xref: /freebsd/sys/netinet/tcp_syncache.c (revision 5b5b7e2ca2fa9a2418dd51749f4ef6f881ae7179)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2001 McAfee, Inc.
5  * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG
6  * All rights reserved.
7  *
8  * This software was developed for the FreeBSD Project by Jonathan Lemon
9  * and McAfee Research, the Security Research Division of McAfee, Inc. under
10  * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
11  * DARPA CHATS research program. [2001 McAfee, Inc.]
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 #include "opt_inet.h"
39 #include "opt_inet6.h"
40 #include "opt_ipsec.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/hash.h>
45 #include <sys/refcount.h>
46 #include <sys/kernel.h>
47 #include <sys/sysctl.h>
48 #include <sys/limits.h>
49 #include <sys/lock.h>
50 #include <sys/mutex.h>
51 #include <sys/malloc.h>
52 #include <sys/mbuf.h>
53 #include <sys/proc.h>		/* for proc0 declaration */
54 #include <sys/random.h>
55 #include <sys/socket.h>
56 #include <sys/socketvar.h>
57 #include <sys/syslog.h>
58 #include <sys/ucred.h>
59 
60 #include <sys/md5.h>
61 #include <crypto/siphash/siphash.h>
62 
63 #include <vm/uma.h>
64 
65 #include <net/if.h>
66 #include <net/if_var.h>
67 #include <net/route.h>
68 #include <net/vnet.h>
69 
70 #include <netinet/in.h>
71 #include <netinet/in_kdtrace.h>
72 #include <netinet/in_systm.h>
73 #include <netinet/ip.h>
74 #include <netinet/in_var.h>
75 #include <netinet/in_pcb.h>
76 #include <netinet/ip_var.h>
77 #include <netinet/ip_options.h>
78 #ifdef INET6
79 #include <netinet/ip6.h>
80 #include <netinet/icmp6.h>
81 #include <netinet6/nd6.h>
82 #include <netinet6/ip6_var.h>
83 #include <netinet6/in6_pcb.h>
84 #endif
85 #include <netinet/tcp.h>
86 #include <netinet/tcp_fastopen.h>
87 #include <netinet/tcp_fsm.h>
88 #include <netinet/tcp_seq.h>
89 #include <netinet/tcp_timer.h>
90 #include <netinet/tcp_var.h>
91 #include <netinet/tcp_syncache.h>
92 #include <netinet/tcp_ecn.h>
93 #ifdef INET6
94 #include <netinet6/tcp6_var.h>
95 #endif
96 #ifdef TCP_OFFLOAD
97 #include <netinet/toecore.h>
98 #endif
99 #include <netinet/udp.h>
100 #include <netinet/udp_var.h>
101 
102 #include <netipsec/ipsec_support.h>
103 
104 #include <machine/in_cksum.h>
105 
106 #include <security/mac/mac_framework.h>
107 
108 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1;
109 #define	V_tcp_syncookies		VNET(tcp_syncookies)
110 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
111     &VNET_NAME(tcp_syncookies), 0,
112     "Use TCP SYN cookies if the syncache overflows");
113 
114 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0;
115 #define	V_tcp_syncookiesonly		VNET(tcp_syncookiesonly)
116 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
117     &VNET_NAME(tcp_syncookiesonly), 0,
118     "Use only TCP SYN cookies");
119 
120 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1;
121 #define V_functions_inherit_listen_socket_stack \
122     VNET(functions_inherit_listen_socket_stack)
123 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack,
124     CTLFLAG_VNET | CTLFLAG_RW,
125     &VNET_NAME(functions_inherit_listen_socket_stack), 0,
126     "Inherit listen socket's stack");
127 
128 #ifdef TCP_OFFLOAD
129 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
130 #endif
131 
132 static void	 syncache_drop(struct syncache *, struct syncache_head *);
133 static void	 syncache_free(struct syncache *);
134 static void	 syncache_insert(struct syncache *, struct syncache_head *);
135 static int	 syncache_respond(struct syncache *, const struct mbuf *, int);
136 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
137 		    struct mbuf *m);
138 static void	 syncache_timeout(struct syncache *sc, struct syncache_head *sch,
139 		    int docallout);
140 static void	 syncache_timer(void *);
141 
142 static uint32_t	 syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
143 		    uint8_t *, uintptr_t);
144 static tcp_seq	 syncookie_generate(struct syncache_head *, struct syncache *);
145 static struct syncache
146 		*syncookie_lookup(struct in_conninfo *, struct syncache_head *,
147 		    struct syncache *, struct tcphdr *, struct tcpopt *,
148 		    struct socket *, uint16_t);
149 static void	syncache_pause(struct in_conninfo *);
150 static void	syncache_unpause(void *);
151 static void	 syncookie_reseed(void *);
152 #ifdef INVARIANTS
153 static int	 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
154 		    struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
155 		    struct socket *lso, uint16_t port);
156 #endif
157 
158 /*
159  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
160  * 3 retransmits corresponds to a timeout with default values of
161  * tcp_rexmit_initial * (             1 +
162  *                       tcp_backoff[1] +
163  *                       tcp_backoff[2] +
164  *                       tcp_backoff[3]) + 3 * tcp_rexmit_slop,
165  * 1000 ms * (1 + 2 + 4 + 8) +  3 * 200 ms = 15600 ms,
166  * the odds are that the user has given up attempting to connect by then.
167  */
168 #define SYNCACHE_MAXREXMTS		3
169 
170 /* Arbitrary values */
171 #define TCP_SYNCACHE_HASHSIZE		512
172 #define TCP_SYNCACHE_BUCKETLIMIT	30
173 
174 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
175 #define	V_tcp_syncache			VNET(tcp_syncache)
176 
177 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache,
178     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
179     "TCP SYN cache");
180 
181 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
182     &VNET_NAME(tcp_syncache.bucket_limit), 0,
183     "Per-bucket hash limit for syncache");
184 
185 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
186     &VNET_NAME(tcp_syncache.cache_limit), 0,
187     "Overall entry limit for syncache");
188 
189 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
190     &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
191 
192 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
193     &VNET_NAME(tcp_syncache.hashsize), 0,
194     "Size of TCP syncache hashtable");
195 
196 SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET |
197     CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0,
198     "All syncache(4) entries are visible, ignoring UID/GID, jail(2) "
199     "and mac(4) checks");
200 
201 static int
202 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
203 {
204 	int error;
205 	u_int new;
206 
207 	new = V_tcp_syncache.rexmt_limit;
208 	error = sysctl_handle_int(oidp, &new, 0, req);
209 	if ((error == 0) && (req->newptr != NULL)) {
210 		if (new > TCP_MAXRXTSHIFT)
211 			error = EINVAL;
212 		else
213 			V_tcp_syncache.rexmt_limit = new;
214 	}
215 	return (error);
216 }
217 
218 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
219     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
220     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
221     sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
222     "Limit on SYN/ACK retransmissions");
223 
224 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
225 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
226     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
227     "Send reset on socket allocation failure");
228 
229 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
230 
231 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
232 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
233 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
234 
235 /*
236  * Requires the syncache entry to be already removed from the bucket list.
237  */
238 static void
239 syncache_free(struct syncache *sc)
240 {
241 
242 	if (sc->sc_ipopts)
243 		(void) m_free(sc->sc_ipopts);
244 	if (sc->sc_cred)
245 		crfree(sc->sc_cred);
246 #ifdef MAC
247 	mac_syncache_destroy(&sc->sc_label);
248 #endif
249 
250 	uma_zfree(V_tcp_syncache.zone, sc);
251 }
252 
253 void
254 syncache_init(void)
255 {
256 	int i;
257 
258 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
259 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
260 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
261 	V_tcp_syncache.hash_secret = arc4random();
262 
263 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
264 	    &V_tcp_syncache.hashsize);
265 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
266 	    &V_tcp_syncache.bucket_limit);
267 	if (!powerof2(V_tcp_syncache.hashsize) ||
268 	    V_tcp_syncache.hashsize == 0) {
269 		printf("WARNING: syncache hash size is not a power of 2.\n");
270 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
271 	}
272 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
273 
274 	/* Set limits. */
275 	V_tcp_syncache.cache_limit =
276 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
277 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
278 	    &V_tcp_syncache.cache_limit);
279 
280 	/* Allocate the hash table. */
281 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
282 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
283 
284 #ifdef VIMAGE
285 	V_tcp_syncache.vnet = curvnet;
286 #endif
287 
288 	/* Initialize the hash buckets. */
289 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
290 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
291 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
292 			 NULL, MTX_DEF);
293 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
294 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
295 		V_tcp_syncache.hashbase[i].sch_length = 0;
296 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
297 		V_tcp_syncache.hashbase[i].sch_last_overflow =
298 		    -(SYNCOOKIE_LIFETIME + 1);
299 	}
300 
301 	/* Create the syncache entry zone. */
302 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
303 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
304 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
305 	    V_tcp_syncache.cache_limit);
306 
307 	/* Start the SYN cookie reseeder callout. */
308 	callout_init(&V_tcp_syncache.secret.reseed, 1);
309 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
310 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
311 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
312 	    syncookie_reseed, &V_tcp_syncache);
313 
314 	/* Initialize the pause machinery. */
315 	mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF);
316 	callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx,
317 	    0);
318 	V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME;
319 	V_tcp_syncache.pause_backoff = 0;
320 	V_tcp_syncache.paused = false;
321 }
322 
323 #ifdef VIMAGE
324 void
325 syncache_destroy(void)
326 {
327 	struct syncache_head *sch;
328 	struct syncache *sc, *nsc;
329 	int i;
330 
331 	/*
332 	 * Stop the re-seed timer before freeing resources.  No need to
333 	 * possibly schedule it another time.
334 	 */
335 	callout_drain(&V_tcp_syncache.secret.reseed);
336 
337 	/* Stop the SYN cache pause callout. */
338 	mtx_lock(&V_tcp_syncache.pause_mtx);
339 	if (callout_stop(&V_tcp_syncache.pause_co) == 0) {
340 		mtx_unlock(&V_tcp_syncache.pause_mtx);
341 		callout_drain(&V_tcp_syncache.pause_co);
342 	} else
343 		mtx_unlock(&V_tcp_syncache.pause_mtx);
344 
345 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
346 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
347 		sch = &V_tcp_syncache.hashbase[i];
348 		callout_drain(&sch->sch_timer);
349 
350 		SCH_LOCK(sch);
351 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
352 			syncache_drop(sc, sch);
353 		SCH_UNLOCK(sch);
354 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
355 		    ("%s: sch->sch_bucket not empty", __func__));
356 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
357 		    __func__, sch->sch_length));
358 		mtx_destroy(&sch->sch_mtx);
359 	}
360 
361 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
362 	    ("%s: cache_count not 0", __func__));
363 
364 	/* Free the allocated global resources. */
365 	uma_zdestroy(V_tcp_syncache.zone);
366 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
367 	mtx_destroy(&V_tcp_syncache.pause_mtx);
368 }
369 #endif
370 
371 /*
372  * Inserts a syncache entry into the specified bucket row.
