xref: /freebsd/sys/netinet/tcp_syncache.c (revision 8d5719aa74f1d1441ee5ee365d45d53f934e81d6)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2001 McAfee, Inc.
5  * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG
6  * All rights reserved.
7  *
8  * This software was developed for the FreeBSD Project by Jonathan Lemon
9  * and McAfee Research, the Security Research Division of McAfee, Inc. under
10  * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
11  * DARPA CHATS research program. [2001 McAfee, Inc.]
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 #include "opt_inet.h"
39 #include "opt_inet6.h"
40 #include "opt_ipsec.h"
41 #include "opt_pcbgroup.h"
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/hash.h>
46 #include <sys/refcount.h>
47 #include <sys/kernel.h>
48 #include <sys/sysctl.h>
49 #include <sys/limits.h>
50 #include <sys/lock.h>
51 #include <sys/mutex.h>
52 #include <sys/malloc.h>
53 #include <sys/mbuf.h>
54 #include <sys/proc.h>		/* for proc0 declaration */
55 #include <sys/random.h>
56 #include <sys/socket.h>
57 #include <sys/socketvar.h>
58 #include <sys/syslog.h>
59 #include <sys/ucred.h>
60 
61 #include <sys/md5.h>
62 #include <crypto/siphash/siphash.h>
63 
64 #include <vm/uma.h>
65 
66 #include <net/if.h>
67 #include <net/if_var.h>
68 #include <net/route.h>
69 #include <net/vnet.h>
70 
71 #include <netinet/in.h>
72 #include <netinet/in_kdtrace.h>
73 #include <netinet/in_systm.h>
74 #include <netinet/ip.h>
75 #include <netinet/in_var.h>
76 #include <netinet/in_pcb.h>
77 #include <netinet/ip_var.h>
78 #include <netinet/ip_options.h>
79 #ifdef INET6
80 #include <netinet/ip6.h>
81 #include <netinet/icmp6.h>
82 #include <netinet6/nd6.h>
83 #include <netinet6/ip6_var.h>
84 #include <netinet6/in6_pcb.h>
85 #endif
86 #include <netinet/tcp.h>
87 #include <netinet/tcp_fastopen.h>
88 #include <netinet/tcp_fsm.h>
89 #include <netinet/tcp_seq.h>
90 #include <netinet/tcp_timer.h>
91 #include <netinet/tcp_var.h>
92 #include <netinet/tcp_syncache.h>
93 #ifdef INET6
94 #include <netinet6/tcp6_var.h>
95 #endif
96 #ifdef TCP_OFFLOAD
97 #include <netinet/toecore.h>
98 #endif
99 
100 #include <netipsec/ipsec_support.h>
101 
102 #include <machine/in_cksum.h>
103 
104 #include <security/mac/mac_framework.h>
105 
106 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1;
107 #define	V_tcp_syncookies		VNET(tcp_syncookies)
108 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
109     &VNET_NAME(tcp_syncookies), 0,
110     "Use TCP SYN cookies if the syncache overflows");
111 
112 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0;
113 #define	V_tcp_syncookiesonly		VNET(tcp_syncookiesonly)
114 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
115     &VNET_NAME(tcp_syncookiesonly), 0,
116     "Use only TCP SYN cookies");
117 
118 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1;
119 #define V_functions_inherit_listen_socket_stack \
120     VNET(functions_inherit_listen_socket_stack)
121 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack,
122     CTLFLAG_VNET | CTLFLAG_RW,
123     &VNET_NAME(functions_inherit_listen_socket_stack), 0,
124     "Inherit listen socket's stack");
125 
126 #ifdef TCP_OFFLOAD
127 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
128 #endif
129 
130 static void	 syncache_drop(struct syncache *, struct syncache_head *);
131 static void	 syncache_free(struct syncache *);
132 static void	 syncache_insert(struct syncache *, struct syncache_head *);
133 static int	 syncache_respond(struct syncache *, const struct mbuf *, int);
134 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
135 		    struct mbuf *m);
136 static void	 syncache_timeout(struct syncache *sc, struct syncache_head *sch,
137 		    int docallout);
138 static void	 syncache_timer(void *);
139 
140 static uint32_t	 syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
141 		    uint8_t *, uintptr_t);
142 static tcp_seq	 syncookie_generate(struct syncache_head *, struct syncache *);
143 static struct syncache
144 		*syncookie_lookup(struct in_conninfo *, struct syncache_head *,
145 		    struct syncache *, struct tcphdr *, struct tcpopt *,
146 		    struct socket *);
147 static void	syncache_pause(struct in_conninfo *);
148 static void	syncache_unpause(void *);
149 static void	 syncookie_reseed(void *);
150 #ifdef INVARIANTS
151 static int	 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
152 		    struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
153 		    struct socket *lso);
154 #endif
155 
156 /*
157  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
158  * 3 retransmits corresponds to a timeout with default values of
159  * tcp_rexmit_initial * (             1 +
160  *                       tcp_backoff[1] +
161  *                       tcp_backoff[2] +
162  *                       tcp_backoff[3]) + 3 * tcp_rexmit_slop,
163  * 1000 ms * (1 + 2 + 4 + 8) +  3 * 200 ms = 15600 ms,
164  * the odds are that the user has given up attempting to connect by then.
165  */
166 #define SYNCACHE_MAXREXMTS		3
167 
168 /* Arbitrary values */
169 #define TCP_SYNCACHE_HASHSIZE		512
170 #define TCP_SYNCACHE_BUCKETLIMIT	30
171 
172 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
173 #define	V_tcp_syncache			VNET(tcp_syncache)
174 
175 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache,
176     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
177     "TCP SYN cache");
178 
179 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
180     &VNET_NAME(tcp_syncache.bucket_limit), 0,
181     "Per-bucket hash limit for syncache");
182 
183 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
184     &VNET_NAME(tcp_syncache.cache_limit), 0,
185     "Overall entry limit for syncache");
186 
187 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
188     &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
189 
190 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
191     &VNET_NAME(tcp_syncache.hashsize), 0,
192     "Size of TCP syncache hashtable");
193 
194 static int
195 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
196 {
197 	int error;
198 	u_int new;
199 
200 	new = V_tcp_syncache.rexmt_limit;
201 	error = sysctl_handle_int(oidp, &new, 0, req);
202 	if ((error == 0) && (req->newptr != NULL)) {
203 		if (new > TCP_MAXRXTSHIFT)
204 			error = EINVAL;
205 		else
206 			V_tcp_syncache.rexmt_limit = new;
207 	}
208 	return (error);
209 }
210 
211 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
212     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
213     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
214     sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
215     "Limit on SYN/ACK retransmissions");
216 
217 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
218 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
219     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
220     "Send reset on socket allocation failure");
221 
222 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
223 
224 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
225 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
226 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
227 
228 /*
229  * Requires the syncache entry to be already removed from the bucket list.
230  */
231 static void
232 syncache_free(struct syncache *sc)
233 {
234 
235 	if (sc->sc_ipopts)
236 		(void) m_free(sc->sc_ipopts);
237 	if (sc->sc_cred)
238 		crfree(sc->sc_cred);
239 #ifdef MAC
240 	mac_syncache_destroy(&sc->sc_label);
241 #endif
242 
243 	uma_zfree(V_tcp_syncache.zone, sc);
244 }
245 
246 void
247 syncache_init(void)
248 {
249 	int i;
250 
251 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
252 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
253 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
254 	V_tcp_syncache.hash_secret = arc4random();
255 
256 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
257 	    &V_tcp_syncache.hashsize);
258 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
259 	    &V_tcp_syncache.bucket_limit);
260 	if (!powerof2(V_tcp_syncache.hashsize) ||
261 	    V_tcp_syncache.hashsize == 0) {
262 		printf("WARNING: syncache hash size is not a power of 2.\n");
263 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
264 	}
265 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
266 
267 	/* Set limits. */
268 	V_tcp_syncache.cache_limit =
269 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
270 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
271 	    &V_tcp_syncache.cache_limit);
272 
273 	/* Allocate the hash table. */
274 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
275 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
276 
277 #ifdef VIMAGE
278 	V_tcp_syncache.vnet = curvnet;
279 #endif
280 
281 	/* Initialize the hash buckets. */
282 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
283 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
284 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
285 			 NULL, MTX_DEF);
286 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
287 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
288 		V_tcp_syncache.hashbase[i].sch_length = 0;
289 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
290 		V_tcp_syncache.hashbase[i].sch_last_overflow =
291 		    -(SYNCOOKIE_LIFETIME + 1);
292 	}
293 
294 	/* Create the syncache entry zone. */
295 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
296 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
297 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
298 	    V_tcp_syncache.cache_limit);
299 
300 	/* Start the SYN cookie reseeder callout. */
301 	callout_init(&V_tcp_syncache.secret.reseed, 1);
302 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
303 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
304 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
305 	    syncookie_reseed, &V_tcp_syncache);
306 
307 	/* Initialize the pause machinery. */
308 	mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF);
309 	callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx,
310 	    0);
311 	V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME;
312 	V_tcp_syncache.pause_backoff = 0;
313 	V_tcp_syncache.paused = false;
314 }
315 
316 #ifdef VIMAGE
317 void
318 syncache_destroy(void)
319 {
320 	struct syncache_head *sch;
321 	struct syncache *sc, *nsc;
322 	int i;
323 
324 	/*
325 	 * Stop the re-seed timer before freeing resources.  No need to
326 	 * possibly schedule it another time.
327 	 */
328 	callout_drain(&V_tcp_syncache.secret.reseed);
329 
330 	/* Stop the SYN cache pause callout. */
331 	mtx_lock(&V_tcp_syncache.pause_mtx);
332 	if (callout_stop(&V_tcp_syncache.pause_co) == 0) {
333 		mtx_unlock(&V_tcp_syncache.pause_mtx);
334 		callout_drain(&V_tcp_syncache.pause_co);
335 	} else
336 		mtx_unlock(&V_tcp_syncache.pause_mtx);
337 
338 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
339 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
340 		sch = &V_tcp_syncache.hashbase[i];
341 		callout_drain(&sch->sch_timer);
342 
343 		SCH_LOCK(sch);
344 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
345 			syncache_drop(sc, sch);
346 		SCH_UNLOCK(sch);
347 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
348 		    ("%s: sch->sch_bucket not empty", __func__));
349 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
350 		    __func__, sch->sch_length));
351 		mtx_destroy(&sch->sch_mtx);
352 	}
353 
354 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
355 	    ("%s: cache_count not 0", __func__));
356 
357 	/* Free the allocated global resources. */
358 	uma_zdestroy(V_tcp_syncache.zone);
359 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
360 	mtx_destroy(&V_tcp_syncache.pause_mtx);
361 }
362 #endif
363 
364 /*
365  * Inserts a syncache entry into the specified bucket row.
