xref: /dflybsd-src/sys/netinet/tcp_syncache.c (revision 4041d91945fc1e4ba1bb1065fd5f9915983f8605)
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_inet6.h"
76 #include "opt_ipsec.h"
77 
78 #include <sys/param.h>
79 #include <sys/systm.h>
80 #include <sys/kernel.h>
81 #include <sys/sysctl.h>
82 #include <sys/malloc.h>
83 #include <sys/mbuf.h>
84 #include <sys/md5.h>
85 #include <sys/proc.h>		/* for proc0 declaration */
86 #include <sys/random.h>
87 #include <sys/socket.h>
88 #include <sys/socketvar.h>
89 #include <sys/in_cksum.h>
90 
91 #include <sys/msgport2.h>
92 #include <net/netmsg2.h>
93 
94 #include <net/if.h>
95 #include <net/route.h>
96 
97 #include <netinet/in.h>
98 #include <netinet/in_systm.h>
99 #include <netinet/ip.h>
100 #include <netinet/in_var.h>
101 #include <netinet/in_pcb.h>
102 #include <netinet/ip_var.h>
103 #include <netinet/ip6.h>
104 #ifdef INET6
105 #include <netinet/icmp6.h>
106 #include <netinet6/nd6.h>
107 #endif
108 #include <netinet6/ip6_var.h>
109 #include <netinet6/in6_pcb.h>
110 #include <netinet/tcp.h>
111 #include <netinet/tcp_fsm.h>
112 #include <netinet/tcp_seq.h>
113 #include <netinet/tcp_timer.h>
114 #include <netinet/tcp_timer2.h>
115 #include <netinet/tcp_var.h>
116 #include <netinet6/tcp6_var.h>
117 
118 #ifdef IPSEC
119 #include <netinet6/ipsec.h>
120 #ifdef INET6
121 #include <netinet6/ipsec6.h>
122 #endif
123 #include <netproto/key/key.h>
124 #endif /*IPSEC*/
125 
126 #ifdef FAST_IPSEC
127 #include <netproto/ipsec/ipsec.h>
128 #ifdef INET6
129 #include <netproto/ipsec/ipsec6.h>
130 #endif
131 #include <netproto/ipsec/key.h>
132 #define	IPSEC
133 #endif /*FAST_IPSEC*/
134 
135 static int tcp_syncookies = 1;
136 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_RW,
137     &tcp_syncookies, 0,
138     "Use TCP SYN cookies if the syncache overflows");
139 
140 static void	 syncache_drop(struct syncache *, struct syncache_head *);
141 static void	 syncache_free(struct syncache *);
142 static void	 syncache_insert(struct syncache *, struct syncache_head *);
143 struct syncache *syncache_lookup(struct in_conninfo *, struct syncache_head **);
144 static int	 syncache_respond(struct syncache *, struct mbuf *);
145 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
146 		    struct mbuf *);
147 static void	 syncache_timer(void *);
148 static u_int32_t syncookie_generate(struct syncache *);
149 static struct syncache *syncookie_lookup(struct in_conninfo *,
150 		    struct tcphdr *, struct socket *);
151 
152 /*
153  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
154  * 3 retransmits corresponds to a timeout of (1 + 2 + 4 + 8 == 15) seconds,
155  * the odds are that the user has given up attempting to connect by then.
156  */
157 #define SYNCACHE_MAXREXMTS		3
158 
159 /* Arbitrary values */
160 #define TCP_SYNCACHE_HASHSIZE		512
161 #define TCP_SYNCACHE_BUCKETLIMIT	30
162 
163 struct netmsg_sc_timer {
164 	struct netmsg nm_netmsg;
165 	struct msgrec *nm_mrec;		/* back pointer to containing msgrec */
166 };
167 
168 struct msgrec {
169 	struct netmsg_sc_timer msg;
170 	lwkt_port_t port;		/* constant after init */
171 	int slot;			/* constant after init */
172 };
173 
174 static void syncache_timer_handler(netmsg_t);
175 
176 struct tcp_syncache {
177 	u_int	hashsize;
178 	u_int	hashmask;
179 	u_int	bucket_limit;
180 	u_int	cache_limit;
181 	u_int	rexmt_limit;
182 	u_int	hash_secret;
183 };
184 static struct tcp_syncache tcp_syncache;
185 
186 TAILQ_HEAD(syncache_list, syncache);
187 
188 struct tcp_syncache_percpu {
189 	struct syncache_head	*hashbase;
190 	u_int			cache_count;
191 	struct syncache_list	timerq[SYNCACHE_MAXREXMTS + 1];
192 	struct callout		tt_timerq[SYNCACHE_MAXREXMTS + 1];
193 	struct msgrec		mrec[SYNCACHE_MAXREXMTS + 1];
194 };
195 static struct tcp_syncache_percpu tcp_syncache_percpu[MAXCPU];
196 
197 static struct lwkt_port syncache_null_rport;
198 
199 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, "TCP SYN cache");
200 
201 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RD,
202      &tcp_syncache.bucket_limit, 0, "Per-bucket hash limit for syncache");
203 
204 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_RD,
205      &tcp_syncache.cache_limit, 0, "Overall entry limit for syncache");
206 
207 /* XXX JH */
208 #if 0
209 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_RD,
210      &tcp_syncache.cache_count, 0, "Current number of entries in syncache");
211 #endif
212 
213 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RD,
214      &tcp_syncache.hashsize, 0, "Size of TCP syncache hashtable");
215 
216 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_RW,
217      &tcp_syncache.rexmt_limit, 0, "Limit on SYN/ACK retransmissions");
218 
219 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
220 
221 #define SYNCACHE_HASH(inc, mask)					\
222 	((tcp_syncache.hash_secret ^					\
223 	  (inc)->inc_faddr.s_addr ^					\
224 	  ((inc)->inc_faddr.s_addr >> 16) ^				\
225 	  (inc)->inc_fport ^ (inc)->inc_lport) & mask)
226 
227 #define SYNCACHE_HASH6(inc, mask)					\
228 	((tcp_syncache.hash_secret ^					\
229 	  (inc)->inc6_faddr.s6_addr32[0] ^				\
230 	  (inc)->inc6_faddr.s6_addr32[3] ^				\
231 	  (inc)->inc_fport ^ (inc)->inc_lport) & mask)
232 
233 #define ENDPTS_EQ(a, b) (						\
234 	(a)->ie_fport == (b)->ie_fport &&				\
235 	(a)->ie_lport == (b)->ie_lport &&				\
236 	(a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr &&			\
237 	(a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr			\
238 )
239 
240 #define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0)
241 
242 static __inline void
243 syncache_timeout(struct tcp_syncache_percpu *syncache_percpu,
244 		 struct syncache *sc, int slot)
245 {
246 	sc->sc_rxtslot = slot;
247 	sc->sc_rxttime = ticks + TCPTV_RTOBASE * tcp_backoff[slot];
248 	TAILQ_INSERT_TAIL(&syncache_percpu->timerq[slot], sc, sc_timerq);
249 	if (!callout_active(&syncache_percpu->tt_timerq[slot])) {
250 		callout_reset(&syncache_percpu->tt_timerq[slot],
251 			      TCPTV_RTOBASE * tcp_backoff[slot],
252 			      syncache_timer,
253 			      &syncache_percpu->mrec[slot]);
254 	}
255 }
256 
257 static void
258 syncache_free(struct syncache *sc)
259 {
260 	struct rtentry *rt;
261 #ifdef INET6
262 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
263 #else
264 	const boolean_t isipv6 = FALSE;
265 #endif
266 
267 	if (sc->sc_ipopts)
268 		m_free(sc->sc_ipopts);
269 
270 	rt = isipv6 ? sc->sc_route6.ro_rt : sc->sc_route.ro_rt;
271 	if (rt != NULL) {
272 		/*
273 		 * If this is the only reference to a protocol-cloned
274 		 * route, remove it immediately.
