1/* ssl/s3_cbc.c */ 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. */ 80static unsigned constant_time_lt(unsigned a, unsigned b) 81 { 82 a -= b; 83 return DUPLICATE_MSB_TO_ALL(a); 84 } 85 86/* constant_time_ge returns 0xff if a>=b and 0x00 otherwise. */ 87static unsigned constant_time_ge(unsigned a, unsigned b) 88 { 89 a -= b; 90 return DUPLICATE_MSB_TO_ALL(~a); 91 } 92 93/* constant_time_eq_8 returns 0xff if a==b and 0x00 otherwise. */ 94static unsigned char constant_time_eq_8(unsigned a, unsigned b) 95 { 96 unsigned c = a ^ b; 97 c--; 98 return DUPLICATE_MSB_TO_ALL_8(c); 99 } 100 101/* ssl3_cbc_remove_padding removes padding from the decrypted, SSLv3, CBC 102 * record in |rec| by updating |rec->length| in constant time. 103 * 104 * block_size: the block size of the cipher used to encrypt the record. 105 * returns: 106 * 0: (in non-constant time) if the record is publicly invalid. 107 * 1: if the padding was valid 108 * -1: otherwise. */ 109int ssl3_cbc_remove_padding(const SSL* s, 110 SSL3_RECORD *rec, 111 unsigned block_size, 112 unsigned mac_size) 113 { 114 unsigned padding_length, good; 115 const unsigned overhead = 1 /* padding length byte */ + mac_size; 116 117 /* These lengths are all public so we can test them in non-constant 118 * time. */ 119 if (overhead > rec->length) 120 return 0; 121 122 padding_length = rec->data[rec->length-1]; 123 good = constant_time_ge(rec->length, padding_length+overhead); 124 /* SSLv3 requires that the padding is minimal. */ 125 good &= constant_time_ge(block_size, padding_length+1); 126 padding_length = good & (padding_length+1); 127 rec->length -= padding_length; 128 rec->type |= padding_length<<8; /* kludge: pass padding length */ 129 return (int)((good & 1) | (~good & -1)); 130} 131 132/* tls1_cbc_remove_padding removes the CBC padding from the decrypted, TLS, CBC 133 * record in |rec| in constant time and returns 1 if the padding is valid and 134 * -1 otherwise. It also removes any explicit IV from the start of the record 135 * without leaking any timing about whether there was enough space after the 136 * padding was removed. 137 * 138 * block_size: the block size of the cipher used to encrypt the record. 139 * returns: 140 * 0: (in non-constant time) if the record is publicly invalid. 141 * 1: if the padding was valid 142 * -1: otherwise. */ 143int tls1_cbc_remove_padding(const SSL* s, 144 SSL3_RECORD *rec, 145 unsigned block_size, 146 unsigned mac_size) 147 { 148 unsigned padding_length, good, to_check, i; 149 const unsigned overhead = 1 /* padding length byte */ + mac_size; 150 /* Check if version requires explicit IV */ 151 if (s->version >= TLS1_1_VERSION || s->version == DTLS1_BAD_VER) 152 { 153 /* These lengths are all public so we can test them in 154 * non-constant time. 155 */ 156 if (overhead + block_size > rec->length) 157 return 0; 158 /* We can now safely skip explicit IV */ 159 rec->data += block_size; 160 rec->input += block_size; 161 rec->length -= block_size; 162 } 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 */ 173 if ( (s->options&SSL_OP_TLS_BLOCK_PADDING_BUG) && !s->expand) 174 { 175 /* First packet is even in size, so check */ 176 if ((memcmp(s->s3->read_sequence, "\0\0\0\0\0\0\0\0",8) == 0) && 177 !(padding_length & 1)) 178 { 179 s->s3->flags|=TLS1_FLAGS_TLS_PADDING_BUG; 180 } 181 if ((s->s3->flags & TLS1_FLAGS_TLS_PADDING_BUG) && 182 padding_length > 0) 183 { 184 padding_length--; 185 } 186 } 187 188 if (EVP_CIPHER_flags(s->enc_read_ctx->cipher)&EVP_CIPH_FLAG_AEAD_CIPHER) 189 { 190 /* padding is already verified */ 191 rec->length -= padding_length + 1; 192 return 1; 193 } 194 195 good = constant_time_ge(rec->length, overhead+padding_length); 196 /* The padding consists of a length byte at the end of the record and 197 * then that many bytes of padding, all with the same value as the 198 * length byte. Thus, with the length byte included, there are i+1 199 * bytes of padding. 200 * 201 * We can't check just |padding_length+1| bytes because that leaks 202 * decrypted information. Therefore we always have to check the maximum 203 * amount of padding possible. (Again, the length of the record is 204 * public information so we can use it.) */ 205 to_check = 255; /* maximum amount of padding. */ 206 if (to_check > rec->length-1) 207 to_check = rec->length-1; 208 209 for (i = 0; i < to_check; i++) 210 { 211 unsigned char mask = constant_time_ge(padding_length, i); 212 unsigned char b = rec->data[rec->length-1-i]; 213 /* The final |padding_length+1| bytes should all have the value 214 * |padding_length|. Therefore the XOR should be zero. */ 215 good &= ~(mask&(padding_length ^ b)); 216 } 217 218 /* If any of the final |padding_length+1| bytes had the wrong value, 219 * one or more of the lower eight bits of |good| will be cleared. We 220 * AND the bottom 8 bits together and duplicate the result to all the 221 * bits. */ 222 good &= good >> 4; 223 good &= good >> 2; 224 good &= good >> 1; 225 good <<= sizeof(good)*8-1; 226 good = DUPLICATE_MSB_TO_ALL(good); 227 228 padding_length = good & (padding_length+1); 229 rec->length -= padding_length; 230 rec->type |= padding_length<<8; /* kludge: pass padding length */ 231 232 return (int)((good & 1) | (~good & -1)); 233 } 234 235/* ssl3_cbc_copy_mac copies |md_size| bytes from the end of |rec| to |out| in 236 * constant time (independent of the concrete value of rec->length, which may 237 * vary within a 256-byte window). 238 * 239 * ssl3_cbc_remove_padding or tls1_cbc_remove_padding must be called prior to 240 * this function. 241 * 242 * On entry: 243 * rec->orig_len >= md_size 244 * md_size <= EVP_MAX_MD_SIZE 245 * 246 * If CBC_MAC_ROTATE_IN_PLACE is defined then the rotation is performed with 247 * variable accesses in a 64-byte-aligned buffer. Assuming that this fits into 248 * a single or pair of cache-lines, then the variable memory accesses don't 249 * actually affect the timing. CPUs with smaller cache-lines [if any] are 250 * not multi-core and are not considered vulnerable to cache-timing attacks. 251 */ 252#define CBC_MAC_ROTATE_IN_PLACE 253 254void ssl3_cbc_copy_mac(unsigned char* out, 255 const SSL3_RECORD *rec, 256 unsigned md_size,unsigned orig_len) 257 { 258#if defined(CBC_MAC_ROTATE_IN_PLACE) 259 unsigned char rotated_mac_buf[64+EVP_MAX_MD_SIZE]; 260 unsigned char *rotated_mac; 261#else 262 unsigned char rotated_mac[EVP_MAX_MD_SIZE]; 263#endif 264 265 /* mac_end is the index of |rec->data| just after the end of the MAC. */ 266 unsigned mac_end = rec->length; 267 unsigned mac_start = mac_end - md_size; 268 /* scan_start contains the number of bytes that we can ignore because 269 * the MAC's position can only vary by 255 bytes. */ 270 unsigned scan_start = 0; 271 unsigned i, j; 272 unsigned div_spoiler; 273 unsigned rotate_offset; 274 275 OPENSSL_assert(orig_len >= md_size); 276 OPENSSL_assert(md_size <= EVP_MAX_MD_SIZE); 277 278#if defined(CBC_MAC_ROTATE_IN_PLACE) 279 rotated_mac = rotated_mac_buf + ((0-(size_t)rotated_mac_buf)&63); 280#endif 281 282 /* This information is public so it's safe to branch based on it. */ 283 if (orig_len > md_size + 255 + 1) 284 scan_start = orig_len - (md_size + 255 + 1); 285 /* div_spoiler contains a multiple of md_size that is used to cause the 286 * modulo operation to be constant time. Without this, the time varies 287 * based on the amount of padding when running on Intel chips at least. 288 * 289 * The aim of right-shifting md_size is so that the compiler doesn't 290 * figure out that it can remove div_spoiler as that would require it 291 * to prove that md_size is always even, which I hope is beyond it. */ 292 div_spoiler = md_size >> 1; 293 div_spoiler <<= (sizeof(div_spoiler)-1)*8; 294 rotate_offset = (div_spoiler + mac_start - scan_start) % md_size; 295 296 memset(rotated_mac, 0, md_size); 297 for (i = scan_start, j = 0; i < orig_len; i++) 298 { 299 unsigned char mac_started = constant_time_ge(i, mac_start); 300 unsigned char mac_ended = constant_time_ge(i, mac_end); 301 unsigned char b = rec->data[i]; 302 rotated_mac[j++] |= b & mac_started & ~mac_ended; 303 j &= constant_time_lt(j,md_size); 304 } 305 306 /* Now rotate the MAC */ 307#if defined(CBC_MAC_ROTATE_IN_PLACE) 308 j = 0; 309 for (i = 0; i < md_size; i++) 310 { 311 /* in case cache-line is 32 bytes, touch second line */ 312 ((volatile unsigned char *)rotated_mac)[rotate_offset^32]; 313 out[j++] = rotated_mac[rotate_offset++]; 314 rotate_offset &= constant_time_lt(rotate_offset,md_size); 315 } 316#else 317 memset(out, 0, md_size); 318 rotate_offset = md_size - rotate_offset; 319 rotate_offset &= constant_time_lt(rotate_offset,md_size); 320 for (i = 0; i < md_size; i++) 321 { 322 for (j = 0; j < md_size; j++) 323 out[j] |= rotated_mac[i] & constant_time_eq_8(j, rotate_offset); 324 rotate_offset++; 325 rotate_offset &= constant_time_lt(rotate_offset,md_size); 326 } 327#endif 328 } 329 330/* u32toLE serialises an unsigned, 32-bit number (n) as four bytes at (p) in 331 * little-endian order. The value of p is advanced by four. */ 332#define u32toLE(n, p) \ 333 (*((p)++)=(unsigned char)(n), \ 334 *((p)++)=(unsigned char)(n>>8), \ 335 *((p)++)=(unsigned char)(n>>16), \ 336 *((p)++)=(unsigned char)(n>>24)) 337 338/* These functions serialize the state of a hash and thus perform the standard 339 * "final" operation without adding the padding and length that such a function 340 * typically does. */ 341static void tls1_md5_final_raw(void* ctx, unsigned char *md_out) 342 { 343 MD5_CTX *md5 = ctx; 344 u32toLE(md5->A, md_out); 345 u32toLE(md5->B, md_out); 346 u32toLE(md5->C, md_out); 347 u32toLE(md5->D, md_out); 348 } 349 350static void tls1_sha1_final_raw(void* ctx, unsigned char *md_out) 351 { 352 SHA_CTX *sha1 = ctx; 353 l2n(sha1->h0, md_out); 354 l2n(sha1->h1, md_out); 355 l2n(sha1->h2, md_out); 356 l2n(sha1->h3, md_out); 357 l2n(sha1->h4, md_out); 358 } 359#define LARGEST_DIGEST_CTX SHA_CTX 360 361#ifndef OPENSSL_NO_SHA256 362static void tls1_sha256_final_raw(void* ctx, unsigned char *md_out) 363 { 364 SHA256_CTX *sha256 = ctx; 365 unsigned i; 366 367 for (i = 0; i < 8; i++) 368 { 369 l2n(sha256->h[i], md_out); 370 } 371 } 372#undef LARGEST_DIGEST_CTX 373#define LARGEST_DIGEST_CTX SHA256_CTX 374#endif 375 376#ifndef OPENSSL_NO_SHA512 377static void tls1_sha512_final_raw(void* ctx, unsigned char *md_out) 378 { 379 SHA512_CTX *sha512 = ctx; 380 unsigned i; 381 382 for (i = 0; i < 8; i++) 383 { 384 l2n8(sha512->h[i], md_out); 385 } 386 } 387#undef LARGEST_DIGEST_CTX 388#define LARGEST_DIGEST_CTX SHA512_CTX 389#endif 390 391/* ssl3_cbc_record_digest_supported returns 1 iff |ctx| uses a hash function 392 * which ssl3_cbc_digest_record supports. */ 393char ssl3_cbc_record_digest_supported(const EVP_MD_CTX *ctx) 394 { 395#ifdef OPENSSL_FIPS 396 if (FIPS_mode()) 397 return 0; 398#endif 399 switch (EVP_MD_CTX_type(ctx)) 400 { 401 case NID_md5: 402 case NID_sha1: 403#ifndef OPENSSL_NO_SHA256 404 case NID_sha224: 405 case NID_sha256: 406#endif 407#ifndef OPENSSL_NO_SHA512 408 case NID_sha384: 409 case NID_sha512: 410#endif 411 return 1; 412 default: 413 return 0; 414 } 415 } 416 417/* ssl3_cbc_digest_record computes the MAC of a decrypted, padded SSLv3/TLS 418 * record. 