1/*
2 * Copyright (c) 2015-2021 Apple Inc. All rights reserved.
3 *
4 * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
5 *
6 * This file contains Original Code and/or Modifications of Original Code
7 * as defined in and that are subject to the Apple Public Source License
8 * Version 2.0 (the 'License'). You may not use this file except in
9 * compliance with the License. The rights granted to you under the License
10 * may not be used to create, or enable the creation or redistribution of,
11 * unlawful or unlicensed copies of an Apple operating system, or to
12 * circumvent, violate, or enable the circumvention or violation of, any
13 * terms of an Apple operating system software license agreement.
14 *
15 * Please obtain a copy of the License at
16 * http://www.opensource.apple.com/apsl/ and read it before using this file.
17 *
18 * The Original Code and all software distributed under the License are
19 * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
20 * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
21 * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
22 * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
23 * Please see the License for the specific language governing rights and
24 * limitations under the License.
25 *
26 * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
27 */
28
29/* TCP-cache to store and retrieve TCP-related information */
30
31#include <net/flowhash.h>
32#include <net/route.h>
33#include <net/necp.h>
34#include <netinet/in_pcb.h>
35#include <netinet/mptcp.h>
36#include <netinet/mptcp_var.h>
37#include <netinet/tcp_cache.h>
38#include <netinet/tcp_seq.h>
39#include <netinet/tcp_var.h>
40#include <kern/locks.h>
41#include <sys/queue.h>
42#include <dev/random/randomdev.h>
43#include <net/sockaddr_utils.h>
44
45typedef union {
46 struct in_addr addr;
47 struct in6_addr addr6;
48} in_4_6_addr;
49
50struct tcp_heuristic_key {
51 union {
52 uint8_t thk_net_signature[IFNET_SIGNATURELEN];
53 in_4_6_addr thk_ip;
54 };
55 sa_family_t thk_family;
56};
57
58struct tcp_heuristic {
59 SLIST_ENTRY(tcp_heuristic) list;
60
61 uint32_t th_last_access;
62
63 struct tcp_heuristic_key th_key;
64
65 char th_val_start[0]; /* Marker for memsetting to 0 */
66
67 uint8_t th_tfo_data_loss; /* The number of times a SYN+data has been lost */
68 uint8_t th_tfo_req_loss; /* The number of times a SYN+cookie-req has been lost */
69 uint8_t th_tfo_data_rst; /* The number of times a SYN+data has received a RST */
70 uint8_t th_tfo_req_rst; /* The number of times a SYN+cookie-req has received a RST */
71 uint8_t th_mptcp_loss; /* The number of times a SYN+MP_CAPABLE has been lost */
72 uint8_t th_mptcp_success; /* The number of times MPTCP-negotiation has been successful */
73 uint8_t th_ecn_loss; /* The number of times a SYN+ecn has been lost */
74 uint8_t th_ecn_aggressive; /* The number of times we did an aggressive fallback */
75 uint8_t th_ecn_droprst; /* The number of times ECN connections received a RST after first data pkt */
76 uint8_t th_ecn_droprxmt; /* The number of times ECN connection is dropped after multiple retransmits */
77 uint8_t th_ecn_synrst; /* number of times RST was received in response to an ECN enabled SYN */
78 uint32_t th_tfo_enabled_time; /* The moment when we reenabled TFO after backing off */
79 uint32_t th_tfo_backoff_until; /* Time until when we should not try out TFO */
80 uint32_t th_tfo_backoff; /* Current backoff timer */
81 uint32_t th_mptcp_backoff; /* Time until when we should not try out MPTCP */
82 uint32_t th_ecn_backoff; /* Time until when we should not try out ECN */
83
84 uint8_t th_tfo_in_backoff:1, /* Are we avoiding TFO due to the backoff timer? */
85 th_mptcp_in_backoff:1, /* Are we avoiding MPTCP due to the backoff timer? */
86 th_mptcp_heuristic_disabled:1; /* Are heuristics disabled? */
87
88 char th_val_end[0]; /* Marker for memsetting to 0 */
89};
90
91struct tcp_heuristics_head {
92 SLIST_HEAD(tcp_heur_bucket, tcp_heuristic) tcp_heuristics;
93
94 /* Per-hashbucket lock to avoid lock-contention */
95 lck_mtx_t thh_mtx;
96};
97
98struct tcp_cache_key {
99 sa_family_t tck_family;
100
101 struct tcp_heuristic_key tck_src;
102 in_4_6_addr tck_dst;
103};
104
105#define MPTCP_VERSION_SUPPORTED 1
106#define MPTCP_VERSION_UNSUPPORTED -1
107#define MPTCP_VERSION_SUPPORTED_UNKNOWN 0
108struct tcp_cache {
109 SLIST_ENTRY(tcp_cache) list;
110
111 uint32_t tc_last_access;
112
113 struct tcp_cache_key tc_key;
114
115 uint8_t tc_tfo_cookie[TFO_COOKIE_LEN_MAX];
116 uint8_t tc_tfo_cookie_len;
117
118 uint8_t tc_mptcp_version_confirmed:1;
119 uint8_t tc_mptcp_version; /* version to use right now */
120 uint32_t tc_mptcp_next_version_try; /* Time, until we try preferred version again */
121};
122
123struct tcp_cache_head {
124 SLIST_HEAD(tcp_cache_bucket, tcp_cache) tcp_caches;
125
126 /* Per-hashbucket lock to avoid lock-contention */
127 lck_mtx_t tch_mtx;
128};
129
130struct tcp_cache_key_src {
131 struct ifnet *ifp;
132 in_4_6_addr laddr;
133 in_4_6_addr faddr;
134 int af;
135};
136
137static uint32_t tcp_cache_hash_seed;
138
139size_t tcp_cache_size;
140
141/*
142 * The maximum depth of the hash-bucket. This way we limit the tcp_cache to
143 * TCP_CACHE_BUCKET_SIZE * tcp_cache_size and have "natural" garbage collection
144 */
145#define TCP_CACHE_BUCKET_SIZE 5
146
147static struct tcp_cache_head *tcp_cache;
148
149static LCK_ATTR_DECLARE(tcp_cache_mtx_attr, 0, 0);
150static LCK_GRP_DECLARE(tcp_cache_mtx_grp, "tcpcache");
151
152static struct tcp_heuristics_head *tcp_heuristics;
153
154static LCK_ATTR_DECLARE(tcp_heuristic_mtx_attr, 0, 0);
155static LCK_GRP_DECLARE(tcp_heuristic_mtx_grp, "tcpheuristic");
156
157static uint32_t tcp_backoff_maximum = 65536;
158
159SYSCTL_UINT(_net_inet_tcp, OID_AUTO, backoff_maximum, CTLFLAG_RW | CTLFLAG_LOCKED,
160 &tcp_backoff_maximum, 0, "Maximum time for which we won't try TFO");
161
162static uint32_t tcp_ecn_timeout = 60;
163
164SYSCTL_UINT(_net_inet_tcp, OID_AUTO, ecn_timeout, CTLFLAG_RW | CTLFLAG_LOCKED,
165 &tcp_ecn_timeout, 60, "Initial minutes to wait before re-trying ECN");
166
167static int disable_tcp_heuristics = 0;
168SYSCTL_INT(_net_inet_tcp, OID_AUTO, disable_tcp_heuristics, CTLFLAG_RW | CTLFLAG_LOCKED,
169 &disable_tcp_heuristics, 0, "Set to 1, to disable all TCP heuristics (TFO, ECN, MPTCP)");
170
171static uint32_t mptcp_version_timeout = 24 * 60;
172
173SYSCTL_UINT(_net_inet_tcp, OID_AUTO, mptcp_version_timeout, CTLFLAG_RW | CTLFLAG_LOCKED,
174 &mptcp_version_timeout, 24 * 60, "Initial minutes to wait before re-trying MPTCP's preferred version");
175
176
177static uint32_t
178tcp_min_to_hz(uint32_t minutes)
179{
180 if (minutes > 65536) {
181 return (uint32_t)65536 * 60 * TCP_RETRANSHZ;
182 }
183
184 return minutes * 60 * TCP_RETRANSHZ;
185}
186
187/*
188 * This number is coupled with tcp_ecn_timeout, because we want to prevent
189 * integer overflow. Need to find an unexpensive way to prevent integer overflow
190 * while still allowing a dynamic sysctl.
191 */
192#define TCP_CACHE_OVERFLOW_PROTECT 9
193
194/* Number of SYN-losses we accept */
195#define TFO_MAX_COOKIE_LOSS 2
196#define ECN_MAX_SYN_LOSS 2
197#define MPTCP_MAX_SYN_LOSS 2
198#define MPTCP_SUCCESS_TRIGGER 10
199#define MPTCP_VERSION_MAX_FAIL 2
200#define ECN_MAX_DROPRST 1
201#define ECN_MAX_DROPRXMT 4
202#define ECN_MAX_SYNRST 4
203
204/* Flags for setting/unsetting loss-heuristics, limited to 4 bytes */
205#define TCPCACHE_F_TFO_REQ 0x01
206#define TCPCACHE_F_TFO_DATA 0x02
207#define TCPCACHE_F_ECN 0x04
208#define TCPCACHE_F_MPTCP 0x08
209#define TCPCACHE_F_ECN_DROPRST 0x10
210#define TCPCACHE_F_ECN_DROPRXMT 0x20
211#define TCPCACHE_F_TFO_REQ_RST 0x40
212#define TCPCACHE_F_TFO_DATA_RST 0x80
213#define TCPCACHE_F_ECN_SYNRST 0x100
214
215/* Always retry ECN after backing off to this level for some heuristics */
216#define ECN_RETRY_LIMIT 9
217
218#define TCP_CACHE_INC_IFNET_STAT(_ifp_, _af_, _stat_) { \
219 if ((_ifp_) != NULL) { \
220 if ((_af_) == AF_INET6) { \
221 (_ifp_)->if_ipv6_stat->_stat_++;\
222 } else { \
223 (_ifp_)->if_ipv4_stat->_stat_++;\
224 }\
225 }\
226}
227
228/*
229 * Round up to next higher power-of 2. See "Bit Twiddling Hacks".
230 *
231 * Might be worth moving this to a library so that others
232 * (e.g., scale_to_powerof2()) can use this as well instead of a while-loop.
233 */
234static uint32_t
235tcp_cache_roundup2(uint32_t a)
236{
237 a--;
238 a |= a >> 1;
239 a |= a >> 2;
240 a |= a >> 4;
241 a |= a >> 8;
242 a |= a >> 16;
243 a++;
244
245 return a;
246}
247
248static void
249tcp_cache_hash_src(struct tcp_cache_key_src *tcks, struct tcp_heuristic_key *key)
250{
251 struct ifnet *ifp = tcks->ifp;
252 uint8_t len = sizeof(key->thk_net_signature);
253 uint16_t flags;
254
255 if (tcks->af == AF_INET6) {
256 int ret;
257
258 key->thk_family = AF_INET6;
259 ret = ifnet_get_netsignature(ifp, AF_INET6, &len, &flags,
260 key->thk_net_signature);
261
262 /*
263 * ifnet_get_netsignature only returns EINVAL if ifn is NULL
264 * (we made sure that in the other cases it does not). So,
265 * in this case we should take the connection's address.
266 */
267 if (ret == ENOENT || ret == EINVAL) {
268 memcpy(dst: &key->thk_ip.addr6, src: &tcks->laddr.addr6, n: sizeof(struct in6_addr));
269 }
270 } else {
271 int ret;
272
273 key->thk_family = AF_INET;
274 ret = ifnet_get_netsignature(ifp, AF_INET, &len, &flags,
275 key->thk_net_signature);
276
277 /*
278 * ifnet_get_netsignature only returns EINVAL if ifn is NULL
279 * (we made sure that in the other cases it does not). So,
280 * in this case we should take the connection's address.
281 */
282 if (ret == ENOENT || ret == EINVAL) {
283 memcpy(dst: &key->thk_ip.addr, src: &tcks->laddr.addr, n: sizeof(struct in_addr));
284 }
285 }
286}
287
288static uint16_t
289tcp_cache_hash(struct tcp_cache_key_src *tcks, struct tcp_cache_key *key)
290{
291 uint32_t hash;
292
293 bzero(s: key, n: sizeof(struct tcp_cache_key));
294
295 tcp_cache_hash_src(tcks, key: &key->tck_src);
296
297 if (tcks->af == AF_INET6) {
298 key->tck_family = AF_INET6;
299 memcpy(dst: &key->tck_dst.addr6, src: &tcks->faddr.addr6,
300 n: sizeof(struct in6_addr));
301 } else {
302 key->tck_family = AF_INET;
303 memcpy(dst: &key->tck_dst.addr, src: &tcks->faddr.addr,
304 n: sizeof(struct in_addr));
305 }
306
307 hash = net_flowhash(key, sizeof(struct tcp_cache_key),
308 tcp_cache_hash_seed);
309
310 return (uint16_t)(hash & (tcp_cache_size - 1));
311}
312
313static void
314tcp_cache_unlock(struct tcp_cache_head *head)
315{
316 lck_mtx_unlock(lck: &head->tch_mtx);
317}
318
319/*
320 * Make sure that everything that happens after tcp_getcache_with_lock()
321 * is short enough to justify that you hold the per-bucket lock!!!
322 *
323 * Otherwise, better build another lookup-function that does not hold the
324 * lock and you copy out the bits and bytes.
325 *
326 * That's why we provide the head as a "return"-pointer so that the caller
327 * can give it back to use for tcp_cache_unlock().
328 */
329static struct tcp_cache *
330tcp_getcache_with_lock(struct tcp_cache_key_src *tcks,
331 int create, struct tcp_cache_head **headarg)
332{
333 struct tcp_cache *tpcache = NULL;
334 struct tcp_cache_head *head;
335 struct tcp_cache_key key;
336 uint16_t hash;
337 int i = 0;
338
339 hash = tcp_cache_hash(tcks, key: &key);
340 head = &tcp_cache[hash];
341
342 lck_mtx_lock(lck: &head->tch_mtx);
343
344 /*** First step: Look for the tcp_cache in our bucket ***/
345 SLIST_FOREACH(tpcache, &head->tcp_caches, list) {
346 if (memcmp(s1: &tpcache->tc_key, s2: &key, n: sizeof(key)) == 0) {
347 break;
348 }
349
350 i++;
351 }
352
353 /*** Second step: If it's not there, create/recycle it ***/
354 if ((tpcache == NULL) && create) {
355 if (i >= TCP_CACHE_BUCKET_SIZE) {
356 struct tcp_cache *oldest_cache = NULL;
357 uint32_t max_age = 0;
358
359 /* Look for the oldest tcp_cache in the bucket */
360 SLIST_FOREACH(tpcache, &head->tcp_caches, list) {
361 uint32_t age = tcp_now - tpcache->tc_last_access;
362 if (age > max_age) {
363 max_age = age;
364 oldest_cache = tpcache;
365 }
366 }
367 VERIFY(oldest_cache != NULL);
368
369 tpcache = oldest_cache;
370
371 /* We recycle, thus let's indicate that there is no cookie */
372 tpcache->tc_tfo_cookie_len = 0;
373 } else {
374 /* Create a new cache and add it to the list */
375 tpcache = kalloc_type(struct tcp_cache, Z_NOPAGEWAIT | Z_ZERO);
376 if (tpcache == NULL) {
377 os_log_error(OS_LOG_DEFAULT, "%s could not allocate cache", __func__);
378 goto out_null;
379 }
380
381 tpcache->tc_mptcp_version = (uint8_t)mptcp_preferred_version;
382 tpcache->tc_mptcp_next_version_try = tcp_now;
383
384 SLIST_INSERT_HEAD(&head->tcp_caches, tpcache, list);
385 }
386
387 memcpy(dst: &tpcache->tc_key, src: &key, n: sizeof(key));
388 }
389
390 if (tpcache == NULL) {
391 goto out_null;
392 }
393
394 /* Update timestamp for garbage collection purposes */
395 tpcache->tc_last_access = tcp_now;
396 *headarg = head;
397
398 return tpcache;
399
400out_null:
401 tcp_cache_unlock(head);
402 return NULL;
403}
404
405static void
406tcp_cache_key_src_create(struct tcpcb *tp, struct tcp_cache_key_src *tcks)
407{
408 struct inpcb *inp = tp->t_inpcb;
409 memset(s: tcks, c: 0, n: sizeof(*tcks));
410
411 tcks->ifp = inp->inp_last_outifp;
412
413 if (inp->inp_vflag & INP_IPV6) {
414 memcpy(dst: &tcks->laddr.addr6, src: &inp->in6p_laddr, n: sizeof(struct in6_addr));
415 memcpy(dst: &tcks->faddr.addr6, src: &inp->in6p_faddr, n: sizeof(struct in6_addr));
416 tcks->af = AF_INET6;
417 } else {
418 memcpy(dst: &tcks->laddr.addr, src: &inp->inp_laddr, n: sizeof(struct in_addr));
419 memcpy(dst: &tcks->faddr.addr, src: &inp->inp_faddr, n: sizeof(struct in_addr));
420 tcks->af = AF_INET;
421 }
422
423 return;
424}
425
426static void
427mptcp_version_cache_key_src_init(struct sockaddr *dst, struct tcp_cache_key_src *tcks)
428{
429 memset(s: tcks, c: 0, n: sizeof(*tcks));
430
431 if (dst->sa_family == AF_INET) {
432 memcpy(dst: &tcks->faddr.addr, src: &SIN(dst)->sin_addr, n: sizeof(struct in_addr));
433 tcks->af = AF_INET;
434 } else {
435 memcpy(dst: &tcks->faddr.addr6, src: &SIN6(dst)->sin6_addr, n: sizeof(struct in6_addr));
436 tcks->af = AF_INET6;
437 }
438
439 return;
440}
441
442static void
443tcp_cache_set_cookie_common(struct tcp_cache_key_src *tcks, u_char *cookie, uint8_t len)
444{
445 struct tcp_cache_head *head;
446 struct tcp_cache *tpcache;
447
448 /* Call lookup/create function */
449 tpcache = tcp_getcache_with_lock(tcks, create: 1, headarg: &head);
450 if (tpcache == NULL) {
451 return;
452 }
453
454 tpcache->tc_tfo_cookie_len = len > TFO_COOKIE_LEN_MAX ?
