Skip to content

Reject invalid API inputs before unsafe work#14

Open
fegge wants to merge 2 commits into
mainfrom
fix/api-input-validation-and-memory-safety
Open

Reject invalid API inputs before unsafe work#14
fegge wants to merge 2 commits into
mainfrom
fix/api-input-validation-and-memory-safety

Conversation

@fegge

@fegge fegge commented Jun 26, 2026

Copy link
Copy Markdown
Collaborator

Summary

  • Reject invalid external-mu lengths before fixed-size copies and allocations
  • Reject invalid signature lengths before verifier allocations
  • Reject invalid context pointer/length pairs before prefix allocation/copying
  • Reject signed-message lengths that overflow MLDSA_BYTES + mlen before output indexing
  • Add allocator and functional regressions for the early rejection paths

Validation

  • clang-format -i mldsa/src/sign.c mldsa/src/sign.h mldsa/mldsa_native.h test/src/test_alloc.c test/src/test_mldsa.c
  • git diff --check
  • make run_alloc run_func -j4

Full repo scripts/format and scripts/lint were not run because this environment lacks shfmt and nixpkgs-fmt.

hanno-becker and others added 2 commits June 26, 2026 15:15
Verify that each assembly kernel preserves the callee-saved registers
its platform calling convention requires:

- AArch64 (AAPCS64):   x19-x28, x29/FP, lower 64 bits of d8-d15.
- x86_64 (System V):   rbx, rbp, r12-r15. No SIMD register callee-saved.
- Armv8.1-M (AAPCS32): r4-r11 plus MVE Q4-Q7 (= D8-D15).

A per-arch assembly call stub loads a random register state, calls the
kernel, and captures the result; a checker then confirms the callee-saved
set is intact. Pointer arguments are backed by correctly-sized buffers
whose layout comes from a per-kernel YAML block; scripts/autogen turns
that YAML into the per-kernel checks. A self-test of hand-written
corrupters confirms the checker actually fires before kernel verdicts
are trusted.

Run via `make run_abicheck OPT=1` or `scripts/tests abicheck`. It needs no
library build, so `scripts/tests all` runs it by default.

Ported from mlkem-native, with the following deviations:

- No PowerPC64/ELFv2 backend, which mldsa-native does not ship.
- Fresh ABI YAML for the ML-DSA kernel set and signatures.
- Fixed the FIPS202 Keccak YAML Name fields to match the exported
  symbols (e.g. keccak_f1600_x1_scalar_aarch64_asm, not
  keccak_f1600_x1_scalar_asm), which the generator maps to mld_<Name>.
- The build force-undefines MLD_CONFIG_NO_{KEYPAIR,SIGN,VERIFY}_API so
  every kernel is present; many ML-DSA kernels are otherwise gated out
  by reduced-API configs, unlike any mlkem kernel.
- Adds MLD_SYS_CAP_MVE to sys.h (mlkem already has MLK_SYS_CAP_MVE), as
  the Armv8.1-M check needs it for its runtime capability gate.

AArch64 ABI YAML lives in dev/aarch64_clean/src/, not dev/aarch64_opt/src/,
which scripts/autogen regenerates from clean.

See test/abicheck/README.md for details.

Signed-off-by: Hanno Becker <beckphan@amazon.co.uk>
@fegge

fegge commented Jun 26, 2026

Copy link
Copy Markdown
Collaborator Author

Upstream issue: pq-code-package#1212
Upstream PR: pq-code-package#1223

@fegge

fegge commented Jun 26, 2026

Copy link
Copy Markdown
Collaborator Author

Replacement upstream PR with DCO-signed commits from an upstream repository branch: pq-code-package#1231

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants