xref: /dragonfly/sys/netinet/tcp_syncache.c (revision 8bf5b238)
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 
77 #include <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/kernel.h>
80 #include <sys/sysctl.h>
81 #include <sys/malloc.h>
82 #include <sys/mbuf.h>
83 #include <sys/md5.h>
84 #include <sys/proc.h>		/* for proc0 declaration */
85 #include <sys/random.h>
86 #include <sys/socket.h>
87 #include <sys/socketvar.h>
88 #include <sys/in_cksum.h>
89 
90 #include <sys/msgport2.h>
91 #include <net/netmsg2.h>
92 #include <net/netisr2.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 static int tcp_syncookies = 1;
119 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_RW,
120     &tcp_syncookies, 0,
121     "Use TCP SYN cookies if the syncache overflows");
122 
123 static void	 syncache_drop(struct syncache *, struct syncache_head *);
124 static void	 syncache_free(struct syncache *);
125 static void	 syncache_insert(struct syncache *, struct syncache_head *);
126 static struct syncache *syncache_lookup(struct in_conninfo *,
127 		    struct syncache_head **);
128 static int	 syncache_respond(struct syncache *, struct mbuf *);
129 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
130 		    struct mbuf *);
131 static void	 syncache_timer(void *);
132 static u_int32_t syncookie_generate(struct syncache *);
133 static struct syncache *syncookie_lookup(struct in_conninfo *,
134 		    struct tcphdr *, struct socket *);
135 
136 /*
137  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
138  * 4 retransmits corresponds to a timeout of (3 + 3 + 3 + 3 + 3 == 15) seconds
139  * or (1 + 1 + 2 + 4 + 8 == 16) seconds if RFC6298 is used, the odds are that
140  * the user has given up attempting to connect by then.
141  */
142 #define SYNCACHE_MAXREXMTS		4
143 
144 /* Arbitrary values */
145 #define TCP_SYNCACHE_HASHSIZE		512
146 #define TCP_SYNCACHE_BUCKETLIMIT	30
147 
148 static void syncache_timer_handler(netmsg_t);
149 static int syncache_sysctl_count(SYSCTL_HANDLER_ARGS);
150 
151 struct tcp_syncache {
152 	u_int	hashsize;
153 	u_int	hashmask;
154 	u_int	bucket_limit;
155 	u_int	cache_limit;
156 	u_int	rexmt_limit;
157 	u_int	hash_secret;
158 };
159 static struct tcp_syncache tcp_syncache;
160 
161 struct syncache_timerq {
162 	TAILQ_HEAD(, syncache)	list;
163 	struct callout		timeo;
164 	struct netmsg_base	nm;
165 };
166 
167 struct tcp_syncache_percpu {
168 	struct syncache_head	*hashbase;
169 	u_int			cache_count;
170 	struct syncache_timerq	timerq[SYNCACHE_MAXREXMTS + 1];
171 };
172 
173 static struct tcp_syncache_percpu *tcp_syncache_percpu[MAXCPU];
174 
175 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, "TCP SYN cache");
176 
177 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RD,
178      &tcp_syncache.bucket_limit, 0, "Per-bucket hash limit for syncache");
179 
180 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_RD,
181      &tcp_syncache.cache_limit, 0, "Overall entry limit for syncache");
182 
183 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, count, (CTLTYPE_INT | CTLFLAG_RD),
184     0, 0, syncache_sysctl_count, "I", "Current number of entries in syncache");
185 
186 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RD,
187      &tcp_syncache.hashsize, 0, "Size of TCP syncache hashtable");
188 
189 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_RW,
190      &tcp_syncache.rexmt_limit, 0, "Limit on SYN/ACK retransmissions");
191 
192 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
193 
194 #define SYNCACHE_HASH(inc, mask)					\
195 	((tcp_syncache.hash_secret ^					\
196 	  (inc)->inc_faddr.s_addr ^					\
197 	  ((inc)->inc_faddr.s_addr >> 16) ^				\
198 	  (inc)->inc_fport ^ (inc)->inc_lport) & mask)
199 
200 #define SYNCACHE_HASH6(inc, mask)					\
201 	((tcp_syncache.hash_secret ^					\
202 	  (inc)->inc6_faddr.s6_addr32[0] ^				\
203 	  (inc)->inc6_faddr.s6_addr32[3] ^				\
204 	  (inc)->inc_fport ^ (inc)->inc_lport) & mask)
205 
206 #define ENDPTS_EQ(a, b) (						\
207 	(a)->ie_fport == (b)->ie_fport &&				\
208 	(a)->ie_lport == (b)->ie_lport &&				\
209 	(a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr &&			\
210 	(a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr			\
211 )
212 
213 #define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0)
214 
215 static __inline int
216 syncache_rto(int slot)
217 {
218 	if (tcp_low_rtobase)
219 		return (TCPTV_RTOBASE * tcp_syn_backoff_low[slot]);
220 	else
221 		return (TCPTV_RTOBASE * tcp_syn_backoff[slot]);
222 }
223 
224 static __inline void
225 syncache_timeout(struct tcp_syncache_percpu *syncache_percpu,
226 		 struct syncache *sc, int slot)
227 {
228 	struct syncache_timerq *tq;
229 	int rto;
230 
231 	KASSERT(slot <= SYNCACHE_MAXREXMTS,
232 	    ("syncache: invalid slot %d", slot));
233 
234 	if (slot > 0) {
235 		/*
236 		 * Record the time that we spent in SYN|ACK
237 		 * retransmition.
238 		 *
239 		 * Needed by RFC3390 and RFC6298.
240 		 */
241 		sc->sc_rxtused += syncache_rto(slot - 1);
242 	}
243 	sc->sc_rxtslot = slot;
244 
245 	rto = syncache_rto(slot);
246 	sc->sc_rxttime = ticks + rto;
247 
248 	tq = &syncache_percpu->timerq[slot];
249 	TAILQ_INSERT_TAIL(&tq->list, sc, sc_timerq);
250 	if (!callout_active(&tq->timeo))
251 		callout_reset(&tq->timeo, rto, syncache_timer, &tq->nm);
252 }
253 
254 static void
255 syncache_free(struct syncache *sc)
256 {
257 	struct rtentry *rt;
258 #ifdef INET6
259 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
260 #else
261 	const boolean_t isipv6 = FALSE;
262 #endif
263 
264 	if (sc->sc_ipopts)
265 		m_free(sc->sc_ipopts);
266 
267 	rt = isipv6 ? sc->sc_route6.ro_rt : sc->sc_route.ro_rt;
268 	if (rt != NULL) {
269 		/*
270 		 * If this is the only reference to a protocol-cloned
271 		 * route, remove it immediately.
