, sch, 1);
1599 TCPSTAT_INC(tcps_sndacks);
1600 TCPSTAT_INC(tcps_sndtotal);
1606 if (tfo_cookie_valid) {
1607 bzero(&scs, sizeof(scs));
1613 * Skip allocating a syncache entry if we are just going to discard
1617 bzero(&scs, sizeof(scs));
1620 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1623 * The zone allocator couldn't provide more entries.
1624 * Treat this as if the cache was full; drop the oldest
1625 * entry and insert the new one.
1627 TCPSTAT_INC(tcps_sc_zonefail);
1628 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1629 sch->sch_last_overflow = time_uptime;
1630 syncache_drop(sc, sch);
1631 syncache_pause(inc);
1633 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1635 if (V_tcp_syncookies) {
1636 bzero(&scs, sizeof(scs));
1640 ("%s: bucket unexpectedly unlocked",
1644 (void) m_free(ipopts);
1651 if (!tfo_cookie_valid && tfo_response_cookie_valid)
1652 sc->sc_tfo_cookie = &tfo_response_cookie;
1655 * Fill in the syncache values.
1658 sc->sc_label = maclabel;
1663 sc->sc_ipopts = ipopts;
1664 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1665 sc->sc_ip_tos = ip_tos;
1666 sc->sc_ip_ttl = ip_ttl;
1669 sc->sc_todctx = todctx;
1671 sc->sc_irs = th->th_seq;
1673 sc->sc_flowlabel = 0;
1676 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1677 * win was derived from socket earlier in the function.
1680 win = imin(win, TCP_MAXWIN);
1683 if (V_tcp_do_rfc1323 &&
1684 !(ltflags & TF_NOOPT)) {
1686 * A timestamp received in a SYN makes
1687 * it ok to send timestamp requests and replies.
1689 if (to->to_flags & TOF_TS) {
1690 sc->sc_tsreflect = to->to_tsval;
1691 sc->sc_flags |= SCF_TIMESTAMP;
1692 sc->sc_tsoff = tcp_new_ts_offset(inc);
1694 if (to->to_flags & TOF_SCALE) {
1698 * Pick the smallest possible scaling factor that
1699 * will still allow us to scale up to sb_max, aka
1700 * kern.ipc.maxsockbuf.
1702 * We do this because there are broken firewalls that
1703 * will corrupt the window scale option, leading to
1704 * the other endpoint believing that our advertised
1705 * window is unscaled. At scale factors larger than
1706 * 5 the unscaled window will drop below 1500 bytes,
1707 * leading to serious problems when traversing these
1710 * With the default maxsockbuf of 256K, a scale factor
1711 * of 3 will be chosen by this algorithm. Those who
1712 * choose a larger maxsockbuf should watch out
1713 * for the compatibility problems mentioned above.
1715 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1716 * or <SYN,ACK>) segment itself is never scaled.
1718 while (wscale < TCP_MAX_WINSHIFT &&
1719 (TCP_MAXWIN << wscale) < sb_max)
1721 sc->sc_requested_r_scale = wscale;
1722 sc->sc_requested_s_scale = to->to_wscale;
1723 sc->sc_flags |= SCF_WINSCALE;
1726 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1728 * If listening socket requested TCP digests, flag this in the
1729 * syncache so that syncache_respond() will do the right thing
1732 if (ltflags & TF_SIGNATURE)
1733 sc->sc_flags |= SCF_SIGNATURE;
1734 #endif /* TCP_SIGNATURE */
1735 if (to->to_flags & TOF_SACKPERM)
1736 sc->sc_flags |= SCF_SACK;
1737 if (to->to_flags & TOF_MSS)
1738 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */
1739 if (ltflags & TF_NOOPT)
1740 sc->sc_flags |= SCF_NOOPT;
1741 if (((th->th_flags & (TH_ECE|TH_CWR)) == (TH_ECE|TH_CWR)) &&
1743 sc->sc_flags |= SCF_ECN;
1745 if (V_tcp_syncookies)
1746 sc->sc_iss = syncookie_generate(sch, sc);
1748 sc->sc_iss = arc4random();
1750 if (autoflowlabel) {
1751 if (V_tcp_syncookies)
1752 sc->sc_flowlabel = sc->sc_iss;
1754 sc->sc_flowlabel = ip6_randomflowlabel();
1755 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1761 if (tfo_cookie_valid) {
1762 rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie);
1763 /* INP_RUNLOCK(inp) will be performed by the caller */
1767 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1769 * Do a standard 3-way handshake.
1771 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1772 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1774 else if (sc != &scs)
1775 syncache_insert(sc, sch); /* locks and unlocks sch */
1776 TCPSTAT_INC(tcps_sndacks);
1777 TCPSTAT_INC(tcps_sndtotal);
1781 TCPSTAT_INC(tcps_sc_dropped);
1786 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1791 * If tfo_pending is not NULL here, then a TFO SYN that did not
1792 * result in a new socket was processed and the associated pending
1793 * counter has not yet been decremented. All such TFO processing paths
1794 * transit this point.
1796 if (tfo_pending != NULL)
1797 tcp_fastopen_decrement_counter(tfo_pending);
1804 mac_syncache_destroy(&maclabel);
1810 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment,
1811 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1814 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1816 struct ip *ip = NULL;
1818 struct tcphdr *th = NULL;
1819 struct udphdr *udp = NULL;
1820 int optlen, error = 0; /* Make compiler happy */
1821 u_int16_t hlen, tlen, mssopt, ulen;
1824 struct ip6_hdr *ip6 = NULL;
1831 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1834 tlen = hlen + sizeof(struct tcphdr);
1836 tlen += sizeof(struct udphdr);
1838 /* Determine MSS we advertize to other end of connection. */
1839 mssopt = tcp_mssopt(&sc->sc_inc);
1841 mssopt -= V_tcp_udp_tunneling_overhead;
1842 mssopt = max(mssopt, V_tcp_minmss);
1844 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1845 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1846 ("syncache: mbuf too small"));
1848 /* Create the IP+TCP header from scratch. */
1849 m = m_gethdr(M_NOWAIT, MT_DATA);
1853 mac_syncache_create_mbuf(sc->sc_label, m);
1855 m->m_data += max_linkhdr;
1857 m->m_pkthdr.len = tlen;
1858 m->m_pkthdr.rcvif = NULL;
1861 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1862 ip6 = mtod(m, struct ip6_hdr *);
1863 ip6->ip6_vfc = IPV6_VERSION;
1864 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1865 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1866 ip6->ip6_plen = htons(tlen - hlen);
1867 /* ip6_hlim is set after checksum */
1868 /* Zero out traffic class and flow label. */
1869 ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1870 ip6->ip6_flow |= sc->sc_flowlabel;
1871 if (sc->sc_port != 0) {
1872 ip6->ip6_nxt = IPPROTO_UDP;
1873 udp = (struct udphdr *)(ip6 + 1);
1874 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1875 udp->uh_dport = sc->sc_port;
1876 ulen = (tlen - sizeof(struct ip6_hdr));
1877 th = (struct tcphdr *)(udp + 1);
1879 ip6->ip6_nxt = IPPROTO_TCP;
1880 th = (struct tcphdr *)(ip6 + 1);
1882 ip6->ip6_flow |= htonl(sc->sc_ip_tos << 20);
1885 #if defined(INET6) && defined(INET)
1890 ip = mtod(m, struct ip *);
1891 ip->ip_v = IPVERSION;
1892 ip->ip_hl = sizeof(struct ip) >> 2;
1893 ip->ip_len = htons(tlen);
1897 ip->ip_src = sc->sc_inc.inc_laddr;
1898 ip->ip_dst = sc->sc_inc.inc_faddr;
1899 ip->ip_ttl = sc->sc_ip_ttl;
1900 ip->ip_tos = sc->sc_ip_tos;
1903 * See if we should do MTU discovery. Route lookups are
1904 * expensive, so we will only unset the DF bit if:
1906 * 1) path_mtu_discovery is disabled
1907 * 2) the SCF_UNREACH flag has been set
1909 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1910 ip->ip_off |= htons(IP_DF);
1911 if (sc->sc_port == 0) {
1912 ip->ip_p = IPPROTO_TCP;
1913 th = (struct tcphdr *)(ip + 1);
1915 ip->ip_p = IPPROTO_UDP;
1916 udp = (struct udphdr *)(ip + 1);
1917 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1918 udp->uh_dport = sc->sc_port;
1919 ulen = (tlen - sizeof(struct ip));
1920 th = (struct tcphdr *)(udp + 1);
1924 th->th_sport = sc->sc_inc.inc_lport;
1925 th->th_dport = sc->sc_inc.inc_fport;
1928 th->th_seq = htonl(sc->sc_iss);
1930 th->th_seq = htonl(sc->sc_iss + 1);
1931 th->th_ack = htonl(sc->sc_irs + 1);
1932 th->th_off = sizeof(struct tcphdr) >> 2;
1934 th->th_flags = flags;
1935 th->th_win = htons(sc->sc_wnd);
1938 if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) {
1939 th->th_flags |= TH_ECE;
1940 TCPSTAT_INC(tcps_ecn_shs);
1943 /* Tack on the TCP options. */
1944 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1947 if (flags & TH_SYN) {
1949 to.to_flags = TOF_MSS;
1950 if (sc->sc_flags & SCF_WINSCALE) {
1951 to.to_wscale = sc->sc_requested_r_scale;
