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 *, struct syncache_head *,
134 const struct mbuf *, int);
135 static struct socket *syncache_socket(struct syncache *, struct socket *,
137 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch,
139 static void syncache_timer(void *);
141 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
142 uint8_t *, uintptr_t);
143 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *);
144 static struct syncache
145 *syncookie_lookup(struct in_conninfo *, struct syncache_head *,
146 struct syncache *, struct tcphdr *, struct tcpopt *,
148 static void syncookie_reseed(void *);
150 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
151 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
156 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
157 * 3 retransmits corresponds to a timeout with default values of
158 * tcp_rexmit_initial * ( 1 +
161 * tcp_backoff[3]) + 3 * tcp_rexmit_slop,
162 * 1000 ms * (1 + 2 + 4 + 8) + 3 * 200 ms = 15600 ms,
163 * the odds are that the user has given up attempting to connect by then.
165 #define SYNCACHE_MAXREXMTS 3
167 /* Arbitrary values */
168 #define TCP_SYNCACHE_HASHSIZE 512
169 #define TCP_SYNCACHE_BUCKETLIMIT 30
171 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
172 #define V_tcp_syncache VNET(tcp_syncache)
174 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0,
177 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
178 &VNET_NAME(tcp_syncache.bucket_limit), 0,
179 "Per-bucket hash limit for syncache");
181 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
182 &VNET_NAME(tcp_syncache.cache_limit), 0,
183 "Overall entry limit for syncache");
185 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
186 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
188 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
189 &VNET_NAME(tcp_syncache.hashsize), 0,
190 "Size of TCP syncache hashtable");
193 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
198 new = V_tcp_syncache.rexmt_limit;
199 error = sysctl_handle_int(oidp, &new, 0, req);
200 if ((error == 0) && (req->newptr != NULL)) {
201 if (new > TCP_MAXRXTSHIFT)
204 V_tcp_syncache.rexmt_limit = new;
209 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
210 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW,
211 &VNET_NAME(tcp_syncache.rexmt_limit), 0,
212 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
213 "Limit on SYN/ACK retransmissions");
215 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
216 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
217 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
218 "Send reset on socket allocation failure");
220 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
222 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx)
223 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx)
224 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED)
227 * Requires the syncache entry to be already removed from the bucket list.
230 syncache_free(struct syncache *sc)
234 (void) m_free(sc->sc_ipopts);
238 mac_syncache_destroy(&sc->sc_label);
241 uma_zfree(V_tcp_syncache.zone, sc);
249 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
250 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
251 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
252 V_tcp_syncache.hash_secret = arc4random();
254 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
255 &V_tcp_syncache.hashsize);
256 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
257 &V_tcp_syncache.bucket_limit);
258 if (!powerof2(V_tcp_syncache.hashsize) ||
259 V_tcp_syncache.hashsize == 0) {
260 printf("WARNING: syncache hash size is not a power of 2.\n");
261 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
263 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
266 V_tcp_syncache.cache_limit =
267 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
268 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
269 &V_tcp_syncache.cache_limit);
271 /* Allocate the hash table. */
272 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
273 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
276 V_tcp_syncache.vnet = curvnet;
279 /* Initialize the hash buckets. */
280 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
281 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
282 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
284 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
285 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
286 V_tcp_syncache.hashbase[i].sch_length = 0;
287 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
288 V_tcp_syncache.hashbase[i].sch_last_overflow =
289 -(SYNCOOKIE_LIFETIME + 1);
292 /* Create the syncache entry zone. */
293 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
294 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
295 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
296 V_tcp_syncache.cache_limit);
298 /* Start the SYN cookie reseeder callout. */
299 callout_init(&V_tcp_syncache.secret.reseed, 1);
300 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
301 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
302 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
303 syncookie_reseed, &V_tcp_syncache);
308 syncache_destroy(void)
310 struct syncache_head *sch;
311 struct syncache *sc, *nsc;
315 * Stop the re-seed timer before freeing resources. No need to
316 * possibly schedule it another time.
318 callout_drain(&V_tcp_syncache.secret.reseed);
320 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */
321 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
323 sch = &V_tcp_syncache.hashbase[i];
324 callout_drain(&sch->sch_timer);
327 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
328 syncache_drop(sc, sch);
330 KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
331 ("%s: sch->sch_bucket not empty", __func__));
332 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
333 __func__, sch->sch_length));
334 mtx_destroy(&sch->sch_mtx);
337 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
338 ("%s: cache_count not 0", __func__));
340 /* Free the allocated global resources. */
341 uma_zdestroy(V_tcp_syncache.zone);
342 free(V_tcp_syncache.hashbase, M_SYNCACHE);
347 * Inserts a syncache entry into the specified bucket row.
348 * Locks and unlocks the syncache_head autonomously.
351 syncache_insert(struct syncache *sc, struct syncache_head *sch)
353 struct syncache *sc2;
358 * Make sure that we don't overflow the per-bucket limit.
359 * If the bucket is full, toss the oldest element.
361 if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
362 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
363 ("sch->sch_length incorrect"));
364 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
365 sch->sch_last_overflow = time_uptime;
366 syncache_drop(sc2, sch);
367 TCPSTAT_INC(tcps_sc_bucketoverflow);
370 /* Put it into the bucket. */
371 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
375 if (ADDED_BY_TOE(sc)) {
376 struct toedev *tod = sc->sc_tod;
378 tod->tod_syncache_added(tod, sc->sc_todctx);
382 /* Reinitialize the bucket row's timer. */
383 if (sch->sch_length == 1)
384 sch->sch_nextc = ticks + INT_MAX;
385 syncache_timeout(sc, sch, 1);
389 TCPSTATES_INC(TCPS_SYN_RECEIVED);
390 TCPSTAT_INC(tcps_sc_added);
394 * Remove and free entry from syncache bucket row.
395 * Expects locked syncache head.
398 syncache_drop(struct syncache *sc, struct syncache_head *sch)
401 SCH_LOCK_ASSERT(sch);
403 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
404 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
408 if (ADDED_BY_TOE(sc)) {
409 struct toedev *tod = sc->sc_tod;
411 tod->tod_syncache_removed(tod, sc->sc_todctx);
419 * Engage/reengage time on bucket row.
422 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
426 if (sc->sc_rxmits == 0)
427 rexmt = tcp_rexmit_initial;
430 tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
431 tcp_rexmit_min, TCPTV_REXMTMAX);
432 sc->sc_rxttime = ticks + rexmt;
434 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
435 sch->sch_nextc = sc->sc_rxttime;
437 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
438 syncache_timer, (void *)sch);
443 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
444 * If we have retransmitted an entry the maximum number of times, expire it.
