xref: /freebsd/sys/netinet/tcp_syncache.c (revision f9218d3d4fd34f082473b3a021c6d4d109fb47cf)
1 /*-
2  * Copyright (c) 2001 Networks Associates Technology, Inc.
3  * All rights reserved.
4  *
5  * This software was developed for the FreeBSD Project by Jonathan Lemon
6  * and NAI Labs, the Security Research Division of Network Associates, Inc.
7  * under DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
8  * DARPA CHATS research program.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. The name of the author may not be used to endorse or promote
19  *    products derived from this software without specific prior written
20  *    permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  * $FreeBSD$
35  */
36 
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 #include "opt_mac.h"
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/kernel.h>
44 #include <sys/sysctl.h>
45 #include <sys/malloc.h>
46 #include <sys/mac.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 <net/if.h>
55 #include <net/route.h>
56 
57 #include <netinet/in.h>
58 #include <netinet/in_systm.h>
59 #include <netinet/ip.h>
60 #include <netinet/in_var.h>
61 #include <netinet/in_pcb.h>
62 #include <netinet/ip_var.h>
63 #ifdef INET6
64 #include <netinet/ip6.h>
65 #include <netinet/icmp6.h>
66 #include <netinet6/nd6.h>
67 #include <netinet6/ip6_var.h>
68 #include <netinet6/in6_pcb.h>
69 #endif
70 #include <netinet/tcp.h>
71 #include <netinet/tcp_fsm.h>
72 #include <netinet/tcp_seq.h>
73 #include <netinet/tcp_timer.h>
74 #include <netinet/tcp_var.h>
75 #ifdef INET6
76 #include <netinet6/tcp6_var.h>
77 #endif
78 
79 #ifdef IPSEC
80 #include <netinet6/ipsec.h>
81 #ifdef INET6
82 #include <netinet6/ipsec6.h>
83 #endif
84 #endif /*IPSEC*/
85 
86 #ifdef FAST_IPSEC
87 #include <netipsec/ipsec.h>
88 #ifdef INET6
89 #include <netipsec/ipsec6.h>
90 #endif
91 #include <netipsec/key.h>
92 #define	IPSEC
93 #endif /*FAST_IPSEC*/
94 
95 #include <machine/in_cksum.h>
96 #include <vm/uma.h>
97 
98 static int tcp_syncookies = 1;
99 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_RW,
100     &tcp_syncookies, 0,
101     "Use TCP SYN cookies if the syncache overflows");
102 
103 static void	 syncache_drop(struct syncache *, struct syncache_head *);
104 static void	 syncache_free(struct syncache *);
105 static void	 syncache_insert(struct syncache *, struct syncache_head *);
106 struct syncache *syncache_lookup(struct in_conninfo *, struct syncache_head **);
107 static int	 syncache_respond(struct syncache *, struct mbuf *);
108 static struct 	 socket *syncache_socket(struct syncache *, struct socket *,
109 		    struct mbuf *m);
110 static void	 syncache_timer(void *);
111 static u_int32_t syncookie_generate(struct syncache *);
112 static struct syncache *syncookie_lookup(struct in_conninfo *,
113 		    struct tcphdr *, struct socket *);
114 
115 /*
116  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
117  * 3 retransmits corresponds to a timeout of (1 + 2 + 4 + 8 == 15) seconds,
118  * the odds are that the user has given up attempting to connect by then.
119  */
120 #define SYNCACHE_MAXREXMTS		3
121 
122 /* Arbitrary values */
123 #define TCP_SYNCACHE_HASHSIZE		512
124 #define TCP_SYNCACHE_BUCKETLIMIT	30
125 
126 struct tcp_syncache {
127 	struct	syncache_head *hashbase;
128 	uma_zone_t zone;
129 	u_int	hashsize;
130 	u_int	hashmask;
131 	u_int	bucket_limit;
132 	u_int	cache_count;
133 	u_int	cache_limit;
134 	u_int	rexmt_limit;
135 	u_int	hash_secret;
136 	u_int	next_reseed;
137 	TAILQ_HEAD(, syncache) timerq[SYNCACHE_MAXREXMTS + 1];
138 	struct	callout tt_timerq[SYNCACHE_MAXREXMTS + 1];
139 };
140 static struct tcp_syncache tcp_syncache;
141 
142 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, "TCP SYN cache");
143 
144 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RD,
145      &tcp_syncache.bucket_limit, 0, "Per-bucket hash limit for syncache");
146 
147 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_RD,
148      &tcp_syncache.cache_limit, 0, "Overall entry limit for syncache");
149 
150 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_RD,
151      &tcp_syncache.cache_count, 0, "Current number of entries in syncache");
152 
153 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RD,
154      &tcp_syncache.hashsize, 0, "Size of TCP syncache hashtable");
155 
156 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_RW,
157      &tcp_syncache.rexmt_limit, 0, "Limit on SYN/ACK retransmissions");
158 
159 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
160 
161 #define SYNCACHE_HASH(inc, mask) 					\
162 	((tcp_syncache.hash_secret ^					\
163 	  (inc)->inc_faddr.s_addr ^					\
164 	  ((inc)->inc_faddr.s_addr >> 16) ^ 				\
165 	  (inc)->inc_fport ^ (inc)->inc_lport) & mask)
166 
167 #define SYNCACHE_HASH6(inc, mask) 					\
168 	((tcp_syncache.hash_secret ^					\
169 	  (inc)->inc6_faddr.s6_addr32[0] ^ 				\
170 	  (inc)->inc6_faddr.s6_addr32[3] ^ 				\
171 	  (inc)->inc_fport ^ (inc)->inc_lport) & mask)
172 
173 #define ENDPTS_EQ(a, b) (						\
174 	(a)->ie_fport == (b)->ie_fport &&				\
175 	(a)->ie_lport == (b)->ie_lport &&				\
176 	(a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr &&			\
177 	(a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr			\
178 )
179 
180 #define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0)
181 
182 #define SYNCACHE_TIMEOUT(sc, slot) do {				\
183 	sc->sc_rxtslot = (slot);					\
184 	sc->sc_rxttime = ticks + TCPTV_RTOBASE * tcp_backoff[(slot)];	\
185 	TAILQ_INSERT_TAIL(&tcp_syncache.timerq[(slot)], sc, sc_timerq);	\
186 	if (!callout_active(&tcp_syncache.tt_timerq[(slot)]))		\
187 		callout_reset(&tcp_syncache.tt_timerq[(slot)],		\
188 		    TCPTV_RTOBASE * tcp_backoff[(slot)],		\
189 		    syncache_timer, (void *)((intptr_t)(slot)));	\
190 } while (0)
191 
192 static void
193 syncache_free(struct syncache *sc)
194 {
195 	struct rtentry *rt;
196 
197 	if (sc->sc_ipopts)
198 		(void) m_free(sc->sc_ipopts);
199 #ifdef INET6
200 	if (sc->sc_inc.inc_isipv6)
201 		rt = sc->sc_route6.ro_rt;
202 	else
203 #endif
204 		rt = sc->sc_route.ro_rt;
205 	if (rt != NULL) {
206 		/*
207 		 * If this is the only reference to a protocol cloned
208 		 * route, remove it immediately.
