xref: /freebsd/sys/netinet/tcp_syncache.c (revision 93a065e7496dfbfbd0a5b0208ef763f37ea975c7)
1 /*-
2  * Copyright (c) 2001 McAfee, Inc.
3  * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG
4  * All rights reserved.
5  *
6  * This software was developed for the FreeBSD Project by Jonathan Lemon
7  * and McAfee Research, the Security Research Division of McAfee, Inc. under
8  * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
9  * DARPA CHATS research program. [2001 McAfee, Inc.]
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  */
32 
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 #include "opt_inet.h"
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 #include "opt_pcbgroup.h"
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/hash.h>
44 #include <sys/refcount.h>
45 #include <sys/kernel.h>
46 #include <sys/sysctl.h>
47 #include <sys/limits.h>
48 #include <sys/lock.h>
49 #include <sys/mutex.h>
50 #include <sys/malloc.h>
51 #include <sys/mbuf.h>
52 #include <sys/proc.h>		/* for proc0 declaration */
53 #include <sys/random.h>
54 #include <sys/socket.h>
55 #include <sys/socketvar.h>
56 #include <sys/syslog.h>
57 #include <sys/ucred.h>
58 
59 #include <sys/md5.h>
60 #include <crypto/siphash/siphash.h>
61 
62 #include <vm/uma.h>
63 
64 #include <net/if.h>
65 #include <net/if_var.h>
66 #include <net/route.h>
67 #include <net/vnet.h>
68 
69 #include <netinet/in.h>
70 #include <netinet/in_systm.h>
71 #include <netinet/ip.h>
72 #include <netinet/in_var.h>
73 #include <netinet/in_pcb.h>
74 #include <netinet/ip_var.h>
75 #include <netinet/ip_options.h>
76 #ifdef INET6
77 #include <netinet/ip6.h>
78 #include <netinet/icmp6.h>
79 #include <netinet6/nd6.h>
80 #include <netinet6/ip6_var.h>
81 #include <netinet6/in6_pcb.h>
82 #endif
83 #include <netinet/tcp.h>
84 #ifdef TCP_RFC7413
85 #include <netinet/tcp_fastopen.h>
86 #endif
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 #ifdef INET6
93 #include <netinet6/tcp6_var.h>
94 #endif
95 #ifdef TCP_OFFLOAD
96 #include <netinet/toecore.h>
97 #endif
98 
99 #ifdef IPSEC
100 #include <netipsec/ipsec.h>
101 #ifdef INET6
102 #include <netipsec/ipsec6.h>
103 #endif
104 #include <netipsec/key.h>
105 #endif /*IPSEC*/
106 
107 #include <machine/in_cksum.h>
108 
109 #include <security/mac/mac_framework.h>
110 
111 static VNET_DEFINE(int, tcp_syncookies) = 1;
112 #define	V_tcp_syncookies		VNET(tcp_syncookies)
113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
114     &VNET_NAME(tcp_syncookies), 0,
115     "Use TCP SYN cookies if the syncache overflows");
116 
117 static VNET_DEFINE(int, tcp_syncookiesonly) = 0;
118 #define	V_tcp_syncookiesonly		VNET(tcp_syncookiesonly)
119 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
120     &VNET_NAME(tcp_syncookiesonly), 0,
121     "Use only TCP SYN cookies");
122 
123 static VNET_DEFINE(int, functions_inherit_listen_socket_stack) = 1;
124 #define V_functions_inherit_listen_socket_stack \
125     VNET(functions_inherit_listen_socket_stack)
126 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack,
127     CTLFLAG_VNET | CTLFLAG_RW,
128     &VNET_NAME(functions_inherit_listen_socket_stack), 0,
129     "Inherit listen socket's stack");
130 
131 #ifdef TCP_OFFLOAD
132 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
133 #endif
134 
135 static void	 syncache_drop(struct syncache *, struct syncache_head *);
136 static void	 syncache_free(struct syncache *);
137 static void	 syncache_insert(struct syncache *, struct syncache_head *);
138 static int	 syncache_respond(struct syncache *, struct syncache_head *, int,
139 		    const struct mbuf *);
140 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
141 		    struct mbuf *m);
142 static void	 syncache_timeout(struct syncache *sc, struct syncache_head *sch,
143 		    int docallout);
144 static void	 syncache_timer(void *);
145 
146 static uint32_t	 syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
147 		    uint8_t *, uintptr_t);
148 static tcp_seq	 syncookie_generate(struct syncache_head *, struct syncache *);
149 static struct syncache
150 		*syncookie_lookup(struct in_conninfo *, struct syncache_head *,
151 		    struct syncache *, struct tcphdr *, struct tcpopt *,
152 		    struct socket *);
153 static void	 syncookie_reseed(void *);
154 #ifdef INVARIANTS
155 static int	 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
156 		    struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
157 		    struct socket *lso);
158 #endif
159 
160 /*
161  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
162  * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds,
163  * the odds are that the user has given up attempting to connect by then.
164  */
165 #define SYNCACHE_MAXREXMTS		3
166 
167 /* Arbitrary values */
168 #define TCP_SYNCACHE_HASHSIZE		512
169 #define TCP_SYNCACHE_BUCKETLIMIT	30
170 
171 static VNET_DEFINE(struct tcp_syncache, tcp_syncache);
172 #define	V_tcp_syncache			VNET(tcp_syncache)
173 
174 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0,
175     "TCP SYN cache");
176 
177 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
178     &VNET_NAME(tcp_syncache.bucket_limit), 0,
179     "Per-bucket hash limit for syncache");
180 
181 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
182     &VNET_NAME(tcp_syncache.cache_limit), 0,
183     "Overall entry limit for syncache");
184 
185 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
186     &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
187 
188 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
189     &VNET_NAME(tcp_syncache.hashsize), 0,
190     "Size of TCP syncache hashtable");
191 
192 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_VNET | CTLFLAG_RW,
193     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
194     "Limit on SYN/ACK retransmissions");
195 
196 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
197 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
198     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
199     "Send reset on socket allocation failure");
200 
201 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
202 
203 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
204 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
205 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
206 
207 /*
208  * Requires the syncache entry to be already removed from the bucket list.
209  */
210 static void
211 syncache_free(struct syncache *sc)
212 {
213 
214 	if (sc->sc_ipopts)
215 		(void) m_free(sc->sc_ipopts);
216 	if (sc->sc_cred)
217 		crfree(sc->sc_cred);
218 #ifdef MAC
219 	mac_syncache_destroy(&sc->sc_label);
220 #endif
221 
222 	uma_zfree(V_tcp_syncache.zone, sc);
223 }
224 
225 void
226 syncache_init(void)
227 {
228 	int i;
229 
230 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
231 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
232 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
233 	V_tcp_syncache.hash_secret = arc4random();
234 
235 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
236 	    &V_tcp_syncache.hashsize);
237 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
238 	    &V_tcp_syncache.bucket_limit);
239 	if (!powerof2(V_tcp_syncache.hashsize) ||
240 	    V_tcp_syncache.hashsize == 0) {
241 		printf("WARNING: syncache hash size is not a power of 2.\n");
242 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
243 	}
244 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
245 
246 	/* Set limits. */
247 	V_tcp_syncache.cache_limit =
248 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
249 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
250 	    &V_tcp_syncache.cache_limit);
251 
252 	/* Allocate the hash table. */
253 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
254 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
255 
256 #ifdef VIMAGE
257 	V_tcp_syncache.vnet = curvnet;
258 #endif
259 
260 	/* Initialize the hash buckets. */
261 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
262 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
263 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
264 			 NULL, MTX_DEF);
265 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
266 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
267 		V_tcp_syncache.hashbase[i].sch_length = 0;
268 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
269 	}
270 
271 	/* Create the syncache entry zone. */
272 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
273 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
274 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
275 	    V_tcp_syncache.cache_limit);
276 
277 	/* Start the SYN cookie reseeder callout. */
278 	callout_init(&V_tcp_syncache.secret.reseed, 1);
279 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
280 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
281 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
282 	    syncookie_reseed, &V_tcp_syncache);
283 }
284 
285 #ifdef VIMAGE
286 void
287 syncache_destroy(void)
288 {
289 	struct syncache_head *sch;
290 	struct syncache *sc, *nsc;
291 	int i;
292 
293 	/*
294 	 * Stop the re-seed timer before freeing resources.  No need to
295 	 * possibly schedule it another time.
296 	 */
297 	callout_drain(&V_tcp_syncache.secret.reseed);
298 
299 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
300 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
301 
302 		sch = &V_tcp_syncache.hashbase[i];
303 		callout_drain(&sch->sch_timer);
304 
305 		SCH_LOCK(sch);
306 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
307 			syncache_drop(sc, sch);
308 		SCH_UNLOCK(sch);
309 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
310 		    ("%s: sch->sch_bucket not empty", __func__));
311 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
312 		    __func__, sch->sch_length));
313 		mtx_destroy(&sch->sch_mtx);
314 	}
315 
316 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
317 	    ("%s: cache_count not 0", __func__));
318 
319 	/* Free the allocated global resources. */
320 	uma_zdestroy(V_tcp_syncache.zone);
321 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
322 }
323 #endif
324 
325 /*
326  * Inserts a syncache entry into the specified bucket row.
327  * Locks and unlocks the syncache_head autonomously.
328  */
329 static void
330 syncache_insert(struct syncache *sc, struct syncache_head *sch)
331 {
332 	struct syncache *sc2;
333 
334 	SCH_LOCK(sch);
335 
336 	/*
337 	 * Make sure that we don't overflow the per-bucket limit.
338 	 * If the bucket is full, toss the oldest element.
339 	 */
340 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
341 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
342 			("sch->sch_length incorrect"));
343 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
344 		syncache_drop(sc2, sch);
345 		TCPSTAT_INC(tcps_sc_bucketoverflow);
346 	}
347 
348 	/* Put it into the bucket. */
349 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
350 	sch->sch_length++;
351 
352 #ifdef TCP_OFFLOAD
353 	if (ADDED_BY_TOE(sc)) {
354 		struct toedev *tod = sc->sc_tod;
355 
356 		tod->tod_syncache_added(tod, sc->sc_todctx);
357 	}
358 #endif
359 
360 	/* Reinitialize the bucket row's timer. */
361 	if (sch->sch_length == 1)
362 		sch->sch_nextc = ticks + INT_MAX;
363 	syncache_timeout(sc, sch, 1);
364 
365 	SCH_UNLOCK(sch);
366 
367 	TCPSTATES_INC(TCPS_SYN_RECEIVED);
368 	TCPSTAT_INC(tcps_sc_added);
369 }
370 
371 /*
372  * Remove and free entry from syncache bucket row.
373  * Expects locked syncache head.
374  */
375 static void
376 syncache_drop(struct syncache *sc, struct syncache_head *sch)
377 {
378 
379 	SCH_LOCK_ASSERT(sch);
380 
381 	TCPSTATES_DEC(TCPS_SYN_RECEIVED);
382 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
383 	sch->sch_length--;
384 
385 #ifdef TCP_OFFLOAD
386 	if (ADDED_BY_TOE(sc)) {
387 		struct toedev *tod = sc->sc_tod;
388 
389 		tod->tod_syncache_removed(tod, sc->sc_todctx);
390 	}
391 #endif
392 
393 	syncache_free(sc);
394 }
395 
396 /*
397  * Engage/reengage time on bucket row.
398  */
399 static void
400 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
401 {
402 	sc->sc_rxttime = ticks +
403 		TCPTV_RTOBASE * (tcp_syn_backoff[sc->sc_rxmits]);
404 	sc->sc_rxmits++;
405 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
406 		sch->sch_nextc = sc->sc_rxttime;
407 		if (docallout)
408 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
409 			    syncache_timer, (void *)sch);
410 	}
411 }
412 
413 /*
414  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
415  * If we have retransmitted an entry the maximum number of times, expire it.
416  * One separate timer for each bucket row.
417  */
418 static void
419 syncache_timer(void *xsch)
420 {
421 	struct syncache_head *sch = (struct syncache_head *)xsch;
422 	struct syncache *sc, *nsc;
423 	int tick = ticks;
424 	char *s;
425 
426 	CURVNET_SET(sch->sch_sc->vnet);
427 
428 	/* NB: syncache_head has already been locked by the callout. */
429 	SCH_LOCK_ASSERT(sch);
430 
431 	/*
432 	 * In the following cycle we may remove some entries and/or
433 	 * advance some timeouts, so re-initialize the bucket timer.
434 	 */
435 	sch->sch_nextc = tick + INT_MAX;
436 
437 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
438 		/*
439 		 * We do not check if the listen socket still exists
440 		 * and accept the case where the listen socket may be
441 		 * gone by the time we resend the SYN/ACK.  We do
442 		 * not expect this to happens often. If it does,
443 		 * then the RST will be sent by the time the remote
444 		 * host does the SYN/ACK->ACK.
445 		 */
446 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
447 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
448 				sch->sch_nextc = sc->sc_rxttime;
449 			continue;
450 		}
451 		if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
452 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
453 				log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
454 				    "giving up and removing syncache entry\n",
455 				    s, __func__);
456 				free(s, M_TCPLOG);
457 			}
458 			syncache_drop(sc, sch);
459 			TCPSTAT_INC(tcps_sc_stale);
460 			continue;
461 		}
462 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
463 			log(LOG_DEBUG, "%s; %s: Response timeout, "
464 			    "retransmitting (%u) SYN|ACK\n",
465 			    s, __func__, sc->sc_rxmits);
466 			free(s, M_TCPLOG);
467 		}
468 
469 		syncache_respond(sc, sch, 1, NULL);
470 		TCPSTAT_INC(tcps_sc_retransmitted);
471 		syncache_timeout(sc, sch, 0);
472 	}
473 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
474 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
475 			syncache_timer, (void *)(sch));
476 	CURVNET_RESTORE();
477 }
478 
479 /*
480  * Find an entry in the syncache.
481  * Returns always with locked syncache_head plus a matching entry or NULL.
482  */
483 static struct syncache *
484 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
485 {
486 	struct syncache *sc;
487 	struct syncache_head *sch;
488 	uint32_t hash;
489 
490 	/*
491 	 * The hash is built on foreign port + local port + foreign address.
492 	 * We rely on the fact that struct in_conninfo starts with 16 bits
493 	 * of foreign port, then 16 bits of local port then followed by 128
494 	 * bits of foreign address.  In case of IPv4 address, the first 3
495 	 * 32-bit words of the address always are zeroes.
496 	 */
497 	hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
498 	    V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
499 
500 	sch = &V_tcp_syncache.hashbase[hash];
501 	*schp = sch;
502 	SCH_LOCK(sch);
503 
504 	/* Circle through bucket row to find matching entry. */
505 	TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
506 		if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
507 		    sizeof(struct in_endpoints)) == 0)
508 			break;
509 
510 	return (sc);	/* Always returns with locked sch. */
511 }
512 
513 /*
514  * This function is called when we get a RST for a
515  * non-existent connection, so that we can see if the
516  * connection is in the syn cache.  If it is, zap it.
517  */
518 void
519 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
520 {
521 	struct syncache *sc;
522 	struct syncache_head *sch;
523 	char *s = NULL;
524 
525 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
526 	SCH_LOCK_ASSERT(sch);
527 
528 	/*
529 	 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
530 	 * See RFC 793 page 65, section SEGMENT ARRIVES.
531 	 */
532 	if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
533 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
534 			log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
535 			    "FIN flag set, segment ignored\n", s, __func__);
536 		TCPSTAT_INC(tcps_badrst);
537 		goto done;
538 	}
539 
540 	/*
541 	 * No corresponding connection was found in syncache.
542 	 * If syncookies are enabled and possibly exclusively
543 	 * used, or we are under memory pressure, a valid RST
544 	 * may not find a syncache entry.  In that case we're
545 	 * done and no SYN|ACK retransmissions will happen.
546 	 * Otherwise the RST was misdirected or spoofed.
547 	 */
548 	if (sc == NULL) {
549 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
550 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
551 			    "syncache entry (possibly syncookie only), "
552 			    "segment ignored\n", s, __func__);
553 		TCPSTAT_INC(tcps_badrst);
554 		goto done;
555 	}
556 
557 	/*
558 	 * If the RST bit is set, check the sequence number to see
559 	 * if this is a valid reset segment.
560 	 * RFC 793 page 37:
561 	 *   In all states except SYN-SENT, all reset (RST) segments
562 	 *   are validated by checking their SEQ-fields.  A reset is
563 	 *   valid if its sequence number is in the window.
564 	 *
565 	 *   The sequence number in the reset segment is normally an
566 	 *   echo of our outgoing acknowlegement numbers, but some hosts
567 	 *   send a reset with the sequence number at the rightmost edge
568 	 *   of our receive window, and we have to handle this case.
569 	 */
570 	if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
571 	    SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
572 		syncache_drop(sc, sch);
573 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
574 			log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, "
575 			    "connection attempt aborted by remote endpoint\n",
576 			    s, __func__);
577 		TCPSTAT_INC(tcps_sc_reset);
578 	} else {
579 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
580 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
581 			    "IRS %u (+WND %u), segment ignored\n",
582 			    s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd);
583 		TCPSTAT_INC(tcps_badrst);
584 	}
585 
586 done:
587 	if (s != NULL)
588 		free(s, M_TCPLOG);
589 	SCH_UNLOCK(sch);
590 }
591 
592 void
593 syncache_badack(struct in_conninfo *inc)
594 {
595 	struct syncache *sc;
596 	struct syncache_head *sch;
597 
598 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
599 	SCH_LOCK_ASSERT(sch);
600 	if (sc != NULL) {
601 		syncache_drop(sc, sch);
602 		TCPSTAT_INC(tcps_sc_badack);
603 	}
604 	SCH_UNLOCK(sch);
605 }
606 
607 void
608 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th)
609 {
610 	struct syncache *sc;
611 	struct syncache_head *sch;
612 
613 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
614 	SCH_LOCK_ASSERT(sch);
615 	if (sc == NULL)
616 		goto done;
617 
618 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
619 	if (ntohl(th->th_seq) != sc->sc_iss)
620 		goto done;
621 
622 	/*
623 	 * If we've rertransmitted 3 times and this is our second error,
624 	 * we remove the entry.  Otherwise, we allow it to continue on.
625 	 * This prevents us from incorrectly nuking an entry during a
626 	 * spurious network outage.
