| 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 |
* [email protected]. |
| 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 |
* ([email protected]). This product includes software written by Tim |
| 51 |
* Hudson ([email protected]). */ |
| 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 |
|
| 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; |
| 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 |
} |
| 164 |
|
| 165 |
/* Ideally the next statement would be: |
| 166 |
* |
| 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 |
/* Below is an SMT-LIB2 verification that the Barrett reductions below are |
| 181 |
* correct within this range: |
| 182 |
* |
| 183 |
* (define-fun barrett ( |
| 184 |
* (x (_ BitVec 32)) |
| 185 |
* (mul (_ BitVec 32)) |
| 186 |
* (shift (_ BitVec 32)) |
| 187 |
* (divisor (_ BitVec 32)) ) (_ BitVec 32) |
| 188 |
* (let ((q (bvsub x (bvmul divisor (bvlshr (bvmul x mul) shift))) )) |
| 189 |
* (ite (bvuge q divisor) |
| 190 |
* (bvsub q divisor) |
| 191 |
* q))) |
| 192 |
* |
| 193 |
* (declare-fun x () (_ BitVec 32)) |
| 194 |
* |
| 195 |
* (assert (or |
| 196 |
* (let ( |
| 197 |
* (divisor (_ bv20 32)) |
| 198 |
* (mul (_ bv25 32)) |
| 199 |
* (shift (_ bv9 32)) |
| 200 |
* (limit (_ bv853 32))) |
| 201 |
* |
| 202 |
* (and (bvule x limit) (not (= (bvurem x divisor) |
| 203 |
* (barrett x mul shift divisor))))) |
| 204 |
* |
| 205 |
* (let ( |
| 206 |
* (divisor (_ bv48 32)) |
| 207 |
* (mul (_ bv10 32)) |
| 208 |
* (shift (_ bv9 32)) |
| 209 |
* (limit (_ bv768 32))) |
| 210 |
* |
| 211 |
* (and (bvule x limit) (not (= (bvurem x divisor) |
| 212 |
* (barrett x mul shift divisor))))) |
| 213 |
* )) |
| 214 |
* |
| 215 |
* (check-sat) |
| 216 |
* (get-model) |
| 217 |
*/ |
| 218 |
|
| 219 |
if (md_size == 16) { |
| 220 |
rotate_offset &= 15; |
| 221 |
} else if (md_size == 20) { |
| 222 |
/* 1/20 is approximated as 25/512 and then Barrett reduction is used. |
| 223 |
* Analytically, this is correct for 0 <= rotate_offset <= 853. */ |
| 224 |
unsigned q = (rotate_offset * 25) >> 9; |
| 225 |
rotate_offset -= q * 20; |
| 226 |
rotate_offset -= |
| 227 |
constant_time_select(constant_time_ge(rotate_offset, 20), 20, 0); |
| 228 |
} else if (md_size == 32) { |
| 229 |
rotate_offset &= 31; |
| 230 |
} else if (md_size == 48) { |
| 231 |
/* 1/48 is approximated as 10/512 and then Barrett reduction is used. |
| 232 |
* Analytically, this is correct for 0 <= rotate_offset <= 768. */ |
| 233 |
unsigned q = (rotate_offset * 10) >> 9; |
| 234 |
rotate_offset -= q * 48; |
| 235 |
rotate_offset -= |
| 236 |
constant_time_select(constant_time_ge(rotate_offset, 48), 48, 0); |
| 237 |
} else { |
| 238 |
/* This should be impossible therefore this path doesn't run in constant |
| 239 |
* time. */ |
| 240 |
assert(0); |
| 241 |
rotate_offset = rotate_offset % md_size; |
| 242 |
} |
| 243 |
|
| 244 |
memset(rotated_mac, 0, md_size); |
| 245 |
for (i = scan_start, j = 0; i < orig_len; i++) { |
| 246 |
uint8_t mac_started = constant_time_ge_8(i, mac_start); |
| 247 |
uint8_t mac_ended = constant_time_ge_8(i, mac_end); |
| 248 |
uint8_t b = in[i]; |
| 249 |
rotated_mac[j++] |= b & mac_started & ~mac_ended; |
| 250 |
j &= constant_time_lt(j, md_size); |
| 251 |
} |
| 252 |
|
| 253 |
/* Now rotate the MAC */ |
| 254 |
#if defined(CBC_MAC_ROTATE_IN_PLACE) |
| 255 |
j = 0; |
| 256 |
for (i = 0; i < md_size; i++) { |
| 257 |
/* in case cache-line is 32 bytes, touch second line */ |
| 258 |
((volatile uint8_t *)rotated_mac)[rotate_offset ^ 32]; |
| 259 |
out[j++] = rotated_mac[rotate_offset++]; |
| 260 |
rotate_offset &= constant_time_lt(rotate_offset, md_size); |
| 261 |
} |
| 262 |
#else |
| 263 |
memset(out, 0, md_size); |
| 264 |
rotate_offset = md_size - rotate_offset; |
| 265 |
rotate_offset &= constant_time_lt(rotate_offset, md_size); |
| 266 |
for (i = 0; i < md_size; i++) { |
| 267 |
for (j = 0; j < md_size; j++) { |
| 268 |
out[j] |= rotated_mac[i] & constant_time_eq_8(j, rotate_offset); |
| 269 |
} |
| 270 |
rotate_offset++; |
| 271 |
rotate_offset &= constant_time_lt(rotate_offset, md_size); |
| 272 |
} |
| 273 |
#endif |
| 274 |
} |
| 275 |
|
| 276 |
/* u32toBE serialises an unsigned, 32-bit number (n) as four bytes at (p) in |
| 277 |
* big-endian order. The value of p is advanced by four. */ |
| 278 |
#define u32toBE(n, p) \ |
| 279 |
(*((p)++)=(uint8_t)(n>>24), \ |
| 280 |
*((p)++)=(uint8_t)(n>>16), \ |
| 281 |
*((p)++)=(uint8_t)(n>>8), \ |
| 282 |
*((p)++)=(uint8_t)(n)) |
| 283 |
|
| 284 |
/* u64toBE serialises an unsigned, 64-bit number (n) as eight bytes at (p) in |
| 285 |
* big-endian order. The value of p is advanced by eight. */ |
| 286 |
#define u64toBE(n, p) \ |
| 287 |
(*((p)++)=(uint8_t)(n>>56), \ |
| 288 |
*((p)++)=(uint8_t)(n>>48), \ |
| 289 |
*((p)++)=(uint8_t)(n>>40), \ |
| 290 |
*((p)++)=(uint8_t)(n>>32), \ |
| 291 |
*((p)++)=(uint8_t)(n>>24), \ |
| 292 |
*((p)++)=(uint8_t)(n>>16), \ |
| 293 |
*((p)++)=(uint8_t)(n>>8), \ |
| 294 |
*((p)++)=(uint8_t)(n)) |
| 295 |
|
| 296 |
/* These functions serialize the state of a hash and thus perform the standard |
| 297 |
* "final" operation without adding the padding and length that such a function |
| 298 |
* typically does. */ |
| 299 |
static void tls1_sha1_final_raw(void *ctx, uint8_t *md_out) { |
| 300 |
SHA_CTX *sha1 = ctx; |
| 301 |
u32toBE(sha1->h[0], md_out); |
| 302 |
u32toBE(sha1->h[1], md_out); |
| 303 |
u32toBE(sha1->h[2], md_out); |
| 304 |
u32toBE(sha1->h[3], md_out); |
| 305 |
u32toBE(sha1->h[4], md_out); |
| 306 |
} |
| 307 |
#define LARGEST_DIGEST_CTX SHA_CTX |
| 308 |
|
| 309 |
static void tls1_sha256_final_raw(void *ctx, uint8_t *md_out) { |
| 310 |
SHA256_CTX *sha256 = ctx; |
| 311 |
unsigned i; |
| 312 |
|
| 313 |
for (i = 0; i < 8; i++) { |
| 314 |
u32toBE(sha256->h[i], md_out); |
| 315 |
} |
| 316 |
} |
| 317 |
#undef LARGEST_DIGEST_CTX |
| 318 |
#define LARGEST_DIGEST_CTX SHA256_CTX |
| 319 |
|
| 320 |
static void tls1_sha512_final_raw(void *ctx, uint8_t *md_out) { |
| 321 |
SHA512_CTX *sha512 = ctx; |
| 322 |
unsigned i; |
| 323 |
|
| 324 |
for (i = 0; i < 8; i++) { |
| 325 |
u64toBE(sha512->h[i], md_out); |
| 326 |
} |
| 327 |
} |
| 328 |
#undef LARGEST_DIGEST_CTX |
| 329 |
#define LARGEST_DIGEST_CTX SHA512_CTX |
| 330 |
|
| 331 |
int EVP_tls_cbc_record_digest_supported(const EVP_MD *md) { |
| 332 |
switch (EVP_MD_type(md)) { |
| 333 |
case NID_sha1: |
| 334 |
case NID_sha256: |
| 335 |
case NID_sha384: |
| 336 |
return 1; |
| 337 |
|
| 338 |
default: |
| 339 |
return 0; |
| 340 |
} |
| 341 |
} |
| 342 |
|
| 343 |
int EVP_tls_cbc_digest_record(const EVP_MD *md, uint8_t *md_out, |
| 344 |
size_t *md_out_size, const uint8_t header[13], |
| 345 |
const uint8_t *data, size_t data_plus_mac_size, |
| 346 |
size_t data_plus_mac_plus_padding_size, |
| 347 |
const uint8_t *mac_secret, |
| 348 |
unsigned mac_secret_length) { |
| 349 |
union { |
| 350 |
double align; |
| 351 |
uint8_t c[sizeof(LARGEST_DIGEST_CTX)]; |
| 352 |
} md_state; |
| 353 |
void (*md_final_raw)(void *ctx, uint8_t *md_out); |
| 354 |
void (*md_transform)(void *ctx, const uint8_t *block); |
| 355 |
unsigned md_size, md_block_size = 64; |
| 356 |
unsigned len, max_mac_bytes, num_blocks, num_starting_blocks, k, |
| 357 |
mac_end_offset, c, index_a, index_b; |
| 358 |
unsigned int bits; /* at most 18 bits */ |
| 359 |
uint8_t length_bytes[MAX_HASH_BIT_COUNT_BYTES]; |
| 360 |
/* hmac_pad is the masked HMAC key. */ |
| 361 |
uint8_t hmac_pad[MAX_HASH_BLOCK_SIZE]; |
| 362 |
uint8_t first_block[MAX_HASH_BLOCK_SIZE]; |
| 363 |
uint8_t mac_out[EVP_MAX_MD_SIZE]; |
| 364 |
unsigned i, j, md_out_size_u; |
| 365 |
EVP_MD_CTX md_ctx; |
| 366 |
/* mdLengthSize is the number of bytes in the length field that terminates |
| 367 |
* the hash. */ |
| 368 |
unsigned md_length_size = 8; |
| 369 |
|
| 370 |
/* This is a, hopefully redundant, check that allows us to forget about |
| 371 |
* many possible overflows later in this function. */ |
| 372 |
assert(data_plus_mac_plus_padding_size < 1024 * 1024); |
| 373 |
|
| 374 |
switch (EVP_MD_type(md)) { |
| 375 |
case NID_sha1: |
| 376 |
SHA1_Init((SHA_CTX *)md_state.c); |
| 377 |
md_final_raw = tls1_sha1_final_raw; |
| 378 |
md_transform = |
| 379 |
(void (*)(void *ctx, const uint8_t *block))SHA1_Transform; |
| 380 |
md_size = 20; |
| 381 |
break; |
| 382 |
|
| 383 |
case NID_sha256: |
| 384 |
SHA256_Init((SHA256_CTX *)md_state.c); |
| 385 |
md_final_raw = tls1_sha256_final_raw; |
| 386 |
md_transform = |
| 387 |
(void (*)(void *ctx, const uint8_t *block))SHA256_Transform; |
| 388 |
md_size = 32; |
| 389 |
break; |
| 390 |
|
| 391 |
case NID_sha384: |
| 392 |
SHA384_Init((SHA512_CTX *)md_state.c); |
| 393 |
md_final_raw = tls1_sha512_final_raw; |
| 394 |
md_transform = |
| 395 |
(void (*)(void *ctx, const uint8_t *block))SHA512_Transform; |
| 396 |
md_size = 384 / 8; |
| 397 |
md_block_size = 128; |
| 398 |
md_length_size = 16; |
| 399 |
break; |
| 400 |
|
| 401 |
default: |
| 402 |
/* EVP_tls_cbc_record_digest_supported should have been called first to |
| 403 |
* check that the hash function is supported. */ |
| 404 |
assert(0); |
| 405 |
*md_out_size = 0; |
| 406 |
return 0; |
| 407 |
} |
| 408 |
|
| 409 |
assert(md_length_size <= MAX_HASH_BIT_COUNT_BYTES); |
| 410 |
assert(md_block_size <= MAX_HASH_BLOCK_SIZE); |
| 411 |
assert(md_size <= EVP_MAX_MD_SIZE); |
| 412 |
|
| 413 |
static const unsigned kHeaderLength = 13; |
| 414 |
|
| 415 |
/* kVarianceBlocks is the number of blocks of the hash that we have to |
| 416 |
* calculate in constant time because they could be altered by the |
| 417 |
* padding value. |
| 418 |
* |
| 419 |
* TLSv1 has MACs up to 48 bytes long (SHA-384) and the padding is not |
| 420 |
* required to be minimal. Therefore we say that the final six blocks |
| 421 |
* can vary based on the padding. */ |
| 422 |
static const unsigned kVarianceBlocks = 6; |
| 423 |
|
| 424 |
/* From now on we're dealing with the MAC, which conceptually has 13 |
| 425 |
* bytes of `header' before the start of the data. */ |
| 426 |
len = data_plus_mac_plus_padding_size + kHeaderLength; |
| 427 |
/* max_mac_bytes contains the maximum bytes of bytes in the MAC, including |
| 428 |
* |header|, assuming that there's no padding. */ |
| 429 |
max_mac_bytes = len - md_size - 1; |
| 430 |
/* num_blocks is the maximum number of hash blocks. */ |
| 431 |
num_blocks = |
| 432 |
(max_mac_bytes + 1 + md_length_size + md_block_size - 1) / md_block_size; |
| 433 |
/* In order to calculate the MAC in constant time we have to handle |
| 434 |
* the final blocks specially because the padding value could cause the |
| 435 |
* end to appear somewhere in the final |kVarianceBlocks| blocks and we |
| 436 |
* can't leak where. However, |num_starting_blocks| worth of data can |
| 437 |
* be hashed right away because no padding value can affect whether |
| 438 |
* they are plaintext. */ |
| 439 |
num_starting_blocks = 0; |
| 440 |
/* k is the starting byte offset into the conceptual header||data where |
| 441 |
* we start processing. */ |
| 442 |
k = 0; |
| 443 |
/* mac_end_offset is the index just past the end of the data to be |
| 444 |
* MACed. */ |
| 445 |
mac_end_offset = data_plus_mac_size + kHeaderLength - md_size; |
| 446 |
/* c is the index of the 0x80 byte in the final hash block that |
| 447 |
* contains application data. */ |
| 448 |
c = mac_end_offset % md_block_size; |
| 449 |
/* index_a is the hash block number that contains the 0x80 terminating |
| 450 |
* value. */ |
| 451 |
index_a = mac_end_offset / md_block_size; |
| 452 |
/* index_b is the hash block number that contains the 64-bit hash |
| 453 |
* length, in bits. */ |
| 454 |
index_b = (mac_end_offset + md_length_size) / md_block_size; |
| 455 |
/* bits is the hash-length in bits. It includes the additional hash |
| 456 |
* block for the masked HMAC key. */ |
| 457 |
|
| 458 |
if (num_blocks > kVarianceBlocks) { |
| 459 |
num_starting_blocks = num_blocks - kVarianceBlocks; |
| 460 |
k = md_block_size * num_starting_blocks; |
| 461 |
} |
| 462 |
|
| 463 |
bits = 8 * mac_end_offset; |
| 464 |
|
| 465 |
/* Compute the initial HMAC block. */ |
| 466 |
bits += 8 * md_block_size; |
| 467 |
memset(hmac_pad, 0, md_block_size); |
| 468 |
assert(mac_secret_length <= sizeof(hmac_pad)); |
| 469 |
memcpy(hmac_pad, mac_secret, mac_secret_length); |
| 470 |
for (i = 0; i < md_block_size; i++) { |
| 471 |
hmac_pad[i] ^= 0x36; |
| 472 |
} |
| 473 |
|
| 474 |
md_transform(md_state.c, hmac_pad); |
| 475 |
|
| 476 |
memset(length_bytes, 0, md_length_size - 4); |
| 477 |
length_bytes[md_length_size - 4] = (uint8_t)(bits >> 24); |
| 478 |
length_bytes[md_length_size - 3] = (uint8_t)(bits >> 16); |
| 479 |
length_bytes[md_length_size - 2] = (uint8_t)(bits >> 8); |
| 480 |
length_bytes[md_length_size - 1] = (uint8_t)bits; |
| 481 |
|
| 482 |
if (k > 0) { |
| 483 |
/* k is a multiple of md_block_size. */ |
| 484 |
memcpy(first_block, header, 13); |
| 485 |
memcpy(first_block + 13, data, md_block_size - 13); |
