xref: /freebsd/sys/netinet/tcp_syncache.c (revision f5ce3f4ef562ea9fc4d8f9c13c268f48a5bacba7)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2001 McAfee, Inc.
5  * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG
6  * All rights reserved.
7  *
8  * This software was developed for the FreeBSD Project by Jonathan Lemon
9  * and McAfee Research, the Security Research Division of McAfee, Inc. under
10  * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
11  * DARPA CHATS research program. [2001 McAfee, Inc.]
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  */
34 
35 #include <sys/cdefs.h>
36 #include "opt_inet.h"
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/hash.h>
43 #include <sys/refcount.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_kdtrace.h>
70 #include <netinet/in_systm.h>
71 #include <netinet/ip.h>
72 #include <netinet/in_var.h>
73 #include <netinet/in_pcb.h>
74 #include <netinet/ip_var.h>
75 #include <netinet/ip_options.h>
76 #ifdef INET6
77 #include <netinet/ip6.h>
78 #include <netinet/icmp6.h>
79 #include <netinet6/nd6.h>
80 #include <netinet6/ip6_var.h>
81 #include <netinet6/in6_pcb.h>
82 #endif
83 #include <netinet/tcp.h>
84 #include <netinet/tcp_fastopen.h>
85 #include <netinet/tcp_fsm.h>
86 #include <netinet/tcp_seq.h>
87 #include <netinet/tcp_timer.h>
88 #include <netinet/tcp_var.h>
89 #include <netinet/tcp_syncache.h>
90 #include <netinet/tcp_ecn.h>
91 #ifdef TCP_BLACKBOX
92 #include <netinet/tcp_log_buf.h>
93 #endif
94 #ifdef TCP_OFFLOAD
95 #include <netinet/toecore.h>
96 #endif
97 #include <netinet/udp.h>
98 
99 #include <netipsec/ipsec_support.h>
100 
101 #include <machine/in_cksum.h>
102 
103 #include <security/mac/mac_framework.h>
104 
105 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1;
106 #define	V_tcp_syncookies		VNET(tcp_syncookies)
107 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
108     &VNET_NAME(tcp_syncookies), 0,
109     "Use TCP SYN cookies if the syncache overflows");
110 
111 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0;
112 #define	V_tcp_syncookiesonly		VNET(tcp_syncookiesonly)
113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
114     &VNET_NAME(tcp_syncookiesonly), 0,
115     "Use only TCP SYN cookies");
116 
117 #ifdef TCP_OFFLOAD
118 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
119 #endif
120 
121 static void	 syncache_drop(struct syncache *, struct syncache_head *);
122 static void	 syncache_free(struct syncache *);
123 static void	 syncache_insert(struct syncache *, struct syncache_head *);
124 static int	 syncache_respond(struct syncache *, const struct mbuf *, int);
125 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
126 		    struct mbuf *m);
127 static void	 syncache_timeout(struct syncache *sc, struct syncache_head *sch,
128 		    int docallout);
129 static void	 syncache_timer(void *);
130 
131 static uint32_t	 syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
132 		    uint8_t *, uintptr_t);
133 static tcp_seq	 syncookie_generate(struct syncache_head *, struct syncache *);
134 static struct syncache
135 		*syncookie_lookup(struct in_conninfo *, struct syncache_head *,
136 		    struct syncache *, struct tcphdr *, struct tcpopt *,
137 		    struct socket *, uint16_t);
138 static void	syncache_pause(struct in_conninfo *);
139 static void	syncache_unpause(void *);
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, uint16_t port);
145 #endif
146 
147 /*
148  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
149  * 3 retransmits corresponds to a timeout with default values of
150  * tcp_rexmit_initial * (             1 +
151  *                       tcp_backoff[1] +
152  *                       tcp_backoff[2] +
153  *                       tcp_backoff[3]) + 3 * tcp_rexmit_slop,
154  * 1000 ms * (1 + 2 + 4 + 8) +  3 * 200 ms = 15600 ms,
155  * the odds are that the user has given up attempting to connect by then.
156  */
157 #define SYNCACHE_MAXREXMTS		3
158 
159 /* Arbitrary values */
160 #define TCP_SYNCACHE_HASHSIZE		512
161 #define TCP_SYNCACHE_BUCKETLIMIT	30
162 
163 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
164 #define	V_tcp_syncache			VNET(tcp_syncache)
165 
166 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache,
167     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
168     "TCP SYN cache");
169 
170 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
171     &VNET_NAME(tcp_syncache.bucket_limit), 0,
172     "Per-bucket hash limit for syncache");
173 
174 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
175     &VNET_NAME(tcp_syncache.cache_limit), 0,
176     "Overall entry limit for syncache");
177 
178 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
179     &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
180 
181 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
182     &VNET_NAME(tcp_syncache.hashsize), 0,
183     "Size of TCP syncache hashtable");
184 
185 SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET |
186     CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0,
187     "All syncache(4) entries are visible, ignoring UID/GID, jail(2) "
188     "and mac(4) checks");
189 
190 static int
191 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
192 {
193 	int error;
194 	u_int new;
195 
196 	new = V_tcp_syncache.rexmt_limit;
197 	error = sysctl_handle_int(oidp, &new, 0, req);
198 	if ((error == 0) && (req->newptr != NULL)) {
199 		if (new > TCP_MAXRXTSHIFT)
200 			error = EINVAL;
201 		else
202 			V_tcp_syncache.rexmt_limit = new;
203 	}
204 	return (error);
205 }
206 
207 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
208     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
209     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
210     sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
211     "Limit on SYN/ACK retransmissions");
212 
213 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
214 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
215     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
216     "Send reset on socket allocation failure");
217 
218 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
219 
220 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
221 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
222 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
223 
224 /*
225  * Requires the syncache entry to be already removed from the bucket list.
226  */
227 static void
228 syncache_free(struct syncache *sc)
229 {
230 
231 	if (sc->sc_ipopts)
232 		(void) m_free(sc->sc_ipopts);
233 	if (sc->sc_cred)
234 		crfree(sc->sc_cred);
235 #ifdef MAC
236 	mac_syncache_destroy(&sc->sc_label);
237 #endif
238 
239 	uma_zfree(V_tcp_syncache.zone, sc);
240 }
241 
242 void
243 syncache_init(void)
244 {
245 	int i;
246 
247 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
248 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
249 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
250 	V_tcp_syncache.hash_secret = arc4random();
251 
252 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
253 	    &V_tcp_syncache.hashsize);
254 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
255 	    &V_tcp_syncache.bucket_limit);
256 	if (!powerof2(V_tcp_syncache.hashsize) ||
257 	    V_tcp_syncache.hashsize == 0) {
258 		printf("WARNING: syncache hash size is not a power of 2.\n");
259 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
260 	}
261 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
262 
263 	/* Set limits. */
264 	V_tcp_syncache.cache_limit =
265 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
266 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
267 	    &V_tcp_syncache.cache_limit);
268 
269 	/* Allocate the hash table. */
270 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
271 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
272 
273 #ifdef VIMAGE
274 	V_tcp_syncache.vnet = curvnet;
275 #endif
276 
277 	/* Initialize the hash buckets. */
278 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
279 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
280 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
281 			 NULL, MTX_DEF);
282 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
283 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
284 		V_tcp_syncache.hashbase[i].sch_length = 0;
285 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
286 		V_tcp_syncache.hashbase[i].sch_last_overflow =
287 		    -(SYNCOOKIE_LIFETIME + 1);
288 	}
289 
290 	/* Create the syncache entry zone. */
291 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
292 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
293 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
294 	    V_tcp_syncache.cache_limit);
295 
296 	/* Start the SYN cookie reseeder callout. */
297 	callout_init(&V_tcp_syncache.secret.reseed, 1);
298 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
299 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
300 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
301 	    syncookie_reseed, &V_tcp_syncache);
302 
303 	/* Initialize the pause machinery. */
304 	mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF);
305 	callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx,
306 	    0);
307 	V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME;
308 	V_tcp_syncache.pause_backoff = 0;
309 	V_tcp_syncache.paused = false;
310 }
311 
312 #ifdef VIMAGE
313 void
314 syncache_destroy(void)
315 {
316 	struct syncache_head *sch;
317 	struct syncache *sc, *nsc;
318 	int i;
319 
320 	/*
321 	 * Stop the re-seed timer before freeing resources.  No need to
322 	 * possibly schedule it another time.
323 	 */
324 	callout_drain(&V_tcp_syncache.secret.reseed);
325 
326 	/* Stop the SYN cache pause callout. */
327 	mtx_lock(&V_tcp_syncache.pause_mtx);
328 	if (callout_stop(&V_tcp_syncache.pause_co) == 0) {
329 		mtx_unlock(&V_tcp_syncache.pause_mtx);
330 		callout_drain(&V_tcp_syncache.pause_co);
331 	} else
332 		mtx_unlock(&V_tcp_syncache.pause_mtx);
333 
334 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
335 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
336 		sch = &V_tcp_syncache.hashbase[i];
337 		callout_drain(&sch->sch_timer);
338 
339 		SCH_LOCK(sch);
340 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
341 			syncache_drop(sc, sch);
342 		SCH_UNLOCK(sch);
343 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
344 		    ("%s: sch->sch_bucket not empty", __func__));
345 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
346 		    __func__, sch->sch_length));
347 		mtx_destroy(&sch->sch_mtx);
348 	}
349 
350 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
351 	    ("%s: cache_count not 0", __func__));
352 
353 	/* Free the allocated global resources. */
354 	uma_zdestroy(V_tcp_syncache.zone);
355 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
356 	mtx_destroy(&V_tcp_syncache.pause_mtx);
357 }
358 #endif
359 
360 /*
361  * Inserts a syncache entry into the specified bucket row.
362  * Locks and unlocks the syncache_head autonomously.
363  */
364 static void
365 syncache_insert(struct syncache *sc, struct syncache_head *sch)
366 {
367 	struct syncache *sc2;
368 
369 	SCH_LOCK(sch);
370 
371 	/*
372 	 * Make sure that we don't overflow the per-bucket limit.
373 	 * If the bucket is full, toss the oldest element.
374 	 */
375 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
376 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
377 			("sch->sch_length incorrect"));
378 		syncache_pause(&sc->sc_inc);
379 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
380 		sch->sch_last_overflow = time_uptime;
381 		syncache_drop(sc2, sch);
382 	}
383 
384 	/* Put it into the bucket. */
385 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
386 	sch->sch_length++;
387 
388 #ifdef TCP_OFFLOAD
389 	if (ADDED_BY_TOE(sc)) {
390 		struct toedev *tod = sc->sc_tod;
391 
392 		tod->tod_syncache_added(tod, sc->sc_todctx);
393 	}
394 #endif
395 
396 	/* Reinitialize the bucket row's timer. */
397 	if (sch->sch_length == 1)
398 		sch->sch_nextc = ticks + INT_MAX;
399 	syncache_timeout(sc, sch, 1);
400 
401 	SCH_UNLOCK(sch);
402 
403 	TCPSTATES_INC(TCPS_SYN_RECEIVED);
404 	TCPSTAT_INC(tcps_sc_added);
405 }
406 
407 /*
408  * Remove and free entry from syncache bucket row.
