xref: /freebsd/sys/netinet/tcp_syncache.c (revision 5ca34122ecdd5abc62bdae39663fec9ac8523d87)
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/hash.h>
44 #include <sys/kernel.h>
45 #include <sys/sysctl.h>
46 #include <sys/limits.h>
47 #include <sys/lock.h>
48 #include <sys/mutex.h>
49 #include <sys/malloc.h>
50 #include <sys/mbuf.h>
51 #include <sys/proc.h>		/* for proc0 declaration */
52 #include <sys/random.h>
53 #include <sys/socket.h>
54 #include <sys/socketvar.h>
55 #include <sys/syslog.h>
56 #include <sys/ucred.h>
57 
58 #include <sys/md5.h>
59 #include <crypto/siphash/siphash.h>
60 
61 #include <vm/uma.h>
62 
63 #include <net/if.h>
64 #include <net/if_var.h>
65 #include <net/route.h>
66 #include <net/vnet.h>
67 
68 #include <netinet/in.h>
69 #include <netinet/in_systm.h>
70 #include <netinet/ip.h>
71 #include <netinet/in_var.h>
72 #include <netinet/in_pcb.h>
73 #include <netinet/ip_var.h>
74 #include <netinet/ip_options.h>
75 #ifdef INET6
76 #include <netinet/ip6.h>
77 #include <netinet/icmp6.h>
78 #include <netinet6/nd6.h>
79 #include <netinet6/ip6_var.h>
80 #include <netinet6/in6_pcb.h>
81 #endif
82 #include <netinet/tcp.h>
83 #include <netinet/tcp_fsm.h>
84 #include <netinet/tcp_seq.h>
85 #include <netinet/tcp_timer.h>
86 #include <netinet/tcp_var.h>
87 #include <netinet/tcp_syncache.h>
88 #ifdef INET6
89 #include <netinet6/tcp6_var.h>
90 #endif
91 #ifdef TCP_OFFLOAD
92 #include <netinet/toecore.h>
93 #endif
94 
95 #ifdef IPSEC
96 #include <netipsec/ipsec.h>
97 #ifdef INET6
98 #include <netipsec/ipsec6.h>
99 #endif
100 #include <netipsec/key.h>
101 #endif /*IPSEC*/
102 
103 #include <machine/in_cksum.h>
104 
105 #include <security/mac/mac_framework.h>
106 
107 static VNET_DEFINE(int, tcp_syncookies) = 1;
108 #define	V_tcp_syncookies		VNET(tcp_syncookies)
109 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
110     &VNET_NAME(tcp_syncookies), 0,
111     "Use TCP SYN cookies if the syncache overflows");
112 
113 static VNET_DEFINE(int, tcp_syncookiesonly) = 0;
114 #define	V_tcp_syncookiesonly		VNET(tcp_syncookiesonly)
115 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
116     &VNET_NAME(tcp_syncookiesonly), 0,
117     "Use only TCP SYN cookies");
118 
119 #ifdef TCP_OFFLOAD
120 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
121 #endif
122 
123 static void	 syncache_drop(struct syncache *, struct syncache_head *);
124 static void	 syncache_free(struct syncache *);
125 static void	 syncache_insert(struct syncache *, struct syncache_head *);
126 static int	 syncache_respond(struct syncache *, struct syncache_head *, int);
127 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
128 		    struct mbuf *m);
129 static void	 syncache_timeout(struct syncache *sc, struct syncache_head *sch,
130 		    int docallout);
131 static void	 syncache_timer(void *);
132 
133 static uint32_t	 syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
134 		    uint8_t *, uintptr_t);
135 static tcp_seq	 syncookie_generate(struct syncache_head *, struct syncache *);
136 static struct syncache
137 		*syncookie_lookup(struct in_conninfo *, struct syncache_head *,
138 		    struct syncache *, struct tcphdr *, struct tcpopt *,
139 		    struct socket *);
140 static void	 syncookie_reseed(void *);
141 #ifdef INVARIANTS
142 static int	 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
143 		    struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
144 		    struct socket *lso);
145 #endif
146 
147 /*
148  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
149  * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds,
150  * the odds are that the user has given up attempting to connect by then.
151  */
152 #define SYNCACHE_MAXREXMTS		3
153 
154 /* Arbitrary values */
155 #define TCP_SYNCACHE_HASHSIZE		512
156 #define TCP_SYNCACHE_BUCKETLIMIT	30
157 
158 static VNET_DEFINE(struct tcp_syncache, tcp_syncache);
159 #define	V_tcp_syncache			VNET(tcp_syncache)
160 
161 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0,
162     "TCP SYN cache");
163 
164 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
165     &VNET_NAME(tcp_syncache.bucket_limit), 0,
166     "Per-bucket hash limit for syncache");
167 
168 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
169     &VNET_NAME(tcp_syncache.cache_limit), 0,
170     "Overall entry limit for syncache");
171 
172 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
173     &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
174 
175 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
176     &VNET_NAME(tcp_syncache.hashsize), 0,
177     "Size of TCP syncache hashtable");
178 
179 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_VNET | CTLFLAG_RW,
180     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
181     "Limit on SYN/ACK retransmissions");
182 
183 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
184 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
185     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
186     "Send reset on socket allocation failure");
187 
188 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
189 
190 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
191 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
192 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
193 
194 /*
195  * Requires the syncache entry to be already removed from the bucket list.
196  */
197 static void
198 syncache_free(struct syncache *sc)
199 {
200 
201 	if (sc->sc_ipopts)
202 		(void) m_free(sc->sc_ipopts);
203 	if (sc->sc_cred)
204 		crfree(sc->sc_cred);
205 #ifdef MAC
206 	mac_syncache_destroy(&sc->sc_label);
207 #endif
208 
209 	uma_zfree(V_tcp_syncache.zone, sc);
210 }
211 
212 void
213 syncache_init(void)
214 {
215 	int i;
216 
217 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
218 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
219 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
220 	V_tcp_syncache.hash_secret = arc4random();
221 
222 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
223 	    &V_tcp_syncache.hashsize);
224 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
225 	    &V_tcp_syncache.bucket_limit);
226 	if (!powerof2(V_tcp_syncache.hashsize) ||
227 	    V_tcp_syncache.hashsize == 0) {
228 		printf("WARNING: syncache hash size is not a power of 2.\n");
229 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
230 	}
231 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
232 
233 	/* Set limits. */
234 	V_tcp_syncache.cache_limit =
235 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
236 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
237 	    &V_tcp_syncache.cache_limit);
238 
239 	/* Allocate the hash table. */
240 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
241 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
242 
243 #ifdef VIMAGE
244 	V_tcp_syncache.vnet = curvnet;
245 #endif
246 
247 	/* Initialize the hash buckets. */
248 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
249 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
250 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
251 			 NULL, MTX_DEF);
252 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
253 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
254 		V_tcp_syncache.hashbase[i].sch_length = 0;
255 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
256 	}
257 
258 	/* Create the syncache entry zone. */
259 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
260 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
261 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
262 	    V_tcp_syncache.cache_limit);
263 
264 	/* Start the SYN cookie reseeder callout. */
265 	callout_init(&V_tcp_syncache.secret.reseed, 1);
266 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
267 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
268 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
269 	    syncookie_reseed, &V_tcp_syncache);
270 }
271 
272 #ifdef VIMAGE
273 void
274 syncache_destroy(void)
275 {
276 	struct syncache_head *sch;
277 	struct syncache *sc, *nsc;
278 	int i;
279 
280 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
281 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
282 
283 		sch = &V_tcp_syncache.hashbase[i];
284 		callout_drain(&sch->sch_timer);
285 
286 		SCH_LOCK(sch);
287 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
288 			syncache_drop(sc, sch);
289 		SCH_UNLOCK(sch);
290 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
291 		    ("%s: sch->sch_bucket not empty", __func__));
292 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
293 		    __func__, sch->sch_length));
294 		mtx_destroy(&sch->sch_mtx);
295 	}
296 
297 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
298 	    ("%s: cache_count not 0", __func__));
299 
300 	/* Free the allocated global resources. */
301 	uma_zdestroy(V_tcp_syncache.zone);
302 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
303 
304 	callout_drain(&V_tcp_syncache.secret.reseed);
305 }
306 #endif
307 
308 /*
309  * Inserts a syncache entry into the specified bucket row.
310  * Locks and unlocks the syncache_head autonomously.
