1 /* $OpenBSD: s3_cbc.c,v 1.9 2014/12/15 00:46:53 doug Exp $ */ 2 /* ==================================================================== 3 * Copyright (c) 2012 The OpenSSL Project. All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in 14 * the documentation and/or other materials provided with the 15 * distribution. 16 * 17 * 3. All advertising materials mentioning features or use of this 18 * software must display the following acknowledgment: 19 * "This product includes software developed by the OpenSSL Project 20 * for use in the OpenSSL Toolkit. (http://www.openssl.org/)" 21 * 22 * 4. The names "OpenSSL Toolkit" and "OpenSSL Project" must not be used to 23 * endorse or promote products derived from this software without 24 * prior written permission. For written permission, please contact 25 * openssl-core@openssl.org. 26 * 27 * 5. Products derived from this software may not be called "OpenSSL" 28 * nor may "OpenSSL" appear in their names without prior written 29 * permission of the OpenSSL Project. 30 * 31 * 6. Redistributions of any form whatsoever must retain the following 32 * acknowledgment: 33 * "This product includes software developed by the OpenSSL Project 34 * for use in the OpenSSL Toolkit (http://www.openssl.org/)" 35 * 36 * THIS SOFTWARE IS PROVIDED BY THE OpenSSL PROJECT ``AS IS'' AND ANY 37 * EXPRESSED OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 38 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 39 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE OpenSSL PROJECT OR 40 * ITS CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 41 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 42 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 43 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 44 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, 45 * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 46 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED 47 * OF THE POSSIBILITY OF SUCH DAMAGE. 48 * ==================================================================== 49 * 50 * This product includes cryptographic software written by Eric Young 51 * (eay@cryptsoft.com). This product includes software written by Tim 52 * Hudson (tjh@cryptsoft.com). 53 * 54 */ 55 56 #include "ssl_locl.h" 57 58 #include <openssl/md5.h> 59 #include <openssl/sha.h> 60 61 /* MAX_HASH_BIT_COUNT_BYTES is the maximum number of bytes in the hash's length 62 * field. (SHA-384/512 have 128-bit length.) */ 63 #define MAX_HASH_BIT_COUNT_BYTES 16 64 65 /* MAX_HASH_BLOCK_SIZE is the maximum hash block size that we'll support. 66 * Currently SHA-384/512 has a 128-byte block size and that's the largest 67 * supported by TLS.) */ 68 #define MAX_HASH_BLOCK_SIZE 128 69 70 /* Some utility functions are needed: 71 * 72 * These macros return the given value with the MSB copied to all the other 73 * bits. They use the fact that arithmetic shift shifts-in the sign bit. 74 * However, this is not ensured by the C standard so you may need to replace 75 * them with something else on odd CPUs. */ 76 #define DUPLICATE_MSB_TO_ALL(x) ((unsigned)((int)(x) >> (sizeof(int) * 8 - 1))) 77 #define DUPLICATE_MSB_TO_ALL_8(x) ((unsigned char)(DUPLICATE_MSB_TO_ALL(x))) 78 79 /* constant_time_lt returns 0xff if a<b and 0x00 otherwise. */ 80 static unsigned 81 constant_time_lt(unsigned a, unsigned b) 82 { 83 a -= b; 84 return DUPLICATE_MSB_TO_ALL(a); 85 } 86 87 /* constant_time_ge returns 0xff if a>=b and 0x00 otherwise. */ 88 static unsigned 89 constant_time_ge(unsigned a, unsigned b) 90 { 91 a -= b; 92 return DUPLICATE_MSB_TO_ALL(~a); 93 } 94 95 /* constant_time_eq_8 returns 0xff if a==b and 0x00 otherwise. */ 96 static unsigned char 97 constant_time_eq_8(unsigned a, unsigned b) 98 { 99 unsigned c = a ^ b; 100 c--; 101 return DUPLICATE_MSB_TO_ALL_8(c); 102 } 103 104 /* ssl3_cbc_remove_padding removes padding from the decrypted, SSLv3, CBC 105 * record in |rec| by updating |rec->length| in constant time. 106 * 107 * block_size: the block size of the cipher used to encrypt the record. 108 * returns: 109 * 0: (in non-constant time) if the record is publicly invalid. 110 * 1: if the padding was valid 111 * -1: otherwise. */ 112 int 113 ssl3_cbc_remove_padding(const SSL* s, SSL3_RECORD *rec, unsigned block_size, 114 unsigned mac_size) 115 { 116 unsigned padding_length, good; 117 const unsigned overhead = 1 /* padding length byte */ + mac_size; 118 119 /* These lengths are all public so we can test them in non-constant 120 * time. */ 121 if (overhead > rec->length) 122 return 0; 123 124 padding_length = rec->data[rec->length - 1]; 125 good = constant_time_ge(rec->length, padding_length + overhead); 126 /* SSLv3 requires that the padding is minimal. */ 127 good &= constant_time_ge(block_size, padding_length + 1); 128 padding_length = good & (padding_length + 1); 129 rec->length -= padding_length; 130 rec->type |= padding_length << 8; /* kludge: pass padding length */ 131 return (int)((good & 1) | (~good & -1)); 132 } 133 134 /* tls1_cbc_remove_padding removes the CBC padding from the decrypted, TLS, CBC 135 * record in |rec| in constant time and returns 1 if the padding is valid and 136 * -1 otherwise. It also removes any explicit IV from the start of the record 137 * without leaking any timing about whether there was enough space after the 138 * padding was removed. 139 * 140 * block_size: the block size of the cipher used to encrypt the record. 141 * returns: 142 * 0: (in non-constant time) if the record is publicly invalid. 143 * 1: if the padding was valid 144 * -1: otherwise. */ 145 int 146 tls1_cbc_remove_padding(const SSL* s, SSL3_RECORD *rec, unsigned block_size, 147 unsigned mac_size) 148 { 149 unsigned padding_length, good, to_check, i; 150 const unsigned overhead = 1 /* padding length byte */ + mac_size; 151 152 /* Check if version requires explicit IV */ 153 if (SSL_USE_EXPLICIT_IV(s)) { 154 /* These lengths are all public so we can test them in 155 * non-constant time. 156 */ 157 if (overhead + block_size > rec->length) 158 return 0; 159 /* We can now safely skip explicit IV */ 160 rec->data += block_size; 161 rec->input += block_size; 162 rec->length -= block_size; 163 } else if (overhead > rec->length) 164 return 0; 165 166 padding_length = rec->data[rec->length - 1]; 167 168 /* NB: if compression is in operation the first packet may not be of 169 * even length so the padding bug check cannot be performed. This bug 170 * workaround has been around since SSLeay so hopefully it is either 171 * fixed now or no buggy implementation supports compression [steve] 172 * (We don't support compression either, so it's not in operation.) 173 */ 174 if ((s->options & SSL_OP_TLS_BLOCK_PADDING_BUG)) { 175 /* First packet is even in size, so check */ 176 if ((memcmp(s->s3->read_sequence, "\0\0\0\0\0\0\0\0", 177 SSL3_SEQUENCE_SIZE) == 0) && !