373  * Locks and unlocks the syncache_head autonomously.
374  */
375 static void
376 syncache_insert(struct syncache *sc, struct syncache_head *sch)
377 {
378 	struct syncache *sc2;
379 
380 	SCH_LOCK(sch);
381 
382 	/*
383 	 * Make sure that we don't overflow the per-bucket limit.
384 	 * If the bucket is full, toss the oldest element.
385 	 */
386 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
387 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
388 			("sch->sch_length incorrect"));
389 		syncache_pause(&sc->sc_inc);
390 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
391 		sch->sch_last_overflow = time_uptime;
392 		syncache_drop(sc2, sch);
393 	}
394 
395 	/* Put it into the bucket. */
396 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
397 	sch->sch_length++;
398 
399 #ifdef TCP_OFFLOAD
400 	if (ADDED_BY_TOE(sc)) {
401 		struct toedev *tod = sc->sc_tod;
402 
403 		tod->tod_syncache_added(tod, sc->sc_todctx);
404 	}
405 #endif
406 
407 	/* Reinitialize the bucket row's timer. */
408 	if (sch->sch_length == 1)
409 		sch->sch_nextc = ticks + INT_MAX;
410 	syncache_timeout(sc, sch, 1);
411 
412 	SCH_UNLOCK(sch);
413 
414 	TCPSTATES_INC(TCPS_SYN_RECEIVED);
415 	TCPSTAT_INC(tcps_sc_added);
416 }
417 
418 /*
419  * Remove and free entry from syncache bucket row.
420  * Expects locked syncache head.
421  */
422 static void
423 syncache_drop(struct syncache *sc, struct syncache_head *sch)
424 {
425 
426 	SCH_LOCK_ASSERT(sch);
427 
428 	TCPSTATES_DEC(TCPS_SYN_RECEIVED);
429 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
430 	sch->sch_length--;
431 
432 #ifdef TCP_OFFLOAD
433 	if (ADDED_BY_TOE(sc)) {
434 		struct toedev *tod = sc->sc_tod;
435 
436 		tod->tod_syncache_removed(tod, sc->sc_todctx);
437 	}
438 #endif
439 
440 	syncache_free(sc);
441 }
442 
443 /*
444  * Engage/reengage time on bucket row.
445  */
446 static void
447 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
448 {
449 	int rexmt;
450 
451 	if (sc->sc_rxmits == 0)
452 		rexmt = tcp_rexmit_initial;
453 	else
454 		TCPT_RANGESET(rexmt,
455 		    tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
456 		    tcp_rexmit_min, TCPTV_REXMTMAX);
457 	sc->sc_rxttime = ticks + rexmt;
458 	sc->sc_rxmits++;
459 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
460 		sch->sch_nextc = sc->sc_rxttime;
461 		if (docallout)
462 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
463 			    syncache_timer, (void *)sch);
464 	}
465 }
466 
467 /*
468  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
469  * If we have retransmitted an entry the maximum number of times, expire it.
470  * One separate timer for each bucket row.
471  */
472 static void
473 syncache_timer(void *xsch)
474 {
475 	struct syncache_head *sch = (struct syncache_head *)xsch;
476 	struct syncache *sc, *nsc;
477 	struct epoch_tracker et;
478 	int tick = ticks;
479 	char *s;
480 	bool paused;
481 
482 	CURVNET_SET(sch->sch_sc->vnet);
483 
484 	/* NB: syncache_head has already been locked by the callout. */
485 	SCH_LOCK_ASSERT(sch);
486 
487 	/*
488 	 * In the following cycle we may remove some entries and/or
489 	 * advance some timeouts, so re-initialize the bucket timer.
490 	 */
491 	sch->sch_nextc = tick + INT_MAX;
492 
493 	/*
494 	 * If we have paused processing, unconditionally remove
495 	 * all syncache entries.
496 	 */
497 	mtx_lock(&V_tcp_syncache.pause_mtx);
498 	paused = V_tcp_syncache.paused;
499 	mtx_unlock(&V_tcp_syncache.pause_mtx);
500 
501 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
502 		if (paused) {
503 			syncache_drop(sc, sch);
504 			continue;
505 		}
506 		/*
507 		 * We do not check if the listen socket still exists
508 		 * and accept the case where the listen socket may be
509 		 * gone by the time we resend the SYN/ACK.  We do
510 		 * not expect this to happens often. If it does,
511 		 * then the RST will be sent by the time the remote
512 		 * host does the SYN/ACK->ACK.
513 		 */
514 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
515 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
516 				sch->sch_nextc = sc->sc_rxttime;
517 			continue;
518 		}
519 		if (sc->sc_rxmits > V_tcp_ecn_maxretries) {
520 			sc->sc_flags &= ~SCF_ECN;
521 		}
522 		if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
523 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
524 				log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
525 				    "giving up and removing syncache entry\n",
526 				    s, __func__);
527 				free(s, M_TCPLOG);
528 			}
529 			syncache_drop(sc, sch);
530 			TCPSTAT_INC(tcps_sc_stale);
531 			continue;
532 		}
533 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
534 			log(LOG_DEBUG, "%s; %s: Response timeout, "
535 			    "retransmitting (%u) SYN|ACK\n",
536 			    s, __func__, sc->sc_rxmits);
537 			free(s, M_TCPLOG);
538 		}
539 
540 		NET_EPOCH_ENTER(et);
541 		syncache_respond(sc, NULL, TH_SYN|TH_ACK);
542 		NET_EPOCH_EXIT(et);
543 		TCPSTAT_INC(tcps_sc_retransmitted);
544 		syncache_timeout(sc, sch, 0);
545 	}
546 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
547 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
548 			syncache_timer, (void *)(sch));
549 	CURVNET_RESTORE();
550 }
551 
552 /*
553  * Returns true if the system is only using cookies at the moment.
554  * This could be due to a sysadmin decision to only use cookies, or it
555  * could be due to the system detecting an attack.
556  */
557 static inline bool
558 syncache_cookiesonly(void)
559 {
560 
561 	return (V_tcp_syncookies && (V_tcp_syncache.paused ||
562 	    V_tcp_syncookiesonly));
563 }
564 
565 /*
566  * Find the hash bucket for the given connection.
567  */
568 static struct syncache_head *
569 syncache_hashbucket(struct in_conninfo *inc)
570 {
571 	uint32_t hash;
572 
573 	/*
574 	 * The hash is built on foreign port + local port + foreign address.
575 	 * We rely on the fact that struct in_conninfo starts with 16 bits
576 	 * of foreign port, then 16 bits of local port then followed by 128
577 	 * bits of foreign address.  In case of IPv4 address, the first 3
578 	 * 32-bit words of the address always are zeroes.
579 	 */
580 	hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
581 	    V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
582 
583 	return (&V_tcp_syncache.hashbase[hash]);
584 }
585 
586 /*
587  * Find an entry in the syncache.
588  * Returns always with locked syncache_head plus a matching entry or NULL.
589  */
590 static struct syncache *
591 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
592 {
593 	struct syncache *sc;
594 	struct syncache_head *sch;
595 
596 	*schp = sch = syncache_hashbucket(inc);
597 	SCH_LOCK(sch);
598 
599 	/* Circle through bucket row to find matching entry. */
600 	TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
601 		if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
602 		    sizeof(struct in_endpoints)) == 0)
603 			break;
604 
605 	return (sc);	/* Always returns with locked sch. */
606 }
607 
608 /*
609  * This function is called when we get a RST for a
610  * non-existent connection, so that we can see if the
611  * connection is in the syn cache.  If it is, zap it.
612  * If required send a challenge ACK.
613  */
614 void
615 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m,
616     uint16_t port)
617 {
618 	struct syncache *sc;
619 	struct syncache_head *sch;
620 	char *s = NULL;
621 
622 	if (syncache_cookiesonly())
623 		return;
624 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
625 	SCH_LOCK_ASSERT(sch);
626 
627 	/*
628 	 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
629 	 * See RFC 793 page 65, section SEGMENT ARRIVES.
630 	 */
631 	if (tcp_get_flags(th) & (TH_ACK|TH_SYN|TH_FIN)) {
632 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
633 			log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
634 			    "FIN flag set, segment ignored\n", s, __func__);
635 		TCPSTAT_INC(tcps_badrst);
636 		goto done;
637 	}
638 
639 	/*
640 	 * No corresponding connection was found in syncache.
641 	 * If syncookies are enabled and possibly exclusively
642 	 * used, or we are under memory pressure, a valid RST
643 	 * may not find a syncache entry.  In that case we're
644 	 * done and no SYN|ACK retransmissions will happen.
645 	 * Otherwise the RST was misdirected or spoofed.
646 	 */
647 	if (sc == NULL) {
648 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
649 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
650 			    "syncache entry (possibly syncookie only), "
651 			    "segment ignored\n", s, __func__);
652 		TCPSTAT_INC(tcps_badrst);
653 		goto done;
654 	}
655 
656 	/* The remote UDP encaps port does not match. */
657 	if (sc->sc_port != port) {
658 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
659 			log(LOG_DEBUG, "%s; %s: Spurious RST with matching "
660 			    "syncache entry but non-matching UDP encaps port, "
661 			    "segment ignored\n", s, __func__);
662 		TCPSTAT_INC(tcps_badrst);
663 		goto done;
664 	}
665 
666 	/*
667 	 * If the RST bit is set, check the sequence number to see
668 	 * if this is a valid reset segment.
669 	 *
670 	 * RFC 793 page 37:
671 	 *   In all states except SYN-SENT, all reset (RST) segments
672 	 *   are validated by checking their SEQ-fields.  A reset is
673 	 *   valid if its sequence number is in the window.
674 	 *
675 	 * RFC 793 page 69:
676 	 *   There are four cases for the acceptability test for an incoming
677 	 *   segment:
678 	 *
679 	 * Segment Receive  Test
680 	 * Length  Window
681 	 * ------- -------  -------------------------------------------
682 	 *    0       0     SEG.SEQ = RCV.NXT
683 	 *    0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
684 	 *   >0       0     not acceptable
685 	 *   >0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
686 	 *               or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
687 	 *
688 	 * Note that when receiving a SYN segment in the LISTEN state,
689 	 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
690 	 * described in RFC 793, page 66.
