xref: /dragonfly/sys/netinet/tcp_syncache.c (revision c3762235)
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  */
73 
74 #include "opt_inet.h"
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 #include <net/netisr2.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  * 4 retransmits corresponds to a timeout of (3 + 3 + 3 + 3 + 3 == 15) seconds
156  * or (1 + 1 + 2 + 4 + 8 == 16) seconds if RFC6298 is used, the odds are that
157  * the user has given up attempting to connect by then.
158  */
159 #define SYNCACHE_MAXREXMTS		4
160 
161 /* Arbitrary values */
162 #define TCP_SYNCACHE_HASHSIZE		512
163 #define TCP_SYNCACHE_BUCKETLIMIT	30
164 
165 struct netmsg_sc_timer {
166 	struct netmsg_base base;
167 	struct msgrec *nm_mrec;		/* back pointer to containing msgrec */
168 };
169 
170 struct msgrec {
171 	struct netmsg_sc_timer msg;
172 	lwkt_port_t port;		/* constant after init */
173 	int slot;			/* constant after init */
174 };
175 
176 static void syncache_timer_handler(netmsg_t);
177 
178 struct tcp_syncache {
179 	u_int	hashsize;
180 	u_int	hashmask;
181 	u_int	bucket_limit;
182 	u_int	cache_limit;
183 	u_int	rexmt_limit;
184 	u_int	hash_secret;
185 };
186 static struct tcp_syncache tcp_syncache;
187 
188 TAILQ_HEAD(syncache_list, syncache);
189 
190 struct tcp_syncache_percpu {
191 	struct syncache_head	*hashbase;
192 	u_int			cache_count;
193 	struct syncache_list	timerq[SYNCACHE_MAXREXMTS + 1];
194 	struct callout		tt_timerq[SYNCACHE_MAXREXMTS + 1];
195 	struct msgrec		mrec[SYNCACHE_MAXREXMTS + 1];
196 };
197 static struct tcp_syncache_percpu tcp_syncache_percpu[MAXCPU];
198 
199 static struct lwkt_port syncache_null_rport;
200 
201 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, "TCP SYN cache");
202 
203 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RD,
204      &tcp_syncache.bucket_limit, 0, "Per-bucket hash limit for syncache");
205 
206 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_RD,
207      &tcp_syncache.cache_limit, 0, "Overall entry limit for syncache");
208 
209 /* XXX JH */
210 #if 0
211 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_RD,
212      &tcp_syncache.cache_count, 0, "Current number of entries in syncache");
213 #endif
214 
215 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RD,
216      &tcp_syncache.hashsize, 0, "Size of TCP syncache hashtable");
217 
218 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_RW,
219      &tcp_syncache.rexmt_limit, 0, "Limit on SYN/ACK retransmissions");
220 
221 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
222 
223 #define SYNCACHE_HASH(inc, mask)					\
224 	((tcp_syncache.hash_secret ^					\
225 	  (inc)->inc_faddr.s_addr ^					\
226 	  ((inc)->inc_faddr.s_addr >> 16) ^				\
227 	  (inc)->inc_fport ^ (inc)->inc_lport) & mask)
228 
229 #define SYNCACHE_HASH6(inc, mask)					\
230 	((tcp_syncache.hash_secret ^					\
231 	  (inc)->inc6_faddr.s6_addr32[0] ^				\
232 	  (inc)->inc6_faddr.s6_addr32[3] ^				\
233 	  (inc)->inc_fport ^ (inc)->inc_lport) & mask)
234 
235 #define ENDPTS_EQ(a, b) (						\
236 	(a)->ie_fport == (b)->ie_fport &&				\
237 	(a)->ie_lport == (b)->ie_lport &&				\
238 	(a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr &&			\
239 	(a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr			\
240 )
241 
242 #define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0)
243 
244 static __inline int
245 syncache_rto(int slot)
246 {
247 	if (tcp_low_rtobase)
248 		return (TCPTV_RTOBASE * tcp_syn_backoff_low[slot]);
249 	else
250 		return (TCPTV_RTOBASE * tcp_syn_backoff[slot]);
251 }
252 
253 static __inline void
254 syncache_timeout(struct tcp_syncache_percpu *syncache_percpu,
255 		 struct syncache *sc, int slot)
256 {
257 	int rto;
258 
259 	if (slot > 0) {
260 		/*
261 		 * Record the time that we spent in SYN|ACK
262 		 * retransmition.
263 		 *
264 		 * Needed by RFC3390 and RFC6298.
265 		 */
266 		sc->sc_rxtused += syncache_rto(slot - 1);
267 	}
268 	sc->sc_rxtslot = slot;
269 
270 	rto = syncache_rto(slot);
271 	sc->sc_rxttime = ticks + rto;
272 
273 	TAILQ_INSERT_TAIL(&syncache_percpu->timerq[slot], sc, sc_timerq);
274 	if (!callout_active(&syncache_percpu->tt_timerq[slot])) {
275 		callout_reset(&syncache_percpu->tt_timerq[slot], rto,
276 		    syncache_timer, &syncache_percpu->mrec[slot]);
277 	}
278 }
279 
280 static void
281 syncache_free(struct syncache *sc)
282 {
283 	struct rtentry *rt;
284 #ifdef INET6
285 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
286 #else
287 	const boolean_t isipv6 = FALSE;
288 #endif
289 
290 	if (sc->sc_ipopts)
291 		m_free(sc->sc_ipopts);
292 
293 	rt = isipv6 ? sc->sc_route6.ro_rt : sc->sc_route.ro_rt;
294 	if (rt != NULL) {
295 		/*
296 		 * If this is the only reference to a protocol-cloned
297 		 * route, remove it immediately.
