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),
  46              void, core, (unsigned r, const salsa20_matrix src,
  47                           salsa20_matrix dest),
  48              (r, src, dest),
  49              pick_core, simple_core);
  50
  51 static void simple_core(unsigned r, const salsa20_matrix src,
  52                         salsa20_matrix dest)
  53   { SALSA20_nR(dest, src, r); SALSA20_FFWD(dest, src); }
  54
  55 #ifdef CPUFAM_X86
  56 extern core__functype salsa20_core_x86_sse2;
  57 #endif
  58
  59 static core__functype *pick_core(void)
  60 {
  61 #ifdef CPUFAM_X86
  62   if (cpu_feature_p(CPUFEAT_X86_SSE2)) return salsa20_core_x86_sse2;
  63 #endif
  64   return simple_core;
  65 }
  66
  67 /* --- @populate@ --- *
  68  *
  69  * Arguments:   @salsa20_matrix a@ = a matrix to fill in
  70  *              @const void *key@ = pointer to key material
  71  *              @size_t ksz@ = size of key
  72  *
  73  * Returns:     ---
  74  *
  75  * Use:         Fills in a Salsa20 matrix from the key, setting the
  76  *              appropriate constants according to the key length.  The nonce
  77  *              and position words are left uninitialized.
  78  */
  79
  80 static void populate(salsa20_matrix a, const void *key, size_t ksz)
  81 {
  82   const octet *k = key;
  83
  84   KSZ_ASSERT(salsa20, ksz);
  85
  86   /* Here's the pattern of key, constant, nonce, and counter pieces in the
  87    * matrix, before and after our permutation.
  88    *
  89    * [ C0 K0 K1 K2 ]       [ C0 C1 C2 C3 ]
  90    * [ K3 C1 N0 N1 ]  -->  [ K3 T1 K7 K2 ]
  91    * [ T0 T1 C2 K4 ]       [ T0 K6 K1 N1 ]
  92    * [ K5 K6 K7 C3 ]       [ K5 K0 N0 K4 ]
  93    */
  94
  95   a[13] = LOAD32_L(k +  0);
  96   a[10] = LOAD32_L(k +  4);
  97   if (ksz == 10) {
  98     a[ 7] = LOAD16_L(k +  8);
  99     a[ 4] = 0;
 100   } else {
 101     a[ 7] = LOAD32_L(k +  8);
 102     a[ 4] = LOAD32_L(k + 12);
 103   }
 104   if (ksz <= 16) {
 105     a[15] = a[13];
 106     a[12] = a[10];
 107     a[ 9] = a[ 7];
 108     a[ 6] = a[ 4];
 109     a[ 0] = SALSA20_A128;
 110     a[ 1] = SALSA20_B128;
 111     a[ 2] = ksz == 10 ? SALSA20_C80 : SALSA20_C128;
 112     a[ 3] = SALSA20_D128;
 113   } else {
 114     a[15] = LOAD32_L(k + 16);
 115     a[12] = LOAD32_L(k + 20);
 116     a[ 9] = LOAD32_L(k + 24);
 117     a[ 6] = LOAD32_L(k + 28);
 118     a[ 0] = SALSA20_A256;
 119     a[ 1] = SALSA20_B256;
 120     a[ 2] = SALSA20_C256;
 121     a[ 3] = SALSA20_D256;
 122   }
 123 }
 124
 125 /*----- Salsa20 implementation --------------------------------------------*/
 126
 127 /* --- @salsa20_init@ --- *
 128  *
 129  * Arguments:   @salsa20_ctx *ctx@ = context to fill in
 130  *              @const void *key@ = pointer to key material
 131  *              @size_t ksz@ = size of key (either 32 or 16)
 132  *              @const void *nonce@ = initial nonce, or null
 133  *
 134  * Returns:     ---
 135  *
 136  * Use:         Initializes a Salsa20 context ready for use.
 137  */
 138
 139 void salsa20_init(salsa20_ctx *ctx, const void *key, size_t ksz,
 140                   const void *nonce)
 141 {
 142   static const octet zerononce[SALSA20_NONCESZ];
 143
 144   populate(ctx->a, key, ksz);
 145   salsa20_setnonce(ctx, nonce ? nonce : zerononce);
 146 }
 147
 148 /* --- @salsa20_setnonce@ --- *
 149  *
 150  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 151  *              @const void *nonce@ = the nonce (@SALSA20_NONCESZ@ bytes)
 152  *
 153  * Returns:     ---
 154  *
 155  * Use:         Set a new nonce in the context @ctx@, e.g., for processing a
 156  *              different message.  The stream position is reset to zero (see
 157  *              @salsa20_seek@ etc.).
