1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3 *
4 * Copyright (c) 2001 McAfee, Inc.
5 * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG
6 * All rights reserved.
7 *
8 * This software was developed for the FreeBSD Project by Jonathan Lemon
9 * and McAfee Research, the Security Research Division of McAfee, Inc. under
10 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
11 * DARPA CHATS research program. [2001 McAfee, Inc.]
12 *
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
15 * are met:
16 * 1. Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution.
21 *
22 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * SUCH DAMAGE.
33 */
34
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD: releng/12.0/sys/netinet/tcp_syncache.c 339430 2018-10-18 19:21:18Z tuexen $");
37
38 #include "opt_inet.h"
39 #include "opt_inet6.h"
40 #include "opt_ipsec.h"
41 #include "opt_pcbgroup.h"
42
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/hash.h>
46 #include <sys/refcount.h>
47 #include <sys/kernel.h>
48 #include <sys/sysctl.h>
49 #include <sys/limits.h>
50 #include <sys/lock.h>
51 #include <sys/mutex.h>
52 #include <sys/malloc.h>
53 #include <sys/mbuf.h>
54 #include <sys/proc.h> /* for proc0 declaration */
55 #include <sys/random.h>
56 #include <sys/socket.h>
57 #include <sys/socketvar.h>
58 #include <sys/syslog.h>
59 #include <sys/ucred.h>
60
61 #include <sys/md5.h>
62 #include <crypto/siphash/siphash.h>
63
64 #include <vm/uma.h>
65
66 #include <net/if.h>
67 #include <net/if_var.h>
68 #include <net/route.h>
69 #include <net/vnet.h>
70
71 #include <netinet/in.h>
72 #include <netinet/in_kdtrace.h>
73 #include <netinet/in_systm.h>
74 #include <netinet/ip.h>
75 #include <netinet/in_var.h>
76 #include <netinet/in_pcb.h>
77 #include <netinet/ip_var.h>
78 #include <netinet/ip_options.h>
79 #ifdef INET6
80 #include <netinet/ip6.h>
81 #include <netinet/icmp6.h>
82 #include <netinet6/nd6.h>
83 #include <netinet6/ip6_var.h>
84 #include <netinet6/in6_pcb.h>
85 #endif
86 #include <netinet/tcp.h>
87 #include <netinet/tcp_fastopen.h>
88 #include <netinet/tcp_fsm.h>
89 #include <netinet/tcp_seq.h>
90 #include <netinet/tcp_timer.h>
91 #include <netinet/tcp_var.h>
92 #include <netinet/tcp_syncache.h>
93 #ifdef INET6
94 #include <netinet6/tcp6_var.h>
95 #endif
96 #ifdef TCP_OFFLOAD
97 #include <netinet/toecore.h>
98 #endif
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(int, tcp_syncookies) = 1;
107 #define V_tcp_syncookies VNET(tcp_syncookies)
108 SYSCTL_INT(_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(int, tcp_syncookiesonly) = 0;
113 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly)
114 SYSCTL_INT(_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 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1;
119 #define V_functions_inherit_listen_socket_stack \
120 VNET(functions_inherit_listen_socket_stack)
121 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack,
122 CTLFLAG_VNET | CTLFLAG_RW,
123 &VNET_NAME(functions_inherit_listen_socket_stack), 0,
124 "Inherit listen socket's stack");
125
126 #ifdef TCP_OFFLOAD
127 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
128 #endif
129
130 static void syncache_drop(struct syncache *, struct syncache_head *);
131 static void syncache_free(struct syncache *);
132 static void syncache_insert(struct syncache *, struct syncache_head *);
133 static int syncache_respond(struct syncache *, struct syncache_head *,
134 const struct mbuf *, int);
135 static struct socket *syncache_socket(struct syncache *, struct socket *,
136 struct mbuf *m);
137 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch,
138 int docallout);
139 static void syncache_timer(void *);
140
141 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
142 uint8_t *, uintptr_t);
143 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *);
144 static struct syncache
145 *syncookie_lookup(struct in_conninfo *, struct syncache_head *,
146 struct syncache *, struct tcphdr *, struct tcpopt *,
147 struct socket *);
148 static void syncookie_reseed(void *);
149 #ifdef INVARIANTS
150 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
151 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
152 struct socket *lso);
153 #endif
154
155 /*
156 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
157 * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds,
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, CTLFLAG_RW, 0,
170 "TCP SYN cache");
171
172 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
173 &VNET_NAME(tcp_syncache.bucket_limit), 0,
174 "Per-bucket hash limit for syncache");
175
176 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
177 &VNET_NAME(tcp_syncache.cache_limit), 0,
178 "Overall entry limit for syncache");
179
180 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
181 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
182
183 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
184 &VNET_NAME(tcp_syncache.hashsize), 0,
185 "Size of TCP syncache hashtable");
186
187 static int
188 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
189 {
190 int error;
191 u_int new;
192
193 new = V_tcp_syncache.rexmt_limit;
194 error = sysctl_handle_int(oidp, &new, 0, req);
195 if ((error == 0) && (req->newptr != NULL)) {
196 if (new > TCP_MAXRXTSHIFT)
197 error = EINVAL;
198 else
199 V_tcp_syncache.rexmt_limit = new;
200 }
201 return (error);
202 }
203
204 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
205 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW,
206 &VNET_NAME(tcp_syncache.rexmt_limit), 0,
207 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
208 "Limit on SYN/ACK retransmissions");
209
210 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
211 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
212 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
213 "Send reset on socket allocation failure");
214
215 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
216
217 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx)
218 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx)
219 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED)
220
221 /*
222 * Requires the syncache entry to be already removed from the bucket list.
223 */
224 static void
225 syncache_free(struct syncache *sc)
226 {
227
228 if (sc->sc_ipopts)
229 (void) m_free(sc->sc_ipopts);
230 if (sc->sc_cred)
231 crfree(sc->sc_cred);
232 #ifdef MAC
233 mac_syncache_destroy(&sc->sc_label);
234 #endif
235
236 uma_zfree(V_tcp_syncache.zone, sc);
237 }
238
239 void
240 syncache_init(void)
241 {
242 int i;
243
244 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
245 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
246 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
247 V_tcp_syncache.hash_secret = arc4random();
248
249 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
250 &V_tcp_syncache.hashsize);
251 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
252 &V_tcp_syncache.bucket_limit);
253 if (!powerof2(V_tcp_syncache.hashsize) ||
254 V_tcp_syncache.hashsize == 0) {
255 printf("WARNING: syncache hash size is not a power of 2.\n");
256 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
257 }
258 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
259
260 /* Set limits. */
261 V_tcp_syncache.cache_limit =
262 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
263 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
264 &V_tcp_syncache.cache_limit);
265
266 /* Allocate the hash table. */
267 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
268 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
269
270 #ifdef VIMAGE
271 V_tcp_syncache.vnet = curvnet;
272 #endif
273
274 /* Initialize the hash buckets. */
275 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
276 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
277 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
278 NULL, MTX_DEF);
279 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
280 &V_tcp_syncache.hashbase[i].sch_mtx, 0);
281 V_tcp_syncache.hashbase[i].sch_length = 0;
282 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
283 V_tcp_syncache.hashbase[i].sch_last_overflow =
284 -(SYNCOOKIE_LIFETIME + 1);
285 }
286
287 /* Create the syncache entry zone. */
288 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
289 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
290 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
291 V_tcp_syncache.cache_limit);
292
293 /* Start the SYN cookie reseeder callout. */
294 callout_init(&V_tcp_syncache.secret.reseed, 1);
295 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
296 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
297 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
298 syncookie_reseed, &V_tcp_syncache);
299 }
300
301 #ifdef VIMAGE
302 void
303 syncache_destroy(void)
304 {
305 struct syncache_head *sch;
306 struct syncache *sc, *nsc;
307 int i;
308
309 /*
310 * Stop the re-seed timer before freeing resources. No need to
311 * possibly schedule it another time.
312 */
313 callout_drain(&V_tcp_syncache.secret.reseed);
314
315 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */
316 for (i = 0; i < V_tcp_syncache.hashsize; i++) {
317
318 sch = &V_tcp_syncache.hashbase[i];
319 callout_drain(&sch->sch_timer);
320
321 SCH_LOCK(sch);
322 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
323 syncache_drop(sc, sch);
324 SCH_UNLOCK(sch);
325 KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
326 ("%s: sch->sch_bucket not empty", __func__));
327 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
328 __func__, sch->sch_length));
329 mtx_destroy(&sch->sch_mtx);
330 }
331
332 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
333 ("%s: cache_count not 0", __func__));
334
335 /* Free the allocated global resources. */
336 uma_zdestroy(V_tcp_syncache.zone);
337 free(V_tcp_syncache.hashbase, M_SYNCACHE);
338 }
339 #endif
340
341 /*
342 * Inserts a syncache entry into the specified bucket row.
343 * Locks and unlocks the syncache_head autonomously.
344 */
345 static void
346 syncache_insert(struct syncache *sc, struct syncache_head *sch)
347 {
348 struct syncache *sc2;
349
350 SCH_LOCK(sch);
351
352 /*
353 * Make sure that we don't overflow the per-bucket limit.
354 * If the bucket is full, toss the oldest element.
355 */
356 if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
357 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
358 ("sch->sch_length incorrect"));
359 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
360 sch->sch_last_overflow = time_uptime;
361 syncache_drop(sc2, sch);
362 TCPSTAT_INC(tcps_sc_bucketoverflow);
363 }
364
365 /* Put it into the bucket. */
366 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
367 sch->sch_length++;
368
369 #ifdef TCP_OFFLOAD
370 if (ADDED_BY_TOE(sc)) {
371 struct toedev *tod = sc->sc_tod;
372
373 tod->tod_syncache_added(tod, sc->sc_todctx);
374 }
375 #endif
376
377 /* Reinitialize the bucket row's timer. */
378 if (sch->sch_length == 1)
379 sch->sch_nextc = ticks + INT_MAX;
380 syncache_timeout(sc, sch, 1);
381
382 SCH_UNLOCK(sch);
383
384 TCPSTATES_INC(TCPS_SYN_RECEIVED);
385 TCPSTAT_INC(tcps_sc_added);
386 }
387
388 /*
389 * Remove and free entry from syncache bucket row.
390 * Expects locked syncache head.
391 */
392 static void
393 syncache_drop(struct syncache *sc, struct syncache_head *sch)
394 {
395
396 SCH_LOCK_ASSERT(sch);
397
398 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
399 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
400 sch->sch_length--;
401
402 #ifdef TCP_OFFLOAD
403 if (ADDED_BY_TOE(sc)) {
404 struct toedev *tod = sc->sc_tod;
405
406 tod->tod_syncache_removed(tod, sc->sc_todctx);
407 }
408 #endif
409
410 syncache_free(sc);
411 }
412
413 /*
414 * Engage/reengage time on bucket row.
415 */
416 static void
417 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
418 {
419 int rexmt;
420
421 if (sc->sc_rxmits == 0)
422 rexmt = TCPTV_RTOBASE;
423 else
424 TCPT_RANGESET(rexmt, TCPTV_RTOBASE * tcp_syn_backoff[sc->sc_rxmits],
425 tcp_rexmit_min, TCPTV_REXMTMAX);
426 sc->sc_rxttime = ticks + rexmt;
427 sc->sc_rxmits++;
428 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
429 sch->sch_nextc = sc->sc_rxttime;
430 if (docallout)
431 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
432 syncache_timer, (void *)sch);
433 }
434 }
435
436 /*
437 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
438 * If we have retransmitted an entry the maximum number of times, expire it.
439 * One separate timer for each bucket row.
440 */
441 static void
442 syncache_timer(void *xsch)
443 {
444 struct syncache_head *sch = (struct syncache_head *)xsch;
445 struct syncache *sc, *nsc;
446 int tick = ticks;
447 char *s;
448
449 CURVNET_SET(sch->sch_sc->vnet);
450
451 /* NB: syncache_head has already been locked by the callout. */
452 SCH_LOCK_ASSERT(sch);
453
454 /*
455 * In the following cycle we may remove some entries and/or
456 * advance some timeouts, so re-initialize the bucket timer.
457 */
458 sch->sch_nextc = tick + INT_MAX;
459
460 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
461 /*
462 * We do not check if the listen socket still exists
463 * and accept the case where the listen socket may be
464 * gone by the time we resend the SYN/ACK. We do
465 * not expect this to happens often. If it does,
466 * then the RST will be sent by the time the remote
467 * host does the SYN/ACK->ACK.
468 */
469 if (TSTMP_GT(sc->sc_rxttime, tick)) {
470 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
471 sch->sch_nextc = sc->sc_rxttime;
472 continue;
473 }
474 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
475 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
476 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
477 "giving up and removing syncache entry\n",
478 s, __func__);
479 free(s, M_TCPLOG);
480 }
481 syncache_drop(sc, sch);
482 TCPSTAT_INC(tcps_sc_stale);
483 continue;
484 }
485 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
486 log(LOG_DEBUG, "%s; %s: Response timeout, "
487 "retransmitting (%u) SYN|ACK\n",
488 s, __func__, sc->sc_rxmits);
489 free(s, M_TCPLOG);
490 }
491
492 syncache_respond(sc, sch, NULL, TH_SYN|TH_ACK);
493 TCPSTAT_INC(tcps_sc_retransmitted);
494 syncache_timeout(sc, sch, 0);
495 }
496 if (!TAILQ_EMPTY(&(sch)->sch_bucket))
497 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
498 syncache_timer, (void *)(sch));
499 CURVNET_RESTORE();
500 }
501
502 /*
503 * Find an entry in the syncache.
504 * Returns always with locked syncache_head plus a matching entry or NULL.
505 */
506 static struct syncache *
507 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
508 {
509 struct syncache *sc;
510 struct syncache_head *sch;
511 uint32_t hash;
512
513 /*
514 * The hash is built on foreign port + local port + foreign address.
515 * We rely on the fact that struct in_conninfo starts with 16 bits
516 * of foreign port, then 16 bits of local port then followed by 128
517 * bits of foreign address. In case of IPv4 address, the first 3
518 * 32-bit words of the address always are zeroes.
519 */
520 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
521 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
522
523 sch = &V_tcp_syncache.hashbase[hash];
524 *schp = sch;
525 SCH_LOCK(sch);
526
527 /* Circle through bucket row to find matching entry. */
528 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
529 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
530 sizeof(struct in_endpoints)) == 0)
531 break;
532
533 return (sc); /* Always returns with locked sch. */
534 }
535
536 /*
537 * This function is called when we get a RST for a
538 * non-existent connection, so that we can see if the
539 * connection is in the syn cache. If it is, zap it.
540 * If required send a challenge ACK.
541 */
542 void
543 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m)
544 {
545 struct syncache *sc;
546 struct syncache_head *sch;
547 char *s = NULL;
548
549 sc = syncache_lookup(inc, &sch); /* returns locked sch */
550 SCH_LOCK_ASSERT(sch);
551
552 /*
553 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
554 * See RFC 793 page 65, section SEGMENT ARRIVES.
555 */
556 if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
557 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
558 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
559 "FIN flag set, segment ignored\n", s, __func__);
560 TCPSTAT_INC(tcps_badrst);
561 goto done;
562 }
563
564 /*
565 * No corresponding connection was found in syncache.
566 * If syncookies are enabled and possibly exclusively
567 * used, or we are under memory pressure, a valid RST
568 * may not find a syncache entry. In that case we're
569 * done and no SYN|ACK retransmissions will happen.
570 * Otherwise the RST was misdirected or spoofed.
571 */
572 if (sc == NULL) {
573 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
574 log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
575 "syncache entry (possibly syncookie only), "
576 "segment ignored\n", s, __func__);
577 TCPSTAT_INC(tcps_badrst);
578 goto done;
579 }
580
581 /*
582 * If the RST bit is set, check the sequence number to see
583 * if this is a valid reset segment.
584 * RFC 793 page 37:
585 * In all states except SYN-SENT, all reset (RST) segments
586 * are validated by checking their SEQ-fields. A reset is
587 * valid if its sequence number is in the window.
588 *
589 * The sequence number in the reset segment is normally an
590 * echo of our outgoing acknowlegement numbers, but some hosts
591 * send a reset with the sequence number at the rightmost edge
592 * of our receive window, and we have to handle this case.
593 */
594 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
595 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
596 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
597 if (V_tcp_insecure_rst ||
598 th->th_seq == sc->sc_irs + 1) {
599 syncache_drop(sc, sch);
600 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
601 log(LOG_DEBUG,
602 "%s; %s: Our SYN|ACK was rejected, "
603 "connection attempt aborted by remote "
604 "endpoint\n",
605 s, __func__);
606 TCPSTAT_INC(tcps_sc_reset);
607 } else {
608 TCPSTAT_INC(tcps_badrst);
609 /* Send challenge ACK. */
610 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
611 log(LOG_DEBUG, "%s; %s: RST with invalid "
612 " SEQ %u != NXT %u (+WND %u), "
613 "sending challenge ACK\n",
614 s, __func__,
615 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
616 syncache_respond(sc, sch, m, TH_ACK);
617 }
618 } else {
619 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
620 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
621 "NXT %u (+WND %u), segment ignored\n",
622 s, __func__,
623 th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
624 TCPSTAT_INC(tcps_badrst);
625 }
626
627 done:
628 if (s != NULL)
629 free(s, M_TCPLOG);
630 SCH_UNLOCK(sch);
631 }
632
633 void
634 syncache_badack(struct in_conninfo *inc)
635 {
636 struct syncache *sc;
637 struct syncache_head *sch;
638
639 sc = syncache_lookup(inc, &sch); /* returns locked sch */
640 SCH_LOCK_ASSERT(sch);
641 if (sc != NULL) {
642 syncache_drop(sc, sch);
643 TCPSTAT_INC(tcps_sc_badack);
644 }
645 SCH_UNLOCK(sch);
646 }
647
648 void
649 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq)
650 {
651 struct syncache *sc;
652 struct syncache_head *sch;
653
654 sc = syncache_lookup(inc, &sch); /* returns locked sch */
655 SCH_LOCK_ASSERT(sch);
656 if (sc == NULL)
657 goto done;
658
659 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
660 if (ntohl(th_seq) != sc->sc_iss)
661 goto done;
662
663 /*
664 * If we've rertransmitted 3 times and this is our second error,
665 * we remove the entry. Otherwise, we allow it to continue on.
