*
*/
-#include <stdint.h>
-
#include "ssl_locl.h"
#include <openssl/md5.h>
#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)
{
}
/* constant_time_eq_8 returns 0xff if a==b and 0x00 otherwise. */
-static unsigned char constant_time_eq_8(unsigned char a, unsigned char b)
+static unsigned char constant_time_eq_8(unsigned a, unsigned b)
{
unsigned c = a ^ b;
c--;
good = constant_time_ge(rec->length, padding_length+overhead);
/* SSLv3 requires that the padding is minimal. */
good &= constant_time_ge(block_size, padding_length+1);
- rec->length -= good & (padding_length+1);
+ 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));
}
unsigned mac_size)
{
unsigned padding_length, good, to_check, i;
- const char has_explicit_iv =
- s->version >= TLS1_1_VERSION || s->version == DTLS1_VERSION;
- const unsigned overhead = 1 /* padding length byte */ +
- mac_size +
- (has_explicit_iv ? block_size : 0);
-
- /* These lengths are all public so we can test them in non-constant
- * time. */
- if (overhead > rec->length)
+ const unsigned overhead = 1 /* padding length byte */ + mac_size;
+ /* Check if version requires explicit IV */
+ if (s->version >= TLS1_1_VERSION || s->version == DTLS1_BAD_VER)
+ {
+ /* These lengths are all public so we can test them in
+ * non-constant time.
+ */
+ if (overhead + block_size > rec->length)
+ return 0;
+ /* We can now safely skip explicit IV */
+ rec->data += block_size;
+ rec->input += block_size;
+ rec->length -= block_size;
+ }
+ else if (overhead > rec->length)
return 0;
padding_length = rec->data[rec->length-1];
}
}
+ if (EVP_CIPHER_flags(s->enc_read_ctx->cipher)&EVP_CIPH_FLAG_AEAD_CIPHER)
+ {
+ /* padding is already verified */
+ rec->length -= padding_length + 1;
+ return 1;
+ }
+
good = constant_time_ge(rec->length, 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
good <<= sizeof(good)*8-1;
good = DUPLICATE_MSB_TO_ALL(good);
- rec->length -= good & (padding_length+1);
-
- /* We can always safely skip the explicit IV. We check at the beginning
- * of this function that the record has at least enough space for the
- * IV, MAC and padding length byte. (These can be checked in
- * non-constant time because it's all public information.) So, if the
- * padding was invalid, then we didn't change |rec->length| and this is
- * safe. If the padding was valid then we know that we have at least
- * overhead+padding_length bytes of space and so this is still safe
- * because overhead accounts for the explicit IV. */
- if (has_explicit_iv)
- {
- rec->data += block_size;
- rec->input += block_size;
- rec->length -= block_size;
- rec->orig_len -= block_size;
- }
+ 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 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];
unsigned div_spoiler;
unsigned rotate_offset;
- OPENSSL_assert(rec->orig_len >= md_size);
+ OPENSSL_assert(orig_len >= 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. */
- if (rec->orig_len > md_size + 255 + 1)
- scan_start = rec->orig_len - (md_size + 255 + 1);
+ 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.
