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