272 		 */
273 		if ((rt->rt_flags & (RTF_WASCLONED | RTF_LLINFO)) ==
274 		    RTF_WASCLONED && rt->rt_refcnt == 1) {
275 			rtrequest(RTM_DELETE, rt_key(rt), rt->rt_gateway,
276 				  rt_mask(rt), rt->rt_flags, NULL);
277 		}
278 		RTFREE(rt);
279 	}
280 	kfree(sc, M_SYNCACHE);
281 }
282 
283 static void
284 syncache_init_dispatch(netmsg_t nm)
285 {
286 	struct tcp_syncache_percpu *syncache_percpu;
287 	int i;
288 
289 	ASSERT_NETISR_NCPUS(mycpuid);
290 
291 	syncache_percpu = kmalloc(sizeof(*syncache_percpu), M_SYNCACHE,
292 				  M_WAITOK | M_ZERO);
293 
294 	/* Allocate the hash table. */
295 	syncache_percpu->hashbase = kmalloc(tcp_syncache.hashsize *
296 					    sizeof(struct syncache_head),
297 					    M_SYNCACHE, M_WAITOK | M_ZERO);
298 
299 	/* Initialize the hash buckets. */
300 	for (i = 0; i < tcp_syncache.hashsize; i++) {
301 		struct syncache_head *bucket;
302 
303 		bucket = &syncache_percpu->hashbase[i];
304 		TAILQ_INIT(&bucket->sch_bucket);
305 		bucket->sch_length = 0;
306 	}
307 
308 	for (i = 0; i <= SYNCACHE_MAXREXMTS; i++) {
309 		struct syncache_timerq *tq =
310 		    &syncache_percpu->timerq[i];
311 
312 		/* Initialize the timer queues. */
313 		TAILQ_INIT(&tq->list);
314 		callout_init_mp(&tq->timeo);
315 
316 		netmsg_init(&tq->nm, NULL, &netisr_adone_rport,
317 			    MSGF_PRIORITY, syncache_timer_handler);
318 		tq->nm.lmsg.u.ms_result = i;
319 	}
320 
321 	tcp_syncache_percpu[mycpuid] = syncache_percpu;
322 
323 	netisr_forwardmsg(&nm->base, mycpuid + 1);
324 }
325 
326 void
327 syncache_init(void)
328 {
329 	struct netmsg_base nm;
330 
331 	tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
332 	tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
333 	tcp_syncache.cache_limit =
334 	    tcp_syncache.hashsize * tcp_syncache.bucket_limit;
335 	tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
336 	tcp_syncache.hash_secret = karc4random();
337 
338 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
339 	    &tcp_syncache.hashsize);
340 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
341 	    &tcp_syncache.cache_limit);
342 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
343 	    &tcp_syncache.bucket_limit);
344 	if (!powerof2(tcp_syncache.hashsize)) {
345 		kprintf("WARNING: syncache hash size is not a power of 2.\n");
346 		tcp_syncache.hashsize = 512;	/* safe default */
347 	}
348 	tcp_syncache.hashmask = tcp_syncache.hashsize - 1;
349 
350 	netmsg_init(&nm, NULL, &curthread->td_msgport, 0,
351 	    syncache_init_dispatch);
352 	netisr_domsg_global(&nm);
353 }
354 
355 static void
356 syncache_insert(struct syncache *sc, struct syncache_head *sch)
357 {
358 	struct tcp_syncache_percpu *syncache_percpu;
359 	struct syncache *sc2;
360 	int i;
361 
362 	syncache_percpu = tcp_syncache_percpu[mycpu->gd_cpuid];
363 
364 	/*
365 	 * Make sure that we don't overflow the per-bucket
366 	 * limit or the total cache size limit.
367 	 */
368 	if (sch->sch_length >= tcp_syncache.bucket_limit) {
369 		/*
370 		 * The bucket is full, toss the oldest element.
371 		 */
372 		sc2 = TAILQ_FIRST(&sch->sch_bucket);
373 		if (sc2->sc_tp != NULL)
374 			sc2->sc_tp->ts_recent = ticks;
375 		syncache_drop(sc2, sch);
376 		tcpstat.tcps_sc_bucketoverflow++;
377 	} else if (syncache_percpu->cache_count >= tcp_syncache.cache_limit) {
378 		/*
379 		 * The cache is full.  Toss the oldest entry in the
380 		 * entire cache.  This is the front entry in the
381 		 * first non-empty timer queue with the largest
382 		 * timeout value.
383 		 */
384 		for (i = SYNCACHE_MAXREXMTS; i >= 0; i--) {
385 			sc2 = TAILQ_FIRST(&syncache_percpu->timerq[i].list);
386 			if (sc2 != NULL)
387 				break;
388 		}
389 		if (sc2->sc_tp != NULL)
390 			sc2->sc_tp->ts_recent = ticks;
391 		syncache_drop(sc2, NULL);
392 		tcpstat.tcps_sc_cacheoverflow++;
393 	}
394 
395 	/* Initialize the entry's timer. */
396 	syncache_timeout(syncache_percpu, sc, 0);
397 
398 	/* Put it into the bucket. */
399 	TAILQ_INSERT_TAIL(&sch->sch_bucket, sc, sc_hash);
400 	sch->sch_length++;
401 	syncache_percpu->cache_count++;
402 	tcpstat.tcps_sc_added++;
403 }
404 
405 void
406 syncache_destroy(struct tcpcb *tp, struct tcpcb *tp_inh)
407 {
408 	struct tcp_syncache_percpu *syncache_percpu;
409 	int i;
410 
411 	ASSERT_NETISR_NCPUS(mycpuid);
412 
413 	syncache_percpu = tcp_syncache_percpu[mycpu->gd_cpuid];
414 	for (i = 0; i < tcp_syncache.hashsize; i++) {
415 		struct syncache_head *bucket;
416 		struct syncache *sc;
417 
418 		bucket = &syncache_percpu->hashbase[i];
419 		TAILQ_FOREACH(sc, &bucket->sch_bucket, sc_hash) {
420 			if (sc->sc_tp == tp)
421 				sc->sc_tp = tp_inh;
422 		}
423 	}
424 }
425 
426 static void
427 syncache_drop(struct syncache *sc, struct syncache_head *sch)
428 {
429 	struct tcp_syncache_percpu *syncache_percpu;
430 #ifdef INET6
431 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
432 #else
433 	const boolean_t isipv6 = FALSE;
434 #endif
435 
436 	syncache_percpu = tcp_syncache_percpu[mycpu->gd_cpuid];
437 
438 	if (sch == NULL) {
439 		if (isipv6) {
440 			sch = &syncache_percpu->hashbase[
441 			    SYNCACHE_HASH6(&sc->sc_inc, tcp_syncache.hashmask)];
442 		} else {
443 			sch = &syncache_percpu->hashbase[
444 			    SYNCACHE_HASH(&sc->sc_inc, tcp_syncache.hashmask)];
445 		}
446 	}
447 
448 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
449 	sch->sch_length--;
450 	syncache_percpu->cache_count--;
451 
452 	/*
453 	 * Cleanup
454 	 */
455 	sc->sc_tp = NULL;
456 
457 	/*
458 	 * Remove the entry from the syncache timer/timeout queue.  Note
459 	 * that we do not try to stop any running timer since we do not know
460 	 * whether the timer's message is in-transit or not.  Since timeouts
461 	 * are fairly long, taking an unneeded callout does not detrimentally
462 	 * effect performance.
463 	 */
464 	TAILQ_REMOVE(&syncache_percpu->timerq[sc->sc_rxtslot].list, sc,
465 	    sc_timerq);
466 
467 	syncache_free(sc);
468 }
469 
470 /*
471  * Place a timeout message on the TCP thread's message queue.
472  * This routine runs in soft interrupt context.