 158  */
 159
 160 void salsa20_setnonce(salsa20_ctx *ctx, const void *nonce)
 161 {
 162   const octet *n = nonce;
 163
 164   ctx->a[14] = LOAD32_L(n + 0);
 165   ctx->a[11] = LOAD32_L(n + 4);
 166   salsa20_seek(ctx, 0);
 167 }
 168
 169 /* --- @salsa20_seek@, @salsa20_seeku64@ --- *
 170  *
 171  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 172  *              @unsigned long i@, @kludge64 i@ = new position to set
 173  *
 174  * Returns:     ---
 175  *
 176  * Use:         Sets a new stream position, in units of Salsa20 output
 177  *              blocks, which are @SALSA20_OUTSZ@ bytes each.  Byte
 178  *              granularity can be achieved by calling @salsa20R_encrypt@
 179  *              appropriately.
 180  */
 181
 182 void salsa20_seek(salsa20_ctx *ctx, unsigned long i)
 183   { kludge64 ii; ASSIGN64(ii, i); salsa20_seeku64(ctx, ii); }
 184
 185 void salsa20_seeku64(salsa20_ctx *ctx, kludge64 i)
 186 {
 187   ctx->a[8] = LO64(i); ctx->a[5] = HI64(i);
 188   ctx->bufi = SALSA20_OUTSZ;
 189 }
 190
 191 /* --- @salsa20_tell@, @salsa20_tellu64@ --- *
 192  *
 193  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 194  *
 195  * Returns:     The current position in the output stream, in blocks,
 196  *              rounding upwards.
 197  */
 198
 199 unsigned long salsa20_tell(salsa20_ctx *ctx)
 200   { kludge64 i = salsa20_tellu64(ctx); return (GET64(unsigned long, i)); }
 201
 202 kludge64 salsa20_tellu64(salsa20_ctx *ctx)
 203   { kludge64 i; SET64(i, ctx->a[5], ctx->a[8]); return (i); }
 204
 205 /* --- @salsa20{,12,8}_encrypt@ --- *
 206  *
 207  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 208  *              @const void *src@ = source buffer (or null)
 209  *              @void *dest@ = destination buffer (or null)
 210  *              @size_t sz@ = size of the buffers
 211  *
 212  * Returns:     ---
 213  *
 214  * Use:         Encrypts or decrypts @sz@ bytes of data from @src@ to @dest@.
 215  *              Salsa20 works by XORing plaintext with a keystream, so
 216  *              encryption and decryption are the same operation.  If @dest@
 217  *              is null then ignore @src@ and skip @sz@ bytes of the
 218  *              keystream.  If @src@ is null, then just write the keystream
 219  *              to @dest@.
 220  */
 221
 222 #define SALSA20_ENCRYPT(r, ctx, src, dest, sz)                          \
 223   SALSA20_DECOR(salsa20, r, _encrypt)(ctx, src, dest, sz)
 224 #define DEFENCRYPT(r)                                                   \
 225   void SALSA20_ENCRYPT(r, salsa20_ctx *ctx, const void *src,            \
 226                        void *dest, size_t sz)                           \
 227   {                                                                     \
 228     salsa20_matrix b;                                                   \
 229     const octet *s = src;                                               \
 230     octet *d = dest;                                                    \
 231     size_t n;                                                           \
 232     kludge64 pos, delta;                                                \
 233                                                                         \
 234     SALSA20_OUTBUF(ctx, d, s, sz);                                      \
 235     if (!sz) return;                                                    \
 236                                                                         \
 237     if (!dest) {                                                        \
 238       n = sz/SALSA20_OUTSZ;                                             \
 239       pos = salsa20_tellu64(ctx);                                       \
 240       ASSIGN64(delta, n);                                               \
 241       ADD64(pos, pos, delta);                                           \
 242       salsa20_seeku64(ctx, pos);                                        \
 243       sz = sz%SALSA20_OUTSZ;                                            \
 244     } else if (!src) {                                                  \
 245       while (sz >= SALSA20_OUTSZ) {                                     \
 246         core(r, ctx->a, b);                                             \
 247         SALSA20_STEP(ctx->a);                                           \
 248         SALSA20_GENFULL(b, d);                                          \
 249         sz -= SALSA20_OUTSZ;                                            \
 250       }                                                                 \
 251     } else {                                                            \
 252       while (sz >= SALSA20_OUTSZ) {                                     \
 253         core(r, ctx->a, b);                                             \
 254         SALSA20_STEP(ctx->a);                                           \
