xref: /freebsd/sys/netinet/tcp_syncache.c (revision f0a75d274af375d15b97b830966b99a02b7db911)
1 /*-
2  * Copyright (c) 2001 McAfee, Inc.
3  * Copyright (c) 2006 Andre Oppermann, Internet Business Solutions AG
4  * All rights reserved.
5  *
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.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
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.
19  *
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
30  * SUCH DAMAGE.
31  *
32  * $FreeBSD$
33  */
34 
35 #include "opt_inet.h"
36 #include "opt_inet6.h"
37 #include "opt_ipsec.h"
38 #include "opt_mac.h"
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/kernel.h>
43 #include <sys/sysctl.h>
44 #include <sys/lock.h>
45 #include <sys/mutex.h>
46 #include <sys/malloc.h>
47 #include <sys/mbuf.h>
48 #include <sys/md5.h>
49 #include <sys/proc.h>		/* for proc0 declaration */
50 #include <sys/random.h>
51 #include <sys/socket.h>
52 #include <sys/socketvar.h>
53 
54 #include <vm/uma.h>
55 
56 #include <net/if.h>
57 #include <net/route.h>
58 
59 #include <netinet/in.h>
60 #include <netinet/in_systm.h>
61 #include <netinet/ip.h>
62 #include <netinet/in_var.h>
63 #include <netinet/in_pcb.h>
64 #include <netinet/ip_var.h>
65 #include <netinet/ip_options.h>
66 #ifdef INET6
67 #include <netinet/ip6.h>
68 #include <netinet/icmp6.h>
69 #include <netinet6/nd6.h>
70 #include <netinet6/ip6_var.h>
71 #include <netinet6/in6_pcb.h>
72 #endif
73 #include <netinet/tcp.h>
74 #include <netinet/tcp_fsm.h>
75 #include <netinet/tcp_seq.h>
76 #include <netinet/tcp_timer.h>
77 #include <netinet/tcp_var.h>
78 #ifdef INET6
79 #include <netinet6/tcp6_var.h>
80 #endif
81 
82 #ifdef IPSEC
83 #include <netinet6/ipsec.h>
84 #ifdef INET6
85 #include <netinet6/ipsec6.h>
86 #endif
87 #endif /*IPSEC*/
88 
89 #ifdef FAST_IPSEC
90 #include <netipsec/ipsec.h>
91 #ifdef INET6
92 #include <netipsec/ipsec6.h>
93 #endif
94 #include <netipsec/key.h>
95 #endif /*FAST_IPSEC*/
96 
97 #include <machine/in_cksum.h>
98 
99 #include <security/mac/mac_framework.h>
100 
101 static int tcp_syncookies = 1;
102 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_RW,
103     &tcp_syncookies, 0,
104     "Use TCP SYN cookies if the syncache overflows");
105 
106 static int tcp_syncookiesonly = 0;
107 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_RW,
108     &tcp_syncookiesonly, 0,
109     "Use only TCP SYN cookies");
110 
111 #define	SYNCOOKIE_SECRET_SIZE	8	/* dwords */
112 #define	SYNCOOKIE_LIFETIME	16	/* seconds */
113 
114 struct syncache {
115 	TAILQ_ENTRY(syncache)	sc_hash;
116 	struct		in_conninfo sc_inc;	/* addresses */
117 	u_long		sc_rxttime;		/* retransmit time */
118 	u_int16_t	sc_rxmits;		/* retransmit counter */
119 
120 	u_int32_t	sc_tsreflect;		/* timestamp to reflect */
121 	u_int32_t	sc_ts;			/* our timestamp to send */
122 	u_int32_t	sc_tsoff;		/* ts offset w/ syncookies */
123 	u_int32_t	sc_flowlabel;		/* IPv6 flowlabel */
124 	tcp_seq		sc_irs;			/* seq from peer */
125 	tcp_seq		sc_iss;			/* our ISS */
126 	struct		mbuf *sc_ipopts;	/* source route */
127 
128 	u_int16_t	sc_peer_mss;		/* peer's MSS */
129 	u_int16_t	sc_wnd;			/* advertised window */
130 	u_int8_t	sc_ip_ttl;		/* IPv4 TTL */
131 	u_int8_t	sc_ip_tos;		/* IPv4 TOS */
132 	u_int8_t	sc_requested_s_scale:4,
133 			sc_requested_r_scale:4;
134 	u_int8_t	sc_flags;
135 #define SCF_NOOPT	0x01			/* no TCP options */
136 #define SCF_WINSCALE	0x02			/* negotiated window scaling */
137 #define SCF_TIMESTAMP	0x04			/* negotiated timestamps */
138 						/* MSS is implicit */
139 #define SCF_UNREACH	0x10			/* icmp unreachable received */
140 #define SCF_SIGNATURE	0x20			/* send MD5 digests */
141 #define SCF_SACK	0x80			/* send SACK option */
142 #ifdef MAC
143 	struct label	*sc_label;		/* MAC label reference */
144 #endif
145 };
146 
147 struct syncache_head {
148 	struct mtx	sch_mtx;
149 	TAILQ_HEAD(sch_head, syncache)	sch_bucket;
150 	struct callout	sch_timer;
151 	int		sch_nextc;
152 	u_int		sch_length;
153 	u_int		sch_oddeven;
154 	u_int32_t	sch_secbits_odd[SYNCOOKIE_SECRET_SIZE];
155 	u_int32_t	sch_secbits_even[SYNCOOKIE_SECRET_SIZE];
156 	u_int		sch_reseed;		/* time_uptime, seconds */
157 };
158 
159 static void	 syncache_drop(struct syncache *, struct syncache_head *);
160 static void	 syncache_free(struct syncache *);
161 static void	 syncache_insert(struct syncache *, struct syncache_head *);
162 struct syncache *syncache_lookup(struct in_conninfo *, struct syncache_head **);
163 static int	 syncache_respond(struct syncache *, struct mbuf *);
164 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
165 		    struct mbuf *m);
166 static void	 syncache_timer(void *);
167 static void	 syncookie_generate(struct syncache_head *, struct syncache *,
168 		    u_int32_t *);
169 static struct syncache
170 		*syncookie_lookup(struct in_conninfo *, struct syncache_head *,
171 		    struct syncache *, struct tcpopt *, struct tcphdr *,
172 		    struct socket *);
173 
174 /*
175  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
176  * 3 retransmits corresponds to a timeout of (1 + 2 + 4 + 8 == 15) seconds,
177  * the odds are that the user has given up attempting to connect by then.
