xref: /freebsd/sys/netinet/tcp_syncache.c (revision 1bea15e601ad4fae5c2912758715fcaa08aa2404)
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,
176     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
177     "TCP SYN cache");
178 
179 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
180     &VNET_NAME(tcp_syncache.bucket_limit), 0,
181     "Per-bucket hash limit for syncache");
182 
183 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
184     &VNET_NAME(tcp_syncache.cache_limit), 0,
185     "Overall entry limit for syncache");
186 
187 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
188     &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
189 
190 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
191     &VNET_NAME(tcp_syncache.hashsize), 0,
192     "Size of TCP syncache hashtable");
193 
194 static int
195 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
196 {
197 	int error;
198 	u_int new;
199 
200 	new = V_tcp_syncache.rexmt_limit;
201 	error = sysctl_handle_int(oidp, &new, 0, req);
202 	if ((error == 0) && (req->newptr != NULL)) {
203 		if (new > TCP_MAXRXTSHIFT)
204 			error = EINVAL;
205 		else
206 			V_tcp_syncache.rexmt_limit = new;
207 	}
208 	return (error);
209 }
210 
211 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
212     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
213     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
214     sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
215     "Limit on SYN/ACK retransmissions");
216 
217 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
218 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
219     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
220     "Send reset on socket allocation failure");
221 
222 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
223 
224 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
225 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
226 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
227 
228 /*
229  * Requires the syncache entry to be already removed from the bucket list.
230  */
231 static void
232 syncache_free(struct syncache *sc)
233 {
234 
235 	if (sc->sc_ipopts)
236 		(void) m_free(sc->sc_ipopts);
237 	if (sc->sc_cred)
238 		crfree(sc->sc_cred);
239 #ifdef MAC
240 	mac_syncache_destroy(&sc->sc_label);
241 #endif
242 
243 	uma_zfree(V_tcp_syncache.zone, sc);
244 }
245 
246 void
247 syncache_init(void)
248 {
249 	int i;
250 
251 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
252 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
253 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
254 	V_tcp_syncache.hash_secret = arc4random();
255 
256 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
257 	    &V_tcp_syncache.hashsize);
258 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
259 	    &V_tcp_syncache.bucket_limit);
260 	if (!powerof2(V_tcp_syncache.hashsize) ||
261 	    V_tcp_syncache.hashsize == 0) {
262 		printf("WARNING: syncache hash size is not a power of 2.\n");
263 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
264 	}
265 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
266 
267 	/* Set limits. */
268 	V_tcp_syncache.cache_limit =
269 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
270 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
271 	    &V_tcp_syncache.cache_limit);
272 
273 	/* Allocate the hash table. */
274 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
275 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
276 
277 #ifdef VIMAGE
278 	V_tcp_syncache.vnet = curvnet;
279 #endif
280 
281 	/* Initialize the hash buckets. */
282 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
283 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
284 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
285 			 NULL, MTX_DEF);
286 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
287 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
288 		V_tcp_syncache.hashbase[i].sch_length = 0;
289 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
290 		V_tcp_syncache.hashbase[i].sch_last_overflow =
291 		    -(SYNCOOKIE_LIFETIME + 1);
292 	}
293 
294 	/* Create the syncache entry zone. */
295 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
296 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
297 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
298 	    V_tcp_syncache.cache_limit);
299 
300 	/* Start the SYN cookie reseeder callout. */
301 	callout_init(&V_tcp_syncache.secret.reseed, 1);
302 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
303 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
304 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
305 	    syncookie_reseed, &V_tcp_syncache);
306 
307 	/* Initialize the pause machinery. */
308 	mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF);
309 	callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx,
310 	    0);
311 	V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME;
312 	V_tcp_syncache.pause_backoff = 0;
313 	V_tcp_syncache.paused = false;
314 }
315 
316 #ifdef VIMAGE
317 void
318 syncache_destroy(void)
319 {
320 	struct syncache_head *sch;
321 	struct syncache *sc, *nsc;
322 	int i;
323 
324 	/*
325 	 * Stop the re-seed timer before freeing resources.  No need to
326 	 * possibly schedule it another time.
327 	 */
328 	callout_drain(&V_tcp_syncache.secret.reseed);
329 
330 	/* Stop the SYN cache pause callout. */
331 	mtx_lock(&V_tcp_syncache.pause_mtx);
332 	if (callout_stop(&V_tcp_syncache.pause_co) == 0) {
333 		mtx_unlock(&V_tcp_syncache.pause_mtx);
334 		callout_drain(&V_tcp_syncache.pause_co);
335 	} else
336 		mtx_unlock(&V_tcp_syncache.pause_mtx);
337 
338 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
339 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
340 
341 		sch = &V_tcp_syncache.hashbase[i];
342 		callout_drain(&sch->sch_timer);
343 
344 		SCH_LOCK(sch);
345 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
346 			syncache_drop(sc, sch);
347 		SCH_UNLOCK(sch);
348 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
349 		    ("%s: sch->sch_bucket not empty", __func__));
350 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
351 		    __func__, sch->sch_length));
352 		mtx_destroy(&sch->sch_mtx);
353 	}
354 
355 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
356 	    ("%s: cache_count not 0", __func__));
357 
358 	/* Free the allocated global resources. */
359 	uma_zdestroy(V_tcp_syncache.zone);
360 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
361 	mtx_destroy(&V_tcp_syncache.pause_mtx);
362 }
363 #endif
364 
365 /*
366  * Inserts a syncache entry into the specified bucket row.
367  * Locks and unlocks the syncache_head autonomously.
368  */
369 static void
370 syncache_insert(struct syncache *sc, struct syncache_head *sch)
371 {
372 	struct syncache *sc2;
373 
374 	SCH_LOCK(sch);
375 
376 	/*
377 	 * Make sure that we don't overflow the per-bucket limit.
378 	 * If the bucket is full, toss the oldest element.
379 	 */
380 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
381 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
382 			("sch->sch_length incorrect"));
383 		syncache_pause(&sc->sc_inc);
384 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
385 		sch->sch_last_overflow = time_uptime;
386 		syncache_drop(sc2, sch);
387 	}
388 
389 	/* Put it into the bucket. */
390 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
391 	sch->sch_length++;
392 
393 #ifdef TCP_OFFLOAD
394 	if (ADDED_BY_TOE(sc)) {
395 		struct toedev *tod = sc->sc_tod;
396 
397 		tod->tod_syncache_added(tod, sc->sc_todctx);
398 	}
399 #endif
400 
401 	/* Reinitialize the bucket row's timer. */
402 	if (sch->sch_length == 1)
403 		sch->sch_nextc = ticks + INT_MAX;
404 	syncache_timeout(sc, sch, 1);
405 
406 	SCH_UNLOCK(sch);
407 
408 	TCPSTATES_INC(TCPS_SYN_RECEIVED);
409 	TCPSTAT_INC(tcps_sc_added);
410 }
411 
412 /*
413  * Remove and free entry from syncache bucket row.
414  * Expects locked syncache head.
415  */
416 static void
417 syncache_drop(struct syncache *sc, struct syncache_head *sch)
418 {
419 
420 	SCH_LOCK_ASSERT(sch);
421 
422 	TCPSTATES_DEC(TCPS_SYN_RECEIVED);
423 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
424 	sch->sch_length--;
425 
426 #ifdef TCP_OFFLOAD
427 	if (ADDED_BY_TOE(sc)) {
428 		struct toedev *tod = sc->sc_tod;
429 
430 		tod->tod_syncache_removed(tod, sc->sc_todctx);
431 	}
432 #endif
433 
434 	syncache_free(sc);
435 }
436 
437 /*
438  * Engage/reengage time on bucket row.
439  */
440 static void
441 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
442 {
443 	int rexmt;
444 
445 	if (sc->sc_rxmits == 0)
446 		rexmt = tcp_rexmit_initial;
447 	else
448 		TCPT_RANGESET(rexmt,
449 		    tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
450 		    tcp_rexmit_min, TCPTV_REXMTMAX);
451 	sc->sc_rxttime = ticks + rexmt;
452 	sc->sc_rxmits++;
453 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
454 		sch->sch_nextc = sc->sc_rxttime;
455 		if (docallout)
456 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
457 			    syncache_timer, (void *)sch);
458 	}
459 }
460 
461 /*
462  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
463  * If we have retransmitted an entry the maximum number of times, expire it.
464  * One separate timer for each bucket row.
465  */
466 static void
467 syncache_timer(void *xsch)
468 {
469 	struct syncache_head *sch = (struct syncache_head *)xsch;
470 	struct syncache *sc, *nsc;
471 	struct epoch_tracker et;
472 	int tick = ticks;
473 	char *s;
474 	bool paused;
475 
476 	CURVNET_SET(sch->sch_sc->vnet);
477 
478 	/* NB: syncache_head has already been locked by the callout. */
479 	SCH_LOCK_ASSERT(sch);
480 
481 	/*
482 	 * In the following cycle we may remove some entries and/or
483 	 * advance some timeouts, so re-initialize the bucket timer.
484 	 */
485 	sch->sch_nextc = tick + INT_MAX;
486 
487 	/*
488 	 * If we have paused processing, unconditionally remove
489 	 * all syncache entries.
490 	 */
491 	mtx_lock(&V_tcp_syncache.pause_mtx);
492 	paused = V_tcp_syncache.paused;
493 	mtx_unlock(&V_tcp_syncache.pause_mtx);
494 
495 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
496 		if (paused) {
497 			syncache_drop(sc, sch);
498 			continue;
499 		}
500 		/*
501 		 * We do not check if the listen socket still exists
502 		 * and accept the case where the listen socket may be
503 		 * gone by the time we resend the SYN/ACK.  We do
504 		 * not expect this to happens often. If it does,
505 		 * then the RST will be sent by the time the remote
506 		 * host does the SYN/ACK->ACK.
