xref: /freebsd/sys/netinet/tcp_syncache.c (revision 4543ef516683042d46f3bd3bb8a4f3f746e00499)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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 #include "opt_inet.h"
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/hash.h>
43 #include <sys/refcount.h>
44 #include <sys/kernel.h>
45 #include <sys/sysctl.h>
46 #include <sys/limits.h>
47 #include <sys/lock.h>
48 #include <sys/mutex.h>
49 #include <sys/malloc.h>
50 #include <sys/mbuf.h>
51 #include <sys/proc.h>		/* for proc0 declaration */
52 #include <sys/random.h>
53 #include <sys/socket.h>
54 #include <sys/socketvar.h>
55 #include <sys/syslog.h>
56 #include <sys/ucred.h>
57 
58 #include <sys/md5.h>
59 #include <crypto/siphash/siphash.h>
60 
61 #include <vm/uma.h>
62 
63 #include <net/if.h>
64 #include <net/if_var.h>
65 #include <net/route.h>
66 #include <net/vnet.h>
67 
68 #include <netinet/in.h>
69 #include <netinet/in_kdtrace.h>
70 #include <netinet/in_systm.h>
71 #include <netinet/ip.h>
72 #include <netinet/in_var.h>
73 #include <netinet/in_pcb.h>
74 #include <netinet/ip_var.h>
75 #include <netinet/ip_options.h>
76 #ifdef INET6
77 #include <netinet/ip6.h>
78 #include <netinet/icmp6.h>
79 #include <netinet6/nd6.h>
80 #include <netinet6/ip6_var.h>
81 #include <netinet6/in6_pcb.h>
82 #endif
83 #include <netinet/tcp.h>
84 #include <netinet/tcp_fastopen.h>
85 #include <netinet/tcp_fsm.h>
86 #include <netinet/tcp_seq.h>
87 #include <netinet/tcp_timer.h>
88 #include <netinet/tcp_var.h>
89 #include <netinet/tcp_syncache.h>
90 #include <netinet/tcp_ecn.h>
91 #ifdef TCP_BLACKBOX
92 #include <netinet/tcp_log_buf.h>
93 #endif
94 #ifdef TCP_OFFLOAD
95 #include <netinet/toecore.h>
96 #endif
97 #include <netinet/udp.h>
98 
99 #include <netipsec/ipsec_support.h>
100 
101 #include <machine/in_cksum.h>
102 
103 #include <security/mac/mac_framework.h>
104 
105 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1;
106 #define	V_tcp_syncookies		VNET(tcp_syncookies)
107 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
108     &VNET_NAME(tcp_syncookies), 0,
109     "Use TCP SYN cookies if the syncache overflows");
110 
111 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0;
112 #define	V_tcp_syncookiesonly		VNET(tcp_syncookiesonly)
113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
114     &VNET_NAME(tcp_syncookiesonly), 0,
115     "Use only TCP SYN cookies");
116 
117 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1;
118 #define V_functions_inherit_listen_socket_stack \
119     VNET(functions_inherit_listen_socket_stack)
120 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack,
121     CTLFLAG_VNET | CTLFLAG_RW,
122     &VNET_NAME(functions_inherit_listen_socket_stack), 0,
123     "Inherit listen socket's stack");
124 
125 #ifdef TCP_OFFLOAD
126 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
127 #endif
128 
129 static void	 syncache_drop(struct syncache *, struct syncache_head *);
130 static void	 syncache_free(struct syncache *);
131 static void	 syncache_insert(struct syncache *, struct syncache_head *);
132 static int	 syncache_respond(struct syncache *, const struct mbuf *, int);
133 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
134 		    struct mbuf *m);
135 static void	 syncache_timeout(struct syncache *sc, struct syncache_head *sch,
136 		    int docallout);
137 static void	 syncache_timer(void *);
138 
139 static uint32_t	 syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
140 		    uint8_t *, uintptr_t);
141 static tcp_seq	 syncookie_generate(struct syncache_head *, struct syncache *);
142 static struct syncache
143 		*syncookie_lookup(struct in_conninfo *, struct syncache_head *,
144 		    struct syncache *, struct tcphdr *, struct tcpopt *,
145 		    struct socket *, uint16_t);
146 static void	syncache_pause(struct in_conninfo *);
147 static void	syncache_unpause(void *);
148 static void	 syncookie_reseed(void *);
149 #ifdef INVARIANTS
150 static int	 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
151 		    struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
152 		    struct socket *lso, uint16_t port);
153 #endif
154 
155 /*
156  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
157  * 3 retransmits corresponds to a timeout with default values of
158  * tcp_rexmit_initial * (             1 +
159  *                       tcp_backoff[1] +
160  *                       tcp_backoff[2] +
161  *                       tcp_backoff[3]) + 3 * tcp_rexmit_slop,
162  * 1000 ms * (1 + 2 + 4 + 8) +  3 * 200 ms = 15600 ms,
163  * the odds are that the user has given up attempting to connect by then.
164  */
165 #define SYNCACHE_MAXREXMTS		3
166 
167 /* Arbitrary values */
168 #define TCP_SYNCACHE_HASHSIZE		512
169 #define TCP_SYNCACHE_BUCKETLIMIT	30
170 
171 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
172 #define	V_tcp_syncache			VNET(tcp_syncache)
173 
174 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache,
175     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
176     "TCP SYN cache");
177 
178 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
179     &VNET_NAME(tcp_syncache.bucket_limit), 0,
180     "Per-bucket hash limit for syncache");
181 
182 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
183     &VNET_NAME(tcp_syncache.cache_limit), 0,
184     "Overall entry limit for syncache");
185 
186 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
187     &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
188 
189 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
190     &VNET_NAME(tcp_syncache.hashsize), 0,
191     "Size of TCP syncache hashtable");
192 
193 SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET |
194     CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0,
195     "All syncache(4) entries are visible, ignoring UID/GID, jail(2) "
196     "and mac(4) checks");
197 
198 static int
199 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
200 {
201 	int error;
202 	u_int new;
203 
204 	new = V_tcp_syncache.rexmt_limit;
205 	error = sysctl_handle_int(oidp, &new, 0, req);
206 	if ((error == 0) && (req->newptr != NULL)) {
207 		if (new > TCP_MAXRXTSHIFT)
208 			error = EINVAL;
209 		else
210 			V_tcp_syncache.rexmt_limit = new;
211 	}
212 	return (error);
213 }
214 
215 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
216     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
217     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
218     sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
219     "Limit on SYN/ACK retransmissions");
220 
221 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
222 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
223     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
224     "Send reset on socket allocation failure");
225 
226 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
227 
228 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
229 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
230 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
231 
232 /*
233  * Requires the syncache entry to be already removed from the bucket list.
234  */
235 static void
236 syncache_free(struct syncache *sc)
237 {
238 
239 	if (sc->sc_ipopts)
240 		(void) m_free(sc->sc_ipopts);
241 	if (sc->sc_cred)
242 		crfree(sc->sc_cred);
243 #ifdef MAC
244 	mac_syncache_destroy(&sc->sc_label);
245 #endif
246 
247 	uma_zfree(V_tcp_syncache.zone, sc);
248 }
249 
250 void
251 syncache_init(void)
252 {
253 	int i;
254 
255 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
256 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
257 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
258 	V_tcp_syncache.hash_secret = arc4random();
259 
260 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
261 	    &V_tcp_syncache.hashsize);
262 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
263 	    &V_tcp_syncache.bucket_limit);
264 	if (!powerof2(V_tcp_syncache.hashsize) ||
265 	    V_tcp_syncache.hashsize == 0) {
266 		printf("WARNING: syncache hash size is not a power of 2.\n");
267 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
268 	}
269 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
270 
271 	/* Set limits. */
272 	V_tcp_syncache.cache_limit =
273 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
274 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
275 	    &V_tcp_syncache.cache_limit);
276 
277 	/* Allocate the hash table. */
278 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
279 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
280 
281 #ifdef VIMAGE
282 	V_tcp_syncache.vnet = curvnet;
283 #endif
284 
285 	/* Initialize the hash buckets. */
286 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
287 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
288 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
289 			 NULL, MTX_DEF);
290 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
291 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
292 		V_tcp_syncache.hashbase[i].sch_length = 0;
293 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
294 		V_tcp_syncache.hashbase[i].sch_last_overflow =
295 		    -(SYNCOOKIE_LIFETIME + 1);
296 	}
297 
298 	/* Create the syncache entry zone. */
299 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
300 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
301 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
302 	    V_tcp_syncache.cache_limit);
303 
304 	/* Start the SYN cookie reseeder callout. */
305 	callout_init(&V_tcp_syncache.secret.reseed, 1);
306 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
307 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
308 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
309 	    syncookie_reseed, &V_tcp_syncache);
310 
311 	/* Initialize the pause machinery. */
312 	mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF);
313 	callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx,
314 	    0);
315 	V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME;
316 	V_tcp_syncache.pause_backoff = 0;
317 	V_tcp_syncache.paused = false;
318 }
319 
320 #ifdef VIMAGE
321 void
322 syncache_destroy(void)
323 {
324 	struct syncache_head *sch;
325 	struct syncache *sc, *nsc;
326 	int i;
327 
328 	/*
329 	 * Stop the re-seed timer before freeing resources.  No need to
330 	 * possibly schedule it another time.
331 	 */
332 	callout_drain(&V_tcp_syncache.secret.reseed);
333 
334 	/* Stop the SYN cache pause callout. */
335 	mtx_lock(&V_tcp_syncache.pause_mtx);
336 	if (callout_stop(&V_tcp_syncache.pause_co) == 0) {
337 		mtx_unlock(&V_tcp_syncache.pause_mtx);
338 		callout_drain(&V_tcp_syncache.pause_co);
339 	} else
340 		mtx_unlock(&V_tcp_syncache.pause_mtx);
341 
342 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
343 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
344 		sch = &V_tcp_syncache.hashbase[i];
345 		callout_drain(&sch->sch_timer);
346 
347 		SCH_LOCK(sch);
348 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
349 			syncache_drop(sc, sch);
350 		SCH_UNLOCK(sch);
351 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
352 		    ("%s: sch->sch_bucket not empty", __func__));
353 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
354 		    __func__, sch->sch_length));
355 		mtx_destroy(&sch->sch_mtx);
356 	}
357 
358 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
359 	    ("%s: cache_count not 0", __func__));
360 
361 	/* Free the allocated global resources. */
362 	uma_zdestroy(V_tcp_syncache.zone);
363 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
364 	mtx_destroy(&V_tcp_syncache.pause_mtx);
365 }
366 #endif
367 
368 /*
369  * Inserts a syncache entry into the specified bucket row.
370  * Locks and unlocks the syncache_head autonomously.
371  */
372 static void
373 syncache_insert(struct syncache *sc, struct syncache_head *sch)
374 {
375 	struct syncache *sc2;
376 
377 	SCH_LOCK(sch);
378 
379 	/*
380 	 * Make sure that we don't overflow the per-bucket limit.
381 	 * If the bucket is full, toss the oldest element.
