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 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);
}
/* Ideally the next statement would be:
*
* rotate_offset = (mac_start - scan_start) % md_size;
*
* However, division is not a constant-time operation (at least on Intel
* chips). Thus we enumerate the possible values of md_size and handle each
* separately. The value of |md_size| is public information (it's determined
* by the cipher suite in the ServerHello) so our timing can vary based on
* its value. */
rotate_offset = mac_start - scan_start;
/* rotate_offset can be, at most, 255 (bytes of padding) + 1 (padding length)
* + md_size = 256 + 48 (since SHA-384 is the largest hash) = 304. */
assert(rotate_offset <= 304);
if (md_size == 16) {
rotate_offset &= 15;
} else if (md_size == 20) {
/* 1/20 is approximated as 25/512 and then Barrett reduction is used.
* Analytically, this is correct for 0 <= rotate_offset <= 853. */
unsigned q = (rotate_offset * 25) >> 9;
rotate_offset -= q * 20;
rotate_offset -=
constant_time_select(constant_time_ge(rotate_offset, 20), 20, 0);
} else if (md_size == 32) {
int EVP_tls_cbc_remove_padding(unsigned *out_len,
const uint8_t *in, unsigned in_len,
unsigned block_size, unsigned mac_size) {
unsigned padding_length, good, to_check, i;
const unsigned overhead = 1 /* padding length byte */ + mac_size;
/* These lengths are all public so we can test them in non-constant time. */
if (overhead > in_len) {
return 0;
}
padding_length = in[in_len - 1];
good = constant_time_ge(in_len, overhead + padding_length);
/* The padding consists of a length byte at the end of the record and
* then that many bytes of padding, all with the same value as the
* length byte. Thus, with the length byte included, there are i+1
* bytes of padding.
*
* We can't check just |padding_length+1| bytes because that leaks
* decrypted information. Therefore we always have to check the maximum
* amount of padding possible. (Again, the length of the record is
* public information so we can use it.) */
to_check = 256; /* maximum amount of padding, inc length byte. */
if (to_check > in_len) {
to_check = in_len;
}
for (i = 0; i < to_check; i++) {
uint8_t mask = constant_time_ge_8(padding_length, i);
uint8_t b = in[in_len - 1 - i];
/* The final |padding_length+1| bytes should all have the value
* |padding_length|. Therefore the XOR should be zero. */
good &= ~(mask & (padding_length ^ b));
}
/* If any of the final |padding_length+1| bytes had the wrong value,
* one or more of the lower eight bits of |good| will be cleared. */
good = constant_time_eq(0xff, good & 0xff);
/* Always treat |padding_length| as zero on error. If, assuming block size of
* 16, a padding of [<15 arbitrary bytes> 15] treated |padding_length| as 16
* and returned -1, distinguishing good MAC and bad padding from bad MAC and
* bad padding would give POODLE's padding oracle. */
padding_length = good & (padding_length + 1);
*out_len = in_len - padding_length;
return constant_time_select_int(good, 1, -1);
}
int RSA_padding_check_PKCS1_type_2(uint8_t *to, unsigned to_len,
const uint8_t *from, unsigned from_len) {
if (from_len == 0) {
OPENSSL_PUT_ERROR(RSA, RSA_R_EMPTY_PUBLIC_KEY);
return -1;
}
/* PKCS#1 v1.5 decryption. See "PKCS #1 v2.2: RSA Cryptography
* Standard", section 7.2.2. */
if (from_len < RSA_PKCS1_PADDING_SIZE) {
/* |from| is zero-padded to the size of the RSA modulus, a public value, so
* this can be rejected in non-constant time. */
OPENSSL_PUT_ERROR(RSA, RSA_R_KEY_SIZE_TOO_SMALL);
return -1;
}
unsigned first_byte_is_zero = constant_time_eq(from[0], 0);
unsigned second_byte_is_two = constant_time_eq(from[1], 2);
unsigned i, zero_index = 0, looking_for_index = ~0u;
for (i = 2; i < from_len; i++) {
unsigned equals0 = constant_time_is_zero(from[i]);
zero_index = constant_time_select(looking_for_index & equals0, (unsigned)i,
zero_index);
looking_for_index = constant_time_select(equals0, 0, looking_for_index);
}
/* The input must begin with 00 02. */
unsigned valid_index = first_byte_is_zero;
valid_index &= second_byte_is_two;
/* We must have found the end of PS. */
valid_index &= ~looking_for_index;
/* PS must be at least 8 bytes long, and it starts two bytes into |from|. */
valid_index &= constant_time_ge(zero_index, 2 + 8);
/* Skip the zero byte. */
zero_index++;
/* NOTE: Although this logic attempts to be constant time, the API contracts
* of this function and |RSA_decrypt| with |RSA_PKCS1_PADDING| make it
* impossible to completely avoid Bleichenbacher's attack. Consumers should
* use |RSA_unpad_key_pkcs1|. */
if (!valid_index) {
OPENSSL_PUT_ERROR(RSA, RSA_R_PKCS_DECODING_ERROR);
return -1;
}
const unsigned msg_len = from_len - zero_index;
if (msg_len > to_len) {
/* This shouldn't happen because this function is always called with
* |to_len| as the key size and |from_len| is bounded by the key size. */
OPENSSL_PUT_ERROR(RSA, RSA_R_PKCS_DECODING_ERROR);
return -1;
}
if (msg_len > INT_MAX) {
OPENSSL_PUT_ERROR(RSA, ERR_R_OVERFLOW);
return -1;
}
memcpy(to, &from[zero_index], msg_len);
return (int)msg_len;
}
int RSA_padding_check_PKCS1_type_2(unsigned char *to, int tlen,
const unsigned char *from, int flen,
int num)
{
int i;
/* |em| is the encoded message, zero-padded to exactly |num| bytes */
unsigned char *em = NULL;
unsigned int good, found_zero_byte;
int zero_index = 0, msg_index, mlen = -1;
if (tlen < 0 || flen < 0)
return -1;
/*
* PKCS#1 v1.5 decryption. See "PKCS #1 v2.2: RSA Cryptography Standard",
* section 7.2.2.