409  * Expects locked syncache head.
410  */
411 static void
412 syncache_drop(struct syncache *sc, struct syncache_head *sch)
413 {
414 
415 	SCH_LOCK_ASSERT(sch);
416 
417 	TCPSTATES_DEC(TCPS_SYN_RECEIVED);
418 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
419 	sch->sch_length--;
420 
421 #ifdef TCP_OFFLOAD
422 	if (ADDED_BY_TOE(sc)) {
423 		struct toedev *tod = sc->sc_tod;
424 
425 		tod->tod_syncache_removed(tod, sc->sc_todctx);
426 	}
427 #endif
428 
429 	syncache_free(sc);
430 }
431 
432 /*
433  * Engage/reengage time on bucket row.
434  */
435 static void
436 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
437 {
438 	int rexmt;
439 
440 	if (sc->sc_rxmits == 0)
441 		rexmt = tcp_rexmit_initial;
442 	else
443 		TCPT_RANGESET(rexmt,
444 		    tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
445 		    tcp_rexmit_min, TCPTV_REXMTMAX);
446 	sc->sc_rxttime = ticks + rexmt;
447 	sc->sc_rxmits++;
448 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
449 		sch->sch_nextc = sc->sc_rxttime;
450 		if (docallout)
451 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
452 			    syncache_timer, (void *)sch);
453 	}
454 }
455 
456 /*
457  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
458  * If we have retransmitted an entry the maximum number of times, expire it.
459  * One separate timer for each bucket row.
460  */
461 static void
462 syncache_timer(void *xsch)
463 {
464 	struct syncache_head *sch = (struct syncache_head *)xsch;
465 	struct syncache *sc, *nsc;
466 	struct epoch_tracker et;
467 	int tick = ticks;
468 	char *s;
469 	bool paused;
470 
471 	CURVNET_SET(sch->sch_sc->vnet);
472 
473 	/* NB: syncache_head has already been locked by the callout. */
474 	SCH_LOCK_ASSERT(sch);
475 
476 	/*
477 	 * In the following cycle we may remove some entries and/or
478 	 * advance some timeouts, so re-initialize the bucket timer.
479 	 */
480 	sch->sch_nextc = tick + INT_MAX;
481 
482 	/*
483 	 * If we have paused processing, unconditionally remove
484 	 * all syncache entries.
485 	 */
486 	mtx_lock(&V_tcp_syncache.pause_mtx);
487 	paused = V_tcp_syncache.paused;
488 	mtx_unlock(&V_tcp_syncache.pause_mtx);
489 
490 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
491 		if (paused) {
492 			syncache_drop(sc, sch);
493 			continue;
494 		}
495 		/*
496 		 * We do not check if the listen socket still exists
497 		 * and accept the case where the listen socket may be
498 		 * gone by the time we resend the SYN/ACK.  We do
499 		 * not expect this to happens often. If it does,
500 		 * then the RST will be sent by the time the remote
501 		 * host does the SYN/ACK->ACK.
502 		 */
503 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
504 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
505 				sch->sch_nextc = sc->sc_rxttime;
506 			continue;
507 		}
508 		if (sc->sc_rxmits > V_tcp_ecn_maxretries) {
509 			sc->sc_flags &= ~SCF_ECN_MASK;
510 		}
511 		if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
512 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
513 				log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
514 				    "giving up and removing syncache entry\n",
515 				    s, __func__);
516 				free(s, M_TCPLOG);
517 			}
518 			syncache_drop(sc, sch);
519 			TCPSTAT_INC(tcps_sc_stale);
520 			continue;
521 		}
522 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
523 			log(LOG_DEBUG, "%s; %s: Response timeout, "
524 			    "retransmitting (%u) SYN|ACK\n",
525 			    s, __func__, sc->sc_rxmits);
526 			free(s, M_TCPLOG);
527 		}
528 
529 		NET_EPOCH_ENTER(et);
530 		syncache_respond(sc, NULL, TH_SYN|TH_ACK);
531 		NET_EPOCH_EXIT(et);
532 		TCPSTAT_INC(tcps_sc_retransmitted);
533 		syncache_timeout(sc, sch, 0);
534 	}
535 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
536 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
537 			syncache_timer, (void *)(sch));
538 	CURVNET_RESTORE();
539 }
540 
541 /*
542  * Returns true if the system is only using cookies at the moment.
543  * This could be due to a sysadmin decision to only use cookies, or it
544  * could be due to the system detecting an attack.
545  */
546 static inline bool
547 syncache_cookiesonly(void)
548 {
549 
550 	return (V_tcp_syncookies && (V_tcp_syncache.paused ||
551 	    V_tcp_syncookiesonly));
552 }
553 
554 /*
555  * Find the hash bucket for the given connection.
556  */
557 static struct syncache_head *
558 syncache_hashbucket(struct in_conninfo *inc)
559 {
560 	uint32_t hash;
561 
562 	/*
563 	 * The hash is built on foreign port + local port + foreign address.
564 	 * We rely on the fact that struct in_conninfo starts with 16 bits
565 	 * of foreign port, then 16 bits of local port then followed by 128
566 	 * bits of foreign address.  In case of IPv4 address, the first 3
567 	 * 32-bit words of the address always are zeroes.
568 	 */
569 	hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
570 	    V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
571 
572 	return (&V_tcp_syncache.hashbase[hash]);
573 }
574 
575 /*
576  * Find an entry in the syncache.
577  * Returns always with locked syncache_head plus a matching entry or NULL.
578  */
579 static struct syncache *
580 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
581 {
582 	struct syncache *sc;
583 	struct syncache_head *sch;
584 
585 	*schp = sch = syncache_hashbucket(inc);
586 	SCH_LOCK(sch);
587 
588 	/* Circle through bucket row to find matching entry. */
589 	TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
590 		if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
591 		    sizeof(struct in_endpoints)) == 0)
592 			break;
593 
594 	return (sc);	/* Always returns with locked sch. */
595 }
596 
597 /*
598  * This function is called when we get a RST for a
599  * non-existent connection, so that we can see if the
600  * connection is in the syn cache.  If it is, zap it.
601  * If required send a challenge ACK.
602  */
603 void
604 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m,
605     uint16_t port)
606 {
607 	struct syncache *sc;
608 	struct syncache_head *sch;
609 	char *s = NULL;
610 
611 	if (syncache_cookiesonly())
612 		return;
613 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
614 	SCH_LOCK_ASSERT(sch);
615 
616 	/*
617 	 * No corresponding connection was found in syncache.
618 	 * If syncookies are enabled and possibly exclusively
619 	 * used, or we are under memory pressure, a valid RST
620 	 * may not find a syncache entry.  In that case we're
621 	 * done and no SYN|ACK retransmissions will happen.
622 	 * Otherwise the RST was misdirected or spoofed.
623 	 */
624 	if (sc == NULL) {
625 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
626 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
627 			    "syncache entry (possibly syncookie only), "
628 			    "segment ignored\n", s, __func__);
629 		TCPSTAT_INC(tcps_badrst);
630 		goto done;
631 	}
632 
633 	/* The remote UDP encaps port does not match. */
634 	if (sc->sc_port != port) {
635 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
636 			log(LOG_DEBUG, "%s; %s: Spurious RST with matching "
637 			    "syncache entry but non-matching UDP encaps port, "
638 			    "segment ignored\n", s, __func__);
639 		TCPSTAT_INC(tcps_badrst);
640 		goto done;
641 	}
642 
643 	/*
644 	 * If the RST bit is set, check the sequence number to see
645 	 * if this is a valid reset segment.
646 	 *
647 	 * RFC 793 page 37:
648 	 *   In all states except SYN-SENT, all reset (RST) segments
649 	 *   are validated by checking their SEQ-fields.  A reset is
650 	 *   valid if its sequence number is in the window.
651 	 *
652 	 * RFC 793 page 69:
653 	 *   There are four cases for the acceptability test for an incoming
654 	 *   segment:
655 	 *
656 	 * Segment Receive  Test
657 	 * Length  Window
658 	 * ------- -------  -------------------------------------------
659 	 *    0       0     SEG.SEQ = RCV.NXT
660 	 *    0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
661 	 *   >0       0     not acceptable
662 	 *   >0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
663 	 *               or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
664 	 *
665 	 * Note that when receiving a SYN segment in the LISTEN state,
666 	 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
667 	 * described in RFC 793, page 66.
668 	 */
669 	if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
670 	    SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
671 	    (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
672 		if (V_tcp_insecure_rst ||
673 		    th->th_seq == sc->sc_irs + 1) {
674 			syncache_drop(sc, sch);
675 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
676 				log(LOG_DEBUG,
677 				    "%s; %s: Our SYN|ACK was rejected, "
678 				    "connection attempt aborted by remote "
679 				    "endpoint\n",
680 				    s, __func__);
681 			TCPSTAT_INC(tcps_sc_reset);
682 		} else {
683 			TCPSTAT_INC(tcps_badrst);
684 			/* Send challenge ACK. */
685 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
686 				log(LOG_DEBUG, "%s; %s: RST with invalid "
687 				    " SEQ %u != NXT %u (+WND %u), "
688 				    "sending challenge ACK\n",
689 				    s, __func__,
690 				    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
691 			syncache_respond(sc, m, TH_ACK);
692 		}
693 	} else {
694 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
695 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
696 			    "NXT %u (+WND %u), segment ignored\n",
697 			    s, __func__,
698 			    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
699 		TCPSTAT_INC(tcps_badrst);
700 	}
701 
702 done:
703 	if (s != NULL)
704 		free(s, M_TCPLOG);
705 	SCH_UNLOCK(sch);
706 }
707 
708 void
709 syncache_badack(struct in_conninfo *inc, uint16_t port)
710 {
711 	struct syncache *sc;
712 	struct syncache_head *sch;
713 
714 	if (syncache_cookiesonly())
715 		return;
716 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
717 	SCH_LOCK_ASSERT(sch);
718 	if ((sc != NULL) && (sc->sc_port == port)) {
719 		syncache_drop(sc, sch);
720 		TCPSTAT_INC(tcps_sc_badack);
721 	}
722 	SCH_UNLOCK(sch);
723 }
724 
725 void
726 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port)
727 {
728 	struct syncache *sc;
729 	struct syncache_head *sch;
730 
731 	if (syncache_cookiesonly())
732 		return;
733 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
734 	SCH_LOCK_ASSERT(sch);
735 	if (sc == NULL)
736 		goto done;
737 
738 	/* If the port != sc_port, then it's a bogus ICMP msg */
739 	if (port != sc->sc_port)
740 		goto done;
741 
742 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
743 	if (ntohl(th_seq) != sc->sc_iss)
744 		goto done;
745 
746 	/*
747 	 * If we've rertransmitted 3 times and this is our second error,
748 	 * we remove the entry.  Otherwise, we allow it to continue on.
