xref: /freebsd/sys/netinet/tcp_syncache.c (revision dd4f32ae62426a10a84b4322756d82c06c202c4e)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2001 McAfee, Inc.
5  * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG
6  * All rights reserved.
7  *
8  * This software was developed for the FreeBSD Project by Jonathan Lemon
9  * and McAfee Research, the Security Research Division of McAfee, Inc. under
10  * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
11  * DARPA CHATS research program. [2001 McAfee, Inc.]
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 #include "opt_inet.h"
39 #include "opt_inet6.h"
40 #include "opt_ipsec.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/hash.h>
45 #include <sys/refcount.h>
46 #include <sys/kernel.h>
47 #include <sys/sysctl.h>
48 #include <sys/limits.h>
49 #include <sys/lock.h>
50 #include <sys/mutex.h>
51 #include <sys/malloc.h>
52 #include <sys/mbuf.h>
53 #include <sys/proc.h>		/* for proc0 declaration */
54 #include <sys/random.h>
55 #include <sys/socket.h>
56 #include <sys/socketvar.h>
57 #include <sys/syslog.h>
58 #include <sys/ucred.h>
59 
60 #include <sys/md5.h>
61 #include <crypto/siphash/siphash.h>
62 
63 #include <vm/uma.h>
64 
65 #include <net/if.h>
66 #include <net/if_var.h>
67 #include <net/route.h>
68 #include <net/vnet.h>
69 
70 #include <netinet/in.h>
71 #include <netinet/in_kdtrace.h>
72 #include <netinet/in_systm.h>
73 #include <netinet/ip.h>
74 #include <netinet/in_var.h>
75 #include <netinet/in_pcb.h>
76 #include <netinet/ip_var.h>
77 #include <netinet/ip_options.h>
78 #ifdef INET6
79 #include <netinet/ip6.h>
80 #include <netinet/icmp6.h>
81 #include <netinet6/nd6.h>
82 #include <netinet6/ip6_var.h>
83 #include <netinet6/in6_pcb.h>
84 #endif
85 #include <netinet/tcp.h>
86 #include <netinet/tcp_fastopen.h>
87 #include <netinet/tcp_fsm.h>
88 #include <netinet/tcp_seq.h>
89 #include <netinet/tcp_timer.h>
90 #include <netinet/tcp_var.h>
91 #include <netinet/tcp_syncache.h>
92 #include <netinet/tcp_ecn.h>
93 #ifdef TCP_BLACKBOX
94 #include <netinet/tcp_log_buf.h>
95 #endif
96 #ifdef TCP_OFFLOAD
97 #include <netinet/toecore.h>
98 #endif
99 #include <netinet/udp.h>
100 
101 #include <netipsec/ipsec_support.h>
102 
103 #include <machine/in_cksum.h>
104 
105 #include <security/mac/mac_framework.h>
106 
107 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1;
108 #define	V_tcp_syncookies		VNET(tcp_syncookies)
109 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
110     &VNET_NAME(tcp_syncookies), 0,
111     "Use TCP SYN cookies if the syncache overflows");
112 
113 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0;
114 #define	V_tcp_syncookiesonly		VNET(tcp_syncookiesonly)
115 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
116     &VNET_NAME(tcp_syncookiesonly), 0,
117     "Use only TCP SYN cookies");
118 
119 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1;
120 #define V_functions_inherit_listen_socket_stack \
121     VNET(functions_inherit_listen_socket_stack)
122 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack,
123     CTLFLAG_VNET | CTLFLAG_RW,
124     &VNET_NAME(functions_inherit_listen_socket_stack), 0,
125     "Inherit listen socket's stack");
126 
127 #ifdef TCP_OFFLOAD
128 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
129 #endif
130 
131 static void	 syncache_drop(struct syncache *, struct syncache_head *);
132 static void	 syncache_free(struct syncache *);
133 static void	 syncache_insert(struct syncache *, struct syncache_head *);
134 static int	 syncache_respond(struct syncache *, 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 *, uint16_t);
148 static void	syncache_pause(struct in_conninfo *);
149 static void	syncache_unpause(void *);
150 static void	 syncookie_reseed(void *);
151 #ifdef INVARIANTS
152 static int	 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
153 		    struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
154 		    struct socket *lso, uint16_t port);
155 #endif
156 
157 /*
158  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
159  * 3 retransmits corresponds to a timeout with default values of
160  * tcp_rexmit_initial * (             1 +
161  *                       tcp_backoff[1] +
162  *                       tcp_backoff[2] +
163  *                       tcp_backoff[3]) + 3 * tcp_rexmit_slop,
164  * 1000 ms * (1 + 2 + 4 + 8) +  3 * 200 ms = 15600 ms,
165  * the odds are that the user has given up attempting to connect by then.
166  */
167 #define SYNCACHE_MAXREXMTS		3
168 
169 /* Arbitrary values */
170 #define TCP_SYNCACHE_HASHSIZE		512
171 #define TCP_SYNCACHE_BUCKETLIMIT	30
172 
173 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
174 #define	V_tcp_syncache			VNET(tcp_syncache)
175 
176 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache,
177     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
178     "TCP SYN cache");
179 
180 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
181     &VNET_NAME(tcp_syncache.bucket_limit), 0,
182     "Per-bucket hash limit for syncache");
183 
184 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
185     &VNET_NAME(tcp_syncache.cache_limit), 0,
186     "Overall entry limit for syncache");
187 
188 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
189     &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
190 
191 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
192     &VNET_NAME(tcp_syncache.hashsize), 0,
193     "Size of TCP syncache hashtable");
194 
195 SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET |
196     CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0,
197     "All syncache(4) entries are visible, ignoring UID/GID, jail(2) "
198     "and mac(4) checks");
199 
200 static int
201 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
202 {
203 	int error;
204 	u_int new;
205 
206 	new = V_tcp_syncache.rexmt_limit;
207 	error = sysctl_handle_int(oidp, &new, 0, req);
208 	if ((error == 0) && (req->newptr != NULL)) {
209 		if (new > TCP_MAXRXTSHIFT)
210 			error = EINVAL;
211 		else
212 			V_tcp_syncache.rexmt_limit = new;
213 	}
214 	return (error);
215 }
216 
217 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
218     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
219     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
220     sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
221     "Limit on SYN/ACK retransmissions");
222 
223 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
224 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
225     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
226     "Send reset on socket allocation failure");
227 
228 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
229 
230 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
231 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
232 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
233 
234 /*
235  * Requires the syncache entry to be already removed from the bucket list.
236  */
237 static void
238 syncache_free(struct syncache *sc)
239 {
240 
241 	if (sc->sc_ipopts)
242 		(void) m_free(sc->sc_ipopts);
243 	if (sc->sc_cred)
244 		crfree(sc->sc_cred);
245 #ifdef MAC
246 	mac_syncache_destroy(&sc->sc_label);
247 #endif
248 
249 	uma_zfree(V_tcp_syncache.zone, sc);
250 }
251 
252 void
253 syncache_init(void)
254 {
255 	int i;
256 
257 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
258 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
259 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
260 	V_tcp_syncache.hash_secret = arc4random();
261 
262 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
263 	    &V_tcp_syncache.hashsize);
264 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
265 	    &V_tcp_syncache.bucket_limit);
266 	if (!powerof2(V_tcp_syncache.hashsize) ||
267 	    V_tcp_syncache.hashsize == 0) {
268 		printf("WARNING: syncache hash size is not a power of 2.\n");
269 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
270 	}
271 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
272 
273 	/* Set limits. */
274 	V_tcp_syncache.cache_limit =
275 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
276 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
277 	    &V_tcp_syncache.cache_limit);
278 
279 	/* Allocate the hash table. */
280 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
281 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
282 
283 #ifdef VIMAGE
284 	V_tcp_syncache.vnet = curvnet;
285 #endif
286 
287 	/* Initialize the hash buckets. */
288 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
289 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
290 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
291 			 NULL, MTX_DEF);
292 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
293 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
294 		V_tcp_syncache.hashbase[i].sch_length = 0;
295 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
296 		V_tcp_syncache.hashbase[i].sch_last_overflow =
297 		    -(SYNCOOKIE_LIFETIME + 1);
298 	}
299 
300 	/* Create the syncache entry zone. */
301 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
302 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
303 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
304 	    V_tcp_syncache.cache_limit);
305 
306 	/* Start the SYN cookie reseeder callout. */
307 	callout_init(&V_tcp_syncache.secret.reseed, 1);
308 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
309 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
310 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
311 	    syncookie_reseed, &V_tcp_syncache);
312 
313 	/* Initialize the pause machinery. */
314 	mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF);
315 	callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx,
316 	    0);
317 	V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME;
318 	V_tcp_syncache.pause_backoff = 0;
319 	V_tcp_syncache.paused = false;
320 }
321 
322 #ifdef VIMAGE
323 void
324 syncache_destroy(void)
325 {
326 	struct syncache_head *sch;
327 	struct syncache *sc, *nsc;
328 	int i;
329 
330 	/*
331 	 * Stop the re-seed timer before freeing resources.  No need to
332 	 * possibly schedule it another time.
333 	 */
334 	callout_drain(&V_tcp_syncache.secret.reseed);
335 
336 	/* Stop the SYN cache pause callout. */
337 	mtx_lock(&V_tcp_syncache.pause_mtx);
338 	if (callout_stop(&V_tcp_syncache.pause_co) == 0) {
339 		mtx_unlock(&V_tcp_syncache.pause_mtx);
340 		callout_drain(&V_tcp_syncache.pause_co);
341 	} else
342 		mtx_unlock(&V_tcp_syncache.pause_mtx);
343 
344 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
345 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
346 		sch = &V_tcp_syncache.hashbase[i];
347 		callout_drain(&sch->sch_timer);
348 
349 		SCH_LOCK(sch);
350 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
351 			syncache_drop(sc, sch);
352 		SCH_UNLOCK(sch);
353 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
354 		    ("%s: sch->sch_bucket not empty", __func__));
355 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
356 		    __func__, sch->sch_length));
357 		mtx_destroy(&sch->sch_mtx);
358 	}
359 
360 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
361 	    ("%s: cache_count not 0", __func__));
362 
363 	/* Free the allocated global resources. */
364 	uma_zdestroy(V_tcp_syncache.zone);
365 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
366 	mtx_destroy(&V_tcp_syncache.pause_mtx);
367 }
368 #endif
369 
370 /*
371  * Inserts a syncache entry into the specified bucket row.
372  * Locks and unlocks the syncache_head autonomously.
373  */
374 static void
375 syncache_insert(struct syncache *sc, struct syncache_head *sch)
376 {
377 	struct syncache *sc2;
378 
379 	SCH_LOCK(sch);
380 
381 	/*
382 	 * Make sure that we don't overflow the per-bucket limit.
383 	 * If the bucket is full, toss the oldest element.