209 		 */
210 		if (rt->rt_flags & RTF_WASCLONED &&
211 		    (sc->sc_flags & SCF_KEEPROUTE) == 0 &&
212 		    rt->rt_refcnt == 1)
213 			rtrequest(RTM_DELETE, rt_key(rt),
214 			    rt->rt_gateway, rt_mask(rt),
215 			    rt->rt_flags, NULL);
216 		RTFREE(rt);
217 	}
218 	uma_zfree(tcp_syncache.zone, sc);
219 }
220 
221 void
222 syncache_init(void)
223 {
224 	int i;
225 
226 	tcp_syncache.cache_count = 0;
227 	tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
228 	tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
229 	tcp_syncache.cache_limit =
230 	    tcp_syncache.hashsize * tcp_syncache.bucket_limit;
231 	tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
232 	tcp_syncache.next_reseed = 0;
233 	tcp_syncache.hash_secret = arc4random();
234 
235         TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
236 	    &tcp_syncache.hashsize);
237         TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
238 	    &tcp_syncache.cache_limit);
239         TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
240 	    &tcp_syncache.bucket_limit);
241 	if (!powerof2(tcp_syncache.hashsize)) {
242                 printf("WARNING: syncache hash size is not a power of 2.\n");
243 		tcp_syncache.hashsize = 512;	/* safe default */
244         }
245 	tcp_syncache.hashmask = tcp_syncache.hashsize - 1;
246 
247 	/* Allocate the hash table. */
248 	MALLOC(tcp_syncache.hashbase, struct syncache_head *,
249 	    tcp_syncache.hashsize * sizeof(struct syncache_head),
250 	    M_SYNCACHE, M_WAITOK);
251 
252 	/* Initialize the hash buckets. */
253 	for (i = 0; i < tcp_syncache.hashsize; i++) {
254 		TAILQ_INIT(&tcp_syncache.hashbase[i].sch_bucket);
255 		tcp_syncache.hashbase[i].sch_length = 0;
256 	}
257 
258 	/* Initialize the timer queues. */
259 	for (i = 0; i <= SYNCACHE_MAXREXMTS; i++) {
260 		TAILQ_INIT(&tcp_syncache.timerq[i]);
261 		callout_init(&tcp_syncache.tt_timerq[i], 0);
262 	}
263 
264 	/*
265 	 * Allocate the syncache entries.  Allow the zone to allocate one
266 	 * more entry than cache limit, so a new entry can bump out an
267 	 * older one.
268 	 */
269 	tcp_syncache.cache_limit -= 1;
270 	tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
271 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
272 	uma_zone_set_max(tcp_syncache.zone, tcp_syncache.cache_limit);
273 }
274 
275 static void
276 syncache_insert(sc, sch)
277 	struct syncache *sc;
278 	struct syncache_head *sch;
279 {
280 	struct syncache *sc2;
281 	int s, i;
282 
283 	/*
284 	 * Make sure that we don't overflow the per-bucket
285 	 * limit or the total cache size limit.
286 	 */
287 	s = splnet();
288 	if (sch->sch_length >= tcp_syncache.bucket_limit) {
289 		/*
290 		 * The bucket is full, toss the oldest element.
291 		 */
292 		sc2 = TAILQ_FIRST(&sch->sch_bucket);
293 		sc2->sc_tp->ts_recent = ticks;
294 		syncache_drop(sc2, sch);
295 		tcpstat.tcps_sc_bucketoverflow++;
296 	} else if (tcp_syncache.cache_count >= tcp_syncache.cache_limit) {
297 		/*
298 		 * The cache is full.  Toss the oldest entry in the
299 		 * entire cache.  This is the front entry in the
300 		 * first non-empty timer queue with the largest
301 		 * timeout value.
302 		 */
303 		for (i = SYNCACHE_MAXREXMTS; i >= 0; i--) {
304 			sc2 = TAILQ_FIRST(&tcp_syncache.timerq[i]);
305 			if (sc2 != NULL)
306 				break;
307 		}
308 		sc2->sc_tp->ts_recent = ticks;
309 		syncache_drop(sc2, NULL);
310 		tcpstat.tcps_sc_cacheoverflow++;
311 	}
312 
313 	/* Initialize the entry's timer. */
314 	SYNCACHE_TIMEOUT(sc, 0);
315 
316 	/* Put it into the bucket. */
317 	TAILQ_INSERT_TAIL(&sch->sch_bucket, sc, sc_hash);
318 	sch->sch_length++;
319 	tcp_syncache.cache_count++;
320 	tcpstat.tcps_sc_added++;
321 	splx(s);
322 }
323 
324 static void
325 syncache_drop(sc, sch)
326 	struct syncache *sc;
327 	struct syncache_head *sch;
328 {
329 	int s;
330 
331 	if (sch == NULL) {
332 #ifdef INET6
333 		if (sc->sc_inc.inc_isipv6) {
334 			sch = &tcp_syncache.hashbase[
335 			    SYNCACHE_HASH6(&sc->sc_inc, tcp_syncache.hashmask)];
336 		} else
337 #endif
338 		{
339 			sch = &tcp_syncache.hashbase[
340 			    SYNCACHE_HASH(&sc->sc_inc, tcp_syncache.hashmask)];
341 		}
342 	}
343 
344 	s = splnet();
345 
346 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
347 	sch->sch_length--;
348 	tcp_syncache.cache_count--;
349 
350 	TAILQ_REMOVE(&tcp_syncache.timerq[sc->sc_rxtslot], sc, sc_timerq);
351 	if (TAILQ_EMPTY(&tcp_syncache.timerq[sc->sc_rxtslot]))
352 		callout_stop(&tcp_syncache.tt_timerq[sc->sc_rxtslot]);
353 	splx(s);
354 
355 	syncache_free(sc);
356 }
357 
358 /*
359  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
360  * If we have retransmitted an entry the maximum number of times, expire it.
361  */
362 static void
363 syncache_timer(xslot)
364 	void *xslot;
365 {
366 	intptr_t slot = (intptr_t)xslot;
367 	struct syncache *sc, *nsc;
368 	struct inpcb *inp;
369 	int s;
370 
371 	s = splnet();
372 	INP_INFO_WLOCK(&tcbinfo);
373         if (callout_pending(&tcp_syncache.tt_timerq[slot]) ||
374             !callout_active(&tcp_syncache.tt_timerq[slot])) {
375 		INP_INFO_WUNLOCK(&tcbinfo);
376                 splx(s);
377                 return;
378         }
379         callout_deactivate(&tcp_syncache.tt_timerq[slot]);
380 
381         nsc = TAILQ_FIRST(&tcp_syncache.timerq[slot]);
382 	while (nsc != NULL) {
383 		if (ticks < nsc->sc_rxttime)
384 			break;
385 		sc = nsc;
386 		inp = sc->sc_tp->t_inpcb;
387 		if (slot == SYNCACHE_MAXREXMTS ||
388 		    slot >= tcp_syncache.rexmt_limit ||
389 		    inp == NULL || inp->inp_gencnt != sc->sc_inp_gencnt) {
390 			nsc = TAILQ_NEXT(sc, sc_timerq);
391 			syncache_drop(sc, NULL);
392 			tcpstat.tcps_sc_stale++;
393 			continue;
394 		}
395 		/*
396 		 * syncache_respond() may call back into the syncache to
397 		 * to modify another entry, so do not obtain the next
398 		 * entry on the timer chain until it has completed.
399 		 */
400 		(void) syncache_respond(sc, NULL);
401 		nsc = TAILQ_NEXT(sc, sc_timerq);
402 		tcpstat.tcps_sc_retransmitted++;
403 		TAILQ_REMOVE(&tcp_syncache.timerq[slot], sc, sc_timerq);
404 		SYNCACHE_TIMEOUT(sc, slot + 1);
405 	}
406 	if (nsc != NULL)
407 		callout_reset(&tcp_syncache.tt_timerq[slot],
408 		    nsc->sc_rxttime - ticks, syncache_timer, (void *)(slot));
409 	INP_INFO_WUNLOCK(&tcbinfo);
410 	splx(s);
411 }
412 
413 /*
414  * Find an entry in the syncache.
415  */
416 struct syncache *
417 syncache_lookup(inc, schp)
418 	struct in_conninfo *inc;
419 	struct syncache_head **schp;
420 {
421 	struct syncache *sc;
422 	struct syncache_head *sch;
423 	int s;
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 		s = splnet();
431 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
432 			if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) {
433 				splx(s);
434 				return (sc);
435 			}
436 		}
437 		splx(s);
438 	} else
439 #endif
440 	{
441 		sch = &tcp_syncache.hashbase[
442 		    SYNCACHE_HASH(inc, tcp_syncache.hashmask)];
443 		*schp = sch;
444 		s = splnet();
445 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
446 #ifdef INET6
447 			if (sc->sc_inc.inc_isipv6)
448 				continue;
449 #endif
450 			if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie)) {
451 				splx(s);
452 				return (sc);
453 			}
454 		}
455 		splx(s);
456 	}
457 	return (NULL);
458 }
459 
460 /*
461  * This function is called when we get a RST for a
462  * non-existent connection, so that we can see if the
463  * connection is in the syn cache.  If it is, zap it.
464  */
465 void
466 syncache_chkrst(inc, th)
467 	struct in_conninfo *inc;
468 	struct tcphdr *th;
469 {
470 	struct syncache *sc;
471 	struct syncache_head *sch;
472 
473 	sc = syncache_lookup(inc, &sch);
474 	if (sc == NULL)
475 		return;
476 	/*
477 	 * If the RST bit is set, check the sequence number to see
478 	 * if this is a valid reset segment.
479 	 * RFC 793 page 37:
480 	 *   In all states except SYN-SENT, all reset (RST) segments
481 	 *   are validated by checking their SEQ-fields.  A reset is
482 	 *   valid if its sequence number is in the window.
483 	 *
484 	 *   The sequence number in the reset segment is normally an
485 	 *   echo of our outgoing acknowlegement numbers, but some hosts
486 	 *   send a reset with the sequence number at the rightmost edge
487 	 *   of our receive window, and we have to handle this case.
488 	 */
489 	if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
490 	    SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
491 		syncache_drop(sc, sch);
492 		tcpstat.tcps_sc_reset++;
493 	}
494 }
495 
496 void
497 syncache_badack(inc)
498 	struct in_conninfo *inc;
499 {
500 	struct syncache *sc;
501 	struct syncache_head *sch;
502 
503 	sc = syncache_lookup(inc, &sch);
504 	if (sc != NULL) {
505 		syncache_drop(sc, sch);
506 		tcpstat.tcps_sc_badack++;
507 	}
508 }
509 
510 void
511 syncache_unreach(inc, th)
512 	struct in_conninfo *inc;
513 	struct tcphdr *th;
514 {
515 	struct syncache *sc;
516 	struct syncache_head *sch;
517 
518 	/* we are called at splnet() here */
519 	sc = syncache_lookup(inc, &sch);
520 	if (sc == NULL)
521 		return;
522 
523 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
524 	if (ntohl(th->th_seq) != sc->sc_iss)
525 		return;
526 
527 	/*
528 	 * If we've rertransmitted 3 times and this is our second error,
529 	 * we remove the entry.  Otherwise, we allow it to continue on.
530 	 * This prevents us from incorrectly nuking an entry during a
531 	 * spurious network outage.
532 	 *
533 	 * See tcp_notify().
534 	 */
535 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxtslot < 3) {
536 		sc->sc_flags |= SCF_UNREACH;
537 		return;
538 	}
539 	syncache_drop(sc, sch);
540 	tcpstat.tcps_sc_unreach++;
541 }
542 
543 /*
544  * Build a new TCP socket structure from a syncache entry.
545  */
546 static struct socket *
547 syncache_socket(sc, lso, m)
548 	struct syncache *sc;
549 	struct socket *lso;
550 	struct mbuf *m;
551 {
552 	struct inpcb *inp = NULL;
553 	struct socket *so;
554 	struct tcpcb *tp;
555 
556 	/*
557 	 * Ok, create the full blown connection, and set things up
558 	 * as they would have been set up if we had created the
559 	 * connection when the SYN arrived.  If we can't create
560 	 * the connection, abort it.
561 	 */
562 	so = sonewconn(lso, SS_ISCONNECTED);
563 	if (so == NULL) {
564 		/*
565 		 * Drop the connection; we will send a RST if the peer
566 		 * retransmits the ACK,
567 		 */
568 		tcpstat.tcps_listendrop++;
569 		goto abort;
570 	}
571 #ifdef MAC
572 	mac_set_socket_peer_from_mbuf(m, so);
573 #endif
574 
575 	inp = sotoinpcb(so);
576 
577 	/*
578 	 * Insert new socket into hash list.
579 	 */
580 	inp->inp_inc.inc_isipv6 = sc->sc_inc.inc_isipv6;
581 #ifdef INET6
582 	if (sc->sc_inc.inc_isipv6) {
583 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
584 	} else {
585 		inp->inp_vflag &= ~INP_IPV6;
586 		inp->inp_vflag |= INP_IPV4;
587 #endif
588 		inp->inp_laddr = sc->sc_inc.inc_laddr;
589 #ifdef INET6
590 	}
591 #endif
592 	inp->inp_lport = sc->sc_inc.inc_lport;
593 	if (in_pcbinshash(inp) != 0) {
594 		/*
595 		 * Undo the assignments above if we failed to
596 		 * put the PCB on the hash lists.
597 		 */
598 #ifdef INET6
599 		if (sc->sc_inc.inc_isipv6)
600 			inp->in6p_laddr = in6addr_any;
601        		else
602 #endif
603 			inp->inp_laddr.s_addr = INADDR_ANY;
604 		inp->inp_lport = 0;
605 		goto abort;
606 	}
607 #ifdef IPSEC
608 	/* copy old policy into new socket's */
609 	if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp))
610 		printf("syncache_expand: could not copy policy\n");
611 #endif
612 #ifdef INET6
613 	if (sc->sc_inc.inc_isipv6) {
614 		struct inpcb *oinp = sotoinpcb(lso);
615 		struct in6_addr laddr6;
616 		struct sockaddr_in6 *sin6;
617 		/*
618 		 * Inherit socket options from the listening socket.
619 		 * Note that in6p_inputopts are not (and should not be)
620 		 * copied, since it stores previously received options and is
621 		 * used to detect if each new option is different than the
622 		 * previous one and hence should be passed to a user.
623                  * If we copied in6p_inputopts, a user would not be able to
624 		 * receive options just after calling the accept system call.
625 		 */
626 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
627 		if (oinp->in6p_outputopts)
628 			inp->in6p_outputopts =
629 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
630 		inp->in6p_route = sc->sc_route6;
631 		sc->sc_route6.ro_rt = NULL;
632 
633 		MALLOC(sin6, struct sockaddr_in6 *, sizeof *sin6,
634 		    M_SONAME, M_NOWAIT | M_ZERO);
635 		if (sin6 == NULL)
636 			goto abort;
637 		sin6->sin6_family = AF_INET6;
638 		sin6->sin6_len = sizeof(*sin6);
639 		sin6->sin6_addr = sc->sc_inc.inc6_faddr;
640 		sin6->sin6_port = sc->sc_inc.inc_fport;
641 		laddr6 = inp->in6p_laddr;
642 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
643 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
644 		if (in6_pcbconnect(inp, (struct sockaddr *)sin6, &thread0)) {
645 			inp->in6p_laddr = laddr6;
646 			FREE(sin6, M_SONAME);
647 			goto abort;
648 		}
649 		FREE(sin6, M_SONAME);
650 	} else
651 #endif
652 	{
653 		struct in_addr laddr;
654 		struct sockaddr_in *sin;
655 
656 		inp->inp_options = ip_srcroute();
657 		if (inp->inp_options == NULL) {
658 			inp->inp_options = sc->sc_ipopts;
659 			sc->sc_ipopts = NULL;
660 		}
661 		inp->inp_route = sc->sc_route;
662 		sc->sc_route.ro_rt = NULL;
663 
664 		MALLOC(sin, struct sockaddr_in *, sizeof *sin,
665 		    M_SONAME, M_NOWAIT | M_ZERO);
666 		if (sin == NULL)
667 			goto abort;
668 		sin->sin_family = AF_INET;
669 		sin->sin_len = sizeof(*sin);
670 		sin->sin_addr = sc->sc_inc.inc_faddr;
671 		sin->sin_port = sc->sc_inc.inc_fport;
672 		bzero((caddr_t)sin->sin_zero, sizeof(sin->sin_zero));
673 		laddr = inp->inp_laddr;
674 		if (inp->inp_laddr.s_addr == INADDR_ANY)
675 			inp->inp_laddr = sc->sc_inc.inc_laddr;
676 		if (in_pcbconnect(inp, (struct sockaddr *)sin, &thread0)) {
677 			inp->inp_laddr = laddr;
678 			FREE(sin, M_SONAME);
679 			goto abort;
680 		}
681 		FREE(sin, M_SONAME);
682 	}
683 
684 	tp = intotcpcb(inp);
685 	tp->t_state = TCPS_SYN_RECEIVED;
686 	tp->iss = sc->sc_iss;
687 	tp->irs = sc->sc_irs;
688 	tcp_rcvseqinit(tp);
689 	tcp_sendseqinit(tp);
690 	tp->snd_wl1 = sc->sc_irs;
691 	tp->rcv_up = sc->sc_irs + 1;
692 	tp->rcv_wnd = sc->sc_wnd;
693 	tp->rcv_adv += tp->rcv_wnd;
694 
695 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
696 	if (sc->sc_flags & SCF_NOOPT)
697 		tp->t_flags |= TF_NOOPT;
698 	if (sc->sc_flags & SCF_WINSCALE) {
699 		tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
700 		tp->requested_s_scale = sc->sc_requested_s_scale;
701 		tp->request_r_scale = sc->sc_request_r_scale;
702 	}
703 	if (sc->sc_flags & SCF_TIMESTAMP) {
704 		tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
705 		tp->ts_recent = sc->sc_tsrecent;
706 		tp->ts_recent_age = ticks;
707 	}
708 	if (sc->sc_flags & SCF_CC) {
709 		/*
710 		 * Initialization of the tcpcb for transaction;
711 		 *   set SND.WND = SEG.WND,
712 		 *   initialize CCsend and CCrecv.
713 		 */
714 		tp->t_flags |= TF_REQ_CC|TF_RCVD_CC;
715 		tp->cc_send = sc->sc_cc_send;
716 		tp->cc_recv = sc->sc_cc_recv;
717 	}
718 
719 	tcp_mss(tp, sc->sc_peer_mss);
720 
721 	/*
722 	 * If the SYN,ACK was retransmitted, reset cwnd to 1 segment.
723 	 */
724 	if (sc->sc_rxtslot != 0)
725                 tp->snd_cwnd = tp->t_maxseg;
726 	callout_reset(tp->tt_keep, tcp_keepinit, tcp_timer_keep, tp);
727 
728 	tcpstat.tcps_accepts++;
729 	return (so);
730 
731 abort:
732 	if (so != NULL)
733 		(void) soabort(so);
734 	return (NULL);
735 }
736 
737 /*
738  * This function gets called when we receive an ACK for a
739  * socket in the LISTEN state.  We look up the connection
740  * in the syncache, and if its there, we pull it out of
741  * the cache and turn it into a full-blown connection in
742  * the SYN-RECEIVED state.
743  */
744 int
745 syncache_expand(inc, th, sop, m)
746 	struct in_conninfo *inc;
747 	struct tcphdr *th;
748 	struct socket **sop;
749 	struct mbuf *m;
750 {
751 	struct syncache *sc;
752 	struct syncache_head *sch;
753 	struct socket *so;
754 
755 	sc = syncache_lookup(inc, &sch);
756 	if (sc == NULL) {
757 		/*
758 		 * There is no syncache entry, so see if this ACK is
759 		 * a returning syncookie.  To do this, first:
760 		 *  A. See if this socket has had a syncache entry dropped in
761 		 *     the past.  We don't want to accept a bogus syncookie
762  		 *     if we've never received a SYN.
763 		 *  B. check that the syncookie is valid.  If it is, then
764 		 *     cobble up a fake syncache entry, and return.
765 		 */
766 		if (!tcp_syncookies)
767 			return (0);
768 		sc = syncookie_lookup(inc, th, *sop);
769 		if (sc == NULL)
770 			return (0);
771 		sch = NULL;
772 		tcpstat.tcps_sc_recvcookie++;
773 	}
774 
775 	/*
776 	 * If seg contains an ACK, but not for our SYN/ACK, send a RST.
777 	 */
778 	if (th->th_ack != sc->sc_iss + 1)
779 		return (0);
780 
781 	so = syncache_socket(sc, *sop, m);
782 	if (so == NULL) {
783 #if 0
784 resetandabort:
785 		/* XXXjlemon check this - is this correct? */
786 		(void) tcp_respond(NULL, m, m, th,
787 		    th->th_seq + tlen, (tcp_seq)0, TH_RST|TH_ACK);
788 #endif
789 		m_freem(m);			/* XXX only needed for above */
790 		tcpstat.tcps_sc_aborted++;
791 	} else {
792 		sc->sc_flags |= SCF_KEEPROUTE;
793 		tcpstat.tcps_sc_completed++;
794 	}
795 	if (sch == NULL)
796 		syncache_free(sc);
797 	else
798 		syncache_drop(sc, sch);
799 	*sop = so;
800 	return (1);
801 }
802 
803 /*
804  * Given a LISTEN socket and an inbound SYN request, add
805  * this to the syn cache, and send back a segment:
806  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
807  * to the source.
808  *
809  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
810  * Doing so would require that we hold onto the data and deliver it
811  * to the application.  However, if we are the target of a SYN-flood
812  * DoS attack, an attacker could send data which would eventually
813  * consume all available buffer space if it were ACKed.  By not ACKing
814  * the data, we avoid this DoS scenario.
815  */
816 int
817 syncache_add(inc, to, th, sop, m)
818 	struct in_conninfo *inc;
819 	struct tcpopt *to;
820 	struct tcphdr *th;
821 	struct socket **sop;
822 	struct mbuf *m;
823 {
824 	struct tcpcb *tp;
825 	struct socket *so;
826 	struct syncache *sc = NULL;
827 	struct syncache_head *sch;
828 	struct mbuf *ipopts = NULL;
829 	struct rmxp_tao *taop;
830 	int i, s, win;
831 
832 	so = *sop;
833 	tp = sototcpcb(so);
834 
835 	/*
836 	 * Remember the IP options, if any.
837 	 */
838 #ifdef INET6
839 	if (!inc->inc_isipv6)
840 #endif
841 		ipopts = ip_srcroute();
842 
843 	/*
844 	 * See if we already have an entry for this connection.
845 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
846 	 *
847 	 * XXX
848 	 * should the syncache be re-initialized with the contents
849 	 * of the new SYN here (which may have different options?)
850 	 */
851 	sc = syncache_lookup(inc, &sch);
852 	if (sc != NULL) {
853 		tcpstat.tcps_sc_dupsyn++;
854 		if (ipopts) {
855 			/*
856 			 * If we were remembering a previous source route,
857 			 * forget it and use the new one we've been given.
858 			 */
859 			if (sc->sc_ipopts)
860 				(void) m_free(sc->sc_ipopts);
861 			sc->sc_ipopts = ipopts;
862 		}
863 		/*
864 		 * Update timestamp if present.
865 		 */
866 		if (sc->sc_flags & SCF_TIMESTAMP)
867 			sc->sc_tsrecent = to->to_tsval;
868 		/*
869 		 * PCB may have changed, pick up new values.
870 		 */
871 		sc->sc_tp = tp;
872 		sc->sc_inp_gencnt = tp->t_inpcb->inp_gencnt;
873 		if (syncache_respond(sc, m) == 0) {
874 		        s = splnet();
875 			TAILQ_REMOVE(&tcp_syncache.timerq[sc->sc_rxtslot],
876 			    sc, sc_timerq);
877 			SYNCACHE_TIMEOUT(sc, sc->sc_rxtslot);
878 		        splx(s);
879 		 	tcpstat.tcps_sndacks++;
880 			tcpstat.tcps_sndtotal++;
881 		}
882 		*sop = NULL;
883 		return (1);
884 	}
885 
886 	sc = uma_zalloc(tcp_syncache.zone, M_NOWAIT);
887 	if (sc == NULL) {
888 		/*
889 		 * The zone allocator couldn't provide more entries.
890 		 * Treat this as if the cache was full; drop the oldest
891 		 * entry and insert the new one.
892 		 */
893 		s = splnet();
894 		for (i = SYNCACHE_MAXREXMTS; i >= 0; i--) {
895 			sc = TAILQ_FIRST(&tcp_syncache.timerq[i]);
896 			if (sc != NULL)
897 				break;
898 		}
899 		sc->sc_tp->ts_recent = ticks;
900 		syncache_drop(sc, NULL);
901 		splx(s);
902 		tcpstat.tcps_sc_zonefail++;
903 		sc = uma_zalloc(tcp_syncache.zone, M_NOWAIT);
904 		if (sc == NULL) {
905 			if (ipopts)
906 				(void) m_free(ipopts);
907 			return (0);
908 		}
909 	}
910 
911 	/*
912 	 * Fill in the syncache values.
913 	 */
914 	bzero(sc, sizeof(*sc));
915 	sc->sc_tp = tp;
916 	sc->sc_inp_gencnt = tp->t_inpcb->inp_gencnt;
917 	sc->sc_ipopts = ipopts;
918 	sc->sc_inc.inc_fport = inc->inc_fport;
919 	sc->sc_inc.inc_lport = inc->inc_lport;
920 #ifdef INET6
921 	sc->sc_inc.inc_isipv6 = inc->inc_isipv6;
922 	if (inc->inc_isipv6) {
923 		sc->sc_inc.inc6_faddr = inc->inc6_faddr;
924 		sc->sc_inc.inc6_laddr = inc->inc6_laddr;
925 		sc->sc_route6.ro_rt = NULL;
926 	} else
927 #endif
928 	{
929 		sc->sc_inc.inc_faddr = inc->inc_faddr;
930 		sc->sc_inc.inc_laddr = inc->inc_laddr;
931 		sc->sc_route.ro_rt = NULL;
932 	}
933 	sc->sc_irs = th->th_seq;
934 	sc->sc_flags = 0;
935 	sc->sc_peer_mss = to->to_flags & TOF_MSS ? to->to_mss : 0;
936 	if (tcp_syncookies)
937 		sc->sc_iss = syncookie_generate(sc);
938 	else
939 		sc->sc_iss = arc4random();
940 
941 	/* Initial receive window: clip sbspace to [0 .. TCP_MAXWIN] */
942 	win = sbspace(&so->so_rcv);
943 	win = imax(win, 0);
944 	win = imin(win, TCP_MAXWIN);
945 	sc->sc_wnd = win;
946 
947 	if (tcp_do_rfc1323) {
948 		/*
949 		 * A timestamp received in a SYN makes
950 		 * it ok to send timestamp requests and replies.
