crypto: add GHASH implementation.

Can be used to implement GCM until GCM support in NSS is widespread.

Review URL: https://codereview.chromium.org/11175015

git-svn-id: svn://svn.chromium.org/chrome/trunk/src@166952 0039d316-1c4b-4281-b951-d872f2087c98

4 files changed, 484 insertions, 0 deletions

diff --git a/crypto/crypto.gyp b/crypto/crypto.gyp
index c9313080..c806873 100644
--- a/crypto/crypto.gyp
+++ b/crypto/crypto.gyp
@@ -182,6 +182,8 @@
         'crypto_module_blocking_password_delegate.h',
         'cssm_init.cc',
         'cssm_init.h',
+        'ghash.cc',
+        'ghash.h',
         'ec_private_key.h',
         'ec_private_key_nss.cc',
         'ec_private_key_openssl.cc',
@@ -257,6 +259,7 @@
         'ec_private_key_unittest.cc',
         'ec_signature_creator_unittest.cc',
         'encryptor_unittest.cc',
+        'ghash_unittest.cc',
         'hmac_unittest.cc',
         'nss_util_unittest.cc',
         'p224_unittest.cc',
diff --git a/crypto/ghash.cc b/crypto/ghash.cc
new file mode 100644
index 0000000..939dd0b
--- /dev/null
+++ b/crypto/ghash.cc
@@ -0,0 +1,257 @@
+// Copyright (c) 2012 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "crypto/ghash.h"
+
+#include "base/logging.h"
+#include "base/sys_byteorder.h"
+
+namespace crypto {
+
+// GaloisHash is a polynomial authenticator that works in GF(2^128).
+//
+// Elements of the field are represented in `little-endian' order (which
+// matches the description in the paper[1]), thus the most significant bit is
+// the right-most bit. (This is backwards from the way that everybody else does
+// it.)
+//
+// We store field elements in a pair of such `little-endian' uint64s. So the
+// value one is represented by {low = 2**63, high = 0} and doubling a value
+// involves a *right* shift.
+//
+// [1] http://csrc.nist.gov/groups/ST/toolkit/BCM/documents/proposedmodes/gcm/gcm-revised-spec.pdf
+
+namespace {
+
+// Get64 reads a 64-bit, big-endian number from |bytes|.
+uint64 Get64(const uint8 bytes[8]) {
+  uint64 t;
+  memcpy(&t, bytes, sizeof(t));
+  return base::NetToHost64(t);
+}
+
+// Put64 writes |x| to |bytes| as a 64-bit, big-endian number.
+void Put64(uint8 bytes[8], uint64 x) {
+  x = base::HostToNet64(x);
+  memcpy(bytes, &x, sizeof(x));
+}
+
+// Reverse reverses the order of the bits of 4-bit number in |i|.
+int Reverse(int i) {
+  i = ((i << 2) & 0xc) | ((i >> 2) & 0x3);
+  i = ((i << 1) & 0xa) | ((i >> 1) & 0x5);
+  return i;
+}
+
+}  // namespace
+
+GaloisHash::GaloisHash(const uint8 key[16]) {
+  Reset();
+
+  // We precompute 16 multiples of |key|. However, when we do lookups into this
+  // table we'll be using bits from a field element and therefore the bits will
+  // be in the reverse order. So normally one would expect, say, 4*key to be in
+  // index 4 of the table but due to this bit ordering it will actually be in
+  // index 0010 (base 2) = 2.
+  FieldElement x = {Get64(key), Get64(key+8)};
+  product_table_[0].low = 0;
+  product_table_[0].hi = 0;
+  product_table_[Reverse(1)] = x;
+
+  for (int i = 0; i < 16; i += 2) {
+    product_table_[Reverse(i)] = Double(product_table_[Reverse(i/2)]);
+    product_table_[Reverse(i+1)] = Add(product_table_[Reverse(i)], x);
+  }
+}
+
+void GaloisHash::Reset() {
+  state_ = kHashingAdditionalData;
+  additional_bytes_ = 0;
+  ciphertext_bytes_ = 0;
+  buf_used_ = 0;
+  y_.low = 0;
+  y_.hi = 0;
+}
+
+void GaloisHash::UpdateAdditional(const uint8* data, size_t length) {
+  DCHECK_EQ(state_, kHashingAdditionalData);
+  additional_bytes_ += length;
+  Update(data, length);
+}
+
+void GaloisHash::UpdateCiphertext(const uint8* data, size_t length) {
+  if (state_ == kHashingAdditionalData) {
+    // If there's any remaining additional data it's zero padded to the next
+    // full block.
