Evolution of pk.h in 4.0

[mpg]

This issue is meant as a place to discuss what we want to do with PK in 4.0.

There are two main options:

1. Keep it as part of the public API until 5.0, but aim to make it a thin wrapper around PSA Crypto (in 4.0 or 4.x).
2. Remove it from the public API in 4.0, and eventually remove it entirely (in 4.0 or 4.x).

Scope of PK

• Crypto operations similar to two PSA Crypto families: asymmetric signature (RSA, ECDSA) and asymmetric encryption (RSA).
  • Note: that's not all asymmetric crypto: key agreement (ECDH) and PAKE (EC J-PAKE) are not covered.
• Writing and parsing of public and private keys (ECC, RSA) in a variety of formats currently not supported by PSA.
• The pk_context structure serves as a container for public/private keys, which tend to be longer-lived and more often user-provided.

Note: for low-level modules related to asymmetric crypto (rsa.h, ecdsa.h covered by PK, but also ecdh.h and ecjpake.h, and underneath them ecp.h and bignum.h) I think the plan is to remove them from the public API, regardless of what we do with PK.

The problem of X.509 vs the PSA key store

With option 1 and pk_context structures being thin wrappers around psa_key_id_t, or with option 2 and blindly replacing all uses of pk_context with psa_key_id_t and no other code change, we'd have the following problem:

• It is not unusual with X.509 to have a more than a hundred trust roots (currently 147 on my Ubuntu laptop).
• When parsing these (for example with mbedtls_x509_crt_parse_path()) a psa_key_id_t would be created to hold each certificate's public key, resulting in 100+ PSA key slots being used simultaneously.
• But currently the default size of the PSA key store (MBEDTLS_PSA_KEY_SLOT_COUNT) is 32.
• Hence this use case would break in the default config.

There are a number of ways to avoid that problem; we can act at any step in the chain leading to that potential issue:

• The easiest and most obvious would be to up the default value of MBEDTLS_PSA_KEY_SLOT_COUNT to something like 256.
• We could also change X.509 so that it doesn't load the key into a pk_context/psa_key_id_t when parsing a certificate, but only when actually doing an operation with the key. This might be desirable for other reasons and have even been done in the baremetal branch as part of "on-demand parsing" which significantly improved RAM usage. But it requires designing new APIs for users to access the public key embedded in a certificate.
• If we keep PK but make it a thin wrapper around PSA, we could go with a hybrid approach where it stores private keys into PSA key slots, but public keys serialized in a way that psa_import() accepts - and then only load them into a temporary PSA key slot when used. (Note: that's the approach currently used by MBEDTLS_PK_USE_PSA_EC_DATA.)

Rationale for option 1 (keep)

This gives users more time to move to PSA. People who were using low-level modules (including those not covered by PK) need to migrate immediately, but those who were only using the generic PK API can migrate at a convenient time between now and 5.0.

Work needed for option 1 (keep)

In 4.0:

• We may still want to remove some parts:
  • strings:mbedtls_pk_get_name() - see Remove parts of md, cipher and pk #8133 item 1
  • access to low-level context: Remove mbedtls_pk_ec and mbedtls_pk_rsa #7572 - this might require providing replacements (see the discussion there).
  • mbedtls_pk_setup_opaque() as the plain pk_setup() should end up having the same effect.
  • mbedtls_pk_setup_rsa_alt() and MBEDTLS_PK_RSA_ALT_SUPPORT as they seem redundant with the driver API.
• We might want to remove mbedtls_pk_info_t from the API in order to make it slicker but that would require users to change their code, so it runs against the stated goal. (Alternatively, we can keep the structure but later make it trivial as it has no public field.)
• We need some investigation (probably including a prototype) to make sure the current API can be implemented efficiently as a thin wrapper around PSA. Note: for crypto operations, this is already mostly the case with USE_PSA_CRYPTO; when using mbedtls_pk_setup_opaque() (or when MBEDTLS_PK_USE_PSA_EC_DATA is defined) that's also the case for (ECC) key storage.

In 4.0 or 4.x:

• Actually make it a thin wrapper around PSA crypto.
• Migrate all internal users (library and tests) to use PSA Crypto directly (note: common with option 2).

Rationale for option 2 (remove)

This leaves us with only one PK API to maintain, document and test. For new users, this also gives more clarity.

Note the current pk.h API has a few shortcomings, the most prominent being it doesn't manage key permissions like PSA does. So, pushing users to the PSA API is also pushing them to a better API.

