code wiki / (root) / nx_hkdf_sha384_wasm.nx
`nx_hkdf_sha384_wasm.nx` source

↩ module page · 408 lines · 16648 B
1// nx_hkdf_sha384_wasm.nx -- HKDF-SHA-384 (RFC 5869) for WAT target.
2//
3// Built on inline HMAC-SHA-384 (= SHA-512 compression + SHA-384 IV +
4// 48-byte truncation).  PRK length = 48 (one HMAC-SHA-384 output);
5// HashLen = 48; max OKM = 255 * 48 = 12,240 bytes.
6//
7// Used in the TLS 1.3 key schedule when the ciphersuite is
8// TLS_AES_256_GCM_SHA384.
9//
10// API:
11//   nx_hkdf384_extract(salt, salt_len, ikm, ikm_len, scratch, prk_out)
12//   nx_hkdf384_expand(prk, info, info_len, okm_len, scratch, okm_out)
13//
14// Verified vs RFC-style cross-check against Node OpenSSL HKDF-SHA-384.
15//
16// license_tier: INDEPENDENT_REDERIVE
17// genealogy_id: international-research-sources/ietf/rfc_5869 + rfc_4231
18// lineage_id:   nishi_hkdf_sha384_wasm_q1
19// safe_shift_audit: rotr-style (shift exponent = nn, mask exponent = 64-nn; differs -> safe)
20
21const W_OFF: i64 = 200
22const HMAC_B: i64 = 128
23
24func _rotr64(x: i64, n: i64) -> i64 {
  let nn: i64 = n & 63
  let low_mask: i64 = (1 << (64 - nn)) - 1
  let lo: i64 = (x >> nn) & low_mask
  let hi: i64 = x << (64 - nn)
  return lo | hi
30}
31
32func _be64_read(buf: *u8, off: i64) -> i64 {
  let b0: i64 = buf[off]
  let b1: i64 = buf[off + 1]
  let b2: i64 = buf[off + 2]
  let b3: i64 = buf[off + 3]
  let b4: i64 = buf[off + 4]
  let b5: i64 = buf[off + 5]
  let b6: i64 = buf[off + 6]
  let b7: i64 = buf[off + 7]
  return (b0 << 56) | (b1 << 48) | (b2 << 40) | (b3 << 32) |
         (b4 << 24) | (b5 << 16) | (b6 << 8)  |  b7
43}
44func _be64_write(buf: *u8, off: i64, v: i64) -> i64 {
  buf[off]     = (v >> 56) & 0xFF
  buf[off + 1] = (v >> 48) & 0xFF
  buf[off + 2] = (v >> 40) & 0xFF
  buf[off + 3] = (v >> 32) & 0xFF
  buf[off + 4] = (v >> 24) & 0xFF
  buf[off + 5] = (v >> 16) & 0xFF
  buf[off + 6] = (v >>  8) & 0xFF
  buf[off + 7] =  v        & 0xFF
  return 0
54}
55
56func _i64_le_get(buf: *u8, idx: i64) -> i64 {
  let off: i64 = idx * 8
  return (buf[off] as i64) | ((buf[off + 1] as i64) << 8) |
         ((buf[off + 2] as i64) << 16) | ((buf[off + 3] as i64) << 24) |
         ((buf[off + 4] as i64) << 32) | ((buf[off + 5] as i64) << 40) |
         ((buf[off + 6] as i64) << 48) | ((buf[off + 7] as i64) << 56)
62}
63func _i64_le_set(buf: *u8, idx: i64, v: i64) -> i64 {
  let off: i64 = idx * 8
  buf[off]     = v & 0xFF
  buf[off + 1] = (v >> 8) & 0xFF
  buf[off + 2] = (v >> 16) & 0xFF
  buf[off + 3] = (v >> 24) & 0xFF
  buf[off + 4] = (v >> 32) & 0xFF
  buf[off + 5] = (v >> 40) & 0xFF
  buf[off + 6] = (v >> 48) & 0xFF
  buf[off + 7] = (v >> 56) & 0xFF
  return 0
74}
75func _h_get(ctx: *u8, i: i64) -> i64 {
  let off: i64 = 128 + i * 8
  return (ctx[off] as i64) | ((ctx[off + 1] as i64) << 8) |
         ((ctx[off + 2] as i64) << 16) | ((ctx[off + 3] as i64) << 24) |
         ((ctx[off + 4] as i64) << 32) | ((ctx[off + 5] as i64) << 40) |
