/* Copyright (c) 2017, Google Inc.
 *
 * Permission to use, copy, modify, and/or distribute this software for any
 * purpose with or without fee is hereby granted, provided that the above
 * copyright notice and this permission notice appear in all copies.
 *
 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY
 * SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION
 * OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN
 * CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE. */

#include <openssl/aead.h>

#include <assert.h>

#include <openssl/cipher.h>
#include <openssl/cpu.h>
#include <openssl/crypto.h>
#include <openssl/err.h>

#include "../fipsmodule/cipher/internal.h"


#define EVP_AEAD_AES_GCM_SIV_NONCE_LEN 12
#define EVP_AEAD_AES_GCM_SIV_TAG_LEN 16

// TODO(davidben): AES-GCM-SIV assembly is not correct for Windows. It must save
// and restore xmm6 through xmm15.
#if defined(OPENSSL_X86_64) && !defined(OPENSSL_NO_ASM) && \
    !defined(OPENSSL_WINDOWS)
#define AES_GCM_SIV_ASM

// Optimised AES-GCM-SIV

struct aead_aes_gcm_siv_asm_ctx {
  alignas(16) uint8_t key[16*15];
  int is_128_bit;
};

// The assembly code assumes 8-byte alignment of the EVP_AEAD_CTX's state, and
// aligns to 16 bytes itself.
OPENSSL_STATIC_ASSERT(sizeof(((EVP_AEAD_CTX *)NULL)->state) + 8 >=
                          sizeof(struct aead_aes_gcm_siv_asm_ctx),
                      "AEAD state is too small");
#if defined(__GNUC__) || defined(__clang__)
OPENSSL_STATIC_ASSERT(alignof(union evp_aead_ctx_st_state) >= 8,
                      "AEAD state has insufficient alignment");
#endif

// asm_ctx_from_ctx returns a 16-byte aligned context pointer from |ctx|.
static struct aead_aes_gcm_siv_asm_ctx *asm_ctx_from_ctx(
    const EVP_AEAD_CTX *ctx) {
  // ctx->state must already be 8-byte aligned. Thus, at most, we may need to
  // add eight to align it to 16 bytes.
  const uintptr_t offset = ((uintptr_t)&ctx->state) & 8;
  return (struct aead_aes_gcm_siv_asm_ctx *)(&ctx->state.opaque[offset]);
}

// aes128gcmsiv_aes_ks writes an AES-128 key schedule for |key| to
// |out_expanded_key|.
extern void aes128gcmsiv_aes_ks(
    const uint8_t key[16], uint8_t out_expanded_key[16*15]);

// aes256gcmsiv_aes_ks writes an AES-256 key schedule for |key| to
// |out_expanded_key|.
extern void aes256gcmsiv_aes_ks(
    const uint8_t key[32], uint8_t out_expanded_key[16*15]);

static int aead_aes_gcm_siv_asm_init(EVP_AEAD_CTX *ctx, const uint8_t *key,
                                     size_t key_len, size_t tag_len) {
  const size_t key_bits = key_len * 8;

  if (key_bits != 128 && key_bits != 256) {
    OPENSSL_PUT_ERROR(CIPHER, CIPHER_R_BAD_KEY_LENGTH);
    return 0;  // EVP_AEAD_CTX_init should catch this.
  }

  if (tag_len == EVP_AEAD_DEFAULT_TAG_LENGTH) {
    tag_len = EVP_AEAD_AES_GCM_SIV_TAG_LEN;
  }

  if (tag_len != EVP_AEAD_AES_GCM_SIV_TAG_LEN) {
    OPENSSL_PUT_ERROR(CIPHER, CIPHER_R_TAG_TOO_LARGE);
    return 0;
  }

  struct aead_aes_gcm_siv_asm_ctx *gcm_siv_ctx = asm_ctx_from_ctx(ctx);
  assert((((uintptr_t)gcm_siv_ctx) & 15) == 0);

  if (key_bits == 128) {
    aes128gcmsiv_aes_ks(key, &gcm_siv_ctx->key[0]);
    gcm_siv_ctx->is_128_bit = 1;
  } else {
    aes256gcmsiv_aes_ks(key, &gcm_siv_ctx->key[0]);
    gcm_siv_ctx->is_128_bit = 0;
  }

  ctx->tag_len = tag_len;

  return 1;
}

static void aead_aes_gcm_siv_asm_cleanup(EVP_AEAD_CTX *ctx) {}

// aesgcmsiv_polyval_horner updates the POLYVAL value in |in_out_poly| to
// include a number (|in_blocks|) of 16-byte blocks of data from |in|, given
// the POLYVAL key in |key|.
extern void aesgcmsiv_polyval_horner(const uint8_t in_out_poly[16],
                                     const uint8_t key[16], const uint8_t *in,
                                     size_t in_blocks);

// aesgcmsiv_htable_init writes powers 1..8 of |auth_key| to |out_htable|.
extern void aesgcmsiv_htable_init(uint8_t out_htable[16 * 8],
                                  const uint8_t auth_key[16]);

// aesgcmsiv_htable6_init writes powers 1..6 of |auth_key| to |out_htable|.
extern void aesgcmsiv_htable6_init(uint8_t out_htable[16 * 6],
                                   const uint8_t auth_key[16]);

