xref: /freebsd/sys/netinet/tcp_syncache.c (revision 84dfba8d183d31e3412639ecb4b8ad4433cf7e80)
1 /*-
2  * Copyright (c) 2001 McAfee, Inc.
3  * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG
4  * All rights reserved.
5  *
6  * This software was developed for the FreeBSD Project by Jonathan Lemon
7  * and McAfee Research, the Security Research Division of McAfee, Inc. under
8  * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
9  * DARPA CHATS research program. [2001 McAfee, Inc.]
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  */
32 
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 #include "opt_inet.h"
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 #include "opt_pcbgroup.h"
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/kernel.h>
44 #include <sys/sysctl.h>
45 #include <sys/limits.h>
46 #include <sys/lock.h>
47 #include <sys/mutex.h>
48 #include <sys/malloc.h>
49 #include <sys/mbuf.h>
50 #include <sys/proc.h>		/* for proc0 declaration */
51 #include <sys/random.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #include <sys/syslog.h>
55 #include <sys/ucred.h>
56 
57 #include <sys/md5.h>
58 #include <crypto/siphash/siphash.h>
59 
60 #include <vm/uma.h>
61 
62 #include <net/if.h>
63 #include <net/if_var.h>
64 #include <net/route.h>
65 #include <net/vnet.h>
66 
67 #include <netinet/in.h>
68 #include <netinet/in_systm.h>
69 #include <netinet/ip.h>
70 #include <netinet/in_var.h>
71 #include <netinet/in_pcb.h>
72 #include <netinet/ip_var.h>
73 #include <netinet/ip_options.h>
74 #ifdef INET6
75 #include <netinet/ip6.h>
76 #include <netinet/icmp6.h>
77 #include <netinet6/nd6.h>
78 #include <netinet6/ip6_var.h>
79 #include <netinet6/in6_pcb.h>
80 #endif
81 #include <netinet/tcp.h>
82 #include <netinet/tcp_fsm.h>
83 #include <netinet/tcp_seq.h>
84 #include <netinet/tcp_timer.h>
85 #include <netinet/tcp_var.h>
86 #include <netinet/tcp_syncache.h>
87 #ifdef INET6
88 #include <netinet6/tcp6_var.h>
89 #endif
90 #ifdef TCP_OFFLOAD
91 #include <netinet/toecore.h>
92 #endif
93 
94 #ifdef IPSEC
95 #include <netipsec/ipsec.h>
96 #ifdef INET6
97 #include <netipsec/ipsec6.h>
98 #endif
99 #include <netipsec/key.h>
100 #endif /*IPSEC*/
101 
102 #include <machine/in_cksum.h>
103 
104 #include <security/mac/mac_framework.h>
105 
106 static VNET_DEFINE(int, tcp_syncookies) = 1;
107 #define	V_tcp_syncookies		VNET(tcp_syncookies)
108 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_RW,
109     &VNET_NAME(tcp_syncookies), 0,
110     "Use TCP SYN cookies if the syncache overflows");
111 
112 static VNET_DEFINE(int, tcp_syncookiesonly) = 0;
113 #define	V_tcp_syncookiesonly		VNET(tcp_syncookiesonly)
114 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_RW,
115     &VNET_NAME(tcp_syncookiesonly), 0,
116     "Use only TCP SYN cookies");
117 
118 #ifdef TCP_OFFLOAD
119 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
120 #endif
121 
122 static void	 syncache_drop(struct syncache *, struct syncache_head *);
123 static void	 syncache_free(struct syncache *);
124 static void	 syncache_insert(struct syncache *, struct syncache_head *);
125 struct syncache *syncache_lookup(struct in_conninfo *, struct syncache_head **);
126 static int	 syncache_respond(struct syncache *);
127 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
128 		    struct mbuf *m);
129 static int	 syncache_sysctl_count(SYSCTL_HANDLER_ARGS);
130 static void	 syncache_timeout(struct syncache *sc, struct syncache_head *sch,
131 		    int docallout);
132 static void	 syncache_timer(void *);
133 
134 static uint32_t	 syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
135 		    uint8_t *, uintptr_t);
136 static tcp_seq	 syncookie_generate(struct syncache_head *, struct syncache *);
137 static struct syncache
138 		*syncookie_lookup(struct in_conninfo *, struct syncache_head *,
139 		    struct syncache *, struct tcphdr *, struct tcpopt *,
140 		    struct socket *);
141 static void	 syncookie_reseed(void *);
142 #ifdef INVARIANTS
143 static int	 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
144 		    struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
145 		    struct socket *lso);
146 #endif
147 
148 /*
149  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
150  * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds,
151  * the odds are that the user has given up attempting to connect by then.
152  */
153 #define SYNCACHE_MAXREXMTS		3
154 
155 /* Arbitrary values */
156 #define TCP_SYNCACHE_HASHSIZE		512
157 #define TCP_SYNCACHE_BUCKETLIMIT	30
158 
159 static VNET_DEFINE(struct tcp_syncache, tcp_syncache);
160 #define	V_tcp_syncache			VNET(tcp_syncache)
161 
162 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0,
163     "TCP SYN cache");
164 
165 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_RDTUN,
166     &VNET_NAME(tcp_syncache.bucket_limit), 0,
167     "Per-bucket hash limit for syncache");
168 
169 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_RDTUN,
170     &VNET_NAME(tcp_syncache.cache_limit), 0,
171     "Overall entry limit for syncache");
172 
173 SYSCTL_VNET_PROC(_net_inet_tcp_syncache, OID_AUTO, count, (CTLTYPE_UINT|CTLFLAG_RD),
174     NULL, 0, &syncache_sysctl_count, "IU",
175     "Current number of entries in syncache");
176 
177 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_RDTUN,
178     &VNET_NAME(tcp_syncache.hashsize), 0,
179     "Size of TCP syncache hashtable");
180 
181 SYSCTL_VNET_UINT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_RW,
182     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
183     "Limit on SYN/ACK retransmissions");
184 
185 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
186 SYSCTL_VNET_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
187     CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
188     "Send reset on socket allocation failure");
189 
190 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
191 
192 #define SYNCACHE_HASH(inc, mask)					\
193 	((V_tcp_syncache.hash_secret ^					\
194 	  (inc)->inc_faddr.s_addr ^					\
195 	  ((inc)->inc_faddr.s_addr >> 16) ^				\
196 	  (inc)->inc_fport ^ (inc)->inc_lport) & mask)
197 
198 #define SYNCACHE_HASH6(inc, mask)					\
199 	((V_tcp_syncache.hash_secret ^					\
200 	  (inc)->inc6_faddr.s6_addr32[0] ^				\
201 	  (inc)->inc6_faddr.s6_addr32[3] ^				\
202 	  (inc)->inc_fport ^ (inc)->inc_lport) & mask)
203 
204 #define ENDPTS_EQ(a, b) (						\
205 	(a)->ie_fport == (b)->ie_fport &&				\
206 	(a)->ie_lport == (b)->ie_lport &&				\
207 	(a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr &&			\
208 	(a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr			\
209 )
210 
211 #define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0)
212 
213 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
214 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
215 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
216 
217 /*
218  * Requires the syncache entry to be already removed from the bucket list.
