2 * Copyright (c) 2001 McAfee, Inc.
3 * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG
6 * This software was developed for the FreeBSD Project by Jonathan Lemon
7 * and McAfee Research, the Security Research Division of McAfee, Inc. under
8 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
9 * DARPA CHATS research program. [2001 McAfee, Inc.]
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in the
18 * documentation and/or other materials provided with the distribution.
20 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 #include "opt_pcbgroup.h"
41 #include <sys/param.h>
42 #include <sys/systm.h>
44 #include <sys/refcount.h>
45 #include <sys/kernel.h>
46 #include <sys/sysctl.h>
47 #include <sys/limits.h>
49 #include <sys/mutex.h>
50 #include <sys/malloc.h>
52 #include <sys/proc.h> /* for proc0 declaration */
53 #include <sys/random.h>
54 #include <sys/socket.h>
55 #include <sys/socketvar.h>
56 #include <sys/syslog.h>
57 #include <sys/ucred.h>
60 #include <crypto/siphash/siphash.h>
65 #include <net/if_var.h>
66 #include <net/route.h>
69 #include <netinet/in.h>
70 #include <netinet/in_systm.h>
71 #include <netinet/ip.h>
72 #include <netinet/in_var.h>
73 #include <netinet/in_pcb.h>
74 #include <netinet/ip_var.h>
75 #include <netinet/ip_options.h>
77 #include <netinet/ip6.h>
78 #include <netinet/icmp6.h>
79 #include <netinet6/nd6.h>
80 #include <netinet6/ip6_var.h>
81 #include <netinet6/in6_pcb.h>
83 #include <netinet/tcp.h>
85 #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");
118 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
121 static void syncache_drop(struct syncache *, struct syncache_head *);
122 static void syncache_free(struct syncache *);
123 static void syncache_insert(struct syncache *, struct syncache_head *);
124 static int syncache_respond(struct syncache *, struct syncache_head *, int,
125 const struct mbuf *);
126 static struct socket *syncache_socket(struct syncache *, struct socket *,
128 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch,
130 static void syncache_timer(void *);
132 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
133 uint8_t *, uintptr_t);
134 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *);
135 static struct syncache
136 *syncookie_lookup(struct in_conninfo *, struct syncache_head *,
137 struct syncache *, struct tcphdr *, struct tcpopt *,
139 static void syncookie_reseed(void *);
141 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
142 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
147 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
148 * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds,
149 * the odds are that the user has given up attempting to connect by then.
151 #define SYNCACHE_MAXREXMTS 3
153 /* Arbitrary values */
154 #define TCP_SYNCACHE_HASHSIZE 512
155 #define TCP_SYNCACHE_BUCKETLIMIT 30
157 static VNET_DEFINE(struct tcp_syncache, tcp_syncache);
158 #define V_tcp_syncache VNET(tcp_syncache)
160 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0,
163 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
164 &VNET_NAME(tcp_syncache.bucket_limit), 0,
165 "Per-bucket hash limit for syncache");
167 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
168 &VNET_NAME(tcp_syncache.cache_limit), 0,
169 "Overall entry limit for syncache");
171 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
172 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
174 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
175 &VNET_NAME(tcp_syncache.hashsize), 0,
176 "Size of TCP syncache hashtable");
179 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
184 new = V_tcp_syncache.rexmt_limit;
185 error = sysctl_handle_int(oidp, &new, 0, req);
186 if ((error == 0) && (req->newptr != NULL)) {
187 if (new > TCP_MAXRXTSHIFT)
190 V_tcp_syncache.rexmt_limit = new;
195 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
196 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW,
197 &VNET_NAME(tcp_syncache.rexmt_limit), 0,
198 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
199 "Limit on SYN/ACK retransmissions");
201 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
202 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
203 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
204 "Send reset on socket allocation failure");
206 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
208 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx)
209 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx)
210 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED)
213 * Requires the syncache entry to be already removed from the bucket list.
216 syncache_free(struct syncache *sc)
220 (void) m_free(sc->sc_ipopts);
224 mac_syncache_destroy(&sc->sc_label);
227 uma_zfree(V_tcp_syncache.zone, sc);
235 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
236 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
237 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
238 V_tcp_syncache.hash_secret = arc4random();
240 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
241 &V_tcp_syncache.hashsize);
242 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
243 &V_tcp_syncache.bucket_limit);
244 if (!powerof2(V_tcp_syncache.hashsize) ||
245 V_tcp_syncache.hashsize == 0) {
246 printf("WARNING: syncache hash size is not a power of 2.\n");
247 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
249 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
252 V_tcp_syncache.cache_limit =
253 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
254 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
255 &V_tcp_syncache.cache_limit);
257 /* Allocate the hash table. */
258 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
259 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
262 V_tcp_syncache.vnet = curvnet;
265 /* Initialize the hash buckets. */
266 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
267 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
268 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
270 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
271 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
272 V_tcp_syncache.hashbase[i].sch_length = 0;
273 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
274 V_tcp_syncache.hashbase[i].sch_last_overflow =
275 -(SYNCOOKIE_LIFETIME + 1);
278 /* Create the syncache entry zone. */
279 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
280 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
281 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
282 V_tcp_syncache.cache_limit);
284 /* Start the SYN cookie reseeder callout. */
285 callout_init(&V_tcp_syncache.secret.reseed, 1);
286 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
287 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
288 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
289 syncookie_reseed, &V_tcp_syncache);
294 syncache_destroy(void)
296 struct syncache_head *sch;
297 struct syncache *sc, *nsc;
301 * Stop the re-seed timer before freeing resources. No need to
302 * possibly schedule it another time.
304 callout_drain(&V_tcp_syncache.secret.reseed);
306 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */
307 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
309 sch = &V_tcp_syncache.hashbase[i];
310 callout_drain(&sch->sch_timer);
313 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
314 syncache_drop(sc, sch);
316 KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
317 ("%s: sch->sch_bucket not empty", __func__));
318 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
319 __func__, sch->sch_length));
320 mtx_destroy(&sch->sch_mtx);
323 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
324 ("%s: cache_count not 0", __func__));
326 /* Free the allocated global resources. */
327 uma_zdestroy(V_tcp_syncache.zone);
328 free(V_tcp_syncache.hashbase, M_SYNCACHE);
333 * Inserts a syncache entry into the specified bucket row.
334 * Locks and unlocks the syncache_head autonomously.
337 syncache_insert(struct syncache *sc, struct syncache_head *sch)
339 struct syncache *sc2;
344 * Make sure that we don't overflow the per-bucket limit.
345 * If the bucket is full, toss the oldest element.