// aesgcmsiv_htable_polyval updates the POLYVAL value in |in_out_poly| to
// include |in_len| bytes of data from |in|. (Where |in_len| must be a multiple
// of 16.) It uses the precomputed powers of the key given in |htable|.
extern void aesgcmsiv_htable_polyval(const uint8_t htable[16 * 8],
                                     const uint8_t *in, size_t in_len,
                                     uint8_t in_out_poly[16]);

// aes128gcmsiv_dec decrypts |in_len| & ~15 bytes from |out| and writes them to
// |in|. (The full value of |in_len| is still used to find the authentication
// tag appended to the ciphertext, however, so must not be pre-masked.)
//
// |in| and |out| may be equal, but must not otherwise overlap.
//
// While decrypting, it updates the POLYVAL value found at the beginning of
// |in_out_calculated_tag_and_scratch| and writes the updated value back before
// return. During executation, it may use the whole of this space for other
// purposes. In order to decrypt and update the POLYVAL value, it uses the
// expanded key from |key| and the table of powers in |htable|.
extern void aes128gcmsiv_dec(const uint8_t *in, uint8_t *out,
                             uint8_t in_out_calculated_tag_and_scratch[16 * 8],
                             const uint8_t htable[16 * 6],
                             const struct aead_aes_gcm_siv_asm_ctx *key,
                             size_t in_len);

// aes256gcmsiv_dec acts like |aes128gcmsiv_dec|, but for AES-256.
extern void aes256gcmsiv_dec(const uint8_t *in, uint8_t *out,
                             uint8_t in_out_calculated_tag_and_scratch[16 * 8],
                             const uint8_t htable[16 * 6],
                             const struct aead_aes_gcm_siv_asm_ctx *key,
                             size_t in_len);

// aes128gcmsiv_kdf performs the AES-GCM-SIV KDF given the expanded key from
// |key_schedule| and the nonce in |nonce|. Note that, while only 12 bytes of
// the nonce are used, 16 bytes are read and so the value must be
// right-padded.
extern void aes128gcmsiv_kdf(const uint8_t nonce[16],
                             uint64_t out_key_material[8],
                             const uint8_t *key_schedule);

// aes256gcmsiv_kdf acts like |aes128gcmsiv_kdf|, but for AES-256.
extern void aes256gcmsiv_kdf(const uint8_t nonce[16],
                             uint64_t out_key_material[12],
                             const uint8_t *key_schedule);

// aes128gcmsiv_aes_ks_enc_x1 performs a key expansion of the AES-128 key in
// |key|, writes the expanded key to |out_expanded_key| and encrypts a single
// block from |in| to |out|.
extern void aes128gcmsiv_aes_ks_enc_x1(const uint8_t in[16], uint8_t out[16],
                                       uint8_t out_expanded_key[16 * 15],
                                       const uint64_t key[2]);

// aes256gcmsiv_aes_ks_enc_x1 acts like |aes128gcmsiv_aes_ks_enc_x1|, but for
// AES-256.
extern void aes256gcmsiv_aes_ks_enc_x1(const uint8_t in[16], uint8_t out[16],
                                       uint8_t out_expanded_key[16 * 15],
                                       const uint64_t key[4]);

// aes128gcmsiv_ecb_enc_block encrypts a single block from |in| to |out| using
// the expanded key in |expanded_key|.
extern void aes128gcmsiv_ecb_enc_block(
    const uint8_t in[16], uint8_t out[16],
    const struct aead_aes_gcm_siv_asm_ctx *expanded_key);

// aes256gcmsiv_ecb_enc_block acts like |aes128gcmsiv_ecb_enc_block|, but for
// AES-256.
extern void aes256gcmsiv_ecb_enc_block(
    const uint8_t in[16], uint8_t out[16],
    const struct aead_aes_gcm_siv_asm_ctx *expanded_key);

// aes128gcmsiv_enc_msg_x4 encrypts |in_len| bytes from |in| to |out| using the
// expanded key from |key|. (The value of |in_len| must be a multiple of 16.)
// The |in| and |out| buffers may be equal but must not otherwise overlap. The
// initial counter is constructed from the given |tag| as required by
// AES-GCM-SIV.
extern void aes128gcmsiv_enc_msg_x4(const uint8_t *in, uint8_t *out,
                                    const uint8_t *tag,
                                    const struct aead_aes_gcm_siv_asm_ctx *key,
                                    size_t in_len);

// aes256gcmsiv_enc_msg_x4 acts like |aes128gcmsiv_enc_msg_x4|, but for
// AES-256.
extern void aes256gcmsiv_enc_msg_x4(const uint8_t *in, uint8_t *out,
                                    const uint8_t *tag,
                                    const struct aead_aes_gcm_siv_asm_ctx *key,
                                    size_t in_len);

// aes128gcmsiv_enc_msg_x8 acts like |aes128gcmsiv_enc_msg_x4|, but is
// optimised for longer messages.
extern void aes128gcmsiv_enc_msg_x8(const uint8_t *in, uint8_t *out,
                                    const uint8_t *tag,
                                    const struct aead_aes_gcm_siv_asm_ctx *key,
                                    size_t in_len);

