xref: /freebsd/sys/netinet/tcp_syncache.c (revision 4ac8f4067096a9d4a00e41cd53bd5c4fa295fd15)
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_OFFLOAD
94 #include <netinet/toecore.h>
95 #endif
96 #include <netinet/udp.h>
97 
98 #include <netipsec/ipsec_support.h>
99 
100 #include <machine/in_cksum.h>
101 
102 #include <security/mac/mac_framework.h>
103 
104 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1;
105 #define	V_tcp_syncookies		VNET(tcp_syncookies)
106 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
107     &VNET_NAME(tcp_syncookies), 0,
108     "Use TCP SYN cookies if the syncache overflows");
109 
110 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0;
111 #define	V_tcp_syncookiesonly		VNET(tcp_syncookiesonly)
112 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
113     &VNET_NAME(tcp_syncookiesonly), 0,
114     "Use only TCP SYN cookies");
115 
116 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1;
117 #define V_functions_inherit_listen_socket_stack \
118     VNET(functions_inherit_listen_socket_stack)
119 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack,
120     CTLFLAG_VNET | CTLFLAG_RW,
121     &VNET_NAME(functions_inherit_listen_socket_stack), 0,
122     "Inherit listen socket's stack");
123 
124 #ifdef TCP_OFFLOAD
125 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
126 #endif
127 
128 static void	 syncache_drop(struct syncache *, struct syncache_head *);
129 static void	 syncache_free(struct syncache *);
130 static void	 syncache_insert(struct syncache *, struct syncache_head *);
131 static int	 syncache_respond(struct syncache *, const struct mbuf *, int);
132 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
133 		    struct mbuf *m);
134 static void	 syncache_timeout(struct syncache *sc, struct syncache_head *sch,
135 		    int docallout);
136 static void	 syncache_timer(void *);
137 
138 static uint32_t	 syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
139 		    uint8_t *, uintptr_t);
140 static tcp_seq	 syncookie_generate(struct syncache_head *, struct syncache *);
141 static struct syncache
142 		*syncookie_lookup(struct in_conninfo *, struct syncache_head *,
143 		    struct syncache *, struct tcphdr *, struct tcpopt *,
144 		    struct socket *, uint16_t);
145 static void	syncache_pause(struct in_conninfo *);
146 static void	syncache_unpause(void *);
147 static void	 syncookie_reseed(void *);
148 #ifdef INVARIANTS
149 static int	 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
150 		    struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
151 		    struct socket *lso, uint16_t port);
152 #endif
153 
154 /*
155  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
156  * 3 retransmits corresponds to a timeout with default values of
157  * tcp_rexmit_initial * (             1 +
158  *                       tcp_backoff[1] +
159  *                       tcp_backoff[2] +
160  *                       tcp_backoff[3]) + 3 * tcp_rexmit_slop,
161  * 1000 ms * (1 + 2 + 4 + 8) +  3 * 200 ms = 15600 ms,
162  * the odds are that the user has given up attempting to connect by then.
163  */
164 #define SYNCACHE_MAXREXMTS		3
165 
166 /* Arbitrary values */
167 #define TCP_SYNCACHE_HASHSIZE		512
168 #define TCP_SYNCACHE_BUCKETLIMIT	30
169 
170 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
171 #define	V_tcp_syncache			VNET(tcp_syncache)
172 
173 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache,
174     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
175     "TCP SYN cache");
176 
177 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
178     &VNET_NAME(tcp_syncache.bucket_limit), 0,
179     "Per-bucket hash limit for syncache");
180 
181 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
182     &VNET_NAME(tcp_syncache.cache_limit), 0,
183     "Overall entry limit for syncache");
184 
185 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
186     &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
187 
188 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
189     &VNET_NAME(tcp_syncache.hashsize), 0,
190     "Size of TCP syncache hashtable");
191 
192 SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET |
193     CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0,
194     "All syncache(4) entries are visible, ignoring UID/GID, jail(2) "
195     "and mac(4) checks");
196 
197 static int
198 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
199 {
200 	int error;
201 	u_int new;
202 
203 	new = V_tcp_syncache.rexmt_limit;
204 	error = sysctl_handle_int(oidp, &new, 0, req);
205 	if ((error == 0) && (req->newptr != NULL)) {
206 		if (new > TCP_MAXRXTSHIFT)
207 			error = EINVAL;
208 		else
209 			V_tcp_syncache.rexmt_limit = new;
210 	}
211 	return (error);
212 }
213 
214 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
215     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
216     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
217     sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
218     "Limit on SYN/ACK retransmissions");
219 
220 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
221 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
222     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
223     "Send reset on socket allocation failure");
224 
225 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
226 
227 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
228 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
229 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
230 
231 /*
232  * Requires the syncache entry to be already removed from the bucket list.
233  */
234 static void
235 syncache_free(struct syncache *sc)
236 {
237 
238 	if (sc->sc_ipopts)
239 		(void) m_free(sc->sc_ipopts);
240 	if (sc->sc_cred)
241 		crfree(sc->sc_cred);
242 #ifdef MAC
243 	mac_syncache_destroy(&sc->sc_label);
244 #endif
245 
246 	uma_zfree(V_tcp_syncache.zone, sc);
247 }
248 
249 void
250 syncache_init(void)
251 {
252 	int i;
253 
254 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
255 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
256 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
257 	V_tcp_syncache.hash_secret = arc4random();
258 
259 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
260 	    &V_tcp_syncache.hashsize);
261 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
262 	    &V_tcp_syncache.bucket_limit);
263 	if (!powerof2(V_tcp_syncache.hashsize) ||
264 	    V_tcp_syncache.hashsize == 0) {
265 		printf("WARNING: syncache hash size is not a power of 2.\n");
266 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
267 	}
268 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
269 
270 	/* Set limits. */
271 	V_tcp_syncache.cache_limit =
272 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
273 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
274 	    &V_tcp_syncache.cache_limit);
275 
276 	/* Allocate the hash table. */
277 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
278 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
279 
280 #ifdef VIMAGE
281 	V_tcp_syncache.vnet = curvnet;
282 #endif
283 
284 	/* Initialize the hash buckets. */
285 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
286 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
287 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
288 			 NULL, MTX_DEF);
289 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
290 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
291 		V_tcp_syncache.hashbase[i].sch_length = 0;
292 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
293 		V_tcp_syncache.hashbase[i].sch_last_overflow =
294 		    -(SYNCOOKIE_LIFETIME + 1);
295 	}
296 
297 	/* Create the syncache entry zone. */
298 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
299 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
300 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
301 	    V_tcp_syncache.cache_limit);
302 
303 	/* Start the SYN cookie reseeder callout. */
304 	callout_init(&V_tcp_syncache.secret.reseed, 1);
305 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
306 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
307 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
308 	    syncookie_reseed, &V_tcp_syncache);
309 
310 	/* Initialize the pause machinery. */
311 	mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF);
312 	callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx,
313 	    0);
314 	V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME;
315 	V_tcp_syncache.pause_backoff = 0;
316 	V_tcp_syncache.paused = false;
317 }
318 
319 #ifdef VIMAGE
320 void
321 syncache_destroy(void)
322 {
323 	struct syncache_head *sch;
324 	struct syncache *sc, *nsc;
325 	int i;
326 
327 	/*
328 	 * Stop the re-seed timer before freeing resources.  No need to
329 	 * possibly schedule it another time.
330 	 */
331 	callout_drain(&V_tcp_syncache.secret.reseed);
332 
333 	/* Stop the SYN cache pause callout. */
334 	mtx_lock(&V_tcp_syncache.pause_mtx);
335 	if (callout_stop(&V_tcp_syncache.pause_co) == 0) {
336 		mtx_unlock(&V_tcp_syncache.pause_mtx);
337 		callout_drain(&V_tcp_syncache.pause_co);
338 	} else
339 		mtx_unlock(&V_tcp_syncache.pause_mtx);
340 
341 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
342 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
343 		sch = &V_tcp_syncache.hashbase[i];
344 		callout_drain(&sch->sch_timer);
345 
346 		SCH_LOCK(sch);
347 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
348 			syncache_drop(sc, sch);
349 		SCH_UNLOCK(sch);
350 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
351 		    ("%s: sch->sch_bucket not empty", __func__));
352 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
353 		    __func__, sch->sch_length));
354 		mtx_destroy(&sch->sch_mtx);
355 	}
356 
357 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
358 	    ("%s: cache_count not 0", __func__));
359 
360 	/* Free the allocated global resources. */
361 	uma_zdestroy(V_tcp_syncache.zone);
362 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
363 	mtx_destroy(&V_tcp_syncache.pause_mtx);
364 }
365 #endif
366 
367 /*
368  * Inserts a syncache entry into the specified bucket row.
369  * Locks and unlocks the syncache_head autonomously.
370  */
371 static void
372 syncache_insert(struct syncache *sc, struct syncache_head *sch)
373 {
374 	struct syncache *sc2;
375 
376 	SCH_LOCK(sch);
377 
378 	/*
379 	 * Make sure that we don't overflow the per-bucket limit.
380 	 * If the bucket is full, toss the oldest element.
381 	 */
382 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
383 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
384 			("sch->sch_length incorrect"));
385 		syncache_pause(&sc->sc_inc);
386 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
387 		sch->sch_last_overflow = time_uptime;
388 		syncache_drop(sc2, sch);
389 	}
390 
391 	/* Put it into the bucket. */
392 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
393 	sch->sch_length++;
394 
395 #ifdef TCP_OFFLOAD
396 	if (ADDED_BY_TOE(sc)) {
397 		struct toedev *tod = sc->sc_tod;
398 
399 		tod->tod_syncache_added(tod, sc->sc_todctx);
400 	}
401 #endif
402 
403 	/* Reinitialize the bucket row's timer. */
404 	if (sch->sch_length == 1)
405 		sch->sch_nextc = ticks + INT_MAX;
406 	syncache_timeout(sc, sch, 1);
407 
408 	SCH_UNLOCK(sch);
409 
410 	TCPSTATES_INC(TCPS_SYN_RECEIVED);
411 	TCPSTAT_INC(tcps_sc_added);
412 }
413 
414 /*
415  * Remove and free entry from syncache bucket row.
416  * Expects locked syncache head.
417  */
418 static void
419 syncache_drop(struct syncache *sc, struct syncache_head *sch)
420 {
421 
422 	SCH_LOCK_ASSERT(sch);
423 
424 	TCPSTATES_DEC(TCPS_SYN_RECEIVED);
425 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
426 	sch->sch_length--;
427 
428 #ifdef TCP_OFFLOAD
429 	if (ADDED_BY_TOE(sc)) {
430 		struct toedev *tod = sc->sc_tod;
431 
432 		tod->tod_syncache_removed(tod, sc->sc_todctx);
433 	}
434 #endif
435 
436 	syncache_free(sc);
437 }
438 
439 /*
440  * Engage/reengage time on bucket row.
