1 /*-
2 * Copyright (c) 2001 McAfee, Inc.
3 * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG
4 * All rights reserved.
5 *
6 * This software was developed for the FreeBSD Project by Jonathan Lemon
7 * and McAfee Research, the Security Research Division of McAfee, Inc. under
8 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
9 * DARPA CHATS research program. [2001 McAfee, Inc.]
10 *
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
13 * are met:
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in the
18 * documentation and/or other materials provided with the distribution.
19 *
20 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * SUCH DAMAGE.
31 */
32
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD: releng/11.1/sys/netinet/tcp_syncache.c 319654 2017-06-07 12:50:54Z tuexen $");
35
36 #include "opt_inet.h"
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 #include "opt_pcbgroup.h"
40
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/hash.h>
44 #include <sys/refcount.h>
45 #include <sys/kernel.h>
46 #include <sys/sysctl.h>
47 #include <sys/limits.h>
48 #include <sys/lock.h>
49 #include <sys/mutex.h>
50 #include <sys/malloc.h>
51 #include <sys/mbuf.h>
52 #include <sys/proc.h> /* for proc0 declaration */
53 #include <sys/random.h>
54 #include <sys/socket.h>
55 #include <sys/socketvar.h>
56 #include <sys/syslog.h>
57 #include <sys/ucred.h>
58
59 #include <sys/md5.h>
60 #include <crypto/siphash/siphash.h>
61
62 #include <vm/uma.h>
63
64 #include <net/if.h>
65 #include <net/if_var.h>
66 #include <net/route.h>
67 #include <net/vnet.h>
68
69 #include <netinet/in.h>
70 #include <netinet/in_systm.h>
71 #include <netinet/ip.h>
72 #include <netinet/in_var.h>
73 #include <netinet/in_pcb.h>
74 #include <netinet/ip_var.h>
75 #include <netinet/ip_options.h>
76 #ifdef INET6
77 #include <netinet/ip6.h>
78 #include <netinet/icmp6.h>
79 #include <netinet6/nd6.h>
80 #include <netinet6/ip6_var.h>
81 #include <netinet6/in6_pcb.h>
82 #endif
83 #include <netinet/tcp.h>
84 #ifdef TCP_RFC7413
85 #include <netinet/tcp_fastopen.h>
86 #endif
87 #include <netinet/tcp_fsm.h>
88 #include <netinet/tcp_seq.h>
89 #include <netinet/tcp_timer.h>
90 #include <netinet/tcp_var.h>
91 #include <netinet/tcp_syncache.h>
92 #ifdef INET6
93 #include <netinet6/tcp6_var.h>
94 #endif
95 #ifdef TCP_OFFLOAD
96 #include <netinet/toecore.h>
97 #endif
98
99 #include <netipsec/ipsec_support.h>
100
101 #include <machine/in_cksum.h>
102
103 #include <security/mac/mac_framework.h>
104
105 static VNET_DEFINE(int, tcp_syncookies) = 1;
106 #define V_tcp_syncookies VNET(tcp_syncookies)
107 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
108 &VNET_NAME(tcp_syncookies), 0,
109 "Use TCP SYN cookies if the syncache overflows");
110
111 static VNET_DEFINE(int, tcp_syncookiesonly) = 0;
112 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly)
113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
114 &VNET_NAME(tcp_syncookiesonly), 0,
115 "Use only TCP SYN cookies");
116
117 #ifdef TCP_OFFLOAD
118 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
119 #endif
120
121 static void syncache_drop(struct syncache *, struct syncache_head *);
122 static void syncache_free(struct syncache *);
123 static void syncache_insert(struct syncache *, struct syncache_head *);
124 static int syncache_respond(struct syncache *, struct syncache_head *, int,
125 const 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 struct syncache
136 *syncookie_lookup(struct in_conninfo *, struct syncache_head *,
137 struct syncache *, struct tcphdr *, struct tcpopt *,
138 struct socket *);
139 static void syncookie_reseed(void *);
140 #ifdef INVARIANTS
141 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
142 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
143 struct socket *lso);
144 #endif
145
146 /*
147 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
148 * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds,
149 * the odds are that the user has given up attempting to connect by then.
150 */
151 #define SYNCACHE_MAXREXMTS 3
152
153 /* Arbitrary values */
154 #define TCP_SYNCACHE_HASHSIZE 512
155 #define TCP_SYNCACHE_BUCKETLIMIT 30
156
157 static VNET_DEFINE(struct tcp_syncache, tcp_syncache);
158 #define V_tcp_syncache VNET(tcp_syncache)
159
160 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0,
161 "TCP SYN cache");
162
163 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
164 &VNET_NAME(tcp_syncache.bucket_limit), 0,
165 "Per-bucket hash limit for syncache");
166
167 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
168 &VNET_NAME(tcp_syncache.cache_limit), 0,
169 "Overall entry limit for syncache");
170
171 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
172 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
173
174 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
175 &VNET_NAME(tcp_syncache.hashsize), 0,
176 "Size of TCP syncache hashtable");
177
178 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_VNET | CTLFLAG_RW,
179 &VNET_NAME(tcp_syncache.rexmt_limit), 0,
180 "Limit on SYN/ACK retransmissions");
181
182 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
183 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
184 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
185 "Send reset on socket allocation failure");
186
187 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
188
189 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx)
190 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx)
191 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED)
192
193 /*
194 * Requires the syncache entry to be already removed from the bucket list.
195 */
196 static void
197 syncache_free(struct syncache *sc)
198 {
199
200 if (sc->sc_ipopts)
201 (void) m_free(sc->sc_ipopts);
202 if (sc->sc_cred)
203 crfree(sc->sc_cred);
204 #ifdef MAC
205 mac_syncache_destroy(&sc->sc_label);
206 #endif
207
208 uma_zfree(V_tcp_syncache.zone, sc);
209 }
210
211 void
212 syncache_init(void)
213 {
214 int i;
215
216 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
217 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
218 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
219 V_tcp_syncache.hash_secret = arc4random();
220
221 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
222 &V_tcp_syncache.hashsize);
223 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
224 &V_tcp_syncache.bucket_limit);
225 if (!powerof2(V_tcp_syncache.hashsize) ||
226 V_tcp_syncache.hashsize == 0) {
227 printf("WARNING: syncache hash size is not a power of 2.\n");
228 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
229 }
230 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
231
232 /* Set limits. */
233 V_tcp_syncache.cache_limit =
234 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
235 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
236 &V_tcp_syncache.cache_limit);
237
238 /* Allocate the hash table. */
239 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
240 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
241
242 #ifdef VIMAGE
243 V_tcp_syncache.vnet = curvnet;
244 #endif
245
246 /* Initialize the hash buckets. */
247 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
248 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
249 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
250 NULL, MTX_DEF);
251 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
252 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
253 V_tcp_syncache.hashbase[i].sch_length = 0;
254 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
255 V_tcp_syncache.hashbase[i].sch_last_overflow =
256 -(SYNCOOKIE_LIFETIME + 1);
257 }
258
259 /* Create the syncache entry zone. */
260 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
261 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
262 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
263 V_tcp_syncache.cache_limit);
264
265 /* Start the SYN cookie reseeder callout. */
266 callout_init(&V_tcp_syncache.secret.reseed, 1);
267 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
268 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
269 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
270 syncookie_reseed, &V_tcp_syncache);
271 }
272
273 #ifdef VIMAGE
274 void
275 syncache_destroy(void)
276 {
277 struct syncache_head *sch;
278 struct syncache *sc, *nsc;
279 int i;
280
281 /*
282 * Stop the re-seed timer before freeing resources. No need to
283 * possibly schedule it another time.
284 */
285 callout_drain(&V_tcp_syncache.secret.reseed);
286
287 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */
288 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
289
290 sch = &V_tcp_syncache.hashbase[i];
291 callout_drain(&sch->sch_timer);
292
293 SCH_LOCK(sch);
294 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
295 syncache_drop(sc, sch);
296 SCH_UNLOCK(sch);
297 KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
298 ("%s: sch->sch_bucket not empty", __func__));
299 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
300 __func__, sch->sch_length));
301 mtx_destroy(&sch->sch_mtx);
302 }
303
304 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
305 ("%s: cache_count not 0", __func__));
306
307 /* Free the allocated global resources. */
308 uma_zdestroy(V_tcp_syncache.zone);
309 free(V_tcp_syncache.hashbase, M_SYNCACHE);
310 }
311 #endif
312
313 /*
314 * Inserts a syncache entry into the specified bucket row.
315 * Locks and unlocks the syncache_head autonomously.
316 */
317 static void
318 syncache_insert(struct syncache *sc, struct syncache_head *sch)
319 {
320 struct syncache *sc2;
321
322 SCH_LOCK(sch);
323
324 /*
325 * Make sure that we don't overflow the per-bucket limit.
326 * If the bucket is full, toss the oldest element.