// aes256gcmsiv_enc_msg_x8 acts like |aes256gcmsiv_enc_msg_x4|, but is
// optimised for longer messages.
extern void aes256gcmsiv_enc_msg_x8(const uint8_t *in, uint8_t *out,
                                    const uint8_t *tag,
                                    const struct aead_aes_gcm_siv_asm_ctx *key,
                                    size_t in_len);

// gcm_siv_asm_polyval evaluates POLYVAL at |auth_key| on the given plaintext
// and AD. The result is written to |out_tag|.
static void gcm_siv_asm_polyval(uint8_t out_tag[16], const uint8_t *in,
                                size_t in_len, const uint8_t *ad, size_t ad_len,
                                const uint8_t auth_key[16],
                                const uint8_t nonce[12]) {
  OPENSSL_memset(out_tag, 0, 16);
  const size_t ad_blocks = ad_len / 16;
  const size_t in_blocks = in_len / 16;
  int htable_init = 0;
  alignas(16) uint8_t htable[16*8];

  if (ad_blocks > 8 || in_blocks > 8) {
    htable_init = 1;
    aesgcmsiv_htable_init(htable, auth_key);
  }

  if (htable_init) {
    aesgcmsiv_htable_polyval(htable, ad, ad_len & ~15, out_tag);
  } else {
    aesgcmsiv_polyval_horner(out_tag, auth_key, ad, ad_blocks);
  }

  uint8_t scratch[16];
  if (ad_len & 15) {
    OPENSSL_memset(scratch, 0, sizeof(scratch));
    OPENSSL_memcpy(scratch, &ad[ad_len & ~15], ad_len & 15);
    aesgcmsiv_polyval_horner(out_tag, auth_key, scratch, 1);
  }

  if (htable_init) {
    aesgcmsiv_htable_polyval(htable, in, in_len & ~15, out_tag);
  } else {
    aesgcmsiv_polyval_horner(out_tag, auth_key, in, in_blocks);
  }

  if (in_len & 15) {
    OPENSSL_memset(scratch, 0, sizeof(scratch));
    OPENSSL_memcpy(scratch, &in[in_len & ~15], in_len & 15);
    aesgcmsiv_polyval_horner(out_tag, auth_key, scratch, 1);
  }

  union {
    uint8_t c[16];
    struct {
      uint64_t ad;
      uint64_t in;
    } bitlens;
  } length_block;

  length_block.bitlens.ad = ad_len * 8;
  length_block.bitlens.in = in_len * 8;
  aesgcmsiv_polyval_horner(out_tag, auth_key, length_block.c, 1);

  for (size_t i = 0; i < 12; i++) {
    out_tag[i] ^= nonce[i];
  }

  out_tag[15] &= 0x7f;
}

// aead_aes_gcm_siv_asm_crypt_last_block handles the encryption/decryption
// (same thing in CTR mode) of the final block of a plaintext/ciphertext. It
// writes |in_len| & 15 bytes to |out| + |in_len|, based on an initial counter
// derived from |tag|.
static void aead_aes_gcm_siv_asm_crypt_last_block(
    int is_128_bit, uint8_t *out, const uint8_t *in, size_t in_len,
    const uint8_t tag[16],
    const struct aead_aes_gcm_siv_asm_ctx *enc_key_expanded) {
  alignas(16) union {
    uint8_t c[16];
    uint32_t u32[4];
  } counter;
  OPENSSL_memcpy(&counter, tag, sizeof(counter));
  counter.c[15] |= 0x80;
  counter.u32[0] += in_len / 16;

  if (is_128_bit) {
    aes128gcmsiv_ecb_enc_block(&counter.c[0], &counter.c[0], enc_key_expanded);
  } else {
    aes256gcmsiv_ecb_enc_block(&counter.c[0], &counter.c[0], enc_key_expanded);
  }

  const size_t last_bytes_offset = in_len & ~15;
  const size_t last_bytes_len = in_len & 15;
  uint8_t *last_bytes_out = &out[last_bytes_offset];
  const uint8_t *last_bytes_in = &in[last_bytes_offset];
  for (size_t i = 0; i < last_bytes_len; i++) {
    last_bytes_out[i] = last_bytes_in[i] ^ counter.c[i];
  }
}

// aead_aes_gcm_siv_kdf calculates the record encryption and authentication
// keys given the |nonce|.
static void aead_aes_gcm_siv_kdf(
    int is_128_bit, const struct aead_aes_gcm_siv_asm_ctx *gcm_siv_ctx,
    uint64_t out_record_auth_key[2], uint64_t out_record_enc_key[4],
    const uint8_t nonce[12]) {
  alignas(16) uint8_t padded_nonce[16];
  OPENSSL_memcpy(padded_nonce, nonce, 12);

  alignas(16) uint64_t key_material[12];
  if (is_128_bit) {
    aes128gcmsiv_kdf(padded_nonce, key_material, &gcm_siv_ctx->key[0]);
    out_record_enc_key[0] = key_material[4];
    out_record_enc_key[1] = key_material[6];
  } else {
    aes256gcmsiv_kdf(padded_nonce, key_material, &gcm_siv_ctx->key[0]);
    out_record_enc_key[0] = key_material[4];
    out_record_enc_key[1] = key_material[6];
    out_record_enc_key[2] = key_material[8];
    out_record_enc_key[3] = key_material[10];
  }

