1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Copyright (C) 2019 Linaro, Ltd. <ard.biesheuvel@linaro.org>
4  */
5 
6 #ifdef CONFIG_ARM64
7 #include <asm/neon-intrinsics.h>
8 
9 #define AES_ROUND	"aese %0.16b, %1.16b \n\t aesmc %0.16b, %0.16b"
10 #else
11 #include <arm_neon.h>
12 
13 #define AES_ROUND	"aese.8 %q0, %q1 \n\t aesmc.8 %q0, %q0"
14 #endif
15 
16 #define AEGIS_BLOCK_SIZE	16
17 
18 #include <stddef.h>
19 
20 extern int aegis128_have_aes_insn;
21 
22 void *memcpy(void *dest, const void *src, size_t n);
23 void *memset(void *s, int c, size_t n);
24 
25 struct aegis128_state {
26 	uint8x16_t v[5];
27 };
28 
29 extern const uint8_t crypto_aes_sbox[];
30 
aegis128_load_state_neon(const void * state)31 static struct aegis128_state aegis128_load_state_neon(const void *state)
32 {
33 	return (struct aegis128_state){ {
34 		vld1q_u8(state),
35 		vld1q_u8(state + 16),
36 		vld1q_u8(state + 32),
37 		vld1q_u8(state + 48),
38 		vld1q_u8(state + 64)
39 	} };
40 }
41 
aegis128_save_state_neon(struct aegis128_state st,void * state)42 static void aegis128_save_state_neon(struct aegis128_state st, void *state)
43 {
44 	vst1q_u8(state, st.v[0]);
45 	vst1q_u8(state + 16, st.v[1]);
46 	vst1q_u8(state + 32, st.v[2]);
47 	vst1q_u8(state + 48, st.v[3]);
48 	vst1q_u8(state + 64, st.v[4]);
49 }
50 
51 static inline __attribute__((always_inline))
aegis_aes_round(uint8x16_t w)52 uint8x16_t aegis_aes_round(uint8x16_t w)
53 {
54 	uint8x16_t z = {};
55 
56 #ifdef CONFIG_ARM64
57 	if (!__builtin_expect(aegis128_have_aes_insn, 1)) {
58 		static const uint8_t shift_rows[] = {
59 			0x0, 0x5, 0xa, 0xf, 0x4, 0x9, 0xe, 0x3,
60 			0x8, 0xd, 0x2, 0x7, 0xc, 0x1, 0x6, 0xb,
61 		};
62 		static const uint8_t ror32by8[] = {
63 			0x1, 0x2, 0x3, 0x0, 0x5, 0x6, 0x7, 0x4,
64 			0x9, 0xa, 0xb, 0x8, 0xd, 0xe, 0xf, 0xc,
65 		};
66 		uint8x16_t v;
67 
68 		// shift rows
69 		w = vqtbl1q_u8(w, vld1q_u8(shift_rows));
70 
71 		// sub bytes
72 #ifndef CONFIG_CC_IS_GCC
73 		v = vqtbl4q_u8(vld1q_u8_x4(crypto_aes_sbox), w);
74 		v = vqtbx4q_u8(v, vld1q_u8_x4(crypto_aes_sbox + 0x40), w - 0x40);
75 		v = vqtbx4q_u8(v, vld1q_u8_x4(crypto_aes_sbox + 0x80), w - 0x80);
76 		v = vqtbx4q_u8(v, vld1q_u8_x4(crypto_aes_sbox + 0xc0), w - 0xc0);
77 #else
78 		asm("tbl %0.16b, {v16.16b-v19.16b}, %1.16b" : "=w"(v) : "w"(w));
79 		w -= 0x40;
80 		asm("tbx %0.16b, {v20.16b-v23.16b}, %1.16b" : "+w"(v) : "w"(w));
81 		w -= 0x40;
82 		asm("tbx %0.16b, {v24.16b-v27.16b}, %1.16b" : "+w"(v) : "w"(w));
83 		w -= 0x40;
84 		asm("tbx %0.16b, {v28.16b-v31.16b}, %1.16b" : "+w"(v) : "w"(w));
85 #endif
86 
87 		// mix columns
88 		w = (v << 1) ^ (uint8x16_t)(((int8x16_t)v >> 7) & 0x1b);
89 		w ^= (uint8x16_t)vrev32q_u16((uint16x8_t)v);
90 		w ^= vqtbl1q_u8(v ^ w, vld1q_u8(ror32by8));
91 
92 		return w;
93 	}
94 #endif
95 
96 	/*
97 	 * We use inline asm here instead of the vaeseq_u8/vaesmcq_u8 intrinsics
98 	 * to force the compiler to issue the aese/aesmc instructions in pairs.