327 */
328 if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
329 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
330 ("sch->sch_length incorrect"));
331 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
332 sch->sch_last_overflow = time_uptime;
333 syncache_drop(sc2, sch);
334 TCPSTAT_INC(tcps_sc_bucketoverflow);
335 }
336
337 /* Put it into the bucket. */
338 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
339 sch->sch_length++;
340
341 #ifdef TCP_OFFLOAD
342 if (ADDED_BY_TOE(sc)) {
343 struct toedev *tod = sc->sc_tod;
344
345 tod->tod_syncache_added(tod, sc->sc_todctx);
346 }
347 #endif
348
349 /* Reinitialize the bucket row's timer. */
350 if (sch->sch_length == 1)
351 sch->sch_nextc = ticks + INT_MAX;
352 syncache_timeout(sc, sch, 1);
353
354 SCH_UNLOCK(sch);
355
356 TCPSTATES_INC(TCPS_SYN_RECEIVED);
357 TCPSTAT_INC(tcps_sc_added);
358 }
359
360 /*
361 * Remove and free entry from syncache bucket row.
362 * Expects locked syncache head.
363 */
364 static void
365 syncache_drop(struct syncache *sc, struct syncache_head *sch)
366 {
367
368 SCH_LOCK_ASSERT(sch);
369
370 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
371 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
372 sch->sch_length--;
373
374 #ifdef TCP_OFFLOAD
375 if (ADDED_BY_TOE(sc)) {
376 struct toedev *tod = sc->sc_tod;
377
378 tod->tod_syncache_removed(tod, sc->sc_todctx);
379 }
380 #endif
381
382 syncache_free(sc);
383 }
384
385 /*
386 * Engage/reengage time on bucket row.
387 */
388 static void
389 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
390 {
391 sc->sc_rxttime = ticks +
392 TCPTV_RTOBASE * (tcp_syn_backoff[sc->sc_rxmits]);
393 sc->sc_rxmits++;
394 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
395 sch->sch_nextc = sc->sc_rxttime;
396 if (docallout)
397 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
398 syncache_timer, (void *)sch);
399 }
400 }
401
402 /*
403 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
404 * If we have retransmitted an entry the maximum number of times, expire it.
405 * One separate timer for each bucket row.
406 */
407 static void
408 syncache_timer(void *xsch)
409 {
410 struct syncache_head *sch = (struct syncache_head *)xsch;
411 struct syncache *sc, *nsc;
412 int tick = ticks;
413 char *s;
414
415 CURVNET_SET(sch->sch_sc->vnet);
416
417 /* NB: syncache_head has already been locked by the callout. */
418 SCH_LOCK_ASSERT(sch);
419
420 /*
421 * In the following cycle we may remove some entries and/or
422 * advance some timeouts, so re-initialize the bucket timer.
423 */
424 sch->sch_nextc = tick + INT_MAX;
425
426 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
427 /*
428 * We do not check if the listen socket still exists
429 * and accept the case where the listen socket may be
430 * gone by the time we resend the SYN/ACK. We do
431 * not expect this to happens often. If it does,
432 * then the RST will be sent by the time the remote
433 * host does the SYN/ACK->ACK.
434 */
435 if (TSTMP_GT(sc->sc_rxttime, tick)) {
436 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
437 sch->sch_nextc = sc->sc_rxttime;
438 continue;
439 }
440 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
441 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
442 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
443 "giving up and removing syncache entry\n",
444 s, __func__);
445 free(s, M_TCPLOG);
446 }
447 syncache_drop(sc, sch);
448 TCPSTAT_INC(tcps_sc_stale);
449 continue;
450 }
451 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
452 log(LOG_DEBUG, "%s; %s: Response timeout, "
453 "retransmitting (%u) SYN|ACK\n",
454 s, __func__, sc->sc_rxmits);
455 free(s, M_TCPLOG);
456 }
457
458 syncache_respond(sc, sch, 1, NULL);
459 TCPSTAT_INC(tcps_sc_retransmitted);
460 syncache_timeout(sc, sch, 0);
461 }
462 if (!TAILQ_EMPTY(&(sch)->sch_bucket))
463 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
464 syncache_timer, (void *)(sch));
465 CURVNET_RESTORE();
466 }
467
468 /*
469 * Find an entry in the syncache.
470 * Returns always with locked syncache_head plus a matching entry or NULL.
471 */
472 static struct syncache *
473 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
474 {
475 struct syncache *sc;
476 struct syncache_head *sch;
477 uint32_t hash;
478
479 /*
480 * The hash is built on foreign port + local port + foreign address.
481 * We rely on the fact that struct in_conninfo starts with 16 bits
482 * of foreign port, then 16 bits of local port then followed by 128
483 * bits of foreign address. In case of IPv4 address, the first 3
484 * 32-bit words of the address always are zeroes.
485 */
486 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
487 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
488
489 sch = &V_tcp_syncache.hashbase[hash];
490 *schp = sch;
491 SCH_LOCK(sch);
492
493 /* Circle through bucket row to find matching entry. */
494 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
495 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
496 sizeof(struct in_endpoints)) == 0)
497 break;
498
499 return (sc); /* Always returns with locked sch. */
500 }
501
502 /*
503 * This function is called when we get a RST for a
504 * non-existent connection, so that we can see if the
505 * connection is in the syn cache. If it is, zap it.
506 */
507 void
508 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th)
509 {
510 struct syncache *sc;
511 struct syncache_head *sch;
512 char *s = NULL;
513
514 sc = syncache_lookup(inc, &sch); /* returns locked sch */
515 SCH_LOCK_ASSERT(sch);
516
517 /*
518 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
519 * See RFC 793 page 65, section SEGMENT ARRIVES.
520 */
521 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
522 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
523 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
524 "FIN flag set, segment ignored\n", s, __func__);
525 TCPSTAT_INC(tcps_badrst);
526 goto done;
527 }
528
529 /*
530 * No corresponding connection was found in syncache.
531 * If syncookies are enabled and possibly exclusively
532 * used, or we are under memory pressure, a valid RST
533 * may not find a syncache entry. In that case we're
534 * done and no SYN|ACK retransmissions will happen.
535 * Otherwise the RST was misdirected or spoofed.
536 */
537 if (sc == NULL) {
538 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
539 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
540 "syncache entry (possibly syncookie only), "
541 "segment ignored\n", s, __func__);
542 TCPSTAT_INC(tcps_badrst);
543 goto done;
544 }
545
546 /*
547 * If the RST bit is set, check the sequence number to see
548 * if this is a valid reset segment.
549 * RFC 793 page 37:
550 * In all states except SYN-SENT, all reset (RST) segments
551 * are validated by checking their SEQ-fields. A reset is
552 * valid if its sequence number is in the window.
553 *
554 * The sequence number in the reset segment is normally an
555 * echo of our outgoing acknowlegement numbers, but some hosts
556 * send a reset with the sequence number at the rightmost edge
557 * of our receive window, and we have to handle this case.
558 */
559 if (SEQ_GEQ(th->th_seq, sc->sc_irs) &&
560 SEQ_LEQ(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
561 syncache_drop(sc, sch);
562 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
563 log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, "
564 "connection attempt aborted by remote endpoint\n",
565 s, __func__);
566 TCPSTAT_INC(tcps_sc_reset);
567 } else {
568 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
569 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
570 "IRS %u (+WND %u), segment ignored\n",
571 s, __func__, th->th_seq, sc->sc_irs, sc->sc_wnd);
572 TCPSTAT_INC(tcps_badrst);
573 }
574
575 done:
576 if (s != NULL)
577 free(s, M_TCPLOG);
578 SCH_UNLOCK(sch);
579 }
580
581 void
582 syncache_badack(struct in_conninfo *inc)
583 {
584 struct syncache *sc;
585 struct syncache_head *sch;
586
587 sc = syncache_lookup(inc, &sch); /* returns locked sch */
588 SCH_LOCK_ASSERT(sch);
589 if (sc != NULL) {
590 syncache_drop(sc, sch);
591 TCPSTAT_INC(tcps_sc_badack);
592 }
593 SCH_UNLOCK(sch);
594 }
595
596 void
597 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq)
598 {
599 struct syncache *sc;
600 struct syncache_head *sch;
601
602 sc = syncache_lookup(inc, &sch); /* returns locked sch */
603 SCH_LOCK_ASSERT(sch);
604 if (sc == NULL)
605 goto done;
606
607 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
608 if (ntohl(th_seq) != sc->sc_iss)
609 goto done;
610
611 /*
612 * If we've rertransmitted 3 times and this is our second error,
613 * we remove the entry. Otherwise, we allow it to continue on.
614 * This prevents us from incorrectly nuking an entry during a
615 * spurious network outage.
616 *
617 * See tcp_notify().
618 */
619 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
620 sc->sc_flags |= SCF_UNREACH;
621 goto done;
622 }
623 syncache_drop(sc, sch);
624 TCPSTAT_INC(tcps_sc_unreach);
625 done:
626 SCH_UNLOCK(sch);
627 }
628
629 /*
630 * Build a new TCP socket structure from a syncache entry.
631 *
632 * On success return the newly created socket with its underlying inp locked.
