xref: /freebsd/sys/netinet/tcp_syncache.c (revision 59c8e88e72633afbc47a4ace0d2170d00d51f7dc)
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 tcp_function_block *blk;
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 ((tp = tcp_newtcpcb(inp)) == NULL) {
806 		in_pcbfree(inp);
807 		sodealloc(so);
808 		goto allocfail;
809 	}
810 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
811 #ifdef INET6
812 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
813 		inp->inp_vflag &= ~INP_IPV4;
814 		inp->inp_vflag |= INP_IPV6;
815 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
816 	} else {
817 		inp->inp_vflag &= ~INP_IPV6;
818 		inp->inp_vflag |= INP_IPV4;
819 #endif
820 		inp->inp_ip_ttl = sc->sc_ip_ttl;
821 		inp->inp_ip_tos = sc->sc_ip_tos;
822 		inp->inp_laddr = sc->sc_inc.inc_laddr;
823 #ifdef INET6
824 	}
825 #endif
826 
827 	/*
828 	 * If there's an mbuf and it has a flowid, then let's initialise the
829 	 * inp with that particular flowid.
830 	 */
831 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
832 		inp->inp_flowid = m->m_pkthdr.flowid;
833 		inp->inp_flowtype = M_HASHTYPE_GET(m);
834 #ifdef NUMA
835 		inp->inp_numa_domain = m->m_pkthdr.numa_domain;
836 #endif
837 	}
838 
839 	inp->inp_lport = sc->sc_inc.inc_lport;
840 #ifdef INET6
841 	if (inp->inp_vflag & INP_IPV6PROTO) {
842 		struct inpcb *oinp = sotoinpcb(lso);
843 
844 		/*
845 		 * Inherit socket options from the listening socket.
846 		 * Note that in6p_inputopts are not (and should not be)
847 		 * copied, since it stores previously received options and is
848 		 * used to detect if each new option is different than the
849 		 * previous one and hence should be passed to a user.
850 		 * If we copied in6p_inputopts, a user would not be able to
851 		 * receive options just after calling the accept system call.
852 		 */
853 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
854 		if (oinp->in6p_outputopts)
855 			inp->in6p_outputopts =
856 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
857 		inp->in6p_hops = oinp->in6p_hops;
858 	}
859 
860 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
861 		struct sockaddr_in6 sin6;
862 
863 		sin6.sin6_family = AF_INET6;
864 		sin6.sin6_len = sizeof(sin6);
865 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
866 		sin6.sin6_port = sc->sc_inc.inc_fport;
867 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
868 		INP_HASH_WLOCK(&V_tcbinfo);
869 		error = in6_pcbconnect(inp, &sin6, thread0.td_ucred, false);
870 		INP_HASH_WUNLOCK(&V_tcbinfo);
871 		if (error != 0)
872 			goto abort;
873 		/* Override flowlabel from in6_pcbconnect. */
874 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
875 		inp->inp_flow |= sc->sc_flowlabel;
876 	}
877 #endif /* INET6 */
878 #if defined(INET) && defined(INET6)
879 	else
880 #endif
881 #ifdef INET
882 	{
883 		struct sockaddr_in sin;
884 
885 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
886 
887 		if (inp->inp_options == NULL) {
888 			inp->inp_options = sc->sc_ipopts;
889 			sc->sc_ipopts = NULL;
890 		}
891 
892 		sin.sin_family = AF_INET;
893 		sin.sin_len = sizeof(sin);
894 		sin.sin_addr = sc->sc_inc.inc_faddr;
895 		sin.sin_port = sc->sc_inc.inc_fport;
896 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
897 		INP_HASH_WLOCK(&V_tcbinfo);
898 		error = in_pcbconnect(inp, &sin, thread0.td_ucred, false);
899 		INP_HASH_WUNLOCK(&V_tcbinfo);
900 		if (error != 0)
901 			goto abort;
902 	}
903 #endif /* INET */
904 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
905 	/* Copy old policy into new socket's. */
906 	if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
907 		printf("syncache_socket: could not copy policy\n");
908 #endif
909 	tp->t_state = TCPS_SYN_RECEIVED;
910 	tp->iss = sc->sc_iss;
911 	tp->irs = sc->sc_irs;
912 	tp->t_port = sc->sc_port;
913 	tcp_rcvseqinit(tp);
914 	tcp_sendseqinit(tp);
915 	blk = sototcpcb(lso)->t_fb;
916 	if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
917 		/*
918 		 * Our parents t_fb was not the default,
919 		 * we need to release our ref on tp->t_fb and
920 		 * pickup one on the new entry.
921 		 */
922 		struct tcp_function_block *rblk;
923 		void *ptr = NULL;
924 
925 		rblk = find_and_ref_tcp_fb(blk);
926 		KASSERT(rblk != NULL,
927 		    ("cannot find blk %p out of syncache?", blk));
928 
929 		if (rblk->tfb_tcp_fb_init == NULL ||
930 		    (*rblk->tfb_tcp_fb_init)(tp, &ptr) == 0) {
931 			/* Release the old stack */
932 			if (tp->t_fb->tfb_tcp_fb_fini != NULL)
933 				(*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
934 			refcount_release(&tp->t_fb->tfb_refcnt);
935 			/* Now set in all the pointers */
936 			tp->t_fb = rblk;
937 			tp->t_fb_ptr = ptr;
938 		} else {
939 			/*
940 			 * Initialization failed. Release the reference count on
941 			 * the looked up default stack.
942 			 */
943 			refcount_release(&rblk->tfb_refcnt);
944 		}
945 	}
946 	tp->snd_wl1 = sc->sc_irs;
947 	tp->snd_max = tp->iss + 1;
948 	tp->snd_nxt = tp->iss + 1;
949 	tp->rcv_up = sc->sc_irs + 1;
950 	tp->rcv_wnd = sc->sc_wnd;
951 	tp->rcv_adv += tp->rcv_wnd;
952 	tp->last_ack_sent = tp->rcv_nxt;
953 
954 	tp->t_flags = sototcpcb(lso)->t_flags &
955 	    (TF_LRD|TF_NOPUSH|TF_NODELAY);
956 	if (sc->sc_flags & SCF_NOOPT)
957 		tp->t_flags |= TF_NOOPT;
958 	else {
959 		if (sc->sc_flags & SCF_WINSCALE) {
960 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
961 			tp->snd_scale = sc->sc_requested_s_scale;
962 			tp->request_r_scale = sc->sc_requested_r_scale;
963 		}
964 		if (sc->sc_flags & SCF_TIMESTAMP) {
965 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
966 			tp->ts_recent = sc->sc_tsreflect;
967 			tp->ts_recent_age = tcp_ts_getticks();
968 			tp->ts_offset = sc->sc_tsoff;
969 		}
970 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
971 		if (sc->sc_flags & SCF_SIGNATURE)
972 			tp->t_flags |= TF_SIGNATURE;
973 #endif
974 		if (sc->sc_flags & SCF_SACK)
975 			tp->t_flags |= TF_SACK_PERMIT;
976 	}
977 
978 	tcp_ecn_syncache_socket(tp, sc);
979 
980 	/*
981 	 * Set up MSS and get cached values from tcp_hostcache.
982 	 * This might overwrite some of the defaults we just set.
983 	 */
984 	tcp_mss(tp, sc->sc_peer_mss);
985 
986 	/*
987 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
988 	 * limited to one segment in cc_conn_init().
989 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
990 	 */
991 	if (sc->sc_rxmits > 1)
992 		tp->snd_cwnd = 1;
993 
994 #ifdef TCP_OFFLOAD
995 	/*
996 	 * Allow a TOE driver to install its hooks.  Note that we hold the
997 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
998 	 * new connection before the TOE driver has done its thing.
999 	 */
1000 	if (ADDED_BY_TOE(sc)) {
1001 		struct toedev *tod = sc->sc_tod;
1002 
1003 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
1004 	}
1005 #endif
1006 #ifdef TCP_BLACKBOX
1007 	/*
1008 	 * Inherit the log state from the listening socket, if
1009 	 * - the log state of the listening socket is not off and
1010 	 * - the listening socket was not auto selected from all sessions and
1011 	 * - a log id is not set on the listening socket.
1012 	 * This avoids inheriting a log state which was automatically set.
1013 	 */
1014 	if ((tcp_get_bblog_state(sototcpcb(lso)) != TCP_LOG_STATE_OFF) &&
1015 	    ((sototcpcb(lso)->t_flags2 & TF2_LOG_AUTO) == 0) &&
1016 	    (sototcpcb(lso)->t_lib == NULL)) {
1017 		tcp_log_state_change(tp, tcp_get_bblog_state(sototcpcb(lso)));
1018 	}
1019 #endif
1020 	/*
1021 	 * Copy and activate timers.
1022 	 */
1023 	tp->t_maxunacktime = sototcpcb(lso)->t_maxunacktime;
1024 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
1025 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
1026 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
1027 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1028 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1029 
1030 	TCPSTAT_INC(tcps_accepts);
1031 	TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, TCPS_LISTEN);
1032 
1033 	if (!solisten_enqueue(so, SS_ISCONNECTED))
1034 		tp->t_flags |= TF_SONOTCONN;
1035 	/* Can we inherit anything from the listener? */
1036 	if (tp->t_fb->tfb_inherit != NULL) {
1037 		(*tp->t_fb->tfb_inherit)(tp, sotoinpcb(lso));
1038 	}
1039 	return (so);
1040 
1041 allocfail:
1042 	/*
1043 	 * Drop the connection; we will either send a RST or have the peer
1044 	 * retransmit its SYN again after its RTO and try again.
1045 	 */
1046 	if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1047 		log(LOG_DEBUG, "%s; %s: Socket create failed "
1048 		    "due to limits or memory shortage\n",
1049 		    s, __func__);
1050 		free(s, M_TCPLOG);
1051 	}
1052 	TCPSTAT_INC(tcps_listendrop);
1053 	return (NULL);
1054 
1055 abort:
1056 	in_pcbfree(inp);
1057 	sodealloc(so);
1058 	if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1059 		log(LOG_DEBUG, "%s; %s: in%s_pcbconnect failed with error %i\n",
1060 		    s, __func__, (sc->sc_inc.inc_flags & INC_ISIPV6) ? "6" : "",
1061 		    error);
1062 		free(s, M_TCPLOG);
1063 	}
1064 	TCPSTAT_INC(tcps_listendrop);
1065 	return (NULL);
1066 }
1067 
1068 /*
1069  * This function gets called when we receive an ACK for a
1070  * socket in the LISTEN state.  We look up the connection
1071  * in the syncache, and if its there, we pull it out of
1072  * the cache and turn it into a full-blown connection in
1073  * the SYN-RECEIVED state.
1074  *
1075  * On syncache_socket() success the newly created socket
1076  * has its underlying inp locked.
1077  */
1078 int
1079 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1080     struct socket **lsop, struct mbuf *m, uint16_t port)
1081 {
1082 	struct syncache *sc;
1083 	struct syncache_head *sch;
1084 	struct syncache scs;
1085 	char *s;
1086 	bool locked;
1087 
1088 	NET_EPOCH_ASSERT();
1089 	KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1090 	    ("%s: can handle only ACK", __func__));
1091 
1092 	if (syncache_cookiesonly()) {
1093 		sc = NULL;
1094 		sch = syncache_hashbucket(inc);
1095 		locked = false;
1096 	} else {
1097 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1098 		locked = true;
1099 		SCH_LOCK_ASSERT(sch);
1100 	}
1101 
1102 #ifdef INVARIANTS
1103 	/*
1104 	 * Test code for syncookies comparing the syncache stored
1105 	 * values with the reconstructed values from the cookie.
1106 	 */
1107 	if (sc != NULL)
1108 		syncookie_cmp(inc, sch, sc, th, to, *lsop, port);
1109 #endif
1110 
1111 	if (sc == NULL) {
1112 		/*
1113 		 * There is no syncache entry, so see if this ACK is
1114 		 * a returning syncookie.  To do this, first:
1115 		 *  A. Check if syncookies are used in case of syncache
1116 		 *     overflows
1117 		 *  B. See if this socket has had a syncache entry dropped in
1118 		 *     the recent past. We don't want to accept a bogus
1119 		 *     syncookie if we've never received a SYN or accept it
1120 		 *     twice.
1121 		 *  C. check that the syncookie is valid.  If it is, then
1122 		 *     cobble up a fake syncache entry, and return.
1123 		 */
1124 		if (locked && !V_tcp_syncookies) {
1125 			SCH_UNLOCK(sch);
1126 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1127 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1128 				    "segment rejected (syncookies disabled)\n",
1129 				    s, __func__);
1130 			goto failed;
1131 		}
1132 		if (locked && !V_tcp_syncookiesonly &&
1133 		    sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1134 			SCH_UNLOCK(sch);
1135 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1136 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1137 				    "segment rejected (no syncache entry)\n",
1138 				    s, __func__);
1139 			goto failed;
1140 		}
1141 		bzero(&scs, sizeof(scs));
1142 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port);
1143 		if (locked)
1144 			SCH_UNLOCK(sch);
1145 		if (sc == NULL) {
1146 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1147 				log(LOG_DEBUG, "%s; %s: Segment failed "
1148 				    "SYNCOOKIE authentication, segment rejected "
1149 				    "(probably spoofed)\n", s, __func__);
1150 			goto failed;
1151 		}
1152 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1153 		/* If received ACK has MD5 signature, check it. */
1154 		if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1155 		    (!TCPMD5_ENABLED() ||
1156 		    TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1157 			/* Drop the ACK. */
1158 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1159 				log(LOG_DEBUG, "%s; %s: Segment rejected, "
1160 				    "MD5 signature doesn't match.\n",
1161 				    s, __func__);
1162 				free(s, M_TCPLOG);
1163 			}
1164 			TCPSTAT_INC(tcps_sig_err_sigopt);
1165 			return (-1); /* Do not send RST */
1166 		}
1167 #endif /* TCP_SIGNATURE */
1168 		TCPSTATES_INC(TCPS_SYN_RECEIVED);
1169 	} else {
1170 		if (sc->sc_port != port) {
1171 			SCH_UNLOCK(sch);
1172 			return (0);
1173 		}
1174 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1175 		/*
1176 		 * If listening socket requested TCP digests, check that
1177 		 * received ACK has signature and it is correct.
1178 		 * If not, drop the ACK and leave sc entry in th cache,
1179 		 * because SYN was received with correct signature.
1180 		 */
1181 		if (sc->sc_flags & SCF_SIGNATURE) {
1182 			if ((to->to_flags & TOF_SIGNATURE) == 0) {
1183 				/* No signature */
1184 				TCPSTAT_INC(tcps_sig_err_nosigopt);
1185 				SCH_UNLOCK(sch);
1186 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1187 					log(LOG_DEBUG, "%s; %s: Segment "
1188 					    "rejected, MD5 signature wasn't "
1189 					    "provided.\n", s, __func__);
1190 					free(s, M_TCPLOG);
1191 				}
1192 				return (-1); /* Do not send RST */
1193 			}
1194 			if (!TCPMD5_ENABLED() ||
1195 			    TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1196 				/* Doesn't match or no SA */
1197 				SCH_UNLOCK(sch);
1198 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1199 					log(LOG_DEBUG, "%s; %s: Segment "
1200 					    "rejected, MD5 signature doesn't "
1201 					    "match.\n", s, __func__);
1202 					free(s, M_TCPLOG);
1203 				}
1204 				return (-1); /* Do not send RST */
1205 			}
1206 		}
1207 #endif /* TCP_SIGNATURE */
1208 
1209 		/*
1210 		 * RFC 7323 PAWS: If we have a timestamp on this segment and
1211 		 * it's less than ts_recent, drop it.
1212 		 * XXXMT: RFC 7323 also requires to send an ACK.
1213 		 *        In tcp_input.c this is only done for TCP segments
1214 		 *        with user data, so be consistent here and just drop
1215 		 *        the segment.
1216 		 */
1217 		if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1218 		    TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1219 			SCH_UNLOCK(sch);
1220 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1221 				log(LOG_DEBUG,
1222 				    "%s; %s: SEG.TSval %u < TS.Recent %u, "
1223 				    "segment dropped\n", s, __func__,
1224 				    to->to_tsval, sc->sc_tsreflect);
1225 				free(s, M_TCPLOG);
1226 			}
1227 			return (-1);  /* Do not send RST */
1228 		}
1229 
1230 		/*
1231 		 * If timestamps were not negotiated during SYN/ACK and a
1232 		 * segment with a timestamp is received, ignore the
1233 		 * timestamp and process the packet normally.
1234 		 * See section 3.2 of RFC 7323.
1235 		 */
1236 		if (!(sc->sc_flags & SCF_TIMESTAMP) &&
1237 		    (to->to_flags & TOF_TS)) {
1238 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1239 				log(LOG_DEBUG, "%s; %s: Timestamp not "
1240 				    "expected, segment processed normally\n",
1241 				    s, __func__);
1242 				free(s, M_TCPLOG);
1243 				s = NULL;
1244 			}
1245 		}
1246 
1247 		/*
1248 		 * If timestamps were negotiated during SYN/ACK and a
1249 		 * segment without a timestamp is received, silently drop
1250 		 * the segment, unless the missing timestamps are tolerated.
1251 		 * See section 3.2 of RFC 7323.
1252 		 */
1253 		if ((sc->sc_flags & SCF_TIMESTAMP) &&
1254 		    !(to->to_flags & TOF_TS)) {
1255 			if (V_tcp_tolerate_missing_ts) {
1256 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1257 					log(LOG_DEBUG,
1258 					    "%s; %s: Timestamp missing, "
1259 					    "segment processed normally\n",
1260 					    s, __func__);
1261 					free(s, M_TCPLOG);
1262 				}
1263 			} else {
1264 				SCH_UNLOCK(sch);
1265 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1266 					log(LOG_DEBUG,
1267 					    "%s; %s: Timestamp missing, "
1268 					    "segment silently dropped\n",
1269 					    s, __func__);
1270 					free(s, M_TCPLOG);
1271 				}
1272 				return (-1);  /* Do not send RST */
1273 			}
1274 		}
1275 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1276 		sch->sch_length--;
1277 #ifdef TCP_OFFLOAD
1278 		if (ADDED_BY_TOE(sc)) {
1279 			struct toedev *tod = sc->sc_tod;
1280 
1281 			tod->tod_syncache_removed(tod, sc->sc_todctx);
1282 		}
1283 #endif
1284 		SCH_UNLOCK(sch);
1285 	}
1286 
1287 	/*
1288 	 * Segment validation:
1289 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1290 	 */
1291 	if (th->th_ack != sc->sc_iss + 1) {
1292 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1293 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1294 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1295 		goto failed;
1296 	}
1297 
1298 	/*
1299 	 * The SEQ must fall in the window starting at the received
1300 	 * initial receive sequence number + 1 (the SYN).
1301 	 */
1302 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1303 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1304 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1305 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1306 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1307 		goto failed;
1308 	}
1309 
1310 	*lsop = syncache_socket(sc, *lsop, m);
1311 
1312 	if (__predict_false(*lsop == NULL)) {
1313 		TCPSTAT_INC(tcps_sc_aborted);
1314 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1315 	} else
1316 		TCPSTAT_INC(tcps_sc_completed);
1317 
1318 /* how do we find the inp for the new socket? */
1319 	if (sc != &scs)
1320 		syncache_free(sc);
1321 	return (1);
1322 failed:
1323 	if (sc != NULL) {
1324 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1325 		if (sc != &scs)
1326 			syncache_free(sc);
1327 	}
1328 	if (s != NULL)
1329 		free(s, M_TCPLOG);
1330 	*lsop = NULL;
1331 	return (0);
1332 }
1333 
1334 static struct socket *
1335 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m,
1336     uint64_t response_cookie)
1337 {
1338 	struct inpcb *inp;
1339 	struct tcpcb *tp;
1340 	unsigned int *pending_counter;
1341 	struct socket *so;
1342 
1343 	NET_EPOCH_ASSERT();
1344 
1345 	pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending;
1346 	so = syncache_socket(sc, lso, m);
1347 	if (so == NULL) {
1348 		TCPSTAT_INC(tcps_sc_aborted);
1349 		atomic_subtract_int(pending_counter, 1);
1350 	} else {
1351 		soisconnected(so);
1352 		inp = sotoinpcb(so);
1353 		tp = intotcpcb(inp);
1354 		tp->t_flags |= TF_FASTOPEN;
1355 		tp->t_tfo_cookie.server = response_cookie;
1356 		tp->snd_max = tp->iss;
1357 		tp->snd_nxt = tp->iss;
1358 		tp->t_tfo_pending = pending_counter;
1359 		TCPSTATES_INC(TCPS_SYN_RECEIVED);
1360 		TCPSTAT_INC(tcps_sc_completed);
1361 	}
1362 
1363 	return (so);
1364 }
1365 
1366 /*
1367  * Given a LISTEN socket and an inbound SYN request, add
1368  * this to the syn cache, and send back a segment:
1369  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1370  * to the source.
1371  *
1372  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1373  * Doing so would require that we hold onto the data and deliver it
1374  * to the application.  However, if we are the target of a SYN-flood
1375  * DoS attack, an attacker could send data which would eventually
1376  * consume all available buffer space if it were ACKed.  By not ACKing
1377  * the data, we avoid this DoS scenario.
1378  *
1379  * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1380  * cookie is processed and a new socket is created.  In this case, any data
1381  * accompanying the SYN will be queued to the socket by tcp_input() and will
1382  * be ACKed either when the application sends response data or the delayed
1383  * ACK timer expires, whichever comes first.
1384  */
1385 struct socket *
1386 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1387     struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod,
1388     void *todctx, uint8_t iptos, uint16_t port)
1389 {
1390 	struct tcpcb *tp;
1391 	struct socket *rv = NULL;
1392 	struct syncache *sc = NULL;
1393 	struct syncache_head *sch;
1394 	struct mbuf *ipopts = NULL;
1395 	u_int ltflags;
1396 	int win, ip_ttl, ip_tos;
1397 	char *s;
1398 #ifdef INET6
1399 	int autoflowlabel = 0;
1400 #endif
1401 #ifdef MAC
1402 	struct label *maclabel;
1403 #endif
1404 	struct syncache scs;
1405 	struct ucred *cred;
1406 	uint64_t tfo_response_cookie;
1407 	unsigned int *tfo_pending = NULL;
1408 	int tfo_cookie_valid = 0;
1409 	int tfo_response_cookie_valid = 0;
1410 	bool locked;
1411 
1412 	INP_RLOCK_ASSERT(inp);			/* listen socket */
1413 	KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1414 	    ("%s: unexpected tcp flags", __func__));
1415 
1416 	/*
1417 	 * Combine all so/tp operations very early to drop the INP lock as
1418 	 * soon as possible.
1419 	 */
1420 	KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1421 	tp = sototcpcb(so);
1422 	cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred);
1423 
1424 #ifdef INET6
1425 	if (inc->inc_flags & INC_ISIPV6) {
1426 		if (inp->inp_flags & IN6P_AUTOFLOWLABEL) {
1427 			autoflowlabel = 1;
1428 		}
1429 		ip_ttl = in6_selecthlim(inp, NULL);
1430 		if ((inp->in6p_outputopts == NULL) ||
1431 		    (inp->in6p_outputopts->ip6po_tclass == -1)) {
1432 			ip_tos = 0;
1433 		} else {
1434 			ip_tos = inp->in6p_outputopts->ip6po_tclass;
1435 		}
1436 	}
1437 #endif
1438 #if defined(INET6) && defined(INET)
1439 	else
1440 #endif
1441 #ifdef INET
1442 	{
1443 		ip_ttl = inp->inp_ip_ttl;
1444 		ip_tos = inp->inp_ip_tos;
1445 	}
1446 #endif
1447 	win = so->sol_sbrcv_hiwat;
1448 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1449 
1450 	if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
1451 	    (tp->t_tfo_pending != NULL) &&
1452 	    (to->to_flags & TOF_FASTOPEN)) {
1453 		/*
1454 		 * Limit the number of pending TFO connections to
1455 		 * approximately half of the queue limit.  This prevents TFO
1456 		 * SYN floods from starving the service by filling the
1457 		 * listen queue with bogus TFO connections.
1458 		 */
1459 		if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1460 		    (so->sol_qlimit / 2)) {
1461 			int result;
1462 
1463 			result = tcp_fastopen_check_cookie(inc,
1464 			    to->to_tfo_cookie, to->to_tfo_len,
1465 			    &tfo_response_cookie);
1466 			tfo_cookie_valid = (result > 0);
1467 			tfo_response_cookie_valid = (result >= 0);
1468 		}
1469 
1470 		/*
1471 		 * Remember the TFO pending counter as it will have to be
1472 		 * decremented below if we don't make it to syncache_tfo_expand().
1473 		 */
1474 		tfo_pending = tp->t_tfo_pending;
1475 	}
1476 
1477 #ifdef MAC
1478 	if (mac_syncache_init(&maclabel) != 0) {
1479 		INP_RUNLOCK(inp);
1480 		goto done;
1481 	} else
1482 		mac_syncache_create(maclabel, inp);
1483 #endif
1484 	if (!tfo_cookie_valid)
1485 		INP_RUNLOCK(inp);
1486 
1487 	/*
1488 	 * Remember the IP options, if any.
1489 	 */
1490 #ifdef INET6
1491 	if (!(inc->inc_flags & INC_ISIPV6))
1492 #endif
1493 #ifdef INET
1494 		ipopts = (m) ? ip_srcroute(m) : NULL;
1495 #else
1496 		ipopts = NULL;
1497 #endif
1498 
1499 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1500 	/*
1501 	 * When the socket is TCP-MD5 enabled check that,
1502 	 *  - a signed packet is valid
1503 	 *  - a non-signed packet does not have a security association
1504 	 *
1505 	 *  If a signed packet fails validation or a non-signed packet has a
1506 	 *  security association, the packet will be dropped.
1507 	 */
1508 	if (ltflags & TF_SIGNATURE) {
1509 		if (to->to_flags & TOF_SIGNATURE) {
1510 			if (!TCPMD5_ENABLED() ||
1511 			    TCPMD5_INPUT(m, th, to->to_signature) != 0)
1512 				goto done;
1513 		} else {
1514 			if (TCPMD5_ENABLED() &&
1515 			    TCPMD5_INPUT(m, NULL, NULL) != ENOENT)
1516 				goto done;
1517 		}
1518 	} else if (to->to_flags & TOF_SIGNATURE)
1519 		goto done;
1520 #endif	/* TCP_SIGNATURE */
1521 	/*
1522 	 * See if we already have an entry for this connection.
1523 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1524 	 *
1525 	 * XXX: should the syncache be re-initialized with the contents
1526 	 * of the new SYN here (which may have different options?)
1527 	 *
1528 	 * XXX: We do not check the sequence number to see if this is a
1529 	 * real retransmit or a new connection attempt.  The question is
1530 	 * how to handle such a case; either ignore it as spoofed, or
1531 	 * drop the current entry and create a new one?
1532 	 */
1533 	if (syncache_cookiesonly()) {
1534 		sc = NULL;
1535 		sch = syncache_hashbucket(inc);
1536 		locked = false;
1537 	} else {
1538 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1539 		locked = true;
1540 		SCH_LOCK_ASSERT(sch);
1541 	}
1542 	if (sc != NULL) {
1543 		if (tfo_cookie_valid)
1544 			INP_RUNLOCK(inp);
1545 		TCPSTAT_INC(tcps_sc_dupsyn);
1546 		if (ipopts) {
1547 			/*
1548 			 * If we were remembering a previous source route,
1549 			 * forget it and use the new one we've been given.
1550 			 */
1551 			if (sc->sc_ipopts)
1552 				(void) m_free(sc->sc_ipopts);
1553 			sc->sc_ipopts = ipopts;
1554 		}
1555 		/*
1556 		 * Update timestamp if present.
1557 		 */
1558 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1559 			sc->sc_tsreflect = to->to_tsval;
1560 		else
1561 			sc->sc_flags &= ~SCF_TIMESTAMP;
1562 		/*
1563 		 * Adjust ECN response if needed, e.g. different
1564 		 * IP ECN field, or a fallback by the remote host.
1565 		 */
1566 		if (sc->sc_flags & SCF_ECN_MASK) {
1567 			sc->sc_flags &= ~SCF_ECN_MASK;
1568 			sc->sc_flags = tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1569 		}
1570 #ifdef MAC
1571 		/*
1572 		 * Since we have already unconditionally allocated label
1573 		 * storage, free it up.  The syncache entry will already
1574 		 * have an initialized label we can use.
1575 		 */
1576 		mac_syncache_destroy(&maclabel);
1577 #endif
1578 		TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1579 		/* Retransmit SYN|ACK and reset retransmit count. */
1580 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1581 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1582 			    "resetting timer and retransmitting SYN|ACK\n",
1583 			    s, __func__);
1584 			free(s, M_TCPLOG);
1585 		}
1586 		if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1587 			sc->sc_rxmits = 0;
1588 			syncache_timeout(sc, sch, 1);
1589 			TCPSTAT_INC(tcps_sndacks);
1590 			TCPSTAT_INC(tcps_sndtotal);
1591 		}
1592 		SCH_UNLOCK(sch);
1593 		goto donenoprobe;
1594 	}
1595 
1596 	if (tfo_cookie_valid) {
1597 		bzero(&scs, sizeof(scs));
1598 		sc = &scs;
1599 		goto skip_alloc;
1600 	}
1601 
1602 	/*
1603 	 * Skip allocating a syncache entry if we are just going to discard
1604 	 * it later.
1605 	 */
1606 	if (!locked) {
1607 		bzero(&scs, sizeof(scs));
1608 		sc = &scs;
1609 	} else
1610 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1611 	if (sc == NULL) {
1612 		/*
1613 		 * The zone allocator couldn't provide more entries.
1614 		 * Treat this as if the cache was full; drop the oldest
1615 		 * entry and insert the new one.
1616 		 */
1617 		TCPSTAT_INC(tcps_sc_zonefail);
1618 		if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1619 			sch->sch_last_overflow = time_uptime;
1620 			syncache_drop(sc, sch);
1621 			syncache_pause(inc);
1622 		}
1623 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1624 		if (sc == NULL) {
1625 			if (V_tcp_syncookies) {
1626 				bzero(&scs, sizeof(scs));
1627 				sc = &scs;
1628 			} else {
1629 				KASSERT(locked,
1630 				    ("%s: bucket unexpectedly unlocked",
1631 				    __func__));
1632 				SCH_UNLOCK(sch);
1633 				if (ipopts)
1634 					(void) m_free(ipopts);
1635 				goto done;
1636 			}
1637 		}
1638 	}
1639 
1640 skip_alloc:
1641 	if (!tfo_cookie_valid && tfo_response_cookie_valid)
1642 		sc->sc_tfo_cookie = &tfo_response_cookie;
1643 
1644 	/*
1645 	 * Fill in the syncache values.
1646 	 */
1647 #ifdef MAC
1648 	sc->sc_label = maclabel;
1649 #endif
1650 	sc->sc_cred = cred;
1651 	sc->sc_port = port;
1652 	cred = NULL;
1653 	sc->sc_ipopts = ipopts;
1654 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1655 	sc->sc_ip_tos = ip_tos;
1656 	sc->sc_ip_ttl = ip_ttl;
1657 #ifdef TCP_OFFLOAD
1658 	sc->sc_tod = tod;
1659 	sc->sc_todctx = todctx;
1660 #endif
1661 	sc->sc_irs = th->th_seq;
1662 	sc->sc_flags = 0;
1663 	sc->sc_flowlabel = 0;
1664 
1665 	/*
1666 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1667 	 * win was derived from socket earlier in the function.
1668 	 */
1669 	win = imax(win, 0);
1670 	win = imin(win, TCP_MAXWIN);
1671 	sc->sc_wnd = win;
1672 
1673 	if (V_tcp_do_rfc1323 &&
1674 	    !(ltflags & TF_NOOPT)) {
1675 		/*
1676 		 * A timestamp received in a SYN makes
1677 		 * it ok to send timestamp requests and replies.
1678 		 */
1679 		if ((to->to_flags & TOF_TS) && (V_tcp_do_rfc1323 != 2)) {
1680 			sc->sc_tsreflect = to->to_tsval;
1681 			sc->sc_flags |= SCF_TIMESTAMP;
1682 			sc->sc_tsoff = tcp_new_ts_offset(inc);
1683 		}
1684 		if ((to->to_flags & TOF_SCALE) && (V_tcp_do_rfc1323 != 3)) {
1685 			int wscale = 0;
1686 
1687 			/*
1688 			 * Pick the smallest possible scaling factor that
1689 			 * will still allow us to scale up to sb_max, aka
1690 			 * kern.ipc.maxsockbuf.
1691 			 *
1692 			 * We do this because there are broken firewalls that
1693 			 * will corrupt the window scale option, leading to
1694 			 * the other endpoint believing that our advertised
1695 			 * window is unscaled.  At scale factors larger than
1696 			 * 5 the unscaled window will drop below 1500 bytes,
1697 			 * leading to serious problems when traversing these
1698 			 * broken firewalls.
1699 			 *
1700 			 * With the default maxsockbuf of 256K, a scale factor
1701 			 * of 3 will be chosen by this algorithm.  Those who
1702 			 * choose a larger maxsockbuf should watch out
1703 			 * for the compatibility problems mentioned above.
1704 			 *
1705 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1706 			 * or <SYN,ACK>) segment itself is never scaled.
1707 			 */
1708 			while (wscale < TCP_MAX_WINSHIFT &&
1709 			    (TCP_MAXWIN << wscale) < sb_max)
1710 				wscale++;
1711 			sc->sc_requested_r_scale = wscale;
1712 			sc->sc_requested_s_scale = to->to_wscale;
1713 			sc->sc_flags |= SCF_WINSCALE;
1714 		}
1715 	}
1716 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1717 	/*
1718 	 * If incoming packet has an MD5 signature, flag this in the
1719 	 * syncache so that syncache_respond() will do the right thing
1720 	 * with the SYN+ACK.
1721 	 */
1722 	if (to->to_flags & TOF_SIGNATURE)
1723 		sc->sc_flags |= SCF_SIGNATURE;
1724 #endif	/* TCP_SIGNATURE */
1725 	if (to->to_flags & TOF_SACKPERM)
1726 		sc->sc_flags |= SCF_SACK;
1727 	if (to->to_flags & TOF_MSS)
1728 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1729 	if (ltflags & TF_NOOPT)
1730 		sc->sc_flags |= SCF_NOOPT;
1731 	/* ECN Handshake */
1732 	if (V_tcp_do_ecn && (tp->t_flags2 & TF2_CANNOT_DO_ECN) == 0)
1733 		sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1734 
1735 	if (V_tcp_syncookies)
1736 		sc->sc_iss = syncookie_generate(sch, sc);
1737 	else
1738 		sc->sc_iss = arc4random();
1739 #ifdef INET6
1740 	if (autoflowlabel) {
1741 		if (V_tcp_syncookies)
1742 			sc->sc_flowlabel = sc->sc_iss;
1743 		else
1744 			sc->sc_flowlabel = ip6_randomflowlabel();
1745 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1746 	}
1747 #endif
1748 	if (locked)
1749 		SCH_UNLOCK(sch);
1750 
1751 	if (tfo_cookie_valid) {
1752 		rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie);
1753 		/* INP_RUNLOCK(inp) will be performed by the caller */
1754 		goto tfo_expanded;
1755 	}
1756 
1757 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1758 	/*
1759 	 * Do a standard 3-way handshake.
1760 	 */
1761 	if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1762 		if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1763 			syncache_free(sc);
1764 		else if (sc != &scs)
1765 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1766 		TCPSTAT_INC(tcps_sndacks);
1767 		TCPSTAT_INC(tcps_sndtotal);
1768 	} else {
1769 		if (sc != &scs)
1770 			syncache_free(sc);
1771 		TCPSTAT_INC(tcps_sc_dropped);
1772 	}
1773 	goto donenoprobe;
1774 
1775 done:
1776 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1777 donenoprobe:
1778 	if (m)
1779 		m_freem(m);
1780 	/*
1781 	 * If tfo_pending is not NULL here, then a TFO SYN that did not
1782 	 * result in a new socket was processed and the associated pending
1783 	 * counter has not yet been decremented.  All such TFO processing paths
1784 	 * transit this point.
1785 	 */
1786 	if (tfo_pending != NULL)
1787 		tcp_fastopen_decrement_counter(tfo_pending);
1788 
1789 tfo_expanded:
1790 	if (cred != NULL)
1791 		crfree(cred);
1792 #ifdef MAC
1793 	if (sc == &scs)
1794 		mac_syncache_destroy(&maclabel);
1795 #endif
1796 	return (rv);
1797 }
1798 
1799 /*
1800  * Send SYN|ACK or ACK to the peer.  Either in response to a peer's segment,
1801  * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1802  */
1803 static int
1804 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1805 {
1806 	struct ip *ip = NULL;
1807 	struct mbuf *m;
1808 	struct tcphdr *th = NULL;
1809 	struct udphdr *udp = NULL;
1810 	int optlen, error = 0;	/* Make compiler happy */
1811 	u_int16_t hlen, tlen, mssopt, ulen;
1812 	struct tcpopt to;
1813 #ifdef INET6
1814 	struct ip6_hdr *ip6 = NULL;
1815 #endif
1816 
1817 	NET_EPOCH_ASSERT();
1818 
1819 	hlen =
1820 #ifdef INET6
1821 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1822 #endif
1823 		sizeof(struct ip);
1824 	tlen = hlen + sizeof(struct tcphdr);
1825 	if (sc->sc_port) {
1826 		tlen += sizeof(struct udphdr);
1827 	}
1828 	/* Determine MSS we advertize to other end of connection. */
1829 	mssopt = tcp_mssopt(&sc->sc_inc);
1830 	if (sc->sc_port)
1831 		mssopt -= V_tcp_udp_tunneling_overhead;
1832 	mssopt = max(mssopt, V_tcp_minmss);
1833 
1834 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1835 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1836 	    ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + "
1837 	    "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port,
1838 	    max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN));
1839 
1840 	/* Create the IP+TCP header from scratch. */
1841 	m = m_gethdr(M_NOWAIT, MT_DATA);
1842 	if (m == NULL)
1843 		return (ENOBUFS);
1844 #ifdef MAC
1845 	mac_syncache_create_mbuf(sc->sc_label, m);
1846 #endif
1847 	m->m_data += max_linkhdr;
1848 	m->m_len = tlen;
1849 	m->m_pkthdr.len = tlen;
1850 	m->m_pkthdr.rcvif = NULL;
1851 
1852 #ifdef INET6
1853 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1854 		ip6 = mtod(m, struct ip6_hdr *);
1855 		ip6->ip6_vfc = IPV6_VERSION;
1856 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1857 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1858 		ip6->ip6_plen = htons(tlen - hlen);
1859 		/* ip6_hlim is set after checksum */
1860 		/* Zero out traffic class and flow label. */
1861 		ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1862 		ip6->ip6_flow |= sc->sc_flowlabel;
1863 		if (sc->sc_port != 0) {
1864 			ip6->ip6_nxt = IPPROTO_UDP;
1865 			udp = (struct udphdr *)(ip6 + 1);
1866 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1867 			udp->uh_dport = sc->sc_port;
1868 			ulen = (tlen - sizeof(struct ip6_hdr));
1869 			th = (struct tcphdr *)(udp + 1);
1870 		} else {
1871 			ip6->ip6_nxt = IPPROTO_TCP;
1872 			th = (struct tcphdr *)(ip6 + 1);
1873 		}
1874 		ip6->ip6_flow |= htonl(sc->sc_ip_tos << IPV6_FLOWLABEL_LEN);
1875 	}
1876 #endif
1877 #if defined(INET6) && defined(INET)
1878 	else
1879 #endif
1880 #ifdef INET
1881 	{
1882 		ip = mtod(m, struct ip *);
1883 		ip->ip_v = IPVERSION;
1884 		ip->ip_hl = sizeof(struct ip) >> 2;
1885 		ip->ip_len = htons(tlen);
1886 		ip->ip_id = 0;
1887 		ip->ip_off = 0;
1888 		ip->ip_sum = 0;
1889 		ip->ip_src = sc->sc_inc.inc_laddr;
1890 		ip->ip_dst = sc->sc_inc.inc_faddr;
1891 		ip->ip_ttl = sc->sc_ip_ttl;
1892 		ip->ip_tos = sc->sc_ip_tos;
1893 
1894 		/*
1895 		 * See if we should do MTU discovery.  Route lookups are
1896 		 * expensive, so we will only unset the DF bit if:
1897 		 *
1898 		 *	1) path_mtu_discovery is disabled
1899 		 *	2) the SCF_UNREACH flag has been set
1900 		 */
1901 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1902 		       ip->ip_off |= htons(IP_DF);
1903 		if (sc->sc_port == 0) {
1904 			ip->ip_p = IPPROTO_TCP;
1905 			th = (struct tcphdr *)(ip + 1);
1906 		} else {
1907 			ip->ip_p = IPPROTO_UDP;
1908 			udp = (struct udphdr *)(ip + 1);
1909 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1910 			udp->uh_dport = sc->sc_port;
1911 			ulen = (tlen - sizeof(struct ip));
1912 			th = (struct tcphdr *)(udp + 1);
1913 		}
1914 	}
1915 #endif /* INET */
1916 	th->th_sport = sc->sc_inc.inc_lport;
1917 	th->th_dport = sc->sc_inc.inc_fport;
1918 
1919 	if (flags & TH_SYN)
1920 		th->th_seq = htonl(sc->sc_iss);
1921 	else
1922 		th->th_seq = htonl(sc->sc_iss + 1);
1923 	th->th_ack = htonl(sc->sc_irs + 1);
1924 	th->th_off = sizeof(struct tcphdr) >> 2;
1925 	th->th_win = htons(sc->sc_wnd);
1926 	th->th_urp = 0;
1927 
1928 	flags = tcp_ecn_syncache_respond(flags, sc);
1929 	tcp_set_flags(th, flags);
1930 
1931 	/* Tack on the TCP options. */
1932 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1933 		to.to_flags = 0;
1934 
1935 		if (flags & TH_SYN) {
1936 			to.to_mss = mssopt;
1937 			to.to_flags = TOF_MSS;
1938 			if (sc->sc_flags & SCF_WINSCALE) {
1939 				to.to_wscale = sc->sc_requested_r_scale;
1940 				to.to_flags |= TOF_SCALE;
1941 			}
1942 			if (sc->sc_flags & SCF_SACK)
1943 				to.to_flags |= TOF_SACKPERM;
1944 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1945 			if (sc->sc_flags & SCF_SIGNATURE)
1946 				to.to_flags |= TOF_SIGNATURE;
1947 #endif
1948 			if (sc->sc_tfo_cookie) {
1949 				to.to_flags |= TOF_FASTOPEN;
1950 				to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1951 				to.to_tfo_cookie = sc->sc_tfo_cookie;
1952 				/* don't send cookie again when retransmitting response */
1953 				sc->sc_tfo_cookie = NULL;
1954 			}
1955 		}
1956 		if (sc->sc_flags & SCF_TIMESTAMP) {
1957 			to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1958 			to.to_tsecr = sc->sc_tsreflect;
1959 			to.to_flags |= TOF_TS;
1960 		}
1961 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1962 
1963 		/* Adjust headers by option size. */
1964 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1965 		m->m_len += optlen;
1966 		m->m_pkthdr.len += optlen;
1967 #ifdef INET6
1968 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1969 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1970 		else
1971 #endif
1972 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1973 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1974 		if (sc->sc_flags & SCF_SIGNATURE) {
1975 			KASSERT(to.to_flags & TOF_SIGNATURE,
1976 			    ("tcp_addoptions() didn't set tcp_signature"));
1977 
1978 			/* NOTE: to.to_signature is inside of mbuf */
1979 			if (!TCPMD5_ENABLED() ||
1980 			    TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1981 				m_freem(m);
1982 				return (EACCES);
1983 			}
1984 		}
1985 #endif
1986 	} else
1987 		optlen = 0;
1988 
1989 	if (udp) {
1990 		ulen += optlen;
1991 		udp->uh_ulen = htons(ulen);
1992 	}
1993 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
1994 	/*
1995 	 * If we have peer's SYN and it has a flowid, then let's assign it to
1996 	 * our SYN|ACK.  ip6_output() and ip_output() will not assign flowid
1997 	 * to SYN|ACK due to lack of inp here.
1998 	 */
1999 	if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
2000 		m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
2001 		M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
2002 	}
2003 #ifdef INET6
2004 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
2005 		if (sc->sc_port) {
2006 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
2007 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2008 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen,
2009 			      IPPROTO_UDP, 0);
2010 			th->th_sum = htons(0);
2011 		} else {
2012 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
2013 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2014 			th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
2015 			    IPPROTO_TCP, 0);
2016 		}
2017 		ip6->ip6_hlim = sc->sc_ip_ttl;
2018 #ifdef TCP_OFFLOAD
2019 		if (ADDED_BY_TOE(sc)) {
2020 			struct toedev *tod = sc->sc_tod;
2021 
2022 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2023 
2024 			return (error);
2025 		}
2026 #endif
2027 		TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
2028 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
2029 	}
2030 #endif
2031 #if defined(INET6) && defined(INET)
2032 	else
2033 #endif
2034 #ifdef INET
2035 	{
2036 		if (sc->sc_port) {
2037 			m->m_pkthdr.csum_flags = CSUM_UDP;
2038 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2039 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
2040 			      ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
2041 			th->th_sum = htons(0);
2042 		} else {
2043 			m->m_pkthdr.csum_flags = CSUM_TCP;
2044 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2045 			th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
2046 			    htons(tlen + optlen - hlen + IPPROTO_TCP));
2047 		}
2048 #ifdef TCP_OFFLOAD
2049 		if (ADDED_BY_TOE(sc)) {
2050 			struct toedev *tod = sc->sc_tod;
2051 
2052 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2053 
2054 			return (error);
2055 		}
2056 #endif
2057 		TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
2058 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
2059 	}
2060 #endif
2061 	return (error);
2062 }
2063 
2064 /*
2065  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
2066  * that exceed the capacity of the syncache by avoiding the storage of any
2067  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
2068  * attacks where the attacker does not have access to our responses.
2069  *
2070  * Syncookies encode and include all necessary information about the
2071  * connection setup within the SYN|ACK that we send back.  That way we
2072  * can avoid keeping any local state until the ACK to our SYN|ACK returns
2073  * (if ever).  Normally the syncache and syncookies are running in parallel
2074  * with the latter taking over when the former is exhausted.  When matching
2075  * syncache entry is found the syncookie is ignored.
2076  *
2077  * The only reliable information persisting the 3WHS is our initial sequence
2078  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
2079  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
2080  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
2081  * returns and signifies a legitimate connection if it matches the ACK.
2082  *
2083  * The available space of 32 bits to store the hash and to encode the SYN
2084  * option information is very tight and we should have at least 24 bits for
2085  * the MAC to keep the number of guesses by blind spoofing reasonably high.
2086  *
2087  * SYN option information we have to encode to fully restore a connection:
2088  * MSS: is imporant to chose an optimal segment size to avoid IP level
2089  *   fragmentation along the path.  The common MSS values can be encoded
2090  *   in a 3-bit table.  Uncommon values are captured by the next lower value
2091  *   in the table leading to a slight increase in packetization overhead.
2092  * WSCALE: is necessary to allow large windows to be used for high delay-
2093  *   bandwidth product links.  Not scaling the window when it was initially
2094  *   negotiated is bad for performance as lack of scaling further decreases
2095  *   the apparent available send window.  We only need to encode the WSCALE
2096  *   we received from the remote end.  Our end can be recalculated at any
2097  *   time.  The common WSCALE values can be encoded in a 3-bit table.