441  */
442 static void
443 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
444 {
445 	int rexmt;
446 
447 	if (sc->sc_rxmits == 0)
448 		rexmt = tcp_rexmit_initial;
449 	else
450 		TCPT_RANGESET(rexmt,
451 		    tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
452 		    tcp_rexmit_min, TCPTV_REXMTMAX);
453 	sc->sc_rxttime = ticks + rexmt;
454 	sc->sc_rxmits++;
455 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
456 		sch->sch_nextc = sc->sc_rxttime;
457 		if (docallout)
458 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
459 			    syncache_timer, (void *)sch);
460 	}
461 }
462 
463 /*
464  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
465  * If we have retransmitted an entry the maximum number of times, expire it.
466  * One separate timer for each bucket row.
467  */
468 static void
469 syncache_timer(void *xsch)
470 {
471 	struct syncache_head *sch = (struct syncache_head *)xsch;
472 	struct syncache *sc, *nsc;
473 	struct epoch_tracker et;
474 	int tick = ticks;
475 	char *s;
476 	bool paused;
477 
478 	CURVNET_SET(sch->sch_sc->vnet);
479 
480 	/* NB: syncache_head has already been locked by the callout. */
481 	SCH_LOCK_ASSERT(sch);
482 
483 	/*
484 	 * In the following cycle we may remove some entries and/or
485 	 * advance some timeouts, so re-initialize the bucket timer.
486 	 */
487 	sch->sch_nextc = tick + INT_MAX;
488 
489 	/*
490 	 * If we have paused processing, unconditionally remove
491 	 * all syncache entries.
492 	 */
493 	mtx_lock(&V_tcp_syncache.pause_mtx);
494 	paused = V_tcp_syncache.paused;
495 	mtx_unlock(&V_tcp_syncache.pause_mtx);
496 
497 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
498 		if (paused) {
499 			syncache_drop(sc, sch);
500 			continue;
501 		}
502 		/*
503 		 * We do not check if the listen socket still exists
504 		 * and accept the case where the listen socket may be
505 		 * gone by the time we resend the SYN/ACK.  We do
506 		 * not expect this to happens often. If it does,
507 		 * then the RST will be sent by the time the remote
508 		 * host does the SYN/ACK->ACK.
509 		 */
510 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
511 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
512 				sch->sch_nextc = sc->sc_rxttime;
513 			continue;
514 		}
515 		if (sc->sc_rxmits > V_tcp_ecn_maxretries) {
516 			sc->sc_flags &= ~SCF_ECN_MASK;
517 		}
518 		if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
519 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
520 				log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
521 				    "giving up and removing syncache entry\n",
522 				    s, __func__);
523 				free(s, M_TCPLOG);
524 			}
525 			syncache_drop(sc, sch);
526 			TCPSTAT_INC(tcps_sc_stale);
527 			continue;
528 		}
529 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
530 			log(LOG_DEBUG, "%s; %s: Response timeout, "
531 			    "retransmitting (%u) SYN|ACK\n",
532 			    s, __func__, sc->sc_rxmits);
533 			free(s, M_TCPLOG);
534 		}
535 
536 		NET_EPOCH_ENTER(et);
537 		syncache_respond(sc, NULL, TH_SYN|TH_ACK);
538 		NET_EPOCH_EXIT(et);
539 		TCPSTAT_INC(tcps_sc_retransmitted);
540 		syncache_timeout(sc, sch, 0);
541 	}
542 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
543 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
544 			syncache_timer, (void *)(sch));
545 	CURVNET_RESTORE();
546 }
547 
548 /*
549  * Returns true if the system is only using cookies at the moment.
550  * This could be due to a sysadmin decision to only use cookies, or it
551  * could be due to the system detecting an attack.
552  */
553 static inline bool
554 syncache_cookiesonly(void)
555 {
556 
557 	return (V_tcp_syncookies && (V_tcp_syncache.paused ||
558 	    V_tcp_syncookiesonly));
559 }
560 
561 /*
562  * Find the hash bucket for the given connection.
563  */
564 static struct syncache_head *
565 syncache_hashbucket(struct in_conninfo *inc)
566 {
567 	uint32_t hash;
568 
569 	/*
570 	 * The hash is built on foreign port + local port + foreign address.
571 	 * We rely on the fact that struct in_conninfo starts with 16 bits
572 	 * of foreign port, then 16 bits of local port then followed by 128
573 	 * bits of foreign address.  In case of IPv4 address, the first 3
574 	 * 32-bit words of the address always are zeroes.
575 	 */
576 	hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
577 	    V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
578 
579 	return (&V_tcp_syncache.hashbase[hash]);
580 }
581 
582 /*
583  * Find an entry in the syncache.
584  * Returns always with locked syncache_head plus a matching entry or NULL.
585  */
586 static struct syncache *
587 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
588 {
589 	struct syncache *sc;
590 	struct syncache_head *sch;
591 
592 	*schp = sch = syncache_hashbucket(inc);
593 	SCH_LOCK(sch);
594 
595 	/* Circle through bucket row to find matching entry. */
596 	TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
597 		if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
598 		    sizeof(struct in_endpoints)) == 0)
599 			break;
600 
601 	return (sc);	/* Always returns with locked sch. */
602 }
603 
604 /*
605  * This function is called when we get a RST for a
606  * non-existent connection, so that we can see if the
607  * connection is in the syn cache.  If it is, zap it.
608  * If required send a challenge ACK.
609  */
610 void
611 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m,
612     uint16_t port)
613 {
614 	struct syncache *sc;
615 	struct syncache_head *sch;
616 	char *s = NULL;
617 
618 	if (syncache_cookiesonly())
619 		return;
620 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
621 	SCH_LOCK_ASSERT(sch);
622 
623 	/*
624 	 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
625 	 * See RFC 793 page 65, section SEGMENT ARRIVES.
626 	 */
627 	if (tcp_get_flags(th) & (TH_ACK|TH_SYN|TH_FIN)) {
628 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
629 			log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
630 			    "FIN flag set, segment ignored\n", s, __func__);
631 		TCPSTAT_INC(tcps_badrst);
632 		goto done;
633 	}
634 
635 	/*
636 	 * No corresponding connection was found in syncache.
637 	 * If syncookies are enabled and possibly exclusively
638 	 * used, or we are under memory pressure, a valid RST
639 	 * may not find a syncache entry.  In that case we're
640 	 * done and no SYN|ACK retransmissions will happen.
641 	 * Otherwise the RST was misdirected or spoofed.
642 	 */
643 	if (sc == NULL) {
644 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
645 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
646 			    "syncache entry (possibly syncookie only), "
647 			    "segment ignored\n", s, __func__);
648 		TCPSTAT_INC(tcps_badrst);
649 		goto done;
650 	}
651 
652 	/* The remote UDP encaps port does not match. */
653 	if (sc->sc_port != port) {
654 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
655 			log(LOG_DEBUG, "%s; %s: Spurious RST with matching "
656 			    "syncache entry but non-matching UDP encaps port, "
657 			    "segment ignored\n", s, __func__);
658 		TCPSTAT_INC(tcps_badrst);
659 		goto done;
660 	}
661 
662 	/*
663 	 * If the RST bit is set, check the sequence number to see
664 	 * if this is a valid reset segment.
665 	 *
666 	 * RFC 793 page 37:
667 	 *   In all states except SYN-SENT, all reset (RST) segments
668 	 *   are validated by checking their SEQ-fields.  A reset is
669 	 *   valid if its sequence number is in the window.
670 	 *
671 	 * RFC 793 page 69:
672 	 *   There are four cases for the acceptability test for an incoming
673 	 *   segment:
674 	 *
675 	 * Segment Receive  Test
676 	 * Length  Window
677 	 * ------- -------  -------------------------------------------
678 	 *    0       0     SEG.SEQ = RCV.NXT
679 	 *    0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
680 	 *   >0       0     not acceptable
681 	 *   >0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
682 	 *               or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
683 	 *
684 	 * Note that when receiving a SYN segment in the LISTEN state,
685 	 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
686 	 * described in RFC 793, page 66.
687 	 */
688 	if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
689 	    SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
690 	    (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
691 		if (V_tcp_insecure_rst ||
692 		    th->th_seq == sc->sc_irs + 1) {
693 			syncache_drop(sc, sch);
694 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
695 				log(LOG_DEBUG,
696 				    "%s; %s: Our SYN|ACK was rejected, "
697 				    "connection attempt aborted by remote "
698 				    "endpoint\n",
699 				    s, __func__);
700 			TCPSTAT_INC(tcps_sc_reset);
701 		} else {
702 			TCPSTAT_INC(tcps_badrst);
703 			/* Send challenge ACK. */
704 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
705 				log(LOG_DEBUG, "%s; %s: RST with invalid "
706 				    " SEQ %u != NXT %u (+WND %u), "
707 				    "sending challenge ACK\n",
708 				    s, __func__,
709 				    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
710 			syncache_respond(sc, m, TH_ACK);
711 		}
712 	} else {
713 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
714 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
715 			    "NXT %u (+WND %u), segment ignored\n",
716 			    s, __func__,
717 			    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
718 		TCPSTAT_INC(tcps_badrst);
719 	}
720 
721 done:
722 	if (s != NULL)
723 		free(s, M_TCPLOG);
724 	SCH_UNLOCK(sch);
725 }
726 
727 void
728 syncache_badack(struct in_conninfo *inc, uint16_t port)
729 {
730 	struct syncache *sc;
731 	struct syncache_head *sch;
732 
733 	if (syncache_cookiesonly())
734 		return;
735 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
736 	SCH_LOCK_ASSERT(sch);
737 	if ((sc != NULL) && (sc->sc_port == port)) {
738 		syncache_drop(sc, sch);
739 		TCPSTAT_INC(tcps_sc_badack);
740 	}
741 	SCH_UNLOCK(sch);
742 }
743 
744 void
745 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port)
746 {
747 	struct syncache *sc;
748 	struct syncache_head *sch;
749 
750 	if (syncache_cookiesonly())
751 		return;
752 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
753 	SCH_LOCK_ASSERT(sch);
754 	if (sc == NULL)
755 		goto done;
756 
757 	/* If the port != sc_port, then it's a bogus ICMP msg */
758 	if (port != sc->sc_port)
759 		goto done;
760 
761 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
762 	if (ntohl(th_seq) != sc->sc_iss)
763 		goto done;
764 
765 	/*
766 	 * If we've rertransmitted 3 times and this is our second error,
767 	 * we remove the entry.  Otherwise, we allow it to continue on.
768 	 * This prevents us from incorrectly nuking an entry during a
769 	 * spurious network outage.
770 	 *
771 	 * See tcp_notify().
