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

   1 /* -*-c-*-
   2  *
   3  * Salsa20 stream cipher
   4  *
   5  * (c) 2015 Straylight/Edgeware
   6  */
   7
   8 /*----- Header files ------------------------------------------------------*/
   9
  10 #include "config.h"
  11
  12 #include <stdarg.h>
  13
  14 #include <mLib/bits.h>
  15
  16 #include "arena.h"
  17 #include "dispatch.h"
  18 #include "gcipher.h"
  19 #include "grand.h"
  20 #include "keysz.h"
  21 #include "paranoia.h"
  22 #include "salsa20.h"
  23 #include "salsa20-core.h"
  24
  25 /*----- Global variables --------------------------------------------------*/
  26
  27 const octet salsa20_keysz[] = { KSZ_SET, 32, 16, 10, 0 };
  28
  29 /*----- The Salsa20 core function and utilities ---------------------------*/
  30
  31 /* --- @core@ --- *
  32  *
  33  * Arguments:   @unsigned r@ = number of rounds
  34  *              @const salsa20_matrix src@ = input matrix
  35  *              @salsa20_matrix dest@ = where to put the output
  36  *
  37  * Returns:     ---
  38  *
  39  *
  40  * Use:         Apply the Salsa20/r core function to @src@, writing the
  41  *              result to @dest@.  This consists of @r@ rounds followed by
  42  *              the feedforward step.
  43  */
  44
  45 CPU_DISPATCH(static, (void), void, core,
  46              (unsigned r, const salsa20_matrix src, salsa20_matrix dest),
  47              (r, src, dest), pick_core, simple_core);
  48
  49 static void simple_core(unsigned r, const salsa20_matrix src,
  50                         salsa20_matrix dest)
  51   { SALSA20_nR(dest, src, r); SALSA20_FFWD(dest, src); }
  52
  53 #if CPUFAM_X86 || CPUFAM_AMD64
  54 extern core__functype salsa20_core_x86ish_sse2;
  55 #endif
  56
  57 #if CPUFAM_ARMEL
  58 extern core__functype salsa20_core_arm_neon;
  59 #endif
  60
  61 static core__functype *pick_core(void)
  62 {
  63 #if CPUFAM_X86 || CPUFAM_AMD64
  64   DISPATCH_PICK_COND(salsa20_core, salsa20_core_x86ish_sse2,
  65                      cpu_feature_p(CPUFEAT_X86_SSE2));
  66 #endif
  67 #if CPUFAM_ARMEL
  68   DISPATCH_PICK_COND(salsa20_core, salsa20_core_arm_neon,
  69                      cpu_feature_p(CPUFEAT_ARM_NEON));
  70 #endif
  71   DISPATCH_PICK_FALLBACK(salsa20_core, simple_core);
  72 }
  73
  74 /* --- @populate@ --- *
  75  *
  76  * Arguments:   @salsa20_matrix a@ = a matrix to fill in
  77  *              @const void *key@ = pointer to key material
  78  *              @size_t ksz@ = size of key
  79  *
  80  * Returns:     ---
  81  *
  82  * Use:         Fills in a Salsa20 matrix from the key, setting the
  83  *              appropriate constants according to the key length.  The nonce
  84  *              and position words are left uninitialized.
  85  */
  86
  87 static void populate(salsa20_matrix a, const void *key, size_t ksz)
  88 {
  89   const octet *k = key;
  90
  91   KSZ_ASSERT(salsa20, ksz);
  92
  93   /* Here's the pattern of key, constant, nonce, and counter pieces in the
  94    * matrix, before and after our permutation.
  95    *
  96    * [ C0 K0 K1 K2 ]       [ C0 C1 C2 C3 ]
  97    * [ K3 C1 N0 N1 ]  -->  [ K3 T1 K7 K2 ]
  98    * [ T0 T1 C2 K4 ]       [ T0 K6 K1 N1 ]
  99    * [ K5 K6 K7 C3 ]       [ K5 K0 N0 K4 ]
 100    */
 101
 102   a[13] = LOAD32_L(k +  0);
 103   a[10] = LOAD32_L(k +  4);
 104   if (ksz == 10) {
 105     a[ 7] = LOAD16_L(k +  8);
 106     a[ 4] = 0;
 107   } else {
 108     a[ 7] = LOAD32_L(k +  8);
 109     a[ 4] = LOAD32_L(k + 12);
 110   }
 111   if (ksz <= 16) {
 112     a[15] = a[13];
 113     a[12] = a[10];
 114     a[ 9] = a[ 7];
 115     a[ 6] = a[ 4];
 116     a[ 0] = SALSA20_A128;
 117     a[ 1] = SALSA20_B128;
 118     a[ 2] = ksz == 10 ? SALSA20_C80 : SALSA20_C128;
 119     a[ 3] = SALSA20_D128;
 120   } else {
 121     a[15] = LOAD32_L(k + 16);
 122     a[12] = LOAD32_L(k + 20);
 123     a[ 9] = LOAD32_L(k + 24);
 124     a[ 6] = LOAD32_L(k + 28);
 125     a[ 0] = SALSA20_A256;
 126     a[ 1] = SALSA20_B256;
 127     a[ 2] = SALSA20_C256;
 128     a[ 3] = SALSA20_D256;
 129   }
 130 }
 131
 132 /*----- Salsa20 implementation --------------------------------------------*/
 133
 134 /* --- @salsa20_init@ --- *
 135  *
 136  * Arguments:   @salsa20_ctx *ctx@ = context to fill in
 137  *              @const void *key@ = pointer to key material
 138  *              @size_t ksz@ = size of key (either 32 or 16)
 139  *              @const void *nonce@ = initial nonce, or null
 140  *
 141  * Returns:     ---
 142  *
 143  * Use:         Initializes a Salsa20 context ready for use.
