|
| 1 | + |
| 2 | +/* |
| 3 | + chacha-merged.c version 20080118 |
| 4 | + D. J. Bernstein |
| 5 | + Public domain. |
| 6 | + */ |
| 7 | + |
| 8 | +#include <stdint.h> |
| 9 | +#include <stdlib.h> |
| 10 | +#include <string.h> |
| 11 | +#include "utils.h" |
| 12 | +#include "chacha20.h" |
| 13 | + |
| 14 | +struct chacha_ctx { |
| 15 | + uint32_t input[16]; |
| 16 | +}; |
| 17 | + |
| 18 | +#define LOAD32_LE(SRC) get_unaligned_le32(SRC) |
| 19 | + |
| 20 | +#define STORE32_LE(DST, W) store32_le((DST), (W)) |
| 21 | + |
| 22 | +static inline void |
| 23 | +store32_le(uint8_t dst[4], uint32_t w) |
| 24 | +{ |
| 25 | + dst[0] = (uint8_t) w; w >>= 8; |
| 26 | + dst[1] = (uint8_t) w; w >>= 8; |
| 27 | + dst[2] = (uint8_t) w; w >>= 8; |
| 28 | + dst[3] = (uint8_t) w; |
| 29 | +} |
| 30 | + |
| 31 | + |
| 32 | +#define ROTL32(X, B) rotl32((X), (B)) |
| 33 | +static inline uint32_t |
| 34 | +rotl32(const uint32_t x, const int b) |
| 35 | +{ |
| 36 | + return (x << b) | (x >> (32 - b)); |
| 37 | +} |
| 38 | + |
| 39 | +typedef struct chacha_ctx chacha_ctx; |
| 40 | + |
| 41 | +#define U32C(v) (v##U) |
| 42 | + |
| 43 | +#define U32V(v) ((uint32_t)(v) &U32C(0xFFFFFFFF)) |
| 44 | + |
| 45 | +#define ROTATE(v, c) (ROTL32(v, c)) |
| 46 | +#define XOR(v, w) ((v) ^ (w)) |
| 47 | +#define PLUS(v, w) (U32V((v) + (w))) |
| 48 | +#define PLUSONE(v) (PLUS((v), 1)) |
| 49 | + |
| 50 | +#define QUARTERROUND(a, b, c, d) \ |
| 51 | + a = PLUS(a, b); \ |
| 52 | + d = ROTATE(XOR(d, a), 16); \ |
| 53 | + c = PLUS(c, d); \ |
| 54 | + b = ROTATE(XOR(b, c), 12); \ |
| 55 | + a = PLUS(a, b); \ |
| 56 | + d = ROTATE(XOR(d, a), 8); \ |
| 57 | + c = PLUS(c, d); \ |
| 58 | + b = ROTATE(XOR(b, c), 7); |
| 59 | + |
| 60 | +static void |
| 61 | +chacha_keysetup(chacha_ctx *ctx, const uint8_t *k) |
| 62 | +{ |
| 63 | + ctx->input[0] = U32C(0x61707865); |
| 64 | + ctx->input[1] = U32C(0x3320646e); |
| 65 | + ctx->input[2] = U32C(0x79622d32); |
| 66 | + ctx->input[3] = U32C(0x6b206574); |
| 67 | + ctx->input[4] = LOAD32_LE(k + 0); |
| 68 | + ctx->input[5] = LOAD32_LE(k + 4); |
| 69 | + ctx->input[6] = LOAD32_LE(k + 8); |
| 70 | + ctx->input[7] = LOAD32_LE(k + 12); |
| 71 | + ctx->input[8] = LOAD32_LE(k + 16); |
| 72 | + ctx->input[9] = LOAD32_LE(k + 20); |
| 73 | + ctx->input[10] = LOAD32_LE(k + 24); |
| 74 | + ctx->input[11] = LOAD32_LE(k + 28); |
| 75 | +} |
| 76 | + |
| 77 | +static void |
| 78 | +chacha_ivsetup(chacha_ctx *ctx, const uint8_t *iv, const uint8_t *counter) |
| 79 | +{ |
| 80 | + ctx->input[12] = counter == NULL ? 0 : LOAD32_LE(counter + 0); |