382 	 */
383 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
384 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
385 			("sch->sch_length incorrect"));
386 		syncache_pause(&sc->sc_inc);
387 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
388 		sch->sch_last_overflow = time_uptime;
389 		syncache_drop(sc2, sch);
390 	}
391 
392 	/* Put it into the bucket. */
393 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
394 	sch->sch_length++;
395 
396 #ifdef TCP_OFFLOAD
397 	if (ADDED_BY_TOE(sc)) {
398 		struct toedev *tod = sc->sc_tod;
399 
400 		tod->tod_syncache_added(tod, sc->sc_todctx);
401 	}
402 #endif
403 
404 	/* Reinitialize the bucket row's timer. */
405 	if (sch->sch_length == 1)
406 		sch->sch_nextc = ticks + INT_MAX;
407 	syncache_timeout(sc, sch, 1);
408 
409 	SCH_UNLOCK(sch);
410 
411 	TCPSTATES_INC(TCPS_SYN_RECEIVED);
412 	TCPSTAT_INC(tcps_sc_added);
413 }
414 
415 /*
416  * Remove and free entry from syncache bucket row.
417  * Expects locked syncache head.
418  */
419 static void
420 syncache_drop(struct syncache *sc, struct syncache_head *sch)
421 {
422 
423 	SCH_LOCK_ASSERT(sch);
424 
425 	TCPSTATES_DEC(TCPS_SYN_RECEIVED);
426 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
427 	sch->sch_length--;
428 
429 #ifdef TCP_OFFLOAD
430 	if (ADDED_BY_TOE(sc)) {
431 		struct toedev *tod = sc->sc_tod;
432 
433 		tod->tod_syncache_removed(tod, sc->sc_todctx);
434 	}
435 #endif
436 
437 	syncache_free(sc);
438 }
439 
440 /*
441  * Engage/reengage time on bucket row.
442  */
443 static void
444 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
445 {
446 	int rexmt;
447 
448 	if (sc->sc_rxmits == 0)
449 		rexmt = tcp_rexmit_initial;
450 	else
451 		TCPT_RANGESET(rexmt,
452 		    tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
453 		    tcp_rexmit_min, TCPTV_REXMTMAX);
454 	sc->sc_rxttime = ticks + rexmt;
455 	sc->sc_rxmits++;
456 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
457 		sch->sch_nextc = sc->sc_rxttime;
458 		if (docallout)
459 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
460 			    syncache_timer, (void *)sch);
461 	}
462 }
463 
464 /*
465  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
466  * If we have retransmitted an entry the maximum number of times, expire it.
467  * One separate timer for each bucket row.
468  */
469 static void
470 syncache_timer(void *xsch)
471 {
472 	struct syncache_head *sch = (struct syncache_head *)xsch;
473 	struct syncache *sc, *nsc;
474 	struct epoch_tracker et;
475 	int tick = ticks;
476 	char *s;
477 	bool paused;
478 
479 	CURVNET_SET(sch->sch_sc->vnet);
480 
481 	/* NB: syncache_head has already been locked by the callout. */
482 	SCH_LOCK_ASSERT(sch);
483 
484 	/*
485 	 * In the following cycle we may remove some entries and/or
486 	 * advance some timeouts, so re-initialize the bucket timer.
487 	 */
488 	sch->sch_nextc = tick + INT_MAX;
489 
490 	/*
491 	 * If we have paused processing, unconditionally remove
492 	 * all syncache entries.
493 	 */
494 	mtx_lock(&V_tcp_syncache.pause_mtx);
495 	paused = V_tcp_syncache.paused;
496 	mtx_unlock(&V_tcp_syncache.pause_mtx);
497 
498 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
499 		if (paused) {
500 			syncache_drop(sc, sch);
501 			continue;
502 		}
503 		/*
504 		 * We do not check if the listen socket still exists
505 		 * and accept the case where the listen socket may be
506 		 * gone by the time we resend the SYN/ACK.  We do
507 		 * not expect this to happens often. If it does,
508 		 * then the RST will be sent by the time the remote
509 		 * host does the SYN/ACK->ACK.
510 		 */
511 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
512 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
513 				sch->sch_nextc = sc->sc_rxttime;
514 			continue;
515 		}
516 		if (sc->sc_rxmits > V_tcp_ecn_maxretries) {
517 			sc->sc_flags &= ~SCF_ECN_MASK;
518 		}
519 		if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
520 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
521 				log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
522 				    "giving up and removing syncache entry\n",
523 				    s, __func__);
524 				free(s, M_TCPLOG);
525 			}
526 			syncache_drop(sc, sch);
527 			TCPSTAT_INC(tcps_sc_stale);
528 			continue;
529 		}
530 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
531 			log(LOG_DEBUG, "%s; %s: Response timeout, "
532 			    "retransmitting (%u) SYN|ACK\n",
533 			    s, __func__, sc->sc_rxmits);
534 			free(s, M_TCPLOG);
535 		}
536 
537 		NET_EPOCH_ENTER(et);
538 		syncache_respond(sc, NULL, TH_SYN|TH_ACK);
539 		NET_EPOCH_EXIT(et);
540 		TCPSTAT_INC(tcps_sc_retransmitted);
541 		syncache_timeout(sc, sch, 0);
542 	}
543 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
544 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
545 			syncache_timer, (void *)(sch));
546 	CURVNET_RESTORE();
547 }
548 
549 /*
550  * Returns true if the system is only using cookies at the moment.
551  * This could be due to a sysadmin decision to only use cookies, or it
552  * could be due to the system detecting an attack.
553  */
554 static inline bool
555 syncache_cookiesonly(void)
556 {
557 
558 	return (V_tcp_syncookies && (V_tcp_syncache.paused ||
559 	    V_tcp_syncookiesonly));
560 }
561 
562 /*
563  * Find the hash bucket for the given connection.
564  */
565 static struct syncache_head *
566 syncache_hashbucket(struct in_conninfo *inc)
567 {
568 	uint32_t hash;
569 
570 	/*
571 	 * The hash is built on foreign port + local port + foreign address.
572 	 * We rely on the fact that struct in_conninfo starts with 16 bits
573 	 * of foreign port, then 16 bits of local port then followed by 128
574 	 * bits of foreign address.  In case of IPv4 address, the first 3
575 	 * 32-bit words of the address always are zeroes.
576 	 */
577 	hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
578 	    V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
579 
580 	return (&V_tcp_syncache.hashbase[hash]);
581 }
582 
583 /*
584  * Find an entry in the syncache.
585  * Returns always with locked syncache_head plus a matching entry or NULL.
586  */
587 static struct syncache *
588 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
589 {
590 	struct syncache *sc;
591 	struct syncache_head *sch;
592 
593 	*schp = sch = syncache_hashbucket(inc);
594 	SCH_LOCK(sch);
595 
596 	/* Circle through bucket row to find matching entry. */
597 	TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
598 		if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
599 		    sizeof(struct in_endpoints)) == 0)
600 			break;
601 
602 	return (sc);	/* Always returns with locked sch. */
603 }
604 
605 /*
606  * This function is called when we get a RST for a
607  * non-existent connection, so that we can see if the
608  * connection is in the syn cache.  If it is, zap it.
609  * If required send a challenge ACK.
610  */
611 void
612 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m,
613     uint16_t port)
614 {
615 	struct syncache *sc;
616 	struct syncache_head *sch;
617 	char *s = NULL;
618 
619 	if (syncache_cookiesonly())
620 		return;
621 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
622 	SCH_LOCK_ASSERT(sch);
623 
624 	/*
625 	 * No corresponding connection was found in syncache.
626 	 * If syncookies are enabled and possibly exclusively
627 	 * used, or we are under memory pressure, a valid RST
628 	 * may not find a syncache entry.  In that case we're
629 	 * done and no SYN|ACK retransmissions will happen.
630 	 * Otherwise the RST was misdirected or spoofed.
631 	 */
632 	if (sc == NULL) {
633 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
634 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
635 			    "syncache entry (possibly syncookie only), "
636 			    "segment ignored\n", s, __func__);
637 		TCPSTAT_INC(tcps_badrst);
638 		goto done;
639 	}
640 
641 	/* The remote UDP encaps port does not match. */
642 	if (sc->sc_port != port) {
643 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
644 			log(LOG_DEBUG, "%s; %s: Spurious RST with matching "
645 			    "syncache entry but non-matching UDP encaps port, "
646 			    "segment ignored\n", s, __func__);
647 		TCPSTAT_INC(tcps_badrst);
648 		goto done;
649 	}
650 
651 	/*
652 	 * If the RST bit is set, check the sequence number to see
653 	 * if this is a valid reset segment.
654 	 *
655 	 * RFC 793 page 37:
656 	 *   In all states except SYN-SENT, all reset (RST) segments
657 	 *   are validated by checking their SEQ-fields.  A reset is
658 	 *   valid if its sequence number is in the window.
659 	 *
660 	 * RFC 793 page 69:
661 	 *   There are four cases for the acceptability test for an incoming
662 	 *   segment:
663 	 *
664 	 * Segment Receive  Test
665 	 * Length  Window
666 	 * ------- -------  -------------------------------------------
667 	 *    0       0     SEG.SEQ = RCV.NXT
668 	 *    0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
669 	 *   >0       0     not acceptable
670 	 *   >0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
671 	 *               or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
672 	 *
673 	 * Note that when receiving a SYN segment in the LISTEN state,
674 	 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
675 	 * described in RFC 793, page 66.
676 	 */
677 	if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
678 	    SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
679 	    (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
680 		if (V_tcp_insecure_rst ||
681 		    th->th_seq == sc->sc_irs + 1) {
682 			syncache_drop(sc, sch);
683 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
684 				log(LOG_DEBUG,
685 				    "%s; %s: Our SYN|ACK was rejected, "
686 				    "connection attempt aborted by remote "
687 				    "endpoint\n",
688 				    s, __func__);
689 			TCPSTAT_INC(tcps_sc_reset);
690 		} else {
691 			TCPSTAT_INC(tcps_badrst);
692 			/* Send challenge ACK. */
693 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
694 				log(LOG_DEBUG, "%s; %s: RST with invalid "
695 				    " SEQ %u != NXT %u (+WND %u), "
696 				    "sending challenge ACK\n",
697 				    s, __func__,
698 				    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
699 			syncache_respond(sc, m, TH_ACK);
700 		}
701 	} else {
702 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
703 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
704 			    "NXT %u (+WND %u), segment ignored\n",
705 			    s, __func__,
706 			    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
707 		TCPSTAT_INC(tcps_badrst);
708 	}
709 
710 done:
711 	if (s != NULL)
712 		free(s, M_TCPLOG);
713 	SCH_UNLOCK(sch);
714 }
715 
716 void
717 syncache_badack(struct in_conninfo *inc, uint16_t port)
718 {
719 	struct syncache *sc;
720 	struct syncache_head *sch;
721 
722 	if (syncache_cookiesonly())
723 		return;
724 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
725 	SCH_LOCK_ASSERT(sch);
726 	if ((sc != NULL) && (sc->sc_port == port)) {
727 		syncache_drop(sc, sch);
728 		TCPSTAT_INC(tcps_sc_badack);
729 	}
730 	SCH_UNLOCK(sch);
731 }
732 
733 void
734 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port)
735 {
736 	struct syncache *sc;
737 	struct syncache_head *sch;
738 
739 	if (syncache_cookiesonly())
740 		return;
741 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
742 	SCH_LOCK_ASSERT(sch);
743 	if (sc == NULL)
744 		goto done;
745 
746 	/* If the port != sc_port, then it's a bogus ICMP msg */
747 	if (port != sc->sc_port)
748 		goto done;
749 
750 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
751 	if (ntohl(th_seq) != sc->sc_iss)
752 		goto done;
753 
754 	/*
755 	 * If we've rertransmitted 3 times and this is our second error,
756 	 * we remove the entry.  Otherwise, we allow it to continue on.
