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");
117 static VNET_DEFINE(int, functions_inherit_listen_socket_stack) = 1;
118 #define V_functions_inherit_listen_socket_stack \
119 VNET(functions_inherit_listen_socket_stack)
120 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack,
121 CTLFLAG_VNET | CTLFLAG_RW,
122 &VNET_NAME(functions_inherit_listen_socket_stack), 0,
123 "Inherit listen socket's stack");
126 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
129 static void syncache_drop(struct syncache *, struct syncache_head *);
130 static void syncache_free(struct syncache *);
131 static void syncache_insert(struct syncache *, struct syncache_head *);
132 static int syncache_respond(struct syncache *, struct syncache_head *, int,
133 const struct mbuf *);
134 static struct socket *syncache_socket(struct syncache *, struct socket *,
136 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch,
138 static void syncache_timer(void *);
140 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
141 uint8_t *, uintptr_t);
142 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *);
143 static struct syncache
144 *syncookie_lookup(struct in_conninfo *, struct syncache_head *,
145 struct syncache *, struct tcphdr *, struct tcpopt *,
147 static void syncookie_reseed(void *);
149 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
150 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
155 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
156 * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds,
157 * the odds are that the user has given up attempting to connect by then.
159 #define SYNCACHE_MAXREXMTS 3
161 /* Arbitrary values */
162 #define TCP_SYNCACHE_HASHSIZE 512
163 #define TCP_SYNCACHE_BUCKETLIMIT 30
165 static VNET_DEFINE(struct tcp_syncache, tcp_syncache);
166 #define V_tcp_syncache VNET(tcp_syncache)
168 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0,
171 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
172 &VNET_NAME(tcp_syncache.bucket_limit), 0,
173 "Per-bucket hash limit for syncache");
175 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
176 &VNET_NAME(tcp_syncache.cache_limit), 0,
177 "Overall entry limit for syncache");
179 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
180 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
182 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
183 &VNET_NAME(tcp_syncache.hashsize), 0,
184 "Size of TCP syncache hashtable");
186 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_VNET | CTLFLAG_RW,
187 &VNET_NAME(tcp_syncache.rexmt_limit), 0,
188 "Limit on SYN/ACK retransmissions");
190 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
191 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
192 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
193 "Send reset on socket allocation failure");
195 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
197 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx)
198 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx)
199 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED)
202 * Requires the syncache entry to be already removed from the bucket list.
205 syncache_free(struct syncache *sc)
209 (void) m_free(sc->sc_ipopts);
213 mac_syncache_destroy(&sc->sc_label);
216 uma_zfree(V_tcp_syncache.zone, sc);
224 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
225 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
226 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
227 V_tcp_syncache.hash_secret = arc4random();
229 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
230 &V_tcp_syncache.hashsize);
231 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
232 &V_tcp_syncache.bucket_limit);
233 if (!powerof2(V_tcp_syncache.hashsize) ||
234 V_tcp_syncache.hashsize == 0) {
235 printf("WARNING: syncache hash size is not a power of 2.\n");
236 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
238 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
241 V_tcp_syncache.cache_limit =
242 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
243 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
244 &V_tcp_syncache.cache_limit);
246 /* Allocate the hash table. */
247 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
248 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
251 V_tcp_syncache.vnet = curvnet;
254 /* Initialize the hash buckets. */
255 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
256 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
257 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
259 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
260 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
261 V_tcp_syncache.hashbase[i].sch_length = 0;
262 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
263 V_tcp_syncache.hashbase[i].sch_last_overflow = INT64_MIN;
266 /* Create the syncache entry zone. */
267 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
268 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
269 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
270 V_tcp_syncache.cache_limit);
272 /* Start the SYN cookie reseeder callout. */
273 callout_init(&V_tcp_syncache.secret.reseed, 1);
274 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
275 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
276 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
277 syncookie_reseed, &V_tcp_syncache);
282 syncache_destroy(void)
284 struct syncache_head *sch;
285 struct syncache *sc, *nsc;
289 * Stop the re-seed timer before freeing resources. No need to
290 * possibly schedule it another time.
292 callout_drain(&V_tcp_syncache.secret.reseed);
294 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */
295 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
297 sch = &V_tcp_syncache.hashbase[i];
298 callout_drain(&sch->sch_timer);
301 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
302 syncache_drop(sc, sch);
304 KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
305 ("%s: sch->sch_bucket not empty", __func__));
306 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
307 __func__, sch->sch_length));
308 mtx_destroy(&sch->sch_mtx);
311 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
312 ("%s: cache_count not 0", __func__));
314 /* Free the allocated global resources. */
315 uma_zdestroy(V_tcp_syncache.zone);
316 free(V_tcp_syncache.hashbase, M_SYNCACHE);
321 * Inserts a syncache entry into the specified bucket row.
322 * Locks and unlocks the syncache_head autonomously.
325 syncache_insert(struct syncache *sc, struct syncache_head *sch)
327 struct syncache *sc2;
332 * Make sure that we don't overflow the per-bucket limit.
333 * If the bucket is full, toss the oldest element.
335 if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
336 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
337 ("sch->sch_length incorrect"));
338 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
339 sch->sch_last_overflow = time_uptime;
340 syncache_drop(sc2, sch);
341 TCPSTAT_INC(tcps_sc_bucketoverflow);
344 /* Put it into the bucket. */
345 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
349 if (ADDED_BY_TOE(sc)) {
350 struct toedev *tod = sc->sc_tod;
352 tod->tod_syncache_added(tod, sc->sc_todctx);
356 /* Reinitialize the bucket row's timer. */
357 if (sch->sch_length == 1)
358 sch->sch_nextc = ticks + INT_MAX;
359 syncache_timeout(sc, sch, 1);
363 TCPSTATES_INC(TCPS_SYN_RECEIVED);
364 TCPSTAT_INC(tcps_sc_added);
368 * Remove and free entry from syncache bucket row.
369 * Expects locked syncache head.
372 syncache_drop(struct syncache *sc, struct syncache_head *sch)
375 SCH_LOCK_ASSERT(sch);
377 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
378 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
382 if (ADDED_BY_TOE(sc)) {
383 struct toedev *tod = sc->sc_tod;
385 tod->tod_syncache_removed(tod, sc->sc_todctx);
393 * Engage/reengage time on bucket row.
396 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
398 sc->sc_rxttime = ticks +
399 TCPTV_RTOBASE * (tcp_syn_backoff[sc->sc_rxmits]);
401 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
402 sch->sch_nextc = sc->sc_rxttime;
404 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
405 syncache_timer, (void *)sch);
410 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
411 * If we have retransmitted an entry the maximum number of times, expire it.
