xref: /dflybsd-src/sys/netinet/tcp_syncache.c (revision 04db30e59799dbe046f0bdb23b82513671b9523e)
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 
94 #include <net/if.h>
95 #include <net/route.h>
96 
97 #include <netinet/in.h>
98 #include <netinet/in_systm.h>
99 #include <netinet/ip.h>
100 #include <netinet/in_var.h>
101 #include <netinet/in_pcb.h>
102 #include <netinet/ip_var.h>
103 #include <netinet/ip6.h>
104 #ifdef INET6
105 #include <netinet/icmp6.h>
106 #include <netinet6/nd6.h>
107 #endif
108 #include <netinet6/ip6_var.h>
109 #include <netinet6/in6_pcb.h>
110 #include <netinet/tcp.h>
111 #include <netinet/tcp_fsm.h>
112 #include <netinet/tcp_seq.h>
113 #include <netinet/tcp_timer.h>
114 #include <netinet/tcp_timer2.h>
115 #include <netinet/tcp_var.h>
116 #include <netinet6/tcp6_var.h>
117 
118 #ifdef IPSEC
119 #include <netinet6/ipsec.h>
120 #ifdef INET6
121 #include <netinet6/ipsec6.h>
122 #endif
123 #include <netproto/key/key.h>
124 #endif /*IPSEC*/
125 
126 #ifdef FAST_IPSEC
127 #include <netproto/ipsec/ipsec.h>
128 #ifdef INET6
129 #include <netproto/ipsec/ipsec6.h>
130 #endif
131 #include <netproto/ipsec/key.h>
132 #define	IPSEC
133 #endif /*FAST_IPSEC*/
134 
135 static int tcp_syncookies = 1;
136 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_RW,
137     &tcp_syncookies, 0,
138     "Use TCP SYN cookies if the syncache overflows");
139 
140 static void	 syncache_drop(struct syncache *, struct syncache_head *);
141 static void	 syncache_free(struct syncache *);
142 static void	 syncache_insert(struct syncache *, struct syncache_head *);
143 struct syncache *syncache_lookup(struct in_conninfo *, struct syncache_head **);
144 static int	 syncache_respond(struct syncache *, struct mbuf *);
145 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
146 		    struct mbuf *);
147 static void	 syncache_timer(void *);
148 static u_int32_t syncookie_generate(struct syncache *);
149 static struct syncache *syncookie_lookup(struct in_conninfo *,
150 		    struct tcphdr *, struct socket *);
151 
152 /*
153  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
154  * 3 retransmits corresponds to a timeout of (1 + 2 + 4 + 8 == 15) seconds,
155  * the odds are that the user has given up attempting to connect by then.
156  */
157 #define SYNCACHE_MAXREXMTS		3
158 
159 /* Arbitrary values */
160 #define TCP_SYNCACHE_HASHSIZE		512
161 #define TCP_SYNCACHE_BUCKETLIMIT	30
162 
163 struct netmsg_sc_timer {
164 	struct netmsg_base base;
165 	struct msgrec *nm_mrec;		/* back pointer to containing msgrec */
166 };
167 
168 struct msgrec {
169 	struct netmsg_sc_timer msg;
170 	lwkt_port_t port;		/* constant after init */
171 	int slot;			/* constant after init */
172 };
173 
174 static void syncache_timer_handler(netmsg_t);
175 
176 struct tcp_syncache {
177 	u_int	hashsize;
178 	u_int	hashmask;
179 	u_int	bucket_limit;
180 	u_int	cache_limit;
181 	u_int	rexmt_limit;
182 	u_int	hash_secret;
183 };
184 static struct tcp_syncache tcp_syncache;
185 
186 TAILQ_HEAD(syncache_list, syncache);
187 
188 struct tcp_syncache_percpu {
189 	struct syncache_head	*hashbase;
190 	u_int			cache_count;
191 	struct syncache_list	timerq[SYNCACHE_MAXREXMTS + 1];
192 	struct callout		tt_timerq[SYNCACHE_MAXREXMTS + 1];
193 	struct msgrec		mrec[SYNCACHE_MAXREXMTS + 1];
194 };
195 static struct tcp_syncache_percpu tcp_syncache_percpu[MAXCPU];
196 
197 static struct lwkt_port syncache_null_rport;
198 
199 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0, "TCP SYN cache");
200 
201 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RD,
202      &tcp_syncache.bucket_limit, 0, "Per-bucket hash limit for syncache");
203 
204 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_RD,
205      &tcp_syncache.cache_limit, 0, "Overall entry limit for syncache");
206 
207 /* XXX JH */
208 #if 0
209 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_RD,
210      &tcp_syncache.cache_count, 0, "Current number of entries in syncache");
211 #endif
212 
213 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RD,
214      &tcp_syncache.hashsize, 0, "Size of TCP syncache hashtable");
215 
216 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_RW,
217      &tcp_syncache.rexmt_limit, 0, "Limit on SYN/ACK retransmissions");
218 
219 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
220 
221 #define SYNCACHE_HASH(inc, mask)					\
222 	((tcp_syncache.hash_secret ^					\
223 	  (inc)->inc_faddr.s_addr ^					\
224 	  ((inc)->inc_faddr.s_addr >> 16) ^				\
225 	  (inc)->inc_fport ^ (inc)->inc_lport) & mask)
226 
227 #define SYNCACHE_HASH6(inc, mask)					\
228 	((tcp_syncache.hash_secret ^					\
229 	  (inc)->inc6_faddr.s6_addr32[0] ^				\
230 	  (inc)->inc6_faddr.s6_addr32[3] ^				\
231 	  (inc)->inc_fport ^ (inc)->inc_lport) & mask)
232 
233 #define ENDPTS_EQ(a, b) (						\
234 	(a)->ie_fport == (b)->ie_fport &&				\
235 	(a)->ie_lport == (b)->ie_lport &&				\
236 	(a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr &&			\
237 	(a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr			\
238 )
239 
240 #define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0)
241 
242 static __inline void
243 syncache_timeout(struct tcp_syncache_percpu *syncache_percpu,
244 		 struct syncache *sc, int slot)
245 {
246 	if (slot > 0) {
247 		/*
248 		 * Record that SYN|ACK was lost.
249 		 * Needed by RFC3390 and RFC6298.
250 		 */
251 		sc->sc_flags |= SCF_SYN_WASLOST;
252 	}
253 	sc->sc_rxtslot = slot;
254 	sc->sc_rxttime = ticks + TCPTV_RTOBASE * tcp_backoff[slot];
255 	TAILQ_INSERT_TAIL(&syncache_percpu->timerq[slot], sc, sc_timerq);
256 	if (!callout_active(&syncache_percpu->tt_timerq[slot])) {
257 		callout_reset(&syncache_percpu->tt_timerq[slot],
258 			      TCPTV_RTOBASE * tcp_backoff[slot],
259 			      syncache_timer,
260 			      &syncache_percpu->mrec[slot]);
261 	}
262 }
263 
264 static void
265 syncache_free(struct syncache *sc)
266 {
267 	struct rtentry *rt;
268 #ifdef INET6
269 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
270 #else
271 	const boolean_t isipv6 = FALSE;
272 #endif
273 
274 	if (sc->sc_ipopts)
275 		m_free(sc->sc_ipopts);
276 
277 	rt = isipv6 ? sc->sc_route6.ro_rt : sc->sc_route.ro_rt;
278 	if (rt != NULL) {
279 		/*
280 		 * If this is the only reference to a protocol-cloned
281 		 * route, remove it immediately.
282 		 */
283 		if ((rt->rt_flags & RTF_WASCLONED) && rt->rt_refcnt == 1)
284 			rtrequest(RTM_DELETE, rt_key(rt), rt->rt_gateway,
285 				  rt_mask(rt), rt->rt_flags, NULL);
286 		RTFREE(rt);
287 	}
288 	kfree(sc, M_SYNCACHE);
289 }
290 
291 void
292 syncache_init(void)
293 {
294 	int i, cpu;
295 
296 	tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
297 	tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
298 	tcp_syncache.cache_limit =
299 	    tcp_syncache.hashsize * tcp_syncache.bucket_limit;
300 	tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
301 	tcp_syncache.hash_secret = karc4random();
302 
303 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
304 	    &tcp_syncache.hashsize);
305 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
306 	    &tcp_syncache.cache_limit);
307 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
308 	    &tcp_syncache.bucket_limit);
309 	if (!powerof2(tcp_syncache.hashsize)) {
310 		kprintf("WARNING: syncache hash size is not a power of 2.\n");
311 		tcp_syncache.hashsize = 512;	/* safe default */
312 	}
313 	tcp_syncache.hashmask = tcp_syncache.hashsize - 1;
314 
315 	lwkt_initport_replyonly_null(&syncache_null_rport);
316 
317 	for (cpu = 0; cpu < ncpus2; cpu++) {
318 		struct tcp_syncache_percpu *syncache_percpu;
319 
320 		syncache_percpu = &tcp_syncache_percpu[cpu];
321 		/* Allocate the hash table. */
322 		syncache_percpu->hashbase = kmalloc(tcp_syncache.hashsize * sizeof(struct syncache_head),
323 						    M_SYNCACHE, M_WAITOK);
324 
325 		/* Initialize the hash buckets. */
326 		for (i = 0; i < tcp_syncache.hashsize; i++) {
327 			struct syncache_head *bucket;
328 
329 			bucket = &syncache_percpu->hashbase[i];
330 			TAILQ_INIT(&bucket->sch_bucket);
331 			bucket->sch_length = 0;
332 		}
333 
334 		for (i = 0; i <= SYNCACHE_MAXREXMTS; i++) {
335 			/* Initialize the timer queues. */
336 			TAILQ_INIT(&syncache_percpu->timerq[i]);
337 			callout_init_mp(&syncache_percpu->tt_timerq[i]);
338 
339 			syncache_percpu->mrec[i].slot = i;
340 			syncache_percpu->mrec[i].port = cpu_portfn(cpu);
341 			syncache_percpu->mrec[i].msg.nm_mrec =
342 				    &syncache_percpu->mrec[i];
343 			netmsg_init(&syncache_percpu->mrec[i].msg.base,
344 				    NULL, &syncache_null_rport,
345 				    0, syncache_timer_handler);
346 		}
347 	}
348 }
349 
350 static void
351 syncache_insert(struct syncache *sc, struct syncache_head *sch)
352 {
353 	struct tcp_syncache_percpu *syncache_percpu;
354 	struct syncache *sc2;
355 	int i;
356 
357 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
358 
359 	/*
360 	 * Make sure that we don't overflow the per-bucket
361 	 * limit or the total cache size limit.
362 	 */
363 	if (sch->sch_length >= tcp_syncache.bucket_limit) {
364 		/*
365 		 * The bucket is full, toss the oldest element.
366 		 */
367 		sc2 = TAILQ_FIRST(&sch->sch_bucket);
368 		sc2->sc_tp->ts_recent = ticks;
369 		syncache_drop(sc2, sch);
370 		tcpstat.tcps_sc_bucketoverflow++;
371 	} else if (syncache_percpu->cache_count >= tcp_syncache.cache_limit) {
372 		/*
373 		 * The cache is full.  Toss the oldest entry in the
374 		 * entire cache.  This is the front entry in the
375 		 * first non-empty timer queue with the largest
376 		 * timeout value.
377 		 */
378 		for (i = SYNCACHE_MAXREXMTS; i >= 0; i--) {
379 			sc2 = TAILQ_FIRST(&syncache_percpu->timerq[i]);
380 			while (sc2 && (sc2->sc_flags & SCF_MARKER))
381 				sc2 = TAILQ_NEXT(sc2, sc_timerq);
382 			if (sc2 != NULL)
383 				break;
384 		}
385 		sc2->sc_tp->ts_recent = ticks;
386 		syncache_drop(sc2, NULL);
387 		tcpstat.tcps_sc_cacheoverflow++;
388 	}
389 
390 	/* Initialize the entry's timer. */
391 	syncache_timeout(syncache_percpu, sc, 0);
392 
393 	/* Put it into the bucket. */
394 	TAILQ_INSERT_TAIL(&sch->sch_bucket, sc, sc_hash);
395 	sch->sch_length++;
396 	syncache_percpu->cache_count++;
397 	tcpstat.tcps_sc_added++;
398 }
399 
400 void
401 syncache_destroy(struct tcpcb *tp)
402 {
403 	struct tcp_syncache_percpu *syncache_percpu;
404 	struct syncache_head *bucket;
405 	struct syncache *sc;
406 	int i;
407 
408 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
409 	sc = NULL;
410 
411 	for (i = 0; i < tcp_syncache.hashsize; i++) {
412 		bucket = &syncache_percpu->hashbase[i];
413 		TAILQ_FOREACH(sc, &bucket->sch_bucket, sc_hash) {
414 			if (sc->sc_tp == tp)
415 				sc->sc_tp = NULL;
416 		}
417 	}
418 }
419 
420 static void
421 syncache_drop(struct syncache *sc, struct syncache_head *sch)
422 {
423 	struct tcp_syncache_percpu *syncache_percpu;
424 #ifdef INET6
425 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
426 #else
427 	const boolean_t isipv6 = FALSE;
428 #endif
429 
430 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
431 
432 	if (sch == NULL) {
433 		if (isipv6) {
434 			sch = &syncache_percpu->hashbase[
435 			    SYNCACHE_HASH6(&sc->sc_inc, tcp_syncache.hashmask)];
436 		} else {
437 			sch = &syncache_percpu->hashbase[
438 			    SYNCACHE_HASH(&sc->sc_inc, tcp_syncache.hashmask)];
439 		}
440 	}
441 
442 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
443 	sch->sch_length--;
444 	syncache_percpu->cache_count--;
445 
446 	/*
447 	 * Cleanup
448 	 */
449 	if (sc->sc_tp)
450 		sc->sc_tp = NULL;
451 
452 	/*
453 	 * Remove the entry from the syncache timer/timeout queue.  Note
454 	 * that we do not try to stop any running timer since we do not know
455 	 * whether the timer's message is in-transit or not.  Since timeouts
456 	 * are fairly long, taking an unneeded callout does not detrimentally
457 	 * effect performance.
