krypteia-arcana 0.1.0

//! X25519 Diffie-Hellman key agreement on Curve25519 (RFC 7748).
//!
//! X25519 is the "EDDH" (Edwards/Montgomery Diffie-Hellman) pairing
//! around Curve25519 — a Montgomery curve over `p = 2^255 - 19`. It
//! is the ECDH variant used by TLS 1.3, Noise, Signal, WireGuard,
//! and most modern protocols that want a fixed, fast, audit-friendly
//! ECDH.
//!
//! Unlike the short-Weierstrass curves exposed through the
//! [`super::curves::Curve`] trait, X25519 works entirely on the
//! **u-coordinate** (the x-coordinate of the Montgomery form). There
//! is no point addition, no compressed/uncompressed distinction, and
//! no branch on the y-coordinate — the Montgomery ladder operates on
//! projective (X:Z) pairs and returns a single 32-byte little-endian
//! u-coordinate.
//!
//! # Side-channel posture
//!
//! Per `arcana/doc/sca/countermeasures/x25519_x448.rst`:
//!
//! | Threat                                            | Status     | Roadmap item                                              |
//! |---------------------------------------------------|------------|-----------------------------------------------------------|
//! | Cache-timing on Montgomery ladder                 | partial    | `T1-G` — audit pass mirroring Weierstrass commit `76191c1`|
//! | SPA on Cortex-M0 (Weissbart-Picek-Batina 2021)    | vulnerable | `T1-G` + `T2-A` (Z-rerand defeats their template attack)  |
//! | DPA on field operations                           | vulnerable | `T2-A` — Z-rerandomization on `(X : Z)`                   |
//! | Template attacks                                  | vulnerable | `T2-A` (alignment break) + `T2-B` (scalar blinding)       |
//! | Invalid-curve attack on peer pubkey               | covered    | RFC 7748 twist security                                   |
//! | Small-subgroup contributory check                 | partial    | `T2-K` — confirm CT all-zero rejection                    |
//!
//! Curve25519 is CT **by construction** (single u-coordinate, no
//! special cases for the neutral element), but the concrete
//! Rust implementation can still leak through the `cswap` mask
//! pattern (see `super::field` `black_box` shielding) and
//! through unmodelled cache-line accesses. The
//! `weissbart2021_curve25519_ml_sca` paper demonstrated
//! deep-learning template attacks on Cortex-M0 even against
//! random-delay defences; Z-rerandomization is the standard
//! answer.
//!
//! # API
//!
//! ```rust,ignore
//! use arcana::ecc::x25519::{x25519_derive_public, x25519_ecdh};
//!
//! // Alice and Bob each draw 32 random bytes as their secret key.
//! let alice_sk: [u8; 32] = /* rng */;
//! let bob_sk:   [u8; 32] = /* rng */;
//!
//! // Derive public keys.
//! let alice_pk = x25519_derive_public(&alice_sk);
//! let bob_pk   = x25519_derive_public(&bob_sk);
//!
//! // Exchange public keys, then each derives the shared secret.
//! let s_ab = x25519_ecdh(&alice_sk, &bob_pk);
//! let s_ba = x25519_ecdh(&bob_sk,   &alice_pk);
//! assert_eq!(s_ab, s_ba);
//! ```
//!
//! # Test vectors
//!
//! The tests at the bottom of this file pin the two §5.2 primitive
//! vectors and the §6.1 full Diffie-Hellman vector directly from
//! RFC 7748. Any future regression in the ladder, the clamping, or
//! the LE byte encoding fails against those bytes immediately.

use super::field::*;

// ============================================================================
// Curve constants (RFC 7748 §4.1)
// ============================================================================

/// `a24 = (486662 - 2) / 4 = 121665` — the constant used inside the
/// Montgomery ladder's doubling step.
///
/// Declared as a `FieldElement<4>` so it can be fed directly into
/// `field_mul` without re-encoding on every call.
fn a24() -> FieldElement<4> {
    let mut fe = FieldElement::<4>::ZERO;
    fe.limbs[0] = 121_665;
    fe
}

/// X25519 base point u-coordinate = 9 (RFC 7748 §4.1).
const BASE_U: [u8; 32] = {
    let mut b = [0u8; 32];
    b[0] = 9;
    b
};

// ============================================================================
// Scalar clamping + u-coordinate decoding (RFC 7748 §5)
// ============================================================================

/// RFC 7748 §5 `decodeScalar25519`: clamps the 32-byte secret scalar
/// so that the resulting integer always lies in [2^254, 2^255) and
/// is a multiple of 8. This is the property that lets the Montgomery
/// ladder always iterate exactly 255 times with the top bit at
/// position 254 guaranteed to be 1.
///
/// Concretely:
/// - Clear the 3 low bits of byte 0 (= make the scalar a multiple of 8)
/// - Clear the high bit of byte 31 (= force bit 255 = 0)
/// - Set bit 6 of byte 31 (= force bit 254 = 1)
fn decode_scalar(scalar: &[u8; 32]) -> [u8; 32] {
    let mut k = *scalar;
    k[0] &= 248; // 0b11111000
    k[31] &= 127; // 0b01111111
    k[31] |= 64; // 0b01000000
    k
}

/// RFC 7748 §5 `decodeUCoordinate` (Curve25519 variant):
/// clear the most significant bit of the last byte, then interpret
/// as a little-endian integer in `Fp`.
///
/// Clearing the top bit matters because some peers (historically TLS)
/// don't mask it themselves; the spec requires the receiver to be
/// permissive.
fn decode_u(u: &[u8; 32]) -> FieldElement<4> {
    let mut buf = *u;
    buf[31] &= 0x7f;
    FieldElement::<4>::from_bytes_le(&buf)
}

/// Encode a field element as 32 little-endian bytes.
fn encode_u(fe: &FieldElement<4>) -> [u8; 32] {
    let v = fe.to_bytes_le();
    let mut out = [0u8; 32];
    out.copy_from_slice(&v);
    out
}

// ============================================================================
// Constant-time conditional swap of two field elements
// ============================================================================

/// Constant-time swap of `a` and `b` iff `swap == 1`.
///
/// `swap` must be 0 or 1. Uses a word-wide mask and XOR trick so the
/// data dependency is the same for both branches (essential: the
/// `swap` bit is derived from the secret scalar).
fn ct_swap_fe(a: &mut FieldElement<4>, b: &mut FieldElement<4>, swap: u64) {
    let mask = 0u64.wrapping_sub(swap);
    for i in 0..4 {
        let t = mask & (a.limbs[i] ^ b.limbs[i]);
        a.limbs[i] ^= t;
        b.limbs[i] ^= t;
    }
}

// ============================================================================
// The X25519 primitive — RFC 7748 §5, Montgomery ladder on (X:Z)
// ============================================================================

/// RFC 7748 §5 `X25519(scalar, u)`.
///
/// Takes a 32-byte little-endian scalar and a 32-byte little-endian
/// u-coordinate, and returns the 32-byte little-endian u-coordinate
/// of `scalar * (u, v)` on Curve25519 (where `v` is uniquely
/// determined by `u` up to sign — we never need it).
///
/// The ladder operates on projective `(X, Z)` pairs where the affine
/// u-coordinate is `X/Z`. 255 ladder steps are performed (bits 254..0
/// of the clamped scalar); the high bit 254 is always 1 after
/// clamping, so the first iteration deterministically initialises
/// (x_2, x_3) = (u, 1), (z_2, z_3) = (1, u) via the cswap.
pub fn x25519(scalar: &[u8; 32], u: &[u8; 32]) -> [u8; 32] {
    let k = decode_scalar(scalar);
    let x1 = decode_u(u);
    let a24 = a24();
    let p = &CURVE25519_P;

    // Ladder state: x_2 = 1, z_2 = 0, x_3 = x1, z_3 = 1.
    let mut x_2 = FieldElement::<4>::one();
    let mut z_2 = FieldElement::<4>::ZERO;
    let mut x_3 = x1;
    let mut z_3 = FieldElement::<4>::one();