445 * One separate timer for each bucket row.
448 syncache_timer(void *xsch)
450 struct syncache_head *sch = (struct syncache_head *)xsch;
451 struct syncache *sc, *nsc;
455 CURVNET_SET(sch->sch_sc->vnet);
457 /* NB: syncache_head has already been locked by the callout. */
458 SCH_LOCK_ASSERT(sch);
461 * In the following cycle we may remove some entries and/or
462 * advance some timeouts, so re-initialize the bucket timer.
464 sch->sch_nextc = tick + INT_MAX;
466 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
468 * We do not check if the listen socket still exists
469 * and accept the case where the listen socket may be
470 * gone by the time we resend the SYN/ACK. We do
471 * not expect this to happens often. If it does,
472 * then the RST will be sent by the time the remote
473 * host does the SYN/ACK->ACK.
475 if (TSTMP_GT(sc->sc_rxttime, tick)) {
476 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
477 sch->sch_nextc = sc->sc_rxttime;
480 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
481 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
482 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
483 "giving up and removing syncache entry\n",
487 syncache_drop(sc, sch);
488 TCPSTAT_INC(tcps_sc_stale);
491 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
492 log(LOG_DEBUG, "%s; %s: Response timeout, "
493 "retransmitting (%u) SYN|ACK\n",
494 s, __func__, sc->sc_rxmits);
498 syncache_respond(sc, sch, NULL, TH_SYN|TH_ACK);
499 TCPSTAT_INC(tcps_sc_retransmitted);
500 syncache_timeout(sc, sch, 0);
502 if (!TAILQ_EMPTY(&(sch)->sch_bucket))
503 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
504 syncache_timer, (void *)(sch));
509 * Find an entry in the syncache.
510 * Returns always with locked syncache_head plus a matching entry or NULL.
512 static struct syncache *
513 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
516 struct syncache_head *sch;
520 * The hash is built on foreign port + local port + foreign address.
521 * We rely on the fact that struct in_conninfo starts with 16 bits
522 * of foreign port, then 16 bits of local port then followed by 128
523 * bits of foreign address. In case of IPv4 address, the first 3
524 * 32-bit words of the address always are zeroes.
526 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
527 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
529 sch = &V_tcp_syncache.hashbase[hash];
533 /* Circle through bucket row to find matching entry. */
534 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
535 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
536 sizeof(struct in_endpoints)) == 0)
539 return (sc); /* Always returns with locked sch. */
543 * This function is called when we get a RST for a
544 * non-existent connection, so that we can see if the
545 * connection is in the syn cache. If it is, zap it.
546 * If required send a challenge ACK.
549 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m)
552 struct syncache_head *sch;
555 sc = syncache_lookup(inc, &sch); /* returns locked sch */
556 SCH_LOCK_ASSERT(sch);
559 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
560 * See RFC 793 page 65, section SEGMENT ARRIVES.
562 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
563 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
564 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
565 "FIN flag set, segment ignored\n", s, __func__);
566 TCPSTAT_INC(tcps_badrst);
571 * No corresponding connection was found in syncache.
572 * If syncookies are enabled and possibly exclusively
573 * used, or we are under memory pressure, a valid RST
574 * may not find a syncache entry. In that case we're
575 * done and no SYN|ACK retransmissions will happen.
576 * Otherwise the RST was misdirected or spoofed.
579 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
580 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
581 "syncache entry (possibly syncookie only), "
582 "segment ignored\n", s, __func__);
583 TCPSTAT_INC(tcps_badrst);
588 * If the RST bit is set, check the sequence number to see
589 * if this is a valid reset segment.
592 * In all states except SYN-SENT, all reset (RST) segments
593 * are validated by checking their SEQ-fields. A reset is
594 * valid if its sequence number is in the window.
597 * There are four cases for the acceptability test for an incoming
600 * Segment Receive Test
602 * ------- ------- -------------------------------------------
603 * 0 0 SEG.SEQ = RCV.NXT
604 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
605 * >0 0 not acceptable
606 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
607 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
609 * Note that when receiving a SYN segment in the LISTEN state,
610 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
611 * described in RFC 793, page 66.
613 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
614 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
615 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
616 if (V_tcp_insecure_rst ||
617 th->th_seq == sc->sc_irs + 1) {
618 syncache_drop(sc, sch);
619 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
621 "%s; %s: Our SYN|ACK was rejected, "
622 "connection attempt aborted by remote "
625 TCPSTAT_INC(tcps_sc_reset);
627 TCPSTAT_INC(tcps_badrst);
628 /* Send challenge ACK. */
629 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
630 log(LOG_DEBUG, "%s; %s: RST with invalid "
631 " SEQ %u != NXT %u (+WND %u), "
632 "sending challenge ACK\n",
634 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
635 syncache_respond(sc, sch, m, TH_ACK);
638 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
639 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
640 "NXT %u (+WND %u), segment ignored\n",
642 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
643 TCPSTAT_INC(tcps_badrst);
653 syncache_badack(struct in_conninfo *inc)
656 struct syncache_head *sch;
658 sc = syncache_lookup(inc, &sch); /* returns locked sch */
659 SCH_LOCK_ASSERT(sch);
661 syncache_drop(sc, sch);
662 TCPSTAT_INC(tcps_sc_badack);
668 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq)
671 struct syncache_head *sch;
673 sc = syncache_lookup(inc, &sch); /* returns locked sch */
674 SCH_LOCK_ASSERT(sch);
678 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
679 if (ntohl(th_seq) != sc->sc_iss)
683 * If we've rertransmitted 3 times and this is our second error,
684 * we remove the entry. Otherwise, we allow it to continue on.
685 * This prevents us from incorrectly nuking an entry during a
686 * spurious network outage.
690 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
691 sc->sc_flags |= SCF_UNREACH;
694 syncache_drop(sc, sch);
695 TCPSTAT_INC(tcps_sc_unreach);
701 * Build a new TCP socket structure from a syncache entry.
703 * On success return the newly created socket with its underlying inp locked.
705 static struct socket *
706 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
708 struct tcp_function_block *blk;
709 struct inpcb *inp = NULL;
715 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
718 * Ok, create the full blown connection, and set things up
719 * as they would have been set up if we had created the
720 * connection when the SYN arrived. If we can't create
721 * the connection, abort it.
