xref: /freebsd-12.1/sys/netinet/tcp_syncache.c (revision cec19e87)
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 *, struct syncache_head *,
134 		    const struct mbuf *, int);
135 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
136 		    struct mbuf *m);
137 static void	 syncache_timeout(struct syncache *sc, struct syncache_head *sch,
138 		    int docallout);
139 static void	 syncache_timer(void *);
140 
141 static uint32_t	 syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
142 		    uint8_t *, uintptr_t);
143 static tcp_seq	 syncookie_generate(struct syncache_head *, struct syncache *);
144 static struct syncache
145 		*syncookie_lookup(struct in_conninfo *, struct syncache_head *,
146 		    struct syncache *, struct tcphdr *, struct tcpopt *,
147 		    struct socket *);
148 static void	 syncookie_reseed(void *);
149 #ifdef INVARIANTS
150 static int	 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
151 		    struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
152 		    struct socket *lso);
153 #endif
154 
155 /*
156  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
157  * 3 retransmits corresponds to a timeout with default values of
158  * tcp_rexmit_initial * (             1 +
159  *                       tcp_backoff[1] +
160  *                       tcp_backoff[2] +
161  *                       tcp_backoff[3]) + 3 * tcp_rexmit_slop,
162  * 1000 ms * (1 + 2 + 4 + 8) +  3 * 200 ms = 15600 ms,
163  * the odds are that the user has given up attempting to connect by then.
164  */
165 #define SYNCACHE_MAXREXMTS		3
166 
167 /* Arbitrary values */
168 #define TCP_SYNCACHE_HASHSIZE		512
169 #define TCP_SYNCACHE_BUCKETLIMIT	30
170 
171 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
172 #define	V_tcp_syncache			VNET(tcp_syncache)
173 
174 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, 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 static int
193 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
194 {
195 	int error;
196 	u_int new;
197 
198 	new = V_tcp_syncache.rexmt_limit;
199 	error = sysctl_handle_int(oidp, &new, 0, req);
200 	if ((error == 0) && (req->newptr != NULL)) {
201 		if (new > TCP_MAXRXTSHIFT)
202 			error = EINVAL;
203 		else
204 			V_tcp_syncache.rexmt_limit = new;
205 	}
206 	return (error);
207 }
208 
209 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
210     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW,
211     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
212     sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
213     "Limit on SYN/ACK retransmissions");
214 
215 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
216 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
217     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
218     "Send reset on socket allocation failure");
219 
220 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
221 
222 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
223 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
224 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
225 
226 /*
227  * Requires the syncache entry to be already removed from the bucket list.
228  */
229 static void
230 syncache_free(struct syncache *sc)
231 {
232 
233 	if (sc->sc_ipopts)
234 		(void) m_free(sc->sc_ipopts);
235 	if (sc->sc_cred)
236 		crfree(sc->sc_cred);
237 #ifdef MAC
238 	mac_syncache_destroy(&sc->sc_label);
239 #endif
240 
241 	uma_zfree(V_tcp_syncache.zone, sc);
242 }
243 
244 void
245 syncache_init(void)
246 {
247 	int i;
248 
249 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
250 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
251 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
252 	V_tcp_syncache.hash_secret = arc4random();
253 
254 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
255 	    &V_tcp_syncache.hashsize);
256 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
257 	    &V_tcp_syncache.bucket_limit);
258 	if (!powerof2(V_tcp_syncache.hashsize) ||
259 	    V_tcp_syncache.hashsize == 0) {
260 		printf("WARNING: syncache hash size is not a power of 2.\n");
261 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
262 	}
263 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
264 
265 	/* Set limits. */
266 	V_tcp_syncache.cache_limit =
267 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
268 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
269 	    &V_tcp_syncache.cache_limit);
270 
271 	/* Allocate the hash table. */
272 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
273 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
274 
275 #ifdef VIMAGE
276 	V_tcp_syncache.vnet = curvnet;
277 #endif
278 
279 	/* Initialize the hash buckets. */
280 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
281 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
282 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
283 			 NULL, MTX_DEF);
284 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
285 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
286 		V_tcp_syncache.hashbase[i].sch_length = 0;
287 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
288 		V_tcp_syncache.hashbase[i].sch_last_overflow =
289 		    -(SYNCOOKIE_LIFETIME + 1);
290 	}
291 
292 	/* Create the syncache entry zone. */
293 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
294 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
295 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
296 	    V_tcp_syncache.cache_limit);
297 
298 	/* Start the SYN cookie reseeder callout. */
299 	callout_init(&V_tcp_syncache.secret.reseed, 1);
300 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
301 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
302 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
303 	    syncookie_reseed, &V_tcp_syncache);
304 }
305 
306 #ifdef VIMAGE
307 void
308 syncache_destroy(void)
309 {
310 	struct syncache_head *sch;
311 	struct syncache *sc, *nsc;
312 	int i;
313 
314 	/*
315 	 * Stop the re-seed timer before freeing resources.  No need to
316 	 * possibly schedule it another time.
317 	 */
318 	callout_drain(&V_tcp_syncache.secret.reseed);
319 
320 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
321 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
322 
323 		sch = &V_tcp_syncache.hashbase[i];
324 		callout_drain(&sch->sch_timer);
325 
326 		SCH_LOCK(sch);
327 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
328 			syncache_drop(sc, sch);
329 		SCH_UNLOCK(sch);
330 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
331 		    ("%s: sch->sch_bucket not empty", __func__));
332 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
333 		    __func__, sch->sch_length));
334 		mtx_destroy(&sch->sch_mtx);
335 	}
336 
337 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
338 	    ("%s: cache_count not 0", __func__));
339 
340 	/* Free the allocated global resources. */
341 	uma_zdestroy(V_tcp_syncache.zone);
342 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
343 }
344 #endif
345 
346 /*
347  * Inserts a syncache entry into the specified bucket row.
348  * Locks and unlocks the syncache_head autonomously.
349  */
350 static void
351 syncache_insert(struct syncache *sc, struct syncache_head *sch)
352 {
353 	struct syncache *sc2;
354 
355 	SCH_LOCK(sch);
356 
357 	/*
358 	 * Make sure that we don't overflow the per-bucket limit.
359 	 * If the bucket is full, toss the oldest element.
360 	 */
361 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
362 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
363 			("sch->sch_length incorrect"));
364 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
365 		sch->sch_last_overflow = time_uptime;
366 		syncache_drop(sc2, sch);
367 		TCPSTAT_INC(tcps_sc_bucketoverflow);
368 	}
369 
370 	/* Put it into the bucket. */
371 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
372 	sch->sch_length++;
373 
374 #ifdef TCP_OFFLOAD
375 	if (ADDED_BY_TOE(sc)) {
376 		struct toedev *tod = sc->sc_tod;
377 
378 		tod->tod_syncache_added(tod, sc->sc_todctx);
379 	}
380 #endif
381 
382 	/* Reinitialize the bucket row's timer. */
383 	if (sch->sch_length == 1)
384 		sch->sch_nextc = ticks + INT_MAX;
385 	syncache_timeout(sc, sch, 1);
386 
387 	SCH_UNLOCK(sch);
388 
389 	TCPSTATES_INC(TCPS_SYN_RECEIVED);
390 	TCPSTAT_INC(tcps_sc_added);
391 }
392 
393 /*
394  * Remove and free entry from syncache bucket row.
395  * Expects locked syncache head.
396  */
397 static void
398 syncache_drop(struct syncache *sc, struct syncache_head *sch)
399 {
400 
401 	SCH_LOCK_ASSERT(sch);
402 
403 	TCPSTATES_DEC(TCPS_SYN_RECEIVED);
404 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
405 	sch->sch_length--;
406 
407 #ifdef TCP_OFFLOAD
408 	if (ADDED_BY_TOE(sc)) {
409 		struct toedev *tod = sc->sc_tod;
410 
411 		tod->tod_syncache_removed(tod, sc->sc_todctx);
412 	}
413 #endif
414 
415 	syncache_free(sc);
416 }
417 
418 /*
419  * Engage/reengage time on bucket row.
420  */
421 static void
422 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
423 {
424 	int rexmt;
425 
426 	if (sc->sc_rxmits == 0)
427 		rexmt = tcp_rexmit_initial;
428 	else
429 		TCPT_RANGESET(rexmt,
430 		    tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
431 		    tcp_rexmit_min, TCPTV_REXMTMAX);
432 	sc->sc_rxttime = ticks + rexmt;
433 	sc->sc_rxmits++;
434 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
435 		sch->sch_nextc = sc->sc_rxttime;
436 		if (docallout)
437 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
438 			    syncache_timer, (void *)sch);
439 	}
440 }
441 
442 /*
443  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
444  * If we have retransmitted an entry the maximum number of times, expire it.
445  * One separate timer for each bucket row.
446  */
447 static void
448 syncache_timer(void *xsch)
449 {
450 	struct syncache_head *sch = (struct syncache_head *)xsch;
451 	struct syncache *sc, *nsc;
452 	int tick = ticks;
453 	char *s;
454 
455 	CURVNET_SET(sch->sch_sc->vnet);
456 
457 	/* NB: syncache_head has already been locked by the callout. */
458 	SCH_LOCK_ASSERT(sch);
459 
460 	/*
461 	 * In the following cycle we may remove some entries and/or
462 	 * advance some timeouts, so re-initialize the bucket timer.