384 	 */
385 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
386 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
387 			("sch->sch_length incorrect"));
388 		syncache_pause(&sc->sc_inc);
389 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
390 		sch->sch_last_overflow = time_uptime;
391 		syncache_drop(sc2, sch);
392 	}
393 
394 	/* Put it into the bucket. */
395 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
396 	sch->sch_length++;
397 
398 #ifdef TCP_OFFLOAD
399 	if (ADDED_BY_TOE(sc)) {
400 		struct toedev *tod = sc->sc_tod;
401 
402 		tod->tod_syncache_added(tod, sc->sc_todctx);
403 	}
404 #endif
405 
406 	/* Reinitialize the bucket row's timer. */
407 	if (sch->sch_length == 1)
408 		sch->sch_nextc = ticks + INT_MAX;
409 	syncache_timeout(sc, sch, 1);
410 
411 	SCH_UNLOCK(sch);
412 
413 	TCPSTATES_INC(TCPS_SYN_RECEIVED);
414 	TCPSTAT_INC(tcps_sc_added);
415 }
416 
417 /*
418  * Remove and free entry from syncache bucket row.
419  * Expects locked syncache head.
420  */
421 static void
422 syncache_drop(struct syncache *sc, struct syncache_head *sch)
423 {
424 
425 	SCH_LOCK_ASSERT(sch);
426 
427 	TCPSTATES_DEC(TCPS_SYN_RECEIVED);
428 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
429 	sch->sch_length--;
430 
431 #ifdef TCP_OFFLOAD
432 	if (ADDED_BY_TOE(sc)) {
433 		struct toedev *tod = sc->sc_tod;
434 
435 		tod->tod_syncache_removed(tod, sc->sc_todctx);
436 	}
437 #endif
438 
439 	syncache_free(sc);
440 }
441 
442 /*
443  * Engage/reengage time on bucket row.
444  */
445 static void
446 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
447 {
448 	int rexmt;
449 
450 	if (sc->sc_rxmits == 0)
451 		rexmt = tcp_rexmit_initial;
452 	else
453 		TCPT_RANGESET(rexmt,
454 		    tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
455 		    tcp_rexmit_min, TCPTV_REXMTMAX);
456 	sc->sc_rxttime = ticks + rexmt;
457 	sc->sc_rxmits++;
458 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
459 		sch->sch_nextc = sc->sc_rxttime;
460 		if (docallout)
461 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
462 			    syncache_timer, (void *)sch);
463 	}
464 }
465 
466 /*
467  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
468  * If we have retransmitted an entry the maximum number of times, expire it.
469  * One separate timer for each bucket row.
470  */
471 static void
472 syncache_timer(void *xsch)
473 {
474 	struct syncache_head *sch = (struct syncache_head *)xsch;
475 	struct syncache *sc, *nsc;
476 	struct epoch_tracker et;
477 	int tick = ticks;
478 	char *s;
479 	bool paused;
480 
481 	CURVNET_SET(sch->sch_sc->vnet);
482 
483 	/* NB: syncache_head has already been locked by the callout. */
484 	SCH_LOCK_ASSERT(sch);
485 
486 	/*
487 	 * In the following cycle we may remove some entries and/or
488 	 * advance some timeouts, so re-initialize the bucket timer.
489 	 */
490 	sch->sch_nextc = tick + INT_MAX;
491 
492 	/*
493 	 * If we have paused processing, unconditionally remove
494 	 * all syncache entries.
495 	 */
496 	mtx_lock(&V_tcp_syncache.pause_mtx);
497 	paused = V_tcp_syncache.paused;
498 	mtx_unlock(&V_tcp_syncache.pause_mtx);
499 
500 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
501 		if (paused) {
502 			syncache_drop(sc, sch);
503 			continue;
504 		}
505 		/*
506 		 * We do not check if the listen socket still exists
507 		 * and accept the case where the listen socket may be
508 		 * gone by the time we resend the SYN/ACK.  We do
509 		 * not expect this to happens often. If it does,
510 		 * then the RST will be sent by the time the remote
511 		 * host does the SYN/ACK->ACK.
512 		 */
513 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
514 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
515 				sch->sch_nextc = sc->sc_rxttime;
516 			continue;
517 		}
518 		if (sc->sc_rxmits > V_tcp_ecn_maxretries) {
519 			sc->sc_flags &= ~SCF_ECN_MASK;
520 		}
521 		if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
522 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
523 				log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
524 				    "giving up and removing syncache entry\n",
525 				    s, __func__);
526 				free(s, M_TCPLOG);
527 			}
528 			syncache_drop(sc, sch);
529 			TCPSTAT_INC(tcps_sc_stale);
530 			continue;
531 		}
532 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
533 			log(LOG_DEBUG, "%s; %s: Response timeout, "
534 			    "retransmitting (%u) SYN|ACK\n",
535 			    s, __func__, sc->sc_rxmits);
536 			free(s, M_TCPLOG);
537 		}
538 
539 		NET_EPOCH_ENTER(et);
540 		syncache_respond(sc, NULL, TH_SYN|TH_ACK);
541 		NET_EPOCH_EXIT(et);
542 		TCPSTAT_INC(tcps_sc_retransmitted);
543 		syncache_timeout(sc, sch, 0);
544 	}
545 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
546 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
547 			syncache_timer, (void *)(sch));
548 	CURVNET_RESTORE();
549 }
550 
551 /*
552  * Returns true if the system is only using cookies at the moment.
553  * This could be due to a sysadmin decision to only use cookies, or it
554  * could be due to the system detecting an attack.
555  */
556 static inline bool
557 syncache_cookiesonly(void)
558 {
559 
560 	return (V_tcp_syncookies && (V_tcp_syncache.paused ||
561 	    V_tcp_syncookiesonly));
562 }
563 
564 /*
565  * Find the hash bucket for the given connection.
566  */
567 static struct syncache_head *
568 syncache_hashbucket(struct in_conninfo *inc)
569 {
570 	uint32_t hash;
571 
572 	/*
573 	 * The hash is built on foreign port + local port + foreign address.
574 	 * We rely on the fact that struct in_conninfo starts with 16 bits
575 	 * of foreign port, then 16 bits of local port then followed by 128
576 	 * bits of foreign address.  In case of IPv4 address, the first 3
577 	 * 32-bit words of the address always are zeroes.
578 	 */
579 	hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
580 	    V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
581 
582 	return (&V_tcp_syncache.hashbase[hash]);
583 }
584 
585 /*
586  * Find an entry in the syncache.
587  * Returns always with locked syncache_head plus a matching entry or NULL.
588  */
589 static struct syncache *
590 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
591 {
592 	struct syncache *sc;
593 	struct syncache_head *sch;
594 
595 	*schp = sch = syncache_hashbucket(inc);
596 	SCH_LOCK(sch);
597 
598 	/* Circle through bucket row to find matching entry. */
599 	TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
600 		if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
601 		    sizeof(struct in_endpoints)) == 0)
602 			break;
603 
604 	return (sc);	/* Always returns with locked sch. */
605 }
606 
607 /*
608  * This function is called when we get a RST for a
609  * non-existent connection, so that we can see if the
610  * connection is in the syn cache.  If it is, zap it.
611  * If required send a challenge ACK.
612  */
613 void
614 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m,
615     uint16_t port)
616 {
617 	struct syncache *sc;
618 	struct syncache_head *sch;
619 	char *s = NULL;
620 
621 	if (syncache_cookiesonly())
622 		return;
623 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
624 	SCH_LOCK_ASSERT(sch);
625 
626 	/*
627 	 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
628 	 * See RFC 793 page 65, section SEGMENT ARRIVES.
629 	 */
630 	if (tcp_get_flags(th) & (TH_ACK|TH_SYN|TH_FIN)) {
631 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
632 			log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
633 			    "FIN flag set, segment ignored\n", s, __func__);
634 		TCPSTAT_INC(tcps_badrst);
635 		goto done;
636 	}
637 
638 	/*
639 	 * No corresponding connection was found in syncache.
640 	 * If syncookies are enabled and possibly exclusively
641 	 * used, or we are under memory pressure, a valid RST
642 	 * may not find a syncache entry.  In that case we're
643 	 * done and no SYN|ACK retransmissions will happen.
644 	 * Otherwise the RST was misdirected or spoofed.
645 	 */
646 	if (sc == NULL) {
647 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
648 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
649 			    "syncache entry (possibly syncookie only), "
650 			    "segment ignored\n", s, __func__);
651 		TCPSTAT_INC(tcps_badrst);
652 		goto done;
653 	}
654 
655 	/* The remote UDP encaps port does not match. */
656 	if (sc->sc_port != port) {
657 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
658 			log(LOG_DEBUG, "%s; %s: Spurious RST with matching "
659 			    "syncache entry but non-matching UDP encaps port, "
660 			    "segment ignored\n", s, __func__);
661 		TCPSTAT_INC(tcps_badrst);
662 		goto done;
663 	}
664 
665 	/*
666 	 * If the RST bit is set, check the sequence number to see
667 	 * if this is a valid reset segment.
668 	 *
669 	 * RFC 793 page 37:
670 	 *   In all states except SYN-SENT, all reset (RST) segments
671 	 *   are validated by checking their SEQ-fields.  A reset is
672 	 *   valid if its sequence number is in the window.
673 	 *
674 	 * RFC 793 page 69:
675 	 *   There are four cases for the acceptability test for an incoming
676 	 *   segment:
677 	 *
678 	 * Segment Receive  Test
679 	 * Length  Window
680 	 * ------- -------  -------------------------------------------
681 	 *    0       0     SEG.SEQ = RCV.NXT
682 	 *    0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
683 	 *   >0       0     not acceptable
684 	 *   >0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
685 	 *               or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
686 	 *
687 	 * Note that when receiving a SYN segment in the LISTEN state,
688 	 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
689 	 * described in RFC 793, page 66.
690 	 */
691 	if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
692 	    SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
693 	    (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
694 		if (V_tcp_insecure_rst ||
695 		    th->th_seq == sc->sc_irs + 1) {
696 			syncache_drop(sc, sch);
697 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
698 				log(LOG_DEBUG,
699 				    "%s; %s: Our SYN|ACK was rejected, "
700 				    "connection attempt aborted by remote "
701 				    "endpoint\n",
702 				    s, __func__);
703 			TCPSTAT_INC(tcps_sc_reset);
704 		} else {
705 			TCPSTAT_INC(tcps_badrst);
706 			/* Send challenge ACK. */
707 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
708 				log(LOG_DEBUG, "%s; %s: RST with invalid "
709 				    " SEQ %u != NXT %u (+WND %u), "
710 				    "sending challenge ACK\n",
711 				    s, __func__,
712 				    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
713 			syncache_respond(sc, m, TH_ACK);
714 		}
715 	} else {
716 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
717 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
718 			    "NXT %u (+WND %u), segment ignored\n",
719 			    s, __func__,
720 			    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
721 		TCPSTAT_INC(tcps_badrst);
722 	}
723 
724 done:
725 	if (s != NULL)
726 		free(s, M_TCPLOG);
727 	SCH_UNLOCK(sch);
728 }
729 
730 void
731 syncache_badack(struct in_conninfo *inc, uint16_t port)
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) && (sc->sc_port == port)) {
741 		syncache_drop(sc, sch);
742 		TCPSTAT_INC(tcps_sc_badack);
743 	}
744 	SCH_UNLOCK(sch);
745 }
746 
747 void
748 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port)
749 {
750 	struct syncache *sc;
751 	struct syncache_head *sch;
752 
753 	if (syncache_cookiesonly())
754 		return;
755 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
756 	SCH_LOCK_ASSERT(sch);
757 	if (sc == NULL)
758 		goto done;
759 
760 	/* If the port != sc_port, then it's a bogus ICMP msg */
761 	if (port != sc->sc_port)
762 		goto done;
763 
764 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
765 	if (ntohl(th_seq) != sc->sc_iss)
766 		goto done;
767 
768 	/*
769 	 * If we've rertransmitted 3 times and this is our second error,
770 	 * we remove the entry.  Otherwise, we allow it to continue on.
