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>
43 #include <sys/kernel.h>
44 #include <sys/sysctl.h>
45 #include <sys/limits.h>
47 #include <sys/mutex.h>
48 #include <sys/malloc.h>
50 #include <sys/proc.h> /* for proc0 declaration */
51 #include <sys/random.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #include <sys/syslog.h>
55 #include <sys/ucred.h>
58 #include <crypto/siphash/siphash.h>
63 #include <net/if_var.h>
64 #include <net/route.h>
67 #include <netinet/in.h>
68 #include <netinet/in_systm.h>
69 #include <netinet/ip.h>
70 #include <netinet/in_var.h>
71 #include <netinet/in_pcb.h>
72 #include <netinet/ip_var.h>
73 #include <netinet/ip_options.h>
75 #include <netinet/ip6.h>
76 #include <netinet/icmp6.h>
77 #include <netinet6/nd6.h>
78 #include <netinet6/ip6_var.h>
79 #include <netinet6/in6_pcb.h>
81 #include <netinet/tcp.h>
82 #include <netinet/tcp_fsm.h>
83 #include <netinet/tcp_seq.h>
84 #include <netinet/tcp_timer.h>
85 #include <netinet/tcp_var.h>
86 #include <netinet/tcp_syncache.h>
88 #include <netinet6/tcp6_var.h>
91 #include <netinet/toecore.h>
95 #include <netipsec/ipsec.h>
97 #include <netipsec/ipsec6.h>
99 #include <netipsec/key.h>
102 #include <machine/in_cksum.h>
104 #include <security/mac/mac_framework.h>
106 static VNET_DEFINE(int, tcp_syncookies) = 1;
107 #define V_tcp_syncookies VNET(tcp_syncookies)
108 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_RW,
109 &VNET_NAME(tcp_syncookies), 0,
110 "Use TCP SYN cookies if the syncache overflows");
112 static VNET_DEFINE(int, tcp_syncookiesonly) = 0;
113 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly)
114 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_RW,
115 &VNET_NAME(tcp_syncookiesonly), 0,
116 "Use only TCP SYN cookies");
119 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
122 static void syncache_drop(struct syncache *, struct syncache_head *);
123 static void syncache_free(struct syncache *);
124 static void syncache_insert(struct syncache *, struct syncache_head *);
125 static int syncache_respond(struct syncache *, struct syncache_head *, int);
126 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_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RDTUN,
164 &VNET_NAME(tcp_syncache.bucket_limit), 0,
165 "Per-bucket hash limit for syncache");
167 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, 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_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RDTUN,
175 &VNET_NAME(tcp_syncache.hashsize), 0,
176 "Size of TCP syncache hashtable");
178 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, 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_VNET_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
184 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 SYNCACHE_HASH(inc, mask) \
190 ((V_tcp_syncache.hash_secret ^ \
191 (inc)->inc_faddr.s_addr ^ \
192 ((inc)->inc_faddr.s_addr >> 16) ^ \
193 (inc)->inc_fport ^ (inc)->inc_lport) & mask)
195 #define SYNCACHE_HASH6(inc, mask) \
196 ((V_tcp_syncache.hash_secret ^ \
197 (inc)->inc6_faddr.s6_addr32[0] ^ \
198 (inc)->inc6_faddr.s6_addr32[3] ^ \
199 (inc)->inc_fport ^ (inc)->inc_lport) & mask)
201 #define ENDPTS_EQ(a, b) ( \
202 (a)->ie_fport == (b)->ie_fport && \
203 (a)->ie_lport == (b)->ie_lport && \
204 (a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr && \
205 (a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr \
208 #define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0)
210 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx)
211 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx)
212 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED)
215 * Requires the syncache entry to be already removed from the bucket list.
218 syncache_free(struct syncache *sc)
222 (void) m_free(sc->sc_ipopts);
226 mac_syncache_destroy(&sc->sc_label);
229 uma_zfree(V_tcp_syncache.zone, sc);
237 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
238 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
239 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
240 V_tcp_syncache.hash_secret = arc4random();
242 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
243 &V_tcp_syncache.hashsize);
244 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
245 &V_tcp_syncache.bucket_limit);
246 if (!powerof2(V_tcp_syncache.hashsize) ||
247 V_tcp_syncache.hashsize == 0) {
248 printf("WARNING: syncache hash size is not a power of 2.\n");
249 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
251 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
254 V_tcp_syncache.cache_limit =
255 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
256 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
257 &V_tcp_syncache.cache_limit);
259 /* Allocate the hash table. */
260 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
261 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
264 V_tcp_syncache.vnet = curvnet;
267 /* Initialize the hash buckets. */
268 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
269 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
270 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
272 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
273 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
274 V_tcp_syncache.hashbase[i].sch_length = 0;
275 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
278 /* Create the syncache entry zone. */
279 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
280 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
281 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
282 V_tcp_syncache.cache_limit);
284 /* Start the SYN cookie reseeder callout. */
285 callout_init(&V_tcp_syncache.secret.reseed, 1);
286 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
287 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
288 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
289 syncookie_reseed, &V_tcp_syncache);
294 syncache_destroy(void)
296 struct syncache_head *sch;
297 struct syncache *sc, *nsc;
300 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */
301 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
303 sch = &V_tcp_syncache.hashbase[i];
304 callout_drain(&sch->sch_timer);
307 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
308 syncache_drop(sc, sch);
310 KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
311 ("%s: sch->sch_bucket not empty", __func__));
312 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
313 __func__, sch->sch_length));
314 mtx_destroy(&sch->sch_mtx);
317 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
318 ("%s: cache_count not 0", __func__));
320 /* Free the allocated global resources. */
321 uma_zdestroy(V_tcp_syncache.zone);
322 free(V_tcp_syncache.hashbase, M_SYNCACHE);
324 callout_drain(&V_tcp_syncache.secret.reseed);
329 * Inserts a syncache entry into the specified bucket row.
330 * Locks and unlocks the syncache_head autonomously.
