crypto: cast5/avx - avoid using temporary stack buffers

Introduce new assembler functions to avoid use temporary stack buffers in glue
code. This also allows use of vector instructions for xoring output in CTR and
CBC modes and construction of IVs for CTR mode.

ECB mode sees ~0.5% decrease in speed because added one extra function
call. CBC mode decryption and CTR mode benefit from vector operations
and gain ~5%.

Signed-off-by: Jussi Kivilinna <jussi.kivilinna@mbnet.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

1 files changed, 23 insertions, 56 deletions

diff --git a/arch/x86/crypto/cast5_avx_glue.c b/arch/x86/crypto/cast5_avx_glue.c
index e0ea14f9547f..c6631813dc11 100644
--- a/arch/x86/crypto/cast5_avx_glue.c
+++ b/arch/x86/crypto/cast5_avx_glue.c
@@ -37,29 +37,14 @@
 
 #define CAST5_PARALLEL_BLOCKS 16
 
-asmlinkage void __cast5_enc_blk_16way(struct cast5_ctx *ctx, u8 *dst,
-				      const u8 *src, bool xor);
-asmlinkage void cast5_dec_blk_16way(struct cast5_ctx *ctx, u8 *dst,
+asmlinkage void cast5_ecb_enc_16way(struct cast5_ctx *ctx, u8 *dst,
 				    const u8 *src);
-
-static inline void cast5_enc_blk_xway(struct cast5_ctx *ctx, u8 *dst,
-				      const u8 *src)
-{
-	__cast5_enc_blk_16way(ctx, dst, src, false);
-}
-
-static inline void cast5_enc_blk_xway_xor(struct cast5_ctx *ctx, u8 *dst,
-					  const u8 *src)
-{
-	__cast5_enc_blk_16way(ctx, dst, src, true);
-}
-
-static inline void cast5_dec_blk_xway(struct cast5_ctx *ctx, u8 *dst,
-				      const u8 *src)
-{
-	cast5_dec_blk_16way(ctx, dst, src);
-}
-
+asmlinkage void cast5_ecb_dec_16way(struct cast5_ctx *ctx, u8 *dst,
+				    const u8 *src);
+asmlinkage void cast5_cbc_dec_16way(struct cast5_ctx *ctx, u8 *dst,
+				    const u8 *src);
+asmlinkage void cast5_ctr_16way(struct cast5_ctx *ctx, u8 *dst, const u8 *src,
+				__be64 *iv);
 
 static inline bool cast5_fpu_begin(bool fpu_enabled, unsigned int nbytes)
 {
@@ -79,8 +64,11 @@ static int ecb_crypt(struct blkcipher_desc *desc, struct blkcipher_walk *walk,
 	struct cast5_ctx *ctx = crypto_blkcipher_ctx(desc->tfm);
 	const unsigned int bsize = CAST5_BLOCK_SIZE;
 	unsigned int nbytes;
+	void (*fn)(struct cast5_ctx *ctx, u8 *dst, const u8 *src);
 	int err;
 
+	fn = (enc) ? cast5_ecb_enc_16way : cast5_ecb_dec_16way;
+
 	err = blkcipher_walk_virt(desc, walk);
 	desc->flags &= ~CRYPTO_TFM_REQ_MAY_SLEEP;
 
@@ -93,10 +81,7 @@ static int ecb_crypt(struct blkcipher_desc *desc, struct blkcipher_walk *walk,
 		/* Process multi-block batch */
 		if (nbytes >= bsize * CAST5_PARALLEL_BLOCKS) {
 			do {
-				if (enc)
-					cast5_enc_blk_xway(ctx, wdst, wsrc);
-				else
-					cast5_dec_blk_xway(ctx, wdst, wsrc);
+				fn(ctx, wdst, wsrc);
 
 				wsrc += bsize * CAST5_PARALLEL_BLOCKS;
 				wdst += bsize * CAST5_PARALLEL_BLOCKS;
@@ -107,12 +92,11 @@ static int ecb_crypt(struct blkcipher_desc *desc, struct blkcipher_walk *walk,
 				goto done;
 		}
 
