xref: /freebsd/sys/netinet/tcp_syncache.c (revision 480093f4440d54b30b3025afeac24b48f2ba7a2e)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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 __FBSDID("$FreeBSD$");
37 
38 #include "opt_inet.h"
39 #include "opt_inet6.h"
40 #include "opt_ipsec.h"
41 #include "opt_pcbgroup.h"
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/hash.h>
46 #include <sys/refcount.h>
47 #include <sys/kernel.h>
48 #include <sys/sysctl.h>
49 #include <sys/limits.h>
50 #include <sys/lock.h>
51 #include <sys/mutex.h>
52 #include <sys/malloc.h>
53 #include <sys/mbuf.h>
54 #include <sys/proc.h>		/* for proc0 declaration */
55 #include <sys/random.h>
56 #include <sys/socket.h>
57 #include <sys/socketvar.h>
58 #include <sys/syslog.h>
59 #include <sys/ucred.h>
60 
61 #include <sys/md5.h>
62 #include <crypto/siphash/siphash.h>
63 
64 #include <vm/uma.h>
65 
66 #include <net/if.h>
67 #include <net/if_var.h>
68 #include <net/route.h>
69 #include <net/vnet.h>
70 
71 #include <netinet/in.h>
72 #include <netinet/in_kdtrace.h>
73 #include <netinet/in_systm.h>
74 #include <netinet/ip.h>
75 #include <netinet/in_var.h>
76 #include <netinet/in_pcb.h>
77 #include <netinet/ip_var.h>
78 #include <netinet/ip_options.h>
79 #ifdef INET6
80 #include <netinet/ip6.h>
81 #include <netinet/icmp6.h>
82 #include <netinet6/nd6.h>
83 #include <netinet6/ip6_var.h>
84 #include <netinet6/in6_pcb.h>
85 #endif
86 #include <netinet/tcp.h>
87 #include <netinet/tcp_fastopen.h>
88 #include <netinet/tcp_fsm.h>
89 #include <netinet/tcp_seq.h>
90 #include <netinet/tcp_timer.h>
91 #include <netinet/tcp_var.h>
92 #include <netinet/tcp_syncache.h>
93 #ifdef INET6
94 #include <netinet6/tcp6_var.h>
95 #endif
96 #ifdef TCP_OFFLOAD
97 #include <netinet/toecore.h>
98 #endif
99 
100 #include <netipsec/ipsec_support.h>
101 
102 #include <machine/in_cksum.h>
103 
104 #include <security/mac/mac_framework.h>
105 
106 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1;
107 #define	V_tcp_syncookies		VNET(tcp_syncookies)
108 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
109     &VNET_NAME(tcp_syncookies), 0,
110     "Use TCP SYN cookies if the syncache overflows");
111 
112 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0;
113 #define	V_tcp_syncookiesonly		VNET(tcp_syncookiesonly)
114 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
115     &VNET_NAME(tcp_syncookiesonly), 0,
116     "Use only TCP SYN cookies");
117 
118 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1;
119 #define V_functions_inherit_listen_socket_stack \
120     VNET(functions_inherit_listen_socket_stack)
121 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack,
122     CTLFLAG_VNET | CTLFLAG_RW,
123     &VNET_NAME(functions_inherit_listen_socket_stack), 0,
124     "Inherit listen socket's stack");
125 
126 #ifdef TCP_OFFLOAD
127 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
128 #endif
129 
130 static void	 syncache_drop(struct syncache *, struct syncache_head *);
131 static void	 syncache_free(struct syncache *);
132 static void	 syncache_insert(struct syncache *, struct syncache_head *);
133 static int	 syncache_respond(struct syncache *, const struct mbuf *, int);
134 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
135 		    struct mbuf *m);
136 static void	 syncache_timeout(struct syncache *sc, struct syncache_head *sch,
137 		    int docallout);
138 static void	 syncache_timer(void *);
139 
140 static uint32_t	 syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
141 		    uint8_t *, uintptr_t);
142 static tcp_seq	 syncookie_generate(struct syncache_head *, struct syncache *);
143 static struct syncache
144 		*syncookie_lookup(struct in_conninfo *, struct syncache_head *,
145 		    struct syncache *, struct tcphdr *, struct tcpopt *,
146 		    struct socket *);
147 static void	syncache_pause(struct in_conninfo *);
148 static void	syncache_unpause(void *);
149 static void	 syncookie_reseed(void *);
150 #ifdef INVARIANTS
151 static int	 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
152 		    struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
153 		    struct socket *lso);
154 #endif
155 
156 /*
157  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
158  * 3 retransmits corresponds to a timeout with default values of
159  * tcp_rexmit_initial * (             1 +
160  *                       tcp_backoff[1] +
161  *                       tcp_backoff[2] +
162  *                       tcp_backoff[3]) + 3 * tcp_rexmit_slop,
163  * 1000 ms * (1 + 2 + 4 + 8) +  3 * 200 ms = 15600 ms,
164  * the odds are that the user has given up attempting to connect by then.
165  */
166 #define SYNCACHE_MAXREXMTS		3
167 
168 /* Arbitrary values */
169 #define TCP_SYNCACHE_HASHSIZE		512
170 #define TCP_SYNCACHE_BUCKETLIMIT	30
171 
172 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
173 #define	V_tcp_syncache			VNET(tcp_syncache)
174 
175 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0,
176     "TCP SYN cache");
177 
178 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
179     &VNET_NAME(tcp_syncache.bucket_limit), 0,
180     "Per-bucket hash limit for syncache");
181 
182 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
183     &VNET_NAME(tcp_syncache.cache_limit), 0,
184     "Overall entry limit for syncache");
185 
186 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
187     &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
188 
189 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
190     &VNET_NAME(tcp_syncache.hashsize), 0,
191     "Size of TCP syncache hashtable");
192 
193 static int
194 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
195 {
196 	int error;
197 	u_int new;
198 
199 	new = V_tcp_syncache.rexmt_limit;
200 	error = sysctl_handle_int(oidp, &new, 0, req);
201 	if ((error == 0) && (req->newptr != NULL)) {
202 		if (new > TCP_MAXRXTSHIFT)
203 			error = EINVAL;
204 		else
205 			V_tcp_syncache.rexmt_limit = new;
206 	}
207 	return (error);
208 }
209 
210 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
211     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW,
212     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
213     sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
214     "Limit on SYN/ACK retransmissions");
215 
216 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
217 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
218     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
219     "Send reset on socket allocation failure");
220 
221 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
222 
223 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
224 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
225 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
226 
227 /*
228  * Requires the syncache entry to be already removed from the bucket list.
229  */
230 static void
231 syncache_free(struct syncache *sc)
232 {
233 
234 	if (sc->sc_ipopts)
235 		(void) m_free(sc->sc_ipopts);
236 	if (sc->sc_cred)
237 		crfree(sc->sc_cred);
238 #ifdef MAC
239 	mac_syncache_destroy(&sc->sc_label);
240 #endif
241 
242 	uma_zfree(V_tcp_syncache.zone, sc);
243 }
244 
245 void
246 syncache_init(void)
247 {
248 	int i;
249 
250 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
251 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
252 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
253 	V_tcp_syncache.hash_secret = arc4random();
254 
255 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
256 	    &V_tcp_syncache.hashsize);
257 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
258 	    &V_tcp_syncache.bucket_limit);
259 	if (!powerof2(V_tcp_syncache.hashsize) ||
260 	    V_tcp_syncache.hashsize == 0) {
261 		printf("WARNING: syncache hash size is not a power of 2.\n");
262 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
263 	}
264 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
265 
266 	/* Set limits. */
267 	V_tcp_syncache.cache_limit =
268 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
269 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
270 	    &V_tcp_syncache.cache_limit);
271 
272 	/* Allocate the hash table. */
273 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
274 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
275 
276 #ifdef VIMAGE
277 	V_tcp_syncache.vnet = curvnet;
278 #endif
279 
280 	/* Initialize the hash buckets. */
281 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
282 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
283 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
284 			 NULL, MTX_DEF);
285 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
286 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
287 		V_tcp_syncache.hashbase[i].sch_length = 0;
288 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
289 		V_tcp_syncache.hashbase[i].sch_last_overflow =
290 		    -(SYNCOOKIE_LIFETIME + 1);
291 	}
292 
293 	/* Create the syncache entry zone. */
294 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
295 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
296 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
297 	    V_tcp_syncache.cache_limit);
298 
299 	/* Start the SYN cookie reseeder callout. */
300 	callout_init(&V_tcp_syncache.secret.reseed, 1);
301 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
302 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
303 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
304 	    syncookie_reseed, &V_tcp_syncache);
305 
306 	/* Initialize the pause machinery. */
307 	mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF);
308 	callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx,
309 	    0);
310 	V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME;
311 	V_tcp_syncache.pause_backoff = 0;
312 	V_tcp_syncache.paused = false;
313 }
314 
315 #ifdef VIMAGE
316 void
317 syncache_destroy(void)
318 {
319 	struct syncache_head *sch;
320 	struct syncache *sc, *nsc;
321 	int i;
322 
323 	/*
324 	 * Stop the re-seed timer before freeing resources.  No need to
325 	 * possibly schedule it another time.
326 	 */
327 	callout_drain(&V_tcp_syncache.secret.reseed);
328 
329 	/* Stop the SYN cache pause callout. */
330 	mtx_lock(&V_tcp_syncache.pause_mtx);
331 	if (callout_stop(&V_tcp_syncache.pause_co) == 0) {
332 		mtx_unlock(&V_tcp_syncache.pause_mtx);
333 		callout_drain(&V_tcp_syncache.pause_co);
334 	} else
335 		mtx_unlock(&V_tcp_syncache.pause_mtx);
336 
337 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
338 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
339 
340 		sch = &V_tcp_syncache.hashbase[i];
341 		callout_drain(&sch->sch_timer);
342 
343 		SCH_LOCK(sch);
344 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
345 			syncache_drop(sc, sch);
346 		SCH_UNLOCK(sch);
347 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
348 		    ("%s: sch->sch_bucket not empty", __func__));
349 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
350 		    __func__, sch->sch_length));
351 		mtx_destroy(&sch->sch_mtx);
352 	}
353 
354 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
355 	    ("%s: cache_count not 0", __func__));
356 
357 	/* Free the allocated global resources. */
358 	uma_zdestroy(V_tcp_syncache.zone);
359 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
360 	mtx_destroy(&V_tcp_syncache.pause_mtx);
361 }
362 #endif
363 
364 /*
365  * Inserts a syncache entry into the specified bucket row.
366  * Locks and unlocks the syncache_head autonomously.
367  */
368 static void
369 syncache_insert(struct syncache *sc, struct syncache_head *sch)
370 {
371 	struct syncache *sc2;
372 
373 	SCH_LOCK(sch);
374 
375 	/*
376 	 * Make sure that we don't overflow the per-bucket limit.
377 	 * If the bucket is full, toss the oldest element.
378 	 */
379 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
380 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
381 			("sch->sch_length incorrect"));
382 		syncache_pause(&sc->sc_inc);
383 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
384 		sch->sch_last_overflow = time_uptime;
385 		syncache_drop(sc2, sch);
386 	}
387 
388 	/* Put it into the bucket. */
389 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
390 	sch->sch_length++;
391 
392 #ifdef TCP_OFFLOAD
393 	if (ADDED_BY_TOE(sc)) {
394 		struct toedev *tod = sc->sc_tod;
395 
396 		tod->tod_syncache_added(tod, sc->sc_todctx);
397 	}
398 #endif
399 
400 	/* Reinitialize the bucket row's timer. */
401 	if (sch->sch_length == 1)
402 		sch->sch_nextc = ticks + INT_MAX;
403 	syncache_timeout(sc, sch, 1);
404 
405 	SCH_UNLOCK(sch);
406 
407 	TCPSTATES_INC(TCPS_SYN_RECEIVED);
408 	TCPSTAT_INC(tcps_sc_added);
409 }
410 
411 /*
412  * Remove and free entry from syncache bucket row.
413  * Expects locked syncache head.
414  */
415 static void
416 syncache_drop(struct syncache *sc, struct syncache_head *sch)
417 {
418 
419 	SCH_LOCK_ASSERT(sch);
420 
421 	TCPSTATES_DEC(TCPS_SYN_RECEIVED);
422 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
423 	sch->sch_length--;
424 
425 #ifdef TCP_OFFLOAD
426 	if (ADDED_BY_TOE(sc)) {
427 		struct toedev *tod = sc->sc_tod;
428 
429 		tod->tod_syncache_removed(tod, sc->sc_todctx);
430 	}
431 #endif
432 
433 	syncache_free(sc);
434 }
435 
436 /*
437  * Engage/reengage time on bucket row.
438  */
439 static void
440 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
441 {
442 	int rexmt;
443 
444 	if (sc->sc_rxmits == 0)
445 		rexmt = tcp_rexmit_initial;
446 	else
447 		TCPT_RANGESET(rexmt,
448 		    tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
449 		    tcp_rexmit_min, TCPTV_REXMTMAX);
450 	sc->sc_rxttime = ticks + rexmt;
451 	sc->sc_rxmits++;
452 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
453 		sch->sch_nextc = sc->sc_rxttime;
454 		if (docallout)
455 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
456 			    syncache_timer, (void *)sch);
457 	}
458 }
459 
460 /*
461  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
462  * If we have retransmitted an entry the maximum number of times, expire it.
463  * One separate timer for each bucket row.
464  */
465 static void
466 syncache_timer(void *xsch)
467 {
468 	struct syncache_head *sch = (struct syncache_head *)xsch;
469 	struct syncache *sc, *nsc;
470 	int tick = ticks;
471 	char *s;
472 	bool paused;
473 
474 	CURVNET_SET(sch->sch_sc->vnet);
475 
476 	/* NB: syncache_head has already been locked by the callout. */
477 	SCH_LOCK_ASSERT(sch);
478 
479 	/*
480 	 * In the following cycle we may remove some entries and/or
481 	 * advance some timeouts, so re-initialize the bucket timer.
482 	 */
483 	sch->sch_nextc = tick + INT_MAX;
484 
485 	/*
486 	 * If we have paused processing, unconditionally remove
487 	 * all syncache entries.
488 	 */
489 	mtx_lock(&V_tcp_syncache.pause_mtx);
490 	paused = V_tcp_syncache.paused;
491 	mtx_unlock(&V_tcp_syncache.pause_mtx);
492 
493 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
494 		if (paused) {
495 			syncache_drop(sc, sch);
496 			continue;
497 		}
498 		/*
499 		 * We do not check if the listen socket still exists
500 		 * and accept the case where the listen socket may be
501 		 * gone by the time we resend the SYN/ACK.  We do
502 		 * not expect this to happens often. If it does,
503 		 * then the RST will be sent by the time the remote
504 		 * host does the SYN/ACK->ACK.
505 		 */
506 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
507 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
508 				sch->sch_nextc = sc->sc_rxttime;
509 			continue;
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 		syncache_respond(sc, NULL, TH_SYN|TH_ACK);
530 		TCPSTAT_INC(tcps_sc_retransmitted);
531 		syncache_timeout(sc, sch, 0);
532 	}
533 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
534 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
535 			syncache_timer, (void *)(sch));
536 	CURVNET_RESTORE();
537 }
538 
539 /*
540  * Returns true if the system is only using cookies at the moment.
541  * This could be due to a sysadmin decision to only use cookies, or it
542  * could be due to the system detecting an attack.
543  */
544 static inline bool
545 syncache_cookiesonly(void)
546 {
547 
548 	return (V_tcp_syncookies && (V_tcp_syncache.paused ||
549 	    V_tcp_syncookiesonly));
550 }
551 
552 /*
553  * Find the hash bucket for the given connection.
554  */
555 static struct syncache_head *
556 syncache_hashbucket(struct in_conninfo *inc)
557 {
558 	uint32_t hash;
559 
560 	/*
561 	 * The hash is built on foreign port + local port + foreign address.
562 	 * We rely on the fact that struct in_conninfo starts with 16 bits
563 	 * of foreign port, then 16 bits of local port then followed by 128
564 	 * bits of foreign address.  In case of IPv4 address, the first 3
565 	 * 32-bit words of the address always are zeroes.
566 	 */
567 	hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
568 	    V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
569 
570 	return (&V_tcp_syncache.hashbase[hash]);
571 }
572 
573 /*
574  * Find an entry in the syncache.
575  * Returns always with locked syncache_head plus a matching entry or NULL.
576  */
577 static struct syncache *
578 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
579 {
580 	struct syncache *sc;
581 	struct syncache_head *sch;
582 
583 	*schp = sch = syncache_hashbucket(inc);
584 	SCH_LOCK(sch);
585 
586 	/* Circle through bucket row to find matching entry. */
587 	TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
588 		if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
589 		    sizeof(struct in_endpoints)) == 0)
590 			break;
591 
592 	return (sc);	/* Always returns with locked sch. */
593 }
594 
595 /*
596  * This function is called when we get a RST for a
597  * non-existent connection, so that we can see if the
598  * connection is in the syn cache.  If it is, zap it.
599  * If required send a challenge ACK.
600  */
601 void
602 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m)
603 {
604 	struct syncache *sc;
605 	struct syncache_head *sch;
606 	char *s = NULL;
607 
608 	if (syncache_cookiesonly())
609 		return;
610 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
611 	SCH_LOCK_ASSERT(sch);
612 
613 	/*
614 	 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
615 	 * See RFC 793 page 65, section SEGMENT ARRIVES.
616 	 */
617 	if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
618 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
619 			log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
620 			    "FIN flag set, segment ignored\n", s, __func__);
621 		TCPSTAT_INC(tcps_badrst);
622 		goto done;
623 	}
624 
625 	/*
626 	 * No corresponding connection was found in syncache.
627 	 * If syncookies are enabled and possibly exclusively
628 	 * used, or we are under memory pressure, a valid RST
629 	 * may not find a syncache entry.  In that case we're
630 	 * done and no SYN|ACK retransmissions will happen.
631 	 * Otherwise the RST was misdirected or spoofed.
632 	 */
633 	if (sc == NULL) {
634 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
635 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
636 			    "syncache entry (possibly syncookie only), "
637 			    "segment ignored\n", s, __func__);
638 		TCPSTAT_INC(tcps_badrst);
639 		goto done;
640 	}
641 
642 	/*
643 	 * If the RST bit is set, check the sequence number to see
644 	 * if this is a valid reset segment.
645 	 *
646 	 * RFC 793 page 37:
647 	 *   In all states except SYN-SENT, all reset (RST) segments
648 	 *   are validated by checking their SEQ-fields.  A reset is
649 	 *   valid if its sequence number is in the window.
650 	 *
651 	 * RFC 793 page 69:
652 	 *   There are four cases for the acceptability test for an incoming
653 	 *   segment:
654 	 *
655 	 * Segment Receive  Test
656 	 * Length  Window
657 	 * ------- -------  -------------------------------------------
658 	 *    0       0     SEG.SEQ = RCV.NXT
659 	 *    0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
660 	 *   >0       0     not acceptable
661 	 *   >0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
662 	 *               or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
663 	 *
664 	 * Note that when receiving a SYN segment in the LISTEN state,
665 	 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
666 	 * described in RFC 793, page 66.