298 		 */
299 		if ((rt->rt_flags & (RTF_WASCLONED | RTF_LLINFO)) ==
300 		    RTF_WASCLONED && rt->rt_refcnt == 1) {
301 			rtrequest(RTM_DELETE, rt_key(rt), rt->rt_gateway,
302 				  rt_mask(rt), rt->rt_flags, NULL);
303 		}
304 		RTFREE(rt);
305 	}
306 	kfree(sc, M_SYNCACHE);
307 }
308 
309 void
310 syncache_init(void)
311 {
312 	int i, cpu;
313 
314 	tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
315 	tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
316 	tcp_syncache.cache_limit =
317 	    tcp_syncache.hashsize * tcp_syncache.bucket_limit;
318 	tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
319 	tcp_syncache.hash_secret = karc4random();
320 
321 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
322 	    &tcp_syncache.hashsize);
323 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
324 	    &tcp_syncache.cache_limit);
325 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
326 	    &tcp_syncache.bucket_limit);
327 	if (!powerof2(tcp_syncache.hashsize)) {
328 		kprintf("WARNING: syncache hash size is not a power of 2.\n");
329 		tcp_syncache.hashsize = 512;	/* safe default */
330 	}
331 	tcp_syncache.hashmask = tcp_syncache.hashsize - 1;
332 
333 	lwkt_initport_replyonly_null(&syncache_null_rport);
334 
335 	for (cpu = 0; cpu < ncpus2; cpu++) {
336 		struct tcp_syncache_percpu *syncache_percpu;
337 
338 		syncache_percpu = &tcp_syncache_percpu[cpu];
339 		/* Allocate the hash table. */
340 		syncache_percpu->hashbase = kmalloc(tcp_syncache.hashsize * sizeof(struct syncache_head),
341 						    M_SYNCACHE, M_WAITOK);
342 
343 		/* Initialize the hash buckets. */
344 		for (i = 0; i < tcp_syncache.hashsize; i++) {
345 			struct syncache_head *bucket;
346 
347 			bucket = &syncache_percpu->hashbase[i];
348 			TAILQ_INIT(&bucket->sch_bucket);
349 			bucket->sch_length = 0;
350 		}
351 
352 		for (i = 0; i <= SYNCACHE_MAXREXMTS; i++) {
353 			/* Initialize the timer queues. */
354 			TAILQ_INIT(&syncache_percpu->timerq[i]);
355 			callout_init_mp(&syncache_percpu->tt_timerq[i]);
356 
357 			syncache_percpu->mrec[i].slot = i;
358 			syncache_percpu->mrec[i].port = netisr_cpuport(cpu);
359 			syncache_percpu->mrec[i].msg.nm_mrec =
360 				    &syncache_percpu->mrec[i];
361 			netmsg_init(&syncache_percpu->mrec[i].msg.base,
362 				    NULL, &syncache_null_rport,
363 				    0, syncache_timer_handler);
364 		}
365 	}
366 }
367 
368 static void
369 syncache_insert(struct syncache *sc, struct syncache_head *sch)
370 {
371 	struct tcp_syncache_percpu *syncache_percpu;
372 	struct syncache *sc2;
373 	int i;
374 
375 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
376 
377 	/*
378 	 * Make sure that we don't overflow the per-bucket
379 	 * limit or the total cache size limit.
380 	 */
381 	if (sch->sch_length >= tcp_syncache.bucket_limit) {
382 		/*
383 		 * The bucket is full, toss the oldest element.
384 		 */
385 		sc2 = TAILQ_FIRST(&sch->sch_bucket);
386 		if (sc2->sc_tp != NULL)
387 			sc2->sc_tp->ts_recent = ticks;
388 		syncache_drop(sc2, sch);
389 		tcpstat.tcps_sc_bucketoverflow++;
390 	} else if (syncache_percpu->cache_count >= tcp_syncache.cache_limit) {
391 		/*
392 		 * The cache is full.  Toss the oldest entry in the
393 		 * entire cache.  This is the front entry in the
394 		 * first non-empty timer queue with the largest
395 		 * timeout value.
396 		 */
397 		for (i = SYNCACHE_MAXREXMTS; i >= 0; i--) {
398 			sc2 = TAILQ_FIRST(&syncache_percpu->timerq[i]);
399 			while (sc2 && (sc2->sc_flags & SCF_MARKER))
400 				sc2 = TAILQ_NEXT(sc2, sc_timerq);
401 			if (sc2 != NULL)
402 				break;
403 		}
404 		if (sc2->sc_tp != NULL)
405 			sc2->sc_tp->ts_recent = ticks;
406 		syncache_drop(sc2, NULL);
407 		tcpstat.tcps_sc_cacheoverflow++;
408 	}
409 
410 	/* Initialize the entry's timer. */
411 	syncache_timeout(syncache_percpu, sc, 0);
412 
413 	/* Put it into the bucket. */
414 	TAILQ_INSERT_TAIL(&sch->sch_bucket, sc, sc_hash);
415 	sch->sch_length++;
416 	syncache_percpu->cache_count++;
417 	tcpstat.tcps_sc_added++;
418 }
419 
420 void
421 syncache_destroy(struct tcpcb *tp, struct tcpcb *tp_inh)
422 {
423 	struct tcp_syncache_percpu *syncache_percpu;
424 	struct syncache_head *bucket;
425 	struct syncache *sc;
426 	int i;
427 
428 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
429 	sc = NULL;
430 
431 	for (i = 0; i < tcp_syncache.hashsize; i++) {
432 		bucket = &syncache_percpu->hashbase[i];
433 		TAILQ_FOREACH(sc, &bucket->sch_bucket, sc_hash) {
434 			if (sc->sc_tp == tp)
435 				sc->sc_tp = tp_inh;
436 		}
437 	}
438 }
439 
440 static void
441 syncache_drop(struct syncache *sc, struct syncache_head *sch)
442 {
443 	struct tcp_syncache_percpu *syncache_percpu;
444 #ifdef INET6
445 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
446 #else
447 	const boolean_t isipv6 = FALSE;
448 #endif
449 
450 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
451 
452 	if (sch == NULL) {
453 		if (isipv6) {
454 			sch = &syncache_percpu->hashbase[
455 			    SYNCACHE_HASH6(&sc->sc_inc, tcp_syncache.hashmask)];
456 		} else {
457 			sch = &syncache_percpu->hashbase[
458 			    SYNCACHE_HASH(&sc->sc_inc, tcp_syncache.hashmask)];
459 		}
460 	}
461 
462 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
463 	sch->sch_length--;
464 	syncache_percpu->cache_count--;
465 
466 	/*
467 	 * Cleanup
468 	 */
469 	sc->sc_tp = NULL;
470 
471 	/*
472 	 * Remove the entry from the syncache timer/timeout queue.  Note
473 	 * that we do not try to stop any running timer since we do not know
474 	 * whether the timer's message is in-transit or not.  Since timeouts
475 	 * are fairly long, taking an unneeded callout does not detrimentally
476 	 * effect performance.
477 	 */
478 	TAILQ_REMOVE(&syncache_percpu->timerq[sc->sc_rxtslot], sc, sc_timerq);
479 
480 	syncache_free(sc);
481 }
482 
483 /*
484  * Place a timeout message on the TCP thread's message queue.
485  * This routine runs in soft interrupt context.
486  *
487  * An invariant is for this routine to be called, the callout must
488  * have been active.  Note that the callout is not deactivated until
489  * after the message has been processed in syncache_timer_handler() below.
490  */
491 static void
492 syncache_timer(void *p)
493 {
494 	struct netmsg_sc_timer *msg = p;
495 
496 	lwkt_sendmsg_oncpu(msg->nm_mrec->port, &msg->base.lmsg);
497 }
498 
499 /*
500  * Service a timer message queued by timer expiration.
501  * This routine runs in the TCP protocol thread.
502  *
503  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
504  * If we have retransmitted an entry the maximum number of times, expire it.
505  *
506  * When we finish processing timed-out entries, we restart the timer if there
507  * are any entries still on the queue and deactivate it otherwise.  Only after
508  * a timer has been deactivated here can it be restarted by syncache_timeout().
