int EVP_tls_cbc_digest_record(const EVP_MD *md, uint8_t *md_out,
size_t *md_out_size, const uint8_t header[13],
const uint8_t *data, size_t data_plus_mac_size,
size_t data_plus_mac_plus_padding_size,
const uint8_t *mac_secret,
unsigned mac_secret_length) {
HASH_CTX md_state;
void (*md_final_raw)(HASH_CTX *ctx, uint8_t *md_out);
void (*md_transform)(HASH_CTX *ctx, const uint8_t *block);
unsigned md_size, md_block_size = 64;
/* md_length_size is the number of bytes in the length field that terminates
* the hash. */
unsigned md_length_size = 8;
/* Bound the acceptable input so we can forget about many possible overflows
* later in this function. This is redundant with the record size limits in
* TLS. */
if (data_plus_mac_plus_padding_size >= 1024 * 1024) {
assert(0);
return 0;
}
switch (EVP_MD_type(md)) {
case NID_sha1:
SHA1_Init(&md_state.sha1);
md_final_raw = tls1_sha1_final_raw;
md_transform = tls1_sha1_transform;
md_size = SHA_DIGEST_LENGTH;
break;
case NID_sha256:
SHA256_Init(&md_state.sha256);
md_final_raw = tls1_sha256_final_raw;
md_transform = tls1_sha256_transform;
md_size = SHA256_DIGEST_LENGTH;
break;
case NID_sha384:
SHA384_Init(&md_state.sha512);
md_final_raw = tls1_sha512_final_raw;
md_transform = tls1_sha512_transform;
md_size = SHA384_DIGEST_LENGTH;
md_block_size = 128;
md_length_size = 16;
break;
default:
/* EVP_tls_cbc_record_digest_supported should have been called first to
* check that the hash function is supported. */
assert(0);
*md_out_size = 0;
return 0;
}
assert(md_length_size <= MAX_HASH_BIT_COUNT_BYTES);
assert(md_block_size <= MAX_HASH_BLOCK_SIZE);
assert(md_size <= EVP_MAX_MD_SIZE);
static const size_t kHeaderLength = 13;
/* kVarianceBlocks is the number of blocks of the hash that we have to
* calculate in constant time because they could be altered by the
* padding value.
*
* TLSv1 has MACs up to 48 bytes long (SHA-384) and the padding is not
* required to be minimal. Therefore we say that the final six blocks
* can vary based on the padding. */
static const size_t kVarianceBlocks = 6;
/* From now on we're dealing with the MAC, which conceptually has 13
* bytes of `header' before the start of the data. */
size_t len = data_plus_mac_plus_padding_size + kHeaderLength;
/* max_mac_bytes contains the maximum bytes of bytes in the MAC, including
* |header|, assuming that there's no padding. */
size_t max_mac_bytes = len - md_size - 1;
/* num_blocks is the maximum number of hash blocks. */
size_t num_blocks =
(max_mac_bytes + 1 + md_length_size + md_block_size - 1) / md_block_size;
/* In order to calculate the MAC in constant time we have to handle
* the final blocks specially because the padding value could cause the
* end to appear somewhere in the final |kVarianceBlocks| blocks and we
* can't leak where. However, |num_starting_blocks| worth of data can
* be hashed right away because no padding value can affect whether
* they are plaintext. */
size_t num_starting_blocks = 0;
/* k is the starting byte offset into the conceptual header||data where
* we start processing. */
size_t k = 0;
/* mac_end_offset is the index just past the end of the data to be
* MACed. */
size_t mac_end_offset = data_plus_mac_size + kHeaderLength - md_size;
/* c is the index of the 0x80 byte in the final hash block that
* contains application data. */
size_t c = mac_end_offset % md_block_size;
/* index_a is the hash block number that contains the 0x80 terminating
* value. */
size_t index_a = mac_end_offset / md_block_size;
/* index_b is the hash block number that contains the 64-bit hash
* length, in bits. */
size_t index_b = (mac_end_offset + md_length_size) / md_block_size;
if (num_blocks > kVarianceBlocks) {
num_starting_blocks = num_blocks - kVarianceBlocks;
k = md_block_size * num_starting_blocks;
}
/* bits is the hash-length in bits. It includes the additional hash
* block for the masked HMAC key. */
size_t bits = 8 * mac_end_offset; /* at most 18 bits to represent */
/* Compute the initial HMAC block. */
bits += 8 * md_block_size;
/* hmac_pad is the masked HMAC key. */
uint8_t hmac_pad[MAX_HASH_BLOCK_SIZE];
OPENSSL_memset(hmac_pad, 0, md_block_size);
assert(mac_secret_length <= sizeof(hmac_pad));
OPENSSL_memcpy(hmac_pad, mac_secret, mac_secret_length);
for (size_t i = 0; i < md_block_size; i++) {
hmac_pad[i] ^= 0x36;
}
md_transform(&md_state, hmac_pad);
/* The length check means |bits| fits in four bytes. */
uint8_t length_bytes[MAX_HASH_BIT_COUNT_BYTES];
OPENSSL_memset(length_bytes, 0, md_length_size - 4);
length_bytes[md_length_size - 4] = (uint8_t)(bits >> 24);
length_bytes[md_length_size - 3] = (uint8_t)(bits >> 16);
length_bytes[md_length_size - 2] = (uint8_t)(bits >> 8);
length_bytes[md_length_size - 1] = (uint8_t)bits;
if (k > 0) {
/* k is a multiple of md_block_size. */
uint8_t first_block[MAX_HASH_BLOCK_SIZE];
OPENSSL_memcpy(first_block, header, 13);
OPENSSL_memcpy(first_block + 13, data, md_block_size - 13);
md_transform(&md_state, first_block);
for (size_t i = 1; i < k / md_block_size; i++) {
md_transform(&md_state, data + md_block_size * i - 13);
}
}
uint8_t mac_out[EVP_MAX_MD_SIZE];
OPENSSL_memset(mac_out, 0, sizeof(mac_out));
/* We now process the final hash blocks. For each block, we construct
* it in constant time. If the |i==index_a| then we'll include the 0x80
* bytes and zero pad etc. For each block we selectively copy it, in
* constant time, to |mac_out|. */
for (size_t i = num_starting_blocks;
i <= num_starting_blocks + kVarianceBlocks; i++) {
uint8_t block[MAX_HASH_BLOCK_SIZE];
uint8_t is_block_a = constant_time_eq_8(i, index_a);
uint8_t is_block_b = constant_time_eq_8(i, index_b);
for (size_t j = 0; j < md_block_size; j++) {
uint8_t b = 0;
if (k < kHeaderLength) {
b = header[k];
} else if (k < data_plus_mac_plus_padding_size + kHeaderLength) {
b = data[k - kHeaderLength];
}
k++;
uint8_t is_past_c = is_block_a & constant_time_ge_8(j, c);
uint8_t is_past_cp1 = is_block_a & constant_time_ge_8(j, c + 1);
/* If this is the block containing the end of the
* application data, and we are at the offset for the
* 0x80 value, then overwrite b with 0x80. */
b = constant_time_select_8(is_past_c, 0x80, b);
/* If this the the block containing the end of the
* application data and we're past the 0x80 value then
* just write zero. */
b = b & ~is_past_cp1;
/* If this is index_b (the final block), but not
* index_a (the end of the data), then the 64-bit
* length didn't fit into index_a and we're having to
* add an extra block of zeros. */
b &= ~is_block_b | is_block_a;
/* The final bytes of one of the blocks contains the
* length. */
if (j >= md_block_size - md_length_size) {
/* If this is index_b, write a length byte. */