178  */
179 #define SYNCACHE_MAXREXMTS		3
180 
181 /* Arbitrary values */
182 #define TCP_SYNCACHE_HASHSIZE		512
183 #define TCP_SYNCACHE_BUCKETLIMIT	30
184 
185 struct tcp_syncache {
186 	struct	syncache_head *hashbase;
187 	uma_zone_t zone;
188 	u_int	hashsize;
189 	u_int	hashmask;
190 	u_int	bucket_limit;
191 	u_int	cache_count;		/* XXX: unprotected */
192 	u_int	cache_limit;
193 	u_int	rexmt_limit;
194 	u_int	hash_secret;
195 };
196 static struct tcp_syncache tcp_syncache;
197 
198 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, "TCP SYN cache");
199 
200 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RDTUN,
201      &tcp_syncache.bucket_limit, 0, "Per-bucket hash limit for syncache");
202 
203 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_RDTUN,
204      &tcp_syncache.cache_limit, 0, "Overall entry limit for syncache");
205 
206 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_RD,
207      &tcp_syncache.cache_count, 0, "Current number of entries in syncache");
208 
209 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RDTUN,
210      &tcp_syncache.hashsize, 0, "Size of TCP syncache hashtable");
211 
212 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_RW,
213      &tcp_syncache.rexmt_limit, 0, "Limit on SYN/ACK retransmissions");
214 
215 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
216 
217 #define SYNCACHE_HASH(inc, mask)					\
218 	((tcp_syncache.hash_secret ^					\
219 	  (inc)->inc_faddr.s_addr ^					\
220 	  ((inc)->inc_faddr.s_addr >> 16) ^				\
221 	  (inc)->inc_fport ^ (inc)->inc_lport) & mask)
222 
223 #define SYNCACHE_HASH6(inc, mask)					\
224 	((tcp_syncache.hash_secret ^					\
225 	  (inc)->inc6_faddr.s6_addr32[0] ^				\
226 	  (inc)->inc6_faddr.s6_addr32[3] ^				\
227 	  (inc)->inc_fport ^ (inc)->inc_lport) & mask)
228 
229 #define ENDPTS_EQ(a, b) (						\
230 	(a)->ie_fport == (b)->ie_fport &&				\
231 	(a)->ie_lport == (b)->ie_lport &&				\
232 	(a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr &&			\
233 	(a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr			\
234 )
235 
236 #define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0)
237 
238 #define SYNCACHE_TIMEOUT(sc, sch, co) do {				\
239 	(sc)->sc_rxmits++;						\
240 	(sc)->sc_rxttime = ticks +					\
241 		TCPTV_RTOBASE * tcp_backoff[(sc)->sc_rxmits - 1];	\
242 	if ((sch)->sch_nextc > (sc)->sc_rxttime)			\
243 		(sch)->sch_nextc = (sc)->sc_rxttime;			\
244 	if (!TAILQ_EMPTY(&(sch)->sch_bucket) && !(co))			\
245 		callout_reset(&(sch)->sch_timer,			\
246 			(sch)->sch_nextc - ticks,			\
247 			syncache_timer, (void *)(sch));			\
248 } while (0)
249 
250 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
251 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
252 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
253 
254 /*
255  * Requires the syncache entry to be already removed from the bucket list.
256  */
257 static void
258 syncache_free(struct syncache *sc)
259 {
260 	if (sc->sc_ipopts)
261 		(void) m_free(sc->sc_ipopts);
262 #ifdef MAC
263 	mac_destroy_syncache(&sc->sc_label);
264 #endif
265 
266 	uma_zfree(tcp_syncache.zone, sc);
267 }
268 
269 void
270 syncache_init(void)
271 {
272 	int i;
273 
274 	tcp_syncache.cache_count = 0;
275 	tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
276 	tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
277 	tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
278 	tcp_syncache.hash_secret = arc4random();
279 
280 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
281 	    &tcp_syncache.hashsize);
282 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
283 	    &tcp_syncache.bucket_limit);
284 	if (!powerof2(tcp_syncache.hashsize) || tcp_syncache.hashsize == 0) {
285 		printf("WARNING: syncache hash size is not a power of 2.\n");
286 		tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
287 	}
288 	tcp_syncache.hashmask = tcp_syncache.hashsize - 1;
289 
290 	/* Set limits. */
291 	tcp_syncache.cache_limit =
292 	    tcp_syncache.hashsize * tcp_syncache.bucket_limit;
293 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
294 	    &tcp_syncache.cache_limit);
295 
296 	/* Allocate the hash table. */
297 	MALLOC(tcp_syncache.hashbase, struct syncache_head *,
298 	    tcp_syncache.hashsize * sizeof(struct syncache_head),
299 	    M_SYNCACHE, M_WAITOK | M_ZERO);
300 
301 	/* Initialize the hash buckets. */
302 	for (i = 0; i < tcp_syncache.hashsize; i++) {
303 		TAILQ_INIT(&tcp_syncache.hashbase[i].sch_bucket);
304 		mtx_init(&tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
305 			 NULL, MTX_DEF);
306 		callout_init_mtx(&tcp_syncache.hashbase[i].sch_timer,
307 			 &tcp_syncache.hashbase[i].sch_mtx, 0);
308 		tcp_syncache.hashbase[i].sch_length = 0;
309 	}
310 
311 	/* Create the syncache entry zone. */
312 	tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
313 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
314 	uma_zone_set_max(tcp_syncache.zone, tcp_syncache.cache_limit);
315 }
316 
317 /*
318  * Inserts a syncache entry into the specified bucket row.
319  * Locks and unlocks the syncache_head autonomously.
320  */
321 static void
322 syncache_insert(struct syncache *sc, struct syncache_head *sch)
323 {
324 	struct syncache *sc2;
325 
326 	SCH_LOCK(sch);
327 
328 	/*
329 	 * Make sure that we don't overflow the per-bucket limit.
330 	 * If the bucket is full, toss the oldest element.
331 	 */
332 	if (sch->sch_length >= tcp_syncache.bucket_limit) {
333 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
334 			("sch->sch_length incorrect"));
335 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
336 		syncache_drop(sc2, sch);
337 		tcpstat.tcps_sc_bucketoverflow++;
338 	}
339 
340 	/* Put it into the bucket. */
341 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
342 	sch->sch_length++;
343 
344 	/* Reinitialize the bucket row's timer. */
345 	SYNCACHE_TIMEOUT(sc, sch, 1);
346 
347 	SCH_UNLOCK(sch);
348 
349 	tcp_syncache.cache_count++;
350 	tcpstat.tcps_sc_added++;
351 }
352 
353 /*
354  * Remove and free entry from syncache bucket row.
355  * Expects locked syncache head.
356  */
357 static void
358 syncache_drop(struct syncache *sc, struct syncache_head *sch)
359 {
360 
361 	SCH_LOCK_ASSERT(sch);
362 
363 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
364 	sch->sch_length--;
365 
366 	syncache_free(sc);
367 	tcp_syncache.cache_count--;
368 }
369 
370 /*
371  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
372  * If we have retransmitted an entry the maximum number of times, expire it.
373  * One separate timer for each bucket row.
374  */
375 static void
376 syncache_timer(void *xsch)
377 {
378 	struct syncache_head *sch = (struct syncache_head *)xsch;
379 	struct syncache *sc, *nsc;
380 	int tick = ticks;
381 
382 	/* NB: syncache_head has already been locked by the callout. */
383 	SCH_LOCK_ASSERT(sch);
384 
385 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
386 		/*
387 		 * We do not check if the listen socket still exists
388 		 * and accept the case where the listen socket may be
389 		 * gone by the time we resend the SYN/ACK.  We do
390 		 * not expect this to happens often. If it does,
391 		 * then the RST will be sent by the time the remote
392 		 * host does the SYN/ACK->ACK.
393 		 */
394 		if (sc->sc_rxttime >= tick) {
395 			if (sc->sc_rxttime < sch->sch_nextc)
396 				sch->sch_nextc = sc->sc_rxttime;
397 			continue;
398 		}
399 
400 		if (sc->sc_rxmits > tcp_syncache.rexmt_limit) {
401 			syncache_drop(sc, sch);
402 			tcpstat.tcps_sc_stale++;
403 			continue;
404 		}
405 
406 		(void) syncache_respond(sc, NULL);
407 		tcpstat.tcps_sc_retransmitted++;
408 		SYNCACHE_TIMEOUT(sc, sch, 0);
409 	}
410 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
411 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
412 			syncache_timer, (void *)(sch));
413 }
414 
415 /*
416  * Find an entry in the syncache.
417  * Returns always with locked syncache_head plus a matching entry or NULL.
418  */
419 struct syncache *
420 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
421 {
422 	struct syncache *sc;
423 	struct syncache_head *sch;
424 
425 #ifdef INET6
426 	if (inc->inc_isipv6) {
427 		sch = &tcp_syncache.hashbase[
428 		    SYNCACHE_HASH6(inc, tcp_syncache.hashmask)];
429 		*schp = sch;
430 
431 		SCH_LOCK(sch);
432 
433 		/* Circle through bucket row to find matching entry. */
434 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
435 			if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
436 				return (sc);
437 		}
438 	} else
439 #endif
440 	{
441 		sch = &tcp_syncache.hashbase[
442 		    SYNCACHE_HASH(inc, tcp_syncache.hashmask)];
443 		*schp = sch;
444 
445 		SCH_LOCK(sch);
446 
447 		/* Circle through bucket row to find matching entry. */
448 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
449 #ifdef INET6
450 			if (sc->sc_inc.inc_isipv6)
451 				continue;
452 #endif
453 			if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
454 				return (sc);
455 		}
456 	}
457 	SCH_LOCK_ASSERT(*schp);
458 	return (NULL);			/* always returns with locked sch */
459 }
460 
461 /*
462  * This function is called when we get a RST for a
463  * non-existent connection, so that we can see if the
464  * connection is in the syn cache.  If it is, zap it.
465  */
466 void
467 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
468 {
469 	struct syncache *sc;
470 	struct syncache_head *sch;
471 
472 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
473 	SCH_LOCK_ASSERT(sch);
474 	if (sc == NULL)
475 		goto done;
476 
477 	/*
478 	 * If the RST bit is set, check the sequence number to see
479 	 * if this is a valid reset segment.
480 	 * RFC 793 page 37:
481 	 *   In all states except SYN-SENT, all reset (RST) segments
482 	 *   are validated by checking their SEQ-fields.  A reset is
483 	 *   valid if its sequence number is in the window.
484 	 *
485 	 *   The sequence number in the reset segment is normally an
486 	 *   echo of our outgoing acknowlegement numbers, but some hosts
487 	 *   send a reset with the sequence number at the rightmost edge
488 	 *   of our receive window, and we have to handle this case.
489 	 */
490 	if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
491 	    SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
492 		syncache_drop(sc, sch);
493 		tcpstat.tcps_sc_reset++;
494 	}
495 done:
496 	SCH_UNLOCK(sch);
497 }
498 
499 void
500 syncache_badack(struct in_conninfo *inc)
501 {
502 	struct syncache *sc;
503 	struct syncache_head *sch;
504 
505 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
506 	SCH_LOCK_ASSERT(sch);
507 	if (sc != NULL) {
508 		syncache_drop(sc, sch);
509 		tcpstat.tcps_sc_badack++;
510 	}
511 	SCH_UNLOCK(sch);
512 }
513 
514 void
515 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th)
516 {
517 	struct syncache *sc;
518 	struct syncache_head *sch;
519 
520 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
521 	SCH_LOCK_ASSERT(sch);
522 	if (sc == NULL)
523 		goto done;
524 
525 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
526 	if (ntohl(th->th_seq) != sc->sc_iss)
527 		goto done;
528 
529 	/*
530 	 * If we've rertransmitted 3 times and this is our second error,
531 	 * we remove the entry.  Otherwise, we allow it to continue on.
532 	 * This prevents us from incorrectly nuking an entry during a
533 	 * spurious network outage.
534 	 *
535 	 * See tcp_notify().
536 	 */
537 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
538 		sc->sc_flags |= SCF_UNREACH;
539 		goto done;
540 	}
541 	syncache_drop(sc, sch);
542 	tcpstat.tcps_sc_unreach++;
543 done:
544 	SCH_UNLOCK(sch);
545 }
546 
547 /*
548  * Build a new TCP socket structure from a syncache entry.
549  */
550 static struct socket *
551 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
552 {
553 	struct inpcb *inp = NULL;
554 	struct socket *so;
555 	struct tcpcb *tp;
556 
557 	NET_ASSERT_GIANT();
558 	INP_INFO_WLOCK_ASSERT(&tcbinfo);
559 
560 	/*
561 	 * Ok, create the full blown connection, and set things up
562 	 * as they would have been set up if we had created the
563 	 * connection when the SYN arrived.  If we can't create
564 	 * the connection, abort it.
565 	 */
566 	so = sonewconn(lso, SS_ISCONNECTED);
567 	if (so == NULL) {
568 		/*
569 		 * Drop the connection; we will send a RST if the peer
570 		 * retransmits the ACK,
571 		 */
572 		tcpstat.tcps_listendrop++;
573 		goto abort2;
574 	}
575 #ifdef MAC
576 	SOCK_LOCK(so);
577 	mac_set_socket_peer_from_mbuf(m, so);
578 	SOCK_UNLOCK(so);
579 #endif
580 
581 	inp = sotoinpcb(so);
582 	INP_LOCK(inp);
583 
584 	/* Insert new socket into PCB hash list. */
585 	inp->inp_inc.inc_isipv6 = sc->sc_inc.inc_isipv6;
586 #ifdef INET6
587 	if (sc->sc_inc.inc_isipv6) {
588 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
589 	} else {
590 		inp->inp_vflag &= ~INP_IPV6;
591 		inp->inp_vflag |= INP_IPV4;
592 #endif
593 		inp->inp_laddr = sc->sc_inc.inc_laddr;
594 #ifdef INET6
595 	}
596 #endif
597 	inp->inp_lport = sc->sc_inc.inc_lport;
598 	if (in_pcbinshash(inp) != 0) {
599 		/*
600 		 * Undo the assignments above if we failed to
601 		 * put the PCB on the hash lists.
602 		 */
603 #ifdef INET6
604 		if (sc->sc_inc.inc_isipv6)
605 			inp->in6p_laddr = in6addr_any;
606 		else
607 #endif
608 			inp->inp_laddr.s_addr = INADDR_ANY;
609 		inp->inp_lport = 0;
610 		goto abort;
611 	}
612 #ifdef IPSEC
613 	/* Copy old policy into new socket's. */
614 	if (ipsec_copy_pcbpolicy(sotoinpcb(lso)->inp_sp, inp->inp_sp))
615 		printf("syncache_socket: could not copy policy\n");
616 #endif
617 #ifdef FAST_IPSEC
618 	/* Copy old policy into new socket's. */
619 	if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp))
620 		printf("syncache_socket: could not copy policy\n");
621 #endif
622 #ifdef INET6
623 	if (sc->sc_inc.inc_isipv6) {
624 		struct inpcb *oinp = sotoinpcb(lso);
625 		struct in6_addr laddr6;
626 		struct sockaddr_in6 sin6;
627 		/*
628 		 * Inherit socket options from the listening socket.