+		fn = (enc) ? __cast5_encrypt : __cast5_decrypt;
+
 		/* Handle leftovers */
 		do {
-			if (enc)
-				__cast5_encrypt(ctx, wdst, wsrc);
-			else
-				__cast5_decrypt(ctx, wdst, wsrc);
+			fn(ctx, wdst, wsrc);
 
 			wsrc += bsize;
 			wdst += bsize;
@@ -194,9 +178,7 @@ static unsigned int __cbc_decrypt(struct blkcipher_desc *desc,
 	unsigned int nbytes = walk->nbytes;
 	u64 *src = (u64 *)walk->src.virt.addr;
 	u64 *dst = (u64 *)walk->dst.virt.addr;
-	u64 ivs[CAST5_PARALLEL_BLOCKS - 1];
 	u64 last_iv;
-	int i;
 
 	/* Start of the last block. */
 	src += nbytes / bsize - 1;
@@ -211,13 +193,7 @@ static unsigned int __cbc_decrypt(struct blkcipher_desc *desc,
 			src -= CAST5_PARALLEL_BLOCKS - 1;
 			dst -= CAST5_PARALLEL_BLOCKS - 1;
 
-			for (i = 0; i < CAST5_PARALLEL_BLOCKS - 1; i++)
-				ivs[i] = src[i];
-
-			cast5_dec_blk_xway(ctx, (u8 *)dst, (u8 *)src);
-
-			for (i = 0; i < CAST5_PARALLEL_BLOCKS - 1; i++)
-				*(dst + (i + 1)) ^= *(ivs + i);
+			cast5_cbc_dec_16way(ctx, (u8 *)dst, (u8 *)src);
 
 			nbytes -= bsize;
 			if (nbytes < bsize)
@@ -298,23 +274,12 @@ static unsigned int __ctr_crypt(struct blkcipher_desc *desc,
 	unsigned int nbytes = walk->nbytes;
 	u64 *src = (u64 *)walk->src.virt.addr;
 	u64 *dst = (u64 *)walk->dst.virt.addr;
-	u64 ctrblk = be64_to_cpu(*(__be64 *)walk->iv);
-	__be64 ctrblocks[CAST5_PARALLEL_BLOCKS];
-	int i;
 
 	/* Process multi-block batch */
 	if (nbytes >= bsize * CAST5_PARALLEL_BLOCKS) {
 		do {
-			/* create ctrblks for parallel encrypt */
-			for (i = 0; i < CAST5_PARALLEL_BLOCKS; i++) {
-				if (dst != src)
-					dst[i] = src[i];
-
-				ctrblocks[i] = cpu_to_be64(ctrblk++);
-			}
-
-			cast5_enc_blk_xway_xor(ctx, (u8 *)dst,
-					       (u8 *)ctrblocks);
+			cast5_ctr_16way(ctx, (u8 *)dst, (u8 *)src,
+					(__be64 *)walk->iv);
 
 			src += CAST5_PARALLEL_BLOCKS;
 			dst += CAST5_PARALLEL_BLOCKS;
@@ -327,13 +292,16 @@ static unsigned int __ctr_crypt(struct blkcipher_desc *desc,
 
 	/* Handle leftovers */
 	do {
+		u64 ctrblk;
+
 		if (dst != src)
 			*dst = *src;
 
-		ctrblocks[0] = cpu_to_be64(ctrblk++);
+		ctrblk = *(u64 *)walk->iv;
+		be64_add_cpu((__be64 *)walk->iv, 1);
 
-		__cast5_encrypt(ctx, (u8 *)ctrblocks, (u8 *)ctrblocks);
-		*dst ^= ctrblocks[0];
+		__cast5_encrypt(ctx, (u8 *)&ctrblk, (u8 *)&ctrblk);
+		*dst ^= ctrblk;
 
 		src += 1;
 		dst += 1;
@@ -341,7 +309,6 @@ static unsigned int __ctr_crypt(struct blkcipher_desc *desc,
 	} while (nbytes >= bsize);
 
 done:
-	*(__be64 *)walk->iv = cpu_to_be64(ctrblk);
 	return nbytes;
 }