We currently only support passing the AlgorithmIdentifier (X509_ALGOR
)
parameter field to symmetric cipher provider implementations.
We do support passing them to legacy implementations of other types of operation algorithms as well, but it's done in a way that can't be supported with providers, because it involves sharing specific structures between libcrypto and the backend implementation.
For a longer background and explanation, see Background / tl;dr at the end of this design.
We already have a parameter key, but it's currently only specified for
EVP_CIPHER
, in support of EVP_CIPHER_param_to_asn1()
and
EVP_CIPHER_asn1_to_param()
.
"alg_id_param", also known as the macro OSSL_CIPHER_PARAM_ALGORITHM_ID_PARAMS
This parameter can be used in the exact same manner with other operations, with the value of the AlgorithmIdentifier parameter as an octet string, to be interpreted by the implementations in whatever way they see fit.
Applications can choose to add these in an OSSL_PARAM
array, to be passed
with the multitude of initialization functions that take such an array, or
using specific operation OSSL_PARAM
setters and getters (such as
EVP_PKEY_CTX_set_params
), or using other available convenience functions
(see below).
This parameter will have to be documented in the following files:
doc/man7/provider-asym_cipher.pod
doc/man7/provider-cipher.pod
doc/man7/provider-digest.pod
doc/man7/provider-kdf.pod
doc/man7/provider-kem.pod
doc/man7/provider-keyexch.pod
doc/man7/provider-mac.pod
doc/man7/provider-signature.pod
That should cover all algorithms that are, or should be possible to fetch by AlgorithmIdentifier.algorithm, and for which there's potentially a relevant AlgorithmIdentifier.parameters field.
We may arguably want to consider doc/man7/provider-keymgmt.pod
too, but
an AlgorithmIdentifier that's attached directly to a key is usually part of
a PrivKeyInfo or SubjectPublicKeyInfo structure, and those are handled by
encoders and decoders as those see fit, and there's no tangible reason why
that would have to change.
For convenience, the following set of functions would be added to pass the AlgorithmIdentifier parameter data to diverse operations, or to retrieve such parameter data from them.
/*
* These two would essentially be aliases for EVP_CIPHER_param_to_asn1()
* and EVP_CIPHER_asn1_to_param().
*/
EVP_CIPHER_CTX_set_algor_param(EVP_PKEY_CTX *ctx, X509_ALGOR *alg);
EVP_CIPHER_CTX_get_algor_param(EVP_PKEY_CTX *ctx, X509_ALGOR *alg);
EVP_MD_CTX_set_algor_param(EVP_PKEY_CTX *ctx, X509_ALGOR *alg);
EVP_MD_CTX_get_algor_param(EVP_PKEY_CTX *ctx, X509_ALGOR *alg);
EVP_MAC_CTX_set_algor_param(EVP_PKEY_CTX *ctx, X509_ALGOR *alg);
EVP_MAC_CTX_get_algor_param(EVP_PKEY_CTX *ctx, X509_ALGOR *alg);
EVP_KDF_CTX_set_algor_param(EVP_PKEY_CTX *ctx, X509_ALGOR *alg);
EVP_KDF_CTX_get_algor_param(EVP_PKEY_CTX *ctx, X509_ALGOR *alg);
EVP_PKEY_CTX_set_algor_param(EVP_PKEY_CTX *ctx, X509_ALGOR *alg);
EVP_PKEY_CTX_get_algor_param(EVP_PKEY_CTX *ctx, X509_ALGOR *alg);
Note that all might not need to be added immediately, depending on if they are considered useful or not. For future proofing, however, they should probably all be added.
Providers that implement ciphers or any operation that uses asymmetric keys will have to implement support for passing AlgorithmIdentifier parameter data, and will have to process that data in whatever manner that's necessary to meet the standards for that operation.
There are no possible fallback strategies, which is fine, considering that
current provider functionality doesn't support passing AlgorithmIdentifier
parameter data at all (except for EVP_CIPHER
), and therefore do not work
at all when such parameter data needs to be passed.
OpenSSL has historically done a few tricks to not have to pass AlgorithmIdenfier parameter data to the backend implementations of cryptographic operations:
RSA
structure can also carry RSA-PSS parameters).EVP_CIPHER
case, there is functionality to pass the parameter
data specifically.EVP_PKEY
ctrls.With providers, some of that support was retained, but not others. Most
crucially, the EVP_PKEY
ctrls for PKCS#7 and CMS were not retained,
because the way they were implemented violated the principle that provider
implementations MUST NOT share complex OpenSSL specific structures with
libcrypto.
Quite a lot of the available examples today revolve around CMS, with a number of RFCs that specify what parameters should be passed with certain operations / algorithms. This list is not exhaustive, the reader is encouraged to research further usages.
XOR-MD5 is experimental, but it does demonstrate the possibility of a parametrized hash algorithm.
Some of it can be claimed to already have support in OpenSSL. However, this is with old libcrypto code that has special knowledge of the algorithms that are involved.