275 		 */
276 		if ((rt->rt_flags & RTF_WASCLONED) && rt->rt_refcnt == 1)
277 			rtrequest(RTM_DELETE, rt_key(rt), rt->rt_gateway,
278 				  rt_mask(rt), rt->rt_flags, NULL);
279 		RTFREE(rt);
280 	}
281 	kfree(sc, M_SYNCACHE);
282 }
283 
284 void
285 syncache_init(void)
286 {
287 	int i, cpu;
288 
289 	tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
290 	tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
291 	tcp_syncache.cache_limit =
292 	    tcp_syncache.hashsize * tcp_syncache.bucket_limit;
293 	tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
294 	tcp_syncache.hash_secret = karc4random();
295 
296 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
297 	    &tcp_syncache.hashsize);
298 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
299 	    &tcp_syncache.cache_limit);
300 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
301 	    &tcp_syncache.bucket_limit);
302 	if (!powerof2(tcp_syncache.hashsize)) {
303 		kprintf("WARNING: syncache hash size is not a power of 2.\n");
304 		tcp_syncache.hashsize = 512;	/* safe default */
305 	}
306 	tcp_syncache.hashmask = tcp_syncache.hashsize - 1;
307 
308 	lwkt_initport_replyonly_null(&syncache_null_rport);
309 
310 	for (cpu = 0; cpu < ncpus2; cpu++) {
311 		struct tcp_syncache_percpu *syncache_percpu;
312 
313 		syncache_percpu = &tcp_syncache_percpu[cpu];
314 		/* Allocate the hash table. */
315 		MALLOC(syncache_percpu->hashbase, struct syncache_head *,
316 		    tcp_syncache.hashsize * sizeof(struct syncache_head),
317 		    M_SYNCACHE, M_WAITOK);
318 
319 		/* Initialize the hash buckets. */
320 		for (i = 0; i < tcp_syncache.hashsize; i++) {
321 			struct syncache_head *bucket;
322 
323 			bucket = &syncache_percpu->hashbase[i];
324 			TAILQ_INIT(&bucket->sch_bucket);
325 			bucket->sch_length = 0;
326 		}
327 
328 		for (i = 0; i <= SYNCACHE_MAXREXMTS; i++) {
329 			/* Initialize the timer queues. */
330 			TAILQ_INIT(&syncache_percpu->timerq[i]);
331 			callout_init(&syncache_percpu->tt_timerq[i]);
332 
333 			syncache_percpu->mrec[i].slot = i;
334 			syncache_percpu->mrec[i].port = tcp_cport(cpu);
335 			syncache_percpu->mrec[i].msg.nm_mrec =
336 			    &syncache_percpu->mrec[i];
337 			netmsg_init(&syncache_percpu->mrec[i].msg.nm_netmsg,
338 				    NULL, &syncache_null_rport,
339 				    0, syncache_timer_handler);
340 		}
341 	}
342 }
343 
344 static void
345 syncache_insert(struct syncache *sc, struct syncache_head *sch)
346 {
347 	struct tcp_syncache_percpu *syncache_percpu;
348 	struct syncache *sc2;
349 	int i;
350 
351 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
352 
353 	/*
354 	 * Make sure that we don't overflow the per-bucket
355 	 * limit or the total cache size limit.
356 	 */
357 	if (sch->sch_length >= tcp_syncache.bucket_limit) {
358 		/*
359 		 * The bucket is full, toss the oldest element.
360 		 */
361 		sc2 = TAILQ_FIRST(&sch->sch_bucket);
362 		sc2->sc_tp->ts_recent = ticks;
363 		syncache_drop(sc2, sch);
364 		tcpstat.tcps_sc_bucketoverflow++;
365 	} else if (syncache_percpu->cache_count >= tcp_syncache.cache_limit) {
366 		/*
367 		 * The cache is full.  Toss the oldest entry in the
368 		 * entire cache.  This is the front entry in the
369 		 * first non-empty timer queue with the largest
370 		 * timeout value.
371 		 */
372 		for (i = SYNCACHE_MAXREXMTS; i >= 0; i--) {
373 			sc2 = TAILQ_FIRST(&syncache_percpu->timerq[i]);
374 			while (sc2 && (sc2->sc_flags & SCF_MARKER))
375 				sc2 = TAILQ_NEXT(sc2, sc_timerq);
376 			if (sc2 != NULL)
377 				break;
378 		}
379 		sc2->sc_tp->ts_recent = ticks;
380 		syncache_drop(sc2, NULL);
381 		tcpstat.tcps_sc_cacheoverflow++;
382 	}
383 
384 	/* Initialize the entry's timer. */
385 	syncache_timeout(syncache_percpu, sc, 0);
386 
387 	/* Put it into the bucket. */
388 	TAILQ_INSERT_TAIL(&sch->sch_bucket, sc, sc_hash);
389 	sch->sch_length++;
390 	syncache_percpu->cache_count++;
391 	tcpstat.tcps_sc_added++;
392 }
393 
394 void
395 syncache_destroy(struct tcpcb *tp)
396 {
397 	struct tcp_syncache_percpu *syncache_percpu;
398 	struct syncache_head *bucket;
399 	struct syncache *sc;
400 	int i;
401 
402 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
403 	sc = NULL;
404 
405 	for (i = 0; i < tcp_syncache.hashsize; i++) {
406 		bucket = &syncache_percpu->hashbase[i];
407 		TAILQ_FOREACH(sc, &bucket->sch_bucket, sc_hash) {
408 			if (sc->sc_tp == tp) {
409 				sc->sc_tp = NULL;
410 				tp->t_flags &= ~TF_SYNCACHE;
411 				break;
412 			}
413 		}
414 	}
415 	kprintf("Warning: delete stale syncache for tp=%p, sc=%p\n", tp, sc);
416 }
417 
418 static void
419 syncache_drop(struct syncache *sc, struct syncache_head *sch)
420 {
421 	struct tcp_syncache_percpu *syncache_percpu;
422 #ifdef INET6
423 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
424 #else
425 	const boolean_t isipv6 = FALSE;
426 #endif
427 
428 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
429 
430 	if (sch == NULL) {
431 		if (isipv6) {
432 			sch = &syncache_percpu->hashbase[
433 			    SYNCACHE_HASH6(&sc->sc_inc, tcp_syncache.hashmask)];
434 		} else {
435 			sch = &syncache_percpu->hashbase[
436 			    SYNCACHE_HASH(&sc->sc_inc, tcp_syncache.hashmask)];
437 		}
438 	}
439 
440 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
441 	sch->sch_length--;
442 	syncache_percpu->cache_count--;
443 
444 	/*
445 	 * Cleanup
446 	 */
447 	if (sc->sc_tp) {
448 		sc->sc_tp->t_flags &= ~TF_SYNCACHE;
449 		sc->sc_tp = NULL;
450 	}
451 
452 	/*
453 	 * Remove the entry from the syncache timer/timeout queue.  Note
454 	 * that we do not try to stop any running timer since we do not know
455 	 * whether the timer's message is in-transit or not.  Since timeouts
456 	 * are fairly long, taking an unneeded callout does not detrimentally
457 	 * effect performance.