473  *
474  * An invariant is for this routine to be called, the callout must
475  * have been active.  Note that the callout is not deactivated until
476  * after the message has been processed in syncache_timer_handler() below.
477  */
478 static void
479 syncache_timer(void *p)
480 {
481 	struct netmsg_base *msg = p;
482 
483 	KKASSERT(mycpuid < netisr_ncpus);
484 
485 	crit_enter();
486 	if (msg->lmsg.ms_flags & MSGF_DONE)
487 		netisr_sendmsg_oncpu(msg);
488 	crit_exit();
489 }
490 
491 /*
492  * Service a timer message queued by timer expiration.
493  * This routine runs in the TCP protocol thread.
494  *
495  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
496  * If we have retransmitted an entry the maximum number of times, expire it.
497  *
498  * When we finish processing timed-out entries, we restart the timer if there
499  * are any entries still on the queue and deactivate it otherwise.  Only after
500  * a timer has been deactivated here can it be restarted by syncache_timeout().
501  */
502 static void
503 syncache_timer_handler(netmsg_t msg)
504 {
505 	struct tcp_syncache_percpu *syncache_percpu;
506 	struct syncache *nsc;
507 	struct syncache_timerq *tq;
508 	int slot;
509 
510 	ASSERT_NETISR_NCPUS(mycpuid);
511 
512 	/* Reply ASAP. */
513 	crit_enter();
514 	netisr_replymsg(&msg->base, 0);
515 	crit_exit();
516 
517 	syncache_percpu = tcp_syncache_percpu[mycpu->gd_cpuid];
518 
519 	slot = msg->lmsg.u.ms_result;
520 	KASSERT(slot <= SYNCACHE_MAXREXMTS,
521 	    ("syncache: invalid slot %d", slot));
522 	tq = &syncache_percpu->timerq[slot];
523 
524 	nsc = TAILQ_FIRST(&tq->list);
525 	while (nsc != NULL) {
526 		struct syncache *sc;
527 
528 		if (ticks < nsc->sc_rxttime)
529 			break;	/* finished because timerq sorted by time */
530 
531 		sc = nsc;
532 		if (sc->sc_tp == NULL) {
533 			nsc = TAILQ_NEXT(sc, sc_timerq);
534 			syncache_drop(sc, NULL);
535 			tcpstat.tcps_sc_stale++;
536 			continue;
537 		}
538 		if (slot == SYNCACHE_MAXREXMTS ||
539 		    slot >= tcp_syncache.rexmt_limit ||
540 		    sc->sc_tp->t_inpcb->inp_gencnt != sc->sc_inp_gencnt) {
541 			nsc = TAILQ_NEXT(sc, sc_timerq);
542 			syncache_drop(sc, NULL);
543 			tcpstat.tcps_sc_stale++;
544 			continue;
545 		}
546 		/*
547 		 * syncache_respond() may call back into the syncache to
548 		 * to modify another entry, so do not obtain the next
549 		 * entry on the timer chain until it has completed.
550 		 */
551 		syncache_respond(sc, NULL);
552 		tcpstat.tcps_sc_retransmitted++;
553 		nsc = TAILQ_NEXT(sc, sc_timerq);
554 		TAILQ_REMOVE(&tq->list, sc, sc_timerq);
555 		syncache_timeout(syncache_percpu, sc, slot + 1);
556 	}
557 
558 	if (nsc != NULL) {
559 		callout_reset(&tq->timeo, nsc->sc_rxttime - ticks,
560 		    syncache_timer, &tq->nm);
561 	} else {
562 		callout_deactivate(&tq->timeo);
563 	}
564 }
565 
566 /*
567  * Find an entry in the syncache.
568  */
569 static struct syncache *
570 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
571 {
572 	struct tcp_syncache_percpu *syncache_percpu;
573 	struct syncache *sc;
574 	struct syncache_head *sch;
575 
576 	syncache_percpu = tcp_syncache_percpu[mycpu->gd_cpuid];
577 #ifdef INET6
578 	if (inc->inc_isipv6) {
579 		sch = &syncache_percpu->hashbase[
580 		    SYNCACHE_HASH6(inc, tcp_syncache.hashmask)];
581 		*schp = sch;
582 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
583 			if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
584 				return (sc);
585 	} else
586 #endif
587 	{
588 		sch = &syncache_percpu->hashbase[
589 		    SYNCACHE_HASH(inc, tcp_syncache.hashmask)];
590 		*schp = sch;
591 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
592 #ifdef INET6
593 			if (sc->sc_inc.inc_isipv6)
594 				continue;
595 #endif
596 			if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
597 				return (sc);
598 		}
599 	}
600 	return (NULL);
601 }
602 
603 /*
604  * This function is called when we get a RST for a
605  * non-existent connection, so that we can see if the
606  * connection is in the syn cache.  If it is, zap it.
607  */
608 void
609 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
610 {
611 	struct syncache *sc;
612 	struct syncache_head *sch;
613 
614 	ASSERT_NETISR_NCPUS(mycpuid);
615 
616 	sc = syncache_lookup(inc, &sch);
617 	if (sc == NULL) {
618 		return;
619 	}
620 	/*
621 	 * If the RST bit is set, check the sequence number to see
622 	 * if this is a valid reset segment.
623 	 * RFC 793 page 37:
624 	 *   In all states except SYN-SENT, all reset (RST) segments
625 	 *   are validated by checking their SEQ-fields.  A reset is
626 	 *   valid if its sequence number is in the window.
627 	 *
628 	 *   The sequence number in the reset segment is normally an
629 	 *   echo of our outgoing acknowlegement numbers, but some hosts
630 	 *   send a reset with the sequence number at the rightmost edge
631 	 *   of our receive window, and we have to handle this case.
632 	 */
633 	if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
634 	    SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
635 		syncache_drop(sc, sch);
636 		tcpstat.tcps_sc_reset++;
637 	}
638 }
639 
640 void
641 syncache_badack(struct in_conninfo *inc)
642 {
643 	struct syncache *sc;
644 	struct syncache_head *sch;
645 
646 	ASSERT_NETISR_NCPUS(mycpuid);
647 
648 	sc = syncache_lookup(inc, &sch);
649 	if (sc != NULL) {
650 		syncache_drop(sc, sch);
651 		tcpstat.tcps_sc_badack++;
652 	}
653 }
654 
655 void
656 syncache_unreach(struct in_conninfo *inc, const struct tcphdr *th)
657 {
658 	struct syncache *sc;
659 	struct syncache_head *sch;
660 
661 	ASSERT_NETISR_NCPUS(mycpuid);
662 
663 	/* we are called at splnet() here */
664 	sc = syncache_lookup(inc, &sch);
665 	if (sc == NULL)
666 		return;
667 
668 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
669 	if (ntohl(th->th_seq) != sc->sc_iss)
670 		return;
671 
672 	/*
673 	 * If we've rertransmitted 3 times and this is our second error,
674 	 * we remove the entry.  Otherwise, we allow it to continue on.
675 	 * This prevents us from incorrectly nuking an entry during a
676 	 * spurious network outage.
677 	 *
678 	 * See tcp_notify().
