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;
474 CURVNET_SET(sch->sch_sc->vnet);
476 /* NB: syncache_head has already been locked by the callout. */
477 SCH_LOCK_ASSERT(sch);
480 * In the following cycle we may remove some entries and/or
481 * advance some timeouts, so re-initialize the bucket timer.
483 sch->sch_nextc = tick + INT_MAX;
486 * If we have paused processing, unconditionally remove
487 * all syncache entries.
489 mtx_lock(&V_tcp_syncache.pause_mtx);
490 paused = V_tcp_syncache.paused;
491 mtx_unlock(&V_tcp_syncache.pause_mtx);
493 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
495 syncache_drop(sc, sch);
499 * We do not check if the listen socket still exists
500 * and accept the case where the listen socket may be
501 * gone by the time we resend the SYN/ACK. We do
502 * not expect this to happens often. If it does,
503 * then the RST will be sent by the time the remote
504 * host does the SYN/ACK->ACK.
506 if (TSTMP_GT(sc->sc_rxttime, tick)) {
507 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
508 sch->sch_nextc = sc->sc_rxttime;
511 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
512 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
513 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
514 "giving up and removing syncache entry\n",
518 syncache_drop(sc, sch);
519 TCPSTAT_INC(tcps_sc_stale);
522 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
523 log(LOG_DEBUG, "%s; %s: Response timeout, "
524 "retransmitting (%u) SYN|ACK\n",
525 s, __func__, sc->sc_rxmits);
529 syncache_respond(sc, NULL, TH_SYN|TH_ACK);
530 TCPSTAT_INC(tcps_sc_retransmitted);
531 syncache_timeout(sc, sch, 0);
533 if (!TAILQ_EMPTY(&(sch)->sch_bucket))
534 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
535 syncache_timer, (void *)(sch));
540 * Returns true if the system is only using cookies at the moment.
541 * This could be due to a sysadmin decision to only use cookies, or it
542 * could be due to the system detecting an attack.
545 syncache_cookiesonly(void)
548 return (V_tcp_syncookies && (V_tcp_syncache.paused ||
549 V_tcp_syncookiesonly));
553 * Find the hash bucket for the given connection.
555 static struct syncache_head *
556 syncache_hashbucket(struct in_conninfo *inc)
561 * The hash is built on foreign port + local port + foreign address.
562 * We rely on the fact that struct in_conninfo starts with 16 bits
563 * of foreign port, then 16 bits of local port then followed by 128
564 * bits of foreign address. In case of IPv4 address, the first 3
565 * 32-bit words of the address always are zeroes.
567 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
568 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
570 return (&V_tcp_syncache.hashbase[hash]);
574 * Find an entry in the syncache.
575 * Returns always with locked syncache_head plus a matching entry or NULL.
577 static struct syncache *
578 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
581 struct syncache_head *sch;
583 *schp = sch = syncache_hashbucket(inc);
586 /* Circle through bucket row to find matching entry. */
587 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
588 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
589 sizeof(struct in_endpoints)) == 0)
592 return (sc); /* Always returns with locked sch. */
596 * This function is called when we get a RST for a
597 * non-existent connection, so that we can see if the
598 * connection is in the syn cache. If it is, zap it.
599 * If required send a challenge ACK.
602 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m)
605 struct syncache_head *sch;
608 if (syncache_cookiesonly())
610 sc = syncache_lookup(inc, &sch); /* returns locked sch */
611 SCH_LOCK_ASSERT(sch);
614 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
615 * See RFC 793 page 65, section SEGMENT ARRIVES.
617 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
618 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
619 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
620 "FIN flag set, segment ignored\n", s, __func__);
621 TCPSTAT_INC(tcps_badrst);
626 * No corresponding connection was found in syncache.
627 * If syncookies are enabled and possibly exclusively
628 * used, or we are under memory pressure, a valid RST
629 * may not find a syncache entry. In that case we're
630 * done and no SYN|ACK retransmissions will happen.
631 * Otherwise the RST was misdirected or spoofed.
634 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
635 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
636 "syncache entry (possibly syncookie only), "
637 "segment ignored\n", s, __func__);
638 TCPSTAT_INC(tcps_badrst);
643 * If the RST bit is set, check the sequence number to see
644 * if this is a valid reset segment.
647 * In all states except SYN-SENT, all reset (RST) segments
648 * are validated by checking their SEQ-fields. A reset is
649 * valid if its sequence number is in the window.
652 * There are four cases for the acceptability test for an incoming
655 * Segment Receive Test
657 * ------- ------- -------------------------------------------
658 * 0 0 SEG.SEQ = RCV.NXT
659 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
660 * >0 0 not acceptable
661 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
662 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
664 * Note that when receiving a SYN segment in the LISTEN state,
665 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
666 * described in RFC 793, page 66.
668 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
669 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
670 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
671 if (V_tcp_insecure_rst ||
672 th->th_seq == sc->sc_irs + 1) {
673 syncache_drop(sc, sch);
674 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
676 "%s; %s: Our SYN|ACK was rejected, "
677 "connection attempt aborted by remote "
680 TCPSTAT_INC(tcps_sc_reset);
682 TCPSTAT_INC(tcps_badrst);
683 /* Send challenge ACK. */
684 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
685 log(LOG_DEBUG, "%s; %s: RST with invalid "
686 " SEQ %u != NXT %u (+WND %u), "
687 "sending challenge ACK\n",
689 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
690 syncache_respond(sc, m, TH_ACK);
693 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
694 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
695 "NXT %u (+WND %u), segment ignored\n",
697 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
698 TCPSTAT_INC(tcps_badrst);
708 syncache_badack(struct in_conninfo *inc)
711 struct syncache_head *sch;
713 if (syncache_cookiesonly())
715 sc = syncache_lookup(inc, &sch); /* returns locked sch */
716 SCH_LOCK_ASSERT(sch);
718 syncache_drop(sc, sch);
719 TCPSTAT_INC(tcps_sc_badack);
725 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq)
728 struct syncache_head *sch;
730 if (syncache_cookiesonly())
732 sc = syncache_lookup(inc, &sch); /* returns locked sch */
733 SCH_LOCK_ASSERT(sch);
737 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
738 if (ntohl(th_seq) != sc->sc_iss)
742 * If we've rertransmitted 3 times and this is our second error,
743 * we remove the entry. Otherwise, we allow it to continue on.