219  */
220 static void
221 syncache_free(struct syncache *sc)
222 {
223 
224 	if (sc->sc_ipopts)
225 		(void) m_free(sc->sc_ipopts);
226 	if (sc->sc_cred)
227 		crfree(sc->sc_cred);
228 #ifdef MAC
229 	mac_syncache_destroy(&sc->sc_label);
230 #endif
231 
232 	uma_zfree(V_tcp_syncache.zone, sc);
233 }
234 
235 void
236 syncache_init(void)
237 {
238 	int i;
239 
240 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
241 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
242 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
243 	V_tcp_syncache.hash_secret = arc4random();
244 
245 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
246 	    &V_tcp_syncache.hashsize);
247 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
248 	    &V_tcp_syncache.bucket_limit);
249 	if (!powerof2(V_tcp_syncache.hashsize) ||
250 	    V_tcp_syncache.hashsize == 0) {
251 		printf("WARNING: syncache hash size is not a power of 2.\n");
252 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
253 	}
254 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
255 
256 	/* Set limits. */
257 	V_tcp_syncache.cache_limit =
258 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
259 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
260 	    &V_tcp_syncache.cache_limit);
261 
262 	/* Allocate the hash table. */
263 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
264 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
265 
266 #ifdef VIMAGE
267 	V_tcp_syncache.vnet = curvnet;
268 #endif
269 
270 	/* Initialize the hash buckets. */
271 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
272 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
273 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
274 			 NULL, MTX_DEF);
275 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
276 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
277 		V_tcp_syncache.hashbase[i].sch_length = 0;
278 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
279 	}
280 
281 	/* Create the syncache entry zone. */
282 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
283 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
284 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
285 	    V_tcp_syncache.cache_limit);
286 
287 	/* Start the SYN cookie reseeder callout. */
288 	callout_init(&V_tcp_syncache.secret.reseed, 1);
289 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
290 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
291 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
292 	    syncookie_reseed, &V_tcp_syncache);
293 }
294 
295 #ifdef VIMAGE
296 void
297 syncache_destroy(void)
298 {
299 	struct syncache_head *sch;
300 	struct syncache *sc, *nsc;
301 	int i;
302 
303 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
304 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
305 
306 		sch = &V_tcp_syncache.hashbase[i];
307 		callout_drain(&sch->sch_timer);
308 
309 		SCH_LOCK(sch);
310 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
311 			syncache_drop(sc, sch);
312 		SCH_UNLOCK(sch);
313 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
314 		    ("%s: sch->sch_bucket not empty", __func__));
315 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
316 		    __func__, sch->sch_length));
317 		mtx_destroy(&sch->sch_mtx);
318 	}
319 
320 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
321 	    ("%s: cache_count not 0", __func__));
322 
323 	/* Free the allocated global resources. */
324 	uma_zdestroy(V_tcp_syncache.zone);
325 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
326 
327 	callout_drain(&V_tcp_syncache.secret.reseed);
328 }
329 #endif
330 
331 static int
332 syncache_sysctl_count(SYSCTL_HANDLER_ARGS)
333 {
334 	int count;
335 
336 	count = uma_zone_get_cur(V_tcp_syncache.zone);
337 	return (sysctl_handle_int(oidp, &count, 0, req));
338 }
339 
340 /*
341  * Inserts a syncache entry into the specified bucket row.
342  * Locks and unlocks the syncache_head autonomously.
343  */
344 static void
345 syncache_insert(struct syncache *sc, struct syncache_head *sch)
346 {
347 	struct syncache *sc2;
348 
349 	SCH_LOCK(sch);
350 
351 	/*
352 	 * Make sure that we don't overflow the per-bucket limit.
353 	 * If the bucket is full, toss the oldest element.
354 	 */
355 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
356 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
357 			("sch->sch_length incorrect"));
358 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
359 		syncache_drop(sc2, sch);
360 		TCPSTAT_INC(tcps_sc_bucketoverflow);
361 	}
362 
363 	/* Put it into the bucket. */
364 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
365 	sch->sch_length++;
366 
367 #ifdef TCP_OFFLOAD
368 	if (ADDED_BY_TOE(sc)) {
369 		struct toedev *tod = sc->sc_tod;
370 
371 		tod->tod_syncache_added(tod, sc->sc_todctx);
372 	}
373 #endif
374 
375 	/* Reinitialize the bucket row's timer. */
376 	if (sch->sch_length == 1)
377 		sch->sch_nextc = ticks + INT_MAX;
378 	syncache_timeout(sc, sch, 1);
379 
380 	SCH_UNLOCK(sch);
381 
382 	TCPSTAT_INC(tcps_sc_added);
383 }
384 
385 /*
386  * Remove and free entry from syncache bucket row.
387  * Expects locked syncache head.
388  */
389 static void
390 syncache_drop(struct syncache *sc, struct syncache_head *sch)
391 {
392 
393 	SCH_LOCK_ASSERT(sch);
394 
395 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
396 	sch->sch_length--;
397 
398 #ifdef TCP_OFFLOAD
399 	if (ADDED_BY_TOE(sc)) {
400 		struct toedev *tod = sc->sc_tod;
401 
402 		tod->tod_syncache_removed(tod, sc->sc_todctx);
403 	}
404 #endif
405 
406 	syncache_free(sc);
407 }
408 
409 /*
410  * Engage/reengage time on bucket row.
411  */
412 static void
413 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
414 {
415 	sc->sc_rxttime = ticks +
416 		TCPTV_RTOBASE * (tcp_syn_backoff[sc->sc_rxmits]);
417 	sc->sc_rxmits++;
418 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
419 		sch->sch_nextc = sc->sc_rxttime;
420 		if (docallout)
421 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
422 			    syncache_timer, (void *)sch);
423 	}
424 }
425 
426 /*
427  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
428  * If we have retransmitted an entry the maximum number of times, expire it.
429  * One separate timer for each bucket row.
430  */
431 static void
432 syncache_timer(void *xsch)
433 {
434 	struct syncache_head *sch = (struct syncache_head *)xsch;
435 	struct syncache *sc, *nsc;
436 	int tick = ticks;
437 	char *s;
438 
439 	CURVNET_SET(sch->sch_sc->vnet);
440 
441 	/* NB: syncache_head has already been locked by the callout. */
442 	SCH_LOCK_ASSERT(sch);
443 
444 	/*
445 	 * In the following cycle we may remove some entries and/or
446 	 * advance some timeouts, so re-initialize the bucket timer.
447 	 */
448 	sch->sch_nextc = tick + INT_MAX;
449 
450 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
451 		/*
452 		 * We do not check if the listen socket still exists
453 		 * and accept the case where the listen socket may be
454 		 * gone by the time we resend the SYN/ACK.  We do
455 		 * not expect this to happens often. If it does,
456 		 * then the RST will be sent by the time the remote
457 		 * host does the SYN/ACK->ACK.
458 		 */
459 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
460 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
461 				sch->sch_nextc = sc->sc_rxttime;
462 			continue;
463 		}
464 		if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
465 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
466 				log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
467 				    "giving up and removing syncache entry\n",
468 				    s, __func__);
469 				free(s, M_TCPLOG);
470 			}
471 			syncache_drop(sc, sch);
472 			TCPSTAT_INC(tcps_sc_stale);
473 			continue;
474 		}
475 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
476 			log(LOG_DEBUG, "%s; %s: Response timeout, "
477 			    "retransmitting (%u) SYN|ACK\n",
478 			    s, __func__, sc->sc_rxmits);
479 			free(s, M_TCPLOG);
480 		}
481 
482 		(void) syncache_respond(sc);
483 		TCPSTAT_INC(tcps_sc_retransmitted);
484 		syncache_timeout(sc, sch, 0);
485 	}
486 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
487 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
488 			syncache_timer, (void *)(sch));
489 	CURVNET_RESTORE();
490 }
491 
492 /*
493  * Find an entry in the syncache.
494  * Returns always with locked syncache_head plus a matching entry or NULL.
495  */
496 struct syncache *
497 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
498 {
499 	struct syncache *sc;
500 	struct syncache_head *sch;
501 
502 #ifdef INET6
503 	if (inc->inc_flags & INC_ISIPV6) {
504 		sch = &V_tcp_syncache.hashbase[
505 		    SYNCACHE_HASH6(inc, V_tcp_syncache.hashmask)];
506 		*schp = sch;
507 
508 		SCH_LOCK(sch);
509 
510 		/* Circle through bucket row to find matching entry. */
511 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
512 			if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
513 				return (sc);
514 		}
515 	} else
516 #endif
517 	{
518 		sch = &V_tcp_syncache.hashbase[
519 		    SYNCACHE_HASH(inc, V_tcp_syncache.hashmask)];
520 		*schp = sch;
521 
522 		SCH_LOCK(sch);
523 
524 		/* Circle through bucket row to find matching entry. */
525 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
526 #ifdef INET6
527 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
528 				continue;
529 #endif
530 			if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
531 				return (sc);
532 		}
533 	}
534 	SCH_LOCK_ASSERT(*schp);
535 	return (NULL);			/* always returns with locked sch */
536 }
537 
538 /*
539  * This function is called when we get a RST for a
540  * non-existent connection, so that we can see if the
541  * connection is in the syn cache.  If it is, zap it.
542  */
543 void
544 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
545 {
546 	struct syncache *sc;
547 	struct syncache_head *sch;
548 	char *s = NULL;
549 
550 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
551 	SCH_LOCK_ASSERT(sch);
552 
553 	/*
554 	 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
555 	 * See RFC 793 page 65, section SEGMENT ARRIVES.
556 	 */
557 	if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
558 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
559 			log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
560 			    "FIN flag set, segment ignored\n", s, __func__);
561 		TCPSTAT_INC(tcps_badrst);
562 		goto done;
563 	}
564 
565 	/*
566 	 * No corresponding connection was found in syncache.