 144  */
 145
 146 void salsa20_init(salsa20_ctx *ctx, const void *key, size_t ksz,
 147                   const void *nonce)
 148 {
 149   static const octet zerononce[SALSA20_NONCESZ];
 150
 151   populate(ctx->a, key, ksz);
 152   salsa20_setnonce(ctx, nonce ? nonce : zerononce);
 153 }
 154
 155 /* --- @salsa20_setnonce@ --- *
 156  *
 157  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 158  *              @const void *nonce@ = the nonce (@SALSA20_NONCESZ@ bytes)
 159  *
 160  * Returns:     ---
 161  *
 162  * Use:         Set a new nonce in the context @ctx@, e.g., for processing a
 163  *              different message.  The stream position is reset to zero (see
 164  *              @salsa20_seek@ etc.).
 165  */
 166
 167 void salsa20_setnonce(salsa20_ctx *ctx, const void *nonce)
 168 {
 169   const octet *n = nonce;
 170
 171   ctx->a[14] = LOAD32_L(n + 0);
 172   ctx->a[11] = LOAD32_L(n + 4);
 173   salsa20_seek(ctx, 0);
 174 }
 175
 176 /* --- @salsa20_seek@, @salsa20_seeku64@ --- *
 177  *
 178  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 179  *              @unsigned long i@, @kludge64 i@ = new position to set
 180  *
 181  * Returns:     ---
 182  *
 183  * Use:         Sets a new stream position, in units of Salsa20 output
 184  *              blocks, which are @SALSA20_OUTSZ@ bytes each.  Byte
 185  *              granularity can be achieved by calling @salsa20R_encrypt@
 186  *              appropriately.
 187  */
 188
 189 void salsa20_seek(salsa20_ctx *ctx, unsigned long i)
 190   { kludge64 ii; ASSIGN64(ii, i); salsa20_seeku64(ctx, ii); }
 191
 192 void salsa20_seeku64(salsa20_ctx *ctx, kludge64 i)
 193 {
 194   ctx->a[8] = LO64(i); ctx->a[5] = HI64(i);
 195   ctx->bufi = SALSA20_OUTSZ;
 196 }
 197
 198 /* --- @salsa20_tell@, @salsa20_tellu64@ --- *
 199  *
 200  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 201  *
 202  * Returns:     The current position in the output stream, in blocks,
 203  *              rounding upwards.
 204  */
 205
 206 unsigned long salsa20_tell(salsa20_ctx *ctx)
 207   { kludge64 i = salsa20_tellu64(ctx); return (GET64(unsigned long, i)); }
 208
 209 kludge64 salsa20_tellu64(salsa20_ctx *ctx)
 210   { kludge64 i; SET64(i, ctx->a[5], ctx->a[8]); return (i); }
 211
 212 /* --- @salsa20{,12,8}_encrypt@ --- *
 213  *
 214  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 215  *              @const void *src@ = source buffer (or null)
 216  *              @void *dest@ = destination buffer (or null)
 217  *              @size_t sz@ = size of the buffers
 218  *
 219  * Returns:     ---
 220  *
 221  * Use:         Encrypts or decrypts @sz@ bytes of data from @src@ to @dest@.
 222  *              Salsa20 works by XORing plaintext with a keystream, so
 223  *              encryption and decryption are the same operation.  If @dest@
 224  *              is null then ignore @src@ and skip @sz@ bytes of the
 225  *              keystream.  If @src@ is null, then just write the keystream
 226  *              to @dest@.
