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 "opt_inet.h"
36 #include "opt_inet6.h"
37 #include "opt_ipsec.h"
38
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/hash.h>
42 #include <sys/refcount.h>
43 #include <sys/kernel.h>
44 #include <sys/sysctl.h>
45 #include <sys/limits.h>
46 #include <sys/lock.h>
47 #include <sys/mutex.h>
48 #include <sys/malloc.h>
49 #include <sys/mbuf.h>
50 #include <sys/proc.h> /* for proc0 declaration */
51 #include <sys/random.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #include <sys/syslog.h>
55 #include <sys/ucred.h>
56
57 #include <sys/md5.h>
58 #include <crypto/siphash/siphash.h>
59
60 #include <vm/uma.h>
61
62 #include <net/if.h>
63 #include <net/if_var.h>
64 #include <net/route.h>
65 #include <net/vnet.h>
66
67 #include <netinet/in.h>
68 #include <netinet/in_kdtrace.h>
69 #include <netinet/in_systm.h>
70 #include <netinet/ip.h>
71 #include <netinet/in_var.h>
72 #include <netinet/in_pcb.h>
73 #include <netinet/in_rss.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 #include <netinet6/in6_rss.h>
83 #endif
84 #include <netinet/tcp.h>
85 #include <netinet/tcp_fastopen.h>
86 #include <netinet/tcp_fsm.h>
87 #include <netinet/tcp_seq.h>
88 #include <netinet/tcp_timer.h>
89 #include <netinet/tcp_var.h>
90 #include <netinet/tcp_syncache.h>
91 #include <netinet/tcp_ecn.h>
92 #ifdef TCP_BLACKBOX
93 #include <netinet/tcp_log_buf.h>
94 #endif
95 #ifdef TCP_OFFLOAD
96 #include <netinet/toecore.h>
97 #endif
98 #include <netinet/udp.h>
99
100 #include <netipsec/ipsec_support.h>
101
102 #include <machine/in_cksum.h>
103
104 #include <security/mac/mac_framework.h>
105
106 VNET_DEFINE_STATIC(bool, tcp_syncookies) = true;
107 #define V_tcp_syncookies VNET(tcp_syncookies)
108 SYSCTL_BOOL(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
109 &VNET_NAME(tcp_syncookies), 0,
110 "Use TCP SYN cookies if the syncache overflows");
111
112 VNET_DEFINE_STATIC(bool, tcp_syncookiesonly) = false;
113 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly)
114 SYSCTL_BOOL(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
115 &VNET_NAME(tcp_syncookiesonly), 0,
116 "Use only TCP SYN cookies");
117
118 #ifdef TCP_OFFLOAD
119 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
120 #endif
121
122 static void syncache_drop(struct syncache *, struct syncache_head *);
123 static void syncache_free(struct syncache *);
124 static void syncache_insert(struct syncache *, struct syncache_head *);
125 static int syncache_respond(struct syncache *, int);
126 static void syncache_send_challenge_ack(struct syncache *);
127 static struct socket *syncache_socket(struct syncache *, struct socket *,
128 struct mbuf *m);
129 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch,
130 int docallout);
131 static void syncache_timer(void *);
132
133 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
134 uint8_t *, uintptr_t);
135 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *);
136 static bool syncookie_expand(struct in_conninfo *,
137 const struct syncache_head *, struct syncache *,
138 struct tcphdr *, struct tcpopt *, struct socket *,
139 uint16_t);
140 static void syncache_pause(struct in_conninfo *);
141 static void syncache_unpause(void *);
142 static void syncookie_reseed(void *);
143 #ifdef INVARIANTS
144 static void syncookie_cmp(struct in_conninfo *,
145 const struct syncache_head *, struct syncache *,
146 struct tcphdr *, struct tcpopt *, struct socket *,
147 uint16_t);
148 #endif
149
150 /*
151 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
152 * 3 retransmits corresponds to a timeout with default values of
153 * tcp_rexmit_initial * ( 1 +
154 * tcp_backoff[1] +
155 * tcp_backoff[2] +
156 * tcp_backoff[3]) + 3 * tcp_rexmit_slop,
157 * 1000 ms * (1 + 2 + 4 + 8) + 3 * 200 ms = 15600 ms,
158 * the odds are that the user has given up attempting to connect by then.
159 */
160 #define SYNCACHE_MAXREXMTS 3
161
162 /* Arbitrary values */
163 #define TCP_SYNCACHE_HASHSIZE 512
164 #define TCP_SYNCACHE_BUCKETLIMIT 30
165
166 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
167 #define V_tcp_syncache VNET(tcp_syncache)
168
169 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache,
170 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
171 "TCP SYN cache");
172
173 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
174 &VNET_NAME(tcp_syncache.bucket_limit), 0,
175 "Per-bucket hash limit for syncache");
176
177 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
178 &VNET_NAME(tcp_syncache.cache_limit), 0,
179 "Overall entry limit for syncache");
180
181 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
182 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
183
184 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
185 &VNET_NAME(tcp_syncache.hashsize), 0,
186 "Size of TCP syncache hashtable");
187
188 SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET |
189 CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0,
190 "All syncache(4) entries are visible, ignoring UID/GID, jail(2) "
191 "and mac(4) checks");
192
193 static int
sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)194 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
195 {
196 int error;
197 u_int new;
198
199 new = V_tcp_syncache.rexmt_limit;
200 error = sysctl_handle_int(oidp, &new, 0, req);
201 if ((error == 0) && (req->newptr != NULL)) {
202 if (new > TCP_MAXRXTSHIFT)
203 error = EINVAL;
204 else
205 V_tcp_syncache.rexmt_limit = new;
206 }
207 return (error);
208 }
209
210 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
211 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
212 &VNET_NAME(tcp_syncache.rexmt_limit), 0,
213 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "IU",
214 "Limit on SYN/ACK retransmissions");
215
216 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
217 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
218 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
219 "Send reset on socket allocation failure");
220
221 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
222
223 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx)
224 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx)
225 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED)
226
227 /*
228 * Requires the syncache entry to be already removed from the bucket list.
229 */
230 static void
syncache_free(struct syncache * sc)231 syncache_free(struct syncache *sc)
232 {
233
234 if (sc->sc_ipopts)
235 (void)m_free(sc->sc_ipopts);
236 if (sc->sc_cred)
237 crfree(sc->sc_cred);
238 #ifdef MAC
239 mac_syncache_destroy(&sc->sc_label);
240 #endif
241
242 uma_zfree(V_tcp_syncache.zone, sc);
243 }
244
245 void
syncache_init(void)246 syncache_init(void)
247 {
248 int i;
249
250 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
251 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
252 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
253 V_tcp_syncache.hash_secret = arc4random();
254
255 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
256 &V_tcp_syncache.hashsize);
257 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
258 &V_tcp_syncache.bucket_limit);
259 if (!powerof2(V_tcp_syncache.hashsize) ||
260 V_tcp_syncache.hashsize == 0) {
261 printf("WARNING: syncache hash size is not a power of 2.\n");
262 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
263 }
264 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
265
266 /* Set limits. */
267 V_tcp_syncache.cache_limit =
268 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
269 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
270 &V_tcp_syncache.cache_limit);
271
272 /* Allocate the hash table. */
273 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
274 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
275
276 #ifdef VIMAGE
277 V_tcp_syncache.vnet = curvnet;
278 #endif
279
280 /* Initialize the hash buckets. */
281 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
282 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
283 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
284 NULL, MTX_DEF);
285 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
286 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
287 V_tcp_syncache.hashbase[i].sch_length = 0;
288 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
289 V_tcp_syncache.hashbase[i].sch_last_overflow =
290 -(SYNCOOKIE_LIFETIME + 1);
291 }
292
293 /* Create the syncache entry zone. */
294 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
295 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
296 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
297 V_tcp_syncache.cache_limit);
298
299 /* Start the SYN cookie reseeder callout. */
300 callout_init(&V_tcp_syncache.secret.reseed, 1);
301 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
302 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
303 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
304 syncookie_reseed, &V_tcp_syncache);
305
306 /* Initialize the pause machinery. */
307 mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF);
308 callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx,
309 0);
310 V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME;
311 V_tcp_syncache.pause_backoff = 0;
312 V_tcp_syncache.paused = false;
313 }
314
315 #ifdef VIMAGE
316 void
syncache_destroy(void)317 syncache_destroy(void)
318 {
319 struct syncache_head *sch;
320 struct syncache *sc, *nsc;
321 int i;
322
323 /*
324 * Stop the re-seed timer before freeing resources. No need to
325 * possibly schedule it another time.
326 */
327 callout_drain(&V_tcp_syncache.secret.reseed);
328
329 /* Stop the SYN cache pause callout. */
330 mtx_lock(&V_tcp_syncache.pause_mtx);
331 if (callout_stop(&V_tcp_syncache.pause_co) == 0) {
332 mtx_unlock(&V_tcp_syncache.pause_mtx);
333 callout_drain(&V_tcp_syncache.pause_co);
334 } else
335 mtx_unlock(&V_tcp_syncache.pause_mtx);
336
337 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */
338 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
339 sch = &V_tcp_syncache.hashbase[i];
340 callout_drain(&sch->sch_timer);
341
342 SCH_LOCK(sch);
343 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
344 syncache_drop(sc, sch);
345 SCH_UNLOCK(sch);
346 KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
347 ("%s: sch->sch_bucket not empty", __func__));
348 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
349 __func__, sch->sch_length));
350 mtx_destroy(&sch->sch_mtx);
351 }
352
353 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
354 ("%s: cache_count not 0", __func__));
355
356 /* Free the allocated global resources. */
357 uma_zdestroy(V_tcp_syncache.zone);
358 free(V_tcp_syncache.hashbase, M_SYNCACHE);
359 mtx_destroy(&V_tcp_syncache.pause_mtx);
360 }
361 #endif
362
363 /*
364 * Inserts a syncache entry into the specified bucket row.
365 * Locks and unlocks the syncache_head autonomously.
366 */
367 static void
syncache_insert(struct syncache * sc,struct syncache_head * sch)368 syncache_insert(struct syncache *sc, struct syncache_head *sch)
369 {
370 struct syncache *sc2;
371
372 SCH_LOCK(sch);
373
374 /*
375 * Make sure that we don't overflow the per-bucket limit.
376 * If the bucket is full, toss the oldest element.
377 */
378 if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
379 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
380 ("sch->sch_length incorrect"));
381 syncache_pause(&sc->sc_inc);
382 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
383 sch->sch_last_overflow = time_uptime;
384 syncache_drop(sc2, sch);
385 }
386
387 /* Put it into the bucket. */
388 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
389 sch->sch_length++;
390
391 #ifdef TCP_OFFLOAD
392 if (ADDED_BY_TOE(sc)) {
393 struct toedev *tod = sc->sc_tod;
394
395 tod->tod_syncache_added(tod, sc->sc_todctx);
396 }
397 #endif
398
399 /* Reinitialize the bucket row's timer. */
400 if (sch->sch_length == 1)
401 sch->sch_nextc = ticks + INT_MAX;
402 syncache_timeout(sc, sch, 1);
403
404 SCH_UNLOCK(sch);
405
406 TCPSTATES_INC(TCPS_SYN_RECEIVED);
407 TCPSTAT_INC(tcps_sc_added);
408 }
409
410 /*
411 * Remove and free entry from syncache bucket row.
412 * Expects locked syncache head.
413 */
414 static void
syncache_drop(struct syncache * sc,struct syncache_head * sch)415 syncache_drop(struct syncache *sc, struct syncache_head *sch)
416 {
417
418 SCH_LOCK_ASSERT(sch);
419
420 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
421 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
422 sch->sch_length--;
423
424 #ifdef TCP_OFFLOAD
425 if (ADDED_BY_TOE(sc)) {
426 struct toedev *tod = sc->sc_tod;
427
428 tod->tod_syncache_removed(tod, sc->sc_todctx);
429 }
430 #endif
431
432 syncache_free(sc);
433 }
434
435 /*
436 * Engage/reengage time on bucket row.
437 */
438 static void
syncache_timeout(struct syncache * sc,struct syncache_head * sch,int docallout)439 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
440 {
441 int rexmt;
442
443 if (sc->sc_rxmits == 0)
444 rexmt = tcp_rexmit_initial;
445 else
446 TCPT_RANGESET(rexmt,
447 tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
448 tcp_rexmit_min, tcp_rexmit_max);
449 sc->sc_rxttime = ticks + rexmt;
450 sc->sc_rxmits++;
451 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
452 sch->sch_nextc = sc->sc_rxttime;
453 if (docallout)
454 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
455 syncache_timer, (void *)sch);
456 }
457 }
458
459 /*
460 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
461 * If we have retransmitted an entry the maximum number of times, expire it.
462 * One separate timer for each bucket row.
463 */
464 static void
syncache_timer(void * xsch)465 syncache_timer(void *xsch)
466 {
467 struct syncache_head *sch = (struct syncache_head *)xsch;
468 struct syncache *sc, *nsc;
469 struct epoch_tracker et;
470 int tick = ticks;
471 char *s;
472 bool paused;
473
474 CURVNET_SET(sch->sch_sc->vnet);
475
476 /* NB: syncache_head has already been locked by the callout. */
477 SCH_LOCK_ASSERT(sch);
478
479 /*
480 * In the following cycle we may remove some entries and/or
481 * advance some timeouts, so re-initialize the bucket timer.
482 */
483 sch->sch_nextc = tick + INT_MAX;
484
485 /*
486 * If we have paused processing, unconditionally remove
487 * all syncache entries.
488 */
489 mtx_lock(&V_tcp_syncache.pause_mtx);
490 paused = V_tcp_syncache.paused;
491 mtx_unlock(&V_tcp_syncache.pause_mtx);
492
493 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
494 if (paused) {
495 syncache_drop(sc, sch);
496 continue;
497 }
498 /*
499 * We do not check if the listen socket still exists
500 * and accept the case where the listen socket may be
501 * gone by the time we resend the SYN/ACK. We do
502 * not expect this to happens often. If it does,
503 * then the RST will be sent by the time the remote
504 * host does the SYN/ACK->ACK.
505 */
506 if (TSTMP_GT(sc->sc_rxttime, tick)) {
507 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
508 sch->sch_nextc = sc->sc_rxttime;
509 continue;
510 }
511 if (sc->sc_rxmits > V_tcp_ecn_maxretries) {
512 sc->sc_flags &= ~SCF_ECN_MASK;
513 }
514 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
515 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
516 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
517 "giving up and removing syncache entry\n",
518 s, __func__);
519 free(s, M_TCPLOG);
520 }
521 syncache_drop(sc, sch);
522 TCPSTAT_INC(tcps_sc_stale);
523 continue;
524 }
525 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
526 log(LOG_DEBUG, "%s; %s: Response timeout, "
527 "retransmitting (%u) SYN|ACK\n",
528 s, __func__, sc->sc_rxmits);
529 free(s, M_TCPLOG);
530 }
531
532 NET_EPOCH_ENTER(et);
533 if (syncache_respond(sc, TH_SYN|TH_ACK) == 0) {
534 syncache_timeout(sc, sch, 0);
535 TCPSTAT_INC(tcps_sndacks);
536 TCPSTAT_INC(tcps_sndtotal);
537 TCPSTAT_INC(tcps_sc_retransmitted);
538 } else {
539 /*
540 * Most likely we are memory constrained, so free
541 * resources.
542 */
543 syncache_drop(sc, sch);
544 TCPSTAT_INC(tcps_sc_dropped);
545 }
546 NET_EPOCH_EXIT(et);
547 }
548 if (!TAILQ_EMPTY(&(sch)->sch_bucket))
549 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
550 syncache_timer, (void *)(sch));
551 CURVNET_RESTORE();
552 }
553
554 /*
555 * Returns true if the system is only using cookies at the moment.
556 * This could be due to a sysadmin decision to only use cookies, or it
557 * could be due to the system detecting an attack.
558 */
559 static inline bool
syncache_cookiesonly(void)560 syncache_cookiesonly(void)
561 {
562 return ((V_tcp_syncookies && V_tcp_syncache.paused) ||
563 V_tcp_syncookiesonly);
564 }
565
566 /*
567 * Find the hash bucket for the given connection.
568 */
569 static struct syncache_head *
syncache_hashbucket(struct in_conninfo * inc)570 syncache_hashbucket(struct in_conninfo *inc)
571 {
572 uint32_t hash;
573
574 /*
575 * The hash is built on foreign port + local port + foreign address.
576 * We rely on the fact that struct in_conninfo starts with 16 bits
577 * of foreign port, then 16 bits of local port then followed by 128
578 * bits of foreign address. In case of IPv4 address, the first 3
579 * 32-bit words of the address always are zeroes.
580 */
581 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
582 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
583
584 return (&V_tcp_syncache.hashbase[hash]);
585 }
586
587 /*
588 * Find an entry in the syncache.
589 * Returns always with locked syncache_head plus a matching entry or NULL.
590 */
591 static struct syncache *
syncache_lookup(struct in_conninfo * inc,struct syncache_head ** schp)592 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
593 {
594 struct syncache *sc;
595 struct syncache_head *sch;
596
597 *schp = sch = syncache_hashbucket(inc);
598 SCH_LOCK(sch);
599
600 /* Circle through bucket row to find matching entry. */
601 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
602 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
603 sizeof(struct in_endpoints)) == 0)
604 break;
605
606 return (sc); /* Always returns with locked sch. */
607 }
608
609 /*
610 * This function is called when we get a RST for a
611 * non-existent connection, so that we can see if the
612 * connection is in the syn cache. If it is, zap it.
613 * If required send a challenge ACK.
614 */
615 void
syncache_chkrst(struct in_conninfo * inc,struct tcphdr * th,uint16_t port)616 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, uint16_t port)
617 {
618 struct syncache *sc;
619 struct syncache_head *sch;
620 char *s = NULL;
621
622 if (syncache_cookiesonly())
623 return;
624 sc = syncache_lookup(inc, &sch); /* returns locked sch */
625 SCH_LOCK_ASSERT(sch);
626
627 /*
628 * No corresponding connection was found in syncache.
629 * If syncookies are enabled and possibly exclusively
630 * used, or we are under memory pressure, a valid RST
631 * may not find a syncache entry. In that case we're
632 * done and no SYN|ACK retransmissions will happen.
633 * Otherwise the RST was misdirected or spoofed.
634 */
635 if (sc == NULL) {
636 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
637 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
638 "syncache entry (possibly syncookie only), "
639 "segment ignored\n", s, __func__);
640 TCPSTAT_INC(tcps_badrst);
641 goto done;
642 }
643
644 /* The remote UDP encaps port does not match. */
645 if (sc->sc_port != port) {
646 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
647 log(LOG_DEBUG, "%s; %s: Spurious RST with matching "
648 "syncache entry but non-matching UDP encaps port, "
649 "segment ignored\n", s, __func__);
650 TCPSTAT_INC(tcps_badrst);
651 goto done;
652 }
653
654 /*
655 * If the RST bit is set, check the sequence number to see
656 * if this is a valid reset segment.
657 *
658 * RFC 793 page 37:
659 * In all states except SYN-SENT, all reset (RST) segments
660 * are validated by checking their SEQ-fields. A reset is
661 * valid if its sequence number is in the window.
662 *
663 * RFC 793 page 69:
664 * There are four cases for the acceptability test for an incoming
665 * segment:
666 *
667 * Segment Receive Test
668 * Length Window
669 * ------- ------- -------------------------------------------
670 * 0 0 SEG.SEQ = RCV.NXT
671 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
672 * >0 0 not acceptable
673 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
674 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
675 *
676 * Note that when receiving a SYN segment in the LISTEN state,
677 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
678 * described in RFC 793, page 66.
679 */
680 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
681 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
682 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
683 if (V_tcp_insecure_rst ||
684 th->th_seq == sc->sc_irs + 1) {
685 syncache_drop(sc, sch);
686 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
687 log(LOG_DEBUG,
688 "%s; %s: Our SYN|ACK was rejected, "
689 "connection attempt aborted by remote "
690 "endpoint\n",
691 s, __func__);
692 TCPSTAT_INC(tcps_sc_reset);
693 } else {
694 TCPSTAT_INC(tcps_badrst);
695 /* Send challenge ACK. */
696 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
697 log(LOG_DEBUG, "%s; %s: RST with invalid "
698 " SEQ %u != NXT %u (+WND %u), "
699 "sending challenge ACK\n",
700 s, __func__,
701 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
702 syncache_send_challenge_ack(sc);
703 }
704 } else {
705 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
706 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
707 "NXT %u (+WND %u), segment ignored\n",
708 s, __func__,
709 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
710 TCPSTAT_INC(tcps_badrst);
711 }
712
713 done:
714 if (s != NULL)
715 free(s, M_TCPLOG);
716 SCH_UNLOCK(sch);
717 }
718
719 void
syncache_unreach(struct in_conninfo * inc,tcp_seq th_seq,uint16_t port)720 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port)
721 {
722 struct syncache *sc;
723 struct syncache_head *sch;
724
725 if (syncache_cookiesonly())
726 return;
727 sc = syncache_lookup(inc, &sch); /* returns locked sch */
728 SCH_LOCK_ASSERT(sch);
729 if (sc == NULL)
730 goto done;
731
732 /* If the port != sc_port, then it's a bogus ICMP msg */
733 if (port != sc->sc_port)
734 goto done;
735
736 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
737 if (ntohl(th_seq) != sc->sc_iss)
738 goto done;
739
740 /*
741 * If we've retransmitted 3 times and this is our second error,
742 * we remove the entry. Otherwise, we allow it to continue on.
743 * This prevents us from incorrectly nuking an entry during a
744 * spurious network outage.