366  * Locks and unlocks the syncache_head autonomously.
367  */
368 static void
369 syncache_insert(struct syncache *sc, struct syncache_head *sch)
370 {
371 	struct syncache *sc2;
372 
373 	SCH_LOCK(sch);
374 
375 	/*
376 	 * Make sure that we don't overflow the per-bucket limit.
377 	 * If the bucket is full, toss the oldest element.
378 	 */
379 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
380 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
381 			("sch->sch_length incorrect"));
382 		syncache_pause(&sc->sc_inc);
383 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
384 		sch->sch_last_overflow = time_uptime;
385 		syncache_drop(sc2, sch);
386 	}
387 
388 	/* Put it into the bucket. */
389 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
390 	sch->sch_length++;
391 
392 #ifdef TCP_OFFLOAD
393 	if (ADDED_BY_TOE(sc)) {
394 		struct toedev *tod = sc->sc_tod;
395 
396 		tod->tod_syncache_added(tod, sc->sc_todctx);
397 	}
398 #endif
399 
400 	/* Reinitialize the bucket row's timer. */
401 	if (sch->sch_length == 1)
402 		sch->sch_nextc = ticks + INT_MAX;
403 	syncache_timeout(sc, sch, 1);
404 
405 	SCH_UNLOCK(sch);
406 
407 	TCPSTATES_INC(TCPS_SYN_RECEIVED);
408 	TCPSTAT_INC(tcps_sc_added);
409 }
410 
411 /*
412  * Remove and free entry from syncache bucket row.
413  * Expects locked syncache head.
414  */
415 static void
416 syncache_drop(struct syncache *sc, struct syncache_head *sch)
417 {
418 
419 	SCH_LOCK_ASSERT(sch);
420 
421 	TCPSTATES_DEC(TCPS_SYN_RECEIVED);
422 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
423 	sch->sch_length--;
424 
425 #ifdef TCP_OFFLOAD
426 	if (ADDED_BY_TOE(sc)) {
427 		struct toedev *tod = sc->sc_tod;
428 
429 		tod->tod_syncache_removed(tod, sc->sc_todctx);
430 	}
431 #endif
432 
433 	syncache_free(sc);
434 }
435 
436 /*
437  * Engage/reengage time on bucket row.
438  */
439 static void
440 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
441 {
442 	int rexmt;
443 
444 	if (sc->sc_rxmits == 0)
445 		rexmt = tcp_rexmit_initial;
446 	else
447 		TCPT_RANGESET(rexmt,
448 		    tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
449 		    tcp_rexmit_min, TCPTV_REXMTMAX);
450 	sc->sc_rxttime = ticks + rexmt;
451 	sc->sc_rxmits++;
452 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
453 		sch->sch_nextc = sc->sc_rxttime;
454 		if (docallout)
455 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
456 			    syncache_timer, (void *)sch);
457 	}
458 }
459 
460 /*
461  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
462  * If we have retransmitted an entry the maximum number of times, expire it.
463  * One separate timer for each bucket row.
464  */
465 static void
466 syncache_timer(void *xsch)
467 {
468 	struct syncache_head *sch = (struct syncache_head *)xsch;
469 	struct syncache *sc, *nsc;
470 	struct epoch_tracker et;
471 	int tick = ticks;
472 	char *s;
473 	bool paused;
474 
475 	CURVNET_SET(sch->sch_sc->vnet);
476 
477 	/* NB: syncache_head has already been locked by the callout. */
478 	SCH_LOCK_ASSERT(sch);
479 
480 	/*
481 	 * In the following cycle we may remove some entries and/or
482 	 * advance some timeouts, so re-initialize the bucket timer.
483 	 */
484 	sch->sch_nextc = tick + INT_MAX;
485 
486 	/*
487 	 * If we have paused processing, unconditionally remove
488 	 * all syncache entries.
489 	 */
490 	mtx_lock(&V_tcp_syncache.pause_mtx);
491 	paused = V_tcp_syncache.paused;
492 	mtx_unlock(&V_tcp_syncache.pause_mtx);
493 
494 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
495 		if (paused) {
496 			syncache_drop(sc, sch);
497 			continue;
498 		}
499 		/*
500 		 * We do not check if the listen socket still exists
501 		 * and accept the case where the listen socket may be
502 		 * gone by the time we resend the SYN/ACK.  We do
503 		 * not expect this to happens often. If it does,
504 		 * then the RST will be sent by the time the remote
505 		 * host does the SYN/ACK->ACK.
506 		 */
507 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
508 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
509 				sch->sch_nextc = sc->sc_rxttime;
510 			continue;
511 		}
512 		if (sc->sc_rxmits > V_tcp_ecn_maxretries) {
513 			sc->sc_flags &= ~SCF_ECN;
514 		}
515 		if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
516 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
517 				log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
518 				    "giving up and removing syncache entry\n",
519 				    s, __func__);
520 				free(s, M_TCPLOG);
521 			}
522 			syncache_drop(sc, sch);
523 			TCPSTAT_INC(tcps_sc_stale);
524 			continue;
525 		}
526 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
527 			log(LOG_DEBUG, "%s; %s: Response timeout, "
528 			    "retransmitting (%u) SYN|ACK\n",
529 			    s, __func__, sc->sc_rxmits);
530 			free(s, M_TCPLOG);
531 		}
532 
533 		NET_EPOCH_ENTER(et);
534 		syncache_respond(sc, NULL, TH_SYN|TH_ACK);
535 		NET_EPOCH_EXIT(et);
536 		TCPSTAT_INC(tcps_sc_retransmitted);
537 		syncache_timeout(sc, sch, 0);
538 	}
539 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
540 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
541 			syncache_timer, (void *)(sch));
542 	CURVNET_RESTORE();
543 }
544 
545 /*
546  * Returns true if the system is only using cookies at the moment.
547  * This could be due to a sysadmin decision to only use cookies, or it
548  * could be due to the system detecting an attack.
549  */
550 static inline bool
551 syncache_cookiesonly(void)
552 {
553 
554 	return (V_tcp_syncookies && (V_tcp_syncache.paused ||
555 	    V_tcp_syncookiesonly));
556 }
557 
558 /*
559  * Find the hash bucket for the given connection.
560  */
561 static struct syncache_head *
562 syncache_hashbucket(struct in_conninfo *inc)
563 {
564 	uint32_t hash;
565 
566 	/*
567 	 * The hash is built on foreign port + local port + foreign address.
568 	 * We rely on the fact that struct in_conninfo starts with 16 bits
569 	 * of foreign port, then 16 bits of local port then followed by 128
570 	 * bits of foreign address.  In case of IPv4 address, the first 3
571 	 * 32-bit words of the address always are zeroes.
572 	 */
573 	hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
574 	    V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
575 
576 	return (&V_tcp_syncache.hashbase[hash]);
577 }
578 
579 /*
580  * Find an entry in the syncache.
581  * Returns always with locked syncache_head plus a matching entry or NULL.
582  */
583 static struct syncache *
584 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
585 {
586 	struct syncache *sc;
587 	struct syncache_head *sch;
588 
589 	*schp = sch = syncache_hashbucket(inc);
590 	SCH_LOCK(sch);
591 
592 	/* Circle through bucket row to find matching entry. */
593 	TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
594 		if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
595 		    sizeof(struct in_endpoints)) == 0)
596 			break;
597 
598 	return (sc);	/* Always returns with locked sch. */
599 }
600 
601 /*
602  * This function is called when we get a RST for a
603  * non-existent connection, so that we can see if the
604  * connection is in the syn cache.  If it is, zap it.
605  * If required send a challenge ACK.
606  */
607 void
608 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m)
609 {
610 	struct syncache *sc;
611 	struct syncache_head *sch;
612 	char *s = NULL;
613 
614 	if (syncache_cookiesonly())
615 		return;
616 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
617 	SCH_LOCK_ASSERT(sch);
618 
619 	/*
620 	 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
621 	 * See RFC 793 page 65, section SEGMENT ARRIVES.
622 	 */
623 	if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
624 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
625 			log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
626 			    "FIN flag set, segment ignored\n", s, __func__);
627 		TCPSTAT_INC(tcps_badrst);
628 		goto done;
629 	}
630 
631 	/*
632 	 * No corresponding connection was found in syncache.
633 	 * If syncookies are enabled and possibly exclusively
634 	 * used, or we are under memory pressure, a valid RST
635 	 * may not find a syncache entry.  In that case we're
636 	 * done and no SYN|ACK retransmissions will happen.
637 	 * Otherwise the RST was misdirected or spoofed.
638 	 */
639 	if (sc == NULL) {
640 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
641 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
642 			    "syncache entry (possibly syncookie only), "
643 			    "segment ignored\n", s, __func__);
644 		TCPSTAT_INC(tcps_badrst);
645 		goto done;
646 	}
647 
648 	/*
649 	 * If the RST bit is set, check the sequence number to see
650 	 * if this is a valid reset segment.
651 	 *
652 	 * RFC 793 page 37:
653 	 *   In all states except SYN-SENT, all reset (RST) segments
654 	 *   are validated by checking their SEQ-fields.  A reset is
655 	 *   valid if its sequence number is in the window.
656 	 *
657 	 * RFC 793 page 69:
658 	 *   There are four cases for the acceptability test for an incoming
659 	 *   segment:
660 	 *
661 	 * Segment Receive  Test
662 	 * Length  Window
663 	 * ------- -------  -------------------------------------------
664 	 *    0       0     SEG.SEQ = RCV.NXT
665 	 *    0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
666 	 *   >0       0     not acceptable
667 	 *   >0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
668 	 *               or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
669 	 *
670 	 * Note that when receiving a SYN segment in the LISTEN state,
671 	 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
672 	 * described in RFC 793, page 66.
