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_systm.h>
73 #include <netinet/ip.h>
74 #include <netinet/in_var.h>
75 #include <netinet/in_pcb.h>
76 #include <netinet/ip_var.h>
77 #include <netinet/ip_options.h>
79 #include <netinet/ip6.h>
80 #include <netinet/icmp6.h>
81 #include <netinet6/nd6.h>
82 #include <netinet6/ip6_var.h>
83 #include <netinet6/in6_pcb.h>
85 #include <netinet/tcp.h>
86 #include <netinet/tcp_fastopen.h>
87 #include <netinet/tcp_fsm.h>
88 #include <netinet/tcp_seq.h>
89 #include <netinet/tcp_timer.h>
90 #include <netinet/tcp_var.h>
91 #include <netinet/tcp_syncache.h>
93 #include <netinet6/tcp6_var.h>
96 #include <netinet/toecore.h>
99 #include <netipsec/ipsec_support.h>
101 #include <machine/in_cksum.h>
103 #include <security/mac/mac_framework.h>
105 static VNET_DEFINE(int, tcp_syncookies) = 1;
106 #define V_tcp_syncookies VNET(tcp_syncookies)
107 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
108 &VNET_NAME(tcp_syncookies), 0,
109 "Use TCP SYN cookies if the syncache overflows");
111 static VNET_DEFINE(int, tcp_syncookiesonly) = 0;
112 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly)
113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
114 &VNET_NAME(tcp_syncookiesonly), 0,
115 "Use only TCP SYN cookies");
117 static VNET_DEFINE(int, functions_inherit_listen_socket_stack) = 1;
118 #define V_functions_inherit_listen_socket_stack \
119 VNET(functions_inherit_listen_socket_stack)
120 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack,
121 CTLFLAG_VNET | CTLFLAG_RW,
122 &VNET_NAME(functions_inherit_listen_socket_stack), 0,
123 "Inherit listen socket's stack");
126 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
129 static void syncache_drop(struct syncache *, struct syncache_head *);
130 static void syncache_free(struct syncache *);
131 static void syncache_insert(struct syncache *, struct syncache_head *);
132 static int syncache_respond(struct syncache *, struct syncache_head *, int,
133 const struct mbuf *);
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 syncookie_reseed(void *);
149 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
150 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
155 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
156 * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds,
157 * the odds are that the user has given up attempting to connect by then.
159 #define SYNCACHE_MAXREXMTS 3
161 /* Arbitrary values */
162 #define TCP_SYNCACHE_HASHSIZE 512
163 #define TCP_SYNCACHE_BUCKETLIMIT 30
165 static VNET_DEFINE(struct tcp_syncache, tcp_syncache);
166 #define V_tcp_syncache VNET(tcp_syncache)
168 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0,
171 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
172 &VNET_NAME(tcp_syncache.bucket_limit), 0,
173 "Per-bucket hash limit for syncache");
175 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
176 &VNET_NAME(tcp_syncache.cache_limit), 0,
177 "Overall entry limit for syncache");
179 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
180 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
182 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
183 &VNET_NAME(tcp_syncache.hashsize), 0,
184 "Size of TCP syncache hashtable");
187 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
192 new = V_tcp_syncache.rexmt_limit;
193 error = sysctl_handle_int(oidp, &new, 0, req);
194 if ((error == 0) && (req->newptr != NULL)) {
195 if (new > TCP_MAXRXTSHIFT)
198 V_tcp_syncache.rexmt_limit = new;
203 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
204 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW,
205 &VNET_NAME(tcp_syncache.rexmt_limit), 0,
206 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
207 "Limit on SYN/ACK retransmissions");
209 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
210 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
211 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
212 "Send reset on socket allocation failure");
214 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
216 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx)
217 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx)
218 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED)
221 * Requires the syncache entry to be already removed from the bucket list.
224 syncache_free(struct syncache *sc)
228 (void) m_free(sc->sc_ipopts);
232 mac_syncache_destroy(&sc->sc_label);
235 uma_zfree(V_tcp_syncache.zone, sc);
243 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
244 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
245 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
246 V_tcp_syncache.hash_secret = arc4random();
248 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
249 &V_tcp_syncache.hashsize);
250 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
251 &V_tcp_syncache.bucket_limit);
252 if (!powerof2(V_tcp_syncache.hashsize) ||
253 V_tcp_syncache.hashsize == 0) {
254 printf("WARNING: syncache hash size is not a power of 2.\n");
255 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
257 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
260 V_tcp_syncache.cache_limit =
261 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
262 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
263 &V_tcp_syncache.cache_limit);
265 /* Allocate the hash table. */
266 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
267 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
270 V_tcp_syncache.vnet = curvnet;
273 /* Initialize the hash buckets. */
274 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
275 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
276 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
278 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
279 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
280 V_tcp_syncache.hashbase[i].sch_length = 0;
281 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
282 V_tcp_syncache.hashbase[i].sch_last_overflow =
283 -(SYNCOOKIE_LIFETIME + 1);
286 /* Create the syncache entry zone. */
287 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
288 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
289 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
290 V_tcp_syncache.cache_limit);
292 /* Start the SYN cookie reseeder callout. */
293 callout_init(&V_tcp_syncache.secret.reseed, 1);
294 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
295 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
296 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
297 syncookie_reseed, &V_tcp_syncache);
302 syncache_destroy(void)
304 struct syncache_head *sch;
305 struct syncache *sc, *nsc;
309 * Stop the re-seed timer before freeing resources. No need to
310 * possibly schedule it another time.
312 callout_drain(&V_tcp_syncache.secret.reseed);
314 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */
315 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
317 sch = &V_tcp_syncache.hashbase[i];
318 callout_drain(&sch->sch_timer);
321 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
322 syncache_drop(sc, sch);
324 KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
325 ("%s: sch->sch_bucket not empty", __func__));
326 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
327 __func__, sch->sch_length));
328 mtx_destroy(&sch->sch_mtx);
331 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
332 ("%s: cache_count not 0", __func__));
334 /* Free the allocated global resources. */
335 uma_zdestroy(V_tcp_syncache.zone);
336 free(V_tcp_syncache.hashbase, M_SYNCACHE);
341 * Inserts a syncache entry into the specified bucket row.
342 * Locks and unlocks the syncache_head autonomously.
345 syncache_insert(struct syncache *sc, struct syncache_head *sch)
347 struct syncache *sc2;
352 * Make sure that we don't overflow the per-bucket limit.
353 * If the bucket is full, toss the oldest element.
355 if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
356 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
357 ("sch->sch_length incorrect"));
358 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
359 sch->sch_last_overflow = time_uptime;
360 syncache_drop(sc2, sch);
361 TCPSTAT_INC(tcps_sc_bucketoverflow);
364 /* Put it into the bucket. */
365 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
369 if (ADDED_BY_TOE(sc)) {
370 struct toedev *tod = sc->sc_tod;
372 tod->tod_syncache_added(tod, sc->sc_todctx);
376 /* Reinitialize the bucket row's timer. */
377 if (sch->sch_length == 1)
378 sch->sch_nextc = ticks + INT_MAX;
379 syncache_timeout(sc, sch, 1);
383 TCPSTATES_INC(TCPS_SYN_RECEIVED);
384 TCPSTAT_INC(tcps_sc_added);
388 * Remove and free entry from syncache bucket row.
389 * Expects locked syncache head.
392 syncache_drop(struct syncache *sc, struct syncache_head *sch)
395 SCH_LOCK_ASSERT(sch);
397 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
398 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
402 if (ADDED_BY_TOE(sc)) {
403 struct toedev *tod = sc->sc_tod;
405 tod->tod_syncache_removed(tod, sc->sc_todctx);
413 * Engage/reengage time on bucket row.
416 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
420 if (sc->sc_rxmits == 0)
421 rexmt = TCPTV_RTOBASE;
423 TCPT_RANGESET(rexmt, TCPTV_RTOBASE * tcp_syn_backoff[sc->sc_rxmits],
424 tcp_rexmit_min, TCPTV_REXMTMAX);
425 sc->sc_rxttime = ticks + rexmt;
427 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
428 sch->sch_nextc = sc->sc_rxttime;
430 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
431 syncache_timer, (void *)sch);
436 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
437 * If we have retransmitted an entry the maximum number of times, expire it.
438 * One separate timer for each bucket row.
441 syncache_timer(void *xsch)
443 struct syncache_head *sch = (struct syncache_head *)xsch;
444 struct syncache *sc, *nsc;
448 CURVNET_SET(sch->sch_sc->vnet);
450 /* NB: syncache_head has already been locked by the callout. */
451 SCH_LOCK_ASSERT(sch);
454 * In the following cycle we may remove some entries and/or
455 * advance some timeouts, so re-initialize the bucket timer.
457 sch->sch_nextc = tick + INT_MAX;
459 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
461 * We do not check if the listen socket still exists
462 * and accept the case where the listen socket may be
463 * gone by the time we resend the SYN/ACK. We do
464 * not expect this to happens often. If it does,
465 * then the RST will be sent by the time the remote
466 * host does the SYN/ACK->ACK.
468 if (TSTMP_GT(sc->sc_rxttime, tick)) {
469 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
470 sch->sch_nextc = sc->sc_rxttime;
473 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
474 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
475 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
476 "giving up and removing syncache entry\n",
480 syncache_drop(sc, sch);
481 TCPSTAT_INC(tcps_sc_stale);
484 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
485 log(LOG_DEBUG, "%s; %s: Response timeout, "
486 "retransmitting (%u) SYN|ACK\n",
487 s, __func__, sc->sc_rxmits);
491 syncache_respond(sc, sch, 1, NULL);
492 TCPSTAT_INC(tcps_sc_retransmitted);
493 syncache_timeout(sc, sch, 0);
495 if (!TAILQ_EMPTY(&(sch)->sch_bucket))
496 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
497 syncache_timer, (void *)(sch));
502 * Find an entry in the syncache.
503 * Returns always with locked syncache_head plus a matching entry or NULL.
505 static struct syncache *
506 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
509 struct syncache_head *sch;
513 * The hash is built on foreign port + local port + foreign address.
514 * We rely on the fact that struct in_conninfo starts with 16 bits
515 * of foreign port, then 16 bits of local port then followed by 128
516 * bits of foreign address. In case of IPv4 address, the first 3
517 * 32-bit words of the address always are zeroes.
519 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
520 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
522 sch = &V_tcp_syncache.hashbase[hash];
526 /* Circle through bucket row to find matching entry. */
527 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
528 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
529 sizeof(struct in_endpoints)) == 0)
532 return (sc); /* Always returns with locked sch. */
536 * This function is called when we get a RST for a
537 * non-existent connection, so that we can see if the
538 * connection is in the syn cache. If it is, zap it.
541 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
544 struct syncache_head *sch;
547 sc = syncache_lookup(inc, &sch); /* returns locked sch */
548 SCH_LOCK_ASSERT(sch);
551 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
552 * See RFC 793 page 65, section SEGMENT ARRIVES.
554 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
555 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
556 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
557 "FIN flag set, segment ignored\n", s, __func__);
558 TCPSTAT_INC(tcps_badrst);
563 * No corresponding connection was found in syncache.
564 * If syncookies are enabled and possibly exclusively
565 * used, or we are under memory pressure, a valid RST
566 * may not find a syncache entry. In that case we're
567 * done and no SYN|ACK retransmissions will happen.
568 * Otherwise the RST was misdirected or spoofed.
571 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
572 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
573 "syncache entry (possibly syncookie only), "
574 "segment ignored\n", s, __func__);
575 TCPSTAT_INC(tcps_badrst);
580 * If the RST bit is set, check the sequence number to see
581 * if this is a valid reset segment.
583 * In all states except SYN-SENT, all reset (RST) segments
584 * are validated by checking their SEQ-fields. A reset is
585 * valid if its sequence number is in the window.
587 * The sequence number in the reset segment is normally an
588 * echo of our outgoing acknowlegement numbers, but some hosts
589 * send a reset with the sequence number at the rightmost edge
590 * of our receive window, and we have to handle this case.
592 if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
593 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
594 syncache_drop(sc, sch);
595 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
596 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, "
597 "connection attempt aborted by remote endpoint\n",
599 TCPSTAT_INC(tcps_sc_reset);
601 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
602 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
603 "IRS %u (+WND %u), segment ignored\n",
604 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd);
605 TCPSTAT_INC(tcps_badrst);
615 syncache_badack(struct in_conninfo *inc)
618 struct syncache_head *sch;
620 sc = syncache_lookup(inc, &sch); /* returns locked sch */
621 SCH_LOCK_ASSERT(sch);
623 syncache_drop(sc, sch);
624 TCPSTAT_INC(tcps_sc_badack);
630 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq)
633 struct syncache_head *sch;
635 sc = syncache_lookup(inc, &sch); /* returns locked sch */
636 SCH_LOCK_ASSERT(sch);
640 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
641 if (ntohl(th_seq) != sc->sc_iss)
645 * If we've rertransmitted 3 times and this is our second error,
646 * we remove the entry. Otherwise, we allow it to continue on.
647 * This prevents us from incorrectly nuking an entry during a
648 * spurious network outage.
