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
44 typedef union {
45 struct in_addr addr;
46 struct in6_addr addr6;
47 } in_4_6_addr;
48
49 struct tcp_heuristic_key {
50 union {
51 uint8_t thk_net_signature[IFNET_SIGNATURELEN];
52 in_4_6_addr thk_ip;
53 };
54 sa_family_t thk_family;
55 };
56
57 struct tcp_heuristic {
58 SLIST_ENTRY(tcp_heuristic) list;
59
60 uint32_t th_last_access;
61
62 struct tcp_heuristic_key th_key;
63
64 char th_val_start[0]; /* Marker for memsetting to 0 */
65
66 uint8_t th_tfo_data_loss; /* The number of times a SYN+data has been lost */
67 uint8_t th_tfo_req_loss; /* The number of times a SYN+cookie-req has been lost */
68 uint8_t th_tfo_data_rst; /* The number of times a SYN+data has received a RST */
69 uint8_t th_tfo_req_rst; /* The number of times a SYN+cookie-req has received a RST */
70 uint8_t th_mptcp_loss; /* The number of times a SYN+MP_CAPABLE has been lost */
71 uint8_t th_mptcp_success; /* The number of times MPTCP-negotiation has been successful */
72 uint8_t th_ecn_loss; /* The number of times a SYN+ecn has been lost */
73 uint8_t th_ecn_aggressive; /* The number of times we did an aggressive fallback */
74 uint8_t th_ecn_droprst; /* The number of times ECN connections received a RST after first data pkt */
75 uint8_t th_ecn_droprxmt; /* The number of times ECN connection is dropped after multiple retransmits */
76 uint8_t th_ecn_synrst; /* number of times RST was received in response to an ECN enabled SYN */
77 uint32_t th_tfo_enabled_time; /* The moment when we reenabled TFO after backing off */
78 uint32_t th_tfo_backoff_until; /* Time until when we should not try out TFO */
79 uint32_t th_tfo_backoff; /* Current backoff timer */
80 uint32_t th_mptcp_backoff; /* Time until when we should not try out MPTCP */
81 uint32_t th_ecn_backoff; /* Time until when we should not try out ECN */
82
83 uint8_t th_tfo_in_backoff:1, /* Are we avoiding TFO due to the backoff timer? */
84 th_mptcp_in_backoff:1, /* Are we avoiding MPTCP due to the backoff timer? */
85 th_mptcp_heuristic_disabled:1; /* Are heuristics disabled? */
86
87 char th_val_end[0]; /* Marker for memsetting to 0 */
88 };
89
90 struct tcp_heuristics_head {
91 SLIST_HEAD(tcp_heur_bucket, tcp_heuristic) tcp_heuristics;
92
93 /* Per-hashbucket lock to avoid lock-contention */
94 lck_mtx_t thh_mtx;
95 };
96
97 struct tcp_cache_key {
98 sa_family_t tck_family;
99
100 struct tcp_heuristic_key tck_src;
101 in_4_6_addr tck_dst;
102 };
103
104 #define MPTCP_VERSION_SUPPORTED 1
105 #define MPTCP_VERSION_UNSUPPORTED -1
106 #define MPTCP_VERSION_SUPPORTED_UNKNOWN 0
107 struct tcp_cache {
108 SLIST_ENTRY(tcp_cache) list;
109
110 uint32_t tc_last_access;
111 uint32_t tc_mptcp_version_discovery_refresh; /* Time until when we should do version discovery again */
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 int8_t tc_mptcp_v0_support; /* -1 unsupported, 0 unknown, 1 supported */
119 int8_t tc_mptcp_v1_support; /* -1 unsupported, 0 unknown, 1 supported */
120 uint8_t tc_mptcp_v0_failed_count;
121 uint8_t tc_mptcp_v1_failed_count;
122 };
123
124 struct tcp_cache_head {
125 SLIST_HEAD(tcp_cache_bucket, tcp_cache) tcp_caches;
126
127 /* Per-hashbucket lock to avoid lock-contention */
128 lck_mtx_t tch_mtx;
129 };
130
131 struct tcp_cache_key_src {
132 struct ifnet *ifp;
133 in_4_6_addr laddr;
134 in_4_6_addr faddr;
135 int af;
136 };
137
138 static uint32_t tcp_cache_hash_seed;
139
140 size_t tcp_cache_size;
141
142 /*
143 * The maximum depth of the hash-bucket. This way we limit the tcp_cache to
144 * TCP_CACHE_BUCKET_SIZE * tcp_cache_size and have "natural" garbage collection
145 */
146 #define TCP_CACHE_BUCKET_SIZE 5
147
148 static struct tcp_cache_head *tcp_cache;
149
150 static LCK_ATTR_DECLARE(tcp_cache_mtx_attr, 0, 0);
151 static LCK_GRP_DECLARE(tcp_cache_mtx_grp, "tcpcache");
152
153 static struct tcp_heuristics_head *tcp_heuristics;
154
155 static LCK_ATTR_DECLARE(tcp_heuristic_mtx_attr, 0, 0);
156 static LCK_GRP_DECLARE(tcp_heuristic_mtx_grp, "tcpheuristic");
157
158 static uint32_t tcp_backoff_maximum = 65536;
159
160 SYSCTL_UINT(_net_inet_tcp, OID_AUTO, backoff_maximum, CTLFLAG_RW | CTLFLAG_LOCKED,
161 &tcp_backoff_maximum, 0, "Maximum time for which we won't try TFO");
162
163 static uint32_t tcp_ecn_timeout = 60;
164
165 SYSCTL_UINT(_net_inet_tcp, OID_AUTO, ecn_timeout, CTLFLAG_RW | CTLFLAG_LOCKED,
166 &tcp_ecn_timeout, 60, "Initial minutes to wait before re-trying ECN");
167
168 static int disable_tcp_heuristics = 0;
169 SYSCTL_INT(_net_inet_tcp, OID_AUTO, disable_tcp_heuristics, CTLFLAG_RW | CTLFLAG_LOCKED,
170 &disable_tcp_heuristics, 0, "Set to 1, to disable all TCP heuristics (TFO, ECN, MPTCP)");
171
172
173 static uint32_t
tcp_min_to_hz(uint32_t minutes)174 tcp_min_to_hz(uint32_t minutes)
175 {
176 if (minutes > 65536) {
177 return (uint32_t)65536 * 60 * TCP_RETRANSHZ;
178 }
179
180 return minutes * 60 * TCP_RETRANSHZ;
181 }
182
183 /*
184 * This number is coupled with tcp_ecn_timeout, because we want to prevent
185 * integer overflow. Need to find an unexpensive way to prevent integer overflow
186 * while still allowing a dynamic sysctl.
187 */
188 #define TCP_CACHE_OVERFLOW_PROTECT 9
189
190 /* Number of SYN-losses we accept */
191 #define TFO_MAX_COOKIE_LOSS 2
192 #define ECN_MAX_SYN_LOSS 2
193 #define MPTCP_MAX_SYN_LOSS 2
194 #define MPTCP_SUCCESS_TRIGGER 10
195 #define MPTCP_VERSION_MAX_FAIL 2
196 #define ECN_MAX_DROPRST 1
197 #define ECN_MAX_DROPRXMT 4
198 #define ECN_MAX_SYNRST 4
199
200 /* Flags for setting/unsetting loss-heuristics, limited to 4 bytes */
201 #define TCPCACHE_F_TFO_REQ 0x01
202 #define TCPCACHE_F_TFO_DATA 0x02
203 #define TCPCACHE_F_ECN 0x04
204 #define TCPCACHE_F_MPTCP 0x08
205 #define TCPCACHE_F_ECN_DROPRST 0x10
206 #define TCPCACHE_F_ECN_DROPRXMT 0x20
207 #define TCPCACHE_F_TFO_REQ_RST 0x40
208 #define TCPCACHE_F_TFO_DATA_RST 0x80
209 #define TCPCACHE_F_ECN_SYNRST 0x100
210
211 /* Always retry ECN after backing off to this level for some heuristics */
212 #define ECN_RETRY_LIMIT 9
213
214 #define TCP_CACHE_INC_IFNET_STAT(_ifp_, _af_, _stat_) { \
215 if ((_ifp_) != NULL) { \
216 if ((_af_) == AF_INET6) { \
217 (_ifp_)->if_ipv6_stat->_stat_++;\
218 } else { \
219 (_ifp_)->if_ipv4_stat->_stat_++;\
220 }\
221 }\
222 }
223
224 /*
225 * Round up to next higher power-of 2. See "Bit Twiddling Hacks".
226 *
227 * Might be worth moving this to a library so that others
228 * (e.g., scale_to_powerof2()) can use this as well instead of a while-loop.
