xref: /freebsd/sys/netinet/tcp_syncache.c (revision 0b4539ee542f7011160c07bf52af7f16739f4a92)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2001 McAfee, Inc.
5  * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG
6  * All rights reserved.
7  *
8  * This software was developed for the FreeBSD Project by Jonathan Lemon
9  * and McAfee Research, the Security Research Division of McAfee, Inc. under
10  * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
11  * DARPA CHATS research program. [2001 McAfee, Inc.]
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  */
34 
35 #include <sys/cdefs.h>
36 #include "opt_inet.h"
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/hash.h>
43 #include <sys/refcount.h>
44 #include <sys/kernel.h>
45 #include <sys/sysctl.h>
46 #include <sys/limits.h>
47 #include <sys/lock.h>
48 #include <sys/mutex.h>
49 #include <sys/malloc.h>
50 #include <sys/mbuf.h>
51 #include <sys/proc.h>		/* for proc0 declaration */
52 #include <sys/random.h>
53 #include <sys/socket.h>
54 #include <sys/socketvar.h>
55 #include <sys/syslog.h>
56 #include <sys/ucred.h>
57 
58 #include <sys/md5.h>
59 #include <crypto/siphash/siphash.h>
60 
61 #include <vm/uma.h>
62 
63 #include <net/if.h>
64 #include <net/if_var.h>
65 #include <net/route.h>
66 #include <net/vnet.h>
67 
68 #include <netinet/in.h>
69 #include <netinet/in_kdtrace.h>
70 #include <netinet/in_systm.h>
71 #include <netinet/ip.h>
72 #include <netinet/in_var.h>
73 #include <netinet/in_pcb.h>
74 #include <netinet/ip_var.h>
75 #include <netinet/ip_options.h>
76 #ifdef INET6
77 #include <netinet/ip6.h>
78 #include <netinet/icmp6.h>
79 #include <netinet6/nd6.h>
80 #include <netinet6/ip6_var.h>
81 #include <netinet6/in6_pcb.h>
82 #endif
83 #include <netinet/tcp.h>
84 #include <netinet/tcp_fastopen.h>
85 #include <netinet/tcp_fsm.h>
86 #include <netinet/tcp_seq.h>
87 #include <netinet/tcp_timer.h>
88 #include <netinet/tcp_var.h>
89 #include <netinet/tcp_syncache.h>
90 #include <netinet/tcp_ecn.h>
91 #ifdef TCP_BLACKBOX
92 #include <netinet/tcp_log_buf.h>
93 #endif
94 #ifdef TCP_OFFLOAD
95 #include <netinet/toecore.h>
96 #endif
97 #include <netinet/udp.h>
98 
99 #include <netipsec/ipsec_support.h>
100 
101 #include <machine/in_cksum.h>
102 
103 #include <security/mac/mac_framework.h>
104 
105 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1;
106 #define	V_tcp_syncookies		VNET(tcp_syncookies)
107 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
108     &VNET_NAME(tcp_syncookies), 0,
109     "Use TCP SYN cookies if the syncache overflows");
110 
111 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0;
112 #define	V_tcp_syncookiesonly		VNET(tcp_syncookiesonly)
113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
114     &VNET_NAME(tcp_syncookiesonly), 0,
115     "Use only TCP SYN cookies");
116 
117 #ifdef TCP_OFFLOAD
118 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
119 #endif
120 
121 static void	 syncache_drop(struct syncache *, struct syncache_head *);
122 static void	 syncache_free(struct syncache *);
123 static void	 syncache_insert(struct syncache *, struct syncache_head *);
124 static int	 syncache_respond(struct syncache *, const struct mbuf *, int);
125 static struct	 socket *syncache_socket(struct syncache *, struct socket *,
126 		    struct mbuf *m);
127 static void	 syncache_timeout(struct syncache *sc, struct syncache_head *sch,
128 		    int docallout);
129 static void	 syncache_timer(void *);
130 
131 static uint32_t	 syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
132 		    uint8_t *, uintptr_t);
133 static tcp_seq	 syncookie_generate(struct syncache_head *, struct syncache *);
134 static struct syncache
135 		*syncookie_lookup(struct in_conninfo *, struct syncache_head *,
136 		    struct syncache *, struct tcphdr *, struct tcpopt *,
137 		    struct socket *, uint16_t);
138 static void	syncache_pause(struct in_conninfo *);
139 static void	syncache_unpause(void *);
140 static void	 syncookie_reseed(void *);
141 #ifdef INVARIANTS
142 static int	 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
143 		    struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
144 		    struct socket *lso, uint16_t port);
145 #endif
146 
147 /*
148  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
149  * 3 retransmits corresponds to a timeout with default values of
150  * tcp_rexmit_initial * (             1 +
151  *                       tcp_backoff[1] +
152  *                       tcp_backoff[2] +
153  *                       tcp_backoff[3]) + 3 * tcp_rexmit_slop,
154  * 1000 ms * (1 + 2 + 4 + 8) +  3 * 200 ms = 15600 ms,
155  * the odds are that the user has given up attempting to connect by then.
156  */
157 #define SYNCACHE_MAXREXMTS		3
158 
159 /* Arbitrary values */
160 #define TCP_SYNCACHE_HASHSIZE		512
161 #define TCP_SYNCACHE_BUCKETLIMIT	30
162 
163 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
164 #define	V_tcp_syncache			VNET(tcp_syncache)
165 
166 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache,
167     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
168     "TCP SYN cache");
169 
170 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
171     &VNET_NAME(tcp_syncache.bucket_limit), 0,
172     "Per-bucket hash limit for syncache");
173 
174 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
175     &VNET_NAME(tcp_syncache.cache_limit), 0,
176     "Overall entry limit for syncache");
177 
178 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
179     &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
180 
181 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
182     &VNET_NAME(tcp_syncache.hashsize), 0,
183     "Size of TCP syncache hashtable");
184 
185 SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET |
186     CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0,
187     "All syncache(4) entries are visible, ignoring UID/GID, jail(2) "
188     "and mac(4) checks");
189 
190 static int
sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)191 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
192 {
193 	int error;
194 	u_int new;
195 
196 	new = V_tcp_syncache.rexmt_limit;
197 	error = sysctl_handle_int(oidp, &new, 0, req);
198 	if ((error == 0) && (req->newptr != NULL)) {
199 		if (new > TCP_MAXRXTSHIFT)
200 			error = EINVAL;
201 		else
202 			V_tcp_syncache.rexmt_limit = new;
203 	}
204 	return (error);
205 }
206 
207 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
208     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
209     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
210     sysctl_net_inet_tcp_syncache_rexmtlimit_check, "IU",
211     "Limit on SYN/ACK retransmissions");
212 
213 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
214 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
215     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
216     "Send reset on socket allocation failure");
217 
218 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
219 
220 #define	SCH_LOCK(sch)		mtx_lock(&(sch)->sch_mtx)
221 #define	SCH_UNLOCK(sch)		mtx_unlock(&(sch)->sch_mtx)
222 #define	SCH_LOCK_ASSERT(sch)	mtx_assert(&(sch)->sch_mtx, MA_OWNED)
223 
224 /*
225  * Requires the syncache entry to be already removed from the bucket list.
226  */
227 static void
syncache_free(struct syncache * sc)228 syncache_free(struct syncache *sc)
229 {
230 
231 	if (sc->sc_ipopts)
232 		(void)m_free(sc->sc_ipopts);
233 	if (sc->sc_cred)
234 		crfree(sc->sc_cred);
235 #ifdef MAC
236 	mac_syncache_destroy(&sc->sc_label);
237 #endif
238 
239 	uma_zfree(V_tcp_syncache.zone, sc);
240 }
241 
242 void
syncache_init(void)243 syncache_init(void)
244 {
245 	int i;
246 
247 	V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
248 	V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
249 	V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
250 	V_tcp_syncache.hash_secret = arc4random();
251 
252 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
253 	    &V_tcp_syncache.hashsize);
254 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
255 	    &V_tcp_syncache.bucket_limit);
256 	if (!powerof2(V_tcp_syncache.hashsize) ||
257 	    V_tcp_syncache.hashsize == 0) {
258 		printf("WARNING: syncache hash size is not a power of 2.\n");
259 		V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
260 	}
261 	V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
262 
263 	/* Set limits. */
264 	V_tcp_syncache.cache_limit =
265 	    V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
266 	TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
267 	    &V_tcp_syncache.cache_limit);
268 
269 	/* Allocate the hash table. */
270 	V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
271 	    sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
272 
273 #ifdef VIMAGE
274 	V_tcp_syncache.vnet = curvnet;
275 #endif
276 
277 	/* Initialize the hash buckets. */
278 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
279 		TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
280 		mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
281 			 NULL, MTX_DEF);
282 		callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
283 			 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
284 		V_tcp_syncache.hashbase[i].sch_length = 0;
285 		V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
286 		V_tcp_syncache.hashbase[i].sch_last_overflow =
287 		    -(SYNCOOKIE_LIFETIME + 1);
288 	}
289 
290 	/* Create the syncache entry zone. */
291 	V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
292 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
293 	V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
294 	    V_tcp_syncache.cache_limit);
295 
296 	/* Start the SYN cookie reseeder callout. */
297 	callout_init(&V_tcp_syncache.secret.reseed, 1);
298 	arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
299 	arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
300 	callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
301 	    syncookie_reseed, &V_tcp_syncache);
302 
303 	/* Initialize the pause machinery. */
304 	mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF);
305 	callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx,
306 	    0);
307 	V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME;
308 	V_tcp_syncache.pause_backoff = 0;
309 	V_tcp_syncache.paused = false;
310 }
311 
312 #ifdef VIMAGE
313 void
syncache_destroy(void)314 syncache_destroy(void)
315 {
316 	struct syncache_head *sch;
317 	struct syncache *sc, *nsc;
318 	int i;
319 
320 	/*
321 	 * Stop the re-seed timer before freeing resources.  No need to
322 	 * possibly schedule it another time.
323 	 */
324 	callout_drain(&V_tcp_syncache.secret.reseed);
325 
326 	/* Stop the SYN cache pause callout. */
327 	mtx_lock(&V_tcp_syncache.pause_mtx);
328 	if (callout_stop(&V_tcp_syncache.pause_co) == 0) {
329 		mtx_unlock(&V_tcp_syncache.pause_mtx);
330 		callout_drain(&V_tcp_syncache.pause_co);
331 	} else
332 		mtx_unlock(&V_tcp_syncache.pause_mtx);
333 
334 	/* Cleanup hash buckets: stop timers, free entries, destroy locks. */
335 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
336 		sch = &V_tcp_syncache.hashbase[i];
337 		callout_drain(&sch->sch_timer);
338 
339 		SCH_LOCK(sch);
340 		TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
341 			syncache_drop(sc, sch);
342 		SCH_UNLOCK(sch);
343 		KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
344 		    ("%s: sch->sch_bucket not empty", __func__));
345 		KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
346 		    __func__, sch->sch_length));
347 		mtx_destroy(&sch->sch_mtx);
348 	}
349 
350 	KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
351 	    ("%s: cache_count not 0", __func__));
352 
353 	/* Free the allocated global resources. */
354 	uma_zdestroy(V_tcp_syncache.zone);
355 	free(V_tcp_syncache.hashbase, M_SYNCACHE);
356 	mtx_destroy(&V_tcp_syncache.pause_mtx);
357 }
358 #endif
359 
360 /*
361  * Inserts a syncache entry into the specified bucket row.
362  * Locks and unlocks the syncache_head autonomously.
363  */
364 static void
syncache_insert(struct syncache * sc,struct syncache_head * sch)365 syncache_insert(struct syncache *sc, struct syncache_head *sch)
366 {
367 	struct syncache *sc2;
368 
369 	SCH_LOCK(sch);
370 
371 	/*
372 	 * Make sure that we don't overflow the per-bucket limit.