629 		 * Note that in6p_inputopts are not (and should not be)
630 		 * copied, since it stores previously received options and is
631 		 * used to detect if each new option is different than the
632 		 * previous one and hence should be passed to a user.
633 		 * If we copied in6p_inputopts, a user would not be able to
634 		 * receive options just after calling the accept system call.
635 		 */
636 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
637 		if (oinp->in6p_outputopts)
638 			inp->in6p_outputopts =
639 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
640 
641 		sin6.sin6_family = AF_INET6;
642 		sin6.sin6_len = sizeof(sin6);
643 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
644 		sin6.sin6_port = sc->sc_inc.inc_fport;
645 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
646 		laddr6 = inp->in6p_laddr;
647 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
648 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
649 		if (in6_pcbconnect(inp, (struct sockaddr *)&sin6,
650 		    thread0.td_ucred)) {
651 			inp->in6p_laddr = laddr6;
652 			goto abort;
653 		}
654 		/* Override flowlabel from in6_pcbconnect. */
655 		inp->in6p_flowinfo &= ~IPV6_FLOWLABEL_MASK;
656 		inp->in6p_flowinfo |= sc->sc_flowlabel;
657 	} else
658 #endif
659 	{
660 		struct in_addr laddr;
661 		struct sockaddr_in sin;
662 
663 		inp->inp_options = ip_srcroute(m);
664 		if (inp->inp_options == NULL) {
665 			inp->inp_options = sc->sc_ipopts;
666 			sc->sc_ipopts = NULL;
667 		}
668 
669 		sin.sin_family = AF_INET;
670 		sin.sin_len = sizeof(sin);
671 		sin.sin_addr = sc->sc_inc.inc_faddr;
672 		sin.sin_port = sc->sc_inc.inc_fport;
673 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
674 		laddr = inp->inp_laddr;
675 		if (inp->inp_laddr.s_addr == INADDR_ANY)
676 			inp->inp_laddr = sc->sc_inc.inc_laddr;
677 		if (in_pcbconnect(inp, (struct sockaddr *)&sin,
678 		    thread0.td_ucred)) {
679 			inp->inp_laddr = laddr;
680 			goto abort;
681 		}
682 	}
683 	tp = intotcpcb(inp);
684 	tp->t_state = TCPS_SYN_RECEIVED;
685 	tp->iss = sc->sc_iss;
686 	tp->irs = sc->sc_irs;
687 	tcp_rcvseqinit(tp);
688 	tcp_sendseqinit(tp);
689 	tp->snd_wl1 = sc->sc_irs;
690 	tp->snd_max = tp->iss + 1;
691 	tp->snd_nxt = tp->iss + 1;
692 	tp->rcv_up = sc->sc_irs + 1;
693 	tp->rcv_wnd = sc->sc_wnd;
694 	tp->rcv_adv += tp->rcv_wnd;
695 	tp->last_ack_sent = tp->rcv_nxt;
696 
697 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
698 	if (sc->sc_flags & SCF_NOOPT)
699 		tp->t_flags |= TF_NOOPT;
700 	else {
701 		if (sc->sc_flags & SCF_WINSCALE) {
702 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
703 			tp->snd_scale = sc->sc_requested_s_scale;
704 			tp->request_r_scale = sc->sc_requested_r_scale;
705 		}
706 		if (sc->sc_flags & SCF_TIMESTAMP) {
707 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
708 			tp->ts_recent = sc->sc_tsreflect;
709 			tp->ts_recent_age = ticks;
710 			tp->ts_offset = sc->sc_tsoff;
711 		}
712 #ifdef TCP_SIGNATURE
713 		if (sc->sc_flags & SCF_SIGNATURE)
714 			tp->t_flags |= TF_SIGNATURE;
715 #endif
716 		if (sc->sc_flags & SCF_SACK) {
717 			tp->sack_enable = 1;
718 			tp->t_flags |= TF_SACK_PERMIT;
719 		}
720 	}
721 
722 	/*
723 	 * Set up MSS and get cached values from tcp_hostcache.
724 	 * This might overwrite some of the defaults we just set.
725 	 */
726 	tcp_mss(tp, sc->sc_peer_mss);
727 
728 	/*
729 	 * If the SYN,ACK was retransmitted, reset cwnd to 1 segment.
730 	 */
731 	if (sc->sc_rxmits > 1)
732 		tp->snd_cwnd = tp->t_maxseg;
733 	callout_reset(tp->tt_keep, tcp_keepinit, tcp_timer_keep, tp);
734 
735 	INP_UNLOCK(inp);
736 
737 	tcpstat.tcps_accepts++;
738 	return (so);
739 
740 abort:
741 	INP_UNLOCK(inp);
742 abort2:
743 	if (so != NULL)
744 		soabort(so);
745 	return (NULL);
746 }
747 
748 /*
749  * This function gets called when we receive an ACK for a
750  * socket in the LISTEN state.  We look up the connection
751  * in the syncache, and if its there, we pull it out of
752  * the cache and turn it into a full-blown connection in
753  * the SYN-RECEIVED state.
754  */
755 int
756 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
757     struct socket **lsop, struct mbuf *m)
758 {
759 	struct syncache *sc;
760 	struct syncache_head *sch;
761 	struct socket *so;
762 	struct syncache scs;
763 
764 	/*
765 	 * Global TCP locks are held because we manipulate the PCB lists
766 	 * and create a new socket.
767 	 */
768 	INP_INFO_WLOCK_ASSERT(&tcbinfo);
769 
770 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
771 	SCH_LOCK_ASSERT(sch);
772 	if (sc == NULL) {
773 		/*
774 		 * There is no syncache entry, so see if this ACK is
775 		 * a returning syncookie.  To do this, first:
776 		 *  A. See if this socket has had a syncache entry dropped in
777 		 *     the past.  We don't want to accept a bogus syncookie
778 		 *     if we've never received a SYN.
779 		 *  B. check that the syncookie is valid.  If it is, then
780 		 *     cobble up a fake syncache entry, and return.
781 		 */
782 		if (!tcp_syncookies) {
783 			SCH_UNLOCK(sch);
784 			goto failed;
785 		}
786 		bzero(&scs, sizeof(scs));
787 		sc = syncookie_lookup(inc, sch, &scs, to, th, *lsop);
788 		SCH_UNLOCK(sch);
789 		if (sc == NULL)
790 			goto failed;
791 		tcpstat.tcps_sc_recvcookie++;
792 	} else {
793 		/* Pull out the entry to unlock the bucket row. */
794 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
795 		sch->sch_length--;
796 		tcp_syncache.cache_count--;
797 		SCH_UNLOCK(sch);
798 	}
799 
800 	/*
801 	 * If seg contains an ACK, but not for our SYN/ACK, send a RST.