         ((ctx[off + 6] as i64) << 48) | ((ctx[off + 7] as i64) << 56)
81}
82func _h_set(ctx: *u8, i: i64, v: i64) -> i64 {
  let off: i64 = 128 + i * 8
  ctx[off]     = v & 0xFF
  ctx[off + 1] = (v >> 8) & 0xFF
  ctx[off + 2] = (v >> 16) & 0xFF
  ctx[off + 3] = (v >> 24) & 0xFF
  ctx[off + 4] = (v >> 32) & 0xFF
  ctx[off + 5] = (v >> 40) & 0xFF
  ctx[off + 6] = (v >> 48) & 0xFF
  ctx[off + 7] = (v >> 56) & 0xFF
  return 0
93}
94
95func _sha512_k(i: i64) -> i64 {
  if i ==  0 { return 0x428a2f98d728ae22 } if i ==  1 { return 0x7137449123ef65cd }
  if i ==  2 { return 0xb5c0fbcfec4d3b2f } if i ==  3 { return 0xe9b5dba58189dbbc }
  if i ==  4 { return 0x3956c25bf348b538 } if i ==  5 { return 0x59f111f1b605d019 }
  if i ==  6 { return 0x923f82a4af194f9b } if i ==  7 { return 0xab1c5ed5da6d8118 }
  if i ==  8 { return 0xd807aa98a3030242 } if i ==  9 { return 0x12835b0145706fbe }
  if i == 10 { return 0x243185be4ee4b28c } if i == 11 { return 0x550c7dc3d5ffb4e2 }
  if i == 12 { return 0x72be5d74f27b896f } if i == 13 { return 0x80deb1fe3b1696b1 }
  if i == 14 { return 0x9bdc06a725c71235 } if i == 15 { return 0xc19bf174cf692694 }
  if i == 16 { return 0xe49b69c19ef14ad2 } if i == 17 { return 0xefbe4786384f25e3 }
  if i == 18 { return 0x0fc19dc68b8cd5b5 } if i == 19 { return 0x240ca1cc77ac9c65 }
  if i == 20 { return 0x2de92c6f592b0275 } if i == 21 { return 0x4a7484aa6ea6e483 }
  if i == 22 { return 0x5cb0a9dcbd41fbd4 } if i == 23 { return 0x76f988da831153b5 }
  if i == 24 { return 0x983e5152ee66dfab } if i == 25 { return 0xa831c66d2db43210 }
  if i == 26 { return 0xb00327c898fb213f } if i == 27 { return 0xbf597fc7beef0ee4 }
  if i == 28 { return 0xc6e00bf33da88fc2 } if i == 29 { return 0xd5a79147930aa725 }
  if i == 30 { return 0x06ca6351e003826f } if i == 31 { return 0x142929670a0e6e70 }
  if i == 32 { return 0x27b70a8546d22ffc } if i == 33 { return 0x2e1b21385c26c926 }
  if i == 34 { return 0x4d2c6dfc5ac42aed } if i == 35 { return 0x53380d139d95b3df }
  if i == 36 { return 0x650a73548baf63de } if i == 37 { return 0x766a0abb3c77b2a8 }
  if i == 38 { return 0x81c2c92e47edaee6 } if i == 39 { return 0x92722c851482353b }
  if i == 40 { return 0xa2bfe8a14cf10364 } if i == 41 { return 0xa81a664bbc423001 }
  if i == 42 { return 0xc24b8b70d0f89791 } if i == 43 { return 0xc76c51a30654be30 }
  if i == 44 { return 0xd192e819d6ef5218 } if i == 45 { return 0xd69906245565a910 }
  if i == 46 { return 0xf40e35855771202a } if i == 47 { return 0x106aa07032bbd1b8 }
  if i == 48 { return 0x19a4c116b8d2d0c8 } if i == 49 { return 0x1e376c085141ab53 }
  if i == 50 { return 0x2748774cdf8eeb99 } if i == 51 { return 0x34b0bcb5e19b48a8 }
  if i == 52 { return 0x391c0cb3c5c95a63 } if i == 53 { return 0x4ed8aa4ae3418acb }
  if i == 54 { return 0x5b9cca4f7763e373 } if i == 55 { return 0x682e6ff3d6b2b8a3 }
  if i == 56 { return 0x748f82ee5defb2fc } if i == 57 { return 0x78a5636f43172f60 }
  if i == 58 { return 0x84c87814a1f0ab72 } if i == 59 { return 0x8cc702081a6439ec }
  if i == 60 { return 0x90befffa23631e28 } if i == 61 { return 0xa4506cebde82bde9 }
  if i == 62 { return 0xbef9a3f7b2c67915 } if i == 63 { return 0xc67178f2e372532b }
  if i == 64 { return 0xca273eceea26619c } if i == 65 { return 0xd186b8c721c0c207 }
  if i == 66 { return 0xeada7dd6cde0eb1e } if i == 67 { return 0xf57d4f7fee6ed178 }
  if i == 68 { return 0x06f067aa72176fba } if i == 69 { return 0x0a637dc5a2c898a6 }
  if i == 70 { return 0x113f9804bef90dae } if i == 71 { return 0x1b710b35131c471b }
  if i == 72 { return 0x28db77f523047d84 } if i == 73 { return 0x32caab7b40c72493 }
  if i == 74 { return 0x3c9ebe0a15c9bebc } if i == 75 { return 0x431d67c49c100d4c }
  if i == 76 { return 0x4cc5d4becb3e42b6 } if i == 77 { return 0x597f299cfc657e2a }
  if i == 78 { return 0x5fcb6fab3ad6faec }
  return 0x6c44198c4a475817
137}
138
139func _sha512_compress(ctx: *u8) -> i64 {
  var i: i64 = 0
  while i < 16 {
      let w: i64 = _be64_read(ctx, i * 8)
      _i64_le_set(ctx, (W_OFF / 8) + i, w)
      i = i + 1
  }
  i = 16
  while i < 80 {
      let w2: i64 = _i64_le_get(ctx, (W_OFF / 8) + i - 2)
      let w15: i64 = _i64_le_get(ctx, (W_OFF / 8) + i - 15)
      let w7: i64 = _i64_le_get(ctx, (W_OFF / 8) + i - 7)
      let w16: i64 = _i64_le_get(ctx, (W_OFF / 8) + i - 16)
      let w2_shr6:  i64 = (w2 >> 6)  & 0x03ffffffffffffff
      let w15_shr7: i64 = (w15 >> 7) & 0x01ffffffffffffff
      let s0: i64 = _rotr64(w15, 1) ^ _rotr64(w15, 8) ^ w15_shr7
      let s1: i64 = _rotr64(w2, 19) ^ _rotr64(w2, 61) ^ w2_shr6
      let wi: i64 = w16 + s0 + w7 + s1
      _i64_le_set(ctx, (W_OFF / 8) + i, wi)
      i = i + 1
  }
  var a: i64 = _h_get(ctx, 0)
  var b: i64 = _h_get(ctx, 1)
  var c: i64 = _h_get(ctx, 2)
  var d: i64 = _h_get(ctx, 3)
  var e: i64 = _h_get(ctx, 4)
  var f: i64 = _h_get(ctx, 5)
  var g: i64 = _h_get(ctx, 6)
  var h: i64 = _h_get(ctx, 7)
  i = 0
  while i < 80 {
      let w: i64 = _i64_le_get(ctx, (W_OFF / 8) + i)
      let s1: i64 = _rotr64(e, 14) ^ _rotr64(e, 18) ^ _rotr64(e, 41)
      let ch: i64 = (e & f) ^ ((~e) & g)
      let temp1: i64 = h + s1 + ch + _sha512_k(i) + w
      let s0: i64 = _rotr64(a, 28) ^ _rotr64(a, 34) ^ _rotr64(a, 39)
      let maj: i64 = (a & b) ^ (a & c) ^ (b & c)
      let temp2: i64 = s0 + maj
      h = g; g = f; f = e
      e = d + temp1
      d = c; c = b; b = a
      a = temp1 + temp2
      i = i + 1
  }
  _h_set(ctx, 0, _h_get(ctx, 0) + a)
  _h_set(ctx, 1, _h_get(ctx, 1) + b)
  _h_set(ctx, 2, _h_get(ctx, 2) + c)
  _h_set(ctx, 3, _h_get(ctx, 3) + d)
  _h_set(ctx, 4, _h_get(ctx, 4) + e)
  _h_set(ctx, 5, _h_get(ctx, 5) + f)
  _h_set(ctx, 6, _h_get(ctx, 6) + g)
  _h_set(ctx, 7, _h_get(ctx, 7) + h)
  return 0
192}
193
194func _sha512_init(ctx: *u8) -> i64 {
  // SHA-384 IV per FIPS 180-4 §5.3.4 (drives this whole module into
  // SHA-384 mode; the SHA-512 compression+pad logic below is unchanged).
