xref: /freebsd/contrib/bearssl/src/ssl/ssl_lru.c (revision 5ca8e32633c4ffbbcd6762e5888b6a4ba0708c6c)
1 /*
2  * Copyright (c) 2016 Thomas Pornin <pornin@bolet.org>
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining
5  * a copy of this software and associated documentation files (the
6  * "Software"), to deal in the Software without restriction, including
7  * without limitation the rights to use, copy, modify, merge, publish,
8  * distribute, sublicense, and/or sell copies of the Software, and to
9  * permit persons to whom the Software is furnished to do so, subject to
10  * the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be
13  * included in all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
16  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
17  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
18  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
19  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
20  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
21  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 
25 #include "inner.h"
26 
27 /*
28  * Each entry consists in a fixed number of bytes. Entries are concatenated
29  * in the store block. "Addresses" are really offsets in the block,
30  * expressed over 32 bits (so the cache may have size at most 4 GB, which
31  * "ought to be enough for everyone"). The "null address" is 0xFFFFFFFF.
32  * Note that since the storage block alignment is in no way guaranteed, we
33  * perform only accesses that can handle unaligned data.
34  *
35  * Two concurrent data structures are maintained:
36  *
37  * -- Entries are organised in a doubly-linked list; saved entries are added
38  * at the head, and loaded entries are moved to the head. Eviction uses
39  * the list tail (this is the LRU algorithm).
40  *
41  * -- Entries are indexed with a binary tree: all left descendants of a
42  * node have a lower session ID (in lexicographic order), while all
43  * right descendants have a higher session ID. The tree is heuristically
44  * balanced.
45  *
46  * Entry format:
47  *
48  *   session ID          32 bytes
49  *   master secret       48 bytes
50  *   protocol version    2 bytes (big endian)
51  *   cipher suite        2 bytes (big endian)
52  *   list prev           4 bytes (big endian)
53  *   list next           4 bytes (big endian)
54  *   tree left child     4 bytes (big endian)
55  *   tree right child    4 bytes (big endian)
56  *
57  * If an entry has a protocol version set to 0, then it is "disabled":
58  * it was a session pushed to the cache at some point, but it has
59  * been explicitly removed.
60  *
61  * We need to keep the tree balanced because an attacker could make
62  * handshakes, selecting some specific sessions (by reusing them) to
63  * try to make us make an imbalanced tree that makes lookups expensive
64  * (a denial-of-service attack that would persist as long as the cache
65  * remains, i.e. even after the attacker made all his connections).
66  * To do that, we replace the session ID (or the start of the session ID)
67  * with a HMAC value computed over the replaced part; the hash function
68  * implementation and the key are obtained from the server context upon
69  * first save() call.
70  *
71  * Theoretically, an attacker could use the exact timing of the lookup
72  * to infer the current tree topology, and try to revive entries to make
73  * it as unbalanced as possible. However, since the session ID are
74  * chosen randomly by the server, and the attacker cannot see the
75  * indexing values and must thus rely on blind selection, it should be
76  * exponentially difficult for the attacker to maintain a large
77  * imbalance.
78  */
79 #define SESSION_ID_LEN       32
80 #define MASTER_SECRET_LEN    48
81 
82 #define SESSION_ID_OFF        0
83 #define MASTER_SECRET_OFF    32
84 #define VERSION_OFF          80
85 #define CIPHER_SUITE_OFF     82
86 #define LIST_PREV_OFF        84
87 #define LIST_NEXT_OFF        88
88 #define TREE_LEFT_OFF        92
89 #define TREE_RIGHT_OFF       96
90 
91 #define LRU_ENTRY_LEN       100
92 
93 #define ADDR_NULL   ((uint32_t)-1)
94 
95 #define GETSET(name, off) \
96 static inline uint32_t get_ ## name(br_ssl_session_cache_lru *cc, uint32_t x) \
97 { \
98 	return br_dec32be(cc->store + x + (off)); \
99 } \
100 static inline void set_ ## name(br_ssl_session_cache_lru *cc, \
101 	uint32_t x, uint32_t val) \
102 { \
103 	br_enc32be(cc->store + x + (off), val); \
104 }
105 
106 GETSET(prev, LIST_PREV_OFF)
107 GETSET(next, LIST_NEXT_OFF)
108 GETSET(left, TREE_LEFT_OFF)
109 GETSET(right, TREE_RIGHT_OFF)
110 
111 /*
112  * Transform the session ID by replacing the first N bytes with a HMAC
113  * value computed over these bytes, using the random key K (the HMAC
114  * value is truncated if needed). HMAC will use the same hash function
115  * as the DRBG in the SSL server context, so with SHA-256, SHA-384,
116  * or SHA-1, depending on what is available.