455 TFO_COOKIE_LEN_MAX : len;
456 memcpy(dst: tpcache->tc_tfo_cookie, src: cookie, n: tpcache->tc_tfo_cookie_len);
457
458 tcp_cache_unlock(head);
459}
460
461void
462tcp_cache_set_cookie(struct tcpcb *tp, u_char *cookie, uint8_t len)
463{
464 struct tcp_cache_key_src tcks;
465
466 tcp_cache_key_src_create(tp, tcks: &tcks);
467 tcp_cache_set_cookie_common(tcks: &tcks, cookie, len);
468}
469
470static int
471tcp_cache_get_cookie_common(struct tcp_cache_key_src *tcks, u_char *cookie, uint8_t *len)
472{
473 struct tcp_cache_head *head;
474 struct tcp_cache *tpcache;
475
476 /* Call lookup/create function */
477 tpcache = tcp_getcache_with_lock(tcks, create: 1, headarg: &head);
478 if (tpcache == NULL) {
479 return 0;
480 }
481
482 if (tpcache->tc_tfo_cookie_len == 0) {
483 tcp_cache_unlock(head);
484 return 0;
485 }
486
487 /*
488 * Not enough space - this should never happen as it has been checked
489 * in tcp_tfo_check. So, fail here!
490 */
491 VERIFY(tpcache->tc_tfo_cookie_len <= *len);
492
493 memcpy(dst: cookie, src: tpcache->tc_tfo_cookie, n: tpcache->tc_tfo_cookie_len);
494 *len = tpcache->tc_tfo_cookie_len;
495
496 tcp_cache_unlock(head);
497
498 return 1;
499}
500
501/*
502 * Get the cookie related to 'tp', and copy it into 'cookie', provided that len
503 * is big enough (len designates the available memory.
504 * Upon return, 'len' is set to the cookie's length.
505 *
506 * Returns 0 if we should request a cookie.
507 * Returns 1 if the cookie has been found and written.
508 */
509int
510tcp_cache_get_cookie(struct tcpcb *tp, u_char *cookie, uint8_t *len)
511{
512 struct tcp_cache_key_src tcks;
513
514 tcp_cache_key_src_create(tp, tcks: &tcks);
515 return tcp_cache_get_cookie_common(tcks: &tcks, cookie, len);
516}
517
518static unsigned int
519tcp_cache_get_cookie_len_common(struct tcp_cache_key_src *tcks)
520{
521 struct tcp_cache_head *head;
522 struct tcp_cache *tpcache;
523 unsigned int cookie_len;
524
525 /* Call lookup/create function */
526 tpcache = tcp_getcache_with_lock(tcks, create: 1, headarg: &head);
527 if (tpcache == NULL) {
528 return 0;
529 }
530
531 cookie_len = tpcache->tc_tfo_cookie_len;
532
533 tcp_cache_unlock(head);
534
535 return cookie_len;
536}
537
538unsigned int
539tcp_cache_get_cookie_len(struct tcpcb *tp)
540{
541 struct tcp_cache_key_src tcks;
542
543 tcp_cache_key_src_create(tp, tcks: &tcks);
544 return tcp_cache_get_cookie_len_common(tcks: &tcks);
545}
546
547/*
548 * @return:
549 * 0 MPTCP_VERSION_0
550 * 1 MPTCP_VERSION_1
551 */
552uint8_t
553tcp_cache_get_mptcp_version(struct sockaddr *dst)
554{
555 struct tcp_cache_key_src tcks;
556 mptcp_version_cache_key_src_init(dst, tcks: &tcks);
557 uint8_t version = (uint8_t) mptcp_preferred_version;
558
559 struct tcp_cache_head *head;
560 struct tcp_cache *tpcache;
561
562 /* Call lookup/create function */
563 tpcache = tcp_getcache_with_lock(tcks: &tcks, create: 1, headarg: &head);
564 if (tpcache == NULL) {
565 return version;
566 }
567
568 version = tpcache->tc_mptcp_version;
569
570 /* Let's see if we should try the preferred version again */
571 if (!tpcache->tc_mptcp_version_confirmed &&
572 version != mptcp_preferred_version &&
573 TSTMP_GEQ(tcp_now, tpcache->tc_mptcp_next_version_try)) {
574 version = (uint8_t) mptcp_preferred_version;
575 }
576
577 tcp_cache_unlock(head);
578 return version;
579}
580
581void
582tcp_cache_update_mptcp_version(struct tcpcb *tp, boolean_t succeeded)
583{
584 uint8_t version = tptomptp(tp)->mpt_version;
585 struct inpcb *inp = tp->t_inpcb;
586 struct tcp_cache_key_src tcks;
587 struct tcp_cache_head *head;
588 struct tcp_cache *tpcache;
589
590 if (inp->inp_vflag & INP_IPV6) {
591 struct sockaddr_in6 dst = {
592 .sin6_len = sizeof(struct sockaddr_in6),
593 .sin6_family = AF_INET6,
594 .sin6_addr = inp->in6p_faddr,
595 };
596 mptcp_version_cache_key_src_init(SA(&dst), tcks: &tcks);
597 } else {
598 struct sockaddr_in dst = {
599 .sin_len = sizeof(struct sockaddr_in),
600 .sin_family = AF_INET,
601 .sin_addr = inp->inp_faddr,
602 };
603 mptcp_version_cache_key_src_init(SA(&dst), tcks: &tcks);
604 }
605
606 /* Call lookup/create function */
607 tpcache = tcp_getcache_with_lock(tcks: &tcks, create: 1, headarg: &head);
608 if (tpcache == NULL) {
609 return;
610 }
611
612 /* We are still in probing phase */
613 if (tpcache->tc_mptcp_version_confirmed) {
614 goto exit;
615 }
616
617 if (succeeded) {
618 if (version == (uint8_t)mptcp_preferred_version) {
619 /* Preferred version succeeded - make it sticky */
620 tpcache->tc_mptcp_version_confirmed = true;
621 tpcache->tc_mptcp_version = version;
622 } else {
623 /* If we are past the next version try, set it
624 * so that we try preferred again in 24h
625 */
626 if (TSTMP_GEQ(tcp_now, tpcache->tc_mptcp_next_version_try)) {
627 tpcache->tc_mptcp_next_version_try = tcp_now + tcp_min_to_hz(minutes: mptcp_version_timeout);
628 }
629 }
630 } else {
631 if (version == (uint8_t)mptcp_preferred_version) {
632 /* Preferred version failed - try the other version */
633 tpcache->tc_mptcp_version = version == MPTCP_VERSION_0 ? MPTCP_VERSION_1 : MPTCP_VERSION_0;
634 }
635 /* Preferred version failed - make sure we give the preferred another
636 * shot in 24h.
637 */
638 if (TSTMP_GEQ(tcp_now, tpcache->tc_mptcp_next_version_try)) {
639 tpcache->tc_mptcp_next_version_try = tcp_now + tcp_min_to_hz(minutes: mptcp_version_timeout);
640 }
641 }
642
643exit:
644 tcp_cache_unlock(head);
645}
646
647static uint16_t
648tcp_heuristics_hash(struct tcp_cache_key_src *tcks, struct tcp_heuristic_key *key)
649{
650 uint32_t hash;
651
652 bzero(s: key, n: sizeof(struct tcp_heuristic_key));
653
654 tcp_cache_hash_src(tcks, key);
655
656 hash = net_flowhash(key, sizeof(struct tcp_heuristic_key),
657 tcp_cache_hash_seed);
658
659 return (uint16_t)(hash & (tcp_cache_size - 1));
660}
661
662static void
663tcp_heuristic_unlock(struct tcp_heuristics_head *head)
664{
665 lck_mtx_unlock(lck: &head->thh_mtx);
666}
667
668/*
669 * Make sure that everything that happens after tcp_getheuristic_with_lock()
670 * is short enough to justify that you hold the per-bucket lock!!!