744 * This prevents us from incorrectly nuking an entry during a
745 * spurious network outage.
749 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
750 sc->sc_flags |= SCF_UNREACH;
753 syncache_drop(sc, sch);
754 TCPSTAT_INC(tcps_sc_unreach);
760 * Build a new TCP socket structure from a syncache entry.
762 * On success return the newly created socket with its underlying inp locked.
764 static struct socket *
765 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
767 struct tcp_function_block *blk;
768 struct inpcb *inp = NULL;
774 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
777 * Ok, create the full blown connection, and set things up
778 * as they would have been set up if we had created the
779 * connection when the SYN arrived. If we can't create
780 * the connection, abort it.
782 so = sonewconn(lso, 0);
785 * Drop the connection; we will either send a RST or
786 * have the peer retransmit its SYN again after its
789 TCPSTAT_INC(tcps_listendrop);
790 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
791 log(LOG_DEBUG, "%s; %s: Socket create failed "
792 "due to limits or memory shortage\n",
799 mac_socketpeer_set_from_mbuf(m, so);
803 inp->inp_inc.inc_fibnum = so->so_fibnum;
806 * Exclusive pcbinfo lock is not required in syncache socket case even
807 * if two inpcb locks can be acquired simultaneously:
808 * - the inpcb in LISTEN state,
809 * - the newly created inp.
811 * In this case, an inp cannot be at same time in LISTEN state and
812 * just created by an accept() call.
814 INP_HASH_WLOCK(&V_tcbinfo);
816 /* Insert new socket into PCB hash list. */
817 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
819 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
820 inp->inp_vflag &= ~INP_IPV4;
821 inp->inp_vflag |= INP_IPV6;
822 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
824 inp->inp_vflag &= ~INP_IPV6;
825 inp->inp_vflag |= INP_IPV4;
827 inp->inp_laddr = sc->sc_inc.inc_laddr;
833 * If there's an mbuf and it has a flowid, then let's initialise the
834 * inp with that particular flowid.
836 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
837 inp->inp_flowid = m->m_pkthdr.flowid;
838 inp->inp_flowtype = M_HASHTYPE_GET(m);
840 inp->inp_numa_domain = m->m_pkthdr.numa_domain;
845 * Install in the reservation hash table for now, but don't yet
846 * install a connection group since the full 4-tuple isn't yet
849 inp->inp_lport = sc->sc_inc.inc_lport;
850 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
852 * Undo the assignments above if we failed to
853 * put the PCB on the hash lists.
856 if (sc->sc_inc.inc_flags & INC_ISIPV6)
857 inp->in6p_laddr = in6addr_any;
860 inp->inp_laddr.s_addr = INADDR_ANY;
862 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
863 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
868 INP_HASH_WUNLOCK(&V_tcbinfo);
872 if (inp->inp_vflag & INP_IPV6PROTO) {
873 struct inpcb *oinp = sotoinpcb(lso);
876 * Inherit socket options from the listening socket.
877 * Note that in6p_inputopts are not (and should not be)
878 * copied, since it stores previously received options and is
879 * used to detect if each new option is different than the
880 * previous one and hence should be passed to a user.
881 * If we copied in6p_inputopts, a user would not be able to
882 * receive options just after calling the accept system call.
884 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
885 if (oinp->in6p_outputopts)
886 inp->in6p_outputopts =
887 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
890 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
891 struct in6_addr laddr6;
892 struct sockaddr_in6 sin6;
894 sin6.sin6_family = AF_INET6;
895 sin6.sin6_len = sizeof(sin6);
896 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
897 sin6.sin6_port = sc->sc_inc.inc_fport;
898 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
899 laddr6 = inp->in6p_laddr;
900 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
901 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
902 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
903 thread0.td_ucred, m)) != 0) {
904 inp->in6p_laddr = laddr6;
905 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
906 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
911 INP_HASH_WUNLOCK(&V_tcbinfo);
914 /* Override flowlabel from in6_pcbconnect. */
915 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
916 inp->inp_flow |= sc->sc_flowlabel;
919 #if defined(INET) && defined(INET6)
924 struct in_addr laddr;
925 struct sockaddr_in sin;
927 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
929 if (inp->inp_options == NULL) {
930 inp->inp_options = sc->sc_ipopts;
931 sc->sc_ipopts = NULL;
934 sin.sin_family = AF_INET;
935 sin.sin_len = sizeof(sin);
936 sin.sin_addr = sc->sc_inc.inc_faddr;
937 sin.sin_port = sc->sc_inc.inc_fport;
938 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
939 laddr = inp->inp_laddr;
940 if (inp->inp_laddr.s_addr == INADDR_ANY)
941 inp->inp_laddr = sc->sc_inc.inc_laddr;
942 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
943 thread0.td_ucred, m)) != 0) {
944 inp->inp_laddr = laddr;
945 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
946 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
951 INP_HASH_WUNLOCK(&V_tcbinfo);
956 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
957 /* Copy old policy into new socket's. */
958 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
959 printf("syncache_socket: could not copy policy\n");
961 INP_HASH_WUNLOCK(&V_tcbinfo);
963 tcp_state_change(tp, TCPS_SYN_RECEIVED);
964 tp->iss = sc->sc_iss;
965 tp->irs = sc->sc_irs;
968 blk = sototcpcb(lso)->t_fb;
969 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
971 * Our parents t_fb was not the default,
972 * we need to release our ref on tp->t_fb and
973 * pickup one on the new entry.
975 struct tcp_function_block *rblk;
977 rblk = find_and_ref_tcp_fb(blk);
978 KASSERT(rblk != NULL,
979 ("cannot find blk %p out of syncache?", blk));
980 if (tp->t_fb->tfb_tcp_fb_fini)
981 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
982 refcount_release(&tp->t_fb->tfb_refcnt);
985 * XXXrrs this is quite dangerous, it is possible
986 * for the new function to fail to init. We also
987 * are not asking if the handoff_is_ok though at
988 * the very start thats probalbly ok.