  _h_set(ctx, 0, 0xcbbb9d5dc1059ed8)
  _h_set(ctx, 1, 0x629a292a367cd507)
  _h_set(ctx, 2, 0x9159015a3070dd17)
  _h_set(ctx, 3, 0x152fecd8f70e5939)
  _h_set(ctx, 4, 0x67332667ffc00b31)
  _h_set(ctx, 5, 0x8eb44a8768581511)
  _h_set(ctx, 6, 0xdb0c2e0d64f98fa7)
  _h_set(ctx, 7, 0x47b5481dbefa4fa4)
  return 0
206}
207
208func _sha512_one_shot(in_ptr: *u8, in_len: i64, ctx: *u8, out_ptr: *u8) -> i64 {
  _sha512_init(ctx)
  var idx: i64 = 0
  var bit_len: i64 = 0
  var i: i64 = 0
  while i < in_len {
      ctx[idx] = in_ptr[i]
      idx = idx + 1
      bit_len = bit_len + 8
      if idx == 128 {
          _sha512_compress(ctx)
          idx = 0
      }
      i = i + 1
  }
  ctx[idx] = 0x80
  idx = idx + 1
  if idx > 112 {
      while idx < 128 { ctx[idx] = 0; idx = idx + 1 }
      _sha512_compress(ctx)
      idx = 0
  }
  while idx < 112 { ctx[idx] = 0; idx = idx + 1 }
  ctx[112] = 0; ctx[113] = 0; ctx[114] = 0; ctx[115] = 0
  ctx[116] = 0; ctx[117] = 0; ctx[118] = 0; ctx[119] = 0
  ctx[120] = (bit_len >> 56) & 0xFF
  ctx[121] = (bit_len >> 48) & 0xFF
  ctx[122] = (bit_len >> 40) & 0xFF
  ctx[123] = (bit_len >> 32) & 0xFF
  ctx[124] = (bit_len >> 24) & 0xFF
  ctx[125] = (bit_len >> 16) & 0xFF
  ctx[126] = (bit_len >>  8) & 0xFF
  ctx[127] =  bit_len        & 0xFF
  _sha512_compress(ctx)
  // SHA-384 truncates state to first 6 H-words = 48 bytes.