652 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
653 sc->sc_flags |= SCF_UNREACH;
656 syncache_drop(sc, sch);
657 TCPSTAT_INC(tcps_sc_unreach);
663 * Build a new TCP socket structure from a syncache entry.
665 * On success return the newly created socket with its underlying inp locked.
667 static struct socket *
668 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
670 struct tcp_function_block *blk;
671 struct inpcb *inp = NULL;
677 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
680 * Ok, create the full blown connection, and set things up
681 * as they would have been set up if we had created the
682 * connection when the SYN arrived. If we can't create
683 * the connection, abort it.
685 so = sonewconn(lso, 0);
688 * Drop the connection; we will either send a RST or
689 * have the peer retransmit its SYN again after its
692 TCPSTAT_INC(tcps_listendrop);
693 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
694 log(LOG_DEBUG, "%s; %s: Socket create failed "
695 "due to limits or memory shortage\n",
702 mac_socketpeer_set_from_mbuf(m, so);
706 inp->inp_inc.inc_fibnum = so->so_fibnum;
709 * Exclusive pcbinfo lock is not required in syncache socket case even
710 * if two inpcb locks can be acquired simultaneously:
711 * - the inpcb in LISTEN state,
712 * - the newly created inp.
714 * In this case, an inp cannot be at same time in LISTEN state and
715 * just created by an accept() call.
717 INP_HASH_WLOCK(&V_tcbinfo);
719 /* Insert new socket into PCB hash list. */
720 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
722 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
723 inp->inp_vflag &= ~INP_IPV4;
724 inp->inp_vflag |= INP_IPV6;
725 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
727 inp->inp_vflag &= ~INP_IPV6;
728 inp->inp_vflag |= INP_IPV4;
730 inp->inp_laddr = sc->sc_inc.inc_laddr;
736 * If there's an mbuf and it has a flowid, then let's initialise the
737 * inp with that particular flowid.
739 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
740 inp->inp_flowid = m->m_pkthdr.flowid;
741 inp->inp_flowtype = M_HASHTYPE_GET(m);
745 * Install in the reservation hash table for now, but don't yet
746 * install a connection group since the full 4-tuple isn't yet
749 inp->inp_lport = sc->sc_inc.inc_lport;
750 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
752 * Undo the assignments above if we failed to
753 * put the PCB on the hash lists.
756 if (sc->sc_inc.inc_flags & INC_ISIPV6)
757 inp->in6p_laddr = in6addr_any;
760 inp->inp_laddr.s_addr = INADDR_ANY;
762 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
763 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
768 INP_HASH_WUNLOCK(&V_tcbinfo);
772 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
773 struct inpcb *oinp = sotoinpcb(lso);
774 struct in6_addr laddr6;
775 struct sockaddr_in6 sin6;
777 * Inherit socket options from the listening socket.
778 * Note that in6p_inputopts are not (and should not be)
779 * copied, since it stores previously received options and is
780 * used to detect if each new option is different than the
781 * previous one and hence should be passed to a user.
782 * If we copied in6p_inputopts, a user would not be able to
783 * receive options just after calling the accept system call.
785 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
786 if (oinp->in6p_outputopts)
787 inp->in6p_outputopts =
788 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
790 sin6.sin6_family = AF_INET6;
791 sin6.sin6_len = sizeof(sin6);
792 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
793 sin6.sin6_port = sc->sc_inc.inc_fport;
794 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
795 laddr6 = inp->in6p_laddr;
796 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
797 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
798 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
799 thread0.td_ucred, m)) != 0) {
800 inp->in6p_laddr = laddr6;
801 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
802 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
807 INP_HASH_WUNLOCK(&V_tcbinfo);
810 /* Override flowlabel from in6_pcbconnect. */
811 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
812 inp->inp_flow |= sc->sc_flowlabel;
815 #if defined(INET) && defined(INET6)
820 struct in_addr laddr;
821 struct sockaddr_in sin;
823 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
825 if (inp->inp_options == NULL) {
826 inp->inp_options = sc->sc_ipopts;
827 sc->sc_ipopts = NULL;
830 sin.sin_family = AF_INET;
831 sin.sin_len = sizeof(sin);
832 sin.sin_addr = sc->sc_inc.inc_faddr;
833 sin.sin_port = sc->sc_inc.inc_fport;
834 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
835 laddr = inp->inp_laddr;
836 if (inp->inp_laddr.s_addr == INADDR_ANY)
837 inp->inp_laddr = sc->sc_inc.inc_laddr;
838 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
839 thread0.td_ucred, m)) != 0) {
840 inp->inp_laddr = laddr;
841 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
842 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
847 INP_HASH_WUNLOCK(&V_tcbinfo);
852 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
853 /* Copy old policy into new socket's. */
854 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
855 printf("syncache_socket: could not copy policy\n");
857 INP_HASH_WUNLOCK(&V_tcbinfo);
859 tcp_state_change(tp, TCPS_SYN_RECEIVED);
860 tp->iss = sc->sc_iss;
861 tp->irs = sc->sc_irs;
864 blk = sototcpcb(lso)->t_fb;
865 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
867 * Our parents t_fb was not the default,
868 * we need to release our ref on tp->t_fb and
869 * pickup one on the new entry.
871 struct tcp_function_block *rblk;
873 rblk = find_and_ref_tcp_fb(blk);
874 KASSERT(rblk != NULL,
875 ("cannot find blk %p out of syncache?", blk));
876 if (tp->t_fb->tfb_tcp_fb_fini)
877 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
878 refcount_release(&tp->t_fb->tfb_refcnt);
881 * XXXrrs this is quite dangerous, it is possible
882 * for the new function to fail to init. We also
883 * are not asking if the handoff_is_ok though at
884 * the very start thats probalbly ok.
886 if (tp->t_fb->tfb_tcp_fb_init) {
887 (*tp->t_fb->tfb_tcp_fb_init)(tp);
890 tp->snd_wl1 = sc->sc_irs;
891 tp->snd_max = tp->iss + 1;
892 tp->snd_nxt = tp->iss + 1;
893 tp->rcv_up = sc->sc_irs + 1;
894 tp->rcv_wnd = sc->sc_wnd;
895 tp->rcv_adv += tp->rcv_wnd;
896 tp->last_ack_sent = tp->rcv_nxt;
898 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
899 if (sc->sc_flags & SCF_NOOPT)
900 tp->t_flags |= TF_NOOPT;
902 if (sc->sc_flags & SCF_WINSCALE) {
903 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
904 tp->snd_scale = sc->sc_requested_s_scale;
905 tp->request_r_scale = sc->sc_requested_r_scale;
907 if (sc->sc_flags & SCF_TIMESTAMP) {
908 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
909 tp->ts_recent = sc->sc_tsreflect;
910 tp->ts_recent_age = tcp_ts_getticks();
911 tp->ts_offset = sc->sc_tsoff;
913 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
914 if (sc->sc_flags & SCF_SIGNATURE)
915 tp->t_flags |= TF_SIGNATURE;
917 if (sc->sc_flags & SCF_SACK)
918 tp->t_flags |= TF_SACK_PERMIT;
921 if (sc->sc_flags & SCF_ECN)
922 tp->t_flags |= TF_ECN_PERMIT;
925 * Set up MSS and get cached values from tcp_hostcache.