458 	 */
459 	TAILQ_REMOVE(&syncache_percpu->timerq[sc->sc_rxtslot], sc, sc_timerq);
460 
461 	syncache_free(sc);
462 }
463 
464 /*
465  * Place a timeout message on the TCP thread's message queue.
466  * This routine runs in soft interrupt context.
467  *
468  * An invariant is for this routine to be called, the callout must
469  * have been active.  Note that the callout is not deactivated until
470  * after the message has been processed in syncache_timer_handler() below.
471  */
472 static void
473 syncache_timer(void *p)
474 {
475 	struct netmsg_sc_timer *msg = p;
476 
477 	lwkt_sendmsg(msg->nm_mrec->port, &msg->nm_netmsg.nm_lmsg);
478 }
479 
480 /*
481  * Service a timer message queued by timer expiration.
482  * This routine runs in the TCP protocol thread.
483  *
484  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
485  * If we have retransmitted an entry the maximum number of times, expire it.
486  *
487  * When we finish processing timed-out entries, we restart the timer if there
488  * are any entries still on the queue and deactivate it otherwise.  Only after
489  * a timer has been deactivated here can it be restarted by syncache_timeout().
490  */
491 static void
492 syncache_timer_handler(netmsg_t netmsg)
493 {
494 	struct tcp_syncache_percpu *syncache_percpu;
495 	struct syncache *sc;
496 	struct syncache marker;
497 	struct syncache_list *list;
498 	struct inpcb *inp;
499 	int slot;
500 
501 	slot = ((struct netmsg_sc_timer *)netmsg)->nm_mrec->slot;
502 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
503 
504 	list = &syncache_percpu->timerq[slot];
505 
506 	/*
507 	 * Use a marker to keep our place in the scan.  syncache_drop()
508 	 * can block and cause any next pointer we cache to become stale.
509 	 */
510 	marker.sc_flags = SCF_MARKER;
511 	TAILQ_INSERT_HEAD(list, &marker, sc_timerq);
512 
513 	while ((sc = TAILQ_NEXT(&marker, sc_timerq)) != NULL) {
514 		/*
515 		 * Move the marker.
516 		 */
517 		TAILQ_REMOVE(list, &marker, sc_timerq);
518 		TAILQ_INSERT_AFTER(list, sc, &marker, sc_timerq);
519 
520 		if (sc->sc_flags & SCF_MARKER)
521 			continue;
522 
523 		if (ticks < sc->sc_rxttime)
524 			break;	/* finished because timerq sorted by time */
525 		if (sc->sc_tp == NULL) {
526 			syncache_drop(sc, NULL);
527 			tcpstat.tcps_sc_stale++;
528 			continue;
529 		}
530 		inp = sc->sc_tp->t_inpcb;
531 		if (slot == SYNCACHE_MAXREXMTS ||
532 		    slot >= tcp_syncache.rexmt_limit ||
533 		    inp == NULL ||
534 		    inp->inp_gencnt != sc->sc_inp_gencnt) {
535 			syncache_drop(sc, NULL);
536 			tcpstat.tcps_sc_stale++;
537 			continue;
538 		}
539 		/*
540 		 * syncache_respond() may call back into the syncache to
541 		 * to modify another entry, so do not obtain the next
542 		 * entry on the timer chain until it has completed.
543 		 */
544 		syncache_respond(sc, NULL);
545 		tcpstat.tcps_sc_retransmitted++;
546 		TAILQ_REMOVE(list, sc, sc_timerq);
547 		syncache_timeout(syncache_percpu, sc, slot + 1);
548 	}
549 	TAILQ_REMOVE(list, &marker, sc_timerq);
550 
551 	if (sc != NULL) {
552 		callout_reset(&syncache_percpu->tt_timerq[slot],
553 			      sc->sc_rxttime - ticks, syncache_timer,
554 			      &syncache_percpu->mrec[slot]);
555 	} else {
556 		callout_deactivate(&syncache_percpu->tt_timerq[slot]);
557 	}
558 	lwkt_replymsg(&netmsg->nm_lmsg, 0);
559 }
560 
561 /*
562  * Find an entry in the syncache.
563  */
564 struct syncache *
565 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
566 {
567 	struct tcp_syncache_percpu *syncache_percpu;
568 	struct syncache *sc;
569 	struct syncache_head *sch;
570 
571 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
572 #ifdef INET6
573 	if (inc->inc_isipv6) {
574 		sch = &syncache_percpu->hashbase[
575 		    SYNCACHE_HASH6(inc, tcp_syncache.hashmask)];
576 		*schp = sch;
577 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
578 			if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
579 				return (sc);
580 	} else
581 #endif
582 	{
583 		sch = &syncache_percpu->hashbase[
584 		    SYNCACHE_HASH(inc, tcp_syncache.hashmask)];
585 		*schp = sch;
586 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
587 #ifdef INET6
588 			if (sc->sc_inc.inc_isipv6)
589 				continue;
590 #endif
591 			if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
592 				return (sc);
593 		}
594 	}
595 	return (NULL);
596 }
597 
598 /*
599  * This function is called when we get a RST for a
600  * non-existent connection, so that we can see if the
601  * connection is in the syn cache.  If it is, zap it.
602  */
603 void
604 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
605 {
606 	struct syncache *sc;
607 	struct syncache_head *sch;
608 
609 	sc = syncache_lookup(inc, &sch);
610 	if (sc == NULL) {
611 		return;
612 	}
613 	/*
614 	 * If the RST bit is set, check the sequence number to see
615 	 * if this is a valid reset segment.
616 	 * RFC 793 page 37:
617 	 *   In all states except SYN-SENT, all reset (RST) segments
618 	 *   are validated by checking their SEQ-fields.  A reset is
619 	 *   valid if its sequence number is in the window.
620 	 *
621 	 *   The sequence number in the reset segment is normally an
622 	 *   echo of our outgoing acknowlegement numbers, but some hosts
623 	 *   send a reset with the sequence number at the rightmost edge
624 	 *   of our receive window, and we have to handle this case.
625 	 */
626 	if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
627 	    SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
628 		syncache_drop(sc, sch);
629 		tcpstat.tcps_sc_reset++;
630 	}
631 }
632 
633 void
634 syncache_badack(struct in_conninfo *inc)
635 {
636 	struct syncache *sc;
637 	struct syncache_head *sch;
638 
639 	sc = syncache_lookup(inc, &sch);
640 	if (sc != NULL) {
641 		syncache_drop(sc, sch);
642 		tcpstat.tcps_sc_badack++;
643 	}
644 }
645 
646 void
647 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th)
648 {
649 	struct syncache *sc;
650 	struct syncache_head *sch;
651 
652 	/* we are called at splnet() here */
653 	sc = syncache_lookup(inc, &sch);
654 	if (sc == NULL)
655 		return;
656 
657 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
658 	if (ntohl(th->th_seq) != sc->sc_iss)
659 		return;
660 
661 	/*
662 	 * If we've rertransmitted 3 times and this is our second error,
663 	 * we remove the entry.  Otherwise, we allow it to continue on.
664 	 * This prevents us from incorrectly nuking an entry during a
665 	 * spurious network outage.
666 	 *
667 	 * See tcp_notify().
668 	 */
669 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxtslot < 3) {
670 		sc->sc_flags |= SCF_UNREACH;
671 		return;
672 	}
673 	syncache_drop(sc, sch);
674 	tcpstat.tcps_sc_unreach++;
675 }
676 
677 /*
678  * Build a new TCP socket structure from a syncache entry.
679  *
680  * This is called from the context of the SYN+ACK
681  */
682 static struct socket *
683 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
684 {
685 	struct inpcb *inp = NULL, *linp;
686 	struct socket *so;
687 	struct tcpcb *tp;
688 	lwkt_port_t port;
689 #ifdef INET6
690 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
691 #else
692 	const boolean_t isipv6 = FALSE;
693 #endif
694 
695 	/*
696 	 * Ok, create the full blown connection, and set things up
697 	 * as they would have been set up if we had created the
698 	 * connection when the SYN arrived.  If we can't create
699 	 * the connection, abort it.
700 	 */
701 	so = sonewconn(lso, SS_ISCONNECTED);
702 	if (so == NULL) {
703 		/*
704 		 * Drop the connection; we will send a RST if the peer
705 		 * retransmits the ACK,
706 		 */
707 		tcpstat.tcps_listendrop++;
708 		goto abort;
709 	}
710 
711 	/*
712 	 * Set the protocol processing port for the socket to the current
713 	 * port (that the connection came in on).
714 	 */
715 	sosetport(so, &curthread->td_msgport);
716 
717 	/*
718 	 * Insert new socket into hash list.
719 	 */
720 	inp = so->so_pcb;
721 	inp->inp_inc.inc_isipv6 = sc->sc_inc.inc_isipv6;
722 	if (isipv6) {
723 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
724 	} else {
725 #ifdef INET6
726 		inp->inp_vflag &= ~INP_IPV6;
727 		inp->inp_vflag |= INP_IPV4;
728 		inp->inp_flags &= ~IN6P_IPV6_V6ONLY;
729 #endif
730 		inp->inp_laddr = sc->sc_inc.inc_laddr;
731 	}
732 	inp->inp_lport = sc->sc_inc.inc_lport;
733 	if (in_pcbinsporthash(inp) != 0) {
734 		/*
735 		 * Undo the assignments above if we failed to
736 		 * put the PCB on the hash lists.
737 		 */
738 		if (isipv6)
739 			inp->in6p_laddr = kin6addr_any;
740 		else
741 			inp->inp_laddr.s_addr = INADDR_ANY;
742 		inp->inp_lport = 0;
743 		goto abort;
744 	}
745 	linp = so->so_pcb;
746 #ifdef IPSEC
747 	/* copy old policy into new socket's */
748 	if (ipsec_copy_policy(linp->inp_sp, inp->inp_sp))
749 		kprintf("syncache_expand: could not copy policy\n");
750 #endif
751 	if (isipv6) {
752 		struct in6_addr laddr6;
753 		struct sockaddr_in6 sin6;
754 		/*
755 		 * Inherit socket options from the listening socket.
756 		 * Note that in6p_inputopts are not (and should not be)
757 		 * copied, since it stores previously received options and is
758 		 * used to detect if each new option is different than the
759 		 * previous one and hence should be passed to a user.
760 		 * If we copied in6p_inputopts, a user would not be able to
761 		 * receive options just after calling the accept system call.
762 		 */
763 		inp->inp_flags |= linp->inp_flags & INP_CONTROLOPTS;
764 		if (linp->in6p_outputopts)
765 			inp->in6p_outputopts =
766 			    ip6_copypktopts(linp->in6p_outputopts, M_INTWAIT);
767 		inp->in6p_route = sc->sc_route6;
768 		sc->sc_route6.ro_rt = NULL;
769 
770 		sin6.sin6_family = AF_INET6;
771 		sin6.sin6_len = sizeof sin6;
772 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
773 		sin6.sin6_port = sc->sc_inc.inc_fport;
774 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
775 		laddr6 = inp->in6p_laddr;
776 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
777 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
778 		if (in6_pcbconnect(inp, (struct sockaddr *)&sin6, &thread0)) {
779 			inp->in6p_laddr = laddr6;
780 			goto abort;
781 		}
782 	} else {
783 		struct in_addr laddr;
784 		struct sockaddr_in sin;
785 
786 		inp->inp_options = ip_srcroute(m);
787 		if (inp->inp_options == NULL) {
788 			inp->inp_options = sc->sc_ipopts;
789 			sc->sc_ipopts = NULL;
790 		}
791 		inp->inp_route = sc->sc_route;
792 		sc->sc_route.ro_rt = NULL;
793 
794 		sin.sin_family = AF_INET;
795 		sin.sin_len = sizeof sin;
796 		sin.sin_addr = sc->sc_inc.inc_faddr;
797 		sin.sin_port = sc->sc_inc.inc_fport;
798 		bzero(sin.sin_zero, sizeof sin.sin_zero);
799 		laddr = inp->inp_laddr;
800 		if (inp->inp_laddr.s_addr == INADDR_ANY)
801 			inp->inp_laddr = sc->sc_inc.inc_laddr;
802 		if (in_pcbconnect(inp, (struct sockaddr *)&sin, &thread0)) {
803 			inp->inp_laddr = laddr;
804 			goto abort;
805 		}
806 	}
807 
808 	/*
809 	 * The current port should be in the context of the SYN+ACK and
810 	 * so should match the tcp address port.
