xref: /xnu-11215/bsd/netinet/tcp_cache.c (revision 8d741a5d)
1 /*
2  * Copyright (c) 2015-2024 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 
45 typedef union {
46 	struct in_addr addr;
47 	struct in6_addr addr6;
48 } in_4_6_addr;
49 
50 struct 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 
58 struct 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 #define th_val_start th_tfo_data_loss
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 	// N.B.: we may sometimes erase ALL values from th_val_start to the end of the structure.
87 };
88 
89 struct tcp_heuristics_head {
90 	SLIST_HEAD(tcp_heur_bucket, tcp_heuristic) tcp_heuristics;
91 
92 	/* Per-hashbucket lock to avoid lock-contention */
93 	lck_mtx_t       thh_mtx;
94 };
95 
96 struct tcp_cache_key {
97 	sa_family_t     tck_family;
98 
99 	struct tcp_heuristic_key tck_src;
100 	in_4_6_addr tck_dst;
101 };
102 
103 #define MPTCP_VERSION_SUPPORTED 1
104 #define MPTCP_VERSION_UNSUPPORTED -1
105 #define MPTCP_VERSION_SUPPORTED_UNKNOWN 0
106 struct tcp_cache {
107 	SLIST_ENTRY(tcp_cache) list;
108 
109 	uint32_t       tc_last_access;
110 
111 	struct tcp_cache_key tc_key;
112 
113 	uint8_t        tc_tfo_cookie[TFO_COOKIE_LEN_MAX];
114 	uint8_t        tc_tfo_cookie_len;
115 
116 	uint8_t        tc_mptcp_version_confirmed:1;
117 	uint8_t        tc_mptcp_version; /* version to use right now */
118 	uint32_t       tc_mptcp_next_version_try; /* Time, until we try preferred version again */
119 };
120 
121 struct tcp_cache_head {
122 	SLIST_HEAD(tcp_cache_bucket, tcp_cache) tcp_caches;
123 
124 	/* Per-hashbucket lock to avoid lock-contention */
125 	lck_mtx_t       tch_mtx;
126 };
127 
128 struct tcp_cache_key_src {
129 	struct ifnet *ifp;
130 	in_4_6_addr laddr;
131 	in_4_6_addr faddr;
132 	int af;
133 };
134 
135 static uint32_t tcp_cache_hash_seed;
136 
137 static size_t tcp_cache_size;
138 static size_t tcp_heuristics_size;
139 
140 /*
141  * The maximum depth of the hash-bucket. This way we limit the tcp_cache to
142  * TCP_CACHE_BUCKET_SIZE * tcp_cache_size and have "natural" garbage collection
143  */
144 #define TCP_CACHE_BUCKET_SIZE 5
145 
146 static struct tcp_cache_head *__counted_by(tcp_cache_size) tcp_cache;
147 
148 static LCK_ATTR_DECLARE(tcp_cache_mtx_attr, 0, 0);
149 static LCK_GRP_DECLARE(tcp_cache_mtx_grp, "tcpcache");
150 
151 static struct tcp_heuristics_head *__counted_by(tcp_heuristics_size) tcp_heuristics;
152 
153 static LCK_ATTR_DECLARE(tcp_heuristic_mtx_attr, 0, 0);
154 static LCK_GRP_DECLARE(tcp_heuristic_mtx_grp, "tcpheuristic");
155 
156 static uint32_t tcp_backoff_maximum = 65536;
157 
158 SYSCTL_UINT(_net_inet_tcp, OID_AUTO, backoff_maximum, CTLFLAG_RW | CTLFLAG_LOCKED,
159     &tcp_backoff_maximum, 0, "Maximum time for which we won't try TFO");
160 
161 static uint32_t tcp_ecn_timeout = 60;
162 
163 SYSCTL_UINT(_net_inet_tcp, OID_AUTO, ecn_timeout, CTLFLAG_RW | CTLFLAG_LOCKED,
164     &tcp_ecn_timeout, 60, "Initial minutes to wait before re-trying ECN");
165 
166 static int disable_tcp_heuristics = 0;
167 SYSCTL_INT(_net_inet_tcp, OID_AUTO, disable_tcp_heuristics, CTLFLAG_RW | CTLFLAG_LOCKED,
168     &disable_tcp_heuristics, 0, "Set to 1, to disable all TCP heuristics (TFO, ECN, MPTCP)");
169 
170 static uint32_t mptcp_version_timeout = 24 * 60;
171 
172 SYSCTL_UINT(_net_inet_tcp, OID_AUTO, mptcp_version_timeout, CTLFLAG_RW | CTLFLAG_LOCKED,
173     &mptcp_version_timeout, 24 * 60, "Initial minutes to wait before re-trying MPTCP's preferred version");
174 
175 
176 static uint32_t
tcp_min_to_hz(uint32_t minutes)177 tcp_min_to_hz(uint32_t minutes)
178 {
179 	if (minutes > 65536) {
180 		return (uint32_t)65536 * 60 * TCP_RETRANSHZ;
181 	}
182 
183 	return minutes * 60 * TCP_RETRANSHZ;
184 }
185 
186 /*
187  * This number is coupled with tcp_ecn_timeout, because we want to prevent
188  * integer overflow. Need to find an unexpensive way to prevent integer overflow
189  * while still allowing a dynamic sysctl.