311  */
312 static void
313 syncache_insert(struct syncache *sc, struct syncache_head *sch)
314 {
315 	struct syncache *sc2;
316 
317 	SCH_LOCK(sch);
318 
319 	/*
320 	 * Make sure that we don't overflow the per-bucket limit.
321 	 * If the bucket is full, toss the oldest element.
322 	 */
323 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
324 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
325 			("sch->sch_length incorrect"));
326 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
327 		syncache_drop(sc2, sch);
328 		TCPSTAT_INC(tcps_sc_bucketoverflow);
329 	}
330 
331 	/* Put it into the bucket. */
332 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
333 	sch->sch_length++;
334 
335 #ifdef TCP_OFFLOAD
336 	if (ADDED_BY_TOE(sc)) {
337 		struct toedev *tod = sc->sc_tod;
338 
339 		tod->tod_syncache_added(tod, sc->sc_todctx);
340 	}
341 #endif
342 
343 	/* Reinitialize the bucket row's timer. */
344 	if (sch->sch_length == 1)
345 		sch->sch_nextc = ticks + INT_MAX;
346 	syncache_timeout(sc, sch, 1);
347 
348 	SCH_UNLOCK(sch);
349 
350 	TCPSTAT_INC(tcps_sc_added);
351 }
352 
353 /*
354  * Remove and free entry from syncache bucket row.
355  * Expects locked syncache head.
356  */
357 static void
358 syncache_drop(struct syncache *sc, struct syncache_head *sch)
359 {
360 
361 	SCH_LOCK_ASSERT(sch);
362 
363 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
364 	sch->sch_length--;
365 
366 #ifdef TCP_OFFLOAD
367 	if (ADDED_BY_TOE(sc)) {
368 		struct toedev *tod = sc->sc_tod;
369 
370 		tod->tod_syncache_removed(tod, sc->sc_todctx);
371 	}
372 #endif
373 
374 	syncache_free(sc);
375 }
376 
377 /*
378  * Engage/reengage time on bucket row.
379  */
380 static void
381 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
382 {
383 	sc->sc_rxttime = ticks +
384 		TCPTV_RTOBASE * (tcp_syn_backoff[sc->sc_rxmits]);
385 	sc->sc_rxmits++;
386 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
387 		sch->sch_nextc = sc->sc_rxttime;
388 		if (docallout)
389 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
390 			    syncache_timer, (void *)sch);
391 	}
392 }
393 
394 /*
395  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
396  * If we have retransmitted an entry the maximum number of times, expire it.
397  * One separate timer for each bucket row.
398  */
399 static void
400 syncache_timer(void *xsch)
401 {
402 	struct syncache_head *sch = (struct syncache_head *)xsch;
403 	struct syncache *sc, *nsc;
404 	int tick = ticks;
405 	char *s;
406 
407 	CURVNET_SET(sch->sch_sc->vnet);
408 
409 	/* NB: syncache_head has already been locked by the callout. */
410 	SCH_LOCK_ASSERT(sch);
411 
412 	/*
413 	 * In the following cycle we may remove some entries and/or
414 	 * advance some timeouts, so re-initialize the bucket timer.
415 	 */
416 	sch->sch_nextc = tick + INT_MAX;
417 
418 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
419 		/*
420 		 * We do not check if the listen socket still exists
421 		 * and accept the case where the listen socket may be
422 		 * gone by the time we resend the SYN/ACK.  We do
423 		 * not expect this to happens often. If it does,
424 		 * then the RST will be sent by the time the remote
425 		 * host does the SYN/ACK->ACK.
426 		 */
427 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
428 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
429 				sch->sch_nextc = sc->sc_rxttime;
430 			continue;
431 		}
432 		if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
433 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
434 				log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
435 				    "giving up and removing syncache entry\n",
436 				    s, __func__);
437 				free(s, M_TCPLOG);
438 			}
439 			syncache_drop(sc, sch);
440 			TCPSTAT_INC(tcps_sc_stale);
441 			continue;
442 		}
443 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
444 			log(LOG_DEBUG, "%s; %s: Response timeout, "
445 			    "retransmitting (%u) SYN|ACK\n",
446 			    s, __func__, sc->sc_rxmits);
447 			free(s, M_TCPLOG);
448 		}
449 
450 		syncache_respond(sc, sch, 1);
451 		TCPSTAT_INC(tcps_sc_retransmitted);
452 		syncache_timeout(sc, sch, 0);
453 	}
454 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
455 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
456 			syncache_timer, (void *)(sch));
457 	CURVNET_RESTORE();
458 }
459 
460 /*
461  * Find an entry in the syncache.
462  * Returns always with locked syncache_head plus a matching entry or NULL.
463  */
464 static struct syncache *
465 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
466 {
467 	struct syncache *sc;
468 	struct syncache_head *sch;
469 	uint32_t hash;
470 
471 	/*
472 	 * The hash is built on foreign port + local port + foreign address.
473 	 * We rely on the fact that struct in_conninfo starts with 16 bits
474 	 * of foreign port, then 16 bits of local port then followed by 128
475 	 * bits of foreign address.  In case of IPv4 address, the first 3
476 	 * 32-bit words of the address always are zeroes.
477 	 */
478 	hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
479 	    V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
480 
481 	sch = &V_tcp_syncache.hashbase[hash];
482 	*schp = sch;
483 	SCH_LOCK(sch);
484 
485 	/* Circle through bucket row to find matching entry. */
486 	TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
487 		if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
488 		    sizeof(struct in_endpoints)) == 0)
489 			break;
490 
491 	return (sc);	/* Always returns with locked sch. */
492 }
493 
494 /*
495  * This function is called when we get a RST for a
496  * non-existent connection, so that we can see if the
497  * connection is in the syn cache.  If it is, zap it.
498  */
499 void
500 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
501 {
502 	struct syncache *sc;
503 	struct syncache_head *sch;
504 	char *s = NULL;
505 
506 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
507 	SCH_LOCK_ASSERT(sch);
508 
509 	/*
510 	 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
511 	 * See RFC 793 page 65, section SEGMENT ARRIVES.
512 	 */
513 	if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
514 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
515 			log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
516 			    "FIN flag set, segment ignored\n", s, __func__);
517 		TCPSTAT_INC(tcps_badrst);
518 		goto done;
519 	}
520 
521 	/*
522 	 * No corresponding connection was found in syncache.
523 	 * If syncookies are enabled and possibly exclusively
524 	 * used, or we are under memory pressure, a valid RST
525 	 * may not find a syncache entry.  In that case we're
526 	 * done and no SYN|ACK retransmissions will happen.
527 	 * Otherwise the RST was misdirected or spoofed.
528 	 */
529 	if (sc == NULL) {
530 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
531 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
532 			    "syncache entry (possibly syncookie only), "
533 			    "segment ignored\n", s, __func__);
534 		TCPSTAT_INC(tcps_badrst);
535 		goto done;
536 	}
537 
538 	/*
539 	 * If the RST bit is set, check the sequence number to see
540 	 * if this is a valid reset segment.
541 	 * RFC 793 page 37:
542 	 *   In all states except SYN-SENT, all reset (RST) segments
543 	 *   are validated by checking their SEQ-fields.  A reset is
544 	 *   valid if its sequence number is in the window.
545 	 *
546 	 *   The sequence number in the reset segment is normally an
547 	 *   echo of our outgoing acknowlegement numbers, but some hosts
548 	 *   send a reset with the sequence number at the rightmost edge
549 	 *   of our receive window, and we have to handle this case.
550 	 */
551 	if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
552 	    SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
553 		syncache_drop(sc, sch);
554 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
555 			log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, "
556 			    "connection attempt aborted by remote endpoint\n",
557 			    s, __func__);
558 		TCPSTAT_INC(tcps_sc_reset);
559 	} else {
560 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
561 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
562 			    "IRS %u (+WND %u), segment ignored\n",
563 			    s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd);
564 		TCPSTAT_INC(tcps_badrst);
565 	}
566 
567 done:
568 	if (s != NULL)
569 		free(s, M_TCPLOG);
570 	SCH_UNLOCK(sch);
571 }
572 
573 void
574 syncache_badack(struct in_conninfo *inc)
575 {
576 	struct syncache *sc;
577 	struct syncache_head *sch;
578 
579 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
580 	SCH_LOCK_ASSERT(sch);
581 	if (sc != NULL) {
582 		syncache_drop(sc, sch);
583 		TCPSTAT_INC(tcps_sc_badack);
584 	}
585 	SCH_UNLOCK(sch);
586 }
587 
588 void
589 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th)
590 {
591 	struct syncache *sc;
592 	struct syncache_head *sch;
593 
594 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
595 	SCH_LOCK_ASSERT(sch);
596 	if (sc == NULL)
597 		goto done;
598 
599 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
600 	if (ntohl(th->th_seq) != sc->sc_iss)
601 		goto done;
602 
603 	/*
604 	 * If we've rertransmitted 3 times and this is our second error,
605 	 * we remove the entry.  Otherwise, we allow it to continue on.
