chiark - git - mdw - catacomb/blob - symm/salsa20.c

   1 /* -*-c-*-
   2  *
   3  * Salsa20 stream cipher
   4  *
   5  * (c) 2015 Straylight/Edgeware
   6  */
   7
   8 /*----- Licensing notice --------------------------------------------------*
   9  *
  10  * This file is part of Catacomb.
  11  *
  12  * Catacomb is free software; you can redistribute it and/or modify
  13  * it under the terms of the GNU Library General Public License as
  14  * published by the Free Software Foundation; either version 2 of the
  15  * License, or (at your option) any later version.
  16  *
  17  * Catacomb is distributed in the hope that it will be useful,
  18  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  20  * GNU Library General Public License for more details.
  21  *
  22  * You should have received a copy of the GNU Library General Public
  23  * License along with Catacomb; if not, write to the Free
  24  * Software Foundation, Inc., 59 Temple Place - Suite 330, Boston,
  25  * MA 02111-1307, USA.
  26  */
  27
  28 /*----- Header files ------------------------------------------------------*/
  29
  30 #include "config.h"
  31
  32 #include <stdarg.h>
  33
  34 #include <mLib/bits.h>
  35
  36 #include "arena.h"
  37 #include "dispatch.h"
  38 #include "gcipher.h"
  39 #include "grand.h"
  40 #include "keysz.h"
  41 #include "paranoia.h"
  42 #include "salsa20.h"
  43 #include "salsa20-core.h"
  44
  45 /*----- Global variables --------------------------------------------------*/
  46
  47 const octet salsa20_keysz[] = { KSZ_SET, 32, 16, 10, 0 };
  48
  49 /*----- The Salsa20 core function and utilities ---------------------------*/
  50
  51 /* --- @core@ --- *
  52  *
  53  * Arguments:   @unsigned r@ = number of rounds
  54  *              @const salsa20_matrix src@ = input matrix
  55  *              @salsa20_matrix dest@ = where to put the output
  56  *
  57  * Returns:     ---
  58  *
  59  *
  60  * Use:         Apply the Salsa20/r core function to @src@, writing the
  61  *              result to @dest@.  This consists of @r@ rounds followed by
  62  *              the feedforward step.
  63  */
  64
  65 CPU_DISPATCH(static, (void), void, core,
  66              (unsigned r, const salsa20_matrix src, salsa20_matrix dest),
  67              (r, src, dest), pick_core, simple_core);
  68
  69 static void simple_core(unsigned r, const salsa20_matrix src,
  70                         salsa20_matrix dest)
  71   { SALSA20_nR(dest, src, r); SALSA20_FFWD(dest, src); }
  72
  73 #if CPUFAM_X86 || CPUFAM_AMD64
  74 extern core__functype salsa20_core_x86ish_sse2;
  75 extern core__functype salsa20_core_x86ish_avx;
  76 #endif
  77
  78 #if CPUFAM_ARMEL
  79 extern core__functype salsa20_core_arm_neon;
  80 #endif
  81
  82 #if CPUFAM_ARM64
  83 extern core__functype salsa20_core_arm64;
  84 #endif
  85
  86 static core__functype *pick_core(void)
  87 {
  88 #if CPUFAM_X86 || CPUFAM_AMD64
  89   DISPATCH_PICK_COND(salsa20_core, salsa20_core_x86ish_avx,
  90                      cpu_feature_p(CPUFEAT_X86_AVX));
  91   DISPATCH_PICK_COND(salsa20_core, salsa20_core_x86ish_sse2,
  92                      cpu_feature_p(CPUFEAT_X86_SSE2));
  93 #endif
  94 #if CPUFAM_ARMEL
  95   DISPATCH_PICK_COND(salsa20_core, salsa20_core_arm_neon,
  96                      cpu_feature_p(CPUFEAT_ARM_NEON));
  97 #endif
  98 #if CPUFAM_ARM64
  99   DISPATCH_PICK_COND(salsa20_core, salsa20_core_arm64, 1);
 100 #endif
 101   DISPATCH_PICK_FALLBACK(salsa20_core, simple_core);
 102 }
 103
 104 /* --- @populate@ --- *
 105  *
 106  * Arguments:   @salsa20_matrix a@ = a matrix to fill in
 107  *              @const void *key@ = pointer to key material
 108  *              @size_t ksz@ = size of key
 109  *
 110  * Returns:     ---
 111  *
 112  * Use:         Fills in a Salsa20 matrix from the key, setting the
 113  *              appropriate constants according to the key length.  The nonce
 114  *              and position words are left uninitialized.
 115  */
 116
 117 static void populate(salsa20_matrix a, const void *key, size_t ksz)
 118 {
 119   const octet *k = key;
 120
 121   KSZ_ASSERT(salsa20, ksz);
 122
 123   /* Here's the pattern of key, constant, nonce, and counter pieces in the
 124    * matrix, before and after our permutation.
 125    *
 126    * [ C0 K0 K1 K2 ]       [ C0 C1 C2 C3 ]
 127    * [ K3 C1 N0 N1 ]  -->  [ K3 T1 K7 K2 ]
 128    * [ T0 T1 C2 K4 ]       [ T0 K6 K1 N1 ]
 129    * [ K5 K6 K7 C3 ]       [ K5 K0 N0 K4 ]
 130    */
 131
 132   a[13] = LOAD32_L(k +  0);
 133   a[10] = LOAD32_L(k +  4);
 134   if (ksz == 10) {
 135     a[ 7] = LOAD16_L(k +  8);
 136     a[ 4] = 0;
 137   } else {
 138     a[ 7] = LOAD32_L(k +  8);
 139     a[ 4] = LOAD32_L(k + 12);
 140   }
 141   if (ksz <= 16) {
 142     a[15] = a[13];
 143     a[12] = a[10];
 144     a[ 9] = a[ 7];
 145     a[ 6] = a[ 4];
 146     a[ 0] = SALSA20_A128;
 147     a[ 1] = SALSA20_B128;
 148     a[ 2] = ksz == 10 ? SALSA20_C80 : SALSA20_C128;
 149     a[ 3] = SALSA20_D128;
 150   } else {
 151     a[15] = LOAD32_L(k + 16);
 152     a[12] = LOAD32_L(k + 20);
 153     a[ 9] = LOAD32_L(k + 24);
 154     a[ 6] = LOAD32_L(k + 28);
 155     a[ 0] = SALSA20_A256;
 156     a[ 1] = SALSA20_B256;
 157     a[ 2] = SALSA20_C256;
 158     a[ 3] = SALSA20_D256;
 159   }
 160 }
 161
 162 /*----- Salsa20 implementation --------------------------------------------*/
 163
 164 /* --- @salsa20_init@ --- *
 165  *
 166  * Arguments:   @salsa20_ctx *ctx@ = context to fill in
 167  *              @const void *key@ = pointer to key material
 168  *              @size_t ksz@ = size of key (either 32 or 16)
 169  *              @const void *nonce@ = initial nonce, or null
 170  *
 171  * Returns:     ---
 172  *
 173  * Use:         Initializes a Salsa20 context ready for use.
 174  */
 175
 176 void salsa20_init(salsa20_ctx *ctx, const void *key, size_t ksz,
 177                   const void *nonce)
 178 {
 179   static const octet zerononce[SALSA20_NONCESZ];
 180
 181   populate(ctx->a, key, ksz);
 182   salsa20_setnonce(ctx, nonce ? nonce : zerononce);
 183 }
 184
 185 /* --- @salsa20_setnonce{,_ietf}@ --- *
 186  *
 187  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 188  *              @const void *nonce@ = the nonce (@SALSA20_NONCESZ@ or
 189  *                      @SALSA20_IETF_NONCESZ@ bytes)
 190  *
 191  * Returns:     ---
 192  *
 193  * Use:         Set a new nonce in the context @ctx@, e.g., for processing a
 194  *              different message.  The stream position is reset to zero (see
 195  *              @salsa20_seek@ etc.).
 196  */
 197
 198 void salsa20_setnonce(salsa20_ctx *ctx, const void *nonce)
 199 {
 200   const octet *n = nonce;
 201
 202   ctx->a[14] = LOAD32_L(n + 0);
 203   ctx->a[11] = LOAD32_L(n + 4);
 204   salsa20_seek(ctx, 0);
 205 }
 206
 207 void salsa20_setnonce_ietf(salsa20_ctx *ctx, const void *nonce)
 208 {
 209   const octet *n = nonce;
 210
 211   ctx->a[ 5] = LOAD32_L(n + 0);
 212   ctx->a[14] = LOAD32_L(n + 4);
 213   ctx->a[11] = LOAD32_L(n + 8);
 214   salsa20_seek_ietf(ctx, 0);
 215 }
 216
 217 /* --- @salsa20_seek{,u64,_ietf}@ --- *
 218  *
 219  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 220  *              @unsigned long i@, @kludge64 i@, @uint32@ = new position
 221  *
 222  * Returns:     ---
 223  *
 224  * Use:         Sets a new stream position, in units of Salsa20 output
 225  *              blocks, which are @SALSA20_OUTSZ@ bytes each.  Byte
 226  *              granularity can be achieved by calling @salsa20R_encrypt@
 227  *              appropriately.
 228  */
 229
 230 void salsa20_seek(salsa20_ctx *ctx, unsigned long i)
 231   { kludge64 ii; ASSIGN64(ii, i); salsa20_seeku64(ctx, ii); }
 232
 233 void salsa20_seeku64(salsa20_ctx *ctx, kludge64 i)
 234 {
 235   ctx->a[8] = LO64(i); ctx->a[5] = HI64(i);
 236   ctx->off = SALSA20_OUTSZ;
 237 }
 238
 239 void salsa20_seek_ietf(salsa20_ctx *ctx, uint32 i)
 240   { ctx->a[8] = i; }
 241
 242 /* --- @salsa20_tell{,u64,_ietf}@ --- *
 243  *
 244  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 245  *
 246  * Returns:     The current position in the output stream, in blocks,
 247  *              rounding upwards.
 248  */
 249
 250 unsigned long salsa20_tell(salsa20_ctx *ctx)
 251   { kludge64 i = salsa20_tellu64(ctx); return (GET64(unsigned long, i)); }
 252
 253 kludge64 salsa20_tellu64(salsa20_ctx *ctx)
 254   { kludge64 i; SET64(i, ctx->a[5], ctx->a[8]); return (i); }
 255
 256 uint32 salsa20_tell_ietf(salsa20_ctx *ctx)
 257   { return (ctx->a[5]); }
 258
 259 /* --- @salsa20{,12,8}_encrypt@ --- *
 260  *
 261  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 262  *              @const void *src@ = source buffer (or null)
 263  *              @void *dest@ = destination buffer (or null)
 264  *              @size_t sz@ = size of the buffers
 265  *
 266  * Returns:     ---
 267  *
 268  * Use:         Encrypts or decrypts @sz@ bytes of data from @src@ to @dest@.
 269  *              Salsa20 works by XORing plaintext with a keystream, so
 270  *              encryption and decryption are the same operation.  If @dest@
 271  *              is null then ignore @src@ and skip @sz@ bytes of the
 272  *              keystream.  If @src@ is null, then just write the keystream
 273  *              to @dest@.