666 * This prevents us from incorrectly nuking an entry during a
667 * spurious network outage.
668 *
669 * See tcp_notify().
670 */
671 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
672 sc->sc_flags |= SCF_UNREACH;
673 goto done;
674 }
675 syncache_drop(sc, sch);
676 TCPSTAT_INC(tcps_sc_unreach);
677 done:
678 SCH_UNLOCK(sch);
679 }
680
681 /*
682 * Build a new TCP socket structure from a syncache entry.
683 *
684 * On success return the newly created socket with its underlying inp locked.
685 */
686 static struct socket *
687 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
688 {
689 struct tcp_function_block *blk;
690 struct inpcb *inp = NULL;
691 struct socket *so;
692 struct tcpcb *tp;
693 int error;
694 char *s;
695
696 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
697
698 /*
699 * Ok, create the full blown connection, and set things up
700 * as they would have been set up if we had created the
701 * connection when the SYN arrived. If we can't create
702 * the connection, abort it.
703 */
704 so = sonewconn(lso, 0);
705 if (so == NULL) {
706 /*
707 * Drop the connection; we will either send a RST or
708 * have the peer retransmit its SYN again after its
709 * RTO and try again.
710 */
711 TCPSTAT_INC(tcps_listendrop);
712 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
713 log(LOG_DEBUG, "%s; %s: Socket create failed "
714 "due to limits or memory shortage\n",
715 s, __func__);
716 free(s, M_TCPLOG);
717 }
718 goto abort2;
719 }
720 #ifdef MAC
721 mac_socketpeer_set_from_mbuf(m, so);
722 #endif
723
724 inp = sotoinpcb(so);
725 inp->inp_inc.inc_fibnum = so->so_fibnum;
726 INP_WLOCK(inp);
727 /*
728 * Exclusive pcbinfo lock is not required in syncache socket case even
729 * if two inpcb locks can be acquired simultaneously:
730 * - the inpcb in LISTEN state,
731 * - the newly created inp.
732 *
733 * In this case, an inp cannot be at same time in LISTEN state and
734 * just created by an accept() call.
735 */
736 INP_HASH_WLOCK(&V_tcbinfo);
737
738 /* Insert new socket into PCB hash list. */
739 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
740 #ifdef INET6
741 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
742 inp->inp_vflag &= ~INP_IPV4;
743 inp->inp_vflag |= INP_IPV6;
744 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
745 } else {
746 inp->inp_vflag &= ~INP_IPV6;
747 inp->inp_vflag |= INP_IPV4;
748 #endif
749 inp->inp_laddr = sc->sc_inc.inc_laddr;
750 #ifdef INET6
751 }
752 #endif
753
754 /*
755 * If there's an mbuf and it has a flowid, then let's initialise the
756 * inp with that particular flowid.
757 */
758 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
759 inp->inp_flowid = m->m_pkthdr.flowid;
760 inp->inp_flowtype = M_HASHTYPE_GET(m);
761 }
762
763 /*
764 * Install in the reservation hash table for now, but don't yet
765 * install a connection group since the full 4-tuple isn't yet
766 * configured.
767 */
768 inp->inp_lport = sc->sc_inc.inc_lport;
769 if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
770 /*
771 * Undo the assignments above if we failed to
772 * put the PCB on the hash lists.
773 */
774 #ifdef INET6
775 if (sc->sc_inc.inc_flags & INC_ISIPV6)
776 inp->in6p_laddr = in6addr_any;
777 else
778 #endif
779 inp->inp_laddr.s_addr = INADDR_ANY;
780 inp->inp_lport = 0;
781 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
782 log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
783 "with error %i\n",
784 s, __func__, error);
785 free(s, M_TCPLOG);
786 }
787 INP_HASH_WUNLOCK(&V_tcbinfo);
788 goto abort;
789 }
790 #ifdef INET6
791 if (inp->inp_vflag & INP_IPV6PROTO) {
792 struct inpcb *oinp = sotoinpcb(lso);
793
794 /*
795 * Inherit socket options from the listening socket.
796 * Note that in6p_inputopts are not (and should not be)
797 * copied, since it stores previously received options and is
798 * used to detect if each new option is different than the
799 * previous one and hence should be passed to a user.
800 * If we copied in6p_inputopts, a user would not be able to
801 * receive options just after calling the accept system call.
802 */
803 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
804 if (oinp->in6p_outputopts)
805 inp->in6p_outputopts =
806 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
807 }
808
809 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
810 struct in6_addr laddr6;
811 struct sockaddr_in6 sin6;
812
813 sin6.sin6_family = AF_INET6;
814 sin6.sin6_len = sizeof(sin6);
815 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
816 sin6.sin6_port = sc->sc_inc.inc_fport;
817 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
818 laddr6 = inp->in6p_laddr;
819 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
820 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
821 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
822 thread0.td_ucred, m)) != 0) {
823 inp->in6p_laddr = laddr6;
824 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
825 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
826 "with error %i\n",
827 s, __func__, error);
828 free(s, M_TCPLOG);
829 }
830 INP_HASH_WUNLOCK(&V_tcbinfo);
831 goto abort;
832 }
833 /* Override flowlabel from in6_pcbconnect. */
834 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
835 inp->inp_flow |= sc->sc_flowlabel;
836 }
837 #endif /* INET6 */
838 #if defined(INET) && defined(INET6)
839 else
840 #endif
841 #ifdef INET
842 {
843 struct in_addr laddr;
844 struct sockaddr_in sin;
845
846 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
847
848 if (inp->inp_options == NULL) {
849 inp->inp_options = sc->sc_ipopts;
850 sc->sc_ipopts = NULL;
851 }
852
853 sin.sin_family = AF_INET;
854 sin.sin_len = sizeof(sin);
855 sin.sin_addr = sc->sc_inc.inc_faddr;
856 sin.sin_port = sc->sc_inc.inc_fport;
857 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
858 laddr = inp->inp_laddr;
859 if (inp->inp_laddr.s_addr == INADDR_ANY)
860 inp->inp_laddr = sc->sc_inc.inc_laddr;
861 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
862 thread0.td_ucred, m)) != 0) {
863 inp->inp_laddr = laddr;
864 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
865 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
866 "with error %i\n",
867 s, __func__, error);
868 free(s, M_TCPLOG);
869 }
870 INP_HASH_WUNLOCK(&V_tcbinfo);
871 goto abort;
872 }
873 }
874 #endif /* INET */
875 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
876 /* Copy old policy into new socket's. */
877 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
878 printf("syncache_socket: could not copy policy\n");
879 #endif
880 INP_HASH_WUNLOCK(&V_tcbinfo);
881 tp = intotcpcb(inp);
882 tcp_state_change(tp, TCPS_SYN_RECEIVED);
883 tp->iss = sc->sc_iss;
884 tp->irs = sc->sc_irs;
885 tcp_rcvseqinit(tp);
886 tcp_sendseqinit(tp);
887 blk = sototcpcb(lso)->t_fb;
888 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
889 /*
890 * Our parents t_fb was not the default,
891 * we need to release our ref on tp->t_fb and
892 * pickup one on the new entry.
893 */
894 struct tcp_function_block *rblk;
895
896 rblk = find_and_ref_tcp_fb(blk);
897 KASSERT(rblk != NULL,
898 ("cannot find blk %p out of syncache?", blk));
899 if (tp->t_fb->tfb_tcp_fb_fini)
900 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
901 refcount_release(&tp->t_fb->tfb_refcnt);
902 tp->t_fb = rblk;
903 /*
904 * XXXrrs this is quite dangerous, it is possible
905 * for the new function to fail to init. We also
906 * are not asking if the handoff_is_ok though at
907 * the very start thats probalbly ok.
908 */
909 if (tp->t_fb->tfb_tcp_fb_init) {
910 (*tp->t_fb->tfb_tcp_fb_init)(tp);
911 }
912 }
913 tp->snd_wl1 = sc->sc_irs;
914 tp->snd_max = tp->iss + 1;
915 tp->snd_nxt = tp->iss + 1;
916 tp->rcv_up = sc->sc_irs + 1;
917 tp->rcv_wnd = sc->sc_wnd;
918 tp->rcv_adv += tp->rcv_wnd;
919 tp->last_ack_sent = tp->rcv_nxt;
920
921 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
922 if (sc->sc_flags & SCF_NOOPT)
923 tp->t_flags |= TF_NOOPT;
924 else {
925 if (sc->sc_flags & SCF_WINSCALE) {
926 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
927 tp->snd_scale = sc->sc_requested_s_scale;
928 tp->request_r_scale = sc->sc_requested_r_scale;
929 }
930 if (sc->sc_flags & SCF_TIMESTAMP) {
931 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
932 tp->ts_recent = sc->sc_tsreflect;
933 tp->ts_recent_age = tcp_ts_getticks();
934 tp->ts_offset = sc->sc_tsoff;
935 }
936 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
937 if (sc->sc_flags & SCF_SIGNATURE)
938 tp->t_flags |= TF_SIGNATURE;
939 #endif
940 if (sc->sc_flags & SCF_SACK)
941 tp->t_flags |= TF_SACK_PERMIT;
942 }
943
944 if (sc->sc_flags & SCF_ECN)
945 tp->t_flags |= TF_ECN_PERMIT;
946
947 /*
948 * Set up MSS and get cached values from tcp_hostcache.