691 	 */
692 	if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
693 	    SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
694 	    (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
695 		if (V_tcp_insecure_rst ||
696 		    th->th_seq == sc->sc_irs + 1) {
697 			syncache_drop(sc, sch);
698 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
699 				log(LOG_DEBUG,
700 				    "%s; %s: Our SYN|ACK was rejected, "
701 				    "connection attempt aborted by remote "
702 				    "endpoint\n",
703 				    s, __func__);
704 			TCPSTAT_INC(tcps_sc_reset);
705 		} else {
706 			TCPSTAT_INC(tcps_badrst);
707 			/* Send challenge ACK. */
708 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
709 				log(LOG_DEBUG, "%s; %s: RST with invalid "
710 				    " SEQ %u != NXT %u (+WND %u), "
711 				    "sending challenge ACK\n",
712 				    s, __func__,
713 				    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
714 			syncache_respond(sc, m, TH_ACK);
715 		}
716 	} else {
717 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
718 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
719 			    "NXT %u (+WND %u), segment ignored\n",
720 			    s, __func__,
721 			    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
722 		TCPSTAT_INC(tcps_badrst);
723 	}
724 
725 done:
726 	if (s != NULL)
727 		free(s, M_TCPLOG);
728 	SCH_UNLOCK(sch);
729 }
730 
731 void
732 syncache_badack(struct in_conninfo *inc, uint16_t port)
733 {
734 	struct syncache *sc;
735 	struct syncache_head *sch;
736 
737 	if (syncache_cookiesonly())
738 		return;
739 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
740 	SCH_LOCK_ASSERT(sch);
741 	if ((sc != NULL) && (sc->sc_port == port)) {
742 		syncache_drop(sc, sch);
743 		TCPSTAT_INC(tcps_sc_badack);
744 	}
745 	SCH_UNLOCK(sch);
746 }
747 
748 void
749 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port)
750 {
751 	struct syncache *sc;
752 	struct syncache_head *sch;
753 
754 	if (syncache_cookiesonly())
755 		return;
756 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
757 	SCH_LOCK_ASSERT(sch);
758 	if (sc == NULL)
759 		goto done;
760 
761 	/* If the port != sc_port, then it's a bogus ICMP msg */
762 	if (port != sc->sc_port)
763 		goto done;
764 
765 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
766 	if (ntohl(th_seq) != sc->sc_iss)
767 		goto done;
768 
769 	/*
770 	 * If we've rertransmitted 3 times and this is our second error,
771 	 * we remove the entry.  Otherwise, we allow it to continue on.
772 	 * This prevents us from incorrectly nuking an entry during a
773 	 * spurious network outage.
774 	 *
775 	 * See tcp_notify().
776 	 */
777 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
778 		sc->sc_flags |= SCF_UNREACH;
779 		goto done;
780 	}
781 	syncache_drop(sc, sch);
782 	TCPSTAT_INC(tcps_sc_unreach);
783 done:
784 	SCH_UNLOCK(sch);
785 }
786 
787 /*
788  * Build a new TCP socket structure from a syncache entry.
789  *
790  * On success return the newly created socket with its underlying inp locked.
791  */
792 static struct socket *
793 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
794 {
795 	struct tcp_function_block *blk;
796 	struct inpcb *inp = NULL;
797 	struct socket *so;
798 	struct tcpcb *tp;
799 	int error;
800 	char *s;
801 
802 	NET_EPOCH_ASSERT();
803 
804 	/*
805 	 * Ok, create the full blown connection, and set things up
806 	 * as they would have been set up if we had created the
807 	 * connection when the SYN arrived.
808 	 */
809 	if ((so = solisten_clone(lso)) == NULL)
810 		goto allocfail;
811 #ifdef MAC
812 	mac_socketpeer_set_from_mbuf(m, so);
813 #endif
814 	error = in_pcballoc(so, &V_tcbinfo);
815 	if (error) {
816 		sodealloc(so);
817 		goto allocfail;
818 	}
819 	inp = sotoinpcb(so);
820 	if ((tp = tcp_newtcpcb(inp)) == NULL) {
821 		in_pcbdetach(inp);
822 		in_pcbfree(inp);
823 		sodealloc(so);
824 		goto allocfail;
825 	}
826 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
827 #ifdef INET6
828 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
829 		inp->inp_vflag &= ~INP_IPV4;
830 		inp->inp_vflag |= INP_IPV6;
831 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
832 	} else {
833 		inp->inp_vflag &= ~INP_IPV6;
834 		inp->inp_vflag |= INP_IPV4;
835 #endif
836 		inp->inp_ip_ttl = sc->sc_ip_ttl;
837 		inp->inp_ip_tos = sc->sc_ip_tos;
838 		inp->inp_laddr = sc->sc_inc.inc_laddr;
839 #ifdef INET6
840 	}
841 #endif
842 
843 	/*
844 	 * If there's an mbuf and it has a flowid, then let's initialise the
845 	 * inp with that particular flowid.
846 	 */
847 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
848 		inp->inp_flowid = m->m_pkthdr.flowid;
849 		inp->inp_flowtype = M_HASHTYPE_GET(m);
850 #ifdef NUMA
851 		inp->inp_numa_domain = m->m_pkthdr.numa_domain;
852 #endif
853 	}
854 
855 	inp->inp_lport = sc->sc_inc.inc_lport;
856 #ifdef INET6
857 	if (inp->inp_vflag & INP_IPV6PROTO) {
858 		struct inpcb *oinp = sotoinpcb(lso);
859 
860 		/*
861 		 * Inherit socket options from the listening socket.
862 		 * Note that in6p_inputopts are not (and should not be)
863 		 * copied, since it stores previously received options and is
864 		 * used to detect if each new option is different than the
865 		 * previous one and hence should be passed to a user.
866 		 * If we copied in6p_inputopts, a user would not be able to
867 		 * receive options just after calling the accept system call.
868 		 */
869 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
870 		if (oinp->in6p_outputopts)
871 			inp->in6p_outputopts =
872 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
873 		inp->in6p_hops = oinp->in6p_hops;
874 	}
875 
876 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
877 		struct in6_addr laddr6;
878 		struct sockaddr_in6 sin6;
879 
880 		sin6.sin6_family = AF_INET6;
881 		sin6.sin6_len = sizeof(sin6);
882 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
883 		sin6.sin6_port = sc->sc_inc.inc_fport;
884 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
885 		laddr6 = inp->in6p_laddr;
886 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
887 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
888 		INP_HASH_WLOCK(&V_tcbinfo);
889 		error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
890 		    thread0.td_ucred, m, false);
891 		INP_HASH_WUNLOCK(&V_tcbinfo);
892 		if (error != 0) {
893 			inp->in6p_laddr = laddr6;
894 			goto abort;
895 		}
896 		/* Override flowlabel from in6_pcbconnect. */
897 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
898 		inp->inp_flow |= sc->sc_flowlabel;
899 	}
900 #endif /* INET6 */
901 #if defined(INET) && defined(INET6)
902 	else
903 #endif
904 #ifdef INET
905 	{
906 		struct in_addr laddr;
907 		struct sockaddr_in sin;
908 
909 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
910 
911 		if (inp->inp_options == NULL) {
912 			inp->inp_options = sc->sc_ipopts;
913 			sc->sc_ipopts = NULL;
914 		}
915 
916 		sin.sin_family = AF_INET;
917 		sin.sin_len = sizeof(sin);
918 		sin.sin_addr = sc->sc_inc.inc_faddr;
919 		sin.sin_port = sc->sc_inc.inc_fport;
920 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
921 		laddr = inp->inp_laddr;
922 		if (inp->inp_laddr.s_addr == INADDR_ANY)
923 			inp->inp_laddr = sc->sc_inc.inc_laddr;
924 		INP_HASH_WLOCK(&V_tcbinfo);
925 		error = in_pcbconnect(inp, (struct sockaddr *)&sin,
926 		    thread0.td_ucred, false);
927 		INP_HASH_WUNLOCK(&V_tcbinfo);
928 		if (error != 0) {
929 			inp->inp_laddr = laddr;
930 			goto abort;
931 		}
932 	}
933 #endif /* INET */
934 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
935 	/* Copy old policy into new socket's. */
936 	if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
937 		printf("syncache_socket: could not copy policy\n");
938 #endif
939 	tp->t_state = TCPS_SYN_RECEIVED;
940 	tp->iss = sc->sc_iss;
941 	tp->irs = sc->sc_irs;
942 	tp->t_port = sc->sc_port;
943 	tcp_rcvseqinit(tp);
944 	tcp_sendseqinit(tp);
945 	blk = sototcpcb(lso)->t_fb;
946 	if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
947 		/*
948 		 * Our parents t_fb was not the default,
949 		 * we need to release our ref on tp->t_fb and
950 		 * pickup one on the new entry.
951 		 */
952 		struct tcp_function_block *rblk;
953 
954 		rblk = find_and_ref_tcp_fb(blk);
955 		KASSERT(rblk != NULL,
956 		    ("cannot find blk %p out of syncache?", blk));
957 		if (tp->t_fb->tfb_tcp_fb_fini)
958 			(*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
959 		refcount_release(&tp->t_fb->tfb_refcnt);
960 		tp->t_fb = rblk;
961 		/*
962 		 * XXXrrs this is quite dangerous, it is possible
963 		 * for the new function to fail to init. We also
964 		 * are not asking if the handoff_is_ok though at
965 		 * the very start thats probalbly ok.
966 		 */
967 		if (tp->t_fb->tfb_tcp_fb_init) {
968 			(*tp->t_fb->tfb_tcp_fb_init)(tp);
969 		}
970 	}
971 	tp->snd_wl1 = sc->sc_irs;
972 	tp->snd_max = tp->iss + 1;
973 	tp->snd_nxt = tp->iss + 1;
974 	tp->rcv_up = sc->sc_irs + 1;
975 	tp->rcv_wnd = sc->sc_wnd;
976 	tp->rcv_adv += tp->rcv_wnd;
977 	tp->last_ack_sent = tp->rcv_nxt;
978 
979 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
980 	if (sc->sc_flags & SCF_NOOPT)
981 		tp->t_flags |= TF_NOOPT;
982 	else {
983 		if (sc->sc_flags & SCF_WINSCALE) {
984 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
985 			tp->snd_scale = sc->sc_requested_s_scale;
986 			tp->request_r_scale = sc->sc_requested_r_scale;
987 		}
988 		if (sc->sc_flags & SCF_TIMESTAMP) {
989 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
990 			tp->ts_recent = sc->sc_tsreflect;
991 			tp->ts_recent_age = tcp_ts_getticks();
992 			tp->ts_offset = sc->sc_tsoff;
993 		}
994 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
995 		if (sc->sc_flags & SCF_SIGNATURE)
996 			tp->t_flags |= TF_SIGNATURE;
997 #endif
998 		if (sc->sc_flags & SCF_SACK)
999 			tp->t_flags |= TF_SACK_PERMIT;
1000 	}
1001 
1002 	tcp_ecn_syncache_socket(tp, sc);
1003 
1004 	/*
1005 	 * Set up MSS and get cached values from tcp_hostcache.
1006 	 * This might overwrite some of the defaults we just set.
1007 	 */
1008 	tcp_mss(tp, sc->sc_peer_mss);
1009 
1010 	/*
1011 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
1012 	 * limited to one segment in cc_conn_init().
1013 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
1014 	 */
1015 	if (sc->sc_rxmits > 1)
1016 		tp->snd_cwnd = 1;
1017 
1018 #ifdef TCP_OFFLOAD
1019 	/*
1020 	 * Allow a TOE driver to install its hooks.  Note that we hold the
1021 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
1022 	 * new connection before the TOE driver has done its thing.
1023 	 */
1024 	if (ADDED_BY_TOE(sc)) {
1025 		struct toedev *tod = sc->sc_tod;
1026 
1027 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
1028 	}
1029 #endif
1030 	/*
1031 	 * Copy and activate timers.