(padding_length & 1)) { 178 s->s3->flags|=TLS1_FLAGS_TLS_PADDING_BUG; 179 } 180 if ((s->s3->flags & TLS1_FLAGS_TLS_PADDING_BUG) && 181 padding_length > 0) { 182 padding_length--; 183 } 184 } 185 186 if (EVP_CIPHER_flags(s->enc_read_ctx->cipher) & EVP_CIPH_FLAG_AEAD_CIPHER) { 187 /* padding is already verified */ 188 rec->length -= padding_length + 1; 189 return 1; 190 } 191 192 good = constant_time_ge(rec->length, overhead + padding_length); 193 /* The padding consists of a length byte at the end of the record and 194 * then that many bytes of padding, all with the same value as the 195 * length byte. Thus, with the length byte included, there are i+1 196 * bytes of padding. 197 * 198 * We can't check just |padding_length+1| bytes because that leaks 199 * decrypted information. Therefore we always have to check the maximum 200 * amount of padding possible. (Again, the length of the record is 201 * public information so we can use it.) */ 202 to_check = 255; /* maximum amount of padding. */ 203 if (to_check > rec->length - 1) 204 to_check = rec->length - 1; 205 206 for (i = 0; i < to_check; i++) { 207 unsigned char mask = constant_time_ge(padding_length, i); 208 unsigned char b = rec->data[rec->length - 1 - i]; 209 /* The final |padding_length+1| bytes should all have the value 210 * |padding_length|. Therefore the XOR should be zero. */ 211 good &= ~(mask&(padding_length ^ b)); 212 } 213 214 /* If any of the final |padding_length+1| bytes had the wrong value, 215 * one or more of the lower eight bits of |good| will be cleared. We 216 * AND the bottom 8 bits together and duplicate the result to all the 217 * bits. */ 218 good &= good >> 4; 219 good &= good >> 2; 220 good &= good >> 1; 221 good <<= sizeof(good)*8 - 1; 222 good = DUPLICATE_MSB_TO_ALL(good); 223 224 padding_length = good & (padding_length + 1); 225 rec->length -= padding_length; 226 rec->type |= padding_length<<8; /* kludge: pass padding length */ 227 228 return (int)((good & 1) | (~good & -1)); 229 } 230 231 /* ssl3_cbc_copy_mac copies |md_size| bytes from the end of |rec| to |out| in 232 * constant time (independent of the concrete value of rec->length, which may 233 * vary within a 256-byte window). 234 * 235 * ssl3_cbc_remove_padding or tls1_cbc_remove_padding must be called prior to 236 * this function. 237 * 238 * On entry: 239 * rec->orig_len >= md_size 240 * md_size <= EVP_MAX_MD_SIZE 241 * 242 * If CBC_MAC_ROTATE_IN_PLACE is defined then the rotation is performed with 243 * variable accesses in a 64-byte-aligned buffer. Assuming that this fits into 244 * a single or pair of cache-lines, then the variable memory accesses don't 245 * actually affect the timing. CPUs with smaller cache-lines [if any] are 246 * not multi-core and are not considered vulnerable to cache-timing attacks. 247 */ 248 #define CBC_MAC_ROTATE_IN_PLACE 249 250 void 251 ssl3_cbc_copy_mac(unsigned char* out, const SSL3_RECORD *rec, 252 unsigned md_size, unsigned orig_len) 253 { 254 #if defined(CBC_MAC_ROTATE_IN_PLACE) 255 unsigned char rotated_mac_buf[64 + EVP_MAX_MD_SIZE]; 256 unsigned char *rotated_mac; 257 #else 258 unsigned char rotated_mac[EVP_MAX_MD_SIZE]; 259 #endif 260 261 /* mac_end is the index of |rec->data| just after the end of the MAC. */ 262 unsigned mac_end = rec->length; 263 unsigned mac_start = mac_end - md_size; 264 /* scan_start