802 	 */
803 	if (th->th_ack != sc->sc_iss + 1)
804 		goto failed;
805 
806 	so = syncache_socket(sc, *lsop, m);
807 
808 	if (so == NULL) {
809 #if 0
810 resetandabort:
811 		/* XXXjlemon check this - is this correct? */
812 		(void) tcp_respond(NULL, m, m, th,
813 		    th->th_seq + tlen, (tcp_seq)0, TH_RST|TH_ACK);
814 #endif
815 		m_freem(m);			/* XXX: only needed for above */
816 		tcpstat.tcps_sc_aborted++;
817 		if (sc != &scs) {
818 			syncache_insert(sc, sch);  /* try again later */
819 			sc = NULL;
820 		}
821 		goto failed;
822 	} else
823 		tcpstat.tcps_sc_completed++;
824 	*lsop = so;
825 
826 	if (sc != &scs)
827 		syncache_free(sc);
828 	return (1);
829 failed:
830 	if (sc != NULL && sc != &scs)
831 		syncache_free(sc);
832 	return (0);
833 }
834 
835 /*
836  * Given a LISTEN socket and an inbound SYN request, add
837  * this to the syn cache, and send back a segment:
838  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
839  * to the source.
840  *
841  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
842  * Doing so would require that we hold onto the data and deliver it
843  * to the application.  However, if we are the target of a SYN-flood
844  * DoS attack, an attacker could send data which would eventually
845  * consume all available buffer space if it were ACKed.  By not ACKing
846  * the data, we avoid this DoS scenario.
847  */
848 int
849 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
850     struct inpcb *inp, struct socket **lsop, struct mbuf *m)
851 {
852 	struct tcpcb *tp;
853 	struct socket *so;
854 	struct syncache *sc = NULL;
855 	struct syncache_head *sch;
856 	struct mbuf *ipopts = NULL;
857 	u_int32_t flowtmp;
858 	int win, sb_hiwat, ip_ttl, ip_tos, noopt;
859 #ifdef INET6
860 	int autoflowlabel = 0;
861 #endif
862 #ifdef MAC
863 	struct label *maclabel;
864 #endif
865 	struct syncache scs;
866 
867 	INP_INFO_WLOCK_ASSERT(&tcbinfo);
868 	INP_LOCK_ASSERT(inp);			/* listen socket */
869 
870 	/*
871 	 * Combine all so/tp operations very early to drop the INP lock as
872 	 * soon as possible.
873 	 */
874 	so = *lsop;
875 	tp = sototcpcb(so);
876 
877 #ifdef INET6
878 	if (inc->inc_isipv6 &&
879 	    (inp->in6p_flags & IN6P_AUTOFLOWLABEL))
880 		autoflowlabel = 1;
881 #endif
882 	ip_ttl = inp->inp_ip_ttl;
883 	ip_tos = inp->inp_ip_tos;
884 	win = sbspace(&so->so_rcv);
885 	sb_hiwat = so->so_rcv.sb_hiwat;
886 	noopt = (tp->t_flags & TF_NOOPT);
887 
888 	so = NULL;
889 	tp = NULL;
890 
891 #ifdef MAC
892 	if (mac_init_syncache(&maclabel) != 0) {
893 		*lsop = NULL;
894 		INP_UNLOCK(inp);
895 		INP_INFO_WUNLOCK(&tcbinfo);
896 		return (1);
897 	} else
898 		mac_init_syncache_from_inpcb(maclabel, inp);
899 #endif
900 	INP_UNLOCK(inp);
901 	INP_INFO_WUNLOCK(&tcbinfo);
902 
903 	/*
904 	 * Remember the IP options, if any.
905 	 */
906 #ifdef INET6
907 	if (!inc->inc_isipv6)
908 #endif
909 		ipopts = ip_srcroute(m);
910 
911 	/*
912 	 * See if we already have an entry for this connection.
913 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
914 	 *
915 	 * XXX: should the syncache be re-initialized with the contents
916 	 * of the new SYN here (which may have different options?)
917 	 */
918 	sc = syncache_lookup(inc, &sch);	/* returns locked entry */
919 	SCH_LOCK_ASSERT(sch);
920 	if (sc != NULL) {
921 		tcpstat.tcps_sc_dupsyn++;
922 		if (ipopts) {
923 			/*
924 			 * If we were remembering a previous source route,
925 			 * forget it and use the new one we've been given.
926 			 */
927 			if (sc->sc_ipopts)
928 				(void) m_free(sc->sc_ipopts);
929 			sc->sc_ipopts = ipopts;
930 		}
931 		/*
932 		 * Update timestamp if present.
933 		 */
934 		if (sc->sc_flags & SCF_TIMESTAMP)
935 			sc->sc_tsreflect = to->to_tsval;
936 #ifdef MAC
937 		/*
938 		 * Since we have already unconditionally allocated label
939 		 * storage, free it up.  The syncache entry will already
940 		 * have an initialized label we can use.
941 		 */
942 		mac_destroy_syncache(&maclabel);
943 		KASSERT(sc->sc_label != NULL,
944 		    ("%s: label not initialized", __func__));
945 #endif
946 		if (syncache_respond(sc, m) == 0) {
947 			SYNCACHE_TIMEOUT(sc, sch, 1);
948 			tcpstat.tcps_sndacks++;
949 			tcpstat.tcps_sndtotal++;
950 		}
951 		SCH_UNLOCK(sch);
952 		goto done;
953 	}
954 
955 	sc = uma_zalloc(tcp_syncache.zone, M_NOWAIT | M_ZERO);
956 	if (sc == NULL) {
957 		/*
958 		 * The zone allocator couldn't provide more entries.
959 		 * Treat this as if the cache was full; drop the oldest
960 		 * entry and insert the new one.
961 		 */
962 		tcpstat.tcps_sc_zonefail++;
963 		sc = TAILQ_LAST(&sch->sch_bucket, sch_head);
964 		syncache_drop(sc, sch);
965 		sc = uma_zalloc(tcp_syncache.zone, M_NOWAIT | M_ZERO);
966 		if (sc == NULL) {
967 			if (tcp_syncookies) {
968 				bzero(&scs, sizeof(scs));
969 				sc = &scs;
970 			} else {
971 				SCH_UNLOCK(sch);
972 				if (ipopts)
973 					(void) m_free(ipopts);
974 				goto done;
975 			}
976 		}
977 	}
978 
979 	/*
980 	 * Fill in the syncache values.
981 	 */
982 #ifdef MAC
983 	sc->sc_label = maclabel;
984 #endif
985 	sc->sc_ipopts = ipopts;
986 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
987 #ifdef INET6
988 	if (!inc->inc_isipv6)
989 #endif
990 	{
991 		sc->sc_ip_tos = ip_tos;
992 		sc->sc_ip_ttl = ip_ttl;
993 	}
994 
995 	sc->sc_irs = th->th_seq;
996 	sc->sc_iss = arc4random();
997 	sc->sc_flags = 0;
998 	sc->sc_flowlabel = 0;
999 
1000 	/*
1001 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1002 	 * win was derived from socket earlier in the function.
1003 	 */
1004 	win = imax(win, 0);
1005 	win = imin(win, TCP_MAXWIN);
1006 	sc->sc_wnd = win;
1007 
1008 	if (tcp_do_rfc1323) {
1009 		/*
1010 		 * A timestamp received in a SYN makes
1011 		 * it ok to send timestamp requests and replies.
1012 		 */
1013 		if (to->to_flags & TOF_TS) {
1014 			sc->sc_tsreflect = to->to_tsval;
1015 			sc->sc_flags |= SCF_TIMESTAMP;
1016 		}
1017 		if (to->to_flags & TOF_SCALE) {
1018 			int wscale = 0;
1019 
1020 			/*
1021 			 * Compute proper scaling value from buffer space.
1022 			 * Leave enough room for the socket buffer to grow
1023 			 * with auto sizing.  This allows us to scale the
1024 			 * receive buffer over a wide range while not losing
1025 			 * any efficiency or fine granularity.