772 	 */
773 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
774 		sc->sc_flags |= SCF_UNREACH;
775 		goto done;
776 	}
777 	syncache_drop(sc, sch);
778 	TCPSTAT_INC(tcps_sc_unreach);
779 done:
780 	SCH_UNLOCK(sch);
781 }
782 
783 /*
784  * Build a new TCP socket structure from a syncache entry.
785  *
786  * On success return the newly created socket with its underlying inp locked.
787  */
788 static struct socket *
789 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
790 {
791 	struct tcp_function_block *blk;
792 	struct inpcb *inp = NULL;
793 	struct socket *so;
794 	struct tcpcb *tp;
795 	int error;
796 	char *s;
797 
798 	NET_EPOCH_ASSERT();
799 
800 	/*
801 	 * Ok, create the full blown connection, and set things up
802 	 * as they would have been set up if we had created the
803 	 * connection when the SYN arrived.
804 	 */
805 	if ((so = solisten_clone(lso)) == NULL)
806 		goto allocfail;
807 #ifdef MAC
808 	mac_socketpeer_set_from_mbuf(m, so);
809 #endif
810 	error = in_pcballoc(so, &V_tcbinfo);
811 	if (error) {
812 		sodealloc(so);
813 		goto allocfail;
814 	}
815 	inp = sotoinpcb(so);
816 	if ((tp = tcp_newtcpcb(inp)) == NULL) {
817 		in_pcbdetach(inp);
818 		in_pcbfree(inp);
819 		sodealloc(so);
820 		goto allocfail;
821 	}
822 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
823 #ifdef INET6
824 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
825 		inp->inp_vflag &= ~INP_IPV4;
826 		inp->inp_vflag |= INP_IPV6;
827 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
828 	} else {
829 		inp->inp_vflag &= ~INP_IPV6;
830 		inp->inp_vflag |= INP_IPV4;
831 #endif
832 		inp->inp_ip_ttl = sc->sc_ip_ttl;
833 		inp->inp_ip_tos = sc->sc_ip_tos;
834 		inp->inp_laddr = sc->sc_inc.inc_laddr;
835 #ifdef INET6
836 	}
837 #endif
838 
839 	/*
840 	 * If there's an mbuf and it has a flowid, then let's initialise the
841 	 * inp with that particular flowid.
842 	 */
843 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
844 		inp->inp_flowid = m->m_pkthdr.flowid;
845 		inp->inp_flowtype = M_HASHTYPE_GET(m);
846 #ifdef NUMA
847 		inp->inp_numa_domain = m->m_pkthdr.numa_domain;
848 #endif
849 	}
850 
851 	inp->inp_lport = sc->sc_inc.inc_lport;
852 #ifdef INET6
853 	if (inp->inp_vflag & INP_IPV6PROTO) {
854 		struct inpcb *oinp = sotoinpcb(lso);
855 
856 		/*
857 		 * Inherit socket options from the listening socket.
858 		 * Note that in6p_inputopts are not (and should not be)
859 		 * copied, since it stores previously received options and is
860 		 * used to detect if each new option is different than the
861 		 * previous one and hence should be passed to a user.
862 		 * If we copied in6p_inputopts, a user would not be able to
863 		 * receive options just after calling the accept system call.
864 		 */
865 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
866 		if (oinp->in6p_outputopts)
867 			inp->in6p_outputopts =
868 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
869 		inp->in6p_hops = oinp->in6p_hops;
870 	}
871 
872 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
873 		struct in6_addr laddr6;
874 		struct sockaddr_in6 sin6;
875 
876 		sin6.sin6_family = AF_INET6;
877 		sin6.sin6_len = sizeof(sin6);
878 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
879 		sin6.sin6_port = sc->sc_inc.inc_fport;
880 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
881 		laddr6 = inp->in6p_laddr;
882 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
883 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
884 		INP_HASH_WLOCK(&V_tcbinfo);
885 		error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
886 		    thread0.td_ucred, m, false);
887 		INP_HASH_WUNLOCK(&V_tcbinfo);
888 		if (error != 0) {
889 			inp->in6p_laddr = laddr6;
890 			goto abort;
891 		}
892 		/* Override flowlabel from in6_pcbconnect. */
893 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
894 		inp->inp_flow |= sc->sc_flowlabel;
895 	}
896 #endif /* INET6 */
897 #if defined(INET) && defined(INET6)
898 	else
899 #endif
900 #ifdef INET
901 	{
902 		struct in_addr laddr;
903 		struct sockaddr_in sin;
904 
905 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
906 
907 		if (inp->inp_options == NULL) {
908 			inp->inp_options = sc->sc_ipopts;
909 			sc->sc_ipopts = NULL;
910 		}
911 
912 		sin.sin_family = AF_INET;
913 		sin.sin_len = sizeof(sin);
914 		sin.sin_addr = sc->sc_inc.inc_faddr;
915 		sin.sin_port = sc->sc_inc.inc_fport;
916 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
917 		laddr = inp->inp_laddr;
918 		if (inp->inp_laddr.s_addr == INADDR_ANY)
919 			inp->inp_laddr = sc->sc_inc.inc_laddr;
920 		INP_HASH_WLOCK(&V_tcbinfo);
921 		error = in_pcbconnect(inp, (struct sockaddr *)&sin,
922 		    thread0.td_ucred, false);
923 		INP_HASH_WUNLOCK(&V_tcbinfo);
924 		if (error != 0) {
925 			inp->inp_laddr = laddr;
926 			goto abort;
927 		}
928 	}
929 #endif /* INET */
930 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
931 	/* Copy old policy into new socket's. */
932 	if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
933 		printf("syncache_socket: could not copy policy\n");
934 #endif
935 	tp->t_state = TCPS_SYN_RECEIVED;
936 	tp->iss = sc->sc_iss;
937 	tp->irs = sc->sc_irs;
938 	tp->t_port = sc->sc_port;
939 	tcp_rcvseqinit(tp);
940 	tcp_sendseqinit(tp);
941 	blk = sototcpcb(lso)->t_fb;
942 	if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
943 		/*
944 		 * Our parents t_fb was not the default,
945 		 * we need to release our ref on tp->t_fb and
946 		 * pickup one on the new entry.
947 		 */
948 		struct tcp_function_block *rblk;
949 
950 		rblk = find_and_ref_tcp_fb(blk);
951 		KASSERT(rblk != NULL,
952 		    ("cannot find blk %p out of syncache?", blk));
953 		if (tp->t_fb->tfb_tcp_fb_fini)
954 			(*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
955 		refcount_release(&tp->t_fb->tfb_refcnt);
956 		tp->t_fb = rblk;
957 		/*
958 		 * XXXrrs this is quite dangerous, it is possible
959 		 * for the new function to fail to init. We also
960 		 * are not asking if the handoff_is_ok though at
961 		 * the very start thats probalbly ok.
962 		 */
963 		if (tp->t_fb->tfb_tcp_fb_init) {
964 			(*tp->t_fb->tfb_tcp_fb_init)(tp);
965 		}
966 	}
967 	tp->snd_wl1 = sc->sc_irs;
968 	tp->snd_max = tp->iss + 1;
969 	tp->snd_nxt = tp->iss + 1;
970 	tp->rcv_up = sc->sc_irs + 1;
971 	tp->rcv_wnd = sc->sc_wnd;
972 	tp->rcv_adv += tp->rcv_wnd;
973 	tp->last_ack_sent = tp->rcv_nxt;
974 
975 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
976 	if (sc->sc_flags & SCF_NOOPT)
977 		tp->t_flags |= TF_NOOPT;
978 	else {
979 		if (sc->sc_flags & SCF_WINSCALE) {
980 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
981 			tp->snd_scale = sc->sc_requested_s_scale;
982 			tp->request_r_scale = sc->sc_requested_r_scale;
983 		}
984 		if (sc->sc_flags & SCF_TIMESTAMP) {
985 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
986 			tp->ts_recent = sc->sc_tsreflect;
987 			tp->ts_recent_age = tcp_ts_getticks();
988 			tp->ts_offset = sc->sc_tsoff;
989 		}
990 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
991 		if (sc->sc_flags & SCF_SIGNATURE)
992 			tp->t_flags |= TF_SIGNATURE;
993 #endif
994 		if (sc->sc_flags & SCF_SACK)
995 			tp->t_flags |= TF_SACK_PERMIT;
996 	}
997 
998 	tcp_ecn_syncache_socket(tp, sc);
999 
1000 	/*
1001 	 * Set up MSS and get cached values from tcp_hostcache.
1002 	 * This might overwrite some of the defaults we just set.
1003 	 */
1004 	tcp_mss(tp, sc->sc_peer_mss);
1005 
1006 	/*
1007 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
1008 	 * limited to one segment in cc_conn_init().
1009 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
1010 	 */
1011 	if (sc->sc_rxmits > 1)
1012 		tp->snd_cwnd = 1;
1013 
1014 #ifdef TCP_OFFLOAD
1015 	/*
1016 	 * Allow a TOE driver to install its hooks.  Note that we hold the
1017 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
1018 	 * new connection before the TOE driver has done its thing.
1019 	 */
1020 	if (ADDED_BY_TOE(sc)) {
1021 		struct toedev *tod = sc->sc_tod;
1022 
1023 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
1024 	}
1025 #endif
1026 	/*
1027 	 * Copy and activate timers.
1028 	 */
1029 	tp->t_maxunacktime = sototcpcb(lso)->t_maxunacktime;
1030 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
1031 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
1032 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
1033 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1034 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1035 
1036 	TCPSTAT_INC(tcps_accepts);
1037 	TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, TCPS_LISTEN);
1038 
1039 	if (!solisten_enqueue(so, SS_ISCONNECTED))
1040 		tp->t_flags |= TF_SONOTCONN;
1041 
1042 	return (so);
1043 
1044 allocfail:
1045 	/*
1046 	 * Drop the connection; we will either send a RST or have the peer
1047 	 * retransmit its SYN again after its RTO and try again.
1048 	 */
1049 	if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1050 		log(LOG_DEBUG, "%s; %s: Socket create failed "
1051 		    "due to limits or memory shortage\n",
1052 		    s, __func__);
1053 		free(s, M_TCPLOG);
1054 	}
1055 	TCPSTAT_INC(tcps_listendrop);
1056 	return (NULL);
1057 
1058 abort:
1059 	in_pcbdetach(inp);
1060 	in_pcbfree(inp);
1061 	sodealloc(so);
1062 	if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1063 		log(LOG_DEBUG, "%s; %s: in%s_pcbconnect failed with error %i\n",
1064 		    s, __func__, (sc->sc_inc.inc_flags & INC_ISIPV6) ? "6" : "",
1065 		    error);
1066 		free(s, M_TCPLOG);
1067 	}
1068 	TCPSTAT_INC(tcps_listendrop);
1069 	return (NULL);
1070 }
1071 
1072 /*
1073  * This function gets called when we receive an ACK for a
1074  * socket in the LISTEN state.  We look up the connection
1075  * in the syncache, and if its there, we pull it out of
1076  * the cache and turn it into a full-blown connection in
1077  * the SYN-RECEIVED state.
