crypto/poly1305/poly1305.c

   1 /* ====================================================================
   2  * Copyright (c) 2015 The OpenSSL Project. All rights reserved.
   3  *
   4  * Rights for redistribution and usage in source and binary
   5  * forms are granted according to the OpenSSL license.
   6  */
   7
   8 #include <stdlib.h>
   9 #include <string.h>
  10
  11 #include "internal/poly1305.h"
  12
  13 typedef void (*poly1305_blocks_f) (void *ctx, const unsigned char *inp,
  14                                    size_t len, unsigned int padbit);
  15 typedef void (*poly1305_emit_f) (void *ctx, unsigned char mac[16],
  16                                  const unsigned int nonce[4]);
  17
  18 struct poly1305_context {
  19     double opaque[24];  /* large enough to hold internal state, declared
  20                          * 'double' to ensure at least 64-bit invariant
  21                          * alignment across all platforms and
  22                          * configurations */
  23     unsigned int nonce[4];
  24     unsigned char data[POLY1305_BLOCK_SIZE];
  25     size_t num;
  26     struct {
  27         poly1305_blocks_f blocks;
  28         poly1305_emit_f emit;
  29     } func;
  30 };
  31
  32 size_t Poly1305_ctx_size ()
  33 {
  34     return sizeof(struct poly1305_context);
  35 }
  36
  37 /* pick 32-bit unsigned integer in little endian order */
  38 static unsigned int U8TOU32(const unsigned char *p)
  39 {
  40     return (((unsigned int)(p[0] & 0xff)) |
  41             ((unsigned int)(p[1] & 0xff) << 8) |
  42             ((unsigned int)(p[2] & 0xff) << 16) |
  43             ((unsigned int)(p[3] & 0xff) << 24));
  44 }
  45
  46 /*
  47  * Implementations can be classified by amount of significant bits in
  48  * words making up the multi-precision value, or in other words radix
  49  * or base of numerical representation, e.g. base 2^64, base 2^32,
  50  * base 2^26. Complementary characteristic is how wide is the result of
  51  * multiplication of pair of digits, e.g. it would take 128 bits to
  52  * accommodate multiplication result in base 2^64 case. These are used
  53  * interchangeably. To describe implementation that is. But interface
  54  * is designed to isolate this so that low-level primitives implemented
  55  * in assembly can be self-contained/self-coherent.
  56  */
  57 #ifndef POLY1305_ASM
  58 /*
  59  * Even though there is __int128 reference implementation targeting
  60  * 64-bit platforms provided below, it's not obvious that it's optimal
  61  * choice for every one of them. Depending on instruction set overall
  62  * amount of instructions can be comparable to one in __int64
  63  * implementation. Amount of multiplication instructions would be lower,
  64  * but not necessarily overall. And in out-of-order execution context,
  65  * it is the latter that can be crucial...
  66  *
  67  * On related note. Poly1305 author, D. J. Bernstein, discusses and
  68  * provides floating-point implementations of the algorithm in question.
  69  * It made a lot of sense by the time of introduction, because most
  70  * then-modern processors didn't have pipelined integer multiplier.
  71  * [Not to mention that some had non-constant timing for integer
  72  * multiplications.] Floating-point instructions on the other hand could
  73  * be issued every cycle, which allowed to achieve better performance.
  74  * Nowadays, with SIMD and/or out-or-order execution, shared or
  75  * even emulated FPU, it's more complicated, and floating-point
  76  * implementation is not necessarily optimal choice in every situation,
  77  * rather contrary...
  78  *
  79  *                                              <appro@openssl.org>
  80  */
  81
  82 typedef unsigned int u32;
  83
  84 /*
  85  * poly1305_blocks processes a multiple of POLY1305_BLOCK_SIZE blocks
  86  * of |inp| no longer than |len|. Behaviour for |len| not divisible by
  87  * block size is unspecified in general case, even though in reference
  88  * implementation the trailing chunk is simply ignored. Per algorithm
  89  * specification, every input block, complete or last partial, is to be
  90  * padded with a bit past most significant byte. The latter kind is then
  91  * padded with zeros till block size. This last partial block padding
  92  * is caller(*)'s responsibility, and because of this the last partial
  93  * block is always processed with separate call with |len| set to
  94  * POLY1305_BLOCK_SIZE and |padbit| to 0. In all other cases |padbit|
  95  * should be set to 1 to perform implicit padding with 128th bit.
  96  * poly1305_blocks does not actually check for this constraint though,
  97  * it's caller(*)'s resposibility to comply.
  98  *
  99  * (*)  In the context "caller" is not application code, but higher
 100  *      level Poly1305_* from this very module, so that quirks are
 101  *      handled locally.
