From e84e9d708e0a7571bc24ae2f443462ed8cddbb84 Mon Sep 17 00:00:00 2001 From: Damjan Marion Date: Thu, 13 Feb 2020 13:11:02 +0100 Subject: crypto-native: refactor CBC code Type: refactor Change-Id: I61e25942de318d03fb3d75689259709d687479bc Signed-off-by: Damjan Marion --- src/plugins/crypto_native/aes.h | 26 ++ src/plugins/crypto_native/aes_cbc.c | 431 ++++++++++++++++++++++++++++-- src/plugins/crypto_native/aes_cbc_aesni.h | 381 -------------------------- src/plugins/crypto_native/aes_cbc_neon.h | 208 -------------- src/vppinfra/vector_avx512.h | 6 + 5 files changed, 446 insertions(+), 606 deletions(-) delete mode 100644 src/plugins/crypto_native/aes_cbc_aesni.h delete mode 100644 src/plugins/crypto_native/aes_cbc_neon.h (limited to 'src') diff --git a/src/plugins/crypto_native/aes.h b/src/plugins/crypto_native/aes.h index b914088e94d..c2c22c6bd16 100644 --- a/src/plugins/crypto_native/aes.h +++ b/src/plugins/crypto_native/aes.h @@ -48,6 +48,32 @@ aes_enc_round (u8x16 a, u8x16 k) #endif } +#if defined (__VAES__) +static_always_inline u8x64 +aes_enc_round_x4 (u8x64 a, u8x64 k) +{ + return (u8x64) _mm512_aesenc_epi128 ((__m512i) a, (__m512i) k); +} + +static_always_inline u8x64 +aes_enc_last_round_x4 (u8x64 a, u8x64 k) +{ + return (u8x64) _mm512_aesenclast_epi128 ((__m512i) a, (__m512i) k); +} + +static_always_inline u8x64 +aes_dec_round_x4 (u8x64 a, u8x64 k) +{ + return (u8x64) _mm512_aesdec_epi128 ((__m512i) a, (__m512i) k); +} + +static_always_inline u8x64 +aes_dec_last_round_x4 (u8x64 a, u8x64 k) +{ + return (u8x64) _mm512_aesdeclast_epi128 ((__m512i) a, (__m512i) k); +} +#endif + static_always_inline u8x16 aes_enc_last_round (u8x16 a, u8x16 k) { diff --git a/src/plugins/crypto_native/aes_cbc.c b/src/plugins/crypto_native/aes_cbc.c index aeec1ea3186..e147ca66511 100644 --- a/src/plugins/crypto_native/aes_cbc.c +++ b/src/plugins/crypto_native/aes_cbc.c @@ -29,17 +29,414 @@ typedef struct { u8x16 encrypt_key[15]; #if __VAES__ - __m512i decrypt_key[15]; + u8x64 decrypt_key[15]; #else u8x16 decrypt_key[15]; #endif } aes_cbc_key_data_t; -#include -#include + +static_always_inline void __clib_unused +aes_cbc_dec (u8x16 * k, u8x16u * src, u8x16u * dst, u8x16u * iv, int count, + int rounds) +{ + u8x16 r[4], c[4], f; + + f = iv[0]; + while (count >= 64) + { + clib_prefetch_load (src + 8); + clib_prefetch_load (dst + 8); + + c[0] = r[0] = src[0]; + c[1] = r[1] = src[1]; + c[2] = r[2] = src[2]; + c[3] = r[3] = src[3]; + +#if __x86_64__ + r[0] ^= k[0]; + r[1] ^= k[0]; + r[2] ^= k[0]; + r[3] ^= k[0]; + + for (int i = 1; i < rounds; i++) + { + r[0] = aes_dec_round (r[0], k[i]); + r[1] = aes_dec_round (r[1], k[i]); + r[2] = aes_dec_round (r[2], k[i]); + r[3] = aes_dec_round (r[3], k[i]); + } + + r[0] = aes_dec_last_round (r[0], k[rounds]); + r[1] = aes_dec_last_round (r[1], k[rounds]); + r[2] = aes_dec_last_round (r[2], k[rounds]); + r[3] = aes_dec_last_round (r[3], k[rounds]); +#else + for (int i = 0; i < rounds - 1; i++) + { + r[0] = vaesimcq_u8 (vaesdq_u8 (r[0], k[i])); + r[1] = vaesimcq_u8 (vaesdq_u8 (r[1], k[i])); + r[2] = vaesimcq_u8 (vaesdq_u8 (r[2], k[i])); + r[3] = vaesimcq_u8 (vaesdq_u8 (r[3], k[i])); + } + r[0] = vaesdq_u8 (r[0], k[rounds - 1]) ^ k[rounds]; + r[1] = vaesdq_u8 (r[1], k[rounds - 1]) ^ k[rounds]; + r[2] = vaesdq_u8 (r[2], k[rounds - 1]) ^ k[rounds]; + r[3] = vaesdq_u8 (r[3], k[rounds - 1]) ^ k[rounds]; +#endif + dst[0] = r[0] ^ f; + dst[1] = r[1] ^ c[0]; + dst[2] = r[2] ^ c[1]; + dst[3] = r[3] ^ c[2]; + f = c[3]; + + count -= 64; + src += 4; + dst += 4; + } + + while (count > 0) + { + c[0] = r[0] = src[0]; +#if __x86_64__ + r[0] ^= k[0]; + for (int i = 1; i < rounds; i++) + r[0] = aes_dec_round (r[0], k[i]); + r[0] = aes_dec_last_round (r[0], k[rounds]); +#else + c[0] = r[0] = src[0]; + for (int i = 0; i < rounds - 1; i++) + r[0] = vaesimcq_u8 (vaesdq_u8 (r[0], k[i])); + r[0] = vaesdq_u8 (r[0], k[rounds - 1]) ^ k[rounds]; +#endif + dst[0] = r[0] ^ f; + f = c[0]; + + count -= 16; + src += 1; + dst += 1; + } +} + +#if __x86_64__ +#ifdef __VAES__ + +static_always_inline u8x64 +aes_block_load_x4 (u8 * src[], int i) +{ + u8x64 r = { }; + r = u8x64_insert_u8x16 (r, aes_block_load (src[0] + i), 0); + r = u8x64_insert_u8x16 (r, aes_block_load (src[1] + i), 1); + r = u8x64_insert_u8x16 (r, aes_block_load (src[2] + i), 2); + r = u8x64_insert_u8x16 (r, aes_block_load (src[3] + i), 3); + return r; +} + +static_always_inline void +aes_block_store_x4 (u8 * dst[], int i, u8x64 r) +{ + aes_block_store (dst[0] + i, u8x64_extract_u8x16 (r, 0)); + aes_block_store (dst[1] + i, u8x64_extract_u8x16 (r, 1)); + aes_block_store (dst[2] + i, u8x64_extract_u8x16 (r, 2)); + aes_block_store (dst[3] + i, u8x64_extract_u8x16 (r, 3)); +} + +static_always_inline u8x64 +aes_cbc_dec_permute (u8x64 a, u8x64 b) +{ + __m512i perm = { 6, 7, 8, 9, 10, 11, 12, 13 }; + return (u8x64) _mm512_permutex2var_epi64 ((__m512i) a, perm, (__m512i) b); +} + +static_always_inline void +vaes_cbc_dec (u8x64 * k, u8x64u * src, u8x64u * dst, u8x16 * iv, int count, + aes_key_size_t rounds) +{ + u8x64 f, r[4], c[4] = { }; + __mmask8 m; + int i, n_blocks = count >> 4; + + f = (u8x64) _mm512_mask_loadu_epi64 (_mm512_setzero_si512 (), 0xc0, + (__m512i *) (iv - 3)); + + while (n_blocks >= 16) + { + c[0] = src[0]; + c[1] = src[1]; + c[2] = src[2]; + c[3] = src[3]; + + r[0] = c[0] ^ k[0]; + r[1] = c[1] ^ k[0]; + r[2] = c[2] ^ k[0]; + r[3] = c[3] ^ k[0]; + + for (i = 1; i < rounds; i++) + { + r[0] = aes_dec_round_x4 (r[0], k[i]); + r[1] = aes_dec_round_x4 (r[1], k[i]); + r[2] = aes_dec_round_x4 (r[2], k[i]); + r[3] = aes_dec_round_x4 (r[3], k[i]); + } + + r[0] = aes_dec_last_round_x4 (r[0], k[i]); + r[1] = aes_dec_last_round_x4 (r[1], k[i]); + r[2] = aes_dec_last_round_x4 (r[2], k[i]); + r[3] = aes_dec_last_round_x4 (r[3], k[i]); + + dst[0] = r[0] ^= aes_cbc_dec_permute (f, c[0]); + dst[1] = r[1] ^= aes_cbc_dec_permute (c[0], c[1]); + dst[2] = r[2] ^= aes_cbc_dec_permute (c[1], c[2]); + dst[4] = r[3] ^= aes_cbc_dec_permute (c[2], c[3]); + f = c[3]; + + n_blocks -= 16; + src += 4; + dst += 4; + } + + while (n_blocks > 0) + { + m = (1 << (n_blocks * 2)) - 1; + c[0] = (u8x64) _mm512_mask_loadu_epi64 ((__m512i) c[0], m, + (__m512i *) src); + f = aes_cbc_dec_permute (f, c[0]); + r[0] = c[0] ^ k[0]; + for (i = 1; i < rounds; i++) + r[0] = aes_dec_round_x4 (r[0], k[i]); + r[0] = aes_dec_last_round_x4 (r[0], k[i]); + _mm512_mask_storeu_epi64 ((__m512i *) dst, m, (__m512i) (r[0] ^ f)); + f = c[0]; + n_blocks -= 4; + src += 1; + dst += 1; + } +} +#endif +#endif + +#ifdef __VAES__ +#define N 16 +#define u32xN u32x16 +#define u32xN_min_scalar u32x16_min_scalar +#define u32xN_is_all_zero u32x16_is_all_zero +#define u32xN_splat u32x16_splat +#else +#define N 4 +#define u32xN u32x4 +#define u32xN_min_scalar u32x4_min_scalar +#define u32xN_is_all_zero u32x4_is_all_zero +#define u32xN_splat u32x4_splat +#endif + +static_always_inline u32 +aes_ops_enc_aes_cbc (vlib_main_t * vm, vnet_crypto_op_t * ops[], + u32 n_ops, aes_key_size_t ks) +{ + crypto_native_main_t *cm = &crypto_native_main; + crypto_native_per_thread_data_t *ptd = + vec_elt_at_index (cm->per_thread_data, vm->thread_index); + int rounds = AES_KEY_ROUNDS (ks); + u8 dummy[8192]; + u32 i, j, count, n_left = n_ops; + u32xN dummy_mask = { }; + u32xN len = { }; + vnet_crypto_key_index_t key_index[N]; + u8 *src[N] = { }; + u8 *dst[N] = { }; +#if __VAES__ + u8x64 r[N / 4] = { }; + u8x64 k[15][N / 4] = { }; + u8x16 *kq, *rq = (u8x16 *) r; +#else + u8x16 r[N] = { }; + u8x16 k[15][N] = { }; +#endif + + for (i = 0; i < N; i++) + key_index[i] = ~0; + +more: + for (i = 0; i < N; i++) + if (len[i] == 0) + { + if (n_left == 0) + { + /* no more work to enqueue, so we are enqueueing dummy buffer */ + src[i] = dst[i] = dummy; + len[i] = sizeof (dummy); + dummy_mask[i] = 0; + } + else + { + u8x16 t; + if (ops[0]->flags & VNET_CRYPTO_OP_FLAG_INIT_IV) + { + t = ptd->cbc_iv[i]; + *(u8x16u *) ops[0]->iv = t; + ptd->cbc_iv[i] = aes_enc_round (t, t); + } + else + t = aes_block_load (ops[0]->iv); +#if __VAES__ + rq[i] = t; +#else + r[i] = t; +#endif + + src[i] = ops[0]->src; + dst[i] = ops[0]->dst; + len[i] = ops[0]->len; + dummy_mask[i] = ~0; + if (key_index[i] != ops[0]->key_index) + { + aes_cbc_key_data_t *kd; + key_index[i] = ops[0]->key_index; + kd = (aes_cbc_key_data_t *) cm->key_data[key_index[i]]; + for (j = 0; j < rounds + 1; j++) + { +#if __VAES__ + kq = (u8x16 *) k[j]; + kq[i] = kd->encrypt_key[j]; +#else + k[j][i] = kd->encrypt_key[j]; +#endif + } + } + ops[0]->status = VNET_CRYPTO_OP_STATUS_COMPLETED; + n_left--; + ops++; + } + } + + count = u32xN_min_scalar (len); + + ASSERT (count % 16 == 0); + + for (i = 0; i < count; i += 16) + { +#ifdef __VAES__ + r[0] = u8x64_xor3 (r[0], aes_block_load_x4 (src, i), k[0][0]); + r[1] = u8x64_xor3 (r[1], aes_block_load_x4 (src, i), k[0][1]); + r[2] = u8x64_xor3 (r[2], aes_block_load_x4 (src, i), k[0][2]); + r[3] = u8x64_xor3 (r[3], aes_block_load_x4 (src, i), k[0][3]); + + for (j = 1; j < rounds; j++) + { + r[0] = aes_enc_round_x4 (r[0], k[j][0]); + r[1] = aes_enc_round_x4 (r[1], k[j][1]); + r[2] = aes_enc_round_x4 (r[2], k[j][2]); + r[3] = aes_enc_round_x4 (r[3], k[j][3]); + } + r[0] = aes_enc_last_round_x4 (r[0], k[j][0]); + r[1] = aes_enc_last_round_x4 (r[1], k[j][1]); + r[2] = aes_enc_last_round_x4 (r[2], k[j][2]); + r[3] = aes_enc_last_round_x4 (r[3], k[j][3]); + + aes_block_store_x4 (dst, i, r[0]); + aes_block_store_x4 (dst + 4, i, r[1]); + aes_block_store_x4 (dst + 8, i, r[2]); + aes_block_store_x4 (dst + 12, i, r[3]); +#else +#if __x86_64__ + r[0] = u8x16_xor3 (r[0], aes_block_load (src[0] + i), k[0][0]); + r[1] = u8x16_xor3 (r[1], aes_block_load (src[1] + i), k[0][1]); + r[2] = u8x16_xor3 (r[2], aes_block_load (src[2] + i), k[0][2]); + r[3] = u8x16_xor3 (r[3], aes_block_load (src[3] + i), k[0][3]); + + for (j = 1; j < rounds; j++) + { + r[0] = aes_enc_round (r[0], k[j][0]); + r[1] = aes_enc_round (r[1], k[j][1]); + r[2] = aes_enc_round (r[2], k[j][2]); + r[3] = aes_enc_round (r[3], k[j][3]); + } + + r[0] = aes_enc_last_round (r[0], k[j][0]); + r[1] = aes_enc_last_round (r[1], k[j][1]); + r[2] = aes_enc_last_round (r[2], k[j][2]); + r[3] = aes_enc_last_round (r[3], k[j][3]); + + aes_block_store (dst[0] + i, r[0]); + aes_block_store (dst[1] + i, r[1]); + aes_block_store (dst[2] + i, r[2]); + aes_block_store (dst[3] + i, r[3]); +#else + r[0] ^= aes_block_load (src[0] + i); + r[1] ^= aes_block_load (src[1] + i); + r[2] ^= aes_block_load (src[2] + i); + r[3] ^= aes_block_load (src[3] + i); + for (j = 0; j < rounds - 1; j++) + { + r[0] = vaesmcq_u8 (vaeseq_u8 (r[0], k[j][0])); + r[1] = vaesmcq_u8 (vaeseq_u8 (r[1], k[j][1])); + r[2] = vaesmcq_u8 (vaeseq_u8 (r[2], k[j][2])); + r[3] = vaesmcq_u8 (vaeseq_u8 (r[3], k[j][3])); + } + r[0] = vaeseq_u8 (r[0], k[j][0]) ^ k[rounds][0]; + r[1] = vaeseq_u8 (r[1], k[j][1]) ^ k[rounds][1]; + r[2] = vaeseq_u8 (r[2], k[j][2]) ^ k[rounds][2]; + r[3] = vaeseq_u8 (r[3], k[j][3]) ^ k[rounds][3]; + aes_block_store (dst[0] + i, r[0]); + aes_block_store (dst[1] + i, r[1]); + aes_block_store (dst[2] + i, r[2]); + aes_block_store (dst[3] + i, r[3]); +#endif +#endif + } + + len -= u32xN_splat (count); + + for (i = 0; i < N; i++) + { + src[i] += count; + dst[i] += count; + } + + if (n_left > 0) + goto more; + + if (!u32xN_is_all_zero (len & dummy_mask)) + goto more; + + return n_ops; +} + + +static_always_inline u32 +aes_ops_dec_aes_cbc (vlib_main_t * vm, vnet_crypto_op_t * ops[], + u32 n_ops, aes_key_size_t ks) +{ + crypto_native_main_t *cm = &crypto_native_main; + int rounds = AES_KEY_ROUNDS (ks); + vnet_crypto_op_t *op = ops[0]; + aes_cbc_key_data_t *kd = (aes_cbc_key_data_t *) cm->key_data[op->key_index]; + u32 n_left = n_ops; + + ASSERT (n_ops >= 1); + +decrypt: +#ifdef __VAES__ + vaes_cbc_dec (kd->decrypt_key, (u8x64u *) op->src, (u8x64u *) op->dst, + (u8x16u *) op->iv, op->len, rounds); +#else + aes_cbc_dec (kd->decrypt_key, (u8x16u *) op->src, (u8x16u *) op->dst, + (u8x16u *) op->iv, op->len, rounds); +#endif + op->status = VNET_CRYPTO_OP_STATUS_COMPLETED; + + if (--n_left) + { + op += 1; + kd = (aes_cbc_key_data_t *) cm->key_data[op->key_index]; + goto decrypt; + } + + return n_ops; +} static_always_inline void * -aesni_cbc_key_exp (vnet_crypto_key_t * key, aes_key_size_t ks) +aes_cbc_key_exp (vnet_crypto_key_t * key, aes_key_size_t ks) { u8x16 e[15], d[15]; aes_cbc_key_data_t *kd; @@ -49,7 +446,7 @@ aesni_cbc_key_exp (vnet_crypto_key_t * key, aes_key_size_t ks) for (int i = 0; i < AES_KEY_ROUNDS (ks) + 1; i++) { #if __VAES__ - kd->decrypt_key[i] = _mm512_broadcast_i64x2 ((__m128i) d[i]); + kd->decrypt_key[i] = (u8x64) _mm512_broadcast_i64x2 ((__m128i) d[i]); #else kd->decrypt_key[i] = d[i]; #endif @@ -58,19 +455,19 @@ aesni_cbc_key_exp (vnet_crypto_key_t * key, aes_key_size_t ks) return kd; } -#define foreach_aesni_cbc_handler_type _(128) _(192) _(256) +#define foreach_aes_cbc_handler_type _(128) _(192) _(256) #define _(x) \ -static u32 aesni_ops_dec_aes_cbc_##x \ +static u32 aes_ops_dec_aes_cbc_##x \ (vlib_main_t * vm, vnet_crypto_op_t * ops[], u32 n_ops) \ -{ return aesni_ops_dec_aes_cbc (vm, ops, n_ops, AES_KEY_##x); } \ -static u32 aesni_ops_enc_aes_cbc_##x \ +{ return aes_ops_dec_aes_cbc (vm, ops, n_ops, AES_KEY_##x); } \ +static u32 aes_ops_enc_aes_cbc_##x \ (vlib_main_t * vm, vnet_crypto_op_t * ops[], u32 n_ops) \ -{ return aesni_ops_enc_aes_cbc (vm, ops, n_ops, AES_KEY_##x); } \ -static void * aesni_cbc_key_exp_##x (vnet_crypto_key_t *key) \ -{ return aesni_cbc_key_exp (key, AES_KEY_##x); } +{ return aes_ops_enc_aes_cbc (vm, ops, n_ops, AES_KEY_##x); } \ +static void * aes_cbc_key_exp_##x (vnet_crypto_key_t *key) \ +{ return aes_cbc_key_exp (key, AES_KEY_##x); } -foreach_aesni_cbc_handler_type; +foreach_aes_cbc_handler_type; #undef _ #include @@ -114,12 +511,12 @@ crypto_native_aes_cbc_init_sse42 (vlib_main_t * vm) #define _(x) \ vnet_crypto_register_ops_handler (vm, cm->crypto_engine_index, \ VNET_CRYPTO_OP_AES_##x##_CBC_ENC, \ - aesni_ops_enc_aes_cbc_##x); \ + aes_ops_enc_aes_cbc_##x); \ vnet_crypto_register_ops_handler (vm, cm->crypto_engine_index, \ VNET_CRYPTO_OP_AES_##x##_CBC_DEC, \ - aesni_ops_dec_aes_cbc_##x); \ - cm->key_fn[VNET_CRYPTO_ALG_AES_##x##_CBC] = aesni_cbc_key_exp_##x; - foreach_aesni_cbc_handler_type; + aes_ops_dec_aes_cbc_##x); \ + cm->key_fn[VNET_CRYPTO_ALG_AES_##x##_CBC] = aes_cbc_key_exp_##x; + foreach_aes_cbc_handler_type; #undef _ error: diff --git a/src/plugins/crypto_native/aes_cbc_aesni.h b/src/plugins/crypto_native/aes_cbc_aesni.h deleted file mode 100644 index 206dace5d56..00000000000 --- a/src/plugins/crypto_native/aes_cbc_aesni.h +++ /dev/null @@ -1,381 +0,0 @@ -/* - *------------------------------------------------------------------ - * Copyright (c) 2020 Cisco and/or its affiliates. - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at: - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - *------------------------------------------------------------------ - */ - -#ifdef __x86_64__ - -static_always_inline