458 	 */
459 	TAILQ_REMOVE(&syncache_percpu->timerq[sc->sc_rxtslot], sc, sc_timerq);
460 
461 	syncache_free(sc);
462 }
463 
464 /*
465  * Place a timeout message on the TCP thread's message queue.
466  * This routine runs in soft interrupt context.
467  *
468  * An invariant is for this routine to be called, the callout must
469  * have been active.  Note that the callout is not deactivated until
470  * after the message has been processed in syncache_timer_handler() below.
471  */
472 static void
473 syncache_timer(void *p)
474 {
475 	struct netmsg_sc_timer *msg = p;
476 
477 	lwkt_sendmsg(msg->nm_mrec->port, &msg->base.lmsg);
478 }
479 
480 /*
481  * Service a timer message queued by timer expiration.
482  * This routine runs in the TCP protocol thread.
483  *
484  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
485  * If we have retransmitted an entry the maximum number of times, expire it.
486  *
487  * When we finish processing timed-out entries, we restart the timer if there
488  * are any entries still on the queue and deactivate it otherwise.  Only after
489  * a timer has been deactivated here can it be restarted by syncache_timeout().
490  */
491 static void
492 syncache_timer_handler(netmsg_t msg)
493 {
494 	struct tcp_syncache_percpu *syncache_percpu;
495 	struct syncache *sc;
496 	struct syncache marker;
497 	struct syncache_list *list;
498 	struct inpcb *inp;
499 	int slot;
500 
501 	slot = ((struct netmsg_sc_timer *)msg)->nm_mrec->slot;
502 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
503 
504 	list = &syncache_percpu->timerq[slot];
505 
506 	/*
507 	 * Use a marker to keep our place in the scan.  syncache_drop()
508 	 * can block and cause any next pointer we cache to become stale.
509 	 */
510 	marker.sc_flags = SCF_MARKER;
511 	TAILQ_INSERT_HEAD(list, &marker, sc_timerq);
512 
513 	while ((sc = TAILQ_NEXT(&marker, sc_timerq)) != NULL) {
514 		/*
515 		 * Move the marker.
516 		 */
517 		TAILQ_REMOVE(list, &marker, sc_timerq);
518 		TAILQ_INSERT_AFTER(list, sc, &marker, sc_timerq);
519 
520 		if (sc->sc_flags & SCF_MARKER)
521 			continue;
522 
523 		if (ticks < sc->sc_rxttime)
524 			break;	/* finished because timerq sorted by time */
525 		if (sc->sc_tp == NULL) {
526 			syncache_drop(sc, NULL);
527 			tcpstat.tcps_sc_stale++;
528 			continue;
529 		}
530 		inp = sc->sc_tp->t_inpcb;
531 		if (slot == SYNCACHE_MAXREXMTS ||
532 		    slot >= tcp_syncache.rexmt_limit ||
533 		    inp == NULL ||
534 		    inp->inp_gencnt != sc->sc_inp_gencnt) {
535 			syncache_drop(sc, NULL);
536 			tcpstat.tcps_sc_stale++;
537 			continue;
538 		}
539 		/*
540 		 * syncache_respond() may call back into the syncache to
541 		 * to modify another entry, so do not obtain the next
542 		 * entry on the timer chain until it has completed.
543 		 */
544 		syncache_respond(sc, NULL);
545 		tcpstat.tcps_sc_retransmitted++;
546 		TAILQ_REMOVE(list, sc, sc_timerq);
547 		syncache_timeout(syncache_percpu, sc, slot + 1);
548 	}
549 	TAILQ_REMOVE(list, &marker, sc_timerq);
550 
551 	if (sc != NULL) {
552 		callout_reset(&syncache_percpu->tt_timerq[slot],
553 			      sc->sc_rxttime - ticks, syncache_timer,
554 			      &syncache_percpu->mrec[slot]);
555 	} else {
556 		callout_deactivate(&syncache_percpu->tt_timerq[slot]);
557 	}
558 	lwkt_replymsg(&msg->base.lmsg, 0);
559 }
560 
561 /*
562  * Find an entry in the syncache.
563  */
564 struct syncache *
565 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
566 {
567 	struct tcp_syncache_percpu *syncache_percpu;
568 	struct syncache *sc;
569 	struct syncache_head *sch;
570 
571 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
572 #ifdef INET6
573 	if (inc->inc_isipv6) {
574 		sch = &syncache_percpu->hashbase[
575 		    SYNCACHE_HASH6(inc, tcp_syncache.hashmask)];
576 		*schp = sch;
577 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
578 			if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
579 				return (sc);
580 	} else
581 #endif
582 	{
583 		sch = &syncache_percpu->hashbase[
584 		    SYNCACHE_HASH(inc, tcp_syncache.hashmask)];
585 		*schp = sch;
586 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
587 #ifdef INET6
588 			if (sc->sc_inc.inc_isipv6)
589 				continue;
590 #endif
591 			if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
592 				return (sc);
593 		}
594 	}
595 	return (NULL);
596 }
597 
598 /*
599  * This function is called when we get a RST for a
600  * non-existent connection, so that we can see if the
601  * connection is in the syn cache.  If it is, zap it.
602  */
603 void
604 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
605 {
606 	struct syncache *sc;
607 	struct syncache_head *sch;
608 
609 	sc = syncache_lookup(inc, &sch);
610 	if (sc == NULL) {
611 		return;
612 	}
613 	/*
614 	 * If the RST bit is set, check the sequence number to see
615 	 * if this is a valid reset segment.
616 	 * RFC 793 page 37:
617 	 *   In all states except SYN-SENT, all reset (RST) segments
618 	 *   are validated by checking their SEQ-fields.  A reset is
619 	 *   valid if its sequence number is in the window.
620 	 *
621 	 *   The sequence number in the reset segment is normally an
622 	 *   echo of our outgoing acknowlegement numbers, but some hosts
623 	 *   send a reset with the sequence number at the rightmost edge
624 	 *   of our receive window, and we have to handle this case.
625 	 */
626 	if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
627 	    SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
628 		syncache_drop(sc, sch);
629 		tcpstat.tcps_sc_reset++;
630 	}
631 }
632 
633 void
634 syncache_badack(struct in_conninfo *inc)
635 {
636 	struct syncache *sc;
637 	struct syncache_head *sch;
638 
639 	sc = syncache_lookup(inc, &sch);
640 	if (sc != NULL) {
641 		syncache_drop(sc, sch);
642 		tcpstat.tcps_sc_badack++;
643 	}
644 }
645 
646 void
647 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th)
648 {
649 	struct syncache *sc;
650 	struct syncache_head *sch;
651 
652 	/* we are called at splnet() here */
653 	sc = syncache_lookup(inc, &sch);
654 	if (sc == NULL)
655 		return;
656 
657 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
658 	if (ntohl(th->th_seq) != sc->sc_iss)
659 		return;
660 
661 	/*
662 	 * If we've rertransmitted 3 times and this is our second error,
663 	 * we remove the entry.  Otherwise, we allow it to continue on.
664 	 * This prevents us from incorrectly nuking an entry during a
665 	 * spurious network outage.
666 	 *
667 	 * See tcp_notify().
