xref: /freebsd/sys/netinet/tcp_syncache.c (revision 357378bbdedf24ce2b90e9bd831af4a9db3ec70a)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2001 McAfee, Inc.
5  * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG
6  * All rights reserved.
7  *
8  * This software was developed for the FreeBSD Project by Jonathan Lemon
9  * and McAfee Research, the Security Research Division of McAfee, Inc. under
10  * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
11  * DARPA CHATS research program. [2001 McAfee, Inc.]
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  */
34 
35 #include <sys/cdefs.h>
36 #include "opt_inet.h"
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/hash.h>
43 #include <sys/refcount.h>
44 #include <sys/kernel.h>
45 #include <sys/sysctl.h>
46 #include <sys/limits.h>
47 #include <sys/lock.h>
48 #include <sys/mutex.h>
49 #include <sys/malloc.h>
50 #include <sys/mbuf.h>
51 #include <sys/proc.h>		/* for proc0 declaration */
52 #include <sys/random.h>
53 #include <sys/socket.h>
54 #include <sys/socketvar.h>
55 #include <sys/syslog.h>
56 #include <sys/ucred.h>
57 
58 #include <sys/md5.h>
59 #include <crypto/siphash/siphash.h>
60 
61 #include <vm/uma.h>
62 
63 #include <net/if.h>
64 #include <net/if_var.h>
65 #include <net/route.h>
66 #include <net/vnet.h>
67 
68 #include <netinet/in.h>
69 #include <netinet/in_kdtrace.h>
70 #include <netinet/in_systm.h>
71 #include <netinet/ip.h>
72 #include <netinet/in_var.h>
73 #include <netinet/in_pcb.h>
74 #include <netinet/ip_var.h>
75 #include <netinet/ip_options.h>
76 #ifdef INET6
77 #include <netinet/ip6.h>
78 #include <netinet/icmp6.h>
79 #include <netinet6/nd6.h>
80 #include <netinet6/ip6_var.h>
81 #include <netinet6/in6_pcb.h>
82 #endif
83 #include <netinet/tcp.h>
84 #include <netinet/tcp_fastopen.h>
85 #include <netinet/tcp_fsm.h>
86 #include <netinet/tcp_seq.h>
87 #include <netinet/tcp_timer.h>
88 #include <netinet/tcp_var.h>
89 #include <netinet/tcp_syncache.h>
90 #include <netinet/tcp_ecn.h>
91 #ifdef TCP_BLACKBOX
92 #include <netinet/tcp_log_buf.h>
93 #endif
94 #ifdef TCP_OFFLOAD
95 #include <netinet/toecore.h>
96 #endif
97 #include <netinet/udp.h>
98 
99 #include <netipsec/ipsec_support.h>
100 
101 #include <machine/in_cksum.h>
102 
103 #include <security/mac/mac_framework.h>
104 
105 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1;
106 #define	V_tcp_syncookies		VNET(tcp_syncookies)
107 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
108     &VNET_NAME(tcp_syncookies), 0,
109     "Use TCP SYN cookies if the syncache overflows");
110 
111 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0;
112 #define	V_tcp_syncookiesonly		VNET(tcp_syncookiesonly)
113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
114     &VNET_NAME(tcp_syncookiesonly), 0,
115     "Use only TCP SYN cookies");
116 
117 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1;
118 #define V_functions_inherit_listen_socket_stack \
119     VNET(functions_inherit_listen_socket_stack)
120 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack,
121     CTLFLAG_VNET | CTLFLAG_RW,
122     &VNET_NAME(functions_inherit_listen_socket_stack), 0,
123     "Inherit listen socket's stack");
124 
125 #ifdef TCP_OFFLOAD
126 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
127 #endif
128 
129 static void	 syncache_drop(struct syncache *, struct syncache_head *);
130 static void	 syncache_free(struct syncache *);
131 static void	 syncache_insert(struct syncache *, struct syncache_head *);
132 static int	 syncache_respond(struct syncache *, const struct mbuf *, int);
133 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
134 		    struct mbuf *m);
135 static void	 syncache_timeout(struct syncache *sc, struct syncache_head *sch,
136 		    int docallout);
137 static void	 syncache_timer(void *);
138 
139 static uint32_t	 syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
140 		    uint8_t *, uintptr_t);
141 static tcp_seq	 syncookie_generate(struct syncache_head *, struct syncache *);
142 static struct syncache
143 		*syncookie_lookup(struct in_conninfo *, struct syncache_head *,
144 		    struct syncache *, struct tcphdr *, struct tcpopt *,
145 		    struct socket *, uint16_t);
146 static void	syncache_pause(struct in_conninfo *);
147 static void	syncache_unpause(void *);
148 static void	 syncookie_reseed(void *);
149 #ifdef INVARIANTS
150 static int	 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
151 		    struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
152 		    struct socket *lso, uint16_t port);
153 #endif
154 
155 /*
156  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
157  * 3 retransmits corresponds to a timeout with default values of
158  * tcp_rexmit_initial * (             1 +
159  *                       tcp_backoff[1] +
160  *                       tcp_backoff[2] +
161  *                       tcp_backoff[3]) + 3 * tcp_rexmit_slop,
162  * 1000 ms * (1 + 2 + 4 + 8) +  3 * 200 ms = 15600 ms,
163  * the odds are that the user has given up attempting to connect by then.
164  */
165 #define SYNCACHE_MAXREXMTS		3
166 
167 /* Arbitrary values */
168 #define TCP_SYNCACHE_HASHSIZE		512
169 #define TCP_SYNCACHE_BUCKETLIMIT	30
170 
171 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
172 #define	V_tcp_syncache			VNET(tcp_syncache)
173 
174 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache,
175     CTLFLAG_RW | CTLFLAG_MPSAFE, 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 SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET |
194     CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0,
195     "All syncache(4) entries are visible, ignoring UID/GID, jail(2) "
196     "and mac(4) checks");
197 
198 static int
199 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
200 {
201 	int error;
202 	u_int new;
203 
204 	new = V_tcp_syncache.rexmt_limit;
205 	error = sysctl_handle_int(oidp, &new, 0, req);
206 	if ((error == 0) && (req->newptr != NULL)) {
207 		if (new > TCP_MAXRXTSHIFT)
208 			error = EINVAL;
209 		else
210 			V_tcp_syncache.rexmt_limit = new;
211 	}
212 	return (error);
213 }
214 
215 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
216     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
217     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
218     sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
219     "Limit on SYN/ACK retransmissions");
220 
221 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
222 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
223     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
224     "Send reset on socket allocation failure");
225 
226 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
227 
228 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
229 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
230 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
231 
232 /*
233  * Requires the syncache entry to be already removed from the bucket list.
234  */
235 static void
236 syncache_free(struct syncache *sc)
237 {
238 
239 	if (sc->sc_ipopts)
240 		(void) m_free(sc->sc_ipopts);
241 	if (sc->sc_cred)
242 		crfree(sc->sc_cred);
243 #ifdef MAC
244 	mac_syncache_destroy(&sc->sc_label);
245 #endif
246 
247 	uma_zfree(V_tcp_syncache.zone, sc);
248 }
249 
250 void
251 syncache_init(void)
252 {
253 	int i;
254 
255 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
256 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
257 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
258 	V_tcp_syncache.hash_secret = arc4random();
259 
260 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
261 	    &V_tcp_syncache.hashsize);
262 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
263 	    &V_tcp_syncache.bucket_limit);
264 	if (!powerof2(V_tcp_syncache.hashsize) ||
265 	    V_tcp_syncache.hashsize == 0) {
266 		printf("WARNING: syncache hash size is not a power of 2.\n");
267 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
268 	}
269 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
270 
271 	/* Set limits. */
272 	V_tcp_syncache.cache_limit =
273 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
274 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
275 	    &V_tcp_syncache.cache_limit);
276 
277 	/* Allocate the hash table. */
278 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
279 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
280 
281 #ifdef VIMAGE
282 	V_tcp_syncache.vnet = curvnet;
283 #endif
284 
285 	/* Initialize the hash buckets. */
286 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
287 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
288 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
289 			 NULL, MTX_DEF);
290 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
291 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
292 		V_tcp_syncache.hashbase[i].sch_length = 0;
293 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
294 		V_tcp_syncache.hashbase[i].sch_last_overflow =
295 		    -(SYNCOOKIE_LIFETIME + 1);
296 	}
297 
298 	/* Create the syncache entry zone. */
299 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
300 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
301 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
302 	    V_tcp_syncache.cache_limit);
303 
304 	/* Start the SYN cookie reseeder callout. */
305 	callout_init(&V_tcp_syncache.secret.reseed, 1);
306 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
307 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
308 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
309 	    syncookie_reseed, &V_tcp_syncache);
310 
311 	/* Initialize the pause machinery. */
312 	mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF);
313 	callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx,
314 	    0);
315 	V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME;
316 	V_tcp_syncache.pause_backoff = 0;
317 	V_tcp_syncache.paused = false;
318 }
319 
320 #ifdef VIMAGE
321 void
322 syncache_destroy(void)
323 {
324 	struct syncache_head *sch;
325 	struct syncache *sc, *nsc;
326 	int i;
327 
328 	/*
329 	 * Stop the re-seed timer before freeing resources.  No need to
330 	 * possibly schedule it another time.
331 	 */
332 	callout_drain(&V_tcp_syncache.secret.reseed);
333 
334 	/* Stop the SYN cache pause callout. */
335 	mtx_lock(&V_tcp_syncache.pause_mtx);
336 	if (callout_stop(&V_tcp_syncache.pause_co) == 0) {
337 		mtx_unlock(&V_tcp_syncache.pause_mtx);
338 		callout_drain(&V_tcp_syncache.pause_co);
339 	} else
340 		mtx_unlock(&V_tcp_syncache.pause_mtx);
341 
342 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
343 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
344 		sch = &V_tcp_syncache.hashbase[i];
345 		callout_drain(&sch->sch_timer);
346 
347 		SCH_LOCK(sch);
348 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
349 			syncache_drop(sc, sch);
350 		SCH_UNLOCK(sch);
351 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
352 		    ("%s: sch->sch_bucket not empty", __func__));
353 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
354 		    __func__, sch->sch_length));
355 		mtx_destroy(&sch->sch_mtx);
356 	}
357 
358 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
359 	    ("%s: cache_count not 0", __func__));
360 
361 	/* Free the allocated global resources. */
362 	uma_zdestroy(V_tcp_syncache.zone);
363 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
364 	mtx_destroy(&V_tcp_syncache.pause_mtx);
365 }
366 #endif
367 
368 /*
369  * Inserts a syncache entry into the specified bucket row.
370  * Locks and unlocks the syncache_head autonomously.
371  */
372 static void
373 syncache_insert(struct syncache *sc, struct syncache_head *sch)
374 {
375 	struct syncache *sc2;
376 
377 	SCH_LOCK(sch);
378 
379 	/*
380 	 * Make sure that we don't overflow the per-bucket limit.