rotate_offset = (div_spoiler + mac_start - scan_start) % md_size;
memset(rotated_mac, 0, md_size);
- for (i = scan_start; i < rec->orig_len;)
+ for (i = scan_start, j = 0; i < orig_len; i++)
{
- for (j = 0; j < md_size && i < rec->orig_len; i++, j++)
- {
- unsigned char mac_started = constant_time_ge(i, mac_start);
- unsigned char mac_ended = constant_time_ge(i, mac_end);
- unsigned char b = 0;
- b = rec->data[i];
- rotated_mac[j] |= b & mac_started & ~mac_ended;
- }
+ unsigned char mac_started = constant_time_ge(i, mac_start);
+ unsigned char mac_ended = constant_time_ge(i, mac_end);
+ unsigned char b = rec->data[i];
+ rotated_mac[j++] |= b & mac_started & ~mac_ended;
+ j &= constant_time_lt(j,md_size);
}
/* Now rotate the MAC */
j = 0;
for (i = 0; i < md_size; i++)
{
- unsigned char offset = (div_spoiler + rotate_offset + i) % md_size;
- out[j++] = rotated_mac[offset];
+ /* 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);
}
#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++)
{
- unsigned char offset = (div_spoiler + md_size - rotate_offset + i) % md_size;
for (j = 0; j < md_size; j++)
- out[j] |= rotated_mac[i] & constant_time_eq_8(j, offset);
+ out[j] |= rotated_mac[i] & constant_time_eq_8(j, rotate_offset);
+ rotate_offset++;
+ rotate_offset &= constant_time_lt(rotate_offset,md_size);
}
#endif
}
+/* u32toLE serialises an unsigned, 32-bit number (n) as four bytes at (p) in
+ * little-endian order. The value of p is advanced by four. */
+#define u32toLE(n, p) \
+ (*((p)++)=(unsigned char)(n), \
+ *((p)++)=(unsigned char)(n>>8), \
+ *((p)++)=(unsigned char)(n>>16), \
+ *((p)++)=(unsigned char)(n>>24))
+
/* These functions serialize the state of a hash and thus perform the standard
* "final" operation without adding the padding and length that such a function
* typically does. */
static void tls1_md5_final_raw(void* ctx, unsigned char *md_out)
{
MD5_CTX *md5 = ctx;
- l2n(md5->A, md_out);
- l2n(md5->B, md_out);
- l2n(md5->C, md_out);
- l2n(md5->D, md_out);
+ u32toLE(md5->A, md_out);
+ u32toLE(md5->B, md_out);
+ u32toLE(md5->C, md_out);
+ u32toLE(md5->D, md_out);
}
static void tls1_sha1_final_raw(void* ctx, unsigned char *md_out)
l2n(sha1->h3, md_out);
l2n(sha1->h4, md_out);
}
+#define LARGEST_DIGEST_CTX SHA_CTX
+#ifndef OPENSSL_NO_SHA256
static void tls1_sha256_final_raw(void* ctx, unsigned char *md_out)
{
SHA256_CTX *sha256 = ctx;
l2n(sha256->h[i], md_out);
}
}
+#undef LARGEST_DIGEST_CTX
+#define LARGEST_DIGEST_CTX SHA256_CTX
+#endif
+#ifndef OPENSSL_NO_SHA512
static void tls1_sha512_final_raw(void* ctx, unsigned char *md_out)
{
SHA512_CTX *sha512 = ctx;
l2n8(sha512->h[i], md_out);
}
}
+#undef LARGEST_DIGEST_CTX
+#define LARGEST_DIGEST_CTX SHA512_CTX
+#endif
/* ssl3_cbc_record_digest_supported returns 1 iff |ctx| uses a hash function
* which ssl3_cbc_digest_record supports. */
if (FIPS_mode())
return 0;
#endif
- switch (ctx->digest->type)
+ switch (EVP_MD_CTX_type(ctx))
{
case NID_md5:
case NID_sha1:
+#ifndef OPENSSL_NO_SHA256
case NID_sha224:
case NID_sha256:
+#endif
+#ifndef OPENSSL_NO_SHA512
case NID_sha384:
case NID_sha512:
+#endif
return 1;
default:
return 0;
unsigned mac_secret_length,
char is_sslv3)
{
- unsigned char md_state[sizeof(SHA512_CTX)];
+ union { double align;
+ unsigned char c[sizeof(LARGEST_DIGEST_CTX)]; } md_state;
void (*md_final_raw)(void *ctx, unsigned char *md_out);
void (*md_transform)(void *ctx, const unsigned char *block);
unsigned md_size, md_block_size = 64;
unsigned sslv3_pad_length = 40, header_length, variance_blocks,
len, max_mac_bytes, num_blocks,
num_starting_blocks, k, mac_end_offset, c, index_a, index_b;
- uint64_t bits;
+ unsigned int bits; /* at most 18 bits */
unsigned char length_bytes[MAX_HASH_BIT_COUNT_BYTES];
/* hmac_pad is the masked HMAC key. */
unsigned char hmac_pad[MAX_HASH_BLOCK_SIZE];
/* mdLengthSize is the number of bytes in the length field that terminates
* the hash. */
unsigned md_length_size = 8;
+ char length_is_big_endian = 1;
