diff options
author | Damjan Marion <damarion@cisco.com> | 2020-02-25 11:51:48 +0100 |
---|---|---|
committer | Damjan Marion <damarion@cisco.com> | 2020-02-25 11:51:48 +0100 |
commit | 47d8f5dcd655f0e01774822fca17dd38e7f557dd (patch) | |
tree | 044235f10e7e53b25abb0bfc13f4aabcbfe77487 /src | |
parent | 8d6d74cdf43d7560eab3cf609cab27e5deb816e0 (diff) |
crypto-native: GCM implementation with vector AESNI instructions
Introduced on intel IceLake uarch.
Type: feature
Change-Id: I1514c76c34e53ce0577666caf32a50f95eb6548f
Signed-off-by: Damjan Marion <damarion@cisco.com>
Diffstat (limited to 'src')
-rw-r--r-- | src/plugins/crypto_native/aes_gcm.c | 483 | ||||
-rw-r--r-- | src/vppinfra/vector_avx512.h | 48 |
2 files changed, 527 insertions, 4 deletions
diff --git a/src/plugins/crypto_native/aes_gcm.c b/src/plugins/crypto_native/aes_gcm.c index cb4590085c0..16bcfc4b01d 100644 --- a/src/plugins/crypto_native/aes_gcm.c +++ b/src/plugins/crypto_native/aes_gcm.c @@ -26,19 +26,31 @@ #pragma GCC optimize ("O3") #endif +#ifdef __VAES__ +#define NUM_HI 32 +#else #define NUM_HI 8 +#endif + typedef struct { /* pre-calculated hash key values */ const u8x16 Hi[NUM_HI]; /* extracted AES key */ const u8x16 Ke[15]; +#ifdef __VAES__ + const u8x64 Ke4[15]; +#endif } aes_gcm_key_data_t; typedef struct { u32 counter; - u32x4 Y; + union + { + u32x4 Y; + u32x16 Y4; + }; } aes_gcm_counter_t; typedef enum @@ -51,8 +63,6 @@ typedef enum static const u32x4 ctr_inv_1 = { 0, 0, 0, 1 << 24 }; - - static_always_inline void aes_gcm_enc_first_round (u8x16 * r, aes_gcm_counter_t * ctr, u8x16 k, int n_blocks) @@ -405,6 +415,334 @@ aes_gcm_ghash_last (u8x16 T, aes_gcm_key_data_t * kd, u8x16 * d, return ghash_final (gd); } +#ifdef __VAES__ +static const u32x16 ctr_inv_1234 = { + 0, 0, 0, 1 << 24, 0, 0, 0, 2 << 24, 0, 0, 0, 3 << 24, 0, 0, 0, 4 << 24, +}; + +static const u32x16 ctr_inv_4444 = { + 0, 0, 0, 4 << 24, 0, 0, 0, 4 << 24, 0, 0, 0, 4 << 24, 0, 0, 0, 4 << 24 +}; + +static const u32x16 ctr_1234 = { + 1, 0, 0, 0, 2, 0, 0, 0, 3, 0, 0, 0, 4, 0, 0, 0, +}; + +static_always_inline void +aes4_gcm_enc_first_round (u8x64 * r, aes_gcm_counter_t * ctr, u8x64 k, int n) +{ + u8 last_byte = (u8) ctr->counter; + int i = 0; + + /* As counter is stored in network byte order for performance reasons we + are incrementing least significant byte only except in case where we + overlow. As we are processing four 512-blocks in parallel except the + last round, overflow can happen only when n == 4 */ + + if (n == 4) + for (; i < 2; i++) + { + r[i] = k ^ (u8x64) ctr->Y4; + ctr->Y4 += ctr_inv_4444; + } + + if (n == 4 && PREDICT_TRUE (last_byte == 241)) + { + u32x16 Yc, Yr = (u32x16) u8x64_reflect_u8x16 ((u8x64) ctr->Y4); + + for (; i < n; i++) + { + r[i] = k ^ (u8x64) ctr->Y4; + Yc = u32x16_splat (ctr->counter + 4 * (i + 1)) + ctr_1234; + Yr = (u32x16) u32x16_mask_blend (Yr, Yc, 0x1111); + ctr->Y4 = (u32x16) u8x64_reflect_u8x16 ((u8x64) Yr); + } + } + else + { + for (; i < n; i++) + { + r[i] = k ^ (u8x64) ctr->Y4; + ctr->Y4 += ctr_inv_4444; + } + } + ctr->counter += n * 4; +} + +static_always_inline void +aes4_gcm_enc_round (u8x64 * r, u8x64 k, int n_blocks) +{ + for (int i = 0; i < n_blocks; i++) + r[i] = aes_enc_round_x4 (r[i], k); +} + +static_always_inline void +aes4_gcm_enc_last_round (u8x64 * r, u8x64 * d, u8x64 const *k, + int rounds, int n_blocks) +{ + + /* additional ronuds for AES-192 and AES-256 */ + for (int i = 10; i < rounds; i++) + aes4_gcm_enc_round (r, k[i], n_blocks); + + for (int i = 0; i < n_blocks; i++) + d[i] ^= aes_enc_last_round_x4 (r[i], k[rounds]); +} + +static_always_inline u8x16 +aes4_gcm_calc (u8x16 T, aes_gcm_key_data_t * kd, u8x64 * d, + aes_gcm_counter_t * ctr, u8x16u * in, u8x16u * out, + int rounds, int n, int last_4block_bytes, aes_gcm_flags_t f) +{ + ghash4_data_t _gd, *gd = &_gd; + const u8x64 *rk = (u8x64 *) kd->Ke4; + int i, ghash_blocks, gc = 1; + u8x64u *Hi4, *inv = (u8x64u *) in, *outv = (u8x64u *) out; + u8x64 r[4]; + u64 byte_mask = _bextr_u64 (-1LL, 0, last_4block_bytes); + + if (f & AES_GCM_F_ENCRYPT) + { + /* during encryption we either hash four 512-bit blocks from previous + round or we don't hash at all */ + ghash_blocks = 4; + Hi4 = (u8x64u *) (kd->Hi + NUM_HI - ghash_blocks * 4); + } + else + { + /* during deccryption we hash 1..4 512-bit blocks from current round */ + ghash_blocks = n; + int n_128bit_blocks = n * 4; + /* if this is last round of decryption, we may have less than 4 + 128-bit blocks in the last 512-bit data block, so we need to adjust + Hi4 pointer accordingly */ + if (f & AES_GCM_F_LAST_ROUND) + n_128bit_blocks += ((last_4block_bytes + 15) >> 4) - 4; + Hi4 = (u8x64u *) (kd->Hi + NUM_HI - n_128bit_blocks); + } + + /* AES rounds 0 and 1 */ + aes4_gcm_enc_first_round (r, ctr, rk[0], n); + aes4_gcm_enc_round (r, rk[1], n); + + /* load 4 blocks of data - decrypt round */ + if (f & AES_GCM_F_DECRYPT) + { + for (i = 0; i < n - ((f & AES_GCM_F_LAST_ROUND) != 0); i++) + d[i] = inv[i]; + + if (f & AES_GCM_F_LAST_ROUND) + d[i] = u8x64_mask_load (u8x64_splat (0), inv + i, byte_mask); + } + + /* GHASH multiply block 0 */ + if (f & AES_GCM_F_WITH_GHASH) + ghash4_mul_first (gd, u8x64_reflect_u8x16 (d[0]) ^ + u8x64_insert_u8x16 (u8x64_splat (0), T, 0), Hi4[0]); + + /* AES rounds 2 and 3 */ + aes4_gcm_enc_round (r, rk[2], n); + aes4_gcm_enc_round (r, rk[3], n); + + /* GHASH multiply block 1 */ + if ((f & AES_GCM_F_WITH_GHASH) && gc++ < ghash_blocks) + ghash4_mul_next (gd, u8x64_reflect_u8x16 (d[1]), Hi4[1]); + + /* AES rounds 4 and 5 */ + aes4_gcm_enc_round (r, rk[4], n); + aes4_gcm_enc_round (r, rk[5], n); + + /* GHASH multiply block 2 */ + if ((f & AES_GCM_F_WITH_GHASH) && gc++ < ghash_blocks) + ghash4_mul_next (gd, u8x64_reflect_u8x16 (d[2]), Hi4[2]); + + /* AES rounds 6 and 7 */ + aes4_gcm_enc_round (r, rk[6], n); + aes4_gcm_enc_round (r, rk[7], n); + + /* GHASH multiply block 3 */ + if ((f & AES_GCM_F_WITH_GHASH) && gc++ < ghash_blocks) + ghash4_mul_next (gd, u8x64_reflect_u8x16 (d[3]), Hi4[3]); + + /* load 4 blocks of data - decrypt round */ + if (f & AES_GCM_F_ENCRYPT) + { + for (i = 0; i < n - ((f & AES_GCM_F_LAST_ROUND) != 0); i++) + d[i] = inv[i]; + + if (f & AES_GCM_F_LAST_ROUND) + d[i] = u8x64_mask_load (u8x64_splat (0), inv + i, byte_mask); + } + + /* AES rounds 8 and 9 */ + aes4_gcm_enc_round (r, rk[8], n); + aes4_gcm_enc_round (r, rk[9], n); + + /* AES last round(s) */ + aes4_gcm_enc_last_round (r, d, rk, rounds, n); + + /* store 4 blocks of data */ + for (i = 0; i < n - ((f & AES_GCM_F_LAST_ROUND) != 0); i++) + outv[i] = d[i]; + + if (f & AES_GCM_F_LAST_ROUND) + u8x64_mask_store (d[i], outv + i, byte_mask); + + /* GHASH reduce 1st step */ + ghash4_reduce (gd); + + /* GHASH reduce 2nd step */ + ghash4_reduce2 (gd); + + /* GHASH final step */ + return ghash4_final (gd); +} + +static_always_inline u8x16 +aes4_gcm_calc_double (u8x16 T, aes_gcm_key_data_t * kd, u8x64 * d, + aes_gcm_counter_t * ctr, u8x16u * in, u8x16u * out, + int rounds, aes_gcm_flags_t f) +{ + u8x64 r[4]; + ghash4_data_t _gd, *gd = &_gd; + const u8x64 *rk = (u8x64 *) kd->Ke4; + u8x64 *Hi4 = (u8x64 *) (kd->Hi + NUM_HI - 32); + u8x64u *inv = (u8x64u *) in, *outv = (u8x64u *) out; + + /* AES rounds 0 and 1 */ + aes4_gcm_enc_first_round (r, ctr, rk[0], 4); + aes4_gcm_enc_round (r, rk[1], 4); + + /* load 4 blocks of data - decrypt round */ + if (f & AES_GCM_F_DECRYPT) + for (int i = 0; i < 4; i++) + d[i] = inv[i]; + + /* GHASH multiply block 0 */ + ghash4_mul_first (gd, u8x64_reflect_u8x16 (d[0]) ^ + u8x64_insert_u8x16 (u8x64_splat (0), T, 0), Hi4[0]); + + /* AES rounds 2 and 3 */ + aes4_gcm_enc_round (r, rk[2], 4); + aes4_gcm_enc_round (r, rk[3], 4); + + /* GHASH multiply block 1 */ + ghash4_mul_next (gd, u8x64_reflect_u8x16 (d[1]), Hi4[1]); + + /* AES rounds 4 and 5 */ + aes4_gcm_enc_round (r, rk[4], 4); + aes4_gcm_enc_round (r, rk[5], 4); + + /* GHASH multiply block 2 */ + ghash4_mul_next (gd, u8x64_reflect_u8x16 (d[2]), Hi4[2]); + + /* AES rounds 6 and 7 */ + aes4_gcm_enc_round (r, rk[6], 4); + aes4_gcm_enc_round (r, rk[7], 4); + + /* GHASH multiply block 3 */ + ghash4_mul_next (gd, u8x64_reflect_u8x16 (d[3]), Hi4[3]); + + /* AES rounds 8 and 9 */ + aes4_gcm_enc_round (r, rk[8], 4); + aes4_gcm_enc_round (r, rk[9], 4); + + /* load 4 blocks of data - encrypt round */ + if (f & AES_GCM_F_ENCRYPT) + for (int i = 0; i < 4; i++) + d[i] = inv[i]; + + /* AES last round(s) */ + aes4_gcm_enc_last_round (r, d, rk, rounds, 4); + + /* store 4 blocks of data */ + for (int i = 0; i < 4; i++) + outv[i] = d[i]; + + /* load 4 blocks of data - decrypt round */ + if (f & AES_GCM_F_DECRYPT) + for (int i = 0; i < 4; i++) + d[i] = inv[i + 4]; + + /* GHASH multiply block 3 */ + ghash4_mul_next (gd, u8x64_reflect_u8x16 (d[0]), Hi4[4]); + + /* AES rounds 0 and 1 */ + aes4_gcm_enc_first_round (r, ctr, rk[0], 4); + aes4_gcm_enc_round (r, rk[1], 4); + + /* GHASH multiply block 5 */ + ghash4_mul_next (gd, u8x64_reflect_u8x16 (d[1]), Hi4[5]); + + /* AES rounds 2 and 3 */ + aes4_gcm_enc_round (r, rk[2], 4); + aes4_gcm_enc_round (r, rk[3], 4); + + /* GHASH multiply block 6 */ + ghash4_mul_next (gd, u8x64_reflect_u8x16 (d[2]), Hi4[6]); + + /* AES rounds 4 and 5 */ + aes4_gcm_enc_round (r, rk[4], 4); + aes4_gcm_enc_round (r, rk[5], 4); + + /* GHASH multiply block 7 */ + ghash4_mul_next (gd, u8x64_reflect_u8x16 (d[3]), Hi4[7]); + + /* AES rounds 6 and 7 */ + aes4_gcm_enc_round (r, rk[6], 4); + aes4_gcm_enc_round (r, rk[7], 4); + + /* GHASH reduce 1st step */ + ghash4_reduce (gd); + + /* AES rounds 8 and 9 */ + aes4_gcm_enc_round (r, rk[8], 4); + aes4_gcm_enc_round (r, rk[9], 4); + + /* GHASH reduce 2nd step */ + ghash4_reduce2 (gd); + + /* load 4 blocks of data - encrypt round */ + if (f & AES_GCM_F_ENCRYPT) + for (int i = 0; i < 4; i++) + d[i] = inv[i + 4]; + + /* AES last round(s) */ + aes4_gcm_enc_last_round (r, d, rk, rounds, 4); + + /* store 4 blocks of data */ + for (int i = 0; i < 4; i++) + outv[i + 4] = d[i]; + + /* GHASH final step */ + return ghash4_final (gd); +} + +static_always_inline u8x16 +aes4_gcm_ghash_last (u8x16 T, aes_gcm_key_data_t * kd, u8x64 * d, + int n, int last_4block_bytes) +{ + ghash4_data_t _gd, *gd = &_gd; + u8x64u *Hi4; + int n_128bit_blocks; + u64 byte_mask = _bextr_u64 (-1LL, 0, last_4block_bytes); + n_128bit_blocks = (n - 1) * 4 + ((last_4block_bytes + 15) >> 4); + Hi4 = (u8x64u *) (kd->Hi + NUM_HI - n_128bit_blocks); + + d[n - 1] = u8x64_mask_blend (u8x64_splat (0), d[n - 1], byte_mask); + ghash4_mul_first (gd, u8x64_reflect_u8x16 (d[0]) ^ + u8x64_insert_u8x16 (u8x64_splat (0), T, 0), Hi4[0]); + if (n > 1) + ghash4_mul_next (gd, u8x64_reflect_u8x16 (d[1]), Hi4[1]); + if (n > 2) + ghash4_mul_next (gd, u8x64_reflect_u8x16 (d[2]), Hi4[2]); + if (n > 3) + ghash4_mul_next (gd, u8x64_reflect_u8x16 (d[3]), Hi4[3]); + ghash4_reduce (gd); + ghash4_reduce2 (gd); + return ghash4_final (gd); +} +#endif static_always_inline u8x16 aes_gcm_enc (u8x16 T, aes_gcm_key_data_t * kd, aes_gcm_counter_t * ctr, @@ -416,6 +754,95 @@ aes_gcm_enc (u8x16 T, aes_gcm_key_data_t * kd, aes_gcm_counter_t * ctr, if (n_left == 0) return T; +#if __VAES__ + u8x64 d4[4]; + if (n_left < 256) + { + f |= AES_GCM_F_LAST_ROUND; + if (n_left > 192) + { + n_left -= 192; + aes4_gcm_calc (T, kd, d4, ctr, inv, outv, rounds, 4, n_left, f); + return aes4_gcm_ghash_last (T, kd, d4, 4, n_left); + } + else if (n_left > 128) + { + n_left -= 128; + aes4_gcm_calc (T, kd, d4, ctr, inv, outv, rounds, 3, n_left, f); + return aes4_gcm_ghash_last (T, kd, d4, 3, n_left); + } + else if (n_left > 64) + { + n_left -= 64; + aes4_gcm_calc (T, kd, d4, ctr, inv, outv, rounds, 2, n_left, f); + return aes4_gcm_ghash_last (T, kd, d4, 2, n_left); + } + else + { + aes4_gcm_calc (T, kd, d4, ctr, inv, outv, rounds, 1, n_left, f); + return aes4_gcm_ghash_last (T, kd, d4, 1, n_left); + } + } + + aes4_gcm_calc (T, kd, d4, ctr, inv, outv, rounds, 4, 0, f); + + /* next */ + n_left -= 256; + outv += 16; + inv += 16; + + f |= AES_GCM_F_WITH_GHASH; + + while (n_left >= 512) + { + T = aes4_gcm_calc_double (T, kd, d4, ctr, inv, outv, rounds, f); + + /* next */ + n_left -= 512; + outv += 32; + inv += 32; + } + + while (n_left >= 256) + { + T = aes4_gcm_calc (T, kd, d4, ctr, inv, outv, rounds, 4, 0, f); + + /* next */ + n_left -= 256; + outv += 16; + inv += 16; + } + + if (n_left == 0) + return aes4_gcm_ghash_last (T, kd, d4, 4, 64); + + f |= AES_GCM_F_LAST_ROUND; + + if (n_left > 192) + { + n_left -= 192; + T = aes4_gcm_calc (T, kd, d4, ctr, inv, outv, rounds, 4, n_left, f); + return aes4_gcm_ghash_last (T, kd, d4, 4, n_left); + } + + if (n_left > 128) + { + n_left -= 128; + T = aes4_gcm_calc (T, kd, d4, ctr, inv, outv, rounds, 3, n_left, f); + return aes4_gcm_ghash_last (T, kd, d4, 3, n_left); + } + + if (n_left > 64) + { + n_left -= 64; + T = aes4_gcm_calc (T, kd, d4, ctr, inv, outv, rounds, 2, n_left, f); + return aes4_gcm_ghash_last (T, kd, d4, 2, n_left); + } + + T = aes4_gcm_calc (T, kd, d4, ctr, inv, outv, rounds, 1, n_left, f); + return aes4_gcm_ghash_last (T, kd, d4, 1, n_left); +#endif + if (n_left < 64) { f |= AES_GCM_F_LAST_ROUND; @@ -507,9 +934,47 @@ static_always_inline u8x16 aes_gcm_dec (u8x16 T, aes_gcm_key_data_t * kd, aes_gcm_counter_t * ctr, u8x16u * inv, u8x16u * outv, u32 n_left, int rounds) { - u8x16 d[8]; aes_gcm_flags_t f = AES_GCM_F_WITH_GHASH | AES_GCM_F_DECRYPT; +#ifdef __VAES__ + u8x64 d4[4] = { }; + + while (n_left >= 512) + { + T = aes4_gcm_calc_double (T, kd, d4, ctr, inv, outv, rounds, f); + + /* next */ + n_left -= 512; + outv += 32; + inv += 32; + } + + while (n_left >= 256) + { + T = aes4_gcm_calc (T, kd, d4, ctr, inv, outv, rounds, 4, 0, f); + + /* next */ + n_left -= 256; + outv += 16; + inv += 16; + } + + if (n_left == 0) + return T; + + f |= AES_GCM_F_LAST_ROUND; + if (n_left > 192) + return aes4_gcm_calc (T, kd, d4, ctr, inv, outv, rounds, 4, + n_left - 192, f); + if (n_left > 128) + return aes4_gcm_calc (T, kd, d4, ctr, inv, outv, rounds, 3, + n_left - 128, f); + if (n_left > 64) + return aes4_gcm_calc (T, kd, d4, ctr, inv, outv, rounds, 2, + n_left - 64, f); + return aes4_gcm_calc (T, kd, d4, ctr, inv, outv, rounds, 1, n_left, f); +#else + u8x16 d[4]; while (n_left >= 128) { T = aes_gcm_calc_double (T, kd, d, ctr, inv, outv, rounds, f); @@ -545,6 +1010,7 @@ aes_gcm_dec (u8x16 T, aes_gcm_key_data_t * kd, aes_gcm_counter_t * ctr, return aes_gcm_calc (T, kd, d, ctr, inv, outv, rounds, 2, n_left - 16, f); return aes_gcm_calc (T, kd, d, ctr, inv, outv, rounds, 1, n_left, f); +#endif } static_always_inline int @@ -573,7 +1039,11 @@ aes_gcm (u8x16u * in, u8x16u * out, u8x16u * addt, u8x16u * iv, u8x16u * tag, /* initalize counter */ ctr->counter = 1; Y0 = (u32x4) aes_load_partial (iv, 12) + ctr_inv_1; +#ifdef __VAES__ + ctr->Y4 = u32x16_splat_u32x4 (Y0) + ctr_inv_1234; +#else ctr->Y = Y0 + ctr_inv_1; +#endif /* ghash and encrypt/edcrypt */ if (is_encrypt) @@ -701,6 +1171,11 @@ aes_gcm_key_exp (vnet_crypto_key_t * key, aes_key_size_t ks) H = aes_encrypt_block (u8x16_splat (0), kd->Ke, ks); H = u8x16_reflect (H); ghash_precompute (H, (u8x16 *) kd->Hi, NUM_HI); +#ifdef __VAES__ + u8x64 *Ke4 = (u8x64 *) kd->Ke4; + for (int i = 0; i < AES_KEY_ROUNDS (ks) + 1; i++) + Ke4[i] = u8x64_splat_u8x16 (kd->Ke[i]); +#endif return kd; } diff --git a/src/vppinfra/vector_avx512.h b/src/vppinfra/vector_avx512.h index efdc78bdc02..0665a23fe77 100644 --- a/src/vppinfra/vector_avx512.h +++ b/src/vppinfra/vector_avx512.h @@ -171,6 +171,54 @@ u8x64_xor3 (u8x64 a, u8x64 b, u8x64 c) (__m512i) c, 0x96); } +static_always_inline u8x64 +u8x64_reflect_u8x16 (u8x64 x) +{ + static const u8x64 mask = { + 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, + 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, + 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, + 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, + }; + return (u8x64) _mm512_shuffle_epi8 ((__m512i) x, (__m512i) mask); +} + +static_always_inline u8x64 +u8x64_mask_load (u8x64 a, void *p, u64 mask) +{ + return (u8x64) _mm512_mask_loadu_epi8 ((__m512i) a, mask, p); +} + +static_always_inline void +u8x64_mask_store (u8x64 a, void *p, u64 mask) +{ + _mm512_mask_storeu_epi8 (p, mask, (__m512i) a); +} + +static_always_inline u8x64 +u8x64_splat_u8x16 (u8x16 a) +{ + return (u8x64) _mm512_broadcast_i64x2 ((__m128i) a); +} + +static_always_inline u32x16 +u32x16_splat_u32x4 (u32x4 a) +{ + return (u32x16) _mm512_broadcast_i64x2 ((__m128i) a); +} + +static_always_inline u32x16 +u32x16_mask_blend (u32x16 a, u32x16 b, u16 mask) +{ + return (u32x16) _mm512_mask_blend_epi32 (mask, (__m512i) a, (__m512i) b); +} + +static_always_inline u8x64 +u8x64_mask_blend (u8x64 a, u8x64 b, u64 mask) +{ + return (u8x64) _mm512_mask_blend_epi8 (mask, (__m512i) a, (__m512i) b); +} + static_always_inline void u32x16_transpose (u32x16 m[16]) { |