771 	 * This prevents us from incorrectly nuking an entry during a
772 	 * spurious network outage.
773 	 *
774 	 * See tcp_notify().
775 	 */
776 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
777 		sc->sc_flags |= SCF_UNREACH;
778 		goto done;
779 	}
780 	syncache_drop(sc, sch);
781 	TCPSTAT_INC(tcps_sc_unreach);
782 done:
783 	SCH_UNLOCK(sch);
784 }
785 
786 /*
787  * Build a new TCP socket structure from a syncache entry.
788  *
789  * On success return the newly created socket with its underlying inp locked.
790  */
791 static struct socket *
792 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
793 {
794 	struct tcp_function_block *blk;
795 	struct inpcb *inp = NULL;
796 	struct socket *so;
797 	struct tcpcb *tp;
798 	int error;
799 	char *s;
800 
801 	NET_EPOCH_ASSERT();
802 
803 	/*
804 	 * Ok, create the full blown connection, and set things up
805 	 * as they would have been set up if we had created the
806 	 * connection when the SYN arrived.
807 	 */
808 	if ((so = solisten_clone(lso)) == NULL)
809 		goto allocfail;
810 #ifdef MAC
811 	mac_socketpeer_set_from_mbuf(m, so);
812 #endif
813 	error = in_pcballoc(so, &V_tcbinfo);
814 	if (error) {
815 		sodealloc(so);
816 		goto allocfail;
817 	}
818 	inp = sotoinpcb(so);
819 	if ((tp = tcp_newtcpcb(inp)) == NULL) {
820 		in_pcbdetach(inp);
821 		in_pcbfree(inp);
822 		sodealloc(so);
823 		goto allocfail;
824 	}
825 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
826 #ifdef INET6
827 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
828 		inp->inp_vflag &= ~INP_IPV4;
829 		inp->inp_vflag |= INP_IPV6;
830 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
831 	} else {
832 		inp->inp_vflag &= ~INP_IPV6;
833 		inp->inp_vflag |= INP_IPV4;
834 #endif
835 		inp->inp_ip_ttl = sc->sc_ip_ttl;
836 		inp->inp_ip_tos = sc->sc_ip_tos;
837 		inp->inp_laddr = sc->sc_inc.inc_laddr;
838 #ifdef INET6
839 	}
840 #endif
841 
842 	/*
843 	 * If there's an mbuf and it has a flowid, then let's initialise the
844 	 * inp with that particular flowid.
845 	 */
846 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
847 		inp->inp_flowid = m->m_pkthdr.flowid;
848 		inp->inp_flowtype = M_HASHTYPE_GET(m);
849 #ifdef NUMA
850 		inp->inp_numa_domain = m->m_pkthdr.numa_domain;
851 #endif
852 	}
853 
854 	inp->inp_lport = sc->sc_inc.inc_lport;
855 #ifdef INET6
856 	if (inp->inp_vflag & INP_IPV6PROTO) {
857 		struct inpcb *oinp = sotoinpcb(lso);
858 
859 		/*
860 		 * Inherit socket options from the listening socket.
861 		 * Note that in6p_inputopts are not (and should not be)
862 		 * copied, since it stores previously received options and is
863 		 * used to detect if each new option is different than the
864 		 * previous one and hence should be passed to a user.
865 		 * If we copied in6p_inputopts, a user would not be able to
866 		 * receive options just after calling the accept system call.
867 		 */
868 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
869 		if (oinp->in6p_outputopts)
870 			inp->in6p_outputopts =
871 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
872 		inp->in6p_hops = oinp->in6p_hops;
873 	}
874 
875 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
876 		struct sockaddr_in6 sin6;
877 
878 		sin6.sin6_family = AF_INET6;
879 		sin6.sin6_len = sizeof(sin6);
880 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
881 		sin6.sin6_port = sc->sc_inc.inc_fport;
882 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
883 		INP_HASH_WLOCK(&V_tcbinfo);
884 		error = in6_pcbconnect(inp, &sin6, thread0.td_ucred, false);
885 		INP_HASH_WUNLOCK(&V_tcbinfo);
886 		if (error != 0)
887 			goto abort;
888 		/* Override flowlabel from in6_pcbconnect. */
889 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
890 		inp->inp_flow |= sc->sc_flowlabel;
891 	}
892 #endif /* INET6 */
893 #if defined(INET) && defined(INET6)
894 	else
895 #endif
896 #ifdef INET
897 	{
898 		struct sockaddr_in sin;
899 
900 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
901 
902 		if (inp->inp_options == NULL) {
903 			inp->inp_options = sc->sc_ipopts;
904 			sc->sc_ipopts = NULL;
905 		}
906 
907 		sin.sin_family = AF_INET;
908 		sin.sin_len = sizeof(sin);
909 		sin.sin_addr = sc->sc_inc.inc_faddr;
910 		sin.sin_port = sc->sc_inc.inc_fport;
911 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
912 		INP_HASH_WLOCK(&V_tcbinfo);
913 		error = in_pcbconnect(inp, &sin, thread0.td_ucred, false);
914 		INP_HASH_WUNLOCK(&V_tcbinfo);
915 		if (error != 0)
916 			goto abort;
917 	}
918 #endif /* INET */
919 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
920 	/* Copy old policy into new socket's. */
921 	if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
922 		printf("syncache_socket: could not copy policy\n");
923 #endif
924 	tp->t_state = TCPS_SYN_RECEIVED;
925 	tp->iss = sc->sc_iss;
926 	tp->irs = sc->sc_irs;
927 	tp->t_port = sc->sc_port;
928 	tcp_rcvseqinit(tp);
929 	tcp_sendseqinit(tp);
930 	blk = sototcpcb(lso)->t_fb;
931 	if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
932 		/*
933 		 * Our parents t_fb was not the default,
934 		 * we need to release our ref on tp->t_fb and
935 		 * pickup one on the new entry.
936 		 */
937 		struct tcp_function_block *rblk;
938 		void *ptr = NULL;
939 
940 		rblk = find_and_ref_tcp_fb(blk);
941 		KASSERT(rblk != NULL,
942 		    ("cannot find blk %p out of syncache?", blk));
943 
944 		if (rblk->tfb_tcp_fb_init == NULL ||
945 		    (*rblk->tfb_tcp_fb_init)(tp, &ptr) == 0) {
946 			/* Release the old stack */
947 			if (tp->t_fb->tfb_tcp_fb_fini != NULL)
948 				(*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
949 			refcount_release(&tp->t_fb->tfb_refcnt);
950 			/* Now set in all the pointers */
951 			tp->t_fb = rblk;
952 			tp->t_fb_ptr = ptr;
953 		} else {
954 			/*
955 			 * Initialization failed. Release the reference count on
956 			 * the looked up default stack.
957 			 */
958 			refcount_release(&rblk->tfb_refcnt);
959 		}
960 	}
961 	tp->snd_wl1 = sc->sc_irs;
962 	tp->snd_max = tp->iss + 1;
963 	tp->snd_nxt = tp->iss + 1;
964 	tp->rcv_up = sc->sc_irs + 1;
965 	tp->rcv_wnd = sc->sc_wnd;
966 	tp->rcv_adv += tp->rcv_wnd;
967 	tp->last_ack_sent = tp->rcv_nxt;
968 
969 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
970 	if (sc->sc_flags & SCF_NOOPT)
971 		tp->t_flags |= TF_NOOPT;
972 	else {
973 		if (sc->sc_flags & SCF_WINSCALE) {
974 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
975 			tp->snd_scale = sc->sc_requested_s_scale;
976 			tp->request_r_scale = sc->sc_requested_r_scale;
977 		}
978 		if (sc->sc_flags & SCF_TIMESTAMP) {
979 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
980 			tp->ts_recent = sc->sc_tsreflect;
981 			tp->ts_recent_age = tcp_ts_getticks();
982 			tp->ts_offset = sc->sc_tsoff;
983 		}
984 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
985 		if (sc->sc_flags & SCF_SIGNATURE)
986 			tp->t_flags |= TF_SIGNATURE;
987 #endif
988 		if (sc->sc_flags & SCF_SACK)
989 			tp->t_flags |= TF_SACK_PERMIT;
990 	}
991 
992 	tcp_ecn_syncache_socket(tp, sc);
993 
994 	/*
995 	 * Set up MSS and get cached values from tcp_hostcache.
996 	 * This might overwrite some of the defaults we just set.
997 	 */
998 	tcp_mss(tp, sc->sc_peer_mss);
999 
1000 	/*
1001 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
1002 	 * limited to one segment in cc_conn_init().
1003 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
1004 	 */
1005 	if (sc->sc_rxmits > 1)
1006 		tp->snd_cwnd = 1;
1007 
1008 #ifdef TCP_OFFLOAD
1009 	/*
1010 	 * Allow a TOE driver to install its hooks.  Note that we hold the
1011 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
1012 	 * new connection before the TOE driver has done its thing.
1013 	 */
1014 	if (ADDED_BY_TOE(sc)) {
1015 		struct toedev *tod = sc->sc_tod;
1016 
1017 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
1018 	}
1019 #endif
1020 #ifdef TCP_BLACKBOX
1021 	/*
1022 	 * Inherit the log state from the listening socket, if
1023 	 * - the log state of the listening socket is not off and
1024 	 * - the listening socket was not auto selected from all sessions and
1025 	 * - a log id is not set on the listening socket.
1026 	 * This avoids inheriting a log state which was automatically set.
1027 	 */
1028 	if ((tcp_get_bblog_state(sototcpcb(lso)) != TCP_LOG_STATE_OFF) &&
1029 	    ((sototcpcb(lso)->t_flags2 & TF2_LOG_AUTO) == 0) &&
1030 	    (sototcpcb(lso)->t_lib == NULL)) {
1031 		tcp_log_state_change(tp, tcp_get_bblog_state(sototcpcb(lso)));
1032 	}
1033 #endif
1034 	/*
1035 	 * Copy and activate timers.
1036 	 */
1037 	tp->t_maxunacktime = sototcpcb(lso)->t_maxunacktime;
1038 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
1039 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
1040 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
1041 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1042 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1043 
1044 	TCPSTAT_INC(tcps_accepts);
1045 	TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, TCPS_LISTEN);
1046 
1047 	if (!solisten_enqueue(so, SS_ISCONNECTED))
1048 		tp->t_flags |= TF_SONOTCONN;
1049 
1050 	return (so);
1051 
1052 allocfail:
1053 	/*
1054 	 * Drop the connection; we will either send a RST or have the peer
1055 	 * retransmit its SYN again after its RTO and try again.