+    if (buf_used_ > 0) {
+      memset(&buf_[buf_used_], 0, sizeof(buf_)-buf_used_);
+      UpdateBlocks(buf_, 1);
+      buf_used_ = 0;
+    }
+    state_ = kHashingCiphertext;
+  }
+
+  DCHECK_EQ(state_, kHashingCiphertext);
+  ciphertext_bytes_ += length;
+  Update(data, length);
+}
+
+void GaloisHash::Finish(void* output, size_t len) {
+  DCHECK(state_ != kComplete);
+
+  if (buf_used_ > 0) {
+    // If there's any remaining data (additional data or ciphertext), it's zero
+    // padded to the next full block.
+    memset(&buf_[buf_used_], 0, sizeof(buf_)-buf_used_);
+    UpdateBlocks(buf_, 1);
+    buf_used_ = 0;
+  }
+
+  state_ = kComplete;
+
+  // The lengths of the additional data and ciphertext are included as the last
+  // block. The lengths are the number of bits.
+  y_.low ^= additional_bytes_*8;
+  y_.hi ^= ciphertext_bytes_*8;
+  MulAfterPrecomputation(product_table_, &y_);
+
+  uint8 *result, result_tmp[16];
+  if (len >= 16) {
+    result = reinterpret_cast<uint8*>(output);
+  } else {
+    result = result_tmp;
+  }
+
+  Put64(result, y_.low);
+  Put64(result + 8, y_.hi);
+
+  if (len < 16)
+    memcpy(output, result_tmp, len);
+}
+
+// static
+GaloisHash::FieldElement GaloisHash::Add(
+    const FieldElement& x,
+    const FieldElement& y) {
+  // Addition in a characteristic 2 field is just XOR.
+  FieldElement z = {x.low^y.low, x.hi^y.hi};
+  return z;
+}
+
+// static
+GaloisHash::FieldElement GaloisHash::Double(const FieldElement& x) {
+  const bool msb_set = x.hi & 1;
+
+  FieldElement xx;
+  // Because of the bit-ordering, doubling is actually a right shift.
+  xx.hi = x.hi >> 1;
+  xx.hi |= x.low << 63;
+  xx.low = x.low >> 1;
+
+  // If the most-significant bit was set before shifting then it, conceptually,
+  // becomes a term of x^128. This is greater than the irreducible polynomial
+  // so the result has to be reduced. The irreducible polynomial is
+  // 1+x+x^2+x^7+x^128. We can subtract that to eliminate the term at x^128
+  // which also means subtracting the other four terms. In characteristic 2
+  // fields, subtraction == addition == XOR.
+  if (msb_set)
+    xx.low ^= 0xe100000000000000ULL;
+
+  return xx;
+}
+
+void GaloisHash::MulAfterPrecomputation(const FieldElement* table,
+                                        FieldElement* x) {
+  FieldElement z = {0, 0};
+
+  // In order to efficiently multiply, we use the precomputed table of i*key,
+  // for i in 0..15, to handle four bits at a time. We could obviously use
+  // larger tables for greater speedups but the next convenient table size is
+  // 4K, which is a little large.
+  //
+  // In other fields one would use bit positions spread out across the field in
+  // order to reduce the number of doublings required. However, in
+  // characteristic 2 fields, repeated doublings are exceptionally cheap and
+  // it's not worth spending more precomputation time to eliminate them.