333 syncache_insert(struct syncache *sc, struct syncache_head *sch)
335 struct syncache *sc2;
340 * Make sure that we don't overflow the per-bucket limit.
341 * If the bucket is full, toss the oldest element.
343 if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
344 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
345 ("sch->sch_length incorrect"));
346 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
347 syncache_drop(sc2, sch);
348 TCPSTAT_INC(tcps_sc_bucketoverflow);
351 /* Put it into the bucket. */
352 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
356 if (ADDED_BY_TOE(sc)) {
357 struct toedev *tod = sc->sc_tod;
359 tod->tod_syncache_added(tod, sc->sc_todctx);
363 /* Reinitialize the bucket row's timer. */
364 if (sch->sch_length == 1)
365 sch->sch_nextc = ticks + INT_MAX;
366 syncache_timeout(sc, sch, 1);
370 TCPSTAT_INC(tcps_sc_added);
374 * Remove and free entry from syncache bucket row.
375 * Expects locked syncache head.
378 syncache_drop(struct syncache *sc, struct syncache_head *sch)
381 SCH_LOCK_ASSERT(sch);
383 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
387 if (ADDED_BY_TOE(sc)) {
388 struct toedev *tod = sc->sc_tod;
390 tod->tod_syncache_removed(tod, sc->sc_todctx);
398 * Engage/reengage time on bucket row.
401 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
403 sc->sc_rxttime = ticks +
404 TCPTV_RTOBASE * (tcp_syn_backoff[sc->sc_rxmits]);
406 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
407 sch->sch_nextc = sc->sc_rxttime;
409 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
410 syncache_timer, (void *)sch);
415 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
416 * If we have retransmitted an entry the maximum number of times, expire it.
417 * One separate timer for each bucket row.
420 syncache_timer(void *xsch)
422 struct syncache_head *sch = (struct syncache_head *)xsch;
423 struct syncache *sc, *nsc;
427 CURVNET_SET(sch->sch_sc->vnet);
429 /* NB: syncache_head has already been locked by the callout. */
430 SCH_LOCK_ASSERT(sch);
433 * In the following cycle we may remove some entries and/or
434 * advance some timeouts, so re-initialize the bucket timer.
436 sch->sch_nextc = tick + INT_MAX;
438 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
440 * We do not check if the listen socket still exists
441 * and accept the case where the listen socket may be
442 * gone by the time we resend the SYN/ACK. We do
443 * not expect this to happens often. If it does,
444 * then the RST will be sent by the time the remote
445 * host does the SYN/ACK->ACK.
447 if (TSTMP_GT(sc->sc_rxttime, tick)) {
448 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
449 sch->sch_nextc = sc->sc_rxttime;
452 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
453 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
454 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
455 "giving up and removing syncache entry\n",
459 syncache_drop(sc, sch);
460 TCPSTAT_INC(tcps_sc_stale);
463 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
464 log(LOG_DEBUG, "%s; %s: Response timeout, "
465 "retransmitting (%u) SYN|ACK\n",
466 s, __func__, sc->sc_rxmits);
470 syncache_respond(sc, sch, 1);
471 TCPSTAT_INC(tcps_sc_retransmitted);
472 syncache_timeout(sc, sch, 0);
474 if (!TAILQ_EMPTY(&(sch)->sch_bucket))
475 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
476 syncache_timer, (void *)(sch));
481 * Find an entry in the syncache.
482 * Returns always with locked syncache_head plus a matching entry or NULL.
484 static struct syncache *
485 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
488 struct syncache_head *sch;
491 if (inc->inc_flags & INC_ISIPV6) {
492 sch = &V_tcp_syncache.hashbase[
493 SYNCACHE_HASH6(inc, V_tcp_syncache.hashmask)];
498 /* Circle through bucket row to find matching entry. */
499 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
500 if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
506 sch = &V_tcp_syncache.hashbase[
507 SYNCACHE_HASH(inc, V_tcp_syncache.hashmask)];
512 /* Circle through bucket row to find matching entry. */
513 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
515 if (sc->sc_inc.inc_flags & INC_ISIPV6)
518 if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
522 SCH_LOCK_ASSERT(*schp);
523 return (NULL); /* always returns with locked sch */
527 * This function is called when we get a RST for a
528 * non-existent connection, so that we can see if the
529 * connection is in the syn cache. If it is, zap it.
532 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
535 struct syncache_head *sch;
538 sc = syncache_lookup(inc, &sch); /* returns locked sch */
539 SCH_LOCK_ASSERT(sch);
542 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
543 * See RFC 793 page 65, section SEGMENT ARRIVES.
545 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
546 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
547 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
548 "FIN flag set, segment ignored\n", s, __func__);
549 TCPSTAT_INC(tcps_badrst);
554 * No corresponding connection was found in syncache.
555 * If syncookies are enabled and possibly exclusively
556 * used, or we are under memory pressure, a valid RST
557 * may not find a syncache entry. In that case we're
558 * done and no SYN|ACK retransmissions will happen.
559 * Otherwise the RST was misdirected or spoofed.
562 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
563 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
564 "syncache entry (possibly syncookie only), "
565 "segment ignored\n", s, __func__);
566 TCPSTAT_INC(tcps_badrst);
571 * If the RST bit is set, check the sequence number to see
572 * if this is a valid reset segment.
574 * In all states except SYN-SENT, all reset (RST) segments
575 * are validated by checking their SEQ-fields. A reset is
576 * valid if its sequence number is in the window.
578 * The sequence number in the reset segment is normally an
579 * echo of our outgoing acknowlegement numbers, but some hosts
580 * send a reset with the sequence number at the rightmost edge
581 * of our receive window, and we have to handle this case.