381 	 * If the bucket is full, toss the oldest element.
382 	 */
383 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
384 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
385 			("sch->sch_length incorrect"));
386 		syncache_pause(&sc->sc_inc);
387 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
388 		sch->sch_last_overflow = time_uptime;
389 		syncache_drop(sc2, sch);
390 	}
391 
392 	/* Put it into the bucket. */
393 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
394 	sch->sch_length++;
395 
396 #ifdef TCP_OFFLOAD
397 	if (ADDED_BY_TOE(sc)) {
398 		struct toedev *tod = sc->sc_tod;
399 
400 		tod->tod_syncache_added(tod, sc->sc_todctx);
401 	}
402 #endif
403 
404 	/* Reinitialize the bucket row's timer. */
405 	if (sch->sch_length == 1)
406 		sch->sch_nextc = ticks + INT_MAX;
407 	syncache_timeout(sc, sch, 1);
408 
409 	SCH_UNLOCK(sch);
410 
411 	TCPSTATES_INC(TCPS_SYN_RECEIVED);
412 	TCPSTAT_INC(tcps_sc_added);
413 }
414 
415 /*
416  * Remove and free entry from syncache bucket row.
417  * Expects locked syncache head.
418  */
419 static void
420 syncache_drop(struct syncache *sc, struct syncache_head *sch)
421 {
422 
423 	SCH_LOCK_ASSERT(sch);
424 
425 	TCPSTATES_DEC(TCPS_SYN_RECEIVED);
426 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
427 	sch->sch_length--;
428 
429 #ifdef TCP_OFFLOAD
430 	if (ADDED_BY_TOE(sc)) {
431 		struct toedev *tod = sc->sc_tod;
432 
433 		tod->tod_syncache_removed(tod, sc->sc_todctx);
434 	}
435 #endif
436 
437 	syncache_free(sc);
438 }
439 
440 /*
441  * Engage/reengage time on bucket row.
442  */
443 static void
444 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
445 {
446 	int rexmt;
447 
448 	if (sc->sc_rxmits == 0)
449 		rexmt = tcp_rexmit_initial;
450 	else
451 		TCPT_RANGESET(rexmt,
452 		    tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
453 		    tcp_rexmit_min, TCPTV_REXMTMAX);
454 	sc->sc_rxttime = ticks + rexmt;
455 	sc->sc_rxmits++;
456 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
457 		sch->sch_nextc = sc->sc_rxttime;
458 		if (docallout)
459 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
460 			    syncache_timer, (void *)sch);
461 	}
462 }
463 
464 /*
465  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
466  * If we have retransmitted an entry the maximum number of times, expire it.
467  * One separate timer for each bucket row.
468  */
469 static void
470 syncache_timer(void *xsch)
471 {
472 	struct syncache_head *sch = (struct syncache_head *)xsch;
473 	struct syncache *sc, *nsc;
474 	struct epoch_tracker et;
475 	int tick = ticks;
476 	char *s;
477 	bool paused;
478 
479 	CURVNET_SET(sch->sch_sc->vnet);
480 
481 	/* NB: syncache_head has already been locked by the callout. */
482 	SCH_LOCK_ASSERT(sch);
483 
484 	/*
485 	 * In the following cycle we may remove some entries and/or
486 	 * advance some timeouts, so re-initialize the bucket timer.
487 	 */
488 	sch->sch_nextc = tick + INT_MAX;
489 
490 	/*
491 	 * If we have paused processing, unconditionally remove
492 	 * all syncache entries.
493 	 */
494 	mtx_lock(&V_tcp_syncache.pause_mtx);
495 	paused = V_tcp_syncache.paused;
496 	mtx_unlock(&V_tcp_syncache.pause_mtx);
497 
498 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
499 		if (paused) {
500 			syncache_drop(sc, sch);
501 			continue;
502 		}
503 		/*
504 		 * We do not check if the listen socket still exists
505 		 * and accept the case where the listen socket may be
506 		 * gone by the time we resend the SYN/ACK.  We do
507 		 * not expect this to happens often. If it does,
508 		 * then the RST will be sent by the time the remote
509 		 * host does the SYN/ACK->ACK.
510 		 */
511 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
512 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
513 				sch->sch_nextc = sc->sc_rxttime;
514 			continue;
515 		}
516 		if (sc->sc_rxmits > V_tcp_ecn_maxretries) {
517 			sc->sc_flags &= ~SCF_ECN_MASK;
518 		}
519 		if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
520 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
521 				log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
522 				    "giving up and removing syncache entry\n",
523 				    s, __func__);
524 				free(s, M_TCPLOG);
525 			}
526 			syncache_drop(sc, sch);
527 			TCPSTAT_INC(tcps_sc_stale);
528 			continue;
529 		}
530 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
531 			log(LOG_DEBUG, "%s; %s: Response timeout, "
532 			    "retransmitting (%u) SYN|ACK\n",
533 			    s, __func__, sc->sc_rxmits);
534 			free(s, M_TCPLOG);
535 		}
536 
537 		NET_EPOCH_ENTER(et);
538 		syncache_respond(sc, NULL, TH_SYN|TH_ACK);
539 		NET_EPOCH_EXIT(et);
540 		TCPSTAT_INC(tcps_sc_retransmitted);
541 		syncache_timeout(sc, sch, 0);
542 	}
543 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
544 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
545 			syncache_timer, (void *)(sch));
546 	CURVNET_RESTORE();
547 }
548 
549 /*
550  * Returns true if the system is only using cookies at the moment.
551  * This could be due to a sysadmin decision to only use cookies, or it
552  * could be due to the system detecting an attack.
553  */
554 static inline bool
555 syncache_cookiesonly(void)
556 {
557 
558 	return (V_tcp_syncookies && (V_tcp_syncache.paused ||
559 	    V_tcp_syncookiesonly));
560 }
561 
562 /*
563  * Find the hash bucket for the given connection.
564  */
565 static struct syncache_head *
566 syncache_hashbucket(struct in_conninfo *inc)
567 {
568 	uint32_t hash;
569 
570 	/*
571 	 * The hash is built on foreign port + local port + foreign address.
572 	 * We rely on the fact that struct in_conninfo starts with 16 bits
573 	 * of foreign port, then 16 bits of local port then followed by 128
574 	 * bits of foreign address.  In case of IPv4 address, the first 3
575 	 * 32-bit words of the address always are zeroes.
576 	 */
577 	hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
578 	    V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
579 
580 	return (&V_tcp_syncache.hashbase[hash]);
581 }
582 
583 /*
584  * Find an entry in the syncache.
585  * Returns always with locked syncache_head plus a matching entry or NULL.
586  */
587 static struct syncache *
588 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
589 {
590 	struct syncache *sc;
591 	struct syncache_head *sch;
592 
593 	*schp = sch = syncache_hashbucket(inc);
594 	SCH_LOCK(sch);
595 
596 	/* Circle through bucket row to find matching entry. */
597 	TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
598 		if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
599 		    sizeof(struct in_endpoints)) == 0)
600 			break;
601 
602 	return (sc);	/* Always returns with locked sch. */
603 }
604 
605 /*
606  * This function is called when we get a RST for a
607  * non-existent connection, so that we can see if the
608  * connection is in the syn cache.  If it is, zap it.
609  * If required send a challenge ACK.
610  */
611 void
612 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m,
613     uint16_t port)
614 {
615 	struct syncache *sc;
616 	struct syncache_head *sch;
617 	char *s = NULL;
618 
619 	if (syncache_cookiesonly())
620 		return;
621 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
622 	SCH_LOCK_ASSERT(sch);
623 
624 	/*
625 	 * No corresponding connection was found in syncache.
626 	 * If syncookies are enabled and possibly exclusively
627 	 * used, or we are under memory pressure, a valid RST
628 	 * may not find a syncache entry.  In that case we're
629 	 * done and no SYN|ACK retransmissions will happen.
630 	 * Otherwise the RST was misdirected or spoofed.
631 	 */
632 	if (sc == NULL) {
633 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
634 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
635 			    "syncache entry (possibly syncookie only), "
636 			    "segment ignored\n", s, __func__);
637 		TCPSTAT_INC(tcps_badrst);
638 		goto done;
639 	}
640 
641 	/* The remote UDP encaps port does not match. */
642 	if (sc->sc_port != port) {
643 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
644 			log(LOG_DEBUG, "%s; %s: Spurious RST with matching "
645 			    "syncache entry but non-matching UDP encaps port, "
646 			    "segment ignored\n", s, __func__);
647 		TCPSTAT_INC(tcps_badrst);
648 		goto done;
649 	}
650 
651 	/*
652 	 * If the RST bit is set, check the sequence number to see
653 	 * if this is a valid reset segment.
654 	 *
655 	 * RFC 793 page 37:
656 	 *   In all states except SYN-SENT, all reset (RST) segments
657 	 *   are validated by checking their SEQ-fields.  A reset is
658 	 *   valid if its sequence number is in the window.
659 	 *
660 	 * RFC 793 page 69:
661 	 *   There are four cases for the acceptability test for an incoming
662 	 *   segment:
663 	 *
664 	 * Segment Receive  Test
665 	 * Length  Window
666 	 * ------- -------  -------------------------------------------
667 	 *    0       0     SEG.SEQ = RCV.NXT
668 	 *    0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
669 	 *   >0       0     not acceptable
670 	 *   >0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
671 	 *               or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
672 	 *
673 	 * Note that when receiving a SYN segment in the LISTEN state,
674 	 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
675 	 * described in RFC 793, page 66.
676 	 */
677 	if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
678 	    SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
679 	    (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
680 		if (V_tcp_insecure_rst ||
681 		    th->th_seq == sc->sc_irs + 1) {
682 			syncache_drop(sc, sch);
683 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
684 				log(LOG_DEBUG,
685 				    "%s; %s: Our SYN|ACK was rejected, "
686 				    "connection attempt aborted by remote "
687 				    "endpoint\n",
688 				    s, __func__);
689 			TCPSTAT_INC(tcps_sc_reset);
690 		} else {
691 			TCPSTAT_INC(tcps_badrst);
692 			/* Send challenge ACK. */
693 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
694 				log(LOG_DEBUG, "%s; %s: RST with invalid "
695 				    " SEQ %u != NXT %u (+WND %u), "
696 				    "sending challenge ACK\n",
697 				    s, __func__,
698 				    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
699 			syncache_respond(sc, m, TH_ACK);
700 		}
701 	} else {
702 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
703 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
704 			    "NXT %u (+WND %u), segment ignored\n",
705 			    s, __func__,
706 			    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
707 		TCPSTAT_INC(tcps_badrst);
708 	}
709 
710 done:
711 	if (s != NULL)
712 		free(s, M_TCPLOG);
713 	SCH_UNLOCK(sch);
714 }
715 
716 void
717 syncache_badack(struct in_conninfo *inc, uint16_t port)
718 {
719 	struct syncache *sc;
720 	struct syncache_head *sch;
721 
722 	if (syncache_cookiesonly())
723 		return;
724 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
725 	SCH_LOCK_ASSERT(sch);
726 	if ((sc != NULL) && (sc->sc_port == port)) {
727 		syncache_drop(sc, sch);
728 		TCPSTAT_INC(tcps_sc_badack);
729 	}
730 	SCH_UNLOCK(sch);
731 }
732 
733 void
734 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port)
735 {
736 	struct syncache *sc;
737 	struct syncache_head *sch;
738 
739 	if (syncache_cookiesonly())
740 		return;
741 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
742 	SCH_LOCK_ASSERT(sch);
743 	if (sc == NULL)
744 		goto done;
745 
746 	/* If the port != sc_port, then it's a bogus ICMP msg */
747 	if (port != sc->sc_port)
748 		goto done;
749 
750 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
751 	if (ntohl(th_seq) != sc->sc_iss)
752 		goto done;
753 
754 	/*
755 	 * If we've rertransmitted 3 times and this is our second error,
756 	 * we remove the entry.  Otherwise, we allow it to continue on.
757 	 * This prevents us from incorrectly nuking an entry during a
758 	 * spurious network outage.
759 	 *
760 	 * See tcp_notify().
761 	 */
762 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
763 		sc->sc_flags |= SCF_UNREACH;
764 		goto done;
765 	}
766 	syncache_drop(sc, sch);
767 	TCPSTAT_INC(tcps_sc_unreach);
768 done:
769 	SCH_UNLOCK(sch);
770 }
771 
772 /*
773  * Build a new TCP socket structure from a syncache entry.
774  *
775  * On success return the newly created socket with its underlying inp locked.
776  */
777 static struct socket *
778 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
779 {
780 	struct tcpcb *listening_tcb;
781 	struct inpcb *inp = NULL;
782 	struct socket *so;
783 	struct tcpcb *tp;
784 	int error;
785 	char *s;
786 
787 	NET_EPOCH_ASSERT();
788 
789 	/*
790 	 * Ok, create the full blown connection, and set things up
791 	 * as they would have been set up if we had created the
792 	 * connection when the SYN arrived.
793 	 */
794 	if ((so = solisten_clone(lso)) == NULL)
795 		goto allocfail;
796 #ifdef MAC
797 	mac_socketpeer_set_from_mbuf(m, so);
798 #endif
799 	error = in_pcballoc(so, &V_tcbinfo);
800 	if (error) {
801 		sodealloc(so);
802 		goto allocfail;
803 	}
804 	inp = sotoinpcb(so);
805 	if (V_functions_inherit_listen_socket_stack)
806 		listening_tcb = sototcpcb(lso);
807 	else
808 		listening_tcb = NULL;
809 	if ((tp = tcp_newtcpcb(inp, listening_tcb)) == NULL) {
810 		in_pcbfree(inp);
811 		sodealloc(so);
812 		goto allocfail;
813 	}
814 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
815 #ifdef INET6
816 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
817 		inp->inp_vflag &= ~INP_IPV4;
818 		inp->inp_vflag |= INP_IPV6;
819 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
820 	} else {
821 		inp->inp_vflag &= ~INP_IPV6;
822 		inp->inp_vflag |= INP_IPV4;
823 #endif
824 		inp->inp_ip_ttl = sc->sc_ip_ttl;
825 		inp->inp_ip_tos = sc->sc_ip_tos;
826 		inp->inp_laddr = sc->sc_inc.inc_laddr;
827 #ifdef INET6
828 	}
829 #endif
830 
831 	/*
832 	 * If there's an mbuf and it has a flowid, then let's initialise the
833 	 * inp with that particular flowid.
834 	 */
835 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
836 		inp->inp_flowid = m->m_pkthdr.flowid;
837 		inp->inp_flowtype = M_HASHTYPE_GET(m);
838 #ifdef NUMA
839 		inp->inp_numa_domain = m->m_pkthdr.numa_domain;
840 #endif
841 	}
842 
843 	inp->inp_lport = sc->sc_inc.inc_lport;
844 #ifdef INET6
845 	if (inp->inp_vflag & INP_IPV6PROTO) {
846 		struct inpcb *oinp = sotoinpcb(lso);
847 
848 		/*
849 		 * Inherit socket options from the listening socket.
850 		 * Note that in6p_inputopts are not (and should not be)
851 		 * copied, since it stores previously received options and is
852 		 * used to detect if each new option is different than the
853 		 * previous one and hence should be passed to a user.
854 		 * If we copied in6p_inputopts, a user would not be able to
855 		 * receive options just after calling the accept system call.
856 		 */
857 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
858 		if (oinp->in6p_outputopts)
859 			inp->in6p_outputopts =
860 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
861 		inp->in6p_hops = oinp->in6p_hops;
862 	}
863 
864 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
865 		struct sockaddr_in6 sin6;
866 
867 		sin6.sin6_family = AF_INET6;
868 		sin6.sin6_len = sizeof(sin6);
869 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
870 		sin6.sin6_port = sc->sc_inc.inc_fport;
871 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
872 		INP_HASH_WLOCK(&V_tcbinfo);
873 		error = in6_pcbconnect(inp, &sin6, thread0.td_ucred, false);
874 		INP_HASH_WUNLOCK(&V_tcbinfo);
875 		if (error != 0)
876 			goto abort;
877 		/* Override flowlabel from in6_pcbconnect. */
878 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
879 		inp->inp_flow |= sc->sc_flowlabel;
880 	}
881 #endif /* INET6 */
882 #if defined(INET) && defined(INET6)
883 	else
884 #endif
885 #ifdef INET
886 	{
887 		struct sockaddr_in sin;
888 
889 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
890 
891 		if (inp->inp_options == NULL) {
892 			inp->inp_options = sc->sc_ipopts;
893 			sc->sc_ipopts = NULL;
894 		}
895 
896 		sin.sin_family = AF_INET;
897 		sin.sin_len = sizeof(sin);
898 		sin.sin_addr = sc->sc_inc.inc_faddr;
899 		sin.sin_port = sc->sc_inc.inc_fport;
900 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
901 		INP_HASH_WLOCK(&V_tcbinfo);
902 		error = in_pcbconnect(inp, &sin, thread0.td_ucred, false);
903 		INP_HASH_WUNLOCK(&V_tcbinfo);
904 		if (error != 0)
905 			goto abort;
906 	}
907 #endif /* INET */
908 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
909 	/* Copy old policy into new socket's. */
910 	if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
911 		printf("syncache_socket: could not copy policy\n");
912 #endif
913 	tp->t_state = TCPS_SYN_RECEIVED;
914 	tp->iss = sc->sc_iss;
915 	tp->irs = sc->sc_irs;
916 	tp->t_port = sc->sc_port;
917 	tcp_rcvseqinit(tp);
918 	tcp_sendseqinit(tp);
919 	tp->snd_wl1 = sc->sc_irs;
920 	tp->snd_max = tp->iss + 1;
921 	tp->snd_nxt = tp->iss + 1;
922 	tp->rcv_up = sc->sc_irs + 1;
923 	tp->rcv_wnd = sc->sc_wnd;
924 	tp->rcv_adv += tp->rcv_wnd;
925 	tp->last_ack_sent = tp->rcv_nxt;
926 
927 	tp->t_flags = sototcpcb(lso)->t_flags &
928 	    (TF_LRD|TF_NOPUSH|TF_NODELAY);
929 	if (sc->sc_flags & SCF_NOOPT)
930 		tp->t_flags |= TF_NOOPT;
931 	else {
932 		if (sc->sc_flags & SCF_WINSCALE) {
933 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
934 			tp->snd_scale = sc->sc_requested_s_scale;
935 			tp->request_r_scale = sc->sc_requested_r_scale;
936 		}
937 		if (sc->sc_flags & SCF_TIMESTAMP) {
938 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
939 			tp->ts_recent = sc->sc_tsreflect;
940 			tp->ts_recent_age = tcp_ts_getticks();
941 			tp->ts_offset = sc->sc_tsoff;
942 		}
943 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
944 		if (sc->sc_flags & SCF_SIGNATURE)
945 			tp->t_flags |= TF_SIGNATURE;
946 #endif
947 		if (sc->sc_flags & SCF_SACK)
948 			tp->t_flags |= TF_SACK_PERMIT;
949 	}
950 
951 	tcp_ecn_syncache_socket(tp, sc);
952 
953 	/*
954 	 * Set up MSS and get cached values from tcp_hostcache.
955 	 * This might overwrite some of the defaults we just set.
956 	 */
957 	tcp_mss(tp, sc->sc_peer_mss);
958 
959 	/*
960 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
961 	 * limited to one segment in cc_conn_init().
962 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
963 	 */
964 	if (sc->sc_rxmits > 1)
965 		tp->snd_cwnd = 1;
966 
967 #ifdef TCP_OFFLOAD
968 	/*
969 	 * Allow a TOE driver to install its hooks.  Note that we hold the
970 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
971 	 * new connection before the TOE driver has done its thing.
972 	 */
973 	if (ADDED_BY_TOE(sc)) {
974 		struct toedev *tod = sc->sc_tod;
975 
976 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
977 	}
978 #endif
979 #ifdef TCP_BLACKBOX
980 	/*
981 	 * Inherit the log state from the listening socket, if
982 	 * - the log state of the listening socket is not off and
983 	 * - the listening socket was not auto selected from all sessions and
984 	 * - a log id is not set on the listening socket.
985 	 * This avoids inheriting a log state which was automatically set.
986 	 */
987 	if ((tcp_get_bblog_state(sototcpcb(lso)) != TCP_LOG_STATE_OFF) &&
988 	    ((sototcpcb(lso)->t_flags2 & TF2_LOG_AUTO) == 0) &&
989 	    (sototcpcb(lso)->t_lib == NULL)) {
990 		tcp_log_state_change(tp, tcp_get_bblog_state(sototcpcb(lso)));
991 	}
992 #endif
993 	/*
994 	 * Copy and activate timers.
995 	 */
996 	tp->t_maxunacktime = sototcpcb(lso)->t_maxunacktime;
997 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
998 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
999 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
1000 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1001 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1002 
1003 	TCPSTAT_INC(tcps_accepts);
1004 	TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, TCPS_LISTEN);
1005 
1006 	if (!solisten_enqueue(so, SS_ISCONNECTED))
1007 		tp->t_flags |= TF_SONOTCONN;
1008 	/* Can we inherit anything from the listener? */
1009 	if (tp->t_fb->tfb_inherit != NULL) {
1010 		(*tp->t_fb->tfb_inherit)(tp, sotoinpcb(lso));
1011 	}
1012 	return (so);
1013 
1014 allocfail:
1015 	/*
1016 	 * Drop the connection; we will either send a RST or have the peer
1017 	 * retransmit its SYN again after its RTO and try again.