347 if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
348 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
349 ("sch->sch_length incorrect"));
350 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
351 sch->sch_last_overflow = time_uptime;
352 syncache_drop(sc2, sch);
353 TCPSTAT_INC(tcps_sc_bucketoverflow);
356 /* Put it into the bucket. */
357 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
361 if (ADDED_BY_TOE(sc)) {
362 struct toedev *tod = sc->sc_tod;
364 tod->tod_syncache_added(tod, sc->sc_todctx);
368 /* Reinitialize the bucket row's timer. */
369 if (sch->sch_length == 1)
370 sch->sch_nextc = ticks + INT_MAX;
371 syncache_timeout(sc, sch, 1);
375 TCPSTATES_INC(TCPS_SYN_RECEIVED);
376 TCPSTAT_INC(tcps_sc_added);
380 * Remove and free entry from syncache bucket row.
381 * Expects locked syncache head.
384 syncache_drop(struct syncache *sc, struct syncache_head *sch)
387 SCH_LOCK_ASSERT(sch);
389 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
390 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
394 if (ADDED_BY_TOE(sc)) {
395 struct toedev *tod = sc->sc_tod;
397 tod->tod_syncache_removed(tod, sc->sc_todctx);
405 * Engage/reengage time on bucket row.
408 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
412 if (sc->sc_rxmits == 0)
413 rexmt = TCPTV_RTOBASE;
415 TCPT_RANGESET(rexmt, TCPTV_RTOBASE * tcp_syn_backoff[sc->sc_rxmits],
416 tcp_rexmit_min, TCPTV_REXMTMAX);
417 sc->sc_rxttime = ticks + rexmt;
419 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
420 sch->sch_nextc = sc->sc_rxttime;
422 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
423 syncache_timer, (void *)sch);
428 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
429 * If we have retransmitted an entry the maximum number of times, expire it.
430 * One separate timer for each bucket row.
433 syncache_timer(void *xsch)
435 struct syncache_head *sch = (struct syncache_head *)xsch;
436 struct syncache *sc, *nsc;
440 CURVNET_SET(sch->sch_sc->vnet);
442 /* NB: syncache_head has already been locked by the callout. */
443 SCH_LOCK_ASSERT(sch);
446 * In the following cycle we may remove some entries and/or
447 * advance some timeouts, so re-initialize the bucket timer.
449 sch->sch_nextc = tick + INT_MAX;
451 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
453 * We do not check if the listen socket still exists
454 * and accept the case where the listen socket may be
455 * gone by the time we resend the SYN/ACK. We do
456 * not expect this to happens often. If it does,
457 * then the RST will be sent by the time the remote
458 * host does the SYN/ACK->ACK.
460 if (TSTMP_GT(sc->sc_rxttime, tick)) {
461 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
462 sch->sch_nextc = sc->sc_rxttime;
465 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
466 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
467 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
468 "giving up and removing syncache entry\n",
472 syncache_drop(sc, sch);
473 TCPSTAT_INC(tcps_sc_stale);
476 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
477 log(LOG_DEBUG, "%s; %s: Response timeout, "
478 "retransmitting (%u) SYN|ACK\n",
479 s, __func__, sc->sc_rxmits);
483 syncache_respond(sc, sch, 1, NULL);
484 TCPSTAT_INC(tcps_sc_retransmitted);
485 syncache_timeout(sc, sch, 0);
487 if (!TAILQ_EMPTY(&(sch)->sch_bucket))
488 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
489 syncache_timer, (void *)(sch));
494 * Find an entry in the syncache.
495 * Returns always with locked syncache_head plus a matching entry or NULL.
497 static struct syncache *
498 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
501 struct syncache_head *sch;
505 * The hash is built on foreign port + local port + foreign address.
506 * We rely on the fact that struct in_conninfo starts with 16 bits
507 * of foreign port, then 16 bits of local port then followed by 128
508 * bits of foreign address. In case of IPv4 address, the first 3
509 * 32-bit words of the address always are zeroes.
511 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
512 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
514 sch = &V_tcp_syncache.hashbase[hash];
518 /* Circle through bucket row to find matching entry. */
519 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
520 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
521 sizeof(struct in_endpoints)) == 0)
524 return (sc); /* Always returns with locked sch. */
528 * This function is called when we get a RST for a
529 * non-existent connection, so that we can see if the
530 * connection is in the syn cache. If it is, zap it.
533 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
536 struct syncache_head *sch;
539 sc = syncache_lookup(inc, &sch); /* returns locked sch */
540 SCH_LOCK_ASSERT(sch);
543 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
544 * See RFC 793 page 65, section SEGMENT ARRIVES.
546 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
547 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
548 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
549 "FIN flag set, segment ignored\n", s, __func__);
550 TCPSTAT_INC(tcps_badrst);
555 * No corresponding connection was found in syncache.
556 * If syncookies are enabled and possibly exclusively
557 * used, or we are under memory pressure, a valid RST
558 * may not find a syncache entry. In that case we're
559 * done and no SYN|ACK retransmissions will happen.
560 * Otherwise the RST was misdirected or spoofed.
563 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
564 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
565 "syncache entry (possibly syncookie only), "
566 "segment ignored\n", s, __func__);
567 TCPSTAT_INC(tcps_badrst);
572 * If the RST bit is set, check the sequence number to see
573 * if this is a valid reset segment.
575 * In all states except SYN-SENT, all reset (RST) segments
576 * are validated by checking their SEQ-fields. A reset is
577 * valid if its sequence number is in the window.
579 * The sequence number in the reset segment is normally an
580 * echo of our outgoing acknowlegement numbers, but some hosts
581 * send a reset with the sequence number at the rightmost edge
582 * of our receive window, and we have to handle this case.
584 if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
585 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
586 syncache_drop(sc, sch);
587 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
588 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, "
589 "connection attempt aborted by remote endpoint\n",
591 TCPSTAT_INC(tcps_sc_reset);
593 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
594 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
595 "IRS %u (+WND %u), segment ignored\n",
596 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd);
597 TCPSTAT_INC(tcps_badrst);
607 syncache_badack(struct in_conninfo *inc)
610 struct syncache_head *sch;
612 sc = syncache_lookup(inc, &sch); /* returns locked sch */
613 SCH_LOCK_ASSERT(sch);
615 syncache_drop(sc, sch);
616 TCPSTAT_INC(tcps_sc_badack);
622 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq)
625 struct syncache_head *sch;
627 sc = syncache_lookup(inc, &sch); /* returns locked sch */
628 SCH_LOCK_ASSERT(sch);
632 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
633 if (ntohl(th_seq) != sc->sc_iss)
637 * If we've rertransmitted 3 times and this is our second error,
638 * we remove the entry. Otherwise, we allow it to continue on.
639 * This prevents us from incorrectly nuking an entry during a
640 * spurious network outage.
644 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
645 sc->sc_flags |= SCF_UNREACH;
648 syncache_drop(sc, sch);
649 TCPSTAT_INC(tcps_sc_unreach);
655 * Build a new TCP socket structure from a syncache entry.
