xref: /freebsd/sys/netinet/tcp_syncache.c (revision 49b49cda41feabe3439f7318e8bf40e3896c7bf4)
1 /*-
2  * Copyright (c) 2001 McAfee, Inc.
3  * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG
4  * All rights reserved.
5  *
6  * This software was developed for the FreeBSD Project by Jonathan Lemon
7  * and McAfee Research, the Security Research Division of McAfee, Inc. under
8  * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
9  * DARPA CHATS research program. [2001 McAfee, Inc.]
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  */
32 
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 #include "opt_inet.h"
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 #include "opt_pcbgroup.h"
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/hash.h>
44 #include <sys/refcount.h>
45 #include <sys/kernel.h>
46 #include <sys/sysctl.h>
47 #include <sys/limits.h>
48 #include <sys/lock.h>
49 #include <sys/mutex.h>
50 #include <sys/malloc.h>
51 #include <sys/mbuf.h>
52 #include <sys/proc.h>		/* for proc0 declaration */
53 #include <sys/random.h>
54 #include <sys/socket.h>
55 #include <sys/socketvar.h>
56 #include <sys/syslog.h>
57 #include <sys/ucred.h>
58 
59 #include <sys/md5.h>
60 #include <crypto/siphash/siphash.h>
61 
62 #include <vm/uma.h>
63 
64 #include <net/if.h>
65 #include <net/if_var.h>
66 #include <net/route.h>
67 #include <net/vnet.h>
68 
69 #include <netinet/in.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 #ifdef TCP_RFC7413
85 #include <netinet/tcp_fastopen.h>
86 #endif
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 #ifdef INET6
93 #include <netinet6/tcp6_var.h>
94 #endif
95 #ifdef TCP_OFFLOAD
96 #include <netinet/toecore.h>
97 #endif
98 
99 #ifdef IPSEC
100 #include <netipsec/ipsec.h>
101 #ifdef INET6
102 #include <netipsec/ipsec6.h>
103 #endif
104 #include <netipsec/key.h>
105 #endif /*IPSEC*/
106 
107 #include <machine/in_cksum.h>
108 
109 #include <security/mac/mac_framework.h>
110 
111 static VNET_DEFINE(int, tcp_syncookies) = 1;
112 #define	V_tcp_syncookies		VNET(tcp_syncookies)
113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
114     &VNET_NAME(tcp_syncookies), 0,
115     "Use TCP SYN cookies if the syncache overflows");
116 
117 static VNET_DEFINE(int, tcp_syncookiesonly) = 0;
118 #define	V_tcp_syncookiesonly		VNET(tcp_syncookiesonly)
119 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
120     &VNET_NAME(tcp_syncookiesonly), 0,
121     "Use only TCP SYN cookies");
122 
123 #ifdef TCP_OFFLOAD
124 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
125 #endif
126 
127 static void	 syncache_drop(struct syncache *, struct syncache_head *);
128 static void	 syncache_free(struct syncache *);
129 static void	 syncache_insert(struct syncache *, struct syncache_head *);
130 static int	 syncache_respond(struct syncache *, struct syncache_head *, int);
131 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
132 		    struct mbuf *m);
133 static void	 syncache_timeout(struct syncache *sc, struct syncache_head *sch,
134 		    int docallout);
135 static void	 syncache_timer(void *);
136 
137 static uint32_t	 syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
138 		    uint8_t *, uintptr_t);
139 static tcp_seq	 syncookie_generate(struct syncache_head *, struct syncache *);
140 static struct syncache
141 		*syncookie_lookup(struct in_conninfo *, struct syncache_head *,
142 		    struct syncache *, struct tcphdr *, struct tcpopt *,
143 		    struct socket *);
144 static void	 syncookie_reseed(void *);
145 #ifdef INVARIANTS
146 static int	 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
147 		    struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
148 		    struct socket *lso);
149 #endif
150 
151 /*
152  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
153  * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds,
154  * the odds are that the user has given up attempting to connect by then.
155  */
156 #define SYNCACHE_MAXREXMTS		3
157 
158 /* Arbitrary values */
159 #define TCP_SYNCACHE_HASHSIZE		512
160 #define TCP_SYNCACHE_BUCKETLIMIT	30
161 
162 static VNET_DEFINE(struct tcp_syncache, tcp_syncache);
163 #define	V_tcp_syncache			VNET(tcp_syncache)
164 
165 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0,
166     "TCP SYN cache");
167 
168 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
169     &VNET_NAME(tcp_syncache.bucket_limit), 0,
170     "Per-bucket hash limit for syncache");
171 
172 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
173     &VNET_NAME(tcp_syncache.cache_limit), 0,
174     "Overall entry limit for syncache");
175 
176 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
177     &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
178 
179 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
180     &VNET_NAME(tcp_syncache.hashsize), 0,
181     "Size of TCP syncache hashtable");
182 
183 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_VNET | CTLFLAG_RW,
184     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
185     "Limit on SYN/ACK retransmissions");
186 
187 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
188 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
189     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
190     "Send reset on socket allocation failure");
191 
192 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
193 
194 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
195 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
196 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
197 
198 /*
199  * Requires the syncache entry to be already removed from the bucket list.
200  */
201 static void
202 syncache_free(struct syncache *sc)
203 {
204 
205 	if (sc->sc_ipopts)
206 		(void) m_free(sc->sc_ipopts);
207 	if (sc->sc_cred)
208 		crfree(sc->sc_cred);
209 #ifdef MAC
210 	mac_syncache_destroy(&sc->sc_label);
211 #endif
212 
213 	uma_zfree(V_tcp_syncache.zone, sc);
214 }
215 
216 void
217 syncache_init(void)
218 {
219 	int i;
220 
221 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
222 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
223 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
224 	V_tcp_syncache.hash_secret = arc4random();
225 
226 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
227 	    &V_tcp_syncache.hashsize);
228 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
229 	    &V_tcp_syncache.bucket_limit);
230 	if (!powerof2(V_tcp_syncache.hashsize) ||
231 	    V_tcp_syncache.hashsize == 0) {
232 		printf("WARNING: syncache hash size is not a power of 2.\n");
233 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
234 	}
235 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
236 
237 	/* Set limits. */
238 	V_tcp_syncache.cache_limit =
239 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
240 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
241 	    &V_tcp_syncache.cache_limit);
242 
243 	/* Allocate the hash table. */
244 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
245 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
246 
247 #ifdef VIMAGE
248 	V_tcp_syncache.vnet = curvnet;
249 #endif
250 
251 	/* Initialize the hash buckets. */
252 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
253 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
254 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
255 			 NULL, MTX_DEF);
256 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
257 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
258 		V_tcp_syncache.hashbase[i].sch_length = 0;
259 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
260 	}
261 
262 	/* Create the syncache entry zone. */
263 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
264 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
265 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
266 	    V_tcp_syncache.cache_limit);
267 
268 	/* Start the SYN cookie reseeder callout. */
269 	callout_init(&V_tcp_syncache.secret.reseed, 1);
270 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
271 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
272 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
273 	    syncookie_reseed, &V_tcp_syncache);
274 }
275 
276 #ifdef VIMAGE
277 void
278 syncache_destroy(void)
279 {
280 	struct syncache_head *sch;
281 	struct syncache *sc, *nsc;
282 	int i;
283 
284 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
285 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
286 
287 		sch = &V_tcp_syncache.hashbase[i];
288 		callout_drain(&sch->sch_timer);
289 
290 		SCH_LOCK(sch);
291 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
292 			syncache_drop(sc, sch);
293 		SCH_UNLOCK(sch);
294 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
295 		    ("%s: sch->sch_bucket not empty", __func__));
296 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
297 		    __func__, sch->sch_length));
298 		mtx_destroy(&sch->sch_mtx);
299 	}
300 
301 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
302 	    ("%s: cache_count not 0", __func__));
303 
304 	/* Free the allocated global resources. */
305 	uma_zdestroy(V_tcp_syncache.zone);
306 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
307 
308 	callout_drain(&V_tcp_syncache.secret.reseed);
309 }
310 #endif
311 
312 /*
313  * Inserts a syncache entry into the specified bucket row.
314  * Locks and unlocks the syncache_head autonomously.
315  */
316 static void
317 syncache_insert(struct syncache *sc, struct syncache_head *sch)
318 {
319 	struct syncache *sc2;
320 
321 	SCH_LOCK(sch);
322 
323 	/*
324 	 * Make sure that we don't overflow the per-bucket limit.
325 	 * If the bucket is full, toss the oldest element.
326 	 */
327 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
328 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
329 			("sch->sch_length incorrect"));
330 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
331 		syncache_drop(sc2, sch);
332 		TCPSTAT_INC(tcps_sc_bucketoverflow);
333 	}
334 
335 	/* Put it into the bucket. */
336 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
337 	sch->sch_length++;
338 
339 #ifdef TCP_OFFLOAD
340 	if (ADDED_BY_TOE(sc)) {
341 		struct toedev *tod = sc->sc_tod;
342 
343 		tod->tod_syncache_added(tod, sc->sc_todctx);
344 	}
345 #endif
346 
347 	/* Reinitialize the bucket row's timer. */
348 	if (sch->sch_length == 1)
349 		sch->sch_nextc = ticks + INT_MAX;
350 	syncache_timeout(sc, sch, 1);
351 
352 	SCH_UNLOCK(sch);
353 
354 	TCPSTAT_INC(tcps_sc_added);
355 }
356 
357 /*
358  * Remove and free entry from syncache bucket row.
