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_kdtrace.h>
71 #include <netinet/in_systm.h>
72 #include <netinet/ip.h>
73 #include <netinet/in_var.h>
74 #include <netinet/in_pcb.h>
75 #include <netinet/ip_var.h>
76 #include <netinet/ip_options.h>
78 #include <netinet/ip6.h>
79 #include <netinet/icmp6.h>
80 #include <netinet6/nd6.h>
81 #include <netinet6/ip6_var.h>
82 #include <netinet6/in6_pcb.h>
84 #include <netinet/tcp.h>
86 #include <netinet/tcp_fastopen.h>
88 #include <netinet/tcp_fsm.h>
89 #include <netinet/tcp_seq.h>
90 #include <netinet/tcp_timer.h>
91 #include <netinet/tcp_var.h>
92 #include <netinet/tcp_syncache.h>
94 #include <netinet6/tcp6_var.h>
97 #include <netinet/toecore.h>
100 #include <netipsec/ipsec_support.h>
102 #include <machine/in_cksum.h>
104 #include <security/mac/mac_framework.h>
106 static VNET_DEFINE(int, tcp_syncookies) = 1;
107 #define V_tcp_syncookies VNET(tcp_syncookies)
108 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
109 &VNET_NAME(tcp_syncookies), 0,
110 "Use TCP SYN cookies if the syncache overflows");
112 static VNET_DEFINE(int, tcp_syncookiesonly) = 0;
113 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly)
114 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
115 &VNET_NAME(tcp_syncookiesonly), 0,
116 "Use only TCP SYN cookies");
119 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
122 static void syncache_drop(struct syncache *, struct syncache_head *);
123 static void syncache_free(struct syncache *);
124 static void syncache_insert(struct syncache *, struct syncache_head *);
125 static int syncache_respond(struct syncache *, struct syncache_head *, int,
126 const struct mbuf *);
127 static struct socket *syncache_socket(struct syncache *, struct socket *,
129 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch,
131 static void syncache_timer(void *);
133 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
134 uint8_t *, uintptr_t);
135 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *);
136 static struct syncache
137 *syncookie_lookup(struct in_conninfo *, struct syncache_head *,
138 struct syncache *, struct tcphdr *, struct tcpopt *,
140 static void syncookie_reseed(void *);
142 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
143 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
148 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
149 * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds,
150 * the odds are that the user has given up attempting to connect by then.
152 #define SYNCACHE_MAXREXMTS 3
154 /* Arbitrary values */
155 #define TCP_SYNCACHE_HASHSIZE 512
156 #define TCP_SYNCACHE_BUCKETLIMIT 30
158 static VNET_DEFINE(struct tcp_syncache, tcp_syncache);
159 #define V_tcp_syncache VNET(tcp_syncache)
161 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0,
164 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
165 &VNET_NAME(tcp_syncache.bucket_limit), 0,
166 "Per-bucket hash limit for syncache");
168 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
169 &VNET_NAME(tcp_syncache.cache_limit), 0,
170 "Overall entry limit for syncache");
172 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
173 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
175 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
176 &VNET_NAME(tcp_syncache.hashsize), 0,
177 "Size of TCP syncache hashtable");
180 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
185 new = V_tcp_syncache.rexmt_limit;
186 error = sysctl_handle_int(oidp, &new, 0, req);
187 if ((error == 0) && (req->newptr != NULL)) {
188 if (new > TCP_MAXRXTSHIFT)
191 V_tcp_syncache.rexmt_limit = new;
196 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
197 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW,
198 &VNET_NAME(tcp_syncache.rexmt_limit), 0,
199 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
200 "Limit on SYN/ACK retransmissions");
202 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
203 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
204 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
205 "Send reset on socket allocation failure");
207 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
209 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx)
210 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx)
211 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED)
214 * Requires the syncache entry to be already removed from the bucket list.
217 syncache_free(struct syncache *sc)
221 (void) m_free(sc->sc_ipopts);
225 mac_syncache_destroy(&sc->sc_label);
228 uma_zfree(V_tcp_syncache.zone, sc);
236 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
237 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
238 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
239 V_tcp_syncache.hash_secret = arc4random();
241 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
242 &V_tcp_syncache.hashsize);
243 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
244 &V_tcp_syncache.bucket_limit);
245 if (!powerof2(V_tcp_syncache.hashsize) ||
246 V_tcp_syncache.hashsize == 0) {
247 printf("WARNING: syncache hash size is not a power of 2.\n");
248 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
250 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
253 V_tcp_syncache.cache_limit =
254 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
255 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
256 &V_tcp_syncache.cache_limit);
258 /* Allocate the hash table. */
259 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
260 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
263 V_tcp_syncache.vnet = curvnet;
266 /* Initialize the hash buckets. */
267 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
268 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
269 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
271 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
272 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
273 V_tcp_syncache.hashbase[i].sch_length = 0;
274 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
275 V_tcp_syncache.hashbase[i].sch_last_overflow =
276 -(SYNCOOKIE_LIFETIME + 1);
279 /* Create the syncache entry zone. */
280 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
281 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
282 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
283 V_tcp_syncache.cache_limit);
285 /* Start the SYN cookie reseeder callout. */
286 callout_init(&V_tcp_syncache.secret.reseed, 1);
287 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
288 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
289 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
290 syncookie_reseed, &V_tcp_syncache);
295 syncache_destroy(void)
297 struct syncache_head *sch;
298 struct syncache *sc, *nsc;
302 * Stop the re-seed timer before freeing resources. No need to
303 * possibly schedule it another time.
305 callout_drain(&V_tcp_syncache.secret.reseed);
307 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */
308 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
310 sch = &V_tcp_syncache.hashbase[i];
311 callout_drain(&sch->sch_timer);
314 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
315 syncache_drop(sc, sch);
317 KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
318 ("%s: sch->sch_bucket not empty", __func__));
319 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
320 __func__, sch->sch_length));
321 mtx_destroy(&sch->sch_mtx);
324 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
325 ("%s: cache_count not 0", __func__));
327 /* Free the allocated global resources. */
328 uma_zdestroy(V_tcp_syncache.zone);
329 free(V_tcp_syncache.hashbase, M_SYNCACHE);
334 * Inserts a syncache entry into the specified bucket row.
335 * Locks and unlocks the syncache_head autonomously.
338 syncache_insert(struct syncache *sc, struct syncache_head *sch)
340 struct syncache *sc2;
345 * Make sure that we don't overflow the per-bucket limit.
346 * If the bucket is full, toss the oldest element.
348 if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
349 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
350 ("sch->sch_length incorrect"));
351 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
352 sch->sch_last_overflow = time_uptime;
353 syncache_drop(sc2, sch);
354 TCPSTAT_INC(tcps_sc_bucketoverflow);
357 /* Put it into the bucket. */
358 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
362 if (ADDED_BY_TOE(sc)) {
363 struct toedev *tod = sc->sc_tod;
365 tod->tod_syncache_added(tod, sc->sc_todctx);
369 /* Reinitialize the bucket row's timer. */
370 if (sch->sch_length == 1)
371 sch->sch_nextc = ticks + INT_MAX;
372 syncache_timeout(sc, sch, 1);
376 TCPSTATES_INC(TCPS_SYN_RECEIVED);
377 TCPSTAT_INC(tcps_sc_added);
381 * Remove and free entry from syncache bucket row.
382 * Expects locked syncache head.
385 syncache_drop(struct syncache *sc, struct syncache_head *sch)
388 SCH_LOCK_ASSERT(sch);
390 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
391 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
395 if (ADDED_BY_TOE(sc)) {
396 struct toedev *tod = sc->sc_tod;
398 tod->tod_syncache_removed(tod, sc->sc_todctx);
406 * Engage/reengage time on bucket row.
409 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
413 if (sc->sc_rxmits == 0)
414 rexmt = TCPTV_RTOBASE;
416 TCPT_RANGESET(rexmt, TCPTV_RTOBASE * tcp_syn_backoff[sc->sc_rxmits],
417 tcp_rexmit_min, TCPTV_REXMTMAX);
418 sc->sc_rxttime = ticks + rexmt;
420 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
421 sch->sch_nextc = sc->sc_rxttime;
423 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
424 syncache_timer, (void *)sch);
429 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
430 * If we have retransmitted an entry the maximum number of times, expire it.
431 * One separate timer for each bucket row.
434 syncache_timer(void *xsch)
436 struct syncache_head *sch = (struct syncache_head *)xsch;
437 struct syncache *sc, *nsc;
441 CURVNET_SET(sch->sch_sc->vnet);
443 /* NB: syncache_head has already been locked by the callout. */
444 SCH_LOCK_ASSERT(sch);
447 * In the following cycle we may remove some entries and/or
448 * advance some timeouts, so re-initialize the bucket timer.
450 sch->sch_nextc = tick + INT_MAX;
452 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
454 * We do not check if the listen socket still exists
455 * and accept the case where the listen socket may be
456 * gone by the time we resend the SYN/ACK. We do
457 * not expect this to happens often. If it does,
458 * then the RST will be sent by the time the remote
459 * host does the SYN/ACK->ACK.
461 if (TSTMP_GT(sc->sc_rxttime, tick)) {
462 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
463 sch->sch_nextc = sc->sc_rxttime;
466 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
467 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
468 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
469 "giving up and removing syncache entry\n",
473 syncache_drop(sc, sch);
474 TCPSTAT_INC(tcps_sc_stale);
477 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
478 log(LOG_DEBUG, "%s; %s: Response timeout, "
479 "retransmitting (%u) SYN|ACK\n",
480 s, __func__, sc->sc_rxmits);
484 syncache_respond(sc, sch, 1, NULL);
485 TCPSTAT_INC(tcps_sc_retransmitted);
486 syncache_timeout(sc, sch, 0);
488 if (!TAILQ_EMPTY(&(sch)->sch_bucket))
489 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
490 syncache_timer, (void *)(sch));
495 * Find an entry in the syncache.
496 * Returns always with locked syncache_head plus a matching entry or NULL.
498 static struct syncache *
499 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
502 struct syncache_head *sch;
506 * The hash is built on foreign port + local port + foreign address.
507 * We rely on the fact that struct in_conninfo starts with 16 bits
508 * of foreign port, then 16 bits of local port then followed by 128
509 * bits of foreign address. In case of IPv4 address, the first 3
510 * 32-bit words of the address always are zeroes.
512 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
513 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
515 sch = &V_tcp_syncache.hashbase[hash];
519 /* Circle through bucket row to find matching entry. */
520 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
521 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
522 sizeof(struct in_endpoints)) == 0)
525 return (sc); /* Always returns with locked sch. */
529 * This function is called when we get a RST for a
530 * non-existent connection, so that we can see if the
531 * connection is in the syn cache. If it is, zap it.
534 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
537 struct syncache_head *sch;
540 sc = syncache_lookup(inc, &sch); /* returns locked sch */
541 SCH_LOCK_ASSERT(sch);
544 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
545 * See RFC 793 page 65, section SEGMENT ARRIVES.
547 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
548 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
549 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
550 "FIN flag set, segment ignored\n", s, __func__);
551 TCPSTAT_INC(tcps_badrst);
556 * No corresponding connection was found in syncache.
557 * If syncookies are enabled and possibly exclusively
558 * used, or we are under memory pressure, a valid RST
559 * may not find a syncache entry. In that case we're
560 * done and no SYN|ACK retransmissions will happen.
561 * Otherwise the RST was misdirected or spoofed.
564 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
565 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
566 "syncache entry (possibly syncookie only), "
567 "segment ignored\n", s, __func__);
568 TCPSTAT_INC(tcps_badrst);
573 * If the RST bit is set, check the sequence number to see
574 * if this is a valid reset segment.
576 * In all states except SYN-SENT, all reset (RST) segments
577 * are validated by checking their SEQ-fields. A reset is
578 * valid if its sequence number is in the window.
580 * The sequence number in the reset segment is normally an
581 * echo of our outgoing acknowlegement numbers, but some hosts
582 * send a reset with the sequence number at the rightmost edge
583 * of our receive window, and we have to handle this case.
585 if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
586 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
587 syncache_drop(sc, sch);
588 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
589 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, "
590 "connection attempt aborted by remote endpoint\n",
592 TCPSTAT_INC(tcps_sc_reset);
594 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
595 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
596 "IRS %u (+WND %u), segment ignored\n",
597 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd);
598 TCPSTAT_INC(tcps_badrst);
608 syncache_badack(struct in_conninfo *inc)
611 struct syncache_head *sch;
613 sc = syncache_lookup(inc, &sch); /* returns locked sch */
614 SCH_LOCK_ASSERT(sch);
616 syncache_drop(sc, sch);
617 TCPSTAT_INC(tcps_sc_badack);
623 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq)
626 struct syncache_head *sch;
628 sc = syncache_lookup(inc, &sch); /* returns locked sch */
629 SCH_LOCK_ASSERT(sch);
633 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
634 if (ntohl(th_seq) != sc->sc_iss)
638 * If we've rertransmitted 3 times and this is our second error,
639 * we remove the entry. Otherwise, we allow it to continue on.
640 * This prevents us from incorrectly nuking an entry during a
641 * spurious network outage.
645 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
646 sc->sc_flags |= SCF_UNREACH;
649 syncache_drop(sc, sch);
650 TCPSTAT_INC(tcps_sc_unreach);
656 * Build a new TCP socket structure from a syncache entry.
658 * On success return the newly created socket with its underlying inp locked.
660 static struct socket *
661 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
663 struct tcp_function_block *blk;
664 struct inpcb *inp = NULL;
670 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
673 * Ok, create the full blown connection, and set things up
674 * as they would have been set up if we had created the
675 * connection when the SYN arrived. If we can't create
676 * the connection, abort it.
678 so = sonewconn(lso, 0);
681 * Drop the connection; we will either send a RST or
682 * have the peer retransmit its SYN again after its
685 TCPSTAT_INC(tcps_listendrop);
686 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
687 log(LOG_DEBUG, "%s; %s: Socket create failed "
688 "due to limits or memory shortage\n",
695 mac_socketpeer_set_from_mbuf(m, so);
699 inp->inp_inc.inc_fibnum = so->so_fibnum;
702 * Exclusive pcbinfo lock is not required in syncache socket case even
703 * if two inpcb locks can be acquired simultaneously:
704 * - the inpcb in LISTEN state,
705 * - the newly created inp.
707 * In this case, an inp cannot be at same time in LISTEN state and
708 * just created by an accept() call.
710 INP_HASH_WLOCK(&V_tcbinfo);
712 /* Insert new socket into PCB hash list. */
713 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
715 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
716 inp->inp_vflag &= ~INP_IPV4;
717 inp->inp_vflag |= INP_IPV6;
718 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
720 inp->inp_vflag &= ~INP_IPV6;
721 inp->inp_vflag |= INP_IPV4;
723 inp->inp_laddr = sc->sc_inc.inc_laddr;
729 * If there's an mbuf and it has a flowid, then let's initialise the
730 * inp with that particular flowid.
732 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
733 inp->inp_flowid = m->m_pkthdr.flowid;
734 inp->inp_flowtype = M_HASHTYPE_GET(m);
738 * Install in the reservation hash table for now, but don't yet
739 * install a connection group since the full 4-tuple isn't yet
742 inp->inp_lport = sc->sc_inc.inc_lport;
743 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
745 * Undo the assignments above if we failed to
746 * put the PCB on the hash lists.
749 if (sc->sc_inc.inc_flags & INC_ISIPV6)
750 inp->in6p_laddr = in6addr_any;
753 inp->inp_laddr.s_addr = INADDR_ANY;
755 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
756 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
761 INP_HASH_WUNLOCK(&V_tcbinfo);
765 if (inp->inp_vflag & INP_IPV6PROTO) {
766 struct inpcb *oinp = sotoinpcb(lso);
769 * Inherit socket options from the listening socket.
770 * Note that in6p_inputopts are not (and should not be)
771 * copied, since it stores previously received options and is
772 * used to detect if each new option is different than the
773 * previous one and hence should be passed to a user.
774 * If we copied in6p_inputopts, a user would not be able to
775 * receive options just after calling the accept system call.
777 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
778 if (oinp->in6p_outputopts)
779 inp->in6p_outputopts =
780 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
783 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
784 struct in6_addr laddr6;
785 struct sockaddr_in6 sin6;
787 sin6.sin6_family = AF_INET6;
788 sin6.sin6_len = sizeof(sin6);
789 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
790 sin6.sin6_port = sc->sc_inc.inc_fport;
791 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
792 laddr6 = inp->in6p_laddr;
793 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
794 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
795 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
796 thread0.td_ucred, m)) != 0) {
797 inp->in6p_laddr = laddr6;
798 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
799 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
804 INP_HASH_WUNLOCK(&V_tcbinfo);
807 /* Override flowlabel from in6_pcbconnect. */
808 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
809 inp->inp_flow |= sc->sc_flowlabel;
812 #if defined(INET) && defined(INET6)
817 struct in_addr laddr;
818 struct sockaddr_in sin;
820 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
822 if (inp->inp_options == NULL) {
823 inp->inp_options = sc->sc_ipopts;
824 sc->sc_ipopts = NULL;
827 sin.sin_family = AF_INET;
828 sin.sin_len = sizeof(sin);
829 sin.sin_addr = sc->sc_inc.inc_faddr;
830 sin.sin_port = sc->sc_inc.inc_fport;
831 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
832 laddr = inp->inp_laddr;
833 if (inp->inp_laddr.s_addr == INADDR_ANY)
834 inp->inp_laddr = sc->sc_inc.inc_laddr;
835 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
836 thread0.td_ucred, m)) != 0) {
837 inp->inp_laddr = laddr;
838 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
839 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
844 INP_HASH_WUNLOCK(&V_tcbinfo);
849 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
850 /* Copy old policy into new socket's. */
851 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
852 printf("syncache_socket: could not copy policy\n");
854 INP_HASH_WUNLOCK(&V_tcbinfo);
856 tcp_state_change(tp, TCPS_SYN_RECEIVED);
857 tp->iss = sc->sc_iss;
858 tp->irs = sc->sc_irs;
861 blk = sototcpcb(lso)->t_fb;
862 if (blk != tp->t_fb) {
864 * Our parents t_fb was not the default,
865 * we need to release our ref on tp->t_fb and
866 * pickup one on the new entry.
868 struct tcp_function_block *rblk;
870 rblk = find_and_ref_tcp_fb(blk);
871 KASSERT(rblk != NULL,
872 ("cannot find blk %p out of syncache?", blk));
873 if (tp->t_fb->tfb_tcp_fb_fini)
874 (*tp->t_fb->tfb_tcp_fb_fini)(tp);
875 refcount_release(&tp->t_fb->tfb_refcnt);
877 if (tp->t_fb->tfb_tcp_fb_init) {
878 (*tp->t_fb->tfb_tcp_fb_init)(tp);
881 tp->snd_wl1 = sc->sc_irs;
882 tp->snd_max = tp->iss + 1;
883 tp->snd_nxt = tp->iss + 1;
884 tp->rcv_up = sc->sc_irs + 1;
885 tp->rcv_wnd = sc->sc_wnd;
886 tp->rcv_adv += tp->rcv_wnd;
887 tp->last_ack_sent = tp->rcv_nxt;
889 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
890 if (sc->sc_flags & SCF_NOOPT)
891 tp->t_flags |= TF_NOOPT;
893 if (sc->sc_flags & SCF_WINSCALE) {
894 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
895 tp->snd_scale = sc->sc_requested_s_scale;
896 tp->request_r_scale = sc->sc_requested_r_scale;
898 if (sc->sc_flags & SCF_TIMESTAMP) {
899 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
900 tp->ts_recent = sc->sc_tsreflect;
901 tp->ts_recent_age = tcp_ts_getticks();
902 tp->ts_offset = sc->sc_tsoff;
904 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
905 if (sc->sc_flags & SCF_SIGNATURE)
906 tp->t_flags |= TF_SIGNATURE;
908 if (sc->sc_flags & SCF_SACK)
909 tp->t_flags |= TF_SACK_PERMIT;
912 if (sc->sc_flags & SCF_ECN)
913 tp->t_flags |= TF_ECN_PERMIT;
916 * Set up MSS and get cached values from tcp_hostcache.
917 * This might overwrite some of the defaults we just set.
919 tcp_mss(tp, sc->sc_peer_mss);
922 * If the SYN,ACK was retransmitted, indicate that CWND to be
923 * limited to one segment in cc_conn_init().
924 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
926 if (sc->sc_rxmits > 1)
931 * Allow a TOE driver to install its hooks. Note that we hold the
932 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
933 * new connection before the TOE driver has done its thing.
935 if (ADDED_BY_TOE(sc)) {
936 struct toedev *tod = sc->sc_tod;
938 tod->tod_offload_socket(tod, sc->sc_todctx, so);
942 * Copy and activate timers.
944 tp->t_keepinit = sototcpcb(lso)->t_keepinit;
945 tp->t_keepidle = sototcpcb(lso)->t_keepidle;
946 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
947 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
948 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
950 TCPSTAT_INC(tcps_accepts);
962 * This function gets called when we receive an ACK for a
963 * socket in the LISTEN state. We look up the connection
964 * in the syncache, and if its there, we pull it out of
965 * the cache and turn it into a full-blown connection in
966 * the SYN-RECEIVED state.
968 * On syncache_socket() success the newly created socket
969 * has its underlying inp locked.
972 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
973 struct socket **lsop, struct mbuf *m)
976 struct syncache_head *sch;
981 * Global TCP locks are held because we manipulate the PCB lists
982 * and create a new socket.
984 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
985 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
986 ("%s: can handle only ACK", __func__));
988 sc = syncache_lookup(inc, &sch); /* returns locked sch */
989 SCH_LOCK_ASSERT(sch);
993 * Test code for syncookies comparing the syncache stored
994 * values with the reconstructed values from the cookie.
997 syncookie_cmp(inc, sch, sc, th, to, *lsop);
1002 * There is no syncache entry, so see if this ACK is
1003 * a returning syncookie. To do this, first:
1004 * A. Check if syncookies are used in case of syncache
1006 * B. See if this socket has had a syncache entry dropped in
1007 * the recent past. We don't want to accept a bogus
1008 * syncookie if we've never received a SYN or accept it
1010 * C. check that the syncookie is valid. If it is, then
1011 * cobble up a fake syncache entry, and return.
1013 if (!V_tcp_syncookies) {
1015 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1016 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1017 "segment rejected (syncookies disabled)\n",
1021 if (!V_tcp_syncookiesonly &&
1022 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1024 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1025 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1026 "segment rejected (no syncache entry)\n",
1030 bzero(&scs, sizeof(scs));
1031 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
1034 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1035 log(LOG_DEBUG, "%s; %s: Segment failed "
1036 "SYNCOOKIE authentication, segment rejected "
1037 "(probably spoofed)\n", s, __func__);
1040 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1041 /* If received ACK has MD5 signature, check it. */
1042 if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1043 (!TCPMD5_ENABLED() ||
1044 TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1046 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1047 log(LOG_DEBUG, "%s; %s: Segment rejected, "
1048 "MD5 signature doesn't match.\n",
1052 TCPSTAT_INC(tcps_sig_err_sigopt);
1053 return (-1); /* Do not send RST */
1055 #endif /* TCP_SIGNATURE */
1057 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1059 * If listening socket requested TCP digests, check that
1060 * received ACK has signature and it is correct.
1061 * If not, drop the ACK and leave sc entry in th cache,
1062 * because SYN was received with correct signature.
1064 if (sc->sc_flags & SCF_SIGNATURE) {
1065 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1067 TCPSTAT_INC(tcps_sig_err_nosigopt);
1069 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1070 log(LOG_DEBUG, "%s; %s: Segment "
1071 "rejected, MD5 signature wasn't "
1072 "provided.\n", s, __func__);
1075 return (-1); /* Do not send RST */
1077 if (!TCPMD5_ENABLED() ||
1078 TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1079 /* Doesn't match or no SA */
1081 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1082 log(LOG_DEBUG, "%s; %s: Segment "
1083 "rejected, MD5 signature doesn't "
1084 "match.\n", s, __func__);
1087 return (-1); /* Do not send RST */
1090 #endif /* TCP_SIGNATURE */
1092 * Pull out the entry to unlock the bucket row.
1094 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1095 * tcp_state_change(). The tcpcb is not existent at this
1096 * moment. A new one will be allocated via syncache_socket->
1097 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1098 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1100 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1101 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1104 if (ADDED_BY_TOE(sc)) {
1105 struct toedev *tod = sc->sc_tod;
1107 tod->tod_syncache_removed(tod, sc->sc_todctx);
1114 * Segment validation:
1115 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1117 if (th->th_ack != sc->sc_iss + 1) {
1118 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1119 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1120 "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1125 * The SEQ must fall in the window starting at the received
1126 * initial receive sequence number + 1 (the SYN).
1128 if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1129 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1130 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1131 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1132 "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1137 * If timestamps were not negotiated during SYN/ACK they
1138 * must not appear on any segment during this session.
1140 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1141 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1142 log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1143 "segment rejected\n", s, __func__);
1148 * If timestamps were negotiated during SYN/ACK they should
1149 * appear on every segment during this session.
1150 * XXXAO: This is only informal as there have been unverified
1151 * reports of non-compliants stacks.
1153 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1154 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1155 log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1156 "no action\n", s, __func__);
1162 *lsop = syncache_socket(sc, *lsop, m);
1165 TCPSTAT_INC(tcps_sc_aborted);
1167 TCPSTAT_INC(tcps_sc_completed);
1169 /* how do we find the inp for the new socket? */
1174 if (sc != NULL && sc != &scs)
1184 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
1185 uint64_t response_cookie)
1189 unsigned int *pending_counter;
1192 * Global TCP locks are held because we manipulate the PCB lists
1193 * and create a new socket.
1195 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1197 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
1198 *lsop = syncache_socket(sc, *lsop, m);
1199 if (*lsop == NULL) {
1200 TCPSTAT_INC(tcps_sc_aborted);
1201 atomic_subtract_int(pending_counter, 1);
1203 inp = sotoinpcb(*lsop);
1204 tp = intotcpcb(inp);
1205 tp->t_flags |= TF_FASTOPEN;
1206 tp->t_tfo_cookie = response_cookie;
1207 tp->snd_max = tp->iss;
1208 tp->snd_nxt = tp->iss;
1209 tp->t_tfo_pending = pending_counter;
1210 TCPSTAT_INC(tcps_sc_completed);
1213 #endif /* TCP_RFC7413 */
1216 * Given a LISTEN socket and an inbound SYN request, add
1217 * this to the syn cache, and send back a segment:
1218 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1221 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1222 * Doing so would require that we hold onto the data and deliver it
1223 * to the application. However, if we are the target of a SYN-flood
1224 * DoS attack, an attacker could send data which would eventually
1225 * consume all available buffer space if it were ACKed. By not ACKing
1226 * the data, we avoid this DoS scenario.
1228 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1229 * cookie is processed, V_tcp_fastopen_enabled set to true, and the
1230 * TCP_FASTOPEN socket option is set. In this case, a new socket is created
1231 * and returned via lsop, the mbuf is not freed so that tcp_input() can
1232 * queue its data to the socket, and 1 is returned to indicate the
1233 * TFO-socket-creation path was taken.
1236 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1237 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1242 struct syncache *sc = NULL;
1243 struct syncache_head *sch;
1244 struct mbuf *ipopts = NULL;
1246 int win, sb_hiwat, ip_ttl, ip_tos;
1250 int autoflowlabel = 0;
1253 struct label *maclabel;
1255 struct syncache scs;
1258 uint64_t tfo_response_cookie;
1259 int tfo_cookie_valid = 0;
1260 int tfo_response_cookie_valid = 0;
1263 INP_WLOCK_ASSERT(inp); /* listen socket */
1264 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1265 ("%s: unexpected tcp flags", __func__));
1268 * Combine all so/tp operations very early to drop the INP lock as
1273 cred = crhold(so->so_cred);
1276 if ((inc->inc_flags & INC_ISIPV6) &&
1277 (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1280 ip_ttl = inp->inp_ip_ttl;
1281 ip_tos = inp->inp_ip_tos;
1282 win = sbspace(&so->so_rcv);
1283 sb_hiwat = so->so_rcv.sb_hiwat;
1284 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1287 if (V_tcp_fastopen_enabled && (tp->t_flags & TF_FASTOPEN) &&
1288 (tp->t_tfo_pending != NULL) && (to->to_flags & TOF_FASTOPEN)) {
1290 * Limit the number of pending TFO connections to
1291 * approximately half of the queue limit. This prevents TFO
1292 * SYN floods from starving the service by filling the
1293 * listen queue with bogus TFO connections.
1295 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1296 (so->so_qlimit / 2)) {
1299 result = tcp_fastopen_check_cookie(inc,
1300 to->to_tfo_cookie, to->to_tfo_len,
1301 &tfo_response_cookie);
1302 tfo_cookie_valid = (result > 0);
1303 tfo_response_cookie_valid = (result >= 0);
1305 atomic_subtract_int(tp->t_tfo_pending, 1);
1309 /* By the time we drop the lock these should no longer be used. */
1314 if (mac_syncache_init(&maclabel) != 0) {
1318 mac_syncache_create(maclabel, inp);
1321 if (!tfo_cookie_valid)
1326 * Remember the IP options, if any.
1329 if (!(inc->inc_flags & INC_ISIPV6))
1332 ipopts = (m) ? ip_srcroute(m) : NULL;
1337 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1339 * If listening socket requested TCP digests, check that received
1340 * SYN has signature and it is correct. If signature doesn't match
1341 * or TCP_SIGNATURE support isn't enabled, drop the packet.
1343 if (ltflags & TF_SIGNATURE) {
1344 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1345 TCPSTAT_INC(tcps_sig_err_nosigopt);
1348 if (!TCPMD5_ENABLED() ||
1349 TCPMD5_INPUT(m, th, to->to_signature) != 0)
1352 #endif /* TCP_SIGNATURE */
1354 * See if we already have an entry for this connection.
1355 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1357 * XXX: should the syncache be re-initialized with the contents
1358 * of the new SYN here (which may have different options?)
1360 * XXX: We do not check the sequence number to see if this is a
1361 * real retransmit or a new connection attempt. The question is
1362 * how to handle such a case; either ignore it as spoofed, or
1363 * drop the current entry and create a new one?
1365 sc = syncache_lookup(inc, &sch); /* returns locked entry */
1366 SCH_LOCK_ASSERT(sch);
1369 if (tfo_cookie_valid)
1372 TCPSTAT_INC(tcps_sc_dupsyn);
1375 * If we were remembering a previous source route,
1376 * forget it and use the new one we've been given.
1379 (void) m_free(sc->sc_ipopts);
1380 sc->sc_ipopts = ipopts;
1383 * Update timestamp if present.
1385 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1386 sc->sc_tsreflect = to->to_tsval;
1388 sc->sc_flags &= ~SCF_TIMESTAMP;
1391 * Since we have already unconditionally allocated label
1392 * storage, free it up. The syncache entry will already
1393 * have an initialized label we can use.
1395 mac_syncache_destroy(&maclabel);
1397 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1398 /* Retransmit SYN|ACK and reset retransmit count. */
1399 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1400 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1401 "resetting timer and retransmitting SYN|ACK\n",
1405 if (syncache_respond(sc, sch, 1, m) == 0) {
1407 syncache_timeout(sc, sch, 1);
1408 TCPSTAT_INC(tcps_sndacks);
1409 TCPSTAT_INC(tcps_sndtotal);
1416 if (tfo_cookie_valid) {
1417 bzero(&scs, sizeof(scs));
1423 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1426 * The zone allocator couldn't provide more entries.
1427 * Treat this as if the cache was full; drop the oldest
1428 * entry and insert the new one.
1430 TCPSTAT_INC(tcps_sc_zonefail);
1431 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1432 sch->sch_last_overflow = time_uptime;
1433 syncache_drop(sc, sch);
1435 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1437 if (V_tcp_syncookies) {
1438 bzero(&scs, sizeof(scs));
1443 (void) m_free(ipopts);
1451 if (!tfo_cookie_valid && tfo_response_cookie_valid)
1452 sc->sc_tfo_cookie = &tfo_response_cookie;
1456 * Fill in the syncache values.
1459 sc->sc_label = maclabel;
1463 sc->sc_ipopts = ipopts;
1464 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1466 if (!(inc->inc_flags & INC_ISIPV6))
1469 sc->sc_ip_tos = ip_tos;
1470 sc->sc_ip_ttl = ip_ttl;
1474 sc->sc_todctx = todctx;
1476 sc->sc_irs = th->th_seq;
1477 sc->sc_iss = arc4random();
1479 sc->sc_flowlabel = 0;
1482 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1483 * win was derived from socket earlier in the function.
1486 win = imin(win, TCP_MAXWIN);
1489 if (V_tcp_do_rfc1323) {
1491 * A timestamp received in a SYN makes
1492 * it ok to send timestamp requests and replies.
1494 if (to->to_flags & TOF_TS) {
1495 sc->sc_tsreflect = to->to_tsval;
1496 sc->sc_flags |= SCF_TIMESTAMP;
1498 if (to->to_flags & TOF_SCALE) {
1502 * Pick the smallest possible scaling factor that
1503 * will still allow us to scale up to sb_max, aka
1504 * kern.ipc.maxsockbuf.
1506 * We do this because there are broken firewalls that
1507 * will corrupt the window scale option, leading to
1508 * the other endpoint believing that our advertised
1509 * window is unscaled. At scale factors larger than
1510 * 5 the unscaled window will drop below 1500 bytes,
1511 * leading to serious problems when traversing these
1514 * With the default maxsockbuf of 256K, a scale factor
1515 * of 3 will be chosen by this algorithm. Those who
1516 * choose a larger maxsockbuf should watch out
1517 * for the compatibility problems mentioned above.
1519 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1520 * or <SYN,ACK>) segment itself is never scaled.
1522 while (wscale < TCP_MAX_WINSHIFT &&
1523 (TCP_MAXWIN << wscale) < sb_max)
1525 sc->sc_requested_r_scale = wscale;
1526 sc->sc_requested_s_scale = to->to_wscale;
1527 sc->sc_flags |= SCF_WINSCALE;
1530 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1532 * If listening socket requested TCP digests, flag this in the
1533 * syncache so that syncache_respond() will do the right thing
1536 if (ltflags & TF_SIGNATURE)
1537 sc->sc_flags |= SCF_SIGNATURE;
1538 #endif /* TCP_SIGNATURE */
1539 if (to->to_flags & TOF_SACKPERM)
1540 sc->sc_flags |= SCF_SACK;
1541 if (to->to_flags & TOF_MSS)
1542 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */
1543 if (ltflags & TF_NOOPT)
1544 sc->sc_flags |= SCF_NOOPT;
1545 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1546 sc->sc_flags |= SCF_ECN;
1548 if (V_tcp_syncookies)
1549 sc->sc_iss = syncookie_generate(sch, sc);
1551 if (autoflowlabel) {
1552 if (V_tcp_syncookies)
1553 sc->sc_flowlabel = sc->sc_iss;
1555 sc->sc_flowlabel = ip6_randomflowlabel();
1556 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 called */
1570 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1572 * Do a standard 3-way handshake.
1574 if (syncache_respond(sc, sch, 0, m) == 0) {
1575 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1577 else if (sc != &scs)
1578 syncache_insert(sc, sch); /* locks and unlocks sch */
1579 TCPSTAT_INC(tcps_sndacks);
1580 TCPSTAT_INC(tcps_sndtotal);
1584 TCPSTAT_INC(tcps_sc_dropped);
1589 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1602 mac_syncache_destroy(&maclabel);
1608 * Send SYN|ACK to the peer. Either in response to the peer's SYN,
1609 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1612 syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked,
1613 const struct mbuf *m0)
1615 struct ip *ip = NULL;
1617 struct tcphdr *th = NULL;
1618 int optlen, error = 0; /* Make compiler happy */
1619 u_int16_t hlen, tlen, mssopt;
1622 struct ip6_hdr *ip6 = NULL;
1626 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1629 tlen = hlen + sizeof(struct tcphdr);
1631 /* Determine MSS we advertize to other end of connection. */
1632 mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss);
1634 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1635 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1636 ("syncache: mbuf too small"));
1638 /* Create the IP+TCP header from scratch. */
1639 m = m_gethdr(M_NOWAIT, MT_DATA);
1643 mac_syncache_create_mbuf(sc->sc_label, m);
1645 m->m_data += max_linkhdr;
1647 m->m_pkthdr.len = tlen;
1648 m->m_pkthdr.rcvif = NULL;
1651 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1652 ip6 = mtod(m, struct ip6_hdr *);
1653 ip6->ip6_vfc = IPV6_VERSION;
1654 ip6->ip6_nxt = IPPROTO_TCP;
1655 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1656 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1657 ip6->ip6_plen = htons(tlen - hlen);
1658 /* ip6_hlim is set after checksum */
1659 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1660 ip6->ip6_flow |= sc->sc_flowlabel;
1662 th = (struct tcphdr *)(ip6 + 1);
1665 #if defined(INET6) && defined(INET)
1670 ip = mtod(m, struct ip *);
1671 ip->ip_v = IPVERSION;
1672 ip->ip_hl = sizeof(struct ip) >> 2;
1673 ip->ip_len = htons(tlen);
1677 ip->ip_p = IPPROTO_TCP;
1678 ip->ip_src = sc->sc_inc.inc_laddr;
1679 ip->ip_dst = sc->sc_inc.inc_faddr;
1680 ip->ip_ttl = sc->sc_ip_ttl;
1681 ip->ip_tos = sc->sc_ip_tos;
1684 * See if we should do MTU discovery. Route lookups are
1685 * expensive, so we will only unset the DF bit if:
1687 * 1) path_mtu_discovery is disabled
1688 * 2) the SCF_UNREACH flag has been set
1690 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1691 ip->ip_off |= htons(IP_DF);
1693 th = (struct tcphdr *)(ip + 1);
1696 th->th_sport = sc->sc_inc.inc_lport;
1697 th->th_dport = sc->sc_inc.inc_fport;
1699 th->th_seq = htonl(sc->sc_iss);
1700 th->th_ack = htonl(sc->sc_irs + 1);
1701 th->th_off = sizeof(struct tcphdr) >> 2;
1703 th->th_flags = TH_SYN|TH_ACK;
1704 th->th_win = htons(sc->sc_wnd);
1707 if (sc->sc_flags & SCF_ECN) {
1708 th->th_flags |= TH_ECE;
1709 TCPSTAT_INC(tcps_ecn_shs);
1712 /* Tack on the TCP options. */
1713 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1717 to.to_flags = TOF_MSS;
1718 if (sc->sc_flags & SCF_WINSCALE) {
1719 to.to_wscale = sc->sc_requested_r_scale;
1720 to.to_flags |= TOF_SCALE;
1722 if (sc->sc_flags & SCF_TIMESTAMP) {
1723 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1724 to.to_tsecr = sc->sc_tsreflect;
1725 to.to_flags |= TOF_TS;
1727 if (sc->sc_flags & SCF_SACK)
1728 to.to_flags |= TOF_SACKPERM;
1729 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1730 if (sc->sc_flags & SCF_SIGNATURE)
1731 to.to_flags |= TOF_SIGNATURE;
1734 if (sc->sc_tfo_cookie) {
1735 to.to_flags |= TOF_FASTOPEN;
1736 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1737 to.to_tfo_cookie = sc->sc_tfo_cookie;
1738 /* don't send cookie again when retransmitting response */
1739 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 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
1797 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1800 #if defined(INET6) && defined(INET)
1805 m->m_pkthdr.csum_flags = CSUM_TCP;
1806 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1807 htons(tlen + optlen - hlen + IPPROTO_TCP));
1809 if (ADDED_BY_TOE(sc)) {
1810 struct toedev *tod = sc->sc_tod;
1812 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1817 TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
1818 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1825 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1826 * that exceed the capacity of the syncache by avoiding the storage of any
1827 * of the SYNs we receive. Syncookies defend against blind SYN flooding
1828 * attacks where the attacker does not have access to our responses.