990 if (tp->t_fb->tfb_tcp_fb_init) {
991 (*tp->t_fb->tfb_tcp_fb_init)(tp);
994 tp->snd_wl1 = sc->sc_irs;
995 tp->snd_max = tp->iss + 1;
996 tp->snd_nxt = tp->iss + 1;
997 tp->rcv_up = sc->sc_irs + 1;
998 tp->rcv_wnd = sc->sc_wnd;
999 tp->rcv_adv += tp->rcv_wnd;
1000 tp->last_ack_sent = tp->rcv_nxt;
1002 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
1003 if (sc->sc_flags & SCF_NOOPT)
1004 tp->t_flags |= TF_NOOPT;
1006 if (sc->sc_flags & SCF_WINSCALE) {
1007 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
1008 tp->snd_scale = sc->sc_requested_s_scale;
1009 tp->request_r_scale = sc->sc_requested_r_scale;
1011 if (sc->sc_flags & SCF_TIMESTAMP) {
1012 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
1013 tp->ts_recent = sc->sc_tsreflect;
1014 tp->ts_recent_age = tcp_ts_getticks();
1015 tp->ts_offset = sc->sc_tsoff;
1017 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1018 if (sc->sc_flags & SCF_SIGNATURE)
1019 tp->t_flags |= TF_SIGNATURE;
1021 if (sc->sc_flags & SCF_SACK)
1022 tp->t_flags |= TF_SACK_PERMIT;
1025 if (sc->sc_flags & SCF_ECN)
1026 tp->t_flags |= TF_ECN_PERMIT;
1029 * Set up MSS and get cached values from tcp_hostcache.
1030 * This might overwrite some of the defaults we just set.
1032 tcp_mss(tp, sc->sc_peer_mss);
1035 * If the SYN,ACK was retransmitted, indicate that CWND to be
1036 * limited to one segment in cc_conn_init().
1037 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
1039 if (sc->sc_rxmits > 1)
1044 * Allow a TOE driver to install its hooks. Note that we hold the
1045 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
1046 * new connection before the TOE driver has done its thing.
1048 if (ADDED_BY_TOE(sc)) {
1049 struct toedev *tod = sc->sc_tod;
1051 tod->tod_offload_socket(tod, sc->sc_todctx, so);
1055 * Copy and activate timers.
1057 tp->t_keepinit = sototcpcb(lso)->t_keepinit;
1058 tp->t_keepidle = sototcpcb(lso)->t_keepidle;
1059 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
1060 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1061 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1063 TCPSTAT_INC(tcps_accepts);
1075 * This function gets called when we receive an ACK for a
1076 * socket in the LISTEN state. We look up the connection
1077 * in the syncache, and if its there, we pull it out of
1078 * the cache and turn it into a full-blown connection in
1079 * the SYN-RECEIVED state.
1081 * On syncache_socket() success the newly created socket
1082 * has its underlying inp locked.
1085 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1086 struct socket **lsop, struct mbuf *m)
1088 struct syncache *sc;
1089 struct syncache_head *sch;
1090 struct syncache scs;
1095 * Global TCP locks are held because we manipulate the PCB lists
1096 * and create a new socket.
1098 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1099 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1100 ("%s: can handle only ACK", __func__));
1102 if (syncache_cookiesonly()) {
1104 sch = syncache_hashbucket(inc);
1107 sc = syncache_lookup(inc, &sch); /* returns locked sch */
1109 SCH_LOCK_ASSERT(sch);
1114 * Test code for syncookies comparing the syncache stored
1115 * values with the reconstructed values from the cookie.
1118 syncookie_cmp(inc, sch, sc, th, to, *lsop);
1123 * There is no syncache entry, so see if this ACK is
1124 * a returning syncookie. To do this, first:
1125 * A. Check if syncookies are used in case of syncache
1127 * B. See if this socket has had a syncache entry dropped in
1128 * the recent past. We don't want to accept a bogus
1129 * syncookie if we've never received a SYN or accept it
1131 * C. check that the syncookie is valid. If it is, then
1132 * cobble up a fake syncache entry, and return.
1134 if (locked && !V_tcp_syncookies) {
1136 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1137 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1138 "segment rejected (syncookies disabled)\n",
1142 if (locked && !V_tcp_syncookiesonly &&
1143 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1145 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1146 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1147 "segment rejected (no syncache entry)\n",
1151 bzero(&scs, sizeof(scs));
1152 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
1156 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1157 log(LOG_DEBUG, "%s; %s: Segment failed "
1158 "SYNCOOKIE authentication, segment rejected "
1159 "(probably spoofed)\n", s, __func__);
1162 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1163 /* If received ACK has MD5 signature, check it. */
1164 if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1165 (!TCPMD5_ENABLED() ||
1166 TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1168 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1169 log(LOG_DEBUG, "%s; %s: Segment rejected, "
1170 "MD5 signature doesn't match.\n",
1174 TCPSTAT_INC(tcps_sig_err_sigopt);
1175 return (-1); /* Do not send RST */
1177 #endif /* TCP_SIGNATURE */
1179 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1181 * If listening socket requested TCP digests, check that
1182 * received ACK has signature and it is correct.
1183 * If not, drop the ACK and leave sc entry in th cache,
1184 * because SYN was received with correct signature.
1186 if (sc->sc_flags & SCF_SIGNATURE) {
1187 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1189 TCPSTAT_INC(tcps_sig_err_nosigopt);
1191 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1192 log(LOG_DEBUG, "%s; %s: Segment "
1193 "rejected, MD5 signature wasn't "
1194 "provided.\n", s, __func__);
1197 return (-1); /* Do not send RST */
1199 if (!TCPMD5_ENABLED() ||
1200 TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1201 /* Doesn't match or no SA */
1203 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1204 log(LOG_DEBUG, "%s; %s: Segment "
1205 "rejected, MD5 signature doesn't "
1206 "match.\n", s, __func__);
1209 return (-1); /* Do not send RST */
1212 #endif /* TCP_SIGNATURE */
1215 * RFC 7323 PAWS: If we have a timestamp on this segment and
1216 * it's less than ts_recent, drop it.
1217 * XXXMT: RFC 7323 also requires to send an ACK.
1218 * In tcp_input.c this is only done for TCP segments
1219 * with user data, so be consistent here and just drop
1222 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1223 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1225 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1227 "%s; %s: SEG.TSval %u < TS.Recent %u, "
1228 "segment dropped\n", s, __func__,
1229 to->to_tsval, sc->sc_tsreflect);
1232 return (-1); /* Do not send RST */
1236 * Pull out the entry to unlock the bucket row.