359  * Expects locked syncache head.
360  */
361 static void
362 syncache_drop(struct syncache *sc, struct syncache_head *sch)
363 {
364 
365 	SCH_LOCK_ASSERT(sch);
366 
367 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
368 	sch->sch_length--;
369 
370 #ifdef TCP_OFFLOAD
371 	if (ADDED_BY_TOE(sc)) {
372 		struct toedev *tod = sc->sc_tod;
373 
374 		tod->tod_syncache_removed(tod, sc->sc_todctx);
375 	}
376 #endif
377 
378 	syncache_free(sc);
379 }
380 
381 /*
382  * Engage/reengage time on bucket row.
383  */
384 static void
385 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
386 {
387 	sc->sc_rxttime = ticks +
388 		TCPTV_RTOBASE * (tcp_syn_backoff[sc->sc_rxmits]);
389 	sc->sc_rxmits++;
390 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
391 		sch->sch_nextc = sc->sc_rxttime;
392 		if (docallout)
393 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
394 			    syncache_timer, (void *)sch);
395 	}
396 }
397 
398 /*
399  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
400  * If we have retransmitted an entry the maximum number of times, expire it.
401  * One separate timer for each bucket row.
402  */
403 static void
404 syncache_timer(void *xsch)
405 {
406 	struct syncache_head *sch = (struct syncache_head *)xsch;
407 	struct syncache *sc, *nsc;
408 	int tick = ticks;
409 	char *s;
410 
411 	CURVNET_SET(sch->sch_sc->vnet);
412 
413 	/* NB: syncache_head has already been locked by the callout. */
414 	SCH_LOCK_ASSERT(sch);
415 
416 	/*
417 	 * In the following cycle we may remove some entries and/or
418 	 * advance some timeouts, so re-initialize the bucket timer.
419 	 */
420 	sch->sch_nextc = tick + INT_MAX;
421 
422 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
423 		/*
424 		 * We do not check if the listen socket still exists
425 		 * and accept the case where the listen socket may be
426 		 * gone by the time we resend the SYN/ACK.  We do
427 		 * not expect this to happens often. If it does,
428 		 * then the RST will be sent by the time the remote
429 		 * host does the SYN/ACK->ACK.
430 		 */
431 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
432 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
433 				sch->sch_nextc = sc->sc_rxttime;
434 			continue;
435 		}
436 		if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
437 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
438 				log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
439 				    "giving up and removing syncache entry\n",
440 				    s, __func__);
441 				free(s, M_TCPLOG);
442 			}
443 			syncache_drop(sc, sch);
444 			TCPSTAT_INC(tcps_sc_stale);
445 			continue;
446 		}
447 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
448 			log(LOG_DEBUG, "%s; %s: Response timeout, "
449 			    "retransmitting (%u) SYN|ACK\n",
450 			    s, __func__, sc->sc_rxmits);
451 			free(s, M_TCPLOG);
452 		}
453 
454 		syncache_respond(sc, sch, 1);
455 		TCPSTAT_INC(tcps_sc_retransmitted);
456 		syncache_timeout(sc, sch, 0);
457 	}
458 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
459 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
460 			syncache_timer, (void *)(sch));
461 	CURVNET_RESTORE();
462 }
463 
464 /*
465  * Find an entry in the syncache.
466  * Returns always with locked syncache_head plus a matching entry or NULL.
467  */
468 static struct syncache *
469 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
470 {
471 	struct syncache *sc;
472 	struct syncache_head *sch;
473 	uint32_t hash;
474 
475 	/*
476 	 * The hash is built on foreign port + local port + foreign address.
477 	 * We rely on the fact that struct in_conninfo starts with 16 bits
478 	 * of foreign port, then 16 bits of local port then followed by 128
479 	 * bits of foreign address.  In case of IPv4 address, the first 3
480 	 * 32-bit words of the address always are zeroes.
481 	 */
482 	hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
483 	    V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
484 
485 	sch = &V_tcp_syncache.hashbase[hash];
486 	*schp = sch;
487 	SCH_LOCK(sch);
488 
489 	/* Circle through bucket row to find matching entry. */
490 	TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
491 		if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
492 		    sizeof(struct in_endpoints)) == 0)
493 			break;
494 
495 	return (sc);	/* Always returns with locked sch. */
496 }
497 
498 /*
499  * This function is called when we get a RST for a
500  * non-existent connection, so that we can see if the
501  * connection is in the syn cache.  If it is, zap it.
502  */
503 void
504 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
505 {
506 	struct syncache *sc;
507 	struct syncache_head *sch;
508 	char *s = NULL;
509 
510 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
511 	SCH_LOCK_ASSERT(sch);
512 
513 	/*
514 	 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
515 	 * See RFC 793 page 65, section SEGMENT ARRIVES.
516 	 */
517 	if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
518 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
519 			log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
520 			    "FIN flag set, segment ignored\n", s, __func__);
521 		TCPSTAT_INC(tcps_badrst);
522 		goto done;
523 	}
524 
525 	/*
526 	 * No corresponding connection was found in syncache.
527 	 * If syncookies are enabled and possibly exclusively
528 	 * used, or we are under memory pressure, a valid RST
529 	 * may not find a syncache entry.  In that case we're
530 	 * done and no SYN|ACK retransmissions will happen.
531 	 * Otherwise the RST was misdirected or spoofed.
532 	 */
533 	if (sc == NULL) {
534 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
535 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
536 			    "syncache entry (possibly syncookie only), "
537 			    "segment ignored\n", s, __func__);
538 		TCPSTAT_INC(tcps_badrst);
539 		goto done;
540 	}
541 
542 	/*
543 	 * If the RST bit is set, check the sequence number to see
544 	 * if this is a valid reset segment.
545 	 * RFC 793 page 37:
546 	 *   In all states except SYN-SENT, all reset (RST) segments
547 	 *   are validated by checking their SEQ-fields.  A reset is
548 	 *   valid if its sequence number is in the window.
549 	 *
550 	 *   The sequence number in the reset segment is normally an
551 	 *   echo of our outgoing acknowlegement numbers, but some hosts
552 	 *   send a reset with the sequence number at the rightmost edge
553 	 *   of our receive window, and we have to handle this case.
554 	 */
555 	if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
556 	    SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
557 		syncache_drop(sc, sch);
558 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
559 			log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, "
560 			    "connection attempt aborted by remote endpoint\n",
561 			    s, __func__);
562 		TCPSTAT_INC(tcps_sc_reset);
563 	} else {
564 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
565 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
566 			    "IRS %u (+WND %u), segment ignored\n",
567 			    s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd);
568 		TCPSTAT_INC(tcps_badrst);
569 	}
570 
571 done:
572 	if (s != NULL)
573 		free(s, M_TCPLOG);
574 	SCH_UNLOCK(sch);
575 }
576 
577 void
578 syncache_badack(struct in_conninfo *inc)
579 {
580 	struct syncache *sc;
581 	struct syncache_head *sch;
582 
583 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
584 	SCH_LOCK_ASSERT(sch);
585 	if (sc != NULL) {
586 		syncache_drop(sc, sch);
587 		TCPSTAT_INC(tcps_sc_badack);
588 	}
589 	SCH_UNLOCK(sch);
590 }
591 
592 void
593 syncache_unreach(struct in_conninfo *inc, struct tcphdr *th)
594 {
595 	struct syncache *sc;
596 	struct syncache_head *sch;
597 
598 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
599 	SCH_LOCK_ASSERT(sch);
600 	if (sc == NULL)
601 		goto done;
602 
603 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
604 	if (ntohl(th->th_seq) != sc->sc_iss)
605 		goto done;
606 
607 	/*
608 	 * If we've rertransmitted 3 times and this is our second error,
609 	 * we remove the entry.  Otherwise, we allow it to continue on.
610 	 * This prevents us from incorrectly nuking an entry during a
611 	 * spurious network outage.
612 	 *
613 	 * See tcp_notify().
614 	 */
615 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
616 		sc->sc_flags |= SCF_UNREACH;
617 		goto done;
618 	}
619 	syncache_drop(sc, sch);
620 	TCPSTAT_INC(tcps_sc_unreach);
621 done:
622 	SCH_UNLOCK(sch);
623 }
624 
625 /*
626  * Build a new TCP socket structure from a syncache entry.
627  *
628  * On success return the newly created socket with its underlying inp locked.
629  */
630 static struct socket *
631 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
632 {
633 	struct tcp_function_block *blk;
634 	struct inpcb *inp = NULL;
635 	struct socket *so;
636 	struct tcpcb *tp;
637 	int error;
638 	char *s;
639 
640 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
641 
642 	/*
643 	 * Ok, create the full blown connection, and set things up
644 	 * as they would have been set up if we had created the
645 	 * connection when the SYN arrived.  If we can't create
646 	 * the connection, abort it.