 274  */
 275
 276 #define SALSA20_ENCRYPT(r, ctx, src, dest, sz)                          \
 277   SALSA20_DECOR(salsa20, r, _encrypt)(ctx, src, dest, sz)
 278 #define DEFENCRYPT(r)                                                   \
 279   void SALSA20_ENCRYPT(r, salsa20_ctx *ctx, const void *src,            \
 280                        void *dest, size_t sz)                           \
 281   {                                                                     \
 282     salsa20_matrix b;                                                   \
 283     const octet *s = src;                                               \
 284     octet *d = dest;                                                    \
 285     size_t n;                                                           \
 286     kludge64 pos, delta;                                                \
 287                                                                         \
 288     SALSA20_OUTBUF(ctx, d, s, sz);                                      \
 289     if (!sz) return;                                                    \
 290                                                                         \
 291     if (!dest) {                                                        \
 292       n = sz/SALSA20_OUTSZ;                                             \
 293       pos = salsa20_tellu64(ctx);                                       \
 294       ASSIGN64(delta, n);                                               \
 295       ADD64(pos, pos, delta);                                           \
 296       salsa20_seeku64(ctx, pos);                                        \
 297       sz = sz%SALSA20_OUTSZ;                                            \
 298     } else if (!src) {                                                  \
 299       while (sz >= SALSA20_OUTSZ) {                                     \
 300         core(r, ctx->a, b);                                             \
 301         SALSA20_STEP(ctx->a);                                           \
 302         SALSA20_GENFULL(b, d);                                          \
 303         sz -= SALSA20_OUTSZ;                                            \
 304       }                                                                 \
 305     } else {                                                            \
 306       while (sz >= SALSA20_OUTSZ) {                                     \
 307         core(r, ctx->a, b);                                             \
 308         SALSA20_STEP(ctx->a);                                           \
 309         SALSA20_MIXFULL(b, d, s);                                       \
 310         sz -= SALSA20_OUTSZ;                                            \
 311       }                                                                 \
 312     }                                                                   \
 313                                                                         \
 314     if (sz) {                                                           \
 315       core(r, ctx->a, b);                                               \
 316       SALSA20_STEP(ctx->a);                                             \
 317       SALSA20_PREPBUF(ctx, b);                                          \
 318       SALSA20_OUTBUF(ctx, d, s, sz);                                    \
 319       assert(!sz);                                                      \
 320     }                                                                   \
 321   }
 322 SALSA20_VARS(DEFENCRYPT)
 323
 324 /*----- HSalsa20 implementation -------------------------------------------*/
 325
 326 #define HSALSA20_RAW(r, ctx, src, dest)                                 \
 327   SALSA20_DECOR(hsalsa20, r, _raw)(ctx, src, dest)
 328 #define HSALSA20_PRF(r, ctx, src, dest)                                 \
 329   SALSA20_DECOR(hsalsa20, r, _prf)(ctx, src, dest)
 330
 331 /* --- @hsalsa20{,12,8}_prf@ --- *
 332  *
 333  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 334  *              @const void *src@ = the input (@HSALSA20_INSZ@ bytes)
 335  *              @void *dest@ = the output (@HSALSA20_OUTSZ@ bytes)
 336  *
 337  * Returns:     ---
 338  *
 339  * Use:         Apply the HSalsa20/r pseudorandom function to @src@, writing
 340  *              the result to @out@.
 341  */
 342
 343 #define DEFHSALSA20(r)                                                  \
 344   static void HSALSA20_RAW(r, salsa20_matrix k,                         \
 345                            const uint32 *src, uint32 *dest)             \
 346   {                                                                     \
 347     salsa20_matrix a;                                                   \
 348     int i;                                                              \
 349                                                                         \
 350     /* --- HSalsa20, computed from full Salsa20 --- *                   \
 351      *                                                                  \
 352      * The security proof makes use of the fact that HSalsa20 (i.e.,    \
 353      * without the final feedforward step) can be computed from full    \
 354      * Salsa20 using only knowledge of the non-secret input.  I don't   \
 355      * want to compromise the performance of the main function by       \
 356      * making the feedforward step separate, but this operation is less \
 357      * speed critical, so we do it the harder way.                      \
 358      */                                                                 \
 359                                                                         \
 360     for (i = 0; i < 4; i++) k[14 - 3*i] = src[i];                       \
 361     core(r, k, a);                                                      \
 362     for (i = 0; i < 4; i++) dest[i] = a[5*i] - k[i];                    \
 363     for (i = 4; i < 8; i++) dest[i] = a[i + 2] - k[26 - 3*i];           \
 364   }                                                                     \
 365                                                                         \
 366   void HSALSA20_PRF(r, salsa20_ctx *ctx, const void *src, void *dest)   \
 367   {                                                                     \
 368     const octet *s = src;                                               \
 369     octet *d = dest;                                                    \
 370     uint32 in[4], out[8];                                               \
 371     int i;                                                              \
 372                                                                         \
 373     for (i = 0; i < 4; i++) in[i] = LOAD32_L(s + 4*i);                  \
 374     HSALSA20_RAW(r, ctx->a, in, out);                                   \
 375     for (i = 0; i < 8; i++) STORE32_L(d + 4*i, out[i]);                 \
 376   }
 377 SALSA20_VARS(DEFHSALSA20)
 378
 379 /*----- XSalsa20 implementation -------------------------------------------*/
 380
 381 /* --- Some convenient macros for naming functions --- *
 382  *
 383  * Because the crypto core is involved in XSalsa20/r's per-nonce setup, we
 384  * need to take an interest in the number of rounds in most of the various
 385  * functions, and it will probably help if we distinguish the context
 386  * structures for the various versions.
 387  */
 388
 389 #define XSALSA20_CTX(r) SALSA20_DECOR(xsalsa20, r, _ctx)
 390 #define XSALSA20_INIT(r, ctx, k, ksz, n)                                \
 391   SALSA20_DECOR(xsalsa20, r, _init)(ctx, k, ksz, n)
 392 #define XSALSA20_SETNONCE(r, ctx, n)                                    \
 393   SALSA20_DECOR(xsalsa20, r, _setnonce)(ctx, n)
 394 #define XSALSA20_SEEK(r, ctx, i)                                        \
 395   SALSA20_DECOR(xsalsa20, r, _seek)(ctx, i)
 396 #define XSALSA20_SEEKU64(r, ctx, i)                                     \
 397   SALSA20_DECOR(xsalsa20, r, _seeku64)(ctx, i)
 398 #define XSALSA20_TELL(r, ctx)                                           \
 399   SALSA20_DECOR(xsalsa20, r, _tell)(ctx)
 400 #define XSALSA20_TELLU64(r, ctx)                                        \
 401   SALSA20_DECOR(xsalsa20, r, _tellu64)(ctx)
 402 #define XSALSA20_ENCRYPT(r, ctx, src, dest, sz)                         \
 403   SALSA20_DECOR(xsalsa20, r, _encrypt)(ctx, src, dest, sz)
 404
 405 /* --- @xsalsa20{,12,8}_init@ --- *
 406  *
 407  * Arguments:   @xsalsa20R_ctx *ctx@ = the context to fill in
 408  *              @const void *key@ = pointer to key material
 409  *              @size_t ksz@ = size of key (either 32 or 16)
 410  *              @const void *nonce@ = initial nonce, or null
 411  *
 412  * Returns:     ---
 413  *
 414  * Use:         Initializes an XSalsa20/r context ready for use.