412 * One separate timer for each bucket row.
415 syncache_timer(void *xsch)
417 struct syncache_head *sch = (struct syncache_head *)xsch;
418 struct syncache *sc, *nsc;
422 CURVNET_SET(sch->sch_sc->vnet);
424 /* NB: syncache_head has already been locked by the callout. */
425 SCH_LOCK_ASSERT(sch);
428 * In the following cycle we may remove some entries and/or
429 * advance some timeouts, so re-initialize the bucket timer.
431 sch->sch_nextc = tick + INT_MAX;
433 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
435 * We do not check if the listen socket still exists
436 * and accept the case where the listen socket may be
437 * gone by the time we resend the SYN/ACK. We do
438 * not expect this to happens often. If it does,
439 * then the RST will be sent by the time the remote
440 * host does the SYN/ACK->ACK.
442 if (TSTMP_GT(sc->sc_rxttime, tick)) {
443 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
444 sch->sch_nextc = sc->sc_rxttime;
447 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
448 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
449 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
450 "giving up and removing syncache entry\n",
454 syncache_drop(sc, sch);
455 TCPSTAT_INC(tcps_sc_stale);
458 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
459 log(LOG_DEBUG, "%s; %s: Response timeout, "
460 "retransmitting (%u) SYN|ACK\n",
461 s, __func__, sc->sc_rxmits);
465 syncache_respond(sc, sch, 1, NULL);
466 TCPSTAT_INC(tcps_sc_retransmitted);
467 syncache_timeout(sc, sch, 0);
469 if (!TAILQ_EMPTY(&(sch)->sch_bucket))
470 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
471 syncache_timer, (void *)(sch));
476 * Find an entry in the syncache.
477 * Returns always with locked syncache_head plus a matching entry or NULL.
479 static struct syncache *
480 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
483 struct syncache_head *sch;
487 * The hash is built on foreign port + local port + foreign address.
488 * We rely on the fact that struct in_conninfo starts with 16 bits
489 * of foreign port, then 16 bits of local port then followed by 128
490 * bits of foreign address. In case of IPv4 address, the first 3
491 * 32-bit words of the address always are zeroes.
493 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
494 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
496 sch = &V_tcp_syncache.hashbase[hash];
500 /* Circle through bucket row to find matching entry. */
501 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
502 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
503 sizeof(struct in_endpoints)) == 0)
506 return (sc); /* Always returns with locked sch. */
510 * This function is called when we get a RST for a
511 * non-existent connection, so that we can see if the
512 * connection is in the syn cache. If it is, zap it.
515 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
518 struct syncache_head *sch;
521 sc = syncache_lookup(inc, &sch); /* returns locked sch */
522 SCH_LOCK_ASSERT(sch);
525 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
526 * See RFC 793 page 65, section SEGMENT ARRIVES.
528 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
529 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
530 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
531 "FIN flag set, segment ignored\n", s, __func__);
532 TCPSTAT_INC(tcps_badrst);
537 * No corresponding connection was found in syncache.
538 * If syncookies are enabled and possibly exclusively
539 * used, or we are under memory pressure, a valid RST
540 * may not find a syncache entry. In that case we're
541 * done and no SYN|ACK retransmissions will happen.
542 * Otherwise the RST was misdirected or spoofed.
545 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
546 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
547 "syncache entry (possibly syncookie only), "
548 "segment ignored\n", s, __func__);
549 TCPSTAT_INC(tcps_badrst);
554 * If the RST bit is set, check the sequence number to see
555 * if this is a valid reset segment.
557 * In all states except SYN-SENT, all reset (RST) segments
558 * are validated by checking their SEQ-fields. A reset is
559 * valid if its sequence number is in the window.
561 * The sequence number in the reset segment is normally an
562 * echo of our outgoing acknowlegement numbers, but some hosts
563 * send a reset with the sequence number at the rightmost edge
564 * of our receive window, and we have to handle this case.
566 if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
567 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
568 syncache_drop(sc, sch);
569 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
570 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, "
571 "connection attempt aborted by remote endpoint\n",
573 TCPSTAT_INC(tcps_sc_reset);
575 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
576 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
577 "IRS %u (+WND %u), segment ignored\n",
578 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd);
579 TCPSTAT_INC(tcps_badrst);
589 syncache_badack(struct in_conninfo *inc)
592 struct syncache_head *sch;
594 sc = syncache_lookup(inc, &sch); /* returns locked sch */
595 SCH_LOCK_ASSERT(sch);
597 syncache_drop(sc, sch);
598 TCPSTAT_INC(tcps_sc_badack);
604 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th)
607 struct syncache_head *sch;
609 sc = syncache_lookup(inc, &sch); /* returns locked sch */
610 SCH_LOCK_ASSERT(sch);
614 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
615 if (ntohl(th->th_seq) != sc->sc_iss)
619 * If we've rertransmitted 3 times and this is our second error,
620 * we remove the entry. Otherwise, we allow it to continue on.
621 * This prevents us from incorrectly nuking an entry during a
622 * spurious network outage.
626 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
627 sc->sc_flags |= SCF_UNREACH;
630 syncache_drop(sc, sch);
631 TCPSTAT_INC(tcps_sc_unreach);
637 * Build a new TCP socket structure from a syncache entry.
639 * On success return the newly created socket with its underlying inp locked.
641 static struct socket *
642 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
644 struct tcp_function_block *blk;
645 struct inpcb *inp = NULL;
651 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
654 * Ok, create the full blown connection, and set things up
655 * as they would have been set up if we had created the
656 * connection when the SYN arrived. If we can't create
657 * the connection, abort it.
659 so = sonewconn(lso, 0);
662 * Drop the connection; we will either send a RST or
663 * have the peer retransmit its SYN again after its
666 TCPSTAT_INC(tcps_listendrop);
667 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
668 log(LOG_DEBUG, "%s; %s: Socket create failed "
669 "due to limits or memory shortage\n",
676 mac_socketpeer_set_from_mbuf(m, so);
680 inp->inp_inc.inc_fibnum = so->so_fibnum;
683 * Exclusive pcbinfo lock is not required in syncache socket case even
684 * if two inpcb locks can be acquired simultaneously:
685 * - the inpcb in LISTEN state,
686 * - the newly created inp.
688 * In this case, an inp cannot be at same time in LISTEN state and
689 * just created by an accept() call.