667 	 */
668 	if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
669 	    SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
670 	    (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
671 		if (V_tcp_insecure_rst ||
672 		    th->th_seq == sc->sc_irs + 1) {
673 			syncache_drop(sc, sch);
674 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
675 				log(LOG_DEBUG,
676 				    "%s; %s: Our SYN|ACK was rejected, "
677 				    "connection attempt aborted by remote "
678 				    "endpoint\n",
679 				    s, __func__);
680 			TCPSTAT_INC(tcps_sc_reset);
681 		} else {
682 			TCPSTAT_INC(tcps_badrst);
683 			/* Send challenge ACK. */
684 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
685 				log(LOG_DEBUG, "%s; %s: RST with invalid "
686 				    " SEQ %u != NXT %u (+WND %u), "
687 				    "sending challenge ACK\n",
688 				    s, __func__,
689 				    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
690 			syncache_respond(sc, m, TH_ACK);
691 		}
692 	} else {
693 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
694 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
695 			    "NXT %u (+WND %u), segment ignored\n",
696 			    s, __func__,
697 			    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
698 		TCPSTAT_INC(tcps_badrst);
699 	}
700 
701 done:
702 	if (s != NULL)
703 		free(s, M_TCPLOG);
704 	SCH_UNLOCK(sch);
705 }
706 
707 void
708 syncache_badack(struct in_conninfo *inc)
709 {
710 	struct syncache *sc;
711 	struct syncache_head *sch;
712 
713 	if (syncache_cookiesonly())
714 		return;
715 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
716 	SCH_LOCK_ASSERT(sch);
717 	if (sc != NULL) {
718 		syncache_drop(sc, sch);
719 		TCPSTAT_INC(tcps_sc_badack);
720 	}
721 	SCH_UNLOCK(sch);
722 }
723 
724 void
725 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq)
726 {
727 	struct syncache *sc;
728 	struct syncache_head *sch;
729 
730 	if (syncache_cookiesonly())
731 		return;
732 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
733 	SCH_LOCK_ASSERT(sch);
734 	if (sc == NULL)
735 		goto done;
736 
737 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
738 	if (ntohl(th_seq) != sc->sc_iss)
739 		goto done;
740 
741 	/*
742 	 * If we've rertransmitted 3 times and this is our second error,
743 	 * we remove the entry.  Otherwise, we allow it to continue on.
744 	 * This prevents us from incorrectly nuking an entry during a
745 	 * spurious network outage.
746 	 *
747 	 * See tcp_notify().
748 	 */
749 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
750 		sc->sc_flags |= SCF_UNREACH;
751 		goto done;
752 	}
753 	syncache_drop(sc, sch);
754 	TCPSTAT_INC(tcps_sc_unreach);
755 done:
756 	SCH_UNLOCK(sch);
757 }
758 
759 /*
760  * Build a new TCP socket structure from a syncache entry.
761  *
762  * On success return the newly created socket with its underlying inp locked.
763  */
764 static struct socket *
765 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
766 {
767 	struct tcp_function_block *blk;
768 	struct inpcb *inp = NULL;
769 	struct socket *so;
770 	struct tcpcb *tp;
771 	int error;
772 	char *s;
773 
774 	NET_EPOCH_ASSERT();
775 
776 	/*
777 	 * Ok, create the full blown connection, and set things up
778 	 * as they would have been set up if we had created the
779 	 * connection when the SYN arrived.  If we can't create
780 	 * the connection, abort it.
781 	 */
782 	so = sonewconn(lso, 0);
783 	if (so == NULL) {
784 		/*
785 		 * Drop the connection; we will either send a RST or
786 		 * have the peer retransmit its SYN again after its
787 		 * RTO and try again.
788 		 */
789 		TCPSTAT_INC(tcps_listendrop);
790 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
791 			log(LOG_DEBUG, "%s; %s: Socket create failed "
792 			    "due to limits or memory shortage\n",
793 			    s, __func__);
794 			free(s, M_TCPLOG);
795 		}
796 		goto abort2;
797 	}
798 #ifdef MAC
799 	mac_socketpeer_set_from_mbuf(m, so);
800 #endif
801 
802 	inp = sotoinpcb(so);
803 	inp->inp_inc.inc_fibnum = so->so_fibnum;
804 	INP_WLOCK(inp);
805 	/*
806 	 * Exclusive pcbinfo lock is not required in syncache socket case even
807 	 * if two inpcb locks can be acquired simultaneously:
808 	 *  - the inpcb in LISTEN state,
809 	 *  - the newly created inp.
810 	 *
811 	 * In this case, an inp cannot be at same time in LISTEN state and
812 	 * just created by an accept() call.
813 	 */
814 	INP_HASH_WLOCK(&V_tcbinfo);
815 
816 	/* Insert new socket into PCB hash list. */
817 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
818 #ifdef INET6
819 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
820 		inp->inp_vflag &= ~INP_IPV4;
821 		inp->inp_vflag |= INP_IPV6;
822 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
823 	} else {
824 		inp->inp_vflag &= ~INP_IPV6;
825 		inp->inp_vflag |= INP_IPV4;
826 #endif
827 		inp->inp_laddr = sc->sc_inc.inc_laddr;
828 #ifdef INET6
829 	}
830 #endif
831 
832 	/*
833 	 * If there's an mbuf and it has a flowid, then let's initialise the
834 	 * inp with that particular flowid.
835 	 */
836 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
837 		inp->inp_flowid = m->m_pkthdr.flowid;
838 		inp->inp_flowtype = M_HASHTYPE_GET(m);
839 #ifdef NUMA
840 		inp->inp_numa_domain = m->m_pkthdr.numa_domain;
841 #endif
842 	}
843 
844 	inp->inp_lport = sc->sc_inc.inc_lport;
845 #ifdef INET6
846 	if (inp->inp_vflag & INP_IPV6PROTO) {
847 		struct inpcb *oinp = sotoinpcb(lso);
848 
849 		/*
850 		 * Inherit socket options from the listening socket.
851 		 * Note that in6p_inputopts are not (and should not be)
852 		 * copied, since it stores previously received options and is
853 		 * used to detect if each new option is different than the
854 		 * previous one and hence should be passed to a user.
855 		 * If we copied in6p_inputopts, a user would not be able to
856 		 * receive options just after calling the accept system call.
857 		 */
858 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
859 		if (oinp->in6p_outputopts)
860 			inp->in6p_outputopts =
861 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
862 	}
863 
864 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
865 		struct in6_addr laddr6;
866 		struct sockaddr_in6 sin6;
867 
868 		sin6.sin6_family = AF_INET6;
869 		sin6.sin6_len = sizeof(sin6);
870 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
871 		sin6.sin6_port = sc->sc_inc.inc_fport;
872 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
873 		laddr6 = inp->in6p_laddr;
874 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
875 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
876 		if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
877 		    thread0.td_ucred, m, false)) != 0) {
878 			inp->in6p_laddr = laddr6;
879 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
880 				log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
881 				    "with error %i\n",
882 				    s, __func__, error);
883 				free(s, M_TCPLOG);
884 			}
885 			INP_HASH_WUNLOCK(&V_tcbinfo);
886 			goto abort;
887 		}
888 		/* Override flowlabel from in6_pcbconnect. */
889 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
890 		inp->inp_flow |= sc->sc_flowlabel;
891 	}
892 #endif /* INET6 */
893 #if defined(INET) && defined(INET6)
894 	else
895 #endif
896 #ifdef INET
897 	{
898 		struct in_addr laddr;
899 		struct sockaddr_in sin;
900 
901 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
902 
903 		if (inp->inp_options == NULL) {
904 			inp->inp_options = sc->sc_ipopts;
905 			sc->sc_ipopts = NULL;
906 		}
907 
908 		sin.sin_family = AF_INET;
909 		sin.sin_len = sizeof(sin);
910 		sin.sin_addr = sc->sc_inc.inc_faddr;
911 		sin.sin_port = sc->sc_inc.inc_fport;
912 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
913 		laddr = inp->inp_laddr;
914 		if (inp->inp_laddr.s_addr == INADDR_ANY)
915 			inp->inp_laddr = sc->sc_inc.inc_laddr;
916 		if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
917 		    thread0.td_ucred, m, false)) != 0) {
918 			inp->inp_laddr = laddr;
919 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
920 				log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
921 				    "with error %i\n",
922 				    s, __func__, error);
923 				free(s, M_TCPLOG);
924 			}
925 			INP_HASH_WUNLOCK(&V_tcbinfo);
926 			goto abort;
927 		}
928 	}
929 #endif /* INET */
930 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
931 	/* Copy old policy into new socket's. */
932 	if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
933 		printf("syncache_socket: could not copy policy\n");
934 #endif
935 	INP_HASH_WUNLOCK(&V_tcbinfo);
936 	tp = intotcpcb(inp);
937 	tcp_state_change(tp, TCPS_SYN_RECEIVED);
938 	tp->iss = sc->sc_iss;
939 	tp->irs = sc->sc_irs;
940 	tcp_rcvseqinit(tp);
941 	tcp_sendseqinit(tp);
942 	blk = sototcpcb(lso)->t_fb;
943 	if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
944 		/*
945 		 * Our parents t_fb was not the default,
946 		 * we need to release our ref on tp->t_fb and
947 		 * pickup one on the new entry.
948 		 */
949 		struct tcp_function_block *rblk;
950 
951 		rblk = find_and_ref_tcp_fb(blk);
952 		KASSERT(rblk != NULL,
953 		    ("cannot find blk %p out of syncache?", blk));
954 		if (tp->t_fb->tfb_tcp_fb_fini)
955 			(*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
956 		refcount_release(&tp->t_fb->tfb_refcnt);
957 		tp->t_fb = rblk;
958 		/*
959 		 * XXXrrs this is quite dangerous, it is possible
960 		 * for the new function to fail to init. We also
961 		 * are not asking if the handoff_is_ok though at
962 		 * the very start thats probalbly ok.
963 		 */
964 		if (tp->t_fb->tfb_tcp_fb_init) {
965 			(*tp->t_fb->tfb_tcp_fb_init)(tp);
966 		}
967 	}
968 	tp->snd_wl1 = sc->sc_irs;
969 	tp->snd_max = tp->iss + 1;
970 	tp->snd_nxt = tp->iss + 1;
971 	tp->rcv_up = sc->sc_irs + 1;
972 	tp->rcv_wnd = sc->sc_wnd;
973 	tp->rcv_adv += tp->rcv_wnd;
974 	tp->last_ack_sent = tp->rcv_nxt;
975 
976 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
977 	if (sc->sc_flags & SCF_NOOPT)
978 		tp->t_flags |= TF_NOOPT;
979 	else {
980 		if (sc->sc_flags & SCF_WINSCALE) {
981 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
982 			tp->snd_scale = sc->sc_requested_s_scale;
983 			tp->request_r_scale = sc->sc_requested_r_scale;
984 		}
985 		if (sc->sc_flags & SCF_TIMESTAMP) {
986 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
987 			tp->ts_recent = sc->sc_tsreflect;
988 			tp->ts_recent_age = tcp_ts_getticks();
989 			tp->ts_offset = sc->sc_tsoff;
990 		}
991 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
992 		if (sc->sc_flags & SCF_SIGNATURE)
993 			tp->t_flags |= TF_SIGNATURE;
994 #endif
995 		if (sc->sc_flags & SCF_SACK)
996 			tp->t_flags |= TF_SACK_PERMIT;
997 	}
998 
999 	if (sc->sc_flags & SCF_ECN)
1000 		tp->t_flags2 |= TF2_ECN_PERMIT;
1001 
1002 	/*
1003 	 * Set up MSS and get cached values from tcp_hostcache.