/* This is a, hopefully redundant, check that allows us to forget about
* many possible overflows later in this function. */
OPENSSL_assert(data_plus_mac_plus_padding_size < 1024*1024);
- switch (ctx->digest->type)
+ switch (EVP_MD_CTX_type(ctx))
{
case NID_md5:
- MD5_Init((MD5_CTX*)md_state);
+ MD5_Init((MD5_CTX*)md_state.c);
md_final_raw = tls1_md5_final_raw;
md_transform = (void(*)(void *ctx, const unsigned char *block)) MD5_Transform;
md_size = 16;
sslv3_pad_length = 48;
+ length_is_big_endian = 0;
break;
case NID_sha1:
- SHA1_Init((SHA_CTX*)md_state);
+ SHA1_Init((SHA_CTX*)md_state.c);
md_final_raw = tls1_sha1_final_raw;
md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA1_Transform;
md_size = 20;
break;
+#ifndef OPENSSL_NO_SHA256
case NID_sha224:
- SHA224_Init((SHA256_CTX*)md_state);
+ SHA224_Init((SHA256_CTX*)md_state.c);
md_final_raw = tls1_sha256_final_raw;
md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA256_Transform;
md_size = 224/8;
break;
case NID_sha256:
- SHA256_Init((SHA256_CTX*)md_state);
+ SHA256_Init((SHA256_CTX*)md_state.c);
md_final_raw = tls1_sha256_final_raw;
md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA256_Transform;
md_size = 32;
break;
+#endif
+#ifndef OPENSSL_NO_SHA512
case NID_sha384:
- SHA384_Init((SHA512_CTX*)md_state);
+ SHA384_Init((SHA512_CTX*)md_state.c);
md_final_raw = tls1_sha512_final_raw;
md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA512_Transform;
md_size = 384/8;
md_length_size = 16;
break;
case NID_sha512:
- SHA512_Init((SHA512_CTX*)md_state);
+ SHA512_Init((SHA512_CTX*)md_state.c);
md_final_raw = tls1_sha512_final_raw;
md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA512_Transform;
md_size = 64;
md_block_size = 128;
md_length_size = 16;
break;
+#endif
default:
/* ssl3_cbc_record_digest_supported should have been
* called first to check that the hash function is
for (i = 0; i < md_block_size; i++)
hmac_pad[i] ^= 0x36;
- md_transform(md_state, hmac_pad);
+ md_transform(md_state.c, hmac_pad);
}
- j = 0;
- if (md_length_size == 16)
+ if (length_is_big_endian)
{
- memset(length_bytes, 0, 8);
- j = 8;
+ memset(length_bytes,0,md_length_size-4);
+ length_bytes[md_length_size-4] = (unsigned char)(bits>>24);
+ length_bytes[md_length_size-3] = (unsigned char)(bits>>16);
+ length_bytes[md_length_size-2] = (unsigned char)(bits>>8);
+ length_bytes[md_length_size-1] = (unsigned char)bits;
+ }
+ else
+ {
+ memset(length_bytes,0,md_length_size);
+ length_bytes[md_length_size-5] = (unsigned char)(bits>>24);
+ length_bytes[md_length_size-6] = (unsigned char)(bits>>16);
+ length_bytes[md_length_size-7] = (unsigned char)(bits>>8);
+ length_bytes[md_length_size-8] = (unsigned char)bits;
}
- for (i = 0; i < 8; i++)
- length_bytes[i+j] = bits >> (8*(7-i));
if (k > 0)
{
* block that the header consumes: either 7 bytes
* (SHA1) or 11 bytes (MD5). */
unsigned overhang = header_length-md_block_size;
- md_transform(md_state, header);
+ md_transform(md_state.c, header);
memcpy(first_block, header + md_block_size, overhang);
memcpy(first_block + overhang, data, md_block_size-overhang);
- md_transform(md_state, first_block);
+ md_transform(md_state.c, first_block);
for (i = 1; i < k/md_block_size - 1; i++)
- md_transform(md_state, data + md_block_size*i - overhang);
+ md_transform(md_state.c, data + md_block_size*i - overhang);
}
else
{
/* k is a multiple of md_block_size. */
memcpy(first_block, header, 13);
memcpy(first_block+13, data, md_block_size-13);
- md_transform(md_state, first_block);
+ md_transform(md_state.c, first_block);
for (i = 1; i < k/md_block_size; i++)
- md_transform(md_state, data + md_block_size*i - 13);
+ md_transform(md_state.c, data + md_block_size*i - 13);
}
}
block[j] = b;
}
- md_transform(md_state, block);
- md_final_raw(md_state, block);
+ md_transform(md_state.c, block);
+ md_final_raw(md_state.c, block);
/* If this is index_b, copy the hash value to |mac_out|. */
for (j = 0; j < md_size; j++)
mac_out[j] |= block[j]&is_block_b;