1952 to.to_flags |= TOF_SCALE;
1954 if (sc->sc_flags & SCF_SACK)
1955 to.to_flags |= TOF_SACKPERM;
1956 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1957 if (sc->sc_flags & SCF_SIGNATURE)
1958 to.to_flags |= TOF_SIGNATURE;
1960 if (sc->sc_tfo_cookie) {
1961 to.to_flags |= TOF_FASTOPEN;
1962 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1963 to.to_tfo_cookie = sc->sc_tfo_cookie;
1964 /* don't send cookie again when retransmitting response */
1965 sc->sc_tfo_cookie = NULL;
1968 if (sc->sc_flags & SCF_TIMESTAMP) {
1969 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1970 to.to_tsecr = sc->sc_tsreflect;
1971 to.to_flags |= TOF_TS;
1973 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1975 /* Adjust headers by option size. */
1976 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1978 m->m_pkthdr.len += optlen;
1980 if (sc->sc_inc.inc_flags & INC_ISIPV6)
1981 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1984 ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1985 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1986 if (sc->sc_flags & SCF_SIGNATURE) {
1987 KASSERT(to.to_flags & TOF_SIGNATURE,
1988 ("tcp_addoptions() didn't set tcp_signature"));
1990 /* NOTE: to.to_signature is inside of mbuf */
1991 if (!TCPMD5_ENABLED() ||
1992 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
2003 udp->uh_ulen = htons(ulen);
2005 M_SETFIB(m, sc->sc_inc.inc_fibnum);
2007 * If we have peer's SYN and it has a flowid, then let's assign it to
2008 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid
2009 * to SYN|ACK due to lack of inp here.
2011 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
2012 m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
2013 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
2016 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
2018 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
2019 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2020 udp->uh_sum = in6_cksum_pseudo(ip6, ulen,
2022 th->th_sum = htons(0);
2024 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
2025 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2026 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
2029 ip6->ip6_hlim = sc->sc_ip_ttl;
2031 if (ADDED_BY_TOE(sc)) {
2032 struct toedev *tod = sc->sc_tod;
2034 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2039 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
2040 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
2043 #if defined(INET6) && defined(INET)
2049 m->m_pkthdr.csum_flags = CSUM_UDP;
2050 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2051 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
2052 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
2053 th->th_sum = htons(0);
2055 m->m_pkthdr.csum_flags = CSUM_TCP;
2056 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2057 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
2058 htons(tlen + optlen - hlen + IPPROTO_TCP));
2061 if (ADDED_BY_TOE(sc)) {
2062 struct toedev *tod = sc->sc_tod;
2064 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2069 TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
2070 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
2077 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
2078 * that exceed the capacity of the syncache by avoiding the storage of any
2079 * of the SYNs we receive. Syncookies defend against blind SYN flooding
2080 * attacks where the attacker does not have access to our responses.
2082 * Syncookies encode and include all necessary information about the
2083 * connection setup within the SYN|ACK that we send back. That way we
2084 * can avoid keeping any local state until the ACK to our SYN|ACK returns
2085 * (if ever). Normally the syncache and syncookies are running in parallel
2086 * with the latter taking over when the former is exhausted. When matching
2087 * syncache entry is found the syncookie is ignored.
2089 * The only reliable information persisting the 3WHS is our initial sequence
2090 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient
2091 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
2092 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK
2093 * returns and signifies a legitimate connection if it matches the ACK.