  out_record_auth_key[0] = key_material[0];
  out_record_auth_key[1] = key_material[2];
}

static int aead_aes_gcm_siv_asm_seal_scatter(
    const EVP_AEAD_CTX *ctx, uint8_t *out, uint8_t *out_tag,
    size_t *out_tag_len, size_t max_out_tag_len, const uint8_t *nonce,
    size_t nonce_len, const uint8_t *in, size_t in_len, const uint8_t *extra_in,
    size_t extra_in_len, const uint8_t *ad, size_t ad_len) {
  const struct aead_aes_gcm_siv_asm_ctx *gcm_siv_ctx = asm_ctx_from_ctx(ctx);
  const uint64_t in_len_64 = in_len;
  const uint64_t ad_len_64 = ad_len;

  if (in_len_64 > (UINT64_C(1) << 36) ||
      ad_len_64 >= (UINT64_C(1) << 61)) {
    OPENSSL_PUT_ERROR(CIPHER, CIPHER_R_TOO_LARGE);
    return 0;
  }

  if (max_out_tag_len < EVP_AEAD_AES_GCM_SIV_TAG_LEN) {
    OPENSSL_PUT_ERROR(CIPHER, CIPHER_R_BUFFER_TOO_SMALL);
    return 0;
  }

  if (nonce_len != EVP_AEAD_AES_GCM_SIV_NONCE_LEN) {
    OPENSSL_PUT_ERROR(CIPHER, CIPHER_R_UNSUPPORTED_NONCE_SIZE);
    return 0;
  }

  alignas(16) uint64_t record_auth_key[2];
  alignas(16) uint64_t record_enc_key[4];
  aead_aes_gcm_siv_kdf(gcm_siv_ctx->is_128_bit, gcm_siv_ctx, record_auth_key,
                       record_enc_key, nonce);

  alignas(16) uint8_t tag[16] = {0};
  gcm_siv_asm_polyval(tag, in, in_len, ad, ad_len,
                      (const uint8_t *)record_auth_key, nonce);

  struct aead_aes_gcm_siv_asm_ctx enc_key_expanded;

  if (gcm_siv_ctx->is_128_bit) {
    aes128gcmsiv_aes_ks_enc_x1(tag, tag, &enc_key_expanded.key[0],
                               record_enc_key);

    if (in_len < 128) {
      aes128gcmsiv_enc_msg_x4(in, out, tag, &enc_key_expanded, in_len & ~15);
    } else {
      aes128gcmsiv_enc_msg_x8(in, out, tag, &enc_key_expanded, in_len & ~15);
    }
  } else {
    aes256gcmsiv_aes_ks_enc_x1(tag, tag, &enc_key_expanded.key[0],
                               record_enc_key);

    if (in_len < 128) {
      aes256gcmsiv_enc_msg_x4(in, out, tag, &enc_key_expanded, in_len & ~15);
    } else {
      aes256gcmsiv_enc_msg_x8(in, out, tag, &enc_key_expanded, in_len & ~15);
    }
  }

  if (in_len & 15) {
    aead_aes_gcm_siv_asm_crypt_last_block(gcm_siv_ctx->is_128_bit, out, in,
                                          in_len, tag, &enc_key_expanded);
  }

  OPENSSL_memcpy(out_tag, tag, sizeof(tag));
  *out_tag_len = EVP_AEAD_AES_GCM_SIV_TAG_LEN;

  return 1;
}

// TODO(martinkr): Add aead_aes_gcm_siv_asm_open_gather. N.B. aes128gcmsiv_dec
// expects ciphertext and tag in a contiguous buffer.

static int aead_aes_gcm_siv_asm_open(const EVP_AEAD_CTX *ctx, uint8_t *out,
                                     size_t *out_len, size_t max_out_len,
                                     const uint8_t *nonce, size_t nonce_len,
                                     const uint8_t *in, size_t in_len,
                                     const uint8_t *ad, size_t ad_len) {
  const uint64_t ad_len_64 = ad_len;
  if (ad_len_64 >= (UINT64_C(1) << 61)) {
    OPENSSL_PUT_ERROR(CIPHER, CIPHER_R_TOO_LARGE);
    return 0;
  }

  const uint64_t in_len_64 = in_len;
  if (in_len < EVP_AEAD_AES_GCM_SIV_TAG_LEN ||
      in_len_64 > (UINT64_C(1) << 36) + AES_BLOCK_SIZE) {
    OPENSSL_PUT_ERROR(CIPHER, CIPHER_R_BAD_DECRYPT);
    return 0;
  }

  if (nonce_len != EVP_AEAD_AES_GCM_SIV_NONCE_LEN) {
    OPENSSL_PUT_ERROR(CIPHER, CIPHER_R_UNSUPPORTED_NONCE_SIZE);
    return 0;
  }

  const struct aead_aes_gcm_siv_asm_ctx *gcm_siv_ctx = asm_ctx_from_ctx(ctx);
  const size_t plaintext_len = in_len - EVP_AEAD_AES_GCM_SIV_TAG_LEN;
  const uint8_t *const given_tag = in + plaintext_len;