373 	 * If the bucket is full, toss the oldest element.
374 	 */
375 	if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
376 		KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
377 			("sch->sch_length incorrect"));
378 		syncache_pause(&sc->sc_inc);
379 		sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
380 		sch->sch_last_overflow = time_uptime;
381 		syncache_drop(sc2, sch);
382 	}
383 
384 	/* Put it into the bucket. */
385 	TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
386 	sch->sch_length++;
387 
388 #ifdef TCP_OFFLOAD
389 	if (ADDED_BY_TOE(sc)) {
390 		struct toedev *tod = sc->sc_tod;
391 
392 		tod->tod_syncache_added(tod, sc->sc_todctx);
393 	}
394 #endif
395 
396 	/* Reinitialize the bucket row's timer. */
397 	if (sch->sch_length == 1)
398 		sch->sch_nextc = ticks + INT_MAX;
399 	syncache_timeout(sc, sch, 1);
400 
401 	SCH_UNLOCK(sch);
402 
403 	TCPSTATES_INC(TCPS_SYN_RECEIVED);
404 	TCPSTAT_INC(tcps_sc_added);
405 }
406 
407 /*
408  * Remove and free entry from syncache bucket row.
409  * Expects locked syncache head.
410  */
411 static void
syncache_drop(struct syncache * sc,struct syncache_head * sch)412 syncache_drop(struct syncache *sc, struct syncache_head *sch)
413 {
414 
415 	SCH_LOCK_ASSERT(sch);
416 
417 	TCPSTATES_DEC(TCPS_SYN_RECEIVED);
418 	TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
419 	sch->sch_length--;
420 
421 #ifdef TCP_OFFLOAD
422 	if (ADDED_BY_TOE(sc)) {
423 		struct toedev *tod = sc->sc_tod;
424 
425 		tod->tod_syncache_removed(tod, sc->sc_todctx);
426 	}
427 #endif
428 
429 	syncache_free(sc);
430 }
431 
432 /*
433  * Engage/reengage time on bucket row.
434  */
435 static void
syncache_timeout(struct syncache * sc,struct syncache_head * sch,int docallout)436 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
437 {
438 	int rexmt;
439 
440 	if (sc->sc_rxmits == 0)
441 		rexmt = tcp_rexmit_initial;
442 	else
443 		TCPT_RANGESET(rexmt,
444 		    tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
445 		    tcp_rexmit_min, TCPTV_REXMTMAX);
446 	sc->sc_rxttime = ticks + rexmt;
447 	sc->sc_rxmits++;
448 	if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
449 		sch->sch_nextc = sc->sc_rxttime;
450 		if (docallout)
451 			callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
452 			    syncache_timer, (void *)sch);
453 	}
454 }
455 
456 /*
457  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
458  * If we have retransmitted an entry the maximum number of times, expire it.
459  * One separate timer for each bucket row.
460  */
461 static void
syncache_timer(void * xsch)462 syncache_timer(void *xsch)
463 {
464 	struct syncache_head *sch = (struct syncache_head *)xsch;
465 	struct syncache *sc, *nsc;
466 	struct epoch_tracker et;
467 	int tick = ticks;
468 	char *s;
469 	bool paused;
470 
471 	CURVNET_SET(sch->sch_sc->vnet);
472 
473 	/* NB: syncache_head has already been locked by the callout. */
474 	SCH_LOCK_ASSERT(sch);
475 
476 	/*
477 	 * In the following cycle we may remove some entries and/or
478 	 * advance some timeouts, so re-initialize the bucket timer.
479 	 */
480 	sch->sch_nextc = tick + INT_MAX;
481 
482 	/*
483 	 * If we have paused processing, unconditionally remove
484 	 * all syncache entries.
485 	 */
486 	mtx_lock(&V_tcp_syncache.pause_mtx);
487 	paused = V_tcp_syncache.paused;
488 	mtx_unlock(&V_tcp_syncache.pause_mtx);
489 
490 	TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
491 		if (paused) {
492 			syncache_drop(sc, sch);
493 			continue;
494 		}
495 		/*
496 		 * We do not check if the listen socket still exists
497 		 * and accept the case where the listen socket may be
498 		 * gone by the time we resend the SYN/ACK.  We do
499 		 * not expect this to happens often. If it does,
500 		 * then the RST will be sent by the time the remote
501 		 * host does the SYN/ACK->ACK.
502 		 */
503 		if (TSTMP_GT(sc->sc_rxttime, tick)) {
504 			if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
505 				sch->sch_nextc = sc->sc_rxttime;
506 			continue;
507 		}
508 		if (sc->sc_rxmits > V_tcp_ecn_maxretries) {
509 			sc->sc_flags &= ~SCF_ECN_MASK;
510 		}
511 		if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
512 			if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
513 				log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
514 				    "giving up and removing syncache entry\n",
515 				    s, __func__);
516 				free(s, M_TCPLOG);
517 			}
518 			syncache_drop(sc, sch);
519 			TCPSTAT_INC(tcps_sc_stale);
520 			continue;
521 		}
522 		if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
523 			log(LOG_DEBUG, "%s; %s: Response timeout, "
524 			    "retransmitting (%u) SYN|ACK\n",
525 			    s, __func__, sc->sc_rxmits);
526 			free(s, M_TCPLOG);
527 		}
528 
529 		NET_EPOCH_ENTER(et);
530 		if (syncache_respond(sc, NULL, TH_SYN|TH_ACK) == 0) {
531 			syncache_timeout(sc, sch, 0);
532 			TCPSTAT_INC(tcps_sndacks);
533 			TCPSTAT_INC(tcps_sndtotal);
534 			TCPSTAT_INC(tcps_sc_retransmitted);
535 		} else {
536 			syncache_drop(sc, sch);
537 			TCPSTAT_INC(tcps_sc_dropped);
538 		}
539 		NET_EPOCH_EXIT(et);
540 	}
541 	if (!TAILQ_EMPTY(&(sch)->sch_bucket))
542 		callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
543 			syncache_timer, (void *)(sch));
544 	CURVNET_RESTORE();
545 }
546 
547 /*
548  * Returns true if the system is only using cookies at the moment.
549  * This could be due to a sysadmin decision to only use cookies, or it
550  * could be due to the system detecting an attack.
551  */
552 static inline bool
syncache_cookiesonly(void)553 syncache_cookiesonly(void)
554 {
555 
556 	return (V_tcp_syncookies && (V_tcp_syncache.paused ||
557 	    V_tcp_syncookiesonly));
558 }
559 
560 /*
561  * Find the hash bucket for the given connection.
562  */
563 static struct syncache_head *
syncache_hashbucket(struct in_conninfo * inc)564 syncache_hashbucket(struct in_conninfo *inc)
565 {
566 	uint32_t hash;
567 
568 	/*
569 	 * The hash is built on foreign port + local port + foreign address.
570 	 * We rely on the fact that struct in_conninfo starts with 16 bits
571 	 * of foreign port, then 16 bits of local port then followed by 128
572 	 * bits of foreign address.  In case of IPv4 address, the first 3
573 	 * 32-bit words of the address always are zeroes.
574 	 */
575 	hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
576 	    V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
577 
578 	return (&V_tcp_syncache.hashbase[hash]);
579 }
580 
581 /*
582  * Find an entry in the syncache.
583  * Returns always with locked syncache_head plus a matching entry or NULL.
584  */
585 static struct syncache *
syncache_lookup(struct in_conninfo * inc,struct syncache_head ** schp)586 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
587 {
588 	struct syncache *sc;
589 	struct syncache_head *sch;
590 
591 	*schp = sch = syncache_hashbucket(inc);
592 	SCH_LOCK(sch);
593 
594 	/* Circle through bucket row to find matching entry. */
595 	TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
596 		if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
597 		    sizeof(struct in_endpoints)) == 0)
598 			break;
599 
600 	return (sc);	/* Always returns with locked sch. */
601 }
602 
603 /*
604  * This function is called when we get a RST for a
605  * non-existent connection, so that we can see if the
606  * connection is in the syn cache.  If it is, zap it.
607  * If required send a challenge ACK.
608  */
609 void
syncache_chkrst(struct in_conninfo * inc,struct tcphdr * th,struct mbuf * m,uint16_t port)610 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m,
611     uint16_t port)
612 {
613 	struct syncache *sc;
614 	struct syncache_head *sch;
615 	char *s = NULL;
616 
617 	if (syncache_cookiesonly())
618 		return;
619 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
620 	SCH_LOCK_ASSERT(sch);
621 
622 	/*
623 	 * No corresponding connection was found in syncache.
624 	 * If syncookies are enabled and possibly exclusively
625 	 * used, or we are under memory pressure, a valid RST
626 	 * may not find a syncache entry.  In that case we're
627 	 * done and no SYN|ACK retransmissions will happen.
628 	 * Otherwise the RST was misdirected or spoofed.
629 	 */
630 	if (sc == NULL) {
631 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
632 			log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
633 			    "syncache entry (possibly syncookie only), "
634 			    "segment ignored\n", s, __func__);
635 		TCPSTAT_INC(tcps_badrst);
636 		goto done;
637 	}
638 
639 	/* The remote UDP encaps port does not match. */
640 	if (sc->sc_port != port) {
641 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
642 			log(LOG_DEBUG, "%s; %s: Spurious RST with matching "
643 			    "syncache entry but non-matching UDP encaps port, "
644 			    "segment ignored\n", s, __func__);
645 		TCPSTAT_INC(tcps_badrst);
646 		goto done;
647 	}
648 
649 	/*
650 	 * If the RST bit is set, check the sequence number to see
651 	 * if this is a valid reset segment.
652 	 *
653 	 * RFC 793 page 37:
654 	 *   In all states except SYN-SENT, all reset (RST) segments
655 	 *   are validated by checking their SEQ-fields.  A reset is
656 	 *   valid if its sequence number is in the window.
657 	 *
658 	 * RFC 793 page 69:
659 	 *   There are four cases for the acceptability test for an incoming
660 	 *   segment:
661 	 *
662 	 * Segment Receive  Test
663 	 * Length  Window
664 	 * ------- -------  -------------------------------------------
665 	 *    0       0     SEG.SEQ = RCV.NXT
666 	 *    0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
667 	 *   >0       0     not acceptable
668 	 *   >0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
669 	 *               or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
670 	 *
671 	 * Note that when receiving a SYN segment in the LISTEN state,
672 	 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
673 	 * described in RFC 793, page 66.
674 	 */
675 	if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
676 	    SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
677 	    (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
678 		if (V_tcp_insecure_rst ||
679 		    th->th_seq == sc->sc_irs + 1) {
680 			syncache_drop(sc, sch);
681 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
682 				log(LOG_DEBUG,
683 				    "%s; %s: Our SYN|ACK was rejected, "
684 				    "connection attempt aborted by remote "
685 				    "endpoint\n",
686 				    s, __func__);
687 			TCPSTAT_INC(tcps_sc_reset);
688 		} else {
689 			TCPSTAT_INC(tcps_badrst);
690 			/* Send challenge ACK. */
691 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
692 				log(LOG_DEBUG, "%s; %s: RST with invalid "
693 				    " SEQ %u != NXT %u (+WND %u), "
694 				    "sending challenge ACK\n",
695 				    s, __func__,
696 				    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
697 			if (syncache_respond(sc, m, TH_ACK) == 0) {
698 				TCPSTAT_INC(tcps_sndacks);
699 				TCPSTAT_INC(tcps_sndtotal);
700 			} else {
701 				syncache_drop(sc, sch);
702 				TCPSTAT_INC(tcps_sc_dropped);
703 			}
704 		}
705 	} else {
706 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
707 			log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
708 			    "NXT %u (+WND %u), segment ignored\n",
709 			    s, __func__,
710 			    th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
711 		TCPSTAT_INC(tcps_badrst);
712 	}
713 
714 done:
715 	if (s != NULL)
716 		free(s, M_TCPLOG);
717 	SCH_UNLOCK(sch);
718 }
719 
720 void
syncache_badack(struct in_conninfo * inc,uint16_t port)721 syncache_badack(struct in_conninfo *inc, uint16_t port)
722 {
723 	struct syncache *sc;
724 	struct syncache_head *sch;
725 
726 	if (syncache_cookiesonly())
727 		return;
728 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
729 	SCH_LOCK_ASSERT(sch);
730 	if ((sc != NULL) && (sc->sc_port == port)) {
731 		syncache_drop(sc, sch);
732 		TCPSTAT_INC(tcps_sc_badack);
733 	}
734 	SCH_UNLOCK(sch);
735 }
736 
737 void
syncache_unreach(struct in_conninfo * inc,tcp_seq th_seq,uint16_t port)738 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port)
739 {
740 	struct syncache *sc;
741 	struct syncache_head *sch;
742 
743 	if (syncache_cookiesonly())
744 		return;
745 	sc = syncache_lookup(inc, &sch);	/* returns locked sch */
746 	SCH_LOCK_ASSERT(sch);
747 	if (sc == NULL)
748 		goto done;
749 
750 	/* If the port != sc_port, then it's a bogus ICMP msg */
751 	if (port != sc->sc_port)
752 		goto done;
753 
754 	/* If the sequence number != sc_iss, then it's a bogus ICMP msg */
755 	if (ntohl(th_seq) != sc->sc_iss)
756 		goto done;
757 
758 	/*
759 	 * If we've rertransmitted 3 times and this is our second error,
760 	 * we remove the entry.  Otherwise, we allow it to continue on.
761 	 * This prevents us from incorrectly nuking an entry during a
762 	 * spurious network outage.
763 	 *
764 	 * See tcp_notify().
765 	 */
766 	if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
767 		sc->sc_flags |= SCF_UNREACH;
768 		goto done;
769 	}
770 	syncache_drop(sc, sch);
771 	TCPSTAT_INC(tcps_sc_unreach);
772 done:
773 	SCH_UNLOCK(sch);
774 }
775 
776 /*
777  * Build a new TCP socket structure from a syncache entry.
778  *
779  * On success return the newly created socket with its underlying inp locked.
780  */
781 static struct socket *
syncache_socket(struct syncache * sc,struct socket * lso,struct mbuf * m)782 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
783 {
784 	struct inpcb *inp = NULL;
785 	struct socket *so;
786 	struct tcpcb *tp;
787 	int error;
788 	char *s;
789 
790 	NET_EPOCH_ASSERT();
791 
792 	/*
793 	 * Ok, create the full blown connection, and set things up
794 	 * as they would have been set up if we had created the
795 	 * connection when the SYN arrived.
796 	 */
797 	if ((so = solisten_clone(lso)) == NULL)
798 		goto allocfail;
799 #ifdef MAC
800 	mac_socketpeer_set_from_mbuf(m, so);
801 #endif
802 	error = in_pcballoc(so, &V_tcbinfo);
803 	if (error) {
804 		sodealloc(so);
805 		goto allocfail;
806 	}
807 	inp = sotoinpcb(so);
808 	if ((tp = tcp_newtcpcb(inp, sototcpcb(lso))) == NULL) {
809 		in_pcbfree(inp);
810 		sodealloc(so);
811 		goto allocfail;
812 	}
813 	inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
814 #ifdef INET6
815 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
816 		inp->inp_vflag &= ~INP_IPV4;
817 		inp->inp_vflag |= INP_IPV6;
818 		inp->in6p_laddr = sc->sc_inc.inc6_laddr;
819 	} else {
820 		inp->inp_vflag &= ~INP_IPV6;
821 		inp->inp_vflag |= INP_IPV4;
822 #endif
823 		inp->inp_ip_ttl = sc->sc_ip_ttl;
824 		inp->inp_ip_tos = sc->sc_ip_tos;
825 		inp->inp_laddr = sc->sc_inc.inc_laddr;
826 #ifdef INET6
827 	}
828 #endif
829 
830 	/*
831 	 * If there's an mbuf and it has a flowid, then let's initialise the
832 	 * inp with that particular flowid.
833 	 */
834 	if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
835 		inp->inp_flowid = m->m_pkthdr.flowid;
836 		inp->inp_flowtype = M_HASHTYPE_GET(m);
837 #ifdef NUMA
838 		inp->inp_numa_domain = m->m_pkthdr.numa_domain;
839 #endif
840 	}
841 
842 	inp->inp_lport = sc->sc_inc.inc_lport;
843 #ifdef INET6
844 	if (inp->inp_vflag & INP_IPV6PROTO) {
845 		struct inpcb *oinp = sotoinpcb(lso);
846 
847 		/*
848 		 * Inherit socket options from the listening socket.
849 		 * Note that in6p_inputopts are not (and should not be)
850 		 * copied, since it stores previously received options and is
851 		 * used to detect if each new option is different than the
852 		 * previous one and hence should be passed to a user.
853 		 * If we copied in6p_inputopts, a user would not be able to
854 		 * receive options just after calling the accept system call.
855 		 */
856 		inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
857 		if (oinp->in6p_outputopts)
858 			inp->in6p_outputopts =
859 			    ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
860 		inp->in6p_hops = oinp->in6p_hops;
861 	}
862 
863 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
864 		struct sockaddr_in6 sin6;
865 
866 		sin6.sin6_family = AF_INET6;
867 		sin6.sin6_len = sizeof(sin6);
868 		sin6.sin6_addr = sc->sc_inc.inc6_faddr;
869 		sin6.sin6_port = sc->sc_inc.inc_fport;
870 		sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
871 		INP_HASH_WLOCK(&V_tcbinfo);
872 		error = in6_pcbconnect(inp, &sin6, thread0.td_ucred, false);
873 		INP_HASH_WUNLOCK(&V_tcbinfo);
874 		if (error != 0)
875 			goto abort;
876 		/* Override flowlabel from in6_pcbconnect. */
877 		inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
878 		inp->inp_flow |= sc->sc_flowlabel;
879 	}
880 #endif /* INET6 */
881 #if defined(INET) && defined(INET6)
882 	else
883 #endif
884 #ifdef INET
885 	{
886 		struct sockaddr_in sin;
887 
888 		inp->inp_options = (m) ? ip_srcroute(m) : NULL;
889 
890 		if (inp->inp_options == NULL) {
891 			inp->inp_options = sc->sc_ipopts;
892 			sc->sc_ipopts = NULL;
893 		}
894 
895 		sin.sin_family = AF_INET;
896 		sin.sin_len = sizeof(sin);
897 		sin.sin_addr = sc->sc_inc.inc_faddr;
898 		sin.sin_port = sc->sc_inc.inc_fport;
899 		bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
900 		INP_HASH_WLOCK(&V_tcbinfo);
901 		error = in_pcbconnect(inp, &sin, thread0.td_ucred);
902 		INP_HASH_WUNLOCK(&V_tcbinfo);
903 		if (error != 0)
904 			goto abort;
905 	}
906 #endif /* INET */
907 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
908 	/* Copy old policy into new socket's. */
909 	if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
910 		printf("syncache_socket: could not copy policy\n");
911 #endif
912 	tp->t_state = TCPS_SYN_RECEIVED;
913 	tp->iss = sc->sc_iss;
914 	tp->irs = sc->sc_irs;
915 	tp->t_port = sc->sc_port;
916 	tcp_rcvseqinit(tp);
917 	tcp_sendseqinit(tp);
918 	tp->snd_wl1 = sc->sc_irs;
919 	tp->snd_max = tp->iss + 1;
920 	tp->snd_nxt = tp->iss + 1;
921 	tp->rcv_up = sc->sc_irs + 1;
922 	tp->rcv_wnd = sc->sc_wnd;
923 	tp->rcv_adv += tp->rcv_wnd;
924 	tp->last_ack_sent = tp->rcv_nxt;
925 
926 	tp->t_flags = sototcpcb(lso)->t_flags &
927 	    (TF_LRD|TF_NOPUSH|TF_NODELAY);
928 	if (sc->sc_flags & SCF_NOOPT)
929 		tp->t_flags |= TF_NOOPT;
930 	else {
931 		if (sc->sc_flags & SCF_WINSCALE) {
932 			tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
933 			tp->snd_scale = sc->sc_requested_s_scale;
934 			tp->request_r_scale = sc->sc_requested_r_scale;
935 		}
936 		if (sc->sc_flags & SCF_TIMESTAMP) {
937 			tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
938 			tp->ts_recent = sc->sc_tsreflect;
939 			tp->ts_recent_age = tcp_ts_getticks();
940 			tp->ts_offset = sc->sc_tsoff;
941 		}
942 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
943 		if (sc->sc_flags & SCF_SIGNATURE)
944 			tp->t_flags |= TF_SIGNATURE;
945 #endif
946 		if (sc->sc_flags & SCF_SACK)
947 			tp->t_flags |= TF_SACK_PERMIT;
948 	}
949 
950 	tcp_ecn_syncache_socket(tp, sc);
951 
952 	/*
953 	 * Set up MSS and get cached values from tcp_hostcache.
954 	 * This might overwrite some of the defaults we just set.
955 	 */
956 	tcp_mss(tp, sc->sc_peer_mss);
957 
958 	/*
959 	 * If the SYN,ACK was retransmitted, indicate that CWND to be
960 	 * limited to one segment in cc_conn_init().
961 	 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
962 	 */
963 	if (sc->sc_rxmits > 1)
964 		tp->snd_cwnd = 1;
965 
966 #ifdef TCP_OFFLOAD
967 	/*
968 	 * Allow a TOE driver to install its hooks.  Note that we hold the
969 	 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
970 	 * new connection before the TOE driver has done its thing.
971 	 */
972 	if (ADDED_BY_TOE(sc)) {
973 		struct toedev *tod = sc->sc_tod;
974 
975 		tod->tod_offload_socket(tod, sc->sc_todctx, so);
976 	}
977 #endif
978 #ifdef TCP_BLACKBOX
979 	/*
980 	 * Inherit the log state from the listening socket, if
981 	 * - the log state of the listening socket is not off and
982 	 * - the listening socket was not auto selected from all sessions and
983 	 * - a log id is not set on the listening socket.
984 	 * This avoids inheriting a log state which was automatically set.
985 	 */
986 	if ((tcp_get_bblog_state(sototcpcb(lso)) != TCP_LOG_STATE_OFF) &&
987 	    ((sototcpcb(lso)->t_flags2 & TF2_LOG_AUTO) == 0) &&
988 	    (sototcpcb(lso)->t_lib == NULL)) {
989 		tcp_log_state_change(tp, tcp_get_bblog_state(sototcpcb(lso)));
990 	}
991 #endif
992 	/*
993 	 * Copy and activate timers.
994 	 */
995 	tp->t_maxunacktime = sototcpcb(lso)->t_maxunacktime;
996 	tp->t_keepinit = sototcpcb(lso)->t_keepinit;
997 	tp->t_keepidle = sototcpcb(lso)->t_keepidle;
998 	tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
999 	tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1000 	tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1001 
1002 	TCPSTAT_INC(tcps_accepts);
1003 	TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, TCPS_LISTEN);
1004 
1005 	if (!solisten_enqueue(so, SS_ISCONNECTED))
1006 		tp->t_flags |= TF_SONOTCONN;
1007 	/* Can we inherit anything from the listener? */
1008 	if (tp->t_fb->tfb_inherit != NULL) {
1009 		(*tp->t_fb->tfb_inherit)(tp, sotoinpcb(lso));
1010 	}
1011 	return (so);
1012 
1013 allocfail:
1014 	/*
1015 	 * Drop the connection; we will either send a RST or have the peer
1016 	 * retransmit its SYN again after its RTO and try again.
1017 	 */
1018 	if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1019 		log(LOG_DEBUG, "%s; %s: Socket create failed "
1020 		    "due to limits or memory shortage\n",
1021 		    s, __func__);
1022 		free(s, M_TCPLOG);
1023 	}
1024 	TCPSTAT_INC(tcps_listendrop);
1025 	return (NULL);
1026 
1027 abort:
1028 	tcp_discardcb(tp);
1029 	in_pcbfree(inp);
1030 	sodealloc(so);
1031 	if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1032 		log(LOG_DEBUG, "%s; %s: in%s_pcbconnect failed with error %i\n",
1033 		    s, __func__, (sc->sc_inc.inc_flags & INC_ISIPV6) ? "6" : "",
1034 		    error);
1035 		free(s, M_TCPLOG);
1036 	}
1037 	TCPSTAT_INC(tcps_listendrop);
1038 	return (NULL);
1039 }
1040 
1041 /*
1042  * This function gets called when we receive an ACK for a
1043  * socket in the LISTEN state.  We look up the connection
1044  * in the syncache, and if its there, we pull it out of
1045  * the cache and turn it into a full-blown connection in
1046  * the SYN-RECEIVED state.
1047  *
1048  * On syncache_socket() success the newly created socket
1049  * has its underlying inp locked.
1050  */
1051 int
syncache_expand(struct in_conninfo * inc,struct tcpopt * to,struct tcphdr * th,struct socket ** lsop,struct mbuf * m,uint16_t port)1052 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1053     struct socket **lsop, struct mbuf *m, uint16_t port)
1054 {
1055 	struct syncache *sc;
1056 	struct syncache_head *sch;
1057 	struct syncache scs;
1058 	char *s;
1059 	bool locked;
1060 
1061 	NET_EPOCH_ASSERT();
1062 	KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1063 	    ("%s: can handle only ACK", __func__));
1064 
1065 	if (syncache_cookiesonly()) {
1066 		sc = NULL;
1067 		sch = syncache_hashbucket(inc);
1068 		locked = false;
1069 	} else {
1070 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1071 		locked = true;
1072 		SCH_LOCK_ASSERT(sch);
1073 	}
1074 
1075 #ifdef INVARIANTS
1076 	/*
1077 	 * Test code for syncookies comparing the syncache stored
1078 	 * values with the reconstructed values from the cookie.
1079 	 */
1080 	if (sc != NULL)
1081 		syncookie_cmp(inc, sch, sc, th, to, *lsop, port);
1082 #endif
1083 
1084 	if (sc == NULL) {
1085 		/*
1086 		 * There is no syncache entry, so see if this ACK is
1087 		 * a returning syncookie.  To do this, first:
1088 		 *  A. Check if syncookies are used in case of syncache
1089 		 *     overflows
1090 		 *  B. See if this socket has had a syncache entry dropped in
1091 		 *     the recent past. We don't want to accept a bogus
1092 		 *     syncookie if we've never received a SYN or accept it
1093 		 *     twice.
1094 		 *  C. check that the syncookie is valid.  If it is, then
1095 		 *     cobble up a fake syncache entry, and return.
1096 		 */
1097 		if (locked && !V_tcp_syncookies) {
1098 			SCH_UNLOCK(sch);
1099 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1100 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1101 				    "segment rejected (syncookies disabled)\n",
1102 				    s, __func__);
1103 			goto failed;
1104 		}
1105 		if (locked && !V_tcp_syncookiesonly &&
1106 		    sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1107 			SCH_UNLOCK(sch);
1108 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1109 				log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1110 				    "segment rejected (no syncache entry)\n",
1111 				    s, __func__);
1112 			goto failed;
1113 		}
1114 		bzero(&scs, sizeof(scs));
1115 		sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port);
1116 		if (locked)
1117 			SCH_UNLOCK(sch);
1118 		if (sc == NULL) {
1119 			if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1120 				log(LOG_DEBUG, "%s; %s: Segment failed "
1121 				    "SYNCOOKIE authentication, segment rejected "
1122 				    "(probably spoofed)\n", s, __func__);
1123 			goto failed;
1124 		}
1125 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1126 		/* If received ACK has MD5 signature, check it. */
1127 		if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1128 		    (!TCPMD5_ENABLED() ||
1129 		    TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1130 			/* Drop the ACK. */
1131 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1132 				log(LOG_DEBUG, "%s; %s: Segment rejected, "
1133 				    "MD5 signature doesn't match.\n",
1134 				    s, __func__);
1135 				free(s, M_TCPLOG);
1136 			}
1137 			TCPSTAT_INC(tcps_sig_err_sigopt);
1138 			return (-1); /* Do not send RST */
1139 		}
1140 #endif /* TCP_SIGNATURE */
1141 		TCPSTATES_INC(TCPS_SYN_RECEIVED);
1142 	} else {
1143 		if (sc->sc_port != port) {
1144 			SCH_UNLOCK(sch);
1145 			return (0);
1146 		}
1147 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1148 		/*
1149 		 * If listening socket requested TCP digests, check that
1150 		 * received ACK has signature and it is correct.
1151 		 * If not, drop the ACK and leave sc entry in th cache,
1152 		 * because SYN was received with correct signature.
1153 		 */
1154 		if (sc->sc_flags & SCF_SIGNATURE) {
1155 			if ((to->to_flags & TOF_SIGNATURE) == 0) {
1156 				/* No signature */
1157 				TCPSTAT_INC(tcps_sig_err_nosigopt);
1158 				SCH_UNLOCK(sch);
1159 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1160 					log(LOG_DEBUG, "%s; %s: Segment "
1161 					    "rejected, MD5 signature wasn't "
1162 					    "provided.\n", s, __func__);
1163 					free(s, M_TCPLOG);
1164 				}
1165 				return (-1); /* Do not send RST */
1166 			}
1167 			if (!TCPMD5_ENABLED() ||
1168 			    TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1169 				/* Doesn't match or no SA */
1170 				SCH_UNLOCK(sch);
1171 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1172 					log(LOG_DEBUG, "%s; %s: Segment "
1173 					    "rejected, MD5 signature doesn't "
1174 					    "match.\n", s, __func__);
1175 					free(s, M_TCPLOG);
1176 				}
1177 				return (-1); /* Do not send RST */
1178 			}
1179 		}
1180 #endif /* TCP_SIGNATURE */
1181 
1182 		/*
1183 		 * RFC 7323 PAWS: If we have a timestamp on this segment and
1184 		 * it's less than ts_recent, drop it.
1185 		 * XXXMT: RFC 7323 also requires to send an ACK.
1186 		 *        In tcp_input.c this is only done for TCP segments
1187 		 *        with user data, so be consistent here and just drop
1188 		 *        the segment.
1189 		 */
1190 		if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1191 		    TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1192 			SCH_UNLOCK(sch);
1193 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1194 				log(LOG_DEBUG,
1195 				    "%s; %s: SEG.TSval %u < TS.Recent %u, "
1196 				    "segment dropped\n", s, __func__,
1197 				    to->to_tsval, sc->sc_tsreflect);
1198 				free(s, M_TCPLOG);
1199 			}
1200 			return (-1);  /* Do not send RST */
1201 		}
1202 
1203 		/*
1204 		 * If timestamps were not negotiated during SYN/ACK and a
1205 		 * segment with a timestamp is received, ignore the
1206 		 * timestamp and process the packet normally.
1207 		 * See section 3.2 of RFC 7323.
1208 		 */
1209 		if (!(sc->sc_flags & SCF_TIMESTAMP) &&
1210 		    (to->to_flags & TOF_TS)) {
1211 			if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1212 				log(LOG_DEBUG, "%s; %s: Timestamp not "
1213 				    "expected, segment processed normally\n",
1214 				    s, __func__);
1215 				free(s, M_TCPLOG);
1216 				s = NULL;
1217 			}
1218 		}
1219 
1220 		/*
1221 		 * If timestamps were negotiated during SYN/ACK and a
1222 		 * segment without a timestamp is received, silently drop
1223 		 * the segment, unless the missing timestamps are tolerated.
1224 		 * See section 3.2 of RFC 7323.
1225 		 */
1226 		if ((sc->sc_flags & SCF_TIMESTAMP) &&
1227 		    !(to->to_flags & TOF_TS)) {
1228 			if (V_tcp_tolerate_missing_ts) {
1229 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1230 					log(LOG_DEBUG,
1231 					    "%s; %s: Timestamp missing, "
1232 					    "segment processed normally\n",
1233 					    s, __func__);
1234 					free(s, M_TCPLOG);
1235 				}
1236 			} else {
1237 				SCH_UNLOCK(sch);
1238 				if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1239 					log(LOG_DEBUG,
1240 					    "%s; %s: Timestamp missing, "
1241 					    "segment silently dropped\n",
1242 					    s, __func__);
1243 					free(s, M_TCPLOG);
1244 				}
1245 				return (-1);  /* Do not send RST */
1246 			}
1247 		}
1248 		TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1249 		sch->sch_length--;
1250 #ifdef TCP_OFFLOAD
1251 		if (ADDED_BY_TOE(sc)) {
1252 			struct toedev *tod = sc->sc_tod;
1253 
1254 			tod->tod_syncache_removed(tod, sc->sc_todctx);
1255 		}
1256 #endif
1257 		SCH_UNLOCK(sch);
1258 	}
1259 
1260 	/*
1261 	 * Segment validation:
1262 	 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1263 	 */
1264 	if (th->th_ack != sc->sc_iss + 1) {
1265 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1266 			log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1267 			    "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1268 		goto failed;
1269 	}
1270 
1271 	/*
1272 	 * The SEQ must fall in the window starting at the received
1273 	 * initial receive sequence number + 1 (the SYN).
1274 	 */
1275 	if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1276 	    SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1277 		if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1278 			log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1279 			    "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1280 		goto failed;
1281 	}
1282 
1283 	*lsop = syncache_socket(sc, *lsop, m);
1284 
1285 	if (__predict_false(*lsop == NULL)) {
1286 		TCPSTAT_INC(tcps_sc_aborted);
1287 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1288 	} else
1289 		TCPSTAT_INC(tcps_sc_completed);
1290 
1291 /* how do we find the inp for the new socket? */
1292 	if (sc != &scs)
1293 		syncache_free(sc);
1294 	return (1);
1295 failed:
1296 	if (sc != NULL) {
1297 		TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1298 		if (sc != &scs)
1299 			syncache_free(sc);
1300 	}
1301 	if (s != NULL)
1302 		free(s, M_TCPLOG);
1303 	*lsop = NULL;
1304 	return (0);
1305 }
1306 
1307 static struct socket *
syncache_tfo_expand(struct syncache * sc,struct socket * lso,struct mbuf * m,uint64_t response_cookie)1308 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m,
1309     uint64_t response_cookie)
1310 {
1311 	struct inpcb *inp;
1312 	struct tcpcb *tp;
1313 	unsigned int *pending_counter;
1314 	struct socket *so;
1315 
1316 	NET_EPOCH_ASSERT();
1317 
1318 	pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending;
1319 	so = syncache_socket(sc, lso, m);
1320 	if (so == NULL) {
1321 		TCPSTAT_INC(tcps_sc_aborted);
1322 		atomic_subtract_int(pending_counter, 1);
1323 	} else {
1324 		soisconnected(so);
1325 		inp = sotoinpcb(so);
1326 		tp = intotcpcb(inp);
1327 		tp->t_flags |= TF_FASTOPEN;
1328 		tp->t_tfo_cookie.server = response_cookie;
1329 		tp->snd_max = tp->iss;
1330 		tp->snd_nxt = tp->iss;
1331 		tp->t_tfo_pending = pending_counter;
1332 		TCPSTATES_INC(TCPS_SYN_RECEIVED);
1333 		TCPSTAT_INC(tcps_sc_completed);
1334 	}
1335 
1336 	return (so);
1337 }
1338 
1339 /*
1340  * Given a LISTEN socket and an inbound SYN request, add
1341  * this to the syn cache, and send back a segment:
1342  *	<SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1343  * to the source.
1344  *
1345  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1346  * Doing so would require that we hold onto the data and deliver it
1347  * to the application.  However, if we are the target of a SYN-flood
1348  * DoS attack, an attacker could send data which would eventually
1349  * consume all available buffer space if it were ACKed.  By not ACKing
1350  * the data, we avoid this DoS scenario.
1351  *
1352  * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1353  * cookie is processed and a new socket is created.  In this case, any data
1354  * accompanying the SYN will be queued to the socket by tcp_input() and will
1355  * be ACKed either when the application sends response data or the delayed
1356  * ACK timer expires, whichever comes first.
1357  */
1358 struct socket *
syncache_add(struct in_conninfo * inc,struct tcpopt * to,struct tcphdr * th,struct inpcb * inp,struct socket * so,struct mbuf * m,void * tod,void * todctx,uint8_t iptos,uint16_t port)1359 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1360     struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod,
1361     void *todctx, uint8_t iptos, uint16_t port)
1362 {
1363 	struct tcpcb *tp;
1364 	struct socket *rv = NULL;
1365 	struct syncache *sc = NULL;
1366 	struct syncache_head *sch;
1367 	struct mbuf *ipopts = NULL;
1368 	u_int ltflags;
1369 	int win, ip_ttl, ip_tos;
1370 	char *s;
1371 #ifdef INET6
1372 	int autoflowlabel = 0;
1373 #endif
1374 #ifdef MAC
1375 	struct label *maclabel = NULL;
1376 #endif
1377 	struct syncache scs;
1378 	uint64_t tfo_response_cookie;
1379 	unsigned int *tfo_pending = NULL;
1380 	int tfo_cookie_valid = 0;
1381 	int tfo_response_cookie_valid = 0;
1382 	bool locked;
1383 
1384 	INP_RLOCK_ASSERT(inp);			/* listen socket */
1385 	KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1386 	    ("%s: unexpected tcp flags", __func__));
1387 
1388 	/*
1389 	 * Combine all so/tp operations very early to drop the INP lock as
1390 	 * soon as possible.
1391 	 */
1392 	KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1393 	tp = sototcpcb(so);
1394 
1395 #ifdef INET6
1396 	if (inc->inc_flags & INC_ISIPV6) {
1397 		if (inp->inp_flags & IN6P_AUTOFLOWLABEL) {
1398 			autoflowlabel = 1;
1399 		}
1400 		ip_ttl = in6_selecthlim(inp, NULL);
1401 		if ((inp->in6p_outputopts == NULL) ||
1402 		    (inp->in6p_outputopts->ip6po_tclass == -1)) {
1403 			ip_tos = 0;
1404 		} else {
1405 			ip_tos = inp->in6p_outputopts->ip6po_tclass;
1406 		}
1407 	}
1408 #endif
1409 #if defined(INET6) && defined(INET)
1410 	else
1411 #endif
1412 #ifdef INET
1413 	{
1414 		ip_ttl = inp->inp_ip_ttl;
1415 		ip_tos = inp->inp_ip_tos;
1416 	}
1417 #endif
1418 	win = so->sol_sbrcv_hiwat;
1419 	ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1420 
1421 	if (V_tcp_fastopen_server_enable && (tp->t_flags & TF_FASTOPEN) &&
1422 	    (tp->t_tfo_pending != NULL) &&
1423 	    (to->to_flags & TOF_FASTOPEN)) {
1424 		/*
1425 		 * Limit the number of pending TFO connections to
1426 		 * approximately half of the queue limit.  This prevents TFO
1427 		 * SYN floods from starving the service by filling the
1428 		 * listen queue with bogus TFO connections.
1429 		 */
1430 		if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1431 		    (so->sol_qlimit / 2)) {
1432 			int result;
1433 
1434 			result = tcp_fastopen_check_cookie(inc,
1435 			    to->to_tfo_cookie, to->to_tfo_len,
1436 			    &tfo_response_cookie);
1437 			tfo_cookie_valid = (result > 0);
1438 			tfo_response_cookie_valid = (result >= 0);
1439 		}
1440 
1441 		/*
1442 		 * Remember the TFO pending counter as it will have to be
1443 		 * decremented below if we don't make it to syncache_tfo_expand().
1444 		 */
1445 		tfo_pending = tp->t_tfo_pending;
1446 	}
1447 
1448 #ifdef MAC
1449 	if (mac_syncache_init(&maclabel) != 0) {
1450 		INP_RUNLOCK(inp);
1451 		goto done;
1452 	} else
1453 		mac_syncache_create(maclabel, inp);
1454 #endif
1455 	if (!tfo_cookie_valid)
1456 		INP_RUNLOCK(inp);
1457 
1458 	/*
1459 	 * Remember the IP options, if any.
1460 	 */
1461 #ifdef INET6
1462 	if (!(inc->inc_flags & INC_ISIPV6))
1463 #endif
1464 #ifdef INET
1465 		ipopts = (m) ? ip_srcroute(m) : NULL;
1466 #else
1467 		ipopts = NULL;
1468 #endif
1469 
1470 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1471 	/*
1472 	 * When the socket is TCP-MD5 enabled check that,
1473 	 *  - a signed packet is valid
1474 	 *  - a non-signed packet does not have a security association
1475 	 *
1476 	 *  If a signed packet fails validation or a non-signed packet has a
1477 	 *  security association, the packet will be dropped.
1478 	 */
1479 	if (ltflags & TF_SIGNATURE) {
1480 		if (to->to_flags & TOF_SIGNATURE) {
1481 			if (!TCPMD5_ENABLED() ||
1482 			    TCPMD5_INPUT(m, th, to->to_signature) != 0)
1483 				goto done;
1484 		} else {
1485 			if (TCPMD5_ENABLED() &&
1486 			    TCPMD5_INPUT(m, NULL, NULL) != ENOENT)
1487 				goto done;
1488 		}
1489 	} else if (to->to_flags & TOF_SIGNATURE)
1490 		goto done;
1491 #endif	/* TCP_SIGNATURE */
1492 	/*
1493 	 * See if we already have an entry for this connection.
1494 	 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1495 	 *
1496 	 * XXX: should the syncache be re-initialized with the contents
1497 	 * of the new SYN here (which may have different options?)
1498 	 *
1499 	 * XXX: We do not check the sequence number to see if this is a
1500 	 * real retransmit or a new connection attempt.  The question is
1501 	 * how to handle such a case; either ignore it as spoofed, or
1502 	 * drop the current entry and create a new one?
1503 	 */
1504 	if (syncache_cookiesonly()) {
1505 		sc = NULL;
1506 		sch = syncache_hashbucket(inc);
1507 		locked = false;
1508 	} else {
1509 		sc = syncache_lookup(inc, &sch);	/* returns locked sch */
1510 		locked = true;
1511 		SCH_LOCK_ASSERT(sch);
1512 	}
1513 	if (sc != NULL) {
1514 		if (tfo_cookie_valid)
1515 			INP_RUNLOCK(inp);
1516 		TCPSTAT_INC(tcps_sc_dupsyn);
1517 		if (ipopts) {
1518 			/*
1519 			 * If we were remembering a previous source route,
1520 			 * forget it and use the new one we've been given.
1521 			 */
1522 			if (sc->sc_ipopts)
1523 				(void)m_free(sc->sc_ipopts);
1524 			sc->sc_ipopts = ipopts;
1525 		}
1526 		/*
1527 		 * Update timestamp if present.
1528 		 */
1529 		if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1530 			sc->sc_tsreflect = to->to_tsval;
1531 		else
1532 			sc->sc_flags &= ~SCF_TIMESTAMP;
1533 		/*
1534 		 * Adjust ECN response if needed, e.g. different
1535 		 * IP ECN field, or a fallback by the remote host.
1536 		 */
1537 		if (sc->sc_flags & SCF_ECN_MASK) {
1538 			sc->sc_flags &= ~SCF_ECN_MASK;
1539 			sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1540 		}
1541 #ifdef MAC
1542 		/*
1543 		 * Since we have already unconditionally allocated label
1544 		 * storage, free it up.  The syncache entry will already
1545 		 * have an initialized label we can use.
1546 		 */
1547 		mac_syncache_destroy(&maclabel);
1548 #endif
1549 		TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1550 		/* Retransmit SYN|ACK and reset retransmit count. */
1551 		if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1552 			log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1553 			    "resetting timer and retransmitting SYN|ACK\n",
1554 			    s, __func__);
1555 			free(s, M_TCPLOG);
1556 		}
1557 		if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1558 			sc->sc_rxmits = 0;
1559 			syncache_timeout(sc, sch, 1);
1560 			TCPSTAT_INC(tcps_sndacks);
1561 			TCPSTAT_INC(tcps_sndtotal);
1562 		} else {
1563 			syncache_drop(sc, sch);
1564 			TCPSTAT_INC(tcps_sc_dropped);
1565 		}
1566 		SCH_UNLOCK(sch);
1567 		goto donenoprobe;
1568 	}
1569 
1570 	KASSERT(sc == NULL, ("sc(%p) != NULL", sc));
1571 	/*
1572 	 * Skip allocating a syncache entry if we are just going to discard
1573 	 * it later.
1574 	 */
1575 	if (!locked || tfo_cookie_valid) {
1576 		bzero(&scs, sizeof(scs));
1577 		sc = &scs;
1578 	} else {
1579 		sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1580 		if (sc == NULL) {
1581 			/*
1582 			 * The zone allocator couldn't provide more entries.
1583 			 * Treat this as if the cache was full; drop the oldest
1584 			 * entry and insert the new one.
1585 			 */
1586 			TCPSTAT_INC(tcps_sc_zonefail);
1587 			sc = TAILQ_LAST(&sch->sch_bucket, sch_head);
1588 			if (sc != NULL) {
1589 				sch->sch_last_overflow = time_uptime;
1590 				syncache_drop(sc, sch);
1591 				syncache_pause(inc);
1592 			}
1593 			sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1594 			if (sc == NULL) {
1595 				if (V_tcp_syncookies) {
1596 					bzero(&scs, sizeof(scs));
1597 					sc = &scs;
1598 				} else {
1599 					KASSERT(locked,
1600 					    ("%s: bucket unexpectedly unlocked",
1601 					    __func__));
1602 					SCH_UNLOCK(sch);
1603 					goto done;
1604 				}
1605 			}
1606 		}
1607 	}
1608 
1609 	KASSERT(sc != NULL, ("sc == NULL"));
1610 	if (!tfo_cookie_valid && tfo_response_cookie_valid)
1611 		sc->sc_tfo_cookie = &tfo_response_cookie;
1612 
1613 	/*
1614 	 * Fill in the syncache values.
1615 	 */
1616 #ifdef MAC
1617 	sc->sc_label = maclabel;
1618 #endif
1619 	/*
1620 	 * sc_cred is only used in syncache_pcblist() to list TCP endpoints in
1621 	 * TCPS_SYN_RECEIVED state when V_tcp_syncache.see_other is false.
1622 	 * Therefore, store the credentials and take a reference count only
1623 	 * when needed:
1624 	 * - sc is allocated from the zone and not using the on stack instance.
1625 	 * - the sysctl variable net.inet.tcp.syncache.see_other is false.