745 *
746 * See tcp_notify().
747 */
748 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
749 sc->sc_flags |= SCF_UNREACH;
750 goto done;
751 }
752 syncache_drop(sc, sch);
753 TCPSTAT_INC(tcps_sc_unreach);
754 done:
755 SCH_UNLOCK(sch);
756 }
757
758 /*
759 * Build a new TCP socket structure from a syncache entry.
760 *
761 * On success return the newly created socket with its underlying inp locked.
762 */
763 static struct socket *
syncache_socket(struct syncache * sc,struct socket * lso,struct mbuf * m)764 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
765 {
766 struct inpcb *inp = NULL;
767 struct socket *so;
768 struct tcpcb *tp;
769 int error;
770 char *s;
771
772 NET_EPOCH_ASSERT();
773
774 /*
775 * Creation of a socket via solisten_clone() bypasses call to pr_attach.
776 * That's why there is some pasted code from soattach() and from
777 * tcp_usr_attach() here. This should improve once TCP is PR_SOCKBUF.
778 */
779 if ((so = solisten_clone(lso)) == NULL)
780 goto allocfail;
781 mtx_init(&so->so_snd_mtx, "so_snd", NULL, MTX_DEF);
782 mtx_init(&so->so_rcv_mtx, "so_rcv", NULL, MTX_DEF);
783 so->so_snd.sb_mtx = &so->so_snd_mtx;
784 so->so_rcv.sb_mtx = &so->so_rcv_mtx;
785 error = soreserve(so, lso->sol_sbsnd_hiwat, lso->sol_sbrcv_hiwat);
786 if (error) {
787 sodealloc(so);
788 goto allocfail;
789 }
790 #ifdef MAC
791 mac_socketpeer_set_from_mbuf(m, so);
792 #endif
793 error = in_pcballoc(so, &V_tcbinfo);
794 if (error) {
795 sodealloc(so);
796 goto allocfail;
797 }
798 inp = sotoinpcb(so);
799 if ((tp = tcp_newtcpcb(inp, sototcpcb(lso))) == NULL) {
800 in_pcbfree(inp);
801 sodealloc(so);
802 goto allocfail;
803 }
804 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
805 #ifdef INET6
806 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
807 inp->inp_vflag &= ~INP_IPV4;
808 inp->inp_vflag |= INP_IPV6;
809 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
810 } else {
811 inp->inp_vflag &= ~INP_IPV6;
812 inp->inp_vflag |= INP_IPV4;
813 #endif
814 inp->inp_ip_ttl = sc->sc_ip_ttl;
815 inp->inp_ip_tos = sc->sc_ip_tos;
816 inp->inp_laddr = sc->sc_inc.inc_laddr;
817 #ifdef INET6
818 }
819 #endif
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 if (sc->sc_flowtype != M_HASHTYPE_NONE) {
891 inp->inp_flowid = sc->sc_flowid;
892 inp->inp_flowtype = sc->sc_flowtype;
893 } else {
894 /* assign flowid by software RSS hash */
895 #ifdef INET6
896 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
897 rss_proto_software_hash_v6(&inp->in6p_faddr,
898 &inp->in6p_laddr,
899 inp->inp_fport,
900 inp->inp_lport,
901 IPPROTO_TCP,
902 &inp->inp_flowid,
903 &inp->inp_flowtype);
904 } else
905 #endif /* INET6 */
906 {
907 #ifdef INET
908 rss_proto_software_hash_v4(inp->inp_faddr,
909 inp->inp_laddr,
910 inp->inp_fport,
911 inp->inp_lport,
912 IPPROTO_TCP,
913 &inp->inp_flowid,
914 &inp->inp_flowtype);
915 #endif /* INET */
916 }
917 }
918 #ifdef NUMA
919 inp->inp_numa_domain = sc->sc_numa_domain;
920 #endif
921
922 tp->t_state = TCPS_SYN_RECEIVED;
923 tp->iss = sc->sc_iss;
924 tp->irs = sc->sc_irs;
925 tp->t_port = sc->sc_port;
926 tcp_rcvseqinit(tp);
927 tcp_sendseqinit(tp);
928 tp->snd_wl1 = sc->sc_irs;
929 tp->snd_max = tp->iss + 1;
930 tp->snd_nxt = tp->iss + 1;
931 tp->rcv_up = sc->sc_irs + 1;
932 tp->rcv_wnd = sc->sc_wnd;
933 tp->rcv_adv += tp->rcv_wnd;
934 tp->last_ack_sent = tp->rcv_nxt;
935
936 tp->t_flags = sototcpcb(lso)->t_flags &
937 (TF_LRD|TF_NOPUSH|TF_NODELAY);
938 if (sc->sc_flags & SCF_NOOPT)
939 tp->t_flags |= TF_NOOPT;
940 else {
941 if (sc->sc_flags & SCF_WINSCALE) {
942 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
943 tp->snd_scale = sc->sc_requested_s_scale;
944 tp->request_r_scale = sc->sc_requested_r_scale;
945 }
946 if (sc->sc_flags & SCF_TIMESTAMP) {
947 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
948 tp->ts_recent = sc->sc_tsreflect;
949 tp->ts_recent_age = tcp_ts_getticks();
950 tp->ts_offset = sc->sc_tsoff;
951 }
952 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
953 if (sc->sc_flags & SCF_SIGNATURE)
954 tp->t_flags |= TF_SIGNATURE;
955 #endif
956 if (sc->sc_flags & SCF_SACK)
957 tp->t_flags |= TF_SACK_PERMIT;
958 }
959
960 tcp_ecn_syncache_socket(tp, sc);
961
962 /*
963 * Set up MSS and get cached values from tcp_hostcache.
964 * This might overwrite some of the defaults we just set.
965 */
966 tcp_mss(tp, sc->sc_peer_mss);
967
968 /*
969 * If the SYN,ACK was retransmitted, indicate that CWND to be
970 * limited to one segment in cc_conn_init().
971 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
972 */
973 if (sc->sc_rxmits > 1)
974 tp->snd_cwnd = 1;
975
976 /* Copy over the challenge ACK state. */
977 tp->t_challenge_ack_end = sc->sc_challenge_ack_end;
978 tp->t_challenge_ack_cnt = sc->sc_challenge_ack_cnt;
979
980 #ifdef TCP_OFFLOAD
981 /*
982 * Allow a TOE driver to install its hooks. Note that we hold the
983 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
984 * new connection before the TOE driver has done its thing.
985 */
986 if (ADDED_BY_TOE(sc)) {
987 struct toedev *tod = sc->sc_tod;
988
989 tod->tod_offload_socket(tod, sc->sc_todctx, so);
990 }
991 #endif
992 #ifdef TCP_BLACKBOX
993 /*
994 * Inherit the log state from the listening socket, if
995 * - the log state of the listening socket is not off and
996 * - the listening socket was not auto selected from all sessions and
997 * - a log id is not set on the listening socket.
998 * This avoids inheriting a log state which was automatically set.
999 */
1000 if ((tcp_get_bblog_state(sototcpcb(lso)) != TCP_LOG_STATE_OFF) &&
1001 ((sototcpcb(lso)->t_flags2 & TF2_LOG_AUTO) == 0) &&
1002 (sototcpcb(lso)->t_lib == NULL)) {
1003 tcp_log_state_change(tp, tcp_get_bblog_state(sototcpcb(lso)));
1004 }
1005 #endif
1006 /*
1007 * Copy and activate timers.
1008 */
1009 tp->t_maxunacktime = sototcpcb(lso)->t_maxunacktime;
1010 tp->t_keepinit = sototcpcb(lso)->t_keepinit;
1011 tp->t_keepidle = sototcpcb(lso)->t_keepidle;
1012 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
1013 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
1014 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
1015
1016 TCPSTAT_INC(tcps_accepts);
1017 TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, TCPS_LISTEN);
1018
1019 if (!solisten_enqueue(so, SS_ISCONNECTED))
1020 tp->t_flags |= TF_SONOTCONN;
1021 /* Can we inherit anything from the listener? */
1022 if (tp->t_fb->tfb_inherit != NULL) {
1023 (*tp->t_fb->tfb_inherit)(tp, sotoinpcb(lso));
1024 }
1025 return (so);
1026
1027 allocfail:
1028 /*
1029 * Drop the connection; we will either send a RST or have the peer
1030 * retransmit its SYN again after its RTO and try again.
1031 */
1032 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1033 log(LOG_DEBUG, "%s; %s: Socket create failed "
1034 "due to limits or memory shortage\n",
1035 s, __func__);
1036 free(s, M_TCPLOG);
1037 }
1038 TCPSTAT_INC(tcps_listendrop);
1039 return (NULL);
1040
1041 abort:
1042 tcp_discardcb(tp);
1043 in_pcbfree(inp);
1044 sodealloc(so);
1045 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
1046 log(LOG_DEBUG, "%s; %s: in%s_pcbconnect failed with error %i\n",
1047 s, __func__, (sc->sc_inc.inc_flags & INC_ISIPV6) ? "6" : "",
1048 error);
1049 free(s, M_TCPLOG);
1050 }
1051 TCPSTAT_INC(tcps_listendrop);
1052 return (NULL);
1053 }
1054
1055 /*
1056 * This function gets called when we receive an ACK for a
1057 * socket in the LISTEN state. We look up the connection
1058 * in the syncache, and if its there, we pull it out of
1059 * the cache and turn it into a full-blown connection in
1060 * the SYN-RECEIVED state.
1061 *
1062 * On syncache_socket() success the newly created socket
1063 * has its underlying inp locked.
1064 *
1065 * *lsop is updated, if and only if 1 is returned.
1066 */
1067 int
syncache_expand(struct in_conninfo * inc,struct tcpopt * to,struct tcphdr * th,struct socket ** lsop,struct mbuf * m,uint16_t port)1068 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1069 struct socket **lsop, struct mbuf *m, uint16_t port)
1070 {
1071 struct syncache *sc;
1072 struct syncache_head *sch;
1073 struct syncache scs;
1074 char *s;
1075 bool locked;
1076
1077 NET_EPOCH_ASSERT();
1078 KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1079 ("%s: can handle only ACK", __func__));
1080
1081 if (syncache_cookiesonly()) {
1082 sc = NULL;
1083 sch = syncache_hashbucket(inc);
1084 locked = false;
1085 } else {
1086 sc = syncache_lookup(inc, &sch); /* returns locked sch */
1087 locked = true;
1088 SCH_LOCK_ASSERT(sch);
1089 }
1090
1091 #ifdef INVARIANTS
1092 /*
1093 * Test code for syncookies comparing the syncache stored
1094 * values with the reconstructed values from the cookie.
1095 */
1096 if (sc != NULL)
1097 syncookie_cmp(inc, sch, sc, th, to, *lsop, port);
1098 #endif
1099
1100 if (sc == NULL) {
1101 if (locked) {
1102 /*
1103 * The syncache is currently in use (neither disabled,
1104 * nor paused), but no entry was found.