668 	 */
669 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxtslot < 3) {
670 		sc->sc_flags |= SCF_UNREACH;
671 		return;
672 	}
673 	syncache_drop(sc, sch);
674 	tcpstat.tcps_sc_unreach++;
675 }
676 
677 /*
678  * Build a new TCP socket structure from a syncache entry.
679  *
680  * This is called from the context of the SYN+ACK
681  */
682 static struct socket *
683 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
684 {
685 	struct inpcb *inp = NULL, *linp;
686 	struct socket *so;
687 	struct tcpcb *tp, *ltp;
688 	lwkt_port_t port;
689 #ifdef INET6
690 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
691 #else
692 	const boolean_t isipv6 = FALSE;
693 #endif
694 	struct sockaddr_in sin_faddr;
695 	struct sockaddr_in6 sin6_faddr;
696 	struct sockaddr *faddr;
697 
698 	if (isipv6) {
699 		faddr = (struct sockaddr *)&sin6_faddr;
700 		sin6_faddr.sin6_family = AF_INET6;
701 		sin6_faddr.sin6_len = sizeof(sin6_faddr);
702 		sin6_faddr.sin6_addr = sc->sc_inc.inc6_faddr;
703 		sin6_faddr.sin6_port = sc->sc_inc.inc_fport;
704 		sin6_faddr.sin6_flowinfo = sin6_faddr.sin6_scope_id = 0;
705 	} else {
706 		faddr = (struct sockaddr *)&sin_faddr;
707 		sin_faddr.sin_family = AF_INET;
708 		sin_faddr.sin_len = sizeof(sin_faddr);
709 		sin_faddr.sin_addr = sc->sc_inc.inc_faddr;
710 		sin_faddr.sin_port = sc->sc_inc.inc_fport;
711 		bzero(sin_faddr.sin_zero, sizeof(sin_faddr.sin_zero));
712 	}
713 
714 	/*
715 	 * Ok, create the full blown connection, and set things up
716 	 * as they would have been set up if we had created the
717 	 * connection when the SYN arrived.  If we can't create
718 	 * the connection, abort it.
719 	 *
720 	 * Set the protocol processing port for the socket to the current
721 	 * port (that the connection came in on).
722 	 */
723 	so = sonewconn_faddr(lso, SS_ISCONNECTED, faddr);
724 	if (so == NULL) {
725 		/*
726 		 * Drop the connection; we will send a RST if the peer
727 		 * retransmits the ACK,
728 		 */
729 		tcpstat.tcps_listendrop++;
730 		goto abort;
731 	}
732 
733 	/*
734 	 * Insert new socket into hash list.
735 	 */
736 	inp = so->so_pcb;
737 	inp->inp_inc.inc_isipv6 = sc->sc_inc.inc_isipv6;
738 	if (isipv6) {
739 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
740 	} else {
741 #ifdef INET6
742 		inp->inp_vflag &= ~INP_IPV6;
743 		inp->inp_vflag |= INP_IPV4;
744 		inp->inp_flags &= ~IN6P_IPV6_V6ONLY;
745 #endif
746 		inp->inp_laddr = sc->sc_inc.inc_laddr;
747 	}
748 	inp->inp_lport = sc->sc_inc.inc_lport;
749 	if (in_pcbinsporthash(inp) != 0) {
750 		/*
751 		 * Undo the assignments above if we failed to
752 		 * put the PCB on the hash lists.
753 		 */
754 		if (isipv6)
755 			inp->in6p_laddr = kin6addr_any;
756 		else
757 			inp->inp_laddr.s_addr = INADDR_ANY;
758 		inp->inp_lport = 0;
759 		goto abort;
760 	}
761 	linp = lso->so_pcb;
762 #ifdef IPSEC
763 	/* copy old policy into new socket's */
764 	if (ipsec_copy_policy(linp->inp_sp, inp->inp_sp))
765 		kprintf("syncache_expand: could not copy policy\n");
766 #endif
767 	if (isipv6) {
768 		struct in6_addr laddr6;
769 		/*
770 		 * Inherit socket options from the listening socket.
771 		 * Note that in6p_inputopts are not (and should not be)
772 		 * copied, since it stores previously received options and is
773 		 * used to detect if each new option is different than the
774 		 * previous one and hence should be passed to a user.
775 		 * If we copied in6p_inputopts, a user would not be able to
776 		 * receive options just after calling the accept system call.
777 		 */
778 		inp->inp_flags |= linp->inp_flags & INP_CONTROLOPTS;
779 		if (linp->in6p_outputopts)
780 			inp->in6p_outputopts =
781 			    ip6_copypktopts(linp->in6p_outputopts, M_INTWAIT);
782 		inp->in6p_route = sc->sc_route6;
783 		sc->sc_route6.ro_rt = NULL;
784 
785 		laddr6 = inp->in6p_laddr;
786 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
787 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
788 		if (in6_pcbconnect(inp, faddr, &thread0)) {
789 			inp->in6p_laddr = laddr6;
790 			goto abort;
791 		}
792 	} else {
793 		struct in_addr laddr;
794 
795 		inp->inp_options = ip_srcroute(m);
796 		if (inp->inp_options == NULL) {
797 			inp->inp_options = sc->sc_ipopts;
798 			sc->sc_ipopts = NULL;
799 		}
800 		inp->inp_route = sc->sc_route;
801 		sc->sc_route.ro_rt = NULL;
802 
803 		laddr = inp->inp_laddr;
804 		if (inp->inp_laddr.s_addr == INADDR_ANY)
805 			inp->inp_laddr = sc->sc_inc.inc_laddr;
806 		if (in_pcbconnect(inp, faddr, &thread0)) {
807 			inp->inp_laddr = laddr;
808 			goto abort;
809 		}
810 	}
811 
812 	/*
813 	 * The current port should be in the context of the SYN+ACK and
814 	 * so should match the tcp address port.
815 	 *
816 	 * XXX we may be running on the netisr thread instead of a tcp
817 	 *     thread, in which case port will not match
818 	 *     curthread->td_msgport.
819 	 */
820 	if (isipv6) {
821 		port = tcp6_addrport();
822 	} else {
823 		port = tcp_addrport(inp->inp_faddr.s_addr, inp->inp_fport,
824 				    inp->inp_laddr.s_addr, inp->inp_lport);
825 	}
826 	if (port != &curthread->td_msgport) {
827 		print_backtrace(-1);
828 		kprintf("TCP PORT MISMATCH %p vs %p\n",
829 			port, &curthread->td_msgport);
830 	}
831 	/*KKASSERT(port == &curthread->td_msgport);*/
832 
833 	tp = intotcpcb(inp);
834 	tp->t_state = TCPS_SYN_RECEIVED;
835 	tp->iss = sc->sc_iss;
836 	tp->irs = sc->sc_irs;
837 	tcp_rcvseqinit(tp);
838 	tcp_sendseqinit(tp);
839 	tp->snd_wl1 = sc->sc_irs;
840 	tp->rcv_up = sc->sc_irs + 1;
841 	tp->rcv_wnd = sc->sc_wnd;
842 	tp->rcv_adv += tp->rcv_wnd;
843 
844 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH | TF_NODELAY);
845 	if (sc->sc_flags & SCF_NOOPT)
846 		tp->t_flags |= TF_NOOPT;
847 	if (sc->sc_flags & SCF_WINSCALE) {
848 		tp->t_flags |= TF_REQ_SCALE | TF_RCVD_SCALE;
849 		tp->requested_s_scale = sc->sc_requested_s_scale;
850 		tp->request_r_scale = sc->sc_request_r_scale;
851 	}
852 	if (sc->sc_flags & SCF_TIMESTAMP) {
853 		tp->t_flags |= TF_REQ_TSTMP | TF_RCVD_TSTMP;
854 		tp->ts_recent = sc->sc_tsrecent;
855 		tp->ts_recent_age = ticks;
856 	}
857 	if (sc->sc_flags & SCF_SACK_PERMITTED)
858 		tp->t_flags |= TF_SACK_PERMITTED;
859 	if (sc->sc_flags & SCF_SYN_WASLOST)
860 		tp->t_flags |= TF_SYN_WASLOST;
861 
862 #ifdef TCP_SIGNATURE
863 	if (sc->sc_flags & SCF_SIGNATURE)
864 		tp->t_flags |= TF_SIGNATURE;
865 #endif /* TCP_SIGNATURE */
866 
867 	tcp_mss(tp, sc->sc_peer_mss);
868 
869 	/*
870 	 * Inherit some properties from the listen socket
871 	 */
872 	ltp = intotcpcb(linp);
873 	tp->t_keepinit = ltp->t_keepinit;
874 	tp->t_keepidle = ltp->t_keepidle;
875 	tp->t_keepintvl = ltp->t_keepintvl;
876 	tp->t_keepcnt = ltp->t_keepcnt;
877 	tp->t_maxidle = ltp->t_maxidle;
878 
879 	tcp_create_timermsg(tp, port);
880 	tcp_callout_reset(tp, tp->tt_keep, tp->t_keepinit, tcp_timer_keep);
881 
882 	tcpstat.tcps_accepts++;
883 	return (so);
884 
885 abort:
886 	if (so != NULL)
887 		soabort_oncpu(so);
888 	return (NULL);
889 }
890 
891 /*
892  * This function gets called when we receive an ACK for a
893  * socket in the LISTEN state.  We look up the connection
894  * in the syncache, and if its there, we pull it out of
895  * the cache and turn it into a full-blown connection in
896  * the SYN-RECEIVED state.
897  */
898 int
899 syncache_expand(struct in_conninfo *inc, struct tcphdr *th, struct socket **sop,
900 		struct mbuf *m)
901 {
902 	struct syncache *sc;
903 	struct syncache_head *sch;
904 	struct socket *so;
905 
906 	sc = syncache_lookup(inc, &sch);
907 	if (sc == NULL) {
908 		/*
909 		 * There is no syncache entry, so see if this ACK is
910 		 * a returning syncookie.  To do this, first:
911 		 *  A. See if this socket has had a syncache entry dropped in
912 		 *     the past.  We don't want to accept a bogus syncookie
913 		 *     if we've never received a SYN.
914 		 *  B. check that the syncookie is valid.  If it is, then
915 		 *     cobble up a fake syncache entry, and return.
916 		 */
917 		if (!tcp_syncookies)
918 			return (0);
919 		sc = syncookie_lookup(inc, th, *sop);
920 		if (sc == NULL)
921 			return (0);
922 		sch = NULL;
923 		tcpstat.tcps_sc_recvcookie++;
924 	}
925 
926 	/*
927 	 * If seg contains an ACK, but not for our SYN/ACK, send a RST.
928 	 */
929 	if (th->th_ack != sc->sc_iss + 1)
930 		return (0);
931 
932 	so = syncache_socket(sc, *sop, m);
933 	if (so == NULL) {
934 #if 0
935 resetandabort:
936 		/* XXXjlemon check this - is this correct? */
937 		tcp_respond(NULL, m, m, th,
938 		    th->th_seq + tlen, (tcp_seq)0, TH_RST | TH_ACK);
939 #endif
940 		m_freem(m);			/* XXX only needed for above */
941 		tcpstat.tcps_sc_aborted++;
942 	} else {
943 		tcpstat.tcps_sc_completed++;
944 	}
945 	if (sch == NULL)
946 		syncache_free(sc);
947 	else
948 		syncache_drop(sc, sch);
949 	*sop = so;
950 	return (1);
951 }
952 
953 /*
954  * Given a LISTEN socket and an inbound SYN request, add
955  * this to the syn cache, and send back a segment:
956  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
957  * to the source.
958  *
959  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
960  * Doing so would require that we hold onto the data and deliver it
961  * to the application.  However, if we are the target of a SYN-flood
962  * DoS attack, an attacker could send data which would eventually
963  * consume all available buffer space if it were ACKed.  By not ACKing
964  * the data, we avoid this DoS scenario.
965  */
966 int
967 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
968 	     struct socket *so, struct mbuf *m)
969 {
970 	struct tcp_syncache_percpu *syncache_percpu;
971 	struct tcpcb *tp;
972 	struct syncache *sc = NULL;
973 	struct syncache_head *sch;
974 	struct mbuf *ipopts = NULL;
975 	int win;
976 
977 	syncache_percpu = &tcp_syncache_percpu[mycpu->gd_cpuid];
978 	tp = sototcpcb(so);
979 
980 	/*
981 	 * Remember the IP options, if any.
982 	 */
983 #ifdef INET6
984 	if (!inc->inc_isipv6)
985 #endif
986 		ipopts = ip_srcroute(m);
987 
988 	/*
989 	 * See if we already have an entry for this connection.
990 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
991 	 *
992 	 * XXX
993 	 * The syncache should be re-initialized with the contents
994 	 * of the new SYN which may have different options.
995 	 */
996 	sc = syncache_lookup(inc, &sch);
997 	if (sc != NULL) {
998 		tcpstat.tcps_sc_dupsyn++;
999 		if (ipopts) {
1000 			/*
1001 			 * If we were remembering a previous source route,
1002 			 * forget it and use the new one we've been given.
1003 			 */
1004 			if (sc->sc_ipopts)
1005 				m_free(sc->sc_ipopts);
1006 			sc->sc_ipopts = ipopts;
1007 		}
1008 		/*
1009 		 * Update timestamp if present.
1010 		 */
1011 		if (sc->sc_flags & SCF_TIMESTAMP)
1012 			sc->sc_tsrecent = to->to_tsval;
1013 
1014 		/* Just update the TOF_SACK_PERMITTED for now. */
1015 		if (tcp_do_sack && (to->to_flags & TOF_SACK_PERMITTED))
1016 			sc->sc_flags |= SCF_SACK_PERMITTED;
1017 		else
1018 			sc->sc_flags &= ~SCF_SACK_PERMITTED;
1019 
1020 		/*
1021 		 * PCB may have changed, pick up new values.
1022 		 */
1023 		sc->sc_tp = tp;
1024 		sc->sc_inp_gencnt = tp->t_inpcb->inp_gencnt;
1025 		if (syncache_respond(sc, m) == 0) {
1026 			TAILQ_REMOVE(&syncache_percpu->timerq[sc->sc_rxtslot],
1027 				     sc, sc_timerq);
1028 			syncache_timeout(syncache_percpu, sc, sc->sc_rxtslot);
1029 			tcpstat.tcps_sndacks++;
1030 			tcpstat.tcps_sndtotal++;
1031 		}
1032 		return (1);
1033 	}
1034 
1035 	/*
1036 	 * Fill in the syncache values.