657 * On success return the newly created socket with its underlying inp locked.
659 static struct socket *
660 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
662 struct tcp_function_block *blk;
663 struct inpcb *inp = NULL;
669 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
672 * Ok, create the full blown connection, and set things up
673 * as they would have been set up if we had created the
674 * connection when the SYN arrived. If we can't create
675 * the connection, abort it.
677 so = sonewconn(lso, 0);
680 * Drop the connection; we will either send a RST or
681 * have the peer retransmit its SYN again after its
684 TCPSTAT_INC(tcps_listendrop);
685 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
686 log(LOG_DEBUG, "%s; %s: Socket create failed "
687 "due to limits or memory shortage\n",
694 mac_socketpeer_set_from_mbuf(m, so);
698 inp->inp_inc.inc_fibnum = so->so_fibnum;
701 * Exclusive pcbinfo lock is not required in syncache socket case even
702 * if two inpcb locks can be acquired simultaneously:
703 * - the inpcb in LISTEN state,
704 * - the newly created inp.
706 * In this case, an inp cannot be at same time in LISTEN state and
707 * just created by an accept() call.
709 INP_HASH_WLOCK(&V_tcbinfo);
711 /* Insert new socket into PCB hash list. */
712 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
714 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
715 inp->inp_vflag &= ~INP_IPV4;
716 inp->inp_vflag |= INP_IPV6;
717 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
719 inp->inp_vflag &= ~INP_IPV6;
720 inp->inp_vflag |= INP_IPV4;
722 inp->inp_laddr = sc->sc_inc.inc_laddr;
728 * If there's an mbuf and it has a flowid, then let's initialise the
729 * inp with that particular flowid.
731 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
732 inp->inp_flowid = m->m_pkthdr.flowid;
733 inp->inp_flowtype = M_HASHTYPE_GET(m);
737 * Install in the reservation hash table for now, but don't yet
738 * install a connection group since the full 4-tuple isn't yet
741 inp->inp_lport = sc->sc_inc.inc_lport;
742 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
744 * Undo the assignments above if we failed to
745 * put the PCB on the hash lists.
748 if (sc->sc_inc.inc_flags & INC_ISIPV6)
749 inp->in6p_laddr = in6addr_any;
752 inp->inp_laddr.s_addr = INADDR_ANY;
754 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
755 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
760 INP_HASH_WUNLOCK(&V_tcbinfo);
764 if (inp->inp_vflag & INP_IPV6PROTO) {
765 struct inpcb *oinp = sotoinpcb(lso);
768 * Inherit socket options from the listening socket.
769 * Note that in6p_inputopts are not (and should not be)
770 * copied, since it stores previously received options and is
771 * used to detect if each new option is different than the
772 * previous one and hence should be passed to a user.
773 * If we copied in6p_inputopts, a user would not be able to
774 * receive options just after calling the accept system call.
776 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
777 if (oinp->in6p_outputopts)
778 inp->in6p_outputopts =
779 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
782 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
783 struct in6_addr laddr6;
784 struct sockaddr_in6 sin6;
786 sin6.sin6_family = AF_INET6;
787 sin6.sin6_len = sizeof(sin6);
788 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
789 sin6.sin6_port = sc->sc_inc.inc_fport;
790 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
791 laddr6 = inp->in6p_laddr;
792 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
793 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
794 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
795 thread0.td_ucred, m)) != 0) {
796 inp->in6p_laddr = laddr6;
797 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
798 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
803 INP_HASH_WUNLOCK(&V_tcbinfo);
806 /* Override flowlabel from in6_pcbconnect. */
807 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
808 inp->inp_flow |= sc->sc_flowlabel;
811 #if defined(INET) && defined(INET6)
816 struct in_addr laddr;
817 struct sockaddr_in sin;
819 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
821 if (inp->inp_options == NULL) {
822 inp->inp_options = sc->sc_ipopts;
823 sc->sc_ipopts = NULL;
826 sin.sin_family = AF_INET;
827 sin.sin_len = sizeof(sin);
828 sin.sin_addr = sc->sc_inc.inc_faddr;
829 sin.sin_port = sc->sc_inc.inc_fport;
830 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
831 laddr = inp->inp_laddr;
832 if (inp->inp_laddr.s_addr == INADDR_ANY)
833 inp->inp_laddr = sc->sc_inc.inc_laddr;
834 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
835 thread0.td_ucred, m)) != 0) {
836 inp->inp_laddr = laddr;
837 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
838 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
843 INP_HASH_WUNLOCK(&V_tcbinfo);
848 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
849 /* Copy old policy into new socket's. */
850 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
851 printf("syncache_socket: could not copy policy\n");
853 INP_HASH_WUNLOCK(&V_tcbinfo);
855 tcp_state_change(tp, TCPS_SYN_RECEIVED);
856 tp->iss = sc->sc_iss;
857 tp->irs = sc->sc_irs;
860 blk = sototcpcb(lso)->t_fb;
861 if (blk != tp->t_fb) {
863 * Our parents t_fb was not the default,
864 * we need to release our ref on tp->t_fb and
865 * pickup one on the new entry.
867 struct tcp_function_block *rblk;
869 rblk = find_and_ref_tcp_fb(blk);
870 KASSERT(rblk != NULL,
871 ("cannot find blk %p out of syncache?", blk));
872 if (tp->t_fb->tfb_tcp_fb_fini)
873 (*tp->t_fb->tfb_tcp_fb_fini)(tp);
874 refcount_release(&tp->t_fb->tfb_refcnt);
876 if (tp->t_fb->tfb_tcp_fb_init) {
877 (*tp->t_fb->tfb_tcp_fb_init)(tp);
880 tp->snd_wl1 = sc->sc_irs;
881 tp->snd_max = tp->iss + 1;
882 tp->snd_nxt = tp->iss + 1;
883 tp->rcv_up = sc->sc_irs + 1;
884 tp->rcv_wnd = sc->sc_wnd;
885 tp->rcv_adv += tp->rcv_wnd;
886 tp->last_ack_sent = tp->rcv_nxt;
888 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
889 if (sc->sc_flags & SCF_NOOPT)
890 tp->t_flags |= TF_NOOPT;
892 if (sc->sc_flags & SCF_WINSCALE) {
893 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
894 tp->snd_scale = sc->sc_requested_s_scale;
895 tp->request_r_scale = sc->sc_requested_r_scale;
897 if (sc->sc_flags & SCF_TIMESTAMP) {
898 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
899 tp->ts_recent = sc->sc_tsreflect;
900 tp->ts_recent_age = tcp_ts_getticks();
901 tp->ts_offset = sc->sc_tsoff;
903 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
904 if (sc->sc_flags & SCF_SIGNATURE)
905 tp->t_flags |= TF_SIGNATURE;
907 if (sc->sc_flags & SCF_SACK)
908 tp->t_flags |= TF_SACK_PERMIT;
911 if (sc->sc_flags & SCF_ECN)
912 tp->t_flags |= TF_ECN_PERMIT;
915 * Set up MSS and get cached values from tcp_hostcache.
916 * This might overwrite some of the defaults we just set.
918 tcp_mss(tp, sc->sc_peer_mss);
921 * If the SYN,ACK was retransmitted, indicate that CWND to be
922 * limited to one segment in cc_conn_init().
923 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
925 if (sc->sc_rxmits > 1)
930 * Allow a TOE driver to install its hooks. Note that we hold the
931 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
932 * new connection before the TOE driver has done its thing.
934 if (ADDED_BY_TOE(sc)) {
935 struct toedev *tod = sc->sc_tod;
937 tod->tod_offload_socket(tod, sc->sc_todctx, so);
941 * Copy and activate timers.
943 tp->t_keepinit = sototcpcb(lso)->t_keepinit;
944 tp->t_keepidle = sototcpcb(lso)->t_keepidle;
945 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
946 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
947 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
949 TCPSTAT_INC(tcps_accepts);
961 * This function gets called when we receive an ACK for a
962 * socket in the LISTEN state. We look up the connection
963 * in the syncache, and if its there, we pull it out of
964 * the cache and turn it into a full-blown connection in
965 * the SYN-RECEIVED state.
967 * On syncache_socket() success the newly created socket
968 * has its underlying inp locked.
971 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
972 struct socket **lsop, struct mbuf *m)
975 struct syncache_head *sch;
980 * Global TCP locks are held because we manipulate the PCB lists
981 * and create a new socket.
983 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
984 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
985 ("%s: can handle only ACK", __func__));
987 sc = syncache_lookup(inc, &sch); /* returns locked sch */
988 SCH_LOCK_ASSERT(sch);
992 * Test code for syncookies comparing the syncache stored
993 * values with the reconstructed values from the cookie.
996 syncookie_cmp(inc, sch, sc, th, to, *lsop);
1001 * There is no syncache entry, so see if this ACK is
1002 * a returning syncookie. To do this, first:
1003 * A. Check if syncookies are used in case of syncache
1005 * B. See if this socket has had a syncache entry dropped in
1006 * the recent past. We don't want to accept a bogus
1007 * syncookie if we've never received a SYN or accept it
1009 * C. check that the syncookie is valid. If it is, then
1010 * cobble up a fake syncache entry, and return.
1012 if (!V_tcp_syncookies) {
1014 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1015 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1016 "segment rejected (syncookies disabled)\n",
1020 if (!V_tcp_syncookiesonly &&
1021 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1023 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1024 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1025 "segment rejected (no syncache entry)\n",
1029 bzero(&scs, sizeof(scs));
1030 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
1033 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1034 log(LOG_DEBUG, "%s; %s: Segment failed "
1035 "SYNCOOKIE authentication, segment rejected "
1036 "(probably spoofed)\n", s, __func__);
1039 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1040 /* If received ACK has MD5 signature, check it. */
1041 if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1042 (!TCPMD5_ENABLED() ||
1043 TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1045 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1046 log(LOG_DEBUG, "%s; %s: Segment rejected, "
1047 "MD5 signature doesn't match.\n",
1051 TCPSTAT_INC(tcps_sig_err_sigopt);
1052 return (-1); /* Do not send RST */
1054 #endif /* TCP_SIGNATURE */
1056 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1058 * If listening socket requested TCP digests, check that
1059 * received ACK has signature and it is correct.
1060 * If not, drop the ACK and leave sc entry in th cache,
1061 * because SYN was received with correct signature.
1063 if (sc->sc_flags & SCF_SIGNATURE) {
1064 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1066 TCPSTAT_INC(tcps_sig_err_nosigopt);
1068 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1069 log(LOG_DEBUG, "%s; %s: Segment "
1070 "rejected, MD5 signature wasn't "
1071 "provided.\n", s, __func__);
1074 return (-1); /* Do not send RST */
1076 if (!TCPMD5_ENABLED() ||
1077 TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1078 /* Doesn't match or no SA */
1080 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1081 log(LOG_DEBUG, "%s; %s: Segment "
1082 "rejected, MD5 signature doesn't "
1083 "match.\n", s, __func__);
1086 return (-1); /* Do not send RST */
1089 #endif /* TCP_SIGNATURE */
1091 * Pull out the entry to unlock the bucket row.
1093 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1094 * tcp_state_change(). The tcpcb is not existent at this
1095 * moment. A new one will be allocated via syncache_socket->
1096 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1097 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1099 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1100 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1103 if (ADDED_BY_TOE(sc)) {
1104 struct toedev *tod = sc->sc_tod;
1106 tod->tod_syncache_removed(tod, sc->sc_todctx);
1113 * Segment validation:
1114 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1116 if (th->th_ack != sc->sc_iss + 1) {
1117 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1118 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1119 "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1124 * The SEQ must fall in the window starting at the received
1125 * initial receive sequence number + 1 (the SYN).
1127 if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1128 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1129 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1130 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1131 "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1136 * If timestamps were not negotiated during SYN/ACK they
1137 * must not appear on any segment during this session.
1139 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1140 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1141 log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1142 "segment rejected\n", s, __func__);
1147 * If timestamps were negotiated during SYN/ACK they should
1148 * appear on every segment during this session.
1149 * XXXAO: This is only informal as there have been unverified
1150 * reports of non-compliants stacks.
1152 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1153 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1154 log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1155 "no action\n", s, __func__);
1162 * If timestamps were negotiated, the reflected timestamp
1163 * must be equal to what we actually sent in the SYN|ACK
1164 * except in the case of 0. Some boxes are known for sending
1165 * broken timestamp replies during the 3whs (and potentially
1166 * during the connection also).
1168 * Accept the final ACK of 3whs with reflected timestamp of 0
1169 * instead of sending a RST and deleting the syncache entry.
1171 if ((to->to_flags & TOF_TS) && to->to_tsecr &&
1172 to->to_tsecr != sc->sc_ts) {
1173 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1174 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, "
1175 "segment rejected\n",
1176 s, __func__, to->to_tsecr, sc->sc_ts);
1180 *lsop = syncache_socket(sc, *lsop, m);
1183 TCPSTAT_INC(tcps_sc_aborted);
1185 TCPSTAT_INC(tcps_sc_completed);
1187 /* how do we find the inp for the new socket? */
1192 if (sc != NULL && sc != &scs)
1202 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
1203 uint64_t response_cookie)
1207 unsigned int *pending_counter;
1210 * Global TCP locks are held because we manipulate the PCB lists
1211 * and create a new socket.
