xref: /freebsd/sys/netinet/tcp_syncache.c (revision 3cbb4cc200f8a0ad7ed08233425ea54524a21f1c)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2001 McAfee, Inc.
5  * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG
6  * All rights reserved.
7  *
8  * This software was developed for the FreeBSD Project by Jonathan Lemon
9  * and McAfee Research, the Security Research Division of McAfee, Inc. under
10  * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
11  * DARPA CHATS research program. [2001 McAfee, Inc.]
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 #include "opt_inet.h"
39 #include "opt_inet6.h"
40 #include "opt_ipsec.h"
41 #include "opt_pcbgroup.h"
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/hash.h>
46 #include <sys/refcount.h>
47 #include <sys/kernel.h>
48 #include <sys/sysctl.h>
49 #include <sys/limits.h>
50 #include <sys/lock.h>
51 #include <sys/mutex.h>
52 #include <sys/malloc.h>
53 #include <sys/mbuf.h>
54 #include <sys/proc.h>		/* for proc0 declaration */
55 #include <sys/random.h>
56 #include <sys/socket.h>
57 #include <sys/socketvar.h>
58 #include <sys/syslog.h>
59 #include <sys/ucred.h>
60 
61 #include <sys/md5.h>
62 #include <crypto/siphash/siphash.h>
63 
64 #include <vm/uma.h>
65 
66 #include <net/if.h>
67 #include <net/if_var.h>
68 #include <net/route.h>
69 #include <net/vnet.h>
70 
71 #include <netinet/in.h>
72 #include <netinet/in_kdtrace.h>
73 #include <netinet/in_systm.h>
74 #include <netinet/ip.h>
75 #include <netinet/in_var.h>
76 #include <netinet/in_pcb.h>
77 #include <netinet/ip_var.h>
78 #include <netinet/ip_options.h>
79 #ifdef INET6
80 #include <netinet/ip6.h>
81 #include <netinet/icmp6.h>
82 #include <netinet6/nd6.h>
83 #include <netinet6/ip6_var.h>
84 #include <netinet6/in6_pcb.h>
85 #endif
86 #include <netinet/tcp.h>
87 #include <netinet/tcp_fastopen.h>
88 #include <netinet/tcp_fsm.h>
89 #include <netinet/tcp_seq.h>
90 #include <netinet/tcp_timer.h>
91 #include <netinet/tcp_var.h>
92 #include <netinet/tcp_syncache.h>
93 #ifdef INET6
94 #include <netinet6/tcp6_var.h>
95 #endif
96 #ifdef TCP_OFFLOAD
97 #include <netinet/toecore.h>
98 #endif
99 
100 #include <netipsec/ipsec_support.h>
101 
102 #include <machine/in_cksum.h>
103 
104 #include <security/mac/mac_framework.h>
105 
106 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1;
107 #define	V_tcp_syncookies		VNET(tcp_syncookies)
108 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
109     &VNET_NAME(tcp_syncookies), 0,
110     "Use TCP SYN cookies if the syncache overflows");
111 
112 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0;
113 #define	V_tcp_syncookiesonly		VNET(tcp_syncookiesonly)
114 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
115     &VNET_NAME(tcp_syncookiesonly), 0,
116     "Use only TCP SYN cookies");
117 
118 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1;
119 #define V_functions_inherit_listen_socket_stack \
120     VNET(functions_inherit_listen_socket_stack)
121 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack,
122     CTLFLAG_VNET | CTLFLAG_RW,
123     &VNET_NAME(functions_inherit_listen_socket_stack), 0,
124     "Inherit listen socket's stack");
125 
126 #ifdef TCP_OFFLOAD
127 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
128 #endif
129 
130 static void	 syncache_drop(struct syncache *, struct syncache_head *);
131 static void	 syncache_free(struct syncache *);
132 static void	 syncache_insert(struct syncache *, struct syncache_head *);
133 static int	 syncache_respond(struct syncache *, const struct mbuf *, int);
134 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
135 		    struct mbuf *m);
136 static void	 syncache_timeout(struct syncache *sc, struct syncache_head *sch,
137 		    int docallout);
138 static void	 syncache_timer(void *);
139 
140 static uint32_t	 syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
141 		    uint8_t *, uintptr_t);
142 static tcp_seq	 syncookie_generate(struct syncache_head *, struct syncache *);
143 static struct syncache
144 		*syncookie_lookup(struct in_conninfo *, struct syncache_head *,
145 		    struct syncache *, struct tcphdr *, struct tcpopt *,
146 		    struct socket *);
147 static void	syncache_pause(struct in_conninfo *);
148 static void	syncache_unpause(void *);
149 static void	 syncookie_reseed(void *);
150 #ifdef INVARIANTS
151 static int	 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
152 		    struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
153 		    struct socket *lso);
154 #endif
155 
156 /*
157  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
158  * 3 retransmits corresponds to a timeout with default values of
159  * tcp_rexmit_initial * (             1 +
160  *                       tcp_backoff[1] +
161  *                       tcp_backoff[2] +
162  *                       tcp_backoff[3]) + 3 * tcp_rexmit_slop,
163  * 1000 ms * (1 + 2 + 4 + 8) +  3 * 200 ms = 15600 ms,
164  * the odds are that the user has given up attempting to connect by then.
165  */
166 #define SYNCACHE_MAXREXMTS		3
167 
168 /* Arbitrary values */
169 #define TCP_SYNCACHE_HASHSIZE		512
170 #define TCP_SYNCACHE_BUCKETLIMIT	30
171 
172 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
173 #define	V_tcp_syncache			VNET(tcp_syncache)
174 
175 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache,
176     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
177     "TCP SYN cache");
178 
179 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
180     &VNET_NAME(tcp_syncache.bucket_limit), 0,
181     "Per-bucket hash limit for syncache");
182 
183 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
184     &VNET_NAME(tcp_syncache.cache_limit), 0,
185     "Overall entry limit for syncache");
186 
187 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
188     &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
189 
190 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
191     &VNET_NAME(tcp_syncache.hashsize), 0,
192     "Size of TCP syncache hashtable");
193 
194 static int
195 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
196 {
197 	int error;
198 	u_int new;
199 
200 	new = V_tcp_syncache.rexmt_limit;
201 	error = sysctl_handle_int(oidp, &new, 0, req);
202 	if ((error == 0) && (req->newptr != NULL)) {
203 		if (new > TCP_MAXRXTSHIFT)
204 			error = EINVAL;
205 		else
206 			V_tcp_syncache.rexmt_limit = new;
207 	}
208 	return (error);
209 }
210 
211 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
212     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
213     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
214     sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
215     "Limit on SYN/ACK retransmissions");
216 
217 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
218 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
219     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
220     "Send reset on socket allocation failure");
221 
222 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
223 
224 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
225 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
226 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
227 
228 /*
229  * Requires the syncache entry to be already removed from the bucket list.
230  */
231 static void
232 syncache_free(struct syncache *sc)
233 {
234 
235 	if (sc->sc_ipopts)
236 		(void) m_free(sc->sc_ipopts);
237 	if (sc->sc_cred)
238 		crfree(sc->sc_cred);
239 #ifdef MAC
240 	mac_syncache_destroy(&sc->sc_label);
241 #endif
242 
243 	uma_zfree(V_tcp_syncache.zone, sc);
244 }
245 
246 void
247 syncache_init(void)
248 {
249 	int i;
250 
251 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
252 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
253 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
254 	V_tcp_syncache.hash_secret = arc4random();
255 
256 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
257 	    &V_tcp_syncache.hashsize);
258 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
259 	    &V_tcp_syncache.bucket_limit);
260 	if (!powerof2(V_tcp_syncache.hashsize) ||
261 	    V_tcp_syncache.hashsize == 0) {
262 		printf("WARNING: syncache hash size is not a power of 2.\n");
263 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
264 	}
265 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
266 
267 	/* Set limits. */
268 	V_tcp_syncache.cache_limit =
269 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
270 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
271 	    &V_tcp_syncache.cache_limit);
272 
273 	/* Allocate the hash table. */
274 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
275 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
276 
277 #ifdef VIMAGE
278 	V_tcp_syncache.vnet = curvnet;
279 #endif
280 
281 	/* Initialize the hash buckets. */
282 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
283 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
284 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
285 			 NULL, MTX_DEF);
286 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
287 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
288 		V_tcp_syncache.hashbase[i].sch_length = 0;
289 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
290 		V_tcp_syncache.hashbase[i].sch_last_overflow =
291 		    -(SYNCOOKIE_LIFETIME + 1);
292 	}
293 
294 	/* Create the syncache entry zone. */
295 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
296 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
297 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
298 	    V_tcp_syncache.cache_limit);
299 
300 	/* Start the SYN cookie reseeder callout. */
301 	callout_init(&V_tcp_syncache.secret.reseed, 1);
302 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
303 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
304 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
305 	    syncookie_reseed, &V_tcp_syncache);
306 
307 	/* Initialize the pause machinery. */
308 	mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF);
309 	callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx,
310 	    0);
311 	V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME;
312 	V_tcp_syncache.pause_backoff = 0;
313 	V_tcp_syncache.paused = false;
314 }
315 
316 #ifdef VIMAGE
317 void
318 syncache_destroy(void)
319 {
320 	struct syncache_head *sch;
321 	struct syncache *sc, *nsc;
322 	int i;
323 
324 	/*
325 	 * Stop the re-seed timer before freeing resources.  No need to
326 	 * possibly schedule it another time.
327 	 */
328 	callout_drain(&V_tcp_syncache.secret.reseed);
329 
330 	/* Stop the SYN cache pause callout. */
331 	mtx_lock(&V_tcp_syncache.pause_mtx);
332 	if (callout_stop(&V_tcp_syncache.pause_co) == 0) {
333 		mtx_unlock(&V_tcp_syncache.pause_mtx);
334 		callout_drain(&V_tcp_syncache.pause_co);
335 	} else
336 		mtx_unlock(&V_tcp_syncache.pause_mtx);
337 
338 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
339 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
340 		sch = &V_tcp_syncache.hashbase[i];
341 		callout_drain(&sch->sch_timer);
342 
343 		SCH_LOCK(sch);
344 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
345 			syncache_drop(sc, sch);
346 		SCH_UNLOCK(sch);
347 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
348 		    ("%s: sch->sch_bucket not empty", __func__));
349 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
350 		    __func__, sch->sch_length));
351 		mtx_destroy(&sch->sch_mtx);
352 	}
353 
354 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
355 	    ("%s: cache_count not 0", __func__));
356 
357 	/* Free the allocated global resources. */
358 	uma_zdestroy(V_tcp_syncache.zone);
359 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
360 	mtx_destroy(&V_tcp_syncache.pause_mtx);
361 }
362 #endif
363 
364 /*
365  * Inserts a syncache entry into the specified bucket row.
366  * Locks and unlocks the syncache_head autonomously.
367  */
368 static void
369 syncache_insert(struct syncache *sc, struct syncache_head *sch)
370 {
371 	struct syncache *sc2;
372 
373 	SCH_LOCK(sch);
374 
375 	/*
376 	 * Make sure that we don't overflow the per-bucket limit.
377 	 * If the bucket is full, toss the oldest element.
378 	 */
379 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
380 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
381 			("sch->sch_length incorrect"));
382 		syncache_pause(&sc->sc_inc);
383 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
384 		sch->sch_last_overflow = time_uptime;
385 		syncache_drop(sc2, sch);
386 	}
387 
388 	/* Put it into the bucket. */
389 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
390 	sch->sch_length++;
391 
392 #ifdef TCP_OFFLOAD
393 	if (ADDED_BY_TOE(sc)) {
394 		struct toedev *tod = sc->sc_tod;
395 
396 		tod->tod_syncache_added(tod, sc->sc_todctx);
397 	}
398 #endif
399 
400 	/* Reinitialize the bucket row's timer. */
401 	if (sch->sch_length == 1)
402 		sch->sch_nextc = ticks + INT_MAX;
403 	syncache_timeout(sc, sch, 1);
404 
405 	SCH_UNLOCK(sch);
406 
407 	TCPSTATES_INC(TCPS_SYN_RECEIVED);
408 	TCPSTAT_INC(tcps_sc_added);
409 }
410 
411 /*
412  * Remove and free entry from syncache bucket row.
413  * Expects locked syncache head.
414  */
415 static void
416 syncache_drop(struct syncache *sc, struct syncache_head *sch)
417 {
418 
419 	SCH_LOCK_ASSERT(sch);
420 
421 	TCPSTATES_DEC(TCPS_SYN_RECEIVED);
422 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
423 	sch->sch_length--;
424 
425 #ifdef TCP_OFFLOAD
426 	if (ADDED_BY_TOE(sc)) {
427 		struct toedev *tod = sc->sc_tod;
428 
429 		tod->tod_syncache_removed(tod, sc->sc_todctx);
430 	}
431 #endif
432 
433 	syncache_free(sc);
434 }
435 
436 /*
437  * Engage/reengage time on bucket row.
438  */
439 static void
440 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
441 {
442 	int rexmt;
443 
444 	if (sc->sc_rxmits == 0)
445 		rexmt = tcp_rexmit_initial;
446 	else
447 		TCPT_RANGESET(rexmt,
448 		    tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
449 		    tcp_rexmit_min, TCPTV_REXMTMAX);
450 	sc->sc_rxttime = ticks + rexmt;
451 	sc->sc_rxmits++;
452 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
453 		sch->sch_nextc = sc->sc_rxttime;
454 		if (docallout)
455 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
456 			    syncache_timer, (void *)sch);
457 	}
458 }
459 
460 /*
461  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
462  * If we have retransmitted an entry the maximum number of times, expire it.
463  * One separate timer for each bucket row.
464  */
465 static void
466 syncache_timer(void *xsch)
467 {
468 	struct syncache_head *sch = (struct syncache_head *)xsch;
469 	struct syncache *sc, *nsc;
470 	struct epoch_tracker et;
471 	int tick = ticks;
472 	char *s;
473 	bool paused;
474 
475 	CURVNET_SET(sch->sch_sc->vnet);
476 
477 	/* NB: syncache_head has already been locked by the callout. */
478 	SCH_LOCK_ASSERT(sch);
479 
480 	/*
481 	 * In the following cycle we may remove some entries and/or
482 	 * advance some timeouts, so re-initialize the bucket timer.
483 	 */
484 	sch->sch_nextc = tick + INT_MAX;
485 
486 	/*
487 	 * If we have paused processing, unconditionally remove
488 	 * all syncache entries.
489 	 */
490 	mtx_lock(&V_tcp_syncache.pause_mtx);
491 	paused = V_tcp_syncache.paused;
492 	mtx_unlock(&V_tcp_syncache.pause_mtx);
493 
494 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
495 		if (paused) {
496 			syncache_drop(sc, sch);
497 			continue;
498 		}
499 		/*
500 		 * We do not check if the listen socket still exists
501 		 * and accept the case where the listen socket may be
502 		 * gone by the time we resend the SYN/ACK.  We do
503 		 * not expect this to happens often. If it does,
504 		 * then the RST will be sent by the time the remote
505 		 * host does the SYN/ACK->ACK.
506 		 */
507 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
508 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
509 				sch->sch_nextc = sc->sc_rxttime;
510 			continue;
511 		}
512 		if (sc->sc_rxmits > V_tcp_ecn_maxretries) {
513 			sc->sc_flags &= ~SCF_ECN;
514 		}
515 		if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
516 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
517 				log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
518 				    "giving up and removing syncache entry\n",
519 				    s, __func__);
520 				free(s, M_TCPLOG);
521 			}
522 			syncache_drop(sc, sch);
523 			TCPSTAT_INC(tcps_sc_stale);
524 			continue;
525 		}
526 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
527 			log(LOG_DEBUG, "%s; %s: Response timeout, "
528 			    "retransmitting (%u) SYN|ACK\n",
529 			    s, __func__, sc->sc_rxmits);
530 			free(s, M_TCPLOG);
531 		}
532 
533 		NET_EPOCH_ENTER(et);
534 		syncache_respond(sc, NULL, TH_SYN|TH_ACK);
535 		NET_EPOCH_EXIT(et);
536 		TCPSTAT_INC(tcps_sc_retransmitted);
537 		syncache_timeout(sc, sch, 0);
538 	}
539 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
540 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
541 			syncache_timer, (void *)(sch));
542 	CURVNET_RESTORE();
543 }
544 
545 /*
546  * Returns true if the system is only using cookies at the moment.
547  * This could be due to a sysadmin decision to only use cookies, or it
548  * could be due to the system detecting an attack.
549  */
550 static inline bool
551 syncache_cookiesonly(void)
552 {
553 
554 	return (V_tcp_syncookies && (V_tcp_syncache.paused ||
555 	    V_tcp_syncookiesonly));
556 }
557 
558 /*
559  * Find the hash bucket for the given connection.
560  */
561 static struct syncache_head *
562 syncache_hashbucket(struct in_conninfo *inc)
563 {
564 	uint32_t hash;
565 
566 	/*
567 	 * The hash is built on foreign port + local port + foreign address.
568 	 * We rely on the fact that struct in_conninfo starts with 16 bits
569 	 * of foreign port, then 16 bits of local port then followed by 128
570 	 * bits of foreign address.  In case of IPv4 address, the first 3
571 	 * 32-bit words of the address always are zeroes.
572 	 */
573 	hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
574 	    V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
575 
576 	return (&V_tcp_syncache.hashbase[hash]);
577 }
578 
579 /*
580  * Find an entry in the syncache.
581  * Returns always with locked syncache_head plus a matching entry or NULL.
582  */
583 static struct syncache *
584 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
585 {
586 	struct syncache *sc;
587 	struct syncache_head *sch;
588 
589 	*schp = sch = syncache_hashbucket(inc);
590 	SCH_LOCK(sch);
591 
592 	/* Circle through bucket row to find matching entry. */
593 	TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
594 		if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
595 		    sizeof(struct in_endpoints)) == 0)
596 			break;
597 
598 	return (sc);	/* Always returns with locked sch. */
599 }
600 
601 /*
602  * This function is called when we get a RST for a
603  * non-existent connection, so that we can see if the
604  * connection is in the syn cache.  If it is, zap it.
605  * If required send a challenge ACK.
606  */
607 void
608 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m)
609 {
610 	struct syncache *sc;
611 	struct syncache_head *sch;
612 	char *s = NULL;
613 
614 	if (syncache_cookiesonly())
615 		return;
616 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
617 	SCH_LOCK_ASSERT(sch);
618 
619 	/*
620 	 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
621 	 * See RFC 793 page 65, section SEGMENT ARRIVES.
622 	 */
623 	if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
624 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
625 			log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
626 			    "FIN flag set, segment ignored\n", s, __func__);
627 		TCPSTAT_INC(tcps_badrst);
628 		goto done;
629 	}
630 
631 	/*
632 	 * No corresponding connection was found in syncache.
633 	 * If syncookies are enabled and possibly exclusively
634 	 * used, or we are under memory pressure, a valid RST
635 	 * may not find a syncache entry.  In that case we're
636 	 * done and no SYN|ACK retransmissions will happen.
637 	 * Otherwise the RST was misdirected or spoofed.
638 	 */
639 	if (sc == NULL) {
640 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
641 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
642 			    "syncache entry (possibly syncookie only), "
643 			    "segment ignored\n", s, __func__);
644 		TCPSTAT_INC(tcps_badrst);
645 		goto done;
646 	}
647 
648 	/*
649 	 * If the RST bit is set, check the sequence number to see
650 	 * if this is a valid reset segment.
651 	 *
652 	 * RFC 793 page 37:
653 	 *   In all states except SYN-SENT, all reset (RST) segments
654 	 *   are validated by checking their SEQ-fields.  A reset is
655 	 *   valid if its sequence number is in the window.
656 	 *
657 	 * RFC 793 page 69:
658 	 *   There are four cases for the acceptability test for an incoming
659 	 *   segment:
660 	 *
661 	 * Segment Receive  Test
662 	 * Length  Window
663 	 * ------- -------  -------------------------------------------
664 	 *    0       0     SEG.SEQ = RCV.NXT
665 	 *    0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
666 	 *   >0       0     not acceptable
667 	 *   >0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
668 	 *               or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
669 	 *
670 	 * Note that when receiving a SYN segment in the LISTEN state,
671 	 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
672 	 * described in RFC 793, page 66.