 255         SALSA20_MIXFULL(b, d, s);                                       \
 256         sz -= SALSA20_OUTSZ;                                            \
 257       }                                                                 \
 258     }                                                                   \
 259                                                                         \
 260     if (sz) {                                                           \
 261       core(r, ctx->a, b);                                               \
 262       SALSA20_STEP(ctx->a);                                             \
 263       SALSA20_PREPBUF(ctx, b);                                          \
 264       SALSA20_OUTBUF(ctx, d, s, sz);                                    \
 265       assert(!sz);                                                      \
 266     }                                                                   \
 267   }
 268 SALSA20_VARS(DEFENCRYPT)
 269
 270 /*----- HSalsa20 implementation -------------------------------------------*/
 271
 272 #define HSALSA20_RAW(r, ctx, src, dest)                                 \
 273   SALSA20_DECOR(hsalsa20, r, _raw)(ctx, src, dest)
 274 #define HSALSA20_PRF(r, ctx, src, dest)                                 \
 275   SALSA20_DECOR(hsalsa20, r, _prf)(ctx, src, dest)
 276
 277 /* --- @hsalsa20{,12,8}_prf@ --- *
 278  *
 279  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 280  *              @const void *src@ = the input (@HSALSA20_INSZ@ bytes)
 281  *              @void *dest@ = the output (@HSALSA20_OUTSZ@ bytes)
 282  *
 283  * Returns:     ---
 284  *
 285  * Use:         Apply the HSalsa20/r pseudorandom function to @src@, writing
 286  *              the result to @out@.
 287  */
 288
 289 #define DEFHSALSA20(r)                                                  \
 290   static void HSALSA20_RAW(r, salsa20_matrix k,                         \
 291                            const uint32 *src, uint32 *dest)             \
 292   {                                                                     \
 293     salsa20_matrix a;                                                   \
 294     int i;                                                              \
 295                                                                         \
 296     /* --- HSalsa20, computed from full Salsa20 --- *                   \
 297      *                                                                  \
 298      * The security proof makes use of the fact that HSalsa20 (i.e.,    \
 299      * without the final feedforward step) can be computed from full    \
 300      * Salsa20 using only knowledge of the non-secret input.  I don't   \
 301      * want to compromise the performance of the main function by       \
 302      * making the feedforward step separate, but this operation is less \
 303      * speed critical, so we do it the harder way.                      \
 304      */                                                                 \
 305                                                                         \
 306     for (i = 0; i < 4; i++) k[14 - 3*i] = src[i];                       \
 307     core(r, k, a);                                                      \
 308     for (i = 0; i < 4; i++) dest[i] = a[5*i] - k[i];                    \
 309     for (i = 4; i < 8; i++) dest[i] = a[i + 2] - k[26 - 3*i];           \
 310   }                                                                     \
 311                                                                         \
 312   void HSALSA20_PRF(r, salsa20_ctx *ctx, const void *src, void *dest)   \
 313   {                                                                     \
 314     const octet *s = src;                                               \
 315     octet *d = dest;                                                    \
 316     uint32 in[4], out[8];                                               \
 317     int i;                                                              \
 318                                                                         \
 319     for (i = 0; i < 4; i++) in[i] = LOAD32_L(s + 4*i);                  \
 320     HSALSA20_RAW(r, ctx->a, in, out);                                   \
 321     for (i = 0; i < 8; i++) STORE32_L(d + 4*i, out[i]);                 \
 322   }
 323 SALSA20_VARS(DEFHSALSA20)
 324
 325 /*----- XSalsa20 implementation -------------------------------------------*/
 326
 327 /* --- Some convenient macros for naming functions --- *
 328  *
 329  * Because the crypto core is involved in XSalsa20/r's per-nonce setup, we
 330  * need to take an interest in the number of rounds in most of the various
 331  * functions, and it will probably help if we distinguish the context
 332  * structures for the various versions.