673 	 */
674 	if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
675 	    SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
676 	    (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
677 		if (V_tcp_insecure_rst ||
678 		    th->th_seq == sc->sc_irs + 1) {
679 			syncache_drop(sc, sch);
680 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
681 				log(LOG_DEBUG,
682 				    "%s; %s: Our SYN|ACK was rejected, "
683 				    "connection attempt aborted by remote "
684 				    "endpoint\n",
685 				    s, __func__);
686 			TCPSTAT_INC(tcps_sc_reset);
687 		} else {
688 			TCPSTAT_INC(tcps_badrst);
689 			/* Send challenge ACK. */
690 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
691 				log(LOG_DEBUG, "%s; %s: RST with invalid "
692 				    " SEQ %u != NXT %u (+WND %u), "
693 				    "sending challenge ACK\n",
694 				    s, __func__,
695 				    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
696 			syncache_respond(sc, m, TH_ACK);
697 		}
698 	} else {
699 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
700 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
701 			    "NXT %u (+WND %u), segment ignored\n",
702 			    s, __func__,
703 			    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
704 		TCPSTAT_INC(tcps_badrst);
705 	}
706 
707 done:
708 	if (s != NULL)
709 		free(s, M_TCPLOG);
710 	SCH_UNLOCK(sch);
711 }
712 
713 void
714 syncache_badack(struct in_conninfo *inc)
715 {
716 	struct syncache *sc;
717 	struct syncache_head *sch;
718 
719 	if (syncache_cookiesonly())
720 		return;
721 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
722 	SCH_LOCK_ASSERT(sch);
723 	if (sc != NULL) {
724 		syncache_drop(sc, sch);
725 		TCPSTAT_INC(tcps_sc_badack);
726 	}
727 	SCH_UNLOCK(sch);
728 }
729 
730 void
731 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq)
732 {
733 	struct syncache *sc;
734 	struct syncache_head *sch;
735 
736 	if (syncache_cookiesonly())
737 		return;
738 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
739 	SCH_LOCK_ASSERT(sch);
740 	if (sc == NULL)
741 		goto done;
742 
743 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
744 	if (ntohl(th_seq) != sc->sc_iss)
745 		goto done;
746 
747 	/*
748 	 * If we've rertransmitted 3 times and this is our second error,
749 	 * we remove the entry.  Otherwise, we allow it to continue on.
750 	 * This prevents us from incorrectly nuking an entry during a
751 	 * spurious network outage.
752 	 *
753 	 * See tcp_notify().
754 	 */
755 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
756 		sc->sc_flags |= SCF_UNREACH;
757 		goto done;
758 	}
759 	syncache_drop(sc, sch);
760 	TCPSTAT_INC(tcps_sc_unreach);
761 done:
762 	SCH_UNLOCK(sch);
763 }
764 
765 /*
766  * Build a new TCP socket structure from a syncache entry.
767  *
768  * On success return the newly created socket with its underlying inp locked.
769  */
770 static struct socket *
771 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
772 {
773 	struct tcp_function_block *blk;
774 	struct inpcb *inp = NULL;
775 	struct socket *so;
776 	struct tcpcb *tp;
777 	int error;
778 	char *s;
779 
780 	NET_EPOCH_ASSERT();
781 
782 	/*
783 	 * Ok, create the full blown connection, and set things up
784 	 * as they would have been set up if we had created the
785 	 * connection when the SYN arrived.  If we can't create
786 	 * the connection, abort it.
787 	 */
788 	so = sonewconn(lso, 0);
789 	if (so == NULL) {
790 		/*
791 		 * Drop the connection; we will either send a RST or
792 		 * have the peer retransmit its SYN again after its
793 		 * RTO and try again.
794 		 */
795 		TCPSTAT_INC(tcps_listendrop);
796 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
797 			log(LOG_DEBUG, "%s; %s: Socket create failed "
798 			    "due to limits or memory shortage\n",
799 			    s, __func__);
800 			free(s, M_TCPLOG);
801 		}
802 		goto abort2;
803 	}
804 #ifdef MAC
805 	mac_socketpeer_set_from_mbuf(m, so);
806 #endif
807 
808 	inp = sotoinpcb(so);
809 	inp->inp_inc.inc_fibnum = so->so_fibnum;
810 	INP_WLOCK(inp);
811 	/*
812 	 * Exclusive pcbinfo lock is not required in syncache socket case even
813 	 * if two inpcb locks can be acquired simultaneously:
814 	 *  - the inpcb in LISTEN state,
815 	 *  - the newly created inp.
816 	 *
817 	 * In this case, an inp cannot be at same time in LISTEN state and
818 	 * just created by an accept() call.
819 	 */
820 	INP_HASH_WLOCK(&V_tcbinfo);
821 
822 	/* Insert new socket into PCB hash list. */
823 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
824 #ifdef INET6
825 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
826 		inp->inp_vflag &= ~INP_IPV4;
827 		inp->inp_vflag |= INP_IPV6;
828 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
829 	} else {
830 		inp->inp_vflag &= ~INP_IPV6;
831 		inp->inp_vflag |= INP_IPV4;
832 #endif
833 		inp->inp_ip_ttl = sc->sc_ip_ttl;
834 		inp->inp_ip_tos = sc->sc_ip_tos;
835 		inp->inp_laddr = sc->sc_inc.inc_laddr;
836 #ifdef INET6
837 	}
838 #endif
839 
840 	/*
841 	 * If there's an mbuf and it has a flowid, then let's initialise the
842 	 * inp with that particular flowid.
843 	 */
844 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
845 		inp->inp_flowid = m->m_pkthdr.flowid;
846 		inp->inp_flowtype = M_HASHTYPE_GET(m);
847 #ifdef NUMA
848 		inp->inp_numa_domain = m->m_pkthdr.numa_domain;
849 #endif
850 	}
851 
852 	inp->inp_lport = sc->sc_inc.inc_lport;
853 #ifdef INET6
854 	if (inp->inp_vflag & INP_IPV6PROTO) {
855 		struct inpcb *oinp = sotoinpcb(lso);
856 
857 		/*
858 		 * Inherit socket options from the listening socket.
859 		 * Note that in6p_inputopts are not (and should not be)
860 		 * copied, since it stores previously received options and is
861 		 * used to detect if each new option is different than the
862 		 * previous one and hence should be passed to a user.
863 		 * If we copied in6p_inputopts, a user would not be able to
864 		 * receive options just after calling the accept system call.
865 		 */
866 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
867 		if (oinp->in6p_outputopts)
868 			inp->in6p_outputopts =
869 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
870 		inp->in6p_hops = oinp->in6p_hops;
871 	}
872 
873 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
874 		struct in6_addr laddr6;
875 		struct sockaddr_in6 sin6;
876 
877 		sin6.sin6_family = AF_INET6;
878 		sin6.sin6_len = sizeof(sin6);
879 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
880 		sin6.sin6_port = sc->sc_inc.inc_fport;
881 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
882 		laddr6 = inp->in6p_laddr;
883 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
884 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
885 		if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
886 		    thread0.td_ucred, m, false)) != 0) {
887 			inp->in6p_laddr = laddr6;
888 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
889 				log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
890 				    "with error %i\n",
891 				    s, __func__, error);
892 				free(s, M_TCPLOG);
893 			}
894 			INP_HASH_WUNLOCK(&V_tcbinfo);
895 			goto abort;
896 		}
897 		/* Override flowlabel from in6_pcbconnect. */
898 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
899 		inp->inp_flow |= sc->sc_flowlabel;
900 	}
901 #endif /* INET6 */
902 #if defined(INET) && defined(INET6)
903 	else
904 #endif
905 #ifdef INET
906 	{
907 		struct in_addr laddr;
908 		struct sockaddr_in sin;
909 
910 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
911 
912 		if (inp->inp_options == NULL) {
913 			inp->inp_options = sc->sc_ipopts;
914 			sc->sc_ipopts = NULL;
915 		}
916 
917 		sin.sin_family = AF_INET;
918 		sin.sin_len = sizeof(sin);
919 		sin.sin_addr = sc->sc_inc.inc_faddr;
920 		sin.sin_port = sc->sc_inc.inc_fport;
921 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
922 		laddr = inp->inp_laddr;
923 		if (inp->inp_laddr.s_addr == INADDR_ANY)
924 			inp->inp_laddr = sc->sc_inc.inc_laddr;
925 		if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
926 		    thread0.td_ucred, m, false)) != 0) {
927 			inp->inp_laddr = laddr;
928 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
929 				log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
930 				    "with error %i\n",
931 				    s, __func__, error);
932 				free(s, M_TCPLOG);
933 			}
934 			INP_HASH_WUNLOCK(&V_tcbinfo);
935 			goto abort;
936 		}
937 	}
938 #endif /* INET */
939 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
940 	/* Copy old policy into new socket's. */
941 	if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
942 		printf("syncache_socket: could not copy policy\n");
943 #endif
944 	INP_HASH_WUNLOCK(&V_tcbinfo);
945 	tp = intotcpcb(inp);
946 	tcp_state_change(tp, TCPS_SYN_RECEIVED);
947 	tp->iss = sc->sc_iss;
948 	tp->irs = sc->sc_irs;
949 	tcp_rcvseqinit(tp);
950 	tcp_sendseqinit(tp);
951 	blk = sototcpcb(lso)->t_fb;
952 	if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
953 		/*
954 		 * Our parents t_fb was not the default,
955 		 * we need to release our ref on tp->t_fb and
956 		 * pickup one on the new entry.
957 		 */
958 		struct tcp_function_block *rblk;
959 
960 		rblk = find_and_ref_tcp_fb(blk);
961 		KASSERT(rblk != NULL,
962 		    ("cannot find blk %p out of syncache?", blk));
963 		if (tp->t_fb->tfb_tcp_fb_fini)
964 			(*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
965 		refcount_release(&tp->t_fb->tfb_refcnt);
966 		tp->t_fb = rblk;
967 		/*
968 		 * XXXrrs this is quite dangerous, it is possible
969 		 * for the new function to fail to init. We also
970 		 * are not asking if the handoff_is_ok though at
971 		 * the very start thats probalbly ok.
972 		 */
973 		if (tp->t_fb->tfb_tcp_fb_init) {
974 			(*tp->t_fb->tfb_tcp_fb_init)(tp);
975 		}
976 	}
977 	tp->snd_wl1 = sc->sc_irs;
978 	tp->snd_max = tp->iss + 1;
979 	tp->snd_nxt = tp->iss + 1;
980 	tp->rcv_up = sc->sc_irs + 1;
981 	tp->rcv_wnd = sc->sc_wnd;
982 	tp->rcv_adv += tp->rcv_wnd;
983 	tp->last_ack_sent = tp->rcv_nxt;
984 
985 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
986 	if (sc->sc_flags & SCF_NOOPT)
987 		tp->t_flags |= TF_NOOPT;
988 	else {
989 		if (sc->sc_flags & SCF_WINSCALE) {
990 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
991 			tp->snd_scale = sc->sc_requested_s_scale;
992 			tp->request_r_scale = sc->sc_requested_r_scale;
993 		}
994 		if (sc->sc_flags & SCF_TIMESTAMP) {
995 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
996 			tp->ts_recent = sc->sc_tsreflect;
997 			tp->ts_recent_age = tcp_ts_getticks();
998 			tp->ts_offset = sc->sc_tsoff;
999 		}
1000 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1001 		if (sc->sc_flags & SCF_SIGNATURE)
1002 			tp->t_flags |= TF_SIGNATURE;
1003 #endif
1004 		if (sc->sc_flags & SCF_SACK)
1005 			tp->t_flags |= TF_SACK_PERMIT;
1006 	}
1007 
1008 	if (sc->sc_flags & SCF_ECN)
1009 		tp->t_flags2 |= TF2_ECN_PERMIT;
1010 
1011 	/*
1012 	 * Set up MSS and get cached values from tcp_hostcache.