757 	 * This prevents us from incorrectly nuking an entry during a
758 	 * spurious network outage.
759 	 *
760 	 * See tcp_notify().
761 	 */
762 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
763 		sc->sc_flags |= SCF_UNREACH;
764 		goto done;
765 	}
766 	syncache_drop(sc, sch);
767 	TCPSTAT_INC(tcps_sc_unreach);
768 done:
769 	SCH_UNLOCK(sch);
770 }
771 
772 /*
773  * Build a new TCP socket structure from a syncache entry.
774  *
775  * On success return the newly created socket with its underlying inp locked.
776  */
777 static struct socket *
778 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
779 {
780 	struct tcp_function_block *blk;
781 	struct inpcb *inp = NULL;
782 	struct socket *so;
783 	struct tcpcb *tp;
784 	int error;
785 	char *s;
786 
787 	NET_EPOCH_ASSERT();
788 
789 	/*
790 	 * Ok, create the full blown connection, and set things up
791 	 * as they would have been set up if we had created the
792 	 * connection when the SYN arrived.
793 	 */
794 	if ((so = solisten_clone(lso)) == NULL)
795 		goto allocfail;
796 #ifdef MAC
797 	mac_socketpeer_set_from_mbuf(m, so);
798 #endif
799 	error = in_pcballoc(so, &V_tcbinfo);
800 	if (error) {
801 		sodealloc(so);
802 		goto allocfail;
803 	}
804 	inp = sotoinpcb(so);
805 	if ((tp = tcp_newtcpcb(inp)) == NULL) {
806 		in_pcbfree(inp);
807 		sodealloc(so);
808 		goto allocfail;
809 	}
810 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
811 #ifdef INET6
812 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
813 		inp->inp_vflag &= ~INP_IPV4;
814 		inp->inp_vflag |= INP_IPV6;
815 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
816 	} else {
817 		inp->inp_vflag &= ~INP_IPV6;
818 		inp->inp_vflag |= INP_IPV4;
819 #endif
820 		inp->inp_ip_ttl = sc->sc_ip_ttl;
821 		inp->inp_ip_tos = sc->sc_ip_tos;
822 		inp->inp_laddr = sc->sc_inc.inc_laddr;
823 #ifdef INET6
824 	}
825 #endif
826 
827 	/*
828 	 * If there's an mbuf and it has a flowid, then let's initialise the
829 	 * inp with that particular flowid.
830 	 */
831 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
832 		inp->inp_flowid = m->m_pkthdr.flowid;
833 		inp->inp_flowtype = M_HASHTYPE_GET(m);
834 #ifdef NUMA
835 		inp->inp_numa_domain = m->m_pkthdr.numa_domain;
836 #endif
837 	}
838 
839 	inp->inp_lport = sc->sc_inc.inc_lport;
840 #ifdef INET6
841 	if (inp->inp_vflag & INP_IPV6PROTO) {
842 		struct inpcb *oinp = sotoinpcb(lso);
843 
844 		/*
845 		 * Inherit socket options from the listening socket.
846 		 * Note that in6p_inputopts are not (and should not be)
847 		 * copied, since it stores previously received options and is
848 		 * used to detect if each new option is different than the
849 		 * previous one and hence should be passed to a user.
850 		 * If we copied in6p_inputopts, a user would not be able to
851 		 * receive options just after calling the accept system call.
852 		 */
853 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
854 		if (oinp->in6p_outputopts)
855 			inp->in6p_outputopts =
856 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
857 		inp->in6p_hops = oinp->in6p_hops;
858 	}
859 
860 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
861 		struct sockaddr_in6 sin6;
862 
863 		sin6.sin6_family = AF_INET6;
864 		sin6.sin6_len = sizeof(sin6);
865 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
866 		sin6.sin6_port = sc->sc_inc.inc_fport;
867 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
868 		INP_HASH_WLOCK(&V_tcbinfo);
869 		error = in6_pcbconnect(inp, &sin6, thread0.td_ucred, false);
870 		INP_HASH_WUNLOCK(&V_tcbinfo);
871 		if (error != 0)
872 			goto abort;
873 		/* Override flowlabel from in6_pcbconnect. */
874 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
875 		inp->inp_flow |= sc->sc_flowlabel;
876 	}
877 #endif /* INET6 */
878 #if defined(INET) && defined(INET6)
879 	else
880 #endif
881 #ifdef INET
882 	{
883 		struct sockaddr_in sin;
884 
885 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
886 
887 		if (inp->inp_options == NULL) {
888 			inp->inp_options = sc->sc_ipopts;
889 			sc->sc_ipopts = NULL;
890 		}
891 
892 		sin.sin_family = AF_INET;
893 		sin.sin_len = sizeof(sin);
894 		sin.sin_addr = sc->sc_inc.inc_faddr;
895 		sin.sin_port = sc->sc_inc.inc_fport;
896 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
897 		INP_HASH_WLOCK(&V_tcbinfo);
898 		error = in_pcbconnect(inp, &sin, thread0.td_ucred, false);
899 		INP_HASH_WUNLOCK(&V_tcbinfo);
900 		if (error != 0)
901 			goto abort;
902 	}
903 #endif /* INET */
904 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
905 	/* Copy old policy into new socket's. */
906 	if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
907 		printf("syncache_socket: could not copy policy\n");
908 #endif
909 	tp->t_state = TCPS_SYN_RECEIVED;
910 	tp->iss = sc->sc_iss;
911 	tp->irs = sc->sc_irs;
912 	tp->t_port = sc->sc_port;
913 	tcp_rcvseqinit(tp);
914 	tcp_sendseqinit(tp);
915 	blk = sototcpcb(lso)->t_fb;
916 	if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
917 		/*
918 		 * Our parents t_fb was not the default,
919 		 * we need to release our ref on tp->t_fb and
920 		 * pickup one on the new entry.
921 		 */
922 		struct tcp_function_block *rblk;
923 		void *ptr = NULL;
924 
925 		rblk = find_and_ref_tcp_fb(blk);
926 		KASSERT(rblk != NULL,
927 		    ("cannot find blk %p out of syncache?", blk));
928 
929 		if (rblk->tfb_tcp_fb_init == NULL ||
930 		    (*rblk->tfb_tcp_fb_init)(tp, &ptr) == 0) {
931 			/* Release the old stack */
932 			if (tp->t_fb->tfb_tcp_fb_fini != NULL)
933 				(*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
934 			refcount_release(&tp->t_fb->tfb_refcnt);
935 			/* Now set in all the pointers */
936 			tp->t_fb = rblk;
937 			tp->t_fb_ptr = ptr;
938 		} else {
939 			/*
940 			 * Initialization failed. Release the reference count on
941 			 * the looked up default stack.
942 			 */
943 			refcount_release(&rblk->tfb_refcnt);
944 		}
945 	}
946 	tp->snd_wl1 = sc->sc_irs;
947 	tp->snd_max = tp->iss + 1;
948 	tp->snd_nxt = tp->iss + 1;
949 	tp->rcv_up = sc->sc_irs + 1;
950 	tp->rcv_wnd = sc->sc_wnd;
951 	tp->rcv_adv += tp->rcv_wnd;
952 	tp->last_ack_sent = tp->rcv_nxt;
953 
954 	tp->t_flags = sototcpcb(lso)->t_flags &
955 	    (TF_LRD|TF_NOPUSH|TF_NODELAY);
956 	if (sc->sc_flags & SCF_NOOPT)
957 		tp->t_flags |= TF_NOOPT;
958 	else {
959 		if (sc->sc_flags & SCF_WINSCALE) {
960 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
961 			tp->snd_scale = sc->sc_requested_s_scale;
962 			tp->request_r_scale = sc->sc_requested_r_scale;
963 		}
964 		if (sc->sc_flags & SCF_TIMESTAMP) {
965 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
966 			tp->ts_recent = sc->sc_tsreflect;
967 			tp->ts_recent_age = tcp_ts_getticks();
968 			tp->ts_offset = sc->sc_tsoff;
969 		}
970 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
971 		if (sc->sc_flags & SCF_SIGNATURE)
972 			tp->t_flags |= TF_SIGNATURE;
973 #endif
974 		if (sc->sc_flags & SCF_SACK)
975 			tp->t_flags |= TF_SACK_PERMIT;
976 	}
977 
978 	tcp_ecn_syncache_socket(tp, sc);
979 
980 	/*
981 	 * Set up MSS and get cached values from tcp_hostcache.
982 	 * This might overwrite some of the defaults we just set.
983 	 */
984 	tcp_mss(tp, sc->sc_peer_mss);
985 
986 	/*
987 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
988 	 * limited to one segment in cc_conn_init().
989 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
990 	 */
991 	if (sc->sc_rxmits > 1)
992 		tp->snd_cwnd = 1;
993 
994 #ifdef TCP_OFFLOAD
995 	/*
996 	 * Allow a TOE driver to install its hooks.  Note that we hold the
997 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
998 	 * new connection before the TOE driver has done its thing.
999 	 */
1000 	if (ADDED_BY_TOE(sc)) {
1001 		struct toedev *tod = sc->sc_tod;
1002 
1003 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
1004 	}
1005 #endif
1006 #ifdef TCP_BLACKBOX
1007 	/*
1008 	 * Inherit the log state from the listening socket, if
1009 	 * - the log state of the listening socket is not off and
1010 	 * - the listening socket was not auto selected from all sessions and
1011 	 * - a log id is not set on the listening socket.
1012 	 * This avoids inheriting a log state which was automatically set.
1013 	 */
1014 	if ((tcp_get_bblog_state(sototcpcb(lso)) != TCP_LOG_STATE_OFF) &&
1015 	    ((sototcpcb(lso)->t_flags2 & TF2_LOG_AUTO) == 0) &&
1016 	    (sototcpcb(lso)->t_lib == NULL)) {
1017 		tcp_log_state_change(tp, tcp_get_bblog_state(sototcpcb(lso)));
1018 	}
1019 #endif
1020 	/*
1021 	 * Copy and activate timers.