1238 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1239 * tcp_state_change(). The tcpcb is not existent at this
1240 * moment. A new one will be allocated via syncache_socket->
1241 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1242 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1244 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1245 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1248 if (ADDED_BY_TOE(sc)) {
1249 struct toedev *tod = sc->sc_tod;
1251 tod->tod_syncache_removed(tod, sc->sc_todctx);
1258 * Segment validation:
1259 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1261 if (th->th_ack != sc->sc_iss + 1) {
1262 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1263 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1264 "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1269 * The SEQ must fall in the window starting at the received
1270 * initial receive sequence number + 1 (the SYN).
1272 if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1273 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1274 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1275 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1276 "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1281 * If timestamps were not negotiated during SYN/ACK they
1282 * must not appear on any segment during this session.
1284 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1285 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1286 log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1287 "segment rejected\n", s, __func__);
1292 * If timestamps were negotiated during SYN/ACK they should
1293 * appear on every segment during this session.
1294 * XXXAO: This is only informal as there have been unverified
1295 * reports of non-compliants stacks.
1297 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1298 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1299 log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1300 "no action\n", s, __func__);
1306 *lsop = syncache_socket(sc, *lsop, m);
1309 TCPSTAT_INC(tcps_sc_aborted);
1311 TCPSTAT_INC(tcps_sc_completed);
1313 /* how do we find the inp for the new socket? */
1318 if (sc != NULL && sc != &scs)
1327 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
1328 uint64_t response_cookie)
1332 unsigned int *pending_counter;
1335 * Global TCP locks are held because we manipulate the PCB lists
1336 * and create a new socket.
1338 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1340 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
1341 *lsop = syncache_socket(sc, *lsop, m);
1342 if (*lsop == NULL) {
1343 TCPSTAT_INC(tcps_sc_aborted);
1344 atomic_subtract_int(pending_counter, 1);
1346 soisconnected(*lsop);
1347 inp = sotoinpcb(*lsop);
1348 tp = intotcpcb(inp);
1349 tp->t_flags |= TF_FASTOPEN;
1350 tp->t_tfo_cookie.server = response_cookie;
1351 tp->snd_max = tp->iss;
1352 tp->snd_nxt = tp->iss;
1353 tp->t_tfo_pending = pending_counter;
1354 TCPSTAT_INC(tcps_sc_completed);
1359 * Given a LISTEN socket and an inbound SYN request, add
1360 * this to the syn cache, and send back a segment:
1361 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1364 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1365 * Doing so would require that we hold onto the data and deliver it
1366 * to the application. However, if we are the target of a SYN-flood
1367 * DoS attack, an attacker could send data which would eventually
1368 * consume all available buffer space if it were ACKed. By not ACKing
1369 * the data, we avoid this DoS scenario.
1371 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1372 * cookie is processed and a new socket is created. In this case, any data
1373 * accompanying the SYN will be queued to the socket by tcp_input() and will
1374 * be ACKed either when the application sends response data or the delayed
1375 * ACK timer expires, whichever comes first.
1378 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1379 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1384 struct syncache *sc = NULL;
1385 struct syncache_head *sch;
1386 struct mbuf *ipopts = NULL;
1388 int win, ip_ttl, ip_tos;
1392 int autoflowlabel = 0;
1395 struct label *maclabel;
1397 struct syncache scs;
1399 uint64_t tfo_response_cookie;
1400 unsigned int *tfo_pending = NULL;
1401 int tfo_cookie_valid = 0;
1402 int tfo_response_cookie_valid = 0;
1405 INP_WLOCK_ASSERT(inp); /* listen socket */
1406 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1407 ("%s: unexpected tcp flags", __func__));
1410 * Combine all so/tp operations very early to drop the INP lock as
1414 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1416 cred = crhold(so->so_cred);
1419 if ((inc->inc_flags & INC_ISIPV6) &&
1420 (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1423 ip_ttl = inp->inp_ip_ttl;
1424 ip_tos = inp->inp_ip_tos;
1425 win = so->sol_sbrcv_hiwat;
1426 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1428 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
1429 (tp->t_tfo_pending != NULL) &&
1430 (to->to_flags & TOF_FASTOPEN)) {
1432 * Limit the number of pending TFO connections to
1433 * approximately half of the queue limit. This prevents TFO
1434 * SYN floods from starving the service by filling the
1435 * listen queue with bogus TFO connections.
1437 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1438 (so->sol_qlimit / 2)) {
1441 result = tcp_fastopen_check_cookie(inc,
1442 to->to_tfo_cookie, to->to_tfo_len,
1443 &tfo_response_cookie);
1444 tfo_cookie_valid = (result > 0);
1445 tfo_response_cookie_valid = (result >= 0);
1449 * Remember the TFO pending counter as it will have to be
1450 * decremented below if we don't make it to syncache_tfo_expand().
1452 tfo_pending = tp->t_tfo_pending;
1455 /* By the time we drop the lock these should no longer be used. */
1460 if (mac_syncache_init(&maclabel) != 0) {
1464 mac_syncache_create(maclabel, inp);
1466 if (!tfo_cookie_valid)
1470 * Remember the IP options, if any.
1473 if (!(inc->inc_flags & INC_ISIPV6))
1476 ipopts = (m) ? ip_srcroute(m) : NULL;
1481 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1483 * If listening socket requested TCP digests, check that received
1484 * SYN has signature and it is correct. If signature doesn't match
1485 * or TCP_SIGNATURE support isn't enabled, drop the packet.
1487 if (ltflags & TF_SIGNATURE) {
1488 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1489 TCPSTAT_INC(tcps_sig_err_nosigopt);
1492 if (!TCPMD5_ENABLED() ||
1493 TCPMD5_INPUT(m, th, to->to_signature) != 0)
1496 #endif /* TCP_SIGNATURE */
1498 * See if we already have an entry for this connection.
1499 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1501 * XXX: should the syncache be re-initialized with the contents
1502 * of the new SYN here (which may have different options?)
1504 * XXX: We do not check the sequence number to see if this is a
1505 * real retransmit or a new connection attempt. The question is
1506 * how to handle such a case; either ignore it as spoofed, or
1507 * drop the current entry and create a new one?