1013 	 * This might overwrite some of the defaults we just set.
1014 	 */
1015 	tcp_mss(tp, sc->sc_peer_mss);
1016 
1017 	/*
1018 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
1019 	 * limited to one segment in cc_conn_init().
1020 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
1021 	 */
1022 	if (sc->sc_rxmits > 1)
1023 		tp->snd_cwnd = 1;
1024 
1025 #ifdef TCP_OFFLOAD
1026 	/*
1027 	 * Allow a TOE driver to install its hooks.  Note that we hold the
1028 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
1029 	 * new connection before the TOE driver has done its thing.
1030 	 */
1031 	if (ADDED_BY_TOE(sc)) {
1032 		struct toedev *tod = sc->sc_tod;
1033 
1034 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
1035 	}
1036 #endif
1037 	/*
1038 	 * Copy and activate timers.
1039 	 */
1040 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
1041 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
1042 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
1043 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1044 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1045 
1046 	TCPSTAT_INC(tcps_accepts);
1047 	return (so);
1048 
1049 abort:
1050 	INP_WUNLOCK(inp);
1051 abort2:
1052 	if (so != NULL)
1053 		soabort(so);
1054 	return (NULL);
1055 }
1056 
1057 /*
1058  * This function gets called when we receive an ACK for a
1059  * socket in the LISTEN state.  We look up the connection
1060  * in the syncache, and if its there, we pull it out of
1061  * the cache and turn it into a full-blown connection in
1062  * the SYN-RECEIVED state.
1063  *
1064  * On syncache_socket() success the newly created socket
1065  * has its underlying inp locked.
1066  */
1067 int
1068 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1069     struct socket **lsop, struct mbuf *m)
1070 {
1071 	struct syncache *sc;
1072 	struct syncache_head *sch;
1073 	struct syncache scs;
1074 	char *s;
1075 	bool locked;
1076 
1077 	NET_EPOCH_ASSERT();
1078 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1079 	    ("%s: can handle only ACK", __func__));
1080 
1081 	if (syncache_cookiesonly()) {
1082 		sc = NULL;
1083 		sch = syncache_hashbucket(inc);
1084 		locked = false;
1085 	} else {
1086 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1087 		locked = true;
1088 		SCH_LOCK_ASSERT(sch);
1089 	}
1090 
1091 #ifdef INVARIANTS
1092 	/*
1093 	 * Test code for syncookies comparing the syncache stored
1094 	 * values with the reconstructed values from the cookie.
1095 	 */
1096 	if (sc != NULL)
1097 		syncookie_cmp(inc, sch, sc, th, to, *lsop);
1098 #endif
1099 
1100 	if (sc == NULL) {
1101 		/*
1102 		 * There is no syncache entry, so see if this ACK is
1103 		 * a returning syncookie.  To do this, first:
1104 		 *  A. Check if syncookies are used in case of syncache
1105 		 *     overflows
1106 		 *  B. See if this socket has had a syncache entry dropped in
1107 		 *     the recent past. We don't want to accept a bogus
1108 		 *     syncookie if we've never received a SYN or accept it
1109 		 *     twice.
1110 		 *  C. check that the syncookie is valid.  If it is, then
1111 		 *     cobble up a fake syncache entry, and return.
1112 		 */
1113 		if (locked && !V_tcp_syncookies) {
1114 			SCH_UNLOCK(sch);
1115 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1116 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1117 				    "segment rejected (syncookies disabled)\n",
1118 				    s, __func__);
1119 			goto failed;
1120 		}
1121 		if (locked && !V_tcp_syncookiesonly &&
1122 		    sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1123 			SCH_UNLOCK(sch);
1124 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1125 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1126 				    "segment rejected (no syncache entry)\n",
1127 				    s, __func__);
1128 			goto failed;
1129 		}
1130 		bzero(&scs, sizeof(scs));
1131 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
1132 		if (locked)
1133 			SCH_UNLOCK(sch);
1134 		if (sc == NULL) {
1135 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1136 				log(LOG_DEBUG, "%s; %s: Segment failed "
1137 				    "SYNCOOKIE authentication, segment rejected "
1138 				    "(probably spoofed)\n", s, __func__);
1139 			goto failed;
1140 		}
1141 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1142 		/* If received ACK has MD5 signature, check it. */
1143 		if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1144 		    (!TCPMD5_ENABLED() ||
1145 		    TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1146 			/* Drop the ACK. */
1147 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1148 				log(LOG_DEBUG, "%s; %s: Segment rejected, "
1149 				    "MD5 signature doesn't match.\n",
1150 				    s, __func__);
1151 				free(s, M_TCPLOG);
1152 			}
1153 			TCPSTAT_INC(tcps_sig_err_sigopt);
1154 			return (-1); /* Do not send RST */
1155 		}
1156 #endif /* TCP_SIGNATURE */
1157 	} else {
1158 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1159 		/*
1160 		 * If listening socket requested TCP digests, check that
1161 		 * received ACK has signature and it is correct.
1162 		 * If not, drop the ACK and leave sc entry in th cache,
1163 		 * because SYN was received with correct signature.
1164 		 */
1165 		if (sc->sc_flags & SCF_SIGNATURE) {
1166 			if ((to->to_flags & TOF_SIGNATURE) == 0) {
1167 				/* No signature */
1168 				TCPSTAT_INC(tcps_sig_err_nosigopt);
1169 				SCH_UNLOCK(sch);
1170 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1171 					log(LOG_DEBUG, "%s; %s: Segment "
1172 					    "rejected, MD5 signature wasn't "
1173 					    "provided.\n", s, __func__);
1174 					free(s, M_TCPLOG);
1175 				}
1176 				return (-1); /* Do not send RST */
1177 			}
1178 			if (!TCPMD5_ENABLED() ||
1179 			    TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1180 				/* Doesn't match or no SA */
1181 				SCH_UNLOCK(sch);
1182 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1183 					log(LOG_DEBUG, "%s; %s: Segment "
1184 					    "rejected, MD5 signature doesn't "
1185 					    "match.\n", s, __func__);
1186 					free(s, M_TCPLOG);
1187 				}
1188 				return (-1); /* Do not send RST */
1189 			}
1190 		}
1191 #endif /* TCP_SIGNATURE */
1192 
1193 		/*
1194 		 * RFC 7323 PAWS: If we have a timestamp on this segment and
1195 		 * it's less than ts_recent, drop it.
1196 		 * XXXMT: RFC 7323 also requires to send an ACK.
1197 		 *        In tcp_input.c this is only done for TCP segments
1198 		 *        with user data, so be consistent here and just drop
1199 		 *        the segment.
1200 		 */
1201 		if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1202 		    TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1203 			SCH_UNLOCK(sch);
1204 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1205 				log(LOG_DEBUG,
1206 				    "%s; %s: SEG.TSval %u < TS.Recent %u, "
1207 				    "segment dropped\n", s, __func__,
1208 				    to->to_tsval, sc->sc_tsreflect);
1209 				free(s, M_TCPLOG);
1210 			}
1211 			return (-1);  /* Do not send RST */
1212 		}
1213 
1214 		/*
1215 		 * If timestamps were not negotiated during SYN/ACK and a
1216 		 * segment with a timestamp is received, ignore the
1217 		 * timestamp and process the packet normally.
1218 		 * See section 3.2 of RFC 7323.
1219 		 */
1220 		if (!(sc->sc_flags & SCF_TIMESTAMP) &&
1221 		    (to->to_flags & TOF_TS)) {
1222 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1223 				log(LOG_DEBUG, "%s; %s: Timestamp not "
1224 				    "expected, segment processed normally\n",
1225 				    s, __func__);
1226 				free(s, M_TCPLOG);
1227 				s = NULL;
1228 			}
1229 		}
1230 
1231 		/*
1232 		 * If timestamps were negotiated during SYN/ACK and a
1233 		 * segment without a timestamp is received, silently drop
1234 		 * the segment, unless the missing timestamps are tolerated.
1235 		 * See section 3.2 of RFC 7323.
1236 		 */
1237 		if ((sc->sc_flags & SCF_TIMESTAMP) &&
1238 		    !(to->to_flags & TOF_TS)) {
1239 			if (V_tcp_tolerate_missing_ts) {
1240 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1241 					log(LOG_DEBUG,
1242 					    "%s; %s: Timestamp missing, "
1243 					    "segment processed normally\n",
1244 					    s, __func__);
1245 					free(s, M_TCPLOG);
1246 				}
1247 			} else {
1248 				SCH_UNLOCK(sch);
1249 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1250 					log(LOG_DEBUG,
1251 					    "%s; %s: Timestamp missing, "
1252 					    "segment silently dropped\n",
1253 					    s, __func__);
1254 					free(s, M_TCPLOG);
1255 				}
1256 				return (-1);  /* Do not send RST */
1257 			}
1258 		}
1259 
1260 		/*
1261 		 * Pull out the entry to unlock the bucket row.
1262 		 *
1263 		 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1264 		 * tcp_state_change().  The tcpcb is not existent at this
1265 		 * moment.  A new one will be allocated via syncache_socket->
1266 		 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1267 		 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1268 		 */
1269 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1270 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1271 		sch->sch_length--;
1272 #ifdef TCP_OFFLOAD
1273 		if (ADDED_BY_TOE(sc)) {
1274 			struct toedev *tod = sc->sc_tod;
1275 
1276 			tod->tod_syncache_removed(tod, sc->sc_todctx);
1277 		}
1278 #endif
1279 		SCH_UNLOCK(sch);
1280 	}
1281 
1282 	/*
1283 	 * Segment validation:
1284 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1285 	 */
1286 	if (th->th_ack != sc->sc_iss + 1) {
1287 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1288 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1289 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1290 		goto failed;
1291 	}
1292 
1293 	/*
1294 	 * The SEQ must fall in the window starting at the received
1295 	 * initial receive sequence number + 1 (the SYN).