1022 	 */
1023 	tp->t_maxunacktime = sototcpcb(lso)->t_maxunacktime;
1024 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
1025 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
1026 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
1027 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1028 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1029 
1030 	TCPSTAT_INC(tcps_accepts);
1031 	TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, TCPS_LISTEN);
1032 
1033 	if (!solisten_enqueue(so, SS_ISCONNECTED))
1034 		tp->t_flags |= TF_SONOTCONN;
1035 
1036 	return (so);
1037 
1038 allocfail:
1039 	/*
1040 	 * Drop the connection; we will either send a RST or have the peer
1041 	 * retransmit its SYN again after its RTO and try again.
1042 	 */
1043 	if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1044 		log(LOG_DEBUG, "%s; %s: Socket create failed "
1045 		    "due to limits or memory shortage\n",
1046 		    s, __func__);
1047 		free(s, M_TCPLOG);
1048 	}
1049 	TCPSTAT_INC(tcps_listendrop);
1050 	return (NULL);
1051 
1052 abort:
1053 	in_pcbfree(inp);
1054 	sodealloc(so);
1055 	if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1056 		log(LOG_DEBUG, "%s; %s: in%s_pcbconnect failed with error %i\n",
1057 		    s, __func__, (sc->sc_inc.inc_flags & INC_ISIPV6) ? "6" : "",
1058 		    error);
1059 		free(s, M_TCPLOG);
1060 	}
1061 	TCPSTAT_INC(tcps_listendrop);
1062 	return (NULL);
1063 }
1064 
1065 /*
1066  * This function gets called when we receive an ACK for a
1067  * socket in the LISTEN state.  We look up the connection
1068  * in the syncache, and if its there, we pull it out of
1069  * the cache and turn it into a full-blown connection in
1070  * the SYN-RECEIVED state.
1071  *
1072  * On syncache_socket() success the newly created socket
1073  * has its underlying inp locked.
1074  */
1075 int
1076 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1077     struct socket **lsop, struct mbuf *m, uint16_t port)
1078 {
1079 	struct syncache *sc;
1080 	struct syncache_head *sch;
1081 	struct syncache scs;
1082 	char *s;
1083 	bool locked;
1084 
1085 	NET_EPOCH_ASSERT();
1086 	KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1087 	    ("%s: can handle only ACK", __func__));
1088 
1089 	if (syncache_cookiesonly()) {
1090 		sc = NULL;
1091 		sch = syncache_hashbucket(inc);
1092 		locked = false;
1093 	} else {
1094 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1095 		locked = true;
1096 		SCH_LOCK_ASSERT(sch);
1097 	}
1098 
1099 #ifdef INVARIANTS
1100 	/*
1101 	 * Test code for syncookies comparing the syncache stored
1102 	 * values with the reconstructed values from the cookie.
1103 	 */
1104 	if (sc != NULL)
1105 		syncookie_cmp(inc, sch, sc, th, to, *lsop, port);
1106 #endif
1107 
1108 	if (sc == NULL) {
1109 		/*
1110 		 * There is no syncache entry, so see if this ACK is
1111 		 * a returning syncookie.  To do this, first:
1112 		 *  A. Check if syncookies are used in case of syncache
1113 		 *     overflows
1114 		 *  B. See if this socket has had a syncache entry dropped in
1115 		 *     the recent past. We don't want to accept a bogus
1116 		 *     syncookie if we've never received a SYN or accept it
1117 		 *     twice.
1118 		 *  C. check that the syncookie is valid.  If it is, then
1119 		 *     cobble up a fake syncache entry, and return.
1120 		 */
1121 		if (locked && !V_tcp_syncookies) {
1122 			SCH_UNLOCK(sch);
1123 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1124 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1125 				    "segment rejected (syncookies disabled)\n",
1126 				    s, __func__);
1127 			goto failed;
1128 		}
1129 		if (locked && !V_tcp_syncookiesonly &&
1130 		    sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1131 			SCH_UNLOCK(sch);
1132 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1133 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1134 				    "segment rejected (no syncache entry)\n",
1135 				    s, __func__);
1136 			goto failed;
1137 		}
1138 		bzero(&scs, sizeof(scs));
1139 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port);
1140 		if (locked)
1141 			SCH_UNLOCK(sch);
1142 		if (sc == NULL) {
1143 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1144 				log(LOG_DEBUG, "%s; %s: Segment failed "
1145 				    "SYNCOOKIE authentication, segment rejected "
1146 				    "(probably spoofed)\n", s, __func__);
1147 			goto failed;
1148 		}
1149 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1150 		/* If received ACK has MD5 signature, check it. */
1151 		if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1152 		    (!TCPMD5_ENABLED() ||
1153 		    TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1154 			/* Drop the ACK. */
1155 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1156 				log(LOG_DEBUG, "%s; %s: Segment rejected, "
1157 				    "MD5 signature doesn't match.\n",
1158 				    s, __func__);
1159 				free(s, M_TCPLOG);
1160 			}
1161 			TCPSTAT_INC(tcps_sig_err_sigopt);
1162 			return (-1); /* Do not send RST */
1163 		}
1164 #endif /* TCP_SIGNATURE */
1165 		TCPSTATES_INC(TCPS_SYN_RECEIVED);
1166 	} else {
1167 		if (sc->sc_port != port) {
1168 			SCH_UNLOCK(sch);
1169 			return (0);
1170 		}
1171 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1172 		/*
1173 		 * If listening socket requested TCP digests, check that
1174 		 * received ACK has signature and it is correct.
1175 		 * If not, drop the ACK and leave sc entry in th cache,
1176 		 * because SYN was received with correct signature.
1177 		 */
1178 		if (sc->sc_flags & SCF_SIGNATURE) {
1179 			if ((to->to_flags & TOF_SIGNATURE) == 0) {
1180 				/* No signature */
1181 				TCPSTAT_INC(tcps_sig_err_nosigopt);
1182 				SCH_UNLOCK(sch);
1183 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1184 					log(LOG_DEBUG, "%s; %s: Segment "
1185 					    "rejected, MD5 signature wasn't "
1186 					    "provided.\n", s, __func__);
1187 					free(s, M_TCPLOG);
1188 				}
1189 				return (-1); /* Do not send RST */
1190 			}
1191 			if (!TCPMD5_ENABLED() ||
1192 			    TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1193 				/* Doesn't match or no SA */
1194 				SCH_UNLOCK(sch);
1195 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1196 					log(LOG_DEBUG, "%s; %s: Segment "
1197 					    "rejected, MD5 signature doesn't "
1198 					    "match.\n", s, __func__);
1199 					free(s, M_TCPLOG);
1200 				}
1201 				return (-1); /* Do not send RST */
1202 			}
1203 		}
1204 #endif /* TCP_SIGNATURE */
1205 
1206 		/*
1207 		 * RFC 7323 PAWS: If we have a timestamp on this segment and
1208 		 * it's less than ts_recent, drop it.
1209 		 * XXXMT: RFC 7323 also requires to send an ACK.
1210 		 *        In tcp_input.c this is only done for TCP segments
1211 		 *        with user data, so be consistent here and just drop
1212 		 *        the segment.
1213 		 */
1214 		if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1215 		    TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1216 			SCH_UNLOCK(sch);
1217 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1218 				log(LOG_DEBUG,
1219 				    "%s; %s: SEG.TSval %u < TS.Recent %u, "
1220 				    "segment dropped\n", s, __func__,
1221 				    to->to_tsval, sc->sc_tsreflect);
1222 				free(s, M_TCPLOG);
1223 			}
1224 			return (-1);  /* Do not send RST */
1225 		}
1226 
1227 		/*
1228 		 * If timestamps were not negotiated during SYN/ACK and a
1229 		 * segment with a timestamp is received, ignore the
1230 		 * timestamp and process the packet normally.
1231 		 * See section 3.2 of RFC 7323.
1232 		 */
1233 		if (!(sc->sc_flags & SCF_TIMESTAMP) &&
1234 		    (to->to_flags & TOF_TS)) {
1235 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1236 				log(LOG_DEBUG, "%s; %s: Timestamp not "
1237 				    "expected, segment processed normally\n",
1238 				    s, __func__);
1239 				free(s, M_TCPLOG);
1240 				s = NULL;
1241 			}
1242 		}
1243 
1244 		/*
1245 		 * If timestamps were negotiated during SYN/ACK and a
1246 		 * segment without a timestamp is received, silently drop
1247 		 * the segment, unless the missing timestamps are tolerated.
1248 		 * See section 3.2 of RFC 7323.
1249 		 */
1250 		if ((sc->sc_flags & SCF_TIMESTAMP) &&
1251 		    !(to->to_flags & TOF_TS)) {
1252 			if (V_tcp_tolerate_missing_ts) {
1253 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1254 					log(LOG_DEBUG,
1255 					    "%s; %s: Timestamp missing, "
1256 					    "segment processed normally\n",
1257 					    s, __func__);
1258 					free(s, M_TCPLOG);
1259 				}
1260 			} else {
1261 				SCH_UNLOCK(sch);
1262 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1263 					log(LOG_DEBUG,
1264 					    "%s; %s: Timestamp missing, "
1265 					    "segment silently dropped\n",
1266 					    s, __func__);
1267 					free(s, M_TCPLOG);
1268 				}
1269 				return (-1);  /* Do not send RST */
1270 			}
1271 		}
1272 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1273 		sch->sch_length--;
1274 #ifdef TCP_OFFLOAD
1275 		if (ADDED_BY_TOE(sc)) {
1276 			struct toedev *tod = sc->sc_tod;
1277 
1278 			tod->tod_syncache_removed(tod, sc->sc_todctx);
1279 		}
1280 #endif
1281 		SCH_UNLOCK(sch);
1282 	}
1283 
1284 	/*
1285 	 * Segment validation:
1286 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1287 	 */
1288 	if (th->th_ack != sc->sc_iss + 1) {
1289 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1290 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1291 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1292 		goto failed;
1293 	}
1294 
1295 	/*
1296 	 * The SEQ must fall in the window starting at the received
1297 	 * initial receive sequence number + 1 (the SYN).
1298 	 */
1299 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1300 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1301 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1302 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1303 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1304 		goto failed;
1305 	}
1306 
1307 	*lsop = syncache_socket(sc, *lsop, m);
1308 
1309 	if (__predict_false(*lsop == NULL)) {
1310 		TCPSTAT_INC(tcps_sc_aborted);
1311 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1312 	} else
1313 		TCPSTAT_INC(tcps_sc_completed);
1314 
1315 /* how do we find the inp for the new socket? */
1316 	if (sc != &scs)
1317 		syncache_free(sc);
1318 	return (1);
1319 failed:
1320 	if (sc != NULL) {
1321 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1322 		if (sc != &scs)
1323 			syncache_free(sc);
1324 	}
1325 	if (s != NULL)
1326 		free(s, M_TCPLOG);
1327 	*lsop = NULL;
1328 	return (0);
1329 }
1330 
1331 static struct socket *
1332 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m,
1333     uint64_t response_cookie)
1334 {
1335 	struct inpcb *inp;
1336 	struct tcpcb *tp;
1337 	unsigned int *pending_counter;
1338 	struct socket *so;
1339 
1340 	NET_EPOCH_ASSERT();
1341 
1342 	pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending;
1343 	so = syncache_socket(sc, lso, m);
1344 	if (so == NULL) {
1345 		TCPSTAT_INC(tcps_sc_aborted);
1346 		atomic_subtract_int(pending_counter, 1);
1347 	} else {
1348 		soisconnected(so);
1349 		inp = sotoinpcb(so);
1350 		tp = intotcpcb(inp);
1351 		tp->t_flags |= TF_FASTOPEN;
1352 		tp->t_tfo_cookie.server = response_cookie;
1353 		tp->snd_max = tp->iss;
1354 		tp->snd_nxt = tp->iss;
1355 		tp->t_tfo_pending = pending_counter;
1356 		TCPSTATES_INC(TCPS_SYN_RECEIVED);
1357 		TCPSTAT_INC(tcps_sc_completed);
1358 	}
1359 
1360 	return (so);
1361 }
1362 
1363 /*
1364  * Given a LISTEN socket and an inbound SYN request, add
1365  * this to the syn cache, and send back a segment:
1366  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1367  * to the source.