723 so = sonewconn(lso, 0);
726 * Drop the connection; we will either send a RST or
727 * have the peer retransmit its SYN again after its
730 TCPSTAT_INC(tcps_listendrop);
731 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
732 log(LOG_DEBUG, "%s; %s: Socket create failed "
733 "due to limits or memory shortage\n",
740 mac_socketpeer_set_from_mbuf(m, so);
744 inp->inp_inc.inc_fibnum = so->so_fibnum;
747 * Exclusive pcbinfo lock is not required in syncache socket case even
748 * if two inpcb locks can be acquired simultaneously:
749 * - the inpcb in LISTEN state,
750 * - the newly created inp.
752 * In this case, an inp cannot be at same time in LISTEN state and
753 * just created by an accept() call.
755 INP_HASH_WLOCK(&V_tcbinfo);
757 /* Insert new socket into PCB hash list. */
758 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
760 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
761 inp->inp_vflag &= ~INP_IPV4;
762 inp->inp_vflag |= INP_IPV6;
763 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
765 inp->inp_vflag &= ~INP_IPV6;
766 inp->inp_vflag |= INP_IPV4;
768 inp->inp_laddr = sc->sc_inc.inc_laddr;
774 * If there's an mbuf and it has a flowid, then let's initialise the
775 * inp with that particular flowid.
777 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
778 inp->inp_flowid = m->m_pkthdr.flowid;
779 inp->inp_flowtype = M_HASHTYPE_GET(m);
781 inp->inp_numa_domain = m->m_pkthdr.numa_domain;
786 * Install in the reservation hash table for now, but don't yet
787 * install a connection group since the full 4-tuple isn't yet
790 inp->inp_lport = sc->sc_inc.inc_lport;
791 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
793 * Undo the assignments above if we failed to
794 * put the PCB on the hash lists.
797 if (sc->sc_inc.inc_flags & INC_ISIPV6)
798 inp->in6p_laddr = in6addr_any;
801 inp->inp_laddr.s_addr = INADDR_ANY;
803 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
804 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
809 INP_HASH_WUNLOCK(&V_tcbinfo);
813 if (inp->inp_vflag & INP_IPV6PROTO) {
814 struct inpcb *oinp = sotoinpcb(lso);
817 * Inherit socket options from the listening socket.
818 * Note that in6p_inputopts are not (and should not be)
819 * copied, since it stores previously received options and is
820 * used to detect if each new option is different than the
821 * previous one and hence should be passed to a user.
822 * If we copied in6p_inputopts, a user would not be able to
823 * receive options just after calling the accept system call.
825 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
826 if (oinp->in6p_outputopts)
827 inp->in6p_outputopts =
828 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
831 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
832 struct in6_addr laddr6;
833 struct sockaddr_in6 sin6;
835 sin6.sin6_family = AF_INET6;
836 sin6.sin6_len = sizeof(sin6);
837 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
838 sin6.sin6_port = sc->sc_inc.inc_fport;
839 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
840 laddr6 = inp->in6p_laddr;
841 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
842 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
843 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
844 thread0.td_ucred, m)) != 0) {
845 inp->in6p_laddr = laddr6;
846 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
847 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
852 INP_HASH_WUNLOCK(&V_tcbinfo);
855 /* Override flowlabel from in6_pcbconnect. */
856 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
857 inp->inp_flow |= sc->sc_flowlabel;
860 #if defined(INET) && defined(INET6)
865 struct in_addr laddr;
866 struct sockaddr_in sin;
868 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
870 if (inp->inp_options == NULL) {
871 inp->inp_options = sc->sc_ipopts;
872 sc->sc_ipopts = NULL;
875 sin.sin_family = AF_INET;
876 sin.sin_len = sizeof(sin);
877 sin.sin_addr = sc->sc_inc.inc_faddr;
878 sin.sin_port = sc->sc_inc.inc_fport;
879 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
880 laddr = inp->inp_laddr;
881 if (inp->inp_laddr.s_addr == INADDR_ANY)
882 inp->inp_laddr = sc->sc_inc.inc_laddr;
883 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
884 thread0.td_ucred, m)) != 0) {
885 inp->inp_laddr = laddr;
886 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
887 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
892 INP_HASH_WUNLOCK(&V_tcbinfo);
897 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
898 /* Copy old policy into new socket's. */
899 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
900 printf("syncache_socket: could not copy policy\n");
902 INP_HASH_WUNLOCK(&V_tcbinfo);
904 tcp_state_change(tp, TCPS_SYN_RECEIVED);
905 tp->iss = sc->sc_iss;
906 tp->irs = sc->sc_irs;
909 blk = sototcpcb(lso)->t_fb;
910 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
912 * Our parents t_fb was not the default,
913 * we need to release our ref on tp->t_fb and
914 * pickup one on the new entry.
916 struct tcp_function_block *rblk;
918 rblk = find_and_ref_tcp_fb(blk);
919 KASSERT(rblk != NULL,
920 ("cannot find blk %p out of syncache?", blk));
921 if (tp->t_fb->tfb_tcp_fb_fini)
922 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
923 refcount_release(&tp->t_fb->tfb_refcnt);
926 * XXXrrs this is quite dangerous, it is possible
927 * for the new function to fail to init. We also
928 * are not asking if the handoff_is_ok though at
929 * the very start thats probalbly ok.
931 if (tp->t_fb->tfb_tcp_fb_init) {
932 (*tp->t_fb->tfb_tcp_fb_init)(tp);
935 tp->snd_wl1 = sc->sc_irs;
936 tp->snd_max = tp->iss + 1;
937 tp->snd_nxt = tp->iss + 1;
938 tp->rcv_up = sc->sc_irs + 1;
939 tp->rcv_wnd = sc->sc_wnd;
940 tp->rcv_adv += tp->rcv_wnd;
941 tp->last_ack_sent = tp->rcv_nxt;
943 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
944 if (sc->sc_flags & SCF_NOOPT)
945 tp->t_flags |= TF_NOOPT;
947 if (sc->sc_flags & SCF_WINSCALE) {
948 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
949 tp->snd_scale = sc->sc_requested_s_scale;
950 tp->request_r_scale = sc->sc_requested_r_scale;
952 if (sc->sc_flags & SCF_TIMESTAMP) {
953 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
954 tp->ts_recent = sc->sc_tsreflect;
955 tp->ts_recent_age = tcp_ts_getticks();
956 tp->ts_offset = sc->sc_tsoff;
958 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
959 if (sc->sc_flags & SCF_SIGNATURE)
960 tp->t_flags |= TF_SIGNATURE;
962 if (sc->sc_flags & SCF_SACK)
963 tp->t_flags |= TF_SACK_PERMIT;
966 if (sc->sc_flags & SCF_ECN)
967 tp->t_flags |= TF_ECN_PERMIT;
970 * Set up MSS and get cached values from tcp_hostcache.