419 * 420 * ctx: the EVP_MD_CTX from which we take the hash function. 421 * ssl3_cbc_record_digest_supported must return true for this EVP_MD_CTX. 422 * md_out: the digest output. At most EVP_MAX_MD_SIZE bytes will be written. 423 * md_out_size: if non-NULL, the number of output bytes is written here. 424 * header: the 13-byte, TLS record header. 425 * data: the record data itself, less any preceeding explicit IV. 426 * data_plus_mac_size: the secret, reported length of the data and MAC 427 * once the padding has been removed. 428 * data_plus_mac_plus_padding_size: the public length of the whole 429 * record, including padding. 430 * is_sslv3: non-zero if we are to use SSLv3. Otherwise, TLS. 431 * 432 * On entry: by virtue of having been through one of the remove_padding 433 * functions, above, we know that data_plus_mac_size is large enough to contain 434 * a padding byte and MAC. (If the padding was invalid, it might contain the 435 * padding too. ) */ 436void ssl3_cbc_digest_record( 437 const EVP_MD_CTX *ctx, 438 unsigned char* md_out, 439 size_t* md_out_size, 440 const unsigned char header[13], 441 const unsigned char *data, 442 size_t data_plus_mac_size, 443 size_t data_plus_mac_plus_padding_size, 444 const unsigned char *mac_secret, 445 unsigned mac_secret_length, 446 char is_sslv3) 447 { 448 union { double align; 449 unsigned char c[sizeof(LARGEST_DIGEST_CTX)]; } md_state; 450 void (*md_final_raw)(void *ctx, unsigned char *md_out); 451 void (*md_transform)(void *ctx, const unsigned char *block); 452 unsigned md_size, md_block_size = 64; 453 unsigned sslv3_pad_length = 40, header_length, variance_blocks, 454 len, max_mac_bytes, num_blocks, 455 num_starting_blocks, k, mac_end_offset, c, index_a, index_b; 456 unsigned int bits; /* at most 18 bits */ 457 unsigned char length_bytes[MAX_HASH_BIT_COUNT_BYTES]; 458 /* hmac_pad is the masked HMAC key. */ 459 unsigned char hmac_pad[MAX_HASH_BLOCK_SIZE]; 460 unsigned char first_block[MAX_HASH_BLOCK_SIZE]; 461 unsigned char mac_out[EVP_MAX_MD_SIZE]; 462 unsigned i, j, md_out_size_u; 463 EVP_MD_CTX md_ctx; 464 /* mdLengthSize is the number of bytes in the length field that terminates 465 * the hash. */ 466 unsigned md_length_size = 8; 467 char length_is_big_endian = 1; 468 469 /* This is a, hopefully redundant, check that allows us to forget about 470 * many possible overflows later in this function. */ 471 OPENSSL_assert(data_plus_mac_plus_padding_size < 1024*1024); 472 473 switch (EVP_MD_CTX_type(ctx)) 474 { 475 case NID_md5: 476 MD5_Init((MD5_CTX*)md_state.c); 477 md_final_raw = tls1_md5_final_raw; 478 md_transform = (void(*)(void *ctx, const unsigned char *block)) MD5_Transform; 479 md_size = 16; 480 sslv3_pad_length = 48; 481 length_is_big_endian = 0; 482 break; 483 case NID_sha1: 484 SHA1_Init((SHA_CTX*)md_state.c); 485 md_final_raw = tls1_sha1_final_raw; 486 md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA1_Transform; 487 md_size = 20; 488 break; 489#ifndef OPENSSL_NO_SHA256 490 case NID_sha224: 491 SHA224_Init((SHA256_CTX*)md_state.c); 492 md_final_raw = tls1_sha256_final_raw; 493 md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA256_Transform; 494 md_size = 224/8; 495 break; 496 case NID_sha256: 497 SHA256_Init((SHA256_CTX*)md_state.c); 498 md_final_raw = tls1_sha256_final_raw; 499 md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA256_Transform; 500 md_size = 32; 501 break; 502#endif 503#ifndef OPENSSL_NO_SHA512 504 case NID_sha384: 505 SHA384_Init((SHA512_CTX*)md_state.c); 506 md_final_raw = tls1_sha512_final_raw; 507 md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA512_Transform; 508 md_size = 384/8; 509 md_block_size = 128; 510 md_length_size = 16; 511 break; 512 case NID_sha512: 513 SHA512_Init((SHA512_CTX*)md_state.c); 514 md_final_raw = tls1_sha512_final_raw; 515 md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA512_Transform; 516 md_size = 64; 517 md_block_size = 128; 518 md_length_size = 16; 519 break; 520#endif 521 default: 522 /* ssl3_cbc_record_digest_supported should have been 523 * called first to check that the hash function is 524 * supported. */ 525 OPENSSL_assert(0); 526 if (md_out_size) 527 *md_out_size = -1; 528 return; 529 } 530 531 OPENSSL_assert(md_length_size <= MAX_HASH_BIT_COUNT_BYTES); 532 OPENSSL_assert(md_block_size <= MAX_HASH_BLOCK_SIZE); 533 OPENSSL_assert(md_size <= EVP_MAX_MD_SIZE); 534 535 header_length = 13; 536 if (is_sslv3) 537 { 538 header_length = 539 mac_secret_length + 540 sslv3_pad_length + 541 8 /* sequence number */ + 542 1 /* record type */ + 543 2 /* record length */; 544 } 545 546 /* variance_blocks is the number of blocks of the hash that we have to 547 * calculate in constant time because they could be altered by the 548 * padding value. 549 * 550 * In SSLv3, the padding must be minimal so the end of the plaintext 551 * varies by, at most, 15+20 = 35 bytes. (We conservatively assume that 552 * the MAC size varies from 0..20 bytes.) In case the 9 bytes of hash 553 * termination (0x80 + 64-bit length) don't fit in the final block, we 554 * say that the final two blocks can vary based on the padding. 555 * 556 * TLSv1 has MACs up to 48 bytes long (SHA-384) and the padding is not 557 * required to be minimal. Therefore we say that the final six blocks 558 * can vary based on the padding. 559 * 560 * Later in the function, if the message is short and there obviously 561 * cannot be this many blocks then variance_blocks can be reduced. */ 562 variance_blocks = is_sslv3 ? 2 : 6; 563 /* From now on we're dealing with the MAC, which conceptually has 13 564 * bytes of `header' before the start of the data (TLS) or 71/75 bytes 565 * (SSLv3) */ 566 len = data_plus_mac_plus_padding_size + header_length; 567 /* max_mac_bytes contains the maximum bytes of bytes in the MAC, including 568 * |header|, assuming that there's no padding. */ 569 max_mac_bytes = len - md_size - 1; 570 /* num_blocks is the maximum number of hash blocks. */ 571 num_blocks = (max_mac_bytes + 1 + md_length_size + md_block_size - 1) / md_block_size; 572 /* In order to calculate the MAC in constant time we have to handle 573 * the final blocks specially because the padding value could cause the 574 * end to appear somewhere in the final |variance_blocks| blocks and we 575 * can't leak where. However, |num_starting_blocks| worth of data can 576 * be hashed right away because no padding value can affect whether 577 * they are plaintext. */ 578 num_starting_blocks = 0; 579 /* k is the starting byte offset into the conceptual header||data where 580 * we start processing. */ 581 k = 0; 582 /* mac_end_offset is the index just past the end of the data to be 583 * MACed. */ 584 mac_end_offset = data_plus_mac_size + header_length - md_size; 585 /* c is the index of the 0x80 byte in the final hash block that 586 * contains application data. */ 587 c = mac_end_offset % md_block_size; 588 /* index_a is the hash block number that contains the 0x80 terminating 589 * value. */ 590 index_a = mac_end_offset / md_block_size; 591 /* index_b is the hash block number that contains the 64-bit hash 592 * length, in bits. */ 593 index_b = (mac_end_offset + md_length_size) / md_block_size; 594 /* bits is the hash-length in bits. It includes the additional hash 595 * block for the masked HMAC key, or whole of |header| in the case of 596 * SSLv3. */ 597 598 /* For SSLv3, if we're going to have any starting blocks then we need 599 * at least two because the header is larger than a single block. */ 600 if (num_blocks > variance_blocks + (is_sslv3 ? 