 102  */
 103 static void
 104 poly1305_blocks(void *ctx, const unsigned char *inp, size_t len, u32 padbit);
 105
 106 /*
 107  * Type-agnostic "rip-off" from constant_time_locl.h
 108  */
 109 # define CONSTANT_TIME_CARRY(a,b) ( \
 110          (a ^ ((a ^ b) | ((a - b) ^ b))) >> (sizeof(a) * 8 - 1) \
 111          )
 112
 113 # if !defined(PEDANTIC) && \
 114      (defined(__SIZEOF_INT128__) && __SIZEOF_INT128__==16) && \
 115      (defined(__SIZEOF_LONG__) && __SIZEOF_LONG__==8)
 116
 117 typedef unsigned long u64;
 118 typedef unsigned __int128 u128;
 119
 120 typedef struct {
 121     u64 h[3];
 122     u64 r[2];
 123 } poly1305_internal;
 124
 125 /* pick 32-bit unsigned integer in little endian order */
 126 static u64 U8TOU64(const unsigned char *p)
 127 {
 128     return (((u64)(p[0] & 0xff)) |
 129             ((u64)(p[1] & 0xff) << 8) |
 130             ((u64)(p[2] & 0xff) << 16) |
 131             ((u64)(p[3] & 0xff) << 24) |
 132             ((u64)(p[4] & 0xff) << 32) |
 133             ((u64)(p[5] & 0xff) << 40) |
 134             ((u64)(p[6] & 0xff) << 48) |
 135             ((u64)(p[7] & 0xff) << 56));
 136 }
 137
 138 /* store a 32-bit unsigned integer in little endian */
 139 static void U64TO8(unsigned char *p, u64 v)
 140 {
 141     p[0] = (unsigned char)((v) & 0xff);
 142     p[1] = (unsigned char)((v >> 8) & 0xff);
 143     p[2] = (unsigned char)((v >> 16) & 0xff);
 144     p[3] = (unsigned char)((v >> 24) & 0xff);
 145     p[4] = (unsigned char)((v >> 32) & 0xff);
 146     p[5] = (unsigned char)((v >> 40) & 0xff);
 147     p[6] = (unsigned char)((v >> 48) & 0xff);
 148     p[7] = (unsigned char)((v >> 56) & 0xff);
 149 }
 150
 151 static void poly1305_init(void *ctx, const unsigned char key[16])
 152 {
 153     poly1305_internal *st = (poly1305_internal *) ctx;
 154
 155     /* h = 0 */
 156     st->h[0] = 0;
 157     st->h[1] = 0;
 158     st->h[2] = 0;
 159
 160     /* r &= 0xffffffc0ffffffc0ffffffc0fffffff */
 161     st->r[0] = U8TOU64(&key[0]) & 0x0ffffffc0fffffff;
 162     st->r[1] = U8TOU64(&key[8]) & 0x0ffffffc0ffffffc;
 163 }
 164
 165 static void
 166 poly1305_blocks(void *ctx, const unsigned char *inp, size_t len, u32 padbit)
 167 {
 168     poly1305_internal *st = (poly1305_internal *)ctx;
 169     u64 r0, r1;
 170     u64 s1;
 171     u64 h0, h1, h2, c;
 172     u128 d0, d1;
 173
 174     r0 = st->r[0];
 175     r1 = st->r[1];
 176
 177     s1 = r1 + (r1 >> 2);
 178
 179     h0 = st->h[0];
 180     h1 = st->h[1];
 181     h2 = st->h[2];
 182
 183     while (len >= POLY1305_BLOCK_SIZE) {
 184         /* h += m[i] */
 185         h0 = (u64)(d0 = (u128)h0 + U8TOU64(inp + 0));
 186         h1 = (u64)(d1 = (u128)h1 + (d0 >> 64) + U8TOU64(inp + 8));
 187         /*
 188          * padbit can be zero only when original len was
 189          * POLY1306_BLOCK_SIZE, but we don't check
 190          */
 191         h2 += (u64)(d1 >> 64) + padbit;
 192
 193         /* h *= r "%" p, where "%" stands for "partial remainder" */
 194         d0 = ((u128)h0 * r0) +
 195              ((u128)h1 * s1);
 196         d1 = ((u128)h0 * r1) +
 197              ((u128)h1 * r0) +
 198              (h2 * s1);
 199         h2 = (h2 * r0);
 200
 201         /* last reduction step: */
 202         /* a) h2:h0 = h2<<128 + d1<<64 + d0 */
 203         h0 = (u64)d0;
 204         h1 = (u64)(d1 += d0 >> 64);
 205         h2 += (u64)(d1 >> 64);
 206         /* b) (h2:h0 += (h2:h0>>130) * 5) %= 2^130 */
 207         c = (h2 >> 2) + (h2 & ~3UL);
 208         h2 &= 3;
 209         h0 += c;
 210         h1 += (c = CONSTANT_TIME_CARRY(h0,c));   /* doesn't overflow */
 211
 212         inp += POLY1305_BLOCK_SIZE;