507 		 */
508 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
509 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
510 				sch->sch_nextc = sc->sc_rxttime;
511 			continue;
512 		}
513 		if (sc->sc_rxmits > V_tcp_ecn_maxretries) {
514 			sc->sc_flags &= ~SCF_ECN;
515 		}
516 		if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
517 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
518 				log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
519 				    "giving up and removing syncache entry\n",
520 				    s, __func__);
521 				free(s, M_TCPLOG);
522 			}
523 			syncache_drop(sc, sch);
524 			TCPSTAT_INC(tcps_sc_stale);
525 			continue;
526 		}
527 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
528 			log(LOG_DEBUG, "%s; %s: Response timeout, "
529 			    "retransmitting (%u) SYN|ACK\n",
530 			    s, __func__, sc->sc_rxmits);
531 			free(s, M_TCPLOG);
532 		}
533 
534 		NET_EPOCH_ENTER(et);
535 		syncache_respond(sc, NULL, TH_SYN|TH_ACK);
536 		NET_EPOCH_EXIT(et);
537 		TCPSTAT_INC(tcps_sc_retransmitted);
538 		syncache_timeout(sc, sch, 0);
539 	}
540 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
541 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
542 			syncache_timer, (void *)(sch));
543 	CURVNET_RESTORE();
544 }
545 
546 /*
547  * Returns true if the system is only using cookies at the moment.
548  * This could be due to a sysadmin decision to only use cookies, or it
549  * could be due to the system detecting an attack.
550  */
551 static inline bool
552 syncache_cookiesonly(void)
553 {
554 
555 	return (V_tcp_syncookies && (V_tcp_syncache.paused ||
556 	    V_tcp_syncookiesonly));
557 }
558 
559 /*
560  * Find the hash bucket for the given connection.
561  */
562 static struct syncache_head *
563 syncache_hashbucket(struct in_conninfo *inc)
564 {
565 	uint32_t hash;
566 
567 	/*
568 	 * The hash is built on foreign port + local port + foreign address.
569 	 * We rely on the fact that struct in_conninfo starts with 16 bits
570 	 * of foreign port, then 16 bits of local port then followed by 128
571 	 * bits of foreign address.  In case of IPv4 address, the first 3
572 	 * 32-bit words of the address always are zeroes.
573 	 */
574 	hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
575 	    V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
576 
577 	return (&V_tcp_syncache.hashbase[hash]);
578 }
579 
580 /*
581  * Find an entry in the syncache.
582  * Returns always with locked syncache_head plus a matching entry or NULL.
583  */
584 static struct syncache *
585 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
586 {
587 	struct syncache *sc;
588 	struct syncache_head *sch;
589 
590 	*schp = sch = syncache_hashbucket(inc);
591 	SCH_LOCK(sch);
592 
593 	/* Circle through bucket row to find matching entry. */
594 	TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
595 		if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
596 		    sizeof(struct in_endpoints)) == 0)
597 			break;
598 
599 	return (sc);	/* Always returns with locked sch. */
600 }
601 
602 /*
603  * This function is called when we get a RST for a
604  * non-existent connection, so that we can see if the
605  * connection is in the syn cache.  If it is, zap it.
606  * If required send a challenge ACK.
607  */
608 void
609 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m)
610 {
611 	struct syncache *sc;
612 	struct syncache_head *sch;
613 	char *s = NULL;
614 
615 	if (syncache_cookiesonly())
616 		return;
617 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
618 	SCH_LOCK_ASSERT(sch);
619 
620 	/*
621 	 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
622 	 * See RFC 793 page 65, section SEGMENT ARRIVES.
623 	 */
624 	if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
625 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
626 			log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
627 			    "FIN flag set, segment ignored\n", s, __func__);
628 		TCPSTAT_INC(tcps_badrst);
629 		goto done;
630 	}
631 
632 	/*
633 	 * No corresponding connection was found in syncache.
634 	 * If syncookies are enabled and possibly exclusively
635 	 * used, or we are under memory pressure, a valid RST
636 	 * may not find a syncache entry.  In that case we're
637 	 * done and no SYN|ACK retransmissions will happen.
638 	 * Otherwise the RST was misdirected or spoofed.
639 	 */
640 	if (sc == NULL) {
641 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
642 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
643 			    "syncache entry (possibly syncookie only), "
644 			    "segment ignored\n", s, __func__);
645 		TCPSTAT_INC(tcps_badrst);
646 		goto done;
647 	}
648 
649 	/*
650 	 * If the RST bit is set, check the sequence number to see
651 	 * if this is a valid reset segment.
652 	 *
653 	 * RFC 793 page 37:
654 	 *   In all states except SYN-SENT, all reset (RST) segments
655 	 *   are validated by checking their SEQ-fields.  A reset is
656 	 *   valid if its sequence number is in the window.
657 	 *
658 	 * RFC 793 page 69:
659 	 *   There are four cases for the acceptability test for an incoming
660 	 *   segment:
661 	 *
662 	 * Segment Receive  Test
663 	 * Length  Window
664 	 * ------- -------  -------------------------------------------
665 	 *    0       0     SEG.SEQ = RCV.NXT
666 	 *    0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
667 	 *   >0       0     not acceptable
668 	 *   >0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
669 	 *               or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
670 	 *
671 	 * Note that when receiving a SYN segment in the LISTEN state,
672 	 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
673 	 * described in RFC 793, page 66.
674 	 */
675 	if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
676 	    SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
677 	    (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
678 		if (V_tcp_insecure_rst ||
679 		    th->th_seq == sc->sc_irs + 1) {
680 			syncache_drop(sc, sch);
681 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
682 				log(LOG_DEBUG,
683 				    "%s; %s: Our SYN|ACK was rejected, "
684 				    "connection attempt aborted by remote "
685 				    "endpoint\n",
686 				    s, __func__);
687 			TCPSTAT_INC(tcps_sc_reset);
688 		} else {
689 			TCPSTAT_INC(tcps_badrst);
690 			/* Send challenge ACK. */
691 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
692 				log(LOG_DEBUG, "%s; %s: RST with invalid "
693 				    " SEQ %u != NXT %u (+WND %u), "
694 				    "sending challenge ACK\n",
695 				    s, __func__,
696 				    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
697 			syncache_respond(sc, m, TH_ACK);
698 		}
699 	} else {
700 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
701 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
702 			    "NXT %u (+WND %u), segment ignored\n",
703 			    s, __func__,
704 			    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
705 		TCPSTAT_INC(tcps_badrst);
706 	}
707 
708 done:
709 	if (s != NULL)
710 		free(s, M_TCPLOG);
711 	SCH_UNLOCK(sch);
712 }
713 
714 void
715 syncache_badack(struct in_conninfo *inc)
716 {
717 	struct syncache *sc;
718 	struct syncache_head *sch;
719 
720 	if (syncache_cookiesonly())
721 		return;
722 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
723 	SCH_LOCK_ASSERT(sch);
724 	if (sc != NULL) {
725 		syncache_drop(sc, sch);
726 		TCPSTAT_INC(tcps_sc_badack);
727 	}
728 	SCH_UNLOCK(sch);
729 }
730 
731 void
732 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq)
733 {
734 	struct syncache *sc;
735 	struct syncache_head *sch;
736 
737 	if (syncache_cookiesonly())
738 		return;
739 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
740 	SCH_LOCK_ASSERT(sch);
741 	if (sc == NULL)
742 		goto done;
743 
744 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
745 	if (ntohl(th_seq) != sc->sc_iss)
746 		goto done;
747 
748 	/*
749 	 * If we've rertransmitted 3 times and this is our second error,
750 	 * we remove the entry.  Otherwise, we allow it to continue on.
751 	 * This prevents us from incorrectly nuking an entry during a
752 	 * spurious network outage.
753 	 *
754 	 * See tcp_notify().
755 	 */
756 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
757 		sc->sc_flags |= SCF_UNREACH;
758 		goto done;
759 	}
760 	syncache_drop(sc, sch);
761 	TCPSTAT_INC(tcps_sc_unreach);
762 done:
763 	SCH_UNLOCK(sch);
764 }
765 
766 /*
767  * Build a new TCP socket structure from a syncache entry.
768  *
769  * On success return the newly created socket with its underlying inp locked.
770  */
771 static struct socket *
772 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
773 {
774 	struct tcp_function_block *blk;
775 	struct inpcb *inp = NULL;
776 	struct socket *so;
777 	struct tcpcb *tp;
778 	int error;
779 	char *s;
780 
781 	NET_EPOCH_ASSERT();
782 
783 	/*
784 	 * Ok, create the full blown connection, and set things up
785 	 * as they would have been set up if we had created the
786 	 * connection when the SYN arrived.  If we can't create
787 	 * the connection, abort it.
788 	 */
789 	so = sonewconn(lso, 0);
790 	if (so == NULL) {
791 		/*
792 		 * Drop the connection; we will either send a RST or
793 		 * have the peer retransmit its SYN again after its
794 		 * RTO and try again.
795 		 */
796 		TCPSTAT_INC(tcps_listendrop);
797 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
798 			log(LOG_DEBUG, "%s; %s: Socket create failed "
799 			    "due to limits or memory shortage\n",
800 			    s, __func__);
801 			free(s, M_TCPLOG);
802 		}
803 		goto abort2;
804 	}
805 #ifdef MAC
806 	mac_socketpeer_set_from_mbuf(m, so);
807 #endif
808 
809 	inp = sotoinpcb(so);
810 	inp->inp_inc.inc_fibnum = so->so_fibnum;
811 	INP_WLOCK(inp);
812 	/*
813 	 * Exclusive pcbinfo lock is not required in syncache socket case even
814 	 * if two inpcb locks can be acquired simultaneously:
815 	 *  - the inpcb in LISTEN state,
816 	 *  - the newly created inp.