926 * This might overwrite some of the defaults we just set.
928 tcp_mss(tp, sc->sc_peer_mss);
931 * If the SYN,ACK was retransmitted, indicate that CWND to be
932 * limited to one segment in cc_conn_init().
933 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
935 if (sc->sc_rxmits > 1)
940 * Allow a TOE driver to install its hooks. Note that we hold the
941 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
942 * new connection before the TOE driver has done its thing.
944 if (ADDED_BY_TOE(sc)) {
945 struct toedev *tod = sc->sc_tod;
947 tod->tod_offload_socket(tod, sc->sc_todctx, so);
951 * Copy and activate timers.
953 tp->t_keepinit = sototcpcb(lso)->t_keepinit;
954 tp->t_keepidle = sototcpcb(lso)->t_keepidle;
955 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
956 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
957 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
959 TCPSTAT_INC(tcps_accepts);
971 * This function gets called when we receive an ACK for a
972 * socket in the LISTEN state. We look up the connection
973 * in the syncache, and if its there, we pull it out of
974 * the cache and turn it into a full-blown connection in
975 * the SYN-RECEIVED state.
977 * On syncache_socket() success the newly created socket
978 * has its underlying inp locked.
981 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
982 struct socket **lsop, struct mbuf *m)
985 struct syncache_head *sch;
990 * Global TCP locks are held because we manipulate the PCB lists
991 * and create a new socket.
993 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
994 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
995 ("%s: can handle only ACK", __func__));
997 sc = syncache_lookup(inc, &sch); /* returns locked sch */
998 SCH_LOCK_ASSERT(sch);
1002 * Test code for syncookies comparing the syncache stored
1003 * values with the reconstructed values from the cookie.
1006 syncookie_cmp(inc, sch, sc, th, to, *lsop);
1011 * There is no syncache entry, so see if this ACK is
1012 * a returning syncookie. To do this, first:
1013 * A. Check if syncookies are used in case of syncache
1015 * B. See if this socket has had a syncache entry dropped in
1016 * the recent past. We don't want to accept a bogus
1017 * syncookie if we've never received a SYN or accept it
1019 * C. check that the syncookie is valid. If it is, then
1020 * cobble up a fake syncache entry, and return.
1022 if (!V_tcp_syncookies) {
1024 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1025 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1026 "segment rejected (syncookies disabled)\n",
1030 if (!V_tcp_syncookiesonly &&
1031 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1033 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1034 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1035 "segment rejected (no syncache entry)\n",
1039 bzero(&scs, sizeof(scs));
1040 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
1043 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1044 log(LOG_DEBUG, "%s; %s: Segment failed "
1045 "SYNCOOKIE authentication, segment rejected "
1046 "(probably spoofed)\n", s, __func__);
1049 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1050 /* If received ACK has MD5 signature, check it. */
1051 if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1052 (!TCPMD5_ENABLED() ||
1053 TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1055 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1056 log(LOG_DEBUG, "%s; %s: Segment rejected, "
1057 "MD5 signature doesn't match.\n",
1061 TCPSTAT_INC(tcps_sig_err_sigopt);
1062 return (-1); /* Do not send RST */
1064 #endif /* TCP_SIGNATURE */
1066 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1068 * If listening socket requested TCP digests, check that
1069 * received ACK has signature and it is correct.
1070 * If not, drop the ACK and leave sc entry in th cache,
1071 * because SYN was received with correct signature.
1073 if (sc->sc_flags & SCF_SIGNATURE) {
1074 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1076 TCPSTAT_INC(tcps_sig_err_nosigopt);
1078 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1079 log(LOG_DEBUG, "%s; %s: Segment "
1080 "rejected, MD5 signature wasn't "
1081 "provided.\n", s, __func__);
1084 return (-1); /* Do not send RST */
1086 if (!TCPMD5_ENABLED() ||
1087 TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1088 /* Doesn't match or no SA */
1090 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1091 log(LOG_DEBUG, "%s; %s: Segment "
1092 "rejected, MD5 signature doesn't "
1093 "match.\n", s, __func__);
1096 return (-1); /* Do not send RST */
1099 #endif /* TCP_SIGNATURE */
1101 * Pull out the entry to unlock the bucket row.
1103 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1104 * tcp_state_change(). The tcpcb is not existent at this
1105 * moment. A new one will be allocated via syncache_socket->
1106 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1107 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1109 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1110 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1113 if (ADDED_BY_TOE(sc)) {
1114 struct toedev *tod = sc->sc_tod;
1116 tod->tod_syncache_removed(tod, sc->sc_todctx);
1123 * Segment validation:
1124 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1126 if (th->th_ack != sc->sc_iss + 1) {
1127 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1128 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1129 "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1134 * The SEQ must fall in the window starting at the received
1135 * initial receive sequence number + 1 (the SYN).
1137 if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1138 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1139 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1140 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1141 "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1146 * If timestamps were not negotiated during SYN/ACK they
1147 * must not appear on any segment during this session.
1149 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1150 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1151 log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1152 "segment rejected\n", s, __func__);
1157 * If timestamps were negotiated during SYN/ACK they should
1158 * appear on every segment during this session.
1159 * XXXAO: This is only informal as there have been unverified
1160 * reports of non-compliants stacks.
1162 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1163 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1164 log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1165 "no action\n", s, __func__);
1171 *lsop = syncache_socket(sc, *lsop, m);
1174 TCPSTAT_INC(tcps_sc_aborted);
1176 TCPSTAT_INC(tcps_sc_completed);
1178 /* how do we find the inp for the new socket? */
1183 if (sc != NULL && sc != &scs)
1192 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
1193 uint64_t response_cookie)
1197 unsigned int *pending_counter;
1200 * Global TCP locks are held because we manipulate the PCB lists
1201 * and create a new socket.
