gm-crypto-rs

Constant-time-designed pure-Rust SM2 / SM3 / SM4 SDK for Chinese national cryptography (GB/T 32905 / 32918 / 32907 / GM/T 0009). Sign / verify, public-key encrypt / decrypt, SM4-CBC, SM4-CTR (single-shot + streaming), length-flexible batched SM4 block encryption, HMAC-SM3, PBKDF2-HMAC-SM3 — all secret-touching paths guarded by an in-CI dudect-bencher detectable-leak regression harness.

Personal project notice: not affiliated with, endorsed by, sponsored by, or certified by any upstream cryptography project, payment gateway, standards body, or vendor.

What this is

A small, auditable, pure-Rust SM2 / SM3 / SM4 SDK whose central differentiating commitment is that secret-touching code paths are constant-time-designed and guarded by an in-CI dudect-bencher detectable-leak regression harness: 17 real ct_* targets (12 always-on + 2 cfg-gated under sm4-bitsliced-simd + 3 cfg-gated under sm4-aead) plus a deliberately-leaky negative_control that proves the harness can detect leaks. Most real targets gate at |tau| < 0.20; ct_sign_k_class and the direct ct_fn_invert / ct_fp_invert invert diagnostics carry target-specific gate policy after the 2026-05-12 recalibration — see SECURITY.md and docs/v0.5-dudect-recalibration.md.

The harness reports timing-leak detection events. It does not prove constant-time. Low |tau| values mean the test could not detect a leak with the budget given, not that no leak exists. Language taken directly from dudect-bencher's own docs.

The harness covers: SM2 sign (split by both private key d and nonce k magnitude, with both retry nonces class-tied), SM2 decrypt (split by recipient d_B), SM4 key schedule + single-block encrypt (split by master key, under default linear-scan and sm4-bitsliced paths), the v0.5 SIMD-packed dispatch (ct_sm4_encrypt_block_bitsliced_simd, cfg-gated), v0.6's batched CBC-decrypt fanout (ct_sm4_cbc_decrypt_fanout, cfg-gated), v0.7's SM4-CTR encrypt (ct_sm4_ctr_encrypt, exercising the public batch path on every cipher matrix entry), v0.8's SM4-GCM + SM4-CCM decrypt (ct_sm4_gcm_decrypt and ct_sm4_ccm_decrypt, cfg-gated on sm4-aead), v0.9's incremental-input buffered SM4-GCM decrypt (ct_sm4_gcm_decrypt_buffered, cfg-gated on sm4-aead), HMAC-SM3 (split by key), encrypted-PKCS#8 decrypt (split by password bytes — both classes' blobs valid for their class's password so both succeed via identical control flow), plus direct Fn::invert and Fp::invert diagnostics. The ct_sign_k_class target closes v0.1's structural blind spot to nonce-only leaks.

The crypto-bigint 0.6 → 0.7.3 upgrade resolved the v0.1-era ConstMontyForm::invert leak directly: on the v0.2 W0 harness both direct invert diagnostics measured under |tau| ≈ 0.01, two orders of magnitude below the gate. Subsequent GH Actions runner-image drift on 2026-05-12 raised the empirical noise floor on ct_fn_invert / ct_fp_invert — both targets moved to PR-smoke telemetry + a nightly gross-regression sentinel at |tau| ≥ 0.55. See docs/v0.5-dudect-recalibration.md for the data and posture. See SECURITY.md for the full constant-time discipline.

The differentiator vs. existing Rust SM2 crates (notably RustCrypto/sm2, which already aims for constant-time secret-dependent operations in its design) is the in-CI regression gate, not the design intent in isolation.

