*
*/
+#include "../crypto/constant_time_locl.h"
#include "ssl_locl.h"
#include <openssl/md5.h>
* supported by TLS.) */
#define MAX_HASH_BLOCK_SIZE 128
-/* Some utility functions are needed:
- *
- * These macros return the given value with the MSB copied to all the other
- * bits. They use the fact that arithmetic shift shifts-in the sign bit.
- * However, this is not ensured by the C standard so you may need to replace
- * them with something else on odd CPUs. */
-#define DUPLICATE_MSB_TO_ALL(x) ( (unsigned)( (int)(x) >> (sizeof(int)*8-1) ) )
-#define DUPLICATE_MSB_TO_ALL_8(x) ((unsigned char)(DUPLICATE_MSB_TO_ALL(x)))
-
-/* constant_time_lt returns 0xff if a<b and 0x00 otherwise. */
-static unsigned constant_time_lt(unsigned a, unsigned b)
- {
- a -= b;
- return DUPLICATE_MSB_TO_ALL(a);
- }
-
-/* constant_time_ge returns 0xff if a>=b and 0x00 otherwise. */
-static unsigned constant_time_ge(unsigned a, unsigned b)
- {
- a -= b;
- return DUPLICATE_MSB_TO_ALL(~a);
- }
-
-/* constant_time_eq_8 returns 0xff if a==b and 0x00 otherwise. */
-static unsigned char constant_time_eq_8(unsigned a, unsigned b)
- {
- unsigned c = a ^ b;
- c--;
- return DUPLICATE_MSB_TO_ALL_8(c);
- }
-
/* ssl3_cbc_remove_padding removes padding from the decrypted, SSLv3, CBC
* record in |rec| by updating |rec->length| in constant time.
*
rec->length -= padding_length;
rec->type |= padding_length<<8; /* kludge: pass padding length */
return (int)((good & 1) | (~good & -1));
-}
+ }
/* tls1_cbc_remove_padding removes the CBC padding from the decrypted, TLS, CBC
* record in |rec| in constant time and returns 1 if the padding is valid and
unsigned padding_length, good, to_check, i;
const unsigned overhead = 1 /* padding length byte */ + mac_size;
/* Check if version requires explicit IV */
- if (s->version >= TLS1_1_VERSION || s->version == DTLS1_VERSION)
+ if (SSL_USE_EXPLICIT_IV(s))
{
/* These lengths are all public so we can test them in
* non-constant time.
for (i = 0; i < to_check; i++)
{
- unsigned char mask = constant_time_ge(padding_length, i);
+ unsigned char mask = constant_time_ge_8(padding_length, i);
unsigned char b = rec->data[rec->length-1-i];
/* The final |padding_length+1| bytes should all have the value
* |padding_length|. Therefore the XOR should be zero. */
}
/* 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. We
- * AND the bottom 8 bits together and duplicate the result to all the
- * bits. */
- good &= good >> 4;
- good &= good >> 2;
- good &= good >> 1;
- good <<= sizeof(good)*8-1;
- good = DUPLICATE_MSB_TO_ALL(good);
-
+ * one or more of the lower eight bits of |good| will be cleared.
+ */
+ good = constant_time_eq(0xff, good & 0xff);
padding_length = good & (padding_length+1);
rec->length -= padding_length;
rec->type |= padding_length<<8; /* kludge: pass padding length */
return (int)((good & 1) | (~good & -1));
}
-#if defined(_M_AMD64) || defined(__x86_64__)
-#define CBC_MAC_ROTATE_IN_PLACE
-#endif
-
/* ssl3_cbc_copy_mac copies |md_size| bytes from the end of |rec| to |out| in
* constant time (independent of the concrete value of rec->length, which may
* vary within a 256-byte window).
*
* If CBC_MAC_ROTATE_IN_PLACE is defined then the rotation is performed with
* variable accesses in a 64-byte-aligned buffer. Assuming that this fits into
- * a single cache-line, then the variable memory accesses don't actually affect
- * the timing. This has been tested to be true on Intel amd64 chips.
+ * a single or pair of cache-lines, then the variable memory accesses don't
+ * actually affect the timing. CPUs with smaller cache-lines [if any] are
+ * not multi-core and are not considered vulnerable to cache-timing attacks.
*/
+#define CBC_MAC_ROTATE_IN_PLACE
+
void ssl3_cbc_copy_mac(unsigned char* out,
const SSL3_RECORD *rec,
unsigned md_size,unsigned orig_len)
{
#if defined(CBC_MAC_ROTATE_IN_PLACE)
- unsigned char rotated_mac_buf[EVP_MAX_MD_SIZE*2];
+ unsigned char rotated_mac_buf[64+EVP_MAX_MD_SIZE];
unsigned char *rotated_mac;
#else
unsigned char rotated_mac[EVP_MAX_MD_SIZE];
OPENSSL_assert(md_size <= EVP_MAX_MD_SIZE);
#if defined(CBC_MAC_ROTATE_IN_PLACE)
- rotated_mac = (unsigned char*) (((intptr_t)(rotated_mac_buf + 64)) & ~63);
+ 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. */
memset(rotated_mac, 0, md_size);
for (i = scan_start, j = 0; i < orig_len; i++)
{
- unsigned char mac_started = constant_time_ge(i, mac_start);
- unsigned char mac_ended = constant_time_ge(i, mac_end);
+ unsigned char mac_started = constant_time_ge_8(i, mac_start);
+ unsigned char mac_ended = constant_time_ge_8(i, mac_end);
unsigned char b = rec->data[i];
rotated_mac[j++] |= b & mac_started & ~mac_ended;
j &= constant_time_lt(j,md_size);
j = 0;
for (i = 0; i < md_size; i++)
{
+ /* in case cache-line is 32 bytes, touch second line */
+ ((volatile unsigned char *)rotated_mac)[rotate_offset^32];
out[j++] = rotated_mac[rotate_offset++];
rotate_offset &= constant_time_lt(rotate_offset,md_size);
}
b = data[k-header_length];
k++;
- is_past_c = is_block_a & constant_time_ge(j, c);
- is_past_cp1 = is_block_a & constant_time_ge(j, c+1);
+ is_past_c = is_block_a & constant_time_ge_8(j, c);
+ 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. */