583 if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
584 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
585 syncache_drop(sc, sch);
586 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
587 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, "
588 "connection attempt aborted by remote endpoint\n",
590 TCPSTAT_INC(tcps_sc_reset);
592 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
593 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
594 "IRS %u (+WND %u), segment ignored\n",
595 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd);
596 TCPSTAT_INC(tcps_badrst);
606 syncache_badack(struct in_conninfo *inc)
609 struct syncache_head *sch;
611 sc = syncache_lookup(inc, &sch); /* returns locked sch */
612 SCH_LOCK_ASSERT(sch);
614 syncache_drop(sc, sch);
615 TCPSTAT_INC(tcps_sc_badack);
621 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th)
624 struct syncache_head *sch;
626 sc = syncache_lookup(inc, &sch); /* returns locked sch */
627 SCH_LOCK_ASSERT(sch);
631 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
632 if (ntohl(th->th_seq) != sc->sc_iss)
636 * If we've rertransmitted 3 times and this is our second error,
637 * we remove the entry. Otherwise, we allow it to continue on.
638 * This prevents us from incorrectly nuking an entry during a
639 * spurious network outage.
643 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
644 sc->sc_flags |= SCF_UNREACH;
647 syncache_drop(sc, sch);
648 TCPSTAT_INC(tcps_sc_unreach);
654 * Build a new TCP socket structure from a syncache entry.
656 static struct socket *
657 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
659 struct inpcb *inp = NULL;
665 INP_INFO_WLOCK_ASSERT(&V_tcbinfo);
668 * Ok, create the full blown connection, and set things up
669 * as they would have been set up if we had created the
670 * connection when the SYN arrived. If we can't create
671 * the connection, abort it.
673 so = sonewconn(lso, 0);
676 * Drop the connection; we will either send a RST or
677 * have the peer retransmit its SYN again after its
680 TCPSTAT_INC(tcps_listendrop);
681 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
682 log(LOG_DEBUG, "%s; %s: Socket create failed "
683 "due to limits or memory shortage\n",
690 mac_socketpeer_set_from_mbuf(m, so);
694 inp->inp_inc.inc_fibnum = so->so_fibnum;
696 INP_HASH_WLOCK(&V_tcbinfo);
698 /* Insert new socket into PCB hash list. */
699 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
701 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
702 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
704 inp->inp_vflag &= ~INP_IPV6;
705 inp->inp_vflag |= INP_IPV4;
707 inp->inp_laddr = sc->sc_inc.inc_laddr;
713 * If there's an mbuf and it has a flowid, then let's initialise the
714 * inp with that particular flowid.
716 if (m != NULL && m->m_flags & M_FLOWID) {
717 inp->inp_flags |= INP_HW_FLOWID;
718 inp->inp_flags &= ~INP_SW_FLOWID;
719 inp->inp_flowid = m->m_pkthdr.flowid;
720 inp->inp_flowtype = M_HASHTYPE_GET(m);
724 * Install in the reservation hash table for now, but don't yet
725 * install a connection group since the full 4-tuple isn't yet
728 inp->inp_lport = sc->sc_inc.inc_lport;
729 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
731 * Undo the assignments above if we failed to
732 * put the PCB on the hash lists.
735 if (sc->sc_inc.inc_flags & INC_ISIPV6)
736 inp->in6p_laddr = in6addr_any;
739 inp->inp_laddr.s_addr = INADDR_ANY;
741 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
742 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
747 INP_HASH_WUNLOCK(&V_tcbinfo);
751 /* Copy old policy into new socket's. */
752 if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp))
753 printf("syncache_socket: could not copy policy\n");
756 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
757 struct inpcb *oinp = sotoinpcb(lso);
758 struct in6_addr laddr6;
759 struct sockaddr_in6 sin6;
761 * Inherit socket options from the listening socket.
762 * Note that in6p_inputopts are not (and should not be)
763 * copied, since it stores previously received options and is
764 * used to detect if each new option is different than the
765 * previous one and hence should be passed to a user.
766 * If we copied in6p_inputopts, a user would not be able to
767 * receive options just after calling the accept system call.
769 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
770 if (oinp->in6p_outputopts)
771 inp->in6p_outputopts =
772 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
774 sin6.sin6_family = AF_INET6;
775 sin6.sin6_len = sizeof(sin6);
776 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
777 sin6.sin6_port = sc->sc_inc.inc_fport;
778 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
779 laddr6 = inp->in6p_laddr;
780 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
781 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
782 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
783 thread0.td_ucred, m)) != 0) {
784 inp->in6p_laddr = laddr6;
785 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
786 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
791 INP_HASH_WUNLOCK(&V_tcbinfo);
794 /* Override flowlabel from in6_pcbconnect. */
795 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
796 inp->inp_flow |= sc->sc_flowlabel;
799 #if defined(INET) && defined(INET6)
804 struct in_addr laddr;
805 struct sockaddr_in sin;
807 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
809 if (inp->inp_options == NULL) {
810 inp->inp_options = sc->sc_ipopts;
811 sc->sc_ipopts = NULL;
814 sin.sin_family = AF_INET;
815 sin.sin_len = sizeof(sin);
816 sin.sin_addr = sc->sc_inc.inc_faddr;
817 sin.sin_port = sc->sc_inc.inc_fport;
818 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
819 laddr = inp->inp_laddr;
820 if (inp->inp_laddr.s_addr == INADDR_ANY)
821 inp->inp_laddr = sc->sc_inc.inc_laddr;
822 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
823 thread0.td_ucred, m)) != 0) {
824 inp->inp_laddr = laddr;
825 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
826 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
831 INP_HASH_WUNLOCK(&V_tcbinfo);
836 INP_HASH_WUNLOCK(&V_tcbinfo);
838 tcp_state_change(tp, TCPS_SYN_RECEIVED);
839 tp->iss = sc->sc_iss;
840 tp->irs = sc->sc_irs;
843 tp->snd_wl1 = sc->sc_irs;
844 tp->snd_max = tp->iss + 1;
845 tp->snd_nxt = tp->iss + 1;
846 tp->rcv_up = sc->sc_irs + 1;
847 tp->rcv_wnd = sc->sc_wnd;
848 tp->rcv_adv += tp->rcv_wnd;
849 tp->last_ack_sent = tp->rcv_nxt;
851 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
852 if (sc->sc_flags & SCF_NOOPT)
853 tp->t_flags |= TF_NOOPT;
855 if (sc->sc_flags & SCF_WINSCALE) {
856 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
857 tp->snd_scale = sc->sc_requested_s_scale;
858 tp->request_r_scale = sc->sc_requested_r_scale;
860 if (sc->sc_flags & SCF_TIMESTAMP) {
861 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
862 tp->ts_recent = sc->sc_tsreflect;
863 tp->ts_recent_age = tcp_ts_getticks();
864 tp->ts_offset = sc->sc_tsoff;
867 if (sc->sc_flags & SCF_SIGNATURE)
868 tp->t_flags |= TF_SIGNATURE;
870 if (sc->sc_flags & SCF_SACK)
871 tp->t_flags |= TF_SACK_PERMIT;
874 if (sc->sc_flags & SCF_ECN)
875 tp->t_flags |= TF_ECN_PERMIT;
878 * Set up MSS and get cached values from tcp_hostcache.
