Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Use mlkem-native as AWS-LC's ML-KEM implementation
This imports mlkem-native (https://github.com/pq-code-package/mlkem-native) into AWS-LC, replacing the reference implementation. This commit focuses on the minimal configuration of mlkem-native: No assembly and no FIPS-202 code are imported. mlkem-native is a high-performance, high-assurance C90 implementation of ML-KEM developed under the Post-Quantum Cryptography Alliance (PQCA) and the Linux Foundation. It is a fork of the reference implementation that AWS-LC previously relied on, and remains close to it. mlkem-native is the default ML-KEM implementation in [libOQS](https://github.com/open-quantum-safe/liboqs). **Import Mechanism** The mlkem-native source code is unmodified and imported using the importer script `crypto/fipsmodule/ml_kem/importer.sh`; the details of the import are in META.yml. Future updates to the source tree would ideally happen through a re-import of a different version of mlkem-native, though a temporary change-log is conceivable, similar to how the changes from the reference implementation were documented so far. **Import Scope** mlkem-native has a C-only version as well as native 'backends' in AVX2 and Neon for high performance. This commit only imports the C-only version. Integration of native backends will be done separately. mlkem-native offers its own FIPS-202 implementation, including fast versions of batched FIPS-202. However, this commit does not import those, but instead provides glue-code around AWS-LC's own FIPS-202 implementation. The path to leveraging the FIPS-202 performance improvements in mlkem-native would be to integrate them directly into [crypto/fipsmodule/sha](crytpo/fipsmodule/sha). **Side-channels** mlkem-native's CI uses a patched version of valgrind to check for various compilers and compile flags that there are no secret-dependent memory accesses, branches, or divisions. The relevant assertions have been kept but are unused unless `MLK_CT_TESTING_ENABLED` is set, which is the case if and only if `BORINGSSL_CONSTANT_TIME_VALIDATION` is set. Similar to AWS-LC, mlkem-native uses value barriers to block potentially harmful compiler reasoning and optimization. Where standard gcc/clang inline assembly is not available, mlkem-native falls back to a slower 'opt blocker' based on a volatile global (an idea by DjB) -- both is described in [verify.h](https://github.com/aws/aws-lc/blob/df5b09029e27d54b2b117eeddb6abd983528ae15/crypto/fipsmodule/ml_kem/mlkem/verify.h). It will be interesting to see if the opt-blocker variant works on all platforms that AWS-LC cares about. **Formal Verification** All C-code imported in this commit is formally verified using the C Bounded Model Checker ([CBMC](https://github.com/diffblue/cbmc/)) to be free of various classes of undefined behaviour, including out-of-bounds memory accesses and arithmetic overflow; the latter is of particular interest for ML-KEM because of the use of lazy modular reduction for improved performance. The heart of the CBMC proofs are function contract and loop annotations to the C-code. Function contracts are denoted `__contract__(...)` clauses and occur at the time of declaration, while loop contracts are denoted `__loop__` and follow the `for` statement. The function contract and loop statements are kept in the source, but removed by the preprocessor so long as the CBMC macro is undefined. Keeping them simplifies the import, and care has been taken to make them readable to the non-expert, and thereby serve as precise documentation of assumptions and guarantees upheld by the code. The CBMC proofs are automatic and don't require further proofs scripts; yet, they come with their own build system and toolchain dependencies, which this commit does not attempt to import. See [proofs/cbmc](https://github.com/pq-code-package/mlkem-native/tree/main/proofs/cbmc) in the mlkem-native repository. Mid-term, however, CI infrastructure should be setup that allows to import and check the CBMC proofs as part of the AWS-LC CI. **FIPS Compliance** The current reference implementation in AWS-LC accommodates FIPS (IG) requirements via: * Adding explicit stack buffer via `OPENSSL_cleanse` * Adding a Pairwise Consistency Test (PCT) after key generation (only for the FIPS-build) mlkem-native unconditionally includes stack zeroization. mlkem-native's default secure `memset` is replaced by `OPENSSL_cleanse`. mlkem-native conditionally includes a PCT, guarded by `MLK_KEYGEN_PCT`. This is set in the config if and only if `AWSLC_FIPS` is set. **Performance** It is expected -- but should be checked! -- that the ML-KEM performance with this PR is comparable to that of the reference implementation. This is because the mlkem-native's fast backends are not yet imported, the FIPS-202 code remains that of AWS-LC, and mlkem-native is otherwise close to the reference implementation. **Multilevel build** At the core, mlkem-native is currently a 'single-level' implementation of ML-KEM: A build of the main source tree provides an implementation of exactly one of ML-KEM-512/768/1024, depending on the MLKEM_K parameter. This property is inherited from the ML-KEM reference implementation, while AWS-LC's fork of the reference implementation has changed this behaviour and passes the security level as a runtime parameter. To build all security levels, level-specific sources are built 3 times, once per security level, and linked with a single build of the level-independent code. The single-compilation-unit approach pursued by AWS-LC makes this process fairly simple since one merely needs to include the single-compilation-unit file provided by mlkem-native three times, and configure it so that the level-independent code is included only once. The final include moreover `#undef`'ines all macros defined by mlkem-native, reducing the risk of name clashes with other parts of crypto/fipsmodule/bcm.c. Signed-off-by: Hanno Becker <beckphan@amazon.co.uk>
- Loading branch information