xref: /dflybsd-src/sys/netinet/tcp_syncache.c (revision 6cef7136f04e2b24a6db289e78720d6d8c60274e)
1 /*
2  * Copyright (c) 2003, 2004 Jeffrey M. Hsu.  All rights reserved.
3  * Copyright (c) 2003, 2004 The DragonFly Project.  All rights reserved.
4  *
5  * This code is derived from software contributed to The DragonFly Project
6  * by Jeffrey M. Hsu.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. Neither the name of The DragonFly Project nor the names of its
17  *    contributors may be used to endorse or promote products derived
18  *    from this software without specific, prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
23  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
24  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
25  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
27  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
29  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
30  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 /*
35  * All advertising materials mentioning features or use of this software
36  * must display the following acknowledgement:
37  *   This product includes software developed by Jeffrey M. Hsu.
38  *
39  * Copyright (c) 2001 Networks Associates Technologies, Inc.
40  * All rights reserved.
41  *
42  * This software was developed for the FreeBSD Project by Jonathan Lemon
43  * and NAI Labs, the Security Research Division of Network Associates, Inc.
44  * under DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
45  * DARPA CHATS research program.
46  *
47  * Redistribution and use in source and binary forms, with or without
48  * modification, are permitted provided that the following conditions
49  * are met:
50  * 1. Redistributions of source code must retain the above copyright
51  *    notice, this list of conditions and the following disclaimer.
52  * 2. Redistributions in binary form must reproduce the above copyright
53  *    notice, this list of conditions and the following disclaimer in the
54  *    documentation and/or other materials provided with the distribution.
55  * 3. The name of the author may not be used to endorse or promote
56  *    products derived from this software without specific prior written
57  *    permission.
58  *
59  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
60  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
61  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
62  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
63  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
64  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
65  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
66  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
67  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
68  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
69  * SUCH DAMAGE.
70  *
71  * $FreeBSD: src/sys/netinet/tcp_syncache.c,v 1.5.2.14 2003/02/24 04:02:27 silby Exp $
72  * $DragonFly: src/sys/netinet/tcp_syncache.c,v 1.35 2008/11/22 11:03:35 sephe Exp $
73  */
74 
75 #include "opt_inet.h"
76 #include "opt_inet6.h"
77 #include "opt_ipsec.h"
78 
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
82 #include <sys/sysctl.h>
83 #include <sys/malloc.h>
84 #include <sys/mbuf.h>
85 #include <sys/md5.h>
86 #include <sys/proc.h>		/* for proc0 declaration */
87 #include <sys/random.h>
88 #include <sys/socket.h>
89 #include <sys/socketvar.h>
90 #include <sys/in_cksum.h>
91 
92 #include <sys/msgport2.h>
93 #include <net/netmsg2.h>
94 
95 #include <net/if.h>
96 #include <net/route.h>
97 
98 #include <netinet/in.h>
99 #include <netinet/in_systm.h>
100 #include <netinet/ip.h>
101 #include <netinet/in_var.h>
102 #include <netinet/in_pcb.h>
103 #include <netinet/ip_var.h>
104 #include <netinet/ip6.h>
105 #ifdef INET6
106 #include <netinet/icmp6.h>
107 #include <netinet6/nd6.h>
108 #endif
109 #include <netinet6/ip6_var.h>
110 #include <netinet6/in6_pcb.h>
111 #include <netinet/tcp.h>
112 #include <netinet/tcp_fsm.h>
113 #include <netinet/tcp_seq.h>
114 #include <netinet/tcp_timer.h>
115 #include <netinet/tcp_timer2.h>
116 #include <netinet/tcp_var.h>
117 #include <netinet6/tcp6_var.h>
118 
119 #ifdef IPSEC
120 #include <netinet6/ipsec.h>
121 #ifdef INET6
122 #include <netinet6/ipsec6.h>
123 #endif
124 #include <netproto/key/key.h>
125 #endif /*IPSEC*/
126 
127 #ifdef FAST_IPSEC
128 #include <netproto/ipsec/ipsec.h>
129 #ifdef INET6
130 #include <netproto/ipsec/ipsec6.h>
131 #endif
132 #include <netproto/ipsec/key.h>
133 #define	IPSEC
134 #endif /*FAST_IPSEC*/
135 
136 static int tcp_syncookies = 1;
137 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_RW,
138     &tcp_syncookies, 0,
139     "Use TCP SYN cookies if the syncache overflows");
140 
141 static void	 syncache_drop(struct syncache *, struct syncache_head *);
142 static void	 syncache_free(struct syncache *);
143 static void	 syncache_insert(struct syncache *, struct syncache_head *);
144 struct syncache *syncache_lookup(struct in_conninfo *, struct syncache_head **);
145 static int	 syncache_respond(struct syncache *, struct mbuf *);
146 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
147 		    struct mbuf *);
148 static void	 syncache_timer(void *);
149 static u_int32_t syncookie_generate(struct syncache *);
150 static struct syncache *syncookie_lookup(struct in_conninfo *,
151 		    struct tcphdr *, struct socket *);
152 
153 /*
154  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
155  * 3 retransmits corresponds to a timeout of (1 + 2 + 4 + 8 == 15) seconds,
156  * the odds are that the user has given up attempting to connect by then.
157  */
158 #define SYNCACHE_MAXREXMTS		3
159 
160 /* Arbitrary values */
161 #define TCP_SYNCACHE_HASHSIZE		512
162 #define TCP_SYNCACHE_BUCKETLIMIT	30
163 
164 struct netmsg_sc_timer {
165 	struct netmsg nm_netmsg;
166 	struct msgrec *nm_mrec;		/* back pointer to containing msgrec */
167 };
168 
169 struct msgrec {
170 	struct netmsg_sc_timer msg;
171 	lwkt_port_t port;		/* constant after init */
172 	int slot;			/* constant after init */
173 };
174 
175 static void syncache_timer_handler(netmsg_t);
176 
177 struct tcp_syncache {
178 	u_int	hashsize;
179 	u_int	hashmask;
180 	u_int	bucket_limit;
181 	u_int	cache_limit;
182 	u_int	rexmt_limit;
183 	u_int	hash_secret;
184 };
185 static struct tcp_syncache tcp_syncache;
186 
187 TAILQ_HEAD(syncache_list, syncache);
188 
189 struct tcp_syncache_percpu {
190 	struct syncache_head	*hashbase;
191 	u_int			cache_count;
192 	struct syncache_list	timerq[SYNCACHE_MAXREXMTS + 1];
193 	struct callout		tt_timerq[SYNCACHE_MAXREXMTS + 1];
194 	struct msgrec		mrec[SYNCACHE_MAXREXMTS + 1];
195 };
196 static struct tcp_syncache_percpu tcp_syncache_percpu[MAXCPU];
197 
198 static struct lwkt_port syncache_null_rport;
199 
200 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, "TCP SYN cache");
201 
202 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RD,
203      &tcp_syncache.bucket_limit, 0, "Per-bucket hash limit for syncache");
204 
205 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_RD,
206      &tcp_syncache.cache_limit, 0, "Overall entry limit for syncache");
207 
208 /* XXX JH */
209 #if 0
210 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_RD,
211      &tcp_syncache.cache_count, 0, "Current number of entries in syncache");
212 #endif
213 
214 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RD,
215      &tcp_syncache.hashsize, 0, "Size of TCP syncache hashtable");
216 
217 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_RW,
218      &tcp_syncache.rexmt_limit, 0, "Limit on SYN/ACK retransmissions");
219 
220 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
221 
222 #define SYNCACHE_HASH(inc, mask)					\
223 	((tcp_syncache.hash_secret ^					\
224 	  (inc)->inc_faddr.s_addr ^					\
225 	  ((inc)->inc_faddr.s_addr >> 16) ^				\
226 	  (inc)->inc_fport ^ (inc)->inc_lport) & mask)
227 
228 #define SYNCACHE_HASH6(inc, mask)					\
229 	((tcp_syncache.hash_secret ^					\
230 	  (inc)->inc6_faddr.s6_addr32[0] ^				\
231 	  (inc)->inc6_faddr.s6_addr32[3] ^				\
232 	  (inc)->inc_fport ^ (inc)->inc_lport) & mask)
233 
234 #define ENDPTS_EQ(a, b) (						\
235 	(a)->ie_fport == (b)->ie_fport &&				\
236 	(a)->ie_lport == (b)->ie_lport &&				\
237 	(a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr &&			\
238 	(a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr			\
239 )
240 
241 #define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0)
242 
243 static __inline void
244 syncache_timeout(struct tcp_syncache_percpu *syncache_percpu,
245 		 struct syncache *sc, int slot)
246 {
247 	sc->sc_rxtslot = slot;
248 	sc->sc_rxttime = ticks + TCPTV_RTOBASE * tcp_backoff[slot];
249 	TAILQ_INSERT_TAIL(&syncache_percpu->timerq[slot], sc, sc_timerq);
250 	if (!callout_active(&syncache_percpu->tt_timerq[slot])) {
251 		callout_reset(&syncache_percpu->tt_timerq[slot],
252 			      TCPTV_RTOBASE * tcp_backoff[slot],
253 			      syncache_timer,
254 			      &syncache_percpu->mrec[slot]);
255 	}
256 }
257 
258 static void
259 syncache_free(struct syncache *sc)
260 {
261 	struct rtentry *rt;
262 #ifdef INET6
263 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
264 #else
265 	const boolean_t isipv6 = FALSE;
266 #endif
267 
268 	if (sc->sc_ipopts)
269 		m_free(sc->sc_ipopts);
270 
271 	rt = isipv6 ? sc->sc_route6.ro_rt : sc->sc_route.ro_rt;
272 	if (rt != NULL) {
273 		/*
274 		 * If this is the only reference to a protocol-cloned
275 		 * route, remove it immediately.
276 		 */
277 		if ((rt->rt_flags & RTF_WASCLONED) && rt->rt_refcnt == 1)
278 			rtrequest(RTM_DELETE, rt_key(rt), rt->rt_gateway,
279 				  rt_mask(rt), rt->rt_flags, NULL);
280 		RTFREE(rt);
281 	}
282 	kfree(sc, M_SYNCACHE);
283 }
284 
285 void
286 syncache_init(void)
287 {
288 	int i, cpu;
289 
290 	tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
291 	tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
292 	tcp_syncache.cache_limit =
293 	    tcp_syncache.hashsize * tcp_syncache.bucket_limit;
294 	tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
295 	tcp_syncache.hash_secret = karc4random();
296 
297 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
298 	    &tcp_syncache.hashsize);
299 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
300 	    &tcp_syncache.cache_limit);
301 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
302 	    &tcp_syncache.bucket_limit);
303 	if (!powerof2(tcp_syncache.hashsize)) {
304 		kprintf("WARNING: syncache hash size is not a power of 2.\n");
305 		tcp_syncache.hashsize = 512;	/* safe default */
306 	}
307 	tcp_syncache.hashmask = tcp_syncache.hashsize - 1;
308 
309 	lwkt_initport_replyonly_null(&syncache_null_rport);
310 
311 	for (cpu = 0; cpu < ncpus2; cpu++) {
312 		struct tcp_syncache_percpu *syncache_percpu;
313 
314 		syncache_percpu = &tcp_syncache_percpu[cpu];
315 		/* Allocate the hash table. */
316 		MALLOC(syncache_percpu->hashbase, struct syncache_head *,
317 		    tcp_syncache.hashsize * sizeof(struct syncache_head),
318 		    M_SYNCACHE, M_WAITOK);
319 
320 		/* Initialize the hash buckets. */
321 		for (i = 0; i < tcp_syncache.hashsize; i++) {
322 			struct syncache_head *bucket;
323 
324 			bucket = &syncache_percpu->hashbase[i];
325 			TAILQ_INIT(&bucket->sch_bucket);
326 			bucket->sch_length = 0;
327 		}
328 
329 		for (i = 0; i <= SYNCACHE_MAXREXMTS; i++) {
330 			/* Initialize the timer queues. */
331 			TAILQ_INIT(&syncache_percpu->timerq[i]);
332 			callout_init(&syncache_percpu->tt_timerq[i]);
333 
334 			syncache_percpu->mrec[i].slot = i;
335 			syncache_percpu->mrec[i].port = cpu_portfn(cpu);
336 			syncache_percpu->mrec[i].msg.nm_mrec =
337 				    &syncache_percpu->mrec[i];
338 			netmsg_init(&syncache_percpu->mrec[i].msg.nm_netmsg,
339 				    NULL, &syncache_null_rport,
340 				    0, syncache_timer_handler);
341 		}
342 	}
343 }
344 
345 static void
346 syncache_insert(struct syncache *sc, struct syncache_head *sch)
347 {
348 	struct tcp_syncache_percpu *syncache_percpu;
349 	struct syncache *sc2;
350 	int i;
351 
352 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
353 
354 	/*
355 	 * Make sure that we don't overflow the per-bucket
356 	 * limit or the total cache size limit.
357 	 */
358 	if (sch->sch_length >= tcp_syncache.bucket_limit) {
359 		/*
360 		 * The bucket is full, toss the oldest element.
361 		 */
362 		sc2 = TAILQ_FIRST(&sch->sch_bucket);
363 		sc2->sc_tp->ts_recent = ticks;
364 		syncache_drop(sc2, sch);
365 		tcpstat.tcps_sc_bucketoverflow++;
366 	} else if (syncache_percpu->cache_count >= tcp_syncache.cache_limit) {
367 		/*
368 		 * The cache is full.  Toss the oldest entry in the
369 		 * entire cache.  This is the front entry in the
370 		 * first non-empty timer queue with the largest
371 		 * timeout value.
