[PATCH v2] crypto: x86/chacha20 - Manually align stack buffer

[Herbert Xu <herbert@gondor.apana.org.au>]

On Wed, Jan 11, 2017 at 12:14:24PM +0000, Ard Biesheuvel wrote:
> 
> I think the old code was fine, actually:
> 
> u32 *state, state_buf[16 + (CHACHA20_STATE_ALIGN / sizeof(u32)) - 1];
> 
> ends up allocating 16 + 3 *words* == 64 + 12 bytes , which given the
> guaranteed 4 byte alignment is sufficient for ensuring the pointer can
> be 16 byte aligned.

Ah yes you're right, it's a u32.

> So [16 + 2] should be sufficient here

Here's an updated version.

---8<---
The kernel on x86-64 cannot use gcc attribute align to align to
a 16-byte boundary.  This patch reverts to the old way of aligning
it by hand.

Fixes: 9ae433bc79f9 ("crypto: chacha20 - convert generic and...")
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

diff --git a/arch/x86/crypto/chacha20_glue.c b/arch/x86/crypto/chacha20_glue.c
index 78f75b0..1e6af1b 100644
--- a/arch/x86/crypto/chacha20_glue.c
+++ b/arch/x86/crypto/chacha20_glue.c
@@ -67,10 +67,13 @@ static int chacha20_simd(struct skcipher_request *req)
 {
 	struct crypto_skcipher *tfm = crypto_skcipher_reqtfm(req);
 	struct chacha20_ctx *ctx = crypto_skcipher_ctx(tfm);
-	u32 state[16] __aligned(CHACHA20_STATE_ALIGN);
+	u32 *state, state_buf[16 + 2] __aligned(8);
 	struct skcipher_walk walk;
 	int err;
 
+	BUILD_BUG_ON(CHACHA20_STATE_ALIGN != 16);
+	state = PTR_ALIGN(state_buf + 0, CHACHA20_STATE_ALIGN);
+
 	if (req->cryptlen <= CHACHA20_BLOCK_SIZE || !may_use_simd())
 		return crypto_chacha20_crypt(req);
 
-- 
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt