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");
178 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_VNET | CTLFLAG_RW,
179 &VNET_NAME(tcp_syncache.rexmt_limit), 0,
180 "Limit on SYN/ACK retransmissions");
182 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
183 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
184 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
185 "Send reset on socket allocation failure");
187 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
189 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx)
190 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx)
191 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED)
194 * Requires the syncache entry to be already removed from the bucket list.
197 syncache_free(struct syncache *sc)
201 (void) m_free(sc->sc_ipopts);
205 mac_syncache_destroy(&sc->sc_label);
208 uma_zfree(V_tcp_syncache.zone, sc);
216 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
217 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
218 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
219 V_tcp_syncache.hash_secret = arc4random();
221 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
222 &V_tcp_syncache.hashsize);
223 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
224 &V_tcp_syncache.bucket_limit);
225 if (!powerof2(V_tcp_syncache.hashsize) ||
226 V_tcp_syncache.hashsize == 0) {
227 printf("WARNING: syncache hash size is not a power of 2.\n");
228 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
230 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
233 V_tcp_syncache.cache_limit =
234 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
235 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
236 &V_tcp_syncache.cache_limit);
238 /* Allocate the hash table. */
239 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
240 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
243 V_tcp_syncache.vnet = curvnet;
246 /* Initialize the hash buckets. */
247 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
248 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
249 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
251 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
252 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
253 V_tcp_syncache.hashbase[i].sch_length = 0;
254 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
255 V_tcp_syncache.hashbase[i].sch_last_overflow = INT64_MIN;
258 /* Create the syncache entry zone. */
259 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
260 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
261 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
262 V_tcp_syncache.cache_limit);
264 /* Start the SYN cookie reseeder callout. */
265 callout_init(&V_tcp_syncache.secret.reseed, 1);
266 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
267 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
268 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
269 syncookie_reseed, &V_tcp_syncache);
274 syncache_destroy(void)
276 struct syncache_head *sch;
277 struct syncache *sc, *nsc;
281 * Stop the re-seed timer before freeing resources. No need to
282 * possibly schedule it another time.
284 callout_drain(&V_tcp_syncache.secret.reseed);
286 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */
287 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
289 sch = &V_tcp_syncache.hashbase[i];
290 callout_drain(&sch->sch_timer);
293 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
294 syncache_drop(sc, sch);
296 KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
297 ("%s: sch->sch_bucket not empty", __func__));
298 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
299 __func__, sch->sch_length));
300 mtx_destroy(&sch->sch_mtx);
303 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
304 ("%s: cache_count not 0", __func__));
306 /* Free the allocated global resources. */
307 uma_zdestroy(V_tcp_syncache.zone);
308 free(V_tcp_syncache.hashbase, M_SYNCACHE);
313 * Inserts a syncache entry into the specified bucket row.
314 * Locks and unlocks the syncache_head autonomously.
317 syncache_insert(struct syncache *sc, struct syncache_head *sch)
319 struct syncache *sc2;
324 * Make sure that we don't overflow the per-bucket limit.
325 * If the bucket is full, toss the oldest element.
327 if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
328 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
329 ("sch->sch_length incorrect"));
330 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
331 sch->sch_last_overflow = time_uptime;
332 syncache_drop(sc2, sch);
333 TCPSTAT_INC(tcps_sc_bucketoverflow);
336 /* Put it into the bucket. */
337 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
341 if (ADDED_BY_TOE(sc)) {
342 struct toedev *tod = sc->sc_tod;
344 tod->tod_syncache_added(tod, sc->sc_todctx);
348 /* Reinitialize the bucket row's timer. */
349 if (sch->sch_length == 1)
350 sch->sch_nextc = ticks + INT_MAX;
351 syncache_timeout(sc, sch, 1);
355 TCPSTATES_INC(TCPS_SYN_RECEIVED);
356 TCPSTAT_INC(tcps_sc_added);
360 * Remove and free entry from syncache bucket row.
361 * Expects locked syncache head.
364 syncache_drop(struct syncache *sc, struct syncache_head *sch)
367 SCH_LOCK_ASSERT(sch);
369 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
370 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
374 if (ADDED_BY_TOE(sc)) {
375 struct toedev *tod = sc->sc_tod;
377 tod->tod_syncache_removed(tod, sc->sc_todctx);
385 * Engage/reengage time on bucket row.
388 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
390 sc->sc_rxttime = ticks +
391 TCPTV_RTOBASE * (tcp_syn_backoff[sc->sc_rxmits]);
393 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
394 sch->sch_nextc = sc->sc_rxttime;
396 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
397 syncache_timer, (void *)sch);
402 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
403 * If we have retransmitted an entry the maximum number of times, expire it.
404 * One separate timer for each bucket row.
407 syncache_timer(void *xsch)
409 struct syncache_head *sch = (struct syncache_head *)xsch;
410 struct syncache *sc, *nsc;
414 CURVNET_SET(sch->sch_sc->vnet);
416 /* NB: syncache_head has already been locked by the callout. */
417 SCH_LOCK_ASSERT(sch);
420 * In the following cycle we may remove some entries and/or
421 * advance some timeouts, so re-initialize the bucket timer.
423 sch->sch_nextc = tick + INT_MAX;
425 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
427 * We do not check if the listen socket still exists
428 * and accept the case where the listen socket may be
429 * gone by the time we resend the SYN/ACK. We do
430 * not expect this to happens often. If it does,
431 * then the RST will be sent by the time the remote
432 * host does the SYN/ACK->ACK.
434 if (TSTMP_GT(sc->sc_rxttime, tick)) {
435 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
436 sch->sch_nextc = sc->sc_rxttime;
439 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
440 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
441 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
442 "giving up and removing syncache entry\n",
446 syncache_drop(sc, sch);
447 TCPSTAT_INC(tcps_sc_stale);
450 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
451 log(LOG_DEBUG, "%s; %s: Response timeout, "
452 "retransmitting (%u) SYN|ACK\n",
453 s, __func__, sc->sc_rxmits);
457 syncache_respond(sc, sch, 1, NULL);
458 TCPSTAT_INC(tcps_sc_retransmitted);
459 syncache_timeout(sc, sch, 0);
461 if (!TAILQ_EMPTY(&(sch)->sch_bucket))
462 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
463 syncache_timer, (void *)(sch));
468 * Find an entry in the syncache.