1105 */
1106 if (!V_tcp_syncookies) {
1107 /*
1108 * Since no syncookies are used in case of
1109 * a bucket overflow, don't even check for
1110 * a valid syncookie.
1111 */
1112 SCH_UNLOCK(sch);
1113 TCPSTAT_INC(tcps_sc_spurcookie);
1114 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1115 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1116 "segment rejected "
1117 "(syncookies disabled)\n",
1118 s, __func__);
1119 free(s, M_TCPLOG);
1120 }
1121 return (0);
1122 }
1123 if (sch->sch_last_overflow <
1124 time_uptime - SYNCOOKIE_LIFETIME) {
1125 /*
1126 * Since the bucket did not overflow recently,
1127 * don't even check for a valid syncookie.
1128 */
1129 SCH_UNLOCK(sch);
1130 TCPSTAT_INC(tcps_sc_spurcookie);
1131 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1132 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1133 "segment rejected "
1134 "(no syncache entry)\n",
1135 s, __func__);
1136 free(s, M_TCPLOG);
1137 }
1138 return (0);
1139 }
1140 SCH_UNLOCK(sch);
1141 }
1142 bzero(&scs, sizeof(scs));
1143 /*
1144 * Now check, if the syncookie is valid. If it is, create an on
1145 * stack syncache entry.
1146 */
1147 if (syncookie_expand(inc, sch, &scs, th, to, *lsop, port)) {
1148 sc = &scs;
1149 TCPSTAT_INC(tcps_sc_recvcookie);
1150 } else {
1151 TCPSTAT_INC(tcps_sc_failcookie);
1152 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1153 log(LOG_DEBUG, "%s; %s: Segment failed "
1154 "SYNCOOKIE authentication, segment rejected "
1155 "(probably spoofed)\n", s, __func__);
1156 free(s, M_TCPLOG);
1157 }
1158 return (0);
1159 }
1160 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1161 /* If received ACK has MD5 signature, check it. */
1162 if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1163 (!TCPMD5_ENABLED() ||
1164 TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1165 /* Drop the ACK. */
1166 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1167 log(LOG_DEBUG, "%s; %s: Segment rejected, "
1168 "MD5 signature doesn't match.\n",
1169 s, __func__);
1170 free(s, M_TCPLOG);
1171 }
1172 TCPSTAT_INC(tcps_sig_err_sigopt);
1173 return (-1); /* Do not send RST */
1174 }
1175 #endif /* TCP_SIGNATURE */
1176 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
1177 sc->sc_flowid = m->m_pkthdr.flowid;
1178 sc->sc_flowtype = M_HASHTYPE_GET(m);
1179 }
1180 #ifdef NUMA
1181 sc->sc_numa_domain = m ? m->m_pkthdr.numa_domain : M_NODOM;
1182 #endif
1183 TCPSTATES_INC(TCPS_SYN_RECEIVED);
1184 } else {
1185 if (sc->sc_port != port) {
1186 SCH_UNLOCK(sch);
1187 return (0);
1188 }
1189 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1190 /*
1191 * If listening socket requested TCP digests, check that
1192 * received ACK has signature and it is correct.
1193 * If not, drop the ACK and leave sc entry in the cache,
1194 * because SYN was received with correct signature.
1195 */
1196 if (sc->sc_flags & SCF_SIGNATURE) {
1197 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1198 /* No signature */
1199 TCPSTAT_INC(tcps_sig_err_nosigopt);
1200 SCH_UNLOCK(sch);
1201 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1202 log(LOG_DEBUG, "%s; %s: Segment "
1203 "rejected, MD5 signature wasn't "
1204 "provided.\n", s, __func__);
1205 free(s, M_TCPLOG);
1206 }
1207 return (-1); /* Do not send RST */
1208 }
1209 if (!TCPMD5_ENABLED() ||
1210 TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1211 /* Doesn't match or no SA */
1212 SCH_UNLOCK(sch);
1213 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1214 log(LOG_DEBUG, "%s; %s: Segment "
1215 "rejected, MD5 signature doesn't "
1216 "match.\n", s, __func__);
1217 free(s, M_TCPLOG);
1218 }
1219 return (-1); /* Do not send RST */
1220 }
1221 }
1222 #endif /* TCP_SIGNATURE */
1223
1224 /*
1225 * RFC 7323 PAWS: If we have a timestamp on this segment and
1226 * it's less than ts_recent, drop it.
1227 * XXXMT: RFC 7323 also requires to send an ACK.
1228 * In tcp_input.c this is only done for TCP segments
1229 * with user data, so be consistent here and just drop
1230 * the segment.
1231 */
1232 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS &&
1233 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) {
1234 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1235 log(LOG_DEBUG,
1236 "%s; %s: SEG.TSval %u < TS.Recent %u, "
1237 "segment dropped\n", s, __func__,
1238 to->to_tsval, sc->sc_tsreflect);
1239 }
1240 SCH_UNLOCK(sch);
1241 free(s, M_TCPLOG);
1242 return (-1); /* Do not send RST */
1243 }
1244
1245 /*
1246 * If timestamps were not negotiated during SYN/ACK and a
1247 * segment with a timestamp is received, ignore the
1248 * timestamp and process the packet normally.
1249 * See section 3.2 of RFC 7323.
1250 */
1251 if (!(sc->sc_flags & SCF_TIMESTAMP) &&
1252 (to->to_flags & TOF_TS)) {
1253 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1254 log(LOG_DEBUG, "%s; %s: Timestamp not "
1255 "expected, segment processed normally\n",
1256 s, __func__);
1257 free(s, M_TCPLOG);
1258 }
1259 }
1260
1261 /*
1262 * If timestamps were negotiated during SYN/ACK and a
1263 * segment without a timestamp is received, silently drop
1264 * the segment, unless the missing timestamps are tolerated.
1265 * See section 3.2 of RFC 7323.
1266 */
1267 if ((sc->sc_flags & SCF_TIMESTAMP) &&
1268 !(to->to_flags & TOF_TS)) {
1269 if (V_tcp_tolerate_missing_ts) {
1270 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1271 log(LOG_DEBUG,
1272 "%s; %s: Timestamp missing, "
1273 "segment processed normally\n",
1274 s, __func__);
1275 free(s, M_TCPLOG);
1276 }
1277 } else {
1278 SCH_UNLOCK(sch);
1279 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1280 log(LOG_DEBUG,
1281 "%s; %s: Timestamp missing, "
1282 "segment silently dropped\n",
1283 s, __func__);
1284 free(s, M_TCPLOG);
1285 }
1286 return (-1); /* Do not send RST */
1287 }
1288 }
1289
1290 /*
1291 * SEG.SEQ validation:
1292 * The SEG.SEQ must be in the window starting at our
1293 * initial receive sequence number + 1.
1294 */
1295 if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1296 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1297 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1298 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, "
1299 "sending challenge ACK\n",
1300 s, __func__, th->th_seq, sc->sc_irs + 1);
1301 syncache_send_challenge_ack(sc);
1302 SCH_UNLOCK(sch);
1303 free(s, M_TCPLOG);
1304 return (-1); /* Do not send RST */
1305 }
1306
1307 /*
1308 * SEG.ACK validation:
1309 * SEG.ACK must match our initial send sequence number + 1.
1310 */
1311 if (th->th_ack != sc->sc_iss + 1) {
1312 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1313 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, "
1314 "segment rejected\n",
1315 s, __func__, th->th_ack, sc->sc_iss + 1);
1316 SCH_UNLOCK(sch);
1317 free(s, M_TCPLOG);
1318 return (0); /* Do send RST, do not free sc. */
1319 }
1320
1321 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1322 sch->sch_length--;
1323 #ifdef TCP_OFFLOAD
1324 if (ADDED_BY_TOE(sc)) {
1325 struct toedev *tod = sc->sc_tod;
1326
1327 tod->tod_syncache_removed(tod, sc->sc_todctx);
1328 }
1329 #endif
1330 SCH_UNLOCK(sch);
1331 }
1332
1333 *lsop = syncache_socket(sc, *lsop, m);
1334
1335 if (__predict_false(*lsop == NULL)) {
1336 TCPSTAT_INC(tcps_sc_aborted);
1337 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1338 } else if (sc != &scs)
1339 TCPSTAT_INC(tcps_sc_completed);
1340
1341 if (sc != &scs)
1342 syncache_free(sc);
1343 return (1);
1344 }
1345
1346 static struct socket *
syncache_tfo_expand(struct syncache * sc,struct socket * lso,struct mbuf * m,uint64_t response_cookie)1347 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m,
1348 uint64_t response_cookie)
1349 {
1350 struct inpcb *inp;
1351 struct tcpcb *tp;
1352 unsigned int *pending_counter;
1353 struct socket *so;
1354
1355 NET_EPOCH_ASSERT();
1356
1357 pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending;
1358 so = syncache_socket(sc, lso, m);
1359 if (so == NULL) {
1360 TCPSTAT_INC(tcps_sc_aborted);
1361 atomic_subtract_int(pending_counter, 1);
1362 } else {
1363 soisconnected(so);
1364 inp = sotoinpcb(so);
1365 tp = intotcpcb(inp);
1366 tp->t_flags |= TF_FASTOPEN;
1367 tp->t_tfo_cookie.server = response_cookie;
1368 tp->snd_max = tp->iss;
1369 tp->snd_nxt = tp->iss;
1370 tp->t_tfo_pending = pending_counter;
1371 TCPSTATES_INC(TCPS_SYN_RECEIVED);
1372 TCPSTAT_INC(tcps_sc_completed);
1373 }
1374
1375 return (so);
1376 }
1377
1378 /*
1379 * Given a LISTEN socket and an inbound SYN request, add
1380 * this to the syn cache, and send back a segment:
1381 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1382 * to the source.
1383 *
1384 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1385 * Doing so would require that we hold onto the data and deliver it
1386 * to the application. However, if we are the target of a SYN-flood
1387 * DoS attack, an attacker could send data which would eventually
1388 * consume all available buffer space if it were ACKed. By not ACKing
1389 * the data, we avoid this DoS scenario.
1390 *
1391 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1392 * cookie is processed and a new socket is created. In this case, any data
1393 * accompanying the SYN will be queued to the socket by tcp_input() and will
1394 * be ACKed either when the application sends response data or the delayed
1395 * ACK timer expires, whichever comes first.
1396 */
1397 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)1398 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1399 struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod,
1400 void *todctx, uint8_t iptos, uint16_t port)
1401 {
1402 struct tcpcb *tp;
1403 struct socket *rv = NULL;
1404 struct syncache *sc = NULL;
1405 struct ucred *cred;
1406 struct syncache_head *sch;
1407 struct mbuf *ipopts = NULL;
1408 u_int ltflags;
1409 int win, ip_ttl, ip_tos;
1410 char *s;
1411 #ifdef INET6
1412 int autoflowlabel = 0;
1413 #endif
1414 #ifdef MAC
1415 struct label *maclabel = NULL;
1416 #endif
1417 struct syncache scs;
1418 uint64_t tfo_response_cookie;
1419 unsigned int *tfo_pending = NULL;
1420 int tfo_cookie_valid = 0;
1421 int tfo_response_cookie_valid = 0;
1422 bool locked;
1423
1424 INP_RLOCK_ASSERT(inp); /* listen socket */
1425 KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1426 ("%s: unexpected tcp flags", __func__));
1427
1428 /*
1429 * Combine all so/tp operations very early to drop the INP lock as
1430 * soon as possible.