1830 * Syncookies encode and include all necessary information about the
1831 * connection setup within the SYN|ACK that we send back. That way we
1832 * can avoid keeping any local state until the ACK to our SYN|ACK returns
1833 * (if ever). Normally the syncache and syncookies are running in parallel
1834 * with the latter taking over when the former is exhausted. When matching
1835 * syncache entry is found the syncookie is ignored.
1837 * The only reliable information persisting the 3WHS is our initial sequence
1838 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient
1839 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1840 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK
1841 * returns and signifies a legitimate connection if it matches the ACK.
1843 * The available space of 32 bits to store the hash and to encode the SYN
1844 * option information is very tight and we should have at least 24 bits for
1845 * the MAC to keep the number of guesses by blind spoofing reasonably high.
1847 * SYN option information we have to encode to fully restore a connection:
1848 * MSS: is imporant to chose an optimal segment size to avoid IP level
1849 * fragmentation along the path. The common MSS values can be encoded
1850 * in a 3-bit table. Uncommon values are captured by the next lower value
1851 * in the table leading to a slight increase in packetization overhead.
1852 * WSCALE: is necessary to allow large windows to be used for high delay-
1853 * bandwidth product links. Not scaling the window when it was initially
1854 * negotiated is bad for performance as lack of scaling further decreases
1855 * the apparent available send window. We only need to encode the WSCALE
1856 * we received from the remote end. Our end can be recalculated at any
1857 * time. The common WSCALE values can be encoded in a 3-bit table.
1858 * Uncommon values are captured by the next lower value in the table
1859 * making us under-estimate the available window size halving our
1860 * theoretically possible maximum throughput for that connection.
1861 * SACK: Greatly assists in packet loss recovery and requires 1 bit.
1862 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
1863 * that are included in all segments on a connection. We enable them when
1866 * Security of syncookies and attack vectors:
1868 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
1869 * together with the gloabl secret to make it unique per connection attempt.
1870 * Thus any change of any of those parameters results in a different MAC output
1871 * in an unpredictable way unless a collision is encountered. 24 bits of the
1872 * MAC are embedded into the ISS.
1874 * To prevent replay attacks two rotating global secrets are updated with a
1875 * new random value every 15 seconds. The life-time of a syncookie is thus
1878 * Vector 1: Attacking the secret. This requires finding a weakness in the
1879 * MAC itself or the way it is used here. The attacker can do a chosen plain
1880 * text attack by varying and testing the all parameters under his control.
1881 * The strength depends on the size and randomness of the secret, and the
1882 * cryptographic security of the MAC function. Due to the constant updating
1883 * of the secret the attacker has at most 29.999 seconds to find the secret
1884 * and launch spoofed connections. After that he has to start all over again.
1886 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC
1887 * size an average of 4,823 attempts are required for a 50% chance of success
1888 * to spoof a single syncookie (birthday collision paradox). However the
1889 * attacker is blind and doesn't know if one of his attempts succeeded unless
1890 * he has a side channel to interfere success from. A single connection setup
1891 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
1892 * This many attempts are required for each one blind spoofed connection. For
1893 * every additional spoofed connection he has to launch another N attempts.
1894 * Thus for a sustained rate 100 spoofed connections per second approximately
1895 * 1,800,000 packets per second would have to be sent.
1897 * NB: The MAC function should be fast so that it doesn't become a CPU
1898 * exhaustion attack vector itself.
1901 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations
1902 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
1903 * http://cr.yp.to/syncookies.html (overview)
1904 * http://cr.yp.to/syncookies/archive (details)
1907 * Schematic construction of a syncookie enabled Initial Sequence Number:
1909 * 12345678901234567890123456789012
1910 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
1912 * x 24 MAC (truncated)
1913 * W 3 Send Window Scale index
1915 * S 1 SACK permitted
1916 * P 1 Odd/even secret
1920 * Distribution and probability of certain MSS values. Those in between are
1921 * rounded down to the next lower one.
1922 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
1923 * .2% .3% 5% 7% 7% 20% 15% 45%
1925 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
1928 * Distribution and probability of certain WSCALE values. We have to map the
1929 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
1930 * bits based on prevalence of certain values. Where we don't have an exact
1931 * match for are rounded down to the next lower one letting us under-estimate
1932 * the true available window. At the moment this would happen only for the
1933 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
1934 * and window size). The absence of the WSCALE option (no scaling in either
1935 * direction) is encoded with index zero.
1936 * [WSCALE values histograms, Allman, 2012]
1937 * X 10 10 35 5 6 14 10% by host
1938 * X 11 4 5 5 18 49 3% by connections
1940 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
1943 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed
1944 * and good cryptographic properties.