1037 	 */
1038 	sc = kmalloc(sizeof(struct syncache), M_SYNCACHE, M_WAITOK|M_ZERO);
1039 	sc->sc_inp_gencnt = tp->t_inpcb->inp_gencnt;
1040 	sc->sc_ipopts = ipopts;
1041 	sc->sc_inc.inc_fport = inc->inc_fport;
1042 	sc->sc_inc.inc_lport = inc->inc_lport;
1043 	sc->sc_tp = tp;
1044 #ifdef INET6
1045 	sc->sc_inc.inc_isipv6 = inc->inc_isipv6;
1046 	if (inc->inc_isipv6) {
1047 		sc->sc_inc.inc6_faddr = inc->inc6_faddr;
1048 		sc->sc_inc.inc6_laddr = inc->inc6_laddr;
1049 		sc->sc_route6.ro_rt = NULL;
1050 	} else
1051 #endif
1052 	{
1053 		sc->sc_inc.inc_faddr = inc->inc_faddr;
1054 		sc->sc_inc.inc_laddr = inc->inc_laddr;
1055 		sc->sc_route.ro_rt = NULL;
1056 	}
1057 	sc->sc_irs = th->th_seq;
1058 	sc->sc_flags = 0;
1059 	sc->sc_peer_mss = to->to_flags & TOF_MSS ? to->to_mss : 0;
1060 	if (tcp_syncookies)
1061 		sc->sc_iss = syncookie_generate(sc);
1062 	else
1063 		sc->sc_iss = karc4random();
1064 
1065 	/* Initial receive window: clip ssb_space to [0 .. TCP_MAXWIN] */
1066 	win = ssb_space(&so->so_rcv);
1067 	win = imax(win, 0);
1068 	win = imin(win, TCP_MAXWIN);
1069 	sc->sc_wnd = win;
1070 
1071 	if (tcp_do_rfc1323) {
1072 		/*
1073 		 * A timestamp received in a SYN makes
1074 		 * it ok to send timestamp requests and replies.
1075 		 */
1076 		if (to->to_flags & TOF_TS) {
1077 			sc->sc_tsrecent = to->to_tsval;
1078 			sc->sc_flags |= SCF_TIMESTAMP;
1079 		}
1080 		if (to->to_flags & TOF_SCALE) {
1081 			int wscale = TCP_MIN_WINSHIFT;
1082 
1083 			/* Compute proper scaling value from buffer space */
1084 			while (wscale < TCP_MAX_WINSHIFT &&
1085 			    (TCP_MAXWIN << wscale) < so->so_rcv.ssb_hiwat) {
1086 				wscale++;
1087 			}
1088 			sc->sc_request_r_scale = wscale;
1089 			sc->sc_requested_s_scale = to->to_requested_s_scale;
1090 			sc->sc_flags |= SCF_WINSCALE;
1091 		}
1092 	}
1093 	if (tcp_do_sack && (to->to_flags & TOF_SACK_PERMITTED))
1094 		sc->sc_flags |= SCF_SACK_PERMITTED;
1095 	if (tp->t_flags & TF_NOOPT)
1096 		sc->sc_flags = SCF_NOOPT;
1097 #ifdef TCP_SIGNATURE
1098 	/*
1099 	 * If listening socket requested TCP digests, and received SYN
1100 	 * contains the option, flag this in the syncache so that
1101 	 * syncache_respond() will do the right thing with the SYN+ACK.
1102 	 * XXX Currently we always record the option by default and will
1103 	 * attempt to use it in syncache_respond().
1104 	 */
1105 	if (to->to_flags & TOF_SIGNATURE)
1106 		sc->sc_flags = SCF_SIGNATURE;
1107 #endif /* TCP_SIGNATURE */
1108 
1109 	if (syncache_respond(sc, m) == 0) {
1110 		syncache_insert(sc, sch);
1111 		tcpstat.tcps_sndacks++;
1112 		tcpstat.tcps_sndtotal++;
1113 	} else {
1114 		syncache_free(sc);
1115 		tcpstat.tcps_sc_dropped++;
1116 	}
1117 	return (1);
1118 }
1119 
1120 static int
1121 syncache_respond(struct syncache *sc, struct mbuf *m)
1122 {
1123 	u_int8_t *optp;
1124 	int optlen, error;
1125 	u_int16_t tlen, hlen, mssopt;
1126 	struct ip *ip = NULL;
1127 	struct rtentry *rt;
1128 	struct tcphdr *th;
1129 	struct ip6_hdr *ip6 = NULL;
1130 #ifdef INET6
1131 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
1132 #else
1133 	const boolean_t isipv6 = FALSE;
1134 #endif
1135 
1136 	if (isipv6) {
1137 		rt = tcp_rtlookup6(&sc->sc_inc);
1138 		if (rt != NULL)
1139 			mssopt = rt->rt_ifp->if_mtu -
1140 			     (sizeof(struct ip6_hdr) + sizeof(struct tcphdr));
1141 		else
1142 			mssopt = tcp_v6mssdflt;
1143 		hlen = sizeof(struct ip6_hdr);
1144 	} else {
1145 		rt = tcp_rtlookup(&sc->sc_inc);
1146 		if (rt != NULL)
1147 			mssopt = rt->rt_ifp->if_mtu -
1148 			     (sizeof(struct ip) + sizeof(struct tcphdr));
1149 		else
1150 			mssopt = tcp_mssdflt;
1151 		hlen = sizeof(struct ip);
1152 	}
1153 
1154 	/* Compute the size of the TCP options. */
1155 	if (sc->sc_flags & SCF_NOOPT) {
1156 		optlen = 0;
1157 	} else {
1158 		optlen = TCPOLEN_MAXSEG +
1159 		    ((sc->sc_flags & SCF_WINSCALE) ? 4 : 0) +
1160 		    ((sc->sc_flags & SCF_TIMESTAMP) ? TCPOLEN_TSTAMP_APPA : 0) +
1161 		    ((sc->sc_flags & SCF_SACK_PERMITTED) ?
1162 			TCPOLEN_SACK_PERMITTED_ALIGNED : 0);
1163 #ifdef TCP_SIGNATURE
1164 		optlen += ((sc->sc_flags & SCF_SIGNATURE) ?
1165 		    (TCPOLEN_SIGNATURE + 2) : 0);
1166 #endif /* TCP_SIGNATURE */
1167 	}
1168 	tlen = hlen + sizeof(struct tcphdr) + optlen;
1169 
1170 	/*
1171 	 * XXX
1172 	 * assume that the entire packet will fit in a header mbuf
1173 	 */
1174 	KASSERT(max_linkhdr + tlen <= MHLEN, ("syncache: mbuf too small"));
1175 
1176 	/*
1177 	 * XXX shouldn't this reuse the mbuf if possible ?
1178 	 * Create the IP+TCP header from scratch.