 415  *
 416  *              There is a different function for each number of rounds,
 417  *              unlike for plain Salsa20.
 418  */
 419
 420 #define DEFXINIT(r)                                                     \
 421   void XSALSA20_INIT(r, XSALSA20_CTX(r) *ctx,                           \
 422                         const void *key, size_t ksz, const void *nonce) \
 423   {                                                                     \
 424     static const octet zerononce[XSALSA20_NONCESZ];                     \
 425                                                                         \
 426     populate(ctx->k, key, ksz);                                         \
 427     ctx->s.a[ 0] = SALSA20_A256;                                        \
 428     ctx->s.a[ 1] = SALSA20_B256;                                        \
 429     ctx->s.a[ 2] = SALSA20_C256;                                        \
 430     ctx->s.a[ 3] = SALSA20_D256;                                        \
 431     XSALSA20_SETNONCE(r, ctx, nonce ? nonce : zerononce);               \
 432   }
 433 SALSA20_VARS(DEFXINIT)
 434
 435 /* --- @xsalsa20{,12,8}_setnonce@ --- *
 436  *
 437  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 438  *              @const void *nonce@ = the nonce (@XSALSA20_NONCESZ@ bytes)
 439  *
 440  * Returns:     ---
 441  *
 442  * Use:         Set a new nonce in the context @ctx@, e.g., for processing a
 443  *              different message.  The stream position is reset to zero (see
 444  *              @salsa20_seek@ etc.).
 445  *
 446  *              There is a different function for each number of rounds,
 447  *              unlike for plain Salsa20.
 448  */
 449
 450 #define DEFXNONCE(r)                                                    \
 451   void XSALSA20_SETNONCE(r, XSALSA20_CTX(r) *ctx, const void *nonce)    \
 452   {                                                                     \
 453     const octet *n = nonce;                                             \
 454     uint32 in[4], out[8];                                               \
 455     int i;                                                              \
 456                                                                         \
 457     for (i = 0; i < 4; i++) in[i] = LOAD32_L(n + 4*i);                  \
 458     HSALSA20_RAW(r, ctx->k, in, out);                                   \
 459     for (i = 0; i < 4; i++) ctx->s.a[13 - 3*i] = out[i];                \
 460     for (i = 4; i < 8; i++) ctx->s.a[27 - 3*i] = out[i];                \
 461     salsa20_setnonce(&ctx->s, n + 16);                                  \
 462   }
 463 SALSA20_VARS(DEFXNONCE)
 464
 465 /* --- @xsalsa20{,12,8}_seek{,u64}@ --- *
 466  *
 467  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 468  *              @unsigned long i@, @kludge64 i@ = new position to set
 469  *
 470  * Returns:     ---
 471  *
 472  * Use:         Sets a new stream position, in units of Salsa20 output
 473  *              blocks, which are @XSALSA20_OUTSZ@ bytes each.  Byte
 474  *              granularity can be achieved by calling @xsalsa20R_encrypt@
 475  *              appropriately.
 476  *
 477  *              There is a different function for each number of rounds,
 478  *              unlike for plain Salsa20, because the context structures are
 479  *              different.
 480  */
 481
 482 /* --- @xsalsa20{,12,8}_tell{,u64}@ --- *
 483  *
 484  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 485  *
 486  * Returns:     The current position in the output stream, in blocks,
 487  *              rounding upwards.
 488  *
 489  *              There is a different function for each number of rounds,
 490  *              unlike for plain Salsa20, because the context structures are
 491  *              different.
 492  */
 493
 494 /* --- @xsalsa20{,12,8}_encrypt@ --- *
 495  *
 496  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 497  *              @const void *src@ = source buffer (or null)
 498  *              @void *dest@ = destination buffer (or null)
 499  *              @size_t sz@ = size of the buffers
 500  *
 501  * Returns:     ---
 502  *
 503  * Use:         Encrypts or decrypts @sz@ bytes of data from @src@ to @dest@.
 504  *              XSalsa20 works by XORing plaintext with a keystream, so
 505  *              encryption and decryption are the same operation.  If @dest@
 506  *              is null then ignore @src@ and skip @sz@ bytes of the
 507  *              keystream.  If @src@ is null, then just write the keystream
 508  *              to @dest@.