| 81 | + ctx->input[13] = counter == NULL ? 0 : LOAD32_LE(counter + 4); |
| 82 | + ctx->input[14] = LOAD32_LE(iv + 0); |
| 83 | + ctx->input[15] = LOAD32_LE(iv + 4); |
| 84 | +} |
| 85 | + |
| 86 | +static void |
| 87 | +chacha20_encrypt_bytes(chacha_ctx *ctx, const uint8_t *m, uint8_t *c, |
| 88 | + unsigned long long bytes) |
| 89 | +{ |
| 90 | + uint32_t x0, x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11, x12, x13, x14, |
| 91 | + x15; |
| 92 | + uint32_t j0, j1, j2, j3, j4, j5, j6, j7, j8, j9, j10, j11, j12, j13, j14, |
| 93 | + j15; |
| 94 | + uint8_t *ctarget = NULL; |
| 95 | + uint8_t tmp[64]; |
| 96 | + unsigned int i; |
| 97 | + |
| 98 | + if (!bytes) { |
| 99 | + return; /* LCOV_EXCL_LINE */ |
| 100 | + } |
| 101 | + j0 = ctx->input[0]; |
| 102 | + j1 = ctx->input[1]; |
| 103 | + j2 = ctx->input[2]; |
| 104 | + j3 = ctx->input[3]; |
| 105 | + j4 = ctx->input[4]; |
| 106 | + j5 = ctx->input[5]; |
| 107 | + j6 = ctx->input[6]; |
| 108 | + j7 = ctx->input[7]; |
| 109 | + j8 = ctx->input[8]; |
| 110 | + j9 = ctx->input[9]; |
| 111 | + j10 = ctx->input[10]; |
| 112 | + j11 = ctx->input[11]; |
| 113 | + j12 = ctx->input[12]; |
| 114 | + j13 = ctx->input[13]; |
| 115 | + j14 = ctx->input[14]; |
| 116 | + j15 = ctx->input[15]; |
| 117 | + |
| 118 | + for (;;) { |
| 119 | + if (bytes < 64) { |
| 120 | + memset(tmp, 0, 64); |
| 121 | + for (i = 0; i < bytes; ++i) { |
| 122 | + tmp[i] = m[i]; |
| 123 | + } |
| 124 | + m = tmp; |
| 125 | + ctarget = c; |
| 126 | + c = tmp; |
| 127 | + } |
| 128 | + x0 = j0; |
| 129 | + x1 = j1; |
| 130 | + x2 = j2; |
| 131 | + x3 = j3; |
| 132 | + x4 = j4; |
| 133 | + x5 = j5; |
| 134 | + x6 = j6; |
| 135 | + x7 = j7; |
| 136 | + x8 = j8; |
| 137 | + x9 = j9; |
| 138 | + x10 = j10; |
| 139 | + x11 = j11; |
| 140 | + x12 = j12; |
| 141 | + x13 = j13; |
| 142 | + x14 = j14; |
| 143 | + x15 = j15; |
| 144 | + for (i = 20; i > 0; i -= 2) { |
| 145 | + QUARTERROUND(x0, x4, x8, x12) |
| 146 | + QUARTERROUND(x1, x5, x9, x13) |
| 147 | + QUARTERROUND(x2, x6, x10, x14) |
| 148 | + QUARTERROUND(x3, x7, x11, x15) |
| 149 | + QUARTERROUND(x0, x5, x10, x15) |
| 150 | + QUARTERROUND(x1, x6, x11, x12) |
| 151 | + QUARTERROUND(x2, x7, x8, x13) |
| 152 | + QUARTERROUND(x3, x4, x9, x14) |
| 153 | + } |
| 154 | + x0 = PLUS(x0, j0); |
| 155 | + x1 = PLUS(x1, j1); |
| 156 | + x2 = PLUS(x2, j2); |
| 157 | + x3 = PLUS(x3, j3); |
| 158 | + x4 = PLUS(x4, j4); |
| 159 | + x5 = PLUS(x5, j5); |
| 160 | + x6 = PLUS(x6, j6); |
| 161 | + x7 = PLUS(x7, j7); |