1032 	 */
1033 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
1034 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
1035 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
1036 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1037 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1038 
1039 	TCPSTAT_INC(tcps_accepts);
1040 	TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, TCPS_LISTEN);
1041 
1042 	if (!solisten_enqueue(so, SS_ISCONNECTED))
1043 		tp->t_flags |= TF_INCQUEUE;
1044 
1045 	return (so);
1046 
1047 allocfail:
1048 	/*
1049 	 * Drop the connection; we will either send a RST or have the peer
1050 	 * retransmit its SYN again after its RTO and try again.
1051 	 */
1052 	if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1053 		log(LOG_DEBUG, "%s; %s: Socket create failed "
1054 		    "due to limits or memory shortage\n",
1055 		    s, __func__);
1056 		free(s, M_TCPLOG);
1057 	}
1058 	TCPSTAT_INC(tcps_listendrop);
1059 	return (NULL);
1060 
1061 abort:
1062 	in_pcbdetach(inp);
1063 	in_pcbfree(inp);
1064 	sodealloc(so);
1065 	if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1066 		log(LOG_DEBUG, "%s; %s: in%s_pcbconnect failed with error %i\n",
1067 		    s, __func__, (sc->sc_inc.inc_flags & INC_ISIPV6) ? "6" : "",
1068 		    error);
1069 		free(s, M_TCPLOG);
1070 	}
1071 	TCPSTAT_INC(tcps_listendrop);
1072 	return (NULL);
1073 }
1074 
1075 /*
1076  * This function gets called when we receive an ACK for a
1077  * socket in the LISTEN state.  We look up the connection
1078  * in the syncache, and if its there, we pull it out of
1079  * the cache and turn it into a full-blown connection in
1080  * the SYN-RECEIVED state.
1081  *
1082  * On syncache_socket() success the newly created socket
1083  * has its underlying inp locked.
1084  */
1085 int
1086 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1087     struct socket **lsop, struct mbuf *m, uint16_t port)
1088 {
1089 	struct syncache *sc;
1090 	struct syncache_head *sch;
1091 	struct syncache scs;
1092 	char *s;
1093 	bool locked;
1094 
1095 	NET_EPOCH_ASSERT();
1096 	KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1097 	    ("%s: can handle only ACK", __func__));
1098 
1099 	if (syncache_cookiesonly()) {
1100 		sc = NULL;
1101 		sch = syncache_hashbucket(inc);
1102 		locked = false;
1103 	} else {
1104 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1105 		locked = true;
1106 		SCH_LOCK_ASSERT(sch);
1107 	}
1108 
1109 #ifdef INVARIANTS
1110 	/*
1111 	 * Test code for syncookies comparing the syncache stored
1112 	 * values with the reconstructed values from the cookie.
1113 	 */
1114 	if (sc != NULL)
1115 		syncookie_cmp(inc, sch, sc, th, to, *lsop, port);
1116 #endif
1117 
1118 	if (sc == NULL) {
1119 		/*
1120 		 * There is no syncache entry, so see if this ACK is
1121 		 * a returning syncookie.  To do this, first:
1122 		 *  A. Check if syncookies are used in case of syncache
1123 		 *     overflows
1124 		 *  B. See if this socket has had a syncache entry dropped in
1125 		 *     the recent past. We don't want to accept a bogus
1126 		 *     syncookie if we've never received a SYN or accept it
1127 		 *     twice.
1128 		 *  C. check that the syncookie is valid.  If it is, then
1129 		 *     cobble up a fake syncache entry, and return.
1130 		 */
1131 		if (locked && !V_tcp_syncookies) {
1132 			SCH_UNLOCK(sch);
1133 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1134 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1135 				    "segment rejected (syncookies disabled)\n",
1136 				    s, __func__);
1137 			goto failed;
1138 		}
1139 		if (locked && !V_tcp_syncookiesonly &&
1140 		    sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1141 			SCH_UNLOCK(sch);
1142 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1143 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1144 				    "segment rejected (no syncache entry)\n",
1145 				    s, __func__);
1146 			goto failed;
1147 		}
1148 		bzero(&scs, sizeof(scs));
1149 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port);
1150 		if (locked)
1151 			SCH_UNLOCK(sch);
1152 		if (sc == NULL) {
1153 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1154 				log(LOG_DEBUG, "%s; %s: Segment failed "
1155 				    "SYNCOOKIE authentication, segment rejected "
1156 				    "(probably spoofed)\n", s, __func__);
1157 			goto failed;
1158 		}
1159 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1160 		/* If received ACK has MD5 signature, check it. */
1161 		if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1162 		    (!TCPMD5_ENABLED() ||
1163 		    TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1164 			/* Drop the ACK. */
1165 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1166 				log(LOG_DEBUG, "%s; %s: Segment rejected, "
1167 				    "MD5 signature doesn't match.\n",
1168 				    s, __func__);
1169 				free(s, M_TCPLOG);
1170 			}
1171 			TCPSTAT_INC(tcps_sig_err_sigopt);
1172 			return (-1); /* Do not send RST */
1173 		}
1174 #endif /* TCP_SIGNATURE */
1175 		TCPSTATES_INC(TCPS_SYN_RECEIVED);
1176 	} else {
1177 		if (sc->sc_port != port) {
1178 			SCH_UNLOCK(sch);
1179 			return (0);
1180 		}
1181 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1182 		/*
1183 		 * If listening socket requested TCP digests, check that
1184 		 * received ACK has signature and it is correct.
1185 		 * If not, drop the ACK and leave sc entry in th cache,
1186 		 * because SYN was received with correct signature.
1187 		 */
1188 		if (sc->sc_flags & SCF_SIGNATURE) {
1189 			if ((to->to_flags & TOF_SIGNATURE) == 0) {
1190 				/* No signature */
1191 				TCPSTAT_INC(tcps_sig_err_nosigopt);
1192 				SCH_UNLOCK(sch);
1193 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1194 					log(LOG_DEBUG, "%s; %s: Segment "
1195 					    "rejected, MD5 signature wasn't "
1196 					    "provided.\n", s, __func__);
1197 					free(s, M_TCPLOG);
1198 				}
1199 				return (-1); /* Do not send RST */
1200 			}
1201 			if (!TCPMD5_ENABLED() ||
1202 			    TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1203 				/* Doesn't match or no SA */
1204 				SCH_UNLOCK(sch);
1205 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1206 					log(LOG_DEBUG, "%s; %s: Segment "
1207 					    "rejected, MD5 signature doesn't "
1208 					    "match.\n", s, __func__);
1209 					free(s, M_TCPLOG);
1210 				}
1211 				return (-1); /* Do not send RST */
1212 			}
1213 		}
1214 #endif /* TCP_SIGNATURE */
1215 
1216 		/*
1217 		 * RFC 7323 PAWS: If we have a timestamp on this segment and
1218 		 * it's less than ts_recent, drop it.
1219 		 * XXXMT: RFC 7323 also requires to send an ACK.
1220 		 *        In tcp_input.c this is only done for TCP segments
1221 		 *        with user data, so be consistent here and just drop
1222 		 *        the segment.
1223 		 */
1224 		if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1225 		    TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1226 			SCH_UNLOCK(sch);
1227 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1228 				log(LOG_DEBUG,
1229 				    "%s; %s: SEG.TSval %u < TS.Recent %u, "
1230 				    "segment dropped\n", s, __func__,
1231 				    to->to_tsval, sc->sc_tsreflect);
1232 				free(s, M_TCPLOG);
1233 			}
1234 			return (-1);  /* Do not send RST */
1235 		}
1236 
1237 		/*
1238 		 * If timestamps were not negotiated during SYN/ACK and a
1239 		 * segment with a timestamp is received, ignore the
1240 		 * timestamp and process the packet normally.
1241 		 * See section 3.2 of RFC 7323.
1242 		 */
1243 		if (!(sc->sc_flags & SCF_TIMESTAMP) &&
1244 		    (to->to_flags & TOF_TS)) {
1245 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1246 				log(LOG_DEBUG, "%s; %s: Timestamp not "
1247 				    "expected, segment processed normally\n",
1248 				    s, __func__);
1249 				free(s, M_TCPLOG);
1250 				s = NULL;
1251 			}
1252 		}
1253 
1254 		/*
1255 		 * If timestamps were negotiated during SYN/ACK and a
1256 		 * segment without a timestamp is received, silently drop
1257 		 * the segment, unless the missing timestamps are tolerated.
1258 		 * See section 3.2 of RFC 7323.
1259 		 */
1260 		if ((sc->sc_flags & SCF_TIMESTAMP) &&
1261 		    !(to->to_flags & TOF_TS)) {
1262 			if (V_tcp_tolerate_missing_ts) {
1263 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1264 					log(LOG_DEBUG,
1265 					    "%s; %s: Timestamp missing, "
1266 					    "segment processed normally\n",
1267 					    s, __func__);
1268 					free(s, M_TCPLOG);
1269 				}
1270 			} else {
1271 				SCH_UNLOCK(sch);
1272 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1273 					log(LOG_DEBUG,
1274 					    "%s; %s: Timestamp missing, "
1275 					    "segment silently dropped\n",
1276 					    s, __func__);
1277 					free(s, M_TCPLOG);
1278 				}
1279 				return (-1);  /* Do not send RST */
1280 			}
1281 		}
1282 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1283 		sch->sch_length--;
1284 #ifdef TCP_OFFLOAD
1285 		if (ADDED_BY_TOE(sc)) {
1286 			struct toedev *tod = sc->sc_tod;
1287 
1288 			tod->tod_syncache_removed(tod, sc->sc_todctx);
1289 		}
1290 #endif
1291 		SCH_UNLOCK(sch);
1292 	}
1293 
1294 	/*
1295 	 * Segment validation:
1296 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1297 	 */
1298 	if (th->th_ack != sc->sc_iss + 1) {
1299 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1300 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1301 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1302 		goto failed;
1303 	}
1304 
1305 	/*
1306 	 * The SEQ must fall in the window starting at the received
1307 	 * initial receive sequence number + 1 (the SYN).
1308 	 */
1309 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1310 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1311 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1312 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1313 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1314 		goto failed;
1315 	}
1316 
1317 	*lsop = syncache_socket(sc, *lsop, m);
1318 
1319 	if (*lsop == NULL)
1320 		TCPSTAT_INC(tcps_sc_aborted);
1321 	else
1322 		TCPSTAT_INC(tcps_sc_completed);
1323 
1324 /* how do we find the inp for the new socket? */
1325 	if (sc != &scs)
1326 		syncache_free(sc);
1327 	return (1);
1328 failed:
1329 	if (sc != NULL) {
1330 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1331 		if (sc != &scs)
1332 			syncache_free(sc);
1333 	}
1334 	if (s != NULL)
1335 		free(s, M_TCPLOG);
1336 	*lsop = NULL;
1337 	return (0);
1338 }
1339 
1340 static struct socket *
1341 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m,
1342     uint64_t response_cookie)
1343 {
1344 	struct inpcb *inp;
1345 	struct tcpcb *tp;
1346 	unsigned int *pending_counter;
1347 	struct socket *so;
1348 
1349 	NET_EPOCH_ASSERT();
1350 
1351 	pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending;
1352 	so = syncache_socket(sc, lso, m);
1353 	if (so == NULL) {
1354 		TCPSTAT_INC(tcps_sc_aborted);
1355 		atomic_subtract_int(pending_counter, 1);
1356 	} else {
1357 		soisconnected(so);
1358 		inp = sotoinpcb(so);
1359 		tp = intotcpcb(inp);
1360 		tp->t_flags |= TF_FASTOPEN;
1361 		tp->t_tfo_cookie.server = response_cookie;
1362 		tp->snd_max = tp->iss;
1363 		tp->snd_nxt = tp->iss;
1364 		tp->t_tfo_pending = pending_counter;
1365 		TCPSTATES_INC(TCPS_SYN_RECEIVED);
1366 		TCPSTAT_INC(tcps_sc_completed);
1367 	}
1368 
1369 	return (so);
1370 }
1371 
1372 /*
1373  * Given a LISTEN socket and an inbound SYN request, add
1374  * this to the syn cache, and send back a segment:
1375  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1376  * to the source.
1377  *
1378  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1379  * Doing so would require that we hold onto the data and deliver it
1380  * to the application.  However, if we are the target of a SYN-flood
1381  * DoS attack, an attacker could send data which would eventually
1382  * consume all available buffer space if it were ACKed.  By not ACKing
1383  * the data, we avoid this DoS scenario.
1384  *
1385  * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1386  * cookie is processed and a new socket is created.  In this case, any data
1387  * accompanying the SYN will be queued to the socket by tcp_input() and will
1388  * be ACKed either when the application sends response data or the delayed
1389  * ACK timer expires, whichever comes first.
1390  */
1391 struct socket *
1392 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1393     struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod,
1394     void *todctx, uint8_t iptos, uint16_t port)
1395 {
1396 	struct tcpcb *tp;
1397 	struct socket *rv = NULL;
1398 	struct syncache *sc = NULL;
1399 	struct syncache_head *sch;
1400 	struct mbuf *ipopts = NULL;
1401 	u_int ltflags;
1402 	int win, ip_ttl, ip_tos;
1403 	char *s;
1404 #ifdef INET6
1405 	int autoflowlabel = 0;
1406 #endif
1407 #ifdef MAC
1408 	struct label *maclabel;
1409 #endif
1410 	struct syncache scs;
1411 	struct ucred *cred;
1412 	uint64_t tfo_response_cookie;
1413 	unsigned int *tfo_pending = NULL;
1414 	int tfo_cookie_valid = 0;
1415 	int tfo_response_cookie_valid = 0;
1416 	bool locked;
1417 
1418 	INP_RLOCK_ASSERT(inp);			/* listen socket */
1419 	KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1420 	    ("%s: unexpected tcp flags", __func__));
1421 
1422 	/*
1423 	 * Combine all so/tp operations very early to drop the INP lock as
1424 	 * soon as possible.
1425 	 */
1426 	KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1427 	tp = sototcpcb(so);
1428 	cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred);
1429 
1430 #ifdef INET6
1431 	if (inc->inc_flags & INC_ISIPV6) {
1432 		if (inp->inp_flags & IN6P_AUTOFLOWLABEL) {
1433 			autoflowlabel = 1;
1434 		}
1435 		ip_ttl = in6_selecthlim(inp, NULL);
1436 		if ((inp->in6p_outputopts == NULL) ||
1437 		    (inp->in6p_outputopts->ip6po_tclass == -1)) {
1438 			ip_tos = 0;
1439 		} else {
1440 			ip_tos = inp->in6p_outputopts->ip6po_tclass;
1441 		}
1442 	}
1443 #endif
1444 #if defined(INET6) && defined(INET)
1445 	else
1446 #endif
1447 #ifdef INET
1448 	{
1449 		ip_ttl = inp->inp_ip_ttl;
1450 		ip_tos = inp->inp_ip_tos;
1451 	}
1452 #endif
1453 	win = so->sol_sbrcv_hiwat;
1454 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1455 
1456 	if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
1457 	    (tp->t_tfo_pending != NULL) &&
1458 	    (to->to_flags & TOF_FASTOPEN)) {
1459 		/*
1460 		 * Limit the number of pending TFO connections to
1461 		 * approximately half of the queue limit.  This prevents TFO
1462 		 * SYN floods from starving the service by filling the
1463 		 * listen queue with bogus TFO connections.
1464 		 */
1465 		if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1466 		    (so->sol_qlimit / 2)) {
1467 			int result;
1468 
1469 			result = tcp_fastopen_check_cookie(inc,
1470 			    to->to_tfo_cookie, to->to_tfo_len,
1471 			    &tfo_response_cookie);
1472 			tfo_cookie_valid = (result > 0);
1473 			tfo_response_cookie_valid = (result >= 0);
1474 		}
1475 
1476 		/*
1477 		 * Remember the TFO pending counter as it will have to be
1478 		 * decremented below if we don't make it to syncache_tfo_expand().
1479 		 */
1480 		tfo_pending = tp->t_tfo_pending;
1481 	}
1482 
1483 #ifdef MAC
1484 	if (mac_syncache_init(&maclabel) != 0) {
1485 		INP_RUNLOCK(inp);
1486 		goto done;
1487 	} else
1488 		mac_syncache_create(maclabel, inp);
1489 #endif
1490 	if (!tfo_cookie_valid)
1491 		INP_RUNLOCK(inp);
1492 
1493 	/*
1494 	 * Remember the IP options, if any.
1495 	 */
1496 #ifdef INET6
1497 	if (!(inc->inc_flags & INC_ISIPV6))
1498 #endif
1499 #ifdef INET
1500 		ipopts = (m) ? ip_srcroute(m) : NULL;
1501 #else
1502 		ipopts = NULL;
1503 #endif
1504 
1505 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1506 	/*
1507 	 * When the socket is TCP-MD5 enabled check that,
1508 	 *  - a signed packet is valid
1509 	 *  - a non-signed packet does not have a security association
1510 	 *
1511 	 *  If a signed packet fails validation or a non-signed packet has a
1512 	 *  security association, the packet will be dropped.
1513 	 */
1514 	if (ltflags & TF_SIGNATURE) {
1515 		if (to->to_flags & TOF_SIGNATURE) {
1516 			if (!TCPMD5_ENABLED() ||
1517 			    TCPMD5_INPUT(m, th, to->to_signature) != 0)
1518 				goto done;
1519 		} else {
1520 			if (TCPMD5_ENABLED() &&
1521 			    TCPMD5_INPUT(m, NULL, NULL) != ENOENT)
1522 				goto done;
1523 		}
1524 	} else if (to->to_flags & TOF_SIGNATURE)
1525 		goto done;
1526 #endif	/* TCP_SIGNATURE */
1527 	/*
1528 	 * See if we already have an entry for this connection.
1529 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1530 	 *
1531 	 * XXX: should the syncache be re-initialized with the contents
1532 	 * of the new SYN here (which may have different options?)
1533 	 *
1534 	 * XXX: We do not check the sequence number to see if this is a
1535 	 * real retransmit or a new connection attempt.  The question is
1536 	 * how to handle such a case; either ignore it as spoofed, or
1537 	 * drop the current entry and create a new one?
1538 	 */
1539 	if (syncache_cookiesonly()) {
1540 		sc = NULL;
1541 		sch = syncache_hashbucket(inc);
1542 		locked = false;
1543 	} else {
1544 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1545 		locked = true;
1546 		SCH_LOCK_ASSERT(sch);
1547 	}
1548 	if (sc != NULL) {
1549 		if (tfo_cookie_valid)
1550 			INP_RUNLOCK(inp);
1551 		TCPSTAT_INC(tcps_sc_dupsyn);
1552 		if (ipopts) {
1553 			/*
1554 			 * If we were remembering a previous source route,
1555 			 * forget it and use the new one we've been given.
1556 			 */
1557 			if (sc->sc_ipopts)
1558 				(void) m_free(sc->sc_ipopts);
1559 			sc->sc_ipopts = ipopts;
1560 		}
1561 		/*
1562 		 * Update timestamp if present.
1563 		 */
1564 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1565 			sc->sc_tsreflect = to->to_tsval;
1566 		else
1567 			sc->sc_flags &= ~SCF_TIMESTAMP;
1568 		/*
1569 		 * Disable ECN if needed.
1570 		 */
1571 		if ((sc->sc_flags & SCF_ECN) &&
1572 		    ((tcp_get_flags(th) & (TH_ECE|TH_CWR)) != (TH_ECE|TH_CWR))) {
1573 			sc->sc_flags &= ~SCF_ECN;
1574 		}
1575 #ifdef MAC
1576 		/*
1577 		 * Since we have already unconditionally allocated label
1578 		 * storage, free it up.  The syncache entry will already
1579 		 * have an initialized label we can use.
1580 		 */
1581 		mac_syncache_destroy(&maclabel);
1582 #endif
1583 		TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1584 		/* Retransmit SYN|ACK and reset retransmit count. */
1585 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1586 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1587 			    "resetting timer and retransmitting SYN|ACK\n",
1588 			    s, __func__);
1589 			free(s, M_TCPLOG);
1590 		}
1591 		if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1592 			sc->sc_rxmits = 0;
1593 			syncache_timeout(sc, sch, 1);
1594 			TCPSTAT_INC(tcps_sndacks);
1595 			TCPSTAT_INC(tcps_sndtotal);
1596 		}
1597 		SCH_UNLOCK(sch);
1598 		goto donenoprobe;
1599 	}
1600 
1601 	if (tfo_cookie_valid) {
1602 		bzero(&scs, sizeof(scs));
1603 		sc = &scs;
1604 		goto skip_alloc;
1605 	}
1606 
1607 	/*
1608 	 * Skip allocating a syncache entry if we are just going to discard
1609 	 * it later.
1610 	 */
1611 	if (!locked) {
1612 		bzero(&scs, sizeof(scs));
1613 		sc = &scs;
1614 	} else
1615 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1616 	if (sc == NULL) {
1617 		/*
1618 		 * The zone allocator couldn't provide more entries.
1619 		 * Treat this as if the cache was full; drop the oldest
1620 		 * entry and insert the new one.
1621 		 */
1622 		TCPSTAT_INC(tcps_sc_zonefail);
1623 		if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1624 			sch->sch_last_overflow = time_uptime;
1625 			syncache_drop(sc, sch);
1626 			syncache_pause(inc);
1627 		}
1628 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1629 		if (sc == NULL) {
1630 			if (V_tcp_syncookies) {
1631 				bzero(&scs, sizeof(scs));
1632 				sc = &scs;
1633 			} else {
1634 				KASSERT(locked,
1635 				    ("%s: bucket unexpectedly unlocked",
1636 				    __func__));
1637 				SCH_UNLOCK(sch);
1638 				if (ipopts)
1639 					(void) m_free(ipopts);
1640 				goto done;
1641 			}
1642 		}
1643 	}
1644 
1645 skip_alloc:
1646 	if (!tfo_cookie_valid && tfo_response_cookie_valid)
1647 		sc->sc_tfo_cookie = &tfo_response_cookie;
1648 
1649 	/*
1650 	 * Fill in the syncache values.