 333  */
 334
 335 #define XSALSA20_CTX(r) SALSA20_DECOR(xsalsa20, r, _ctx)
 336 #define XSALSA20_INIT(r, ctx, k, ksz, n)                                \
 337   SALSA20_DECOR(xsalsa20, r, _init)(ctx, k, ksz, n)
 338 #define XSALSA20_SETNONCE(r, ctx, n)                                    \
 339   SALSA20_DECOR(xsalsa20, r, _setnonce)(ctx, n)
 340 #define XSALSA20_SEEK(r, ctx, i)                                        \
 341   SALSA20_DECOR(xsalsa20, r, _seek)(ctx, i)
 342 #define XSALSA20_SEEKU64(r, ctx, i)                                     \
 343   SALSA20_DECOR(xsalsa20, r, _seeku64)(ctx, i)
 344 #define XSALSA20_TELL(r, ctx)                                           \
 345   SALSA20_DECOR(xsalsa20, r, _tell)(ctx)
 346 #define XSALSA20_TELLU64(r, ctx)                                        \
 347   SALSA20_DECOR(xsalsa20, r, _tellu64)(ctx)
 348 #define XSALSA20_ENCRYPT(r, ctx, src, dest, sz)                         \
 349   SALSA20_DECOR(xsalsa20, r, _encrypt)(ctx, src, dest, sz)
 350
 351 /* --- @xsalsa20{,12,8}_init@ --- *
 352  *
 353  * Arguments:   @xsalsa20R_ctx *ctx@ = the context to fill in
 354  *              @const void *key@ = pointer to key material
 355  *              @size_t ksz@ = size of key (either 32 or 16)
 356  *              @const void *nonce@ = initial nonce, or null
 357  *
 358  * Returns:     ---
 359  *
 360  * Use:         Initializes an XSalsa20/r context ready for use.
 361  *
 362  *              There is a different function for each number of rounds,
 363  *              unlike for plain Salsa20.
 364  */
 365
 366 #define DEFXINIT(r)                                                     \
 367   void XSALSA20_INIT(r, XSALSA20_CTX(r) *ctx,                           \
 368                         const void *key, size_t ksz, const void *nonce) \
 369   {                                                                     \
 370     static const octet zerononce[XSALSA20_NONCESZ];                     \
 371                                                                         \
 372     populate(ctx->k, key, ksz);                                         \
 373     ctx->s.a[ 0] = SALSA20_A256;                                        \
 374     ctx->s.a[ 1] = SALSA20_B256;                                        \
 375     ctx->s.a[ 2] = SALSA20_C256;                                        \
 376     ctx->s.a[ 3] = SALSA20_D256;                                        \
 377     XSALSA20_SETNONCE(r, ctx, nonce ? nonce : zerononce);               \
 378   }
 379 SALSA20_VARS(DEFXINIT)
 380
 381 /* --- @xsalsa20{,12,8}_setnonce@ --- *
 382  *
 383  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 384  *              @const void *nonce@ = the nonce (@XSALSA20_NONCESZ@ bytes)
 385  *
 386  * Returns:     ---
 387  *
 388  * Use:         Set a new nonce in the context @ctx@, e.g., for processing a
 389  *              different message.  The stream position is reset to zero (see
 390  *              @salsa20_seek@ etc.).
 391  *
 392  *              There is a different function for each number of rounds,
 393  *              unlike for plain Salsa20.
 394  */
 395
 396 #define DEFXNONCE(r)                                                    \
 397   void XSALSA20_SETNONCE(r, XSALSA20_CTX(r) *ctx, const void *nonce)    \
 398   {                                                                     \
 399     const octet *n = nonce;                                             \
 400     uint32 in[4], out[8];                                               \
 401     int i;                                                              \
 402                                                                         \
 403     for (i = 0; i < 4; i++) in[i] = LOAD32_L(n + 4*i);                  \
 404     HSALSA20_RAW(r, ctx->k, in, out);                                   \
 405     for (i = 0; i < 4; i++) ctx->s.a[13 - 3*i] = out[i];                \
 406     for (i = 4; i < 8; i++) ctx->s.a[27 - 3*i] = out[i];                \
 407     salsa20_setnonce(&ctx->s, n + 16);                                  \
 408   }
 409 SALSA20_VARS(DEFXNONCE)
 410
 411 /* --- @xsalsa20{,12,8}_seek@, @xsalsa20{,12,8}_seeku64@ --- *
 412  *
 413  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 414  *              @unsigned long i@, @kludge64 i@ = new position to set
 415  *
 416  * Returns:     ---
 417  *
 418  * Use:         Sets a new stream position, in units of Salsa20 output
 419  *              blocks, which are @XSALSA20_OUTSZ@ bytes each.  Byte
 420  *              granularity can be achieved by calling @xsalsa20R_encrypt@
 421  *              appropriately.