+  for (unsigned i = 0; i < 2; i++) {
+    uint64 word;
+    if (i == 0) {
+      word = x->hi;
+    } else {
+      word = x->low;
+    }
+
+    for (unsigned j = 0; j < 64; j += 4) {
+      Mul16(&z);
+      // the values in |table| are ordered for little-endian bit positions. See
+      // the comment in the constructor.
+      const FieldElement& t = table[word & 0xf];
+      z.low ^= t.low;
+      z.hi ^= t.hi;
+      word >>= 4;
+    }
+  }
+
+  *x = z;
+}
+
+// kReductionTable allows for rapid multiplications by 16. A multiplication by
+// 16 is a right shift by four bits, which results in four bits at 2**128.
+// These terms have to be eliminated by dividing by the irreducible polynomial.
+// In GHASH, the polynomial is such that all the terms occur in the
+// least-significant 8 bits, save for the term at x^128. Therefore we can
+// precompute the value to be added to the field element for each of the 16 bit
+// patterns at 2**128 and the values fit within 12 bits.
+static const uint16 kReductionTable[16] = {
+  0x0000, 0x1c20, 0x3840, 0x2460, 0x7080, 0x6ca0, 0x48c0, 0x54e0,
+  0xe100, 0xfd20, 0xd940, 0xc560, 0x9180, 0x8da0, 0xa9c0, 0xb5e0,
+};
+
+// static
+void GaloisHash::Mul16(FieldElement* x) {
+  const unsigned msw = x->hi & 0xf;
+  x->hi >>= 4;
+  x->hi |= x->low << 60;
+  x->low >>= 4;
+  x->low ^= static_cast<uint64>(kReductionTable[msw]) << 48;
+}
+
+void GaloisHash::UpdateBlocks(const uint8* bytes, size_t num_blocks) {
+  for (size_t i = 0; i < num_blocks; i++) {
+    y_.low ^= Get64(bytes);
+    bytes += 8;
+    y_.hi ^= Get64(bytes);
+    bytes += 8;
+    MulAfterPrecomputation(product_table_, &y_);
+  }
+}
+
+void GaloisHash::Update(const uint8* data, size_t length) {
+  if (buf_used_ > 0) {
+    const size_t n = std::min(length, buf_used_);
+    memcpy(&buf_[buf_used_], data, n);
+    buf_used_ += n;
+    length -= n;
+    data += n;
+
+    if (buf_used_ == sizeof(buf_)) {
+      UpdateBlocks(buf_, 1);
+      buf_used_ = 0;
+    }
+  }
+
+  if (length >= 16) {
+    const size_t n = length / 16;
+    UpdateBlocks(data, n);
+    length -= n*16;
+    data += n*16;
+  }
+
+  if (length > 0) {
+    memcpy(buf_, data, length);
+    buf_used_ = length;
+  }
+}
+
+}  // namespace crypto
diff --git a/crypto/ghash.h b/crypto/ghash.h
new file mode 100644
index 0000000..6dc247b
--- /dev/null
+++ b/crypto/ghash.h
@@ -0,0 +1,86 @@
+// Copyright (c) 2012 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "base/basictypes.h"
+#include "crypto/crypto_export.h"
+
+namespace crypto {
+
+// GaloisHash implements the polynomial authenticator part of GCM as specified
+// in http://csrc.nist.gov/groups/ST/toolkit/BCM/documents/proposedmodes/gcm/gcm-revised-spec.pdf
+// Specifically it implements the GHASH function, defined in section 2.3 of
+// that document.
+//
+// In SP-800-38D, GHASH is defined differently and takes only a single data
+// argument. But it is always called with an argument of a certain form:
+//   GHASH_H (A || 0^v || C || 0^u || [len(A)]_64 || [len(C)]_64)
+// This mirrors how the gcm-revised-spec.pdf version of GHASH handles its two
+// data arguments. The two GHASH functions therefore differ only in whether the
+// data is formatted inside or outside of the function.
+//
+// WARNING: do not use this as a generic authenticator. Polynomial
+// authenticators must be used in the correct manner and any use outside of GCM
+// requires careful consideration.
+//
+// WARNING: this code is not constant time. However, in all likelihood, nor is
+// the implementation of AES that is used.