1651 	 */
1652 #ifdef MAC
1653 	sc->sc_label = maclabel;
1654 #endif
1655 	sc->sc_cred = cred;
1656 	sc->sc_port = port;
1657 	cred = NULL;
1658 	sc->sc_ipopts = ipopts;
1659 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1660 	sc->sc_ip_tos = ip_tos;
1661 	sc->sc_ip_ttl = ip_ttl;
1662 #ifdef TCP_OFFLOAD
1663 	sc->sc_tod = tod;
1664 	sc->sc_todctx = todctx;
1665 #endif
1666 	sc->sc_irs = th->th_seq;
1667 	sc->sc_flags = 0;
1668 	sc->sc_flowlabel = 0;
1669 
1670 	/*
1671 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1672 	 * win was derived from socket earlier in the function.
1673 	 */
1674 	win = imax(win, 0);
1675 	win = imin(win, TCP_MAXWIN);
1676 	sc->sc_wnd = win;
1677 
1678 	if (V_tcp_do_rfc1323 &&
1679 	    !(ltflags & TF_NOOPT)) {
1680 		/*
1681 		 * A timestamp received in a SYN makes
1682 		 * it ok to send timestamp requests and replies.
1683 		 */
1684 		if (to->to_flags & TOF_TS) {
1685 			sc->sc_tsreflect = to->to_tsval;
1686 			sc->sc_flags |= SCF_TIMESTAMP;
1687 			sc->sc_tsoff = tcp_new_ts_offset(inc);
1688 		}
1689 		if (to->to_flags & TOF_SCALE) {
1690 			int wscale = 0;
1691 
1692 			/*
1693 			 * Pick the smallest possible scaling factor that
1694 			 * will still allow us to scale up to sb_max, aka
1695 			 * kern.ipc.maxsockbuf.
1696 			 *
1697 			 * We do this because there are broken firewalls that
1698 			 * will corrupt the window scale option, leading to
1699 			 * the other endpoint believing that our advertised
1700 			 * window is unscaled.  At scale factors larger than
1701 			 * 5 the unscaled window will drop below 1500 bytes,
1702 			 * leading to serious problems when traversing these
1703 			 * broken firewalls.
1704 			 *
1705 			 * With the default maxsockbuf of 256K, a scale factor
1706 			 * of 3 will be chosen by this algorithm.  Those who
1707 			 * choose a larger maxsockbuf should watch out
1708 			 * for the compatibility problems mentioned above.
1709 			 *
1710 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1711 			 * or <SYN,ACK>) segment itself is never scaled.
1712 			 */
1713 			while (wscale < TCP_MAX_WINSHIFT &&
1714 			    (TCP_MAXWIN << wscale) < sb_max)
1715 				wscale++;
1716 			sc->sc_requested_r_scale = wscale;
1717 			sc->sc_requested_s_scale = to->to_wscale;
1718 			sc->sc_flags |= SCF_WINSCALE;
1719 		}
1720 	}
1721 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1722 	/*
1723 	 * If incoming packet has an MD5 signature, flag this in the
1724 	 * syncache so that syncache_respond() will do the right thing
1725 	 * with the SYN+ACK.
1726 	 */
1727 	if (to->to_flags & TOF_SIGNATURE)
1728 		sc->sc_flags |= SCF_SIGNATURE;
1729 #endif	/* TCP_SIGNATURE */
1730 	if (to->to_flags & TOF_SACKPERM)
1731 		sc->sc_flags |= SCF_SACK;
1732 	if (to->to_flags & TOF_MSS)
1733 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1734 	if (ltflags & TF_NOOPT)
1735 		sc->sc_flags |= SCF_NOOPT;
1736 	/* ECN Handshake */
1737 	if (V_tcp_do_ecn)
1738 		sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1739 
1740 	if (V_tcp_syncookies)
1741 		sc->sc_iss = syncookie_generate(sch, sc);
1742 	else
1743 		sc->sc_iss = arc4random();
1744 #ifdef INET6
1745 	if (autoflowlabel) {
1746 		if (V_tcp_syncookies)
1747 			sc->sc_flowlabel = sc->sc_iss;
1748 		else
1749 			sc->sc_flowlabel = ip6_randomflowlabel();
1750 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1751 	}
1752 #endif
1753 	if (locked)
1754 		SCH_UNLOCK(sch);
1755 
1756 	if (tfo_cookie_valid) {
1757 		rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie);
1758 		/* INP_RUNLOCK(inp) will be performed by the caller */
1759 		goto tfo_expanded;
1760 	}
1761 
1762 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1763 	/*
1764 	 * Do a standard 3-way handshake.
1765 	 */
1766 	if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1767 		if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1768 			syncache_free(sc);
1769 		else if (sc != &scs)
1770 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1771 		TCPSTAT_INC(tcps_sndacks);
1772 		TCPSTAT_INC(tcps_sndtotal);
1773 	} else {
1774 		if (sc != &scs)
1775 			syncache_free(sc);
1776 		TCPSTAT_INC(tcps_sc_dropped);
1777 	}
1778 	goto donenoprobe;
1779 
1780 done:
1781 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1782 donenoprobe:
1783 	if (m)
1784 		m_freem(m);
1785 	/*
1786 	 * If tfo_pending is not NULL here, then a TFO SYN that did not
1787 	 * result in a new socket was processed and the associated pending
1788 	 * counter has not yet been decremented.  All such TFO processing paths
1789 	 * transit this point.
1790 	 */
1791 	if (tfo_pending != NULL)
1792 		tcp_fastopen_decrement_counter(tfo_pending);
1793 
1794 tfo_expanded:
1795 	if (cred != NULL)
1796 		crfree(cred);
1797 #ifdef MAC
1798 	if (sc == &scs)
1799 		mac_syncache_destroy(&maclabel);
1800 #endif
1801 	return (rv);
1802 }
1803 
1804 /*
1805  * Send SYN|ACK or ACK to the peer.  Either in response to a peer's segment,
1806  * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1807  */
1808 static int
1809 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1810 {
1811 	struct ip *ip = NULL;
1812 	struct mbuf *m;
1813 	struct tcphdr *th = NULL;
1814 	struct udphdr *udp = NULL;
1815 	int optlen, error = 0;	/* Make compiler happy */
1816 	u_int16_t hlen, tlen, mssopt, ulen;
1817 	struct tcpopt to;
1818 #ifdef INET6
1819 	struct ip6_hdr *ip6 = NULL;
1820 #endif
1821 
1822 	NET_EPOCH_ASSERT();
1823 
1824 	hlen =
1825 #ifdef INET6
1826 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1827 #endif
1828 		sizeof(struct ip);
1829 	tlen = hlen + sizeof(struct tcphdr);
1830 	if (sc->sc_port) {
1831 		tlen += sizeof(struct udphdr);
1832 	}
1833 	/* Determine MSS we advertize to other end of connection. */
1834 	mssopt = tcp_mssopt(&sc->sc_inc);
1835 	if (sc->sc_port)
1836 		mssopt -= V_tcp_udp_tunneling_overhead;
1837 	mssopt = max(mssopt, V_tcp_minmss);
1838 
1839 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1840 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1841 	    ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + "
1842 	    "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port,
1843 	    max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN));
1844 
1845 	/* Create the IP+TCP header from scratch. */
1846 	m = m_gethdr(M_NOWAIT, MT_DATA);
1847 	if (m == NULL)
1848 		return (ENOBUFS);
1849 #ifdef MAC
1850 	mac_syncache_create_mbuf(sc->sc_label, m);
1851 #endif
1852 	m->m_data += max_linkhdr;
1853 	m->m_len = tlen;
1854 	m->m_pkthdr.len = tlen;
1855 	m->m_pkthdr.rcvif = NULL;
1856 
1857 #ifdef INET6
1858 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1859 		ip6 = mtod(m, struct ip6_hdr *);
1860 		ip6->ip6_vfc = IPV6_VERSION;
1861 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1862 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1863 		ip6->ip6_plen = htons(tlen - hlen);
1864 		/* ip6_hlim is set after checksum */
1865 		/* Zero out traffic class and flow label. */
1866 		ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1867 		ip6->ip6_flow |= sc->sc_flowlabel;
1868 		if (sc->sc_port != 0) {
1869 			ip6->ip6_nxt = IPPROTO_UDP;
1870 			udp = (struct udphdr *)(ip6 + 1);
1871 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1872 			udp->uh_dport = sc->sc_port;
1873 			ulen = (tlen - sizeof(struct ip6_hdr));
1874 			th = (struct tcphdr *)(udp + 1);
1875 		} else {
1876 			ip6->ip6_nxt = IPPROTO_TCP;
1877 			th = (struct tcphdr *)(ip6 + 1);
1878 		}
1879 		ip6->ip6_flow |= htonl(sc->sc_ip_tos << 20);
1880 	}
1881 #endif
1882 #if defined(INET6) && defined(INET)
1883 	else
1884 #endif
1885 #ifdef INET
1886 	{
1887 		ip = mtod(m, struct ip *);
1888 		ip->ip_v = IPVERSION;
1889 		ip->ip_hl = sizeof(struct ip) >> 2;
1890 		ip->ip_len = htons(tlen);
1891 		ip->ip_id = 0;
1892 		ip->ip_off = 0;
1893 		ip->ip_sum = 0;
1894 		ip->ip_src = sc->sc_inc.inc_laddr;
1895 		ip->ip_dst = sc->sc_inc.inc_faddr;
1896 		ip->ip_ttl = sc->sc_ip_ttl;
1897 		ip->ip_tos = sc->sc_ip_tos;
1898 
1899 		/*
1900 		 * See if we should do MTU discovery.  Route lookups are
1901 		 * expensive, so we will only unset the DF bit if:
1902 		 *
1903 		 *	1) path_mtu_discovery is disabled
1904 		 *	2) the SCF_UNREACH flag has been set
1905 		 */
1906 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1907 		       ip->ip_off |= htons(IP_DF);
1908 		if (sc->sc_port == 0) {
1909 			ip->ip_p = IPPROTO_TCP;
1910 			th = (struct tcphdr *)(ip + 1);
1911 		} else {
1912 			ip->ip_p = IPPROTO_UDP;
1913 			udp = (struct udphdr *)(ip + 1);
1914 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1915 			udp->uh_dport = sc->sc_port;
1916 			ulen = (tlen - sizeof(struct ip));
1917 			th = (struct tcphdr *)(udp + 1);
1918 		}
1919 	}
1920 #endif /* INET */
1921 	th->th_sport = sc->sc_inc.inc_lport;
1922 	th->th_dport = sc->sc_inc.inc_fport;
1923 
1924 	if (flags & TH_SYN)
1925 		th->th_seq = htonl(sc->sc_iss);
1926 	else
1927 		th->th_seq = htonl(sc->sc_iss + 1);
1928 	th->th_ack = htonl(sc->sc_irs + 1);
1929 	th->th_off = sizeof(struct tcphdr) >> 2;
1930 	th->th_win = htons(sc->sc_wnd);
1931 	th->th_urp = 0;
1932 
1933 	flags = tcp_ecn_syncache_respond(flags, sc);
1934 	tcp_set_flags(th, flags);
1935 
1936 	/* Tack on the TCP options. */
1937 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1938 		to.to_flags = 0;
1939 
1940 		if (flags & TH_SYN) {
1941 			to.to_mss = mssopt;
1942 			to.to_flags = TOF_MSS;
1943 			if (sc->sc_flags & SCF_WINSCALE) {
1944 				to.to_wscale = sc->sc_requested_r_scale;
1945 				to.to_flags |= TOF_SCALE;
1946 			}
1947 			if (sc->sc_flags & SCF_SACK)
1948 				to.to_flags |= TOF_SACKPERM;
1949 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1950 			if (sc->sc_flags & SCF_SIGNATURE)
1951 				to.to_flags |= TOF_SIGNATURE;
1952 #endif
1953 			if (sc->sc_tfo_cookie) {
1954 				to.to_flags |= TOF_FASTOPEN;
1955 				to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1956 				to.to_tfo_cookie = sc->sc_tfo_cookie;
1957 				/* don't send cookie again when retransmitting response */
1958 				sc->sc_tfo_cookie = NULL;
1959 			}
1960 		}
1961 		if (sc->sc_flags & SCF_TIMESTAMP) {
1962 			to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1963 			to.to_tsecr = sc->sc_tsreflect;
1964 			to.to_flags |= TOF_TS;
1965 		}
1966 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1967 
1968 		/* Adjust headers by option size. */
1969 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1970 		m->m_len += optlen;
1971 		m->m_pkthdr.len += optlen;
1972 #ifdef INET6
1973 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1974 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1975 		else
1976 #endif
1977 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1978 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1979 		if (sc->sc_flags & SCF_SIGNATURE) {
1980 			KASSERT(to.to_flags & TOF_SIGNATURE,
1981 			    ("tcp_addoptions() didn't set tcp_signature"));
1982 
1983 			/* NOTE: to.to_signature is inside of mbuf */
1984 			if (!TCPMD5_ENABLED() ||
1985 			    TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1986 				m_freem(m);
1987 				return (EACCES);
1988 			}
1989 		}
1990 #endif
1991 	} else
1992 		optlen = 0;
1993 
1994 	if (udp) {
1995 		ulen += optlen;
1996 		udp->uh_ulen = htons(ulen);
1997 	}
1998 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
1999 	/*
2000 	 * If we have peer's SYN and it has a flowid, then let's assign it to
2001 	 * our SYN|ACK.  ip6_output() and ip_output() will not assign flowid
2002 	 * to SYN|ACK due to lack of inp here.