817 	 *
818 	 * In this case, an inp cannot be at same time in LISTEN state and
819 	 * just created by an accept() call.
820 	 */
821 	INP_HASH_WLOCK(&V_tcbinfo);
822 
823 	/* Insert new socket into PCB hash list. */
824 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
825 #ifdef INET6
826 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
827 		inp->inp_vflag &= ~INP_IPV4;
828 		inp->inp_vflag |= INP_IPV6;
829 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
830 	} else {
831 		inp->inp_vflag &= ~INP_IPV6;
832 		inp->inp_vflag |= INP_IPV4;
833 #endif
834 		inp->inp_laddr = sc->sc_inc.inc_laddr;
835 #ifdef INET6
836 	}
837 #endif
838 
839 	/*
840 	 * If there's an mbuf and it has a flowid, then let's initialise the
841 	 * inp with that particular flowid.
842 	 */
843 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
844 		inp->inp_flowid = m->m_pkthdr.flowid;
845 		inp->inp_flowtype = M_HASHTYPE_GET(m);
846 #ifdef NUMA
847 		inp->inp_numa_domain = m->m_pkthdr.numa_domain;
848 #endif
849 	}
850 
851 	inp->inp_lport = sc->sc_inc.inc_lport;
852 #ifdef INET6
853 	if (inp->inp_vflag & INP_IPV6PROTO) {
854 		struct inpcb *oinp = sotoinpcb(lso);
855 
856 		/*
857 		 * Inherit socket options from the listening socket.
858 		 * Note that in6p_inputopts are not (and should not be)
859 		 * copied, since it stores previously received options and is
860 		 * used to detect if each new option is different than the
861 		 * previous one and hence should be passed to a user.
862 		 * If we copied in6p_inputopts, a user would not be able to
863 		 * receive options just after calling the accept system call.
864 		 */
865 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
866 		if (oinp->in6p_outputopts)
867 			inp->in6p_outputopts =
868 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
869 	}
870 
871 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
872 		struct in6_addr laddr6;
873 		struct sockaddr_in6 sin6;
874 
875 		sin6.sin6_family = AF_INET6;
876 		sin6.sin6_len = sizeof(sin6);
877 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
878 		sin6.sin6_port = sc->sc_inc.inc_fport;
879 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
880 		laddr6 = inp->in6p_laddr;
881 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
882 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
883 		if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
884 		    thread0.td_ucred, m, false)) != 0) {
885 			inp->in6p_laddr = laddr6;
886 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
887 				log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
888 				    "with error %i\n",
889 				    s, __func__, error);
890 				free(s, M_TCPLOG);
891 			}
892 			INP_HASH_WUNLOCK(&V_tcbinfo);
893 			goto abort;
894 		}
895 		/* Override flowlabel from in6_pcbconnect. */
896 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
897 		inp->inp_flow |= sc->sc_flowlabel;
898 	}
899 #endif /* INET6 */
900 #if defined(INET) && defined(INET6)
901 	else
902 #endif
903 #ifdef INET
904 	{
905 		struct in_addr laddr;
906 		struct sockaddr_in sin;
907 
908 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
909 
910 		if (inp->inp_options == NULL) {
911 			inp->inp_options = sc->sc_ipopts;
912 			sc->sc_ipopts = NULL;
913 		}
914 
915 		sin.sin_family = AF_INET;
916 		sin.sin_len = sizeof(sin);
917 		sin.sin_addr = sc->sc_inc.inc_faddr;
918 		sin.sin_port = sc->sc_inc.inc_fport;
919 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
920 		laddr = inp->inp_laddr;
921 		if (inp->inp_laddr.s_addr == INADDR_ANY)
922 			inp->inp_laddr = sc->sc_inc.inc_laddr;
923 		if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
924 		    thread0.td_ucred, m, false)) != 0) {
925 			inp->inp_laddr = laddr;
926 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
927 				log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
928 				    "with error %i\n",
929 				    s, __func__, error);
930 				free(s, M_TCPLOG);
931 			}
932 			INP_HASH_WUNLOCK(&V_tcbinfo);
933 			goto abort;
934 		}
935 	}
936 #endif /* INET */
937 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
938 	/* Copy old policy into new socket's. */
939 	if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
940 		printf("syncache_socket: could not copy policy\n");
941 #endif
942 	INP_HASH_WUNLOCK(&V_tcbinfo);
943 	tp = intotcpcb(inp);
944 	tcp_state_change(tp, TCPS_SYN_RECEIVED);
945 	tp->iss = sc->sc_iss;
946 	tp->irs = sc->sc_irs;
947 	tcp_rcvseqinit(tp);
948 	tcp_sendseqinit(tp);
949 	blk = sototcpcb(lso)->t_fb;
950 	if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
951 		/*
952 		 * Our parents t_fb was not the default,
953 		 * we need to release our ref on tp->t_fb and
954 		 * pickup one on the new entry.
955 		 */
956 		struct tcp_function_block *rblk;
957 
958 		rblk = find_and_ref_tcp_fb(blk);
959 		KASSERT(rblk != NULL,
960 		    ("cannot find blk %p out of syncache?", blk));
961 		if (tp->t_fb->tfb_tcp_fb_fini)
962 			(*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
963 		refcount_release(&tp->t_fb->tfb_refcnt);
964 		tp->t_fb = rblk;
965 		/*
966 		 * XXXrrs this is quite dangerous, it is possible
967 		 * for the new function to fail to init. We also
968 		 * are not asking if the handoff_is_ok though at
969 		 * the very start thats probalbly ok.
970 		 */
971 		if (tp->t_fb->tfb_tcp_fb_init) {
972 			(*tp->t_fb->tfb_tcp_fb_init)(tp);
973 		}
974 	}
975 	tp->snd_wl1 = sc->sc_irs;
976 	tp->snd_max = tp->iss + 1;
977 	tp->snd_nxt = tp->iss + 1;
978 	tp->rcv_up = sc->sc_irs + 1;
979 	tp->rcv_wnd = sc->sc_wnd;
980 	tp->rcv_adv += tp->rcv_wnd;
981 	tp->last_ack_sent = tp->rcv_nxt;
982 
983 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
984 	if (sc->sc_flags & SCF_NOOPT)
985 		tp->t_flags |= TF_NOOPT;
986 	else {
987 		if (sc->sc_flags & SCF_WINSCALE) {
988 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
989 			tp->snd_scale = sc->sc_requested_s_scale;
990 			tp->request_r_scale = sc->sc_requested_r_scale;
991 		}
992 		if (sc->sc_flags & SCF_TIMESTAMP) {
993 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
994 			tp->ts_recent = sc->sc_tsreflect;
995 			tp->ts_recent_age = tcp_ts_getticks();
996 			tp->ts_offset = sc->sc_tsoff;
997 		}
998 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
999 		if (sc->sc_flags & SCF_SIGNATURE)
1000 			tp->t_flags |= TF_SIGNATURE;
1001 #endif
1002 		if (sc->sc_flags & SCF_SACK)
1003 			tp->t_flags |= TF_SACK_PERMIT;
1004 	}
1005 
1006 	if (sc->sc_flags & SCF_ECN)
1007 		tp->t_flags2 |= TF2_ECN_PERMIT;
1008 
1009 	/*
1010 	 * Set up MSS and get cached values from tcp_hostcache.
1011 	 * This might overwrite some of the defaults we just set.
1012 	 */
1013 	tcp_mss(tp, sc->sc_peer_mss);
1014 
1015 	/*
1016 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
1017 	 * limited to one segment in cc_conn_init().
1018 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
1019 	 */
1020 	if (sc->sc_rxmits > 1)
1021 		tp->snd_cwnd = 1;
1022 
1023 #ifdef TCP_OFFLOAD
1024 	/*
1025 	 * Allow a TOE driver to install its hooks.  Note that we hold the
1026 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
1027 	 * new connection before the TOE driver has done its thing.
1028 	 */
1029 	if (ADDED_BY_TOE(sc)) {
1030 		struct toedev *tod = sc->sc_tod;
1031 
1032 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
1033 	}
1034 #endif
1035 	/*
1036 	 * Copy and activate timers.
1037 	 */
1038 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
1039 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
1040 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
1041 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1042 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1043 
1044 	TCPSTAT_INC(tcps_accepts);
1045 	return (so);
1046 
1047 abort:
1048 	INP_WUNLOCK(inp);
1049 abort2:
1050 	if (so != NULL)
1051 		soabort(so);
1052 	return (NULL);
1053 }
1054 
1055 /*
1056  * This function gets called when we receive an ACK for a
1057  * socket in the LISTEN state.  We look up the connection
1058  * in the syncache, and if its there, we pull it out of
1059  * the cache and turn it into a full-blown connection in
1060  * the SYN-RECEIVED state.
1061  *
1062  * On syncache_socket() success the newly created socket
1063  * has its underlying inp locked.
1064  */
1065 int
1066 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1067     struct socket **lsop, struct mbuf *m)
1068 {
1069 	struct syncache *sc;
1070 	struct syncache_head *sch;
1071 	struct syncache scs;
1072 	char *s;
1073 	bool locked;
1074 
1075 	NET_EPOCH_ASSERT();
1076 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1077 	    ("%s: can handle only ACK", __func__));
1078 
1079 	if (syncache_cookiesonly()) {
1080 		sc = NULL;
1081 		sch = syncache_hashbucket(inc);
1082 		locked = false;
1083 	} else {
1084 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1085 		locked = true;
1086 		SCH_LOCK_ASSERT(sch);
1087 	}
1088 
1089 #ifdef INVARIANTS
1090 	/*
1091 	 * Test code for syncookies comparing the syncache stored
1092 	 * values with the reconstructed values from the cookie.