2095 * The available space of 32 bits to store the hash and to encode the SYN
2096 * option information is very tight and we should have at least 24 bits for
2097 * the MAC to keep the number of guesses by blind spoofing reasonably high.
2099 * SYN option information we have to encode to fully restore a connection:
2100 * MSS: is imporant to chose an optimal segment size to avoid IP level
2101 * fragmentation along the path. The common MSS values can be encoded
2102 * in a 3-bit table. Uncommon values are captured by the next lower value
2103 * in the table leading to a slight increase in packetization overhead.
2104 * WSCALE: is necessary to allow large windows to be used for high delay-
2105 * bandwidth product links. Not scaling the window when it was initially
2106 * negotiated is bad for performance as lack of scaling further decreases
2107 * the apparent available send window. We only need to encode the WSCALE
2108 * we received from the remote end. Our end can be recalculated at any
2109 * time. The common WSCALE values can be encoded in a 3-bit table.
2110 * Uncommon values are captured by the next lower value in the table
2111 * making us under-estimate the available window size halving our
2112 * theoretically possible maximum throughput for that connection.
2113 * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2114 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2115 * that are included in all segments on a connection. We enable them when
2118 * Security of syncookies and attack vectors:
2120 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2121 * together with the gloabl secret to make it unique per connection attempt.
2122 * Thus any change of any of those parameters results in a different MAC output
2123 * in an unpredictable way unless a collision is encountered. 24 bits of the
2124 * MAC are embedded into the ISS.
2126 * To prevent replay attacks two rotating global secrets are updated with a
2127 * new random value every 15 seconds. The life-time of a syncookie is thus
2130 * Vector 1: Attacking the secret. This requires finding a weakness in the
2131 * MAC itself or the way it is used here. The attacker can do a chosen plain
2132 * text attack by varying and testing the all parameters under his control.
2133 * The strength depends on the size and randomness of the secret, and the
2134 * cryptographic security of the MAC function. Due to the constant updating
2135 * of the secret the attacker has at most 29.999 seconds to find the secret
2136 * and launch spoofed connections. After that he has to start all over again.
2138 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC
2139 * size an average of 4,823 attempts are required for a 50% chance of success
2140 * to spoof a single syncookie (birthday collision paradox). However the
2141 * attacker is blind and doesn't know if one of his attempts succeeded unless
2142 * he has a side channel to interfere success from. A single connection setup
2143 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2144 * This many attempts are required for each one blind spoofed connection. For
2145 * every additional spoofed connection he has to launch another N attempts.
2146 * Thus for a sustained rate 100 spoofed connections per second approximately
2147 * 1,800,000 packets per second would have to be sent.
2149 * NB: The MAC function should be fast so that it doesn't become a CPU
2150 * exhaustion attack vector itself.
2153 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2154 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2155 * http://cr.yp.to/syncookies.html (overview)
2156 * http://cr.yp.to/syncookies/archive (details)
2159 * Schematic construction of a syncookie enabled Initial Sequence Number:
2161 * 12345678901234567890123456789012
2162 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2164 * x 24 MAC (truncated)
2165 * W 3 Send Window Scale index
2167 * S 1 SACK permitted
2168 * P 1 Odd/even secret
2172 * Distribution and probability of certain MSS values. Those in between are
2173 * rounded down to the next lower one.
2174 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2175 * .2% .3% 5% 7% 7% 20% 15% 45%
2177 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2180 * Distribution and probability of certain WSCALE values. We have to map the
2181 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2182 * bits based on prevalence of certain values. Where we don't have an exact
2183 * match for are rounded down to the next lower one letting us under-estimate
2184 * the true available window. At the moment this would happen only for the
2185 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2186 * and window size). The absence of the WSCALE option (no scaling in either
2187 * direction) is encoded with index zero.
2188 * [WSCALE values histograms, Allman, 2012]
2189 * X 10 10 35 5 6 14 10% by host
2190 * X 11 4 5 5 18 49 3% by connections
2192 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2195 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed
2196 * and good cryptographic properties.