1004 	 * This might overwrite some of the defaults we just set.
1005 	 */
1006 	tcp_mss(tp, sc->sc_peer_mss);
1007 
1008 	/*
1009 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
1010 	 * limited to one segment in cc_conn_init().
1011 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
1012 	 */
1013 	if (sc->sc_rxmits > 1)
1014 		tp->snd_cwnd = 1;
1015 
1016 #ifdef TCP_OFFLOAD
1017 	/*
1018 	 * Allow a TOE driver to install its hooks.  Note that we hold the
1019 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
1020 	 * new connection before the TOE driver has done its thing.
1021 	 */
1022 	if (ADDED_BY_TOE(sc)) {
1023 		struct toedev *tod = sc->sc_tod;
1024 
1025 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
1026 	}
1027 #endif
1028 	/*
1029 	 * Copy and activate timers.
1030 	 */
1031 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
1032 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
1033 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
1034 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1035 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1036 
1037 	TCPSTAT_INC(tcps_accepts);
1038 	return (so);
1039 
1040 abort:
1041 	INP_WUNLOCK(inp);
1042 abort2:
1043 	if (so != NULL)
1044 		soabort(so);
1045 	return (NULL);
1046 }
1047 
1048 /*
1049  * This function gets called when we receive an ACK for a
1050  * socket in the LISTEN state.  We look up the connection
1051  * in the syncache, and if its there, we pull it out of
1052  * the cache and turn it into a full-blown connection in
1053  * the SYN-RECEIVED state.
1054  *
1055  * On syncache_socket() success the newly created socket
1056  * has its underlying inp locked.
1057  */
1058 int
1059 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1060     struct socket **lsop, struct mbuf *m)
1061 {
1062 	struct syncache *sc;
1063 	struct syncache_head *sch;
1064 	struct syncache scs;
1065 	char *s;
1066 	bool locked;
1067 
1068 	NET_EPOCH_ASSERT();
1069 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1070 	    ("%s: can handle only ACK", __func__));
1071 
1072 	if (syncache_cookiesonly()) {
1073 		sc = NULL;
1074 		sch = syncache_hashbucket(inc);
1075 		locked = false;
1076 	} else {
1077 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1078 		locked = true;
1079 		SCH_LOCK_ASSERT(sch);
1080 	}
1081 
1082 #ifdef INVARIANTS
1083 	/*
1084 	 * Test code for syncookies comparing the syncache stored
1085 	 * values with the reconstructed values from the cookie.
1086 	 */
1087 	if (sc != NULL)
1088 		syncookie_cmp(inc, sch, sc, th, to, *lsop);
1089 #endif
1090 
1091 	if (sc == NULL) {
1092 		/*
1093 		 * There is no syncache entry, so see if this ACK is
1094 		 * a returning syncookie.  To do this, first:
1095 		 *  A. Check if syncookies are used in case of syncache
1096 		 *     overflows
1097 		 *  B. See if this socket has had a syncache entry dropped in
1098 		 *     the recent past. We don't want to accept a bogus
1099 		 *     syncookie if we've never received a SYN or accept it
1100 		 *     twice.
1101 		 *  C. check that the syncookie is valid.  If it is, then
1102 		 *     cobble up a fake syncache entry, and return.
1103 		 */
1104 		if (locked && !V_tcp_syncookies) {
1105 			SCH_UNLOCK(sch);
1106 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1107 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1108 				    "segment rejected (syncookies disabled)\n",
1109 				    s, __func__);
1110 			goto failed;
1111 		}
1112 		if (locked && !V_tcp_syncookiesonly &&
1113 		    sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1114 			SCH_UNLOCK(sch);
1115 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1116 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1117 				    "segment rejected (no syncache entry)\n",
1118 				    s, __func__);
1119 			goto failed;
1120 		}
1121 		bzero(&scs, sizeof(scs));
1122 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
1123 		if (locked)
1124 			SCH_UNLOCK(sch);
1125 		if (sc == NULL) {
1126 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1127 				log(LOG_DEBUG, "%s; %s: Segment failed "
1128 				    "SYNCOOKIE authentication, segment rejected "
1129 				    "(probably spoofed)\n", s, __func__);
1130 			goto failed;
1131 		}
1132 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1133 		/* If received ACK has MD5 signature, check it. */
1134 		if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1135 		    (!TCPMD5_ENABLED() ||
1136 		    TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1137 			/* Drop the ACK. */
1138 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1139 				log(LOG_DEBUG, "%s; %s: Segment rejected, "
1140 				    "MD5 signature doesn't match.\n",
1141 				    s, __func__);
1142 				free(s, M_TCPLOG);
1143 			}
1144 			TCPSTAT_INC(tcps_sig_err_sigopt);
1145 			return (-1); /* Do not send RST */
1146 		}
1147 #endif /* TCP_SIGNATURE */
1148 	} else {
1149 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1150 		/*
1151 		 * If listening socket requested TCP digests, check that
1152 		 * received ACK has signature and it is correct.
1153 		 * If not, drop the ACK and leave sc entry in th cache,
1154 		 * because SYN was received with correct signature.
1155 		 */
1156 		if (sc->sc_flags & SCF_SIGNATURE) {
1157 			if ((to->to_flags & TOF_SIGNATURE) == 0) {
1158 				/* No signature */
1159 				TCPSTAT_INC(tcps_sig_err_nosigopt);
1160 				SCH_UNLOCK(sch);
1161 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1162 					log(LOG_DEBUG, "%s; %s: Segment "
1163 					    "rejected, MD5 signature wasn't "
1164 					    "provided.\n", s, __func__);
1165 					free(s, M_TCPLOG);
1166 				}
1167 				return (-1); /* Do not send RST */
1168 			}
1169 			if (!TCPMD5_ENABLED() ||
1170 			    TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1171 				/* Doesn't match or no SA */
1172 				SCH_UNLOCK(sch);
1173 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1174 					log(LOG_DEBUG, "%s; %s: Segment "
1175 					    "rejected, MD5 signature doesn't "
1176 					    "match.\n", s, __func__);
1177 					free(s, M_TCPLOG);
1178 				}
1179 				return (-1); /* Do not send RST */
1180 			}
1181 		}
1182 #endif /* TCP_SIGNATURE */
1183 
1184 		/*
1185 		 * RFC 7323 PAWS: If we have a timestamp on this segment and
1186 		 * it's less than ts_recent, drop it.
1187 		 * XXXMT: RFC 7323 also requires to send an ACK.
1188 		 *        In tcp_input.c this is only done for TCP segments
1189 		 *        with user data, so be consistent here and just drop
1190 		 *        the segment.
1191 		 */
1192 		if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1193 		    TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1194 			SCH_UNLOCK(sch);
1195 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1196 				log(LOG_DEBUG,
1197 				    "%s; %s: SEG.TSval %u < TS.Recent %u, "
1198 				    "segment dropped\n", s, __func__,
1199 				    to->to_tsval, sc->sc_tsreflect);
1200 				free(s, M_TCPLOG);
1201 			}
1202 			return (-1);  /* Do not send RST */
1203 		}
1204 
1205 		/*
1206 		 * Pull out the entry to unlock the bucket row.
1207 		 *
1208 		 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1209 		 * tcp_state_change().  The tcpcb is not existent at this
1210 		 * moment.  A new one will be allocated via syncache_socket->
1211 		 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1212 		 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1213 		 */
1214 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1215 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1216 		sch->sch_length--;
1217 #ifdef TCP_OFFLOAD
1218 		if (ADDED_BY_TOE(sc)) {
1219 			struct toedev *tod = sc->sc_tod;
1220 
1221 			tod->tod_syncache_removed(tod, sc->sc_todctx);
1222 		}
1223 #endif
1224 		SCH_UNLOCK(sch);
1225 	}
1226 
1227 	/*
1228 	 * Segment validation:
1229 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1230 	 */
1231 	if (th->th_ack != sc->sc_iss + 1) {
1232 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1233 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1234 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1235 		goto failed;
1236 	}
1237 
1238 	/*
1239 	 * The SEQ must fall in the window starting at the received
1240 	 * initial receive sequence number + 1 (the SYN).
1241 	 */
1242 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1243 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1244 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1245 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1246 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1247 		goto failed;
1248 	}
1249 
1250 	/*
1251 	 * If timestamps were not negotiated during SYN/ACK they
1252 	 * must not appear on any segment during this session.
1253 	 */
1254 	if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1255 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1256 			log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1257 			    "segment rejected\n", s, __func__);
1258 		goto failed;
1259 	}
1260 
1261 	/*
1262 	 * If timestamps were negotiated during SYN/ACK they should
1263 	 * appear on every segment during this session.
1264 	 * XXXAO: This is only informal as there have been unverified
1265 	 * reports of non-compliants stacks.
1266 	 */
1267 	if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1268 		if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1269 			log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1270 			    "no action\n", s, __func__);
1271 			free(s, M_TCPLOG);
1272 			s = NULL;
1273 		}
1274 	}
1275 
1276 	*lsop = syncache_socket(sc, *lsop, m);
1277 
1278 	if (*lsop == NULL)
1279 		TCPSTAT_INC(tcps_sc_aborted);
1280 	else
1281 		TCPSTAT_INC(tcps_sc_completed);
1282 
1283 /* how do we find the inp for the new socket? */
1284 	if (sc != &scs)
1285 		syncache_free(sc);
1286 	return (1);
1287 failed:
1288 	if (sc != NULL && sc != &scs)
1289 		syncache_free(sc);
1290 	if (s != NULL)
1291 		free(s, M_TCPLOG);
1292 	*lsop = NULL;
1293 	return (0);
1294 }
1295 
1296 static void
1297 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
1298     uint64_t response_cookie)
1299 {
1300 	struct inpcb *inp;
1301 	struct tcpcb *tp;
1302 	unsigned int *pending_counter;
1303 
1304 	NET_EPOCH_ASSERT();
1305 
1306 	pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
1307 	*lsop = syncache_socket(sc, *lsop, m);
1308 	if (*lsop == NULL) {
1309 		TCPSTAT_INC(tcps_sc_aborted);
1310 		atomic_subtract_int(pending_counter, 1);
1311 	} else {
1312 		soisconnected(*lsop);
1313 		inp = sotoinpcb(*lsop);
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 		TCPSTAT_INC(tcps_sc_completed);
1321 	}
1322 }
1323 
1324 /*
1325  * Given a LISTEN socket and an inbound SYN request, add
1326  * this to the syn cache, and send back a segment:
1327  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1328  * to the source.