 227  */
 228
 229 #define SALSA20_ENCRYPT(r, ctx, src, dest, sz)                          \
 230   SALSA20_DECOR(salsa20, r, _encrypt)(ctx, src, dest, sz)
 231 #define DEFENCRYPT(r)                                                   \
 232   void SALSA20_ENCRYPT(r, salsa20_ctx *ctx, const void *src,            \
 233                        void *dest, size_t sz)                           \
 234   {                                                                     \
 235     salsa20_matrix b;                                                   \
 236     const octet *s = src;                                               \
 237     octet *d = dest;                                                    \
 238     size_t n;                                                           \
 239     kludge64 pos, delta;                                                \
 240                                                                         \
 241     SALSA20_OUTBUF(ctx, d, s, sz);                                      \
 242     if (!sz) return;                                                    \
 243                                                                         \
 244     if (!dest) {                                                        \
 245       n = sz/SALSA20_OUTSZ;                                             \
 246       pos = salsa20_tellu64(ctx);                                       \
 247       ASSIGN64(delta, n);                                               \
 248       ADD64(pos, pos, delta);                                           \
 249       salsa20_seeku64(ctx, pos);                                        \
 250       sz = sz%SALSA20_OUTSZ;                                            \
 251     } else if (!src) {                                                  \
 252       while (sz >= SALSA20_OUTSZ) {                                     \
 253         core(r, ctx->a, b);                                             \
 254         SALSA20_STEP(ctx->a);                                           \
 255         SALSA20_GENFULL(b, d);                                          \
 256         sz -= SALSA20_OUTSZ;                                            \
 257       }                                                                 \
 258     } else {                                                            \
 259       while (sz >= SALSA20_OUTSZ) {                                     \
 260         core(r, ctx->a, b);                                             \
 261         SALSA20_STEP(ctx->a);                                           \
 262         SALSA20_MIXFULL(b, d, s);                                       \
 263         sz -= SALSA20_OUTSZ;                                            \
 264       }                                                                 \
 265     }                                                                   \
 266                                                                         \
 267     if (sz) {                                                           \
 268       core(r, ctx->a, b);                                               \
 269       SALSA20_STEP(ctx->a);                                             \
 270       SALSA20_PREPBUF(ctx, b);                                          \
 271       SALSA20_OUTBUF(ctx, d, s, sz);                                    \
 272       assert(!sz);                                                      \
 273     }                                                                   \
 274   }
 275 SALSA20_VARS(DEFENCRYPT)
 276
 277 /*----- HSalsa20 implementation -------------------------------------------*/
 278
 279 #define HSALSA20_RAW(r, ctx, src, dest)                                 \
 280   SALSA20_DECOR(hsalsa20, r, _raw)(ctx, src, dest)
 281 #define HSALSA20_PRF(r, ctx, src, dest)                                 \
 282   SALSA20_DECOR(hsalsa20, r, _prf)(ctx, src, dest)
 283
 284 /* --- @hsalsa20{,12,8}_prf@ --- *
 285  *
 286  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 287  *              @const void *src@ = the input (@HSALSA20_INSZ@ bytes)
 288  *              @void *dest@ = the output (@HSALSA20_OUTSZ@ bytes)
 289  *
 290  * Returns:     ---
 291  *
 292  * Use:         Apply the HSalsa20/r pseudorandom function to @src@, writing
 293  *              the result to @out@.
 294  */
 295
 296 #define DEFHSALSA20(r)                                                  \
 297   static void HSALSA20_RAW(r, salsa20_matrix k,                         \
 298                            const uint32 *src, uint32 *dest)             \
 299   {                                                                     \
 300     salsa20_matrix a;                                                   \
 301     int i;                                                              \
 302                                                                         \
 303     /* --- HSalsa20, computed from full Salsa20 --- *                   \
 304      *                                                                  \
 305      * The security proof makes use of the fact that HSalsa20 (i.e.,    \
 306      * without the final feedforward step) can be computed from full    \
 307      * Salsa20 using only knowledge of the non-secret input.  I don't   \
 308      * want to compromise the performance of the main function by       \
 309      * making the feedforward step separate, but this operation is less \
 310      * speed critical, so we do it the harder way.                      \
 311      */                                                                 \
 312                                                                         \
 313     for (i = 0; i < 4; i++) k[14 - 3*i] = src[i];                       \
 314     core(r, k, a);                                                      \
 315     for (i = 0; i < 4; i++) dest[i] = a[5*i] - k[i];                    \
 316     for (i = 4; i < 8; i++) dest[i] = a[i + 2] - k[26 - 3*i];           \
 317   }                                                                     \
 318                                                                         \
 319   void HSALSA20_PRF(r, salsa20_ctx *ctx, const void *src, void *dest)   \
 320   {                                                                     \
 321     const octet *s = src;                                               \
 322     octet *d = dest;                                                    \
 323     uint32 in[4], out[8];                                               \
 324     int i;                                                              \
 325                                                                         \
 326     for (i = 0; i < 4; i++) in[i] = LOAD32_L(s + 4*i);                  \
 327     HSALSA20_RAW(r, ctx->a, in, out);                                   \
 328     for (i = 0; i < 8; i++) STORE32_L(d + 4*i, out[i]);                 \
 329   }
 330 SALSA20_VARS(DEFHSALSA20)
 331
 332 /*----- XSalsa20 implementation -------------------------------------------*/
 333
 334 /* --- Some convenient macros for naming functions --- *
 335  *
 336  * Because the crypto core is involved in XSalsa20/r's per-nonce setup, we
 337  * need to take an interest in the number of rounds in most of the various
 338  * functions, and it will probably help if we distinguish the context
 339  * structures for the various versions.
 340  */
 341
 342 #define XSALSA20_CTX(r) SALSA20_DECOR(xsalsa20, r, _ctx)
 343 #define XSALSA20_INIT(r, ctx, k, ksz, n)                                \
 344   SALSA20_DECOR(xsalsa20, r, _init)(ctx, k, ksz, n)
 345 #define XSALSA20_SETNONCE(r, ctx, n)                                    \
 346   SALSA20_DECOR(xsalsa20, r, _setnonce)(ctx, n)
 347 #define XSALSA20_SEEK(r, ctx, i)                                        \
 348   SALSA20_DECOR(xsalsa20, r, _seek)(ctx, i)
 349 #define XSALSA20_SEEKU64(r, ctx, i)                                     \
 350   SALSA20_DECOR(xsalsa20, r, _seeku64)(ctx, i)
 351 #define XSALSA20_TELL(r, ctx)                                           \
 352   SALSA20_DECOR(xsalsa20, r, _tell)(ctx)
 353 #define XSALSA20_TELLU64(r, ctx)                                        \
 354   SALSA20_DECOR(xsalsa20, r, _tellu64)(ctx)
 355 #define XSALSA20_ENCRYPT(r, ctx, src, dest, sz)                         \
 356   SALSA20_DECOR(xsalsa20, r, _encrypt)(ctx, src, dest, sz)
 357
 358 /* --- @xsalsa20{,12,8}_init@ --- *
 359  *
 360  * Arguments:   @xsalsa20R_ctx *ctx@ = the context to fill in
 361  *              @const void *key@ = pointer to key material
 362  *              @size_t ksz@ = size of key (either 32 or 16)
 363  *              @const void *nonce@ = initial nonce, or null
 364  *
 365  * Returns:     ---
 366  *
 367  * Use:         Initializes an XSalsa20/r context ready for use.