1056 	 */
1057 	if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1058 		log(LOG_DEBUG, "%s; %s: Socket create failed "
1059 		    "due to limits or memory shortage\n",
1060 		    s, __func__);
1061 		free(s, M_TCPLOG);
1062 	}
1063 	TCPSTAT_INC(tcps_listendrop);
1064 	return (NULL);
1065 
1066 abort:
1067 	in_pcbdetach(inp);
1068 	in_pcbfree(inp);
1069 	sodealloc(so);
1070 	if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1071 		log(LOG_DEBUG, "%s; %s: in%s_pcbconnect failed with error %i\n",
1072 		    s, __func__, (sc->sc_inc.inc_flags & INC_ISIPV6) ? "6" : "",
1073 		    error);
1074 		free(s, M_TCPLOG);
1075 	}
1076 	TCPSTAT_INC(tcps_listendrop);
1077 	return (NULL);
1078 }
1079 
1080 /*
1081  * This function gets called when we receive an ACK for a
1082  * socket in the LISTEN state.  We look up the connection
1083  * in the syncache, and if its there, we pull it out of
1084  * the cache and turn it into a full-blown connection in
1085  * the SYN-RECEIVED state.
1086  *
1087  * On syncache_socket() success the newly created socket
1088  * has its underlying inp locked.
1089  */
1090 int
1091 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1092     struct socket **lsop, struct mbuf *m, uint16_t port)
1093 {
1094 	struct syncache *sc;
1095 	struct syncache_head *sch;
1096 	struct syncache scs;
1097 	char *s;
1098 	bool locked;
1099 
1100 	NET_EPOCH_ASSERT();
1101 	KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1102 	    ("%s: can handle only ACK", __func__));
1103 
1104 	if (syncache_cookiesonly()) {
1105 		sc = NULL;
1106 		sch = syncache_hashbucket(inc);
1107 		locked = false;
1108 	} else {
1109 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1110 		locked = true;
1111 		SCH_LOCK_ASSERT(sch);
1112 	}
1113 
1114 #ifdef INVARIANTS
1115 	/*
1116 	 * Test code for syncookies comparing the syncache stored
1117 	 * values with the reconstructed values from the cookie.
1118 	 */
1119 	if (sc != NULL)
1120 		syncookie_cmp(inc, sch, sc, th, to, *lsop, port);
1121 #endif
1122 
1123 	if (sc == NULL) {
1124 		/*
1125 		 * There is no syncache entry, so see if this ACK is
1126 		 * a returning syncookie.  To do this, first:
1127 		 *  A. Check if syncookies are used in case of syncache
1128 		 *     overflows
1129 		 *  B. See if this socket has had a syncache entry dropped in
1130 		 *     the recent past. We don't want to accept a bogus
1131 		 *     syncookie if we've never received a SYN or accept it
1132 		 *     twice.
1133 		 *  C. check that the syncookie is valid.  If it is, then
1134 		 *     cobble up a fake syncache entry, and return.
1135 		 */
1136 		if (locked && !V_tcp_syncookies) {
1137 			SCH_UNLOCK(sch);
1138 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1139 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1140 				    "segment rejected (syncookies disabled)\n",
1141 				    s, __func__);
1142 			goto failed;
1143 		}
1144 		if (locked && !V_tcp_syncookiesonly &&
1145 		    sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1146 			SCH_UNLOCK(sch);
1147 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1148 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1149 				    "segment rejected (no syncache entry)\n",
1150 				    s, __func__);
1151 			goto failed;
1152 		}
1153 		bzero(&scs, sizeof(scs));
1154 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port);
1155 		if (locked)
1156 			SCH_UNLOCK(sch);
1157 		if (sc == NULL) {
1158 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1159 				log(LOG_DEBUG, "%s; %s: Segment failed "
1160 				    "SYNCOOKIE authentication, segment rejected "
1161 				    "(probably spoofed)\n", s, __func__);
1162 			goto failed;
1163 		}
1164 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1165 		/* If received ACK has MD5 signature, check it. */
1166 		if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1167 		    (!TCPMD5_ENABLED() ||
1168 		    TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1169 			/* Drop the ACK. */
1170 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1171 				log(LOG_DEBUG, "%s; %s: Segment rejected, "
1172 				    "MD5 signature doesn't match.\n",
1173 				    s, __func__);
1174 				free(s, M_TCPLOG);
1175 			}
1176 			TCPSTAT_INC(tcps_sig_err_sigopt);
1177 			return (-1); /* Do not send RST */
1178 		}
1179 #endif /* TCP_SIGNATURE */
1180 		TCPSTATES_INC(TCPS_SYN_RECEIVED);
1181 	} else {
1182 		if (sc->sc_port != port) {
1183 			SCH_UNLOCK(sch);
1184 			return (0);
1185 		}
1186 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1187 		/*
1188 		 * If listening socket requested TCP digests, check that
1189 		 * received ACK has signature and it is correct.
1190 		 * If not, drop the ACK and leave sc entry in th cache,
1191 		 * because SYN was received with correct signature.
1192 		 */
1193 		if (sc->sc_flags & SCF_SIGNATURE) {
1194 			if ((to->to_flags & TOF_SIGNATURE) == 0) {
1195 				/* No signature */
1196 				TCPSTAT_INC(tcps_sig_err_nosigopt);
1197 				SCH_UNLOCK(sch);
1198 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1199 					log(LOG_DEBUG, "%s; %s: Segment "
1200 					    "rejected, MD5 signature wasn't "
1201 					    "provided.\n", s, __func__);
1202 					free(s, M_TCPLOG);
1203 				}
1204 				return (-1); /* Do not send RST */
1205 			}
1206 			if (!TCPMD5_ENABLED() ||
1207 			    TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1208 				/* Doesn't match or no SA */
1209 				SCH_UNLOCK(sch);
1210 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1211 					log(LOG_DEBUG, "%s; %s: Segment "
1212 					    "rejected, MD5 signature doesn't "
1213 					    "match.\n", s, __func__);
1214 					free(s, M_TCPLOG);
1215 				}
1216 				return (-1); /* Do not send RST */
1217 			}
1218 		}
1219 #endif /* TCP_SIGNATURE */
1220 
1221 		/*
1222 		 * RFC 7323 PAWS: If we have a timestamp on this segment and
1223 		 * it's less than ts_recent, drop it.
1224 		 * XXXMT: RFC 7323 also requires to send an ACK.
1225 		 *        In tcp_input.c this is only done for TCP segments
1226 		 *        with user data, so be consistent here and just drop
1227 		 *        the segment.
1228 		 */
1229 		if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1230 		    TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1231 			SCH_UNLOCK(sch);
1232 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1233 				log(LOG_DEBUG,
1234 				    "%s; %s: SEG.TSval %u < TS.Recent %u, "
1235 				    "segment dropped\n", s, __func__,
1236 				    to->to_tsval, sc->sc_tsreflect);
1237 				free(s, M_TCPLOG);
1238 			}
1239 			return (-1);  /* Do not send RST */
1240 		}
1241 
1242 		/*
1243 		 * If timestamps were not negotiated during SYN/ACK and a
1244 		 * segment with a timestamp is received, ignore the
1245 		 * timestamp and process the packet normally.
1246 		 * See section 3.2 of RFC 7323.
1247 		 */
1248 		if (!(sc->sc_flags & SCF_TIMESTAMP) &&
1249 		    (to->to_flags & TOF_TS)) {
1250 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1251 				log(LOG_DEBUG, "%s; %s: Timestamp not "
1252 				    "expected, segment processed normally\n",
1253 				    s, __func__);
1254 				free(s, M_TCPLOG);
1255 				s = NULL;
1256 			}
1257 		}
1258 
1259 		/*
1260 		 * If timestamps were negotiated during SYN/ACK and a
1261 		 * segment without a timestamp is received, silently drop
1262 		 * the segment, unless the missing timestamps are tolerated.
1263 		 * See section 3.2 of RFC 7323.
1264 		 */
1265 		if ((sc->sc_flags & SCF_TIMESTAMP) &&
1266 		    !(to->to_flags & TOF_TS)) {
1267 			if (V_tcp_tolerate_missing_ts) {
1268 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1269 					log(LOG_DEBUG,
1270 					    "%s; %s: Timestamp missing, "
1271 					    "segment processed normally\n",
1272 					    s, __func__);
1273 					free(s, M_TCPLOG);
1274 				}
1275 			} else {
1276 				SCH_UNLOCK(sch);
1277 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1278 					log(LOG_DEBUG,
1279 					    "%s; %s: Timestamp missing, "
1280 					    "segment silently dropped\n",
1281 					    s, __func__);
1282 					free(s, M_TCPLOG);
1283 				}
1284 				return (-1);  /* Do not send RST */
1285 			}
1286 		}
1287 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1288 		sch->sch_length--;
1289 #ifdef TCP_OFFLOAD
1290 		if (ADDED_BY_TOE(sc)) {
1291 			struct toedev *tod = sc->sc_tod;
1292 
1293 			tod->tod_syncache_removed(tod, sc->sc_todctx);
1294 		}
1295 #endif
1296 		SCH_UNLOCK(sch);
1297 	}
1298 
1299 	/*
1300 	 * Segment validation:
1301 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1302 	 */
1303 	if (th->th_ack != sc->sc_iss + 1) {
1304 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1305 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1306 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1307 		goto failed;
1308 	}
1309 
1310 	/*
1311 	 * The SEQ must fall in the window starting at the received
1312 	 * initial receive sequence number + 1 (the SYN).
1313 	 */
1314 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1315 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1316 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1317 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1318 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1319 		goto failed;
1320 	}
1321 
1322 	*lsop = syncache_socket(sc, *lsop, m);
1323 
1324 	if (__predict_false(*lsop == NULL)) {
1325 		TCPSTAT_INC(tcps_sc_aborted);
1326 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1327 	} else
1328 		TCPSTAT_INC(tcps_sc_completed);
1329 
1330 /* how do we find the inp for the new socket? */
1331 	if (sc != &scs)
1332 		syncache_free(sc);
1333 	return (1);
1334 failed:
1335 	if (sc != NULL) {
1336 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1337 		if (sc != &scs)
1338 			syncache_free(sc);
1339 	}
1340 	if (s != NULL)
1341 		free(s, M_TCPLOG);
1342 	*lsop = NULL;
1343 	return (0);
1344 }
1345 
1346 static struct socket *
1347 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m,
1348     uint64_t response_cookie)
1349 {
1350 	struct inpcb *inp;
1351 	struct tcpcb *tp;
1352 	unsigned int *pending_counter;
1353 	struct socket *so;
1354 
1355 	NET_EPOCH_ASSERT();
1356 
1357 	pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending;
1358 	so = syncache_socket(sc, lso, m);
1359 	if (so == NULL) {
1360 		TCPSTAT_INC(tcps_sc_aborted);
1361 		atomic_subtract_int(pending_counter, 1);
1362 	} else {
1363 		soisconnected(so);
1364 		inp = sotoinpcb(so);
1365 		tp = intotcpcb(inp);
1366 		tp->t_flags |= TF_FASTOPEN;
1367 		tp->t_tfo_cookie.server = response_cookie;
1368 		tp->snd_max = tp->iss;
1369 		tp->snd_nxt = tp->iss;
1370 		tp->t_tfo_pending = pending_counter;
1371 		TCPSTATES_INC(TCPS_SYN_RECEIVED);
1372 		TCPSTAT_INC(tcps_sc_completed);
1373 	}
1374 
1375 	return (so);
1376 }
1377 
1378 /*
1379  * Given a LISTEN socket and an inbound SYN request, add
1380  * this to the syn cache, and send back a segment:
1381  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1382  * to the source.