879 * This might overwrite some of the defaults we just set.
881 tcp_mss(tp, sc->sc_peer_mss);
884 * If the SYN,ACK was retransmitted, indicate that CWND to be
885 * limited to one segment in cc_conn_init().
886 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
888 if (sc->sc_rxmits > 1)
893 * Allow a TOE driver to install its hooks. Note that we hold the
894 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
895 * new connection before the TOE driver has done its thing.
897 if (ADDED_BY_TOE(sc)) {
898 struct toedev *tod = sc->sc_tod;
900 tod->tod_offload_socket(tod, sc->sc_todctx, so);
904 * Copy and activate timers.
906 tp->t_keepinit = sototcpcb(lso)->t_keepinit;
907 tp->t_keepidle = sototcpcb(lso)->t_keepidle;
908 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
909 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
910 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
916 TCPSTAT_INC(tcps_accepts);
928 * This function gets called when we receive an ACK for a
929 * socket in the LISTEN state. We look up the connection
930 * in the syncache, and if its there, we pull it out of
931 * the cache and turn it into a full-blown connection in
932 * the SYN-RECEIVED state.
935 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
936 struct socket **lsop, struct mbuf *m)
939 struct syncache_head *sch;
944 * Global TCP locks are held because we manipulate the PCB lists
945 * and create a new socket.
947 INP_INFO_WLOCK_ASSERT(&V_tcbinfo);
948 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
949 ("%s: can handle only ACK", __func__));
951 sc = syncache_lookup(inc, &sch); /* returns locked sch */
952 SCH_LOCK_ASSERT(sch);
956 * Test code for syncookies comparing the syncache stored
957 * values with the reconstructed values from the cookie.
960 syncookie_cmp(inc, sch, sc, th, to, *lsop);
965 * There is no syncache entry, so see if this ACK is
966 * a returning syncookie. To do this, first:
967 * A. See if this socket has had a syncache entry dropped in
968 * the past. We don't want to accept a bogus syncookie
969 * if we've never received a SYN.
970 * B. check that the syncookie is valid. If it is, then
971 * cobble up a fake syncache entry, and return.
973 if (!V_tcp_syncookies) {
975 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
976 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
977 "segment rejected (syncookies disabled)\n",
981 bzero(&scs, sizeof(scs));
982 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
985 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
986 log(LOG_DEBUG, "%s; %s: Segment failed "
987 "SYNCOOKIE authentication, segment rejected "
988 "(probably spoofed)\n", s, __func__);
992 /* Pull out the entry to unlock the bucket row. */
993 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
996 if (ADDED_BY_TOE(sc)) {
997 struct toedev *tod = sc->sc_tod;
999 tod->tod_syncache_removed(tod, sc->sc_todctx);
1006 * Segment validation:
1007 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1009 if (th->th_ack != sc->sc_iss + 1) {
1010 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1011 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1012 "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1017 * The SEQ must fall in the window starting at the received
1018 * initial receive sequence number + 1 (the SYN).
1020 if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1021 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1022 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1023 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1024 "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1029 * If timestamps were not negotiated during SYN/ACK they
1030 * must not appear on any segment during this session.
1032 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1033 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1034 log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1035 "segment rejected\n", s, __func__);
1040 * If timestamps were negotiated during SYN/ACK they should
1041 * appear on every segment during this session.
1042 * XXXAO: This is only informal as there have been unverified
1043 * reports of non-compliants stacks.
1045 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1046 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1047 log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1048 "no action\n", s, __func__);
1055 * If timestamps were negotiated the reflected timestamp
1056 * must be equal to what we actually sent in the SYN|ACK.
1058 if ((to->to_flags & TOF_TS) && to->to_tsecr != sc->sc_ts) {
1059 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1060 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, "
1061 "segment rejected\n",
1062 s, __func__, to->to_tsecr, sc->sc_ts);
1066 *lsop = syncache_socket(sc, *lsop, m);
1069 TCPSTAT_INC(tcps_sc_aborted);
1071 TCPSTAT_INC(tcps_sc_completed);
1073 /* how do we find the inp for the new socket? */
1078 if (sc != NULL && sc != &scs)
1087 * Given a LISTEN socket and an inbound SYN request, add
1088 * this to the syn cache, and send back a segment:
1089 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1092 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1093 * Doing so would require that we hold onto the data and deliver it
1094 * to the application. However, if we are the target of a SYN-flood
1095 * DoS attack, an attacker could send data which would eventually
1096 * consume all available buffer space if it were ACKed. By not ACKing
1097 * the data, we avoid this DoS scenario.