749 	 * This prevents us from incorrectly nuking an entry during a
750 	 * spurious network outage.
751 	 *
752 	 * See tcp_notify().
753 	 */
754 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
755 		sc->sc_flags |= SCF_UNREACH;
756 		goto done;
757 	}
758 	syncache_drop(sc, sch);
759 	TCPSTAT_INC(tcps_sc_unreach);
760 done:
761 	SCH_UNLOCK(sch);
762 }
763 
764 /*
765  * Build a new TCP socket structure from a syncache entry.
766  *
767  * On success return the newly created socket with its underlying inp locked.
768  */
769 static struct socket *
770 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
771 {
772 	struct inpcb *inp = NULL;
773 	struct socket *so;
774 	struct tcpcb *tp;
775 	int error;
776 	char *s;
777 
778 	NET_EPOCH_ASSERT();
779 
780 	/*
781 	 * Ok, create the full blown connection, and set things up
782 	 * as they would have been set up if we had created the
783 	 * connection when the SYN arrived.
784 	 */
785 	if ((so = solisten_clone(lso)) == NULL)
786 		goto allocfail;
787 #ifdef MAC
788 	mac_socketpeer_set_from_mbuf(m, so);
789 #endif
790 	error = in_pcballoc(so, &V_tcbinfo);
791 	if (error) {
792 		sodealloc(so);
793 		goto allocfail;
794 	}
795 	inp = sotoinpcb(so);
796 	if ((tp = tcp_newtcpcb(inp, sototcpcb(lso))) == NULL) {
797 		in_pcbfree(inp);
798 		sodealloc(so);
799 		goto allocfail;
800 	}
801 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
802 #ifdef INET6
803 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
804 		inp->inp_vflag &= ~INP_IPV4;
805 		inp->inp_vflag |= INP_IPV6;
806 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
807 	} else {
808 		inp->inp_vflag &= ~INP_IPV6;
809 		inp->inp_vflag |= INP_IPV4;
810 #endif
811 		inp->inp_ip_ttl = sc->sc_ip_ttl;
812 		inp->inp_ip_tos = sc->sc_ip_tos;
813 		inp->inp_laddr = sc->sc_inc.inc_laddr;
814 #ifdef INET6
815 	}
816 #endif
817 
818 	/*
819 	 * If there's an mbuf and it has a flowid, then let's initialise the
820 	 * inp with that particular flowid.
821 	 */
822 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
823 		inp->inp_flowid = m->m_pkthdr.flowid;
824 		inp->inp_flowtype = M_HASHTYPE_GET(m);
825 #ifdef NUMA
826 		inp->inp_numa_domain = m->m_pkthdr.numa_domain;
827 #endif
828 	}
829 
830 	inp->inp_lport = sc->sc_inc.inc_lport;
831 #ifdef INET6
832 	if (inp->inp_vflag & INP_IPV6PROTO) {
833 		struct inpcb *oinp = sotoinpcb(lso);
834 
835 		/*
836 		 * Inherit socket options from the listening socket.
837 		 * Note that in6p_inputopts are not (and should not be)
838 		 * copied, since it stores previously received options and is
839 		 * used to detect if each new option is different than the
840 		 * previous one and hence should be passed to a user.
841 		 * If we copied in6p_inputopts, a user would not be able to
842 		 * receive options just after calling the accept system call.
843 		 */
844 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
845 		if (oinp->in6p_outputopts)
846 			inp->in6p_outputopts =
847 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
848 		inp->in6p_hops = oinp->in6p_hops;
849 	}
850 
851 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
852 		struct sockaddr_in6 sin6;
853 
854 		sin6.sin6_family = AF_INET6;
855 		sin6.sin6_len = sizeof(sin6);
856 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
857 		sin6.sin6_port = sc->sc_inc.inc_fport;
858 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
859 		INP_HASH_WLOCK(&V_tcbinfo);
860 		error = in6_pcbconnect(inp, &sin6, thread0.td_ucred, false);
861 		INP_HASH_WUNLOCK(&V_tcbinfo);
862 		if (error != 0)
863 			goto abort;
864 		/* Override flowlabel from in6_pcbconnect. */
865 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
866 		inp->inp_flow |= sc->sc_flowlabel;
867 	}
868 #endif /* INET6 */
869 #if defined(INET) && defined(INET6)
870 	else
871 #endif
872 #ifdef INET
873 	{
874 		struct sockaddr_in sin;
875 
876 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
877 
878 		if (inp->inp_options == NULL) {
879 			inp->inp_options = sc->sc_ipopts;
880 			sc->sc_ipopts = NULL;
881 		}
882 
883 		sin.sin_family = AF_INET;
884 		sin.sin_len = sizeof(sin);
885 		sin.sin_addr = sc->sc_inc.inc_faddr;
886 		sin.sin_port = sc->sc_inc.inc_fport;
887 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
888 		INP_HASH_WLOCK(&V_tcbinfo);
889 		error = in_pcbconnect(inp, &sin, thread0.td_ucred, false);
890 		INP_HASH_WUNLOCK(&V_tcbinfo);
891 		if (error != 0)
892 			goto abort;
893 	}
894 #endif /* INET */
895 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
896 	/* Copy old policy into new socket's. */
897 	if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
898 		printf("syncache_socket: could not copy policy\n");
899 #endif
900 	tp->t_state = TCPS_SYN_RECEIVED;
901 	tp->iss = sc->sc_iss;
902 	tp->irs = sc->sc_irs;
903 	tp->t_port = sc->sc_port;
904 	tcp_rcvseqinit(tp);
905 	tcp_sendseqinit(tp);
906 	tp->snd_wl1 = sc->sc_irs;
907 	tp->snd_max = tp->iss + 1;
908 	tp->snd_nxt = tp->iss + 1;
909 	tp->rcv_up = sc->sc_irs + 1;
910 	tp->rcv_wnd = sc->sc_wnd;
911 	tp->rcv_adv += tp->rcv_wnd;
912 	tp->last_ack_sent = tp->rcv_nxt;
913 
914 	tp->t_flags = sototcpcb(lso)->t_flags &
915 	    (TF_LRD|TF_NOPUSH|TF_NODELAY);
916 	if (sc->sc_flags & SCF_NOOPT)
917 		tp->t_flags |= TF_NOOPT;
918 	else {
919 		if (sc->sc_flags & SCF_WINSCALE) {
920 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
921 			tp->snd_scale = sc->sc_requested_s_scale;
922 			tp->request_r_scale = sc->sc_requested_r_scale;
923 		}
924 		if (sc->sc_flags & SCF_TIMESTAMP) {
925 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
926 			tp->ts_recent = sc->sc_tsreflect;
927 			tp->ts_recent_age = tcp_ts_getticks();
928 			tp->ts_offset = sc->sc_tsoff;
929 		}
930 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
931 		if (sc->sc_flags & SCF_SIGNATURE)
932 			tp->t_flags |= TF_SIGNATURE;
933 #endif
934 		if (sc->sc_flags & SCF_SACK)
935 			tp->t_flags |= TF_SACK_PERMIT;
936 	}
937 
938 	tcp_ecn_syncache_socket(tp, sc);
939 
940 	/*
941 	 * Set up MSS and get cached values from tcp_hostcache.
942 	 * This might overwrite some of the defaults we just set.
943 	 */
944 	tcp_mss(tp, sc->sc_peer_mss);
945 
946 	/*
947 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
948 	 * limited to one segment in cc_conn_init().
949 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
950 	 */
951 	if (sc->sc_rxmits > 1)
952 		tp->snd_cwnd = 1;
953 
954 #ifdef TCP_OFFLOAD
955 	/*
956 	 * Allow a TOE driver to install its hooks.  Note that we hold the
957 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
958 	 * new connection before the TOE driver has done its thing.
959 	 */
960 	if (ADDED_BY_TOE(sc)) {
961 		struct toedev *tod = sc->sc_tod;
962 
963 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
964 	}
965 #endif
966 #ifdef TCP_BLACKBOX
967 	/*
968 	 * Inherit the log state from the listening socket, if
969 	 * - the log state of the listening socket is not off and
970 	 * - the listening socket was not auto selected from all sessions and
971 	 * - a log id is not set on the listening socket.
972 	 * This avoids inheriting a log state which was automatically set.
973 	 */
974 	if ((tcp_get_bblog_state(sototcpcb(lso)) != TCP_LOG_STATE_OFF) &&
975 	    ((sototcpcb(lso)->t_flags2 & TF2_LOG_AUTO) == 0) &&
976 	    (sototcpcb(lso)->t_lib == NULL)) {
977 		tcp_log_state_change(tp, tcp_get_bblog_state(sototcpcb(lso)));
978 	}
979 #endif
980 	/*
981 	 * Copy and activate timers.
982 	 */
983 	tp->t_maxunacktime = sototcpcb(lso)->t_maxunacktime;
984 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
985 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
986 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
987 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
988 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
989 
990 	TCPSTAT_INC(tcps_accepts);
991 	TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, TCPS_LISTEN);
992 
993 	if (!solisten_enqueue(so, SS_ISCONNECTED))
994 		tp->t_flags |= TF_SONOTCONN;
995 	/* Can we inherit anything from the listener? */
996 	if (tp->t_fb->tfb_inherit != NULL) {
997 		(*tp->t_fb->tfb_inherit)(tp, sotoinpcb(lso));
998 	}
999 	return (so);
1000 
1001 allocfail:
1002 	/*
1003 	 * Drop the connection; we will either send a RST or have the peer
1004 	 * retransmit its SYN again after its RTO and try again.
1005 	 */
1006 	if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1007 		log(LOG_DEBUG, "%s; %s: Socket create failed "
1008 		    "due to limits or memory shortage\n",
1009 		    s, __func__);
1010 		free(s, M_TCPLOG);
1011 	}
1012 	TCPSTAT_INC(tcps_listendrop);
1013 	return (NULL);
1014 
1015 abort:
1016 	tcp_discardcb(tp);
1017 	in_pcbfree(inp);
1018 	sodealloc(so);
1019 	if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1020 		log(LOG_DEBUG, "%s; %s: in%s_pcbconnect failed with error %i\n",
1021 		    s, __func__, (sc->sc_inc.inc_flags & INC_ISIPV6) ? "6" : "",
1022 		    error);
1023 		free(s, M_TCPLOG);
1024 	}
1025 	TCPSTAT_INC(tcps_listendrop);
1026 	return (NULL);
1027 }
1028 
1029 /*
1030  * This function gets called when we receive an ACK for a
1031  * socket in the LISTEN state.  We look up the connection
1032  * in the syncache, and if its there, we pull it out of
1033  * the cache and turn it into a full-blown connection in
1034  * the SYN-RECEIVED state.