 368  *
 369  *              There is a different function for each number of rounds,
 370  *              unlike for plain Salsa20.
 371  */
 372
 373 #define DEFXINIT(r)                                                     \
 374   void XSALSA20_INIT(r, XSALSA20_CTX(r) *ctx,                           \
 375                         const void *key, size_t ksz, const void *nonce) \
 376   {                                                                     \
 377     static const octet zerononce[XSALSA20_NONCESZ];                     \
 378                                                                         \
 379     populate(ctx->k, key, ksz);                                         \
 380     ctx->s.a[ 0] = SALSA20_A256;                                        \
 381     ctx->s.a[ 1] = SALSA20_B256;                                        \
 382     ctx->s.a[ 2] = SALSA20_C256;                                        \
 383     ctx->s.a[ 3] = SALSA20_D256;                                        \
 384     XSALSA20_SETNONCE(r, ctx, nonce ? nonce : zerononce);               \
 385   }
 386 SALSA20_VARS(DEFXINIT)
 387
 388 /* --- @xsalsa20{,12,8}_setnonce@ --- *
 389  *
 390  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 391  *              @const void *nonce@ = the nonce (@XSALSA20_NONCESZ@ bytes)
 392  *
 393  * Returns:     ---
 394  *
 395  * Use:         Set a new nonce in the context @ctx@, e.g., for processing a
 396  *              different message.  The stream position is reset to zero (see
 397  *              @salsa20_seek@ etc.).
 398  *
 399  *              There is a different function for each number of rounds,
 400  *              unlike for plain Salsa20.
 401  */
 402
 403 #define DEFXNONCE(r)                                                    \
 404   void XSALSA20_SETNONCE(r, XSALSA20_CTX(r) *ctx, const void *nonce)    \
 405   {                                                                     \
 406     const octet *n = nonce;                                             \
 407     uint32 in[4], out[8];                                               \
 408     int i;                                                              \
 409                                                                         \
 410     for (i = 0; i < 4; i++) in[i] = LOAD32_L(n + 4*i);                  \
 411     HSALSA20_RAW(r, ctx->k, in, out);                                   \
 412     for (i = 0; i < 4; i++) ctx->s.a[13 - 3*i] = out[i];                \
 413     for (i = 4; i < 8; i++) ctx->s.a[27 - 3*i] = out[i];                \
 414     salsa20_setnonce(&ctx->s, n + 16);                                  \
 415   }
 416 SALSA20_VARS(DEFXNONCE)
 417
 418 /* --- @xsalsa20{,12,8}_seek@, @xsalsa20{,12,8}_seeku64@ --- *
 419  *
 420  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 421  *              @unsigned long i@, @kludge64 i@ = new position to set
 422  *
 423  * Returns:     ---
 424  *
 425  * Use:         Sets a new stream position, in units of Salsa20 output
 426  *              blocks, which are @XSALSA20_OUTSZ@ bytes each.  Byte
 427  *              granularity can be achieved by calling @xsalsa20R_encrypt@
 428  *              appropriately.
 429  *
 430  *              There is a different function for each number of rounds,
 431  *              unlike for plain Salsa20, because the context structures are
 432  *              different.
 433  */
 434
 435 /* --- @xsalsa20{,12,8}_tell@, @xsalsa20{,12,8}_tellu64@ --- *
 436  *
 437  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 438  *
 439  * Returns:     The current position in the output stream, in blocks,
 440  *              rounding upwards.
 441  *
 442  *              There is a different function for each number of rounds,
 443  *              unlike for plain Salsa20, because the context structures are
 444  *              different.
 445  */
 446
 447 /* --- @xsalsa20{,12,8}_encrypt@ --- *
 448  *
 449  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 450  *              @const void *src@ = source buffer (or null)
 451  *              @void *dest@ = destination buffer (or null)
 452  *              @size_t sz@ = size of the buffers
 453  *
 454  * Returns:     ---
 455  *
 456  * Use:         Encrypts or decrypts @sz@ bytes of data from @src@ to @dest@.
 457  *              XSalsa20 works by XORing plaintext with a keystream, so
 458  *              encryption and decryption are the same operation.  If @dest@
 459  *              is null then ignore @src@ and skip @sz@ bytes of the
 460  *              keystream.  If @src@ is null, then just write the keystream
 461  *              to @dest@.