1100 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1101 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1106 struct syncache *sc = NULL;
1107 struct syncache_head *sch;
1108 struct mbuf *ipopts = NULL;
1110 int win, sb_hiwat, ip_ttl, ip_tos;
1113 int autoflowlabel = 0;
1116 struct label *maclabel;
1118 struct syncache scs;
1121 INP_WLOCK_ASSERT(inp); /* listen socket */
1122 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1123 ("%s: unexpected tcp flags", __func__));
1126 * Combine all so/tp operations very early to drop the INP lock as
1131 cred = crhold(so->so_cred);
1134 if ((inc->inc_flags & INC_ISIPV6) &&
1135 (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1138 ip_ttl = inp->inp_ip_ttl;
1139 ip_tos = inp->inp_ip_tos;
1140 win = sbspace(&so->so_rcv);
1141 sb_hiwat = so->so_rcv.sb_hiwat;
1142 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1144 /* By the time we drop the lock these should no longer be used. */
1149 if (mac_syncache_init(&maclabel) != 0) {
1153 mac_syncache_create(maclabel, inp);
1158 * Remember the IP options, if any.
1161 if (!(inc->inc_flags & INC_ISIPV6))
1164 ipopts = (m) ? ip_srcroute(m) : NULL;
1170 * See if we already have an entry for this connection.
1171 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1173 * XXX: should the syncache be re-initialized with the contents
1174 * of the new SYN here (which may have different options?)
1176 * XXX: We do not check the sequence number to see if this is a
1177 * real retransmit or a new connection attempt. The question is
1178 * how to handle such a case; either ignore it as spoofed, or
1179 * drop the current entry and create a new one?
1181 sc = syncache_lookup(inc, &sch); /* returns locked entry */
1182 SCH_LOCK_ASSERT(sch);
1184 TCPSTAT_INC(tcps_sc_dupsyn);
1187 * If we were remembering a previous source route,
1188 * forget it and use the new one we've been given.
1191 (void) m_free(sc->sc_ipopts);
1192 sc->sc_ipopts = ipopts;
1195 * Update timestamp if present.
1197 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1198 sc->sc_tsreflect = to->to_tsval;
1200 sc->sc_flags &= ~SCF_TIMESTAMP;
1203 * Since we have already unconditionally allocated label
1204 * storage, free it up. The syncache entry will already
1205 * have an initialized label we can use.
1207 mac_syncache_destroy(&maclabel);
1209 /* Retransmit SYN|ACK and reset retransmit count. */
1210 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1211 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1212 "resetting timer and retransmitting SYN|ACK\n",
1216 if (syncache_respond(sc, sch, 1) == 0) {
1218 syncache_timeout(sc, sch, 1);
1219 TCPSTAT_INC(tcps_sndacks);
1220 TCPSTAT_INC(tcps_sndtotal);
1226 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1229 * The zone allocator couldn't provide more entries.
1230 * Treat this as if the cache was full; drop the oldest
1231 * entry and insert the new one.
1233 TCPSTAT_INC(tcps_sc_zonefail);
1234 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL)
1235 syncache_drop(sc, sch);
1236 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1238 if (V_tcp_syncookies) {
1239 bzero(&scs, sizeof(scs));
1244 (void) m_free(ipopts);
1251 * Fill in the syncache values.
1254 sc->sc_label = maclabel;
1258 sc->sc_ipopts = ipopts;
1259 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1261 if (!(inc->inc_flags & INC_ISIPV6))
1264 sc->sc_ip_tos = ip_tos;
1265 sc->sc_ip_ttl = ip_ttl;
1269 sc->sc_todctx = todctx;
1271 sc->sc_irs = th->th_seq;
1272 sc->sc_iss = arc4random();
1274 sc->sc_flowlabel = 0;
1277 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1278 * win was derived from socket earlier in the function.
1281 win = imin(win, TCP_MAXWIN);
1284 if (V_tcp_do_rfc1323) {
1286 * A timestamp received in a SYN makes
1287 * it ok to send timestamp requests and replies.
1289 if (to->to_flags & TOF_TS) {
1290 sc->sc_tsreflect = to->to_tsval;
1291 sc->sc_ts = tcp_ts_getticks();
1292 sc->sc_flags |= SCF_TIMESTAMP;
1294 if (to->to_flags & TOF_SCALE) {
1298 * Pick the smallest possible scaling factor that
1299 * will still allow us to scale up to sb_max, aka
1300 * kern.ipc.maxsockbuf.
1302 * We do this because there are broken firewalls that
1303 * will corrupt the window scale option, leading to
1304 * the other endpoint believing that our advertised
1305 * window is unscaled. At scale factors larger than
1306 * 5 the unscaled window will drop below 1500 bytes,
1307 * leading to serious problems when traversing these
1310 * With the default maxsockbuf of 256K, a scale factor
1311 * of 3 will be chosen by this algorithm. Those who
1312 * choose a larger maxsockbuf should watch out
1313 * for the compatiblity problems mentioned above.
1315 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1316 * or <SYN,ACK>) segment itself is never scaled.
1318 while (wscale < TCP_MAX_WINSHIFT &&
1319 (TCP_MAXWIN << wscale) < sb_max)
1321 sc->sc_requested_r_scale = wscale;
1322 sc->sc_requested_s_scale = to->to_wscale;
1323 sc->sc_flags |= SCF_WINSCALE;
1326 #ifdef TCP_SIGNATURE
1328 * If listening socket requested TCP digests, OR received SYN
1329 * contains the option, flag this in the syncache so that
1330 * syncache_respond() will do the right thing with the SYN+ACK.
1332 if (to->to_flags & TOF_SIGNATURE || ltflags & TF_SIGNATURE)
1333 sc->sc_flags |= SCF_SIGNATURE;
1335 if (to->to_flags & TOF_SACKPERM)
1336 sc->sc_flags |= SCF_SACK;
1337 if (to->to_flags & TOF_MSS)
1338 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */
1339 if (ltflags & TF_NOOPT)
1340 sc->sc_flags |= SCF_NOOPT;
1341 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1342 sc->sc_flags |= SCF_ECN;
1344 if (V_tcp_syncookies)
1345 sc->sc_iss = syncookie_generate(sch, sc);
1347 if (autoflowlabel) {
1348 if (V_tcp_syncookies)
1349 sc->sc_flowlabel = sc->sc_iss;
1351 sc->sc_flowlabel = ip6_randomflowlabel();
1352 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1358 * Do a standard 3-way handshake.