949 * This might overwrite some of the defaults we just set.
950 */
951 tcp_mss(tp, sc->sc_peer_mss);
952
953 /*
954 * If the SYN,ACK was retransmitted, indicate that CWND to be
955 * limited to one segment in cc_conn_init().
956 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
957 */
958 if (sc->sc_rxmits > 1)
959 tp->snd_cwnd = 1;
960
961 #ifdef TCP_OFFLOAD
962 /*
963 * Allow a TOE driver to install its hooks. Note that we hold the
964 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
965 * new connection before the TOE driver has done its thing.
966 */
967 if (ADDED_BY_TOE(sc)) {
968 struct toedev *tod = sc->sc_tod;
969
970 tod->tod_offload_socket(tod, sc->sc_todctx, so);
971 }
972 #endif
973 /*
974 * Copy and activate timers.
975 */
976 tp->t_keepinit = sototcpcb(lso)->t_keepinit;
977 tp->t_keepidle = sototcpcb(lso)->t_keepidle;
978 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
979 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
980 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
981
982 TCPSTAT_INC(tcps_accepts);
983 return (so);
984
985 abort:
986 INP_WUNLOCK(inp);
987 abort2:
988 if (so != NULL)
989 soabort(so);
990 return (NULL);
991 }
992
993 /*
994 * This function gets called when we receive an ACK for a
995 * socket in the LISTEN state. We look up the connection
996 * in the syncache, and if its there, we pull it out of
997 * the cache and turn it into a full-blown connection in
998 * the SYN-RECEIVED state.
999 *
1000 * On syncache_socket() success the newly created socket
1001 * has its underlying inp locked.
1002 */
1003 int
1004 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1005 struct socket **lsop, struct mbuf *m)
1006 {
1007 struct syncache *sc;
1008 struct syncache_head *sch;
1009 struct syncache scs;
1010 char *s;
1011
1012 /*
1013 * Global TCP locks are held because we manipulate the PCB lists
1014 * and create a new socket.
1015 */
1016 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1017 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
1018 ("%s: can handle only ACK", __func__));
1019
1020 sc = syncache_lookup(inc, &sch); /* returns locked sch */
1021 SCH_LOCK_ASSERT(sch);
1022
1023 #ifdef INVARIANTS
1024 /*
1025 * Test code for syncookies comparing the syncache stored
1026 * values with the reconstructed values from the cookie.
1027 */
1028 if (sc != NULL)
1029 syncookie_cmp(inc, sch, sc, th, to, *lsop);
1030 #endif
1031
1032 if (sc == NULL) {
1033 /*
1034 * There is no syncache entry, so see if this ACK is
1035 * a returning syncookie. To do this, first:
1036 * A. Check if syncookies are used in case of syncache
1037 * overflows
1038 * B. See if this socket has had a syncache entry dropped in
1039 * the recent past. We don't want to accept a bogus
1040 * syncookie if we've never received a SYN or accept it
1041 * twice.
1042 * C. check that the syncookie is valid. If it is, then
1043 * cobble up a fake syncache entry, and return.
1044 */
1045 if (!V_tcp_syncookies) {
1046 SCH_UNLOCK(sch);
1047 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1048 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1049 "segment rejected (syncookies disabled)\n",
1050 s, __func__);
1051 goto failed;
1052 }
1053 if (!V_tcp_syncookiesonly &&
1054 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
1055 SCH_UNLOCK(sch);
1056 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1057 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
1058 "segment rejected (no syncache entry)\n",
1059 s, __func__);
1060 goto failed;
1061 }
1062 bzero(&scs, sizeof(scs));
1063 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
1064 SCH_UNLOCK(sch);
1065 if (sc == NULL) {
1066 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1067 log(LOG_DEBUG, "%s; %s: Segment failed "
1068 "SYNCOOKIE authentication, segment rejected "
1069 "(probably spoofed)\n", s, __func__);
1070 goto failed;
1071 }
1072 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1073 /* If received ACK has MD5 signature, check it. */
1074 if ((to->to_flags & TOF_SIGNATURE) != 0 &&
1075 (!TCPMD5_ENABLED() ||
1076 TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
1077 /* Drop the ACK. */
1078 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1079 log(LOG_DEBUG, "%s; %s: Segment rejected, "
1080 "MD5 signature doesn't match.\n",
1081 s, __func__);
1082 free(s, M_TCPLOG);
1083 }
1084 TCPSTAT_INC(tcps_sig_err_sigopt);
1085 return (-1); /* Do not send RST */
1086 }
1087 #endif /* TCP_SIGNATURE */
1088 } else {
1089 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1090 /*
1091 * If listening socket requested TCP digests, check that
1092 * received ACK has signature and it is correct.
1093 * If not, drop the ACK and leave sc entry in th cache,
1094 * because SYN was received with correct signature.
1095 */
1096 if (sc->sc_flags & SCF_SIGNATURE) {
1097 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1098 /* No signature */
1099 TCPSTAT_INC(tcps_sig_err_nosigopt);
1100 SCH_UNLOCK(sch);
1101 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1102 log(LOG_DEBUG, "%s; %s: Segment "
1103 "rejected, MD5 signature wasn't "
1104 "provided.\n", s, __func__);
1105 free(s, M_TCPLOG);
1106 }
1107 return (-1); /* Do not send RST */
1108 }
1109 if (!TCPMD5_ENABLED() ||
1110 TCPMD5_INPUT(m, th, to->to_signature) != 0) {
1111 /* Doesn't match or no SA */
1112 SCH_UNLOCK(sch);
1113 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1114 log(LOG_DEBUG, "%s; %s: Segment "
1115 "rejected, MD5 signature doesn't "
1116 "match.\n", s, __func__);
1117 free(s, M_TCPLOG);
1118 }
1119 return (-1); /* Do not send RST */
1120 }
1121 }
1122 #endif /* TCP_SIGNATURE */
1123 /*
1124 * Pull out the entry to unlock the bucket row.
1125 *
1126 * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
1127 * tcp_state_change(). The tcpcb is not existent at this
1128 * moment. A new one will be allocated via syncache_socket->
1129 * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
1130 * syncache_socket() will change it to TCPS_SYN_RECEIVED.
1131 */
1132 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
1133 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
1134 sch->sch_length--;
1135 #ifdef TCP_OFFLOAD
1136 if (ADDED_BY_TOE(sc)) {
1137 struct toedev *tod = sc->sc_tod;
1138
1139 tod->tod_syncache_removed(tod, sc->sc_todctx);
1140 }
1141 #endif
1142 SCH_UNLOCK(sch);
1143 }
1144
1145 /*
1146 * Segment validation:
1147 * ACK must match our initial sequence number + 1 (the SYN|ACK).
1148 */
1149 if (th->th_ack != sc->sc_iss + 1) {
1150 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1151 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
1152 "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
1153 goto failed;
1154 }
1155
1156 /*
1157 * The SEQ must fall in the window starting at the received
1158 * initial receive sequence number + 1 (the SYN).
1159 */
1160 if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
1161 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
1162 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1163 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
1164 "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
1165 goto failed;
1166 }
1167
1168 /*
1169 * If timestamps were not negotiated during SYN/ACK they
1170 * must not appear on any segment during this session.
1171 */
1172 if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
1173 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
1174 log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
1175 "segment rejected\n", s, __func__);
1176 goto failed;
1177 }
1178
1179 /*
1180 * If timestamps were negotiated during SYN/ACK they should
1181 * appear on every segment during this session.
1182 * XXXAO: This is only informal as there have been unverified
1183 * reports of non-compliants stacks.
1184 */
1185 if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
1186 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
1187 log(LOG_DEBUG, "%s; %s: Timestamp missing, "
1188 "no action\n", s, __func__);
1189 free(s, M_TCPLOG);
1190 s = NULL;
1191 }
1192 }
1193
1194 *lsop = syncache_socket(sc, *lsop, m);
1195
1196 if (*lsop == NULL)
1197 TCPSTAT_INC(tcps_sc_aborted);
1198 else
1199 TCPSTAT_INC(tcps_sc_completed);
1200
1201 /* how do we find the inp for the new socket? */
1202 if (sc != &scs)
1203 syncache_free(sc);
1204 return (1);
1205 failed:
1206 if (sc != NULL && sc != &scs)
1207 syncache_free(sc);
1208 if (s != NULL)
1209 free(s, M_TCPLOG);
1210 *lsop = NULL;
1211 return (0);
1212 }
1213
1214 static void
1215 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
1216 uint64_t response_cookie)
1217 {
1218 struct inpcb *inp;
1219 struct tcpcb *tp;
1220 unsigned int *pending_counter;
1221
1222 /*
1223 * Global TCP locks are held because we manipulate the PCB lists
1224 * and create a new socket.
