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