What this isn't

• Not a TLS/TLCP implementation.
• Not SM9, ZUC, post-quantum.
• Not an HSM/SDF/SKF integration.
• Not a certified cryptographic module.
• Not constant-time on CPUs with data-dependent multiply latencies (some older x86, some embedded).
• Not a comprehensive SM-crypto library yet — see the milestone roadmap.

v0.12 scope (shipping)

SM4-XTS — tweakable mode for disk/sector encryption. v0.12 adds sm4::mode_xts behind the new opt-in sm4-xts feature: single-shot, full ciphertext stealing, GB/T 17964-2021 (GM-T OID 1.2.156.10197.1.104.10), byte-identical to OpenSSL 3.x EVP SM4-XTS (xts_standard=GB). Design rationale: docs/v0.12-scope.md; KAT sourcing: docs/v0.12-xts-kat-sourcing.md.

• Default-features users are unaffected — additive, opt-in, no new dependency (the XTS tweak doubling is a trivial bit-reflected multiply-by-x, not GHASH, so no gmcrypto-simd dep).
• GB/T 17964, not IEEE 1619 — the two standards differ in the GF(2¹²⁸) tweak-doubling convention (GB is the bit-reflected / GHASH-style one), so they produce different ciphertext for multi-block / non-aligned data. v0.12 targets GB (the SM4 national standard + OpenSSL's default for SM4-XTS).
• Confidentiality only — no authentication. XTS has no tag; callers needing integrity use an AEAD mode (GCM/CCM). The per-data-unit tweak-uniqueness contract is the caller's responsibility.
• 32-byte key (Key1 ‖ Key2) + raw 16-byte tweak; lengths [16 B, 16 MiB]; single None failure mode. New dudect target ct_sm4_xts_decrypt. The whole- block bulk rides the Sm4Cipher::encrypt_blocks batch API (picks up the SIMD fanout under sm4-bitsliced-simd).

Deferred to v0.13 (per docs/v0.12-scope.md §5/§6): C FFI for SM4-XTS, RustCrypto aead trait fit, pinned/noise-isolated dudect runner, AVX-512 sbox_x64, CCM incremental input.

v0.11 scope (shipped)

RustCrypto trait-fit modernization. v0.11 migrates the opt-in digest-traits / cipher-traits impls from digest 0.10 / cipher 0.4 to digest 0.11 / cipher 0.5 (the crypto-common 0.2 / hybrid-array generation), in-place. Design rationale: docs/v0.11-scope.md.

• Default-features users are unaffected — the trait fit is opt-in; generic-array / hybrid-array never enter the default dep graph, and every SM2 / SM3 / SM4 / HMAC / AEAD output is byte-identical (validated against the full KAT suite + gmssl 3.1.1 interop).
• BREAKING for trait-fit consumers only: code enabling digest-traits / cipher-traits must bump its own digest / cipher deps to 0.11 / 0.5. HMAC construction via the Mac trait moves to digest::KeyInit::new_from_slice (digest 0.11's Mac dropped KeyInit); the cipher block traits renamed BlockEncrypt / BlockDecrypt → BlockCipherEncrypt / BlockCipherDecrypt.
• MSRV stays 1.85. The RustCrypto aead 0.6 trait fit remains deferred (still 0.6.0-rc.10); v0.11 lands the crypto-common 0.2 line it will need.

Deferred to v0.12 (per docs/v0.11-scope.md §5/§6): RustCrypto aead trait fit, pinned/noise-isolated dudect runner, AVX-512 sbox_x64, SM4-XTS, CCM incremental input, Argon2-with-SM3.

v0.10 scope (shipped)

Streaming AEAD FFI. v0.10 exposes the v0.9 incremental-input buffered SM4-GCM encryptor/decryptor through the gmcrypto-c C ABI — the item v0.9 deferred (Q9.6) now that the Rust streaming API is proven. Additive behind the existing sm4-aead feature. Design rationale: docs/v0.10-scope.md.