1225 */
1226 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1227
1228 pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
1229 *lsop = syncache_socket(sc, *lsop, m);
1230 if (*lsop == NULL) {
1231 TCPSTAT_INC(tcps_sc_aborted);
1232 atomic_subtract_int(pending_counter, 1);
1233 } else {
1234 soisconnected(*lsop);
1235 inp = sotoinpcb(*lsop);
1236 tp = intotcpcb(inp);
1237 tp->t_flags |= TF_FASTOPEN;
1238 tp->t_tfo_cookie.server = response_cookie;
1239 tp->snd_max = tp->iss;
1240 tp->snd_nxt = tp->iss;
1241 tp->t_tfo_pending = pending_counter;
1242 TCPSTAT_INC(tcps_sc_completed);
1243 }
1244 }
1245
1246 /*
1247 * Given a LISTEN socket and an inbound SYN request, add
1248 * this to the syn cache, and send back a segment:
1249 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
1250 * to the source.
1251 *
1252 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
1253 * Doing so would require that we hold onto the data and deliver it
1254 * to the application. However, if we are the target of a SYN-flood
1255 * DoS attack, an attacker could send data which would eventually
1256 * consume all available buffer space if it were ACKed. By not ACKing
1257 * the data, we avoid this DoS scenario.
1258 *
1259 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
1260 * cookie is processed and a new socket is created. In this case, any data
1261 * accompanying the SYN will be queued to the socket by tcp_input() and will
1262 * be ACKed either when the application sends response data or the delayed
1263 * ACK timer expires, whichever comes first.
1264 */
1265 int
1266 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
1267 struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
1268 void *todctx)
1269 {
1270 struct tcpcb *tp;
1271 struct socket *so;
1272 struct syncache *sc = NULL;
1273 struct syncache_head *sch;
1274 struct mbuf *ipopts = NULL;
1275 u_int ltflags;
1276 int win, ip_ttl, ip_tos;
1277 char *s;
1278 int rv = 0;
1279 #ifdef INET6
1280 int autoflowlabel = 0;
1281 #endif
1282 #ifdef MAC
1283 struct label *maclabel;
1284 #endif
1285 struct syncache scs;
1286 struct ucred *cred;
1287 uint64_t tfo_response_cookie;
1288 unsigned int *tfo_pending = NULL;
1289 int tfo_cookie_valid = 0;
1290 int tfo_response_cookie_valid = 0;
1291
1292 INP_WLOCK_ASSERT(inp); /* listen socket */
1293 KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
1294 ("%s: unexpected tcp flags", __func__));
1295
1296 /*
1297 * Combine all so/tp operations very early to drop the INP lock as
1298 * soon as possible.
1299 */
1300 so = *lsop;
1301 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
1302 tp = sototcpcb(so);
1303 cred = crhold(so->so_cred);
1304
1305 #ifdef INET6
1306 if ((inc->inc_flags & INC_ISIPV6) &&
1307 (inp->inp_flags & IN6P_AUTOFLOWLABEL))
1308 autoflowlabel = 1;
1309 #endif
1310 ip_ttl = inp->inp_ip_ttl;
1311 ip_tos = inp->inp_ip_tos;
1312 win = so->sol_sbrcv_hiwat;
1313 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
1314
1315 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
1316 (tp->t_tfo_pending != NULL) &&
1317 (to->to_flags & TOF_FASTOPEN)) {
1318 /*
1319 * Limit the number of pending TFO connections to
1320 * approximately half of the queue limit. This prevents TFO
1321 * SYN floods from starving the service by filling the
1322 * listen queue with bogus TFO connections.
1323 */
1324 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
1325 (so->sol_qlimit / 2)) {
1326 int result;
1327
1328 result = tcp_fastopen_check_cookie(inc,
1329 to->to_tfo_cookie, to->to_tfo_len,
1330 &tfo_response_cookie);
1331 tfo_cookie_valid = (result > 0);
1332 tfo_response_cookie_valid = (result >= 0);
1333 }
1334
1335 /*
1336 * Remember the TFO pending counter as it will have to be
1337 * decremented below if we don't make it to syncache_tfo_expand().
1338 */
1339 tfo_pending = tp->t_tfo_pending;
1340 }
1341
1342 /* By the time we drop the lock these should no longer be used. */
1343 so = NULL;
1344 tp = NULL;
1345
1346 #ifdef MAC
1347 if (mac_syncache_init(&maclabel) != 0) {
1348 INP_WUNLOCK(inp);
1349 goto done;
1350 } else
1351 mac_syncache_create(maclabel, inp);
1352 #endif
1353 if (!tfo_cookie_valid)
1354 INP_WUNLOCK(inp);
1355
1356 /*
1357 * Remember the IP options, if any.
1358 */
1359 #ifdef INET6
1360 if (!(inc->inc_flags & INC_ISIPV6))
1361 #endif
1362 #ifdef INET
1363 ipopts = (m) ? ip_srcroute(m) : NULL;
1364 #else
1365 ipopts = NULL;
1366 #endif
1367
1368 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1369 /*
1370 * If listening socket requested TCP digests, check that received
1371 * SYN has signature and it is correct. If signature doesn't match
1372 * or TCP_SIGNATURE support isn't enabled, drop the packet.
1373 */
1374 if (ltflags & TF_SIGNATURE) {
1375 if ((to->to_flags & TOF_SIGNATURE) == 0) {
1376 TCPSTAT_INC(tcps_sig_err_nosigopt);
1377 goto done;
1378 }
1379 if (!TCPMD5_ENABLED() ||
1380 TCPMD5_INPUT(m, th, to->to_signature) != 0)
1381 goto done;
1382 }
1383 #endif /* TCP_SIGNATURE */
1384 /*
1385 * See if we already have an entry for this connection.
1386 * If we do, resend the SYN,ACK, and reset the retransmit timer.
1387 *
1388 * XXX: should the syncache be re-initialized with the contents
1389 * of the new SYN here (which may have different options?)
1390 *
1391 * XXX: We do not check the sequence number to see if this is a
1392 * real retransmit or a new connection attempt. The question is
1393 * how to handle such a case; either ignore it as spoofed, or
1394 * drop the current entry and create a new one?
1395 */
1396 sc = syncache_lookup(inc, &sch); /* returns locked entry */
1397 SCH_LOCK_ASSERT(sch);
1398 if (sc != NULL) {
1399 if (tfo_cookie_valid)
1400 INP_WUNLOCK(inp);
1401 TCPSTAT_INC(tcps_sc_dupsyn);
1402 if (ipopts) {
1403 /*
1404 * If we were remembering a previous source route,
1405 * forget it and use the new one we've been given.
1406 */
1407 if (sc->sc_ipopts)
1408 (void) m_free(sc->sc_ipopts);
1409 sc->sc_ipopts = ipopts;
1410 }
1411 /*
1412 * Update timestamp if present.
1413 */
1414 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
1415 sc->sc_tsreflect = to->to_tsval;
1416 else
1417 sc->sc_flags &= ~SCF_TIMESTAMP;
1418 #ifdef MAC
1419 /*
1420 * Since we have already unconditionally allocated label
1421 * storage, free it up. The syncache entry will already
1422 * have an initialized label we can use.
1423 */
1424 mac_syncache_destroy(&maclabel);
1425 #endif
1426 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1427 /* Retransmit SYN|ACK and reset retransmit count. */
1428 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
1429 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
1430 "resetting timer and retransmitting SYN|ACK\n",
1431 s, __func__);
1432 free(s, M_TCPLOG);
1433 }
1434 if (syncache_respond(sc, sch, m, TH_SYN|TH_ACK) == 0) {
1435 sc->sc_rxmits = 0;
1436 syncache_timeout(sc, sch, 1);
1437 TCPSTAT_INC(tcps_sndacks);
1438 TCPSTAT_INC(tcps_sndtotal);
1439 }
1440 SCH_UNLOCK(sch);
1441 goto donenoprobe;
1442 }
1443
1444 if (tfo_cookie_valid) {
1445 bzero(&scs, sizeof(scs));
1446 sc = &scs;
1447 goto skip_alloc;
1448 }
1449
1450 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1451 if (sc == NULL) {
1452 /*
1453 * The zone allocator couldn't provide more entries.
1454 * Treat this as if the cache was full; drop the oldest
1455 * entry and insert the new one.
1456 */
1457 TCPSTAT_INC(tcps_sc_zonefail);
1458 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
1459 sch->sch_last_overflow = time_uptime;
1460 syncache_drop(sc, sch);
1461 }
1462 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
1463 if (sc == NULL) {
1464 if (V_tcp_syncookies) {
1465 bzero(&scs, sizeof(scs));
1466 sc = &scs;
1467 } else {
1468 SCH_UNLOCK(sch);
1469 if (ipopts)
1470 (void) m_free(ipopts);
1471 goto done;
1472 }
1473 }
1474 }
1475
1476 skip_alloc:
1477 if (!tfo_cookie_valid && tfo_response_cookie_valid)
1478 sc->sc_tfo_cookie = &tfo_response_cookie;
1479
1480 /*
1481 * Fill in the syncache values.
