crypto: aes/fixed-time - align key schedule with other implementations

The fixed time AES code mangles the key schedule so that xoring the
first round key with values at fixed offsets across the Sbox produces
the correct value. This primes the D-cache with the entire Sbox before
any data dependent lookups are done, making it more difficult to infer
key bits from timing variances when the plaintext is known.

The downside of this approach is that it renders the key schedule
incompatible with other implementations of AES in the kernel, which
makes it cumbersome to use this implementation as a fallback for SIMD
based AES in contexts where this is not allowed.

So let's tweak the fixed Sbox indexes so that they add up to zero under
the xor operation. While at it, increase the granularity to 16 bytes so
we cover the entire Sbox even on systems with 16 byte cachelines.

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

diff --git a/crypto/aes_ti.c b/crypto/aes_ti.c
index 798fc9a..b3ebdc5 100644 (file)
--- a/crypto/aes_ti.c
+++ b/crypto/aes_ti.c
@@ -234,30 +234,8 @@ static int aesti_set_key(struct crypto_tfm *tfm, const u8 *in_key,
                         unsigned int key_len)
 {
        struct crypto_aes_ctx *ctx = crypto_tfm_ctx(tfm);
-       int err;
 
-       err = aesti_expand_key(ctx, in_key, key_len);
-       if (err)
-               return err;
-
-       /*
-        * In order to force the compiler to emit data independent Sbox lookups
-        * at the start of each block, xor the first round key with values at
-        * fixed indexes in the Sbox. This will need to be repeated each time
-        * the key is used, which will pull the entire Sbox into the D-cache
-        * before any data dependent Sbox lookups are performed.
-        */
-       ctx->key_enc[0] ^= __aesti_sbox[ 0] ^ __aesti_sbox[128];
-       ctx->key_enc[1] ^= __aesti_sbox[32] ^ __aesti_sbox[160];
-       ctx->key_enc[2] ^= __aesti_sbox[64] ^ __aesti_sbox[192];
-       ctx->key_enc[3] ^= __aesti_sbox[96] ^ __aesti_sbox[224];
-
-       ctx->key_dec[0] ^= __aesti_inv_sbox[ 0] ^ __aesti_inv_sbox[128];
-       ctx->key_dec[1] ^= __aesti_inv_sbox[32] ^ __aesti_inv_sbox[160];
-       ctx->key_dec[2] ^= __aesti_inv_sbox[64] ^ __aesti_inv_sbox[192];
-       ctx->key_dec[3] ^= __aesti_inv_sbox[96] ^ __aesti_inv_sbox[224];
-
-       return 0;
+       return aesti_expand_key(ctx, in_key, key_len);
 }
 
 static void aesti_encrypt(struct crypto_tfm *tfm, u8 *out, const u8 *in)
@@ -280,10 +258,16 @@ static void aesti_encrypt(struct crypto_tfm *tfm, u8 *out, const u8 *in)
         */
        local_irq_save(flags);
 
-       st0[0] ^= __aesti_sbox[ 0] ^ __aesti_sbox[128];
-       st0[1] ^= __aesti_sbox[32] ^ __aesti_sbox[160];
-       st0[2] ^= __aesti_sbox[64] ^ __aesti_sbox[192];
-       st0[3] ^= __aesti_sbox[96] ^ __aesti_sbox[224];
+       /*
+        * Force the compiler to emit data independent Sbox references,
+        * by xoring the input with Sbox values that are known to add up
+        * to zero. This pulls the entire Sbox into the D-cache before any
+        * data dependent lookups are done.
+        */
+       st0[0] ^= __aesti_sbox[ 0] ^ __aesti_sbox[ 64] ^ __aesti_sbox[134] ^ __aesti_sbox[195];
+       st0[1] ^= __aesti_sbox[16] ^ __aesti_sbox[ 82] ^ __aesti_sbox[158] ^ __aesti_sbox[221];
+       st0[2] ^= __aesti_sbox[32] ^ __aesti_sbox[ 96] ^ __aesti_sbox[160] ^ __aesti_sbox[234];
+       st0[3] ^= __aesti_sbox[48] ^ __aesti_sbox[112] ^ __aesti_sbox[186] ^ __aesti_sbox[241];
 
        for (round = 0;; round += 2, rkp += 8) {
                st1[0] = mix_columns(subshift(st0, 0)) ^ rkp[0];
@@ -328,10 +312,16 @@ static void aesti_decrypt(struct crypto_tfm *tfm, u8 *out, const u8 *in)
         */
        local_irq_save(flags);
 
-       st0[0] ^= __aesti_inv_sbox[ 0] ^ __aesti_inv_sbox[128];
-       st0[1] ^= __aesti_inv_sbox[32] ^ __aesti_inv_sbox[160];
-       st0[2] ^= __aesti_inv_sbox[64] ^ __aesti_inv_sbox[192];
-       st0[3] ^= __aesti_inv_sbox[96] ^ __aesti_inv_sbox[224];
+       /*
+        * Force the compiler to emit data independent Sbox references,
+        * by xoring the input with Sbox values that are known to add up
+        * to zero. This pulls the entire Sbox into the D-cache before any
+        * data dependent lookups are done.
+        */
+       st0[0] ^= __aesti_inv_sbox[ 0] ^ __aesti_inv_sbox[ 64] ^ __aesti_inv_sbox[129] ^ __aesti_inv_sbox[200];
+       st0[1] ^= __aesti_inv_sbox[16] ^ __aesti_inv_sbox[ 83] ^ __aesti_inv_sbox[150] ^ __aesti_inv_sbox[212];
+       st0[2] ^= __aesti_inv_sbox[32] ^ __aesti_inv_sbox[ 96] ^ __aesti_inv_sbox[160] ^ __aesti_inv_sbox[236];
+       st0[3] ^= __aesti_inv_sbox[48] ^ __aesti_inv_sbox[112] ^ __aesti_inv_sbox[187] ^ __aesti_inv_sbox[247];
 
        for (round = 0;; round += 2, rkp += 8) {
                st1[0] = inv_mix_columns(inv_subshift(st0, 0)) ^ rkp[0];