xref: /freebsd/sys/netinet/tcp_syncache.c (revision bd81e07d2761cf1c13063eb49a5c0cb4a6951318)
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_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | 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_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | 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 static int	 syncache_respond(struct syncache *, struct syncache_head *, int);
126 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
127 		    struct mbuf *m);
128 static void	 syncache_timeout(struct syncache *sc, struct syncache_head *sch,
129 		    int docallout);
130 static void	 syncache_timer(void *);
131 
132 static uint32_t	 syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
133 		    uint8_t *, uintptr_t);
134 static tcp_seq	 syncookie_generate(struct syncache_head *, struct syncache *);
135 static struct syncache
136 		*syncookie_lookup(struct in_conninfo *, struct syncache_head *,
137 		    struct syncache *, struct tcphdr *, struct tcpopt *,
138 		    struct socket *);
139 static void	 syncookie_reseed(void *);
140 #ifdef INVARIANTS
141 static int	 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
142 		    struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
143 		    struct socket *lso);
144 #endif
145 
146 /*
147  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
148  * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds,
149  * the odds are that the user has given up attempting to connect by then.
150  */
151 #define SYNCACHE_MAXREXMTS		3
152 
153 /* Arbitrary values */
154 #define TCP_SYNCACHE_HASHSIZE		512
155 #define TCP_SYNCACHE_BUCKETLIMIT	30
156 
157 static VNET_DEFINE(struct tcp_syncache, tcp_syncache);
158 #define	V_tcp_syncache			VNET(tcp_syncache)
159 
160 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0,
161     "TCP SYN cache");
162 
163 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
164     &VNET_NAME(tcp_syncache.bucket_limit), 0,
165     "Per-bucket hash limit for syncache");
166 
167 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
168     &VNET_NAME(tcp_syncache.cache_limit), 0,
169     "Overall entry limit for syncache");
170 
171 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
172     &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
173 
174 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
175     &VNET_NAME(tcp_syncache.hashsize), 0,
176     "Size of TCP syncache hashtable");
177 
178 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_VNET | CTLFLAG_RW,
179     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
180     "Limit on SYN/ACK retransmissions");
181 
182 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
183 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
184     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
185     "Send reset on socket allocation failure");
186 
187 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
188 
189 #define SYNCACHE_HASH(inc, mask)					\
190 	((V_tcp_syncache.hash_secret ^					\
191 	  (inc)->inc_faddr.s_addr ^					\
192 	  ((inc)->inc_faddr.s_addr >> 16) ^				\
193 	  (inc)->inc_fport ^ (inc)->inc_lport) & mask)
194 
195 #define SYNCACHE_HASH6(inc, mask)					\
196 	((V_tcp_syncache.hash_secret ^					\
197 	  (inc)->inc6_faddr.s6_addr32[0] ^				\
198 	  (inc)->inc6_faddr.s6_addr32[3] ^				\
199 	  (inc)->inc_fport ^ (inc)->inc_lport) & mask)
200 
201 #define ENDPTS_EQ(a, b) (						\
202 	(a)->ie_fport == (b)->ie_fport &&				\
203 	(a)->ie_lport == (b)->ie_lport &&				\
204 	(a)->ie_faddr.s_addr == (b)->ie_faddr.s_addr &&			\
205 	(a)->ie_laddr.s_addr == (b)->ie_laddr.s_addr			\
206 )
207 
208 #define ENDPTS6_EQ(a, b) (memcmp(a, b, sizeof(*a)) == 0)
209 
210 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
211 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
212 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
213 
214 /*
215  * Requires the syncache entry to be already removed from the bucket list.
216  */
217 static void
218 syncache_free(struct syncache *sc)
219 {
220 
221 	if (sc->sc_ipopts)
222 		(void) m_free(sc->sc_ipopts);
223 	if (sc->sc_cred)
224 		crfree(sc->sc_cred);
225 #ifdef MAC
226 	mac_syncache_destroy(&sc->sc_label);
227 #endif
228 
229 	uma_zfree(V_tcp_syncache.zone, sc);
230 }
231 
232 void
233 syncache_init(void)
234 {
235 	int i;
236 
237 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
238 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
239 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
240 	V_tcp_syncache.hash_secret = arc4random();
241 
242 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
243 	    &V_tcp_syncache.hashsize);
244 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
245 	    &V_tcp_syncache.bucket_limit);
246 	if (!powerof2(V_tcp_syncache.hashsize) ||
247 	    V_tcp_syncache.hashsize == 0) {
248 		printf("WARNING: syncache hash size is not a power of 2.\n");
249 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
250 	}
251 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
252 
253 	/* Set limits. */
254 	V_tcp_syncache.cache_limit =
255 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
256 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
257 	    &V_tcp_syncache.cache_limit);
258 
259 	/* Allocate the hash table. */
260 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
261 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
262 
263 #ifdef VIMAGE
264 	V_tcp_syncache.vnet = curvnet;
265 #endif
266 
267 	/* Initialize the hash buckets. */
268 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
269 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
270 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
271 			 NULL, MTX_DEF);
272 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
273 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
274 		V_tcp_syncache.hashbase[i].sch_length = 0;
275 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
276 	}
277 
278 	/* Create the syncache entry zone. */
279 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
280 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
281 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
282 	    V_tcp_syncache.cache_limit);
283 
284 	/* Start the SYN cookie reseeder callout. */
285 	callout_init(&V_tcp_syncache.secret.reseed, 1);
286 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
287 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
288 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
289 	    syncookie_reseed, &V_tcp_syncache);
290 }
291 
292 #ifdef VIMAGE
293 void
294 syncache_destroy(void)
295 {
296 	struct syncache_head *sch;
297 	struct syncache *sc, *nsc;
298 	int i;
299 
300 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
301 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
302 
303 		sch = &V_tcp_syncache.hashbase[i];
304 		callout_drain(&sch->sch_timer);
305 
306 		SCH_LOCK(sch);
307 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
308 			syncache_drop(sc, sch);
309 		SCH_UNLOCK(sch);
310 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
311 		    ("%s: sch->sch_bucket not empty", __func__));
312 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
313 		    __func__, sch->sch_length));
314 		mtx_destroy(&sch->sch_mtx);
315 	}
316 
317 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
318 	    ("%s: cache_count not 0", __func__));
319 
320 	/* Free the allocated global resources. */
321 	uma_zdestroy(V_tcp_syncache.zone);
322 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
323 
324 	callout_drain(&V_tcp_syncache.secret.reseed);
325 }
326 #endif
327 
328 /*
329  * Inserts a syncache entry into the specified bucket row.
330  * Locks and unlocks the syncache_head autonomously.
331  */
332 static void
333 syncache_insert(struct syncache *sc, struct syncache_head *sch)
334 {
335 	struct syncache *sc2;
336 
337 	SCH_LOCK(sch);
338 
339 	/*
340 	 * Make sure that we don't overflow the per-bucket limit.
341 	 * If the bucket is full, toss the oldest element.
342 	 */
343 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
344 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
345 			("sch->sch_length incorrect"));
346 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
347 		syncache_drop(sc2, sch);
348 		TCPSTAT_INC(tcps_sc_bucketoverflow);
349 	}
350 
351 	/* Put it into the bucket. */
352 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
353 	sch->sch_length++;
354 
355 #ifdef TCP_OFFLOAD
356 	if (ADDED_BY_TOE(sc)) {
357 		struct toedev *tod = sc->sc_tod;
358 
359 		tod->tod_syncache_added(tod, sc->sc_todctx);
360 	}
361 #endif
362 
363 	/* Reinitialize the bucket row's timer. */
364 	if (sch->sch_length == 1)
365 		sch->sch_nextc = ticks + INT_MAX;
366 	syncache_timeout(sc, sch, 1);
367 
368 	SCH_UNLOCK(sch);
369 
370 	TCPSTAT_INC(tcps_sc_added);
371 }
372 
373 /*
374  * Remove and free entry from syncache bucket row.
375  * Expects locked syncache head.
376  */
377 static void
378 syncache_drop(struct syncache *sc, struct syncache_head *sch)
379 {
380 
381 	SCH_LOCK_ASSERT(sch);
382 
383 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
384 	sch->sch_length--;
385 
386 #ifdef TCP_OFFLOAD
387 	if (ADDED_BY_TOE(sc)) {
388 		struct toedev *tod = sc->sc_tod;
389 
390 		tod->tod_syncache_removed(tod, sc->sc_todctx);
391 	}
392 #endif
393 
394 	syncache_free(sc);
395 }
396 
397 /*
398  * Engage/reengage time on bucket row.
399  */
400 static void
401 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
402 {
403 	sc->sc_rxttime = ticks +
404 		TCPTV_RTOBASE * (tcp_syn_backoff[sc->sc_rxmits]);
405 	sc->sc_rxmits++;
406 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
407 		sch->sch_nextc = sc->sc_rxttime;
408 		if (docallout)
409 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
410 			    syncache_timer, (void *)sch);
411 	}
412 }
413 
414 /*
415  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
416  * If we have retransmitted an entry the maximum number of times, expire it.
417  * One separate timer for each bucket row.
418  */
419 static void
420 syncache_timer(void *xsch)
421 {
422 	struct syncache_head *sch = (struct syncache_head *)xsch;
423 	struct syncache *sc, *nsc;
424 	int tick = ticks;
425 	char *s;
426 
427 	CURVNET_SET(sch->sch_sc->vnet);
428 
429 	/* NB: syncache_head has already been locked by the callout. */
430 	SCH_LOCK_ASSERT(sch);
431 
432 	/*
433 	 * In the following cycle we may remove some entries and/or
434 	 * advance some timeouts, so re-initialize the bucket timer.
435 	 */
436 	sch->sch_nextc = tick + INT_MAX;
437 
438 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
439 		/*
440 		 * We do not check if the listen socket still exists
441 		 * and accept the case where the listen socket may be
442 		 * gone by the time we resend the SYN/ACK.  We do
443 		 * not expect this to happens often. If it does,
444 		 * then the RST will be sent by the time the remote
445 		 * host does the SYN/ACK->ACK.
446 		 */
447 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
448 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
449 				sch->sch_nextc = sc->sc_rxttime;
450 			continue;
451 		}
452 		if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
453 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
454 				log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
455 				    "giving up and removing syncache entry\n",
456 				    s, __func__);
457 				free(s, M_TCPLOG);
458 			}
459 			syncache_drop(sc, sch);
460 			TCPSTAT_INC(tcps_sc_stale);
461 			continue;
462 		}
463 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
464 			log(LOG_DEBUG, "%s; %s: Response timeout, "
465 			    "retransmitting (%u) SYN|ACK\n",
466 			    s, __func__, sc->sc_rxmits);
467 			free(s, M_TCPLOG);
468 		}
469 
470 		syncache_respond(sc, sch, 1);
471 		TCPSTAT_INC(tcps_sc_retransmitted);
472 		syncache_timeout(sc, sch, 0);
473 	}
474 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
475 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
476 			syncache_timer, (void *)(sch));
477 	CURVNET_RESTORE();
478 }
479 
480 /*
481  * Find an entry in the syncache.
482  * Returns always with locked syncache_head plus a matching entry or NULL.
483  */
484 static struct syncache *
485 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
486 {
487 	struct syncache *sc;
488 	struct syncache_head *sch;
489 
490 #ifdef INET6
491 	if (inc->inc_flags & INC_ISIPV6) {
492 		sch = &V_tcp_syncache.hashbase[
493 		    SYNCACHE_HASH6(inc, V_tcp_syncache.hashmask)];
494 		*schp = sch;
495 
496 		SCH_LOCK(sch);
497 
498 		/* Circle through bucket row to find matching entry. */
499 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
500 			if (ENDPTS6_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
501 				return (sc);
502 		}
503 	} else
504 #endif
505 	{
506 		sch = &V_tcp_syncache.hashbase[
507 		    SYNCACHE_HASH(inc, V_tcp_syncache.hashmask)];
508 		*schp = sch;
509 
510 		SCH_LOCK(sch);
511 
512 		/* Circle through bucket row to find matching entry. */
513 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
514 #ifdef INET6
515 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
516 				continue;
517 #endif
518 			if (ENDPTS_EQ(&inc->inc_ie, &sc->sc_inc.inc_ie))
519 				return (sc);
520 		}
521 	}
522 	SCH_LOCK_ASSERT(*schp);
523 	return (NULL);			/* always returns with locked sch */
524 }
525 
526 /*
527  * This function is called when we get a RST for a
528  * non-existent connection, so that we can see if the
529  * connection is in the syn cache.  If it is, zap it.
530  */
531 void
532 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
533 {
534 	struct syncache *sc;
535 	struct syncache_head *sch;
536 	char *s = NULL;
537 
538 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
539 	SCH_LOCK_ASSERT(sch);
540 
541 	/*
542 	 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
543 	 * See RFC 793 page 65, section SEGMENT ARRIVES.
544 	 */
545 	if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
546 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
547 			log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
548 			    "FIN flag set, segment ignored\n", s, __func__);
549 		TCPSTAT_INC(tcps_badrst);
550 		goto done;
551 	}
552 
553 	/*
554 	 * No corresponding connection was found in syncache.
555 	 * If syncookies are enabled and possibly exclusively
556 	 * used, or we are under memory pressure, a valid RST
557 	 * may not find a syncache entry.  In that case we're
558 	 * done and no SYN|ACK retransmissions will happen.
559 	 * Otherwise the RST was misdirected or spoofed.
560 	 */
561 	if (sc == NULL) {
562 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
563 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
564 			    "syncache entry (possibly syncookie only), "
565 			    "segment ignored\n", s, __func__);
566 		TCPSTAT_INC(tcps_badrst);
567 		goto done;
568 	}
569 
570 	/*
571 	 * If the RST bit is set, check the sequence number to see
572 	 * if this is a valid reset segment.
573 	 * RFC 793 page 37:
574 	 *   In all states except SYN-SENT, all reset (RST) segments
575 	 *   are validated by checking their SEQ-fields.  A reset is
576 	 *   valid if its sequence number is in the window.
577 	 *
578 	 *   The sequence number in the reset segment is normally an
579 	 *   echo of our outgoing acknowlegement numbers, but some hosts
580 	 *   send a reset with the sequence number at the rightmost edge
581 	 *   of our receive window, and we have to handle this case.
582 	 */
583 	if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
584 	    SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
585 		syncache_drop(sc, sch);
586 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
587 			log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, "
588 			    "connection attempt aborted by remote endpoint\n",
589 			    s, __func__);
590 		TCPSTAT_INC(tcps_sc_reset);
591 	} else {
592 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
593 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
594 			    "IRS %u (+WND %u), segment ignored\n",
595 			    s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd);
596 		TCPSTAT_INC(tcps_badrst);
597 	}
598 
599 done:
600 	if (s != NULL)
601 		free(s, M_TCPLOG);
602 	SCH_UNLOCK(sch);
603 }
604 
605 void
606 syncache_badack(struct in_conninfo *inc)
607 {
608 	struct syncache *sc;
609 	struct syncache_head *sch;
610 
611 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
612 	SCH_LOCK_ASSERT(sch);
613 	if (sc != NULL) {
614 		syncache_drop(sc, sch);
615 		TCPSTAT_INC(tcps_sc_badack);
616 	}
617 	SCH_UNLOCK(sch);
618 }
619 
620 void
621 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th)
622 {
623 	struct syncache *sc;
624 	struct syncache_head *sch;
625 
626 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
627 	SCH_LOCK_ASSERT(sch);
628 	if (sc == NULL)
629 		goto done;
630 
631 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
632 	if (ntohl(th->th_seq) != sc->sc_iss)
633 		goto done;
634 
635 	/*
636 	 * If we've rertransmitted 3 times and this is our second error,
637 	 * we remove the entry.  Otherwise, we allow it to continue on.
638 	 * This prevents us from incorrectly nuking an entry during a
639 	 * spurious network outage.
640 	 *
641 	 * See tcp_notify().
642 	 */
643 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
644 		sc->sc_flags |= SCF_UNREACH;
645 		goto done;
646 	}
647 	syncache_drop(sc, sch);
648 	TCPSTAT_INC(tcps_sc_unreach);
649 done:
650 	SCH_UNLOCK(sch);
651 }
652 
653 /*
654  * Build a new TCP socket structure from a syncache entry.
655  *
656  * On success return the newly created socket with its underlying inp locked.
657  */
658 static struct socket *
659 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
660 {
661 	struct inpcb *inp = NULL;
662 	struct socket *so;
663 	struct tcpcb *tp;
664 	int error;
665 	char *s;
666 
667 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
668 
669 	/*
670 	 * Ok, create the full blown connection, and set things up
671 	 * as they would have been set up if we had created the
672 	 * connection when the SYN arrived.  If we can't create
673 	 * the connection, abort it.
674 	 */
675 	so = sonewconn(lso, 0);
676 	if (so == NULL) {
677 		/*
678 		 * Drop the connection; we will either send a RST or
679 		 * have the peer retransmit its SYN again after its
680 		 * RTO and try again.
681 		 */
682 		TCPSTAT_INC(tcps_listendrop);
683 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
684 			log(LOG_DEBUG, "%s; %s: Socket create failed "
685 			    "due to limits or memory shortage\n",
686 			    s, __func__);
687 			free(s, M_TCPLOG);
688 		}
689 		goto abort2;
690 	}
691 #ifdef MAC
692 	mac_socketpeer_set_from_mbuf(m, so);
693 #endif
694 
695 	inp = sotoinpcb(so);
696 	inp->inp_inc.inc_fibnum = so->so_fibnum;
697 	INP_WLOCK(inp);
698 	/*
699 	 * Exclusive pcbinfo lock is not required in syncache socket case even
700 	 * if two inpcb locks can be acquired simultaneously:
701 	 *  - the inpcb in LISTEN state,
702 	 *  - the newly created inp.
703 	 *
704 	 * In this case, an inp cannot be at same time in LISTEN state and
705 	 * just created by an accept() call.
706 	 */
707 	INP_HASH_WLOCK(&V_tcbinfo);
708 
709 	/* Insert new socket into PCB hash list. */
710 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
711 #ifdef INET6
712 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
713 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
714 	} else {
715 		inp->inp_vflag &= ~INP_IPV6;
716 		inp->inp_vflag |= INP_IPV4;
717 #endif
718 		inp->inp_laddr = sc->sc_inc.inc_laddr;
719 #ifdef INET6
720 	}
721 #endif
722 
723 	/*
724 	 * If there's an mbuf and it has a flowid, then let's initialise the
725 	 * inp with that particular flowid.
726 	 */
727 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
728 		inp->inp_flowid = m->m_pkthdr.flowid;
729 		inp->inp_flowtype = M_HASHTYPE_GET(m);
730 	}
731 
732 	/*
733 	 * Install in the reservation hash table for now, but don't yet
734 	 * install a connection group since the full 4-tuple isn't yet
735 	 * configured.
736 	 */
737 	inp->inp_lport = sc->sc_inc.inc_lport;
738 	if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
739 		/*
740 		 * Undo the assignments above if we failed to
741 		 * put the PCB on the hash lists.
742 		 */
743 #ifdef INET6
744 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
745 			inp->in6p_laddr = in6addr_any;
746 		else
747 #endif
748 			inp->inp_laddr.s_addr = INADDR_ANY;
749 		inp->inp_lport = 0;
750 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
751 			log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
752 			    "with error %i\n",
753 			    s, __func__, error);
754 			free(s, M_TCPLOG);
755 		}
756 		INP_HASH_WUNLOCK(&V_tcbinfo);
757 		goto abort;
758 	}
759 #ifdef IPSEC
760 	/* Copy old policy into new socket's. */
761 	if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp))
762 		printf("syncache_socket: could not copy policy\n");
763 #endif
764 #ifdef INET6
765 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
766 		struct inpcb *oinp = sotoinpcb(lso);
767 		struct in6_addr laddr6;
768 		struct sockaddr_in6 sin6;
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 |= oinp->inp_flags & INP_CONTROLOPTS;
779 		if (oinp->in6p_outputopts)
780 			inp->in6p_outputopts =
781 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
782 
783 		sin6.sin6_family = AF_INET6;
784 		sin6.sin6_len = sizeof(sin6);
785 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
786 		sin6.sin6_port = sc->sc_inc.inc_fport;
787 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
788 		laddr6 = inp->in6p_laddr;
789 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
790 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
791 		if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
792 		    thread0.td_ucred, m)) != 0) {
793 			inp->in6p_laddr = laddr6;
794 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
795 				log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
796 				    "with error %i\n",
797 				    s, __func__, error);
798 				free(s, M_TCPLOG);
799 			}
800 			INP_HASH_WUNLOCK(&V_tcbinfo);
801 			goto abort;
802 		}
803 		/* Override flowlabel from in6_pcbconnect. */
804 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
805 		inp->inp_flow |= sc->sc_flowlabel;
806 	}
807 #endif /* INET6 */
808 #if defined(INET) && defined(INET6)
809 	else
810 #endif
811 #ifdef INET
812 	{
813 		struct in_addr laddr;
814 		struct sockaddr_in sin;
815 
816 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
817 
818 		if (inp->inp_options == NULL) {
819 			inp->inp_options = sc->sc_ipopts;
820 			sc->sc_ipopts = NULL;
821 		}
822 
823 		sin.sin_family = AF_INET;
824 		sin.sin_len = sizeof(sin);
825 		sin.sin_addr = sc->sc_inc.inc_faddr;
826 		sin.sin_port = sc->sc_inc.inc_fport;
827 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
828 		laddr = inp->inp_laddr;
829 		if (inp->inp_laddr.s_addr == INADDR_ANY)
830 			inp->inp_laddr = sc->sc_inc.inc_laddr;
831 		if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
832 		    thread0.td_ucred, m)) != 0) {
833 			inp->inp_laddr = laddr;
834 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
835 				log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
836 				    "with error %i\n",
837 				    s, __func__, error);
838 				free(s, M_TCPLOG);
839 			}
840 			INP_HASH_WUNLOCK(&V_tcbinfo);
841 			goto abort;
842 		}
843 	}
844 #endif /* INET */
845 	INP_HASH_WUNLOCK(&V_tcbinfo);
846 	tp = intotcpcb(inp);
847 	tcp_state_change(tp, TCPS_SYN_RECEIVED);
848 	tp->iss = sc->sc_iss;
849 	tp->irs = sc->sc_irs;
850 	tcp_rcvseqinit(tp);
851 	tcp_sendseqinit(tp);
852 	tp->snd_wl1 = sc->sc_irs;
853 	tp->snd_max = tp->iss + 1;
854 	tp->snd_nxt = tp->iss + 1;
855 	tp->rcv_up = sc->sc_irs + 1;
856 	tp->rcv_wnd = sc->sc_wnd;
857 	tp->rcv_adv += tp->rcv_wnd;
858 	tp->last_ack_sent = tp->rcv_nxt;
859 
860 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
861 	if (sc->sc_flags & SCF_NOOPT)
862 		tp->t_flags |= TF_NOOPT;
863 	else {
864 		if (sc->sc_flags & SCF_WINSCALE) {
865 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
866 			tp->snd_scale = sc->sc_requested_s_scale;
867 			tp->request_r_scale = sc->sc_requested_r_scale;
868 		}
869 		if (sc->sc_flags & SCF_TIMESTAMP) {
870 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
871 			tp->ts_recent = sc->sc_tsreflect;
872 			tp->ts_recent_age = tcp_ts_getticks();
873 			tp->ts_offset = sc->sc_tsoff;
874 		}
875 #ifdef TCP_SIGNATURE
876 		if (sc->sc_flags & SCF_SIGNATURE)
877 			tp->t_flags |= TF_SIGNATURE;
878 #endif
879 		if (sc->sc_flags & SCF_SACK)
880 			tp->t_flags |= TF_SACK_PERMIT;
881 	}
882 
883 	if (sc->sc_flags & SCF_ECN)
884 		tp->t_flags |= TF_ECN_PERMIT;
885 
886 	/*
887 	 * Set up MSS and get cached values from tcp_hostcache.
888 	 * This might overwrite some of the defaults we just set.