117  *
118  * The risk of collision is considered too small to be a concern; and
119  * the impact of a collision is low (the handshake won't succeed). This
120  * risk is much lower than any transmission error, which would lead to
121  * the same consequences.
122  *
123  * Source and destination arrays msut be disjoint.
124  */
125 static void
126 mask_id(br_ssl_session_cache_lru *cc,
127 	const unsigned char *src, unsigned char *dst)
128 {
129 	br_hmac_key_context hkc;
130 	br_hmac_context hc;
131 
132 	memcpy(dst, src, SESSION_ID_LEN);
133 	br_hmac_key_init(&hkc, cc->hash, cc->index_key, sizeof cc->index_key);
134 	br_hmac_init(&hc, &hkc, SESSION_ID_LEN);
135 	br_hmac_update(&hc, src, SESSION_ID_LEN);
136 	br_hmac_out(&hc, dst);
137 }
138 
139 /*
140  * Find a node by ID. Returned value is the node address, or ADDR_NULL if
141  * the node is not found.
142  *
143  * If addr_link is not NULL, then '*addr_link' is set to the address of the
144  * last followed link. If the found node is the root, or if the tree is
145  * empty, then '*addr_link' is set to ADDR_NULL.
146  */
147 static uint32_t
148 find_node(br_ssl_session_cache_lru *cc, const unsigned char *id,
149 	uint32_t *addr_link)
150 {
151 	uint32_t x, y;
152 
153 	x = cc->root;
154 	y = ADDR_NULL;
155 	while (x != ADDR_NULL) {
156 		int r;
157 
158 		r = memcmp(id, cc->store + x + SESSION_ID_OFF, SESSION_ID_LEN);
159 		if (r < 0) {
160 			y = x + TREE_LEFT_OFF;
161 			x = get_left(cc, x);
162 		} else if (r == 0) {
163 			if (addr_link != NULL) {
164 				*addr_link = y;
165 			}
166 			return x;
167 		} else {
168 			y = x + TREE_RIGHT_OFF;
169 			x = get_right(cc, x);
170 		}
171 	}
172 	if (addr_link != NULL) {
173 		*addr_link = y;
174 	}
175 	return ADDR_NULL;
176 }
177 
178 /*
179  * For node x, find its replacement upon removal.
180  *
181  *  -- If node x has no child, then this returns ADDR_NULL.
182  *  -- Otherwise, if node x has a left child, then the replacement is the
183  *     rightmost left-descendent.
184  *  -- Otherwise, the replacement is the leftmost right-descendent.
185  *
186  * If a node is returned, then '*al' is set to the address of the field
187  * that points to that node. Otherwise (node x has no child), '*al' is
188  * set to ADDR_NULL.
189  *
190  * Note that the replacement node, when found, is always a descendent
191  * of node 'x', so it cannot be the tree root. Thus, '*al' can be set
192  * to ADDR_NULL only when no node is found and ADDR_NULL is returned.
