crypto: arm64/aegis128 - implement plain NEON version
authorArd Biesheuvel <ard.biesheuvel@linaro.org>
Sun, 11 Aug 2019 22:59:12 +0000 (01:59 +0300)
committerHerbert Xu <herbert@gondor.apana.org.au>
Thu, 15 Aug 2019 11:52:15 +0000 (21:52 +1000)
Provide a version of the core AES transform to the aegis128 SIMD
code that does not rely on the special AES instructions, but uses
plain NEON instructions instead. This allows the SIMD version of
the aegis128 driver to be used on arm64 systems that do not
implement those instructions (which are not mandatory in the
architecture), such as the Raspberry Pi 3.

Since GCC makes a mess of this when using the tbl/tbx intrinsics
to perform the sbox substitution, preload the Sbox into v16..v31
in this case and use inline asm to emit the tbl/tbx instructions.
Clang does not support this approach, nor does it require it, since
it does a much better job at code generation, so there we use the
intrinsics as usual.

Cc: Nick Desaulniers <ndesaulniers@google.com>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Acked-by: Nick Desaulniers <ndesaulniers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
crypto/Makefile
crypto/aegis128-neon-inner.c
crypto/aegis128-neon.c

index 99a9fa9087d1fcb89c289a7590a9aecb2b2cd476..0d2cdd523fd982cf532451d07ff32e7a3d6e9802 100644 (file)
@@ -98,7 +98,14 @@ CFLAGS_aegis128-neon-inner.o += -mfpu=crypto-neon-fp-armv8
 aegis128-$(CONFIG_CRYPTO_AEGIS128_SIMD) += aegis128-neon.o aegis128-neon-inner.o
 endif
 ifeq ($(ARCH),arm64)
-CFLAGS_aegis128-neon-inner.o += -ffreestanding -mcpu=generic+crypto
+aegis128-cflags-y := -ffreestanding -mcpu=generic+crypto
+aegis128-cflags-$(CONFIG_CC_IS_GCC) += -ffixed-q16 -ffixed-q17 -ffixed-q18 \
+                                      -ffixed-q19 -ffixed-q20 -ffixed-q21 \
+                                      -ffixed-q22 -ffixed-q23 -ffixed-q24 \
+                                      -ffixed-q25 -ffixed-q26 -ffixed-q27 \
+                                      -ffixed-q28 -ffixed-q29 -ffixed-q30 \
+                                      -ffixed-q31
+CFLAGS_aegis128-neon-inner.o += $(aegis128-cflags-y)
 CFLAGS_REMOVE_aegis128-neon-inner.o += -mgeneral-regs-only
 aegis128-$(CONFIG_CRYPTO_AEGIS128_SIMD) += aegis128-neon.o aegis128-neon-inner.o
 endif
index 3d8043c4832b35c84100a88edeb3b70aa0baa2e8..ed55568afd1be7ab184015fe625ecc10f46740ba 100644 (file)
@@ -17,6 +17,8 @@
 
 #include <stddef.h>
 
+extern int aegis128_have_aes_insn;
+
 void *memcpy(void *dest, const void *src, size_t n);
 void *memset(void *s, int c, size_t n);
 
@@ -24,6 +26,8 @@ struct aegis128_state {
        uint8x16_t v[5];
 };
 
+extern const uint8x16x4_t crypto_aes_sbox[];
+
 static struct aegis128_state aegis128_load_state_neon(const void *state)
 {
        return (struct aegis128_state){ {
@@ -49,6 +53,46 @@ uint8x16_t aegis_aes_round(uint8x16_t w)
 {
        uint8x16_t z = {};
 