 509  */
 510
 511 #define DEFXPASSTHRU(r)                                                 \
 512   void XSALSA20_SEEK(r, XSALSA20_CTX(r) *ctx, unsigned long i)          \
 513     { salsa20_seek(&ctx->s, i); }                                       \
 514   void XSALSA20_SEEKU64(r, XSALSA20_CTX(r) *ctx, kludge64 i)            \
 515     { salsa20_seeku64(&ctx->s, i); }                                    \
 516   unsigned long XSALSA20_TELL(r, XSALSA20_CTX(r) *ctx)                  \
 517     { return salsa20_tell(&ctx->s); }                                   \
 518   kludge64 XSALSA20_TELLU64(r, XSALSA20_CTX(r) *ctx)                    \
 519     { return salsa20_tellu64(&ctx->s); }                                \
 520   void XSALSA20_ENCRYPT(r, XSALSA20_CTX(r) *ctx,                        \
 521                         const void *src, void *dest, size_t sz)         \
 522     { SALSA20_ENCRYPT(r, &ctx->s, src, dest, sz); }
 523 SALSA20_VARS(DEFXPASSTHRU)
 524
 525 /*----- Generic cipher interface ------------------------------------------*/
 526
 527 typedef struct gctx { gcipher c; salsa20_ctx ctx; } gctx;
 528
 529 static void gsetiv(gcipher *c, const void *iv)
 530   { gctx *g = (gctx *)c; salsa20_setnonce(&g->ctx, iv); }
 531
 532 static void gsetiv_ietf(gcipher *c, const void *iv)
 533   { gctx *g = (gctx *)c; salsa20_setnonce_ietf(&g->ctx, iv); }
 534
 535 static void gdestroy(gcipher *c)
 536   { gctx *g = (gctx *)c; BURN(*g); S_DESTROY(g); }
 537
 538 static gcipher *ginit(const void *k, size_t sz, const gcipher_ops *ops)
 539 {
 540   gctx *g = S_CREATE(gctx);
 541   g->c.ops = ops;
 542   salsa20_init(&g->ctx, k, sz, 0);
 543   return (&g->c);
 544 }
 545
 546 #define DEFGCIPHER(r)                                                   \
 547                                                                         \
 548   static const gcipher_ops gops_##r, gops_##r##_ietf;                   \
 549                                                                         \
 550   static gcipher *ginit_##r(const void *k, size_t sz)                   \
 551     { return (ginit(k, sz, &gops_##r)); }                               \
 552                                                                         \
 553   static gcipher *ginit_##r##_ietf(const void *k, size_t sz)            \
 554     { return (ginit(k, sz, &gops_##r##_ietf)); }                        \
 555                                                                         \
 556   static void gencrypt_##r(gcipher *c, const void *s,                   \
 557                            void *t, size_t sz)                          \
 558     { gctx *g = (gctx *)c; SALSA20_ENCRYPT(r, &g->ctx, s, t, sz); }     \
 559                                                                         \
 560   static const gcipher_ops gops_##r = {                                 \
 561     &SALSA20_DECOR(salsa20, r, ),                                       \
 562     gencrypt_##r, gencrypt_##r, gdestroy, gsetiv, 0                     \
 563   };                                                                    \
 564                                                                         \
 565   static const gcipher_ops gops_##r##_ietf = {                          \
 566     &SALSA20_DECOR(salsa20, r, _ietf),                                  \
 567     gencrypt_##r, gencrypt_##r, gdestroy, gsetiv_ietf, 0                \
 568   };                                                                    \
 569                                                                         \
 570   const gccipher SALSA20_DECOR(salsa20, r, ) = {                        \
 571     SALSA20_NAME_##r, salsa20_keysz,                                    \
 572     SALSA20_NONCESZ, ginit_##r                                          \
 573   };                                                                    \
 574                                                                         \
 575   const gccipher SALSA20_DECOR(salsa20, r, _ietf) = {                   \
 576     SALSA20_NAME_##r "-ietf", salsa20_keysz,                            \
 577     SALSA20_IETF_NONCESZ, ginit_##r##_ietf                              \
 578   };
 579
 580 SALSA20_VARS(DEFGCIPHER)
 581
 582 #define DEFGXCIPHER(r)                                                  \
 583                                                                         \
 584   typedef struct { gcipher c; XSALSA20_CTX(r) ctx; } gxctx_##r;         \
 585                                                                         \
 586   static void gxsetiv_##r(gcipher *c, const void *iv)                   \
 587     { gxctx_##r *g = (gxctx_##r *)c; XSALSA20_SETNONCE(r, &g->ctx, iv); } \
 588                                                                         \
 589   static void gxdestroy_##r(gcipher *c)                                 \
 590     { gxctx_##r *g = (gxctx_##r *)c; BURN(*g); S_DESTROY(g); }          \
 591                                                                         \
 592   static const gcipher_ops gxops_##r;                                   \
 593                                                                         \
 594   static gcipher *gxinit_##r(const void *k, size_t sz)                  \
 595   {                                                                     \
 596     gxctx_##r *g = S_CREATE(gxctx_##r);                                 \
 597     g->c.ops = &gxops_##r;                                              \
 598     XSALSA20_INIT(r, &g->ctx, k, sz, 0);                                \
 599     return (&g->c);                                                     \
 600   }                                                                     \
 601                                                                         \
 602   static void gxencrypt_##r(gcipher *c, const void *s,                  \
 603                             void *t, size_t sz)                         \
 604   {                                                                     \
 605     gxctx_##r *g = (gxctx_##r *)c;                                      \
 606     XSALSA20_ENCRYPT(r, &g->ctx, s, t, sz);                             \
 607   }                                                                     \
 608                                                                         \
 609   static const gcipher_ops gxops_##r = {                                \
 610     &SALSA20_DECOR(xsalsa20, r, ),                                      \
 611     gxencrypt_##r, gxencrypt_##r, gxdestroy_##r, gxsetiv_##r, 0         \
 612   };                                                                    \
 613                                                                         \
 614   const gccipher SALSA20_DECOR(xsalsa20, r, ) = {                       \
 615     "x" SALSA20_NAME_##r, salsa20_keysz,                                \
 616     XSALSA20_NONCESZ, gxinit_##r                                        \
 617   };
 618
 619 SALSA20_VARS(DEFGXCIPHER)
 620
 621 /*----- Generic random number generator interface -------------------------*/
 622
 623 typedef struct grops {
 624   size_t noncesz;
 625   void (*seek)(void *, kludge64);
 626   kludge64 (*tell)(void *);
 627   void (*setnonce)(void *, const void *);
 628   void (*generate)(void *, void *, size_t);
 629 } grops;
 630
 631 typedef struct grbasectx {
 632   grand r;
 633   const grops *ops;
 634 } grbasectx;
 635
 636 static int grmisc(grand *r, unsigned op, ...)
