Skip to content

Should a key derivation continue working after a verify failure? #112

@gilles-peskine-arm

Description

@gilles-peskine-arm

In a nutshell: if psa_key_derivation_verify_bytes detects incorrect output (INVALID_SIGNATURE condition), may this abort the operation?

(Everything about verify_bytes here applies equally to verify_key.)


Out-of-capacity state

Generally, when a function fails in a multipart operation, the multipart operation enters an error state.

The specification of psa_key_derivation_output_bytes suggests, but does not clearly mandate a different behavior for PSA_ERROR_INSUFFICIENT_DATA:

The operation’s capacity was less than output_length bytes. In this case, the following occurs:

  • No output is written to the output buffer.
  • The operation’s capacity is set to zero — subsequent calls to this function will not succeed, even with a smaller output buffer.

And in general:

If this function returns an error status other than PSA_ERROR_INSUFFICIENT_DATA, the operation enters an error state and must be aborted by calling psa_key_derivation_abort().

The intent is that on an INSUFFICIENT_DATA condition, the operation object should enter an out-of-capacity state. In that state, subsequent calls to functions that request more output (output_bytes, output_key, verify_bytes, etc.) will return PSA_ERROR_INSUFFICIENT_DATA, rather than PSA_ERROR_BAD_STATE. However the specification does not explicitly state that it would be wrong for an implementation to return PSA_ERROR_BAD_STATE for subsequent output calls in this scenario.

Proposed action: explicitly describe the out-of-capacity state and require that an INSUFFICIENT_DATA condition enters it (absent other errors). Applies to psa_key_derivation_output_bytes, psa_key_derivation_output_key, psa_key_derivation_verify_bytes, psa_key_derivation_verify_key.

Note that there is an edge case of requesting 0 bytes when the remaining capacity is 0. Should this succeeds because the whole requested output can be produced? Or should this fail so as not to make a distinction between capacity-0 and capacity-exceeded? In other words, if the initial capacity is 1 then what happens with the following sequence of calls?

  1. output_bytes(1) → success
  2. output_bytes(0) → success or INSUFFICIENT_DATA?
  3. output_bytes(0) → success or INSUFFICIENT_DATA?

Behavior after a failed verification

Generally, when a function fails in a multipart operation, the multipart operation enters an error state.

The specification of psa_key_derivation_verify_bytes suggests, but does not actually mandate different behavior on an INVALID_SIGNATURE condition:

PSA_ERROR_INVALID_SIGNATURE

The output of the key derivation operation does not match the value in expected_output.

(…)

If this function returns an error status other than PSA_ERROR_INSUFFICIENT_DATA or PSA_ERROR_INVALID_SIGNATURE, the operation enters an error state and must be aborted by calling psa_key_derivation_abort().

Based on this sentence, and on the code snippet that approximates the intended behavior of psa_key_derivation_verify_bytes, it seems that the intent is that an INVALID_SIGNATURE condition consumes the output, but does not affect the state of the object.

However, the specification does not state this clearly. It seems to me that an implementation could make the operation object enter an error state when it reaches an INVALID_SIGNATURE condition.

Which behavior is desirable? I can see the appeal for an implementation to treat verify_bytes as a wrapper around output_bytes which does not otherwise affect the state of the object. However, this isn't quite right since a PSA_ERROR_NOT_PERMITTED condition does enter an error state, and the permission requirements are different for output_bytes and verify_bytes. Furthermore, it's inconsistent with the fact that the specification does require putting the operation in an error state if psa_key_derivation_verify_key fails to access the key to compare against (e.g. PSA_ERROR_INVALID_HANDLE condition).

From a usage perspective, I can't think of a use case where it's desirable to keep the operation going after an INVALID_SIGNATURE condition. It's somewhat uncommon to use the same inputs to both validate part of the output, and construct a key or other material with another part of the output. A typical case where it does happen is using the first half of the output of a password hashing function as a password hash, and the second half as a key encryption key). In this case, a verification failure indicates a wrong password, so the user is not authenticated and the putative key encryption key should not be used.

Proposed action: don't give any special treatment to PSA_ERROR_INVALID_SIGNATURE. Thus it will cause the operation to enter the error state. Applies to psa_key_derivation_verify_bytes, psa_key_derivation_verify_key.

It's not clear to me whether the specification intended to mandate or allow implementations to preserve the operation's validity on INVALID_SIGNATURE. So I'm not sure if this proposed action is merely restricting the permitted behaviors, or if it would be an incompatible change.

Metadata

Metadata

Assignees

No one assigned

    Labels

    API designRelated the design of the APICrypto APIIssue or PR related to the Cryptography APIclarificationSomething is confusing or missing in the documentationenhancementNew feature or request

    Type

    No type

    Projects

    Status

    Done

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions