crypto: chacha-generic - fix use as arm64 no-NEON fallback

The arm64 implementations of ChaCha and XChaCha are failing the extra
crypto self-tests following my patches to test the !may_use_simd() code
paths, which previously were untested.  The problem is as follows:

When !may_use_simd(), the arm64 NEON implementations fall back to the
generic implementation, which uses the skcipher_walk API to iterate
through the src/dst scatterlists.  Due to how the skcipher_walk API
works, walk.stride is set from the skcipher_alg actually being used,
which in this case is the arm64 NEON algorithm.  Thus walk.stride is
5*CHACHA_BLOCK_SIZE, not CHACHA_BLOCK_SIZE.

This unnecessarily large stride shouldn't cause an actual problem.
However, the generic implementation computes round_down(nbytes,
walk.stride).  round_down() assumes the round amount is a power of 2,
which 5*CHACHA_BLOCK_SIZE is not, so it gives the wrong result.

This causes the following case in skcipher_walk_done() to be hit,
causing a WARN() and failing the encryption operation:

	if (WARN_ON(err)) {
		/* unexpected case; didn't process all bytes */
		err = -EINVAL;
		goto finish;
	}

Fix it by rounding down to CHACHA_BLOCK_SIZE instead of walk.stride.

(Or we could replace round_down() with rounddown(), but that would add a
slow division operation every time, which I think we should avoid.)

Fixes: 2fe55987 ("crypto: arm64/chacha - use combined SIMD/ALU routine for more speed")
Cc: <stable@vger.kernel.org> # v5.0+
Signed-off-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

crypto/chacha_generic.c

@@ -52,7 +52,7 @@ static int chacha_stream_xor(struct skcipher_request *req,
 		unsigned int nbytes = walk.nbytes;
 
 		if (nbytes < walk.total)
-			nbytes = round_down(nbytes, walk.stride);
+			nbytes = round_down(nbytes, CHACHA_BLOCK_SIZE);
 
 		chacha_docrypt(state, walk.dst.virt.addr, walk.src.virt.addr,
 			       nbytes, ctx->nrounds);