469 * Returns always with locked syncache_head plus a matching entry or NULL.
471 static struct syncache *
472 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
475 struct syncache_head *sch;
479 * The hash is built on foreign port + local port + foreign address.
480 * We rely on the fact that struct in_conninfo starts with 16 bits
481 * of foreign port, then 16 bits of local port then followed by 128
482 * bits of foreign address. In case of IPv4 address, the first 3
483 * 32-bit words of the address always are zeroes.
485 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
486 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
488 sch = &V_tcp_syncache.hashbase[hash];
492 /* Circle through bucket row to find matching entry. */
493 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
494 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
495 sizeof(struct in_endpoints)) == 0)
498 return (sc); /* Always returns with locked sch. */
502 * This function is called when we get a RST for a
503 * non-existent connection, so that we can see if the
504 * connection is in the syn cache. If it is, zap it.
507 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
510 struct syncache_head *sch;
513 sc = syncache_lookup(inc, &sch); /* returns locked sch */
514 SCH_LOCK_ASSERT(sch);
517 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
518 * See RFC 793 page 65, section SEGMENT ARRIVES.
520 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
521 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
522 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
523 "FIN flag set, segment ignored\n", s, __func__);
524 TCPSTAT_INC(tcps_badrst);
529 * No corresponding connection was found in syncache.
530 * If syncookies are enabled and possibly exclusively
531 * used, or we are under memory pressure, a valid RST
532 * may not find a syncache entry. In that case we're
533 * done and no SYN|ACK retransmissions will happen.
534 * Otherwise the RST was misdirected or spoofed.
537 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
538 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
539 "syncache entry (possibly syncookie only), "
540 "segment ignored\n", s, __func__);
541 TCPSTAT_INC(tcps_badrst);
546 * If the RST bit is set, check the sequence number to see
547 * if this is a valid reset segment.
549 * In all states except SYN-SENT, all reset (RST) segments
550 * are validated by checking their SEQ-fields. A reset is
551 * valid if its sequence number is in the window.
553 * The sequence number in the reset segment is normally an
554 * echo of our outgoing acknowlegement numbers, but some hosts
555 * send a reset with the sequence number at the rightmost edge
556 * of our receive window, and we have to handle this case.
558 if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
559 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
560 syncache_drop(sc, sch);
561 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
562 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, "
563 "connection attempt aborted by remote endpoint\n",
565 TCPSTAT_INC(tcps_sc_reset);
567 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
568 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
569 "IRS %u (+WND %u), segment ignored\n",
570 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd);
571 TCPSTAT_INC(tcps_badrst);
581 syncache_badack(struct in_conninfo *inc)
584 struct syncache_head *sch;
586 sc = syncache_lookup(inc, &sch); /* returns locked sch */
587 SCH_LOCK_ASSERT(sch);
589 syncache_drop(sc, sch);
590 TCPSTAT_INC(tcps_sc_badack);
596 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th)
599 struct syncache_head *sch;
601 sc = syncache_lookup(inc, &sch); /* returns locked sch */
602 SCH_LOCK_ASSERT(sch);
606 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
607 if (ntohl(th->th_seq) != sc->sc_iss)
611 * If we've rertransmitted 3 times and this is our second error,
612 * we remove the entry. Otherwise, we allow it to continue on.
613 * This prevents us from incorrectly nuking an entry during a
614 * spurious network outage.
618 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
619 sc->sc_flags |= SCF_UNREACH;
622 syncache_drop(sc, sch);
623 TCPSTAT_INC(tcps_sc_unreach);
629 * Build a new TCP socket structure from a syncache entry.
631 * On success return the newly created socket with its underlying inp locked.
633 static struct socket *
634 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
636 struct tcp_function_block *blk;
637 struct inpcb *inp = NULL;
643 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
646 * Ok, create the full blown connection, and set things up
647 * as they would have been set up if we had created the
648 * connection when the SYN arrived. If we can't create
649 * the connection, abort it.
651 so = sonewconn(lso, 0);
654 * Drop the connection; we will either send a RST or
655 * have the peer retransmit its SYN again after its
658 TCPSTAT_INC(tcps_listendrop);
659 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
660 log(LOG_DEBUG, "%s; %s: Socket create failed "
661 "due to limits or memory shortage\n",
668 mac_socketpeer_set_from_mbuf(m, so);
672 inp->inp_inc.inc_fibnum = so->so_fibnum;
675 * Exclusive pcbinfo lock is not required in syncache socket case even
676 * if two inpcb locks can be acquired simultaneously:
677 * - the inpcb in LISTEN state,
678 * - the newly created inp.
680 * In this case, an inp cannot be at same time in LISTEN state and
681 * just created by an accept() call.
683 INP_HASH_WLOCK(&V_tcbinfo);
685 /* Insert new socket into PCB hash list. */
686 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
688 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
689 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
691 inp->inp_vflag &= ~INP_IPV6;
692 inp->inp_vflag |= INP_IPV4;
694 inp->inp_laddr = sc->sc_inc.inc_laddr;
700 * If there's an mbuf and it has a flowid, then let's initialise the
701 * inp with that particular flowid.
703 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
704 inp->inp_flowid = m->m_pkthdr.flowid;
705 inp->inp_flowtype = M_HASHTYPE_GET(m);
709 * Install in the reservation hash table for now, but don't yet
710 * install a connection group since the full 4-tuple isn't yet
713 inp->inp_lport = sc->sc_inc.inc_lport;
714 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
716 * Undo the assignments above if we failed to
717 * put the PCB on the hash lists.
720 if (sc->sc_inc.inc_flags & INC_ISIPV6)
721 inp->in6p_laddr = in6addr_any;
724 inp->inp_laddr.s_addr = INADDR_ANY;
726 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
727 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
732 INP_HASH_WUNLOCK(&V_tcbinfo);
736 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
737 struct inpcb *oinp = sotoinpcb(lso);
738 struct in6_addr laddr6;
739 struct sockaddr_in6 sin6;
741 * Inherit socket options from the listening socket.
742 * Note that in6p_inputopts are not (and should not be)
743 * copied, since it stores previously received options and is
744 * used to detect if each new option is different than the
745 * previous one and hence should be passed to a user.
746 * If we copied in6p_inputopts, a user would not be able to
747 * receive options just after calling the accept system call.