1179 	 */
1180 	if (m)
1181 		m_freem(m);
1182 
1183 	m = m_gethdr(MB_DONTWAIT, MT_HEADER);
1184 	if (m == NULL)
1185 		return (ENOBUFS);
1186 	m->m_data += max_linkhdr;
1187 	m->m_len = tlen;
1188 	m->m_pkthdr.len = tlen;
1189 	m->m_pkthdr.rcvif = NULL;
1190 
1191 	if (isipv6) {
1192 		ip6 = mtod(m, struct ip6_hdr *);
1193 		ip6->ip6_vfc = IPV6_VERSION;
1194 		ip6->ip6_nxt = IPPROTO_TCP;
1195 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1196 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1197 		ip6->ip6_plen = htons(tlen - hlen);
1198 		/* ip6_hlim is set after checksum */
1199 		/* ip6_flow = ??? */
1200 
1201 		th = (struct tcphdr *)(ip6 + 1);
1202 	} else {
1203 		ip = mtod(m, struct ip *);
1204 		ip->ip_v = IPVERSION;
1205 		ip->ip_hl = sizeof(struct ip) >> 2;
1206 		ip->ip_len = tlen;
1207 		ip->ip_id = 0;
1208 		ip->ip_off = 0;
1209 		ip->ip_sum = 0;
1210 		ip->ip_p = IPPROTO_TCP;
1211 		ip->ip_src = sc->sc_inc.inc_laddr;
1212 		ip->ip_dst = sc->sc_inc.inc_faddr;
1213 		ip->ip_ttl = sc->sc_tp->t_inpcb->inp_ip_ttl;   /* XXX */
1214 		ip->ip_tos = sc->sc_tp->t_inpcb->inp_ip_tos;   /* XXX */
1215 
1216 		/*
1217 		 * See if we should do MTU discovery.  Route lookups are
1218 		 * expensive, so we will only unset the DF bit if:
1219 		 *
1220 		 *	1) path_mtu_discovery is disabled
1221 		 *	2) the SCF_UNREACH flag has been set
1222 		 */
1223 		if (path_mtu_discovery
1224 		    && ((sc->sc_flags & SCF_UNREACH) == 0)) {
1225 		       ip->ip_off |= IP_DF;
1226 		}
1227 
1228 		th = (struct tcphdr *)(ip + 1);
1229 	}
1230 	th->th_sport = sc->sc_inc.inc_lport;
1231 	th->th_dport = sc->sc_inc.inc_fport;
1232 
1233 	th->th_seq = htonl(sc->sc_iss);
1234 	th->th_ack = htonl(sc->sc_irs + 1);
1235 	th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1236 	th->th_x2 = 0;
1237 	th->th_flags = TH_SYN | TH_ACK;
1238 	th->th_win = htons(sc->sc_wnd);
1239 	th->th_urp = 0;
1240 
1241 	/* Tack on the TCP options. */
1242 	if (optlen == 0)
1243 		goto no_options;
1244 	optp = (u_int8_t *)(th + 1);
1245 	*optp++ = TCPOPT_MAXSEG;
1246 	*optp++ = TCPOLEN_MAXSEG;
1247 	*optp++ = (mssopt >> 8) & 0xff;
1248 	*optp++ = mssopt & 0xff;
1249 
1250 	if (sc->sc_flags & SCF_WINSCALE) {
1251 		*((u_int32_t *)optp) = htonl(TCPOPT_NOP << 24 |
1252 		    TCPOPT_WINDOW << 16 | TCPOLEN_WINDOW << 8 |
1253 		    sc->sc_request_r_scale);
1254 		optp += 4;
1255 	}
1256 
1257 	if (sc->sc_flags & SCF_TIMESTAMP) {
1258 		u_int32_t *lp = (u_int32_t *)(optp);
1259 
1260 		/* Form timestamp option as shown in appendix A of RFC 1323. */
1261 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
1262 		*lp++ = htonl(ticks);
1263 		*lp   = htonl(sc->sc_tsrecent);
1264 		optp += TCPOLEN_TSTAMP_APPA;
1265 	}
1266 
1267 #ifdef TCP_SIGNATURE
1268 	/*
1269 	 * Handle TCP-MD5 passive opener response.
1270 	 */
1271 	if (sc->sc_flags & SCF_SIGNATURE) {
1272 		u_int8_t *bp = optp;
1273 		int i;
1274 
1275 		*bp++ = TCPOPT_SIGNATURE;
1276 		*bp++ = TCPOLEN_SIGNATURE;
1277 		for (i = 0; i < TCP_SIGLEN; i++)
1278 			*bp++ = 0;
1279 		tcpsignature_compute(m, 0, optlen,
1280 				optp + 2, IPSEC_DIR_OUTBOUND);
1281 		*bp++ = TCPOPT_NOP;
1282 		*bp++ = TCPOPT_EOL;
1283 		optp += TCPOLEN_SIGNATURE + 2;
1284 	}
1285 #endif /* TCP_SIGNATURE */
1286 
1287 	if (sc->sc_flags & SCF_SACK_PERMITTED) {
1288 		*((u_int32_t *)optp) = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
1289 		optp += TCPOLEN_SACK_PERMITTED_ALIGNED;
1290 	}
1291 
1292 no_options:
1293 	if (isipv6) {
1294 		struct route_in6 *ro6 = &sc->sc_route6;
1295 
1296 		th->th_sum = 0;
1297 		th->th_sum = in6_cksum(m, IPPROTO_TCP, hlen, tlen - hlen);
1298 		ip6->ip6_hlim = in6_selecthlim(NULL,
1299 		    ro6->ro_rt ? ro6->ro_rt->rt_ifp : NULL);
1300 		error = ip6_output(m, NULL, ro6, 0, NULL, NULL,
1301 				sc->sc_tp->t_inpcb);
1302 	} else {
1303 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1304 				       htons(tlen - hlen + IPPROTO_TCP));
1305 		m->m_pkthdr.csum_flags = CSUM_TCP;
1306 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1307 		error = ip_output(m, sc->sc_ipopts, &sc->sc_route,
1308 				  IP_DEBUGROUTE, NULL, sc->sc_tp->t_inpcb);
1309 	}
1310 	return (error);
1311 }
1312 
1313 /*
1314  * cookie layers:
1315  *
1316  *	|. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .|
1317  *	| peer iss                                                      |
1318  *	| MD5(laddr,faddr,secret,lport,fport)             |. . . . . . .|
1319  *	|                     0                       |(A)|             |
1320  * (A): peer mss index
1321  */
1322 
1323 /*
1324  * The values below are chosen to minimize the size of the tcp_secret
1325  * table, as well as providing roughly a 16 second lifetime for the cookie.