99 	 * This is much faster on many cores, where the instruction pair can
100 	 * execute in a single cycle.
101 	 */
102 	asm(AES_ROUND : "+w"(w) : "w"(z));
103 	return w;
104 }
105 
106 static inline __attribute__((always_inline))
aegis128_update_neon(struct aegis128_state st,uint8x16_t m)107 struct aegis128_state aegis128_update_neon(struct aegis128_state st,
108 					   uint8x16_t m)
109 {
110 	m       ^= aegis_aes_round(st.v[4]);
111 	st.v[4] ^= aegis_aes_round(st.v[3]);
112 	st.v[3] ^= aegis_aes_round(st.v[2]);
113 	st.v[2] ^= aegis_aes_round(st.v[1]);
114 	st.v[1] ^= aegis_aes_round(st.v[0]);
115 	st.v[0] ^= m;
116 
117 	return st;
118 }
119 
120 static inline __attribute__((always_inline))
preload_sbox(void)121 void preload_sbox(void)
122 {
123 	if (!IS_ENABLED(CONFIG_ARM64) ||
124 	    !IS_ENABLED(CONFIG_CC_IS_GCC) ||
125 	    __builtin_expect(aegis128_have_aes_insn, 1))
126 		return;
127 
128 	asm("ld1	{v16.16b-v19.16b}, [%0], #64	\n\t"
129 	    "ld1	{v20.16b-v23.16b}, [%0], #64	\n\t"
130 	    "ld1	{v24.16b-v27.16b}, [%0], #64	\n\t"
131 	    "ld1	{v28.16b-v31.16b}, [%0]		\n\t"
132 	    :: "r"(crypto_aes_sbox));
133 }
134 
crypto_aegis128_update_neon(void * state,const void * msg)135 void crypto_aegis128_update_neon(void *state, const void *msg)
136 {
137 	struct aegis128_state st = aegis128_load_state_neon(state);
138 
139 	preload_sbox();
140 
141 	st = aegis128_update_neon(st, vld1q_u8(msg));
142 
143 	aegis128_save_state_neon(st, state);
144 }
145 
crypto_aegis128_encrypt_chunk_neon(void * state,void * dst,const void * src,unsigned int size)146 void crypto_aegis128_encrypt_chunk_neon(void *state, void *dst, const void *src,
147 					unsigned int size)
148 {
149 	struct aegis128_state st = aegis128_load_state_neon(state);
150 	uint8x16_t msg;
151 
152 	preload_sbox();
153 
154 	while (size >= AEGIS_BLOCK_SIZE) {
155 		uint8x16_t s = st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
156 
157 		msg = vld1q_u8(src);
158 		st = aegis128_update_neon(st, msg);
159 		vst1q_u8(dst, msg ^ s);
160 
161 		size -= AEGIS_BLOCK_SIZE;
162 		src += AEGIS_BLOCK_SIZE;
163 		dst += AEGIS_BLOCK_SIZE;
164 	}
165 
166 	if (size > 0) {
167 		uint8x16_t s = st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
168 		uint8_t buf[AEGIS_BLOCK_SIZE] = {};
169 
170 		memcpy(buf, src, size);
171 		msg = vld1q_u8(buf);
172 		st = aegis128_update_neon(st, msg);
173 		vst1q_u8(buf, msg ^ s);
174 		memcpy(dst, buf, size);
175 	}
176 
177 	aegis128_save_state_neon(st, state);
178 }
179 
crypto_aegis128_decrypt_chunk_neon(void * state,void * dst,const void * src,unsigned int size)180 void crypto_aegis128_decrypt_chunk_neon(void *state, void *dst, const void *src,
181 					unsigned int size)
182 {
183 	struct aegis128_state st = aegis128_load_state_neon(state);
184 	uint8x16_t msg;
185 
186 	preload_sbox();
187 
188 	while (size >= AEGIS_BLOCK_SIZE) {
189 		msg = vld1q_u8(src) ^ st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
190 		st = aegis128_update_neon(st, msg);
191 		vst1q_u8(dst, msg);
192 
193 		size -= AEGIS_BLOCK_SIZE;
194 		src += AEGIS_BLOCK_SIZE;
195 		dst += AEGIS_BLOCK_SIZE;
196 	}
197 
198 	if (size > 0) {
199 		uint8x16_t s = st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
200 		uint8_t buf[AEGIS_BLOCK_SIZE];
201 
202 		vst1q_u8(buf, s);
203 		memcpy(buf, src, size);
204 		msg = vld1q_u8(buf) ^ s;
205 		vst1q_u8(buf, msg);
206 		memcpy(dst, buf, size);
207 
208 		st = aegis128_update_neon(st, msg);
209 	}
210 
211 	aegis128_save_state_neon(st, state);
212 }
213