971 * This might overwrite some of the defaults we just set.
973 tcp_mss(tp, sc->sc_peer_mss);
976 * If the SYN,ACK was retransmitted, indicate that CWND to be
977 * limited to one segment in cc_conn_init().
978 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
980 if (sc->sc_rxmits > 1)
985 * Allow a TOE driver to install its hooks. Note that we hold the
986 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
987 * new connection before the TOE driver has done its thing.
989 if (ADDED_BY_TOE(sc)) {
990 struct toedev *tod = sc->sc_tod;
992 tod->tod_offload_socket(tod, sc->sc_todctx, so);
996 * Copy and activate timers.
998 tp->t_keepinit = sototcpcb(lso)->t_keepinit;
999 tp->t_keepidle = sototcpcb(lso)->t_keepidle;
1000 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
1001 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1002 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1004 TCPSTAT_INC(tcps_accepts);
1016 * This function gets called when we receive an ACK for a
1017 * socket in the LISTEN state. We look up the connection
1018 * in the syncache, and if its there, we pull it out of
1019 * the cache and turn it into a full-blown connection in
1020 * the SYN-RECEIVED state.
1022 * On syncache_socket() success the newly created socket
1023 * has its underlying inp locked.
1026 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1027 struct socket **lsop, struct mbuf *m)
1029 struct syncache *sc;
1030 struct syncache_head *sch;
1031 struct syncache scs;
1035 * Global TCP locks are held because we manipulate the PCB lists
1036 * and create a new socket.
1038 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1039 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1040 ("%s: can handle only ACK", __func__));
1042 sc = syncache_lookup(inc, &sch); /* returns locked sch */
1043 SCH_LOCK_ASSERT(sch);
1047 * Test code for syncookies comparing the syncache stored
1048 * values with the reconstructed values from the cookie.
1051 syncookie_cmp(inc, sch, sc, th, to, *lsop);
1056 * There is no syncache entry, so see if this ACK is
1057 * a returning syncookie. To do this, first:
1058 * A. Check if syncookies are used in case of syncache
1060 * B. See if this socket has had a syncache entry dropped in
1061 * the recent past. We don't want to accept a bogus
1062 * syncookie if we've never received a SYN or accept it
1064 * C. check that the syncookie is valid. If it is, then
1065 * cobble up a fake syncache entry, and return.
1067 if (!V_tcp_syncookies) {
1069 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1070 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1071 "segment rejected (syncookies disabled)\n",
1075 if (!V_tcp_syncookiesonly &&
1076 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1078 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1079 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1080 "segment rejected (no syncache entry)\n",
1084 bzero(&scs, sizeof(scs));
1085 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
1088 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1089 log(LOG_DEBUG, "%s; %s: Segment failed "
1090 "SYNCOOKIE authentication, segment rejected "
1091 "(probably spoofed)\n", s, __func__);
1094 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1095 /* If received ACK has MD5 signature, check it. */
1096 if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1097 (!TCPMD5_ENABLED() ||
1098 TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1100 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1101 log(LOG_DEBUG, "%s; %s: Segment rejected, "
1102 "MD5 signature doesn't match.\n",
1106 TCPSTAT_INC(tcps_sig_err_sigopt);
1107 return (-1); /* Do not send RST */
1109 #endif /* TCP_SIGNATURE */
1111 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1113 * If listening socket requested TCP digests, check that
1114 * received ACK has signature and it is correct.
1115 * If not, drop the ACK and leave sc entry in th cache,
1116 * because SYN was received with correct signature.
1118 if (sc->sc_flags & SCF_SIGNATURE) {
1119 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1121 TCPSTAT_INC(tcps_sig_err_nosigopt);
1123 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1124 log(LOG_DEBUG, "%s; %s: Segment "
1125 "rejected, MD5 signature wasn't "
1126 "provided.\n", s, __func__);
1129 return (-1); /* Do not send RST */
1131 if (!TCPMD5_ENABLED() ||
1132 TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1133 /* Doesn't match or no SA */
1135 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1136 log(LOG_DEBUG, "%s; %s: Segment "
1137 "rejected, MD5 signature doesn't "
1138 "match.\n", s, __func__);
1141 return (-1); /* Do not send RST */
1144 #endif /* TCP_SIGNATURE */
1146 * Pull out the entry to unlock the bucket row.
1148 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1149 * tcp_state_change(). The tcpcb is not existent at this
1150 * moment. A new one will be allocated via syncache_socket->
1151 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1152 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1154 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1155 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1158 if (ADDED_BY_TOE(sc)) {
1159 struct toedev *tod = sc->sc_tod;
1161 tod->tod_syncache_removed(tod, sc->sc_todctx);
1168 * Segment validation:
1169 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1171 if (th->th_ack != sc->sc_iss + 1) {
1172 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1173 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1174 "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1179 * The SEQ must fall in the window starting at the received
1180 * initial receive sequence number + 1 (the SYN).
1182 if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1183 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1184 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1185 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1186 "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1191 * If timestamps were not negotiated during SYN/ACK they
1192 * must not appear on any segment during this session.
1194 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1195 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1196 log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1197 "segment rejected\n", s, __func__);
1202 * If timestamps were negotiated during SYN/ACK they should
1203 * appear on every segment during this session.
1204 * XXXAO: This is only informal as there have been unverified
1205 * reports of non-compliants stacks.
1207 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1208 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1209 log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1210 "no action\n", s, __func__);
1216 *lsop = syncache_socket(sc, *lsop, m);
1219 TCPSTAT_INC(tcps_sc_aborted);
1221 TCPSTAT_INC(tcps_sc_completed);
1223 /* how do we find the inp for the new socket? */
1228 if (sc != NULL && sc != &scs)
1237 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
1238 uint64_t response_cookie)
1242 unsigned int *pending_counter;
1245 * Global TCP locks are held because we manipulate the PCB lists
1246 * and create a new socket.
1248 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1250 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
1251 *lsop = syncache_socket(sc, *lsop, m);
1252 if (*lsop == NULL) {
1253 TCPSTAT_INC(tcps_sc_aborted);
1254 atomic_subtract_int(pending_counter, 1);
1256 soisconnected(*lsop);
1257 inp = sotoinpcb(*lsop);
1258 tp = intotcpcb(inp);
1259 tp->t_flags |= TF_FASTOPEN;
1260 tp->t_tfo_cookie.server = response_cookie;
1261 tp->snd_max = tp->iss;
1262 tp->snd_nxt = tp->iss;
1263 tp->t_tfo_pending = pending_counter;
1264 TCPSTAT_INC(tcps_sc_completed);
1269 * Given a LISTEN socket and an inbound SYN request, add
1270 * this to the syn cache, and send back a segment:
1271 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1274 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1275 * Doing so would require that we hold onto the data and deliver it
1276 * to the application. However, if we are the target of a SYN-flood
1277 * DoS attack, an attacker could send data which would eventually
1278 * consume all available buffer space if it were ACKed. By not ACKing
1279 * the data, we avoid this DoS scenario.