| 486 |
md_transform(md_state.c, first_block); |
| 487 |
for (i = 1; i < k / md_block_size; i++) { |
| 488 |
md_transform(md_state.c, data + md_block_size * i - 13); |
| 489 |
} |
| 490 |
} |
| 491 |
|
| 492 |
memset(mac_out, 0, sizeof(mac_out)); |
| 493 |
|
| 494 |
/* We now process the final hash blocks. For each block, we construct |
| 495 |
* it in constant time. If the |i==index_a| then we'll include the 0x80 |
| 496 |
* bytes and zero pad etc. For each block we selectively copy it, in |
| 497 |
* constant time, to |mac_out|. */ |
| 498 |
for (i = num_starting_blocks; i <= num_starting_blocks + kVarianceBlocks; |
| 499 |
i++) { |
| 500 |
uint8_t block[MAX_HASH_BLOCK_SIZE]; |
| 501 |
uint8_t is_block_a = constant_time_eq_8(i, index_a); |
| 502 |
uint8_t is_block_b = constant_time_eq_8(i, index_b); |
| 503 |
for (j = 0; j < md_block_size; j++) { |
| 504 |
uint8_t b = 0, is_past_c, is_past_cp1; |
| 505 |
if (k < kHeaderLength) { |
| 506 |
b = header[k]; |
| 507 |
} else if (k < data_plus_mac_plus_padding_size + kHeaderLength) { |
| 508 |
b = data[k - kHeaderLength]; |
| 509 |
} |
| 510 |
k++; |
| 511 |
|
| 512 |
is_past_c = is_block_a & constant_time_ge_8(j, c); |
| 513 |
is_past_cp1 = is_block_a & constant_time_ge_8(j, c + 1); |
| 514 |
/* If this is the block containing the end of the |
| 515 |
* application data, and we are at the offset for the |
| 516 |
* 0x80 value, then overwrite b with 0x80. */ |
| 517 |
b = constant_time_select_8(is_past_c, 0x80, b); |
| 518 |
/* If this the the block containing the end of the |
| 519 |
* application data and we're past the 0x80 value then |
| 520 |
* just write zero. */ |
| 521 |
b = b & ~is_past_cp1; |
| 522 |
/* If this is index_b (the final block), but not |
| 523 |
* index_a (the end of the data), then the 64-bit |
| 524 |
* length didn't fit into index_a and we're having to |
| 525 |
* add an extra block of zeros. */ |
| 526 |
b &= ~is_block_b | is_block_a; |
| 527 |
|
| 528 |
/* The final bytes of one of the blocks contains the |
| 529 |
* length. */ |
| 530 |
if (j >= md_block_size - md_length_size) { |
| 531 |
/* If this is index_b, write a length byte. */ |
| 532 |
b = constant_time_select_8( |
| 533 |
is_block_b, length_bytes[j - (md_block_size - md_length_size)], b); |
| 534 |
} |
| 535 |
block[j] = b; |
| 536 |
} |
| 537 |
|
| 538 |
md_transform(md_state.c, block); |
| 539 |
md_final_raw(md_state.c, block); |
| 540 |
/* If this is index_b, copy the hash value to |mac_out|. */ |
| 541 |
for (j = 0; j < md_size; j++) { |
| 542 |
mac_out[j] |= block[j] & is_block_b; |
| 543 |
} |
| 544 |
} |
| 545 |
|
| 546 |
EVP_MD_CTX_init(&md_ctx); |
| 547 |
if (!EVP_DigestInit_ex(&md_ctx, md, NULL /* engine */)) { |
| 548 |
EVP_MD_CTX_cleanup(&md_ctx); |
| 549 |
return 0; |
| 550 |
} |
| 551 |
|
| 552 |
/* Complete the HMAC in the standard manner. */ |
| 553 |
for (i = 0; i < md_block_size; i++) { |
| 554 |
hmac_pad[i] ^= 0x6a; |
| 555 |
} |
| 556 |
|
| 557 |
EVP_DigestUpdate(&md_ctx, hmac_pad, md_block_size); |
| 558 |
EVP_DigestUpdate(&md_ctx, mac_out, md_size); |
| 559 |
EVP_DigestFinal(&md_ctx, md_out, &md_out_size_u); |
| 560 |
*md_out_size = md_out_size_u; |
| 561 |
EVP_MD_CTX_cleanup(&md_ctx); |
| 562 |
|
| 563 |
return 1; |
| 564 |
} |