1368  *
1369  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1370  * Doing so would require that we hold onto the data and deliver it
1371  * to the application.  However, if we are the target of a SYN-flood
1372  * DoS attack, an attacker could send data which would eventually
1373  * consume all available buffer space if it were ACKed.  By not ACKing
1374  * the data, we avoid this DoS scenario.
1375  *
1376  * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1377  * cookie is processed and a new socket is created.  In this case, any data
1378  * accompanying the SYN will be queued to the socket by tcp_input() and will
1379  * be ACKed either when the application sends response data or the delayed
1380  * ACK timer expires, whichever comes first.
1381  */
1382 struct socket *
1383 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1384     struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod,
1385     void *todctx, uint8_t iptos, uint16_t port)
1386 {
1387 	struct tcpcb *tp;
1388 	struct socket *rv = NULL;
1389 	struct syncache *sc = NULL;
1390 	struct syncache_head *sch;
1391 	struct mbuf *ipopts = NULL;
1392 	u_int ltflags;
1393 	int win, ip_ttl, ip_tos;
1394 	char *s;
1395 #ifdef INET6
1396 	int autoflowlabel = 0;
1397 #endif
1398 #ifdef MAC
1399 	struct label *maclabel;
1400 #endif
1401 	struct syncache scs;
1402 	struct ucred *cred;
1403 	uint64_t tfo_response_cookie;
1404 	unsigned int *tfo_pending = NULL;
1405 	int tfo_cookie_valid = 0;
1406 	int tfo_response_cookie_valid = 0;
1407 	bool locked;
1408 
1409 	INP_RLOCK_ASSERT(inp);			/* listen socket */
1410 	KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1411 	    ("%s: unexpected tcp flags", __func__));
1412 
1413 	/*
1414 	 * Combine all so/tp operations very early to drop the INP lock as
1415 	 * soon as possible.
1416 	 */
1417 	KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1418 	tp = sototcpcb(so);
1419 	cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred);
1420 
1421 #ifdef INET6
1422 	if (inc->inc_flags & INC_ISIPV6) {
1423 		if (inp->inp_flags & IN6P_AUTOFLOWLABEL) {
1424 			autoflowlabel = 1;
1425 		}
1426 		ip_ttl = in6_selecthlim(inp, NULL);
1427 		if ((inp->in6p_outputopts == NULL) ||
1428 		    (inp->in6p_outputopts->ip6po_tclass == -1)) {
1429 			ip_tos = 0;
1430 		} else {
1431 			ip_tos = inp->in6p_outputopts->ip6po_tclass;
1432 		}
1433 	}
1434 #endif
1435 #if defined(INET6) && defined(INET)
1436 	else
1437 #endif
1438 #ifdef INET
1439 	{
1440 		ip_ttl = inp->inp_ip_ttl;
1441 		ip_tos = inp->inp_ip_tos;
1442 	}
1443 #endif
1444 	win = so->sol_sbrcv_hiwat;
1445 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1446 
1447 	if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
1448 	    (tp->t_tfo_pending != NULL) &&
1449 	    (to->to_flags & TOF_FASTOPEN)) {
1450 		/*
1451 		 * Limit the number of pending TFO connections to
1452 		 * approximately half of the queue limit.  This prevents TFO
1453 		 * SYN floods from starving the service by filling the
1454 		 * listen queue with bogus TFO connections.
1455 		 */
1456 		if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1457 		    (so->sol_qlimit / 2)) {
1458 			int result;
1459 
1460 			result = tcp_fastopen_check_cookie(inc,
1461 			    to->to_tfo_cookie, to->to_tfo_len,
1462 			    &tfo_response_cookie);
1463 			tfo_cookie_valid = (result > 0);
1464 			tfo_response_cookie_valid = (result >= 0);
1465 		}
1466 
1467 		/*
1468 		 * Remember the TFO pending counter as it will have to be
1469 		 * decremented below if we don't make it to syncache_tfo_expand().
1470 		 */
1471 		tfo_pending = tp->t_tfo_pending;
1472 	}
1473 
1474 #ifdef MAC
1475 	if (mac_syncache_init(&maclabel) != 0) {
1476 		INP_RUNLOCK(inp);
1477 		goto done;
1478 	} else
1479 		mac_syncache_create(maclabel, inp);
1480 #endif
1481 	if (!tfo_cookie_valid)
1482 		INP_RUNLOCK(inp);
1483 
1484 	/*
1485 	 * Remember the IP options, if any.
1486 	 */
1487 #ifdef INET6
1488 	if (!(inc->inc_flags & INC_ISIPV6))
1489 #endif
1490 #ifdef INET
1491 		ipopts = (m) ? ip_srcroute(m) : NULL;
1492 #else
1493 		ipopts = NULL;
1494 #endif
1495 
1496 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1497 	/*
1498 	 * When the socket is TCP-MD5 enabled check that,
1499 	 *  - a signed packet is valid
1500 	 *  - a non-signed packet does not have a security association
1501 	 *
1502 	 *  If a signed packet fails validation or a non-signed packet has a
1503 	 *  security association, the packet will be dropped.
1504 	 */
1505 	if (ltflags & TF_SIGNATURE) {
1506 		if (to->to_flags & TOF_SIGNATURE) {
1507 			if (!TCPMD5_ENABLED() ||
1508 			    TCPMD5_INPUT(m, th, to->to_signature) != 0)
1509 				goto done;
1510 		} else {
1511 			if (TCPMD5_ENABLED() &&
1512 			    TCPMD5_INPUT(m, NULL, NULL) != ENOENT)
1513 				goto done;
1514 		}
1515 	} else if (to->to_flags & TOF_SIGNATURE)
1516 		goto done;
1517 #endif	/* TCP_SIGNATURE */
1518 	/*
1519 	 * See if we already have an entry for this connection.
1520 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1521 	 *
1522 	 * XXX: should the syncache be re-initialized with the contents
1523 	 * of the new SYN here (which may have different options?)
1524 	 *
1525 	 * XXX: We do not check the sequence number to see if this is a
1526 	 * real retransmit or a new connection attempt.  The question is
1527 	 * how to handle such a case; either ignore it as spoofed, or
1528 	 * drop the current entry and create a new one?
1529 	 */
1530 	if (syncache_cookiesonly()) {
1531 		sc = NULL;
1532 		sch = syncache_hashbucket(inc);
1533 		locked = false;
1534 	} else {
1535 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1536 		locked = true;
1537 		SCH_LOCK_ASSERT(sch);
1538 	}
1539 	if (sc != NULL) {
1540 		if (tfo_cookie_valid)
1541 			INP_RUNLOCK(inp);
1542 		TCPSTAT_INC(tcps_sc_dupsyn);
1543 		if (ipopts) {
1544 			/*
1545 			 * If we were remembering a previous source route,
1546 			 * forget it and use the new one we've been given.
1547 			 */
1548 			if (sc->sc_ipopts)
1549 				(void) m_free(sc->sc_ipopts);
1550 			sc->sc_ipopts = ipopts;
1551 		}
1552 		/*
1553 		 * Update timestamp if present.
1554 		 */
1555 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1556 			sc->sc_tsreflect = to->to_tsval;
1557 		else
1558 			sc->sc_flags &= ~SCF_TIMESTAMP;
1559 		/*
1560 		 * Adjust ECN response if needed, e.g. different
1561 		 * IP ECN field, or a fallback by the remote host.
1562 		 */
1563 		if (sc->sc_flags & SCF_ECN_MASK) {
1564 			sc->sc_flags &= ~SCF_ECN_MASK;
1565 			sc->sc_flags = tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1566 		}
1567 #ifdef MAC
1568 		/*
1569 		 * Since we have already unconditionally allocated label
1570 		 * storage, free it up.  The syncache entry will already
1571 		 * have an initialized label we can use.
1572 		 */
1573 		mac_syncache_destroy(&maclabel);
1574 #endif
1575 		TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1576 		/* Retransmit SYN|ACK and reset retransmit count. */
1577 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1578 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1579 			    "resetting timer and retransmitting SYN|ACK\n",
1580 			    s, __func__);
1581 			free(s, M_TCPLOG);
1582 		}
1583 		if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1584 			sc->sc_rxmits = 0;
1585 			syncache_timeout(sc, sch, 1);
1586 			TCPSTAT_INC(tcps_sndacks);
1587 			TCPSTAT_INC(tcps_sndtotal);
1588 		}
1589 		SCH_UNLOCK(sch);
1590 		goto donenoprobe;
1591 	}
1592 
1593 	if (tfo_cookie_valid) {
1594 		bzero(&scs, sizeof(scs));
1595 		sc = &scs;
1596 		goto skip_alloc;
1597 	}
1598 
1599 	/*
1600 	 * Skip allocating a syncache entry if we are just going to discard
1601 	 * it later.
1602 	 */
1603 	if (!locked) {
1604 		bzero(&scs, sizeof(scs));
1605 		sc = &scs;
1606 	} else
1607 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1608 	if (sc == NULL) {
1609 		/*
1610 		 * The zone allocator couldn't provide more entries.
1611 		 * Treat this as if the cache was full; drop the oldest
1612 		 * entry and insert the new one.
1613 		 */
1614 		TCPSTAT_INC(tcps_sc_zonefail);
1615 		if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1616 			sch->sch_last_overflow = time_uptime;
1617 			syncache_drop(sc, sch);
1618 			syncache_pause(inc);
1619 		}
1620 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1621 		if (sc == NULL) {
1622 			if (V_tcp_syncookies) {
1623 				bzero(&scs, sizeof(scs));
1624 				sc = &scs;
1625 			} else {
1626 				KASSERT(locked,
1627 				    ("%s: bucket unexpectedly unlocked",
1628 				    __func__));
1629 				SCH_UNLOCK(sch);
1630 				if (ipopts)
1631 					(void) m_free(ipopts);
1632 				goto done;
1633 			}
1634 		}
1635 	}
1636 
1637 skip_alloc:
1638 	if (!tfo_cookie_valid && tfo_response_cookie_valid)
1639 		sc->sc_tfo_cookie = &tfo_response_cookie;
1640 
1641 	/*
1642 	 * Fill in the syncache values.
