Adam Langley | d9e397b | 2015-01-22 14:27:53 -0800 | [diff] [blame] | 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> |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame^] | 57 | #include <openssl/nid.h> |
Adam Langley | d9e397b | 2015-01-22 14:27:53 -0800 | [diff] [blame] | 58 | #include <openssl/sha.h> |
| 59 | |
| 60 | #include "../internal.h" |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame^] | 61 | #include "internal.h" |
Adam Langley | d9e397b | 2015-01-22 14:27:53 -0800 | [diff] [blame] | 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 | |
| 76 | int EVP_tls_cbc_remove_padding(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 | |
| 123 | return constant_time_select_int(good, 1, -1); |
| 124 | } |
| 125 | |
| 126 | /* If CBC_MAC_ROTATE_IN_PLACE is defined then EVP_tls_cbc_copy_mac is performed |
| 127 | * with variable accesses in a 64-byte-aligned buffer. Assuming that this fits |
| 128 | * into a single or pair of cache-lines, then the variable memory accesses don't |
| 129 | * actually affect the timing. CPUs with smaller cache-lines [if any] are not |
| 130 | * multi-core and are not considered vulnerable to cache-timing attacks. */ |
| 131 | #define CBC_MAC_ROTATE_IN_PLACE |
| 132 | |
| 133 | void EVP_tls_cbc_copy_mac(uint8_t *out, unsigned md_size, |
| 134 | const uint8_t *in, unsigned in_len, |
| 135 | unsigned orig_len) { |
| 136 | #if defined(CBC_MAC_ROTATE_IN_PLACE) |
| 137 | uint8_t rotated_mac_buf[64 + EVP_MAX_MD_SIZE]; |
| 138 | uint8_t *rotated_mac; |
| 139 | #else |
| 140 | uint8_t rotated_mac[EVP_MAX_MD_SIZE]; |
| 141 | #endif |
| 142 | |
| 143 | /* mac_end is the index of |in| just after the end of the MAC. */ |
| 144 | unsigned mac_end = in_len; |
| 145 | unsigned mac_start = mac_end - md_size; |
| 146 | /* scan_start contains the number of bytes that we can ignore because |
| 147 | * the MAC's position can only vary by 255 bytes. */ |
| 148 | unsigned scan_start = 0; |
| 149 | unsigned i, j; |
Adam Langley | d9e397b | 2015-01-22 14:27:53 -0800 | [diff] [blame] | 150 | unsigned rotate_offset; |
| 151 | |
| 152 | assert(orig_len >= in_len); |
| 153 | assert(in_len >= md_size); |
| 154 | assert(md_size <= EVP_MAX_MD_SIZE); |
| 155 | |
| 156 | #if defined(CBC_MAC_ROTATE_IN_PLACE) |
| 157 | rotated_mac = rotated_mac_buf + ((0 - (size_t)rotated_mac_buf) & 63); |
| 158 | #endif |
| 159 | |
| 160 | /* This information is public so it's safe to branch based on it. */ |
| 161 | if (orig_len > md_size + 255 + 1) { |
| 162 | scan_start = orig_len - (md_size + 255 + 1); |
| 163 | } |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame^] | 164 | |
| 165 | /* Ideally the next statement would be: |
Adam Langley | d9e397b | 2015-01-22 14:27:53 -0800 | [diff] [blame] | 166 | * |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame^] | 167 | * rotate_offset = (mac_start - scan_start) % md_size; |
| 168 | * |
| 169 | * However, division is not a constant-time operation (at least on Intel |
| 170 | * chips). Thus we enumerate the possible values of md_size and handle each |
| 171 | * separately. The value of |md_size| is public information (it's determined |
| 172 | * by the cipher suite in the ServerHello) so our timing can vary based on |
| 173 | * its value. */ |
| 174 | |
| 175 | rotate_offset = mac_start - scan_start; |
| 176 | /* rotate_offset can be, at most, 255 (bytes of padding) + 1 (padding length) |
| 177 | * + md_size = 256 + 48 (since SHA-384 is the largest hash) = 304. */ |
| 178 | assert(rotate_offset <= 304); |
| 179 | |
| 180 | if (md_size == 16) { |
| 181 | rotate_offset &= 15; |
| 182 | } else if (md_size == 20) { |
| 183 | /* 1/20 is approximated as 25/512 and then Barrett reduction is used. |
| 184 | * Analytically, this is correct for 0 <= rotate_offset <= 853. */ |
| 185 | unsigned q = (rotate_offset * 25) >> 9; |
| 186 | rotate_offset -= q * 20; |
| 187 | rotate_offset -= |
| 188 | constant_time_select(constant_time_ge(rotate_offset, 20), 20, 0); |
| 189 | } else if (md_size == 32) { |
| 190 | rotate_offset &= 31; |
| 191 | } else if (md_size == 48) { |
| 192 | /* 1/48 is approximated as 10/512 and then Barrett reduction is used. |
| 193 | * Analytically, this is correct for 0 <= rotate_offset <= 768. */ |
| 194 | unsigned q = (rotate_offset * 10) >> 9; |
| 195 | rotate_offset -= q * 48; |
| 196 | rotate_offset -= |
| 197 | constant_time_select(constant_time_ge(rotate_offset, 48), 48, 0); |
| 198 | } else { |
| 199 | /* This should be impossible therefore this path doesn't run in constant |
| 200 | * time. */ |
| 201 | assert(0); |
| 202 | rotate_offset = rotate_offset % md_size; |
| 203 | } |
Adam Langley | d9e397b | 2015-01-22 14:27:53 -0800 | [diff] [blame] | 204 | |
| 205 | memset(rotated_mac, 0, md_size); |
| 206 | for (i = scan_start, j = 0; i < orig_len; i++) { |
| 207 | uint8_t mac_started = constant_time_ge_8(i, mac_start); |
| 208 | uint8_t mac_ended = constant_time_ge_8(i, mac_end); |
| 209 | uint8_t b = in[i]; |
| 210 | rotated_mac[j++] |= b & mac_started & ~mac_ended; |
| 211 | j &= constant_time_lt(j, md_size); |
| 212 | } |
| 213 | |
| 214 | /* Now rotate the MAC */ |
| 215 | #if defined(CBC_MAC_ROTATE_IN_PLACE) |
| 216 | j = 0; |
| 217 | for (i = 0; i < md_size; i++) { |
| 218 | /* in case cache-line is 32 bytes, touch second line */ |
| 219 | ((volatile uint8_t *)rotated_mac)[rotate_offset ^ 32]; |
| 220 | out[j++] = rotated_mac[rotate_offset++]; |
| 221 | rotate_offset &= constant_time_lt(rotate_offset, md_size); |
| 222 | } |
| 223 | #else |
| 224 | memset(out, 0, md_size); |
| 225 | rotate_offset = md_size - rotate_offset; |
| 226 | rotate_offset &= constant_time_lt(rotate_offset, md_size); |
| 227 | for (i = 0; i < md_size; i++) { |
| 228 | for (j = 0; j < md_size; j++) { |
| 229 | out[j] |= rotated_mac[i] & constant_time_eq_8(j, rotate_offset); |
| 230 | } |
| 231 | rotate_offset++; |
| 232 | rotate_offset &= constant_time_lt(rotate_offset, md_size); |
| 233 | } |
| 234 | #endif |
| 235 | } |
| 236 | |
| 237 | /* u32toBE serialises an unsigned, 32-bit number (n) as four bytes at (p) in |