691 INP_HASH_WLOCK(&V_tcbinfo);
693 /* Insert new socket into PCB hash list. */
694 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
696 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
697 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
699 inp->inp_vflag &= ~INP_IPV6;
700 inp->inp_vflag |= INP_IPV4;
702 inp->inp_laddr = sc->sc_inc.inc_laddr;
708 * If there's an mbuf and it has a flowid, then let's initialise the
709 * inp with that particular flowid.
711 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
712 inp->inp_flowid = m->m_pkthdr.flowid;
713 inp->inp_flowtype = M_HASHTYPE_GET(m);
717 * Install in the reservation hash table for now, but don't yet
718 * install a connection group since the full 4-tuple isn't yet
721 inp->inp_lport = sc->sc_inc.inc_lport;
722 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
724 * Undo the assignments above if we failed to
725 * put the PCB on the hash lists.
728 if (sc->sc_inc.inc_flags & INC_ISIPV6)
729 inp->in6p_laddr = in6addr_any;
732 inp->inp_laddr.s_addr = INADDR_ANY;
734 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
735 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
740 INP_HASH_WUNLOCK(&V_tcbinfo);
744 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
745 struct inpcb *oinp = sotoinpcb(lso);
746 struct in6_addr laddr6;
747 struct sockaddr_in6 sin6;
749 * Inherit socket options from the listening socket.
750 * Note that in6p_inputopts are not (and should not be)
751 * copied, since it stores previously received options and is
752 * used to detect if each new option is different than the
753 * previous one and hence should be passed to a user.
754 * If we copied in6p_inputopts, a user would not be able to
755 * receive options just after calling the accept system call.
757 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
758 if (oinp->in6p_outputopts)
759 inp->in6p_outputopts =
760 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
762 sin6.sin6_family = AF_INET6;
763 sin6.sin6_len = sizeof(sin6);
764 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
765 sin6.sin6_port = sc->sc_inc.inc_fport;
766 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
767 laddr6 = inp->in6p_laddr;
768 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
769 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
770 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
771 thread0.td_ucred, m)) != 0) {
772 inp->in6p_laddr = laddr6;
773 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
774 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
779 INP_HASH_WUNLOCK(&V_tcbinfo);
782 /* Override flowlabel from in6_pcbconnect. */
783 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
784 inp->inp_flow |= sc->sc_flowlabel;
787 #if defined(INET) && defined(INET6)
792 struct in_addr laddr;
793 struct sockaddr_in sin;
795 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
797 if (inp->inp_options == NULL) {
798 inp->inp_options = sc->sc_ipopts;
799 sc->sc_ipopts = NULL;
802 sin.sin_family = AF_INET;
803 sin.sin_len = sizeof(sin);
804 sin.sin_addr = sc->sc_inc.inc_faddr;
805 sin.sin_port = sc->sc_inc.inc_fport;
806 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
807 laddr = inp->inp_laddr;
808 if (inp->inp_laddr.s_addr == INADDR_ANY)
809 inp->inp_laddr = sc->sc_inc.inc_laddr;
810 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
811 thread0.td_ucred, m)) != 0) {
812 inp->inp_laddr = laddr;
813 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
814 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
819 INP_HASH_WUNLOCK(&V_tcbinfo);
824 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
825 /* Copy old policy into new socket's. */
826 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
827 printf("syncache_socket: could not copy policy\n");
829 INP_HASH_WUNLOCK(&V_tcbinfo);
831 tcp_state_change(tp, TCPS_SYN_RECEIVED);
832 tp->iss = sc->sc_iss;
833 tp->irs = sc->sc_irs;
836 blk = sototcpcb(lso)->t_fb;
837 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
839 * Our parents t_fb was not the default,
840 * we need to release our ref on tp->t_fb and
841 * pickup one on the new entry.
843 struct tcp_function_block *rblk;
845 rblk = find_and_ref_tcp_fb(blk);
846 KASSERT(rblk != NULL,
847 ("cannot find blk %p out of syncache?", blk));
848 if (tp->t_fb->tfb_tcp_fb_fini)
849 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
850 refcount_release(&tp->t_fb->tfb_refcnt);
852 if (tp->t_fb->tfb_tcp_fb_init) {
853 (*tp->t_fb->tfb_tcp_fb_init)(tp);
856 tp->snd_wl1 = sc->sc_irs;
857 tp->snd_max = tp->iss + 1;
858 tp->snd_nxt = tp->iss + 1;
859 tp->rcv_up = sc->sc_irs + 1;
860 tp->rcv_wnd = sc->sc_wnd;
861 tp->rcv_adv += tp->rcv_wnd;
862 tp->last_ack_sent = tp->rcv_nxt;
864 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
865 if (sc->sc_flags & SCF_NOOPT)
866 tp->t_flags |= TF_NOOPT;
868 if (sc->sc_flags & SCF_WINSCALE) {
869 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
870 tp->snd_scale = sc->sc_requested_s_scale;
871 tp->request_r_scale = sc->sc_requested_r_scale;
873 if (sc->sc_flags & SCF_TIMESTAMP) {
874 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
875 tp->ts_recent = sc->sc_tsreflect;
876 tp->ts_recent_age = tcp_ts_getticks();
877 tp->ts_offset = sc->sc_tsoff;
879 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
880 if (sc->sc_flags & SCF_SIGNATURE)
881 tp->t_flags |= TF_SIGNATURE;
883 if (sc->sc_flags & SCF_SACK)
884 tp->t_flags |= TF_SACK_PERMIT;
887 if (sc->sc_flags & SCF_ECN)
888 tp->t_flags |= TF_ECN_PERMIT;
891 * Set up MSS and get cached values from tcp_hostcache.
892 * This might overwrite some of the defaults we just set.
894 tcp_mss(tp, sc->sc_peer_mss);
897 * If the SYN,ACK was retransmitted, indicate that CWND to be
898 * limited to one segment in cc_conn_init().
899 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
901 if (sc->sc_rxmits > 1)
906 * Allow a TOE driver to install its hooks. Note that we hold the
907 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
908 * new connection before the TOE driver has done its thing.
910 if (ADDED_BY_TOE(sc)) {
911 struct toedev *tod = sc->sc_tod;
913 tod->tod_offload_socket(tod, sc->sc_todctx, so);
917 * Copy and activate timers.
919 tp->t_keepinit = sototcpcb(lso)->t_keepinit;
920 tp->t_keepidle = sototcpcb(lso)->t_keepidle;
921 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
922 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
923 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
925 TCPSTAT_INC(tcps_accepts);
937 * This function gets called when we receive an ACK for a
938 * socket in the LISTEN state. We look up the connection
939 * in the syncache, and if its there, we pull it out of
940 * the cache and turn it into a full-blown connection in
941 * the SYN-RECEIVED state.
