1 /* $OpenBSD$ */ 2 /* 3 * Public Domain poly1305 from Andrew Moon 4 * Based on poly1305-donna.c, poly1305-donna-32.h and poly1305-donna.h from: 5 * https://github.com/floodyberry/poly1305-donna 6 */ 7 8 #include <stddef.h> 9 10 static inline void poly1305_init(poly1305_context *ctx, 11 const unsigned char key[32]); 12 static inline void poly1305_update(poly1305_context *ctx, 13 const unsigned char *m, size_t bytes); 14 static inline void poly1305_finish(poly1305_context *ctx, 15 unsigned char mac[16]); 16 17 /* 18 * poly1305 implementation using 32 bit * 32 bit = 64 bit multiplication 19 * and 64 bit addition. 20 */ 21 22 #define poly1305_block_size 16 23 24 /* 17 + sizeof(size_t) + 14*sizeof(unsigned long) */ 25 typedef struct poly1305_state_internal_t { 26 unsigned long r[5]; 27 unsigned long h[5]; 28 unsigned long pad[4]; 29 size_t leftover; 30 unsigned char buffer[poly1305_block_size]; 31 unsigned char final; 32 } poly1305_state_internal_t; 33 34 /* interpret four 8 bit unsigned integers as a 32 bit unsigned integer in little endian */ 35 static unsigned long 36 U8TO32(const unsigned char *p) 37 { 38 return (((unsigned long)(p[0] & 0xff)) | 39 ((unsigned long)(p[1] & 0xff) << 8) | 40 ((unsigned long)(p[2] & 0xff) << 16) | 41 ((unsigned long)(p[3] & 0xff) << 24)); 42 } 43 44 /* store a 32 bit unsigned integer as four 8 bit unsigned integers in little endian */ 45 static void 46 U32TO8(unsigned char *p, unsigned long v) 47 { 48 p[0] = (v) & 0xff; 49 p[1] = (v >> 8) & 0xff; 50 p[2] = (v >> 16) & 0xff; 51 p[3] = (v >> 24) & 0xff; 52 } 53 54 static inline void 55 poly1305_init(poly1305_context *ctx, const unsigned char key[32]) 56 { 57 poly1305_state_internal_t *st = (poly1305_state_internal_t *)ctx; 58 59 /* r &= 0xffffffc0ffffffc0ffffffc0fffffff */ 60 st->r[0] = (U8TO32(&key[0])) & 0x3ffffff; 61 st->r[1] = (U8TO32(&key[3]) >> 2) & 0x3ffff03; 62 st->r[2] = (U8TO32(&key[6]) >> 4) & 0x3ffc0ff; 63 st->r[3] = (U8TO32(&key[9]) >> 6) & 0x3f03fff; 64 st->r[4] = (U8TO32(&key[12]) >> 8) & 0x00fffff; 65 66 /* h = 0 */ 67 st->h[0] = 0; 68 st->h[1] = 0; 69 st->h[2] = 0; 70 st->h[3] = 0; 71 st->h[4] = 0; 72 73 /* save pad for later */ 74 st->pad[0] = U8TO32(&key[16]); 75 st->pad[1] = U8TO32(&key[20]); 76 st->pad[2] = U8TO32(&key[24]); 77 st->pad[3] = U8TO32(&key[28]); 78 79 st->leftover = 0; 80 st->final = 0; 81 } 82 83 static void 84 poly1305_blocks(poly1305_state_internal_t *st, const unsigned char *m, size_t bytes) 85 { 86 const unsigned long hibit = (st->final) ? 