lib/crypto: optimize aes_cmac_128

- We avoid variables in order to do a lazy cleanup
  in aes_cmac_128_final() via ZERO_STRUCTP(ctx)
- We avoid unused memcpy() calls
- We use the optimized aes_block_{xor,lshift}() functions
- Align AES_BLOCK_SIZE arrays to 8 bytes

BUG: https://bugzilla.samba.org/show_bug.cgi?id=11451

Signed-off-by: Stefan Metzmacher <metze@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
(cherry picked from commit 8795ad2030ade771145264be7c64d960dcfdbed7)

2 files changed, 25 insertions, 85 deletions

diff --git a/lib/crypto/aes_cmac_128.c b/lib/crypto/aes_cmac_128.c
index b630eeacd4e..5d71e82de0d 100644
--- a/lib/crypto/aes_cmac_128.c
+++ b/lib/crypto/aes_cmac_128.c
@@ -33,92 +33,42 @@ static const uint8_t const_Rb[] = {
 
 #define _MSB(x) (((x)[0] & 0x80)?1:0)
 
-static inline void aes_cmac_128_left_shift_1(const uint8_t in[AES_BLOCK_SIZE],
-					     uint8_t out[AES_BLOCK_SIZE])
-{
-	uint8_t overflow = 0;
-	int8_t i;
-
-	for (i = AES_BLOCK_SIZE - 1; i >= 0; i--) {
-		out[i] = in[i] << 1;
-
-		out[i] |= overflow;
-
-		overflow = _MSB(&in[i]);
-	}
-}
-
-static inline void aes_cmac_128_xor(const uint8_t in1[AES_BLOCK_SIZE],
-				    const uint8_t in2[AES_BLOCK_SIZE],
-				    uint8_t out[AES_BLOCK_SIZE])
-{
-	uint8_t i;
-
-	for (i = 0; i < AES_BLOCK_SIZE; i++) {
-		out[i] = in1[i] ^ in2[i];
-	}
-}
-
 void aes_cmac_128_init(struct aes_cmac_128_context *ctx,
 		       const uint8_t K[AES_BLOCK_SIZE])
 {
-	uint8_t L[AES_BLOCK_SIZE];
-
 	ZERO_STRUCTP(ctx);
 
 	AES_set_encrypt_key(K, 128, &ctx->aes_key);
 
 	/* step 1 - generate subkeys k1 and k2 */
 
-	AES_encrypt(const_Zero, L, &ctx->aes_key);
+	AES_encrypt(const_Zero, ctx->L, &ctx->aes_key);
 
-	if (_MSB(L) == 0) {
-		aes_cmac_128_left_shift_1(L, ctx->K1);
+	if (_MSB(ctx->L) == 0) {
+		aes_block_lshift(ctx->L, ctx->K1);
 	} else {
-		uint8_t tmp_block[AES_BLOCK_SIZE];
-
-		aes_cmac_128_left_shift_1(L, tmp_block);
-		aes_cmac_128_xor(tmp_block, const_Rb, ctx->K1);
-		ZERO_STRUCT(tmp_block);
+		aes_block_lshift(ctx->L, ctx->tmp);
+		aes_block_xor(ctx->tmp, const_Rb, ctx->K1);
 	}
 
 	if (_MSB(ctx->K1) == 0) {
-		aes_cmac_128_left_shift_1(ctx->K1, ctx->K2);
+		aes_block_lshift(ctx->K1, ctx->K2);
 	} else {
-		uint8_t tmp_block[AES_BLOCK_SIZE];
-
-		aes_cmac_128_left_shift_1(ctx->K1, tmp_block);
-		aes_cmac_128_xor(tmp_block, const_Rb, ctx->K2);
-		ZERO_STRUCT(tmp_block);
+		aes_block_lshift(ctx->K1, ctx->tmp);
+		aes_block_xor(ctx->tmp, const_Rb, ctx->K2);
 	}
-
-	ZERO_STRUCT(L);
 }
 
 void aes_cmac_128_update(struct aes_cmac_128_context *ctx,
-			 const uint8_t *_msg, size_t _msg_len)
+			 const uint8_t *msg, size_t msg_len)
 {
-	uint8_t tmp_block[AES_BLOCK_SIZE];
-	uint8_t Y[AES_BLOCK_SIZE];
-	const uint8_t *msg = _msg;
-	size_t msg_len = _msg_len;
-
-	/*
-	 * copy the remembered last block
-	 */
-	ZERO_STRUCT(tmp_block);
-	if (ctx->last_len) {
-		memcpy(tmp_block, ctx->last, ctx->last_len);
-	}
-
 	/*
 	 * check if we expand the block
 	 */
 	if (ctx->last_len < AES_BLOCK_SIZE) {
 		size_t len = MIN(AES_BLOCK_SIZE - ctx->last_len, msg_len);
 