372 		 */
373 		for (i = SYNCACHE_MAXREXMTS; i >= 0; i--) {
374 			sc2 = TAILQ_FIRST(&syncache_percpu->timerq[i]);
375 			while (sc2 && (sc2->sc_flags & SCF_MARKER))
376 				sc2 = TAILQ_NEXT(sc2, sc_timerq);
377 			if (sc2 != NULL)
378 				break;
379 		}
380 		sc2->sc_tp->ts_recent = ticks;
381 		syncache_drop(sc2, NULL);
382 		tcpstat.tcps_sc_cacheoverflow++;
383 	}
384 
385 	/* Initialize the entry's timer. */
386 	syncache_timeout(syncache_percpu, sc, 0);
387 
388 	/* Put it into the bucket. */
389 	TAILQ_INSERT_TAIL(&sch->sch_bucket, sc, sc_hash);
390 	sch->sch_length++;
391 	syncache_percpu->cache_count++;
392 	tcpstat.tcps_sc_added++;
393 }
394 
395 void
396 syncache_destroy(struct tcpcb *tp)
397 {
398 	struct tcp_syncache_percpu *syncache_percpu;
399 	struct syncache_head *bucket;
400 	struct syncache *sc;
401 	int i;
402 
403 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
404 	sc = NULL;
405 
406 	for (i = 0; i < tcp_syncache.hashsize; i++) {
407 		bucket = &syncache_percpu->hashbase[i];
408 		TAILQ_FOREACH(sc, &bucket->sch_bucket, sc_hash) {
409 			if (sc->sc_tp == tp) {
410 				sc->sc_tp = NULL;
411 				tp->t_flags &= ~TF_SYNCACHE;
412 				break;
413 			}
414 		}
415 	}
416 	kprintf("Warning: delete stale syncache for tp=%p, sc=%p\n", tp, sc);
417 }
418 
419 static void
420 syncache_drop(struct syncache *sc, struct syncache_head *sch)
421 {
422 	struct tcp_syncache_percpu *syncache_percpu;
423 #ifdef INET6
424 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
425 #else
426 	const boolean_t isipv6 = FALSE;
427 #endif
428 
429 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
430 
431 	if (sch == NULL) {
432 		if (isipv6) {
433 			sch = &syncache_percpu->hashbase[
434 			    SYNCACHE_HASH6(&sc->sc_inc, tcp_syncache.hashmask)];
435 		} else {
436 			sch = &syncache_percpu->hashbase[
437 			    SYNCACHE_HASH(&sc->sc_inc, tcp_syncache.hashmask)];
438 		}
439 	}
440 
441 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
442 	sch->sch_length--;
443 	syncache_percpu->cache_count--;
444 
445 	/*
446 	 * Cleanup
447 	 */
448 	if (sc->sc_tp) {
449 		sc->sc_tp->t_flags &= ~TF_SYNCACHE;
450 		sc->sc_tp = NULL;
451 	}
452 
453 	/*
454 	 * Remove the entry from the syncache timer/timeout queue.  Note
455 	 * that we do not try to stop any running timer since we do not know
456 	 * whether the timer's message is in-transit or not.  Since timeouts
457 	 * are fairly long, taking an unneeded callout does not detrimentally
458 	 * effect performance.
459 	 */
460 	TAILQ_REMOVE(&syncache_percpu->timerq[sc->sc_rxtslot], sc, sc_timerq);
461 
462 	syncache_free(sc);
463 }
464 
465 /*
466  * Place a timeout message on the TCP thread's message queue.
467  * This routine runs in soft interrupt context.
468  *
469  * An invariant is for this routine to be called, the callout must
470  * have been active.  Note that the callout is not deactivated until
471  * after the message has been processed in syncache_timer_handler() below.
472  */
473 static void
474 syncache_timer(void *p)
475 {
476 	struct netmsg_sc_timer *msg = p;
477 
478 	lwkt_sendmsg(msg->nm_mrec->port, &msg->nm_netmsg.nm_lmsg);
479 }
480 
481 /*
482  * Service a timer message queued by timer expiration.
483  * This routine runs in the TCP protocol thread.
484  *
485  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
486  * If we have retransmitted an entry the maximum number of times, expire it.
487  *
488  * When we finish processing timed-out entries, we restart the timer if there
489  * are any entries still on the queue and deactivate it otherwise.  Only after
490  * a timer has been deactivated here can it be restarted by syncache_timeout().
491  */
492 static void
493 syncache_timer_handler(netmsg_t netmsg)
494 {
495 	struct tcp_syncache_percpu *syncache_percpu;
496 	struct syncache *sc;
497 	struct syncache marker;
498 	struct syncache_list *list;
499 	struct inpcb *inp;
500 	int slot;
501 
502 	slot = ((struct netmsg_sc_timer *)netmsg)->nm_mrec->slot;
503 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
504 
505 	list = &syncache_percpu->timerq[slot];
506 
507 	/*
508 	 * Use a marker to keep our place in the scan.  syncache_drop()
509 	 * can block and cause any next pointer we cache to become stale.
510 	 */
511 	marker.sc_flags = SCF_MARKER;
512 	TAILQ_INSERT_HEAD(list, &marker, sc_timerq);
513 
514 	while ((sc = TAILQ_NEXT(&marker, sc_timerq)) != NULL) {
515 		/*
516 		 * Move the marker.
517 		 */
518 		TAILQ_REMOVE(list, &marker, sc_timerq);
519 		TAILQ_INSERT_AFTER(list, sc, &marker, sc_timerq);
520 
521 		if (sc->sc_flags & SCF_MARKER)
522 			continue;
523 
524 		if (ticks < sc->sc_rxttime)
525 			break;	/* finished because timerq sorted by time */
526 		if (sc->sc_tp == NULL) {
527 			syncache_drop(sc, NULL);
528 			tcpstat.tcps_sc_stale++;
529 			continue;
530 		}
531 		inp = sc->sc_tp->t_inpcb;
532 		if (slot == SYNCACHE_MAXREXMTS ||
533 		    slot >= tcp_syncache.rexmt_limit ||
534 		    inp == NULL ||
535 		    inp->inp_gencnt != sc->sc_inp_gencnt) {
536 			syncache_drop(sc, NULL);
537 			tcpstat.tcps_sc_stale++;
538 			continue;
539 		}
540 		/*
541 		 * syncache_respond() may call back into the syncache to
542 		 * to modify another entry, so do not obtain the next
543 		 * entry on the timer chain until it has completed.
544 		 */
545 		syncache_respond(sc, NULL);
546 		tcpstat.tcps_sc_retransmitted++;
547 		TAILQ_REMOVE(list, sc, sc_timerq);
548 		syncache_timeout(syncache_percpu, sc, slot + 1);
549 	}
550 	TAILQ_REMOVE(list, &marker, sc_timerq);
551 
552 	if (sc != NULL) {
553 		callout_reset(&syncache_percpu->tt_timerq[slot],
554 			      sc->sc_rxttime - ticks, syncache_timer,
555 			      &syncache_percpu->mrec[slot]);
556 	} else {
557 		callout_deactivate(&syncache_percpu->tt_timerq[slot]);
558 	}
559 	lwkt_replymsg(&netmsg->nm_lmsg, 0);
560 }
561 
562 /*
563  * Find an entry in the syncache.
564  */
565 struct syncache *
566 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
567 {
568 	struct tcp_syncache_percpu *syncache_percpu;
569 	struct syncache *sc;
570 	struct syncache_head *sch;
571 
572 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
573 #ifdef INET6
574 	if (inc->inc_isipv6) {
575 		sch = &syncache_percpu->hashbase[
576 		    SYNCACHE_HASH6(inc, tcp_syncache.hashmask)];
577 		*schp = sch;
578 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
579 			if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
580 				return (sc);
581 	} else
582 #endif
583 	{
584 		sch = &syncache_percpu->hashbase[
585 		    SYNCACHE_HASH(inc, tcp_syncache.hashmask)];
586 		*schp = sch;
587 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
588 #ifdef INET6
589 			if (sc->sc_inc.inc_isipv6)
590 				continue;
591 #endif
592 			if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
593 				return (sc);
594 		}
595 	}
596 	return (NULL);
597 }
598 
599 /*
600  * This function is called when we get a RST for a
601  * non-existent connection, so that we can see if the
602  * connection is in the syn cache.  If it is, zap it.
