diff options
Diffstat (limited to 'src/plugins/crypto_native/aes_cbc.c')
| -rw-r--r-- | src/plugins/crypto_native/aes_cbc.c | 305 |
1 files changed, 81 insertions, 224 deletions
diff --git a/src/plugins/crypto_native/aes_cbc.c b/src/plugins/crypto_native/aes_cbc.c index c84390c3108..c981897783f 100644 --- a/src/plugins/crypto_native/aes_cbc.c +++ b/src/plugins/crypto_native/aes_cbc.c @@ -25,191 +25,40 @@ #pragma GCC optimize ("O3") #endif -#if defined(__VAES__) && defined(__AVX512F__) -#define u8xN u8x64 -#define u32xN u32x16 -#define u32xN_min_scalar u32x16_min_scalar -#define u32xN_is_all_zero u32x16_is_all_zero -#define u32xN_splat u32x16_splat -#elif defined(__VAES__) -#define u8xN u8x32 -#define u32xN u32x8 -#define u32xN_min_scalar u32x8_min_scalar -#define u32xN_is_all_zero u32x8_is_all_zero -#define u32xN_splat u32x8_splat -#else -#define u8xN u8x16 -#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 +#define CRYPTO_NATIVE_AES_CBC_ENC_VEC_SIZE 256 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; - int rounds = AES_KEY_ROUNDS (ks); - u8 placeholder[8192]; - u32 i, j, count, n_left = n_ops; - u32xN placeholder_mask = { }; - u32xN len = { }; - vnet_crypto_key_index_t key_index[4 * N_AES_LANES]; - u8 *src[4 * N_AES_LANES] = {}; - u8 *dst[4 * N_AES_LANES] = {}; - u8xN r[4] = {}; - u8xN k[15][4] = {}; - - for (i = 0; i < 4 * N_AES_LANES; i++) - key_index[i] = ~0; - -more: - for (i = 0; i < 4 * N_AES_LANES; i++) - if (len[i] == 0) - { - if (n_left == 0) - { - /* no more work to enqueue, so we are enqueueing placeholder buffer */ - src[i] = dst[i] = placeholder; - len[i] = sizeof (placeholder); - placeholder_mask[i] = 0; - } - else - { - u8x16 t = aes_block_load (ops[0]->iv); - ((u8x16 *) r)[i] = t; - - src[i] = ops[0]->src; - dst[i] = ops[0]->dst; - len[i] = ops[0]->len; - placeholder_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++) - ((u8x16 *) k[j])[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) + u32 i, n_left = n_ops; + uword key_indices[CRYPTO_NATIVE_AES_CBC_ENC_VEC_SIZE] = {}; + u8 *plaintext[CRYPTO_NATIVE_AES_CBC_ENC_VEC_SIZE] = {}; + uword oplen[CRYPTO_NATIVE_AES_CBC_ENC_VEC_SIZE] = {}; + u8 *iv[CRYPTO_NATIVE_AES_CBC_ENC_VEC_SIZE] = {}; + u8 *ciphertext[CRYPTO_NATIVE_AES_CBC_ENC_VEC_SIZE] = {}; + + while (n_left) { -#if defined(__VAES__) && defined(__AVX512F__) - 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 + 4, i), k[0][1]); - r[2] = u8x64_xor3 (r[2], aes_block_load_x4 (src + 8, i), k[0][2]); - r[3] = u8x64_xor3 (r[3], aes_block_load_x4 (src + 12, i), k[0][3]); - - for (j = 1; j < rounds; j++) + i = 0; + while (n_left && i < CRYPTO_NATIVE_AES_CBC_ENC_VEC_SIZE) { - 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]); + key_indices[i] = ops[0]->key_index; + plaintext[i] = ops[0]->src; + ciphertext[i] = ops[0]->dst; + oplen[i] = ops[0]->len; + iv[i] = ops[0]->iv; + ops[0]->status = VNET_CRYPTO_OP_STATUS_COMPLETED; + + ops++; + n_left--; + i++; } - 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]); -#elif defined(__VAES__) - r[0] = u8x32_xor3 (r[0], aes_block_load_x2 (src, i), k[0][0]); - r[1] = u8x32_xor3 (r[1], aes_block_load_x2 (src + 2, i), k[0][1]); - r[2] = u8x32_xor3 (r[2], aes_block_load_x2 (src + 4, i), k[0][2]); - r[3] = u8x32_xor3 (r[3], aes_block_load_x2 (src + 6, i), k[0][3]); - - for (j = 1; j < rounds; j++) - { - r[0] = aes_enc_round_x2 (r[0], k[j][0]); - r[1] = aes_enc_round_x2 (r[1], k[j][1]); - r[2] = aes_enc_round_x2 (r[2], k[j][2]); - r[3] = aes_enc_round_x2 (r[3], k[j][3]); - } - r[0] = aes_enc_last_round_x2 (r[0], k[j][0]); - r[1] = aes_enc_last_round_x2 (r[1], k[j][1]); - r[2] = aes_enc_last_round_x2 (r[2], k[j][2]); - r[3] = aes_enc_last_round_x2 (r[3], k[j][3]); - - aes_block_store_x2 (dst, i, r[0]); - aes_block_store_x2 (dst + 2, i, r[1]); - aes_block_store_x2 (dst + 4, i, r[2]); - aes_block_store_x2 (dst + 6, 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_x1 (r[0], k[j][0]); - r[1] = aes_enc_round_x1 (r[1], k[j][1]); - r[2] = aes_enc_round_x1 (r[2], k[j][2]); - r[3] = aes_enc_round_x1 (r[3], k[j][3]); - } - - r[0] = aes_enc_last_round_x1 (r[0], k[j][0]); - r[1] = aes_enc_last_round_x1 (r[1], k[j][1]); - r[2] = aes_enc_last_round_x1 (r[2], k[j][2]); - r[3] = aes_enc_last_round_x1 (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 + clib_aes_cbc_encrypt_multi ((aes_cbc_key_data_t **) cm->key_data, + key_indices, plaintext, oplen, iv, ks, + ciphertext, i); } - - len -= u32xN_splat (count); - - for (i = 0; i < 4 * N_AES_LANES; i++) - { - src[i] += count; - dst[i] += count; - } - - if (n_left > 0) - goto more; - - if (!u32xN_is_all_zero (len & placeholder_mask)) - goto more; - return n_ops; } @@ -249,18 +98,30 @@ decrypt: return n_ops; } -#define foreach_aes_cbc_handler_type _(128) _(192) _(256) - -#define _(x) \ -static u32 aes_ops_dec_aes_cbc_##x \ -(vlib_main_t * vm, vnet_crypto_op_t * ops[], u32 n_ops) \ -{ 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 aes_ops_enc_aes_cbc (vm, ops, n_ops, AES_KEY_##x); } \ - -foreach_aes_cbc_handler_type; -#undef _ +static int +aes_cbc_cpu_probe () +{ +#if defined(__VAES__) && defined(__AVX512F__) + if (clib_cpu_supports_vaes () && clib_cpu_supports_avx512f ()) + return 50; +#elif defined(__VAES__) + if (clib_cpu_supports_vaes ()) + return 40; +#elif defined(__AVX512F__) + if (clib_cpu_supports_avx512f ()) + return 30; +#elif defined(__AVX2__) + if (clib_cpu_supports_avx2 ()) + return 20; +#elif __AES__ + if (clib_cpu_supports_aes ()) + return 10; +#elif __aarch64__ + if (clib_cpu_supports_aarch64_aes ()) + return 10; +#endif + return -1; +} static void * aes_cbc_key_exp_128 (vnet_crypto_key_t *key) @@ -289,43 +150,39 @@ aes_cbc_key_exp_256 (vnet_crypto_key_t *key) return kd; } -#include <fcntl.h> +#define foreach_aes_cbc_handler_type _ (128) _ (192) _ (256) + +#define _(x) \ + static u32 aes_ops_enc_aes_cbc_##x (vlib_main_t *vm, \ + vnet_crypto_op_t *ops[], u32 n_ops) \ + { \ + return aes_ops_enc_aes_cbc (vm, ops, n_ops, AES_KEY_##x); \ + } \ + \ + CRYPTO_NATIVE_OP_HANDLER (aes_##x##_cbc_enc) = { \ + .op_id = VNET_CRYPTO_OP_AES_##x##_CBC_ENC, \ + .fn = aes_ops_enc_aes_cbc_##x, \ + .probe = aes_cbc_cpu_probe, \ + }; \ + \ + static u32 aes_ops_dec_aes_cbc_##x (vlib_main_t *vm, \ + vnet_crypto_op_t *ops[], u32 n_ops) \ + { \ + return aes_ops_dec_aes_cbc (vm, ops, n_ops, AES_KEY_##x); \ + } \ + \ + CRYPTO_NATIVE_OP_HANDLER (aes_##x##_cbc_dec) = { \ + .op_id = VNET_CRYPTO_OP_AES_##x##_CBC_DEC, \ + .fn = aes_ops_dec_aes_cbc_##x, \ + .probe = aes_cbc_cpu_probe, \ + }; \ + \ + CRYPTO_NATIVE_KEY_HANDLER (aes_##x##_cbc) = { \ + .alg_id = VNET_CRYPTO_ALG_AES_##x##_CBC, \ + .key_fn = aes_cbc_key_exp_##x, \ + .probe = aes_cbc_cpu_probe, \ + }; -clib_error_t * -#if defined(__VAES__) && defined(__AVX512F__) -crypto_native_aes_cbc_init_icl (vlib_main_t *vm) -#elif defined(__VAES__) -crypto_native_aes_cbc_init_adl (vlib_main_t *vm) -#elif __AVX512F__ -crypto_native_aes_cbc_init_skx (vlib_main_t * vm) -#elif __aarch64__ -crypto_native_aes_cbc_init_neon (vlib_main_t * vm) -#elif __AVX2__ -crypto_native_aes_cbc_init_hsw (vlib_main_t * vm) -#else -crypto_native_aes_cbc_init_slm (vlib_main_t * vm) -#endif -{ - crypto_native_main_t *cm = &crypto_native_main; - -#define _(x) \ - vnet_crypto_register_ops_handler (vm, cm->crypto_engine_index, \ - VNET_CRYPTO_OP_AES_##x##_CBC_ENC, \ - aes_ops_enc_aes_cbc_##x); \ - vnet_crypto_register_ops_handler (vm, cm->crypto_engine_index, \ - VNET_CRYPTO_OP_AES_##x##_CBC_DEC, \ - 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; +foreach_aes_cbc_handler_type; #undef _ - return 0; -} - -/* - * fd.io coding-style-patch-verification: ON - * - * Local Variables: - * eval: (c-set-style "gnu") - * End: - */ |