#include <stdbool.h>
#include <stdint.h>
#include <string.h>

#include "cipher.h"
#include "lilliput-ae.h"


static const uint8_t _0n[BLOCK_BYTES] = {
    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
};


static uint8_t _upper_nibble(uint8_t i)
{
    return i >> 4;
}

static uint8_t _lower_nibble(uint8_t i)
{
    return i & 0x0f;
}

static void _encrypt(const uint8_t K[KEY_BYTES],
                     const uint8_t T[TWEAK_BYTES],
                     const uint8_t M[BLOCK_BYTES],
                     uint8_t C[BLOCK_BYTES])
{
    lilliput_tbc_encrypt(K, T, M, C, NULL);
}

static void _decrypt(const uint8_t K[KEY_BYTES],
                     const uint8_t T[TWEAK_BYTES],
                     const uint8_t C[BLOCK_BYTES],
                     uint8_t M[BLOCK_BYTES])
{
    lilliput_tbc_decrypt(K, T, C, M, NULL);
}

static void _xor_into(uint8_t dest[BLOCK_BYTES], const uint8_t src[BLOCK_BYTES])
{
    for (size_t i=0; i<BLOCK_BYTES; i++)
        dest[i] ^= src[i];
}

static void _xor_arrays(size_t len, uint8_t out[len], const uint8_t a[len], const uint8_t b[len])
{
    for (size_t i=0; i<len; i++)
        out[i] = a[i] ^ b[i];
}

static void _pad10(size_t len, const uint8_t buf[len], uint8_t padded[BLOCK_BYTES])
{
    /* Assume that len<BLOCK_BYTES. */
    memcpy(padded, buf, len);
    padded[len] = 0x80;

    if (len+1 < BLOCK_BYTES)
    {
        memset(padded+len+1, 0, BLOCK_BYTES-len-1);
    }
}

static void _fill_ad_tweak(
    uint8_t  prefix,
    uint64_t block_nb,
    uint8_t  tweak[TWEAK_BYTES]
)
{
    /* The 192-bit tweak is filled as follows:
     *
     * - bits   1-188: block number
     *          1- 64: actual 64-bit block number
     *         65-188: 0-padding
     * - bits 189-192: constant 4-bit prefix
     */

    for (size_t i=0; i<sizeof(block_nb); i++)
    {
        uint64_t mask = (uint64_t)0xff << 8*i;
        uint8_t b = (mask & block_nb) >> 8*i;

        tweak[i] = b;
    }

    /* Assume padding bytes have already been memset to 0. */

    tweak[TWEAK_BYTES-1] |= prefix << 4;
}

static void _fill_msg_tweak(
    uint8_t       prefix,
    const uint8_t N[NONCE_BYTES],
    uint64_t      block_nb,
    uint8_t       tweak[TWEAK_BYTES]
)
{
    /* The 192-bit tweak is filled as follows:
     *
     * - bits   1- 68: block number
     *          1- 64: actual 64-bit block number
     *         64- 68: 0-padding
     * - bits  67-188: nonce
     * - bits 189-192: constant 4-bit prefix
     */

    for (size_t i=0; i<sizeof(block_nb); i++)
    {
        uint64_t mask = (uint64_t)0xff << 8*i;
        uint8_t b = (mask & block_nb) >> 8*i;

        tweak[i] = b;
    }

    tweak[sizeof(block_nb)] = _lower_nibble(N[0]) << 4;

    for (size_t i=1; i<NONCE_BYTES; i++)
    {
        tweak[sizeof(block_nb)+i] = _lower_nibble(N[i]) << 4 ^ _upper_nibble(N[i-1]);
    }

    tweak[TWEAK_BYTES-1] = prefix << 4 ^ _upper_nibble(N[NONCE_BYTES-1]);
}

static void _process_associated_data(
    const uint8_t key[KEY_BYTES],
    size_t        A_len,
    const uint8_t A[A_len],
    uint8_t       Auth[BLOCK_BYTES]
)
{
    uint8_t Ek_Ai[BLOCK_BYTES];
    uint8_t tweak[TWEAK_BYTES];

    memset(tweak, 0, TWEAK_BYTES);
    memset(Auth, 0, BLOCK_BYTES);

    size_t l_a = A_len / BLOCK_BYTES;
    size_t rest = A_len % BLOCK_BYTES;

    for (size_t i=0; i<l_a; i++)
    {
        _fill_ad_tweak(0x2, i, tweak);
        _encrypt(key, tweak, &A[i*BLOCK_BYTES], Ek_Ai);
        _xor_into(Auth, Ek_Ai);
    }

    if (rest != 0)
    {
        uint8_t A_rest[BLOCK_BYTES];
        _pad10(rest, &A[l_a*BLOCK_BYTES], A_rest);
        _fill_ad_tweak(0x6, l_a, tweak);
        _encrypt(key, tweak, A_rest, Ek_Ai);
        _xor_into(Auth, Ek_Ai);
    }
}