951 		 */
952 		if (to->to_flags & TOF_TS) {
953 			sc->sc_tsrecent = to->to_tsval;
954 			sc->sc_flags |= SCF_TIMESTAMP;
955 		}
956 		if (to->to_flags & TOF_SCALE) {
957 			int wscale = 0;
958 
959 			/* Compute proper scaling value from buffer space */
960 			while (wscale < TCP_MAX_WINSHIFT &&
961 			    (TCP_MAXWIN << wscale) < so->so_rcv.sb_hiwat)
962 				wscale++;
963 			sc->sc_request_r_scale = wscale;
964 			sc->sc_requested_s_scale = to->to_requested_s_scale;
965 			sc->sc_flags |= SCF_WINSCALE;
966 		}
967 	}
968 	if (tcp_do_rfc1644) {
969 		/*
970 		 * A CC or CC.new option received in a SYN makes
971 		 * it ok to send CC in subsequent segments.
972 		 */
973 		if (to->to_flags & (TOF_CC|TOF_CCNEW)) {
974 			sc->sc_cc_recv = to->to_cc;
975 			sc->sc_cc_send = CC_INC(tcp_ccgen);
976 			sc->sc_flags |= SCF_CC;
977 		}
978 	}
979 	if (tp->t_flags & TF_NOOPT)
980 		sc->sc_flags = SCF_NOOPT;
981 
982 	/*
983 	 * XXX
984 	 * We have the option here of not doing TAO (even if the segment
985 	 * qualifies) and instead fall back to a normal 3WHS via the syncache.
986 	 * This allows us to apply synflood protection to TAO-qualifying SYNs
987 	 * also. However, there should be a hueristic to determine when to
988 	 * do this, and is not present at the moment.
989 	 */
990 
991 	/*
992 	 * Perform TAO test on incoming CC (SEG.CC) option, if any.
993 	 * - compare SEG.CC against cached CC from the same host, if any.
994 	 * - if SEG.CC > chached value, SYN must be new and is accepted
995 	 *	immediately: save new CC in the cache, mark the socket
996 	 *	connected, enter ESTABLISHED state, turn on flag to
997 	 *	send a SYN in the next segment.
998 	 *	A virtual advertised window is set in rcv_adv to
999 	 *	initialize SWS prevention.  Then enter normal segment
1000 	 *	processing: drop SYN, process data and FIN.
1001 	 * - otherwise do a normal 3-way handshake.
1002 	 */
1003 	taop = tcp_gettaocache(&sc->sc_inc);
1004 	if ((to->to_flags & TOF_CC) != 0) {
1005 		if (((tp->t_flags & TF_NOPUSH) != 0) &&
1006 		    sc->sc_flags & SCF_CC &&
1007 		    taop != NULL && taop->tao_cc != 0 &&
1008 		    CC_GT(to->to_cc, taop->tao_cc)) {
1009 			sc->sc_rxtslot = 0;
1010 			so = syncache_socket(sc, *sop, m);
1011 			if (so != NULL) {
1012 				sc->sc_flags |= SCF_KEEPROUTE;
1013 				taop->tao_cc = to->to_cc;
1014 				*sop = so;
1015 			}
1016 			syncache_free(sc);
1017 			return (so != NULL);
1018 		}
1019 	} else {
1020 		/*
1021 		 * No CC option, but maybe CC.NEW: invalidate cached value.
1022 		 */
1023 		if (taop != NULL)
1024 			taop->tao_cc = 0;
1025 	}
1026 	/*
1027 	 * TAO test failed or there was no CC option,
1028 	 *    do a standard 3-way handshake.
1029 	 */
1030 	if (syncache_respond(sc, m) == 0) {
1031 		syncache_insert(sc, sch);
1032 		tcpstat.tcps_sndacks++;
1033 		tcpstat.tcps_sndtotal++;
1034 	} else {
1035 		syncache_free(sc);
1036 		tcpstat.tcps_sc_dropped++;
1037 	}
1038 	*sop = NULL;
1039 	return (1);
1040 }
1041 
1042 static int
1043 syncache_respond(sc, m)
1044 	struct syncache *sc;
1045 	struct mbuf *m;
1046 {
1047 	u_int8_t *optp;
1048 	int optlen, error;
1049 	u_int16_t tlen, hlen, mssopt;
1050 	struct ip *ip = NULL;
1051 	struct rtentry *rt;
1052 	struct tcphdr *th;
1053 #ifdef INET6
1054 	struct ip6_hdr *ip6 = NULL;
1055 #endif
1056 
1057 #ifdef INET6
1058 	if (sc->sc_inc.inc_isipv6) {
1059 		rt = tcp_rtlookup6(&sc->sc_inc);
1060 		if (rt != NULL)
1061 			mssopt = rt->rt_ifp->if_mtu -
1062 			     (sizeof(struct ip6_hdr) + sizeof(struct tcphdr));
1063 		else
1064 			mssopt = tcp_v6mssdflt;
1065 		hlen = sizeof(struct ip6_hdr);
1066 	} else
1067 #endif
1068 	{
1069 		rt = tcp_rtlookup(&sc->sc_inc);
1070 		if (rt != NULL)
1071 			mssopt = rt->rt_ifp->if_mtu -
1072 			     (sizeof(struct ip) + sizeof(struct tcphdr));
1073 		else
1074 			mssopt = tcp_mssdflt;
1075 		hlen = sizeof(struct ip);
1076 	}
1077 
1078 	/* Compute the size of the TCP options. */
1079 	if (sc->sc_flags & SCF_NOOPT) {
1080 		optlen = 0;
1081 	} else {
1082 		optlen = TCPOLEN_MAXSEG +
1083 		    ((sc->sc_flags & SCF_WINSCALE) ? 4 : 0) +
1084 		    ((sc->sc_flags & SCF_TIMESTAMP) ? TCPOLEN_TSTAMP_APPA : 0) +
1085 		    ((sc->sc_flags & SCF_CC) ? TCPOLEN_CC_APPA * 2 : 0);
1086 	}
1087 	tlen = hlen + sizeof(struct tcphdr) + optlen;
1088 
1089 	/*
1090 	 * XXX
1091 	 * assume that the entire packet will fit in a header mbuf
1092 	 */
1093 	KASSERT(max_linkhdr + tlen <= MHLEN, ("syncache: mbuf too small"));
1094 
1095 	/*
1096 	 * XXX shouldn't this reuse the mbuf if possible ?
1097 	 * Create the IP+TCP header from scratch.