671 *
672 * Otherwise, better build another lookup-function that does not hold the
673 * lock and you copy out the bits and bytes.
674 *
675 * That's why we provide the head as a "return"-pointer so that the caller
676 * can give it back to use for tcp_heur_unlock().
677 *
678 *
679 * ToDo - way too much code-duplication. We should create an interface to handle
680 * bucketized hashtables with recycling of the oldest element.
681 */
682static struct tcp_heuristic *
683tcp_getheuristic_with_lock(struct tcp_cache_key_src *tcks,
684 int create, struct tcp_heuristics_head **headarg)
685{
686 struct tcp_heuristic *tpheur = NULL;
687 struct tcp_heuristics_head *head;
688 struct tcp_heuristic_key key;
689 uint16_t hash;
690 int i = 0;
691
692 hash = tcp_heuristics_hash(tcks, key: &key);
693 head = &tcp_heuristics[hash];
694
695 lck_mtx_lock(lck: &head->thh_mtx);
696
697 /*** First step: Look for the tcp_heur in our bucket ***/
698 SLIST_FOREACH(tpheur, &head->tcp_heuristics, list) {
699 if (memcmp(s1: &tpheur->th_key, s2: &key, n: sizeof(key)) == 0) {
700 break;
701 }
702
703 i++;
704 }
705
706 /*** Second step: If it's not there, create/recycle it ***/
707 if ((tpheur == NULL) && create) {
708 if (i >= TCP_CACHE_BUCKET_SIZE) {
709 struct tcp_heuristic *oldest_heur = NULL;
710 uint32_t max_age = 0;
711
712 /* Look for the oldest tcp_heur in the bucket */
713 SLIST_FOREACH(tpheur, &head->tcp_heuristics, list) {
714 uint32_t age = tcp_now - tpheur->th_last_access;
715 if (age > max_age) {
716 max_age = age;
717 oldest_heur = tpheur;
718 }
719 }
720 VERIFY(oldest_heur != NULL);
721
722 tpheur = oldest_heur;
723
724 /* We recycle - set everything to 0 */
725 bzero(s: tpheur->th_val_start,
726 n: tpheur->th_val_end - tpheur->th_val_start);
727 } else {
728 /* Create a new heuristic and add it to the list */
729 tpheur = kalloc_type(struct tcp_heuristic, Z_NOPAGEWAIT | Z_ZERO);
730 if (tpheur == NULL) {
731 os_log_error(OS_LOG_DEFAULT, "%s could not allocate heuristic", __func__);
732 goto out_null;
733 }
734
735 SLIST_INSERT_HEAD(&head->tcp_heuristics, tpheur, list);
736 }
737
738 /*
739 * Set to tcp_now, to make sure it won't be > than tcp_now in the
740 * near future.
741 */
742 tpheur->th_ecn_backoff = tcp_now;
743 tpheur->th_tfo_backoff_until = tcp_now;
744 tpheur->th_mptcp_backoff = tcp_now;
745 tpheur->th_tfo_backoff = tcp_min_to_hz(minutes: tcp_ecn_timeout);
746
747 memcpy(dst: &tpheur->th_key, src: &key, n: sizeof(key));
748 }
749
750 if (tpheur == NULL) {
751 goto out_null;
752 }
753
754 /* Update timestamp for garbage collection purposes */
755 tpheur->th_last_access = tcp_now;
756 *headarg = head;
757
758 return tpheur;
759
760out_null:
761 tcp_heuristic_unlock(head);
762 return NULL;
763}
764
765static void
766tcp_heuristic_reset_counters(struct tcp_cache_key_src *tcks, uint8_t flags)
767{
768 struct tcp_heuristics_head *head;
769 struct tcp_heuristic *tpheur;
770
771 /*
772 * Always create heuristics here because MPTCP needs to write success
773 * into it. Thus, we always end up creating them.
774 */
775 tpheur = tcp_getheuristic_with_lock(tcks, create: 1, headarg: &head);
776 if (tpheur == NULL) {
777 return;
778 }
779
780 if (flags & TCPCACHE_F_TFO_DATA) {
781 if (tpheur->th_tfo_data_loss >= TFO_MAX_COOKIE_LOSS) {
782 os_log(OS_LOG_DEFAULT, "%s: Resetting TFO-data loss to 0 from %u on heur %lx\n",
783 __func__, tpheur->th_tfo_data_loss, (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
784 }
785 tpheur->th_tfo_data_loss = 0;
786 }
787
788 if (flags & TCPCACHE_F_TFO_REQ) {
789 if (tpheur->th_tfo_req_loss >= TFO_MAX_COOKIE_LOSS) {
790 os_log(OS_LOG_DEFAULT, "%s: Resetting TFO-req loss to 0 from %u on heur %lx\n",
791 __func__, tpheur->th_tfo_req_loss, (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
792 }
793 tpheur->th_tfo_req_loss = 0;
794 }
795
796 if (flags & TCPCACHE_F_TFO_DATA_RST) {
797 if (tpheur->th_tfo_data_rst >= TFO_MAX_COOKIE_LOSS) {
798 os_log(OS_LOG_DEFAULT, "%s: Resetting TFO-data RST to 0 from %u on heur %lx\n",
799 __func__, tpheur->th_tfo_data_rst, (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
800 }
801 tpheur->th_tfo_data_rst = 0;
802 }
803
804 if (flags & TCPCACHE_F_TFO_REQ_RST) {
805 if (tpheur->th_tfo_req_rst >= TFO_MAX_COOKIE_LOSS) {
806 os_log(OS_LOG_DEFAULT, "%s: Resetting TFO-req RST to 0 from %u on heur %lx\n",
807 __func__, tpheur->th_tfo_req_rst, (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
808 }
809 tpheur->th_tfo_req_rst = 0;
810 }
811
812 if (flags & TCPCACHE_F_ECN) {
813 if (tpheur->th_ecn_loss >= ECN_MAX_SYN_LOSS || tpheur->th_ecn_synrst >= ECN_MAX_SYNRST) {
814 os_log(OS_LOG_DEFAULT, "%s: Resetting ECN-loss to 0 from %u and synrst from %u on heur %lx\n",
815 __func__, tpheur->th_ecn_loss, tpheur->th_ecn_synrst, (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
816 }
817 tpheur->th_ecn_loss = 0;
818 tpheur->th_ecn_synrst = 0;
819 }
820
821 if (flags & TCPCACHE_F_MPTCP) {
822 tpheur->th_mptcp_loss = 0;
823 if (tpheur->th_mptcp_success < MPTCP_SUCCESS_TRIGGER) {
824 tpheur->th_mptcp_success++;
825
826 if (tpheur->th_mptcp_success == MPTCP_SUCCESS_TRIGGER) {
827 os_log(mptcp_log_handle, "%s disabling heuristics for 12 hours", __func__);
828 tpheur->th_mptcp_heuristic_disabled = 1;
829 /* Disable heuristics for 12 hours */
830 tpheur->th_mptcp_backoff = tcp_now + tcp_min_to_hz(minutes: tcp_ecn_timeout * 12);
831 }
832 }
833 }
