1 /* ====================================================================
2 * Copyright (c) 2012 The OpenSSL Project. All rights reserved.
3 *
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
6 * are met:
7 *
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 *
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in
13 * the documentation and/or other materials provided with the
14 * distribution.
15 *
16 * 3. All advertising materials mentioning features or use of this
17 * software must display the following acknowledgment:
18 * "This product includes software developed by the OpenSSL Project
19 * for use in the OpenSSL Toolkit. (http://www.openssl.org/)"
20 *
21 * 4. The names "OpenSSL Toolkit" and "OpenSSL Project" must not be used to
22 * endorse or promote products derived from this software without
23 * prior written permission. For written permission, please contact
24 * openssl-core@openssl.org.
25 *
26 * 5. Products derived from this software may not be called "OpenSSL"
27 * nor may "OpenSSL" appear in their names without prior written
28 * permission of the OpenSSL Project.
29 *
30 * 6. Redistributions of any form whatsoever must retain the following
31 * acknowledgment:
32 * "This product includes software developed by the OpenSSL Project
33 * for use in the OpenSSL Toolkit (http://www.openssl.org/)"
34 *
35 * THIS SOFTWARE IS PROVIDED BY THE OpenSSL PROJECT ``AS IS'' AND ANY
36 * EXPRESSED OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
37 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
38 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE OpenSSL PROJECT OR
39 * ITS CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
40 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
41 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
42 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
43 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
44 * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
45 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
46 * OF THE POSSIBILITY OF SUCH DAMAGE.
47 * ====================================================================
48 *
49 * This product includes cryptographic software written by Eric Young
50 * (eay@cryptsoft.com). This product includes software written by Tim
51 * Hudson (tjh@cryptsoft.com). */
52
53 #include <assert.h>
54 #include <string.h>
55
56 #include <openssl/digest.h>
57 #include <openssl/nid.h>
58 #include <openssl/sha.h>
59
60 #include "../internal.h"
61 #include "internal.h"
62
63
64 /* TODO(davidben): unsigned should be size_t. The various constant_time
65 * functions need to be switched to size_t. */
66
67 /* MAX_HASH_BIT_COUNT_BYTES is the maximum number of bytes in the hash's length
68 * field. (SHA-384/512 have 128-bit length.) */
69 #define MAX_HASH_BIT_COUNT_BYTES 16
70
71 /* MAX_HASH_BLOCK_SIZE is the maximum hash block size that we'll support.
72 * Currently SHA-384/512 has a 128-byte block size and that's the largest
73 * supported by TLS.) */
74 #define MAX_HASH_BLOCK_SIZE 128
75
EVP_tls_cbc_remove_padding(unsigned * out_padding_ok,unsigned * out_len,const uint8_t * in,unsigned in_len,unsigned block_size,unsigned mac_size)76 int EVP_tls_cbc_remove_padding(unsigned *out_padding_ok, unsigned *out_len,
77 const uint8_t *in, unsigned in_len,
78 unsigned block_size, unsigned mac_size) {
79 unsigned padding_length, good, to_check, i;
80 const unsigned overhead = 1 /* padding length byte */ + mac_size;
81
82 /* These lengths are all public so we can test them in non-constant time. */
83 if (overhead > in_len) {
84 return 0;
85 }
86
87 padding_length = in[in_len - 1];
88
89 good = constant_time_ge(in_len, overhead + padding_length);
90 /* The padding consists of a length byte at the end of the record and
91 * then that many bytes of padding, all with the same value as the
92 * length byte. Thus, with the length byte included, there are i+1
93 * bytes of padding.
94 *
95 * We can't check just |padding_length+1| bytes because that leaks
96 * decrypted information. Therefore we always have to check the maximum
97 * amount of padding possible. (Again, the length of the record is
98 * public information so we can use it.) */
99 to_check = 256; /* maximum amount of padding, inc length byte. */
100 if (to_check > in_len) {
101 to_check = in_len;
102 }
103
104 for (i = 0; i < to_check; i++) {
105 uint8_t mask = constant_time_ge_8(padding_length, i);
106 uint8_t b = in[in_len - 1 - i];
107 /* The final |padding_length+1| bytes should all have the value
108 * |padding_length|. Therefore the XOR should be zero. */
109 good &= ~(mask & (padding_length ^ b));
110 }
111
112 /* If any of the final |padding_length+1| bytes had the wrong value,
113 * one or more of the lower eight bits of |good| will be cleared. */
114 good = constant_time_eq(0xff, good & 0xff);
115
116 /* Always treat |padding_length| as zero on error. If, assuming block size of
117 * 16, a padding of [<15 arbitrary bytes> 15] treated |padding_length| as 16
118 * and returned -1, distinguishing good MAC and bad padding from bad MAC and
119 * bad padding would give POODLE's padding oracle. */
120 padding_length = good & (padding_length + 1);
121 *out_len = in_len - padding_length;
122 *out_padding_ok = good;
123 return 1;
124 }
125
EVP_tls_cbc_copy_mac(uint8_t * out,unsigned md_size,const uint8_t * in,unsigned in_len,unsigned orig_len)126 void EVP_tls_cbc_copy_mac(uint8_t *out, unsigned md_size,
127 const uint8_t *in, unsigned in_len,
128 unsigned orig_len) {
129 uint8_t rotated_mac1[EVP_MAX_MD_SIZE], rotated_mac2[EVP_MAX_MD_SIZE];
130 uint8_t *rotated_mac = rotated_mac1;
131 uint8_t *rotated_mac_tmp = rotated_mac2;
132
133 /* mac_end is the index of |in| just after the end of the MAC. */
134 unsigned mac_end = in_len;
135 unsigned mac_start = mac_end - md_size;
136
137 assert(orig_len >= in_len);
138 assert(in_len >= md_size);
139 assert(md_size <= EVP_MAX_MD_SIZE);
140
141 /* scan_start contains the number of bytes that we can ignore because
142 * the MAC's position can only vary by 255 bytes. */
143 unsigned scan_start = 0;
144 /* This information is public so it's safe to branch based on it. */
145 if (orig_len > md_size + 255 + 1) {
146 scan_start = orig_len - (md_size + 255 + 1);
147 }
148
149 unsigned rotate_offset = 0;
150 uint8_t mac_started = 0;
151 OPENSSL_memset(rotated_mac, 0, md_size);
152 for (unsigned i = scan_start, j = 0; i < orig_len; i++, j++) {
153 if (j >= md_size) {
154 j -= md_size;
155 }
156 unsigned is_mac_start = constant_time_eq(i, mac_start);
157 mac_started |= is_mac_start;
158 uint8_t mac_ended = constant_time_ge_8(i, mac_end);
159 rotated_mac[j] |= in[i] & mac_started & ~mac_ended;
160 /* Save the offset that |mac_start| is mapped to. */
161 rotate_offset |= j & is_mac_start;
162 }
163
164 /* Now rotate the MAC. We rotate in log(md_size) steps, one for each bit
165 * position. */
166 for (unsigned offset = 1; offset < md_size;
167 offset <<= 1, rotate_offset >>= 1) {
168 /* Rotate by |offset| iff the corresponding bit is set in
169 * |rotate_offset|, placing the result in |rotated_mac_tmp|. */
170 const uint8_t skip_rotate = (rotate_offset & 1) - 1;
171 for (unsigned i = 0, j = offset; i < md_size; i++, j++) {
172 if (j >= md_size) {
173 j -= md_size;
174 }
175 rotated_mac_tmp[i] =
176 constant_time_select_8(skip_rotate, rotated_mac[i], rotated_mac[j]);
177 }
178
179 /* Swap pointers so |rotated_mac| contains the (possibly) rotated value.