889 	 */
890 	tcp_mss(tp, sc->sc_peer_mss);
891 
892 	/*
893 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
894 	 * limited to one segment in cc_conn_init().
895 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
896 	 */
897 	if (sc->sc_rxmits > 1)
898 		tp->snd_cwnd = 1;
899 
900 #ifdef TCP_OFFLOAD
901 	/*
902 	 * Allow a TOE driver to install its hooks.  Note that we hold the
903 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
904 	 * new connection before the TOE driver has done its thing.
905 	 */
906 	if (ADDED_BY_TOE(sc)) {
907 		struct toedev *tod = sc->sc_tod;
908 
909 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
910 	}
911 #endif
912 	/*
913 	 * Copy and activate timers.
914 	 */
915 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
916 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
917 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
918 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
919 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
920 
921 	soisconnected(so);
922 
923 	TCPSTAT_INC(tcps_accepts);
924 	return (so);
925 
926 abort:
927 	INP_WUNLOCK(inp);
928 abort2:
929 	if (so != NULL)
930 		soabort(so);
931 	return (NULL);
932 }
933 
934 /*
935  * This function gets called when we receive an ACK for a
936  * socket in the LISTEN state.  We look up the connection
937  * in the syncache, and if its there, we pull it out of
938  * the cache and turn it into a full-blown connection in
939  * the SYN-RECEIVED state.
940  *
941  * On syncache_socket() success the newly created socket
942  * has its underlying inp locked.
943  */
944 int
945 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
946     struct socket **lsop, struct mbuf *m)
947 {
948 	struct syncache *sc;
949 	struct syncache_head *sch;
950 	struct syncache scs;
951 	char *s;
952 
953 	/*
954 	 * Global TCP locks are held because we manipulate the PCB lists
955 	 * and create a new socket.
956 	 */
957 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
958 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
959 	    ("%s: can handle only ACK", __func__));
960 
961 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
962 	SCH_LOCK_ASSERT(sch);
963 
964 #ifdef INVARIANTS
965 	/*
966 	 * Test code for syncookies comparing the syncache stored
967 	 * values with the reconstructed values from the cookie.
968 	 */
969 	if (sc != NULL)
970 		syncookie_cmp(inc, sch, sc, th, to, *lsop);
971 #endif
972 
973 	if (sc == NULL) {
974 		/*
975 		 * There is no syncache entry, so see if this ACK is
976 		 * a returning syncookie.  To do this, first:
977 		 *  A. See if this socket has had a syncache entry dropped in
978 		 *     the past.  We don't want to accept a bogus syncookie
979 		 *     if we've never received a SYN.
980 		 *  B. check that the syncookie is valid.  If it is, then
981 		 *     cobble up a fake syncache entry, and return.
982 		 */
983 		if (!V_tcp_syncookies) {
984 			SCH_UNLOCK(sch);
985 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
986 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
987 				    "segment rejected (syncookies disabled)\n",
988 				    s, __func__);
989 			goto failed;
990 		}
991 		bzero(&scs, sizeof(scs));
992 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
993 		SCH_UNLOCK(sch);
994 		if (sc == NULL) {
995 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
996 				log(LOG_DEBUG, "%s; %s: Segment failed "
997 				    "SYNCOOKIE authentication, segment rejected "
998 				    "(probably spoofed)\n", s, __func__);
999 			goto failed;
1000 		}
1001 	} else {
1002 		/* Pull out the entry to unlock the bucket row. */
1003 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1004 		sch->sch_length--;
1005 #ifdef TCP_OFFLOAD
1006 		if (ADDED_BY_TOE(sc)) {
1007 			struct toedev *tod = sc->sc_tod;
1008 
1009 			tod->tod_syncache_removed(tod, sc->sc_todctx);
1010 		}
1011 #endif
1012 		SCH_UNLOCK(sch);
1013 	}
1014 
1015 	/*
1016 	 * Segment validation:
1017 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1018 	 */
1019 	if (th->th_ack != sc->sc_iss + 1) {
1020 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1021 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1022 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1023 		goto failed;
1024 	}
1025 
1026 	/*
1027 	 * The SEQ must fall in the window starting at the received
1028 	 * initial receive sequence number + 1 (the SYN).
1029 	 */
1030 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1031 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1032 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1033 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1034 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1035 		goto failed;
1036 	}
1037 
1038 	/*
1039 	 * If timestamps were not negotiated during SYN/ACK they
1040 	 * must not appear on any segment during this session.
1041 	 */
1042 	if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1043 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1044 			log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1045 			    "segment rejected\n", s, __func__);
1046 		goto failed;
1047 	}
1048 
1049 	/*
1050 	 * If timestamps were negotiated during SYN/ACK they should
1051 	 * appear on every segment during this session.
1052 	 * XXXAO: This is only informal as there have been unverified
1053 	 * reports of non-compliants stacks.
1054 	 */
1055 	if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1056 		if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1057 			log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1058 			    "no action\n", s, __func__);
1059 			free(s, M_TCPLOG);
1060 			s = NULL;
1061 		}
1062 	}
1063 
1064 	/*
1065 	 * If timestamps were negotiated the reflected timestamp
1066 	 * must be equal to what we actually sent in the SYN|ACK.
1067 	 */
1068 	if ((to->to_flags & TOF_TS) && to->to_tsecr != sc->sc_ts) {
1069 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1070 			log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, "
1071 			    "segment rejected\n",
1072 			    s, __func__, to->to_tsecr, sc->sc_ts);
1073 		goto failed;
1074 	}
1075 
1076 	*lsop = syncache_socket(sc, *lsop, m);
1077 
1078 	if (*lsop == NULL)
1079 		TCPSTAT_INC(tcps_sc_aborted);
1080 	else
1081 		TCPSTAT_INC(tcps_sc_completed);
1082 
1083 /* how do we find the inp for the new socket? */
1084 	if (sc != &scs)
1085 		syncache_free(sc);
1086 	return (1);
1087 failed:
1088 	if (sc != NULL && sc != &scs)
1089 		syncache_free(sc);
1090 	if (s != NULL)
1091 		free(s, M_TCPLOG);
1092 	*lsop = NULL;
1093 	return (0);
1094 }
1095 
1096 /*
1097  * Given a LISTEN socket and an inbound SYN request, add
1098  * this to the syn cache, and send back a segment:
1099  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1100  * to the source.
1101  *
1102  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1103  * Doing so would require that we hold onto the data and deliver it
1104  * to the application.  However, if we are the target of a SYN-flood
1105  * DoS attack, an attacker could send data which would eventually
1106  * consume all available buffer space if it were ACKed.  By not ACKing
1107  * the data, we avoid this DoS scenario.
1108  */
1109 void
1110 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1111     struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1112     void *todctx)
1113 {
1114 	struct tcpcb *tp;
1115 	struct socket *so;
1116 	struct syncache *sc = NULL;
1117 	struct syncache_head *sch;
1118 	struct mbuf *ipopts = NULL;
1119 	u_int ltflags;
1120 	int win, sb_hiwat, ip_ttl, ip_tos;
1121 	char *s;
1122 #ifdef INET6
1123 	int autoflowlabel = 0;
1124 #endif
1125 #ifdef MAC
1126 	struct label *maclabel;
1127 #endif
1128 	struct syncache scs;
1129 	struct ucred *cred;
1130 
1131 	INP_WLOCK_ASSERT(inp);			/* listen socket */
1132 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1133 	    ("%s: unexpected tcp flags", __func__));
1134 
1135 	/*
1136 	 * Combine all so/tp operations very early to drop the INP lock as
1137 	 * soon as possible.
1138 	 */
1139 	so = *lsop;
1140 	tp = sototcpcb(so);
1141 	cred = crhold(so->so_cred);
1142 
1143 #ifdef INET6
1144 	if ((inc->inc_flags & INC_ISIPV6) &&
1145 	    (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1146 		autoflowlabel = 1;
1147 #endif
1148 	ip_ttl = inp->inp_ip_ttl;
1149 	ip_tos = inp->inp_ip_tos;
1150 	win = sbspace(&so->so_rcv);
1151 	sb_hiwat = so->so_rcv.sb_hiwat;
1152 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1153 
1154 	/* By the time we drop the lock these should no longer be used. */
1155 	so = NULL;
1156 	tp = NULL;
1157 
1158 #ifdef MAC
1159 	if (mac_syncache_init(&maclabel) != 0) {
1160 		INP_WUNLOCK(inp);
1161 		goto done;
1162 	} else
1163 		mac_syncache_create(maclabel, inp);
1164 #endif
1165 	INP_WUNLOCK(inp);
1166 
1167 	/*
1168 	 * Remember the IP options, if any.
1169 	 */
1170 #ifdef INET6
1171 	if (!(inc->inc_flags & INC_ISIPV6))
1172 #endif
1173 #ifdef INET
1174 		ipopts = (m) ? ip_srcroute(m) : NULL;
1175 #else
1176 		ipopts = NULL;
1177 #endif
1178 
1179 	/*
1180 	 * See if we already have an entry for this connection.
1181 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1182 	 *
1183 	 * XXX: should the syncache be re-initialized with the contents
1184 	 * of the new SYN here (which may have different options?)
1185 	 *
1186 	 * XXX: We do not check the sequence number to see if this is a
1187 	 * real retransmit or a new connection attempt.  The question is
1188 	 * how to handle such a case; either ignore it as spoofed, or
1189 	 * drop the current entry and create a new one?
1190 	 */
1191 	sc = syncache_lookup(inc, &sch);	/* returns locked entry */
1192 	SCH_LOCK_ASSERT(sch);
1193 	if (sc != NULL) {
1194 		TCPSTAT_INC(tcps_sc_dupsyn);
1195 		if (ipopts) {
1196 			/*
1197 			 * If we were remembering a previous source route,
1198 			 * forget it and use the new one we've been given.
1199 			 */
1200 			if (sc->sc_ipopts)
1201 				(void) m_free(sc->sc_ipopts);
1202 			sc->sc_ipopts = ipopts;
1203 		}
1204 		/*
1205 		 * Update timestamp if present.
1206 		 */
1207 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1208 			sc->sc_tsreflect = to->to_tsval;
1209 		else
1210 			sc->sc_flags &= ~SCF_TIMESTAMP;
1211 #ifdef MAC
1212 		/*
1213 		 * Since we have already unconditionally allocated label
1214 		 * storage, free it up.  The syncache entry will already
1215 		 * have an initialized label we can use.
1216 		 */
1217 		mac_syncache_destroy(&maclabel);
1218 #endif
1219 		/* Retransmit SYN|ACK and reset retransmit count. */
1220 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1221 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1222 			    "resetting timer and retransmitting SYN|ACK\n",
1223 			    s, __func__);
1224 			free(s, M_TCPLOG);
1225 		}
1226 		if (syncache_respond(sc, sch, 1) == 0) {
1227 			sc->sc_rxmits = 0;
1228 			syncache_timeout(sc, sch, 1);
1229 			TCPSTAT_INC(tcps_sndacks);
1230 			TCPSTAT_INC(tcps_sndtotal);
1231 		}
1232 		SCH_UNLOCK(sch);
1233 		goto done;
1234 	}
1235 
1236 	sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1237 	if (sc == NULL) {
1238 		/*
1239 		 * The zone allocator couldn't provide more entries.
1240 		 * Treat this as if the cache was full; drop the oldest
1241 		 * entry and insert the new one.
1242 		 */
1243 		TCPSTAT_INC(tcps_sc_zonefail);
1244 		if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL)
1245 			syncache_drop(sc, sch);
1246 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1247 		if (sc == NULL) {
1248 			if (V_tcp_syncookies) {
1249 				bzero(&scs, sizeof(scs));
1250 				sc = &scs;
1251 			} else {
1252 				SCH_UNLOCK(sch);
1253 				if (ipopts)
1254 					(void) m_free(ipopts);
1255 				goto done;
1256 			}
1257 		}
1258 	}
1259 
1260 	/*
1261 	 * Fill in the syncache values.
1262 	 */
1263 #ifdef MAC
1264 	sc->sc_label = maclabel;
1265 #endif
1266 	sc->sc_cred = cred;
1267 	cred = NULL;
1268 	sc->sc_ipopts = ipopts;
1269 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1270 #ifdef INET6
1271 	if (!(inc->inc_flags & INC_ISIPV6))
1272 #endif
1273 	{
1274 		sc->sc_ip_tos = ip_tos;
1275 		sc->sc_ip_ttl = ip_ttl;
1276 	}
1277 #ifdef TCP_OFFLOAD
1278 	sc->sc_tod = tod;
1279 	sc->sc_todctx = todctx;
1280 #endif
1281 	sc->sc_irs = th->th_seq;
1282 	sc->sc_iss = arc4random();
1283 	sc->sc_flags = 0;
1284 	sc->sc_flowlabel = 0;
1285 
1286 	/*
1287 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1288 	 * win was derived from socket earlier in the function.
1289 	 */
1290 	win = imax(win, 0);
1291 	win = imin(win, TCP_MAXWIN);
1292 	sc->sc_wnd = win;
1293 
1294 	if (V_tcp_do_rfc1323) {
1295 		/*
1296 		 * A timestamp received in a SYN makes
1297 		 * it ok to send timestamp requests and replies.
1298 		 */
1299 		if (to->to_flags & TOF_TS) {
1300 			sc->sc_tsreflect = to->to_tsval;
1301 			sc->sc_ts = tcp_ts_getticks();
1302 			sc->sc_flags |= SCF_TIMESTAMP;
1303 		}
1304 		if (to->to_flags & TOF_SCALE) {
1305 			int wscale = 0;
1306 
1307 			/*
1308 			 * Pick the smallest possible scaling factor that
1309 			 * will still allow us to scale up to sb_max, aka
1310 			 * kern.ipc.maxsockbuf.
1311 			 *
1312 			 * We do this because there are broken firewalls that
1313 			 * will corrupt the window scale option, leading to
1314 			 * the other endpoint believing that our advertised
1315 			 * window is unscaled.  At scale factors larger than
1316 			 * 5 the unscaled window will drop below 1500 bytes,
1317 			 * leading to serious problems when traversing these
1318 			 * broken firewalls.
1319 			 *
1320 			 * With the default maxsockbuf of 256K, a scale factor
1321 			 * of 3 will be chosen by this algorithm.  Those who
1322 			 * choose a larger maxsockbuf should watch out
1323 			 * for the compatiblity problems mentioned above.
1324 			 *
1325 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1326 			 * or <SYN,ACK>) segment itself is never scaled.
1327 			 */
1328 			while (wscale < TCP_MAX_WINSHIFT &&
1329 			    (TCP_MAXWIN << wscale) < sb_max)
1330 				wscale++;
1331 			sc->sc_requested_r_scale = wscale;
1332 			sc->sc_requested_s_scale = to->to_wscale;
1333 			sc->sc_flags |= SCF_WINSCALE;
1334 		}
1335 	}
1336 #ifdef TCP_SIGNATURE
1337 	/*
1338 	 * If listening socket requested TCP digests, OR received SYN
1339 	 * contains the option, flag this in the syncache so that
1340 	 * syncache_respond() will do the right thing with the SYN+ACK.
1341 	 */
1342 	if (to->to_flags & TOF_SIGNATURE || ltflags & TF_SIGNATURE)
1343 		sc->sc_flags |= SCF_SIGNATURE;
1344 #endif
1345 	if (to->to_flags & TOF_SACKPERM)
1346 		sc->sc_flags |= SCF_SACK;
1347 	if (to->to_flags & TOF_MSS)
1348 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1349 	if (ltflags & TF_NOOPT)
1350 		sc->sc_flags |= SCF_NOOPT;
1351 	if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1352 		sc->sc_flags |= SCF_ECN;
1353 
1354 	if (V_tcp_syncookies)
1355 		sc->sc_iss = syncookie_generate(sch, sc);
1356 #ifdef INET6
1357 	if (autoflowlabel) {
1358 		if (V_tcp_syncookies)
1359 			sc->sc_flowlabel = sc->sc_iss;
1360 		else
1361 			sc->sc_flowlabel = ip6_randomflowlabel();
1362 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1363 	}
1364 #endif
1365 	SCH_UNLOCK(sch);
1366 
1367 	/*
1368 	 * Do a standard 3-way handshake.
1369 	 */
1370 	if (syncache_respond(sc, sch, 0) == 0) {
1371 		if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1372 			syncache_free(sc);
1373 		else if (sc != &scs)
1374 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1375 		TCPSTAT_INC(tcps_sndacks);
1376 		TCPSTAT_INC(tcps_sndtotal);
1377 	} else {
1378 		if (sc != &scs)
1379 			syncache_free(sc);
1380 		TCPSTAT_INC(tcps_sc_dropped);
1381 	}
1382 
1383 done:
1384 	if (cred != NULL)
1385 		crfree(cred);
1386 #ifdef MAC
1387 	if (sc == &scs)
1388 		mac_syncache_destroy(&maclabel);
1389 #endif
1390 	if (m) {
1391 
1392 		*lsop = NULL;
1393 		m_freem(m);
1394 	}
1395 }
1396 
1397 static int
1398 syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked)
1399 {
1400 	struct ip *ip = NULL;
1401 	struct mbuf *m;
1402 	struct tcphdr *th = NULL;
1403 	int optlen, error = 0;	/* Make compiler happy */
1404 	u_int16_t hlen, tlen, mssopt;
1405 	struct tcpopt to;
1406 #ifdef INET6
1407 	struct ip6_hdr *ip6 = NULL;
1408 #endif
1409 #ifdef TCP_SIGNATURE
1410 	struct secasvar *sav;
1411 #endif
1412 
1413 	hlen =
1414 #ifdef INET6
1415 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1416 #endif
1417 		sizeof(struct ip);
1418 	tlen = hlen + sizeof(struct tcphdr);
1419 
1420 	/* Determine MSS we advertize to other end of connection. */
1421 	mssopt = tcp_mssopt(&sc->sc_inc);
1422 	if (sc->sc_peer_mss)
1423 		mssopt = max( min(sc->sc_peer_mss, mssopt), V_tcp_minmss);
1424 
1425 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1426 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1427 	    ("syncache: mbuf too small"));
1428 
1429 	/* Create the IP+TCP header from scratch. */
1430 	m = m_gethdr(M_NOWAIT, MT_DATA);
1431 	if (m == NULL)
1432 		return (ENOBUFS);
1433 #ifdef MAC
1434 	mac_syncache_create_mbuf(sc->sc_label, m);
1435 #endif
1436 	m->m_data += max_linkhdr;
1437 	m->m_len = tlen;
1438 	m->m_pkthdr.len = tlen;
1439 	m->m_pkthdr.rcvif = NULL;
1440 
1441 #ifdef INET6
1442 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1443 		ip6 = mtod(m, struct ip6_hdr *);
1444 		ip6->ip6_vfc = IPV6_VERSION;
1445 		ip6->ip6_nxt = IPPROTO_TCP;
1446 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1447 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1448 		ip6->ip6_plen = htons(tlen - hlen);
1449 		/* ip6_hlim is set after checksum */
1450 		ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1451 		ip6->ip6_flow |= sc->sc_flowlabel;
1452 
1453 		th = (struct tcphdr *)(ip6 + 1);
1454 	}
1455 #endif
1456 #if defined(INET6) && defined(INET)
1457 	else
1458 #endif
1459 #ifdef INET
1460 	{
1461 		ip = mtod(m, struct ip *);
1462 		ip->ip_v = IPVERSION;
1463 		ip->ip_hl = sizeof(struct ip) >> 2;
1464 		ip->ip_len = htons(tlen);
1465 		ip->ip_id = 0;
1466 		ip->ip_off = 0;
1467 		ip->ip_sum = 0;
1468 		ip->ip_p = IPPROTO_TCP;
1469 		ip->ip_src = sc->sc_inc.inc_laddr;
1470 		ip->ip_dst = sc->sc_inc.inc_faddr;
1471 		ip->ip_ttl = sc->sc_ip_ttl;
1472 		ip->ip_tos = sc->sc_ip_tos;
1473 
1474 		/*
1475 		 * See if we should do MTU discovery.  Route lookups are
1476 		 * expensive, so we will only unset the DF bit if:
1477 		 *
1478 		 *	1) path_mtu_discovery is disabled
1479 		 *	2) the SCF_UNREACH flag has been set
1480 		 */
1481 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1482 		       ip->ip_off |= htons(IP_DF);
1483 
1484 		th = (struct tcphdr *)(ip + 1);
1485 	}
1486 #endif /* INET */
1487 	th->th_sport = sc->sc_inc.inc_lport;
1488 	th->th_dport = sc->sc_inc.inc_fport;
1489 
1490 	th->th_seq = htonl(sc->sc_iss);
1491 	th->th_ack = htonl(sc->sc_irs + 1);
1492 	th->th_off = sizeof(struct tcphdr) >> 2;
1493 	th->th_x2 = 0;
1494 	th->th_flags = TH_SYN|TH_ACK;
1495 	th->th_win = htons(sc->sc_wnd);
1496 	th->th_urp = 0;
1497 
1498 	if (sc->sc_flags & SCF_ECN) {
1499 		th->th_flags |= TH_ECE;
1500 		TCPSTAT_INC(tcps_ecn_shs);
1501 	}
1502 
1503 	/* Tack on the TCP options. */
1504 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1505 		to.to_flags = 0;
1506 
1507 		to.to_mss = mssopt;
1508 		to.to_flags = TOF_MSS;
1509 		if (sc->sc_flags & SCF_WINSCALE) {
1510 			to.to_wscale = sc->sc_requested_r_scale;
1511 			to.to_flags |= TOF_SCALE;
1512 		}
1513 		if (sc->sc_flags & SCF_TIMESTAMP) {
1514 			/* Virgin timestamp or TCP cookie enhanced one. */
1515 			to.to_tsval = sc->sc_ts;
1516 			to.to_tsecr = sc->sc_tsreflect;
1517 			to.to_flags |= TOF_TS;
1518 		}
1519 		if (sc->sc_flags & SCF_SACK)
1520 			to.to_flags |= TOF_SACKPERM;
1521 #ifdef TCP_SIGNATURE
1522 		sav = NULL;
1523 		if (sc->sc_flags & SCF_SIGNATURE) {
1524 			sav = tcp_get_sav(m, IPSEC_DIR_OUTBOUND);
1525 			if (sav != NULL)
1526 				to.to_flags |= TOF_SIGNATURE;
1527 			else {
1528 
1529 				/*
1530 				 * We've got SCF_SIGNATURE flag
1531 				 * inherited from listening socket,
1532 				 * but no SADB key for given source
1533 				 * address. Assume signature is not
1534 				 * required and remove signature flag
1535 				 * instead of silently dropping
1536 				 * connection.
1537 				 */
1538 				if (locked == 0)
1539 					SCH_LOCK(sch);
1540 				sc->sc_flags &= ~SCF_SIGNATURE;
1541 				if (locked == 0)
1542 					SCH_UNLOCK(sch);
1543 			}
1544 		}
1545 #endif
1546 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1547 
1548 		/* Adjust headers by option size. */
1549 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1550 		m->m_len += optlen;
1551 		m->m_pkthdr.len += optlen;
1552 
1553 #ifdef TCP_SIGNATURE
1554 		if (sc->sc_flags & SCF_SIGNATURE)
1555 			tcp_signature_do_compute(m, 0, optlen,
1556 			    to.to_signature, sav);
1557 #endif
1558 #ifdef INET6
1559 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1560 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1561 		else
1562 #endif
1563 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1564 	} else
1565 		optlen = 0;
1566 
1567 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
1568 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1569 #ifdef INET6
1570 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1571 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1572 		th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1573 		    IPPROTO_TCP, 0);
1574 		ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1575 #ifdef TCP_OFFLOAD
1576 		if (ADDED_BY_TOE(sc)) {
1577 			struct toedev *tod = sc->sc_tod;
1578 
1579 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1580 
1581 			return (error);
1582 		}
1583 #endif
1584 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1585 	}
1586 #endif
1587 #if defined(INET6) && defined(INET)
1588 	else
1589 #endif
1590 #ifdef INET
1591 	{
1592 		m->m_pkthdr.csum_flags = CSUM_TCP;
1593 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1594 		    htons(tlen + optlen - hlen + IPPROTO_TCP));
1595 #ifdef TCP_OFFLOAD
1596 		if (ADDED_BY_TOE(sc)) {
1597 			struct toedev *tod = sc->sc_tod;
1598 
1599 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1600 
1601 			return (error);
1602 		}
1603 #endif
1604 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1605 	}
1606 #endif
1607 	return (error);
1608 }
1609 
1610 /*
1611  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1612  * that exceed the capacity of the syncache by avoiding the storage of any
1613  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
1614  * attacks where the attacker does not have access to our responses.