2098  *   Uncommon values are captured by the next lower value in the table
2099  *   making us under-estimate the available window size halving our
2100  *   theoretically possible maximum throughput for that connection.
2101  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2102  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2103  *   that are included in all segments on a connection.  We enable them when
2104  *   the ACK has them.
2105  *
2106  * Security of syncookies and attack vectors:
2107  *
2108  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2109  * together with the gloabl secret to make it unique per connection attempt.
2110  * Thus any change of any of those parameters results in a different MAC output
2111  * in an unpredictable way unless a collision is encountered.  24 bits of the
2112  * MAC are embedded into the ISS.
2113  *
2114  * To prevent replay attacks two rotating global secrets are updated with a
2115  * new random value every 15 seconds.  The life-time of a syncookie is thus
2116  * 15-30 seconds.
2117  *
2118  * Vector 1: Attacking the secret.  This requires finding a weakness in the
2119  * MAC itself or the way it is used here.  The attacker can do a chosen plain
2120  * text attack by varying and testing the all parameters under his control.
2121  * The strength depends on the size and randomness of the secret, and the
2122  * cryptographic security of the MAC function.  Due to the constant updating
2123  * of the secret the attacker has at most 29.999 seconds to find the secret
2124  * and launch spoofed connections.  After that he has to start all over again.
2125  *
2126  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
2127  * size an average of 4,823 attempts are required for a 50% chance of success
2128  * to spoof a single syncookie (birthday collision paradox).  However the
2129  * attacker is blind and doesn't know if one of his attempts succeeded unless
2130  * he has a side channel to interfere success from.  A single connection setup
2131  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2132  * This many attempts are required for each one blind spoofed connection.  For
2133  * every additional spoofed connection he has to launch another N attempts.
2134  * Thus for a sustained rate 100 spoofed connections per second approximately
2135  * 1,800,000 packets per second would have to be sent.
2136  *
2137  * NB: The MAC function should be fast so that it doesn't become a CPU
2138  * exhaustion attack vector itself.
2139  *
2140  * References:
2141  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2142  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2143  *   http://cr.yp.to/syncookies.html    (overview)
2144  *   http://cr.yp.to/syncookies/archive (details)
2145  *
2146  *
2147  * Schematic construction of a syncookie enabled Initial Sequence Number:
2148  *  0        1         2         3
2149  *  12345678901234567890123456789012
2150  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2151  *
2152  *  x 24 MAC (truncated)
2153  *  W  3 Send Window Scale index
2154  *  M  3 MSS index
2155  *  S  1 SACK permitted
2156  *  P  1 Odd/even secret
2157  */
2158 
2159 /*
2160  * Distribution and probability of certain MSS values.  Those in between are
2161  * rounded down to the next lower one.
2162  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2163  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
2164  */
2165 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2166 
2167 /*
2168  * Distribution and probability of certain WSCALE values.  We have to map the
2169  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2170  * bits based on prevalence of certain values.  Where we don't have an exact
2171  * match for are rounded down to the next lower one letting us under-estimate
2172  * the true available window.  At the moment this would happen only for the
2173  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2174  * and window size).  The absence of the WSCALE option (no scaling in either
2175  * direction) is encoded with index zero.
2176  * [WSCALE values histograms, Allman, 2012]
2177  *                            X 10 10 35  5  6 14 10%   by host
2178  *                            X 11  4  5  5 18 49  3%   by connections
2179  */
2180 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2181 
2182 /*
2183  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
2184  * and good cryptographic properties.
2185  */
2186 static uint32_t
2187 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2188     uint8_t *secbits, uintptr_t secmod)
2189 {
2190 	SIPHASH_CTX ctx;
2191 	uint32_t siphash[2];
2192 
2193 	SipHash24_Init(&ctx);
2194 	SipHash_SetKey(&ctx, secbits);
2195 	switch (inc->inc_flags & INC_ISIPV6) {
2196 #ifdef INET
2197 	case 0:
2198 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2199 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2200 		break;
2201 #endif
2202 #ifdef INET6
2203 	case INC_ISIPV6:
2204 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2205 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2206 		break;
2207 #endif
2208 	}
2209 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2210 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2211 	SipHash_Update(&ctx, &irs, sizeof(irs));
2212 	SipHash_Update(&ctx, &flags, sizeof(flags));
2213 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
2214 	SipHash_Final((u_int8_t *)&siphash, &ctx);
2215 
2216 	return (siphash[0] ^ siphash[1]);
2217 }
2218 
2219 static tcp_seq
2220 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2221 {
2222 	u_int i, secbit, wscale;
2223 	uint32_t iss, hash;
2224 	uint8_t *secbits;
2225 	union syncookie cookie;
2226 
2227 	cookie.cookie = 0;
2228 
2229 	/* Map our computed MSS into the 3-bit index. */
2230 	for (i = nitems(tcp_sc_msstab) - 1;
2231 	     tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2232 	     i--)
2233 		;
2234 	cookie.flags.mss_idx = i;
2235 
2236 	/*
2237 	 * Map the send window scale into the 3-bit index but only if
2238 	 * the wscale option was received.
2239 	 */
2240 	if (sc->sc_flags & SCF_WINSCALE) {
2241 		wscale = sc->sc_requested_s_scale;
2242 		for (i = nitems(tcp_sc_wstab) - 1;
2243 		    tcp_sc_wstab[i] > wscale && i > 0;
2244 		     i--)
2245 			;
2246 		cookie.flags.wscale_idx = i;
2247 	}
2248 
2249 	/* Can we do SACK? */
2250 	if (sc->sc_flags & SCF_SACK)
2251 		cookie.flags.sack_ok = 1;
2252 
2253 	/* Which of the two secrets to use. */
2254 	secbit = V_tcp_syncache.secret.oddeven & 0x1;
2255 	cookie.flags.odd_even = secbit;
2256 
2257 	secbits = V_tcp_syncache.secret.key[secbit];
2258 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2259 	    (uintptr_t)sch);
2260 
2261 	/*
2262 	 * Put the flags into the hash and XOR them to get better ISS number
2263 	 * variance.  This doesn't enhance the cryptographic strength and is
2264 	 * done to prevent the 8 cookie bits from showing up directly on the
2265 	 * wire.
2266 	 */
2267 	iss = hash & ~0xff;
2268 	iss |= cookie.cookie ^ (hash >> 24);
2269 
2270 	TCPSTAT_INC(tcps_sc_sendcookie);
2271 	return (iss);
2272 }
2273 
2274 static struct syncache *
2275 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2276     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2277     struct socket *lso, uint16_t port)
2278 {
2279 	uint32_t hash;
2280 	uint8_t *secbits;
2281 	tcp_seq ack, seq;
2282 	int wnd, wscale = 0;
2283 	union syncookie cookie;
2284 
2285 	/*
2286 	 * Pull information out of SYN-ACK/ACK and revert sequence number
2287 	 * advances.
2288 	 */
2289 	ack = th->th_ack - 1;
2290 	seq = th->th_seq - 1;
2291 
2292 	/*
2293 	 * Unpack the flags containing enough information to restore the
2294 	 * connection.
2295 	 */
2296 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2297 
2298 	/* Which of the two secrets to use. */
2299 	secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2300 
2301 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2302 
2303 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
2304 	if ((ack & ~0xff) != (hash & ~0xff))
2305 		return (NULL);
2306 
2307 	/* Fill in the syncache values. */
2308 	sc->sc_flags = 0;
2309 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2310 	sc->sc_ipopts = NULL;
2311 
2312 	sc->sc_irs = seq;
2313 	sc->sc_iss = ack;
2314 
2315 	switch (inc->inc_flags & INC_ISIPV6) {
2316 #ifdef INET
2317 	case 0:
2318 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2319 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2320 		break;
2321 #endif
2322 #ifdef INET6
2323 	case INC_ISIPV6:
2324 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2325 			sc->sc_flowlabel =
2326 			    htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2327 		break;
2328 #endif
2329 	}
2330 
2331 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2332 
2333 	/* We can simply recompute receive window scale we sent earlier. */
2334 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2335 		wscale++;
2336 
2337 	/* Only use wscale if it was enabled in the orignal SYN. */
2338 	if (cookie.flags.wscale_idx > 0) {
2339 		sc->sc_requested_r_scale = wscale;
2340 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2341 		sc->sc_flags |= SCF_WINSCALE;
2342 	}
2343 
2344 	wnd = lso->sol_sbrcv_hiwat;
2345 	wnd = imax(wnd, 0);
2346 	wnd = imin(wnd, TCP_MAXWIN);
2347 	sc->sc_wnd = wnd;
2348 
2349 	if (cookie.flags.sack_ok)
2350 		sc->sc_flags |= SCF_SACK;
2351 
2352 	if (to->to_flags & TOF_TS) {
2353 		sc->sc_flags |= SCF_TIMESTAMP;
2354 		sc->sc_tsreflect = to->to_tsval;
2355 		sc->sc_tsoff = tcp_new_ts_offset(inc);
2356 	}
2357 
2358 	if (to->to_flags & TOF_SIGNATURE)
2359 		sc->sc_flags |= SCF_SIGNATURE;
2360 
2361 	sc->sc_rxmits = 0;
2362 
2363 	sc->sc_port = port;
2364 
2365 	TCPSTAT_INC(tcps_sc_recvcookie);
2366 	return (sc);
2367 }
2368 
2369 #ifdef INVARIANTS
2370 static int
2371 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2372     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2373     struct socket *lso, uint16_t port)
2374 {
2375 	struct syncache scs, *scx;
2376 	char *s;
2377 
2378 	bzero(&scs, sizeof(scs));
2379 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port);
2380 
2381 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2382 		return (0);
2383 
2384 	if (scx != NULL) {
2385 		if (sc->sc_peer_mss != scx->sc_peer_mss)
2386 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2387 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2388 
2389 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2390 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2391 			    s, __func__, sc->sc_requested_r_scale,
2392 			    scx->sc_requested_r_scale);
2393 
2394 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2395 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2396 			    s, __func__, sc->sc_requested_s_scale,
2397 			    scx->sc_requested_s_scale);
2398 
2399 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2400 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2401 	}
2402 
2403 	if (s != NULL)
2404 		free(s, M_TCPLOG);
2405 	return (0);
2406 }
2407 #endif /* INVARIANTS */
2408 
2409 static void
2410 syncookie_reseed(void *arg)
2411 {
2412 	struct tcp_syncache *sc = arg;
2413 	uint8_t *secbits;
2414 	int secbit;
2415 
2416 	/*
2417 	 * Reseeding the secret doesn't have to be protected by a lock.
2418 	 * It only must be ensured that the new random values are visible
2419 	 * to all CPUs in a SMP environment.  The atomic with release
2420 	 * semantics ensures that.
2421 	 */
2422 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2423 	secbits = sc->secret.key[secbit];
2424 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2425 	atomic_add_rel_int(&sc->secret.oddeven, 1);
2426 
2427 	/* Reschedule ourself. */
2428 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2429 }
2430 
2431 /*
2432  * We have overflowed a bucket. Let's pause dealing with the syncache.
2433  * This function will increment the bucketoverflow statistics appropriately
2434  * (once per pause when pausing is enabled; otherwise, once per overflow).
2435  */
2436 static void
2437 syncache_pause(struct in_conninfo *inc)
2438 {
2439 	time_t delta;
2440 	const char *s;
2441 
2442 	/* XXX:
2443 	 * 2. Add sysctl read here so we don't get the benefit of this
2444 	 * change without the new sysctl.
2445 	 */
2446 
2447 	/*
2448 	 * Try an unlocked read. If we already know that another thread
2449 	 * has activated the feature, there is no need to proceed.
2450 	 */
2451 	if (V_tcp_syncache.paused)
2452 		return;
2453 
2454 	/* Are cookied enabled? If not, we can't pause. */
2455 	if (!V_tcp_syncookies) {
2456 		TCPSTAT_INC(tcps_sc_bucketoverflow);
2457 		return;
2458 	}
2459 
2460 	/*
2461 	 * We may be the first thread to find an overflow. Get the lock
2462 	 * and evaluate if we need to take action.
2463 	 */
2464 	mtx_lock(&V_tcp_syncache.pause_mtx);
2465 	if (V_tcp_syncache.paused) {
2466 		mtx_unlock(&V_tcp_syncache.pause_mtx);
2467 		return;
2468 	}
2469 
2470 	/* Activate protection. */
2471 	V_tcp_syncache.paused = true;
2472 	TCPSTAT_INC(tcps_sc_bucketoverflow);
2473 
2474 	/*
2475 	 * Determine the last backoff time. If we are seeing a re-newed
2476 	 * attack within that same time after last reactivating the syncache,
2477 	 * consider it an extension of the same attack.
2478 	 */
2479 	delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2480 	if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2481 		if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2482 			delta <<= 1;
2483 			V_tcp_syncache.pause_backoff++;
2484 		}
2485 	} else {
2486 		delta = TCP_SYNCACHE_PAUSE_TIME;
2487 		V_tcp_syncache.pause_backoff = 0;
2488 	}
2489 
2490 	/* Log a warning, including IP addresses, if able. */
2491 	if (inc != NULL)
2492 		s = tcp_log_addrs(inc, NULL, NULL, NULL);
2493 	else
2494 		s = (const char *)NULL;
2495 	log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2496 	    "the next %lld seconds%s%s%s\n", (long long)delta,
2497 	    (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2498 	    (s != NULL) ? ")" : "");
2499 	free(__DECONST(void *, s), M_TCPLOG);
2500 
2501 	/* Use the calculated delta to set a new pause time. */
2502 	V_tcp_syncache.pause_until = time_uptime + delta;
2503 	callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2504 	    &V_tcp_syncache);
2505 	mtx_unlock(&V_tcp_syncache.pause_mtx);
2506 }
2507 
2508 /* Evaluate whether we need to unpause. */
2509 static void
2510 syncache_unpause(void *arg)
2511 {
2512 	struct tcp_syncache *sc;
2513 	time_t delta;
2514 
2515 	sc = arg;
2516 	mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2517 	callout_deactivate(&sc->pause_co);
2518 
2519 	/*
2520 	 * Check to make sure we are not running early. If the pause
2521 	 * time has expired, then deactivate the protection.
2522 	 */
2523 	if ((delta = sc->pause_until - time_uptime) > 0)
2524 		callout_schedule(&sc->pause_co, delta * hz);
2525 	else
2526 		sc->paused = false;
2527 }
2528 
2529 /*
2530  * Exports the syncache entries to userland so that netstat can display
2531  * them alongside the other sockets.  This function is intended to be
2532  * called only from tcp_pcblist.
2533  *
2534  * Due to concurrency on an active system, the number of pcbs exported
2535  * may have no relation to max_pcbs.  max_pcbs merely indicates the
2536  * amount of space the caller allocated for this function to use.
2537  */
2538 int
2539 syncache_pcblist(struct sysctl_req *req)
2540 {
2541 	struct xtcpcb xt;
2542 	struct syncache *sc;
2543 	struct syncache_head *sch;
2544 	int error, i;
2545 
2546 	bzero(&xt, sizeof(xt));
2547 	xt.xt_len = sizeof(xt);
2548 	xt.t_state = TCPS_SYN_RECEIVED;
2549 	xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2550 	xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2551 	xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2552 	xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2553 
2554 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2555 		sch = &V_tcp_syncache.hashbase[i];
2556 		SCH_LOCK(sch);
2557 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2558 			if (sc->sc_cred != NULL &&
2559 			    cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2560 				continue;
2561 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2562 				xt.xt_inp.inp_vflag = INP_IPV6;
2563 			else
2564 				xt.xt_inp.inp_vflag = INP_IPV4;
2565 			xt.xt_encaps_port = sc->sc_port;
2566 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2567 			    sizeof (struct in_conninfo));
2568 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2569 			if (error) {
2570 				SCH_UNLOCK(sch);
2571 				return (0);
2572 			}
2573 		}
2574 		SCH_UNLOCK(sch);
2575 	}
2576 
2577 	return (0);
2578 }
2579