1093 	 */
1094 	if (sc != NULL)
1095 		syncookie_cmp(inc, sch, sc, th, to, *lsop);
1096 #endif
1097 
1098 	if (sc == NULL) {
1099 		/*
1100 		 * There is no syncache entry, so see if this ACK is
1101 		 * a returning syncookie.  To do this, first:
1102 		 *  A. Check if syncookies are used in case of syncache
1103 		 *     overflows
1104 		 *  B. See if this socket has had a syncache entry dropped in
1105 		 *     the recent past. We don't want to accept a bogus
1106 		 *     syncookie if we've never received a SYN or accept it
1107 		 *     twice.
1108 		 *  C. check that the syncookie is valid.  If it is, then
1109 		 *     cobble up a fake syncache entry, and return.
1110 		 */
1111 		if (locked && !V_tcp_syncookies) {
1112 			SCH_UNLOCK(sch);
1113 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1114 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1115 				    "segment rejected (syncookies disabled)\n",
1116 				    s, __func__);
1117 			goto failed;
1118 		}
1119 		if (locked && !V_tcp_syncookiesonly &&
1120 		    sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1121 			SCH_UNLOCK(sch);
1122 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1123 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1124 				    "segment rejected (no syncache entry)\n",
1125 				    s, __func__);
1126 			goto failed;
1127 		}
1128 		bzero(&scs, sizeof(scs));
1129 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
1130 		if (locked)
1131 			SCH_UNLOCK(sch);
1132 		if (sc == NULL) {
1133 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1134 				log(LOG_DEBUG, "%s; %s: Segment failed "
1135 				    "SYNCOOKIE authentication, segment rejected "
1136 				    "(probably spoofed)\n", s, __func__);
1137 			goto failed;
1138 		}
1139 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1140 		/* If received ACK has MD5 signature, check it. */
1141 		if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1142 		    (!TCPMD5_ENABLED() ||
1143 		    TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1144 			/* Drop the ACK. */
1145 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1146 				log(LOG_DEBUG, "%s; %s: Segment rejected, "
1147 				    "MD5 signature doesn't match.\n",
1148 				    s, __func__);
1149 				free(s, M_TCPLOG);
1150 			}
1151 			TCPSTAT_INC(tcps_sig_err_sigopt);
1152 			return (-1); /* Do not send RST */
1153 		}
1154 #endif /* TCP_SIGNATURE */
1155 	} else {
1156 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1157 		/*
1158 		 * If listening socket requested TCP digests, check that
1159 		 * received ACK has signature and it is correct.
1160 		 * If not, drop the ACK and leave sc entry in th cache,
1161 		 * because SYN was received with correct signature.
1162 		 */
1163 		if (sc->sc_flags & SCF_SIGNATURE) {
1164 			if ((to->to_flags & TOF_SIGNATURE) == 0) {
1165 				/* No signature */
1166 				TCPSTAT_INC(tcps_sig_err_nosigopt);
1167 				SCH_UNLOCK(sch);
1168 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1169 					log(LOG_DEBUG, "%s; %s: Segment "
1170 					    "rejected, MD5 signature wasn't "
1171 					    "provided.\n", s, __func__);
1172 					free(s, M_TCPLOG);
1173 				}
1174 				return (-1); /* Do not send RST */
1175 			}
1176 			if (!TCPMD5_ENABLED() ||
1177 			    TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1178 				/* Doesn't match or no SA */
1179 				SCH_UNLOCK(sch);
1180 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1181 					log(LOG_DEBUG, "%s; %s: Segment "
1182 					    "rejected, MD5 signature doesn't "
1183 					    "match.\n", s, __func__);
1184 					free(s, M_TCPLOG);
1185 				}
1186 				return (-1); /* Do not send RST */
1187 			}
1188 		}
1189 #endif /* TCP_SIGNATURE */
1190 
1191 		/*
1192 		 * RFC 7323 PAWS: If we have a timestamp on this segment and
1193 		 * it's less than ts_recent, drop it.
1194 		 * XXXMT: RFC 7323 also requires to send an ACK.
1195 		 *        In tcp_input.c this is only done for TCP segments
1196 		 *        with user data, so be consistent here and just drop
1197 		 *        the segment.
1198 		 */
1199 		if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1200 		    TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1201 			SCH_UNLOCK(sch);
1202 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1203 				log(LOG_DEBUG,
1204 				    "%s; %s: SEG.TSval %u < TS.Recent %u, "
1205 				    "segment dropped\n", s, __func__,
1206 				    to->to_tsval, sc->sc_tsreflect);
1207 				free(s, M_TCPLOG);
1208 			}
1209 			return (-1);  /* Do not send RST */
1210 		}
1211 
1212 		/*
1213 		 * Pull out the entry to unlock the bucket row.
1214 		 *
1215 		 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1216 		 * tcp_state_change().  The tcpcb is not existent at this
1217 		 * moment.  A new one will be allocated via syncache_socket->
1218 		 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1219 		 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1220 		 */
1221 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1222 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1223 		sch->sch_length--;
1224 #ifdef TCP_OFFLOAD
1225 		if (ADDED_BY_TOE(sc)) {
1226 			struct toedev *tod = sc->sc_tod;
1227 
1228 			tod->tod_syncache_removed(tod, sc->sc_todctx);
1229 		}
1230 #endif
1231 		SCH_UNLOCK(sch);
1232 	}
1233 
1234 	/*
1235 	 * Segment validation:
1236 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1237 	 */
1238 	if (th->th_ack != sc->sc_iss + 1) {
1239 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1240 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1241 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1242 		goto failed;
1243 	}
1244 
1245 	/*
1246 	 * The SEQ must fall in the window starting at the received
1247 	 * initial receive sequence number + 1 (the SYN).
1248 	 */
1249 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1250 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1251 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1252 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1253 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1254 		goto failed;
1255 	}
1256 
1257 	/*
1258 	 * If timestamps were not negotiated during SYN/ACK they
1259 	 * must not appear on any segment during this session.
1260 	 */
1261 	if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1262 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1263 			log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1264 			    "segment rejected\n", s, __func__);
1265 		goto failed;
1266 	}
1267 
1268 	/*
1269 	 * If timestamps were negotiated during SYN/ACK they should
1270 	 * appear on every segment during this session.
1271 	 * XXXAO: This is only informal as there have been unverified
1272 	 * reports of non-compliants stacks.
1273 	 */
1274 	if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1275 		if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1276 			log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1277 			    "no action\n", s, __func__);
1278 			free(s, M_TCPLOG);
1279 			s = NULL;
1280 		}
1281 	}
1282 
1283 	*lsop = syncache_socket(sc, *lsop, m);
1284 
1285 	if (*lsop == NULL)
1286 		TCPSTAT_INC(tcps_sc_aborted);
1287 	else
1288 		TCPSTAT_INC(tcps_sc_completed);
1289 
1290 /* how do we find the inp for the new socket? */
1291 	if (sc != &scs)
1292 		syncache_free(sc);
1293 	return (1);
1294 failed:
1295 	if (sc != NULL && sc != &scs)
1296 		syncache_free(sc);
1297 	if (s != NULL)
1298 		free(s, M_TCPLOG);
1299 	*lsop = NULL;
1300 	return (0);
1301 }
1302 
1303 static void
1304 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
1305     uint64_t response_cookie)
1306 {
1307 	struct inpcb *inp;
1308 	struct tcpcb *tp;
1309 	unsigned int *pending_counter;
1310 
1311 	NET_EPOCH_ASSERT();
1312 
1313 	pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
1314 	*lsop = syncache_socket(sc, *lsop, m);
1315 	if (*lsop == NULL) {
1316 		TCPSTAT_INC(tcps_sc_aborted);
1317 		atomic_subtract_int(pending_counter, 1);
1318 	} else {
1319 		soisconnected(*lsop);
1320 		inp = sotoinpcb(*lsop);
1321 		tp = intotcpcb(inp);
1322 		tp->t_flags |= TF_FASTOPEN;
1323 		tp->t_tfo_cookie.server = response_cookie;
1324 		tp->snd_max = tp->iss;
1325 		tp->snd_nxt = tp->iss;
1326 		tp->t_tfo_pending = pending_counter;
1327 		TCPSTAT_INC(tcps_sc_completed);
1328 	}
1329 }
1330 
1331 /*
1332  * Given a LISTEN socket and an inbound SYN request, add
1333  * this to the syn cache, and send back a segment:
1334  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1335  * to the source.
1336  *
1337  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1338  * Doing so would require that we hold onto the data and deliver it
1339  * to the application.  However, if we are the target of a SYN-flood
1340  * DoS attack, an attacker could send data which would eventually
1341  * consume all available buffer space if it were ACKed.  By not ACKing
1342  * the data, we avoid this DoS scenario.
1343  *
1344  * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1345  * cookie is processed and a new socket is created.  In this case, any data
1346  * accompanying the SYN will be queued to the socket by tcp_input() and will
1347  * be ACKed either when the application sends response data or the delayed
1348  * ACK timer expires, whichever comes first.
1349  */
1350 int
1351 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1352     struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1353     void *todctx, uint8_t iptos)
1354 {
1355 	struct tcpcb *tp;
1356 	struct socket *so;
1357 	struct syncache *sc = NULL;
1358 	struct syncache_head *sch;
1359 	struct mbuf *ipopts = NULL;
1360 	u_int ltflags;
1361 	int win, ip_ttl, ip_tos;
1362 	char *s;
1363 	int rv = 0;
1364 #ifdef INET6
1365 	int autoflowlabel = 0;
1366 #endif
1367 #ifdef MAC
1368 	struct label *maclabel;
1369 #endif
1370 	struct syncache scs;
1371 	struct ucred *cred;
1372 	uint64_t tfo_response_cookie;
1373 	unsigned int *tfo_pending = NULL;
1374 	int tfo_cookie_valid = 0;
1375 	int tfo_response_cookie_valid = 0;
1376 	bool locked;
1377 
1378 	INP_WLOCK_ASSERT(inp);			/* listen socket */
1379 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1380 	    ("%s: unexpected tcp flags", __func__));
1381 
1382 	/*
1383 	 * Combine all so/tp operations very early to drop the INP lock as
1384 	 * soon as possible.