1078  *
1079  * On syncache_socket() success the newly created socket
1080  * has its underlying inp locked.
1081  */
1082 int
1083 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1084     struct socket **lsop, struct mbuf *m, uint16_t port)
1085 {
1086 	struct syncache *sc;
1087 	struct syncache_head *sch;
1088 	struct syncache scs;
1089 	char *s;
1090 	bool locked;
1091 
1092 	NET_EPOCH_ASSERT();
1093 	KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1094 	    ("%s: can handle only ACK", __func__));
1095 
1096 	if (syncache_cookiesonly()) {
1097 		sc = NULL;
1098 		sch = syncache_hashbucket(inc);
1099 		locked = false;
1100 	} else {
1101 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1102 		locked = true;
1103 		SCH_LOCK_ASSERT(sch);
1104 	}
1105 
1106 #ifdef INVARIANTS
1107 	/*
1108 	 * Test code for syncookies comparing the syncache stored
1109 	 * values with the reconstructed values from the cookie.
1110 	 */
1111 	if (sc != NULL)
1112 		syncookie_cmp(inc, sch, sc, th, to, *lsop, port);
1113 #endif
1114 
1115 	if (sc == NULL) {
1116 		/*
1117 		 * There is no syncache entry, so see if this ACK is
1118 		 * a returning syncookie.  To do this, first:
1119 		 *  A. Check if syncookies are used in case of syncache
1120 		 *     overflows
1121 		 *  B. See if this socket has had a syncache entry dropped in
1122 		 *     the recent past. We don't want to accept a bogus
1123 		 *     syncookie if we've never received a SYN or accept it
1124 		 *     twice.
1125 		 *  C. check that the syncookie is valid.  If it is, then
1126 		 *     cobble up a fake syncache entry, and return.
1127 		 */
1128 		if (locked && !V_tcp_syncookies) {
1129 			SCH_UNLOCK(sch);
1130 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1131 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1132 				    "segment rejected (syncookies disabled)\n",
1133 				    s, __func__);
1134 			goto failed;
1135 		}
1136 		if (locked && !V_tcp_syncookiesonly &&
1137 		    sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1138 			SCH_UNLOCK(sch);
1139 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1140 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1141 				    "segment rejected (no syncache entry)\n",
1142 				    s, __func__);
1143 			goto failed;
1144 		}
1145 		bzero(&scs, sizeof(scs));
1146 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port);
1147 		if (locked)
1148 			SCH_UNLOCK(sch);
1149 		if (sc == NULL) {
1150 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1151 				log(LOG_DEBUG, "%s; %s: Segment failed "
1152 				    "SYNCOOKIE authentication, segment rejected "
1153 				    "(probably spoofed)\n", s, __func__);
1154 			goto failed;
1155 		}
1156 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1157 		/* If received ACK has MD5 signature, check it. */
1158 		if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1159 		    (!TCPMD5_ENABLED() ||
1160 		    TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1161 			/* Drop the ACK. */
1162 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1163 				log(LOG_DEBUG, "%s; %s: Segment rejected, "
1164 				    "MD5 signature doesn't match.\n",
1165 				    s, __func__);
1166 				free(s, M_TCPLOG);
1167 			}
1168 			TCPSTAT_INC(tcps_sig_err_sigopt);
1169 			return (-1); /* Do not send RST */
1170 		}
1171 #endif /* TCP_SIGNATURE */
1172 		TCPSTATES_INC(TCPS_SYN_RECEIVED);
1173 	} else {
1174 		if (sc->sc_port != port) {
1175 			SCH_UNLOCK(sch);
1176 			return (0);
1177 		}
1178 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1179 		/*
1180 		 * If listening socket requested TCP digests, check that
1181 		 * received ACK has signature and it is correct.
1182 		 * If not, drop the ACK and leave sc entry in th cache,
1183 		 * because SYN was received with correct signature.
1184 		 */
1185 		if (sc->sc_flags & SCF_SIGNATURE) {
1186 			if ((to->to_flags & TOF_SIGNATURE) == 0) {
1187 				/* No signature */
1188 				TCPSTAT_INC(tcps_sig_err_nosigopt);
1189 				SCH_UNLOCK(sch);
1190 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1191 					log(LOG_DEBUG, "%s; %s: Segment "
1192 					    "rejected, MD5 signature wasn't "
1193 					    "provided.\n", s, __func__);
1194 					free(s, M_TCPLOG);
1195 				}
1196 				return (-1); /* Do not send RST */
1197 			}
1198 			if (!TCPMD5_ENABLED() ||
1199 			    TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1200 				/* Doesn't match or no SA */
1201 				SCH_UNLOCK(sch);
1202 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1203 					log(LOG_DEBUG, "%s; %s: Segment "
1204 					    "rejected, MD5 signature doesn't "
1205 					    "match.\n", s, __func__);
1206 					free(s, M_TCPLOG);
1207 				}
1208 				return (-1); /* Do not send RST */
1209 			}
1210 		}
1211 #endif /* TCP_SIGNATURE */
1212 
1213 		/*
1214 		 * RFC 7323 PAWS: If we have a timestamp on this segment and
1215 		 * it's less than ts_recent, drop it.
1216 		 * XXXMT: RFC 7323 also requires to send an ACK.
1217 		 *        In tcp_input.c this is only done for TCP segments
1218 		 *        with user data, so be consistent here and just drop
1219 		 *        the segment.
1220 		 */
1221 		if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1222 		    TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1223 			SCH_UNLOCK(sch);
1224 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1225 				log(LOG_DEBUG,
1226 				    "%s; %s: SEG.TSval %u < TS.Recent %u, "
1227 				    "segment dropped\n", s, __func__,
1228 				    to->to_tsval, sc->sc_tsreflect);
1229 				free(s, M_TCPLOG);
1230 			}
1231 			return (-1);  /* Do not send RST */
1232 		}
1233 
1234 		/*
1235 		 * If timestamps were not negotiated during SYN/ACK and a
1236 		 * segment with a timestamp is received, ignore the
1237 		 * timestamp and process the packet normally.
1238 		 * See section 3.2 of RFC 7323.
1239 		 */
1240 		if (!(sc->sc_flags & SCF_TIMESTAMP) &&
1241 		    (to->to_flags & TOF_TS)) {
1242 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1243 				log(LOG_DEBUG, "%s; %s: Timestamp not "
1244 				    "expected, segment processed normally\n",
1245 				    s, __func__);
1246 				free(s, M_TCPLOG);
1247 				s = NULL;
1248 			}
1249 		}
1250 
1251 		/*
1252 		 * If timestamps were negotiated during SYN/ACK and a
1253 		 * segment without a timestamp is received, silently drop
1254 		 * the segment, unless the missing timestamps are tolerated.
1255 		 * See section 3.2 of RFC 7323.
1256 		 */
1257 		if ((sc->sc_flags & SCF_TIMESTAMP) &&
1258 		    !(to->to_flags & TOF_TS)) {
1259 			if (V_tcp_tolerate_missing_ts) {
1260 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1261 					log(LOG_DEBUG,
1262 					    "%s; %s: Timestamp missing, "
1263 					    "segment processed normally\n",
1264 					    s, __func__);
1265 					free(s, M_TCPLOG);
1266 				}
1267 			} else {
1268 				SCH_UNLOCK(sch);
1269 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1270 					log(LOG_DEBUG,
1271 					    "%s; %s: Timestamp missing, "
1272 					    "segment silently dropped\n",
1273 					    s, __func__);
1274 					free(s, M_TCPLOG);
1275 				}
1276 				return (-1);  /* Do not send RST */
1277 			}
1278 		}
1279 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1280 		sch->sch_length--;
1281 #ifdef TCP_OFFLOAD
1282 		if (ADDED_BY_TOE(sc)) {
1283 			struct toedev *tod = sc->sc_tod;
1284 
1285 			tod->tod_syncache_removed(tod, sc->sc_todctx);
1286 		}
1287 #endif
1288 		SCH_UNLOCK(sch);
1289 	}
1290 
1291 	/*
1292 	 * Segment validation:
1293 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1294 	 */
1295 	if (th->th_ack != sc->sc_iss + 1) {
1296 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1297 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1298 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1299 		goto failed;
1300 	}
1301 
1302 	/*
1303 	 * The SEQ must fall in the window starting at the received
1304 	 * initial receive sequence number + 1 (the SYN).
1305 	 */
1306 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1307 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1308 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1309 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1310 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1311 		goto failed;
1312 	}
1313 
1314 	*lsop = syncache_socket(sc, *lsop, m);
1315 
1316 	if (__predict_false(*lsop == NULL)) {
1317 		TCPSTAT_INC(tcps_sc_aborted);
1318 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1319 	} else
1320 		TCPSTAT_INC(tcps_sc_completed);
1321 
1322 /* how do we find the inp for the new socket? */
1323 	if (sc != &scs)
1324 		syncache_free(sc);
1325 	return (1);
1326 failed:
1327 	if (sc != NULL) {
1328 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1329 		if (sc != &scs)
1330 			syncache_free(sc);
1331 	}
1332 	if (s != NULL)
1333 		free(s, M_TCPLOG);
1334 	*lsop = NULL;
1335 	return (0);
1336 }
1337 
1338 static struct socket *
1339 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m,
1340     uint64_t response_cookie)
1341 {
1342 	struct inpcb *inp;
1343 	struct tcpcb *tp;
1344 	unsigned int *pending_counter;
1345 	struct socket *so;
1346 
1347 	NET_EPOCH_ASSERT();
1348 
1349 	pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending;
1350 	so = syncache_socket(sc, lso, m);
1351 	if (so == NULL) {
1352 		TCPSTAT_INC(tcps_sc_aborted);
1353 		atomic_subtract_int(pending_counter, 1);
1354 	} else {
1355 		soisconnected(so);
1356 		inp = sotoinpcb(so);
1357 		tp = intotcpcb(inp);
1358 		tp->t_flags |= TF_FASTOPEN;
1359 		tp->t_tfo_cookie.server = response_cookie;
1360 		tp->snd_max = tp->iss;
1361 		tp->snd_nxt = tp->iss;
1362 		tp->t_tfo_pending = pending_counter;
1363 		TCPSTATES_INC(TCPS_SYN_RECEIVED);
1364 		TCPSTAT_INC(tcps_sc_completed);
1365 	}
1366 
1367 	return (so);
1368 }
1369 
1370 /*
1371  * Given a LISTEN socket and an inbound SYN request, add
1372  * this to the syn cache, and send back a segment:
1373  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1374  * to the source.