    // `swap` tracks the conditional swap state between iterations.
    // See RFC 7748 §5 for the reference Python pseudocode.
    let mut swap: u64 = 0;

    for t in (0..=254).rev() {
        let k_t = ((k[t >> 3] >> (t & 7)) & 1) as u64;
        swap ^= k_t;
        ct_swap_fe(&mut x_2, &mut x_3, swap);
        ct_swap_fe(&mut z_2, &mut z_3, swap);
        swap = k_t;

        // A  = x_2 + z_2
        // AA = A^2
        // B  = x_2 - z_2
        // BB = B^2
        // E  = AA - BB
        // C  = x_3 + z_3
        // D  = x_3 - z_3
        // DA = D * A
        // CB = C * B
        // x_3 = (DA + CB)^2
        // z_3 = x_1 * (DA - CB)^2
        // x_2 = AA * BB
        // z_2 = E * (AA + a24 * E)
        let a = field_add(&x_2, &z_2, p);
        let aa = field_sqr(&a, p);
        let b = field_sub(&x_2, &z_2, p);
        let bb = field_sqr(&b, p);
        let e = field_sub(&aa, &bb, p);
        let c = field_add(&x_3, &z_3, p);
        let d = field_sub(&x_3, &z_3, p);
        let da = field_mul(&d, &a, p);
        let cb = field_mul(&c, &b, p);

        let da_plus_cb = field_add(&da, &cb, p);
        x_3 = field_sqr(&da_plus_cb, p);

        let da_minus_cb = field_sub(&da, &cb, p);
        let da_minus_cb_sq = field_sqr(&da_minus_cb, p);
        z_3 = field_mul(&x1, &da_minus_cb_sq, p);

        x_2 = field_mul(&aa, &bb, p);

        let a24_e = field_mul(&a24, &e, p);
        let aa_plus_a24e = field_add(&aa, &a24_e, p);
        z_2 = field_mul(&e, &aa_plus_a24e, p);
    }

    // Final swap based on the residual `swap` state.
    ct_swap_fe(&mut x_2, &mut x_3, swap);
    ct_swap_fe(&mut z_2, &mut z_3, swap);

    // Return x_2 / z_2 = x_2 * z_2^{p-2} mod p as 32 LE bytes.
    let z_inv = field_inv(&z_2, p);
    let result = field_mul(&x_2, &z_inv, p);
    encode_u(&result)
}

// ============================================================================
// Public API — keygen + ECDH convenience wrappers
// ============================================================================

/// Derive the X25519 **public key** from a 32-byte secret key.
///
/// The secret key is any 32 random bytes; the caller is responsible
/// for drawing them from a suitable CSPRNG. The returned public key
/// is the 32-byte u-coordinate of `sk * base` on Curve25519.
///
/// This is `X25519(sk, 9)` per RFC 7748 §5.
pub fn x25519_derive_public(sk: &[u8; 32]) -> [u8; 32] {
    x25519(sk, &BASE_U)
}

/// X25519 Diffie-Hellman: derive a shared secret from our secret key
/// and the peer's public key.
///
/// Returns the 32-byte u-coordinate of `sk * peer_pk`. This is the
/// raw shared secret (NIST SP 800-56A "Z"); pass it to an HKDF or
/// similar KDF before using it for symmetric keying.
///
/// # Small-subgroup attack note
///
/// RFC 7748 §6.1 warns that X25519 accepts certain "contributory"
/// public keys (points in small subgroups) that collapse the shared
/// secret to a fixed value. A defensive implementation may want to
/// reject the result if it is all-zero, which signals the peer sent
/// a low-order point. We deliberately do **not** reject here because
/// (a) the spec allows it, and (b) a downstream KDF with context
/// binding (TLS 1.3 `transcript_hash`, Noise `HKDF`, ...) is the
/// standard mitigation. Callers who want the low-order check can
/// run it themselves:
///
/// ```rust,ignore
/// let shared = x25519_ecdh(&sk, &peer_pk);
/// if shared.iter().all(|&b| b == 0) {
///     return Err("contributory shared secret");
/// }
/// ```
pub fn x25519_ecdh(sk: &[u8; 32], peer_pk: &[u8; 32]) -> [u8; 32] {
    x25519(sk, peer_pk)
}

// ============================================================================
// Tests (RFC 7748 pinned vectors)
// ============================================================================

#[cfg(test)]
mod tests {
    use super::*;

    fn hex32(h: &str) -> [u8; 32] {
        assert_eq!(h.len(), 64);
        let mut out = [0u8; 32];
        for i in 0..32 {
            out[i] = u8::from_str_radix(&h[2 * i..2 * i + 2], 16).unwrap();
        }
        out
    }

    // ----- RFC 7748 §5.2 primitive test vector #1 -----
    //
    // Scalar:      a546e36bf0527c9d3b16154b82465edd62144c0ac1fc5a18506a2244ba449ac4
    // u:           e6db6867583030db3594c1a424b15f7c726624ec26b3353b10a903a6d0ab1c4c
    // X25519(k,u): c3da55379de9c6908e94ea4df28d084f32eccf03491c71f754b4075577a28552
    #[test]
    fn rfc7748_section_5_2_vector_1() {
        let scalar = hex32("a546e36bf0527c9d3b16154b82465edd62144c0ac1fc5a18506a2244ba449ac4");
        let u = hex32("e6db6867583030db3594c1a424b15f7c726624ec26b3353b10a903a6d0ab1c4c");
        let expected = hex32("c3da55379de9c6908e94ea4df28d084f32eccf03491c71f754b4075577a28552");
        let got = x25519(&scalar, &u);
        assert_eq!(got, expected);
    }

    // ----- RFC 7748 §5.2 primitive test vector #2 -----
    //
    // Scalar:      4b66e9d4d1b4673c5ad22691957d6af5c11b6421e0ea01d42ca4169e7918ba0d
    // u:           e5210f12786811d3f4b7959d0538ae2c31dbe7106fc03c3efc4cd549c715a493
    // X25519(k,u): 95cbde9476e8907d7aade45cb4b873f88b595a68799fa152e6f8f7647aac7957
    #[test]
    fn rfc7748_section_5_2_vector_2() {
        let scalar = hex32("4b66e9d4d1b4673c5ad22691957d6af5c11b6421e0ea01d42ca4169e7918ba0d");
        let u = hex32("e5210f12786811d3f4b7959d0538ae2c31dbe7106fc03c3efc4cd549c715a493");
        let expected = hex32("95cbde9476e8907d7aade45cb4b873f88b595a68799fa152e6f8f7647aac7957");
        let got = x25519(&scalar, &u);
        assert_eq!(got, expected);
    }

    // ----- RFC 7748 §6.1 full Diffie-Hellman test vector -----
    //
    // Alice's private key: 77076d0a7318a57d3c16c17251b26645df4c2f87ebc0992ab177fba51db92c2a
    // Alice's public key:  8520f0098930a754748b7ddcb43ef75a0dbf3a0d26381af4eba4a98eaa9b4e6a
    // Bob's private key:   5dab087e624a8a4b79e17f8b83800ee66f3bb1292618b6fd1c2f8b27ff88e0eb
    // Bob's public key:    de9edb7d7b7dc1b4d35b61c2ece435373f8343c85b78674dadfc7e146f882b4f
    // Shared secret (K):   4a5d9d5ba4ce2de1728e3bf480350f25e07e21c947d19e3376f09b3c1e161742
    #[test]
    fn rfc7748_section_6_1_alice_pk() {
        let alice_sk = hex32("77076d0a7318a57d3c16c17251b26645df4c2f87ebc0992ab177fba51db92c2a");
        let expected = hex32("8520f0098930a754748b7ddcb43ef75a0dbf3a0d26381af4eba4a98eaa9b4e6a");
        assert_eq!(x25519_derive_public(&alice_sk), expected);
    }