647 	 */
648 	so = sonewconn(lso, 0);
649 	if (so == NULL) {
650 		/*
651 		 * Drop the connection; we will either send a RST or
652 		 * have the peer retransmit its SYN again after its
653 		 * RTO and try again.
654 		 */
655 		TCPSTAT_INC(tcps_listendrop);
656 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
657 			log(LOG_DEBUG, "%s; %s: Socket create failed "
658 			    "due to limits or memory shortage\n",
659 			    s, __func__);
660 			free(s, M_TCPLOG);
661 		}
662 		goto abort2;
663 	}
664 #ifdef MAC
665 	mac_socketpeer_set_from_mbuf(m, so);
666 #endif
667 
668 	inp = sotoinpcb(so);
669 	inp->inp_inc.inc_fibnum = so->so_fibnum;
670 	INP_WLOCK(inp);
671 	/*
672 	 * Exclusive pcbinfo lock is not required in syncache socket case even
673 	 * if two inpcb locks can be acquired simultaneously:
674 	 *  - the inpcb in LISTEN state,
675 	 *  - the newly created inp.
676 	 *
677 	 * In this case, an inp cannot be at same time in LISTEN state and
678 	 * just created by an accept() call.
679 	 */
680 	INP_HASH_WLOCK(&V_tcbinfo);
681 
682 	/* Insert new socket into PCB hash list. */
683 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
684 #ifdef INET6
685 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
686 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
687 	} else {
688 		inp->inp_vflag &= ~INP_IPV6;
689 		inp->inp_vflag |= INP_IPV4;
690 #endif
691 		inp->inp_laddr = sc->sc_inc.inc_laddr;
692 #ifdef INET6
693 	}
694 #endif
695 
696 	/*
697 	 * If there's an mbuf and it has a flowid, then let's initialise the
698 	 * inp with that particular flowid.
699 	 */
700 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
701 		inp->inp_flowid = m->m_pkthdr.flowid;
702 		inp->inp_flowtype = M_HASHTYPE_GET(m);
703 	}
704 
705 	/*
706 	 * Install in the reservation hash table for now, but don't yet
707 	 * install a connection group since the full 4-tuple isn't yet
708 	 * configured.
709 	 */
710 	inp->inp_lport = sc->sc_inc.inc_lport;
711 	if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
712 		/*
713 		 * Undo the assignments above if we failed to
714 		 * put the PCB on the hash lists.
715 		 */
716 #ifdef INET6
717 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
718 			inp->in6p_laddr = in6addr_any;
719 		else
720 #endif
721 			inp->inp_laddr.s_addr = INADDR_ANY;
722 		inp->inp_lport = 0;
723 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
724 			log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
725 			    "with error %i\n",
726 			    s, __func__, error);
727 			free(s, M_TCPLOG);
728 		}
729 		INP_HASH_WUNLOCK(&V_tcbinfo);
730 		goto abort;
731 	}
732 #ifdef IPSEC
733 	/* Copy old policy into new socket's. */
734 	if (ipsec_copy_policy(sotoinpcb(lso)->inp_sp, inp->inp_sp))
735 		printf("syncache_socket: could not copy policy\n");
736 #endif
737 #ifdef INET6
738 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
739 		struct inpcb *oinp = sotoinpcb(lso);
740 		struct in6_addr laddr6;
741 		struct sockaddr_in6 sin6;
742 		/*
743 		 * Inherit socket options from the listening socket.
744 		 * Note that in6p_inputopts are not (and should not be)
745 		 * copied, since it stores previously received options and is
746 		 * used to detect if each new option is different than the
747 		 * previous one and hence should be passed to a user.
748 		 * If we copied in6p_inputopts, a user would not be able to
749 		 * receive options just after calling the accept system call.
750 		 */
751 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
752 		if (oinp->in6p_outputopts)
753 			inp->in6p_outputopts =
754 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
755 
756 		sin6.sin6_family = AF_INET6;
757 		sin6.sin6_len = sizeof(sin6);
758 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
759 		sin6.sin6_port = sc->sc_inc.inc_fport;
760 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
761 		laddr6 = inp->in6p_laddr;
762 		if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
763 			inp->in6p_laddr = sc->sc_inc.inc6_laddr;
764 		if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
765 		    thread0.td_ucred, m)) != 0) {
766 			inp->in6p_laddr = laddr6;
767 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
768 				log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
769 				    "with error %i\n",
770 				    s, __func__, error);
771 				free(s, M_TCPLOG);
772 			}
773 			INP_HASH_WUNLOCK(&V_tcbinfo);
774 			goto abort;
775 		}
776 		/* Override flowlabel from in6_pcbconnect. */
777 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
778 		inp->inp_flow |= sc->sc_flowlabel;
779 	}
780 #endif /* INET6 */
781 #if defined(INET) && defined(INET6)
782 	else
783 #endif
784 #ifdef INET
785 	{
786 		struct in_addr laddr;
787 		struct sockaddr_in sin;
788 
789 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
790 
791 		if (inp->inp_options == NULL) {
792 			inp->inp_options = sc->sc_ipopts;
793 			sc->sc_ipopts = NULL;
794 		}
795 
796 		sin.sin_family = AF_INET;
797 		sin.sin_len = sizeof(sin);
798 		sin.sin_addr = sc->sc_inc.inc_faddr;
799 		sin.sin_port = sc->sc_inc.inc_fport;
800 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
801 		laddr = inp->inp_laddr;
802 		if (inp->inp_laddr.s_addr == INADDR_ANY)
803 			inp->inp_laddr = sc->sc_inc.inc_laddr;
804 		if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
805 		    thread0.td_ucred, m)) != 0) {
806 			inp->inp_laddr = laddr;
807 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
808 				log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
809 				    "with error %i\n",
810 				    s, __func__, error);
811 				free(s, M_TCPLOG);
812 			}
813 			INP_HASH_WUNLOCK(&V_tcbinfo);
814 			goto abort;
815 		}
816 	}
817 #endif /* INET */
818 	INP_HASH_WUNLOCK(&V_tcbinfo);
819 	tp = intotcpcb(inp);
820 	tcp_state_change(tp, TCPS_SYN_RECEIVED);
821 	tp->iss = sc->sc_iss;
822 	tp->irs = sc->sc_irs;
823 	tcp_rcvseqinit(tp);
824 	tcp_sendseqinit(tp);
825 	blk = sototcpcb(lso)->t_fb;
826 	if (blk != tp->t_fb) {
827 		/*
828 		 * Our parents t_fb was not the default,
829 		 * we need to release our ref on tp->t_fb and
830 		 * pickup one on the new entry.
831 		 */
832 		struct tcp_function_block *rblk;
833 
834 		rblk = find_and_ref_tcp_fb(blk);
835 		KASSERT(rblk != NULL,
836 		    ("cannot find blk %p out of syncache?", blk));
837 		if (tp->t_fb->tfb_tcp_fb_fini)
838 			(*tp->t_fb->tfb_tcp_fb_fini)(tp);
839 		refcount_release(&tp->t_fb->tfb_refcnt);
840 		tp->t_fb = rblk;
841 		if (tp->t_fb->tfb_tcp_fb_init) {
842 			(*tp->t_fb->tfb_tcp_fb_init)(tp);
843 		}
844 	}
845 	tp->snd_wl1 = sc->sc_irs;
846 	tp->snd_max = tp->iss + 1;
847 	tp->snd_nxt = tp->iss + 1;
848 	tp->rcv_up = sc->sc_irs + 1;
849 	tp->rcv_wnd = sc->sc_wnd;
850 	tp->rcv_adv += tp->rcv_wnd;
851 	tp->last_ack_sent = tp->rcv_nxt;
852 
853 	tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
854 	if (sc->sc_flags & SCF_NOOPT)
855 		tp->t_flags |= TF_NOOPT;
856 	else {
857 		if (sc->sc_flags & SCF_WINSCALE) {
858 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
859 			tp->snd_scale = sc->sc_requested_s_scale;
860 			tp->request_r_scale = sc->sc_requested_r_scale;
861 		}
862 		if (sc->sc_flags & SCF_TIMESTAMP) {
863 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
864 			tp->ts_recent = sc->sc_tsreflect;
865 			tp->ts_recent_age = tcp_ts_getticks();
866 			tp->ts_offset = sc->sc_tsoff;
867 		}
868 #ifdef TCP_SIGNATURE
869 		if (sc->sc_flags & SCF_SIGNATURE)
870 			tp->t_flags |= TF_SIGNATURE;
871 #endif
872 		if (sc->sc_flags & SCF_SACK)
873 			tp->t_flags |= TF_SACK_PERMIT;
874 	}
875 
876 	if (sc->sc_flags & SCF_ECN)
877 		tp->t_flags |= TF_ECN_PERMIT;
878 
879 	/*
880 	 * Set up MSS and get cached values from tcp_hostcache.