1296 	 */
1297 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1298 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1299 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1300 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1301 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1302 		goto failed;
1303 	}
1304 
1305 	*lsop = syncache_socket(sc, *lsop, m);
1306 
1307 	if (*lsop == NULL)
1308 		TCPSTAT_INC(tcps_sc_aborted);
1309 	else
1310 		TCPSTAT_INC(tcps_sc_completed);
1311 
1312 /* how do we find the inp for the new socket? */
1313 	if (sc != &scs)
1314 		syncache_free(sc);
1315 	return (1);
1316 failed:
1317 	if (sc != NULL && sc != &scs)
1318 		syncache_free(sc);
1319 	if (s != NULL)
1320 		free(s, M_TCPLOG);
1321 	*lsop = NULL;
1322 	return (0);
1323 }
1324 
1325 static struct socket *
1326 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m,
1327     uint64_t response_cookie)
1328 {
1329 	struct inpcb *inp;
1330 	struct tcpcb *tp;
1331 	unsigned int *pending_counter;
1332 	struct socket *so;
1333 
1334 	NET_EPOCH_ASSERT();
1335 
1336 	pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending;
1337 	so = syncache_socket(sc, lso, m);
1338 	if (so == NULL) {
1339 		TCPSTAT_INC(tcps_sc_aborted);
1340 		atomic_subtract_int(pending_counter, 1);
1341 	} else {
1342 		soisconnected(so);
1343 		inp = sotoinpcb(so);
1344 		tp = intotcpcb(inp);
1345 		tp->t_flags |= TF_FASTOPEN;
1346 		tp->t_tfo_cookie.server = response_cookie;
1347 		tp->snd_max = tp->iss;
1348 		tp->snd_nxt = tp->iss;
1349 		tp->t_tfo_pending = pending_counter;
1350 		TCPSTAT_INC(tcps_sc_completed);
1351 	}
1352 
1353 	return (so);
1354 }
1355 
1356 /*
1357  * Given a LISTEN socket and an inbound SYN request, add
1358  * this to the syn cache, and send back a segment:
1359  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1360  * to the source.
1361  *
1362  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1363  * Doing so would require that we hold onto the data and deliver it
1364  * to the application.  However, if we are the target of a SYN-flood
1365  * DoS attack, an attacker could send data which would eventually
1366  * consume all available buffer space if it were ACKed.  By not ACKing
1367  * the data, we avoid this DoS scenario.
1368  *
1369  * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1370  * cookie is processed and a new socket is created.  In this case, any data
1371  * accompanying the SYN will be queued to the socket by tcp_input() and will
1372  * be ACKed either when the application sends response data or the delayed
1373  * ACK timer expires, whichever comes first.
1374  */
1375 struct socket *
1376 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1377     struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod,
1378     void *todctx, uint8_t iptos)
1379 {
1380 	struct tcpcb *tp;
1381 	struct socket *rv = NULL;
1382 	struct syncache *sc = NULL;
1383 	struct syncache_head *sch;
1384 	struct mbuf *ipopts = NULL;
1385 	u_int ltflags;
1386 	int win, ip_ttl, ip_tos;
1387 	char *s;
1388 #ifdef INET6
1389 	int autoflowlabel = 0;
1390 #endif
1391 #ifdef MAC
1392 	struct label *maclabel;
1393 #endif
1394 	struct syncache scs;
1395 	struct ucred *cred;
1396 	uint64_t tfo_response_cookie;
1397 	unsigned int *tfo_pending = NULL;
1398 	int tfo_cookie_valid = 0;
1399 	int tfo_response_cookie_valid = 0;
1400 	bool locked;
1401 
1402 	INP_RLOCK_ASSERT(inp);			/* listen socket */
1403 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1404 	    ("%s: unexpected tcp flags", __func__));
1405 
1406 	/*
1407 	 * Combine all so/tp operations very early to drop the INP lock as
1408 	 * soon as possible.
1409 	 */
1410 	KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1411 	tp = sototcpcb(so);
1412 	cred = crhold(so->so_cred);
1413 
1414 #ifdef INET6
1415 	if (inc->inc_flags & INC_ISIPV6) {
1416 		if (inp->inp_flags & IN6P_AUTOFLOWLABEL) {
1417 			autoflowlabel = 1;
1418 		}
1419 		ip_ttl = in6_selecthlim(inp, NULL);
1420 		if ((inp->in6p_outputopts == NULL) ||
1421 		    (inp->in6p_outputopts->ip6po_tclass == -1)) {
1422 			ip_tos = 0;
1423 		} else {
1424 			ip_tos = inp->in6p_outputopts->ip6po_tclass;
1425 		}
1426 	}
1427 #endif
1428 #if defined(INET6) && defined(INET)
1429 	else
1430 #endif
1431 #ifdef INET
1432 	{
1433 		ip_ttl = inp->inp_ip_ttl;
1434 		ip_tos = inp->inp_ip_tos;
1435 	}
1436 #endif
1437 	win = so->sol_sbrcv_hiwat;
1438 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1439 
1440 	if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
1441 	    (tp->t_tfo_pending != NULL) &&
1442 	    (to->to_flags & TOF_FASTOPEN)) {
1443 		/*
1444 		 * Limit the number of pending TFO connections to
1445 		 * approximately half of the queue limit.  This prevents TFO
1446 		 * SYN floods from starving the service by filling the
1447 		 * listen queue with bogus TFO connections.
1448 		 */
1449 		if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1450 		    (so->sol_qlimit / 2)) {
1451 			int result;
1452 
1453 			result = tcp_fastopen_check_cookie(inc,
1454 			    to->to_tfo_cookie, to->to_tfo_len,
1455 			    &tfo_response_cookie);
1456 			tfo_cookie_valid = (result > 0);
1457 			tfo_response_cookie_valid = (result >= 0);
1458 		}
1459 
1460 		/*
1461 		 * Remember the TFO pending counter as it will have to be
1462 		 * decremented below if we don't make it to syncache_tfo_expand().
1463 		 */
1464 		tfo_pending = tp->t_tfo_pending;
1465 	}
1466 
1467 	/* By the time we drop the lock these should no longer be used. */
1468 	so = NULL;
1469 	tp = NULL;
1470 
1471 #ifdef MAC
1472 	if (mac_syncache_init(&maclabel) != 0) {
1473 		INP_RUNLOCK(inp);
1474 		goto done;
1475 	} else
1476 		mac_syncache_create(maclabel, inp);
1477 #endif
1478 	if (!tfo_cookie_valid)
1479 		INP_RUNLOCK(inp);
1480 
1481 	/*
1482 	 * Remember the IP options, if any.
1483 	 */
1484 #ifdef INET6
1485 	if (!(inc->inc_flags & INC_ISIPV6))
1486 #endif
1487 #ifdef INET
1488 		ipopts = (m) ? ip_srcroute(m) : NULL;
1489 #else
1490 		ipopts = NULL;
1491 #endif
1492 
1493 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1494 	/*
1495 	 * If listening socket requested TCP digests, check that received
1496 	 * SYN has signature and it is correct. If signature doesn't match
1497 	 * or TCP_SIGNATURE support isn't enabled, drop the packet.
1498 	 */
1499 	if (ltflags & TF_SIGNATURE) {
1500 		if ((to->to_flags & TOF_SIGNATURE) == 0) {
1501 			TCPSTAT_INC(tcps_sig_err_nosigopt);
1502 			goto done;
1503 		}
1504 		if (!TCPMD5_ENABLED() ||
1505 		    TCPMD5_INPUT(m, th, to->to_signature) != 0)
1506 			goto done;
1507 	}
1508 #endif	/* TCP_SIGNATURE */
1509 	/*
1510 	 * See if we already have an entry for this connection.
1511 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1512 	 *
1513 	 * XXX: should the syncache be re-initialized with the contents
1514 	 * of the new SYN here (which may have different options?)
1515 	 *
1516 	 * XXX: We do not check the sequence number to see if this is a
1517 	 * real retransmit or a new connection attempt.  The question is
1518 	 * how to handle such a case; either ignore it as spoofed, or
1519 	 * drop the current entry and create a new one?
1520 	 */
1521 	if (syncache_cookiesonly()) {
1522 		sc = NULL;
1523 		sch = syncache_hashbucket(inc);
1524 		locked = false;
1525 	} else {
1526 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1527 		locked = true;
1528 		SCH_LOCK_ASSERT(sch);
1529 	}
1530 	if (sc != NULL) {
1531 		if (tfo_cookie_valid)
1532 			INP_RUNLOCK(inp);
1533 		TCPSTAT_INC(tcps_sc_dupsyn);
1534 		if (ipopts) {
1535 			/*
1536 			 * If we were remembering a previous source route,
1537 			 * forget it and use the new one we've been given.
1538 			 */
1539 			if (sc->sc_ipopts)
1540 				(void) m_free(sc->sc_ipopts);
1541 			sc->sc_ipopts = ipopts;
1542 		}
1543 		/*
1544 		 * Update timestamp if present.
1545 		 */
1546 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1547 			sc->sc_tsreflect = to->to_tsval;
1548 		else
1549 			sc->sc_flags &= ~SCF_TIMESTAMP;
1550 		/*
1551 		 * Disable ECN if needed.
1552 		 */
1553 		if ((sc->sc_flags & SCF_ECN) &&
1554 		    ((th->th_flags & (TH_ECE|TH_CWR)) != (TH_ECE|TH_CWR))) {
1555 			sc->sc_flags &= ~SCF_ECN;
1556 		}
1557 #ifdef MAC
1558 		/*
1559 		 * Since we have already unconditionally allocated label
1560 		 * storage, free it up.  The syncache entry will already
1561 		 * have an initialized label we can use.
1562 		 */
1563 		mac_syncache_destroy(&maclabel);
1564 #endif
1565 		TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1566 		/* Retransmit SYN|ACK and reset retransmit count. */
1567 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1568 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1569 			    "resetting timer and retransmitting SYN|ACK\n",
1570 			    s, __func__);
1571 			free(s, M_TCPLOG);
1572 		}
1573 		if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1574 			sc->sc_rxmits = 0;
1575 			syncache_timeout(sc, sch, 1);
1576 			TCPSTAT_INC(tcps_sndacks);
1577 			TCPSTAT_INC(tcps_sndtotal);
1578 		}
1579 		SCH_UNLOCK(sch);
1580 		goto donenoprobe;
1581 	}
1582 
1583 	if (tfo_cookie_valid) {
1584 		bzero(&scs, sizeof(scs));
1585 		sc = &scs;
1586 		goto skip_alloc;
1587 	}
1588 
1589 	/*
1590 	 * Skip allocating a syncache entry if we are just going to discard
1591 	 * it later.
1592 	 */
1593 	if (!locked) {
1594 		bzero(&scs, sizeof(scs));
1595 		sc = &scs;
1596 	} else
1597 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1598 	if (sc == NULL) {
1599 		/*
1600 		 * The zone allocator couldn't provide more entries.
1601 		 * Treat this as if the cache was full; drop the oldest
1602 		 * entry and insert the new one.