180 * Note the number of iterations and thus the identity of these pointers is
181 * public information. */
182 uint8_t *tmp = rotated_mac;
183 rotated_mac = rotated_mac_tmp;
184 rotated_mac_tmp = tmp;
185 }
186
187 OPENSSL_memcpy(out, rotated_mac, md_size);
188 }
189
190 /* u32toBE serialises an unsigned, 32-bit number (n) as four bytes at (p) in
191 * big-endian order. The value of p is advanced by four. */
192 #define u32toBE(n, p) \
193 do { \
194 *((p)++) = (uint8_t)((n) >> 24); \
195 *((p)++) = (uint8_t)((n) >> 16); \
196 *((p)++) = (uint8_t)((n) >> 8); \
197 *((p)++) = (uint8_t)((n)); \
198 } while (0)
199
200 /* u64toBE serialises an unsigned, 64-bit number (n) as eight bytes at (p) in
201 * big-endian order. The value of p is advanced by eight. */
202 #define u64toBE(n, p) \
203 do { \
204 *((p)++) = (uint8_t)((n) >> 56); \
205 *((p)++) = (uint8_t)((n) >> 48); \
206 *((p)++) = (uint8_t)((n) >> 40); \
207 *((p)++) = (uint8_t)((n) >> 32); \
208 *((p)++) = (uint8_t)((n) >> 24); \
209 *((p)++) = (uint8_t)((n) >> 16); \
210 *((p)++) = (uint8_t)((n) >> 8); \
211 *((p)++) = (uint8_t)((n)); \
212 } while (0)
213
214 /* These functions serialize the state of a hash and thus perform the standard
215 * "final" operation without adding the padding and length that such a function
216 * typically does. */
tls1_sha1_final_raw(void * ctx,uint8_t * md_out)217 static void tls1_sha1_final_raw(void *ctx, uint8_t *md_out) {
218 SHA_CTX *sha1 = ctx;
219 u32toBE(sha1->h[0], md_out);
220 u32toBE(sha1->h[1], md_out);
221 u32toBE(sha1->h[2], md_out);
222 u32toBE(sha1->h[3], md_out);
223 u32toBE(sha1->h[4], md_out);
224 }
225 #define LARGEST_DIGEST_CTX SHA_CTX
226
tls1_sha256_final_raw(void * ctx,uint8_t * md_out)227 static void tls1_sha256_final_raw(void *ctx, uint8_t *md_out) {
228 SHA256_CTX *sha256 = ctx;
229 unsigned i;
230
231 for (i = 0; i < 8; i++) {
232 u32toBE(sha256->h[i], md_out);
233 }
234 }
235 #undef LARGEST_DIGEST_CTX
236 #define LARGEST_DIGEST_CTX SHA256_CTX
237
tls1_sha512_final_raw(void * ctx,uint8_t * md_out)238 static void tls1_sha512_final_raw(void *ctx, uint8_t *md_out) {
239 SHA512_CTX *sha512 = ctx;
240 unsigned i;
241
242 for (i = 0; i < 8; i++) {
243 u64toBE(sha512->h[i], md_out);
244 }
245 }
246 #undef LARGEST_DIGEST_CTX
247 #define LARGEST_DIGEST_CTX SHA512_CTX
248
EVP_tls_cbc_record_digest_supported(const EVP_MD * md)249 int EVP_tls_cbc_record_digest_supported(const EVP_MD *md) {
250 switch (EVP_MD_type(md)) {
251 case NID_sha1:
252 case NID_sha256:
253 case NID_sha384:
254 return 1;
255
256 default:
257 return 0;
258 }
259 }
260
EVP_tls_cbc_digest_record(const EVP_MD * md,uint8_t * md_out,size_t * md_out_size,const uint8_t header[13],const uint8_t * data,size_t data_plus_mac_size,size_t data_plus_mac_plus_padding_size,const uint8_t * mac_secret,unsigned mac_secret_length)261 int EVP_tls_cbc_digest_record(const EVP_MD *md, uint8_t *md_out,
262 size_t *md_out_size, const uint8_t header[13],
263 const uint8_t *data, size_t data_plus_mac_size,
264 size_t data_plus_mac_plus_padding_size,
265 const uint8_t *mac_secret,
266 unsigned mac_secret_length) {
267 union {
268 double align;
269 uint8_t c[sizeof(LARGEST_DIGEST_CTX)];
270 } md_state;
271 void (*md_final_raw)(void *ctx, uint8_t *md_out);
272 void (*md_transform)(void *ctx, const uint8_t *block);
273 unsigned md_size, md_block_size = 64;
274 unsigned len, max_mac_bytes, num_blocks, num_starting_blocks, k,
275 mac_end_offset, c, index_a, index_b;
276 unsigned int bits; /* at most 18 bits */
277 uint8_t length_bytes[MAX_HASH_BIT_COUNT_BYTES];
278 /* hmac_pad is the masked HMAC key. */
279 uint8_t hmac_pad[MAX_HASH_BLOCK_SIZE];
280 uint8_t first_block[MAX_HASH_BLOCK_SIZE];
281 uint8_t mac_out[EVP_MAX_MD_SIZE];