229 */
230 static uint32_t
tcp_cache_roundup2(uint32_t a)231 tcp_cache_roundup2(uint32_t a)
232 {
233 a--;
234 a |= a >> 1;
235 a |= a >> 2;
236 a |= a >> 4;
237 a |= a >> 8;
238 a |= a >> 16;
239 a++;
240
241 return a;
242 }
243
244 static void
tcp_cache_hash_src(struct tcp_cache_key_src * tcks,struct tcp_heuristic_key * key)245 tcp_cache_hash_src(struct tcp_cache_key_src *tcks, struct tcp_heuristic_key *key)
246 {
247 struct ifnet *ifp = tcks->ifp;
248 uint8_t len = sizeof(key->thk_net_signature);
249 uint16_t flags;
250
251 if (tcks->af == AF_INET6) {
252 int ret;
253
254 key->thk_family = AF_INET6;
255 ret = ifnet_get_netsignature(ifp, AF_INET6, &len, &flags,
256 key->thk_net_signature);
257
258 /*
259 * ifnet_get_netsignature only returns EINVAL if ifn is NULL
260 * (we made sure that in the other cases it does not). So,
261 * in this case we should take the connection's address.
262 */
263 if (ret == ENOENT || ret == EINVAL) {
264 memcpy(&key->thk_ip.addr6, &tcks->laddr.addr6, sizeof(struct in6_addr));
265 }
266 } else {
267 int ret;
268
269 key->thk_family = AF_INET;
270 ret = ifnet_get_netsignature(ifp, AF_INET, &len, &flags,
271 key->thk_net_signature);
272
273 /*
274 * ifnet_get_netsignature only returns EINVAL if ifn is NULL
275 * (we made sure that in the other cases it does not). So,
276 * in this case we should take the connection's address.
277 */
278 if (ret == ENOENT || ret == EINVAL) {
279 memcpy(&key->thk_ip.addr, &tcks->laddr.addr, sizeof(struct in_addr));
280 }
281 }
282 }
283
284 static uint16_t
tcp_cache_hash(struct tcp_cache_key_src * tcks,struct tcp_cache_key * key)285 tcp_cache_hash(struct tcp_cache_key_src *tcks, struct tcp_cache_key *key)
286 {
287 uint32_t hash;
288
289 bzero(key, sizeof(struct tcp_cache_key));
290
291 tcp_cache_hash_src(tcks, &key->tck_src);
292
293 if (tcks->af == AF_INET6) {
294 key->tck_family = AF_INET6;
295 memcpy(&key->tck_dst.addr6, &tcks->faddr.addr6,
296 sizeof(struct in6_addr));
297 } else {
298 key->tck_family = AF_INET;
299 memcpy(&key->tck_dst.addr, &tcks->faddr.addr,
300 sizeof(struct in_addr));
301 }
302
303 hash = net_flowhash(key, sizeof(struct tcp_cache_key),
304 tcp_cache_hash_seed);
305
306 return (uint16_t)(hash & (tcp_cache_size - 1));
307 }
308
309 static void
tcp_cache_unlock(struct tcp_cache_head * head)310 tcp_cache_unlock(struct tcp_cache_head *head)
311 {
312 lck_mtx_unlock(&head->tch_mtx);
313 }
314
315 /*
316 * Make sure that everything that happens after tcp_getcache_with_lock()
317 * is short enough to justify that you hold the per-bucket lock!!!
318 *
319 * Otherwise, better build another lookup-function that does not hold the
320 * lock and you copy out the bits and bytes.
321 *
322 * That's why we provide the head as a "return"-pointer so that the caller
323 * can give it back to use for tcp_cache_unlock().
324 */
325 static struct tcp_cache *
tcp_getcache_with_lock(struct tcp_cache_key_src * tcks,int create,struct tcp_cache_head ** headarg)326 tcp_getcache_with_lock(struct tcp_cache_key_src *tcks,
327 int create, struct tcp_cache_head **headarg)
328 {
329 struct tcp_cache *tpcache = NULL;
330 struct tcp_cache_head *head;
331 struct tcp_cache_key key;
332 uint16_t hash;
333 int i = 0;
334
335 hash = tcp_cache_hash(tcks, &key);
336 head = &tcp_cache[hash];
337
338 lck_mtx_lock(&head->tch_mtx);
339
340 /*** First step: Look for the tcp_cache in our bucket ***/
341 SLIST_FOREACH(tpcache, &head->tcp_caches, list) {
342 if (memcmp(&tpcache->tc_key, &key, sizeof(key)) == 0) {
343 break;
344 }
345
346 i++;
347 }
348
349 /*** Second step: If it's not there, create/recycle it ***/
350 if ((tpcache == NULL) && create) {
351 if (i >= TCP_CACHE_BUCKET_SIZE) {
352 struct tcp_cache *oldest_cache = NULL;
353 uint32_t max_age = 0;
354
355 /* Look for the oldest tcp_cache in the bucket */
356 SLIST_FOREACH(tpcache, &head->tcp_caches, list) {
357 uint32_t age = tcp_now - tpcache->tc_last_access;
358 if (age > max_age) {
359 max_age = age;
360 oldest_cache = tpcache;
361 }
362 }
363 VERIFY(oldest_cache != NULL);
364
365 tpcache = oldest_cache;
366
367 /* We recycle, thus let's indicate that there is no cookie */
368 tpcache->tc_tfo_cookie_len = 0;
369 } else {
370 /* Create a new cache and add it to the list */
371 tpcache = kalloc_type(struct tcp_cache, Z_NOWAIT | Z_ZERO);
372 if (tpcache == NULL) {
373 os_log_error(OS_LOG_DEFAULT, "%s could not allocate cache", __func__);
374 goto out_null;
375 }
376
377 SLIST_INSERT_HEAD(&head->tcp_caches, tpcache, list);
378 }
379
380 memcpy(&tpcache->tc_key, &key, sizeof(key));
381 }
382
383 if (tpcache == NULL) {
384 goto out_null;
385 }
386
387 /* Update timestamp for garbage collection purposes */
388 tpcache->tc_last_access = tcp_now;
389 *headarg = head;
390
391 return tpcache;
392
393 out_null:
394 tcp_cache_unlock(head);
395 return NULL;
396 }
397
398 static void
tcp_cache_key_src_create(struct tcpcb * tp,struct tcp_cache_key_src * tcks)399 tcp_cache_key_src_create(struct tcpcb *tp, struct tcp_cache_key_src *tcks)
400 {
401 struct inpcb *inp = tp->t_inpcb;
402 memset(tcks, 0, sizeof(*tcks));
403
404 tcks->ifp = inp->inp_last_outifp;
405
406 if (inp->inp_vflag & INP_IPV6) {
407 memcpy(&tcks->laddr.addr6, &inp->in6p_laddr, sizeof(struct in6_addr));
408 memcpy(&tcks->faddr.addr6, &inp->in6p_faddr, sizeof(struct in6_addr));
409 tcks->af = AF_INET6;
410 } else {
411 memcpy(&tcks->laddr.addr, &inp->inp_laddr, sizeof(struct in_addr));
412 memcpy(&tcks->faddr.addr, &inp->inp_faddr, sizeof(struct in_addr));
413 tcks->af = AF_INET;
414 }
415
416 return;
417 }
418
419 static void
mptcp_version_cache_key_src_init(struct sockaddr * dst,struct tcp_cache_key_src * tcks)420 mptcp_version_cache_key_src_init(struct sockaddr *dst, struct tcp_cache_key_src *tcks)
421 {
422 memset(tcks, 0, sizeof(*tcks));
423
424 if (dst->sa_family == AF_INET) {
425 memcpy(&tcks->faddr.addr, &SIN(dst)->sin_addr, sizeof(struct in_addr));
426 tcks->af = AF_INET;
427 } else {
428 memcpy(&tcks->faddr.addr6, &SIN6(dst)->sin6_addr, sizeof(struct in6_addr));
429 tcks->af = AF_INET6;
430 }
431
432 return;
433 }
434
435 static void
tcp_cache_set_cookie_common(struct tcp_cache_key_src * tcks,u_char * cookie,uint8_t len)436 tcp_cache_set_cookie_common(struct tcp_cache_key_src *tcks, u_char *cookie, uint8_t len)
437 {
438 struct tcp_cache_head *head;
439 struct tcp_cache *tpcache;
440
441 /* Call lookup/create function */
442 tpcache = tcp_getcache_with_lock(tcks, 1, &head);
443 if (tpcache == NULL) {
444 return;
445 }
446
447 tpcache->tc_tfo_cookie_len = len > TFO_COOKIE_LEN_MAX ?