2199 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2200 uint8_t *secbits, uintptr_t secmod)
2203 uint32_t siphash[2];
2205 SipHash24_Init(&ctx);
2206 SipHash_SetKey(&ctx, secbits);
2207 switch (inc->inc_flags & INC_ISIPV6) {
2210 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2211 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2216 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2217 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2221 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2222 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2223 SipHash_Update(&ctx, &irs, sizeof(irs));
2224 SipHash_Update(&ctx, &flags, sizeof(flags));
2225 SipHash_Update(&ctx, &secmod, sizeof(secmod));
2226 SipHash_Final((u_int8_t *)&siphash, &ctx);
2228 return (siphash[0] ^ siphash[1]);
2232 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2234 u_int i, secbit, wscale;
2237 union syncookie cookie;
2241 /* Map our computed MSS into the 3-bit index. */
2242 for (i = nitems(tcp_sc_msstab) - 1;
2243 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2246 cookie.flags.mss_idx = i;
2249 * Map the send window scale into the 3-bit index but only if
2250 * the wscale option was received.
2252 if (sc->sc_flags & SCF_WINSCALE) {
2253 wscale = sc->sc_requested_s_scale;
2254 for (i = nitems(tcp_sc_wstab) - 1;
2255 tcp_sc_wstab[i] > wscale && i > 0;
2258 cookie.flags.wscale_idx = i;
2261 /* Can we do SACK? */
2262 if (sc->sc_flags & SCF_SACK)
2263 cookie.flags.sack_ok = 1;
2265 /* Which of the two secrets to use. */
2266 secbit = V_tcp_syncache.secret.oddeven & 0x1;
2267 cookie.flags.odd_even = secbit;
2269 secbits = V_tcp_syncache.secret.key[secbit];
2270 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2274 * Put the flags into the hash and XOR them to get better ISS number
2275 * variance. This doesn't enhance the cryptographic strength and is
2276 * done to prevent the 8 cookie bits from showing up directly on the
2280 iss |= cookie.cookie ^ (hash >> 24);
2282 TCPSTAT_INC(tcps_sc_sendcookie);
2286 static struct syncache *
2287 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2288 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2289 struct socket *lso, uint16_t port)
2294 int wnd, wscale = 0;
2295 union syncookie cookie;
2298 * Pull information out of SYN-ACK/ACK and revert sequence number
2301 ack = th->th_ack - 1;
2302 seq = th->th_seq - 1;
2305 * Unpack the flags containing enough information to restore the
2308 cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2310 /* Which of the two secrets to use. */
2311 secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2313 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2315 /* The recomputed hash matches the ACK if this was a genuine cookie. */
2316 if ((ack & ~0xff) != (hash & ~0xff))
2319 /* Fill in the syncache values. */
2321 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2322 sc->sc_ipopts = NULL;
2327 switch (inc->inc_flags & INC_ISIPV6) {
2330 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2331 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2336 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2338 htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2343 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2345 /* We can simply recompute receive window scale we sent earlier. */
2346 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2349 /* Only use wscale if it was enabled in the orignal SYN. */
2350 if (cookie.flags.wscale_idx > 0) {
2351 sc->sc_requested_r_scale = wscale;
2352 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2353 sc->sc_flags |= SCF_WINSCALE;
2356 wnd = lso->sol_sbrcv_hiwat;
2358 wnd = imin(wnd, TCP_MAXWIN);
2361 if (cookie.flags.sack_ok)
2362 sc->sc_flags |= SCF_SACK;
2364 if (to->to_flags & TOF_TS) {
2365 sc->sc_flags |= SCF_TIMESTAMP;
2366 sc->sc_tsreflect = to->to_tsval;
2367 sc->sc_tsoff = tcp_new_ts_offset(inc);
2370 if (to->to_flags & TOF_SIGNATURE)
2371 sc->sc_flags |= SCF_SIGNATURE;
2377 TCPSTAT_INC(tcps_sc_recvcookie);
2383 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2384 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2385 struct socket *lso, uint16_t port)
2387 struct syncache scs, *scx;
2390 bzero(&scs, sizeof(scs));
2391 scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port);
2393 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2397 if (sc->sc_peer_mss != scx->sc_peer_mss)
2398 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2399 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2401 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2402 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2403 s, __func__, sc->sc_requested_r_scale,
2404 scx->sc_requested_r_scale);
2406 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2407 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2408 s, __func__, sc->sc_requested_s_scale,
2409 scx->sc_requested_s_scale);
2411 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2412 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2419 #endif /* INVARIANTS */
2422 syncookie_reseed(void *arg)
2424 struct tcp_syncache *sc = arg;
2429 * Reseeding the secret doesn't have to be protected by a lock.