749 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
750 if (oinp->in6p_outputopts)
751 inp->in6p_outputopts =
752 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
754 sin6.sin6_family = AF_INET6;
755 sin6.sin6_len = sizeof(sin6);
756 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
757 sin6.sin6_port = sc->sc_inc.inc_fport;
758 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
759 laddr6 = inp->in6p_laddr;
760 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
761 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
762 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
763 thread0.td_ucred, m)) != 0) {
764 inp->in6p_laddr = laddr6;
765 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
766 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
771 INP_HASH_WUNLOCK(&V_tcbinfo);
774 /* Override flowlabel from in6_pcbconnect. */
775 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
776 inp->inp_flow |= sc->sc_flowlabel;
779 #if defined(INET) && defined(INET6)
784 struct in_addr laddr;
785 struct sockaddr_in sin;
787 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
789 if (inp->inp_options == NULL) {
790 inp->inp_options = sc->sc_ipopts;
791 sc->sc_ipopts = NULL;
794 sin.sin_family = AF_INET;
795 sin.sin_len = sizeof(sin);
796 sin.sin_addr = sc->sc_inc.inc_faddr;
797 sin.sin_port = sc->sc_inc.inc_fport;
798 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
799 laddr = inp->inp_laddr;
800 if (inp->inp_laddr.s_addr == INADDR_ANY)
801 inp->inp_laddr = sc->sc_inc.inc_laddr;
802 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
803 thread0.td_ucred, m)) != 0) {
804 inp->inp_laddr = laddr;
805 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
806 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
811 INP_HASH_WUNLOCK(&V_tcbinfo);
816 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
817 /* Copy old policy into new socket's. */
818 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
819 printf("syncache_socket: could not copy policy\n");
821 INP_HASH_WUNLOCK(&V_tcbinfo);
823 tcp_state_change(tp, TCPS_SYN_RECEIVED);
824 tp->iss = sc->sc_iss;
825 tp->irs = sc->sc_irs;
828 blk = sototcpcb(lso)->t_fb;
829 if (blk != tp->t_fb) {
831 * Our parents t_fb was not the default,
832 * we need to release our ref on tp->t_fb and
833 * pickup one on the new entry.
835 struct tcp_function_block *rblk;
837 rblk = find_and_ref_tcp_fb(blk);
838 KASSERT(rblk != NULL,
839 ("cannot find blk %p out of syncache?", blk));
840 if (tp->t_fb->tfb_tcp_fb_fini)
841 (*tp->t_fb->tfb_tcp_fb_fini)(tp);
842 refcount_release(&tp->t_fb->tfb_refcnt);
844 if (tp->t_fb->tfb_tcp_fb_init) {
845 (*tp->t_fb->tfb_tcp_fb_init)(tp);
848 tp->snd_wl1 = sc->sc_irs;
849 tp->snd_max = tp->iss + 1;
850 tp->snd_nxt = tp->iss + 1;
851 tp->rcv_up = sc->sc_irs + 1;
852 tp->rcv_wnd = sc->sc_wnd;
853 tp->rcv_adv += tp->rcv_wnd;
854 tp->last_ack_sent = tp->rcv_nxt;
856 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
857 if (sc->sc_flags & SCF_NOOPT)
858 tp->t_flags |= TF_NOOPT;
860 if (sc->sc_flags & SCF_WINSCALE) {
861 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
862 tp->snd_scale = sc->sc_requested_s_scale;
863 tp->request_r_scale = sc->sc_requested_r_scale;
865 if (sc->sc_flags & SCF_TIMESTAMP) {
866 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
867 tp->ts_recent = sc->sc_tsreflect;
868 tp->ts_recent_age = tcp_ts_getticks();
869 tp->ts_offset = sc->sc_tsoff;
871 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
872 if (sc->sc_flags & SCF_SIGNATURE)
873 tp->t_flags |= TF_SIGNATURE;
875 if (sc->sc_flags & SCF_SACK)
876 tp->t_flags |= TF_SACK_PERMIT;
879 if (sc->sc_flags & SCF_ECN)
880 tp->t_flags |= TF_ECN_PERMIT;
883 * Set up MSS and get cached values from tcp_hostcache.
884 * This might overwrite some of the defaults we just set.
886 tcp_mss(tp, sc->sc_peer_mss);
889 * If the SYN,ACK was retransmitted, indicate that CWND to be
890 * limited to one segment in cc_conn_init().
891 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
893 if (sc->sc_rxmits > 1)
898 * Allow a TOE driver to install its hooks. Note that we hold the
899 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
900 * new connection before the TOE driver has done its thing.
902 if (ADDED_BY_TOE(sc)) {
903 struct toedev *tod = sc->sc_tod;
905 tod->tod_offload_socket(tod, sc->sc_todctx, so);
909 * Copy and activate timers.
911 tp->t_keepinit = sototcpcb(lso)->t_keepinit;
912 tp->t_keepidle = sototcpcb(lso)->t_keepidle;
913 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
914 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
915 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
917 TCPSTAT_INC(tcps_accepts);
929 * This function gets called when we receive an ACK for a
930 * socket in the LISTEN state. We look up the connection
931 * in the syncache, and if its there, we pull it out of
932 * the cache and turn it into a full-blown connection in
933 * the SYN-RECEIVED state.
935 * On syncache_socket() success the newly created socket
936 * has its underlying inp locked.
939 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
940 struct socket **lsop, struct mbuf *m)
943 struct syncache_head *sch;
948 * Global TCP locks are held because we manipulate the PCB lists
949 * and create a new socket.
951 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
952 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
953 ("%s: can handle only ACK", __func__));
955 sc = syncache_lookup(inc, &sch); /* returns locked sch */
956 SCH_LOCK_ASSERT(sch);
960 * Test code for syncookies comparing the syncache stored
961 * values with the reconstructed values from the cookie.
964 syncookie_cmp(inc, sch, sc, th, to, *lsop);
969 * There is no syncache entry, so see if this ACK is
970 * a returning syncookie. To do this, first:
971 * A. Check if syncookies are used in case of syncache
973 * B. See if this socket has had a syncache entry dropped in
974 * the recent past. We don't want to accept a bogus
975 * syncookie if we've never received a SYN or accept it
977 * C. check that the syncookie is valid. If it is, then
978 * cobble up a fake syncache entry, and return.
980 if (!V_tcp_syncookies) {
982 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
983 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
984 "segment rejected (syncookies disabled)\n",
988 if (!V_tcp_syncookiesonly &&
989 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
991 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
992 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
993 "segment rejected (no syncache entry)\n",
997 bzero(&scs, sizeof(scs));
998 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
1001 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1002 log(LOG_DEBUG, "%s; %s: Segment failed "
1003 "SYNCOOKIE authentication, segment rejected "
1004 "(probably spoofed)\n", s, __func__);
1007 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1008 /* If received ACK has MD5 signature, check it. */
1009 if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1010 (!TCPMD5_ENABLED() ||
1011 TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1013 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1014 log(LOG_DEBUG, "%s; %s: Segment rejected, "
1015 "MD5 signature doesn't match.\n",
1019 TCPSTAT_INC(tcps_sig_err_sigopt);
1020 return (-1); /* Do not send RST */
1022 #endif /* TCP_SIGNATURE */
1024 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1026 * If listening socket requested TCP digests, check that
1027 * received ACK has signature and it is correct.
1028 * If not, drop the ACK and leave sc entry in th cache,
1029 * because SYN was received with correct signature.
1031 if (sc->sc_flags & SCF_SIGNATURE) {
1032 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1034 TCPSTAT_INC(tcps_sig_err_nosigopt);
1036 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1037 log(LOG_DEBUG, "%s; %s: Segment "
1038 "rejected, MD5 signature wasn't "
1039 "provided.\n", s, __func__);
1042 return (-1); /* Do not send RST */
1044 if (!TCPMD5_ENABLED() ||
1045 TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1046 /* Doesn't match or no SA */
1048 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1049 log(LOG_DEBUG, "%s; %s: Segment "
1050 "rejected, MD5 signature doesn't "
1051 "match.\n", s, __func__);
1054 return (-1); /* Do not send RST */
1057 #endif /* TCP_SIGNATURE */
1059 * Pull out the entry to unlock the bucket row.
1061 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1062 * tcp_state_change(). The tcpcb is not existent at this
1063 * moment. A new one will be allocated via syncache_socket->
1064 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1065 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1067 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1068 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1071 if (ADDED_BY_TOE(sc)) {
1072 struct toedev *tod = sc->sc_tod;
1074 tod->tod_syncache_removed(tod, sc->sc_todctx);
1081 * Segment validation:
1082 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1084 if (th->th_ack != sc->sc_iss + 1) {
1085 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1086 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1087 "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1092 * The SEQ must fall in the window starting at the received
1093 * initial receive sequence number + 1 (the SYN).
1095 if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1096 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1097 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1098 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1099 "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1104 * If timestamps were not negotiated during SYN/ACK they
1105 * must not appear on any segment during this session.
1107 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1108 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1109 log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1110 "segment rejected\n", s, __func__);
1115 * If timestamps were negotiated during SYN/ACK they should
1116 * appear on every segment during this session.
1117 * XXXAO: This is only informal as there have been unverified
1118 * reports of non-compliants stacks.
1120 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1121 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1122 log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1123 "no action\n", s, __func__);
1130 * If timestamps were negotiated, the reflected timestamp
1131 * must be equal to what we actually sent in the SYN|ACK
1132 * except in the case of 0. Some boxes are known for sending
1133 * broken timestamp replies during the 3whs (and potentially
1134 * during the connection also).
1136 * Accept the final ACK of 3whs with reflected timestamp of 0
1137 * instead of sending a RST and deleting the syncache entry.
1139 if ((to->to_flags & TOF_TS) && to->to_tsecr &&
1140 to->to_tsecr != sc->sc_ts) {
1141 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1142 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, "
1143 "segment rejected\n",
1144 s, __func__, to->to_tsecr, sc->sc_ts);
1148 *lsop = syncache_socket(sc, *lsop, m);
1151 TCPSTAT_INC(tcps_sc_aborted);
1153 TCPSTAT_INC(tcps_sc_completed);
1155 /* how do we find the inp for the new socket? */
1160 if (sc != NULL && sc != &scs)
1170 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
1171 uint64_t response_cookie)
1175 unsigned int *pending_counter;
1178 * Global TCP locks are held because we manipulate the PCB lists
1179 * and create a new socket.
1181 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1183 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
1184 *lsop = syncache_socket(sc, *lsop, m);
1185 if (*lsop == NULL) {
1186 TCPSTAT_INC(tcps_sc_aborted);
1187 atomic_subtract_int(pending_counter, 1);
1189 inp = sotoinpcb(*lsop);
1190 tp = intotcpcb(inp);
1191 tp->t_flags |= TF_FASTOPEN;
1192 tp->t_tfo_cookie = response_cookie;
1193 tp->snd_max = tp->iss;
1194 tp->snd_nxt = tp->iss;
1195 tp->t_tfo_pending = pending_counter;
1196 TCPSTAT_INC(tcps_sc_completed);
1199 #endif /* TCP_RFC7413 */
1202 * Given a LISTEN socket and an inbound SYN request, add
1203 * this to the syn cache, and send back a segment:
1204 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1207 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1208 * Doing so would require that we hold onto the data and deliver it
1209 * to the application. However, if we are the target of a SYN-flood
1210 * DoS attack, an attacker could send data which would eventually
1211 * consume all available buffer space if it were ACKed. By not ACKing
1212 * the data, we avoid this DoS scenario.
1214 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1215 * cookie is processed, V_tcp_fastopen_enabled set to true, and the
1216 * TCP_FASTOPEN socket option is set. In this case, a new socket is created
1217 * and returned via lsop, the mbuf is not freed so that tcp_input() can
1218 * queue its data to the socket, and 1 is returned to indicate the
1219 * TFO-socket-creation path was taken.
1222 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1223 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1228 struct syncache *sc = NULL;
1229 struct syncache_head *sch;
1230 struct mbuf *ipopts = NULL;
1232 int win, sb_hiwat, ip_ttl, ip_tos;
1236 int autoflowlabel = 0;
1239 struct label *maclabel;
1241 struct syncache scs;
1244 uint64_t tfo_response_cookie;
1245 int tfo_cookie_valid = 0;
1246 int tfo_response_cookie_valid = 0;
1249 INP_WLOCK_ASSERT(inp); /* listen socket */
1250 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1251 ("%s: unexpected tcp flags", __func__));
1254 * Combine all so/tp operations very early to drop the INP lock as
1259 cred = crhold(so->so_cred);
1262 if ((inc->inc_flags & INC_ISIPV6) &&
1263 (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1266 ip_ttl = inp->inp_ip_ttl;
1267 ip_tos = inp->inp_ip_tos;
1268 win = sbspace(&so->so_rcv);
1269 sb_hiwat = so->so_rcv.sb_hiwat;
1270 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1273 if (V_tcp_fastopen_enabled && (tp->t_flags & TF_FASTOPEN) &&
1274 (tp->t_tfo_pending != NULL) && (to->to_flags & TOF_FASTOPEN)) {
1276 * Limit the number of pending TFO connections to
1277 * approximately half of the queue limit. This prevents TFO
1278 * SYN floods from starving the service by filling the
1279 * listen queue with bogus TFO connections.