1947 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
1948 uint8_t *secbits, uintptr_t secmod)
1951 uint32_t siphash[2];
1953 SipHash24_Init(&ctx);
1954 SipHash_SetKey(&ctx, secbits);
1955 switch (inc->inc_flags & INC_ISIPV6) {
1958 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
1959 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
1964 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
1965 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
1969 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
1970 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
1971 SipHash_Update(&ctx, &irs, sizeof(irs));
1972 SipHash_Update(&ctx, &flags, sizeof(flags));
1973 SipHash_Update(&ctx, &secmod, sizeof(secmod));
1974 SipHash_Final((u_int8_t *)&siphash, &ctx);
1976 return (siphash[0] ^ siphash[1]);
1980 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
1982 u_int i, secbit, wscale;
1985 union syncookie cookie;
1987 SCH_LOCK_ASSERT(sch);
1991 /* Map our computed MSS into the 3-bit index. */
1992 for (i = nitems(tcp_sc_msstab) - 1;
1993 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
1996 cookie.flags.mss_idx = i;
1999 * Map the send window scale into the 3-bit index but only if
2000 * the wscale option was received.
2002 if (sc->sc_flags & SCF_WINSCALE) {
2003 wscale = sc->sc_requested_s_scale;
2004 for (i = nitems(tcp_sc_wstab) - 1;
2005 tcp_sc_wstab[i] > wscale && i > 0;
2008 cookie.flags.wscale_idx = i;
2011 /* Can we do SACK? */
2012 if (sc->sc_flags & SCF_SACK)
2013 cookie.flags.sack_ok = 1;
2015 /* Which of the two secrets to use. */
2016 secbit = sch->sch_sc->secret.oddeven & 0x1;
2017 cookie.flags.odd_even = secbit;
2019 secbits = sch->sch_sc->secret.key[secbit];
2020 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2024 * Put the flags into the hash and XOR them to get better ISS number
2025 * variance. This doesn't enhance the cryptographic strength and is
2026 * done to prevent the 8 cookie bits from showing up directly on the
2030 iss |= cookie.cookie ^ (hash >> 24);
2032 /* Randomize the timestamp. */
2033 if (sc->sc_flags & SCF_TIMESTAMP) {
2034 sc->sc_tsoff = arc4random() - tcp_ts_getticks();
2037 TCPSTAT_INC(tcps_sc_sendcookie);
2041 static struct syncache *
2042 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2043 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2049 int wnd, wscale = 0;
2050 union syncookie cookie;
2052 SCH_LOCK_ASSERT(sch);
2055 * Pull information out of SYN-ACK/ACK and revert sequence number
2058 ack = th->th_ack - 1;
2059 seq = th->th_seq - 1;
2062 * Unpack the flags containing enough information to restore the
2065 cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2067 /* Which of the two secrets to use. */
2068 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
2070 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2072 /* The recomputed hash matches the ACK if this was a genuine cookie. */
2073 if ((ack & ~0xff) != (hash & ~0xff))
2076 /* Fill in the syncache values. */
2078 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2079 sc->sc_ipopts = NULL;
2084 switch (inc->inc_flags & INC_ISIPV6) {
2087 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2088 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2093 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2094 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
2099 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2101 /* We can simply recompute receive window scale we sent earlier. */
2102 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2105 /* Only use wscale if it was enabled in the orignal SYN. */
2106 if (cookie.flags.wscale_idx > 0) {
2107 sc->sc_requested_r_scale = wscale;
2108 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2109 sc->sc_flags |= SCF_WINSCALE;
2112 wnd = sbspace(&lso->so_rcv);
2114 wnd = imin(wnd, TCP_MAXWIN);
2117 if (cookie.flags.sack_ok)
2118 sc->sc_flags |= SCF_SACK;
2120 if (to->to_flags & TOF_TS) {
2121 sc->sc_flags |= SCF_TIMESTAMP;
2122 sc->sc_tsreflect = to->to_tsval;
2123 sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks();
2126 if (to->to_flags & TOF_SIGNATURE)
2127 sc->sc_flags |= SCF_SIGNATURE;
2131 TCPSTAT_INC(tcps_sc_recvcookie);
2137 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2138 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2141 struct syncache scs, *scx;
2144 bzero(&scs, sizeof(scs));
2145 scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
2147 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2151 if (sc->sc_peer_mss != scx->sc_peer_mss)
2152 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2153 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2155 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2156 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2157 s, __func__, sc->sc_requested_r_scale,
2158 scx->sc_requested_r_scale);
2160 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2161 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2162 s, __func__, sc->sc_requested_s_scale,
2163 scx->sc_requested_s_scale);
2165 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2166 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2173 #endif /* INVARIANTS */
2176 syncookie_reseed(void *arg)
2178 struct tcp_syncache *sc = arg;
2183 * Reseeding the secret doesn't have to be protected by a lock.
2184 * It only must be ensured that the new random values are visible
2185 * to all CPUs in a SMP environment. The atomic with release
2186 * semantics ensures that.
2188 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2189 secbits = sc->secret.key[secbit];
2190 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2191 atomic_add_rel_int(&sc->secret.oddeven, 1);
2193 /* Reschedule ourself. */
2194 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2198 * Exports the syncache entries to userland so that netstat can display
2199 * them alongside the other sockets. This function is intended to be
2200 * called only from tcp_pcblist.
2202 * Due to concurrency on an active system, the number of pcbs exported
2203 * may have no relation to max_pcbs. max_pcbs merely indicates the
2204 * amount of space the caller allocated for this function to use.
2207 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
2210 struct syncache *sc;
2211 struct syncache_head *sch;
2212 int count, error, i;
2214 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
2215 sch = &V_tcp_syncache.hashbase[i];
2217 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2218 if (count >= max_pcbs) {
2222 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2224 bzero(&xt, sizeof(xt));
2225 xt.xt_len = sizeof(xt);
2226 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2227 xt.xt_inp.inp_vflag = INP_IPV6;
2229 xt.xt_inp.inp_vflag = INP_IPV4;
2230 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo));
2231 xt.xt_tp.t_inpcb = &xt.xt_inp;
2232 xt.xt_tp.t_state = TCPS_SYN_RECEIVED;
2233 xt.xt_socket.xso_protocol = IPPROTO_TCP;
2234 xt.xt_socket.xso_len = sizeof (struct xsocket);
2235 xt.xt_socket.so_type = SOCK_STREAM;
2236 xt.xt_socket.so_state = SS_ISCONNECTING;
2237 error = SYSCTL_OUT(req, &xt, sizeof xt);
2247 *pcbs_exported = count;