[PATCH 0/2] relax crypto Kconfig dependencies for fsverity/fscrypt

[PATCH 0/2] relax crypto Kconfig dependencies for fsverity/fscrypt

[Ard Biesheuvel]

Relax 'select' dependencies to 'imply' for crypto algorithms that are
fulfilled only at runtime, and which may be implemented by other drivers
than the generic ones implemented in C. This permits, e.g., arm64 builds
to omit the generic CRYPTO_SHA256 and CRYPTO_AES drivers, both of which
are superseded by optimized scalar versions at the very least,

Cc: "Theodore Y. Ts'o" <tytso@mit.edu>
Cc: Jaegeuk Kim <jaegeuk@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>

Ard Biesheuvel (2):
  fscrypt: relax Kconfig dependencies for crypto API algorithms
  fsverity: relax build time dependency on CRYPTO_SHA256

 fs/crypto/Kconfig | 23 ++++++++++++++------
 fs/verity/Kconfig |  8 +++++--
 2 files changed, 22 insertions(+), 9 deletions(-)

-- 
2.30.2

[PATCH 1/2] fscrypt: relax Kconfig dependencies for crypto API algorithms

[Ard Biesheuvel]

Even if FS encryption has strict functional dependencies on various
crypto algorithms and chaining modes. those dependencies could potentially
be satisified by other implementations than the generic ones, and no link
time dependency exists on the 'depends on' claused defined by
CONFIG_FS_ENCRYPTION_ALGS.

So let's relax these clauses to 'imply', so that the default behavior
is still to pull in those generic algorithms, but in a way that permits
them to be disabled again in Kconfig.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
---
 fs/crypto/Kconfig | 23 ++++++++++++++------
 1 file changed, 16 insertions(+), 7 deletions(-)

diff --git a/fs/crypto/Kconfig b/fs/crypto/Kconfig
index a5f5c30368a2..1e6c11de95c8 100644
--- a/fs/crypto/Kconfig
+++ b/fs/crypto/Kconfig
@@ -17,13 +17,22 @@ config FS_ENCRYPTION
 # allows the algorithms to be built as modules when all the filesystems are.
 config FS_ENCRYPTION_ALGS
 	tristate
-	select CRYPTO_AES
-	select CRYPTO_CBC
-	select CRYPTO_CTS
-	select CRYPTO_ECB
-	select CRYPTO_HMAC
-	select CRYPTO_SHA512
-	select CRYPTO_XTS
+	imply CRYPTO_AES
+	imply CRYPTO_CBC
+	imply CRYPTO_CTS
+	imply CRYPTO_ECB
+	imply CRYPTO_HMAC
+	imply CRYPTO_SHA512
+	imply CRYPTO_XTS
+	help
+	  This pulls in the generic implementations of the various
+	  cryptographic algorithms and chaining modes that filesystem
+	  encryption relies on. These are 'soft' dependencies only, as
+	  architectures may supersede these generic implementations with
+	  special, optimized ones.
+
+	  If unsure, keep the generic algorithms enabled, as they can
+	  happily co-exist with per-architecture implementations.
 
 config FS_ENCRYPTION_INLINE_CRYPT
 	bool "Enable fscrypt to use inline crypto"
-- 
2.30.2

Re: [PATCH 1/2] fscrypt: relax Kconfig dependencies for crypto API algorithms

[Eric Biggers]

On Fri, Apr 16, 2021 at 06:06:41PM +0200, Ard Biesheuvel wrote:
> Even if FS encryption has strict functional dependencies on various
> crypto algorithms and chaining modes. those dependencies could potentially
> be satisified by other implementations than the generic ones, and no link
> time dependency exists on the 'depends on' claused defined by
> CONFIG_FS_ENCRYPTION_ALGS.
> 
> So let's relax these clauses to 'imply', so that the default behavior
> is still to pull in those generic algorithms, but in a way that permits
> them to be disabled again in Kconfig.
> 
> Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
> ---
>  fs/crypto/Kconfig | 23 ++++++++++++++------
>  1 file changed, 16 insertions(+), 7 deletions(-)
> 
> diff --git a/fs/crypto/Kconfig b/fs/crypto/Kconfig
> index a5f5c30368a2..1e6c11de95c8 100644
> --- a/fs/crypto/Kconfig
> +++ b/fs/crypto/Kconfig
> @@ -17,13 +17,22 @@ config FS_ENCRYPTION
>  # allows the algorithms to be built as modules when all the filesystems are.
>  config FS_ENCRYPTION_ALGS
>  	tristate
> -	select CRYPTO_AES
> -	select CRYPTO_CBC
> -	select CRYPTO_CTS
> -	select CRYPTO_ECB
> -	select CRYPTO_HMAC
> -	select CRYPTO_SHA512
> -	select CRYPTO_XTS
> +	imply CRYPTO_AES
> +	imply CRYPTO_CBC
> +	imply CRYPTO_CTS
> +	imply CRYPTO_ECB
> +	imply CRYPTO_HMAC
> +	imply CRYPTO_SHA512
> +	imply CRYPTO_XTS
> +	help
> +	  This pulls in the generic implementations of the various
> +	  cryptographic algorithms and chaining modes that filesystem
> +	  encryption relies on. These are 'soft' dependencies only, as
> +	  architectures may supersede these generic implementations with
> +	  special, optimized ones.
> +
> +	  If unsure, keep the generic algorithms enabled, as they can
> +	  happily co-exist with per-architecture implementations.
>  

This seems reasonable to me.  It does have the disadvantage of allowing
misconfigurations where algorithms that are supposed to be available are not
actually made available.  But it's probably better to allow the flexibility to
disable the generic implementations.