881 	 * This might overwrite some of the defaults we just set.
882 	 */
883 	tcp_mss(tp, sc->sc_peer_mss);
884 
885 	/*
886 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
887 	 * limited to one segment in cc_conn_init().
888 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
889 	 */
890 	if (sc->sc_rxmits > 1)
891 		tp->snd_cwnd = 1;
892 
893 #ifdef TCP_OFFLOAD
894 	/*
895 	 * Allow a TOE driver to install its hooks.  Note that we hold the
896 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
897 	 * new connection before the TOE driver has done its thing.
898 	 */
899 	if (ADDED_BY_TOE(sc)) {
900 		struct toedev *tod = sc->sc_tod;
901 
902 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
903 	}
904 #endif
905 	/*
906 	 * Copy and activate timers.
907 	 */
908 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
909 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
910 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
911 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
912 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
913 
914 	soisconnected(so);
915 
916 	TCPSTAT_INC(tcps_accepts);
917 	return (so);
918 
919 abort:
920 	INP_WUNLOCK(inp);
921 abort2:
922 	if (so != NULL)
923 		soabort(so);
924 	return (NULL);
925 }
926 
927 /*
928  * This function gets called when we receive an ACK for a
929  * socket in the LISTEN state.  We look up the connection
930  * in the syncache, and if its there, we pull it out of
931  * the cache and turn it into a full-blown connection in
932  * the SYN-RECEIVED state.
933  *
934  * On syncache_socket() success the newly created socket
935  * has its underlying inp locked.
936  */
937 int
938 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
939     struct socket **lsop, struct mbuf *m)
940 {
941 	struct syncache *sc;
942 	struct syncache_head *sch;
943 	struct syncache scs;
944 	char *s;
945 
946 	/*
947 	 * Global TCP locks are held because we manipulate the PCB lists
948 	 * and create a new socket.
949 	 */
950 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
951 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
952 	    ("%s: can handle only ACK", __func__));
953 
954 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
955 	SCH_LOCK_ASSERT(sch);
956 
957 #ifdef INVARIANTS
958 	/*
959 	 * Test code for syncookies comparing the syncache stored
960 	 * values with the reconstructed values from the cookie.
961 	 */
962 	if (sc != NULL)
963 		syncookie_cmp(inc, sch, sc, th, to, *lsop);
964 #endif
965 
966 	if (sc == NULL) {
967 		/*
968 		 * There is no syncache entry, so see if this ACK is
969 		 * a returning syncookie.  To do this, first:
970 		 *  A. See if this socket has had a syncache entry dropped in
971 		 *     the past.  We don't want to accept a bogus syncookie
972 		 *     if we've never received a SYN.
973 		 *  B. check that the syncookie is valid.  If it is, then
974 		 *     cobble up a fake syncache entry, and return.
975 		 */
976 		if (!V_tcp_syncookies) {
977 			SCH_UNLOCK(sch);
978 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
979 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
980 				    "segment rejected (syncookies disabled)\n",
981 				    s, __func__);
982 			goto failed;
983 		}
984 		bzero(&scs, sizeof(scs));
985 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
986 		SCH_UNLOCK(sch);
987 		if (sc == NULL) {
988 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
989 				log(LOG_DEBUG, "%s; %s: Segment failed "
990 				    "SYNCOOKIE authentication, segment rejected "
991 				    "(probably spoofed)\n", s, __func__);
992 			goto failed;
993 		}
994 	} else {
995 		/* Pull out the entry to unlock the bucket row. */
996 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
997 		sch->sch_length--;
998 #ifdef TCP_OFFLOAD
999 		if (ADDED_BY_TOE(sc)) {
1000 			struct toedev *tod = sc->sc_tod;
1001 
1002 			tod->tod_syncache_removed(tod, sc->sc_todctx);
1003 		}
1004 #endif
1005 		SCH_UNLOCK(sch);
1006 	}
1007 
1008 	/*
1009 	 * Segment validation:
1010 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1011 	 */
1012 	if (th->th_ack != sc->sc_iss + 1) {
1013 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1014 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1015 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1016 		goto failed;
1017 	}
1018 
1019 	/*
1020 	 * The SEQ must fall in the window starting at the received
1021 	 * initial receive sequence number + 1 (the SYN).
1022 	 */
1023 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1024 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1025 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1026 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1027 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1028 		goto failed;
1029 	}
1030 
1031 	/*
1032 	 * If timestamps were not negotiated during SYN/ACK they
1033 	 * must not appear on any segment during this session.
1034 	 */
1035 	if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1036 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1037 			log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1038 			    "segment rejected\n", s, __func__);
1039 		goto failed;
1040 	}
1041 
1042 	/*
1043 	 * If timestamps were negotiated during SYN/ACK they should
1044 	 * appear on every segment during this session.
1045 	 * XXXAO: This is only informal as there have been unverified
1046 	 * reports of non-compliants stacks.
1047 	 */
1048 	if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1049 		if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1050 			log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1051 			    "no action\n", s, __func__);
1052 			free(s, M_TCPLOG);
1053 			s = NULL;
1054 		}
1055 	}
1056 
1057 	/*
1058 	 * If timestamps were negotiated the reflected timestamp
1059 	 * must be equal to what we actually sent in the SYN|ACK.
1060 	 */
1061 	if ((to->to_flags & TOF_TS) && to->to_tsecr != sc->sc_ts) {
1062 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1063 			log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, "
1064 			    "segment rejected\n",
1065 			    s, __func__, to->to_tsecr, sc->sc_ts);
1066 		goto failed;
1067 	}
1068 
1069 	*lsop = syncache_socket(sc, *lsop, m);
1070 
1071 	if (*lsop == NULL)
1072 		TCPSTAT_INC(tcps_sc_aborted);
1073 	else
1074 		TCPSTAT_INC(tcps_sc_completed);
1075 
1076 /* how do we find the inp for the new socket? */
1077 	if (sc != &scs)
1078 		syncache_free(sc);
1079 	return (1);
1080 failed:
1081 	if (sc != NULL && sc != &scs)
1082 		syncache_free(sc);
1083 	if (s != NULL)
1084 		free(s, M_TCPLOG);
1085 	*lsop = NULL;
1086 	return (0);
1087 }
1088 
1089 #ifdef TCP_RFC7413
1090 static void
1091 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
1092     uint64_t response_cookie)
1093 {
1094 	struct inpcb *inp;
1095 	struct tcpcb *tp;
1096 	unsigned int *pending_counter;
1097 
1098 	/*
1099 	 * Global TCP locks are held because we manipulate the PCB lists
1100 	 * and create a new socket.
1101 	 */
1102 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1103 
1104 	pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
1105 	*lsop = syncache_socket(sc, *lsop, m);
1106 	if (*lsop == NULL) {
1107 		TCPSTAT_INC(tcps_sc_aborted);
1108 		atomic_subtract_int(pending_counter, 1);
1109 	} else {
1110 		inp = sotoinpcb(*lsop);
1111 		tp = intotcpcb(inp);
1112 		tp->t_flags |= TF_FASTOPEN;
1113 		tp->t_tfo_cookie = response_cookie;
1114 		tp->snd_max = tp->iss;
1115 		tp->snd_nxt = tp->iss;
1116 		tp->t_tfo_pending = pending_counter;
1117 		TCPSTAT_INC(tcps_sc_completed);
1118 	}
1119 }
1120 #endif /* TCP_RFC7413 */
1121 
1122 /*
1123  * Given a LISTEN socket and an inbound SYN request, add
1124  * this to the syn cache, and send back a segment:
1125  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1126  * to the source.
1127  *
1128  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1129  * Doing so would require that we hold onto the data and deliver it
1130  * to the application.  However, if we are the target of a SYN-flood
1131  * DoS attack, an attacker could send data which would eventually
1132  * consume all available buffer space if it were ACKed.  By not ACKing
1133  * the data, we avoid this DoS scenario.
1134  *
1135  * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1136  * cookie is processed, V_tcp_fastopen_enabled set to true, and the
1137  * TCP_FASTOPEN socket option is set.  In this case, a new socket is created
1138  * and returned via lsop, the mbuf is not freed so that tcp_input() can
1139  * queue its data to the socket, and 1 is returned to indicate the
1140  * TFO-socket-creation path was taken.
1141  */
1142 int
1143 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1144     struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1145     void *todctx)
1146 {
1147 	struct tcpcb *tp;
1148 	struct socket *so;
1149 	struct syncache *sc = NULL;
1150 	struct syncache_head *sch;
1151 	struct mbuf *ipopts = NULL;
1152 	u_int ltflags;
1153 	int win, sb_hiwat, ip_ttl, ip_tos;
1154 	char *s;
1155 	int rv = 0;
1156 #ifdef INET6
1157 	int autoflowlabel = 0;
1158 #endif
1159 #ifdef MAC
1160 	struct label *maclabel;
1161 #endif
1162 	struct syncache scs;
1163 	struct ucred *cred;
1164 #ifdef TCP_RFC7413
1165 	uint64_t tfo_response_cookie;
1166 	int tfo_cookie_valid = 0;
1167 	int tfo_response_cookie_valid = 0;
1168 #endif
1169 
1170 	INP_WLOCK_ASSERT(inp);			/* listen socket */
1171 	KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1172 	    ("%s: unexpected tcp flags", __func__));
1173 
1174 	/*
1175 	 * Combine all so/tp operations very early to drop the INP lock as
1176 	 * soon as possible.