509  */
510 static void
511 syncache_timer_handler(netmsg_t msg)
512 {
513 	struct tcp_syncache_percpu *syncache_percpu;
514 	struct syncache *sc;
515 	struct syncache marker;
516 	struct syncache_list *list;
517 	struct inpcb *inp;
518 	int slot;
519 
520 	slot = ((struct netmsg_sc_timer *)msg)->nm_mrec->slot;
521 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
522 
523 	list = &syncache_percpu->timerq[slot];
524 
525 	/*
526 	 * Use a marker to keep our place in the scan.  syncache_drop()
527 	 * can block and cause any next pointer we cache to become stale.
528 	 */
529 	marker.sc_flags = SCF_MARKER;
530 	TAILQ_INSERT_HEAD(list, &marker, sc_timerq);
531 
532 	while ((sc = TAILQ_NEXT(&marker, sc_timerq)) != NULL) {
533 		/*
534 		 * Move the marker.
535 		 */
536 		TAILQ_REMOVE(list, &marker, sc_timerq);
537 		TAILQ_INSERT_AFTER(list, sc, &marker, sc_timerq);
538 
539 		if (sc->sc_flags & SCF_MARKER)
540 			continue;
541 
542 		if (ticks < sc->sc_rxttime)
543 			break;	/* finished because timerq sorted by time */
544 		if (sc->sc_tp == NULL) {
545 			syncache_drop(sc, NULL);
546 			tcpstat.tcps_sc_stale++;
547 			continue;
548 		}
549 		inp = sc->sc_tp->t_inpcb;
550 		if (slot == SYNCACHE_MAXREXMTS ||
551 		    slot >= tcp_syncache.rexmt_limit ||
552 		    inp == NULL ||
553 		    inp->inp_gencnt != sc->sc_inp_gencnt) {
554 			syncache_drop(sc, NULL);
555 			tcpstat.tcps_sc_stale++;
556 			continue;
557 		}
558 		/*
559 		 * syncache_respond() may call back into the syncache to
560 		 * to modify another entry, so do not obtain the next
561 		 * entry on the timer chain until it has completed.
562 		 */
563 		syncache_respond(sc, NULL);
564 		tcpstat.tcps_sc_retransmitted++;
565 		TAILQ_REMOVE(list, sc, sc_timerq);
566 		syncache_timeout(syncache_percpu, sc, slot + 1);
567 	}
568 	TAILQ_REMOVE(list, &marker, sc_timerq);
569 
570 	if (sc != NULL) {
571 		callout_reset(&syncache_percpu->tt_timerq[slot],
572 			      sc->sc_rxttime - ticks, syncache_timer,
573 			      &syncache_percpu->mrec[slot]);
574 	} else {
575 		callout_deactivate(&syncache_percpu->tt_timerq[slot]);
576 	}
577 	lwkt_replymsg(&msg->base.lmsg, 0);
578 }
579 
580 /*
581  * Find an entry in the syncache.
582  */
583 struct syncache *
584 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
585 {
586 	struct tcp_syncache_percpu *syncache_percpu;
587 	struct syncache *sc;
588 	struct syncache_head *sch;
589 
590 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
591 #ifdef INET6
592 	if (inc->inc_isipv6) {
593 		sch = &syncache_percpu->hashbase[
594 		    SYNCACHE_HASH6(inc, tcp_syncache.hashmask)];
595 		*schp = sch;
596 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
597 			if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
598 				return (sc);
599 	} else
600 #endif
601 	{
602 		sch = &syncache_percpu->hashbase[
603 		    SYNCACHE_HASH(inc, tcp_syncache.hashmask)];
604 		*schp = sch;
605 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
606 #ifdef INET6
607 			if (sc->sc_inc.inc_isipv6)
608 				continue;
609 #endif
610 			if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
611 				return (sc);
612 		}
613 	}
614 	return (NULL);
615 }
616 
617 /*
618  * This function is called when we get a RST for a
619  * non-existent connection, so that we can see if the
620  * connection is in the syn cache.  If it is, zap it.
621  */
622 void
623 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
624 {
625 	struct syncache *sc;
626 	struct syncache_head *sch;
627 
628 	sc = syncache_lookup(inc, &sch);
629 	if (sc == NULL) {
630 		return;
631 	}
632 	/*
633 	 * If the RST bit is set, check the sequence number to see
634 	 * if this is a valid reset segment.
635 	 * RFC 793 page 37:
636 	 *   In all states except SYN-SENT, all reset (RST) segments
637 	 *   are validated by checking their SEQ-fields.  A reset is
638 	 *   valid if its sequence number is in the window.
639 	 *
640 	 *   The sequence number in the reset segment is normally an
641 	 *   echo of our outgoing acknowlegement numbers, but some hosts
642 	 *   send a reset with the sequence number at the rightmost edge
643 	 *   of our receive window, and we have to handle this case.
644 	 */
645 	if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
646 	    SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
647 		syncache_drop(sc, sch);
648 		tcpstat.tcps_sc_reset++;
649 	}
650 }
651 
652 void
653 syncache_badack(struct in_conninfo *inc)
654 {
655 	struct syncache *sc;
656 	struct syncache_head *sch;
657 
658 	sc = syncache_lookup(inc, &sch);
659 	if (sc != NULL) {
660 		syncache_drop(sc, sch);
661 		tcpstat.tcps_sc_badack++;
662 	}
663 }
664 
665 void
666 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th)
667 {
668 	struct syncache *sc;
669 	struct syncache_head *sch;
670 
671 	/* we are called at splnet() here */
672 	sc = syncache_lookup(inc, &sch);
673 	if (sc == NULL)
674 		return;
675 
676 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
677 	if (ntohl(th->th_seq) != sc->sc_iss)
678 		return;
679 
680 	/*
681 	 * If we've rertransmitted 3 times and this is our second error,
682 	 * we remove the entry.  Otherwise, we allow it to continue on.
683 	 * This prevents us from incorrectly nuking an entry during a
684 	 * spurious network outage.
685 	 *
686 	 * See tcp_notify().
687 	 */
688 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxtslot < 3) {
689 		sc->sc_flags |= SCF_UNREACH;
690 		return;
691 	}
692 	syncache_drop(sc, sch);
693 	tcpstat.tcps_sc_unreach++;
694 }
695 
696 /*
697  * Build a new TCP socket structure from a syncache entry.