1360 if (syncache_respond(sc, sch, 0) == 0) {
1361 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1363 else if (sc != &scs)
1364 syncache_insert(sc, sch); /* locks and unlocks sch */
1365 TCPSTAT_INC(tcps_sndacks);
1366 TCPSTAT_INC(tcps_sndtotal);
1370 TCPSTAT_INC(tcps_sc_dropped);
1378 mac_syncache_destroy(&maclabel);
1388 syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked)
1390 struct ip *ip = NULL;
1392 struct tcphdr *th = NULL;
1393 int optlen, error = 0; /* Make compiler happy */
1394 u_int16_t hlen, tlen, mssopt;
1397 struct ip6_hdr *ip6 = NULL;
1399 #ifdef TCP_SIGNATURE
1400 struct secasvar *sav;
1405 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1408 tlen = hlen + sizeof(struct tcphdr);
1410 /* Determine MSS we advertize to other end of connection. */
1411 mssopt = tcp_mssopt(&sc->sc_inc);
1412 if (sc->sc_peer_mss)
1413 mssopt = max( min(sc->sc_peer_mss, mssopt), V_tcp_minmss);
1415 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1416 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1417 ("syncache: mbuf too small"));
1419 /* Create the IP+TCP header from scratch. */
1420 m = m_gethdr(M_NOWAIT, MT_DATA);
1424 mac_syncache_create_mbuf(sc->sc_label, m);
1426 m->m_data += max_linkhdr;
1428 m->m_pkthdr.len = tlen;
1429 m->m_pkthdr.rcvif = NULL;
1432 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1433 ip6 = mtod(m, struct ip6_hdr *);
1434 ip6->ip6_vfc = IPV6_VERSION;
1435 ip6->ip6_nxt = IPPROTO_TCP;
1436 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1437 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1438 ip6->ip6_plen = htons(tlen - hlen);
1439 /* ip6_hlim is set after checksum */
1440 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1441 ip6->ip6_flow |= sc->sc_flowlabel;
1443 th = (struct tcphdr *)(ip6 + 1);
1446 #if defined(INET6) && defined(INET)
1451 ip = mtod(m, struct ip *);
1452 ip->ip_v = IPVERSION;
1453 ip->ip_hl = sizeof(struct ip) >> 2;
1454 ip->ip_len = htons(tlen);
1458 ip->ip_p = IPPROTO_TCP;
1459 ip->ip_src = sc->sc_inc.inc_laddr;
1460 ip->ip_dst = sc->sc_inc.inc_faddr;
1461 ip->ip_ttl = sc->sc_ip_ttl;
1462 ip->ip_tos = sc->sc_ip_tos;
1465 * See if we should do MTU discovery. Route lookups are
1466 * expensive, so we will only unset the DF bit if:
1468 * 1) path_mtu_discovery is disabled
1469 * 2) the SCF_UNREACH flag has been set
1471 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1472 ip->ip_off |= htons(IP_DF);
1474 th = (struct tcphdr *)(ip + 1);
1477 th->th_sport = sc->sc_inc.inc_lport;
1478 th->th_dport = sc->sc_inc.inc_fport;
1480 th->th_seq = htonl(sc->sc_iss);
1481 th->th_ack = htonl(sc->sc_irs + 1);
1482 th->th_off = sizeof(struct tcphdr) >> 2;
1484 th->th_flags = TH_SYN|TH_ACK;
1485 th->th_win = htons(sc->sc_wnd);
1488 if (sc->sc_flags & SCF_ECN) {
1489 th->th_flags |= TH_ECE;
1490 TCPSTAT_INC(tcps_ecn_shs);
1493 /* Tack on the TCP options. */
1494 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1498 to.to_flags = TOF_MSS;
1499 if (sc->sc_flags & SCF_WINSCALE) {
1500 to.to_wscale = sc->sc_requested_r_scale;
1501 to.to_flags |= TOF_SCALE;
1503 if (sc->sc_flags & SCF_TIMESTAMP) {
1504 /* Virgin timestamp or TCP cookie enhanced one. */
1505 to.to_tsval = sc->sc_ts;
1506 to.to_tsecr = sc->sc_tsreflect;
1507 to.to_flags |= TOF_TS;
1509 if (sc->sc_flags & SCF_SACK)
1510 to.to_flags |= TOF_SACKPERM;
1511 #ifdef TCP_SIGNATURE
1513 if (sc->sc_flags & SCF_SIGNATURE) {
1514 sav = tcp_get_sav(m, IPSEC_DIR_OUTBOUND);
1516 to.to_flags |= TOF_SIGNATURE;
1520 * We've got SCF_SIGNATURE flag
1521 * inherited from listening socket,
1522 * but to SADB key for given source
1523 * address. Assume signature is not
1524 * required and remove signature flag
1525 * instead of silently dropping
1530 sc->sc_flags &= ~SCF_SIGNATURE;
1536 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1538 /* Adjust headers by option size. */
1539 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1541 m->m_pkthdr.len += optlen;
1543 #ifdef TCP_SIGNATURE
1544 if (sc->sc_flags & SCF_SIGNATURE)
1545 tcp_signature_do_compute(m, 0, optlen,
1546 to.to_signature, sav);
1549 if (sc->sc_inc.inc_flags & INC_ISIPV6)
1550 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1553 ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1557 M_SETFIB(m, sc->sc_inc.inc_fibnum);
1558 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1560 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1561 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1562 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1564 ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1566 if (ADDED_BY_TOE(sc)) {
1567 struct toedev *tod = sc->sc_tod;
1569 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1574 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1577 #if defined(INET6) && defined(INET)
1582 m->m_pkthdr.csum_flags = CSUM_TCP;
1583 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1584 htons(tlen + optlen - hlen + IPPROTO_TCP));
1586 if (ADDED_BY_TOE(sc)) {
1587 struct toedev *tod = sc->sc_tod;
1589 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1594 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1601 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1602 * that exceed the capacity of the syncache by avoiding the storage of any
1603 * of the SYNs we receive. Syncookies defend against blind SYN flooding
1604 * attacks where the attacker does not have access to our responses.