1385 	 */
1386 	so = *lsop;
1387 	KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1388 	tp = sototcpcb(so);
1389 	cred = crhold(so->so_cred);
1390 
1391 #ifdef INET6
1392 	if ((inc->inc_flags & INC_ISIPV6) &&
1393 	    (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1394 		autoflowlabel = 1;
1395 #endif
1396 	ip_ttl = inp->inp_ip_ttl;
1397 	ip_tos = inp->inp_ip_tos;
1398 	win = so->sol_sbrcv_hiwat;
1399 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1400 
1401 	if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
1402 	    (tp->t_tfo_pending != NULL) &&
1403 	    (to->to_flags & TOF_FASTOPEN)) {
1404 		/*
1405 		 * Limit the number of pending TFO connections to
1406 		 * approximately half of the queue limit.  This prevents TFO
1407 		 * SYN floods from starving the service by filling the
1408 		 * listen queue with bogus TFO connections.
1409 		 */
1410 		if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1411 		    (so->sol_qlimit / 2)) {
1412 			int result;
1413 
1414 			result = tcp_fastopen_check_cookie(inc,
1415 			    to->to_tfo_cookie, to->to_tfo_len,
1416 			    &tfo_response_cookie);
1417 			tfo_cookie_valid = (result > 0);
1418 			tfo_response_cookie_valid = (result >= 0);
1419 		}
1420 
1421 		/*
1422 		 * Remember the TFO pending counter as it will have to be
1423 		 * decremented below if we don't make it to syncache_tfo_expand().
1424 		 */
1425 		tfo_pending = tp->t_tfo_pending;
1426 	}
1427 
1428 	/* By the time we drop the lock these should no longer be used. */
1429 	so = NULL;
1430 	tp = NULL;
1431 
1432 #ifdef MAC
1433 	if (mac_syncache_init(&maclabel) != 0) {
1434 		INP_WUNLOCK(inp);
1435 		goto done;
1436 	} else
1437 		mac_syncache_create(maclabel, inp);
1438 #endif
1439 	if (!tfo_cookie_valid)
1440 		INP_WUNLOCK(inp);
1441 
1442 	/*
1443 	 * Remember the IP options, if any.
1444 	 */
1445 #ifdef INET6
1446 	if (!(inc->inc_flags & INC_ISIPV6))
1447 #endif
1448 #ifdef INET
1449 		ipopts = (m) ? ip_srcroute(m) : NULL;
1450 #else
1451 		ipopts = NULL;
1452 #endif
1453 
1454 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1455 	/*
1456 	 * If listening socket requested TCP digests, check that received
1457 	 * SYN has signature and it is correct. If signature doesn't match
1458 	 * or TCP_SIGNATURE support isn't enabled, drop the packet.
1459 	 */
1460 	if (ltflags & TF_SIGNATURE) {
1461 		if ((to->to_flags & TOF_SIGNATURE) == 0) {
1462 			TCPSTAT_INC(tcps_sig_err_nosigopt);
1463 			goto done;
1464 		}
1465 		if (!TCPMD5_ENABLED() ||
1466 		    TCPMD5_INPUT(m, th, to->to_signature) != 0)
1467 			goto done;
1468 	}
1469 #endif	/* TCP_SIGNATURE */
1470 	/*
1471 	 * See if we already have an entry for this connection.
1472 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1473 	 *
1474 	 * XXX: should the syncache be re-initialized with the contents
1475 	 * of the new SYN here (which may have different options?)
1476 	 *
1477 	 * XXX: We do not check the sequence number to see if this is a
1478 	 * real retransmit or a new connection attempt.  The question is
1479 	 * how to handle such a case; either ignore it as spoofed, or
1480 	 * drop the current entry and create a new one?
1481 	 */
1482 	if (syncache_cookiesonly()) {
1483 		sc = NULL;
1484 		sch = syncache_hashbucket(inc);
1485 		locked = false;
1486 	} else {
1487 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1488 		locked = true;
1489 		SCH_LOCK_ASSERT(sch);
1490 	}
1491 	if (sc != NULL) {
1492 		if (tfo_cookie_valid)
1493 			INP_WUNLOCK(inp);
1494 		TCPSTAT_INC(tcps_sc_dupsyn);
1495 		if (ipopts) {
1496 			/*
1497 			 * If we were remembering a previous source route,
1498 			 * forget it and use the new one we've been given.
1499 			 */
1500 			if (sc->sc_ipopts)
1501 				(void) m_free(sc->sc_ipopts);
1502 			sc->sc_ipopts = ipopts;
1503 		}
1504 		/*
1505 		 * Update timestamp if present.
1506 		 */
1507 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1508 			sc->sc_tsreflect = to->to_tsval;
1509 		else
1510 			sc->sc_flags &= ~SCF_TIMESTAMP;
1511 		/*
1512 		 * Disable ECN if needed.
1513 		 */
1514 		if ((sc->sc_flags & SCF_ECN) &&
1515 		    ((th->th_flags & (TH_ECE|TH_CWR)) != (TH_ECE|TH_CWR))) {
1516 			sc->sc_flags &= ~SCF_ECN;
1517 		}
1518 #ifdef MAC
1519 		/*
1520 		 * Since we have already unconditionally allocated label
1521 		 * storage, free it up.  The syncache entry will already
1522 		 * have an initialized label we can use.
1523 		 */
1524 		mac_syncache_destroy(&maclabel);
1525 #endif
1526 		TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1527 		/* Retransmit SYN|ACK and reset retransmit count. */
1528 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1529 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1530 			    "resetting timer and retransmitting SYN|ACK\n",
1531 			    s, __func__);
1532 			free(s, M_TCPLOG);
1533 		}
1534 		if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1535 			sc->sc_rxmits = 0;
1536 			syncache_timeout(sc, sch, 1);
1537 			TCPSTAT_INC(tcps_sndacks);
1538 			TCPSTAT_INC(tcps_sndtotal);
1539 		}
1540 		SCH_UNLOCK(sch);
1541 		goto donenoprobe;
1542 	}
1543 
1544 	if (tfo_cookie_valid) {
1545 		bzero(&scs, sizeof(scs));
1546 		sc = &scs;
1547 		goto skip_alloc;
1548 	}
1549 
1550 	/*
1551 	 * Skip allocating a syncache entry if we are just going to discard
1552 	 * it later.
1553 	 */
1554 	if (!locked) {
1555 		bzero(&scs, sizeof(scs));
1556 		sc = &scs;
1557 	} else
1558 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1559 	if (sc == NULL) {
1560 		/*
1561 		 * The zone allocator couldn't provide more entries.
1562 		 * Treat this as if the cache was full; drop the oldest
1563 		 * entry and insert the new one.
1564 		 */
1565 		TCPSTAT_INC(tcps_sc_zonefail);
1566 		if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1567 			sch->sch_last_overflow = time_uptime;
1568 			syncache_drop(sc, sch);
1569 			syncache_pause(inc);
1570 		}
1571 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1572 		if (sc == NULL) {
1573 			if (V_tcp_syncookies) {
1574 				bzero(&scs, sizeof(scs));
1575 				sc = &scs;
1576 			} else {
1577 				KASSERT(locked,
1578 				    ("%s: bucket unexpectedly unlocked",
1579 				    __func__));
1580 				SCH_UNLOCK(sch);
1581 				if (ipopts)
1582 					(void) m_free(ipopts);
1583 				goto done;
1584 			}
1585 		}
1586 	}
1587 
1588 skip_alloc:
1589 	if (!tfo_cookie_valid && tfo_response_cookie_valid)
1590 		sc->sc_tfo_cookie = &tfo_response_cookie;
1591 
1592 	/*
1593 	 * Fill in the syncache values.
1594 	 */
1595 #ifdef MAC
1596 	sc->sc_label = maclabel;
1597 #endif
1598 	sc->sc_cred = cred;
1599 	cred = NULL;
1600 	sc->sc_ipopts = ipopts;
1601 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1602 #ifdef INET6
1603 	if (!(inc->inc_flags & INC_ISIPV6))
1604 #endif
1605 	{
1606 		sc->sc_ip_tos = ip_tos;
1607 		sc->sc_ip_ttl = ip_ttl;
1608 	}
1609 #ifdef TCP_OFFLOAD
1610 	sc->sc_tod = tod;
1611 	sc->sc_todctx = todctx;
1612 #endif
1613 	sc->sc_irs = th->th_seq;
1614 	sc->sc_flags = 0;
1615 	sc->sc_flowlabel = 0;
1616 
1617 	/*
1618 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1619 	 * win was derived from socket earlier in the function.
1620 	 */
1621 	win = imax(win, 0);
1622 	win = imin(win, TCP_MAXWIN);
1623 	sc->sc_wnd = win;
1624 
1625 	if (V_tcp_do_rfc1323) {
1626 		/*
1627 		 * A timestamp received in a SYN makes
1628 		 * it ok to send timestamp requests and replies.
1629 		 */
1630 		if (to->to_flags & TOF_TS) {
1631 			sc->sc_tsreflect = to->to_tsval;
1632 			sc->sc_flags |= SCF_TIMESTAMP;
1633 			sc->sc_tsoff = tcp_new_ts_offset(inc);
1634 		}
1635 		if (to->to_flags & TOF_SCALE) {
1636 			int wscale = 0;
1637 
1638 			/*
1639 			 * Pick the smallest possible scaling factor that
1640 			 * will still allow us to scale up to sb_max, aka
1641 			 * kern.ipc.maxsockbuf.