698  *
699  * This is called from the context of the SYN+ACK
700  */
701 static struct socket *
702 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
703 {
704 	struct inpcb *inp = NULL, *linp;
705 	struct socket *so;
706 	struct tcpcb *tp, *ltp;
707 	lwkt_port_t port;
708 #ifdef INET6
709 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
710 #else
711 	const boolean_t isipv6 = FALSE;
712 #endif
713 	struct sockaddr_in sin_faddr;
714 	struct sockaddr_in6 sin6_faddr;
715 	struct sockaddr *faddr;
716 
717 	if (isipv6) {
718 		faddr = (struct sockaddr *)&sin6_faddr;
719 		sin6_faddr.sin6_family = AF_INET6;
720 		sin6_faddr.sin6_len = sizeof(sin6_faddr);
721 		sin6_faddr.sin6_addr = sc->sc_inc.inc6_faddr;
722 		sin6_faddr.sin6_port = sc->sc_inc.inc_fport;
723 		sin6_faddr.sin6_flowinfo = sin6_faddr.sin6_scope_id = 0;
724 	} else {
725 		faddr = (struct sockaddr *)&sin_faddr;
726 		sin_faddr.sin_family = AF_INET;
727 		sin_faddr.sin_len = sizeof(sin_faddr);
728 		sin_faddr.sin_addr = sc->sc_inc.inc_faddr;
729 		sin_faddr.sin_port = sc->sc_inc.inc_fport;
730 		bzero(sin_faddr.sin_zero, sizeof(sin_faddr.sin_zero));
731 	}
732 
733 	/*
734 	 * Ok, create the full blown connection, and set things up
735 	 * as they would have been set up if we had created the
736 	 * connection when the SYN arrived.  If we can't create
737 	 * the connection, abort it.
738 	 *
739 	 * Set the protocol processing port for the socket to the current
740 	 * port (that the connection came in on).
741 	 */
742 	so = sonewconn_faddr(lso, SS_ISCONNECTED, faddr);
743 	if (so == NULL) {
744 		/*
745 		 * Drop the connection; we will send a RST if the peer
746 		 * retransmits the ACK,
747 		 */
748 		tcpstat.tcps_listendrop++;
749 		goto abort;
750 	}
751 
752 	/*
753 	 * Insert new socket into hash list.
754 	 */
755 	inp = so->so_pcb;
756 	inp->inp_inc.inc_isipv6 = sc->sc_inc.inc_isipv6;
757 	if (isipv6) {
758 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
759 	} else {
760 #ifdef INET6
761 		inp->inp_vflag &= ~INP_IPV6;
762 		inp->inp_vflag |= INP_IPV4;
763 		inp->inp_flags &= ~IN6P_IPV6_V6ONLY;
764 #endif
765 		inp->inp_laddr = sc->sc_inc.inc_laddr;
766 	}
767 	inp->inp_lport = sc->sc_inc.inc_lport;
768 	if (in_pcbinsporthash(inp) != 0) {
769 		/*
770 		 * Undo the assignments above if we failed to
771 		 * put the PCB on the hash lists.
772 		 */
773 		if (isipv6)
774 			inp->in6p_laddr = kin6addr_any;
775 		else
776 			inp->inp_laddr.s_addr = INADDR_ANY;
777 		inp->inp_lport = 0;
778 		goto abort;
779 	}
780 	linp = lso->so_pcb;
781 #ifdef IPSEC
782 	/* copy old policy into new socket's */
783 	if (ipsec_copy_policy(linp->inp_sp, inp->inp_sp))
784 		kprintf("syncache_expand: could not copy policy\n");
785 #endif
786 	if (isipv6) {
787 		struct in6_addr laddr6;
788 		/*
789 		 * Inherit socket options from the listening socket.
790 		 * Note that in6p_inputopts are not (and should not be)
791 		 * copied, since it stores previously received options and is
792 		 * used to detect if each new option is different than the
793 		 * previous one and hence should be passed to a user.
794 		 * If we copied in6p_inputopts, a user would not be able to
795 		 * receive options just after calling the accept system call.
796 		 */
797 		inp->inp_flags |= linp->inp_flags & INP_CONTROLOPTS;
798 		if (linp->in6p_outputopts)
799 			inp->in6p_outputopts =
800 			    ip6_copypktopts(linp->in6p_outputopts, M_INTWAIT);
801 		inp->in6p_route = sc->sc_route6;
802 		sc->sc_route6.ro_rt = NULL;
803 
804 		laddr6 = inp->in6p_laddr;
805 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
806 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
807 		if (in6_pcbconnect(inp, faddr, &thread0)) {
808 			inp->in6p_laddr = laddr6;
809 			goto abort;
810 		}
811 	} else {
812 		struct in_addr laddr;
813 
814 		inp->inp_options = ip_srcroute(m);
815 		if (inp->inp_options == NULL) {
816 			inp->inp_options = sc->sc_ipopts;
817 			sc->sc_ipopts = NULL;
818 		}
819 		inp->inp_route = sc->sc_route;
820 		sc->sc_route.ro_rt = NULL;
821 
822 		laddr = inp->inp_laddr;
823 		if (inp->inp_laddr.s_addr == INADDR_ANY)
824 			inp->inp_laddr = sc->sc_inc.inc_laddr;
825 		if (in_pcbconnect(inp, faddr, &thread0)) {
826 			inp->inp_laddr = laddr;
827 			goto abort;
828 		}
829 	}
830 
831 	/*
832 	 * The current port should be in the context of the SYN+ACK and
833 	 * so should match the tcp address port.
834 	 */
835 	if (isipv6) {
836 		port = tcp6_addrport();
837 	} else {
838 		port = tcp_addrport(inp->inp_faddr.s_addr, inp->inp_fport,
839 				    inp->inp_laddr.s_addr, inp->inp_lport);
840 	}
841 	KASSERT(port == &curthread->td_msgport,
842 	    ("TCP PORT MISMATCH %p vs %p\n", port, &curthread->td_msgport));
843 
844 	tp = intotcpcb(inp);
845 	tp->t_state = TCPS_SYN_RECEIVED;
846 	tp->iss = sc->sc_iss;
847 	tp->irs = sc->sc_irs;
848 	tcp_rcvseqinit(tp);
849 	tcp_sendseqinit(tp);
850 	tp->snd_wnd = sc->sc_sndwnd;
851 	tp->snd_wl1 = sc->sc_irs;
852 	tp->rcv_up = sc->sc_irs + 1;
853 	tp->rcv_wnd = sc->sc_wnd;
854 	tp->rcv_adv += tp->rcv_wnd;
855 
856 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH | TF_NODELAY);
857 	if (sc->sc_flags & SCF_NOOPT)
858 		tp->t_flags |= TF_NOOPT;
859 	if (sc->sc_flags & SCF_WINSCALE) {
860 		tp->t_flags |= TF_REQ_SCALE | TF_RCVD_SCALE;
861 		tp->snd_scale = sc->sc_requested_s_scale;
862 		tp->request_r_scale = sc->sc_request_r_scale;
863 	}
864 	if (sc->sc_flags & SCF_TIMESTAMP) {
865 		tp->t_flags |= TF_REQ_TSTMP | TF_RCVD_TSTMP;
866 		tp->ts_recent = sc->sc_tsrecent;
867 		tp->ts_recent_age = ticks;
868 	}
869 	if (sc->sc_flags & SCF_SACK_PERMITTED)
870 		tp->t_flags |= TF_SACK_PERMITTED;
871 
872 #ifdef TCP_SIGNATURE
873 	if (sc->sc_flags & SCF_SIGNATURE)
874 		tp->t_flags |= TF_SIGNATURE;
875 #endif /* TCP_SIGNATURE */
876 
877 	tp->t_rxtsyn = sc->sc_rxtused;
878 	tcp_mss(tp, sc->sc_peer_mss);
879 
880 	/*
881 	 * Inherit some properties from the listen socket
882 	 */
883 	ltp = intotcpcb(linp);
884 	tp->t_keepinit = ltp->t_keepinit;
885 	tp->t_keepidle = ltp->t_keepidle;
886 	tp->t_keepintvl = ltp->t_keepintvl;
887 	tp->t_keepcnt = ltp->t_keepcnt;
888 	tp->t_maxidle = ltp->t_maxidle;
889 
890 	tcp_create_timermsg(tp, port);
891 	tcp_callout_reset(tp, tp->tt_keep, tp->t_keepinit, tcp_timer_keep);
892 
893 	tcpstat.tcps_accepts++;
894 	return (so);
895 
896 abort:
897 	if (so != NULL)
898 		soabort_oncpu(so);
899 	return (NULL);
900 }
901 
902 /*
903  * This function gets called when we receive an ACK for a
904  * socket in the LISTEN state.  We look up the connection
905  * in the syncache, and if its there, we pull it out of
906  * the cache and turn it into a full-blown connection in
907  * the SYN-RECEIVED state.