*/
if (flen > num)
goto err;
if (num < 11)
goto err;
em = OPENSSL_zalloc(num);
if (em == NULL) {
RSAerr(RSA_F_RSA_PADDING_CHECK_PKCS1_TYPE_2, ERR_R_MALLOC_FAILURE);
return -1;
}
/*
* Always do this zero-padding copy (even when num == flen) to avoid
* leaking that information. The copy still leaks some side-channel
* information, but it's impossible to have a fixed memory access
* pattern since we can't read out of the bounds of |from|.
*
* TODO(emilia): Consider porting BN_bn2bin_padded from BoringSSL.
*/
memcpy(em + num - flen, from, flen);
good = constant_time_is_zero(em[0]);
good &= constant_time_eq(em[1], 2);
found_zero_byte = 0;
for (i = 2; i < num; i++) {
unsigned int equals0 = constant_time_is_zero(em[i]);
zero_index =
constant_time_select_int(~found_zero_byte & equals0, i,
zero_index);
found_zero_byte |= equals0;
}
/*
* PS must be at least 8 bytes long, and it starts two bytes into |em|.
* If we never found a 0-byte, then |zero_index| is 0 and the check
* also fails.
*/
good &= constant_time_ge((unsigned int)(zero_index), 2 + 8);
/*
* Skip the zero byte. This is incorrect if we never found a zero-byte
* but in this case we also do not copy the message out.
*/
msg_index = zero_index + 1;
mlen = num - msg_index;
/*
* For good measure, do this check in constant time as well; it could
* leak something if |tlen| was assuming valid padding.
*/
good &= constant_time_ge((unsigned int)(tlen), (unsigned int)(mlen));
/*
* We can't continue in constant-time because we need to copy the result
* and we cannot fake its length. This unavoidably leaks timing
* information at the API boundary.
* TODO(emilia): this could be addressed at the call site,
* see BoringSSL commit 0aa0767340baf925bda4804882aab0cb974b2d26.
*/
if (!good) {
mlen = -1;
goto err;
}
memcpy(to, em + msg_index, mlen);
err:
OPENSSL_free(em);
if (mlen == -1)
RSAerr(RSA_F_RSA_PADDING_CHECK_PKCS1_TYPE_2,
RSA_R_PKCS_DECODING_ERROR);
return mlen;
}
int RSA_padding_check_PKCS1_type_2(unsigned char *to, int tlen,
const unsigned char *from, int flen,
int num)
{
int i;
/* |em| is the encoded message, zero-padded to exactly |num| bytes */
unsigned char *em = NULL;
unsigned int good, found_zero_byte, mask;
int zero_index = 0, msg_index, mlen = -1;
if (tlen < 0 || flen < 0)
return -1;
/*
* PKCS#1 v1.5 decryption. See "PKCS #1 v2.2: RSA Cryptography Standard",
* section 7.2.2.
*/
if (flen > num || num < 11) {
RSAerr(RSA_F_RSA_PADDING_CHECK_PKCS1_TYPE_2,
RSA_R_PKCS_DECODING_ERROR);
return -1;
}
em = OPENSSL_malloc(num);
if (em == NULL) {
RSAerr(RSA_F_RSA_PADDING_CHECK_PKCS1_TYPE_2, ERR_R_MALLOC_FAILURE);
return -1;
}
/*
* Caller is encouraged to pass zero-padded message created with
* BN_bn2binpad. Trouble is that since we can't read out of |from|'s
* bounds, it's impossible to have an invariant memory access pattern
* in case |from| was not zero-padded in advance.