1203 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1205 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
1206 *lsop = syncache_socket(sc, *lsop, m);
1207 if (*lsop == NULL) {
1208 TCPSTAT_INC(tcps_sc_aborted);
1209 atomic_subtract_int(pending_counter, 1);
1211 soisconnected(*lsop);
1212 inp = sotoinpcb(*lsop);
1213 tp = intotcpcb(inp);
1214 tp->t_flags |= TF_FASTOPEN;
1215 tp->t_tfo_cookie.server = response_cookie;
1216 tp->snd_max = tp->iss;
1217 tp->snd_nxt = tp->iss;
1218 tp->t_tfo_pending = pending_counter;
1219 TCPSTAT_INC(tcps_sc_completed);
1224 * Given a LISTEN socket and an inbound SYN request, add
1225 * this to the syn cache, and send back a segment:
1226 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1229 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1230 * Doing so would require that we hold onto the data and deliver it
1231 * to the application. However, if we are the target of a SYN-flood
1232 * DoS attack, an attacker could send data which would eventually
1233 * consume all available buffer space if it were ACKed. By not ACKing
1234 * the data, we avoid this DoS scenario.
1236 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1237 * cookie is processed and a new socket is created. In this case, any data
1238 * accompanying the SYN will be queued to the socket by tcp_input() and will
1239 * be ACKed either when the application sends response data or the delayed
1240 * ACK timer expires, whichever comes first.
1243 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1244 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1249 struct syncache *sc = NULL;
1250 struct syncache_head *sch;
1251 struct mbuf *ipopts = NULL;
1253 int win, ip_ttl, ip_tos;
1257 int autoflowlabel = 0;
1260 struct label *maclabel;
1262 struct syncache scs;
1264 uint64_t tfo_response_cookie;
1265 unsigned int *tfo_pending = NULL;
1266 int tfo_cookie_valid = 0;
1267 int tfo_response_cookie_valid = 0;
1269 INP_WLOCK_ASSERT(inp); /* listen socket */
1270 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1271 ("%s: unexpected tcp flags", __func__));
1274 * Combine all so/tp operations very early to drop the INP lock as
1278 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1280 cred = crhold(so->so_cred);
1283 if ((inc->inc_flags & INC_ISIPV6) &&
1284 (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1287 ip_ttl = inp->inp_ip_ttl;
1288 ip_tos = inp->inp_ip_tos;
1289 win = so->sol_sbrcv_hiwat;
1290 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1292 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
1293 (tp->t_tfo_pending != NULL) &&
1294 (to->to_flags & TOF_FASTOPEN)) {
1296 * Limit the number of pending TFO connections to
1297 * approximately half of the queue limit. This prevents TFO
1298 * SYN floods from starving the service by filling the
1299 * listen queue with bogus TFO connections.
1301 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1302 (so->sol_qlimit / 2)) {
1305 result = tcp_fastopen_check_cookie(inc,
1306 to->to_tfo_cookie, to->to_tfo_len,
1307 &tfo_response_cookie);
1308 tfo_cookie_valid = (result > 0);
1309 tfo_response_cookie_valid = (result >= 0);
1313 * Remember the TFO pending counter as it will have to be
1314 * decremented below if we don't make it to syncache_tfo_expand().
1316 tfo_pending = tp->t_tfo_pending;
1319 /* By the time we drop the lock these should no longer be used. */
1324 if (mac_syncache_init(&maclabel) != 0) {
1328 mac_syncache_create(maclabel, inp);
1330 if (!tfo_cookie_valid)
1334 * Remember the IP options, if any.
1337 if (!(inc->inc_flags & INC_ISIPV6))
1340 ipopts = (m) ? ip_srcroute(m) : NULL;
1345 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1347 * If listening socket requested TCP digests, check that received
1348 * SYN has signature and it is correct. If signature doesn't match
1349 * or TCP_SIGNATURE support isn't enabled, drop the packet.
1351 if (ltflags & TF_SIGNATURE) {
1352 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1353 TCPSTAT_INC(tcps_sig_err_nosigopt);
1356 if (!TCPMD5_ENABLED() ||
1357 TCPMD5_INPUT(m, th, to->to_signature) != 0)
1360 #endif /* TCP_SIGNATURE */
1362 * See if we already have an entry for this connection.
1363 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1365 * XXX: should the syncache be re-initialized with the contents
1366 * of the new SYN here (which may have different options?)
1368 * XXX: We do not check the sequence number to see if this is a
1369 * real retransmit or a new connection attempt. The question is
1370 * how to handle such a case; either ignore it as spoofed, or
1371 * drop the current entry and create a new one?
1373 sc = syncache_lookup(inc, &sch); /* returns locked entry */
1374 SCH_LOCK_ASSERT(sch);
1376 if (tfo_cookie_valid)
1378 TCPSTAT_INC(tcps_sc_dupsyn);
1381 * If we were remembering a previous source route,
1382 * forget it and use the new one we've been given.
1385 (void) m_free(sc->sc_ipopts);
1386 sc->sc_ipopts = ipopts;
1389 * Update timestamp if present.
1391 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1392 sc->sc_tsreflect = to->to_tsval;
1394 sc->sc_flags &= ~SCF_TIMESTAMP;
1397 * Since we have already unconditionally allocated label
1398 * storage, free it up. The syncache entry will already
1399 * have an initialized label we can use.
1401 mac_syncache_destroy(&maclabel);
1403 /* Retransmit SYN|ACK and reset retransmit count. */
1404 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1405 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1406 "resetting timer and retransmitting SYN|ACK\n",
1410 if (syncache_respond(sc, sch, 1, m) == 0) {
1412 syncache_timeout(sc, sch, 1);
1413 TCPSTAT_INC(tcps_sndacks);
1414 TCPSTAT_INC(tcps_sndtotal);
1420 if (tfo_cookie_valid) {
1421 bzero(&scs, sizeof(scs));
1426 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1429 * The zone allocator couldn't provide more entries.
1430 * Treat this as if the cache was full; drop the oldest
1431 * entry and insert the new one.
1433 TCPSTAT_INC(tcps_sc_zonefail);
1434 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1435 sch->sch_last_overflow = time_uptime;
1436 syncache_drop(sc, sch);
1438 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1440 if (V_tcp_syncookies) {
1441 bzero(&scs, sizeof(scs));
1446 (void) m_free(ipopts);
1453 if (!tfo_cookie_valid && tfo_response_cookie_valid)
1454 sc->sc_tfo_cookie = &tfo_response_cookie;
1457 * Fill in the syncache values.
1460 sc->sc_label = maclabel;
1464 sc->sc_ipopts = ipopts;
1465 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1467 if (!(inc->inc_flags & INC_ISIPV6))
1470 sc->sc_ip_tos = ip_tos;
1471 sc->sc_ip_ttl = ip_ttl;
1475 sc->sc_todctx = todctx;
1477 sc->sc_irs = th->th_seq;
1478 sc->sc_iss = arc4random();
1480 sc->sc_flowlabel = 0;
1483 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1484 * win was derived from socket earlier in the function.
1487 win = imin(win, TCP_MAXWIN);
1490 if (V_tcp_do_rfc1323) {
1492 * A timestamp received in a SYN makes
1493 * it ok to send timestamp requests and replies.
1495 if (to->to_flags & TOF_TS) {
1496 sc->sc_tsreflect = to->to_tsval;
1497 sc->sc_flags |= SCF_TIMESTAMP;
1499 if (to->to_flags & TOF_SCALE) {
1503 * Pick the smallest possible scaling factor that
1504 * will still allow us to scale up to sb_max, aka
1505 * kern.ipc.maxsockbuf.
1507 * We do this because there are broken firewalls that
1508 * will corrupt the window scale option, leading to
1509 * the other endpoint believing that our advertised
1510 * window is unscaled. At scale factors larger than
1511 * 5 the unscaled window will drop below 1500 bytes,
1512 * leading to serious problems when traversing these
1515 * With the default maxsockbuf of 256K, a scale factor
1516 * of 3 will be chosen by this algorithm. Those who
1517 * choose a larger maxsockbuf should watch out
1518 * for the compatibility problems mentioned above.
1520 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1521 * or <SYN,ACK>) segment itself is never scaled.
1523 while (wscale < TCP_MAX_WINSHIFT &&
1524 (TCP_MAXWIN << wscale) < sb_max)
1526 sc->sc_requested_r_scale = wscale;
1527 sc->sc_requested_s_scale = to->to_wscale;
1528 sc->sc_flags |= SCF_WINSCALE;
1531 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1533 * If listening socket requested TCP digests, flag this in the
1534 * syncache so that syncache_respond() will do the right thing
1537 if (ltflags & TF_SIGNATURE)
1538 sc->sc_flags |= SCF_SIGNATURE;
1539 #endif /* TCP_SIGNATURE */
1540 if (to->to_flags & TOF_SACKPERM)
1541 sc->sc_flags |= SCF_SACK;
1542 if (to->to_flags & TOF_MSS)
1543 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */
1544 if (ltflags & TF_NOOPT)
1545 sc->sc_flags |= SCF_NOOPT;
1546 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1547 sc->sc_flags |= SCF_ECN;
1549 if (V_tcp_syncookies)
1550 sc->sc_iss = syncookie_generate(sch, sc);
1552 if (autoflowlabel) {
1553 if (V_tcp_syncookies)
1554 sc->sc_flowlabel = sc->sc_iss;
1556 sc->sc_flowlabel = ip6_randomflowlabel();
1557 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1562 if (tfo_cookie_valid) {
1563 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie);
1564 /* INP_WUNLOCK(inp) will be performed by the caller */
1570 * Do a standard 3-way handshake.
1572 if (syncache_respond(sc, sch, 0, m) == 0) {
1573 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1575 else if (sc != &scs)
1576 syncache_insert(sc, sch); /* locks and unlocks sch */
1577 TCPSTAT_INC(tcps_sndacks);
1578 TCPSTAT_INC(tcps_sndtotal);
1582 TCPSTAT_INC(tcps_sc_dropped);
1591 * If tfo_pending is not NULL here, then a TFO SYN that did not
1592 * result in a new socket was processed and the associated pending
1593 * counter has not yet been decremented. All such TFO processing paths
1594 * transit this point.
1596 if (tfo_pending != NULL)
1597 tcp_fastopen_decrement_counter(tfo_pending);
1604 mac_syncache_destroy(&maclabel);
1610 * Send SYN|ACK to the peer. Either in response to the peer's SYN,
1611 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1614 syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked,
1615 const struct mbuf *m0)
1617 struct ip *ip = NULL;
1619 struct tcphdr *th = NULL;
1620 int optlen, error = 0; /* Make compiler happy */
1621 u_int16_t hlen, tlen, mssopt;
1624 struct ip6_hdr *ip6 = NULL;
1628 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1631 tlen = hlen + sizeof(struct tcphdr);
1633 /* Determine MSS we advertize to other end of connection. */
1634 mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss);
1636 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1637 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1638 ("syncache: mbuf too small"));
1640 /* Create the IP+TCP header from scratch. */
1641 m = m_gethdr(M_NOWAIT, MT_DATA);
1645 mac_syncache_create_mbuf(sc->sc_label, m);
1647 m->m_data += max_linkhdr;
1649 m->m_pkthdr.len = tlen;
1650 m->m_pkthdr.rcvif = NULL;
1653 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1654 ip6 = mtod(m, struct ip6_hdr *);
1655 ip6->ip6_vfc = IPV6_VERSION;
1656 ip6->ip6_nxt = IPPROTO_TCP;
1657 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1658 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1659 ip6->ip6_plen = htons(tlen - hlen);
1660 /* ip6_hlim is set after checksum */
1661 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1662 ip6->ip6_flow |= sc->sc_flowlabel;
1664 th = (struct tcphdr *)(ip6 + 1);
1667 #if defined(INET6) && defined(INET)
1672 ip = mtod(m, struct ip *);
1673 ip->ip_v = IPVERSION;
1674 ip->ip_hl = sizeof(struct ip) >> 2;
1675 ip->ip_len = htons(tlen);
1679 ip->ip_p = IPPROTO_TCP;
1680 ip->ip_src = sc->sc_inc.inc_laddr;
1681 ip->ip_dst = sc->sc_inc.inc_faddr;
1682 ip->ip_ttl = sc->sc_ip_ttl;
1683 ip->ip_tos = sc->sc_ip_tos;
1686 * See if we should do MTU discovery. Route lookups are
1687 * expensive, so we will only unset the DF bit if:
1689 * 1) path_mtu_discovery is disabled
1690 * 2) the SCF_UNREACH flag has been set
1692 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1693 ip->ip_off |= htons(IP_DF);
1695 th = (struct tcphdr *)(ip + 1);
1698 th->th_sport = sc->sc_inc.inc_lport;
1699 th->th_dport = sc->sc_inc.inc_fport;
1701 th->th_seq = htonl(sc->sc_iss);
1702 th->th_ack = htonl(sc->sc_irs + 1);
1703 th->th_off = sizeof(struct tcphdr) >> 2;
1705 th->th_flags = TH_SYN|TH_ACK;
1706 th->th_win = htons(sc->sc_wnd);
1709 if (sc->sc_flags & SCF_ECN) {
1710 th->th_flags |= TH_ECE;
1711 TCPSTAT_INC(tcps_ecn_shs);
1714 /* Tack on the TCP options. */
1715 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1719 to.to_flags = TOF_MSS;
1720 if (sc->sc_flags & SCF_WINSCALE) {
1721 to.to_wscale = sc->sc_requested_r_scale;
1722 to.to_flags |= TOF_SCALE;
1724 if (sc->sc_flags & SCF_TIMESTAMP) {
1725 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1726 to.to_tsecr = sc->sc_tsreflect;
1727 to.to_flags |= TOF_TS;
1729 if (sc->sc_flags & SCF_SACK)
1730 to.to_flags |= TOF_SACKPERM;
1731 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1732 if (sc->sc_flags & SCF_SIGNATURE)
1733 to.to_flags |= TOF_SIGNATURE;
1735 if (sc->sc_tfo_cookie) {
1736 to.to_flags |= TOF_FASTOPEN;
1737 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1738 to.to_tfo_cookie = sc->sc_tfo_cookie;
1739 /* don't send cookie again when retransmitting response */
1740 sc->sc_tfo_cookie = NULL;
1742 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1744 /* Adjust headers by option size. */
1745 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1747 m->m_pkthdr.len += optlen;
1749 if (sc->sc_inc.inc_flags & INC_ISIPV6)
1750 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1753 ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1754 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1755 if (sc->sc_flags & SCF_SIGNATURE) {
1756 KASSERT(to.to_flags & TOF_SIGNATURE,
1757 ("tcp_addoptions() didn't set tcp_signature"));
1759 /* NOTE: to.to_signature is inside of mbuf */
1760 if (!TCPMD5_ENABLED() ||
1761 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1770 M_SETFIB(m, sc->sc_inc.inc_fibnum);
1771 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1773 * If we have peer's SYN and it has a flowid, then let's assign it to
1774 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid
1775 * to SYN|ACK due to lack of inp here.
