[design] Make it possible to pass AlgorithmIdentifier parameter data

This design is to allow passing AlgorithmIdentifier parameter data to and
from any cryptograpfic operation, with convenience functions for them all,
not just for symmetric ciphers.

This is crucial to support CMS, among others.

Ref: openssl/project#172

Reviewed-by: Tomas Mraz <tomas@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Paul Dale <pauli@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/22162)
This commit is contained in:
Richard Levitte 2023-09-21 14:39:30 +02:00 committed by Pauli
parent fb20e66c6b
commit 11f69aa507

View File

@ -0,0 +1,156 @@
Passing AlgorithmIdentifier parameters to operations
====================================================
Quick background
----------------
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](#background-tldr) at the end of this design.
Establish an OSSL_PARAM key that any algorithms may become aware of
-------------------------------------------------------------------
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.
Public convenience API
----------------------
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.
``` C
/*
* 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.
Requirements on the providers
-----------------------------
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.
Fallback strategies
-------------------
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.
-----
-----
Background / tl;dr
------------------
### AlgorithmIdenfier parameter and how it's used
OpenSSL has historically done a few tricks to not have to pass
AlgorithmIdenfier parameter data to the backend implementations of
cryptographic operations:
- In some cases, they were passed as part of the lower level key structure
(for example, the `RSA` structure can also carry RSA-PSS parameters).
- In the `EVP_CIPHER` case, there is functionality to pass the parameter
data specifically.
- For asymmetric key operations, PKCS#7 and CMS support was added as
`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.
### Usage examples
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.
- [DSA](https://www.rfc-editor.org/rfc/rfc3370#section-3.1) signatures
typically have the domain parameters *p*, *q* and *g*.
- [RC2 key wrap](https://www.rfc-editor.org/rfc/rfc3370#section-4.3.2)
- [PBKDF2](https://www.rfc-editor.org/rfc/rfc3370#section-4.4.1)
- [3DES-CBC](https://www.rfc-editor.org/rfc/rfc3370#section-5.1)
- [RC2-CBC](https://www.rfc-editor.org/rfc/rfc3370#section-5.2)
- [GOST 28147-89](https://www.rfc-editor.org/rfc/rfc4490.html#section-5.1)
- [RSA-OAEP](https://www.rfc-editor.org/rfc/rfc8017#appendix-A.2.1)
- [RSA-PSS](https://www.rfc-editor.org/rfc/rfc8017#appendix-A.2.3)
- [XOR-MD5](https://www.rfc-editor.org/rfc/rfc6210.html) 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.