811 	 *
812 	 * XXX we may be running on the netisr thread instead of a tcp
813 	 *     thread, in which case port will not match
814 	 *     curthread->td_msgport.
815 	 */
816 	if (isipv6) {
817 		port = tcp6_addrport();
818 	} else {
819 		port = tcp_addrport(inp->inp_faddr.s_addr, inp->inp_fport,
820 				    inp->inp_laddr.s_addr, inp->inp_lport);
821 	}
822 	/*KKASSERT(port == &curthread->td_msgport);*/
823 
824 	tp = intotcpcb(inp);
825 	tp->t_state = TCPS_SYN_RECEIVED;
826 	tp->iss = sc->sc_iss;
827 	tp->irs = sc->sc_irs;
828 	tcp_rcvseqinit(tp);
829 	tcp_sendseqinit(tp);
830 	tp->snd_wl1 = sc->sc_irs;
831 	tp->rcv_up = sc->sc_irs + 1;
832 	tp->rcv_wnd = sc->sc_wnd;
833 	tp->rcv_adv += tp->rcv_wnd;
834 
835 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH | TF_NODELAY);
836 	if (sc->sc_flags & SCF_NOOPT)
837 		tp->t_flags |= TF_NOOPT;
838 	if (sc->sc_flags & SCF_WINSCALE) {
839 		tp->t_flags |= TF_REQ_SCALE | TF_RCVD_SCALE;
840 		tp->requested_s_scale = sc->sc_requested_s_scale;
841 		tp->request_r_scale = sc->sc_request_r_scale;
842 	}
843 	if (sc->sc_flags & SCF_TIMESTAMP) {
844 		tp->t_flags |= TF_REQ_TSTMP | TF_RCVD_TSTMP;
845 		tp->ts_recent = sc->sc_tsrecent;
846 		tp->ts_recent_age = ticks;
847 	}
848 	if (sc->sc_flags & SCF_SACK_PERMITTED)
849 		tp->t_flags |= TF_SACK_PERMITTED;
850 
851 	tcp_mss(tp, sc->sc_peer_mss);
852 
853 	/*
854 	 * If the SYN,ACK was retransmitted, reset cwnd to 1 segment.
855 	 */
856 	if (sc->sc_rxtslot != 0)
857 		tp->snd_cwnd = tp->t_maxseg;
858 	tcp_create_timermsg(tp, port);
859 	tcp_callout_reset(tp, tp->tt_keep, tcp_keepinit, tcp_timer_keep);
860 
861 	tcpstat.tcps_accepts++;
862 	return (so);
863 
864 abort:
865 	if (so != NULL)
866 		soabort_oncpu(so);
867 	return (NULL);
868 }
869 
870 /*
871  * This function gets called when we receive an ACK for a
872  * socket in the LISTEN state.  We look up the connection
873  * in the syncache, and if its there, we pull it out of
874  * the cache and turn it into a full-blown connection in
875  * the SYN-RECEIVED state.
876  */
877 int
878 syncache_expand(struct in_conninfo *inc, struct tcphdr *th, struct socket **sop,
879 		struct mbuf *m)
880 {
881 	struct syncache *sc;
882 	struct syncache_head *sch;
883 	struct socket *so;
884 
885 	sc = syncache_lookup(inc, &sch);
886 	if (sc == NULL) {
887 		/*
888 		 * There is no syncache entry, so see if this ACK is
889 		 * a returning syncookie.  To do this, first:
890 		 *  A. See if this socket has had a syncache entry dropped in
891 		 *     the past.  We don't want to accept a bogus syncookie
892 		 *     if we've never received a SYN.
893 		 *  B. check that the syncookie is valid.  If it is, then
894 		 *     cobble up a fake syncache entry, and return.
895 		 */
896 		if (!tcp_syncookies)
897 			return (0);
898 		sc = syncookie_lookup(inc, th, *sop);
899 		if (sc == NULL)
900 			return (0);
901 		sch = NULL;
902 		tcpstat.tcps_sc_recvcookie++;
903 	}
904 
905 	/*
906 	 * If seg contains an ACK, but not for our SYN/ACK, send a RST.
907 	 */
908 	if (th->th_ack != sc->sc_iss + 1)
909 		return (0);
910 
911 	so = syncache_socket(sc, *sop, m);
912 	if (so == NULL) {
913 #if 0
914 resetandabort:
915 		/* XXXjlemon check this - is this correct? */
916 		tcp_respond(NULL, m, m, th,
917 		    th->th_seq + tlen, (tcp_seq)0, TH_RST | TH_ACK);
918 #endif
919 		m_freem(m);			/* XXX only needed for above */
920 		tcpstat.tcps_sc_aborted++;
921 	} else {
922 		tcpstat.tcps_sc_completed++;
923 	}
924 	if (sch == NULL)
925 		syncache_free(sc);
926 	else
927 		syncache_drop(sc, sch);
928 	*sop = so;
929 	return (1);
930 }
931 
932 /*
933  * Given a LISTEN socket and an inbound SYN request, add
934  * this to the syn cache, and send back a segment:
935  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
936  * to the source.
937  *
938  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
939  * Doing so would require that we hold onto the data and deliver it
940  * to the application.  However, if we are the target of a SYN-flood
941  * DoS attack, an attacker could send data which would eventually
942  * consume all available buffer space if it were ACKed.  By not ACKing
943  * the data, we avoid this DoS scenario.
944  */
945 int
946 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
947 	     struct socket **sop, struct mbuf *m)
948 {
949 	struct tcp_syncache_percpu *syncache_percpu;
950 	struct tcpcb *tp;
951 	struct socket *so;
952 	struct syncache *sc = NULL;
953 	struct syncache_head *sch;
954 	struct mbuf *ipopts = NULL;
955 	int win;
956 
957 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
958 	so = *sop;
959 	tp = sototcpcb(so);
960 
961 	/*
962 	 * Remember the IP options, if any.
963 	 */
964 #ifdef INET6
965 	if (!inc->inc_isipv6)
966 #endif
967 		ipopts = ip_srcroute(m);
968 
969 	/*
970 	 * See if we already have an entry for this connection.
971 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
972 	 *
973 	 * XXX
974 	 * The syncache should be re-initialized with the contents
975 	 * of the new SYN which may have different options.
976 	 */
977 	sc = syncache_lookup(inc, &sch);
978 	if (sc != NULL) {
979 		tcpstat.tcps_sc_dupsyn++;
980 		if (ipopts) {
981 			/*
982 			 * If we were remembering a previous source route,
983 			 * forget it and use the new one we've been given.
984 			 */
985 			if (sc->sc_ipopts)
986 				m_free(sc->sc_ipopts);
987 			sc->sc_ipopts = ipopts;
988 		}
989 		/*
990 		 * Update timestamp if present.