943 * On syncache_socket() success the newly created socket
944 * has its underlying inp locked.
947 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
948 struct socket **lsop, struct mbuf *m)
951 struct syncache_head *sch;
956 * Global TCP locks are held because we manipulate the PCB lists
957 * and create a new socket.
959 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
960 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
961 ("%s: can handle only ACK", __func__));
963 sc = syncache_lookup(inc, &sch); /* returns locked sch */
964 SCH_LOCK_ASSERT(sch);
968 * Test code for syncookies comparing the syncache stored
969 * values with the reconstructed values from the cookie.
972 syncookie_cmp(inc, sch, sc, th, to, *lsop);
977 * There is no syncache entry, so see if this ACK is
978 * a returning syncookie. To do this, first:
979 * A. Check if syncookies are used in case of syncache
981 * B. See if this socket has had a syncache entry dropped in
982 * the recent past. We don't want to accept a bogus
983 * syncookie if we've never received a SYN or accept it
985 * C. check that the syncookie is valid. If it is, then
986 * cobble up a fake syncache entry, and return.
988 if (!V_tcp_syncookies) {
990 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
991 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
992 "segment rejected (syncookies disabled)\n",
996 if (!V_tcp_syncookiesonly &&
997 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
999 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1000 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1001 "segment rejected (no syncache entry)\n",
1005 bzero(&scs, sizeof(scs));
1006 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
1009 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1010 log(LOG_DEBUG, "%s; %s: Segment failed "
1011 "SYNCOOKIE authentication, segment rejected "
1012 "(probably spoofed)\n", s, __func__);
1015 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1016 /* If received ACK has MD5 signature, check it. */
1017 if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1018 (!TCPMD5_ENABLED() ||
1019 TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1021 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1022 log(LOG_DEBUG, "%s; %s: Segment rejected, "
1023 "MD5 signature doesn't match.\n",
1027 TCPSTAT_INC(tcps_sig_err_sigopt);
1028 return (-1); /* Do not send RST */
1030 #endif /* TCP_SIGNATURE */
1032 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1034 * If listening socket requested TCP digests, check that
1035 * received ACK has signature and it is correct.
1036 * If not, drop the ACK and leave sc entry in th cache,
1037 * because SYN was received with correct signature.
1039 if (sc->sc_flags & SCF_SIGNATURE) {
1040 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1042 TCPSTAT_INC(tcps_sig_err_nosigopt);
1044 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1045 log(LOG_DEBUG, "%s; %s: Segment "
1046 "rejected, MD5 signature wasn't "
1047 "provided.\n", s, __func__);
1050 return (-1); /* Do not send RST */
1052 if (!TCPMD5_ENABLED() ||
1053 TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1054 /* Doesn't match or no SA */
1056 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1057 log(LOG_DEBUG, "%s; %s: Segment "
1058 "rejected, MD5 signature doesn't "
1059 "match.\n", s, __func__);
1062 return (-1); /* Do not send RST */
1065 #endif /* TCP_SIGNATURE */
1067 * Pull out the entry to unlock the bucket row.
1069 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1070 * tcp_state_change(). The tcpcb is not existent at this
1071 * moment. A new one will be allocated via syncache_socket->
1072 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1073 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1075 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1076 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1079 if (ADDED_BY_TOE(sc)) {
1080 struct toedev *tod = sc->sc_tod;
1082 tod->tod_syncache_removed(tod, sc->sc_todctx);
1089 * Segment validation:
1090 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1092 if (th->th_ack != sc->sc_iss + 1) {
1093 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1094 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1095 "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1100 * The SEQ must fall in the window starting at the received
1101 * initial receive sequence number + 1 (the SYN).
1103 if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1104 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1105 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1106 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1107 "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1112 * If timestamps were not negotiated during SYN/ACK they
1113 * must not appear on any segment during this session.
1115 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1116 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1117 log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1118 "segment rejected\n", s, __func__);
1123 * If timestamps were negotiated during SYN/ACK they should
1124 * appear on every segment during this session.
1125 * XXXAO: This is only informal as there have been unverified
1126 * reports of non-compliants stacks.
1128 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1129 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1130 log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1131 "no action\n", s, __func__);
1138 * If timestamps were negotiated, the reflected timestamp
1139 * must be equal to what we actually sent in the SYN|ACK
1140 * except in the case of 0. Some boxes are known for sending
1141 * broken timestamp replies during the 3whs (and potentially
1142 * during the connection also).
1144 * Accept the final ACK of 3whs with reflected timestamp of 0
1145 * instead of sending a RST and deleting the syncache entry.
1147 if ((to->to_flags & TOF_TS) && to->to_tsecr &&
1148 to->to_tsecr != sc->sc_ts) {
1149 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1150 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, "
1151 "segment rejected\n",
1152 s, __func__, to->to_tsecr, sc->sc_ts);
1156 *lsop = syncache_socket(sc, *lsop, m);
1159 TCPSTAT_INC(tcps_sc_aborted);
1161 TCPSTAT_INC(tcps_sc_completed);
1163 /* how do we find the inp for the new socket? */
1168 if (sc != NULL && sc != &scs)
1178 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
1179 uint64_t response_cookie)
1183 unsigned int *pending_counter;
1186 * Global TCP locks are held because we manipulate the PCB lists
1187 * and create a new socket.
1189 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1191 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
1192 *lsop = syncache_socket(sc, *lsop, m);
1193 if (*lsop == NULL) {
1194 TCPSTAT_INC(tcps_sc_aborted);
1195 atomic_subtract_int(pending_counter, 1);
1197 inp = sotoinpcb(*lsop);
1198 tp = intotcpcb(inp);
1199 tp->t_flags |= TF_FASTOPEN;
1200 tp->t_tfo_cookie = response_cookie;
1201 tp->snd_max = tp->iss;
1202 tp->snd_nxt = tp->iss;
1203 tp->t_tfo_pending = pending_counter;
1204 TCPSTAT_INC(tcps_sc_completed);
1207 #endif /* TCP_RFC7413 */
1210 * Given a LISTEN socket and an inbound SYN request, add
1211 * this to the syn cache, and send back a segment:
1212 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1215 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1216 * Doing so would require that we hold onto the data and deliver it
1217 * to the application. However, if we are the target of a SYN-flood
1218 * DoS attack, an attacker could send data which would eventually
1219 * consume all available buffer space if it were ACKed. By not ACKing
1220 * the data, we avoid this DoS scenario.
1222 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1223 * cookie is processed and a new socket is created. In this case, any data
1224 * accompanying the SYN will be queued to the socket by tcp_input() and will
1225 * be ACKed either when the application sends response data or the delayed
1226 * ACK timer expires, whichever comes first.
1229 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1230 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1235 struct syncache *sc = NULL;
1236 struct syncache_head *sch;
1237 struct mbuf *ipopts = NULL;
1239 int win, ip_ttl, ip_tos;
1243 int autoflowlabel = 0;
1246 struct label *maclabel;
1248 struct syncache scs;
1251 uint64_t tfo_response_cookie;
1252 unsigned int *tfo_pending = NULL;
1253 int tfo_cookie_valid = 0;
1254 int tfo_response_cookie_valid = 0;
1257 INP_WLOCK_ASSERT(inp); /* listen socket */
1258 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1259 ("%s: unexpected tcp flags", __func__));
1262 * Combine all so/tp operations very early to drop the INP lock as
1267 cred = crhold(so->so_cred);
1270 if ((inc->inc_flags & INC_ISIPV6) &&
1271 (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1274 ip_ttl = inp->inp_ip_ttl;
1275 ip_tos = inp->inp_ip_tos;
1276 win = sbspace(&so->so_rcv);
1277 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1280 if (V_tcp_fastopen_enabled && IS_FASTOPEN(tp->t_flags) &&
1281 (tp->t_tfo_pending != NULL) && (to->to_flags & TOF_FASTOPEN)) {
1283 * Limit the number of pending TFO connections to
1284 * approximately half of the queue limit. This prevents TFO
1285 * SYN floods from starving the service by filling the
1286 * listen queue with bogus TFO connections.
1288 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1289 (so->so_qlimit / 2)) {
1292 result = tcp_fastopen_check_cookie(inc,
1293 to->to_tfo_cookie, to->to_tfo_len,
1294 &tfo_response_cookie);
1295 tfo_cookie_valid = (result > 0);
1296 tfo_response_cookie_valid = (result >= 0);
1300 * Remember the TFO pending counter as it will have to be
1301 * decremented below if we don't make it to syncache_tfo_expand().
1303 tfo_pending = tp->t_tfo_pending;
1307 /* By the time we drop the lock these should no longer be used. */
1312 if (mac_syncache_init(&maclabel) != 0) {
1316 mac_syncache_create(maclabel, inp);
1319 if (!tfo_cookie_valid)
1324 * Remember the IP options, if any.
1327 if (!(inc->inc_flags & INC_ISIPV6))
1330 ipopts = (m) ? ip_srcroute(m) : NULL;
1335 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1337 * If listening socket requested TCP digests, check that received
1338 * SYN has signature and it is correct. If signature doesn't match
1339 * or TCP_SIGNATURE support isn't enabled, drop the packet.
1341 if (ltflags & TF_SIGNATURE) {
1342 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1343 TCPSTAT_INC(tcps_sig_err_nosigopt);
1346 if (!TCPMD5_ENABLED() ||
1347 TCPMD5_INPUT(m, th, to->to_signature) != 0)
1350 #endif /* TCP_SIGNATURE */
1352 * See if we already have an entry for this connection.
1353 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1355 * XXX: should the syncache be re-initialized with the contents
1356 * of the new SYN here (which may have different options?)
1358 * XXX: We do not check the sequence number to see if this is a
1359 * real retransmit or a new connection attempt. The question is
1360 * how to handle such a case; either ignore it as spoofed, or
1361 * drop the current entry and create a new one?
1363 sc = syncache_lookup(inc, &sch); /* returns locked entry */
1364 SCH_LOCK_ASSERT(sch);
1367 if (tfo_cookie_valid)
1370 TCPSTAT_INC(tcps_sc_dupsyn);
1373 * If we were remembering a previous source route,
1374 * forget it and use the new one we've been given.
1377 (void) m_free(sc->sc_ipopts);
1378 sc->sc_ipopts = ipopts;
1381 * Update timestamp if present.
1383 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1384 sc->sc_tsreflect = to->to_tsval;
1386 sc->sc_flags &= ~SCF_TIMESTAMP;
1389 * Since we have already unconditionally allocated label
1390 * storage, free it up. The syncache entry will already
1391 * have an initialized label we can use.
1393 mac_syncache_destroy(&maclabel);
1395 /* Retransmit SYN|ACK and reset retransmit count. */
1396 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1397 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1398 "resetting timer and retransmitting SYN|ACK\n",
1402 if (syncache_respond(sc, sch, 1, m) == 0) {
1404 syncache_timeout(sc, sch, 1);
1405 TCPSTAT_INC(tcps_sndacks);
1406 TCPSTAT_INC(tcps_sndtotal);
1413 if (tfo_cookie_valid) {
1414 bzero(&scs, sizeof(scs));
1420 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1423 * The zone allocator couldn't provide more entries.
1424 * Treat this as if the cache was full; drop the oldest
1425 * entry and insert the new one.
1427 TCPSTAT_INC(tcps_sc_zonefail);
1428 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1429 sch->sch_last_overflow = time_uptime;
1430 syncache_drop(sc, sch);
1432 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1434 if (V_tcp_syncookies) {
1435 bzero(&scs, sizeof(scs));
1440 (void) m_free(ipopts);
1448 if (!tfo_cookie_valid && tfo_response_cookie_valid)
1449 sc->sc_tfo_cookie = &tfo_response_cookie;
1453 * Fill in the syncache values.
1456 sc->sc_label = maclabel;
1460 sc->sc_ipopts = ipopts;
1461 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1463 if (!(inc->inc_flags & INC_ISIPV6))
1466 sc->sc_ip_tos = ip_tos;
1467 sc->sc_ip_ttl = ip_ttl;
1471 sc->sc_todctx = todctx;
1473 sc->sc_irs = th->th_seq;
1474 sc->sc_iss = arc4random();
1476 sc->sc_flowlabel = 0;
1479 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1480 * win was derived from socket earlier in the function.
1483 win = imin(win, TCP_MAXWIN);
1486 if (V_tcp_do_rfc1323) {
1488 * A timestamp received in a SYN makes
1489 * it ok to send timestamp requests and replies.
1491 if (to->to_flags & TOF_TS) {
1492 sc->sc_tsreflect = to->to_tsval;
1493 sc->sc_ts = tcp_ts_getticks();
1494 sc->sc_flags |= SCF_TIMESTAMP;
1496 if (to->to_flags & TOF_SCALE) {
1500 * Pick the smallest possible scaling factor that
1501 * will still allow us to scale up to sb_max, aka
1502 * kern.ipc.maxsockbuf.
1504 * We do this because there are broken firewalls that
1505 * will corrupt the window scale option, leading to
1506 * the other endpoint believing that our advertised
1507 * window is unscaled. At scale factors larger than
1508 * 5 the unscaled window will drop below 1500 bytes,
1509 * leading to serious problems when traversing these
1512 * With the default maxsockbuf of 256K, a scale factor
1513 * of 3 will be chosen by this algorithm. Those who
1514 * choose a larger maxsockbuf should watch out
1515 * for the compatibility problems mentioned above.
1517 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1518 * or <SYN,ACK>) segment itself is never scaled.
1520 while (wscale < TCP_MAX_WINSHIFT &&
1521 (TCP_MAXWIN << wscale) < sb_max)
1523 sc->sc_requested_r_scale = wscale;
1524 sc->sc_requested_s_scale = to->to_wscale;
1525 sc->sc_flags |= SCF_WINSCALE;
1528 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1530 * If listening socket requested TCP digests, flag this in the
1531 * syncache so that syncache_respond() will do the right thing
1534 if (ltflags & TF_SIGNATURE)
1535 sc->sc_flags |= SCF_SIGNATURE;
1536 #endif /* TCP_SIGNATURE */
1537 if (to->to_flags & TOF_SACKPERM)
1538 sc->sc_flags |= SCF_SACK;
1539 if (to->to_flags & TOF_MSS)
1540 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */
1541 if (ltflags & TF_NOOPT)
1542 sc->sc_flags |= SCF_NOOPT;
1543 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1544 sc->sc_flags |= SCF_ECN;
1546 if (V_tcp_syncookies)
1547 sc->sc_iss = syncookie_generate(sch, sc);
1549 if (autoflowlabel) {
1550 if (V_tcp_syncookies)
1551 sc->sc_flowlabel = sc->sc_iss;
1553 sc->sc_flowlabel = ip6_randomflowlabel();
1554 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1560 if (tfo_cookie_valid) {
1561 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie);
1562 /* INP_WUNLOCK(inp) will be performed by the caller */
1569 * Do a standard 3-way handshake.
1571 if (syncache_respond(sc, sch, 0, m) == 0) {
1572 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1574 else if (sc != &scs)
1575 syncache_insert(sc, sch); /* locks and unlocks sch */
1576 TCPSTAT_INC(tcps_sndacks);
1577 TCPSTAT_INC(tcps_sndtotal);
1581 TCPSTAT_INC(tcps_sc_dropped);
1591 * If tfo_pending is not NULL here, then a TFO SYN that did not
1592 * result in a new socket was processed and the associated pending
1593 * counter has not yet been decremented. All such TFO processing paths
1594 * transit this point.
1596 if (tfo_pending != NULL)
1597 tcp_fastopen_decrement_counter(tfo_pending);
1605 mac_syncache_destroy(&maclabel);
1611 * Send SYN|ACK to the peer. Either in response to the peer's SYN,
1612 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1615 syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked,
1616 const struct mbuf *m0)
1618 struct ip *ip = NULL;
1620 struct tcphdr *th = NULL;
1621 int optlen, error = 0; /* Make compiler happy */
1622 u_int16_t hlen, tlen, mssopt;
1625 struct ip6_hdr *ip6 = NULL;
1629 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1632 tlen = hlen + sizeof(struct tcphdr);
1634 /* Determine MSS we advertize to other end of connection. */
1635 mssopt = tcp_mssopt(&sc->sc_inc);
1636 if (sc->sc_peer_mss)
1637 mssopt = max( min(sc->sc_peer_mss, mssopt), V_tcp_minmss);
1639 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1640 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1641 ("syncache: mbuf too small"));
1643 /* Create the IP+TCP header from scratch. */
1644 m = m_gethdr(M_NOWAIT, MT_DATA);
1648 mac_syncache_create_mbuf(sc->sc_label, m);
1650 m->m_data += max_linkhdr;
1652 m->m_pkthdr.len = tlen;
1653 m->m_pkthdr.rcvif = NULL;
1656 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1657 ip6 = mtod(m, struct ip6_hdr *);
1658 ip6->ip6_vfc = IPV6_VERSION;
1659 ip6->ip6_nxt = IPPROTO_TCP;
1660 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1661 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1662 ip6->ip6_plen = htons(tlen - hlen);
1663 /* ip6_hlim is set after checksum */
1664 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1665 ip6->ip6_flow |= sc->sc_flowlabel;
1667 th = (struct tcphdr *)(ip6 + 1);
1670 #if defined(INET6) && defined(INET)
1675 ip = mtod(m, struct ip *);
1676 ip->ip_v = IPVERSION;
1677 ip->ip_hl = sizeof(struct ip) >> 2;
1678 ip->ip_len = htons(tlen);
1682 ip->ip_p = IPPROTO_TCP;
1683 ip->ip_src = sc->sc_inc.inc_laddr;
1684 ip->ip_dst = sc->sc_inc.inc_faddr;
1685 ip->ip_ttl = sc->sc_ip_ttl;
1686 ip->ip_tos = sc->sc_ip_tos;
1689 * See if we should do MTU discovery. Route lookups are
1690 * expensive, so we will only unset the DF bit if:
1692 * 1) path_mtu_discovery is disabled
1693 * 2) the SCF_UNREACH flag has been set
1695 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1696 ip->ip_off |= htons(IP_DF);
1698 th = (struct tcphdr *)(ip + 1);
1701 th->th_sport = sc->sc_inc.inc_lport;
1702 th->th_dport = sc->sc_inc.inc_fport;
1704 th->th_seq = htonl(sc->sc_iss);
1705 th->th_ack = htonl(sc->sc_irs + 1);
1706 th->th_off = sizeof(struct tcphdr) >> 2;
1708 th->th_flags = TH_SYN|TH_ACK;
1709 th->th_win = htons(sc->sc_wnd);
1712 if (sc->sc_flags & SCF_ECN) {
1713 th->th_flags |= TH_ECE;
1714 TCPSTAT_INC(tcps_ecn_shs);
1717 /* Tack on the TCP options. */
1718 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1722 to.to_flags = TOF_MSS;
1723 if (sc->sc_flags & SCF_WINSCALE) {
1724 to.to_wscale = sc->sc_requested_r_scale;
1725 to.to_flags |= TOF_SCALE;
1727 if (sc->sc_flags & SCF_TIMESTAMP) {
1728 /* Virgin timestamp or TCP cookie enhanced one. */
1729 to.to_tsval = sc->sc_ts;
1730 to.to_tsecr = sc->sc_tsreflect;
1731 to.to_flags |= TOF_TS;
1733 if (sc->sc_flags & SCF_SACK)
1734 to.to_flags |= TOF_SACKPERM;
1735 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1736 if (sc->sc_flags & SCF_SIGNATURE)
1737 to.to_flags |= TOF_SIGNATURE;
1740 if (sc->sc_tfo_cookie) {
1741 to.to_flags |= TOF_FASTOPEN;
1742 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1743 to.to_tfo_cookie = sc->sc_tfo_cookie;
1744 /* don't send cookie again when retransmitting response */
1745 sc->sc_tfo_cookie = NULL;
1748 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1750 /* Adjust headers by option size. */
1751 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1753 m->m_pkthdr.len += optlen;
1755 if (sc->sc_inc.inc_flags & INC_ISIPV6)
1756 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1759 ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1760 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1761 if (sc->sc_flags & SCF_SIGNATURE) {
1762 KASSERT(to.to_flags & TOF_SIGNATURE,
1763 ("tcp_addoptions() didn't set tcp_signature"));
1765 /* NOTE: to.to_signature is inside of mbuf */
1766 if (!TCPMD5_ENABLED() ||
1767 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1776 M_SETFIB(m, sc->sc_inc.inc_fibnum);
1777 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1779 * If we have peer's SYN and it has a flowid, then let's assign it to
1780 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid
1781 * to SYN|ACK due to lack of inp here.