673 	 */
674 	if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
675 	    SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
676 	    (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
677 		if (V_tcp_insecure_rst ||
678 		    th->th_seq == sc->sc_irs + 1) {
679 			syncache_drop(sc, sch);
680 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
681 				log(LOG_DEBUG,
682 				    "%s; %s: Our SYN|ACK was rejected, "
683 				    "connection attempt aborted by remote "
684 				    "endpoint\n",
685 				    s, __func__);
686 			TCPSTAT_INC(tcps_sc_reset);
687 		} else {
688 			TCPSTAT_INC(tcps_badrst);
689 			/* Send challenge ACK. */
690 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
691 				log(LOG_DEBUG, "%s; %s: RST with invalid "
692 				    " SEQ %u != NXT %u (+WND %u), "
693 				    "sending challenge ACK\n",
694 				    s, __func__,
695 				    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
696 			syncache_respond(sc, m, TH_ACK);
697 		}
698 	} else {
699 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
700 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
701 			    "NXT %u (+WND %u), segment ignored\n",
702 			    s, __func__,
703 			    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
704 		TCPSTAT_INC(tcps_badrst);
705 	}
706 
707 done:
708 	if (s != NULL)
709 		free(s, M_TCPLOG);
710 	SCH_UNLOCK(sch);
711 }
712 
713 void
714 syncache_badack(struct in_conninfo *inc)
715 {
716 	struct syncache *sc;
717 	struct syncache_head *sch;
718 
719 	if (syncache_cookiesonly())
720 		return;
721 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
722 	SCH_LOCK_ASSERT(sch);
723 	if (sc != NULL) {
724 		syncache_drop(sc, sch);
725 		TCPSTAT_INC(tcps_sc_badack);
726 	}
727 	SCH_UNLOCK(sch);
728 }
729 
730 void
731 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq)
732 {
733 	struct syncache *sc;
734 	struct syncache_head *sch;
735 
736 	if (syncache_cookiesonly())
737 		return;
738 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
739 	SCH_LOCK_ASSERT(sch);
740 	if (sc == NULL)
741 		goto done;
742 
743 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
744 	if (ntohl(th_seq) != sc->sc_iss)
745 		goto done;
746 
747 	/*
748 	 * If we've rertransmitted 3 times and this is our second error,
749 	 * we remove the entry.  Otherwise, we allow it to continue on.
750 	 * This prevents us from incorrectly nuking an entry during a
751 	 * spurious network outage.
752 	 *
753 	 * See tcp_notify().
754 	 */
755 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
756 		sc->sc_flags |= SCF_UNREACH;
757 		goto done;
758 	}
759 	syncache_drop(sc, sch);
760 	TCPSTAT_INC(tcps_sc_unreach);
761 done:
762 	SCH_UNLOCK(sch);
763 }
764 
765 /*
766  * Build a new TCP socket structure from a syncache entry.
767  *
768  * On success return the newly created socket with its underlying inp locked.
769  */
770 static struct socket *
771 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
772 {
773 	struct tcp_function_block *blk;
774 	struct inpcb *inp = NULL;
775 	struct socket *so;
776 	struct tcpcb *tp;
777 	int error;
778 	char *s;
779 
780 	NET_EPOCH_ASSERT();
781 
782 	/*
783 	 * Ok, create the full blown connection, and set things up
784 	 * as they would have been set up if we had created the
785 	 * connection when the SYN arrived.  If we can't create
786 	 * the connection, abort it.
787 	 */
788 	so = sonewconn(lso, 0);
789 	if (so == NULL) {
790 		/*
791 		 * Drop the connection; we will either send a RST or
792 		 * have the peer retransmit its SYN again after its
793 		 * RTO and try again.
794 		 */
795 		TCPSTAT_INC(tcps_listendrop);
796 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
797 			log(LOG_DEBUG, "%s; %s: Socket create failed "
798 			    "due to limits or memory shortage\n",
799 			    s, __func__);
800 			free(s, M_TCPLOG);
801 		}
802 		goto abort2;
803 	}
804 #ifdef MAC
805 	mac_socketpeer_set_from_mbuf(m, so);
806 #endif
807 
808 	inp = sotoinpcb(so);
809 	inp->inp_inc.inc_fibnum = so->so_fibnum;
810 	INP_WLOCK(inp);
811 	/*
812 	 * Exclusive pcbinfo lock is not required in syncache socket case even
813 	 * if two inpcb locks can be acquired simultaneously:
814 	 *  - the inpcb in LISTEN state,
815 	 *  - the newly created inp.
816 	 *
817 	 * In this case, an inp cannot be at same time in LISTEN state and
818 	 * just created by an accept() call.
819 	 */
820 	INP_HASH_WLOCK(&V_tcbinfo);
821 
822 	/* Insert new socket into PCB hash list. */
823 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
824 #ifdef INET6
825 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
826 		inp->inp_vflag &= ~INP_IPV4;
827 		inp->inp_vflag |= INP_IPV6;
828 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
829 	} else {
830 		inp->inp_vflag &= ~INP_IPV6;
831 		inp->inp_vflag |= INP_IPV4;
832 #endif
833 		inp->inp_ip_ttl = sc->sc_ip_ttl;
834 		inp->inp_ip_tos = sc->sc_ip_tos;
835 		inp->inp_laddr = sc->sc_inc.inc_laddr;
836 #ifdef INET6
837 	}
838 #endif
839 
840 	/*
841 	 * If there's an mbuf and it has a flowid, then let's initialise the
842 	 * inp with that particular flowid.
843 	 */
844 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
845 		inp->inp_flowid = m->m_pkthdr.flowid;
846 		inp->inp_flowtype = M_HASHTYPE_GET(m);
847 #ifdef NUMA
848 		inp->inp_numa_domain = m->m_pkthdr.numa_domain;
849 #endif
850 	}
851 
852 	inp->inp_lport = sc->sc_inc.inc_lport;
853 #ifdef INET6
854 	if (inp->inp_vflag & INP_IPV6PROTO) {
855 		struct inpcb *oinp = sotoinpcb(lso);
856 
857 		/*
858 		 * Inherit socket options from the listening socket.
859 		 * Note that in6p_inputopts are not (and should not be)
860 		 * copied, since it stores previously received options and is
861 		 * used to detect if each new option is different than the
862 		 * previous one and hence should be passed to a user.
863 		 * If we copied in6p_inputopts, a user would not be able to
864 		 * receive options just after calling the accept system call.
865 		 */
866 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
867 		if (oinp->in6p_outputopts)
868 			inp->in6p_outputopts =
869 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
870 		inp->in6p_hops = oinp->in6p_hops;
871 	}
872 
873 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
874 		struct in6_addr laddr6;
875 		struct sockaddr_in6 sin6;
876 
877 		sin6.sin6_family = AF_INET6;
878 		sin6.sin6_len = sizeof(sin6);
879 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
880 		sin6.sin6_port = sc->sc_inc.inc_fport;
881 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
882 		laddr6 = inp->in6p_laddr;
883 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
884 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
885 		if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
886 		    thread0.td_ucred, m, false)) != 0) {
887 			inp->in6p_laddr = laddr6;
888 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
889 				log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
890 				    "with error %i\n",
891 				    s, __func__, error);
892 				free(s, M_TCPLOG);
893 			}
894 			INP_HASH_WUNLOCK(&V_tcbinfo);
895 			goto abort;
896 		}
897 		/* Override flowlabel from in6_pcbconnect. */
898 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
899 		inp->inp_flow |= sc->sc_flowlabel;
900 	}
901 #endif /* INET6 */
902 #if defined(INET) && defined(INET6)
903 	else
904 #endif
905 #ifdef INET
906 	{
907 		struct in_addr laddr;
908 		struct sockaddr_in sin;
909 
910 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
911 
912 		if (inp->inp_options == NULL) {
913 			inp->inp_options = sc->sc_ipopts;
914 			sc->sc_ipopts = NULL;
915 		}
916 
917 		sin.sin_family = AF_INET;
918 		sin.sin_len = sizeof(sin);
919 		sin.sin_addr = sc->sc_inc.inc_faddr;
920 		sin.sin_port = sc->sc_inc.inc_fport;
921 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
922 		laddr = inp->inp_laddr;
923 		if (inp->inp_laddr.s_addr == INADDR_ANY)
924 			inp->inp_laddr = sc->sc_inc.inc_laddr;
925 		if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
926 		    thread0.td_ucred, m, false)) != 0) {
927 			inp->inp_laddr = laddr;
928 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
929 				log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
930 				    "with error %i\n",
931 				    s, __func__, error);
932 				free(s, M_TCPLOG);
933 			}
934 			INP_HASH_WUNLOCK(&V_tcbinfo);
935 			goto abort;
936 		}
937 	}
938 #endif /* INET */
939 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
940 	/* Copy old policy into new socket's. */
941 	if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
942 		printf("syncache_socket: could not copy policy\n");
943 #endif
944 	INP_HASH_WUNLOCK(&V_tcbinfo);
945 	tp = intotcpcb(inp);
946 	tcp_state_change(tp, TCPS_SYN_RECEIVED);
947 	tp->iss = sc->sc_iss;
948 	tp->irs = sc->sc_irs;
949 	tcp_rcvseqinit(tp);
950 	tcp_sendseqinit(tp);
951 	blk = sototcpcb(lso)->t_fb;
952 	if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
953 		/*
954 		 * Our parents t_fb was not the default,
955 		 * we need to release our ref on tp->t_fb and
956 		 * pickup one on the new entry.
957 		 */
958 		struct tcp_function_block *rblk;
959 
960 		rblk = find_and_ref_tcp_fb(blk);
961 		KASSERT(rblk != NULL,
962 		    ("cannot find blk %p out of syncache?", blk));
963 		if (tp->t_fb->tfb_tcp_fb_fini)
964 			(*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
965 		refcount_release(&tp->t_fb->tfb_refcnt);
966 		tp->t_fb = rblk;
967 		/*
968 		 * XXXrrs this is quite dangerous, it is possible
969 		 * for the new function to fail to init. We also
970 		 * are not asking if the handoff_is_ok though at
971 		 * the very start thats probalbly ok.
972 		 */
973 		if (tp->t_fb->tfb_tcp_fb_init) {
974 			(*tp->t_fb->tfb_tcp_fb_init)(tp);
975 		}
976 	}
977 	tp->snd_wl1 = sc->sc_irs;
978 	tp->snd_max = tp->iss + 1;
979 	tp->snd_nxt = tp->iss + 1;
980 	tp->rcv_up = sc->sc_irs + 1;
981 	tp->rcv_wnd = sc->sc_wnd;
982 	tp->rcv_adv += tp->rcv_wnd;
983 	tp->last_ack_sent = tp->rcv_nxt;
984 
985 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
986 	if (sc->sc_flags & SCF_NOOPT)
987 		tp->t_flags |= TF_NOOPT;
988 	else {
989 		if (sc->sc_flags & SCF_WINSCALE) {
990 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
991 			tp->snd_scale = sc->sc_requested_s_scale;
992 			tp->request_r_scale = sc->sc_requested_r_scale;
993 		}
994 		if (sc->sc_flags & SCF_TIMESTAMP) {
995 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
996 			tp->ts_recent = sc->sc_tsreflect;
997 			tp->ts_recent_age = tcp_ts_getticks();
998 			tp->ts_offset = sc->sc_tsoff;
999 		}
1000 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1001 		if (sc->sc_flags & SCF_SIGNATURE)
1002 			tp->t_flags |= TF_SIGNATURE;
1003 #endif
1004 		if (sc->sc_flags & SCF_SACK)
1005 			tp->t_flags |= TF_SACK_PERMIT;
1006 	}
1007 
1008 	if (sc->sc_flags & SCF_ECN)
1009 		tp->t_flags2 |= TF2_ECN_PERMIT;
1010 
1011 	/*
1012 	 * Set up MSS and get cached values from tcp_hostcache.