Work needed for option 2 (remove)

In 4.0:

• Design and implement a PSA replacement for PK parse. Note: that's probably an EPIC on its own, but also something we want to do anyway.
• Update all TLS and X.509 APIs that handle a pk_context:
  • X.509: mbedtls_x509write_crt_set_subject_key(), mbedtls_x509write_crt_set_issuer_key() and mbedtls_x509write_csr_set_key() + public fields in mbedtls_x509_crt and mbedtls_x509_csr structure, and some private fields too.
  • TLS: mbedtls_ssl_conf_own_cert() and mbedtls_ssl_set_hs_own_cert().
• Move pk.h to an internal location and adapt all files that #include it.
• Change or remove all example programs that used it.
• Provide a migration guide.

In 4.0 or 4.x:

• Refactor psa_crypto_cipher.c so that it no longer
• Migrate all remaining internal users (library and tests) to use PSA Crypto directly (note: common with option 2).
• Then remove Cipher from the code base entirely.

Other options

• We could go for a mix: keep some parts but remove others. For example, we could keep PK parse or pk_context but remove the part about doing operations?
• Other?

[gilles-peskine-arm]

The general wisdom for API transitions is to have one full major version cycle where the new API is stable and the old API is present but deprecated. In Mbed TLS 3.6, we won't have a new API that covers the major features of the legacy pk API. In particular, PSA is missing the equivalent of pkparse and pkwrite.

Thus I don't see any choice. We need to keep pk (without its rough edges such as mbedtls_pk_ec and mbedtls_pk_rsa) in 4.x.

[mpg]

The general wisdom for API transitions is to have one full major version cycle where the new API is stable and the old API is present but deprecated.

I we want to follow that rule (which I think is generally speaking a good guideline), then even if we're keeping PK in 4.0, we should still provide PSA-based replacements for the X.509 and TLS APIs that use a key. So, for example, change mbedtls_ssl_conf_own_cert() to accept a PSA key slot instead of a pk_context and add a new, and immediately deprecated, function mbedtls_ssl_conf_own_cert_pk() - and similarly for the other APIs.

(Considering TLS and X.509 both use PK internally for now, the easiest way to implement this initially is for the new, PSA-based API to just internally wrap the PSA key in a PK context, and store the result.)

[mpg]

I forgot to note some optional features that PK parse has, for which it's unclear whether the upcoming PSA import extension will have support:

• parsing ECC keys where the public key is in compressed format (Weierstrass curves);
• parsing ECC keys where the curve is encoded as SpecifiedECDomain rather than NamedCurve.

(Note: PK parse only supports these options when at least some of the built-in implementation (ECP_LIGHT) is enabled.)

[gilles-peskine-arm]

parsing ECC keys where the public key is in compressed format (Weierstrass curves);
parsing ECC keys where the curve is encoded as SpecifiedECDomain rather than NamedCurve.

Yes, we will use the PSA formatted import API to support those. But I fear that it won't be in 4.0, only in 4.1 or so.

[mpg]

If we don't have a complete replacement in 4.0 but only in 4.1, I think that's a rather strong argument for keeping PK public in 4.0.

[mpg]

Note (just for reference, I don't think it influences the decision to keep PK or not): currently, internal users of PK are:

• TLS (even 1.3 and even with USE_PSA);
• X.509 and PKCS7 (even with USE_PSA);
• the built-in implementation of RSA in PSA (psa_crypto_rsa.c) uses PK parse and PK write.

[mpg]

I think there's a pretty strong argument for keeping PK in 4.0 so assuming we do, I think we need to be careful about its implementation and whether TLS & X.509 depend on it, in order to not make things harder for us in the future.

I think we want to avoid a situation where using TLS & X.509 require using PK and PK relies on psa_set_key_enrollment_algorithm() (I think it's OK if only one of this conditions is true), see part of the discussion at #8449.

More generally, I think currently our story about policies when PK internally stores key in PSA key slots (that is, for EC keys in builds where MBEDTLS_PK_USE_PSA_EC_DATA is defined) is not clean and we should clarify it. Might also be related with part of the discussion that happened here: #7930 (comment)

If we keep PK in 4.0, we probably need some hard thinking about its API and semantics.

[mpg]

We could go for a mix: keep some parts but remove others. For example, we could keep PK parse or pk_context but remove the part about doing operations?

The more I think about it, the more I like this option: keep PK-parse/write but remove the sign/verify, encrypt/decrypt APIs (hereunder PK-ops). Instead, the only thing you can to with a pk_context is make a PSA key out of it - and then you can use PSA APIs for crypto operations.

Rationale:

• The strongest argument for keeping PK is that we don't have a replacement yet - this only applies to PK-parse/write.
• The design of PK-ops is flawed and inconsistent when it comes to policies, key types and operations. Improving it requires significant work, and since it will involve incompatible API changes, would reduce the backward compatibility advantage of keeping it.