• 9 streaming AEAD C FFI symbols + 2 opaque handle types — gmcrypto_sm4_gcm_encryptor_t (output-streaming: new / update → ciphertext per chunk / finalize + finalize_with_tag_len → tag / free) and gmcrypto_sm4_gcm_decryptor_t (commit-on-verify: new / update buffers and emits nothing / finalize_verify releases plaintext only after the constant-time tag check / free). _finalize* consume+free the handle; single GMCRYPTO_ERR on every failure (no tag-/length-oracle across the boundary). Mirrors the v0.5 CBC-streaming lifecycle. C example: examples/sm4_gcm_streaming.c.

No public API breakage — purely additive. v0.9.0 callers can cargo update to v0.10.0 without migration. No new gmcrypto-core API; no new dudect target (the FFI is a thin wrapper over the v0.9 ct_sm4_gcm_decrypt_buffered-gated path).

Deferred to v0.11 (per docs/v0.10-scope.md §5/§6): streaming/incremental CCM, RustCrypto aead trait fit (upstream still 0.6.0-rc.10), pinned dudect runner, AVX-512 sbox_x64, SM4-XTS, Argon2-with-SM3.

v0.9 scope (shipped)

AEAD ergonomics. v0.9 extends the v0.8 AEAD core with the three items v0.8 deferred: GCM tag-length parameterization, incremental-input buffered SM4-GCM, and single-shot AEAD C FFI. All additive behind the existing sm4-aead flag. Design rationale: docs/v0.9-scope.md.

• sm4::GcmTagLen + mode_gcm::encrypt_with_tag_len / decrypt_with_tag_len — W1. Caller-chosen GCM tag length per NIST SP 800-38D §5.2.1.2 ({4, 8, 12, 13, 14, 15, 16} bytes; truncated tag = MSB_t(full_tag)). GcmTagLen::new(usize) -> Option<Self> centralizes the valid-length policy. The fixed-16-byte encrypt / decrypt are unchanged.
• sm4::Sm4GcmEncryptor / Sm4GcmDecryptor — W2. Incremental- input buffered SM4-GCM (deliberately NOT "streaming"). The encryptor is output-streaming: update(chunk) -> Option<Vec<u8>> emits each chunk's ciphertext (None once the cumulative plaintext would exceed the NIST §5.2.1.1 ceiling 2^36 − 32 bytes); finalize() / finalize_with_tag_len() emit the tag. The decryptor is input-incremental but output-BUFFERED: update(chunk) buffers ciphertext + folds GHASH, and finalize_verify(tag) -> Option<Vec<u8>> releases the plaintext only after the constant-time tag check (commit-on-verify — never leaks pre-verify bytes). AAD is supplied at construction. Driven with any chunking, both reproduce the single-shot path byte-for-byte.
• 6 single-shot AEAD C FFI entry points — W4. gmcrypto_sm4_gcm_ encrypt / _decrypt / _encrypt_with_tag_len / _decrypt_with_tag_ len + gmcrypto_sm4_ccm_encrypt / _decrypt, behind a new forwarding sm4-aead feature on gmcrypto-c. Every error path returns GMCRYPTO_ERR (single failure code). Streaming AEAD FFI is deferred to v0.10.
• New dudect target ct_sm4_gcm_decrypt_buffered — W3. Class-split by master key, drives Sm4GcmDecryptor; |tau| < 0.20 (5K-sample smoke |τ| ≈ 0.029). No new CI matrix slot — rides the existing sm4-aead entries.

No public API breakage — purely additive. v0.8.0 callers can cargo update to v0.9.0 without migration; sm4-aead is opt-in.

Deferred to v0.10 (per docs/v0.9-scope.md §5/§6): CCM incremental input, streaming AEAD FFI, RustCrypto aead trait fit (upstream still on 0.6.0-rc), pinned dudect runner, AVX-512 sbox_x64, SM4-XTS, Argon2-with-SM3.

v0.8 scope (shipped)

The AEAD core. v0.8 cashed in the cipher-mode surface that v0.7 opened up: SM4-GCM and SM4-CCM single-shot, plus a constant-time GHASH primitive in gmcrypto-simd.

• sm4::mode_gcm::encrypt / decrypt — W2. Single-shot SM4-GCM per NIST SP 800-38D / GM/T 0009 / RFC 8998. encrypt(key, nonce, aad, pt) -> (Vec<u8>, [u8; 16]) returns (ciphertext, tag). decrypt(key, nonce, aad, ct, tag) -> Option<Vec<u8>> — Some(plaintext) only when the tag verifies (constant-time compare via subtle::ConstantTimeEq). Both 12-byte canonical and arbitrary-length nonce paths supported. Tag length fixed at 128 bits in v0.8 (parameterized in v0.9 via GcmTagLen). Byte-identical to gmssl 3.1.1 sm4 -gcm — bidirectional interop validated.
• sm4::mode_ccm::encrypt / decrypt — W3. Single-shot SM4-CCM per NIST SP 800-38C / RFC 3610 / GM/T 0009 (OID 1.2.156.10197.1.104.9). encrypt(key, nonce, aad, pt, tag_len) -> Option<Vec<u8>> (output: ciphertext ‖ tag). tag_len ∈ {4, 6, 8, 10, 12, 14, 16} per spec, validated at API entry. nonce.len() ∈ [7, 13]. Pure-Rust CBC-MAC + CTR over the existing Sm4Cipher path — no GHASH. Byte-identical to OpenSSL 3.x EVP SM4-CCM across 8 KAT scenarios (gmssl 3.1.1 doesn't ship sm4 -ccm so the CCM reference oracle comes from OpenSSL; see docs/v0.8-ccm-kat-sourcing.md).
• gmcrypto_simd::ghash::ghash_mul(h, x) -> [u8; 16] — W1. Constant-time GHASH multiplication over GF(2^128) / (x^128 + x^7 + x^2 + x + 1). Single dispatch entry point:
  • ghash_mul_clmul on x86_64 (PCLMULQDQ + SSE2; runtime cpufeatures detect; Intel Westmere+ / AMD Bulldozer+).
  • ghash_mul_pmull on aarch64 (ARMv8.0 AES extension vmull_p64; runtime cpufeatures detect; Apple Silicon / most modern ARM chips).
  • ghash_mul_software (bit-serial mask-XOR; constant-time over both inputs; available everywhere as fallback).
• New sm4-aead feature flag — default-off; opt-in. sm4-aead = ["dep:gmcrypto-simd"] activates mode_gcm and mode_ccm. Additive on the default-features build.
• New dudect targets ct_sm4_gcm_decrypt + ct_sm4_ccm_decrypt — W4. Class-split by master key over a fixed 256-byte plaintext + 16-byte AAD. Both classes' (ct, tag) pairs are valid encrypts under their own keys, so both decrypt paths reach the tag-compare via identical control flow. Same |tau| < 0.20 gate as the rest of the SM4 surface; new CI matrix slot sm4-bitsliced-simd,sm4-aead exercises the most-demanding cipher-stack combination.

No public API breakage — purely additive. v0.7.0 callers can cargo update to v0.8.0 without migration; sm4-aead is opt-in.

Everything v0.4 shipped (wasm32-unknown-unknown build, RustCrypto trait fit behind digest-traits / cipher-traits, bitsliced SM4 S-box behind sm4-bitsliced, gmcrypto-c C ABI crate) is unchanged — see the Roadmap row for the compact reference and CHANGELOG.md [0.4.0] for detail.