+class CRYPTO_EXPORT_PRIVATE GaloisHash {
+ public:
+  explicit GaloisHash(const uint8 key[16]);
+
+  // Reset prepares to digest a fresh message with the same key. This is more
+  // efficient than creating a fresh object.
+  void Reset();
+
+  // UpdateAdditional hashes in `additional' data. This is data that is not
+  // encrypted, but is covered by the authenticator. All additional data must
+  // be written before any ciphertext is written.
+  void UpdateAdditional(const uint8* data, size_t length);
+
+  // UpdateCiphertext hashes in ciphertext to be authenticated.
+  void UpdateCiphertext(const uint8* data, size_t length);
+
+  // Finish completes the hash computation and writes at most |len| bytes of
+  // the result to |output|.
+  void Finish(void* output, size_t len);
+
+ private:
+  enum State {
+    kHashingAdditionalData,
+    kHashingCiphertext,
+    kComplete,
+  };
+
+  struct FieldElement {
+    uint64 low, hi;
+  };
+
+  // Add returns |x|+|y|.
+  static FieldElement Add(const FieldElement& x, const FieldElement& y);
+  // Double returns 2*|x|.
+  static FieldElement Double(const FieldElement& x);
+  // MulAfterPrecomputation sets |x| = |x|*h where h is |table[1]| and
+  // table[i] = i*h for i=0..15.
+  static void MulAfterPrecomputation(const FieldElement* table,
+                                     FieldElement* x);
+  // Mul16 sets |x| = 16*|x|.
+  static void Mul16(FieldElement* x);
+
+  // UpdateBlocks processes |num_blocks| 16-bytes blocks from |bytes|.
+  void UpdateBlocks(const uint8* bytes, size_t num_blocks);
+  // Update processes |length| bytes from |bytes| and calls UpdateBlocks on as
+  // much data as possible. It uses |buf_| to buffer any remaining data and
+  // always consumes all of |bytes|.
+  void Update(const uint8* bytes, size_t length);
+
+  FieldElement y_;
+  State state_;
+  size_t additional_bytes_;
+  size_t ciphertext_bytes_;
+  uint8 buf_[16];
+  size_t buf_used_;
+  FieldElement product_table_[16];
+};
+
+}  // namespace crypto
diff --git a/crypto/ghash_unittest.cc b/crypto/ghash_unittest.cc
new file mode 100644
index 0000000..c491f76
--- /dev/null
+++ b/crypto/ghash_unittest.cc
@@ -0,0 +1,138 @@
+// Copyright (c) 2012 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "crypto/ghash.h"
+
+#include "testing/gtest/include/gtest/gtest.h"
+
+namespace crypto {
+
+namespace {
+
+// Test vectors are taken from Appendix B of
+// http://csrc.nist.gov/groups/ST/toolkit/BCM/documents/proposedmodes/gcm/gcm-revised-spec.pdf
+
+static const uint8 kKey1[16] = {
+  0x66, 0xe9, 0x4b, 0xd4, 0xef, 0x8a, 0x2c, 0x3b,
+  0x88, 0x4c, 0xfa, 0x59, 0xca, 0x34, 0x2b, 0x2e,
+};
+
+static const uint8 kCiphertext2[] = {
+  0x03, 0x88, 0xda, 0xce, 0x60, 0xb6, 0xa3, 0x92,
+  0xf3, 0x28, 0xc2, 0xb9, 0x71, 0xb2, 0xfe, 0x78,
+};
+
+static const uint8 kKey3[16] = {
+  0xb8, 0x3b, 0x53, 0x37, 0x08, 0xbf, 0x53, 0x5d,