1213 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1215 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
1216 *lsop = syncache_socket(sc, *lsop, m);
1217 if (*lsop == NULL) {
1218 TCPSTAT_INC(tcps_sc_aborted);
1219 atomic_subtract_int(pending_counter, 1);
1221 inp = sotoinpcb(*lsop);
1222 tp = intotcpcb(inp);
1223 tp->t_flags |= TF_FASTOPEN;
1224 tp->t_tfo_cookie = response_cookie;
1225 tp->snd_max = tp->iss;
1226 tp->snd_nxt = tp->iss;
1227 tp->t_tfo_pending = pending_counter;
1228 TCPSTAT_INC(tcps_sc_completed);
1231 #endif /* TCP_RFC7413 */
1234 * Given a LISTEN socket and an inbound SYN request, add
1235 * this to the syn cache, and send back a segment:
1236 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1239 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1240 * Doing so would require that we hold onto the data and deliver it
1241 * to the application. However, if we are the target of a SYN-flood
1242 * DoS attack, an attacker could send data which would eventually
1243 * consume all available buffer space if it were ACKed. By not ACKing
1244 * the data, we avoid this DoS scenario.
1246 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1247 * cookie is processed, V_tcp_fastopen_enabled set to true, and the
1248 * TCP_FASTOPEN socket option is set. In this case, a new socket is created
1249 * and returned via lsop, the mbuf is not freed so that tcp_input() can
1250 * queue its data to the socket, and 1 is returned to indicate the
1251 * TFO-socket-creation path was taken.
1254 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1255 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1260 struct syncache *sc = NULL;
1261 struct syncache_head *sch;
1262 struct mbuf *ipopts = NULL;
1264 int win, sb_hiwat, ip_ttl, ip_tos;
1268 int autoflowlabel = 0;
1271 struct label *maclabel;
1273 struct syncache scs;
1276 uint64_t tfo_response_cookie;
1277 int tfo_cookie_valid = 0;
1278 int tfo_response_cookie_valid = 0;
1281 INP_WLOCK_ASSERT(inp); /* listen socket */
1282 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1283 ("%s: unexpected tcp flags", __func__));
1286 * Combine all so/tp operations very early to drop the INP lock as
1291 cred = crhold(so->so_cred);
1294 if ((inc->inc_flags & INC_ISIPV6) &&
1295 (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1298 ip_ttl = inp->inp_ip_ttl;
1299 ip_tos = inp->inp_ip_tos;
1300 win = sbspace(&so->so_rcv);
1301 sb_hiwat = so->so_rcv.sb_hiwat;
1302 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1305 if (V_tcp_fastopen_enabled && (tp->t_flags & TF_FASTOPEN) &&
1306 (tp->t_tfo_pending != NULL) && (to->to_flags & TOF_FASTOPEN)) {
1308 * Limit the number of pending TFO connections to
1309 * approximately half of the queue limit. This prevents TFO
1310 * SYN floods from starving the service by filling the
1311 * listen queue with bogus TFO connections.
1313 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1314 (so->so_qlimit / 2)) {
1317 result = tcp_fastopen_check_cookie(inc,
1318 to->to_tfo_cookie, to->to_tfo_len,
1319 &tfo_response_cookie);
1320 tfo_cookie_valid = (result > 0);
1321 tfo_response_cookie_valid = (result >= 0);
1323 atomic_subtract_int(tp->t_tfo_pending, 1);
1327 /* By the time we drop the lock these should no longer be used. */
1332 if (mac_syncache_init(&maclabel) != 0) {
1336 mac_syncache_create(maclabel, inp);
1339 if (!tfo_cookie_valid)
1344 * Remember the IP options, if any.
1347 if (!(inc->inc_flags & INC_ISIPV6))
1350 ipopts = (m) ? ip_srcroute(m) : NULL;
1355 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1357 * If listening socket requested TCP digests, check that received
1358 * SYN has signature and it is correct. If signature doesn't match
1359 * or TCP_SIGNATURE support isn't enabled, drop the packet.
1361 if (ltflags & TF_SIGNATURE) {
1362 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1363 TCPSTAT_INC(tcps_sig_err_nosigopt);
1366 if (!TCPMD5_ENABLED() ||
1367 TCPMD5_INPUT(m, th, to->to_signature) != 0)
1370 #endif /* TCP_SIGNATURE */
1372 * See if we already have an entry for this connection.
1373 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1375 * XXX: should the syncache be re-initialized with the contents
1376 * of the new SYN here (which may have different options?)
1378 * XXX: We do not check the sequence number to see if this is a
1379 * real retransmit or a new connection attempt. The question is
1380 * how to handle such a case; either ignore it as spoofed, or
1381 * drop the current entry and create a new one?
1383 sc = syncache_lookup(inc, &sch); /* returns locked entry */
1384 SCH_LOCK_ASSERT(sch);
1387 if (tfo_cookie_valid)
1390 TCPSTAT_INC(tcps_sc_dupsyn);
1393 * If we were remembering a previous source route,
1394 * forget it and use the new one we've been given.
1397 (void) m_free(sc->sc_ipopts);
1398 sc->sc_ipopts = ipopts;
1401 * Update timestamp if present.
1403 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1404 sc->sc_tsreflect = to->to_tsval;
1406 sc->sc_flags &= ~SCF_TIMESTAMP;
1409 * Since we have already unconditionally allocated label
1410 * storage, free it up. The syncache entry will already
1411 * have an initialized label we can use.
1413 mac_syncache_destroy(&maclabel);
1415 /* Retransmit SYN|ACK and reset retransmit count. */
1416 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1417 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1418 "resetting timer and retransmitting SYN|ACK\n",
1422 if (syncache_respond(sc, sch, 1, m) == 0) {
1424 syncache_timeout(sc, sch, 1);
1425 TCPSTAT_INC(tcps_sndacks);
1426 TCPSTAT_INC(tcps_sndtotal);
1433 if (tfo_cookie_valid) {
1434 bzero(&scs, sizeof(scs));
1440 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1443 * The zone allocator couldn't provide more entries.
1444 * Treat this as if the cache was full; drop the oldest
1445 * entry and insert the new one.
1447 TCPSTAT_INC(tcps_sc_zonefail);
1448 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1449 sch->sch_last_overflow = time_uptime;
1450 syncache_drop(sc, sch);
1452 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1454 if (V_tcp_syncookies) {
1455 bzero(&scs, sizeof(scs));
1460 (void) m_free(ipopts);
1468 if (!tfo_cookie_valid && tfo_response_cookie_valid)
1469 sc->sc_tfo_cookie = &tfo_response_cookie;
1473 * Fill in the syncache values.
1476 sc->sc_label = maclabel;
1480 sc->sc_ipopts = ipopts;
1481 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1483 if (!(inc->inc_flags & INC_ISIPV6))
1486 sc->sc_ip_tos = ip_tos;
1487 sc->sc_ip_ttl = ip_ttl;
1491 sc->sc_todctx = todctx;
1493 sc->sc_irs = th->th_seq;
1494 sc->sc_iss = arc4random();
1496 sc->sc_flowlabel = 0;
1499 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1500 * win was derived from socket earlier in the function.
1503 win = imin(win, TCP_MAXWIN);
1506 if (V_tcp_do_rfc1323) {
1508 * A timestamp received in a SYN makes
1509 * it ok to send timestamp requests and replies.
1511 if (to->to_flags & TOF_TS) {
1512 sc->sc_tsreflect = to->to_tsval;
1513 sc->sc_ts = tcp_ts_getticks();
1514 sc->sc_flags |= SCF_TIMESTAMP;
1516 if (to->to_flags & TOF_SCALE) {
1520 * Pick the smallest possible scaling factor that
1521 * will still allow us to scale up to sb_max, aka
1522 * kern.ipc.maxsockbuf.
1524 * We do this because there are broken firewalls that
1525 * will corrupt the window scale option, leading to
1526 * the other endpoint believing that our advertised
1527 * window is unscaled. At scale factors larger than
1528 * 5 the unscaled window will drop below 1500 bytes,
1529 * leading to serious problems when traversing these
1532 * With the default maxsockbuf of 256K, a scale factor
1533 * of 3 will be chosen by this algorithm. Those who
1534 * choose a larger maxsockbuf should watch out
1535 * for the compatibility problems mentioned above.
1537 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1538 * or <SYN,ACK>) segment itself is never scaled.
1540 while (wscale < TCP_MAX_WINSHIFT &&
1541 (TCP_MAXWIN << wscale) < sb_max)
1543 sc->sc_requested_r_scale = wscale;
1544 sc->sc_requested_s_scale = to->to_wscale;
1545 sc->sc_flags |= SCF_WINSCALE;
1548 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1550 * If listening socket requested TCP digests, flag this in the
1551 * syncache so that syncache_respond() will do the right thing
1554 if (ltflags & TF_SIGNATURE)
1555 sc->sc_flags |= SCF_SIGNATURE;
1556 #endif /* TCP_SIGNATURE */
1557 if (to->to_flags & TOF_SACKPERM)
1558 sc->sc_flags |= SCF_SACK;
1559 if (to->to_flags & TOF_MSS)
1560 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */
1561 if (ltflags & TF_NOOPT)
1562 sc->sc_flags |= SCF_NOOPT;
1563 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1564 sc->sc_flags |= SCF_ECN;
1566 if (V_tcp_syncookies)
1567 sc->sc_iss = syncookie_generate(sch, sc);
1569 if (autoflowlabel) {
1570 if (V_tcp_syncookies)
1571 sc->sc_flowlabel = sc->sc_iss;
1573 sc->sc_flowlabel = ip6_randomflowlabel();
1574 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1580 if (tfo_cookie_valid) {
1581 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie);
1582 /* INP_WUNLOCK(inp) will be performed by the called */
1589 * Do a standard 3-way handshake.
1591 if (syncache_respond(sc, sch, 0, m) == 0) {
1592 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1594 else if (sc != &scs)
1595 syncache_insert(sc, sch); /* locks and unlocks sch */
1596 TCPSTAT_INC(tcps_sndacks);
1597 TCPSTAT_INC(tcps_sndtotal);
1601 TCPSTAT_INC(tcps_sc_dropped);
1616 mac_syncache_destroy(&maclabel);
1622 * Send SYN|ACK to the peer. Either in response to the peer's SYN,
1623 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1626 syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked,
1627 const struct mbuf *m0)
1629 struct ip *ip = NULL;
1631 struct tcphdr *th = NULL;
1632 int optlen, error = 0; /* Make compiler happy */
1633 u_int16_t hlen, tlen, mssopt;
1636 struct ip6_hdr *ip6 = NULL;
1640 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1643 tlen = hlen + sizeof(struct tcphdr);
1645 /* Determine MSS we advertize to other end of connection. */
1646 mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss);
1648 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1649 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1650 ("syncache: mbuf too small"));
1652 /* Create the IP+TCP header from scratch. */
1653 m = m_gethdr(M_NOWAIT, MT_DATA);
1657 mac_syncache_create_mbuf(sc->sc_label, m);
1659 m->m_data += max_linkhdr;
1661 m->m_pkthdr.len = tlen;
1662 m->m_pkthdr.rcvif = NULL;
1665 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1666 ip6 = mtod(m, struct ip6_hdr *);
1667 ip6->ip6_vfc = IPV6_VERSION;
1668 ip6->ip6_nxt = IPPROTO_TCP;
1669 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1670 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1671 ip6->ip6_plen = htons(tlen - hlen);
1672 /* ip6_hlim is set after checksum */
1673 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1674 ip6->ip6_flow |= sc->sc_flowlabel;
1676 th = (struct tcphdr *)(ip6 + 1);
1679 #if defined(INET6) && defined(INET)
1684 ip = mtod(m, struct ip *);
1685 ip->ip_v = IPVERSION;
1686 ip->ip_hl = sizeof(struct ip) >> 2;
1687 ip->ip_len = htons(tlen);
1691 ip->ip_p = IPPROTO_TCP;
1692 ip->ip_src = sc->sc_inc.inc_laddr;
1693 ip->ip_dst = sc->sc_inc.inc_faddr;
1694 ip->ip_ttl = sc->sc_ip_ttl;
1695 ip->ip_tos = sc->sc_ip_tos;
1698 * See if we should do MTU discovery. Route lookups are
1699 * expensive, so we will only unset the DF bit if:
1701 * 1) path_mtu_discovery is disabled
1702 * 2) the SCF_UNREACH flag has been set
1704 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1705 ip->ip_off |= htons(IP_DF);
1707 th = (struct tcphdr *)(ip + 1);
1710 th->th_sport = sc->sc_inc.inc_lport;
1711 th->th_dport = sc->sc_inc.inc_fport;
1713 th->th_seq = htonl(sc->sc_iss);
1714 th->th_ack = htonl(sc->sc_irs + 1);
1715 th->th_off = sizeof(struct tcphdr) >> 2;
1717 th->th_flags = TH_SYN|TH_ACK;
1718 th->th_win = htons(sc->sc_wnd);
1721 if (sc->sc_flags & SCF_ECN) {
1722 th->th_flags |= TH_ECE;
1723 TCPSTAT_INC(tcps_ecn_shs);
1726 /* Tack on the TCP options. */
1727 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1731 to.to_flags = TOF_MSS;
1732 if (sc->sc_flags & SCF_WINSCALE) {
1733 to.to_wscale = sc->sc_requested_r_scale;
1734 to.to_flags |= TOF_SCALE;
1736 if (sc->sc_flags & SCF_TIMESTAMP) {
1737 /* Virgin timestamp or TCP cookie enhanced one. */
1738 to.to_tsval = sc->sc_ts;
1739 to.to_tsecr = sc->sc_tsreflect;
1740 to.to_flags |= TOF_TS;
1742 if (sc->sc_flags & SCF_SACK)
1743 to.to_flags |= TOF_SACKPERM;
1744 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1745 if (sc->sc_flags & SCF_SIGNATURE)
1746 to.to_flags |= TOF_SIGNATURE;
1749 if (sc->sc_tfo_cookie) {
1750 to.to_flags |= TOF_FASTOPEN;
1751 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1752 to.to_tfo_cookie = sc->sc_tfo_cookie;
1753 /* don't send cookie again when retransmitting response */
1754 sc->sc_tfo_cookie = NULL;
1757 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1759 /* Adjust headers by option size. */
1760 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1762 m->m_pkthdr.len += optlen;
1764 if (sc->sc_inc.inc_flags & INC_ISIPV6)
1765 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1768 ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1769 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1770 if (sc->sc_flags & SCF_SIGNATURE) {
1771 KASSERT(to.to_flags & TOF_SIGNATURE,
1772 ("tcp_addoptions() didn't set tcp_signature"));
1774 /* NOTE: to.to_signature is inside of mbuf */
1775 if (!TCPMD5_ENABLED() ||
1776 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1785 M_SETFIB(m, sc->sc_inc.inc_fibnum);
1786 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1788 * If we have peer's SYN and it has a flowid, then let's assign it to
1789 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid
1790 * to SYN|ACK due to lack of inp here.
1792 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1793 m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1794 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1797 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1798 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1799 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1801 ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1803 if (ADDED_BY_TOE(sc)) {
1804 struct toedev *tod = sc->sc_tod;
1806 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1811 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1814 #if defined(INET6) && defined(INET)
1819 m->m_pkthdr.csum_flags = CSUM_TCP;
1820 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1821 htons(tlen + optlen - hlen + IPPROTO_TCP));
1823 if (ADDED_BY_TOE(sc)) {
1824 struct toedev *tod = sc->sc_tod;
1826 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1831 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1838 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1839 * that exceed the capacity of the syncache by avoiding the storage of any
1840 * of the SYNs we receive. Syncookies defend against blind SYN flooding
1841 * attacks where the attacker does not have access to our responses.
1843 * Syncookies encode and include all necessary information about the
1844 * connection setup within the SYN|ACK that we send back. That way we
1845 * can avoid keeping any local state until the ACK to our SYN|ACK returns
1846 * (if ever). Normally the syncache and syncookies are running in parallel
1847 * with the latter taking over when the former is exhausted. When matching
1848 * syncache entry is found the syncookie is ignored.
1850 * The only reliable information persisting the 3WHS is our initial sequence
1851 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient
1852 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1853 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK
1854 * returns and signifies a legitimate connection if it matches the ACK.
1856 * The available space of 32 bits to store the hash and to encode the SYN
1857 * option information is very tight and we should have at least 24 bits for
1858 * the MAC to keep the number of guesses by blind spoofing reasonably high.
1860 * SYN option information we have to encode to fully restore a connection:
1861 * MSS: is imporant to chose an optimal segment size to avoid IP level
1862 * fragmentation along the path. The common MSS values can be encoded
1863 * in a 3-bit table. Uncommon values are captured by the next lower value
1864 * in the table leading to a slight increase in packetization overhead.
1865 * WSCALE: is necessary to allow large windows to be used for high delay-
1866 * bandwidth product links. Not scaling the window when it was initially
1867 * negotiated is bad for performance as lack of scaling further decreases
1868 * the apparent available send window. We only need to encode the WSCALE
1869 * we received from the remote end. Our end can be recalculated at any
1870 * time. The common WSCALE values can be encoded in a 3-bit table.
1871 * Uncommon values are captured by the next lower value in the table
1872 * making us under-estimate the available window size halving our
1873 * theoretically possible maximum throughput for that connection.
1874 * SACK: Greatly assists in packet loss recovery and requires 1 bit.
1875 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
1876 * that are included in all segments on a connection. We enable them when
1879 * Security of syncookies and attack vectors:
1881 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
1882 * together with the gloabl secret to make it unique per connection attempt.
1883 * Thus any change of any of those parameters results in a different MAC output
1884 * in an unpredictable way unless a collision is encountered. 24 bits of the
1885 * MAC are embedded into the ISS.
1887 * To prevent replay attacks two rotating global secrets are updated with a
1888 * new random value every 15 seconds. The life-time of a syncookie is thus
1891 * Vector 1: Attacking the secret. This requires finding a weakness in the
1892 * MAC itself or the way it is used here. The attacker can do a chosen plain
1893 * text attack by varying and testing the all parameters under his control.
1894 * The strength depends on the size and randomness of the secret, and the
1895 * cryptographic security of the MAC function. Due to the constant updating
1896 * of the secret the attacker has at most 29.999 seconds to find the secret
1897 * and launch spoofed connections. After that he has to start all over again.
1899 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC
1900 * size an average of 4,823 attempts are required for a 50% chance of success
1901 * to spoof a single syncookie (birthday collision paradox). However the
1902 * attacker is blind and doesn't know if one of his attempts succeeded unless
1903 * he has a side channel to interfere success from. A single connection setup
1904 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
1905 * This many attempts are required for each one blind spoofed connection. For
1906 * every additional spoofed connection he has to launch another N attempts.
1907 * Thus for a sustained rate 100 spoofed connections per second approximately
1908 * 1,800,000 packets per second would have to be sent.
1910 * NB: The MAC function should be fast so that it doesn't become a CPU
1911 * exhaustion attack vector itself.
1914 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations
1915 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
1916 * http://cr.yp.to/syncookies.html (overview)
1917 * http://cr.yp.to/syncookies/archive (details)
1920 * Schematic construction of a syncookie enabled Initial Sequence Number:
1922 * 12345678901234567890123456789012
1923 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
1925 * x 24 MAC (truncated)
1926 * W 3 Send Window Scale index
1928 * S 1 SACK permitted
1929 * P 1 Odd/even secret
1933 * Distribution and probability of certain MSS values. Those in between are
1934 * rounded down to the next lower one.
1935 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
1936 * .2% .3% 5% 7% 7% 20% 15% 45%
1938 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
1941 * Distribution and probability of certain WSCALE values. We have to map the
1942 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
1943 * bits based on prevalence of certain values. Where we don't have an exact
1944 * match for are rounded down to the next lower one letting us under-estimate
1945 * the true available window. At the moment this would happen only for the
1946 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
1947 * and window size). The absence of the WSCALE option (no scaling in either
1948 * direction) is encoded with index zero.
1949 * [WSCALE values histograms, Allman, 2012]
1950 * X 10 10 35 5 6 14 10% by host
1951 * X 11 4 5 5 18 49 3% by connections
1953 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
1956 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed
1957 * and good cryptographic properties.
1960 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
1961 uint8_t *secbits, uintptr_t secmod)
1964 uint32_t siphash[2];
1966 SipHash24_Init(&ctx);
1967 SipHash_SetKey(&ctx, secbits);
1968 switch (inc->inc_flags & INC_ISIPV6) {
1971 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
1972 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
1977 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
1978 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
1982 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
1983 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
1984 SipHash_Update(&ctx, &irs, sizeof(irs));
1985 SipHash_Update(&ctx, &flags, sizeof(flags));
1986 SipHash_Update(&ctx, &secmod, sizeof(secmod));
1987 SipHash_Final((u_int8_t *)&siphash, &ctx);
1989 return (siphash[0] ^ siphash[1]);
1993 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
1995 u_int i, secbit, wscale;
1998 union syncookie cookie;
2000 SCH_LOCK_ASSERT(sch);
2004 /* Map our computed MSS into the 3-bit index. */
2005 for (i = nitems(tcp_sc_msstab) - 1;
2006 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2009 cookie.flags.mss_idx = i;
2012 * Map the send window scale into the 3-bit index but only if
2013 * the wscale option was received.
2015 if (sc->sc_flags & SCF_WINSCALE) {
2016 wscale = sc->sc_requested_s_scale;
2017 for (i = nitems(tcp_sc_wstab) - 1;
2018 tcp_sc_wstab[i] > wscale && i > 0;
2021 cookie.flags.wscale_idx = i;
2024 /* Can we do SACK? */
2025 if (sc->sc_flags & SCF_SACK)
2026 cookie.flags.sack_ok = 1;
2028 /* Which of the two secrets to use. */
2029 secbit = sch->sch_sc->secret.oddeven & 0x1;
2030 cookie.flags.odd_even = secbit;
2032 secbits = sch->sch_sc->secret.key[secbit];
2033 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2037 * Put the flags into the hash and XOR them to get better ISS number
2038 * variance. This doesn't enhance the cryptographic strength and is
2039 * done to prevent the 8 cookie bits from showing up directly on the
2043 iss |= cookie.cookie ^ (hash >> 24);
2045 /* Randomize the timestamp. */
2046 if (sc->sc_flags & SCF_TIMESTAMP) {
2047 sc->sc_ts = arc4random();
2048 sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks();
2051 TCPSTAT_INC(tcps_sc_sendcookie);
2055 static struct syncache *
2056 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2057 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2063 int wnd, wscale = 0;
2064 union syncookie cookie;
2066 SCH_LOCK_ASSERT(sch);
2069 * Pull information out of SYN-ACK/ACK and revert sequence number
2072 ack = th->th_ack - 1;
2073 seq = th->th_seq - 1;
2076 * Unpack the flags containing enough information to restore the
2079 cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2081 /* Which of the two secrets to use. */
2082 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
2084 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2086 /* The recomputed hash matches the ACK if this was a genuine cookie. */
2087 if ((ack & ~0xff) != (hash & ~0xff))
2090 /* Fill in the syncache values. */
2092 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2093 sc->sc_ipopts = NULL;
2098 switch (inc->inc_flags & INC_ISIPV6) {
2101 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2102 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2107 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2108 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
2113 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2115 /* We can simply recompute receive window scale we sent earlier. */
2116 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2119 /* Only use wscale if it was enabled in the orignal SYN. */
2120 if (cookie.flags.wscale_idx > 0) {
2121 sc->sc_requested_r_scale = wscale;
2122 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2123 sc->sc_flags |= SCF_WINSCALE;
2126 wnd = sbspace(&lso->so_rcv);
2128 wnd = imin(wnd, TCP_MAXWIN);
2131 if (cookie.flags.sack_ok)
2132 sc->sc_flags |= SCF_SACK;
2134 if (to->to_flags & TOF_TS) {
2135 sc->sc_flags |= SCF_TIMESTAMP;
2136 sc->sc_tsreflect = to->to_tsval;
2137 sc->sc_ts = to->to_tsecr;
2138 sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks();
2141 if (to->to_flags & TOF_SIGNATURE)
2142 sc->sc_flags |= SCF_SIGNATURE;
2146 TCPSTAT_INC(tcps_sc_recvcookie);
2152 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2153 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2156 struct syncache scs, *scx;
2159 bzero(&scs, sizeof(scs));
2160 scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
2162 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2166 if (sc->sc_peer_mss != scx->sc_peer_mss)
2167 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2168 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2170 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2171 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2172 s, __func__, sc->sc_requested_r_scale,
2173 scx->sc_requested_r_scale);
2175 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2176 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2177 s, __func__, sc->sc_requested_s_scale,
2178 scx->sc_requested_s_scale);
2180 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2181 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2188 #endif /* INVARIANTS */
2191 syncookie_reseed(void *arg)
2193 struct tcp_syncache *sc = arg;
2198 * Reseeding the secret doesn't have to be protected by a lock.
2199 * It only must be ensured that the new random values are visible
2200 * to all CPUs in a SMP environment. The atomic with release
2201 * semantics ensures that.
2203 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2204 secbits = sc->secret.key[secbit];
2205 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2206 atomic_add_rel_int(&sc->secret.oddeven, 1);
2208 /* Reschedule ourself. */
2209 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2213 * Exports the syncache entries to userland so that netstat can display
2214 * them alongside the other sockets. This function is intended to be
2215 * called only from tcp_pcblist.
2217 * Due to concurrency on an active system, the number of pcbs exported
2218 * may have no relation to max_pcbs. max_pcbs merely indicates the
2219 * amount of space the caller allocated for this function to use.
2222 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
2225 struct syncache *sc;
2226 struct syncache_head *sch;
2227 int count, error, i;
2229 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
2230 sch = &V_tcp_syncache.hashbase[i];
2232 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2233 if (count >= max_pcbs) {
2237 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2239 bzero(&xt, sizeof(xt));
2240 xt.xt_len = sizeof(xt);
2241 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2242 xt.xt_inp.inp_vflag = INP_IPV6;
2244 xt.xt_inp.inp_vflag = INP_IPV4;
2245 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo));
2246 xt.xt_tp.t_inpcb = &xt.xt_inp;
2247 xt.xt_tp.t_state = TCPS_SYN_RECEIVED;
2248 xt.xt_socket.xso_protocol = IPPROTO_TCP;
2249 xt.xt_socket.xso_len = sizeof (struct xsocket);
2250 xt.xt_socket.so_type = SOCK_STREAM;
2251 xt.xt_socket.so_state = SS_ISCONNECTING;
2252 error = SYSCTL_OUT(req, &xt, sizeof xt);
2262 *pcbs_exported = count;