908  */
909 int
910 syncache_expand(struct in_conninfo *inc, struct tcphdr *th, struct socket **sop,
911 		struct mbuf *m)
912 {
913 	struct syncache *sc;
914 	struct syncache_head *sch;
915 	struct socket *so;
916 
917 	sc = syncache_lookup(inc, &sch);
918 	if (sc == NULL) {
919 		/*
920 		 * There is no syncache entry, so see if this ACK is
921 		 * a returning syncookie.  To do this, first:
922 		 *  A. See if this socket has had a syncache entry dropped in
923 		 *     the past.  We don't want to accept a bogus syncookie
924 		 *     if we've never received a SYN.
925 		 *  B. check that the syncookie is valid.  If it is, then
926 		 *     cobble up a fake syncache entry, and return.
927 		 */
928 		if (!tcp_syncookies)
929 			return (0);
930 		sc = syncookie_lookup(inc, th, *sop);
931 		if (sc == NULL)
932 			return (0);
933 		sch = NULL;
934 		tcpstat.tcps_sc_recvcookie++;
935 	}
936 
937 	/*
938 	 * If seg contains an ACK, but not for our SYN/ACK, send a RST.
939 	 */
940 	if (th->th_ack != sc->sc_iss + 1)
941 		return (0);
942 
943 	so = syncache_socket(sc, *sop, m);
944 	if (so == NULL) {
945 #if 0
946 resetandabort:
947 		/* XXXjlemon check this - is this correct? */
948 		tcp_respond(NULL, m, m, th,
949 		    th->th_seq + tlen, (tcp_seq)0, TH_RST | TH_ACK);
950 #endif
951 		m_freem(m);			/* XXX only needed for above */
952 		tcpstat.tcps_sc_aborted++;
953 	} else {
954 		tcpstat.tcps_sc_completed++;
955 	}
956 	if (sch == NULL)
957 		syncache_free(sc);
958 	else
959 		syncache_drop(sc, sch);
960 	*sop = so;
961 	return (1);
962 }
963 
964 /*
965  * Given a LISTEN socket and an inbound SYN request, add
966  * this to the syn cache, and send back a segment:
967  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
968  * to the source.
969  *
970  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
971  * Doing so would require that we hold onto the data and deliver it
972  * to the application.  However, if we are the target of a SYN-flood
973  * DoS attack, an attacker could send data which would eventually
974  * consume all available buffer space if it were ACKed.  By not ACKing
975  * the data, we avoid this DoS scenario.
976  */
977 int
978 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
979 	     struct socket *so, struct mbuf *m)
980 {
981 	struct tcp_syncache_percpu *syncache_percpu;
982 	struct tcpcb *tp;
983 	struct syncache *sc = NULL;
984 	struct syncache_head *sch;
985 	struct mbuf *ipopts = NULL;
986 	int win;
987 
988 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
989 	tp = sototcpcb(so);
990 
991 	/*
992 	 * Remember the IP options, if any.
993 	 */
994 #ifdef INET6
995 	if (!inc->inc_isipv6)
996 #endif
997 		ipopts = ip_srcroute(m);
998 
999 	/*
1000 	 * See if we already have an entry for this connection.
1001 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1002 	 *
1003 	 * XXX
1004 	 * The syncache should be re-initialized with the contents
1005 	 * of the new SYN which may have different options.
1006 	 */
1007 	sc = syncache_lookup(inc, &sch);
1008 	if (sc != NULL) {
1009 		tcpstat.tcps_sc_dupsyn++;
1010 		if (ipopts) {
1011 			/*
1012 			 * If we were remembering a previous source route,
1013 			 * forget it and use the new one we've been given.
1014 			 */
1015 			if (sc->sc_ipopts)
1016 				m_free(sc->sc_ipopts);
1017 			sc->sc_ipopts = ipopts;
1018 		}
1019 		/*
1020 		 * Update timestamp if present.
1021 		 */
1022 		if (sc->sc_flags & SCF_TIMESTAMP)
1023 			sc->sc_tsrecent = to->to_tsval;
1024 
1025 		/* Just update the TOF_SACK_PERMITTED for now. */
1026 		if (tcp_do_sack && (to->to_flags & TOF_SACK_PERMITTED))
1027 			sc->sc_flags |= SCF_SACK_PERMITTED;
1028 		else
1029 			sc->sc_flags &= ~SCF_SACK_PERMITTED;
1030 
1031 		/* Update initial send window */
1032 		sc->sc_sndwnd = th->th_win;
1033 
1034 		/*
1035 		 * PCB may have changed, pick up new values.
1036 		 */
1037 		sc->sc_tp = tp;
1038 		sc->sc_inp_gencnt = tp->t_inpcb->inp_gencnt;
1039 		if (syncache_respond(sc, m) == 0) {
1040 			TAILQ_REMOVE(&syncache_percpu->timerq[sc->sc_rxtslot],
1041 				     sc, sc_timerq);
1042 			syncache_timeout(syncache_percpu, sc, sc->sc_rxtslot);
1043 			tcpstat.tcps_sndacks++;
1044 			tcpstat.tcps_sndtotal++;
1045 		}
1046 		return (1);
1047 	}
1048 
1049 	/*
1050 	 * Fill in the syncache values.
