Re: [PATCH v3] crypto: aes - Fix undesired override of some optimized AES modes

[Stian Halseth <stian@itx.no>]

[-- Attachment #1: Type: text/plain, Size: 6215 bytes --]


On Tue, 2026-09-29 at 15:27 -0700, Eric Biggers wrote:
> The new library APIs for AES encryption modes were wired up to the
> traditional crypto API via crypto/aes.c.  However, for now the kernel
> is
> still in a transitional state where various architectures still have
> architecture-optimized implementations of AES modes in
> arch/*/crypto/,
> wired up to the traditional crypto API only.  Because of that, the
> crypto/aes.c algorithms were given a cra_priority of only 110 to
> prevent
> them from overriding arch/*/crypto/ in the traditional crypto API.
> 
> However, because of how the traditional crypto API works, the
> cra_priority trick doesn't work in cases where the relevant algorithm
> isn't directly implemented by arch/*/crypto/ but rather is provided
> by a
> template instance using other code in arch/*/crypto/.
> 
> For example, x86 doesn't have its own "ccm(aes)" but rather relies on
> the "ccm" template constructing it from the x86-optimized "ctr(aes)".
> The existence of the library-based "ccm(aes)" prevents that, even
> though
> its priority is lower than what the template would produce.
> 
> Thus, "ccm(aes)" ends up using the slower single-block AES code.
> 
> Therefore, skip wiring up the relevant library-based code to the
> traditional crypto API on architectures where this problem can occur,
> as
> determined by what exists in arch/*/crypto/ for each architecture.
> 
> This is ugly, but it's also temporary: these conditions will go away
> as
> architecture-optimized implementations of AES modes are migrated into
> the library.  But until then, we need to prevent performance
> regressions
> by ensuring that the optimized code continues to be used.
> 
> Fixes: 20df21a482aa ("crypto: aes - Add CBC and CBC-CTS support using
> library")
> Fixes: 8ca62072faa1 ("crypto: aes - Add GCM support using library")
> Fixes: f70ad727d1d6 ("crypto: aes - Add CCM support using library")
> Fixes: 94efa0c9fb36 ("crypto: aes - Add XTS support using library")
> Closes: https://github.com/sparclinux/issues/issues/106
> Signed-off-by: Eric Biggers <ebiggers@kernel.org>
> ---
> 
> This patch is intended to taken through libcrypto-fixes
> 
> v3: Also suppress xts(aes) on SPARC, and improved comments
> v2: Fixed PowerPC config option, and resent as standalone patch
> 
> 
Tested on a SPARC T7-1 (M7) on top of v7.3-rc5. xts(aes) resolves to
xts(ecb-aes-sparc64) again. cryptsetup benchmark aes-xts goes from
152/147 MiB/s back to 429/404 (AES-128/AES-256), the 7.2 figures.
CRYPTO_SELFTESTS_FULL passes and XTS matches OpenSSL.