190  */
191 #define TCP_CACHE_OVERFLOW_PROTECT      9
192 
193 /* Number of SYN-losses we accept */
194 #define TFO_MAX_COOKIE_LOSS     2
195 #define ECN_MAX_SYN_LOSS        2
196 #define MPTCP_MAX_SYN_LOSS      2
197 #define MPTCP_SUCCESS_TRIGGER   10
198 #define MPTCP_VERSION_MAX_FAIL  2
199 #define ECN_MAX_DROPRST         1
200 #define ECN_MAX_DROPRXMT        4
201 #define ECN_MAX_SYNRST          4
202 
203 /* Flags for setting/unsetting loss-heuristics, limited to 4 bytes */
204 #define TCPCACHE_F_TFO_REQ      0x01
205 #define TCPCACHE_F_TFO_DATA     0x02
206 #define TCPCACHE_F_ECN          0x04
207 #define TCPCACHE_F_MPTCP        0x08
208 #define TCPCACHE_F_ECN_DROPRST  0x10
209 #define TCPCACHE_F_ECN_DROPRXMT 0x20
210 #define TCPCACHE_F_TFO_REQ_RST  0x40
211 #define TCPCACHE_F_TFO_DATA_RST 0x80
212 #define TCPCACHE_F_ECN_SYNRST   0x100
213 
214 /* Always retry ECN after backing off to this level for some heuristics */
215 #define ECN_RETRY_LIMIT 9
216 
217 #define TCP_CACHE_INC_IFNET_STAT(_ifp_, _af_, _stat_) { \
218 	if ((_ifp_) != NULL) { \
219 	        if ((_af_) == AF_INET6) { \
220 	                (_ifp_)->if_ipv6_stat->_stat_++;\
221 	        } else { \
222 	                (_ifp_)->if_ipv4_stat->_stat_++;\
223 	        }\
224 	}\
225 }
226 
227 /*
228  * Round up to next higher power-of 2.  See "Bit Twiddling Hacks".
229  *
230  * Might be worth moving this to a library so that others
231  * (e.g., scale_to_powerof2()) can use this as well instead of a while-loop.
232  */
233 static uint32_t
tcp_cache_roundup2(uint32_t a)234 tcp_cache_roundup2(uint32_t a)
235 {
236 	a--;
237 	a |= a >> 1;
238 	a |= a >> 2;
239 	a |= a >> 4;
240 	a |= a >> 8;
241 	a |= a >> 16;
242 	a++;
243 
244 	return a;
245 }
246 
247 static void
tcp_cache_hash_src(struct tcp_cache_key_src * tcks,struct tcp_heuristic_key * key)248 tcp_cache_hash_src(struct tcp_cache_key_src *tcks, struct tcp_heuristic_key *key)
249 {
250 	struct ifnet *ifp = tcks->ifp;
251 	uint8_t len = sizeof(key->thk_net_signature);
252 	uint16_t flags;
253 
254 	if (tcks->af == AF_INET6) {
255 		int ret;
256 
257 		key->thk_family = AF_INET6;
258 		ret = ifnet_get_netsignature(ifp, AF_INET6, &len, &flags,
259 		    key->thk_net_signature);
260 
261 		/*
262 		 * ifnet_get_netsignature only returns EINVAL if ifn is NULL
263 		 * (we made sure that in the other cases it does not). So,
264 		 * in this case we should take the connection's address.
265 		 */
266 		if (ret == ENOENT || ret == EINVAL) {
267 			memcpy(&key->thk_ip.addr6, &tcks->laddr.addr6, sizeof(struct in6_addr));
268 		}
269 	} else {
270 		int ret;
271 
272 		key->thk_family = AF_INET;
273 		ret = ifnet_get_netsignature(ifp, AF_INET, &len, &flags,
274 		    key->thk_net_signature);
275 
276 		/*
277 		 * ifnet_get_netsignature only returns EINVAL if ifn is NULL
278 		 * (we made sure that in the other cases it does not). So,
279 		 * in this case we should take the connection's address.
