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