1482 */
1483 #ifdef MAC
1484 sc->sc_label = maclabel;
1485 #endif
1486 sc->sc_cred = cred;
1487 cred = NULL;
1488 sc->sc_ipopts = ipopts;
1489 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
1490 #ifdef INET6
1491 if (!(inc->inc_flags & INC_ISIPV6))
1492 #endif
1493 {
1494 sc->sc_ip_tos = ip_tos;
1495 sc->sc_ip_ttl = ip_ttl;
1496 }
1497 #ifdef TCP_OFFLOAD
1498 sc->sc_tod = tod;
1499 sc->sc_todctx = todctx;
1500 #endif
1501 sc->sc_irs = th->th_seq;
1502 sc->sc_iss = arc4random();
1503 sc->sc_flags = 0;
1504 sc->sc_flowlabel = 0;
1505
1506 /*
1507 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
1508 * win was derived from socket earlier in the function.
1509 */
1510 win = imax(win, 0);
1511 win = imin(win, TCP_MAXWIN);
1512 sc->sc_wnd = win;
1513
1514 if (V_tcp_do_rfc1323) {
1515 /*
1516 * A timestamp received in a SYN makes
1517 * it ok to send timestamp requests and replies.
1518 */
1519 if (to->to_flags & TOF_TS) {
1520 sc->sc_tsreflect = to->to_tsval;
1521 sc->sc_flags |= SCF_TIMESTAMP;
1522 sc->sc_tsoff = tcp_new_ts_offset(inc);
1523 }
1524 if (to->to_flags & TOF_SCALE) {
1525 int wscale = 0;
1526
1527 /*
1528 * Pick the smallest possible scaling factor that
1529 * will still allow us to scale up to sb_max, aka
1530 * kern.ipc.maxsockbuf.
1531 *
1532 * We do this because there are broken firewalls that
1533 * will corrupt the window scale option, leading to
1534 * the other endpoint believing that our advertised
1535 * window is unscaled. At scale factors larger than
1536 * 5 the unscaled window will drop below 1500 bytes,
1537 * leading to serious problems when traversing these
1538 * broken firewalls.
1539 *
1540 * With the default maxsockbuf of 256K, a scale factor
1541 * of 3 will be chosen by this algorithm. Those who
1542 * choose a larger maxsockbuf should watch out
1543 * for the compatibility problems mentioned above.
1544 *
1545 * RFC1323: The Window field in a SYN (i.e., a <SYN>
1546 * or <SYN,ACK>) segment itself is never scaled.
1547 */
1548 while (wscale < TCP_MAX_WINSHIFT &&
1549 (TCP_MAXWIN << wscale) < sb_max)
1550 wscale++;
1551 sc->sc_requested_r_scale = wscale;
1552 sc->sc_requested_s_scale = to->to_wscale;
1553 sc->sc_flags |= SCF_WINSCALE;
1554 }
1555 }
1556 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1557 /*
1558 * If listening socket requested TCP digests, flag this in the
1559 * syncache so that syncache_respond() will do the right thing
1560 * with the SYN+ACK.
1561 */
1562 if (ltflags & TF_SIGNATURE)
1563 sc->sc_flags |= SCF_SIGNATURE;
1564 #endif /* TCP_SIGNATURE */
1565 if (to->to_flags & TOF_SACKPERM)
1566 sc->sc_flags |= SCF_SACK;
1567 if (to->to_flags & TOF_MSS)
1568 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */
1569 if (ltflags & TF_NOOPT)
1570 sc->sc_flags |= SCF_NOOPT;
1571 if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
1572 sc->sc_flags |= SCF_ECN;
1573
1574 if (V_tcp_syncookies)
1575 sc->sc_iss = syncookie_generate(sch, sc);
1576 #ifdef INET6
1577 if (autoflowlabel) {
1578 if (V_tcp_syncookies)
1579 sc->sc_flowlabel = sc->sc_iss;
1580 else
1581 sc->sc_flowlabel = ip6_randomflowlabel();
1582 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
1583 }
1584 #endif
1585 SCH_UNLOCK(sch);
1586
1587 if (tfo_cookie_valid) {
1588 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie);
1589 /* INP_WUNLOCK(inp) will be performed by the caller */
1590 rv = 1;
1591 goto tfo_expanded;
1592 }
1593
1594 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1595 /*
1596 * Do a standard 3-way handshake.
1597 */
1598 if (syncache_respond(sc, sch, m, TH_SYN|TH_ACK) == 0) {
1599 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
1600 syncache_free(sc);
1601 else if (sc != &scs)
1602 syncache_insert(sc, sch); /* locks and unlocks sch */
1603 TCPSTAT_INC(tcps_sndacks);
1604 TCPSTAT_INC(tcps_sndtotal);
1605 } else {
1606 if (sc != &scs)
1607 syncache_free(sc);
1608 TCPSTAT_INC(tcps_sc_dropped);
1609 }
1610 goto donenoprobe;
1611
1612 done:
1613 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
1614 donenoprobe:
1615 if (m) {
1616 *lsop = NULL;
1617 m_freem(m);
1618 }
1619 /*
1620 * If tfo_pending is not NULL here, then a TFO SYN that did not
1621 * result in a new socket was processed and the associated pending
1622 * counter has not yet been decremented. All such TFO processing paths
1623 * transit this point.
1624 */
1625 if (tfo_pending != NULL)
1626 tcp_fastopen_decrement_counter(tfo_pending);
1627
1628 tfo_expanded:
1629 if (cred != NULL)
1630 crfree(cred);
1631 #ifdef MAC
1632 if (sc == &scs)
1633 mac_syncache_destroy(&maclabel);
1634 #endif
1635 return (rv);
1636 }
1637
1638 /*
1639 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment,
1640 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
1641 */
1642 static int
1643 syncache_respond(struct syncache *sc, struct syncache_head *sch,
1644 const struct mbuf *m0, int flags)
1645 {
1646 struct ip *ip = NULL;
1647 struct mbuf *m;
1648 struct tcphdr *th = NULL;
1649 int optlen, error = 0; /* Make compiler happy */
1650 u_int16_t hlen, tlen, mssopt;
1651 struct tcpopt to;
1652 #ifdef INET6
1653 struct ip6_hdr *ip6 = NULL;
1654 #endif
1655 hlen =
1656 #ifdef INET6
1657 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
1658 #endif
1659 sizeof(struct ip);
1660 tlen = hlen + sizeof(struct tcphdr);
1661
1662 /* Determine MSS we advertize to other end of connection. */
1663 mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss);
1664
1665 /* XXX: Assume that the entire packet will fit in a header mbuf. */
1666 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
1667 ("syncache: mbuf too small"));
1668
1669 /* Create the IP+TCP header from scratch. */
1670 m = m_gethdr(M_NOWAIT, MT_DATA);
1671 if (m == NULL)
1672 return (ENOBUFS);
1673 #ifdef MAC
1674 mac_syncache_create_mbuf(sc->sc_label, m);
1675 #endif
1676 m->m_data += max_linkhdr;
1677 m->m_len = tlen;
1678 m->m_pkthdr.len = tlen;
1679 m->m_pkthdr.rcvif = NULL;
1680
1681 #ifdef INET6
1682 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1683 ip6 = mtod(m, struct ip6_hdr *);
1684 ip6->ip6_vfc = IPV6_VERSION;
1685 ip6->ip6_nxt = IPPROTO_TCP;
1686 ip6->ip6_src = sc->sc_inc.inc6_laddr;
1687 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
1688 ip6->ip6_plen = htons(tlen - hlen);
1689 /* ip6_hlim is set after checksum */
1690 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
1691 ip6->ip6_flow |= sc->sc_flowlabel;
1692
1693 th = (struct tcphdr *)(ip6 + 1);
1694 }
1695 #endif
1696 #if defined(INET6) && defined(INET)
1697 else
1698 #endif
1699 #ifdef INET
1700 {
1701 ip = mtod(m, struct ip *);
1702 ip->ip_v = IPVERSION;
1703 ip->ip_hl = sizeof(struct ip) >> 2;
1704 ip->ip_len = htons(tlen);
1705 ip->ip_id = 0;
1706 ip->ip_off = 0;
1707 ip->ip_sum = 0;
1708 ip->ip_p = IPPROTO_TCP;
1709 ip->ip_src = sc->sc_inc.inc_laddr;
1710 ip->ip_dst = sc->sc_inc.inc_faddr;
1711 ip->ip_ttl = sc->sc_ip_ttl;
1712 ip->ip_tos = sc->sc_ip_tos;
1713
1714 /*
1715 * See if we should do MTU discovery. Route lookups are
1716 * expensive, so we will only unset the DF bit if:
1717 *
1718 * 1) path_mtu_discovery is disabled
1719 * 2) the SCF_UNREACH flag has been set
1720 */
1721 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
1722 ip->ip_off |= htons(IP_DF);
1723
1724 th = (struct tcphdr *)(ip + 1);
1725 }
1726 #endif /* INET */
1727 th->th_sport = sc->sc_inc.inc_lport;
1728 th->th_dport = sc->sc_inc.inc_fport;
1729
1730 if (flags & TH_SYN)
1731 th->th_seq = htonl(sc->sc_iss);
1732 else
1733 th->th_seq = htonl(sc->sc_iss + 1);
1734 th->th_ack = htonl(sc->sc_irs + 1);
1735 th->th_off = sizeof(struct tcphdr) >> 2;
1736 th->th_x2 = 0;
1737 th->th_flags = flags;
1738 th->th_win = htons(sc->sc_wnd);
1739 th->th_urp = 0;
1740
1741 if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) {
1742 th->th_flags |= TH_ECE;
1743 TCPSTAT_INC(tcps_ecn_shs);
1744 }
1745
1746 /* Tack on the TCP options. */
1747 if ((sc->sc_flags & SCF_NOOPT) == 0) {
1748 to.to_flags = 0;
1749
1750 if (flags & TH_SYN) {
1751 to.to_mss = mssopt;
1752 to.to_flags = TOF_MSS;
1753 if (sc->sc_flags & SCF_WINSCALE) {
1754 to.to_wscale = sc->sc_requested_r_scale;
1755 to.to_flags |= TOF_SCALE;
1756 }
1757 if (sc->sc_flags & SCF_SACK)
1758 to.to_flags |= TOF_SACKPERM;
1759 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1760 if (sc->sc_flags & SCF_SIGNATURE)
1761 to.to_flags |= TOF_SIGNATURE;
1762 #endif
1763 if (sc->sc_tfo_cookie) {
1764 to.to_flags |= TOF_FASTOPEN;
1765 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
1766 to.to_tfo_cookie = sc->sc_tfo_cookie;
1767 /* don't send cookie again when retransmitting response */
1768 sc->sc_tfo_cookie = NULL;
1769 }
1770 }
1771 if (sc->sc_flags & SCF_TIMESTAMP) {
1772 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
1773 to.to_tsecr = sc->sc_tsreflect;
1774 to.to_flags |= TOF_TS;
1775 }
1776 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
1777