280 		 */
281 		if (ret == ENOENT || ret == EINVAL) {
282 			memcpy(&key->thk_ip.addr, &tcks->laddr.addr, sizeof(struct in_addr));
283 		}
284 	}
285 }
286 
287 static uint16_t
tcp_cache_hash(struct tcp_cache_key_src * tcks,struct tcp_cache_key * key)288 tcp_cache_hash(struct tcp_cache_key_src *tcks, struct tcp_cache_key *key)
289 {
290 	uint32_t hash;
291 
292 	bzero(key, sizeof(struct tcp_cache_key));
293 
294 	tcp_cache_hash_src(tcks, &key->tck_src);
295 
296 	if (tcks->af == AF_INET6) {
297 		key->tck_family = AF_INET6;
298 		memcpy(&key->tck_dst.addr6, &tcks->faddr.addr6,
299 		    sizeof(struct in6_addr));
300 	} else {
301 		key->tck_family = AF_INET;
302 		memcpy(&key->tck_dst.addr, &tcks->faddr.addr,
303 		    sizeof(struct in_addr));
304 	}
305 
306 	hash = net_flowhash(key, sizeof(struct tcp_cache_key),
307 	    tcp_cache_hash_seed);
308 
309 	return (uint16_t)(hash & (tcp_cache_size - 1));
310 }
311 
312 static void
tcp_cache_unlock(struct tcp_cache_head * head)313 tcp_cache_unlock(struct tcp_cache_head *head)
314 {
315 	lck_mtx_unlock(&head->tch_mtx);
316 }
317 
318 /*
319  * Make sure that everything that happens after tcp_getcache_with_lock()
320  * is short enough to justify that you hold the per-bucket lock!!!
321  *
322  * Otherwise, better build another lookup-function that does not hold the
323  * lock and you copy out the bits and bytes.
324  *
325  * That's why we provide the head as a "return"-pointer so that the caller
326  * can give it back to use for tcp_cache_unlock().
327  */
328 static struct tcp_cache *
tcp_getcache_with_lock(struct tcp_cache_key_src * tcks,int create,struct tcp_cache_head ** headarg)329 tcp_getcache_with_lock(struct tcp_cache_key_src *tcks,
330     int create, struct tcp_cache_head **headarg)
331 {
332 	struct tcp_cache *__single tpcache = NULL;
333 	struct tcp_cache_head *__single head;
334 	struct tcp_cache_key key;
335 	uint16_t hash;
336 	int i = 0;
337 
338 	hash = tcp_cache_hash(tcks, &key);
339 	head = &tcp_cache[hash];
340 
341 	lck_mtx_lock(&head->tch_mtx);
342 
343 	/*** First step: Look for the tcp_cache in our bucket ***/
344 	SLIST_FOREACH(tpcache, &head->tcp_caches, list) {
345 		if (memcmp(&tpcache->tc_key, &key, sizeof(key)) == 0) {
346 			break;
347 		}
348 
349 		i++;
350 	}
351 
352 	/*** Second step: If it's not there, create/recycle it ***/
353 	if ((tpcache == NULL) && create) {
354 		if (i >= TCP_CACHE_BUCKET_SIZE) {
355 			struct tcp_cache *oldest_cache = NULL;
356 			uint32_t max_age = 0;
357 
358 			/* Look for the oldest tcp_cache in the bucket */
359 			SLIST_FOREACH(tpcache, &head->tcp_caches, list) {
360 				uint32_t age = tcp_now - tpcache->tc_last_access;
361 				if (age >= max_age) {
362 					max_age = age;
363 					oldest_cache = tpcache;
364 				}
365 			}
366 			VERIFY(oldest_cache != NULL);
367 
368 			tpcache = oldest_cache;
369 
370 			/* We recycle, thus let's indicate that there is no cookie */
371 			tpcache->tc_tfo_cookie_len = 0;
372 		} else {
373 			/* Create a new cache and add it to the list */
374 			tpcache = kalloc_type(struct tcp_cache, Z_NOPAGEWAIT | Z_ZERO);
375 			if (tpcache == NULL) {
376 				os_log_error(OS_LOG_DEFAULT, "%s could not allocate cache", __func__);
377 				goto out_null;
378 			}
379 
380 			tpcache->tc_mptcp_version = (uint8_t)mptcp_preferred_version;
381 			tpcache->tc_mptcp_next_version_try = tcp_now;
382 
383 			SLIST_INSERT_HEAD(&head->tcp_caches, tpcache, list);