679 	 */
680 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxtslot < 3) {
681 		sc->sc_flags |= SCF_UNREACH;
682 		return;
683 	}
684 	syncache_drop(sc, sch);
685 	tcpstat.tcps_sc_unreach++;
686 }
687 
688 /*
689  * Build a new TCP socket structure from a syncache entry.
690  *
691  * This is called from the context of the SYN+ACK
692  */
693 static struct socket *
694 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
695 {
696 	struct inpcb *inp = NULL, *linp;
697 	struct socket *so;
698 	struct tcpcb *tp, *ltp;
699 	lwkt_port_t port;
700 #ifdef INET6
701 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
702 #else
703 	const boolean_t isipv6 = FALSE;
704 #endif
705 	struct sockaddr_in sin_faddr;
706 	struct sockaddr_in6 sin6_faddr;
707 	struct sockaddr *faddr;
708 
709 	KASSERT(m->m_flags & M_HASH, ("mbuf has no hash"));
710 
711 	if (isipv6) {
712 		faddr = (struct sockaddr *)&sin6_faddr;
713 		sin6_faddr.sin6_family = AF_INET6;
714 		sin6_faddr.sin6_len = sizeof(sin6_faddr);
715 		sin6_faddr.sin6_addr = sc->sc_inc.inc6_faddr;
716 		sin6_faddr.sin6_port = sc->sc_inc.inc_fport;
717 		sin6_faddr.sin6_flowinfo = sin6_faddr.sin6_scope_id = 0;
718 	} else {
719 		faddr = (struct sockaddr *)&sin_faddr;
720 		sin_faddr.sin_family = AF_INET;
721 		sin_faddr.sin_len = sizeof(sin_faddr);
722 		sin_faddr.sin_addr = sc->sc_inc.inc_faddr;
723 		sin_faddr.sin_port = sc->sc_inc.inc_fport;
724 		bzero(sin_faddr.sin_zero, sizeof(sin_faddr.sin_zero));
725 	}
726 
727 	/*
728 	 * Ok, create the full blown connection, and set things up
729 	 * as they would have been set up if we had created the
730 	 * connection when the SYN arrived.  If we can't create
731 	 * the connection, abort it.
732 	 *
733 	 * Set the protocol processing port for the socket to the current
734 	 * port (that the connection came in on).
735 	 *
736 	 * NOTE:
737 	 * We don't keep a reference on the new socket, since its
738 	 * destruction will run in this thread (netisrN); there is no
739 	 * race here.
740 	 */
741 	so = sonewconn_faddr(lso, SS_ISCONNECTED, faddr,
742 	    FALSE /* don't ref */);
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 		KASSERT(INP_ISIPV4(inp), ("not inet pcb"));
761 		inp->inp_laddr = sc->sc_inc.inc_laddr;
762 	}
763 	inp->inp_lport = sc->sc_inc.inc_lport;
764 
765 	linp = lso->so_pcb;
766 	ltp = intotcpcb(linp);
767 
768 	tcp_pcbport_insert(ltp, inp);
769 
770 	if (isipv6) {
771 		struct in6_addr laddr6;
772 		/*
773 		 * Inherit socket options from the listening socket.
774 		 * Note that in6p_inputopts are not (and should not be)
775 		 * copied, since it stores previously received options and is
776 		 * used to detect if each new option is different than the
777 		 * previous one and hence should be passed to a user.
778 		 * If we copied in6p_inputopts, a user would not be able to
779 		 * receive options just after calling the accept system call.
780 		 */
781 		inp->inp_flags |= linp->inp_flags & INP_CONTROLOPTS;
782 		if (linp->in6p_outputopts)
783 			inp->in6p_outputopts =
784 			    ip6_copypktopts(linp->in6p_outputopts, M_INTWAIT);
785 		inp->in6p_route = sc->sc_route6;
786 		sc->sc_route6.ro_rt = NULL;
787 
788 		laddr6 = inp->in6p_laddr;
789 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
790 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
791 		if (in6_pcbconnect(inp, faddr, &thread0)) {
792 			inp->in6p_laddr = laddr6;
793 			goto abort;
794 		}
795 		port = tcp6_addrport();
796 	} else {
797 		struct in_addr laddr;
798 
799 		inp->inp_options = ip_srcroute(m);
800 		if (inp->inp_options == NULL) {
801 			inp->inp_options = sc->sc_ipopts;
802 			sc->sc_ipopts = NULL;
803 		}
804 		inp->inp_route = sc->sc_route;
805 		sc->sc_route.ro_rt = NULL;
806 
807 		laddr = inp->inp_laddr;
808 		if (inp->inp_laddr.s_addr == INADDR_ANY)
809 			inp->inp_laddr = sc->sc_inc.inc_laddr;
810 		if (in_pcbconnect(inp, faddr, &thread0)) {
811 			inp->inp_laddr = laddr;
812 			goto abort;
813 		}
814 
815 		inp->inp_flags |= INP_HASH;
816 		inp->inp_hashval = m->m_pkthdr.hash;
817 		port = netisr_hashport(inp->inp_hashval);
818 	}
819 
820 	/*
821 	 * The current port should be in the context of the SYN+ACK and
822 	 * so should match the tcp address port.
823 	 */
824 	KASSERT(port == &curthread->td_msgport,
825 	    ("TCP PORT MISMATCH %p vs %p\n", port, &curthread->td_msgport));
826 
827 	tp = intotcpcb(inp);
828 	TCP_STATE_CHANGE(tp, TCPS_SYN_RECEIVED);
829 	tp->iss = sc->sc_iss;
830 	tp->irs = sc->sc_irs;
831 	tcp_rcvseqinit(tp);
832 	tcp_sendseqinit(tp);
833 	tp->snd_wnd = sc->sc_sndwnd;
834 	tp->snd_wl1 = sc->sc_irs;
835 	tp->rcv_up = sc->sc_irs + 1;
836 	tp->rcv_wnd = sc->sc_wnd;
837 	tp->rcv_adv += tp->rcv_wnd;
838 
839 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH | TF_NODELAY);
840 	if (sc->sc_flags & SCF_NOOPT)
841 		tp->t_flags |= TF_NOOPT;
842 	if (sc->sc_flags & SCF_WINSCALE) {
843 		tp->t_flags |= TF_REQ_SCALE | TF_RCVD_SCALE;
844 		tp->snd_scale = sc->sc_requested_s_scale;
845 		tp->request_r_scale = sc->sc_request_r_scale;
846 	}
847 	if (sc->sc_flags & SCF_TIMESTAMP) {
848 		tp->t_flags |= TF_REQ_TSTMP | TF_RCVD_TSTMP;
849 		tp->ts_recent = sc->sc_tsrecent;
850 		tp->ts_recent_age = ticks;
851 	}
852 	if (sc->sc_flags & SCF_SACK_PERMITTED)
853 		tp->t_flags |= TF_SACK_PERMITTED;
854 
855 #ifdef TCP_SIGNATURE
856 	if (sc->sc_flags & SCF_SIGNATURE)
857 		tp->t_flags |= TF_SIGNATURE;
858 #endif /* TCP_SIGNATURE */
859 
860 	tp->t_rxtsyn = sc->sc_rxtused;
861 	tcp_rmx_init(tp, sc->sc_peer_mss);
862 
863 	/*
864 	 * Inherit some properties from the listen socket
865 	 */
866 	tp->t_keepinit = ltp->t_keepinit;
867 	tp->t_keepidle = ltp->t_keepidle;
868 	tp->t_keepintvl = ltp->t_keepintvl;
869 	tp->t_keepcnt = ltp->t_keepcnt;
870 	tp->t_maxidle = ltp->t_maxidle;
871 
872 	tcp_create_timermsg(tp, port);
873 	tcp_callout_reset(tp, tp->tt_keep, tp->t_keepinit, tcp_timer_keep);
874 
875 	tcpstat.tcps_accepts++;
876 	return (so);
877 
878 abort:
879 	if (so != NULL)
880 		soabort_direct(so);
881 	return (NULL);
882 }
883 
884 /*
885  * This function gets called when we receive an ACK for a
886  * socket in the LISTEN state.  We look up the connection
887  * in the syncache, and if its there, we pull it out of
888  * the cache and turn it into a full-blown connection in
889  * the SYN-RECEIVED state.