    #[test]
    fn rfc7748_section_6_1_bob_pk() {
        let bob_sk = hex32("5dab087e624a8a4b79e17f8b83800ee66f3bb1292618b6fd1c2f8b27ff88e0eb");
        let expected = hex32("de9edb7d7b7dc1b4d35b61c2ece435373f8343c85b78674dadfc7e146f882b4f");
        assert_eq!(x25519_derive_public(&bob_sk), expected);
    }

    #[test]
    fn rfc7748_section_6_1_shared_secret() {
        let alice_sk = hex32("77076d0a7318a57d3c16c17251b26645df4c2f87ebc0992ab177fba51db92c2a");
        let bob_sk = hex32("5dab087e624a8a4b79e17f8b83800ee66f3bb1292618b6fd1c2f8b27ff88e0eb");
        let alice_pk = x25519_derive_public(&alice_sk);
        let bob_pk = x25519_derive_public(&bob_sk);
        let expected = hex32("4a5d9d5ba4ce2de1728e3bf480350f25e07e21c947d19e3376f09b3c1e161742");
        assert_eq!(x25519_ecdh(&alice_sk, &bob_pk), expected);
        assert_eq!(x25519_ecdh(&bob_sk, &alice_pk), expected);
    }

    // ----- Roundtrip on an arbitrary key pair -----
    #[test]
    fn x25519_roundtrip_custom_keys() {
        let alice_sk = hex32("0101010101010101010101010101010101010101010101010101010101010101");
        let bob_sk = hex32("0202020202020202020202020202020202020202020202020202020202020202");
        let alice_pk = x25519_derive_public(&alice_sk);
        let bob_pk = x25519_derive_public(&bob_sk);
        let s_ab = x25519_ecdh(&alice_sk, &bob_pk);
        let s_ba = x25519_ecdh(&bob_sk, &alice_pk);
        assert_eq!(s_ab, s_ba);
        // Sanity: the shared secret is non-zero (an all-zero output
        // would signal a low-order input point; our TestRng keys are
        // not in a small subgroup).
        assert!(s_ab.iter().any(|&b| b != 0));
    }

    // ----- Clamping test -----
    //
    // Changing any of the bits that `decode_scalar` forces (low 3
    // bits of byte 0, top 2 bits of byte 31) must not change the
    // output. This exercises the clamping path in isolation.
    #[test]
    fn clamping_is_idempotent() {
        let base = hex32("a546e36bf0527c9d3b16154b82465edd62144c0ac1fc5a18506a2244ba449ac4");
        let u = hex32("e6db6867583030db3594c1a424b15f7c726624ec26b3353b10a903a6d0ab1c4c");
        let ref_out = x25519(&base, &u);

        // Force-dirty the bits that clamping will clear anyway.
        let mut dirty = base;
        dirty[0] |= 0b0000_0111; // set low 3 bits (will be cleared)
        dirty[31] |= 0b1000_0000; // set high bit (will be cleared)
        dirty[31] &= !0b0100_0000; // clear bit 6 (will be set)
        let dirty_out = x25519(&dirty, &u);
        assert_eq!(ref_out, dirty_out);
    }

    // ----- High-bit-of-u decoding test -----
    //
    // decode_u clears the top bit of byte 31 per RFC 7748, so toggling
    // that bit on the input u must not change the output.
    #[test]
    fn u_high_bit_is_ignored() {
        let scalar = hex32("a546e36bf0527c9d3b16154b82465edd62144c0ac1fc5a18506a2244ba449ac4");
        let u = hex32("e6db6867583030db3594c1a424b15f7c726624ec26b3353b10a903a6d0ab1c4c");
        let ref_out = x25519(&scalar, &u);

        let mut u_dirty = u;
        u_dirty[31] |= 0x80;
        let dirty_out = x25519(&scalar, &u_dirty);
        assert_eq!(ref_out, dirty_out);
    }
}