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(bool, tcp_syncookies) = true;
106 #define V_tcp_syncookies VNET(tcp_syncookies)
107 SYSCTL_BOOL(_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(bool, tcp_syncookiesonly) = false;
112 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly)
113 SYSCTL_BOOL(_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 void syncache_send_challenge_ack(struct syncache *, struct mbuf *);
126 static struct socket *syncache_socket(struct syncache *, struct socket *,
127 struct mbuf *m);
128 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch,
129 int docallout);
130 static void syncache_timer(void *);
131
132 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
133 uint8_t *, uintptr_t);
134 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *);
135 static bool syncookie_expand(struct in_conninfo *,
136 const struct syncache_head *, struct syncache *,
137 struct tcphdr *, struct tcpopt *, struct socket *,
138 uint16_t);
139 static void syncache_pause(struct in_conninfo *);
140 static void syncache_unpause(void *);
141 static void syncookie_reseed(void *);
142 #ifdef INVARIANTS
143 static void syncookie_cmp(struct in_conninfo *,
144 const struct syncache_head *, struct syncache *,
145 struct tcphdr *, struct tcpopt *, struct socket *,
146 uint16_t);
147 #endif
148
149 /*
150 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
151 * 3 retransmits corresponds to a timeout with default values of
152 * tcp_rexmit_initial * ( 1 +
153 * tcp_backoff[1] +
154 * tcp_backoff[2] +
155 * tcp_backoff[3]) + 3 * tcp_rexmit_slop,
156 * 1000 ms * (1 + 2 + 4 + 8) + 3 * 200 ms = 15600 ms,
157 * the odds are that the user has given up attempting to connect by then.
158 */
159 #define SYNCACHE_MAXREXMTS 3
160
161 /* Arbitrary values */
162 #define TCP_SYNCACHE_HASHSIZE 512
163 #define TCP_SYNCACHE_BUCKETLIMIT 30
164
165 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
166 #define V_tcp_syncache VNET(tcp_syncache)
167
168 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache,
169 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
170 "TCP SYN cache");
171
172 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
173 &VNET_NAME(tcp_syncache.bucket_limit), 0,
174 "Per-bucket hash limit for syncache");
175
176 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
177 &VNET_NAME(tcp_syncache.cache_limit), 0,
178 "Overall entry limit for syncache");
179
180 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
181 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
182
183 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
184 &VNET_NAME(tcp_syncache.hashsize), 0,
185 "Size of TCP syncache hashtable");
186
187 SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET |
188 CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0,
189 "All syncache(4) entries are visible, ignoring UID/GID, jail(2) "
190 "and mac(4) checks");
191
192 static int
sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)193 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
194 {
195 int error;
196 u_int new;
197
198 new = V_tcp_syncache.rexmt_limit;
199 error = sysctl_handle_int(oidp, &new, 0, req);
200 if ((error == 0) && (req->newptr != NULL)) {
201 if (new > TCP_MAXRXTSHIFT)
202 error = EINVAL;
203 else
204 V_tcp_syncache.rexmt_limit = new;
205 }
206 return (error);
207 }
208
209 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
210 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
211 &VNET_NAME(tcp_syncache.rexmt_limit), 0,
212 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "IU",
213 "Limit on SYN/ACK retransmissions");
214
215 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
216 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
217 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
218 "Send reset on socket allocation failure");
219
220 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
221
222 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx)
223 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx)
224 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED)
225
226 /*
227 * Requires the syncache entry to be already removed from the bucket list.
228 */
229 static void
syncache_free(struct syncache * sc)230 syncache_free(struct syncache *sc)
231 {
232
233 if (sc->sc_ipopts)
234 (void)m_free(sc->sc_ipopts);
235 if (sc->sc_cred)
236 crfree(sc->sc_cred);
237 #ifdef MAC
238 mac_syncache_destroy(&sc->sc_label);
239 #endif
240
241 uma_zfree(V_tcp_syncache.zone, sc);
242 }
243
244 void
syncache_init(void)245 syncache_init(void)
246 {
247 int i;
248
249 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
250 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
251 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
252 V_tcp_syncache.hash_secret = arc4random();
253
254 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
255 &V_tcp_syncache.hashsize);
256 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
257 &V_tcp_syncache.bucket_limit);
258 if (!powerof2(V_tcp_syncache.hashsize) ||
259 V_tcp_syncache.hashsize == 0) {
260 printf("WARNING: syncache hash size is not a power of 2.\n");
261 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
262 }
263 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
264
265 /* Set limits. */
266 V_tcp_syncache.cache_limit =
267 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
268 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
269 &V_tcp_syncache.cache_limit);
270
271 /* Allocate the hash table. */
272 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
273 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
274
275 #ifdef VIMAGE
276 V_tcp_syncache.vnet = curvnet;
277 #endif
278
279 /* Initialize the hash buckets. */
280 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
281 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
282 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
283 NULL, MTX_DEF);
284 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
285 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
286 V_tcp_syncache.hashbase[i].sch_length = 0;
287 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
288 V_tcp_syncache.hashbase[i].sch_last_overflow =
289 -(SYNCOOKIE_LIFETIME + 1);
290 }
291
292 /* Create the syncache entry zone. */
293 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
294 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
295 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
296 V_tcp_syncache.cache_limit);
297
298 /* Start the SYN cookie reseeder callout. */
299 callout_init(&V_tcp_syncache.secret.reseed, 1);
300 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
301 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
302 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
303 syncookie_reseed, &V_tcp_syncache);
304
305 /* Initialize the pause machinery. */
306 mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF);
307 callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx,
308 0);
309 V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME;
310 V_tcp_syncache.pause_backoff = 0;
311 V_tcp_syncache.paused = false;
312 }
313
314 #ifdef VIMAGE
315 void
syncache_destroy(void)316 syncache_destroy(void)
317 {
318 struct syncache_head *sch;
319 struct syncache *sc, *nsc;
320 int i;
321
322 /*
323 * Stop the re-seed timer before freeing resources. No need to
324 * possibly schedule it another time.
325 */
326 callout_drain(&V_tcp_syncache.secret.reseed);
327
328 /* Stop the SYN cache pause callout. */
329 mtx_lock(&V_tcp_syncache.pause_mtx);
330 if (callout_stop(&V_tcp_syncache.pause_co) == 0) {
331 mtx_unlock(&V_tcp_syncache.pause_mtx);
332 callout_drain(&V_tcp_syncache.pause_co);
333 } else
334 mtx_unlock(&V_tcp_syncache.pause_mtx);
335
336 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */
337 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
338 sch = &V_tcp_syncache.hashbase[i];
339 callout_drain(&sch->sch_timer);
340
341 SCH_LOCK(sch);
342 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
343 syncache_drop(sc, sch);
344 SCH_UNLOCK(sch);
345 KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
346 ("%s: sch->sch_bucket not empty", __func__));
347 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
348 __func__, sch->sch_length));
349 mtx_destroy(&sch->sch_mtx);
350 }
351
352 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
353 ("%s: cache_count not 0", __func__));
354
355 /* Free the allocated global resources. */
356 uma_zdestroy(V_tcp_syncache.zone);
357 free(V_tcp_syncache.hashbase, M_SYNCACHE);
358 mtx_destroy(&V_tcp_syncache.pause_mtx);
359 }
360 #endif
361
362 /*
363 * Inserts a syncache entry into the specified bucket row.
364 * Locks and unlocks the syncache_head autonomously.
365 */
366 static void
syncache_insert(struct syncache * sc,struct syncache_head * sch)367 syncache_insert(struct syncache *sc, struct syncache_head *sch)
368 {
369 struct syncache *sc2;
370
371 SCH_LOCK(sch);
372
373 /*
374 * Make sure that we don't overflow the per-bucket limit.
375 * If the bucket is full, toss the oldest element.
376 */
377 if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
378 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
379 ("sch->sch_length incorrect"));
380 syncache_pause(&sc->sc_inc);
381 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
382 sch->sch_last_overflow = time_uptime;
383 syncache_drop(sc2, sch);
384 }
385
386 /* Put it into the bucket. */
387 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
388 sch->sch_length++;
389
390 #ifdef TCP_OFFLOAD
391 if (ADDED_BY_TOE(sc)) {
392 struct toedev *tod = sc->sc_tod;
393
394 tod->tod_syncache_added(tod, sc->sc_todctx);
395 }
396 #endif
397
398 /* Reinitialize the bucket row's timer. */
399 if (sch->sch_length == 1)
400 sch->sch_nextc = ticks + INT_MAX;
401 syncache_timeout(sc, sch, 1);
402
403 SCH_UNLOCK(sch);
404
405 TCPSTATES_INC(TCPS_SYN_RECEIVED);
406 TCPSTAT_INC(tcps_sc_added);
407 }
408
409 /*
410 * Remove and free entry from syncache bucket row.
411 * Expects locked syncache head.
412 */
413 static void
syncache_drop(struct syncache * sc,struct syncache_head * sch)414 syncache_drop(struct syncache *sc, struct syncache_head *sch)
415 {
416
417 SCH_LOCK_ASSERT(sch);
418
419 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
420 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
421 sch->sch_length--;
422
423 #ifdef TCP_OFFLOAD
424 if (ADDED_BY_TOE(sc)) {
425 struct toedev *tod = sc->sc_tod;
426
427 tod->tod_syncache_removed(tod, sc->sc_todctx);
428 }
429 #endif
430
431 syncache_free(sc);
432 }
433
434 /*
435 * Engage/reengage time on bucket row.
436 */
437 static void
syncache_timeout(struct syncache * sc,struct syncache_head * sch,int docallout)438 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
439 {
440 int rexmt;
441
442 if (sc->sc_rxmits == 0)
443 rexmt = tcp_rexmit_initial;
444 else
445 TCPT_RANGESET(rexmt,
446 tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
447 tcp_rexmit_min, tcp_rexmit_max);
448 sc->sc_rxttime = ticks + rexmt;
449 sc->sc_rxmits++;
450 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
451 sch->sch_nextc = sc->sc_rxttime;
452 if (docallout)
453 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
454 syncache_timer, (void *)sch);
455 }
456 }
457
458 /*
459 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
460 * If we have retransmitted an entry the maximum number of times, expire it.
461 * One separate timer for each bucket row.
462 */
463 static void
syncache_timer(void * xsch)464 syncache_timer(void *xsch)
465 {
466 struct syncache_head *sch = (struct syncache_head *)xsch;
467 struct syncache *sc, *nsc;
468 struct epoch_tracker et;
469 int tick = ticks;
470 char *s;
471 bool paused;
472
473 CURVNET_SET(sch->sch_sc->vnet);
474
475 /* NB: syncache_head has already been locked by the callout. */
476 SCH_LOCK_ASSERT(sch);
477
478 /*
479 * In the following cycle we may remove some entries and/or
480 * advance some timeouts, so re-initialize the bucket timer.
481 */
482 sch->sch_nextc = tick + INT_MAX;
483
484 /*
485 * If we have paused processing, unconditionally remove
486 * all syncache entries.
487 */
488 mtx_lock(&V_tcp_syncache.pause_mtx);
489 paused = V_tcp_syncache.paused;
490 mtx_unlock(&V_tcp_syncache.pause_mtx);
491
492 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
493 if (paused) {
494 syncache_drop(sc, sch);
495 continue;
496 }
497 /*
498 * We do not check if the listen socket still exists
499 * and accept the case where the listen socket may be
500 * gone by the time we resend the SYN/ACK. We do
501 * not expect this to happens often. If it does,
502 * then the RST will be sent by the time the remote
503 * host does the SYN/ACK->ACK.
504 */
505 if (TSTMP_GT(sc->sc_rxttime, tick)) {
506 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
507 sch->sch_nextc = sc->sc_rxttime;
508 continue;
509 }
510 if (sc->sc_rxmits > V_tcp_ecn_maxretries) {
511 sc->sc_flags &= ~SCF_ECN_MASK;
512 }
513 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
514 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
515 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
516 "giving up and removing syncache entry\n",
517 s, __func__);
518 free(s, M_TCPLOG);
519 }
520 syncache_drop(sc, sch);
521 TCPSTAT_INC(tcps_sc_stale);
522 continue;
523 }
524 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
525 log(LOG_DEBUG, "%s; %s: Response timeout, "
526 "retransmitting (%u) SYN|ACK\n",
527 s, __func__, sc->sc_rxmits);
528 free(s, M_TCPLOG);
529 }
530
531 NET_EPOCH_ENTER(et);
532 if (syncache_respond(sc, NULL, TH_SYN|TH_ACK) == 0) {
533 syncache_timeout(sc, sch, 0);
534 TCPSTAT_INC(tcps_sndacks);
535 TCPSTAT_INC(tcps_sndtotal);
536 TCPSTAT_INC(tcps_sc_retransmitted);
537 } else {
538 syncache_drop(sc, sch);
539 TCPSTAT_INC(tcps_sc_dropped);
540 }
541 NET_EPOCH_EXIT(et);
542 }
543 if (!TAILQ_EMPTY(&(sch)->sch_bucket))
544 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
545 syncache_timer, (void *)(sch));
546 CURVNET_RESTORE();
547 }
548
549 /*
550 * Returns true if the system is only using cookies at the moment.
551 * This could be due to a sysadmin decision to only use cookies, or it
552 * could be due to the system detecting an attack.
553 */
554 static inline bool
syncache_cookiesonly(void)555 syncache_cookiesonly(void)
556 {
557 return ((V_tcp_syncookies && V_tcp_syncache.paused) ||
558 V_tcp_syncookiesonly);
559 }
560
561 /*
562 * Find the hash bucket for the given connection.
563 */
564 static struct syncache_head *
syncache_hashbucket(struct in_conninfo * inc)565 syncache_hashbucket(struct in_conninfo *inc)
566 {
567 uint32_t hash;
568
569 /*
570 * The hash is built on foreign port + local port + foreign address.
571 * We rely on the fact that struct in_conninfo starts with 16 bits
572 * of foreign port, then 16 bits of local port then followed by 128
573 * bits of foreign address. In case of IPv4 address, the first 3
574 * 32-bit words of the address always are zeroes.
575 */
576 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
577 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
578
579 return (&V_tcp_syncache.hashbase[hash]);
580 }
581
582 /*
583 * Find an entry in the syncache.
584 * Returns always with locked syncache_head plus a matching entry or NULL.
585 */
586 static struct syncache *
syncache_lookup(struct in_conninfo * inc,struct syncache_head ** schp)587 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
588 {
589 struct syncache *sc;
590 struct syncache_head *sch;
591
592 *schp = sch = syncache_hashbucket(inc);
593 SCH_LOCK(sch);
594
595 /* Circle through bucket row to find matching entry. */
596 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
597 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
598 sizeof(struct in_endpoints)) == 0)
599 break;
600
601 return (sc); /* Always returns with locked sch. */
602 }
603
604 /*
605 * This function is called when we get a RST for a
606 * non-existent connection, so that we can see if the
607 * connection is in the syn cache. If it is, zap it.
608 * If required send a challenge ACK.
609 */
610 void
syncache_chkrst(struct in_conninfo * inc,struct tcphdr * th,struct mbuf * m,uint16_t port)611 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m,
612 uint16_t port)
613 {
614 struct syncache *sc;
615 struct syncache_head *sch;
616 char *s = NULL;
617
618 if (syncache_cookiesonly())
619 return;
620 sc = syncache_lookup(inc, &sch); /* returns locked sch */
621 SCH_LOCK_ASSERT(sch);
622
623 /*
624 * No corresponding connection was found in syncache.
625 * If syncookies are enabled and possibly exclusively
626 * used, or we are under memory pressure, a valid RST
627 * may not find a syncache entry. In that case we're
628 * done and no SYN|ACK retransmissions will happen.
629 * Otherwise the RST was misdirected or spoofed.
630 */
631 if (sc == NULL) {
632 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
633 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
634 "syncache entry (possibly syncookie only), "
635 "segment ignored\n", s, __func__);
636 TCPSTAT_INC(tcps_badrst);
637 goto done;
638 }
639
640 /* The remote UDP encaps port does not match. */
641 if (sc->sc_port != port) {
642 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
643 log(LOG_DEBUG, "%s; %s: Spurious RST with matching "
644 "syncache entry but non-matching UDP encaps port, "
645 "segment ignored\n", s, __func__);
646 TCPSTAT_INC(tcps_badrst);
647 goto done;
648 }
649
650 /*
651 * If the RST bit is set, check the sequence number to see
652 * if this is a valid reset segment.
653 *
654 * RFC 793 page 37:
655 * In all states except SYN-SENT, all reset (RST) segments
656 * are validated by checking their SEQ-fields. A reset is
657 * valid if its sequence number is in the window.
658 *
659 * RFC 793 page 69:
660 * There are four cases for the acceptability test for an incoming
661 * segment:
662 *
663 * Segment Receive Test
664 * Length Window
665 * ------- ------- -------------------------------------------
666 * 0 0 SEG.SEQ = RCV.NXT
667 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
668 * >0 0 not acceptable
669 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
670 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
671 *
672 * Note that when receiving a SYN segment in the LISTEN state,
673 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
674 * described in RFC 793, page 66.
675 */
676 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
677 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
678 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
679 if (V_tcp_insecure_rst ||
680 th->th_seq == sc->sc_irs + 1) {
681 syncache_drop(sc, sch);
682 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
683 log(LOG_DEBUG,
684 "%s; %s: Our SYN|ACK was rejected, "
685 "connection attempt aborted by remote "
686 "endpoint\n",
687 s, __func__);
688 TCPSTAT_INC(tcps_sc_reset);
689 } else {
690 TCPSTAT_INC(tcps_badrst);
691 /* Send challenge ACK. */
692 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
693 log(LOG_DEBUG, "%s; %s: RST with invalid "
694 " SEQ %u != NXT %u (+WND %u), "
695 "sending challenge ACK\n",
696 s, __func__,
697 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
698 syncache_send_challenge_ack(sc, m);
699 }
700 } else {
701 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
702 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
703 "NXT %u (+WND %u), segment ignored\n",
704 s, __func__,
705 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
706 TCPSTAT_INC(tcps_badrst);
707 }
708
709 done:
710 if (s != NULL)
711 free(s, M_TCPLOG);
712 SCH_UNLOCK(sch);
713 }
714
715 void
syncache_unreach(struct in_conninfo * inc,tcp_seq th_seq,uint16_t port)716 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port)
717 {
718 struct syncache *sc;
719 struct syncache_head *sch;
720
721 if (syncache_cookiesonly())
722 return;
723 sc = syncache_lookup(inc, &sch); /* returns locked sch */
724 SCH_LOCK_ASSERT(sch);
725 if (sc == NULL)
726 goto done;
727
728 /* If the port != sc_port, then it's a bogus ICMP msg */
729 if (port != sc->sc_port)
730 goto done;
731
732 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
733 if (ntohl(th_seq) != sc->sc_iss)
734 goto done;
735
736 /*
737 * If we've rertransmitted 3 times and this is our second error,
738 * we remove the entry. Otherwise, we allow it to continue on.
739 * This prevents us from incorrectly nuking an entry during a
740 * spurious network outage.
741 *
742 * See tcp_notify().
743 */
744 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
745 sc->sc_flags |= SCF_UNREACH;
746 goto done;
747 }
748 syncache_drop(sc, sch);
749 TCPSTAT_INC(tcps_sc_unreach);
750 done:
751 SCH_UNLOCK(sch);
752 }
753
754 /*
755 * Build a new TCP socket structure from a syncache entry.
756 *
757 * On success return the newly created socket with its underlying inp locked.
758 */
759 static struct socket *
syncache_socket(struct syncache * sc,struct socket * lso,struct mbuf * m)760 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
761 {
762 struct inpcb *inp = NULL;
763 struct socket *so;
764 struct tcpcb *tp;
765 int error;
766 char *s;
767
768 NET_EPOCH_ASSERT();
769
770 /*
771 * Ok, create the full blown connection, and set things up
772 * as they would have been set up if we had created the
773 * connection when the SYN arrived.
774 */
775 if ((so = solisten_clone(lso)) == NULL)
776 goto allocfail;
777 #ifdef MAC
778 mac_socketpeer_set_from_mbuf(m, so);
779 #endif
780 error = in_pcballoc(so, &V_tcbinfo);
781 if (error) {
782 sodealloc(so);
783 goto allocfail;
784 }
785 inp = sotoinpcb(so);
786 if ((tp = tcp_newtcpcb(inp, sototcpcb(lso))) == NULL) {
787 in_pcbfree(inp);
788 sodealloc(so);
789 goto allocfail;
790 }
791 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
792 #ifdef INET6
793 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
794 inp->inp_vflag &= ~INP_IPV4;
795 inp->inp_vflag |= INP_IPV6;
796 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
797 } else {
798 inp->inp_vflag &= ~INP_IPV6;
799 inp->inp_vflag |= INP_IPV4;
800 #endif
801 inp->inp_ip_ttl = sc->sc_ip_ttl;
802 inp->inp_ip_tos = sc->sc_ip_tos;
803 inp->inp_laddr = sc->sc_inc.inc_laddr;
804 #ifdef INET6
805 }
806 #endif
807
808 /*
809 * If there's an mbuf and it has a flowid, then let's initialise the
810 * inp with that particular flowid.
811 */
812 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
813 inp->inp_flowid = m->m_pkthdr.flowid;
814 inp->inp_flowtype = M_HASHTYPE_GET(m);
815 #ifdef NUMA
816 inp->inp_numa_domain = m->m_pkthdr.numa_domain;
817 #endif
818 }
819
820 inp->inp_lport = sc->sc_inc.inc_lport;
821 #ifdef INET6
822 if (inp->inp_vflag & INP_IPV6PROTO) {
823 struct inpcb *oinp = sotoinpcb(lso);
824
825 /*
826 * Inherit socket options from the listening socket.
827 * Note that in6p_inputopts are not (and should not be)
828 * copied, since it stores previously received options and is
829 * used to detect if each new option is different than the
830 * previous one and hence should be passed to a user.
831 * If we copied in6p_inputopts, a user would not be able to
832 * receive options just after calling the accept system call.
833 */
834 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
835 if (oinp->in6p_outputopts)
836 inp->in6p_outputopts =
837 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
838 inp->in6p_hops = oinp->in6p_hops;
839 }
840
841 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
842 struct sockaddr_in6 sin6;
843
844 sin6.sin6_family = AF_INET6;
845 sin6.sin6_len = sizeof(sin6);
846 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
847 sin6.sin6_port = sc->sc_inc.inc_fport;
848 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
849 INP_HASH_WLOCK(&V_tcbinfo);
850 error = in6_pcbconnect(inp, &sin6, thread0.td_ucred, false);
851 INP_HASH_WUNLOCK(&V_tcbinfo);
852 if (error != 0)
853 goto abort;
854 /* Override flowlabel from in6_pcbconnect. */
855 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
856 inp->inp_flow |= sc->sc_flowlabel;
857 }
858 #endif /* INET6 */
859 #if defined(INET) && defined(INET6)
860 else
861 #endif
862 #ifdef INET
863 {
864 struct sockaddr_in sin;
865
866 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
867
868 if (inp->inp_options == NULL) {
869 inp->inp_options = sc->sc_ipopts;
870 sc->sc_ipopts = NULL;
871 }
872
873 sin.sin_family = AF_INET;
874 sin.sin_len = sizeof(sin);
875 sin.sin_addr = sc->sc_inc.inc_faddr;
876 sin.sin_port = sc->sc_inc.inc_fport;
877 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
878 INP_HASH_WLOCK(&V_tcbinfo);
879 error = in_pcbconnect(inp, &sin, thread0.td_ucred);
880 INP_HASH_WUNLOCK(&V_tcbinfo);
881 if (error != 0)
882 goto abort;
883 }
884 #endif /* INET */
885 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
886 /* Copy old policy into new socket's. */
887 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
888 printf("syncache_socket: could not copy policy\n");
889 #endif
890 tp->t_state = TCPS_SYN_RECEIVED;
891 tp->iss = sc->sc_iss;
892 tp->irs = sc->sc_irs;
893 tp->t_port = sc->sc_port;
894 tcp_rcvseqinit(tp);
895 tcp_sendseqinit(tp);
896 tp->snd_wl1 = sc->sc_irs;
897 tp->snd_max = tp->iss + 1;
898 tp->snd_nxt = tp->iss + 1;
899 tp->rcv_up = sc->sc_irs + 1;
900 tp->rcv_wnd = sc->sc_wnd;
901 tp->rcv_adv += tp->rcv_wnd;
902 tp->last_ack_sent = tp->rcv_nxt;
903
904 tp->t_flags = sototcpcb(lso)->t_flags &
905 (TF_LRD|TF_NOPUSH|TF_NODELAY);
906 if (sc->sc_flags & SCF_NOOPT)
907 tp->t_flags |= TF_NOOPT;
908 else {
909 if (sc->sc_flags & SCF_WINSCALE) {
910 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
911 tp->snd_scale = sc->sc_requested_s_scale;
912 tp->request_r_scale = sc->sc_requested_r_scale;
913 }
914 if (sc->sc_flags & SCF_TIMESTAMP) {
915 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
916 tp->ts_recent = sc->sc_tsreflect;
917 tp->ts_recent_age = tcp_ts_getticks();
918 tp->ts_offset = sc->sc_tsoff;
919 }
920 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
921 if (sc->sc_flags & SCF_SIGNATURE)
922 tp->t_flags |= TF_SIGNATURE;
923 #endif
924 if (sc->sc_flags & SCF_SACK)
925 tp->t_flags |= TF_SACK_PERMIT;
926 }
927
928 tcp_ecn_syncache_socket(tp, sc);
929
930 /*
931 * Set up MSS and get cached values from tcp_hostcache.
932 * This might overwrite some of the defaults we just set.
933 */
934 tcp_mss(tp, sc->sc_peer_mss);
935
936 /*
937 * If the SYN,ACK was retransmitted, indicate that CWND to be
938 * limited to one segment in cc_conn_init().
939 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
940 */
941 if (sc->sc_rxmits > 1)
942 tp->snd_cwnd = 1;
943
944 /* Copy over the challenge ACK state. */
945 tp->t_challenge_ack_end = sc->sc_challenge_ack_end;
946 tp->t_challenge_ack_cnt = sc->sc_challenge_ack_cnt;
947
948 #ifdef TCP_OFFLOAD
949 /*
950 * Allow a TOE driver to install its hooks. Note that we hold the
951 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
952 * new connection before the TOE driver has done its thing.
953 */
954 if (ADDED_BY_TOE(sc)) {
955 struct toedev *tod = sc->sc_tod;
956
957 tod->tod_offload_socket(tod, sc->sc_todctx, so);
958 }
959 #endif
960 #ifdef TCP_BLACKBOX
961 /*
962 * Inherit the log state from the listening socket, if
963 * - the log state of the listening socket is not off and
964 * - the listening socket was not auto selected from all sessions and
965 * - a log id is not set on the listening socket.
966 * This avoids inheriting a log state which was automatically set.
967 */
968 if ((tcp_get_bblog_state(sototcpcb(lso)) != TCP_LOG_STATE_OFF) &&
969 ((sototcpcb(lso)->t_flags2 & TF2_LOG_AUTO) == 0) &&
970 (sototcpcb(lso)->t_lib == NULL)) {
971 tcp_log_state_change(tp, tcp_get_bblog_state(sototcpcb(lso)));
972 }
973 #endif
974 /*
975 * Copy and activate timers.
976 */
977 tp->t_maxunacktime = sototcpcb(lso)->t_maxunacktime;
978 tp->t_keepinit = sototcpcb(lso)->t_keepinit;
979 tp->t_keepidle = sototcpcb(lso)->t_keepidle;
980 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
981 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
982 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
983
984 TCPSTAT_INC(tcps_accepts);
985 TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, TCPS_LISTEN);
986
987 if (!solisten_enqueue(so, SS_ISCONNECTED))
988 tp->t_flags |= TF_SONOTCONN;
989 /* Can we inherit anything from the listener? */
990 if (tp->t_fb->tfb_inherit != NULL) {
991 (*tp->t_fb->tfb_inherit)(tp, sotoinpcb(lso));
992 }
993 return (so);
994
995 allocfail:
996 /*
997 * Drop the connection; we will either send a RST or have the peer
998 * retransmit its SYN again after its RTO and try again.
999 */
1000 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1001 log(LOG_DEBUG, "%s; %s: Socket create failed "
1002 "due to limits or memory shortage\n",
1003 s, __func__);
1004 free(s, M_TCPLOG);
1005 }
1006 TCPSTAT_INC(tcps_listendrop);
1007 return (NULL);
1008
1009 abort:
1010 tcp_discardcb(tp);
1011 in_pcbfree(inp);
1012 sodealloc(so);
1013 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1014 log(LOG_DEBUG, "%s; %s: in%s_pcbconnect failed with error %i\n",
1015 s, __func__, (sc->sc_inc.inc_flags & INC_ISIPV6) ? "6" : "",
1016 error);
1017 free(s, M_TCPLOG);
1018 }
1019 TCPSTAT_INC(tcps_listendrop);
1020 return (NULL);
1021 }
1022
1023 /*
1024 * This function gets called when we receive an ACK for a
1025 * socket in the LISTEN state. We look up the connection
1026 * in the syncache, and if its there, we pull it out of
1027 * the cache and turn it into a full-blown connection in
1028 * the SYN-RECEIVED state.
1029 *
1030 * On syncache_socket() success the newly created socket
1031 * has its underlying inp locked.
1032 *
1033 * *lsop is updated, if and only if 1 is returned.
1034 */
1035 int
syncache_expand(struct in_conninfo * inc,struct tcpopt * to,struct tcphdr * th,struct socket ** lsop,struct mbuf * m,uint16_t port)1036 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1037 struct socket **lsop, struct mbuf *m, uint16_t port)
1038 {
1039 struct syncache *sc;
1040 struct syncache_head *sch;
1041 struct syncache scs;
1042 char *s;
1043 bool locked;
1044
1045 NET_EPOCH_ASSERT();
1046 KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1047 ("%s: can handle only ACK", __func__));
1048
1049 if (syncache_cookiesonly()) {
1050 sc = NULL;
1051 sch = syncache_hashbucket(inc);
1052 locked = false;
1053 } else {
1054 sc = syncache_lookup(inc, &sch); /* returns locked sch */
1055 locked = true;
1056 SCH_LOCK_ASSERT(sch);
1057 }
1058
1059 #ifdef INVARIANTS
1060 /*
1061 * Test code for syncookies comparing the syncache stored
1062 * values with the reconstructed values from the cookie.
1063 */
1064 if (sc != NULL)
1065 syncookie_cmp(inc, sch, sc, th, to, *lsop, port);
1066 #endif
1067
1068 if (sc == NULL) {
1069 if (locked) {
1070 /*
1071 * The syncache is currently in use (neither disabled,
1072 * nor paused), but no entry was found.
1073 */
1074 if (!V_tcp_syncookies) {
1075 /*
1076 * Since no syncookies are used in case of
1077 * a bucket overflow, don't even check for
1078 * a valid syncookie.
1079 */
1080 SCH_UNLOCK(sch);
1081 TCPSTAT_INC(tcps_sc_spurcookie);
1082 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1083 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1084 "segment rejected "
1085 "(syncookies disabled)\n",
1086 s, __func__);
1087 free(s, M_TCPLOG);
1088 }
1089 return (0);
1090 }
1091 if (sch->sch_last_overflow <
1092 time_uptime - SYNCOOKIE_LIFETIME) {
1093 /*
1094 * Since the bucket did not overflow recently,
1095 * don't even check for a valid syncookie.
1096 */
1097 SCH_UNLOCK(sch);
1098 TCPSTAT_INC(tcps_sc_spurcookie);
1099 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1100 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1101 "segment rejected "
1102 "(no syncache entry)\n",
1103 s, __func__);
1104 free(s, M_TCPLOG);
1105 }
1106 return (0);
1107 }
1108 SCH_UNLOCK(sch);
1109 }
1110 bzero(&scs, sizeof(scs));
1111 /*
1112 * Now check, if the syncookie is valid. If it is, create an on
1113 * stack syncache entry.
1114 */
1115 if (syncookie_expand(inc, sch, &scs, th, to, *lsop, port)) {
1116 sc = &scs;
1117 TCPSTAT_INC(tcps_sc_recvcookie);
1118 } else {
1119 TCPSTAT_INC(tcps_sc_failcookie);
1120 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1121 log(LOG_DEBUG, "%s; %s: Segment failed "
1122 "SYNCOOKIE authentication, segment rejected "
1123 "(probably spoofed)\n", s, __func__);
1124 free(s, M_TCPLOG);
1125 }
1126 return (0);
1127 }
1128 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1129 /* If received ACK has MD5 signature, check it. */
1130 if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1131 (!TCPMD5_ENABLED() ||
1132 TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1133 /* Drop the ACK. */
1134 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1135 log(LOG_DEBUG, "%s; %s: Segment rejected, "
1136 "MD5 signature doesn't match.\n",
1137 s, __func__);
1138 free(s, M_TCPLOG);
1139 }
1140 TCPSTAT_INC(tcps_sig_err_sigopt);
1141 return (-1); /* Do not send RST */
1142 }
1143 #endif /* TCP_SIGNATURE */
1144 TCPSTATES_INC(TCPS_SYN_RECEIVED);
1145 } else {
1146 if (sc->sc_port != port) {
1147 SCH_UNLOCK(sch);
1148 return (0);
1149 }
1150 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1151 /*
1152 * If listening socket requested TCP digests, check that
1153 * received ACK has signature and it is correct.
1154 * If not, drop the ACK and leave sc entry in the cache,
1155 * because SYN was received with correct signature.
1156 */
1157 if (sc->sc_flags & SCF_SIGNATURE) {
1158 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1159 /* No signature */
1160 TCPSTAT_INC(tcps_sig_err_nosigopt);
1161 SCH_UNLOCK(sch);
1162 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1163 log(LOG_DEBUG, "%s; %s: Segment "
1164 "rejected, MD5 signature wasn't "
1165 "provided.\n", s, __func__);
1166 free(s, M_TCPLOG);
1167 }
1168 return (-1); /* Do not send RST */
1169 }
1170 if (!TCPMD5_ENABLED() ||
1171 TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1172 /* Doesn't match or no SA */
1173 SCH_UNLOCK(sch);
1174 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1175 log(LOG_DEBUG, "%s; %s: Segment "
1176 "rejected, MD5 signature doesn't "
1177 "match.\n", s, __func__);
1178 free(s, M_TCPLOG);
1179 }
1180 return (-1); /* Do not send RST */
1181 }
1182 }
1183 #endif /* TCP_SIGNATURE */
1184
1185 /*
1186 * RFC 7323 PAWS: If we have a timestamp on this segment and
1187 * it's less than ts_recent, drop it.
1188 * XXXMT: RFC 7323 also requires to send an ACK.
1189 * In tcp_input.c this is only done for TCP segments
1190 * with user data, so be consistent here and just drop
1191 * the segment.
1192 */
1193 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1194 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1195 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1196 log(LOG_DEBUG,
1197 "%s; %s: SEG.TSval %u < TS.Recent %u, "
1198 "segment dropped\n", s, __func__,
1199 to->to_tsval, sc->sc_tsreflect);
1200 }
1201 SCH_UNLOCK(sch);
1202 free(s, M_TCPLOG);
1203 return (-1); /* Do not send RST */
1204 }
1205
1206 /*
1207 * If timestamps were not negotiated during SYN/ACK and a
1208 * segment with a timestamp is received, ignore the
1209 * timestamp and process the packet normally.
1210 * See section 3.2 of RFC 7323.
1211 */
1212 if (!(sc->sc_flags & SCF_TIMESTAMP) &&
1213 (to->to_flags & TOF_TS)) {
1214 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1215 log(LOG_DEBUG, "%s; %s: Timestamp not "
1216 "expected, segment processed normally\n",
1217 s, __func__);
1218 free(s, M_TCPLOG);
1219 }
1220 }
1221
1222 /*
1223 * If timestamps were negotiated during SYN/ACK and a
1224 * segment without a timestamp is received, silently drop
1225 * the segment, unless the missing timestamps are tolerated.
1226 * See section 3.2 of RFC 7323.
1227 */
1228 if ((sc->sc_flags & SCF_TIMESTAMP) &&
1229 !(to->to_flags & TOF_TS)) {
1230 if (V_tcp_tolerate_missing_ts) {
1231 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1232 log(LOG_DEBUG,
1233 "%s; %s: Timestamp missing, "
1234 "segment processed normally\n",
1235 s, __func__);
1236 free(s, M_TCPLOG);
1237 }
1238 } else {
1239 SCH_UNLOCK(sch);
1240 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1241 log(LOG_DEBUG,
1242 "%s; %s: Timestamp missing, "
1243 "segment silently dropped\n",
1244 s, __func__);
1245 free(s, M_TCPLOG);
1246 }
1247 return (-1); /* Do not send RST */
1248 }
1249 }
1250
1251 /*
1252 * SEG.SEQ validation:
1253 * The SEG.SEQ must be in the window starting at our
1254 * initial receive sequence number + 1.
1255 */
1256 if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1257 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1258 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1259 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, "
1260 "sending challenge ACK\n",
1261 s, __func__, th->th_seq, sc->sc_irs + 1);
1262 syncache_send_challenge_ack(sc, m);
1263 SCH_UNLOCK(sch);
1264 free(s, M_TCPLOG);
1265 return (-1); /* Do not send RST */
1266 }
1267
1268 /*
1269 * SEG.ACK validation:
1270 * SEG.ACK must match our initial send sequence number + 1.
1271 */
1272 if (th->th_ack != sc->sc_iss + 1) {
1273 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1274 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, "
1275 "segment rejected\n",
1276 s, __func__, th->th_ack, sc->sc_iss + 1);
1277 SCH_UNLOCK(sch);
1278 free(s, M_TCPLOG);
1279 return (0); /* Do send RST, do not free sc. */
1280 }
1281
1282 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1283 sch->sch_length--;
1284 #ifdef TCP_OFFLOAD
1285 if (ADDED_BY_TOE(sc)) {
1286 struct toedev *tod = sc->sc_tod;
1287
1288 tod->tod_syncache_removed(tod, sc->sc_todctx);
1289 }
1290 #endif
1291 SCH_UNLOCK(sch);
1292 }
1293
1294 *lsop = syncache_socket(sc, *lsop, m);
1295
1296 if (__predict_false(*lsop == NULL)) {
1297 TCPSTAT_INC(tcps_sc_aborted);
1298 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1299 } else if (sc != &scs)
1300 TCPSTAT_INC(tcps_sc_completed);
1301
1302 if (sc != &scs)
1303 syncache_free(sc);
1304 return (1);
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 ucred *cred;
1367 struct syncache_head *sch;
1368 struct mbuf *ipopts = NULL;
1369 u_int ltflags;
1370 int win, ip_ttl, ip_tos;
1371 char *s;
1372 #ifdef INET6
1373 int autoflowlabel = 0;
1374 #endif
1375 #ifdef MAC
1376 struct label *maclabel = NULL;
1377 #endif
1378 struct syncache scs;
1379 uint64_t tfo_response_cookie;
1380 unsigned int *tfo_pending = NULL;
1381 int tfo_cookie_valid = 0;
1382 int tfo_response_cookie_valid = 0;
1383 bool locked;
1384
1385 INP_RLOCK_ASSERT(inp); /* listen socket */
1386 KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1387 ("%s: unexpected tcp flags", __func__));
1388
1389 /*
1390 * Combine all so/tp operations very early to drop the INP lock as
1391 * soon as possible.
1392 */
1393 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1394 tp = sototcpcb(so);
1395 cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred);
1396
1397 #ifdef INET6
1398 if (inc->inc_flags & INC_ISIPV6) {
1399 if (inp->inp_flags & IN6P_AUTOFLOWLABEL) {
1400 autoflowlabel = 1;
1401 }
1402 ip_ttl = in6_selecthlim(inp, NULL);
1403 if ((inp->in6p_outputopts == NULL) ||
1404 (inp->in6p_outputopts->ip6po_tclass == -1)) {
1405 ip_tos = 0;
1406 } else {
1407 ip_tos = inp->in6p_outputopts->ip6po_tclass;
1408 }
1409 }
1410 #endif
1411 #if defined(INET6) && defined(INET)
1412 else
1413 #endif
1414 #ifdef INET
1415 {
1416 ip_ttl = inp->inp_ip_ttl;
1417 ip_tos = inp->inp_ip_tos;
1418 }
1419 #endif
1420 win = so->sol_sbrcv_hiwat;
1421 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1422
1423 if (V_tcp_fastopen_server_enable && (tp->t_flags & TF_FASTOPEN) &&
1424 (tp->t_tfo_pending != NULL) &&
1425 (to->to_flags & TOF_FASTOPEN)) {
1426 /*
1427 * Limit the number of pending TFO connections to
1428 * approximately half of the queue limit. This prevents TFO
1429 * SYN floods from starving the service by filling the
1430 * listen queue with bogus TFO connections.