834
835 tcp_heuristic_unlock(head);
836}
837
838void
839tcp_heuristic_tfo_success(struct tcpcb *tp)
840{
841 struct tcp_cache_key_src tcks;
842 uint8_t flag = 0;
843
844 tcp_cache_key_src_create(tp, tcks: &tcks);
845
846 if (tp->t_tfo_stats & TFO_S_SYN_DATA_SENT) {
847 flag = (TCPCACHE_F_TFO_DATA | TCPCACHE_F_TFO_REQ |
848 TCPCACHE_F_TFO_DATA_RST | TCPCACHE_F_TFO_REQ_RST);
849 }
850 if (tp->t_tfo_stats & TFO_S_COOKIE_REQ) {
851 flag = (TCPCACHE_F_TFO_REQ | TCPCACHE_F_TFO_REQ_RST);
852 }
853
854 tcp_heuristic_reset_counters(tcks: &tcks, flags: flag);
855}
856
857void
858tcp_heuristic_mptcp_success(struct tcpcb *tp)
859{
860 struct tcp_cache_key_src tcks;
861
862 tcp_cache_key_src_create(tp, tcks: &tcks);
863 tcp_heuristic_reset_counters(tcks: &tcks, TCPCACHE_F_MPTCP);
864}
865
866void
867tcp_heuristic_ecn_success(struct tcpcb *tp)
868{
869 struct tcp_cache_key_src tcks;
870
871 tcp_cache_key_src_create(tp, tcks: &tcks);
872 tcp_heuristic_reset_counters(tcks: &tcks, TCPCACHE_F_ECN);
873}
874
875static void
876__tcp_heuristic_tfo_middlebox_common(struct tcp_heuristic *tpheur)
877{
878 if (tpheur->th_tfo_in_backoff) {
879 return;
880 }
881
882 tpheur->th_tfo_in_backoff = 1;
883
884 if (tpheur->th_tfo_enabled_time) {
885 uint32_t old_backoff = tpheur->th_tfo_backoff;
886
887 tpheur->th_tfo_backoff -= (tcp_now - tpheur->th_tfo_enabled_time);
888 if (tpheur->th_tfo_backoff > old_backoff) {
889 tpheur->th_tfo_backoff = tcp_min_to_hz(minutes: tcp_ecn_timeout);
890 }
891 }
892
893 tpheur->th_tfo_backoff_until = tcp_now + tpheur->th_tfo_backoff;
894
895 /* Then, increase the backoff time */
896 tpheur->th_tfo_backoff *= 2;
897
898 if (tpheur->th_tfo_backoff > tcp_min_to_hz(minutes: tcp_backoff_maximum)) {
899 tpheur->th_tfo_backoff = tcp_min_to_hz(minutes: tcp_ecn_timeout);
900 }
901
902 os_log(OS_LOG_DEFAULT, "%s disable TFO until %u now %u on %lx\n", __func__,
903 tpheur->th_tfo_backoff_until, tcp_now, (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
904}
905
906static void
907tcp_heuristic_tfo_middlebox_common(struct tcp_cache_key_src *tcks)
908{
909 struct tcp_heuristics_head *head;
910 struct tcp_heuristic *tpheur;
911
912 tpheur = tcp_getheuristic_with_lock(tcks, create: 1, headarg: &head);
913 if (tpheur == NULL) {
914 return;
915 }
916
917 __tcp_heuristic_tfo_middlebox_common(tpheur);
918
919 tcp_heuristic_unlock(head);
920}
921
922static void
923tcp_heuristic_inc_counters(struct tcp_cache_key_src *tcks,
924 uint32_t flags)
925{
926 struct tcp_heuristics_head *head;
927 struct tcp_heuristic *tpheur;
928
929 tpheur = tcp_getheuristic_with_lock(tcks, create: 1, headarg: &head);
930 if (tpheur == NULL) {
931 return;
932 }
933
934 /* Limit to prevent integer-overflow during exponential backoff */
935 if ((flags & TCPCACHE_F_TFO_DATA) && tpheur->th_tfo_data_loss < TCP_CACHE_OVERFLOW_PROTECT) {
936 tpheur->th_tfo_data_loss++;
937
938 if (tpheur->th_tfo_data_loss >= TFO_MAX_COOKIE_LOSS) {
939 __tcp_heuristic_tfo_middlebox_common(tpheur);
940 }
941 }
942
943 if ((flags & TCPCACHE_F_TFO_REQ) && tpheur->th_tfo_req_loss < TCP_CACHE_OVERFLOW_PROTECT) {
944 tpheur->th_tfo_req_loss++;
945
946 if (tpheur->th_tfo_req_loss >= TFO_MAX_COOKIE_LOSS) {
947 __tcp_heuristic_tfo_middlebox_common(tpheur);
948 }
949 }
950
951 if ((flags & TCPCACHE_F_TFO_DATA_RST) && tpheur->th_tfo_data_rst < TCP_CACHE_OVERFLOW_PROTECT) {
952 tpheur->th_tfo_data_rst++;
953
954 if (tpheur->th_tfo_data_rst >= TFO_MAX_COOKIE_LOSS) {
955 __tcp_heuristic_tfo_middlebox_common(tpheur);
956 }
957 }
958
959 if ((flags & TCPCACHE_F_TFO_REQ_RST) && tpheur->th_tfo_req_rst < TCP_CACHE_OVERFLOW_PROTECT) {
960 tpheur->th_tfo_req_rst++;
961
962 if (tpheur->th_tfo_req_rst >= TFO_MAX_COOKIE_LOSS) {
963 __tcp_heuristic_tfo_middlebox_common(tpheur);
964 }
965 }
966
967 if ((flags & TCPCACHE_F_ECN) &&
968 tpheur->th_ecn_loss < TCP_CACHE_OVERFLOW_PROTECT &&
969 TSTMP_LEQ(tpheur->th_ecn_backoff, tcp_now)) {
970 tpheur->th_ecn_loss++;
971 if (tpheur->th_ecn_loss >= ECN_MAX_SYN_LOSS) {
972 tcpstat.tcps_ecn_fallback_synloss++;
973 TCP_CACHE_INC_IFNET_STAT(tcks->ifp, tcks->af, ecn_fallback_synloss);
974 tpheur->th_ecn_backoff = tcp_now +
975 (tcp_min_to_hz(minutes: tcp_ecn_timeout) <<
976 (tpheur->th_ecn_loss - ECN_MAX_SYN_LOSS));
977
978 os_log(OS_LOG_DEFAULT, "%s disable ECN until %u now %u on %lx for SYN-loss\n",
979 __func__, tpheur->th_ecn_backoff, tcp_now,
980 (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
981 }
982 }
983
984 if ((flags & TCPCACHE_F_MPTCP) &&
985 tpheur->th_mptcp_loss < TCP_CACHE_OVERFLOW_PROTECT &&
986 tpheur->th_mptcp_heuristic_disabled == 0) {
987 tpheur->th_mptcp_loss++;
988 if (tpheur->th_mptcp_loss >= MPTCP_MAX_SYN_LOSS) {
989 /*
990 * Yes, we take tcp_ecn_timeout, to avoid adding yet
991 * another sysctl that is just used for testing.
992 */
993 tpheur->th_mptcp_backoff = tcp_now +
994 (tcp_min_to_hz(minutes: tcp_ecn_timeout) <<
995 (tpheur->th_mptcp_loss - MPTCP_MAX_SYN_LOSS));
996 tpheur->th_mptcp_in_backoff = 1;
997
998 os_log(OS_LOG_DEFAULT, "%s disable MPTCP until %u now %u on %lx\n",
999 __func__, tpheur->th_mptcp_backoff, tcp_now,
1000 (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
1001 }
1002 }
1003
1004 if ((flags & TCPCACHE_F_ECN_DROPRST) &&
1005 tpheur->th_ecn_droprst < TCP_CACHE_OVERFLOW_PROTECT &&
1006 TSTMP_LEQ(tpheur->th_ecn_backoff, tcp_now)) {
1007 tpheur->th_ecn_droprst++;