1383  *
1384  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1385  * Doing so would require that we hold onto the data and deliver it
1386  * to the application.  However, if we are the target of a SYN-flood
1387  * DoS attack, an attacker could send data which would eventually
1388  * consume all available buffer space if it were ACKed.  By not ACKing
1389  * the data, we avoid this DoS scenario.
1390  *
1391  * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1392  * cookie is processed and a new socket is created.  In this case, any data
1393  * accompanying the SYN will be queued to the socket by tcp_input() and will
1394  * be ACKed either when the application sends response data or the delayed
1395  * ACK timer expires, whichever comes first.
1396  */
1397 struct socket *
1398 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1399     struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod,
1400     void *todctx, uint8_t iptos, uint16_t port)
1401 {
1402 	struct tcpcb *tp;
1403 	struct socket *rv = NULL;
1404 	struct syncache *sc = NULL;
1405 	struct syncache_head *sch;
1406 	struct mbuf *ipopts = NULL;
1407 	u_int ltflags;
1408 	int win, ip_ttl, ip_tos;
1409 	char *s;
1410 #ifdef INET6
1411 	int autoflowlabel = 0;
1412 #endif
1413 #ifdef MAC
1414 	struct label *maclabel;
1415 #endif
1416 	struct syncache scs;
1417 	struct ucred *cred;
1418 	uint64_t tfo_response_cookie;
1419 	unsigned int *tfo_pending = NULL;
1420 	int tfo_cookie_valid = 0;
1421 	int tfo_response_cookie_valid = 0;
1422 	bool locked;
1423 
1424 	INP_RLOCK_ASSERT(inp);			/* listen socket */
1425 	KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1426 	    ("%s: unexpected tcp flags", __func__));
1427 
1428 	/*
1429 	 * Combine all so/tp operations very early to drop the INP lock as
1430 	 * soon as possible.
1431 	 */
1432 	KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1433 	tp = sototcpcb(so);
1434 	cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred);
1435 
1436 #ifdef INET6
1437 	if (inc->inc_flags & INC_ISIPV6) {
1438 		if (inp->inp_flags & IN6P_AUTOFLOWLABEL) {
1439 			autoflowlabel = 1;
1440 		}
1441 		ip_ttl = in6_selecthlim(inp, NULL);
1442 		if ((inp->in6p_outputopts == NULL) ||
1443 		    (inp->in6p_outputopts->ip6po_tclass == -1)) {
1444 			ip_tos = 0;
1445 		} else {
1446 			ip_tos = inp->in6p_outputopts->ip6po_tclass;
1447 		}
1448 	}
1449 #endif
1450 #if defined(INET6) && defined(INET)
1451 	else
1452 #endif
1453 #ifdef INET
1454 	{
1455 		ip_ttl = inp->inp_ip_ttl;
1456 		ip_tos = inp->inp_ip_tos;
1457 	}
1458 #endif
1459 	win = so->sol_sbrcv_hiwat;
1460 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1461 
1462 	if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
1463 	    (tp->t_tfo_pending != NULL) &&
1464 	    (to->to_flags & TOF_FASTOPEN)) {
1465 		/*
1466 		 * Limit the number of pending TFO connections to
1467 		 * approximately half of the queue limit.  This prevents TFO
1468 		 * SYN floods from starving the service by filling the
1469 		 * listen queue with bogus TFO connections.
1470 		 */
1471 		if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1472 		    (so->sol_qlimit / 2)) {
1473 			int result;
1474 
1475 			result = tcp_fastopen_check_cookie(inc,
1476 			    to->to_tfo_cookie, to->to_tfo_len,
1477 			    &tfo_response_cookie);
1478 			tfo_cookie_valid = (result > 0);
1479 			tfo_response_cookie_valid = (result >= 0);
1480 		}
1481 
1482 		/*
1483 		 * Remember the TFO pending counter as it will have to be
1484 		 * decremented below if we don't make it to syncache_tfo_expand().
1485 		 */
1486 		tfo_pending = tp->t_tfo_pending;
1487 	}
1488 
1489 #ifdef MAC
1490 	if (mac_syncache_init(&maclabel) != 0) {
1491 		INP_RUNLOCK(inp);
1492 		goto done;
1493 	} else
1494 		mac_syncache_create(maclabel, inp);
1495 #endif
1496 	if (!tfo_cookie_valid)
1497 		INP_RUNLOCK(inp);
1498 
1499 	/*
1500 	 * Remember the IP options, if any.
1501 	 */
1502 #ifdef INET6
1503 	if (!(inc->inc_flags & INC_ISIPV6))
1504 #endif
1505 #ifdef INET
1506 		ipopts = (m) ? ip_srcroute(m) : NULL;
1507 #else
1508 		ipopts = NULL;
1509 #endif
1510 
1511 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1512 	/*
1513 	 * When the socket is TCP-MD5 enabled check that,
1514 	 *  - a signed packet is valid
1515 	 *  - a non-signed packet does not have a security association
1516 	 *
1517 	 *  If a signed packet fails validation or a non-signed packet has a
1518 	 *  security association, the packet will be dropped.
1519 	 */
1520 	if (ltflags & TF_SIGNATURE) {
1521 		if (to->to_flags & TOF_SIGNATURE) {
1522 			if (!TCPMD5_ENABLED() ||
1523 			    TCPMD5_INPUT(m, th, to->to_signature) != 0)
1524 				goto done;
1525 		} else {
1526 			if (TCPMD5_ENABLED() &&
1527 			    TCPMD5_INPUT(m, NULL, NULL) != ENOENT)
1528 				goto done;
1529 		}
1530 	} else if (to->to_flags & TOF_SIGNATURE)
1531 		goto done;
1532 #endif	/* TCP_SIGNATURE */
1533 	/*
1534 	 * See if we already have an entry for this connection.
1535 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1536 	 *
1537 	 * XXX: should the syncache be re-initialized with the contents
1538 	 * of the new SYN here (which may have different options?)
1539 	 *
1540 	 * XXX: We do not check the sequence number to see if this is a
1541 	 * real retransmit or a new connection attempt.  The question is
1542 	 * how to handle such a case; either ignore it as spoofed, or
1543 	 * drop the current entry and create a new one?
1544 	 */
1545 	if (syncache_cookiesonly()) {
1546 		sc = NULL;
1547 		sch = syncache_hashbucket(inc);
1548 		locked = false;
1549 	} else {
1550 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1551 		locked = true;
1552 		SCH_LOCK_ASSERT(sch);
1553 	}
1554 	if (sc != NULL) {
1555 		if (tfo_cookie_valid)
1556 			INP_RUNLOCK(inp);
1557 		TCPSTAT_INC(tcps_sc_dupsyn);
1558 		if (ipopts) {
1559 			/*
1560 			 * If we were remembering a previous source route,
1561 			 * forget it and use the new one we've been given.
1562 			 */
1563 			if (sc->sc_ipopts)
1564 				(void) m_free(sc->sc_ipopts);
1565 			sc->sc_ipopts = ipopts;
1566 		}
1567 		/*
1568 		 * Update timestamp if present.
1569 		 */
1570 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1571 			sc->sc_tsreflect = to->to_tsval;
1572 		else
1573 			sc->sc_flags &= ~SCF_TIMESTAMP;
1574 		/*
1575 		 * Adjust ECN response if needed, e.g. different
1576 		 * IP ECN field, or a fallback by the remote host.
1577 		 */
1578 		if (sc->sc_flags & SCF_ECN_MASK) {
1579 			sc->sc_flags &= ~SCF_ECN_MASK;
1580 			sc->sc_flags = tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1581 		}
1582 #ifdef MAC
1583 		/*
1584 		 * Since we have already unconditionally allocated label
1585 		 * storage, free it up.  The syncache entry will already
1586 		 * have an initialized label we can use.
1587 		 */
1588 		mac_syncache_destroy(&maclabel);
1589 #endif
1590 		TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1591 		/* Retransmit SYN|ACK and reset retransmit count. */
1592 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1593 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1594 			    "resetting timer and retransmitting SYN|ACK\n",
1595 			    s, __func__);
1596 			free(s, M_TCPLOG);
1597 		}
1598 		if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1599 			sc->sc_rxmits = 0;
1600 			syncache_timeout(sc, sch, 1);
1601 			TCPSTAT_INC(tcps_sndacks);
1602 			TCPSTAT_INC(tcps_sndtotal);
1603 		}
1604 		SCH_UNLOCK(sch);
1605 		goto donenoprobe;
1606 	}
1607 
1608 	if (tfo_cookie_valid) {
1609 		bzero(&scs, sizeof(scs));
1610 		sc = &scs;
1611 		goto skip_alloc;
1612 	}
1613 
1614 	/*
1615 	 * Skip allocating a syncache entry if we are just going to discard
1616 	 * it later.
1617 	 */
1618 	if (!locked) {
1619 		bzero(&scs, sizeof(scs));
1620 		sc = &scs;
1621 	} else
1622 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1623 	if (sc == NULL) {
1624 		/*
1625 		 * The zone allocator couldn't provide more entries.
1626 		 * Treat this as if the cache was full; drop the oldest
1627 		 * entry and insert the new one.
1628 		 */
1629 		TCPSTAT_INC(tcps_sc_zonefail);
1630 		if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1631 			sch->sch_last_overflow = time_uptime;
1632 			syncache_drop(sc, sch);
1633 			syncache_pause(inc);
1634 		}
1635 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1636 		if (sc == NULL) {
1637 			if (V_tcp_syncookies) {
1638 				bzero(&scs, sizeof(scs));
1639 				sc = &scs;
1640 			} else {
1641 				KASSERT(locked,
1642 				    ("%s: bucket unexpectedly unlocked",
1643 				    __func__));
1644 				SCH_UNLOCK(sch);
1645 				if (ipopts)
1646 					(void) m_free(ipopts);
1647 				goto done;
1648 			}
1649 		}
1650 	}
1651 
1652 skip_alloc:
1653 	if (!tfo_cookie_valid && tfo_response_cookie_valid)
1654 		sc->sc_tfo_cookie = &tfo_response_cookie;
1655 
1656 	/*
1657 	 * Fill in the syncache values.