 637 {
 638   octet buf[XSALSA20_NONCESZ];
 639   grbasectx *g = (grbasectx *)r;
 640   grand *rr;
 641   const octet *p;
 642   size_t sz;
 643   uint32 i;
 644   unsigned long ul;
 645   kludge64 pos;
 646   va_list ap;
 647   int rc = 0;
 648
 649   va_start(ap, op);
 650
 651   switch (op) {
 652     case GRAND_CHECK:
 653       switch (va_arg(ap, unsigned)) {
 654         case GRAND_CHECK:
 655         case GRAND_SEEDINT:
 656         case GRAND_SEEDUINT32:
 657         case GRAND_SEEDBLOCK:
 658         case GRAND_SEEDRAND:
 659         case SALSA20_SEEK:
 660         case SALSA20_SEEKU64:
 661         case SALSA20_TELL:
 662         case SALSA20_TELLU64:
 663           rc = 1;
 664           break;
 665         default:
 666           rc = 0;
 667           break;
 668       }
 669       break;
 670
 671     case GRAND_SEEDINT:
 672       i = va_arg(ap, unsigned); STORE32_L(buf, i);
 673       memset(buf + 4, 0, g->ops->noncesz - 4);
 674       g->ops->setnonce(g, buf);
 675       break;
 676     case GRAND_SEEDUINT32:
 677       i = va_arg(ap, uint32); STORE32_L(buf, i);
 678       memset(buf + 4, 0, g->ops->noncesz - 4);
 679       g->ops->setnonce(g, buf);
 680       break;
 681     case GRAND_SEEDBLOCK:
 682       p = va_arg(ap, const void *);
 683       sz = va_arg(ap, size_t);
 684       if (sz < g->ops->noncesz) {
 685         memcpy(buf, p, sz);
 686         memset(buf + sz, 0, g->ops->noncesz - sz);
 687         p = buf;
 688       }
 689       g->ops->setnonce(g, p);
 690       break;
 691     case GRAND_SEEDRAND:
 692       rr = va_arg(ap, grand *);
 693       rr->ops->fill(rr, buf, g->ops->noncesz);
 694       g->ops->setnonce(g, buf);
 695       break;
 696     case SALSA20_SEEK:
 697       ul = va_arg(ap, unsigned long); ASSIGN64(pos, ul);
 698       g->ops->seek(g, pos);
 699       break;
 700     case SALSA20_SEEKU64:
 701       pos = va_arg(ap, kludge64);
 702       g->ops->seek(g, pos);
 703       break;
 704     case SALSA20_TELL:
 705       pos = g->ops->tell(g);
 706       *va_arg(ap, unsigned long *) = GET64(unsigned long, pos);
 707       break;
 708     case SALSA20_TELLU64:
 709       *va_arg(ap, kludge64 *) = g->ops->tell(g);
 710       break;
 711     default:
 712       GRAND_BADOP;
 713       break;
 714   }
 715
 716   return (rc);
 717 }
 718
 719 static octet grbyte(grand *r)
 720 {
 721   grbasectx *g = (grbasectx *)r;
 722   octet o;
 723   g->ops->generate(g, &o, 1);
 724   return (o);
 725 }
 726
 727 static uint32 grword(grand *r)
 728 {
 729   grbasectx *g = (grbasectx *)r;
 730   octet b[4];
 731   g->ops->generate(g, b, sizeof(b));
 732   return (LOAD32_L(b));
 733 }
 734
 735 static void grfill(grand *r, void *p, size_t sz)
 736 {
 737   grbasectx *g = (grbasectx *)r;
 738   g->ops->generate(r, p, sz);
 739 }
 740
 741 typedef struct grctx {
 742   grbasectx r;
 743   salsa20_ctx ctx;
 744 } grctx;
 745
 746 static void gr_seek(void *r, kludge64 pos)
 747   { grctx *g = r; salsa20_seeku64(&g->ctx, pos); }
 748
 749 static void gr_seek_ietf(void *r, kludge64 pos)
 750   { grctx *g = r; salsa20_seek_ietf(&g->ctx, LO64(pos)); }
 751
 752 static kludge64 gr_tell(void *r)
 753   { grctx *g = r; return (salsa20_tellu64(&g->ctx)); }
 754
 755 static kludge64 gr_tell_ietf(void *r)
 756 {
 757   grctx *g = r;
 758   kludge64 pos;
 759
 760   SET64(pos, 0, salsa20_tell_ietf(&g->ctx));
 761   return (pos);
 762 }
 763
 764 static void gr_setnonce(void *r, const void *n)
 765   { grctx *g = r; salsa20_setnonce(&g->ctx, n); }
 766
 767 static void gr_setnonce_ietf(void *r, const void *n)
 768   { grctx *g = r; salsa20_setnonce(&g->ctx, n); }
 769
 770 static void grdestroy(grand *r)
 771   { grctx *g = (grctx *)r; BURN(*g); S_DESTROY(g); }
 772
 773 static grand *grinit(const void *k, size_t ksz, const void *n,
 774                      const grand_ops *ops, const grops *myops)
 775 {
 776     grctx *g = S_CREATE(grctx);
 777     g->r.r.ops = ops;
 778     g->r.ops = myops;
 779     salsa20_init(&g->ctx, k, ksz, 0);
 780     if (n) myops->setnonce(g, n);
 781     return (&g->r.r);
 782 }
 783
 784 #define DEFGRAND(rr)                                                    \
 785                                                                         \
 786   static void gr_generate_##rr(void *r, void *b, size_t sz)             \
 787     { grctx *g = r; SALSA20_ENCRYPT(rr, &g->ctx, 0, b, sz); }           \
 788                                                                         \
 789   static const grops grops_##rr =                                       \
 790     { SALSA20_NONCESZ, gr_seek, gr_tell,                                \
 791       gr_setnonce, gr_generate_##rr };                                  \
 792                                                                         \
 793   static const grops grops_##rr##_ietf =                                \
 794     { SALSA20_IETF_NONCESZ, gr_seek_ietf, gr_tell_ietf,                 \
 795       gr_setnonce_ietf, gr_generate_##rr };                             \
 796                                                                         \
 797   static const grand_ops grops_rand_##rr = {                            \
 798     SALSA20_NAME_##rr, GRAND_CRYPTO, 0,                                 \
 799     grmisc, grdestroy, grword,                                          \
 800     grbyte, grword, grand_defaultrange, grfill                          \
 801   };                                                                    \
 802                                                                         \
 803   static const grand_ops grops_rand_##rr##_ietf = {                     \
 804     SALSA20_NAME_##rr "-ietf", GRAND_CRYPTO, 0,                         \
 805     grmisc, grdestroy, grword,                                          \
 806     grbyte, grword, grand_defaultrange, grfill                          \
 807   };                                                                    \
 808                                                                         \
 809   grand *SALSA20_DECOR(salsa20, rr, _rand)                              \
 810     (const void *k, size_t ksz, const void *n)                          \
 811     { return (grinit(k, ksz, n, &grops_rand_##rr, &grops_##rr)); }      \
 812                                                                         \
 813   grand *SALSA20_DECOR(salsa20, rr, _ietf_rand)                         \
 814     (const void *k, size_t ksz, const void *n)                          \
 815   {                                                                     \
 816     return (grinit(k, ksz, n,                                           \
 817                    &grops_rand_##rr##_ietf,                             \
 818                    &grops_##rr##_ietf));                                \
 819   }
 820
 821 SALSA20_VARS(DEFGRAND)
 822
 823 #define DEFXGRAND(rr)                                                   \
 824                                                                         \
 825   typedef struct grxctx_##rr {                                          \
 826     grbasectx r;                                                        \
 827     XSALSA20_CTX(rr) ctx;                                               \
 828   } grxctx_##rr;                                                        \
 829                                                                         \
 830   static void grx_seek_##rr(void *r, kludge64 pos)                      \
 831     { grxctx_##rr *g = r; XSALSA20_SEEKU64(rr, &g->ctx, pos); }         \
 832                                                                         \
 833   static kludge64 grx_tell_##rr(void *r)                                \
 834     { grxctx_##rr *g = r; return (XSALSA20_TELLU64(rr, &g->ctx)); }     \
 835                                                                         \
 836   static void grx_setnonce_##rr(void *r, const void *n)                 \
 837     { grxctx_##rr *g = r; XSALSA20_SETNONCE(rr, &g->ctx, n); }          \
 838                                                                         \
 839   static void grxdestroy_##rr(grand *r)                                 \
 840     { grxctx_##rr *g = (grxctx_##rr *)r; BURN(*g); S_DESTROY(g); }      \
 841                                                                         \
 842   static void grx_generate_##rr(void *r, void *b, size_t sz)            \
 843     { grxctx_##rr *g = r; XSALSA20_ENCRYPT(rr, &g->ctx, 0, b, sz); }    \
 844                                                                         \
 845   static const grops grxops_##rr =                                      \
 846   { XSALSA20_NONCESZ, grx_seek_##rr, grx_tell_##rr,                     \
 847       grx_setnonce_##rr, grx_generate_##rr };                           \
 848                                                                         \
 849   static const grand_ops grxops_rand_##rr = {                           \
 850     "x" SALSA20_NAME_##rr, GRAND_CRYPTO, 0,                             \
 851     grmisc, grxdestroy_##rr, grword,                                    \
 852     grbyte, grword, grand_defaultrange, grfill                          \
 853   };                                                                    \
 854                                                                         \
 855   grand *SALSA20_DECOR(xsalsa20, rr, _rand)                             \
 856     (const void *k, size_t ksz, const void *n)                          \
 857   {                                                                     \
 858     grxctx_##rr *g = S_CREATE(grxctx_##rr);                             \
 859     g->r.r.ops = &grxops_rand_##rr;                                     \
 860     g->r.ops = &grxops_##rr;                                            \
 861     XSALSA20_INIT(rr, &g->ctx, k, ksz, n);                              \
 862     return (&g->r.r);                                                   \
 863   }
 864 SALSA20_VARS(DEFXGRAND)
 865
 866 /*----- Test rig ----------------------------------------------------------*/
 867
 868 #ifdef TEST_RIG
 869
 870 #include <stdio.h>
 871 #include <string.h>
 872
 873 #include <mLib/quis.h>
 874 #include <mLib/testrig.h>
 875
 876 static const int perm[] = {
 877    0, 13, 10,  7,
 878    4,  1, 14, 11,
 879    8,  5,  2, 15,
 880   12,  9,  6,  3
 881 };
 882
 883 #define DEFVCORE(r)                                                     \
 884   static int v_core_##r(dstr *v)                                        \
 885   {                                                                     \
 886     salsa20_matrix a, b;                                                \
 887     dstr d = DSTR_INIT;                                                 \
 888     int i, j, n;                                                        \
 889     int ok = 1;                                                         \
 890                                                                         \
 891     DENSURE(&d, SALSA20_OUTSZ); d.len = SALSA20_OUTSZ;                  \
 892     n = *(int *)v[0].buf;                                               \
 893     for (i = 0; i < SALSA20_OUTSZ/4; i++)                               \
 894       b[i] = LOAD32_L(v[1].buf + 4*i);                                  \
 895     for (i = 0; i < n; i++) {                                           \
 896       for (j = 0; j < 16; j++) a[perm[j]] = b[j];                       \
 897       core(r, a, b);                                                    \
 898       memcpy(a, b, sizeof(a));                                          \
 899     }                                                                   \
 900     for (i = 0; i < SALSA20_OUTSZ/4; i++) STORE32_L(d.buf + 4*i, b[i]); \
 901                                                                         \
 902     if (d.len != v[2].len || memcmp(d.buf, v[2].buf, v[2].len) != 0) {  \
 903       ok = 0;                                                           \
 904       printf("\nfail core:"                                             \
 905              "\n\titerations = %d"                                      \
 906              "\n\tin       = ", n);                                     \
 907       type_hex.dump(&v[1], stdout);                                     \
 908       printf("\n\texpected   = ");                                      \
 909       type_hex.dump(&v[2], stdout);                                     \
 910       printf("\n\tcalculated = ");                                      \
 911       type_hex.dump(&d, stdout);                                        \
 912       putchar('\n');                                                    \
 913     }                                                                   \
 914                                                                         \
 915     dstr_destroy(&d);                                                   \
 916     return (ok);                                                        \
 917   }
 918 SALSA20_VARS(DEFVCORE)
 919
 920 #define SALSA20_CTX(r) salsa20_ctx
 921
 922 #define SALSA20_TESTSETUP(r, ctx, k, ksz, n, nsz, p, psz) do {          \
 923   kludge64 pos64;                                                       \
 924   salsa20_init(ctx, k, ksz, 0);                                         \
 925   if (nsz == 8) salsa20_setnonce(ctx, n);                               \
 926   else if (nsz == 12) salsa20_setnonce_ietf(ctx, n);                    \
 927   if (psz == 8) { LOAD64_(pos64, p); salsa20_seeku64(ctx, pos64); }     \
 928   else if (psz == 4) salsa20_seek_ietf(ctx, LOAD32(p));                 \
 929 } while (0)
 930
 931 #define XSALSA20_TESTSETUP(r, ctx, k, ksz, n, nsz, p, psz) do {         \
 932   kludge64 pos64;                                                       \
 933   XSALSA20_INIT(r, ctx, k, ksz, 0);                                     \
 934   if (nsz == 24) XSALSA20_SETNONCE(r, ctx, n);                          \
 935   if (psz == 8) { LOAD64_(pos64, p); XSALSA20_SEEKU64(r, ctx, pos64); } \
 936 } while (0)
 937
 938 #define DEFxVENC(base, BASE, r)                                         \
 939   static int v_encrypt_##base##_##r(dstr *v)                            \
 940   {                                                                     \
 941     BASE##_CTX(r) ctx;                                                  \
 942     dstr d = DSTR_INIT;                                                 \
 943     const octet *p, *p0;                                                \
 944     octet *q;                                                           \
 945     size_t sz, sz0, step;                                               \
 946     unsigned long skip;                                                 \
 947     int ok = 1;                                                         \
 948                                                                         \
 949     if (v[4].len) { p0 = (const octet *)v[4].buf; sz0 = v[4].len; }     \
 950     else { p0 = 0; sz0 = v[5].len; }                                    \
 951     DENSURE(&d, sz0); d.len = sz0;                                      \
 952     skip = *(unsigned long *)v[3].buf;                                  \
 953                                                                         \
 954     step = 0;                                                           \
 955     while (step < sz0 + skip) {                                         \
 956       step = step ? 3*step + 4 : 1;                                     \
 957       if (step > sz0 + skip) step = sz0 + skip;                         \
 958       BASE##_TESTSETUP(r, &ctx, v[0].buf, v[0].len,                     \
 959                        v[1].buf, v[1].len, v[2].buf, v[2].len);         \
 960                                                                         \
 961       for (sz = skip; sz >= step; sz -= step)                           \
 962         BASE##_ENCRYPT(r, &ctx, 0, 0, step);                            \
 963       if (sz) BASE##_ENCRYPT(r, &ctx, 0, 0, sz);                        \
 964       for (p = p0, q = (octet *)d.buf, sz = sz0;                        \
 965            sz >= step;                                                  \
 966            sz -= step, q += step) {                                     \
 967         BASE##_ENCRYPT(r, &ctx, p, q, step);                            \
 968         if (p) p += step;                                               \
 969       }                                                                 \
 970       if (sz) BASE##_ENCRYPT(r, &ctx, p, q, sz);                        \
 971                                                                         \
 972       if (d.len != v[5].len || memcmp(d.buf, v[5].buf, v[5].len) != 0) { \
 973         ok = 0;                                                         \
 974         printf("\nfail encrypt:"                                        \
 975                "\n\tstep           = %lu"                               \
 976                "\n\tkey    = ", (unsigned long)step);                   \
 977         type_hex.dump(&v[0], stdout);                                   \
 978         printf("\n\tnonce          = ");                                \
 979         type_hex.dump(&v[1], stdout);                                   \
 980         printf("\n\tposition   = ");                                    \
 981         type_hex.dump(&v[2], stdout);                                   \
 982         printf("\n\tskip           = %lu", skip);                       \
 983         printf("\n\tmessage    = ");                                    \
 984         type_hex.dump(&v[4], stdout);                                   \
 985         printf("\n\texpected   = ");                                    \
 986         type_hex.dump(&v[5], stdout);                                   \
 987         printf("\n\tcalculated = ");                                    \
 988         type_hex.dump(&d, stdout);                                      \
 989         putchar('\n');                                                  \
 990       }                                                                 \
 991     }                                                                   \
 992                                                                         \
 993     dstr_destroy(&d);                                                   \
 994     return (ok);                                                        \
 995   }
 996 #define DEFVENC(r) DEFxVENC(salsa20, SALSA20, r)
 997 #define DEFXVENC(r) DEFxVENC(xsalsa20, XSALSA20, r)
 998 SALSA20_VARS(DEFVENC)
 999 SALSA20_VARS(DEFXVENC)
1000
1001 static test_chunk defs[] = {
1002 #define DEFxTAB(pre, base, r)                                           \
1003   { pre SALSA20_NAME_##r, v_encrypt_##base##_##r,                       \
1004     { &type_hex, &type_hex, &type_hex, &type_ulong,                     \
1005       &type_hex, &type_hex, 0 } },
1006 #define DEFTAB(r)                                                       \
1007   { SALSA20_NAME_##r "-core", v_core_##r,                               \
1008     { &type_int, &type_hex, &type_hex, 0 } },                           \
1009   DEFxTAB("", salsa20, r)
1010 #define DEFXTAB(r) DEFxTAB("x", xsalsa20, r)
1011 SALSA20_VARS(DEFTAB)
1012 SALSA20_VARS(DEFXTAB)
1013   { 0, 0, { 0 } }
1014 };
1015
1016 int main(int argc, char *argv[])
1017 {
1018   test_run(argc, argv, defs, SRCDIR"/t/salsa20");
1019   return (0);
1020 }
1021
1022 #endif
1023
1024 /*----- That's all, folks -------------------------------------------------*/