991 		 */
992 		if (sc->sc_flags & SCF_TIMESTAMP)
993 			sc->sc_tsrecent = to->to_tsval;
994 
995 		/* Just update the TOF_SACK_PERMITTED for now. */
996 		if (tcp_do_sack && (to->to_flags & TOF_SACK_PERMITTED))
997 			sc->sc_flags |= SCF_SACK_PERMITTED;
998 		else
999 			sc->sc_flags &= ~SCF_SACK_PERMITTED;
1000 
1001 		/*
1002 		 * PCB may have changed, pick up new values.
1003 		 */
1004 		if (sc->sc_tp) {
1005 			sc->sc_tp->t_flags &= ~TF_SYNCACHE;
1006 			tp->t_flags |= TF_SYNCACHE;
1007 		}
1008 		sc->sc_tp = tp;
1009 		sc->sc_inp_gencnt = tp->t_inpcb->inp_gencnt;
1010 		if (syncache_respond(sc, m) == 0) {
1011 			TAILQ_REMOVE(&syncache_percpu->timerq[sc->sc_rxtslot],
1012 				     sc, sc_timerq);
1013 			syncache_timeout(syncache_percpu, sc, sc->sc_rxtslot);
1014 			tcpstat.tcps_sndacks++;
1015 			tcpstat.tcps_sndtotal++;
1016 		}
1017 		*sop = NULL;
1018 		return (1);
1019 	}
1020 
1021 	/*
1022 	 * Fill in the syncache values.
1023 	 */
1024 	sc = kmalloc(sizeof(struct syncache), M_SYNCACHE, M_WAITOK|M_ZERO);
1025 	sc->sc_inp_gencnt = tp->t_inpcb->inp_gencnt;
1026 	sc->sc_ipopts = ipopts;
1027 	sc->sc_inc.inc_fport = inc->inc_fport;
1028 	sc->sc_inc.inc_lport = inc->inc_lport;
1029 	sc->sc_tp = tp;
1030 	tp->t_flags |= TF_SYNCACHE;
1031 #ifdef INET6
1032 	sc->sc_inc.inc_isipv6 = inc->inc_isipv6;
1033 	if (inc->inc_isipv6) {
1034 		sc->sc_inc.inc6_faddr = inc->inc6_faddr;
1035 		sc->sc_inc.inc6_laddr = inc->inc6_laddr;
1036 		sc->sc_route6.ro_rt = NULL;
1037 	} else
1038 #endif
1039 	{
1040 		sc->sc_inc.inc_faddr = inc->inc_faddr;
1041 		sc->sc_inc.inc_laddr = inc->inc_laddr;
1042 		sc->sc_route.ro_rt = NULL;
1043 	}
1044 	sc->sc_irs = th->th_seq;
1045 	sc->sc_flags = 0;
1046 	sc->sc_peer_mss = to->to_flags & TOF_MSS ? to->to_mss : 0;
1047 	if (tcp_syncookies)
1048 		sc->sc_iss = syncookie_generate(sc);
1049 	else
1050 		sc->sc_iss = karc4random();
1051 
1052 	/* Initial receive window: clip ssb_space to [0 .. TCP_MAXWIN] */
1053 	win = ssb_space(&so->so_rcv);
1054 	win = imax(win, 0);
1055 	win = imin(win, TCP_MAXWIN);
1056 	sc->sc_wnd = win;
1057 
1058 	if (tcp_do_rfc1323) {
1059 		/*
1060 		 * A timestamp received in a SYN makes
1061 		 * it ok to send timestamp requests and replies.
1062 		 */
1063 		if (to->to_flags & TOF_TS) {
1064 			sc->sc_tsrecent = to->to_tsval;
1065 			sc->sc_flags |= SCF_TIMESTAMP;
1066 		}
1067 		if (to->to_flags & TOF_SCALE) {
1068 			int wscale = TCP_MIN_WINSHIFT;
1069 
1070 			/* Compute proper scaling value from buffer space */
1071 			while (wscale < TCP_MAX_WINSHIFT &&
1072 			    (TCP_MAXWIN << wscale) < so->so_rcv.ssb_hiwat) {
1073 				wscale++;
1074 			}
1075 			sc->sc_request_r_scale = wscale;
1076 			sc->sc_requested_s_scale = to->to_requested_s_scale;
1077 			sc->sc_flags |= SCF_WINSCALE;
1078 		}
1079 	}
1080 	if (tcp_do_sack && (to->to_flags & TOF_SACK_PERMITTED))
1081 		sc->sc_flags |= SCF_SACK_PERMITTED;
1082 	if (tp->t_flags & TF_NOOPT)
1083 		sc->sc_flags = SCF_NOOPT;
1084 
1085 	if (syncache_respond(sc, m) == 0) {
1086 		syncache_insert(sc, sch);
1087 		tcpstat.tcps_sndacks++;
1088 		tcpstat.tcps_sndtotal++;
1089 	} else {
1090 		syncache_free(sc);
1091 		tcpstat.tcps_sc_dropped++;
1092 	}
1093 	*sop = NULL;
1094 	return (1);
1095 }
1096 
1097 static int
1098 syncache_respond(struct syncache *sc, struct mbuf *m)
1099 {
1100 	u_int8_t *optp;
1101 	int optlen, error;
1102 	u_int16_t tlen, hlen, mssopt;
1103 	struct ip *ip = NULL;
1104 	struct rtentry *rt;
1105 	struct tcphdr *th;
1106 	struct ip6_hdr *ip6 = NULL;
1107 #ifdef INET6
1108 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
1109 #else
1110 	const boolean_t isipv6 = FALSE;
1111 #endif
1112 
1113 	if (isipv6) {
1114 		rt = tcp_rtlookup6(&sc->sc_inc);
1115 		if (rt != NULL)
1116 			mssopt = rt->rt_ifp->if_mtu -
1117 			     (sizeof(struct ip6_hdr) + sizeof(struct tcphdr));
1118 		else
1119 			mssopt = tcp_v6mssdflt;
1120 		hlen = sizeof(struct ip6_hdr);
1121 	} else {
1122 		rt = tcp_rtlookup(&sc->sc_inc);
1123 		if (rt != NULL)
1124 			mssopt = rt->rt_ifp->if_mtu -
1125 			     (sizeof(struct ip) + sizeof(struct tcphdr));
1126 		else
1127 			mssopt = tcp_mssdflt;
1128 		hlen = sizeof(struct ip);
1129 	}
1130 
1131 	/* Compute the size of the TCP options. */
1132 	if (sc->sc_flags & SCF_NOOPT) {
1133 		optlen = 0;
1134 	} else {
1135 		optlen = TCPOLEN_MAXSEG +
1136 		    ((sc->sc_flags & SCF_WINSCALE) ? 4 : 0) +
1137 		    ((sc->sc_flags & SCF_TIMESTAMP) ? TCPOLEN_TSTAMP_APPA : 0) +
1138 		    ((sc->sc_flags & SCF_SACK_PERMITTED) ?