1375  *
1376  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1377  * Doing so would require that we hold onto the data and deliver it
1378  * to the application.  However, if we are the target of a SYN-flood
1379  * DoS attack, an attacker could send data which would eventually
1380  * consume all available buffer space if it were ACKed.  By not ACKing
1381  * the data, we avoid this DoS scenario.
1382  *
1383  * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1384  * cookie is processed and a new socket is created.  In this case, any data
1385  * accompanying the SYN will be queued to the socket by tcp_input() and will
1386  * be ACKed either when the application sends response data or the delayed
1387  * ACK timer expires, whichever comes first.
1388  */
1389 struct socket *
1390 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1391     struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod,
1392     void *todctx, uint8_t iptos, uint16_t port)
1393 {
1394 	struct tcpcb *tp;
1395 	struct socket *rv = NULL;
1396 	struct syncache *sc = NULL;
1397 	struct syncache_head *sch;
1398 	struct mbuf *ipopts = NULL;
1399 	u_int ltflags;
1400 	int win, ip_ttl, ip_tos;
1401 	char *s;
1402 #ifdef INET6
1403 	int autoflowlabel = 0;
1404 #endif
1405 #ifdef MAC
1406 	struct label *maclabel;
1407 #endif
1408 	struct syncache scs;
1409 	struct ucred *cred;
1410 	uint64_t tfo_response_cookie;
1411 	unsigned int *tfo_pending = NULL;
1412 	int tfo_cookie_valid = 0;
1413 	int tfo_response_cookie_valid = 0;
1414 	bool locked;
1415 
1416 	INP_RLOCK_ASSERT(inp);			/* listen socket */
1417 	KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1418 	    ("%s: unexpected tcp flags", __func__));
1419 
1420 	/*
1421 	 * Combine all so/tp operations very early to drop the INP lock as
1422 	 * soon as possible.
1423 	 */
1424 	KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1425 	tp = sototcpcb(so);
1426 	cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred);
1427 
1428 #ifdef INET6
1429 	if (inc->inc_flags & INC_ISIPV6) {
1430 		if (inp->inp_flags & IN6P_AUTOFLOWLABEL) {
1431 			autoflowlabel = 1;
1432 		}
1433 		ip_ttl = in6_selecthlim(inp, NULL);
1434 		if ((inp->in6p_outputopts == NULL) ||
1435 		    (inp->in6p_outputopts->ip6po_tclass == -1)) {
1436 			ip_tos = 0;
1437 		} else {
1438 			ip_tos = inp->in6p_outputopts->ip6po_tclass;
1439 		}
1440 	}
1441 #endif
1442 #if defined(INET6) && defined(INET)
1443 	else
1444 #endif
1445 #ifdef INET
1446 	{
1447 		ip_ttl = inp->inp_ip_ttl;
1448 		ip_tos = inp->inp_ip_tos;
1449 	}
1450 #endif
1451 	win = so->sol_sbrcv_hiwat;
1452 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1453 
1454 	if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
1455 	    (tp->t_tfo_pending != NULL) &&
1456 	    (to->to_flags & TOF_FASTOPEN)) {
1457 		/*
1458 		 * Limit the number of pending TFO connections to
1459 		 * approximately half of the queue limit.  This prevents TFO
1460 		 * SYN floods from starving the service by filling the
1461 		 * listen queue with bogus TFO connections.
1462 		 */
1463 		if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1464 		    (so->sol_qlimit / 2)) {
1465 			int result;
1466 
1467 			result = tcp_fastopen_check_cookie(inc,
1468 			    to->to_tfo_cookie, to->to_tfo_len,
1469 			    &tfo_response_cookie);
1470 			tfo_cookie_valid = (result > 0);
1471 			tfo_response_cookie_valid = (result >= 0);
1472 		}
1473 
1474 		/*
1475 		 * Remember the TFO pending counter as it will have to be
1476 		 * decremented below if we don't make it to syncache_tfo_expand().
1477 		 */
1478 		tfo_pending = tp->t_tfo_pending;
1479 	}
1480 
1481 #ifdef MAC
1482 	if (mac_syncache_init(&maclabel) != 0) {
1483 		INP_RUNLOCK(inp);
1484 		goto done;
1485 	} else
1486 		mac_syncache_create(maclabel, inp);
1487 #endif
1488 	if (!tfo_cookie_valid)
1489 		INP_RUNLOCK(inp);
1490 
1491 	/*
1492 	 * Remember the IP options, if any.
1493 	 */
1494 #ifdef INET6
1495 	if (!(inc->inc_flags & INC_ISIPV6))
1496 #endif
1497 #ifdef INET
1498 		ipopts = (m) ? ip_srcroute(m) : NULL;
1499 #else
1500 		ipopts = NULL;
1501 #endif
1502 
1503 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1504 	/*
1505 	 * When the socket is TCP-MD5 enabled check that,
1506 	 *  - a signed packet is valid
1507 	 *  - a non-signed packet does not have a security association
1508 	 *
1509 	 *  If a signed packet fails validation or a non-signed packet has a
1510 	 *  security association, the packet will be dropped.
1511 	 */
1512 	if (ltflags & TF_SIGNATURE) {
1513 		if (to->to_flags & TOF_SIGNATURE) {
1514 			if (!TCPMD5_ENABLED() ||
1515 			    TCPMD5_INPUT(m, th, to->to_signature) != 0)
1516 				goto done;
1517 		} else {
1518 			if (TCPMD5_ENABLED() &&
1519 			    TCPMD5_INPUT(m, NULL, NULL) != ENOENT)
1520 				goto done;
1521 		}
1522 	} else if (to->to_flags & TOF_SIGNATURE)
1523 		goto done;
1524 #endif	/* TCP_SIGNATURE */
1525 	/*
1526 	 * See if we already have an entry for this connection.
1527 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1528 	 *
1529 	 * XXX: should the syncache be re-initialized with the contents
1530 	 * of the new SYN here (which may have different options?)
1531 	 *
1532 	 * XXX: We do not check the sequence number to see if this is a
1533 	 * real retransmit or a new connection attempt.  The question is
1534 	 * how to handle such a case; either ignore it as spoofed, or
1535 	 * drop the current entry and create a new one?
1536 	 */
1537 	if (syncache_cookiesonly()) {
1538 		sc = NULL;
1539 		sch = syncache_hashbucket(inc);
1540 		locked = false;
1541 	} else {
1542 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1543 		locked = true;
1544 		SCH_LOCK_ASSERT(sch);
1545 	}
1546 	if (sc != NULL) {
1547 		if (tfo_cookie_valid)
1548 			INP_RUNLOCK(inp);
1549 		TCPSTAT_INC(tcps_sc_dupsyn);
1550 		if (ipopts) {
1551 			/*
1552 			 * If we were remembering a previous source route,
1553 			 * forget it and use the new one we've been given.
1554 			 */
1555 			if (sc->sc_ipopts)
1556 				(void) m_free(sc->sc_ipopts);
1557 			sc->sc_ipopts = ipopts;
1558 		}
1559 		/*
1560 		 * Update timestamp if present.
1561 		 */
1562 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1563 			sc->sc_tsreflect = to->to_tsval;
1564 		else
1565 			sc->sc_flags &= ~SCF_TIMESTAMP;
1566 		/*
1567 		 * Adjust ECN response if needed, e.g. different
1568 		 * IP ECN field, or a fallback by the remote host.
1569 		 */
1570 		if (sc->sc_flags & SCF_ECN_MASK) {
1571 			sc->sc_flags &= ~SCF_ECN_MASK;
1572 			sc->sc_flags = tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1573 		}
1574 #ifdef MAC
1575 		/*
1576 		 * Since we have already unconditionally allocated label
1577 		 * storage, free it up.  The syncache entry will already
1578 		 * have an initialized label we can use.
1579 		 */
1580 		mac_syncache_destroy(&maclabel);
1581 #endif
1582 		TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1583 		/* Retransmit SYN|ACK and reset retransmit count. */
1584 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1585 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1586 			    "resetting timer and retransmitting SYN|ACK\n",
1587 			    s, __func__);
1588 			free(s, M_TCPLOG);
1589 		}
1590 		if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1591 			sc->sc_rxmits = 0;
1592 			syncache_timeout(sc, sch, 1);
1593 			TCPSTAT_INC(tcps_sndacks);
1594 			TCPSTAT_INC(tcps_sndtotal);
1595 		}
1596 		SCH_UNLOCK(sch);
1597 		goto donenoprobe;
1598 	}
1599 
1600 	if (tfo_cookie_valid) {
1601 		bzero(&scs, sizeof(scs));
1602 		sc = &scs;
1603 		goto skip_alloc;
1604 	}
1605 
1606 	/*
1607 	 * Skip allocating a syncache entry if we are just going to discard
1608 	 * it later.
1609 	 */
1610 	if (!locked) {
1611 		bzero(&scs, sizeof(scs));
1612 		sc = &scs;
1613 	} else
1614 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1615 	if (sc == NULL) {
1616 		/*
1617 		 * The zone allocator couldn't provide more entries.
1618 		 * Treat this as if the cache was full; drop the oldest
1619 		 * entry and insert the new one.
1620 		 */
1621 		TCPSTAT_INC(tcps_sc_zonefail);
1622 		if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1623 			sch->sch_last_overflow = time_uptime;
1624 			syncache_drop(sc, sch);
1625 			syncache_pause(inc);
1626 		}
1627 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1628 		if (sc == NULL) {
1629 			if (V_tcp_syncookies) {
1630 				bzero(&scs, sizeof(scs));
1631 				sc = &scs;
1632 			} else {
1633 				KASSERT(locked,
1634 				    ("%s: bucket unexpectedly unlocked",
1635 				    __func__));
1636 				SCH_UNLOCK(sch);
1637 				if (ipopts)
1638 					(void) m_free(ipopts);
1639 				goto done;
1640 			}
1641 		}
1642 	}
1643 
1644 skip_alloc:
1645 	if (!tfo_cookie_valid && tfo_response_cookie_valid)
1646 		sc->sc_tfo_cookie = &tfo_response_cookie;
1647 
1648 	/*
1649 	 * Fill in the syncache values.
1650 	 */
1651 #ifdef MAC
1652 	sc->sc_label = maclabel;
1653 #endif
1654 	sc->sc_cred = cred;
1655 	sc->sc_port = port;
1656 	cred = NULL;
1657 	sc->sc_ipopts = ipopts;
1658 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1659 	sc->sc_ip_tos = ip_tos;
1660 	sc->sc_ip_ttl = ip_ttl;
1661 #ifdef TCP_OFFLOAD
1662 	sc->sc_tod = tod;
1663 	sc->sc_todctx = todctx;
1664 #endif
1665 	sc->sc_irs = th->th_seq;
1666 	sc->sc_flags = 0;
1667 	sc->sc_flowlabel = 0;
1668 
1669 	/*
1670 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1671 	 * win was derived from socket earlier in the function.