1026 			 *
1027 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1028 			 * or <SYN,ACK>) segment itself is never scaled.
1029 			 */
1030 			while (wscale < TCP_MAX_WINSHIFT &&
1031 			    (0x1 << wscale) < tcp_minmss)
1032 				wscale++;
1033 			sc->sc_requested_r_scale = wscale;
1034 			sc->sc_requested_s_scale = to->to_wscale;
1035 			sc->sc_flags |= SCF_WINSCALE;
1036 		}
1037 	}
1038 #ifdef TCP_SIGNATURE
1039 	/*
1040 	 * If listening socket requested TCP digests, and received SYN
1041 	 * contains the option, flag this in the syncache so that
1042 	 * syncache_respond() will do the right thing with the SYN+ACK.
1043 	 * XXX: Currently we always record the option by default and will
1044 	 * attempt to use it in syncache_respond().
1045 	 */
1046 	if (to->to_flags & TOF_SIGNATURE)
1047 		sc->sc_flags |= SCF_SIGNATURE;
1048 #endif
1049 	if (to->to_flags & TOF_SACK)
1050 		sc->sc_flags |= SCF_SACK;
1051 	if (to->to_flags & TOF_MSS)
1052 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1053 	if (noopt)
1054 		sc->sc_flags |= SCF_NOOPT;
1055 
1056 	if (tcp_syncookies) {
1057 		syncookie_generate(sch, sc, &flowtmp);
1058 #ifdef INET6
1059 		if (autoflowlabel)
1060 			sc->sc_flowlabel = flowtmp;
1061 #endif
1062 	} else {
1063 #ifdef INET6
1064 		if (autoflowlabel)
1065 			sc->sc_flowlabel =
1066 			    (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK);
1067 #endif
1068 	}
1069 	SCH_UNLOCK(sch);
1070 
1071 	/*
1072 	 * Do a standard 3-way handshake.
1073 	 */
1074 	if (syncache_respond(sc, m) == 0) {
1075 		if (tcp_syncookies && tcp_syncookiesonly && sc != &scs)
1076 			syncache_free(sc);
1077 		else if (sc != &scs)
1078 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1079 #ifdef MAC
1080 		else
1081 			mac_destroy_syncache(&sc->sc_label);
1082 #endif
1083 		tcpstat.tcps_sndacks++;
1084 		tcpstat.tcps_sndtotal++;
1085 	} else {
1086 		if (sc != &scs)
1087 			syncache_free(sc);
1088 #ifdef MAC
1089 		else
1090 			mac_destroy_syncache(&sc->sc_label);
1091 #endif
1092 		tcpstat.tcps_sc_dropped++;
1093 	}
1094 
1095 done:
1096 	*lsop = NULL;
1097 	return (1);
1098 }
1099 
1100 static int
1101 syncache_respond(struct syncache *sc, struct mbuf *m)
1102 {
1103 	struct ip *ip = NULL;
1104 	struct tcphdr *th;
1105 	int optlen, error;
1106 	u_int16_t hlen, tlen, mssopt;
1107 	struct tcpopt to;
1108 #ifdef INET6
1109 	struct ip6_hdr *ip6 = NULL;
1110 #endif
1111 
1112 	hlen =
1113 #ifdef INET6
1114 	       (sc->sc_inc.inc_isipv6) ? sizeof(struct ip6_hdr) :
1115 #endif
1116 		sizeof(struct ip);
1117 	tlen = hlen + sizeof(struct tcphdr);
1118 
1119 	/* Determine MSS we advertize to other end of connection. */
1120 	mssopt = tcp_mssopt(&sc->sc_inc);
1121 	if (sc->sc_peer_mss)
1122 		mssopt = max( min(sc->sc_peer_mss, mssopt), tcp_minmss);
1123 
1124 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1125 	KASSERT(max_linkhdr + tlen + MAX_TCPOPTLEN <= MHLEN,
1126 	    ("syncache: mbuf too small"));
1127 
1128 	/* Create the IP+TCP header from scratch. */
1129 	if (m)
1130 		m_freem(m);
1131 
1132 	m = m_gethdr(M_DONTWAIT, MT_DATA);
1133 	if (m == NULL)
1134 		return (ENOBUFS);
1135 #ifdef MAC
1136 	mac_create_mbuf_from_syncache(sc->sc_label, m);
1137 #endif
1138 	m->m_data += max_linkhdr;
1139 	m->m_len = tlen;
1140 	m->m_pkthdr.len = tlen;
1141 	m->m_pkthdr.rcvif = NULL;
1142 
1143 #ifdef INET6
1144 	if (sc->sc_inc.inc_isipv6) {
1145 		ip6 = mtod(m, struct ip6_hdr *);
1146 		ip6->ip6_vfc = IPV6_VERSION;
1147 		ip6->ip6_nxt = IPPROTO_TCP;
1148 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1149 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1150 		ip6->ip6_plen = htons(tlen - hlen);
1151 		/* ip6_hlim is set after checksum */
1152 		ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1153 		ip6->ip6_flow |= sc->sc_flowlabel;
1154 
1155 		th = (struct tcphdr *)(ip6 + 1);
1156 	} else
1157 #endif
1158 	{
1159 		ip = mtod(m, struct ip *);
1160 		ip->ip_v = IPVERSION;
1161 		ip->ip_hl = sizeof(struct ip) >> 2;
1162 		ip->ip_len = tlen;
1163 		ip->ip_id = 0;
1164 		ip->ip_off = 0;
1165 		ip->ip_sum = 0;
1166 		ip->ip_p = IPPROTO_TCP;
1167 		ip->ip_src = sc->sc_inc.inc_laddr;
1168 		ip->ip_dst = sc->sc_inc.inc_faddr;
1169 		ip->ip_ttl = sc->sc_ip_ttl;
1170 		ip->ip_tos = sc->sc_ip_tos;
1171 
1172 		/*
1173 		 * See if we should do MTU discovery.  Route lookups are
1174 		 * expensive, so we will only unset the DF bit if:
1175 		 *
1176 		 *	1) path_mtu_discovery is disabled
1177 		 *	2) the SCF_UNREACH flag has been set
1178 		 */
1179 		if (path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1180 		       ip->ip_off |= IP_DF;
1181 
1182 		th = (struct tcphdr *)(ip + 1);
1183 	}
1184 	th->th_sport = sc->sc_inc.inc_lport;
1185 	th->th_dport = sc->sc_inc.inc_fport;
1186 
1187 	th->th_seq = htonl(sc->sc_iss);
1188 	th->th_ack = htonl(sc->sc_irs + 1);
1189 	th->th_off = sizeof(struct tcphdr) >> 2;
1190 	th->th_x2 = 0;
1191 	th->th_flags = TH_SYN|TH_ACK;
1192 	th->th_win = htons(sc->sc_wnd);
1193 	th->th_urp = 0;
1194 
1195 	/* Tack on the TCP options. */
1196 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1197 		to.to_flags = 0;
1198 
1199 		to.to_mss = mssopt;
1200 		to.to_flags = TOF_MSS;
1201 		if (sc->sc_flags & SCF_WINSCALE) {
1202 			to.to_wscale = sc->sc_requested_r_scale;
1203 			to.to_flags |= TOF_SCALE;
1204 		}
1205 		if (sc->sc_flags & SCF_TIMESTAMP) {
1206 			/* Virgin timestamp or TCP cookie enhanced one. */
1207 			to.to_tsval = sc->sc_ts ? sc->sc_ts : ticks;
1208 			to.to_tsecr = sc->sc_tsreflect;
1209 			to.to_flags |= TOF_TS;
1210 		}
1211 		if (sc->sc_flags & SCF_SACK)
1212 			to.to_flags |= TOF_SACKPERM;
1213 #ifdef TCP_SIGNATURE
1214 		if (sc->sc_flags & SCF_SIGNATURE)
1215 			to.to_flags |= TOF_SIGNATURE;
1216 #endif
1217 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1218 
1219 #ifdef TCP_SIGNATURE
1220 		tcp_signature_compute(m, sizeof(struct ip), 0, optlen,