1177 	 */
1178 	so = *lsop;
1179 	tp = sototcpcb(so);
1180 	cred = crhold(so->so_cred);
1181 
1182 #ifdef INET6
1183 	if ((inc->inc_flags & INC_ISIPV6) &&
1184 	    (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1185 		autoflowlabel = 1;
1186 #endif
1187 	ip_ttl = inp->inp_ip_ttl;
1188 	ip_tos = inp->inp_ip_tos;
1189 	win = sbspace(&so->so_rcv);
1190 	sb_hiwat = so->so_rcv.sb_hiwat;
1191 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1192 
1193 #ifdef TCP_RFC7413
1194 	if (V_tcp_fastopen_enabled && (tp->t_flags & TF_FASTOPEN) &&
1195 	    (tp->t_tfo_pending != NULL) && (to->to_flags & TOF_FASTOPEN)) {
1196 		/*
1197 		 * Limit the number of pending TFO connections to
1198 		 * approximately half of the queue limit.  This prevents TFO
1199 		 * SYN floods from starving the service by filling the
1200 		 * listen queue with bogus TFO connections.
1201 		 */
1202 		if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1203 		    (so->so_qlimit / 2)) {
1204 			int result;
1205 
1206 			result = tcp_fastopen_check_cookie(inc,
1207 			    to->to_tfo_cookie, to->to_tfo_len,
1208 			    &tfo_response_cookie);
1209 			tfo_cookie_valid = (result > 0);
1210 			tfo_response_cookie_valid = (result >= 0);
1211 		} else
1212 			atomic_subtract_int(tp->t_tfo_pending, 1);
1213 	}
1214 #endif
1215 
1216 	/* By the time we drop the lock these should no longer be used. */
1217 	so = NULL;
1218 	tp = NULL;
1219 
1220 #ifdef MAC
1221 	if (mac_syncache_init(&maclabel) != 0) {
1222 		INP_WUNLOCK(inp);
1223 		goto done;
1224 	} else
1225 		mac_syncache_create(maclabel, inp);
1226 #endif
1227 #ifdef TCP_RFC7413
1228 	if (!tfo_cookie_valid)
1229 #endif
1230 		INP_WUNLOCK(inp);
1231 
1232 	/*
1233 	 * Remember the IP options, if any.
1234 	 */
1235 #ifdef INET6
1236 	if (!(inc->inc_flags & INC_ISIPV6))
1237 #endif
1238 #ifdef INET
1239 		ipopts = (m) ? ip_srcroute(m) : NULL;
1240 #else
1241 		ipopts = NULL;
1242 #endif
1243 
1244 	/*
1245 	 * See if we already have an entry for this connection.
1246 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1247 	 *
1248 	 * XXX: should the syncache be re-initialized with the contents
1249 	 * of the new SYN here (which may have different options?)
1250 	 *
1251 	 * XXX: We do not check the sequence number to see if this is a
1252 	 * real retransmit or a new connection attempt.  The question is
1253 	 * how to handle such a case; either ignore it as spoofed, or
1254 	 * drop the current entry and create a new one?
1255 	 */
1256 	sc = syncache_lookup(inc, &sch);	/* returns locked entry */
1257 	SCH_LOCK_ASSERT(sch);
1258 	if (sc != NULL) {
1259 #ifdef TCP_RFC7413
1260 		if (tfo_cookie_valid)
1261 			INP_WUNLOCK(inp);
1262 #endif
1263 		TCPSTAT_INC(tcps_sc_dupsyn);
1264 		if (ipopts) {
1265 			/*
1266 			 * If we were remembering a previous source route,
1267 			 * forget it and use the new one we've been given.
1268 			 */
1269 			if (sc->sc_ipopts)
1270 				(void) m_free(sc->sc_ipopts);
1271 			sc->sc_ipopts = ipopts;
1272 		}
1273 		/*
1274 		 * Update timestamp if present.
1275 		 */
1276 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1277 			sc->sc_tsreflect = to->to_tsval;
1278 		else
1279 			sc->sc_flags &= ~SCF_TIMESTAMP;
1280 #ifdef MAC
1281 		/*
1282 		 * Since we have already unconditionally allocated label
1283 		 * storage, free it up.  The syncache entry will already
1284 		 * have an initialized label we can use.
1285 		 */
1286 		mac_syncache_destroy(&maclabel);
1287 #endif
1288 		/* Retransmit SYN|ACK and reset retransmit count. */
1289 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1290 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1291 			    "resetting timer and retransmitting SYN|ACK\n",
1292 			    s, __func__);
1293 			free(s, M_TCPLOG);
1294 		}
1295 		if (syncache_respond(sc, sch, 1) == 0) {
1296 			sc->sc_rxmits = 0;
1297 			syncache_timeout(sc, sch, 1);
1298 			TCPSTAT_INC(tcps_sndacks);
1299 			TCPSTAT_INC(tcps_sndtotal);
1300 		}
1301 		SCH_UNLOCK(sch);
1302 		goto done;
1303 	}
1304 
1305 #ifdef TCP_RFC7413
1306 	if (tfo_cookie_valid) {
1307 		bzero(&scs, sizeof(scs));
1308 		sc = &scs;
1309 		goto skip_alloc;
1310 	}
1311 #endif
1312 
1313 	sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1314 	if (sc == NULL) {
1315 		/*
1316 		 * The zone allocator couldn't provide more entries.
1317 		 * Treat this as if the cache was full; drop the oldest
1318 		 * entry and insert the new one.
1319 		 */
1320 		TCPSTAT_INC(tcps_sc_zonefail);
1321 		if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL)
1322 			syncache_drop(sc, sch);
1323 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1324 		if (sc == NULL) {
1325 			if (V_tcp_syncookies) {
1326 				bzero(&scs, sizeof(scs));
1327 				sc = &scs;
1328 			} else {
1329 				SCH_UNLOCK(sch);
1330 				if (ipopts)
1331 					(void) m_free(ipopts);
1332 				goto done;
1333 			}
1334 		}
1335 	}
1336 
1337 #ifdef TCP_RFC7413
1338 skip_alloc:
1339 	if (!tfo_cookie_valid && tfo_response_cookie_valid)
1340 		sc->sc_tfo_cookie = &tfo_response_cookie;
1341 #endif
1342 
1343 	/*
1344 	 * Fill in the syncache values.
1345 	 */
1346 #ifdef MAC
1347 	sc->sc_label = maclabel;
1348 #endif
1349 	sc->sc_cred = cred;
1350 	cred = NULL;
1351 	sc->sc_ipopts = ipopts;
1352 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1353 #ifdef INET6
1354 	if (!(inc->inc_flags & INC_ISIPV6))
1355 #endif
1356 	{
1357 		sc->sc_ip_tos = ip_tos;
1358 		sc->sc_ip_ttl = ip_ttl;
1359 	}
1360 #ifdef TCP_OFFLOAD
1361 	sc->sc_tod = tod;
1362 	sc->sc_todctx = todctx;
1363 #endif
1364 	sc->sc_irs = th->th_seq;
1365 	sc->sc_iss = arc4random();
1366 	sc->sc_flags = 0;
1367 	sc->sc_flowlabel = 0;
1368 
1369 	/*
1370 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1371 	 * win was derived from socket earlier in the function.
1372 	 */
1373 	win = imax(win, 0);
1374 	win = imin(win, TCP_MAXWIN);
1375 	sc->sc_wnd = win;
1376 
1377 	if (V_tcp_do_rfc1323) {
1378 		/*
1379 		 * A timestamp received in a SYN makes
1380 		 * it ok to send timestamp requests and replies.
1381 		 */
1382 		if (to->to_flags & TOF_TS) {
1383 			sc->sc_tsreflect = to->to_tsval;
1384 			sc->sc_ts = tcp_ts_getticks();
1385 			sc->sc_flags |= SCF_TIMESTAMP;
1386 		}
1387 		if (to->to_flags & TOF_SCALE) {
1388 			int wscale = 0;
1389 
1390 			/*
1391 			 * Pick the smallest possible scaling factor that
1392 			 * will still allow us to scale up to sb_max, aka
1393 			 * kern.ipc.maxsockbuf.
1394 			 *
1395 			 * We do this because there are broken firewalls that
1396 			 * will corrupt the window scale option, leading to
1397 			 * the other endpoint believing that our advertised
1398 			 * window is unscaled.  At scale factors larger than
1399 			 * 5 the unscaled window will drop below 1500 bytes,
1400 			 * leading to serious problems when traversing these
1401 			 * broken firewalls.
1402 			 *
1403 			 * With the default maxsockbuf of 256K, a scale factor
1404 			 * of 3 will be chosen by this algorithm.  Those who
1405 			 * choose a larger maxsockbuf should watch out
1406 			 * for the compatiblity problems mentioned above.