1281 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1282 * cookie is processed and a new socket is created. In this case, any data
1283 * accompanying the SYN will be queued to the socket by tcp_input() and will
1284 * be ACKed either when the application sends response data or the delayed
1285 * ACK timer expires, whichever comes first.
1288 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1289 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1294 struct syncache *sc = NULL;
1295 struct syncache_head *sch;
1296 struct mbuf *ipopts = NULL;
1298 int win, ip_ttl, ip_tos;
1302 int autoflowlabel = 0;
1305 struct label *maclabel;
1307 struct syncache scs;
1309 uint64_t tfo_response_cookie;
1310 unsigned int *tfo_pending = NULL;
1311 int tfo_cookie_valid = 0;
1312 int tfo_response_cookie_valid = 0;
1314 INP_WLOCK_ASSERT(inp); /* listen socket */
1315 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1316 ("%s: unexpected tcp flags", __func__));
1319 * Combine all so/tp operations very early to drop the INP lock as
1323 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1325 cred = crhold(so->so_cred);
1328 if ((inc->inc_flags & INC_ISIPV6) &&
1329 (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1332 ip_ttl = inp->inp_ip_ttl;
1333 ip_tos = inp->inp_ip_tos;
1334 win = so->sol_sbrcv_hiwat;
1335 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1337 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
1338 (tp->t_tfo_pending != NULL) &&
1339 (to->to_flags & TOF_FASTOPEN)) {
1341 * Limit the number of pending TFO connections to
1342 * approximately half of the queue limit. This prevents TFO
1343 * SYN floods from starving the service by filling the
1344 * listen queue with bogus TFO connections.
1346 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1347 (so->sol_qlimit / 2)) {
1350 result = tcp_fastopen_check_cookie(inc,
1351 to->to_tfo_cookie, to->to_tfo_len,
1352 &tfo_response_cookie);
1353 tfo_cookie_valid = (result > 0);
1354 tfo_response_cookie_valid = (result >= 0);
1358 * Remember the TFO pending counter as it will have to be
1359 * decremented below if we don't make it to syncache_tfo_expand().
1361 tfo_pending = tp->t_tfo_pending;
1364 /* By the time we drop the lock these should no longer be used. */
1369 if (mac_syncache_init(&maclabel) != 0) {
1373 mac_syncache_create(maclabel, inp);
1375 if (!tfo_cookie_valid)
1379 * Remember the IP options, if any.
1382 if (!(inc->inc_flags & INC_ISIPV6))
1385 ipopts = (m) ? ip_srcroute(m) : NULL;
1390 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1392 * If listening socket requested TCP digests, check that received
1393 * SYN has signature and it is correct. If signature doesn't match
1394 * or TCP_SIGNATURE support isn't enabled, drop the packet.
1396 if (ltflags & TF_SIGNATURE) {
1397 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1398 TCPSTAT_INC(tcps_sig_err_nosigopt);
1401 if (!TCPMD5_ENABLED() ||
1402 TCPMD5_INPUT(m, th, to->to_signature) != 0)
1405 #endif /* TCP_SIGNATURE */
1407 * See if we already have an entry for this connection.
1408 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1410 * XXX: should the syncache be re-initialized with the contents
1411 * of the new SYN here (which may have different options?)
1413 * XXX: We do not check the sequence number to see if this is a
1414 * real retransmit or a new connection attempt. The question is
1415 * how to handle such a case; either ignore it as spoofed, or
1416 * drop the current entry and create a new one?
1418 sc = syncache_lookup(inc, &sch); /* returns locked entry */
1419 SCH_LOCK_ASSERT(sch);
1421 if (tfo_cookie_valid)
1423 TCPSTAT_INC(tcps_sc_dupsyn);
1426 * If we were remembering a previous source route,
1427 * forget it and use the new one we've been given.
1430 (void) m_free(sc->sc_ipopts);
1431 sc->sc_ipopts = ipopts;
1434 * Update timestamp if present.
1436 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1437 sc->sc_tsreflect = to->to_tsval;
1439 sc->sc_flags &= ~SCF_TIMESTAMP;
1442 * Since we have already unconditionally allocated label
1443 * storage, free it up. The syncache entry will already
1444 * have an initialized label we can use.
1446 mac_syncache_destroy(&maclabel);
1448 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1449 /* Retransmit SYN|ACK and reset retransmit count. */
1450 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1451 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1452 "resetting timer and retransmitting SYN|ACK\n",
1456 if (syncache_respond(sc, sch, m, TH_SYN|TH_ACK) == 0) {
1458 syncache_timeout(sc, sch, 1);
1459 TCPSTAT_INC(tcps_sndacks);
1460 TCPSTAT_INC(tcps_sndtotal);
1466 if (tfo_cookie_valid) {
1467 bzero(&scs, sizeof(scs));
1472 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1475 * The zone allocator couldn't provide more entries.
1476 * Treat this as if the cache was full; drop the oldest
1477 * entry and insert the new one.
1479 TCPSTAT_INC(tcps_sc_zonefail);
1480 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1481 sch->sch_last_overflow = time_uptime;
1482 syncache_drop(sc, sch);
1484 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1486 if (V_tcp_syncookies) {
1487 bzero(&scs, sizeof(scs));
1492 (void) m_free(ipopts);
1499 if (!tfo_cookie_valid && tfo_response_cookie_valid)
1500 sc->sc_tfo_cookie = &tfo_response_cookie;
1503 * Fill in the syncache values.
1506 sc->sc_label = maclabel;
1510 sc->sc_ipopts = ipopts;
1511 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1513 if (!(inc->inc_flags & INC_ISIPV6))
1516 sc->sc_ip_tos = ip_tos;
1517 sc->sc_ip_ttl = ip_ttl;
1521 sc->sc_todctx = todctx;
1523 sc->sc_irs = th->th_seq;
1524 sc->sc_iss = arc4random();
1526 sc->sc_flowlabel = 0;
1529 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1530 * win was derived from socket earlier in the function.
1533 win = imin(win, TCP_MAXWIN);
1536 if (V_tcp_do_rfc1323) {
1538 * A timestamp received in a SYN makes
1539 * it ok to send timestamp requests and replies.