1018 	 */
1019 	if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1020 		log(LOG_DEBUG, "%s; %s: Socket create failed "
1021 		    "due to limits or memory shortage\n",
1022 		    s, __func__);
1023 		free(s, M_TCPLOG);
1024 	}
1025 	TCPSTAT_INC(tcps_listendrop);
1026 	return (NULL);
1027 
1028 abort:
1029 	tcp_discardcb(tp);
1030 	in_pcbfree(inp);
1031 	sodealloc(so);
1032 	if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1033 		log(LOG_DEBUG, "%s; %s: in%s_pcbconnect failed with error %i\n",
1034 		    s, __func__, (sc->sc_inc.inc_flags & INC_ISIPV6) ? "6" : "",
1035 		    error);
1036 		free(s, M_TCPLOG);
1037 	}
1038 	TCPSTAT_INC(tcps_listendrop);
1039 	return (NULL);
1040 }
1041 
1042 /*
1043  * This function gets called when we receive an ACK for a
1044  * socket in the LISTEN state.  We look up the connection
1045  * in the syncache, and if its there, we pull it out of
1046  * the cache and turn it into a full-blown connection in
1047  * the SYN-RECEIVED state.
1048  *
1049  * On syncache_socket() success the newly created socket
1050  * has its underlying inp locked.
1051  */
1052 int
1053 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1054     struct socket **lsop, struct mbuf *m, uint16_t port)
1055 {
1056 	struct syncache *sc;
1057 	struct syncache_head *sch;
1058 	struct syncache scs;
1059 	char *s;
1060 	bool locked;
1061 
1062 	NET_EPOCH_ASSERT();
1063 	KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1064 	    ("%s: can handle only ACK", __func__));
1065 
1066 	if (syncache_cookiesonly()) {
1067 		sc = NULL;
1068 		sch = syncache_hashbucket(inc);
1069 		locked = false;
1070 	} else {
1071 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1072 		locked = true;
1073 		SCH_LOCK_ASSERT(sch);
1074 	}
1075 
1076 #ifdef INVARIANTS
1077 	/*
1078 	 * Test code for syncookies comparing the syncache stored
1079 	 * values with the reconstructed values from the cookie.
1080 	 */
1081 	if (sc != NULL)
1082 		syncookie_cmp(inc, sch, sc, th, to, *lsop, port);
1083 #endif
1084 
1085 	if (sc == NULL) {
1086 		/*
1087 		 * There is no syncache entry, so see if this ACK is
1088 		 * a returning syncookie.  To do this, first:
1089 		 *  A. Check if syncookies are used in case of syncache
1090 		 *     overflows
1091 		 *  B. See if this socket has had a syncache entry dropped in
1092 		 *     the recent past. We don't want to accept a bogus
1093 		 *     syncookie if we've never received a SYN or accept it
1094 		 *     twice.
1095 		 *  C. check that the syncookie is valid.  If it is, then
1096 		 *     cobble up a fake syncache entry, and return.
1097 		 */
1098 		if (locked && !V_tcp_syncookies) {
1099 			SCH_UNLOCK(sch);
1100 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1101 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1102 				    "segment rejected (syncookies disabled)\n",
1103 				    s, __func__);
1104 			goto failed;
1105 		}
1106 		if (locked && !V_tcp_syncookiesonly &&
1107 		    sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1108 			SCH_UNLOCK(sch);
1109 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1110 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1111 				    "segment rejected (no syncache entry)\n",
1112 				    s, __func__);
1113 			goto failed;
1114 		}
1115 		bzero(&scs, sizeof(scs));
1116 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port);
1117 		if (locked)
1118 			SCH_UNLOCK(sch);
1119 		if (sc == NULL) {
1120 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1121 				log(LOG_DEBUG, "%s; %s: Segment failed "
1122 				    "SYNCOOKIE authentication, segment rejected "
1123 				    "(probably spoofed)\n", s, __func__);
1124 			goto failed;
1125 		}
1126 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1127 		/* If received ACK has MD5 signature, check it. */
1128 		if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1129 		    (!TCPMD5_ENABLED() ||
1130 		    TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1131 			/* Drop the ACK. */
1132 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1133 				log(LOG_DEBUG, "%s; %s: Segment rejected, "
1134 				    "MD5 signature doesn't match.\n",
1135 				    s, __func__);
1136 				free(s, M_TCPLOG);
1137 			}
1138 			TCPSTAT_INC(tcps_sig_err_sigopt);
1139 			return (-1); /* Do not send RST */
1140 		}
1141 #endif /* TCP_SIGNATURE */
1142 		TCPSTATES_INC(TCPS_SYN_RECEIVED);
1143 	} else {
1144 		if (sc->sc_port != port) {
1145 			SCH_UNLOCK(sch);
1146 			return (0);
1147 		}
1148 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1149 		/*
1150 		 * If listening socket requested TCP digests, check that
1151 		 * received ACK has signature and it is correct.
1152 		 * If not, drop the ACK and leave sc entry in th cache,
1153 		 * because SYN was received with correct signature.
1154 		 */
1155 		if (sc->sc_flags & SCF_SIGNATURE) {
1156 			if ((to->to_flags & TOF_SIGNATURE) == 0) {
1157 				/* No signature */
1158 				TCPSTAT_INC(tcps_sig_err_nosigopt);
1159 				SCH_UNLOCK(sch);
1160 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1161 					log(LOG_DEBUG, "%s; %s: Segment "
1162 					    "rejected, MD5 signature wasn't "
1163 					    "provided.\n", s, __func__);
1164 					free(s, M_TCPLOG);
1165 				}
1166 				return (-1); /* Do not send RST */
1167 			}
1168 			if (!TCPMD5_ENABLED() ||
1169 			    TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1170 				/* Doesn't match or no SA */
1171 				SCH_UNLOCK(sch);
1172 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1173 					log(LOG_DEBUG, "%s; %s: Segment "
1174 					    "rejected, MD5 signature doesn't "
1175 					    "match.\n", s, __func__);
1176 					free(s, M_TCPLOG);
1177 				}
1178 				return (-1); /* Do not send RST */
1179 			}
1180 		}
1181 #endif /* TCP_SIGNATURE */
1182 
1183 		/*
1184 		 * RFC 7323 PAWS: If we have a timestamp on this segment and
1185 		 * it's less than ts_recent, drop it.
1186 		 * XXXMT: RFC 7323 also requires to send an ACK.
1187 		 *        In tcp_input.c this is only done for TCP segments
1188 		 *        with user data, so be consistent here and just drop
1189 		 *        the segment.
1190 		 */
1191 		if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1192 		    TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1193 			SCH_UNLOCK(sch);
1194 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1195 				log(LOG_DEBUG,
1196 				    "%s; %s: SEG.TSval %u < TS.Recent %u, "
1197 				    "segment dropped\n", s, __func__,
1198 				    to->to_tsval, sc->sc_tsreflect);
1199 				free(s, M_TCPLOG);
1200 			}
1201 			return (-1);  /* Do not send RST */
1202 		}
1203 
1204 		/*
1205 		 * If timestamps were not negotiated during SYN/ACK and a
1206 		 * segment with a timestamp is received, ignore the
1207 		 * timestamp and process the packet normally.
1208 		 * See section 3.2 of RFC 7323.
1209 		 */
1210 		if (!(sc->sc_flags & SCF_TIMESTAMP) &&
1211 		    (to->to_flags & TOF_TS)) {
1212 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1213 				log(LOG_DEBUG, "%s; %s: Timestamp not "
1214 				    "expected, segment processed normally\n",
1215 				    s, __func__);
1216 				free(s, M_TCPLOG);
1217 				s = NULL;
1218 			}
1219 		}
1220 
1221 		/*
1222 		 * If timestamps were negotiated during SYN/ACK and a
1223 		 * segment without a timestamp is received, silently drop
1224 		 * the segment, unless the missing timestamps are tolerated.
1225 		 * See section 3.2 of RFC 7323.
1226 		 */
1227 		if ((sc->sc_flags & SCF_TIMESTAMP) &&
1228 		    !(to->to_flags & TOF_TS)) {
1229 			if (V_tcp_tolerate_missing_ts) {
1230 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1231 					log(LOG_DEBUG,
1232 					    "%s; %s: Timestamp missing, "
1233 					    "segment processed normally\n",
1234 					    s, __func__);
1235 					free(s, M_TCPLOG);
1236 				}
1237 			} else {
1238 				SCH_UNLOCK(sch);
1239 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1240 					log(LOG_DEBUG,
1241 					    "%s; %s: Timestamp missing, "
1242 					    "segment silently dropped\n",
1243 					    s, __func__);
1244 					free(s, M_TCPLOG);
1245 				}
1246 				return (-1);  /* Do not send RST */
1247 			}
1248 		}
1249 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1250 		sch->sch_length--;
1251 #ifdef TCP_OFFLOAD
1252 		if (ADDED_BY_TOE(sc)) {
1253 			struct toedev *tod = sc->sc_tod;
1254 
1255 			tod->tod_syncache_removed(tod, sc->sc_todctx);
1256 		}
1257 #endif
1258 		SCH_UNLOCK(sch);
1259 	}
1260 
1261 	/*
1262 	 * Segment validation:
1263 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1264 	 */
1265 	if (th->th_ack != sc->sc_iss + 1) {
1266 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1267 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1268 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1269 		goto failed;
1270 	}
1271 
1272 	/*
1273 	 * The SEQ must fall in the window starting at the received
1274 	 * initial receive sequence number + 1 (the SYN).
1275 	 */
1276 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1277 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1278 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1279 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1280 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1281 		goto failed;
1282 	}
1283 
1284 	*lsop = syncache_socket(sc, *lsop, m);
1285 
1286 	if (__predict_false(*lsop == NULL)) {
1287 		TCPSTAT_INC(tcps_sc_aborted);
1288 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1289 	} else
1290 		TCPSTAT_INC(tcps_sc_completed);
1291 
1292 /* how do we find the inp for the new socket? */
1293 	if (sc != &scs)
1294 		syncache_free(sc);
1295 	return (1);
1296 failed:
1297 	if (sc != NULL) {
1298 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1299 		if (sc != &scs)
1300 			syncache_free(sc);
1301 	}
1302 	if (s != NULL)
1303 		free(s, M_TCPLOG);
1304 	*lsop = NULL;
1305 	return (0);
1306 }
1307 
1308 static struct socket *
1309 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m,
1310     uint64_t response_cookie)
1311 {
1312 	struct inpcb *inp;
1313 	struct tcpcb *tp;
1314 	unsigned int *pending_counter;
1315 	struct socket *so;
1316 
1317 	NET_EPOCH_ASSERT();
1318 
1319 	pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending;
1320 	so = syncache_socket(sc, lso, m);
1321 	if (so == NULL) {
1322 		TCPSTAT_INC(tcps_sc_aborted);
1323 		atomic_subtract_int(pending_counter, 1);
1324 	} else {
1325 		soisconnected(so);
1326 		inp = sotoinpcb(so);
1327 		tp = intotcpcb(inp);
1328 		tp->t_flags |= TF_FASTOPEN;
1329 		tp->t_tfo_cookie.server = response_cookie;
1330 		tp->snd_max = tp->iss;
1331 		tp->snd_nxt = tp->iss;
1332 		tp->t_tfo_pending = pending_counter;
1333 		TCPSTATES_INC(TCPS_SYN_RECEIVED);
1334 		TCPSTAT_INC(tcps_sc_completed);
1335 	}
1336 
1337 	return (so);
1338 }
1339 
1340 /*
1341  * Given a LISTEN socket and an inbound SYN request, add
1342  * this to the syn cache, and send back a segment:
1343  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1344  * to the source.