193  */
194 static uint32_t
195 find_replacement_node(br_ssl_session_cache_lru *cc, uint32_t x, uint32_t *al)
196 {
197 	uint32_t y1, y2;
198 
199 	y1 = get_left(cc, x);
200 	if (y1 != ADDR_NULL) {
201 		y2 = x + TREE_LEFT_OFF;
202 		for (;;) {
203 			uint32_t z;
204 
205 			z = get_right(cc, y1);
206 			if (z == ADDR_NULL) {
207 				*al = y2;
208 				return y1;
209 			}
210 			y2 = y1 + TREE_RIGHT_OFF;
211 			y1 = z;
212 		}
213 	}
214 	y1 = get_right(cc, x);
215 	if (y1 != ADDR_NULL) {
216 		y2 = x + TREE_RIGHT_OFF;
217 		for (;;) {
218 			uint32_t z;
219 
220 			z = get_left(cc, y1);
221 			if (z == ADDR_NULL) {
222 				*al = y2;
223 				return y1;
224 			}
225 			y2 = y1 + TREE_LEFT_OFF;
226 			y1 = z;
227 		}
228 	}
229 	*al = ADDR_NULL;
230 	return ADDR_NULL;
231 }
232 
233 /*
234  * Set the link at address 'alx' to point to node 'x'. If 'alx' is
235  * ADDR_NULL, then this sets the tree root to 'x'.
236  */
237 static inline void
238 set_link(br_ssl_session_cache_lru *cc, uint32_t alx, uint32_t x)
239 {
240 	if (alx == ADDR_NULL) {
241 		cc->root = x;
242 	} else {
243 		br_enc32be(cc->store + alx, x);
244 	}
245 }
246 
247 /*
248  * Remove node 'x' from the tree. This function shall not be called if
249  * node 'x' is not part of the tree.
250  */
251 static void
252 remove_node(br_ssl_session_cache_lru *cc, uint32_t x)
253 {
254 	uint32_t alx, y, aly;
255 
256 	/*
257 	 * Removal algorithm:
258 	 * ------------------
259 	 *
260 	 * - If we remove the root, then the tree becomes empty.
261 	 *
262 	 * - If the removed node has no child, then we can simply remove
263 	 *   it, with nothing else to do.
264 	 *
265 	 * - Otherwise, the removed node must be replaced by either its
266 	 *   rightmost left-descendent, or its leftmost right-descendent.
267 	 *   The replacement node itself must be removed from its current
268 	 *   place. By definition, that replacement node has either no
269 	 *   child, or at most a single child that will replace it in the
270 	 *   tree.
271 	 */
272 
273 	/*
274 	 * Find node back and its ancestor link. If the node was the
275 	 * root, then alx is set to ADDR_NULL.
276 	 */
277 	find_node(cc, cc->store + x + SESSION_ID_OFF, &alx);
278 
279 	/*
280 	 * Find replacement node 'y', and 'aly' is set to the address of
281 	 * the link to that replacement node. If the removed node has no
282 	 * child, then both 'y' and 'aly' are set to ADDR_NULL.
283 	 */
284 	y = find_replacement_node(cc, x, &aly);
285 
286 	if (y != ADDR_NULL) {
287 		uint32_t z;
288 
289 		/*
290 		 * The unlinked replacement node may have one child (but
291 		 * not two) that takes its place.
292 		 */
293 		z = get_left(cc, y);
294 		if (z == ADDR_NULL) {
295 			z = get_right(cc, y);
296 		}
297 		set_link(cc, aly, z);
298 
299 		/*
300 		 * Link the replacement node in its new place, overwriting
301 		 * the current link to the node 'x' (which removes 'x').
302 		 */
303 		set_link(cc, alx, y);
304 
305 		/*
306 		 * The replacement node adopts the left and right children
307 		 * of the removed node. Note that this also works even if
308 		 * the replacement node was a direct descendent of the
309 		 * removed node, since we unlinked it previously.
310 		 */
311 		set_left(cc, y, get_left(cc, x));
312 		set_right(cc, y, get_right(cc, x));
313 	} else {
314 		/*
315 		 * No replacement, we simply unlink the node 'x'.