603  */
604 void
605 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
606 {
607 	struct syncache *sc;
608 	struct syncache_head *sch;
609 
610 	sc = syncache_lookup(inc, &sch);
611 	if (sc == NULL) {
612 		return;
613 	}
614 	/*
615 	 * If the RST bit is set, check the sequence number to see
616 	 * if this is a valid reset segment.
617 	 * RFC 793 page 37:
618 	 *   In all states except SYN-SENT, all reset (RST) segments
619 	 *   are validated by checking their SEQ-fields.  A reset is
620 	 *   valid if its sequence number is in the window.
621 	 *
622 	 *   The sequence number in the reset segment is normally an
623 	 *   echo of our outgoing acknowlegement numbers, but some hosts
624 	 *   send a reset with the sequence number at the rightmost edge
625 	 *   of our receive window, and we have to handle this case.
626 	 */
627 	if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
628 	    SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
629 		syncache_drop(sc, sch);
630 		tcpstat.tcps_sc_reset++;
631 	}
632 }
633 
634 void
635 syncache_badack(struct in_conninfo *inc)
636 {
637 	struct syncache *sc;
638 	struct syncache_head *sch;
639 
640 	sc = syncache_lookup(inc, &sch);
641 	if (sc != NULL) {
642 		syncache_drop(sc, sch);
643 		tcpstat.tcps_sc_badack++;
644 	}
645 }
646 
647 void
648 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th)
649 {
650 	struct syncache *sc;
651 	struct syncache_head *sch;
652 
653 	/* we are called at splnet() here */
654 	sc = syncache_lookup(inc, &sch);
655 	if (sc == NULL)
656 		return;
657 
658 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
659 	if (ntohl(th->th_seq) != sc->sc_iss)
660 		return;
661 
662 	/*
663 	 * If we've rertransmitted 3 times and this is our second error,
664 	 * we remove the entry.  Otherwise, we allow it to continue on.
665 	 * This prevents us from incorrectly nuking an entry during a
666 	 * spurious network outage.
667 	 *
668 	 * See tcp_notify().
669 	 */
670 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxtslot < 3) {
671 		sc->sc_flags |= SCF_UNREACH;
672 		return;
673 	}
674 	syncache_drop(sc, sch);
675 	tcpstat.tcps_sc_unreach++;
676 }
677 
678 /*
679  * Build a new TCP socket structure from a syncache entry.
680  *
681  * This is called from the context of the SYN+ACK
682  */
683 static struct socket *
684 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
685 {
686 	struct inpcb *inp = NULL, *linp;
687 	struct socket *so;
688 	struct tcpcb *tp;
689 	lwkt_port_t port;
690 #ifdef INET6
691 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
692 #else
693 	const boolean_t isipv6 = FALSE;
694 #endif
695 
696 	/*
697 	 * Ok, create the full blown connection, and set things up
698 	 * as they would have been set up if we had created the
699 	 * connection when the SYN arrived.  If we can't create
700 	 * the connection, abort it.
701 	 */
702 	so = sonewconn(lso, SS_ISCONNECTED);
703 	if (so == NULL) {
704 		/*
705 		 * Drop the connection; we will send a RST if the peer
706 		 * retransmits the ACK,
707 		 */
708 		tcpstat.tcps_listendrop++;
709 		goto abort;
710 	}
711 
712 	/*
713 	 * Set the protocol processing port for the socket to the current
714 	 * port (that the connection came in on).
715 	 */
716 	sosetport(so, &curthread->td_msgport);
717 
718 	/*
719 	 * Insert new socket into hash list.
720 	 */
721 	inp = so->so_pcb;
722 	inp->inp_inc.inc_isipv6 = sc->sc_inc.inc_isipv6;
723 	if (isipv6) {
724 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
725 	} else {
726 #ifdef INET6
727 		inp->inp_vflag &= ~INP_IPV6;
728 		inp->inp_vflag |= INP_IPV4;
729 		inp->inp_flags &= ~IN6P_IPV6_V6ONLY;
730 #endif
731 		inp->inp_laddr = sc->sc_inc.inc_laddr;
732 	}
733 	inp->inp_lport = sc->sc_inc.inc_lport;
734 	if (in_pcbinsporthash(inp) != 0) {
735 		/*
736 		 * Undo the assignments above if we failed to
737 		 * put the PCB on the hash lists.
738 		 */
739 		if (isipv6)
740 			inp->in6p_laddr = kin6addr_any;
741 		else
742 			inp->inp_laddr.s_addr = INADDR_ANY;
743 		inp->inp_lport = 0;
744 		goto abort;
745 	}
746 	linp = so->so_pcb;
747 #ifdef IPSEC
748 	/* copy old policy into new socket's */
749 	if (ipsec_copy_policy(linp->inp_sp, inp->inp_sp))
750 		kprintf("syncache_expand: could not copy policy\n");
751 #endif
752 	if (isipv6) {
753 		struct in6_addr laddr6;
754 		struct sockaddr_in6 sin6;
755 		/*
756 		 * Inherit socket options from the listening socket.
757 		 * Note that in6p_inputopts are not (and should not be)
758 		 * copied, since it stores previously received options and is
759 		 * used to detect if each new option is different than the
760 		 * previous one and hence should be passed to a user.
761 		 * If we copied in6p_inputopts, a user would not be able to
762 		 * receive options just after calling the accept system call.
763 		 */
764 		inp->inp_flags |= linp->inp_flags & INP_CONTROLOPTS;
765 		if (linp->in6p_outputopts)
766 			inp->in6p_outputopts =
767 			    ip6_copypktopts(linp->in6p_outputopts, M_INTWAIT);
768 		inp->in6p_route = sc->sc_route6;
769 		sc->sc_route6.ro_rt = NULL;
770 
771 		sin6.sin6_family = AF_INET6;
772 		sin6.sin6_len = sizeof sin6;
773 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
774 		sin6.sin6_port = sc->sc_inc.inc_fport;
775 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
776 		laddr6 = inp->in6p_laddr;
777 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
778 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
779 		if (in6_pcbconnect(inp, (struct sockaddr *)&sin6, &thread0)) {
780 			inp->in6p_laddr = laddr6;
781 			goto abort;
782 		}
783 	} else {
784 		struct in_addr laddr;
785 		struct sockaddr_in sin;
786 
787 		inp->inp_options = ip_srcroute(m);
788 		if (inp->inp_options == NULL) {
789 			inp->inp_options = sc->sc_ipopts;
790 			sc->sc_ipopts = NULL;
791 		}
792 		inp->inp_route = sc->sc_route;
793 		sc->sc_route.ro_rt = NULL;
794 
795 		sin.sin_family = AF_INET;
796 		sin.sin_len = sizeof sin;
797 		sin.sin_addr = sc->sc_inc.inc_faddr;
798 		sin.sin_port = sc->sc_inc.inc_fport;
799 		bzero(sin.sin_zero, sizeof sin.sin_zero);
800 		laddr = inp->inp_laddr;
801 		if (inp->inp_laddr.s_addr == INADDR_ANY)
802 			inp->inp_laddr = sc->sc_inc.inc_laddr;
803 		if (in_pcbconnect(inp, (struct sockaddr *)&sin, &thread0)) {
804 			inp->inp_laddr = laddr;
805 			goto abort;
806 		}
807 	}
808 
809 	/*
810 	 * The current port should be in the context of the SYN+ACK and
811 	 * so should match the tcp address port.
812 	 *
813 	 * XXX we may be running on the netisr thread instead of a tcp
814 	 *     thread, in which case port will not match
815 	 *     curthread->td_msgport.