1139 			TCPOLEN_SACK_PERMITTED_ALIGNED : 0);
1140 	}
1141 	tlen = hlen + sizeof(struct tcphdr) + optlen;
1142 
1143 	/*
1144 	 * XXX
1145 	 * assume that the entire packet will fit in a header mbuf
1146 	 */
1147 	KASSERT(max_linkhdr + tlen <= MHLEN, ("syncache: mbuf too small"));
1148 
1149 	/*
1150 	 * XXX shouldn't this reuse the mbuf if possible ?
1151 	 * Create the IP+TCP header from scratch.
1152 	 */
1153 	if (m)
1154 		m_freem(m);
1155 
1156 	m = m_gethdr(MB_DONTWAIT, MT_HEADER);
1157 	if (m == NULL)
1158 		return (ENOBUFS);
1159 	m->m_data += max_linkhdr;
1160 	m->m_len = tlen;
1161 	m->m_pkthdr.len = tlen;
1162 	m->m_pkthdr.rcvif = NULL;
1163 
1164 	if (isipv6) {
1165 		ip6 = mtod(m, struct ip6_hdr *);
1166 		ip6->ip6_vfc = IPV6_VERSION;
1167 		ip6->ip6_nxt = IPPROTO_TCP;
1168 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1169 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1170 		ip6->ip6_plen = htons(tlen - hlen);
1171 		/* ip6_hlim is set after checksum */
1172 		/* ip6_flow = ??? */
1173 
1174 		th = (struct tcphdr *)(ip6 + 1);
1175 	} else {
1176 		ip = mtod(m, struct ip *);
1177 		ip->ip_v = IPVERSION;
1178 		ip->ip_hl = sizeof(struct ip) >> 2;
1179 		ip->ip_len = tlen;
1180 		ip->ip_id = 0;
1181 		ip->ip_off = 0;
1182 		ip->ip_sum = 0;
1183 		ip->ip_p = IPPROTO_TCP;
1184 		ip->ip_src = sc->sc_inc.inc_laddr;
1185 		ip->ip_dst = sc->sc_inc.inc_faddr;
1186 		ip->ip_ttl = sc->sc_tp->t_inpcb->inp_ip_ttl;   /* XXX */
1187 		ip->ip_tos = sc->sc_tp->t_inpcb->inp_ip_tos;   /* XXX */
1188 
1189 		/*
1190 		 * See if we should do MTU discovery.  Route lookups are
1191 		 * expensive, so we will only unset the DF bit if:
1192 		 *
1193 		 *	1) path_mtu_discovery is disabled
1194 		 *	2) the SCF_UNREACH flag has been set
1195 		 */
1196 		if (path_mtu_discovery
1197 		    && ((sc->sc_flags & SCF_UNREACH) == 0)) {
1198 		       ip->ip_off |= IP_DF;
1199 		}
1200 
1201 		th = (struct tcphdr *)(ip + 1);
1202 	}
1203 	th->th_sport = sc->sc_inc.inc_lport;
1204 	th->th_dport = sc->sc_inc.inc_fport;
1205 
1206 	th->th_seq = htonl(sc->sc_iss);
1207 	th->th_ack = htonl(sc->sc_irs + 1);
1208 	th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1209 	th->th_x2 = 0;
1210 	th->th_flags = TH_SYN | TH_ACK;
1211 	th->th_win = htons(sc->sc_wnd);
1212 	th->th_urp = 0;
1213 
1214 	/* Tack on the TCP options. */
1215 	if (optlen == 0)
1216 		goto no_options;
1217 	optp = (u_int8_t *)(th + 1);
1218 	*optp++ = TCPOPT_MAXSEG;
1219 	*optp++ = TCPOLEN_MAXSEG;
1220 	*optp++ = (mssopt >> 8) & 0xff;
1221 	*optp++ = mssopt & 0xff;
1222 
1223 	if (sc->sc_flags & SCF_WINSCALE) {
1224 		*((u_int32_t *)optp) = htonl(TCPOPT_NOP << 24 |
1225 		    TCPOPT_WINDOW << 16 | TCPOLEN_WINDOW << 8 |
1226 		    sc->sc_request_r_scale);
1227 		optp += 4;
1228 	}
1229 
1230 	if (sc->sc_flags & SCF_TIMESTAMP) {
1231 		u_int32_t *lp = (u_int32_t *)(optp);
1232 
1233 		/* Form timestamp option as shown in appendix A of RFC 1323. */
1234 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
1235 		*lp++ = htonl(ticks);
1236 		*lp   = htonl(sc->sc_tsrecent);
1237 		optp += TCPOLEN_TSTAMP_APPA;
1238 	}
1239 
1240 	if (sc->sc_flags & SCF_SACK_PERMITTED) {
1241 		*((u_int32_t *)optp) = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
1242 		optp += TCPOLEN_SACK_PERMITTED_ALIGNED;
1243 	}
1244 
1245 no_options:
1246 	if (isipv6) {
1247 		struct route_in6 *ro6 = &sc->sc_route6;
1248 
1249 		th->th_sum = 0;
1250 		th->th_sum = in6_cksum(m, IPPROTO_TCP, hlen, tlen - hlen);
1251 		ip6->ip6_hlim = in6_selecthlim(NULL,
1252 		    ro6->ro_rt ? ro6->ro_rt->rt_ifp : NULL);
1253 		error = ip6_output(m, NULL, ro6, 0, NULL, NULL,
1254 				sc->sc_tp->t_inpcb);
1255 	} else {
1256 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1257 				       htons(tlen - hlen + IPPROTO_TCP));
1258 		m->m_pkthdr.csum_flags = CSUM_TCP;
1259 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1260 		error = ip_output(m, sc->sc_ipopts, &sc->sc_route,
1261 				  IP_DEBUGROUTE, NULL, sc->sc_tp->t_inpcb);
1262 	}
1263 	return (error);
1264 }
1265 
1266 /*
1267  * cookie layers:
1268  *
1269  *	|. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .|
1270  *	| peer iss                                                      |
1271  *	| MD5(laddr,faddr,secret,lport,fport)             |. . . . . . .|
1272  *	|                     0                       |(A)|             |
1273  * (A): peer mss index
1274  */
1275 
1276 /*
1277  * The values below are chosen to minimize the size of the tcp_secret
1278  * table, as well as providing roughly a 16 second lifetime for the cookie.