 462  */
 463
 464 #define DEFXPASSTHRU(r)                                                 \
 465   void XSALSA20_SEEK(r, XSALSA20_CTX(r) *ctx, unsigned long i)          \
 466     { salsa20_seek(&ctx->s, i); }                                       \
 467   void XSALSA20_SEEKU64(r, XSALSA20_CTX(r) *ctx, kludge64 i)            \
 468     { salsa20_seeku64(&ctx->s, i); }                                    \
 469   unsigned long XSALSA20_TELL(r, XSALSA20_CTX(r) *ctx)                  \
 470     { return salsa20_tell(&ctx->s); }                                   \
 471   kludge64 XSALSA20_TELLU64(r, XSALSA20_CTX(r) *ctx)                    \
 472     { return salsa20_tellu64(&ctx->s); }                                \
 473   void XSALSA20_ENCRYPT(r, XSALSA20_CTX(r) *ctx,                        \
 474                         const void *src, void *dest, size_t sz)         \
 475     { SALSA20_ENCRYPT(r, &ctx->s, src, dest, sz); }
 476 SALSA20_VARS(DEFXPASSTHRU)
 477
 478 /*----- Generic cipher interface ------------------------------------------*/
 479
 480 typedef struct gctx { gcipher c; salsa20_ctx ctx; } gctx;
 481
 482 static void gsetiv(gcipher *c, const void *iv)
 483   { gctx *g = (gctx *)c; salsa20_setnonce(&g->ctx, iv); }
 484
 485 static void gdestroy(gcipher *c)
 486   { gctx *g = (gctx *)c; BURN(*g); S_DESTROY(g); }
 487
 488 #define DEFGCIPHER(r)                                                   \
 489                                                                         \
 490   static const gcipher_ops gops_##r;                                    \
 491                                                                         \
 492   static gcipher *ginit_##r(const void *k, size_t sz)                   \
 493   {                                                                     \
 494     gctx *g = S_CREATE(gctx);                                           \
 495     g->c.ops = &gops_##r;                                               \
 496     salsa20_init(&g->ctx, k, sz, 0);                                    \
 497     return (&g->c);                                                     \
 498   }                                                                     \
 499                                                                         \
 500   static void gencrypt_##r(gcipher *c, const void *s,                   \
 501                            void *t, size_t sz)                          \
 502     { gctx *g = (gctx *)c; SALSA20_ENCRYPT(r, &g->ctx, s, t, sz); }     \
 503                                                                         \
 504   static const gcipher_ops gops_##r = {                                 \
 505     &SALSA20_DECOR(salsa20, r, ),                                       \
 506     gencrypt_##r, gencrypt_##r, gdestroy, gsetiv, 0                     \
 507   };                                                                    \
 508                                                                         \
 509   const gccipher SALSA20_DECOR(salsa20, r, ) = {                        \
 510     SALSA20_NAME_##r, salsa20_keysz,                                    \
 511     SALSA20_NONCESZ, ginit_##r                                          \
 512   };
 513
 514 SALSA20_VARS(DEFGCIPHER)
 515
 516 #define DEFGXCIPHER(r)                                                  \
 517                                                                         \
 518   typedef struct { gcipher c; XSALSA20_CTX(r) ctx; } gxctx_##r;         \
 519                                                                         \
 520   static void gxsetiv_##r(gcipher *c, const void *iv)                   \
 521     { gxctx_##r *g = (gxctx_##r *)c; XSALSA20_SETNONCE(r, &g->ctx, iv); } \
 522                                                                         \
 523   static void gxdestroy_##r(gcipher *c)                                 \
 524     { gxctx_##r *g = (gxctx_##r *)c; BURN(*g); S_DESTROY(g); }          \
 525                                                                         \
 526   static const gcipher_ops gxops_##r;                                   \
 527                                                                         \
 528   static gcipher *gxinit_##r(const void *k, size_t sz)                  \
 529   {                                                                     \
 530     gxctx_##r *g = S_CREATE(gxctx_##r);                                 \
 531     g->c.ops = &gxops_##r;                                              \
 532     XSALSA20_INIT(r, &g->ctx, k, sz, 0);                                \
 533     return (&g->c);                                                     \
 534   }                                                                     \
 535                                                                         \
 536   static void gxencrypt_##r(gcipher *c, const void *s,                  \
 537                             void *t, size_t sz)                         \
 538   {                                                                     \
 539     gxctx_##r *g = (gxctx_##r *)c;                                      \
 540     XSALSA20_ENCRYPT(r, &g->ctx, s, t, sz);                             \
 541   }                                                                     \
 542                                                                         \
 543   static const gcipher_ops gxops_##r = {                                \
 544     &SALSA20_DECOR(xsalsa20, r, ),                                      \
 545     gxencrypt_##r, gxencrypt_##r, gxdestroy_##r, gxsetiv_##r, 0         \
 546   };                                                                    \
 547                                                                         \
 548   const gccipher SALSA20_DECOR(xsalsa20, r, ) = {                       \
 549     "x" SALSA20_NAME_##r, salsa20_keysz,                                \
 550     XSALSA20_NONCESZ, gxinit_##r                                        \
 551   };
 552
 553 SALSA20_VARS(DEFGXCIPHER)
 554
 555 /*----- Generic random number generator interface -------------------------*/
 556
 557 typedef struct grops {
 558   size_t noncesz;
 559   void (*seek)(void *, kludge64);
 560   kludge64 (*tell)(void *);
 561   void (*setnonce)(void *, const void *);
 562   void (*generate)(void *, void *, size_t);
 563 } grops;
 564
 565 typedef struct grbasectx {
 566   grand r;
 567   const grops *ops;
 568 } grbasectx;
 569
 570 static int grmisc(grand *r, unsigned op, ...)