1626 	 * The reference count is decremented when a zone allocated sc is
1627 	 * freed in syncache_free().
1628 	 */
1629 	if (sc != &scs && !V_tcp_syncache.see_other)
1630 		sc->sc_cred = crhold(so->so_cred);
1631 	else
1632 		sc->sc_cred = NULL;
1633 	sc->sc_port = port;
1634 	sc->sc_ipopts = ipopts;
1635 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1636 	sc->sc_ip_tos = ip_tos;
1637 	sc->sc_ip_ttl = ip_ttl;
1638 #ifdef TCP_OFFLOAD
1639 	sc->sc_tod = tod;
1640 	sc->sc_todctx = todctx;
1641 #endif
1642 	sc->sc_irs = th->th_seq;
1643 	sc->sc_flags = 0;
1644 	sc->sc_flowlabel = 0;
1645 
1646 	/*
1647 	 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1648 	 * win was derived from socket earlier in the function.
1649 	 */
1650 	win = imax(win, 0);
1651 	win = imin(win, TCP_MAXWIN);
1652 	sc->sc_wnd = win;
1653 
1654 	if (V_tcp_do_rfc1323 &&
1655 	    !(ltflags & TF_NOOPT)) {
1656 		/*
1657 		 * A timestamp received in a SYN makes
1658 		 * it ok to send timestamp requests and replies.
1659 		 */
1660 		if ((to->to_flags & TOF_TS) && (V_tcp_do_rfc1323 != 2)) {
1661 			sc->sc_tsreflect = to->to_tsval;
1662 			sc->sc_flags |= SCF_TIMESTAMP;
1663 			sc->sc_tsoff = tcp_new_ts_offset(inc);
1664 		}
1665 		if ((to->to_flags & TOF_SCALE) && (V_tcp_do_rfc1323 != 3)) {
1666 			int wscale = 0;
1667 
1668 			/*
1669 			 * Pick the smallest possible scaling factor that
1670 			 * will still allow us to scale up to sb_max, aka
1671 			 * kern.ipc.maxsockbuf.
1672 			 *
1673 			 * We do this because there are broken firewalls that
1674 			 * will corrupt the window scale option, leading to
1675 			 * the other endpoint believing that our advertised
1676 			 * window is unscaled.  At scale factors larger than
1677 			 * 5 the unscaled window will drop below 1500 bytes,
1678 			 * leading to serious problems when traversing these
1679 			 * broken firewalls.
1680 			 *
1681 			 * With the default maxsockbuf of 256K, a scale factor
1682 			 * of 3 will be chosen by this algorithm.  Those who
1683 			 * choose a larger maxsockbuf should watch out
1684 			 * for the compatibility problems mentioned above.
1685 			 *
1686 			 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1687 			 * or <SYN,ACK>) segment itself is never scaled.
1688 			 */
1689 			while (wscale < TCP_MAX_WINSHIFT &&
1690 			    (TCP_MAXWIN << wscale) < sb_max)
1691 				wscale++;
1692 			sc->sc_requested_r_scale = wscale;
1693 			sc->sc_requested_s_scale = to->to_wscale;
1694 			sc->sc_flags |= SCF_WINSCALE;
1695 		}
1696 	}
1697 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1698 	/*
1699 	 * If incoming packet has an MD5 signature, flag this in the
1700 	 * syncache so that syncache_respond() will do the right thing
1701 	 * with the SYN+ACK.
1702 	 */
1703 	if (to->to_flags & TOF_SIGNATURE)
1704 		sc->sc_flags |= SCF_SIGNATURE;
1705 #endif	/* TCP_SIGNATURE */
1706 	if (to->to_flags & TOF_SACKPERM)
1707 		sc->sc_flags |= SCF_SACK;
1708 	if (to->to_flags & TOF_MSS)
1709 		sc->sc_peer_mss = to->to_mss;	/* peer mss may be zero */
1710 	if (ltflags & TF_NOOPT)
1711 		sc->sc_flags |= SCF_NOOPT;
1712 	/* ECN Handshake */
1713 	if (V_tcp_do_ecn && (tp->t_flags2 & TF2_CANNOT_DO_ECN) == 0)
1714 		sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1715 
1716 	if (V_tcp_syncookies)
1717 		sc->sc_iss = syncookie_generate(sch, sc);
1718 	else
1719 		sc->sc_iss = arc4random();
1720 #ifdef INET6
1721 	if (autoflowlabel) {
1722 		if (V_tcp_syncookies)
1723 			sc->sc_flowlabel = sc->sc_iss;
1724 		else
1725 			sc->sc_flowlabel = ip6_randomflowlabel();
1726 		sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1727 	}
1728 #endif
1729 	if (locked)
1730 		SCH_UNLOCK(sch);
1731 
1732 	if (tfo_cookie_valid) {
1733 		rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie);
1734 		/* INP_RUNLOCK(inp) will be performed by the caller */
1735 		goto tfo_expanded;
1736 	}
1737 
1738 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1739 	/*
1740 	 * Do a standard 3-way handshake.
1741 	 */
1742 	if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1743 		if (sc != &scs)
1744 			syncache_insert(sc, sch);   /* locks and unlocks sch */
1745 		TCPSTAT_INC(tcps_sndacks);
1746 		TCPSTAT_INC(tcps_sndtotal);
1747 	} else {
1748 		if (sc != &scs)
1749 			syncache_free(sc);
1750 		TCPSTAT_INC(tcps_sc_dropped);
1751 	}
1752 	goto donenoprobe;
1753 
1754 done:
1755 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1756 donenoprobe:
1757 	if (m)
1758 		m_freem(m);
1759 	/*
1760 	 * If tfo_pending is not NULL here, then a TFO SYN that did not
1761 	 * result in a new socket was processed and the associated pending
1762 	 * counter has not yet been decremented.  All such TFO processing paths
1763 	 * transit this point.
1764 	 */
1765 	if (tfo_pending != NULL)
1766 		tcp_fastopen_decrement_counter(tfo_pending);
1767 
1768 tfo_expanded:
1769 	if (sc == NULL || sc == &scs) {
1770 #ifdef MAC
1771 		mac_syncache_destroy(&maclabel);
1772 #endif
1773 		if (ipopts)
1774 			(void)m_free(ipopts);
1775 	}
1776 	return (rv);
1777 }
1778 
1779 /*
1780  * Send SYN|ACK or ACK to the peer.  Either in response to a peer's segment,
1781  * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1782  */
1783 static int
syncache_respond(struct syncache * sc,const struct mbuf * m0,int flags)1784 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1785 {
1786 	struct ip *ip = NULL;
1787 	struct mbuf *m;
1788 	struct tcphdr *th = NULL;
1789 	struct udphdr *udp = NULL;
1790 	int optlen, error = 0;	/* Make compiler happy */
1791 	u_int16_t hlen, tlen, mssopt, ulen;
1792 	struct tcpopt to;
1793 #ifdef INET6
1794 	struct ip6_hdr *ip6 = NULL;
1795 #endif
1796 
1797 	NET_EPOCH_ASSERT();
1798 
1799 	hlen =
1800 #ifdef INET6
1801 	       (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1802 #endif
1803 		sizeof(struct ip);
1804 	tlen = hlen + sizeof(struct tcphdr);
1805 	if (sc->sc_port) {
1806 		tlen += sizeof(struct udphdr);
1807 	}
1808 	/* Determine MSS we advertize to other end of connection. */
1809 	mssopt = tcp_mssopt(&sc->sc_inc);
1810 	if (sc->sc_port)
1811 		mssopt -= V_tcp_udp_tunneling_overhead;
1812 	mssopt = max(mssopt, V_tcp_minmss);
1813 
1814 	/* XXX: Assume that the entire packet will fit in a header mbuf. */
1815 	KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1816 	    ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + "
1817 	    "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port,
1818 	    max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN));
1819 
1820 	/* Create the IP+TCP header from scratch. */
1821 	m = m_gethdr(M_NOWAIT, MT_DATA);
1822 	if (m == NULL)
1823 		return (ENOBUFS);
1824 #ifdef MAC
1825 	mac_syncache_create_mbuf(sc->sc_label, m);
1826 #endif
1827 	m->m_data += max_linkhdr;
1828 	m->m_len = tlen;
1829 	m->m_pkthdr.len = tlen;
1830 	m->m_pkthdr.rcvif = NULL;
1831 
1832 #ifdef INET6
1833 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1834 		ip6 = mtod(m, struct ip6_hdr *);
1835 		ip6->ip6_vfc = IPV6_VERSION;
1836 		ip6->ip6_src = sc->sc_inc.inc6_laddr;
1837 		ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1838 		ip6->ip6_plen = htons(tlen - hlen);
1839 		/* ip6_hlim is set after checksum */
1840 		/* Zero out traffic class and flow label. */
1841 		ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1842 		ip6->ip6_flow |= sc->sc_flowlabel;
1843 		if (sc->sc_port != 0) {
1844 			ip6->ip6_nxt = IPPROTO_UDP;
1845 			udp = (struct udphdr *)(ip6 + 1);
1846 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1847 			udp->uh_dport = sc->sc_port;
1848 			ulen = (tlen - sizeof(struct ip6_hdr));
1849 			th = (struct tcphdr *)(udp + 1);
1850 		} else {
1851 			ip6->ip6_nxt = IPPROTO_TCP;
1852 			th = (struct tcphdr *)(ip6 + 1);
1853 		}
1854 		ip6->ip6_flow |= htonl(sc->sc_ip_tos << IPV6_FLOWLABEL_LEN);
1855 	}
1856 #endif
1857 #if defined(INET6) && defined(INET)
1858 	else
1859 #endif
1860 #ifdef INET
1861 	{
1862 		ip = mtod(m, struct ip *);
1863 		ip->ip_v = IPVERSION;
1864 		ip->ip_hl = sizeof(struct ip) >> 2;
1865 		ip->ip_len = htons(tlen);
1866 		ip->ip_id = 0;
1867 		ip->ip_off = 0;
1868 		ip->ip_sum = 0;
1869 		ip->ip_src = sc->sc_inc.inc_laddr;
1870 		ip->ip_dst = sc->sc_inc.inc_faddr;
1871 		ip->ip_ttl = sc->sc_ip_ttl;
1872 		ip->ip_tos = sc->sc_ip_tos;
1873 
1874 		/*
1875 		 * See if we should do MTU discovery.  Route lookups are
1876 		 * expensive, so we will only unset the DF bit if:
1877 		 *
1878 		 *	1) path_mtu_discovery is disabled
1879 		 *	2) the SCF_UNREACH flag has been set
1880 		 */
1881 		if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1882 		       ip->ip_off |= htons(IP_DF);
1883 		if (sc->sc_port == 0) {
1884 			ip->ip_p = IPPROTO_TCP;
1885 			th = (struct tcphdr *)(ip + 1);
1886 		} else {
1887 			ip->ip_p = IPPROTO_UDP;
1888 			udp = (struct udphdr *)(ip + 1);
1889 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1890 			udp->uh_dport = sc->sc_port;
1891 			ulen = (tlen - sizeof(struct ip));
1892 			th = (struct tcphdr *)(udp + 1);
1893 		}
1894 	}
1895 #endif /* INET */
1896 	th->th_sport = sc->sc_inc.inc_lport;
1897 	th->th_dport = sc->sc_inc.inc_fport;
1898 
1899 	if (flags & TH_SYN)
1900 		th->th_seq = htonl(sc->sc_iss);
1901 	else
1902 		th->th_seq = htonl(sc->sc_iss + 1);
1903 	th->th_ack = htonl(sc->sc_irs + 1);
1904 	th->th_off = sizeof(struct tcphdr) >> 2;
1905 	th->th_win = htons(sc->sc_wnd);
1906 	th->th_urp = 0;
1907 
1908 	flags = tcp_ecn_syncache_respond(flags, sc);
1909 	tcp_set_flags(th, flags);
1910 
1911 	/* Tack on the TCP options. */
1912 	if ((sc->sc_flags & SCF_NOOPT) == 0) {
1913 		to.to_flags = 0;
1914 
1915 		if (flags & TH_SYN) {
1916 			to.to_mss = mssopt;
1917 			to.to_flags = TOF_MSS;
1918 			if (sc->sc_flags & SCF_WINSCALE) {
1919 				to.to_wscale = sc->sc_requested_r_scale;
1920 				to.to_flags |= TOF_SCALE;
1921 			}
1922 			if (sc->sc_flags & SCF_SACK)
1923 				to.to_flags |= TOF_SACKPERM;
1924 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1925 			if (sc->sc_flags & SCF_SIGNATURE)
1926 				to.to_flags |= TOF_SIGNATURE;
1927 #endif
1928 			if (sc->sc_tfo_cookie) {
1929 				to.to_flags |= TOF_FASTOPEN;
1930 				to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1931 				to.to_tfo_cookie = sc->sc_tfo_cookie;
1932 				/* don't send cookie again when retransmitting response */
1933 				sc->sc_tfo_cookie = NULL;
1934 			}
1935 		}
1936 		if (sc->sc_flags & SCF_TIMESTAMP) {
1937 			to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1938 			to.to_tsecr = sc->sc_tsreflect;
1939 			to.to_flags |= TOF_TS;
1940 		}
1941 		optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1942 
1943 		/* Adjust headers by option size. */
1944 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1945 		m->m_len += optlen;
1946 		m->m_pkthdr.len += optlen;
1947 #ifdef INET6
1948 		if (sc->sc_inc.inc_flags & INC_ISIPV6)
1949 			ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1950 		else
1951 #endif
1952 			ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1953 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1954 		if (sc->sc_flags & SCF_SIGNATURE) {
1955 			KASSERT(to.to_flags & TOF_SIGNATURE,
1956 			    ("tcp_addoptions() didn't set tcp_signature"));
1957 
1958 			/* NOTE: to.to_signature is inside of mbuf */
1959 			if (!TCPMD5_ENABLED() ||
1960 			    TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1961 				m_freem(m);
1962 				return (EACCES);
1963 			}
1964 		}
1965 #endif
1966 	} else
1967 		optlen = 0;
1968 
1969 	if (udp) {
1970 		ulen += optlen;
1971 		udp->uh_ulen = htons(ulen);
1972 	}
1973 	M_SETFIB(m, sc->sc_inc.inc_fibnum);
1974 	/*
1975 	 * If we have peer's SYN and it has a flowid, then let's assign it to
1976 	 * our SYN|ACK.  ip6_output() and ip_output() will not assign flowid
1977 	 * to SYN|ACK due to lack of inp here.