1051 	 */
1052 	sc = kmalloc(sizeof(struct syncache), M_SYNCACHE, M_WAITOK|M_ZERO);
1053 	sc->sc_inp_gencnt = tp->t_inpcb->inp_gencnt;
1054 	sc->sc_ipopts = ipopts;
1055 	sc->sc_inc.inc_fport = inc->inc_fport;
1056 	sc->sc_inc.inc_lport = inc->inc_lport;
1057 	sc->sc_tp = tp;
1058 #ifdef INET6
1059 	sc->sc_inc.inc_isipv6 = inc->inc_isipv6;
1060 	if (inc->inc_isipv6) {
1061 		sc->sc_inc.inc6_faddr = inc->inc6_faddr;
1062 		sc->sc_inc.inc6_laddr = inc->inc6_laddr;
1063 		sc->sc_route6.ro_rt = NULL;
1064 	} else
1065 #endif
1066 	{
1067 		sc->sc_inc.inc_faddr = inc->inc_faddr;
1068 		sc->sc_inc.inc_laddr = inc->inc_laddr;
1069 		sc->sc_route.ro_rt = NULL;
1070 	}
1071 	sc->sc_irs = th->th_seq;
1072 	sc->sc_flags = 0;
1073 	sc->sc_peer_mss = to->to_flags & TOF_MSS ? to->to_mss : 0;
1074 	if (tcp_syncookies)
1075 		sc->sc_iss = syncookie_generate(sc);
1076 	else
1077 		sc->sc_iss = karc4random();
1078 
1079 	/* Initial receive window: clip ssb_space to [0 .. TCP_MAXWIN] */
1080 	win = ssb_space(&so->so_rcv);
1081 	win = imax(win, 0);
1082 	win = imin(win, TCP_MAXWIN);
1083 	sc->sc_wnd = win;
1084 
1085 	if (tcp_do_rfc1323) {
1086 		/*
1087 		 * A timestamp received in a SYN makes
1088 		 * it ok to send timestamp requests and replies.
1089 		 */
1090 		if (to->to_flags & TOF_TS) {
1091 			sc->sc_tsrecent = to->to_tsval;
1092 			sc->sc_flags |= SCF_TIMESTAMP;
1093 		}
1094 		if (to->to_flags & TOF_SCALE) {
1095 			int wscale = TCP_MIN_WINSHIFT;
1096 
1097 			/* Compute proper scaling value from buffer space */
1098 			while (wscale < TCP_MAX_WINSHIFT &&
1099 			    (TCP_MAXWIN << wscale) < so->so_rcv.ssb_hiwat) {
1100 				wscale++;
1101 			}
1102 			sc->sc_request_r_scale = wscale;
1103 			sc->sc_requested_s_scale = to->to_requested_s_scale;
1104 			sc->sc_flags |= SCF_WINSCALE;
1105 		}
1106 	}
1107 	if (tcp_do_sack && (to->to_flags & TOF_SACK_PERMITTED))
1108 		sc->sc_flags |= SCF_SACK_PERMITTED;
1109 	if (tp->t_flags & TF_NOOPT)
1110 		sc->sc_flags = SCF_NOOPT;
1111 #ifdef TCP_SIGNATURE
1112 	/*
1113 	 * If listening socket requested TCP digests, and received SYN
1114 	 * contains the option, flag this in the syncache so that
1115 	 * syncache_respond() will do the right thing with the SYN+ACK.
1116 	 * XXX Currently we always record the option by default and will
1117 	 * attempt to use it in syncache_respond().
1118 	 */
1119 	if (to->to_flags & TOF_SIGNATURE)
1120 		sc->sc_flags = SCF_SIGNATURE;
1121 #endif /* TCP_SIGNATURE */
1122 	sc->sc_sndwnd = th->th_win;
1123 
1124 	if (syncache_respond(sc, m) == 0) {
1125 		syncache_insert(sc, sch);
1126 		tcpstat.tcps_sndacks++;
1127 		tcpstat.tcps_sndtotal++;
1128 	} else {
1129 		syncache_free(sc);
1130 		tcpstat.tcps_sc_dropped++;
1131 	}
1132 	return (1);
1133 }
1134 
1135 static int
1136 syncache_respond(struct syncache *sc, struct mbuf *m)
1137 {
1138 	u_int8_t *optp;
1139 	int optlen, error;
1140 	u_int16_t tlen, hlen, mssopt;
1141 	struct ip *ip = NULL;
1142 	struct rtentry *rt;
1143 	struct tcphdr *th;
1144 	struct ip6_hdr *ip6 = NULL;
1145 #ifdef INET6
1146 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
1147 #else
1148 	const boolean_t isipv6 = FALSE;
1149 #endif
1150 
1151 	if (isipv6) {
1152 		rt = tcp_rtlookup6(&sc->sc_inc);
1153 		if (rt != NULL)
1154 			mssopt = rt->rt_ifp->if_mtu -
1155 			     (sizeof(struct ip6_hdr) + sizeof(struct tcphdr));
1156 		else
1157 			mssopt = tcp_v6mssdflt;
1158 		hlen = sizeof(struct ip6_hdr);
1159 	} else {
1160 		rt = tcp_rtlookup(&sc->sc_inc);
1161 		if (rt != NULL)
1162 			mssopt = rt->rt_ifp->if_mtu -
1163 			     (sizeof(struct ip) + sizeof(struct tcphdr));
1164 		else
1165 			mssopt = tcp_mssdflt;
1166 		hlen = sizeof(struct ip);
1167 	}
1168 
1169 	/* Compute the size of the TCP options. */
1170 	if (sc->sc_flags & SCF_NOOPT) {
1171 		optlen = 0;
1172 	} else {
1173 		optlen = TCPOLEN_MAXSEG +
1174 		    ((sc->sc_flags & SCF_WINSCALE) ? 4 : 0) +
1175 		    ((sc->sc_flags & SCF_TIMESTAMP) ? TCPOLEN_TSTAMP_APPA : 0) +
1176 		    ((sc->sc_flags & SCF_SACK_PERMITTED) ?
1177 			TCPOLEN_SACK_PERMITTED_ALIGNED : 0);
1178 #ifdef TCP_SIGNATURE
1179 		optlen += ((sc->sc_flags & SCF_SIGNATURE) ?
1180 		    (TCPOLEN_SIGNATURE + 2) : 0);
1181 #endif /* TCP_SIGNATURE */
1182 	}
1183 	tlen = hlen + sizeof(struct tcphdr) + optlen;
1184 
1185 	/*
1186 	 * XXX
1187 	 * assume that the entire packet will fit in a header mbuf
1188 	 */
1189 	KASSERT(max_linkhdr + tlen <= MHLEN, ("syncache: mbuf too small"));
1190 
1191 	/*
1192 	 * XXX shouldn't this reuse the mbuf if possible ?
1193 	 * Create the IP+TCP header from scratch.