1407 			 *
1408 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1409 			 * or <SYN,ACK>) segment itself is never scaled.
1410 			 */
1411 			while (wscale < TCP_MAX_WINSHIFT &&
1412 			    (TCP_MAXWIN << wscale) < sb_max)
1413 				wscale++;
1414 			sc->sc_requested_r_scale = wscale;
1415 			sc->sc_requested_s_scale = to->to_wscale;
1416 			sc->sc_flags |= SCF_WINSCALE;
1417 		}
1418 	}
1419 #ifdef TCP_SIGNATURE
1420 	/*
1421 	 * If listening socket requested TCP digests, OR received SYN
1422 	 * contains the option, flag this in the syncache so that
1423 	 * syncache_respond() will do the right thing with the SYN+ACK.
1424 	 */
1425 	if (to->to_flags & TOF_SIGNATURE || ltflags & TF_SIGNATURE)
1426 		sc->sc_flags |= SCF_SIGNATURE;
1427 #endif
1428 	if (to->to_flags & TOF_SACKPERM)
1429 		sc->sc_flags |= SCF_SACK;
1430 	if (to->to_flags & TOF_MSS)
1431 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1432 	if (ltflags & TF_NOOPT)
1433 		sc->sc_flags |= SCF_NOOPT;
1434 	if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1435 		sc->sc_flags |= SCF_ECN;
1436 
1437 	if (V_tcp_syncookies)
1438 		sc->sc_iss = syncookie_generate(sch, sc);
1439 #ifdef INET6
1440 	if (autoflowlabel) {
1441 		if (V_tcp_syncookies)
1442 			sc->sc_flowlabel = sc->sc_iss;
1443 		else
1444 			sc->sc_flowlabel = ip6_randomflowlabel();
1445 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1446 	}
1447 #endif
1448 	SCH_UNLOCK(sch);
1449 
1450 #ifdef TCP_RFC7413
1451 	if (tfo_cookie_valid) {
1452 		syncache_tfo_expand(sc, lsop, m, tfo_response_cookie);
1453 		/* INP_WUNLOCK(inp) will be performed by the called */
1454 		rv = 1;
1455 		goto tfo_done;
1456 	}
1457 #endif
1458 
1459 	/*
1460 	 * Do a standard 3-way handshake.
1461 	 */
1462 	if (syncache_respond(sc, sch, 0) == 0) {
1463 		if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1464 			syncache_free(sc);
1465 		else if (sc != &scs)
1466 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1467 		TCPSTAT_INC(tcps_sndacks);
1468 		TCPSTAT_INC(tcps_sndtotal);
1469 	} else {
1470 		if (sc != &scs)
1471 			syncache_free(sc);
1472 		TCPSTAT_INC(tcps_sc_dropped);
1473 	}
1474 
1475 done:
1476 	if (m) {
1477 		*lsop = NULL;
1478 		m_freem(m);
1479 	}
1480 #ifdef TCP_RFC7413
1481 tfo_done:
1482 #endif
1483 	if (cred != NULL)
1484 		crfree(cred);
1485 #ifdef MAC
1486 	if (sc == &scs)
1487 		mac_syncache_destroy(&maclabel);
1488 #endif
1489 	return (rv);
1490 }
1491 
1492 static int
1493 syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked)
1494 {
1495 	struct ip *ip = NULL;
1496 	struct mbuf *m;
1497 	struct tcphdr *th = NULL;
1498 	int optlen, error = 0;	/* Make compiler happy */
1499 	u_int16_t hlen, tlen, mssopt;
1500 	struct tcpopt to;
1501 #ifdef INET6
1502 	struct ip6_hdr *ip6 = NULL;
1503 #endif
1504 #ifdef TCP_SIGNATURE
1505 	struct secasvar *sav;
1506 #endif
1507 
1508 	hlen =
1509 #ifdef INET6
1510 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1511 #endif
1512 		sizeof(struct ip);
1513 	tlen = hlen + sizeof(struct tcphdr);
1514 
1515 	/* Determine MSS we advertize to other end of connection. */
1516 	mssopt = tcp_mssopt(&sc->sc_inc);
1517 	if (sc->sc_peer_mss)
1518 		mssopt = max( min(sc->sc_peer_mss, mssopt), V_tcp_minmss);
1519 
1520 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1521 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1522 	    ("syncache: mbuf too small"));
1523 
1524 	/* Create the IP+TCP header from scratch. */
1525 	m = m_gethdr(M_NOWAIT, MT_DATA);
1526 	if (m == NULL)
1527 		return (ENOBUFS);
1528 #ifdef MAC
1529 	mac_syncache_create_mbuf(sc->sc_label, m);
1530 #endif
1531 	m->m_data += max_linkhdr;
1532 	m->m_len = tlen;
1533 	m->m_pkthdr.len = tlen;
1534 	m->m_pkthdr.rcvif = NULL;
1535 
1536 #ifdef INET6
1537 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1538 		ip6 = mtod(m, struct ip6_hdr *);
1539 		ip6->ip6_vfc = IPV6_VERSION;
1540 		ip6->ip6_nxt = IPPROTO_TCP;
1541 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1542 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1543 		ip6->ip6_plen = htons(tlen - hlen);
1544 		/* ip6_hlim is set after checksum */
1545 		ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1546 		ip6->ip6_flow |= sc->sc_flowlabel;
1547 
1548 		th = (struct tcphdr *)(ip6 + 1);
1549 	}
1550 #endif
1551 #if defined(INET6) && defined(INET)
1552 	else
1553 #endif
1554 #ifdef INET
1555 	{
1556 		ip = mtod(m, struct ip *);
1557 		ip->ip_v = IPVERSION;
1558 		ip->ip_hl = sizeof(struct ip) >> 2;
1559 		ip->ip_len = htons(tlen);
1560 		ip->ip_id = 0;
1561 		ip->ip_off = 0;
1562 		ip->ip_sum = 0;
1563 		ip->ip_p = IPPROTO_TCP;
1564 		ip->ip_src = sc->sc_inc.inc_laddr;
1565 		ip->ip_dst = sc->sc_inc.inc_faddr;
1566 		ip->ip_ttl = sc->sc_ip_ttl;
1567 		ip->ip_tos = sc->sc_ip_tos;
1568 
1569 		/*
1570 		 * See if we should do MTU discovery.  Route lookups are
1571 		 * expensive, so we will only unset the DF bit if:
1572 		 *
1573 		 *	1) path_mtu_discovery is disabled
1574 		 *	2) the SCF_UNREACH flag has been set
1575 		 */
1576 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1577 		       ip->ip_off |= htons(IP_DF);
1578 
1579 		th = (struct tcphdr *)(ip + 1);
1580 	}
1581 #endif /* INET */
1582 	th->th_sport = sc->sc_inc.inc_lport;
1583 	th->th_dport = sc->sc_inc.inc_fport;
1584 
1585 	th->th_seq = htonl(sc->sc_iss);
1586 	th->th_ack = htonl(sc->sc_irs + 1);
1587 	th->th_off = sizeof(struct tcphdr) >> 2;
1588 	th->th_x2 = 0;
1589 	th->th_flags = TH_SYN|TH_ACK;
1590 	th->th_win = htons(sc->sc_wnd);
1591 	th->th_urp = 0;
1592 
1593 	if (sc->sc_flags & SCF_ECN) {
1594 		th->th_flags |= TH_ECE;
1595 		TCPSTAT_INC(tcps_ecn_shs);
1596 	}
1597 
1598 	/* Tack on the TCP options. */
1599 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1600 		to.to_flags = 0;
1601 
1602 		to.to_mss = mssopt;
1603 		to.to_flags = TOF_MSS;
1604 		if (sc->sc_flags & SCF_WINSCALE) {
1605 			to.to_wscale = sc->sc_requested_r_scale;
1606 			to.to_flags |= TOF_SCALE;
1607 		}
1608 		if (sc->sc_flags & SCF_TIMESTAMP) {
1609 			/* Virgin timestamp or TCP cookie enhanced one. */
1610 			to.to_tsval = sc->sc_ts;
1611 			to.to_tsecr = sc->sc_tsreflect;
1612 			to.to_flags |= TOF_TS;
1613 		}
1614 		if (sc->sc_flags & SCF_SACK)
1615 			to.to_flags |= TOF_SACKPERM;
1616 #ifdef TCP_SIGNATURE
1617 		sav = NULL;
1618 		if (sc->sc_flags & SCF_SIGNATURE) {
1619 			sav = tcp_get_sav(m, IPSEC_DIR_OUTBOUND);
1620 			if (sav != NULL)
1621 				to.to_flags |= TOF_SIGNATURE;
1622 			else {
1623 
1624 				/*
1625 				 * We've got SCF_SIGNATURE flag
1626 				 * inherited from listening socket,
1627 				 * but no SADB key for given source
1628 				 * address. Assume signature is not
1629 				 * required and remove signature flag
1630 				 * instead of silently dropping
1631 				 * connection.
1632 				 */
1633 				if (locked == 0)
1634 					SCH_LOCK(sch);
1635 				sc->sc_flags &= ~SCF_SIGNATURE;
1636 				if (locked == 0)
1637 					SCH_UNLOCK(sch);
1638 			}
1639 		}
1640 #endif
1641 
1642 #ifdef TCP_RFC7413
1643 		if (sc->sc_tfo_cookie) {
1644 			to.to_flags |= TOF_FASTOPEN;
1645 			to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1646 			to.to_tfo_cookie = sc->sc_tfo_cookie;
1647 			/* don't send cookie again when retransmitting response */
1648 			sc->sc_tfo_cookie = NULL;
1649 		}
1650 #endif
1651 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1652 
1653 		/* Adjust headers by option size. */
1654 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1655 		m->m_len += optlen;
1656 		m->m_pkthdr.len += optlen;
1657 
1658 #ifdef TCP_SIGNATURE
1659 		if (sc->sc_flags & SCF_SIGNATURE)
1660 			tcp_signature_do_compute(m, 0, optlen,
1661 			    to.to_signature, sav);
1662 #endif
1663 #ifdef INET6
1664 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1665 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1666 		else
1667 #endif
1668 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1669 	} else
1670 		optlen = 0;
1671 
1672 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
1673 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1674 #ifdef INET6
1675 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1676 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1677 		th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1678 		    IPPROTO_TCP, 0);
1679 		ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1680 #ifdef TCP_OFFLOAD
1681 		if (ADDED_BY_TOE(sc)) {
1682 			struct toedev *tod = sc->sc_tod;
1683 
1684 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1685 
1686 			return (error);
1687 		}
1688 #endif
1689 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1690 	}
1691 #endif
1692 #if defined(INET6) && defined(INET)
1693 	else
1694 #endif
1695 #ifdef INET
1696 	{
1697 		m->m_pkthdr.csum_flags = CSUM_TCP;
1698 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1699 		    htons(tlen + optlen - hlen + IPPROTO_TCP));
1700 #ifdef TCP_OFFLOAD
1701 		if (ADDED_BY_TOE(sc)) {
1702 			struct toedev *tod = sc->sc_tod;
1703 
1704 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1705 
1706 			return (error);
1707 		}
1708 #endif
1709 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1710 	}
1711 #endif
1712 	return (error);
1713 }
1714 
1715 /*
1716  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1717  * that exceed the capacity of the syncache by avoiding the storage of any
1718  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
1719  * attacks where the attacker does not have access to our responses.
1720  *
1721  * Syncookies encode and include all necessary information about the
1722  * connection setup within the SYN|ACK that we send back.  That way we
1723  * can avoid keeping any local state until the ACK to our SYN|ACK returns
1724  * (if ever).  Normally the syncache and syncookies are running in parallel
1725  * with the latter taking over when the former is exhausted.  When matching
1726  * syncache entry is found the syncookie is ignored.
1727  *
1728  * The only reliable information persisting the 3WHS is our inital sequence
1729  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
1730  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1731  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
1732  * returns and signifies a legitimate connection if it matches the ACK.
1733  *
1734  * The available space of 32 bits to store the hash and to encode the SYN
1735  * option information is very tight and we should have at least 24 bits for
1736  * the MAC to keep the number of guesses by blind spoofing reasonably high.
1737  *
1738  * SYN option information we have to encode to fully restore a connection:
1739  * MSS: is imporant to chose an optimal segment size to avoid IP level
1740  *   fragmentation along the path.  The common MSS values can be encoded
1741  *   in a 3-bit table.  Uncommon values are captured by the next lower value
1742  *   in the table leading to a slight increase in packetization overhead.
1743  * WSCALE: is necessary to allow large windows to be used for high delay-
1744  *   bandwidth product links.  Not scaling the window when it was initially
1745  *   negotiated is bad for performance as lack of scaling further decreases
1746  *   the apparent available send window.  We only need to encode the WSCALE
1747  *   we received from the remote end.  Our end can be recalculated at any
1748  *   time.  The common WSCALE values can be encoded in a 3-bit table.
1749  *   Uncommon values are captured by the next lower value in the table
1750  *   making us under-estimate the available window size halving our
1751  *   theoretically possible maximum throughput for that connection.
1752  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
1753  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
1754  *   that are included in all segments on a connection.  We enable them when
1755  *   the ACK has them.
1756  *
1757  * Security of syncookies and attack vectors:
1758  *
1759  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
1760  * together with the gloabl secret to make it unique per connection attempt.
1761  * Thus any change of any of those parameters results in a different MAC output
1762  * in an unpredictable way unless a collision is encountered.  24 bits of the
1763  * MAC are embedded into the ISS.
1764  *
1765  * To prevent replay attacks two rotating global secrets are updated with a
1766  * new random value every 15 seconds.  The life-time of a syncookie is thus
1767  * 15-30 seconds.
1768  *
1769  * Vector 1: Attacking the secret.  This requires finding a weakness in the
1770  * MAC itself or the way it is used here.  The attacker can do a chosen plain
1771  * text attack by varying and testing the all parameters under his control.
1772  * The strength depends on the size and randomness of the secret, and the
1773  * cryptographic security of the MAC function.  Due to the constant updating
1774  * of the secret the attacker has at most 29.999 seconds to find the secret
1775  * and launch spoofed connections.  After that he has to start all over again.
1776  *
1777  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
1778  * size an average of 4,823 attempts are required for a 50% chance of success
1779  * to spoof a single syncookie (birthday collision paradox).  However the
1780  * attacker is blind and doesn't know if one of his attempts succeeded unless
1781  * he has a side channel to interfere success from.  A single connection setup
1782  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
1783  * This many attempts are required for each one blind spoofed connection.  For
1784  * every additional spoofed connection he has to launch another N attempts.
1785  * Thus for a sustained rate 100 spoofed connections per second approximately
1786  * 1,800,000 packets per second would have to be sent.
1787  *
1788  * NB: The MAC function should be fast so that it doesn't become a CPU
1789  * exhaustion attack vector itself.
1790  *
1791  * References:
1792  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
1793  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
1794  *   http://cr.yp.to/syncookies.html    (overview)
1795  *   http://cr.yp.to/syncookies/archive (details)
1796  *
1797  *
1798  * Schematic construction of a syncookie enabled Initial Sequence Number:
1799  *  0        1         2         3
1800  *  12345678901234567890123456789012
1801  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
1802  *
1803  *  x 24 MAC (truncated)
1804  *  W  3 Send Window Scale index
1805  *  M  3 MSS index
1806  *  S  1 SACK permitted
1807  *  P  1 Odd/even secret
1808  */
1809 
1810 /*
1811  * Distribution and probability of certain MSS values.  Those in between are
1812  * rounded down to the next lower one.
1813  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
1814  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
1815  */
1816 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
1817 
1818 /*
1819  * Distribution and probability of certain WSCALE values.  We have to map the
1820  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
1821  * bits based on prevalence of certain values.  Where we don't have an exact
1822  * match for are rounded down to the next lower one letting us under-estimate
1823  * the true available window.  At the moment this would happen only for the
1824  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
1825  * and window size).  The absence of the WSCALE option (no scaling in either
1826  * direction) is encoded with index zero.
1827  * [WSCALE values histograms, Allman, 2012]
1828  *                            X 10 10 35  5  6 14 10%   by host
1829  *                            X 11  4  5  5 18 49  3%   by connections
1830  */
1831 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
1832 
1833 /*
1834  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
1835  * and good cryptographic properties.
1836  */
1837 static uint32_t
1838 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
1839     uint8_t *secbits, uintptr_t secmod)
1840 {
1841 	SIPHASH_CTX ctx;
1842 	uint32_t siphash[2];
1843 
1844 	SipHash24_Init(&ctx);
1845 	SipHash_SetKey(&ctx, secbits);
1846 	switch (inc->inc_flags & INC_ISIPV6) {
1847 #ifdef INET
1848 	case 0:
1849 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
1850 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
1851 		break;
1852 #endif
1853 #ifdef INET6
1854 	case INC_ISIPV6:
1855 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
1856 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
1857 		break;
1858 #endif
1859 	}
1860 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
1861 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
1862 	SipHash_Update(&ctx, &irs, sizeof(irs));
1863 	SipHash_Update(&ctx, &flags, sizeof(flags));
1864 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
1865 	SipHash_Final((u_int8_t *)&siphash, &ctx);
1866 
1867 	return (siphash[0] ^ siphash[1]);
1868 }
1869 
1870 static tcp_seq
1871 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
1872 {
1873 	u_int i, mss, secbit, wscale;
1874 	uint32_t iss, hash;
1875 	uint8_t *secbits;
1876 	union syncookie cookie;
1877 
1878 	SCH_LOCK_ASSERT(sch);
1879 
1880 	cookie.cookie = 0;
1881 
1882 	/* Map our computed MSS into the 3-bit index. */
1883 	mss = min(tcp_mssopt(&sc->sc_inc), max(sc->sc_peer_mss, V_tcp_minmss));
1884 	for (i = sizeof(tcp_sc_msstab) / sizeof(*tcp_sc_msstab) - 1;
1885 	     tcp_sc_msstab[i] > mss && i > 0;
1886 	     i--)
1887 		;
1888 	cookie.flags.mss_idx = i;
1889 
1890 	/*
1891 	 * Map the send window scale into the 3-bit index but only if
1892 	 * the wscale option was received.
1893 	 */
1894 	if (sc->sc_flags & SCF_WINSCALE) {
1895 		wscale = sc->sc_requested_s_scale;
1896 		for (i = sizeof(tcp_sc_wstab) / sizeof(*tcp_sc_wstab) - 1;
1897 		     tcp_sc_wstab[i] > wscale && i > 0;
1898 		     i--)
1899 			;
1900 		cookie.flags.wscale_idx = i;
1901 	}
1902 
1903 	/* Can we do SACK? */
1904 	if (sc->sc_flags & SCF_SACK)
1905 		cookie.flags.sack_ok = 1;
1906 
1907 	/* Which of the two secrets to use. */
1908 	secbit = sch->sch_sc->secret.oddeven & 0x1;
1909 	cookie.flags.odd_even = secbit;
1910 
1911 	secbits = sch->sch_sc->secret.key[secbit];
1912 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
1913 	    (uintptr_t)sch);
1914 
1915 	/*
1916 	 * Put the flags into the hash and XOR them to get better ISS number
1917 	 * variance.  This doesn't enhance the cryptographic strength and is
1918 	 * done to prevent the 8 cookie bits from showing up directly on the
1919 	 * wire.
1920 	 */
1921 	iss = hash & ~0xff;
1922 	iss |= cookie.cookie ^ (hash >> 24);
1923 
1924 	/* Randomize the timestamp. */
1925 	if (sc->sc_flags & SCF_TIMESTAMP) {
1926 		sc->sc_ts = arc4random();
1927 		sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks();
1928 	}
1929 
1930 	TCPSTAT_INC(tcps_sc_sendcookie);
1931 	return (iss);
1932 }
1933 
1934 static struct syncache *
1935 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
1936     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
1937     struct socket *lso)
1938 {
1939 	uint32_t hash;
1940 	uint8_t *secbits;
1941 	tcp_seq ack, seq;
1942 	int wnd, wscale = 0;
1943 	union syncookie cookie;
1944 
1945 	SCH_LOCK_ASSERT(sch);
1946 
1947 	/*
1948 	 * Pull information out of SYN-ACK/ACK and revert sequence number
1949 	 * advances.
1950 	 */
1951 	ack = th->th_ack - 1;
1952 	seq = th->th_seq - 1;
1953 
1954 	/*
1955 	 * Unpack the flags containing enough information to restore the
1956 	 * connection.
1957 	 */
1958 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
1959 
1960 	/* Which of the two secrets to use. */
1961 	secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
1962 
1963 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
1964 
1965 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
1966 	if ((ack & ~0xff) != (hash & ~0xff))
1967 		return (NULL);
1968 
1969 	/* Fill in the syncache values. */
1970 	sc->sc_flags = 0;
1971 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1972 	sc->sc_ipopts = NULL;
1973 
1974 	sc->sc_irs = seq;
1975 	sc->sc_iss = ack;
1976 
1977 	switch (inc->inc_flags & INC_ISIPV6) {
1978 #ifdef INET
1979 	case 0:
1980 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
1981 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
1982 		break;
1983 #endif
1984 #ifdef INET6
1985 	case INC_ISIPV6:
1986 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
1987 			sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
1988 		break;
1989 #endif
1990 	}
1991 
1992 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
1993 
1994 	/* We can simply recompute receive window scale we sent earlier. */
1995 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
1996 		wscale++;
1997 
1998 	/* Only use wscale if it was enabled in the orignal SYN. */
1999 	if (cookie.flags.wscale_idx > 0) {
2000 		sc->sc_requested_r_scale = wscale;
2001 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2002 		sc->sc_flags |= SCF_WINSCALE;
2003 	}
2004 
2005 	wnd = sbspace(&lso->so_rcv);
2006 	wnd = imax(wnd, 0);
2007 	wnd = imin(wnd, TCP_MAXWIN);
2008 	sc->sc_wnd = wnd;
2009 
2010 	if (cookie.flags.sack_ok)
2011 		sc->sc_flags |= SCF_SACK;
2012 
2013 	if (to->to_flags & TOF_TS) {
2014 		sc->sc_flags |= SCF_TIMESTAMP;
2015 		sc->sc_tsreflect = to->to_tsval;
2016 		sc->sc_ts = to->to_tsecr;
2017 		sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks();
2018 	}
2019 
2020 	if (to->to_flags & TOF_SIGNATURE)
2021 		sc->sc_flags |= SCF_SIGNATURE;
2022 
2023 	sc->sc_rxmits = 0;
2024 
2025 	TCPSTAT_INC(tcps_sc_recvcookie);
2026 	return (sc);
2027 }
2028 
2029 #ifdef INVARIANTS
2030 static int
2031 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2032     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2033     struct socket *lso)
2034 {
2035 	struct syncache scs, *scx;
2036 	char *s;
2037 
2038 	bzero(&scs, sizeof(scs));
2039 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
2040 
2041 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2042 		return (0);
2043 
2044 	if (scx != NULL) {
2045 		if (sc->sc_peer_mss != scx->sc_peer_mss)
2046 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2047 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2048 
2049 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2050 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2051 			    s, __func__, sc->sc_requested_r_scale,
2052 			    scx->sc_requested_r_scale);
2053 
2054 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2055 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2056 			    s, __func__, sc->sc_requested_s_scale,
2057 			    scx->sc_requested_s_scale);
2058 
2059 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2060 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2061 	}
2062 
2063 	if (s != NULL)
2064 		free(s, M_TCPLOG);
2065 	return (0);
2066 }
2067 #endif /* INVARIANTS */
2068 
2069 static void
2070 syncookie_reseed(void *arg)
2071 {
2072 	struct tcp_syncache *sc = arg;
2073 	uint8_t *secbits;
2074 	int secbit;
2075 
2076 	/*
2077 	 * Reseeding the secret doesn't have to be protected by a lock.
2078 	 * It only must be ensured that the new random values are visible
2079 	 * to all CPUs in a SMP environment.  The atomic with release
2080 	 * semantics ensures that.
2081 	 */
2082 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2083 	secbits = sc->secret.key[secbit];
2084 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2085 	atomic_add_rel_int(&sc->secret.oddeven, 1);
2086 
2087 	/* Reschedule ourself. */
2088 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2089 }
2090 
2091 /*
2092  * Returns the current number of syncache entries.  This number
2093  * will probably change before you get around to calling
2094  * syncache_pcblist.
2095  */
2096 int
2097 syncache_pcbcount(void)
2098 {
2099 	struct syncache_head *sch;
2100 	int count, i;
2101 
2102 	for (count = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
2103 		/* No need to lock for a read. */
2104 		sch = &V_tcp_syncache.hashbase[i];
2105 		count += sch->sch_length;
2106 	}
2107 	return count;
2108 }
2109 
2110 /*
2111  * Exports the syncache entries to userland so that netstat can display
2112  * them alongside the other sockets.  This function is intended to be
2113  * called only from tcp_pcblist.
2114  *
2115  * Due to concurrency on an active system, the number of pcbs exported
2116  * may have no relation to max_pcbs.  max_pcbs merely indicates the
2117  * amount of space the caller allocated for this function to use.
2118  */
2119 int
2120 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
2121 {
2122 	struct xtcpcb xt;
2123 	struct syncache *sc;
2124 	struct syncache_head *sch;
2125 	int count, error, i;
2126 
2127 	for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
2128 		sch = &V_tcp_syncache.hashbase[i];
2129 		SCH_LOCK(sch);
2130 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2131 			if (count >= max_pcbs) {
2132 				SCH_UNLOCK(sch);
2133 				goto exit;
2134 			}
2135 			if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2136 				continue;
2137 			bzero(&xt, sizeof(xt));
2138 			xt.xt_len = sizeof(xt);
2139 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2140 				xt.xt_inp.inp_vflag = INP_IPV6;
2141 			else
2142 				xt.xt_inp.inp_vflag = INP_IPV4;
2143 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo));
2144 			xt.xt_tp.t_inpcb = &xt.xt_inp;
2145 			xt.xt_tp.t_state = TCPS_SYN_RECEIVED;
2146 			xt.xt_socket.xso_protocol = IPPROTO_TCP;
2147 			xt.xt_socket.xso_len = sizeof (struct xsocket);
2148 			xt.xt_socket.so_type = SOCK_STREAM;
2149 			xt.xt_socket.so_state = SS_ISCONNECTING;
2150 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2151 			if (error) {
2152 				SCH_UNLOCK(sch);
2153 				goto exit;
2154 			}
2155 			count++;
2156 		}
2157 		SCH_UNLOCK(sch);
2158 	}
2159 exit:
2160 	*pcbs_exported = count;
2161 	return error;
2162 }
2163