1329  *
1330  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1331  * Doing so would require that we hold onto the data and deliver it
1332  * to the application.  However, if we are the target of a SYN-flood
1333  * DoS attack, an attacker could send data which would eventually
1334  * consume all available buffer space if it were ACKed.  By not ACKing
1335  * the data, we avoid this DoS scenario.
1336  *
1337  * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1338  * cookie is processed and a new socket is created.  In this case, any data
1339  * accompanying the SYN will be queued to the socket by tcp_input() and will
1340  * be ACKed either when the application sends response data or the delayed
1341  * ACK timer expires, whichever comes first.
1342  */
1343 int
1344 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1345     struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1346     void *todctx, uint8_t iptos)
1347 {
1348 	struct tcpcb *tp;
1349 	struct socket *so;
1350 	struct syncache *sc = NULL;
1351 	struct syncache_head *sch;
1352 	struct mbuf *ipopts = NULL;
1353 	u_int ltflags;
1354 	int win, ip_ttl, ip_tos;
1355 	char *s;
1356 	int rv = 0;
1357 #ifdef INET6
1358 	int autoflowlabel = 0;
1359 #endif
1360 #ifdef MAC
1361 	struct label *maclabel;
1362 #endif
1363 	struct syncache scs;
1364 	struct ucred *cred;
1365 	uint64_t tfo_response_cookie;
1366 	unsigned int *tfo_pending = NULL;
1367 	int tfo_cookie_valid = 0;
1368 	int tfo_response_cookie_valid = 0;
1369 	bool locked;
1370 
1371 	INP_WLOCK_ASSERT(inp);			/* listen socket */
1372 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1373 	    ("%s: unexpected tcp flags", __func__));
1374 
1375 	/*
1376 	 * Combine all so/tp operations very early to drop the INP lock as
1377 	 * soon as possible.
1378 	 */
1379 	so = *lsop;
1380 	KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1381 	tp = sototcpcb(so);
1382 	cred = crhold(so->so_cred);
1383 
1384 #ifdef INET6
1385 	if ((inc->inc_flags & INC_ISIPV6) &&
1386 	    (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1387 		autoflowlabel = 1;
1388 #endif
1389 	ip_ttl = inp->inp_ip_ttl;
1390 	ip_tos = inp->inp_ip_tos;
1391 	win = so->sol_sbrcv_hiwat;
1392 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1393 
1394 	if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
1395 	    (tp->t_tfo_pending != NULL) &&
1396 	    (to->to_flags & TOF_FASTOPEN)) {
1397 		/*
1398 		 * Limit the number of pending TFO connections to
1399 		 * approximately half of the queue limit.  This prevents TFO
1400 		 * SYN floods from starving the service by filling the
1401 		 * listen queue with bogus TFO connections.
1402 		 */
1403 		if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1404 		    (so->sol_qlimit / 2)) {
1405 			int result;
1406 
1407 			result = tcp_fastopen_check_cookie(inc,
1408 			    to->to_tfo_cookie, to->to_tfo_len,
1409 			    &tfo_response_cookie);
1410 			tfo_cookie_valid = (result > 0);
1411 			tfo_response_cookie_valid = (result >= 0);
1412 		}
1413 
1414 		/*
1415 		 * Remember the TFO pending counter as it will have to be
1416 		 * decremented below if we don't make it to syncache_tfo_expand().
1417 		 */
1418 		tfo_pending = tp->t_tfo_pending;
1419 	}
1420 
1421 	/* By the time we drop the lock these should no longer be used. */
1422 	so = NULL;
1423 	tp = NULL;
1424 
1425 #ifdef MAC
1426 	if (mac_syncache_init(&maclabel) != 0) {
1427 		INP_WUNLOCK(inp);
1428 		goto done;
1429 	} else
1430 		mac_syncache_create(maclabel, inp);
1431 #endif
1432 	if (!tfo_cookie_valid)
1433 		INP_WUNLOCK(inp);
1434 
1435 	/*
1436 	 * Remember the IP options, if any.
1437 	 */
1438 #ifdef INET6
1439 	if (!(inc->inc_flags & INC_ISIPV6))
1440 #endif
1441 #ifdef INET
1442 		ipopts = (m) ? ip_srcroute(m) : NULL;
1443 #else
1444 		ipopts = NULL;
1445 #endif
1446 
1447 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1448 	/*
1449 	 * If listening socket requested TCP digests, check that received
1450 	 * SYN has signature and it is correct. If signature doesn't match
1451 	 * or TCP_SIGNATURE support isn't enabled, drop the packet.
1452 	 */
1453 	if (ltflags & TF_SIGNATURE) {
1454 		if ((to->to_flags & TOF_SIGNATURE) == 0) {
1455 			TCPSTAT_INC(tcps_sig_err_nosigopt);
1456 			goto done;
1457 		}
1458 		if (!TCPMD5_ENABLED() ||
1459 		    TCPMD5_INPUT(m, th, to->to_signature) != 0)
1460 			goto done;
1461 	}
1462 #endif	/* TCP_SIGNATURE */
1463 	/*
1464 	 * See if we already have an entry for this connection.
1465 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1466 	 *
1467 	 * XXX: should the syncache be re-initialized with the contents
1468 	 * of the new SYN here (which may have different options?)
1469 	 *
1470 	 * XXX: We do not check the sequence number to see if this is a
1471 	 * real retransmit or a new connection attempt.  The question is
1472 	 * how to handle such a case; either ignore it as spoofed, or
1473 	 * drop the current entry and create a new one?
1474 	 */
1475 	if (syncache_cookiesonly()) {
1476 		sc = NULL;
1477 		sch = syncache_hashbucket(inc);
1478 		locked = false;
1479 	} else {
1480 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1481 		locked = true;
1482 		SCH_LOCK_ASSERT(sch);
1483 	}
1484 	if (sc != NULL) {
1485 		if (tfo_cookie_valid)
1486 			INP_WUNLOCK(inp);
1487 		TCPSTAT_INC(tcps_sc_dupsyn);
1488 		if (ipopts) {
1489 			/*
1490 			 * If we were remembering a previous source route,
1491 			 * forget it and use the new one we've been given.
1492 			 */
1493 			if (sc->sc_ipopts)
1494 				(void) m_free(sc->sc_ipopts);
1495 			sc->sc_ipopts = ipopts;
1496 		}
1497 		/*
1498 		 * Update timestamp if present.
1499 		 */
1500 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1501 			sc->sc_tsreflect = to->to_tsval;
1502 		else
1503 			sc->sc_flags &= ~SCF_TIMESTAMP;
1504 #ifdef MAC
1505 		/*
1506 		 * Since we have already unconditionally allocated label
1507 		 * storage, free it up.  The syncache entry will already
1508 		 * have an initialized label we can use.
1509 		 */
1510 		mac_syncache_destroy(&maclabel);
1511 #endif
1512 		TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1513 		/* Retransmit SYN|ACK and reset retransmit count. */
1514 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1515 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1516 			    "resetting timer and retransmitting SYN|ACK\n",
1517 			    s, __func__);
1518 			free(s, M_TCPLOG);
1519 		}
1520 		if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1521 			sc->sc_rxmits = 0;
1522 			syncache_timeout(sc, sch, 1);
1523 			TCPSTAT_INC(tcps_sndacks);
1524 			TCPSTAT_INC(tcps_sndtotal);
1525 		}
1526 		SCH_UNLOCK(sch);
1527 		goto donenoprobe;
1528 	}
1529 
1530 	if (tfo_cookie_valid) {
1531 		bzero(&scs, sizeof(scs));
1532 		sc = &scs;
1533 		goto skip_alloc;
1534 	}
1535 
1536 	/*
1537 	 * Skip allocating a syncache entry if we are just going to discard
1538 	 * it later.
1539 	 */
1540 	if (!locked) {
1541 		bzero(&scs, sizeof(scs));
1542 		sc = &scs;
1543 	} else
1544 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1545 	if (sc == NULL) {
1546 		/*
1547 		 * The zone allocator couldn't provide more entries.
1548 		 * Treat this as if the cache was full; drop the oldest
1549 		 * entry and insert the new one.
1550 		 */
1551 		TCPSTAT_INC(tcps_sc_zonefail);
1552 		if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1553 			sch->sch_last_overflow = time_uptime;
1554 			syncache_drop(sc, sch);
1555 			syncache_pause(inc);
1556 		}
1557 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1558 		if (sc == NULL) {
1559 			if (V_tcp_syncookies) {
1560 				bzero(&scs, sizeof(scs));
1561 				sc = &scs;
1562 			} else {
1563 				KASSERT(locked,
1564 				    ("%s: bucket unexpectedly unlocked",
1565 				    __func__));
1566 				SCH_UNLOCK(sch);
1567 				if (ipopts)
1568 					(void) m_free(ipopts);
1569 				goto done;
1570 			}
1571 		}
1572 	}
1573 
1574 skip_alloc:
1575 	if (!tfo_cookie_valid && tfo_response_cookie_valid)
1576 		sc->sc_tfo_cookie = &tfo_response_cookie;
1577 
1578 	/*
1579 	 * Fill in the syncache values.