1035  *
1036  * On syncache_socket() success the newly created socket
1037  * has its underlying inp locked.
1038  */
1039 int
1040 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1041     struct socket **lsop, struct mbuf *m, uint16_t port)
1042 {
1043 	struct syncache *sc;
1044 	struct syncache_head *sch;
1045 	struct syncache scs;
1046 	char *s;
1047 	bool locked;
1048 
1049 	NET_EPOCH_ASSERT();
1050 	KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1051 	    ("%s: can handle only ACK", __func__));
1052 
1053 	if (syncache_cookiesonly()) {
1054 		sc = NULL;
1055 		sch = syncache_hashbucket(inc);
1056 		locked = false;
1057 	} else {
1058 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1059 		locked = true;
1060 		SCH_LOCK_ASSERT(sch);
1061 	}
1062 
1063 #ifdef INVARIANTS
1064 	/*
1065 	 * Test code for syncookies comparing the syncache stored
1066 	 * values with the reconstructed values from the cookie.
1067 	 */
1068 	if (sc != NULL)
1069 		syncookie_cmp(inc, sch, sc, th, to, *lsop, port);
1070 #endif
1071 
1072 	if (sc == NULL) {
1073 		/*
1074 		 * There is no syncache entry, so see if this ACK is
1075 		 * a returning syncookie.  To do this, first:
1076 		 *  A. Check if syncookies are used in case of syncache
1077 		 *     overflows
1078 		 *  B. See if this socket has had a syncache entry dropped in
1079 		 *     the recent past. We don't want to accept a bogus
1080 		 *     syncookie if we've never received a SYN or accept it
1081 		 *     twice.
1082 		 *  C. check that the syncookie is valid.  If it is, then
1083 		 *     cobble up a fake syncache entry, and return.
1084 		 */
1085 		if (locked && !V_tcp_syncookies) {
1086 			SCH_UNLOCK(sch);
1087 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1088 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1089 				    "segment rejected (syncookies disabled)\n",
1090 				    s, __func__);
1091 			goto failed;
1092 		}
1093 		if (locked && !V_tcp_syncookiesonly &&
1094 		    sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1095 			SCH_UNLOCK(sch);
1096 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1097 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1098 				    "segment rejected (no syncache entry)\n",
1099 				    s, __func__);
1100 			goto failed;
1101 		}
1102 		bzero(&scs, sizeof(scs));
1103 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port);
1104 		if (locked)
1105 			SCH_UNLOCK(sch);
1106 		if (sc == NULL) {
1107 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1108 				log(LOG_DEBUG, "%s; %s: Segment failed "
1109 				    "SYNCOOKIE authentication, segment rejected "
1110 				    "(probably spoofed)\n", s, __func__);
1111 			goto failed;
1112 		}
1113 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1114 		/* If received ACK has MD5 signature, check it. */
1115 		if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1116 		    (!TCPMD5_ENABLED() ||
1117 		    TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1118 			/* Drop the ACK. */
1119 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1120 				log(LOG_DEBUG, "%s; %s: Segment rejected, "
1121 				    "MD5 signature doesn't match.\n",
1122 				    s, __func__);
1123 				free(s, M_TCPLOG);
1124 			}
1125 			TCPSTAT_INC(tcps_sig_err_sigopt);
1126 			return (-1); /* Do not send RST */
1127 		}
1128 #endif /* TCP_SIGNATURE */
1129 		TCPSTATES_INC(TCPS_SYN_RECEIVED);
1130 	} else {
1131 		if (sc->sc_port != port) {
1132 			SCH_UNLOCK(sch);
1133 			return (0);
1134 		}
1135 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1136 		/*
1137 		 * If listening socket requested TCP digests, check that
1138 		 * received ACK has signature and it is correct.
1139 		 * If not, drop the ACK and leave sc entry in th cache,
1140 		 * because SYN was received with correct signature.
1141 		 */
1142 		if (sc->sc_flags & SCF_SIGNATURE) {
1143 			if ((to->to_flags & TOF_SIGNATURE) == 0) {
1144 				/* No signature */
1145 				TCPSTAT_INC(tcps_sig_err_nosigopt);
1146 				SCH_UNLOCK(sch);
1147 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1148 					log(LOG_DEBUG, "%s; %s: Segment "
1149 					    "rejected, MD5 signature wasn't "
1150 					    "provided.\n", s, __func__);
1151 					free(s, M_TCPLOG);
1152 				}
1153 				return (-1); /* Do not send RST */
1154 			}
1155 			if (!TCPMD5_ENABLED() ||
1156 			    TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1157 				/* Doesn't match or no SA */
1158 				SCH_UNLOCK(sch);
1159 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1160 					log(LOG_DEBUG, "%s; %s: Segment "
1161 					    "rejected, MD5 signature doesn't "
1162 					    "match.\n", s, __func__);
1163 					free(s, M_TCPLOG);
1164 				}
1165 				return (-1); /* Do not send RST */
1166 			}
1167 		}
1168 #endif /* TCP_SIGNATURE */
1169 
1170 		/*
1171 		 * RFC 7323 PAWS: If we have a timestamp on this segment and
1172 		 * it's less than ts_recent, drop it.
1173 		 * XXXMT: RFC 7323 also requires to send an ACK.
1174 		 *        In tcp_input.c this is only done for TCP segments
1175 		 *        with user data, so be consistent here and just drop
1176 		 *        the segment.
1177 		 */
1178 		if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1179 		    TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1180 			SCH_UNLOCK(sch);
1181 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1182 				log(LOG_DEBUG,
1183 				    "%s; %s: SEG.TSval %u < TS.Recent %u, "
1184 				    "segment dropped\n", s, __func__,
1185 				    to->to_tsval, sc->sc_tsreflect);
1186 				free(s, M_TCPLOG);
1187 			}
1188 			return (-1);  /* Do not send RST */
1189 		}
1190 
1191 		/*
1192 		 * If timestamps were not negotiated during SYN/ACK and a
1193 		 * segment with a timestamp is received, ignore the
1194 		 * timestamp and process the packet normally.
1195 		 * See section 3.2 of RFC 7323.
1196 		 */
1197 		if (!(sc->sc_flags & SCF_TIMESTAMP) &&
1198 		    (to->to_flags & TOF_TS)) {
1199 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1200 				log(LOG_DEBUG, "%s; %s: Timestamp not "
1201 				    "expected, segment processed normally\n",
1202 				    s, __func__);
1203 				free(s, M_TCPLOG);
1204 				s = NULL;
1205 			}
1206 		}
1207 
1208 		/*
1209 		 * If timestamps were negotiated during SYN/ACK and a
1210 		 * segment without a timestamp is received, silently drop
1211 		 * the segment, unless the missing timestamps are tolerated.
1212 		 * See section 3.2 of RFC 7323.
1213 		 */
1214 		if ((sc->sc_flags & SCF_TIMESTAMP) &&
1215 		    !(to->to_flags & TOF_TS)) {
1216 			if (V_tcp_tolerate_missing_ts) {
1217 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1218 					log(LOG_DEBUG,
1219 					    "%s; %s: Timestamp missing, "
1220 					    "segment processed normally\n",
1221 					    s, __func__);
1222 					free(s, M_TCPLOG);
1223 				}
1224 			} else {
1225 				SCH_UNLOCK(sch);
1226 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1227 					log(LOG_DEBUG,
1228 					    "%s; %s: Timestamp missing, "
1229 					    "segment silently dropped\n",
1230 					    s, __func__);
1231 					free(s, M_TCPLOG);
1232 				}
1233 				return (-1);  /* Do not send RST */
1234 			}
1235 		}
1236 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1237 		sch->sch_length--;
1238 #ifdef TCP_OFFLOAD
1239 		if (ADDED_BY_TOE(sc)) {
1240 			struct toedev *tod = sc->sc_tod;
1241 
1242 			tod->tod_syncache_removed(tod, sc->sc_todctx);
1243 		}
1244 #endif
1245 		SCH_UNLOCK(sch);
1246 	}
1247 
1248 	/*
1249 	 * Segment validation:
1250 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1251 	 */
1252 	if (th->th_ack != sc->sc_iss + 1) {
1253 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1254 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1255 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1256 		goto failed;
1257 	}
1258 
1259 	/*
1260 	 * The SEQ must fall in the window starting at the received
1261 	 * initial receive sequence number + 1 (the SYN).
1262 	 */
1263 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1264 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1265 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1266 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1267 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1268 		goto failed;
1269 	}
1270 
1271 	*lsop = syncache_socket(sc, *lsop, m);
1272 
1273 	if (__predict_false(*lsop == NULL)) {
1274 		TCPSTAT_INC(tcps_sc_aborted);
1275 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1276 	} else
1277 		TCPSTAT_INC(tcps_sc_completed);
1278 
1279 /* how do we find the inp for the new socket? */
1280 	if (sc != &scs)
1281 		syncache_free(sc);
1282 	return (1);
1283 failed:
1284 	if (sc != NULL) {
1285 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1286 		if (sc != &scs)
1287 			syncache_free(sc);
1288 	}
1289 	if (s != NULL)
1290 		free(s, M_TCPLOG);
1291 	*lsop = NULL;
1292 	return (0);
1293 }
1294 
1295 static struct socket *
1296 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m,
1297     uint64_t response_cookie)
1298 {
1299 	struct inpcb *inp;
1300 	struct tcpcb *tp;
1301 	unsigned int *pending_counter;
1302 	struct socket *so;
1303 
1304 	NET_EPOCH_ASSERT();
1305 
1306 	pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending;
1307 	so = syncache_socket(sc, lso, m);
1308 	if (so == NULL) {
1309 		TCPSTAT_INC(tcps_sc_aborted);
1310 		atomic_subtract_int(pending_counter, 1);
1311 	} else {
1312 		soisconnected(so);
1313 		inp = sotoinpcb(so);
1314 		tp = intotcpcb(inp);
1315 		tp->t_flags |= TF_FASTOPEN;
1316 		tp->t_tfo_cookie.server = response_cookie;
1317 		tp->snd_max = tp->iss;
1318 		tp->snd_nxt = tp->iss;
1319 		tp->t_tfo_pending = pending_counter;
1320 		TCPSTATES_INC(TCPS_SYN_RECEIVED);
1321 		TCPSTAT_INC(tcps_sc_completed);
1322 	}
1323 
1324 	return (so);
1325 }
1326 
1327 /*
1328  * Given a LISTEN socket and an inbound SYN request, add
1329  * this to the syn cache, and send back a segment:
1330  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1331  * to the source.
1332  *
1333  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1334  * Doing so would require that we hold onto the data and deliver it
1335  * to the application.  However, if we are the target of a SYN-flood
1336  * DoS attack, an attacker could send data which would eventually
1337  * consume all available buffer space if it were ACKed.  By not ACKing
1338  * the data, we avoid this DoS scenario.