contains the number of bytes that we can ignore because 265 * the MAC's position can only vary by 255 bytes. */ 266 unsigned scan_start = 0; 267 unsigned i, j; 268 unsigned div_spoiler; 269 unsigned rotate_offset; 270 271 OPENSSL_assert(orig_len >= md_size); 272 OPENSSL_assert(md_size <= EVP_MAX_MD_SIZE); 273 274 #if defined(CBC_MAC_ROTATE_IN_PLACE) 275 rotated_mac = rotated_mac_buf + ((0 - (size_t)rotated_mac_buf)&63); 276 #endif 277 278 /* This information is public so it's safe to branch based on it. */ 279 if (orig_len > md_size + 255 + 1) 280 scan_start = orig_len - (md_size + 255 + 1); 281 /* div_spoiler contains a multiple of md_size that is used to cause the 282 * modulo operation to be constant time. Without this, the time varies 283 * based on the amount of padding when running on Intel chips at least. 284 * 285 * The aim of right-shifting md_size is so that the compiler doesn't 286 * figure out that it can remove div_spoiler as that would require it 287 * to prove that md_size is always even, which I hope is beyond it. */ 288 div_spoiler = md_size >> 1; 289 div_spoiler <<= (sizeof(div_spoiler) - 1) * 8; 290 rotate_offset = (div_spoiler + mac_start - scan_start) % md_size; 291 292 memset(rotated_mac, 0, md_size); 293 for (i = scan_start, j = 0; i < orig_len; i++) { 294 unsigned char mac_started = constant_time_ge(i, mac_start); 295 unsigned char mac_ended = constant_time_ge(i, mac_end); 296 unsigned char b = rec->data[i]; 297 rotated_mac[j++] |= b & mac_started & ~mac_ended; 298 j &= constant_time_lt(j, md_size); 299 } 300 301 /* Now rotate the MAC */ 302 #if defined(CBC_MAC_ROTATE_IN_PLACE) 303 j = 0; 304 for (i = 0; i < md_size; i++) { 305 /* in case cache-line is 32 bytes, touch second line */ 306 ((volatile unsigned char *)rotated_mac)[rotate_offset^32]; 307 out[j++] = rotated_mac[rotate_offset++]; 308 rotate_offset &= constant_time_lt(rotate_offset, md_size); 309 } 310 #else 311 memset(out, 0, md_size); 312 rotate_offset = md_size - rotate_offset; 313 rotate_offset &= constant_time_lt(rotate_offset, md_size); 314 for (i = 0; i < md_size; i++) { 315 for (j = 0; j < md_size; j++) 316 out[j] |= rotated_mac[i] & constant_time_eq_8(j, rotate_offset); 317 rotate_offset++; 318 rotate_offset &= constant_time_lt(rotate_offset, md_size); 319 } 320 #endif 321 } 322 323 /* u32toLE serialises an unsigned, 32-bit number (n) as four bytes at (p) in 324 * little-endian order. The value of p is advanced by four. */ 325 #define u32toLE(n, p) \ 326 (*((p)++)=(unsigned char)(n), \ 327 *((p)++)=(unsigned char)(n>>8), \ 328 *((p)++)=(unsigned char)(n>>16), \ 329 *((p)++)=(unsigned char)(n>>24)) 330 331 /* These functions serialize the state of a hash and thus perform the standard 332 * "final" operation without adding the padding and length that such a function 333 * typically does. */ 334 static void 335 tls1_md5_final_raw(void* ctx, unsigned char *md_out) 336 { 337 MD5_CTX *md5 = ctx; 338 u32toLE(md5->A, md_out); 339 u32toLE(md5->B, md_out); 340 u32toLE(md5->C, md_out); 341 u32toLE(md5->D, md_out); 342 } 343 344 static void 345 tls1_sha1_final_raw(void* ctx, unsigned char *md_out) 346 { 347 SHA_CTX *sha1 = ctx; 348 l2n(sha1->h0, md_out); 349 l2n(sha1->h1, md_out); 350 l2n(sha1->h2, md_out); 351 l2n(sha1->h3, md_out); 352 l2n(sha1->h4, md_out); 353 } 354 #define LARGEST_DIGEST_CTX SHA_CTX 355 356 static void 357 tls1_sha256_final_raw(void* ctx, unsigned