1098 	 */
1099 	if (m)
1100 		m_freem(m);
1101 
1102 	m = m_gethdr(M_DONTWAIT, MT_HEADER);
1103 	if (m == NULL)
1104 		return (ENOBUFS);
1105 	m->m_data += max_linkhdr;
1106 	m->m_len = tlen;
1107 	m->m_pkthdr.len = tlen;
1108 	m->m_pkthdr.rcvif = NULL;
1109 #ifdef MAC
1110 	mac_create_mbuf_from_socket(sc->sc_tp->t_inpcb->inp_socket, m);
1111 #endif
1112 
1113 #ifdef INET6
1114 	if (sc->sc_inc.inc_isipv6) {
1115 		ip6 = mtod(m, struct ip6_hdr *);
1116 		ip6->ip6_vfc = IPV6_VERSION;
1117 		ip6->ip6_nxt = IPPROTO_TCP;
1118 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1119 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1120 		ip6->ip6_plen = htons(tlen - hlen);
1121 		/* ip6_hlim is set after checksum */
1122 		/* ip6_flow = ??? */
1123 
1124 		th = (struct tcphdr *)(ip6 + 1);
1125 	} else
1126 #endif
1127 	{
1128 		ip = mtod(m, struct ip *);
1129 		ip->ip_v = IPVERSION;
1130 		ip->ip_hl = sizeof(struct ip) >> 2;
1131 		ip->ip_len = tlen;
1132 		ip->ip_id = 0;
1133 		ip->ip_off = 0;
1134 		ip->ip_sum = 0;
1135 		ip->ip_p = IPPROTO_TCP;
1136 		ip->ip_src = sc->sc_inc.inc_laddr;
1137 		ip->ip_dst = sc->sc_inc.inc_faddr;
1138 		ip->ip_ttl = sc->sc_tp->t_inpcb->inp_ip_ttl;   /* XXX */
1139 		ip->ip_tos = sc->sc_tp->t_inpcb->inp_ip_tos;   /* XXX */
1140 
1141 		/*
1142 		 * See if we should do MTU discovery.  Route lookups are
1143 		 * expensive, so we will only unset the DF bit if:
1144 		 *
1145 		 *	1) path_mtu_discovery is disabled
1146 		 *	2) the SCF_UNREACH flag has been set
1147 		 */
1148 		if (path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1149 		       ip->ip_off |= IP_DF;
1150 
1151 		th = (struct tcphdr *)(ip + 1);
1152 	}
1153 	th->th_sport = sc->sc_inc.inc_lport;
1154 	th->th_dport = sc->sc_inc.inc_fport;
1155 
1156 	th->th_seq = htonl(sc->sc_iss);
1157 	th->th_ack = htonl(sc->sc_irs + 1);
1158 	th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1159 	th->th_x2 = 0;
1160 	th->th_flags = TH_SYN|TH_ACK;
1161 	th->th_win = htons(sc->sc_wnd);
1162 	th->th_urp = 0;
1163 
1164 	/* Tack on the TCP options. */
1165 	if (optlen != 0) {
1166 		optp = (u_int8_t *)(th + 1);
1167 		*optp++ = TCPOPT_MAXSEG;
1168 		*optp++ = TCPOLEN_MAXSEG;
1169 		*optp++ = (mssopt >> 8) & 0xff;
1170 		*optp++ = mssopt & 0xff;
1171 
1172 		if (sc->sc_flags & SCF_WINSCALE) {
1173 			*((u_int32_t *)optp) = htonl(TCPOPT_NOP << 24 |
1174 			    TCPOPT_WINDOW << 16 | TCPOLEN_WINDOW << 8 |
1175 			    sc->sc_request_r_scale);
1176 			optp += 4;
1177 		}
1178 
1179 		if (sc->sc_flags & SCF_TIMESTAMP) {
1180 			u_int32_t *lp = (u_int32_t *)(optp);
1181 
1182 			/* Form timestamp option per appendix A of RFC 1323. */
1183 			*lp++ = htonl(TCPOPT_TSTAMP_HDR);
1184 			*lp++ = htonl(ticks);
1185 			*lp   = htonl(sc->sc_tsrecent);
1186 			optp += TCPOLEN_TSTAMP_APPA;
1187 		}
1188 
1189 		/*
1190 		 * Send CC and CC.echo if we received CC from our peer.
1191 		 */
1192 		if (sc->sc_flags & SCF_CC) {
1193 			u_int32_t *lp = (u_int32_t *)(optp);
1194 
1195 			*lp++ = htonl(TCPOPT_CC_HDR(TCPOPT_CC));
1196 			*lp++ = htonl(sc->sc_cc_send);
1197 			*lp++ = htonl(TCPOPT_CC_HDR(TCPOPT_CCECHO));
1198 			*lp   = htonl(sc->sc_cc_recv);
1199 			optp += TCPOLEN_CC_APPA * 2;
1200 		}
1201 	}
1202 
1203 #ifdef INET6
1204 	if (sc->sc_inc.inc_isipv6) {
1205 		struct route_in6 *ro6 = &sc->sc_route6;
1206 
1207 		th->th_sum = 0;
1208 		th->th_sum = in6_cksum(m, IPPROTO_TCP, hlen, tlen - hlen);
1209 		ip6->ip6_hlim = in6_selecthlim(NULL,
1210 		    ro6->ro_rt ? ro6->ro_rt->rt_ifp : NULL);
1211 		error = ip6_output(m, NULL, ro6, 0, NULL, NULL,
1212 				sc->sc_tp->t_inpcb);
1213 	} else
1214 #endif
1215 	{
1216         	th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1217 		    htons(tlen - hlen + IPPROTO_TCP));
1218 		m->m_pkthdr.csum_flags = CSUM_TCP;
1219 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1220 		error = ip_output(m, sc->sc_ipopts, &sc->sc_route, 0, NULL,
1221 				sc->sc_tp->t_inpcb);
1222 	}
1223 	return (error);
1224 }
1225 
1226 /*
1227  * cookie layers:
1228  *
1229  *	|. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .|
1230  *	| peer iss                                                      |
1231  *	| MD5(laddr,faddr,secret,lport,fport)             |. . . . . . .|
1232  *	|                     0                       |(A)|             |
1233  * (A): peer mss index
1234  */
1235 
1236 /*
1237  * The values below are chosen to minimize the size of the tcp_secret
1238  * table, as well as providing roughly a 16 second lifetime for the cookie.
1239  */
1240 
1241 #define SYNCOOKIE_WNDBITS	5	/* exposed bits for window indexing */
1242 #define SYNCOOKIE_TIMESHIFT	1	/* scale ticks to window time units */
1243 
1244 #define SYNCOOKIE_WNDMASK	((1 << SYNCOOKIE_WNDBITS) - 1)
1245 #define SYNCOOKIE_NSECRETS	(1 << SYNCOOKIE_WNDBITS)
1246 #define SYNCOOKIE_TIMEOUT \
1247     (hz * (1 << SYNCOOKIE_WNDBITS) / (1 << SYNCOOKIE_TIMESHIFT))
1248 #define SYNCOOKIE_DATAMASK 	((3 << SYNCOOKIE_WNDBITS) | SYNCOOKIE_WNDMASK)
1249 
1250 static struct {
1251 	u_int32_t	ts_secbits[4];
1252 	u_int		ts_expire;
1253 } tcp_secret[SYNCOOKIE_NSECRETS];
1254 
1255 static int tcp_msstab[] = { 0, 536, 1460, 8960 };
1256 
1257 static MD5_CTX syn_ctx;
1258 
1259 #define MD5Add(v)	MD5Update(&syn_ctx, (u_char *)&v, sizeof(v))
1260 
1261 struct md5_add {
1262 	u_int32_t laddr, faddr;
1263 	u_int32_t secbits[4];
1264 	u_int16_t lport, fport;
1265 };
1266 
1267 #ifdef CTASSERT
1268 CTASSERT(sizeof(struct md5_add) == 28);
1269 #endif
1270 
1271 /*
1272  * Consider the problem of a recreated (and retransmitted) cookie.  If the
1273  * original SYN was accepted, the connection is established.  The second
1274  * SYN is inflight, and if it arrives with an ISN that falls within the
1275  * receive window, the connection is killed.
