crypto-native: add AArch64 AES-GCM native implementation

Type: feature

Change-Id: I4f96b0af13b875d491704b010328a1814e1dbda1
Signed-off-by: Damjan Marion <dmarion@me.com>

7 files changed, 78 insertions, 55 deletions

diff --git a/src/plugins/crypto_native/CMakeLists.txt b/src/plugins/crypto_native/CMakeLists.txt
index 9ac29462815..c6d916ddbbd 100644
--- a/src/plugins/crypto_native/CMakeLists.txt
+++ b/src/plugins/crypto_native/CMakeLists.txt
@@ -26,7 +26,7 @@ endif()
 
 if(CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64.*|AARCH64.*)")
   list(APPEND VARIANTS "armv8\;-march=armv8.1-a+crc+crypto")
-  set (COMPILE_FILES aes_cbc.c)
+  set (COMPILE_FILES aes_cbc.c aes_gcm.c)
   set (COMPILE_OPTS -Wall -fno-common)
 endif()
 
diff --git a/src/plugins/crypto_native/aes.h b/src/plugins/crypto_native/aes.h
index aacbf8ae34d..b914088e94d 100644
--- a/src/plugins/crypto_native/aes.h
+++ b/src/plugins/crypto_native/aes.h
@@ -28,8 +28,6 @@ typedef enum
 #define AES_KEY_ROUNDS(x)		(10 + x * 2)
 #define AES_KEY_BYTES(x)		(16 + x * 8)
 
-#ifdef __x86_64__
-
 static const u8x16 byte_mask_scale = {
   0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
 };
@@ -37,21 +35,31 @@ static const u8x16 byte_mask_scale = {
 static_always_inline u8x16
 aes_block_load (u8 * p)
 {
-  return (u8x16) _mm_loadu_si128 ((__m128i *) p);
+  return *(u8x16u *) p;
 }
 
 static_always_inline u8x16
 aes_enc_round (u8x16 a, u8x16 k)
 {
+#if defined (__AES__)
   return (u8x16) _mm_aesenc_si128 ((__m128i) a, (__m128i) k);
+#elif defined (__ARM_FEATURE_AES)
+  return vaesmcq_u8 (vaeseq_u8 (a, u8x16_splat (0))) ^ k;
+#endif
 }
 
 static_always_inline u8x16
 aes_enc_last_round (u8x16 a, u8x16 k)
 {
+#if defined (__AES__)
   return (u8x16) _mm_aesenclast_si128 ((__m128i) a, (__m128i) k);
+#elif defined (__ARM_FEATURE_AES)
+  return vaeseq_u8 (a, u8x16_splat (0)) ^ k;
+#endif
 }
 
+#ifdef __x86_64__
+
 static_always_inline u8x16
 aes_dec_round (u8x16 a, u8x16 k)
 {
@@ -63,11 +71,12 @@ aes_dec_last_round (u8x16 a, u8x16 k)
 {
   return (u8x16) _mm_aesdeclast_si128 ((__m128i) a, (__m128i) k);
 }
+#endif
 
 static_always_inline void
 aes_block_store (u8 * p, u8x16 r)
 {
-  _mm_storeu_si128 ((__m128i *) p, (__m128i) r);
+  *(u8x16u *) p = r;
 }
 
 static_always_inline u8x16
@@ -91,31 +100,40 @@ aes_load_partial (u8x16u * p, int n_bytes)
 static_always_inline void
 aes_store_partial (void *p, u8x16 r, int n_bytes)
 {
+#if __aarch64__
+  clib_memcpy_fast (p, &r, n_bytes);
+#else
 #ifdef __AVX512F__
   _mm_mask_storeu_epi8 (p, (1 << n_bytes) - 1, (__m128i) r);
 #else
   u8x16 mask = u8x16_is_greater (u8x16_splat (n_bytes), byte_mask_scale);
   _mm_maskmoveu_si128 ((__m128i) r, (__m128i) mask, p);
 #endif
+#endif
 }
 
 
 static_always_inline u8x16
 aes_encrypt_block (u8x16 block, const u8x16 * round_keys, aes_key_size_t ks)
 {
-  int i;
+  int rounds = AES_KEY_ROUNDS (ks);
   block ^= round_keys[0];
-  for (i = 1; i < AES_KEY_ROUNDS (ks); i += 1)
+  for (int i = 1; i < rounds; i += 1)
     block = aes_enc_round (block, round_keys[i]);
-  return aes_enc_last_round (block, round_keys[i]);
+  return aes_enc_last_round (block, round_keys[rounds]);
 }
 
 static_always_inline u8x16
 aes_inv_mix_column (u8x16 a)
 {
+#if defined (__AES__)
   return (u8x16) _mm_aesimc_si128 ((__m128i) a);
+#elif defined (__ARM_FEATURE_AES)
+  return vaesimcq_u8 (a);
+#endif
 }
 