1339  *
1340  * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1341  * cookie is processed and a new socket is created.  In this case, any data
1342  * accompanying the SYN will be queued to the socket by tcp_input() and will
1343  * be ACKed either when the application sends response data or the delayed
1344  * ACK timer expires, whichever comes first.
1345  */
1346 struct socket *
1347 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1348     struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod,
1349     void *todctx, uint8_t iptos, uint16_t port)
1350 {
1351 	struct tcpcb *tp;
1352 	struct socket *rv = NULL;
1353 	struct syncache *sc = NULL;
1354 	struct syncache_head *sch;
1355 	struct mbuf *ipopts = NULL;
1356 	u_int ltflags;
1357 	int win, ip_ttl, ip_tos;
1358 	char *s;
1359 #ifdef INET6
1360 	int autoflowlabel = 0;
1361 #endif
1362 #ifdef MAC
1363 	struct label *maclabel;
1364 #endif
1365 	struct syncache scs;
1366 	struct ucred *cred;
1367 	uint64_t tfo_response_cookie;
1368 	unsigned int *tfo_pending = NULL;
1369 	int tfo_cookie_valid = 0;
1370 	int tfo_response_cookie_valid = 0;
1371 	bool locked;
1372 
1373 	INP_RLOCK_ASSERT(inp);			/* listen socket */
1374 	KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1375 	    ("%s: unexpected tcp flags", __func__));
1376 
1377 	/*
1378 	 * Combine all so/tp operations very early to drop the INP lock as
1379 	 * soon as possible.
1380 	 */
1381 	KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1382 	tp = sototcpcb(so);
1383 	cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred);
1384 
1385 #ifdef INET6
1386 	if (inc->inc_flags & INC_ISIPV6) {
1387 		if (inp->inp_flags & IN6P_AUTOFLOWLABEL) {
1388 			autoflowlabel = 1;
1389 		}
1390 		ip_ttl = in6_selecthlim(inp, NULL);
1391 		if ((inp->in6p_outputopts == NULL) ||
1392 		    (inp->in6p_outputopts->ip6po_tclass == -1)) {
1393 			ip_tos = 0;
1394 		} else {
1395 			ip_tos = inp->in6p_outputopts->ip6po_tclass;
1396 		}
1397 	}
1398 #endif
1399 #if defined(INET6) && defined(INET)
1400 	else
1401 #endif
1402 #ifdef INET
1403 	{
1404 		ip_ttl = inp->inp_ip_ttl;
1405 		ip_tos = inp->inp_ip_tos;
1406 	}
1407 #endif
1408 	win = so->sol_sbrcv_hiwat;
1409 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1410 
1411 	if (V_tcp_fastopen_server_enable && (tp->t_flags & TF_FASTOPEN) &&
1412 	    (tp->t_tfo_pending != NULL) &&
1413 	    (to->to_flags & TOF_FASTOPEN)) {
1414 		/*
1415 		 * Limit the number of pending TFO connections to
1416 		 * approximately half of the queue limit.  This prevents TFO
1417 		 * SYN floods from starving the service by filling the
1418 		 * listen queue with bogus TFO connections.
1419 		 */
1420 		if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1421 		    (so->sol_qlimit / 2)) {
1422 			int result;
1423 
1424 			result = tcp_fastopen_check_cookie(inc,
1425 			    to->to_tfo_cookie, to->to_tfo_len,
1426 			    &tfo_response_cookie);
1427 			tfo_cookie_valid = (result > 0);
1428 			tfo_response_cookie_valid = (result >= 0);
1429 		}
1430 
1431 		/*
1432 		 * Remember the TFO pending counter as it will have to be
1433 		 * decremented below if we don't make it to syncache_tfo_expand().
1434 		 */
1435 		tfo_pending = tp->t_tfo_pending;
1436 	}
1437 
1438 #ifdef MAC
1439 	if (mac_syncache_init(&maclabel) != 0) {
1440 		INP_RUNLOCK(inp);
1441 		goto done;
1442 	} else
1443 		mac_syncache_create(maclabel, inp);
1444 #endif
1445 	if (!tfo_cookie_valid)
1446 		INP_RUNLOCK(inp);
1447 
1448 	/*
1449 	 * Remember the IP options, if any.
1450 	 */
1451 #ifdef INET6
1452 	if (!(inc->inc_flags & INC_ISIPV6))
1453 #endif
1454 #ifdef INET
1455 		ipopts = (m) ? ip_srcroute(m) : NULL;
1456 #else
1457 		ipopts = NULL;
1458 #endif
1459 
1460 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1461 	/*
1462 	 * When the socket is TCP-MD5 enabled check that,
1463 	 *  - a signed packet is valid
1464 	 *  - a non-signed packet does not have a security association
1465 	 *
1466 	 *  If a signed packet fails validation or a non-signed packet has a
1467 	 *  security association, the packet will be dropped.
1468 	 */
1469 	if (ltflags & TF_SIGNATURE) {
1470 		if (to->to_flags & TOF_SIGNATURE) {
1471 			if (!TCPMD5_ENABLED() ||
1472 			    TCPMD5_INPUT(m, th, to->to_signature) != 0)
1473 				goto done;
1474 		} else {
1475 			if (TCPMD5_ENABLED() &&
1476 			    TCPMD5_INPUT(m, NULL, NULL) != ENOENT)
1477 				goto done;
1478 		}
1479 	} else if (to->to_flags & TOF_SIGNATURE)
1480 		goto done;
1481 #endif	/* TCP_SIGNATURE */
1482 	/*
1483 	 * See if we already have an entry for this connection.
1484 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1485 	 *
1486 	 * XXX: should the syncache be re-initialized with the contents
1487 	 * of the new SYN here (which may have different options?)
1488 	 *
1489 	 * XXX: We do not check the sequence number to see if this is a
1490 	 * real retransmit or a new connection attempt.  The question is
1491 	 * how to handle such a case; either ignore it as spoofed, or
1492 	 * drop the current entry and create a new one?
1493 	 */
1494 	if (syncache_cookiesonly()) {
1495 		sc = NULL;
1496 		sch = syncache_hashbucket(inc);
1497 		locked = false;
1498 	} else {
1499 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1500 		locked = true;
1501 		SCH_LOCK_ASSERT(sch);
1502 	}
1503 	if (sc != NULL) {
1504 		if (tfo_cookie_valid)
1505 			INP_RUNLOCK(inp);
1506 		TCPSTAT_INC(tcps_sc_dupsyn);
1507 		if (ipopts) {
1508 			/*
1509 			 * If we were remembering a previous source route,
1510 			 * forget it and use the new one we've been given.
1511 			 */
1512 			if (sc->sc_ipopts)
1513 				(void) m_free(sc->sc_ipopts);
1514 			sc->sc_ipopts = ipopts;
1515 		}
1516 		/*
1517 		 * Update timestamp if present.
1518 		 */
1519 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1520 			sc->sc_tsreflect = to->to_tsval;
1521 		else
1522 			sc->sc_flags &= ~SCF_TIMESTAMP;
1523 		/*
1524 		 * Adjust ECN response if needed, e.g. different
1525 		 * IP ECN field, or a fallback by the remote host.
1526 		 */
1527 		if (sc->sc_flags & SCF_ECN_MASK) {
1528 			sc->sc_flags &= ~SCF_ECN_MASK;
1529 			sc->sc_flags = tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1530 		}
1531 #ifdef MAC
1532 		/*
1533 		 * Since we have already unconditionally allocated label
1534 		 * storage, free it up.  The syncache entry will already
1535 		 * have an initialized label we can use.
1536 		 */
1537 		mac_syncache_destroy(&maclabel);
1538 #endif
1539 		TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1540 		/* Retransmit SYN|ACK and reset retransmit count. */
1541 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1542 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1543 			    "resetting timer and retransmitting SYN|ACK\n",
1544 			    s, __func__);
1545 			free(s, M_TCPLOG);
1546 		}
1547 		if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1548 			sc->sc_rxmits = 0;
1549 			syncache_timeout(sc, sch, 1);
1550 			TCPSTAT_INC(tcps_sndacks);
1551 			TCPSTAT_INC(tcps_sndtotal);
1552 		}
1553 		SCH_UNLOCK(sch);
1554 		goto donenoprobe;
1555 	}
1556 
1557 	if (tfo_cookie_valid) {
1558 		bzero(&scs, sizeof(scs));
1559 		sc = &scs;
1560 		goto skip_alloc;
1561 	}
1562 
1563 	/*
1564 	 * Skip allocating a syncache entry if we are just going to discard
1565 	 * it later.
1566 	 */
1567 	if (!locked) {
1568 		bzero(&scs, sizeof(scs));
1569 		sc = &scs;
1570 	} else
1571 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1572 	if (sc == NULL) {
1573 		/*
1574 		 * The zone allocator couldn't provide more entries.
1575 		 * Treat this as if the cache was full; drop the oldest
1576 		 * entry and insert the new one.
1577 		 */
1578 		TCPSTAT_INC(tcps_sc_zonefail);
1579 		if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1580 			sch->sch_last_overflow = time_uptime;
1581 			syncache_drop(sc, sch);
1582 			syncache_pause(inc);
1583 		}
1584 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1585 		if (sc == NULL) {
1586 			if (V_tcp_syncookies) {
1587 				bzero(&scs, sizeof(scs));
1588 				sc = &scs;
1589 			} else {
1590 				KASSERT(locked,
1591 				    ("%s: bucket unexpectedly unlocked",
1592 				    __func__));
1593 				SCH_UNLOCK(sch);
1594 				if (ipopts)
1595 					(void) m_free(ipopts);
1596 				goto done;
1597 			}
1598 		}
1599 	}
1600 
1601 skip_alloc:
1602 	if (!tfo_cookie_valid && tfo_response_cookie_valid)
1603 		sc->sc_tfo_cookie = &tfo_response_cookie;
1604 
1605 	/*
1606 	 * Fill in the syncache values.
1607 	 */
1608 #ifdef MAC
1609 	sc->sc_label = maclabel;
1610 #endif
1611 	sc->sc_cred = cred;
1612 	sc->sc_port = port;
1613 	cred = NULL;
1614 	sc->sc_ipopts = ipopts;
1615 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1616 	sc->sc_ip_tos = ip_tos;
1617 	sc->sc_ip_ttl = ip_ttl;
1618 #ifdef TCP_OFFLOAD
1619 	sc->sc_tod = tod;
1620 	sc->sc_todctx = todctx;
1621 #endif
1622 	sc->sc_irs = th->th_seq;
1623 	sc->sc_flags = 0;
1624 	sc->sc_flowlabel = 0;
1625 
1626 	/*
1627 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1628 	 * win was derived from socket earlier in the function.