1 : 0)) 601 { 602 num_starting_blocks = num_blocks - variance_blocks; 603 k = md_block_size*num_starting_blocks; 604 } 605 606 bits = 8*mac_end_offset; 607 if (!is_sslv3) 608 { 609 /* Compute the initial HMAC block. For SSLv3, the padding and 610 * secret bytes are included in |header| because they take more 611 * than a single block. */ 612 bits += 8*md_block_size; 613 memset(hmac_pad, 0, md_block_size); 614 OPENSSL_assert(mac_secret_length <= sizeof(hmac_pad)); 615 memcpy(hmac_pad, mac_secret, mac_secret_length); 616 for (i = 0; i < md_block_size; i++) 617 hmac_pad[i] ^= 0x36; 618 619 md_transform(md_state.c, hmac_pad); 620 } 621 622 if (length_is_big_endian) 623 { 624 memset(length_bytes,0,md_length_size-4); 625 length_bytes[md_length_size-4] = (unsigned char)(bits>>24); 626 length_bytes[md_length_size-3] = (unsigned char)(bits>>16); 627 length_bytes[md_length_size-2] = (unsigned char)(bits>>8); 628 length_bytes[md_length_size-1] = (unsigned char)bits; 629 } 630 else 631 { 632 memset(length_bytes,0,md_length_size); 633 length_bytes[md_length_size-5] = (unsigned char)(bits>>24); 634 length_bytes[md_length_size-6] = (unsigned char)(bits>>16); 635 length_bytes[md_length_size-7] = (unsigned char)(bits>>8); 636 length_bytes[md_length_size-8] = (unsigned char)bits; 637 } 638 639 if (k > 0) 640 { 641 if (is_sslv3) 642 { 643 /* The SSLv3 header is larger than a single block. 644 * overhang is the number of bytes beyond a single 645 * block that the header consumes: either 7 bytes 646 * (SHA1) or 11 bytes (MD5). */ 647 unsigned overhang = header_length-md_block_size; 648 md_transform(md_state.c, header); 649 memcpy(first_block, header + md_block_size, overhang); 650 memcpy(first_block + overhang, data, md_block_size-overhang); 651 md_transform(md_state.c, first_block); 652 for (i = 1; i < k/md_block_size - 1; i++) 653 md_transform(md_state.c, data + md_block_size*i - overhang); 654 } 655 else 656 { 657 /* k is a multiple of md_block_size. */ 658 memcpy(first_block, header, 13); 659 memcpy(first_block+13, data, md_block_size-13); 660 md_transform(md_state.c, first_block); 661 for (i = 1; i < k/md_block_size; i++) 662 md_transform(md_state.c, data + md_block_size*i - 13); 663 } 664 } 665 666 memset(mac_out, 0, sizeof(mac_out)); 667 668 /* We now process the final hash blocks. For each block, we construct 669 * it in constant time. If the |i==index_a| then we'll include the 0x80 670 * bytes and zero pad etc. For each block we selectively copy it, in 671 * constant time, to |mac_out|. */ 672 for (i = num_starting_blocks; i <= num_starting_blocks+variance_blocks; i++) 673 { 674 unsigned char block[MAX_HASH_BLOCK_SIZE]; 675 unsigned char is_block_a = constant_time_eq_8(i, index_a); 676 unsigned char is_block_b = constant_time_eq_8(i, index_b); 677 for (j = 0; j < md_block_size; j++) 678 { 679 unsigned char b = 0, is_past_c, is_past_cp1; 680 if (k < header_length) 681 b = header[k]; 682 else if (k < data_plus_mac_plus_padding_size + header_length) 683 b = data[k-header_length]; 684 k++; 685 686 is_past_c = is_block_a & constant_time_ge(j, c); 687 is_past_cp1 = is_block_a & constant_time_ge(j, c+1); 688 /* If this is the block containing the end of the 689 * application data, and we are at the offset for the 690 * 0x80 value, then overwrite b with 0x80. */ 691 b = (b&~is_past_c) | (0x80&is_past_c); 692 /* If