 571 {
 572   octet buf[XSALSA20_NONCESZ];
 573   grbasectx *g = (grbasectx *)r;
 574   grand *rr;
 575   const octet *p;
 576   size_t sz;
 577   uint32 i;
 578   unsigned long ul;
 579   kludge64 pos;
 580   va_list ap;
 581   int rc = 0;
 582
 583   va_start(ap, op);
 584
 585   switch (op) {
 586     case GRAND_CHECK:
 587       switch (va_arg(ap, unsigned)) {
 588         case GRAND_CHECK:
 589         case GRAND_SEEDINT:
 590         case GRAND_SEEDUINT32:
 591         case GRAND_SEEDBLOCK:
 592         case GRAND_SEEDRAND:
 593         case SALSA20_SEEK:
 594         case SALSA20_SEEKU64:
 595         case SALSA20_TELL:
 596         case SALSA20_TELLU64:
 597           rc = 1;
 598           break;
 599         default:
 600           rc = 0;
 601           break;
 602       }
 603       break;
 604
 605     case GRAND_SEEDINT:
 606       i = va_arg(ap, unsigned); STORE32_L(buf, i);
 607       memset(buf + 4, 0, g->ops->noncesz - 4);
 608       g->ops->setnonce(g, buf);
 609       break;
 610     case GRAND_SEEDUINT32:
 611       i = va_arg(ap, uint32); STORE32_L(buf, i);
 612       memset(buf + 4, 0, g->ops->noncesz - 4);
 613       g->ops->setnonce(g, buf);
 614       break;
 615     case GRAND_SEEDBLOCK:
 616       p = va_arg(ap, const void *);
 617       sz = va_arg(ap, size_t);
 618       if (sz < g->ops->noncesz) {
 619         memcpy(buf, p, sz);
 620         memset(buf + sz, 0, g->ops->noncesz - sz);
 621         p = buf;
 622       }
 623       g->ops->setnonce(g, p);
 624       break;
 625     case GRAND_SEEDRAND:
 626       rr = va_arg(ap, grand *);
 627       rr->ops->fill(rr, buf, g->ops->noncesz);
 628       g->ops->setnonce(g, buf);
 629       break;
 630     case SALSA20_SEEK:
 631       ul = va_arg(ap, unsigned long); ASSIGN64(pos, ul);
 632       g->ops->seek(g, pos);
 633       break;
 634     case SALSA20_SEEKU64:
 635       pos = va_arg(ap, kludge64);
 636       g->ops->seek(g, pos);
 637       break;
 638     case SALSA20_TELL:
 639       pos = g->ops->tell(g);
 640       *va_arg(ap, unsigned long *) = GET64(unsigned long, pos);
 641       break;
 642     case SALSA20_TELLU64:
 643       *va_arg(ap, kludge64 *) = g->ops->tell(g);
 644       break;
 645     default:
 646       GRAND_BADOP;
 647       break;
 648   }
 649
 650   return (rc);
 651 }
 652
 653 static octet grbyte(grand *r)
 654 {
 655   grbasectx *g = (grbasectx *)r;
 656   octet o;
 657   g->ops->generate(g, &o, 1);
 658   return (o);
 659 }
 660
 661 static uint32 grword(grand *r)
 662 {
 663   grbasectx *g = (grbasectx *)r;
 664   octet b[4];
 665   g->ops->generate(g, b, sizeof(b));
 666   return (LOAD32_L(b));
 667 }
 668
 669 static void grfill(grand *r, void *p, size_t sz)
 670 {
 671   grbasectx *g = (grbasectx *)r;
 672   g->ops->generate(r, p, sz);
 673 }
 674
 675 typedef struct grctx {
 676   grbasectx r;
 677   salsa20_ctx ctx;
 678 } grctx;
 679
 680 static void gr_seek(void *r, kludge64 pos)
 681   { grctx *g = r; salsa20_seeku64(&g->ctx, pos); }
 682
 683 static kludge64 gr_tell(void *r)
 684   { grctx *g = r; return (salsa20_tellu64(&g->ctx)); }
 685
 686 static void gr_setnonce(void *r, const void *n)
 687   { grctx *g = r; salsa20_setnonce(&g->ctx, n); }
 688
 689 static void grdestroy(grand *r)
 690   { grctx *g = (grctx *)r; BURN(*g); S_DESTROY(g); }
 691
 692 #define DEFGRAND(rr)                                                    \
 693                                                                         \
 694   static void gr_generate_##rr(void *r, void *b, size_t sz)             \
 695     { grctx *g = r; SALSA20_ENCRYPT(rr, &g->ctx, 0, b, sz); }           \
 696                                                                         \
 697   static const grops grops_##rr =                                       \
 698     { SALSA20_NONCESZ, gr_seek, gr_tell,                                \
 699       gr_setnonce, gr_generate_##rr };                                  \
 700                                                                         \
 701   static const grand_ops grops_rand_##rr = {                            \
 702     SALSA20_NAME_##rr, GRAND_CRYPTO, 0,                                 \
 703     grmisc, grdestroy, grword,                                          \
 704     grbyte, grword, grand_defaultrange, grfill                          \
 705   };                                                                    \
 706                                                                         \
 707   grand *SALSA20_DECOR(salsa20, rr, _rand)                              \
 708     (const void *k, size_t ksz, const void *n)                          \
 709   {                                                                     \
 710     grctx *g = S_CREATE(g);                                             \
 711     g->r.r.ops = &grops_rand_##rr;                                      \
 712     g->r.ops = &grops_##rr;                                             \
 713     salsa20_init(&g->ctx, k, ksz, n);                                   \
 714     return (&g->r.r);                                                   \
 715   }
 716 SALSA20_VARS(DEFGRAND)
 717
 718 #define DEFXGRAND(rr)                                                   \
 719                                                                         \
 720   typedef struct grxctx_##rr {                                          \
 721     grbasectx r;                                                        \
 722     XSALSA20_CTX(rr) ctx;                                               \
 723   } grxctx_##rr;                                                        \
 724                                                                         \
 725   static void grx_seek_##rr(void *r, kludge64 pos)                      \
 726     { grxctx_##rr *g = r; XSALSA20_SEEKU64(rr, &g->ctx, pos); }         \