1431 */
1432 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1433 (so->sol_qlimit / 2)) {
1434 int result;
1435
1436 result = tcp_fastopen_check_cookie(inc,
1437 to->to_tfo_cookie, to->to_tfo_len,
1438 &tfo_response_cookie);
1439 tfo_cookie_valid = (result > 0);
1440 tfo_response_cookie_valid = (result >= 0);
1441 }
1442
1443 /*
1444 * Remember the TFO pending counter as it will have to be
1445 * decremented below if we don't make it to syncache_tfo_expand().
1446 */
1447 tfo_pending = tp->t_tfo_pending;
1448 }
1449
1450 #ifdef MAC
1451 if (mac_syncache_init(&maclabel) != 0) {
1452 INP_RUNLOCK(inp);
1453 goto done;
1454 } else
1455 mac_syncache_create(maclabel, inp);
1456 #endif
1457 if (!tfo_cookie_valid)
1458 INP_RUNLOCK(inp);
1459
1460 /*
1461 * Remember the IP options, if any.
1462 */
1463 #ifdef INET6
1464 if (!(inc->inc_flags & INC_ISIPV6))
1465 #endif
1466 #ifdef INET
1467 ipopts = (m) ? ip_srcroute(m) : NULL;
1468 #else
1469 ipopts = NULL;
1470 #endif
1471
1472 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1473 /*
1474 * When the socket is TCP-MD5 enabled check that,
1475 * - a signed packet is valid
1476 * - a non-signed packet does not have a security association
1477 *
1478 * If a signed packet fails validation or a non-signed packet has a
1479 * security association, the packet will be dropped.
1480 */
1481 if (ltflags & TF_SIGNATURE) {
1482 if (to->to_flags & TOF_SIGNATURE) {
1483 if (!TCPMD5_ENABLED() ||
1484 TCPMD5_INPUT(m, th, to->to_signature) != 0)
1485 goto done;
1486 } else {
1487 if (TCPMD5_ENABLED() &&
1488 TCPMD5_INPUT(m, NULL, NULL) != ENOENT)
1489 goto done;
1490 }
1491 } else if (to->to_flags & TOF_SIGNATURE)
1492 goto done;
1493 #endif /* TCP_SIGNATURE */
1494 /*
1495 * See if we already have an entry for this connection.
1496 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1497 *
1498 * XXX: should the syncache be re-initialized with the contents
1499 * of the new SYN here (which may have different options?)
1500 *
1501 * XXX: We do not check the sequence number to see if this is a
1502 * real retransmit or a new connection attempt. The question is
1503 * how to handle such a case; either ignore it as spoofed, or
1504 * drop the current entry and create a new one?
1505 */
1506 if (syncache_cookiesonly()) {
1507 sc = NULL;
1508 sch = syncache_hashbucket(inc);
1509 locked = false;
1510 } else {
1511 sc = syncache_lookup(inc, &sch); /* returns locked sch */
1512 locked = true;
1513 SCH_LOCK_ASSERT(sch);
1514 }
1515 if (sc != NULL) {
1516 if (tfo_cookie_valid)
1517 INP_RUNLOCK(inp);
1518 TCPSTAT_INC(tcps_sc_dupsyn);
1519 if (ipopts) {
1520 /*
1521 * If we were remembering a previous source route,
1522 * forget it and use the new one we've been given.
1523 */
1524 if (sc->sc_ipopts)
1525 (void)m_free(sc->sc_ipopts);
1526 sc->sc_ipopts = ipopts;
1527 }
1528 /*
1529 * Update timestamp if present.
1530 */
1531 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1532 sc->sc_tsreflect = to->to_tsval;
1533 else
1534 sc->sc_flags &= ~SCF_TIMESTAMP;
1535 /*
1536 * Adjust ECN response if needed, e.g. different
1537 * IP ECN field, or a fallback by the remote host.
1538 */
1539 if (sc->sc_flags & SCF_ECN_MASK) {
1540 sc->sc_flags &= ~SCF_ECN_MASK;
1541 sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1542 }
1543 #ifdef MAC
1544 /*
1545 * Since we have already unconditionally allocated label
1546 * storage, free it up. The syncache entry will already
1547 * have an initialized label we can use.
1548 */
1549 mac_syncache_destroy(&maclabel);
1550 #endif
1551 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1552 /* Retransmit SYN|ACK and reset retransmit count. */
1553 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1554 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1555 "resetting timer and retransmitting SYN|ACK\n",
1556 s, __func__);
1557 free(s, M_TCPLOG);
1558 }
1559 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1560 sc->sc_rxmits = 0;
1561 syncache_timeout(sc, sch, 1);
1562 TCPSTAT_INC(tcps_sndacks);
1563 TCPSTAT_INC(tcps_sndtotal);
1564 } else {
1565 syncache_drop(sc, sch);
1566 TCPSTAT_INC(tcps_sc_dropped);
1567 }
1568 SCH_UNLOCK(sch);
1569 goto donenoprobe;
1570 }
1571
1572 KASSERT(sc == NULL, ("sc(%p) != NULL", sc));
1573 /*
1574 * Skip allocating a syncache entry if we are just going to discard
1575 * it later.
1576 */
1577 if (!locked || tfo_cookie_valid) {
1578 bzero(&scs, sizeof(scs));
1579 sc = &scs;
1580 } else {
1581 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1582 if (sc == NULL) {
1583 /*
1584 * The zone allocator couldn't provide more entries.
1585 * Treat this as if the cache was full; drop the oldest
1586 * entry and insert the new one.
1587 */
1588 TCPSTAT_INC(tcps_sc_zonefail);
1589 sc = TAILQ_LAST(&sch->sch_bucket, sch_head);
1590 if (sc != NULL) {
1591 sch->sch_last_overflow = time_uptime;
1592 syncache_drop(sc, sch);
1593 syncache_pause(inc);
1594 }
1595 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1596 if (sc == NULL) {
1597 if (V_tcp_syncookies) {
1598 bzero(&scs, sizeof(scs));
1599 sc = &scs;
1600 } else {
1601 KASSERT(locked,
1602 ("%s: bucket unexpectedly unlocked",
1603 __func__));
1604 SCH_UNLOCK(sch);
1605 goto done;
1606 }
1607 }
1608 }
1609 }
1610
1611 KASSERT(sc != NULL, ("sc == NULL"));
1612 if (!tfo_cookie_valid && tfo_response_cookie_valid)
1613 sc->sc_tfo_cookie = &tfo_response_cookie;
1614
1615 /*
1616 * Fill in the syncache values.
1617 */
1618 #ifdef MAC
1619 sc->sc_label = maclabel;
1620 #endif
1621 /*
1622 * sc_cred is only used in syncache_pcblist() to list TCP endpoints in
1623 * TCPS_SYN_RECEIVED state when V_tcp_syncache.see_other is false.
1624 * Therefore, store the credentials only when needed:
1625 * - sc is allocated from the zone and not using the on stack instance.
1626 * - the sysctl variable net.inet.tcp.syncache.see_other is false.
1627 * The reference count is decremented when a zone allocated sc is
1628 * freed in syncache_free().
1629 */
1630 if (sc != &scs && !V_tcp_syncache.see_other) {
1631 sc->sc_cred = cred;
1632 cred = NULL;
1633 } else
1634 sc->sc_cred = NULL;
1635 sc->sc_port = port;
1636 sc->sc_ipopts = ipopts;
1637 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1638 sc->sc_ip_tos = ip_tos;
1639 sc->sc_ip_ttl = ip_ttl;
1640 #ifdef TCP_OFFLOAD
1641 sc->sc_tod = tod;
1642 sc->sc_todctx = todctx;
1643 #endif
1644 sc->sc_irs = th->th_seq;
1645 sc->sc_flags = 0;
1646 sc->sc_flowlabel = 0;
1647
1648 /*
1649 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1650 * win was derived from socket earlier in the function.
1651 */
1652 win = imax(win, 0);
1653 win = imin(win, TCP_MAXWIN);
1654 sc->sc_wnd = win;
1655
1656 if (V_tcp_do_rfc1323 &&
1657 !(ltflags & TF_NOOPT)) {
1658 /*
1659 * A timestamp received in a SYN makes
1660 * it ok to send timestamp requests and replies.
1661 */
1662 if ((to->to_flags & TOF_TS) && (V_tcp_do_rfc1323 != 2)) {
1663 sc->sc_tsreflect = to->to_tsval;
1664 sc->sc_flags |= SCF_TIMESTAMP;
1665 sc->sc_tsoff = tcp_new_ts_offset(inc);
1666 }
1667 if ((to->to_flags & TOF_SCALE) && (V_tcp_do_rfc1323 != 3)) {
1668 u_int wscale = 0;
1669
1670 /*
1671 * Pick the smallest possible scaling factor that
1672 * will still allow us to scale up to sb_max, aka
1673 * kern.ipc.maxsockbuf.
1674 *
1675 * We do this because there are broken firewalls that
1676 * will corrupt the window scale option, leading to
1677 * the other endpoint believing that our advertised
1678 * window is unscaled. At scale factors larger than
1679 * 5 the unscaled window will drop below 1500 bytes,
1680 * leading to serious problems when traversing these
1681 * broken firewalls.
1682 *
1683 * With the default maxsockbuf of 256K, a scale factor
1684 * of 3 will be chosen by this algorithm. Those who
1685 * choose a larger maxsockbuf should watch out
1686 * for the compatibility problems mentioned above.
1687 *
1688 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1689 * or <SYN,ACK>) segment itself is never scaled.
1690 */
1691 while (wscale < TCP_MAX_WINSHIFT &&
1692 (TCP_MAXWIN << wscale) < sb_max)
1693 wscale++;
1694 sc->sc_requested_r_scale = wscale;
1695 sc->sc_requested_s_scale = to->to_wscale;
1696 sc->sc_flags |= SCF_WINSCALE;
1697 }
1698 }
1699 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1700 /*
1701 * If incoming packet has an MD5 signature, flag this in the
1702 * syncache so that syncache_respond() will do the right thing
1703 * with the SYN+ACK.
1704 */
1705 if (to->to_flags & TOF_SIGNATURE)
1706 sc->sc_flags |= SCF_SIGNATURE;
1707 #endif /* TCP_SIGNATURE */
1708 if (to->to_flags & TOF_SACKPERM)
1709 sc->sc_flags |= SCF_SACK;
1710 if (to->to_flags & TOF_MSS)
1711 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */
1712 if (ltflags & TF_NOOPT)
1713 sc->sc_flags |= SCF_NOOPT;
1714 /* ECN Handshake */
1715 if (V_tcp_do_ecn && (tp->t_flags2 & TF2_CANNOT_DO_ECN) == 0)
1716 sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1717
1718 if (V_tcp_syncookies || V_tcp_syncookiesonly)
1719 sc->sc_iss = syncookie_generate(sch, sc);
1720 else
1721 sc->sc_iss = arc4random();
1722 #ifdef INET6
1723 if (autoflowlabel) {
1724 if (V_tcp_syncookies || V_tcp_syncookiesonly)
1725 sc->sc_flowlabel = sc->sc_iss;
1726 else
1727 sc->sc_flowlabel = ip6_randomflowlabel();
1728 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1729 }
1730 #endif
1731 if (locked)
1732 SCH_UNLOCK(sch);
1733
1734 if (tfo_cookie_valid) {
1735 rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie);
1736 /* INP_RUNLOCK(inp) will be performed by the caller */
1737 goto tfo_expanded;
1738 }
1739
1740 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1741 /*
1742 * Do a standard 3-way handshake.
1743 */
1744 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) {
1745 if (sc != &scs)
1746 syncache_insert(sc, sch); /* locks and unlocks sch */
1747 TCPSTAT_INC(tcps_sndacks);
1748 TCPSTAT_INC(tcps_sndtotal);
1749 } else {
1750 if (sc != &scs)
1751 syncache_free(sc);
1752 TCPSTAT_INC(tcps_sc_dropped);
1753 }
1754 goto donenoprobe;
1755
1756 done:
1757 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1758 donenoprobe:
1759 if (m)
1760 m_freem(m);
1761 /*
1762 * If tfo_pending is not NULL here, then a TFO SYN that did not
1763 * result in a new socket was processed and the associated pending
1764 * counter has not yet been decremented. All such TFO processing paths
1765 * transit this point.
1766 */
1767 if (tfo_pending != NULL)
1768 tcp_fastopen_decrement_counter(tfo_pending);
1769
1770 tfo_expanded:
1771 if (cred != NULL)
1772 crfree(cred);
1773 if (sc == NULL || sc == &scs) {
1774 #ifdef MAC
1775 mac_syncache_destroy(&maclabel);
1776 #endif
1777 if (ipopts)
1778 (void)m_free(ipopts);
1779 }
1780 return (rv);
1781 }
1782
1783 /*
1784 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment,
1785 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1786 */
1787 static int
syncache_respond(struct syncache * sc,const struct mbuf * m0,int flags)1788 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags)
1789 {
1790 struct ip *ip = NULL;
1791 struct mbuf *m;
1792 struct tcphdr *th = NULL;
1793 struct udphdr *udp = NULL;
1794 int optlen, error = 0; /* Make compiler happy */
1795 u_int16_t hlen, tlen, mssopt, ulen;
1796 struct tcpopt to;
1797 #ifdef INET6
1798 struct ip6_hdr *ip6 = NULL;
1799 #endif
1800
1801 NET_EPOCH_ASSERT();
1802
1803 hlen =
1804 #ifdef INET6
1805 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1806 #endif
1807 sizeof(struct ip);
1808 tlen = hlen + sizeof(struct tcphdr);
1809 if (sc->sc_port) {
1810 tlen += sizeof(struct udphdr);
1811 }
1812 /* Determine MSS we advertize to other end of connection. */
1813 mssopt = tcp_mssopt(&sc->sc_inc);
1814 if (sc->sc_port)
1815 mssopt -= V_tcp_udp_tunneling_overhead;
1816 mssopt = max(mssopt, V_tcp_minmss);
1817
1818 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1819 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1820 ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + "
1821 "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port,
1822 max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN));
1823
1824 /* Create the IP+TCP header from scratch. */
1825 m = m_gethdr(M_NOWAIT, MT_DATA);
1826 if (m == NULL)
1827 return (ENOBUFS);
1828 #ifdef MAC
1829 mac_syncache_create_mbuf(sc->sc_label, m);
1830 #endif
1831 m->m_data += max_linkhdr;
1832 m->m_len = tlen;
1833 m->m_pkthdr.len = tlen;
1834 m->m_pkthdr.rcvif = NULL;
1835
1836 #ifdef INET6
1837 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1838 ip6 = mtod(m, struct ip6_hdr *);
1839 ip6->ip6_vfc = IPV6_VERSION;
1840 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1841 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1842 ip6->ip6_plen = htons(tlen - hlen);
1843 /* ip6_hlim is set after checksum */
1844 /* Zero out traffic class and flow label. */
1845 ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1846 ip6->ip6_flow |= sc->sc_flowlabel;
1847 if (sc->sc_port != 0) {
1848 ip6->ip6_nxt = IPPROTO_UDP;
1849 udp = (struct udphdr *)(ip6 + 1);
1850 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1851 udp->uh_dport = sc->sc_port;
1852 ulen = (tlen - sizeof(struct ip6_hdr));
1853 th = (struct tcphdr *)(udp + 1);
1854 } else {
1855 ip6->ip6_nxt = IPPROTO_TCP;
1856 th = (struct tcphdr *)(ip6 + 1);
1857 }
1858 ip6->ip6_flow |= htonl(sc->sc_ip_tos << IPV6_FLOWLABEL_LEN);
1859 }
1860 #endif
1861 #if defined(INET6) && defined(INET)
1862 else
1863 #endif
1864 #ifdef INET
1865 {
1866 ip = mtod(m, struct ip *);
1867 ip->ip_v = IPVERSION;
1868 ip->ip_hl = sizeof(struct ip) >> 2;
1869 ip->ip_len = htons(tlen);
1870 ip->ip_id = 0;
1871 ip->ip_off = 0;
1872 ip->ip_sum = 0;
1873 ip->ip_src = sc->sc_inc.inc_laddr;
1874 ip->ip_dst = sc->sc_inc.inc_faddr;
1875 ip->ip_ttl = sc->sc_ip_ttl;
1876 ip->ip_tos = sc->sc_ip_tos;
1877
1878 /*
1879 * See if we should do MTU discovery. Route lookups are
1880 * expensive, so we will only unset the DF bit if:
1881 *
1882 * 1) path_mtu_discovery is disabled
1883 * 2) the SCF_UNREACH flag has been set
1884 */
1885 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1886 ip->ip_off |= htons(IP_DF);
1887 if (sc->sc_port == 0) {
1888 ip->ip_p = IPPROTO_TCP;
1889 th = (struct tcphdr *)(ip + 1);
1890 } else {
1891 ip->ip_p = IPPROTO_UDP;
1892 udp = (struct udphdr *)(ip + 1);
1893 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1894 udp->uh_dport = sc->sc_port;
1895 ulen = (tlen - sizeof(struct ip));
1896 th = (struct tcphdr *)(udp + 1);
1897 }
1898 }
1899 #endif /* INET */
1900 th->th_sport = sc->sc_inc.inc_lport;
1901 th->th_dport = sc->sc_inc.inc_fport;
1902
1903 if (flags & TH_SYN)
1904 th->th_seq = htonl(sc->sc_iss);
1905 else
1906 th->th_seq = htonl(sc->sc_iss + 1);
1907 th->th_ack = htonl(sc->sc_irs + 1);
1908 th->th_off = sizeof(struct tcphdr) >> 2;
1909 th->th_win = htons(sc->sc_wnd);
1910 th->th_urp = 0;
1911
1912 flags = tcp_ecn_syncache_respond(flags, sc);
1913 tcp_set_flags(th, flags);
1914
1915 /* Tack on the TCP options. */
1916 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1917 to.to_flags = 0;
1918
1919 if (flags & TH_SYN) {
1920 to.to_mss = mssopt;
1921 to.to_flags = TOF_MSS;
1922 if (sc->sc_flags & SCF_WINSCALE) {
1923 to.to_wscale = sc->sc_requested_r_scale;
1924 to.to_flags |= TOF_SCALE;
1925 }
1926 if (sc->sc_flags & SCF_SACK)
1927 to.to_flags |= TOF_SACKPERM;
1928 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1929 if (sc->sc_flags & SCF_SIGNATURE)
1930 to.to_flags |= TOF_SIGNATURE;
1931 #endif
1932 if (sc->sc_tfo_cookie) {
1933 to.to_flags |= TOF_FASTOPEN;
1934 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1935 to.to_tfo_cookie = sc->sc_tfo_cookie;
1936 /* don't send cookie again when retransmitting response */
1937 sc->sc_tfo_cookie = NULL;
1938 }
1939 }
1940 if (sc->sc_flags & SCF_TIMESTAMP) {
1941 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1942 to.to_tsecr = sc->sc_tsreflect;
1943 to.to_flags |= TOF_TS;
1944 }
1945 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1946
1947 /* Adjust headers by option size. */
1948 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1949 m->m_len += optlen;
1950 m->m_pkthdr.len += optlen;
1951 #ifdef INET6
1952 if (sc->sc_inc.inc_flags & INC_ISIPV6)
1953 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1954 else
1955 #endif
1956 ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1957 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1958 if (sc->sc_flags & SCF_SIGNATURE) {
1959 KASSERT(to.to_flags & TOF_SIGNATURE,
1960 ("tcp_addoptions() didn't set tcp_signature"));
1961
1962 /* NOTE: to.to_signature is inside of mbuf */
1963 if (!TCPMD5_ENABLED() ||
1964 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1965 m_freem(m);
1966 return (EACCES);
1967 }
1968 }
1969 #endif
1970 } else
1971 optlen = 0;
1972
1973 if (udp) {
1974 ulen += optlen;
1975 udp->uh_ulen = htons(ulen);
1976 }
1977 M_SETFIB(m, sc->sc_inc.inc_fibnum);
1978 /*
1979 * If we have peer's SYN and it has a flowid, then let's assign it to
1980 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid
1981 * to SYN|ACK due to lack of inp here.