1013 	 * This might overwrite some of the defaults we just set.
1014 	 */
1015 	tcp_mss(tp, sc->sc_peer_mss);
1016 
1017 	/*
1018 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
1019 	 * limited to one segment in cc_conn_init().
1020 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
1021 	 */
1022 	if (sc->sc_rxmits > 1)
1023 		tp->snd_cwnd = 1;
1024 
1025 #ifdef TCP_OFFLOAD
1026 	/*
1027 	 * Allow a TOE driver to install its hooks.  Note that we hold the
1028 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
1029 	 * new connection before the TOE driver has done its thing.
1030 	 */
1031 	if (ADDED_BY_TOE(sc)) {
1032 		struct toedev *tod = sc->sc_tod;
1033 
1034 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
1035 	}
1036 #endif
1037 	/*
1038 	 * Copy and activate timers.
1039 	 */
1040 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
1041 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
1042 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
1043 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1044 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1045 
1046 	TCPSTAT_INC(tcps_accepts);
1047 	return (so);
1048 
1049 abort:
1050 	INP_WUNLOCK(inp);
1051 abort2:
1052 	if (so != NULL)
1053 		soabort(so);
1054 	return (NULL);
1055 }
1056 
1057 /*
1058  * This function gets called when we receive an ACK for a
1059  * socket in the LISTEN state.  We look up the connection
1060  * in the syncache, and if its there, we pull it out of
1061  * the cache and turn it into a full-blown connection in
1062  * the SYN-RECEIVED state.
1063  *
1064  * On syncache_socket() success the newly created socket
1065  * has its underlying inp locked.
1066  */
1067 int
1068 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1069     struct socket **lsop, struct mbuf *m)
1070 {
1071 	struct syncache *sc;
1072 	struct syncache_head *sch;
1073 	struct syncache scs;
1074 	char *s;
1075 	bool locked;
1076 
1077 	NET_EPOCH_ASSERT();
1078 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1079 	    ("%s: can handle only ACK", __func__));
1080 
1081 	if (syncache_cookiesonly()) {
1082 		sc = NULL;
1083 		sch = syncache_hashbucket(inc);
1084 		locked = false;
1085 	} else {
1086 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1087 		locked = true;
1088 		SCH_LOCK_ASSERT(sch);
1089 	}
1090 
1091 #ifdef INVARIANTS
1092 	/*
1093 	 * Test code for syncookies comparing the syncache stored
1094 	 * values with the reconstructed values from the cookie.
1095 	 */
1096 	if (sc != NULL)
1097 		syncookie_cmp(inc, sch, sc, th, to, *lsop);
1098 #endif
1099 
1100 	if (sc == NULL) {
1101 		/*
1102 		 * There is no syncache entry, so see if this ACK is
1103 		 * a returning syncookie.  To do this, first:
1104 		 *  A. Check if syncookies are used in case of syncache
1105 		 *     overflows
1106 		 *  B. See if this socket has had a syncache entry dropped in
1107 		 *     the recent past. We don't want to accept a bogus
1108 		 *     syncookie if we've never received a SYN or accept it
1109 		 *     twice.
1110 		 *  C. check that the syncookie is valid.  If it is, then
1111 		 *     cobble up a fake syncache entry, and return.
1112 		 */
1113 		if (locked && !V_tcp_syncookies) {
1114 			SCH_UNLOCK(sch);
1115 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1116 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1117 				    "segment rejected (syncookies disabled)\n",
1118 				    s, __func__);
1119 			goto failed;
1120 		}
1121 		if (locked && !V_tcp_syncookiesonly &&
1122 		    sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1123 			SCH_UNLOCK(sch);
1124 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1125 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1126 				    "segment rejected (no syncache entry)\n",
1127 				    s, __func__);
1128 			goto failed;
1129 		}
1130 		bzero(&scs, sizeof(scs));
1131 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
1132 		if (locked)
1133 			SCH_UNLOCK(sch);
1134 		if (sc == NULL) {
1135 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1136 				log(LOG_DEBUG, "%s; %s: Segment failed "
1137 				    "SYNCOOKIE authentication, segment rejected "
1138 				    "(probably spoofed)\n", s, __func__);
1139 			goto failed;
1140 		}
1141 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1142 		/* If received ACK has MD5 signature, check it. */
1143 		if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1144 		    (!TCPMD5_ENABLED() ||
1145 		    TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1146 			/* Drop the ACK. */
1147 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1148 				log(LOG_DEBUG, "%s; %s: Segment rejected, "
1149 				    "MD5 signature doesn't match.\n",
1150 				    s, __func__);
1151 				free(s, M_TCPLOG);
1152 			}
1153 			TCPSTAT_INC(tcps_sig_err_sigopt);
1154 			return (-1); /* Do not send RST */
1155 		}
1156 #endif /* TCP_SIGNATURE */
1157 	} else {
1158 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1159 		/*
1160 		 * If listening socket requested TCP digests, check that
1161 		 * received ACK has signature and it is correct.
1162 		 * If not, drop the ACK and leave sc entry in th cache,
1163 		 * because SYN was received with correct signature.
1164 		 */
1165 		if (sc->sc_flags & SCF_SIGNATURE) {
1166 			if ((to->to_flags & TOF_SIGNATURE) == 0) {
1167 				/* No signature */
1168 				TCPSTAT_INC(tcps_sig_err_nosigopt);
1169 				SCH_UNLOCK(sch);
1170 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1171 					log(LOG_DEBUG, "%s; %s: Segment "
1172 					    "rejected, MD5 signature wasn't "
1173 					    "provided.\n", s, __func__);
1174 					free(s, M_TCPLOG);
1175 				}
1176 				return (-1); /* Do not send RST */
1177 			}
1178 			if (!TCPMD5_ENABLED() ||
1179 			    TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1180 				/* Doesn't match or no SA */
1181 				SCH_UNLOCK(sch);
1182 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1183 					log(LOG_DEBUG, "%s; %s: Segment "
1184 					    "rejected, MD5 signature doesn't "
1185 					    "match.\n", s, __func__);
1186 					free(s, M_TCPLOG);
1187 				}
1188 				return (-1); /* Do not send RST */
1189 			}
1190 		}
1191 #endif /* TCP_SIGNATURE */
1192 
1193 		/*
1194 		 * RFC 7323 PAWS: If we have a timestamp on this segment and
1195 		 * it's less than ts_recent, drop it.
1196 		 * XXXMT: RFC 7323 also requires to send an ACK.
1197 		 *        In tcp_input.c this is only done for TCP segments
1198 		 *        with user data, so be consistent here and just drop
1199 		 *        the segment.
1200 		 */
1201 		if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1202 		    TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1203 			SCH_UNLOCK(sch);
1204 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1205 				log(LOG_DEBUG,
1206 				    "%s; %s: SEG.TSval %u < TS.Recent %u, "
1207 				    "segment dropped\n", s, __func__,
1208 				    to->to_tsval, sc->sc_tsreflect);
1209 				free(s, M_TCPLOG);
1210 			}
1211 			return (-1);  /* Do not send RST */
1212 		}
1213 
1214 		/*
1215 		 * If timestamps were not negotiated during SYN/ACK and a
1216 		 * segment without a timestamp is received, ignore the
1217 		 * timestamp and process the packet normally.
1218 		 * See section 3.2 of RFC 7323.
1219 		 */
1220 		if (!(sc->sc_flags & SCF_TIMESTAMP) &&
1221 		    (to->to_flags & TOF_TS)) {
1222 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1223 				log(LOG_DEBUG, "%s; %s: Timestamp not "
1224 				    "expected, segment processed normally\n",
1225 				    s, __func__);
1226 				free(s, M_TCPLOG);
1227 				s = NULL;
1228 			}
1229 		}
1230 
1231 		/*
1232 		 * If timestamps were negotiated during SYN/ACK and a
1233 		 * segment without a timestamp is received, silently drop
1234 		 * the segment.
1235 		 * See section 3.2 of RFC 7323.
1236 		 */
1237 		if ((sc->sc_flags & SCF_TIMESTAMP) &&
1238 		    !(to->to_flags & TOF_TS)) {
1239 			SCH_UNLOCK(sch);
1240 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1241 				log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1242 				    "segment silently dropped\n", s, __func__);
1243 				free(s, M_TCPLOG);
1244 			}
1245 			return (-1);  /* Do not send RST */
1246 		}
1247 
1248 		/*
1249 		 * Pull out the entry to unlock the bucket row.
1250 		 *
1251 		 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1252 		 * tcp_state_change().  The tcpcb is not existent at this
1253 		 * moment.  A new one will be allocated via syncache_socket->
1254 		 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1255 		 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1256 		 */
1257 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1258 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1259 		sch->sch_length--;
1260 #ifdef TCP_OFFLOAD
1261 		if (ADDED_BY_TOE(sc)) {
1262 			struct toedev *tod = sc->sc_tod;
1263 
1264 			tod->tod_syncache_removed(tod, sc->sc_todctx);
1265 		}
1266 #endif
1267 		SCH_UNLOCK(sch);
1268 	}
1269 
1270 	/*
1271 	 * Segment validation:
1272 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1273 	 */
1274 	if (th->th_ack != sc->sc_iss + 1) {
1275 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1276 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1277 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1278 		goto failed;
1279 	}
1280 
1281 	/*
1282 	 * The SEQ must fall in the window starting at the received
1283 	 * initial receive sequence number + 1 (the SYN).
1284 	 */
1285 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1286 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1287 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1288 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1289 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1290 		goto failed;
1291 	}
1292 
1293 	*lsop = syncache_socket(sc, *lsop, m);
1294 
1295 	if (*lsop == NULL)
1296 		TCPSTAT_INC(tcps_sc_aborted);
1297 	else
1298 		TCPSTAT_INC(tcps_sc_completed);
1299 
1300 /* how do we find the inp for the new socket? */
1301 	if (sc != &scs)
1302 		syncache_free(sc);
1303 	return (1);
1304 failed:
1305 	if (sc != NULL && sc != &scs)
1306 		syncache_free(sc);
1307 	if (s != NULL)
1308 		free(s, M_TCPLOG);
1309 	*lsop = NULL;
1310 	return (0);
1311 }
1312 
1313 static void
1314 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
1315     uint64_t response_cookie)
1316 {
1317 	struct inpcb *inp;
1318 	struct tcpcb *tp;
1319 	unsigned int *pending_counter;
1320 
1321 	NET_EPOCH_ASSERT();
1322 
1323 	pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
1324 	*lsop = syncache_socket(sc, *lsop, m);
1325 	if (*lsop == NULL) {
1326 		TCPSTAT_INC(tcps_sc_aborted);
1327 		atomic_subtract_int(pending_counter, 1);
1328 	} else {
1329 		soisconnected(*lsop);
1330 		inp = sotoinpcb(*lsop);
1331 		tp = intotcpcb(inp);
1332 		tp->t_flags |= TF_FASTOPEN;
1333 		tp->t_tfo_cookie.server = response_cookie;
1334 		tp->snd_max = tp->iss;
1335 		tp->snd_nxt = tp->iss;
1336 		tp->t_tfo_pending = pending_counter;
1337 		TCPSTAT_INC(tcps_sc_completed);
1338 	}
1339 }
1340 
1341 /*
1342  * Given a LISTEN socket and an inbound SYN request, add
1343  * this to the syn cache, and send back a segment:
1344  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1345  * to the source.
1346  *
1347  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1348  * Doing so would require that we hold onto the data and deliver it
1349  * to the application.  However, if we are the target of a SYN-flood
1350  * DoS attack, an attacker could send data which would eventually
1351  * consume all available buffer space if it were ACKed.  By not ACKing
1352  * the data, we avoid this DoS scenario.
1353  *
1354  * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1355  * cookie is processed and a new socket is created.  In this case, any data
1356  * accompanying the SYN will be queued to the socket by tcp_input() and will
1357  * be ACKed either when the application sends response data or the delayed
1358  * ACK timer expires, whichever comes first.
1359  */
1360 int
1361 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1362     struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1363     void *todctx, uint8_t iptos)
1364 {
1365 	struct tcpcb *tp;
1366 	struct socket *so;
1367 	struct syncache *sc = NULL;
1368 	struct syncache_head *sch;
1369 	struct mbuf *ipopts = NULL;
1370 	u_int ltflags;
1371 	int win, ip_ttl, ip_tos;
1372 	char *s;
1373 	int rv = 0;
1374 #ifdef INET6
1375 	int autoflowlabel = 0;
1376 #endif
1377 #ifdef MAC
1378 	struct label *maclabel;
1379 #endif
1380 	struct syncache scs;
1381 	struct ucred *cred;
1382 	uint64_t tfo_response_cookie;
1383 	unsigned int *tfo_pending = NULL;
1384 	int tfo_cookie_valid = 0;
1385 	int tfo_response_cookie_valid = 0;
1386 	bool locked;
1387 
1388 	INP_WLOCK_ASSERT(inp);			/* listen socket */
1389 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1390 	    ("%s: unexpected tcp flags", __func__));
1391 
1392 	/*
1393 	 * Combine all so/tp operations very early to drop the INP lock as
1394 	 * soon as possible.
1395 	 */
1396 	so = *lsop;
1397 	KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1398 	tp = sototcpcb(so);
1399 	cred = crhold(so->so_cred);
1400 
1401 #ifdef INET6
1402 	if (inc->inc_flags & INC_ISIPV6) {
1403 		if (inp->inp_flags & IN6P_AUTOFLOWLABEL) {
1404 			autoflowlabel = 1;
1405 		}
1406 		ip_ttl = in6_selecthlim(inp, NULL);
1407 		if ((inp->in6p_outputopts == NULL) ||
1408 		    (inp->in6p_outputopts->ip6po_tclass == -1)) {
1409 			ip_tos = 0;
1410 		} else {
1411 			ip_tos = inp->in6p_outputopts->ip6po_tclass;
1412 		}
1413 	}
1414 #endif
1415 #if defined(INET6) && defined(INET)
1416 	else
1417 #endif
1418 #ifdef INET
1419 	{
1420 		ip_ttl = inp->inp_ip_ttl;
1421 		ip_tos = inp->inp_ip_tos;
1422 	}
1423 #endif
1424 	win = so->sol_sbrcv_hiwat;
1425 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1426 
1427 	if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
1428 	    (tp->t_tfo_pending != NULL) &&
1429 	    (to->to_flags & TOF_FASTOPEN)) {
1430 		/*
1431 		 * Limit the number of pending TFO connections to
1432 		 * approximately half of the queue limit.  This prevents TFO
1433 		 * SYN floods from starving the service by filling the
1434 		 * listen queue with bogus TFO connections.
1435 		 */
1436 		if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1437 		    (so->sol_qlimit / 2)) {
1438 			int result;
1439 
1440 			result = tcp_fastopen_check_cookie(inc,
1441 			    to->to_tfo_cookie, to->to_tfo_len,
1442 			    &tfo_response_cookie);
1443 			tfo_cookie_valid = (result > 0);
1444 			tfo_response_cookie_valid = (result >= 0);
1445 		}
1446 
1447 		/*
1448 		 * Remember the TFO pending counter as it will have to be
1449 		 * decremented below if we don't make it to syncache_tfo_expand().
1450 		 */
1451 		tfo_pending = tp->t_tfo_pending;
1452 	}
1453 
1454 	/* By the time we drop the lock these should no longer be used. */
1455 	so = NULL;
1456 	tp = NULL;
1457 
1458 #ifdef MAC
1459 	if (mac_syncache_init(&maclabel) != 0) {
1460 		INP_WUNLOCK(inp);
1461 		goto done;
1462 	} else
1463 		mac_syncache_create(maclabel, inp);
1464 #endif
1465 	if (!tfo_cookie_valid)
1466 		INP_WUNLOCK(inp);
1467 
1468 	/*
1469 	 * Remember the IP options, if any.
1470 	 */
1471 #ifdef INET6
1472 	if (!(inc->inc_flags & INC_ISIPV6))
1473 #endif
1474 #ifdef INET
1475 		ipopts = (m) ? ip_srcroute(m) : NULL;
1476 #else
1477 		ipopts = NULL;
1478 #endif
1479 
1480 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1481 	/*
1482 	 * If listening socket requested TCP digests, check that received
1483 	 * SYN has signature and it is correct. If signature doesn't match
1484 	 * or TCP_SIGNATURE support isn't enabled, drop the packet.
1485 	 */
1486 	if (ltflags & TF_SIGNATURE) {
1487 		if ((to->to_flags & TOF_SIGNATURE) == 0) {
1488 			TCPSTAT_INC(tcps_sig_err_nosigopt);
1489 			goto done;
1490 		}
1491 		if (!TCPMD5_ENABLED() ||
1492 		    TCPMD5_INPUT(m, th, to->to_signature) != 0)
1493 			goto done;
1494 	}
1495 #endif	/* TCP_SIGNATURE */
1496 	/*
1497 	 * See if we already have an entry for this connection.
1498 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1499 	 *
1500 	 * XXX: should the syncache be re-initialized with the contents
1501 	 * of the new SYN here (which may have different options?)
1502 	 *
1503 	 * XXX: We do not check the sequence number to see if this is a
1504 	 * real retransmit or a new connection attempt.  The question is
1505 	 * how to handle such a case; either ignore it as spoofed, or
1506 	 * drop the current entry and create a new one?
1507 	 */
1508 	if (syncache_cookiesonly()) {
1509 		sc = NULL;
1510 		sch = syncache_hashbucket(inc);
1511 		locked = false;
1512 	} else {
1513 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1514 		locked = true;
1515 		SCH_LOCK_ASSERT(sch);
1516 	}
1517 	if (sc != NULL) {
1518 		if (tfo_cookie_valid)
1519 			INP_WUNLOCK(inp);
1520 		TCPSTAT_INC(tcps_sc_dupsyn);
1521 		if (ipopts) {
1522 			/*
1523 			 * If we were remembering a previous source route,
1524 			 * forget it and use the new one we've been given.
1525 			 */
1526 			if (sc->sc_ipopts)
1527 				(void) m_free(sc->sc_ipopts);
1528 			sc->sc_ipopts = ipopts;
1529 		}
1530 		/*
1531 		 * Update timestamp if present.
1532 		 */
1533 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1534 			sc->sc_tsreflect = to->to_tsval;
1535 		else
1536 			sc->sc_flags &= ~SCF_TIMESTAMP;
1537 		/*
1538 		 * Disable ECN if needed.
1539 		 */
1540 		if ((sc->sc_flags & SCF_ECN) &&
1541 		    ((th->th_flags & (TH_ECE|TH_CWR)) != (TH_ECE|TH_CWR))) {
1542 			sc->sc_flags &= ~SCF_ECN;
1543 		}
1544 #ifdef MAC
1545 		/*
1546 		 * Since we have already unconditionally allocated label
1547 		 * storage, free it up.  The syncache entry will already
1548 		 * have an initialized label we can use.
1549 		 */
1550 		mac_syncache_destroy(&maclabel);
1551 #endif
1552 		TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1553 		/* Retransmit SYN|ACK and reset retransmit count. */
1554 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1555 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1556 			    "resetting timer and retransmitting SYN|ACK\n",
1557 			    s, __func__);
1558 			free(s, M_TCPLOG);
1559 		}
1560 		if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1561 			sc->sc_rxmits = 0;
1562 			syncache_timeout(sc, sch, 1);
1563 			TCPSTAT_INC(tcps_sndacks);
1564 			TCPSTAT_INC(tcps_sndtotal);
1565 		}
1566 		SCH_UNLOCK(sch);
1567 		goto donenoprobe;
1568 	}
1569 
1570 	if (tfo_cookie_valid) {
1571 		bzero(&scs, sizeof(scs));
1572 		sc = &scs;
1573 		goto skip_alloc;
1574 	}
1575 
1576 	/*
1577 	 * Skip allocating a syncache entry if we are just going to discard
1578 	 * it later.
1579 	 */
1580 	if (!locked) {
1581 		bzero(&scs, sizeof(scs));
1582 		sc = &scs;
1583 	} else
1584 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1585 	if (sc == NULL) {
1586 		/*
1587 		 * The zone allocator couldn't provide more entries.
1588 		 * Treat this as if the cache was full; drop the oldest
1589 		 * entry and insert the new one.