1629 	 */
1630 	win = imax(win, 0);
1631 	win = imin(win, TCP_MAXWIN);
1632 	sc->sc_wnd = win;
1633 
1634 	if (V_tcp_do_rfc1323 &&
1635 	    !(ltflags & TF_NOOPT)) {
1636 		/*
1637 		 * A timestamp received in a SYN makes
1638 		 * it ok to send timestamp requests and replies.
1639 		 */
1640 		if ((to->to_flags & TOF_TS) && (V_tcp_do_rfc1323 != 2)) {
1641 			sc->sc_tsreflect = to->to_tsval;
1642 			sc->sc_flags |= SCF_TIMESTAMP;
1643 			sc->sc_tsoff = tcp_new_ts_offset(inc);
1644 		}
1645 		if ((to->to_flags & TOF_SCALE) && (V_tcp_do_rfc1323 != 3)) {
1646 			int wscale = 0;
1647 
1648 			/*
1649 			 * Pick the smallest possible scaling factor that
1650 			 * will still allow us to scale up to sb_max, aka
1651 			 * kern.ipc.maxsockbuf.
1652 			 *
1653 			 * We do this because there are broken firewalls that
1654 			 * will corrupt the window scale option, leading to
1655 			 * the other endpoint believing that our advertised
1656 			 * window is unscaled.  At scale factors larger than
1657 			 * 5 the unscaled window will drop below 1500 bytes,
1658 			 * leading to serious problems when traversing these
1659 			 * broken firewalls.
1660 			 *
1661 			 * With the default maxsockbuf of 256K, a scale factor
1662 			 * of 3 will be chosen by this algorithm.  Those who
1663 			 * choose a larger maxsockbuf should watch out
1664 			 * for the compatibility problems mentioned above.
1665 			 *
1666 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1667 			 * or <SYN,ACK>) segment itself is never scaled.
1668 			 */
1669 			while (wscale < TCP_MAX_WINSHIFT &&
1670 			    (TCP_MAXWIN << wscale) < sb_max)
1671 				wscale++;
1672 			sc->sc_requested_r_scale = wscale;
1673 			sc->sc_requested_s_scale = to->to_wscale;
1674 			sc->sc_flags |= SCF_WINSCALE;
1675 		}
1676 	}
1677 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1678 	/*
1679 	 * If incoming packet has an MD5 signature, flag this in the
1680 	 * syncache so that syncache_respond() will do the right thing
1681 	 * with the SYN+ACK.
1682 	 */
1683 	if (to->to_flags & TOF_SIGNATURE)
1684 		sc->sc_flags |= SCF_SIGNATURE;
1685 #endif	/* TCP_SIGNATURE */
1686 	if (to->to_flags & TOF_SACKPERM)
1687 		sc->sc_flags |= SCF_SACK;
1688 	if (to->to_flags & TOF_MSS)
1689 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1690 	if (ltflags & TF_NOOPT)
1691 		sc->sc_flags |= SCF_NOOPT;
1692 	/* ECN Handshake */
1693 	if (V_tcp_do_ecn && (tp->t_flags2 & TF2_CANNOT_DO_ECN) == 0)
1694 		sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1695 
1696 	if (V_tcp_syncookies)
1697 		sc->sc_iss = syncookie_generate(sch, sc);
1698 	else
1699 		sc->sc_iss = arc4random();
1700 #ifdef INET6
1701 	if (autoflowlabel) {
1702 		if (V_tcp_syncookies)
1703 			sc->sc_flowlabel = sc->sc_iss;
1704 		else
1705 			sc->sc_flowlabel = ip6_randomflowlabel();
1706 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1707 	}
1708 #endif
1709 	if (locked)
1710 		SCH_UNLOCK(sch);
1711 
1712 	if (tfo_cookie_valid) {
1713 		rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie);
1714 		/* INP_RUNLOCK(inp) will be performed by the caller */
1715 		goto tfo_expanded;
1716 	}
1717 
1718 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1719 	/*
1720 	 * Do a standard 3-way handshake.
1721 	 */
1722 	if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1723 		if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1724 			syncache_free(sc);
1725 		else if (sc != &scs)
1726 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1727 		TCPSTAT_INC(tcps_sndacks);
1728 		TCPSTAT_INC(tcps_sndtotal);
1729 	} else {
1730 		if (sc != &scs)
1731 			syncache_free(sc);
1732 		TCPSTAT_INC(tcps_sc_dropped);
1733 	}
1734 	goto donenoprobe;
1735 
1736 done:
1737 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1738 donenoprobe:
1739 	if (m)
1740 		m_freem(m);
1741 	/*
1742 	 * If tfo_pending is not NULL here, then a TFO SYN that did not
1743 	 * result in a new socket was processed and the associated pending
1744 	 * counter has not yet been decremented.  All such TFO processing paths
1745 	 * transit this point.
1746 	 */
1747 	if (tfo_pending != NULL)
1748 		tcp_fastopen_decrement_counter(tfo_pending);
1749 
1750 tfo_expanded:
1751 	if (cred != NULL)
1752 		crfree(cred);
1753 #ifdef MAC
1754 	if (sc == &scs)
1755 		mac_syncache_destroy(&maclabel);
1756 #endif
1757 	return (rv);
1758 }
1759 
1760 /*
1761  * Send SYN|ACK or ACK to the peer.  Either in response to a peer's segment,
1762  * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1763  */
1764 static int
1765 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1766 {
1767 	struct ip *ip = NULL;
1768 	struct mbuf *m;
1769 	struct tcphdr *th = NULL;
1770 	struct udphdr *udp = NULL;
1771 	int optlen, error = 0;	/* Make compiler happy */
1772 	u_int16_t hlen, tlen, mssopt, ulen;
1773 	struct tcpopt to;
1774 #ifdef INET6
1775 	struct ip6_hdr *ip6 = NULL;
1776 #endif
1777 
1778 	NET_EPOCH_ASSERT();
1779 
1780 	hlen =
1781 #ifdef INET6
1782 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1783 #endif
1784 		sizeof(struct ip);
1785 	tlen = hlen + sizeof(struct tcphdr);
1786 	if (sc->sc_port) {
1787 		tlen += sizeof(struct udphdr);
1788 	}
1789 	/* Determine MSS we advertize to other end of connection. */
1790 	mssopt = tcp_mssopt(&sc->sc_inc);
1791 	if (sc->sc_port)
1792 		mssopt -= V_tcp_udp_tunneling_overhead;
1793 	mssopt = max(mssopt, V_tcp_minmss);
1794 
1795 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1796 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1797 	    ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + "
1798 	    "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port,
1799 	    max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN));
1800 
1801 	/* Create the IP+TCP header from scratch. */
1802 	m = m_gethdr(M_NOWAIT, MT_DATA);
1803 	if (m == NULL)
1804 		return (ENOBUFS);
1805 #ifdef MAC
1806 	mac_syncache_create_mbuf(sc->sc_label, m);
1807 #endif
1808 	m->m_data += max_linkhdr;
1809 	m->m_len = tlen;
1810 	m->m_pkthdr.len = tlen;
1811 	m->m_pkthdr.rcvif = NULL;
1812 
1813 #ifdef INET6
1814 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1815 		ip6 = mtod(m, struct ip6_hdr *);
1816 		ip6->ip6_vfc = IPV6_VERSION;
1817 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1818 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1819 		ip6->ip6_plen = htons(tlen - hlen);
1820 		/* ip6_hlim is set after checksum */
1821 		/* Zero out traffic class and flow label. */
1822 		ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1823 		ip6->ip6_flow |= sc->sc_flowlabel;
1824 		if (sc->sc_port != 0) {
1825 			ip6->ip6_nxt = IPPROTO_UDP;
1826 			udp = (struct udphdr *)(ip6 + 1);
1827 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1828 			udp->uh_dport = sc->sc_port;
1829 			ulen = (tlen - sizeof(struct ip6_hdr));
1830 			th = (struct tcphdr *)(udp + 1);
1831 		} else {
1832 			ip6->ip6_nxt = IPPROTO_TCP;
1833 			th = (struct tcphdr *)(ip6 + 1);
1834 		}
1835 		ip6->ip6_flow |= htonl(sc->sc_ip_tos << IPV6_FLOWLABEL_LEN);
1836 	}
1837 #endif
1838 #if defined(INET6) && defined(INET)
1839 	else
1840 #endif
1841 #ifdef INET
1842 	{
1843 		ip = mtod(m, struct ip *);
1844 		ip->ip_v = IPVERSION;
1845 		ip->ip_hl = sizeof(struct ip) >> 2;
1846 		ip->ip_len = htons(tlen);
1847 		ip->ip_id = 0;
1848 		ip->ip_off = 0;
1849 		ip->ip_sum = 0;
1850 		ip->ip_src = sc->sc_inc.inc_laddr;
1851 		ip->ip_dst = sc->sc_inc.inc_faddr;
1852 		ip->ip_ttl = sc->sc_ip_ttl;
1853 		ip->ip_tos = sc->sc_ip_tos;
1854 
1855 		/*
1856 		 * See if we should do MTU discovery.  Route lookups are
1857 		 * expensive, so we will only unset the DF bit if:
1858 		 *
1859 		 *	1) path_mtu_discovery is disabled
1860 		 *	2) the SCF_UNREACH flag has been set
1861 		 */
1862 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1863 		       ip->ip_off |= htons(IP_DF);
1864 		if (sc->sc_port == 0) {
1865 			ip->ip_p = IPPROTO_TCP;
1866 			th = (struct tcphdr *)(ip + 1);
1867 		} else {
1868 			ip->ip_p = IPPROTO_UDP;
1869 			udp = (struct udphdr *)(ip + 1);
1870 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1871 			udp->uh_dport = sc->sc_port;
1872 			ulen = (tlen - sizeof(struct ip));
1873 			th = (struct tcphdr *)(udp + 1);
1874 		}
1875 	}
1876 #endif /* INET */
1877 	th->th_sport = sc->sc_inc.inc_lport;
1878 	th->th_dport = sc->sc_inc.inc_fport;
1879 
1880 	if (flags & TH_SYN)
1881 		th->th_seq = htonl(sc->sc_iss);
1882 	else
1883 		th->th_seq = htonl(sc->sc_iss + 1);
1884 	th->th_ack = htonl(sc->sc_irs + 1);
1885 	th->th_off = sizeof(struct tcphdr) >> 2;
1886 	th->th_win = htons(sc->sc_wnd);
1887 	th->th_urp = 0;
1888 
1889 	flags = tcp_ecn_syncache_respond(flags, sc);
1890 	tcp_set_flags(th, flags);
1891 
1892 	/* Tack on the TCP options. */
1893 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1894 		to.to_flags = 0;
1895 
1896 		if (flags & TH_SYN) {
1897 			to.to_mss = mssopt;
1898 			to.to_flags = TOF_MSS;
1899 			if (sc->sc_flags & SCF_WINSCALE) {
1900 				to.to_wscale = sc->sc_requested_r_scale;
1901 				to.to_flags |= TOF_SCALE;
1902 			}
1903 			if (sc->sc_flags & SCF_SACK)
1904 				to.to_flags |= TOF_SACKPERM;
1905 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1906 			if (sc->sc_flags & SCF_SIGNATURE)
1907 				to.to_flags |= TOF_SIGNATURE;
1908 #endif
1909 			if (sc->sc_tfo_cookie) {
1910 				to.to_flags |= TOF_FASTOPEN;
1911 				to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1912 				to.to_tfo_cookie = sc->sc_tfo_cookie;
1913 				/* don't send cookie again when retransmitting response */
1914 				sc->sc_tfo_cookie = NULL;
1915 			}
1916 		}
1917 		if (sc->sc_flags & SCF_TIMESTAMP) {
1918 			to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1919 			to.to_tsecr = sc->sc_tsreflect;
1920 			to.to_flags |= TOF_TS;
1921 		}
1922 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1923 
1924 		/* Adjust headers by option size. */
1925 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1926 		m->m_len += optlen;
1927 		m->m_pkthdr.len += optlen;
1928 #ifdef INET6
1929 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1930 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1931 		else
1932 #endif
1933 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1934 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1935 		if (sc->sc_flags & SCF_SIGNATURE) {
1936 			KASSERT(to.to_flags & TOF_SIGNATURE,
1937 			    ("tcp_addoptions() didn't set tcp_signature"));
1938 
1939 			/* NOTE: to.to_signature is inside of mbuf */
1940 			if (!TCPMD5_ENABLED() ||
1941 			    TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1942 				m_freem(m);
1943 				return (EACCES);
1944 			}
1945 		}
1946 #endif
1947 	} else
1948 		optlen = 0;
1949 
1950 	if (udp) {
1951 		ulen += optlen;
1952 		udp->uh_ulen = htons(ulen);
1953 	}
1954 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
1955 	/*
1956 	 * If we have peer's SYN and it has a flowid, then let's assign it to
1957 	 * our SYN|ACK.  ip6_output() and ip_output() will not assign flowid
1958 	 * to SYN|ACK due to lack of inp here.
1959 	 */
1960 	if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1961 		m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1962 		M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1963 	}
1964 #ifdef INET6
1965 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1966 		if (sc->sc_port) {
1967 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
1968 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1969 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen,
1970 			      IPPROTO_UDP, 0);
1971 			th->th_sum = htons(0);
1972 		} else {
1973 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1974 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1975 			th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1976 			    IPPROTO_TCP, 0);
1977 		}
1978 		ip6->ip6_hlim = sc->sc_ip_ttl;
1979 #ifdef TCP_OFFLOAD
1980 		if (ADDED_BY_TOE(sc)) {
1981 			struct toedev *tod = sc->sc_tod;
1982 
1983 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1984 
1985 			return (error);
1986 		}
1987 #endif
1988 		TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
1989 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1990 	}
1991 #endif
1992 #if defined(INET6) && defined(INET)
1993 	else
1994 #endif
1995 #ifdef INET
1996 	{
1997 		if (sc->sc_port) {
1998 			m->m_pkthdr.csum_flags = CSUM_UDP;
1999 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2000 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
2001 			      ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
2002 			th->th_sum = htons(0);
2003 		} else {
2004 			m->m_pkthdr.csum_flags = CSUM_TCP;
2005 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2006 			th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
2007 			    htons(tlen + optlen - hlen + IPPROTO_TCP));
2008 		}
2009 #ifdef TCP_OFFLOAD
2010 		if (ADDED_BY_TOE(sc)) {
2011 			struct toedev *tod = sc->sc_tod;
2012 
2013 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2014 
2015 			return (error);
2016 		}
2017 #endif
2018 		TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
2019 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
2020 	}
2021 #endif
2022 	return (error);
2023 }
2024 
2025 /*
2026  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
2027  * that exceed the capacity of the syncache by avoiding the storage of any
2028  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
2029  * attacks where the attacker does not have access to our responses.
2030  *
2031  * Syncookies encode and include all necessary information about the
2032  * connection setup within the SYN|ACK that we send back.  That way we
2033  * can avoid keeping any local state until the ACK to our SYN|ACK returns
2034  * (if ever).  Normally the syncache and syncookies are running in parallel
2035  * with the latter taking over when the former is exhausted.  When matching
2036  * syncache entry is found the syncookie is ignored.
2037  *
2038  * The only reliable information persisting the 3WHS is our initial sequence
2039  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
2040  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
2041  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
2042  * returns and signifies a legitimate connection if it matches the ACK.
2043  *
2044  * The available space of 32 bits to store the hash and to encode the SYN
2045  * option information is very tight and we should have at least 24 bits for
2046  * the MAC to keep the number of guesses by blind spoofing reasonably high.
2047  *
2048  * SYN option information we have to encode to fully restore a connection:
2049  * MSS: is imporant to chose an optimal segment size to avoid IP level
2050  *   fragmentation along the path.  The common MSS values can be encoded
2051  *   in a 3-bit table.  Uncommon values are captured by the next lower value
2052  *   in the table leading to a slight increase in packetization overhead.
2053  * WSCALE: is necessary to allow large windows to be used for high delay-
2054  *   bandwidth product links.  Not scaling the window when it was initially
2055  *   negotiated is bad for performance as lack of scaling further decreases
2056  *   the apparent available send window.  We only need to encode the WSCALE
2057  *   we received from the remote end.  Our end can be recalculated at any
2058  *   time.  The common WSCALE values can be encoded in a 3-bit table.
2059  *   Uncommon values are captured by the next lower value in the table
2060  *   making us under-estimate the available window size halving our
2061  *   theoretically possible maximum throughput for that connection.
2062  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2063  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2064  *   that are included in all segments on a connection.  We enable them when
2065  *   the ACK has them.
2066  *
2067  * Security of syncookies and attack vectors:
2068  *
2069  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2070  * together with the gloabl secret to make it unique per connection attempt.
2071  * Thus any change of any of those parameters results in a different MAC output
2072  * in an unpredictable way unless a collision is encountered.  24 bits of the
2073  * MAC are embedded into the ISS.
2074  *
2075  * To prevent replay attacks two rotating global secrets are updated with a
2076  * new random value every 15 seconds.  The life-time of a syncookie is thus
2077  * 15-30 seconds.
2078  *
2079  * Vector 1: Attacking the secret.  This requires finding a weakness in the
2080  * MAC itself or the way it is used here.  The attacker can do a chosen plain
2081  * text attack by varying and testing the all parameters under his control.
2082  * The strength depends on the size and randomness of the secret, and the
2083  * cryptographic security of the MAC function.  Due to the constant updating
2084  * of the secret the attacker has at most 29.999 seconds to find the secret
2085  * and launch spoofed connections.  After that he has to start all over again.
2086  *
2087  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
2088  * size an average of 4,823 attempts are required for a 50% chance of success
2089  * to spoof a single syncookie (birthday collision paradox).  However the
2090  * attacker is blind and doesn't know if one of his attempts succeeded unless
2091  * he has a side channel to interfere success from.  A single connection setup
2092  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2093  * This many attempts are required for each one blind spoofed connection.  For
2094  * every additional spoofed connection he has to launch another N attempts.
2095  * Thus for a sustained rate 100 spoofed connections per second approximately
2096  * 1,800,000 packets per second would have to be sent.
2097  *
2098  * NB: The MAC function should be fast so that it doesn't become a CPU
2099  * exhaustion attack vector itself.
2100  *
2101  * References:
2102  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2103  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2104  *   http://cr.yp.to/syncookies.html    (overview)
2105  *   http://cr.yp.to/syncookies/archive (details)
2106  *
2107  *
2108  * Schematic construction of a syncookie enabled Initial Sequence Number:
2109  *  0        1         2         3
2110  *  12345678901234567890123456789012
2111  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2112  *
2113  *  x 24 MAC (truncated)
2114  *  W  3 Send Window Scale index
2115  *  M  3 MSS index
2116  *  S  1 SACK permitted
2117  *  P  1 Odd/even secret
2118  */
2119 
2120 /*
2121  * Distribution and probability of certain MSS values.  Those in between are
2122  * rounded down to the next lower one.
2123  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2124  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
2125  */
2126 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2127 
2128 /*
2129  * Distribution and probability of certain WSCALE values.  We have to map the
2130  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2131  * bits based on prevalence of certain values.  Where we don't have an exact
2132  * match for are rounded down to the next lower one letting us under-estimate
2133  * the true available window.  At the moment this would happen only for the
2134  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2135  * and window size).  The absence of the WSCALE option (no scaling in either
2136  * direction) is encoded with index zero.
2137  * [WSCALE values histograms, Allman, 2012]
2138  *                            X 10 10 35  5  6 14 10%   by host
2139  *                            X 11  4  5  5 18 49  3%   by connections
2140  */
2141 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2142 
2143 /*
2144  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
2145  * and good cryptographic properties.
2146  */
2147 static uint32_t
2148 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2149     uint8_t *secbits, uintptr_t secmod)
2150 {
2151 	SIPHASH_CTX ctx;
2152 	uint32_t siphash[2];
2153 
2154 	SipHash24_Init(&ctx);
2155 	SipHash_SetKey(&ctx, secbits);
2156 	switch (inc->inc_flags & INC_ISIPV6) {
2157 #ifdef INET
2158 	case 0:
2159 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2160 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2161 		break;
2162 #endif
2163 #ifdef INET6
2164 	case INC_ISIPV6:
2165 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2166 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2167 		break;
2168 #endif
2169 	}
2170 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2171 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2172 	SipHash_Update(&ctx, &irs, sizeof(irs));
2173 	SipHash_Update(&ctx, &flags, sizeof(flags));
2174 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
2175 	SipHash_Final((u_int8_t *)&siphash, &ctx);
2176 
2177 	return (siphash[0] ^ siphash[1]);
2178 }
2179 
2180 static tcp_seq
2181 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2182 {
2183 	u_int i, secbit, wscale;
2184 	uint32_t iss, hash;
2185 	uint8_t *secbits;
2186 	union syncookie cookie;
2187 
2188 	cookie.cookie = 0;
2189 
2190 	/* Map our computed MSS into the 3-bit index. */
2191 	for (i = nitems(tcp_sc_msstab) - 1;
2192 	     tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2193 	     i--)
2194 		;
2195 	cookie.flags.mss_idx = i;
2196 
2197 	/*
2198 	 * Map the send window scale into the 3-bit index but only if
2199 	 * the wscale option was received.
2200 	 */
2201 	if (sc->sc_flags & SCF_WINSCALE) {
2202 		wscale = sc->sc_requested_s_scale;
2203 		for (i = nitems(tcp_sc_wstab) - 1;
2204 		    tcp_sc_wstab[i] > wscale && i > 0;
2205 		     i--)
2206 			;
2207 		cookie.flags.wscale_idx = i;
2208 	}
2209 
2210 	/* Can we do SACK? */
2211 	if (sc->sc_flags & SCF_SACK)
2212 		cookie.flags.sack_ok = 1;
2213 
2214 	/* Which of the two secrets to use. */
2215 	secbit = V_tcp_syncache.secret.oddeven & 0x1;
2216 	cookie.flags.odd_even = secbit;
2217 
2218 	secbits = V_tcp_syncache.secret.key[secbit];
2219 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2220 	    (uintptr_t)sch);
2221 
2222 	/*
2223 	 * Put the flags into the hash and XOR them to get better ISS number
2224 	 * variance.  This doesn't enhance the cryptographic strength and is
2225 	 * done to prevent the 8 cookie bits from showing up directly on the
2226 	 * wire.
2227 	 */
2228 	iss = hash & ~0xff;
2229 	iss |= cookie.cookie ^ (hash >> 24);
2230 
2231 	TCPSTAT_INC(tcps_sc_sendcookie);
2232 	return (iss);
2233 }
2234 
2235 static struct syncache *
2236 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2237     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2238     struct socket *lso, uint16_t port)
2239 {
2240 	uint32_t hash;
2241 	uint8_t *secbits;
2242 	tcp_seq ack, seq;
2243 	int wnd, wscale = 0;
2244 	union syncookie cookie;
2245 
2246 	/*
2247 	 * Pull information out of SYN-ACK/ACK and revert sequence number
2248 	 * advances.
2249 	 */
2250 	ack = th->th_ack - 1;
2251 	seq = th->th_seq - 1;
2252 
2253 	/*
2254 	 * Unpack the flags containing enough information to restore the
2255 	 * connection.
2256 	 */
2257 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2258 
2259 	/* Which of the two secrets to use. */
2260 	secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2261 
2262 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2263 
2264 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
2265 	if ((ack & ~0xff) != (hash & ~0xff))
2266 		return (NULL);
2267 
2268 	/* Fill in the syncache values. */
2269 	sc->sc_flags = 0;
2270 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2271 	sc->sc_ipopts = NULL;
2272 
2273 	sc->sc_irs = seq;
2274 	sc->sc_iss = ack;
2275 
2276 	switch (inc->inc_flags & INC_ISIPV6) {
2277 #ifdef INET
2278 	case 0:
2279 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2280 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2281 		break;
2282 #endif
2283 #ifdef INET6
2284 	case INC_ISIPV6:
2285 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2286 			sc->sc_flowlabel =
2287 			    htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2288 		break;
2289 #endif
2290 	}
2291 
2292 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2293 
2294 	/* We can simply recompute receive window scale we sent earlier. */
2295 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2296 		wscale++;
2297 
2298 	/* Only use wscale if it was enabled in the orignal SYN. */
2299 	if (cookie.flags.wscale_idx > 0) {
2300 		sc->sc_requested_r_scale = wscale;
2301 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2302 		sc->sc_flags |= SCF_WINSCALE;
2303 	}
2304 
2305 	wnd = lso->sol_sbrcv_hiwat;
2306 	wnd = imax(wnd, 0);
2307 	wnd = imin(wnd, TCP_MAXWIN);
2308 	sc->sc_wnd = wnd;
2309 
2310 	if (cookie.flags.sack_ok)
2311 		sc->sc_flags |= SCF_SACK;
2312 
2313 	if (to->to_flags & TOF_TS) {
2314 		sc->sc_flags |= SCF_TIMESTAMP;
2315 		sc->sc_tsreflect = to->to_tsval;
2316 		sc->sc_tsoff = tcp_new_ts_offset(inc);
2317 	}
2318 
2319 	if (to->to_flags & TOF_SIGNATURE)
2320 		sc->sc_flags |= SCF_SIGNATURE;
2321 
2322 	sc->sc_rxmits = 0;
2323 
2324 	sc->sc_port = port;
2325 
2326 	TCPSTAT_INC(tcps_sc_recvcookie);
2327 	return (sc);
2328 }
2329 
2330 #ifdef INVARIANTS
2331 static int
2332 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2333     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2334     struct socket *lso, uint16_t port)
2335 {
2336 	struct syncache scs, *scx;
2337 	char *s;
2338 
2339 	bzero(&scs, sizeof(scs));
2340 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port);
2341 
2342 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2343 		return (0);
2344 
2345 	if (scx != NULL) {
2346 		if (sc->sc_peer_mss != scx->sc_peer_mss)
2347 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2348 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2349 
2350 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2351 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2352 			    s, __func__, sc->sc_requested_r_scale,
2353 			    scx->sc_requested_r_scale);
2354 
2355 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2356 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2357 			    s, __func__, sc->sc_requested_s_scale,
2358 			    scx->sc_requested_s_scale);
2359 
2360 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2361 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2362 	}
2363 
2364 	if (s != NULL)
2365 		free(s, M_TCPLOG);
2366 	return (0);
2367 }
2368 #endif /* INVARIANTS */
2369 
2370 static void
2371 syncookie_reseed(void *arg)
2372 {
2373 	struct tcp_syncache *sc = arg;
2374 	uint8_t *secbits;
2375 	int secbit;
2376 
2377 	/*
2378 	 * Reseeding the secret doesn't have to be protected by a lock.
2379 	 * It only must be ensured that the new random values are visible
2380 	 * to all CPUs in a SMP environment.  The atomic with release
2381 	 * semantics ensures that.
2382 	 */
2383 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2384 	secbits = sc->secret.key[secbit];
2385 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2386 	atomic_add_rel_int(&sc->secret.oddeven, 1);
2387 
2388 	/* Reschedule ourself. */
2389 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2390 }
2391 
2392 /*
2393  * We have overflowed a bucket. Let's pause dealing with the syncache.
2394  * This function will increment the bucketoverflow statistics appropriately
2395  * (once per pause when pausing is enabled; otherwise, once per overflow).
2396  */
2397 static void
2398 syncache_pause(struct in_conninfo *inc)
2399 {
2400 	time_t delta;
2401 	const char *s;
2402 
2403 	/* XXX:
2404 	 * 2. Add sysctl read here so we don't get the benefit of this
2405 	 * change without the new sysctl.
2406 	 */
2407 
2408 	/*
2409 	 * Try an unlocked read. If we already know that another thread
2410 	 * has activated the feature, there is no need to proceed.
2411 	 */
2412 	if (V_tcp_syncache.paused)
2413 		return;
2414 
2415 	/* Are cookied enabled? If not, we can't pause. */
2416 	if (!V_tcp_syncookies) {
2417 		TCPSTAT_INC(tcps_sc_bucketoverflow);
2418 		return;
2419 	}
2420 
2421 	/*
2422 	 * We may be the first thread to find an overflow. Get the lock
2423 	 * and evaluate if we need to take action.
2424 	 */
2425 	mtx_lock(&V_tcp_syncache.pause_mtx);
2426 	if (V_tcp_syncache.paused) {
2427 		mtx_unlock(&V_tcp_syncache.pause_mtx);
2428 		return;
2429 	}
2430 
2431 	/* Activate protection. */
2432 	V_tcp_syncache.paused = true;
2433 	TCPSTAT_INC(tcps_sc_bucketoverflow);
2434 
2435 	/*
2436 	 * Determine the last backoff time. If we are seeing a re-newed
2437 	 * attack within that same time after last reactivating the syncache,
2438 	 * consider it an extension of the same attack.
2439 	 */
2440 	delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2441 	if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2442 		if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2443 			delta <<= 1;
2444 			V_tcp_syncache.pause_backoff++;
2445 		}
2446 	} else {
2447 		delta = TCP_SYNCACHE_PAUSE_TIME;
2448 		V_tcp_syncache.pause_backoff = 0;
2449 	}
2450 
2451 	/* Log a warning, including IP addresses, if able. */
2452 	if (inc != NULL)
2453 		s = tcp_log_addrs(inc, NULL, NULL, NULL);
2454 	else
2455 		s = (const char *)NULL;
2456 	log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2457 	    "the next %lld seconds%s%s%s\n", (long long)delta,
2458 	    (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2459 	    (s != NULL) ? ")" : "");
2460 	free(__DECONST(void *, s), M_TCPLOG);
2461 
2462 	/* Use the calculated delta to set a new pause time. */
2463 	V_tcp_syncache.pause_until = time_uptime + delta;
2464 	callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2465 	    &V_tcp_syncache);
2466 	mtx_unlock(&V_tcp_syncache.pause_mtx);
2467 }
2468 
2469 /* Evaluate whether we need to unpause. */
2470 static void
2471 syncache_unpause(void *arg)
2472 {
2473 	struct tcp_syncache *sc;
2474 	time_t delta;
2475 
2476 	sc = arg;
2477 	mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2478 	callout_deactivate(&sc->pause_co);
2479 
2480 	/*
2481 	 * Check to make sure we are not running early. If the pause
2482 	 * time has expired, then deactivate the protection.
2483 	 */
2484 	if ((delta = sc->pause_until - time_uptime) > 0)
2485 		callout_schedule(&sc->pause_co, delta * hz);
2486 	else
2487 		sc->paused = false;
2488 }
2489 
2490 /*
2491  * Exports the syncache entries to userland so that netstat can display
2492  * them alongside the other sockets.  This function is intended to be
2493  * called only from tcp_pcblist.
2494  *
2495  * Due to concurrency on an active system, the number of pcbs exported
2496  * may have no relation to max_pcbs.  max_pcbs merely indicates the
2497  * amount of space the caller allocated for this function to use.
2498  */
2499 int
2500 syncache_pcblist(struct sysctl_req *req)
2501 {
2502 	struct xtcpcb xt;
2503 	struct syncache *sc;
2504 	struct syncache_head *sch;
2505 	int error, i;
2506 
2507 	bzero(&xt, sizeof(xt));
2508 	xt.xt_len = sizeof(xt);
2509 	xt.t_state = TCPS_SYN_RECEIVED;
2510 	xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2511 	xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2512 	xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2513 	xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2514 
2515 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2516 		sch = &V_tcp_syncache.hashbase[i];
2517 		SCH_LOCK(sch);
2518 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2519 			if (sc->sc_cred != NULL &&
2520 			    cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2521 				continue;
2522 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2523 				xt.xt_inp.inp_vflag = INP_IPV6;
2524 			else
2525 				xt.xt_inp.inp_vflag = INP_IPV4;
2526 			xt.xt_encaps_port = sc->sc_port;
2527 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2528 			    sizeof (struct in_conninfo));
2529 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2530 			if (error) {
2531 				SCH_UNLOCK(sch);
2532 				return (0);
2533 			}
2534 		}
2535 		SCH_UNLOCK(sch);
2536 	}
2537 
2538 	return (0);
2539 }
2540