1642 			 *
1643 			 * We do this because there are broken firewalls that
1644 			 * will corrupt the window scale option, leading to
1645 			 * the other endpoint believing that our advertised
1646 			 * window is unscaled.  At scale factors larger than
1647 			 * 5 the unscaled window will drop below 1500 bytes,
1648 			 * leading to serious problems when traversing these
1649 			 * broken firewalls.
1650 			 *
1651 			 * With the default maxsockbuf of 256K, a scale factor
1652 			 * of 3 will be chosen by this algorithm.  Those who
1653 			 * choose a larger maxsockbuf should watch out
1654 			 * for the compatibility problems mentioned above.
1655 			 *
1656 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1657 			 * or <SYN,ACK>) segment itself is never scaled.
1658 			 */
1659 			while (wscale < TCP_MAX_WINSHIFT &&
1660 			    (TCP_MAXWIN << wscale) < sb_max)
1661 				wscale++;
1662 			sc->sc_requested_r_scale = wscale;
1663 			sc->sc_requested_s_scale = to->to_wscale;
1664 			sc->sc_flags |= SCF_WINSCALE;
1665 		}
1666 	}
1667 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1668 	/*
1669 	 * If listening socket requested TCP digests, flag this in the
1670 	 * syncache so that syncache_respond() will do the right thing
1671 	 * with the SYN+ACK.
1672 	 */
1673 	if (ltflags & TF_SIGNATURE)
1674 		sc->sc_flags |= SCF_SIGNATURE;
1675 #endif	/* TCP_SIGNATURE */
1676 	if (to->to_flags & TOF_SACKPERM)
1677 		sc->sc_flags |= SCF_SACK;
1678 	if (to->to_flags & TOF_MSS)
1679 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1680 	if (ltflags & TF_NOOPT)
1681 		sc->sc_flags |= SCF_NOOPT;
1682 	if (((th->th_flags & (TH_ECE|TH_CWR)) == (TH_ECE|TH_CWR)) &&
1683 	    V_tcp_do_ecn)
1684 		sc->sc_flags |= SCF_ECN;
1685 
1686 	if (V_tcp_syncookies)
1687 		sc->sc_iss = syncookie_generate(sch, sc);
1688 	else
1689 		sc->sc_iss = arc4random();
1690 #ifdef INET6
1691 	if (autoflowlabel) {
1692 		if (V_tcp_syncookies)
1693 			sc->sc_flowlabel = sc->sc_iss;
1694 		else
1695 			sc->sc_flowlabel = ip6_randomflowlabel();
1696 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1697 	}
1698 #endif
1699 	if (locked)
1700 		SCH_UNLOCK(sch);
1701 
1702 	if (tfo_cookie_valid) {
1703 		syncache_tfo_expand(sc, lsop, m, tfo_response_cookie);
1704 		/* INP_WUNLOCK(inp) will be performed by the caller */
1705 		rv = 1;
1706 		goto tfo_expanded;
1707 	}
1708 
1709 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1710 	/*
1711 	 * Do a standard 3-way handshake.
1712 	 */
1713 	if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1714 		if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1715 			syncache_free(sc);
1716 		else if (sc != &scs)
1717 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1718 		TCPSTAT_INC(tcps_sndacks);
1719 		TCPSTAT_INC(tcps_sndtotal);
1720 	} else {
1721 		if (sc != &scs)
1722 			syncache_free(sc);
1723 		TCPSTAT_INC(tcps_sc_dropped);
1724 	}
1725 	goto donenoprobe;
1726 
1727 done:
1728 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1729 donenoprobe:
1730 	if (m) {
1731 		*lsop = NULL;
1732 		m_freem(m);
1733 	}
1734 	/*
1735 	 * If tfo_pending is not NULL here, then a TFO SYN that did not
1736 	 * result in a new socket was processed and the associated pending
1737 	 * counter has not yet been decremented.  All such TFO processing paths
1738 	 * transit this point.
1739 	 */
1740 	if (tfo_pending != NULL)
1741 		tcp_fastopen_decrement_counter(tfo_pending);
1742 
1743 tfo_expanded:
1744 	if (cred != NULL)
1745 		crfree(cred);
1746 #ifdef MAC
1747 	if (sc == &scs)
1748 		mac_syncache_destroy(&maclabel);
1749 #endif
1750 	return (rv);
1751 }
1752 
1753 /*
1754  * Send SYN|ACK or ACK to the peer.  Either in response to a peer's segment,
1755  * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1756  */
1757 static int
1758 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1759 {
1760 	struct ip *ip = NULL;
1761 	struct mbuf *m;
1762 	struct tcphdr *th = NULL;
1763 	int optlen, error = 0;	/* Make compiler happy */
1764 	u_int16_t hlen, tlen, mssopt;
1765 	struct tcpopt to;
1766 #ifdef INET6
1767 	struct ip6_hdr *ip6 = NULL;
1768 #endif
1769 
1770 	NET_EPOCH_ASSERT();
1771 
1772 	hlen =
1773 #ifdef INET6
1774 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1775 #endif
1776 		sizeof(struct ip);
1777 	tlen = hlen + sizeof(struct tcphdr);
1778 
1779 	/* Determine MSS we advertize to other end of connection. */
1780 	mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss);
1781 
1782 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1783 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1784 	    ("syncache: mbuf too small"));
1785 
1786 	/* Create the IP+TCP header from scratch. */
1787 	m = m_gethdr(M_NOWAIT, MT_DATA);
1788 	if (m == NULL)
1789 		return (ENOBUFS);
1790 #ifdef MAC
1791 	mac_syncache_create_mbuf(sc->sc_label, m);
1792 #endif
1793 	m->m_data += max_linkhdr;
1794 	m->m_len = tlen;
1795 	m->m_pkthdr.len = tlen;
1796 	m->m_pkthdr.rcvif = NULL;
1797 
1798 #ifdef INET6
1799 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1800 		ip6 = mtod(m, struct ip6_hdr *);
1801 		ip6->ip6_vfc = IPV6_VERSION;
1802 		ip6->ip6_nxt = IPPROTO_TCP;
1803 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1804 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1805 		ip6->ip6_plen = htons(tlen - hlen);
1806 		/* ip6_hlim is set after checksum */
1807 		/* Zero out traffic class and flow label. */
1808 		ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1809 		ip6->ip6_flow |= sc->sc_flowlabel;
1810 
1811 		th = (struct tcphdr *)(ip6 + 1);
1812 	}
1813 #endif
1814 #if defined(INET6) && defined(INET)
1815 	else
1816 #endif
1817 #ifdef INET
1818 	{
1819 		ip = mtod(m, struct ip *);
1820 		ip->ip_v = IPVERSION;
1821 		ip->ip_hl = sizeof(struct ip) >> 2;
1822 		ip->ip_len = htons(tlen);
1823 		ip->ip_id = 0;
1824 		ip->ip_off = 0;
1825 		ip->ip_sum = 0;
1826 		ip->ip_p = IPPROTO_TCP;
1827 		ip->ip_src = sc->sc_inc.inc_laddr;
1828 		ip->ip_dst = sc->sc_inc.inc_faddr;
1829 		ip->ip_ttl = sc->sc_ip_ttl;
1830 		ip->ip_tos = sc->sc_ip_tos;
1831 
1832 		/*
1833 		 * See if we should do MTU discovery.  Route lookups are
1834 		 * expensive, so we will only unset the DF bit if:
1835 		 *
1836 		 *	1) path_mtu_discovery is disabled
1837 		 *	2) the SCF_UNREACH flag has been set
1838 		 */
1839 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1840 		       ip->ip_off |= htons(IP_DF);
1841 
1842 		th = (struct tcphdr *)(ip + 1);
1843 	}
1844 #endif /* INET */
1845 	th->th_sport = sc->sc_inc.inc_lport;
1846 	th->th_dport = sc->sc_inc.inc_fport;
1847 
1848 	if (flags & TH_SYN)
1849 		th->th_seq = htonl(sc->sc_iss);
1850 	else
1851 		th->th_seq = htonl(sc->sc_iss + 1);
1852 	th->th_ack = htonl(sc->sc_irs + 1);
1853 	th->th_off = sizeof(struct tcphdr) >> 2;
1854 	th->th_x2 = 0;
1855 	th->th_flags = flags;
1856 	th->th_win = htons(sc->sc_wnd);
1857 	th->th_urp = 0;
1858 
1859 	if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) {
1860 		th->th_flags |= TH_ECE;
1861 		TCPSTAT_INC(tcps_ecn_shs);
1862 	}
1863 
1864 	/* Tack on the TCP options. */
1865 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1866 		to.to_flags = 0;
1867 
1868 		if (flags & TH_SYN) {
1869 			to.to_mss = mssopt;
1870 			to.to_flags = TOF_MSS;
1871 			if (sc->sc_flags & SCF_WINSCALE) {
1872 				to.to_wscale = sc->sc_requested_r_scale;
1873 				to.to_flags |= TOF_SCALE;
1874 			}
1875 			if (sc->sc_flags & SCF_SACK)
1876 				to.to_flags |= TOF_SACKPERM;
1877 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1878 			if (sc->sc_flags & SCF_SIGNATURE)
1879 				to.to_flags |= TOF_SIGNATURE;
1880 #endif
1881 			if (sc->sc_tfo_cookie) {
1882 				to.to_flags |= TOF_FASTOPEN;
1883 				to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1884 				to.to_tfo_cookie = sc->sc_tfo_cookie;
1885 				/* don't send cookie again when retransmitting response */
1886 				sc->sc_tfo_cookie = NULL;
1887 			}
1888 		}
1889 		if (sc->sc_flags & SCF_TIMESTAMP) {
1890 			to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1891 			to.to_tsecr = sc->sc_tsreflect;
1892 			to.to_flags |= TOF_TS;
1893 		}
1894 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1895 
1896 		/* Adjust headers by option size. */
1897 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1898 		m->m_len += optlen;
1899 		m->m_pkthdr.len += optlen;
1900 #ifdef INET6
1901 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1902 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1903 		else
1904 #endif
1905 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1906 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1907 		if (sc->sc_flags & SCF_SIGNATURE) {
1908 			KASSERT(to.to_flags & TOF_SIGNATURE,
1909 			    ("tcp_addoptions() didn't set tcp_signature"));
1910 
1911 			/* NOTE: to.to_signature is inside of mbuf */
1912 			if (!TCPMD5_ENABLED() ||
1913 			    TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1914 				m_freem(m);
1915 				return (EACCES);
1916 			}
1917 		}
1918 #endif
1919 	} else
1920 		optlen = 0;
1921 
1922 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
1923 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1924 	/*
1925 	 * If we have peer's SYN and it has a flowid, then let's assign it to
1926 	 * our SYN|ACK.  ip6_output() and ip_output() will not assign flowid
1927 	 * to SYN|ACK due to lack of inp here.