1281 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1282 (so->so_qlimit / 2)) {
1285 result = tcp_fastopen_check_cookie(inc,
1286 to->to_tfo_cookie, to->to_tfo_len,
1287 &tfo_response_cookie);
1288 tfo_cookie_valid = (result > 0);
1289 tfo_response_cookie_valid = (result >= 0);
1291 atomic_subtract_int(tp->t_tfo_pending, 1);
1295 /* By the time we drop the lock these should no longer be used. */
1300 if (mac_syncache_init(&maclabel) != 0) {
1304 mac_syncache_create(maclabel, inp);
1307 if (!tfo_cookie_valid)
1312 * Remember the IP options, if any.
1315 if (!(inc->inc_flags & INC_ISIPV6))
1318 ipopts = (m) ? ip_srcroute(m) : NULL;
1323 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1325 * If listening socket requested TCP digests, check that received
1326 * SYN has signature and it is correct. If signature doesn't match
1327 * or TCP_SIGNATURE support isn't enabled, drop the packet.
1329 if (ltflags & TF_SIGNATURE) {
1330 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1331 TCPSTAT_INC(tcps_sig_err_nosigopt);
1334 if (!TCPMD5_ENABLED() ||
1335 TCPMD5_INPUT(m, th, to->to_signature) != 0)
1338 #endif /* TCP_SIGNATURE */
1340 * See if we already have an entry for this connection.
1341 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1343 * XXX: should the syncache be re-initialized with the contents
1344 * of the new SYN here (which may have different options?)
1346 * XXX: We do not check the sequence number to see if this is a
1347 * real retransmit or a new connection attempt. The question is
1348 * how to handle such a case; either ignore it as spoofed, or
1349 * drop the current entry and create a new one?
1351 sc = syncache_lookup(inc, &sch); /* returns locked entry */
1352 SCH_LOCK_ASSERT(sch);
1355 if (tfo_cookie_valid)
1358 TCPSTAT_INC(tcps_sc_dupsyn);
1361 * If we were remembering a previous source route,
1362 * forget it and use the new one we've been given.
1365 (void) m_free(sc->sc_ipopts);
1366 sc->sc_ipopts = ipopts;
1369 * Update timestamp if present.
1371 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1372 sc->sc_tsreflect = to->to_tsval;
1374 sc->sc_flags &= ~SCF_TIMESTAMP;
1377 * Since we have already unconditionally allocated label
1378 * storage, free it up. The syncache entry will already
1379 * have an initialized label we can use.
1381 mac_syncache_destroy(&maclabel);
1383 /* Retransmit SYN|ACK and reset retransmit count. */
1384 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1385 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1386 "resetting timer and retransmitting SYN|ACK\n",
1390 if (syncache_respond(sc, sch, 1, m) == 0) {
1392 syncache_timeout(sc, sch, 1);
1393 TCPSTAT_INC(tcps_sndacks);
1394 TCPSTAT_INC(tcps_sndtotal);
1401 if (tfo_cookie_valid) {
1402 bzero(&scs, sizeof(scs));
1408 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1411 * The zone allocator couldn't provide more entries.
1412 * Treat this as if the cache was full; drop the oldest
1413 * entry and insert the new one.
1415 TCPSTAT_INC(tcps_sc_zonefail);
1416 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1417 sch->sch_last_overflow = time_uptime;
1418 syncache_drop(sc, sch);
1420 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1422 if (V_tcp_syncookies) {
1423 bzero(&scs, sizeof(scs));
1428 (void) m_free(ipopts);
1436 if (!tfo_cookie_valid && tfo_response_cookie_valid)
1437 sc->sc_tfo_cookie = &tfo_response_cookie;
1441 * Fill in the syncache values.
1444 sc->sc_label = maclabel;
1448 sc->sc_ipopts = ipopts;
1449 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1451 if (!(inc->inc_flags & INC_ISIPV6))
1454 sc->sc_ip_tos = ip_tos;
1455 sc->sc_ip_ttl = ip_ttl;
1459 sc->sc_todctx = todctx;
1461 sc->sc_irs = th->th_seq;
1462 sc->sc_iss = arc4random();
1464 sc->sc_flowlabel = 0;
1467 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1468 * win was derived from socket earlier in the function.
1471 win = imin(win, TCP_MAXWIN);
1474 if (V_tcp_do_rfc1323) {
1476 * A timestamp received in a SYN makes
1477 * it ok to send timestamp requests and replies.
1479 if (to->to_flags & TOF_TS) {
1480 sc->sc_tsreflect = to->to_tsval;
1481 sc->sc_ts = tcp_ts_getticks();
1482 sc->sc_flags |= SCF_TIMESTAMP;
1484 if (to->to_flags & TOF_SCALE) {
1488 * Pick the smallest possible scaling factor that
1489 * will still allow us to scale up to sb_max, aka
1490 * kern.ipc.maxsockbuf.
1492 * We do this because there are broken firewalls that
1493 * will corrupt the window scale option, leading to
1494 * the other endpoint believing that our advertised
1495 * window is unscaled. At scale factors larger than
1496 * 5 the unscaled window will drop below 1500 bytes,
1497 * leading to serious problems when traversing these
1500 * With the default maxsockbuf of 256K, a scale factor
1501 * of 3 will be chosen by this algorithm. Those who
1502 * choose a larger maxsockbuf should watch out
1503 * for the compatibility problems mentioned above.
1505 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1506 * or <SYN,ACK>) segment itself is never scaled.