| 238 | * big-endian order. The value of p is advanced by four. */ |
| 239 | #define u32toBE(n, p) \ |
| 240 | (*((p)++)=(uint8_t)(n>>24), \ |
| 241 | *((p)++)=(uint8_t)(n>>16), \ |
| 242 | *((p)++)=(uint8_t)(n>>8), \ |
| 243 | *((p)++)=(uint8_t)(n)) |
| 244 | |
| 245 | /* u64toBE serialises an unsigned, 64-bit number (n) as eight bytes at (p) in |
| 246 | * big-endian order. The value of p is advanced by eight. */ |
| 247 | #define u64toBE(n, p) \ |
| 248 | (*((p)++)=(uint8_t)(n>>56), \ |
| 249 | *((p)++)=(uint8_t)(n>>48), \ |
| 250 | *((p)++)=(uint8_t)(n>>40), \ |
| 251 | *((p)++)=(uint8_t)(n>>32), \ |
| 252 | *((p)++)=(uint8_t)(n>>24), \ |
| 253 | *((p)++)=(uint8_t)(n>>16), \ |
| 254 | *((p)++)=(uint8_t)(n>>8), \ |
| 255 | *((p)++)=(uint8_t)(n)) |
| 256 | |
| 257 | /* These functions serialize the state of a hash and thus perform the standard |
| 258 | * "final" operation without adding the padding and length that such a function |
| 259 | * typically does. */ |
| 260 | static void tls1_sha1_final_raw(void *ctx, uint8_t *md_out) { |
| 261 | SHA_CTX *sha1 = ctx; |
Adam Langley | fad6327 | 2015-11-12 12:15:39 -0800 | [diff] [blame] | 262 | u32toBE(sha1->h[0], md_out); |
| 263 | u32toBE(sha1->h[1], md_out); |
| 264 | u32toBE(sha1->h[2], md_out); |
| 265 | u32toBE(sha1->h[3], md_out); |
| 266 | u32toBE(sha1->h[4], md_out); |
Adam Langley | d9e397b | 2015-01-22 14:27:53 -0800 | [diff] [blame] | 267 | } |
| 268 | #define LARGEST_DIGEST_CTX SHA_CTX |
| 269 | |
| 270 | static void tls1_sha256_final_raw(void *ctx, uint8_t *md_out) { |
| 271 | SHA256_CTX *sha256 = ctx; |
| 272 | unsigned i; |
| 273 | |
| 274 | for (i = 0; i < 8; i++) { |
| 275 | u32toBE(sha256->h[i], md_out); |
| 276 | } |
| 277 | } |
| 278 | #undef LARGEST_DIGEST_CTX |
| 279 | #define LARGEST_DIGEST_CTX SHA256_CTX |
| 280 | |
| 281 | static void tls1_sha512_final_raw(void *ctx, uint8_t *md_out) { |
| 282 | SHA512_CTX *sha512 = ctx; |
| 283 | unsigned i; |
| 284 | |
| 285 | for (i = 0; i < 8; i++) { |
| 286 | u64toBE(sha512->h[i], md_out); |
| 287 | } |
| 288 | } |
| 289 | #undef LARGEST_DIGEST_CTX |
| 290 | #define LARGEST_DIGEST_CTX SHA512_CTX |
| 291 | |
| 292 | int EVP_tls_cbc_record_digest_supported(const EVP_MD *md) { |
| 293 | switch (EVP_MD_type(md)) { |
| 294 | case NID_sha1: |
| 295 | case NID_sha256: |
| 296 | case NID_sha384: |
| 297 | return 1; |
| 298 | |
| 299 | default: |
| 300 | return 0; |
| 301 | } |
| 302 | } |
| 303 | |
| 304 | int EVP_tls_cbc_digest_record(const EVP_MD *md, uint8_t *md_out, |
| 305 | size_t *md_out_size, const uint8_t header[13], |
| 306 | const uint8_t *data, size_t data_plus_mac_size, |
| 307 | size_t data_plus_mac_plus_padding_size, |
| 308 | const uint8_t *mac_secret, |
| 309 | unsigned mac_secret_length) { |
| 310 | union { |
| 311 | double align; |
| 312 | uint8_t c[sizeof(LARGEST_DIGEST_CTX)]; |
| 313 | } md_state; |
| 314 | void (*md_final_raw)(void *ctx, uint8_t *md_out); |
| 315 | void (*md_transform)(void *ctx, const uint8_t *block); |
| 316 | unsigned md_size, md_block_size = 64; |
| 317 | unsigned len, max_mac_bytes, num_blocks, num_starting_blocks, k, |
| 318 | mac_end_offset, c, index_a, index_b; |
| 319 | unsigned int bits; /* at most 18 bits */ |
| 320 | uint8_t length_bytes[MAX_HASH_BIT_COUNT_BYTES]; |
| 321 | /* hmac_pad is the masked HMAC key. */ |
| 322 | uint8_t hmac_pad[MAX_HASH_BLOCK_SIZE]; |
| 323 | uint8_t first_block[MAX_HASH_BLOCK_SIZE]; |
| 324 | uint8_t mac_out[EVP_MAX_MD_SIZE]; |
| 325 | unsigned i, j, md_out_size_u; |
| 326 | EVP_MD_CTX md_ctx; |
| 327 | /* mdLengthSize is the number of bytes in the length field that terminates |
| 328 | * the hash. */ |
| 329 | unsigned md_length_size = 8; |
| 330 | |
| 331 | /* This is a, hopefully redundant, check that allows us to forget about |
| 332 | * many possible overflows later in this function. */ |
| 333 | assert(data_plus_mac_plus_padding_size < 1024 * 1024); |
| 334 | |
| 335 | switch (EVP_MD_type(md)) { |
| 336 | case NID_sha1: |
| 337 | SHA1_Init((SHA_CTX *)md_state.c); |
| 338 | md_final_raw = tls1_sha1_final_raw; |
| 339 | md_transform = |
| 340 | (void (*)(void *ctx, const uint8_t *block))SHA1_Transform; |
| 341 | md_size = 20; |
| 342 | break; |
| 343 | |
| 344 | case NID_sha256: |
| 345 | SHA256_Init((SHA256_CTX *)md_state.c); |
| 346 | md_final_raw = tls1_sha256_final_raw; |
| 347 | md_transform = |
| 348 | (void (*)(void *ctx, const uint8_t *block))SHA256_Transform; |
| 349 | md_size = 32; |
| 350 | break; |
| 351 | |
| 352 | case NID_sha384: |
| 353 | SHA384_Init((SHA512_CTX *)md_state.c); |
| 354 | md_final_raw = tls1_sha512_final_raw; |
| 355 | md_transform = |
| 356 | (void (*)(void *ctx, const uint8_t *block))SHA512_Transform; |
| 357 | md_size = 384 / 8; |
| 358 | md_block_size = 128; |
| 359 | md_length_size = 16; |
| 360 | break; |
| 361 | |
| 362 | default: |
| 363 | /* EVP_tls_cbc_record_digest_supported should have been called first to |
| 364 | * check that the hash function is supported. */ |
| 365 | assert(0); |
| 366 | *md_out_size = 0; |
| 367 | return 0; |
| 368 | } |
| 369 | |
| 370 | assert(md_length_size <= MAX_HASH_BIT_COUNT_BYTES); |
| 371 | assert(md_block_size <= MAX_HASH_BLOCK_SIZE); |
| 372 | assert(md_size <= EVP_MAX_MD_SIZE); |
| 373 | |
| 374 | static const unsigned kHeaderLength = 13; |
| 375 | |
| 376 | /* kVarianceBlocks is the number of blocks of the hash that we have to |
| 377 | * calculate in constant time because they could be altered by the |
| 378 | * padding value. |
| 379 | * |
| 380 | * TLSv1 has MACs up to 48 bytes long (SHA-384) and the padding is not |
| 381 | * required to be minimal. Therefore we say that the final six blocks |
| 382 | * can vary based on the padding. */ |
| 383 | static const unsigned kVarianceBlocks = 6; |
| 384 | |
| 385 | /* From now on we're dealing with the MAC, which conceptually has 13 |
| 386 | * bytes of `header' before the start of the data. */ |
| 387 | len = data_plus_mac_plus_padding_size + kHeaderLength; |
| 388 | /* max_mac_bytes contains the maximum bytes of bytes in the MAC, including |