2003 	 */
2004 	if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
2005 		m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
2006 		M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
2007 	}
2008 #ifdef INET6
2009 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
2010 		if (sc->sc_port) {
2011 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
2012 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2013 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen,
2014 			      IPPROTO_UDP, 0);
2015 			th->th_sum = htons(0);
2016 		} else {
2017 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
2018 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2019 			th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
2020 			    IPPROTO_TCP, 0);
2021 		}
2022 		ip6->ip6_hlim = sc->sc_ip_ttl;
2023 #ifdef TCP_OFFLOAD
2024 		if (ADDED_BY_TOE(sc)) {
2025 			struct toedev *tod = sc->sc_tod;
2026 
2027 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2028 
2029 			return (error);
2030 		}
2031 #endif
2032 		TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
2033 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
2034 	}
2035 #endif
2036 #if defined(INET6) && defined(INET)
2037 	else
2038 #endif
2039 #ifdef INET
2040 	{
2041 		if (sc->sc_port) {
2042 			m->m_pkthdr.csum_flags = CSUM_UDP;
2043 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2044 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
2045 			      ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
2046 			th->th_sum = htons(0);
2047 		} else {
2048 			m->m_pkthdr.csum_flags = CSUM_TCP;
2049 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2050 			th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
2051 			    htons(tlen + optlen - hlen + IPPROTO_TCP));
2052 		}
2053 #ifdef TCP_OFFLOAD
2054 		if (ADDED_BY_TOE(sc)) {
2055 			struct toedev *tod = sc->sc_tod;
2056 
2057 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2058 
2059 			return (error);
2060 		}
2061 #endif
2062 		TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
2063 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
2064 	}
2065 #endif
2066 	return (error);
2067 }
2068 
2069 /*
2070  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
2071  * that exceed the capacity of the syncache by avoiding the storage of any
2072  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
2073  * attacks where the attacker does not have access to our responses.
2074  *
2075  * Syncookies encode and include all necessary information about the
2076  * connection setup within the SYN|ACK that we send back.  That way we
2077  * can avoid keeping any local state until the ACK to our SYN|ACK returns
2078  * (if ever).  Normally the syncache and syncookies are running in parallel
2079  * with the latter taking over when the former is exhausted.  When matching
2080  * syncache entry is found the syncookie is ignored.
2081  *
2082  * The only reliable information persisting the 3WHS is our initial sequence
2083  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
2084  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
2085  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
2086  * returns and signifies a legitimate connection if it matches the ACK.
2087  *
2088  * The available space of 32 bits to store the hash and to encode the SYN
2089  * option information is very tight and we should have at least 24 bits for
2090  * the MAC to keep the number of guesses by blind spoofing reasonably high.
2091  *
2092  * SYN option information we have to encode to fully restore a connection:
2093  * MSS: is imporant to chose an optimal segment size to avoid IP level
2094  *   fragmentation along the path.  The common MSS values can be encoded
2095  *   in a 3-bit table.  Uncommon values are captured by the next lower value
2096  *   in the table leading to a slight increase in packetization overhead.
2097  * WSCALE: is necessary to allow large windows to be used for high delay-
2098  *   bandwidth product links.  Not scaling the window when it was initially
2099  *   negotiated is bad for performance as lack of scaling further decreases
2100  *   the apparent available send window.  We only need to encode the WSCALE
2101  *   we received from the remote end.  Our end can be recalculated at any
2102  *   time.  The common WSCALE values can be encoded in a 3-bit table.
2103  *   Uncommon values are captured by the next lower value in the table
2104  *   making us under-estimate the available window size halving our
2105  *   theoretically possible maximum throughput for that connection.
2106  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2107  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2108  *   that are included in all segments on a connection.  We enable them when
2109  *   the ACK has them.
2110  *
2111  * Security of syncookies and attack vectors:
2112  *
2113  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2114  * together with the gloabl secret to make it unique per connection attempt.
2115  * Thus any change of any of those parameters results in a different MAC output
2116  * in an unpredictable way unless a collision is encountered.  24 bits of the
2117  * MAC are embedded into the ISS.
2118  *
2119  * To prevent replay attacks two rotating global secrets are updated with a
2120  * new random value every 15 seconds.  The life-time of a syncookie is thus
2121  * 15-30 seconds.
2122  *
2123  * Vector 1: Attacking the secret.  This requires finding a weakness in the
2124  * MAC itself or the way it is used here.  The attacker can do a chosen plain
2125  * text attack by varying and testing the all parameters under his control.
2126  * The strength depends on the size and randomness of the secret, and the
2127  * cryptographic security of the MAC function.  Due to the constant updating
2128  * of the secret the attacker has at most 29.999 seconds to find the secret
2129  * and launch spoofed connections.  After that he has to start all over again.
2130  *
2131  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
2132  * size an average of 4,823 attempts are required for a 50% chance of success
2133  * to spoof a single syncookie (birthday collision paradox).  However the
2134  * attacker is blind and doesn't know if one of his attempts succeeded unless
2135  * he has a side channel to interfere success from.  A single connection setup
2136  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2137  * This many attempts are required for each one blind spoofed connection.  For
2138  * every additional spoofed connection he has to launch another N attempts.
2139  * Thus for a sustained rate 100 spoofed connections per second approximately
2140  * 1,800,000 packets per second would have to be sent.
2141  *
2142  * NB: The MAC function should be fast so that it doesn't become a CPU
2143  * exhaustion attack vector itself.
2144  *
2145  * References:
2146  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2147  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2148  *   http://cr.yp.to/syncookies.html    (overview)
2149  *   http://cr.yp.to/syncookies/archive (details)
2150  *
2151  *
2152  * Schematic construction of a syncookie enabled Initial Sequence Number:
2153  *  0        1         2         3
2154  *  12345678901234567890123456789012
2155  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2156  *
2157  *  x 24 MAC (truncated)
2158  *  W  3 Send Window Scale index
2159  *  M  3 MSS index
2160  *  S  1 SACK permitted
2161  *  P  1 Odd/even secret
2162  */
2163 
2164 /*
2165  * Distribution and probability of certain MSS values.  Those in between are
2166  * rounded down to the next lower one.
2167  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2168  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
2169  */
2170 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2171 
2172 /*
2173  * Distribution and probability of certain WSCALE values.  We have to map the
2174  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2175  * bits based on prevalence of certain values.  Where we don't have an exact
2176  * match for are rounded down to the next lower one letting us under-estimate
2177  * the true available window.  At the moment this would happen only for the
2178  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2179  * and window size).  The absence of the WSCALE option (no scaling in either
2180  * direction) is encoded with index zero.
2181  * [WSCALE values histograms, Allman, 2012]
2182  *                            X 10 10 35  5  6 14 10%   by host
2183  *                            X 11  4  5  5 18 49  3%   by connections
2184  */
2185 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2186 
2187 /*
2188  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
2189  * and good cryptographic properties.
2190  */
2191 static uint32_t
2192 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2193     uint8_t *secbits, uintptr_t secmod)
2194 {
2195 	SIPHASH_CTX ctx;
2196 	uint32_t siphash[2];
2197 
2198 	SipHash24_Init(&ctx);
2199 	SipHash_SetKey(&ctx, secbits);
2200 	switch (inc->inc_flags & INC_ISIPV6) {
2201 #ifdef INET
2202 	case 0:
2203 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2204 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2205 		break;
2206 #endif
2207 #ifdef INET6
2208 	case INC_ISIPV6:
2209 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2210 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2211 		break;
2212 #endif
2213 	}
2214 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2215 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2216 	SipHash_Update(&ctx, &irs, sizeof(irs));
2217 	SipHash_Update(&ctx, &flags, sizeof(flags));
2218 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
2219 	SipHash_Final((u_int8_t *)&siphash, &ctx);
2220 
2221 	return (siphash[0] ^ siphash[1]);
2222 }
2223 
2224 static tcp_seq
2225 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2226 {
2227 	u_int i, secbit, wscale;
2228 	uint32_t iss, hash;
2229 	uint8_t *secbits;
2230 	union syncookie cookie;
2231 
2232 	cookie.cookie = 0;
2233 
2234 	/* Map our computed MSS into the 3-bit index. */
2235 	for (i = nitems(tcp_sc_msstab) - 1;
2236 	     tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2237 	     i--)
2238 		;
2239 	cookie.flags.mss_idx = i;
2240 
2241 	/*
2242 	 * Map the send window scale into the 3-bit index but only if
2243 	 * the wscale option was received.