1431 */
1432 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1433 tp = sototcpcb(so);
1434 cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred);
1435
1436 #ifdef INET6
1437 if (inc->inc_flags & INC_ISIPV6) {
1438 if (inp->inp_flags & IN6P_AUTOFLOWLABEL) {
1439 autoflowlabel = 1;
1440 }
1441 ip_ttl = in6_selecthlim(inp, NULL);
1442 if ((inp->in6p_outputopts == NULL) ||
1443 (inp->in6p_outputopts->ip6po_tclass == -1)) {
1444 ip_tos = 0;
1445 } else {
1446 ip_tos = inp->in6p_outputopts->ip6po_tclass;
1447 }
1448 }
1449 #endif
1450 #if defined(INET6) && defined(INET)
1451 else
1452 #endif
1453 #ifdef INET
1454 {
1455 ip_ttl = inp->inp_ip_ttl;
1456 ip_tos = inp->inp_ip_tos;
1457 }
1458 #endif
1459 win = so->sol_sbrcv_hiwat;
1460 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1461
1462 if (V_tcp_fastopen_server_enable && (tp->t_flags & TF_FASTOPEN) &&
1463 (tp->t_tfo_pending != NULL) &&
1464 (to->to_flags & TOF_FASTOPEN)) {
1465 /*
1466 * Limit the number of pending TFO connections to
1467 * approximately half of the queue limit. This prevents TFO
1468 * SYN floods from starving the service by filling the
1469 * listen queue with bogus TFO connections.
1470 */
1471 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1472 (so->sol_qlimit / 2)) {
1473 int result;
1474
1475 result = tcp_fastopen_check_cookie(inc,
1476 to->to_tfo_cookie, to->to_tfo_len,
1477 &tfo_response_cookie);
1478 tfo_cookie_valid = (result > 0);
1479 tfo_response_cookie_valid = (result >= 0);
1480 }
1481
1482 /*
1483 * Remember the TFO pending counter as it will have to be
1484 * decremented below if we don't make it to syncache_tfo_expand().
1485 */
1486 tfo_pending = tp->t_tfo_pending;
1487 }
1488
1489 #ifdef MAC
1490 if (mac_syncache_init(&maclabel) != 0) {
1491 INP_RUNLOCK(inp);
1492 goto done;
1493 } else
1494 mac_syncache_create(maclabel, inp);
1495 #endif
1496 if (!tfo_cookie_valid)
1497 INP_RUNLOCK(inp);
1498
1499 /*
1500 * Remember the IP options, if any.
1501 */
1502 #ifdef INET6
1503 if (!(inc->inc_flags & INC_ISIPV6))
1504 #endif
1505 #ifdef INET
1506 ipopts = (m) ? ip_srcroute(m) : NULL;
1507 #else
1508 ipopts = NULL;
1509 #endif
1510
1511 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1512 /*
1513 * When the socket is TCP-MD5 enabled check that,
1514 * - a signed packet is valid
1515 * - a non-signed packet does not have a security association
1516 *
1517 * If a signed packet fails validation or a non-signed packet has a
1518 * security association, the packet will be dropped.
1519 */
1520 if (ltflags & TF_SIGNATURE) {
1521 if (to->to_flags & TOF_SIGNATURE) {
1522 if (!TCPMD5_ENABLED() ||
1523 TCPMD5_INPUT(m, th, to->to_signature) != 0)
1524 goto done;
1525 } else {
1526 if (TCPMD5_ENABLED() &&
1527 TCPMD5_INPUT(m, NULL, NULL) != ENOENT)
1528 goto done;
1529 }
1530 } else if (to->to_flags & TOF_SIGNATURE)
1531 goto done;
1532 #endif /* TCP_SIGNATURE */
1533 /*
1534 * See if we already have an entry for this connection.
1535 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1536 *
1537 * XXX: should the syncache be re-initialized with the contents
1538 * of the new SYN here (which may have different options?)
1539 *
1540 * XXX: We do not check the sequence number to see if this is a
1541 * real retransmit or a new connection attempt. The question is
1542 * how to handle such a case; either ignore it as spoofed, or
1543 * drop the current entry and create a new one?
1544 */
1545 if (syncache_cookiesonly()) {
1546 sc = NULL;
1547 sch = syncache_hashbucket(inc);
1548 locked = false;
1549 } else {
1550 sc = syncache_lookup(inc, &sch); /* returns locked sch */
1551 locked = true;
1552 SCH_LOCK_ASSERT(sch);
1553 }
1554 if (sc != NULL) {
1555 if (tfo_cookie_valid)
1556 INP_RUNLOCK(inp);
1557 TCPSTAT_INC(tcps_sc_dupsyn);
1558 if (ipopts) {
1559 /*
1560 * If we were remembering a previous source route,
1561 * forget it and use the new one we've been given.
1562 */
1563 if (sc->sc_ipopts)
1564 (void)m_free(sc->sc_ipopts);
1565 sc->sc_ipopts = ipopts;
1566 }
1567 /*
1568 * Update timestamp if present.
1569 */
1570 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1571 sc->sc_tsreflect = to->to_tsval;
1572 else
1573 sc->sc_flags &= ~SCF_TIMESTAMP;
1574 /*
1575 * Adjust ECN response if needed, e.g. different
1576 * IP ECN field, or a fallback by the remote host.
1577 */
1578 if (sc->sc_flags & SCF_ECN_MASK) {
1579 sc->sc_flags &= ~SCF_ECN_MASK;
1580 sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1581 }
1582 #ifdef MAC
1583 /*
1584 * Since we have already unconditionally allocated label
1585 * storage, free it up. The syncache entry will already
1586 * have an initialized label we can use.
1587 */
1588 mac_syncache_destroy(&maclabel);
1589 #endif
1590 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1591 /* Retransmit SYN|ACK and reset retransmit count. */
1592 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1593 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1594 "resetting timer and retransmitting SYN|ACK\n",
1595 s, __func__);
1596 free(s, M_TCPLOG);
1597 }
1598 if (syncache_respond(sc, TH_SYN|TH_ACK) == 0) {
1599 sc->sc_rxmits = 0;
1600 syncache_timeout(sc, sch, 1);
1601 TCPSTAT_INC(tcps_sndacks);
1602 TCPSTAT_INC(tcps_sndtotal);
1603 } else {
1604 /*
1605 * Most likely we are memory constrained, so free
1606 * resources.
1607 */
1608 syncache_drop(sc, sch);
1609 TCPSTAT_INC(tcps_sc_dropped);
1610 }
1611 SCH_UNLOCK(sch);
1612 goto donenoprobe;
1613 }
1614
1615 KASSERT(sc == NULL, ("sc(%p) != NULL", sc));
1616 /*
1617 * Skip allocating a syncache entry if we are just going to discard
1618 * it later.
1619 */
1620 if (!locked || tfo_cookie_valid) {
1621 bzero(&scs, sizeof(scs));
1622 sc = &scs;
1623 } else {
1624 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1625 if (sc == NULL) {
1626 /*
1627 * The zone allocator couldn't provide more entries.
1628 * Treat this as if the cache was full; drop the oldest
1629 * entry and insert the new one.
1630 */
1631 TCPSTAT_INC(tcps_sc_zonefail);
1632 sc = TAILQ_LAST(&sch->sch_bucket, sch_head);
1633 if (sc != NULL) {
1634 sch->sch_last_overflow = time_uptime;
1635 syncache_drop(sc, sch);
1636 syncache_pause(inc);
1637 }
1638 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1639 if (sc == NULL) {
1640 if (V_tcp_syncookies) {
1641 bzero(&scs, sizeof(scs));
1642 sc = &scs;
1643 } else {
1644 KASSERT(locked,
1645 ("%s: bucket unexpectedly unlocked",
1646 __func__));
1647 SCH_UNLOCK(sch);
1648 goto done;
1649 }
1650 }
1651 }
1652 }
1653
1654 KASSERT(sc != NULL, ("sc == NULL"));
1655 if (!tfo_cookie_valid && tfo_response_cookie_valid)
1656 sc->sc_tfo_cookie = &tfo_response_cookie;
1657
1658 /*
1659 * Fill in the syncache values.
1660 */
1661 #ifdef MAC
1662 sc->sc_label = maclabel;
1663 #endif
1664 /*
1665 * sc_cred is only used in syncache_pcblist() to list TCP endpoints in
1666 * TCPS_SYN_RECEIVED state when V_tcp_syncache.see_other is false.
1667 * Therefore, store the credentials only when needed:
1668 * - sc is allocated from the zone and not using the on stack instance.
1669 * - the sysctl variable net.inet.tcp.syncache.see_other is false.
1670 * The reference count is decremented when a zone allocated sc is
1671 * freed in syncache_free().
1672 */
1673 if (sc != &scs && !V_tcp_syncache.see_other) {
1674 sc->sc_cred = cred;
1675 cred = NULL;
1676 } else
1677 sc->sc_cred = NULL;
1678 sc->sc_port = port;
1679 sc->sc_ipopts = ipopts;
1680 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1681 sc->sc_ip_tos = ip_tos;
1682 sc->sc_ip_ttl = ip_ttl;
1683 #ifdef TCP_OFFLOAD
1684 sc->sc_tod = tod;
1685 sc->sc_todctx = todctx;
1686 #endif
1687 sc->sc_irs = th->th_seq;
1688 sc->sc_flags = 0;
1689 sc->sc_flowlabel = 0;
1690
1691 /*
1692 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1693 * win was derived from socket earlier in the function.
1694 */
1695 win = imax(win, 0);
1696 win = imin(win, TCP_MAXWIN);
1697 sc->sc_wnd = win;
1698
1699 if (V_tcp_do_rfc1323 &&
1700 !(ltflags & TF_NOOPT)) {
1701 /*
1702 * A timestamp received in a SYN makes
1703 * it ok to send timestamp requests and replies.
1704 */
1705 if ((to->to_flags & TOF_TS) && (V_tcp_do_rfc1323 != 2)) {
1706 sc->sc_tsreflect = to->to_tsval;
1707 sc->sc_flags |= SCF_TIMESTAMP;
1708 sc->sc_tsoff = tcp_new_ts_offset(inc);
1709 }
1710 if ((to->to_flags & TOF_SCALE) && (V_tcp_do_rfc1323 != 3)) {
1711 u_int wscale = 0;
1712
1713 /*
1714 * Pick the smallest possible scaling factor that
1715 * will still allow us to scale up to sb_max, aka
1716 * kern.ipc.maxsockbuf.