1783 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1784 m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1785 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1788 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1789 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1790 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1792 ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1794 if (ADDED_BY_TOE(sc)) {
1795 struct toedev *tod = sc->sc_tod;
1797 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1802 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1805 #if defined(INET6) && defined(INET)
1810 m->m_pkthdr.csum_flags = CSUM_TCP;
1811 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1812 htons(tlen + optlen - hlen + IPPROTO_TCP));
1814 if (ADDED_BY_TOE(sc)) {
1815 struct toedev *tod = sc->sc_tod;
1817 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1822 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1829 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1830 * that exceed the capacity of the syncache by avoiding the storage of any
1831 * of the SYNs we receive. Syncookies defend against blind SYN flooding
1832 * attacks where the attacker does not have access to our responses.
1834 * Syncookies encode and include all necessary information about the
1835 * connection setup within the SYN|ACK that we send back. That way we
1836 * can avoid keeping any local state until the ACK to our SYN|ACK returns
1837 * (if ever). Normally the syncache and syncookies are running in parallel
1838 * with the latter taking over when the former is exhausted. When matching
1839 * syncache entry is found the syncookie is ignored.
1841 * The only reliable information persisting the 3WHS is our initial sequence
1842 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient
1843 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1844 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK
1845 * returns and signifies a legitimate connection if it matches the ACK.
1847 * The available space of 32 bits to store the hash and to encode the SYN
1848 * option information is very tight and we should have at least 24 bits for
1849 * the MAC to keep the number of guesses by blind spoofing reasonably high.
1851 * SYN option information we have to encode to fully restore a connection:
1852 * MSS: is imporant to chose an optimal segment size to avoid IP level
1853 * fragmentation along the path. The common MSS values can be encoded
1854 * in a 3-bit table. Uncommon values are captured by the next lower value
1855 * in the table leading to a slight increase in packetization overhead.
1856 * WSCALE: is necessary to allow large windows to be used for high delay-
1857 * bandwidth product links. Not scaling the window when it was initially
1858 * negotiated is bad for performance as lack of scaling further decreases
1859 * the apparent available send window. We only need to encode the WSCALE
1860 * we received from the remote end. Our end can be recalculated at any
1861 * time. The common WSCALE values can be encoded in a 3-bit table.
1862 * Uncommon values are captured by the next lower value in the table
1863 * making us under-estimate the available window size halving our
1864 * theoretically possible maximum throughput for that connection.
1865 * SACK: Greatly assists in packet loss recovery and requires 1 bit.
1866 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
1867 * that are included in all segments on a connection. We enable them when
1870 * Security of syncookies and attack vectors:
1872 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
1873 * together with the gloabl secret to make it unique per connection attempt.
1874 * Thus any change of any of those parameters results in a different MAC output
1875 * in an unpredictable way unless a collision is encountered. 24 bits of the
1876 * MAC are embedded into the ISS.
1878 * To prevent replay attacks two rotating global secrets are updated with a
1879 * new random value every 15 seconds. The life-time of a syncookie is thus
1882 * Vector 1: Attacking the secret. This requires finding a weakness in the
1883 * MAC itself or the way it is used here. The attacker can do a chosen plain
1884 * text attack by varying and testing the all parameters under his control.
1885 * The strength depends on the size and randomness of the secret, and the
1886 * cryptographic security of the MAC function. Due to the constant updating
1887 * of the secret the attacker has at most 29.999 seconds to find the secret
1888 * and launch spoofed connections. After that he has to start all over again.
1890 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC
1891 * size an average of 4,823 attempts are required for a 50% chance of success
1892 * to spoof a single syncookie (birthday collision paradox). However the
1893 * attacker is blind and doesn't know if one of his attempts succeeded unless
1894 * he has a side channel to interfere success from. A single connection setup
1895 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
1896 * This many attempts are required for each one blind spoofed connection. For
1897 * every additional spoofed connection he has to launch another N attempts.
1898 * Thus for a sustained rate 100 spoofed connections per second approximately
1899 * 1,800,000 packets per second would have to be sent.
1901 * NB: The MAC function should be fast so that it doesn't become a CPU
1902 * exhaustion attack vector itself.
1905 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations
1906 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
1907 * http://cr.yp.to/syncookies.html (overview)
1908 * http://cr.yp.to/syncookies/archive (details)
1911 * Schematic construction of a syncookie enabled Initial Sequence Number:
1913 * 12345678901234567890123456789012
1914 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
1916 * x 24 MAC (truncated)
1917 * W 3 Send Window Scale index
1919 * S 1 SACK permitted
1920 * P 1 Odd/even secret
1924 * Distribution and probability of certain MSS values. Those in between are
1925 * rounded down to the next lower one.
1926 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
1927 * .2% .3% 5% 7% 7% 20% 15% 45%
1929 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
1932 * Distribution and probability of certain WSCALE values. We have to map the
1933 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
1934 * bits based on prevalence of certain values. Where we don't have an exact
1935 * match for are rounded down to the next lower one letting us under-estimate
1936 * the true available window. At the moment this would happen only for the
1937 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
1938 * and window size). The absence of the WSCALE option (no scaling in either
1939 * direction) is encoded with index zero.
1940 * [WSCALE values histograms, Allman, 2012]
1941 * X 10 10 35 5 6 14 10% by host
1942 * X 11 4 5 5 18 49 3% by connections
1944 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
1947 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed
1948 * and good cryptographic properties.