1541 if (to->to_flags & TOF_TS) {
1542 sc->sc_tsreflect = to->to_tsval;
1543 sc->sc_flags |= SCF_TIMESTAMP;
1544 sc->sc_tsoff = tcp_new_ts_offset(inc);
1546 if (to->to_flags & TOF_SCALE) {
1550 * Pick the smallest possible scaling factor that
1551 * will still allow us to scale up to sb_max, aka
1552 * kern.ipc.maxsockbuf.
1554 * We do this because there are broken firewalls that
1555 * will corrupt the window scale option, leading to
1556 * the other endpoint believing that our advertised
1557 * window is unscaled. At scale factors larger than
1558 * 5 the unscaled window will drop below 1500 bytes,
1559 * leading to serious problems when traversing these
1562 * With the default maxsockbuf of 256K, a scale factor
1563 * of 3 will be chosen by this algorithm. Those who
1564 * choose a larger maxsockbuf should watch out
1565 * for the compatibility problems mentioned above.
1567 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1568 * or <SYN,ACK>) segment itself is never scaled.
1570 while (wscale < TCP_MAX_WINSHIFT &&
1571 (TCP_MAXWIN << wscale) < sb_max)
1573 sc->sc_requested_r_scale = wscale;
1574 sc->sc_requested_s_scale = to->to_wscale;
1575 sc->sc_flags |= SCF_WINSCALE;
1578 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1580 * If listening socket requested TCP digests, flag this in the
1581 * syncache so that syncache_respond() will do the right thing
1584 if (ltflags & TF_SIGNATURE)
1585 sc->sc_flags |= SCF_SIGNATURE;
1586 #endif /* TCP_SIGNATURE */
1587 if (to->to_flags & TOF_SACKPERM)
1588 sc->sc_flags |= SCF_SACK;
1589 if (to->to_flags & TOF_MSS)
1590 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */
1591 if (ltflags & TF_NOOPT)
1592 sc->sc_flags |= SCF_NOOPT;
1593 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1594 sc->sc_flags |= SCF_ECN;
1596 if (V_tcp_syncookies)
1597 sc->sc_iss = syncookie_generate(sch, sc);
1599 if (autoflowlabel) {
1600 if (V_tcp_syncookies)
1601 sc->sc_flowlabel = sc->sc_iss;
1603 sc->sc_flowlabel = ip6_randomflowlabel();
1604 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1609 if (tfo_cookie_valid) {
1610 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie);
1611 /* INP_WUNLOCK(inp) will be performed by the caller */
1616 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1618 * Do a standard 3-way handshake.
1620 if (syncache_respond(sc, sch, m, TH_SYN|TH_ACK) == 0) {
1621 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1623 else if (sc != &scs)
1624 syncache_insert(sc, sch); /* locks and unlocks sch */
1625 TCPSTAT_INC(tcps_sndacks);
1626 TCPSTAT_INC(tcps_sndtotal);
1630 TCPSTAT_INC(tcps_sc_dropped);
1635 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1642 * If tfo_pending is not NULL here, then a TFO SYN that did not
1643 * result in a new socket was processed and the associated pending
1644 * counter has not yet been decremented. All such TFO processing paths
1645 * transit this point.
1647 if (tfo_pending != NULL)
1648 tcp_fastopen_decrement_counter(tfo_pending);
1655 mac_syncache_destroy(&maclabel);
1661 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment,
1662 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1665 syncache_respond(struct syncache *sc, struct syncache_head *sch,
1666 const struct mbuf *m0, int flags)
1668 struct ip *ip = NULL;
1670 struct tcphdr *th = NULL;
1671 int optlen, error = 0; /* Make compiler happy */
1672 u_int16_t hlen, tlen, mssopt;
1675 struct ip6_hdr *ip6 = NULL;
1679 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1682 tlen = hlen + sizeof(struct tcphdr);
1684 /* Determine MSS we advertize to other end of connection. */
1685 mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss);
1687 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1688 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1689 ("syncache: mbuf too small"));
1691 /* Create the IP+TCP header from scratch. */
1692 m = m_gethdr(M_NOWAIT, MT_DATA);
1696 mac_syncache_create_mbuf(sc->sc_label, m);
1698 m->m_data += max_linkhdr;
1700 m->m_pkthdr.len = tlen;
1701 m->m_pkthdr.rcvif = NULL;
1704 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1705 ip6 = mtod(m, struct ip6_hdr *);
1706 ip6->ip6_vfc = IPV6_VERSION;
1707 ip6->ip6_nxt = IPPROTO_TCP;
1708 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1709 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1710 ip6->ip6_plen = htons(tlen - hlen);
1711 /* ip6_hlim is set after checksum */
1712 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1713 ip6->ip6_flow |= sc->sc_flowlabel;
1715 th = (struct tcphdr *)(ip6 + 1);
1718 #if defined(INET6) && defined(INET)
1723 ip = mtod(m, struct ip *);
1724 ip->ip_v = IPVERSION;
1725 ip->ip_hl = sizeof(struct ip) >> 2;
1726 ip->ip_len = htons(tlen);
1730 ip->ip_p = IPPROTO_TCP;
1731 ip->ip_src = sc->sc_inc.inc_laddr;
1732 ip->ip_dst = sc->sc_inc.inc_faddr;
1733 ip->ip_ttl = sc->sc_ip_ttl;
1734 ip->ip_tos = sc->sc_ip_tos;
1737 * See if we should do MTU discovery. Route lookups are
1738 * expensive, so we will only unset the DF bit if:
1740 * 1) path_mtu_discovery is disabled
1741 * 2) the SCF_UNREACH flag has been set
1743 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1744 ip->ip_off |= htons(IP_DF);
1746 th = (struct tcphdr *)(ip + 1);
1749 th->th_sport = sc->sc_inc.inc_lport;
1750 th->th_dport = sc->sc_inc.inc_fport;
1753 th->th_seq = htonl(sc->sc_iss);
1755 th->th_seq = htonl(sc->sc_iss + 1);
1756 th->th_ack = htonl(sc->sc_irs + 1);
1757 th->th_off = sizeof(struct tcphdr) >> 2;
1759 th->th_flags = flags;
1760 th->th_win = htons(sc->sc_wnd);
1763 if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) {
1764 th->th_flags |= TH_ECE;
1765 TCPSTAT_INC(tcps_ecn_shs);
1768 /* Tack on the TCP options. */
1769 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1772 if (flags & TH_SYN) {
1774 to.to_flags = TOF_MSS;
1775 if (sc->sc_flags & SCF_WINSCALE) {
1776 to.to_wscale = sc->sc_requested_r_scale;
1777 to.to_flags |= TOF_SCALE;
1779 if (sc->sc_flags & SCF_SACK)
1780 to.to_flags |= TOF_SACKPERM;
1781 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1782 if (sc->sc_flags & SCF_SIGNATURE)
1783 to.to_flags |= TOF_SIGNATURE;
1785 if (sc->sc_tfo_cookie) {
1786 to.to_flags |= TOF_FASTOPEN;
1787 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1788 to.to_tfo_cookie = sc->sc_tfo_cookie;
1789 /* don't send cookie again when retransmitting response */
1790 sc->sc_tfo_cookie = NULL;
1793 if (sc->sc_flags & SCF_TIMESTAMP) {
1794 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1795 to.to_tsecr = sc->sc_tsreflect;
1796 to.to_flags |= TOF_TS;
1798 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1800 /* Adjust headers by option size. */
1801 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1803 m->m_pkthdr.len += optlen;
1805 if (sc->sc_inc.inc_flags & INC_ISIPV6)
1806 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1809 ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1810 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1811 if (sc->sc_flags & SCF_SIGNATURE) {
1812 KASSERT(to.to_flags & TOF_SIGNATURE,
1813 ("tcp_addoptions() didn't set tcp_signature"));
1815 /* NOTE: to.to_signature is inside of mbuf */
1816 if (!TCPMD5_ENABLED() ||
1817 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1826 M_SETFIB(m, sc->sc_inc.inc_fibnum);
1827 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1829 * If we have peer's SYN and it has a flowid, then let's assign it to
1830 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid
1831 * to SYN|ACK due to lack of inp here.