 213         len -= POLY1305_BLOCK_SIZE;
 214     }
 215
 216     st->h[0] = h0;
 217     st->h[1] = h1;
 218     st->h[2] = h2;
 219 }
 220
 221 static void poly1305_emit(void *ctx, unsigned char mac[16],
 222                           const u32 nonce[4])
 223 {
 224     poly1305_internal *st = (poly1305_internal *) ctx;
 225     u64 h0, h1, h2;
 226     u64 g0, g1, g2;
 227     u128 t;
 228     u64 mask;
 229
 230     h0 = st->h[0];
 231     h1 = st->h[1];
 232     h2 = st->h[2];
 233
 234     /* compute h + -p */
 235     g0 = (u64)(t = (u128)h0 + 5);
 236     g1 = (u64)(t = (u128)h1 + (t >> 64));
 237     g2 = h2 + (u64)(t >> 64);
 238
 239     /* if there was carry into 130th bit, h1:h0 = g1:g0 */
 240     mask = 0 - (g2 >> 2);
 241     g0 &= mask;
 242     g1 &= mask;
 243     mask = ~mask;
 244     h0 = (h0 & mask) | g0;
 245     h1 = (h1 & mask) | g1;
 246
 247     /* mac = (h + nonce) % (2^128) */
 248     h0 = (u64)(t = (u128)h0 + nonce[0] + ((u64)nonce[1]<<32));
 249     h1 = (u64)(t = (u128)h1 + nonce[2] + ((u64)nonce[3]<<32) + (t >> 64));
 250
 251     U64TO8(mac + 0, h0);
 252     U64TO8(mac + 8, h1);
 253 }
 254
 255 # else
 256
 257 #  if defined(_WIN32) && !defined(__MINGW32__)
 258 typedef unsigned __int64 u64;
 259 #  elif defined(__arch64__)
 260 typedef unsigned long u64;
 261 #  else
 262 typedef unsigned long long u64;
 263 #  endif
 264
 265 typedef struct {
 266     u32 h[5];
 267     u32 r[4];
 268 } poly1305_internal;
 269
 270 /* store a 32-bit unsigned integer in little endian */
 271 static void U32TO8(unsigned char *p, unsigned int v)
 272 {
 273     p[0] = (unsigned char)((v) & 0xff);
 274     p[1] = (unsigned char)((v >> 8) & 0xff);
 275     p[2] = (unsigned char)((v >> 16) & 0xff);
 276     p[3] = (unsigned char)((v >> 24) & 0xff);
 277 }
 278
 279 static void poly1305_init(void *ctx, const unsigned char key[16])
 280 {
 281     poly1305_internal *st = (poly1305_internal *) ctx;
 282
 283     /* h = 0 */
 284     st->h[0] = 0;
 285     st->h[1] = 0;
 286     st->h[2] = 0;
 287     st->h[3] = 0;
 288     st->h[4] = 0;
 289
 290     /* r &= 0xffffffc0ffffffc0ffffffc0fffffff */
 291     st->r[0] = U8TOU32(&key[0]) & 0x0fffffff;
 292     st->r[1] = U8TOU32(&key[4]) & 0x0ffffffc;
 293     st->r[2] = U8TOU32(&key[8]) & 0x0ffffffc;
 294     st->r[3] = U8TOU32(&key[12]) & 0x0ffffffc;
 295 }
 296
 297 static void
 298 poly1305_blocks(void *ctx, const unsigned char *inp, size_t len, u32 padbit)
 299 {
 300     poly1305_internal *st = (poly1305_internal *)ctx;
 301     u32 r0, r1, r2, r3;
 302     u32 s1, s2, s3;
 303     u32 h0, h1, h2, h3, h4, c;
 304     u64 d0, d1, d2, d3;
 305
 306     r0 = st->r[0];
 307     r1 = st->r[1];
 308     r2 = st->r[2];
 309     r3 = st->r[3];
 310
 311     s1 = r1 + (r1 >> 2);
 312     s2 = r2 + (r2 >> 2);
 313     s3 = r3 + (r3 >> 2);
 314
 315     h0 = st->h[0];
 316     h1 = st->h[1];
 317     h2 = st->h[2];
 318     h3 = st->h[3];
 319     h4 = st->h[4];
 320
 321     while (len >= POLY1305_BLOCK_SIZE) {
 322         /* h += m[i] */
 323         h0 = (u32)(d0 = (u64)h0 + U8TOU32(inp + 0));
 324         h1 = (u32)(d1 = (u64)h1 + (d0 >> 32) + U8TOU32(inp + 4));
 325         h2 = (u32)(d2 = (u64)h2 + (d1 >> 32) + U8TOU32(inp + 8));
 326         h3 = (u32)(d3 = (u64)h3 + (d2 >> 32) + U8TOU32(inp + 12));
 327         h4 += (u32)(d3 >> 32) + padbit;
 328
 329         /* h *= r "%" p, where "%" stands for "partial remainder" */
 330         d0 = ((u64)h0 * r0) +
 331              ((u64)h1 * s3) +
 332              ((u64)h2 * s2) +
 333              ((u64)h3 * s1);
 334         d1 = ((u64)h0 * r1) +
 335              ((u64)h1 * r0) +
 336              ((u64)h2 * s3) +