1928 	 */
1929 	if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1930 		m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1931 		M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1932 	}
1933 #ifdef INET6
1934 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1935 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1936 		th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1937 		    IPPROTO_TCP, 0);
1938 		ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1939 #ifdef TCP_OFFLOAD
1940 		if (ADDED_BY_TOE(sc)) {
1941 			struct toedev *tod = sc->sc_tod;
1942 
1943 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1944 
1945 			return (error);
1946 		}
1947 #endif
1948 		TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
1949 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1950 	}
1951 #endif
1952 #if defined(INET6) && defined(INET)
1953 	else
1954 #endif
1955 #ifdef INET
1956 	{
1957 		m->m_pkthdr.csum_flags = CSUM_TCP;
1958 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1959 		    htons(tlen + optlen - hlen + IPPROTO_TCP));
1960 #ifdef TCP_OFFLOAD
1961 		if (ADDED_BY_TOE(sc)) {
1962 			struct toedev *tod = sc->sc_tod;
1963 
1964 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1965 
1966 			return (error);
1967 		}
1968 #endif
1969 		TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
1970 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1971 	}
1972 #endif
1973 	return (error);
1974 }
1975 
1976 /*
1977  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1978  * that exceed the capacity of the syncache by avoiding the storage of any
1979  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
1980  * attacks where the attacker does not have access to our responses.
1981  *
1982  * Syncookies encode and include all necessary information about the
1983  * connection setup within the SYN|ACK that we send back.  That way we
1984  * can avoid keeping any local state until the ACK to our SYN|ACK returns
1985  * (if ever).  Normally the syncache and syncookies are running in parallel
1986  * with the latter taking over when the former is exhausted.  When matching
1987  * syncache entry is found the syncookie is ignored.
1988  *
1989  * The only reliable information persisting the 3WHS is our initial sequence
1990  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
1991  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1992  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
1993  * returns and signifies a legitimate connection if it matches the ACK.
1994  *
1995  * The available space of 32 bits to store the hash and to encode the SYN
1996  * option information is very tight and we should have at least 24 bits for
1997  * the MAC to keep the number of guesses by blind spoofing reasonably high.
1998  *
1999  * SYN option information we have to encode to fully restore a connection:
2000  * MSS: is imporant to chose an optimal segment size to avoid IP level
2001  *   fragmentation along the path.  The common MSS values can be encoded
2002  *   in a 3-bit table.  Uncommon values are captured by the next lower value
2003  *   in the table leading to a slight increase in packetization overhead.
2004  * WSCALE: is necessary to allow large windows to be used for high delay-
2005  *   bandwidth product links.  Not scaling the window when it was initially
2006  *   negotiated is bad for performance as lack of scaling further decreases
2007  *   the apparent available send window.  We only need to encode the WSCALE
2008  *   we received from the remote end.  Our end can be recalculated at any
2009  *   time.  The common WSCALE values can be encoded in a 3-bit table.
2010  *   Uncommon values are captured by the next lower value in the table
2011  *   making us under-estimate the available window size halving our
2012  *   theoretically possible maximum throughput for that connection.
2013  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2014  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2015  *   that are included in all segments on a connection.  We enable them when
2016  *   the ACK has them.
2017  *
2018  * Security of syncookies and attack vectors:
2019  *
2020  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2021  * together with the gloabl secret to make it unique per connection attempt.
2022  * Thus any change of any of those parameters results in a different MAC output
2023  * in an unpredictable way unless a collision is encountered.  24 bits of the
2024  * MAC are embedded into the ISS.
2025  *
2026  * To prevent replay attacks two rotating global secrets are updated with a
2027  * new random value every 15 seconds.  The life-time of a syncookie is thus
2028  * 15-30 seconds.
2029  *
2030  * Vector 1: Attacking the secret.  This requires finding a weakness in the
2031  * MAC itself or the way it is used here.  The attacker can do a chosen plain
2032  * text attack by varying and testing the all parameters under his control.
2033  * The strength depends on the size and randomness of the secret, and the
2034  * cryptographic security of the MAC function.  Due to the constant updating
2035  * of the secret the attacker has at most 29.999 seconds to find the secret
2036  * and launch spoofed connections.  After that he has to start all over again.
2037  *
2038  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
2039  * size an average of 4,823 attempts are required for a 50% chance of success
2040  * to spoof a single syncookie (birthday collision paradox).  However the
2041  * attacker is blind and doesn't know if one of his attempts succeeded unless
2042  * he has a side channel to interfere success from.  A single connection setup
2043  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2044  * This many attempts are required for each one blind spoofed connection.  For
2045  * every additional spoofed connection he has to launch another N attempts.
2046  * Thus for a sustained rate 100 spoofed connections per second approximately
2047  * 1,800,000 packets per second would have to be sent.
2048  *
2049  * NB: The MAC function should be fast so that it doesn't become a CPU
2050  * exhaustion attack vector itself.
2051  *
2052  * References:
2053  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2054  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2055  *   http://cr.yp.to/syncookies.html    (overview)
2056  *   http://cr.yp.to/syncookies/archive (details)
2057  *
2058  *
2059  * Schematic construction of a syncookie enabled Initial Sequence Number:
2060  *  0        1         2         3
2061  *  12345678901234567890123456789012
2062  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2063  *
2064  *  x 24 MAC (truncated)
2065  *  W  3 Send Window Scale index
2066  *  M  3 MSS index
2067  *  S  1 SACK permitted
2068  *  P  1 Odd/even secret
2069  */
2070 
2071 /*
2072  * Distribution and probability of certain MSS values.  Those in between are
2073  * rounded down to the next lower one.
2074  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2075  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
2076  */
2077 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2078 
2079 /*
2080  * Distribution and probability of certain WSCALE values.  We have to map the
2081  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2082  * bits based on prevalence of certain values.  Where we don't have an exact
2083  * match for are rounded down to the next lower one letting us under-estimate
2084  * the true available window.  At the moment this would happen only for the
2085  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2086  * and window size).  The absence of the WSCALE option (no scaling in either
2087  * direction) is encoded with index zero.
2088  * [WSCALE values histograms, Allman, 2012]
2089  *                            X 10 10 35  5  6 14 10%   by host
2090  *                            X 11  4  5  5 18 49  3%   by connections
2091  */
2092 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2093 
2094 /*
2095  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
2096  * and good cryptographic properties.
2097  */
2098 static uint32_t
2099 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2100     uint8_t *secbits, uintptr_t secmod)
2101 {
2102 	SIPHASH_CTX ctx;
2103 	uint32_t siphash[2];
2104 
2105 	SipHash24_Init(&ctx);
2106 	SipHash_SetKey(&ctx, secbits);
2107 	switch (inc->inc_flags & INC_ISIPV6) {
2108 #ifdef INET
2109 	case 0:
2110 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2111 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2112 		break;
2113 #endif
2114 #ifdef INET6
2115 	case INC_ISIPV6:
2116 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2117 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2118 		break;
2119 #endif
2120 	}
2121 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2122 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2123 	SipHash_Update(&ctx, &irs, sizeof(irs));
2124 	SipHash_Update(&ctx, &flags, sizeof(flags));
2125 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
2126 	SipHash_Final((u_int8_t *)&siphash, &ctx);
2127 
2128 	return (siphash[0] ^ siphash[1]);
2129 }
2130 
2131 static tcp_seq
2132 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2133 {
2134 	u_int i, secbit, wscale;
2135 	uint32_t iss, hash;
2136 	uint8_t *secbits;
2137 	union syncookie cookie;
2138 
2139 	cookie.cookie = 0;
2140 
2141 	/* Map our computed MSS into the 3-bit index. */
2142 	for (i = nitems(tcp_sc_msstab) - 1;
2143 	     tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2144 	     i--)
2145 		;
2146 	cookie.flags.mss_idx = i;
2147 
2148 	/*
2149 	 * Map the send window scale into the 3-bit index but only if
2150 	 * the wscale option was received.
2151 	 */
2152 	if (sc->sc_flags & SCF_WINSCALE) {
2153 		wscale = sc->sc_requested_s_scale;
2154 		for (i = nitems(tcp_sc_wstab) - 1;
2155 		    tcp_sc_wstab[i] > wscale && i > 0;
2156 		     i--)
2157 			;
2158 		cookie.flags.wscale_idx = i;
2159 	}
2160 
2161 	/* Can we do SACK? */
2162 	if (sc->sc_flags & SCF_SACK)
2163 		cookie.flags.sack_ok = 1;
2164 
2165 	/* Which of the two secrets to use. */
2166 	secbit = V_tcp_syncache.secret.oddeven & 0x1;
2167 	cookie.flags.odd_even = secbit;
2168 
2169 	secbits = V_tcp_syncache.secret.key[secbit];
2170 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2171 	    (uintptr_t)sch);
2172 
2173 	/*
2174 	 * Put the flags into the hash and XOR them to get better ISS number
2175 	 * variance.  This doesn't enhance the cryptographic strength and is
2176 	 * done to prevent the 8 cookie bits from showing up directly on the
2177 	 * wire.