1658 	 */
1659 #ifdef MAC
1660 	sc->sc_label = maclabel;
1661 #endif
1662 	sc->sc_cred = cred;
1663 	sc->sc_port = port;
1664 	cred = NULL;
1665 	sc->sc_ipopts = ipopts;
1666 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1667 	sc->sc_ip_tos = ip_tos;
1668 	sc->sc_ip_ttl = ip_ttl;
1669 #ifdef TCP_OFFLOAD
1670 	sc->sc_tod = tod;
1671 	sc->sc_todctx = todctx;
1672 #endif
1673 	sc->sc_irs = th->th_seq;
1674 	sc->sc_flags = 0;
1675 	sc->sc_flowlabel = 0;
1676 
1677 	/*
1678 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1679 	 * win was derived from socket earlier in the function.
1680 	 */
1681 	win = imax(win, 0);
1682 	win = imin(win, TCP_MAXWIN);
1683 	sc->sc_wnd = win;
1684 
1685 	if (V_tcp_do_rfc1323 &&
1686 	    !(ltflags & TF_NOOPT)) {
1687 		/*
1688 		 * A timestamp received in a SYN makes
1689 		 * it ok to send timestamp requests and replies.
1690 		 */
1691 		if ((to->to_flags & TOF_TS) && (V_tcp_do_rfc1323 != 2)) {
1692 			sc->sc_tsreflect = to->to_tsval;
1693 			sc->sc_flags |= SCF_TIMESTAMP;
1694 			sc->sc_tsoff = tcp_new_ts_offset(inc);
1695 		}
1696 		if ((to->to_flags & TOF_SCALE) && (V_tcp_do_rfc1323 != 3)) {
1697 			int wscale = 0;
1698 
1699 			/*
1700 			 * Pick the smallest possible scaling factor that
1701 			 * will still allow us to scale up to sb_max, aka
1702 			 * kern.ipc.maxsockbuf.
1703 			 *
1704 			 * We do this because there are broken firewalls that
1705 			 * will corrupt the window scale option, leading to
1706 			 * the other endpoint believing that our advertised
1707 			 * window is unscaled.  At scale factors larger than
1708 			 * 5 the unscaled window will drop below 1500 bytes,
1709 			 * leading to serious problems when traversing these
1710 			 * broken firewalls.
1711 			 *
1712 			 * With the default maxsockbuf of 256K, a scale factor
1713 			 * of 3 will be chosen by this algorithm.  Those who
1714 			 * choose a larger maxsockbuf should watch out
1715 			 * for the compatibility problems mentioned above.
1716 			 *
1717 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1718 			 * or <SYN,ACK>) segment itself is never scaled.
1719 			 */
1720 			while (wscale < TCP_MAX_WINSHIFT &&
1721 			    (TCP_MAXWIN << wscale) < sb_max)
1722 				wscale++;
1723 			sc->sc_requested_r_scale = wscale;
1724 			sc->sc_requested_s_scale = to->to_wscale;
1725 			sc->sc_flags |= SCF_WINSCALE;
1726 		}
1727 	}
1728 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1729 	/*
1730 	 * If incoming packet has an MD5 signature, flag this in the
1731 	 * syncache so that syncache_respond() will do the right thing
1732 	 * with the SYN+ACK.
1733 	 */
1734 	if (to->to_flags & TOF_SIGNATURE)
1735 		sc->sc_flags |= SCF_SIGNATURE;
1736 #endif	/* TCP_SIGNATURE */
1737 	if (to->to_flags & TOF_SACKPERM)
1738 		sc->sc_flags |= SCF_SACK;
1739 	if (to->to_flags & TOF_MSS)
1740 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1741 	if (ltflags & TF_NOOPT)
1742 		sc->sc_flags |= SCF_NOOPT;
1743 	/* ECN Handshake */
1744 	if (V_tcp_do_ecn && (tp->t_flags2 & TF2_CANNOT_DO_ECN) == 0)
1745 		sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1746 
1747 	if (V_tcp_syncookies)
1748 		sc->sc_iss = syncookie_generate(sch, sc);
1749 	else
1750 		sc->sc_iss = arc4random();
1751 #ifdef INET6
1752 	if (autoflowlabel) {
1753 		if (V_tcp_syncookies)
1754 			sc->sc_flowlabel = sc->sc_iss;
1755 		else
1756 			sc->sc_flowlabel = ip6_randomflowlabel();
1757 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1758 	}
1759 #endif
1760 	if (locked)
1761 		SCH_UNLOCK(sch);
1762 
1763 	if (tfo_cookie_valid) {
1764 		rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie);
1765 		/* INP_RUNLOCK(inp) will be performed by the caller */
1766 		goto tfo_expanded;
1767 	}
1768 
1769 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1770 	/*
1771 	 * Do a standard 3-way handshake.
1772 	 */
1773 	if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1774 		if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1775 			syncache_free(sc);
1776 		else if (sc != &scs)
1777 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1778 		TCPSTAT_INC(tcps_sndacks);
1779 		TCPSTAT_INC(tcps_sndtotal);
1780 	} else {
1781 		if (sc != &scs)
1782 			syncache_free(sc);
1783 		TCPSTAT_INC(tcps_sc_dropped);
1784 	}
1785 	goto donenoprobe;
1786 
1787 done:
1788 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1789 donenoprobe:
1790 	if (m)
1791 		m_freem(m);
1792 	/*
1793 	 * If tfo_pending is not NULL here, then a TFO SYN that did not
1794 	 * result in a new socket was processed and the associated pending
1795 	 * counter has not yet been decremented.  All such TFO processing paths
1796 	 * transit this point.
1797 	 */
1798 	if (tfo_pending != NULL)
1799 		tcp_fastopen_decrement_counter(tfo_pending);
1800 
1801 tfo_expanded:
1802 	if (cred != NULL)
1803 		crfree(cred);
1804 #ifdef MAC
1805 	if (sc == &scs)
1806 		mac_syncache_destroy(&maclabel);
1807 #endif
1808 	return (rv);
1809 }
1810 
1811 /*
1812  * Send SYN|ACK or ACK to the peer.  Either in response to a peer's segment,
1813  * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1814  */
1815 static int
1816 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1817 {
1818 	struct ip *ip = NULL;
1819 	struct mbuf *m;
1820 	struct tcphdr *th = NULL;
1821 	struct udphdr *udp = NULL;
1822 	int optlen, error = 0;	/* Make compiler happy */
1823 	u_int16_t hlen, tlen, mssopt, ulen;
1824 	struct tcpopt to;
1825 #ifdef INET6
1826 	struct ip6_hdr *ip6 = NULL;
1827 #endif
1828 
1829 	NET_EPOCH_ASSERT();
1830 
1831 	hlen =
1832 #ifdef INET6
1833 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1834 #endif
1835 		sizeof(struct ip);
1836 	tlen = hlen + sizeof(struct tcphdr);
1837 	if (sc->sc_port) {
1838 		tlen += sizeof(struct udphdr);
1839 	}
1840 	/* Determine MSS we advertize to other end of connection. */
1841 	mssopt = tcp_mssopt(&sc->sc_inc);
1842 	if (sc->sc_port)
1843 		mssopt -= V_tcp_udp_tunneling_overhead;
1844 	mssopt = max(mssopt, V_tcp_minmss);
1845 
1846 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1847 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1848 	    ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + "
1849 	    "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port,
1850 	    max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN));
1851 
1852 	/* Create the IP+TCP header from scratch. */
1853 	m = m_gethdr(M_NOWAIT, MT_DATA);
1854 	if (m == NULL)
1855 		return (ENOBUFS);
1856 #ifdef MAC
1857 	mac_syncache_create_mbuf(sc->sc_label, m);
1858 #endif
1859 	m->m_data += max_linkhdr;
1860 	m->m_len = tlen;
1861 	m->m_pkthdr.len = tlen;
1862 	m->m_pkthdr.rcvif = NULL;
1863 
1864 #ifdef INET6
1865 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1866 		ip6 = mtod(m, struct ip6_hdr *);
1867 		ip6->ip6_vfc = IPV6_VERSION;
1868 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1869 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1870 		ip6->ip6_plen = htons(tlen - hlen);
1871 		/* ip6_hlim is set after checksum */
1872 		/* Zero out traffic class and flow label. */
1873 		ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1874 		ip6->ip6_flow |= sc->sc_flowlabel;
1875 		if (sc->sc_port != 0) {
1876 			ip6->ip6_nxt = IPPROTO_UDP;
1877 			udp = (struct udphdr *)(ip6 + 1);
1878 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1879 			udp->uh_dport = sc->sc_port;
1880 			ulen = (tlen - sizeof(struct ip6_hdr));
1881 			th = (struct tcphdr *)(udp + 1);
1882 		} else {
1883 			ip6->ip6_nxt = IPPROTO_TCP;
1884 			th = (struct tcphdr *)(ip6 + 1);
1885 		}
1886 		ip6->ip6_flow |= htonl(sc->sc_ip_tos << IPV6_FLOWLABEL_LEN);
1887 	}
1888 #endif
1889 #if defined(INET6) && defined(INET)
1890 	else
1891 #endif
1892 #ifdef INET
1893 	{
1894 		ip = mtod(m, struct ip *);
1895 		ip->ip_v = IPVERSION;
1896 		ip->ip_hl = sizeof(struct ip) >> 2;
1897 		ip->ip_len = htons(tlen);
1898 		ip->ip_id = 0;
1899 		ip->ip_off = 0;
1900 		ip->ip_sum = 0;
1901 		ip->ip_src = sc->sc_inc.inc_laddr;
1902 		ip->ip_dst = sc->sc_inc.inc_faddr;
1903 		ip->ip_ttl = sc->sc_ip_ttl;
1904 		ip->ip_tos = sc->sc_ip_tos;
1905 
1906 		/*
1907 		 * See if we should do MTU discovery.  Route lookups are
1908 		 * expensive, so we will only unset the DF bit if:
1909 		 *
1910 		 *	1) path_mtu_discovery is disabled
1911 		 *	2) the SCF_UNREACH flag has been set
1912 		 */
1913 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1914 		       ip->ip_off |= htons(IP_DF);
1915 		if (sc->sc_port == 0) {
1916 			ip->ip_p = IPPROTO_TCP;
1917 			th = (struct tcphdr *)(ip + 1);
1918 		} else {
1919 			ip->ip_p = IPPROTO_UDP;
1920 			udp = (struct udphdr *)(ip + 1);
1921 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1922 			udp->uh_dport = sc->sc_port;
1923 			ulen = (tlen - sizeof(struct ip));
1924 			th = (struct tcphdr *)(udp + 1);
1925 		}
1926 	}
1927 #endif /* INET */
1928 	th->th_sport = sc->sc_inc.inc_lport;
1929 	th->th_dport = sc->sc_inc.inc_fport;
1930 
1931 	if (flags & TH_SYN)
1932 		th->th_seq = htonl(sc->sc_iss);
1933 	else
1934 		th->th_seq = htonl(sc->sc_iss + 1);
1935 	th->th_ack = htonl(sc->sc_irs + 1);
1936 	th->th_off = sizeof(struct tcphdr) >> 2;
1937 	th->th_win = htons(sc->sc_wnd);
1938 	th->th_urp = 0;
1939 
1940 	flags = tcp_ecn_syncache_respond(flags, sc);
1941 	tcp_set_flags(th, flags);
1942 
1943 	/* Tack on the TCP options. */
1944 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1945 		to.to_flags = 0;
1946 
1947 		if (flags & TH_SYN) {
1948 			to.to_mss = mssopt;
1949 			to.to_flags = TOF_MSS;
1950 			if (sc->sc_flags & SCF_WINSCALE) {
1951 				to.to_wscale = sc->sc_requested_r_scale;
1952 				to.to_flags |= TOF_SCALE;
1953 			}
1954 			if (sc->sc_flags & SCF_SACK)
1955 				to.to_flags |= TOF_SACKPERM;
1956 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1957 			if (sc->sc_flags & SCF_SIGNATURE)
1958 				to.to_flags |= TOF_SIGNATURE;
1959 #endif
1960 			if (sc->sc_tfo_cookie) {
1961 				to.to_flags |= TOF_FASTOPEN;
1962 				to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1963 				to.to_tfo_cookie = sc->sc_tfo_cookie;
1964 				/* don't send cookie again when retransmitting response */
1965 				sc->sc_tfo_cookie = NULL;
1966 			}
1967 		}
1968 		if (sc->sc_flags & SCF_TIMESTAMP) {
1969 			to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1970 			to.to_tsecr = sc->sc_tsreflect;
1971 			to.to_flags |= TOF_TS;
1972 		}
1973 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1974 
1975 		/* Adjust headers by option size. */
1976 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1977 		m->m_len += optlen;
1978 		m->m_pkthdr.len += optlen;
1979 #ifdef INET6
1980 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1981 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1982 		else
1983 #endif
1984 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1985 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1986 		if (sc->sc_flags & SCF_SIGNATURE) {
1987 			KASSERT(to.to_flags & TOF_SIGNATURE,
1988 			    ("tcp_addoptions() didn't set tcp_signature"));
1989 
1990 			/* NOTE: to.to_signature is inside of mbuf */
1991 			if (!TCPMD5_ENABLED() ||
1992 			    TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1993 				m_freem(m);
1994 				return (EACCES);
1995 			}
1996 		}
1997 #endif
1998 	} else
1999 		optlen = 0;
2000 
2001 	if (udp) {
2002 		ulen += optlen;
2003 		udp->uh_ulen = htons(ulen);
2004 	}
2005 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
2006 	/*
2007 	 * If we have peer's SYN and it has a flowid, then let's assign it to
2008 	 * our SYN|ACK.  ip6_output() and ip_output() will not assign flowid
2009 	 * to SYN|ACK due to lack of inp here.