1833 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1834 m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1835 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1838 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1839 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1840 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1842 ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1844 if (ADDED_BY_TOE(sc)) {
1845 struct toedev *tod = sc->sc_tod;
1847 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1852 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
1853 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1856 #if defined(INET6) && defined(INET)
1861 m->m_pkthdr.csum_flags = CSUM_TCP;
1862 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1863 htons(tlen + optlen - hlen + IPPROTO_TCP));
1865 if (ADDED_BY_TOE(sc)) {
1866 struct toedev *tod = sc->sc_tod;
1868 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1873 TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
1874 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1881 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1882 * that exceed the capacity of the syncache by avoiding the storage of any
1883 * of the SYNs we receive. Syncookies defend against blind SYN flooding
1884 * attacks where the attacker does not have access to our responses.
1886 * Syncookies encode and include all necessary information about the
1887 * connection setup within the SYN|ACK that we send back. That way we
1888 * can avoid keeping any local state until the ACK to our SYN|ACK returns
1889 * (if ever). Normally the syncache and syncookies are running in parallel
1890 * with the latter taking over when the former is exhausted. When matching
1891 * syncache entry is found the syncookie is ignored.
1893 * The only reliable information persisting the 3WHS is our initial sequence
1894 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient
1895 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1896 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK
1897 * returns and signifies a legitimate connection if it matches the ACK.
1899 * The available space of 32 bits to store the hash and to encode the SYN
1900 * option information is very tight and we should have at least 24 bits for
1901 * the MAC to keep the number of guesses by blind spoofing reasonably high.
1903 * SYN option information we have to encode to fully restore a connection:
1904 * MSS: is imporant to chose an optimal segment size to avoid IP level
1905 * fragmentation along the path. The common MSS values can be encoded
1906 * in a 3-bit table. Uncommon values are captured by the next lower value
1907 * in the table leading to a slight increase in packetization overhead.
1908 * WSCALE: is necessary to allow large windows to be used for high delay-
1909 * bandwidth product links. Not scaling the window when it was initially
1910 * negotiated is bad for performance as lack of scaling further decreases
1911 * the apparent available send window. We only need to encode the WSCALE
1912 * we received from the remote end. Our end can be recalculated at any
1913 * time. The common WSCALE values can be encoded in a 3-bit table.
1914 * Uncommon values are captured by the next lower value in the table
1915 * making us under-estimate the available window size halving our
1916 * theoretically possible maximum throughput for that connection.
1917 * SACK: Greatly assists in packet loss recovery and requires 1 bit.
1918 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
1919 * that are included in all segments on a connection. We enable them when
1922 * Security of syncookies and attack vectors:
1924 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
1925 * together with the gloabl secret to make it unique per connection attempt.
1926 * Thus any change of any of those parameters results in a different MAC output
1927 * in an unpredictable way unless a collision is encountered. 24 bits of the
1928 * MAC are embedded into the ISS.
1930 * To prevent replay attacks two rotating global secrets are updated with a
1931 * new random value every 15 seconds. The life-time of a syncookie is thus
1934 * Vector 1: Attacking the secret. This requires finding a weakness in the
1935 * MAC itself or the way it is used here. The attacker can do a chosen plain
1936 * text attack by varying and testing the all parameters under his control.
1937 * The strength depends on the size and randomness of the secret, and the
1938 * cryptographic security of the MAC function. Due to the constant updating
1939 * of the secret the attacker has at most 29.999 seconds to find the secret
1940 * and launch spoofed connections. After that he has to start all over again.
1942 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC
1943 * size an average of 4,823 attempts are required for a 50% chance of success
1944 * to spoof a single syncookie (birthday collision paradox). However the
1945 * attacker is blind and doesn't know if one of his attempts succeeded unless
1946 * he has a side channel to interfere success from. A single connection setup
1947 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
1948 * This many attempts are required for each one blind spoofed connection. For
1949 * every additional spoofed connection he has to launch another N attempts.
1950 * Thus for a sustained rate 100 spoofed connections per second approximately
1951 * 1,800,000 packets per second would have to be sent.
1953 * NB: The MAC function should be fast so that it doesn't become a CPU
1954 * exhaustion attack vector itself.
1957 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations
1958 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
1959 * http://cr.yp.to/syncookies.html (overview)
1960 * http://cr.yp.to/syncookies/archive (details)
1963 * Schematic construction of a syncookie enabled Initial Sequence Number:
1965 * 12345678901234567890123456789012
1966 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
1968 * x 24 MAC (truncated)
1969 * W 3 Send Window Scale index
1971 * S 1 SACK permitted
1972 * P 1 Odd/even secret
1976 * Distribution and probability of certain MSS values. Those in between are
1977 * rounded down to the next lower one.
1978 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
1979 * .2% .3% 5% 7% 7% 20% 15% 45%
1981 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
1984 * Distribution and probability of certain WSCALE values. We have to map the
1985 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
1986 * bits based on prevalence of certain values. Where we don't have an exact
1987 * match for are rounded down to the next lower one letting us under-estimate
1988 * the true available window. At the moment this would happen only for the
1989 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
1990 * and window size). The absence of the WSCALE option (no scaling in either
1991 * direction) is encoded with index zero.