448 TFO_COOKIE_LEN_MAX : len;
449 memcpy(tpcache->tc_tfo_cookie, cookie, tpcache->tc_tfo_cookie_len);
450
451 tcp_cache_unlock(head);
452 }
453
454 void
tcp_cache_set_cookie(struct tcpcb * tp,u_char * cookie,uint8_t len)455 tcp_cache_set_cookie(struct tcpcb *tp, u_char *cookie, uint8_t len)
456 {
457 struct tcp_cache_key_src tcks;
458
459 tcp_cache_key_src_create(tp, &tcks);
460 tcp_cache_set_cookie_common(&tcks, cookie, len);
461 }
462
463 static int
tcp_cache_get_cookie_common(struct tcp_cache_key_src * tcks,u_char * cookie,uint8_t * len)464 tcp_cache_get_cookie_common(struct tcp_cache_key_src *tcks, u_char *cookie, uint8_t *len)
465 {
466 struct tcp_cache_head *head;
467 struct tcp_cache *tpcache;
468
469 /* Call lookup/create function */
470 tpcache = tcp_getcache_with_lock(tcks, 1, &head);
471 if (tpcache == NULL) {
472 return 0;
473 }
474
475 if (tpcache->tc_tfo_cookie_len == 0) {
476 tcp_cache_unlock(head);
477 return 0;
478 }
479
480 /*
481 * Not enough space - this should never happen as it has been checked
482 * in tcp_tfo_check. So, fail here!
483 */
484 VERIFY(tpcache->tc_tfo_cookie_len <= *len);
485
486 memcpy(cookie, tpcache->tc_tfo_cookie, tpcache->tc_tfo_cookie_len);
487 *len = tpcache->tc_tfo_cookie_len;
488
489 tcp_cache_unlock(head);
490
491 return 1;
492 }
493
494 /*
495 * Get the cookie related to 'tp', and copy it into 'cookie', provided that len
496 * is big enough (len designates the available memory.
497 * Upon return, 'len' is set to the cookie's length.
498 *
499 * Returns 0 if we should request a cookie.
500 * Returns 1 if the cookie has been found and written.
501 */
502 int
tcp_cache_get_cookie(struct tcpcb * tp,u_char * cookie,uint8_t * len)503 tcp_cache_get_cookie(struct tcpcb *tp, u_char *cookie, uint8_t *len)
504 {
505 struct tcp_cache_key_src tcks;
506
507 tcp_cache_key_src_create(tp, &tcks);
508 return tcp_cache_get_cookie_common(&tcks, cookie, len);
509 }
510
511 static unsigned int
tcp_cache_get_cookie_len_common(struct tcp_cache_key_src * tcks)512 tcp_cache_get_cookie_len_common(struct tcp_cache_key_src *tcks)
513 {
514 struct tcp_cache_head *head;
515 struct tcp_cache *tpcache;
516 unsigned int cookie_len;
517
518 /* Call lookup/create function */
519 tpcache = tcp_getcache_with_lock(tcks, 1, &head);
520 if (tpcache == NULL) {
521 return 0;
522 }
523
524 cookie_len = tpcache->tc_tfo_cookie_len;
525
526 tcp_cache_unlock(head);
527
528 return cookie_len;
529 }
530
531 unsigned int
tcp_cache_get_cookie_len(struct tcpcb * tp)532 tcp_cache_get_cookie_len(struct tcpcb *tp)
533 {
534 struct tcp_cache_key_src tcks;
535
536 tcp_cache_key_src_create(tp, &tcks);
537 return tcp_cache_get_cookie_len_common(&tcks);
538 }
539
540 /*
541 * @return:
542 * 0 MPTCP_VERSION_0
543 * 1 MPTCP_VERSION_1
544 */
545 uint8_t
tcp_cache_get_mptcp_version(struct sockaddr * dst)546 tcp_cache_get_mptcp_version(struct sockaddr *dst)
547 {
548 struct tcp_cache_key_src tcks;
549 mptcp_version_cache_key_src_init(dst, &tcks);
550 uint8_t version = (uint8_t) mptcp_preferred_version;
551
552 struct tcp_cache_head *head;
553 struct tcp_cache *tpcache;
554
555 /* Call lookup/create function */
556 tpcache = tcp_getcache_with_lock(&tcks, 1, &head);
557 if (tpcache == NULL) {
558 return version;
559 }
560
561 if (tpcache->tc_mptcp_v1_support == MPTCP_VERSION_UNSUPPORTED &&
562 tpcache->tc_mptcp_v0_support != MPTCP_VERSION_UNSUPPORTED) {
563 version = MPTCP_VERSION_0;
564 }
565 if (tpcache->tc_mptcp_v0_support == MPTCP_VERSION_UNSUPPORTED &&
566 tpcache->tc_mptcp_v1_support != MPTCP_VERSION_UNSUPPORTED) {
567 version = MPTCP_VERSION_1;
568 }
569
570 tcp_cache_unlock(head);
571 return version;
572 }
573
574 void
tcp_cache_update_mptcp_version(struct tcpcb * tp,boolean_t succeeded)575 tcp_cache_update_mptcp_version(struct tcpcb *tp, boolean_t succeeded)
576 {
577 struct mptcb *mp_tp = tptomptp(tp);
578 uint8_t version = mp_tp->mpt_version;
579 int8_t record = succeeded ? MPTCP_VERSION_SUPPORTED : MPTCP_VERSION_UNSUPPORTED;
580
581 struct tcp_cache_key_src tcks;
582
583 struct inpcb *inp = tp->t_inpcb;
584 if (inp->inp_vflag & INP_IPV6) {
585 struct sockaddr_in6 dst = {
586 .sin6_len = sizeof(struct sockaddr_in6),
587 .sin6_family = AF_INET6,
588 .sin6_addr = inp->in6p_faddr,
589 };
590 mptcp_version_cache_key_src_init((struct sockaddr *)&dst, &tcks);
591 } else {
592 struct sockaddr_in dst = {
593 .sin_len = sizeof(struct sockaddr_in),
594 .sin_family = AF_INET,
595 .sin_addr = inp->inp_faddr,
596 };
597 mptcp_version_cache_key_src_init((struct sockaddr *)&dst, &tcks);
598 }
599
600 struct tcp_cache_head *head;
601 struct tcp_cache *tpcache;
602 /* Call lookup/create function */
603 tpcache = tcp_getcache_with_lock(&tcks, 1, &head);
604 if (tpcache == NULL) {
605 return;
606 }
607
608 if (version == MPTCP_VERSION_0) {
609 if (tpcache->tc_mptcp_v0_support == MPTCP_VERSION_SUPPORTED_UNKNOWN) {
610 tpcache->tc_mptcp_v0_support = record;
611 } else {
612 if (succeeded) {
613 tpcache->tc_mptcp_v0_failed_count = 0;
614 tpcache->tc_mptcp_v0_support = MPTCP_VERSION_SUPPORTED;
615 } else {
616 tpcache->tc_mptcp_v0_failed_count += 1;
617 // flip the record, so the other version is attempted the next time
618 if (tpcache->tc_mptcp_v0_failed_count >= MPTCP_VERSION_MAX_FAIL) {
619 tpcache->tc_mptcp_v0_failed_count = 0;
620 tpcache->tc_mptcp_v1_failed_count = 0;
621 tpcache->tc_mptcp_v0_support = MPTCP_VERSION_UNSUPPORTED;
622 tpcache->tc_mptcp_v1_support = MPTCP_VERSION_SUPPORTED_UNKNOWN;
623 }
624 }
625 }
626 }
627
628 if (version == MPTCP_VERSION_1) {
629 if (tpcache->tc_mptcp_v1_support == MPTCP_VERSION_SUPPORTED_UNKNOWN) {
630 tpcache->tc_mptcp_v1_support = record;
631 } else {
632 if (succeeded) {
633 tpcache->tc_mptcp_v1_failed_count = 0;
634 tpcache->tc_mptcp_v1_support = MPTCP_VERSION_SUPPORTED;
635 } else {
636 tpcache->tc_mptcp_v1_failed_count += 1;
637 // flip the record, so the other version is attempted the next time
638 if (tpcache->tc_mptcp_v1_failed_count >= MPTCP_VERSION_MAX_FAIL) {
639 tpcache->tc_mptcp_v0_failed_count = 0;
640 tpcache->tc_mptcp_v1_failed_count = 0;
641 tpcache->tc_mptcp_v1_support = MPTCP_VERSION_UNSUPPORTED;
642 tpcache->tc_mptcp_v0_support = MPTCP_VERSION_SUPPORTED_UNKNOWN;
643 }
644 }
645 }
646 }
647
648 tcp_cache_unlock(head);
649 }
650
651 static uint16_t
tcp_heuristics_hash(struct tcp_cache_key_src * tcks,struct tcp_heuristic_key * key)652 tcp_heuristics_hash(struct tcp_cache_key_src *tcks, struct tcp_heuristic_key *key)
653 {
654 uint32_t hash;
655
656 bzero(key, sizeof(struct tcp_heuristic_key));
657
658 tcp_cache_hash_src(tcks, key);
659
660 hash = net_flowhash(key, sizeof(struct tcp_heuristic_key),
661 tcp_cache_hash_seed);
662
663 return (uint16_t)(hash & (tcp_cache_size - 1));
664 }
665
666 static void
tcp_heuristic_unlock(struct tcp_heuristics_head * head)667 tcp_heuristic_unlock(struct tcp_heuristics_head *head)
668 {
669 lck_mtx_unlock(&head->thh_mtx);
670 }
671
672 /*
673 * Make sure that everything that happens after tcp_getheuristic_with_lock()
674 * is short enough to justify that you hold the per-bucket lock!!!