463 	 */
464 	sch->sch_nextc = tick + INT_MAX;
465 
466 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
467 		/*
468 		 * We do not check if the listen socket still exists
469 		 * and accept the case where the listen socket may be
470 		 * gone by the time we resend the SYN/ACK.  We do
471 		 * not expect this to happens often. If it does,
472 		 * then the RST will be sent by the time the remote
473 		 * host does the SYN/ACK->ACK.
474 		 */
475 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
476 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
477 				sch->sch_nextc = sc->sc_rxttime;
478 			continue;
479 		}
480 		if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
481 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
482 				log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
483 				    "giving up and removing syncache entry\n",
484 				    s, __func__);
485 				free(s, M_TCPLOG);
486 			}
487 			syncache_drop(sc, sch);
488 			TCPSTAT_INC(tcps_sc_stale);
489 			continue;
490 		}
491 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
492 			log(LOG_DEBUG, "%s; %s: Response timeout, "
493 			    "retransmitting (%u) SYN|ACK\n",
494 			    s, __func__, sc->sc_rxmits);
495 			free(s, M_TCPLOG);
496 		}
497 
498 		syncache_respond(sc, sch, NULL, TH_SYN|TH_ACK);
499 		TCPSTAT_INC(tcps_sc_retransmitted);
500 		syncache_timeout(sc, sch, 0);
501 	}
502 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
503 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
504 			syncache_timer, (void *)(sch));
505 	CURVNET_RESTORE();
506 }
507 
508 /*
509  * Find an entry in the syncache.
510  * Returns always with locked syncache_head plus a matching entry or NULL.
511  */
512 static struct syncache *
513 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
514 {
515 	struct syncache *sc;
516 	struct syncache_head *sch;
517 	uint32_t hash;
518 
519 	/*
520 	 * The hash is built on foreign port + local port + foreign address.
521 	 * We rely on the fact that struct in_conninfo starts with 16 bits
522 	 * of foreign port, then 16 bits of local port then followed by 128
523 	 * bits of foreign address.  In case of IPv4 address, the first 3
524 	 * 32-bit words of the address always are zeroes.
525 	 */
526 	hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
527 	    V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
528 
529 	sch = &V_tcp_syncache.hashbase[hash];
530 	*schp = sch;
531 	SCH_LOCK(sch);
532 
533 	/* Circle through bucket row to find matching entry. */
534 	TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
535 		if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
536 		    sizeof(struct in_endpoints)) == 0)
537 			break;
538 
539 	return (sc);	/* Always returns with locked sch. */
540 }
541 
542 /*
543  * This function is called when we get a RST for a
544  * non-existent connection, so that we can see if the
545  * connection is in the syn cache.  If it is, zap it.
546  * If required send a challenge ACK.
547  */
548 void
549 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m)
550 {
551 	struct syncache *sc;
552 	struct syncache_head *sch;
553 	char *s = NULL;
554 
555 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
556 	SCH_LOCK_ASSERT(sch);
557 
558 	/*
559 	 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
560 	 * See RFC 793 page 65, section SEGMENT ARRIVES.
561 	 */
562 	if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
563 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
564 			log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
565 			    "FIN flag set, segment ignored\n", s, __func__);
566 		TCPSTAT_INC(tcps_badrst);
567 		goto done;
568 	}
569 
570 	/*
571 	 * No corresponding connection was found in syncache.
572 	 * If syncookies are enabled and possibly exclusively
573 	 * used, or we are under memory pressure, a valid RST
574 	 * may not find a syncache entry.  In that case we're
575 	 * done and no SYN|ACK retransmissions will happen.
576 	 * Otherwise the RST was misdirected or spoofed.
577 	 */
578 	if (sc == NULL) {
579 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
580 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
581 			    "syncache entry (possibly syncookie only), "
582 			    "segment ignored\n", s, __func__);
583 		TCPSTAT_INC(tcps_badrst);
584 		goto done;
585 	}
586 
587 	/*
588 	 * If the RST bit is set, check the sequence number to see
589 	 * if this is a valid reset segment.
590 	 *
591 	 * RFC 793 page 37:
592 	 *   In all states except SYN-SENT, all reset (RST) segments
593 	 *   are validated by checking their SEQ-fields.  A reset is
594 	 *   valid if its sequence number is in the window.
595 	 *
596 	 * RFC 793 page 69:
597 	 *   There are four cases for the acceptability test for an incoming
598 	 *   segment:
599 	 *
600 	 * Segment Receive  Test
601 	 * Length  Window
602 	 * ------- -------  -------------------------------------------
603 	 *    0       0     SEG.SEQ = RCV.NXT
604 	 *    0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
605 	 *   >0       0     not acceptable
606 	 *   >0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
607 	 *               or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
608 	 *
609 	 * Note that when receiving a SYN segment in the LISTEN state,
610 	 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
611 	 * described in RFC 793, page 66.
612 	 */
613 	if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
614 	    SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
615 	    (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
616 		if (V_tcp_insecure_rst ||
617 		    th->th_seq == sc->sc_irs + 1) {
618 			syncache_drop(sc, sch);
619 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
620 				log(LOG_DEBUG,
621 				    "%s; %s: Our SYN|ACK was rejected, "
622 				    "connection attempt aborted by remote "
623 				    "endpoint\n",
624 				    s, __func__);
625 			TCPSTAT_INC(tcps_sc_reset);
626 		} else {
627 			TCPSTAT_INC(tcps_badrst);
628 			/* Send challenge ACK. */
629 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
630 				log(LOG_DEBUG, "%s; %s: RST with invalid "
631 				    " SEQ %u != NXT %u (+WND %u), "
632 				    "sending challenge ACK\n",
633 				    s, __func__,
634 				    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
635 			syncache_respond(sc, sch, m, TH_ACK);
636 		}
637 	} else {
638 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
639 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
640 			    "NXT %u (+WND %u), segment ignored\n",
641 			    s, __func__,
642 			    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
643 		TCPSTAT_INC(tcps_badrst);
644 	}
645 
646 done:
647 	if (s != NULL)
648 		free(s, M_TCPLOG);
649 	SCH_UNLOCK(sch);
650 }
651 
652 void
653 syncache_badack(struct in_conninfo *inc)
654 {
655 	struct syncache *sc;
656 	struct syncache_head *sch;
657 
658 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
659 	SCH_LOCK_ASSERT(sch);
660 	if (sc != NULL) {
661 		syncache_drop(sc, sch);
662 		TCPSTAT_INC(tcps_sc_badack);
663 	}
664 	SCH_UNLOCK(sch);
665 }
666 
667 void
668 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq)
669 {
670 	struct syncache *sc;
671 	struct syncache_head *sch;
672 
673 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
674 	SCH_LOCK_ASSERT(sch);
675 	if (sc == NULL)
676 		goto done;
677 
678 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
679 	if (ntohl(th_seq) != sc->sc_iss)
680 		goto done;
681 
682 	/*
683 	 * If we've rertransmitted 3 times and this is our second error,
684 	 * we remove the entry.  Otherwise, we allow it to continue on.
685 	 * This prevents us from incorrectly nuking an entry during a
686 	 * spurious network outage.
687 	 *
688 	 * See tcp_notify().
689 	 */
690 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
691 		sc->sc_flags |= SCF_UNREACH;
692 		goto done;
693 	}
694 	syncache_drop(sc, sch);
695 	TCPSTAT_INC(tcps_sc_unreach);
696 done:
697 	SCH_UNLOCK(sch);
698 }
699 
700 /*
701  * Build a new TCP socket structure from a syncache entry.
702  *
703  * On success return the newly created socket with its underlying inp locked.
704  */
705 static struct socket *
706 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
707 {
708 	struct tcp_function_block *blk;
709 	struct inpcb *inp = NULL;
710 	struct socket *so;
711 	struct tcpcb *tp;
712 	int error;
713 	char *s;
714 
715 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
716 
717 	/*
718 	 * Ok, create the full blown connection, and set things up
719 	 * as they would have been set up if we had created the
720 	 * connection when the SYN arrived.  If we can't create
721 	 * the connection, abort it.
722 	 */
723 	so = sonewconn(lso, 0);
724 	if (so == NULL) {
725 		/*
726 		 * Drop the connection; we will either send a RST or
727 		 * have the peer retransmit its SYN again after its
728 		 * RTO and try again.
729 		 */
730 		TCPSTAT_INC(tcps_listendrop);
731 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
732 			log(LOG_DEBUG, "%s; %s: Socket create failed "
733 			    "due to limits or memory shortage\n",
734 			    s, __func__);
735 			free(s, M_TCPLOG);
736 		}
737 		goto abort2;
738 	}
739 #ifdef MAC
740 	mac_socketpeer_set_from_mbuf(m, so);
741 #endif
742 
743 	inp = sotoinpcb(so);
744 	inp->inp_inc.inc_fibnum = so->so_fibnum;
745 	INP_WLOCK(inp);
746 	/*
747 	 * Exclusive pcbinfo lock is not required in syncache socket case even
748 	 * if two inpcb locks can be acquired simultaneously:
749 	 *  - the inpcb in LISTEN state,
750 	 *  - the newly created inp.
751 	 *
752 	 * In this case, an inp cannot be at same time in LISTEN state and
753 	 * just created by an accept() call.
754 	 */
755 	INP_HASH_WLOCK(&V_tcbinfo);
756 
757 	/* Insert new socket into PCB hash list. */
758 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
759 #ifdef INET6
760 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
761 		inp->inp_vflag &= ~INP_IPV4;
762 		inp->inp_vflag |= INP_IPV6;
763 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
764 	} else {
765 		inp->inp_vflag &= ~INP_IPV6;
766 		inp->inp_vflag |= INP_IPV4;
767 #endif
768 		inp->inp_laddr = sc->sc_inc.inc_laddr;
769 #ifdef INET6
770 	}
771 #endif
772 
773 	/*
774 	 * If there's an mbuf and it has a flowid, then let's initialise the
775 	 * inp with that particular flowid.
776 	 */
777 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
778 		inp->inp_flowid = m->m_pkthdr.flowid;
779 		inp->inp_flowtype = M_HASHTYPE_GET(m);
780 	}
781 
782 	/*
783 	 * Install in the reservation hash table for now, but don't yet
784 	 * install a connection group since the full 4-tuple isn't yet
785 	 * configured.
786 	 */
787 	inp->inp_lport = sc->sc_inc.inc_lport;
788 	if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
789 		/*
790 		 * Undo the assignments above if we failed to
791 		 * put the PCB on the hash lists.
792 		 */
793 #ifdef INET6
794 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
795 			inp->in6p_laddr = in6addr_any;
796 		else
797 #endif
798 			inp->inp_laddr.s_addr = INADDR_ANY;
799 		inp->inp_lport = 0;
800 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
801 			log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
802 			    "with error %i\n",
803 			    s, __func__, error);
804 			free(s, M_TCPLOG);
805 		}
806 		INP_HASH_WUNLOCK(&V_tcbinfo);
807 		goto abort;
808 	}
809 #ifdef INET6
810 	if (inp->inp_vflag & INP_IPV6PROTO) {
811 		struct inpcb *oinp = sotoinpcb(lso);
812 
813 		/*
814 		 * Inherit socket options from the listening socket.
815 		 * Note that in6p_inputopts are not (and should not be)
816 		 * copied, since it stores previously received options and is
817 		 * used to detect if each new option is different than the
818 		 * previous one and hence should be passed to a user.
819 		 * If we copied in6p_inputopts, a user would not be able to
820 		 * receive options just after calling the accept system call.
821 		 */
822 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
823 		if (oinp->in6p_outputopts)
824 			inp->in6p_outputopts =
825 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
826 	}
827 
828 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
829 		struct in6_addr laddr6;
830 		struct sockaddr_in6 sin6;
831 
832 		sin6.sin6_family = AF_INET6;
833 		sin6.sin6_len = sizeof(sin6);
834 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
835 		sin6.sin6_port = sc->sc_inc.inc_fport;
836 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
837 		laddr6 = inp->in6p_laddr;
838 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
839 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
840 		if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
841 		    thread0.td_ucred, m)) != 0) {
842 			inp->in6p_laddr = laddr6;
843 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
844 				log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
845 				    "with error %i\n",
846 				    s, __func__, error);
847 				free(s, M_TCPLOG);
848 			}
849 			INP_HASH_WUNLOCK(&V_tcbinfo);
850 			goto abort;
851 		}
852 		/* Override flowlabel from in6_pcbconnect. */
853 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
854 		inp->inp_flow |= sc->sc_flowlabel;
855 	}
856 #endif /* INET6 */
857 #if defined(INET) && defined(INET6)
858 	else
859 #endif
860 #ifdef INET
861 	{
862 		struct in_addr laddr;
863 		struct sockaddr_in sin;
864 
865 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
866 
867 		if (inp->inp_options == NULL) {
868 			inp->inp_options = sc->sc_ipopts;
869 			sc->sc_ipopts = NULL;
870 		}
871 
872 		sin.sin_family = AF_INET;
873 		sin.sin_len = sizeof(sin);
874 		sin.sin_addr = sc->sc_inc.inc_faddr;
875 		sin.sin_port = sc->sc_inc.inc_fport;
876 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
877 		laddr = inp->inp_laddr;
878 		if (inp->inp_laddr.s_addr == INADDR_ANY)
879 			inp->inp_laddr = sc->sc_inc.inc_laddr;
880 		if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
881 		    thread0.td_ucred, m)) != 0) {
882 			inp->inp_laddr = laddr;
883 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
884 				log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
885 				    "with error %i\n",
886 				    s, __func__, error);
887 				free(s, M_TCPLOG);
888 			}
889 			INP_HASH_WUNLOCK(&V_tcbinfo);
890 			goto abort;
891 		}
892 	}
893 #endif /* INET */
894 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
895 	/* Copy old policy into new socket's. */
896 	if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
897 		printf("syncache_socket: could not copy policy\n");
898 #endif
899 	INP_HASH_WUNLOCK(&V_tcbinfo);
900 	tp = intotcpcb(inp);
901 	tcp_state_change(tp, TCPS_SYN_RECEIVED);
902 	tp->iss = sc->sc_iss;
903 	tp->irs = sc->sc_irs;
904 	tcp_rcvseqinit(tp);
905 	tcp_sendseqinit(tp);
906 	blk = sototcpcb(lso)->t_fb;
907 	if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
908 		/*
909 		 * Our parents t_fb was not the default,
910 		 * we need to release our ref on tp->t_fb and
911 		 * pickup one on the new entry.
912 		 */
913 		struct tcp_function_block *rblk;
914 
915 		rblk = find_and_ref_tcp_fb(blk);
916 		KASSERT(rblk != NULL,
917 		    ("cannot find blk %p out of syncache?", blk));
918 		if (tp->t_fb->tfb_tcp_fb_fini)
919 			(*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
920 		refcount_release(&tp->t_fb->tfb_refcnt);
921 		tp->t_fb = rblk;
922 		/*
923 		 * XXXrrs this is quite dangerous, it is possible
924 		 * for the new function to fail to init. We also
925 		 * are not asking if the handoff_is_ok though at
926 		 * the very start thats probalbly ok.
927 		 */
928 		if (tp->t_fb->tfb_tcp_fb_init) {
929 			(*tp->t_fb->tfb_tcp_fb_init)(tp);
930 		}
931 	}
932 	tp->snd_wl1 = sc->sc_irs;
933 	tp->snd_max = tp->iss + 1;
934 	tp->snd_nxt = tp->iss + 1;
935 	tp->rcv_up = sc->sc_irs + 1;
936 	tp->rcv_wnd = sc->sc_wnd;
937 	tp->rcv_adv += tp->rcv_wnd;
938 	tp->last_ack_sent = tp->rcv_nxt;
939 
940 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
941 	if (sc->sc_flags & SCF_NOOPT)
942 		tp->t_flags |= TF_NOOPT;
943 	else {
944 		if (sc->sc_flags & SCF_WINSCALE) {
945 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
946 			tp->snd_scale = sc->sc_requested_s_scale;
947 			tp->request_r_scale = sc->sc_requested_r_scale;
948 		}
949 		if (sc->sc_flags & SCF_TIMESTAMP) {
950 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
951 			tp->ts_recent = sc->sc_tsreflect;
952 			tp->ts_recent_age = tcp_ts_getticks();
953 			tp->ts_offset = sc->sc_tsoff;
954 		}
955 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
956 		if (sc->sc_flags & SCF_SIGNATURE)
957 			tp->t_flags |= TF_SIGNATURE;
958 #endif
959 		if (sc->sc_flags & SCF_SACK)
960 			tp->t_flags |= TF_SACK_PERMIT;
961 	}
962 
963 	if (sc->sc_flags & SCF_ECN)
964 		tp->t_flags |= TF_ECN_PERMIT;
965 
966 	/*
967 	 * Set up MSS and get cached values from tcp_hostcache.
968 	 * This might overwrite some of the defaults we just set.
969 	 */
970 	tcp_mss(tp, sc->sc_peer_mss);
971 
972 	/*
973 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
974 	 * limited to one segment in cc_conn_init().
975 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
976 	 */
977 	if (sc->sc_rxmits > 1)
978 		tp->snd_cwnd = 1;
979 
980 #ifdef TCP_OFFLOAD
981 	/*
982 	 * Allow a TOE driver to install its hooks.  Note that we hold the
983 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
984 	 * new connection before the TOE driver has done its thing.
985 	 */
986 	if (ADDED_BY_TOE(sc)) {
987 		struct toedev *tod = sc->sc_tod;
988 
989 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
990 	}
991 #endif
992 	/*
993 	 * Copy and activate timers.
994 	 */
995 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
996 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
997 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
998 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
999 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1000 
1001 	TCPSTAT_INC(tcps_accepts);
1002 	return (so);
1003 
1004 abort:
1005 	INP_WUNLOCK(inp);
1006 abort2:
1007 	if (so != NULL)
1008 		soabort(so);
1009 	return (NULL);
1010 }
1011 
1012 /*
1013  * This function gets called when we receive an ACK for a
1014  * socket in the LISTEN state.  We look up the connection
1015  * in the syncache, and if its there, we pull it out of
1016  * the cache and turn it into a full-blown connection in
1017  * the SYN-RECEIVED state.
1018  *
1019  * On syncache_socket() success the newly created socket
1020  * has its underlying inp locked.
1021  */
1022 int
1023 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1024     struct socket **lsop, struct mbuf *m)
1025 {
1026 	struct syncache *sc;
1027 	struct syncache_head *sch;
1028 	struct syncache scs;
1029 	char *s;
1030 
1031 	/*
1032 	 * Global TCP locks are held because we manipulate the PCB lists
1033 	 * and create a new socket.
1034 	 */
1035 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1036 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1037 	    ("%s: can handle only ACK", __func__));
1038 
1039 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1040 	SCH_LOCK_ASSERT(sch);
1041 
1042 #ifdef INVARIANTS
1043 	/*
1044 	 * Test code for syncookies comparing the syncache stored
1045 	 * values with the reconstructed values from the cookie.
1046 	 */
1047 	if (sc != NULL)
1048 		syncookie_cmp(inc, sch, sc, th, to, *lsop);
1049 #endif
1050 
1051 	if (sc == NULL) {
1052 		/*
1053 		 * There is no syncache entry, so see if this ACK is
1054 		 * a returning syncookie.  To do this, first:
1055 		 *  A. Check if syncookies are used in case of syncache
1056 		 *     overflows
1057 		 *  B. See if this socket has had a syncache entry dropped in
1058 		 *     the recent past. We don't want to accept a bogus
1059 		 *     syncookie if we've never received a SYN or accept it
1060 		 *     twice.
1061 		 *  C. check that the syncookie is valid.  If it is, then
1062 		 *     cobble up a fake syncache entry, and return.
1063 		 */
1064 		if (!V_tcp_syncookies) {
1065 			SCH_UNLOCK(sch);
1066 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1067 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1068 				    "segment rejected (syncookies disabled)\n",
1069 				    s, __func__);
1070 			goto failed;
1071 		}
1072 		if (!V_tcp_syncookiesonly &&
1073 		    sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1074 			SCH_UNLOCK(sch);
1075 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1076 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1077 				    "segment rejected (no syncache entry)\n",
1078 				    s, __func__);
1079 			goto failed;
1080 		}
1081 		bzero(&scs, sizeof(scs));
1082 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
1083 		SCH_UNLOCK(sch);
1084 		if (sc == NULL) {
1085 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1086 				log(LOG_DEBUG, "%s; %s: Segment failed "
1087 				    "SYNCOOKIE authentication, segment rejected "
1088 				    "(probably spoofed)\n", s, __func__);
1089 			goto failed;
1090 		}
1091 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1092 		/* If received ACK has MD5 signature, check it. */
1093 		if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1094 		    (!TCPMD5_ENABLED() ||
1095 		    TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1096 			/* Drop the ACK. */
1097 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1098 				log(LOG_DEBUG, "%s; %s: Segment rejected, "
1099 				    "MD5 signature doesn't match.\n",
1100 				    s, __func__);
1101 				free(s, M_TCPLOG);
1102 			}
1103 			TCPSTAT_INC(tcps_sig_err_sigopt);
1104 			return (-1); /* Do not send RST */
1105 		}
1106 #endif /* TCP_SIGNATURE */
1107 	} else {
1108 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1109 		/*
1110 		 * If listening socket requested TCP digests, check that
1111 		 * received ACK has signature and it is correct.
1112 		 * If not, drop the ACK and leave sc entry in th cache,
1113 		 * because SYN was received with correct signature.
1114 		 */
1115 		if (sc->sc_flags & SCF_SIGNATURE) {
1116 			if ((to->to_flags & TOF_SIGNATURE) == 0) {
1117 				/* No signature */
1118 				TCPSTAT_INC(tcps_sig_err_nosigopt);
1119 				SCH_UNLOCK(sch);
1120 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1121 					log(LOG_DEBUG, "%s; %s: Segment "
1122 					    "rejected, MD5 signature wasn't "
1123 					    "provided.\n", s, __func__);
1124 					free(s, M_TCPLOG);
1125 				}
1126 				return (-1); /* Do not send RST */
1127 			}
1128 			if (!TCPMD5_ENABLED() ||
1129 			    TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1130 				/* Doesn't match or no SA */
1131 				SCH_UNLOCK(sch);
1132 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1133 					log(LOG_DEBUG, "%s; %s: Segment "
1134 					    "rejected, MD5 signature doesn't "
1135 					    "match.\n", s, __func__);
1136 					free(s, M_TCPLOG);
1137 				}
1138 				return (-1); /* Do not send RST */
1139 			}
1140 		}
1141 #endif /* TCP_SIGNATURE */
1142 
1143 		/*
1144 		 * RFC 7323 PAWS: If we have a timestamp on this segment and
1145 		 * it's less than ts_recent, drop it.
1146 		 * XXXMT: RFC 7323 also requires to send an ACK.
1147 		 *        In tcp_input.c this is only done for TCP segments
1148 		 *        with user data, so be consistent here and just drop
1149 		 *        the segment.
1150 		 */
1151 		if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1152 		    TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1153 			SCH_UNLOCK(sch);
1154 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1155 				log(LOG_DEBUG,
1156 				    "%s; %s: SEG.TSval %u < TS.Recent %u, "
1157 				    "segment dropped\n", s, __func__,
1158 				    to->to_tsval, sc->sc_tsreflect);
1159 				free(s, M_TCPLOG);
1160 			}
1161 			return (-1);  /* Do not send RST */
1162 		}
1163 
1164 		/*
1165 		 * Pull out the entry to unlock the bucket row.
1166 		 *
1167 		 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1168 		 * tcp_state_change().  The tcpcb is not existent at this
1169 		 * moment.  A new one will be allocated via syncache_socket->
1170 		 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1171 		 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1172 		 */
1173 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1174 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1175 		sch->sch_length--;
1176 #ifdef TCP_OFFLOAD
1177 		if (ADDED_BY_TOE(sc)) {
1178 			struct toedev *tod = sc->sc_tod;
1179 
1180 			tod->tod_syncache_removed(tod, sc->sc_todctx);
1181 		}
1182 #endif
1183 		SCH_UNLOCK(sch);
1184 	}
1185 
1186 	/*
1187 	 * Segment validation:
1188 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1189 	 */
1190 	if (th->th_ack != sc->sc_iss + 1) {
1191 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1192 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1193 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1194 		goto failed;
1195 	}
1196 
1197 	/*
1198 	 * The SEQ must fall in the window starting at the received
1199 	 * initial receive sequence number + 1 (the SYN).
1200 	 */
1201 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1202 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1203 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1204 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1205 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1206 		goto failed;
1207 	}
1208 
1209 	/*
1210 	 * If timestamps were not negotiated during SYN/ACK they
1211 	 * must not appear on any segment during this session.
1212 	 */
1213 	if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1214 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1215 			log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1216 			    "segment rejected\n", s, __func__);
1217 		goto failed;
1218 	}
1219 
1220 	/*
1221 	 * If timestamps were negotiated during SYN/ACK they should
1222 	 * appear on every segment during this session.
1223 	 * XXXAO: This is only informal as there have been unverified
1224 	 * reports of non-compliants stacks.
1225 	 */
1226 	if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1227 		if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1228 			log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1229 			    "no action\n", s, __func__);
1230 			free(s, M_TCPLOG);
1231 			s = NULL;
1232 		}
1233 	}
1234 
1235 	*lsop = syncache_socket(sc, *lsop, m);
1236 
1237 	if (*lsop == NULL)
1238 		TCPSTAT_INC(tcps_sc_aborted);
1239 	else
1240 		TCPSTAT_INC(tcps_sc_completed);
1241 
1242 /* how do we find the inp for the new socket? */
1243 	if (sc != &scs)
1244 		syncache_free(sc);
1245 	return (1);
1246 failed:
1247 	if (sc != NULL && sc != &scs)
1248 		syncache_free(sc);
1249 	if (s != NULL)
1250 		free(s, M_TCPLOG);
1251 	*lsop = NULL;
1252 	return (0);
1253 }
1254 
1255 static void
1256 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
1257     uint64_t response_cookie)
1258 {
1259 	struct inpcb *inp;
1260 	struct tcpcb *tp;
1261 	unsigned int *pending_counter;
1262 
1263 	/*
1264 	 * Global TCP locks are held because we manipulate the PCB lists
1265 	 * and create a new socket.
1266 	 */
1267 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1268 
1269 	pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
1270 	*lsop = syncache_socket(sc, *lsop, m);
1271 	if (*lsop == NULL) {
1272 		TCPSTAT_INC(tcps_sc_aborted);
1273 		atomic_subtract_int(pending_counter, 1);
1274 	} else {
1275 		soisconnected(*lsop);
1276 		inp = sotoinpcb(*lsop);
1277 		tp = intotcpcb(inp);
1278 		tp->t_flags |= TF_FASTOPEN;
1279 		tp->t_tfo_cookie.server = response_cookie;
1280 		tp->snd_max = tp->iss;
1281 		tp->snd_nxt = tp->iss;
1282 		tp->t_tfo_pending = pending_counter;
1283 		TCPSTAT_INC(tcps_sc_completed);
1284 	}
1285 }
1286 
1287 /*
1288  * Given a LISTEN socket and an inbound SYN request, add
1289  * this to the syn cache, and send back a segment:
1290  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1291  * to the source.
1292  *
1293  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1294  * Doing so would require that we hold onto the data and deliver it
1295  * to the application.  However, if we are the target of a SYN-flood
1296  * DoS attack, an attacker could send data which would eventually
1297  * consume all available buffer space if it were ACKed.  By not ACKing
1298  * the data, we avoid this DoS scenario.
1299  *
1300  * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1301  * cookie is processed and a new socket is created.  In this case, any data
1302  * accompanying the SYN will be queued to the socket by tcp_input() and will
1303  * be ACKed either when the application sends response data or the delayed
1304  * ACK timer expires, whichever comes first.
1305  */
1306 int
1307 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1308     struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1309     void *todctx)
1310 {
1311 	struct tcpcb *tp;
1312 	struct socket *so;
1313 	struct syncache *sc = NULL;
1314 	struct syncache_head *sch;
1315 	struct mbuf *ipopts = NULL;
1316 	u_int ltflags;
1317 	int win, ip_ttl, ip_tos;
1318 	char *s;
1319 	int rv = 0;
1320 #ifdef INET6
1321 	int autoflowlabel = 0;
1322 #endif
1323 #ifdef MAC
1324 	struct label *maclabel;
1325 #endif
1326 	struct syncache scs;
1327 	struct ucred *cred;
1328 	uint64_t tfo_response_cookie;
1329 	unsigned int *tfo_pending = NULL;
1330 	int tfo_cookie_valid = 0;
1331 	int tfo_response_cookie_valid = 0;
1332 
1333 	INP_WLOCK_ASSERT(inp);			/* listen socket */
1334 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1335 	    ("%s: unexpected tcp flags", __func__));
1336 
1337 	/*
1338 	 * Combine all so/tp operations very early to drop the INP lock as
1339 	 * soon as possible.
1340 	 */
1341 	so = *lsop;
1342 	KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1343 	tp = sototcpcb(so);
1344 	cred = crhold(so->so_cred);
1345 
1346 #ifdef INET6
1347 	if ((inc->inc_flags & INC_ISIPV6) &&
1348 	    (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1349 		autoflowlabel = 1;
1350 #endif
1351 	ip_ttl = inp->inp_ip_ttl;
1352 	ip_tos = inp->inp_ip_tos;
1353 	win = so->sol_sbrcv_hiwat;
1354 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1355 
1356 	if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
1357 	    (tp->t_tfo_pending != NULL) &&
1358 	    (to->to_flags & TOF_FASTOPEN)) {
1359 		/*
1360 		 * Limit the number of pending TFO connections to
1361 		 * approximately half of the queue limit.  This prevents TFO
1362 		 * SYN floods from starving the service by filling the
1363 		 * listen queue with bogus TFO connections.
1364 		 */
1365 		if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1366 		    (so->sol_qlimit / 2)) {
1367 			int result;
1368 
1369 			result = tcp_fastopen_check_cookie(inc,
1370 			    to->to_tfo_cookie, to->to_tfo_len,
1371 			    &tfo_response_cookie);
1372 			tfo_cookie_valid = (result > 0);
1373 			tfo_response_cookie_valid = (result >= 0);
1374 		}
1375 
1376 		/*
1377 		 * Remember the TFO pending counter as it will have to be
1378 		 * decremented below if we don't make it to syncache_tfo_expand().
1379 		 */
1380 		tfo_pending = tp->t_tfo_pending;
1381 	}
1382 
1383 	/* By the time we drop the lock these should no longer be used. */
1384 	so = NULL;
1385 	tp = NULL;
1386 
1387 #ifdef MAC
1388 	if (mac_syncache_init(&maclabel) != 0) {
1389 		INP_WUNLOCK(inp);
1390 		goto done;
1391 	} else
1392 		mac_syncache_create(maclabel, inp);
1393 #endif
1394 	if (!tfo_cookie_valid)
1395 		INP_WUNLOCK(inp);
1396 
1397 	/*
1398 	 * Remember the IP options, if any.
1399 	 */
1400 #ifdef INET6
1401 	if (!(inc->inc_flags & INC_ISIPV6))
1402 #endif
1403 #ifdef INET
1404 		ipopts = (m) ? ip_srcroute(m) : NULL;
1405 #else
1406 		ipopts = NULL;
1407 #endif
1408 
1409 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1410 	/*
1411 	 * If listening socket requested TCP digests, check that received
1412 	 * SYN has signature and it is correct. If signature doesn't match
1413 	 * or TCP_SIGNATURE support isn't enabled, drop the packet.
1414 	 */
1415 	if (ltflags & TF_SIGNATURE) {
1416 		if ((to->to_flags & TOF_SIGNATURE) == 0) {
1417 			TCPSTAT_INC(tcps_sig_err_nosigopt);
1418 			goto done;
1419 		}
1420 		if (!TCPMD5_ENABLED() ||
1421 		    TCPMD5_INPUT(m, th, to->to_signature) != 0)
1422 			goto done;
1423 	}
1424 #endif	/* TCP_SIGNATURE */
1425 	/*
1426 	 * See if we already have an entry for this connection.
1427 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1428 	 *
1429 	 * XXX: should the syncache be re-initialized with the contents
1430 	 * of the new SYN here (which may have different options?)
1431 	 *
1432 	 * XXX: We do not check the sequence number to see if this is a
1433 	 * real retransmit or a new connection attempt.  The question is
1434 	 * how to handle such a case; either ignore it as spoofed, or
1435 	 * drop the current entry and create a new one?
1436 	 */
1437 	sc = syncache_lookup(inc, &sch);	/* returns locked entry */
1438 	SCH_LOCK_ASSERT(sch);
1439 	if (sc != NULL) {
1440 		if (tfo_cookie_valid)
1441 			INP_WUNLOCK(inp);
1442 		TCPSTAT_INC(tcps_sc_dupsyn);
1443 		if (ipopts) {
1444 			/*
1445 			 * If we were remembering a previous source route,
1446 			 * forget it and use the new one we've been given.
1447 			 */
1448 			if (sc->sc_ipopts)
1449 				(void) m_free(sc->sc_ipopts);
1450 			sc->sc_ipopts = ipopts;
1451 		}
1452 		/*
1453 		 * Update timestamp if present.
1454 		 */
1455 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1456 			sc->sc_tsreflect = to->to_tsval;
1457 		else
1458 			sc->sc_flags &= ~SCF_TIMESTAMP;
1459 #ifdef MAC
1460 		/*
1461 		 * Since we have already unconditionally allocated label
1462 		 * storage, free it up.  The syncache entry will already
1463 		 * have an initialized label we can use.
1464 		 */
1465 		mac_syncache_destroy(&maclabel);
1466 #endif
1467 		TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1468 		/* Retransmit SYN|ACK and reset retransmit count. */
1469 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1470 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1471 			    "resetting timer and retransmitting SYN|ACK\n",
1472 			    s, __func__);
1473 			free(s, M_TCPLOG);
1474 		}
1475 		if (syncache_respond(sc, sch, m, TH_SYN|TH_ACK) == 0) {
1476 			sc->sc_rxmits = 0;
1477 			syncache_timeout(sc, sch, 1);
1478 			TCPSTAT_INC(tcps_sndacks);
1479 			TCPSTAT_INC(tcps_sndtotal);
1480 		}
1481 		SCH_UNLOCK(sch);
1482 		goto donenoprobe;
1483 	}
1484 
1485 	if (tfo_cookie_valid) {
1486 		bzero(&scs, sizeof(scs));
1487 		sc = &scs;
1488 		goto skip_alloc;
1489 	}
1490 
1491 	sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1492 	if (sc == NULL) {
1493 		/*
1494 		 * The zone allocator couldn't provide more entries.
1495 		 * Treat this as if the cache was full; drop the oldest
1496 		 * entry and insert the new one.
1497 		 */
1498 		TCPSTAT_INC(tcps_sc_zonefail);
1499 		if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1500 			sch->sch_last_overflow = time_uptime;
1501 			syncache_drop(sc, sch);
1502 		}
1503 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1504 		if (sc == NULL) {
1505 			if (V_tcp_syncookies) {
1506 				bzero(&scs, sizeof(scs));
1507 				sc = &scs;
1508 			} else {
1509 				SCH_UNLOCK(sch);
1510 				if (ipopts)
1511 					(void) m_free(ipopts);
1512 				goto done;
1513 			}
1514 		}
1515 	}
1516 
1517 skip_alloc:
1518 	if (!tfo_cookie_valid && tfo_response_cookie_valid)
1519 		sc->sc_tfo_cookie = &tfo_response_cookie;
1520 
1521 	/*
1522 	 * Fill in the syncache values.
1523 	 */
1524 #ifdef MAC
1525 	sc->sc_label = maclabel;
1526 #endif
1527 	sc->sc_cred = cred;
1528 	cred = NULL;
1529 	sc->sc_ipopts = ipopts;
1530 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1531 #ifdef INET6
1532 	if (!(inc->inc_flags & INC_ISIPV6))
1533 #endif
1534 	{
1535 		sc->sc_ip_tos = ip_tos;
1536 		sc->sc_ip_ttl = ip_ttl;
1537 	}
1538 #ifdef TCP_OFFLOAD
1539 	sc->sc_tod = tod;
1540 	sc->sc_todctx = todctx;
1541 #endif
1542 	sc->sc_irs = th->th_seq;
1543 	sc->sc_iss = arc4random();
1544 	sc->sc_flags = 0;
1545 	sc->sc_flowlabel = 0;
1546 
1547 	/*
1548 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1549 	 * win was derived from socket earlier in the function.
1550 	 */
1551 	win = imax(win, 0);
1552 	win = imin(win, TCP_MAXWIN);
1553 	sc->sc_wnd = win;
1554 
1555 	if (V_tcp_do_rfc1323) {
1556 		/*
1557 		 * A timestamp received in a SYN makes
1558 		 * it ok to send timestamp requests and replies.
1559 		 */
1560 		if (to->to_flags & TOF_TS) {
1561 			sc->sc_tsreflect = to->to_tsval;
1562 			sc->sc_flags |= SCF_TIMESTAMP;
1563 			sc->sc_tsoff = tcp_new_ts_offset(inc);
1564 		}
1565 		if (to->to_flags & TOF_SCALE) {
1566 			int wscale = 0;
1567 
1568 			/*
1569 			 * Pick the smallest possible scaling factor that
1570 			 * will still allow us to scale up to sb_max, aka
1571 			 * kern.ipc.maxsockbuf.
1572 			 *
1573 			 * We do this because there are broken firewalls that
1574 			 * will corrupt the window scale option, leading to
1575 			 * the other endpoint believing that our advertised
1576 			 * window is unscaled.  At scale factors larger than
1577 			 * 5 the unscaled window will drop below 1500 bytes,
1578 			 * leading to serious problems when traversing these
1579 			 * broken firewalls.
1580 			 *
1581 			 * With the default maxsockbuf of 256K, a scale factor
1582 			 * of 3 will be chosen by this algorithm.  Those who
1583 			 * choose a larger maxsockbuf should watch out
1584 			 * for the compatibility problems mentioned above.
1585 			 *
1586 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1587 			 * or <SYN,ACK>) segment itself is never scaled.
1588 			 */
1589 			while (wscale < TCP_MAX_WINSHIFT &&
1590 			    (TCP_MAXWIN << wscale) < sb_max)
1591 				wscale++;
1592 			sc->sc_requested_r_scale = wscale;
1593 			sc->sc_requested_s_scale = to->to_wscale;
1594 			sc->sc_flags |= SCF_WINSCALE;
1595 		}
1596 	}
1597 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1598 	/*
1599 	 * If listening socket requested TCP digests, flag this in the
1600 	 * syncache so that syncache_respond() will do the right thing
1601 	 * with the SYN+ACK.
1602 	 */
1603 	if (ltflags & TF_SIGNATURE)
1604 		sc->sc_flags |= SCF_SIGNATURE;
1605 #endif	/* TCP_SIGNATURE */
1606 	if (to->to_flags & TOF_SACKPERM)
1607 		sc->sc_flags |= SCF_SACK;
1608 	if (to->to_flags & TOF_MSS)
1609 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1610 	if (ltflags & TF_NOOPT)
1611 		sc->sc_flags |= SCF_NOOPT;
1612 	if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1613 		sc->sc_flags |= SCF_ECN;
1614 
1615 	if (V_tcp_syncookies)
1616 		sc->sc_iss = syncookie_generate(sch, sc);
1617 #ifdef INET6
1618 	if (autoflowlabel) {
1619 		if (V_tcp_syncookies)
1620 			sc->sc_flowlabel = sc->sc_iss;
1621 		else
1622 			sc->sc_flowlabel = ip6_randomflowlabel();
1623 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1624 	}
1625 #endif
1626 	SCH_UNLOCK(sch);
1627 
1628 	if (tfo_cookie_valid) {
1629 		syncache_tfo_expand(sc, lsop, m, tfo_response_cookie);
1630 		/* INP_WUNLOCK(inp) will be performed by the caller */
1631 		rv = 1;
1632 		goto tfo_expanded;
1633 	}
1634 
1635 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1636 	/*
1637 	 * Do a standard 3-way handshake.
1638 	 */
1639 	if (syncache_respond(sc, sch, m, TH_SYN|TH_ACK) == 0) {
1640 		if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1641 			syncache_free(sc);
1642 		else if (sc != &scs)
1643 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1644 		TCPSTAT_INC(tcps_sndacks);
1645 		TCPSTAT_INC(tcps_sndtotal);
1646 	} else {
1647 		if (sc != &scs)
1648 			syncache_free(sc);
1649 		TCPSTAT_INC(tcps_sc_dropped);
1650 	}
1651 	goto donenoprobe;
1652 
1653 done:
1654 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1655 donenoprobe:
1656 	if (m) {
1657 		*lsop = NULL;
1658 		m_freem(m);
1659 	}
1660 	/*
1661 	 * If tfo_pending is not NULL here, then a TFO SYN that did not
1662 	 * result in a new socket was processed and the associated pending
1663 	 * counter has not yet been decremented.  All such TFO processing paths
1664 	 * transit this point.
1665 	 */
1666 	if (tfo_pending != NULL)
1667 		tcp_fastopen_decrement_counter(tfo_pending);
1668 
1669 tfo_expanded:
1670 	if (cred != NULL)
1671 		crfree(cred);
1672 #ifdef MAC
1673 	if (sc == &scs)
1674 		mac_syncache_destroy(&maclabel);
1675 #endif
1676 	return (rv);
1677 }
1678 
1679 /*
1680  * Send SYN|ACK or ACK to the peer.  Either in response to a peer's segment,
1681  * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1682  */
1683 static int
1684 syncache_respond(struct syncache *sc, struct syncache_head *sch,
1685     const struct mbuf *m0, int flags)
1686 {
1687 	struct ip *ip = NULL;
1688 	struct mbuf *m;
1689 	struct tcphdr *th = NULL;
1690 	int optlen, error = 0;	/* Make compiler happy */
1691 	u_int16_t hlen, tlen, mssopt;
1692 	struct tcpopt to;
1693 #ifdef INET6
1694 	struct ip6_hdr *ip6 = NULL;
1695 #endif
1696 	hlen =
1697 #ifdef INET6
1698 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1699 #endif
1700 		sizeof(struct ip);
1701 	tlen = hlen + sizeof(struct tcphdr);
1702 
1703 	/* Determine MSS we advertize to other end of connection. */
1704 	mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss);
1705 
1706 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1707 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1708 	    ("syncache: mbuf too small"));
1709 
1710 	/* Create the IP+TCP header from scratch. */
1711 	m = m_gethdr(M_NOWAIT, MT_DATA);
1712 	if (m == NULL)
1713 		return (ENOBUFS);
1714 #ifdef MAC
1715 	mac_syncache_create_mbuf(sc->sc_label, m);
1716 #endif
1717 	m->m_data += max_linkhdr;
1718 	m->m_len = tlen;
1719 	m->m_pkthdr.len = tlen;
1720 	m->m_pkthdr.rcvif = NULL;
1721 
1722 #ifdef INET6
1723 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1724 		ip6 = mtod(m, struct ip6_hdr *);
1725 		ip6->ip6_vfc = IPV6_VERSION;
1726 		ip6->ip6_nxt = IPPROTO_TCP;
1727 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1728 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1729 		ip6->ip6_plen = htons(tlen - hlen);
1730 		/* ip6_hlim is set after checksum */
1731 		ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1732 		ip6->ip6_flow |= sc->sc_flowlabel;
1733 
1734 		th = (struct tcphdr *)(ip6 + 1);
1735 	}
1736 #endif
1737 #if defined(INET6) && defined(INET)
1738 	else
1739 #endif
1740 #ifdef INET
1741 	{
1742 		ip = mtod(m, struct ip *);
1743 		ip->ip_v = IPVERSION;
1744 		ip->ip_hl = sizeof(struct ip) >> 2;
1745 		ip->ip_len = htons(tlen);
1746 		ip->ip_id = 0;
1747 		ip->ip_off = 0;
1748 		ip->ip_sum = 0;
1749 		ip->ip_p = IPPROTO_TCP;
1750 		ip->ip_src = sc->sc_inc.inc_laddr;
1751 		ip->ip_dst = sc->sc_inc.inc_faddr;
1752 		ip->ip_ttl = sc->sc_ip_ttl;
1753 		ip->ip_tos = sc->sc_ip_tos;
1754 
1755 		/*
1756 		 * See if we should do MTU discovery.  Route lookups are
1757 		 * expensive, so we will only unset the DF bit if:
1758 		 *
1759 		 *	1) path_mtu_discovery is disabled
1760 		 *	2) the SCF_UNREACH flag has been set
1761 		 */
1762 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1763 		       ip->ip_off |= htons(IP_DF);
1764 
1765 		th = (struct tcphdr *)(ip + 1);
1766 	}
1767 #endif /* INET */
1768 	th->th_sport = sc->sc_inc.inc_lport;
1769 	th->th_dport = sc->sc_inc.inc_fport;
1770 
1771 	if (flags & TH_SYN)
1772 		th->th_seq = htonl(sc->sc_iss);
1773 	else
1774 		th->th_seq = htonl(sc->sc_iss + 1);
1775 	th->th_ack = htonl(sc->sc_irs + 1);
1776 	th->th_off = sizeof(struct tcphdr) >> 2;
1777 	th->th_x2 = 0;
1778 	th->th_flags = flags;
1779 	th->th_win = htons(sc->sc_wnd);
1780 	th->th_urp = 0;
1781 
1782 	if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) {
1783 		th->th_flags |= TH_ECE;
1784 		TCPSTAT_INC(tcps_ecn_shs);
1785 	}
1786 
1787 	/* Tack on the TCP options. */
1788 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1789 		to.to_flags = 0;
1790 
1791 		if (flags & TH_SYN) {
1792 			to.to_mss = mssopt;
1793 			to.to_flags = TOF_MSS;
1794 			if (sc->sc_flags & SCF_WINSCALE) {
1795 				to.to_wscale = sc->sc_requested_r_scale;
1796 				to.to_flags |= TOF_SCALE;
1797 			}
1798 			if (sc->sc_flags & SCF_SACK)
1799 				to.to_flags |= TOF_SACKPERM;
1800 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1801 			if (sc->sc_flags & SCF_SIGNATURE)
1802 				to.to_flags |= TOF_SIGNATURE;
1803 #endif
1804 			if (sc->sc_tfo_cookie) {
1805 				to.to_flags |= TOF_FASTOPEN;
1806 				to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1807 				to.to_tfo_cookie = sc->sc_tfo_cookie;
1808 				/* don't send cookie again when retransmitting response */
1809 				sc->sc_tfo_cookie = NULL;
1810 			}
1811 		}
1812 		if (sc->sc_flags & SCF_TIMESTAMP) {
1813 			to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1814 			to.to_tsecr = sc->sc_tsreflect;
1815 			to.to_flags |= TOF_TS;
1816 		}
1817 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1818 
1819 		/* Adjust headers by option size. */
1820 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1821 		m->m_len += optlen;
1822 		m->m_pkthdr.len += optlen;
1823 #ifdef INET6
1824 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1825 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1826 		else
1827 #endif
1828 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1829 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1830 		if (sc->sc_flags & SCF_SIGNATURE) {
1831 			KASSERT(to.to_flags & TOF_SIGNATURE,
1832 			    ("tcp_addoptions() didn't set tcp_signature"));
1833 
1834 			/* NOTE: to.to_signature is inside of mbuf */
1835 			if (!TCPMD5_ENABLED() ||
1836 			    TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1837 				m_freem(m);
1838 				return (EACCES);
1839 			}
1840 		}
1841 #endif
1842 	} else
1843 		optlen = 0;
1844 
1845 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
1846 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1847 	/*
1848 	 * If we have peer's SYN and it has a flowid, then let's assign it to
1849 	 * our SYN|ACK.  ip6_output() and ip_output() will not assign flowid
1850 	 * to SYN|ACK due to lack of inp here.
1851 	 */
1852 	if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1853 		m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1854 		M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1855 	}
1856 #ifdef INET6
1857 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1858 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1859 		th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1860 		    IPPROTO_TCP, 0);
1861 		ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1862 #ifdef TCP_OFFLOAD
1863 		if (ADDED_BY_TOE(sc)) {
1864 			struct toedev *tod = sc->sc_tod;
1865 
1866 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1867 
1868 			return (error);
1869 		}
1870 #endif
1871 		TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
1872 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1873 	}
1874 #endif
1875 #if defined(INET6) && defined(INET)
1876 	else
1877 #endif
1878 #ifdef INET
1879 	{
1880 		m->m_pkthdr.csum_flags = CSUM_TCP;
1881 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1882 		    htons(tlen + optlen - hlen + IPPROTO_TCP));
1883 #ifdef TCP_OFFLOAD
1884 		if (ADDED_BY_TOE(sc)) {
1885 			struct toedev *tod = sc->sc_tod;
1886 
1887 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1888 
1889 			return (error);
1890 		}
1891 #endif
1892 		TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
1893 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1894 	}
1895 #endif
1896 	return (error);
1897 }
1898 
1899 /*
1900  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1901  * that exceed the capacity of the syncache by avoiding the storage of any
1902  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
1903  * attacks where the attacker does not have access to our responses.
1904  *
1905  * Syncookies encode and include all necessary information about the
1906  * connection setup within the SYN|ACK that we send back.  That way we
1907  * can avoid keeping any local state until the ACK to our SYN|ACK returns
1908  * (if ever).  Normally the syncache and syncookies are running in parallel
1909  * with the latter taking over when the former is exhausted.  When matching
1910  * syncache entry is found the syncookie is ignored.
1911  *
1912  * The only reliable information persisting the 3WHS is our initial sequence
1913  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
1914  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1915  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
1916  * returns and signifies a legitimate connection if it matches the ACK.
1917  *
1918  * The available space of 32 bits to store the hash and to encode the SYN
1919  * option information is very tight and we should have at least 24 bits for
1920  * the MAC to keep the number of guesses by blind spoofing reasonably high.
1921  *
1922  * SYN option information we have to encode to fully restore a connection:
1923  * MSS: is imporant to chose an optimal segment size to avoid IP level
1924  *   fragmentation along the path.  The common MSS values can be encoded
1925  *   in a 3-bit table.  Uncommon values are captured by the next lower value
1926  *   in the table leading to a slight increase in packetization overhead.
1927  * WSCALE: is necessary to allow large windows to be used for high delay-
1928  *   bandwidth product links.  Not scaling the window when it was initially
1929  *   negotiated is bad for performance as lack of scaling further decreases
1930  *   the apparent available send window.  We only need to encode the WSCALE
1931  *   we received from the remote end.  Our end can be recalculated at any
1932  *   time.  The common WSCALE values can be encoded in a 3-bit table.
1933  *   Uncommon values are captured by the next lower value in the table
1934  *   making us under-estimate the available window size halving our
1935  *   theoretically possible maximum throughput for that connection.
1936  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
1937  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
1938  *   that are included in all segments on a connection.  We enable them when
1939  *   the ACK has them.
1940  *
1941  * Security of syncookies and attack vectors:
1942  *
1943  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
1944  * together with the gloabl secret to make it unique per connection attempt.
1945  * Thus any change of any of those parameters results in a different MAC output
1946  * in an unpredictable way unless a collision is encountered.  24 bits of the
1947  * MAC are embedded into the ISS.
1948  *
1949  * To prevent replay attacks two rotating global secrets are updated with a
1950  * new random value every 15 seconds.  The life-time of a syncookie is thus
1951  * 15-30 seconds.
1952  *
1953  * Vector 1: Attacking the secret.  This requires finding a weakness in the
1954  * MAC itself or the way it is used here.  The attacker can do a chosen plain
1955  * text attack by varying and testing the all parameters under his control.
1956  * The strength depends on the size and randomness of the secret, and the
1957  * cryptographic security of the MAC function.  Due to the constant updating
1958  * of the secret the attacker has at most 29.999 seconds to find the secret
1959  * and launch spoofed connections.  After that he has to start all over again.
1960  *
1961  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
1962  * size an average of 4,823 attempts are required for a 50% chance of success
1963  * to spoof a single syncookie (birthday collision paradox).  However the
1964  * attacker is blind and doesn't know if one of his attempts succeeded unless
1965  * he has a side channel to interfere success from.  A single connection setup
1966  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
1967  * This many attempts are required for each one blind spoofed connection.  For
1968  * every additional spoofed connection he has to launch another N attempts.
1969  * Thus for a sustained rate 100 spoofed connections per second approximately
1970  * 1,800,000 packets per second would have to be sent.
1971  *
1972  * NB: The MAC function should be fast so that it doesn't become a CPU
1973  * exhaustion attack vector itself.
1974  *
1975  * References:
1976  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
1977  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
1978  *   http://cr.yp.to/syncookies.html    (overview)
1979  *   http://cr.yp.to/syncookies/archive (details)
1980  *
1981  *
1982  * Schematic construction of a syncookie enabled Initial Sequence Number:
1983  *  0        1         2         3
1984  *  12345678901234567890123456789012
1985  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
1986  *
1987  *  x 24 MAC (truncated)
1988  *  W  3 Send Window Scale index
1989  *  M  3 MSS index
1990  *  S  1 SACK permitted
1991  *  P  1 Odd/even secret
1992  */
1993 
1994 /*
1995  * Distribution and probability of certain MSS values.  Those in between are
1996  * rounded down to the next lower one.
1997  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
1998  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
1999  */
2000 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2001 
2002 /*
2003  * Distribution and probability of certain WSCALE values.  We have to map the
2004  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2005  * bits based on prevalence of certain values.  Where we don't have an exact
2006  * match for are rounded down to the next lower one letting us under-estimate
2007  * the true available window.  At the moment this would happen only for the
2008  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2009  * and window size).  The absence of the WSCALE option (no scaling in either
2010  * direction) is encoded with index zero.
2011  * [WSCALE values histograms, Allman, 2012]
2012  *                            X 10 10 35  5  6 14 10%   by host
2013  *                            X 11  4  5  5 18 49  3%   by connections
2014  */
2015 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2016 
2017 /*
2018  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
2019  * and good cryptographic properties.
2020  */
2021 static uint32_t
2022 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2023     uint8_t *secbits, uintptr_t secmod)
2024 {
2025 	SIPHASH_CTX ctx;
2026 	uint32_t siphash[2];
2027 
2028 	SipHash24_Init(&ctx);
2029 	SipHash_SetKey(&ctx, secbits);
2030 	switch (inc->inc_flags & INC_ISIPV6) {
2031 #ifdef INET
2032 	case 0:
2033 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2034 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2035 		break;
2036 #endif
2037 #ifdef INET6
2038 	case INC_ISIPV6:
2039 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2040 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2041 		break;
2042 #endif
2043 	}
2044 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2045 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2046 	SipHash_Update(&ctx, &irs, sizeof(irs));
2047 	SipHash_Update(&ctx, &flags, sizeof(flags));
2048 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
2049 	SipHash_Final((u_int8_t *)&siphash, &ctx);
2050 
2051 	return (siphash[0] ^ siphash[1]);
2052 }
2053 
2054 static tcp_seq
2055 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2056 {
2057 	u_int i, secbit, wscale;
2058 	uint32_t iss, hash;
2059 	uint8_t *secbits;
2060 	union syncookie cookie;
2061 
2062 	SCH_LOCK_ASSERT(sch);
2063 
2064 	cookie.cookie = 0;
2065 
2066 	/* Map our computed MSS into the 3-bit index. */
2067 	for (i = nitems(tcp_sc_msstab) - 1;
2068 	     tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2069 	     i--)
2070 		;
2071 	cookie.flags.mss_idx = i;
2072 
2073 	/*
2074 	 * Map the send window scale into the 3-bit index but only if
2075 	 * the wscale option was received.
2076 	 */
2077 	if (sc->sc_flags & SCF_WINSCALE) {
2078 		wscale = sc->sc_requested_s_scale;
2079 		for (i = nitems(tcp_sc_wstab) - 1;
2080 		    tcp_sc_wstab[i] > wscale && i > 0;
2081 		     i--)
2082 			;
2083 		cookie.flags.wscale_idx = i;
2084 	}
2085 
2086 	/* Can we do SACK? */
2087 	if (sc->sc_flags & SCF_SACK)
2088 		cookie.flags.sack_ok = 1;
2089 
2090 	/* Which of the two secrets to use. */
2091 	secbit = sch->sch_sc->secret.oddeven & 0x1;
2092 	cookie.flags.odd_even = secbit;
2093 
2094 	secbits = sch->sch_sc->secret.key[secbit];
2095 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2096 	    (uintptr_t)sch);
2097 
2098 	/*
2099 	 * Put the flags into the hash and XOR them to get better ISS number
2100 	 * variance.  This doesn't enhance the cryptographic strength and is
2101 	 * done to prevent the 8 cookie bits from showing up directly on the
2102 	 * wire.
2103 	 */
2104 	iss = hash & ~0xff;
2105 	iss |= cookie.cookie ^ (hash >> 24);
2106 
2107 	TCPSTAT_INC(tcps_sc_sendcookie);
2108 	return (iss);
2109 }
2110 
2111 static struct syncache *
2112 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2113     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2114     struct socket *lso)
2115 {
2116 	uint32_t hash;
2117 	uint8_t *secbits;
2118 	tcp_seq ack, seq;
2119 	int wnd, wscale = 0;
2120 	union syncookie cookie;
2121 
2122 	SCH_LOCK_ASSERT(sch);
2123 
2124 	/*
2125 	 * Pull information out of SYN-ACK/ACK and revert sequence number
2126 	 * advances.
2127 	 */
2128 	ack = th->th_ack - 1;
2129 	seq = th->th_seq - 1;
2130 
2131 	/*
2132 	 * Unpack the flags containing enough information to restore the
2133 	 * connection.
2134 	 */
2135 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2136 
2137 	/* Which of the two secrets to use. */
2138 	secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
2139 
2140 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2141 
2142 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
2143 	if ((ack & ~0xff) != (hash & ~0xff))
2144 		return (NULL);
2145 
2146 	/* Fill in the syncache values. */
2147 	sc->sc_flags = 0;
2148 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2149 	sc->sc_ipopts = NULL;
2150 
2151 	sc->sc_irs = seq;
2152 	sc->sc_iss = ack;
2153 
2154 	switch (inc->inc_flags & INC_ISIPV6) {
2155 #ifdef INET
2156 	case 0:
2157 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2158 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2159 		break;
2160 #endif
2161 #ifdef INET6
2162 	case INC_ISIPV6:
2163 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2164 			sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
2165 		break;
2166 #endif
2167 	}
2168 
2169 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2170 
2171 	/* We can simply recompute receive window scale we sent earlier. */
2172 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2173 		wscale++;
2174 
2175 	/* Only use wscale if it was enabled in the orignal SYN. */
2176 	if (cookie.flags.wscale_idx > 0) {
2177 		sc->sc_requested_r_scale = wscale;
2178 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2179 		sc->sc_flags |= SCF_WINSCALE;
2180 	}
2181 
2182 	wnd = lso->sol_sbrcv_hiwat;
2183 	wnd = imax(wnd, 0);
2184 	wnd = imin(wnd, TCP_MAXWIN);
2185 	sc->sc_wnd = wnd;
2186 
2187 	if (cookie.flags.sack_ok)
2188 		sc->sc_flags |= SCF_SACK;
2189 
2190 	if (to->to_flags & TOF_TS) {
2191 		sc->sc_flags |= SCF_TIMESTAMP;
2192 		sc->sc_tsreflect = to->to_tsval;
2193 		sc->sc_tsoff = tcp_new_ts_offset(inc);
2194 	}
2195 
2196 	if (to->to_flags & TOF_SIGNATURE)
2197 		sc->sc_flags |= SCF_SIGNATURE;
2198 
2199 	sc->sc_rxmits = 0;
2200 
2201 	TCPSTAT_INC(tcps_sc_recvcookie);
2202 	return (sc);
2203 }
2204 
2205 #ifdef INVARIANTS
2206 static int
2207 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2208     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2209     struct socket *lso)
2210 {
2211 	struct syncache scs, *scx;
2212 	char *s;
2213 
2214 	bzero(&scs, sizeof(scs));
2215 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
2216 
2217 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2218 		return (0);
2219 
2220 	if (scx != NULL) {
2221 		if (sc->sc_peer_mss != scx->sc_peer_mss)
2222 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2223 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2224 
2225 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2226 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2227 			    s, __func__, sc->sc_requested_r_scale,
2228 			    scx->sc_requested_r_scale);
2229 
2230 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2231 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2232 			    s, __func__, sc->sc_requested_s_scale,
2233 			    scx->sc_requested_s_scale);
2234 
2235 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2236 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2237 	}
2238 
2239 	if (s != NULL)
2240 		free(s, M_TCPLOG);
2241 	return (0);
2242 }
2243 #endif /* INVARIANTS */
2244 
2245 static void
2246 syncookie_reseed(void *arg)
2247 {
2248 	struct tcp_syncache *sc = arg;
2249 	uint8_t *secbits;
2250 	int secbit;
2251 
2252 	/*
2253 	 * Reseeding the secret doesn't have to be protected by a lock.
2254 	 * It only must be ensured that the new random values are visible
2255 	 * to all CPUs in a SMP environment.  The atomic with release
2256 	 * semantics ensures that.
2257 	 */
2258 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2259 	secbits = sc->secret.key[secbit];
2260 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2261 	atomic_add_rel_int(&sc->secret.oddeven, 1);
2262 
2263 	/* Reschedule ourself. */
2264 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2265 }
2266 
2267 /*
2268  * Exports the syncache entries to userland so that netstat can display
2269  * them alongside the other sockets.  This function is intended to be
2270  * called only from tcp_pcblist.
2271  *
2272  * Due to concurrency on an active system, the number of pcbs exported
2273  * may have no relation to max_pcbs.  max_pcbs merely indicates the
2274  * amount of space the caller allocated for this function to use.
2275  */
2276 int
2277 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
2278 {
2279 	struct xtcpcb xt;
2280 	struct syncache *sc;
2281 	struct syncache_head *sch;
2282 	int count, error, i;
2283 
2284 	for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
2285 		sch = &V_tcp_syncache.hashbase[i];
2286 		SCH_LOCK(sch);
2287 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2288 			if (count >= max_pcbs) {
2289 				SCH_UNLOCK(sch);
2290 				goto exit;
2291 			}
2292 			if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2293 				continue;
2294 			bzero(&xt, sizeof(xt));
2295 			xt.xt_len = sizeof(xt);
2296 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2297 				xt.xt_inp.inp_vflag = INP_IPV6;
2298 			else
2299 				xt.xt_inp.inp_vflag = INP_IPV4;
2300 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2301 			    sizeof (struct in_conninfo));
2302 			xt.t_state = TCPS_SYN_RECEIVED;
2303 			xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2304 			xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2305 			xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2306 			xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2307 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2308 			if (error) {
2309 				SCH_UNLOCK(sch);
2310 				goto exit;
2311 			}
2312 			count++;
2313 		}
2314 		SCH_UNLOCK(sch);
2315 	}
2316 exit:
2317 	*pcbs_exported = count;
2318 	return error;
2319 }
2320