1194 	 */
1195 	if (m)
1196 		m_freem(m);
1197 
1198 	m = m_gethdr(MB_DONTWAIT, MT_HEADER);
1199 	if (m == NULL)
1200 		return (ENOBUFS);
1201 	m->m_data += max_linkhdr;
1202 	m->m_len = tlen;
1203 	m->m_pkthdr.len = tlen;
1204 	m->m_pkthdr.rcvif = NULL;
1205 	if (tcp_prio_synack)
1206 		m->m_flags |= M_PRIO;
1207 
1208 	if (isipv6) {
1209 		ip6 = mtod(m, struct ip6_hdr *);
1210 		ip6->ip6_vfc = IPV6_VERSION;
1211 		ip6->ip6_nxt = IPPROTO_TCP;
1212 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1213 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1214 		ip6->ip6_plen = htons(tlen - hlen);
1215 		/* ip6_hlim is set after checksum */
1216 		/* ip6_flow = ??? */
1217 
1218 		th = (struct tcphdr *)(ip6 + 1);
1219 	} else {
1220 		ip = mtod(m, struct ip *);
1221 		ip->ip_v = IPVERSION;
1222 		ip->ip_hl = sizeof(struct ip) >> 2;
1223 		ip->ip_len = tlen;
1224 		ip->ip_id = 0;
1225 		ip->ip_off = 0;
1226 		ip->ip_sum = 0;
1227 		ip->ip_p = IPPROTO_TCP;
1228 		ip->ip_src = sc->sc_inc.inc_laddr;
1229 		ip->ip_dst = sc->sc_inc.inc_faddr;
1230 		ip->ip_ttl = sc->sc_tp->t_inpcb->inp_ip_ttl;   /* XXX */
1231 		ip->ip_tos = sc->sc_tp->t_inpcb->inp_ip_tos;   /* XXX */
1232 
1233 		/*
1234 		 * See if we should do MTU discovery.  Route lookups are
1235 		 * expensive, so we will only unset the DF bit if:
1236 		 *
1237 		 *	1) path_mtu_discovery is disabled
1238 		 *	2) the SCF_UNREACH flag has been set
1239 		 */
1240 		if (path_mtu_discovery
1241 		    && ((sc->sc_flags & SCF_UNREACH) == 0)) {
1242 		       ip->ip_off |= IP_DF;
1243 		}
1244 
1245 		th = (struct tcphdr *)(ip + 1);
1246 	}
1247 	th->th_sport = sc->sc_inc.inc_lport;
1248 	th->th_dport = sc->sc_inc.inc_fport;
1249 
1250 	th->th_seq = htonl(sc->sc_iss);
1251 	th->th_ack = htonl(sc->sc_irs + 1);
1252 	th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1253 	th->th_x2 = 0;
1254 	th->th_flags = TH_SYN | TH_ACK;
1255 	th->th_win = htons(sc->sc_wnd);
1256 	th->th_urp = 0;
1257 
1258 	/* Tack on the TCP options. */
1259 	if (optlen == 0)
1260 		goto no_options;
1261 	optp = (u_int8_t *)(th + 1);
1262 	*optp++ = TCPOPT_MAXSEG;
1263 	*optp++ = TCPOLEN_MAXSEG;
1264 	*optp++ = (mssopt >> 8) & 0xff;
1265 	*optp++ = mssopt & 0xff;
1266 
1267 	if (sc->sc_flags & SCF_WINSCALE) {
1268 		*((u_int32_t *)optp) = htonl(TCPOPT_NOP << 24 |
1269 		    TCPOPT_WINDOW << 16 | TCPOLEN_WINDOW << 8 |
1270 		    sc->sc_request_r_scale);
1271 		optp += 4;
1272 	}
1273 
1274 	if (sc->sc_flags & SCF_TIMESTAMP) {
1275 		u_int32_t *lp = (u_int32_t *)(optp);
1276 
1277 		/* Form timestamp option as shown in appendix A of RFC 1323. */
1278 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
1279 		*lp++ = htonl(ticks);
1280 		*lp   = htonl(sc->sc_tsrecent);
1281 		optp += TCPOLEN_TSTAMP_APPA;
1282 	}
1283 
1284 #ifdef TCP_SIGNATURE
1285 	/*
1286 	 * Handle TCP-MD5 passive opener response.
1287 	 */
1288 	if (sc->sc_flags & SCF_SIGNATURE) {
1289 		u_int8_t *bp = optp;
1290 		int i;
1291 
1292 		*bp++ = TCPOPT_SIGNATURE;
1293 		*bp++ = TCPOLEN_SIGNATURE;
1294 		for (i = 0; i < TCP_SIGLEN; i++)
1295 			*bp++ = 0;
1296 		tcpsignature_compute(m, 0, optlen,
1297 				optp + 2, IPSEC_DIR_OUTBOUND);
1298 		*bp++ = TCPOPT_NOP;
1299 		*bp++ = TCPOPT_EOL;
1300 		optp += TCPOLEN_SIGNATURE + 2;
1301 	}
1302 #endif /* TCP_SIGNATURE */
1303 
1304 	if (sc->sc_flags & SCF_SACK_PERMITTED) {
1305 		*((u_int32_t *)optp) = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
1306 		optp += TCPOLEN_SACK_PERMITTED_ALIGNED;
1307 	}
1308 
1309 no_options:
1310 	if (isipv6) {
1311 		struct route_in6 *ro6 = &sc->sc_route6;
1312 
1313 		th->th_sum = 0;
1314 		th->th_sum = in6_cksum(m, IPPROTO_TCP, hlen, tlen - hlen);
1315 		ip6->ip6_hlim = in6_selecthlim(NULL,
1316 		    ro6->ro_rt ? ro6->ro_rt->rt_ifp : NULL);
1317 		error = ip6_output(m, NULL, ro6, 0, NULL, NULL,
1318 				sc->sc_tp->t_inpcb);
1319 	} else {
1320 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1321 				       htons(tlen - hlen + IPPROTO_TCP));
1322 		m->m_pkthdr.csum_flags = CSUM_TCP;
1323 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1324 		m->m_pkthdr.csum_thlen = sizeof(struct tcphdr) + optlen;
1325 		error = ip_output(m, sc->sc_ipopts, &sc->sc_route,
1326 				  IP_DEBUGROUTE, NULL, sc->sc_tp->t_inpcb);
1327 	}
1328 	return (error);
1329 }
1330 
1331 /*
1332  * cookie layers:
1333  *
1334  *	|. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .|
1335  *	| peer iss                                                      |
1336  *	| MD5(laddr,faddr,secret,lport,fport)             |. . . . . . .|
1337  *	|                     0                       |(A)|             |
1338  * (A): peer mss index
1339  */
1340 
1341 /*
1342  * The values below are chosen to minimize the size of the tcp_secret
1343  * table, as well as providing roughly a 16 second lifetime for the cookie.