| 162 | + x8 = PLUS(x8, j8); |
| 163 | + x9 = PLUS(x9, j9); |
| 164 | + x10 = PLUS(x10, j10); |
| 165 | + x11 = PLUS(x11, j11); |
| 166 | + x12 = PLUS(x12, j12); |
| 167 | + x13 = PLUS(x13, j13); |
| 168 | + x14 = PLUS(x14, j14); |
| 169 | + x15 = PLUS(x15, j15); |
| 170 | + |
| 171 | + x0 = XOR(x0, LOAD32_LE(m + 0)); |
| 172 | + x1 = XOR(x1, LOAD32_LE(m + 4)); |
| 173 | + x2 = XOR(x2, LOAD32_LE(m + 8)); |
| 174 | + x3 = XOR(x3, LOAD32_LE(m + 12)); |
| 175 | + x4 = XOR(x4, LOAD32_LE(m + 16)); |
| 176 | + x5 = XOR(x5, LOAD32_LE(m + 20)); |
| 177 | + x6 = XOR(x6, LOAD32_LE(m + 24)); |
| 178 | + x7 = XOR(x7, LOAD32_LE(m + 28)); |
| 179 | + x8 = XOR(x8, LOAD32_LE(m + 32)); |
| 180 | + x9 = XOR(x9, LOAD32_LE(m + 36)); |
| 181 | + x10 = XOR(x10, LOAD32_LE(m + 40)); |
| 182 | + x11 = XOR(x11, LOAD32_LE(m + 44)); |
| 183 | + x12 = XOR(x12, LOAD32_LE(m + 48)); |
| 184 | + x13 = XOR(x13, LOAD32_LE(m + 52)); |
| 185 | + x14 = XOR(x14, LOAD32_LE(m + 56)); |
| 186 | + x15 = XOR(x15, LOAD32_LE(m + 60)); |
| 187 | + |
| 188 | + j12 = PLUSONE(j12); |
| 189 | + /* LCOV_EXCL_START */ |
| 190 | + if (!j12) { |
| 191 | + j13 = PLUSONE(j13); |
| 192 | + } |
| 193 | + /* LCOV_EXCL_STOP */ |
| 194 | + |
| 195 | + STORE32_LE(c + 0, x0); |
| 196 | + STORE32_LE(c + 4, x1); |
| 197 | + STORE32_LE(c + 8, x2); |
| 198 | + STORE32_LE(c + 12, x3); |
| 199 | + STORE32_LE(c + 16, x4); |
| 200 | + STORE32_LE(c + 20, x5); |
| 201 | + STORE32_LE(c + 24, x6); |
| 202 | + STORE32_LE(c + 28, x7); |
| 203 | + STORE32_LE(c + 32, x8); |
| 204 | + STORE32_LE(c + 36, x9); |
| 205 | + STORE32_LE(c + 40, x10); |
| 206 | + STORE32_LE(c + 44, x11); |
| 207 | + STORE32_LE(c + 48, x12); |
| 208 | + STORE32_LE(c + 52, x13); |
| 209 | + STORE32_LE(c + 56, x14); |
| 210 | + STORE32_LE(c + 60, x15); |
| 211 | + |
| 212 | + if (bytes <= 64) { |
| 213 | + if (bytes < 64) { |
| 214 | + for (i = 0; i < (unsigned int) bytes; ++i) { |
| 215 | + ctarget[i] = c[i]; /* ctarget cannot be NULL */ |
| 216 | + } |
| 217 | + } |
| 218 | + ctx->input[12] = j12; |
| 219 | + ctx->input[13] = j13; |
| 220 | + |
| 221 | + return; |
| 222 | + } |
| 223 | + bytes -= 64; |
| 224 | + c += 64; |
| 225 | + m += 64; |
| 226 | + } |
| 227 | +} |
| 228 | + |
| 229 | +void chacha20_encrypt_msg(void *msg, size_t len, const void *nonce, const void *key) |
| 230 | +{ |
| 231 | + struct chacha_ctx ctx; |
| 232 | + |
| 233 | + chacha_keysetup(&ctx, key); |
| 234 | + chacha_ivsetup(&ctx, nonce, NULL); |
| 235 | + chacha20_encrypt_bytes(&ctx, msg, msg, len); |
| 236 | +} |
0 commit comments