1008 if (tpheur->th_ecn_droprst >= ECN_MAX_DROPRST) {
1009 tcpstat.tcps_ecn_fallback_droprst++;
1010 TCP_CACHE_INC_IFNET_STAT(tcks->ifp, tcks->af,
1011 ecn_fallback_droprst);
1012 tpheur->th_ecn_backoff = tcp_now +
1013 (tcp_min_to_hz(minutes: tcp_ecn_timeout) <<
1014 (tpheur->th_ecn_droprst - ECN_MAX_DROPRST));
1015
1016 os_log(OS_LOG_DEFAULT, "%s disable ECN until %u now %u on %lx for drop-RST\n",
1017 __func__, tpheur->th_ecn_backoff, tcp_now,
1018 (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
1019 }
1020 }
1021
1022 if ((flags & TCPCACHE_F_ECN_DROPRXMT) &&
1023 tpheur->th_ecn_droprxmt < TCP_CACHE_OVERFLOW_PROTECT &&
1024 TSTMP_LEQ(tpheur->th_ecn_backoff, tcp_now)) {
1025 tpheur->th_ecn_droprxmt++;
1026 if (tpheur->th_ecn_droprxmt >= ECN_MAX_DROPRXMT) {
1027 tcpstat.tcps_ecn_fallback_droprxmt++;
1028 TCP_CACHE_INC_IFNET_STAT(tcks->ifp, tcks->af,
1029 ecn_fallback_droprxmt);
1030 tpheur->th_ecn_backoff = tcp_now +
1031 (tcp_min_to_hz(minutes: tcp_ecn_timeout) <<
1032 (tpheur->th_ecn_droprxmt - ECN_MAX_DROPRXMT));
1033
1034 os_log(OS_LOG_DEFAULT, "%s disable ECN until %u now %u on %lx for drop-Rxmit\n",
1035 __func__, tpheur->th_ecn_backoff, tcp_now,
1036 (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
1037 }
1038 }
1039 if ((flags & TCPCACHE_F_ECN_SYNRST) &&
1040 tpheur->th_ecn_synrst < TCP_CACHE_OVERFLOW_PROTECT) {
1041 tpheur->th_ecn_synrst++;
1042 if (tpheur->th_ecn_synrst >= ECN_MAX_SYNRST) {
1043 tcpstat.tcps_ecn_fallback_synrst++;
1044 TCP_CACHE_INC_IFNET_STAT(tcks->ifp, tcks->af,
1045 ecn_fallback_synrst);
1046 tpheur->th_ecn_backoff = tcp_now +
1047 (tcp_min_to_hz(minutes: tcp_ecn_timeout) <<
1048 (tpheur->th_ecn_synrst - ECN_MAX_SYNRST));
1049
1050 os_log(OS_LOG_DEFAULT, "%s disable ECN until %u now %u on %lx for SYN-RST\n",
1051 __func__, tpheur->th_ecn_backoff, tcp_now,
1052 (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
1053 }
1054 }
1055 tcp_heuristic_unlock(head);
1056}
1057
1058void
1059tcp_heuristic_tfo_loss(struct tcpcb *tp)
1060{
1061 struct tcp_cache_key_src tcks;
1062 uint32_t flag = 0;
1063
1064 if (symptoms_is_wifi_lossy() &&
1065 IFNET_IS_WIFI(tp->t_inpcb->inp_last_outifp)) {
1066 return;
1067 }
1068
1069 tcp_cache_key_src_create(tp, tcks: &tcks);
1070
1071 if (tp->t_tfo_stats & TFO_S_SYN_DATA_SENT) {
1072 flag = (TCPCACHE_F_TFO_DATA | TCPCACHE_F_TFO_REQ);
1073 }
1074 if (tp->t_tfo_stats & TFO_S_COOKIE_REQ) {
1075 flag = TCPCACHE_F_TFO_REQ;
1076 }
1077
1078 tcp_heuristic_inc_counters(tcks: &tcks, flags: flag);
1079}
1080
1081void
1082tcp_heuristic_tfo_rst(struct tcpcb *tp)
1083{
1084 struct tcp_cache_key_src tcks;
1085 uint32_t flag = 0;
1086
1087 tcp_cache_key_src_create(tp, tcks: &tcks);
1088
1089 if (tp->t_tfo_stats & TFO_S_SYN_DATA_SENT) {
1090 flag = (TCPCACHE_F_TFO_DATA_RST | TCPCACHE_F_TFO_REQ_RST);
1091 }
1092 if (tp->t_tfo_stats & TFO_S_COOKIE_REQ) {
1093 flag = TCPCACHE_F_TFO_REQ_RST;
1094 }
1095
1096 tcp_heuristic_inc_counters(tcks: &tcks, flags: flag);
1097}
1098
1099void
1100tcp_heuristic_mptcp_loss(struct tcpcb *tp)
1101{
1102 struct tcp_cache_key_src tcks;
1103
1104 if (symptoms_is_wifi_lossy() &&
1105 IFNET_IS_WIFI(tp->t_inpcb->inp_last_outifp)) {
1106 return;
1107 }
1108
1109 tcp_cache_key_src_create(tp, tcks: &tcks);
1110
1111 tcp_heuristic_inc_counters(tcks: &tcks, TCPCACHE_F_MPTCP);
1112}
1113
1114void
1115tcp_heuristic_ecn_loss(struct tcpcb *tp)
1116{
1117 struct tcp_cache_key_src tcks;
1118
1119 if (symptoms_is_wifi_lossy() &&
1120 IFNET_IS_WIFI(tp->t_inpcb->inp_last_outifp)) {
1121 return;
1122 }
1123
1124 tcp_cache_key_src_create(tp, tcks: &tcks);
1125
1126 tcp_heuristic_inc_counters(tcks: &tcks, TCPCACHE_F_ECN);
1127}
1128
1129void
1130tcp_heuristic_ecn_droprst(struct tcpcb *tp)
1131{
1132 struct tcp_cache_key_src tcks;
1133
1134 tcp_cache_key_src_create(tp, tcks: &tcks);
1135
1136 tcp_heuristic_inc_counters(tcks: &tcks, TCPCACHE_F_ECN_DROPRST);
1137}
1138
1139void
1140tcp_heuristic_ecn_droprxmt(struct tcpcb *tp)
1141{
1142 struct tcp_cache_key_src tcks;
1143
1144 tcp_cache_key_src_create(tp, tcks: &tcks);
1145
1146 tcp_heuristic_inc_counters(tcks: &tcks, TCPCACHE_F_ECN_DROPRXMT);
1147}
1148
1149void
1150tcp_heuristic_ecn_synrst(struct tcpcb *tp)
1151{
1152 struct tcp_cache_key_src tcks;
1153
1154 tcp_cache_key_src_create(tp, tcks: &tcks);
1155
1156 tcp_heuristic_inc_counters(tcks: &tcks, TCPCACHE_F_ECN_SYNRST);
1157}
1158
1159void
1160tcp_heuristic_tfo_middlebox(struct tcpcb *tp)
1161{
1162 struct tcp_cache_key_src tcks;
1163
1164 tp->t_tfo_flags |= TFO_F_HEURISTIC_DONE;
1165
1166 tcp_cache_key_src_create(tp, tcks: &tcks);
1167 tcp_heuristic_tfo_middlebox_common(tcks: &tcks);
1168}
1169
1170static void
1171tcp_heuristic_ecn_aggressive_common(struct tcp_cache_key_src *tcks)
1172{
1173 struct tcp_heuristics_head *head;
1174 struct tcp_heuristic *tpheur;
1175
1176 tpheur = tcp_getheuristic_with_lock(tcks, create: 1, headarg: &head);
1177 if (tpheur == NULL) {
1178 return;
1179 }
1180
1181 if (TSTMP_GT(tpheur->th_ecn_backoff, tcp_now)) {
1182 /* We are already in aggressive mode */
1183 tcp_heuristic_unlock(head);
1184 return;
1185 }
1186
1187 /* Must be done before, otherwise we will start off with expo-backoff */
1188 tpheur->th_ecn_backoff = tcp_now +
1189 (tcp_min_to_hz(minutes: tcp_ecn_timeout) << (tpheur->th_ecn_aggressive));
1190
1191 /*
1192 * Ugly way to prevent integer overflow... limit to prevent in
1193 * overflow during exp. backoff.