  var k: i64 = 0
  while k < 6 {
      _be64_write(out_ptr, k * 8, _h_get(ctx, k))
      k = k + 1
  }
  return 0
249}
250
251// ===== HMAC-SHA-512 =====
252//
253// scratch layout (>= 2048 bytes):
254//   0..127     : K' (block-sized key buffer)
255//   128..255   : K' XOR ipad
256//   256..383   : K' XOR opad
257//   384..447   : inner-hash output (64 bytes)
258//   512..1431  : SHA-512 ctx scratch (920 bytes; needs 128 + 64 + 8 + 640 = 840)
259//   1536..1727 : opad_block || inner_hash concat for outer (192 bytes)
260func _hmac_sha512(key_ptr: *u8, key_len: i64,
                 msg_ptr: *u8, msg_len: i64,
                 scratch_ptr: *u8, out_ptr: *u8) -> i64 {
  let kprime: *u8 = scratch_ptr
  var j: i64 = 0
  while j < HMAC_B { kprime[j] = 0; j = j + 1 }
  if key_len > HMAC_B {
      let h_ctx: *u8 = (scratch_ptr as i64 + 512) as *u8
      _sha512_one_shot(key_ptr, key_len, h_ctx, kprime)
  } else {
      var ki: i64 = 0
      while ki < key_len { kprime[ki] = key_ptr[ki]; ki = ki + 1 }
  }
  let ipad_blk: *u8 = (scratch_ptr as i64 + 128) as *u8
  var ii: i64 = 0
  while ii < HMAC_B { ipad_blk[ii] = kprime[ii] ^ 0x36; ii = ii + 1 }
  let opad_blk: *u8 = (scratch_ptr as i64 + 256) as *u8
  var oi: i64 = 0
  while oi < HMAC_B { opad_blk[oi] = kprime[oi] ^ 0x5c; oi = oi + 1 }
  // Inner hash: streaming SHA-512 of (ipad || msg).
  let inner_ctx: *u8 = (scratch_ptr as i64 + 512) as *u8
  _sha512_init(inner_ctx)
  var inner_idx: i64 = 0
  var inner_bits: i64 = 0
  var ip: i64 = 0
  while ip < HMAC_B {
      inner_ctx[inner_idx] = ipad_blk[ip]
      inner_idx = inner_idx + 1
      inner_bits = inner_bits + 8
      if inner_idx == 128 { _sha512_compress(inner_ctx); inner_idx = 0 }
      ip = ip + 1
  }
  var mp: i64 = 0
  while mp < msg_len {
      inner_ctx[inner_idx] = msg_ptr[mp]
      inner_idx = inner_idx + 1
      inner_bits = inner_bits + 8
      if inner_idx == 128 { _sha512_compress(inner_ctx); inner_idx = 0 }
      mp = mp + 1
  }
  inner_ctx[inner_idx] = 0x80
  inner_idx = inner_idx + 1
  if inner_idx > 112 {
      while inner_idx < 128 { inner_ctx[inner_idx] = 0; inner_idx = inner_idx + 1 }
      _sha512_compress(inner_ctx)
      inner_idx = 0
  }
  while inner_idx < 112 { inner_ctx[inner_idx] = 0; inner_idx = inner_idx + 1 }
  inner_ctx[112] = 0; inner_ctx[113] = 0; inner_ctx[114] = 0; inner_ctx[115] = 0
  inner_ctx[116] = 0; inner_ctx[117] = 0; inner_ctx[118] = 0; inner_ctx[119] = 0
  inner_ctx[120] = (inner_bits >> 56) & 0xFF
  inner_ctx[121] = (inner_bits >> 48) & 0xFF
  inner_ctx[122] = (inner_bits >> 40) & 0xFF
  inner_ctx[123] = (inner_bits >> 32) & 0xFF
  inner_ctx[124] = (inner_bits >> 24) & 0xFF
  inner_ctx[125] = (inner_bits >> 16) & 0xFF
  inner_ctx[126] = (inner_bits >>  8) & 0xFF
  inner_ctx[127] =  inner_bits        & 0xFF
  _sha512_compress(inner_ctx)
  // SHA-384 truncates inner state to first 6 H-words = 48 bytes.
  let inner_out: *u8 = (scratch_ptr as i64 + 384) as *u8
  var dk: i64 = 0
  while dk < 6 {
      _be64_write(inner_out, dk * 8, _h_get(inner_ctx, dk))
      dk = dk + 1
  }
  // Outer hash: SHA-384(opad_blk || inner_out).  _sha512_one_shot already
  // truncates to 48 bytes (via the same `while k < 6` change above), so
  // out_ptr receives exactly 48 HMAC-SHA-384 bytes.
  let outer_in: *u8 = (scratch_ptr as i64 + 1536) as *u8
  var op: i64 = 0
  while op < HMAC_B { outer_in[op] = opad_blk[op]; op = op + 1 }
  var ohi: i64 = 0
  while ohi < 48 { outer_in[HMAC_B + ohi] = inner_out[ohi]; ohi = ohi + 1 }
  _sha512_one_shot(outer_in, HMAC_B + 48, inner_ctx, out_ptr)
  return 0
336}
337
338// ===== HKDF proper =====
339
340// HKDF-Extract(salt, IKM) = HMAC-SHA-384(salt, IKM) -> 48-byte PRK.