627 	 *
628 	 * See tcp_notify().
629 	 */
630 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
631 		sc->sc_flags |= SCF_UNREACH;
632 		goto done;
633 	}
634 	syncache_drop(sc, sch);
635 	TCPSTAT_INC(tcps_sc_unreach);
636 done:
637 	SCH_UNLOCK(sch);
638 }
639 
640 /*
641  * Build a new TCP socket structure from a syncache entry.
642  *
643  * On success return the newly created socket with its underlying inp locked.
644  */
645 static struct socket *
646 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
647 {
648 	struct tcp_function_block *blk;
649 	struct inpcb *inp = NULL;
650 	struct socket *so;
651 	struct tcpcb *tp;
652 	int error;
653 	char *s;
654 
655 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
656 
657 	/*
658 	 * Ok, create the full blown connection, and set things up
659 	 * as they would have been set up if we had created the
660 	 * connection when the SYN arrived.  If we can't create
661 	 * the connection, abort it.
662 	 */
663 	so = sonewconn(lso, 0);
664 	if (so == NULL) {
665 		/*
666 		 * Drop the connection; we will either send a RST or
667 		 * have the peer retransmit its SYN again after its
668 		 * RTO and try again.
669 		 */
670 		TCPSTAT_INC(tcps_listendrop);
671 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
672 			log(LOG_DEBUG, "%s; %s: Socket create failed "
673 			    "due to limits or memory shortage\n",
674 			    s, __func__);
675 			free(s, M_TCPLOG);
676 		}
677 		goto abort2;
678 	}
679 #ifdef MAC
680 	mac_socketpeer_set_from_mbuf(m, so);
681 #endif
682 
683 	inp = sotoinpcb(so);
684 	inp->inp_inc.inc_fibnum = so->so_fibnum;
685 	INP_WLOCK(inp);
686 	/*
687 	 * Exclusive pcbinfo lock is not required in syncache socket case even
688 	 * if two inpcb locks can be acquired simultaneously:
689 	 *  - the inpcb in LISTEN state,
690 	 *  - the newly created inp.
691 	 *
692 	 * In this case, an inp cannot be at same time in LISTEN state and
693 	 * just created by an accept() call.
694 	 */
695 	INP_HASH_WLOCK(&V_tcbinfo);
696 
697 	/* Insert new socket into PCB hash list. */
698 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
699 #ifdef INET6
700 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
701 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
702 	} else {
703 		inp->inp_vflag &= ~INP_IPV6;
704 		inp->inp_vflag |= INP_IPV4;
705 #endif
706 		inp->inp_laddr = sc->sc_inc.inc_laddr;
707 #ifdef INET6
708 	}
709 #endif
710 
711 	/*
712 	 * If there's an mbuf and it has a flowid, then let's initialise the
713 	 * inp with that particular flowid.
714 	 */
715 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
716 		inp->inp_flowid = m->m_pkthdr.flowid;
717 		inp->inp_flowtype = M_HASHTYPE_GET(m);
718 	}
719 
720 	/*
721 	 * Install in the reservation hash table for now, but don't yet
722 	 * install a connection group since the full 4-tuple isn't yet
723 	 * configured.
724 	 */
725 	inp->inp_lport = sc->sc_inc.inc_lport;
726 	if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
727 		/*
728 		 * Undo the assignments above if we failed to
729 		 * put the PCB on the hash lists.
730 		 */
731 #ifdef INET6
732 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
733 			inp->in6p_laddr = in6addr_any;
734 		else
735 #endif
736 			inp->inp_laddr.s_addr = INADDR_ANY;
737 		inp->inp_lport = 0;
738 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
739 			log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
740 			    "with error %i\n",
741 			    s, __func__, error);
742 			free(s, M_TCPLOG);
743 		}
744 		INP_HASH_WUNLOCK(&V_tcbinfo);
745 		goto abort;
746 	}
747 #ifdef IPSEC
748 	/* Copy old policy into new socket's. */
749 	if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp))
750 		printf("syncache_socket: could not copy policy\n");
751 #endif
752 #ifdef INET6
753 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
754 		struct inpcb *oinp = sotoinpcb(lso);
755 		struct in6_addr laddr6;
756 		struct sockaddr_in6 sin6;
757 		/*
758 		 * Inherit socket options from the listening socket.
759 		 * Note that in6p_inputopts are not (and should not be)
760 		 * copied, since it stores previously received options and is
761 		 * used to detect if each new option is different than the
762 		 * previous one and hence should be passed to a user.
763 		 * If we copied in6p_inputopts, a user would not be able to
764 		 * receive options just after calling the accept system call.
765 		 */
766 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
767 		if (oinp->in6p_outputopts)
768 			inp->in6p_outputopts =
769 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
770 
771 		sin6.sin6_family = AF_INET6;
772 		sin6.sin6_len = sizeof(sin6);
773 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
774 		sin6.sin6_port = sc->sc_inc.inc_fport;
775 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
776 		laddr6 = inp->in6p_laddr;
777 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
778 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
779 		if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
780 		    thread0.td_ucred, m)) != 0) {
781 			inp->in6p_laddr = laddr6;
782 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
783 				log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
784 				    "with error %i\n",
785 				    s, __func__, error);
786 				free(s, M_TCPLOG);
787 			}
788 			INP_HASH_WUNLOCK(&V_tcbinfo);
789 			goto abort;
790 		}
791 		/* Override flowlabel from in6_pcbconnect. */
792 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
793 		inp->inp_flow |= sc->sc_flowlabel;
794 	}
795 #endif /* INET6 */
796 #if defined(INET) && defined(INET6)
797 	else
798 #endif
799 #ifdef INET
800 	{
801 		struct in_addr laddr;
802 		struct sockaddr_in sin;
803 
804 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
805 
806 		if (inp->inp_options == NULL) {
807 			inp->inp_options = sc->sc_ipopts;
808 			sc->sc_ipopts = NULL;
809 		}
810 
811 		sin.sin_family = AF_INET;
812 		sin.sin_len = sizeof(sin);
813 		sin.sin_addr = sc->sc_inc.inc_faddr;
814 		sin.sin_port = sc->sc_inc.inc_fport;
815 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
816 		laddr = inp->inp_laddr;
817 		if (inp->inp_laddr.s_addr == INADDR_ANY)
818 			inp->inp_laddr = sc->sc_inc.inc_laddr;
819 		if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
820 		    thread0.td_ucred, m)) != 0) {
821 			inp->inp_laddr = laddr;
822 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
823 				log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
824 				    "with error %i\n",
825 				    s, __func__, error);
826 				free(s, M_TCPLOG);
827 			}
828 			INP_HASH_WUNLOCK(&V_tcbinfo);
829 			goto abort;
830 		}
831 	}
832 #endif /* INET */
833 	INP_HASH_WUNLOCK(&V_tcbinfo);
834 	tp = intotcpcb(inp);
835 	tcp_state_change(tp, TCPS_SYN_RECEIVED);
836 	tp->iss = sc->sc_iss;
837 	tp->irs = sc->sc_irs;
838 	tcp_rcvseqinit(tp);
839 	tcp_sendseqinit(tp);
840 	blk = sototcpcb(lso)->t_fb;
841 	if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
842 		/*
843 		 * Our parents t_fb was not the default,
844 		 * we need to release our ref on tp->t_fb and
845 		 * pickup one on the new entry.
846 		 */
847 		struct tcp_function_block *rblk;
848 
849 		rblk = find_and_ref_tcp_fb(blk);
850 		KASSERT(rblk != NULL,
851 		    ("cannot find blk %p out of syncache?", blk));
852 		if (tp->t_fb->tfb_tcp_fb_fini)
853 			(*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
854 		refcount_release(&tp->t_fb->tfb_refcnt);
855 		tp->t_fb = rblk;
856 		if (tp->t_fb->tfb_tcp_fb_init) {
857 			(*tp->t_fb->tfb_tcp_fb_init)(tp);
858 		}
859 	}
860 	tp->snd_wl1 = sc->sc_irs;
861 	tp->snd_max = tp->iss + 1;
862 	tp->snd_nxt = tp->iss + 1;
863 	tp->rcv_up = sc->sc_irs + 1;
864 	tp->rcv_wnd = sc->sc_wnd;
865 	tp->rcv_adv += tp->rcv_wnd;
866 	tp->last_ack_sent = tp->rcv_nxt;
867 
868 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
869 	if (sc->sc_flags & SCF_NOOPT)
870 		tp->t_flags |= TF_NOOPT;
871 	else {
872 		if (sc->sc_flags & SCF_WINSCALE) {
873 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
874 			tp->snd_scale = sc->sc_requested_s_scale;
875 			tp->request_r_scale = sc->sc_requested_r_scale;
876 		}
877 		if (sc->sc_flags & SCF_TIMESTAMP) {
878 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
879 			tp->ts_recent = sc->sc_tsreflect;
880 			tp->ts_recent_age = tcp_ts_getticks();
881 			tp->ts_offset = sc->sc_tsoff;
882 		}
883 #ifdef TCP_SIGNATURE
884 		if (sc->sc_flags & SCF_SIGNATURE)
885 			tp->t_flags |= TF_SIGNATURE;
886 #endif
887 		if (sc->sc_flags & SCF_SACK)
888 			tp->t_flags |= TF_SACK_PERMIT;
889 	}
890 
891 	if (sc->sc_flags & SCF_ECN)
892 		tp->t_flags |= TF_ECN_PERMIT;
893 
894 	/*
895 	 * Set up MSS and get cached values from tcp_hostcache.
896 	 * This might overwrite some of the defaults we just set.
897 	 */
898 	tcp_mss(tp, sc->sc_peer_mss);
899 
900 	/*
901 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
902 	 * limited to one segment in cc_conn_init().
903 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
904 	 */
905 	if (sc->sc_rxmits > 1)
906 		tp->snd_cwnd = 1;
907 
908 #ifdef TCP_OFFLOAD
909 	/*
910 	 * Allow a TOE driver to install its hooks.  Note that we hold the
911 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
912 	 * new connection before the TOE driver has done its thing.
913 	 */
914 	if (ADDED_BY_TOE(sc)) {
915 		struct toedev *tod = sc->sc_tod;
916 
917 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
918 	}
919 #endif
920 	/*
921 	 * Copy and activate timers.
922 	 */
923 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
924 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
925 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
926 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
927 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
928 
929 	TCPSTAT_INC(tcps_accepts);
930 	return (so);
931 
932 abort:
933 	INP_WUNLOCK(inp);
934 abort2:
935 	if (so != NULL)
936 		soabort(so);
937 	return (NULL);
938 }
939 
940 /*
941  * This function gets called when we receive an ACK for a
942  * socket in the LISTEN state.  We look up the connection
943  * in the syncache, and if its there, we pull it out of
944  * the cache and turn it into a full-blown connection in
945  * the SYN-RECEIVED state.
946  *
947  * On syncache_socket() success the newly created socket
948  * has its underlying inp locked.
949  */
950 int
951 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
952     struct socket **lsop, struct mbuf *m)
953 {
954 	struct syncache *sc;
955 	struct syncache_head *sch;
956 	struct syncache scs;
957 	char *s;
958 
959 	/*
960 	 * Global TCP locks are held because we manipulate the PCB lists
961 	 * and create a new socket.
962 	 */
963 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
964 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
965 	    ("%s: can handle only ACK", __func__));
966 
967 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
968 	SCH_LOCK_ASSERT(sch);
969 
970 #ifdef INVARIANTS
971 	/*
972 	 * Test code for syncookies comparing the syncache stored
973 	 * values with the reconstructed values from the cookie.
974 	 */
975 	if (sc != NULL)
976 		syncookie_cmp(inc, sch, sc, th, to, *lsop);
977 #endif
978 
979 	if (sc == NULL) {
980 		/*
981 		 * There is no syncache entry, so see if this ACK is
982 		 * a returning syncookie.  To do this, first:
983 		 *  A. See if this socket has had a syncache entry dropped in
984 		 *     the past.  We don't want to accept a bogus syncookie
985 		 *     if we've never received a SYN.
986 		 *  B. check that the syncookie is valid.  If it is, then
987 		 *     cobble up a fake syncache entry, and return.
988 		 */
989 		if (!V_tcp_syncookies) {
990 			SCH_UNLOCK(sch);
991 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
992 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
993 				    "segment rejected (syncookies disabled)\n",
994 				    s, __func__);
995 			goto failed;
996 		}
997 		bzero(&scs, sizeof(scs));
998 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
999 		SCH_UNLOCK(sch);
1000 		if (sc == NULL) {
1001 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1002 				log(LOG_DEBUG, "%s; %s: Segment failed "
1003 				    "SYNCOOKIE authentication, segment rejected "
1004 				    "(probably spoofed)\n", s, __func__);
1005 			goto failed;
1006 		}
1007 	} else {
1008 		/*
1009 		 * Pull out the entry to unlock the bucket row.
1010 		 *
1011 		 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1012 		 * tcp_state_change().  The tcpcb is not existent at this
1013 		 * moment.  A new one will be allocated via syncache_socket->
1014 		 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1015 		 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1016 		 */
1017 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1018 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1019 		sch->sch_length--;
1020 #ifdef TCP_OFFLOAD
1021 		if (ADDED_BY_TOE(sc)) {
1022 			struct toedev *tod = sc->sc_tod;
1023 
1024 			tod->tod_syncache_removed(tod, sc->sc_todctx);
1025 		}
1026 #endif
1027 		SCH_UNLOCK(sch);
1028 	}
1029 
1030 	/*
1031 	 * Segment validation:
1032 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1033 	 */
1034 	if (th->th_ack != sc->sc_iss + 1) {
1035 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1036 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1037 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1038 		goto failed;
1039 	}
1040 
1041 	/*
1042 	 * The SEQ must fall in the window starting at the received
1043 	 * initial receive sequence number + 1 (the SYN).
1044 	 */
1045 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1046 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1047 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1048 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1049 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1050 		goto failed;
1051 	}
1052 
1053 	/*
1054 	 * If timestamps were not negotiated during SYN/ACK they
1055 	 * must not appear on any segment during this session.
1056 	 */
1057 	if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1058 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1059 			log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1060 			    "segment rejected\n", s, __func__);
1061 		goto failed;
1062 	}
1063 
1064 	/*
1065 	 * If timestamps were negotiated during SYN/ACK they should
1066 	 * appear on every segment during this session.
1067 	 * XXXAO: This is only informal as there have been unverified
1068 	 * reports of non-compliants stacks.
1069 	 */
1070 	if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1071 		if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1072 			log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1073 			    "no action\n", s, __func__);
1074 			free(s, M_TCPLOG);
1075 			s = NULL;
1076 		}
1077 	}
1078 
1079 	/*
1080 	 * If timestamps were negotiated, the reflected timestamp
1081 	 * must be equal to what we actually sent in the SYN|ACK
1082 	 * except in the case of 0. Some boxes are known for sending
1083 	 * broken timestamp replies during the 3whs (and potentially
1084 	 * during the connection also).
1085 	 *
1086 	 * Accept the final ACK of 3whs with reflected timestamp of 0
1087 	 * instead of sending a RST and deleting the syncache entry.
1088 	 */
1089 	if ((to->to_flags & TOF_TS) && to->to_tsecr &&
1090 	    to->to_tsecr != sc->sc_ts) {
1091 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1092 			log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, "
1093 			    "segment rejected\n",
1094 			    s, __func__, to->to_tsecr, sc->sc_ts);
1095 		goto failed;
1096 	}
1097 
1098 	*lsop = syncache_socket(sc, *lsop, m);
1099 
1100 	if (*lsop == NULL)
1101 		TCPSTAT_INC(tcps_sc_aborted);
1102 	else
1103 		TCPSTAT_INC(tcps_sc_completed);
1104 
1105 /* how do we find the inp for the new socket? */
1106 	if (sc != &scs)
1107 		syncache_free(sc);
1108 	return (1);
1109 failed:
1110 	if (sc != NULL && sc != &scs)
1111 		syncache_free(sc);
1112 	if (s != NULL)
1113 		free(s, M_TCPLOG);
1114 	*lsop = NULL;
1115 	return (0);
1116 }
1117 
1118 #ifdef TCP_RFC7413
1119 static void
1120 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
1121     uint64_t response_cookie)
1122 {
1123 	struct inpcb *inp;
1124 	struct tcpcb *tp;
1125 	unsigned int *pending_counter;
1126 
1127 	/*
1128 	 * Global TCP locks are held because we manipulate the PCB lists
1129 	 * and create a new socket.
1130 	 */
1131 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1132 
1133 	pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
1134 	*lsop = syncache_socket(sc, *lsop, m);
1135 	if (*lsop == NULL) {
1136 		TCPSTAT_INC(tcps_sc_aborted);
1137 		atomic_subtract_int(pending_counter, 1);
1138 	} else {
1139 		inp = sotoinpcb(*lsop);
1140 		tp = intotcpcb(inp);
1141 		tp->t_flags |= TF_FASTOPEN;
1142 		tp->t_tfo_cookie = response_cookie;
1143 		tp->snd_max = tp->iss;
1144 		tp->snd_nxt = tp->iss;
1145 		tp->t_tfo_pending = pending_counter;
1146 		TCPSTAT_INC(tcps_sc_completed);
1147 	}
1148 }
1149 #endif /* TCP_RFC7413 */
1150 
1151 /*
1152  * Given a LISTEN socket and an inbound SYN request, add
1153  * this to the syn cache, and send back a segment:
1154  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1155  * to the source.
1156  *
1157  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1158  * Doing so would require that we hold onto the data and deliver it
1159  * to the application.  However, if we are the target of a SYN-flood
1160  * DoS attack, an attacker could send data which would eventually
1161  * consume all available buffer space if it were ACKed.  By not ACKing
1162  * the data, we avoid this DoS scenario.
1163  *
1164  * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1165  * cookie is processed and a new socket is created.  In this case, any data
1166  * accompanying the SYN will be queued to the socket by tcp_input() and will
1167  * be ACKed either when the application sends response data or the delayed
1168  * ACK timer expires, whichever comes first.