675 *
676 * Otherwise, better build another lookup-function that does not hold the
677 * lock and you copy out the bits and bytes.
678 *
679 * That's why we provide the head as a "return"-pointer so that the caller
680 * can give it back to use for tcp_heur_unlock().
681 *
682 *
683 * ToDo - way too much code-duplication. We should create an interface to handle
684 * bucketized hashtables with recycling of the oldest element.
685 */
686 static struct tcp_heuristic *
tcp_getheuristic_with_lock(struct tcp_cache_key_src * tcks,int create,struct tcp_heuristics_head ** headarg)687 tcp_getheuristic_with_lock(struct tcp_cache_key_src *tcks,
688 int create, struct tcp_heuristics_head **headarg)
689 {
690 struct tcp_heuristic *tpheur = NULL;
691 struct tcp_heuristics_head *head;
692 struct tcp_heuristic_key key;
693 uint16_t hash;
694 int i = 0;
695
696 hash = tcp_heuristics_hash(tcks, &key);
697 head = &tcp_heuristics[hash];
698
699 lck_mtx_lock(&head->thh_mtx);
700
701 /*** First step: Look for the tcp_heur in our bucket ***/
702 SLIST_FOREACH(tpheur, &head->tcp_heuristics, list) {
703 if (memcmp(&tpheur->th_key, &key, sizeof(key)) == 0) {
704 break;
705 }
706
707 i++;
708 }
709
710 /*** Second step: If it's not there, create/recycle it ***/
711 if ((tpheur == NULL) && create) {
712 if (i >= TCP_CACHE_BUCKET_SIZE) {
713 struct tcp_heuristic *oldest_heur = NULL;
714 uint32_t max_age = 0;
715
716 /* Look for the oldest tcp_heur in the bucket */
717 SLIST_FOREACH(tpheur, &head->tcp_heuristics, list) {
718 uint32_t age = tcp_now - tpheur->th_last_access;
719 if (age > max_age) {
720 max_age = age;
721 oldest_heur = tpheur;
722 }
723 }
724 VERIFY(oldest_heur != NULL);
725
726 tpheur = oldest_heur;
727
728 /* We recycle - set everything to 0 */
729 bzero(tpheur->th_val_start,
730 tpheur->th_val_end - tpheur->th_val_start);
731 } else {
732 /* Create a new heuristic and add it to the list */
733 tpheur = kalloc_type(struct tcp_heuristic, Z_NOWAIT | Z_ZERO);
734 if (tpheur == NULL) {
735 os_log_error(OS_LOG_DEFAULT, "%s could not allocate cache", __func__);
736 goto out_null;
737 }
738
739 SLIST_INSERT_HEAD(&head->tcp_heuristics, tpheur, list);
740 }
741
742 /*
743 * Set to tcp_now, to make sure it won't be > than tcp_now in the
744 * near future.
745 */
746 tpheur->th_ecn_backoff = tcp_now;
747 tpheur->th_tfo_backoff_until = tcp_now;
748 tpheur->th_mptcp_backoff = tcp_now;
749 tpheur->th_tfo_backoff = tcp_min_to_hz(tcp_ecn_timeout);
750
751 memcpy(&tpheur->th_key, &key, sizeof(key));
752 }
753
754 if (tpheur == NULL) {
755 goto out_null;
756 }
757
758 /* Update timestamp for garbage collection purposes */
759 tpheur->th_last_access = tcp_now;
760 *headarg = head;
761
762 return tpheur;
763
764 out_null:
765 tcp_heuristic_unlock(head);
766 return NULL;
767 }
768
769 static void
tcp_heuristic_reset_counters(struct tcp_cache_key_src * tcks,uint8_t flags)770 tcp_heuristic_reset_counters(struct tcp_cache_key_src *tcks, uint8_t flags)
771 {
772 struct tcp_heuristics_head *head;
773 struct tcp_heuristic *tpheur;
774
775 /*
776 * Always create heuristics here because MPTCP needs to write success
777 * into it. Thus, we always end up creating them.
778 */
779 tpheur = tcp_getheuristic_with_lock(tcks, 1, &head);
780 if (tpheur == NULL) {
781 return;
782 }
783
784 if (flags & TCPCACHE_F_TFO_DATA) {
785 if (tpheur->th_tfo_data_loss >= TFO_MAX_COOKIE_LOSS) {
786 os_log(OS_LOG_DEFAULT, "%s: Resetting TFO-data loss to 0 from %u on heur %lx\n",
787 __func__, tpheur->th_tfo_data_loss, (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
788 }
789 tpheur->th_tfo_data_loss = 0;
790 }
791
792 if (flags & TCPCACHE_F_TFO_REQ) {
793 if (tpheur->th_tfo_req_loss >= TFO_MAX_COOKIE_LOSS) {
794 os_log(OS_LOG_DEFAULT, "%s: Resetting TFO-req loss to 0 from %u on heur %lx\n",
795 __func__, tpheur->th_tfo_req_loss, (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
796 }
797 tpheur->th_tfo_req_loss = 0;
798 }
799
800 if (flags & TCPCACHE_F_TFO_DATA_RST) {
801 if (tpheur->th_tfo_data_rst >= TFO_MAX_COOKIE_LOSS) {
802 os_log(OS_LOG_DEFAULT, "%s: Resetting TFO-data RST to 0 from %u on heur %lx\n",
803 __func__, tpheur->th_tfo_data_rst, (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
804 }
805 tpheur->th_tfo_data_rst = 0;
806 }
807
808 if (flags & TCPCACHE_F_TFO_REQ_RST) {
809 if (tpheur->th_tfo_req_rst >= TFO_MAX_COOKIE_LOSS) {
810 os_log(OS_LOG_DEFAULT, "%s: Resetting TFO-req RST to 0 from %u on heur %lx\n",
811 __func__, tpheur->th_tfo_req_rst, (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
812 }
813 tpheur->th_tfo_req_rst = 0;
814 }
815
816 if (flags & TCPCACHE_F_ECN) {
817 if (tpheur->th_ecn_loss >= ECN_MAX_SYN_LOSS || tpheur->th_ecn_synrst >= ECN_MAX_SYNRST) {
818 os_log(OS_LOG_DEFAULT, "%s: Resetting ECN-loss to 0 from %u and synrst from %u on heur %lx\n",
819 __func__, tpheur->th_ecn_loss, tpheur->th_ecn_synrst, (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
820 }
821 tpheur->th_ecn_loss = 0;
822 tpheur->th_ecn_synrst = 0;
823 }
824
825 if (flags & TCPCACHE_F_MPTCP) {
826 tpheur->th_mptcp_loss = 0;
827 if (tpheur->th_mptcp_success < MPTCP_SUCCESS_TRIGGER) {
828 tpheur->th_mptcp_success++;
829
830 if (tpheur->th_mptcp_success == MPTCP_SUCCESS_TRIGGER) {
831 os_log(mptcp_log_handle, "%s disabling heuristics for 12 hours", __func__);
832 tpheur->th_mptcp_heuristic_disabled = 1;
833 /* Disable heuristics for 12 hours */
834 tpheur->th_mptcp_backoff = tcp_now + tcp_min_to_hz(tcp_ecn_timeout * 12);
835 }
836 }
837 }
838
839 tcp_heuristic_unlock(head);
840 }
841
842 void
tcp_heuristic_tfo_success(struct tcpcb * tp)843 tcp_heuristic_tfo_success(struct tcpcb *tp)