+  0x0a, 0xa6, 0xe5, 0x29, 0x80, 0xd5, 0x3b, 0x78,
+};
+
+static const uint8 kCiphertext3[] = {
+  0x42, 0x83, 0x1e, 0xc2, 0x21, 0x77, 0x74, 0x24,
+  0x4b, 0x72, 0x21, 0xb7, 0x84, 0xd0, 0xd4, 0x9c,
+  0xe3, 0xaa, 0x21, 0x2f, 0x2c, 0x02, 0xa4, 0xe0,
+  0x35, 0xc1, 0x7e, 0x23, 0x29, 0xac, 0xa1, 0x2e,
+  0x21, 0xd5, 0x14, 0xb2, 0x54, 0x66, 0x93, 0x1c,
+  0x7d, 0x8f, 0x6a, 0x5a, 0xac, 0x84, 0xaa, 0x05,
+  0x1b, 0xa3, 0x0b, 0x39, 0x6a, 0x0a, 0xac, 0x97,
+  0x3d, 0x58, 0xe0, 0x91, 0x47, 0x3f, 0x59, 0x85,
+};
+
+static const uint8 kAdditional4[] = {
+  0xfe, 0xed, 0xfa, 0xce, 0xde, 0xad, 0xbe, 0xef,
+  0xfe, 0xed, 0xfa, 0xce, 0xde, 0xad, 0xbe, 0xef,
+  0xab, 0xad, 0xda, 0xd2,
+};
+
+struct TestCase {
+  const uint8* key;
+  const uint8* additional;
+  unsigned additional_length;
+  const uint8* ciphertext;
+  unsigned ciphertext_length;
+  const uint8 expected[16];
+};
+
+static const TestCase kTestCases[] = {
+  {
+    kKey1,
+    NULL,
+    0,
+    NULL,
+    0,
+    {
+      0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+      0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    },
+  },
+  {
+    kKey1,
+    NULL,
+    0,
+    kCiphertext2,
+    sizeof(kCiphertext2),
+    {
+      0xf3, 0x8c, 0xbb, 0x1a, 0xd6, 0x92, 0x23, 0xdc,
+      0xc3, 0x45, 0x7a, 0xe5, 0xb6, 0xb0, 0xf8, 0x85,
+    },
+  },
+  {
+    kKey3,
+    NULL,
+    0,
+    kCiphertext3,
+    sizeof(kCiphertext3),
+    {
+      0x7f, 0x1b, 0x32, 0xb8, 0x1b, 0x82, 0x0d, 0x02,
+      0x61, 0x4f, 0x88, 0x95, 0xac, 0x1d, 0x4e, 0xac,
+    },
+  },
+  {
+    kKey3,
+    kAdditional4,
+    sizeof(kAdditional4),
+    kCiphertext3,
+    sizeof(kCiphertext3) - 4,
+    {
+      0x69, 0x8e, 0x57, 0xf7, 0x0e, 0x6e, 0xcc, 0x7f,
+      0xd9, 0x46, 0x3b, 0x72, 0x60, 0xa9, 0xae, 0x5f,
+    },
+  },
+};
+
+TEST(GaloisHash, TestCases) {
+  uint8 out[16];
+
+  for (size_t i = 0; i < arraysize(kTestCases); ++i) {
+    const TestCase& test = kTestCases[i];
+
+    GaloisHash hash(test.key);
+    if (test.additional_length)
+      hash.UpdateAdditional(test.additional, test.additional_length);
+    if (test.ciphertext_length)
+      hash.UpdateCiphertext(test.ciphertext, test.ciphertext_length);
+    hash.Finish(out, sizeof(out));
+    EXPECT_TRUE(0 == memcmp(out, test.expected, 16));
+  }
+}
+
+TEST(GaloisHash, TestCasesByteAtATime) {
+  uint8 out[16];
+
+  for (size_t i = 0; i < arraysize(kTestCases); ++i) {
+    const TestCase& test = kTestCases[i];
+
+    GaloisHash hash(test.key);
+    for (size_t i = 0; i < test.additional_length; ++i)
+      hash.UpdateAdditional(test.additional + i, 1);
+    for (size_t i = 0; i < test.ciphertext_length; ++i)
+      hash.UpdateCiphertext(test.ciphertext + i, 1);
+    hash.Finish(out, sizeof(out));
+    EXPECT_TRUE(0 == memcmp(out, test.expected, 16));
+  }
+}
+
+}  // namespace
+
+}  // namespace crypto