I don't really like the help text, though.  First, the description of
FS_ENCRYPTION_ALGS should be either in a comment *or* in a 'help' block, not
split between both.  I think a comment would make more sense since
FS_ENCRYPTION_ALGS isn't a user-selectable symbol, so the 'help' would only be
seen by reading the Kconfig file anyway.

Second, "algorithms that filesystem encryption relies on" is too vague.  We
should clarify that only the "default" algorithms get automatically selected,
and the user may still need to explicitly select more.

Here's a suggested comment which I think would explain things a lot better:

# Filesystems supporting encryption must select this if FS_ENCRYPTION.  This
# allows the algorithms to be built as modules when all the filesystems are,
# whereas selecting them from FS_ENCRYPTION would force them to be built-in.
#
# Note: this option only pulls in the algorithms that filesystem encryption
# needs "by default".  If userspace will use "non-default" encryption modes such
# as Adiantum encryption, then those other modes need to be explicitly enabled
# in the crypto API; see Documentation/filesystems/fscrypt.rst for details.
#
# Also note that this option only pulls in the generic implementations of the
# algorithms, not any per-architecture optimized implementations.  It is
# strongly recommended to enable optimized implementations too.  It is safe to
# disable these generic implementations if corresponding optimized
# implementations will always be available too; for this reason, these are soft
# dependencies ('imply' rather than 'select').  Only disable these generic
# implementations if you're sure they will never be needed, though.

[PATCH 2/2] fsverity: relax build time dependency on CRYPTO_SHA256

[Ard Biesheuvel]

CONFIG_CRYPTO_SHA256 denotes the generic C implementation of the SHA-256
shash algorithm, which is selected as the default crypto shash provider
for fsverity. However, fsverity has no strict link time dependency, and
the same shash could be exposed by an optimized implementation, and arm64
has a number of those (scalar, NEON-based and one based on special crypto
instructions). In such cases, it makes little sense to require that the
generic C implementation is incorporated as well, given that it will never
be called.

To address this, relax the 'select' clause to 'imply' so that the generic
driver can be omitted from the build if desired.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
---
 fs/verity/Kconfig | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/fs/verity/Kconfig b/fs/verity/Kconfig
index 88fb25119899..24d1b54de807 100644
--- a/fs/verity/Kconfig
+++ b/fs/verity/Kconfig
@@ -3,9 +3,13 @@
 config FS_VERITY
 	bool "FS Verity (read-only file-based authenticity protection)"
 	select CRYPTO
-	# SHA-256 is selected as it's intended to be the default hash algorithm.
+	# SHA-256 is implied as it's intended to be the default hash algorithm.
 	# To avoid bloat, other wanted algorithms must be selected explicitly.
-	select CRYPTO_SHA256
+	# Note that CRYPTO_SHA256 denotes the generic C implementation, but
+	# some architectures provided optimized implementations of the same
+	# algorithm that may be used instead. In this case, CRYPTO_SHA256 may
+	# be omitted even if SHA-256 is being used.
+	imply CRYPTO_SHA256
 	help
 	  This option enables fs-verity.  fs-verity is the dm-verity
 	  mechanism implemented at the file level.  On supported
-- 
2.30.2

Re: [PATCH 2/2] fsverity: relax build time dependency on CRYPTO_SHA256

[Eric Biggers]

On Fri, Apr 16, 2021 at 06:06:42PM +0200, Ard Biesheuvel wrote:
> CONFIG_CRYPTO_SHA256 denotes the generic C implementation of the SHA-256
> shash algorithm, which is selected as the default crypto shash provider
> for fsverity. However, fsverity has no strict link time dependency, and
> the same shash could be exposed by an optimized implementation, and arm64
> has a number of those (scalar, NEON-based and one based on special crypto
> instructions). In such cases, it makes little sense to require that the
> generic C implementation is incorporated as well, given that it will never
> be called.
> 
> To address this, relax the 'select' clause to 'imply' so that the generic
> driver can be omitted from the build if desired.
> 
> Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
> ---
>  fs/verity/Kconfig | 8 ++++++--
>  1 file changed, 6 insertions(+), 2 deletions(-)
> 
> diff --git a/fs/verity/Kconfig b/fs/verity/Kconfig
> index 88fb25119899..24d1b54de807 100644
> --- a/fs/verity/Kconfig
> +++ b/fs/verity/Kconfig
> @@ -3,9 +3,13 @@
>  config FS_VERITY
>  	bool "FS Verity (read-only file-based authenticity protection)"
>  	select CRYPTO
> -	# SHA-256 is selected as it's intended to be the default hash algorithm.
> +	# SHA-256 is implied as it's intended to be the default hash algorithm.
>  	# To avoid bloat, other wanted algorithms must be selected explicitly.
> -	select CRYPTO_SHA256
> +	# Note that CRYPTO_SHA256 denotes the generic C implementation, but
> +	# some architectures provided optimized implementations of the same
> +	# algorithm that may be used instead. In this case, CRYPTO_SHA256 may
> +	# be omitted even if SHA-256 is being used.
> +	imply CRYPTO_SHA256
>  	help
>  	  This option enables fs-verity.  fs-verity is the dm-verity
>  	  mechanism implemented at the file level.  On supported

Looks fine,

Acked-by: Eric Biggers <ebiggers@google.com>

- Eric