384 		}
385 
386 		memcpy(&tpcache->tc_key, &key, sizeof(key));
387 	}
388 
389 	if (tpcache == NULL) {
390 		goto out_null;
391 	}
392 
393 	/* Update timestamp for garbage collection purposes */
394 	tpcache->tc_last_access = tcp_now;
395 	*headarg = head;
396 
397 	return tpcache;
398 
399 out_null:
400 	tcp_cache_unlock(head);
401 	return NULL;
402 }
403 
404 static void
tcp_cache_key_src_create(struct tcpcb * tp,struct tcp_cache_key_src * tcks)405 tcp_cache_key_src_create(struct tcpcb *tp, struct tcp_cache_key_src *tcks)
406 {
407 	struct inpcb *inp = tp->t_inpcb;
408 	memset(tcks, 0, sizeof(*tcks));
409 
410 	tcks->ifp = inp->inp_last_outifp;
411 
412 	if (inp->inp_vflag & INP_IPV6) {
413 		memcpy(&tcks->laddr.addr6, &inp->in6p_laddr, sizeof(struct in6_addr));
414 		memcpy(&tcks->faddr.addr6, &inp->in6p_faddr, sizeof(struct in6_addr));
415 		tcks->af = AF_INET6;
416 	} else {
417 		memcpy(&tcks->laddr.addr, &inp->inp_laddr, sizeof(struct in_addr));
418 		memcpy(&tcks->faddr.addr, &inp->inp_faddr, sizeof(struct in_addr));
419 		tcks->af = AF_INET;
420 	}
421 
422 	return;
423 }
424 
425 static void
mptcp_version_cache_key_src_init(struct sockaddr * dst,struct tcp_cache_key_src * tcks)426 mptcp_version_cache_key_src_init(struct sockaddr *dst, struct tcp_cache_key_src *tcks)
427 {
428 	memset(tcks, 0, sizeof(*tcks));
429 
430 	if (dst->sa_family == AF_INET) {
431 		memcpy(&tcks->faddr.addr, &SIN(dst)->sin_addr, sizeof(struct in_addr));
432 		tcks->af = AF_INET;
433 	} else {
434 		memcpy(&tcks->faddr.addr6, &SIN6(dst)->sin6_addr, sizeof(struct in6_addr));
435 		tcks->af = AF_INET6;
436 	}
437 
438 	return;
439 }
440 
441 static void
tcp_cache_set_cookie_common(struct tcp_cache_key_src * tcks,u_char * __counted_by (len)cookie,uint8_t len)442 tcp_cache_set_cookie_common(struct tcp_cache_key_src *tcks, u_char *__counted_by(len) cookie, uint8_t len)
443 {
444 	struct tcp_cache_head *__single head;
445 	struct tcp_cache *__single tpcache;
446 
447 	/* Call lookup/create function */
448 	tpcache = tcp_getcache_with_lock(tcks, 1, &head);
449 	if (tpcache == NULL) {
450 		return;
451 	}
452 
453 	tpcache->tc_tfo_cookie_len = len > TFO_COOKIE_LEN_MAX ?
454 	    TFO_COOKIE_LEN_MAX : len;
455 	memcpy(tpcache->tc_tfo_cookie, cookie, tpcache->tc_tfo_cookie_len);
456 
457 	tcp_cache_unlock(head);
458 }
459 
460 void
tcp_cache_set_cookie(struct tcpcb * tp,u_char * __counted_by (len)cookie,uint8_t len)461 tcp_cache_set_cookie(struct tcpcb *tp, u_char *__counted_by(len) cookie, uint8_t len)
462 {
463 	struct tcp_cache_key_src tcks;
464 
465 	tcp_cache_key_src_create(tp, &tcks);
466 	tcp_cache_set_cookie_common(&tcks, cookie, len);
467 }
468 
469 static int
tcp_cache_get_cookie_common(struct tcp_cache_key_src * tcks,u_char * __counted_by (maxlen)cookie,uint8_t maxlen,uint8_t * len)470 tcp_cache_get_cookie_common(struct tcp_cache_key_src *tcks,
471     u_char *__counted_by(maxlen) cookie, uint8_t maxlen, uint8_t *len)
472 {
473 #pragma unused(maxlen)
474 	struct tcp_cache_head *__single head;
475 	struct tcp_cache *__single tpcache;
476 
477 	/* Call lookup/create function */
478 	tpcache = tcp_getcache_with_lock(tcks, 1, &head);
479 	if (tpcache == NULL) {
480 		return 0;
481 	}
482 
483 	if (tpcache->tc_tfo_cookie_len == 0) {
484 		tcp_cache_unlock(head);
485 		return 0;
486 	}
487 
488 	/*
489 	 * Not enough space - this should never happen as it has been checked
490 	 * in tcp_tfo_check. So, fail here!