1508 while (wscale < TCP_MAX_WINSHIFT &&
1509 (TCP_MAXWIN << wscale) < sb_max)
1511 sc->sc_requested_r_scale = wscale;
1512 sc->sc_requested_s_scale = to->to_wscale;
1513 sc->sc_flags |= SCF_WINSCALE;
1516 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1518 * If listening socket requested TCP digests, flag this in the
1519 * syncache so that syncache_respond() will do the right thing
1522 if (ltflags & TF_SIGNATURE)
1523 sc->sc_flags |= SCF_SIGNATURE;
1524 #endif /* TCP_SIGNATURE */
1525 if (to->to_flags & TOF_SACKPERM)
1526 sc->sc_flags |= SCF_SACK;
1527 if (to->to_flags & TOF_MSS)
1528 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */
1529 if (ltflags & TF_NOOPT)
1530 sc->sc_flags |= SCF_NOOPT;
1531 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1532 sc->sc_flags |= SCF_ECN;
1534 if (V_tcp_syncookies)
1535 sc->sc_iss = syncookie_generate(sch, sc);
1537 if (autoflowlabel) {
1538 if (V_tcp_syncookies)
1539 sc->sc_flowlabel = sc->sc_iss;
1541 sc->sc_flowlabel = ip6_randomflowlabel();
1542 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1548 if (tfo_cookie_valid) {
1549 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie);
1550 /* INP_WUNLOCK(inp) will be performed by the called */
1557 * Do a standard 3-way handshake.
1559 if (syncache_respond(sc, sch, 0, m) == 0) {
1560 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1562 else if (sc != &scs)
1563 syncache_insert(sc, sch); /* locks and unlocks sch */
1564 TCPSTAT_INC(tcps_sndacks);
1565 TCPSTAT_INC(tcps_sndtotal);
1569 TCPSTAT_INC(tcps_sc_dropped);
1584 mac_syncache_destroy(&maclabel);
1590 * Send SYN|ACK to the peer. Either in response to the peer's SYN,
1591 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1594 syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked,
1595 const struct mbuf *m0)
1597 struct ip *ip = NULL;
1599 struct tcphdr *th = NULL;
1600 int optlen, error = 0; /* Make compiler happy */
1601 u_int16_t hlen, tlen, mssopt;
1604 struct ip6_hdr *ip6 = NULL;
1608 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1611 tlen = hlen + sizeof(struct tcphdr);
1613 /* Determine MSS we advertize to other end of connection. */
1614 mssopt = tcp_mssopt(&sc->sc_inc);
1615 if (sc->sc_peer_mss)
1616 mssopt = max( min(sc->sc_peer_mss, mssopt), V_tcp_minmss);
1618 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1619 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1620 ("syncache: mbuf too small"));
1622 /* Create the IP+TCP header from scratch. */
1623 m = m_gethdr(M_NOWAIT, MT_DATA);
1627 mac_syncache_create_mbuf(sc->sc_label, m);
1629 m->m_data += max_linkhdr;
1631 m->m_pkthdr.len = tlen;
1632 m->m_pkthdr.rcvif = NULL;
1635 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1636 ip6 = mtod(m, struct ip6_hdr *);
1637 ip6->ip6_vfc = IPV6_VERSION;
1638 ip6->ip6_nxt = IPPROTO_TCP;
1639 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1640 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1641 ip6->ip6_plen = htons(tlen - hlen);
1642 /* ip6_hlim is set after checksum */
1643 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1644 ip6->ip6_flow |= sc->sc_flowlabel;
1646 th = (struct tcphdr *)(ip6 + 1);
1649 #if defined(INET6) && defined(INET)
1654 ip = mtod(m, struct ip *);
1655 ip->ip_v = IPVERSION;
1656 ip->ip_hl = sizeof(struct ip) >> 2;
1657 ip->ip_len = htons(tlen);
1661 ip->ip_p = IPPROTO_TCP;
1662 ip->ip_src = sc->sc_inc.inc_laddr;
1663 ip->ip_dst = sc->sc_inc.inc_faddr;
1664 ip->ip_ttl = sc->sc_ip_ttl;
1665 ip->ip_tos = sc->sc_ip_tos;
1668 * See if we should do MTU discovery. Route lookups are
1669 * expensive, so we will only unset the DF bit if:
1671 * 1) path_mtu_discovery is disabled
1672 * 2) the SCF_UNREACH flag has been set
1674 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1675 ip->ip_off |= htons(IP_DF);
1677 th = (struct tcphdr *)(ip + 1);
1680 th->th_sport = sc->sc_inc.inc_lport;
1681 th->th_dport = sc->sc_inc.inc_fport;
1683 th->th_seq = htonl(sc->sc_iss);
1684 th->th_ack = htonl(sc->sc_irs + 1);
1685 th->th_off = sizeof(struct tcphdr) >> 2;
1687 th->th_flags = TH_SYN|TH_ACK;
1688 th->th_win = htons(sc->sc_wnd);
1691 if (sc->sc_flags & SCF_ECN) {
1692 th->th_flags |= TH_ECE;
1693 TCPSTAT_INC(tcps_ecn_shs);
1696 /* Tack on the TCP options. */
1697 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1701 to.to_flags = TOF_MSS;
1702 if (sc->sc_flags & SCF_WINSCALE) {
1703 to.to_wscale = sc->sc_requested_r_scale;
1704 to.to_flags |= TOF_SCALE;
1706 if (sc->sc_flags & SCF_TIMESTAMP) {
1707 /* Virgin timestamp or TCP cookie enhanced one. */
1708 to.to_tsval = sc->sc_ts;
1709 to.to_tsecr = sc->sc_tsreflect;
1710 to.to_flags |= TOF_TS;
1712 if (sc->sc_flags & SCF_SACK)
1713 to.to_flags |= TOF_SACKPERM;
1714 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1715 if (sc->sc_flags & SCF_SIGNATURE)
1716 to.to_flags |= TOF_SIGNATURE;
1719 if (sc->sc_tfo_cookie) {
1720 to.to_flags |= TOF_FASTOPEN;
1721 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1722 to.to_tfo_cookie = sc->sc_tfo_cookie;
1723 /* don't send cookie again when retransmitting response */
1724 sc->sc_tfo_cookie = NULL;
1727 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1729 /* Adjust headers by option size. */
1730 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1732 m->m_pkthdr.len += optlen;
1734 if (sc->sc_inc.inc_flags & INC_ISIPV6)
1735 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1738 ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1739 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1740 if (sc->sc_flags & SCF_SIGNATURE) {
1741 KASSERT(to.to_flags & TOF_SIGNATURE,
1742 ("tcp_addoptions() didn't set tcp_signature"));
1744 /* NOTE: to.to_signature is inside of mbuf */
1745 if (!TCPMD5_ENABLED() ||
1746 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1755 M_SETFIB(m, sc->sc_inc.inc_fibnum);
1756 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1758 * If we have peer's SYN and it has a flowid, then let's assign it to
1759 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid
1760 * to SYN|ACK due to lack of inp here.