633 */
634 static struct socket *
635 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
636 {
637 struct tcp_function_block *blk;
638 struct inpcb *inp = NULL;
639 struct socket *so;
640 struct tcpcb *tp;
641 int error;
642 char *s;
643
644 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
645
646 /*
647 * Ok, create the full blown connection, and set things up
648 * as they would have been set up if we had created the
649 * connection when the SYN arrived. If we can't create
650 * the connection, abort it.
651 */
652 so = sonewconn(lso, 0);
653 if (so == NULL) {
654 /*
655 * Drop the connection; we will either send a RST or
656 * have the peer retransmit its SYN again after its
657 * RTO and try again.
658 */
659 TCPSTAT_INC(tcps_listendrop);
660 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
661 log(LOG_DEBUG, "%s; %s: Socket create failed "
662 "due to limits or memory shortage\n",
663 s, __func__);
664 free(s, M_TCPLOG);
665 }
666 goto abort2;
667 }
668 #ifdef MAC
669 mac_socketpeer_set_from_mbuf(m, so);
670 #endif
671
672 inp = sotoinpcb(so);
673 inp->inp_inc.inc_fibnum = so->so_fibnum;
674 INP_WLOCK(inp);
675 /*
676 * Exclusive pcbinfo lock is not required in syncache socket case even
677 * if two inpcb locks can be acquired simultaneously:
678 * - the inpcb in LISTEN state,
679 * - the newly created inp.
680 *
681 * In this case, an inp cannot be at same time in LISTEN state and
682 * just created by an accept() call.
683 */
684 INP_HASH_WLOCK(&V_tcbinfo);
685
686 /* Insert new socket into PCB hash list. */
687 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
688 #ifdef INET6
689 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
690 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
691 } else {
692 inp->inp_vflag &= ~INP_IPV6;
693 inp->inp_vflag |= INP_IPV4;
694 #endif
695 inp->inp_laddr = sc->sc_inc.inc_laddr;
696 #ifdef INET6
697 }
698 #endif
699
700 /*
701 * If there's an mbuf and it has a flowid, then let's initialise the
702 * inp with that particular flowid.
703 */
704 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
705 inp->inp_flowid = m->m_pkthdr.flowid;
706 inp->inp_flowtype = M_HASHTYPE_GET(m);
707 }
708
709 /*
710 * Install in the reservation hash table for now, but don't yet
711 * install a connection group since the full 4-tuple isn't yet
712 * configured.
713 */
714 inp->inp_lport = sc->sc_inc.inc_lport;
715 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
716 /*
717 * Undo the assignments above if we failed to
718 * put the PCB on the hash lists.
719 */
720 #ifdef INET6
721 if (sc->sc_inc.inc_flags & INC_ISIPV6)
722 inp->in6p_laddr = in6addr_any;
723 else
724 #endif
725 inp->inp_laddr.s_addr = INADDR_ANY;
726 inp->inp_lport = 0;
727 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
728 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
729 "with error %i\n",
730 s, __func__, error);
731 free(s, M_TCPLOG);
732 }
733 INP_HASH_WUNLOCK(&V_tcbinfo);
734 goto abort;
735 }
736 #ifdef INET6
737 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
738 struct inpcb *oinp = sotoinpcb(lso);
739 struct in6_addr laddr6;
740 struct sockaddr_in6 sin6;
741 /*
742 * Inherit socket options from the listening socket.
743 * Note that in6p_inputopts are not (and should not be)
744 * copied, since it stores previously received options and is
745 * used to detect if each new option is different than the
746 * previous one and hence should be passed to a user.
747 * If we copied in6p_inputopts, a user would not be able to
748 * receive options just after calling the accept system call.
749 */
750 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
751 if (oinp->in6p_outputopts)
752 inp->in6p_outputopts =
753 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
754
755 sin6.sin6_family = AF_INET6;
756 sin6.sin6_len = sizeof(sin6);
757 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
758 sin6.sin6_port = sc->sc_inc.inc_fport;
759 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
760 laddr6 = inp->in6p_laddr;
761 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
762 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
763 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
764 thread0.td_ucred, m)) != 0) {
765 inp->in6p_laddr = laddr6;
766 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
767 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
768 "with error %i\n",
769 s, __func__, error);
770 free(s, M_TCPLOG);
771 }
772 INP_HASH_WUNLOCK(&V_tcbinfo);
773 goto abort;
774 }
775 /* Override flowlabel from in6_pcbconnect. */
776 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
777 inp->inp_flow |= sc->sc_flowlabel;
778 }
779 #endif /* INET6 */
780 #if defined(INET) && defined(INET6)
781 else
782 #endif
783 #ifdef INET
784 {
785 struct in_addr laddr;
786 struct sockaddr_in sin;
787
788 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
789
790 if (inp->inp_options == NULL) {
791 inp->inp_options = sc->sc_ipopts;
792 sc->sc_ipopts = NULL;
793 }
794
795 sin.sin_family = AF_INET;
796 sin.sin_len = sizeof(sin);
797 sin.sin_addr = sc->sc_inc.inc_faddr;
798 sin.sin_port = sc->sc_inc.inc_fport;
799 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
800 laddr = inp->inp_laddr;
801 if (inp->inp_laddr.s_addr == INADDR_ANY)
802 inp->inp_laddr = sc->sc_inc.inc_laddr;
803 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
804 thread0.td_ucred, m)) != 0) {
805 inp->inp_laddr = laddr;
806 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
807 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
808 "with error %i\n",
809 s, __func__, error);
810 free(s, M_TCPLOG);
811 }
812 INP_HASH_WUNLOCK(&V_tcbinfo);
813 goto abort;
814 }
815 }
816 #endif /* INET */
817 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
818 /* Copy old policy into new socket's. */
819 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
820 printf("syncache_socket: could not copy policy\n");
821 #endif
822 INP_HASH_WUNLOCK(&V_tcbinfo);
823 tp = intotcpcb(inp);
824 tcp_state_change(tp, TCPS_SYN_RECEIVED);
825 tp->iss = sc->sc_iss;
826 tp->irs = sc->sc_irs;
827 tcp_rcvseqinit(tp);
828 tcp_sendseqinit(tp);
829 blk = sototcpcb(lso)->t_fb;
830 if (blk != tp->t_fb) {
831 /*
832 * Our parents t_fb was not the default,
833 * we need to release our ref on tp->t_fb and
834 * pickup one on the new entry.
835 */
836 struct tcp_function_block *rblk;
837
838 rblk = find_and_ref_tcp_fb(blk);
839 KASSERT(rblk != NULL,
840 ("cannot find blk %p out of syncache?", blk));
841 if (tp->t_fb->tfb_tcp_fb_fini)
842 (*tp->t_fb->tfb_tcp_fb_fini)(tp);
843 refcount_release(&tp->t_fb->tfb_refcnt);
844 tp->t_fb = rblk;
845 if (tp->t_fb->tfb_tcp_fb_init) {
846 (*tp->t_fb->tfb_tcp_fb_init)(tp);
847 }
848 }
849 tp->snd_wl1 = sc->sc_irs;
850 tp->snd_max = tp->iss + 1;
851 tp->snd_nxt = tp->iss + 1;
852 tp->rcv_up = sc->sc_irs + 1;
853 tp->rcv_wnd = sc->sc_wnd;
854 tp->rcv_adv += tp->rcv_wnd;
855 tp->last_ack_sent = tp->rcv_nxt;
856
857 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
858 if (sc->sc_flags & SCF_NOOPT)
859 tp->t_flags |= TF_NOOPT;
860 else {
861 if (sc->sc_flags & SCF_WINSCALE) {
862 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
863 tp->snd_scale = sc->sc_requested_s_scale;
864 tp->request_r_scale = sc->sc_requested_r_scale;
865 }
866 if (sc->sc_flags & SCF_TIMESTAMP) {
867 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
868 tp->ts_recent = sc->sc_tsreflect;
869 tp->ts_recent_age = tcp_ts_getticks();
870 tp->ts_offset = sc->sc_tsoff;
871 }
872 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
873 if (sc->sc_flags & SCF_SIGNATURE)
874 tp->t_flags |= TF_SIGNATURE;
875 #endif
876 if (sc->sc_flags & SCF_SACK)
877 tp->t_flags |= TF_SACK_PERMIT;
878 }
879
880 if (sc->sc_flags & SCF_ECN)
881 tp->t_flags |= TF_ECN_PERMIT;
882
883 /*
884 * Set up MSS and get cached values from tcp_hostcache.
885 * This might overwrite some of the defaults we just set.
886 */
887 tcp_mss(tp, sc->sc_peer_mss);
888
889 /*
890 * If the SYN,ACK was retransmitted, indicate that CWND to be
891 * limited to one segment in cc_conn_init().
892 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
893 */
894 if (sc->sc_rxmits > 1)
895 tp->snd_cwnd = 1;
896
897 #ifdef TCP_OFFLOAD
898 /*
899 * Allow a TOE driver to install its hooks. Note that we hold the
900 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
901 * new connection before the TOE driver has done its thing.
902 */
903 if (ADDED_BY_TOE(sc)) {
904 struct toedev *tod = sc->sc_tod;
905
906 tod->tod_offload_socket(tod, sc->sc_todctx, so);
907 }
908 #endif
909 /*
910 * Copy and activate timers.