1992 * [WSCALE values histograms, Allman, 2012]
1993 * X 10 10 35 5 6 14 10% by host
1994 * X 11 4 5 5 18 49 3% by connections
1996 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
1999 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed
2000 * and good cryptographic properties.
2003 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2004 uint8_t *secbits, uintptr_t secmod)
2007 uint32_t siphash[2];
2009 SipHash24_Init(&ctx);
2010 SipHash_SetKey(&ctx, secbits);
2011 switch (inc->inc_flags & INC_ISIPV6) {
2014 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2015 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2020 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2021 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2025 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2026 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2027 SipHash_Update(&ctx, &irs, sizeof(irs));
2028 SipHash_Update(&ctx, &flags, sizeof(flags));
2029 SipHash_Update(&ctx, &secmod, sizeof(secmod));
2030 SipHash_Final((u_int8_t *)&siphash, &ctx);
2032 return (siphash[0] ^ siphash[1]);
2036 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2038 u_int i, secbit, wscale;
2041 union syncookie cookie;
2043 SCH_LOCK_ASSERT(sch);
2047 /* Map our computed MSS into the 3-bit index. */
2048 for (i = nitems(tcp_sc_msstab) - 1;
2049 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2052 cookie.flags.mss_idx = i;
2055 * Map the send window scale into the 3-bit index but only if
2056 * the wscale option was received.
2058 if (sc->sc_flags & SCF_WINSCALE) {
2059 wscale = sc->sc_requested_s_scale;
2060 for (i = nitems(tcp_sc_wstab) - 1;
2061 tcp_sc_wstab[i] > wscale && i > 0;
2064 cookie.flags.wscale_idx = i;
2067 /* Can we do SACK? */
2068 if (sc->sc_flags & SCF_SACK)
2069 cookie.flags.sack_ok = 1;
2071 /* Which of the two secrets to use. */
2072 secbit = sch->sch_sc->secret.oddeven & 0x1;
2073 cookie.flags.odd_even = secbit;
2075 secbits = sch->sch_sc->secret.key[secbit];
2076 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2080 * Put the flags into the hash and XOR them to get better ISS number
2081 * variance. This doesn't enhance the cryptographic strength and is
2082 * done to prevent the 8 cookie bits from showing up directly on the
2086 iss |= cookie.cookie ^ (hash >> 24);
2088 TCPSTAT_INC(tcps_sc_sendcookie);
2092 static struct syncache *
2093 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2094 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2100 int wnd, wscale = 0;
2101 union syncookie cookie;
2103 SCH_LOCK_ASSERT(sch);
2106 * Pull information out of SYN-ACK/ACK and revert sequence number
2109 ack = th->th_ack - 1;
2110 seq = th->th_seq - 1;
2113 * Unpack the flags containing enough information to restore the
2116 cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2118 /* Which of the two secrets to use. */
2119 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
2121 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2123 /* The recomputed hash matches the ACK if this was a genuine cookie. */
2124 if ((ack & ~0xff) != (hash & ~0xff))
2127 /* Fill in the syncache values. */
2129 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2130 sc->sc_ipopts = NULL;
2135 switch (inc->inc_flags & INC_ISIPV6) {
2138 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2139 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2144 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2145 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
2150 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2152 /* We can simply recompute receive window scale we sent earlier. */
2153 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2156 /* Only use wscale if it was enabled in the orignal SYN. */
2157 if (cookie.flags.wscale_idx > 0) {
2158 sc->sc_requested_r_scale = wscale;
2159 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2160 sc->sc_flags |= SCF_WINSCALE;
2163 wnd = lso->sol_sbrcv_hiwat;
2165 wnd = imin(wnd, TCP_MAXWIN);
2168 if (cookie.flags.sack_ok)
2169 sc->sc_flags |= SCF_SACK;
2171 if (to->to_flags & TOF_TS) {
2172 sc->sc_flags |= SCF_TIMESTAMP;
2173 sc->sc_tsreflect = to->to_tsval;
2174 sc->sc_tsoff = tcp_new_ts_offset(inc);
2177 if (to->to_flags & TOF_SIGNATURE)
2178 sc->sc_flags |= SCF_SIGNATURE;
2182 TCPSTAT_INC(tcps_sc_recvcookie);
2188 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2189 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2192 struct syncache scs, *scx;
2195 bzero(&scs, sizeof(scs));
2196 scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
2198 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2202 if (sc->sc_peer_mss != scx->sc_peer_mss)
2203 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2204 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2206 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2207 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2208 s, __func__, sc->sc_requested_r_scale,
2209 scx->sc_requested_r_scale);
2211 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2212 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2213 s, __func__, sc->sc_requested_s_scale,
2214 scx->sc_requested_s_scale);
2216 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2217 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2224 #endif /* INVARIANTS */
2227 syncookie_reseed(void *arg)
2229 struct tcp_syncache *sc = arg;
2234 * Reseeding the secret doesn't have to be protected by a lock.
2235 * It only must be ensured that the new random values are visible
2236 * to all CPUs in a SMP environment. The atomic with release
2237 * semantics ensures that.
2239 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2240 secbits = sc->secret.key[secbit];
2241 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2242 atomic_add_rel_int(&sc->secret.oddeven, 1);
2244 /* Reschedule ourself. */
2245 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2249 * Exports the syncache entries to userland so that netstat can display
2250 * them alongside the other sockets. This function is intended to be
2251 * called only from tcp_pcblist.
2253 * Due to concurrency on an active system, the number of pcbs exported
2254 * may have no relation to max_pcbs. max_pcbs merely indicates the
2255 * amount of space the caller allocated for this function to use.
2258 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
2261 struct syncache *sc;
2262 struct syncache_head *sch;
2263 int count, error, i;
2265 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
2266 sch = &V_tcp_syncache.hashbase[i];
2268 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2269 if (count >= max_pcbs) {
2273 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2275 bzero(&xt, sizeof(xt));
2276 xt.xt_len = sizeof(xt);
2277 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2278 xt.xt_inp.inp_vflag = INP_IPV6;
2280 xt.xt_inp.inp_vflag = INP_IPV4;
2281 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2282 sizeof (struct in_conninfo));
2283 xt.t_state = TCPS_SYN_RECEIVED;
2284 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2285 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2286 xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2287 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2288 error = SYSCTL_OUT(req, &xt, sizeof xt);
2298 *pcbs_exported = count;