1603 		 */
1604 		TCPSTAT_INC(tcps_sc_zonefail);
1605 		if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1606 			sch->sch_last_overflow = time_uptime;
1607 			syncache_drop(sc, sch);
1608 			syncache_pause(inc);
1609 		}
1610 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1611 		if (sc == NULL) {
1612 			if (V_tcp_syncookies) {
1613 				bzero(&scs, sizeof(scs));
1614 				sc = &scs;
1615 			} else {
1616 				KASSERT(locked,
1617 				    ("%s: bucket unexpectedly unlocked",
1618 				    __func__));
1619 				SCH_UNLOCK(sch);
1620 				if (ipopts)
1621 					(void) m_free(ipopts);
1622 				goto done;
1623 			}
1624 		}
1625 	}
1626 
1627 skip_alloc:
1628 	if (!tfo_cookie_valid && tfo_response_cookie_valid)
1629 		sc->sc_tfo_cookie = &tfo_response_cookie;
1630 
1631 	/*
1632 	 * Fill in the syncache values.
1633 	 */
1634 #ifdef MAC
1635 	sc->sc_label = maclabel;
1636 #endif
1637 	sc->sc_cred = cred;
1638 	cred = NULL;
1639 	sc->sc_ipopts = ipopts;
1640 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1641 	sc->sc_ip_tos = ip_tos;
1642 	sc->sc_ip_ttl = ip_ttl;
1643 #ifdef TCP_OFFLOAD
1644 	sc->sc_tod = tod;
1645 	sc->sc_todctx = todctx;
1646 #endif
1647 	sc->sc_irs = th->th_seq;
1648 	sc->sc_flags = 0;
1649 	sc->sc_flowlabel = 0;
1650 
1651 	/*
1652 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1653 	 * win was derived from socket earlier in the function.
1654 	 */
1655 	win = imax(win, 0);
1656 	win = imin(win, TCP_MAXWIN);
1657 	sc->sc_wnd = win;
1658 
1659 	if (V_tcp_do_rfc1323 &&
1660 	    !(ltflags & TF_NOOPT)) {
1661 		/*
1662 		 * A timestamp received in a SYN makes
1663 		 * it ok to send timestamp requests and replies.
1664 		 */
1665 		if (to->to_flags & TOF_TS) {
1666 			sc->sc_tsreflect = to->to_tsval;
1667 			sc->sc_flags |= SCF_TIMESTAMP;
1668 			sc->sc_tsoff = tcp_new_ts_offset(inc);
1669 		}
1670 		if (to->to_flags & TOF_SCALE) {
1671 			int wscale = 0;
1672 
1673 			/*
1674 			 * Pick the smallest possible scaling factor that
1675 			 * will still allow us to scale up to sb_max, aka
1676 			 * kern.ipc.maxsockbuf.
1677 			 *
1678 			 * We do this because there are broken firewalls that
1679 			 * will corrupt the window scale option, leading to
1680 			 * the other endpoint believing that our advertised
1681 			 * window is unscaled.  At scale factors larger than
1682 			 * 5 the unscaled window will drop below 1500 bytes,
1683 			 * leading to serious problems when traversing these
1684 			 * broken firewalls.
1685 			 *
1686 			 * With the default maxsockbuf of 256K, a scale factor
1687 			 * of 3 will be chosen by this algorithm.  Those who
1688 			 * choose a larger maxsockbuf should watch out
1689 			 * for the compatibility problems mentioned above.
1690 			 *
1691 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1692 			 * or <SYN,ACK>) segment itself is never scaled.
1693 			 */
1694 			while (wscale < TCP_MAX_WINSHIFT &&
1695 			    (TCP_MAXWIN << wscale) < sb_max)
1696 				wscale++;
1697 			sc->sc_requested_r_scale = wscale;
1698 			sc->sc_requested_s_scale = to->to_wscale;
1699 			sc->sc_flags |= SCF_WINSCALE;
1700 		}
1701 	}
1702 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1703 	/*
1704 	 * If listening socket requested TCP digests, flag this in the
1705 	 * syncache so that syncache_respond() will do the right thing
1706 	 * with the SYN+ACK.
1707 	 */
1708 	if (ltflags & TF_SIGNATURE)
1709 		sc->sc_flags |= SCF_SIGNATURE;
1710 #endif	/* TCP_SIGNATURE */
1711 	if (to->to_flags & TOF_SACKPERM)
1712 		sc->sc_flags |= SCF_SACK;
1713 	if (to->to_flags & TOF_MSS)
1714 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1715 	if (ltflags & TF_NOOPT)
1716 		sc->sc_flags |= SCF_NOOPT;
1717 	if (((th->th_flags & (TH_ECE|TH_CWR)) == (TH_ECE|TH_CWR)) &&
1718 	    V_tcp_do_ecn)
1719 		sc->sc_flags |= SCF_ECN;
1720 
1721 	if (V_tcp_syncookies)
1722 		sc->sc_iss = syncookie_generate(sch, sc);
1723 	else
1724 		sc->sc_iss = arc4random();
1725 #ifdef INET6
1726 	if (autoflowlabel) {
1727 		if (V_tcp_syncookies)
1728 			sc->sc_flowlabel = sc->sc_iss;
1729 		else
1730 			sc->sc_flowlabel = ip6_randomflowlabel();
1731 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1732 	}
1733 #endif
1734 	if (locked)
1735 		SCH_UNLOCK(sch);
1736 
1737 	if (tfo_cookie_valid) {
1738 		rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie);
1739 		/* INP_RUNLOCK(inp) will be performed by the caller */
1740 		goto tfo_expanded;
1741 	}
1742 
1743 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1744 	/*
1745 	 * Do a standard 3-way handshake.
1746 	 */
1747 	if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1748 		if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1749 			syncache_free(sc);
1750 		else if (sc != &scs)
1751 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1752 		TCPSTAT_INC(tcps_sndacks);
1753 		TCPSTAT_INC(tcps_sndtotal);
1754 	} else {
1755 		if (sc != &scs)
1756 			syncache_free(sc);
1757 		TCPSTAT_INC(tcps_sc_dropped);
1758 	}
1759 	goto donenoprobe;
1760 
1761 done:
1762 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1763 donenoprobe:
1764 	if (m)
1765 		m_freem(m);
1766 	/*
1767 	 * If tfo_pending is not NULL here, then a TFO SYN that did not
1768 	 * result in a new socket was processed and the associated pending
1769 	 * counter has not yet been decremented.  All such TFO processing paths
1770 	 * transit this point.
1771 	 */
1772 	if (tfo_pending != NULL)
1773 		tcp_fastopen_decrement_counter(tfo_pending);
1774 
1775 tfo_expanded:
1776 	if (cred != NULL)
1777 		crfree(cred);
1778 #ifdef MAC
1779 	if (sc == &scs)
1780 		mac_syncache_destroy(&maclabel);
1781 #endif
1782 	return (rv);
1783 }
1784 
1785 /*
1786  * Send SYN|ACK or ACK to the peer.  Either in response to a peer's segment,
1787  * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1788  */
1789 static int
1790 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1791 {
1792 	struct ip *ip = NULL;
1793 	struct mbuf *m;
1794 	struct tcphdr *th = NULL;
1795 	int optlen, error = 0;	/* Make compiler happy */
1796 	u_int16_t hlen, tlen, mssopt;
1797 	struct tcpopt to;
1798 #ifdef INET6
1799 	struct ip6_hdr *ip6 = NULL;
1800 #endif
1801 
1802 	NET_EPOCH_ASSERT();
1803 
1804 	hlen =
1805 #ifdef INET6
1806 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1807 #endif
1808 		sizeof(struct ip);
1809 	tlen = hlen + sizeof(struct tcphdr);
1810 
1811 	/* Determine MSS we advertize to other end of connection. */
1812 	mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss);
1813 
1814 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1815 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1816 	    ("syncache: mbuf too small"));
1817 
1818 	/* Create the IP+TCP header from scratch. */
1819 	m = m_gethdr(M_NOWAIT, MT_DATA);
1820 	if (m == NULL)
1821 		return (ENOBUFS);
1822 #ifdef MAC
1823 	mac_syncache_create_mbuf(sc->sc_label, m);
1824 #endif
1825 	m->m_data += max_linkhdr;
1826 	m->m_len = tlen;
1827 	m->m_pkthdr.len = tlen;
1828 	m->m_pkthdr.rcvif = NULL;
1829 
1830 #ifdef INET6
1831 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1832 		ip6 = mtod(m, struct ip6_hdr *);
1833 		ip6->ip6_vfc = IPV6_VERSION;
1834 		ip6->ip6_nxt = IPPROTO_TCP;
1835 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1836 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1837 		ip6->ip6_plen = htons(tlen - hlen);
1838 		/* ip6_hlim is set after checksum */
1839 		/* Zero out traffic class and flow label. */
1840 		ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1841 		ip6->ip6_flow |= sc->sc_flowlabel;
1842 		ip6->ip6_flow |= htonl(sc->sc_ip_tos << 20);
1843 
1844 		th = (struct tcphdr *)(ip6 + 1);
1845 	}
1846 #endif
1847 #if defined(INET6) && defined(INET)
1848 	else
1849 #endif
1850 #ifdef INET
1851 	{
1852 		ip = mtod(m, struct ip *);
1853 		ip->ip_v = IPVERSION;
1854 		ip->ip_hl = sizeof(struct ip) >> 2;
1855 		ip->ip_len = htons(tlen);
1856 		ip->ip_id = 0;
1857 		ip->ip_off = 0;
1858 		ip->ip_sum = 0;
1859 		ip->ip_p = IPPROTO_TCP;
1860 		ip->ip_src = sc->sc_inc.inc_laddr;
1861 		ip->ip_dst = sc->sc_inc.inc_faddr;
1862 		ip->ip_ttl = sc->sc_ip_ttl;
1863 		ip->ip_tos = sc->sc_ip_tos;
1864 
1865 		/*
1866 		 * See if we should do MTU discovery.  Route lookups are
1867 		 * expensive, so we will only unset the DF bit if:
1868 		 *
1869 		 *	1) path_mtu_discovery is disabled
1870 		 *	2) the SCF_UNREACH flag has been set
1871 		 */
1872 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1873 		       ip->ip_off |= htons(IP_DF);
1874 
1875 		th = (struct tcphdr *)(ip + 1);
1876 	}
1877 #endif /* INET */
1878 	th->th_sport = sc->sc_inc.inc_lport;
1879 	th->th_dport = sc->sc_inc.inc_fport;
1880 
1881 	if (flags & TH_SYN)
1882 		th->th_seq = htonl(sc->sc_iss);
1883 	else
1884 		th->th_seq = htonl(sc->sc_iss + 1);
1885 	th->th_ack = htonl(sc->sc_irs + 1);
1886 	th->th_off = sizeof(struct tcphdr) >> 2;
1887 	th->th_x2 = 0;
1888 	th->th_flags = flags;
1889 	th->th_win = htons(sc->sc_wnd);
1890 	th->th_urp = 0;
1891 
1892 	if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) {
1893 		th->th_flags |= TH_ECE;
1894 		TCPSTAT_INC(tcps_ecn_shs);
1895 	}
1896 
1897 	/* Tack on the TCP options. */
1898 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1899 		to.to_flags = 0;
1900 
1901 		if (flags & TH_SYN) {
1902 			to.to_mss = mssopt;
1903 			to.to_flags = TOF_MSS;
1904 			if (sc->sc_flags & SCF_WINSCALE) {
1905 				to.to_wscale = sc->sc_requested_r_scale;
1906 				to.to_flags |= TOF_SCALE;
1907 			}
1908 			if (sc->sc_flags & SCF_SACK)
1909 				to.to_flags |= TOF_SACKPERM;
1910 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1911 			if (sc->sc_flags & SCF_SIGNATURE)
1912 				to.to_flags |= TOF_SIGNATURE;
1913 #endif
1914 			if (sc->sc_tfo_cookie) {
1915 				to.to_flags |= TOF_FASTOPEN;
1916 				to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1917 				to.to_tfo_cookie = sc->sc_tfo_cookie;
1918 				/* don't send cookie again when retransmitting response */
1919 				sc->sc_tfo_cookie = NULL;
1920 			}
1921 		}
1922 		if (sc->sc_flags & SCF_TIMESTAMP) {
1923 			to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1924 			to.to_tsecr = sc->sc_tsreflect;
1925 			to.to_flags |= TOF_TS;
1926 		}
1927 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1928 
1929 		/* Adjust headers by option size. */
1930 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1931 		m->m_len += optlen;
1932 		m->m_pkthdr.len += optlen;
1933 #ifdef INET6
1934 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1935 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1936 		else
1937 #endif
1938 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1939 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1940 		if (sc->sc_flags & SCF_SIGNATURE) {
1941 			KASSERT(to.to_flags & TOF_SIGNATURE,
1942 			    ("tcp_addoptions() didn't set tcp_signature"));
1943 
1944 			/* NOTE: to.to_signature is inside of mbuf */
1945 			if (!TCPMD5_ENABLED() ||
1946 			    TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1947 				m_freem(m);
1948 				return (EACCES);
1949 			}
1950 		}
1951 #endif
1952 	} else
1953 		optlen = 0;
1954 
1955 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
1956 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1957 	/*
1958 	 * If we have peer's SYN and it has a flowid, then let's assign it to
1959 	 * our SYN|ACK.  ip6_output() and ip_output() will not assign flowid
1960 	 * to SYN|ACK due to lack of inp here.