Everything v0.3 shipped is unchanged:

• Reusable strict-canonical DER reader / writer subset (gmcrypto_core::asn1::{reader, writer, oid}).
• PEM + encrypted PKCS#8 + X.509 SPKI + SEC1 codecs (gmcrypto_core::{pem, pkcs8, spki, sec1}).
• Full bidirectional gmssl 3.1.1 interop (SM2 sign / verify, SM2 encrypt / decrypt, SM4-CBC). Gated on GMCRYPTO_GMSSL=1.
• Raw byte-concat SM2 ciphertext helpers (gmcrypto_core::sm2::raw_ciphertext): C1 || C3 || C2 emit + decode; legacy C1 || C2 || C3 decrypt-only.
• Streaming HmacSm3 + Sm4Cbc{En,De}cryptor. In-crate Hash / Mac / BlockCipher traits (gmcrypto_core::traits).
• Comb-table mul_g (~5× sign-side speedup). 64 sub-tables of 16 entries each, lazily built once per process via spin::Once.

Everything v0.2 shipped is unchanged:

• SM3 hash function (#![no_std] + alloc).
• SM2 sign / verify with custom signer ID (default 1234567812345678 per GM/T 0009).
• SM2 public-key encrypt / decrypt with GM/T 0009-2012 ciphertext DER (SEQUENCE { x, y, hash, ciphertext }). Invalid-curve attack defense via on-curve check on C1 before scalar mult; non-branching KDF-zero detection so a chosen-ciphertext attacker cannot distinguish it from a normal MAC failure.
• SM4 block cipher (GB/T 32907-2016) and SM4-CBC (PKCS#7 padding, caller-supplied unpredictable IV per NIST SP 800-38A Appendix C). Constant-time-designed subtle linear-scan S-box (~1-2M blocks/s); opt-in bitsliced (table-less, gate-only) S-box via the sm4-bitsliced feature (v0.4 W3). PKCS#7 strip uses a constant-time scan over the final block; decrypt collapses every failure mode to a single None against padding-oracle attacks.
• HMAC-SM3 per RFC 2104, gmssl-cross-validated KAT vectors. Hash-first long-key path. v0.3 adds the streaming HmacSm3 shape alongside single-shot hmac_sm3.
• PBKDF2-HMAC-SM3 per RFC 8018 §5.2. Caller-supplied output buffer (no internal allocation, no iteration-count default).
• Constant-time-designed Fp and Fn field arithmetic via crypto-bigint = 0.7.3.
• Renes-Costello-Batina complete addition formulas for the SM2 curve (a=-3 specialized).
• Fixed-base (v0.3 comb-table) and variable-base scalar multiplication, both constant-time-designed with subtle::ConditionallySelectable linear-scan table lookup.
• Fixed-K masked-select signing retry: the retry loop runs K=2 iterations unconditionally, regardless of which iteration produced a valid signature. The constant-time contract holds for any RNG that respects CryptoRng; pathological RNGs cannot leak the secret via observable retry count.
• Strict canonical ASN.1 DER for SEQUENCE { r, s } (signatures), the GM/T 0009 SM2 ciphertext SEQUENCE, and all v0.3 PEM / PKCS#8 / SPKI / SEC1 wire formats. Rejects non-canonical leading-zero padding, sign-bit-set first bytes, empty content, and (for ciphertext coordinates) values ≥ p.
• KAT vectors from GB/T 32905-2016 (SM3), GB/T 32918.2-2017 / .5-2017 (SM2), GB/T 32907-2016 Appendix A.1 (SM4 single-block + 1M-round), GM/T 0042-2015 (HMAC-SM3), GM/T 0091-2020 (PBKDF2-HMAC-SM3).
• gmssl CLI cross-validation for HMAC-SM3, PBKDF2-HMAC-SM3, and (new in v0.3) SM2 sign/verify, SM2 encrypt/decrypt, and SM4-CBC in both directions. Gated on GMCRYPTO_GMSSL=1.
• dudect-bencher harness — 17 real ct_* targets (12 always-on + 2 cfg-gated under sm4-bitsliced-simd + 3 cfg-gated under sm4-aead) plus a deliberately-leaky negative_control that proves the harness can detect leaks. Matrix-run under features=default, sm4-bitsliced, sm4-bitsliced-simd, and sm4-bitsliced-simd,sm4-aead — PR-smoke 10⁴ samples; nightly 10⁵ samples (more samples = tighter empirical confidence at the same threshold). Most real targets gate at |tau| < 0.20; per-target policy in SECURITY.md.
• Failure-mode invariant: every Result-returning public API uses the workspace-wide gmcrypto_core::Error (single Failed variant, #[non_exhaustive]); per-module aliases sm2::Error, pem::Error, pkcs8::Error all point at the same type. verify_with_id returns bool; DER decode returns Option. Defense against padding-oracle, malleability, and invalid-curve attacks.
• Zeroization on private keys, SM4 round keys, HMAC K' / K' XOR ipad / K' XOR opad, PBKDF2 intermediates, SM2 KDF buffers, and PKCS#8 inner-key scratch.