2178 	 */
2179 	iss = hash & ~0xff;
2180 	iss |= cookie.cookie ^ (hash >> 24);
2181 
2182 	TCPSTAT_INC(tcps_sc_sendcookie);
2183 	return (iss);
2184 }
2185 
2186 static struct syncache *
2187 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2188     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2189     struct socket *lso)
2190 {
2191 	uint32_t hash;
2192 	uint8_t *secbits;
2193 	tcp_seq ack, seq;
2194 	int wnd, wscale = 0;
2195 	union syncookie cookie;
2196 
2197 	/*
2198 	 * Pull information out of SYN-ACK/ACK and revert sequence number
2199 	 * advances.
2200 	 */
2201 	ack = th->th_ack - 1;
2202 	seq = th->th_seq - 1;
2203 
2204 	/*
2205 	 * Unpack the flags containing enough information to restore the
2206 	 * connection.
2207 	 */
2208 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2209 
2210 	/* Which of the two secrets to use. */
2211 	secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2212 
2213 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2214 
2215 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
2216 	if ((ack & ~0xff) != (hash & ~0xff))
2217 		return (NULL);
2218 
2219 	/* Fill in the syncache values. */
2220 	sc->sc_flags = 0;
2221 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2222 	sc->sc_ipopts = NULL;
2223 
2224 	sc->sc_irs = seq;
2225 	sc->sc_iss = ack;
2226 
2227 	switch (inc->inc_flags & INC_ISIPV6) {
2228 #ifdef INET
2229 	case 0:
2230 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2231 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2232 		break;
2233 #endif
2234 #ifdef INET6
2235 	case INC_ISIPV6:
2236 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2237 			sc->sc_flowlabel =
2238 			    htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2239 		break;
2240 #endif
2241 	}
2242 
2243 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2244 
2245 	/* We can simply recompute receive window scale we sent earlier. */
2246 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2247 		wscale++;
2248 
2249 	/* Only use wscale if it was enabled in the orignal SYN. */
2250 	if (cookie.flags.wscale_idx > 0) {
2251 		sc->sc_requested_r_scale = wscale;
2252 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2253 		sc->sc_flags |= SCF_WINSCALE;
2254 	}
2255 
2256 	wnd = lso->sol_sbrcv_hiwat;
2257 	wnd = imax(wnd, 0);
2258 	wnd = imin(wnd, TCP_MAXWIN);
2259 	sc->sc_wnd = wnd;
2260 
2261 	if (cookie.flags.sack_ok)
2262 		sc->sc_flags |= SCF_SACK;
2263 
2264 	if (to->to_flags & TOF_TS) {
2265 		sc->sc_flags |= SCF_TIMESTAMP;
2266 		sc->sc_tsreflect = to->to_tsval;
2267 		sc->sc_tsoff = tcp_new_ts_offset(inc);
2268 	}
2269 
2270 	if (to->to_flags & TOF_SIGNATURE)
2271 		sc->sc_flags |= SCF_SIGNATURE;
2272 
2273 	sc->sc_rxmits = 0;
2274 
2275 	TCPSTAT_INC(tcps_sc_recvcookie);
2276 	return (sc);
2277 }
2278 
2279 #ifdef INVARIANTS
2280 static int
2281 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2282     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2283     struct socket *lso)
2284 {
2285 	struct syncache scs, *scx;
2286 	char *s;
2287 
2288 	bzero(&scs, sizeof(scs));
2289 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
2290 
2291 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2292 		return (0);
2293 
2294 	if (scx != NULL) {
2295 		if (sc->sc_peer_mss != scx->sc_peer_mss)
2296 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2297 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2298 
2299 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2300 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2301 			    s, __func__, sc->sc_requested_r_scale,
2302 			    scx->sc_requested_r_scale);
2303 
2304 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2305 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2306 			    s, __func__, sc->sc_requested_s_scale,
2307 			    scx->sc_requested_s_scale);
2308 
2309 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2310 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2311 	}
2312 
2313 	if (s != NULL)
2314 		free(s, M_TCPLOG);
2315 	return (0);
2316 }
2317 #endif /* INVARIANTS */
2318 
2319 static void
2320 syncookie_reseed(void *arg)
2321 {
2322 	struct tcp_syncache *sc = arg;
2323 	uint8_t *secbits;
2324 	int secbit;
2325 
2326 	/*
2327 	 * Reseeding the secret doesn't have to be protected by a lock.
2328 	 * It only must be ensured that the new random values are visible
2329 	 * to all CPUs in a SMP environment.  The atomic with release
2330 	 * semantics ensures that.
2331 	 */
2332 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2333 	secbits = sc->secret.key[secbit];
2334 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2335 	atomic_add_rel_int(&sc->secret.oddeven, 1);
2336 
2337 	/* Reschedule ourself. */
2338 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2339 }
2340 
2341 /*
2342  * We have overflowed a bucket. Let's pause dealing with the syncache.
2343  * This function will increment the bucketoverflow statistics appropriately
2344  * (once per pause when pausing is enabled; otherwise, once per overflow).
2345  */
2346 static void
2347 syncache_pause(struct in_conninfo *inc)
2348 {
2349 	time_t delta;
2350 	const char *s;
2351 
2352 	/* XXX:
2353 	 * 2. Add sysctl read here so we don't get the benefit of this
2354 	 * change without the new sysctl.
2355 	 */
2356 
2357 	/*
2358 	 * Try an unlocked read. If we already know that another thread
2359 	 * has activated the feature, there is no need to proceed.
2360 	 */
2361 	if (V_tcp_syncache.paused)
2362 		return;
2363 
2364 	/* Are cookied enabled? If not, we can't pause. */
2365 	if (!V_tcp_syncookies) {
2366 		TCPSTAT_INC(tcps_sc_bucketoverflow);
2367 		return;
2368 	}
2369 
2370 	/*
2371 	 * We may be the first thread to find an overflow. Get the lock
2372 	 * and evaluate if we need to take action.
2373 	 */
2374 	mtx_lock(&V_tcp_syncache.pause_mtx);
2375 	if (V_tcp_syncache.paused) {
2376 		mtx_unlock(&V_tcp_syncache.pause_mtx);
2377 		return;
2378 	}
2379 
2380 	/* Activate protection. */
2381 	V_tcp_syncache.paused = true;
2382 	TCPSTAT_INC(tcps_sc_bucketoverflow);
2383 
2384 	/*
2385 	 * Determine the last backoff time. If we are seeing a re-newed
2386 	 * attack within that same time after last reactivating the syncache,
2387 	 * consider it an extension of the same attack.
2388 	 */
2389 	delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2390 	if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2391 		if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2392 			delta <<= 1;
2393 			V_tcp_syncache.pause_backoff++;
2394 		}
2395 	} else {
2396 		delta = TCP_SYNCACHE_PAUSE_TIME;
2397 		V_tcp_syncache.pause_backoff = 0;
2398 	}
2399 
2400 	/* Log a warning, including IP addresses, if able. */
2401 	if (inc != NULL)
2402 		s = tcp_log_addrs(inc, NULL, NULL, NULL);
2403 	else
2404 		s = (const char *)NULL;
2405 	log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2406 	    "the next %lld seconds%s%s%s\n", (long long)delta,
2407 	    (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2408 	    (s != NULL) ? ")" : "");
2409 	free(__DECONST(void *, s), M_TCPLOG);
2410 
2411 	/* Use the calculated delta to set a new pause time. */
2412 	V_tcp_syncache.pause_until = time_uptime + delta;
2413 	callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2414 	    &V_tcp_syncache);
2415 	mtx_unlock(&V_tcp_syncache.pause_mtx);
2416 }
2417 
2418 /* Evaluate whether we need to unpause. */
2419 static void
2420 syncache_unpause(void *arg)
2421 {
2422 	struct tcp_syncache *sc;
2423 	time_t delta;
2424 
2425 	sc = arg;
2426 	mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2427 	callout_deactivate(&sc->pause_co);
2428 
2429 	/*
2430 	 * Check to make sure we are not running early. If the pause
2431 	 * time has expired, then deactivate the protection.
2432 	 */
2433 	if ((delta = sc->pause_until - time_uptime) > 0)
2434 		callout_schedule(&sc->pause_co, delta * hz);
2435 	else
2436 		sc->paused = false;
2437 }
2438 
2439 /*
2440  * Exports the syncache entries to userland so that netstat can display
2441  * them alongside the other sockets.  This function is intended to be
2442  * called only from tcp_pcblist.
2443  *
2444  * Due to concurrency on an active system, the number of pcbs exported
2445  * may have no relation to max_pcbs.  max_pcbs merely indicates the
2446  * amount of space the caller allocated for this function to use.
2447  */
2448 int
2449 syncache_pcblist(struct sysctl_req *req)
2450 {
2451 	struct xtcpcb xt;
2452 	struct syncache *sc;
2453 	struct syncache_head *sch;
2454 	int error, i;
2455 
2456 	bzero(&xt, sizeof(xt));
2457 	xt.xt_len = sizeof(xt);
2458 	xt.t_state = TCPS_SYN_RECEIVED;
2459 	xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2460 	xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2461 	xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2462 	xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2463 
2464 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2465 		sch = &V_tcp_syncache.hashbase[i];
2466 		SCH_LOCK(sch);
2467 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2468 			if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2469 				continue;
2470 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2471 				xt.xt_inp.inp_vflag = INP_IPV6;
2472 			else
2473 				xt.xt_inp.inp_vflag = INP_IPV4;
2474 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2475 			    sizeof (struct in_conninfo));
2476 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2477 			if (error) {
2478 				SCH_UNLOCK(sch);
2479 				return (0);
2480 			}
2481 		}
2482 		SCH_UNLOCK(sch);
2483 	}
2484 
2485 	return (0);
2486 }
2487