1777 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1778 m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1779 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1782 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1783 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1784 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1786 ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1788 if (ADDED_BY_TOE(sc)) {
1789 struct toedev *tod = sc->sc_tod;
1791 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1796 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1799 #if defined(INET6) && defined(INET)
1804 m->m_pkthdr.csum_flags = CSUM_TCP;
1805 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1806 htons(tlen + optlen - hlen + IPPROTO_TCP));
1808 if (ADDED_BY_TOE(sc)) {
1809 struct toedev *tod = sc->sc_tod;
1811 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1816 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1823 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1824 * that exceed the capacity of the syncache by avoiding the storage of any
1825 * of the SYNs we receive. Syncookies defend against blind SYN flooding
1826 * attacks where the attacker does not have access to our responses.
1828 * Syncookies encode and include all necessary information about the
1829 * connection setup within the SYN|ACK that we send back. That way we
1830 * can avoid keeping any local state until the ACK to our SYN|ACK returns
1831 * (if ever). Normally the syncache and syncookies are running in parallel
1832 * with the latter taking over when the former is exhausted. When matching
1833 * syncache entry is found the syncookie is ignored.
1835 * The only reliable information persisting the 3WHS is our initial sequence
1836 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient
1837 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1838 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK
1839 * returns and signifies a legitimate connection if it matches the ACK.
1841 * The available space of 32 bits to store the hash and to encode the SYN
1842 * option information is very tight and we should have at least 24 bits for
1843 * the MAC to keep the number of guesses by blind spoofing reasonably high.
1845 * SYN option information we have to encode to fully restore a connection:
1846 * MSS: is imporant to chose an optimal segment size to avoid IP level
1847 * fragmentation along the path. The common MSS values can be encoded
1848 * in a 3-bit table. Uncommon values are captured by the next lower value
1849 * in the table leading to a slight increase in packetization overhead.
1850 * WSCALE: is necessary to allow large windows to be used for high delay-
1851 * bandwidth product links. Not scaling the window when it was initially
1852 * negotiated is bad for performance as lack of scaling further decreases
1853 * the apparent available send window. We only need to encode the WSCALE
1854 * we received from the remote end. Our end can be recalculated at any
1855 * time. The common WSCALE values can be encoded in a 3-bit table.
1856 * Uncommon values are captured by the next lower value in the table
1857 * making us under-estimate the available window size halving our
1858 * theoretically possible maximum throughput for that connection.
1859 * SACK: Greatly assists in packet loss recovery and requires 1 bit.
1860 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
1861 * that are included in all segments on a connection. We enable them when
1864 * Security of syncookies and attack vectors:
1866 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
1867 * together with the gloabl secret to make it unique per connection attempt.
1868 * Thus any change of any of those parameters results in a different MAC output
1869 * in an unpredictable way unless a collision is encountered. 24 bits of the
1870 * MAC are embedded into the ISS.
1872 * To prevent replay attacks two rotating global secrets are updated with a
1873 * new random value every 15 seconds. The life-time of a syncookie is thus
1876 * Vector 1: Attacking the secret. This requires finding a weakness in the
1877 * MAC itself or the way it is used here. The attacker can do a chosen plain
1878 * text attack by varying and testing the all parameters under his control.
1879 * The strength depends on the size and randomness of the secret, and the
1880 * cryptographic security of the MAC function. Due to the constant updating
1881 * of the secret the attacker has at most 29.999 seconds to find the secret
1882 * and launch spoofed connections. After that he has to start all over again.
1884 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC
1885 * size an average of 4,823 attempts are required for a 50% chance of success
1886 * to spoof a single syncookie (birthday collision paradox). However the
1887 * attacker is blind and doesn't know if one of his attempts succeeded unless
1888 * he has a side channel to interfere success from. A single connection setup
1889 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
1890 * This many attempts are required for each one blind spoofed connection. For
1891 * every additional spoofed connection he has to launch another N attempts.
1892 * Thus for a sustained rate 100 spoofed connections per second approximately
1893 * 1,800,000 packets per second would have to be sent.
1895 * NB: The MAC function should be fast so that it doesn't become a CPU
1896 * exhaustion attack vector itself.
1899 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations
1900 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
1901 * http://cr.yp.to/syncookies.html (overview)
1902 * http://cr.yp.to/syncookies/archive (details)
1905 * Schematic construction of a syncookie enabled Initial Sequence Number:
1907 * 12345678901234567890123456789012
1908 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
1910 * x 24 MAC (truncated)
1911 * W 3 Send Window Scale index
1913 * S 1 SACK permitted
1914 * P 1 Odd/even secret
1918 * Distribution and probability of certain MSS values. Those in between are
1919 * rounded down to the next lower one.
1920 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
1921 * .2% .3% 5% 7% 7% 20% 15% 45%
1923 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
1926 * Distribution and probability of certain WSCALE values. We have to map the
1927 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
1928 * bits based on prevalence of certain values. Where we don't have an exact
1929 * match for are rounded down to the next lower one letting us under-estimate
1930 * the true available window. At the moment this would happen only for the
1931 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
1932 * and window size). The absence of the WSCALE option (no scaling in either
1933 * direction) is encoded with index zero.