1717 *
1718 * We do this because there are broken firewalls that
1719 * will corrupt the window scale option, leading to
1720 * the other endpoint believing that our advertised
1721 * window is unscaled. At scale factors larger than
1722 * 5 the unscaled window will drop below 1500 bytes,
1723 * leading to serious problems when traversing these
1724 * broken firewalls.
1725 *
1726 * With the default maxsockbuf of 256K, a scale factor
1727 * of 3 will be chosen by this algorithm. Those who
1728 * choose a larger maxsockbuf should watch out
1729 * for the compatibility problems mentioned above.
1730 *
1731 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1732 * or <SYN,ACK>) segment itself is never scaled.
1733 */
1734 while (wscale < TCP_MAX_WINSHIFT &&
1735 (TCP_MAXWIN << wscale) < sb_max)
1736 wscale++;
1737 sc->sc_requested_r_scale = wscale;
1738 sc->sc_requested_s_scale = to->to_wscale;
1739 sc->sc_flags |= SCF_WINSCALE;
1740 }
1741 }
1742 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1743 /*
1744 * If incoming packet has an MD5 signature, flag this in the
1745 * syncache so that syncache_respond() will do the right thing
1746 * with the SYN+ACK.
1747 */
1748 if (to->to_flags & TOF_SIGNATURE)
1749 sc->sc_flags |= SCF_SIGNATURE;
1750 #endif /* TCP_SIGNATURE */
1751 if (to->to_flags & TOF_SACKPERM)
1752 sc->sc_flags |= SCF_SACK;
1753 if (to->to_flags & TOF_MSS)
1754 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */
1755 if (ltflags & TF_NOOPT)
1756 sc->sc_flags |= SCF_NOOPT;
1757 /* ECN Handshake */
1758 if (V_tcp_do_ecn && (tp->t_flags2 & TF2_CANNOT_DO_ECN) == 0)
1759 sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos);
1760
1761 if (V_tcp_syncookies || V_tcp_syncookiesonly)
1762 sc->sc_iss = syncookie_generate(sch, sc);
1763 else
1764 sc->sc_iss = arc4random();
1765 #ifdef INET6
1766 if (autoflowlabel) {
1767 if (V_tcp_syncookies || V_tcp_syncookiesonly)
1768 sc->sc_flowlabel = sc->sc_iss;
1769 else
1770 sc->sc_flowlabel = ip6_randomflowlabel();
1771 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1772 }
1773 #endif
1774 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
1775 sc->sc_flowid = m->m_pkthdr.flowid;
1776 sc->sc_flowtype = M_HASHTYPE_GET(m);
1777 }
1778 #ifdef NUMA
1779 sc->sc_numa_domain = m ? m->m_pkthdr.numa_domain : M_NODOM;
1780 #endif
1781 if (locked)
1782 SCH_UNLOCK(sch);
1783
1784 if (tfo_cookie_valid) {
1785 rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie);
1786 /* INP_RUNLOCK(inp) will be performed by the caller */
1787 goto tfo_expanded;
1788 }
1789
1790 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1791 /*
1792 * Do a standard 3-way handshake.
1793 */
1794 if (syncache_respond(sc, TH_SYN|TH_ACK) == 0) {
1795 if (sc != &scs)
1796 syncache_insert(sc, sch); /* locks and unlocks sch */
1797 TCPSTAT_INC(tcps_sndacks);
1798 TCPSTAT_INC(tcps_sndtotal);
1799 } else {
1800 /*
1801 * Most likely we are memory constrained, so free resources.
1802 */
1803 if (sc != &scs)
1804 syncache_free(sc);
1805 TCPSTAT_INC(tcps_sc_dropped);
1806 }
1807 goto donenoprobe;
1808
1809 done:
1810 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1811 donenoprobe:
1812 if (m)
1813 m_freem(m);
1814 /*
1815 * If tfo_pending is not NULL here, then a TFO SYN that did not
1816 * result in a new socket was processed and the associated pending
1817 * counter has not yet been decremented. All such TFO processing paths
1818 * transit this point.
1819 */
1820 if (tfo_pending != NULL)
1821 tcp_fastopen_decrement_counter(tfo_pending);
1822
1823 tfo_expanded:
1824 if (cred != NULL)
1825 crfree(cred);
1826 if (sc == NULL || sc == &scs) {
1827 #ifdef MAC
1828 mac_syncache_destroy(&maclabel);
1829 #endif
1830 if (ipopts)
1831 (void)m_free(ipopts);
1832 }
1833 return (rv);
1834 }
1835
1836 /*
1837 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment
1838 * or upon 3WHS ACK timeout.
1839 */
1840 static int
syncache_respond(struct syncache * sc,int flags)1841 syncache_respond(struct syncache *sc, int flags)
1842 {
1843 struct ip *ip = NULL;
1844 struct mbuf *m;
1845 struct tcphdr *th = NULL;
1846 struct udphdr *udp = NULL;
1847 int optlen, error = 0; /* Make compiler happy */
1848 u_int16_t hlen, tlen, mssopt, ulen;
1849 struct tcpopt to;
1850 #ifdef INET6
1851 struct ip6_hdr *ip6 = NULL;
1852 #endif
1853
1854 NET_EPOCH_ASSERT();
1855
1856 hlen =
1857 #ifdef INET6
1858 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1859 #endif
1860 sizeof(struct ip);
1861 tlen = hlen + sizeof(struct tcphdr);
1862 if (sc->sc_port) {
1863 tlen += sizeof(struct udphdr);
1864 }
1865 /* Determine MSS we advertize to other end of connection. */
1866 mssopt = tcp_mssopt(&sc->sc_inc);
1867 if (sc->sc_port)
1868 mssopt -= V_tcp_udp_tunneling_overhead;
1869 mssopt = max(mssopt, V_tcp_minmss);
1870
1871 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1872 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1873 ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + "
1874 "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port,
1875 max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN));
1876
1877 /* Create the IP+TCP header from scratch. */
1878 m = m_gethdr(M_NOWAIT, MT_DATA);
1879 if (m == NULL)
1880 return (ENOBUFS);
1881 #ifdef MAC
1882 mac_syncache_create_mbuf(sc->sc_label, m);
1883 #endif
1884 m->m_data += max_linkhdr;
1885 m->m_len = tlen;
1886 m->m_pkthdr.len = tlen;
1887 m->m_pkthdr.rcvif = NULL;
1888
1889 #ifdef INET6
1890 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1891 ip6 = mtod(m, struct ip6_hdr *);
1892 ip6->ip6_vfc = IPV6_VERSION;
1893 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1894 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1895 ip6->ip6_plen = htons(tlen - hlen);
1896 /* ip6_hlim is set after checksum */
1897 /* Zero out traffic class and flow label. */
1898 ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK;
1899 ip6->ip6_flow |= sc->sc_flowlabel;
1900 if (sc->sc_port != 0) {
1901 ip6->ip6_nxt = IPPROTO_UDP;
1902 udp = (struct udphdr *)(ip6 + 1);
1903 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1904 udp->uh_dport = sc->sc_port;
1905 ulen = (tlen - sizeof(struct ip6_hdr));
1906 th = (struct tcphdr *)(udp + 1);
1907 } else {
1908 ip6->ip6_nxt = IPPROTO_TCP;
1909 th = (struct tcphdr *)(ip6 + 1);
1910 }
1911 ip6->ip6_flow |= htonl(sc->sc_ip_tos << IPV6_FLOWLABEL_LEN);
1912 }
1913 #endif
1914 #if defined(INET6) && defined(INET)
1915 else
1916 #endif
1917 #ifdef INET
1918 {
1919 ip = mtod(m, struct ip *);
1920 ip->ip_v = IPVERSION;
1921 ip->ip_hl = sizeof(struct ip) >> 2;
1922 ip->ip_len = htons(tlen);
1923 ip->ip_id = 0;
1924 ip->ip_off = 0;
1925 ip->ip_sum = 0;
1926 ip->ip_src = sc->sc_inc.inc_laddr;
1927 ip->ip_dst = sc->sc_inc.inc_faddr;
1928 ip->ip_ttl = sc->sc_ip_ttl;
1929 ip->ip_tos = sc->sc_ip_tos;
1930
1931 /*
1932 * See if we should do MTU discovery. Route lookups are
1933 * expensive, so we will only unset the DF bit if:
1934 *
1935 * 1) path_mtu_discovery is disabled
1936 * 2) the SCF_UNREACH flag has been set
1937 */
1938 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1939 ip->ip_off |= htons(IP_DF);
1940 if (sc->sc_port == 0) {
1941 ip->ip_p = IPPROTO_TCP;
1942 th = (struct tcphdr *)(ip + 1);
1943 } else {
1944 ip->ip_p = IPPROTO_UDP;
1945 udp = (struct udphdr *)(ip + 1);
1946 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1947 udp->uh_dport = sc->sc_port;
1948 ulen = (tlen - sizeof(struct ip));
1949 th = (struct tcphdr *)(udp + 1);
1950 }
1951 }
1952 #endif /* INET */
1953 th->th_sport = sc->sc_inc.inc_lport;
1954 th->th_dport = sc->sc_inc.inc_fport;
1955
1956 if (flags & TH_SYN)
1957 th->th_seq = htonl(sc->sc_iss);
1958 else
1959 th->th_seq = htonl(sc->sc_iss + 1);
1960 th->th_ack = htonl(sc->sc_irs + 1);
1961 th->th_off = sizeof(struct tcphdr) >> 2;
1962 th->th_win = htons(sc->sc_wnd);
1963 th->th_urp = 0;
1964
1965 flags = tcp_ecn_syncache_respond(flags, sc);
1966 tcp_set_flags(th, flags);
1967
1968 /* Tack on the TCP options. */
1969 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1970 to.to_flags = 0;
1971
1972 if (flags & TH_SYN) {
1973 to.to_mss = mssopt;
1974 to.to_flags = TOF_MSS;
1975 if (sc->sc_flags & SCF_WINSCALE) {
1976 to.to_wscale = sc->sc_requested_r_scale;
1977 to.to_flags |= TOF_SCALE;
1978 }
1979 if (sc->sc_flags & SCF_SACK)
1980 to.to_flags |= TOF_SACKPERM;
1981 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1982 if (sc->sc_flags & SCF_SIGNATURE)
1983 to.to_flags |= TOF_SIGNATURE;
1984 #endif
1985 if (sc->sc_tfo_cookie) {
1986 to.to_flags |= TOF_FASTOPEN;
1987 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1988 to.to_tfo_cookie = sc->sc_tfo_cookie;
1989 /* don't send cookie again when retransmitting response */
1990 sc->sc_tfo_cookie = NULL;
1991 }
1992 }
1993 if (sc->sc_flags & SCF_TIMESTAMP) {
1994 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1995 to.to_tsecr = sc->sc_tsreflect;
1996 to.to_flags |= TOF_TS;
1997 }
1998 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1999
2000 /* Adjust headers by option size. */
2001 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
2002 m->m_len += optlen;
2003 m->m_pkthdr.len += optlen;
2004 #ifdef INET6
2005 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2006 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
2007 else
2008 #endif
2009 ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
2010 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
2011 if (sc->sc_flags & SCF_SIGNATURE) {
2012 KASSERT(to.to_flags & TOF_SIGNATURE,
2013 ("tcp_addoptions() didn't set tcp_signature"));
2014
2015 /* NOTE: to.to_signature is inside of mbuf */
2016 if (!TCPMD5_ENABLED() ||
2017 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
2018 m_freem(m);
2019 return (EACCES);
2020 }
2021 }
2022 #endif
2023 } else
2024 optlen = 0;
2025
2026 if (udp) {
2027 ulen += optlen;
2028 udp->uh_ulen = htons(ulen);
2029 }
2030 M_SETFIB(m, sc->sc_inc.inc_fibnum);
2031 m->m_pkthdr.flowid = sc->sc_flowid;
2032 M_HASHTYPE_SET(m, sc->sc_flowtype);
2033 #ifdef NUMA
2034 m->m_pkthdr.numa_domain = sc->sc_numa_domain;
2035 #endif
2036 #ifdef INET6
2037 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
2038 if (sc->sc_port) {
2039 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
2040 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2041 udp->uh_sum = in6_cksum_pseudo(ip6, ulen,
2042 IPPROTO_UDP, 0);
2043 th->th_sum = htons(0);
2044 } else {
2045 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
2046 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2047 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
2048 IPPROTO_TCP, 0);
2049 }
2050 ip6->ip6_hlim = sc->sc_ip_ttl;
2051 #ifdef TCP_OFFLOAD
2052 if (ADDED_BY_TOE(sc)) {
2053 struct toedev *tod = sc->sc_tod;
2054
2055 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2056
2057 return (error);
2058 }
2059 #endif
2060 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
2061 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
2062 }
2063 #endif
2064 #if defined(INET6) && defined(INET)
2065 else
2066 #endif
2067 #ifdef INET
2068 {
2069 if (sc->sc_port) {
2070 m->m_pkthdr.csum_flags = CSUM_UDP;
2071 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2072 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
2073 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
2074 th->th_sum = htons(0);
2075 } else {
2076 m->m_pkthdr.csum_flags = CSUM_TCP;
2077 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2078 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
2079 htons(tlen + optlen - hlen + IPPROTO_TCP));
2080 }
2081 #ifdef TCP_OFFLOAD
2082 if (ADDED_BY_TOE(sc)) {
2083 struct toedev *tod = sc->sc_tod;
2084
2085 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
2086
2087 return (error);
2088 }
2089 #endif
2090 TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
2091 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
2092 }
2093 #endif
2094 return (error);
2095 }
2096
2097 static void
syncache_send_challenge_ack(struct syncache * sc)2098 syncache_send_challenge_ack(struct syncache *sc)
2099 {
2100 if (tcp_challenge_ack_check(&sc->sc_challenge_ack_end,
2101 &sc->sc_challenge_ack_cnt)) {
2102 if (syncache_respond(sc, TH_ACK) == 0) {
2103 TCPSTAT_INC(tcps_sndacks);
2104 TCPSTAT_INC(tcps_sndtotal);
2105 }
2106 }
2107 }
2108
2109 /*
2110 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
2111 * that exceed the capacity of the syncache by avoiding the storage of any
2112 * of the SYNs we receive. Syncookies defend against blind SYN flooding
2113 * attacks where the attacker does not have access to our responses.
2114 *
2115 * Syncookies encode and include all necessary information about the
2116 * connection setup within the SYN|ACK that we send back. That way we
2117 * can avoid keeping any local state until the ACK to our SYN|ACK returns
2118 * (if ever). Normally the syncache and syncookies are running in parallel
2119 * with the latter taking over when the former is exhausted. When matching
2120 * syncache entry is found the syncookie is ignored.
2121 *
2122 * The only reliable information persisting the 3WHS is our initial sequence
2123 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient
2124 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
2125 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK
2126 * returns and signifies a legitimate connection if it matches the ACK.
2127 *
2128 * The available space of 32 bits to store the hash and to encode the SYN
2129 * option information is very tight and we should have at least 24 bits for
2130 * the MAC to keep the number of guesses by blind spoofing reasonably high.
2131 *
2132 * SYN option information we have to encode to fully restore a connection:
2133 * MSS: is imporant to chose an optimal segment size to avoid IP level
2134 * fragmentation along the path. The common MSS values can be encoded
2135 * in a 3-bit table. Uncommon values are captured by the next lower value
2136 * in the table leading to a slight increase in packetization overhead.
2137 * WSCALE: is necessary to allow large windows to be used for high delay-
2138 * bandwidth product links. Not scaling the window when it was initially
2139 * negotiated is bad for performance as lack of scaling further decreases
2140 * the apparent available send window. We only need to encode the WSCALE
2141 * we received from the remote end. Our end can be recalculated at any
2142 * time. The common WSCALE values can be encoded in a 3-bit table.
2143 * Uncommon values are captured by the next lower value in the table
2144 * making us under-estimate the available window size halving our
2145 * theoretically possible maximum throughput for that connection.
2146 * SACK: Greatly assists in packet loss recovery and requires 1 bit.
2147 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
2148 * that are included in all segments on a connection. We enable them when
2149 * the ACK has them.
2150 *
2151 * Security of syncookies and attack vectors:
2152 *
2153 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
2154 * together with the gloabl secret to make it unique per connection attempt.
2155 * Thus any change of any of those parameters results in a different MAC output
2156 * in an unpredictable way unless a collision is encountered. 24 bits of the
2157 * MAC are embedded into the ISS.
2158 *
2159 * To prevent replay attacks two rotating global secrets are updated with a
2160 * new random value every 15 seconds. The life-time of a syncookie is thus
2161 * 15-30 seconds.
2162 *
2163 * Vector 1: Attacking the secret. This requires finding a weakness in the
2164 * MAC itself or the way it is used here. The attacker can do a chosen plain
2165 * text attack by varying and testing the all parameters under his control.
2166 * The strength depends on the size and randomness of the secret, and the
2167 * cryptographic security of the MAC function. Due to the constant updating
2168 * of the secret the attacker has at most 29.999 seconds to find the secret
2169 * and launch spoofed connections. After that he has to start all over again.
2170 *
2171 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC
2172 * size an average of 4,823 attempts are required for a 50% chance of success
2173 * to spoof a single syncookie (birthday collision paradox). However the
2174 * attacker is blind and doesn't know if one of his attempts succeeded unless
2175 * he has a side channel to interfere success from. A single connection setup
2176 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
2177 * This many attempts are required for each one blind spoofed connection. For
2178 * every additional spoofed connection he has to launch another N attempts.
2179 * Thus for a sustained rate 100 spoofed connections per second approximately
2180 * 1,800,000 packets per second would have to be sent.
2181 *
2182 * NB: The MAC function should be fast so that it doesn't become a CPU
2183 * exhaustion attack vector itself.
2184 *
2185 * References:
2186 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations
2187 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
2188 * http://cr.yp.to/syncookies.html (overview)
2189 * http://cr.yp.to/syncookies/archive (details)
2190 *
2191 *
2192 * Schematic construction of a syncookie enabled Initial Sequence Number:
2193 * 0 1 2 3
2194 * 12345678901234567890123456789012
2195 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
2196 *
2197 * x 24 MAC (truncated)
2198 * W 3 Send Window Scale index
2199 * M 3 MSS index
2200 * S 1 SACK permitted
2201 * P 1 Odd/even secret
2202 */
2203
2204 /*
2205 * Distribution and probability of certain MSS values. Those in between are
2206 * rounded down to the next lower one.
2207 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
2208 * .2% .3% 5% 7% 7% 20% 15% 45%
2209 */
2210 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
2211
2212 /*
2213 * Distribution and probability of certain WSCALE values. We have to map the
2214 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
2215 * bits based on prevalence of certain values. Where we don't have an exact
2216 * match for are rounded down to the next lower one letting us under-estimate
2217 * the true available window. At the moment this would happen only for the
2218 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
2219 * and window size). The absence of the WSCALE option (no scaling in either
2220 * direction) is encoded with index zero.
2221 * [WSCALE values histograms, Allman, 2012]
2222 * X 10 10 35 5 6 14 10% by host
2223 * X 11 4 5 5 18 49 3% by connections
2224 */
2225 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
2226
2227 /*
2228 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed
2229 * and good cryptographic properties.
2230 */
2231 static uint32_t
syncookie_mac(struct in_conninfo * inc,tcp_seq irs,uint8_t flags,uint8_t * secbits,uintptr_t secmod)2232 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
2233 uint8_t *secbits, uintptr_t secmod)
2234 {
2235 SIPHASH_CTX ctx;
2236 uint32_t siphash[2];
2237
2238 SipHash24_Init(&ctx);
2239 SipHash_SetKey(&ctx, secbits);
2240 switch (inc->inc_flags & INC_ISIPV6) {
2241 #ifdef INET
2242 case 0:
2243 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
2244 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
2245 break;
2246 #endif
2247 #ifdef INET6
2248 case INC_ISIPV6:
2249 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
2250 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2251 break;
2252 #endif
2253 }
2254 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2255 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2256 SipHash_Update(&ctx, &irs, sizeof(irs));
2257 SipHash_Update(&ctx, &flags, sizeof(flags));
2258 SipHash_Update(&ctx, &secmod, sizeof(secmod));
2259 SipHash_Final((u_int8_t *)&siphash, &ctx);
2260
2261 return (siphash[0] ^ siphash[1]);
2262 }
2263
2264 static tcp_seq
syncookie_generate(struct syncache_head * sch,struct syncache * sc)2265 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2266 {
2267 u_int i, secbit, wscale;
2268 uint32_t iss, hash;
2269 uint8_t *secbits;
2270 union syncookie cookie;
2271
2272 cookie.cookie = 0;
2273
2274 /* Map our computed MSS into the 3-bit index. */
2275 for (i = nitems(tcp_sc_msstab) - 1;
2276 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2277 i--)
2278 ;
2279 cookie.flags.mss_idx = i;
2280
2281 /*
2282 * Map the send window scale into the 3-bit index but only if
2283 * the wscale option was received.
2284 */
2285 if (sc->sc_flags & SCF_WINSCALE) {
2286 wscale = sc->sc_requested_s_scale;
2287 for (i = nitems(tcp_sc_wstab) - 1;
2288 tcp_sc_wstab[i] > wscale && i > 0;
2289 i--)
2290 ;
2291 cookie.flags.wscale_idx = i;
2292 }
2293
2294 /* Can we do SACK? */
2295 if (sc->sc_flags & SCF_SACK)
2296 cookie.flags.sack_ok = 1;
2297
2298 /* Which of the two secrets to use. */
2299 secbit = V_tcp_syncache.secret.oddeven & 0x1;
2300 cookie.flags.odd_even = secbit;
2301
2302 secbits = V_tcp_syncache.secret.key[secbit];
2303 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2304 (uintptr_t)sch);
2305
2306 /*
2307 * Put the flags into the hash and XOR them to get better ISS number
2308 * variance. This doesn't enhance the cryptographic strength and is
2309 * done to prevent the 8 cookie bits from showing up directly on the
2310 * wire.