1326  */
1327 
1328 #define SYNCOOKIE_WNDBITS	5	/* exposed bits for window indexing */
1329 #define SYNCOOKIE_TIMESHIFT	1	/* scale ticks to window time units */
1330 
1331 #define SYNCOOKIE_WNDMASK	((1 << SYNCOOKIE_WNDBITS) - 1)
1332 #define SYNCOOKIE_NSECRETS	(1 << SYNCOOKIE_WNDBITS)
1333 #define SYNCOOKIE_TIMEOUT \
1334     (hz * (1 << SYNCOOKIE_WNDBITS) / (1 << SYNCOOKIE_TIMESHIFT))
1335 #define SYNCOOKIE_DATAMASK	((3 << SYNCOOKIE_WNDBITS) | SYNCOOKIE_WNDMASK)
1336 
1337 static struct {
1338 	u_int32_t	ts_secbits[4];
1339 	u_int		ts_expire;
1340 } tcp_secret[SYNCOOKIE_NSECRETS];
1341 
1342 static int tcp_msstab[] = { 0, 536, 1460, 8960 };
1343 
1344 static MD5_CTX syn_ctx;
1345 
1346 #define MD5Add(v)	MD5Update(&syn_ctx, (u_char *)&v, sizeof(v))
1347 
1348 struct md5_add {
1349 	u_int32_t laddr, faddr;
1350 	u_int32_t secbits[4];
1351 	u_int16_t lport, fport;
1352 };
1353 
1354 #ifdef CTASSERT
1355 CTASSERT(sizeof(struct md5_add) == 28);
1356 #endif
1357 
1358 /*
1359  * Consider the problem of a recreated (and retransmitted) cookie.  If the
1360  * original SYN was accepted, the connection is established.  The second
1361  * SYN is inflight, and if it arrives with an ISN that falls within the
1362  * receive window, the connection is killed.
1363  *
1364  * However, since cookies have other problems, this may not be worth
1365  * worrying about.
1366  */
1367 
1368 static u_int32_t
1369 syncookie_generate(struct syncache *sc)
1370 {
1371 	u_int32_t md5_buffer[4];
1372 	u_int32_t data;
1373 	int idx, i;
1374 	struct md5_add add;
1375 #ifdef INET6
1376 	const boolean_t isipv6 = sc->sc_inc.inc_isipv6;
1377 #else
1378 	const boolean_t isipv6 = FALSE;
1379 #endif
1380 
1381 	idx = ((ticks << SYNCOOKIE_TIMESHIFT) / hz) & SYNCOOKIE_WNDMASK;
1382 	if (tcp_secret[idx].ts_expire < ticks) {
1383 		for (i = 0; i < 4; i++)
1384 			tcp_secret[idx].ts_secbits[i] = karc4random();
1385 		tcp_secret[idx].ts_expire = ticks + SYNCOOKIE_TIMEOUT;
1386 	}
1387 	for (data = NELEM(tcp_msstab) - 1; data > 0; data--)
1388 		if (tcp_msstab[data] <= sc->sc_peer_mss)
1389 			break;
1390 	data = (data << SYNCOOKIE_WNDBITS) | idx;
1391 	data ^= sc->sc_irs;				/* peer's iss */
1392 	MD5Init(&syn_ctx);
1393 	if (isipv6) {
1394 		MD5Add(sc->sc_inc.inc6_laddr);
1395 		MD5Add(sc->sc_inc.inc6_faddr);
1396 		add.laddr = 0;
1397 		add.faddr = 0;
1398 	} else {
1399 		add.laddr = sc->sc_inc.inc_laddr.s_addr;
1400 		add.faddr = sc->sc_inc.inc_faddr.s_addr;
1401 	}
1402 	add.lport = sc->sc_inc.inc_lport;
1403 	add.fport = sc->sc_inc.inc_fport;
1404 	add.secbits[0] = tcp_secret[idx].ts_secbits[0];
1405 	add.secbits[1] = tcp_secret[idx].ts_secbits[1];
1406 	add.secbits[2] = tcp_secret[idx].ts_secbits[2];
1407 	add.secbits[3] = tcp_secret[idx].ts_secbits[3];
1408 	MD5Add(add);
1409 	MD5Final((u_char *)&md5_buffer, &syn_ctx);
1410 	data ^= (md5_buffer[0] & ~SYNCOOKIE_WNDMASK);
1411 	return (data);
1412 }
1413 
1414 static struct syncache *
1415 syncookie_lookup(struct in_conninfo *inc, struct tcphdr *th, struct socket *so)
1416 {
1417 	u_int32_t md5_buffer[4];
1418 	struct syncache *sc;
1419 	u_int32_t data;
1420 	int wnd, idx;
1421 	struct md5_add add;
1422 
1423 	data = (th->th_ack - 1) ^ (th->th_seq - 1);	/* remove ISS */
1424 	idx = data & SYNCOOKIE_WNDMASK;
1425 	if (tcp_secret[idx].ts_expire < ticks ||
1426 	    sototcpcb(so)->ts_recent + SYNCOOKIE_TIMEOUT < ticks)
1427 		return (NULL);
1428 	MD5Init(&syn_ctx);
1429 #ifdef INET6
1430 	if (inc->inc_isipv6) {
1431 		MD5Add(inc->inc6_laddr);
1432 		MD5Add(inc->inc6_faddr);
1433 		add.laddr = 0;
1434 		add.faddr = 0;
1435 	} else
1436 #endif
1437 	{
1438 		add.laddr = inc->inc_laddr.s_addr;
1439 		add.faddr = inc->inc_faddr.s_addr;
1440 	}
1441 	add.lport = inc->inc_lport;
1442 	add.fport = inc->inc_fport;
1443 	add.secbits[0] = tcp_secret[idx].ts_secbits[0];
1444 	add.secbits[1] = tcp_secret[idx].ts_secbits[1];
1445 	add.secbits[2] = tcp_secret[idx].ts_secbits[2];
1446 	add.secbits[3] = tcp_secret[idx].ts_secbits[3];
1447 	MD5Add(add);
1448 	MD5Final((u_char *)&md5_buffer, &syn_ctx);
1449 	data ^= md5_buffer[0];
1450 	if (data & ~SYNCOOKIE_DATAMASK)
1451 		return (NULL);
1452 	data = data >> SYNCOOKIE_WNDBITS;
1453 
1454 	/*
1455 	 * Fill in the syncache values.
1456 	 * XXX duplicate code from syncache_add
1457 	 */
1458 	sc = kmalloc(sizeof(struct syncache), M_SYNCACHE, M_WAITOK|M_ZERO);
1459 	sc->sc_ipopts = NULL;
1460 	sc->sc_inc.inc_fport = inc->inc_fport;
1461 	sc->sc_inc.inc_lport = inc->inc_lport;
1462 #ifdef INET6
1463 	sc->sc_inc.inc_isipv6 = inc->inc_isipv6;
1464 	if (inc->inc_isipv6) {
1465 		sc->sc_inc.inc6_faddr = inc->inc6_faddr;
1466 		sc->sc_inc.inc6_laddr = inc->inc6_laddr;
1467 		sc->sc_route6.ro_rt = NULL;
1468 	} else
1469 #endif
1470 	{
1471 		sc->sc_inc.inc_faddr = inc->inc_faddr;
1472 		sc->sc_inc.inc_laddr = inc->inc_laddr;
1473 		sc->sc_route.ro_rt = NULL;
1474 	}
1475 	sc->sc_irs = th->th_seq - 1;
1476 	sc->sc_iss = th->th_ack - 1;
1477 	wnd = ssb_space(&so->so_rcv);
1478 	wnd = imax(wnd, 0);
1479 	wnd = imin(wnd, TCP_MAXWIN);
1480 	sc->sc_wnd = wnd;
1481 	sc->sc_flags = 0;
1482 	sc->sc_rxtslot = 0;
1483 	sc->sc_peer_mss = tcp_msstab[data];
1484 	return (sc);
1485 }
1486