816 	 */
817 	if (isipv6) {
818 		port = tcp6_addrport();
819 	} else {
820 		port = tcp_addrport(inp->inp_faddr.s_addr, inp->inp_fport,
821 				    inp->inp_laddr.s_addr, inp->inp_lport);
822 	}
823 	/*KKASSERT(port == &curthread->td_msgport);*/
824 
825 	tp = intotcpcb(inp);
826 	tp->t_state = TCPS_SYN_RECEIVED;
827 	tp->iss = sc->sc_iss;
828 	tp->irs = sc->sc_irs;
829 	tcp_rcvseqinit(tp);
830 	tcp_sendseqinit(tp);
831 	tp->snd_wl1 = sc->sc_irs;
832 	tp->rcv_up = sc->sc_irs + 1;
833 	tp->rcv_wnd = sc->sc_wnd;
834 	tp->rcv_adv += tp->rcv_wnd;
835 
836 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH | TF_NODELAY);
837 	if (sc->sc_flags & SCF_NOOPT)
838 		tp->t_flags |= TF_NOOPT;
839 	if (sc->sc_flags & SCF_WINSCALE) {
840 		tp->t_flags |= TF_REQ_SCALE | TF_RCVD_SCALE;
841 		tp->requested_s_scale = sc->sc_requested_s_scale;
842 		tp->request_r_scale = sc->sc_request_r_scale;
843 	}
844 	if (sc->sc_flags & SCF_TIMESTAMP) {
845 		tp->t_flags |= TF_REQ_TSTMP | TF_RCVD_TSTMP;
846 		tp->ts_recent = sc->sc_tsrecent;
847 		tp->ts_recent_age = ticks;
848 	}
849 	if (sc->sc_flags & SCF_SACK_PERMITTED)
850 		tp->t_flags |= TF_SACK_PERMITTED;
851 
852 #ifdef TCP_SIGNATURE
853 	if (sc->sc_flags & SCF_SIGNATURE)
854 		tp->t_flags |= TF_SIGNATURE;
855 #endif /* TCP_SIGNATURE */
856 
857 
858 	tcp_mss(tp, sc->sc_peer_mss);
859 
860 	/*
861 	 * If the SYN,ACK was retransmitted, reset cwnd to 1 segment.
862 	 */
863 	if (sc->sc_rxtslot != 0)
864 		tp->snd_cwnd = tp->t_maxseg;
865 	tcp_create_timermsg(tp, port);
866 	tcp_callout_reset(tp, tp->tt_keep, tcp_keepinit, tcp_timer_keep);
867 
868 	tcpstat.tcps_accepts++;
869 	return (so);
870 
871 abort:
872 	if (so != NULL)
873 		soabort_oncpu(so);
874 	return (NULL);
875 }
876 
877 /*
878  * This function gets called when we receive an ACK for a
879  * socket in the LISTEN state.  We look up the connection
880  * in the syncache, and if its there, we pull it out of
881  * the cache and turn it into a full-blown connection in
882  * the SYN-RECEIVED state.
883  */
884 int
885 syncache_expand(struct in_conninfo *inc, struct tcphdr *th, struct socket **sop,
886 		struct mbuf *m)
887 {
888 	struct syncache *sc;
889 	struct syncache_head *sch;
890 	struct socket *so;
891 
892 	sc = syncache_lookup(inc, &sch);
893 	if (sc == NULL) {
894 		/*
895 		 * There is no syncache entry, so see if this ACK is
896 		 * a returning syncookie.  To do this, first:
897 		 *  A. See if this socket has had a syncache entry dropped in
898 		 *     the past.  We don't want to accept a bogus syncookie
899 		 *     if we've never received a SYN.
900 		 *  B. check that the syncookie is valid.  If it is, then
901 		 *     cobble up a fake syncache entry, and return.
902 		 */
903 		if (!tcp_syncookies)
904 			return (0);
905 		sc = syncookie_lookup(inc, th, *sop);
906 		if (sc == NULL)
907 			return (0);
908 		sch = NULL;
909 		tcpstat.tcps_sc_recvcookie++;
910 	}
911 
912 	/*
913 	 * If seg contains an ACK, but not for our SYN/ACK, send a RST.
914 	 */
915 	if (th->th_ack != sc->sc_iss + 1)
916 		return (0);
917 
918 	so = syncache_socket(sc, *sop, m);
919 	if (so == NULL) {
920 #if 0
921 resetandabort:
922 		/* XXXjlemon check this - is this correct? */
923 		tcp_respond(NULL, m, m, th,
924 		    th->th_seq + tlen, (tcp_seq)0, TH_RST | TH_ACK);
925 #endif
926 		m_freem(m);			/* XXX only needed for above */
927 		tcpstat.tcps_sc_aborted++;
928 	} else {
929 		tcpstat.tcps_sc_completed++;
930 	}
931 	if (sch == NULL)
932 		syncache_free(sc);
933 	else
934 		syncache_drop(sc, sch);
935 	*sop = so;
936 	return (1);
937 }
938 
939 /*
940  * Given a LISTEN socket and an inbound SYN request, add
941  * this to the syn cache, and send back a segment:
942  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
943  * to the source.
944  *
945  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
946  * Doing so would require that we hold onto the data and deliver it
947  * to the application.  However, if we are the target of a SYN-flood
948  * DoS attack, an attacker could send data which would eventually
949  * consume all available buffer space if it were ACKed.  By not ACKing
950  * the data, we avoid this DoS scenario.
951  */
952 int
953 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
954 	     struct socket **sop, struct mbuf *m)
955 {
956 	struct tcp_syncache_percpu *syncache_percpu;
957 	struct tcpcb *tp;
958 	struct socket *so;
959 	struct syncache *sc = NULL;
960 	struct syncache_head *sch;
961 	struct mbuf *ipopts = NULL;
962 	int win;
963 
964 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
965 	so = *sop;
966 	tp = sototcpcb(so);
967 
968 	/*
969 	 * Remember the IP options, if any.
970 	 */
971 #ifdef INET6
972 	if (!inc->inc_isipv6)
973 #endif
974 		ipopts = ip_srcroute(m);
975 
976 	/*
977 	 * See if we already have an entry for this connection.
978 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
979 	 *
980 	 * XXX
981 	 * The syncache should be re-initialized with the contents
982 	 * of the new SYN which may have different options.
983 	 */
984 	sc = syncache_lookup(inc, &sch);
985 	if (sc != NULL) {
986 		tcpstat.tcps_sc_dupsyn++;
987 		if (ipopts) {
988 			/*
989 			 * If we were remembering a previous source route,
990 			 * forget it and use the new one we've been given.
991 			 */
992 			if (sc->sc_ipopts)
993 				m_free(sc->sc_ipopts);
994 			sc->sc_ipopts = ipopts;
995 		}
996 		/*
997 		 * Update timestamp if present.
998 		 */
999 		if (sc->sc_flags & SCF_TIMESTAMP)
1000 			sc->sc_tsrecent = to->to_tsval;
1001 
1002 		/* Just update the TOF_SACK_PERMITTED for now. */
1003 		if (tcp_do_sack && (to->to_flags & TOF_SACK_PERMITTED))
1004 			sc->sc_flags |= SCF_SACK_PERMITTED;
1005 		else
1006 			sc->sc_flags &= ~SCF_SACK_PERMITTED;
1007 
1008 		/*
1009 		 * PCB may have changed, pick up new values.