2311 */
2312 iss = hash & ~0xff;
2313 iss |= cookie.cookie ^ (hash >> 24);
2314
2315 TCPSTAT_INC(tcps_sc_sendcookie);
2316 return (iss);
2317 }
2318
2319 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)2320 syncookie_expand(struct in_conninfo *inc, const struct syncache_head *sch,
2321 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2322 struct socket *lso, uint16_t port)
2323 {
2324 uint32_t hash;
2325 uint8_t *secbits;
2326 tcp_seq ack, seq;
2327 int wnd;
2328 union syncookie cookie;
2329
2330 /*
2331 * Pull information out of SYN-ACK/ACK and revert sequence number
2332 * advances.
2333 */
2334 ack = th->th_ack - 1;
2335 seq = th->th_seq - 1;
2336
2337 /*
2338 * Unpack the flags containing enough information to restore the
2339 * connection.
2340 */
2341 cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2342
2343 /* Which of the two secrets to use. */
2344 secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even];
2345
2346 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2347
2348 /* The recomputed hash matches the ACK if this was a genuine cookie. */
2349 if ((ack & ~0xff) != (hash & ~0xff))
2350 return (false);
2351
2352 /* Fill in the syncache values. */
2353 sc->sc_flags = 0;
2354 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2355 sc->sc_ipopts = NULL;
2356
2357 sc->sc_irs = seq;
2358 sc->sc_iss = ack;
2359
2360 switch (inc->inc_flags & INC_ISIPV6) {
2361 #ifdef INET
2362 case 0:
2363 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2364 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2365 break;
2366 #endif
2367 #ifdef INET6
2368 case INC_ISIPV6:
2369 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2370 sc->sc_flowlabel =
2371 htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK;
2372 break;
2373 #endif
2374 }
2375
2376 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2377
2378 /* Only use wscale if it was enabled in the orignal SYN. */
2379 if (cookie.flags.wscale_idx > 0) {
2380 u_int wscale = 0;
2381
2382 /* Recompute the receive window scale that was sent earlier. */
2383 while (wscale < TCP_MAX_WINSHIFT &&
2384 (TCP_MAXWIN << wscale) < sb_max)
2385 wscale++;
2386 sc->sc_requested_r_scale = wscale;
2387 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2388 sc->sc_flags |= SCF_WINSCALE;
2389 }
2390
2391 wnd = lso->sol_sbrcv_hiwat;
2392 wnd = imax(wnd, 0);
2393 wnd = imin(wnd, TCP_MAXWIN);
2394 sc->sc_wnd = wnd;
2395
2396 if (cookie.flags.sack_ok)
2397 sc->sc_flags |= SCF_SACK;
2398
2399 if (to->to_flags & TOF_TS) {
2400 sc->sc_flags |= SCF_TIMESTAMP;
2401 sc->sc_tsreflect = to->to_tsval;
2402 sc->sc_tsoff = tcp_new_ts_offset(inc);
2403 }
2404
2405 if (to->to_flags & TOF_SIGNATURE)
2406 sc->sc_flags |= SCF_SIGNATURE;
2407
2408 sc->sc_rxmits = 0;
2409
2410 sc->sc_port = port;
2411
2412 return (true);
2413 }
2414
2415 #ifdef INVARIANTS
2416 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)2417 syncookie_cmp(struct in_conninfo *inc, const struct syncache_head *sch,
2418 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2419 struct socket *lso, uint16_t port)
2420 {
2421 struct syncache scs;
2422 char *s;
2423
2424 bzero(&scs, sizeof(scs));
2425 if (syncookie_expand(inc, sch, &scs, th, to, lso, port) &&
2426 (sc->sc_peer_mss != scs.sc_peer_mss ||
2427 sc->sc_requested_r_scale != scs.sc_requested_r_scale ||
2428 sc->sc_requested_s_scale != scs.sc_requested_s_scale ||
2429 (sc->sc_flags & SCF_SACK) != (scs.sc_flags & SCF_SACK))) {
2430
2431 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2432 return;
2433
2434 if (sc->sc_peer_mss != scs.sc_peer_mss)
2435 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2436 s, __func__, sc->sc_peer_mss, scs.sc_peer_mss);
2437
2438 if (sc->sc_requested_r_scale != scs.sc_requested_r_scale)
2439 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2440 s, __func__, sc->sc_requested_r_scale,
2441 scs.sc_requested_r_scale);
2442
2443 if (sc->sc_requested_s_scale != scs.sc_requested_s_scale)
2444 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2445 s, __func__, sc->sc_requested_s_scale,
2446 scs.sc_requested_s_scale);
2447
2448 if ((sc->sc_flags & SCF_SACK) != (scs.sc_flags & SCF_SACK))
2449 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2450
2451 free(s, M_TCPLOG);
2452 }
2453 }
2454 #endif /* INVARIANTS */
2455
2456 static void
syncookie_reseed(void * arg)2457 syncookie_reseed(void *arg)
2458 {
2459 struct tcp_syncache *sc = arg;
2460 uint8_t *secbits;
2461 int secbit;
2462
2463 /*
2464 * Reseeding the secret doesn't have to be protected by a lock.
2465 * It only must be ensured that the new random values are visible
2466 * to all CPUs in a SMP environment. The atomic with release
2467 * semantics ensures that.
2468 */
2469 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2470 secbits = sc->secret.key[secbit];
2471 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2472 atomic_add_rel_int(&sc->secret.oddeven, 1);
2473
2474 /* Reschedule ourself. */
2475 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2476 }
2477
2478 /*
2479 * We have overflowed a bucket. Let's pause dealing with the syncache.
2480 * This function will increment the bucketoverflow statistics appropriately
2481 * (once per pause when pausing is enabled; otherwise, once per overflow).
2482 */
2483 static void
syncache_pause(struct in_conninfo * inc)2484 syncache_pause(struct in_conninfo *inc)
2485 {
2486 time_t delta;
2487 const char *s;
2488
2489 /* XXX:
2490 * 2. Add sysctl read here so we don't get the benefit of this
2491 * change without the new sysctl.
2492 */
2493
2494 /*
2495 * Try an unlocked read. If we already know that another thread
2496 * has activated the feature, there is no need to proceed.
2497 */
2498 if (V_tcp_syncache.paused)
2499 return;
2500
2501 /* Are cookied enabled? If not, we can't pause. */
2502 if (!V_tcp_syncookies) {
2503 TCPSTAT_INC(tcps_sc_bucketoverflow);
2504 return;
2505 }
2506
2507 /*
2508 * We may be the first thread to find an overflow. Get the lock
2509 * and evaluate if we need to take action.
2510 */
2511 mtx_lock(&V_tcp_syncache.pause_mtx);
2512 if (V_tcp_syncache.paused) {
2513 mtx_unlock(&V_tcp_syncache.pause_mtx);
2514 return;
2515 }
2516
2517 /* Activate protection. */
2518 V_tcp_syncache.paused = true;
2519 TCPSTAT_INC(tcps_sc_bucketoverflow);
2520
2521 /*
2522 * Determine the last backoff time. If we are seeing a re-newed
2523 * attack within that same time after last reactivating the syncache,
2524 * consider it an extension of the same attack.
2525 */
2526 delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff;
2527 if (V_tcp_syncache.pause_until + delta - time_uptime > 0) {
2528 if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) {
2529 delta <<= 1;
2530 V_tcp_syncache.pause_backoff++;
2531 }
2532 } else {
2533 delta = TCP_SYNCACHE_PAUSE_TIME;
2534 V_tcp_syncache.pause_backoff = 0;
2535 }
2536
2537 /* Log a warning, including IP addresses, if able. */
2538 if (inc != NULL)
2539 s = tcp_log_addrs(inc, NULL, NULL, NULL);
2540 else
2541 s = (const char *)NULL;
2542 log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for "
2543 "the next %lld seconds%s%s%s\n", (long long)delta,
2544 (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "",
2545 (s != NULL) ? ")" : "");
2546 free(__DECONST(void *, s), M_TCPLOG);
2547
2548 /* Use the calculated delta to set a new pause time. */
2549 V_tcp_syncache.pause_until = time_uptime + delta;
2550 callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause,
2551 &V_tcp_syncache);
2552 mtx_unlock(&V_tcp_syncache.pause_mtx);
2553 }
2554
2555 /* Evaluate whether we need to unpause. */
2556 static void
syncache_unpause(void * arg)2557 syncache_unpause(void *arg)
2558 {
2559 struct tcp_syncache *sc;
2560 time_t delta;
2561
2562 sc = arg;
2563 mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED);
2564 callout_deactivate(&sc->pause_co);
2565
2566 /*
2567 * Check to make sure we are not running early. If the pause
2568 * time has expired, then deactivate the protection.
2569 */
2570 if ((delta = sc->pause_until - time_uptime) > 0)
2571 callout_schedule(&sc->pause_co, delta * hz);
2572 else
2573 sc->paused = false;
2574 }
2575
2576 /*
2577 * Exports the syncache entries to userland so that netstat can display
2578 * them alongside the other sockets. This function is intended to be
2579 * called only from tcp_pcblist.
2580 *
2581 * Due to concurrency on an active system, the number of pcbs exported
2582 * may have no relation to max_pcbs. max_pcbs merely indicates the
2583 * amount of space the caller allocated for this function to use.
2584 */
2585 int
syncache_pcblist(struct sysctl_req * req)2586 syncache_pcblist(struct sysctl_req *req)
2587 {
2588 struct xtcpcb xt;
2589 struct syncache *sc;
2590 struct syncache_head *sch;
2591 int error, i;
2592
2593 bzero(&xt, sizeof(xt));
2594 xt.xt_len = sizeof(xt);
2595 xt.t_state = TCPS_SYN_RECEIVED;
2596 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2597 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2598 xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2599 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2600
2601 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
2602 sch = &V_tcp_syncache.hashbase[i];
2603 SCH_LOCK(sch);
2604 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2605 if (sc->sc_cred != NULL &&
2606 cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2607 continue;
2608 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2609 xt.xt_inp.inp_vflag = INP_IPV6;
2610 else
2611 xt.xt_inp.inp_vflag = INP_IPV4;
2612 xt.xt_encaps_port = sc->sc_port;
2613 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2614 sizeof (struct in_conninfo));
2615 error = SYSCTL_OUT(req, &xt, sizeof xt);
2616 if (error) {
2617 SCH_UNLOCK(sch);
2618 return (0);
2619 }
2620 }
2621 SCH_UNLOCK(sch);
2622 }
2623
2624 return (0);
2625 }
2626