1978 	 */
1979 	if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1980 		m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1981 		M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1982 	}
1983 #ifdef INET6
1984 	if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1985 		if (sc->sc_port) {
1986 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
1987 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1988 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen,
1989 			      IPPROTO_UDP, 0);
1990 			th->th_sum = htons(0);
1991 		} else {
1992 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1993 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1994 			th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1995 			    IPPROTO_TCP, 0);
1996 		}
1997 		ip6->ip6_hlim = sc->sc_ip_ttl;
1998 #ifdef TCP_OFFLOAD
1999 		if (ADDED_BY_TOE(sc)) {
2000 			struct toedev *tod = sc->sc_tod;
2001 
2002 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2003 
2004 			return (error);
2005 		}
2006 #endif
2007 		TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
2008 		error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
2009 	}
2010 #endif
2011 #if defined(INET6) && defined(INET)
2012 	else
2013 #endif
2014 #ifdef INET
2015 	{
2016 		if (sc->sc_port) {
2017 			m->m_pkthdr.csum_flags = CSUM_UDP;
2018 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2019 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
2020 			      ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
2021 			th->th_sum = htons(0);
2022 		} else {
2023 			m->m_pkthdr.csum_flags = CSUM_TCP;
2024 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2025 			th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
2026 			    htons(tlen + optlen - hlen + IPPROTO_TCP));
2027 		}
2028 #ifdef TCP_OFFLOAD
2029 		if (ADDED_BY_TOE(sc)) {
2030 			struct toedev *tod = sc->sc_tod;
2031 
2032 			error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2033 
2034 			return (error);
2035 		}
2036 #endif
2037 		TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
2038 		error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
2039 	}
2040 #endif
2041 	return (error);
2042 }
2043 
2044 /*
2045  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
2046  * that exceed the capacity of the syncache by avoiding the storage of any
2047  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
2048  * attacks where the attacker does not have access to our responses.
2049  *
2050  * Syncookies encode and include all necessary information about the
2051  * connection setup within the SYN|ACK that we send back.  That way we
2052  * can avoid keeping any local state until the ACK to our SYN|ACK returns
2053  * (if ever).  Normally the syncache and syncookies are running in parallel
2054  * with the latter taking over when the former is exhausted.  When matching
2055  * syncache entry is found the syncookie is ignored.
2056  *
2057  * The only reliable information persisting the 3WHS is our initial sequence
2058  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
2059  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
2060  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
2061  * returns and signifies a legitimate connection if it matches the ACK.
2062  *
2063  * The available space of 32 bits to store the hash and to encode the SYN
2064  * option information is very tight and we should have at least 24 bits for
2065  * the MAC to keep the number of guesses by blind spoofing reasonably high.
2066  *
2067  * SYN option information we have to encode to fully restore a connection:
2068  * MSS: is imporant to chose an optimal segment size to avoid IP level
2069  *   fragmentation along the path.  The common MSS values can be encoded
2070  *   in a 3-bit table.  Uncommon values are captured by the next lower value
2071  *   in the table leading to a slight increase in packetization overhead.
2072  * WSCALE: is necessary to allow large windows to be used for high delay-
2073  *   bandwidth product links.  Not scaling the window when it was initially
2074  *   negotiated is bad for performance as lack of scaling further decreases
2075  *   the apparent available send window.  We only need to encode the WSCALE
2076  *   we received from the remote end.  Our end can be recalculated at any
2077  *   time.  The common WSCALE values can be encoded in a 3-bit table.
2078  *   Uncommon values are captured by the next lower value in the table
2079  *   making us under-estimate the available window size halving our
2080  *   theoretically possible maximum throughput for that connection.
2081  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2082  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2083  *   that are included in all segments on a connection.  We enable them when
2084  *   the ACK has them.
2085  *
2086  * Security of syncookies and attack vectors:
2087  *
2088  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2089  * together with the gloabl secret to make it unique per connection attempt.
2090  * Thus any change of any of those parameters results in a different MAC output
2091  * in an unpredictable way unless a collision is encountered.  24 bits of the
2092  * MAC are embedded into the ISS.
2093  *
2094  * To prevent replay attacks two rotating global secrets are updated with a
2095  * new random value every 15 seconds.  The life-time of a syncookie is thus
2096  * 15-30 seconds.
2097  *
2098  * Vector 1: Attacking the secret.  This requires finding a weakness in the
2099  * MAC itself or the way it is used here.  The attacker can do a chosen plain
2100  * text attack by varying and testing the all parameters under his control.
2101  * The strength depends on the size and randomness of the secret, and the
2102  * cryptographic security of the MAC function.  Due to the constant updating
2103  * of the secret the attacker has at most 29.999 seconds to find the secret
2104  * and launch spoofed connections.  After that he has to start all over again.
2105  *
2106  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
2107  * size an average of 4,823 attempts are required for a 50% chance of success
2108  * to spoof a single syncookie (birthday collision paradox).  However the
2109  * attacker is blind and doesn't know if one of his attempts succeeded unless
2110  * he has a side channel to interfere success from.  A single connection setup
2111  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2112  * This many attempts are required for each one blind spoofed connection.  For
2113  * every additional spoofed connection he has to launch another N attempts.
2114  * Thus for a sustained rate 100 spoofed connections per second approximately
2115  * 1,800,000 packets per second would have to be sent.
2116  *
2117  * NB: The MAC function should be fast so that it doesn't become a CPU
2118  * exhaustion attack vector itself.
2119  *
2120  * References:
2121  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2122  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2123  *   http://cr.yp.to/syncookies.html    (overview)
2124  *   http://cr.yp.to/syncookies/archive (details)
2125  *
2126  *
2127  * Schematic construction of a syncookie enabled Initial Sequence Number:
2128  *  0        1         2         3
2129  *  12345678901234567890123456789012
2130  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2131  *
2132  *  x 24 MAC (truncated)
2133  *  W  3 Send Window Scale index
2134  *  M  3 MSS index
2135  *  S  1 SACK permitted
2136  *  P  1 Odd/even secret
2137  */
2138 
2139 /*
2140  * Distribution and probability of certain MSS values.  Those in between are
2141  * rounded down to the next lower one.
2142  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2143  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
2144  */
2145 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2146 
2147 /*
2148  * Distribution and probability of certain WSCALE values.  We have to map the
2149  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2150  * bits based on prevalence of certain values.  Where we don't have an exact
2151  * match for are rounded down to the next lower one letting us under-estimate
2152  * the true available window.  At the moment this would happen only for the
2153  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2154  * and window size).  The absence of the WSCALE option (no scaling in either
2155  * direction) is encoded with index zero.
2156  * [WSCALE values histograms, Allman, 2012]
2157  *                            X 10 10 35  5  6 14 10%   by host
2158  *                            X 11  4  5  5 18 49  3%   by connections
2159  */
2160 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2161 
2162 /*
2163  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
2164  * and good cryptographic properties.
2165  */
2166 static uint32_t
syncookie_mac(struct in_conninfo * inc,tcp_seq irs,uint8_t flags,uint8_t * secbits,uintptr_t secmod)2167 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2168     uint8_t *secbits, uintptr_t secmod)
2169 {
2170 	SIPHASH_CTX ctx;
2171 	uint32_t siphash[2];
2172 
2173 	SipHash24_Init(&ctx);
2174 	SipHash_SetKey(&ctx, secbits);
2175 	switch (inc->inc_flags & INC_ISIPV6) {
2176 #ifdef INET
2177 	case 0:
2178 		SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2179 		SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2180 		break;
2181 #endif
2182 #ifdef INET6
2183 	case INC_ISIPV6:
2184 		SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2185 		SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2186 		break;
2187 #endif
2188 	}
2189 	SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2190 	SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2191 	SipHash_Update(&ctx, &irs, sizeof(irs));
2192 	SipHash_Update(&ctx, &flags, sizeof(flags));
2193 	SipHash_Update(&ctx, &secmod, sizeof(secmod));
2194 	SipHash_Final((u_int8_t *)&siphash, &ctx);
2195 
2196 	return (siphash[0] ^ siphash[1]);
2197 }
2198 
2199 static tcp_seq
syncookie_generate(struct syncache_head * sch,struct syncache * sc)2200 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2201 {
2202 	u_int i, secbit, wscale;
2203 	uint32_t iss, hash;
2204 	uint8_t *secbits;
2205 	union syncookie cookie;
2206 
2207 	cookie.cookie = 0;
2208 
2209 	/* Map our computed MSS into the 3-bit index. */
2210 	for (i = nitems(tcp_sc_msstab) - 1;
2211 	     tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2212 	     i--)
2213 		;
2214 	cookie.flags.mss_idx = i;
2215 
2216 	/*
2217 	 * Map the send window scale into the 3-bit index but only if
2218 	 * the wscale option was received.