+#ifdef __x86_64__
 #define aes_keygen_assist(a, b) \
   (u8x16) _mm_aeskeygenassist_si128((__m128i) a, b)
 
@@ -244,12 +262,6 @@ aes256_key_expand (u8x16 * rk, u8x16u const *k)
 
 #ifdef __aarch64__
 
-static_always_inline u8x16
-aes_inv_mix_column (u8x16 a)
-{
-  return vaesimcq_u8 (a);
-}
-
 static const u8x16 aese_prep_mask1 =
   { 13, 14, 15, 12, 13, 14, 15, 12, 13, 14, 15, 12, 13, 14, 15, 12 };
 static const u8x16 aese_prep_mask2 =
diff --git a/src/plugins/crypto_native/aes_gcm.c b/src/plugins/crypto_native/aes_gcm.c
index f2dec629359..41a9d480269 100644
--- a/src/plugins/crypto_native/aes_gcm.c
+++ b/src/plugins/crypto_native/aes_gcm.c
@@ -18,7 +18,6 @@
 #include <vlib/vlib.h>
 #include <vnet/plugin/plugin.h>
 #include <vnet/crypto/crypto.h>
-#include <x86intrin.h>
 #include <crypto_native/crypto_native.h>
 #include <crypto_native/aes.h>
 #include <crypto_native/ghash.h>
@@ -35,18 +34,6 @@ typedef struct
   const u8x16 Ke[15];
 } aes_gcm_key_data_t;
 
-static const u32x4 last_byte_one = { 0, 0, 0, 1 << 24 };
-
-static const u8x16 bswap_mask = {
-  15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0
-};
-
-static_always_inline u8x16
-aesni_gcm_bswap (u8x16 x)
-{
-  return (u8x16) _mm_shuffle_epi8 ((__m128i) x, (__m128i) bswap_mask);
-}
-
 static_always_inline void
 aesni_gcm_load (u8x16 * d, u8x16u * inv, int n, int n_bytes)
 {
@@ -70,6 +57,8 @@ static_always_inline void
 aesni_gcm_enc_first_round (u8x16 * r, u32x4 * Y, u32 * ctr, u8x16 k,
 			   int n_blocks)
 {
+  static const u32x4 last_byte_one = { 0, 0, 0, 1 << 24 };
+
   if (PREDICT_TRUE ((u8) ctr[0] < (256 - n_blocks)))
     {
       for (int i = 0; i < n_blocks; i++)
@@ -115,9 +104,9 @@ aesni_gcm_ghash_blocks (u8x16 T, aes_gcm_key_data_t * kd,
 {
   ghash_data_t _gd, *gd = &_gd;
   const u8x16 *Hi = kd->Hi + n_blocks - 1;
-  ghash_mul_first (gd, aesni_gcm_bswap (in[0]) ^ T, Hi[0]);
+  ghash_mul_first (gd, u8x16_reflect (in[0]) ^ T, Hi[0]);
   for (int i = 1; i < n_blocks; i++)
-    ghash_mul_next (gd, aesni_gcm_bswap ((in[i])), Hi[-i]);
+    ghash_mul_next (gd, u8x16_reflect ((in[i])), Hi[-i]);
   ghash_reduce (gd);
   ghash_reduce2 (gd);
   return ghash_final (gd);
@@ -158,7 +147,7 @@ aesni_gcm_ghash (u8x16 T, aes_gcm_key_data_t * kd, u8x16u * in, u32 n_left)
   if (n_left)
     {
       u8x16 r = aes_load_partial (in, n_left);
-      T = ghash_mul (aesni_gcm_bswap (r) ^ T, kd->Hi[0]);
+      T = ghash_mul (u8x16_reflect (r) ^ T, kd->Hi[0]);
     }
   return T;
 }
@@ -174,7 +163,7 @@ aesni_gcm_calc (u8x16 T, aes_gcm_key_data_t * kd, u8x16 * d,
   const u8x16 *rk = (u8x16 *) kd->Ke;
   int hidx = is_encrypt ? 4 : n, didx = 0;
 
-  _mm_prefetch (inv + 4, _MM_HINT_T0);
+  clib_prefetch_load (inv + 4);
 