char *md_out) 358 { 359 SHA256_CTX *sha256 = ctx; 360 unsigned i; 361 362 for (i = 0; i < 8; i++) { 363 l2n(sha256->h[i], md_out); 364 } 365 } 366 #undef LARGEST_DIGEST_CTX 367 #define LARGEST_DIGEST_CTX SHA256_CTX 368 369 static void 370 tls1_sha512_final_raw(void* ctx, unsigned char *md_out) 371 { 372 SHA512_CTX *sha512 = ctx; 373 unsigned i; 374 375 for (i = 0; i < 8; i++) { 376 l2n8(sha512->h[i], md_out); 377 } 378 } 379 #undef LARGEST_DIGEST_CTX 380 #define LARGEST_DIGEST_CTX SHA512_CTX 381 382 /* ssl3_cbc_record_digest_supported returns 1 iff |ctx| uses a hash function 383 * which ssl3_cbc_digest_record supports. */ 384 char 385 ssl3_cbc_record_digest_supported(const EVP_MD_CTX *ctx) 386 { 387 switch (EVP_MD_CTX_type(ctx)) { 388 case NID_md5: 389 case NID_sha1: 390 case NID_sha224: 391 case NID_sha256: 392 case NID_sha384: 393 case NID_sha512: 394 return 1; 395 default: 396 return 0; 397 } 398 } 399 400 /* ssl3_cbc_digest_record computes the MAC of a decrypted, padded SSLv3/TLS 401 * record. 402 * 403 * ctx: the EVP_MD_CTX from which we take the hash function. 404 * ssl3_cbc_record_digest_supported must return true for this EVP_MD_CTX. 405 * md_out: the digest output. At most EVP_MAX_MD_SIZE bytes will be written. 406 * md_out_size: if non-NULL, the number of output bytes is written here. 407 * header: the 13-byte, TLS record header. 408 * data: the record data itself, less any preceeding explicit IV. 409 * data_plus_mac_size: the secret, reported length of the data and MAC 410 * once the padding has been removed. 411 * data_plus_mac_plus_padding_size: the public length of the whole 412 * record, including padding. 413 * is_sslv3: non-zero if we are to use SSLv3. Otherwise, TLS. 414 * 415 * On entry: by virtue of having been through one of the remove_padding 416 * functions, above, we know that data_plus_mac_size is large enough to contain 417 * a padding byte and MAC. (If the padding was invalid, it might contain the 418 * padding too. ) */ 419 int 420 ssl3_cbc_digest_record(const EVP_MD_CTX *ctx, unsigned char* md_out, 421 size_t* md_out_size, const unsigned char header[13], 422 const unsigned char *data, size_t data_plus_mac_size, 423 size_t data_plus_mac_plus_padding_size, const unsigned char *mac_secret, 424 unsigned mac_secret_length, char is_sslv3) 425 { 426 union { double align; 427 unsigned char c[sizeof(LARGEST_DIGEST_CTX)]; 428 } md_state; 429 void (*md_final_raw)(void *ctx, unsigned char *md_out); 430 void (*md_transform)(void *ctx, const unsigned char *block); 431 unsigned md_size, md_block_size = 64; 432 unsigned sslv3_pad_length = 40, header_length, variance_blocks, 433 len, max_mac_bytes, num_blocks, 434 num_starting_blocks, k, mac_end_offset, c, index_a, index_b; 435 unsigned int bits; /* at most 18 bits */ 436 unsigned char length_bytes[MAX_HASH_BIT_COUNT_BYTES]; 437 /* hmac_pad is the masked HMAC key. */ 438 unsigned char hmac_pad[MAX_HASH_BLOCK_SIZE]; 439 unsigned char first_block[MAX_HASH_BLOCK_SIZE]; 440 unsigned char mac_out[EVP_MAX_MD_SIZE]; 441 unsigned i, j, md_out_size_u; 442 EVP_MD_CTX md_ctx; 443 /* mdLengthSize is the number of bytes in the length field that terminates 444 * the hash. */ 445 unsigned md_length_size = 8; 446 char length_is_big_endian = 1; 447 448 /* This is a, hopefully redundant, check that allows us to forget about 449 * many possible overflows later in this function. */ 450 