1344  */
1345 
1346 #define SYNCOOKIE_WNDBITS	5	/* exposed bits for window indexing */
1347 #define SYNCOOKIE_TIMESHIFT	1	/* scale ticks to window time units */
1348 
1349 #define SYNCOOKIE_WNDMASK	((1 << SYNCOOKIE_WNDBITS) - 1)
1350 #define SYNCOOKIE_NSECRETS	(1 << SYNCOOKIE_WNDBITS)
1351 #define SYNCOOKIE_TIMEOUT \
1352     (hz * (1 << SYNCOOKIE_WNDBITS) / (1 << SYNCOOKIE_TIMESHIFT))
1353 #define SYNCOOKIE_DATAMASK	((3 << SYNCOOKIE_WNDBITS) | SYNCOOKIE_WNDMASK)
1354 
1355 static struct {
1356 	u_int32_t	ts_secbits[4];
1357 	u_int		ts_expire;
1358 } tcp_secret[SYNCOOKIE_NSECRETS];
1359 
1360 static int tcp_msstab[] = { 0, 536, 1460, 8960 };
1361 
1362 static MD5_CTX syn_ctx;
1363 
1364 #define MD5Add(v)	MD5Update(&syn_ctx, (u_char *)&v, sizeof(v))
1365 
1366 struct md5_add {
1367 	u_int32_t laddr, faddr;
1368 	u_int32_t secbits[4];
1369 	u_int16_t lport, fport;
1370 };
1371 
1372 #ifdef CTASSERT
1373 CTASSERT(sizeof(struct md5_add) == 28);
1374 #endif
1375 
1376 /*
1377  * Consider the problem of a recreated (and retransmitted) cookie.  If the
1378  * original SYN was accepted, the connection is established.  The second
1379  * SYN is inflight, and if it arrives with an ISN that falls within the
1380  * receive window, the connection is killed.
1381  *
1382  * However, since cookies have other problems, this may not be worth
1383  * worrying about.
1384  */
1385 
1386 static u_int32_t
1387 syncookie_generate(struct syncache *sc)
1388 {
1389 	u_int32_t md5_buffer[4];
1390 	u_int32_t data;
1391 	int idx, i;
1392 	struct md5_add add;
1393 #ifdef INET6
1394 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
1395 #else
1396 	const boolean_t isipv6 = FALSE;
1397 #endif
1398 
1399 	idx = ((ticks << SYNCOOKIE_TIMESHIFT) / hz) & SYNCOOKIE_WNDMASK;
1400 	if (tcp_secret[idx].ts_expire < ticks) {
1401 		for (i = 0; i < 4; i++)
1402 			tcp_secret[idx].ts_secbits[i] = karc4random();
1403 		tcp_secret[idx].ts_expire = ticks + SYNCOOKIE_TIMEOUT;
1404 	}
1405 	for (data = NELEM(tcp_msstab) - 1; data > 0; data--)
1406 		if (tcp_msstab[data] <= sc->sc_peer_mss)
1407 			break;
1408 	data = (data << SYNCOOKIE_WNDBITS) | idx;
1409 	data ^= sc->sc_irs;				/* peer's iss */
1410 	MD5Init(&syn_ctx);
1411 	if (isipv6) {
1412 		MD5Add(sc->sc_inc.inc6_laddr);
1413 		MD5Add(sc->sc_inc.inc6_faddr);
1414 		add.laddr = 0;
1415 		add.faddr = 0;
1416 	} else {
1417 		add.laddr = sc->sc_inc.inc_laddr.s_addr;
1418 		add.faddr = sc->sc_inc.inc_faddr.s_addr;
1419 	}
1420 	add.lport = sc->sc_inc.inc_lport;
1421 	add.fport = sc->sc_inc.inc_fport;
1422 	add.secbits[0] = tcp_secret[idx].ts_secbits[0];
1423 	add.secbits[1] = tcp_secret[idx].ts_secbits[1];
1424 	add.secbits[2] = tcp_secret[idx].ts_secbits[2];
1425 	add.secbits[3] = tcp_secret[idx].ts_secbits[3];
1426 	MD5Add(add);
1427 	MD5Final((u_char *)&md5_buffer, &syn_ctx);
1428 	data ^= (md5_buffer[0] & ~SYNCOOKIE_WNDMASK);
1429 	return (data);
1430 }
1431 
1432 static struct syncache *
1433 syncookie_lookup(struct in_conninfo *inc, struct tcphdr *th, struct socket *so)
1434 {
1435 	u_int32_t md5_buffer[4];
1436 	struct syncache *sc;
1437 	u_int32_t data;
1438 	int wnd, idx;
1439 	struct md5_add add;
1440 
1441 	data = (th->th_ack - 1) ^ (th->th_seq - 1);	/* remove ISS */
1442 	idx = data & SYNCOOKIE_WNDMASK;
1443 	if (tcp_secret[idx].ts_expire < ticks ||
1444 	    sototcpcb(so)->ts_recent + SYNCOOKIE_TIMEOUT < ticks)
1445 		return (NULL);
1446 	MD5Init(&syn_ctx);
1447 #ifdef INET6
1448 	if (inc->inc_isipv6) {
1449 		MD5Add(inc->inc6_laddr);
1450 		MD5Add(inc->inc6_faddr);
1451 		add.laddr = 0;
1452 		add.faddr = 0;
1453 	} else
1454 #endif
1455 	{
1456 		add.laddr = inc->inc_laddr.s_addr;
1457 		add.faddr = inc->inc_faddr.s_addr;
1458 	}
1459 	add.lport = inc->inc_lport;
1460 	add.fport = inc->inc_fport;
1461 	add.secbits[0] = tcp_secret[idx].ts_secbits[0];
1462 	add.secbits[1] = tcp_secret[idx].ts_secbits[1];
1463 	add.secbits[2] = tcp_secret[idx].ts_secbits[2];
1464 	add.secbits[3] = tcp_secret[idx].ts_secbits[3];
1465 	MD5Add(add);
1466 	MD5Final((u_char *)&md5_buffer, &syn_ctx);
1467 	data ^= md5_buffer[0];
1468 	if (data & ~SYNCOOKIE_DATAMASK)
1469 		return (NULL);
1470 	data = data >> SYNCOOKIE_WNDBITS;
1471 
1472 	/*
1473 	 * Fill in the syncache values.
1474 	 * XXX duplicate code from syncache_add
1475 	 */
1476 	sc = kmalloc(sizeof(struct syncache), M_SYNCACHE, M_WAITOK|M_ZERO);
1477 	sc->sc_ipopts = NULL;
1478 	sc->sc_inc.inc_fport = inc->inc_fport;
1479 	sc->sc_inc.inc_lport = inc->inc_lport;
1480 #ifdef INET6
1481 	sc->sc_inc.inc_isipv6 = inc->inc_isipv6;
1482 	if (inc->inc_isipv6) {
1483 		sc->sc_inc.inc6_faddr = inc->inc6_faddr;
1484 		sc->sc_inc.inc6_laddr = inc->inc6_laddr;
1485 		sc->sc_route6.ro_rt = NULL;
1486 	} else
1487 #endif
1488 	{
1489 		sc->sc_inc.inc_faddr = inc->inc_faddr;
1490 		sc->sc_inc.inc_laddr = inc->inc_laddr;
1491 		sc->sc_route.ro_rt = NULL;
1492 	}
1493 	sc->sc_irs = th->th_seq - 1;
1494 	sc->sc_iss = th->th_ack - 1;
1495 	wnd = ssb_space(&so->so_rcv);
1496 	wnd = imax(wnd, 0);
1497 	wnd = imin(wnd, TCP_MAXWIN);
1498 	sc->sc_wnd = wnd;
1499 	sc->sc_flags = 0;
1500 	sc->sc_rxtslot = 0;
1501 	sc->sc_peer_mss = tcp_msstab[data];
1502 	return (sc);
1503 }
1504