1221 		    to.to_signature, IPSEC_DIR_OUTBOUND);
1222 #endif
1223 
1224 		/* Adjust headers by option size. */
1225 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1226 		m->m_len += optlen;
1227 		m->m_pkthdr.len += optlen;
1228 #ifdef INET6
1229 		if (sc->sc_inc.inc_isipv6)
1230 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1231 		else
1232 #endif
1233 			ip->ip_len += optlen;
1234 	} else
1235 		optlen = 0;
1236 
1237 #ifdef INET6
1238 	if (sc->sc_inc.inc_isipv6) {
1239 		th->th_sum = 0;
1240 		th->th_sum = in6_cksum(m, IPPROTO_TCP, hlen,
1241 				       tlen + optlen - hlen);
1242 		ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1243 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1244 	} else
1245 #endif
1246 	{
1247 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1248 		    htons(tlen + optlen - hlen + IPPROTO_TCP));
1249 		m->m_pkthdr.csum_flags = CSUM_TCP;
1250 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1251 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1252 	}
1253 	return (error);
1254 }
1255 
1256 /*
1257  * The purpose of SYN cookies is to avoid keeping track of all SYN's we
1258  * receive and to be able to handle SYN floods from bogus source addresses
1259  * (where we will never receive any reply).  SYN floods try to exhaust all
1260  * our memory and available slots in the SYN cache table to cause a denial
1261  * of service to legitimate users of the local host.
1262  *
1263  * The idea of SYN cookies is to encode and include all necessary information
1264  * about the connection setup state within the SYN-ACK we send back and thus
1265  * to get along without keeping any local state until the ACK to the SYN-ACK
1266  * arrives (if ever).  Everything we need to know should be available from
1267  * the information we encoded in the SYN-ACK.
1268  *
1269  * More information about the theory behind SYN cookies and its first
1270  * discussion and specification can be found at:
1271  *  http://cr.yp.to/syncookies.html    (overview)
1272  *  http://cr.yp.to/syncookies/archive (gory details)
1273  *
1274  * This implementation extends the orginal idea and first implementation
1275  * of FreeBSD by using not only the initial sequence number field to store
1276  * information but also the timestamp field if present.  This way we can
1277  * keep track of the entire state we need to know to recreate the session in
1278  * its original form.  Almost all TCP speakers implement RFC1323 timestamps
1279  * these days.  For those that do not we still have to live with the known
1280  * shortcomings of the ISN only SYN cookies.
1281  *
1282  * Cookie layers:
1283  *
1284  * Initial sequence number we send:
1285  * 31|................................|0
1286  *    DDDDDDDDDDDDDDDDDDDDDDDDDMMMRRRP
1287  *    D = MD5 Digest (first dword)
1288  *    M = MSS index
1289  *    R = Rotation of secret
1290  *    P = Odd or Even secret
1291  *
1292  * The MD5 Digest is computed with over following parameters:
1293  *  a) randomly rotated secret
1294  *  b) struct in_conninfo containing the remote/local ip/port (IPv4&IPv6)
1295  *  c) the received initial sequence number from remote host
1296  *  d) the rotation offset and odd/even bit
1297  *
1298  * Timestamp we send:
1299  * 31|................................|0
1300  *    DDDDDDDDDDDDDDDDDDDDDDSSSSRRRRA5
1301  *    D = MD5 Digest (third dword) (only as filler)
1302  *    S = Requested send window scale
1303  *    R = Requested receive window scale
1304  *    A = SACK allowed
1305  *    5 = TCP-MD5 enabled (not implemented yet)
1306  *    XORed with MD5 Digest (forth dword)
1307  *
1308  * The timestamp isn't cryptographically secure and doesn't need to be.
1309  * The double use of the MD5 digest dwords ties it to a specific remote/
1310  * local host/port, remote initial sequence number and our local time
1311  * limited secret.  A received timestamp is reverted (XORed) and then
1312  * the contained MD5 dword is compared to the computed one to ensure the
1313  * timestamp belongs to the SYN-ACK we sent.  The other parameters may
1314  * have been tampered with but this isn't different from supplying bogus
1315  * values in the SYN in the first place.
1316  *
1317  * Some problems with SYN cookies remain however:
1318  * Consider the problem of a recreated (and retransmitted) cookie.  If the
1319  * original SYN was accepted, the connection is established.  The second
1320  * SYN is inflight, and if it arrives with an ISN that falls within the
1321  * receive window, the connection is killed.
1322  *
1323  * Notes:
1324  * A heuristic to determine when to accept syn cookies is not necessary.
1325  * An ACK flood would cause the syncookie verification to be attempted,
1326  * but a SYN flood causes syncookies to be generated.  Both are of equal
1327  * cost, so there's no point in trying to optimize the ACK flood case.
1328  * Also, if you don't process certain ACKs for some reason, then all someone
1329  * would have to do is launch a SYN and ACK flood at the same time, which
1330  * would stop cookie verification and defeat the entire purpose of syncookies.
1331  */
1332 static int tcp_sc_msstab[] = { 0, 256, 468, 536, 996, 1452, 1460, 8960 };
1333 
1334 static void
1335 syncookie_generate(struct syncache_head *sch, struct syncache *sc,
1336     u_int32_t *flowlabel)
1337 {
1338 	MD5_CTX ctx;
1339 	u_int32_t md5_buffer[MD5_DIGEST_LENGTH / sizeof(u_int32_t)];
1340 	u_int32_t data;
1341 	u_int32_t *secbits;
1342 	u_int off, pmss, mss;
1343 	int i;
1344 
1345 	SCH_LOCK_ASSERT(sch);
1346 
1347 	/* Which of the two secrets to use. */
1348 	secbits = sch->sch_oddeven ?
1349 			sch->sch_secbits_odd : sch->sch_secbits_even;
1350 
1351 	/* Reseed secret if too old. */
1352 	if (sch->sch_reseed < time_uptime) {
1353 		sch->sch_oddeven = sch->sch_oddeven ? 0 : 1;	/* toggle */
1354 		secbits = sch->sch_oddeven ?