 337              ((u64)h3 * s2) +
 338              (h4 * s1);
 339         d2 = ((u64)h0 * r2) +
 340              ((u64)h1 * r1) +
 341              ((u64)h2 * r0) +
 342              ((u64)h3 * s3) +
 343              (h4 * s2);
 344         d3 = ((u64)h0 * r3) +
 345              ((u64)h1 * r2) +
 346              ((u64)h2 * r1) +
 347              ((u64)h3 * r0) +
 348              (h4 * s3);
 349         h4 = (h4 * r0);
 350
 351         /* last reduction step: */
 352         /* a) h4:h0 = h4<<128 + d3<<96 + d2<<64 + d1<<32 + d0 */
 353         h0 = (u32)d0;
 354         h1 = (u32)(d1 += d0 >> 32);
 355         h2 = (u32)(d2 += d1 >> 32);
 356         h3 = (u32)(d3 += d2 >> 32);
 357         h4 += (u32)(d3 >> 32);
 358         /* b) (h4:h0 += (h4:h0>>130) * 5) %= 2^130 */
 359         c = (h4 >> 2) + (h4 & ~3U);
 360         h4 &= 3;
 361         h0 += c;
 362         h1 += (c = CONSTANT_TIME_CARRY(h0,c));
 363         h2 += (c = CONSTANT_TIME_CARRY(h1,c));
 364         h3 += (c = CONSTANT_TIME_CARRY(h2,c));   /* doesn't overflow */
 365
 366         inp += POLY1305_BLOCK_SIZE;
 367         len -= POLY1305_BLOCK_SIZE;
 368     }
 369
 370     st->h[0] = h0;
 371     st->h[1] = h1;
 372     st->h[2] = h2;
 373     st->h[3] = h3;
 374     st->h[4] = h4;
 375 }
 376
 377 static void poly1305_emit(void *ctx, unsigned char mac[16],
 378                           const u32 nonce[4])
 379 {
 380     poly1305_internal *st = (poly1305_internal *) ctx;
 381     u32 h0, h1, h2, h3, h4;
 382     u32 g0, g1, g2, g3, g4;
 383     u64 t;
 384     u32 mask;
 385
 386     h0 = st->h[0];
 387     h1 = st->h[1];
 388     h2 = st->h[2];
 389     h3 = st->h[3];
 390     h4 = st->h[4];
 391
 392     /* compute h + -p */
 393     g0 = (u32)(t = (u64)h0 + 5);
 394     g1 = (u32)(t = (u64)h1 + (t >> 32));
 395     g2 = (u32)(t = (u64)h2 + (t >> 32));
 396     g3 = (u32)(t = (u64)h3 + (t >> 32));
 397     g4 = h4 + (u32)(t >> 32);
 398
 399     /* if there was carry into 130th bit, h3:h0 = g3:g0 */
 400     mask = 0 - (g4 >> 2);
 401     g0 &= mask;
 402     g1 &= mask;
 403     g2 &= mask;
 404     g3 &= mask;
 405     mask = ~mask;
 406     h0 = (h0 & mask) | g0;
 407     h1 = (h1 & mask) | g1;
 408     h2 = (h2 & mask) | g2;
 409     h3 = (h3 & mask) | g3;
 410
 411     /* mac = (h + nonce) % (2^128) */
 412     h0 = (u32)(t = (u64)h0 + nonce[0]);
 413     h1 = (u32)(t = (u64)h1 + (t >> 32) + nonce[1]);
 414     h2 = (u32)(t = (u64)h2 + (t >> 32) + nonce[2]);
 415     h3 = (u32)(t = (u64)h3 + (t >> 32) + nonce[3]);
 416
 417     U32TO8(mac + 0, h0);
 418     U32TO8(mac + 4, h1);
 419     U32TO8(mac + 8, h2);
 420     U32TO8(mac + 12, h3);
 421 }
 422 # endif
 423 #else
 424 int poly1305_init(void *ctx, const unsigned char key[16], void *func);
 425 void poly1305_blocks(void *ctx, const unsigned char *inp, size_t len,
 426                      unsigned int padbit);
 427 void poly1305_emit(void *ctx, unsigned char mac[16],
 428                    const unsigned int nonce[4]);
 429 #endif
 430
 431 void Poly1305_Init(POLY1305 *ctx, const unsigned char key[32])
 432 {
 433     ctx->nonce[0] = U8TOU32(&key[16]);
 434     ctx->nonce[1] = U8TOU32(&key[20]);
 435     ctx->nonce[2] = U8TOU32(&key[24]);
 436     ctx->nonce[3] = U8TOU32(&key[28]);
 437
 438 #ifndef POLY1305_ASM
 439     poly1305_init(ctx->opaque, key);
 440 #else
 441     /*
 442      * Unlike reference poly1305_init assembly counterpart is expected
 443      * to return a value: non-zero if it initializes ctx->func, and zero
 444      * otherwise. Latter is to simplify assembly in cases when there no
 445      * multiple code paths to switch between.