911 */
912 tp->t_keepinit = sototcpcb(lso)->t_keepinit;
913 tp->t_keepidle = sototcpcb(lso)->t_keepidle;
914 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
915 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
916 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
917
918 TCPSTAT_INC(tcps_accepts);
919 return (so);
920
921 abort:
922 INP_WUNLOCK(inp);
923 abort2:
924 if (so != NULL)
925 soabort(so);
926 return (NULL);
927 }
928
929 /*
930 * This function gets called when we receive an ACK for a
931 * socket in the LISTEN state. We look up the connection
932 * in the syncache, and if its there, we pull it out of
933 * the cache and turn it into a full-blown connection in
934 * the SYN-RECEIVED state.
935 *
936 * On syncache_socket() success the newly created socket
937 * has its underlying inp locked.
938 */
939 int
940 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
941 struct socket **lsop, struct mbuf *m)
942 {
943 struct syncache *sc;
944 struct syncache_head *sch;
945 struct syncache scs;
946 char *s;
947
948 /*
949 * Global TCP locks are held because we manipulate the PCB lists
950 * and create a new socket.
951 */
952 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
953 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
954 ("%s: can handle only ACK", __func__));
955
956 sc = syncache_lookup(inc, &sch); /* returns locked sch */
957 SCH_LOCK_ASSERT(sch);
958
959 #ifdef INVARIANTS
960 /*
961 * Test code for syncookies comparing the syncache stored
962 * values with the reconstructed values from the cookie.
963 */
964 if (sc != NULL)
965 syncookie_cmp(inc, sch, sc, th, to, *lsop);
966 #endif
967
968 if (sc == NULL) {
969 /*
970 * There is no syncache entry, so see if this ACK is
971 * a returning syncookie. To do this, first:
972 * A. Check if syncookies are used in case of syncache
973 * overflows
974 * B. See if this socket has had a syncache entry dropped in
975 * the recent past. We don't want to accept a bogus
976 * syncookie if we've never received a SYN or accept it
977 * twice.
978 * C. check that the syncookie is valid. If it is, then
979 * cobble up a fake syncache entry, and return.
980 */
981 if (!V_tcp_syncookies) {
982 SCH_UNLOCK(sch);
983 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
984 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
985 "segment rejected (syncookies disabled)\n",
986 s, __func__);
987 goto failed;
988 }
989 if (!V_tcp_syncookiesonly &&
990 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
991 SCH_UNLOCK(sch);
992 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
993 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
994 "segment rejected (no syncache entry)\n",
995 s, __func__);
996 goto failed;
997 }
998 bzero(&scs, sizeof(scs));
999 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
1000 SCH_UNLOCK(sch);
1001 if (sc == NULL) {
1002 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1003 log(LOG_DEBUG, "%s; %s: Segment failed "
1004 "SYNCOOKIE authentication, segment rejected "
1005 "(probably spoofed)\n", s, __func__);
1006 goto failed;
1007 }
1008 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1009 /* If received ACK has MD5 signature, check it. */
1010 if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1011 (!TCPMD5_ENABLED() ||
1012 TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1013 /* Drop the ACK. */
1014 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1015 log(LOG_DEBUG, "%s; %s: Segment rejected, "
1016 "MD5 signature doesn't match.\n",
1017 s, __func__);
1018 free(s, M_TCPLOG);
1019 }
1020 TCPSTAT_INC(tcps_sig_err_sigopt);
1021 return (-1); /* Do not send RST */
1022 }
1023 #endif /* TCP_SIGNATURE */
1024 } else {
1025 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1026 /*
1027 * If listening socket requested TCP digests, check that
1028 * received ACK has signature and it is correct.
1029 * If not, drop the ACK and leave sc entry in th cache,
1030 * because SYN was received with correct signature.
1031 */
1032 if (sc->sc_flags & SCF_SIGNATURE) {
1033 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1034 /* No signature */
1035 TCPSTAT_INC(tcps_sig_err_nosigopt);
1036 SCH_UNLOCK(sch);
1037 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1038 log(LOG_DEBUG, "%s; %s: Segment "
1039 "rejected, MD5 signature wasn't "
1040 "provided.\n", s, __func__);
1041 free(s, M_TCPLOG);
1042 }
1043 return (-1); /* Do not send RST */
1044 }
1045 if (!TCPMD5_ENABLED() ||
1046 TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1047 /* Doesn't match or no SA */
1048 SCH_UNLOCK(sch);
1049 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1050 log(LOG_DEBUG, "%s; %s: Segment "
1051 "rejected, MD5 signature doesn't "
1052 "match.\n", s, __func__);
1053 free(s, M_TCPLOG);
1054 }
1055 return (-1); /* Do not send RST */
1056 }
1057 }
1058 #endif /* TCP_SIGNATURE */
1059 /*
1060 * Pull out the entry to unlock the bucket row.
1061 *
1062 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1063 * tcp_state_change(). The tcpcb is not existent at this
1064 * moment. A new one will be allocated via syncache_socket->
1065 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1066 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1067 */
1068 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1069 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1070 sch->sch_length--;
1071 #ifdef TCP_OFFLOAD
1072 if (ADDED_BY_TOE(sc)) {
1073 struct toedev *tod = sc->sc_tod;
1074
1075 tod->tod_syncache_removed(tod, sc->sc_todctx);
1076 }
1077 #endif
1078 SCH_UNLOCK(sch);
1079 }
1080
1081 /*
1082 * Segment validation:
1083 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1084 */
1085 if (th->th_ack != sc->sc_iss + 1) {
1086 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1087 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1088 "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1089 goto failed;
1090 }
1091
1092 /*
1093 * The SEQ must fall in the window starting at the received
1094 * initial receive sequence number + 1 (the SYN).
1095 */
1096 if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1097 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1098 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1099 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1100 "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1101 goto failed;
1102 }
1103
1104 /*
1105 * If timestamps were not negotiated during SYN/ACK they
1106 * must not appear on any segment during this session.
1107 */
1108 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1109 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1110 log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1111 "segment rejected\n", s, __func__);
1112 goto failed;
1113 }
1114
1115 /*
1116 * If timestamps were negotiated during SYN/ACK they should
1117 * appear on every segment during this session.
1118 * XXXAO: This is only informal as there have been unverified
1119 * reports of non-compliants stacks.
1120 */
1121 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1122 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1123 log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1124 "no action\n", s, __func__);
1125 free(s, M_TCPLOG);
1126 s = NULL;
1127 }
1128 }
1129
1130 /*
1131 * If timestamps were negotiated, the reflected timestamp
1132 * must be equal to what we actually sent in the SYN|ACK
1133 * except in the case of 0. Some boxes are known for sending
1134 * broken timestamp replies during the 3whs (and potentially
1135 * during the connection also).
1136 *
1137 * Accept the final ACK of 3whs with reflected timestamp of 0
1138 * instead of sending a RST and deleting the syncache entry.
1139 */
1140 if ((to->to_flags & TOF_TS) && to->to_tsecr &&
1141 to->to_tsecr != sc->sc_ts) {
1142 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1143 log(LOG_DEBUG, "%s; %s: TSECR %u != TS %u, "
1144 "segment rejected\n",
1145 s, __func__, to->to_tsecr, sc->sc_ts);
1146 goto failed;
1147 }
1148
1149 *lsop = syncache_socket(sc, *lsop, m);
1150
1151 if (*lsop == NULL)
1152 TCPSTAT_INC(tcps_sc_aborted);
1153 else
1154 TCPSTAT_INC(tcps_sc_completed);
1155
1156 /* how do we find the inp for the new socket? */
1157 if (sc != &scs)
1158 syncache_free(sc);
1159 return (1);
1160 failed:
1161 if (sc != NULL && sc != &scs)
1162 syncache_free(sc);
1163 if (s != NULL)
1164 free(s, M_TCPLOG);
1165 *lsop = NULL;
1166 return (0);
1167 }
1168
1169 #ifdef TCP_RFC7413
1170 static void
1171 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
1172 uint64_t response_cookie)
1173 {
1174 struct inpcb *inp;
1175 struct tcpcb *tp;
1176 unsigned int *pending_counter;
1177
1178 /*
1179 * Global TCP locks are held because we manipulate the PCB lists
1180 * and create a new socket.
1181 */
1182 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1183
1184 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
1185 *lsop = syncache_socket(sc, *lsop, m);
1186 if (*lsop == NULL) {
1187 TCPSTAT_INC(tcps_sc_aborted);
1188 atomic_subtract_int(pending_counter, 1);
1189 } else {
1190 inp = sotoinpcb(*lsop);
1191 tp = intotcpcb(inp);
1192 tp->t_flags |= TF_FASTOPEN;
1193 tp->t_tfo_cookie = response_cookie;
1194 tp->snd_max = tp->iss;
1195 tp->snd_nxt = tp->iss;
1196 tp->t_tfo_pending = pending_counter;
1197 TCPSTAT_INC(tcps_sc_completed);
1198 }
1199 }
1200 #endif /* TCP_RFC7413 */
1201
1202 /*
1203 * Given a LISTEN socket and an inbound SYN request, add
1204 * this to the syn cache, and send back a segment:
1205 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1206 * to the source.