1580 	 */
1581 #ifdef MAC
1582 	sc->sc_label = maclabel;
1583 #endif
1584 	sc->sc_cred = cred;
1585 	cred = NULL;
1586 	sc->sc_ipopts = ipopts;
1587 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1588 #ifdef INET6
1589 	if (!(inc->inc_flags & INC_ISIPV6))
1590 #endif
1591 	{
1592 		sc->sc_ip_tos = ip_tos;
1593 		sc->sc_ip_ttl = ip_ttl;
1594 	}
1595 #ifdef TCP_OFFLOAD
1596 	sc->sc_tod = tod;
1597 	sc->sc_todctx = todctx;
1598 #endif
1599 	sc->sc_irs = th->th_seq;
1600 	sc->sc_flags = 0;
1601 	sc->sc_flowlabel = 0;
1602 
1603 	/*
1604 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1605 	 * win was derived from socket earlier in the function.
1606 	 */
1607 	win = imax(win, 0);
1608 	win = imin(win, TCP_MAXWIN);
1609 	sc->sc_wnd = win;
1610 
1611 	if (V_tcp_do_rfc1323) {
1612 		/*
1613 		 * A timestamp received in a SYN makes
1614 		 * it ok to send timestamp requests and replies.
1615 		 */
1616 		if (to->to_flags & TOF_TS) {
1617 			sc->sc_tsreflect = to->to_tsval;
1618 			sc->sc_flags |= SCF_TIMESTAMP;
1619 			sc->sc_tsoff = tcp_new_ts_offset(inc);
1620 		}
1621 		if (to->to_flags & TOF_SCALE) {
1622 			int wscale = 0;
1623 
1624 			/*
1625 			 * Pick the smallest possible scaling factor that
1626 			 * will still allow us to scale up to sb_max, aka
1627 			 * kern.ipc.maxsockbuf.
1628 			 *
1629 			 * We do this because there are broken firewalls that
1630 			 * will corrupt the window scale option, leading to
1631 			 * the other endpoint believing that our advertised
1632 			 * window is unscaled.  At scale factors larger than
1633 			 * 5 the unscaled window will drop below 1500 bytes,
1634 			 * leading to serious problems when traversing these
1635 			 * broken firewalls.
1636 			 *
1637 			 * With the default maxsockbuf of 256K, a scale factor
1638 			 * of 3 will be chosen by this algorithm.  Those who
1639 			 * choose a larger maxsockbuf should watch out
1640 			 * for the compatibility problems mentioned above.
1641 			 *
1642 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1643 			 * or <SYN,ACK>) segment itself is never scaled.
1644 			 */
1645 			while (wscale < TCP_MAX_WINSHIFT &&
1646 			    (TCP_MAXWIN << wscale) < sb_max)
1647 				wscale++;
1648 			sc->sc_requested_r_scale = wscale;
1649 			sc->sc_requested_s_scale = to->to_wscale;
1650 			sc->sc_flags |= SCF_WINSCALE;
1651 		}
1652 	}
1653 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1654 	/*
1655 	 * If listening socket requested TCP digests, flag this in the
1656 	 * syncache so that syncache_respond() will do the right thing
1657 	 * with the SYN+ACK.
1658 	 */
1659 	if (ltflags & TF_SIGNATURE)
1660 		sc->sc_flags |= SCF_SIGNATURE;
1661 #endif	/* TCP_SIGNATURE */
1662 	if (to->to_flags & TOF_SACKPERM)
1663 		sc->sc_flags |= SCF_SACK;
1664 	if (to->to_flags & TOF_MSS)
1665 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1666 	if (ltflags & TF_NOOPT)
1667 		sc->sc_flags |= SCF_NOOPT;
1668 	if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1669 		sc->sc_flags |= SCF_ECN;
1670 
1671 	if (V_tcp_syncookies)
1672 		sc->sc_iss = syncookie_generate(sch, sc);
1673 	else
1674 		sc->sc_iss = arc4random();
1675 #ifdef INET6
1676 	if (autoflowlabel) {
1677 		if (V_tcp_syncookies)
1678 			sc->sc_flowlabel = sc->sc_iss;
1679 		else
1680 			sc->sc_flowlabel = ip6_randomflowlabel();
1681 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1682 	}
1683 #endif
1684 	if (locked)
1685 		SCH_UNLOCK(sch);
1686 
1687 	if (tfo_cookie_valid) {
1688 		syncache_tfo_expand(sc, lsop, m, tfo_response_cookie);
1689 		/* INP_WUNLOCK(inp) will be performed by the caller */
1690 		rv = 1;
1691 		goto tfo_expanded;
1692 	}
1693 
1694 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1695 	/*
1696 	 * Do a standard 3-way handshake.
1697 	 */
1698 	if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1699 		if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1700 			syncache_free(sc);
1701 		else if (sc != &scs)
1702 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1703 		TCPSTAT_INC(tcps_sndacks);
1704 		TCPSTAT_INC(tcps_sndtotal);
1705 	} else {
1706 		if (sc != &scs)
1707 			syncache_free(sc);
1708 		TCPSTAT_INC(tcps_sc_dropped);
1709 	}
1710 	goto donenoprobe;
1711 
1712 done:
1713 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1714 donenoprobe:
1715 	if (m) {
1716 		*lsop = NULL;
1717 		m_freem(m);
1718 	}
1719 	/*
1720 	 * If tfo_pending is not NULL here, then a TFO SYN that did not
1721 	 * result in a new socket was processed and the associated pending
1722 	 * counter has not yet been decremented.  All such TFO processing paths
1723 	 * transit this point.
1724 	 */
1725 	if (tfo_pending != NULL)
1726 		tcp_fastopen_decrement_counter(tfo_pending);
1727 
1728 tfo_expanded:
1729 	if (cred != NULL)
1730 		crfree(cred);
1731 #ifdef MAC
1732 	if (sc == &scs)
1733 		mac_syncache_destroy(&maclabel);
1734 #endif
1735 	return (rv);
1736 }
1737 
1738 /*
1739  * Send SYN|ACK or ACK to the peer.  Either in response to a peer's segment,
1740  * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1741  */
1742 static int
1743 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1744 {
1745 	struct ip *ip = NULL;
1746 	struct mbuf *m;
1747 	struct tcphdr *th = NULL;
1748 	int optlen, error = 0;	/* Make compiler happy */
1749 	u_int16_t hlen, tlen, mssopt;
1750 	struct tcpopt to;
1751 #ifdef INET6
1752 	struct ip6_hdr *ip6 = NULL;
1753 #endif
1754 	hlen =
1755 #ifdef INET6
1756 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1757 #endif
1758 		sizeof(struct ip);
1759 	tlen = hlen + sizeof(struct tcphdr);
1760 
1761 	/* Determine MSS we advertize to other end of connection. */
1762 	mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss);
1763 
1764 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1765 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1766 	    ("syncache: mbuf too small"));
1767 
1768 	/* Create the IP+TCP header from scratch. */
1769 	m = m_gethdr(M_NOWAIT, MT_DATA);
1770 	if (m == NULL)
1771 		return (ENOBUFS);
1772 #ifdef MAC
1773 	mac_syncache_create_mbuf(sc->sc_label, m);
1774 #endif
1775 	m->m_data += max_linkhdr;
1776 	m->m_len = tlen;
1777 	m->m_pkthdr.len = tlen;
1778 	m->m_pkthdr.rcvif = NULL;
1779 
1780 #ifdef INET6
1781 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1782 		ip6 = mtod(m, struct ip6_hdr *);
1783 		ip6->ip6_vfc = IPV6_VERSION;
1784 		ip6->ip6_nxt = IPPROTO_TCP;
1785 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1786 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1787 		ip6->ip6_plen = htons(tlen - hlen);
1788 		/* ip6_hlim is set after checksum */
1789 		ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1790 		ip6->ip6_flow |= sc->sc_flowlabel;
1791 
1792 		th = (struct tcphdr *)(ip6 + 1);
1793 	}
1794 #endif
1795 #if defined(INET6) && defined(INET)
1796 	else
1797 #endif
1798 #ifdef INET
1799 	{
1800 		ip = mtod(m, struct ip *);
1801 		ip->ip_v = IPVERSION;
1802 		ip->ip_hl = sizeof(struct ip) >> 2;
1803 		ip->ip_len = htons(tlen);
1804 		ip->ip_id = 0;
1805 		ip->ip_off = 0;
1806 		ip->ip_sum = 0;
1807 		ip->ip_p = IPPROTO_TCP;
1808 		ip->ip_src = sc->sc_inc.inc_laddr;
1809 		ip->ip_dst = sc->sc_inc.inc_faddr;
1810 		ip->ip_ttl = sc->sc_ip_ttl;
1811 		ip->ip_tos = sc->sc_ip_tos;
1812 
1813 		/*
1814 		 * See if we should do MTU discovery.  Route lookups are
1815 		 * expensive, so we will only unset the DF bit if:
1816 		 *
1817 		 *	1) path_mtu_discovery is disabled
1818 		 *	2) the SCF_UNREACH flag has been set
1819 		 */
1820 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1821 		       ip->ip_off |= htons(IP_DF);
1822 
1823 		th = (struct tcphdr *)(ip + 1);
1824 	}
1825 #endif /* INET */
1826 	th->th_sport = sc->sc_inc.inc_lport;
1827 	th->th_dport = sc->sc_inc.inc_fport;
1828 
1829 	if (flags & TH_SYN)
1830 		th->th_seq = htonl(sc->sc_iss);
1831 	else
1832 		th->th_seq = htonl(sc->sc_iss + 1);
1833 	th->th_ack = htonl(sc->sc_irs + 1);
1834 	th->th_off = sizeof(struct tcphdr) >> 2;
1835 	th->th_x2 = 0;
1836 	th->th_flags = flags;
1837 	th->th_win = htons(sc->sc_wnd);
1838 	th->th_urp = 0;
1839 
1840 	if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) {
1841 		th->th_flags |= TH_ECE;
1842 		TCPSTAT_INC(tcps_ecn_shs);
1843 	}
1844 
1845 	/* Tack on the TCP options. */
1846 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1847 		to.to_flags = 0;
1848 
1849 		if (flags & TH_SYN) {
1850 			to.to_mss = mssopt;
1851 			to.to_flags = TOF_MSS;
1852 			if (sc->sc_flags & SCF_WINSCALE) {
1853 				to.to_wscale = sc->sc_requested_r_scale;
1854 				to.to_flags |= TOF_SCALE;
1855 			}
1856 			if (sc->sc_flags & SCF_SACK)
1857 				to.to_flags |= TOF_SACKPERM;
1858 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1859 			if (sc->sc_flags & SCF_SIGNATURE)
1860 				to.to_flags |= TOF_SIGNATURE;
1861 #endif
1862 			if (sc->sc_tfo_cookie) {
1863 				to.to_flags |= TOF_FASTOPEN;
1864 				to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1865 				to.to_tfo_cookie = sc->sc_tfo_cookie;
1866 				/* don't send cookie again when retransmitting response */
1867 				sc->sc_tfo_cookie = NULL;
1868 			}
1869 		}
1870 		if (sc->sc_flags & SCF_TIMESTAMP) {
1871 			to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1872 			to.to_tsecr = sc->sc_tsreflect;
1873 			to.to_flags |= TOF_TS;
1874 		}
1875 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1876 
1877 		/* Adjust headers by option size. */
1878 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1879 		m->m_len += optlen;
1880 		m->m_pkthdr.len += optlen;
1881 #ifdef INET6
1882 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1883 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1884 		else
1885 #endif
1886 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1887 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1888 		if (sc->sc_flags & SCF_SIGNATURE) {
1889 			KASSERT(to.to_flags & TOF_SIGNATURE,
1890 			    ("tcp_addoptions() didn't set tcp_signature"));
1891 
1892 			/* NOTE: to.to_signature is inside of mbuf */
1893 			if (!TCPMD5_ENABLED() ||
1894 			    TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1895 				m_freem(m);
1896 				return (EACCES);
1897 			}
1898 		}
1899 #endif
1900 	} else
1901 		optlen = 0;
1902 
1903 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
1904 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1905 	/*
1906 	 * If we have peer's SYN and it has a flowid, then let's assign it to
1907 	 * our SYN|ACK.  ip6_output() and ip_output() will not assign flowid
1908 	 * to SYN|ACK due to lack of inp here.