1194 */
1195 if (tpheur->th_ecn_aggressive < TCP_CACHE_OVERFLOW_PROTECT) {
1196 tpheur->th_ecn_aggressive++;
1197 }
1198
1199 tcp_heuristic_unlock(head);
1200
1201 os_log(OS_LOG_DEFAULT, "%s disable ECN until %u now %u on %lx\n", __func__,
1202 tpheur->th_ecn_backoff, tcp_now, (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
1203}
1204
1205void
1206tcp_heuristic_ecn_aggressive(struct tcpcb *tp)
1207{
1208 struct tcp_cache_key_src tcks;
1209
1210 tcp_cache_key_src_create(tp, tcks: &tcks);
1211 tcp_heuristic_ecn_aggressive_common(tcks: &tcks);
1212}
1213
1214static boolean_t
1215tcp_heuristic_do_tfo_common(struct tcp_cache_key_src *tcks)
1216{
1217 struct tcp_heuristics_head *head;
1218 struct tcp_heuristic *tpheur;
1219
1220 if (disable_tcp_heuristics) {
1221 return TRUE;
1222 }
1223
1224 /* Get the tcp-heuristic. */
1225 tpheur = tcp_getheuristic_with_lock(tcks, create: 0, headarg: &head);
1226 if (tpheur == NULL) {
1227 return TRUE;
1228 }
1229
1230 if (tpheur->th_tfo_in_backoff == 0) {
1231 goto tfo_ok;
1232 }
1233
1234 if (TSTMP_GT(tcp_now, tpheur->th_tfo_backoff_until)) {
1235 tpheur->th_tfo_in_backoff = 0;
1236 tpheur->th_tfo_enabled_time = tcp_now;
1237
1238 goto tfo_ok;
1239 }
1240
1241 tcp_heuristic_unlock(head);
1242 return FALSE;
1243
1244tfo_ok:
1245 tcp_heuristic_unlock(head);
1246 return TRUE;
1247}
1248
1249boolean_t
1250tcp_heuristic_do_tfo(struct tcpcb *tp)
1251{
1252 struct tcp_cache_key_src tcks;
1253
1254 tcp_cache_key_src_create(tp, tcks: &tcks);
1255 if (tcp_heuristic_do_tfo_common(tcks: &tcks)) {
1256 return TRUE;
1257 }
1258
1259 return FALSE;
1260}
1261/*
1262 * @return:
1263 * 0 Enable MPTCP (we are still discovering middleboxes)
1264 * -1 Enable MPTCP (heuristics have been temporarily disabled)
1265 * 1 Disable MPTCP
1266 */
1267int
1268tcp_heuristic_do_mptcp(struct tcpcb *tp)
1269{
1270 struct tcp_cache_key_src tcks;
1271 struct tcp_heuristics_head *head = NULL;
1272 struct tcp_heuristic *tpheur;
1273 int ret = 0;
1274
1275 if (disable_tcp_heuristics ||
1276 (tptomptp(tp)->mpt_mpte->mpte_flags & MPTE_FORCE_ENABLE)) {
1277 return 0;
1278 }
1279
1280 tcp_cache_key_src_create(tp, tcks: &tcks);
1281
1282 /* Get the tcp-heuristic. */
1283 tpheur = tcp_getheuristic_with_lock(tcks: &tcks, create: 0, headarg: &head);
1284 if (tpheur == NULL) {
1285 return 0;
1286 }
1287
1288 if (tpheur->th_mptcp_in_backoff == 0 ||
1289 tpheur->th_mptcp_heuristic_disabled == 1) {
1290 goto mptcp_ok;
1291 }
1292
1293 if (TSTMP_GT(tpheur->th_mptcp_backoff, tcp_now)) {
1294 goto fallback;
1295 }
1296
1297 tpheur->th_mptcp_in_backoff = 0;
1298
1299mptcp_ok:
1300 if (tpheur->th_mptcp_heuristic_disabled) {
1301 ret = -1;
1302
1303 if (TSTMP_GT(tcp_now, tpheur->th_mptcp_backoff)) {
1304 tpheur->th_mptcp_heuristic_disabled = 0;
1305 tpheur->th_mptcp_success = 0;
1306 }
1307 }
1308
1309 tcp_heuristic_unlock(head);
1310 return ret;
1311
1312fallback:
1313 if (head) {
1314 tcp_heuristic_unlock(head);
1315 }
1316
1317 if (tptomptp(tp)->mpt_mpte->mpte_flags & MPTE_FIRSTPARTY) {
1318 tcpstat.tcps_mptcp_fp_heuristic_fallback++;
1319 } else {
1320 tcpstat.tcps_mptcp_heuristic_fallback++;
1321 }
1322
1323 return 1;
1324}
1325
1326static boolean_t
1327tcp_heuristic_do_ecn_common(struct tcp_cache_key_src *tcks)
1328{
1329 struct tcp_heuristics_head *head;
1330 struct tcp_heuristic *tpheur;
1331 boolean_t ret = TRUE;
1332
1333 if (disable_tcp_heuristics) {
1334 return TRUE;
1335 }
1336
1337 /* Get the tcp-heuristic. */
1338 tpheur = tcp_getheuristic_with_lock(tcks, create: 0, headarg: &head);
1339 if (tpheur == NULL) {
1340 return ret;
1341 }
1342
1343 if (TSTMP_GT(tpheur->th_ecn_backoff, tcp_now)) {
1344 ret = FALSE;
1345 } else {
1346 /* Reset the following counters to start re-evaluating */
1347 if (tpheur->th_ecn_droprst >= ECN_RETRY_LIMIT) {
1348 tpheur->th_ecn_droprst = 0;
1349 }
1350 if (tpheur->th_ecn_droprxmt >= ECN_RETRY_LIMIT) {
1351 tpheur->th_ecn_droprxmt = 0;
1352 }
1353 if (tpheur->th_ecn_synrst >= ECN_RETRY_LIMIT) {
1354 tpheur->th_ecn_synrst = 0;
1355 }
1356
1357 /* Make sure it follows along */
1358 tpheur->th_ecn_backoff = tcp_now;
1359 }
1360
1361 tcp_heuristic_unlock(head);
1362
1363 return ret;
1364}
1365
1366boolean_t
1367tcp_heuristic_do_ecn(struct tcpcb *tp)
1368{
1369 struct tcp_cache_key_src tcks;
1370
1371 tcp_cache_key_src_create(tp, tcks: &tcks);
1372 return tcp_heuristic_do_ecn_common(tcks: &tcks);
1373}
1374
1375boolean_t
1376tcp_heuristic_do_ecn_with_address(struct ifnet *ifp,
1377 union sockaddr_in_4_6 *local_address)
1378{
1379 struct tcp_cache_key_src tcks;
1380
1381 memset(s: &tcks, c: 0, n: sizeof(tcks));
1382 tcks.ifp = ifp;
1383
1384 calculate_tcp_clock();
1385
1386 if (local_address->sa.sa_family == AF_INET6) {
1387 memcpy(dst: &tcks.laddr.addr6, src: &local_address->sin6.sin6_addr, n: sizeof(struct in6_addr));
1388 tcks.af = AF_INET6;
1389 } else if (local_address->sa.sa_family == AF_INET) {
1390 memcpy(dst: &tcks.laddr.addr, src: &local_address->sin.sin_addr, n: sizeof(struct in_addr));
1391 tcks.af = AF_INET;
1392 }
1393
1394 return tcp_heuristic_do_ecn_common(tcks: &tcks);
1395}
1396
1397void
1398tcp_heuristics_ecn_update(struct necp_tcp_ecn_cache *necp_buffer,
1399 struct ifnet *ifp, union sockaddr_in_4_6 *local_address)
1400{
1401 struct tcp_cache_key_src tcks;
1402
1403 memset(s: &tcks, c: 0, n: sizeof(tcks));
1404 tcks.ifp = ifp;
1405
1406 calculate_tcp_clock();
1407
1408 if (local_address->sa.sa_family == AF_INET6) {
1409 memcpy(dst: &tcks.laddr.addr6, src: &local_address->sin6.sin6_addr, n: sizeof(struct in6_addr));
1410 tcks.af = AF_INET6;
1411 } else if (local_address->sa.sa_family == AF_INET) {
1412 memcpy(dst: &tcks.laddr.addr, src: &local_address->sin.sin_addr, n: sizeof(struct in_addr));
1413 tcks.af = AF_INET;
1414 }
1415
1416 if (necp_buffer->necp_tcp_ecn_heuristics_success) {
1417 tcp_heuristic_reset_counters(tcks: &tcks, TCPCACHE_F_ECN);
1418 } else if (necp_buffer->necp_tcp_ecn_heuristics_loss) {
1419 tcp_heuristic_inc_counters(tcks: &tcks, TCPCACHE_F_ECN);
1420 } else if (necp_buffer->necp_tcp_ecn_heuristics_drop_rst) {
1421 tcp_heuristic_inc_counters(tcks: &tcks, TCPCACHE_F_ECN_DROPRST);
1422 } else if (necp_buffer->necp_tcp_ecn_heuristics_drop_rxmt) {
1423 tcp_heuristic_inc_counters(tcks: &tcks, TCPCACHE_F_ECN_DROPRXMT);
1424 } else if (necp_buffer->necp_tcp_ecn_heuristics_syn_rst) {
1425 tcp_heuristic_inc_counters(tcks: &tcks, TCPCACHE_F_ECN_SYNRST);
1426 } else if (necp_buffer->necp_tcp_ecn_heuristics_aggressive) {
1427 tcp_heuristic_ecn_aggressive_common(tcks: &tcks);
1428 }
1429
1430 return;
1431}
1432
1433boolean_t
1434tcp_heuristic_do_tfo_with_address(struct ifnet *ifp,