1762 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1763 m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1764 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1767 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1768 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1769 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1771 ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1773 if (ADDED_BY_TOE(sc)) {
1774 struct toedev *tod = sc->sc_tod;
1776 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1781 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1784 #if defined(INET6) && defined(INET)
1789 m->m_pkthdr.csum_flags = CSUM_TCP;
1790 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1791 htons(tlen + optlen - hlen + IPPROTO_TCP));
1793 if (ADDED_BY_TOE(sc)) {
1794 struct toedev *tod = sc->sc_tod;
1796 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1801 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1808 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1809 * that exceed the capacity of the syncache by avoiding the storage of any
1810 * of the SYNs we receive. Syncookies defend against blind SYN flooding
1811 * attacks where the attacker does not have access to our responses.
1813 * Syncookies encode and include all necessary information about the
1814 * connection setup within the SYN|ACK that we send back. That way we
1815 * can avoid keeping any local state until the ACK to our SYN|ACK returns
1816 * (if ever). Normally the syncache and syncookies are running in parallel
1817 * with the latter taking over when the former is exhausted. When matching
1818 * syncache entry is found the syncookie is ignored.
1820 * The only reliable information persisting the 3WHS is our initial sequence
1821 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient
1822 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1823 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK
1824 * returns and signifies a legitimate connection if it matches the ACK.
1826 * The available space of 32 bits to store the hash and to encode the SYN
1827 * option information is very tight and we should have at least 24 bits for
1828 * the MAC to keep the number of guesses by blind spoofing reasonably high.
1830 * SYN option information we have to encode to fully restore a connection:
1831 * MSS: is imporant to chose an optimal segment size to avoid IP level
1832 * fragmentation along the path. The common MSS values can be encoded
1833 * in a 3-bit table. Uncommon values are captured by the next lower value
1834 * in the table leading to a slight increase in packetization overhead.
1835 * WSCALE: is necessary to allow large windows to be used for high delay-
1836 * bandwidth product links. Not scaling the window when it was initially
1837 * negotiated is bad for performance as lack of scaling further decreases
1838 * the apparent available send window. We only need to encode the WSCALE
1839 * we received from the remote end. Our end can be recalculated at any
1840 * time. The common WSCALE values can be encoded in a 3-bit table.
1841 * Uncommon values are captured by the next lower value in the table
1842 * making us under-estimate the available window size halving our
1843 * theoretically possible maximum throughput for that connection.
1844 * SACK: Greatly assists in packet loss recovery and requires 1 bit.
1845 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
1846 * that are included in all segments on a connection. We enable them when
1849 * Security of syncookies and attack vectors:
1851 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
1852 * together with the gloabl secret to make it unique per connection attempt.
1853 * Thus any change of any of those parameters results in a different MAC output
1854 * in an unpredictable way unless a collision is encountered. 24 bits of the
1855 * MAC are embedded into the ISS.
1857 * To prevent replay attacks two rotating global secrets are updated with a
1858 * new random value every 15 seconds. The life-time of a syncookie is thus
1861 * Vector 1: Attacking the secret. This requires finding a weakness in the
1862 * MAC itself or the way it is used here. The attacker can do a chosen plain
1863 * text attack by varying and testing the all parameters under his control.
1864 * The strength depends on the size and randomness of the secret, and the
1865 * cryptographic security of the MAC function. Due to the constant updating
1866 * of the secret the attacker has at most 29.999 seconds to find the secret
1867 * and launch spoofed connections. After that he has to start all over again.
1869 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC
1870 * size an average of 4,823 attempts are required for a 50% chance of success
1871 * to spoof a single syncookie (birthday collision paradox). However the
1872 * attacker is blind and doesn't know if one of his attempts succeeded unless
1873 * he has a side channel to interfere success from. A single connection setup
1874 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
1875 * This many attempts are required for each one blind spoofed connection. For
1876 * every additional spoofed connection he has to launch another N attempts.
1877 * Thus for a sustained rate 100 spoofed connections per second approximately
1878 * 1,800,000 packets per second would have to be sent.
1880 * NB: The MAC function should be fast so that it doesn't become a CPU
1881 * exhaustion attack vector itself.
1884 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations
1885 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
1886 * http://cr.yp.to/syncookies.html (overview)
1887 * http://cr.yp.to/syncookies/archive (details)
1890 * Schematic construction of a syncookie enabled Initial Sequence Number:
1892 * 12345678901234567890123456789012
1893 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
1895 * x 24 MAC (truncated)
1896 * W 3 Send Window Scale index
1898 * S 1 SACK permitted
1899 * P 1 Odd/even secret
1903 * Distribution and probability of certain MSS values. Those in between are
1904 * rounded down to the next lower one.
1905 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
1906 * .2% .3% 5% 7% 7% 20% 15% 45%
1908 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
1911 * Distribution and probability of certain WSCALE values. We have to map the
1912 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
1913 * bits based on prevalence of certain values. Where we don't have an exact
1914 * match for are rounded down to the next lower one letting us under-estimate
1915 * the true available window. At the moment this would happen only for the
1916 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
1917 * and window size). The absence of the WSCALE option (no scaling in either
1918 * direction) is encoded with index zero.
1919 * [WSCALE values histograms, Allman, 2012]
1920 * X 10 10 35 5 6 14 10% by host
1921 * X 11 4 5 5 18 49 3% by connections
1923 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
1926 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed
1927 * and good cryptographic properties.