2219 	 */
2220 	if (sc->sc_flags & SCF_WINSCALE) {
2221 		wscale = sc->sc_requested_s_scale;
2222 		for (i = nitems(tcp_sc_wstab) - 1;
2223 		    tcp_sc_wstab[i] > wscale && i > 0;
2224 		     i--)
2225 			;
2226 		cookie.flags.wscale_idx = i;
2227 	}
2228 
2229 	/* Can we do SACK? */
2230 	if (sc->sc_flags & SCF_SACK)
2231 		cookie.flags.sack_ok = 1;
2232 
2233 	/* Which of the two secrets to use. */
2234 	secbit = V_tcp_syncache.secret.oddeven & 0x1;
2235 	cookie.flags.odd_even = secbit;
2236 
2237 	secbits = V_tcp_syncache.secret.key[secbit];
2238 	hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2239 	    (uintptr_t)sch);
2240 
2241 	/*
2242 	 * Put the flags into the hash and XOR them to get better ISS number
2243 	 * variance.  This doesn't enhance the cryptographic strength and is
2244 	 * done to prevent the 8 cookie bits from showing up directly on the
2245 	 * wire.
2246 	 */
2247 	iss = hash & ~0xff;
2248 	iss |= cookie.cookie ^ (hash >> 24);
2249 
2250 	TCPSTAT_INC(tcps_sc_sendcookie);
2251 	return (iss);
2252 }
2253 
2254 static struct syncache *
syncookie_lookup(struct in_conninfo * inc,struct syncache_head * sch,struct syncache * sc,struct tcphdr * th,struct tcpopt * to,struct socket * lso,uint16_t port)2255 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2256     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2257     struct socket *lso, uint16_t port)
2258 {
2259 	uint32_t hash;
2260 	uint8_t *secbits;
2261 	tcp_seq ack, seq;
2262 	int wnd, wscale = 0;
2263 	union syncookie cookie;
2264 
2265 	/*
2266 	 * Pull information out of SYN-ACK/ACK and revert sequence number
2267 	 * advances.
2268 	 */
2269 	ack = th->th_ack - 1;
2270 	seq = th->th_seq - 1;
2271 
2272 	/*
2273 	 * Unpack the flags containing enough information to restore the
2274 	 * connection.
2275 	 */
2276 	cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2277 
2278 	/* Which of the two secrets to use. */
2279 	secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2280 
2281 	hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2282 
2283 	/* The recomputed hash matches the ACK if this was a genuine cookie. */
2284 	if ((ack & ~0xff) != (hash & ~0xff))
2285 		return (NULL);
2286 
2287 	/* Fill in the syncache values. */
2288 	sc->sc_flags = 0;
2289 	bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2290 	sc->sc_ipopts = NULL;
2291 
2292 	sc->sc_irs = seq;
2293 	sc->sc_iss = ack;
2294 
2295 	switch (inc->inc_flags & INC_ISIPV6) {
2296 #ifdef INET
2297 	case 0:
2298 		sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2299 		sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2300 		break;
2301 #endif
2302 #ifdef INET6
2303 	case INC_ISIPV6:
2304 		if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2305 			sc->sc_flowlabel =
2306 			    htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2307 		break;
2308 #endif
2309 	}
2310 
2311 	sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2312 
2313 	/* We can simply recompute receive window scale we sent earlier. */
2314 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2315 		wscale++;
2316 
2317 	/* Only use wscale if it was enabled in the orignal SYN. */
2318 	if (cookie.flags.wscale_idx > 0) {
2319 		sc->sc_requested_r_scale = wscale;
2320 		sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2321 		sc->sc_flags |= SCF_WINSCALE;
2322 	}
2323 
2324 	wnd = lso->sol_sbrcv_hiwat;
2325 	wnd = imax(wnd, 0);
2326 	wnd = imin(wnd, TCP_MAXWIN);
2327 	sc->sc_wnd = wnd;
2328 
2329 	if (cookie.flags.sack_ok)
2330 		sc->sc_flags |= SCF_SACK;
2331 
2332 	if (to->to_flags & TOF_TS) {
2333 		sc->sc_flags |= SCF_TIMESTAMP;
2334 		sc->sc_tsreflect = to->to_tsval;
2335 		sc->sc_tsoff = tcp_new_ts_offset(inc);
2336 	}
2337 
2338 	if (to->to_flags & TOF_SIGNATURE)
2339 		sc->sc_flags |= SCF_SIGNATURE;
2340 
2341 	sc->sc_rxmits = 0;
2342 
2343 	sc->sc_port = port;
2344 
2345 	TCPSTAT_INC(tcps_sc_recvcookie);
2346 	return (sc);
2347 }
2348 
2349 #ifdef INVARIANTS
2350 static int
syncookie_cmp(struct in_conninfo * inc,struct syncache_head * sch,struct syncache * sc,struct tcphdr * th,struct tcpopt * to,struct socket * lso,uint16_t port)2351 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2352     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2353     struct socket *lso, uint16_t port)
2354 {
2355 	struct syncache scs, *scx;
2356 	char *s;
2357 
2358 	bzero(&scs, sizeof(scs));
2359 	scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port);
2360 
2361 	if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2362 		return (0);
2363 
2364 	if (scx != NULL) {
2365 		if (sc->sc_peer_mss != scx->sc_peer_mss)
2366 			log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2367 			    s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2368 
2369 		if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2370 			log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2371 			    s, __func__, sc->sc_requested_r_scale,
2372 			    scx->sc_requested_r_scale);
2373 
2374 		if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2375 			log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2376 			    s, __func__, sc->sc_requested_s_scale,
2377 			    scx->sc_requested_s_scale);
2378 
2379 		if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2380 			log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2381 	}
2382 
2383 	if (s != NULL)
2384 		free(s, M_TCPLOG);
2385 	return (0);
2386 }
2387 #endif /* INVARIANTS */
2388 
2389 static void
syncookie_reseed(void * arg)2390 syncookie_reseed(void *arg)
2391 {
2392 	struct tcp_syncache *sc = arg;
2393 	uint8_t *secbits;
2394 	int secbit;
2395 
2396 	/*
2397 	 * Reseeding the secret doesn't have to be protected by a lock.
2398 	 * It only must be ensured that the new random values are visible
2399 	 * to all CPUs in a SMP environment.  The atomic with release
2400 	 * semantics ensures that.
2401 	 */
2402 	secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2403 	secbits = sc->secret.key[secbit];
2404 	arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2405 	atomic_add_rel_int(&sc->secret.oddeven, 1);
2406 
2407 	/* Reschedule ourself. */
2408 	callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2409 }
2410 
2411 /*
2412  * We have overflowed a bucket. Let's pause dealing with the syncache.
2413  * This function will increment the bucketoverflow statistics appropriately
2414  * (once per pause when pausing is enabled; otherwise, once per overflow).
2415  */
2416 static void
syncache_pause(struct in_conninfo * inc)2417 syncache_pause(struct in_conninfo *inc)
2418 {
2419 	time_t delta;
2420 	const char *s;
2421 
2422 	/* XXX:
2423 	 * 2. Add sysctl read here so we don't get the benefit of this
2424 	 * change without the new sysctl.
2425 	 */
2426 
2427 	/*
2428 	 * Try an unlocked read. If we already know that another thread
2429 	 * has activated the feature, there is no need to proceed.
2430 	 */
2431 	if (V_tcp_syncache.paused)
2432 		return;
2433 
2434 	/* Are cookied enabled? If not, we can't pause. */
2435 	if (!V_tcp_syncookies) {
2436 		TCPSTAT_INC(tcps_sc_bucketoverflow);
2437 		return;
2438 	}
2439 
2440 	/*
2441 	 * We may be the first thread to find an overflow. Get the lock
2442 	 * and evaluate if we need to take action.
2443 	 */
2444 	mtx_lock(&V_tcp_syncache.pause_mtx);
2445 	if (V_tcp_syncache.paused) {
2446 		mtx_unlock(&V_tcp_syncache.pause_mtx);
2447 		return;
2448 	}
2449 
2450 	/* Activate protection. */
2451 	V_tcp_syncache.paused = true;
2452 	TCPSTAT_INC(tcps_sc_bucketoverflow);
2453 
2454 	/*
2455 	 * Determine the last backoff time. If we are seeing a re-newed
2456 	 * attack within that same time after last reactivating the syncache,
2457 	 * consider it an extension of the same attack.
2458 	 */
2459 	delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2460 	if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2461 		if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2462 			delta <<= 1;
2463 			V_tcp_syncache.pause_backoff++;
2464 		}
2465 	} else {
2466 		delta = TCP_SYNCACHE_PAUSE_TIME;
2467 		V_tcp_syncache.pause_backoff = 0;
2468 	}
2469 
2470 	/* Log a warning, including IP addresses, if able. */
2471 	if (inc != NULL)
2472 		s = tcp_log_addrs(inc, NULL, NULL, NULL);
2473 	else
2474 		s = (const char *)NULL;
2475 	log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2476 	    "the next %lld seconds%s%s%s\n", (long long)delta,
2477 	    (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2478 	    (s != NULL) ? ")" : "");
2479 	free(__DECONST(void *, s), M_TCPLOG);
2480 
2481 	/* Use the calculated delta to set a new pause time. */
2482 	V_tcp_syncache.pause_until = time_uptime + delta;
2483 	callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2484 	    &V_tcp_syncache);
2485 	mtx_unlock(&V_tcp_syncache.pause_mtx);
2486 }
2487 
2488 /* Evaluate whether we need to unpause. */
2489 static void
syncache_unpause(void * arg)2490 syncache_unpause(void *arg)
2491 {
2492 	struct tcp_syncache *sc;
2493 	time_t delta;
2494 
2495 	sc = arg;
2496 	mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2497 	callout_deactivate(&sc->pause_co);
2498 
2499 	/*
2500 	 * Check to make sure we are not running early. If the pause
2501 	 * time has expired, then deactivate the protection.
2502 	 */
2503 	if ((delta = sc->pause_until - time_uptime) > 0)
2504 		callout_schedule(&sc->pause_co, delta * hz);
2505 	else
2506 		sc->paused = false;
2507 }
2508 
2509 /*
2510  * Exports the syncache entries to userland so that netstat can display
2511  * them alongside the other sockets.  This function is intended to be
2512  * called only from tcp_pcblist.
2513  *
2514  * Due to concurrency on an active system, the number of pcbs exported
2515  * may have no relation to max_pcbs.  max_pcbs merely indicates the
2516  * amount of space the caller allocated for this function to use.
2517  */
2518 int
syncache_pcblist(struct sysctl_req * req)2519 syncache_pcblist(struct sysctl_req *req)
2520 {
2521 	struct xtcpcb xt;
2522 	struct syncache *sc;
2523 	struct syncache_head *sch;
2524 	int error, i;
2525 
2526 	bzero(&xt, sizeof(xt));
2527 	xt.xt_len = sizeof(xt);
2528 	xt.t_state = TCPS_SYN_RECEIVED;
2529 	xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2530 	xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2531 	xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2532 	xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2533 
2534 	for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2535 		sch = &V_tcp_syncache.hashbase[i];
2536 		SCH_LOCK(sch);
2537 		TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2538 			if (sc->sc_cred != NULL &&
2539 			    cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2540 				continue;
2541 			if (sc->sc_inc.inc_flags & INC_ISIPV6)
2542 				xt.xt_inp.inp_vflag = INP_IPV6;
2543 			else
2544 				xt.xt_inp.inp_vflag = INP_IPV4;
2545 			xt.xt_encaps_port = sc->sc_port;
2546 			bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2547 			    sizeof (struct in_conninfo));
2548 			error = SYSCTL_OUT(req, &xt, sizeof xt);
2549 			if (error) {
2550 				SCH_UNLOCK(sch);
2551 				return (0);
2552 			}
2553 		}
2554 		SCH_UNLOCK(sch);
2555 	}
2556 
2557 	return (0);
2558 }
2559