1509 if (syncache_cookiesonly()) {
1511 sch = syncache_hashbucket(inc);
1514 sc = syncache_lookup(inc, &sch); /* returns locked sch */
1516 SCH_LOCK_ASSERT(sch);
1519 if (tfo_cookie_valid)
1521 TCPSTAT_INC(tcps_sc_dupsyn);
1524 * If we were remembering a previous source route,
1525 * forget it and use the new one we've been given.
1528 (void) m_free(sc->sc_ipopts);
1529 sc->sc_ipopts = ipopts;
1532 * Update timestamp if present.
1534 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1535 sc->sc_tsreflect = to->to_tsval;
1537 sc->sc_flags &= ~SCF_TIMESTAMP;
1540 * Since we have already unconditionally allocated label
1541 * storage, free it up. The syncache entry will already
1542 * have an initialized label we can use.
1544 mac_syncache_destroy(&maclabel);
1546 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1547 /* Retransmit SYN|ACK and reset retransmit count. */
1548 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1549 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1550 "resetting timer and retransmitting SYN|ACK\n",
1554 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1556 syncache_timeout(sc, sch, 1);
1557 TCPSTAT_INC(tcps_sndacks);
1558 TCPSTAT_INC(tcps_sndtotal);
1564 if (tfo_cookie_valid) {
1565 bzero(&scs, sizeof(scs));
1571 * Skip allocating a syncache entry if we are just going to discard
1575 bzero(&scs, sizeof(scs));
1578 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1581 * The zone allocator couldn't provide more entries.
1582 * Treat this as if the cache was full; drop the oldest
1583 * entry and insert the new one.
1585 TCPSTAT_INC(tcps_sc_zonefail);
1586 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1587 sch->sch_last_overflow = time_uptime;
1588 syncache_drop(sc, sch);
1589 syncache_pause(inc);
1591 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1593 if (V_tcp_syncookies) {
1594 bzero(&scs, sizeof(scs));
1598 ("%s: bucket unexpectedly unlocked",
1602 (void) m_free(ipopts);
1609 if (!tfo_cookie_valid && tfo_response_cookie_valid)
1610 sc->sc_tfo_cookie = &tfo_response_cookie;
1613 * Fill in the syncache values.
1616 sc->sc_label = maclabel;
1620 sc->sc_ipopts = ipopts;
1621 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1623 if (!(inc->inc_flags & INC_ISIPV6))
1626 sc->sc_ip_tos = ip_tos;
1627 sc->sc_ip_ttl = ip_ttl;
1631 sc->sc_todctx = todctx;
1633 sc->sc_irs = th->th_seq;
1635 sc->sc_flowlabel = 0;
1638 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1639 * win was derived from socket earlier in the function.
1642 win = imin(win, TCP_MAXWIN);
1645 if (V_tcp_do_rfc1323) {
1647 * A timestamp received in a SYN makes
1648 * it ok to send timestamp requests and replies.
1650 if (to->to_flags & TOF_TS) {
1651 sc->sc_tsreflect = to->to_tsval;
1652 sc->sc_flags |= SCF_TIMESTAMP;
1653 sc->sc_tsoff = tcp_new_ts_offset(inc);
1655 if (to->to_flags & TOF_SCALE) {
1659 * Pick the smallest possible scaling factor that
1660 * will still allow us to scale up to sb_max, aka
1661 * kern.ipc.maxsockbuf.
1663 * We do this because there are broken firewalls that
1664 * will corrupt the window scale option, leading to
1665 * the other endpoint believing that our advertised
1666 * window is unscaled. At scale factors larger than
1667 * 5 the unscaled window will drop below 1500 bytes,
1668 * leading to serious problems when traversing these
1671 * With the default maxsockbuf of 256K, a scale factor
1672 * of 3 will be chosen by this algorithm. Those who
1673 * choose a larger maxsockbuf should watch out
1674 * for the compatibility problems mentioned above.
1676 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1677 * or <SYN,ACK>) segment itself is never scaled.
1679 while (wscale < TCP_MAX_WINSHIFT &&
1680 (TCP_MAXWIN << wscale) < sb_max)
1682 sc->sc_requested_r_scale = wscale;
1683 sc->sc_requested_s_scale = to->to_wscale;
1684 sc->sc_flags |= SCF_WINSCALE;
1687 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1689 * If listening socket requested TCP digests, flag this in the
1690 * syncache so that syncache_respond() will do the right thing
1693 if (ltflags & TF_SIGNATURE)
1694 sc->sc_flags |= SCF_SIGNATURE;
1695 #endif /* TCP_SIGNATURE */
1696 if (to->to_flags & TOF_SACKPERM)
1697 sc->sc_flags |= SCF_SACK;
1698 if (to->to_flags & TOF_MSS)
1699 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */
1700 if (ltflags & TF_NOOPT)
1701 sc->sc_flags |= SCF_NOOPT;
1702 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1703 sc->sc_flags |= SCF_ECN;
1705 if (V_tcp_syncookies)
1706 sc->sc_iss = syncookie_generate(sch, sc);
1708 sc->sc_iss = arc4random();
1710 if (autoflowlabel) {
1711 if (V_tcp_syncookies)
1712 sc->sc_flowlabel = sc->sc_iss;
1714 sc->sc_flowlabel = ip6_randomflowlabel();
1715 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1721 if (tfo_cookie_valid) {
1722 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie);
1723 /* INP_WUNLOCK(inp) will be performed by the caller */
1728 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1730 * Do a standard 3-way handshake.
1732 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1733 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1735 else if (sc != &scs)
1736 syncache_insert(sc, sch); /* locks and unlocks sch */
1737 TCPSTAT_INC(tcps_sndacks);
1738 TCPSTAT_INC(tcps_sndtotal);
1742 TCPSTAT_INC(tcps_sc_dropped);
1747 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1754 * If tfo_pending is not NULL here, then a TFO SYN that did not
1755 * result in a new socket was processed and the associated pending
1756 * counter has not yet been decremented. All such TFO processing paths
1757 * transit this point.