1934 * [WSCALE values histograms, Allman, 2012]
1935 * X 10 10 35 5 6 14 10% by host
1936 * X 11 4 5 5 18 49 3% by connections
1938 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
1941 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed
1942 * and good cryptographic properties.
1945 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
1946 uint8_t *secbits, uintptr_t secmod)
1949 uint32_t siphash[2];
1951 SipHash24_Init(&ctx);
1952 SipHash_SetKey(&ctx, secbits);
1953 switch (inc->inc_flags & INC_ISIPV6) {
1956 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
1957 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
1962 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
1963 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
1967 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
1968 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
1969 SipHash_Update(&ctx, &irs, sizeof(irs));
1970 SipHash_Update(&ctx, &flags, sizeof(flags));
1971 SipHash_Update(&ctx, &secmod, sizeof(secmod));
1972 SipHash_Final((u_int8_t *)&siphash, &ctx);
1974 return (siphash[0] ^ siphash[1]);
1978 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
1980 u_int i, secbit, wscale;
1983 union syncookie cookie;
1985 SCH_LOCK_ASSERT(sch);
1989 /* Map our computed MSS into the 3-bit index. */
1990 for (i = nitems(tcp_sc_msstab) - 1;
1991 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
1994 cookie.flags.mss_idx = i;
1997 * Map the send window scale into the 3-bit index but only if
1998 * the wscale option was received.
2000 if (sc->sc_flags & SCF_WINSCALE) {
2001 wscale = sc->sc_requested_s_scale;
2002 for (i = nitems(tcp_sc_wstab) - 1;
2003 tcp_sc_wstab[i] > wscale && i > 0;
2006 cookie.flags.wscale_idx = i;
2009 /* Can we do SACK? */
2010 if (sc->sc_flags & SCF_SACK)
2011 cookie.flags.sack_ok = 1;
2013 /* Which of the two secrets to use. */
2014 secbit = sch->sch_sc->secret.oddeven & 0x1;
2015 cookie.flags.odd_even = secbit;
2017 secbits = sch->sch_sc->secret.key[secbit];
2018 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2022 * Put the flags into the hash and XOR them to get better ISS number
2023 * variance. This doesn't enhance the cryptographic strength and is
2024 * done to prevent the 8 cookie bits from showing up directly on the
2028 iss |= cookie.cookie ^ (hash >> 24);
2030 /* Randomize the timestamp. */
2031 if (sc->sc_flags & SCF_TIMESTAMP) {
2032 sc->sc_tsoff = arc4random() - tcp_ts_getticks();
2035 TCPSTAT_INC(tcps_sc_sendcookie);
2039 static struct syncache *
2040 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2041 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2047 int wnd, wscale = 0;
2048 union syncookie cookie;
2050 SCH_LOCK_ASSERT(sch);
2053 * Pull information out of SYN-ACK/ACK and revert sequence number
2056 ack = th->th_ack - 1;
2057 seq = th->th_seq - 1;
2060 * Unpack the flags containing enough information to restore the
2063 cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2065 /* Which of the two secrets to use. */
2066 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
2068 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2070 /* The recomputed hash matches the ACK if this was a genuine cookie. */
2071 if ((ack & ~0xff) != (hash & ~0xff))
2074 /* Fill in the syncache values. */
2076 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2077 sc->sc_ipopts = NULL;
2082 switch (inc->inc_flags & INC_ISIPV6) {
2085 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2086 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2091 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2092 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
2097 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2099 /* We can simply recompute receive window scale we sent earlier. */
2100 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2103 /* Only use wscale if it was enabled in the orignal SYN. */
2104 if (cookie.flags.wscale_idx > 0) {
2105 sc->sc_requested_r_scale = wscale;
2106 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2107 sc->sc_flags |= SCF_WINSCALE;
2110 wnd = lso->sol_sbrcv_hiwat;
2112 wnd = imin(wnd, TCP_MAXWIN);
2115 if (cookie.flags.sack_ok)
2116 sc->sc_flags |= SCF_SACK;
2118 if (to->to_flags & TOF_TS) {
2119 sc->sc_flags |= SCF_TIMESTAMP;
2120 sc->sc_tsreflect = to->to_tsval;
2121 sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks();
2124 if (to->to_flags & TOF_SIGNATURE)
2125 sc->sc_flags |= SCF_SIGNATURE;
2129 TCPSTAT_INC(tcps_sc_recvcookie);
2135 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2136 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2139 struct syncache scs, *scx;
2142 bzero(&scs, sizeof(scs));
2143 scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
2145 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2149 if (sc->sc_peer_mss != scx->sc_peer_mss)
2150 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2151 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2153 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2154 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2155 s, __func__, sc->sc_requested_r_scale,
2156 scx->sc_requested_r_scale);
2158 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2159 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2160 s, __func__, sc->sc_requested_s_scale,
2161 scx->sc_requested_s_scale);
2163 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2164 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2171 #endif /* INVARIANTS */
2174 syncookie_reseed(void *arg)
2176 struct tcp_syncache *sc = arg;
2181 * Reseeding the secret doesn't have to be protected by a lock.
2182 * It only must be ensured that the new random values are visible
2183 * to all CPUs in a SMP environment. The atomic with release
2184 * semantics ensures that.
2186 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2187 secbits = sc->secret.key[secbit];
2188 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2189 atomic_add_rel_int(&sc->secret.oddeven, 1);
2191 /* Reschedule ourself. */
2192 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2196 * Exports the syncache entries to userland so that netstat can display
2197 * them alongside the other sockets. This function is intended to be
2198 * called only from tcp_pcblist.
2200 * Due to concurrency on an active system, the number of pcbs exported
2201 * may have no relation to max_pcbs. max_pcbs merely indicates the
2202 * amount of space the caller allocated for this function to use.
2205 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
2208 struct syncache *sc;
2209 struct syncache_head *sch;
2210 int count, error, i;
2212 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
2213 sch = &V_tcp_syncache.hashbase[i];
2215 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2216 if (count >= max_pcbs) {
2220 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2222 bzero(&xt, sizeof(xt));
2223 xt.xt_len = sizeof(xt);
2224 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2225 xt.xt_inp.inp_vflag = INP_IPV6;
2227 xt.xt_inp.inp_vflag = INP_IPV4;
2228 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2229 sizeof (struct in_conninfo));
2230 xt.t_state = TCPS_SYN_RECEIVED;
2231 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2232 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2233 xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2234 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2235 error = SYSCTL_OUT(req, &xt, sizeof xt);
2245 *pcbs_exported = count;