567 	 * If syncookies are enabled and possibly exclusively
568 	 * used, or we are under memory pressure, a valid RST
569 	 * may not find a syncache entry.  In that case we're
570 	 * done and no SYN|ACK retransmissions will happen.
571 	 * Otherwise the RST was misdirected or spoofed.
572 	 */
573 	if (sc == NULL) {
574 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
575 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
576 			    "syncache entry (possibly syncookie only), "
577 			    "segment ignored\n", s, __func__);
578 		TCPSTAT_INC(tcps_badrst);
579 		goto done;
580 	}
581 
582 	/*
583 	 * If the RST bit is set, check the sequence number to see
584 	 * if this is a valid reset segment.
585 	 * RFC 793 page 37:
586 	 *   In all states except SYN-SENT, all reset (RST) segments
587 	 *   are validated by checking their SEQ-fields.  A reset is
588 	 *   valid if its sequence number is in the window.
589 	 *
590 	 *   The sequence number in the reset segment is normally an
591 	 *   echo of our outgoing acknowlegement numbers, but some hosts
592 	 *   send a reset with the sequence number at the rightmost edge
593 	 *   of our receive window, and we have to handle this case.
594 	 */
595 	if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
596 	    SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
597 		syncache_drop(sc, sch);
598 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
599 			log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, "
600 			    "connection attempt aborted by remote endpoint\n",
601 			    s, __func__);
602 		TCPSTAT_INC(tcps_sc_reset);
603 	} else {
604 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
605 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
606 			    "IRS %u (+WND %u), segment ignored\n",
607 			    s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd);
608 		TCPSTAT_INC(tcps_badrst);
609 	}
610 
611 done:
612 	if (s != NULL)
613 		free(s, M_TCPLOG);
614 	SCH_UNLOCK(sch);
615 }
616 
617 void
618 syncache_badack(struct in_conninfo *inc)
619 {
620 	struct syncache *sc;
621 	struct syncache_head *sch;
622 
623 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
624 	SCH_LOCK_ASSERT(sch);
625 	if (sc != NULL) {
626 		syncache_drop(sc, sch);
627 		TCPSTAT_INC(tcps_sc_badack);
628 	}
629 	SCH_UNLOCK(sch);
630 }
631 
632 void
633 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th)
634 {
635 	struct syncache *sc;
636 	struct syncache_head *sch;
637 
638 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
639 	SCH_LOCK_ASSERT(sch);
640 	if (sc == NULL)
641 		goto done;
642 
643 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
644 	if (ntohl(th->th_seq) != sc->sc_iss)
645 		goto done;
646 
647 	/*
648 	 * If we've rertransmitted 3 times and this is our second error,
649 	 * we remove the entry.  Otherwise, we allow it to continue on.
650 	 * This prevents us from incorrectly nuking an entry during a
651 	 * spurious network outage.
652 	 *
653 	 * See tcp_notify().
654 	 */
655 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
656 		sc->sc_flags |= SCF_UNREACH;
657 		goto done;
658 	}
659 	syncache_drop(sc, sch);
660 	TCPSTAT_INC(tcps_sc_unreach);
661 done:
662 	SCH_UNLOCK(sch);
663 }
664 
665 /*
666  * Build a new TCP socket structure from a syncache entry.
667  */
668 static struct socket *
669 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
670 {
671 	struct inpcb *inp = NULL;
672 	struct socket *so;
673 	struct tcpcb *tp;
674 	int error;
675 	char *s;
676 
677 	INP_INFO_WLOCK_ASSERT(&V_tcbinfo);
678 
679 	/*
680 	 * Ok, create the full blown connection, and set things up
681 	 * as they would have been set up if we had created the
682 	 * connection when the SYN arrived.  If we can't create
683 	 * the connection, abort it.
684 	 */
685 	so = sonewconn(lso, SS_ISCONNECTED);
686 	if (so == NULL) {
687 		/*
688 		 * Drop the connection; we will either send a RST or
689 		 * have the peer retransmit its SYN again after its
690 		 * RTO and try again.
691 		 */
692 		TCPSTAT_INC(tcps_listendrop);
693 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
694 			log(LOG_DEBUG, "%s; %s: Socket create failed "
695 			    "due to limits or memory shortage\n",
696 			    s, __func__);
697 			free(s, M_TCPLOG);
698 		}
699 		goto abort2;
700 	}
701 #ifdef MAC
702 	mac_socketpeer_set_from_mbuf(m, so);
703 #endif
704 
705 	inp = sotoinpcb(so);
706 	inp->inp_inc.inc_fibnum = so->so_fibnum;
707 	INP_WLOCK(inp);
708 	INP_HASH_WLOCK(&V_tcbinfo);
709 
710 	/* Insert new socket into PCB hash list. */
711 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
712 #ifdef INET6
713 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
714 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
715 	} else {
716 		inp->inp_vflag &= ~INP_IPV6;
717 		inp->inp_vflag |= INP_IPV4;
718 #endif
719 		inp->inp_laddr = sc->sc_inc.inc_laddr;
720 #ifdef INET6
721 	}
722 #endif
723 
724 	/*
725 	 * Install in the reservation hash table for now, but don't yet
726 	 * install a connection group since the full 4-tuple isn't yet
727 	 * configured.
728 	 */
729 	inp->inp_lport = sc->sc_inc.inc_lport;
730 	if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
731 		/*
732 		 * Undo the assignments above if we failed to
733 		 * put the PCB on the hash lists.
734 		 */
735 #ifdef INET6
736 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
737 			inp->in6p_laddr = in6addr_any;
738 		else
739 #endif
740 			inp->inp_laddr.s_addr = INADDR_ANY;
741 		inp->inp_lport = 0;
742 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
743 			log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
744 			    "with error %i\n",
745 			    s, __func__, error);
746 			free(s, M_TCPLOG);
747 		}
748 		INP_HASH_WUNLOCK(&V_tcbinfo);
749 		goto abort;
750 	}
751 #ifdef IPSEC
752 	/* Copy old policy into new socket's. */
753 	if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp))
754 		printf("syncache_socket: could not copy policy\n");
755 #endif
756 #ifdef INET6
757 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
758 		struct inpcb *oinp = sotoinpcb(lso);
759 		struct in6_addr laddr6;
760 		struct sockaddr_in6 sin6;
761 		/*
762 		 * Inherit socket options from the listening socket.
763 		 * Note that in6p_inputopts are not (and should not be)
764 		 * copied, since it stores previously received options and is
765 		 * used to detect if each new option is different than the
766 		 * previous one and hence should be passed to a user.
767 		 * If we copied in6p_inputopts, a user would not be able to
768 		 * receive options just after calling the accept system call.