1759 if (tfo_pending != NULL)
1760 tcp_fastopen_decrement_counter(tfo_pending);
1767 mac_syncache_destroy(&maclabel);
1773 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment,
1774 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1777 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1779 struct ip *ip = NULL;
1781 struct tcphdr *th = NULL;
1782 int optlen, error = 0; /* Make compiler happy */
1783 u_int16_t hlen, tlen, mssopt;
1786 struct ip6_hdr *ip6 = NULL;
1790 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1793 tlen = hlen + sizeof(struct tcphdr);
1795 /* Determine MSS we advertize to other end of connection. */
1796 mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss);
1798 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1799 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1800 ("syncache: mbuf too small"));
1802 /* Create the IP+TCP header from scratch. */
1803 m = m_gethdr(M_NOWAIT, MT_DATA);
1807 mac_syncache_create_mbuf(sc->sc_label, m);
1809 m->m_data += max_linkhdr;
1811 m->m_pkthdr.len = tlen;
1812 m->m_pkthdr.rcvif = NULL;
1815 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1816 ip6 = mtod(m, struct ip6_hdr *);
1817 ip6->ip6_vfc = IPV6_VERSION;
1818 ip6->ip6_nxt = IPPROTO_TCP;
1819 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1820 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1821 ip6->ip6_plen = htons(tlen - hlen);
1822 /* ip6_hlim is set after checksum */
1823 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1824 ip6->ip6_flow |= sc->sc_flowlabel;
1826 th = (struct tcphdr *)(ip6 + 1);
1829 #if defined(INET6) && defined(INET)
1834 ip = mtod(m, struct ip *);
1835 ip->ip_v = IPVERSION;
1836 ip->ip_hl = sizeof(struct ip) >> 2;
1837 ip->ip_len = htons(tlen);
1841 ip->ip_p = IPPROTO_TCP;
1842 ip->ip_src = sc->sc_inc.inc_laddr;
1843 ip->ip_dst = sc->sc_inc.inc_faddr;
1844 ip->ip_ttl = sc->sc_ip_ttl;
1845 ip->ip_tos = sc->sc_ip_tos;
1848 * See if we should do MTU discovery. Route lookups are
1849 * expensive, so we will only unset the DF bit if:
1851 * 1) path_mtu_discovery is disabled
1852 * 2) the SCF_UNREACH flag has been set
1854 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1855 ip->ip_off |= htons(IP_DF);
1857 th = (struct tcphdr *)(ip + 1);
1860 th->th_sport = sc->sc_inc.inc_lport;
1861 th->th_dport = sc->sc_inc.inc_fport;
1864 th->th_seq = htonl(sc->sc_iss);
1866 th->th_seq = htonl(sc->sc_iss + 1);
1867 th->th_ack = htonl(sc->sc_irs + 1);
1868 th->th_off = sizeof(struct tcphdr) >> 2;
1870 th->th_flags = flags;
1871 th->th_win = htons(sc->sc_wnd);
1874 if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) {
1875 th->th_flags |= TH_ECE;
1876 TCPSTAT_INC(tcps_ecn_shs);
1879 /* Tack on the TCP options. */
1880 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1883 if (flags & TH_SYN) {
1885 to.to_flags = TOF_MSS;
1886 if (sc->sc_flags & SCF_WINSCALE) {
1887 to.to_wscale = sc->sc_requested_r_scale;
1888 to.to_flags |= TOF_SCALE;
1890 if (sc->sc_flags & SCF_SACK)
1891 to.to_flags |= TOF_SACKPERM;
1892 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1893 if (sc->sc_flags & SCF_SIGNATURE)
1894 to.to_flags |= TOF_SIGNATURE;
1896 if (sc->sc_tfo_cookie) {
1897 to.to_flags |= TOF_FASTOPEN;
1898 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1899 to.to_tfo_cookie = sc->sc_tfo_cookie;
1900 /* don't send cookie again when retransmitting response */
1901 sc->sc_tfo_cookie = NULL;
1904 if (sc->sc_flags & SCF_TIMESTAMP) {
1905 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1906 to.to_tsecr = sc->sc_tsreflect;
1907 to.to_flags |= TOF_TS;
1909 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1911 /* Adjust headers by option size. */
1912 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1914 m->m_pkthdr.len += optlen;
1916 if (sc->sc_inc.inc_flags & INC_ISIPV6)
1917 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1920 ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1921 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1922 if (sc->sc_flags & SCF_SIGNATURE) {
1923 KASSERT(to.to_flags & TOF_SIGNATURE,
1924 ("tcp_addoptions() didn't set tcp_signature"));
1926 /* NOTE: to.to_signature is inside of mbuf */
1927 if (!TCPMD5_ENABLED() ||
1928 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1937 M_SETFIB(m, sc->sc_inc.inc_fibnum);
1938 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1940 * If we have peer's SYN and it has a flowid, then let's assign it to
1941 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid
1942 * to SYN|ACK due to lack of inp here.
1944 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1945 m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1946 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1949 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1950 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1951 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1953 ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1955 if (ADDED_BY_TOE(sc)) {
1956 struct toedev *tod = sc->sc_tod;
1958 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1963 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
1964 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1967 #if defined(INET6) && defined(INET)
1972 m->m_pkthdr.csum_flags = CSUM_TCP;
1973 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1974 htons(tlen + optlen - hlen + IPPROTO_TCP));
1976 if (ADDED_BY_TOE(sc)) {
1977 struct toedev *tod = sc->sc_tod;
1979 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1984 TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
1985 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1992 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1993 * that exceed the capacity of the syncache by avoiding the storage of any
1994 * of the SYNs we receive. Syncookies defend against blind SYN flooding
1995 * attacks where the attacker does not have access to our responses.
1997 * Syncookies encode and include all necessary information about the
1998 * connection setup within the SYN|ACK that we send back. That way we
1999 * can avoid keeping any local state until the ACK to our SYN|ACK returns
2000 * (if ever). Normally the syncache and syncookies are running in parallel
2001 * with the latter taking over when the former is exhausted. When matching
2002 * syncache entry is found the syncookie is ignored.
2004 * The only reliable information persisting the 3WHS is our initial sequence
2005 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient
2006 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
2007 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK
2008 * returns and signifies a legitimate connection if it matches the ACK.
2010 * The available space of 32 bits to store the hash and to encode the SYN
2011 * option information is very tight and we should have at least 24 bits for
2012 * the MAC to keep the number of guesses by blind spoofing reasonably high.
2014 * SYN option information we have to encode to fully restore a connection:
2015 * MSS: is imporant to chose an optimal segment size to avoid IP level
2016 * fragmentation along the path. The common MSS values can be encoded
2017 * in a 3-bit table. Uncommon values are captured by the next lower value
2018 * in the table leading to a slight increase in packetization overhead.
2019 * WSCALE: is necessary to allow large windows to be used for high delay-
2020 * bandwidth product links. Not scaling the window when it was initially
2021 * negotiated is bad for performance as lack of scaling further decreases
2022 * the apparent available send window. We only need to encode the WSCALE
2023 * we received from the remote end. Our end can be recalculated at any
2024 * time. The common WSCALE values can be encoded in a 3-bit table.
2025 * Uncommon values are captured by the next lower value in the table
2026 * making us under-estimate the available window size halving our
2027 * theoretically possible maximum throughput for that connection.
2028 * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2029 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2030 * that are included in all segments on a connection. We enable them when
2033 * Security of syncookies and attack vectors:
2035 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2036 * together with the gloabl secret to make it unique per connection attempt.
2037 * Thus any change of any of those parameters results in a different MAC output
2038 * in an unpredictable way unless a collision is encountered. 24 bits of the
2039 * MAC are embedded into the ISS.
2041 * To prevent replay attacks two rotating global secrets are updated with a
2042 * new random value every 15 seconds. The life-time of a syncookie is thus
2045 * Vector 1: Attacking the secret. This requires finding a weakness in the
2046 * MAC itself or the way it is used here. The attacker can do a chosen plain
2047 * text attack by varying and testing the all parameters under his control.
2048 * The strength depends on the size and randomness of the secret, and the
2049 * cryptographic security of the MAC function. Due to the constant updating
2050 * of the secret the attacker has at most 29.999 seconds to find the secret
2051 * and launch spoofed connections. After that he has to start all over again.
2053 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC
2054 * size an average of 4,823 attempts are required for a 50% chance of success
2055 * to spoof a single syncookie (birthday collision paradox). However the
2056 * attacker is blind and doesn't know if one of his attempts succeeded unless
2057 * he has a side channel to interfere success from. A single connection setup
2058 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2059 * This many attempts are required for each one blind spoofed connection. For
2060 * every additional spoofed connection he has to launch another N attempts.
2061 * Thus for a sustained rate 100 spoofed connections per second approximately
2062 * 1,800,000 packets per second would have to be sent.
2064 * NB: The MAC function should be fast so that it doesn't become a CPU
2065 * exhaustion attack vector itself.
2068 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2069 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2070 * http://cr.yp.to/syncookies.html (overview)
2071 * http://cr.yp.to/syncookies/archive (details)
2074 * Schematic construction of a syncookie enabled Initial Sequence Number:
2076 * 12345678901234567890123456789012
2077 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2079 * x 24 MAC (truncated)
2080 * W 3 Send Window Scale index
2082 * S 1 SACK permitted
2083 * P 1 Odd/even secret
2087 * Distribution and probability of certain MSS values. Those in between are
2088 * rounded down to the next lower one.
2089 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2090 * .2% .3% 5% 7% 7% 20% 15% 45%
2092 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2095 * Distribution and probability of certain WSCALE values. We have to map the
2096 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2097 * bits based on prevalence of certain values. Where we don't have an exact
2098 * match for are rounded down to the next lower one letting us under-estimate
2099 * the true available window. At the moment this would happen only for the
2100 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2101 * and window size). The absence of the WSCALE option (no scaling in either
2102 * direction) is encoded with index zero.
2103 * [WSCALE values histograms, Allman, 2012]
2104 * X 10 10 35 5 6 14 10% by host
2105 * X 11 4 5 5 18 49 3% by connections
2107 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2110 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed
2111 * and good cryptographic properties.
2114 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2115 uint8_t *secbits, uintptr_t secmod)
2118 uint32_t siphash[2];
2120 SipHash24_Init(&ctx);
2121 SipHash_SetKey(&ctx, secbits);
2122 switch (inc->inc_flags & INC_ISIPV6) {
2125 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2126 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2131 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2132 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2136 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2137 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2138 SipHash_Update(&ctx, &irs, sizeof(irs));
2139 SipHash_Update(&ctx, &flags, sizeof(flags));
2140 SipHash_Update(&ctx, &secmod, sizeof(secmod));
2141 SipHash_Final((u_int8_t *)&siphash, &ctx);
2143 return (siphash[0] ^ siphash[1]);
2147 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2149 u_int i, secbit, wscale;
2152 union syncookie cookie;
2156 /* Map our computed MSS into the 3-bit index. */
2157 for (i = nitems(tcp_sc_msstab) - 1;
2158 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2161 cookie.flags.mss_idx = i;
2164 * Map the send window scale into the 3-bit index but only if
2165 * the wscale option was received.