1355 				sch->sch_secbits_odd : sch->sch_secbits_even;
1356 		for (i = 0; i < SYNCOOKIE_SECRET_SIZE; i++)
1357 			secbits[i] = arc4random();
1358 		sch->sch_reseed = time_uptime + SYNCOOKIE_LIFETIME;
1359 	}
1360 
1361 	/* Secret rotation offset. */
1362 	off = sc->sc_iss & 0x7;			/* iss was randomized before */
1363 
1364 	/* Maximum segment size calculation. */
1365 	pmss = max( min(sc->sc_peer_mss, tcp_mssopt(&sc->sc_inc)), tcp_minmss);
1366 	for (mss = sizeof(tcp_sc_msstab) / sizeof(int) - 1; mss > 0; mss--)
1367 		if (tcp_sc_msstab[mss] <= pmss)
1368 			break;
1369 
1370 	/* Fold parameters and MD5 digest into the ISN we will send. */
1371 	data = sch->sch_oddeven;/* odd or even secret, 1 bit */
1372 	data |= off << 1;	/* secret offset, derived from iss, 3 bits */
1373 	data |= mss << 4;	/* mss, 3 bits */
1374 
1375 	MD5Init(&ctx);
1376 	MD5Update(&ctx, ((u_int8_t *)secbits) + off,
1377 	    SYNCOOKIE_SECRET_SIZE * sizeof(*secbits) - off);
1378 	MD5Update(&ctx, secbits, off);
1379 	MD5Update(&ctx, &sc->sc_inc, sizeof(sc->sc_inc));
1380 	MD5Update(&ctx, &sc->sc_irs, sizeof(sc->sc_irs));
1381 	MD5Update(&ctx, &data, sizeof(data));
1382 	MD5Final((u_int8_t *)&md5_buffer, &ctx);
1383 
1384 	data |= (md5_buffer[0] << 7);
1385 	sc->sc_iss = data;
1386 
1387 #ifdef INET6
1388 	*flowlabel = md5_buffer[1] & IPV6_FLOWLABEL_MASK;
1389 #endif
1390 
1391 	/* Additional parameters are stored in the timestamp if present. */
1392 	if (sc->sc_flags & SCF_TIMESTAMP) {
1393 		data =  ((sc->sc_flags & SCF_SIGNATURE) ? 1 : 0); /* TCP-MD5, 1 bit */
1394 		data |= ((sc->sc_flags & SCF_SACK) ? 1 : 0) << 1; /* SACK, 1 bit */
1395 		data |= sc->sc_requested_s_scale << 2;  /* SWIN scale, 4 bits */
1396 		data |= sc->sc_requested_r_scale << 6;  /* RWIN scale, 4 bits */
1397 		data |= md5_buffer[2] << 10;		/* more digest bits */
1398 		data ^= md5_buffer[3];
1399 		sc->sc_ts = data;
1400 		sc->sc_tsoff = data - ticks;		/* after XOR */
1401 	} else
1402 		sc->sc_ts = 0;
1403 
1404 	return;
1405 }
1406 
1407 static struct syncache *
1408 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
1409     struct syncache *sc, struct tcpopt *to, struct tcphdr *th,
1410     struct socket *so)
1411 {
1412 	MD5_CTX ctx;
1413 	u_int32_t md5_buffer[MD5_DIGEST_LENGTH / sizeof(u_int32_t)];
1414 	u_int32_t data = 0;
1415 	u_int32_t *secbits;
1416 	tcp_seq ack, seq;
1417 	int off, mss, wnd, flags;
1418 
1419 	SCH_LOCK_ASSERT(sch);
1420 
1421 	/*
1422 	 * Pull information out of SYN-ACK/ACK and
1423 	 * revert sequence number advances.
1424 	 */
1425 	ack = th->th_ack - 1;
1426 	seq = th->th_seq - 1;
1427 	off = (ack >> 1) & 0x7;
1428 	mss = (ack >> 4) & 0x7;
1429 	flags = ack & 0x7f;
1430 
1431 	/* Which of the two secrets to use. */
1432 	secbits = (flags & 0x1) ? sch->sch_secbits_odd : sch->sch_secbits_even;
1433 
1434 	/*
1435 	 * The secret wasn't updated for the lifetime of a syncookie,
1436 	 * so this SYN-ACK/ACK is either too old (replay) or totally bogus.
1437 	 */
1438 	if (sch->sch_reseed < time_uptime) {
1439 		return (NULL);
1440 	}
1441 
1442 	/* Recompute the digest so we can compare it. */
1443 	MD5Init(&ctx);
1444 	MD5Update(&ctx, ((u_int8_t *)secbits) + off,
1445 	    SYNCOOKIE_SECRET_SIZE * sizeof(*secbits) - off);
1446 	MD5Update(&ctx, secbits, off);
1447 	MD5Update(&ctx, inc, sizeof(*inc));
1448 	MD5Update(&ctx, &seq, sizeof(seq));
1449 	MD5Update(&ctx, &flags, sizeof(flags));
1450 	MD5Final((u_int8_t *)&md5_buffer, &ctx);
1451 
1452 	/* Does the digest part of or ACK'ed ISS match? */
1453 	if ((ack & (~0x7f)) != (md5_buffer[0] << 7))
1454 		return (NULL);
1455 
1456 	/* Does the digest part of our reflected timestamp match? */
1457 	if (to->to_flags & TOF_TS) {
1458 		data = md5_buffer[3] ^ to->to_tsecr;
1459 		if ((data & (~0x3ff)) != (md5_buffer[2] << 10))
1460 			return (NULL);
1461 	}
1462 
1463 	/* Fill in the syncache values. */
1464 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1465 	sc->sc_ipopts = NULL;
1466 
1467 	sc->sc_irs = seq;
1468 	sc->sc_iss = ack;
1469 
1470 #ifdef INET6
1471 	if (inc->inc_isipv6) {
1472 		if (sotoinpcb(so)->in6p_flags & IN6P_AUTOFLOWLABEL)
1473 			sc->sc_flowlabel = md5_buffer[1] & IPV6_FLOWLABEL_MASK;
1474 	} else
1475 #endif
1476 	{
1477 		sc->sc_ip_ttl = sotoinpcb(so)->inp_ip_ttl;
1478 		sc->sc_ip_tos = sotoinpcb(so)->inp_ip_tos;
1479 	}
1480 
1481 	/* Additional parameters that were encoded in the timestamp. */
1482 	if (data) {
1483 		sc->sc_flags |= SCF_TIMESTAMP;
1484 		sc->sc_tsreflect = to->to_tsval;
1485 		sc->sc_tsoff = to->to_tsecr - ticks;
1486 		sc->sc_flags |= (data & 0x1) ? SCF_SIGNATURE : 0;
1487 		sc->sc_flags |= ((data >> 1) & 0x1) ? SCF_SACK : 0;
1488 		sc->sc_requested_s_scale = min((data >> 2) & 0xf,
1489 		    TCP_MAX_WINSHIFT);
1490 		sc->sc_requested_r_scale = min((data >> 6) & 0xf,
1491 		    TCP_MAX_WINSHIFT);
1492 		if (sc->sc_requested_s_scale || sc->sc_requested_r_scale)
1493 			sc->sc_flags |= SCF_WINSCALE;
1494 	} else
1495 		sc->sc_flags |= SCF_NOOPT;
1496 
1497 	wnd = sbspace(&so->so_rcv);
1498 	wnd = imax(wnd, 0);
1499 	wnd = imin(wnd, TCP_MAXWIN);
1500 	sc->sc_wnd = wnd;
1501 
1502 	sc->sc_rxmits = 0;
1503 	sc->sc_peer_mss = tcp_sc_msstab[mss];
1504 
1505 	return (sc);
1506 }
1507