1930 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
1931 uint8_t *secbits, uintptr_t secmod)
1934 uint32_t siphash[2];
1936 SipHash24_Init(&ctx);
1937 SipHash_SetKey(&ctx, secbits);
1938 switch (inc->inc_flags & INC_ISIPV6) {
1941 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
1942 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
1947 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
1948 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
1952 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
1953 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
1954 SipHash_Update(&ctx, &irs, sizeof(irs));
1955 SipHash_Update(&ctx, &flags, sizeof(flags));
1956 SipHash_Update(&ctx, &secmod, sizeof(secmod));
1957 SipHash_Final((u_int8_t *)&siphash, &ctx);
1959 return (siphash[0] ^ siphash[1]);
1963 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
1965 u_int i, mss, secbit, wscale;
1968 union syncookie cookie;
1970 SCH_LOCK_ASSERT(sch);
1974 /* Map our computed MSS into the 3-bit index. */
1975 mss = min(tcp_mssopt(&sc->sc_inc), max(sc->sc_peer_mss, V_tcp_minmss));
1976 for (i = nitems(tcp_sc_msstab) - 1; tcp_sc_msstab[i] > mss && i > 0;
1979 cookie.flags.mss_idx = i;
1982 * Map the send window scale into the 3-bit index but only if
1983 * the wscale option was received.
1985 if (sc->sc_flags & SCF_WINSCALE) {
1986 wscale = sc->sc_requested_s_scale;
1987 for (i = nitems(tcp_sc_wstab) - 1;
1988 tcp_sc_wstab[i] > wscale && i > 0;
1991 cookie.flags.wscale_idx = i;
1994 /* Can we do SACK? */
1995 if (sc->sc_flags & SCF_SACK)
1996 cookie.flags.sack_ok = 1;
1998 /* Which of the two secrets to use. */
1999 secbit = sch->sch_sc->secret.oddeven & 0x1;
2000 cookie.flags.odd_even = secbit;
2002 secbits = sch->sch_sc->secret.key[secbit];
2003 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2007 * Put the flags into the hash and XOR them to get better ISS number
2008 * variance. This doesn't enhance the cryptographic strength and is
2009 * done to prevent the 8 cookie bits from showing up directly on the
2013 iss |= cookie.cookie ^ (hash >> 24);
2015 /* Randomize the timestamp. */
2016 if (sc->sc_flags & SCF_TIMESTAMP) {
2017 sc->sc_ts = arc4random();
2018 sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks();
2021 TCPSTAT_INC(tcps_sc_sendcookie);
2025 static struct syncache *
2026 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2027 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2033 int wnd, wscale = 0;
2034 union syncookie cookie;
2036 SCH_LOCK_ASSERT(sch);
2039 * Pull information out of SYN-ACK/ACK and revert sequence number
2042 ack = th->th_ack - 1;
2043 seq = th->th_seq - 1;
2046 * Unpack the flags containing enough information to restore the
2049 cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2051 /* Which of the two secrets to use. */
2052 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
2054 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2056 /* The recomputed hash matches the ACK if this was a genuine cookie. */
2057 if ((ack & ~0xff) != (hash & ~0xff))
2060 /* Fill in the syncache values. */
2062 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2063 sc->sc_ipopts = NULL;
2068 switch (inc->inc_flags & INC_ISIPV6) {
2071 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2072 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2077 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2078 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
2083 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2085 /* We can simply recompute receive window scale we sent earlier. */
2086 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2089 /* Only use wscale if it was enabled in the orignal SYN. */
2090 if (cookie.flags.wscale_idx > 0) {
2091 sc->sc_requested_r_scale = wscale;
2092 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2093 sc->sc_flags |= SCF_WINSCALE;
2096 wnd = sbspace(&lso->so_rcv);
2098 wnd = imin(wnd, TCP_MAXWIN);
2101 if (cookie.flags.sack_ok)
2102 sc->sc_flags |= SCF_SACK;
2104 if (to->to_flags & TOF_TS) {
2105 sc->sc_flags |= SCF_TIMESTAMP;
2106 sc->sc_tsreflect = to->to_tsval;
2107 sc->sc_ts = to->to_tsecr;
2108 sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks();
2111 if (to->to_flags & TOF_SIGNATURE)
2112 sc->sc_flags |= SCF_SIGNATURE;
2116 TCPSTAT_INC(tcps_sc_recvcookie);
2122 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2123 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2126 struct syncache scs, *scx;
2129 bzero(&scs, sizeof(scs));
2130 scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
2132 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2136 if (sc->sc_peer_mss != scx->sc_peer_mss)
2137 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2138 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2140 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2141 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2142 s, __func__, sc->sc_requested_r_scale,
2143 scx->sc_requested_r_scale);
2145 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2146 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2147 s, __func__, sc->sc_requested_s_scale,
2148 scx->sc_requested_s_scale);
2150 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2151 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2158 #endif /* INVARIANTS */
2161 syncookie_reseed(void *arg)
2163 struct tcp_syncache *sc = arg;
2168 * Reseeding the secret doesn't have to be protected by a lock.
2169 * It only must be ensured that the new random values are visible
2170 * to all CPUs in a SMP environment. The atomic with release
2171 * semantics ensures that.
2173 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2174 secbits = sc->secret.key[secbit];
2175 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2176 atomic_add_rel_int(&sc->secret.oddeven, 1);
2178 /* Reschedule ourself. */
2179 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2183 * Exports the syncache entries to userland so that netstat can display
2184 * them alongside the other sockets. This function is intended to be
2185 * called only from tcp_pcblist.
2187 * Due to concurrency on an active system, the number of pcbs exported
2188 * may have no relation to max_pcbs. max_pcbs merely indicates the
2189 * amount of space the caller allocated for this function to use.
2192 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
2195 struct syncache *sc;
2196 struct syncache_head *sch;
2197 int count, error, i;
2199 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
2200 sch = &V_tcp_syncache.hashbase[i];
2202 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2203 if (count >= max_pcbs) {
2207 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2209 bzero(&xt, sizeof(xt));
2210 xt.xt_len = sizeof(xt);
2211 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2212 xt.xt_inp.inp_vflag = INP_IPV6;
2214 xt.xt_inp.inp_vflag = INP_IPV4;
2215 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo));
2216 xt.xt_tp.t_inpcb = &xt.xt_inp;
2217 xt.xt_tp.t_state = TCPS_SYN_RECEIVED;
2218 xt.xt_socket.xso_protocol = IPPROTO_TCP;
2219 xt.xt_socket.xso_len = sizeof (struct xsocket);
2220 xt.xt_socket.so_type = SOCK_STREAM;
2221 xt.xt_socket.so_state = SS_ISCONNECTING;
2222 error = SYSCTL_OUT(req, &xt, sizeof xt);
2232 *pcbs_exported = count;