b = constant_time_select_8(
is_block_b, length_bytes[j - (md_block_size - md_length_size)], b);
}
block[j] = b;
}
md_transform(&md_state, block);
md_final_raw(&md_state, block);
/* If this is index_b, copy the hash value to |mac_out|. */
for (size_t j = 0; j < md_size; j++) {
mac_out[j] |= block[j] & is_block_b;
}
}
EVP_MD_CTX md_ctx;
EVP_MD_CTX_init(&md_ctx);
if (!EVP_DigestInit_ex(&md_ctx, md, NULL /* engine */)) {
EVP_MD_CTX_cleanup(&md_ctx);
return 0;
}
/* Complete the HMAC in the standard manner. */
for (size_t i = 0; i < md_block_size; i++) {
hmac_pad[i] ^= 0x6a;
}
EVP_DigestUpdate(&md_ctx, hmac_pad, md_block_size);
EVP_DigestUpdate(&md_ctx, mac_out, md_size);
unsigned md_out_size_u;
EVP_DigestFinal(&md_ctx, md_out, &md_out_size_u);
*md_out_size = md_out_size_u;
EVP_MD_CTX_cleanup(&md_ctx);
return 1;
}
void EVP_tls_cbc_copy_mac(uint8_t *out, unsigned md_size,
const uint8_t *in, unsigned in_len,
unsigned orig_len) {
#if defined(CBC_MAC_ROTATE_IN_PLACE)
uint8_t rotated_mac_buf[64 + EVP_MAX_MD_SIZE];
uint8_t *rotated_mac;
#else
uint8_t rotated_mac[EVP_MAX_MD_SIZE];
#endif
/* mac_end is the index of |in| just after the end of the MAC. */
unsigned mac_end = in_len;
unsigned mac_start = mac_end - md_size;
/* scan_start contains the number of bytes that we can ignore because
* the MAC's position can only vary by 255 bytes. */
unsigned scan_start = 0;
unsigned i, j;
unsigned div_spoiler;
unsigned rotate_offset;
assert(orig_len >= in_len);
assert(in_len >= md_size);
assert(md_size <= EVP_MAX_MD_SIZE);
#if defined(CBC_MAC_ROTATE_IN_PLACE)
rotated_mac = rotated_mac_buf + ((0 - (size_t)rotated_mac_buf) & 63);
#endif
/* This information is public so it's safe to branch based on it. */
if (orig_len > md_size + 255 + 1) {
scan_start = orig_len - (md_size + 255 + 1);
}
/* div_spoiler contains a multiple of md_size that is used to cause the
* modulo operation to be constant time. Without this, the time varies
* based on the amount of padding when running on Intel chips at least.
*
* The aim of right-shifting md_size is so that the compiler doesn't
* figure out that it can remove div_spoiler as that would require it
* to prove that md_size is always even, which I hope is beyond it. */
div_spoiler = md_size >> 1;
div_spoiler <<= (sizeof(div_spoiler) - 1) * 8;
rotate_offset = (div_spoiler + mac_start - scan_start) % md_size;
memset(rotated_mac, 0, md_size);
for (i = scan_start, j = 0; i < orig_len; i++) {
uint8_t mac_started = constant_time_ge_8(i, mac_start);
uint8_t mac_ended = constant_time_ge_8(i, mac_end);
uint8_t b = in[i];
rotated_mac[j++] |= b & mac_started & ~mac_ended;
j &= constant_time_lt(j, md_size);
}
/* Now rotate the MAC */
#if defined(CBC_MAC_ROTATE_IN_PLACE)
j = 0;
for (i = 0; i < md_size; i++) {
/* in case cache-line is 32 bytes, touch second line */
((volatile uint8_t *)rotated_mac)[rotate_offset ^ 32];
out[j++] = rotated_mac[rotate_offset++];
rotate_offset &= constant_time_lt(rotate_offset, md_size);
}
#else
memset(out, 0, md_size);
rotate_offset = md_size - rotate_offset;
rotate_offset &= constant_time_lt(rotate_offset, md_size);
for (i = 0; i < md_size; i++) {
for (j = 0; j < md_size; j++) {
out[j] |= rotated_mac[i] & constant_time_eq_8(j, rotate_offset);
}
rotate_offset++;
rotate_offset &= constant_time_lt(rotate_offset, md_size);
}
#endif
}
