Re: [PATCH 04/12] RISC-V: crypto: add Zvkned accelerated AES implementation

[Eric Biggers]

On Thu, Oct 26, 2023 at 02:36:36AM +0800, Jerry Shih wrote:
> diff --git a/arch/riscv/crypto/Kconfig b/arch/riscv/crypto/Kconfig
> index 10d60edc0110..500938317e71 100644
> --- a/arch/riscv/crypto/Kconfig
> +++ b/arch/riscv/crypto/Kconfig
> @@ -2,4 +2,16 @@
>  
>  menu "Accelerated Cryptographic Algorithms for CPU (riscv)"
>  
> +config CRYPTO_AES_RISCV64
> +	default y if RISCV_ISA_V
> +	tristate "Ciphers: AES"
> +	depends on 64BIT && RISCV_ISA_V
> +	select CRYPTO_AES
> +	select CRYPTO_ALGAPI
> +	help
> +	  Block ciphers: AES cipher algorithms (FIPS-197)
> +
> +	  Architecture: riscv64 using:
> +	  - Zvkned vector crypto extension

kconfig options should default to off.

I.e., remove the line "default y if RISCV_ISA_V"

> + *
> + * All zvkned-based functions use encryption expending keys for both encryption
> + * and decryption.
> + */

The above comment is a bit confusing.  It's describing the 'key' field of struct
aes_key; maybe there should be a comment there instead:

    struct aes_key {
            u32 key[AES_MAX_KEYLENGTH_U32]; /* round keys in encryption order */
            u32 rounds;
    };

> +int riscv64_aes_setkey(struct riscv64_aes_ctx *ctx, const u8 *key,
> +		       unsigned int keylen)
> +{
> +	/*
> +	 * The RISC-V AES vector crypto key expending doesn't support AES-192.
> +	 * We just use the generic software key expending here to simplify the key
> +	 * expending flow.
> +	 */

expending => expanding

> +	u32 aes_rounds;
> +	u32 key_length;
> +	int ret;
> +
> +	ret = aes_expandkey(&ctx->fallback_ctx, key, keylen);
> +	if (ret < 0)
> +		return -EINVAL;
> +
> +	/*
> +	 * Copy the key from `crypto_aes_ctx` to `aes_key` for zvkned-based AES
> +	 * implementations.
> +	 */
> +	aes_rounds = aes_round_num(keylen);
> +	ctx->key.rounds = aes_rounds;
> +	key_length = AES_BLOCK_SIZE * (aes_rounds + 1);
> +	memcpy(ctx->key.key, ctx->fallback_ctx.key_enc, key_length);
> +
> +	return 0;
> +}

Ideally this would use the same crypto_aes_ctx for both the fallback and the
assembly code.  I suppose we don't want to diverge from the OpenSSL code (unless
it gets rewritten), though.  So I guess this is fine for now.

> void riscv64_aes_encrypt_zvkned(const struct riscv64_aes_ctx *ctx, u8 *dst,
>                                const u8 *src)

These functions can be called from a different module (aes-block-riscv64), so
they need EXPORT_SYMBOL_GPL.

> +static inline bool check_aes_ext(void)
> +{
> +	return riscv_isa_extension_available(NULL, ZVKNED) &&
> +	       riscv_vector_vlen() >= 128;
> +}
> +
> +static int __init riscv64_aes_mod_init(void)
> +{
> +	if (check_aes_ext())
> +		return crypto_register_alg(&riscv64_aes_alg_zvkned);
> +
> +	return -ENODEV;
> +}
> +
> +static void __exit riscv64_aes_mod_fini(void)
> +{
> +	if (check_aes_ext())
> +		crypto_unregister_alg(&riscv64_aes_alg_zvkned);
> +}
> +
> +module_init(riscv64_aes_mod_init);
> +module_exit(riscv64_aes_mod_fini);

module_exit can only run if module_init succeeded.  So, in cases like this it's
not necessary to check for CPU features before unregistering the algorithm.

- Eric