| 389 | * |header|, assuming that there's no padding. */ |
| 390 | max_mac_bytes = len - md_size - 1; |
| 391 | /* num_blocks is the maximum number of hash blocks. */ |
| 392 | num_blocks = |
| 393 | (max_mac_bytes + 1 + md_length_size + md_block_size - 1) / md_block_size; |
| 394 | /* In order to calculate the MAC in constant time we have to handle |
| 395 | * the final blocks specially because the padding value could cause the |
| 396 | * end to appear somewhere in the final |kVarianceBlocks| blocks and we |
| 397 | * can't leak where. However, |num_starting_blocks| worth of data can |
| 398 | * be hashed right away because no padding value can affect whether |
| 399 | * they are plaintext. */ |
| 400 | num_starting_blocks = 0; |
| 401 | /* k is the starting byte offset into the conceptual header||data where |
| 402 | * we start processing. */ |
| 403 | k = 0; |
| 404 | /* mac_end_offset is the index just past the end of the data to be |
| 405 | * MACed. */ |
| 406 | mac_end_offset = data_plus_mac_size + kHeaderLength - md_size; |
| 407 | /* c is the index of the 0x80 byte in the final hash block that |
| 408 | * contains application data. */ |
| 409 | c = mac_end_offset % md_block_size; |
| 410 | /* index_a is the hash block number that contains the 0x80 terminating |
| 411 | * value. */ |
| 412 | index_a = mac_end_offset / md_block_size; |
| 413 | /* index_b is the hash block number that contains the 64-bit hash |
| 414 | * length, in bits. */ |
| 415 | index_b = (mac_end_offset + md_length_size) / md_block_size; |
| 416 | /* bits is the hash-length in bits. It includes the additional hash |
| 417 | * block for the masked HMAC key. */ |
| 418 | |
| 419 | if (num_blocks > kVarianceBlocks) { |
| 420 | num_starting_blocks = num_blocks - kVarianceBlocks; |
| 421 | k = md_block_size * num_starting_blocks; |
| 422 | } |
| 423 | |
| 424 | bits = 8 * mac_end_offset; |
| 425 | |
| 426 | /* Compute the initial HMAC block. */ |
| 427 | bits += 8 * md_block_size; |
| 428 | memset(hmac_pad, 0, md_block_size); |
| 429 | assert(mac_secret_length <= sizeof(hmac_pad)); |
| 430 | memcpy(hmac_pad, mac_secret, mac_secret_length); |
| 431 | for (i = 0; i < md_block_size; i++) { |
| 432 | hmac_pad[i] ^= 0x36; |
| 433 | } |
| 434 | |
| 435 | md_transform(md_state.c, hmac_pad); |
| 436 | |
| 437 | memset(length_bytes, 0, md_length_size - 4); |
| 438 | length_bytes[md_length_size - 4] = (uint8_t)(bits >> 24); |
| 439 | length_bytes[md_length_size - 3] = (uint8_t)(bits >> 16); |
| 440 | length_bytes[md_length_size - 2] = (uint8_t)(bits >> 8); |
| 441 | length_bytes[md_length_size - 1] = (uint8_t)bits; |
| 442 | |
| 443 | if (k > 0) { |
| 444 | /* k is a multiple of md_block_size. */ |
| 445 | memcpy(first_block, header, 13); |
| 446 | memcpy(first_block + 13, data, md_block_size - 13); |
| 447 | md_transform(md_state.c, first_block); |
| 448 | for (i = 1; i < k / md_block_size; i++) { |
| 449 | md_transform(md_state.c, data + md_block_size * i - 13); |
| 450 | } |
| 451 | } |
| 452 | |