1909 	 */
1910 	if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1911 		m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1912 		M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1913 	}
1914 #ifdef INET6
1915 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1916 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1917 		th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1918 		    IPPROTO_TCP, 0);
1919 		ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1920 #ifdef TCP_OFFLOAD
1921 		if (ADDED_BY_TOE(sc)) {
1922 			struct toedev *tod = sc->sc_tod;
1923 
1924 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1925 
1926 			return (error);
1927 		}
1928 #endif
1929 		TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
1930 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1931 	}
1932 #endif
1933 #if defined(INET6) && defined(INET)
1934 	else
1935 #endif
1936 #ifdef INET
1937 	{
1938 		m->m_pkthdr.csum_flags = CSUM_TCP;
1939 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1940 		    htons(tlen + optlen - hlen + IPPROTO_TCP));
1941 #ifdef TCP_OFFLOAD
1942 		if (ADDED_BY_TOE(sc)) {
1943 			struct toedev *tod = sc->sc_tod;
1944 
1945 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1946 
1947 			return (error);
1948 		}
1949 #endif
1950 		TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
1951 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1952 	}
1953 #endif
1954 	return (error);
1955 }
1956 
1957 /*
1958  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1959  * that exceed the capacity of the syncache by avoiding the storage of any
1960  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
1961  * attacks where the attacker does not have access to our responses.
1962  *
1963  * Syncookies encode and include all necessary information about the
1964  * connection setup within the SYN|ACK that we send back.  That way we
1965  * can avoid keeping any local state until the ACK to our SYN|ACK returns
1966  * (if ever).  Normally the syncache and syncookies are running in parallel
1967  * with the latter taking over when the former is exhausted.  When matching
1968  * syncache entry is found the syncookie is ignored.
1969  *
1970  * The only reliable information persisting the 3WHS is our initial sequence
1971  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
1972  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1973  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
1974  * returns and signifies a legitimate connection if it matches the ACK.
1975  *
1976  * The available space of 32 bits to store the hash and to encode the SYN
1977  * option information is very tight and we should have at least 24 bits for
1978  * the MAC to keep the number of guesses by blind spoofing reasonably high.
1979  *
1980  * SYN option information we have to encode to fully restore a connection:
1981  * MSS: is imporant to chose an optimal segment size to avoid IP level
1982  *   fragmentation along the path.  The common MSS values can be encoded
1983  *   in a 3-bit table.  Uncommon values are captured by the next lower value
1984  *   in the table leading to a slight increase in packetization overhead.
1985  * WSCALE: is necessary to allow large windows to be used for high delay-
1986  *   bandwidth product links.  Not scaling the window when it was initially
1987  *   negotiated is bad for performance as lack of scaling further decreases
1988  *   the apparent available send window.  We only need to encode the WSCALE
1989  *   we received from the remote end.  Our end can be recalculated at any
1990  *   time.  The common WSCALE values can be encoded in a 3-bit table.
1991  *   Uncommon values are captured by the next lower value in the table
1992  *   making us under-estimate the available window size halving our
1993  *   theoretically possible maximum throughput for that connection.
1994  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
1995  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
1996  *   that are included in all segments on a connection.  We enable them when
1997  *   the ACK has them.
1998  *
1999  * Security of syncookies and attack vectors:
2000  *
2001  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2002  * together with the gloabl secret to make it unique per connection attempt.
2003  * Thus any change of any of those parameters results in a different MAC output
2004  * in an unpredictable way unless a collision is encountered.  24 bits of the
2005  * MAC are embedded into the ISS.
2006  *
2007  * To prevent replay attacks two rotating global secrets are updated with a
2008  * new random value every 15 seconds.  The life-time of a syncookie is thus
2009  * 15-30 seconds.
2010  *
2011  * Vector 1: Attacking the secret.  This requires finding a weakness in the
2012  * MAC itself or the way it is used here.  The attacker can do a chosen plain
2013  * text attack by varying and testing the all parameters under his control.
2014  * The strength depends on the size and randomness of the secret, and the
2015  * cryptographic security of the MAC function.  Due to the constant updating
2016  * of the secret the attacker has at most 29.999 seconds to find the secret
2017  * and launch spoofed connections.  After that he has to start all over again.
2018  *
2019  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
2020  * size an average of 4,823 attempts are required for a 50% chance of success
2021  * to spoof a single syncookie (birthday collision paradox).  However the
2022  * attacker is blind and doesn't know if one of his attempts succeeded unless
2023  * he has a side channel to interfere success from.  A single connection setup
2024  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2025  * This many attempts are required for each one blind spoofed connection.  For
2026  * every additional spoofed connection he has to launch another N attempts.
2027  * Thus for a sustained rate 100 spoofed connections per second approximately
2028  * 1,800,000 packets per second would have to be sent.
2029  *
2030  * NB: The MAC function should be fast so that it doesn't become a CPU
2031  * exhaustion attack vector itself.
2032  *
2033  * References:
2034  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2035  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2036  *   http://cr.yp.to/syncookies.html    (overview)
2037  *   http://cr.yp.to/syncookies/archive (details)
2038  *
2039  *
2040  * Schematic construction of a syncookie enabled Initial Sequence Number:
2041  *  0        1         2         3
2042  *  12345678901234567890123456789012
2043  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2044  *
2045  *  x 24 MAC (truncated)
2046  *  W  3 Send Window Scale index
2047  *  M  3 MSS index
2048  *  S  1 SACK permitted
2049  *  P  1 Odd/even secret
2050  */
2051 
2052 /*
2053  * Distribution and probability of certain MSS values.  Those in between are
2054  * rounded down to the next lower one.
2055  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2056  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
2057  */
2058 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2059 
2060 /*
2061  * Distribution and probability of certain WSCALE values.  We have to map the
2062  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2063  * bits based on prevalence of certain values.  Where we don't have an exact
2064  * match for are rounded down to the next lower one letting us under-estimate
2065  * the true available window.  At the moment this would happen only for the
2066  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2067  * and window size).  The absence of the WSCALE option (no scaling in either
2068  * direction) is encoded with index zero.
2069  * [WSCALE values histograms, Allman, 2012]
2070  *                            X 10 10 35  5  6 14 10%   by host
2071  *                            X 11  4  5  5 18 49  3%   by connections
2072  */
2073 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2074 
2075 /*
2076  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
2077  * and good cryptographic properties.
2078  */
2079 static uint32_t
2080 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2081     uint8_t *secbits, uintptr_t secmod)
2082 {
2083 	SIPHASH_CTX ctx;
2084 	uint32_t siphash[2];
2085 
2086 	SipHash24_Init(&ctx);
2087 	SipHash_SetKey(&ctx, secbits);
2088 	switch (inc->inc_flags & INC_ISIPV6) {
2089 #ifdef INET
2090 	case 0:
2091 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2092 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2093 		break;
2094 #endif
2095 #ifdef INET6
2096 	case INC_ISIPV6:
2097 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2098 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2099 		break;
2100 #endif
2101 	}
2102 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2103 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2104 	SipHash_Update(&ctx, &irs, sizeof(irs));
2105 	SipHash_Update(&ctx, &flags, sizeof(flags));
2106 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
2107 	SipHash_Final((u_int8_t *)&siphash, &ctx);
2108 
2109 	return (siphash[0] ^ siphash[1]);
2110 }
2111 
2112 static tcp_seq
2113 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2114 {
2115 	u_int i, secbit, wscale;
2116 	uint32_t iss, hash;
2117 	uint8_t *secbits;
2118 	union syncookie cookie;
2119 
2120 	cookie.cookie = 0;
2121 
2122 	/* Map our computed MSS into the 3-bit index. */
2123 	for (i = nitems(tcp_sc_msstab) - 1;
2124 	     tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2125 	     i--)
2126 		;
2127 	cookie.flags.mss_idx = i;
2128 
2129 	/*
2130 	 * Map the send window scale into the 3-bit index but only if
2131 	 * the wscale option was received.
2132 	 */
2133 	if (sc->sc_flags & SCF_WINSCALE) {
2134 		wscale = sc->sc_requested_s_scale;
2135 		for (i = nitems(tcp_sc_wstab) - 1;
2136 		    tcp_sc_wstab[i] > wscale && i > 0;
2137 		     i--)
2138 			;
2139 		cookie.flags.wscale_idx = i;
2140 	}
2141 
2142 	/* Can we do SACK? */
2143 	if (sc->sc_flags & SCF_SACK)
2144 		cookie.flags.sack_ok = 1;
2145 
2146 	/* Which of the two secrets to use. */
2147 	secbit = V_tcp_syncache.secret.oddeven & 0x1;
2148 	cookie.flags.odd_even = secbit;
2149 
2150 	secbits = V_tcp_syncache.secret.key[secbit];
2151 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2152 	    (uintptr_t)sch);
2153 
2154 	/*
2155 	 * Put the flags into the hash and XOR them to get better ISS number
2156 	 * variance.  This doesn't enhance the cryptographic strength and is
2157 	 * done to prevent the 8 cookie bits from showing up directly on the
2158 	 * wire.