1951 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
1952 uint8_t *secbits, uintptr_t secmod)
1955 uint32_t siphash[2];
1957 SipHash24_Init(&ctx);
1958 SipHash_SetKey(&ctx, secbits);
1959 switch (inc->inc_flags & INC_ISIPV6) {
1962 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
1963 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
1968 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
1969 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
1973 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
1974 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
1975 SipHash_Update(&ctx, &irs, sizeof(irs));
1976 SipHash_Update(&ctx, &flags, sizeof(flags));
1977 SipHash_Update(&ctx, &secmod, sizeof(secmod));
1978 SipHash_Final((u_int8_t *)&siphash, &ctx);
1980 return (siphash[0] ^ siphash[1]);
1984 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
1986 u_int i, mss, secbit, wscale;
1989 union syncookie cookie;
1991 SCH_LOCK_ASSERT(sch);
1995 /* Map our computed MSS into the 3-bit index. */
1996 mss = min(tcp_mssopt(&sc->sc_inc), max(sc->sc_peer_mss, V_tcp_minmss));
1997 for (i = nitems(tcp_sc_msstab) - 1; tcp_sc_msstab[i] > mss && i > 0;
2000 cookie.flags.mss_idx = i;
2003 * Map the send window scale into the 3-bit index but only if
2004 * the wscale option was received.
2006 if (sc->sc_flags & SCF_WINSCALE) {
2007 wscale = sc->sc_requested_s_scale;
2008 for (i = nitems(tcp_sc_wstab) - 1;
2009 tcp_sc_wstab[i] > wscale && i > 0;
2012 cookie.flags.wscale_idx = i;
2015 /* Can we do SACK? */
2016 if (sc->sc_flags & SCF_SACK)
2017 cookie.flags.sack_ok = 1;
2019 /* Which of the two secrets to use. */
2020 secbit = sch->sch_sc->secret.oddeven & 0x1;
2021 cookie.flags.odd_even = secbit;
2023 secbits = sch->sch_sc->secret.key[secbit];
2024 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2028 * Put the flags into the hash and XOR them to get better ISS number
2029 * variance. This doesn't enhance the cryptographic strength and is
2030 * done to prevent the 8 cookie bits from showing up directly on the
2034 iss |= cookie.cookie ^ (hash >> 24);
2036 /* Randomize the timestamp. */
2037 if (sc->sc_flags & SCF_TIMESTAMP) {
2038 sc->sc_ts = arc4random();
2039 sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks();
2042 TCPSTAT_INC(tcps_sc_sendcookie);
2046 static struct syncache *
2047 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2048 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2054 int wnd, wscale = 0;
2055 union syncookie cookie;
2057 SCH_LOCK_ASSERT(sch);
2060 * Pull information out of SYN-ACK/ACK and revert sequence number
2063 ack = th->th_ack - 1;
2064 seq = th->th_seq - 1;
2067 * Unpack the flags containing enough information to restore the
2070 cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2072 /* Which of the two secrets to use. */
2073 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
2075 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2077 /* The recomputed hash matches the ACK if this was a genuine cookie. */
2078 if ((ack & ~0xff) != (hash & ~0xff))
2081 /* Fill in the syncache values. */
2083 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2084 sc->sc_ipopts = NULL;
2089 switch (inc->inc_flags & INC_ISIPV6) {
2092 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2093 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2098 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2099 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
2104 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2106 /* We can simply recompute receive window scale we sent earlier. */
2107 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2110 /* Only use wscale if it was enabled in the orignal SYN. */
2111 if (cookie.flags.wscale_idx > 0) {
2112 sc->sc_requested_r_scale = wscale;
2113 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2114 sc->sc_flags |= SCF_WINSCALE;
2117 wnd = sbspace(&lso->so_rcv);
2119 wnd = imin(wnd, TCP_MAXWIN);
2122 if (cookie.flags.sack_ok)
2123 sc->sc_flags |= SCF_SACK;
2125 if (to->to_flags & TOF_TS) {
2126 sc->sc_flags |= SCF_TIMESTAMP;
2127 sc->sc_tsreflect = to->to_tsval;
2128 sc->sc_ts = to->to_tsecr;
2129 sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks();
2132 if (to->to_flags & TOF_SIGNATURE)
2133 sc->sc_flags |= SCF_SIGNATURE;
2137 TCPSTAT_INC(tcps_sc_recvcookie);
2143 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2144 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2147 struct syncache scs, *scx;
2150 bzero(&scs, sizeof(scs));
2151 scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
2153 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2157 if (sc->sc_peer_mss != scx->sc_peer_mss)
2158 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2159 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2161 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2162 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2163 s, __func__, sc->sc_requested_r_scale,
2164 scx->sc_requested_r_scale);
2166 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2167 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2168 s, __func__, sc->sc_requested_s_scale,
2169 scx->sc_requested_s_scale);
2171 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2172 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2179 #endif /* INVARIANTS */
2182 syncookie_reseed(void *arg)
2184 struct tcp_syncache *sc = arg;
2189 * Reseeding the secret doesn't have to be protected by a lock.
2190 * It only must be ensured that the new random values are visible
2191 * to all CPUs in a SMP environment. The atomic with release
2192 * semantics ensures that.
2194 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2195 secbits = sc->secret.key[secbit];
2196 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2197 atomic_add_rel_int(&sc->secret.oddeven, 1);
2199 /* Reschedule ourself. */
2200 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2204 * Exports the syncache entries to userland so that netstat can display
2205 * them alongside the other sockets. This function is intended to be
2206 * called only from tcp_pcblist.
2208 * Due to concurrency on an active system, the number of pcbs exported
2209 * may have no relation to max_pcbs. max_pcbs merely indicates the
2210 * amount of space the caller allocated for this function to use.
2213 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
2216 struct syncache *sc;
2217 struct syncache_head *sch;
2218 int count, error, i;
2220 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
2221 sch = &V_tcp_syncache.hashbase[i];
2223 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2224 if (count >= max_pcbs) {
2228 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2230 bzero(&xt, sizeof(xt));
2231 xt.xt_len = sizeof(xt);
2232 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2233 xt.xt_inp.inp_vflag = INP_IPV6;
2235 xt.xt_inp.inp_vflag = INP_IPV4;
2236 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2237 sizeof (struct in_conninfo));
2238 xt.t_state = TCPS_SYN_RECEIVED;
2239 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2240 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2241 xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2242 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2243 error = SYSCTL_OUT(req, &xt, sizeof xt);
2253 *pcbs_exported = count;