1276  *
1277  * However, since cookies have other problems, this may not be worth
1278  * worrying about.
1279  */
1280 
1281 static u_int32_t
1282 syncookie_generate(struct syncache *sc)
1283 {
1284 	u_int32_t md5_buffer[4];
1285 	u_int32_t data;
1286 	int idx, i;
1287 	struct md5_add add;
1288 
1289 	idx = ((ticks << SYNCOOKIE_TIMESHIFT) / hz) & SYNCOOKIE_WNDMASK;
1290 	if (tcp_secret[idx].ts_expire < ticks) {
1291 		for (i = 0; i < 4; i++)
1292 			tcp_secret[idx].ts_secbits[i] = arc4random();
1293 		tcp_secret[idx].ts_expire = ticks + SYNCOOKIE_TIMEOUT;
1294 	}
1295 	for (data = sizeof(tcp_msstab) / sizeof(int) - 1; data > 0; data--)
1296 		if (tcp_msstab[data] <= sc->sc_peer_mss)
1297 			break;
1298 	data = (data << SYNCOOKIE_WNDBITS) | idx;
1299 	data ^= sc->sc_irs;				/* peer's iss */
1300 	MD5Init(&syn_ctx);
1301 #ifdef INET6
1302 	if (sc->sc_inc.inc_isipv6) {
1303 		MD5Add(sc->sc_inc.inc6_laddr);
1304 		MD5Add(sc->sc_inc.inc6_faddr);
1305 		add.laddr = 0;
1306 		add.faddr = 0;
1307 	} else
1308 #endif
1309 	{
1310 		add.laddr = sc->sc_inc.inc_laddr.s_addr;
1311 		add.faddr = sc->sc_inc.inc_faddr.s_addr;
1312 	}
1313 	add.lport = sc->sc_inc.inc_lport;
1314 	add.fport = sc->sc_inc.inc_fport;
1315 	add.secbits[0] = tcp_secret[idx].ts_secbits[0];
1316 	add.secbits[1] = tcp_secret[idx].ts_secbits[1];
1317 	add.secbits[2] = tcp_secret[idx].ts_secbits[2];
1318 	add.secbits[3] = tcp_secret[idx].ts_secbits[3];
1319 	MD5Add(add);
1320 	MD5Final((u_char *)&md5_buffer, &syn_ctx);
1321 	data ^= (md5_buffer[0] & ~SYNCOOKIE_WNDMASK);
1322 	return (data);
1323 }
1324 
1325 static struct syncache *
1326 syncookie_lookup(inc, th, so)
1327 	struct in_conninfo *inc;
1328 	struct tcphdr *th;
1329 	struct socket *so;
1330 {
1331 	u_int32_t md5_buffer[4];
1332 	struct syncache *sc;
1333 	u_int32_t data;
1334 	int wnd, idx;
1335 	struct md5_add add;
1336 
1337 	data = (th->th_ack - 1) ^ (th->th_seq - 1);	/* remove ISS */
1338 	idx = data & SYNCOOKIE_WNDMASK;
1339 	if (tcp_secret[idx].ts_expire < ticks ||
1340 	    sototcpcb(so)->ts_recent + SYNCOOKIE_TIMEOUT < ticks)
1341 		return (NULL);
1342 	MD5Init(&syn_ctx);
1343 #ifdef INET6
1344 	if (inc->inc_isipv6) {
1345 		MD5Add(inc->inc6_laddr);
1346 		MD5Add(inc->inc6_faddr);
1347 		add.laddr = 0;
1348 		add.faddr = 0;
1349 	} else
1350 #endif
1351 	{
1352 		add.laddr = inc->inc_laddr.s_addr;
1353 		add.faddr = inc->inc_faddr.s_addr;
1354 	}
1355 	add.lport = inc->inc_lport;
1356 	add.fport = inc->inc_fport;
1357 	add.secbits[0] = tcp_secret[idx].ts_secbits[0];
1358 	add.secbits[1] = tcp_secret[idx].ts_secbits[1];
1359 	add.secbits[2] = tcp_secret[idx].ts_secbits[2];
1360 	add.secbits[3] = tcp_secret[idx].ts_secbits[3];
1361 	MD5Add(add);
1362 	MD5Final((u_char *)&md5_buffer, &syn_ctx);
1363 	data ^= md5_buffer[0];
1364 	if ((data & ~SYNCOOKIE_DATAMASK) != 0)
1365 		return (NULL);
1366 	data = data >> SYNCOOKIE_WNDBITS;
1367 
1368 	sc = uma_zalloc(tcp_syncache.zone, M_NOWAIT);
1369 	if (sc == NULL)
1370 		return (NULL);
1371 	/*
1372 	 * Fill in the syncache values.
1373 	 * XXX duplicate code from syncache_add
1374 	 */
1375 	sc->sc_ipopts = NULL;
1376 	sc->sc_inc.inc_fport = inc->inc_fport;
1377 	sc->sc_inc.inc_lport = inc->inc_lport;
1378 #ifdef INET6
1379 	sc->sc_inc.inc_isipv6 = inc->inc_isipv6;
1380 	if (inc->inc_isipv6) {
1381 		sc->sc_inc.inc6_faddr = inc->inc6_faddr;
1382 		sc->sc_inc.inc6_laddr = inc->inc6_laddr;
1383 		sc->sc_route6.ro_rt = NULL;
1384 	} else
1385 #endif
1386 	{
1387 		sc->sc_inc.inc_faddr = inc->inc_faddr;
1388 		sc->sc_inc.inc_laddr = inc->inc_laddr;
1389 		sc->sc_route.ro_rt = NULL;
1390 	}
1391 	sc->sc_irs = th->th_seq - 1;
1392 	sc->sc_iss = th->th_ack - 1;
1393 	wnd = sbspace(&so->so_rcv);
1394 	wnd = imax(wnd, 0);
1395 	wnd = imin(wnd, TCP_MAXWIN);
1396 	sc->sc_wnd = wnd;
1397 	sc->sc_flags = 0;
1398 	sc->sc_rxtslot = 0;
1399 	sc->sc_peer_mss = tcp_msstab[data];
1400 	return (sc);
1401 }
1402