1010 		 */
1011 		if (sc->sc_tp) {
1012 			sc->sc_tp->t_flags &= ~TF_SYNCACHE;
1013 			tp->t_flags |= TF_SYNCACHE;
1014 		}
1015 		sc->sc_tp = tp;
1016 		sc->sc_inp_gencnt = tp->t_inpcb->inp_gencnt;
1017 		if (syncache_respond(sc, m) == 0) {
1018 			TAILQ_REMOVE(&syncache_percpu->timerq[sc->sc_rxtslot],
1019 				     sc, sc_timerq);
1020 			syncache_timeout(syncache_percpu, sc, sc->sc_rxtslot);
1021 			tcpstat.tcps_sndacks++;
1022 			tcpstat.tcps_sndtotal++;
1023 		}
1024 		*sop = NULL;
1025 		return (1);
1026 	}
1027 
1028 	/*
1029 	 * Fill in the syncache values.
1030 	 */
1031 	sc = kmalloc(sizeof(struct syncache), M_SYNCACHE, M_WAITOK|M_ZERO);
1032 	sc->sc_inp_gencnt = tp->t_inpcb->inp_gencnt;
1033 	sc->sc_ipopts = ipopts;
1034 	sc->sc_inc.inc_fport = inc->inc_fport;
1035 	sc->sc_inc.inc_lport = inc->inc_lport;
1036 	sc->sc_tp = tp;
1037 	tp->t_flags |= TF_SYNCACHE;
1038 #ifdef INET6
1039 	sc->sc_inc.inc_isipv6 = inc->inc_isipv6;
1040 	if (inc->inc_isipv6) {
1041 		sc->sc_inc.inc6_faddr = inc->inc6_faddr;
1042 		sc->sc_inc.inc6_laddr = inc->inc6_laddr;
1043 		sc->sc_route6.ro_rt = NULL;
1044 	} else
1045 #endif
1046 	{
1047 		sc->sc_inc.inc_faddr = inc->inc_faddr;
1048 		sc->sc_inc.inc_laddr = inc->inc_laddr;
1049 		sc->sc_route.ro_rt = NULL;
1050 	}
1051 	sc->sc_irs = th->th_seq;
1052 	sc->sc_flags = 0;
1053 	sc->sc_peer_mss = to->to_flags & TOF_MSS ? to->to_mss : 0;
1054 	if (tcp_syncookies)
1055 		sc->sc_iss = syncookie_generate(sc);
1056 	else
1057 		sc->sc_iss = karc4random();
1058 
1059 	/* Initial receive window: clip ssb_space to [0 .. TCP_MAXWIN] */
1060 	win = ssb_space(&so->so_rcv);
1061 	win = imax(win, 0);
1062 	win = imin(win, TCP_MAXWIN);
1063 	sc->sc_wnd = win;
1064 
1065 	if (tcp_do_rfc1323) {
1066 		/*
1067 		 * A timestamp received in a SYN makes
1068 		 * it ok to send timestamp requests and replies.
1069 		 */
1070 		if (to->to_flags & TOF_TS) {
1071 			sc->sc_tsrecent = to->to_tsval;
1072 			sc->sc_flags |= SCF_TIMESTAMP;
1073 		}
1074 		if (to->to_flags & TOF_SCALE) {
1075 			int wscale = TCP_MIN_WINSHIFT;
1076 
1077 			/* Compute proper scaling value from buffer space */
1078 			while (wscale < TCP_MAX_WINSHIFT &&
1079 			    (TCP_MAXWIN << wscale) < so->so_rcv.ssb_hiwat) {
1080 				wscale++;
1081 			}
1082 			sc->sc_request_r_scale = wscale;
1083 			sc->sc_requested_s_scale = to->to_requested_s_scale;
1084 			sc->sc_flags |= SCF_WINSCALE;
1085 		}
1086 	}
1087 	if (tcp_do_sack && (to->to_flags & TOF_SACK_PERMITTED))
1088 		sc->sc_flags |= SCF_SACK_PERMITTED;
1089 	if (tp->t_flags & TF_NOOPT)
1090 		sc->sc_flags = SCF_NOOPT;
1091 #ifdef TCP_SIGNATURE
1092 	/*
1093 	 * If listening socket requested TCP digests, and received SYN
1094 	 * contains the option, flag this in the syncache so that
1095 	 * syncache_respond() will do the right thing with the SYN+ACK.
1096 	 * XXX Currently we always record the option by default and will
1097 	 * attempt to use it in syncache_respond().
1098 	 */
1099 	if (to->to_flags & TOF_SIGNATURE)
1100 		sc->sc_flags = SCF_SIGNATURE;
1101 #endif /* TCP_SIGNATURE */
1102 
1103 	if (syncache_respond(sc, m) == 0) {
1104 		syncache_insert(sc, sch);
1105 		tcpstat.tcps_sndacks++;
1106 		tcpstat.tcps_sndtotal++;
1107 	} else {
1108 		syncache_free(sc);
1109 		tcpstat.tcps_sc_dropped++;
1110 	}
1111 	*sop = NULL;
1112 	return (1);
1113 }
1114 
1115 static int
1116 syncache_respond(struct syncache *sc, struct mbuf *m)
1117 {
1118 	u_int8_t *optp;
1119 	int optlen, error;
1120 	u_int16_t tlen, hlen, mssopt;
1121 	struct ip *ip = NULL;
1122 	struct rtentry *rt;
1123 	struct tcphdr *th;
1124 	struct ip6_hdr *ip6 = NULL;
1125 #ifdef INET6
1126 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
1127 #else
1128 	const boolean_t isipv6 = FALSE;
1129 #endif
1130 
1131 	if (isipv6) {
1132 		rt = tcp_rtlookup6(&sc->sc_inc);
1133 		if (rt != NULL)
1134 			mssopt = rt->rt_ifp->if_mtu -
1135 			     (sizeof(struct ip6_hdr) + sizeof(struct tcphdr));
1136 		else
1137 			mssopt = tcp_v6mssdflt;
1138 		hlen = sizeof(struct ip6_hdr);
1139 	} else {
1140 		rt = tcp_rtlookup(&sc->sc_inc);
1141 		if (rt != NULL)
1142 			mssopt = rt->rt_ifp->if_mtu -
1143 			     (sizeof(struct ip) + sizeof(struct tcphdr));
1144 		else
1145 			mssopt = tcp_mssdflt;
1146 		hlen = sizeof(struct ip);
1147 	}
1148 
1149 	/* Compute the size of the TCP options. */
1150 	if (sc->sc_flags & SCF_NOOPT) {
1151 		optlen = 0;
1152 	} else {
1153 		optlen = TCPOLEN_MAXSEG +
1154 		    ((sc->sc_flags & SCF_WINSCALE) ? 4 : 0) +
1155 		    ((sc->sc_flags & SCF_TIMESTAMP) ? TCPOLEN_TSTAMP_APPA : 0) +
1156 		    ((sc->sc_flags & SCF_SACK_PERMITTED) ?
1157 			TCPOLEN_SACK_PERMITTED_ALIGNED : 0);
1158 #ifdef TCP_SIGNATURE
1159 				optlen += ((sc->sc_flags & SCF_SIGNATURE) ?
1160 						(TCPOLEN_SIGNATURE + 2) : 0);
1161 #endif /* TCP_SIGNATURE */
1162 	}
1163 	tlen = hlen + sizeof(struct tcphdr) + optlen;
1164 
1165 	/*
1166 	 * XXX
1167 	 * assume that the entire packet will fit in a header mbuf
1168 	 */
1169 	KASSERT(max_linkhdr + tlen <= MHLEN, ("syncache: mbuf too small"));
1170 
1171 	/*
1172 	 * XXX shouldn't this reuse the mbuf if possible ?
1173 	 * Create the IP+TCP header from scratch.
1174 	 */
1175 	if (m)
1176 		m_freem(m);
1177 
1178 	m = m_gethdr(MB_DONTWAIT, MT_HEADER);
1179 	if (m == NULL)
1180 		return (ENOBUFS);
1181 	m->m_data += max_linkhdr;
1182 	m->m_len = tlen;
1183 	m->m_pkthdr.len = tlen;
1184 	m->m_pkthdr.rcvif = NULL;
1185 
1186 	if (isipv6) {
1187 		ip6 = mtod(m, struct ip6_hdr *);
1188 		ip6->ip6_vfc = IPV6_VERSION;
1189 		ip6->ip6_nxt = IPPROTO_TCP;
1190 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1191 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1192 		ip6->ip6_plen = htons(tlen - hlen);
1193 		/* ip6_hlim is set after checksum */
1194 		/* ip6_flow = ??? */
1195 
1196 		th = (struct tcphdr *)(ip6 + 1);
1197 	} else {
1198 		ip = mtod(m, struct ip *);
1199 		ip->ip_v = IPVERSION;
1200 		ip->ip_hl = sizeof(struct ip) >> 2;
1201 		ip->ip_len = tlen;
1202 		ip->ip_id = 0;
1203 		ip->ip_off = 0;
1204 		ip->ip_sum = 0;
1205 		ip->ip_p = IPPROTO_TCP;
1206 		ip->ip_src = sc->sc_inc.inc_laddr;
1207 		ip->ip_dst = sc->sc_inc.inc_faddr;
1208 		ip->ip_ttl = sc->sc_tp->t_inpcb->inp_ip_ttl;   /* XXX */
1209 		ip->ip_tos = sc->sc_tp->t_inpcb->inp_ip_tos;   /* XXX */
1210 
1211 		/*
1212 		 * See if we should do MTU discovery.  Route lookups are
1213 		 * expensive, so we will only unset the DF bit if:
1214 		 *
1215 		 *	1) path_mtu_discovery is disabled
1216 		 *	2) the SCF_UNREACH flag has been set
1217 		 */
1218 		if (path_mtu_discovery
1219 		    && ((sc->sc_flags & SCF_UNREACH) == 0)) {
1220 		       ip->ip_off |= IP_DF;
1221 		}
1222 
1223 		th = (struct tcphdr *)(ip + 1);
1224 	}
1225 	th->th_sport = sc->sc_inc.inc_lport;
1226 	th->th_dport = sc->sc_inc.inc_fport;
1227 
1228 	th->th_seq = htonl(sc->sc_iss);
1229 	th->th_ack = htonl(sc->sc_irs + 1);
1230 	th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1231 	th->th_x2 = 0;
1232 	th->th_flags = TH_SYN | TH_ACK;
1233 	th->th_win = htons(sc->sc_wnd);
1234 	th->th_urp = 0;
1235 
1236 	/* Tack on the TCP options. */
1237 	if (optlen == 0)
1238 		goto no_options;
1239 	optp = (u_int8_t *)(th + 1);
1240 	*optp++ = TCPOPT_MAXSEG;
1241 	*optp++ = TCPOLEN_MAXSEG;
1242 	*optp++ = (mssopt >> 8) & 0xff;
1243 	*optp++ = mssopt & 0xff;
1244 
1245 	if (sc->sc_flags & SCF_WINSCALE) {
1246 		*((u_int32_t *)optp) = htonl(TCPOPT_NOP << 24 |
1247 		    TCPOPT_WINDOW << 16 | TCPOLEN_WINDOW << 8 |
1248 		    sc->sc_request_r_scale);
1249 		optp += 4;
1250 	}
1251 
1252 	if (sc->sc_flags & SCF_TIMESTAMP) {
1253 		u_int32_t *lp = (u_int32_t *)(optp);
1254 
1255 		/* Form timestamp option as shown in appendix A of RFC 1323. */
1256 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
1257 		*lp++ = htonl(ticks);
1258 		*lp   = htonl(sc->sc_tsrecent);
1259 		optp += TCPOLEN_TSTAMP_APPA;
1260 	}
1261 
1262 #ifdef TCP_SIGNATURE
1263 	/*
1264 	 * Handle TCP-MD5 passive opener response.
1265 	 */
1266 	if (sc->sc_flags & SCF_SIGNATURE) {
1267 		u_int8_t *bp = optp;
1268 		int i;
1269 
1270 		*bp++ = TCPOPT_SIGNATURE;
1271 		*bp++ = TCPOLEN_SIGNATURE;
1272 		for (i = 0; i < TCP_SIGLEN; i++)
1273 			*bp++ = 0;
1274 		tcpsignature_compute(m, 0, optlen,
1275 				optp + 2, IPSEC_DIR_OUTBOUND);
1276 		*bp++ = TCPOPT_NOP;
1277 		*bp++ = TCPOPT_EOL;
1278 		optp += TCPOLEN_SIGNATURE + 2;
1279 }
1280 #endif /* TCP_SIGNATURE */
1281 
1282 	if (sc->sc_flags & SCF_SACK_PERMITTED) {
1283 		*((u_int32_t *)optp) = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
1284 		optp += TCPOLEN_SACK_PERMITTED_ALIGNED;
1285 	}
1286 
1287 no_options:
1288 	if (isipv6) {
1289 		struct route_in6 *ro6 = &sc->sc_route6;
1290 
1291 		th->th_sum = 0;
1292 		th->th_sum = in6_cksum(m, IPPROTO_TCP, hlen, tlen - hlen);
1293 		ip6->ip6_hlim = in6_selecthlim(NULL,
1294 		    ro6->ro_rt ? ro6->ro_rt->rt_ifp : NULL);
1295 		error = ip6_output(m, NULL, ro6, 0, NULL, NULL,
1296 				sc->sc_tp->t_inpcb);
1297 	} else {
1298 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1299 				       htons(tlen - hlen + IPPROTO_TCP));
1300 		m->m_pkthdr.csum_flags = CSUM_TCP;
1301 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1302 		error = ip_output(m, sc->sc_ipopts, &sc->sc_route,
1303 				  IP_DEBUGROUTE, NULL, sc->sc_tp->t_inpcb);
1304 	}
1305 	return (error);
1306 }
1307 
1308 /*
1309  * cookie layers:
1310  *
1311  *	|. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .|
1312  *	| peer iss                                                      |
1313  *	| MD5(laddr,faddr,secret,lport,fport)             |. . . . . . .|
1314  *	|                     0                       |(A)|             |
1315  * (A): peer mss index
1316  */
1317 
1318 /*
1319  * The values below are chosen to minimize the size of the tcp_secret
1320  * table, as well as providing roughly a 16 second lifetime for the cookie.