Tested-by: Stian Halseth stian@itx.no
>  crypto/aes.c | 51 +++++++++++++++++++++++++++++++++++++++++++++++---
> -
>  1 file changed, 47 insertions(+), 4 deletions(-)
> 
> diff --git a/crypto/aes.c b/crypto/aes.c
> index 94791f481e98..51eee78396ea 100644
> --- a/crypto/aes.c
> +++ b/crypto/aes.c
> @@ -637,7 +637,18 @@ static struct skcipher_alg skcipher_algs[] = {
>  		.decrypt = crypto_aes_cbc_decrypt,
>  	},
>  #endif
> -#if IS_ENABLED(CONFIG_CRYPTO_CTS)
> +	/*
> +	 * Don't register library-based "cts(cbc(aes))" on
> architectures where
> +	 * it might block a "better" implementation from being
> instantiated via
> +	 * the "cts" template.  These exclusions are temporary and
> will go away
> +	 * as the arch-optimized AES code is migrated into the
> library.
> +	 */
> +#if IS_ENABLED(CONFIG_CRYPTO_CTS) && \
> +	!(IS_ENABLED(CONFIG_ARM) || \
> +	  IS_ENABLED(CONFIG_ARM64) || \
> +	  IS_ENABLED(CONFIG_PPC) || \
> +	  IS_ENABLED(CONFIG_S390) || \
> +	  IS_ENABLED(CONFIG_SPARC))
>  	{
>  		.base.cra_name = "cts(cbc(aes))",
>  		.base.cra_driver_name = "cts-cbc-aes-lib",
> @@ -687,7 +698,13 @@ static struct skcipher_alg skcipher_algs[] = {
>  		.decrypt = crypto_aes_xctr_crypt,
>  	},
>  #endif
> -#if IS_ENABLED(CONFIG_CRYPTO_XTS)
> +	/*
> +	 * Don't register library-based "xts(aes)" on architectures
> where it
> +	 * might block a "better" implementation from being
> instantiated via the
> +	 * "xts" template.  This exclusion is temporary and will go
> away when
> +	 * the library AES-XTS is optimized for SPARC.
> +	 */
> +#if IS_ENABLED(CONFIG_CRYPTO_XTS) && !IS_ENABLED(CONFIG_SPARC)
>  	{
>  		.base.cra_name = "xts(aes)",
>  		.base.cra_driver_name = "xts-aes-lib",
> @@ -980,7 +997,20 @@ static __maybe_unused int
> crypto_aes_ccm_decrypt(struct aead_request *req)
>  }
>  
>  static struct aead_alg aead_algs[] = {
> -#if IS_ENABLED(CONFIG_CRYPTO_GCM)
> +	/*
> +	 * Don't register library-based "gcm(aes)" and
> "rfc4106(gcm(aes))" on
> +	 * architectures where they might block a "better"
> implementation from
> +	 * being instantiated via the "gcm" and "rfc4106"
> templates.  These
> +	 * exclusions are temporary and will go away as the arch-
> optimized AES
> +	 * code is migrated into the library.
> +	 */
> +#if IS_ENABLED(CONFIG_CRYPTO_GCM) && \
> +	!(IS_ENABLED(CONFIG_ARM) || \
> +	  IS_ENABLED(CONFIG_ARM64) || \
> +	  IS_ENABLED(CONFIG_PPC) || \
> +	  IS_ENABLED(CONFIG_RISCV) || \
> +	  IS_ENABLED(CONFIG_S390) || \
> +	  IS_ENABLED(CONFIG_SPARC))
>  	{
>  		.base.cra_name = "gcm(aes)",
>  		.base.cra_driver_name = "gcm-aes-lib",
> @@ -1012,7 +1042,20 @@ static struct aead_alg aead_algs[] = {
>  		.chunksize = AES_BLOCK_SIZE,
>  	},
>  #endif /* CONFIG_CRYPTO_GCM */
> -#if IS_ENABLED(CONFIG_CRYPTO_CCM)
> +	/*
> +	 * Don't register library-based "ccm(aes)" on architectures
> where it
> +	 * might block a "better" implementation from being
> instantiated via the
> +	 * "ccm" template.  These exclusions are temporary and will
> go away as
> +	 * the arch-optimized AES code is migrated into the library.
> +	 */
> +#if IS_ENABLED(CONFIG_CRYPTO_CCM) && \
> +	!(IS_ENABLED(CONFIG_ARM) || \
> +	  IS_ENABLED(CONFIG_ARM64) || \
> +	  IS_ENABLED(CONFIG_PPC) || \
> +	  IS_ENABLED(CONFIG_RISCV) || \
> +	  IS_ENABLED(CONFIG_S390) || \
> +	  IS_ENABLED(CONFIG_SPARC) || \
> +	  IS_ENABLED(CONFIG_X86))
>  	{
>  		.base.cra_name = "ccm(aes)",
>  		.base.cra_driver_name = "ccm-aes-lib",
> 
> base-commit: 93f51579e7df248780214094418f205253383cc5

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 228 bytes --]