890  */
891 int
892 syncache_expand(struct in_conninfo *inc, struct tcphdr *th, struct socket **sop,
893 		struct mbuf *m)
894 {
895 	struct syncache *sc;
896 	struct syncache_head *sch;
897 	struct socket *so;
898 
899 	ASSERT_NETISR_NCPUS(mycpuid);
900 
901 	sc = syncache_lookup(inc, &sch);
902 	if (sc == NULL) {
903 		/*
904 		 * There is no syncache entry, so see if this ACK is
905 		 * a returning syncookie.  To do this, first:
906 		 *  A. See if this socket has had a syncache entry dropped in
907 		 *     the past.  We don't want to accept a bogus syncookie
908 		 *     if we've never received a SYN.
909 		 *  B. check that the syncookie is valid.  If it is, then
910 		 *     cobble up a fake syncache entry, and return.
911 		 */
912 		if (!tcp_syncookies)
913 			return (0);
914 		sc = syncookie_lookup(inc, th, *sop);
915 		if (sc == NULL)
916 			return (0);
917 		sch = NULL;
918 		tcpstat.tcps_sc_recvcookie++;
919 	}
920 
921 	/*
922 	 * If seg contains an ACK, but not for our SYN/ACK, send a RST.
923 	 */
924 	if (th->th_ack != sc->sc_iss + 1)
925 		return (0);
926 
927 	so = syncache_socket(sc, *sop, m);
928 	if (so == NULL) {
929 #if 0
930 resetandabort:
931 		/* XXXjlemon check this - is this correct? */
932 		tcp_respond(NULL, m, m, th,
933 		    th->th_seq + tlen, (tcp_seq)0, TH_RST | TH_ACK);
934 #endif
935 		m_freem(m);			/* XXX only needed for above */
936 		tcpstat.tcps_sc_aborted++;
937 	} else {
938 		tcpstat.tcps_sc_completed++;
939 	}
940 	if (sch == NULL)
941 		syncache_free(sc);
942 	else
943 		syncache_drop(sc, sch);
944 	*sop = so;
945 	return (1);
946 }
947 
948 /*
949  * Given a LISTEN socket and an inbound SYN request, add
950  * this to the syn cache, and send back a segment:
951  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
952  * to the source.
953  *
954  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
955  * Doing so would require that we hold onto the data and deliver it
956  * to the application.  However, if we are the target of a SYN-flood
957  * DoS attack, an attacker could send data which would eventually
958  * consume all available buffer space if it were ACKed.  By not ACKing
959  * the data, we avoid this DoS scenario.
960  */
961 int
962 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
963 	     struct socket *so, struct mbuf *m)
964 {
965 	struct tcp_syncache_percpu *syncache_percpu;
966 	struct tcpcb *tp;
967 	struct syncache *sc = NULL;
968 	struct syncache_head *sch;
969 	struct mbuf *ipopts = NULL;
970 	int win;
971 
972 	ASSERT_NETISR_NCPUS(mycpuid);
973 	KASSERT(m->m_flags & M_HASH, ("mbuf has no hash"));
974 
975 	syncache_percpu = tcp_syncache_percpu[mycpu->gd_cpuid];
976 	tp = sototcpcb(so);
977 
978 	/*
979 	 * Remember the IP options, if any.
980 	 */
981 #ifdef INET6
982 	if (!inc->inc_isipv6)
983 #endif
984 		ipopts = ip_srcroute(m);
985 
986 	/*
987 	 * See if we already have an entry for this connection.
988 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
989 	 *
990 	 * XXX
991 	 * The syncache should be re-initialized with the contents
992 	 * of the new SYN which may have different options.
993 	 */
994 	sc = syncache_lookup(inc, &sch);
995 	if (sc != NULL) {
996 		KASSERT(sc->sc_flags & SCF_HASH, ("syncache has no hash"));
997 		KASSERT(sc->sc_hashval == m->m_pkthdr.hash,
998 		    ("syncache/mbuf hash mismatches"));
999 
1000 		tcpstat.tcps_sc_dupsyn++;
1001 		if (ipopts) {
1002 			/*
1003 			 * If we were remembering a previous source route,
1004 			 * forget it and use the new one we've been given.
1005 			 */
1006 			if (sc->sc_ipopts)
1007 				m_free(sc->sc_ipopts);
1008 			sc->sc_ipopts = ipopts;
1009 		}
1010 		/*
1011 		 * Update timestamp if present.
1012 		 */
1013 		if (sc->sc_flags & SCF_TIMESTAMP)
1014 			sc->sc_tsrecent = to->to_tsval;
1015 
1016 		/* Just update the TOF_SACK_PERMITTED for now. */
1017 		if (tcp_do_sack && (to->to_flags & TOF_SACK_PERMITTED))
1018 			sc->sc_flags |= SCF_SACK_PERMITTED;
1019 		else
1020 			sc->sc_flags &= ~SCF_SACK_PERMITTED;
1021 
1022 		/* Update initial send window */
1023 		sc->sc_sndwnd = th->th_win;
1024 
1025 		/*
1026 		 * PCB may have changed, pick up new values.
1027 		 */
1028 		sc->sc_tp = tp;
1029 		sc->sc_inp_gencnt = tp->t_inpcb->inp_gencnt;
1030 		if (syncache_respond(sc, m) == 0) {
1031 			TAILQ_REMOVE(
1032 			    &syncache_percpu->timerq[sc->sc_rxtslot].list,
1033 			    sc, sc_timerq);
1034 			syncache_timeout(syncache_percpu, sc, sc->sc_rxtslot);
1035 			tcpstat.tcps_sndacks++;
1036 			tcpstat.tcps_sndtotal++;
1037 		}
1038 		return (1);
1039 	}
1040 
1041 	/*
1042 	 * Fill in the syncache values.
1043 	 */
1044 	sc = kmalloc(sizeof(struct syncache), M_SYNCACHE, M_WAITOK|M_ZERO);
1045 	sc->sc_inp_gencnt = tp->t_inpcb->inp_gencnt;
1046 	sc->sc_ipopts = ipopts;
1047 	sc->sc_inc.inc_fport = inc->inc_fport;
1048 	sc->sc_inc.inc_lport = inc->inc_lport;
1049 	sc->sc_tp = tp;
1050 #ifdef INET6
1051 	sc->sc_inc.inc_isipv6 = inc->inc_isipv6;
1052 	if (inc->inc_isipv6) {
1053 		sc->sc_inc.inc6_faddr = inc->inc6_faddr;
1054 		sc->sc_inc.inc6_laddr = inc->inc6_laddr;
1055 		sc->sc_route6.ro_rt = NULL;
1056 	} else
1057 #endif
1058 	{
1059 		sc->sc_inc.inc_faddr = inc->inc_faddr;
1060 		sc->sc_inc.inc_laddr = inc->inc_laddr;
1061 		sc->sc_route.ro_rt = NULL;
1062 	}
1063 	sc->sc_irs = th->th_seq;
1064 	sc->sc_flags = SCF_HASH;
1065 	sc->sc_hashval = m->m_pkthdr.hash;
1066 	sc->sc_peer_mss = to->to_flags & TOF_MSS ? to->to_mss : 0;
1067 	if (tcp_syncookies)
1068 		sc->sc_iss = syncookie_generate(sc);
1069 	else
1070 		sc->sc_iss = karc4random();
1071 
1072 	/* Initial receive window: clip ssb_space to [0 .. TCP_MAXWIN] */
1073 	win = ssb_space(&so->so_rcv);
1074 	win = imax(win, 0);
1075 	win = imin(win, TCP_MAXWIN);
1076 	sc->sc_wnd = win;
1077 
1078 	if (tcp_do_rfc1323) {
1079 		/*
1080 		 * A timestamp received in a SYN makes
1081 		 * it ok to send timestamp requests and replies.
1082 		 */
1083 		if (to->to_flags & TOF_TS) {
1084 			sc->sc_tsrecent = to->to_tsval;
1085 			sc->sc_flags |= SCF_TIMESTAMP;
1086 		}
1087 		if (to->to_flags & TOF_SCALE) {
1088 			int wscale = TCP_MIN_WINSHIFT;
1089 
1090 			/* Compute proper scaling value from buffer space */
1091 			while (wscale < TCP_MAX_WINSHIFT &&
1092 			    (TCP_MAXWIN << wscale) < so->so_rcv.ssb_hiwat) {
1093 				wscale++;
1094 			}
1095 			sc->sc_request_r_scale = wscale;
1096 			sc->sc_requested_s_scale = to->to_requested_s_scale;
1097 			sc->sc_flags |= SCF_WINSCALE;
1098 		}
1099 	}
1100 	if (tcp_do_sack && (to->to_flags & TOF_SACK_PERMITTED))
1101 		sc->sc_flags |= SCF_SACK_PERMITTED;
1102 	if (tp->t_flags & TF_NOOPT)
1103 		sc->sc_flags = SCF_NOOPT;
1104 #ifdef TCP_SIGNATURE
1105 	/*
1106 	 * If listening socket requested TCP digests, and received SYN
1107 	 * contains the option, flag this in the syncache so that
1108 	 * syncache_respond() will do the right thing with the SYN+ACK.
1109 	 * XXX Currently we always record the option by default and will
1110 	 * attempt to use it in syncache_respond().
1111 	 */
1112 	if (to->to_flags & TOF_SIGNATURE)
1113 		sc->sc_flags = SCF_SIGNATURE;
1114 #endif /* TCP_SIGNATURE */
1115 	sc->sc_sndwnd = th->th_win;
1116 
1117 	if (syncache_respond(sc, m) == 0) {
1118 		syncache_insert(sc, sch);
1119 		tcpstat.tcps_sndacks++;
1120 		tcpstat.tcps_sndtotal++;
1121 	} else {
1122 		syncache_free(sc);
1123 		tcpstat.tcps_sc_dropped++;
1124 	}
1125 	return (1);
1126 }
1127 
1128 static int
1129 syncache_respond(struct syncache *sc, struct mbuf *m)
1130 {
1131 	u_int8_t *optp;
1132 	int optlen, error;
1133 	u_int16_t tlen, hlen, mssopt;
1134 	struct ip *ip = NULL;
1135 	struct rtentry *rt;
1136 	struct tcphdr *th;
1137 	struct ip6_hdr *ip6 = NULL;
1138 #ifdef INET6
1139 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
1140 #else
1141 	const boolean_t isipv6 = FALSE;
1142 #endif
1143 
1144 	if (isipv6) {
1145 		rt = tcp_rtlookup6(&sc->sc_inc);
1146 		if (rt != NULL)
1147 			mssopt = rt->rt_ifp->if_mtu -
1148 			     (sizeof(struct ip6_hdr) + sizeof(struct tcphdr));
1149 		else
1150 			mssopt = tcp_v6mssdflt;
1151 		hlen = sizeof(struct ip6_hdr);
1152 	} else {
1153 		rt = tcp_rtlookup(&sc->sc_inc);
1154 		if (rt != NULL)
1155 			mssopt = rt->rt_ifp->if_mtu -
1156 			     (sizeof(struct ip) + sizeof(struct tcphdr));
1157 		else
1158 			mssopt = tcp_mssdflt;
1159 		hlen = sizeof(struct ip);
1160 	}
1161 
1162 	/* Compute the size of the TCP options. */
1163 	if (sc->sc_flags & SCF_NOOPT) {
1164 		optlen = 0;
1165 	} else {
1166 		optlen = TCPOLEN_MAXSEG +
1167 		    ((sc->sc_flags & SCF_WINSCALE) ? 4 : 0) +
1168 		    ((sc->sc_flags & SCF_TIMESTAMP) ? TCPOLEN_TSTAMP_APPA : 0) +
1169 		    ((sc->sc_flags & SCF_SACK_PERMITTED) ?
1170 			TCPOLEN_SACK_PERMITTED_ALIGNED : 0);
1171 #ifdef TCP_SIGNATURE
1172 		optlen += ((sc->sc_flags & SCF_SIGNATURE) ?
1173 		    (TCPOLEN_SIGNATURE + 2) : 0);
1174 #endif /* TCP_SIGNATURE */
1175 	}
1176 	tlen = hlen + sizeof(struct tcphdr) + optlen;
1177 
1178 	/*
1179 	 * XXX
1180 	 * assume that the entire packet will fit in a header mbuf
1181 	 */
1182 	KASSERT(max_linkhdr + tlen <= MHLEN, ("syncache: mbuf too small"));
1183 
1184 	/*
1185 	 * XXX shouldn't this reuse the mbuf if possible ?
1186 	 * Create the IP+TCP header from scratch.
1187 	 */
1188 	if (m)
1189 		m_freem(m);
1190 
1191 	m = m_gethdr(M_NOWAIT, MT_HEADER);
1192 	if (m == NULL)
1193 		return (ENOBUFS);
1194 	m->m_data += max_linkhdr;
1195 	m->m_len = tlen;
1196 	m->m_pkthdr.len = tlen;
1197 	m->m_pkthdr.rcvif = NULL;
1198 	if (tcp_prio_synack)
1199 		m->m_flags |= M_PRIO;
1200 
1201 	if (isipv6) {
1202 		ip6 = mtod(m, struct ip6_hdr *);
1203 		ip6->ip6_vfc = IPV6_VERSION;
1204 		ip6->ip6_nxt = IPPROTO_TCP;
1205 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1206 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1207 		ip6->ip6_plen = htons(tlen - hlen);
1208 		/* ip6_hlim is set after checksum */
1209 		/* ip6_flow = ??? */
1210 
1211 		th = (struct tcphdr *)(ip6 + 1);
1212 	} else {
1213 		ip = mtod(m, struct ip *);
1214 		ip->ip_v = IPVERSION;
1215 		ip->ip_hl = sizeof(struct ip) >> 2;
1216 		ip->ip_len = tlen;
1217 		ip->ip_id = 0;
1218 		ip->ip_off = 0;
1219 		ip->ip_sum = 0;
1220 		ip->ip_p = IPPROTO_TCP;
1221 		ip->ip_src = sc->sc_inc.inc_laddr;
1222 		ip->ip_dst = sc->sc_inc.inc_faddr;
1223 		ip->ip_ttl = sc->sc_tp->t_inpcb->inp_ip_ttl;   /* XXX */
1224 		ip->ip_tos = sc->sc_tp->t_inpcb->inp_ip_tos;   /* XXX */
1225 
1226 		/*
1227 		 * See if we should do MTU discovery.  Route lookups are
1228 		 * expensive, so we will only unset the DF bit if:
1229 		 *
1230 		 *	1) path_mtu_discovery is disabled
1231 		 *	2) the SCF_UNREACH flag has been set
1232 		 */
1233 		if (path_mtu_discovery
1234 		    && ((sc->sc_flags & SCF_UNREACH) == 0)) {
1235 		       ip->ip_off |= IP_DF;
1236 		}
1237 
1238 		th = (struct tcphdr *)(ip + 1);
1239 	}
1240 	th->th_sport = sc->sc_inc.inc_lport;
1241 	th->th_dport = sc->sc_inc.inc_fport;
1242 
1243 	th->th_seq = htonl(sc->sc_iss);
1244 	th->th_ack = htonl(sc->sc_irs + 1);
1245 	th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1246 	th->th_x2 = 0;
1247 	th->th_flags = TH_SYN | TH_ACK;
1248 	th->th_win = htons(sc->sc_wnd);
1249 	th->th_urp = 0;
1250 
1251 	/* Tack on the TCP options. */
1252 	if (optlen == 0)
1253 		goto no_options;
1254 	optp = (u_int8_t *)(th + 1);
1255 	*optp++ = TCPOPT_MAXSEG;
1256 	*optp++ = TCPOLEN_MAXSEG;
1257 	*optp++ = (mssopt >> 8) & 0xff;
1258 	*optp++ = mssopt & 0xff;
1259 
1260 	if (sc->sc_flags & SCF_WINSCALE) {
1261 		*((u_int32_t *)optp) = htonl(TCPOPT_NOP << 24 |
1262 		    TCPOPT_WINDOW << 16 | TCPOLEN_WINDOW << 8 |
1263 		    sc->sc_request_r_scale);
1264 		optp += 4;
1265 	}
1266 
1267 	if (sc->sc_flags & SCF_TIMESTAMP) {
1268 		u_int32_t *lp = (u_int32_t *)(optp);
1269 
1270 		/* Form timestamp option as shown in appendix A of RFC 1323. */
1271 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
1272 		*lp++ = htonl(ticks);
1273 		*lp   = htonl(sc->sc_tsrecent);
1274 		optp += TCPOLEN_TSTAMP_APPA;
1275 	}
1276 
1277 #ifdef TCP_SIGNATURE
1278 	/*
1279 	 * Handle TCP-MD5 passive opener response.
1280 	 */
1281 	if (sc->sc_flags & SCF_SIGNATURE) {
1282 		u_int8_t *bp = optp;
1283 		int i;
1284 
1285 		*bp++ = TCPOPT_SIGNATURE;
1286 		*bp++ = TCPOLEN_SIGNATURE;
1287 		for (i = 0; i < TCP_SIGLEN; i++)
1288 			*bp++ = 0;
1289 		tcpsignature_compute(m, 0, optlen,
1290 				optp + 2, IPSEC_DIR_OUTBOUND);
1291 		*bp++ = TCPOPT_NOP;
1292 		*bp++ = TCPOPT_EOL;
1293 		optp += TCPOLEN_SIGNATURE + 2;
1294 	}
1295 #endif /* TCP_SIGNATURE */
1296 
1297 	if (sc->sc_flags & SCF_SACK_PERMITTED) {
1298 		*((u_int32_t *)optp) = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
1299 		optp += TCPOLEN_SACK_PERMITTED_ALIGNED;
1300 	}
1301 
1302 no_options:
1303 	if (isipv6) {
1304 		struct route_in6 *ro6 = &sc->sc_route6;
1305 
1306 		th->th_sum = 0;
1307 		th->th_sum = in6_cksum(m, IPPROTO_TCP, hlen, tlen - hlen);
1308 		ip6->ip6_hlim = in6_selecthlim(NULL,
1309 		    ro6->ro_rt ? ro6->ro_rt->rt_ifp : NULL);
1310 		error = ip6_output(m, NULL, ro6, 0, NULL, NULL,
1311 				sc->sc_tp->t_inpcb);
1312 	} else {
1313 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1314 				       htons(tlen - hlen + IPPROTO_TCP));
1315 		m->m_pkthdr.csum_flags = CSUM_TCP;
1316 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1317 		m->m_pkthdr.csum_thlen = sizeof(struct tcphdr) + optlen;
1318 		KASSERT(sc->sc_flags & SCF_HASH, ("syncache has no hash"));
1319 		m_sethash(m, sc->sc_hashval);
1320 		error = ip_output(m, sc->sc_ipopts, &sc->sc_route,
1321 				  IP_DEBUGROUTE, NULL, sc->sc_tp->t_inpcb);
1322 	}
1323 	return (error);
1324 }
1325 
1326 /*
1327  * cookie layers:
1328  *
1329  *	|. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .|
1330  *	| peer iss                                                      |
1331  *	| MD5(laddr,faddr,secret,lport,fport)             |. . . . . . .|
1332  *	|                     0                       |(A)|             |
1333  * (A): peer mss index
1334  */
1335 
1336 /*
1337  * The values below are chosen to minimize the size of the tcp_secret
1338  * table, as well as providing roughly a 16 second lifetime for the cookie.
1339  */
1340 
1341 #define SYNCOOKIE_WNDBITS	5	/* exposed bits for window indexing */
1342 #define SYNCOOKIE_TIMESHIFT	1	/* scale ticks to window time units */
1343 
1344 #define SYNCOOKIE_WNDMASK	((1 << SYNCOOKIE_WNDBITS) - 1)
1345 #define SYNCOOKIE_NSECRETS	(1 << SYNCOOKIE_WNDBITS)
1346 #define SYNCOOKIE_TIMEOUT \
1347     (hz * (1 << SYNCOOKIE_WNDBITS) / (1 << SYNCOOKIE_TIMESHIFT))
1348 #define SYNCOOKIE_DATAMASK	((3 << SYNCOOKIE_WNDBITS) | SYNCOOKIE_WNDMASK)
1349 
1350 static struct {
1351 	u_int32_t	ts_secbits[4];
1352 	u_int		ts_expire;
1353 } tcp_secret[SYNCOOKIE_NSECRETS];
1354 
1355 static int tcp_msstab[] = { 0, 536, 1460, 8960 };
1356 
1357 static MD5_CTX syn_ctx;
1358 
1359 #define MD5Add(v)	MD5Update(&syn_ctx, (u_char *)&v, sizeof(v))
1360 
1361 struct md5_add {
1362 	u_int32_t laddr, faddr;
1363 	u_int32_t secbits[4];
1364 	u_int16_t lport, fport;
1365 };
1366 
1367 #ifdef CTASSERT
1368 CTASSERT(sizeof(struct md5_add) == 28);
1369 #endif
1370 
1371 /*
1372  * Consider the problem of a recreated (and retransmitted) cookie.  If the
1373  * original SYN was accepted, the connection is established.  The second
1374  * SYN is inflight, and if it arrives with an ISN that falls within the
1375  * receive window, the connection is killed.
1376  *
1377  * However, since cookies have other problems, this may not be worth
1378  * worrying about.
1379  */
1380 
1381 static u_int32_t
1382 syncookie_generate(struct syncache *sc)
1383 {
1384 	u_int32_t md5_buffer[4];
1385 	u_int32_t data;
1386 	int idx, i;
1387 	struct md5_add add;
1388 #ifdef INET6
1389 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
1390 #else
1391 	const boolean_t isipv6 = FALSE;
1392 #endif
1393 
1394 	idx = ((ticks << SYNCOOKIE_TIMESHIFT) / hz) & SYNCOOKIE_WNDMASK;
1395 	if (tcp_secret[idx].ts_expire < ticks) {
1396 		for (i = 0; i < 4; i++)
1397 			tcp_secret[idx].ts_secbits[i] = karc4random();
1398 		tcp_secret[idx].ts_expire = ticks + SYNCOOKIE_TIMEOUT;
1399 	}
1400 	for (data = NELEM(tcp_msstab) - 1; data > 0; data--)
1401 		if (tcp_msstab[data] <= sc->sc_peer_mss)
1402 			break;
1403 	data = (data << SYNCOOKIE_WNDBITS) | idx;
1404 	data ^= sc->sc_irs;				/* peer's iss */
1405 	MD5Init(&syn_ctx);
1406 	if (isipv6) {
1407 		MD5Add(sc->sc_inc.inc6_laddr);
1408 		MD5Add(sc->sc_inc.inc6_faddr);
1409 		add.laddr = 0;
1410 		add.faddr = 0;
1411 	} else {
1412 		add.laddr = sc->sc_inc.inc_laddr.s_addr;
1413 		add.faddr = sc->sc_inc.inc_faddr.s_addr;
1414 	}
1415 	add.lport = sc->sc_inc.inc_lport;
1416 	add.fport = sc->sc_inc.inc_fport;
1417 	add.secbits[0] = tcp_secret[idx].ts_secbits[0];
1418 	add.secbits[1] = tcp_secret[idx].ts_secbits[1];
1419 	add.secbits[2] = tcp_secret[idx].ts_secbits[2];
1420 	add.secbits[3] = tcp_secret[idx].ts_secbits[3];
1421 	MD5Add(add);
1422 	MD5Final((u_char *)&md5_buffer, &syn_ctx);
1423 	data ^= (md5_buffer[0] & ~SYNCOOKIE_WNDMASK);
1424 	return (data);
1425 }
1426 
1427 static struct syncache *
1428 syncookie_lookup(struct in_conninfo *inc, struct tcphdr *th, struct socket *so)
1429 {
1430 	u_int32_t md5_buffer[4];
1431 	struct syncache *sc;
1432 	u_int32_t data;
1433 	int wnd, idx;
1434 	struct md5_add add;
1435 
1436 	data = (th->th_ack - 1) ^ (th->th_seq - 1);	/* remove ISS */
1437 	idx = data & SYNCOOKIE_WNDMASK;
1438 	if (tcp_secret[idx].ts_expire < ticks ||
1439 	    sototcpcb(so)->ts_recent + SYNCOOKIE_TIMEOUT < ticks)
1440 		return (NULL);
1441 	MD5Init(&syn_ctx);
1442 #ifdef INET6
1443 	if (inc->inc_isipv6) {
1444 		MD5Add(inc->inc6_laddr);
1445 		MD5Add(inc->inc6_faddr);
1446 		add.laddr = 0;
1447 		add.faddr = 0;
1448 	} else
1449 #endif
1450 	{
1451 		add.laddr = inc->inc_laddr.s_addr;
1452 		add.faddr = inc->inc_faddr.s_addr;
1453 	}
1454 	add.lport = inc->inc_lport;
1455 	add.fport = inc->inc_fport;
1456 	add.secbits[0] = tcp_secret[idx].ts_secbits[0];
1457 	add.secbits[1] = tcp_secret[idx].ts_secbits[1];
1458 	add.secbits[2] = tcp_secret[idx].ts_secbits[2];
1459 	add.secbits[3] = tcp_secret[idx].ts_secbits[3];
1460 	MD5Add(add);
1461 	MD5Final((u_char *)&md5_buffer, &syn_ctx);
1462 	data ^= md5_buffer[0];
1463 	if (data & ~SYNCOOKIE_DATAMASK)
1464 		return (NULL);
1465 	data = data >> SYNCOOKIE_WNDBITS;
1466 
1467 	/*
1468 	 * Fill in the syncache values.
1469 	 * XXX duplicate code from syncache_add
1470 	 */
1471 	sc = kmalloc(sizeof(struct syncache), M_SYNCACHE, M_WAITOK|M_ZERO);
1472 	sc->sc_ipopts = NULL;
1473 	sc->sc_inc.inc_fport = inc->inc_fport;
1474 	sc->sc_inc.inc_lport = inc->inc_lport;
1475 #ifdef INET6
1476 	sc->sc_inc.inc_isipv6 = inc->inc_isipv6;
1477 	if (inc->inc_isipv6) {
1478 		sc->sc_inc.inc6_faddr = inc->inc6_faddr;
1479 		sc->sc_inc.inc6_laddr = inc->inc6_laddr;
1480 		sc->sc_route6.ro_rt = NULL;
1481 	} else
1482 #endif
1483 	{
1484 		sc->sc_inc.inc_faddr = inc->inc_faddr;
1485 		sc->sc_inc.inc_laddr = inc->inc_laddr;
1486 		sc->sc_route.ro_rt = NULL;
1487 	}
1488 	sc->sc_irs = th->th_seq - 1;
1489 	sc->sc_iss = th->th_ack - 1;
1490 	wnd = ssb_space(&so->so_rcv);
1491 	wnd = imax(wnd, 0);
1492 	wnd = imin(wnd, TCP_MAXWIN);
1493 	sc->sc_wnd = wnd;
1494 	sc->sc_flags = 0;
1495 	sc->sc_rxtslot = 0;
1496 	sc->sc_peer_mss = tcp_msstab[data];
1497 	return (sc);
1498 }
1499 
1500 static int
1501 syncache_sysctl_count(SYSCTL_HANDLER_ARGS)
1502 {
1503 	u_int count = 0;
1504 	int cpu;
1505 
1506 	for (cpu = 0; cpu < netisr_ncpus; ++cpu)
1507 		count += tcp_syncache_percpu[cpu]->cache_count;
1508 	return sysctl_handle_int(oidp, &count, 0, req);
1509 }
1510