1345  *
1346  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1347  * Doing so would require that we hold onto the data and deliver it
1348  * to the application.  However, if we are the target of a SYN-flood
1349  * DoS attack, an attacker could send data which would eventually
1350  * consume all available buffer space if it were ACKed.  By not ACKing
1351  * the data, we avoid this DoS scenario.
1352  *
1353  * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1354  * cookie is processed and a new socket is created.  In this case, any data
1355  * accompanying the SYN will be queued to the socket by tcp_input() and will
1356  * be ACKed either when the application sends response data or the delayed
1357  * ACK timer expires, whichever comes first.
1358  */
1359 struct socket *
1360 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1361     struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod,
1362     void *todctx, uint8_t iptos, uint16_t port)
1363 {
1364 	struct tcpcb *tp;
1365 	struct socket *rv = NULL;
1366 	struct syncache *sc = NULL;
1367 	struct syncache_head *sch;
1368 	struct mbuf *ipopts = NULL;
1369 	u_int ltflags;
1370 	int win, ip_ttl, ip_tos;
1371 	char *s;
1372 #ifdef INET6
1373 	int autoflowlabel = 0;
1374 #endif
1375 #ifdef MAC
1376 	struct label *maclabel;
1377 #endif
1378 	struct syncache scs;
1379 	struct ucred *cred;
1380 	uint64_t tfo_response_cookie;
1381 	unsigned int *tfo_pending = NULL;
1382 	int tfo_cookie_valid = 0;
1383 	int tfo_response_cookie_valid = 0;
1384 	bool locked;
1385 
1386 	INP_RLOCK_ASSERT(inp);			/* listen socket */
1387 	KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1388 	    ("%s: unexpected tcp flags", __func__));
1389 
1390 	/*
1391 	 * Combine all so/tp operations very early to drop the INP lock as
1392 	 * soon as possible.
1393 	 */
1394 	KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1395 	tp = sototcpcb(so);
1396 	cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred);
1397 
1398 #ifdef INET6
1399 	if (inc->inc_flags & INC_ISIPV6) {
1400 		if (inp->inp_flags & IN6P_AUTOFLOWLABEL) {
1401 			autoflowlabel = 1;
1402 		}
1403 		ip_ttl = in6_selecthlim(inp, NULL);
1404 		if ((inp->in6p_outputopts == NULL) ||
1405 		    (inp->in6p_outputopts->ip6po_tclass == -1)) {
1406 			ip_tos = 0;
1407 		} else {
1408 			ip_tos = inp->in6p_outputopts->ip6po_tclass;
1409 		}
1410 	}
1411 #endif
1412 #if defined(INET6) && defined(INET)
1413 	else
1414 #endif
1415 #ifdef INET
1416 	{
1417 		ip_ttl = inp->inp_ip_ttl;
1418 		ip_tos = inp->inp_ip_tos;
1419 	}
1420 #endif
1421 	win = so->sol_sbrcv_hiwat;
1422 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1423 
1424 	if (V_tcp_fastopen_server_enable && (tp->t_flags & TF_FASTOPEN) &&
1425 	    (tp->t_tfo_pending != NULL) &&
1426 	    (to->to_flags & TOF_FASTOPEN)) {
1427 		/*
1428 		 * Limit the number of pending TFO connections to
1429 		 * approximately half of the queue limit.  This prevents TFO
1430 		 * SYN floods from starving the service by filling the
1431 		 * listen queue with bogus TFO connections.
1432 		 */
1433 		if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1434 		    (so->sol_qlimit / 2)) {
1435 			int result;
1436 
1437 			result = tcp_fastopen_check_cookie(inc,
1438 			    to->to_tfo_cookie, to->to_tfo_len,
1439 			    &tfo_response_cookie);
1440 			tfo_cookie_valid = (result > 0);
1441 			tfo_response_cookie_valid = (result >= 0);
1442 		}
1443 
1444 		/*
1445 		 * Remember the TFO pending counter as it will have to be
1446 		 * decremented below if we don't make it to syncache_tfo_expand().
1447 		 */
1448 		tfo_pending = tp->t_tfo_pending;
1449 	}
1450 
1451 #ifdef MAC
1452 	if (mac_syncache_init(&maclabel) != 0) {
1453 		INP_RUNLOCK(inp);
1454 		goto done;
1455 	} else
1456 		mac_syncache_create(maclabel, inp);
1457 #endif
1458 	if (!tfo_cookie_valid)
1459 		INP_RUNLOCK(inp);
1460 
1461 	/*
1462 	 * Remember the IP options, if any.
1463 	 */
1464 #ifdef INET6
1465 	if (!(inc->inc_flags & INC_ISIPV6))
1466 #endif
1467 #ifdef INET
1468 		ipopts = (m) ? ip_srcroute(m) : NULL;
1469 #else
1470 		ipopts = NULL;
1471 #endif
1472 
1473 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1474 	/*
1475 	 * When the socket is TCP-MD5 enabled check that,
1476 	 *  - a signed packet is valid
1477 	 *  - a non-signed packet does not have a security association
1478 	 *
1479 	 *  If a signed packet fails validation or a non-signed packet has a
1480 	 *  security association, the packet will be dropped.
1481 	 */
1482 	if (ltflags & TF_SIGNATURE) {
1483 		if (to->to_flags & TOF_SIGNATURE) {
1484 			if (!TCPMD5_ENABLED() ||
1485 			    TCPMD5_INPUT(m, th, to->to_signature) != 0)
1486 				goto done;
1487 		} else {
1488 			if (TCPMD5_ENABLED() &&
1489 			    TCPMD5_INPUT(m, NULL, NULL) != ENOENT)
1490 				goto done;
1491 		}
1492 	} else if (to->to_flags & TOF_SIGNATURE)
1493 		goto done;
1494 #endif	/* TCP_SIGNATURE */
1495 	/*
1496 	 * See if we already have an entry for this connection.
1497 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1498 	 *
1499 	 * XXX: should the syncache be re-initialized with the contents
1500 	 * of the new SYN here (which may have different options?)
1501 	 *
1502 	 * XXX: We do not check the sequence number to see if this is a
1503 	 * real retransmit or a new connection attempt.  The question is
1504 	 * how to handle such a case; either ignore it as spoofed, or
1505 	 * drop the current entry and create a new one?
1506 	 */
1507 	if (syncache_cookiesonly()) {
1508 		sc = NULL;
1509 		sch = syncache_hashbucket(inc);
1510 		locked = false;
1511 	} else {
1512 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1513 		locked = true;
1514 		SCH_LOCK_ASSERT(sch);
1515 	}
1516 	if (sc != NULL) {
1517 		if (tfo_cookie_valid)
1518 			INP_RUNLOCK(inp);
1519 		TCPSTAT_INC(tcps_sc_dupsyn);
1520 		if (ipopts) {
1521 			/*
1522 			 * If we were remembering a previous source route,
1523 			 * forget it and use the new one we've been given.
1524 			 */
1525 			if (sc->sc_ipopts)
1526 				(void) m_free(sc->sc_ipopts);
1527 			sc->sc_ipopts = ipopts;
1528 		}
1529 		/*
1530 		 * Update timestamp if present.
1531 		 */
1532 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1533 			sc->sc_tsreflect = to->to_tsval;
1534 		else
1535 			sc->sc_flags &= ~SCF_TIMESTAMP;
1536 		/*
1537 		 * Adjust ECN response if needed, e.g. different
1538 		 * IP ECN field, or a fallback by the remote host.
1539 		 */
1540 		if (sc->sc_flags & SCF_ECN_MASK) {
1541 			sc->sc_flags &= ~SCF_ECN_MASK;
1542 			sc->sc_flags = tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1543 		}
1544 #ifdef MAC
1545 		/*
1546 		 * Since we have already unconditionally allocated label
1547 		 * storage, free it up.  The syncache entry will already
1548 		 * have an initialized label we can use.
1549 		 */
1550 		mac_syncache_destroy(&maclabel);
1551 #endif
1552 		TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1553 		/* Retransmit SYN|ACK and reset retransmit count. */
1554 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1555 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1556 			    "resetting timer and retransmitting SYN|ACK\n",
1557 			    s, __func__);
1558 			free(s, M_TCPLOG);
1559 		}
1560 		if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1561 			sc->sc_rxmits = 0;
1562 			syncache_timeout(sc, sch, 1);
1563 			TCPSTAT_INC(tcps_sndacks);
1564 			TCPSTAT_INC(tcps_sndtotal);
1565 		}
1566 		SCH_UNLOCK(sch);
1567 		goto donenoprobe;
1568 	}
1569 
1570 	if (tfo_cookie_valid) {
1571 		bzero(&scs, sizeof(scs));
1572 		sc = &scs;
1573 		goto skip_alloc;
1574 	}
1575 
1576 	/*
1577 	 * Skip allocating a syncache entry if we are just going to discard
1578 	 * it later.
1579 	 */
1580 	if (!locked) {
1581 		bzero(&scs, sizeof(scs));
1582 		sc = &scs;
1583 	} else
1584 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1585 	if (sc == NULL) {
1586 		/*
1587 		 * The zone allocator couldn't provide more entries.
1588 		 * Treat this as if the cache was full; drop the oldest
1589 		 * entry and insert the new one.
1590 		 */
1591 		TCPSTAT_INC(tcps_sc_zonefail);
1592 		if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1593 			sch->sch_last_overflow = time_uptime;
1594 			syncache_drop(sc, sch);
1595 			syncache_pause(inc);
1596 		}
1597 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1598 		if (sc == NULL) {
1599 			if (V_tcp_syncookies) {
1600 				bzero(&scs, sizeof(scs));
1601 				sc = &scs;
1602 			} else {
1603 				KASSERT(locked,
1604 				    ("%s: bucket unexpectedly unlocked",
1605 				    __func__));
1606 				SCH_UNLOCK(sch);
1607 				if (ipopts)
1608 					(void) m_free(ipopts);
1609 				goto done;
1610 			}
1611 		}
1612 	}
1613 
1614 skip_alloc:
1615 	if (!tfo_cookie_valid && tfo_response_cookie_valid)
1616 		sc->sc_tfo_cookie = &tfo_response_cookie;
1617 
1618 	/*
1619 	 * Fill in the syncache values.
1620 	 */
1621 #ifdef MAC
1622 	sc->sc_label = maclabel;
1623 #endif
1624 	sc->sc_cred = cred;
1625 	sc->sc_port = port;
1626 	cred = NULL;
1627 	sc->sc_ipopts = ipopts;
1628 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1629 	sc->sc_ip_tos = ip_tos;
1630 	sc->sc_ip_ttl = ip_ttl;
1631 #ifdef TCP_OFFLOAD
1632 	sc->sc_tod = tod;
1633 	sc->sc_todctx = todctx;
1634 #endif
1635 	sc->sc_irs = th->th_seq;
1636 	sc->sc_flags = 0;
1637 	sc->sc_flowlabel = 0;
1638 
1639 	/*
1640 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1641 	 * win was derived from socket earlier in the function.
1642 	 */
1643 	win = imax(win, 0);
1644 	win = imin(win, TCP_MAXWIN);
1645 	sc->sc_wnd = win;
1646 
1647 	if (V_tcp_do_rfc1323 &&
1648 	    !(ltflags & TF_NOOPT)) {
1649 		/*
1650 		 * A timestamp received in a SYN makes
1651 		 * it ok to send timestamp requests and replies.
1652 		 */
1653 		if ((to->to_flags & TOF_TS) && (V_tcp_do_rfc1323 != 2)) {
1654 			sc->sc_tsreflect = to->to_tsval;
1655 			sc->sc_flags |= SCF_TIMESTAMP;
1656 			sc->sc_tsoff = tcp_new_ts_offset(inc);
1657 		}
1658 		if ((to->to_flags & TOF_SCALE) && (V_tcp_do_rfc1323 != 3)) {
1659 			int wscale = 0;
1660 
1661 			/*
1662 			 * Pick the smallest possible scaling factor that
1663 			 * will still allow us to scale up to sb_max, aka
1664 			 * kern.ipc.maxsockbuf.
1665 			 *
1666 			 * We do this because there are broken firewalls that
1667 			 * will corrupt the window scale option, leading to
1668 			 * the other endpoint believing that our advertised
1669 			 * window is unscaled.  At scale factors larger than
1670 			 * 5 the unscaled window will drop below 1500 bytes,
1671 			 * leading to serious problems when traversing these
1672 			 * broken firewalls.
1673 			 *
1674 			 * With the default maxsockbuf of 256K, a scale factor
1675 			 * of 3 will be chosen by this algorithm.  Those who
1676 			 * choose a larger maxsockbuf should watch out
1677 			 * for the compatibility problems mentioned above.
1678 			 *
1679 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1680 			 * or <SYN,ACK>) segment itself is never scaled.
1681 			 */
1682 			while (wscale < TCP_MAX_WINSHIFT &&
1683 			    (TCP_MAXWIN << wscale) < sb_max)
1684 				wscale++;
1685 			sc->sc_requested_r_scale = wscale;
1686 			sc->sc_requested_s_scale = to->to_wscale;
1687 			sc->sc_flags |= SCF_WINSCALE;
1688 		}
1689 	}
1690 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1691 	/*
1692 	 * If incoming packet has an MD5 signature, flag this in the
1693 	 * syncache so that syncache_respond() will do the right thing
1694 	 * with the SYN+ACK.
1695 	 */
1696 	if (to->to_flags & TOF_SIGNATURE)
1697 		sc->sc_flags |= SCF_SIGNATURE;
1698 #endif	/* TCP_SIGNATURE */
1699 	if (to->to_flags & TOF_SACKPERM)
1700 		sc->sc_flags |= SCF_SACK;
1701 	if (to->to_flags & TOF_MSS)
1702 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1703 	if (ltflags & TF_NOOPT)
1704 		sc->sc_flags |= SCF_NOOPT;
1705 	/* ECN Handshake */
1706 	if (V_tcp_do_ecn && (tp->t_flags2 & TF2_CANNOT_DO_ECN) == 0)
1707 		sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1708 
1709 	if (V_tcp_syncookies)
1710 		sc->sc_iss = syncookie_generate(sch, sc);
1711 	else
1712 		sc->sc_iss = arc4random();
1713 #ifdef INET6
1714 	if (autoflowlabel) {
1715 		if (V_tcp_syncookies)
1716 			sc->sc_flowlabel = sc->sc_iss;
1717 		else
1718 			sc->sc_flowlabel = ip6_randomflowlabel();
1719 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1720 	}
1721 #endif
1722 	if (locked)
1723 		SCH_UNLOCK(sch);
1724 
1725 	if (tfo_cookie_valid) {
1726 		rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie);
1727 		/* INP_RUNLOCK(inp) will be performed by the caller */
1728 		goto tfo_expanded;
1729 	}
1730 
1731 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1732 	/*
1733 	 * Do a standard 3-way handshake.
1734 	 */
1735 	if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1736 		if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1737 			syncache_free(sc);
1738 		else if (sc != &scs)
1739 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1740 		TCPSTAT_INC(tcps_sndacks);
1741 		TCPSTAT_INC(tcps_sndtotal);
1742 	} else {
1743 		if (sc != &scs)
1744 			syncache_free(sc);
1745 		TCPSTAT_INC(tcps_sc_dropped);
1746 	}
1747 	goto donenoprobe;
1748 
1749 done:
1750 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1751 donenoprobe:
1752 	if (m)
1753 		m_freem(m);
1754 	/*
1755 	 * If tfo_pending is not NULL here, then a TFO SYN that did not
1756 	 * result in a new socket was processed and the associated pending
1757 	 * counter has not yet been decremented.  All such TFO processing paths
1758 	 * transit this point.
1759 	 */
1760 	if (tfo_pending != NULL)
1761 		tcp_fastopen_decrement_counter(tfo_pending);
1762 
1763 tfo_expanded:
1764 	if (cred != NULL)
1765 		crfree(cred);
1766 #ifdef MAC
1767 	if (sc == &scs)
1768 		mac_syncache_destroy(&maclabel);
1769 #endif
1770 	return (rv);
1771 }
1772 
1773 /*
1774  * Send SYN|ACK or ACK to the peer.  Either in response to a peer's segment,
1775  * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1776  */
1777 static int
1778 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1779 {
1780 	struct ip *ip = NULL;
1781 	struct mbuf *m;
1782 	struct tcphdr *th = NULL;
1783 	struct udphdr *udp = NULL;
1784 	int optlen, error = 0;	/* Make compiler happy */
1785 	u_int16_t hlen, tlen, mssopt, ulen;
1786 	struct tcpopt to;
1787 #ifdef INET6
1788 	struct ip6_hdr *ip6 = NULL;
1789 #endif
1790 
1791 	NET_EPOCH_ASSERT();
1792 
1793 	hlen =
1794 #ifdef INET6
1795 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1796 #endif
1797 		sizeof(struct ip);
1798 	tlen = hlen + sizeof(struct tcphdr);
1799 	if (sc->sc_port) {
1800 		tlen += sizeof(struct udphdr);
1801 	}
1802 	/* Determine MSS we advertize to other end of connection. */
1803 	mssopt = tcp_mssopt(&sc->sc_inc);
1804 	if (sc->sc_port)
1805 		mssopt -= V_tcp_udp_tunneling_overhead;
1806 	mssopt = max(mssopt, V_tcp_minmss);
1807 
1808 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1809 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1810 	    ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + "
1811 	    "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port,
1812 	    max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN));
1813 
1814 	/* Create the IP+TCP header from scratch. */
1815 	m = m_gethdr(M_NOWAIT, MT_DATA);
1816 	if (m == NULL)
1817 		return (ENOBUFS);
1818 #ifdef MAC
1819 	mac_syncache_create_mbuf(sc->sc_label, m);
1820 #endif
1821 	m->m_data += max_linkhdr;
1822 	m->m_len = tlen;
1823 	m->m_pkthdr.len = tlen;
1824 	m->m_pkthdr.rcvif = NULL;
1825 
1826 #ifdef INET6
1827 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1828 		ip6 = mtod(m, struct ip6_hdr *);
1829 		ip6->ip6_vfc = IPV6_VERSION;
1830 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1831 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1832 		ip6->ip6_plen = htons(tlen - hlen);
1833 		/* ip6_hlim is set after checksum */
1834 		/* Zero out traffic class and flow label. */
1835 		ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1836 		ip6->ip6_flow |= sc->sc_flowlabel;
1837 		if (sc->sc_port != 0) {
1838 			ip6->ip6_nxt = IPPROTO_UDP;
1839 			udp = (struct udphdr *)(ip6 + 1);
1840 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1841 			udp->uh_dport = sc->sc_port;
1842 			ulen = (tlen - sizeof(struct ip6_hdr));
1843 			th = (struct tcphdr *)(udp + 1);
1844 		} else {
1845 			ip6->ip6_nxt = IPPROTO_TCP;
1846 			th = (struct tcphdr *)(ip6 + 1);
1847 		}
1848 		ip6->ip6_flow |= htonl(sc->sc_ip_tos << IPV6_FLOWLABEL_LEN);
1849 	}
1850 #endif
1851 #if defined(INET6) && defined(INET)
1852 	else
1853 #endif
1854 #ifdef INET
1855 	{
1856 		ip = mtod(m, struct ip *);
1857 		ip->ip_v = IPVERSION;
1858 		ip->ip_hl = sizeof(struct ip) >> 2;
1859 		ip->ip_len = htons(tlen);
1860 		ip->ip_id = 0;
1861 		ip->ip_off = 0;
1862 		ip->ip_sum = 0;
1863 		ip->ip_src = sc->sc_inc.inc_laddr;
1864 		ip->ip_dst = sc->sc_inc.inc_faddr;
1865 		ip->ip_ttl = sc->sc_ip_ttl;
1866 		ip->ip_tos = sc->sc_ip_tos;
1867 
1868 		/*
1869 		 * See if we should do MTU discovery.  Route lookups are
1870 		 * expensive, so we will only unset the DF bit if:
1871 		 *
1872 		 *	1) path_mtu_discovery is disabled
1873 		 *	2) the SCF_UNREACH flag has been set
1874 		 */
1875 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1876 		       ip->ip_off |= htons(IP_DF);
1877 		if (sc->sc_port == 0) {
1878 			ip->ip_p = IPPROTO_TCP;
1879 			th = (struct tcphdr *)(ip + 1);
1880 		} else {
1881 			ip->ip_p = IPPROTO_UDP;
1882 			udp = (struct udphdr *)(ip + 1);
1883 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1884 			udp->uh_dport = sc->sc_port;
1885 			ulen = (tlen - sizeof(struct ip));
1886 			th = (struct tcphdr *)(udp + 1);
1887 		}
1888 	}
1889 #endif /* INET */
1890 	th->th_sport = sc->sc_inc.inc_lport;
1891 	th->th_dport = sc->sc_inc.inc_fport;
1892 
1893 	if (flags & TH_SYN)
1894 		th->th_seq = htonl(sc->sc_iss);
1895 	else
1896 		th->th_seq = htonl(sc->sc_iss + 1);
1897 	th->th_ack = htonl(sc->sc_irs + 1);
1898 	th->th_off = sizeof(struct tcphdr) >> 2;
1899 	th->th_win = htons(sc->sc_wnd);
1900 	th->th_urp = 0;
1901 
1902 	flags = tcp_ecn_syncache_respond(flags, sc);
1903 	tcp_set_flags(th, flags);
1904 
1905 	/* Tack on the TCP options. */
1906 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1907 		to.to_flags = 0;
1908 
1909 		if (flags & TH_SYN) {
1910 			to.to_mss = mssopt;
1911 			to.to_flags = TOF_MSS;
1912 			if (sc->sc_flags & SCF_WINSCALE) {
1913 				to.to_wscale = sc->sc_requested_r_scale;
1914 				to.to_flags |= TOF_SCALE;
1915 			}
1916 			if (sc->sc_flags & SCF_SACK)
1917 				to.to_flags |= TOF_SACKPERM;
1918 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1919 			if (sc->sc_flags & SCF_SIGNATURE)
1920 				to.to_flags |= TOF_SIGNATURE;
1921 #endif
1922 			if (sc->sc_tfo_cookie) {
1923 				to.to_flags |= TOF_FASTOPEN;
1924 				to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1925 				to.to_tfo_cookie = sc->sc_tfo_cookie;
1926 				/* don't send cookie again when retransmitting response */
1927 				sc->sc_tfo_cookie = NULL;
1928 			}
1929 		}
1930 		if (sc->sc_flags & SCF_TIMESTAMP) {
1931 			to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1932 			to.to_tsecr = sc->sc_tsreflect;
1933 			to.to_flags |= TOF_TS;
1934 		}
1935 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1936 
1937 		/* Adjust headers by option size. */
1938 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1939 		m->m_len += optlen;
1940 		m->m_pkthdr.len += optlen;
1941 #ifdef INET6
1942 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1943 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1944 		else
1945 #endif
1946 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1947 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1948 		if (sc->sc_flags & SCF_SIGNATURE) {
1949 			KASSERT(to.to_flags & TOF_SIGNATURE,
1950 			    ("tcp_addoptions() didn't set tcp_signature"));
1951 
1952 			/* NOTE: to.to_signature is inside of mbuf */
1953 			if (!TCPMD5_ENABLED() ||
1954 			    TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1955 				m_freem(m);
1956 				return (EACCES);
1957 			}
1958 		}
1959 #endif
1960 	} else
1961 		optlen = 0;
1962 
1963 	if (udp) {
1964 		ulen += optlen;
1965 		udp->uh_ulen = htons(ulen);
1966 	}
1967 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
1968 	/*
1969 	 * If we have peer's SYN and it has a flowid, then let's assign it to
1970 	 * our SYN|ACK.  ip6_output() and ip_output() will not assign flowid
1971 	 * to SYN|ACK due to lack of inp here.