1778 /* Adjust headers by option size. */
1779 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1780 m->m_len += optlen;
1781 m->m_pkthdr.len += optlen;
1782 #ifdef INET6
1783 if (sc->sc_inc.inc_flags & INC_ISIPV6)
1784 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
1785 else
1786 #endif
1787 ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
1788 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1789 if (sc->sc_flags & SCF_SIGNATURE) {
1790 KASSERT(to.to_flags & TOF_SIGNATURE,
1791 ("tcp_addoptions() didn't set tcp_signature"));
1792
1793 /* NOTE: to.to_signature is inside of mbuf */
1794 if (!TCPMD5_ENABLED() ||
1795 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
1796 m_freem(m);
1797 return (EACCES);
1798 }
1799 }
1800 #endif
1801 } else
1802 optlen = 0;
1803
1804 M_SETFIB(m, sc->sc_inc.inc_fibnum);
1805 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1806 /*
1807 * If we have peer's SYN and it has a flowid, then let's assign it to
1808 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid
1809 * to SYN|ACK due to lack of inp here.
1810 */
1811 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
1812 m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
1813 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
1814 }
1815 #ifdef INET6
1816 if (sc->sc_inc.inc_flags & INC_ISIPV6) {
1817 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1818 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
1819 IPPROTO_TCP, 0);
1820 ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
1821 #ifdef TCP_OFFLOAD
1822 if (ADDED_BY_TOE(sc)) {
1823 struct toedev *tod = sc->sc_tod;
1824
1825 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1826
1827 return (error);
1828 }
1829 #endif
1830 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
1831 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
1832 }
1833 #endif
1834 #if defined(INET6) && defined(INET)
1835 else
1836 #endif
1837 #ifdef INET
1838 {
1839 m->m_pkthdr.csum_flags = CSUM_TCP;
1840 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1841 htons(tlen + optlen - hlen + IPPROTO_TCP));
1842 #ifdef TCP_OFFLOAD
1843 if (ADDED_BY_TOE(sc)) {
1844 struct toedev *tod = sc->sc_tod;
1845
1846 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
1847
1848 return (error);
1849 }
1850 #endif
1851 TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
1852 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
1853 }
1854 #endif
1855 return (error);
1856 }
1857
1858 /*
1859 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
1860 * that exceed the capacity of the syncache by avoiding the storage of any
1861 * of the SYNs we receive. Syncookies defend against blind SYN flooding
1862 * attacks where the attacker does not have access to our responses.
1863 *
1864 * Syncookies encode and include all necessary information about the
1865 * connection setup within the SYN|ACK that we send back. That way we
1866 * can avoid keeping any local state until the ACK to our SYN|ACK returns
1867 * (if ever). Normally the syncache and syncookies are running in parallel
1868 * with the latter taking over when the former is exhausted. When matching
1869 * syncache entry is found the syncookie is ignored.
1870 *
1871 * The only reliable information persisting the 3WHS is our initial sequence
1872 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient
1873 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
1874 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK
1875 * returns and signifies a legitimate connection if it matches the ACK.
1876 *
1877 * The available space of 32 bits to store the hash and to encode the SYN
1878 * option information is very tight and we should have at least 24 bits for
1879 * the MAC to keep the number of guesses by blind spoofing reasonably high.
1880 *
1881 * SYN option information we have to encode to fully restore a connection:
1882 * MSS: is imporant to chose an optimal segment size to avoid IP level
1883 * fragmentation along the path. The common MSS values can be encoded
1884 * in a 3-bit table. Uncommon values are captured by the next lower value
1885 * in the table leading to a slight increase in packetization overhead.
1886 * WSCALE: is necessary to allow large windows to be used for high delay-
1887 * bandwidth product links. Not scaling the window when it was initially
1888 * negotiated is bad for performance as lack of scaling further decreases
1889 * the apparent available send window. We only need to encode the WSCALE
1890 * we received from the remote end. Our end can be recalculated at any
1891 * time. The common WSCALE values can be encoded in a 3-bit table.
1892 * Uncommon values are captured by the next lower value in the table
1893 * making us under-estimate the available window size halving our
1894 * theoretically possible maximum throughput for that connection.
1895 * SACK: Greatly assists in packet loss recovery and requires 1 bit.
1896 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
1897 * that are included in all segments on a connection. We enable them when
1898 * the ACK has them.
1899 *
1900 * Security of syncookies and attack vectors:
1901 *
1902 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
1903 * together with the gloabl secret to make it unique per connection attempt.
1904 * Thus any change of any of those parameters results in a different MAC output
1905 * in an unpredictable way unless a collision is encountered. 24 bits of the
1906 * MAC are embedded into the ISS.
1907 *
1908 * To prevent replay attacks two rotating global secrets are updated with a
1909 * new random value every 15 seconds. The life-time of a syncookie is thus
1910 * 15-30 seconds.
1911 *
1912 * Vector 1: Attacking the secret. This requires finding a weakness in the
1913 * MAC itself or the way it is used here. The attacker can do a chosen plain
1914 * text attack by varying and testing the all parameters under his control.
1915 * The strength depends on the size and randomness of the secret, and the
1916 * cryptographic security of the MAC function. Due to the constant updating
1917 * of the secret the attacker has at most 29.999 seconds to find the secret
1918 * and launch spoofed connections. After that he has to start all over again.
1919 *
1920 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC
1921 * size an average of 4,823 attempts are required for a 50% chance of success
1922 * to spoof a single syncookie (birthday collision paradox). However the
1923 * attacker is blind and doesn't know if one of his attempts succeeded unless
1924 * he has a side channel to interfere success from. A single connection setup
1925 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
1926 * This many attempts are required for each one blind spoofed connection. For
1927 * every additional spoofed connection he has to launch another N attempts.
1928 * Thus for a sustained rate 100 spoofed connections per second approximately
1929 * 1,800,000 packets per second would have to be sent.
1930 *
1931 * NB: The MAC function should be fast so that it doesn't become a CPU
1932 * exhaustion attack vector itself.
1933 *
1934 * References:
1935 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations
1936 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
1937 * http://cr.yp.to/syncookies.html (overview)
1938 * http://cr.yp.to/syncookies/archive (details)
1939 *
1940 *
1941 * Schematic construction of a syncookie enabled Initial Sequence Number:
1942 * 0 1 2 3
1943 * 12345678901234567890123456789012
1944 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
1945 *
1946 * x 24 MAC (truncated)
1947 * W 3 Send Window Scale index
1948 * M 3 MSS index
1949 * S 1 SACK permitted
1950 * P 1 Odd/even secret
1951 */
1952
1953 /*
1954 * Distribution and probability of certain MSS values. Those in between are
1955 * rounded down to the next lower one.
1956 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
1957 * .2% .3% 5% 7% 7% 20% 15% 45%
1958 */
1959 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
1960
1961 /*
1962 * Distribution and probability of certain WSCALE values. We have to map the
1963 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
1964 * bits based on prevalence of certain values. Where we don't have an exact
1965 * match for are rounded down to the next lower one letting us under-estimate
1966 * the true available window. At the moment this would happen only for the
1967 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
1968 * and window size). The absence of the WSCALE option (no scaling in either
1969 * direction) is encoded with index zero.
1970 * [WSCALE values histograms, Allman, 2012]
1971 * X 10 10 35 5 6 14 10% by host
1972 * X 11 4 5 5 18 49 3% by connections
1973 */
1974 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
1975
1976 /*
1977 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed
1978 * and good cryptographic properties.