1982 */
1983 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1984 m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1985 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1986 }
1987 #ifdef INET6
1988 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1989 if (sc->sc_port) {
1990 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
1991 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1992 udp->uh_sum = in6_cksum_pseudo(ip6, ulen,
1993 IPPROTO_UDP, 0);
1994 th->th_sum = htons(0);
1995 } else {
1996 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1997 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1998 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1999 IPPROTO_TCP, 0);
2000 }
2001 ip6->ip6_hlim = sc->sc_ip_ttl;
2002 #ifdef TCP_OFFLOAD
2003 if (ADDED_BY_TOE(sc)) {
2004 struct toedev *tod = sc->sc_tod;
2005
2006 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2007
2008 return (error);
2009 }
2010 #endif
2011 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
2012 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
2013 }
2014 #endif
2015 #if defined(INET6) && defined(INET)
2016 else
2017 #endif
2018 #ifdef INET
2019 {
2020 if (sc->sc_port) {
2021 m->m_pkthdr.csum_flags = CSUM_UDP;
2022 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2023 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
2024 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
2025 th->th_sum = htons(0);
2026 } else {
2027 m->m_pkthdr.csum_flags = CSUM_TCP;
2028 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2029 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
2030 htons(tlen + optlen - hlen + IPPROTO_TCP));
2031 }
2032 #ifdef TCP_OFFLOAD
2033 if (ADDED_BY_TOE(sc)) {
2034 struct toedev *tod = sc->sc_tod;
2035
2036 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2037
2038 return (error);
2039 }
2040 #endif
2041 TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
2042 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
2043 }
2044 #endif
2045 return (error);
2046 }
2047
2048 static void
syncache_send_challenge_ack(struct syncache * sc,struct mbuf * m)2049 syncache_send_challenge_ack(struct syncache *sc, struct mbuf *m)
2050 {
2051 if (tcp_challenge_ack_check(&sc->sc_challenge_ack_end,
2052 &sc->sc_challenge_ack_cnt)) {
2053 if (syncache_respond(sc, m, TH_ACK) == 0) {
2054 TCPSTAT_INC(tcps_sndacks);
2055 TCPSTAT_INC(tcps_sndtotal);
2056 }
2057 }
2058 }
2059
2060 /*
2061 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
2062 * that exceed the capacity of the syncache by avoiding the storage of any
2063 * of the SYNs we receive. Syncookies defend against blind SYN flooding
2064 * attacks where the attacker does not have access to our responses.
2065 *
2066 * Syncookies encode and include all necessary information about the
2067 * connection setup within the SYN|ACK that we send back. That way we
2068 * can avoid keeping any local state until the ACK to our SYN|ACK returns
2069 * (if ever). Normally the syncache and syncookies are running in parallel
2070 * with the latter taking over when the former is exhausted. When matching
2071 * syncache entry is found the syncookie is ignored.
2072 *
2073 * The only reliable information persisting the 3WHS is our initial sequence
2074 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient
2075 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
2076 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK
2077 * returns and signifies a legitimate connection if it matches the ACK.
2078 *
2079 * The available space of 32 bits to store the hash and to encode the SYN
2080 * option information is very tight and we should have at least 24 bits for
2081 * the MAC to keep the number of guesses by blind spoofing reasonably high.
2082 *
2083 * SYN option information we have to encode to fully restore a connection:
2084 * MSS: is imporant to chose an optimal segment size to avoid IP level
2085 * fragmentation along the path. The common MSS values can be encoded
2086 * in a 3-bit table. Uncommon values are captured by the next lower value
2087 * in the table leading to a slight increase in packetization overhead.
2088 * WSCALE: is necessary to allow large windows to be used for high delay-
2089 * bandwidth product links. Not scaling the window when it was initially
2090 * negotiated is bad for performance as lack of scaling further decreases
2091 * the apparent available send window. We only need to encode the WSCALE
2092 * we received from the remote end. Our end can be recalculated at any
2093 * time. The common WSCALE values can be encoded in a 3-bit table.
2094 * Uncommon values are captured by the next lower value in the table
2095 * making us under-estimate the available window size halving our
2096 * theoretically possible maximum throughput for that connection.
2097 * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2098 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2099 * that are included in all segments on a connection. We enable them when
2100 * the ACK has them.
2101 *
2102 * Security of syncookies and attack vectors:
2103 *
2104 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2105 * together with the gloabl secret to make it unique per connection attempt.
2106 * Thus any change of any of those parameters results in a different MAC output
2107 * in an unpredictable way unless a collision is encountered. 24 bits of the
2108 * MAC are embedded into the ISS.
2109 *
2110 * To prevent replay attacks two rotating global secrets are updated with a
2111 * new random value every 15 seconds. The life-time of a syncookie is thus
2112 * 15-30 seconds.
2113 *
2114 * Vector 1: Attacking the secret. This requires finding a weakness in the
2115 * MAC itself or the way it is used here. The attacker can do a chosen plain
2116 * text attack by varying and testing the all parameters under his control.
2117 * The strength depends on the size and randomness of the secret, and the
2118 * cryptographic security of the MAC function. Due to the constant updating
2119 * of the secret the attacker has at most 29.999 seconds to find the secret
2120 * and launch spoofed connections. After that he has to start all over again.
2121 *
2122 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC
2123 * size an average of 4,823 attempts are required for a 50% chance of success
2124 * to spoof a single syncookie (birthday collision paradox). However the
2125 * attacker is blind and doesn't know if one of his attempts succeeded unless
2126 * he has a side channel to interfere success from. A single connection setup
2127 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2128 * This many attempts are required for each one blind spoofed connection. For
2129 * every additional spoofed connection he has to launch another N attempts.
2130 * Thus for a sustained rate 100 spoofed connections per second approximately
2131 * 1,800,000 packets per second would have to be sent.
2132 *
2133 * NB: The MAC function should be fast so that it doesn't become a CPU
2134 * exhaustion attack vector itself.
2135 *
2136 * References:
2137 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2138 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2139 * http://cr.yp.to/syncookies.html (overview)
2140 * http://cr.yp.to/syncookies/archive (details)
2141 *
2142 *
2143 * Schematic construction of a syncookie enabled Initial Sequence Number:
2144 * 0 1 2 3
2145 * 12345678901234567890123456789012
2146 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2147 *
2148 * x 24 MAC (truncated)
2149 * W 3 Send Window Scale index
2150 * M 3 MSS index
2151 * S 1 SACK permitted
2152 * P 1 Odd/even secret
2153 */
2154
2155 /*
2156 * Distribution and probability of certain MSS values. Those in between are
2157 * rounded down to the next lower one.
2158 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2159 * .2% .3% 5% 7% 7% 20% 15% 45%
2160 */
2161 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2162
2163 /*
2164 * Distribution and probability of certain WSCALE values. We have to map the
2165 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2166 * bits based on prevalence of certain values. Where we don't have an exact
2167 * match for are rounded down to the next lower one letting us under-estimate
2168 * the true available window. At the moment this would happen only for the
2169 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2170 * and window size). The absence of the WSCALE option (no scaling in either
2171 * direction) is encoded with index zero.
2172 * [WSCALE values histograms, Allman, 2012]
2173 * X 10 10 35 5 6 14 10% by host
2174 * X 11 4 5 5 18 49 3% by connections
2175 */
2176 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2177
2178 /*
2179 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed
2180 * and good cryptographic properties.
2181 */
2182 static uint32_t
syncookie_mac(struct in_conninfo * inc,tcp_seq irs,uint8_t flags,uint8_t * secbits,uintptr_t secmod)2183 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2184 uint8_t *secbits, uintptr_t secmod)
2185 {
2186 SIPHASH_CTX ctx;
2187 uint32_t siphash[2];
2188
2189 SipHash24_Init(&ctx);
2190 SipHash_SetKey(&ctx, secbits);
2191 switch (inc->inc_flags & INC_ISIPV6) {
2192 #ifdef INET
2193 case 0:
2194 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2195 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2196 break;
2197 #endif
2198 #ifdef INET6
2199 case INC_ISIPV6:
2200 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2201 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2202 break;
2203 #endif
2204 }
2205 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2206 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2207 SipHash_Update(&ctx, &irs, sizeof(irs));
2208 SipHash_Update(&ctx, &flags, sizeof(flags));
2209 SipHash_Update(&ctx, &secmod, sizeof(secmod));
2210 SipHash_Final((u_int8_t *)&siphash, &ctx);
2211
2212 return (siphash[0] ^ siphash[1]);
2213 }
2214
2215 static tcp_seq
syncookie_generate(struct syncache_head * sch,struct syncache * sc)2216 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2217 {
2218 u_int i, secbit, wscale;
2219 uint32_t iss, hash;
2220 uint8_t *secbits;
2221 union syncookie cookie;
2222
2223 cookie.cookie = 0;
2224
2225 /* Map our computed MSS into the 3-bit index. */
2226 for (i = nitems(tcp_sc_msstab) - 1;
2227 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2228 i--)
2229 ;
2230 cookie.flags.mss_idx = i;
2231
2232 /*
2233 * Map the send window scale into the 3-bit index but only if
2234 * the wscale option was received.
2235 */
2236 if (sc->sc_flags & SCF_WINSCALE) {
2237 wscale = sc->sc_requested_s_scale;
2238 for (i = nitems(tcp_sc_wstab) - 1;
2239 tcp_sc_wstab[i] > wscale && i > 0;
2240 i--)
2241 ;
2242 cookie.flags.wscale_idx = i;
2243 }
2244
2245 /* Can we do SACK? */
2246 if (sc->sc_flags & SCF_SACK)
2247 cookie.flags.sack_ok = 1;
2248
2249 /* Which of the two secrets to use. */
2250 secbit = V_tcp_syncache.secret.oddeven & 0x1;
2251 cookie.flags.odd_even = secbit;
2252
2253 secbits = V_tcp_syncache.secret.key[secbit];
2254 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2255 (uintptr_t)sch);
2256
2257 /*
2258 * Put the flags into the hash and XOR them to get better ISS number
2259 * variance. This doesn't enhance the cryptographic strength and is
2260 * done to prevent the 8 cookie bits from showing up directly on the
2261 * wire.