2167 if (sc->sc_flags & SCF_WINSCALE) {
2168 wscale = sc->sc_requested_s_scale;
2169 for (i = nitems(tcp_sc_wstab) - 1;
2170 tcp_sc_wstab[i] > wscale && i > 0;
2173 cookie.flags.wscale_idx = i;
2176 /* Can we do SACK? */
2177 if (sc->sc_flags & SCF_SACK)
2178 cookie.flags.sack_ok = 1;
2180 /* Which of the two secrets to use. */
2181 secbit = V_tcp_syncache.secret.oddeven & 0x1;
2182 cookie.flags.odd_even = secbit;
2184 secbits = V_tcp_syncache.secret.key[secbit];
2185 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2189 * Put the flags into the hash and XOR them to get better ISS number
2190 * variance. This doesn't enhance the cryptographic strength and is
2191 * done to prevent the 8 cookie bits from showing up directly on the
2195 iss |= cookie.cookie ^ (hash >> 24);
2197 TCPSTAT_INC(tcps_sc_sendcookie);
2201 static struct syncache *
2202 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2203 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2209 int wnd, wscale = 0;
2210 union syncookie cookie;
2213 * Pull information out of SYN-ACK/ACK and revert sequence number
2216 ack = th->th_ack - 1;
2217 seq = th->th_seq - 1;
2220 * Unpack the flags containing enough information to restore the
2223 cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2225 /* Which of the two secrets to use. */
2226 secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2228 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2230 /* The recomputed hash matches the ACK if this was a genuine cookie. */
2231 if ((ack & ~0xff) != (hash & ~0xff))
2234 /* Fill in the syncache values. */
2236 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2237 sc->sc_ipopts = NULL;
2242 switch (inc->inc_flags & INC_ISIPV6) {
2245 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2246 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2251 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2252 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
2257 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2259 /* We can simply recompute receive window scale we sent earlier. */
2260 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2263 /* Only use wscale if it was enabled in the orignal SYN. */
2264 if (cookie.flags.wscale_idx > 0) {
2265 sc->sc_requested_r_scale = wscale;
2266 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2267 sc->sc_flags |= SCF_WINSCALE;
2270 wnd = lso->sol_sbrcv_hiwat;
2272 wnd = imin(wnd, TCP_MAXWIN);
2275 if (cookie.flags.sack_ok)
2276 sc->sc_flags |= SCF_SACK;
2278 if (to->to_flags & TOF_TS) {
2279 sc->sc_flags |= SCF_TIMESTAMP;
2280 sc->sc_tsreflect = to->to_tsval;
2281 sc->sc_tsoff = tcp_new_ts_offset(inc);
2284 if (to->to_flags & TOF_SIGNATURE)
2285 sc->sc_flags |= SCF_SIGNATURE;
2289 TCPSTAT_INC(tcps_sc_recvcookie);
2295 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2296 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2299 struct syncache scs, *scx;
2302 bzero(&scs, sizeof(scs));
2303 scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
2305 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2309 if (sc->sc_peer_mss != scx->sc_peer_mss)
2310 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2311 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2313 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2314 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2315 s, __func__, sc->sc_requested_r_scale,
2316 scx->sc_requested_r_scale);
2318 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2319 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2320 s, __func__, sc->sc_requested_s_scale,
2321 scx->sc_requested_s_scale);
2323 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2324 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2331 #endif /* INVARIANTS */
2334 syncookie_reseed(void *arg)
2336 struct tcp_syncache *sc = arg;
2341 * Reseeding the secret doesn't have to be protected by a lock.
2342 * It only must be ensured that the new random values are visible
2343 * to all CPUs in a SMP environment. The atomic with release
2344 * semantics ensures that.
2346 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2347 secbits = sc->secret.key[secbit];
2348 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2349 atomic_add_rel_int(&sc->secret.oddeven, 1);
2351 /* Reschedule ourself. */
2352 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2356 * We have overflowed a bucket. Let's pause dealing with the syncache.
2357 * This function will increment the bucketoverflow statistics appropriately
2358 * (once per pause when pausing is enabled; otherwise, once per overflow).
2361 syncache_pause(struct in_conninfo *inc)
2367 * 2. Add sysctl read here so we don't get the benefit of this
2368 * change without the new sysctl.
2372 * Try an unlocked read. If we already know that another thread
2373 * has activated the feature, there is no need to proceed.
2375 if (V_tcp_syncache.paused)
2378 /* Are cookied enabled? If not, we can't pause. */
2379 if (!V_tcp_syncookies) {
2380 TCPSTAT_INC(tcps_sc_bucketoverflow);
2385 * We may be the first thread to find an overflow. Get the lock
2386 * and evaluate if we need to take action.
2388 mtx_lock(&V_tcp_syncache.pause_mtx);
2389 if (V_tcp_syncache.paused) {
2390 mtx_unlock(&V_tcp_syncache.pause_mtx);
2394 /* Activate protection. */
2395 V_tcp_syncache.paused = true;
2396 TCPSTAT_INC(tcps_sc_bucketoverflow);
2399 * Determine the last backoff time. If we are seeing a re-newed
2400 * attack within that same time after last reactivating the syncache,
2401 * consider it an extension of the same attack.
2403 delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2404 if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2405 if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2407 V_tcp_syncache.pause_backoff++;
2410 delta = TCP_SYNCACHE_PAUSE_TIME;
2411 V_tcp_syncache.pause_backoff = 0;
2414 /* Log a warning, including IP addresses, if able. */
2416 s = tcp_log_addrs(inc, NULL, NULL, NULL);
2418 s = (const char *)NULL;
2419 log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2420 "the next %lld seconds%s%s%s\n", (long long)delta,
2421 (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2422 (s != NULL) ? ")" : "");
2423 free(__DECONST(void *, s), M_TCPLOG);
2425 /* Use the calculated delta to set a new pause time. */
2426 V_tcp_syncache.pause_until = time_uptime + delta;
2427 callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2429 mtx_unlock(&V_tcp_syncache.pause_mtx);
2432 /* Evaluate whether we need to unpause. */
2434 syncache_unpause(void *arg)
2436 struct tcp_syncache *sc;
2440 mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2441 callout_deactivate(&sc->pause_co);
2444 * Check to make sure we are not running early. If the pause
2445 * time has expired, then deactivate the protection.
2447 if ((delta = sc->pause_until - time_uptime) > 0)
2448 callout_schedule(&sc->pause_co, delta * hz);
2454 * Exports the syncache entries to userland so that netstat can display
2455 * them alongside the other sockets. This function is intended to be
2456 * called only from tcp_pcblist.
2458 * Due to concurrency on an active system, the number of pcbs exported
2459 * may have no relation to max_pcbs. max_pcbs merely indicates the
2460 * amount of space the caller allocated for this function to use.
2463 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
2466 struct syncache *sc;
2467 struct syncache_head *sch;
2468 int count, error, i;
2470 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
2471 sch = &V_tcp_syncache.hashbase[i];
2473 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2474 if (count >= max_pcbs) {
2478 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2480 bzero(&xt, sizeof(xt));
2481 xt.xt_len = sizeof(xt);
2482 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2483 xt.xt_inp.inp_vflag = INP_IPV6;
2485 xt.xt_inp.inp_vflag = INP_IPV4;
2486 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2487 sizeof (struct in_conninfo));
2488 xt.t_state = TCPS_SYN_RECEIVED;
2489 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2490 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2491 xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2492 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2493 error = SYSCTL_OUT(req, &xt, sizeof xt);
2503 *pcbs_exported = count;