1615  *
1616  * Syncookies encode and include all necessary information about the
1617  * connection setup within the SYN|ACK that we send back.  That way we
1618  * can avoid keeping any local state until the ACK to our SYN|ACK returns
1619  * (if ever).  Normally the syncache and syncookies are running in parallel
1620  * with the latter taking over when the former is exhausted.  When matching
1621  * syncache entry is found the syncookie is ignored.
1622  *
1623  * The only reliable information persisting the 3WHS is our inital sequence
1624  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
1625  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1626  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
1627  * returns and signifies a legitimate connection if it matches the ACK.
1628  *
1629  * The available space of 32 bits to store the hash and to encode the SYN
1630  * option information is very tight and we should have at least 24 bits for
1631  * the MAC to keep the number of guesses by blind spoofing reasonably high.
1632  *
1633  * SYN option information we have to encode to fully restore a connection:
1634  * MSS: is imporant to chose an optimal segment size to avoid IP level
1635  *   fragmentation along the path.  The common MSS values can be encoded
1636  *   in a 3-bit table.  Uncommon values are captured by the next lower value
1637  *   in the table leading to a slight increase in packetization overhead.
1638  * WSCALE: is necessary to allow large windows to be used for high delay-
1639  *   bandwidth product links.  Not scaling the window when it was initially
1640  *   negotiated is bad for performance as lack of scaling further decreases
1641  *   the apparent available send window.  We only need to encode the WSCALE
1642  *   we received from the remote end.  Our end can be recalculated at any
1643  *   time.  The common WSCALE values can be encoded in a 3-bit table.
1644  *   Uncommon values are captured by the next lower value in the table
1645  *   making us under-estimate the available window size halving our
1646  *   theoretically possible maximum throughput for that connection.
1647  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
1648  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
1649  *   that are included in all segments on a connection.  We enable them when
1650  *   the ACK has them.
1651  *
1652  * Security of syncookies and attack vectors:
1653  *
1654  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
1655  * together with the gloabl secret to make it unique per connection attempt.
1656  * Thus any change of any of those parameters results in a different MAC output
1657  * in an unpredictable way unless a collision is encountered.  24 bits of the
1658  * MAC are embedded into the ISS.
1659  *
1660  * To prevent replay attacks two rotating global secrets are updated with a
1661  * new random value every 15 seconds.  The life-time of a syncookie is thus
1662  * 15-30 seconds.
1663  *
1664  * Vector 1: Attacking the secret.  This requires finding a weakness in the
1665  * MAC itself or the way it is used here.  The attacker can do a chosen plain
1666  * text attack by varying and testing the all parameters under his control.
1667  * The strength depends on the size and randomness of the secret, and the
1668  * cryptographic security of the MAC function.  Due to the constant updating
1669  * of the secret the attacker has at most 29.999 seconds to find the secret
1670  * and launch spoofed connections.  After that he has to start all over again.
1671  *
1672  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
1673  * size an average of 4,823 attempts are required for a 50% chance of success
1674  * to spoof a single syncookie (birthday collision paradox).  However the
1675  * attacker is blind and doesn't know if one of his attempts succeeded unless
1676  * he has a side channel to interfere success from.  A single connection setup
1677  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
1678  * This many attempts are required for each one blind spoofed connection.  For
1679  * every additional spoofed connection he has to launch another N attempts.
1680  * Thus for a sustained rate 100 spoofed connections per second approximately
1681  * 1,800,000 packets per second would have to be sent.
1682  *
1683  * NB: The MAC function should be fast so that it doesn't become a CPU
1684  * exhaustion attack vector itself.
1685  *
1686  * References:
1687  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
1688  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
1689  *   http://cr.yp.to/syncookies.html    (overview)
1690  *   http://cr.yp.to/syncookies/archive (details)
1691  *
1692  *
1693  * Schematic construction of a syncookie enabled Initial Sequence Number:
1694  *  0        1         2         3
1695  *  12345678901234567890123456789012
1696  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
1697  *
1698  *  x 24 MAC (truncated)
1699  *  W  3 Send Window Scale index
1700  *  M  3 MSS index
1701  *  S  1 SACK permitted
1702  *  P  1 Odd/even secret
1703  */
1704 
1705 /*
1706  * Distribution and probability of certain MSS values.  Those in between are
1707  * rounded down to the next lower one.
1708  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
1709  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
1710  */
1711 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
1712 
1713 /*
1714  * Distribution and probability of certain WSCALE values.  We have to map the
1715  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
1716  * bits based on prevalence of certain values.  Where we don't have an exact
1717  * match for are rounded down to the next lower one letting us under-estimate
1718  * the true available window.  At the moment this would happen only for the
1719  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
1720  * and window size).  The absence of the WSCALE option (no scaling in either
1721  * direction) is encoded with index zero.
1722  * [WSCALE values histograms, Allman, 2012]
1723  *                            X 10 10 35  5  6 14 10%   by host
1724  *                            X 11  4  5  5 18 49  3%   by connections
1725  */
1726 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
1727 
1728 /*
1729  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
1730  * and good cryptographic properties.
1731  */
1732 static uint32_t
1733 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
1734     uint8_t *secbits, uintptr_t secmod)
1735 {
1736 	SIPHASH_CTX ctx;
1737 	uint32_t siphash[2];
1738 
1739 	SipHash24_Init(&ctx);
1740 	SipHash_SetKey(&ctx, secbits);
1741 	switch (inc->inc_flags & INC_ISIPV6) {
1742 #ifdef INET
1743 	case 0:
1744 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
1745 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
1746 		break;
1747 #endif
1748 #ifdef INET6
1749 	case INC_ISIPV6:
1750 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
1751 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
1752 		break;
1753 #endif
1754 	}
1755 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
1756 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
1757 	SipHash_Update(&ctx, &irs, sizeof(irs));
1758 	SipHash_Update(&ctx, &flags, sizeof(flags));
1759 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
1760 	SipHash_Final((u_int8_t *)&siphash, &ctx);
1761 
1762 	return (siphash[0] ^ siphash[1]);
1763 }
1764 
1765 static tcp_seq
1766 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
1767 {
1768 	u_int i, mss, secbit, wscale;
1769 	uint32_t iss, hash;
1770 	uint8_t *secbits;
1771 	union syncookie cookie;
1772 
1773 	SCH_LOCK_ASSERT(sch);
1774 
1775 	cookie.cookie = 0;
1776 
1777 	/* Map our computed MSS into the 3-bit index. */
1778 	mss = min(tcp_mssopt(&sc->sc_inc), max(sc->sc_peer_mss, V_tcp_minmss));
1779 	for (i = sizeof(tcp_sc_msstab) / sizeof(*tcp_sc_msstab) - 1;
1780 	     tcp_sc_msstab[i] > mss && i > 0;
1781 	     i--)
1782 		;
1783 	cookie.flags.mss_idx = i;
1784 
1785 	/*
1786 	 * Map the send window scale into the 3-bit index but only if
1787 	 * the wscale option was received.