491 	 */
492 	VERIFY(tpcache->tc_tfo_cookie_len <= *len);
493 
494 	memcpy(cookie, tpcache->tc_tfo_cookie, tpcache->tc_tfo_cookie_len);
495 	*len = tpcache->tc_tfo_cookie_len;
496 
497 	tcp_cache_unlock(head);
498 
499 	return 1;
500 }
501 
502 /*
503  * Get the cookie related to 'tp', and copy it into 'cookie', provided that len
504  * is big enough (len designates the available memory.
505  * Upon return, 'len' is set to the cookie's length.
506  *
507  * Returns 0 if we should request a cookie.
508  * Returns 1 if the cookie has been found and written.
509  */
510 int
tcp_cache_get_cookie(struct tcpcb * tp,u_char * __counted_by (maxlen)cookie,uint8_t maxlen,uint8_t * len)511 tcp_cache_get_cookie(struct tcpcb *tp, u_char *__counted_by(maxlen) cookie,
512     uint8_t maxlen, uint8_t *len)
513 {
514 	struct tcp_cache_key_src tcks;
515 
516 	tcp_cache_key_src_create(tp, &tcks);
517 	return tcp_cache_get_cookie_common(&tcks, cookie, maxlen, len);
518 }
519 
520 static unsigned int
tcp_cache_get_cookie_len_common(struct tcp_cache_key_src * tcks)521 tcp_cache_get_cookie_len_common(struct tcp_cache_key_src *tcks)
522 {
523 	struct tcp_cache_head *__single head;
524 	struct tcp_cache *__single tpcache;
525 	unsigned int cookie_len;
526 
527 	/* Call lookup/create function */
528 	tpcache = tcp_getcache_with_lock(tcks, 1, &head);
529 	if (tpcache == NULL) {
530 		return 0;
531 	}
532 
533 	cookie_len = tpcache->tc_tfo_cookie_len;
534 
535 	tcp_cache_unlock(head);
536 
537 	return cookie_len;
538 }
539 
540 unsigned int
tcp_cache_get_cookie_len(struct tcpcb * tp)541 tcp_cache_get_cookie_len(struct tcpcb *tp)
542 {
543 	struct tcp_cache_key_src tcks;
544 
545 	tcp_cache_key_src_create(tp, &tcks);
546 	return tcp_cache_get_cookie_len_common(&tcks);
547 }
548 
549 /*
550  * @return:
551  *         0	MPTCP_VERSION_0
552  *         1	MPTCP_VERSION_1
553  */
554 uint8_t
tcp_cache_get_mptcp_version(struct sockaddr * dst)555 tcp_cache_get_mptcp_version(struct sockaddr *dst)
556 {
557 	struct tcp_cache_key_src tcks;
558 	mptcp_version_cache_key_src_init(dst, &tcks);
559 	uint8_t version = (uint8_t) mptcp_preferred_version;
560 
561 	struct tcp_cache_head *__single head;
562 	struct tcp_cache *__single tpcache;
563 
564 	/* Call lookup/create function */
565 	tpcache = tcp_getcache_with_lock(&tcks, 1, &head);
566 	if (tpcache == NULL) {
567 		return version;
568 	}
569 
570 	version = tpcache->tc_mptcp_version;
571 
572 	/* Let's see if we should try the preferred version again */
573 	if (!tpcache->tc_mptcp_version_confirmed &&
574 	    version != mptcp_preferred_version &&
575 	    TSTMP_GEQ(tcp_now, tpcache->tc_mptcp_next_version_try)) {
576 		version = (uint8_t) mptcp_preferred_version;
577 	}
578 
579 	tcp_cache_unlock(head);
580 	return version;
581 }
582 
583 void
tcp_cache_update_mptcp_version(struct tcpcb * tp,boolean_t succeeded)584 tcp_cache_update_mptcp_version(struct tcpcb *tp, boolean_t succeeded)
585 {
586 	uint8_t version = tptomptp(tp)->mpt_version;
587 	struct inpcb *inp = tp->t_inpcb;
588 	struct tcp_cache_key_src tcks;
589 	struct tcp_cache_head *__single head;
590 	struct tcp_cache *__single tpcache;