+#ifdef CONFIG_ARM64
+       if (!__builtin_expect(aegis128_have_aes_insn, 1)) {
+               static const uint8x16_t shift_rows = {
+                       0x0, 0x5, 0xa, 0xf, 0x4, 0x9, 0xe, 0x3,
+                       0x8, 0xd, 0x2, 0x7, 0xc, 0x1, 0x6, 0xb,
+               };
+               static const uint8x16_t ror32by8 = {
+                       0x1, 0x2, 0x3, 0x0, 0x5, 0x6, 0x7, 0x4,
+                       0x9, 0xa, 0xb, 0x8, 0xd, 0xe, 0xf, 0xc,
+               };
+               uint8x16_t v;
+
+               // shift rows
+               w = vqtbl1q_u8(w, shift_rows);
+
+               // sub bytes
+               if (!IS_ENABLED(CONFIG_CC_IS_GCC)) {
+                       v = vqtbl4q_u8(crypto_aes_sbox[0], w);
+                       v = vqtbx4q_u8(v, crypto_aes_sbox[1], w - 0x40);
+                       v = vqtbx4q_u8(v, crypto_aes_sbox[2], w - 0x80);
+                       v = vqtbx4q_u8(v, crypto_aes_sbox[3], w - 0xc0);
+               } else {
+                       asm("tbl %0.16b, {v16.16b-v19.16b}, %1.16b" : "=w"(v) : "w"(w));
+                       w -= 0x40;
+                       asm("tbx %0.16b, {v20.16b-v23.16b}, %1.16b" : "+w"(v) : "w"(w));
+                       w -= 0x40;
+                       asm("tbx %0.16b, {v24.16b-v27.16b}, %1.16b" : "+w"(v) : "w"(w));
+                       w -= 0x40;
+                       asm("tbx %0.16b, {v28.16b-v31.16b}, %1.16b" : "+w"(v) : "w"(w));
+               }
+
+               // mix columns
+               w = (v << 1) ^ (uint8x16_t)(((int8x16_t)v >> 7) & 0x1b);
+               w ^= (uint8x16_t)vrev32q_u16((uint16x8_t)v);
+               w ^= vqtbl1q_u8(v ^ w, ror32by8);
+
+               return w;
+       }
+#endif
+
        /*
         * We use inline asm here instead of the vaeseq_u8/vaesmcq_u8 intrinsics
         * to force the compiler to issue the aese/aesmc instructions in pairs.
@@ -73,10 +117,27 @@ struct aegis128_state aegis128_update_neon(struct aegis128_state st,
        return st;
 }
 
+static inline __attribute__((always_inline))
+void preload_sbox(void)
+{
+       if (!IS_ENABLED(CONFIG_ARM64) ||
+           !IS_ENABLED(CONFIG_CC_IS_GCC) ||
+           __builtin_expect(aegis128_have_aes_insn, 1))
+               return;
+
+       asm("ld1        {v16.16b-v19.16b}, [%0], #64    \n\t"
+           "ld1        {v20.16b-v23.16b}, [%0], #64    \n\t"
+           "ld1        {v24.16b-v27.16b}, [%0], #64    \n\t"
+           "ld1        {v28.16b-v31.16b}, [%0]         \n\t"
+           :: "r"(crypto_aes_sbox));
+}
+
 void crypto_aegis128_update_neon(void *state, const void *msg)
 {
        struct aegis128_state st = aegis128_load_state_neon(state);
 
+       preload_sbox();
+
        st = aegis128_update_neon(st, vld1q_u8(msg));
 
        aegis128_save_state_neon(st, state);
@@ -88,6 +149,8 @@ void crypto_aegis128_encrypt_chunk_neon(void *state, void *dst, const void *src,
        struct aegis128_state st = aegis128_load_state_neon(state);
        uint8x16_t msg;
 
+       preload_sbox();
+
        while (size >= AEGIS_BLOCK_SIZE) {
                uint8x16_t s = st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
 
@@ -120,6 +183,8 @@ void crypto_aegis128_decrypt_chunk_neon(void *state, void *dst, const void *src,
        struct aegis128_state st = aegis128_load_state_neon(state);
        uint8x16_t msg;
 
+       preload_sbox();
+
        while (size >= AEGIS_BLOCK_SIZE) {
                msg = vld1q_u8(src) ^ st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
                st = aegis128_update_neon(st, msg);
index c1c0a1686f67a1511ebb06f82f237c2eef43d745..751f9c195aa42a0b43916c16362e15b32497000c 100644 (file)
@@ -14,9 +14,15 @@ void crypto_aegis128_encrypt_chunk_neon(void *state, void *dst, const void *src,
 void crypto_aegis128_decrypt_chunk_neon(void *state, void *dst, const void *src,
                                        unsigned int size);
 
+int aegis128_have_aes_insn __ro_after_init;
+
 bool crypto_aegis128_have_simd(void)
 {
-       return cpu_have_feature(cpu_feature(AES));
+       if (cpu_have_feature(cpu_feature(AES))) {
+               aegis128_have_aes_insn = 1;
+               return true;
+       }
+       return IS_ENABLED(CONFIG_ARM64);
 }
 
 void crypto_aegis128_update_simd(union aegis_block *state, const void *msg)