1972 	 */
1973 	if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1974 		m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1975 		M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1976 	}
1977 #ifdef INET6
1978 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1979 		if (sc->sc_port) {
1980 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
1981 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1982 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen,
1983 			      IPPROTO_UDP, 0);
1984 			th->th_sum = htons(0);
1985 		} else {
1986 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1987 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1988 			th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1989 			    IPPROTO_TCP, 0);
1990 		}
1991 		ip6->ip6_hlim = sc->sc_ip_ttl;
1992 #ifdef TCP_OFFLOAD
1993 		if (ADDED_BY_TOE(sc)) {
1994 			struct toedev *tod = sc->sc_tod;
1995 
1996 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1997 
1998 			return (error);
1999 		}
2000 #endif
2001 		TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
2002 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
2003 	}
2004 #endif
2005 #if defined(INET6) && defined(INET)
2006 	else
2007 #endif
2008 #ifdef INET
2009 	{
2010 		if (sc->sc_port) {
2011 			m->m_pkthdr.csum_flags = CSUM_UDP;
2012 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2013 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
2014 			      ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
2015 			th->th_sum = htons(0);
2016 		} else {
2017 			m->m_pkthdr.csum_flags = CSUM_TCP;
2018 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2019 			th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
2020 			    htons(tlen + optlen - hlen + IPPROTO_TCP));
2021 		}
2022 #ifdef TCP_OFFLOAD
2023 		if (ADDED_BY_TOE(sc)) {
2024 			struct toedev *tod = sc->sc_tod;
2025 
2026 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2027 
2028 			return (error);
2029 		}
2030 #endif
2031 		TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
2032 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
2033 	}
2034 #endif
2035 	return (error);
2036 }
2037 
2038 /*
2039  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
2040  * that exceed the capacity of the syncache by avoiding the storage of any
2041  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
2042  * attacks where the attacker does not have access to our responses.
2043  *
2044  * Syncookies encode and include all necessary information about the
2045  * connection setup within the SYN|ACK that we send back.  That way we
2046  * can avoid keeping any local state until the ACK to our SYN|ACK returns
2047  * (if ever).  Normally the syncache and syncookies are running in parallel
2048  * with the latter taking over when the former is exhausted.  When matching
2049  * syncache entry is found the syncookie is ignored.
2050  *
2051  * The only reliable information persisting the 3WHS is our initial sequence
2052  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
2053  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
2054  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
2055  * returns and signifies a legitimate connection if it matches the ACK.
2056  *
2057  * The available space of 32 bits to store the hash and to encode the SYN
2058  * option information is very tight and we should have at least 24 bits for
2059  * the MAC to keep the number of guesses by blind spoofing reasonably high.
2060  *
2061  * SYN option information we have to encode to fully restore a connection:
2062  * MSS: is imporant to chose an optimal segment size to avoid IP level
2063  *   fragmentation along the path.  The common MSS values can be encoded
2064  *   in a 3-bit table.  Uncommon values are captured by the next lower value
2065  *   in the table leading to a slight increase in packetization overhead.
2066  * WSCALE: is necessary to allow large windows to be used for high delay-
2067  *   bandwidth product links.  Not scaling the window when it was initially
2068  *   negotiated is bad for performance as lack of scaling further decreases
2069  *   the apparent available send window.  We only need to encode the WSCALE
2070  *   we received from the remote end.  Our end can be recalculated at any
2071  *   time.  The common WSCALE values can be encoded in a 3-bit table.
2072  *   Uncommon values are captured by the next lower value in the table
2073  *   making us under-estimate the available window size halving our
2074  *   theoretically possible maximum throughput for that connection.
2075  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2076  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2077  *   that are included in all segments on a connection.  We enable them when
2078  *   the ACK has them.
2079  *
2080  * Security of syncookies and attack vectors:
2081  *
2082  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2083  * together with the gloabl secret to make it unique per connection attempt.
2084  * Thus any change of any of those parameters results in a different MAC output
2085  * in an unpredictable way unless a collision is encountered.  24 bits of the
2086  * MAC are embedded into the ISS.
2087  *
2088  * To prevent replay attacks two rotating global secrets are updated with a
2089  * new random value every 15 seconds.  The life-time of a syncookie is thus
2090  * 15-30 seconds.
2091  *
2092  * Vector 1: Attacking the secret.  This requires finding a weakness in the
2093  * MAC itself or the way it is used here.  The attacker can do a chosen plain
2094  * text attack by varying and testing the all parameters under his control.
2095  * The strength depends on the size and randomness of the secret, and the
2096  * cryptographic security of the MAC function.  Due to the constant updating
2097  * of the secret the attacker has at most 29.999 seconds to find the secret
2098  * and launch spoofed connections.  After that he has to start all over again.
2099  *
2100  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
2101  * size an average of 4,823 attempts are required for a 50% chance of success
2102  * to spoof a single syncookie (birthday collision paradox).  However the
2103  * attacker is blind and doesn't know if one of his attempts succeeded unless
2104  * he has a side channel to interfere success from.  A single connection setup
2105  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2106  * This many attempts are required for each one blind spoofed connection.  For
2107  * every additional spoofed connection he has to launch another N attempts.
2108  * Thus for a sustained rate 100 spoofed connections per second approximately
2109  * 1,800,000 packets per second would have to be sent.
2110  *
2111  * NB: The MAC function should be fast so that it doesn't become a CPU
2112  * exhaustion attack vector itself.
2113  *
2114  * References:
2115  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2116  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2117  *   http://cr.yp.to/syncookies.html    (overview)
2118  *   http://cr.yp.to/syncookies/archive (details)
2119  *
2120  *
2121  * Schematic construction of a syncookie enabled Initial Sequence Number:
2122  *  0        1         2         3
2123  *  12345678901234567890123456789012
2124  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2125  *
2126  *  x 24 MAC (truncated)
2127  *  W  3 Send Window Scale index
2128  *  M  3 MSS index
2129  *  S  1 SACK permitted
2130  *  P  1 Odd/even secret
2131  */
2132 
2133 /*
2134  * Distribution and probability of certain MSS values.  Those in between are
2135  * rounded down to the next lower one.
2136  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2137  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
2138  */
2139 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2140 
2141 /*
2142  * Distribution and probability of certain WSCALE values.  We have to map the
2143  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2144  * bits based on prevalence of certain values.  Where we don't have an exact
2145  * match for are rounded down to the next lower one letting us under-estimate
2146  * the true available window.  At the moment this would happen only for the
2147  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2148  * and window size).  The absence of the WSCALE option (no scaling in either
2149  * direction) is encoded with index zero.
2150  * [WSCALE values histograms, Allman, 2012]
2151  *                            X 10 10 35  5  6 14 10%   by host
2152  *                            X 11  4  5  5 18 49  3%   by connections
2153  */
2154 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2155 
2156 /*
2157  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
2158  * and good cryptographic properties.
2159  */
2160 static uint32_t
2161 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2162     uint8_t *secbits, uintptr_t secmod)
2163 {
2164 	SIPHASH_CTX ctx;
2165 	uint32_t siphash[2];
2166 
2167 	SipHash24_Init(&ctx);
2168 	SipHash_SetKey(&ctx, secbits);
2169 	switch (inc->inc_flags & INC_ISIPV6) {
2170 #ifdef INET
2171 	case 0:
2172 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2173 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2174 		break;
2175 #endif
2176 #ifdef INET6
2177 	case INC_ISIPV6:
2178 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2179 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2180 		break;
2181 #endif
2182 	}
2183 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2184 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2185 	SipHash_Update(&ctx, &irs, sizeof(irs));
2186 	SipHash_Update(&ctx, &flags, sizeof(flags));
2187 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
2188 	SipHash_Final((u_int8_t *)&siphash, &ctx);
2189 
2190 	return (siphash[0] ^ siphash[1]);
2191 }
2192 
2193 static tcp_seq
2194 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2195 {
2196 	u_int i, secbit, wscale;
2197 	uint32_t iss, hash;
2198 	uint8_t *secbits;
2199 	union syncookie cookie;
2200 
2201 	cookie.cookie = 0;
2202 
2203 	/* Map our computed MSS into the 3-bit index. */
2204 	for (i = nitems(tcp_sc_msstab) - 1;
2205 	     tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2206 	     i--)
2207 		;
2208 	cookie.flags.mss_idx = i;
2209 
2210 	/*
2211 	 * Map the send window scale into the 3-bit index but only if
2212 	 * the wscale option was received.