Roadmap

Quick-start

use gmcrypto_core::sm2::{
    sign_with_id, verify_with_id, Sm2PrivateKey, Sm2PublicKey, DEFAULT_SIGNER_ID,
};
use getrandom::SysRng;
use hex_literal::hex;
use rand_core::UnwrapErr;

// v0.5 W5 — `from_bytes_be` is the recommended public constructor
// (always-on, doesn't expose `crypto_bigint::U256` to callers).
let d_be: [u8; 32] = hex!(
    "3945208F7B2144B13F36E38AC6D39F95889393692860B51A42FB81EF4DF7C5B8"
);
let key = Sm2PrivateKey::from_bytes_be(&d_be).expect("d in [1, n-2]");
let public = Sm2PublicKey::from_point(key.public_key());

let mut rng = UnwrapErr(SysRng);
let sig = sign_with_id(&key, DEFAULT_SIGNER_ID, b"hello", &mut rng).unwrap();
assert!(verify_with_id(&public, DEFAULT_SIGNER_ID, b"hello", &sig));

Threat model

See SECURITY.md. Briefly: server-side use, dedicated host, operator-trusted, network MITM in scope, side-channel attacks beyond what the dudect harness covers are NOT in scope.

Build & test

cargo test --workspace                                                          # unit + integration
cargo bench --bench timing_leaks --features crypto-bigint-scalar                # local timing harness (~75s)
DUDECT_SAMPLES=10000 cargo bench --bench timing_leaks --features crypto-bigint-scalar  # match CI smoke budget

gmssl interop test (gated; install gmssl v3.1.1 to enable):

GMCRYPTO_GMSSL=1 cargo test --test interop_gmssl

wasm32 support

gmcrypto-core builds on wasm32-unknown-unknown as of v0.4. CI gates both stable and MSRV (1.85) builds on the target.

rustup target add wasm32-unknown-unknown
cargo build -p gmcrypto-core --target wasm32-unknown-unknown --no-default-features

The crate is no_std + alloc only and does NOT pull getrandom's wasm_js backend or wasm-bindgen / js-sys into its default dep graph. Wasm callers wire their own rand_core::Rng impl — typically by enabling getrandom's wasm_js feature in their Cargo.toml:

[dependencies]
gmcrypto-core = "0.7"
rand_core = { version = "0.10", default-features = false }
getrandom = { version = "0.4", default-features = false, features = ["wasm_js"] }

use gmcrypto_core::sm2::{sign_with_id, Sm2PrivateKey, DEFAULT_SIGNER_ID};
use rand_core::UnwrapErr;
use getrandom::SysRng;

let mut rng = UnwrapErr(SysRng); // wasm_js-backed when targeting wasm32
let sig = sign_with_id(&priv_key, DEFAULT_SIGNER_ID, b"msg", &mut rng).unwrap();

A wasm-bindgen-test-driven test runner (running KAT vectors under Node or a headless browser) is post-v0.4 — v0.4 ships the build-target gate only.

License

Apache-2.0. See LICENSE.

Some reference outputs use the upstream gmssl tool. This project is independent of that project.