1979 */
1980 static uint32_t
1981 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
1982 uint8_t *secbits, uintptr_t secmod)
1983 {
1984 SIPHASH_CTX ctx;
1985 uint32_t siphash[2];
1986
1987 SipHash24_Init(&ctx);
1988 SipHash_SetKey(&ctx, secbits);
1989 switch (inc->inc_flags & INC_ISIPV6) {
1990 #ifdef INET
1991 case 0:
1992 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
1993 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
1994 break;
1995 #endif
1996 #ifdef INET6
1997 case INC_ISIPV6:
1998 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
1999 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
2000 break;
2001 #endif
2002 }
2003 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
2004 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
2005 SipHash_Update(&ctx, &irs, sizeof(irs));
2006 SipHash_Update(&ctx, &flags, sizeof(flags));
2007 SipHash_Update(&ctx, &secmod, sizeof(secmod));
2008 SipHash_Final((u_int8_t *)&siphash, &ctx);
2009
2010 return (siphash[0] ^ siphash[1]);
2011 }
2012
2013 static tcp_seq
2014 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
2015 {
2016 u_int i, secbit, wscale;
2017 uint32_t iss, hash;
2018 uint8_t *secbits;
2019 union syncookie cookie;
2020
2021 SCH_LOCK_ASSERT(sch);
2022
2023 cookie.cookie = 0;
2024
2025 /* Map our computed MSS into the 3-bit index. */
2026 for (i = nitems(tcp_sc_msstab) - 1;
2027 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
2028 i--)
2029 ;
2030 cookie.flags.mss_idx = i;
2031
2032 /*
2033 * Map the send window scale into the 3-bit index but only if
2034 * the wscale option was received.
2035 */
2036 if (sc->sc_flags & SCF_WINSCALE) {
2037 wscale = sc->sc_requested_s_scale;
2038 for (i = nitems(tcp_sc_wstab) - 1;
2039 tcp_sc_wstab[i] > wscale && i > 0;
2040 i--)
2041 ;
2042 cookie.flags.wscale_idx = i;
2043 }
2044
2045 /* Can we do SACK? */
2046 if (sc->sc_flags & SCF_SACK)
2047 cookie.flags.sack_ok = 1;
2048
2049 /* Which of the two secrets to use. */
2050 secbit = sch->sch_sc->secret.oddeven & 0x1;
2051 cookie.flags.odd_even = secbit;
2052
2053 secbits = sch->sch_sc->secret.key[secbit];
2054 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
2055 (uintptr_t)sch);
2056
2057 /*
2058 * Put the flags into the hash and XOR them to get better ISS number
2059 * variance. This doesn't enhance the cryptographic strength and is
2060 * done to prevent the 8 cookie bits from showing up directly on the
2061 * wire.
2062 */
2063 iss = hash & ~0xff;
2064 iss |= cookie.cookie ^ (hash >> 24);
2065
2066 TCPSTAT_INC(tcps_sc_sendcookie);
2067 return (iss);
2068 }
2069
2070 static struct syncache *
2071 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch,
2072 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2073 struct socket *lso)
2074 {
2075 uint32_t hash;
2076 uint8_t *secbits;
2077 tcp_seq ack, seq;
2078 int wnd, wscale = 0;
2079 union syncookie cookie;
2080
2081 SCH_LOCK_ASSERT(sch);
2082
2083 /*
2084 * Pull information out of SYN-ACK/ACK and revert sequence number
2085 * advances.
2086 */
2087 ack = th->th_ack - 1;
2088 seq = th->th_seq - 1;
2089
2090 /*
2091 * Unpack the flags containing enough information to restore the
2092 * connection.
2093 */
2094 cookie.cookie = (ack & 0xff) ^ (ack >> 24);
2095
2096 /* Which of the two secrets to use. */
2097 secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
2098
2099 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
2100
2101 /* The recomputed hash matches the ACK if this was a genuine cookie. */
2102 if ((ack & ~0xff) != (hash & ~0xff))
2103 return (NULL);
2104
2105 /* Fill in the syncache values. */
2106 sc->sc_flags = 0;
2107 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
2108 sc->sc_ipopts = NULL;
2109
2110 sc->sc_irs = seq;
2111 sc->sc_iss = ack;
2112
2113 switch (inc->inc_flags & INC_ISIPV6) {
2114 #ifdef INET
2115 case 0:
2116 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
2117 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
2118 break;
2119 #endif
2120 #ifdef INET6
2121 case INC_ISIPV6:
2122 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
2123 sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
2124 break;
2125 #endif
2126 }
2127
2128 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
2129
2130 /* We can simply recompute receive window scale we sent earlier. */
2131 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
2132 wscale++;
2133
2134 /* Only use wscale if it was enabled in the orignal SYN. */
2135 if (cookie.flags.wscale_idx > 0) {
2136 sc->sc_requested_r_scale = wscale;
2137 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
2138 sc->sc_flags |= SCF_WINSCALE;
2139 }
2140
2141 wnd = lso->sol_sbrcv_hiwat;
2142 wnd = imax(wnd, 0);
2143 wnd = imin(wnd, TCP_MAXWIN);
2144 sc->sc_wnd = wnd;
2145
2146 if (cookie.flags.sack_ok)
2147 sc->sc_flags |= SCF_SACK;
2148
2149 if (to->to_flags & TOF_TS) {
2150 sc->sc_flags |= SCF_TIMESTAMP;
2151 sc->sc_tsreflect = to->to_tsval;
2152 sc->sc_tsoff = tcp_new_ts_offset(inc);
2153 }
2154
2155 if (to->to_flags & TOF_SIGNATURE)
2156 sc->sc_flags |= SCF_SIGNATURE;
2157
2158 sc->sc_rxmits = 0;
2159
2160 TCPSTAT_INC(tcps_sc_recvcookie);
2161 return (sc);
2162 }
2163
2164 #ifdef INVARIANTS
2165 static int
2166 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
2167 struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
2168 struct socket *lso)
2169 {
2170 struct syncache scs, *scx;
2171 char *s;
2172
2173 bzero(&scs, sizeof(scs));
2174 scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
2175
2176 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
2177 return (0);
2178
2179 if (scx != NULL) {
2180 if (sc->sc_peer_mss != scx->sc_peer_mss)
2181 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
2182 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
2183
2184 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
2185 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
2186 s, __func__, sc->sc_requested_r_scale,
2187 scx->sc_requested_r_scale);
2188
2189 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
2190 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
2191 s, __func__, sc->sc_requested_s_scale,
2192 scx->sc_requested_s_scale);
2193
2194 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
2195 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
2196 }
2197
2198 if (s != NULL)
2199 free(s, M_TCPLOG);
2200 return (0);
2201 }
2202 #endif /* INVARIANTS */
2203
2204 static void
2205 syncookie_reseed(void *arg)
2206 {
2207 struct tcp_syncache *sc = arg;
2208 uint8_t *secbits;
2209 int secbit;
2210
2211 /*
2212 * Reseeding the secret doesn't have to be protected by a lock.
2213 * It only must be ensured that the new random values are visible
2214 * to all CPUs in a SMP environment. The atomic with release
2215 * semantics ensures that.
2216 */
2217 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
2218 secbits = sc->secret.key[secbit];
2219 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
2220 atomic_add_rel_int(&sc->secret.oddeven, 1);
2221
2222 /* Reschedule ourself. */
2223 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
2224 }
2225
2226 /*
2227 * Exports the syncache entries to userland so that netstat can display
2228 * them alongside the other sockets. This function is intended to be
2229 * called only from tcp_pcblist.
2230 *
2231 * Due to concurrency on an active system, the number of pcbs exported
2232 * may have no relation to max_pcbs. max_pcbs merely indicates the
2233 * amount of space the caller allocated for this function to use.
2234 */
2235 int
2236 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
2237 {
2238 struct xtcpcb xt;
2239 struct syncache *sc;
2240 struct syncache_head *sch;
2241 int count, error, i;
2242
2243 for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
2244 sch = &V_tcp_syncache.hashbase[i];
2245 SCH_LOCK(sch);
2246 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
2247 if (count >= max_pcbs) {
2248 SCH_UNLOCK(sch);
2249 goto exit;
2250 }
2251 if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
2252 continue;
2253 bzero(&xt, sizeof(xt));
2254 xt.xt_len = sizeof(xt);
2255 if (sc->sc_inc.inc_flags & INC_ISIPV6)
2256 xt.xt_inp.inp_vflag = INP_IPV6;
2257 else
2258 xt.xt_inp.inp_vflag = INP_IPV4;
2259 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
2260 sizeof (struct in_conninfo));
2261 xt.t_state = TCPS_SYN_RECEIVED;
2262 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
2263 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
2264 xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
2265 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
2266 error = SYSCTL_OUT(req, &xt, sizeof xt);
2267 if (error) {
2268 SCH_UNLOCK(sch);
2269 goto exit;
2270 }
2271 count++;
2272 }
2273 SCH_UNLOCK(sch);
2274 }
2275 exit:
2276 *pcbs_exported = count;
2277 return error;
2278 }
Cache object: 67035c94c8a328adcf36c5cf0b42bdeb
|