static void _encrypt_message(
    const uint8_t key[KEY_BYTES],
    size_t        M_len,
    const uint8_t M[M_len],
    const uint8_t N[NONCE_BYTES],
    uint8_t       C[M_len+BLOCK_BYTES],
    uint8_t       Final[BLOCK_BYTES]
)
{
    size_t l = M_len / BLOCK_BYTES;
    size_t rest = M_len % BLOCK_BYTES;

    uint8_t tweak[TWEAK_BYTES];
    uint8_t checksum[BLOCK_BYTES];

    memset(tweak, 0, TWEAK_BYTES);
    memset(checksum, 0, BLOCK_BYTES);

    for (size_t j=0; j<l; j++)
    {
        _xor_into(checksum, &M[j*BLOCK_BYTES]);
        _fill_msg_tweak(0x0, N, j, tweak);
        _encrypt(key, tweak, &M[j*BLOCK_BYTES], &C[j*BLOCK_BYTES]);
    }

    if (rest == 0)
    {
        _fill_msg_tweak(0x1, N, l-1, tweak);
        _encrypt(key, tweak, checksum, Final);
    }
    else
    {
        uint8_t M_rest[BLOCK_BYTES];
        uint8_t Pad[BLOCK_BYTES];

        _pad10(rest, &M[l*BLOCK_BYTES], M_rest);
        _xor_into(checksum, M_rest);

        _fill_msg_tweak(0x4, N, l, tweak);
        _encrypt(key, tweak, _0n, Pad);
        _xor_arrays(rest, &C[l*BLOCK_BYTES], &M[l*BLOCK_BYTES], Pad);

        _fill_msg_tweak(0x5, N, l, tweak);
        _encrypt(key, tweak, checksum, Final);
    }
}

static void _decrypt_message(
    const uint8_t key[KEY_BYTES],
    size_t        C_len,
    const uint8_t C[C_len],
    const uint8_t N[NONCE_BYTES],
    uint8_t       M[C_len],
    uint8_t       Final[BLOCK_BYTES]
)
{
    size_t l = C_len / BLOCK_BYTES;
    size_t rest = C_len % BLOCK_BYTES;

    uint8_t tweak[TWEAK_BYTES];
    uint8_t checksum[BLOCK_BYTES];

    memset(tweak, 0, TWEAK_BYTES);
    memset(checksum, 0, BLOCK_BYTES);

    for (size_t j=0; j<l; j++)
    {
        _fill_msg_tweak(0x0, N, j, tweak);
        _decrypt(key, tweak, &C[j*BLOCK_BYTES], &M[j*BLOCK_BYTES]);
        _xor_into(checksum, &M[j*BLOCK_BYTES]);
    }

    if (rest == 0)
    {
        _fill_msg_tweak(0x1, N, l-1, tweak);
        _encrypt(key, tweak, checksum, Final);
    }
    else
    {
        uint8_t M_rest[BLOCK_BYTES];
        uint8_t Pad[BLOCK_BYTES];

        _fill_msg_tweak(0x4, N, l, tweak);
        _encrypt(key, tweak, _0n, Pad);
        _xor_arrays(rest, &M[l*BLOCK_BYTES], &C[l*BLOCK_BYTES], Pad);

        _pad10(rest, &M[l*BLOCK_BYTES], M_rest);
        _xor_into(checksum, M_rest);

        _fill_msg_tweak(0x5, N, l, tweak);
        _encrypt(key, tweak, checksum, Final);
    }
}

static void _generate_tag(
    const uint8_t Final[BLOCK_BYTES],
    const uint8_t Auth[BLOCK_BYTES],
    uint8_t       tag[TAG_BYTES]
)
{
    _xor_arrays(TAG_BYTES, tag, Final, Auth);
}


void lilliput_ae_encrypt(
    size_t        message_len,
    const uint8_t message[message_len],
    size_t        auth_data_len,
    const uint8_t auth_data[auth_data_len],
    const uint8_t key[KEY_BYTES],
    const uint8_t nonce[NONCE_BYTES],
    uint8_t       ciphertext[message_len],
    uint8_t       tag[TAG_BYTES]
)
{
    uint8_t auth[BLOCK_BYTES];
    _process_associated_data(key, auth_data_len, auth_data, auth);

    uint8_t final[BLOCK_BYTES];
    _encrypt_message(key, message_len, message, nonce, ciphertext, final);

    _generate_tag(final, auth, tag);
}

bool lilliput_ae_decrypt(
    size_t        ciphertext_len,
    const uint8_t ciphertext[ciphertext_len],
    size_t        auth_data_len,
    const uint8_t auth_data[auth_data_len],
    const uint8_t key[KEY_BYTES],
    const uint8_t nonce[NONCE_BYTES],
    const uint8_t tag[TAG_BYTES],
    uint8_t       message[ciphertext_len]
)
{
    uint8_t auth[BLOCK_BYTES];
    _process_associated_data(key, auth_data_len, auth_data, auth);

    uint8_t final[BLOCK_BYTES];
    _decrypt_message(key, ciphertext_len, ciphertext, nonce, message, final);

    uint8_t effective_tag[TAG_BYTES];
    _generate_tag(final, auth, effective_tag);

    return memcmp(tag, effective_tag, TAG_BYTES) == 0;
}