282 unsigned i, j, md_out_size_u;
283 EVP_MD_CTX md_ctx;
284 /* mdLengthSize is the number of bytes in the length field that terminates
285 * the hash. */
286 unsigned md_length_size = 8;
287
288 /* This is a, hopefully redundant, check that allows us to forget about
289 * many possible overflows later in this function. */
290 assert(data_plus_mac_plus_padding_size < 1024 * 1024);
291
292 switch (EVP_MD_type(md)) {
293 case NID_sha1:
294 SHA1_Init((SHA_CTX *)md_state.c);
295 md_final_raw = tls1_sha1_final_raw;
296 md_transform =
297 (void (*)(void *ctx, const uint8_t *block))SHA1_Transform;
298 md_size = 20;
299 break;
300
301 case NID_sha256:
302 SHA256_Init((SHA256_CTX *)md_state.c);
303 md_final_raw = tls1_sha256_final_raw;
304 md_transform =
305 (void (*)(void *ctx, const uint8_t *block))SHA256_Transform;
306 md_size = 32;
307 break;
308
309 case NID_sha384:
310 SHA384_Init((SHA512_CTX *)md_state.c);
311 md_final_raw = tls1_sha512_final_raw;
312 md_transform =
313 (void (*)(void *ctx, const uint8_t *block))SHA512_Transform;
314 md_size = 384 / 8;
315 md_block_size = 128;
316 md_length_size = 16;
317 break;
318
319 default:
320 /* EVP_tls_cbc_record_digest_supported should have been called first to
321 * check that the hash function is supported. */
322 assert(0);
323 *md_out_size = 0;
324 return 0;
325 }
326
327 assert(md_length_size <= MAX_HASH_BIT_COUNT_BYTES);
328 assert(md_block_size <= MAX_HASH_BLOCK_SIZE);
329 assert(md_size <= EVP_MAX_MD_SIZE);
330
331 static const unsigned kHeaderLength = 13;
332
333 /* kVarianceBlocks is the number of blocks of the hash that we have to
334 * calculate in constant time because they could be altered by the
335 * padding value.
336 *
337 * TLSv1 has MACs up to 48 bytes long (SHA-384) and the padding is not
338 * required to be minimal. Therefore we say that the final six blocks
339 * can vary based on the padding. */
340 static const unsigned kVarianceBlocks = 6;
341
342 /* From now on we're dealing with the MAC, which conceptually has 13
343 * bytes of `header' before the start of the data. */
344 len = data_plus_mac_plus_padding_size + kHeaderLength;
345 /* max_mac_bytes contains the maximum bytes of bytes in the MAC, including
346 * |header|, assuming that there's no padding. */
347 max_mac_bytes = len - md_size - 1;
348 /* num_blocks is the maximum number of hash blocks. */
349 num_blocks =
350 (max_mac_bytes + 1 + md_length_size + md_block_size - 1) / md_block_size;
351 /* In order to calculate the MAC in constant time we have to handle
352 * the final blocks specially because the padding value could cause the
353 * end to appear somewhere in the final |kVarianceBlocks| blocks and we
354 * can't leak where. However, |num_starting_blocks| worth of data can
355 * be hashed right away because no padding value can affect whether
356 * they are plaintext. */
357 num_starting_blocks = 0;
358 /* k is the starting byte offset into the conceptual header||data where
359 * we start processing. */
360 k = 0;
361 /* mac_end_offset is the index just past the end of the data to be
362 * MACed. */
363 mac_end_offset = data_plus_mac_size + kHeaderLength - md_size;
364 /* c is the index of the 0x80 byte in the final hash block that
365 * contains application data. */
366 c = mac_end_offset % md_block_size;
367 /* index_a is the hash block number that contains the 0x80 terminating
368 * value. */
369 index_a = mac_end_offset / md_block_size;
370 /* index_b is the hash block number that contains the 64-bit hash
371 * length, in bits. */
372 index_b = (mac_end_offset + md_length_size) / md_block_size;
373 /* bits is the hash-length in bits. It includes the additional hash
374 * block for the masked HMAC key. */
375
376 if (num_blocks > kVarianceBlocks) {