OPENSSL_assert(data_plus_mac_plus_padding_size < 1024*1024); 451 452 switch (EVP_MD_CTX_type(ctx)) { 453 case NID_md5: 454 MD5_Init((MD5_CTX*)md_state.c); 455 md_final_raw = tls1_md5_final_raw; 456 md_transform = (void(*)(void *ctx, const unsigned char *block)) MD5_Transform; 457 md_size = 16; 458 sslv3_pad_length = 48; 459 length_is_big_endian = 0; 460 break; 461 case NID_sha1: 462 SHA1_Init((SHA_CTX*)md_state.c); 463 md_final_raw = tls1_sha1_final_raw; 464 md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA1_Transform; 465 md_size = 20; 466 break; 467 case NID_sha224: 468 SHA224_Init((SHA256_CTX*)md_state.c); 469 md_final_raw = tls1_sha256_final_raw; 470 md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA256_Transform; 471 md_size = 224/8; 472 break; 473 case NID_sha256: 474 SHA256_Init((SHA256_CTX*)md_state.c); 475 md_final_raw = tls1_sha256_final_raw; 476 md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA256_Transform; 477 md_size = 32; 478 break; 479 case NID_sha384: 480 SHA384_Init((SHA512_CTX*)md_state.c); 481 md_final_raw = tls1_sha512_final_raw; 482 md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA512_Transform; 483 md_size = 384/8; 484 md_block_size = 128; 485 md_length_size = 16; 486 break; 487 case NID_sha512: 488 SHA512_Init((SHA512_CTX*)md_state.c); 489 md_final_raw = tls1_sha512_final_raw; 490 md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA512_Transform; 491 md_size = 64; 492 md_block_size = 128; 493 md_length_size = 16; 494 break; 495 default: 496 /* ssl3_cbc_record_digest_supported should have been 497 * called first to check that the hash function is 498 * supported. */ 499 OPENSSL_assert(0); 500 if (md_out_size) 501 *md_out_size = 0; 502 return 0; 503 } 504 505 OPENSSL_assert(md_length_size <= MAX_HASH_BIT_COUNT_BYTES); 506 OPENSSL_assert(md_block_size <= MAX_HASH_BLOCK_SIZE); 507 OPENSSL_assert(md_size <= EVP_MAX_MD_SIZE); 508 509 header_length = 13; 510 if (is_sslv3) { 511 header_length = mac_secret_length + sslv3_pad_length + 512 8 /* sequence number */ + 513 1 /* record type */ + 514 2 /* record length */; 515 } 516 517 /* variance_blocks is the number of blocks of the hash that we have to 518 * calculate in constant time because they could be altered by the 519 * padding value. 520 * 521 * In SSLv3, the padding must be minimal so the end of the plaintext 522 * varies by, at most, 15+20 = 35 bytes. (We conservatively assume that 523 * the MAC size varies from 0..20 bytes.) In case the 9 bytes of hash 524 * termination (0x80 + 64-bit length) don't fit in the final block, we 525 * say that the final two blocks can vary based on the padding. 526 * 527 * TLSv1 has MACs up to 48 bytes long (SHA-384) and the padding is not 528 * required to be minimal. Therefore we say that the final six blocks 529 * can vary based on the padding. 530 * 531 * Later in the function, if the message is short and there obviously 532 * cannot be this many blocks then variance_blocks can be reduced. */ 533 variance_blocks = is_sslv3 ? 