2262 */
2263 iss = hash & ~0xff;
2264 iss |= cookie.cookie ^ (hash >> 24);
2265
2266 TCPSTAT_INC(tcps_sc_sendcookie);
2267 return (iss);
2268 }
2269
2270 static bool
syncookie_expand(struct in_conninfo * inc,const struct syncache_head * sch,struct syncache * sc,struct tcphdr * th,struct tcpopt * to,struct socket * lso,uint16_t port)2271 syncookie_expand(struct in_conninfo *inc, const struct syncache_head *sch,
2272 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2273 struct socket *lso, uint16_t port)
2274 {
2275 uint32_t hash;
2276 uint8_t *secbits;
2277 tcp_seq ack, seq;
2278 int wnd;
2279 union syncookie cookie;
2280
2281 /*
2282 * Pull information out of SYN-ACK/ACK and revert sequence number
2283 * advances.
2284 */
2285 ack = th->th_ack - 1;
2286 seq = th->th_seq - 1;
2287
2288 /*
2289 * Unpack the flags containing enough information to restore the
2290 * connection.
2291 */
2292 cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2293
2294 /* Which of the two secrets to use. */
2295 secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2296
2297 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2298
2299 /* The recomputed hash matches the ACK if this was a genuine cookie. */
2300 if ((ack & ~0xff) != (hash & ~0xff))
2301 return (false);
2302
2303 /* Fill in the syncache values. */
2304 sc->sc_flags = 0;
2305 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2306 sc->sc_ipopts = NULL;
2307
2308 sc->sc_irs = seq;
2309 sc->sc_iss = ack;
2310
2311 switch (inc->inc_flags & INC_ISIPV6) {
2312 #ifdef INET
2313 case 0:
2314 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2315 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2316 break;
2317 #endif
2318 #ifdef INET6
2319 case INC_ISIPV6:
2320 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2321 sc->sc_flowlabel =
2322 htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2323 break;
2324 #endif
2325 }
2326
2327 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2328
2329 /* Only use wscale if it was enabled in the orignal SYN. */
2330 if (cookie.flags.wscale_idx > 0) {
2331 u_int wscale = 0;
2332
2333 /* Recompute the receive window scale that was sent earlier. */
2334 while (wscale < TCP_MAX_WINSHIFT &&
2335 (TCP_MAXWIN << wscale) < sb_max)
2336 wscale++;
2337 sc->sc_requested_r_scale = wscale;
2338 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2339 sc->sc_flags |= SCF_WINSCALE;
2340 }
2341
2342 wnd = lso->sol_sbrcv_hiwat;
2343 wnd = imax(wnd, 0);
2344 wnd = imin(wnd, TCP_MAXWIN);
2345 sc->sc_wnd = wnd;
2346
2347 if (cookie.flags.sack_ok)
2348 sc->sc_flags |= SCF_SACK;
2349
2350 if (to->to_flags & TOF_TS) {
2351 sc->sc_flags |= SCF_TIMESTAMP;
2352 sc->sc_tsreflect = to->to_tsval;
2353 sc->sc_tsoff = tcp_new_ts_offset(inc);
2354 }
2355
2356 if (to->to_flags & TOF_SIGNATURE)
2357 sc->sc_flags |= SCF_SIGNATURE;
2358
2359 sc->sc_rxmits = 0;
2360
2361 sc->sc_port = port;
2362
2363 return (true);
2364 }
2365
2366 #ifdef INVARIANTS
2367 static void
syncookie_cmp(struct in_conninfo * inc,const struct syncache_head * sch,struct syncache * sc,struct tcphdr * th,struct tcpopt * to,struct socket * lso,uint16_t port)2368 syncookie_cmp(struct in_conninfo *inc, const struct syncache_head *sch,
2369 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2370 struct socket *lso, uint16_t port)
2371 {
2372 struct syncache scs;
2373 char *s;
2374
2375 bzero(&scs, sizeof(scs));
2376 if (syncookie_expand(inc, sch, &scs, th, to, lso, port) &&
2377 (sc->sc_peer_mss != scs.sc_peer_mss ||
2378 sc->sc_requested_r_scale != scs.sc_requested_r_scale ||
2379 sc->sc_requested_s_scale != scs.sc_requested_s_scale ||
2380 (sc->sc_flags & SCF_SACK) != (scs.sc_flags & SCF_SACK))) {
2381
2382 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2383 return;
2384
2385 if (sc->sc_peer_mss != scs.sc_peer_mss)
2386 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2387 s, __func__, sc->sc_peer_mss, scs.sc_peer_mss);
2388
2389 if (sc->sc_requested_r_scale != scs.sc_requested_r_scale)
2390 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2391 s, __func__, sc->sc_requested_r_scale,
2392 scs.sc_requested_r_scale);
2393
2394 if (sc->sc_requested_s_scale != scs.sc_requested_s_scale)
2395 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2396 s, __func__, sc->sc_requested_s_scale,
2397 scs.sc_requested_s_scale);
2398
2399 if ((sc->sc_flags & SCF_SACK) != (scs.sc_flags & SCF_SACK))
2400 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2401
2402 free(s, M_TCPLOG);
2403 }
2404 }
2405 #endif /* INVARIANTS */
2406
2407 static void
syncookie_reseed(void * arg)2408 syncookie_reseed(void *arg)
2409 {
2410 struct tcp_syncache *sc = arg;
2411 uint8_t *secbits;
2412 int secbit;
2413
2414 /*
2415 * Reseeding the secret doesn't have to be protected by a lock.
2416 * It only must be ensured that the new random values are visible
2417 * to all CPUs in a SMP environment. The atomic with release
2418 * semantics ensures that.
2419 */
2420 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2421 secbits = sc->secret.key[secbit];
2422 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2423 atomic_add_rel_int(&sc->secret.oddeven, 1);
2424
2425 /* Reschedule ourself. */
2426 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2427 }
2428
2429 /*
2430 * We have overflowed a bucket. Let's pause dealing with the syncache.
2431 * This function will increment the bucketoverflow statistics appropriately
2432 * (once per pause when pausing is enabled; otherwise, once per overflow).
2433 */
2434 static void
syncache_pause(struct in_conninfo * inc)2435 syncache_pause(struct in_conninfo *inc)
2436 {
2437 time_t delta;
2438 const char *s;
2439
2440 /* XXX:
2441 * 2. Add sysctl read here so we don't get the benefit of this
2442 * change without the new sysctl.
2443 */
2444
2445 /*
2446 * Try an unlocked read. If we already know that another thread
2447 * has activated the feature, there is no need to proceed.
2448 */
2449 if (V_tcp_syncache.paused)
2450 return;
2451
2452 /* Are cookied enabled? If not, we can't pause. */
2453 if (!V_tcp_syncookies) {
2454 TCPSTAT_INC(tcps_sc_bucketoverflow);
2455 return;
2456 }
2457
2458 /*
2459 * We may be the first thread to find an overflow. Get the lock
2460 * and evaluate if we need to take action.
2461 */
2462 mtx_lock(&V_tcp_syncache.pause_mtx);
2463 if (V_tcp_syncache.paused) {
2464 mtx_unlock(&V_tcp_syncache.pause_mtx);
2465 return;
2466 }
2467
2468 /* Activate protection. */
2469 V_tcp_syncache.paused = true;
2470 TCPSTAT_INC(tcps_sc_bucketoverflow);
2471
2472 /*
2473 * Determine the last backoff time. If we are seeing a re-newed
2474 * attack within that same time after last reactivating the syncache,
2475 * consider it an extension of the same attack.
2476 */
2477 delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2478 if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2479 if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2480 delta <<= 1;
2481 V_tcp_syncache.pause_backoff++;
2482 }
2483 } else {
2484 delta = TCP_SYNCACHE_PAUSE_TIME;
2485 V_tcp_syncache.pause_backoff = 0;
2486 }
2487
2488 /* Log a warning, including IP addresses, if able. */
2489 if (inc != NULL)
2490 s = tcp_log_addrs(inc, NULL, NULL, NULL);
2491 else
2492 s = (const char *)NULL;
2493 log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2494 "the next %lld seconds%s%s%s\n", (long long)delta,
2495 (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2496 (s != NULL) ? ")" : "");
2497 free(__DECONST(void *, s), M_TCPLOG);
2498
2499 /* Use the calculated delta to set a new pause time. */
2500 V_tcp_syncache.pause_until = time_uptime + delta;
2501 callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2502 &V_tcp_syncache);
2503 mtx_unlock(&V_tcp_syncache.pause_mtx);
2504 }
2505
2506 /* Evaluate whether we need to unpause. */
2507 static void
syncache_unpause(void * arg)2508 syncache_unpause(void *arg)
2509 {
2510 struct tcp_syncache *sc;
2511 time_t delta;
2512
2513 sc = arg;
2514 mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2515 callout_deactivate(&sc->pause_co);
2516
2517 /*
2518 * Check to make sure we are not running early. If the pause
2519 * time has expired, then deactivate the protection.
2520 */
2521 if ((delta = sc->pause_until - time_uptime) > 0)
2522 callout_schedule(&sc->pause_co, delta * hz);
2523 else
2524 sc->paused = false;
2525 }
2526
2527 /*
2528 * Exports the syncache entries to userland so that netstat can display
2529 * them alongside the other sockets. This function is intended to be
2530 * called only from tcp_pcblist.
2531 *
2532 * Due to concurrency on an active system, the number of pcbs exported
2533 * may have no relation to max_pcbs. max_pcbs merely indicates the
2534 * amount of space the caller allocated for this function to use.
2535 */
2536 int
syncache_pcblist(struct sysctl_req * req)2537 syncache_pcblist(struct sysctl_req *req)
2538 {
2539 struct xtcpcb xt;
2540 struct syncache *sc;
2541 struct syncache_head *sch;
2542 int error, i;
2543
2544 bzero(&xt, sizeof(xt));
2545 xt.xt_len = sizeof(xt);
2546 xt.t_state = TCPS_SYN_RECEIVED;
2547 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2548 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2549 xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2550 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2551
2552 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2553 sch = &V_tcp_syncache.hashbase[i];
2554 SCH_LOCK(sch);
2555 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2556 if (sc->sc_cred != NULL &&
2557 cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2558 continue;
2559 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2560 xt.xt_inp.inp_vflag = INP_IPV6;
2561 else
2562 xt.xt_inp.inp_vflag = INP_IPV4;
2563 xt.xt_encaps_port = sc->sc_port;
2564 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2565 sizeof (struct in_conninfo));
2566 error = SYSCTL_OUT(req, &xt, sizeof xt);
2567 if (error) {
2568 SCH_UNLOCK(sch);
2569 return (0);
2570 }
2571 }
2572 SCH_UNLOCK(sch);
2573 }
2574
2575 return (0);
2576 }
2577