crypto: aesni - disable "by8" AVX CTR optimization

The "by8" implementation introduced in commit 22cddcc ("crypto: aes
- AES CTR x86_64 "by8" AVX optimization") is failing crypto tests as it
handles counter block overflows differently. It only accounts the right
most 32 bit as a counter -- not the whole block as all other
implementations do. This makes it fail the cryptomgr test #4 that
specifically tests this corner case.

As we're quite late in the release cycle, just disable the "by8" variant
for now.

Reported-by: Romain Francoise <romain@orebokech.com>
Signed-off-by: Mathias Krause <minipli@googlemail.com>
Cc: Chandramouli Narayanan <mouli@linux.intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

arch/x86/crypto/aesni-intel_glue.c

@@ -481,7 +481,7 @@ static void ctr_crypt_final(struct crypto_aes_ctx *ctx,
 	crypto_inc(ctrblk, AES_BLOCK_SIZE);
 }
 
-#ifdef CONFIG_AS_AVX
+#if 0	/* temporary disabled due to failing crypto tests */
 static void aesni_ctr_enc_avx_tfm(struct crypto_aes_ctx *ctx, u8 *out,
 			      const u8 *in, unsigned int len, u8 *iv)
 {
@@ -1522,7 +1522,7 @@ static int __init aesni_init(void)
 		aesni_gcm_dec_tfm = aesni_gcm_dec;
 	}
 	aesni_ctr_enc_tfm = aesni_ctr_enc;
-#ifdef CONFIG_AS_AVX
+#if 0	/* temporary disabled due to failing crypto tests */
 	if (cpu_has_avx) {
 		/* optimize performance of ctr mode encryption transform */
 		aesni_ctr_enc_tfm = aesni_ctr_enc_avx_tfm;