   /* AES rounds 0 and 1 */
   aesni_gcm_enc_first_round (r, Y, ctr, rk[0], n);
@@ -186,7 +175,7 @@ aesni_gcm_calc (u8x16 T, aes_gcm_key_data_t * kd, u8x16 * d,
 
   /* GHASH multiply block 1 */
   if (with_ghash)
-    ghash_mul_first (gd, aesni_gcm_bswap (d[didx++]) ^ T, kd->Hi[--hidx]);
+    ghash_mul_first (gd, u8x16_reflect (d[didx++]) ^ T, kd->Hi[--hidx]);
 
   /* AES rounds 2 and 3 */
   aesni_gcm_enc_round (r, rk[2], n);
@@ -194,7 +183,7 @@ aesni_gcm_calc (u8x16 T, aes_gcm_key_data_t * kd, u8x16 * d,
 
   /* GHASH multiply block 2 */
   if (with_ghash && hidx)
-    ghash_mul_next (gd, aesni_gcm_bswap (d[didx++]), kd->Hi[--hidx]);
+    ghash_mul_next (gd, u8x16_reflect (d[didx++]), kd->Hi[--hidx]);
 
   /* AES rounds 4 and 5 */
   aesni_gcm_enc_round (r, rk[4], n);
@@ -202,7 +191,7 @@ aesni_gcm_calc (u8x16 T, aes_gcm_key_data_t * kd, u8x16 * d,
 
   /* GHASH multiply block 3 */
   if (with_ghash && hidx)
-    ghash_mul_next (gd, aesni_gcm_bswap (d[didx++]), kd->Hi[--hidx]);
+    ghash_mul_next (gd, u8x16_reflect (d[didx++]), kd->Hi[--hidx]);
 
   /* AES rounds 6 and 7 */
   aesni_gcm_enc_round (r, rk[6], n);
@@ -210,7 +199,7 @@ aesni_gcm_calc (u8x16 T, aes_gcm_key_data_t * kd, u8x16 * d,
 
   /* GHASH multiply block 4 */
   if (with_ghash && hidx)
-    ghash_mul_next (gd, aesni_gcm_bswap (d[didx++]), kd->Hi[--hidx]);
+    ghash_mul_next (gd, u8x16_reflect (d[didx++]), kd->Hi[--hidx]);
 
   /* AES rounds 8 and 9 */
   aesni_gcm_enc_round (r, rk[8], n);
@@ -259,28 +248,28 @@ aesni_gcm_calc_double (u8x16 T, aes_gcm_key_data_t * kd, u8x16 * d,
     aesni_gcm_load (d, inv, 4, 0);
 
   /* GHASH multiply block 0 */
-  ghash_mul_first (gd, aesni_gcm_bswap (d[0]) ^ T, kd->Hi[7]);
+  ghash_mul_first (gd, u8x16_reflect (d[0]) ^ T, kd->Hi[7]);
 
   /* AES rounds 2 and 3 */
   aesni_gcm_enc_round (r, rk[2], 4);
   aesni_gcm_enc_round (r, rk[3], 4);
 
   /* GHASH multiply block 1 */
-  ghash_mul_next (gd, aesni_gcm_bswap (d[1]), kd->Hi[6]);
+  ghash_mul_next (gd, u8x16_reflect (d[1]), kd->Hi[6]);
 
   /* AES rounds 4 and 5 */
   aesni_gcm_enc_round (r, rk[4], 4);
   aesni_gcm_enc_round (r, rk[5], 4);
 
   /* GHASH multiply block 2 */
-  ghash_mul_next (gd, aesni_gcm_bswap (d[2]), kd->Hi[5]);
+  ghash_mul_next (gd, u8x16_reflect (d[2]), kd->Hi[5]);
 
   /* AES rounds 6 and 7 */
   aesni_gcm_enc_round (r, rk[6], 4);
   aesni_gcm_enc_round (r, rk[7], 4);
 
   /* GHASH multiply block 3 */
-  ghash_mul_next (gd, aesni_gcm_bswap (d[3]), kd->Hi[4]);
+  ghash_mul_next (gd, u8x16_reflect (d[3]), kd->Hi[4]);
 
   /* AES rounds 8 and 9 */
   aesni_gcm_enc_round (r, rk[8], 4);
@@ -301,7 +290,7 @@ aesni_gcm_calc_double (u8x16 T, aes_gcm_key_data_t * kd, u8x16 * d,
     aesni_gcm_load (d, inv + 4, 4, 0);
 
   /* GHASH multiply block 4 */
-  ghash_mul_next (gd, aesni_gcm_bswap (d[0]), kd->Hi[3]);
+  ghash_mul_next (gd, u8x16_reflect (d[0]), kd->Hi[3]);
 
   /* AES rounds 0, 1 and 2 */
   aesni_gcm_enc_first_round (r, Y, ctr, rk[0], 4);
@@ -309,21 +298,21 @@ aesni_gcm_calc_double (u8x16 T, aes_gcm_key_data_t * kd, u8x16 * d,
   aesni_gcm_enc_round (r, rk[2], 4);
 
   /* GHASH multiply block 5 */
-  ghash_mul_next (gd, aesni_gcm_bswap (d[1]), kd->Hi[2]);
+  ghash_mul_next (gd, u8x16_reflect (d[1]), kd->Hi[2]);
 
   /* AES rounds 3 and 4 */
   aesni_gcm_enc_round (r, rk[3], 4);
   aesni_gcm_enc_round (r, rk[4], 4);
 
   /* GHASH multiply block 6 */
-  ghash_mul_next (gd, aesni_gcm_bswap (d[2]), kd->Hi[1]);
+  ghash_mul_next (gd, u8x16_reflect (d[2]), kd->Hi[1]);
 
   /* AES rounds 5 and 6 */
   aesni_gcm_enc_round (r, rk[5], 4);
   aesni_gcm_enc_round (r, rk[6], 4);
 
   /* GHASH multiply block 7 */
-  ghash_mul_next (gd, aesni_gcm_bswap (d[3]), kd->Hi[0]);
+  ghash_mul_next (gd, u8x16_reflect (d[3]), kd->Hi[0]);
 
   /* AES rounds 7 and 8 */
   aesni_gcm_enc_round (r, rk[7], 4);
@@ -361,13 +350,13 @@ aesni_gcm_ghash_last (u8x16 T, aes_gcm_key_data_t * kd, u8x16 * d,
   if (n_bytes)
     d[n_blocks - 1] = aes_byte_mask (d[n_blocks - 1], n_bytes);
 
-  ghash_mul_first (gd, aesni_gcm_bswap (d[0]) ^ T, kd->Hi[n_blocks - 1]);
+  ghash_mul_first (gd, u8x16_reflect (d[0]) ^ T, kd->Hi[n_blocks - 1]);
   if (n_blocks > 1)
-    ghash_mul_next (gd, aesni_gcm_bswap (d[1]), kd->Hi[n_blocks - 2]);
+    ghash_mul_next (gd, u8x16_reflect (d[1]), kd->Hi[n_blocks - 2]);
   if (n_blocks > 2)
-    ghash_mul_next (gd, aesni_gcm_bswap (d[2]), kd->Hi[n_blocks - 3]);
+    ghash_mul_next (gd, u8x16_reflect (d[2]), kd->Hi[n_blocks - 3]);
   if (n_blocks > 3)
-    ghash_mul_next (gd, aesni_gcm_bswap (d[3]), kd->Hi[n_blocks - 4]);
+    ghash_mul_next (gd, u8x16_reflect (d[3]), kd->Hi[n_blocks - 4]);
   ghash_reduce (gd);
   ghash_reduce2 (gd);
   return ghash_final (gd);
@@ -539,9 +528,9 @@ aes_gcm (u8x16u * in, u8x16u * out, u8x16u * addt, u8x16u * iv, u8x16u * tag,
   u32x4 Y0;
   ghash_data_t _gd, *gd = &_gd;
 
-  _mm_prefetch (iv, _MM_HINT_T0);
-  _mm_prefetch (in, _MM_HINT_T0);
-  _mm_prefetch (in + CLIB_CACHE_LINE_BYTES, _MM_HINT_T0);
+  clib_prefetch_load (iv);
+  clib_prefetch_load (in);
+  clib_prefetch_load (in + 4);
 
   /* calculate ghash for AAD - optimized for ipsec common cases */
   if (aad_bytes == 8)
@@ -561,7 +550,7 @@ aes_gcm (u8x16u * in, u8x16u * out, u8x16u * addt, u8x16u * iv, u8x16u * tag,
   else
     T = aesni_gcm_dec (T, kd, Y0, in, out, data_bytes, aes_rounds);
 
-  _mm_prefetch (tag, _MM_HINT_T0);
+  clib_prefetch_load (tag);
 