2244 	 */
2245 	if (sc->sc_flags & SCF_WINSCALE) {
2246 		wscale = sc->sc_requested_s_scale;
2247 		for (i = nitems(tcp_sc_wstab) - 1;
2248 		    tcp_sc_wstab[i] > wscale && i > 0;
2249 		     i--)
2250 			;
2251 		cookie.flags.wscale_idx = i;
2252 	}
2253 
2254 	/* Can we do SACK? */
2255 	if (sc->sc_flags & SCF_SACK)
2256 		cookie.flags.sack_ok = 1;
2257 
2258 	/* Which of the two secrets to use. */
2259 	secbit = V_tcp_syncache.secret.oddeven & 0x1;
2260 	cookie.flags.odd_even = secbit;
2261 
2262 	secbits = V_tcp_syncache.secret.key[secbit];
2263 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2264 	    (uintptr_t)sch);
2265 
2266 	/*
2267 	 * Put the flags into the hash and XOR them to get better ISS number
2268 	 * variance.  This doesn't enhance the cryptographic strength and is
2269 	 * done to prevent the 8 cookie bits from showing up directly on the
2270 	 * wire.
2271 	 */
2272 	iss = hash & ~0xff;
2273 	iss |= cookie.cookie ^ (hash >> 24);
2274 
2275 	TCPSTAT_INC(tcps_sc_sendcookie);
2276 	return (iss);
2277 }
2278 
2279 static struct syncache *
2280 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2281     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2282     struct socket *lso, uint16_t port)
2283 {
2284 	uint32_t hash;
2285 	uint8_t *secbits;
2286 	tcp_seq ack, seq;
2287 	int wnd, wscale = 0;
2288 	union syncookie cookie;
2289 
2290 	/*
2291 	 * Pull information out of SYN-ACK/ACK and revert sequence number
2292 	 * advances.
2293 	 */
2294 	ack = th->th_ack - 1;
2295 	seq = th->th_seq - 1;
2296 
2297 	/*
2298 	 * Unpack the flags containing enough information to restore the
2299 	 * connection.
2300 	 */
2301 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2302 
2303 	/* Which of the two secrets to use. */
2304 	secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2305 
2306 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2307 
2308 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
2309 	if ((ack & ~0xff) != (hash & ~0xff))
2310 		return (NULL);
2311 
2312 	/* Fill in the syncache values. */
2313 	sc->sc_flags = 0;
2314 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2315 	sc->sc_ipopts = NULL;
2316 
2317 	sc->sc_irs = seq;
2318 	sc->sc_iss = ack;
2319 
2320 	switch (inc->inc_flags & INC_ISIPV6) {
2321 #ifdef INET
2322 	case 0:
2323 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2324 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2325 		break;
2326 #endif
2327 #ifdef INET6
2328 	case INC_ISIPV6:
2329 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2330 			sc->sc_flowlabel =
2331 			    htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2332 		break;
2333 #endif
2334 	}
2335 
2336 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2337 
2338 	/* We can simply recompute receive window scale we sent earlier. */
2339 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2340 		wscale++;
2341 
2342 	/* Only use wscale if it was enabled in the orignal SYN. */
2343 	if (cookie.flags.wscale_idx > 0) {
2344 		sc->sc_requested_r_scale = wscale;
2345 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2346 		sc->sc_flags |= SCF_WINSCALE;
2347 	}
2348 
2349 	wnd = lso->sol_sbrcv_hiwat;
2350 	wnd = imax(wnd, 0);
2351 	wnd = imin(wnd, TCP_MAXWIN);
2352 	sc->sc_wnd = wnd;
2353 
2354 	if (cookie.flags.sack_ok)
2355 		sc->sc_flags |= SCF_SACK;
2356 
2357 	if (to->to_flags & TOF_TS) {
2358 		sc->sc_flags |= SCF_TIMESTAMP;
2359 		sc->sc_tsreflect = to->to_tsval;
2360 		sc->sc_tsoff = tcp_new_ts_offset(inc);
2361 	}
2362 
2363 	if (to->to_flags & TOF_SIGNATURE)
2364 		sc->sc_flags |= SCF_SIGNATURE;
2365 
2366 	sc->sc_rxmits = 0;
2367 
2368 	sc->sc_port = port;
2369 
2370 	TCPSTAT_INC(tcps_sc_recvcookie);
2371 	return (sc);
2372 }
2373 
2374 #ifdef INVARIANTS
2375 static int
2376 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2377     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2378     struct socket *lso, uint16_t port)
2379 {
2380 	struct syncache scs, *scx;
2381 	char *s;
2382 
2383 	bzero(&scs, sizeof(scs));
2384 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port);
2385 
2386 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2387 		return (0);
2388 
2389 	if (scx != NULL) {
2390 		if (sc->sc_peer_mss != scx->sc_peer_mss)
2391 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2392 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2393 
2394 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2395 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2396 			    s, __func__, sc->sc_requested_r_scale,
2397 			    scx->sc_requested_r_scale);
2398 
2399 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2400 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2401 			    s, __func__, sc->sc_requested_s_scale,
2402 			    scx->sc_requested_s_scale);
2403 
2404 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2405 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2406 	}
2407 
2408 	if (s != NULL)
2409 		free(s, M_TCPLOG);
2410 	return (0);
2411 }
2412 #endif /* INVARIANTS */
2413 
2414 static void
2415 syncookie_reseed(void *arg)
2416 {
2417 	struct tcp_syncache *sc = arg;
2418 	uint8_t *secbits;
2419 	int secbit;
2420 
2421 	/*
2422 	 * Reseeding the secret doesn't have to be protected by a lock.
2423 	 * It only must be ensured that the new random values are visible
2424 	 * to all CPUs in a SMP environment.  The atomic with release
2425 	 * semantics ensures that.
2426 	 */
2427 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2428 	secbits = sc->secret.key[secbit];
2429 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2430 	atomic_add_rel_int(&sc->secret.oddeven, 1);
2431 
2432 	/* Reschedule ourself. */
2433 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2434 }
2435 
2436 /*
2437  * We have overflowed a bucket. Let's pause dealing with the syncache.
2438  * This function will increment the bucketoverflow statistics appropriately
2439  * (once per pause when pausing is enabled; otherwise, once per overflow).
2440  */
2441 static void
2442 syncache_pause(struct in_conninfo *inc)
2443 {
2444 	time_t delta;
2445 	const char *s;
2446 
2447 	/* XXX:
2448 	 * 2. Add sysctl read here so we don't get the benefit of this
2449 	 * change without the new sysctl.
2450 	 */
2451 
2452 	/*
2453 	 * Try an unlocked read. If we already know that another thread
2454 	 * has activated the feature, there is no need to proceed.
2455 	 */
2456 	if (V_tcp_syncache.paused)
2457 		return;
2458 
2459 	/* Are cookied enabled? If not, we can't pause. */
2460 	if (!V_tcp_syncookies) {
2461 		TCPSTAT_INC(tcps_sc_bucketoverflow);
2462 		return;
2463 	}
2464 
2465 	/*
2466 	 * We may be the first thread to find an overflow. Get the lock
2467 	 * and evaluate if we need to take action.
2468 	 */
2469 	mtx_lock(&V_tcp_syncache.pause_mtx);
2470 	if (V_tcp_syncache.paused) {
2471 		mtx_unlock(&V_tcp_syncache.pause_mtx);
2472 		return;
2473 	}
2474 
2475 	/* Activate protection. */
2476 	V_tcp_syncache.paused = true;
2477 	TCPSTAT_INC(tcps_sc_bucketoverflow);
2478 
2479 	/*
2480 	 * Determine the last backoff time. If we are seeing a re-newed
2481 	 * attack within that same time after last reactivating the syncache,
2482 	 * consider it an extension of the same attack.
2483 	 */
2484 	delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2485 	if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2486 		if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2487 			delta <<= 1;
2488 			V_tcp_syncache.pause_backoff++;
2489 		}
2490 	} else {
2491 		delta = TCP_SYNCACHE_PAUSE_TIME;
2492 		V_tcp_syncache.pause_backoff = 0;
2493 	}
2494 
2495 	/* Log a warning, including IP addresses, if able. */
2496 	if (inc != NULL)
2497 		s = tcp_log_addrs(inc, NULL, NULL, NULL);
2498 	else
2499 		s = (const char *)NULL;
2500 	log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2501 	    "the next %lld seconds%s%s%s\n", (long long)delta,
2502 	    (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2503 	    (s != NULL) ? ")" : "");
2504 	free(__DECONST(void *, s), M_TCPLOG);
2505 
2506 	/* Use the calculated delta to set a new pause time. */
2507 	V_tcp_syncache.pause_until = time_uptime + delta;
2508 	callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2509 	    &V_tcp_syncache);
2510 	mtx_unlock(&V_tcp_syncache.pause_mtx);
2511 }
2512 
2513 /* Evaluate whether we need to unpause. */
2514 static void
2515 syncache_unpause(void *arg)
2516 {
2517 	struct tcp_syncache *sc;
2518 	time_t delta;
2519 
2520 	sc = arg;
2521 	mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2522 	callout_deactivate(&sc->pause_co);
2523 
2524 	/*
2525 	 * Check to make sure we are not running early. If the pause
2526 	 * time has expired, then deactivate the protection.
2527 	 */
2528 	if ((delta = sc->pause_until - time_uptime) > 0)
2529 		callout_schedule(&sc->pause_co, delta * hz);
2530 	else
2531 		sc->paused = false;
2532 }
2533 
2534 /*
2535  * Exports the syncache entries to userland so that netstat can display
2536  * them alongside the other sockets.  This function is intended to be
2537  * called only from tcp_pcblist.
2538  *
2539  * Due to concurrency on an active system, the number of pcbs exported
2540  * may have no relation to max_pcbs.  max_pcbs merely indicates the
2541  * amount of space the caller allocated for this function to use.
2542  */
2543 int
2544 syncache_pcblist(struct sysctl_req *req)
2545 {
2546 	struct xtcpcb xt;
2547 	struct syncache *sc;
2548 	struct syncache_head *sch;
2549 	int error, i;
2550 
2551 	bzero(&xt, sizeof(xt));
2552 	xt.xt_len = sizeof(xt);
2553 	xt.t_state = TCPS_SYN_RECEIVED;
2554 	xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2555 	xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2556 	xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2557 	xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2558 
2559 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2560 		sch = &V_tcp_syncache.hashbase[i];
2561 		SCH_LOCK(sch);
2562 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2563 			if (sc->sc_cred != NULL &&
2564 			    cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2565 				continue;
2566 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2567 				xt.xt_inp.inp_vflag = INP_IPV6;
2568 			else
2569 				xt.xt_inp.inp_vflag = INP_IPV4;
2570 			xt.xt_encaps_port = sc->sc_port;
2571 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2572 			    sizeof (struct in_conninfo));
2573 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2574 			if (error) {
2575 				SCH_UNLOCK(sch);
2576 				return (0);
2577 			}
2578 		}
2579 		SCH_UNLOCK(sch);
2580 	}
2581 
2582 	return (0);
2583 }
2584