2010 	 */
2011 	if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
2012 		m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
2013 		M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
2014 	}
2015 #ifdef INET6
2016 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
2017 		if (sc->sc_port) {
2018 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
2019 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2020 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen,
2021 			      IPPROTO_UDP, 0);
2022 			th->th_sum = htons(0);
2023 		} else {
2024 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
2025 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2026 			th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
2027 			    IPPROTO_TCP, 0);
2028 		}
2029 		ip6->ip6_hlim = sc->sc_ip_ttl;
2030 #ifdef TCP_OFFLOAD
2031 		if (ADDED_BY_TOE(sc)) {
2032 			struct toedev *tod = sc->sc_tod;
2033 
2034 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2035 
2036 			return (error);
2037 		}
2038 #endif
2039 		TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
2040 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
2041 	}
2042 #endif
2043 #if defined(INET6) && defined(INET)
2044 	else
2045 #endif
2046 #ifdef INET
2047 	{
2048 		if (sc->sc_port) {
2049 			m->m_pkthdr.csum_flags = CSUM_UDP;
2050 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2051 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
2052 			      ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
2053 			th->th_sum = htons(0);
2054 		} else {
2055 			m->m_pkthdr.csum_flags = CSUM_TCP;
2056 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2057 			th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
2058 			    htons(tlen + optlen - hlen + IPPROTO_TCP));
2059 		}
2060 #ifdef TCP_OFFLOAD
2061 		if (ADDED_BY_TOE(sc)) {
2062 			struct toedev *tod = sc->sc_tod;
2063 
2064 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2065 
2066 			return (error);
2067 		}
2068 #endif
2069 		TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
2070 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
2071 	}
2072 #endif
2073 	return (error);
2074 }
2075 
2076 /*
2077  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
2078  * that exceed the capacity of the syncache by avoiding the storage of any
2079  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
2080  * attacks where the attacker does not have access to our responses.
2081  *
2082  * Syncookies encode and include all necessary information about the
2083  * connection setup within the SYN|ACK that we send back.  That way we
2084  * can avoid keeping any local state until the ACK to our SYN|ACK returns
2085  * (if ever).  Normally the syncache and syncookies are running in parallel
2086  * with the latter taking over when the former is exhausted.  When matching
2087  * syncache entry is found the syncookie is ignored.
2088  *
2089  * The only reliable information persisting the 3WHS is our initial sequence
2090  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
2091  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
2092  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
2093  * returns and signifies a legitimate connection if it matches the ACK.
2094  *
2095  * The available space of 32 bits to store the hash and to encode the SYN
2096  * option information is very tight and we should have at least 24 bits for
2097  * the MAC to keep the number of guesses by blind spoofing reasonably high.
2098  *
2099  * SYN option information we have to encode to fully restore a connection:
2100  * MSS: is imporant to chose an optimal segment size to avoid IP level
2101  *   fragmentation along the path.  The common MSS values can be encoded
2102  *   in a 3-bit table.  Uncommon values are captured by the next lower value
2103  *   in the table leading to a slight increase in packetization overhead.
2104  * WSCALE: is necessary to allow large windows to be used for high delay-
2105  *   bandwidth product links.  Not scaling the window when it was initially
2106  *   negotiated is bad for performance as lack of scaling further decreases
2107  *   the apparent available send window.  We only need to encode the WSCALE
2108  *   we received from the remote end.  Our end can be recalculated at any
2109  *   time.  The common WSCALE values can be encoded in a 3-bit table.
2110  *   Uncommon values are captured by the next lower value in the table
2111  *   making us under-estimate the available window size halving our
2112  *   theoretically possible maximum throughput for that connection.
2113  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2114  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2115  *   that are included in all segments on a connection.  We enable them when
2116  *   the ACK has them.
2117  *
2118  * Security of syncookies and attack vectors:
2119  *
2120  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2121  * together with the gloabl secret to make it unique per connection attempt.
2122  * Thus any change of any of those parameters results in a different MAC output
2123  * in an unpredictable way unless a collision is encountered.  24 bits of the
2124  * MAC are embedded into the ISS.
2125  *
2126  * To prevent replay attacks two rotating global secrets are updated with a
2127  * new random value every 15 seconds.  The life-time of a syncookie is thus
2128  * 15-30 seconds.
2129  *
2130  * Vector 1: Attacking the secret.  This requires finding a weakness in the
2131  * MAC itself or the way it is used here.  The attacker can do a chosen plain
2132  * text attack by varying and testing the all parameters under his control.
2133  * The strength depends on the size and randomness of the secret, and the
2134  * cryptographic security of the MAC function.  Due to the constant updating
2135  * of the secret the attacker has at most 29.999 seconds to find the secret
2136  * and launch spoofed connections.  After that he has to start all over again.
2137  *
2138  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
2139  * size an average of 4,823 attempts are required for a 50% chance of success
2140  * to spoof a single syncookie (birthday collision paradox).  However the
2141  * attacker is blind and doesn't know if one of his attempts succeeded unless
2142  * he has a side channel to interfere success from.  A single connection setup
2143  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2144  * This many attempts are required for each one blind spoofed connection.  For
2145  * every additional spoofed connection he has to launch another N attempts.
2146  * Thus for a sustained rate 100 spoofed connections per second approximately
2147  * 1,800,000 packets per second would have to be sent.
2148  *
2149  * NB: The MAC function should be fast so that it doesn't become a CPU
2150  * exhaustion attack vector itself.
2151  *
2152  * References:
2153  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2154  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2155  *   http://cr.yp.to/syncookies.html    (overview)
2156  *   http://cr.yp.to/syncookies/archive (details)
2157  *
2158  *
2159  * Schematic construction of a syncookie enabled Initial Sequence Number:
2160  *  0        1         2         3
2161  *  12345678901234567890123456789012
2162  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2163  *
2164  *  x 24 MAC (truncated)
2165  *  W  3 Send Window Scale index
2166  *  M  3 MSS index
2167  *  S  1 SACK permitted
2168  *  P  1 Odd/even secret
2169  */
2170 
2171 /*
2172  * Distribution and probability of certain MSS values.  Those in between are
2173  * rounded down to the next lower one.
2174  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2175  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
2176  */
2177 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2178 
2179 /*
2180  * Distribution and probability of certain WSCALE values.  We have to map the
2181  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2182  * bits based on prevalence of certain values.  Where we don't have an exact
2183  * match for are rounded down to the next lower one letting us under-estimate
2184  * the true available window.  At the moment this would happen only for the
2185  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2186  * and window size).  The absence of the WSCALE option (no scaling in either
2187  * direction) is encoded with index zero.
2188  * [WSCALE values histograms, Allman, 2012]
2189  *                            X 10 10 35  5  6 14 10%   by host
2190  *                            X 11  4  5  5 18 49  3%   by connections
2191  */
2192 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2193 
2194 /*
2195  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
2196  * and good cryptographic properties.
2197  */
2198 static uint32_t
2199 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2200     uint8_t *secbits, uintptr_t secmod)
2201 {
2202 	SIPHASH_CTX ctx;
2203 	uint32_t siphash[2];
2204 
2205 	SipHash24_Init(&ctx);
2206 	SipHash_SetKey(&ctx, secbits);
2207 	switch (inc->inc_flags & INC_ISIPV6) {
2208 #ifdef INET
2209 	case 0:
2210 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2211 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2212 		break;
2213 #endif
2214 #ifdef INET6
2215 	case INC_ISIPV6:
2216 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2217 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2218 		break;
2219 #endif
2220 	}
2221 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2222 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2223 	SipHash_Update(&ctx, &irs, sizeof(irs));
2224 	SipHash_Update(&ctx, &flags, sizeof(flags));
2225 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
2226 	SipHash_Final((u_int8_t *)&siphash, &ctx);
2227 
2228 	return (siphash[0] ^ siphash[1]);
2229 }
2230 
2231 static tcp_seq
2232 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2233 {
2234 	u_int i, secbit, wscale;
2235 	uint32_t iss, hash;
2236 	uint8_t *secbits;
2237 	union syncookie cookie;
2238 
2239 	cookie.cookie = 0;
2240 
2241 	/* Map our computed MSS into the 3-bit index. */
2242 	for (i = nitems(tcp_sc_msstab) - 1;
2243 	     tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2244 	     i--)
2245 		;
2246 	cookie.flags.mss_idx = i;
2247 
2248 	/*
2249 	 * Map the send window scale into the 3-bit index but only if
2250 	 * the wscale option was received.
2251 	 */
2252 	if (sc->sc_flags & SCF_WINSCALE) {
2253 		wscale = sc->sc_requested_s_scale;
2254 		for (i = nitems(tcp_sc_wstab) - 1;
2255 		    tcp_sc_wstab[i] > wscale && i > 0;
2256 		     i--)
2257 			;
2258 		cookie.flags.wscale_idx = i;
2259 	}
2260 
2261 	/* Can we do SACK? */
2262 	if (sc->sc_flags & SCF_SACK)
2263 		cookie.flags.sack_ok = 1;
2264 
2265 	/* Which of the two secrets to use. */
2266 	secbit = V_tcp_syncache.secret.oddeven & 0x1;
2267 	cookie.flags.odd_even = secbit;
2268 
2269 	secbits = V_tcp_syncache.secret.key[secbit];
2270 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2271 	    (uintptr_t)sch);
2272 
2273 	/*
2274 	 * Put the flags into the hash and XOR them to get better ISS number
2275 	 * variance.  This doesn't enhance the cryptographic strength and is
2276 	 * done to prevent the 8 cookie bits from showing up directly on the
2277 	 * wire.
2278 	 */
2279 	iss = hash & ~0xff;
2280 	iss |= cookie.cookie ^ (hash >> 24);
2281 
2282 	TCPSTAT_INC(tcps_sc_sendcookie);
2283 	return (iss);
2284 }
2285 
2286 static struct syncache *
2287 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2288     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2289     struct socket *lso, uint16_t port)
2290 {
2291 	uint32_t hash;
2292 	uint8_t *secbits;
2293 	tcp_seq ack, seq;
2294 	int wnd, wscale = 0;
2295 	union syncookie cookie;
2296 
2297 	/*
2298 	 * Pull information out of SYN-ACK/ACK and revert sequence number
2299 	 * advances.
2300 	 */
2301 	ack = th->th_ack - 1;
2302 	seq = th->th_seq - 1;
2303 
2304 	/*
2305 	 * Unpack the flags containing enough information to restore the
2306 	 * connection.
2307 	 */
2308 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2309 
2310 	/* Which of the two secrets to use. */
2311 	secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2312 
2313 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2314 
2315 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
2316 	if ((ack & ~0xff) != (hash & ~0xff))
2317 		return (NULL);
2318 
2319 	/* Fill in the syncache values. */
2320 	sc->sc_flags = 0;
2321 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2322 	sc->sc_ipopts = NULL;
2323 
2324 	sc->sc_irs = seq;
2325 	sc->sc_iss = ack;
2326 
2327 	switch (inc->inc_flags & INC_ISIPV6) {
2328 #ifdef INET
2329 	case 0:
2330 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2331 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2332 		break;
2333 #endif
2334 #ifdef INET6
2335 	case INC_ISIPV6:
2336 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2337 			sc->sc_flowlabel =
2338 			    htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2339 		break;
2340 #endif
2341 	}
2342 
2343 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2344 
2345 	/* We can simply recompute receive window scale we sent earlier. */
2346 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2347 		wscale++;
2348 
2349 	/* Only use wscale if it was enabled in the orignal SYN. */
2350 	if (cookie.flags.wscale_idx > 0) {
2351 		sc->sc_requested_r_scale = wscale;
2352 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2353 		sc->sc_flags |= SCF_WINSCALE;
2354 	}
2355 
2356 	wnd = lso->sol_sbrcv_hiwat;
2357 	wnd = imax(wnd, 0);
2358 	wnd = imin(wnd, TCP_MAXWIN);
2359 	sc->sc_wnd = wnd;
2360 
2361 	if (cookie.flags.sack_ok)
2362 		sc->sc_flags |= SCF_SACK;
2363 
2364 	if (to->to_flags & TOF_TS) {
2365 		sc->sc_flags |= SCF_TIMESTAMP;
2366 		sc->sc_tsreflect = to->to_tsval;
2367 		sc->sc_tsoff = tcp_new_ts_offset(inc);
2368 	}
2369 
2370 	if (to->to_flags & TOF_SIGNATURE)
2371 		sc->sc_flags |= SCF_SIGNATURE;
2372 
2373 	sc->sc_rxmits = 0;
2374 
2375 	sc->sc_port = port;
2376 
2377 	TCPSTAT_INC(tcps_sc_recvcookie);
2378 	return (sc);
2379 }
2380 
2381 #ifdef INVARIANTS
2382 static int
2383 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2384     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2385     struct socket *lso, uint16_t port)
2386 {
2387 	struct syncache scs, *scx;
2388 	char *s;
2389 
2390 	bzero(&scs, sizeof(scs));
2391 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port);
2392 
2393 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2394 		return (0);
2395 
2396 	if (scx != NULL) {
2397 		if (sc->sc_peer_mss != scx->sc_peer_mss)
2398 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2399 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2400 
2401 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2402 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2403 			    s, __func__, sc->sc_requested_r_scale,
2404 			    scx->sc_requested_r_scale);
2405 
2406 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2407 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2408 			    s, __func__, sc->sc_requested_s_scale,
2409 			    scx->sc_requested_s_scale);
2410 
2411 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2412 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2413 	}
2414 
2415 	if (s != NULL)
2416 		free(s, M_TCPLOG);
2417 	return (0);
2418 }
2419 #endif /* INVARIANTS */
2420 
2421 static void
2422 syncookie_reseed(void *arg)
2423 {
2424 	struct tcp_syncache *sc = arg;
2425 	uint8_t *secbits;
2426 	int secbit;
2427 
2428 	/*
2429 	 * Reseeding the secret doesn't have to be protected by a lock.
2430 	 * It only must be ensured that the new random values are visible
2431 	 * to all CPUs in a SMP environment.  The atomic with release
2432 	 * semantics ensures that.
2433 	 */
2434 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2435 	secbits = sc->secret.key[secbit];
2436 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2437 	atomic_add_rel_int(&sc->secret.oddeven, 1);
2438 
2439 	/* Reschedule ourself. */
2440 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2441 }
2442 
2443 /*
2444  * We have overflowed a bucket. Let's pause dealing with the syncache.
2445  * This function will increment the bucketoverflow statistics appropriately
2446  * (once per pause when pausing is enabled; otherwise, once per overflow).
2447  */
2448 static void
2449 syncache_pause(struct in_conninfo *inc)
2450 {
2451 	time_t delta;
2452 	const char *s;
2453 
2454 	/* XXX:
2455 	 * 2. Add sysctl read here so we don't get the benefit of this
2456 	 * change without the new sysctl.
2457 	 */
2458 
2459 	/*
2460 	 * Try an unlocked read. If we already know that another thread
2461 	 * has activated the feature, there is no need to proceed.
2462 	 */
2463 	if (V_tcp_syncache.paused)
2464 		return;
2465 
2466 	/* Are cookied enabled? If not, we can't pause. */
2467 	if (!V_tcp_syncookies) {
2468 		TCPSTAT_INC(tcps_sc_bucketoverflow);
2469 		return;
2470 	}
2471 
2472 	/*
2473 	 * We may be the first thread to find an overflow. Get the lock
2474 	 * and evaluate if we need to take action.
2475 	 */
2476 	mtx_lock(&V_tcp_syncache.pause_mtx);
2477 	if (V_tcp_syncache.paused) {
2478 		mtx_unlock(&V_tcp_syncache.pause_mtx);
2479 		return;
2480 	}
2481 
2482 	/* Activate protection. */
2483 	V_tcp_syncache.paused = true;
2484 	TCPSTAT_INC(tcps_sc_bucketoverflow);
2485 
2486 	/*
2487 	 * Determine the last backoff time. If we are seeing a re-newed
2488 	 * attack within that same time after last reactivating the syncache,
2489 	 * consider it an extension of the same attack.
2490 	 */
2491 	delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2492 	if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2493 		if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2494 			delta <<= 1;
2495 			V_tcp_syncache.pause_backoff++;
2496 		}
2497 	} else {
2498 		delta = TCP_SYNCACHE_PAUSE_TIME;
2499 		V_tcp_syncache.pause_backoff = 0;
2500 	}
2501 
2502 	/* Log a warning, including IP addresses, if able. */
2503 	if (inc != NULL)
2504 		s = tcp_log_addrs(inc, NULL, NULL, NULL);
2505 	else
2506 		s = (const char *)NULL;
2507 	log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2508 	    "the next %lld seconds%s%s%s\n", (long long)delta,
2509 	    (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2510 	    (s != NULL) ? ")" : "");
2511 	free(__DECONST(void *, s), M_TCPLOG);
2512 
2513 	/* Use the calculated delta to set a new pause time. */
2514 	V_tcp_syncache.pause_until = time_uptime + delta;
2515 	callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2516 	    &V_tcp_syncache);
2517 	mtx_unlock(&V_tcp_syncache.pause_mtx);
2518 }
2519 
2520 /* Evaluate whether we need to unpause. */
2521 static void
2522 syncache_unpause(void *arg)
2523 {
2524 	struct tcp_syncache *sc;
2525 	time_t delta;
2526 
2527 	sc = arg;
2528 	mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2529 	callout_deactivate(&sc->pause_co);
2530 
2531 	/*
2532 	 * Check to make sure we are not running early. If the pause
2533 	 * time has expired, then deactivate the protection.
2534 	 */
2535 	if ((delta = sc->pause_until - time_uptime) > 0)
2536 		callout_schedule(&sc->pause_co, delta * hz);
2537 	else
2538 		sc->paused = false;
2539 }
2540 
2541 /*
2542  * Exports the syncache entries to userland so that netstat can display
2543  * them alongside the other sockets.  This function is intended to be
2544  * called only from tcp_pcblist.
2545  *
2546  * Due to concurrency on an active system, the number of pcbs exported
2547  * may have no relation to max_pcbs.  max_pcbs merely indicates the
2548  * amount of space the caller allocated for this function to use.
2549  */
2550 int
2551 syncache_pcblist(struct sysctl_req *req)
2552 {
2553 	struct xtcpcb xt;
2554 	struct syncache *sc;
2555 	struct syncache_head *sch;
2556 	int error, i;
2557 
2558 	bzero(&xt, sizeof(xt));
2559 	xt.xt_len = sizeof(xt);
2560 	xt.t_state = TCPS_SYN_RECEIVED;
2561 	xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2562 	xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2563 	xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2564 	xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2565 
2566 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2567 		sch = &V_tcp_syncache.hashbase[i];
2568 		SCH_LOCK(sch);
2569 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2570 			if (sc->sc_cred != NULL &&
2571 			    cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2572 				continue;
2573 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2574 				xt.xt_inp.inp_vflag = INP_IPV6;
2575 			else
2576 				xt.xt_inp.inp_vflag = INP_IPV4;
2577 			xt.xt_encaps_port = sc->sc_port;
2578 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2579 			    sizeof (struct in_conninfo));
2580 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2581 			if (error) {
2582 				SCH_UNLOCK(sch);
2583 				return (0);
2584 			}
2585 		}
2586 		SCH_UNLOCK(sch);
2587 	}
2588 
2589 	return (0);
2590 }
2591