2213 	 */
2214 	if (sc->sc_flags & SCF_WINSCALE) {
2215 		wscale = sc->sc_requested_s_scale;
2216 		for (i = nitems(tcp_sc_wstab) - 1;
2217 		    tcp_sc_wstab[i] > wscale && i > 0;
2218 		     i--)
2219 			;
2220 		cookie.flags.wscale_idx = i;
2221 	}
2222 
2223 	/* Can we do SACK? */
2224 	if (sc->sc_flags & SCF_SACK)
2225 		cookie.flags.sack_ok = 1;
2226 
2227 	/* Which of the two secrets to use. */
2228 	secbit = V_tcp_syncache.secret.oddeven & 0x1;
2229 	cookie.flags.odd_even = secbit;
2230 
2231 	secbits = V_tcp_syncache.secret.key[secbit];
2232 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2233 	    (uintptr_t)sch);
2234 
2235 	/*
2236 	 * Put the flags into the hash and XOR them to get better ISS number
2237 	 * variance.  This doesn't enhance the cryptographic strength and is
2238 	 * done to prevent the 8 cookie bits from showing up directly on the
2239 	 * wire.
2240 	 */
2241 	iss = hash & ~0xff;
2242 	iss |= cookie.cookie ^ (hash >> 24);
2243 
2244 	TCPSTAT_INC(tcps_sc_sendcookie);
2245 	return (iss);
2246 }
2247 
2248 static struct syncache *
2249 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2250     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2251     struct socket *lso, uint16_t port)
2252 {
2253 	uint32_t hash;
2254 	uint8_t *secbits;
2255 	tcp_seq ack, seq;
2256 	int wnd, wscale = 0;
2257 	union syncookie cookie;
2258 
2259 	/*
2260 	 * Pull information out of SYN-ACK/ACK and revert sequence number
2261 	 * advances.
2262 	 */
2263 	ack = th->th_ack - 1;
2264 	seq = th->th_seq - 1;
2265 
2266 	/*
2267 	 * Unpack the flags containing enough information to restore the
2268 	 * connection.
2269 	 */
2270 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2271 
2272 	/* Which of the two secrets to use. */
2273 	secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2274 
2275 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2276 
2277 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
2278 	if ((ack & ~0xff) != (hash & ~0xff))
2279 		return (NULL);
2280 
2281 	/* Fill in the syncache values. */
2282 	sc->sc_flags = 0;
2283 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2284 	sc->sc_ipopts = NULL;
2285 
2286 	sc->sc_irs = seq;
2287 	sc->sc_iss = ack;
2288 
2289 	switch (inc->inc_flags & INC_ISIPV6) {
2290 #ifdef INET
2291 	case 0:
2292 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2293 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2294 		break;
2295 #endif
2296 #ifdef INET6
2297 	case INC_ISIPV6:
2298 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2299 			sc->sc_flowlabel =
2300 			    htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2301 		break;
2302 #endif
2303 	}
2304 
2305 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2306 
2307 	/* We can simply recompute receive window scale we sent earlier. */
2308 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2309 		wscale++;
2310 
2311 	/* Only use wscale if it was enabled in the orignal SYN. */
2312 	if (cookie.flags.wscale_idx > 0) {
2313 		sc->sc_requested_r_scale = wscale;
2314 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2315 		sc->sc_flags |= SCF_WINSCALE;
2316 	}
2317 
2318 	wnd = lso->sol_sbrcv_hiwat;
2319 	wnd = imax(wnd, 0);
2320 	wnd = imin(wnd, TCP_MAXWIN);
2321 	sc->sc_wnd = wnd;
2322 
2323 	if (cookie.flags.sack_ok)
2324 		sc->sc_flags |= SCF_SACK;
2325 
2326 	if (to->to_flags & TOF_TS) {
2327 		sc->sc_flags |= SCF_TIMESTAMP;
2328 		sc->sc_tsreflect = to->to_tsval;
2329 		sc->sc_tsoff = tcp_new_ts_offset(inc);
2330 	}
2331 
2332 	if (to->to_flags & TOF_SIGNATURE)
2333 		sc->sc_flags |= SCF_SIGNATURE;
2334 
2335 	sc->sc_rxmits = 0;
2336 
2337 	sc->sc_port = port;
2338 
2339 	TCPSTAT_INC(tcps_sc_recvcookie);
2340 	return (sc);
2341 }
2342 
2343 #ifdef INVARIANTS
2344 static int
2345 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2346     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2347     struct socket *lso, uint16_t port)
2348 {
2349 	struct syncache scs, *scx;
2350 	char *s;
2351 
2352 	bzero(&scs, sizeof(scs));
2353 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port);
2354 
2355 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2356 		return (0);
2357 
2358 	if (scx != NULL) {
2359 		if (sc->sc_peer_mss != scx->sc_peer_mss)
2360 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2361 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2362 
2363 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2364 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2365 			    s, __func__, sc->sc_requested_r_scale,
2366 			    scx->sc_requested_r_scale);
2367 
2368 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2369 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2370 			    s, __func__, sc->sc_requested_s_scale,
2371 			    scx->sc_requested_s_scale);
2372 
2373 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2374 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2375 	}
2376 
2377 	if (s != NULL)
2378 		free(s, M_TCPLOG);
2379 	return (0);
2380 }
2381 #endif /* INVARIANTS */
2382 
2383 static void
2384 syncookie_reseed(void *arg)
2385 {
2386 	struct tcp_syncache *sc = arg;
2387 	uint8_t *secbits;
2388 	int secbit;
2389 
2390 	/*
2391 	 * Reseeding the secret doesn't have to be protected by a lock.
2392 	 * It only must be ensured that the new random values are visible
2393 	 * to all CPUs in a SMP environment.  The atomic with release
2394 	 * semantics ensures that.
2395 	 */
2396 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2397 	secbits = sc->secret.key[secbit];
2398 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2399 	atomic_add_rel_int(&sc->secret.oddeven, 1);
2400 
2401 	/* Reschedule ourself. */
2402 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2403 }
2404 
2405 /*
2406  * We have overflowed a bucket. Let's pause dealing with the syncache.
2407  * This function will increment the bucketoverflow statistics appropriately
2408  * (once per pause when pausing is enabled; otherwise, once per overflow).
2409  */
2410 static void
2411 syncache_pause(struct in_conninfo *inc)
2412 {
2413 	time_t delta;
2414 	const char *s;
2415 
2416 	/* XXX:
2417 	 * 2. Add sysctl read here so we don't get the benefit of this
2418 	 * change without the new sysctl.
2419 	 */
2420 
2421 	/*
2422 	 * Try an unlocked read. If we already know that another thread
2423 	 * has activated the feature, there is no need to proceed.
2424 	 */
2425 	if (V_tcp_syncache.paused)
2426 		return;
2427 
2428 	/* Are cookied enabled? If not, we can't pause. */
2429 	if (!V_tcp_syncookies) {
2430 		TCPSTAT_INC(tcps_sc_bucketoverflow);
2431 		return;
2432 	}
2433 
2434 	/*
2435 	 * We may be the first thread to find an overflow. Get the lock
2436 	 * and evaluate if we need to take action.
2437 	 */
2438 	mtx_lock(&V_tcp_syncache.pause_mtx);
2439 	if (V_tcp_syncache.paused) {
2440 		mtx_unlock(&V_tcp_syncache.pause_mtx);
2441 		return;
2442 	}
2443 
2444 	/* Activate protection. */
2445 	V_tcp_syncache.paused = true;
2446 	TCPSTAT_INC(tcps_sc_bucketoverflow);
2447 
2448 	/*
2449 	 * Determine the last backoff time. If we are seeing a re-newed
2450 	 * attack within that same time after last reactivating the syncache,
2451 	 * consider it an extension of the same attack.
2452 	 */
2453 	delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2454 	if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2455 		if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2456 			delta <<= 1;
2457 			V_tcp_syncache.pause_backoff++;
2458 		}
2459 	} else {
2460 		delta = TCP_SYNCACHE_PAUSE_TIME;
2461 		V_tcp_syncache.pause_backoff = 0;
2462 	}
2463 
2464 	/* Log a warning, including IP addresses, if able. */
2465 	if (inc != NULL)
2466 		s = tcp_log_addrs(inc, NULL, NULL, NULL);
2467 	else
2468 		s = (const char *)NULL;
2469 	log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2470 	    "the next %lld seconds%s%s%s\n", (long long)delta,
2471 	    (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2472 	    (s != NULL) ? ")" : "");
2473 	free(__DECONST(void *, s), M_TCPLOG);
2474 
2475 	/* Use the calculated delta to set a new pause time. */
2476 	V_tcp_syncache.pause_until = time_uptime + delta;
2477 	callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2478 	    &V_tcp_syncache);
2479 	mtx_unlock(&V_tcp_syncache.pause_mtx);
2480 }
2481 
2482 /* Evaluate whether we need to unpause. */
2483 static void
2484 syncache_unpause(void *arg)
2485 {
2486 	struct tcp_syncache *sc;
2487 	time_t delta;
2488 
2489 	sc = arg;
2490 	mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2491 	callout_deactivate(&sc->pause_co);
2492 
2493 	/*
2494 	 * Check to make sure we are not running early. If the pause
2495 	 * time has expired, then deactivate the protection.
2496 	 */
2497 	if ((delta = sc->pause_until - time_uptime) > 0)
2498 		callout_schedule(&sc->pause_co, delta * hz);
2499 	else
2500 		sc->paused = false;
2501 }
2502 
2503 /*
2504  * Exports the syncache entries to userland so that netstat can display
2505  * them alongside the other sockets.  This function is intended to be
2506  * called only from tcp_pcblist.
2507  *
2508  * Due to concurrency on an active system, the number of pcbs exported
2509  * may have no relation to max_pcbs.  max_pcbs merely indicates the
2510  * amount of space the caller allocated for this function to use.
2511  */
2512 int
2513 syncache_pcblist(struct sysctl_req *req)
2514 {
2515 	struct xtcpcb xt;
2516 	struct syncache *sc;
2517 	struct syncache_head *sch;
2518 	int error, i;
2519 
2520 	bzero(&xt, sizeof(xt));
2521 	xt.xt_len = sizeof(xt);
2522 	xt.t_state = TCPS_SYN_RECEIVED;
2523 	xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2524 	xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2525 	xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2526 	xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2527 
2528 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2529 		sch = &V_tcp_syncache.hashbase[i];
2530 		SCH_LOCK(sch);
2531 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2532 			if (sc->sc_cred != NULL &&
2533 			    cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2534 				continue;
2535 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2536 				xt.xt_inp.inp_vflag = INP_IPV6;
2537 			else
2538 				xt.xt_inp.inp_vflag = INP_IPV4;
2539 			xt.xt_encaps_port = sc->sc_port;
2540 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2541 			    sizeof (struct in_conninfo));
2542 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2543 			if (error) {
2544 				SCH_UNLOCK(sch);
2545 				return (0);
2546 			}
2547 		}
2548 		SCH_UNLOCK(sch);
2549 	}
2550 
2551 	return (0);
2552 }
2553