844 {
845 struct tcp_cache_key_src tcks;
846 uint8_t flag = 0;
847
848 tcp_cache_key_src_create(tp, &tcks);
849
850 if (tp->t_tfo_stats & TFO_S_SYN_DATA_SENT) {
851 flag = (TCPCACHE_F_TFO_DATA | TCPCACHE_F_TFO_REQ |
852 TCPCACHE_F_TFO_DATA_RST | TCPCACHE_F_TFO_REQ_RST);
853 }
854 if (tp->t_tfo_stats & TFO_S_COOKIE_REQ) {
855 flag = (TCPCACHE_F_TFO_REQ | TCPCACHE_F_TFO_REQ_RST);
856 }
857
858 tcp_heuristic_reset_counters(&tcks, flag);
859 }
860
861 void
tcp_heuristic_mptcp_success(struct tcpcb * tp)862 tcp_heuristic_mptcp_success(struct tcpcb *tp)
863 {
864 struct tcp_cache_key_src tcks;
865
866 tcp_cache_key_src_create(tp, &tcks);
867 tcp_heuristic_reset_counters(&tcks, TCPCACHE_F_MPTCP);
868 }
869
870 void
tcp_heuristic_ecn_success(struct tcpcb * tp)871 tcp_heuristic_ecn_success(struct tcpcb *tp)
872 {
873 struct tcp_cache_key_src tcks;
874
875 tcp_cache_key_src_create(tp, &tcks);
876 tcp_heuristic_reset_counters(&tcks, TCPCACHE_F_ECN);
877 }
878
879 static void
__tcp_heuristic_tfo_middlebox_common(struct tcp_heuristic * tpheur)880 __tcp_heuristic_tfo_middlebox_common(struct tcp_heuristic *tpheur)
881 {
882 if (tpheur->th_tfo_in_backoff) {
883 return;
884 }
885
886 tpheur->th_tfo_in_backoff = 1;
887
888 if (tpheur->th_tfo_enabled_time) {
889 uint32_t old_backoff = tpheur->th_tfo_backoff;
890
891 tpheur->th_tfo_backoff -= (tcp_now - tpheur->th_tfo_enabled_time);
892 if (tpheur->th_tfo_backoff > old_backoff) {
893 tpheur->th_tfo_backoff = tcp_min_to_hz(tcp_ecn_timeout);
894 }
895 }
896
897 tpheur->th_tfo_backoff_until = tcp_now + tpheur->th_tfo_backoff;
898
899 /* Then, increase the backoff time */
900 tpheur->th_tfo_backoff *= 2;
901
902 if (tpheur->th_tfo_backoff > tcp_min_to_hz(tcp_backoff_maximum)) {
903 tpheur->th_tfo_backoff = tcp_min_to_hz(tcp_ecn_timeout);
904 }
905
906 os_log(OS_LOG_DEFAULT, "%s disable TFO until %u now %u on %lx\n", __func__,
907 tpheur->th_tfo_backoff_until, tcp_now, (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
908 }
909
910 static void
tcp_heuristic_tfo_middlebox_common(struct tcp_cache_key_src * tcks)911 tcp_heuristic_tfo_middlebox_common(struct tcp_cache_key_src *tcks)
912 {
913 struct tcp_heuristics_head *head;
914 struct tcp_heuristic *tpheur;
915
916 tpheur = tcp_getheuristic_with_lock(tcks, 1, &head);
917 if (tpheur == NULL) {
918 return;
919 }
920
921 __tcp_heuristic_tfo_middlebox_common(tpheur);
922
923 tcp_heuristic_unlock(head);
924 }
925
926 static void
tcp_heuristic_inc_counters(struct tcp_cache_key_src * tcks,uint32_t flags)927 tcp_heuristic_inc_counters(struct tcp_cache_key_src *tcks,
928 uint32_t flags)
929 {
930 struct tcp_heuristics_head *head;
931 struct tcp_heuristic *tpheur;
932
933 tpheur = tcp_getheuristic_with_lock(tcks, 1, &head);
934 if (tpheur == NULL) {
935 return;
936 }
937
938 /* Limit to prevent integer-overflow during exponential backoff */
939 if ((flags & TCPCACHE_F_TFO_DATA) && tpheur->th_tfo_data_loss < TCP_CACHE_OVERFLOW_PROTECT) {
940 tpheur->th_tfo_data_loss++;
941
942 if (tpheur->th_tfo_data_loss >= TFO_MAX_COOKIE_LOSS) {
943 __tcp_heuristic_tfo_middlebox_common(tpheur);
944 }
945 }
946
947 if ((flags & TCPCACHE_F_TFO_REQ) && tpheur->th_tfo_req_loss < TCP_CACHE_OVERFLOW_PROTECT) {
948 tpheur->th_tfo_req_loss++;
949
950 if (tpheur->th_tfo_req_loss >= TFO_MAX_COOKIE_LOSS) {
951 __tcp_heuristic_tfo_middlebox_common(tpheur);
952 }
953 }
954
955 if ((flags & TCPCACHE_F_TFO_DATA_RST) && tpheur->th_tfo_data_rst < TCP_CACHE_OVERFLOW_PROTECT) {
956 tpheur->th_tfo_data_rst++;
957
958 if (tpheur->th_tfo_data_rst >= TFO_MAX_COOKIE_LOSS) {
959 __tcp_heuristic_tfo_middlebox_common(tpheur);
960 }
961 }
962
963 if ((flags & TCPCACHE_F_TFO_REQ_RST) && tpheur->th_tfo_req_rst < TCP_CACHE_OVERFLOW_PROTECT) {
964 tpheur->th_tfo_req_rst++;
965
966 if (tpheur->th_tfo_req_rst >= TFO_MAX_COOKIE_LOSS) {
967 __tcp_heuristic_tfo_middlebox_common(tpheur);
968 }
969 }
970
971 if ((flags & TCPCACHE_F_ECN) &&
972 tpheur->th_ecn_loss < TCP_CACHE_OVERFLOW_PROTECT &&
973 TSTMP_LEQ(tpheur->th_ecn_backoff, tcp_now)) {
974 tpheur->th_ecn_loss++;
975 if (tpheur->th_ecn_loss >= ECN_MAX_SYN_LOSS) {
976 tcpstat.tcps_ecn_fallback_synloss++;
977 TCP_CACHE_INC_IFNET_STAT(tcks->ifp, tcks->af, ecn_fallback_synloss);
978 tpheur->th_ecn_backoff = tcp_now +
979 (tcp_min_to_hz(tcp_ecn_timeout) <<
980 (tpheur->th_ecn_loss - ECN_MAX_SYN_LOSS));
981
982 os_log(OS_LOG_DEFAULT, "%s disable ECN until %u now %u on %lx for SYN-loss\n",
983 __func__, tpheur->th_ecn_backoff, tcp_now,
984 (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
985 }
986 }
987
988 if ((flags & TCPCACHE_F_MPTCP) &&
989 tpheur->th_mptcp_loss < TCP_CACHE_OVERFLOW_PROTECT &&
990 tpheur->th_mptcp_heuristic_disabled == 0) {
991 tpheur->th_mptcp_loss++;
992 if (tpheur->th_mptcp_loss >= MPTCP_MAX_SYN_LOSS) {
993 /*
994 * Yes, we take tcp_ecn_timeout, to avoid adding yet
995 * another sysctl that is just used for testing.
996 */
997 tpheur->th_mptcp_backoff = tcp_now +
998 (tcp_min_to_hz(tcp_ecn_timeout) <<
999 (tpheur->th_mptcp_loss - MPTCP_MAX_SYN_LOSS));
1000 tpheur->th_mptcp_in_backoff = 1;
1001
1002 os_log(OS_LOG_DEFAULT, "%s disable MPTCP until %u now %u on %lx\n",
1003 __func__, tpheur->th_mptcp_backoff, tcp_now,
1004 (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
1005 }
1006 }
1007
1008 if ((flags & TCPCACHE_F_ECN_DROPRST) &&
1009 tpheur->th_ecn_droprst < TCP_CACHE_OVERFLOW_PROTECT &&
1010 TSTMP_LEQ(tpheur->th_ecn_backoff, tcp_now)) {
1011 tpheur->th_ecn_droprst++;
1012 if (tpheur->th_ecn_droprst >= ECN_MAX_DROPRST) {
1013 tcpstat.tcps_ecn_fallback_droprst++;