 446      */
 447     if (!poly1305_init(ctx->opaque, key, &ctx->func)) {
 448         ctx->func.blocks = poly1305_blocks;
 449         ctx->func.emit = poly1305_emit;
 450     }
 451 #endif
 452
 453     ctx->num = 0;
 454
 455 }
 456
 457 void Poly1305_Update(POLY1305 *ctx, const unsigned char *inp, size_t len)
 458 {
 459 #ifdef POLY1305_ASM
 460     /*
 461      * As documented, poly1305_blocks is never called with input
 462      * longer than single block and padbit argument set to 0. This
 463      * property is fluently used in assembly modules to optimize
 464      * padbit handling on loop boundary.
 465      */
 466     poly1305_blocks_f poly1305_blocks = ctx->func.blocks;
 467 #endif
 468     size_t rem, num;
 469
 470     if ((num = ctx->num)) {
 471         rem = POLY1305_BLOCK_SIZE - num;
 472         if (len >= rem) {
 473             memcpy(ctx->data + num, inp, rem);
 474             poly1305_blocks(ctx->opaque, ctx->data, POLY1305_BLOCK_SIZE, 1);
 475             inp += rem;
 476             len -= rem;
 477         } else {
 478             /* Still not enough data to process a block. */
 479             memcpy(ctx->data + num, inp, len);
 480             ctx->num = num + len;
 481             return;
 482         }
 483     }
 484
 485     rem = len % POLY1305_BLOCK_SIZE;
 486     len -= rem;
 487
 488     if (len >= POLY1305_BLOCK_SIZE) {
 489         poly1305_blocks(ctx->opaque, inp, len, 1);
 490         inp += len;
 491     }
 492
 493     if (rem)
 494         memcpy(ctx->data, inp, rem);
 495
 496     ctx->num = rem;
 497 }
 498
 499 void Poly1305_Final(POLY1305 *ctx, unsigned char mac[16])
 500 {
 501 #ifdef POLY1305_ASM
 502     poly1305_blocks_f poly1305_blocks = ctx->func.blocks;
 503     poly1305_emit_f poly1305_emit = ctx->func.emit;
 504 #endif
 505     size_t num;
 506
 507     if ((num = ctx->num)) {
 508         ctx->data[num++] = 1;   /* pad bit */
 509         while (num < POLY1305_BLOCK_SIZE)
 510             ctx->data[num++] = 0;
 511         poly1305_blocks(ctx->opaque, ctx->data, POLY1305_BLOCK_SIZE, 0);
 512     }
 513
 514     poly1305_emit(ctx->opaque, mac, ctx->nonce);
 515
 516     /* zero out the state */
 517     memset(ctx, 0, sizeof(*ctx));
 518 }
 519
 520 #ifdef SELFTEST
 521 #include <stdio.h>
 522
 523 struct poly1305_test {
 524     const char *inputhex;
 525     const char *keyhex;
 526     const char *outhex;
 527 };
 528
 529 static const struct poly1305_test poly1305_tests[] = {
 530     /*
 531      * RFC7539
 532      */
 533     {
 534      "43727970746f6772617068696320466f72756d2052657365617263682047726f"
 535      "7570",
 536      "85d6be7857556d337f4452fe42d506a8""0103808afb0db2fd4abff6af4149f51b",
 537      "a8061dc1305136c6c22b8baf0c0127a9"
 538     },
 539     /*
 540      * test vectors from "The Poly1305-AES message-authentication code"
 541      */
 542     {
 543      "f3f6",
 544      "851fc40c3467ac0be05cc20404f3f700""580b3b0f9447bb1e69d095b5928b6dbc",
 545      "f4c633c3044fc145f84f335cb81953de"
 546     },
 547     {
 548      "",
 549      "a0f3080000f46400d0c7e9076c834403""dd3fab2251f11ac759f0887129cc2ee7",
 550      "dd3fab2251f11ac759f0887129cc2ee7"
 551     },
 552     {
 553      "663cea190ffb83d89593f3f476b6bc24d7e679107ea26adb8caf6652d0656136",
 554      "48443d0bb0d21109c89a100b5ce2c208""83149c69b561dd88298a1798b10716ef",
 555      "0ee1c16bb73f0f4fd19881753c01cdbe"
 556     },
 557     {
 558      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 559      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9",
 560      "12976a08c4426d0ce8a82407c4f48207""80f8c20aa71202d1e29179cbcb555a57",
 561      "5154ad0d2cb26e01274fc51148491f1b"
 562     },
 563     /*
 564      * self-generated
 565      */
 566     {
 567      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 568      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9af",
 569      "12976a08c4426d0ce8a82407c4f48207""80f8c20aa71202d1e29179cbcb555a57",
 570      "812059a5da198637cac7c4a631bee466"
 571     },
 572     {
 573      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 574      "990c62e48b8018b2c3e4a0fa3134cb67",