2159 	 */
2160 	iss = hash & ~0xff;
2161 	iss |= cookie.cookie ^ (hash >> 24);
2162 
2163 	TCPSTAT_INC(tcps_sc_sendcookie);
2164 	return (iss);
2165 }
2166 
2167 static struct syncache *
2168 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2169     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2170     struct socket *lso)
2171 {
2172 	uint32_t hash;
2173 	uint8_t *secbits;
2174 	tcp_seq ack, seq;
2175 	int wnd, wscale = 0;
2176 	union syncookie cookie;
2177 
2178 	/*
2179 	 * Pull information out of SYN-ACK/ACK and revert sequence number
2180 	 * advances.
2181 	 */
2182 	ack = th->th_ack - 1;
2183 	seq = th->th_seq - 1;
2184 
2185 	/*
2186 	 * Unpack the flags containing enough information to restore the
2187 	 * connection.
2188 	 */
2189 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2190 
2191 	/* Which of the two secrets to use. */
2192 	secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2193 
2194 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2195 
2196 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
2197 	if ((ack & ~0xff) != (hash & ~0xff))
2198 		return (NULL);
2199 
2200 	/* Fill in the syncache values. */
2201 	sc->sc_flags = 0;
2202 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2203 	sc->sc_ipopts = NULL;
2204 
2205 	sc->sc_irs = seq;
2206 	sc->sc_iss = ack;
2207 
2208 	switch (inc->inc_flags & INC_ISIPV6) {
2209 #ifdef INET
2210 	case 0:
2211 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2212 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2213 		break;
2214 #endif
2215 #ifdef INET6
2216 	case INC_ISIPV6:
2217 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2218 			sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
2219 		break;
2220 #endif
2221 	}
2222 
2223 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2224 
2225 	/* We can simply recompute receive window scale we sent earlier. */
2226 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2227 		wscale++;
2228 
2229 	/* Only use wscale if it was enabled in the orignal SYN. */
2230 	if (cookie.flags.wscale_idx > 0) {
2231 		sc->sc_requested_r_scale = wscale;
2232 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2233 		sc->sc_flags |= SCF_WINSCALE;
2234 	}
2235 
2236 	wnd = lso->sol_sbrcv_hiwat;
2237 	wnd = imax(wnd, 0);
2238 	wnd = imin(wnd, TCP_MAXWIN);
2239 	sc->sc_wnd = wnd;
2240 
2241 	if (cookie.flags.sack_ok)
2242 		sc->sc_flags |= SCF_SACK;
2243 
2244 	if (to->to_flags & TOF_TS) {
2245 		sc->sc_flags |= SCF_TIMESTAMP;
2246 		sc->sc_tsreflect = to->to_tsval;
2247 		sc->sc_tsoff = tcp_new_ts_offset(inc);
2248 	}
2249 
2250 	if (to->to_flags & TOF_SIGNATURE)
2251 		sc->sc_flags |= SCF_SIGNATURE;
2252 
2253 	sc->sc_rxmits = 0;
2254 
2255 	TCPSTAT_INC(tcps_sc_recvcookie);
2256 	return (sc);
2257 }
2258 
2259 #ifdef INVARIANTS
2260 static int
2261 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2262     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2263     struct socket *lso)
2264 {
2265 	struct syncache scs, *scx;
2266 	char *s;
2267 
2268 	bzero(&scs, sizeof(scs));
2269 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
2270 
2271 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2272 		return (0);
2273 
2274 	if (scx != NULL) {
2275 		if (sc->sc_peer_mss != scx->sc_peer_mss)
2276 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2277 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2278 
2279 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2280 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2281 			    s, __func__, sc->sc_requested_r_scale,
2282 			    scx->sc_requested_r_scale);
2283 
2284 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2285 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2286 			    s, __func__, sc->sc_requested_s_scale,
2287 			    scx->sc_requested_s_scale);
2288 
2289 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2290 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2291 	}
2292 
2293 	if (s != NULL)
2294 		free(s, M_TCPLOG);
2295 	return (0);
2296 }
2297 #endif /* INVARIANTS */
2298 
2299 static void
2300 syncookie_reseed(void *arg)
2301 {
2302 	struct tcp_syncache *sc = arg;
2303 	uint8_t *secbits;
2304 	int secbit;
2305 
2306 	/*
2307 	 * Reseeding the secret doesn't have to be protected by a lock.
2308 	 * It only must be ensured that the new random values are visible
2309 	 * to all CPUs in a SMP environment.  The atomic with release
2310 	 * semantics ensures that.
2311 	 */
2312 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2313 	secbits = sc->secret.key[secbit];
2314 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2315 	atomic_add_rel_int(&sc->secret.oddeven, 1);
2316 
2317 	/* Reschedule ourself. */
2318 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2319 }
2320 
2321 /*
2322  * We have overflowed a bucket. Let's pause dealing with the syncache.
2323  * This function will increment the bucketoverflow statistics appropriately
2324  * (once per pause when pausing is enabled; otherwise, once per overflow).
2325  */
2326 static void
2327 syncache_pause(struct in_conninfo *inc)
2328 {
2329 	time_t delta;
2330 	const char *s;
2331 
2332 	/* XXX:
2333 	 * 2. Add sysctl read here so we don't get the benefit of this
2334 	 * change without the new sysctl.
2335 	 */
2336 
2337 	/*
2338 	 * Try an unlocked read. If we already know that another thread
2339 	 * has activated the feature, there is no need to proceed.
2340 	 */
2341 	if (V_tcp_syncache.paused)
2342 		return;
2343 
2344 	/* Are cookied enabled? If not, we can't pause. */
2345 	if (!V_tcp_syncookies) {
2346 		TCPSTAT_INC(tcps_sc_bucketoverflow);
2347 		return;
2348 	}
2349 
2350 	/*
2351 	 * We may be the first thread to find an overflow. Get the lock
2352 	 * and evaluate if we need to take action.
2353 	 */
2354 	mtx_lock(&V_tcp_syncache.pause_mtx);
2355 	if (V_tcp_syncache.paused) {
2356 		mtx_unlock(&V_tcp_syncache.pause_mtx);
2357 		return;
2358 	}
2359 
2360 	/* Activate protection. */
2361 	V_tcp_syncache.paused = true;
2362 	TCPSTAT_INC(tcps_sc_bucketoverflow);
2363 
2364 	/*
2365 	 * Determine the last backoff time. If we are seeing a re-newed
2366 	 * attack within that same time after last reactivating the syncache,
2367 	 * consider it an extension of the same attack.
2368 	 */
2369 	delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2370 	if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2371 		if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2372 			delta <<= 1;
2373 			V_tcp_syncache.pause_backoff++;
2374 		}
2375 	} else {
2376 		delta = TCP_SYNCACHE_PAUSE_TIME;
2377 		V_tcp_syncache.pause_backoff = 0;
2378 	}
2379 
2380 	/* Log a warning, including IP addresses, if able. */
2381 	if (inc != NULL)
2382 		s = tcp_log_addrs(inc, NULL, NULL, NULL);
2383 	else
2384 		s = (const char *)NULL;
2385 	log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2386 	    "the next %lld seconds%s%s%s\n", (long long)delta,
2387 	    (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2388 	    (s != NULL) ? ")" : "");
2389 	free(__DECONST(void *, s), M_TCPLOG);
2390 
2391 	/* Use the calculated delta to set a new pause time. */
2392 	V_tcp_syncache.pause_until = time_uptime + delta;
2393 	callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2394 	    &V_tcp_syncache);
2395 	mtx_unlock(&V_tcp_syncache.pause_mtx);
2396 }
2397 
2398 /* Evaluate whether we need to unpause. */
2399 static void
2400 syncache_unpause(void *arg)
2401 {
2402 	struct tcp_syncache *sc;
2403 	time_t delta;
2404 
2405 	sc = arg;
2406 	mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2407 	callout_deactivate(&sc->pause_co);
2408 
2409 	/*
2410 	 * Check to make sure we are not running early. If the pause
2411 	 * time has expired, then deactivate the protection.
2412 	 */
2413 	if ((delta = sc->pause_until - time_uptime) > 0)
2414 		callout_schedule(&sc->pause_co, delta * hz);
2415 	else
2416 		sc->paused = false;
2417 }
2418 
2419 /*
2420  * Exports the syncache entries to userland so that netstat can display
2421  * them alongside the other sockets.  This function is intended to be
2422  * called only from tcp_pcblist.
2423  *
2424  * Due to concurrency on an active system, the number of pcbs exported
2425  * may have no relation to max_pcbs.  max_pcbs merely indicates the
2426  * amount of space the caller allocated for this function to use.
2427  */
2428 int
2429 syncache_pcblist(struct sysctl_req *req)
2430 {
2431 	struct xtcpcb xt;
2432 	struct syncache *sc;
2433 	struct syncache_head *sch;
2434 	int error, i;
2435 
2436 	bzero(&xt, sizeof(xt));
2437 	xt.xt_len = sizeof(xt);
2438 	xt.t_state = TCPS_SYN_RECEIVED;
2439 	xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2440 	xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2441 	xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2442 	xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2443 
2444 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2445 		sch = &V_tcp_syncache.hashbase[i];
2446 		SCH_LOCK(sch);
2447 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2448 			if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2449 				continue;
2450 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2451 				xt.xt_inp.inp_vflag = INP_IPV6;
2452 			else
2453 				xt.xt_inp.inp_vflag = INP_IPV4;
2454 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2455 			    sizeof (struct in_conninfo));
2456 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2457 			if (error) {
2458 				SCH_UNLOCK(sch);
2459 				return (0);
2460 			}
2461 		}
2462 		SCH_UNLOCK(sch);
2463 	}
2464 
2465 	return (0);
2466 }
2467