1672 	 */
1673 	win = imax(win, 0);
1674 	win = imin(win, TCP_MAXWIN);
1675 	sc->sc_wnd = win;
1676 
1677 	if (V_tcp_do_rfc1323 &&
1678 	    !(ltflags & TF_NOOPT)) {
1679 		/*
1680 		 * A timestamp received in a SYN makes
1681 		 * it ok to send timestamp requests and replies.
1682 		 */
1683 		if ((to->to_flags & TOF_TS) && (V_tcp_do_rfc1323 != 2)) {
1684 			sc->sc_tsreflect = to->to_tsval;
1685 			sc->sc_flags |= SCF_TIMESTAMP;
1686 			sc->sc_tsoff = tcp_new_ts_offset(inc);
1687 		}
1688 		if ((to->to_flags & TOF_SCALE) && (V_tcp_do_rfc1323 != 3)) {
1689 			int wscale = 0;
1690 
1691 			/*
1692 			 * Pick the smallest possible scaling factor that
1693 			 * will still allow us to scale up to sb_max, aka
1694 			 * kern.ipc.maxsockbuf.
1695 			 *
1696 			 * We do this because there are broken firewalls that
1697 			 * will corrupt the window scale option, leading to
1698 			 * the other endpoint believing that our advertised
1699 			 * window is unscaled.  At scale factors larger than
1700 			 * 5 the unscaled window will drop below 1500 bytes,
1701 			 * leading to serious problems when traversing these
1702 			 * broken firewalls.
1703 			 *
1704 			 * With the default maxsockbuf of 256K, a scale factor
1705 			 * of 3 will be chosen by this algorithm.  Those who
1706 			 * choose a larger maxsockbuf should watch out
1707 			 * for the compatibility problems mentioned above.
1708 			 *
1709 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1710 			 * or <SYN,ACK>) segment itself is never scaled.
1711 			 */
1712 			while (wscale < TCP_MAX_WINSHIFT &&
1713 			    (TCP_MAXWIN << wscale) < sb_max)
1714 				wscale++;
1715 			sc->sc_requested_r_scale = wscale;
1716 			sc->sc_requested_s_scale = to->to_wscale;
1717 			sc->sc_flags |= SCF_WINSCALE;
1718 		}
1719 	}
1720 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1721 	/*
1722 	 * If incoming packet has an MD5 signature, flag this in the
1723 	 * syncache so that syncache_respond() will do the right thing
1724 	 * with the SYN+ACK.
1725 	 */
1726 	if (to->to_flags & TOF_SIGNATURE)
1727 		sc->sc_flags |= SCF_SIGNATURE;
1728 #endif	/* TCP_SIGNATURE */
1729 	if (to->to_flags & TOF_SACKPERM)
1730 		sc->sc_flags |= SCF_SACK;
1731 	if (to->to_flags & TOF_MSS)
1732 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1733 	if (ltflags & TF_NOOPT)
1734 		sc->sc_flags |= SCF_NOOPT;
1735 	/* ECN Handshake */
1736 	if (V_tcp_do_ecn)
1737 		sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1738 
1739 	if (V_tcp_syncookies)
1740 		sc->sc_iss = syncookie_generate(sch, sc);
1741 	else
1742 		sc->sc_iss = arc4random();
1743 #ifdef INET6
1744 	if (autoflowlabel) {
1745 		if (V_tcp_syncookies)
1746 			sc->sc_flowlabel = sc->sc_iss;
1747 		else
1748 			sc->sc_flowlabel = ip6_randomflowlabel();
1749 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1750 	}
1751 #endif
1752 	if (locked)
1753 		SCH_UNLOCK(sch);
1754 
1755 	if (tfo_cookie_valid) {
1756 		rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie);
1757 		/* INP_RUNLOCK(inp) will be performed by the caller */
1758 		goto tfo_expanded;
1759 	}
1760 
1761 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1762 	/*
1763 	 * Do a standard 3-way handshake.
1764 	 */
1765 	if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1766 		if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1767 			syncache_free(sc);
1768 		else if (sc != &scs)
1769 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1770 		TCPSTAT_INC(tcps_sndacks);
1771 		TCPSTAT_INC(tcps_sndtotal);
1772 	} else {
1773 		if (sc != &scs)
1774 			syncache_free(sc);
1775 		TCPSTAT_INC(tcps_sc_dropped);
1776 	}
1777 	goto donenoprobe;
1778 
1779 done:
1780 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1781 donenoprobe:
1782 	if (m)
1783 		m_freem(m);
1784 	/*
1785 	 * If tfo_pending is not NULL here, then a TFO SYN that did not
1786 	 * result in a new socket was processed and the associated pending
1787 	 * counter has not yet been decremented.  All such TFO processing paths
1788 	 * transit this point.
1789 	 */
1790 	if (tfo_pending != NULL)
1791 		tcp_fastopen_decrement_counter(tfo_pending);
1792 
1793 tfo_expanded:
1794 	if (cred != NULL)
1795 		crfree(cred);
1796 #ifdef MAC
1797 	if (sc == &scs)
1798 		mac_syncache_destroy(&maclabel);
1799 #endif
1800 	return (rv);
1801 }
1802 
1803 /*
1804  * Send SYN|ACK or ACK to the peer.  Either in response to a peer's segment,
1805  * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1806  */
1807 static int
1808 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1809 {
1810 	struct ip *ip = NULL;
1811 	struct mbuf *m;
1812 	struct tcphdr *th = NULL;
1813 	struct udphdr *udp = NULL;
1814 	int optlen, error = 0;	/* Make compiler happy */
1815 	u_int16_t hlen, tlen, mssopt, ulen;
1816 	struct tcpopt to;
1817 #ifdef INET6
1818 	struct ip6_hdr *ip6 = NULL;
1819 #endif
1820 
1821 	NET_EPOCH_ASSERT();
1822 
1823 	hlen =
1824 #ifdef INET6
1825 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1826 #endif
1827 		sizeof(struct ip);
1828 	tlen = hlen + sizeof(struct tcphdr);
1829 	if (sc->sc_port) {
1830 		tlen += sizeof(struct udphdr);
1831 	}
1832 	/* Determine MSS we advertize to other end of connection. */
1833 	mssopt = tcp_mssopt(&sc->sc_inc);
1834 	if (sc->sc_port)
1835 		mssopt -= V_tcp_udp_tunneling_overhead;
1836 	mssopt = max(mssopt, V_tcp_minmss);
1837 
1838 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1839 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1840 	    ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + "
1841 	    "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port,
1842 	    max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN));
1843 
1844 	/* Create the IP+TCP header from scratch. */
1845 	m = m_gethdr(M_NOWAIT, MT_DATA);
1846 	if (m == NULL)
1847 		return (ENOBUFS);
1848 #ifdef MAC
1849 	mac_syncache_create_mbuf(sc->sc_label, m);
1850 #endif
1851 	m->m_data += max_linkhdr;
1852 	m->m_len = tlen;
1853 	m->m_pkthdr.len = tlen;
1854 	m->m_pkthdr.rcvif = NULL;
1855 
1856 #ifdef INET6
1857 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1858 		ip6 = mtod(m, struct ip6_hdr *);
1859 		ip6->ip6_vfc = IPV6_VERSION;
1860 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1861 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1862 		ip6->ip6_plen = htons(tlen - hlen);
1863 		/* ip6_hlim is set after checksum */
1864 		/* Zero out traffic class and flow label. */
1865 		ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1866 		ip6->ip6_flow |= sc->sc_flowlabel;
1867 		if (sc->sc_port != 0) {
1868 			ip6->ip6_nxt = IPPROTO_UDP;
1869 			udp = (struct udphdr *)(ip6 + 1);
1870 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1871 			udp->uh_dport = sc->sc_port;
1872 			ulen = (tlen - sizeof(struct ip6_hdr));
1873 			th = (struct tcphdr *)(udp + 1);
1874 		} else {
1875 			ip6->ip6_nxt = IPPROTO_TCP;
1876 			th = (struct tcphdr *)(ip6 + 1);
1877 		}
1878 		ip6->ip6_flow |= htonl(sc->sc_ip_tos << 20);
1879 	}
1880 #endif
1881 #if defined(INET6) && defined(INET)
1882 	else
1883 #endif
1884 #ifdef INET
1885 	{
1886 		ip = mtod(m, struct ip *);
1887 		ip->ip_v = IPVERSION;
1888 		ip->ip_hl = sizeof(struct ip) >> 2;
1889 		ip->ip_len = htons(tlen);
1890 		ip->ip_id = 0;
1891 		ip->ip_off = 0;
1892 		ip->ip_sum = 0;
1893 		ip->ip_src = sc->sc_inc.inc_laddr;
1894 		ip->ip_dst = sc->sc_inc.inc_faddr;
1895 		ip->ip_ttl = sc->sc_ip_ttl;
1896 		ip->ip_tos = sc->sc_ip_tos;
1897 
1898 		/*
1899 		 * See if we should do MTU discovery.  Route lookups are
1900 		 * expensive, so we will only unset the DF bit if:
1901 		 *
1902 		 *	1) path_mtu_discovery is disabled
1903 		 *	2) the SCF_UNREACH flag has been set
1904 		 */
1905 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1906 		       ip->ip_off |= htons(IP_DF);
1907 		if (sc->sc_port == 0) {
1908 			ip->ip_p = IPPROTO_TCP;
1909 			th = (struct tcphdr *)(ip + 1);
1910 		} else {
1911 			ip->ip_p = IPPROTO_UDP;
1912 			udp = (struct udphdr *)(ip + 1);
1913 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1914 			udp->uh_dport = sc->sc_port;
1915 			ulen = (tlen - sizeof(struct ip));
1916 			th = (struct tcphdr *)(udp + 1);
1917 		}
1918 	}
1919 #endif /* INET */
1920 	th->th_sport = sc->sc_inc.inc_lport;
1921 	th->th_dport = sc->sc_inc.inc_fport;
1922 
1923 	if (flags & TH_SYN)
1924 		th->th_seq = htonl(sc->sc_iss);
1925 	else
1926 		th->th_seq = htonl(sc->sc_iss + 1);
1927 	th->th_ack = htonl(sc->sc_irs + 1);
1928 	th->th_off = sizeof(struct tcphdr) >> 2;
1929 	th->th_win = htons(sc->sc_wnd);
1930 	th->th_urp = 0;
1931 
1932 	flags = tcp_ecn_syncache_respond(flags, sc);
1933 	tcp_set_flags(th, flags);
1934 
1935 	/* Tack on the TCP options. */
1936 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1937 		to.to_flags = 0;
1938 
1939 		if (flags & TH_SYN) {
1940 			to.to_mss = mssopt;
1941 			to.to_flags = TOF_MSS;
1942 			if (sc->sc_flags & SCF_WINSCALE) {
1943 				to.to_wscale = sc->sc_requested_r_scale;
1944 				to.to_flags |= TOF_SCALE;
1945 			}
1946 			if (sc->sc_flags & SCF_SACK)
1947 				to.to_flags |= TOF_SACKPERM;
1948 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1949 			if (sc->sc_flags & SCF_SIGNATURE)
1950 				to.to_flags |= TOF_SIGNATURE;
1951 #endif
1952 			if (sc->sc_tfo_cookie) {
1953 				to.to_flags |= TOF_FASTOPEN;
1954 				to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1955 				to.to_tfo_cookie = sc->sc_tfo_cookie;
1956 				/* don't send cookie again when retransmitting response */
1957 				sc->sc_tfo_cookie = NULL;
1958 			}
1959 		}
1960 		if (sc->sc_flags & SCF_TIMESTAMP) {
1961 			to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1962 			to.to_tsecr = sc->sc_tsreflect;
1963 			to.to_flags |= TOF_TS;
1964 		}
1965 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1966 
1967 		/* Adjust headers by option size. */
1968 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1969 		m->m_len += optlen;
1970 		m->m_pkthdr.len += optlen;
1971 #ifdef INET6
1972 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1973 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1974 		else
1975 #endif
1976 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1977 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1978 		if (sc->sc_flags & SCF_SIGNATURE) {
1979 			KASSERT(to.to_flags & TOF_SIGNATURE,
1980 			    ("tcp_addoptions() didn't set tcp_signature"));
1981 
1982 			/* NOTE: to.to_signature is inside of mbuf */
1983 			if (!TCPMD5_ENABLED() ||
1984 			    TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1985 				m_freem(m);
1986 				return (EACCES);
1987 			}
1988 		}
1989 #endif
1990 	} else
1991 		optlen = 0;
1992 
1993 	if (udp) {
1994 		ulen += optlen;
1995 		udp->uh_ulen = htons(ulen);
1996 	}
1997 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
1998 	/*
1999 	 * If we have peer's SYN and it has a flowid, then let's assign it to
2000 	 * our SYN|ACK.  ip6_output() and ip_output() will not assign flowid
2001 	 * to SYN|ACK due to lack of inp here.