this the the block containing the end of the 693 * application data and we're past the 0x80 value then 694 * just write zero. */ 695 b = b&~is_past_cp1; 696 /* If this is index_b (the final block), but not 697 * index_a (the end of the data), then the 64-bit 698 * length didn't fit into index_a and we're having to 699 * add an extra block of zeros. */ 700 b &= ~is_block_b | is_block_a; 701 702 /* The final bytes of one of the blocks contains the 703 * length. */ 704 if (j >= md_block_size - md_length_size) 705 { 706 /* If this is index_b, write a length byte. */ 707 b = (b&~is_block_b) | (is_block_b&length_bytes[j-(md_block_size-md_length_size)]); 708 } 709 block[j] = b; 710 } 711 712 md_transform(md_state.c, block); 713 md_final_raw(md_state.c, block); 714 /* If this is index_b, copy the hash value to |mac_out|. */ 715 for (j = 0; j < md_size; j++) 716 mac_out[j] |= block[j]&is_block_b; 717 } 718 719 EVP_MD_CTX_init(&md_ctx); 720 EVP_DigestInit_ex(&md_ctx, ctx->digest, NULL /* engine */); 721 if (is_sslv3) 722 { 723 /* We repurpose |hmac_pad| to contain the SSLv3 pad2 block. */ 724 memset(hmac_pad, 0x5c, sslv3_pad_length); 725 726 EVP_DigestUpdate(&md_ctx, mac_secret, mac_secret_length); 727 EVP_DigestUpdate(&md_ctx, hmac_pad, sslv3_pad_length); 728 EVP_DigestUpdate(&md_ctx, mac_out, md_size); 729 } 730 else 731 { 732 /* Complete the HMAC in the standard manner. */ 733 for (i = 0; i < md_block_size; i++) 734 hmac_pad[i] ^= 0x6a; 735 736 EVP_DigestUpdate(&md_ctx, hmac_pad, md_block_size); 737 EVP_DigestUpdate(&md_ctx, mac_out, md_size); 738 } 739 EVP_DigestFinal(&md_ctx, md_out, &md_out_size_u); 740 if (md_out_size) 741 *md_out_size = md_out_size_u; 742 EVP_MD_CTX_cleanup(&md_ctx); 743 } 744 745#ifdef OPENSSL_FIPS 746 747/* Due to the need to use EVP in FIPS mode we can't reimplement digests but 748 * we can ensure the number of blocks processed is equal for all cases 749 * by digesting additional data. 750 */ 751 752void tls_fips_digest_extra( 753 const EVP_CIPHER_CTX *cipher_ctx, EVP_MD_CTX *mac_ctx, 754 const unsigned char *data, size_t data_len, size_t orig_len) 755 { 756 size_t block_size, digest_pad, blocks_data, blocks_orig; 757 if (EVP_CIPHER_CTX_mode(cipher_ctx) != EVP_CIPH_CBC_MODE) 758 return; 759 block_size = EVP_MD_CTX_block_size(mac_ctx); 760 /* We are in FIPS mode if we get this far so we know we have only SHA* 761 * digests and TLS to deal with. 762 * Minimum digest padding length is 17 for SHA384/SHA512 and 9 763 * otherwise. 764 * Additional header is 13 bytes. To get the number of digest blocks 765 * processed round up the amount of data plus padding to the nearest 766 * block length. Block length is 128 for SHA384/SHA512 and 64 otherwise. 767 * So we have: 768 * blocks = (payload_len + digest_pad + 13 + block_size - 1)/block_size 769 * equivalently: 770 * blocks = (payload_len + digest_pad + 12)/block_size + 1 771 * HMAC adds a constant overhead. 772 * We're ultimately only interested in differences so this becomes 773 * blocks = (payload_len + 29)/128 774 * for SHA384/SHA512 and 775 * blocks = (payload_len + 21)/64 776 * otherwise. 777 */ 778 digest_pad = block_size == 64 ? 21 : 29; 779 blocks_orig = (orig_len + digest_pad)/block_size; 780 blocks_data = (data_len + digest_pad)/block_size; 781 /* MAC enough blocks to make up the difference between the original 782 * and actual lengths plus one extra block to ensure this is never a 783 * no op. The "data" pointer should always have enough space to 784 * perform this operation as it is large enough for a maximum 785 * length TLS buffer. 786 */ 787 EVP_DigestSignUpdate(mac_ctx, data, 788 (blocks_orig - blocks_data + 1) * block_size); 789 } 790#endif 791