1435 union sockaddr_in_4_6 *local_address, union sockaddr_in_4_6 *remote_address,
1436 uint8_t *cookie, uint8_t *cookie_len)
1437{
1438 struct tcp_cache_key_src tcks;
1439
1440 memset(s: &tcks, c: 0, n: sizeof(tcks));
1441 tcks.ifp = ifp;
1442
1443 calculate_tcp_clock();
1444
1445 if (remote_address->sa.sa_family == AF_INET6) {
1446 memcpy(dst: &tcks.laddr.addr6, src: &local_address->sin6.sin6_addr, n: sizeof(struct in6_addr));
1447 memcpy(dst: &tcks.faddr.addr6, src: &remote_address->sin6.sin6_addr, n: sizeof(struct in6_addr));
1448 tcks.af = AF_INET6;
1449 } else if (remote_address->sa.sa_family == AF_INET) {
1450 memcpy(dst: &tcks.laddr.addr, src: &local_address->sin.sin_addr, n: sizeof(struct in_addr));
1451 memcpy(dst: &tcks.faddr.addr, src: &remote_address->sin.sin_addr, n: sizeof(struct in_addr));
1452 tcks.af = AF_INET;
1453 }
1454
1455 if (tcp_heuristic_do_tfo_common(tcks: &tcks)) {
1456 if (!tcp_cache_get_cookie_common(tcks: &tcks, cookie, len: cookie_len)) {
1457 *cookie_len = 0;
1458 }
1459 return TRUE;
1460 }
1461
1462 return FALSE;
1463}
1464
1465void
1466tcp_heuristics_tfo_update(struct necp_tcp_tfo_cache *necp_buffer,
1467 struct ifnet *ifp, union sockaddr_in_4_6 *local_address,
1468 union sockaddr_in_4_6 *remote_address)
1469{
1470 struct tcp_cache_key_src tcks;
1471
1472 memset(s: &tcks, c: 0, n: sizeof(tcks));
1473 tcks.ifp = ifp;
1474
1475 calculate_tcp_clock();
1476
1477 if (remote_address->sa.sa_family == AF_INET6) {
1478 memcpy(dst: &tcks.laddr.addr6, src: &local_address->sin6.sin6_addr, n: sizeof(struct in6_addr));
1479 memcpy(dst: &tcks.faddr.addr6, src: &remote_address->sin6.sin6_addr, n: sizeof(struct in6_addr));
1480 tcks.af = AF_INET6;
1481 } else if (remote_address->sa.sa_family == AF_INET) {
1482 memcpy(dst: &tcks.laddr.addr, src: &local_address->sin.sin_addr, n: sizeof(struct in_addr));
1483 memcpy(dst: &tcks.faddr.addr, src: &remote_address->sin.sin_addr, n: sizeof(struct in_addr));
1484 tcks.af = AF_INET;
1485 }
1486
1487 if (necp_buffer->necp_tcp_tfo_heuristics_success) {
1488 tcp_heuristic_reset_counters(tcks: &tcks, TCPCACHE_F_TFO_REQ | TCPCACHE_F_TFO_DATA |
1489 TCPCACHE_F_TFO_REQ_RST | TCPCACHE_F_TFO_DATA_RST);
1490 }
1491
1492 if (necp_buffer->necp_tcp_tfo_heuristics_success_req) {
1493 tcp_heuristic_reset_counters(tcks: &tcks, TCPCACHE_F_TFO_REQ | TCPCACHE_F_TFO_REQ_RST);
1494 }
1495
1496 if (necp_buffer->necp_tcp_tfo_heuristics_loss) {
1497 tcp_heuristic_inc_counters(tcks: &tcks, TCPCACHE_F_TFO_REQ | TCPCACHE_F_TFO_DATA);
1498 }
1499
1500 if (necp_buffer->necp_tcp_tfo_heuristics_loss_req) {
1501 tcp_heuristic_inc_counters(tcks: &tcks, TCPCACHE_F_TFO_REQ);
1502 }
1503
1504 if (necp_buffer->necp_tcp_tfo_heuristics_rst_data) {
1505 tcp_heuristic_inc_counters(tcks: &tcks, TCPCACHE_F_TFO_REQ_RST | TCPCACHE_F_TFO_DATA_RST);
1506 }
1507
1508 if (necp_buffer->necp_tcp_tfo_heuristics_rst_req) {
1509 tcp_heuristic_inc_counters(tcks: &tcks, TCPCACHE_F_TFO_REQ_RST);
1510 }
1511
1512 if (necp_buffer->necp_tcp_tfo_heuristics_middlebox) {
1513 tcp_heuristic_tfo_middlebox_common(tcks: &tcks);
1514 }
1515
1516 if (necp_buffer->necp_tcp_tfo_cookie_len != 0) {
1517 tcp_cache_set_cookie_common(tcks: &tcks,
1518 cookie: necp_buffer->necp_tcp_tfo_cookie, len: necp_buffer->necp_tcp_tfo_cookie_len);
1519 }
1520
1521 return;
1522}
1523
1524static void
1525sysctl_cleartfocache(void)
1526{
1527 int i;
1528
1529 for (i = 0; i < tcp_cache_size; i++) {
1530 struct tcp_cache_head *head = &tcp_cache[i];
1531 struct tcp_cache *tpcache, *tmp;
1532 struct tcp_heuristics_head *hhead = &tcp_heuristics[i];
1533 struct tcp_heuristic *tpheur, *htmp;
1534
1535 lck_mtx_lock(lck: &head->tch_mtx);
1536 SLIST_FOREACH_SAFE(tpcache, &head->tcp_caches, list, tmp) {
1537 SLIST_REMOVE(&head->tcp_caches, tpcache, tcp_cache, list);
1538 kfree_type(struct tcp_cache, tpcache);
1539 }
1540 lck_mtx_unlock(lck: &head->tch_mtx);
1541
1542 lck_mtx_lock(lck: &hhead->thh_mtx);
1543 SLIST_FOREACH_SAFE(tpheur, &hhead->tcp_heuristics, list, htmp) {
1544 SLIST_REMOVE(&hhead->tcp_heuristics, tpheur, tcp_heuristic, list);
1545 kfree_type(struct tcp_heuristic, tpheur);
1546 }
1547 lck_mtx_unlock(lck: &hhead->thh_mtx);
1548 }
1549}
1550
1551/* This sysctl is useful for testing purposes only */
1552static int tcpcleartfo = 0;
1553
1554static int sysctl_cleartfo SYSCTL_HANDLER_ARGS
1555{
1556#pragma unused(arg1, arg2)
1557 int error = 0, val, oldval = tcpcleartfo;
1558
1559 val = oldval;
1560 error = sysctl_handle_int(oidp, arg1: &val, arg2: 0, req);
1561 if (error || !req->newptr) {
1562 if (error) {
1563 os_log_error(OS_LOG_DEFAULT, "%s could not parse int: %d", __func__, error);
1564 }
1565 return error;
1566 }
1567
1568 /*
1569 * The actual value does not matter. If the value is set, it triggers
1570 * the clearing of the TFO cache. If a future implementation does not
1571 * use the route entry to hold the TFO cache, replace the route sysctl.
1572 */
1573
1574 if (val != oldval) {
1575 sysctl_cleartfocache();
1576 }
1577
1578 tcpcleartfo = val;
1579
1580 return error;
1581}
1582
1583SYSCTL_PROC(_net_inet_tcp, OID_AUTO, clear_tfocache, CTLTYPE_INT | CTLFLAG_RW |
1584 CTLFLAG_LOCKED, &tcpcleartfo, 0, &sysctl_cleartfo, "I",
1585 "Toggle to clear the TFO destination based heuristic cache");
1586
1587void
1588tcp_cache_init(void)
1589{
1590 uint64_t sane_size_meg = sane_size / 1024 / 1024;
1591
1592 /*
1593 * On machines with <100MB of memory this will result in a (full) cache-size
1594 * of 32 entries, thus 32 * 5 * 64bytes = 10KB. (about 0.01 %)
1595 * On machines with > 4GB of memory, we have a cache-size of 1024 entries,
1596 * thus about 327KB.
1597 *
1598 * Side-note: we convert to uint32_t. If sane_size is more than
1599 * 16000 TB, we loose precision. But, who cares? :)
1600 */
1601 tcp_cache_size = tcp_cache_roundup2(a: (uint32_t)(sane_size_meg >> 2));
1602 if (tcp_cache_size < 32) {
1603 tcp_cache_size = 32;
1604 } else if (tcp_cache_size > 1024) {
1605 tcp_cache_size = 1024;
1606 }
1607
1608 tcp_cache = zalloc_permanent(sizeof(struct tcp_cache_head) * tcp_cache_size,
1609 ZALIGN(struct tcp_cache_head));
1610
1611 tcp_heuristics = zalloc_permanent(sizeof(struct tcp_heuristics_head) * tcp_cache_size,
1612 ZALIGN(struct tcp_heuristics_head));
1613
1614 for (int i = 0; i < tcp_cache_size; i++) {
1615 lck_mtx_init(lck: &tcp_cache[i].tch_mtx, grp: &tcp_cache_mtx_grp,
1616 attr: &tcp_cache_mtx_attr);
1617 SLIST_INIT(&tcp_cache[i].tcp_caches);
1618
1619 lck_mtx_init(lck: &tcp_heuristics[i].thh_mtx, grp: &tcp_heuristic_mtx_grp,
1620 attr: &tcp_heuristic_mtx_attr);
1621 SLIST_INIT(&tcp_heuristics[i].tcp_heuristics);
1622 }
1623
1624 tcp_cache_hash_seed = RandomULong();
1625}
1626