 727                                                                         \
 728   static kludge64 grx_tell_##rr(void *r)                                \
 729     { grxctx_##rr *g = r; return (XSALSA20_TELLU64(rr, &g->ctx)); }     \
 730                                                                         \
 731   static void grx_setnonce_##rr(void *r, const void *n)                 \
 732     { grxctx_##rr *g = r; XSALSA20_SETNONCE(rr, &g->ctx, n); }          \
 733                                                                         \
 734   static void grxdestroy_##rr(grand *r)                                 \
 735     { grxctx_##rr *g = (grxctx_##rr *)r; BURN(*g); S_DESTROY(g); }      \
 736                                                                         \
 737   static void grx_generate_##rr(void *r, void *b, size_t sz)            \
 738     { grxctx_##rr *g = r; XSALSA20_ENCRYPT(rr, &g->ctx, 0, b, sz); }    \
 739                                                                         \
 740   static const grops grxops_##rr =                                      \
 741   { XSALSA20_NONCESZ, grx_seek_##rr, grx_tell_##rr,                     \
 742       grx_setnonce_##rr, grx_generate_##rr };                           \
 743                                                                         \
 744   static const grand_ops grxops_rand_##rr = {                           \
 745     "x" SALSA20_NAME_##rr, GRAND_CRYPTO, 0,                             \
 746     grmisc, grxdestroy_##rr, grword,                                    \
 747     grbyte, grword, grand_defaultrange, grfill                          \
 748   };                                                                    \
 749                                                                         \
 750   grand *SALSA20_DECOR(xsalsa20, rr, _rand)                             \
 751     (const void *k, size_t ksz, const void *n)                          \
 752   {                                                                     \
 753     grxctx_##rr *g = S_CREATE(g);                                       \
 754     g->r.r.ops = &grxops_rand_##rr;                                     \
 755     g->r.ops = &grxops_##rr;                                            \
 756     XSALSA20_INIT(rr, &g->ctx, k, ksz, n);                              \
 757     return (&g->r.r);                                                   \
 758   }
 759 SALSA20_VARS(DEFXGRAND)
 760
 761 /*----- Test rig ----------------------------------------------------------*/
 762
 763 #ifdef TEST_RIG
 764
 765 #include <stdio.h>
 766 #include <string.h>
 767
 768 #include <mLib/quis.h>
 769 #include <mLib/testrig.h>
 770
 771 static const int perm[] = {
 772    0, 13, 10,  7,
 773    4,  1, 14, 11,
 774    8,  5,  2, 15,
 775   12,  9,  6,  3
 776 };
 777
 778 #define DEFVCORE(r)                                                     \
 779   static int v_core_##r(dstr *v)                                        \
 780   {                                                                     \
 781     salsa20_matrix a, b;                                                \
 782     dstr d = DSTR_INIT;                                                 \
 783     int i, j, n;                                                        \
 784     int ok = 1;                                                         \
 785                                                                         \
 786     DENSURE(&d, SALSA20_OUTSZ); d.len = SALSA20_OUTSZ;                  \
 787     n = *(int *)v[0].buf;                                               \
 788     for (i = 0; i < SALSA20_OUTSZ/4; i++)                               \
 789       b[i] = LOAD32_L(v[1].buf + 4*i);                                  \
 790     for (i = 0; i < n; i++) {                                           \
 791       for (j = 0; j < 16; j++) a[perm[j]] = b[j];                       \
 792       core(r, a, b);                                                    \
 793       memcpy(a, b, sizeof(a));                                          \
 794     }                                                                   \
 795     for (i = 0; i < SALSA20_OUTSZ/4; i++) STORE32_L(d.buf + 4*i, b[i]); \
 796                                                                         \
 797     if (d.len != v[2].len || memcmp(d.buf, v[2].buf, v[2].len) != 0) {  \
 798       ok = 0;                                                           \
 799       printf("\nfail core:"                                             \
 800              "\n\titerations = %d"                                      \
 801              "\n\tin       = ", n);                                     \
 802       type_hex.dump(&v[1], stdout);                                     \
 803       printf("\n\texpected   = ");                                      \
 804       type_hex.dump(&v[2], stdout);                                     \
 805       printf("\n\tcalculated = ");                                      \
 806       type_hex.dump(&d, stdout);                                        \
 807       putchar('\n');                                                    \
 808     }                                                                   \
 809                                                                         \
 810     dstr_destroy(&d);                                                   \
 811     return (ok);                                                        \
 812   }
 813 SALSA20_VARS(DEFVCORE)
 814
 815 #define SALSA20_CTX(r) salsa20_ctx
 816 #define SALSA20_INIT(r, ctx, k, ksz, n) salsa20_init(ctx, k, ksz, n)
 817 #define SALSA20_SEEKU64(r, ctx, i) salsa20_seeku64(ctx, i)
 818
 819 #define DEFxVENC(base, BASE, r)                                         \