1207 *
1208 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1209 * Doing so would require that we hold onto the data and deliver it
1210 * to the application. However, if we are the target of a SYN-flood
1211 * DoS attack, an attacker could send data which would eventually
1212 * consume all available buffer space if it were ACKed. By not ACKing
1213 * the data, we avoid this DoS scenario.
1214 *
1215 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1216 * cookie is processed, V_tcp_fastopen_enabled set to true, and the
1217 * TCP_FASTOPEN socket option is set. In this case, a new socket is created
1218 * and returned via lsop, the mbuf is not freed so that tcp_input() can
1219 * queue its data to the socket, and 1 is returned to indicate the
1220 * TFO-socket-creation path was taken.
1221 */
1222 int
1223 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1224 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1225 void *todctx)
1226 {
1227 struct tcpcb *tp;
1228 struct socket *so;
1229 struct syncache *sc = NULL;
1230 struct syncache_head *sch;
1231 struct mbuf *ipopts = NULL;
1232 u_int ltflags;
1233 int win, sb_hiwat, ip_ttl, ip_tos;
1234 char *s;
1235 int rv = 0;
1236 #ifdef INET6
1237 int autoflowlabel = 0;
1238 #endif
1239 #ifdef MAC
1240 struct label *maclabel;
1241 #endif
1242 struct syncache scs;
1243 struct ucred *cred;
1244 #ifdef TCP_RFC7413
1245 uint64_t tfo_response_cookie;
1246 int tfo_cookie_valid = 0;
1247 int tfo_response_cookie_valid = 0;
1248 #endif
1249
1250 INP_WLOCK_ASSERT(inp); /* listen socket */
1251 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1252 ("%s: unexpected tcp flags", __func__));
1253
1254 /*
1255 * Combine all so/tp operations very early to drop the INP lock as
1256 * soon as possible.
1257 */
1258 so = *lsop;
1259 tp = sototcpcb(so);
1260 cred = crhold(so->so_cred);
1261
1262 #ifdef INET6
1263 if ((inc->inc_flags & INC_ISIPV6) &&
1264 (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1265 autoflowlabel = 1;
1266 #endif
1267 ip_ttl = inp->inp_ip_ttl;
1268 ip_tos = inp->inp_ip_tos;
1269 win = sbspace(&so->so_rcv);
1270 sb_hiwat = so->so_rcv.sb_hiwat;
1271 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1272
1273 #ifdef TCP_RFC7413
1274 if (V_tcp_fastopen_enabled && (tp->t_flags & TF_FASTOPEN) &&
1275 (tp->t_tfo_pending != NULL) && (to->to_flags & TOF_FASTOPEN)) {
1276 /*
1277 * Limit the number of pending TFO connections to
1278 * approximately half of the queue limit. This prevents TFO
1279 * SYN floods from starving the service by filling the
1280 * listen queue with bogus TFO connections.
1281 */
1282 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1283 (so->so_qlimit / 2)) {
1284 int result;
1285
1286 result = tcp_fastopen_check_cookie(inc,
1287 to->to_tfo_cookie, to->to_tfo_len,
1288 &tfo_response_cookie);
1289 tfo_cookie_valid = (result > 0);
1290 tfo_response_cookie_valid = (result >= 0);
1291 } else
1292 atomic_subtract_int(tp->t_tfo_pending, 1);
1293 }
1294 #endif
1295
1296 /* By the time we drop the lock these should no longer be used. */
1297 so = NULL;
1298 tp = NULL;
1299
1300 #ifdef MAC
1301 if (mac_syncache_init(&maclabel) != 0) {
1302 INP_WUNLOCK(inp);
1303 goto done;
1304 } else
1305 mac_syncache_create(maclabel, inp);
1306 #endif
1307 #ifdef TCP_RFC7413
1308 if (!tfo_cookie_valid)
1309 #endif
1310 INP_WUNLOCK(inp);
1311
1312 /*
1313 * Remember the IP options, if any.
1314 */
1315 #ifdef INET6
1316 if (!(inc->inc_flags & INC_ISIPV6))
1317 #endif
1318 #ifdef INET
1319 ipopts = (m) ? ip_srcroute(m) : NULL;
1320 #else
1321 ipopts = NULL;
1322 #endif
1323
1324 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1325 /*
1326 * If listening socket requested TCP digests, check that received
1327 * SYN has signature and it is correct. If signature doesn't match
1328 * or TCP_SIGNATURE support isn't enabled, drop the packet.
1329 */
1330 if (ltflags & TF_SIGNATURE) {
1331 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1332 TCPSTAT_INC(tcps_sig_err_nosigopt);
1333 goto done;
1334 }
1335 if (!TCPMD5_ENABLED() ||
1336 TCPMD5_INPUT(m, th, to->to_signature) != 0)
1337 goto done;
1338 }
1339 #endif /* TCP_SIGNATURE */
1340 /*
1341 * See if we already have an entry for this connection.
1342 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1343 *
1344 * XXX: should the syncache be re-initialized with the contents
1345 * of the new SYN here (which may have different options?)
1346 *
1347 * XXX: We do not check the sequence number to see if this is a
1348 * real retransmit or a new connection attempt. The question is
1349 * how to handle such a case; either ignore it as spoofed, or
1350 * drop the current entry and create a new one?
1351 */
1352 sc = syncache_lookup(inc, &sch); /* returns locked entry */
1353 SCH_LOCK_ASSERT(sch);
1354 if (sc != NULL) {
1355 #ifdef TCP_RFC7413
1356 if (tfo_cookie_valid)
1357 INP_WUNLOCK(inp);
1358 #endif
1359 TCPSTAT_INC(tcps_sc_dupsyn);
1360 if (ipopts) {
1361 /*
1362 * If we were remembering a previous source route,
1363 * forget it and use the new one we've been given.
1364 */
1365 if (sc->sc_ipopts)
1366 (void) m_free(sc->sc_ipopts);
1367 sc->sc_ipopts = ipopts;
1368 }
1369 /*
1370 * Update timestamp if present.
1371 */
1372 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1373 sc->sc_tsreflect = to->to_tsval;
1374 else
1375 sc->sc_flags &= ~SCF_TIMESTAMP;
1376 #ifdef MAC
1377 /*
1378 * Since we have already unconditionally allocated label
1379 * storage, free it up. The syncache entry will already
1380 * have an initialized label we can use.
1381 */
1382 mac_syncache_destroy(&maclabel);
1383 #endif
1384 /* Retransmit SYN|ACK and reset retransmit count. */
1385 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1386 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1387 "resetting timer and retransmitting SYN|ACK\n",
1388 s, __func__);
1389 free(s, M_TCPLOG);
1390 }
1391 if (syncache_respond(sc, sch, 1, m) == 0) {
1392 sc->sc_rxmits = 0;
1393 syncache_timeout(sc, sch, 1);
1394 TCPSTAT_INC(tcps_sndacks);
1395 TCPSTAT_INC(tcps_sndtotal);
1396 }
1397 SCH_UNLOCK(sch);
1398 goto done;
1399 }
1400
1401 #ifdef TCP_RFC7413
1402 if (tfo_cookie_valid) {
1403 bzero(&scs, sizeof(scs));
1404 sc = &scs;
1405 goto skip_alloc;
1406 }
1407 #endif
1408
1409 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1410 if (sc == NULL) {
1411 /*
1412 * The zone allocator couldn't provide more entries.
1413 * Treat this as if the cache was full; drop the oldest
1414 * entry and insert the new one.
1415 */
1416 TCPSTAT_INC(tcps_sc_zonefail);
1417 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1418 sch->sch_last_overflow = time_uptime;
1419 syncache_drop(sc, sch);
1420 }
1421 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1422 if (sc == NULL) {
1423 if (V_tcp_syncookies) {
1424 bzero(&scs, sizeof(scs));
1425 sc = &scs;
1426 } else {
1427 SCH_UNLOCK(sch);
1428 if (ipopts)
1429 (void) m_free(ipopts);
1430 goto done;
1431 }
1432 }
1433 }
1434
1435 #ifdef TCP_RFC7413
1436 skip_alloc:
1437 if (!tfo_cookie_valid && tfo_response_cookie_valid)
1438 sc->sc_tfo_cookie = &tfo_response_cookie;
1439 #endif
1440
1441 /*
1442 * Fill in the syncache values.
1443 */
1444 #ifdef MAC
1445 sc->sc_label = maclabel;
1446 #endif
1447 sc->sc_cred = cred;
1448 cred = NULL;
1449 sc->sc_ipopts = ipopts;
1450 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1451 #ifdef INET6
1452 if (!(inc->inc_flags & INC_ISIPV6))
1453 #endif
1454 {
1455 sc->sc_ip_tos = ip_tos;
1456 sc->sc_ip_ttl = ip_ttl;
1457 }
1458 #ifdef TCP_OFFLOAD
1459 sc->sc_tod = tod;
1460 sc->sc_todctx = todctx;
1461 #endif
1462 sc->sc_irs = th->th_seq;
1463 sc->sc_iss = arc4random();
1464 sc->sc_flags = 0;
1465 sc->sc_flowlabel = 0;
1466
1467 /*
1468 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1469 * win was derived from socket earlier in the function.
1470 */
1471 win = imax(win, 0);
1472 win = imin(win, TCP_MAXWIN);
1473 sc->sc_wnd = win;
1474
1475 if (V_tcp_do_rfc1323) {
1476 /*
1477 * A timestamp received in a SYN makes
1478 * it ok to send timestamp requests and replies.
1479 */
1480 if (to->to_flags & TOF_TS) {
1481 sc->sc_tsreflect = to->to_tsval;
1482 sc->sc_ts = tcp_ts_getticks();
1483 sc->sc_flags |= SCF_TIMESTAMP;
1484 }
1485 if (to->to_flags & TOF_SCALE) {
1486 int wscale = 0;
1487
1488 /*
1489 * Pick the smallest possible scaling factor that
1490 * will still allow us to scale up to sb_max, aka
1491 * kern.ipc.maxsockbuf.
1492 *
1493 * We do this because there are broken firewalls that
1494 * will corrupt the window scale option, leading to
1495 * the other endpoint believing that our advertised
1496 * window is unscaled. At scale factors larger than
1497 * 5 the unscaled window will drop below 1500 bytes,
1498 * leading to serious problems when traversing these
1499 * broken firewalls.
1500 *
1501 * With the default maxsockbuf of 256K, a scale factor
1502 * of 3 will be chosen by this algorithm. Those who
1503 * choose a larger maxsockbuf should watch out
1504 * for the compatibility problems mentioned above.
1505 *
1506 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1507 * or <SYN,ACK>) segment itself is never scaled.
1508 */
1509 while (wscale < TCP_MAX_WINSHIFT &&
1510 (TCP_MAXWIN << wscale) < sb_max)
1511 wscale++;
1512 sc->sc_requested_r_scale = wscale;
1513 sc->sc_requested_s_scale = to->to_wscale;
1514 sc->sc_flags |= SCF_WINSCALE;
1515 }
1516 }
1517 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1518 /*
1519 * If listening socket requested TCP digests, flag this in the
1520 * syncache so that syncache_respond() will do the right thing
1521 * with the SYN+ACK.
1522 */
1523 if (ltflags & TF_SIGNATURE)
1524 sc->sc_flags |= SCF_SIGNATURE;
1525 #endif /* TCP_SIGNATURE */
1526 if (to->to_flags & TOF_SACKPERM)
1527 sc->sc_flags |= SCF_SACK;
1528 if (to->to_flags & TOF_MSS)
1529 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */
1530 if (ltflags & TF_NOOPT)
1531 sc->sc_flags |= SCF_NOOPT;
1532 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1533 sc->sc_flags |= SCF_ECN;
1534
1535 if (V_tcp_syncookies)
1536 sc->sc_iss = syncookie_generate(sch, sc);
1537 #ifdef INET6
1538 if (autoflowlabel) {
1539 if (V_tcp_syncookies)
1540 sc->sc_flowlabel = sc->sc_iss;
1541 else
1542 sc->sc_flowlabel = ip6_randomflowlabel();
1543 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1544 }
1545 #endif
1546 SCH_UNLOCK(sch);
1547
1548 #ifdef TCP_RFC7413
1549 if (tfo_cookie_valid) {
1550 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie);
1551 /* INP_WUNLOCK(inp) will be performed by the called */
1552 rv = 1;
1553 goto tfo_done;
1554 }
1555 #endif
1556
1557 /*
1558 * Do a standard 3-way handshake.
1559 */
1560 if (syncache_respond(sc, sch, 0, m) == 0) {
1561 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1562 syncache_free(sc);
1563 else if (sc != &scs)
1564 syncache_insert(sc, sch); /* locks and unlocks sch */
1565 TCPSTAT_INC(tcps_sndacks);
1566 TCPSTAT_INC(tcps_sndtotal);
1567 } else {
1568 if (sc != &scs)
1569 syncache_free(sc);
1570 TCPSTAT_INC(tcps_sc_dropped);
1571 }
1572
1573 done:
1574 if (m) {
1575 *lsop = NULL;
1576 m_freem(m);
1577 }
1578 #ifdef TCP_RFC7413
1579 tfo_done:
1580 #endif
1581 if (cred != NULL)
1582 crfree(cred);
1583 #ifdef MAC
1584 if (sc == &scs)
1585 mac_syncache_destroy(&maclabel);
1586 #endif
1587 return (rv);
1588 }
1589
1590 /*
1591 * Send SYN|ACK to the peer. Either in response to the peer's SYN,
1592 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1593 */
1594 static int
1595 syncache_respond(struct syncache *sc, struct syncache_head *sch, int locked,
1596 const struct mbuf *m0)
1597 {
1598 struct ip *ip = NULL;
1599 struct mbuf *m;
1600 struct tcphdr *th = NULL;
1601 int optlen, error = 0; /* Make compiler happy */
1602 u_int16_t hlen, tlen, mssopt;
1603 struct tcpopt to;
1604 #ifdef INET6
1605 struct ip6_hdr *ip6 = NULL;
1606 #endif
1607 hlen =
1608 #ifdef INET6
1609 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1610 #endif
1611 sizeof(struct ip);
1612 tlen = hlen + sizeof(struct tcphdr);
1613
1614 /* Determine MSS we advertize to other end of connection. */
1615 mssopt = tcp_mssopt(&sc->sc_inc);
1616 if (sc->sc_peer_mss)
1617 mssopt = max( min(sc->sc_peer_mss, mssopt), V_tcp_minmss);
1618
1619 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1620 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1621 ("syncache: mbuf too small"));
1622
1623 /* Create the IP+TCP header from scratch. */
1624 m = m_gethdr(M_NOWAIT, MT_DATA);
1625 if (m == NULL)
1626 return (ENOBUFS);
1627 #ifdef MAC
1628 mac_syncache_create_mbuf(sc->sc_label, m);
1629 #endif
1630 m->m_data += max_linkhdr;
1631 m->m_len = tlen;
1632 m->m_pkthdr.len = tlen;
1633 m->m_pkthdr.rcvif = NULL;
1634
1635 #ifdef INET6
1636 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1637 ip6 = mtod(m, struct ip6_hdr *);
1638 ip6->ip6_vfc = IPV6_VERSION;
1639 ip6->ip6_nxt = IPPROTO_TCP;
1640 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1641 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1642 ip6->ip6_plen = htons(tlen - hlen);
1643 /* ip6_hlim is set after checksum */
1644 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1645 ip6->ip6_flow |= sc->sc_flowlabel;
1646
1647 th = (struct tcphdr *)(ip6 + 1);
1648 }
1649 #endif
1650 #if defined(INET6) && defined(INET)
1651 else
1652 #endif
1653 #ifdef INET
1654 {
1655 ip = mtod(m, struct ip *);
1656 ip->ip_v = IPVERSION;
1657 ip->ip_hl = sizeof(struct ip) >> 2;
1658 ip->ip_len = htons(tlen);
1659 ip->ip_id = 0;
1660 ip->ip_off = 0;
1661 ip->ip_sum = 0;
1662 ip->ip_p = IPPROTO_TCP;
1663 ip->ip_src = sc->sc_inc.inc_laddr;
1664 ip->ip_dst = sc->sc_inc.inc_faddr;
1665 ip->ip_ttl = sc->sc_ip_ttl;
1666 ip->ip_tos = sc->sc_ip_tos;
1667
1668 /*
1669 * See if we should do MTU discovery. Route lookups are
1670 * expensive, so we will only unset the DF bit if:
1671 *
1672 * 1) path_mtu_discovery is disabled
1673 * 2) the SCF_UNREACH flag has been set
1674 */
1675 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1676 ip->ip_off |= htons(IP_DF);
1677
1678 th = (struct tcphdr *)(ip + 1);
1679 }
1680 #endif /* INET */
1681 th->th_sport = sc->sc_inc.inc_lport;
1682 th->th_dport = sc->sc_inc.inc_fport;
1683
1684 th->th_seq = htonl(sc->sc_iss);
1685 th->th_ack = htonl(sc->sc_irs + 1);
1686 th->th_off = sizeof(struct tcphdr) >> 2;
1687 th->th_x2 = 0;
1688 th->th_flags = TH_SYN|TH_ACK;
1689 th->th_win = htons(sc->sc_wnd);
1690 th->th_urp = 0;
1691
1692 if (sc->sc_flags & SCF_ECN) {
1693 th->th_flags |= TH_ECE;
1694 TCPSTAT_INC(tcps_ecn_shs);
1695 }
1696
1697 /* Tack on the TCP options. */
1698 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1699 to.to_flags = 0;
1700
1701 to.to_mss = mssopt;
1702 to.to_flags = TOF_MSS;
1703 if (sc->sc_flags & SCF_WINSCALE) {
1704 to.to_wscale = sc->sc_requested_r_scale;
1705 to.to_flags |= TOF_SCALE;
1706 }
1707 if (sc->sc_flags & SCF_TIMESTAMP) {
1708 /* Virgin timestamp or TCP cookie enhanced one. */
1709 to.to_tsval = sc->sc_ts;
1710 to.to_tsecr = sc->sc_tsreflect;
1711 to.to_flags |= TOF_TS;
1712 }
1713 if (sc->sc_flags & SCF_SACK)
1714 to.to_flags |= TOF_SACKPERM;
1715 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1716 if (sc->sc_flags & SCF_SIGNATURE)
1717 to.to_flags |= TOF_SIGNATURE;
1718 #endif
1719 #ifdef TCP_RFC7413
1720 if (sc->sc_tfo_cookie) {
1721 to.to_flags |= TOF_FASTOPEN;
1722 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1723 to.to_tfo_cookie = sc->sc_tfo_cookie;
1724 /* don't send cookie again when retransmitting response */
1725 sc->sc_tfo_cookie = NULL;
1726 }
1727 #endif
1728 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1729
1730 /* Adjust headers by option size. */
1731 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1732 m->m_len += optlen;
1733 m->m_pkthdr.len += optlen;
1734 #ifdef INET6
1735 if (sc->sc_inc.inc_flags & INC_ISIPV6)
1736 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1737 else
1738 #endif
1739 ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1740 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1741 if (sc->sc_flags & SCF_SIGNATURE) {
1742 KASSERT(to.to_flags & TOF_SIGNATURE,
1743 ("tcp_addoptions() didn't set tcp_signature"));
1744
1745 /* NOTE: to.to_signature is inside of mbuf */
1746 if (!TCPMD5_ENABLED() ||
1747 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1748 m_freem(m);
1749 return (EACCES);
1750 }
1751 }
1752 #endif
1753 } else
1754 optlen = 0;
1755
1756 M_SETFIB(m, sc->sc_inc.inc_fibnum);
1757 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1758 /*
1759 * If we have peer's SYN and it has a flowid, then let's assign it to
1760 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid
1761 * to SYN|ACK due to lack of inp here.
1762 */
1763 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1764 m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1765 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1766 }
1767 #ifdef INET6
1768 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1769 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1770 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1771 IPPROTO_TCP, 0);
1772 ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1773 #ifdef TCP_OFFLOAD
1774 if (ADDED_BY_TOE(sc)) {
1775 struct toedev *tod = sc->sc_tod;
1776
1777 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1778
1779 return (error);
1780 }
1781 #endif
1782 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1783 }
1784 #endif
1785 #if defined(INET6) && defined(INET)
1786 else
1787 #endif
1788 #ifdef INET
1789 {
1790 m->m_pkthdr.csum_flags = CSUM_TCP;
1791 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1792 htons(tlen + optlen - hlen + IPPROTO_TCP));
1793 #ifdef TCP_OFFLOAD
1794 if (ADDED_BY_TOE(sc)) {
1795 struct toedev *tod = sc->sc_tod;
1796
1797 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1798
1799 return (error);
1800 }
1801 #endif
1802 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1803 }
1804 #endif
1805 return (error);
1806 }
1807
1808 /*
1809 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1810 * that exceed the capacity of the syncache by avoiding the storage of any
1811 * of the SYNs we receive. Syncookies defend against blind SYN flooding
1812 * attacks where the attacker does not have access to our responses.
1813 *
1814 * Syncookies encode and include all necessary information about the
1815 * connection setup within the SYN|ACK that we send back. That way we
1816 * can avoid keeping any local state until the ACK to our SYN|ACK returns
1817 * (if ever). Normally the syncache and syncookies are running in parallel
1818 * with the latter taking over when the former is exhausted. When matching
1819 * syncache entry is found the syncookie is ignored.
1820 *
1821 * The only reliable information persisting the 3WHS is our initial sequence
1822 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient
1823 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1824 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK
1825 * returns and signifies a legitimate connection if it matches the ACK.
1826 *
1827 * The available space of 32 bits to store the hash and to encode the SYN
1828 * option information is very tight and we should have at least 24 bits for
1829 * the MAC to keep the number of guesses by blind spoofing reasonably high.
1830 *
1831 * SYN option information we have to encode to fully restore a connection:
1832 * MSS: is imporant to chose an optimal segment size to avoid IP level
1833 * fragmentation along the path. The common MSS values can be encoded
1834 * in a 3-bit table. Uncommon values are captured by the next lower value
1835 * in the table leading to a slight increase in packetization overhead.
1836 * WSCALE: is necessary to allow large windows to be used for high delay-
1837 * bandwidth product links. Not scaling the window when it was initially
1838 * negotiated is bad for performance as lack of scaling further decreases
1839 * the apparent available send window. We only need to encode the WSCALE
1840 * we received from the remote end. Our end can be recalculated at any
1841 * time. The common WSCALE values can be encoded in a 3-bit table.
1842 * Uncommon values are captured by the next lower value in the table
1843 * making us under-estimate the available window size halving our
1844 * theoretically possible maximum throughput for that connection.
1845 * SACK: Greatly assists in packet loss recovery and requires 1 bit.
1846 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
1847 * that are included in all segments on a connection. We enable them when
1848 * the ACK has them.
1849 *
1850 * Security of syncookies and attack vectors:
1851 *
1852 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
1853 * together with the gloabl secret to make it unique per connection attempt.
1854 * Thus any change of any of those parameters results in a different MAC output
1855 * in an unpredictable way unless a collision is encountered. 24 bits of the
1856 * MAC are embedded into the ISS.
1857 *
1858 * To prevent replay attacks two rotating global secrets are updated with a
1859 * new random value every 15 seconds. The life-time of a syncookie is thus
1860 * 15-30 seconds.
1861 *
1862 * Vector 1: Attacking the secret. This requires finding a weakness in the
1863 * MAC itself or the way it is used here. The attacker can do a chosen plain
1864 * text attack by varying and testing the all parameters under his control.
1865 * The strength depends on the size and randomness of the secret, and the
1866 * cryptographic security of the MAC function. Due to the constant updating
1867 * of the secret the attacker has at most 29.999 seconds to find the secret
1868 * and launch spoofed connections. After that he has to start all over again.
1869 *
1870 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC
1871 * size an average of 4,823 attempts are required for a 50% chance of success
1872 * to spoof a single syncookie (birthday collision paradox). However the
1873 * attacker is blind and doesn't know if one of his attempts succeeded unless
1874 * he has a side channel to interfere success from. A single connection setup
1875 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
1876 * This many attempts are required for each one blind spoofed connection. For
1877 * every additional spoofed connection he has to launch another N attempts.
1878 * Thus for a sustained rate 100 spoofed connections per second approximately
1879 * 1,800,000 packets per second would have to be sent.
1880 *
1881 * NB: The MAC function should be fast so that it doesn't become a CPU
1882 * exhaustion attack vector itself.
1883 *
1884 * References:
1885 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations
1886 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
1887 * http://cr.yp.to/syncookies.html (overview)
1888 * http://cr.yp.to/syncookies/archive (details)
1889 *
1890 *
1891 * Schematic construction of a syncookie enabled Initial Sequence Number:
1892 * 0 1 2 3
1893 * 12345678901234567890123456789012
1894 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
1895 *
1896 * x 24 MAC (truncated)
1897 * W 3 Send Window Scale index
1898 * M 3 MSS index
1899 * S 1 SACK permitted
1900 * P 1 Odd/even secret
1901 */
1902
1903 /*
1904 * Distribution and probability of certain MSS values. Those in between are
1905 * rounded down to the next lower one.
1906 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
1907 * .2% .3% 5% 7% 7% 20% 15% 45%
1908 */
1909 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
1910
1911 /*
1912 * Distribution and probability of certain WSCALE values. We have to map the
1913 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
1914 * bits based on prevalence of certain values. Where we don't have an exact
1915 * match for are rounded down to the next lower one letting us under-estimate
1916 * the true available window. At the moment this would happen only for the
1917 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
1918 * and window size). The absence of the WSCALE option (no scaling in either
1919 * direction) is encoded with index zero.
1920 * [WSCALE values histograms, Allman, 2012]
1921 * X 10 10 35 5 6 14 10% by host
1922 * X 11 4 5 5 18 49 3% by connections
1923 */
1924 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
1925
1926 /*
1927 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed
1928 * and good cryptographic properties.