1590 		 */
1591 		TCPSTAT_INC(tcps_sc_zonefail);
1592 		if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1593 			sch->sch_last_overflow = time_uptime;
1594 			syncache_drop(sc, sch);
1595 			syncache_pause(inc);
1596 		}
1597 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1598 		if (sc == NULL) {
1599 			if (V_tcp_syncookies) {
1600 				bzero(&scs, sizeof(scs));
1601 				sc = &scs;
1602 			} else {
1603 				KASSERT(locked,
1604 				    ("%s: bucket unexpectedly unlocked",
1605 				    __func__));
1606 				SCH_UNLOCK(sch);
1607 				if (ipopts)
1608 					(void) m_free(ipopts);
1609 				goto done;
1610 			}
1611 		}
1612 	}
1613 
1614 skip_alloc:
1615 	if (!tfo_cookie_valid && tfo_response_cookie_valid)
1616 		sc->sc_tfo_cookie = &tfo_response_cookie;
1617 
1618 	/*
1619 	 * Fill in the syncache values.
1620 	 */
1621 #ifdef MAC
1622 	sc->sc_label = maclabel;
1623 #endif
1624 	sc->sc_cred = cred;
1625 	cred = NULL;
1626 	sc->sc_ipopts = ipopts;
1627 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1628 	sc->sc_ip_tos = ip_tos;
1629 	sc->sc_ip_ttl = ip_ttl;
1630 #ifdef TCP_OFFLOAD
1631 	sc->sc_tod = tod;
1632 	sc->sc_todctx = todctx;
1633 #endif
1634 	sc->sc_irs = th->th_seq;
1635 	sc->sc_flags = 0;
1636 	sc->sc_flowlabel = 0;
1637 
1638 	/*
1639 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1640 	 * win was derived from socket earlier in the function.
1641 	 */
1642 	win = imax(win, 0);
1643 	win = imin(win, TCP_MAXWIN);
1644 	sc->sc_wnd = win;
1645 
1646 	if (V_tcp_do_rfc1323) {
1647 		/*
1648 		 * A timestamp received in a SYN makes
1649 		 * it ok to send timestamp requests and replies.
1650 		 */
1651 		if (to->to_flags & TOF_TS) {
1652 			sc->sc_tsreflect = to->to_tsval;
1653 			sc->sc_flags |= SCF_TIMESTAMP;
1654 			sc->sc_tsoff = tcp_new_ts_offset(inc);
1655 		}
1656 		if (to->to_flags & TOF_SCALE) {
1657 			int wscale = 0;
1658 
1659 			/*
1660 			 * Pick the smallest possible scaling factor that
1661 			 * will still allow us to scale up to sb_max, aka
1662 			 * kern.ipc.maxsockbuf.
1663 			 *
1664 			 * We do this because there are broken firewalls that
1665 			 * will corrupt the window scale option, leading to
1666 			 * the other endpoint believing that our advertised
1667 			 * window is unscaled.  At scale factors larger than
1668 			 * 5 the unscaled window will drop below 1500 bytes,
1669 			 * leading to serious problems when traversing these
1670 			 * broken firewalls.
1671 			 *
1672 			 * With the default maxsockbuf of 256K, a scale factor
1673 			 * of 3 will be chosen by this algorithm.  Those who
1674 			 * choose a larger maxsockbuf should watch out
1675 			 * for the compatibility problems mentioned above.
1676 			 *
1677 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1678 			 * or <SYN,ACK>) segment itself is never scaled.
1679 			 */
1680 			while (wscale < TCP_MAX_WINSHIFT &&
1681 			    (TCP_MAXWIN << wscale) < sb_max)
1682 				wscale++;
1683 			sc->sc_requested_r_scale = wscale;
1684 			sc->sc_requested_s_scale = to->to_wscale;
1685 			sc->sc_flags |= SCF_WINSCALE;
1686 		}
1687 	}
1688 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1689 	/*
1690 	 * If listening socket requested TCP digests, flag this in the
1691 	 * syncache so that syncache_respond() will do the right thing
1692 	 * with the SYN+ACK.
1693 	 */
1694 	if (ltflags & TF_SIGNATURE)
1695 		sc->sc_flags |= SCF_SIGNATURE;
1696 #endif	/* TCP_SIGNATURE */
1697 	if (to->to_flags & TOF_SACKPERM)
1698 		sc->sc_flags |= SCF_SACK;
1699 	if (to->to_flags & TOF_MSS)
1700 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1701 	if (ltflags & TF_NOOPT)
1702 		sc->sc_flags |= SCF_NOOPT;
1703 	if (((th->th_flags & (TH_ECE|TH_CWR)) == (TH_ECE|TH_CWR)) &&
1704 	    V_tcp_do_ecn)
1705 		sc->sc_flags |= SCF_ECN;
1706 
1707 	if (V_tcp_syncookies)
1708 		sc->sc_iss = syncookie_generate(sch, sc);
1709 	else
1710 		sc->sc_iss = arc4random();
1711 #ifdef INET6
1712 	if (autoflowlabel) {
1713 		if (V_tcp_syncookies)
1714 			sc->sc_flowlabel = sc->sc_iss;
1715 		else
1716 			sc->sc_flowlabel = ip6_randomflowlabel();
1717 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1718 	}
1719 #endif
1720 	if (locked)
1721 		SCH_UNLOCK(sch);
1722 
1723 	if (tfo_cookie_valid) {
1724 		syncache_tfo_expand(sc, lsop, m, tfo_response_cookie);
1725 		/* INP_WUNLOCK(inp) will be performed by the caller */
1726 		rv = 1;
1727 		goto tfo_expanded;
1728 	}
1729 
1730 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1731 	/*
1732 	 * Do a standard 3-way handshake.
1733 	 */
1734 	if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1735 		if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1736 			syncache_free(sc);
1737 		else if (sc != &scs)
1738 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1739 		TCPSTAT_INC(tcps_sndacks);
1740 		TCPSTAT_INC(tcps_sndtotal);
1741 	} else {
1742 		if (sc != &scs)
1743 			syncache_free(sc);
1744 		TCPSTAT_INC(tcps_sc_dropped);
1745 	}
1746 	goto donenoprobe;
1747 
1748 done:
1749 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1750 donenoprobe:
1751 	if (m) {
1752 		*lsop = NULL;
1753 		m_freem(m);
1754 	}
1755 	/*
1756 	 * If tfo_pending is not NULL here, then a TFO SYN that did not
1757 	 * result in a new socket was processed and the associated pending
1758 	 * counter has not yet been decremented.  All such TFO processing paths
1759 	 * transit this point.
1760 	 */
1761 	if (tfo_pending != NULL)
1762 		tcp_fastopen_decrement_counter(tfo_pending);
1763 
1764 tfo_expanded:
1765 	if (cred != NULL)
1766 		crfree(cred);
1767 #ifdef MAC
1768 	if (sc == &scs)
1769 		mac_syncache_destroy(&maclabel);
1770 #endif
1771 	return (rv);
1772 }
1773 
1774 /*
1775  * Send SYN|ACK or ACK to the peer.  Either in response to a peer's segment,
1776  * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1777  */
1778 static int
1779 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1780 {
1781 	struct ip *ip = NULL;
1782 	struct mbuf *m;
1783 	struct tcphdr *th = NULL;
1784 	int optlen, error = 0;	/* Make compiler happy */
1785 	u_int16_t hlen, tlen, mssopt;
1786 	struct tcpopt to;
1787 #ifdef INET6
1788 	struct ip6_hdr *ip6 = NULL;
1789 #endif
1790 
1791 	NET_EPOCH_ASSERT();
1792 
1793 	hlen =
1794 #ifdef INET6
1795 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1796 #endif
1797 		sizeof(struct ip);
1798 	tlen = hlen + sizeof(struct tcphdr);
1799 
1800 	/* Determine MSS we advertize to other end of connection. */
1801 	mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss);
1802 
1803 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1804 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1805 	    ("syncache: mbuf too small"));
1806 
1807 	/* Create the IP+TCP header from scratch. */
1808 	m = m_gethdr(M_NOWAIT, MT_DATA);
1809 	if (m == NULL)
1810 		return (ENOBUFS);
1811 #ifdef MAC
1812 	mac_syncache_create_mbuf(sc->sc_label, m);
1813 #endif
1814 	m->m_data += max_linkhdr;
1815 	m->m_len = tlen;
1816 	m->m_pkthdr.len = tlen;
1817 	m->m_pkthdr.rcvif = NULL;
1818 
1819 #ifdef INET6
1820 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1821 		ip6 = mtod(m, struct ip6_hdr *);
1822 		ip6->ip6_vfc = IPV6_VERSION;
1823 		ip6->ip6_nxt = IPPROTO_TCP;
1824 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1825 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1826 		ip6->ip6_plen = htons(tlen - hlen);
1827 		/* ip6_hlim is set after checksum */
1828 		/* Zero out traffic class and flow label. */
1829 		ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1830 		ip6->ip6_flow |= sc->sc_flowlabel;
1831 		ip6->ip6_flow |= htonl(sc->sc_ip_tos << 20);
1832 
1833 		th = (struct tcphdr *)(ip6 + 1);
1834 	}
1835 #endif
1836 #if defined(INET6) && defined(INET)
1837 	else
1838 #endif
1839 #ifdef INET
1840 	{
1841 		ip = mtod(m, struct ip *);
1842 		ip->ip_v = IPVERSION;
1843 		ip->ip_hl = sizeof(struct ip) >> 2;
1844 		ip->ip_len = htons(tlen);
1845 		ip->ip_id = 0;
1846 		ip->ip_off = 0;
1847 		ip->ip_sum = 0;
1848 		ip->ip_p = IPPROTO_TCP;
1849 		ip->ip_src = sc->sc_inc.inc_laddr;
1850 		ip->ip_dst = sc->sc_inc.inc_faddr;
1851 		ip->ip_ttl = sc->sc_ip_ttl;
1852 		ip->ip_tos = sc->sc_ip_tos;
1853 
1854 		/*
1855 		 * See if we should do MTU discovery.  Route lookups are
1856 		 * expensive, so we will only unset the DF bit if:
1857 		 *
1858 		 *	1) path_mtu_discovery is disabled
1859 		 *	2) the SCF_UNREACH flag has been set
1860 		 */
1861 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1862 		       ip->ip_off |= htons(IP_DF);
1863 
1864 		th = (struct tcphdr *)(ip + 1);
1865 	}
1866 #endif /* INET */
1867 	th->th_sport = sc->sc_inc.inc_lport;
1868 	th->th_dport = sc->sc_inc.inc_fport;
1869 
1870 	if (flags & TH_SYN)
1871 		th->th_seq = htonl(sc->sc_iss);
1872 	else
1873 		th->th_seq = htonl(sc->sc_iss + 1);
1874 	th->th_ack = htonl(sc->sc_irs + 1);
1875 	th->th_off = sizeof(struct tcphdr) >> 2;
1876 	th->th_x2 = 0;
1877 	th->th_flags = flags;
1878 	th->th_win = htons(sc->sc_wnd);
1879 	th->th_urp = 0;
1880 
1881 	if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) {
1882 		th->th_flags |= TH_ECE;
1883 		TCPSTAT_INC(tcps_ecn_shs);
1884 	}
1885 
1886 	/* Tack on the TCP options. */
1887 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1888 		to.to_flags = 0;
1889 
1890 		if (flags & TH_SYN) {
1891 			to.to_mss = mssopt;
1892 			to.to_flags = TOF_MSS;
1893 			if (sc->sc_flags & SCF_WINSCALE) {
1894 				to.to_wscale = sc->sc_requested_r_scale;
1895 				to.to_flags |= TOF_SCALE;
1896 			}
1897 			if (sc->sc_flags & SCF_SACK)
1898 				to.to_flags |= TOF_SACKPERM;
1899 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1900 			if (sc->sc_flags & SCF_SIGNATURE)
1901 				to.to_flags |= TOF_SIGNATURE;
1902 #endif
1903 			if (sc->sc_tfo_cookie) {
1904 				to.to_flags |= TOF_FASTOPEN;
1905 				to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1906 				to.to_tfo_cookie = sc->sc_tfo_cookie;
1907 				/* don't send cookie again when retransmitting response */
1908 				sc->sc_tfo_cookie = NULL;
1909 			}
1910 		}
1911 		if (sc->sc_flags & SCF_TIMESTAMP) {
1912 			to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1913 			to.to_tsecr = sc->sc_tsreflect;
1914 			to.to_flags |= TOF_TS;
1915 		}
1916 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1917 
1918 		/* Adjust headers by option size. */
1919 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1920 		m->m_len += optlen;
1921 		m->m_pkthdr.len += optlen;
1922 #ifdef INET6
1923 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1924 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1925 		else
1926 #endif
1927 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1928 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1929 		if (sc->sc_flags & SCF_SIGNATURE) {
1930 			KASSERT(to.to_flags & TOF_SIGNATURE,
1931 			    ("tcp_addoptions() didn't set tcp_signature"));
1932 
1933 			/* NOTE: to.to_signature is inside of mbuf */
1934 			if (!TCPMD5_ENABLED() ||
1935 			    TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1936 				m_freem(m);
1937 				return (EACCES);
1938 			}
1939 		}
1940 #endif
1941 	} else
1942 		optlen = 0;
1943 
1944 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
1945 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1946 	/*
1947 	 * If we have peer's SYN and it has a flowid, then let's assign it to
1948 	 * our SYN|ACK.  ip6_output() and ip_output() will not assign flowid
1949 	 * to SYN|ACK due to lack of inp here.
1950 	 */
1951 	if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1952 		m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1953 		M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1954 	}
1955 #ifdef INET6
1956 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1957 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1958 		th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1959 		    IPPROTO_TCP, 0);
1960 		ip6->ip6_hlim = sc->sc_ip_ttl;
1961 #ifdef TCP_OFFLOAD
1962 		if (ADDED_BY_TOE(sc)) {
1963 			struct toedev *tod = sc->sc_tod;
1964 
1965 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1966 
1967 			return (error);
1968 		}
1969 #endif
1970 		TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
1971 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1972 	}
1973 #endif
1974 #if defined(INET6) && defined(INET)
1975 	else
1976 #endif
1977 #ifdef INET
1978 	{
1979 		m->m_pkthdr.csum_flags = CSUM_TCP;
1980 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1981 		    htons(tlen + optlen - hlen + IPPROTO_TCP));
1982 #ifdef TCP_OFFLOAD
1983 		if (ADDED_BY_TOE(sc)) {
1984 			struct toedev *tod = sc->sc_tod;
1985 
1986 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1987 
1988 			return (error);
1989 		}
1990 #endif
1991 		TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
1992 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1993 	}
1994 #endif
1995 	return (error);
1996 }
1997 
1998 /*
1999  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
2000  * that exceed the capacity of the syncache by avoiding the storage of any
2001  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
2002  * attacks where the attacker does not have access to our responses.
2003  *
2004  * Syncookies encode and include all necessary information about the
2005  * connection setup within the SYN|ACK that we send back.  That way we
2006  * can avoid keeping any local state until the ACK to our SYN|ACK returns
2007  * (if ever).  Normally the syncache and syncookies are running in parallel
2008  * with the latter taking over when the former is exhausted.  When matching
2009  * syncache entry is found the syncookie is ignored.
2010  *
2011  * The only reliable information persisting the 3WHS is our initial sequence
2012  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
2013  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
2014  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
2015  * returns and signifies a legitimate connection if it matches the ACK.
2016  *
2017  * The available space of 32 bits to store the hash and to encode the SYN
2018  * option information is very tight and we should have at least 24 bits for
2019  * the MAC to keep the number of guesses by blind spoofing reasonably high.
2020  *
2021  * SYN option information we have to encode to fully restore a connection:
2022  * MSS: is imporant to chose an optimal segment size to avoid IP level
2023  *   fragmentation along the path.  The common MSS values can be encoded
2024  *   in a 3-bit table.  Uncommon values are captured by the next lower value
2025  *   in the table leading to a slight increase in packetization overhead.
2026  * WSCALE: is necessary to allow large windows to be used for high delay-
2027  *   bandwidth product links.  Not scaling the window when it was initially
2028  *   negotiated is bad for performance as lack of scaling further decreases
2029  *   the apparent available send window.  We only need to encode the WSCALE
2030  *   we received from the remote end.  Our end can be recalculated at any
2031  *   time.  The common WSCALE values can be encoded in a 3-bit table.
2032  *   Uncommon values are captured by the next lower value in the table
2033  *   making us under-estimate the available window size halving our
2034  *   theoretically possible maximum throughput for that connection.
2035  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2036  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2037  *   that are included in all segments on a connection.  We enable them when
2038  *   the ACK has them.
2039  *
2040  * Security of syncookies and attack vectors:
2041  *
2042  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2043  * together with the gloabl secret to make it unique per connection attempt.
2044  * Thus any change of any of those parameters results in a different MAC output
2045  * in an unpredictable way unless a collision is encountered.  24 bits of the
2046  * MAC are embedded into the ISS.
2047  *
2048  * To prevent replay attacks two rotating global secrets are updated with a
2049  * new random value every 15 seconds.  The life-time of a syncookie is thus
2050  * 15-30 seconds.
2051  *
2052  * Vector 1: Attacking the secret.  This requires finding a weakness in the
2053  * MAC itself or the way it is used here.  The attacker can do a chosen plain
2054  * text attack by varying and testing the all parameters under his control.
2055  * The strength depends on the size and randomness of the secret, and the
2056  * cryptographic security of the MAC function.  Due to the constant updating
2057  * of the secret the attacker has at most 29.999 seconds to find the secret
2058  * and launch spoofed connections.  After that he has to start all over again.
2059  *
2060  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
2061  * size an average of 4,823 attempts are required for a 50% chance of success
2062  * to spoof a single syncookie (birthday collision paradox).  However the
2063  * attacker is blind and doesn't know if one of his attempts succeeded unless
2064  * he has a side channel to interfere success from.  A single connection setup
2065  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2066  * This many attempts are required for each one blind spoofed connection.  For
2067  * every additional spoofed connection he has to launch another N attempts.
2068  * Thus for a sustained rate 100 spoofed connections per second approximately
2069  * 1,800,000 packets per second would have to be sent.
2070  *
2071  * NB: The MAC function should be fast so that it doesn't become a CPU
2072  * exhaustion attack vector itself.
2073  *
2074  * References:
2075  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2076  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2077  *   http://cr.yp.to/syncookies.html    (overview)
2078  *   http://cr.yp.to/syncookies/archive (details)
2079  *
2080  *
2081  * Schematic construction of a syncookie enabled Initial Sequence Number:
2082  *  0        1         2         3
2083  *  12345678901234567890123456789012
2084  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2085  *
2086  *  x 24 MAC (truncated)
2087  *  W  3 Send Window Scale index
2088  *  M  3 MSS index
2089  *  S  1 SACK permitted
2090  *  P  1 Odd/even secret
2091  */
2092 
2093 /*
2094  * Distribution and probability of certain MSS values.  Those in between are
2095  * rounded down to the next lower one.
2096  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2097  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
2098  */
2099 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2100 
2101 /*
2102  * Distribution and probability of certain WSCALE values.  We have to map the
2103  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2104  * bits based on prevalence of certain values.  Where we don't have an exact
2105  * match for are rounded down to the next lower one letting us under-estimate
2106  * the true available window.  At the moment this would happen only for the
2107  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2108  * and window size).  The absence of the WSCALE option (no scaling in either
2109  * direction) is encoded with index zero.
2110  * [WSCALE values histograms, Allman, 2012]
2111  *                            X 10 10 35  5  6 14 10%   by host
2112  *                            X 11  4  5  5 18 49  3%   by connections
2113  */
2114 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2115 
2116 /*
2117  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
2118  * and good cryptographic properties.
2119  */
2120 static uint32_t
2121 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2122     uint8_t *secbits, uintptr_t secmod)
2123 {
2124 	SIPHASH_CTX ctx;
2125 	uint32_t siphash[2];
2126 
2127 	SipHash24_Init(&ctx);
2128 	SipHash_SetKey(&ctx, secbits);
2129 	switch (inc->inc_flags & INC_ISIPV6) {
2130 #ifdef INET
2131 	case 0:
2132 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2133 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2134 		break;
2135 #endif
2136 #ifdef INET6
2137 	case INC_ISIPV6:
2138 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2139 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2140 		break;
2141 #endif
2142 	}
2143 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2144 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2145 	SipHash_Update(&ctx, &irs, sizeof(irs));
2146 	SipHash_Update(&ctx, &flags, sizeof(flags));
2147 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
2148 	SipHash_Final((u_int8_t *)&siphash, &ctx);
2149 
2150 	return (siphash[0] ^ siphash[1]);
2151 }
2152 
2153 static tcp_seq
2154 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2155 {
2156 	u_int i, secbit, wscale;
2157 	uint32_t iss, hash;
2158 	uint8_t *secbits;
2159 	union syncookie cookie;
2160 
2161 	cookie.cookie = 0;
2162 
2163 	/* Map our computed MSS into the 3-bit index. */
2164 	for (i = nitems(tcp_sc_msstab) - 1;
2165 	     tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2166 	     i--)
2167 		;
2168 	cookie.flags.mss_idx = i;
2169 
2170 	/*
2171 	 * Map the send window scale into the 3-bit index but only if
2172 	 * the wscale option was received.
2173 	 */
2174 	if (sc->sc_flags & SCF_WINSCALE) {
2175 		wscale = sc->sc_requested_s_scale;
2176 		for (i = nitems(tcp_sc_wstab) - 1;
2177 		    tcp_sc_wstab[i] > wscale && i > 0;
2178 		     i--)
2179 			;
2180 		cookie.flags.wscale_idx = i;
2181 	}
2182 
2183 	/* Can we do SACK? */
2184 	if (sc->sc_flags & SCF_SACK)
2185 		cookie.flags.sack_ok = 1;
2186 
2187 	/* Which of the two secrets to use. */
2188 	secbit = V_tcp_syncache.secret.oddeven & 0x1;
2189 	cookie.flags.odd_even = secbit;
2190 
2191 	secbits = V_tcp_syncache.secret.key[secbit];
2192 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2193 	    (uintptr_t)sch);
2194 
2195 	/*
2196 	 * Put the flags into the hash and XOR them to get better ISS number
2197 	 * variance.  This doesn't enhance the cryptographic strength and is
2198 	 * done to prevent the 8 cookie bits from showing up directly on the
2199 	 * wire.
2200 	 */
2201 	iss = hash & ~0xff;
2202 	iss |= cookie.cookie ^ (hash >> 24);
2203 
2204 	TCPSTAT_INC(tcps_sc_sendcookie);
2205 	return (iss);
2206 }
2207 
2208 static struct syncache *
2209 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2210     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2211     struct socket *lso)
2212 {
2213 	uint32_t hash;
2214 	uint8_t *secbits;
2215 	tcp_seq ack, seq;
2216 	int wnd, wscale = 0;
2217 	union syncookie cookie;
2218 
2219 	/*
2220 	 * Pull information out of SYN-ACK/ACK and revert sequence number
2221 	 * advances.
2222 	 */
2223 	ack = th->th_ack - 1;
2224 	seq = th->th_seq - 1;
2225 
2226 	/*
2227 	 * Unpack the flags containing enough information to restore the
2228 	 * connection.
2229 	 */
2230 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2231 
2232 	/* Which of the two secrets to use. */
2233 	secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2234 
2235 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2236 
2237 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
2238 	if ((ack & ~0xff) != (hash & ~0xff))
2239 		return (NULL);
2240 
2241 	/* Fill in the syncache values. */
2242 	sc->sc_flags = 0;
2243 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2244 	sc->sc_ipopts = NULL;
2245 
2246 	sc->sc_irs = seq;
2247 	sc->sc_iss = ack;
2248 
2249 	switch (inc->inc_flags & INC_ISIPV6) {
2250 #ifdef INET
2251 	case 0:
2252 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2253 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2254 		break;
2255 #endif
2256 #ifdef INET6
2257 	case INC_ISIPV6:
2258 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2259 			sc->sc_flowlabel =
2260 			    htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2261 		break;
2262 #endif
2263 	}
2264 
2265 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2266 
2267 	/* We can simply recompute receive window scale we sent earlier. */
2268 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2269 		wscale++;
2270 
2271 	/* Only use wscale if it was enabled in the orignal SYN. */
2272 	if (cookie.flags.wscale_idx > 0) {
2273 		sc->sc_requested_r_scale = wscale;
2274 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2275 		sc->sc_flags |= SCF_WINSCALE;
2276 	}
2277 
2278 	wnd = lso->sol_sbrcv_hiwat;
2279 	wnd = imax(wnd, 0);
2280 	wnd = imin(wnd, TCP_MAXWIN);
2281 	sc->sc_wnd = wnd;
2282 
2283 	if (cookie.flags.sack_ok)
2284 		sc->sc_flags |= SCF_SACK;
2285 
2286 	if (to->to_flags & TOF_TS) {
2287 		sc->sc_flags |= SCF_TIMESTAMP;
2288 		sc->sc_tsreflect = to->to_tsval;
2289 		sc->sc_tsoff = tcp_new_ts_offset(inc);
2290 	}
2291 
2292 	if (to->to_flags & TOF_SIGNATURE)
2293 		sc->sc_flags |= SCF_SIGNATURE;
2294 
2295 	sc->sc_rxmits = 0;
2296 
2297 	TCPSTAT_INC(tcps_sc_recvcookie);
2298 	return (sc);
2299 }
2300 
2301 #ifdef INVARIANTS
2302 static int
2303 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2304     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2305     struct socket *lso)
2306 {
2307 	struct syncache scs, *scx;
2308 	char *s;
2309 
2310 	bzero(&scs, sizeof(scs));
2311 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
2312 
2313 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2314 		return (0);
2315 
2316 	if (scx != NULL) {
2317 		if (sc->sc_peer_mss != scx->sc_peer_mss)
2318 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2319 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2320 
2321 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2322 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2323 			    s, __func__, sc->sc_requested_r_scale,
2324 			    scx->sc_requested_r_scale);
2325 
2326 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2327 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2328 			    s, __func__, sc->sc_requested_s_scale,
2329 			    scx->sc_requested_s_scale);
2330 
2331 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2332 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2333 	}
2334 
2335 	if (s != NULL)
2336 		free(s, M_TCPLOG);
2337 	return (0);
2338 }
2339 #endif /* INVARIANTS */
2340 
2341 static void
2342 syncookie_reseed(void *arg)
2343 {
2344 	struct tcp_syncache *sc = arg;
2345 	uint8_t *secbits;
2346 	int secbit;
2347 
2348 	/*
2349 	 * Reseeding the secret doesn't have to be protected by a lock.
2350 	 * It only must be ensured that the new random values are visible
2351 	 * to all CPUs in a SMP environment.  The atomic with release
2352 	 * semantics ensures that.
2353 	 */
2354 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2355 	secbits = sc->secret.key[secbit];
2356 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2357 	atomic_add_rel_int(&sc->secret.oddeven, 1);
2358 
2359 	/* Reschedule ourself. */
2360 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2361 }
2362 
2363 /*
2364  * We have overflowed a bucket. Let's pause dealing with the syncache.
2365  * This function will increment the bucketoverflow statistics appropriately
2366  * (once per pause when pausing is enabled; otherwise, once per overflow).
2367  */
2368 static void
2369 syncache_pause(struct in_conninfo *inc)
2370 {
2371 	time_t delta;
2372 	const char *s;
2373 
2374 	/* XXX:
2375 	 * 2. Add sysctl read here so we don't get the benefit of this
2376 	 * change without the new sysctl.
2377 	 */
2378 
2379 	/*
2380 	 * Try an unlocked read. If we already know that another thread
2381 	 * has activated the feature, there is no need to proceed.
2382 	 */
2383 	if (V_tcp_syncache.paused)
2384 		return;
2385 
2386 	/* Are cookied enabled? If not, we can't pause. */
2387 	if (!V_tcp_syncookies) {
2388 		TCPSTAT_INC(tcps_sc_bucketoverflow);
2389 		return;
2390 	}
2391 
2392 	/*
2393 	 * We may be the first thread to find an overflow. Get the lock
2394 	 * and evaluate if we need to take action.
2395 	 */
2396 	mtx_lock(&V_tcp_syncache.pause_mtx);
2397 	if (V_tcp_syncache.paused) {
2398 		mtx_unlock(&V_tcp_syncache.pause_mtx);
2399 		return;
2400 	}
2401 
2402 	/* Activate protection. */
2403 	V_tcp_syncache.paused = true;
2404 	TCPSTAT_INC(tcps_sc_bucketoverflow);
2405 
2406 	/*
2407 	 * Determine the last backoff time. If we are seeing a re-newed
2408 	 * attack within that same time after last reactivating the syncache,
2409 	 * consider it an extension of the same attack.
2410 	 */
2411 	delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2412 	if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2413 		if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2414 			delta <<= 1;
2415 			V_tcp_syncache.pause_backoff++;
2416 		}
2417 	} else {
2418 		delta = TCP_SYNCACHE_PAUSE_TIME;
2419 		V_tcp_syncache.pause_backoff = 0;
2420 	}
2421 
2422 	/* Log a warning, including IP addresses, if able. */
2423 	if (inc != NULL)
2424 		s = tcp_log_addrs(inc, NULL, NULL, NULL);
2425 	else
2426 		s = (const char *)NULL;
2427 	log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2428 	    "the next %lld seconds%s%s%s\n", (long long)delta,
2429 	    (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2430 	    (s != NULL) ? ")" : "");
2431 	free(__DECONST(void *, s), M_TCPLOG);
2432 
2433 	/* Use the calculated delta to set a new pause time. */
2434 	V_tcp_syncache.pause_until = time_uptime + delta;
2435 	callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2436 	    &V_tcp_syncache);
2437 	mtx_unlock(&V_tcp_syncache.pause_mtx);
2438 }
2439 
2440 /* Evaluate whether we need to unpause. */
2441 static void
2442 syncache_unpause(void *arg)
2443 {
2444 	struct tcp_syncache *sc;
2445 	time_t delta;
2446 
2447 	sc = arg;
2448 	mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2449 	callout_deactivate(&sc->pause_co);
2450 
2451 	/*
2452 	 * Check to make sure we are not running early. If the pause
2453 	 * time has expired, then deactivate the protection.
2454 	 */
2455 	if ((delta = sc->pause_until - time_uptime) > 0)
2456 		callout_schedule(&sc->pause_co, delta * hz);
2457 	else
2458 		sc->paused = false;
2459 }
2460 
2461 /*
2462  * Exports the syncache entries to userland so that netstat can display
2463  * them alongside the other sockets.  This function is intended to be
2464  * called only from tcp_pcblist.
2465  *
2466  * Due to concurrency on an active system, the number of pcbs exported
2467  * may have no relation to max_pcbs.  max_pcbs merely indicates the
2468  * amount of space the caller allocated for this function to use.
2469  */
2470 int
2471 syncache_pcblist(struct sysctl_req *req)
2472 {
2473 	struct xtcpcb xt;
2474 	struct syncache *sc;
2475 	struct syncache_head *sch;
2476 	int error, i;
2477 
2478 	bzero(&xt, sizeof(xt));
2479 	xt.xt_len = sizeof(xt);
2480 	xt.t_state = TCPS_SYN_RECEIVED;
2481 	xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2482 	xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2483 	xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2484 	xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2485 
2486 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2487 		sch = &V_tcp_syncache.hashbase[i];
2488 		SCH_LOCK(sch);
2489 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2490 			if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2491 				continue;
2492 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2493 				xt.xt_inp.inp_vflag = INP_IPV6;
2494 			else
2495 				xt.xt_inp.inp_vflag = INP_IPV4;
2496 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2497 			    sizeof (struct in_conninfo));
2498 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2499 			if (error) {
2500 				SCH_UNLOCK(sch);
2501 				return (0);
2502 			}
2503 		}
2504 		SCH_UNLOCK(sch);
2505 	}
2506 
2507 	return (0);
2508 }
2509