0 : (1 << 24); /* 1 << 128 */ 87 unsigned long r0, r1, r2, r3, r4; 88 unsigned long s1, s2, s3, s4; 89 unsigned long h0, h1, h2, h3, h4; 90 unsigned long long d0, d1, d2, d3, d4; 91 unsigned long c; 92 93 r0 = st->r[0]; 94 r1 = st->r[1]; 95 r2 = st->r[2]; 96 r3 = st->r[3]; 97 r4 = st->r[4]; 98 99 s1 = r1 * 5; 100 s2 = r2 * 5; 101 s3 = r3 * 5; 102 s4 = r4 * 5; 103 104 h0 = st->h[0]; 105 h1 = st->h[1]; 106 h2 = st->h[2]; 107 h3 = st->h[3]; 108 h4 = st->h[4]; 109 110 while (bytes >= poly1305_block_size) { 111 /* h += m[i] */ 112 h0 += (U8TO32(m + 0)) & 0x3ffffff; 113 h1 += (U8TO32(m + 3) >> 2) & 0x3ffffff; 114 h2 += (U8TO32(m + 6) >> 4) & 0x3ffffff; 115 h3 += (U8TO32(m + 9) >> 6) & 0x3ffffff; 116 h4 += (U8TO32(m + 12) >> 8) | hibit; 117 118 /* h *= r */ 119 d0 = ((unsigned long long)h0 * r0) + 120 ((unsigned long long)h1 * s4) + 121 ((unsigned long long)h2 * s3) + 122 ((unsigned long long)h3 * s2) + 123 ((unsigned long long)h4 * s1); 124 d1 = ((unsigned long long)h0 * r1) + 125 ((unsigned long long)h1 * r0) + 126 ((unsigned long long)h2 * s4) + 127 ((unsigned long long)h3 * s3) + 128 ((unsigned long long)h4 * s2); 129 d2 = ((unsigned long long)h0 * r2) + 130 ((unsigned long long)h1 * r1) + 131 ((unsigned long long)h2 * r0) + 132 ((unsigned long long)h3 * s4) + 133 ((unsigned long long)h4 * s3); 134 d3 = ((unsigned long long)h0 * r3) + 135 ((unsigned long long)h1 * r2) + 136 ((unsigned long long)h2 * r1) + 137 ((unsigned long long)h3 * r0) + 138 ((unsigned long long)h4 * s4); 139 d4 = ((unsigned long long)h0 * r4) + 140 ((unsigned long long)h1 * r3) + 141 ((unsigned long long)h2 * r2) + 142 ((unsigned long long)h3 * r1) + 143 ((unsigned long long)h4 * r0); 144 145 /* (partial) h %= p */ 146 c = (unsigned long)(d0 >> 26); 147 h0 = (unsigned long)d0 & 0x3ffffff; 148 d1 += c; 149 c = (unsigned long)(d1 >> 26); 150 h1 = (unsigned long)d1 & 0x3ffffff; 151 d2 += c; 152 c = (unsigned long)(d2 >> 26); 153 h2 = (unsigned long)d2 & 0x3ffffff; 154 d3 += c; 155 c = (unsigned long)(d3 >> 26); 156 h3 = (unsigned long)d3 & 0x3ffffff; 157 d4 += c; 158 c = (unsigned long)(d4 >> 26); 159 h4 = (unsigned long)d4 & 0x3ffffff; 160 h0 += c * 5; 161 c = (h0 >> 26); 162 h0 = h0 & 0x3ffffff; 163 h1 += c; 164 165 m += poly1305_block_size; 166 bytes -= poly1305_block_size; 167 } 168 169 st->h[0] = h0; 170 st->h[1] = h1; 171 st->h[2] = h2; 172 st->h[3] = h3; 173 st->h[4] = h4; 174 } 175 176 static inline void 177 poly1305_update(poly1305_context *ctx, const unsigned char *m, size_t bytes) 178 { 179 poly1305_state_internal_t *st = (poly1305_state_internal_t *)ctx; 180 size_t i; 181 182 /* handle leftover */ 183 if (st->leftover) { 184 size_t want = (poly1305_block_size - st->leftover); 185 if (want > bytes) 186 want = bytes; 187 for (i = 0; i < want; i++) 188 st->buffer[st->leftover + i] = m[i]; 189 bytes -= want; 190 m += want; 191 st->leftover += want; 192 if (st->leftover < poly1305_block_size) 193 return; 194 poly1305_blocks(st, st->buffer, poly1305_block_size); 195 st->leftover = 0; 196 } 197 198 /* process full blocks */ 199 if (bytes >= poly1305_block_size) { 200 size_t want = (bytes & ~(poly1305_block_size - 1)); 