769 		 */
770 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
771 		if (oinp->in6p_outputopts)
772 			inp->in6p_outputopts =
773 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
774 
775 		sin6.sin6_family = AF_INET6;
776 		sin6.sin6_len = sizeof(sin6);
777 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
778 		sin6.sin6_port = sc->sc_inc.inc_fport;
779 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
780 		laddr6 = inp->in6p_laddr;
781 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
782 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
783 		if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
784 		    thread0.td_ucred, m)) != 0) {
785 			inp->in6p_laddr = laddr6;
786 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
787 				log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
788 				    "with error %i\n",
789 				    s, __func__, error);
790 				free(s, M_TCPLOG);
791 			}
792 			INP_HASH_WUNLOCK(&V_tcbinfo);
793 			goto abort;
794 		}
795 		/* Override flowlabel from in6_pcbconnect. */
796 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
797 		inp->inp_flow |= sc->sc_flowlabel;
798 	}
799 #endif /* INET6 */
800 #if defined(INET) && defined(INET6)
801 	else
802 #endif
803 #ifdef INET
804 	{
805 		struct in_addr laddr;
806 		struct sockaddr_in sin;
807 
808 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
809 
810 		if (inp->inp_options == NULL) {
811 			inp->inp_options = sc->sc_ipopts;
812 			sc->sc_ipopts = NULL;
813 		}
814 
815 		sin.sin_family = AF_INET;
816 		sin.sin_len = sizeof(sin);
817 		sin.sin_addr = sc->sc_inc.inc_faddr;
818 		sin.sin_port = sc->sc_inc.inc_fport;
819 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
820 		laddr = inp->inp_laddr;
821 		if (inp->inp_laddr.s_addr == INADDR_ANY)
822 			inp->inp_laddr = sc->sc_inc.inc_laddr;
823 		if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
824 		    thread0.td_ucred, m)) != 0) {
825 			inp->inp_laddr = laddr;
826 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
827 				log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
828 				    "with error %i\n",
829 				    s, __func__, error);
830 				free(s, M_TCPLOG);
831 			}
832 			INP_HASH_WUNLOCK(&V_tcbinfo);
833 			goto abort;
834 		}
835 	}
836 #endif /* INET */
837 	INP_HASH_WUNLOCK(&V_tcbinfo);
838 	tp = intotcpcb(inp);
839 	tcp_state_change(tp, TCPS_SYN_RECEIVED);
840 	tp->iss = sc->sc_iss;
841 	tp->irs = sc->sc_irs;
842 	tcp_rcvseqinit(tp);
843 	tcp_sendseqinit(tp);
844 	tp->snd_wl1 = sc->sc_irs;
845 	tp->snd_max = tp->iss + 1;
846 	tp->snd_nxt = tp->iss + 1;
847 	tp->rcv_up = sc->sc_irs + 1;
848 	tp->rcv_wnd = sc->sc_wnd;
849 	tp->rcv_adv += tp->rcv_wnd;
850 	tp->last_ack_sent = tp->rcv_nxt;
851 
852 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
853 	if (sc->sc_flags & SCF_NOOPT)
854 		tp->t_flags |= TF_NOOPT;
855 	else {
856 		if (sc->sc_flags & SCF_WINSCALE) {
857 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
858 			tp->snd_scale = sc->sc_requested_s_scale;
859 			tp->request_r_scale = sc->sc_requested_r_scale;
860 		}
861 		if (sc->sc_flags & SCF_TIMESTAMP) {
862 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
863 			tp->ts_recent = sc->sc_tsreflect;
864 			tp->ts_recent_age = tcp_ts_getticks();
865 			tp->ts_offset = sc->sc_tsoff;
866 		}
867 #ifdef TCP_SIGNATURE
868 		if (sc->sc_flags & SCF_SIGNATURE)
869 			tp->t_flags |= TF_SIGNATURE;
870 #endif
871 		if (sc->sc_flags & SCF_SACK)
872 			tp->t_flags |= TF_SACK_PERMIT;
873 	}
874 
875 	if (sc->sc_flags & SCF_ECN)
876 		tp->t_flags |= TF_ECN_PERMIT;
877 
878 	/*
879 	 * Set up MSS and get cached values from tcp_hostcache.
880 	 * This might overwrite some of the defaults we just set.
881 	 */
882 	tcp_mss(tp, sc->sc_peer_mss);
883 
884 	/*
885 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
886 	 * limited to one segment in cc_conn_init().
887 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
888 	 */
889 	if (sc->sc_rxmits > 1)
890 		tp->snd_cwnd = 1;
891 
892 #ifdef TCP_OFFLOAD
893 	/*
894 	 * Allow a TOE driver to install its hooks.  Note that we hold the
895 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
896 	 * new connection before the TOE driver has done its thing.
897 	 */
898 	if (ADDED_BY_TOE(sc)) {
899 		struct toedev *tod = sc->sc_tod;
900 
901 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
902 	}
903 #endif
904 	/*
905 	 * Copy and activate timers.
906 	 */
907 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
908 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
909 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
910 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
911 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
912 
913 	INP_WUNLOCK(inp);
914 
915 	TCPSTAT_INC(tcps_accepts);
916 	return (so);
917 
918 abort:
919 	INP_WUNLOCK(inp);
920 abort2:
921 	if (so != NULL)
922 		soabort(so);
923 	return (NULL);
924 }
925 
926 /*
927  * This function gets called when we receive an ACK for a
928  * socket in the LISTEN state.  We look up the connection
929  * in the syncache, and if its there, we pull it out of
930  * the cache and turn it into a full-blown connection in
931  * the SYN-RECEIVED state.
932  */
933 int
934 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
935     struct socket **lsop, struct mbuf *m)
936 {
937 	struct syncache *sc;
938 	struct syncache_head *sch;
939 	struct syncache scs;
940 	char *s;
941 
942 	/*
943 	 * Global TCP locks are held because we manipulate the PCB lists
944 	 * and create a new socket.
945 	 */
946 	INP_INFO_WLOCK_ASSERT(&V_tcbinfo);
947 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
948 	    ("%s: can handle only ACK", __func__));
949 
950 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
951 	SCH_LOCK_ASSERT(sch);
952 
953 #ifdef INVARIANTS
954 	/*
955 	 * Test code for syncookies comparing the syncache stored
956 	 * values with the reconstructed values from the cookie.
957 	 */
958 	if (sc != NULL)
959 		syncookie_cmp(inc, sch, sc, th, to, *lsop);
960 #endif
961 
962 	if (sc == NULL) {
963 		/*
964 		 * There is no syncache entry, so see if this ACK is
965 		 * a returning syncookie.  To do this, first:
966 		 *  A. See if this socket has had a syncache entry dropped in
967 		 *     the past.  We don't want to accept a bogus syncookie
968 		 *     if we've never received a SYN.
969 		 *  B. check that the syncookie is valid.  If it is, then
970 		 *     cobble up a fake syncache entry, and return.
971 		 */
972 		if (!V_tcp_syncookies) {
973 			SCH_UNLOCK(sch);
974 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
975 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
976 				    "segment rejected (syncookies disabled)\n",
977 				    s, __func__);
978 			goto failed;
979 		}
980 		bzero(&scs, sizeof(scs));
981 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
982 		SCH_UNLOCK(sch);
983 		if (sc == NULL) {
984 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
985 				log(LOG_DEBUG, "%s; %s: Segment failed "
986 				    "SYNCOOKIE authentication, segment rejected "
987 				    "(probably spoofed)\n", s, __func__);
988 			goto failed;
989 		}
990 	} else {
991 		/* Pull out the entry to unlock the bucket row. */
992 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
993 		sch->sch_length--;
994 #ifdef TCP_OFFLOAD
995 		if (ADDED_BY_TOE(sc)) {
996 			struct toedev *tod = sc->sc_tod;
997 
998 			tod->tod_syncache_removed(tod, sc->sc_todctx);
999 		}
1000 #endif
1001 		SCH_UNLOCK(sch);
1002 	}
1003 
1004 	/*
1005 	 * Segment validation:
1006 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1007 	 */
1008 	if (th->th_ack != sc->sc_iss + 1) {
1009 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1010 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1011 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1012 		goto failed;
1013 	}
1014 
1015 	/*
1016 	 * The SEQ must fall in the window starting at the received
1017 	 * initial receive sequence number + 1 (the SYN).
1018 	 */
1019 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1020 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1021 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1022 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1023 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1024 		goto failed;
1025 	}
1026 
1027 	/*
1028 	 * If timestamps were not negotiated during SYN/ACK they
1029 	 * must not appear on any segment during this session.
1030 	 */
1031 	if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1032 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1033 			log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1034 			    "segment rejected\n", s, __func__);
1035 		goto failed;
1036 	}
1037 
1038 	/*
1039 	 * If timestamps were negotiated during SYN/ACK they should
1040 	 * appear on every segment during this session.
1041 	 * XXXAO: This is only informal as there have been unverified
1042 	 * reports of non-compliants stacks.
1043 	 */
1044 	if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1045 		if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1046 			log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1047 			    "no action\n", s, __func__);
1048 			free(s, M_TCPLOG);
1049 			s = NULL;
1050 		}
1051 	}
1052 
1053 	/*
1054 	 * If timestamps were negotiated the reflected timestamp
1055 	 * must be equal to what we actually sent in the SYN|ACK.
1056 	 */
1057 	if ((to->to_flags & TOF_TS) && to->to_tsecr != sc->sc_ts) {
1058 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1059 			log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, "
1060 			    "segment rejected\n",
1061 			    s, __func__, to->to_tsecr, sc->sc_ts);
1062 		goto failed;
1063 	}
1064 
1065 	*lsop = syncache_socket(sc, *lsop, m);
1066 
1067 	if (*lsop == NULL)
1068 		TCPSTAT_INC(tcps_sc_aborted);
1069 	else
1070 		TCPSTAT_INC(tcps_sc_completed);
1071 
1072 /* how do we find the inp for the new socket? */
1073 	if (sc != &scs)
1074 		syncache_free(sc);
1075 	return (1);
1076 failed:
1077 	if (sc != NULL && sc != &scs)
1078 		syncache_free(sc);
1079 	if (s != NULL)
1080 		free(s, M_TCPLOG);
1081 	*lsop = NULL;
1082 	return (0);
1083 }
1084 
1085 /*
1086  * Given a LISTEN socket and an inbound SYN request, add
1087  * this to the syn cache, and send back a segment:
1088  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1089  * to the source.
1090  *
1091  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1092  * Doing so would require that we hold onto the data and deliver it
1093  * to the application.  However, if we are the target of a SYN-flood
1094  * DoS attack, an attacker could send data which would eventually
1095  * consume all available buffer space if it were ACKed.  By not ACKing
1096  * the data, we avoid this DoS scenario.
1097  */
1098 void
1099 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1100     struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1101     void *todctx)
1102 {
1103 	struct tcpcb *tp;
1104 	struct socket *so;
1105 	struct syncache *sc = NULL;
1106 	struct syncache_head *sch;
1107 	struct mbuf *ipopts = NULL;
1108 	u_int ltflags;
1109 	int win, sb_hiwat, ip_ttl, ip_tos;
1110 	char *s;
1111 #ifdef INET6
1112 	int autoflowlabel = 0;
1113 #endif
1114 #ifdef MAC
1115 	struct label *maclabel;
1116 #endif
1117 	struct syncache scs;
1118 	struct ucred *cred;
1119 
1120 	INP_INFO_WLOCK_ASSERT(&V_tcbinfo);
1121 	INP_WLOCK_ASSERT(inp);			/* listen socket */
1122 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1123 	    ("%s: unexpected tcp flags", __func__));
1124 
1125 	/*
1126 	 * Combine all so/tp operations very early to drop the INP lock as
1127 	 * soon as possible.
1128 	 */
1129 	so = *lsop;
1130 	tp = sototcpcb(so);
1131 	cred = crhold(so->so_cred);
1132 
1133 #ifdef INET6
1134 	if ((inc->inc_flags & INC_ISIPV6) &&
1135 	    (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1136 		autoflowlabel = 1;
1137 #endif
1138 	ip_ttl = inp->inp_ip_ttl;
1139 	ip_tos = inp->inp_ip_tos;
1140 	win = sbspace(&so->so_rcv);
1141 	sb_hiwat = so->so_rcv.sb_hiwat;
1142 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1143 
1144 	/* By the time we drop the lock these should no longer be used. */
1145 	so = NULL;
1146 	tp = NULL;
1147 
1148 #ifdef MAC
1149 	if (mac_syncache_init(&maclabel) != 0) {
1150 		INP_WUNLOCK(inp);
1151 		INP_INFO_WUNLOCK(&V_tcbinfo);
1152 		goto done;
1153 	} else
1154 		mac_syncache_create(maclabel, inp);
1155 #endif
1156 	INP_WUNLOCK(inp);
1157 	INP_INFO_WUNLOCK(&V_tcbinfo);
1158 
1159 	/*
1160 	 * Remember the IP options, if any.
1161 	 */
1162 #ifdef INET6
1163 	if (!(inc->inc_flags & INC_ISIPV6))
1164 #endif
1165 #ifdef INET
1166 		ipopts = (m) ? ip_srcroute(m) : NULL;
1167 #else
1168 		ipopts = NULL;
1169 #endif
1170 
1171 	/*
1172 	 * See if we already have an entry for this connection.
1173 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1174 	 *
1175 	 * XXX: should the syncache be re-initialized with the contents
1176 	 * of the new SYN here (which may have different options?)
1177 	 *
1178 	 * XXX: We do not check the sequence number to see if this is a
1179 	 * real retransmit or a new connection attempt.  The question is
1180 	 * how to handle such a case; either ignore it as spoofed, or
1181 	 * drop the current entry and create a new one?
1182 	 */
1183 	sc = syncache_lookup(inc, &sch);	/* returns locked entry */
1184 	SCH_LOCK_ASSERT(sch);
1185 	if (sc != NULL) {
1186 		TCPSTAT_INC(tcps_sc_dupsyn);
1187 		if (ipopts) {
1188 			/*
1189 			 * If we were remembering a previous source route,
1190 			 * forget it and use the new one we've been given.
1191 			 */
1192 			if (sc->sc_ipopts)
1193 				(void) m_free(sc->sc_ipopts);
1194 			sc->sc_ipopts = ipopts;
1195 		}
1196 		/*
1197 		 * Update timestamp if present.
1198 		 */
1199 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1200 			sc->sc_tsreflect = to->to_tsval;
1201 		else
1202 			sc->sc_flags &= ~SCF_TIMESTAMP;
1203 #ifdef MAC
1204 		/*
1205 		 * Since we have already unconditionally allocated label
1206 		 * storage, free it up.  The syncache entry will already
1207 		 * have an initialized label we can use.
1208 		 */
1209 		mac_syncache_destroy(&maclabel);
1210 #endif
1211 		/* Retransmit SYN|ACK and reset retransmit count. */
1212 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1213 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1214 			    "resetting timer and retransmitting SYN|ACK\n",
1215 			    s, __func__);
1216 			free(s, M_TCPLOG);
1217 		}
1218 		if (syncache_respond(sc) == 0) {
1219 			sc->sc_rxmits = 0;
1220 			syncache_timeout(sc, sch, 1);
1221 			TCPSTAT_INC(tcps_sndacks);
1222 			TCPSTAT_INC(tcps_sndtotal);
1223 		}
1224 		SCH_UNLOCK(sch);
1225 		goto done;
1226 	}
1227 
1228 	sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1229 	if (sc == NULL) {
1230 		/*
1231 		 * The zone allocator couldn't provide more entries.
1232 		 * Treat this as if the cache was full; drop the oldest
1233 		 * entry and insert the new one.
1234 		 */
1235 		TCPSTAT_INC(tcps_sc_zonefail);
1236 		if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL)
1237 			syncache_drop(sc, sch);
1238 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1239 		if (sc == NULL) {
1240 			if (V_tcp_syncookies) {
1241 				bzero(&scs, sizeof(scs));
1242 				sc = &scs;
1243 			} else {
1244 				SCH_UNLOCK(sch);
1245 				if (ipopts)
1246 					(void) m_free(ipopts);
1247 				goto done;
1248 			}
1249 		}
1250 	}
1251 
1252 	/*
1253 	 * Fill in the syncache values.
1254 	 */
1255 #ifdef MAC
1256 	sc->sc_label = maclabel;
1257 #endif
1258 	sc->sc_cred = cred;
1259 	cred = NULL;
1260 	sc->sc_ipopts = ipopts;
1261 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1262 #ifdef INET6
1263 	if (!(inc->inc_flags & INC_ISIPV6))
1264 #endif
1265 	{
1266 		sc->sc_ip_tos = ip_tos;
1267 		sc->sc_ip_ttl = ip_ttl;
1268 	}
1269 #ifdef TCP_OFFLOAD
1270 	sc->sc_tod = tod;
1271 	sc->sc_todctx = todctx;
1272 #endif
1273 	sc->sc_irs = th->th_seq;
1274 	sc->sc_iss = arc4random();
1275 	sc->sc_flags = 0;
1276 	sc->sc_flowlabel = 0;
1277 
1278 	/*
1279 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1280 	 * win was derived from socket earlier in the function.
1281 	 */
1282 	win = imax(win, 0);
1283 	win = imin(win, TCP_MAXWIN);
1284 	sc->sc_wnd = win;
1285 
1286 	if (V_tcp_do_rfc1323) {
1287 		/*
1288 		 * A timestamp received in a SYN makes
1289 		 * it ok to send timestamp requests and replies.
1290 		 */
1291 		if (to->to_flags & TOF_TS) {
1292 			sc->sc_tsreflect = to->to_tsval;
1293 			sc->sc_ts = tcp_ts_getticks();
1294 			sc->sc_flags |= SCF_TIMESTAMP;
1295 		}
1296 		if (to->to_flags & TOF_SCALE) {
1297 			int wscale = 0;
1298 
1299 			/*
1300 			 * Pick the smallest possible scaling factor that
1301 			 * will still allow us to scale up to sb_max, aka
1302 			 * kern.ipc.maxsockbuf.
1303 			 *
1304 			 * We do this because there are broken firewalls that
1305 			 * will corrupt the window scale option, leading to
1306 			 * the other endpoint believing that our advertised
1307 			 * window is unscaled.  At scale factors larger than
1308 			 * 5 the unscaled window will drop below 1500 bytes,
1309 			 * leading to serious problems when traversing these
1310 			 * broken firewalls.
1311 			 *
1312 			 * With the default maxsockbuf of 256K, a scale factor
1313 			 * of 3 will be chosen by this algorithm.  Those who
1314 			 * choose a larger maxsockbuf should watch out
1315 			 * for the compatiblity problems mentioned above.
1316 			 *
1317 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1318 			 * or <SYN,ACK>) segment itself is never scaled.
1319 			 */
1320 			while (wscale < TCP_MAX_WINSHIFT &&
1321 			    (TCP_MAXWIN << wscale) < sb_max)
1322 				wscale++;
1323 			sc->sc_requested_r_scale = wscale;
1324 			sc->sc_requested_s_scale = to->to_wscale;
1325 			sc->sc_flags |= SCF_WINSCALE;
1326 		}
1327 	}
1328 #ifdef TCP_SIGNATURE
1329 	/*
1330 	 * If listening socket requested TCP digests, and received SYN
1331 	 * contains the option, flag this in the syncache so that
1332 	 * syncache_respond() will do the right thing with the SYN+ACK.
1333 	 * XXX: Currently we always record the option by default and will
1334 	 * attempt to use it in syncache_respond().
1335 	 */
1336 	if (to->to_flags & TOF_SIGNATURE || ltflags & TF_SIGNATURE)
1337 		sc->sc_flags |= SCF_SIGNATURE;
1338 #endif
1339 	if (to->to_flags & TOF_SACKPERM)
1340 		sc->sc_flags |= SCF_SACK;
1341 	if (to->to_flags & TOF_MSS)
1342 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1343 	if (ltflags & TF_NOOPT)
1344 		sc->sc_flags |= SCF_NOOPT;
1345 	if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1346 		sc->sc_flags |= SCF_ECN;
1347 
1348 	if (V_tcp_syncookies)
1349 		sc->sc_iss = syncookie_generate(sch, sc);
1350 #ifdef INET6
1351 	if (autoflowlabel) {
1352 		if (V_tcp_syncookies)
1353 			sc->sc_flowlabel = sc->sc_iss;
1354 		else
1355 			sc->sc_flowlabel = ip6_randomflowlabel();
1356 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1357 	}
1358 #endif
1359 	SCH_UNLOCK(sch);
1360 
1361 	/*
1362 	 * Do a standard 3-way handshake.
1363 	 */
1364 	if (syncache_respond(sc) == 0) {
1365 		if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1366 			syncache_free(sc);
1367 		else if (sc != &scs)
1368 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1369 		TCPSTAT_INC(tcps_sndacks);
1370 		TCPSTAT_INC(tcps_sndtotal);
1371 	} else {
1372 		if (sc != &scs)
1373 			syncache_free(sc);
1374 		TCPSTAT_INC(tcps_sc_dropped);
1375 	}
1376 
1377 done:
1378 	if (cred != NULL)
1379 		crfree(cred);
1380 #ifdef MAC
1381 	if (sc == &scs)
1382 		mac_syncache_destroy(&maclabel);
1383 #endif
1384 	if (m) {
1385 
1386 		*lsop = NULL;
1387 		m_freem(m);
1388 	}
1389 }
1390 
1391 static int
1392 syncache_respond(struct syncache *sc)
1393 {
1394 	struct ip *ip = NULL;
1395 	struct mbuf *m;
1396 	struct tcphdr *th = NULL;
1397 	int optlen, error = 0;	/* Make compiler happy */
1398 	u_int16_t hlen, tlen, mssopt;
1399 	struct tcpopt to;
1400 #ifdef INET6
1401 	struct ip6_hdr *ip6 = NULL;
1402 #endif
1403 
1404 	hlen =
1405 #ifdef INET6
1406 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1407 #endif
1408 		sizeof(struct ip);
1409 	tlen = hlen + sizeof(struct tcphdr);
1410 
1411 	/* Determine MSS we advertize to other end of connection. */
1412 	mssopt = tcp_mssopt(&sc->sc_inc);
1413 	if (sc->sc_peer_mss)
1414 		mssopt = max( min(sc->sc_peer_mss, mssopt), V_tcp_minmss);
1415 
1416 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1417 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1418 	    ("syncache: mbuf too small"));
1419 
1420 	/* Create the IP+TCP header from scratch. */
1421 	m = m_gethdr(M_NOWAIT, MT_DATA);
1422 	if (m == NULL)
1423 		return (ENOBUFS);
1424 #ifdef MAC
1425 	mac_syncache_create_mbuf(sc->sc_label, m);
1426 #endif
1427 	m->m_data += max_linkhdr;
1428 	m->m_len = tlen;
1429 	m->m_pkthdr.len = tlen;
1430 	m->m_pkthdr.rcvif = NULL;
1431 
1432 #ifdef INET6
1433 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1434 		ip6 = mtod(m, struct ip6_hdr *);
1435 		ip6->ip6_vfc = IPV6_VERSION;
1436 		ip6->ip6_nxt = IPPROTO_TCP;
1437 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1438 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1439 		ip6->ip6_plen = htons(tlen - hlen);
1440 		/* ip6_hlim is set after checksum */
1441 		ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1442 		ip6->ip6_flow |= sc->sc_flowlabel;
1443 
1444 		th = (struct tcphdr *)(ip6 + 1);
1445 	}
1446 #endif
1447 #if defined(INET6) && defined(INET)
1448 	else
1449 #endif
1450 #ifdef INET
1451 	{
1452 		ip = mtod(m, struct ip *);
1453 		ip->ip_v = IPVERSION;
1454 		ip->ip_hl = sizeof(struct ip) >> 2;
1455 		ip->ip_len = htons(tlen);
1456 		ip->ip_id = 0;
1457 		ip->ip_off = 0;
1458 		ip->ip_sum = 0;
1459 		ip->ip_p = IPPROTO_TCP;
1460 		ip->ip_src = sc->sc_inc.inc_laddr;
1461 		ip->ip_dst = sc->sc_inc.inc_faddr;
1462 		ip->ip_ttl = sc->sc_ip_ttl;
1463 		ip->ip_tos = sc->sc_ip_tos;
1464 
1465 		/*
1466 		 * See if we should do MTU discovery.  Route lookups are
1467 		 * expensive, so we will only unset the DF bit if:
1468 		 *
1469 		 *	1) path_mtu_discovery is disabled
1470 		 *	2) the SCF_UNREACH flag has been set
1471 		 */
1472 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1473 		       ip->ip_off |= htons(IP_DF);
1474 
1475 		th = (struct tcphdr *)(ip + 1);
1476 	}
1477 #endif /* INET */
1478 	th->th_sport = sc->sc_inc.inc_lport;
1479 	th->th_dport = sc->sc_inc.inc_fport;
1480 
1481 	th->th_seq = htonl(sc->sc_iss);
1482 	th->th_ack = htonl(sc->sc_irs + 1);
1483 	th->th_off = sizeof(struct tcphdr) >> 2;
1484 	th->th_x2 = 0;
1485 	th->th_flags = TH_SYN|TH_ACK;
1486 	th->th_win = htons(sc->sc_wnd);
1487 	th->th_urp = 0;
1488 
1489 	if (sc->sc_flags & SCF_ECN) {
1490 		th->th_flags |= TH_ECE;
1491 		TCPSTAT_INC(tcps_ecn_shs);
1492 	}
1493 
1494 	/* Tack on the TCP options. */
1495 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1496 		to.to_flags = 0;
1497 
1498 		to.to_mss = mssopt;
1499 		to.to_flags = TOF_MSS;
1500 		if (sc->sc_flags & SCF_WINSCALE) {
1501 			to.to_wscale = sc->sc_requested_r_scale;
1502 			to.to_flags |= TOF_SCALE;
1503 		}
1504 		if (sc->sc_flags & SCF_TIMESTAMP) {
1505 			/* Virgin timestamp or TCP cookie enhanced one. */
1506 			to.to_tsval = sc->sc_ts;
1507 			to.to_tsecr = sc->sc_tsreflect;
1508 			to.to_flags |= TOF_TS;
1509 		}
1510 		if (sc->sc_flags & SCF_SACK)
1511 			to.to_flags |= TOF_SACKPERM;
1512 #ifdef TCP_SIGNATURE
1513 		if (sc->sc_flags & SCF_SIGNATURE)
1514 			to.to_flags |= TOF_SIGNATURE;
1515 #endif
1516 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1517 
1518 		/* Adjust headers by option size. */
1519 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1520 		m->m_len += optlen;
1521 		m->m_pkthdr.len += optlen;
1522 
1523 #ifdef TCP_SIGNATURE
1524 		if (sc->sc_flags & SCF_SIGNATURE)
1525 			tcp_signature_compute(m, 0, 0, optlen,
1526 			    to.to_signature, IPSEC_DIR_OUTBOUND);
1527 #endif
1528 #ifdef INET6
1529 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1530 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1531 		else
1532 #endif
1533 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1534 	} else
1535 		optlen = 0;
1536 
1537 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
1538 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1539 #ifdef INET6
1540 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1541 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1542 		th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1543 		    IPPROTO_TCP, 0);
1544 		ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1545 #ifdef TCP_OFFLOAD
1546 		if (ADDED_BY_TOE(sc)) {
1547 			struct toedev *tod = sc->sc_tod;
1548 
1549 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1550 
1551 			return (error);
1552 		}
1553 #endif
1554 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1555 	}
1556 #endif
1557 #if defined(INET6) && defined(INET)
1558 	else
1559 #endif
1560 #ifdef INET
1561 	{
1562 		m->m_pkthdr.csum_flags = CSUM_TCP;
1563 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1564 		    htons(tlen + optlen - hlen + IPPROTO_TCP));
1565 #ifdef TCP_OFFLOAD
1566 		if (ADDED_BY_TOE(sc)) {
1567 			struct toedev *tod = sc->sc_tod;
1568 
1569 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1570 
1571 			return (error);
1572 		}
1573 #endif
1574 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1575 	}
1576 #endif
1577 	return (error);
1578 }
1579 
1580 /*
1581  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1582  * that exceed the capacity of the syncache by avoiding the storage of any
1583  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
1584  * attacks where the attacker does not have access to our responses.
1585  *
1586  * Syncookies encode and include all necessary information about the
1587  * connection setup within the SYN|ACK that we send back.  That way we
1588  * can avoid keeping any local state until the ACK to our SYN|ACK returns
1589  * (if ever).  Normally the syncache and syncookies are running in parallel
1590  * with the latter taking over when the former is exhausted.  When matching
1591  * syncache entry is found the syncookie is ignored.
1592  *
1593  * The only reliable information persisting the 3WHS is our inital sequence
1594  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
1595  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1596  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
1597  * returns and signifies a legitimate connection if it matches the ACK.
1598  *
1599  * The available space of 32 bits to store the hash and to encode the SYN
1600  * option information is very tight and we should have at least 24 bits for
1601  * the MAC to keep the number of guesses by blind spoofing reasonably high.
1602  *
1603  * SYN option information we have to encode to fully restore a connection:
1604  * MSS: is imporant to chose an optimal segment size to avoid IP level
1605  *   fragmentation along the path.  The common MSS values can be encoded
1606  *   in a 3-bit table.  Uncommon values are captured by the next lower value
1607  *   in the table leading to a slight increase in packetization overhead.
1608  * WSCALE: is necessary to allow large windows to be used for high delay-
1609  *   bandwidth product links.  Not scaling the window when it was initially
1610  *   negotiated is bad for performance as lack of scaling further decreases
1611  *   the apparent available send window.  We only need to encode the WSCALE
1612  *   we received from the remote end.  Our end can be recalculated at any
1613  *   time.  The common WSCALE values can be encoded in a 3-bit table.
1614  *   Uncommon values are captured by the next lower value in the table
1615  *   making us under-estimate the available window size halving our
1616  *   theoretically possible maximum throughput for that connection.
1617  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
1618  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
1619  *   that are included in all segments on a connection.  We enable them when
1620  *   the ACK has them.
1621  *
1622  * Security of syncookies and attack vectors:
1623  *
1624  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
1625  * together with the gloabl secret to make it unique per connection attempt.
1626  * Thus any change of any of those parameters results in a different MAC output
1627  * in an unpredictable way unless a collision is encountered.  24 bits of the
1628  * MAC are embedded into the ISS.
1629  *
1630  * To prevent replay attacks two rotating global secrets are updated with a
1631  * new random value every 15 seconds.  The life-time of a syncookie is thus
1632  * 15-30 seconds.
1633  *
1634  * Vector 1: Attacking the secret.  This requires finding a weakness in the
1635  * MAC itself or the way it is used here.  The attacker can do a chosen plain
1636  * text attack by varying and testing the all parameters under his control.
1637  * The strength depends on the size and randomness of the secret, and the
1638  * cryptographic security of the MAC function.  Due to the constant updating
1639  * of the secret the attacker has at most 29.999 seconds to find the secret
1640  * and launch spoofed connections.  After that he has to start all over again.
1641  *
1642  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
1643  * size an average of 4,823 attempts are required for a 50% chance of success
1644  * to spoof a single syncookie (birthday collision paradox).  However the
1645  * attacker is blind and doesn't know if one of his attempts succeeded unless
1646  * he has a side channel to interfere success from.  A single connection setup
1647  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
1648  * This many attempts are required for each one blind spoofed connection.  For
1649  * every additional spoofed connection he has to launch another N attempts.
1650  * Thus for a sustained rate 100 spoofed connections per second approximately
1651  * 1,800,000 packets per second would have to be sent.
1652  *
1653  * NB: The MAC function should be fast so that it doesn't become a CPU
1654  * exhaustion attack vector itself.
1655  *
1656  * References:
1657  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
1658  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
1659  *   http://cr.yp.to/syncookies.html    (overview)
1660  *   http://cr.yp.to/syncookies/archive (details)
1661  *
1662  *
1663  * Schematic construction of a syncookie enabled Initial Sequence Number:
1664  *  0        1         2         3
1665  *  12345678901234567890123456789012
1666  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
1667  *
1668  *  x 24 MAC (truncated)
1669  *  W  3 Send Window Scale index
1670  *  M  3 MSS index
1671  *  S  1 SACK permitted
1672  *  P  1 Odd/even secret
1673  */
1674 
1675 /*
1676  * Distribution and probability of certain MSS values.  Those in between are
1677  * rounded down to the next lower one.
1678  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
1679  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
1680  */
1681 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
1682 
1683 /*
1684  * Distribution and probability of certain WSCALE values.  We have to map the
1685  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
1686  * bits based on prevalence of certain values.  Where we don't have an exact
1687  * match for are rounded down to the next lower one letting us under-estimate
1688  * the true available window.  At the moment this would happen only for the
1689  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
1690  * and window size).  The absence of the WSCALE option (no scaling in either
1691  * direction) is encoded with index zero.
1692  * [WSCALE values histograms, Allman, 2012]
1693  *                            X 10 10 35  5  6 14 10%   by host
1694  *                            X 11  4  5  5 18 49  3%   by connections
1695  */
1696 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
1697 
1698 /*
1699  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
1700  * and good cryptographic properties.
1701  */
1702 static uint32_t
1703 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
1704     uint8_t *secbits, uintptr_t secmod)
1705 {
1706 	SIPHASH_CTX ctx;
1707 	uint32_t siphash[2];
1708 
1709 	SipHash24_Init(&ctx);
1710 	SipHash_SetKey(&ctx, secbits);
1711 	switch (inc->inc_flags & INC_ISIPV6) {
1712 #ifdef INET
1713 	case 0:
1714 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
1715 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
1716 		break;
1717 #endif
1718 #ifdef INET6
1719 	case INC_ISIPV6:
1720 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
1721 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
1722 		break;
1723 #endif
1724 	}
1725 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
1726 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
1727 	SipHash_Update(&ctx, &flags, sizeof(flags));
1728 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
1729 	SipHash_Final((u_int8_t *)&siphash, &ctx);
1730 
1731 	return (siphash[0] ^ siphash[1]);
1732 }
1733 
1734 static tcp_seq
1735 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
1736 {
1737 	u_int i, mss, secbit, wscale;
1738 	uint32_t iss, hash;
1739 	uint8_t *secbits;
1740 	union syncookie cookie;
1741 
1742 	SCH_LOCK_ASSERT(sch);
1743 
1744 	cookie.cookie = 0;
1745 
1746 	/* Map our computed MSS into the 3-bit index. */
1747 	mss = min(tcp_mssopt(&sc->sc_inc), max(sc->sc_peer_mss, V_tcp_minmss));
1748 	for (i = sizeof(tcp_sc_msstab) / sizeof(*tcp_sc_msstab) - 1;
1749 	     tcp_sc_msstab[i] > mss && i > 0;
1750 	     i--)
1751 		;
1752 	cookie.flags.mss_idx = i;
1753 
1754 	/*
1755 	 * Map the send window scale into the 3-bit index but only if
1756 	 * the wscale option was received.
1757 	 */
1758 	if (sc->sc_flags & SCF_WINSCALE) {
1759 		wscale = sc->sc_requested_s_scale;
1760 		for (i = sizeof(tcp_sc_wstab) / sizeof(*tcp_sc_wstab) - 1;
1761 		     tcp_sc_wstab[i] > wscale && i > 0;
1762 		     i--)
1763 			;
1764 		cookie.flags.wscale_idx = i;
1765 	}
1766 
1767 	/* Can we do SACK? */
1768 	if (sc->sc_flags & SCF_SACK)
1769 		cookie.flags.sack_ok = 1;
1770 
1771 	/* Which of the two secrets to use. */
1772 	secbit = sch->sch_sc->secret.oddeven & 0x1;
1773 	cookie.flags.odd_even = secbit;
1774 
1775 	secbits = sch->sch_sc->secret.key[secbit];
1776 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
1777 	    (uintptr_t)sch);
1778 
1779 	/*
1780 	 * Put the flags into the hash and XOR them to get better ISS number
1781 	 * variance.  This doesn't enhance the cryptographic strength and is
1782 	 * done to prevent the 8 cookie bits from showing up directly on the
1783 	 * wire.
1784 	 */
1785 	iss = hash & ~0xff;
1786 	iss |= cookie.cookie ^ (hash >> 24);
1787 
1788 	/* Randomize the timestamp. */
1789 	if (sc->sc_flags & SCF_TIMESTAMP) {
1790 		sc->sc_ts = arc4random();
1791 		sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks();
1792 	}
1793 
1794 	TCPSTAT_INC(tcps_sc_sendcookie);
1795 	return (iss);
1796 }
1797 
1798 static struct syncache *
1799 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
1800     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
1801     struct socket *lso)
1802 {
1803 	uint32_t hash;
1804 	uint8_t *secbits;
1805 	tcp_seq ack, seq;
1806 	int wnd, wscale = 0;
1807 	union syncookie cookie;
1808 
1809 	SCH_LOCK_ASSERT(sch);
1810 
1811 	/*
1812 	 * Pull information out of SYN-ACK/ACK and revert sequence number
1813 	 * advances.
1814 	 */
1815 	ack = th->th_ack - 1;
1816 	seq = th->th_seq - 1;
1817 
1818 	/*
1819 	 * Unpack the flags containing enough information to restore the
1820 	 * connection.
1821 	 */
1822 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
1823 
1824 	/* Which of the two secrets to use. */
1825 	secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
1826 
1827 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
1828 
1829 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
1830 	if ((ack & ~0xff) != (hash & ~0xff))
1831 		return (NULL);
1832 
1833 	/* Fill in the syncache values. */
1834 	sc->sc_flags = 0;
1835 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1836 	sc->sc_ipopts = NULL;
1837 
1838 	sc->sc_irs = seq;
1839 	sc->sc_iss = ack;
1840 
1841 	switch (inc->inc_flags & INC_ISIPV6) {
1842 #ifdef INET
1843 	case 0:
1844 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
1845 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
1846 		break;
1847 #endif
1848 #ifdef INET6
1849 	case INC_ISIPV6:
1850 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
1851 			sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
1852 		break;
1853 #endif
1854 	}
1855 
1856 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
1857 
1858 	/* We can simply recompute receive window scale we sent earlier. */
1859 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
1860 		wscale++;
1861 
1862 	/* Only use wscale if it was enabled in the orignal SYN. */
1863 	if (cookie.flags.wscale_idx > 0) {
1864 		sc->sc_requested_r_scale = wscale;
1865 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
1866 		sc->sc_flags |= SCF_WINSCALE;
1867 	}
1868 
1869 	wnd = sbspace(&lso->so_rcv);
1870 	wnd = imax(wnd, 0);
1871 	wnd = imin(wnd, TCP_MAXWIN);
1872 	sc->sc_wnd = wnd;
1873 
1874 	if (cookie.flags.sack_ok)
1875 		sc->sc_flags |= SCF_SACK;
1876 
1877 	if (to->to_flags & TOF_TS) {
1878 		sc->sc_flags |= SCF_TIMESTAMP;
1879 		sc->sc_tsreflect = to->to_tsval;
1880 		sc->sc_ts = to->to_tsecr;
1881 		sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks();
1882 	}
1883 
1884 	if (to->to_flags & TOF_SIGNATURE)
1885 		sc->sc_flags |= SCF_SIGNATURE;
1886 
1887 	sc->sc_rxmits = 0;
1888 
1889 	TCPSTAT_INC(tcps_sc_recvcookie);
1890 	return (sc);
1891 }
1892 
1893 #ifdef INVARIANTS
1894 static int
1895 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
1896     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
1897     struct socket *lso)
1898 {
1899 	struct syncache scs, *scx;
1900 	char *s;
1901 
1902 	bzero(&scs, sizeof(scs));
1903 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
1904 
1905 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
1906 		return (0);
1907 
1908 	if (scx != NULL) {
1909 		if (sc->sc_peer_mss != scx->sc_peer_mss)
1910 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
1911 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
1912 
1913 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
1914 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
1915 			    s, __func__, sc->sc_requested_r_scale,
1916 			    scx->sc_requested_r_scale);
1917 
1918 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
1919 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
1920 			    s, __func__, sc->sc_requested_s_scale,
1921 			    scx->sc_requested_s_scale);
1922 
1923 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
1924 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
1925 	}
1926 
1927 	if (s != NULL)
1928 		free(s, M_TCPLOG);
1929 	return (0);
1930 }
1931 #endif /* INVARIANTS */
1932 
1933 static void
1934 syncookie_reseed(void *arg)
1935 {
1936 	struct tcp_syncache *sc = arg;
1937 	uint8_t *secbits;
1938 	int secbit;
1939 
1940 	/*
1941 	 * Reseeding the secret doesn't have to be protected by a lock.
1942 	 * It only must be ensured that the new random values are visible
1943 	 * to all CPUs in a SMP environment.  The atomic with release
1944 	 * semantics ensures that.
1945 	 */
1946 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
1947 	secbits = sc->secret.key[secbit];
1948 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
1949 	atomic_add_rel_int(&sc->secret.oddeven, 1);
1950 
1951 	/* Reschedule ourself. */
1952 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
1953 }
1954 
1955 /*
1956  * Returns the current number of syncache entries.  This number
1957  * will probably change before you get around to calling
1958  * syncache_pcblist.
1959  */
1960 int
1961 syncache_pcbcount(void)
1962 {
1963 	struct syncache_head *sch;
1964 	int count, i;
1965 
1966 	for (count = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
1967 		/* No need to lock for a read. */
1968 		sch = &V_tcp_syncache.hashbase[i];
1969 		count += sch->sch_length;
1970 	}
1971 	return count;
1972 }
1973 
1974 /*
1975  * Exports the syncache entries to userland so that netstat can display
1976  * them alongside the other sockets.  This function is intended to be
1977  * called only from tcp_pcblist.
1978  *
1979  * Due to concurrency on an active system, the number of pcbs exported
1980  * may have no relation to max_pcbs.  max_pcbs merely indicates the
1981  * amount of space the caller allocated for this function to use.
1982  */
1983 int
1984 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
1985 {
1986 	struct xtcpcb xt;
1987 	struct syncache *sc;
1988 	struct syncache_head *sch;
1989 	int count, error, i;
1990 
1991 	for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
1992 		sch = &V_tcp_syncache.hashbase[i];
1993 		SCH_LOCK(sch);
1994 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
1995 			if (count >= max_pcbs) {
1996 				SCH_UNLOCK(sch);
1997 				goto exit;
1998 			}
1999 			if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2000 				continue;
2001 			bzero(&xt, sizeof(xt));
2002 			xt.xt_len = sizeof(xt);
2003 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2004 				xt.xt_inp.inp_vflag = INP_IPV6;
2005 			else
2006 				xt.xt_inp.inp_vflag = INP_IPV4;
2007 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo));
2008 			xt.xt_tp.t_inpcb = &xt.xt_inp;
2009 			xt.xt_tp.t_state = TCPS_SYN_RECEIVED;
2010 			xt.xt_socket.xso_protocol = IPPROTO_TCP;
2011 			xt.xt_socket.xso_len = sizeof (struct xsocket);
2012 			xt.xt_socket.so_type = SOCK_STREAM;
2013 			xt.xt_socket.so_state = SS_ISCONNECTING;
2014 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2015 			if (error) {
2016 				SCH_UNLOCK(sch);
2017 				goto exit;
2018 			}
2019 			count++;
2020 		}
2021 		SCH_UNLOCK(sch);
2022 	}
2023 exit:
2024 	*pcbs_exported = count;
2025 	return error;
2026 }
2027