606 	 * This prevents us from incorrectly nuking an entry during a
607 	 * spurious network outage.
608 	 *
609 	 * See tcp_notify().
610 	 */
611 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
612 		sc->sc_flags |= SCF_UNREACH;
613 		goto done;
614 	}
615 	syncache_drop(sc, sch);
616 	TCPSTAT_INC(tcps_sc_unreach);
617 done:
618 	SCH_UNLOCK(sch);
619 }
620 
621 /*
622  * Build a new TCP socket structure from a syncache entry.
623  *
624  * On success return the newly created socket with its underlying inp locked.
625  */
626 static struct socket *
627 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
628 {
629 	struct inpcb *inp = NULL;
630 	struct socket *so;
631 	struct tcpcb *tp;
632 	int error;
633 	char *s;
634 
635 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
636 
637 	/*
638 	 * Ok, create the full blown connection, and set things up
639 	 * as they would have been set up if we had created the
640 	 * connection when the SYN arrived.  If we can't create
641 	 * the connection, abort it.
642 	 */
643 	so = sonewconn(lso, 0);
644 	if (so == NULL) {
645 		/*
646 		 * Drop the connection; we will either send a RST or
647 		 * have the peer retransmit its SYN again after its
648 		 * RTO and try again.
649 		 */
650 		TCPSTAT_INC(tcps_listendrop);
651 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
652 			log(LOG_DEBUG, "%s; %s: Socket create failed "
653 			    "due to limits or memory shortage\n",
654 			    s, __func__);
655 			free(s, M_TCPLOG);
656 		}
657 		goto abort2;
658 	}
659 #ifdef MAC
660 	mac_socketpeer_set_from_mbuf(m, so);
661 #endif
662 
663 	inp = sotoinpcb(so);
664 	inp->inp_inc.inc_fibnum = so->so_fibnum;
665 	INP_WLOCK(inp);
666 	/*
667 	 * Exclusive pcbinfo lock is not required in syncache socket case even
668 	 * if two inpcb locks can be acquired simultaneously:
669 	 *  - the inpcb in LISTEN state,
670 	 *  - the newly created inp.
671 	 *
672 	 * In this case, an inp cannot be at same time in LISTEN state and
673 	 * just created by an accept() call.
674 	 */
675 	INP_HASH_WLOCK(&V_tcbinfo);
676 
677 	/* Insert new socket into PCB hash list. */
678 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
679 #ifdef INET6
680 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
681 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
682 	} else {
683 		inp->inp_vflag &= ~INP_IPV6;
684 		inp->inp_vflag |= INP_IPV4;
685 #endif
686 		inp->inp_laddr = sc->sc_inc.inc_laddr;
687 #ifdef INET6
688 	}
689 #endif
690 
691 	/*
692 	 * If there's an mbuf and it has a flowid, then let's initialise the
693 	 * inp with that particular flowid.
694 	 */
695 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
696 		inp->inp_flowid = m->m_pkthdr.flowid;
697 		inp->inp_flowtype = M_HASHTYPE_GET(m);
698 	}
699 
700 	/*
701 	 * Install in the reservation hash table for now, but don't yet
702 	 * install a connection group since the full 4-tuple isn't yet
703 	 * configured.
704 	 */
705 	inp->inp_lport = sc->sc_inc.inc_lport;
706 	if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
707 		/*
708 		 * Undo the assignments above if we failed to
709 		 * put the PCB on the hash lists.
710 		 */
711 #ifdef INET6
712 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
713 			inp->in6p_laddr = in6addr_any;
714 		else
715 #endif
716 			inp->inp_laddr.s_addr = INADDR_ANY;
717 		inp->inp_lport = 0;
718 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
719 			log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
720 			    "with error %i\n",
721 			    s, __func__, error);
722 			free(s, M_TCPLOG);
723 		}
724 		INP_HASH_WUNLOCK(&V_tcbinfo);
725 		goto abort;
726 	}
727 #ifdef IPSEC
728 	/* Copy old policy into new socket's. */
729 	if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp))
730 		printf("syncache_socket: could not copy policy\n");
731 #endif
732 #ifdef INET6
733 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
734 		struct inpcb *oinp = sotoinpcb(lso);
735 		struct in6_addr laddr6;
736 		struct sockaddr_in6 sin6;
737 		/*
738 		 * Inherit socket options from the listening socket.
739 		 * Note that in6p_inputopts are not (and should not be)
740 		 * copied, since it stores previously received options and is
741 		 * used to detect if each new option is different than the
742 		 * previous one and hence should be passed to a user.
743 		 * If we copied in6p_inputopts, a user would not be able to
744 		 * receive options just after calling the accept system call.
745 		 */
746 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
747 		if (oinp->in6p_outputopts)
748 			inp->in6p_outputopts =
749 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
750 
751 		sin6.sin6_family = AF_INET6;
752 		sin6.sin6_len = sizeof(sin6);
753 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
754 		sin6.sin6_port = sc->sc_inc.inc_fport;
755 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
756 		laddr6 = inp->in6p_laddr;
757 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
758 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
759 		if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
760 		    thread0.td_ucred, m)) != 0) {
761 			inp->in6p_laddr = laddr6;
762 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
763 				log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
764 				    "with error %i\n",
765 				    s, __func__, error);
766 				free(s, M_TCPLOG);
767 			}
768 			INP_HASH_WUNLOCK(&V_tcbinfo);
769 			goto abort;
770 		}
771 		/* Override flowlabel from in6_pcbconnect. */
772 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
773 		inp->inp_flow |= sc->sc_flowlabel;
774 	}
775 #endif /* INET6 */
776 #if defined(INET) && defined(INET6)
777 	else
778 #endif
779 #ifdef INET
780 	{
781 		struct in_addr laddr;
782 		struct sockaddr_in sin;
783 
784 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
785 
786 		if (inp->inp_options == NULL) {
787 			inp->inp_options = sc->sc_ipopts;
788 			sc->sc_ipopts = NULL;
789 		}
790 
791 		sin.sin_family = AF_INET;
792 		sin.sin_len = sizeof(sin);
793 		sin.sin_addr = sc->sc_inc.inc_faddr;
794 		sin.sin_port = sc->sc_inc.inc_fport;
795 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
796 		laddr = inp->inp_laddr;
797 		if (inp->inp_laddr.s_addr == INADDR_ANY)
798 			inp->inp_laddr = sc->sc_inc.inc_laddr;
799 		if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
800 		    thread0.td_ucred, m)) != 0) {
801 			inp->inp_laddr = laddr;
802 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
803 				log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
804 				    "with error %i\n",
805 				    s, __func__, error);
806 				free(s, M_TCPLOG);
807 			}
808 			INP_HASH_WUNLOCK(&V_tcbinfo);
809 			goto abort;
810 		}
811 	}
812 #endif /* INET */
813 	INP_HASH_WUNLOCK(&V_tcbinfo);
814 	tp = intotcpcb(inp);
815 	tcp_state_change(tp, TCPS_SYN_RECEIVED);
816 	tp->iss = sc->sc_iss;
817 	tp->irs = sc->sc_irs;
818 	tcp_rcvseqinit(tp);
819 	tcp_sendseqinit(tp);
820 	tp->snd_wl1 = sc->sc_irs;
821 	tp->snd_max = tp->iss + 1;
822 	tp->snd_nxt = tp->iss + 1;
823 	tp->rcv_up = sc->sc_irs + 1;
824 	tp->rcv_wnd = sc->sc_wnd;
825 	tp->rcv_adv += tp->rcv_wnd;
826 	tp->last_ack_sent = tp->rcv_nxt;
827 
828 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
829 	if (sc->sc_flags & SCF_NOOPT)
830 		tp->t_flags |= TF_NOOPT;
831 	else {
832 		if (sc->sc_flags & SCF_WINSCALE) {
833 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
834 			tp->snd_scale = sc->sc_requested_s_scale;
835 			tp->request_r_scale = sc->sc_requested_r_scale;
836 		}
837 		if (sc->sc_flags & SCF_TIMESTAMP) {
838 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
839 			tp->ts_recent = sc->sc_tsreflect;
840 			tp->ts_recent_age = tcp_ts_getticks();
841 			tp->ts_offset = sc->sc_tsoff;
842 		}
843 #ifdef TCP_SIGNATURE
844 		if (sc->sc_flags & SCF_SIGNATURE)
845 			tp->t_flags |= TF_SIGNATURE;
846 #endif
847 		if (sc->sc_flags & SCF_SACK)
848 			tp->t_flags |= TF_SACK_PERMIT;
849 	}
850 
851 	if (sc->sc_flags & SCF_ECN)
852 		tp->t_flags |= TF_ECN_PERMIT;
853 
854 	/*
855 	 * Set up MSS and get cached values from tcp_hostcache.