 575      "12976a08c4426d0ce8a82407c4f48207""80f8c20aa71202d1e29179cbcb555a57",
 576      "5b88d7f6228b11e2e28579a5c0c1f761"
 577     },
 578     {
 579      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 580      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9af"
 581      "663cea190ffb83d89593f3f476b6bc24d7e679107ea26adb8caf6652d0656136",
 582      "12976a08c4426d0ce8a82407c4f48207""80f8c20aa71202d1e29179cbcb555a57",
 583      "bbb613b2b6d753ba07395b916aaece15"
 584     },
 585     {
 586      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 587      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9af"
 588      "48443d0bb0d21109c89a100b5ce2c20883149c69b561dd88298a1798b10716ef"
 589      "663cea190ffb83d89593f3f476b6bc24",
 590      "12976a08c4426d0ce8a82407c4f48207""80f8c20aa71202d1e29179cbcb555a57",
 591      "c794d7057d1778c4bbee0a39b3d97342"
 592     },
 593     {
 594      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 595      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9af"
 596      "48443d0bb0d21109c89a100b5ce2c20883149c69b561dd88298a1798b10716ef"
 597      "663cea190ffb83d89593f3f476b6bc24d7e679107ea26adb8caf6652d0656136",
 598      "12976a08c4426d0ce8a82407c4f48207""80f8c20aa71202d1e29179cbcb555a57",
 599      "ffbcb9b371423152d7fca5ad042fbaa9"
 600     },
 601     {
 602      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 603      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9af"
 604      "48443d0bb0d21109c89a100b5ce2c20883149c69b561dd88298a1798b10716ef"
 605      "663cea190ffb83d89593f3f476b6bc24d7e679107ea26adb8caf6652d0656136"
 606      "812059a5da198637cac7c4a631bee466",
 607      "12976a08c4426d0ce8a82407c4f48207""80f8c20aa71202d1e29179cbcb555a57",
 608      "069ed6b8ef0f207b3e243bb1019fe632"
 609     },
 610     {
 611      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 612      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9af"
 613      "48443d0bb0d21109c89a100b5ce2c20883149c69b561dd88298a1798b10716ef"
 614      "663cea190ffb83d89593f3f476b6bc24d7e679107ea26adb8caf6652d0656136"
 615      "812059a5da198637cac7c4a631bee4665b88d7f6228b11e2e28579a5c0c1f761",
 616      "12976a08c4426d0ce8a82407c4f48207""80f8c20aa71202d1e29179cbcb555a57",
 617      "cca339d9a45fa2368c2c68b3a4179133"
 618     },
 619     {
 620      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 621      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9af"
 622      "48443d0bb0d21109c89a100b5ce2c20883149c69b561dd88298a1798b10716ef"
 623      "663cea190ffb83d89593f3f476b6bc24d7e679107ea26adb8caf6652d0656136"
 624      "812059a5da198637cac7c4a631bee4665b88d7f6228b11e2e28579a5c0c1f761"
 625      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 626      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9af"
 627      "48443d0bb0d21109c89a100b5ce2c20883149c69b561dd88298a1798b10716ef"
 628      "663cea190ffb83d89593f3f476b6bc24d7e679107ea26adb8caf6652d0656136",
 629      "12976a08c4426d0ce8a82407c4f48207""80f8c20aa71202d1e29179cbcb555a57",
 630      "53f6e828a2f0fe0ee815bf0bd5841a34"
 631     },
 632     {
 633      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 634      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9af"
 635      "48443d0bb0d21109c89a100b5ce2c20883149c69b561dd88298a1798b10716ef"
 636      "663cea190ffb83d89593f3f476b6bc24d7e679107ea26adb8caf6652d0656136"
 637      "812059a5da198637cac7c4a631bee4665b88d7f6228b11e2e28579a5c0c1f761"
 638      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 639      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9af"
 640      "48443d0bb0d21109c89a100b5ce2c20883149c69b561dd88298a1798b10716ef"
 641      "663cea190ffb83d89593f3f476b6bc24d7e679107ea26adb8caf6652d0656136"
 642      "812059a5da198637cac7c4a631bee4665b88d7f6228b11e2e28579a5c0c1f761",
 643      "12976a08c4426d0ce8a82407c4f48207""80f8c20aa71202d1e29179cbcb555a57",
 644      "b846d44e9bbd53cedffbfbb6b7fa4933"
 645     },
 646     {
 647     /*
 648      * poly1305_ieee754.c failed this in final stage
 649      */
 650      "842364e156336c0998b933a6237726180d9e3fdcbde4cd5d17080fc3beb49614"
 651      "d7122c037463ff104d73f19c12704628d417c4c54a3fe30d3c3d7714382d43b0"