341// If salt_len == 0, use 48 zero bytes as salt per RFC 5869 §2.2.
342//
343// scratch_ptr must be >=2048 bytes (HMAC needs that much).  PRK out
344// destination is 48 bytes.
345func nx_hkdf384_extract(salt_ptr: *u8, salt_len: i64,
                       ikm_ptr: *u8, ikm_len: i64,
                       scratch_ptr: *u8, prk_out: *u8) -> i64 {
  if salt_len == 0 {
      let zero_salt: *u8 = (scratch_ptr as i64 + 1856) as *u8
      var z: i64 = 0
      while z < 48 { zero_salt[z] = 0; z = z + 1 }
      _hmac_sha512(zero_salt, 48, ikm_ptr, ikm_len, scratch_ptr, prk_out)
  } else {
      _hmac_sha512(salt_ptr, salt_len, ikm_ptr, ikm_len, scratch_ptr, prk_out)
  }
  return 0
357}
358
359// HKDF-Expand(PRK, info, L) -> first L bytes of T(1) || T(2) || ...
360//   T(i) = HMAC-SHA-384(PRK, T(i-1) || info || i_byte)
361//   T(0) = empty
362// L MUST be <= 255 * 48 = 12240 bytes.
363//
364// scratch_ptr layout (>=4096 bytes):
365//   0..2047    : HMAC scratch (used by _hmac_sha512 internally; SHA-384 IV)
366//   2048..2095 : T(i-1) buffer (48 B)
367//   2112..2159 : T(i) output buffer (48 B)
368//   2176..3199 : HMAC input buffer (T(i-1) || info || counter; <= 48 + 950 + 1)
369func nx_hkdf384_expand(prk_ptr: *u8, info_ptr: *u8, info_len: i64,
                      okm_len: i64, scratch_ptr: *u8, okm_out: *u8) -> i64 {
  if okm_len > 12240 { return -1 }
  let hmac_scratch: *u8 = scratch_ptr
  let t_prev: *u8 = (scratch_ptr as i64 + 2048) as *u8
  let t_cur:  *u8 = (scratch_ptr as i64 + 2112) as *u8
  let hmac_in: *u8 = (scratch_ptr as i64 + 2176) as *u8
376
  let n: i64 = (okm_len + 47) / 48       // number of T blocks (HashLen=48)
  var written: i64 = 0
  var i: i64 = 1
  var prev_len: i64 = 0
  while i <= n {
      var p: i64 = 0
      var pp: i64 = 0
      while pp < prev_len { hmac_in[p] = t_prev[pp]; p = p + 1; pp = pp + 1 }
      var ii: i64 = 0
      while ii < info_len { hmac_in[p] = info_ptr[ii]; p = p + 1; ii = ii + 1 }
      hmac_in[p] = i & 0xFF
      p = p + 1
389
      _hmac_sha512(prk_ptr, 48, hmac_in, p, hmac_scratch, t_cur)
391
      let need: i64 = okm_len - written
      var take: i64 = 48
      if need < 48 { take = need }
      var ki: i64 = 0
      while ki < take {
          okm_out[written + ki] = t_cur[ki]
          ki = ki + 1
      }
      written = written + take
401
      var ci: i64 = 0
      while ci < 48 { t_prev[ci] = t_cur[ci]; ci = ci + 1 }
      prev_len = 48
      i = i + 1
  }
  return 0
408}
nx_hkdf_sha384_wasm.nx source

`nx_hkdf_sha384_wasm.nx` source