| 453 | memset(mac_out, 0, sizeof(mac_out)); |
| 454 | |
| 455 | /* We now process the final hash blocks. For each block, we construct |
| 456 | * it in constant time. If the |i==index_a| then we'll include the 0x80 |
| 457 | * bytes and zero pad etc. For each block we selectively copy it, in |
| 458 | * constant time, to |mac_out|. */ |
| 459 | for (i = num_starting_blocks; i <= num_starting_blocks + kVarianceBlocks; |
| 460 | i++) { |
| 461 | uint8_t block[MAX_HASH_BLOCK_SIZE]; |
| 462 | uint8_t is_block_a = constant_time_eq_8(i, index_a); |
| 463 | uint8_t is_block_b = constant_time_eq_8(i, index_b); |
| 464 | for (j = 0; j < md_block_size; j++) { |
| 465 | uint8_t b = 0, is_past_c, is_past_cp1; |
| 466 | if (k < kHeaderLength) { |
| 467 | b = header[k]; |
| 468 | } else if (k < data_plus_mac_plus_padding_size + kHeaderLength) { |
| 469 | b = data[k - kHeaderLength]; |
| 470 | } |
| 471 | k++; |
| 472 | |
| 473 | is_past_c = is_block_a & constant_time_ge_8(j, c); |
| 474 | is_past_cp1 = is_block_a & constant_time_ge_8(j, c + 1); |
| 475 | /* If this is the block containing the end of the |
| 476 | * application data, and we are at the offset for the |
| 477 | * 0x80 value, then overwrite b with 0x80. */ |
| 478 | b = constant_time_select_8(is_past_c, 0x80, b); |
| 479 | /* If this the the block containing the end of the |
| 480 | * application data and we're past the 0x80 value then |
| 481 | * just write zero. */ |
| 482 | b = b & ~is_past_cp1; |
| 483 | /* If this is index_b (the final block), but not |
| 484 | * index_a (the end of the data), then the 64-bit |
| 485 | * length didn't fit into index_a and we're having to |
| 486 | * add an extra block of zeros. */ |
| 487 | b &= ~is_block_b | is_block_a; |
| 488 | |
| 489 | /* The final bytes of one of the blocks contains the |
| 490 | * length. */ |
| 491 | if (j >= md_block_size - md_length_size) { |
| 492 | /* If this is index_b, write a length byte. */ |
| 493 | b = constant_time_select_8( |
| 494 | is_block_b, length_bytes[j - (md_block_size - md_length_size)], b); |
| 495 | } |
| 496 | block[j] = b; |
| 497 | } |
| 498 | |
| 499 | md_transform(md_state.c, block); |
| 500 | md_final_raw(md_state.c, block); |
| 501 | /* If this is index_b, copy the hash value to |mac_out|. */ |
| 502 | for (j = 0; j < md_size; j++) { |
| 503 | mac_out[j] |= block[j] & is_block_b; |
| 504 | } |
| 505 | } |
| 506 | |
| 507 | EVP_MD_CTX_init(&md_ctx); |
| 508 | if (!EVP_DigestInit_ex(&md_ctx, md, NULL /* engine */)) { |
| 509 | EVP_MD_CTX_cleanup(&md_ctx); |
| 510 | return 0; |
| 511 | } |
| 512 | |
| 513 | /* Complete the HMAC in the standard manner. */ |
| 514 | for (i = 0; i < md_block_size; i++) { |
| 515 | hmac_pad[i] ^= 0x6a; |
| 516 | } |
| 517 | |
| 518 | EVP_DigestUpdate(&md_ctx, hmac_pad, md_block_size); |
| 519 | EVP_DigestUpdate(&md_ctx, mac_out, md_size); |
| 520 | EVP_DigestFinal(&md_ctx, md_out, &md_out_size_u); |
| 521 | *md_out_size = md_out_size_u; |
| 522 | EVP_MD_CTX_cleanup(&md_ctx); |
| 523 | |
| 524 | return 1; |
| 525 | } |