1961 	 */
1962 	if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1963 		m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1964 		M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1965 	}
1966 #ifdef INET6
1967 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1968 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1969 		th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1970 		    IPPROTO_TCP, 0);
1971 		ip6->ip6_hlim = sc->sc_ip_ttl;
1972 #ifdef TCP_OFFLOAD
1973 		if (ADDED_BY_TOE(sc)) {
1974 			struct toedev *tod = sc->sc_tod;
1975 
1976 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1977 
1978 			return (error);
1979 		}
1980 #endif
1981 		TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
1982 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1983 	}
1984 #endif
1985 #if defined(INET6) && defined(INET)
1986 	else
1987 #endif
1988 #ifdef INET
1989 	{
1990 		m->m_pkthdr.csum_flags = CSUM_TCP;
1991 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1992 		    htons(tlen + optlen - hlen + IPPROTO_TCP));
1993 #ifdef TCP_OFFLOAD
1994 		if (ADDED_BY_TOE(sc)) {
1995 			struct toedev *tod = sc->sc_tod;
1996 
1997 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1998 
1999 			return (error);
2000 		}
2001 #endif
2002 		TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
2003 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
2004 	}
2005 #endif
2006 	return (error);
2007 }
2008 
2009 /*
2010  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
2011  * that exceed the capacity of the syncache by avoiding the storage of any
2012  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
2013  * attacks where the attacker does not have access to our responses.
2014  *
2015  * Syncookies encode and include all necessary information about the
2016  * connection setup within the SYN|ACK that we send back.  That way we
2017  * can avoid keeping any local state until the ACK to our SYN|ACK returns
2018  * (if ever).  Normally the syncache and syncookies are running in parallel
2019  * with the latter taking over when the former is exhausted.  When matching
2020  * syncache entry is found the syncookie is ignored.
2021  *
2022  * The only reliable information persisting the 3WHS is our initial sequence
2023  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
2024  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
2025  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
2026  * returns and signifies a legitimate connection if it matches the ACK.
2027  *
2028  * The available space of 32 bits to store the hash and to encode the SYN
2029  * option information is very tight and we should have at least 24 bits for
2030  * the MAC to keep the number of guesses by blind spoofing reasonably high.
2031  *
2032  * SYN option information we have to encode to fully restore a connection:
2033  * MSS: is imporant to chose an optimal segment size to avoid IP level
2034  *   fragmentation along the path.  The common MSS values can be encoded
2035  *   in a 3-bit table.  Uncommon values are captured by the next lower value
2036  *   in the table leading to a slight increase in packetization overhead.
2037  * WSCALE: is necessary to allow large windows to be used for high delay-
2038  *   bandwidth product links.  Not scaling the window when it was initially
2039  *   negotiated is bad for performance as lack of scaling further decreases
2040  *   the apparent available send window.  We only need to encode the WSCALE
2041  *   we received from the remote end.  Our end can be recalculated at any
2042  *   time.  The common WSCALE values can be encoded in a 3-bit table.
2043  *   Uncommon values are captured by the next lower value in the table
2044  *   making us under-estimate the available window size halving our
2045  *   theoretically possible maximum throughput for that connection.
2046  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2047  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2048  *   that are included in all segments on a connection.  We enable them when
2049  *   the ACK has them.
2050  *
2051  * Security of syncookies and attack vectors:
2052  *
2053  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2054  * together with the gloabl secret to make it unique per connection attempt.
2055  * Thus any change of any of those parameters results in a different MAC output
2056  * in an unpredictable way unless a collision is encountered.  24 bits of the
2057  * MAC are embedded into the ISS.
2058  *
2059  * To prevent replay attacks two rotating global secrets are updated with a
2060  * new random value every 15 seconds.  The life-time of a syncookie is thus
2061  * 15-30 seconds.
2062  *
2063  * Vector 1: Attacking the secret.  This requires finding a weakness in the
2064  * MAC itself or the way it is used here.  The attacker can do a chosen plain
2065  * text attack by varying and testing the all parameters under his control.
2066  * The strength depends on the size and randomness of the secret, and the
2067  * cryptographic security of the MAC function.  Due to the constant updating
2068  * of the secret the attacker has at most 29.999 seconds to find the secret
2069  * and launch spoofed connections.  After that he has to start all over again.
2070  *
2071  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
2072  * size an average of 4,823 attempts are required for a 50% chance of success
2073  * to spoof a single syncookie (birthday collision paradox).  However the
2074  * attacker is blind and doesn't know if one of his attempts succeeded unless
2075  * he has a side channel to interfere success from.  A single connection setup
2076  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2077  * This many attempts are required for each one blind spoofed connection.  For
2078  * every additional spoofed connection he has to launch another N attempts.
2079  * Thus for a sustained rate 100 spoofed connections per second approximately
2080  * 1,800,000 packets per second would have to be sent.
2081  *
2082  * NB: The MAC function should be fast so that it doesn't become a CPU
2083  * exhaustion attack vector itself.
2084  *
2085  * References:
2086  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2087  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2088  *   http://cr.yp.to/syncookies.html    (overview)
2089  *   http://cr.yp.to/syncookies/archive (details)
2090  *
2091  *
2092  * Schematic construction of a syncookie enabled Initial Sequence Number:
2093  *  0        1         2         3
2094  *  12345678901234567890123456789012
2095  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2096  *
2097  *  x 24 MAC (truncated)
2098  *  W  3 Send Window Scale index
2099  *  M  3 MSS index
2100  *  S  1 SACK permitted
2101  *  P  1 Odd/even secret
2102  */
2103 
2104 /*
2105  * Distribution and probability of certain MSS values.  Those in between are
2106  * rounded down to the next lower one.
2107  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2108  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
2109  */
2110 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2111 
2112 /*
2113  * Distribution and probability of certain WSCALE values.  We have to map the
2114  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2115  * bits based on prevalence of certain values.  Where we don't have an exact
2116  * match for are rounded down to the next lower one letting us under-estimate
2117  * the true available window.  At the moment this would happen only for the
2118  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2119  * and window size).  The absence of the WSCALE option (no scaling in either
2120  * direction) is encoded with index zero.
2121  * [WSCALE values histograms, Allman, 2012]
2122  *                            X 10 10 35  5  6 14 10%   by host
2123  *                            X 11  4  5  5 18 49  3%   by connections
2124  */
2125 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2126 
2127 /*
2128  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
2129  * and good cryptographic properties.
2130  */
2131 static uint32_t
2132 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2133     uint8_t *secbits, uintptr_t secmod)
2134 {
2135 	SIPHASH_CTX ctx;
2136 	uint32_t siphash[2];
2137 
2138 	SipHash24_Init(&ctx);
2139 	SipHash_SetKey(&ctx, secbits);
2140 	switch (inc->inc_flags & INC_ISIPV6) {
2141 #ifdef INET
2142 	case 0:
2143 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2144 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2145 		break;
2146 #endif
2147 #ifdef INET6
2148 	case INC_ISIPV6:
2149 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2150 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2151 		break;
2152 #endif
2153 	}
2154 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2155 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2156 	SipHash_Update(&ctx, &irs, sizeof(irs));
2157 	SipHash_Update(&ctx, &flags, sizeof(flags));
2158 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
2159 	SipHash_Final((u_int8_t *)&siphash, &ctx);
2160 
2161 	return (siphash[0] ^ siphash[1]);
2162 }
2163 
2164 static tcp_seq
2165 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2166 {
2167 	u_int i, secbit, wscale;
2168 	uint32_t iss, hash;
2169 	uint8_t *secbits;
2170 	union syncookie cookie;
2171 
2172 	cookie.cookie = 0;
2173 
2174 	/* Map our computed MSS into the 3-bit index. */
2175 	for (i = nitems(tcp_sc_msstab) - 1;
2176 	     tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2177 	     i--)
2178 		;
2179 	cookie.flags.mss_idx = i;
2180 
2181 	/*
2182 	 * Map the send window scale into the 3-bit index but only if
2183 	 * the wscale option was received.
2184 	 */
2185 	if (sc->sc_flags & SCF_WINSCALE) {
2186 		wscale = sc->sc_requested_s_scale;
2187 		for (i = nitems(tcp_sc_wstab) - 1;
2188 		    tcp_sc_wstab[i] > wscale && i > 0;
2189 		     i--)
2190 			;
2191 		cookie.flags.wscale_idx = i;
2192 	}
2193 
2194 	/* Can we do SACK? */
2195 	if (sc->sc_flags & SCF_SACK)
2196 		cookie.flags.sack_ok = 1;
2197 
2198 	/* Which of the two secrets to use. */
2199 	secbit = V_tcp_syncache.secret.oddeven & 0x1;
2200 	cookie.flags.odd_even = secbit;
2201 
2202 	secbits = V_tcp_syncache.secret.key[secbit];
2203 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2204 	    (uintptr_t)sch);
2205 
2206 	/*
2207 	 * Put the flags into the hash and XOR them to get better ISS number
2208 	 * variance.  This doesn't enhance the cryptographic strength and is
2209 	 * done to prevent the 8 cookie bits from showing up directly on the
2210 	 * wire.
2211 	 */
2212 	iss = hash & ~0xff;
2213 	iss |= cookie.cookie ^ (hash >> 24);
2214 
2215 	TCPSTAT_INC(tcps_sc_sendcookie);
2216 	return (iss);
2217 }
2218 
2219 static struct syncache *
2220 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2221     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2222     struct socket *lso)
2223 {
2224 	uint32_t hash;
2225 	uint8_t *secbits;
2226 	tcp_seq ack, seq;
2227 	int wnd, wscale = 0;
2228 	union syncookie cookie;
2229 
2230 	/*
2231 	 * Pull information out of SYN-ACK/ACK and revert sequence number
2232 	 * advances.
2233 	 */
2234 	ack = th->th_ack - 1;
2235 	seq = th->th_seq - 1;
2236 
2237 	/*
2238 	 * Unpack the flags containing enough information to restore the
2239 	 * connection.
2240 	 */
2241 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2242 
2243 	/* Which of the two secrets to use. */
2244 	secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2245 
2246 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2247 
2248 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
2249 	if ((ack & ~0xff) != (hash & ~0xff))
2250 		return (NULL);
2251 
2252 	/* Fill in the syncache values. */
2253 	sc->sc_flags = 0;
2254 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2255 	sc->sc_ipopts = NULL;
2256 
2257 	sc->sc_irs = seq;
2258 	sc->sc_iss = ack;
2259 
2260 	switch (inc->inc_flags & INC_ISIPV6) {
2261 #ifdef INET
2262 	case 0:
2263 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2264 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2265 		break;
2266 #endif
2267 #ifdef INET6
2268 	case INC_ISIPV6:
2269 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2270 			sc->sc_flowlabel =
2271 			    htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2272 		break;
2273 #endif
2274 	}
2275 
2276 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2277 
2278 	/* We can simply recompute receive window scale we sent earlier. */
2279 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2280 		wscale++;
2281 
2282 	/* Only use wscale if it was enabled in the orignal SYN. */
2283 	if (cookie.flags.wscale_idx > 0) {
2284 		sc->sc_requested_r_scale = wscale;
2285 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2286 		sc->sc_flags |= SCF_WINSCALE;
2287 	}
2288 
2289 	wnd = lso->sol_sbrcv_hiwat;
2290 	wnd = imax(wnd, 0);
2291 	wnd = imin(wnd, TCP_MAXWIN);
2292 	sc->sc_wnd = wnd;
2293 
2294 	if (cookie.flags.sack_ok)
2295 		sc->sc_flags |= SCF_SACK;
2296 
2297 	if (to->to_flags & TOF_TS) {
2298 		sc->sc_flags |= SCF_TIMESTAMP;
2299 		sc->sc_tsreflect = to->to_tsval;
2300 		sc->sc_tsoff = tcp_new_ts_offset(inc);
2301 	}
2302 
2303 	if (to->to_flags & TOF_SIGNATURE)
2304 		sc->sc_flags |= SCF_SIGNATURE;
2305 
2306 	sc->sc_rxmits = 0;
2307 
2308 	TCPSTAT_INC(tcps_sc_recvcookie);
2309 	return (sc);
2310 }
2311 
2312 #ifdef INVARIANTS
2313 static int
2314 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2315     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2316     struct socket *lso)
2317 {
2318 	struct syncache scs, *scx;
2319 	char *s;
2320 
2321 	bzero(&scs, sizeof(scs));
2322 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
2323 
2324 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2325 		return (0);
2326 
2327 	if (scx != NULL) {
2328 		if (sc->sc_peer_mss != scx->sc_peer_mss)
2329 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2330 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2331 
2332 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2333 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2334 			    s, __func__, sc->sc_requested_r_scale,
2335 			    scx->sc_requested_r_scale);
2336 
2337 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2338 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2339 			    s, __func__, sc->sc_requested_s_scale,
2340 			    scx->sc_requested_s_scale);
2341 
2342 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2343 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2344 	}
2345 
2346 	if (s != NULL)
2347 		free(s, M_TCPLOG);
2348 	return (0);
2349 }
2350 #endif /* INVARIANTS */
2351 
2352 static void
2353 syncookie_reseed(void *arg)
2354 {
2355 	struct tcp_syncache *sc = arg;
2356 	uint8_t *secbits;
2357 	int secbit;
2358 
2359 	/*
2360 	 * Reseeding the secret doesn't have to be protected by a lock.
2361 	 * It only must be ensured that the new random values are visible
2362 	 * to all CPUs in a SMP environment.  The atomic with release
2363 	 * semantics ensures that.
2364 	 */
2365 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2366 	secbits = sc->secret.key[secbit];
2367 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2368 	atomic_add_rel_int(&sc->secret.oddeven, 1);
2369 
2370 	/* Reschedule ourself. */
2371 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2372 }
2373 
2374 /*
2375  * We have overflowed a bucket. Let's pause dealing with the syncache.
2376  * This function will increment the bucketoverflow statistics appropriately
2377  * (once per pause when pausing is enabled; otherwise, once per overflow).
2378  */
2379 static void
2380 syncache_pause(struct in_conninfo *inc)
2381 {
2382 	time_t delta;
2383 	const char *s;
2384 
2385 	/* XXX:
2386 	 * 2. Add sysctl read here so we don't get the benefit of this
2387 	 * change without the new sysctl.
2388 	 */
2389 
2390 	/*
2391 	 * Try an unlocked read. If we already know that another thread
2392 	 * has activated the feature, there is no need to proceed.
2393 	 */
2394 	if (V_tcp_syncache.paused)
2395 		return;
2396 
2397 	/* Are cookied enabled? If not, we can't pause. */
2398 	if (!V_tcp_syncookies) {
2399 		TCPSTAT_INC(tcps_sc_bucketoverflow);
2400 		return;
2401 	}
2402 
2403 	/*
2404 	 * We may be the first thread to find an overflow. Get the lock
2405 	 * and evaluate if we need to take action.
2406 	 */
2407 	mtx_lock(&V_tcp_syncache.pause_mtx);
2408 	if (V_tcp_syncache.paused) {
2409 		mtx_unlock(&V_tcp_syncache.pause_mtx);
2410 		return;
2411 	}
2412 
2413 	/* Activate protection. */
2414 	V_tcp_syncache.paused = true;
2415 	TCPSTAT_INC(tcps_sc_bucketoverflow);
2416 
2417 	/*
2418 	 * Determine the last backoff time. If we are seeing a re-newed
2419 	 * attack within that same time after last reactivating the syncache,
2420 	 * consider it an extension of the same attack.
2421 	 */
2422 	delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2423 	if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2424 		if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2425 			delta <<= 1;
2426 			V_tcp_syncache.pause_backoff++;
2427 		}
2428 	} else {
2429 		delta = TCP_SYNCACHE_PAUSE_TIME;
2430 		V_tcp_syncache.pause_backoff = 0;
2431 	}
2432 
2433 	/* Log a warning, including IP addresses, if able. */
2434 	if (inc != NULL)
2435 		s = tcp_log_addrs(inc, NULL, NULL, NULL);
2436 	else
2437 		s = (const char *)NULL;
2438 	log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2439 	    "the next %lld seconds%s%s%s\n", (long long)delta,
2440 	    (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2441 	    (s != NULL) ? ")" : "");
2442 	free(__DECONST(void *, s), M_TCPLOG);
2443 
2444 	/* Use the calculated delta to set a new pause time. */
2445 	V_tcp_syncache.pause_until = time_uptime + delta;
2446 	callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2447 	    &V_tcp_syncache);
2448 	mtx_unlock(&V_tcp_syncache.pause_mtx);
2449 }
2450 
2451 /* Evaluate whether we need to unpause. */
2452 static void
2453 syncache_unpause(void *arg)
2454 {
2455 	struct tcp_syncache *sc;
2456 	time_t delta;
2457 
2458 	sc = arg;
2459 	mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2460 	callout_deactivate(&sc->pause_co);
2461 
2462 	/*
2463 	 * Check to make sure we are not running early. If the pause
2464 	 * time has expired, then deactivate the protection.
2465 	 */
2466 	if ((delta = sc->pause_until - time_uptime) > 0)
2467 		callout_schedule(&sc->pause_co, delta * hz);
2468 	else
2469 		sc->paused = false;
2470 }
2471 
2472 /*
2473  * Exports the syncache entries to userland so that netstat can display
2474  * them alongside the other sockets.  This function is intended to be
2475  * called only from tcp_pcblist.
2476  *
2477  * Due to concurrency on an active system, the number of pcbs exported
2478  * may have no relation to max_pcbs.  max_pcbs merely indicates the
2479  * amount of space the caller allocated for this function to use.
2480  */
2481 int
2482 syncache_pcblist(struct sysctl_req *req)
2483 {
2484 	struct xtcpcb xt;
2485 	struct syncache *sc;
2486 	struct syncache_head *sch;
2487 	int error, i;
2488 
2489 	bzero(&xt, sizeof(xt));
2490 	xt.xt_len = sizeof(xt);
2491 	xt.t_state = TCPS_SYN_RECEIVED;
2492 	xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2493 	xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2494 	xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2495 	xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2496 
2497 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2498 		sch = &V_tcp_syncache.hashbase[i];
2499 		SCH_LOCK(sch);
2500 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2501 			if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2502 				continue;
2503 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2504 				xt.xt_inp.inp_vflag = INP_IPV6;
2505 			else
2506 				xt.xt_inp.inp_vflag = INP_IPV4;
2507 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2508 			    sizeof (struct in_conninfo));
2509 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2510 			if (error) {
2511 				SCH_UNLOCK(sch);
2512 				return (0);
2513 			}
2514 		}
2515 		SCH_UNLOCK(sch);
2516 	}
2517 
2518 	return (0);
2519 }
2520