1279  */
1280 
1281 #define SYNCOOKIE_WNDBITS	5	/* exposed bits for window indexing */
1282 #define SYNCOOKIE_TIMESHIFT	1	/* scale ticks to window time units */
1283 
1284 #define SYNCOOKIE_WNDMASK	((1 << SYNCOOKIE_WNDBITS) - 1)
1285 #define SYNCOOKIE_NSECRETS	(1 << SYNCOOKIE_WNDBITS)
1286 #define SYNCOOKIE_TIMEOUT \
1287     (hz * (1 << SYNCOOKIE_WNDBITS) / (1 << SYNCOOKIE_TIMESHIFT))
1288 #define SYNCOOKIE_DATAMASK	((3 << SYNCOOKIE_WNDBITS) | SYNCOOKIE_WNDMASK)
1289 
1290 static struct {
1291 	u_int32_t	ts_secbits[4];
1292 	u_int		ts_expire;
1293 } tcp_secret[SYNCOOKIE_NSECRETS];
1294 
1295 static int tcp_msstab[] = { 0, 536, 1460, 8960 };
1296 
1297 static MD5_CTX syn_ctx;
1298 
1299 #define MD5Add(v)	MD5Update(&syn_ctx, (u_char *)&v, sizeof(v))
1300 
1301 struct md5_add {
1302 	u_int32_t laddr, faddr;
1303 	u_int32_t secbits[4];
1304 	u_int16_t lport, fport;
1305 };
1306 
1307 #ifdef CTASSERT
1308 CTASSERT(sizeof(struct md5_add) == 28);
1309 #endif
1310 
1311 /*
1312  * Consider the problem of a recreated (and retransmitted) cookie.  If the
1313  * original SYN was accepted, the connection is established.  The second
1314  * SYN is inflight, and if it arrives with an ISN that falls within the
1315  * receive window, the connection is killed.
1316  *
1317  * However, since cookies have other problems, this may not be worth
1318  * worrying about.
1319  */
1320 
1321 static u_int32_t
1322 syncookie_generate(struct syncache *sc)
1323 {
1324 	u_int32_t md5_buffer[4];
1325 	u_int32_t data;
1326 	int idx, i;
1327 	struct md5_add add;
1328 #ifdef INET6
1329 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
1330 #else
1331 	const boolean_t isipv6 = FALSE;
1332 #endif
1333 
1334 	idx = ((ticks << SYNCOOKIE_TIMESHIFT) / hz) & SYNCOOKIE_WNDMASK;
1335 	if (tcp_secret[idx].ts_expire < ticks) {
1336 		for (i = 0; i < 4; i++)
1337 			tcp_secret[idx].ts_secbits[i] = karc4random();
1338 		tcp_secret[idx].ts_expire = ticks + SYNCOOKIE_TIMEOUT;
1339 	}
1340 	for (data = sizeof(tcp_msstab) / sizeof(int) - 1; data > 0; data--)
1341 		if (tcp_msstab[data] <= sc->sc_peer_mss)
1342 			break;
1343 	data = (data << SYNCOOKIE_WNDBITS) | idx;
1344 	data ^= sc->sc_irs;				/* peer's iss */
1345 	MD5Init(&syn_ctx);
1346 	if (isipv6) {
1347 		MD5Add(sc->sc_inc.inc6_laddr);
1348 		MD5Add(sc->sc_inc.inc6_faddr);
1349 		add.laddr = 0;
1350 		add.faddr = 0;
1351 	} else {
1352 		add.laddr = sc->sc_inc.inc_laddr.s_addr;
1353 		add.faddr = sc->sc_inc.inc_faddr.s_addr;
1354 	}
1355 	add.lport = sc->sc_inc.inc_lport;
1356 	add.fport = sc->sc_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] & ~SYNCOOKIE_WNDMASK);
1364 	return (data);
1365 }
1366 
1367 static struct syncache *
1368 syncookie_lookup(struct in_conninfo *inc, struct tcphdr *th, struct socket *so)
1369 {
1370 	u_int32_t md5_buffer[4];
1371 	struct syncache *sc;
1372 	u_int32_t data;
1373 	int wnd, idx;
1374 	struct md5_add add;
1375 
1376 	data = (th->th_ack - 1) ^ (th->th_seq - 1);	/* remove ISS */
1377 	idx = data & SYNCOOKIE_WNDMASK;
1378 	if (tcp_secret[idx].ts_expire < ticks ||
1379 	    sototcpcb(so)->ts_recent + SYNCOOKIE_TIMEOUT < ticks)
1380 		return (NULL);
1381 	MD5Init(&syn_ctx);
1382 #ifdef INET6
1383 	if (inc->inc_isipv6) {
1384 		MD5Add(inc->inc6_laddr);
1385 		MD5Add(inc->inc6_faddr);
1386 		add.laddr = 0;
1387 		add.faddr = 0;
1388 	} else
1389 #endif
1390 	{
1391 		add.laddr = inc->inc_laddr.s_addr;
1392 		add.faddr = inc->inc_faddr.s_addr;
1393 	}
1394 	add.lport = inc->inc_lport;
1395 	add.fport = inc->inc_fport;
1396 	add.secbits[0] = tcp_secret[idx].ts_secbits[0];
1397 	add.secbits[1] = tcp_secret[idx].ts_secbits[1];
1398 	add.secbits[2] = tcp_secret[idx].ts_secbits[2];
1399 	add.secbits[3] = tcp_secret[idx].ts_secbits[3];
1400 	MD5Add(add);
1401 	MD5Final((u_char *)&md5_buffer, &syn_ctx);
1402 	data ^= md5_buffer[0];
1403 	if (data & ~SYNCOOKIE_DATAMASK)
1404 		return (NULL);
1405 	data = data >> SYNCOOKIE_WNDBITS;
1406 
1407 	/*
1408 	 * Fill in the syncache values.
1409 	 * XXX duplicate code from syncache_add
1410 	 */
1411 	sc = kmalloc(sizeof(struct syncache), M_SYNCACHE, M_WAITOK|M_ZERO);
1412 	sc->sc_ipopts = NULL;
1413 	sc->sc_inc.inc_fport = inc->inc_fport;
1414 	sc->sc_inc.inc_lport = inc->inc_lport;
1415 #ifdef INET6
1416 	sc->sc_inc.inc_isipv6 = inc->inc_isipv6;
1417 	if (inc->inc_isipv6) {
1418 		sc->sc_inc.inc6_faddr = inc->inc6_faddr;
1419 		sc->sc_inc.inc6_laddr = inc->inc6_laddr;
1420 		sc->sc_route6.ro_rt = NULL;
1421 	} else
1422 #endif
1423 	{
1424 		sc->sc_inc.inc_faddr = inc->inc_faddr;
1425 		sc->sc_inc.inc_laddr = inc->inc_laddr;
1426 		sc->sc_route.ro_rt = NULL;
1427 	}
1428 	sc->sc_irs = th->th_seq - 1;
1429 	sc->sc_iss = th->th_ack - 1;
1430 	wnd = ssb_space(&so->so_rcv);
1431 	wnd = imax(wnd, 0);
1432 	wnd = imin(wnd, TCP_MAXWIN);
1433 	sc->sc_wnd = wnd;
1434 	sc->sc_flags = 0;
1435 	sc->sc_rxtslot = 0;
1436 	sc->sc_peer_mss = tcp_msstab[data];
1437 	return (sc);
1438 }
1439