2430 * It only must be ensured that the new random values are visible
2431 * to all CPUs in a SMP environment. The atomic with release
2432 * semantics ensures that.
2434 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2435 secbits = sc->secret.key[secbit];
2436 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2437 atomic_add_rel_int(&sc->secret.oddeven, 1);
2439 /* Reschedule ourself. */
2440 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2444 * We have overflowed a bucket. Let's pause dealing with the syncache.
2445 * This function will increment the bucketoverflow statistics appropriately
2446 * (once per pause when pausing is enabled; otherwise, once per overflow).
2449 syncache_pause(struct in_conninfo *inc)
2455 * 2. Add sysctl read here so we don't get the benefit of this
2456 * change without the new sysctl.
2460 * Try an unlocked read. If we already know that another thread
2461 * has activated the feature, there is no need to proceed.
2463 if (V_tcp_syncache.paused)
2466 /* Are cookied enabled? If not, we can't pause. */
2467 if (!V_tcp_syncookies) {
2468 TCPSTAT_INC(tcps_sc_bucketoverflow);
2473 * We may be the first thread to find an overflow. Get the lock
2474 * and evaluate if we need to take action.
2476 mtx_lock(&V_tcp_syncache.pause_mtx);
2477 if (V_tcp_syncache.paused) {
2478 mtx_unlock(&V_tcp_syncache.pause_mtx);
2482 /* Activate protection. */
2483 V_tcp_syncache.paused = true;
2484 TCPSTAT_INC(tcps_sc_bucketoverflow);
2487 * Determine the last backoff time. If we are seeing a re-newed
2488 * attack within that same time after last reactivating the syncache,
2489 * consider it an extension of the same attack.
2491 delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2492 if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2493 if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2495 V_tcp_syncache.pause_backoff++;
2498 delta = TCP_SYNCACHE_PAUSE_TIME;
2499 V_tcp_syncache.pause_backoff = 0;
2502 /* Log a warning, including IP addresses, if able. */
2504 s = tcp_log_addrs(inc, NULL, NULL, NULL);
2506 s = (const char *)NULL;
2507 log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2508 "the next %lld seconds%s%s%s\n", (long long)delta,
2509 (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2510 (s != NULL) ? ")" : "");
2511 free(__DECONST(void *, s), M_TCPLOG);
2513 /* Use the calculated delta to set a new pause time. */
2514 V_tcp_syncache.pause_until = time_uptime + delta;
2515 callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2517 mtx_unlock(&V_tcp_syncache.pause_mtx);
2520 /* Evaluate whether we need to unpause. */
2522 syncache_unpause(void *arg)
2524 struct tcp_syncache *sc;
2528 mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2529 callout_deactivate(&sc->pause_co);
2532 * Check to make sure we are not running early. If the pause
2533 * time has expired, then deactivate the protection.
2535 if ((delta = sc->pause_until - time_uptime) > 0)
2536 callout_schedule(&sc->pause_co, delta * hz);
2542 * Exports the syncache entries to userland so that netstat can display
2543 * them alongside the other sockets. This function is intended to be
2544 * called only from tcp_pcblist.
2546 * Due to concurrency on an active system, the number of pcbs exported
2547 * may have no relation to max_pcbs. max_pcbs merely indicates the
2548 * amount of space the caller allocated for this function to use.
2551 syncache_pcblist(struct sysctl_req *req)
2554 struct syncache *sc;
2555 struct syncache_head *sch;
2558 bzero(&xt, sizeof(xt));
2559 xt.xt_len = sizeof(xt);
2560 xt.t_state = TCPS_SYN_RECEIVED;
2561 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2562 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2563 xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2564 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2566 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2567 sch = &V_tcp_syncache.hashbase[i];
2569 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2570 if (sc->sc_cred != NULL &&
2571 cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2573 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2574 xt.xt_inp.inp_vflag = INP_IPV6;
2576 xt.xt_inp.inp_vflag = INP_IPV4;
2577 xt.xt_encaps_port = sc->sc_port;
2578 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2579 sizeof (struct in_conninfo));
2580 error = SYSCTL_OUT(req, &xt, sizeof xt);