1321  */
1322 
1323 #define SYNCOOKIE_WNDBITS	5	/* exposed bits for window indexing */
1324 #define SYNCOOKIE_TIMESHIFT	1	/* scale ticks to window time units */
1325 
1326 #define SYNCOOKIE_WNDMASK	((1 << SYNCOOKIE_WNDBITS) - 1)
1327 #define SYNCOOKIE_NSECRETS	(1 << SYNCOOKIE_WNDBITS)
1328 #define SYNCOOKIE_TIMEOUT \
1329     (hz * (1 << SYNCOOKIE_WNDBITS) / (1 << SYNCOOKIE_TIMESHIFT))
1330 #define SYNCOOKIE_DATAMASK	((3 << SYNCOOKIE_WNDBITS) | SYNCOOKIE_WNDMASK)
1331 
1332 static struct {
1333 	u_int32_t	ts_secbits[4];
1334 	u_int		ts_expire;
1335 } tcp_secret[SYNCOOKIE_NSECRETS];
1336 
1337 static int tcp_msstab[] = { 0, 536, 1460, 8960 };
1338 
1339 static MD5_CTX syn_ctx;
1340 
1341 #define MD5Add(v)	MD5Update(&syn_ctx, (u_char *)&v, sizeof(v))
1342 
1343 struct md5_add {
1344 	u_int32_t laddr, faddr;
1345 	u_int32_t secbits[4];
1346 	u_int16_t lport, fport;
1347 };
1348 
1349 #ifdef CTASSERT
1350 CTASSERT(sizeof(struct md5_add) == 28);
1351 #endif
1352 
1353 /*
1354  * Consider the problem of a recreated (and retransmitted) cookie.  If the
1355  * original SYN was accepted, the connection is established.  The second
1356  * SYN is inflight, and if it arrives with an ISN that falls within the
1357  * receive window, the connection is killed.
1358  *
1359  * However, since cookies have other problems, this may not be worth
1360  * worrying about.
1361  */
1362 
1363 static u_int32_t
1364 syncookie_generate(struct syncache *sc)
1365 {
1366 	u_int32_t md5_buffer[4];
1367 	u_int32_t data;
1368 	int idx, i;
1369 	struct md5_add add;
1370 #ifdef INET6
1371 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
1372 #else
1373 	const boolean_t isipv6 = FALSE;
1374 #endif
1375 
1376 	idx = ((ticks << SYNCOOKIE_TIMESHIFT) / hz) & SYNCOOKIE_WNDMASK;
1377 	if (tcp_secret[idx].ts_expire < ticks) {
1378 		for (i = 0; i < 4; i++)
1379 			tcp_secret[idx].ts_secbits[i] = karc4random();
1380 		tcp_secret[idx].ts_expire = ticks + SYNCOOKIE_TIMEOUT;
1381 	}
1382 	for (data = sizeof(tcp_msstab) / sizeof(int) - 1; data > 0; data--)
1383 		if (tcp_msstab[data] <= sc->sc_peer_mss)
1384 			break;
1385 	data = (data << SYNCOOKIE_WNDBITS) | idx;
1386 	data ^= sc->sc_irs;				/* peer's iss */
1387 	MD5Init(&syn_ctx);
1388 	if (isipv6) {
1389 		MD5Add(sc->sc_inc.inc6_laddr);
1390 		MD5Add(sc->sc_inc.inc6_faddr);
1391 		add.laddr = 0;
1392 		add.faddr = 0;
1393 	} else {
1394 		add.laddr = sc->sc_inc.inc_laddr.s_addr;
1395 		add.faddr = sc->sc_inc.inc_faddr.s_addr;
1396 	}
1397 	add.lport = sc->sc_inc.inc_lport;
1398 	add.fport = sc->sc_inc.inc_fport;
1399 	add.secbits[0] = tcp_secret[idx].ts_secbits[0];
1400 	add.secbits[1] = tcp_secret[idx].ts_secbits[1];
1401 	add.secbits[2] = tcp_secret[idx].ts_secbits[2];
1402 	add.secbits[3] = tcp_secret[idx].ts_secbits[3];
1403 	MD5Add(add);
1404 	MD5Final((u_char *)&md5_buffer, &syn_ctx);
1405 	data ^= (md5_buffer[0] & ~SYNCOOKIE_WNDMASK);
1406 	return (data);
1407 }
1408 
1409 static struct syncache *
1410 syncookie_lookup(struct in_conninfo *inc, struct tcphdr *th, struct socket *so)
1411 {
1412 	u_int32_t md5_buffer[4];
1413 	struct syncache *sc;
1414 	u_int32_t data;
1415 	int wnd, idx;
1416 	struct md5_add add;
1417 
1418 	data = (th->th_ack - 1) ^ (th->th_seq - 1);	/* remove ISS */
1419 	idx = data & SYNCOOKIE_WNDMASK;
1420 	if (tcp_secret[idx].ts_expire < ticks ||
1421 	    sototcpcb(so)->ts_recent + SYNCOOKIE_TIMEOUT < ticks)
1422 		return (NULL);
1423 	MD5Init(&syn_ctx);
1424 #ifdef INET6
1425 	if (inc->inc_isipv6) {
1426 		MD5Add(inc->inc6_laddr);
1427 		MD5Add(inc->inc6_faddr);
1428 		add.laddr = 0;
1429 		add.faddr = 0;
1430 	} else
1431 #endif
1432 	{
1433 		add.laddr = inc->inc_laddr.s_addr;
1434 		add.faddr = inc->inc_faddr.s_addr;
1435 	}
1436 	add.lport = inc->inc_lport;
1437 	add.fport = inc->inc_fport;
1438 	add.secbits[0] = tcp_secret[idx].ts_secbits[0];
1439 	add.secbits[1] = tcp_secret[idx].ts_secbits[1];
1440 	add.secbits[2] = tcp_secret[idx].ts_secbits[2];
1441 	add.secbits[3] = tcp_secret[idx].ts_secbits[3];
1442 	MD5Add(add);
1443 	MD5Final((u_char *)&md5_buffer, &syn_ctx);
1444 	data ^= md5_buffer[0];
1445 	if (data & ~SYNCOOKIE_DATAMASK)
1446 		return (NULL);
1447 	data = data >> SYNCOOKIE_WNDBITS;
1448 
1449 	/*
1450 	 * Fill in the syncache values.
1451 	 * XXX duplicate code from syncache_add
1452 	 */
1453 	sc = kmalloc(sizeof(struct syncache), M_SYNCACHE, M_WAITOK|M_ZERO);
1454 	sc->sc_ipopts = NULL;
1455 	sc->sc_inc.inc_fport = inc->inc_fport;
1456 	sc->sc_inc.inc_lport = inc->inc_lport;
1457 #ifdef INET6
1458 	sc->sc_inc.inc_isipv6 = inc->inc_isipv6;
1459 	if (inc->inc_isipv6) {
1460 		sc->sc_inc.inc6_faddr = inc->inc6_faddr;
1461 		sc->sc_inc.inc6_laddr = inc->inc6_laddr;
1462 		sc->sc_route6.ro_rt = NULL;
1463 	} else
1464 #endif
1465 	{
1466 		sc->sc_inc.inc_faddr = inc->inc_faddr;
1467 		sc->sc_inc.inc_laddr = inc->inc_laddr;
1468 		sc->sc_route.ro_rt = NULL;
1469 	}
1470 	sc->sc_irs = th->th_seq - 1;
1471 	sc->sc_iss = th->th_ack - 1;
1472 	wnd = ssb_space(&so->so_rcv);
1473 	wnd = imax(wnd, 0);
1474 	wnd = imin(wnd, TCP_MAXWIN);
1475 	sc->sc_wnd = wnd;
1476 	sc->sc_flags = 0;
1477 	sc->sc_rxtslot = 0;
1478 	sc->sc_peer_mss = tcp_msstab[data];
1479 	return (sc);
1480 }
1481