1606 * Syncookies encode and include all necessary information about the
1607 * connection setup within the SYN|ACK that we send back. That way we
1608 * can avoid keeping any local state until the ACK to our SYN|ACK returns
1609 * (if ever). Normally the syncache and syncookies are running in parallel
1610 * with the latter taking over when the former is exhausted. When matching
1611 * syncache entry is found the syncookie is ignored.
1613 * The only reliable information persisting the 3WHS is our inital sequence
1614 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient
1615 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1616 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK
1617 * returns and signifies a legitimate connection if it matches the ACK.
1619 * The available space of 32 bits to store the hash and to encode the SYN
1620 * option information is very tight and we should have at least 24 bits for
1621 * the MAC to keep the number of guesses by blind spoofing reasonably high.
1623 * SYN option information we have to encode to fully restore a connection:
1624 * MSS: is imporant to chose an optimal segment size to avoid IP level
1625 * fragmentation along the path. The common MSS values can be encoded
1626 * in a 3-bit table. Uncommon values are captured by the next lower value
1627 * in the table leading to a slight increase in packetization overhead.
1628 * WSCALE: is necessary to allow large windows to be used for high delay-
1629 * bandwidth product links. Not scaling the window when it was initially
1630 * negotiated is bad for performance as lack of scaling further decreases
1631 * the apparent available send window. We only need to encode the WSCALE
1632 * we received from the remote end. Our end can be recalculated at any
1633 * time. The common WSCALE values can be encoded in a 3-bit table.
1634 * Uncommon values are captured by the next lower value in the table
1635 * making us under-estimate the available window size halving our
1636 * theoretically possible maximum throughput for that connection.
1637 * SACK: Greatly assists in packet loss recovery and requires 1 bit.
1638 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
1639 * that are included in all segments on a connection. We enable them when
1642 * Security of syncookies and attack vectors:
1644 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
1645 * together with the gloabl secret to make it unique per connection attempt.
1646 * Thus any change of any of those parameters results in a different MAC output
1647 * in an unpredictable way unless a collision is encountered. 24 bits of the
1648 * MAC are embedded into the ISS.
1650 * To prevent replay attacks two rotating global secrets are updated with a
1651 * new random value every 15 seconds. The life-time of a syncookie is thus
1654 * Vector 1: Attacking the secret. This requires finding a weakness in the
1655 * MAC itself or the way it is used here. The attacker can do a chosen plain
1656 * text attack by varying and testing the all parameters under his control.
1657 * The strength depends on the size and randomness of the secret, and the
1658 * cryptographic security of the MAC function. Due to the constant updating
1659 * of the secret the attacker has at most 29.999 seconds to find the secret
1660 * and launch spoofed connections. After that he has to start all over again.
1662 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC
1663 * size an average of 4,823 attempts are required for a 50% chance of success
1664 * to spoof a single syncookie (birthday collision paradox). However the
1665 * attacker is blind and doesn't know if one of his attempts succeeded unless
1666 * he has a side channel to interfere success from. A single connection setup
1667 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
1668 * This many attempts are required for each one blind spoofed connection. For
1669 * every additional spoofed connection he has to launch another N attempts.
1670 * Thus for a sustained rate 100 spoofed connections per second approximately
1671 * 1,800,000 packets per second would have to be sent.
1673 * NB: The MAC function should be fast so that it doesn't become a CPU
1674 * exhaustion attack vector itself.
1677 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations
1678 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
1679 * http://cr.yp.to/syncookies.html (overview)
1680 * http://cr.yp.to/syncookies/archive (details)
1683 * Schematic construction of a syncookie enabled Initial Sequence Number:
1685 * 12345678901234567890123456789012
1686 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
1688 * x 24 MAC (truncated)
1689 * W 3 Send Window Scale index
1691 * S 1 SACK permitted
1692 * P 1 Odd/even secret
1696 * Distribution and probability of certain MSS values. Those in between are
1697 * rounded down to the next lower one.
1698 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
1699 * .2% .3% 5% 7% 7% 20% 15% 45%
1701 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
1704 * Distribution and probability of certain WSCALE values. We have to map the
1705 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
1706 * bits based on prevalence of certain values. Where we don't have an exact
1707 * match for are rounded down to the next lower one letting us under-estimate
1708 * the true available window. At the moment this would happen only for the
1709 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
1710 * and window size). The absence of the WSCALE option (no scaling in either
1711 * direction) is encoded with index zero.
1712 * [WSCALE values histograms, Allman, 2012]
1713 * X 10 10 35 5 6 14 10% by host
1714 * X 11 4 5 5 18 49 3% by connections
1716 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
1719 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed
1720 * and good cryptographic properties.
1723 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
1724 uint8_t *secbits, uintptr_t secmod)
1727 uint32_t siphash[2];
1729 SipHash24_Init(&ctx);
1730 SipHash_SetKey(&ctx, secbits);
1731 switch (inc->inc_flags & INC_ISIPV6) {
1734 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
1735 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
1740 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
1741 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
1745 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
1746 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
1747 SipHash_Update(&ctx, &flags, sizeof(flags));
1748 SipHash_Update(&ctx, &secmod, sizeof(secmod));
1749 SipHash_Final((u_int8_t *)&siphash, &ctx);
1751 return (siphash[0] ^ siphash[1]);
1755 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
1757 u_int i, mss, secbit, wscale;
1760 union syncookie cookie;
1762 SCH_LOCK_ASSERT(sch);
1766 /* Map our computed MSS into the 3-bit index. */
1767 mss = min(tcp_mssopt(&sc->sc_inc), max(sc->sc_peer_mss, V_tcp_minmss));
1768 for (i = sizeof(tcp_sc_msstab) / sizeof(*tcp_sc_msstab) - 1;
1769 tcp_sc_msstab[i] > mss && i > 0;
1772 cookie.flags.mss_idx = i;
1775 * Map the send window scale into the 3-bit index but only if
1776 * the wscale option was received.