1643 	 */
1644 #ifdef MAC
1645 	sc->sc_label = maclabel;
1646 #endif
1647 	sc->sc_cred = cred;
1648 	sc->sc_port = port;
1649 	cred = NULL;
1650 	sc->sc_ipopts = ipopts;
1651 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1652 	sc->sc_ip_tos = ip_tos;
1653 	sc->sc_ip_ttl = ip_ttl;
1654 #ifdef TCP_OFFLOAD
1655 	sc->sc_tod = tod;
1656 	sc->sc_todctx = todctx;
1657 #endif
1658 	sc->sc_irs = th->th_seq;
1659 	sc->sc_flags = 0;
1660 	sc->sc_flowlabel = 0;
1661 
1662 	/*
1663 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1664 	 * win was derived from socket earlier in the function.
1665 	 */
1666 	win = imax(win, 0);
1667 	win = imin(win, TCP_MAXWIN);
1668 	sc->sc_wnd = win;
1669 
1670 	if (V_tcp_do_rfc1323 &&
1671 	    !(ltflags & TF_NOOPT)) {
1672 		/*
1673 		 * A timestamp received in a SYN makes
1674 		 * it ok to send timestamp requests and replies.
1675 		 */
1676 		if ((to->to_flags & TOF_TS) && (V_tcp_do_rfc1323 != 2)) {
1677 			sc->sc_tsreflect = to->to_tsval;
1678 			sc->sc_flags |= SCF_TIMESTAMP;
1679 			sc->sc_tsoff = tcp_new_ts_offset(inc);
1680 		}
1681 		if ((to->to_flags & TOF_SCALE) && (V_tcp_do_rfc1323 != 3)) {
1682 			int wscale = 0;
1683 
1684 			/*
1685 			 * Pick the smallest possible scaling factor that
1686 			 * will still allow us to scale up to sb_max, aka
1687 			 * kern.ipc.maxsockbuf.
1688 			 *
1689 			 * We do this because there are broken firewalls that
1690 			 * will corrupt the window scale option, leading to
1691 			 * the other endpoint believing that our advertised
1692 			 * window is unscaled.  At scale factors larger than
1693 			 * 5 the unscaled window will drop below 1500 bytes,
1694 			 * leading to serious problems when traversing these
1695 			 * broken firewalls.
1696 			 *
1697 			 * With the default maxsockbuf of 256K, a scale factor
1698 			 * of 3 will be chosen by this algorithm.  Those who
1699 			 * choose a larger maxsockbuf should watch out
1700 			 * for the compatibility problems mentioned above.
1701 			 *
1702 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1703 			 * or <SYN,ACK>) segment itself is never scaled.
1704 			 */
1705 			while (wscale < TCP_MAX_WINSHIFT &&
1706 			    (TCP_MAXWIN << wscale) < sb_max)
1707 				wscale++;
1708 			sc->sc_requested_r_scale = wscale;
1709 			sc->sc_requested_s_scale = to->to_wscale;
1710 			sc->sc_flags |= SCF_WINSCALE;
1711 		}
1712 	}
1713 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1714 	/*
1715 	 * If incoming packet has an MD5 signature, flag this in the
1716 	 * syncache so that syncache_respond() will do the right thing
1717 	 * with the SYN+ACK.
1718 	 */
1719 	if (to->to_flags & TOF_SIGNATURE)
1720 		sc->sc_flags |= SCF_SIGNATURE;
1721 #endif	/* TCP_SIGNATURE */
1722 	if (to->to_flags & TOF_SACKPERM)
1723 		sc->sc_flags |= SCF_SACK;
1724 	if (to->to_flags & TOF_MSS)
1725 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1726 	if (ltflags & TF_NOOPT)
1727 		sc->sc_flags |= SCF_NOOPT;
1728 	/* ECN Handshake */
1729 	if (V_tcp_do_ecn && (tp->t_flags2 & TF2_CANNOT_DO_ECN) == 0)
1730 		sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1731 
1732 	if (V_tcp_syncookies)
1733 		sc->sc_iss = syncookie_generate(sch, sc);
1734 	else
1735 		sc->sc_iss = arc4random();
1736 #ifdef INET6
1737 	if (autoflowlabel) {
1738 		if (V_tcp_syncookies)
1739 			sc->sc_flowlabel = sc->sc_iss;
1740 		else
1741 			sc->sc_flowlabel = ip6_randomflowlabel();
1742 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1743 	}
1744 #endif
1745 	if (locked)
1746 		SCH_UNLOCK(sch);
1747 
1748 	if (tfo_cookie_valid) {
1749 		rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie);
1750 		/* INP_RUNLOCK(inp) will be performed by the caller */
1751 		goto tfo_expanded;
1752 	}
1753 
1754 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1755 	/*
1756 	 * Do a standard 3-way handshake.
1757 	 */
1758 	if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1759 		if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1760 			syncache_free(sc);
1761 		else if (sc != &scs)
1762 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1763 		TCPSTAT_INC(tcps_sndacks);
1764 		TCPSTAT_INC(tcps_sndtotal);
1765 	} else {
1766 		if (sc != &scs)
1767 			syncache_free(sc);
1768 		TCPSTAT_INC(tcps_sc_dropped);
1769 	}
1770 	goto donenoprobe;
1771 
1772 done:
1773 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1774 donenoprobe:
1775 	if (m)
1776 		m_freem(m);
1777 	/*
1778 	 * If tfo_pending is not NULL here, then a TFO SYN that did not
1779 	 * result in a new socket was processed and the associated pending
1780 	 * counter has not yet been decremented.  All such TFO processing paths
1781 	 * transit this point.
1782 	 */
1783 	if (tfo_pending != NULL)
1784 		tcp_fastopen_decrement_counter(tfo_pending);
1785 
1786 tfo_expanded:
1787 	if (cred != NULL)
1788 		crfree(cred);
1789 #ifdef MAC
1790 	if (sc == &scs)
1791 		mac_syncache_destroy(&maclabel);
1792 #endif
1793 	return (rv);
1794 }
1795 
1796 /*
1797  * Send SYN|ACK or ACK to the peer.  Either in response to a peer's segment,
1798  * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1799  */
1800 static int
1801 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1802 {
1803 	struct ip *ip = NULL;
1804 	struct mbuf *m;
1805 	struct tcphdr *th = NULL;
1806 	struct udphdr *udp = NULL;
1807 	int optlen, error = 0;	/* Make compiler happy */
1808 	u_int16_t hlen, tlen, mssopt, ulen;
1809 	struct tcpopt to;
1810 #ifdef INET6
1811 	struct ip6_hdr *ip6 = NULL;
1812 #endif
1813 
1814 	NET_EPOCH_ASSERT();
1815 
1816 	hlen =
1817 #ifdef INET6
1818 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1819 #endif
1820 		sizeof(struct ip);
1821 	tlen = hlen + sizeof(struct tcphdr);
1822 	if (sc->sc_port) {
1823 		tlen += sizeof(struct udphdr);
1824 	}
1825 	/* Determine MSS we advertize to other end of connection. */
1826 	mssopt = tcp_mssopt(&sc->sc_inc);
1827 	if (sc->sc_port)
1828 		mssopt -= V_tcp_udp_tunneling_overhead;
1829 	mssopt = max(mssopt, V_tcp_minmss);
1830 
1831 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1832 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1833 	    ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + "
1834 	    "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port,
1835 	    max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN));
1836 
1837 	/* Create the IP+TCP header from scratch. */
1838 	m = m_gethdr(M_NOWAIT, MT_DATA);
1839 	if (m == NULL)
1840 		return (ENOBUFS);
1841 #ifdef MAC
1842 	mac_syncache_create_mbuf(sc->sc_label, m);
1843 #endif
1844 	m->m_data += max_linkhdr;
1845 	m->m_len = tlen;
1846 	m->m_pkthdr.len = tlen;
1847 	m->m_pkthdr.rcvif = NULL;
1848 
1849 #ifdef INET6
1850 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1851 		ip6 = mtod(m, struct ip6_hdr *);
1852 		ip6->ip6_vfc = IPV6_VERSION;
1853 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1854 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1855 		ip6->ip6_plen = htons(tlen - hlen);
1856 		/* ip6_hlim is set after checksum */
1857 		/* Zero out traffic class and flow label. */
1858 		ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1859 		ip6->ip6_flow |= sc->sc_flowlabel;
1860 		if (sc->sc_port != 0) {
1861 			ip6->ip6_nxt = IPPROTO_UDP;
1862 			udp = (struct udphdr *)(ip6 + 1);
1863 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1864 			udp->uh_dport = sc->sc_port;
1865 			ulen = (tlen - sizeof(struct ip6_hdr));
1866 			th = (struct tcphdr *)(udp + 1);
1867 		} else {
1868 			ip6->ip6_nxt = IPPROTO_TCP;
1869 			th = (struct tcphdr *)(ip6 + 1);
1870 		}
1871 		ip6->ip6_flow |= htonl(sc->sc_ip_tos << IPV6_FLOWLABEL_LEN);
1872 	}
1873 #endif
1874 #if defined(INET6) && defined(INET)
1875 	else
1876 #endif
1877 #ifdef INET
1878 	{
1879 		ip = mtod(m, struct ip *);
1880 		ip->ip_v = IPVERSION;
1881 		ip->ip_hl = sizeof(struct ip) >> 2;
1882 		ip->ip_len = htons(tlen);
1883 		ip->ip_id = 0;
1884 		ip->ip_off = 0;
1885 		ip->ip_sum = 0;
1886 		ip->ip_src = sc->sc_inc.inc_laddr;
1887 		ip->ip_dst = sc->sc_inc.inc_faddr;
1888 		ip->ip_ttl = sc->sc_ip_ttl;
1889 		ip->ip_tos = sc->sc_ip_tos;
1890 
1891 		/*
1892 		 * See if we should do MTU discovery.  Route lookups are
1893 		 * expensive, so we will only unset the DF bit if:
1894 		 *
1895 		 *	1) path_mtu_discovery is disabled
1896 		 *	2) the SCF_UNREACH flag has been set
1897 		 */
1898 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1899 		       ip->ip_off |= htons(IP_DF);
1900 		if (sc->sc_port == 0) {
1901 			ip->ip_p = IPPROTO_TCP;
1902 			th = (struct tcphdr *)(ip + 1);
1903 		} else {
1904 			ip->ip_p = IPPROTO_UDP;
1905 			udp = (struct udphdr *)(ip + 1);
1906 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1907 			udp->uh_dport = sc->sc_port;
1908 			ulen = (tlen - sizeof(struct ip));
1909 			th = (struct tcphdr *)(udp + 1);
1910 		}
1911 	}
1912 #endif /* INET */
1913 	th->th_sport = sc->sc_inc.inc_lport;
1914 	th->th_dport = sc->sc_inc.inc_fport;
1915 
1916 	if (flags & TH_SYN)
1917 		th->th_seq = htonl(sc->sc_iss);
1918 	else
1919 		th->th_seq = htonl(sc->sc_iss + 1);
1920 	th->th_ack = htonl(sc->sc_irs + 1);
1921 	th->th_off = sizeof(struct tcphdr) >> 2;
1922 	th->th_win = htons(sc->sc_wnd);
1923 	th->th_urp = 0;
1924 
1925 	flags = tcp_ecn_syncache_respond(flags, sc);
1926 	tcp_set_flags(th, flags);
1927 
1928 	/* Tack on the TCP options. */
1929 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1930 		to.to_flags = 0;
1931 
1932 		if (flags & TH_SYN) {
1933 			to.to_mss = mssopt;
1934 			to.to_flags = TOF_MSS;
1935 			if (sc->sc_flags & SCF_WINSCALE) {
1936 				to.to_wscale = sc->sc_requested_r_scale;
1937 				to.to_flags |= TOF_SCALE;
1938 			}
1939 			if (sc->sc_flags & SCF_SACK)
1940 				to.to_flags |= TOF_SACKPERM;
1941 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1942 			if (sc->sc_flags & SCF_SIGNATURE)
1943 				to.to_flags |= TOF_SIGNATURE;
1944 #endif
1945 			if (sc->sc_tfo_cookie) {
1946 				to.to_flags |= TOF_FASTOPEN;
1947 				to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1948 				to.to_tfo_cookie = sc->sc_tfo_cookie;
1949 				/* don't send cookie again when retransmitting response */
1950 				sc->sc_tfo_cookie = NULL;
1951 			}
1952 		}
1953 		if (sc->sc_flags & SCF_TIMESTAMP) {
1954 			to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1955 			to.to_tsecr = sc->sc_tsreflect;
1956 			to.to_flags |= TOF_TS;
1957 		}
1958 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1959 
1960 		/* Adjust headers by option size. */
1961 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1962 		m->m_len += optlen;
1963 		m->m_pkthdr.len += optlen;
1964 #ifdef INET6
1965 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1966 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1967 		else
1968 #endif
1969 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1970 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1971 		if (sc->sc_flags & SCF_SIGNATURE) {
1972 			KASSERT(to.to_flags & TOF_SIGNATURE,
1973 			    ("tcp_addoptions() didn't set tcp_signature"));
1974 
1975 			/* NOTE: to.to_signature is inside of mbuf */
1976 			if (!TCPMD5_ENABLED() ||
1977 			    TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1978 				m_freem(m);
1979 				return (EACCES);
1980 			}
1981 		}
1982 #endif
1983 	} else
1984 		optlen = 0;
1985 
1986 	if (udp) {
1987 		ulen += optlen;
1988 		udp->uh_ulen = htons(ulen);
1989 	}
1990 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
1991 	/*
1992 	 * If we have peer's SYN and it has a flowid, then let's assign it to
1993 	 * our SYN|ACK.  ip6_output() and ip_output() will not assign flowid
1994 	 * to SYN|ACK due to lack of inp here.