1014 TCP_CACHE_INC_IFNET_STAT(tcks->ifp, tcks->af,
1015 ecn_fallback_droprst);
1016 tpheur->th_ecn_backoff = tcp_now +
1017 (tcp_min_to_hz(tcp_ecn_timeout) <<
1018 (tpheur->th_ecn_droprst - ECN_MAX_DROPRST));
1019
1020 os_log(OS_LOG_DEFAULT, "%s disable ECN until %u now %u on %lx for drop-RST\n",
1021 __func__, tpheur->th_ecn_backoff, tcp_now,
1022 (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
1023 }
1024 }
1025
1026 if ((flags & TCPCACHE_F_ECN_DROPRXMT) &&
1027 tpheur->th_ecn_droprxmt < TCP_CACHE_OVERFLOW_PROTECT &&
1028 TSTMP_LEQ(tpheur->th_ecn_backoff, tcp_now)) {
1029 tpheur->th_ecn_droprxmt++;
1030 if (tpheur->th_ecn_droprxmt >= ECN_MAX_DROPRXMT) {
1031 tcpstat.tcps_ecn_fallback_droprxmt++;
1032 TCP_CACHE_INC_IFNET_STAT(tcks->ifp, tcks->af,
1033 ecn_fallback_droprxmt);
1034 tpheur->th_ecn_backoff = tcp_now +
1035 (tcp_min_to_hz(tcp_ecn_timeout) <<
1036 (tpheur->th_ecn_droprxmt - ECN_MAX_DROPRXMT));
1037
1038 os_log(OS_LOG_DEFAULT, "%s disable ECN until %u now %u on %lx for drop-Rxmit\n",
1039 __func__, tpheur->th_ecn_backoff, tcp_now,
1040 (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
1041 }
1042 }
1043 if ((flags & TCPCACHE_F_ECN_SYNRST) &&
1044 tpheur->th_ecn_synrst < TCP_CACHE_OVERFLOW_PROTECT) {
1045 tpheur->th_ecn_synrst++;
1046 if (tpheur->th_ecn_synrst >= ECN_MAX_SYNRST) {
1047 tcpstat.tcps_ecn_fallback_synrst++;
1048 TCP_CACHE_INC_IFNET_STAT(tcks->ifp, tcks->af,
1049 ecn_fallback_synrst);
1050 tpheur->th_ecn_backoff = tcp_now +
1051 (tcp_min_to_hz(tcp_ecn_timeout) <<
1052 (tpheur->th_ecn_synrst - ECN_MAX_SYNRST));
1053
1054 os_log(OS_LOG_DEFAULT, "%s disable ECN until %u now %u on %lx for SYN-RST\n",
1055 __func__, tpheur->th_ecn_backoff, tcp_now,
1056 (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
1057 }
1058 }
1059 tcp_heuristic_unlock(head);
1060 }
1061
1062 void
tcp_heuristic_tfo_loss(struct tcpcb * tp)1063 tcp_heuristic_tfo_loss(struct tcpcb *tp)
1064 {
1065 struct tcp_cache_key_src tcks;
1066 uint32_t flag = 0;
1067
1068 if (symptoms_is_wifi_lossy() &&
1069 IFNET_IS_WIFI(tp->t_inpcb->inp_last_outifp)) {
1070 return;
1071 }
1072
1073 tcp_cache_key_src_create(tp, &tcks);
1074
1075 if (tp->t_tfo_stats & TFO_S_SYN_DATA_SENT) {
1076 flag = (TCPCACHE_F_TFO_DATA | TCPCACHE_F_TFO_REQ);
1077 }
1078 if (tp->t_tfo_stats & TFO_S_COOKIE_REQ) {
1079 flag = TCPCACHE_F_TFO_REQ;
1080 }
1081
1082 tcp_heuristic_inc_counters(&tcks, flag);
1083 }
1084
1085 void
tcp_heuristic_tfo_rst(struct tcpcb * tp)1086 tcp_heuristic_tfo_rst(struct tcpcb *tp)
1087 {
1088 struct tcp_cache_key_src tcks;
1089 uint32_t flag = 0;
1090
1091 tcp_cache_key_src_create(tp, &tcks);
1092
1093 if (tp->t_tfo_stats & TFO_S_SYN_DATA_SENT) {
1094 flag = (TCPCACHE_F_TFO_DATA_RST | TCPCACHE_F_TFO_REQ_RST);
1095 }
1096 if (tp->t_tfo_stats & TFO_S_COOKIE_REQ) {
1097 flag = TCPCACHE_F_TFO_REQ_RST;
1098 }
1099
1100 tcp_heuristic_inc_counters(&tcks, flag);
1101 }
1102
1103 void
tcp_heuristic_mptcp_loss(struct tcpcb * tp)1104 tcp_heuristic_mptcp_loss(struct tcpcb *tp)
1105 {
1106 struct tcp_cache_key_src tcks;
1107
1108 if (symptoms_is_wifi_lossy() &&
1109 IFNET_IS_WIFI(tp->t_inpcb->inp_last_outifp)) {
1110 return;
1111 }
1112
1113 tcp_cache_key_src_create(tp, &tcks);
1114
1115 tcp_heuristic_inc_counters(&tcks, TCPCACHE_F_MPTCP);
1116 }
1117
1118 void
tcp_heuristic_ecn_loss(struct tcpcb * tp)1119 tcp_heuristic_ecn_loss(struct tcpcb *tp)
1120 {
1121 struct tcp_cache_key_src tcks;
1122
1123 if (symptoms_is_wifi_lossy() &&
1124 IFNET_IS_WIFI(tp->t_inpcb->inp_last_outifp)) {
1125 return;
1126 }
1127
1128 tcp_cache_key_src_create(tp, &tcks);
1129
1130 tcp_heuristic_inc_counters(&tcks, TCPCACHE_F_ECN);
1131 }
1132
1133 void
tcp_heuristic_ecn_droprst(struct tcpcb * tp)1134 tcp_heuristic_ecn_droprst(struct tcpcb *tp)
1135 {
1136 struct tcp_cache_key_src tcks;
1137
1138 tcp_cache_key_src_create(tp, &tcks);
1139
1140 tcp_heuristic_inc_counters(&tcks, TCPCACHE_F_ECN_DROPRST);
1141 }
1142
1143 void
tcp_heuristic_ecn_droprxmt(struct tcpcb * tp)1144 tcp_heuristic_ecn_droprxmt(struct tcpcb *tp)
1145 {
1146 struct tcp_cache_key_src tcks;
1147
1148 tcp_cache_key_src_create(tp, &tcks);
1149
1150 tcp_heuristic_inc_counters(&tcks, TCPCACHE_F_ECN_DROPRXMT);
1151 }
1152
1153 void
tcp_heuristic_ecn_synrst(struct tcpcb * tp)1154 tcp_heuristic_ecn_synrst(struct tcpcb *tp)
1155 {
1156 struct tcp_cache_key_src tcks;
1157
1158 tcp_cache_key_src_create(tp, &tcks);
1159
1160 tcp_heuristic_inc_counters(&tcks, TCPCACHE_F_ECN_SYNRST);
1161 }
1162
1163 void
tcp_heuristic_tfo_middlebox(struct tcpcb * tp)1164 tcp_heuristic_tfo_middlebox(struct tcpcb *tp)
1165 {
1166 struct tcp_cache_key_src tcks;
1167
1168 tp->t_tfo_flags |= TFO_F_HEURISTIC_DONE;
1169
1170 tcp_cache_key_src_create(tp, &tcks);
1171 tcp_heuristic_tfo_middlebox_common(&tcks);
1172 }
1173
1174 static void
tcp_heuristic_ecn_aggressive_common(struct tcp_cache_key_src * tcks)1175 tcp_heuristic_ecn_aggressive_common(struct tcp_cache_key_src *tcks)
1176 {
1177 struct tcp_heuristics_head *head;
1178 struct tcp_heuristic *tpheur;
1179
1180 tpheur = tcp_getheuristic_with_lock(tcks, 1, &head);
1181 if (tpheur == NULL) {
1182 return;
1183 }
1184
1185 if (TSTMP_GT(tpheur->th_ecn_backoff, tcp_now)) {
1186 /* We are already in aggressive mode */
1187 tcp_heuristic_unlock(head);
1188 return;
1189 }
1190
1191 /* Must be done before, otherwise we will start off with expo-backoff */
1192 tpheur->th_ecn_backoff = tcp_now +
1193 (tcp_min_to_hz(tcp_ecn_timeout) << (tpheur->th_ecn_aggressive));
1194
1195 /*
1196 * Ugly way to prevent integer overflow... limit to prevent in
1197 * overflow during exp. backoff.
1198 */
1199 if (tpheur->th_ecn_aggressive < TCP_CACHE_OVERFLOW_PROTECT) {
1200 tpheur->th_ecn_aggressive++;
1201 }
1202
1203 tcp_heuristic_unlock(head);
1204
1205 os_log(OS_LOG_DEFAULT, "%s disable ECN until %u now %u on %lx\n", __func__,
1206 tpheur->th_ecn_backoff, tcp_now, (unsigned long)VM_KERNEL_ADDRPERM(tpheur));
1207 }
1208
1209 void
tcp_heuristic_ecn_aggressive(struct tcpcb * tp)1210 tcp_heuristic_ecn_aggressive(struct tcpcb *tp)
1211 {
1212 struct tcp_cache_key_src tcks;
1213
1214 tcp_cache_key_src_create(tp, &tcks);
1215 tcp_heuristic_ecn_aggressive_common(&tcks);
1216 }
1217
1218 static boolean_t
tcp_heuristic_do_tfo_common(struct tcp_cache_key_src * tcks)1219 tcp_heuristic_do_tfo_common(struct tcp_cache_key_src *tcks)
1220 {
1221 struct tcp_heuristics_head *head;
1222 struct tcp_heuristic *tpheur;
1223
1224 if (disable_tcp_heuristics) {
1225 return TRUE;
1226 }
1227
1228 /* Get the tcp-heuristic. */
1229 tpheur = tcp_getheuristic_with_lock(tcks, 0, &head);
1230 if (tpheur == NULL) {
1231 return TRUE;
1232 }
1233
1234 if (tpheur->th_tfo_in_backoff == 0) {
1235 goto tfo_ok;
1236 }
1237
1238 if (TSTMP_GT(tcp_now, tpheur->th_tfo_backoff_until)) {
1239 tpheur->th_tfo_in_backoff = 0;
1240 tpheur->th_tfo_enabled_time = tcp_now;
1241
1242 goto tfo_ok;
1243 }
1244
1245 tcp_heuristic_unlock(head);
1246 return FALSE;
1247
1248 tfo_ok:
1249 tcp_heuristic_unlock(head);
1250 return TRUE;
1251 }
1252
1253 boolean_t
tcp_heuristic_do_tfo(struct tcpcb * tp)1254 tcp_heuristic_do_tfo(struct tcpcb *tp)
1255 {
1256 struct tcp_cache_key_src tcks;
1257
1258 tcp_cache_key_src_create(tp, &tcks);
1259 if (tcp_heuristic_do_tfo_common(&tcks)) {
1260 return TRUE;
1261 }
1262
1263 return FALSE;
1264 }
1265 /*
1266 * @return:
1267 * 0 Enable MPTCP (we are still discovering middleboxes)
1268 * -1 Enable MPTCP (heuristics have been temporarily disabled)
1269 * 1 Disable MPTCP
1270 */
1271 int
tcp_heuristic_do_mptcp(struct tcpcb * tp)1272 tcp_heuristic_do_mptcp(struct tcpcb *tp)
1273 {
1274 struct tcp_cache_key_src tcks;
1275 struct tcp_heuristics_head *head = NULL;
1276 struct tcp_heuristic *tpheur;
1277 int ret = 0;
1278
1279 if (disable_tcp_heuristics ||
1280 (tptomptp(tp)->mpt_mpte->mpte_flags & MPTE_FORCE_ENABLE)) {
1281 return 0;
1282 }
1283
1284 tcp_cache_key_src_create(tp, &tcks);
1285
1286 /* Get the tcp-heuristic. */
1287 tpheur = tcp_getheuristic_with_lock(&tcks, 0, &head);
1288 if (tpheur == NULL) {
1289 return 0;
1290 }
1291
1292 if (tpheur->th_mptcp_in_backoff == 0 ||
1293 tpheur->th_mptcp_heuristic_disabled == 1) {
1294 goto mptcp_ok;
1295 }
1296
1297 if (TSTMP_GT(tpheur->th_mptcp_backoff, tcp_now)) {
1298 goto fallback;
1299 }
1300
1301 tpheur->th_mptcp_in_backoff = 0;
1302
1303 mptcp_ok:
1304 if (tpheur->th_mptcp_heuristic_disabled) {
1305 ret = -1;
1306
1307 if (TSTMP_GT(tcp_now, tpheur->th_mptcp_backoff)) {
1308 tpheur->th_mptcp_heuristic_disabled = 0;
1309 tpheur->th_mptcp_success = 0;
1310 }
1311 }
1312
1313 tcp_heuristic_unlock(head);
1314 return ret;
1315
1316 fallback:
1317 if (head) {
1318 tcp_heuristic_unlock(head);
1319 }
1320
1321 if (tptomptp(tp)->mpt_mpte->mpte_flags & MPTE_FIRSTPARTY) {
1322 tcpstat.tcps_mptcp_fp_heuristic_fallback++;
1323 } else {
1324 tcpstat.tcps_mptcp_heuristic_fallback++;
1325 }
1326
1327 return 1;
1328 }
1329
1330 static boolean_t
tcp_heuristic_do_ecn_common(struct tcp_cache_key_src * tcks)1331 tcp_heuristic_do_ecn_common(struct tcp_cache_key_src *tcks)
1332 {
1333 struct tcp_heuristics_head *head;
1334 struct tcp_heuristic *tpheur;
1335 boolean_t ret = TRUE;
1336
1337 if (disable_tcp_heuristics) {
1338 return TRUE;
1339 }
1340
1341 /* Get the tcp-heuristic. */
1342 tpheur = tcp_getheuristic_with_lock(tcks, 0, &head);
1343 if (tpheur == NULL) {
1344 return ret;
1345 }
1346
1347 if (TSTMP_GT(tpheur->th_ecn_backoff, tcp_now)) {
1348 ret = FALSE;
1349 } else {
1350 /* Reset the following counters to start re-evaluating */
1351 if (tpheur->th_ecn_droprst >= ECN_RETRY_LIMIT) {
1352 tpheur->th_ecn_droprst = 0;
1353 }
1354 if (tpheur->th_ecn_droprxmt >= ECN_RETRY_LIMIT) {
1355 tpheur->th_ecn_droprxmt = 0;
1356 }
1357 if (tpheur->th_ecn_synrst >= ECN_RETRY_LIMIT) {
1358 tpheur->th_ecn_synrst = 0;
1359 }
1360
1361 /* Make sure it follows along */
1362 tpheur->th_ecn_backoff = tcp_now;
1363 }
1364
1365 tcp_heuristic_unlock(head);
1366
1367 return ret;
1368 }
1369
1370 boolean_t
tcp_heuristic_do_ecn(struct tcpcb * tp)1371 tcp_heuristic_do_ecn(struct tcpcb *tp)
1372 {
1373 struct tcp_cache_key_src tcks;
1374
1375 tcp_cache_key_src_create(tp, &tcks);
1376 return tcp_heuristic_do_ecn_common(&tcks);
1377 }
1378
1379 boolean_t
tcp_heuristic_do_ecn_with_address(struct ifnet * ifp,union sockaddr_in_4_6 * local_address)1380 tcp_heuristic_do_ecn_with_address(struct ifnet *ifp,
1381 union sockaddr_in_4_6 *local_address)
1382 {
1383 struct tcp_cache_key_src tcks;
1384
1385 memset(&tcks, 0, sizeof(tcks));
1386 tcks.ifp = ifp;
1387
1388 calculate_tcp_clock();
1389
1390 if (local_address->sa.sa_family == AF_INET6) {
1391 memcpy(&tcks.laddr.addr6, &local_address->sin6.sin6_addr, sizeof(struct in6_addr));
1392 tcks.af = AF_INET6;
1393 } else if (local_address->sa.sa_family == AF_INET) {
1394 memcpy(&tcks.laddr.addr, &local_address->sin.sin_addr, sizeof(struct in_addr));
1395 tcks.af = AF_INET;
1396 }
1397
1398 return tcp_heuristic_do_ecn_common(&tcks);
1399 }
1400
1401 void
tcp_heuristics_ecn_update(struct necp_tcp_ecn_cache * necp_buffer,struct ifnet * ifp,union sockaddr_in_4_6 * local_address)1402 tcp_heuristics_ecn_update(struct necp_tcp_ecn_cache *necp_buffer,
1403 struct ifnet *ifp, union sockaddr_in_4_6 *local_address)
1404 {
1405 struct tcp_cache_key_src tcks;
1406
1407 memset(&tcks, 0, sizeof(tcks));
1408 tcks.ifp = ifp;
1409
1410 calculate_tcp_clock();
1411
1412 if (local_address->sa.sa_family == AF_INET6) {
1413 memcpy(&tcks.laddr.addr6, &local_address->sin6.sin6_addr, sizeof(struct in6_addr));
1414 tcks.af = AF_INET6;
1415 } else if (local_address->sa.sa_family == AF_INET) {
1416 memcpy(&tcks.laddr.addr, &local_address->sin.sin_addr, sizeof(struct in_addr));
1417 tcks.af = AF_INET;
1418 }
1419
1420 if (necp_buffer->necp_tcp_ecn_heuristics_success) {
1421 tcp_heuristic_reset_counters(&tcks, TCPCACHE_F_ECN);
1422 } else if (necp_buffer->necp_tcp_ecn_heuristics_loss) {
1423 tcp_heuristic_inc_counters(&tcks, TCPCACHE_F_ECN);
1424 } else if (necp_buffer->necp_tcp_ecn_heuristics_drop_rst) {
1425 tcp_heuristic_inc_counters(&tcks, TCPCACHE_F_ECN_DROPRST);
1426 } else if (necp_buffer->necp_tcp_ecn_heuristics_drop_rxmt) {
1427 tcp_heuristic_inc_counters(&tcks, TCPCACHE_F_ECN_DROPRXMT);
1428 } else if (necp_buffer->necp_tcp_ecn_heuristics_syn_rst) {
1429 tcp_heuristic_inc_counters(&tcks, TCPCACHE_F_ECN_SYNRST);
1430 } else if (necp_buffer->necp_tcp_ecn_heuristics_aggressive) {
1431 tcp_heuristic_ecn_aggressive_common(&tcks);
1432 }
1433
1434 return;
1435 }
1436
1437 boolean_t
tcp_heuristic_do_tfo_with_address(struct ifnet * ifp,union sockaddr_in_4_6 * local_address,union sockaddr_in_4_6 * remote_address,uint8_t * cookie,uint8_t * cookie_len)1438 tcp_heuristic_do_tfo_with_address(struct ifnet *ifp,
1439 union sockaddr_in_4_6 *local_address, union sockaddr_in_4_6 *remote_address,
1440 uint8_t *cookie, uint8_t *cookie_len)
1441 {
1442 struct tcp_cache_key_src tcks;
1443
1444 memset(&tcks, 0, sizeof(tcks));
1445 tcks.ifp = ifp;
1446
1447 calculate_tcp_clock();
1448
1449 if (remote_address->sa.sa_family == AF_INET6) {
1450 memcpy(&tcks.laddr.addr6, &local_address->sin6.sin6_addr, sizeof(struct in6_addr));
1451 memcpy(&tcks.faddr.addr6, &remote_address->sin6.sin6_addr, sizeof(struct in6_addr));
1452 tcks.af = AF_INET6;
1453 } else if (remote_address->sa.sa_family == AF_INET) {
1454 memcpy(&tcks.laddr.addr, &local_address->sin.sin_addr, sizeof(struct in_addr));
1455 memcpy(&tcks.faddr.addr, &remote_address->sin.sin_addr, sizeof(struct in_addr));
1456 tcks.af = AF_INET;
1457 }
1458
1459 if (tcp_heuristic_do_tfo_common(&tcks)) {
1460 if (!tcp_cache_get_cookie_common(&tcks, cookie, cookie_len)) {
1461 *cookie_len = 0;
1462 }
1463 return TRUE;
1464 }
1465
1466 return FALSE;
1467 }
1468
1469 void
tcp_heuristics_tfo_update(struct necp_tcp_tfo_cache * necp_buffer,struct ifnet * ifp,union sockaddr_in_4_6 * local_address,union sockaddr_in_4_6 * remote_address)1470 tcp_heuristics_tfo_update(struct necp_tcp_tfo_cache *necp_buffer,
1471 struct ifnet *ifp, union sockaddr_in_4_6 *local_address,
1472 union sockaddr_in_4_6 *remote_address)
1473 {
1474 struct tcp_cache_key_src tcks;
1475
1476 memset(&tcks, 0, sizeof(tcks));
1477 tcks.ifp = ifp;
1478
1479 calculate_tcp_clock();
1480
1481 if (remote_address->sa.sa_family == AF_INET6) {
1482 memcpy(&tcks.laddr.addr6, &local_address->sin6.sin6_addr, sizeof(struct in6_addr));
1483 memcpy(&tcks.faddr.addr6, &remote_address->sin6.sin6_addr, sizeof(struct in6_addr));
1484 tcks.af = AF_INET6;
1485 } else if (remote_address->sa.sa_family == AF_INET) {
1486 memcpy(&tcks.laddr.addr, &local_address->sin.sin_addr, sizeof(struct in_addr));
1487 memcpy(&tcks.faddr.addr, &remote_address->sin.sin_addr, sizeof(struct in_addr));
1488 tcks.af = AF_INET;
1489 }
1490
1491 if (necp_buffer->necp_tcp_tfo_heuristics_success) {
1492 tcp_heuristic_reset_counters(&tcks, TCPCACHE_F_TFO_REQ | TCPCACHE_F_TFO_DATA |
1493 TCPCACHE_F_TFO_REQ_RST | TCPCACHE_F_TFO_DATA_RST);
1494 }
1495
1496 if (necp_buffer->necp_tcp_tfo_heuristics_success_req) {
1497 tcp_heuristic_reset_counters(&tcks, TCPCACHE_F_TFO_REQ | TCPCACHE_F_TFO_REQ_RST);
1498 }
1499
1500 if (necp_buffer->necp_tcp_tfo_heuristics_loss) {
1501 tcp_heuristic_inc_counters(&tcks, TCPCACHE_F_TFO_REQ | TCPCACHE_F_TFO_DATA);
1502 }
1503
1504 if (necp_buffer->necp_tcp_tfo_heuristics_loss_req) {
1505 tcp_heuristic_inc_counters(&tcks, TCPCACHE_F_TFO_REQ);
1506 }
1507
1508 if (necp_buffer->necp_tcp_tfo_heuristics_rst_data) {
1509 tcp_heuristic_inc_counters(&tcks, TCPCACHE_F_TFO_REQ_RST | TCPCACHE_F_TFO_DATA_RST);
1510 }
1511
1512 if (necp_buffer->necp_tcp_tfo_heuristics_rst_req) {
1513 tcp_heuristic_inc_counters(&tcks, TCPCACHE_F_TFO_REQ_RST);
1514 }
1515
1516 if (necp_buffer->necp_tcp_tfo_heuristics_middlebox) {
1517 tcp_heuristic_tfo_middlebox_common(&tcks);
1518 }
1519
1520 if (necp_buffer->necp_tcp_tfo_cookie_len != 0) {
1521 tcp_cache_set_cookie_common(&tcks,
1522 necp_buffer->necp_tcp_tfo_cookie, necp_buffer->necp_tcp_tfo_cookie_len);
1523 }
1524
1525 return;
1526 }
1527
1528 static void
sysctl_cleartfocache(void)1529 sysctl_cleartfocache(void)
1530 {
1531 int i;
1532
1533 for (i = 0; i < tcp_cache_size; i++) {
1534 struct tcp_cache_head *head = &tcp_cache[i];
1535 struct tcp_cache *tpcache, *tmp;
1536 struct tcp_heuristics_head *hhead = &tcp_heuristics[i];
1537 struct tcp_heuristic *tpheur, *htmp;
1538
1539 lck_mtx_lock(&head->tch_mtx);
1540 SLIST_FOREACH_SAFE(tpcache, &head->tcp_caches, list, tmp) {
1541 SLIST_REMOVE(&head->tcp_caches, tpcache, tcp_cache, list);
1542 kfree_type(struct tcp_cache, tpcache);
1543 }
1544 lck_mtx_unlock(&head->tch_mtx);
1545
1546 lck_mtx_lock(&hhead->thh_mtx);
1547 SLIST_FOREACH_SAFE(tpheur, &hhead->tcp_heuristics, list, htmp) {
1548 SLIST_REMOVE(&hhead->tcp_heuristics, tpheur, tcp_heuristic, list);
1549 kfree_type(struct tcp_heuristic, tpheur);
1550 }
1551 lck_mtx_unlock(&hhead->thh_mtx);
1552 }
1553 }
1554
1555 /* This sysctl is useful for testing purposes only */
1556 static int tcpcleartfo = 0;
1557
1558 static int sysctl_cleartfo SYSCTL_HANDLER_ARGS
1559 {
1560 #pragma unused(arg1, arg2)
1561 int error = 0, val, oldval = tcpcleartfo;
1562
1563 val = oldval;
1564 error = sysctl_handle_int(oidp, &val, 0, req);
1565 if (error || !req->newptr) {
1566 if (error) {
1567 os_log_error(OS_LOG_DEFAULT, "%s could not parse int: %d", __func__, error);
1568 }
1569 return error;
1570 }
1571
1572 /*
1573 * The actual value does not matter. If the value is set, it triggers
1574 * the clearing of the TFO cache. If a future implementation does not
1575 * use the route entry to hold the TFO cache, replace the route sysctl.
1576 */
1577
1578 if (val != oldval) {
1579 sysctl_cleartfocache();
1580 }
1581
1582 tcpcleartfo = val;
1583
1584 return error;
1585 }
1586
1587 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, clear_tfocache, CTLTYPE_INT | CTLFLAG_RW |
1588 CTLFLAG_LOCKED, &tcpcleartfo, 0, &sysctl_cleartfo, "I",
1589 "Toggle to clear the TFO destination based heuristic cache");
1590
1591 void
tcp_cache_init(void)1592 tcp_cache_init(void)
1593 {
1594 uint64_t sane_size_meg = sane_size / 1024 / 1024;
1595
1596 /*
1597 * On machines with <100MB of memory this will result in a (full) cache-size
1598 * of 32 entries, thus 32 * 5 * 64bytes = 10KB. (about 0.01 %)
1599 * On machines with > 4GB of memory, we have a cache-size of 1024 entries,
1600 * thus about 327KB.
1601 *
1602 * Side-note: we convert to uint32_t. If sane_size is more than
1603 * 16000 TB, we loose precision. But, who cares? :)
1604 */
1605 tcp_cache_size = tcp_cache_roundup2((uint32_t)(sane_size_meg >> 2));
1606 if (tcp_cache_size < 32) {
1607 tcp_cache_size = 32;
1608 } else if (tcp_cache_size > 1024) {
1609 tcp_cache_size = 1024;
1610 }
1611
1612 tcp_cache = zalloc_permanent(sizeof(struct tcp_cache_head) * tcp_cache_size,
1613 ZALIGN(struct tcp_cache_head));
1614
1615 tcp_heuristics = zalloc_permanent(sizeof(struct tcp_heuristics_head) * tcp_cache_size,
1616 ZALIGN(struct tcp_heuristics_head));
1617
1618 for (int i = 0; i < tcp_cache_size; i++) {
1619 lck_mtx_init(&tcp_cache[i].tch_mtx, &tcp_cache_mtx_grp,
1620 &tcp_cache_mtx_attr);
1621 SLIST_INIT(&tcp_cache[i].tcp_caches);
1622
1623 lck_mtx_init(&tcp_heuristics[i].thh_mtx, &tcp_heuristic_mtx_grp,
1624 &tcp_heuristic_mtx_attr);
1625 SLIST_INIT(&tcp_heuristics[i].tcp_heuristics);
1626 }
1627
1628 tcp_cache_hash_seed = RandomULong();
1629 }
1630