2002 	 */
2003 	if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
2004 		m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
2005 		M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
2006 	}
2007 #ifdef INET6
2008 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
2009 		if (sc->sc_port) {
2010 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
2011 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2012 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen,
2013 			      IPPROTO_UDP, 0);
2014 			th->th_sum = htons(0);
2015 		} else {
2016 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
2017 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2018 			th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
2019 			    IPPROTO_TCP, 0);
2020 		}
2021 		ip6->ip6_hlim = sc->sc_ip_ttl;
2022 #ifdef TCP_OFFLOAD
2023 		if (ADDED_BY_TOE(sc)) {
2024 			struct toedev *tod = sc->sc_tod;
2025 
2026 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2027 
2028 			return (error);
2029 		}
2030 #endif
2031 		TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
2032 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
2033 	}
2034 #endif
2035 #if defined(INET6) && defined(INET)
2036 	else
2037 #endif
2038 #ifdef INET
2039 	{
2040 		if (sc->sc_port) {
2041 			m->m_pkthdr.csum_flags = CSUM_UDP;
2042 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2043 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
2044 			      ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
2045 			th->th_sum = htons(0);
2046 		} else {
2047 			m->m_pkthdr.csum_flags = CSUM_TCP;
2048 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2049 			th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
2050 			    htons(tlen + optlen - hlen + IPPROTO_TCP));
2051 		}
2052 #ifdef TCP_OFFLOAD
2053 		if (ADDED_BY_TOE(sc)) {
2054 			struct toedev *tod = sc->sc_tod;
2055 
2056 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2057 
2058 			return (error);
2059 		}
2060 #endif
2061 		TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
2062 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
2063 	}
2064 #endif
2065 	return (error);
2066 }
2067 
2068 /*
2069  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
2070  * that exceed the capacity of the syncache by avoiding the storage of any
2071  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
2072  * attacks where the attacker does not have access to our responses.
2073  *
2074  * Syncookies encode and include all necessary information about the
2075  * connection setup within the SYN|ACK that we send back.  That way we
2076  * can avoid keeping any local state until the ACK to our SYN|ACK returns
2077  * (if ever).  Normally the syncache and syncookies are running in parallel
2078  * with the latter taking over when the former is exhausted.  When matching
2079  * syncache entry is found the syncookie is ignored.
2080  *
2081  * The only reliable information persisting the 3WHS is our initial sequence
2082  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
2083  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
2084  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
2085  * returns and signifies a legitimate connection if it matches the ACK.
2086  *
2087  * The available space of 32 bits to store the hash and to encode the SYN
2088  * option information is very tight and we should have at least 24 bits for
2089  * the MAC to keep the number of guesses by blind spoofing reasonably high.
2090  *
2091  * SYN option information we have to encode to fully restore a connection:
2092  * MSS: is imporant to chose an optimal segment size to avoid IP level
2093  *   fragmentation along the path.  The common MSS values can be encoded
2094  *   in a 3-bit table.  Uncommon values are captured by the next lower value
2095  *   in the table leading to a slight increase in packetization overhead.
2096  * WSCALE: is necessary to allow large windows to be used for high delay-
2097  *   bandwidth product links.  Not scaling the window when it was initially
2098  *   negotiated is bad for performance as lack of scaling further decreases
2099  *   the apparent available send window.  We only need to encode the WSCALE
2100  *   we received from the remote end.  Our end can be recalculated at any
2101  *   time.  The common WSCALE values can be encoded in a 3-bit table.
2102  *   Uncommon values are captured by the next lower value in the table
2103  *   making us under-estimate the available window size halving our
2104  *   theoretically possible maximum throughput for that connection.
2105  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2106  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2107  *   that are included in all segments on a connection.  We enable them when
2108  *   the ACK has them.
2109  *
2110  * Security of syncookies and attack vectors:
2111  *
2112  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2113  * together with the gloabl secret to make it unique per connection attempt.
2114  * Thus any change of any of those parameters results in a different MAC output
2115  * in an unpredictable way unless a collision is encountered.  24 bits of the
2116  * MAC are embedded into the ISS.
2117  *
2118  * To prevent replay attacks two rotating global secrets are updated with a
2119  * new random value every 15 seconds.  The life-time of a syncookie is thus
2120  * 15-30 seconds.
2121  *
2122  * Vector 1: Attacking the secret.  This requires finding a weakness in the
2123  * MAC itself or the way it is used here.  The attacker can do a chosen plain
2124  * text attack by varying and testing the all parameters under his control.
2125  * The strength depends on the size and randomness of the secret, and the
2126  * cryptographic security of the MAC function.  Due to the constant updating
2127  * of the secret the attacker has at most 29.999 seconds to find the secret
2128  * and launch spoofed connections.  After that he has to start all over again.
2129  *
2130  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
2131  * size an average of 4,823 attempts are required for a 50% chance of success
2132  * to spoof a single syncookie (birthday collision paradox).  However the
2133  * attacker is blind and doesn't know if one of his attempts succeeded unless
2134  * he has a side channel to interfere success from.  A single connection setup
2135  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2136  * This many attempts are required for each one blind spoofed connection.  For
2137  * every additional spoofed connection he has to launch another N attempts.
2138  * Thus for a sustained rate 100 spoofed connections per second approximately
2139  * 1,800,000 packets per second would have to be sent.
2140  *
2141  * NB: The MAC function should be fast so that it doesn't become a CPU
2142  * exhaustion attack vector itself.
2143  *
2144  * References:
2145  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2146  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2147  *   http://cr.yp.to/syncookies.html    (overview)
2148  *   http://cr.yp.to/syncookies/archive (details)
2149  *
2150  *
2151  * Schematic construction of a syncookie enabled Initial Sequence Number:
2152  *  0        1         2         3
2153  *  12345678901234567890123456789012
2154  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2155  *
2156  *  x 24 MAC (truncated)
2157  *  W  3 Send Window Scale index
2158  *  M  3 MSS index
2159  *  S  1 SACK permitted
2160  *  P  1 Odd/even secret
2161  */
2162 
2163 /*
2164  * Distribution and probability of certain MSS values.  Those in between are
2165  * rounded down to the next lower one.
2166  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2167  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
2168  */
2169 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2170 
2171 /*
2172  * Distribution and probability of certain WSCALE values.  We have to map the
2173  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2174  * bits based on prevalence of certain values.  Where we don't have an exact
2175  * match for are rounded down to the next lower one letting us under-estimate
2176  * the true available window.  At the moment this would happen only for the
2177  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2178  * and window size).  The absence of the WSCALE option (no scaling in either
2179  * direction) is encoded with index zero.
2180  * [WSCALE values histograms, Allman, 2012]
2181  *                            X 10 10 35  5  6 14 10%   by host
2182  *                            X 11  4  5  5 18 49  3%   by connections
2183  */
2184 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2185 
2186 /*
2187  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
2188  * and good cryptographic properties.
2189  */
2190 static uint32_t
2191 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2192     uint8_t *secbits, uintptr_t secmod)
2193 {
2194 	SIPHASH_CTX ctx;
2195 	uint32_t siphash[2];
2196 
2197 	SipHash24_Init(&ctx);
2198 	SipHash_SetKey(&ctx, secbits);
2199 	switch (inc->inc_flags & INC_ISIPV6) {
2200 #ifdef INET
2201 	case 0:
2202 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2203 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2204 		break;
2205 #endif
2206 #ifdef INET6
2207 	case INC_ISIPV6:
2208 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2209 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2210 		break;
2211 #endif
2212 	}
2213 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2214 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2215 	SipHash_Update(&ctx, &irs, sizeof(irs));
2216 	SipHash_Update(&ctx, &flags, sizeof(flags));
2217 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
2218 	SipHash_Final((u_int8_t *)&siphash, &ctx);
2219 
2220 	return (siphash[0] ^ siphash[1]);
2221 }
2222 
2223 static tcp_seq
2224 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2225 {
2226 	u_int i, secbit, wscale;
2227 	uint32_t iss, hash;
2228 	uint8_t *secbits;
2229 	union syncookie cookie;
2230 
2231 	cookie.cookie = 0;
2232 
2233 	/* Map our computed MSS into the 3-bit index. */
2234 	for (i = nitems(tcp_sc_msstab) - 1;
2235 	     tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2236 	     i--)
2237 		;
2238 	cookie.flags.mss_idx = i;
2239 
2240 	/*
2241 	 * Map the send window scale into the 3-bit index but only if
2242 	 * the wscale option was received.