1995 	 */
1996 	if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1997 		m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1998 		M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1999 	}
2000 #ifdef INET6
2001 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
2002 		if (sc->sc_port) {
2003 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
2004 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2005 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen,
2006 			      IPPROTO_UDP, 0);
2007 			th->th_sum = htons(0);
2008 		} else {
2009 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
2010 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2011 			th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
2012 			    IPPROTO_TCP, 0);
2013 		}
2014 		ip6->ip6_hlim = sc->sc_ip_ttl;
2015 #ifdef TCP_OFFLOAD
2016 		if (ADDED_BY_TOE(sc)) {
2017 			struct toedev *tod = sc->sc_tod;
2018 
2019 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2020 
2021 			return (error);
2022 		}
2023 #endif
2024 		TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
2025 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
2026 	}
2027 #endif
2028 #if defined(INET6) && defined(INET)
2029 	else
2030 #endif
2031 #ifdef INET
2032 	{
2033 		if (sc->sc_port) {
2034 			m->m_pkthdr.csum_flags = CSUM_UDP;
2035 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2036 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
2037 			      ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
2038 			th->th_sum = htons(0);
2039 		} else {
2040 			m->m_pkthdr.csum_flags = CSUM_TCP;
2041 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2042 			th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
2043 			    htons(tlen + optlen - hlen + IPPROTO_TCP));
2044 		}
2045 #ifdef TCP_OFFLOAD
2046 		if (ADDED_BY_TOE(sc)) {
2047 			struct toedev *tod = sc->sc_tod;
2048 
2049 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2050 
2051 			return (error);
2052 		}
2053 #endif
2054 		TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
2055 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
2056 	}
2057 #endif
2058 	return (error);
2059 }
2060 
2061 /*
2062  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
2063  * that exceed the capacity of the syncache by avoiding the storage of any
2064  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
2065  * attacks where the attacker does not have access to our responses.
2066  *
2067  * Syncookies encode and include all necessary information about the
2068  * connection setup within the SYN|ACK that we send back.  That way we
2069  * can avoid keeping any local state until the ACK to our SYN|ACK returns
2070  * (if ever).  Normally the syncache and syncookies are running in parallel
2071  * with the latter taking over when the former is exhausted.  When matching
2072  * syncache entry is found the syncookie is ignored.
2073  *
2074  * The only reliable information persisting the 3WHS is our initial sequence
2075  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
2076  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
2077  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
2078  * returns and signifies a legitimate connection if it matches the ACK.
2079  *
2080  * The available space of 32 bits to store the hash and to encode the SYN
2081  * option information is very tight and we should have at least 24 bits for
2082  * the MAC to keep the number of guesses by blind spoofing reasonably high.
2083  *
2084  * SYN option information we have to encode to fully restore a connection:
2085  * MSS: is imporant to chose an optimal segment size to avoid IP level
2086  *   fragmentation along the path.  The common MSS values can be encoded
2087  *   in a 3-bit table.  Uncommon values are captured by the next lower value
2088  *   in the table leading to a slight increase in packetization overhead.
2089  * WSCALE: is necessary to allow large windows to be used for high delay-
2090  *   bandwidth product links.  Not scaling the window when it was initially
2091  *   negotiated is bad for performance as lack of scaling further decreases
2092  *   the apparent available send window.  We only need to encode the WSCALE
2093  *   we received from the remote end.  Our end can be recalculated at any
2094  *   time.  The common WSCALE values can be encoded in a 3-bit table.
2095  *   Uncommon values are captured by the next lower value in the table
2096  *   making us under-estimate the available window size halving our
2097  *   theoretically possible maximum throughput for that connection.
2098  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2099  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2100  *   that are included in all segments on a connection.  We enable them when
2101  *   the ACK has them.
2102  *
2103  * Security of syncookies and attack vectors:
2104  *
2105  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2106  * together with the gloabl secret to make it unique per connection attempt.
2107  * Thus any change of any of those parameters results in a different MAC output
2108  * in an unpredictable way unless a collision is encountered.  24 bits of the
2109  * MAC are embedded into the ISS.
2110  *
2111  * To prevent replay attacks two rotating global secrets are updated with a
2112  * new random value every 15 seconds.  The life-time of a syncookie is thus
2113  * 15-30 seconds.
2114  *
2115  * Vector 1: Attacking the secret.  This requires finding a weakness in the
2116  * MAC itself or the way it is used here.  The attacker can do a chosen plain
2117  * text attack by varying and testing the all parameters under his control.
2118  * The strength depends on the size and randomness of the secret, and the
2119  * cryptographic security of the MAC function.  Due to the constant updating
2120  * of the secret the attacker has at most 29.999 seconds to find the secret
2121  * and launch spoofed connections.  After that he has to start all over again.
2122  *
2123  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
2124  * size an average of 4,823 attempts are required for a 50% chance of success
2125  * to spoof a single syncookie (birthday collision paradox).  However the
2126  * attacker is blind and doesn't know if one of his attempts succeeded unless
2127  * he has a side channel to interfere success from.  A single connection setup
2128  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2129  * This many attempts are required for each one blind spoofed connection.  For
2130  * every additional spoofed connection he has to launch another N attempts.
2131  * Thus for a sustained rate 100 spoofed connections per second approximately
2132  * 1,800,000 packets per second would have to be sent.
2133  *
2134  * NB: The MAC function should be fast so that it doesn't become a CPU
2135  * exhaustion attack vector itself.
2136  *
2137  * References:
2138  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2139  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2140  *   http://cr.yp.to/syncookies.html    (overview)
2141  *   http://cr.yp.to/syncookies/archive (details)
2142  *
2143  *
2144  * Schematic construction of a syncookie enabled Initial Sequence Number:
2145  *  0        1         2         3
2146  *  12345678901234567890123456789012
2147  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2148  *
2149  *  x 24 MAC (truncated)
2150  *  W  3 Send Window Scale index
2151  *  M  3 MSS index
2152  *  S  1 SACK permitted
2153  *  P  1 Odd/even secret
2154  */
2155 
2156 /*
2157  * Distribution and probability of certain MSS values.  Those in between are
2158  * rounded down to the next lower one.
2159  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2160  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
2161  */
2162 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2163 
2164 /*
2165  * Distribution and probability of certain WSCALE values.  We have to map the
2166  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2167  * bits based on prevalence of certain values.  Where we don't have an exact
2168  * match for are rounded down to the next lower one letting us under-estimate
2169  * the true available window.  At the moment this would happen only for the
2170  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2171  * and window size).  The absence of the WSCALE option (no scaling in either
2172  * direction) is encoded with index zero.
2173  * [WSCALE values histograms, Allman, 2012]
2174  *                            X 10 10 35  5  6 14 10%   by host
2175  *                            X 11  4  5  5 18 49  3%   by connections
2176  */
2177 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2178 
2179 /*
2180  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
2181  * and good cryptographic properties.
2182  */
2183 static uint32_t
2184 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2185     uint8_t *secbits, uintptr_t secmod)
2186 {
2187 	SIPHASH_CTX ctx;
2188 	uint32_t siphash[2];
2189 
2190 	SipHash24_Init(&ctx);
2191 	SipHash_SetKey(&ctx, secbits);
2192 	switch (inc->inc_flags & INC_ISIPV6) {
2193 #ifdef INET
2194 	case 0:
2195 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2196 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2197 		break;
2198 #endif
2199 #ifdef INET6
2200 	case INC_ISIPV6:
2201 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2202 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2203 		break;
2204 #endif
2205 	}
2206 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2207 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2208 	SipHash_Update(&ctx, &irs, sizeof(irs));
2209 	SipHash_Update(&ctx, &flags, sizeof(flags));
2210 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
2211 	SipHash_Final((u_int8_t *)&siphash, &ctx);
2212 
2213 	return (siphash[0] ^ siphash[1]);
2214 }
2215 
2216 static tcp_seq
2217 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2218 {
2219 	u_int i, secbit, wscale;
2220 	uint32_t iss, hash;
2221 	uint8_t *secbits;
2222 	union syncookie cookie;
2223 
2224 	cookie.cookie = 0;
2225 
2226 	/* Map our computed MSS into the 3-bit index. */
2227 	for (i = nitems(tcp_sc_msstab) - 1;
2228 	     tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2229 	     i--)
2230 		;
2231 	cookie.flags.mss_idx = i;
2232 
2233 	/*
2234 	 * Map the send window scale into the 3-bit index but only if
2235 	 * the wscale option was received.
2236 	 */
2237 	if (sc->sc_flags & SCF_WINSCALE) {
2238 		wscale = sc->sc_requested_s_scale;
2239 		for (i = nitems(tcp_sc_wstab) - 1;
2240 		    tcp_sc_wstab[i] > wscale && i > 0;
2241 		     i--)
2242 			;
2243 		cookie.flags.wscale_idx = i;
2244 	}
2245 
2246 	/* Can we do SACK? */
2247 	if (sc->sc_flags & SCF_SACK)
2248 		cookie.flags.sack_ok = 1;
2249 
2250 	/* Which of the two secrets to use. */
2251 	secbit = V_tcp_syncache.secret.oddeven & 0x1;
2252 	cookie.flags.odd_even = secbit;
2253 
2254 	secbits = V_tcp_syncache.secret.key[secbit];
2255 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2256 	    (uintptr_t)sch);
2257 
2258 	/*
2259 	 * Put the flags into the hash and XOR them to get better ISS number
2260 	 * variance.  This doesn't enhance the cryptographic strength and is
2261 	 * done to prevent the 8 cookie bits from showing up directly on the
2262 	 * wire.
2263 	 */
2264 	iss = hash & ~0xff;
2265 	iss |= cookie.cookie ^ (hash >> 24);
2266 
2267 	TCPSTAT_INC(tcps_sc_sendcookie);
2268 	return (iss);
2269 }
2270 
2271 static struct syncache *
2272 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2273     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2274     struct socket *lso, uint16_t port)
2275 {
2276 	uint32_t hash;
2277 	uint8_t *secbits;
2278 	tcp_seq ack, seq;
2279 	int wnd, wscale = 0;
2280 	union syncookie cookie;
2281 
2282 	/*
2283 	 * Pull information out of SYN-ACK/ACK and revert sequence number
2284 	 * advances.
2285 	 */
2286 	ack = th->th_ack - 1;
2287 	seq = th->th_seq - 1;
2288 
2289 	/*
2290 	 * Unpack the flags containing enough information to restore the
2291 	 * connection.
2292 	 */
2293 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2294 
2295 	/* Which of the two secrets to use. */
2296 	secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2297 
2298 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2299 
2300 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
2301 	if ((ack & ~0xff) != (hash & ~0xff))
2302 		return (NULL);
2303 
2304 	/* Fill in the syncache values. */
2305 	sc->sc_flags = 0;
2306 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2307 	sc->sc_ipopts = NULL;
2308 
2309 	sc->sc_irs = seq;
2310 	sc->sc_iss = ack;
2311 
2312 	switch (inc->inc_flags & INC_ISIPV6) {
2313 #ifdef INET
2314 	case 0:
2315 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2316 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2317 		break;
2318 #endif
2319 #ifdef INET6
2320 	case INC_ISIPV6:
2321 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2322 			sc->sc_flowlabel =
2323 			    htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2324 		break;
2325 #endif
2326 	}
2327 
2328 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2329 
2330 	/* We can simply recompute receive window scale we sent earlier. */
2331 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2332 		wscale++;
2333 
2334 	/* Only use wscale if it was enabled in the orignal SYN. */
2335 	if (cookie.flags.wscale_idx > 0) {
2336 		sc->sc_requested_r_scale = wscale;
2337 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2338 		sc->sc_flags |= SCF_WINSCALE;
2339 	}
2340 
2341 	wnd = lso->sol_sbrcv_hiwat;
2342 	wnd = imax(wnd, 0);
2343 	wnd = imin(wnd, TCP_MAXWIN);
2344 	sc->sc_wnd = wnd;
2345 
2346 	if (cookie.flags.sack_ok)
2347 		sc->sc_flags |= SCF_SACK;
2348 
2349 	if (to->to_flags & TOF_TS) {
2350 		sc->sc_flags |= SCF_TIMESTAMP;
2351 		sc->sc_tsreflect = to->to_tsval;
2352 		sc->sc_tsoff = tcp_new_ts_offset(inc);
2353 	}
2354 
2355 	if (to->to_flags & TOF_SIGNATURE)
2356 		sc->sc_flags |= SCF_SIGNATURE;
2357 
2358 	sc->sc_rxmits = 0;
2359 
2360 	sc->sc_port = port;
2361 
2362 	TCPSTAT_INC(tcps_sc_recvcookie);
2363 	return (sc);
2364 }
2365 
2366 #ifdef INVARIANTS
2367 static int
2368 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2369     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2370     struct socket *lso, uint16_t port)
2371 {
2372 	struct syncache scs, *scx;
2373 	char *s;
2374 
2375 	bzero(&scs, sizeof(scs));
2376 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port);
2377 
2378 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2379 		return (0);
2380 
2381 	if (scx != NULL) {
2382 		if (sc->sc_peer_mss != scx->sc_peer_mss)
2383 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2384 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2385 
2386 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2387 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2388 			    s, __func__, sc->sc_requested_r_scale,
2389 			    scx->sc_requested_r_scale);
2390 
2391 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2392 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2393 			    s, __func__, sc->sc_requested_s_scale,
2394 			    scx->sc_requested_s_scale);
2395 
2396 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2397 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2398 	}
2399 
2400 	if (s != NULL)
2401 		free(s, M_TCPLOG);
2402 	return (0);
2403 }
2404 #endif /* INVARIANTS */
2405 
2406 static void
2407 syncookie_reseed(void *arg)
2408 {
2409 	struct tcp_syncache *sc = arg;
2410 	uint8_t *secbits;
2411 	int secbit;
2412 
2413 	/*
2414 	 * Reseeding the secret doesn't have to be protected by a lock.
2415 	 * It only must be ensured that the new random values are visible
2416 	 * to all CPUs in a SMP environment.  The atomic with release
2417 	 * semantics ensures that.
2418 	 */
2419 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2420 	secbits = sc->secret.key[secbit];
2421 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2422 	atomic_add_rel_int(&sc->secret.oddeven, 1);
2423 
2424 	/* Reschedule ourself. */
2425 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2426 }
2427 
2428 /*
2429  * We have overflowed a bucket. Let's pause dealing with the syncache.
2430  * This function will increment the bucketoverflow statistics appropriately
2431  * (once per pause when pausing is enabled; otherwise, once per overflow).
2432  */
2433 static void
2434 syncache_pause(struct in_conninfo *inc)
2435 {
2436 	time_t delta;
2437 	const char *s;
2438 
2439 	/* XXX:
2440 	 * 2. Add sysctl read here so we don't get the benefit of this
2441 	 * change without the new sysctl.
2442 	 */
2443 
2444 	/*
2445 	 * Try an unlocked read. If we already know that another thread
2446 	 * has activated the feature, there is no need to proceed.
2447 	 */
2448 	if (V_tcp_syncache.paused)
2449 		return;
2450 
2451 	/* Are cookied enabled? If not, we can't pause. */
2452 	if (!V_tcp_syncookies) {
2453 		TCPSTAT_INC(tcps_sc_bucketoverflow);
2454 		return;
2455 	}
2456 
2457 	/*
2458 	 * We may be the first thread to find an overflow. Get the lock
2459 	 * and evaluate if we need to take action.
2460 	 */
2461 	mtx_lock(&V_tcp_syncache.pause_mtx);
2462 	if (V_tcp_syncache.paused) {
2463 		mtx_unlock(&V_tcp_syncache.pause_mtx);
2464 		return;
2465 	}
2466 
2467 	/* Activate protection. */
2468 	V_tcp_syncache.paused = true;
2469 	TCPSTAT_INC(tcps_sc_bucketoverflow);
2470 
2471 	/*
2472 	 * Determine the last backoff time. If we are seeing a re-newed
2473 	 * attack within that same time after last reactivating the syncache,
2474 	 * consider it an extension of the same attack.
2475 	 */
2476 	delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2477 	if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2478 		if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2479 			delta <<= 1;
2480 			V_tcp_syncache.pause_backoff++;
2481 		}
2482 	} else {
2483 		delta = TCP_SYNCACHE_PAUSE_TIME;
2484 		V_tcp_syncache.pause_backoff = 0;
2485 	}
2486 
2487 	/* Log a warning, including IP addresses, if able. */
2488 	if (inc != NULL)
2489 		s = tcp_log_addrs(inc, NULL, NULL, NULL);
2490 	else
2491 		s = (const char *)NULL;
2492 	log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2493 	    "the next %lld seconds%s%s%s\n", (long long)delta,
2494 	    (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2495 	    (s != NULL) ? ")" : "");
2496 	free(__DECONST(void *, s), M_TCPLOG);
2497 
2498 	/* Use the calculated delta to set a new pause time. */
2499 	V_tcp_syncache.pause_until = time_uptime + delta;
2500 	callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2501 	    &V_tcp_syncache);
2502 	mtx_unlock(&V_tcp_syncache.pause_mtx);
2503 }
2504 
2505 /* Evaluate whether we need to unpause. */
2506 static void
2507 syncache_unpause(void *arg)
2508 {
2509 	struct tcp_syncache *sc;
2510 	time_t delta;
2511 
2512 	sc = arg;
2513 	mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2514 	callout_deactivate(&sc->pause_co);
2515 
2516 	/*
2517 	 * Check to make sure we are not running early. If the pause
2518 	 * time has expired, then deactivate the protection.
2519 	 */
2520 	if ((delta = sc->pause_until - time_uptime) > 0)
2521 		callout_schedule(&sc->pause_co, delta * hz);
2522 	else
2523 		sc->paused = false;
2524 }
2525 
2526 /*
2527  * Exports the syncache entries to userland so that netstat can display
2528  * them alongside the other sockets.  This function is intended to be
2529  * called only from tcp_pcblist.
2530  *
2531  * Due to concurrency on an active system, the number of pcbs exported
2532  * may have no relation to max_pcbs.  max_pcbs merely indicates the
2533  * amount of space the caller allocated for this function to use.
2534  */
2535 int
2536 syncache_pcblist(struct sysctl_req *req)
2537 {
2538 	struct xtcpcb xt;
2539 	struct syncache *sc;
2540 	struct syncache_head *sch;
2541 	int error, i;
2542 
2543 	bzero(&xt, sizeof(xt));
2544 	xt.xt_len = sizeof(xt);
2545 	xt.t_state = TCPS_SYN_RECEIVED;
2546 	xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2547 	xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2548 	xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2549 	xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2550 
2551 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2552 		sch = &V_tcp_syncache.hashbase[i];
2553 		SCH_LOCK(sch);
2554 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2555 			if (sc->sc_cred != NULL &&
2556 			    cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2557 				continue;
2558 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2559 				xt.xt_inp.inp_vflag = INP_IPV6;
2560 			else
2561 				xt.xt_inp.inp_vflag = INP_IPV4;
2562 			xt.xt_encaps_port = sc->sc_port;
2563 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2564 			    sizeof (struct in_conninfo));
2565 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2566 			if (error) {
2567 				SCH_UNLOCK(sch);
2568 				return (0);
2569 			}
2570 		}
2571 		SCH_UNLOCK(sch);
2572 	}
2573 
2574 	return (0);
2575 }
2576