2 : 6; 534 /* From now on we're dealing with the MAC, which conceptually has 13 535 * bytes of `header' before the start of the data (TLS) or 71/75 bytes 536 * (SSLv3) */ 537 len = data_plus_mac_plus_padding_size + header_length; 538 /* max_mac_bytes contains the maximum bytes of bytes in the MAC, including 539 * |header|, assuming that there's no padding. */ 540 max_mac_bytes = len - md_size - 1; 541 /* num_blocks is the maximum number of hash blocks. */ 542 num_blocks = (max_mac_bytes + 1 + md_length_size + md_block_size - 1) / md_block_size; 543 /* In order to calculate the MAC in constant time we have to handle 544 * the final blocks specially because the padding value could cause the 545 * end to appear somewhere in the final |variance_blocks| blocks and we 546 * can't leak where. However, |num_starting_blocks| worth of data can 547 * be hashed right away because no padding value can affect whether 548 * they are plaintext. */ 549 num_starting_blocks = 0; 550 /* k is the starting byte offset into the conceptual header||data where 551 * we start processing. */ 552 k = 0; 553 /* mac_end_offset is the index just past the end of the data to be 554 * MACed. */ 555 mac_end_offset = data_plus_mac_size + header_length - md_size; 556 /* c is the index of the 0x80 byte in the final hash block that 557 * contains application data. */ 558 c = mac_end_offset % md_block_size; 559 /* index_a is the hash block number that contains the 0x80 terminating 560 * value. */ 561 index_a = mac_end_offset / md_block_size; 562 /* index_b is the hash block number that contains the 64-bit hash 563 * length, in bits. */ 564 index_b = (mac_end_offset + md_length_size) / md_block_size; 565 /* bits is the hash-length in bits. It includes the additional hash 566 * block for the masked HMAC key, or whole of |header| in the case of 567 * SSLv3. */ 568 569 /* For SSLv3, if we're going to have any starting blocks then we need 570 * at least two because the header is larger than a single block. */ 571 if (num_blocks > variance_blocks + (is_sslv3 ? 1 : 0)) { 572 num_starting_blocks = num_blocks - variance_blocks; 573 k = md_block_size*num_starting_blocks; 574 } 575 576 bits = 8*mac_end_offset; 577 if (!is_sslv3) { 578 /* Compute the initial HMAC block. For SSLv3, the padding and 579 * secret bytes are included in |header| because they take more 580 * than a single block. */ 581 bits += 8*md_block_size; 582 memset(hmac_pad, 0, md_block_size); 583 OPENSSL_assert(mac_secret_length <= sizeof(hmac_pad)); 584 memcpy(hmac_pad, mac_secret, mac_secret_length); 585 for (i = 0; i < md_block_size; i++) 586 hmac_pad[i] ^= 0x36; 587 588 md_transform(md_state.c, hmac_pad); 589 } 590 591 if (length_is_big_endian) { 592 memset(length_bytes, 0, md_length_size - 4); 593 length_bytes[md_length_size - 4] = (unsigned char)(bits >> 24); 594 length_bytes[md_length_size - 3] = (unsigned char)(bits >> 16); 595 length_bytes[md_length_size - 2] = (unsigned char)(bits >> 8); 596 length_bytes[md_length_size - 1] = (unsigned char)bits; 597 } else { 598 memset(length_bytes, 0, md_length_size); 599 length_bytes[md_length_size - 5] = (unsigned char)(bits >> 24); 600 length_bytes[md_length_size - 6] = (unsigned char)(bits >> 16); 601 length_bytes[md_length_size - 7] = (unsigned char)(bits >> 8); 602 length_bytes[md_length_size - 8] = (unsigned char)bits; 603 } 604 605 if (k > 0) { 606 if (is_sslv3) { 607 /* The SSLv3 header is larger than a single block. 