201 poly1305_blocks(st, m, want); 202 m += want; 203 bytes -= want; 204 } 205 206 /* store leftover */ 207 if (bytes) { 208 for (i = 0; i < bytes; i++) 209 st->buffer[st->leftover + i] = m[i]; 210 st->leftover += bytes; 211 } 212 } 213 214 static inline void 215 poly1305_finish(poly1305_context *ctx, unsigned char mac[16]) 216 { 217 poly1305_state_internal_t *st = (poly1305_state_internal_t *)ctx; 218 unsigned long h0, h1, h2, h3, h4, c; 219 unsigned long g0, g1, g2, g3, g4; 220 unsigned long long f; 221 unsigned long mask; 222 223 /* process the remaining block */ 224 if (st->leftover) { 225 size_t i = st->leftover; 226 st->buffer[i++] = 1; 227 for (; i < poly1305_block_size; i++) 228 st->buffer[i] = 0; 229 st->final = 1; 230 poly1305_blocks(st, st->buffer, poly1305_block_size); 231 } 232 233 /* fully carry h */ 234 h0 = st->h[0]; 235 h1 = st->h[1]; 236 h2 = st->h[2]; 237 h3 = st->h[3]; 238 h4 = st->h[4]; 239 240 c = h1 >> 26; 241 h1 = h1 & 0x3ffffff; 242 h2 += c; 243 c = h2 >> 26; 244 h2 = h2 & 0x3ffffff; 245 h3 += c; 246 c = h3 >> 26; 247 h3 = h3 & 0x3ffffff; 248 h4 += c; 249 c = h4 >> 26; 250 h4 = h4 & 0x3ffffff; 251 h0 += c * 5; 252 c = h0 >> 26; 253 h0 = h0 & 0x3ffffff; 254 h1 += c; 255 256 /* compute h + -p */ 257 g0 = h0 + 5; 258 c = g0 >> 26; 259 g0 &= 0x3ffffff; 260 g1 = h1 + c; 261 c = g1 >> 26; 262 g1 &= 0x3ffffff; 263 g2 = h2 + c; 264 c = g2 >> 26; 265 g2 &= 0x3ffffff; 266 g3 = h3 + c; 267 c = g3 >> 26; 268 g3 &= 0x3ffffff; 269 g4 = h4 + c - (1 << 26); 270 271 /* select h if h < p, or h + -p if h >= p */ 272 mask = (g4 >> ((sizeof(unsigned long) * 8) - 1)) - 1; 273 g0 &= mask; 274 g1 &= mask; 275 g2 &= mask; 276 g3 &= mask; 277 g4 &= mask; 278 mask = ~mask; 279 h0 = (h0 & mask) | g0; 280 h1 = (h1 & mask) | g1; 281 h2 = (h2 & mask) | g2; 282 h3 = (h3 & mask) | g3; 283 h4 = (h4 & mask) | g4; 284 285 /* h = h % (2^128) */ 286 h0 = ((h0) | (h1 << 26)) & 0xffffffff; 287 h1 = ((h1 >> 6) | (h2 << 20)) & 0xffffffff; 288 h2 = ((h2 >> 12) | (h3 << 14)) & 0xffffffff; 289 h3 = ((h3 >> 18) | (h4 << 8)) & 0xffffffff; 290 291 /* mac = (h + pad) % (2^128) */ 292 f = (unsigned long long)h0 + st->pad[0]; 293 h0 = (unsigned long)f; 294 f = (unsigned long long)h1 + st->pad[1] + (f >> 32); 295 h1 = (unsigned long)f; 296 f = (unsigned long long)h2 + st->pad[2] + (f >> 32); 297 h2 = (unsigned long)f; 298 f = (unsigned long long)h3 + st->pad[3] + (f >> 32); 299 h3 = (unsigned long)f; 300 301 U32TO8(mac + 0, h0); 302 U32TO8(mac + 4, h1); 303 U32TO8(mac + 8, h2); 304 U32TO8(mac + 12, h3); 305 306 /* zero out the state */ 307 st->h[0] = 0; 308 st->h[1] = 0; 309 st->h[2] = 0; 310 st->h[3] = 0; 311 st->h[4] = 0; 312 st->r[0] = 0; 313 st->r[1] = 0; 314 st->r[2] = 0; 315 st->r[3] = 0; 316 st->r[4] = 0; 317 st->pad[0] = 0; 318 st->pad[1] = 0; 319 st->pad[2] = 0; 320 st->pad[3] = 0; 321 } 322