*/
for (from += flen, em += num, i = 0; i < num; i++) {
mask = ~constant_time_is_zero(flen);
flen -= 1 & mask;
from -= 1 & mask;
*--em = *from & mask;
}
from = em;
good = constant_time_is_zero(from[0]);
good &= constant_time_eq(from[1], 2);
/* scan over padding data */
found_zero_byte = 0;
for (i = 2; i < num; i++) {
unsigned int equals0 = constant_time_is_zero(from[i]);
zero_index = constant_time_select_int(~found_zero_byte & equals0,
i, zero_index);
found_zero_byte |= equals0;
}
/*
* PS must be at least 8 bytes long, and it starts two bytes into |from|.
* If we never found a 0-byte, then |zero_index| is 0 and the check
* also fails.
*/
good &= constant_time_ge(zero_index, 2 + 8);
/*
* Skip the zero byte. This is incorrect if we never found a zero-byte
* but in this case we also do not copy the message out.
*/
msg_index = zero_index + 1;
mlen = num - msg_index;
/*
* For good measure, do this check in constant time as well.
*/
good &= constant_time_ge(tlen, mlen);
/*
* Even though we can't fake result's length, we can pretend copying
* |tlen| bytes where |mlen| bytes would be real. Last |tlen| of |num|
* bytes are viewed as circular buffer with start at |tlen|-|mlen'|,
* where |mlen'| is "saturated" |mlen| value. Deducing information
* about failure or |mlen| would take attacker's ability to observe
* memory access pattern with byte granularity *as it occurs*. It
* should be noted that failure is indistinguishable from normal
* operation if |tlen| is fixed by protocol.
*/
tlen = constant_time_select_int(constant_time_lt(num, tlen), num, tlen);
msg_index = constant_time_select_int(good, msg_index, num - tlen);
mlen = num - msg_index;
for (from += msg_index, mask = good, i = 0; i < tlen; i++) {
unsigned int equals = constant_time_eq(i, mlen);
from -= tlen & equals; /* if (i == mlen) rewind */
mask &= mask ^ equals; /* if (i == mlen) mask = 0 */
to[i] = constant_time_select_8(mask, from[i], to[i]);
}
OPENSSL_cleanse(em, num);
OPENSSL_free(em);
RSAerr(RSA_F_RSA_PADDING_CHECK_PKCS1_TYPE_2, RSA_R_PKCS_DECODING_ERROR);
err_clear_last_constant_time(1 & good);
return constant_time_select_int(good, mlen, -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 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);
}
/* Ideally the next statement would be:
*
* rotate_offset = (mac_start - scan_start) % md_size;
*
* However, division is not a constant-time operation (at least on Intel
* chips). Thus we enumerate the possible values of md_size and handle each
* separately. The value of |md_size| is public information (it's determined
* by the cipher suite in the ServerHello) so our timing can vary based on
* its value. */
rotate_offset = mac_start - scan_start;
/* rotate_offset can be, at most, 255 (bytes of padding) + 1 (padding length)
* + md_size = 256 + 48 (since SHA-384 is the largest hash) = 304. */
assert(rotate_offset <= 304);
/* Below is an SMT-LIB2 verification that the Barrett reductions below are
* correct within this range:
*
* (define-fun barrett (
* (x (_ BitVec 32))
* (mul (_ BitVec 32))
* (shift (_ BitVec 32))
* (divisor (_ BitVec 32)) ) (_ BitVec 32)
* (let ((q (bvsub x (bvmul divisor (bvlshr (bvmul x mul) shift))) ))
* (ite (bvuge q divisor)
* (bvsub q divisor)
* q)))
*
* (declare-fun x () (_ BitVec 32))
*
* (assert (or
* (let (
* (divisor (_ bv20 32))
* (mul (_ bv25 32))
* (shift (_ bv9 32))
* (limit (_ bv853 32)))
*
* (and (bvule x limit) (not (= (bvurem x divisor)
* (barrett x mul shift divisor)))))
*
* (let (
* (divisor (_ bv48 32))
* (mul (_ bv10 32))
* (shift (_ bv9 32))
* (limit (_ bv768 32)))
*
* (and (bvule x limit) (not (= (bvurem x divisor)
* (barrett x mul shift divisor)))))
* ))
*
* (check-sat)
* (get-model)
*/
if (md_size == 16) {
rotate_offset &= 15;
} else if (md_size == 20) {
/* 1/20 is approximated as 25/512 and then Barrett reduction is used.
* Analytically, this is correct for 0 <= rotate_offset <= 853. */
unsigned q = (rotate_offset * 25) >> 9;
rotate_offset -= q * 20;
rotate_offset -=
constant_time_select(constant_time_ge(rotate_offset, 20), 20, 0);
} else if (md_size == 32) {