 820   static int v_encrypt_##base##_##r(dstr *v)                            \
 821   {                                                                     \
 822     BASE##_CTX(r) ctx;                                                  \
 823     dstr d = DSTR_INIT;                                                 \
 824     kludge64 pos;                                                       \
 825     const octet *p, *p0;                                                \
 826     octet *q;                                                           \
 827     size_t sz, sz0, step;                                               \
 828     unsigned long skip;                                                 \
 829     int ok = 1;                                                         \
 830                                                                         \
 831     if (v[4].len) { p0 = (const octet *)v[4].buf; sz0 = v[4].len; }     \
 832     else { p0 = 0; sz0 = v[5].len; }                                    \
 833     DENSURE(&d, sz0); d.len = sz0;                                      \
 834     skip = *(unsigned long *)v[3].buf;                                  \
 835                                                                         \
 836     step = 0;                                                           \
 837     while (step < sz0 + skip) {                                         \
 838       step = step ? 3*step + 4 : 1;                                     \
 839       if (step > sz0 + skip) step = sz0 + skip;                         \
 840       BASE##_INIT(r, &ctx, v[0].buf, v[0].len, v[1].buf);               \
 841       if (v[2].len) {                                                   \
 842         LOAD64_(pos, v[2].buf);                                         \
 843         BASE##_SEEKU64(r, &ctx, pos);                                   \
 844       }                                                                 \
 845                                                                         \
 846       for (sz = skip; sz >= step; sz -= step)                           \
 847         BASE##_ENCRYPT(r, &ctx, 0, 0, step);                            \
 848       if (sz) BASE##_ENCRYPT(r, &ctx, 0, 0, sz);                        \
 849       for (p = p0, q = (octet *)d.buf, sz = sz0;                        \
 850            sz >= step;                                                  \
 851            sz -= step, q += step) {                                     \
 852         BASE##_ENCRYPT(r, &ctx, p, q, step);                            \
 853         if (p) p += step;                                               \
 854       }                                                                 \
 855       if (sz) BASE##_ENCRYPT(r, &ctx, p, q, sz);                        \
 856                                                                         \
 857       if (d.len != v[5].len || memcmp(d.buf, v[5].buf, v[5].len) != 0) { \
 858         ok = 0;                                                         \
 859         printf("\nfail encrypt:"                                        \
 860                "\n\tstep           = %lu"                               \
 861                "\n\tkey    = ", (unsigned long)step);                   \
 862         type_hex.dump(&v[0], stdout);                                   \
 863         printf("\n\tnonce          = ");                                \
 864         type_hex.dump(&v[1], stdout);                                   \
 865         printf("\n\tposition   = ");                                    \
 866         type_hex.dump(&v[2], stdout);                                   \
 867         printf("\n\tskip           = %lu", skip);                       \
 868         printf("\n\tmessage    = ");                                    \
 869         type_hex.dump(&v[4], stdout);                                   \
 870         printf("\n\texpected   = ");                                    \
 871         type_hex.dump(&v[5], stdout);                                   \
 872         printf("\n\tcalculated = ");                                    \
 873         type_hex.dump(&d, stdout);                                      \
 874         putchar('\n');                                                  \
 875       }                                                                 \
 876     }                                                                   \
 877                                                                         \
 878     dstr_destroy(&d);                                                   \
 879     return (ok);                                                        \
 880   }
 881 #define DEFVENC(r) DEFxVENC(salsa20, SALSA20, r)
 882 #define DEFXVENC(r) DEFxVENC(xsalsa20, XSALSA20, r)
 883 SALSA20_VARS(DEFVENC)
 884 SALSA20_VARS(DEFXVENC)
 885
 886 static test_chunk defs[] = {
 887 #define DEFxTAB(pre, base, r)                                           \
 888   { pre SALSA20_NAME_##r, v_encrypt_##base##_##r,                       \
 889     { &type_hex, &type_hex, &type_hex, &type_ulong,                     \
 890       &type_hex, &type_hex, 0 } },
 891 #define DEFTAB(r)                                                       \
 892   { SALSA20_NAME_##r "-core", v_core_##r,                               \
 893     { &type_int, &type_hex, &type_hex, 0 } },                           \
 894   DEFxTAB("", salsa20, r)
 895 #define DEFXTAB(r) DEFxTAB("x", xsalsa20, r)
 896 SALSA20_VARS(DEFTAB)
 897 SALSA20_VARS(DEFXTAB)
 898   { 0, 0, { 0 } }
 899 };
 900
 901 int main(int argc, char *argv[])
 902 {
 903   test_run(argc, argv, defs, SRCDIR"/t/salsa20");
 904   return (0);
 905 }
 906
 907 #endif
 908
 909 /*----- That's all, folks -------------------------------------------------*/