   /* Finalize ghash  - data bytes and aad bytes converted to bits */
   /* *INDENT-OFF* */
@@ -581,7 +570,7 @@ aes_gcm (u8x16u * in, u8x16u * out, u8x16u * addt, u8x16u * iv, u8x16u * tag,
   for (; i < aes_rounds; i += 1)
     r = aes_enc_round (r, kd->Ke[i]);
   r = aes_enc_last_round (r, kd->Ke[aes_rounds]);
-  T = aesni_gcm_bswap (T) ^ r;
+  T = u8x16_reflect (T) ^ r;
 
   /* tag_len 16 -> 0 */
   tag_len &= 0xf;
@@ -679,7 +668,7 @@ aesni_gcm_key_exp (vnet_crypto_key_t * key, aes_key_size_t ks)
 
   /* pre-calculate H */
   H = aes_encrypt_block (u8x16_splat (0), kd->Ke, ks);
-  H = aesni_gcm_bswap (H);
+  H = u8x16_reflect (H);
   ghash_precompute (H, (u8x16 *) kd->Hi, 8);
   return kd;
 }
@@ -706,6 +695,8 @@ crypto_native_aes_gcm_init_vaes (vlib_main_t * vm)
 crypto_native_aes_gcm_init_avx512 (vlib_main_t * vm)
 #elif __AVX2__
 crypto_native_aes_gcm_init_avx2 (vlib_main_t * vm)
+#elif __aarch64__
+crypto_native_aes_gcm_init_neon (vlib_main_t * vm)
 #else
 crypto_native_aes_gcm_init_sse42 (vlib_main_t * vm)
 #endif
diff --git a/src/plugins/crypto_native/crypto_native.h b/src/plugins/crypto_native/crypto_native.h
index 0b6116a7d38..279684b8b8e 100644
--- a/src/plugins/crypto_native/crypto_native.h
+++ b/src/plugins/crypto_native/crypto_native.h
@@ -45,6 +45,7 @@ clib_error_t *crypto_native_aes_gcm_init_sse42 (vlib_main_t * vm);
 clib_error_t *crypto_native_aes_gcm_init_avx2 (vlib_main_t * vm);
 clib_error_t *crypto_native_aes_gcm_init_avx512 (vlib_main_t * vm);
 clib_error_t *crypto_native_aes_gcm_init_vaes (vlib_main_t * vm);
+clib_error_t *crypto_native_aes_gcm_init_neon (vlib_main_t * vm);
 #endif /* __crypto_native_h__ */
 
 /*
diff --git a/src/plugins/crypto_native/main.c b/src/plugins/crypto_native/main.c
index d338ab6d910..45d3d8dfe4e 100644
--- a/src/plugins/crypto_native/main.c
+++ b/src/plugins/crypto_native/main.c
@@ -102,9 +102,10 @@ crypto_native_init (vlib_main_t * vm)
     }
 #endif
 #if __aarch64__
-  error = crypto_native_aes_cbc_init_neon (vm);
+  if ((error = crypto_native_aes_cbc_init_neon (vm)))
+    goto error;
 
-  if (error)
+  if ((error = crypto_native_aes_gcm_init_neon (vm)))
     goto error;
 #endif
 
diff --git a/src/vppinfra/vector_neon.h b/src/vppinfra/vector_neon.h
index 307fbc541b2..81d99a64f05 100644
--- a/src/vppinfra/vector_neon.h
+++ b/src/vppinfra/vector_neon.h
@@ -194,6 +194,15 @@ u8x16_word_shift_right (u8x16 x, const int n)
   return vextq_u8 (x, u8x16_splat (0), n);
 }
 
+static_always_inline u8x16
+u8x16_reflect (u8x16 v)
+{
+  u8x16 mask = {
+    15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0
+  };
+  return (u8x16) vqtbl1q_u8 (v, mask);
+}
+
 #define CLIB_HAVE_VEC128_MSB_MASK
 
 #define CLIB_HAVE_VEC128_UNALIGNED_LOAD_STORE
diff --git a/src/vppinfra/vector_sse42.h b/src/vppinfra/vector_sse42.h
index 0c1b2f0324f..c22e86e7437 100644
--- a/src/vppinfra/vector_sse42.h
+++ b/src/vppinfra/vector_sse42.h
@@ -635,6 +635,15 @@ u16x8_byte_swap (u16x8 v)
   return (u16x8) _mm_shuffle_epi8 ((__m128i) v, (__m128i) swap);
 }
 
+static_always_inline u8x16
+u8x16_reflect (u8x16 v)
+{
+  u8x16 mask = {
+    15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0
+  };
+  return (u8x16) _mm_shuffle_epi8 ((__m128i) v, (__m128i) mask);
+}
+
 static_always_inline u32x4
 u32x4_hadd (u32x4 v1, u32x4 v2)
 {