1788 	 */
1789 	if (sc->sc_flags & SCF_WINSCALE) {
1790 		wscale = sc->sc_requested_s_scale;
1791 		for (i = sizeof(tcp_sc_wstab) / sizeof(*tcp_sc_wstab) - 1;
1792 		     tcp_sc_wstab[i] > wscale && i > 0;
1793 		     i--)
1794 			;
1795 		cookie.flags.wscale_idx = i;
1796 	}
1797 
1798 	/* Can we do SACK? */
1799 	if (sc->sc_flags & SCF_SACK)
1800 		cookie.flags.sack_ok = 1;
1801 
1802 	/* Which of the two secrets to use. */
1803 	secbit = sch->sch_sc->secret.oddeven & 0x1;
1804 	cookie.flags.odd_even = secbit;
1805 
1806 	secbits = sch->sch_sc->secret.key[secbit];
1807 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
1808 	    (uintptr_t)sch);
1809 
1810 	/*
1811 	 * Put the flags into the hash and XOR them to get better ISS number
1812 	 * variance.  This doesn't enhance the cryptographic strength and is
1813 	 * done to prevent the 8 cookie bits from showing up directly on the
1814 	 * wire.
1815 	 */
1816 	iss = hash & ~0xff;
1817 	iss |= cookie.cookie ^ (hash >> 24);
1818 
1819 	/* Randomize the timestamp. */
1820 	if (sc->sc_flags & SCF_TIMESTAMP) {
1821 		sc->sc_ts = arc4random();
1822 		sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks();
1823 	}
1824 
1825 	TCPSTAT_INC(tcps_sc_sendcookie);
1826 	return (iss);
1827 }
1828 
1829 static struct syncache *
1830 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
1831     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
1832     struct socket *lso)
1833 {
1834 	uint32_t hash;
1835 	uint8_t *secbits;
1836 	tcp_seq ack, seq;
1837 	int wnd, wscale = 0;
1838 	union syncookie cookie;
1839 
1840 	SCH_LOCK_ASSERT(sch);
1841 
1842 	/*
1843 	 * Pull information out of SYN-ACK/ACK and revert sequence number
1844 	 * advances.
1845 	 */
1846 	ack = th->th_ack - 1;
1847 	seq = th->th_seq - 1;
1848 
1849 	/*
1850 	 * Unpack the flags containing enough information to restore the
1851 	 * connection.
1852 	 */
1853 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
1854 
1855 	/* Which of the two secrets to use. */
1856 	secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
1857 
1858 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
1859 
1860 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
1861 	if ((ack & ~0xff) != (hash & ~0xff))
1862 		return (NULL);
1863 
1864 	/* Fill in the syncache values. */
1865 	sc->sc_flags = 0;
1866 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1867 	sc->sc_ipopts = NULL;
1868 
1869 	sc->sc_irs = seq;
1870 	sc->sc_iss = ack;
1871 
1872 	switch (inc->inc_flags & INC_ISIPV6) {
1873 #ifdef INET
1874 	case 0:
1875 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
1876 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
1877 		break;
1878 #endif
1879 #ifdef INET6
1880 	case INC_ISIPV6:
1881 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
1882 			sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
1883 		break;
1884 #endif
1885 	}
1886 
1887 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
1888 
1889 	/* We can simply recompute receive window scale we sent earlier. */
1890 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
1891 		wscale++;
1892 
1893 	/* Only use wscale if it was enabled in the orignal SYN. */
1894 	if (cookie.flags.wscale_idx > 0) {
1895 		sc->sc_requested_r_scale = wscale;
1896 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
1897 		sc->sc_flags |= SCF_WINSCALE;
1898 	}
1899 
1900 	wnd = sbspace(&lso->so_rcv);
1901 	wnd = imax(wnd, 0);
1902 	wnd = imin(wnd, TCP_MAXWIN);
1903 	sc->sc_wnd = wnd;
1904 
1905 	if (cookie.flags.sack_ok)
1906 		sc->sc_flags |= SCF_SACK;
1907 
1908 	if (to->to_flags & TOF_TS) {
1909 		sc->sc_flags |= SCF_TIMESTAMP;
1910 		sc->sc_tsreflect = to->to_tsval;
1911 		sc->sc_ts = to->to_tsecr;
1912 		sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks();
1913 	}
1914 
1915 	if (to->to_flags & TOF_SIGNATURE)
1916 		sc->sc_flags |= SCF_SIGNATURE;
1917 
1918 	sc->sc_rxmits = 0;
1919 
1920 	TCPSTAT_INC(tcps_sc_recvcookie);
1921 	return (sc);
1922 }
1923 
1924 #ifdef INVARIANTS
1925 static int
1926 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
1927     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
1928     struct socket *lso)
1929 {
1930 	struct syncache scs, *scx;
1931 	char *s;
1932 
1933 	bzero(&scs, sizeof(scs));
1934 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
1935 
1936 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
1937 		return (0);
1938 
1939 	if (scx != NULL) {
1940 		if (sc->sc_peer_mss != scx->sc_peer_mss)
1941 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
1942 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
1943 
1944 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
1945 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
1946 			    s, __func__, sc->sc_requested_r_scale,
1947 			    scx->sc_requested_r_scale);
1948 
1949 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
1950 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
1951 			    s, __func__, sc->sc_requested_s_scale,
1952 			    scx->sc_requested_s_scale);
1953 
1954 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
1955 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
1956 	}
1957 
1958 	if (s != NULL)
1959 		free(s, M_TCPLOG);
1960 	return (0);
1961 }
1962 #endif /* INVARIANTS */
1963 
1964 static void
1965 syncookie_reseed(void *arg)
1966 {
1967 	struct tcp_syncache *sc = arg;
1968 	uint8_t *secbits;
1969 	int secbit;
1970 
1971 	/*
1972 	 * Reseeding the secret doesn't have to be protected by a lock.
1973 	 * It only must be ensured that the new random values are visible
1974 	 * to all CPUs in a SMP environment.  The atomic with release
1975 	 * semantics ensures that.
1976 	 */
1977 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
1978 	secbits = sc->secret.key[secbit];
1979 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
1980 	atomic_add_rel_int(&sc->secret.oddeven, 1);
1981 
1982 	/* Reschedule ourself. */
1983 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
1984 }
1985 
1986 /*
1987  * Returns the current number of syncache entries.  This number
1988  * will probably change before you get around to calling
1989  * syncache_pcblist.
1990  */
1991 int
1992 syncache_pcbcount(void)
1993 {
1994 	struct syncache_head *sch;
1995 	int count, i;
1996 
1997 	for (count = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
1998 		/* No need to lock for a read. */
1999 		sch = &V_tcp_syncache.hashbase[i];
2000 		count += sch->sch_length;
2001 	}
2002 	return count;
2003 }
2004 
2005 /*
2006  * Exports the syncache entries to userland so that netstat can display
2007  * them alongside the other sockets.  This function is intended to be
2008  * called only from tcp_pcblist.
2009  *
2010  * Due to concurrency on an active system, the number of pcbs exported
2011  * may have no relation to max_pcbs.  max_pcbs merely indicates the
2012  * amount of space the caller allocated for this function to use.
2013  */
2014 int
2015 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
2016 {
2017 	struct xtcpcb xt;
2018 	struct syncache *sc;
2019 	struct syncache_head *sch;
2020 	int count, error, i;
2021 
2022 	for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
2023 		sch = &V_tcp_syncache.hashbase[i];
2024 		SCH_LOCK(sch);
2025 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2026 			if (count >= max_pcbs) {
2027 				SCH_UNLOCK(sch);
2028 				goto exit;
2029 			}
2030 			if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2031 				continue;
2032 			bzero(&xt, sizeof(xt));
2033 			xt.xt_len = sizeof(xt);
2034 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2035 				xt.xt_inp.inp_vflag = INP_IPV6;
2036 			else
2037 				xt.xt_inp.inp_vflag = INP_IPV4;
2038 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo));
2039 			xt.xt_tp.t_inpcb = &xt.xt_inp;
2040 			xt.xt_tp.t_state = TCPS_SYN_RECEIVED;
2041 			xt.xt_socket.xso_protocol = IPPROTO_TCP;
2042 			xt.xt_socket.xso_len = sizeof (struct xsocket);
2043 			xt.xt_socket.so_type = SOCK_STREAM;
2044 			xt.xt_socket.so_state = SS_ISCONNECTING;
2045 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2046 			if (error) {
2047 				SCH_UNLOCK(sch);
2048 				goto exit;
2049 			}
2050 			count++;
2051 		}
2052 		SCH_UNLOCK(sch);
2053 	}
2054 exit:
2055 	*pcbs_exported = count;
2056 	return error;
2057 }
2058