1778 if (sc->sc_flags & SCF_WINSCALE) {
1779 wscale = sc->sc_requested_s_scale;
1780 for (i = sizeof(tcp_sc_wstab) / sizeof(*tcp_sc_wstab) - 1;
1781 tcp_sc_wstab[i] > wscale && i > 0;
1784 cookie.flags.wscale_idx = i;
1787 /* Can we do SACK? */
1788 if (sc->sc_flags & SCF_SACK)
1789 cookie.flags.sack_ok = 1;
1791 /* Which of the two secrets to use. */
1792 secbit = sch->sch_sc->secret.oddeven & 0x1;
1793 cookie.flags.odd_even = secbit;
1795 secbits = sch->sch_sc->secret.key[secbit];
1796 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
1800 * Put the flags into the hash and XOR them to get better ISS number
1801 * variance. This doesn't enhance the cryptographic strength and is
1802 * done to prevent the 8 cookie bits from showing up directly on the
1806 iss |= cookie.cookie ^ (hash >> 24);
1808 /* Randomize the timestamp. */
1809 if (sc->sc_flags & SCF_TIMESTAMP) {
1810 sc->sc_ts = arc4random();
1811 sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks();
1814 TCPSTAT_INC(tcps_sc_sendcookie);
1818 static struct syncache *
1819 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
1820 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
1826 int wnd, wscale = 0;
1827 union syncookie cookie;
1829 SCH_LOCK_ASSERT(sch);
1832 * Pull information out of SYN-ACK/ACK and revert sequence number
1835 ack = th->th_ack - 1;
1836 seq = th->th_seq - 1;
1839 * Unpack the flags containing enough information to restore the
1842 cookie.cookie = (ack & 0xff) ^ (ack >> 24);
1844 /* Which of the two secrets to use. */
1845 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
1847 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
1849 /* The recomputed hash matches the ACK if this was a genuine cookie. */
1850 if ((ack & ~0xff) != (hash & ~0xff))
1853 /* Fill in the syncache values. */
1855 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1856 sc->sc_ipopts = NULL;
1861 switch (inc->inc_flags & INC_ISIPV6) {
1864 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
1865 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
1870 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
1871 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
1876 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
1878 /* We can simply recompute receive window scale we sent earlier. */
1879 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
1882 /* Only use wscale if it was enabled in the orignal SYN. */
1883 if (cookie.flags.wscale_idx > 0) {
1884 sc->sc_requested_r_scale = wscale;
1885 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
1886 sc->sc_flags |= SCF_WINSCALE;
1889 wnd = sbspace(&lso->so_rcv);
1891 wnd = imin(wnd, TCP_MAXWIN);
1894 if (cookie.flags.sack_ok)
1895 sc->sc_flags |= SCF_SACK;
1897 if (to->to_flags & TOF_TS) {
1898 sc->sc_flags |= SCF_TIMESTAMP;
1899 sc->sc_tsreflect = to->to_tsval;
1900 sc->sc_ts = to->to_tsecr;
1901 sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks();
1904 if (to->to_flags & TOF_SIGNATURE)
1905 sc->sc_flags |= SCF_SIGNATURE;
1909 TCPSTAT_INC(tcps_sc_recvcookie);
1915 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
1916 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
1919 struct syncache scs, *scx;
1922 bzero(&scs, sizeof(scs));
1923 scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
1925 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
1929 if (sc->sc_peer_mss != scx->sc_peer_mss)
1930 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
1931 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
1933 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
1934 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
1935 s, __func__, sc->sc_requested_r_scale,
1936 scx->sc_requested_r_scale);
1938 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
1939 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
1940 s, __func__, sc->sc_requested_s_scale,
1941 scx->sc_requested_s_scale);
1943 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
1944 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
1951 #endif /* INVARIANTS */
1954 syncookie_reseed(void *arg)
1956 struct tcp_syncache *sc = arg;
1961 * Reseeding the secret doesn't have to be protected by a lock.
1962 * It only must be ensured that the new random values are visible
1963 * to all CPUs in a SMP environment. The atomic with release
1964 * semantics ensures that.
1966 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
1967 secbits = sc->secret.key[secbit];
1968 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
1969 atomic_add_rel_int(&sc->secret.oddeven, 1);
1971 /* Reschedule ourself. */
1972 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
1976 * Returns the current number of syncache entries. This number
1977 * will probably change before you get around to calling
1981 syncache_pcbcount(void)
1983 struct syncache_head *sch;
1986 for (count = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
1987 /* No need to lock for a read. */
1988 sch = &V_tcp_syncache.hashbase[i];
1989 count += sch->sch_length;
1995 * Exports the syncache entries to userland so that netstat can display
1996 * them alongside the other sockets. This function is intended to be
1997 * called only from tcp_pcblist.
1999 * Due to concurrency on an active system, the number of pcbs exported
2000 * may have no relation to max_pcbs. max_pcbs merely indicates the
2001 * amount of space the caller allocated for this function to use.
2004 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
2007 struct syncache *sc;
2008 struct syncache_head *sch;
2009 int count, error, i;
2011 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
2012 sch = &V_tcp_syncache.hashbase[i];
2014 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2015 if (count >= max_pcbs) {
2019 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2021 bzero(&xt, sizeof(xt));
2022 xt.xt_len = sizeof(xt);
2023 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2024 xt.xt_inp.inp_vflag = INP_IPV6;
2026 xt.xt_inp.inp_vflag = INP_IPV4;
2027 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo));
2028 xt.xt_tp.t_inpcb = &xt.xt_inp;
2029 xt.xt_tp.t_state = TCPS_SYN_RECEIVED;
2030 xt.xt_socket.xso_protocol = IPPROTO_TCP;
2031 xt.xt_socket.xso_len = sizeof (struct xsocket);
2032 xt.xt_socket.so_type = SOCK_STREAM;
2033 xt.xt_socket.so_state = SS_ISCONNECTING;
2034 error = SYSCTL_OUT(req, &xt, sizeof xt);
2044 *pcbs_exported = count;