316 		 */
317 		set_link(cc, alx, ADDR_NULL);
318 	}
319 }
320 
321 static void
322 lru_save(const br_ssl_session_cache_class **ctx,
323 	br_ssl_server_context *server_ctx,
324 	const br_ssl_session_parameters *params)
325 {
326 	br_ssl_session_cache_lru *cc;
327 	unsigned char id[SESSION_ID_LEN];
328 	uint32_t x, alx;
329 
330 	cc = (br_ssl_session_cache_lru *)ctx;
331 
332 	/*
333 	 * If the buffer is too small, we don't record anything. This
334 	 * test avoids problems in subsequent code.
335 	 */
336 	if (cc->store_len < LRU_ENTRY_LEN) {
337 		return;
338 	}
339 
340 	/*
341 	 * Upon the first save in a session cache instance, we obtain
342 	 * a random key for our indexing.
343 	 */
344 	if (!cc->init_done) {
345 		br_hmac_drbg_generate(&server_ctx->eng.rng,
346 			cc->index_key, sizeof cc->index_key);
347 		cc->hash = br_hmac_drbg_get_hash(&server_ctx->eng.rng);
348 		cc->init_done = 1;
349 	}
350 	mask_id(cc, params->session_id, id);
351 
352 	/*
353 	 * Look for the node in the tree. If the same ID is already used,
354 	 * then reject it. This is a collision event, which should be
355 	 * exceedingly rare.
356 	 * Note: we do NOT record the emplacement here, because the
357 	 * removal of an entry may change the tree topology.
358 	 */
359 	if (find_node(cc, id, NULL) != ADDR_NULL) {
360 		return;
361 	}
362 
363 	/*
364 	 * Find some room for the new parameters. If the cache is not
365 	 * full yet, add it to the end of the area and bump the pointer up.
366 	 * Otherwise, evict the list tail entry. Note that we already
367 	 * filtered out the case of a ridiculously small buffer that
368 	 * cannot hold any entry at all; thus, if there is no room for an
369 	 * extra entry, then the cache cannot be empty.
370 	 */
371 	if (cc->store_ptr > (cc->store_len - LRU_ENTRY_LEN)) {
372 		/*
373 		 * Evict tail. If the buffer has room for a single entry,
374 		 * then this may also be the head.
375 		 */
376 		x = cc->tail;
377 		cc->tail = get_prev(cc, x);
378 		if (cc->tail == ADDR_NULL) {
379 			cc->head = ADDR_NULL;
380 		} else {
381 			set_next(cc, cc->tail, ADDR_NULL);
382 		}
383 
384 		/*
385 		 * Remove the node from the tree.
386 		 */
387 		remove_node(cc, x);
388 	} else {
389 		/*
390 		 * Allocate room for new node.
391 		 */
392 		x = cc->store_ptr;
393 		cc->store_ptr += LRU_ENTRY_LEN;
394 	}
395 
396 	/*
397 	 * Find the emplacement for the new node, and link it.
398 	 */
399 	find_node(cc, id, &alx);
400 	set_link(cc, alx, x);
401 	set_left(cc, x, ADDR_NULL);
402 	set_right(cc, x, ADDR_NULL);
403 
404 	/*
405 	 * New entry becomes new list head. It may also become the list
406 	 * tail if the cache was empty at that point.
407 	 */
408 	if (cc->head == ADDR_NULL) {
409 		cc->tail = x;
410 	} else {
411 		set_prev(cc, cc->head, x);
412 	}
413 	set_prev(cc, x, ADDR_NULL);
414 	set_next(cc, x, cc->head);
415 	cc->head = x;
416 
417 	/*
418 	 * Fill data in the entry.
419 	 */
420 	memcpy(cc->store + x + SESSION_ID_OFF, id, SESSION_ID_LEN);
421 	memcpy(cc->store + x + MASTER_SECRET_OFF,
422 		params->master_secret, MASTER_SECRET_LEN);
423 	br_enc16be(cc->store + x + VERSION_OFF, params->version);
424 	br_enc16be(cc->store + x + CIPHER_SUITE_OFF, params->cipher_suite);
425 }
426 
427 static int
428 lru_load(const br_ssl_session_cache_class **ctx,
429 	br_ssl_server_context *server_ctx,
430 	br_ssl_session_parameters *params)
431 {
432 	br_ssl_session_cache_lru *cc;
433 	unsigned char id[SESSION_ID_LEN];
434 	uint32_t x;
435 
436 	(void)server_ctx;
437 	cc = (br_ssl_session_cache_lru *)ctx;
438 	if (!cc->init_done) {
439 		return 0;
440 	}
441 	mask_id(cc, params->session_id, id);
442 	x = find_node(cc, id, NULL);
443 	if (x != ADDR_NULL) {
444 		unsigned version;
445 
446 		version = br_dec16be(cc->store + x + VERSION_OFF);
447 		if (version == 0) {
448 			/*
449 			 * Entry is disabled, we pretend we did not find it.
450 			 * Notably, we don't move it to the front of the
451 			 * LRU list.
452 			 */
453 			return 0;
454 		}
455 		params->version = version;
456 		params->cipher_suite = br_dec16be(
457 			cc->store + x + CIPHER_SUITE_OFF);
458 		memcpy(params->master_secret,
459 			cc->store + x + MASTER_SECRET_OFF,
460 			MASTER_SECRET_LEN);
461 		if (x != cc->head) {
462 			/*
463 			 * Found node is not at list head, so move
464 			 * it to the head.
465 			 */
466 			uint32_t p, n;
467 
468 			p = get_prev(cc, x);
469 			n = get_next(cc, x);
470 			set_next(cc, p, n);
471 			if (n == ADDR_NULL) {
472 				cc->tail = p;
473 			} else {
474 				set_prev(cc, n, p);
475 			}
476 			set_prev(cc, cc->head, x);
477 			set_next(cc, x, cc->head);
478 			set_prev(cc, x, ADDR_NULL);
479 			cc->head = x;
480 		}
481 		return 1;
482 	}
483 	return 0;
484 }
485 
486 static const br_ssl_session_cache_class lru_class = {
487 	sizeof(br_ssl_session_cache_lru),
488 	&lru_save,
489 	&lru_load
490 };
491 
492 /* see inner.h */
493 void
494 br_ssl_session_cache_lru_init(br_ssl_session_cache_lru *cc,
495 	unsigned char *store, size_t store_len)
496 {
497 	cc->vtable = &lru_class;
498 	cc->store = store;
499 	cc->store_len = store_len;
500 	cc->store_ptr = 0;
501 	cc->init_done = 0;
502 	cc->head = ADDR_NULL;
503 	cc->tail = ADDR_NULL;
504 	cc->root = ADDR_NULL;
505 }
506 
507 /* see bearssl_ssl.h */
508 void br_ssl_session_cache_lru_forget(
509 	br_ssl_session_cache_lru *cc, const unsigned char *id)
510 {
511 	unsigned char mid[SESSION_ID_LEN];
512 	uint32_t addr;
513 
514 	/*
515 	 * If the cache is not initialised yet, then it is empty, and
516 	 * there is nothing to forget.
517 	 */
518 	if (!cc->init_done) {
519 		return;
520 	}
521 
522 	/*
523 	 * Look for the node in the tree. If found, the entry is marked
524 	 * as "disabled"; it will be reused in due course, as it ages
525 	 * through the list.
526 	 *
527 	 * We do not go through the complex moves of actually releasing
528 	 * the entry right away because explicitly forgetting sessions
529 	 * should be a rare event, meant mostly for testing purposes,
530 	 * so this is not worth the extra code size.
531 	 */
532 	mask_id(cc, id, mid);
533 	addr = find_node(cc, mid, NULL);
534 	if (addr != ADDR_NULL) {
535 		br_enc16be(cc->store + addr + VERSION_OFF, 0);
536 	}
537 }
538