1169  */
1170 int
1171 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1172     struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1173     void *todctx)
1174 {
1175 	struct tcpcb *tp;
1176 	struct socket *so;
1177 	struct syncache *sc = NULL;
1178 	struct syncache_head *sch;
1179 	struct mbuf *ipopts = NULL;
1180 	u_int ltflags;
1181 	int win, ip_ttl, ip_tos;
1182 	char *s;
1183 	int rv = 0;
1184 #ifdef INET6
1185 	int autoflowlabel = 0;
1186 #endif
1187 #ifdef MAC
1188 	struct label *maclabel;
1189 #endif
1190 	struct syncache scs;
1191 	struct ucred *cred;
1192 #ifdef TCP_RFC7413
1193 	uint64_t tfo_response_cookie;
1194 	unsigned int *tfo_pending = NULL;
1195 	int tfo_cookie_valid = 0;
1196 	int tfo_response_cookie_valid = 0;
1197 #endif
1198 
1199 	INP_WLOCK_ASSERT(inp);			/* listen socket */
1200 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1201 	    ("%s: unexpected tcp flags", __func__));
1202 
1203 	/*
1204 	 * Combine all so/tp operations very early to drop the INP lock as
1205 	 * soon as possible.
1206 	 */
1207 	so = *lsop;
1208 	tp = sototcpcb(so);
1209 	cred = crhold(so->so_cred);
1210 
1211 #ifdef INET6
1212 	if ((inc->inc_flags & INC_ISIPV6) &&
1213 	    (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1214 		autoflowlabel = 1;
1215 #endif
1216 	ip_ttl = inp->inp_ip_ttl;
1217 	ip_tos = inp->inp_ip_tos;
1218 	win = sbspace(&so->so_rcv);
1219 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1220 
1221 #ifdef TCP_RFC7413
1222 	if (V_tcp_fastopen_enabled && IS_FASTOPEN(tp->t_flags) &&
1223 	    (tp->t_tfo_pending != NULL) && (to->to_flags & TOF_FASTOPEN)) {
1224 		/*
1225 		 * Limit the number of pending TFO connections to
1226 		 * approximately half of the queue limit.  This prevents TFO
1227 		 * SYN floods from starving the service by filling the
1228 		 * listen queue with bogus TFO connections.
1229 		 */
1230 		if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1231 		    (so->so_qlimit / 2)) {
1232 			int result;
1233 
1234 			result = tcp_fastopen_check_cookie(inc,
1235 			    to->to_tfo_cookie, to->to_tfo_len,
1236 			    &tfo_response_cookie);
1237 			tfo_cookie_valid = (result > 0);
1238 			tfo_response_cookie_valid = (result >= 0);
1239 		}
1240 
1241 		/*
1242 		 * Remember the TFO pending counter as it will have to be
1243 		 * decremented below if we don't make it to syncache_tfo_expand().
1244 		 */
1245 		tfo_pending = tp->t_tfo_pending;
1246 	}
1247 #endif
1248 
1249 	/* By the time we drop the lock these should no longer be used. */
1250 	so = NULL;
1251 	tp = NULL;
1252 
1253 #ifdef MAC
1254 	if (mac_syncache_init(&maclabel) != 0) {
1255 		INP_WUNLOCK(inp);
1256 		goto done;
1257 	} else
1258 		mac_syncache_create(maclabel, inp);
1259 #endif
1260 #ifdef TCP_RFC7413
1261 	if (!tfo_cookie_valid)
1262 #endif
1263 		INP_WUNLOCK(inp);
1264 
1265 	/*
1266 	 * Remember the IP options, if any.
1267 	 */
1268 #ifdef INET6
1269 	if (!(inc->inc_flags & INC_ISIPV6))
1270 #endif
1271 #ifdef INET
1272 		ipopts = (m) ? ip_srcroute(m) : NULL;
1273 #else
1274 		ipopts = NULL;
1275 #endif
1276 
1277 	/*
1278 	 * See if we already have an entry for this connection.
1279 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1280 	 *
1281 	 * XXX: should the syncache be re-initialized with the contents
1282 	 * of the new SYN here (which may have different options?)
1283 	 *
1284 	 * XXX: We do not check the sequence number to see if this is a
1285 	 * real retransmit or a new connection attempt.  The question is
1286 	 * how to handle such a case; either ignore it as spoofed, or
1287 	 * drop the current entry and create a new one?
1288 	 */
1289 	sc = syncache_lookup(inc, &sch);	/* returns locked entry */
1290 	SCH_LOCK_ASSERT(sch);
1291 	if (sc != NULL) {
1292 #ifdef TCP_RFC7413
1293 		if (tfo_cookie_valid)
1294 			INP_WUNLOCK(inp);
1295 #endif
1296 		TCPSTAT_INC(tcps_sc_dupsyn);
1297 		if (ipopts) {
1298 			/*
1299 			 * If we were remembering a previous source route,
1300 			 * forget it and use the new one we've been given.
1301 			 */
1302 			if (sc->sc_ipopts)
1303 				(void) m_free(sc->sc_ipopts);
1304 			sc->sc_ipopts = ipopts;
1305 		}
1306 		/*
1307 		 * Update timestamp if present.
1308 		 */
1309 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1310 			sc->sc_tsreflect = to->to_tsval;
1311 		else
1312 			sc->sc_flags &= ~SCF_TIMESTAMP;
1313 #ifdef MAC
1314 		/*
1315 		 * Since we have already unconditionally allocated label
1316 		 * storage, free it up.  The syncache entry will already
1317 		 * have an initialized label we can use.
1318 		 */
1319 		mac_syncache_destroy(&maclabel);
1320 #endif
1321 		/* Retransmit SYN|ACK and reset retransmit count. */
1322 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1323 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1324 			    "resetting timer and retransmitting SYN|ACK\n",
1325 			    s, __func__);
1326 			free(s, M_TCPLOG);
1327 		}
1328 		if (syncache_respond(sc, sch, 1, m) == 0) {
1329 			sc->sc_rxmits = 0;
1330 			syncache_timeout(sc, sch, 1);
1331 			TCPSTAT_INC(tcps_sndacks);
1332 			TCPSTAT_INC(tcps_sndtotal);
1333 		}
1334 		SCH_UNLOCK(sch);
1335 		goto done;
1336 	}
1337 
1338 #ifdef TCP_RFC7413
1339 	if (tfo_cookie_valid) {
1340 		bzero(&scs, sizeof(scs));
1341 		sc = &scs;
1342 		goto skip_alloc;
1343 	}
1344 #endif
1345 
1346 	sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1347 	if (sc == NULL) {
1348 		/*
1349 		 * The zone allocator couldn't provide more entries.
1350 		 * Treat this as if the cache was full; drop the oldest
1351 		 * entry and insert the new one.
1352 		 */
1353 		TCPSTAT_INC(tcps_sc_zonefail);
1354 		if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL)
1355 			syncache_drop(sc, sch);
1356 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1357 		if (sc == NULL) {
1358 			if (V_tcp_syncookies) {
1359 				bzero(&scs, sizeof(scs));
1360 				sc = &scs;
1361 			} else {
1362 				SCH_UNLOCK(sch);
1363 				if (ipopts)
1364 					(void) m_free(ipopts);
1365 				goto done;
1366 			}
1367 		}
1368 	}
1369 
1370 #ifdef TCP_RFC7413
1371 skip_alloc:
1372 	if (!tfo_cookie_valid && tfo_response_cookie_valid)
1373 		sc->sc_tfo_cookie = &tfo_response_cookie;
1374 #endif
1375 
1376 	/*
1377 	 * Fill in the syncache values.
1378 	 */
1379 #ifdef MAC
1380 	sc->sc_label = maclabel;
1381 #endif
1382 	sc->sc_cred = cred;
1383 	cred = NULL;
1384 	sc->sc_ipopts = ipopts;
1385 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1386 #ifdef INET6
1387 	if (!(inc->inc_flags & INC_ISIPV6))
1388 #endif
1389 	{
1390 		sc->sc_ip_tos = ip_tos;
1391 		sc->sc_ip_ttl = ip_ttl;
1392 	}
1393 #ifdef TCP_OFFLOAD
1394 	sc->sc_tod = tod;
1395 	sc->sc_todctx = todctx;
1396 #endif
1397 	sc->sc_irs = th->th_seq;
1398 	sc->sc_iss = arc4random();
1399 	sc->sc_flags = 0;
1400 	sc->sc_flowlabel = 0;
1401 
1402 	/*
1403 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1404 	 * win was derived from socket earlier in the function.
1405 	 */
1406 	win = imax(win, 0);
1407 	win = imin(win, TCP_MAXWIN);
1408 	sc->sc_wnd = win;
1409 
1410 	if (V_tcp_do_rfc1323) {
1411 		/*
1412 		 * A timestamp received in a SYN makes
1413 		 * it ok to send timestamp requests and replies.
1414 		 */
1415 		if (to->to_flags & TOF_TS) {
1416 			sc->sc_tsreflect = to->to_tsval;
1417 			sc->sc_ts = tcp_ts_getticks();
1418 			sc->sc_flags |= SCF_TIMESTAMP;
1419 		}
1420 		if (to->to_flags & TOF_SCALE) {
1421 			int wscale = 0;
1422 
1423 			/*
1424 			 * Pick the smallest possible scaling factor that
1425 			 * will still allow us to scale up to sb_max, aka
1426 			 * kern.ipc.maxsockbuf.
1427 			 *
1428 			 * We do this because there are broken firewalls that
1429 			 * will corrupt the window scale option, leading to
1430 			 * the other endpoint believing that our advertised
1431 			 * window is unscaled.  At scale factors larger than
1432 			 * 5 the unscaled window will drop below 1500 bytes,
1433 			 * leading to serious problems when traversing these
1434 			 * broken firewalls.
1435 			 *
1436 			 * With the default maxsockbuf of 256K, a scale factor
1437 			 * of 3 will be chosen by this algorithm.  Those who
1438 			 * choose a larger maxsockbuf should watch out
1439 			 * for the compatibility problems mentioned above.
1440 			 *
1441 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1442 			 * or <SYN,ACK>) segment itself is never scaled.
1443 			 */
1444 			while (wscale < TCP_MAX_WINSHIFT &&
1445 			    (TCP_MAXWIN << wscale) < sb_max)
1446 				wscale++;
1447 			sc->sc_requested_r_scale = wscale;
1448 			sc->sc_requested_s_scale = to->to_wscale;
1449 			sc->sc_flags |= SCF_WINSCALE;
1450 		}
1451 	}
1452 #ifdef TCP_SIGNATURE
1453 	/*
1454 	 * If listening socket requested TCP digests, OR received SYN
1455 	 * contains the option, flag this in the syncache so that
1456 	 * syncache_respond() will do the right thing with the SYN+ACK.
1457 	 */
1458 	if (to->to_flags & TOF_SIGNATURE || ltflags & TF_SIGNATURE)
1459 		sc->sc_flags |= SCF_SIGNATURE;
1460 #endif
1461 	if (to->to_flags & TOF_SACKPERM)
1462 		sc->sc_flags |= SCF_SACK;
1463 	if (to->to_flags & TOF_MSS)
1464 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1465 	if (ltflags & TF_NOOPT)
1466 		sc->sc_flags |= SCF_NOOPT;
1467 	if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1468 		sc->sc_flags |= SCF_ECN;
1469 
1470 	if (V_tcp_syncookies)
1471 		sc->sc_iss = syncookie_generate(sch, sc);
1472 #ifdef INET6
1473 	if (autoflowlabel) {
1474 		if (V_tcp_syncookies)
1475 			sc->sc_flowlabel = sc->sc_iss;
1476 		else
1477 			sc->sc_flowlabel = ip6_randomflowlabel();
1478 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1479 	}
1480 #endif
1481 	SCH_UNLOCK(sch);
1482 
1483 #ifdef TCP_RFC7413
1484 	if (tfo_cookie_valid) {
1485 		syncache_tfo_expand(sc, lsop, m, tfo_response_cookie);
1486 		/* INP_WUNLOCK(inp) will be performed by the caller */
1487 		rv = 1;
1488 		goto tfo_expanded;
1489 	}
1490 #endif
1491 
1492 	/*
1493 	 * Do a standard 3-way handshake.
1494 	 */
1495 	if (syncache_respond(sc, sch, 0, m) == 0) {
1496 		if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1497 			syncache_free(sc);
1498 		else if (sc != &scs)
1499 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1500 		TCPSTAT_INC(tcps_sndacks);
1501 		TCPSTAT_INC(tcps_sndtotal);
1502 	} else {
1503 		if (sc != &scs)
1504 			syncache_free(sc);
1505 		TCPSTAT_INC(tcps_sc_dropped);
1506 	}
1507 
1508 done:
1509 	if (m) {
1510 		*lsop = NULL;
1511 		m_freem(m);
1512 	}
1513 #ifdef TCP_RFC7413
1514 	/*
1515 	 * If tfo_pending is not NULL here, then a TFO SYN that did not
1516 	 * result in a new socket was processed and the associated pending
1517 	 * counter has not yet been decremented.  All such TFO processing paths
1518 	 * transit this point.
1519 	 */
1520 	if (tfo_pending != NULL)
1521 		tcp_fastopen_decrement_counter(tfo_pending);
1522 
1523 tfo_expanded:
1524 #endif
1525 	if (cred != NULL)
1526 		crfree(cred);
1527 #ifdef MAC
1528 	if (sc == &scs)
1529 		mac_syncache_destroy(&maclabel);
1530 #endif
1531 	return (rv);
1532 }
1533 
1534 /*
1535  * Send SYN|ACK to the peer.  Either in response to the peer's SYN,
1536  * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1537  */
1538 static int
1539 syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked,
1540     const struct mbuf *m0)
1541 {
1542 	struct ip *ip = NULL;
1543 	struct mbuf *m;
1544 	struct tcphdr *th = NULL;
1545 	int optlen, error = 0;	/* Make compiler happy */
1546 	u_int16_t hlen, tlen, mssopt;
1547 	struct tcpopt to;
1548 #ifdef INET6
1549 	struct ip6_hdr *ip6 = NULL;
1550 #endif
1551 #ifdef TCP_SIGNATURE
1552 	struct secasvar *sav;
1553 #endif
1554 
1555 	hlen =
1556 #ifdef INET6
1557 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1558 #endif
1559 		sizeof(struct ip);
1560 	tlen = hlen + sizeof(struct tcphdr);
1561 
1562 	/* Determine MSS we advertize to other end of connection. */
1563 	mssopt = tcp_mssopt(&sc->sc_inc);
1564 	if (sc->sc_peer_mss)
1565 		mssopt = max( min(sc->sc_peer_mss, mssopt), V_tcp_minmss);
1566 
1567 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1568 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1569 	    ("syncache: mbuf too small"));
1570 
1571 	/* Create the IP+TCP header from scratch. */
1572 	m = m_gethdr(M_NOWAIT, MT_DATA);
1573 	if (m == NULL)
1574 		return (ENOBUFS);
1575 #ifdef MAC
1576 	mac_syncache_create_mbuf(sc->sc_label, m);
1577 #endif
1578 	m->m_data += max_linkhdr;
1579 	m->m_len = tlen;
1580 	m->m_pkthdr.len = tlen;
1581 	m->m_pkthdr.rcvif = NULL;
1582 
1583 #ifdef INET6
1584 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1585 		ip6 = mtod(m, struct ip6_hdr *);
1586 		ip6->ip6_vfc = IPV6_VERSION;
1587 		ip6->ip6_nxt = IPPROTO_TCP;
1588 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1589 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1590 		ip6->ip6_plen = htons(tlen - hlen);
1591 		/* ip6_hlim is set after checksum */
1592 		ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1593 		ip6->ip6_flow |= sc->sc_flowlabel;
1594 
1595 		th = (struct tcphdr *)(ip6 + 1);
1596 	}
1597 #endif
1598 #if defined(INET6) && defined(INET)
1599 	else
1600 #endif
1601 #ifdef INET
1602 	{
1603 		ip = mtod(m, struct ip *);
1604 		ip->ip_v = IPVERSION;
1605 		ip->ip_hl = sizeof(struct ip) >> 2;
1606 		ip->ip_len = htons(tlen);
1607 		ip->ip_id = 0;
1608 		ip->ip_off = 0;
1609 		ip->ip_sum = 0;
1610 		ip->ip_p = IPPROTO_TCP;
1611 		ip->ip_src = sc->sc_inc.inc_laddr;
1612 		ip->ip_dst = sc->sc_inc.inc_faddr;
1613 		ip->ip_ttl = sc->sc_ip_ttl;
1614 		ip->ip_tos = sc->sc_ip_tos;
1615 
1616 		/*
1617 		 * See if we should do MTU discovery.  Route lookups are
1618 		 * expensive, so we will only unset the DF bit if:
1619 		 *
1620 		 *	1) path_mtu_discovery is disabled
1621 		 *	2) the SCF_UNREACH flag has been set
1622 		 */
1623 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1624 		       ip->ip_off |= htons(IP_DF);
1625 
1626 		th = (struct tcphdr *)(ip + 1);
1627 	}
1628 #endif /* INET */
1629 	th->th_sport = sc->sc_inc.inc_lport;
1630 	th->th_dport = sc->sc_inc.inc_fport;
1631 
1632 	th->th_seq = htonl(sc->sc_iss);
1633 	th->th_ack = htonl(sc->sc_irs + 1);
1634 	th->th_off = sizeof(struct tcphdr) >> 2;
1635 	th->th_x2 = 0;
1636 	th->th_flags = TH_SYN|TH_ACK;
1637 	th->th_win = htons(sc->sc_wnd);
1638 	th->th_urp = 0;
1639 
1640 	if (sc->sc_flags & SCF_ECN) {
1641 		th->th_flags |= TH_ECE;
1642 		TCPSTAT_INC(tcps_ecn_shs);
1643 	}
1644 
1645 	/* Tack on the TCP options. */
1646 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1647 		to.to_flags = 0;
1648 
1649 		to.to_mss = mssopt;
1650 		to.to_flags = TOF_MSS;
1651 		if (sc->sc_flags & SCF_WINSCALE) {
1652 			to.to_wscale = sc->sc_requested_r_scale;
1653 			to.to_flags |= TOF_SCALE;
1654 		}
1655 		if (sc->sc_flags & SCF_TIMESTAMP) {
1656 			/* Virgin timestamp or TCP cookie enhanced one. */
1657 			to.to_tsval = sc->sc_ts;
1658 			to.to_tsecr = sc->sc_tsreflect;
1659 			to.to_flags |= TOF_TS;
1660 		}
1661 		if (sc->sc_flags & SCF_SACK)
1662 			to.to_flags |= TOF_SACKPERM;
1663 #ifdef TCP_SIGNATURE
1664 		sav = NULL;
1665 		if (sc->sc_flags & SCF_SIGNATURE) {
1666 			sav = tcp_get_sav(m, IPSEC_DIR_OUTBOUND);
1667 			if (sav != NULL)
1668 				to.to_flags |= TOF_SIGNATURE;
1669 			else {
1670 
1671 				/*
1672 				 * We've got SCF_SIGNATURE flag
1673 				 * inherited from listening socket,
1674 				 * but no SADB key for given source
1675 				 * address. Assume signature is not
1676 				 * required and remove signature flag
1677 				 * instead of silently dropping
1678 				 * connection.
1679 				 */
1680 				if (locked == 0)
1681 					SCH_LOCK(sch);
1682 				sc->sc_flags &= ~SCF_SIGNATURE;
1683 				if (locked == 0)
1684 					SCH_UNLOCK(sch);
1685 			}
1686 		}
1687 #endif
1688 
1689 #ifdef TCP_RFC7413
1690 		if (sc->sc_tfo_cookie) {
1691 			to.to_flags |= TOF_FASTOPEN;
1692 			to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1693 			to.to_tfo_cookie = sc->sc_tfo_cookie;
1694 			/* don't send cookie again when retransmitting response */
1695 			sc->sc_tfo_cookie = NULL;
1696 		}
1697 #endif
1698 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1699 
1700 		/* Adjust headers by option size. */
1701 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1702 		m->m_len += optlen;
1703 		m->m_pkthdr.len += optlen;
1704 
1705 #ifdef TCP_SIGNATURE
1706 		if (sc->sc_flags & SCF_SIGNATURE)
1707 			tcp_signature_do_compute(m, 0, optlen,
1708 			    to.to_signature, sav);
1709 #endif
1710 #ifdef INET6
1711 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1712 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1713 		else
1714 #endif
1715 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1716 	} else
1717 		optlen = 0;
1718 
1719 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
1720 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1721 	/*
1722 	 * If we have peer's SYN and it has a flowid, then let's assign it to
1723 	 * our SYN|ACK.  ip6_output() and ip_output() will not assign flowid
1724 	 * to SYN|ACK due to lack of inp here.
1725 	 */
1726 	if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1727 		m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1728 		M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1729 	}
1730 #ifdef INET6
1731 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1732 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1733 		th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1734 		    IPPROTO_TCP, 0);
1735 		ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1736 #ifdef TCP_OFFLOAD
1737 		if (ADDED_BY_TOE(sc)) {
1738 			struct toedev *tod = sc->sc_tod;
1739 
1740 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1741 
1742 			return (error);
1743 		}
1744 #endif
1745 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1746 	}
1747 #endif
1748 #if defined(INET6) && defined(INET)
1749 	else
1750 #endif
1751 #ifdef INET
1752 	{
1753 		m->m_pkthdr.csum_flags = CSUM_TCP;
1754 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1755 		    htons(tlen + optlen - hlen + IPPROTO_TCP));
1756 #ifdef TCP_OFFLOAD
1757 		if (ADDED_BY_TOE(sc)) {
1758 			struct toedev *tod = sc->sc_tod;
1759 
1760 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1761 
1762 			return (error);
1763 		}
1764 #endif
1765 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1766 	}
1767 #endif
1768 	return (error);
1769 }
1770 
1771 /*
1772  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1773  * that exceed the capacity of the syncache by avoiding the storage of any
1774  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
1775  * attacks where the attacker does not have access to our responses.
1776  *
1777  * Syncookies encode and include all necessary information about the
1778  * connection setup within the SYN|ACK that we send back.  That way we
1779  * can avoid keeping any local state until the ACK to our SYN|ACK returns
1780  * (if ever).  Normally the syncache and syncookies are running in parallel
1781  * with the latter taking over when the former is exhausted.  When matching
1782  * syncache entry is found the syncookie is ignored.
1783  *
1784  * The only reliable information persisting the 3WHS is our initial sequence
1785  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
1786  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1787  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
1788  * returns and signifies a legitimate connection if it matches the ACK.
1789  *
1790  * The available space of 32 bits to store the hash and to encode the SYN
1791  * option information is very tight and we should have at least 24 bits for
1792  * the MAC to keep the number of guesses by blind spoofing reasonably high.
1793  *
1794  * SYN option information we have to encode to fully restore a connection:
1795  * MSS: is imporant to chose an optimal segment size to avoid IP level
1796  *   fragmentation along the path.  The common MSS values can be encoded
1797  *   in a 3-bit table.  Uncommon values are captured by the next lower value
1798  *   in the table leading to a slight increase in packetization overhead.
1799  * WSCALE: is necessary to allow large windows to be used for high delay-
1800  *   bandwidth product links.  Not scaling the window when it was initially
1801  *   negotiated is bad for performance as lack of scaling further decreases
1802  *   the apparent available send window.  We only need to encode the WSCALE
1803  *   we received from the remote end.  Our end can be recalculated at any
1804  *   time.  The common WSCALE values can be encoded in a 3-bit table.
1805  *   Uncommon values are captured by the next lower value in the table
1806  *   making us under-estimate the available window size halving our
1807  *   theoretically possible maximum throughput for that connection.
1808  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
1809  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
1810  *   that are included in all segments on a connection.  We enable them when
1811  *   the ACK has them.
1812  *
1813  * Security of syncookies and attack vectors:
1814  *
1815  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
1816  * together with the gloabl secret to make it unique per connection attempt.
1817  * Thus any change of any of those parameters results in a different MAC output
1818  * in an unpredictable way unless a collision is encountered.  24 bits of the
1819  * MAC are embedded into the ISS.
1820  *
1821  * To prevent replay attacks two rotating global secrets are updated with a
1822  * new random value every 15 seconds.  The life-time of a syncookie is thus
1823  * 15-30 seconds.
1824  *
1825  * Vector 1: Attacking the secret.  This requires finding a weakness in the
1826  * MAC itself or the way it is used here.  The attacker can do a chosen plain
1827  * text attack by varying and testing the all parameters under his control.
1828  * The strength depends on the size and randomness of the secret, and the
1829  * cryptographic security of the MAC function.  Due to the constant updating
1830  * of the secret the attacker has at most 29.999 seconds to find the secret
1831  * and launch spoofed connections.  After that he has to start all over again.
1832  *
1833  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
1834  * size an average of 4,823 attempts are required for a 50% chance of success
1835  * to spoof a single syncookie (birthday collision paradox).  However the
1836  * attacker is blind and doesn't know if one of his attempts succeeded unless
1837  * he has a side channel to interfere success from.  A single connection setup
1838  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
1839  * This many attempts are required for each one blind spoofed connection.  For
1840  * every additional spoofed connection he has to launch another N attempts.
1841  * Thus for a sustained rate 100 spoofed connections per second approximately
1842  * 1,800,000 packets per second would have to be sent.
1843  *
1844  * NB: The MAC function should be fast so that it doesn't become a CPU
1845  * exhaustion attack vector itself.
1846  *
1847  * References:
1848  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
1849  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
1850  *   http://cr.yp.to/syncookies.html    (overview)
1851  *   http://cr.yp.to/syncookies/archive (details)
1852  *
1853  *
1854  * Schematic construction of a syncookie enabled Initial Sequence Number:
1855  *  0        1         2         3
1856  *  12345678901234567890123456789012
1857  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
1858  *
1859  *  x 24 MAC (truncated)
1860  *  W  3 Send Window Scale index
1861  *  M  3 MSS index
1862  *  S  1 SACK permitted
1863  *  P  1 Odd/even secret
1864  */
1865 
1866 /*
1867  * Distribution and probability of certain MSS values.  Those in between are
1868  * rounded down to the next lower one.
1869  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
1870  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
1871  */
1872 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
1873 
1874 /*
1875  * Distribution and probability of certain WSCALE values.  We have to map the
1876  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
1877  * bits based on prevalence of certain values.  Where we don't have an exact
1878  * match for are rounded down to the next lower one letting us under-estimate
1879  * the true available window.  At the moment this would happen only for the
1880  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
1881  * and window size).  The absence of the WSCALE option (no scaling in either
1882  * direction) is encoded with index zero.
1883  * [WSCALE values histograms, Allman, 2012]
1884  *                            X 10 10 35  5  6 14 10%   by host
1885  *                            X 11  4  5  5 18 49  3%   by connections
1886  */
1887 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
1888 
1889 /*
1890  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
1891  * and good cryptographic properties.
1892  */
1893 static uint32_t
1894 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
1895     uint8_t *secbits, uintptr_t secmod)
1896 {
1897 	SIPHASH_CTX ctx;
1898 	uint32_t siphash[2];
1899 
1900 	SipHash24_Init(&ctx);
1901 	SipHash_SetKey(&ctx, secbits);
1902 	switch (inc->inc_flags & INC_ISIPV6) {
1903 #ifdef INET
1904 	case 0:
1905 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
1906 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
1907 		break;
1908 #endif
1909 #ifdef INET6
1910 	case INC_ISIPV6:
1911 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
1912 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
1913 		break;
1914 #endif
1915 	}
1916 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
1917 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
1918 	SipHash_Update(&ctx, &irs, sizeof(irs));
1919 	SipHash_Update(&ctx, &flags, sizeof(flags));
1920 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
1921 	SipHash_Final((u_int8_t *)&siphash, &ctx);
1922 
1923 	return (siphash[0] ^ siphash[1]);
1924 }
1925 
1926 static tcp_seq
1927 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
1928 {
1929 	u_int i, mss, secbit, wscale;
1930 	uint32_t iss, hash;
1931 	uint8_t *secbits;
1932 	union syncookie cookie;
1933 
1934 	SCH_LOCK_ASSERT(sch);
1935 
1936 	cookie.cookie = 0;
1937 
1938 	/* Map our computed MSS into the 3-bit index. */
1939 	mss = min(tcp_mssopt(&sc->sc_inc), max(sc->sc_peer_mss, V_tcp_minmss));
1940 	for (i = nitems(tcp_sc_msstab) - 1; tcp_sc_msstab[i] > mss && i > 0;
1941 	     i--)
1942 		;
1943 	cookie.flags.mss_idx = i;
1944 
1945 	/*
1946 	 * Map the send window scale into the 3-bit index but only if
1947 	 * the wscale option was received.
1948 	 */
1949 	if (sc->sc_flags & SCF_WINSCALE) {
1950 		wscale = sc->sc_requested_s_scale;
1951 		for (i = nitems(tcp_sc_wstab) - 1;
1952 		    tcp_sc_wstab[i] > wscale && i > 0;
1953 		     i--)
1954 			;
1955 		cookie.flags.wscale_idx = i;
1956 	}
1957 
1958 	/* Can we do SACK? */
1959 	if (sc->sc_flags & SCF_SACK)
1960 		cookie.flags.sack_ok = 1;
1961 
1962 	/* Which of the two secrets to use. */
1963 	secbit = sch->sch_sc->secret.oddeven & 0x1;
1964 	cookie.flags.odd_even = secbit;
1965 
1966 	secbits = sch->sch_sc->secret.key[secbit];
1967 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
1968 	    (uintptr_t)sch);
1969 
1970 	/*
1971 	 * Put the flags into the hash and XOR them to get better ISS number
1972 	 * variance.  This doesn't enhance the cryptographic strength and is
1973 	 * done to prevent the 8 cookie bits from showing up directly on the
1974 	 * wire.
1975 	 */
1976 	iss = hash & ~0xff;
1977 	iss |= cookie.cookie ^ (hash >> 24);
1978 
1979 	/* Randomize the timestamp. */
1980 	if (sc->sc_flags & SCF_TIMESTAMP) {
1981 		sc->sc_ts = arc4random();
1982 		sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks();
1983 	}
1984 
1985 	TCPSTAT_INC(tcps_sc_sendcookie);
1986 	return (iss);
1987 }
1988 
1989 static struct syncache *
1990 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
1991     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
1992     struct socket *lso)
1993 {
1994 	uint32_t hash;
1995 	uint8_t *secbits;
1996 	tcp_seq ack, seq;
1997 	int wnd, wscale = 0;
1998 	union syncookie cookie;
1999 
2000 	SCH_LOCK_ASSERT(sch);
2001 
2002 	/*
2003 	 * Pull information out of SYN-ACK/ACK and revert sequence number
2004 	 * advances.
2005 	 */
2006 	ack = th->th_ack - 1;
2007 	seq = th->th_seq - 1;
2008 
2009 	/*
2010 	 * Unpack the flags containing enough information to restore the
2011 	 * connection.
2012 	 */
2013 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2014 
2015 	/* Which of the two secrets to use. */
2016 	secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
2017 
2018 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2019 
2020 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
2021 	if ((ack & ~0xff) != (hash & ~0xff))
2022 		return (NULL);
2023 
2024 	/* Fill in the syncache values. */
2025 	sc->sc_flags = 0;
2026 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2027 	sc->sc_ipopts = NULL;
2028 
2029 	sc->sc_irs = seq;
2030 	sc->sc_iss = ack;
2031 
2032 	switch (inc->inc_flags & INC_ISIPV6) {
2033 #ifdef INET
2034 	case 0:
2035 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2036 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2037 		break;
2038 #endif
2039 #ifdef INET6
2040 	case INC_ISIPV6:
2041 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2042 			sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
2043 		break;
2044 #endif
2045 	}
2046 
2047 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2048 
2049 	/* We can simply recompute receive window scale we sent earlier. */
2050 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2051 		wscale++;
2052 
2053 	/* Only use wscale if it was enabled in the orignal SYN. */
2054 	if (cookie.flags.wscale_idx > 0) {
2055 		sc->sc_requested_r_scale = wscale;
2056 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2057 		sc->sc_flags |= SCF_WINSCALE;
2058 	}
2059 
2060 	wnd = sbspace(&lso->so_rcv);
2061 	wnd = imax(wnd, 0);
2062 	wnd = imin(wnd, TCP_MAXWIN);
2063 	sc->sc_wnd = wnd;
2064 
2065 	if (cookie.flags.sack_ok)
2066 		sc->sc_flags |= SCF_SACK;
2067 
2068 	if (to->to_flags & TOF_TS) {
2069 		sc->sc_flags |= SCF_TIMESTAMP;
2070 		sc->sc_tsreflect = to->to_tsval;
2071 		sc->sc_ts = to->to_tsecr;
2072 		sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks();
2073 	}
2074 
2075 	if (to->to_flags & TOF_SIGNATURE)
2076 		sc->sc_flags |= SCF_SIGNATURE;
2077 
2078 	sc->sc_rxmits = 0;
2079 
2080 	TCPSTAT_INC(tcps_sc_recvcookie);
2081 	return (sc);
2082 }
2083 
2084 #ifdef INVARIANTS
2085 static int
2086 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2087     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2088     struct socket *lso)
2089 {
2090 	struct syncache scs, *scx;
2091 	char *s;
2092 
2093 	bzero(&scs, sizeof(scs));
2094 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
2095 
2096 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2097 		return (0);
2098 
2099 	if (scx != NULL) {
2100 		if (sc->sc_peer_mss != scx->sc_peer_mss)
2101 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2102 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2103 
2104 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2105 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2106 			    s, __func__, sc->sc_requested_r_scale,
2107 			    scx->sc_requested_r_scale);
2108 
2109 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2110 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2111 			    s, __func__, sc->sc_requested_s_scale,
2112 			    scx->sc_requested_s_scale);
2113 
2114 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2115 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2116 	}
2117 
2118 	if (s != NULL)
2119 		free(s, M_TCPLOG);
2120 	return (0);
2121 }
2122 #endif /* INVARIANTS */
2123 
2124 static void
2125 syncookie_reseed(void *arg)
2126 {
2127 	struct tcp_syncache *sc = arg;
2128 	uint8_t *secbits;
2129 	int secbit;
2130 
2131 	/*
2132 	 * Reseeding the secret doesn't have to be protected by a lock.
2133 	 * It only must be ensured that the new random values are visible
2134 	 * to all CPUs in a SMP environment.  The atomic with release
2135 	 * semantics ensures that.
2136 	 */
2137 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2138 	secbits = sc->secret.key[secbit];
2139 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2140 	atomic_add_rel_int(&sc->secret.oddeven, 1);
2141 
2142 	/* Reschedule ourself. */
2143 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2144 }
2145 
2146 /*
2147  * Exports the syncache entries to userland so that netstat can display
2148  * them alongside the other sockets.  This function is intended to be
2149  * called only from tcp_pcblist.
2150  *
2151  * Due to concurrency on an active system, the number of pcbs exported
2152  * may have no relation to max_pcbs.  max_pcbs merely indicates the
2153  * amount of space the caller allocated for this function to use.
2154  */
2155 int
2156 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
2157 {
2158 	struct xtcpcb xt;
2159 	struct syncache *sc;
2160 	struct syncache_head *sch;
2161 	int count, error, i;
2162 
2163 	for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
2164 		sch = &V_tcp_syncache.hashbase[i];
2165 		SCH_LOCK(sch);
2166 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2167 			if (count >= max_pcbs) {
2168 				SCH_UNLOCK(sch);
2169 				goto exit;
2170 			}
2171 			if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2172 				continue;
2173 			bzero(&xt, sizeof(xt));
2174 			xt.xt_len = sizeof(xt);
2175 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2176 				xt.xt_inp.inp_vflag = INP_IPV6;
2177 			else
2178 				xt.xt_inp.inp_vflag = INP_IPV4;
2179 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo));
2180 			xt.xt_tp.t_inpcb = &xt.xt_inp;
2181 			xt.xt_tp.t_state = TCPS_SYN_RECEIVED;
2182 			xt.xt_socket.xso_protocol = IPPROTO_TCP;
2183 			xt.xt_socket.xso_len = sizeof (struct xsocket);
2184 			xt.xt_socket.so_type = SOCK_STREAM;
2185 			xt.xt_socket.so_state = SS_ISCONNECTING;
2186 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2187 			if (error) {
2188 				SCH_UNLOCK(sch);
2189 				goto exit;
2190 			}
2191 			count++;
2192 		}
2193 		SCH_UNLOCK(sch);
2194 	}
2195 exit:
2196 	*pcbs_exported = count;
2197 	return error;
2198 }
2199