591 
592 	if (inp->inp_vflag & INP_IPV6) {
593 		struct sockaddr_in6 dst = {
594 			.sin6_len = sizeof(struct sockaddr_in6),
595 			.sin6_family = AF_INET6,
596 			.sin6_addr = inp->in6p_faddr,
597 		};
598 		mptcp_version_cache_key_src_init(SA(&dst), &tcks);
599 	} else {
600 		struct sockaddr_in dst = {
601 			.sin_len = sizeof(struct sockaddr_in),
602 			.sin_family = AF_INET,
603 			.sin_addr = inp->inp_faddr,
604 		};
605 		mptcp_version_cache_key_src_init(SA(&dst), &tcks);
606 	}
607 
608 	/* Call lookup/create function */
609 	tpcache = tcp_getcache_with_lock(&tcks, 1, &head);
610 	if (tpcache == NULL) {
611 		return;
612 	}
613 
614 	/* We are still in probing phase */
615 	if (tpcache->tc_mptcp_version_confirmed) {
616 		goto exit;
617 	}
618 
619 	if (succeeded) {
620 		if (version == (uint8_t)mptcp_preferred_version) {
621 			/* Preferred version succeeded - make it sticky */
622 			tpcache->tc_mptcp_version_confirmed = true;
623 			tpcache->tc_mptcp_version = version;
624 		} else {
625 			/* If we are past the next version try, set it
626 			 * so that we try preferred again in 24h
627 			 */
628 			if (TSTMP_GEQ(tcp_now, tpcache->tc_mptcp_next_version_try)) {
629 				tpcache->tc_mptcp_next_version_try = tcp_now + tcp_min_to_hz(mptcp_version_timeout);
630 			}
631 		}
632 	} else {
633 		if (version == (uint8_t)mptcp_preferred_version) {
634 			/* Preferred version failed - try the other version */
635 			tpcache->tc_mptcp_version = version == MPTCP_VERSION_0 ? MPTCP_VERSION_1 : MPTCP_VERSION_0;
636 		}
637 		/* Preferred version failed - make sure we give the preferred another
638 		 * shot in 24h.
639 		 */
640 		if (TSTMP_GEQ(tcp_now, tpcache->tc_mptcp_next_version_try)) {
641 			tpcache->tc_mptcp_next_version_try = tcp_now + tcp_min_to_hz(mptcp_version_timeout);
642 		}
643 	}
644 
645 exit:
646 	tcp_cache_unlock(head);
647 }
648 
649 static uint16_t
tcp_heuristics_hash(struct tcp_cache_key_src * tcks,struct tcp_heuristic_key * key)650 tcp_heuristics_hash(struct tcp_cache_key_src *tcks, struct tcp_heuristic_key *key)
651 {
652 	uint32_t hash;
653 
654 	bzero(key, sizeof(struct tcp_heuristic_key));
655 
656 	tcp_cache_hash_src(tcks, key);
657 
658 	hash = net_flowhash(key, sizeof(struct tcp_heuristic_key),
659 	    tcp_cache_hash_seed);
660 
661 	return (uint16_t)(hash & (tcp_cache_size - 1));
662 }
663 
664 static void
tcp_heuristic_unlock(struct tcp_heuristics_head * head)665 tcp_heuristic_unlock(struct tcp_heuristics_head *head)
666 {
667 	lck_mtx_unlock(&head->thh_mtx);
668 }
669 
670 /*
671  * Make sure that everything that happens after tcp_getheuristic_with_lock()
672  * is short enough to justify that you hold the per-bucket lock!!!
673  *
674  * Otherwise, better build another lookup-function that does not hold the
675  * lock and you copy out the bits and bytes.
676  *
677  * That's why we provide the head as a "return"-pointer so that the caller
678  * can give it back to use for tcp_heur_unlock().
679  *
680  *
681  * ToDo - way too much code-duplication. We should create an interface to handle
682  * bucketized hashtables with recycling of the oldest element.
683  */
684 static struct tcp_heuristic *
tcp_getheuristic_with_lock(struct tcp_cache_key_src * tcks,int create,struct tcp_heuristics_head ** headarg)685 tcp_getheuristic_with_lock(struct tcp_cache_key_src *tcks,
686     int create, struct tcp_heuristics_head **headarg)
687 {
688 	struct tcp_heuristic *__single tpheur = NULL;
689 	struct tcp_heuristics_head *__single head;
690 	struct tcp_heuristic_key key;
691 	uint16_t hash;
692 	int i = 0;
693 
694 	hash = tcp_heuristics_hash(tcks, &key);
695 	head = &tcp_heuristics[hash];
696 
697 	lck_mtx_lock(&head->thh_mtx);
698 
699 	/*** First step: Look for the tcp_heur in our bucket ***/
700 	SLIST_FOREACH(tpheur, &head->tcp_heuristics, list) {
701 		if (memcmp(&tpheur->th_key, &key, sizeof(key)) == 0) {
702 			break;
703 		}
704 
705 		i++;
706 	}
707 
708 	/*** Second step: If it's not there, create/recycle it ***/
709 	if ((tpheur == NULL) && create) {
710 		if (i >= TCP_CACHE_BUCKET_SIZE) {
711 			struct tcp_heuristic *__single oldest_heur = NULL;
712 			uint32_t max_age = 0;
713 
714 			/* Look for the oldest tcp_heur in the bucket */
715 			SLIST_FOREACH(tpheur, &head->tcp_heuristics, list) {
716 				uint32_t age = tcp_now - tpheur->th_last_access;
717 				if (age >= max_age) {
718 					max_age = age;
719 					oldest_heur = tpheur;
720 				}
721 			}
722 			VERIFY(oldest_heur != NULL);
723 
724 			tpheur = oldest_heur;
725 
726 			/* We recycle - set everything to 0 */
727 			uint8_t *ptr = (uint8_t *)(struct tcp_heuristic *__indexable)tpheur;
728 			const size_t preamble = offsetof(struct tcp_heuristic, th_val_start);
729 			const size_t size = sizeof(struct tcp_heuristic) - preamble;
730 			bzero(ptr + preamble, size);
731 		} else {
732 			/* Create a new heuristic and add it to the list */
733 			tpheur = kalloc_type(struct tcp_heuristic, Z_NOPAGEWAIT | Z_ZERO);
734 			if (tpheur == NULL) {
735 				os_log_error(OS_LOG_DEFAULT, "%s could not allocate heuristic", __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 *__single head;
773 	struct tcp_heuristic *__single 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 *__single head;
914 	struct tcp_heuristic *__single 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 *__single head;
931 	struct tcp_heuristic *__single 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 *__single head;
1178 	struct tcp_heuristic *__single 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 *__single head;
1222 	struct tcp_heuristic *__single 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 *__single head = NULL;
1276 	struct tcp_heuristic *__single 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 *__single head;
1334 	struct tcp_heuristic *__single 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 * __counted_by (maxlen)cookie,uint8_t maxlen,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 *__counted_by(maxlen) cookie, uint8_t maxlen, 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, maxlen, 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 #if (DEVELOPMENT || DEBUG)
1529 /*
1530  * This test sysctl forces the hash table to be full which will force us to
1531  * erase portions of it.
1532  */
1533 static int
1534 sysctl_fill_hashtable SYSCTL_HANDLER_ARGS
1535 {
1536 #pragma unused(arg1, arg2)
1537 	int error = 0, val;
1538 
1539 	val = 0;
1540 	error = sysctl_handle_int(oidp, &val, 0, req);
1541 	if (error || !req->newptr) {
1542 		return error;
1543 	}
1544 	if (val == 1) {
1545 		struct necp_tcp_tfo_cache necp_buffer = {};
1546 		union sockaddr_in_4_6 local_address = {}, remote_address = {};
1547 
1548 		necp_buffer.necp_tcp_tfo_heuristics_success = 1;
1549 		necp_buffer.necp_tcp_tfo_heuristics_loss = 1;
1550 		necp_buffer.necp_tcp_tfo_heuristics_middlebox = 1;
1551 
1552 		for (unsigned i = 0; i < 1024; i++) {
1553 			local_address.sin.sin_family = AF_INET;
1554 			local_address.sin.sin_len = sizeof(struct sockaddr_in);
1555 			local_address.sin.sin_port = random() % UINT16_MAX;
1556 			local_address.sin.sin_addr.s_addr = random();
1557 
1558 			remote_address.sin.sin_family = AF_INET;
1559 			remote_address.sin.sin_len = sizeof(struct sockaddr_in);
1560 			remote_address.sin.sin_port = random() % UINT16_MAX;