856 	 * This might overwrite some of the defaults we just set.
857 	 */
858 	tcp_mss(tp, sc->sc_peer_mss);
859 
860 	/*
861 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
862 	 * limited to one segment in cc_conn_init().
863 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
864 	 */
865 	if (sc->sc_rxmits > 1)
866 		tp->snd_cwnd = 1;
867 
868 #ifdef TCP_OFFLOAD
869 	/*
870 	 * Allow a TOE driver to install its hooks.  Note that we hold the
871 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
872 	 * new connection before the TOE driver has done its thing.
873 	 */
874 	if (ADDED_BY_TOE(sc)) {
875 		struct toedev *tod = sc->sc_tod;
876 
877 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
878 	}
879 #endif
880 	/*
881 	 * Copy and activate timers.
882 	 */
883 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
884 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
885 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
886 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
887 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
888 
889 	soisconnected(so);
890 
891 	TCPSTAT_INC(tcps_accepts);
892 	return (so);
893 
894 abort:
895 	INP_WUNLOCK(inp);
896 abort2:
897 	if (so != NULL)
898 		soabort(so);
899 	return (NULL);
900 }
901 
902 /*
903  * This function gets called when we receive an ACK for a
904  * socket in the LISTEN state.  We look up the connection
905  * in the syncache, and if its there, we pull it out of
906  * the cache and turn it into a full-blown connection in
907  * the SYN-RECEIVED state.
908  *
909  * On syncache_socket() success the newly created socket
910  * has its underlying inp locked.
911  */
912 int
913 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
914     struct socket **lsop, struct mbuf *m)
915 {
916 	struct syncache *sc;
917 	struct syncache_head *sch;
918 	struct syncache scs;
919 	char *s;
920 
921 	/*
922 	 * Global TCP locks are held because we manipulate the PCB lists
923 	 * and create a new socket.
924 	 */
925 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
926 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
927 	    ("%s: can handle only ACK", __func__));
928 
929 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
930 	SCH_LOCK_ASSERT(sch);
931 
932 #ifdef INVARIANTS
933 	/*
934 	 * Test code for syncookies comparing the syncache stored
935 	 * values with the reconstructed values from the cookie.
936 	 */
937 	if (sc != NULL)
938 		syncookie_cmp(inc, sch, sc, th, to, *lsop);
939 #endif
940 
941 	if (sc == NULL) {
942 		/*
943 		 * There is no syncache entry, so see if this ACK is
944 		 * a returning syncookie.  To do this, first:
945 		 *  A. See if this socket has had a syncache entry dropped in
946 		 *     the past.  We don't want to accept a bogus syncookie
947 		 *     if we've never received a SYN.
948 		 *  B. check that the syncookie is valid.  If it is, then
949 		 *     cobble up a fake syncache entry, and return.
950 		 */
951 		if (!V_tcp_syncookies) {
952 			SCH_UNLOCK(sch);
953 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
954 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
955 				    "segment rejected (syncookies disabled)\n",
956 				    s, __func__);
957 			goto failed;
958 		}
959 		bzero(&scs, sizeof(scs));
960 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
961 		SCH_UNLOCK(sch);
962 		if (sc == NULL) {
963 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
964 				log(LOG_DEBUG, "%s; %s: Segment failed "
965 				    "SYNCOOKIE authentication, segment rejected "
966 				    "(probably spoofed)\n", s, __func__);
967 			goto failed;
968 		}
969 	} else {
970 		/* Pull out the entry to unlock the bucket row. */
971 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
972 		sch->sch_length--;
973 #ifdef TCP_OFFLOAD
974 		if (ADDED_BY_TOE(sc)) {
975 			struct toedev *tod = sc->sc_tod;
976 
977 			tod->tod_syncache_removed(tod, sc->sc_todctx);
978 		}
979 #endif
980 		SCH_UNLOCK(sch);
981 	}
982 
983 	/*
984 	 * Segment validation:
985 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
986 	 */
987 	if (th->th_ack != sc->sc_iss + 1) {
988 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
989 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
990 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
991 		goto failed;
992 	}
993 
994 	/*
995 	 * The SEQ must fall in the window starting at the received
996 	 * initial receive sequence number + 1 (the SYN).
997 	 */
998 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
999 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1000 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1001 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1002 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1003 		goto failed;
1004 	}
1005 
1006 	/*
1007 	 * If timestamps were not negotiated during SYN/ACK they
1008 	 * must not appear on any segment during this session.
1009 	 */
1010 	if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1011 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1012 			log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1013 			    "segment rejected\n", s, __func__);
1014 		goto failed;
1015 	}
1016 
1017 	/*
1018 	 * If timestamps were negotiated during SYN/ACK they should
1019 	 * appear on every segment during this session.
1020 	 * XXXAO: This is only informal as there have been unverified
1021 	 * reports of non-compliants stacks.
1022 	 */
1023 	if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1024 		if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1025 			log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1026 			    "no action\n", s, __func__);
1027 			free(s, M_TCPLOG);
1028 			s = NULL;
1029 		}
1030 	}
1031 
1032 	/*
1033 	 * If timestamps were negotiated the reflected timestamp
1034 	 * must be equal to what we actually sent in the SYN|ACK.
1035 	 */
1036 	if ((to->to_flags & TOF_TS) && to->to_tsecr != sc->sc_ts) {
1037 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1038 			log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, "
1039 			    "segment rejected\n",
1040 			    s, __func__, to->to_tsecr, sc->sc_ts);
1041 		goto failed;
1042 	}
1043 
1044 	*lsop = syncache_socket(sc, *lsop, m);
1045 
1046 	if (*lsop == NULL)
1047 		TCPSTAT_INC(tcps_sc_aborted);
1048 	else
1049 		TCPSTAT_INC(tcps_sc_completed);
1050 
1051 /* how do we find the inp for the new socket? */
1052 	if (sc != &scs)
1053 		syncache_free(sc);
1054 	return (1);
1055 failed:
1056 	if (sc != NULL && sc != &scs)
1057 		syncache_free(sc);
1058 	if (s != NULL)
1059 		free(s, M_TCPLOG);
1060 	*lsop = NULL;
1061 	return (0);
1062 }
1063 
1064 /*
1065  * Given a LISTEN socket and an inbound SYN request, add
1066  * this to the syn cache, and send back a segment:
1067  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1068  * to the source.
1069  *
1070  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1071  * Doing so would require that we hold onto the data and deliver it
1072  * to the application.  However, if we are the target of a SYN-flood
1073  * DoS attack, an attacker could send data which would eventually
1074  * consume all available buffer space if it were ACKed.  By not ACKing
1075  * the data, we avoid this DoS scenario.
1076  */
1077 void
1078 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1079     struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1080     void *todctx)
1081 {
1082 	struct tcpcb *tp;
1083 	struct socket *so;
1084 	struct syncache *sc = NULL;
1085 	struct syncache_head *sch;
1086 	struct mbuf *ipopts = NULL;
1087 	u_int ltflags;
1088 	int win, sb_hiwat, ip_ttl, ip_tos;
1089 	char *s;
1090 #ifdef INET6
1091 	int autoflowlabel = 0;
1092 #endif
1093 #ifdef MAC
1094 	struct label *maclabel;
1095 #endif
1096 	struct syncache scs;
1097 	struct ucred *cred;
1098 
1099 	INP_WLOCK_ASSERT(inp);			/* listen socket */
1100 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1101 	    ("%s: unexpected tcp flags", __func__));
1102 
1103 	/*
1104 	 * Combine all so/tp operations very early to drop the INP lock as
1105 	 * soon as possible.
1106 	 */
1107 	so = *lsop;
1108 	tp = sototcpcb(so);
1109 	cred = crhold(so->so_cred);
1110 
1111 #ifdef INET6
1112 	if ((inc->inc_flags & INC_ISIPV6) &&
1113 	    (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1114 		autoflowlabel = 1;
1115 #endif
1116 	ip_ttl = inp->inp_ip_ttl;
1117 	ip_tos = inp->inp_ip_tos;
1118 	win = sbspace(&so->so_rcv);
1119 	sb_hiwat = so->so_rcv.sb_hiwat;
1120 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1121 
1122 	/* By the time we drop the lock these should no longer be used. */
1123 	so = NULL;
1124 	tp = NULL;
1125 
1126 #ifdef MAC
1127 	if (mac_syncache_init(&maclabel) != 0) {
1128 		INP_WUNLOCK(inp);
1129 		goto done;
1130 	} else
1131 		mac_syncache_create(maclabel, inp);
1132 #endif
1133 	INP_WUNLOCK(inp);
1134 
1135 	/*
1136 	 * Remember the IP options, if any.
1137 	 */
1138 #ifdef INET6
1139 	if (!(inc->inc_flags & INC_ISIPV6))
1140 #endif
1141 #ifdef INET
1142 		ipopts = (m) ? ip_srcroute(m) : NULL;
1143 #else
1144 		ipopts = NULL;
1145 #endif
1146 
1147 	/*
1148 	 * See if we already have an entry for this connection.
1149 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1150 	 *
1151 	 * XXX: should the syncache be re-initialized with the contents
1152 	 * of the new SYN here (which may have different options?)
1153 	 *
1154 	 * XXX: We do not check the sequence number to see if this is a
1155 	 * real retransmit or a new connection attempt.  The question is
1156 	 * how to handle such a case; either ignore it as spoofed, or
1157 	 * drop the current entry and create a new one?
1158 	 */
1159 	sc = syncache_lookup(inc, &sch);	/* returns locked entry */
1160 	SCH_LOCK_ASSERT(sch);
1161 	if (sc != NULL) {
1162 		TCPSTAT_INC(tcps_sc_dupsyn);
1163 		if (ipopts) {
1164 			/*
1165 			 * If we were remembering a previous source route,
1166 			 * forget it and use the new one we've been given.
1167 			 */
1168 			if (sc->sc_ipopts)
1169 				(void) m_free(sc->sc_ipopts);
1170 			sc->sc_ipopts = ipopts;
1171 		}
1172 		/*
1173 		 * Update timestamp if present.
1174 		 */
1175 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1176 			sc->sc_tsreflect = to->to_tsval;
1177 		else
1178 			sc->sc_flags &= ~SCF_TIMESTAMP;
1179 #ifdef MAC
1180 		/*
1181 		 * Since we have already unconditionally allocated label
1182 		 * storage, free it up.  The syncache entry will already
1183 		 * have an initialized label we can use.
1184 		 */
1185 		mac_syncache_destroy(&maclabel);
1186 #endif
1187 		/* Retransmit SYN|ACK and reset retransmit count. */
1188 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1189 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1190 			    "resetting timer and retransmitting SYN|ACK\n",
1191 			    s, __func__);
1192 			free(s, M_TCPLOG);
1193 		}
1194 		if (syncache_respond(sc, sch, 1) == 0) {
1195 			sc->sc_rxmits = 0;
1196 			syncache_timeout(sc, sch, 1);
1197 			TCPSTAT_INC(tcps_sndacks);
1198 			TCPSTAT_INC(tcps_sndtotal);
1199 		}
1200 		SCH_UNLOCK(sch);
1201 		goto done;
1202 	}
1203 
1204 	sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1205 	if (sc == NULL) {
1206 		/*
1207 		 * The zone allocator couldn't provide more entries.
1208 		 * Treat this as if the cache was full; drop the oldest
1209 		 * entry and insert the new one.
1210 		 */
1211 		TCPSTAT_INC(tcps_sc_zonefail);
1212 		if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL)
1213 			syncache_drop(sc, sch);
1214 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1215 		if (sc == NULL) {
1216 			if (V_tcp_syncookies) {
1217 				bzero(&scs, sizeof(scs));
1218 				sc = &scs;
1219 			} else {
1220 				SCH_UNLOCK(sch);
1221 				if (ipopts)
1222 					(void) m_free(ipopts);
1223 				goto done;
1224 			}
1225 		}
1226 	}
1227 
1228 	/*
1229 	 * Fill in the syncache values.
1230 	 */
1231 #ifdef MAC
1232 	sc->sc_label = maclabel;
1233 #endif
1234 	sc->sc_cred = cred;
1235 	cred = NULL;
1236 	sc->sc_ipopts = ipopts;
1237 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1238 #ifdef INET6
1239 	if (!(inc->inc_flags & INC_ISIPV6))
1240 #endif
1241 	{
1242 		sc->sc_ip_tos = ip_tos;
1243 		sc->sc_ip_ttl = ip_ttl;
1244 	}
1245 #ifdef TCP_OFFLOAD
1246 	sc->sc_tod = tod;
1247 	sc->sc_todctx = todctx;
1248 #endif
1249 	sc->sc_irs = th->th_seq;
1250 	sc->sc_iss = arc4random();
1251 	sc->sc_flags = 0;
1252 	sc->sc_flowlabel = 0;
1253 
1254 	/*
1255 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1256 	 * win was derived from socket earlier in the function.
1257 	 */
1258 	win = imax(win, 0);
1259 	win = imin(win, TCP_MAXWIN);
1260 	sc->sc_wnd = win;
1261 
1262 	if (V_tcp_do_rfc1323) {
1263 		/*
1264 		 * A timestamp received in a SYN makes
1265 		 * it ok to send timestamp requests and replies.
1266 		 */
1267 		if (to->to_flags & TOF_TS) {
1268 			sc->sc_tsreflect = to->to_tsval;
1269 			sc->sc_ts = tcp_ts_getticks();
1270 			sc->sc_flags |= SCF_TIMESTAMP;
1271 		}
1272 		if (to->to_flags & TOF_SCALE) {
1273 			int wscale = 0;
1274 
1275 			/*
1276 			 * Pick the smallest possible scaling factor that
1277 			 * will still allow us to scale up to sb_max, aka
1278 			 * kern.ipc.maxsockbuf.
1279 			 *
1280 			 * We do this because there are broken firewalls that
1281 			 * will corrupt the window scale option, leading to
1282 			 * the other endpoint believing that our advertised
1283 			 * window is unscaled.  At scale factors larger than
1284 			 * 5 the unscaled window will drop below 1500 bytes,
1285 			 * leading to serious problems when traversing these
1286 			 * broken firewalls.
1287 			 *
1288 			 * With the default maxsockbuf of 256K, a scale factor
1289 			 * of 3 will be chosen by this algorithm.  Those who
1290 			 * choose a larger maxsockbuf should watch out
1291 			 * for the compatiblity problems mentioned above.
1292 			 *
1293 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1294 			 * or <SYN,ACK>) segment itself is never scaled.
1295 			 */
1296 			while (wscale < TCP_MAX_WINSHIFT &&
1297 			    (TCP_MAXWIN << wscale) < sb_max)
1298 				wscale++;
1299 			sc->sc_requested_r_scale = wscale;
1300 			sc->sc_requested_s_scale = to->to_wscale;
1301 			sc->sc_flags |= SCF_WINSCALE;
1302 		}
1303 	}
1304 #ifdef TCP_SIGNATURE
1305 	/*
1306 	 * If listening socket requested TCP digests, OR received SYN
1307 	 * contains the option, flag this in the syncache so that
1308 	 * syncache_respond() will do the right thing with the SYN+ACK.
1309 	 */
1310 	if (to->to_flags & TOF_SIGNATURE || ltflags & TF_SIGNATURE)
1311 		sc->sc_flags |= SCF_SIGNATURE;
1312 #endif
1313 	if (to->to_flags & TOF_SACKPERM)
1314 		sc->sc_flags |= SCF_SACK;
1315 	if (to->to_flags & TOF_MSS)
1316 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1317 	if (ltflags & TF_NOOPT)
1318 		sc->sc_flags |= SCF_NOOPT;
1319 	if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1320 		sc->sc_flags |= SCF_ECN;
1321 
1322 	if (V_tcp_syncookies)
1323 		sc->sc_iss = syncookie_generate(sch, sc);
1324 #ifdef INET6
1325 	if (autoflowlabel) {
1326 		if (V_tcp_syncookies)
1327 			sc->sc_flowlabel = sc->sc_iss;
1328 		else
1329 			sc->sc_flowlabel = ip6_randomflowlabel();
1330 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1331 	}
1332 #endif
1333 	SCH_UNLOCK(sch);
1334 
1335 	/*
1336 	 * Do a standard 3-way handshake.
1337 	 */
1338 	if (syncache_respond(sc, sch, 0) == 0) {
1339 		if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1340 			syncache_free(sc);
1341 		else if (sc != &scs)
1342 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1343 		TCPSTAT_INC(tcps_sndacks);
1344 		TCPSTAT_INC(tcps_sndtotal);
1345 	} else {
1346 		if (sc != &scs)
1347 			syncache_free(sc);
1348 		TCPSTAT_INC(tcps_sc_dropped);
1349 	}
1350 
1351 done:
1352 	if (cred != NULL)
1353 		crfree(cred);
1354 #ifdef MAC
1355 	if (sc == &scs)
1356 		mac_syncache_destroy(&maclabel);
1357 #endif
1358 	if (m) {
1359 
1360 		*lsop = NULL;
1361 		m_freem(m);
1362 	}
1363 }
1364 
1365 static int
1366 syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked)
1367 {
1368 	struct ip *ip = NULL;
1369 	struct mbuf *m;
1370 	struct tcphdr *th = NULL;
1371 	int optlen, error = 0;	/* Make compiler happy */
1372 	u_int16_t hlen, tlen, mssopt;
1373 	struct tcpopt to;
1374 #ifdef INET6
1375 	struct ip6_hdr *ip6 = NULL;
1376 #endif
1377 #ifdef TCP_SIGNATURE
1378 	struct secasvar *sav;
1379 #endif
1380 
1381 	hlen =
1382 #ifdef INET6
1383 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1384 #endif
1385 		sizeof(struct ip);
1386 	tlen = hlen + sizeof(struct tcphdr);
1387 
1388 	/* Determine MSS we advertize to other end of connection. */
1389 	mssopt = tcp_mssopt(&sc->sc_inc);
1390 	if (sc->sc_peer_mss)
1391 		mssopt = max( min(sc->sc_peer_mss, mssopt), V_tcp_minmss);
1392 
1393 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1394 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1395 	    ("syncache: mbuf too small"));
1396 
1397 	/* Create the IP+TCP header from scratch. */
1398 	m = m_gethdr(M_NOWAIT, MT_DATA);
1399 	if (m == NULL)
1400 		return (ENOBUFS);
1401 #ifdef MAC
1402 	mac_syncache_create_mbuf(sc->sc_label, m);
1403 #endif
1404 	m->m_data += max_linkhdr;
1405 	m->m_len = tlen;
1406 	m->m_pkthdr.len = tlen;
1407 	m->m_pkthdr.rcvif = NULL;
1408 
1409 #ifdef INET6
1410 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1411 		ip6 = mtod(m, struct ip6_hdr *);
1412 		ip6->ip6_vfc = IPV6_VERSION;
1413 		ip6->ip6_nxt = IPPROTO_TCP;
1414 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1415 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1416 		ip6->ip6_plen = htons(tlen - hlen);
1417 		/* ip6_hlim is set after checksum */
1418 		ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1419 		ip6->ip6_flow |= sc->sc_flowlabel;
1420 
1421 		th = (struct tcphdr *)(ip6 + 1);
1422 	}
1423 #endif
1424 #if defined(INET6) && defined(INET)
1425 	else
1426 #endif
1427 #ifdef INET
1428 	{
1429 		ip = mtod(m, struct ip *);
1430 		ip->ip_v = IPVERSION;
1431 		ip->ip_hl = sizeof(struct ip) >> 2;
1432 		ip->ip_len = htons(tlen);
1433 		ip->ip_id = 0;
1434 		ip->ip_off = 0;
1435 		ip->ip_sum = 0;
1436 		ip->ip_p = IPPROTO_TCP;
1437 		ip->ip_src = sc->sc_inc.inc_laddr;
1438 		ip->ip_dst = sc->sc_inc.inc_faddr;
1439 		ip->ip_ttl = sc->sc_ip_ttl;
1440 		ip->ip_tos = sc->sc_ip_tos;
1441 
1442 		/*
1443 		 * See if we should do MTU discovery.  Route lookups are
1444 		 * expensive, so we will only unset the DF bit if:
1445 		 *
1446 		 *	1) path_mtu_discovery is disabled
1447 		 *	2) the SCF_UNREACH flag has been set
1448 		 */
1449 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1450 		       ip->ip_off |= htons(IP_DF);
1451 
1452 		th = (struct tcphdr *)(ip + 1);
1453 	}
1454 #endif /* INET */
1455 	th->th_sport = sc->sc_inc.inc_lport;
1456 	th->th_dport = sc->sc_inc.inc_fport;
1457 
1458 	th->th_seq = htonl(sc->sc_iss);
1459 	th->th_ack = htonl(sc->sc_irs + 1);
1460 	th->th_off = sizeof(struct tcphdr) >> 2;
1461 	th->th_x2 = 0;
1462 	th->th_flags = TH_SYN|TH_ACK;
1463 	th->th_win = htons(sc->sc_wnd);
1464 	th->th_urp = 0;
1465 
1466 	if (sc->sc_flags & SCF_ECN) {
1467 		th->th_flags |= TH_ECE;
1468 		TCPSTAT_INC(tcps_ecn_shs);
1469 	}
1470 
1471 	/* Tack on the TCP options. */
1472 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1473 		to.to_flags = 0;
1474 
1475 		to.to_mss = mssopt;
1476 		to.to_flags = TOF_MSS;
1477 		if (sc->sc_flags & SCF_WINSCALE) {
1478 			to.to_wscale = sc->sc_requested_r_scale;
1479 			to.to_flags |= TOF_SCALE;
1480 		}
1481 		if (sc->sc_flags & SCF_TIMESTAMP) {
1482 			/* Virgin timestamp or TCP cookie enhanced one. */
1483 			to.to_tsval = sc->sc_ts;
1484 			to.to_tsecr = sc->sc_tsreflect;
1485 			to.to_flags |= TOF_TS;
1486 		}
1487 		if (sc->sc_flags & SCF_SACK)
1488 			to.to_flags |= TOF_SACKPERM;
1489 #ifdef TCP_SIGNATURE
1490 		sav = NULL;
1491 		if (sc->sc_flags & SCF_SIGNATURE) {
1492 			sav = tcp_get_sav(m, IPSEC_DIR_OUTBOUND);
1493 			if (sav != NULL)
1494 				to.to_flags |= TOF_SIGNATURE;
1495 			else {
1496 
1497 				/*
1498 				 * We've got SCF_SIGNATURE flag
1499 				 * inherited from listening socket,
1500 				 * but no SADB key for given source
1501 				 * address. Assume signature is not
1502 				 * required and remove signature flag
1503 				 * instead of silently dropping
1504 				 * connection.
1505 				 */
1506 				if (locked == 0)
1507 					SCH_LOCK(sch);
1508 				sc->sc_flags &= ~SCF_SIGNATURE;
1509 				if (locked == 0)
1510 					SCH_UNLOCK(sch);
1511 			}
1512 		}
1513 #endif
1514 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1515 
1516 		/* Adjust headers by option size. */
1517 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1518 		m->m_len += optlen;
1519 		m->m_pkthdr.len += optlen;
1520 
1521 #ifdef TCP_SIGNATURE
1522 		if (sc->sc_flags & SCF_SIGNATURE)
1523 			tcp_signature_do_compute(m, 0, optlen,
1524 			    to.to_signature, sav);
1525 #endif
1526 #ifdef INET6
1527 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1528 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1529 		else
1530 #endif
1531 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1532 	} else
1533 		optlen = 0;
1534 
1535 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
1536 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1537 #ifdef INET6
1538 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1539 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1540 		th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1541 		    IPPROTO_TCP, 0);
1542 		ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1543 #ifdef TCP_OFFLOAD
1544 		if (ADDED_BY_TOE(sc)) {
1545 			struct toedev *tod = sc->sc_tod;
1546 
1547 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1548 
1549 			return (error);
1550 		}
1551 #endif
1552 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1553 	}
1554 #endif
1555 #if defined(INET6) && defined(INET)
1556 	else
1557 #endif
1558 #ifdef INET
1559 	{
1560 		m->m_pkthdr.csum_flags = CSUM_TCP;
1561 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1562 		    htons(tlen + optlen - hlen + IPPROTO_TCP));
1563 #ifdef TCP_OFFLOAD
1564 		if (ADDED_BY_TOE(sc)) {
1565 			struct toedev *tod = sc->sc_tod;
1566 
1567 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1568 
1569 			return (error);
1570 		}
1571 #endif
1572 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1573 	}
1574 #endif
1575 	return (error);
1576 }
1577 
1578 /*
1579  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1580  * that exceed the capacity of the syncache by avoiding the storage of any
1581  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
1582  * attacks where the attacker does not have access to our responses.
1583  *
1584  * Syncookies encode and include all necessary information about the
1585  * connection setup within the SYN|ACK that we send back.  That way we
1586  * can avoid keeping any local state until the ACK to our SYN|ACK returns
1587  * (if ever).  Normally the syncache and syncookies are running in parallel
1588  * with the latter taking over when the former is exhausted.  When matching
1589  * syncache entry is found the syncookie is ignored.
1590  *
1591  * The only reliable information persisting the 3WHS is our inital sequence
1592  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
1593  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1594  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
1595  * returns and signifies a legitimate connection if it matches the ACK.
1596  *
1597  * The available space of 32 bits to store the hash and to encode the SYN
1598  * option information is very tight and we should have at least 24 bits for
1599  * the MAC to keep the number of guesses by blind spoofing reasonably high.
1600  *
1601  * SYN option information we have to encode to fully restore a connection:
1602  * MSS: is imporant to chose an optimal segment size to avoid IP level
1603  *   fragmentation along the path.  The common MSS values can be encoded
1604  *   in a 3-bit table.  Uncommon values are captured by the next lower value
1605  *   in the table leading to a slight increase in packetization overhead.
1606  * WSCALE: is necessary to allow large windows to be used for high delay-
1607  *   bandwidth product links.  Not scaling the window when it was initially
1608  *   negotiated is bad for performance as lack of scaling further decreases
1609  *   the apparent available send window.  We only need to encode the WSCALE
1610  *   we received from the remote end.  Our end can be recalculated at any
1611  *   time.  The common WSCALE values can be encoded in a 3-bit table.
1612  *   Uncommon values are captured by the next lower value in the table
1613  *   making us under-estimate the available window size halving our
1614  *   theoretically possible maximum throughput for that connection.
1615  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
1616  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
1617  *   that are included in all segments on a connection.  We enable them when
1618  *   the ACK has them.
1619  *
1620  * Security of syncookies and attack vectors:
1621  *
1622  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
1623  * together with the gloabl secret to make it unique per connection attempt.
1624  * Thus any change of any of those parameters results in a different MAC output
1625  * in an unpredictable way unless a collision is encountered.  24 bits of the
1626  * MAC are embedded into the ISS.
1627  *
1628  * To prevent replay attacks two rotating global secrets are updated with a
1629  * new random value every 15 seconds.  The life-time of a syncookie is thus
1630  * 15-30 seconds.
1631  *
1632  * Vector 1: Attacking the secret.  This requires finding a weakness in the
1633  * MAC itself or the way it is used here.  The attacker can do a chosen plain
1634  * text attack by varying and testing the all parameters under his control.
1635  * The strength depends on the size and randomness of the secret, and the
1636  * cryptographic security of the MAC function.  Due to the constant updating
1637  * of the secret the attacker has at most 29.999 seconds to find the secret
1638  * and launch spoofed connections.  After that he has to start all over again.
1639  *
1640  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
1641  * size an average of 4,823 attempts are required for a 50% chance of success
1642  * to spoof a single syncookie (birthday collision paradox).  However the
1643  * attacker is blind and doesn't know if one of his attempts succeeded unless
1644  * he has a side channel to interfere success from.  A single connection setup
1645  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
1646  * This many attempts are required for each one blind spoofed connection.  For
1647  * every additional spoofed connection he has to launch another N attempts.
1648  * Thus for a sustained rate 100 spoofed connections per second approximately
1649  * 1,800,000 packets per second would have to be sent.
1650  *
1651  * NB: The MAC function should be fast so that it doesn't become a CPU
1652  * exhaustion attack vector itself.
1653  *
1654  * References:
1655  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
1656  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
1657  *   http://cr.yp.to/syncookies.html    (overview)
1658  *   http://cr.yp.to/syncookies/archive (details)
1659  *
1660  *
1661  * Schematic construction of a syncookie enabled Initial Sequence Number:
1662  *  0        1         2         3
1663  *  12345678901234567890123456789012
1664  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
1665  *
1666  *  x 24 MAC (truncated)
1667  *  W  3 Send Window Scale index
1668  *  M  3 MSS index
1669  *  S  1 SACK permitted
1670  *  P  1 Odd/even secret
1671  */
1672 
1673 /*
1674  * Distribution and probability of certain MSS values.  Those in between are
1675  * rounded down to the next lower one.
1676  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
1677  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
1678  */
1679 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
1680 
1681 /*
1682  * Distribution and probability of certain WSCALE values.  We have to map the
1683  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
1684  * bits based on prevalence of certain values.  Where we don't have an exact
1685  * match for are rounded down to the next lower one letting us under-estimate
1686  * the true available window.  At the moment this would happen only for the
1687  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
1688  * and window size).  The absence of the WSCALE option (no scaling in either
1689  * direction) is encoded with index zero.
1690  * [WSCALE values histograms, Allman, 2012]
1691  *                            X 10 10 35  5  6 14 10%   by host
1692  *                            X 11  4  5  5 18 49  3%   by connections
1693  */
1694 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
1695 
1696 /*
1697  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
1698  * and good cryptographic properties.
1699  */
1700 static uint32_t
1701 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
1702     uint8_t *secbits, uintptr_t secmod)
1703 {
1704 	SIPHASH_CTX ctx;
1705 	uint32_t siphash[2];
1706 
1707 	SipHash24_Init(&ctx);
1708 	SipHash_SetKey(&ctx, secbits);
1709 	switch (inc->inc_flags & INC_ISIPV6) {
1710 #ifdef INET
1711 	case 0:
1712 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
1713 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
1714 		break;
1715 #endif
1716 #ifdef INET6
1717 	case INC_ISIPV6:
1718 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
1719 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
1720 		break;
1721 #endif
1722 	}
1723 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
1724 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
1725 	SipHash_Update(&ctx, &irs, sizeof(irs));
1726 	SipHash_Update(&ctx, &flags, sizeof(flags));
1727 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
1728 	SipHash_Final((u_int8_t *)&siphash, &ctx);
1729 
1730 	return (siphash[0] ^ siphash[1]);
1731 }
1732 
1733 static tcp_seq
1734 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
1735 {
1736 	u_int i, mss, secbit, wscale;
1737 	uint32_t iss, hash;
1738 	uint8_t *secbits;
1739 	union syncookie cookie;
1740 
1741 	SCH_LOCK_ASSERT(sch);
1742 
1743 	cookie.cookie = 0;
1744 
1745 	/* Map our computed MSS into the 3-bit index. */
1746 	mss = min(tcp_mssopt(&sc->sc_inc), max(sc->sc_peer_mss, V_tcp_minmss));
1747 	for (i = sizeof(tcp_sc_msstab) / sizeof(*tcp_sc_msstab) - 1;
1748 	     tcp_sc_msstab[i] > mss && i > 0;
1749 	     i--)
1750 		;
1751 	cookie.flags.mss_idx = i;
1752 
1753 	/*
1754 	 * Map the send window scale into the 3-bit index but only if
1755 	 * the wscale option was received.
1756 	 */
1757 	if (sc->sc_flags & SCF_WINSCALE) {
1758 		wscale = sc->sc_requested_s_scale;
1759 		for (i = sizeof(tcp_sc_wstab) / sizeof(*tcp_sc_wstab) - 1;
1760 		     tcp_sc_wstab[i] > wscale && i > 0;
1761 		     i--)
1762 			;
1763 		cookie.flags.wscale_idx = i;
1764 	}
1765 
1766 	/* Can we do SACK? */
1767 	if (sc->sc_flags & SCF_SACK)
1768 		cookie.flags.sack_ok = 1;
1769 
1770 	/* Which of the two secrets to use. */
1771 	secbit = sch->sch_sc->secret.oddeven & 0x1;
1772 	cookie.flags.odd_even = secbit;
1773 
1774 	secbits = sch->sch_sc->secret.key[secbit];
1775 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
1776 	    (uintptr_t)sch);
1777 
1778 	/*
1779 	 * Put the flags into the hash and XOR them to get better ISS number
1780 	 * variance.  This doesn't enhance the cryptographic strength and is
1781 	 * done to prevent the 8 cookie bits from showing up directly on the
1782 	 * wire.
1783 	 */
1784 	iss = hash & ~0xff;
1785 	iss |= cookie.cookie ^ (hash >> 24);
1786 
1787 	/* Randomize the timestamp. */
1788 	if (sc->sc_flags & SCF_TIMESTAMP) {
1789 		sc->sc_ts = arc4random();
1790 		sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks();
1791 	}
1792 
1793 	TCPSTAT_INC(tcps_sc_sendcookie);
1794 	return (iss);
1795 }
1796 
1797 static struct syncache *
1798 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
1799     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
1800     struct socket *lso)
1801 {
1802 	uint32_t hash;
1803 	uint8_t *secbits;
1804 	tcp_seq ack, seq;
1805 	int wnd, wscale = 0;
1806 	union syncookie cookie;
1807 
1808 	SCH_LOCK_ASSERT(sch);
1809 
1810 	/*
1811 	 * Pull information out of SYN-ACK/ACK and revert sequence number
1812 	 * advances.
1813 	 */
1814 	ack = th->th_ack - 1;
1815 	seq = th->th_seq - 1;
1816 
1817 	/*
1818 	 * Unpack the flags containing enough information to restore the
1819 	 * connection.
1820 	 */
1821 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
1822 
1823 	/* Which of the two secrets to use. */
1824 	secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
1825 
1826 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
1827 
1828 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
1829 	if ((ack & ~0xff) != (hash & ~0xff))
1830 		return (NULL);
1831 
1832 	/* Fill in the syncache values. */
1833 	sc->sc_flags = 0;
1834 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1835 	sc->sc_ipopts = NULL;
1836 
1837 	sc->sc_irs = seq;
1838 	sc->sc_iss = ack;
1839 
1840 	switch (inc->inc_flags & INC_ISIPV6) {
1841 #ifdef INET
1842 	case 0:
1843 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
1844 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
1845 		break;
1846 #endif
1847 #ifdef INET6
1848 	case INC_ISIPV6:
1849 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
1850 			sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
1851 		break;
1852 #endif
1853 	}
1854 
1855 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
1856 
1857 	/* We can simply recompute receive window scale we sent earlier. */
1858 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
1859 		wscale++;
1860 
1861 	/* Only use wscale if it was enabled in the orignal SYN. */
1862 	if (cookie.flags.wscale_idx > 0) {
1863 		sc->sc_requested_r_scale = wscale;
1864 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
1865 		sc->sc_flags |= SCF_WINSCALE;
1866 	}
1867 
1868 	wnd = sbspace(&lso->so_rcv);
1869 	wnd = imax(wnd, 0);
1870 	wnd = imin(wnd, TCP_MAXWIN);
1871 	sc->sc_wnd = wnd;
1872 
1873 	if (cookie.flags.sack_ok)
1874 		sc->sc_flags |= SCF_SACK;
1875 
1876 	if (to->to_flags & TOF_TS) {
1877 		sc->sc_flags |= SCF_TIMESTAMP;
1878 		sc->sc_tsreflect = to->to_tsval;
1879 		sc->sc_ts = to->to_tsecr;
1880 		sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks();
1881 	}
1882 
1883 	if (to->to_flags & TOF_SIGNATURE)
1884 		sc->sc_flags |= SCF_SIGNATURE;
1885 
1886 	sc->sc_rxmits = 0;
1887 
1888 	TCPSTAT_INC(tcps_sc_recvcookie);
1889 	return (sc);
1890 }
1891 
1892 #ifdef INVARIANTS
1893 static int
1894 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
1895     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
1896     struct socket *lso)
1897 {
1898 	struct syncache scs, *scx;
1899 	char *s;
1900 
1901 	bzero(&scs, sizeof(scs));
1902 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
1903 
1904 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
1905 		return (0);
1906 
1907 	if (scx != NULL) {
1908 		if (sc->sc_peer_mss != scx->sc_peer_mss)
1909 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
1910 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
1911 
1912 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
1913 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
1914 			    s, __func__, sc->sc_requested_r_scale,
1915 			    scx->sc_requested_r_scale);
1916 
1917 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
1918 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
1919 			    s, __func__, sc->sc_requested_s_scale,
1920 			    scx->sc_requested_s_scale);
1921 
1922 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
1923 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
1924 	}
1925 
1926 	if (s != NULL)
1927 		free(s, M_TCPLOG);
1928 	return (0);
1929 }
1930 #endif /* INVARIANTS */
1931 
1932 static void
1933 syncookie_reseed(void *arg)
1934 {
1935 	struct tcp_syncache *sc = arg;
1936 	uint8_t *secbits;
1937 	int secbit;
1938 
1939 	/*
1940 	 * Reseeding the secret doesn't have to be protected by a lock.
1941 	 * It only must be ensured that the new random values are visible
1942 	 * to all CPUs in a SMP environment.  The atomic with release
1943 	 * semantics ensures that.
1944 	 */
1945 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
1946 	secbits = sc->secret.key[secbit];
1947 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
1948 	atomic_add_rel_int(&sc->secret.oddeven, 1);
1949 
1950 	/* Reschedule ourself. */
1951 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
1952 }
1953 
1954 /*
1955  * Returns the current number of syncache entries.  This number
1956  * will probably change before you get around to calling
1957  * syncache_pcblist.
1958  */
1959 int
1960 syncache_pcbcount(void)
1961 {
1962 	struct syncache_head *sch;
1963 	int count, i;
1964 
1965 	for (count = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
1966 		/* No need to lock for a read. */
1967 		sch = &V_tcp_syncache.hashbase[i];
1968 		count += sch->sch_length;
1969 	}
1970 	return count;
1971 }
1972 
1973 /*
1974  * Exports the syncache entries to userland so that netstat can display
1975  * them alongside the other sockets.  This function is intended to be
1976  * called only from tcp_pcblist.
1977  *
1978  * Due to concurrency on an active system, the number of pcbs exported
1979  * may have no relation to max_pcbs.  max_pcbs merely indicates the
1980  * amount of space the caller allocated for this function to use.
1981  */
1982 int
1983 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
1984 {
1985 	struct xtcpcb xt;
1986 	struct syncache *sc;
1987 	struct syncache_head *sch;
1988 	int count, error, i;
1989 
1990 	for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
1991 		sch = &V_tcp_syncache.hashbase[i];
1992 		SCH_LOCK(sch);
1993 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
1994 			if (count >= max_pcbs) {
1995 				SCH_UNLOCK(sch);
1996 				goto exit;
1997 			}
1998 			if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
1999 				continue;
2000 			bzero(&xt, sizeof(xt));
2001 			xt.xt_len = sizeof(xt);
2002 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2003 				xt.xt_inp.inp_vflag = INP_IPV6;
2004 			else
2005 				xt.xt_inp.inp_vflag = INP_IPV4;
2006 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo));
2007 			xt.xt_tp.t_inpcb = &xt.xt_inp;
2008 			xt.xt_tp.t_state = TCPS_SYN_RECEIVED;
2009 			xt.xt_socket.xso_protocol = IPPROTO_TCP;
2010 			xt.xt_socket.xso_len = sizeof (struct xsocket);
2011 			xt.xt_socket.so_type = SOCK_STREAM;
2012 			xt.xt_socket.so_state = SS_ISCONNECTING;
2013 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2014 			if (error) {
2015 				SCH_UNLOCK(sch);
2016 				goto exit;
2017 			}
2018 			count++;
2019 		}
2020 		SCH_UNLOCK(sch);
2021 	}
2022 exit:
2023 	*pcbs_exported = count;
2024 	return error;
2025 }
2026