 652      "382a50a5dee54be844b076e8df88201a1cd43b90eb21643fa96f39b518aa8340"
 653      "c942ff3c31baf7c9bdbf0f31ae3fa096bf8c63030609829fe72e179824890bc8"
 654      "e08c315c1cce2a83144dbbff09f74e3efc770b54d0984a8f19b14719e6363564"
 655      "1d6b1eedf63efbf080e1783d32445412114c20de0b837a0dfa33d6b82825fff4"
 656      "4c9a70ea54ce47f07df698e6b03323b53079364a5fc3e9dd034392bdde86dccd"
 657      "da94321c5e44060489336cb65bf3989c36f7282c2f5d2b882c171e74",
 658      "95d5c005503e510d8cd0aa072c4a4d06""6eabc52d11653df47fbf63ab198bcc26",
 659      "f248312e578d9d58f8b7bb4d19105431"
 660     },
 661     /*
 662      * test vectors from Google
 663      */
 664     {
 665      "",
 666      "c8afaac331ee372cd6082de134943b17""4710130e9f6fea8d72293850a667d86c",
 667      "4710130e9f6fea8d72293850a667d86c",
 668     },
 669     {
 670      "48656c6c6f20776f726c6421",
 671      "746869732069732033322d6279746520""6b657920666f7220506f6c7931333035",
 672      "a6f745008f81c916a20dcc74eef2b2f0"
 673     },
 674     {
 675      "0000000000000000000000000000000000000000000000000000000000000000",
 676      "746869732069732033322d6279746520""6b657920666f7220506f6c7931333035",
 677      "49ec78090e481ec6c26b33b91ccc0307"
 678     },
 679     /*
 680      * test vectors from Andrew Moon
 681      */
 682     {   /* nacl */
 683      "8e993b9f48681273c29650ba32fc76ce48332ea7164d96a4476fb8c531a1186a"
 684      "c0dfc17c98dce87b4da7f011ec48c97271d2c20f9b928fe2270d6fb863d51738"
 685      "b48eeee314a7cc8ab932164548e526ae90224368517acfeabd6bb3732bc0e9da"
 686      "99832b61ca01b6de56244a9e88d5f9b37973f622a43d14a6599b1f654cb45a74"
 687      "e355a5",
 688      "eea6a7251c1e72916d11c2cb214d3c25""2539121d8e234e652d651fa4c8cff880",
 689      "f3ffc7703f9400e52a7dfb4b3d3305d9"
 690     },
 691     {   /* wrap 2^130-5 */
 692      "ffffffffffffffffffffffffffffffff",
 693      "02000000000000000000000000000000""00000000000000000000000000000000",
 694      "03000000000000000000000000000000"
 695     },
 696     {   /* wrap 2^128 */
 697      "02000000000000000000000000000000",
 698      "02000000000000000000000000000000""ffffffffffffffffffffffffffffffff",
 699      "03000000000000000000000000000000"
 700     },
 701     {   /* limb carry */
 702      "fffffffffffffffffffffffffffffffff0ffffffffffffffffffffffffffffff"
 703      "11000000000000000000000000000000",
 704      "01000000000000000000000000000000""00000000000000000000000000000000",
 705      "05000000000000000000000000000000"
 706     },
 707     {   /* 2^130-5 */
 708      "fffffffffffffffffffffffffffffffffbfefefefefefefefefefefefefefefe"
 709      "01010101010101010101010101010101",
 710      "01000000000000000000000000000000""00000000000000000000000000000000",
 711      "00000000000000000000000000000000"
 712     },
 713     {   /* 2^130-6 */
 714      "fdffffffffffffffffffffffffffffff",
 715      "02000000000000000000000000000000""00000000000000000000000000000000",
 716      "faffffffffffffffffffffffffffffff"
 717     },
 718     {   /* 5*H+L reduction intermediate */
 719      "e33594d7505e43b900000000000000003394d7505e4379cd0100000000000000"
 720      "0000000000000000000000000000000001000000000000000000000000000000",
 721      "01000000000000000400000000000000""00000000000000000000000000000000",
 722      "14000000000000005500000000000000"
 723     },
 724     {   /* 5*H+L reduction final */
 725      "e33594d7505e43b900000000000000003394d7505e4379cd0100000000000000"
 726      "00000000000000000000000000000000",
 727      "01000000000000000400000000000000""00000000000000000000000000000000",
 728      "13000000000000000000000000000000"
 729     }
 730 };
 731
 732 static unsigned char hex_digit(char h)
 733 {
 734     if (h >= '0' && h <= '9')
 735         return h - '0';
 736     else if (h >= 'a' && h <= 'f')
 737         return h - 'a' + 10;
 738     else if (h >= 'A' && h <= 'F')
 739         return h - 'A' + 10;
 740     else
 741         abort();
 742 }
 743
 744 static void hex_decode(unsigned char *out, const char *hex)
 745 {
 746     size_t j = 0;
 747
 748     while (*hex != 0) {
 749         unsigned char v = hex_digit(*hex++);
 750         v <<= 4;