2243 	 */
2244 	if (sc->sc_flags & SCF_WINSCALE) {
2245 		wscale = sc->sc_requested_s_scale;
2246 		for (i = nitems(tcp_sc_wstab) - 1;
2247 		    tcp_sc_wstab[i] > wscale && i > 0;
2248 		     i--)
2249 			;
2250 		cookie.flags.wscale_idx = i;
2251 	}
2252 
2253 	/* Can we do SACK? */
2254 	if (sc->sc_flags & SCF_SACK)
2255 		cookie.flags.sack_ok = 1;
2256 
2257 	/* Which of the two secrets to use. */
2258 	secbit = V_tcp_syncache.secret.oddeven & 0x1;
2259 	cookie.flags.odd_even = secbit;
2260 
2261 	secbits = V_tcp_syncache.secret.key[secbit];
2262 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2263 	    (uintptr_t)sch);
2264 
2265 	/*
2266 	 * Put the flags into the hash and XOR them to get better ISS number
2267 	 * variance.  This doesn't enhance the cryptographic strength and is
2268 	 * done to prevent the 8 cookie bits from showing up directly on the
2269 	 * wire.
2270 	 */
2271 	iss = hash & ~0xff;
2272 	iss |= cookie.cookie ^ (hash >> 24);
2273 
2274 	TCPSTAT_INC(tcps_sc_sendcookie);
2275 	return (iss);
2276 }
2277 
2278 static struct syncache *
2279 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2280     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2281     struct socket *lso, uint16_t port)
2282 {
2283 	uint32_t hash;
2284 	uint8_t *secbits;
2285 	tcp_seq ack, seq;
2286 	int wnd, wscale = 0;
2287 	union syncookie cookie;
2288 
2289 	/*
2290 	 * Pull information out of SYN-ACK/ACK and revert sequence number
2291 	 * advances.
2292 	 */
2293 	ack = th->th_ack - 1;
2294 	seq = th->th_seq - 1;
2295 
2296 	/*
2297 	 * Unpack the flags containing enough information to restore the
2298 	 * connection.
2299 	 */
2300 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2301 
2302 	/* Which of the two secrets to use. */
2303 	secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2304 
2305 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2306 
2307 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
2308 	if ((ack & ~0xff) != (hash & ~0xff))
2309 		return (NULL);
2310 
2311 	/* Fill in the syncache values. */
2312 	sc->sc_flags = 0;
2313 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2314 	sc->sc_ipopts = NULL;
2315 
2316 	sc->sc_irs = seq;
2317 	sc->sc_iss = ack;
2318 
2319 	switch (inc->inc_flags & INC_ISIPV6) {
2320 #ifdef INET
2321 	case 0:
2322 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2323 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2324 		break;
2325 #endif
2326 #ifdef INET6
2327 	case INC_ISIPV6:
2328 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2329 			sc->sc_flowlabel =
2330 			    htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2331 		break;
2332 #endif
2333 	}
2334 
2335 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2336 
2337 	/* We can simply recompute receive window scale we sent earlier. */
2338 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2339 		wscale++;
2340 
2341 	/* Only use wscale if it was enabled in the orignal SYN. */
2342 	if (cookie.flags.wscale_idx > 0) {
2343 		sc->sc_requested_r_scale = wscale;
2344 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2345 		sc->sc_flags |= SCF_WINSCALE;
2346 	}
2347 
2348 	wnd = lso->sol_sbrcv_hiwat;
2349 	wnd = imax(wnd, 0);
2350 	wnd = imin(wnd, TCP_MAXWIN);
2351 	sc->sc_wnd = wnd;
2352 
2353 	if (cookie.flags.sack_ok)
2354 		sc->sc_flags |= SCF_SACK;
2355 
2356 	if (to->to_flags & TOF_TS) {
2357 		sc->sc_flags |= SCF_TIMESTAMP;
2358 		sc->sc_tsreflect = to->to_tsval;
2359 		sc->sc_tsoff = tcp_new_ts_offset(inc);
2360 	}
2361 
2362 	if (to->to_flags & TOF_SIGNATURE)
2363 		sc->sc_flags |= SCF_SIGNATURE;
2364 
2365 	sc->sc_rxmits = 0;
2366 
2367 	sc->sc_port = port;
2368 
2369 	TCPSTAT_INC(tcps_sc_recvcookie);
2370 	return (sc);
2371 }
2372 
2373 #ifdef INVARIANTS
2374 static int
2375 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2376     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2377     struct socket *lso, uint16_t port)
2378 {
2379 	struct syncache scs, *scx;
2380 	char *s;
2381 
2382 	bzero(&scs, sizeof(scs));
2383 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port);
2384 
2385 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2386 		return (0);
2387 
2388 	if (scx != NULL) {
2389 		if (sc->sc_peer_mss != scx->sc_peer_mss)
2390 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2391 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2392 
2393 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2394 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2395 			    s, __func__, sc->sc_requested_r_scale,
2396 			    scx->sc_requested_r_scale);
2397 
2398 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2399 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2400 			    s, __func__, sc->sc_requested_s_scale,
2401 			    scx->sc_requested_s_scale);
2402 
2403 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2404 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2405 	}
2406 
2407 	if (s != NULL)
2408 		free(s, M_TCPLOG);
2409 	return (0);
2410 }
2411 #endif /* INVARIANTS */
2412 
2413 static void
2414 syncookie_reseed(void *arg)
2415 {
2416 	struct tcp_syncache *sc = arg;
2417 	uint8_t *secbits;
2418 	int secbit;
2419 
2420 	/*
2421 	 * Reseeding the secret doesn't have to be protected by a lock.
2422 	 * It only must be ensured that the new random values are visible
2423 	 * to all CPUs in a SMP environment.  The atomic with release
2424 	 * semantics ensures that.
2425 	 */
2426 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2427 	secbits = sc->secret.key[secbit];
2428 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2429 	atomic_add_rel_int(&sc->secret.oddeven, 1);
2430 
2431 	/* Reschedule ourself. */
2432 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2433 }
2434 
2435 /*
2436  * We have overflowed a bucket. Let's pause dealing with the syncache.
2437  * This function will increment the bucketoverflow statistics appropriately
2438  * (once per pause when pausing is enabled; otherwise, once per overflow).
2439  */
2440 static void
2441 syncache_pause(struct in_conninfo *inc)
2442 {
2443 	time_t delta;
2444 	const char *s;
2445 
2446 	/* XXX:
2447 	 * 2. Add sysctl read here so we don't get the benefit of this
2448 	 * change without the new sysctl.
2449 	 */
2450 
2451 	/*
2452 	 * Try an unlocked read. If we already know that another thread
2453 	 * has activated the feature, there is no need to proceed.
2454 	 */
2455 	if (V_tcp_syncache.paused)
2456 		return;
2457 
2458 	/* Are cookied enabled? If not, we can't pause. */
2459 	if (!V_tcp_syncookies) {
2460 		TCPSTAT_INC(tcps_sc_bucketoverflow);
2461 		return;
2462 	}
2463 
2464 	/*
2465 	 * We may be the first thread to find an overflow. Get the lock
2466 	 * and evaluate if we need to take action.
2467 	 */
2468 	mtx_lock(&V_tcp_syncache.pause_mtx);
2469 	if (V_tcp_syncache.paused) {
2470 		mtx_unlock(&V_tcp_syncache.pause_mtx);
2471 		return;
2472 	}
2473 
2474 	/* Activate protection. */
2475 	V_tcp_syncache.paused = true;
2476 	TCPSTAT_INC(tcps_sc_bucketoverflow);
2477 
2478 	/*
2479 	 * Determine the last backoff time. If we are seeing a re-newed
2480 	 * attack within that same time after last reactivating the syncache,
2481 	 * consider it an extension of the same attack.
2482 	 */
2483 	delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2484 	if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2485 		if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2486 			delta <<= 1;
2487 			V_tcp_syncache.pause_backoff++;
2488 		}
2489 	} else {
2490 		delta = TCP_SYNCACHE_PAUSE_TIME;
2491 		V_tcp_syncache.pause_backoff = 0;
2492 	}
2493 
2494 	/* Log a warning, including IP addresses, if able. */
2495 	if (inc != NULL)
2496 		s = tcp_log_addrs(inc, NULL, NULL, NULL);
2497 	else
2498 		s = (const char *)NULL;
2499 	log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2500 	    "the next %lld seconds%s%s%s\n", (long long)delta,
2501 	    (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2502 	    (s != NULL) ? ")" : "");
2503 	free(__DECONST(void *, s), M_TCPLOG);
2504 
2505 	/* Use the calculated delta to set a new pause time. */
2506 	V_tcp_syncache.pause_until = time_uptime + delta;
2507 	callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2508 	    &V_tcp_syncache);
2509 	mtx_unlock(&V_tcp_syncache.pause_mtx);
2510 }
2511 
2512 /* Evaluate whether we need to unpause. */
2513 static void
2514 syncache_unpause(void *arg)
2515 {
2516 	struct tcp_syncache *sc;
2517 	time_t delta;
2518 
2519 	sc = arg;
2520 	mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2521 	callout_deactivate(&sc->pause_co);
2522 
2523 	/*
2524 	 * Check to make sure we are not running early. If the pause
2525 	 * time has expired, then deactivate the protection.
2526 	 */
2527 	if ((delta = sc->pause_until - time_uptime) > 0)
2528 		callout_schedule(&sc->pause_co, delta * hz);
2529 	else
2530 		sc->paused = false;
2531 }
2532 
2533 /*
2534  * Exports the syncache entries to userland so that netstat can display
2535  * them alongside the other sockets.  This function is intended to be
2536  * called only from tcp_pcblist.
2537  *
2538  * Due to concurrency on an active system, the number of pcbs exported
2539  * may have no relation to max_pcbs.  max_pcbs merely indicates the
2540  * amount of space the caller allocated for this function to use.
2541  */
2542 int
2543 syncache_pcblist(struct sysctl_req *req)
2544 {
2545 	struct xtcpcb xt;
2546 	struct syncache *sc;
2547 	struct syncache_head *sch;
2548 	int error, i;
2549 
2550 	bzero(&xt, sizeof(xt));
2551 	xt.xt_len = sizeof(xt);
2552 	xt.t_state = TCPS_SYN_RECEIVED;
2553 	xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2554 	xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2555 	xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2556 	xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2557 
2558 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2559 		sch = &V_tcp_syncache.hashbase[i];
2560 		SCH_LOCK(sch);
2561 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2562 			if (sc->sc_cred != NULL &&
2563 			    cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2564 				continue;
2565 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2566 				xt.xt_inp.inp_vflag = INP_IPV6;
2567 			else
2568 				xt.xt_inp.inp_vflag = INP_IPV4;
2569 			xt.xt_encaps_port = sc->sc_port;
2570 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2571 			    sizeof (struct in_conninfo));
2572 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2573 			if (error) {
2574 				SCH_UNLOCK(sch);
2575 				return (0);
2576 			}
2577 		}
2578 		SCH_UNLOCK(sch);
2579 	}
2580 
2581 	return (0);
2582 }
2583