u8x16 __clib_unused -xor3 (u8x16 a, u8x16 b, u8x16 c) -{ -#if __AVX512F__ - return (u8x16) _mm_ternarylogic_epi32 ((__m128i) a, (__m128i) b, - (__m128i) c, 0x96); -#endif - return a ^ b ^ c; -} - -#if __VAES__ -static_always_inline __m512i -xor3_x4 (__m512i a, __m512i b, __m512i c) -{ - return _mm512_ternarylogic_epi32 (a, b, c, 0x96); -} - -static_always_inline __m512i -aes_block_load_x4 (u8 * src[], int i) -{ - __m512i r = { }; - r = _mm512_inserti64x2 (r, (__m128i) aes_block_load (src[0] + i), 0); - r = _mm512_inserti64x2 (r, (__m128i) aes_block_load (src[1] + i), 1); - r = _mm512_inserti64x2 (r, (__m128i) aes_block_load (src[2] + i), 2); - r = _mm512_inserti64x2 (r, (__m128i) aes_block_load (src[3] + i), 3); - return r; -} - -static_always_inline void -aes_block_store_x4 (u8 * dst[], int i, __m512i r) -{ - aes_block_store (dst[0] + i, (u8x16) _mm512_extracti64x2_epi64 (r, 0)); - aes_block_store (dst[1] + i, (u8x16) _mm512_extracti64x2_epi64 (r, 1)); - aes_block_store (dst[2] + i, (u8x16) _mm512_extracti64x2_epi64 (r, 2)); - aes_block_store (dst[3] + i, (u8x16) _mm512_extracti64x2_epi64 (r, 3)); -} -#endif - -static_always_inline void __clib_unused -aes_cbc_dec (u8x16 * k, u8 * src, u8 * dst, u8 * iv, int count, - aes_key_size_t rounds) -{ - u8x16 r0, r1, r2, r3, c0, c1, c2, c3, f; - int i; - - f = aes_block_load (iv); - - while (count >= 64) - { - _mm_prefetch (src + 128, _MM_HINT_T0); - _mm_prefetch (dst + 128, _MM_HINT_T0); - - c0 = aes_block_load (src); - c1 = aes_block_load (src + 16); - c2 = aes_block_load (src + 32); - c3 = aes_block_load (src + 48); - - r0 = c0 ^ k[0]; - r1 = c1 ^ k[0]; - r2 = c2 ^ k[0]; - r3 = c3 ^ k[0]; - - for (i = 1; i < rounds; i++) - { - r0 = aes_dec_round (r0, k[i]); - r1 = aes_dec_round (r1, k[i]); - r2 = aes_dec_round (r2, k[i]); - r3 = aes_dec_round (r3, k[i]); - } - - r0 = aes_dec_last_round (r0, k[i]); - r1 = aes_dec_last_round (r1, k[i]); - r2 = aes_dec_last_round (r2, k[i]); - r3 = aes_dec_last_round (r3, k[i]); - - aes_block_store (dst, r0 ^ f); - aes_block_store (dst + 16, r1 ^ c0); - aes_block_store (dst + 32, r2 ^ c1); - aes_block_store (dst + 48, r3 ^ c2); - - f = c3; - - count -= 64; - src += 64; - dst += 64; - } - - while (count > 0) - { - c0 = aes_block_load (src); - r0 = c0 ^ k[0]; - for (i = 1; i < rounds; i++) - r0 = aes_dec_round (r0, k[i]); - r0 = aes_dec_last_round (r0, k[i]); - aes_block_store (dst, r0 ^ f); - f = c0; - count -= 16; - src += 16; - dst += 16; - } -} - -#ifdef __VAES__ -static_always_inline void -vaes_cbc_dec (__m512i * k, u8 * src, u8 * dst, u8 * iv, int count, - aes_key_size_t rounds) -{ - __m512i permute = { 6, 7, 8, 9, 10, 11, 12, 13 }; - __m512i r0, r1, r2, r3, c0, c1, c2, c3, f = { }; - __mmask8 m; - int i, n_blocks = count >> 4; - - f = _mm512_mask_loadu_epi64 (f, 0xc0, (__m512i *) (iv - 48)); - - while (n_blocks >= 16) - { - c0 = _mm512_loadu_si512 ((__m512i *) src); - c1 = _mm512_loadu_si512 ((__m512i *) (src + 64)); - c2 = _mm512_loadu_si512 ((__m512i *) (src + 128)); - c3 = _mm512_loadu_si512 ((__m512i *) (src + 192)); - - r0 = c0 ^ k[0]; - r1 = c1 ^ k[0]; - r2 = c2 ^ k[0]; - r3 = c3 ^ k[0]; - - for (i = 1; i < rounds; i++) - { - r0 = _mm512_aesdec_epi128 (r0, k[i]); - r1 = _mm512_aesdec_epi128 (r1, k[i]); - r2 = _mm512_aesdec_epi128 (r2, k[i]); - r3 = _mm512_aesdec_epi128 (r3, k[i]); - } - - r0 = _mm512_aesdeclast_epi128 (r0, k[i]); - r1 = _mm512_aesdeclast_epi128 (r1, k[i]); - r2 = _mm512_aesdeclast_epi128 (r2, k[i]); - r3 = _mm512_aesdeclast_epi128 (r3, k[i]); - - r0 ^= _mm512_permutex2var_epi64 (f, permute, c0); - _mm512_storeu_si512 ((__m512i *) dst, r0); - - r1 ^= _mm512_permutex2var_epi64 (c0, permute, c1); - _mm512_storeu_si512 ((__m512i *) (dst + 64), r1); - - r2 ^= _mm512_permutex2var_epi64 (c1, permute, c2); - _mm512_storeu_si512 ((__m512i *) (dst + 128), r2); - - r3 ^= _mm512_permutex2var_epi64 (c2, permute, c3); - _mm512_storeu_si512 ((__m512i *) (dst + 192), r3); - f = c3; - - n_blocks -= 16; - src += 256; - dst += 256; - } - - while (n_blocks > 0) - { - m = (1 << (n_blocks * 2)) - 1; - c0 = _mm512_mask_loadu_epi64 (c0, m, (__m512i *) src); - f = _mm512_permutex2var_epi64 (f, permute, c0); - r0 = c0 ^ k[0]; - for (i = 1; i < rounds; i++) - r0 = _mm512_aesdec_epi128 (r0, k[i]); - r0 = _mm512_aesdeclast_epi128 (r0, k[i]); - _mm512_mask_storeu_epi64 ((__m512i *) dst, m, r0 ^ f); - f = c0; - n_blocks -= 4; - src += 64; - dst += 64; - } -} -#endif - -#ifdef __VAES__ -#define N 16 -#define u32xN u32x16 -#define u32xN_min_scalar u32x16_min_scalar -#define u32xN_is_all_zero u32x16_is_all_zero -#else -#define N 4 -#define u32xN u32x4 -#define u32xN_min_scalar u32x4_min_scalar -#define u32xN_is_all_zero u32x4_is_all_zero -#endif - -static_always_inline u32 -aesni_ops_enc_aes_cbc (vlib_main_t * vm, vnet_crypto_op_t * ops[], - u32 n_ops, aes_key_size_t ks) -{ - crypto_native_main_t *cm = &crypto_native_main; - crypto_native_per_thread_data_t *ptd = - vec_elt_at_index (cm->per_thread_data, vm->thread_index); - int rounds = AES_KEY_ROUNDS (ks); - u8 dummy[8192]; - u32 i, j, count, n_left = n_ops; - u32xN dummy_mask = { }; - u32xN len = { }; - vnet_crypto_key_index_t key_index[N]; - u8 *src[N] = { }; - u8 *dst[N] = { }; - /* *INDENT-OFF* */ - union - { - u8x16 x1[N]; - __m512i x4[N / 4]; - } r = { }, k[15] = { }; - /* *INDENT-ON* */ - - for (i = 0; i < N; i++) - key_index[i] = ~0; - -more: - for (i = 0; i < N; i++) - if (len[i] == 0) - { - if (n_left == 0) - { - /* no more work to enqueue, so we are enqueueing dummy buffer */ - src[i] = dst[i] = dummy; - len[i] = sizeof (dummy); - dummy_mask[i] = 0; - } - else - { - if (ops[0]->flags & VNET_CRYPTO_OP_FLAG_INIT_IV) - { - r.x1[i] = ptd->cbc_iv[i]; - aes_block_store (ops[0]->iv, r.x1[i]); - ptd->cbc_iv[i] = aes_enc_round (r.x1[i], r.x1[i]); - } - else - r.x1[i] = aes_block_load (ops[0]->iv); - - src[i] = ops[0]->src; - dst[i] = ops[0]->dst; - len[i] = ops[0]->len; - dummy_mask[i] = ~0; - if (key_index[i] != ops[0]->key_index) - { - aes_cbc_key_data_t *kd; - key_index[i] = ops[0]->key_index; - kd = (aes_cbc_key_data_t *) cm->key_data[key_index[i]]; - for (j = 0; j < rounds + 1; j++) - k[j].x1[i] = kd->encrypt_key[j]; - } - ops[0]->status = VNET_CRYPTO_OP_STATUS_COMPLETED; - n_left--; - ops++; - } - } - - count = u32xN_min_scalar (len); - - ASSERT (count % 16 == 0); - - for (i = 0; i < count; i += 16) - { -#ifdef __VAES__ - r.x4[0] = xor3_x4 (r.x4[0], aes_block_load_x4 (src, i), k[0].x4[0]); - r.x4[1] = xor3_x4 (r.x4[1], aes_block_load_x4 (src, i), k[0].x4[1]); - r.x4[2] = xor3_x4 (r.x4[2], aes_block_load_x4 (src, i), k[0].x4[2]); - r.x4[3] = xor3_x4 (r.x4[3], aes_block_load_x4 (src, i), k[0].x4[3]); - - for (j = 1; j < rounds; j++) - { - r.x4[0] = _mm512_aesenc_epi128 (r.x4[0], k[j].x4[0]); - r.x4[1] = _mm512_aesenc_epi128 (r.x4[1], k[j].x4[1]); - r.x4[2] = _mm512_aesenc_epi128 (r.x4[2], k[j].x4[2]); - r.x4[3] = _mm512_aesenc_epi128 (r.x4[3], k[j].x4[3]); - } - r.x4[0] = _mm512_aesenclast_epi128 (r.x4[0], k[j].x4[0]); - r.x4[1] = _mm512_aesenclast_epi128 (r.x4[1], k[j].x4[1]); - r.x4[2] = _mm512_aesenclast_epi128 (r.x4[2], k[j].x4[2]); - r.x4[3] = _mm512_aesenclast_epi128 (r.x4[3], k[j].x4[3]); - - aes_block_store_x4 (dst, i, r.x4[0]); - aes_block_store_x4 (dst + 4, i, r.x4[1]); - aes_block_store_x4 (dst + 8, i, r.x4[2]); - aes_block_store_x4 (dst + 12, i, r.x4[3]); -#else - r.x1[0] = xor3 (r.x1[0], aes_block_load (src[0] + i), k[0].x1[0]); - r.x1[1] = xor3 (r.x1[1], aes_block_load (src[1] + i), k[0].x1[1]); - r.x1[2] = xor3 (r.x1[2], aes_block_load (src[2] + i), k[0].x1[2]); - r.x1[3] = xor3 (r.x1[3], aes_block_load (src[3] + i), k[0].x1[3]); - - for (j = 1; j < rounds; j++) - { - r.x1[0] = aes_enc_round (r.x1[0], k[j].x1[0]); - r.x1[1] = aes_enc_round (r.x1[1], k[j].x1[1]); - r.x1[2] = aes_enc_round (r.x1[2], k[j].x1[2]); - r.x1[3] = aes_enc_round (r.x1[3], k[j].x1[3]); - } - - r.x1[0] = aes_enc_last_round (r.x1[0], k[j].x1[0]); - r.x1[1] = aes_enc_last_round (r.x1[1], k[j].x1[1]); - r.x1[2] = aes_enc_last_round (r.x1[2], k[j].x1[2]); - r.x1[3] = aes_enc_last_round (r.x1[3], k[j].x1[3]); - - aes_block_store (dst[0] + i, r.x1[0]); - aes_block_store (dst[1] + i, r.x1[1]); - aes_block_store (dst[2] + i, r.x1[2]); - aes_block_store (dst[3] + i, r.x1[3]); -#endif - } - - for (i = 0; i < N; i++) - { - src[i] += count; - dst[i] += count; - len[i] -= count; - } - - if (n_left > 0) - goto more; - - if (!u32xN_is_all_zero (len & dummy_mask)) - goto more; - - return n_ops; -} - -static_always_inline u32 -aesni_ops_dec_aes_cbc (vlib_main_t * vm, vnet_crypto_op_t * ops[], - u32 n_ops, aes_key_size_t ks) -{ - crypto_native_main_t *cm = &crypto_native_main; - int rounds = AES_KEY_ROUNDS (ks); - vnet_crypto_op_t *op = ops[0]; - aes_cbc_key_data_t *kd = (aes_cbc_key_data_t *) cm->key_data[op->key_index]; - u32 n_left = n_ops; - - ASSERT (n_ops >= 1); - -decrypt: -#ifdef __VAES__ - vaes_cbc_dec (kd->decrypt_key, op->src, op->dst, op->iv, op->len, rounds); -#else - aes_cbc_dec (kd->decrypt_key, op->src, op->dst, op->iv, op->len, rounds); -#endif - op->status = VNET_CRYPTO_OP_STATUS_COMPLETED; - - if (--n_left) - { - op += 1; - kd = (aes_cbc_key_data_t *) cm->key_data[op->key_index]; - goto decrypt; - } - - return n_ops; -} - -#endif - -/* - * fd.io coding-style-patch-verification: ON - * - * Local Variables: - * eval: (c-set-style "gnu") - * End: - */ diff --git a/src/plugins/crypto_native/aes_cbc_neon.h b/src/plugins/crypto_native/aes_cbc_neon.h deleted file mode 100644 index 4a9c3cd3986..00000000000 --- a/src/plugins/crypto_native/aes_cbc_neon.h +++ /dev/null @@ -1,208 +0,0 @@ -/* - *------------------------------------------------------------------ - * Copyright (c) 2020 Cisco and/or its affiliates. - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at: - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - *------------------------------------------------------------------ - */ - -#ifdef __aarch64__ - -static_always_inline void -aes_cbc_dec (u8x16 * k, u8 * src, u8 * dst, u8 * iv, int count, int rounds) -{ - u8x16 r0, r1, r2, r3, c0, c1, c2, c3, f; - - f = vld1q_u8 (iv); - - while (count >= 64) - { - c0 = r0 = vld1q_u8 (src); - c1 = r1 = vld1q_u8 (src + 16); - c2 = r2 = vld1q_u8 (src + 32); - c3 = r3 = vld1q_u8 (src + 48); - for (int i = 0; i < rounds - 1; i++) - { - r0 = vaesimcq_u8 (vaesdq_u8 (r0, k[i])); - r1 = vaesimcq_u8 (vaesdq_u8 (r1, k[i])); - r2 = vaesimcq_u8 (vaesdq_u8 (r2, k[i])); - r3 = vaesimcq_u8 (vaesdq_u8 (r3, k[i])); - } - r0 = vaesdq_u8 (r0, k[rounds - 1]) ^ k[rounds]; - r1 = vaesdq_u8 (r1, k[rounds - 1]) ^ k[rounds]; - r2 = vaesdq_u8 (r2, k[rounds - 1]) ^ k[rounds]; - r3 = vaesdq_u8 (r3, k[rounds - 1]) ^ k[rounds]; - vst1q_u8 (dst, r0 ^ f); - vst1q_u8 (dst + 16, r1 ^ c0); - vst1q_u8 (dst + 32, r2 ^ c1); - vst1q_u8 (dst + 48, r3 ^ c2); - f = c3; - - src += 64; - dst += 64; - count -= 64; - } - - while (count >= 16) - { - c0 = r0 = vld1q_u8 (src); - for (int i = 0; i < rounds - 1; i++) - r0 = vaesimcq_u8 (vaesdq_u8 (r0, k[i])); - r0 = vaesdq_u8 (r0, k[rounds - 1]) ^ k[rounds]; - vst1q_u8 (dst, r0 ^ f); - f = c0; - - src += 16; - dst += 16; - count -= 16; - } -} - -static_always_inline u32 -aesni_ops_enc_aes_cbc (vlib_main_t * vm, vnet_crypto_op_t * ops[], - u32 n_ops, aes_key_size_t ks) -{ - crypto_native_main_t *cm = &crypto_native_main; - crypto_native_per_thread_data_t *ptd = - vec_elt_at_index (cm->per_thread_data, vm->thread_index); - int rounds = AES_KEY_ROUNDS (ks); - u8 dummy[8192]; - u32 i, j, count, n_left = n_ops; - u32x4 dummy_mask = { }; - u32x4 len = { }; - vnet_crypto_key_index_t key_index[4]; - u8 *src[4] = { }; - u8 *dst[4] = { }; - u8x16 r[4] = { }; - u8x16 k[15][4] = { }; - - for (i = 0; i < 4; i++) - key_index[i] = ~0; - -more: - for (i = 0; i < 4; i++) - if (len[i] == 0) - { - if (n_left == 0) - { - /* no more work to enqueue, so we are enqueueing dummy buffer */ - src[i] = dst[i] = dummy; - len[i] = sizeof (dummy); - dummy_mask[i] = 0; - } - else - { - if (ops[0]->flags & VNET_CRYPTO_OP_FLAG_INIT_IV) - { - r[i] = ptd->cbc_iv[i]; - vst1q_u8 (ops[0]->iv, r[i]); - ptd->cbc_iv[i] = vaeseq_u8 (r[i], r[i]); - } - else - r[i] = vld1q_u8 (ops[0]->iv); - - src[i] = ops[0]->src; - dst[i] = ops[0]->dst; - len[i] = ops[0]->len; - dummy_mask[i] = ~0; - if (key_index[i] != ops[0]->key_index) - { - aes_cbc_key_data_t *kd; - key_index[i] = ops[0]->key_index; - kd = (aes_cbc_key_data_t *) cm->key_data[key_index[i]]; - for (j = 0; j < rounds + 1; j++) - k[j][i] = kd->encrypt_key[j]; - } - ops[0]->status = VNET_CRYPTO_OP_STATUS_COMPLETED; - n_left--; - ops++; - } - } - - count = u32x4_min_scalar (len); - - ASSERT (count % 16 == 0); - - for (i = 0; i < count; i += 16) - { - r[0] ^= vld1q_u8 (src[0] + i); - r[1] ^= vld1q_u8 (src[1] + i); - r[2] ^= vld1q_u8 (src[2] + i); - r[3] ^= vld1q_u8 (src[3] + i); - for (j = 0; j < rounds - 1; j++) - { - r[0] = vaesmcq_u8 (vaeseq_u8 (r[0], k[j][0])); - r[1] = vaesmcq_u8 (vaeseq_u8 (r[1], k[j][1])); - r[2] = vaesmcq_u8 (vaeseq_u8 (r[2], k[j][2])); - r[3] = vaesmcq_u8 (vaeseq_u8 (r[3], k[j][3])); - } - r[0] = vaeseq_u8 (r[0], k[j][0]) ^ k[rounds][0]; - r[1] = vaeseq_u8 (r[1], k[j][1]) ^ k[rounds][1]; - r[2] = vaeseq_u8 (r[2], k[j][2]) ^ k[rounds][2]; - r[3] = vaeseq_u8 (r[3], k[j][3]) ^ k[rounds][3]; - vst1q_u8 (dst[0] + i, r[0]); - vst1q_u8 (dst[1] + i, r[1]); - vst1q_u8 (dst[2] + i, r[2]); - vst1q_u8 (dst[3] + i, r[3]); - } - - for (i = 0; i < 4; i++) - { - src[i] += count; - dst[i] += count; - len[i] -= count; - } - - if (n_left > 0) - goto more; - - if (!u32x4_is_all_zero (len & dummy_mask)) - goto more; - - return n_ops; -} - -static_always_inline u32 -aesni_ops_dec_aes_cbc (vlib_main_t * vm, vnet_crypto_op_t * ops[], - u32 n_ops, aes_key_size_t ks) -{ - crypto_native_main_t *cm = &crypto_native_main; - int rounds = AES_KEY_ROUNDS (ks); - vnet_crypto_op_t *op = ops[0]; - aes_cbc_key_data_t *kd = (aes_cbc_key_data_t *) cm->key_data[op->key_index]; - u32 n_left = n_ops; - - ASSERT (n_ops >= 1); - -decrypt: - aes_cbc_dec (kd->decrypt_key, op->src, op->dst, op->iv, op->len, rounds); - op->status = VNET_CRYPTO_OP_STATUS_COMPLETED; - - if (--n_left) - { - op += 1; - kd = (aes_cbc_key_data_t *) cm->key_data[op->key_index]; - goto decrypt; - } - - return n_ops; -} - -#endif - -/* - * fd.io coding-style-patch-verification: ON - * - * Local Variables: - * eval: (c-set-style "gnu") - * End: - */ diff --git a/src/vppinfra/vector_avx512.h b/src/vppinfra/vector_avx512.h index 29d96f85ce9..e27c7c041e4 100644 --- a/src/vppinfra/vector_avx512.h +++ b/src/vppinfra/vector_avx512.h @@ -143,6 +143,12 @@ u64x8_permute (u64x8 a, u64x8 b, u64x8 mask) #define u32x16_ternary_logic(a, b, c, d) \ (u32x16) _mm512_ternarylogic_epi32 ((__m512i) a, (__m512i) b, (__m512i) c, d) +#define u8x64_insert_u8x16(a, b, n) \ + (u8x64) _mm512_inserti64x2 ((__m512i) (a), (__m128i) (b), n) + +#define u8x64_extract_u8x16(a, n) \ + (u8x16) _mm512_extracti64x2_epi64 ((__m512i) (a), n) + static_always_inline u8x64 u8x64_xor3 (u8x64 a, u8x64 b, u8x64 c) { -- cgit 1.2.3-korg