  if (max_out_len < plaintext_len) {
    OPENSSL_PUT_ERROR(CIPHER, CIPHER_R_BUFFER_TOO_SMALL);
    return 0;
  }

  alignas(16) uint64_t record_auth_key[2];
  alignas(16) uint64_t record_enc_key[4];
  aead_aes_gcm_siv_kdf(gcm_siv_ctx->is_128_bit, gcm_siv_ctx, record_auth_key,
                       record_enc_key, nonce);

  struct aead_aes_gcm_siv_asm_ctx expanded_key;
  if (gcm_siv_ctx->is_128_bit) {
    aes128gcmsiv_aes_ks((const uint8_t *) record_enc_key, &expanded_key.key[0]);
  } else {
    aes256gcmsiv_aes_ks((const uint8_t *) record_enc_key, &expanded_key.key[0]);
  }
  // calculated_tag is 16*8 bytes, rather than 16 bytes, because
  // aes[128|256]gcmsiv_dec uses the extra as scratch space.
  alignas(16) uint8_t calculated_tag[16 * 8] = {0};

  OPENSSL_memset(calculated_tag, 0, EVP_AEAD_AES_GCM_SIV_TAG_LEN);
  const size_t ad_blocks = ad_len / 16;
  aesgcmsiv_polyval_horner(calculated_tag, (const uint8_t *)record_auth_key, ad,
                           ad_blocks);

  uint8_t scratch[16];
  if (ad_len & 15) {
    OPENSSL_memset(scratch, 0, sizeof(scratch));
    OPENSSL_memcpy(scratch, &ad[ad_len & ~15], ad_len & 15);
    aesgcmsiv_polyval_horner(calculated_tag, (const uint8_t *)record_auth_key,
                             scratch, 1);
  }

  alignas(16) uint8_t htable[16 * 6];
  aesgcmsiv_htable6_init(htable, (const uint8_t *)record_auth_key);

  if (gcm_siv_ctx->is_128_bit) {
    aes128gcmsiv_dec(in, out, calculated_tag, htable, &expanded_key,
                     plaintext_len);
  } else {
    aes256gcmsiv_dec(in, out, calculated_tag, htable, &expanded_key,
                     plaintext_len);
  }

  if (plaintext_len & 15) {
    aead_aes_gcm_siv_asm_crypt_last_block(gcm_siv_ctx->is_128_bit, out, in,
                                          plaintext_len, given_tag,
                                          &expanded_key);
    OPENSSL_memset(scratch, 0, sizeof(scratch));
    OPENSSL_memcpy(scratch, out + (plaintext_len & ~15), plaintext_len & 15);
    aesgcmsiv_polyval_horner(calculated_tag, (const uint8_t *)record_auth_key,
                             scratch, 1);
  }

  union {
    uint8_t c[16];
    struct {
      uint64_t ad;
      uint64_t in;
    } bitlens;
  } length_block;

  length_block.bitlens.ad = ad_len * 8;
  length_block.bitlens.in = plaintext_len * 8;
  aesgcmsiv_polyval_horner(calculated_tag, (const uint8_t *)record_auth_key,
                           length_block.c, 1);

  for (size_t i = 0; i < 12; i++) {
    calculated_tag[i] ^= nonce[i];
  }

  calculated_tag[15] &= 0x7f;

  if (gcm_siv_ctx->is_128_bit) {
    aes128gcmsiv_ecb_enc_block(calculated_tag, calculated_tag, &expanded_key);
  } else {
    aes256gcmsiv_ecb_enc_block(calculated_tag, calculated_tag, &expanded_key);
  }

  if (CRYPTO_memcmp(calculated_tag, given_tag, EVP_AEAD_AES_GCM_SIV_TAG_LEN) !=
      0) {
    OPENSSL_PUT_ERROR(CIPHER, CIPHER_R_BAD_DECRYPT);
    return 0;
  }

  *out_len = in_len - EVP_AEAD_AES_GCM_SIV_TAG_LEN;
  return 1;
}

static const EVP_AEAD aead_aes_128_gcm_siv_asm = {
    16,                              // key length
    EVP_AEAD_AES_GCM_SIV_NONCE_LEN,  // nonce length
    EVP_AEAD_AES_GCM_SIV_TAG_LEN,    // overhead
    EVP_AEAD_AES_GCM_SIV_TAG_LEN,    // max tag length
    0,                               // seal_scatter_supports_extra_in

    aead_aes_gcm_siv_asm_init,
    NULL /* init_with_direction */,
    aead_aes_gcm_siv_asm_cleanup,
    aead_aes_gcm_siv_asm_open,
    aead_aes_gcm_siv_asm_seal_scatter,
    NULL /* open_gather */,
    NULL /* get_iv */,
    NULL /* tag_len */,
};

static const EVP_AEAD aead_aes_256_gcm_siv_asm = {
    32,                              // key length
    EVP_AEAD_AES_GCM_SIV_NONCE_LEN,  // nonce length
    EVP_AEAD_AES_GCM_SIV_TAG_LEN,    // overhead
    EVP_AEAD_AES_GCM_SIV_TAG_LEN,    // max tag length
    0,                               // seal_scatter_supports_extra_in

    aead_aes_gcm_siv_asm_init,
    NULL /* init_with_direction */,
    aead_aes_gcm_siv_asm_cleanup,
    aead_aes_gcm_siv_asm_open,
    aead_aes_gcm_siv_asm_seal_scatter,
    NULL /* open_gather */,
    NULL /* get_iv */,
    NULL /* tag_len */,
};

#endif  // X86_64 && !NO_ASM && !WINDOWS

struct aead_aes_gcm_siv_ctx {
  union {
    double align;
    AES_KEY ks;
  } ks;
  block128_f kgk_block;
  unsigned is_256:1;
};

OPENSSL_STATIC_ASSERT(sizeof(((EVP_AEAD_CTX *)NULL)->state) >=
                          sizeof(struct aead_aes_gcm_siv_ctx),
                      "AEAD state is too small");
#if defined(__GNUC__) || defined(__clang__)
OPENSSL_STATIC_ASSERT(alignof(union evp_aead_ctx_st_state) >=
                          alignof(struct aead_aes_gcm_siv_ctx),
                      "AEAD state has insufficient alignment");
#endif

static int aead_aes_gcm_siv_init(EVP_AEAD_CTX *ctx, const uint8_t *key,
                                 size_t key_len, size_t tag_len) {
  const size_t key_bits = key_len * 8;

  if (key_bits != 128 && key_bits != 256) {
    OPENSSL_PUT_ERROR(CIPHER, CIPHER_R_BAD_KEY_LENGTH);
    return 0;  // EVP_AEAD_CTX_init should catch this.
  }

  if (tag_len == EVP_AEAD_DEFAULT_TAG_LENGTH) {
    tag_len = EVP_AEAD_AES_GCM_SIV_TAG_LEN;
  }
  if (tag_len != EVP_AEAD_AES_GCM_SIV_TAG_LEN) {
    OPENSSL_PUT_ERROR(CIPHER, CIPHER_R_TAG_TOO_LARGE);
    return 0;
  }

  struct aead_aes_gcm_siv_ctx *gcm_siv_ctx =
      (struct aead_aes_gcm_siv_ctx *)&ctx->state;
  OPENSSL_memset(gcm_siv_ctx, 0, sizeof(struct aead_aes_gcm_siv_ctx));

  aes_ctr_set_key(&gcm_siv_ctx->ks.ks, NULL, &gcm_siv_ctx->kgk_block, key,
                  key_len);
  gcm_siv_ctx->is_256 = (key_len == 32);
  ctx->tag_len = tag_len;

  return 1;
}

static void aead_aes_gcm_siv_cleanup(EVP_AEAD_CTX *ctx) {}

// gcm_siv_crypt encrypts (or decrypts—it's the same thing) |in_len| bytes from
// |in| to |out|, using the block function |enc_block| with |key| in counter
// mode, starting at |initial_counter|. This differs from the traditional
// counter mode code in that the counter is handled little-endian, only the
// first four bytes are used and the GCM-SIV tweak to the final byte is
// applied. The |in| and |out| pointers may be equal but otherwise must not
// alias.
static void gcm_siv_crypt(uint8_t *out, const uint8_t *in, size_t in_len,
                          const uint8_t initial_counter[AES_BLOCK_SIZE],
                          block128_f enc_block, const AES_KEY *key) {
  union {
    uint32_t w[4];
    uint8_t c[16];
  } counter;

  OPENSSL_memcpy(counter.c, initial_counter, AES_BLOCK_SIZE);
  counter.c[15] |= 0x80;

  for (size_t done = 0; done < in_len;) {
    uint8_t keystream[AES_BLOCK_SIZE];
    enc_block(counter.c, keystream, key);
    counter.w[0]++;

    size_t todo = AES_BLOCK_SIZE;
    if (in_len - done < todo) {
      todo = in_len - done;
    }

    for (size_t i = 0; i < todo; i++) {
      out[done + i] = keystream[i] ^ in[done + i];
    }

    done += todo;
  }
}

// gcm_siv_polyval evaluates POLYVAL at |auth_key| on the given plaintext and
// AD. The result is written to |out_tag|.
static void gcm_siv_polyval(
    uint8_t out_tag[16], const uint8_t *in, size_t in_len, const uint8_t *ad,
    size_t ad_len, const uint8_t auth_key[16],
    const uint8_t nonce[EVP_AEAD_AES_GCM_SIV_NONCE_LEN]) {
  struct polyval_ctx polyval_ctx;
  CRYPTO_POLYVAL_init(&polyval_ctx, auth_key);

  CRYPTO_POLYVAL_update_blocks(&polyval_ctx, ad, ad_len & ~15);

  uint8_t scratch[16];
  if (ad_len & 15) {
    OPENSSL_memset(scratch, 0, sizeof(scratch));
    OPENSSL_memcpy(scratch, &ad[ad_len & ~15], ad_len & 15);
    CRYPTO_POLYVAL_update_blocks(&polyval_ctx, scratch, sizeof(scratch));
  }

  CRYPTO_POLYVAL_update_blocks(&polyval_ctx, in, in_len & ~15);
  if (in_len & 15) {
    OPENSSL_memset(scratch, 0, sizeof(scratch));
    OPENSSL_memcpy(scratch, &in[in_len & ~15], in_len & 15);
    CRYPTO_POLYVAL_update_blocks(&polyval_ctx, scratch, sizeof(scratch));
  }

  union {
    uint8_t c[16];
    struct {
      uint64_t ad;
      uint64_t in;
    } bitlens;
  } length_block;

  length_block.bitlens.ad = ad_len * 8;
  length_block.bitlens.in = in_len * 8;
  CRYPTO_POLYVAL_update_blocks(&polyval_ctx, length_block.c,
                               sizeof(length_block));

  CRYPTO_POLYVAL_finish(&polyval_ctx, out_tag);
  for (size_t i = 0; i < EVP_AEAD_AES_GCM_SIV_NONCE_LEN; i++) {
    out_tag[i] ^= nonce[i];
  }
  out_tag[15] &= 0x7f;
}

// gcm_siv_record_keys contains the keys used for a specific GCM-SIV record.
struct gcm_siv_record_keys {
  uint8_t auth_key[16];
  union {
    double align;
    AES_KEY ks;
  } enc_key;
  block128_f enc_block;
};

// gcm_siv_keys calculates the keys for a specific GCM-SIV record with the
// given nonce and writes them to |*out_keys|.
static void gcm_siv_keys(
    const struct aead_aes_gcm_siv_ctx *gcm_siv_ctx,
    struct gcm_siv_record_keys *out_keys,
    const uint8_t nonce[EVP_AEAD_AES_GCM_SIV_NONCE_LEN]) {
  const AES_KEY *const key = &gcm_siv_ctx->ks.ks;
  uint8_t key_material[(128 /* POLYVAL key */ + 256 /* max AES key */) / 8];
  const size_t blocks_needed = gcm_siv_ctx->is_256 ? 6 : 4;

  uint8_t counter[AES_BLOCK_SIZE];
  OPENSSL_memset(counter, 0, AES_BLOCK_SIZE - EVP_AEAD_AES_GCM_SIV_NONCE_LEN);
  OPENSSL_memcpy(counter + AES_BLOCK_SIZE - EVP_AEAD_AES_GCM_SIV_NONCE_LEN,
                 nonce, EVP_AEAD_AES_GCM_SIV_NONCE_LEN);
  for (size_t i = 0; i < blocks_needed; i++) {
    counter[0] = i;

    uint8_t ciphertext[AES_BLOCK_SIZE];
    gcm_siv_ctx->kgk_block(counter, ciphertext, key);
    OPENSSL_memcpy(&key_material[i * 8], ciphertext, 8);
  }

  OPENSSL_memcpy(out_keys->auth_key, key_material, 16);
  // Note the |ctr128_f| function uses a big-endian couner, while AES-GCM-SIV
  // uses a little-endian counter. We ignore the return value and only use
  // |block128_f|. This has a significant performance cost for the fallback
  // bitsliced AES implementations (bsaes and aes_nohw).
  //
  // We currently do not consider AES-GCM-SIV to be performance-sensitive on
  // client hardware. If this changes, we can write little-endian |ctr128_f|
  // functions.
  aes_ctr_set_key(&out_keys->enc_key.ks, NULL, &out_keys->enc_block,
                  key_material + 16, gcm_siv_ctx->is_256 ? 32 : 16);
}

static int aead_aes_gcm_siv_seal_scatter(
    const EVP_AEAD_CTX *ctx, uint8_t *out, uint8_t *out_tag,
    size_t *out_tag_len, size_t max_out_tag_len, const uint8_t *nonce,
    size_t nonce_len, const uint8_t *in, size_t in_len, const uint8_t *extra_in,
    size_t extra_in_len, const uint8_t *ad, size_t ad_len) {
  const struct aead_aes_gcm_siv_ctx *gcm_siv_ctx =
      (struct aead_aes_gcm_siv_ctx *)&ctx->state;
  const uint64_t in_len_64 = in_len;
  const uint64_t ad_len_64 = ad_len;

  if (in_len + EVP_AEAD_AES_GCM_SIV_TAG_LEN < in_len ||
      in_len_64 > (UINT64_C(1) << 36) ||
      ad_len_64 >= (UINT64_C(1) << 61)) {
    OPENSSL_PUT_ERROR(CIPHER, CIPHER_R_TOO_LARGE);
    return 0;
  }

  if (max_out_tag_len < EVP_AEAD_AES_GCM_SIV_TAG_LEN) {
    OPENSSL_PUT_ERROR(CIPHER, CIPHER_R_BUFFER_TOO_SMALL);
    return 0;
  }

  if (nonce_len != EVP_AEAD_AES_GCM_SIV_NONCE_LEN) {
    OPENSSL_PUT_ERROR(CIPHER, CIPHER_R_UNSUPPORTED_NONCE_SIZE);
    return 0;
  }

  struct gcm_siv_record_keys keys;
  gcm_siv_keys(gcm_siv_ctx, &keys, nonce);

  uint8_t tag[16];
  gcm_siv_polyval(tag, in, in_len, ad, ad_len, keys.auth_key, nonce);
  keys.enc_block(tag, tag, &keys.enc_key.ks);

  gcm_siv_crypt(out, in, in_len, tag, keys.enc_block, &keys.enc_key.ks);

  OPENSSL_memcpy(out_tag, tag, EVP_AEAD_AES_GCM_SIV_TAG_LEN);
  *out_tag_len = EVP_AEAD_AES_GCM_SIV_TAG_LEN;

  return 1;
}

static int aead_aes_gcm_siv_open_gather(const EVP_AEAD_CTX *ctx, uint8_t *out,
                                        const uint8_t *nonce, size_t nonce_len,
                                        const uint8_t *in, size_t in_len,
                                        const uint8_t *in_tag,
                                        size_t in_tag_len, const uint8_t *ad,
                                        size_t ad_len) {
  const uint64_t ad_len_64 = ad_len;
  if (ad_len_64 >= (UINT64_C(1) << 61)) {
    OPENSSL_PUT_ERROR(CIPHER, CIPHER_R_TOO_LARGE);
    return 0;
  }

  const uint64_t in_len_64 = in_len;
  if (in_tag_len != EVP_AEAD_AES_GCM_SIV_TAG_LEN ||
      in_len_64 > (UINT64_C(1) << 36) + AES_BLOCK_SIZE) {
    OPENSSL_PUT_ERROR(CIPHER, CIPHER_R_BAD_DECRYPT);
    return 0;
  }

  if (nonce_len != EVP_AEAD_AES_GCM_SIV_NONCE_LEN) {
    OPENSSL_PUT_ERROR(CIPHER, CIPHER_R_UNSUPPORTED_NONCE_SIZE);
    return 0;
  }

  const struct aead_aes_gcm_siv_ctx *gcm_siv_ctx =
      (struct aead_aes_gcm_siv_ctx *)&ctx->state;

  struct gcm_siv_record_keys keys;
  gcm_siv_keys(gcm_siv_ctx, &keys, nonce);

  gcm_siv_crypt(out, in, in_len, in_tag, keys.enc_block, &keys.enc_key.ks);

  uint8_t expected_tag[EVP_AEAD_AES_GCM_SIV_TAG_LEN];
  gcm_siv_polyval(expected_tag, out, in_len, ad, ad_len, keys.auth_key, nonce);
  keys.enc_block(expected_tag, expected_tag, &keys.enc_key.ks);

  if (CRYPTO_memcmp(expected_tag, in_tag, sizeof(expected_tag)) != 0) {
    OPENSSL_PUT_ERROR(CIPHER, CIPHER_R_BAD_DECRYPT);
    return 0;
  }

  return 1;
}

static const EVP_AEAD aead_aes_128_gcm_siv = {
    16,                              // key length
    EVP_AEAD_AES_GCM_SIV_NONCE_LEN,  // nonce length
    EVP_AEAD_AES_GCM_SIV_TAG_LEN,    // overhead
    EVP_AEAD_AES_GCM_SIV_TAG_LEN,    // max tag length
    0,                               // seal_scatter_supports_extra_in

    aead_aes_gcm_siv_init,
    NULL /* init_with_direction */,
    aead_aes_gcm_siv_cleanup,
    NULL /* open */,
    aead_aes_gcm_siv_seal_scatter,
    aead_aes_gcm_siv_open_gather,
    NULL /* get_iv */,
    NULL /* tag_len */,
};

static const EVP_AEAD aead_aes_256_gcm_siv = {
    32,                              // key length
    EVP_AEAD_AES_GCM_SIV_NONCE_LEN,  // nonce length
    EVP_AEAD_AES_GCM_SIV_TAG_LEN,    // overhead
    EVP_AEAD_AES_GCM_SIV_TAG_LEN,    // max tag length
    0,                               // seal_scatter_supports_extra_in

    aead_aes_gcm_siv_init,
    NULL /* init_with_direction */,
    aead_aes_gcm_siv_cleanup,
    NULL /* open */,
    aead_aes_gcm_siv_seal_scatter,
    aead_aes_gcm_siv_open_gather,
    NULL /* get_iv */,
    NULL /* tag_len */,
};

#if defined(AES_GCM_SIV_ASM)

static char avx_aesni_capable(void) {
  const uint32_t ecx = OPENSSL_ia32cap_P[1];

  return (ecx & (1 << (57 - 32))) != 0 /* AESNI */ &&
         (ecx & (1 << 28)) != 0 /* AVX */;
}

const EVP_AEAD *EVP_aead_aes_128_gcm_siv(void) {
  if (avx_aesni_capable()) {
    return &aead_aes_128_gcm_siv_asm;
  }
  return &aead_aes_128_gcm_siv;
}

const EVP_AEAD *EVP_aead_aes_256_gcm_siv(void) {
  if (avx_aesni_capable()) {
    return &aead_aes_256_gcm_siv_asm;
  }
  return &aead_aes_256_gcm_siv;
}

#else

const EVP_AEAD *EVP_aead_aes_128_gcm_siv(void) {
  return &aead_aes_128_gcm_siv;
}

const EVP_AEAD *EVP_aead_aes_256_gcm_siv(void) {
  return &aead_aes_256_gcm_siv;
}

#endif  // AES_GCM_SIV_ASM