608 * overhang is the number of bytes beyond a single 609 * block that the header consumes: either 7 bytes 610 * (SHA1) or 11 bytes (MD5). */ 611 unsigned overhang = header_length - md_block_size; 612 md_transform(md_state.c, header); 613 memcpy(first_block, header + md_block_size, overhang); 614 memcpy(first_block + overhang, data, md_block_size - overhang); 615 md_transform(md_state.c, first_block); 616 for (i = 1; i < k/md_block_size - 1; i++) 617 md_transform(md_state.c, data + md_block_size*i - overhang); 618 } else { 619 /* k is a multiple of md_block_size. */ 620 memcpy(first_block, header, 13); 621 memcpy(first_block + 13, data, md_block_size - 13); 622 md_transform(md_state.c, first_block); 623 for (i = 1; i < k/md_block_size; i++) 624 md_transform(md_state.c, data + md_block_size*i - 13); 625 } 626 } 627 628 memset(mac_out, 0, sizeof(mac_out)); 629 630 /* We now process the final hash blocks. For each block, we construct 631 * it in constant time. If the |i==index_a| then we'll include the 0x80 632 * bytes and zero pad etc. For each block we selectively copy it, in 633 * constant time, to |mac_out|. */ 634 for (i = num_starting_blocks; i <= num_starting_blocks + variance_blocks; i++) { 635 unsigned char block[MAX_HASH_BLOCK_SIZE]; 636 unsigned char is_block_a = constant_time_eq_8(i, index_a); 637 unsigned char is_block_b = constant_time_eq_8(i, index_b); 638 for (j = 0; j < md_block_size; j++) { 639 unsigned char b = 0, is_past_c, is_past_cp1; 640 if (k < header_length) 641 b = header[k]; 642 else if (k < data_plus_mac_plus_padding_size + header_length) 643 b = data[k - header_length]; 644 k++; 645 646 is_past_c = is_block_a & constant_time_ge(j, c); 647 is_past_cp1 = is_block_a & constant_time_ge(j, c + 1); 648 /* If this is the block containing the end of the 649 * application data, and we are at the offset for the 650 * 0x80 value, then overwrite b with 0x80. */ 651 b = (b&~is_past_c) | (0x80&is_past_c); 652 /* If this the the block containing the end of the 653 * application data and we're past the 0x80 value then 654 * just write zero. */ 655 b = b&~is_past_cp1; 656 /* If this is index_b (the final block), but not 657 * index_a (the end of the data), then the 64-bit 658 * length didn't fit into index_a and we're having to 659 * add an extra block of zeros. */ 660 b &= ~is_block_b | is_block_a; 661 662 /* The final bytes of one of the blocks contains the 663 * length. */ 664 if (j >= md_block_size - md_length_size) { 665 /* If this is index_b, write a length byte. */ 666 b = (b&~is_block_b) | (is_block_b&length_bytes[j - (md_block_size - md_length_size)]); 667 } 668 block[j] = b; 669 } 670 671 md_transform(md_state.c, block); 672 md_final_raw(md_state.c, block); 673 /* If this is index_b, copy the hash value to |mac_out|. */ 674 for (j = 0; j < md_size; j++) 675 mac_out[j] |= block[j]&is_block_b; 676 } 677 678 EVP_MD_CTX_init(&md_ctx); 679 if (!EVP_DigestInit_ex(&md_ctx, ctx->digest, NULL /* engine */)) { 680 EVP_MD_CTX_cleanup(&md_ctx); 681 return 0; 682 } 683 if (is_sslv3) { 684 /* We repurpose |hmac_pad| to contain the SSLv3 pad2 block. */ 685 memset(hmac_pad, 0x5c, sslv3_pad_length); 686 687 EVP_DigestUpdate(&md_ctx, mac_secret, mac_secret_length); 688 EVP_DigestUpdate(&md_ctx, hmac_pad, sslv3_pad_length); 689 EVP_DigestUpdate(&md_ctx, mac_out, md_size); 690 } else { 691 /* Complete the HMAC in the standard manner. */ 692 for (i = 0; i < md_block_size; i++) 693 hmac_pad[i] ^= 0x6a; 694 695 EVP_DigestUpdate(&md_ctx, hmac_pad, md_block_size); 696 EVP_DigestUpdate(&md_ctx, mac_out, md_size); 697 } 698 EVP_DigestFinal(&md_ctx, md_out, &md_out_size_u); 699 if (md_out_size) 700 *md_out_size = md_out_size_u; 701 EVP_MD_CTX_cleanup(&md_ctx); 702 703 return 1; 704 } 705