 751         v |= hex_digit(*hex++);
 752         out[j++] = v;
 753     }
 754 }
 755
 756 static void hexdump(unsigned char *a, size_t len)
 757 {
 758     size_t i;
 759
 760     for (i = 0; i < len; i++)
 761         printf("%02x", a[i]);
 762 }
 763
 764 int main()
 765 {
 766     static const unsigned num_tests =
 767         sizeof(poly1305_tests) / sizeof(struct poly1305_test);
 768     unsigned i;
 769     unsigned char key[32], out[16], expected[16];
 770     POLY1305 poly1305;
 771
 772     for (i = 0; i < num_tests; i++) {
 773         const struct poly1305_test *test = &poly1305_tests[i];
 774         unsigned char *in;
 775         size_t inlen = strlen(test->inputhex);
 776
 777         if (strlen(test->keyhex) != sizeof(key) * 2 ||
 778             strlen(test->outhex) != sizeof(out) * 2 || (inlen & 1) == 1)
 779             return 1;
 780
 781         inlen /= 2;
 782
 783         hex_decode(key, test->keyhex);
 784         hex_decode(expected, test->outhex);
 785
 786         in = malloc(inlen);
 787
 788         hex_decode(in, test->inputhex);
 789
 790         Poly1305_Init(&poly1305, key);
 791         Poly1305_Update(&poly1305, in, inlen);
 792         Poly1305_Final(&poly1305, out);
 793
 794         if (memcmp(out, expected, sizeof(expected)) != 0) {
 795             printf("Poly1305 test #%d failed.\n", i);
 796             printf("got:      ");
 797             hexdump(out, sizeof(out));
 798             printf("\nexpected: ");
 799             hexdump(expected, sizeof(expected));
 800             printf("\n");
 801             return 1;
 802         }
 803
 804         if (inlen > 16) {
 805             Poly1305_Init(&poly1305, key);
 806             Poly1305_Update(&poly1305, in, 1);
 807             Poly1305_Update(&poly1305, in+1, inlen-1);
 808             Poly1305_Final(&poly1305, out);
 809
 810             if (memcmp(out, expected, sizeof(expected)) != 0) {
 811                 printf("Poly1305 test #%d/1+(N-1) failed.\n", i);
 812                 printf("got:      ");
 813                 hexdump(out, sizeof(out));
 814                 printf("\nexpected: ");
 815                 hexdump(expected, sizeof(expected));
 816                 printf("\n");
 817                 return 1;
 818             }
 819         }
 820
 821         if (inlen > 32) {
 822             size_t half = inlen / 2;
 823
 824             Poly1305_Init(&poly1305, key);
 825             Poly1305_Update(&poly1305, in, half);
 826             Poly1305_Update(&poly1305, in+half, inlen-half);
 827             Poly1305_Final(&poly1305, out);
 828
 829             if (memcmp(out, expected, sizeof(expected)) != 0) {
 830                 printf("Poly1305 test #%d/2 failed.\n", i);
 831                 printf("got:      ");
 832                 hexdump(out, sizeof(out));
 833                 printf("\nexpected: ");
 834                 hexdump(expected, sizeof(expected));
 835                 printf("\n");
 836                 return 1;
 837             }
 838         }
 839
 840         free(in);
 841     }
 842
 843     printf("PASS\n");
 844
 845 # ifdef OPENSSL_CPUID_OBJ
 846     {
 847         unsigned char buf[8192];
 848         unsigned long long stopwatch;
 849         unsigned long long OPENSSL_rdtsc();
 850
 851         memset (buf,0x55,sizeof(buf));
 852         memset (key,0xAA,sizeof(key));
 853
 854         Poly1305_Init(&poly1305, key);
 855
 856         for (i=0;i<100000;i++)
 857             Poly1305_Update(&poly1305,buf,sizeof(buf));
 858
 859         stopwatch = OPENSSL_rdtsc();
 860         for (i=0;i<10000;i++)
 861             Poly1305_Update(&poly1305,buf,sizeof(buf));
 862         stopwatch = OPENSSL_rdtsc() - stopwatch;
 863
 864         printf("%g\n",stopwatch/(double)(i*sizeof(buf)));
 865
 866         stopwatch = OPENSSL_rdtsc();
 867         for (i=0;i<10000;i++) {
 868             Poly1305_Init(&poly1305, key);
 869             Poly1305_Update(&poly1305,buf,16);
 870             Poly1305_Final(&poly1305,buf);
 871         }
 872         stopwatch = OPENSSL_rdtsc() - stopwatch;
 873
 874         printf("%g\n",stopwatch/(double)(i));
 875     }
 876 # endif
 877     return 0;
 878 }
 879 #endif