-		memcpy(&tmp_block[ctx->last_len], msg, len);
-		memcpy(ctx->last, tmp_block, AES_BLOCK_SIZE);
+		memcpy(&ctx->last[ctx->last_len], msg, len);
 		msg += len;
 		msg_len -= len;
 		ctx->last_len += len;
@@ -126,59 +76,43 @@ void aes_cmac_128_update(struct aes_cmac_128_context *ctx,
 
 	if (msg_len == 0) {
 		/* if it is still the last block, we are done */
-		ZERO_STRUCT(tmp_block);
 		return;
 	}
 
 	/*
-	 * It is not the last block anymore
-	 */
-	ZERO_STRUCT(ctx->last);
-	ctx->last_len = 0;
-
-	/*
 	 * now checksum everything but the last block
 	 */
-	aes_cmac_128_xor(ctx->X, tmp_block, Y);
-	AES_encrypt(Y, ctx->X, &ctx->aes_key);
+	aes_block_xor(ctx->X, ctx->last, ctx->Y);
+	AES_encrypt(ctx->Y, ctx->X, &ctx->aes_key);
 
 	while (msg_len > AES_BLOCK_SIZE) {
-		memcpy(tmp_block, msg, AES_BLOCK_SIZE);
+		aes_block_xor(ctx->X, msg, ctx->Y);
+		AES_encrypt(ctx->Y, ctx->X, &ctx->aes_key);
 		msg += AES_BLOCK_SIZE;
 		msg_len -= AES_BLOCK_SIZE;
-
-		aes_cmac_128_xor(ctx->X, tmp_block, Y);
-		AES_encrypt(Y, ctx->X, &ctx->aes_key);
 	}
 
 	/*
 	 * copy the last block, it will be processed in
 	 * aes_cmac_128_final().
 	 */
+	ZERO_STRUCT(ctx->last);
 	memcpy(ctx->last, msg, msg_len);
 	ctx->last_len = msg_len;
-
-	ZERO_STRUCT(tmp_block);
-	ZERO_STRUCT(Y);
 }
 
 void aes_cmac_128_final(struct aes_cmac_128_context *ctx,
 			uint8_t T[AES_BLOCK_SIZE])
 {
-	uint8_t tmp_block[AES_BLOCK_SIZE];
-	uint8_t Y[AES_BLOCK_SIZE];
-
 	if (ctx->last_len < AES_BLOCK_SIZE) {
 		ctx->last[ctx->last_len] = 0x80;
-		aes_cmac_128_xor(ctx->last, ctx->K2, tmp_block);
+		aes_block_xor(ctx->last, ctx->K2, ctx->tmp);
 	} else {
-		aes_cmac_128_xor(ctx->last, ctx->K1, tmp_block);
+		aes_block_xor(ctx->last, ctx->K1, ctx->tmp);
 	}
 
-	aes_cmac_128_xor(tmp_block, ctx->X, Y);
-	AES_encrypt(Y, T, &ctx->aes_key);
+	aes_block_xor(ctx->tmp, ctx->X, ctx->Y);
+	AES_encrypt(ctx->Y, T, &ctx->aes_key);
 
-	ZERO_STRUCT(tmp_block);
-	ZERO_STRUCT(Y);
 	ZERO_STRUCTP(ctx);
 }
diff --git a/lib/crypto/aes_cmac_128.h b/lib/crypto/aes_cmac_128.h
index 28117a06b3e..9bcf951171c 100644
--- a/lib/crypto/aes_cmac_128.h
+++ b/lib/crypto/aes_cmac_128.h
@@ -22,10 +22,16 @@
 struct aes_cmac_128_context {
 	AES_KEY aes_key;
 
+	uint64_t __align;
+
 	uint8_t K1[AES_BLOCK_SIZE];
 	uint8_t K2[AES_BLOCK_SIZE];
 
+	uint8_t L[AES_BLOCK_SIZE];
 	uint8_t X[AES_BLOCK_SIZE];
+	uint8_t Y[AES_BLOCK_SIZE];
+
+	uint8_t tmp[AES_BLOCK_SIZE];
 
 	uint8_t last[AES_BLOCK_SIZE];
 	size_t last_len;