377 num_starting_blocks = num_blocks - kVarianceBlocks;
378 k = md_block_size * num_starting_blocks;
379 }
380
381 bits = 8 * mac_end_offset;
382
383 /* Compute the initial HMAC block. */
384 bits += 8 * md_block_size;
385 OPENSSL_memset(hmac_pad, 0, md_block_size);
386 assert(mac_secret_length <= sizeof(hmac_pad));
387 OPENSSL_memcpy(hmac_pad, mac_secret, mac_secret_length);
388 for (i = 0; i < md_block_size; i++) {
389 hmac_pad[i] ^= 0x36;
390 }
391
392 md_transform(md_state.c, hmac_pad);
393
394 OPENSSL_memset(length_bytes, 0, md_length_size - 4);
395 length_bytes[md_length_size - 4] = (uint8_t)(bits >> 24);
396 length_bytes[md_length_size - 3] = (uint8_t)(bits >> 16);
397 length_bytes[md_length_size - 2] = (uint8_t)(bits >> 8);
398 length_bytes[md_length_size - 1] = (uint8_t)bits;
399
400 if (k > 0) {
401 /* k is a multiple of md_block_size. */
402 OPENSSL_memcpy(first_block, header, 13);
403 OPENSSL_memcpy(first_block + 13, data, md_block_size - 13);
404 md_transform(md_state.c, first_block);
405 for (i = 1; i < k / md_block_size; i++) {
406 md_transform(md_state.c, data + md_block_size * i - 13);
407 }
408 }
409
410 OPENSSL_memset(mac_out, 0, sizeof(mac_out));
411
412 /* We now process the final hash blocks. For each block, we construct
413 * it in constant time. If the |i==index_a| then we'll include the 0x80
414 * bytes and zero pad etc. For each block we selectively copy it, in
415 * constant time, to |mac_out|. */
416 for (i = num_starting_blocks; i <= num_starting_blocks + kVarianceBlocks;
417 i++) {
418 uint8_t block[MAX_HASH_BLOCK_SIZE];
419 uint8_t is_block_a = constant_time_eq_8(i, index_a);
420 uint8_t is_block_b = constant_time_eq_8(i, index_b);
421 for (j = 0; j < md_block_size; j++) {
422 uint8_t b = 0, is_past_c, is_past_cp1;
423 if (k < kHeaderLength) {
424 b = header[k];
425 } else if (k < data_plus_mac_plus_padding_size + kHeaderLength) {
426 b = data[k - kHeaderLength];
427 }
428 k++;
429
430 is_past_c = is_block_a & constant_time_ge_8(j, c);
431 is_past_cp1 = is_block_a & constant_time_ge_8(j, c + 1);
432 /* If this is the block containing the end of the
433 * application data, and we are at the offset for the
434 * 0x80 value, then overwrite b with 0x80. */
435 b = constant_time_select_8(is_past_c, 0x80, b);
436 /* If this the the block containing the end of the
437 * application data and we're past the 0x80 value then
438 * just write zero. */
439 b = b & ~is_past_cp1;
440 /* If this is index_b (the final block), but not
441 * index_a (the end of the data), then the 64-bit
442 * length didn't fit into index_a and we're having to
443 * add an extra block of zeros. */
444 b &= ~is_block_b | is_block_a;
445
446 /* The final bytes of one of the blocks contains the
447 * length. */
448 if (j >= md_block_size - md_length_size) {
449 /* If this is index_b, write a length byte. */
450 b = constant_time_select_8(
451 is_block_b, length_bytes[j - (md_block_size - md_length_size)], b);
452 }
453 block[j] = b;
454 }
455
456 md_transform(md_state.c, block);
457 md_final_raw(md_state.c, block);
458 /* If this is index_b, copy the hash value to |mac_out|. */
459 for (j = 0; j < md_size; j++) {
460 mac_out[j] |= block[j] & is_block_b;
461 }
462 }
463
464 EVP_MD_CTX_init(&md_ctx);
465 if (!EVP_DigestInit_ex(&md_ctx, md, NULL /* engine */)) {
466 EVP_MD_CTX_cleanup(&md_ctx);
467 return 0;
468 }
469
470 /* Complete the HMAC in the standard manner. */
471 for (i = 0; i < md_block_size; i++) {
472 hmac_pad[i] ^= 0x6a;
473 }
474
475 EVP_DigestUpdate(&md_ctx, hmac_pad, md_block_size);
476 EVP_DigestUpdate(&md_ctx, mac_out, md_size);
477 EVP_DigestFinal(&md_ctx, md_out, &md_out_size_u);
478 *md_out_size = md_out_size_u;
479 EVP_MD_CTX_cleanup(&md_ctx);
480
481 return 1;
482 }
483