1561 			remote_address.sin.sin_addr.s_addr = random();
1562 
1563 			tcp_heuristics_tfo_update(&necp_buffer, lo_ifp,
1564 			    &local_address,
1565 			    &remote_address);
1566 		}
1567 	}
1568 
1569 	return error;
1570 }
1571 
1572 static int fill_hash_table = 0;
1573 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, test_cache, CTLTYPE_INT | CTLFLAG_RW |
1574     CTLFLAG_LOCKED, &fill_hash_table, 0, &sysctl_fill_hashtable, "I",
1575     "Tests the hash table erasing procedures");
1576 #endif /* DEVELOPMENT || DEBUG */
1577 
1578 static void
sysctl_cleartfocache(void)1579 sysctl_cleartfocache(void)
1580 {
1581 	int i;
1582 
1583 	for (i = 0; i < tcp_cache_size; i++) {
1584 		struct tcp_cache_head *__single head = &tcp_cache[i];
1585 		struct tcp_cache *__single tpcache, *__single tmp;
1586 		struct tcp_heuristics_head *__single hhead = &tcp_heuristics[i];
1587 		struct tcp_heuristic *__single tpheur, *__single htmp;
1588 
1589 		lck_mtx_lock(&head->tch_mtx);
1590 		SLIST_FOREACH_SAFE(tpcache, &head->tcp_caches, list, tmp) {
1591 			SLIST_REMOVE(&head->tcp_caches, tpcache, tcp_cache, list);
1592 			kfree_type(struct tcp_cache, tpcache);
1593 		}
1594 		lck_mtx_unlock(&head->tch_mtx);
1595 
1596 		lck_mtx_lock(&hhead->thh_mtx);
1597 		SLIST_FOREACH_SAFE(tpheur, &hhead->tcp_heuristics, list, htmp) {
1598 			SLIST_REMOVE(&hhead->tcp_heuristics, tpheur, tcp_heuristic, list);
1599 			kfree_type(struct tcp_heuristic, tpheur);
1600 		}
1601 		lck_mtx_unlock(&hhead->thh_mtx);
1602 	}
1603 }
1604 
1605 /* This sysctl is useful for testing purposes only */
1606 static int tcpcleartfo = 0;
1607 
1608 static int sysctl_cleartfo SYSCTL_HANDLER_ARGS
1609 {
1610 #pragma unused(arg1, arg2)
1611 	int error = 0, val, oldval = tcpcleartfo;
1612 
1613 	val = oldval;
1614 	error = sysctl_handle_int(oidp, &val, 0, req);
1615 	if (error || !req->newptr) {
1616 		if (error) {
1617 			os_log_error(OS_LOG_DEFAULT, "%s could not parse int: %d", __func__, error);
1618 		}
1619 		return error;
1620 	}
1621 
1622 	/*
1623 	 * The actual value does not matter. If the value is set, it triggers
1624 	 * the clearing of the TFO cache. If a future implementation does not
1625 	 * use the route entry to hold the TFO cache, replace the route sysctl.
1626 	 */
1627 
1628 	if (val != oldval) {
1629 		sysctl_cleartfocache();
1630 	}
1631 
1632 	tcpcleartfo = val;
1633 
1634 	return error;
1635 }
1636 
1637 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, clear_tfocache, CTLTYPE_INT | CTLFLAG_RW |
1638     CTLFLAG_LOCKED, &tcpcleartfo, 0, &sysctl_cleartfo, "I",
1639     "Toggle to clear the TFO destination based heuristic cache");
1640 
1641 void
tcp_cache_init(void)1642 tcp_cache_init(void)
1643 {
1644 	uint64_t sane_size_meg = sane_size / 1024 / 1024;
1645 	size_t cache_size;
1646 	/*
1647 	 * On machines with <100MB of memory this will result in a (full) cache-size
1648 	 * of 32 entries, thus 32 * 5 * 64bytes = 10KB. (about 0.01 %)
1649 	 * On machines with > 4GB of memory, we have a cache-size of 1024 entries,
1650 	 * thus about 327KB.
1651 	 *
1652 	 * Side-note: we convert to uint32_t. If sane_size is more than
1653 	 * 16000 TB, we loose precision. But, who cares? :)
1654 	 */
1655 	cache_size = tcp_cache_roundup2((uint32_t)(sane_size_meg >> 2));
1656 	if (cache_size < 32) {
1657 		cache_size = 32;
1658 	} else if (cache_size > 1024) {
1659 		cache_size = 1024;
1660 	}
1661 
1662 	tcp_cache = zalloc_permanent(sizeof(struct tcp_cache_head) * cache_size,
1663 	    ZALIGN(struct tcp_cache_head));
1664 	tcp_cache_size = cache_size;
1665 	tcp_heuristics = zalloc_permanent(sizeof(struct tcp_heuristics_head) * cache_size,
1666 	    ZALIGN(struct tcp_heuristics_head));
1667 	tcp_heuristics_size = cache_size;
1668 
1669 	for (int i = 0; i < tcp_cache_size; i++) {
1670 		lck_mtx_init(&tcp_cache[i].tch_mtx, &tcp_cache_mtx_grp,
1671 		    &tcp_cache_mtx_attr);
1672 		SLIST_INIT(&tcp_cache[i].tcp_caches);
1673 
1674 		lck_mtx_init(&tcp_heuristics[i].thh_mtx, &tcp_heuristic_mtx_grp,
1675 		    &tcp_heuristic_mtx_attr);
1676 		SLIST_INIT(&tcp_heuristics[i].tcp_heuristics);
1677 	}
1678 
1679 	tcp_cache_hash_seed = RandomULong();
1680 }
1681