1929 */
1930 static uint32_t
1931 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
1932 uint8_t *secbits, uintptr_t secmod)
1933 {
1934 SIPHASH_CTX ctx;
1935 uint32_t siphash[2];
1936
1937 SipHash24_Init(&ctx);
1938 SipHash_SetKey(&ctx, secbits);
1939 switch (inc->inc_flags & INC_ISIPV6) {
1940 #ifdef INET
1941 case 0:
1942 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
1943 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
1944 break;
1945 #endif
1946 #ifdef INET6
1947 case INC_ISIPV6:
1948 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
1949 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
1950 break;
1951 #endif
1952 }
1953 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
1954 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
1955 SipHash_Update(&ctx, &irs, sizeof(irs));
1956 SipHash_Update(&ctx, &flags, sizeof(flags));
1957 SipHash_Update(&ctx, &secmod, sizeof(secmod));
1958 SipHash_Final((u_int8_t *)&siphash, &ctx);
1959
1960 return (siphash[0] ^ siphash[1]);
1961 }
1962
1963 static tcp_seq
1964 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
1965 {
1966 u_int i, mss, secbit, wscale;
1967 uint32_t iss, hash;
1968 uint8_t *secbits;
1969 union syncookie cookie;
1970
1971 SCH_LOCK_ASSERT(sch);
1972
1973 cookie.cookie = 0;
1974
1975 /* Map our computed MSS into the 3-bit index. */
1976 mss = min(tcp_mssopt(&sc->sc_inc), max(sc->sc_peer_mss, V_tcp_minmss));
1977 for (i = nitems(tcp_sc_msstab) - 1; tcp_sc_msstab[i] > mss && i > 0;
1978 i--)
1979 ;
1980 cookie.flags.mss_idx = i;
1981
1982 /*
1983 * Map the send window scale into the 3-bit index but only if
1984 * the wscale option was received.
1985 */
1986 if (sc->sc_flags & SCF_WINSCALE) {
1987 wscale = sc->sc_requested_s_scale;
1988 for (i = nitems(tcp_sc_wstab) - 1;
1989 tcp_sc_wstab[i] > wscale && i > 0;
1990 i--)
1991 ;
1992 cookie.flags.wscale_idx = i;
1993 }
1994
1995 /* Can we do SACK? */
1996 if (sc->sc_flags & SCF_SACK)
1997 cookie.flags.sack_ok = 1;
1998
1999 /* Which of the two secrets to use. */
2000 secbit = sch->sch_sc->secret.oddeven & 0x1;
2001 cookie.flags.odd_even = secbit;
2002
2003 secbits = sch->sch_sc->secret.key[secbit];
2004 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2005 (uintptr_t)sch);
2006
2007 /*
2008 * Put the flags into the hash and XOR them to get better ISS number
2009 * variance. This doesn't enhance the cryptographic strength and is
2010 * done to prevent the 8 cookie bits from showing up directly on the
2011 * wire.
2012 */
2013 iss = hash & ~0xff;
2014 iss |= cookie.cookie ^ (hash >> 24);
2015
2016 /* Randomize the timestamp. */
2017 if (sc->sc_flags & SCF_TIMESTAMP) {
2018 sc->sc_ts = arc4random();
2019 sc->sc_tsoff = sc->sc_ts - tcp_ts_getticks();
2020 }
2021
2022 TCPSTAT_INC(tcps_sc_sendcookie);
2023 return (iss);
2024 }
2025
2026 static struct syncache *
2027 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2028 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2029 struct socket *lso)
2030 {
2031 uint32_t hash;
2032 uint8_t *secbits;
2033 tcp_seq ack, seq;
2034 int wnd, wscale = 0;
2035 union syncookie cookie;
2036
2037 SCH_LOCK_ASSERT(sch);
2038
2039 /*
2040 * Pull information out of SYN-ACK/ACK and revert sequence number
2041 * advances.
2042 */
2043 ack = th->th_ack - 1;
2044 seq = th->th_seq - 1;
2045
2046 /*
2047 * Unpack the flags containing enough information to restore the
2048 * connection.
2049 */
2050 cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2051
2052 /* Which of the two secrets to use. */
2053 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
2054
2055 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2056
2057 /* The recomputed hash matches the ACK if this was a genuine cookie. */
2058 if ((ack & ~0xff) != (hash & ~0xff))
2059 return (NULL);
2060
2061 /* Fill in the syncache values. */
2062 sc->sc_flags = 0;
2063 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2064 sc->sc_ipopts = NULL;
2065
2066 sc->sc_irs = seq;
2067 sc->sc_iss = ack;
2068
2069 switch (inc->inc_flags & INC_ISIPV6) {
2070 #ifdef INET
2071 case 0:
2072 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2073 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2074 break;
2075 #endif
2076 #ifdef INET6
2077 case INC_ISIPV6:
2078 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2079 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
2080 break;
2081 #endif
2082 }
2083
2084 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2085
2086 /* We can simply recompute receive window scale we sent earlier. */
2087 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2088 wscale++;
2089
2090 /* Only use wscale if it was enabled in the orignal SYN. */
2091 if (cookie.flags.wscale_idx > 0) {
2092 sc->sc_requested_r_scale = wscale;
2093 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2094 sc->sc_flags |= SCF_WINSCALE;
2095 }
2096
2097 wnd = sbspace(&lso->so_rcv);
2098 wnd = imax(wnd, 0);
2099 wnd = imin(wnd, TCP_MAXWIN);
2100 sc->sc_wnd = wnd;
2101
2102 if (cookie.flags.sack_ok)
2103 sc->sc_flags |= SCF_SACK;
2104
2105 if (to->to_flags & TOF_TS) {
2106 sc->sc_flags |= SCF_TIMESTAMP;
2107 sc->sc_tsreflect = to->to_tsval;
2108 sc->sc_ts = to->to_tsecr;
2109 sc->sc_tsoff = to->to_tsecr - tcp_ts_getticks();
2110 }
2111
2112 if (to->to_flags & TOF_SIGNATURE)
2113 sc->sc_flags |= SCF_SIGNATURE;
2114
2115 sc->sc_rxmits = 0;
2116
2117 TCPSTAT_INC(tcps_sc_recvcookie);
2118 return (sc);
2119 }
2120
2121 #ifdef INVARIANTS
2122 static int
2123 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2124 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2125 struct socket *lso)
2126 {
2127 struct syncache scs, *scx;
2128 char *s;
2129
2130 bzero(&scs, sizeof(scs));
2131 scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
2132
2133 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2134 return (0);
2135
2136 if (scx != NULL) {
2137 if (sc->sc_peer_mss != scx->sc_peer_mss)
2138 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2139 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2140
2141 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2142 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2143 s, __func__, sc->sc_requested_r_scale,
2144 scx->sc_requested_r_scale);
2145
2146 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2147 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2148 s, __func__, sc->sc_requested_s_scale,
2149 scx->sc_requested_s_scale);
2150
2151 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2152 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2153 }
2154
2155 if (s != NULL)
2156 free(s, M_TCPLOG);
2157 return (0);
2158 }
2159 #endif /* INVARIANTS */
2160
2161 static void
2162 syncookie_reseed(void *arg)
2163 {
2164 struct tcp_syncache *sc = arg;
2165 uint8_t *secbits;
2166 int secbit;
2167
2168 /*
2169 * Reseeding the secret doesn't have to be protected by a lock.
2170 * It only must be ensured that the new random values are visible
2171 * to all CPUs in a SMP environment. The atomic with release
2172 * semantics ensures that.
2173 */
2174 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2175 secbits = sc->secret.key[secbit];
2176 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2177 atomic_add_rel_int(&sc->secret.oddeven, 1);
2178
2179 /* Reschedule ourself. */
2180 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2181 }
2182
2183 /*
2184 * Exports the syncache entries to userland so that netstat can display
2185 * them alongside the other sockets. This function is intended to be
2186 * called only from tcp_pcblist.
2187 *
2188 * Due to concurrency on an active system, the number of pcbs exported
2189 * may have no relation to max_pcbs. max_pcbs merely indicates the
2190 * amount of space the caller allocated for this function to use.
2191 */
2192 int
2193 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
2194 {
2195 struct xtcpcb xt;
2196 struct syncache *sc;
2197 struct syncache_head *sch;
2198 int count, error, i;
2199
2200 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
2201 sch = &V_tcp_syncache.hashbase[i];
2202 SCH_LOCK(sch);
2203 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2204 if (count >= max_pcbs) {
2205 SCH_UNLOCK(sch);
2206 goto exit;
2207 }
2208 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2209 continue;
2210 bzero(&xt, sizeof(xt));
2211 xt.xt_len = sizeof(xt);
2212 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2213 xt.xt_inp.inp_vflag = INP_IPV6;
2214 else
2215 xt.xt_inp.inp_vflag = INP_IPV4;
2216 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo));
2217 xt.xt_tp.t_inpcb = &xt.xt_inp;
2218 xt.xt_tp.t_state = TCPS_SYN_RECEIVED;
2219 xt.xt_socket.xso_protocol = IPPROTO_TCP;
2220 xt.xt_socket.xso_len = sizeof (struct xsocket);
2221 xt.xt_socket.so_type = SOCK_STREAM;
2222 xt.xt_socket.so_state = SS_ISCONNECTING;
2223 error = SYSCTL_OUT(req, &xt, sizeof xt);
2224 if (error) {
2225 SCH_UNLOCK(sch);
2226 goto exit;
2227 }
2228 count++;
2229 }
2230 SCH_UNLOCK(sch);
2231 }
2232 exit:
2233 *pcbs_exported = count;
2234 return error;
2235 }
Cache object: 4cc887cd9b71f42403ade666c76e365f
|