 422  *
 423  *              There is a different function for each number of rounds,
 424  *              unlike for plain Salsa20, because the context structures are
 425  *              different.
 426  */
 427
 428 /* --- @xsalsa20{,12,8}_tell@, @xsalsa20{,12,8}_tellu64@ --- *
 429  *
 430  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 431  *
 432  * Returns:     The current position in the output stream, in blocks,
 433  *              rounding upwards.
 434  *
 435  *              There is a different function for each number of rounds,
 436  *              unlike for plain Salsa20, because the context structures are
 437  *              different.
 438  */
 439
 440 /* --- @xsalsa20{,12,8}_encrypt@ --- *
 441  *
 442  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 443  *              @const void *src@ = source buffer (or null)
 444  *              @void *dest@ = destination buffer (or null)
 445  *              @size_t sz@ = size of the buffers
 446  *
 447  * Returns:     ---
 448  *
 449  * Use:         Encrypts or decrypts @sz@ bytes of data from @src@ to @dest@.
 450  *              XSalsa20 works by XORing plaintext with a keystream, so
 451  *              encryption and decryption are the same operation.  If @dest@
 452  *              is null then ignore @src@ and skip @sz@ bytes of the
 453  *              keystream.  If @src@ is null, then just write the keystream
 454  *              to @dest@.
 455  */
 456
 457 #define DEFXPASSTHRU(r)                                                 \
 458   void XSALSA20_SEEK(r, XSALSA20_CTX(r) *ctx, unsigned long i)          \
 459     { salsa20_seek(&ctx->s, i); }                                       \
 460   void XSALSA20_SEEKU64(r, XSALSA20_CTX(r) *ctx, kludge64 i)            \
 461     { salsa20_seeku64(&ctx->s, i); }                                    \
 462   unsigned long XSALSA20_TELL(r, XSALSA20_CTX(r) *ctx)                  \
 463     { return salsa20_tell(&ctx->s); }                                   \
 464   kludge64 XSALSA20_TELLU64(r, XSALSA20_CTX(r) *ctx)                    \
 465     { return salsa20_tellu64(&ctx->s); }                                \
 466   void XSALSA20_ENCRYPT(r, XSALSA20_CTX(r) *ctx,                        \
 467                         const void *src, void *dest, size_t sz)         \
 468     { SALSA20_ENCRYPT(r, &ctx->s, src, dest, sz); }
 469 SALSA20_VARS(DEFXPASSTHRU)
 470
 471 /*----- Generic cipher interface ------------------------------------------*/
 472
 473 typedef struct gctx { gcipher c; salsa20_ctx ctx; } gctx;
 474
 475 static void gsetiv(gcipher *c, const void *iv)
 476   { gctx *g = (gctx *)c; salsa20_setnonce(&g->ctx, iv); }
 477
 478 static void gdestroy(gcipher *c)
 479   { gctx *g = (gctx *)c; BURN(*g); S_DESTROY(g); }
 480
 481 #define DEFGCIPHER(r)                                                   \
 482                                                                         \
 483   static const gcipher_ops gops_##r;                                    \
 484                                                                         \
 485   static gcipher *ginit_##r(const void *k, size_t sz)                   \
 486   {                                                                     \
 487     gctx *g = S_CREATE(gctx);                                           \
 488     g->c.ops = &gops_##r;                                               \
 489     salsa20_init(&g->ctx, k, sz, 0);                                    \
 490     return (&g->c);                                                     \
 491   }                                                                     \
 492                                                                         \
 493   static void gencrypt_##r(gcipher *c, const void *s,                   \
 494                            void *t, size_t sz)                          \
 495     { gctx *g = (gctx *)c; SALSA20_ENCRYPT(r, &g->ctx, s, t, sz); }     \
 496                                                                         \
 497   static const gcipher_ops gops_##r = {                                 \
 498     &SALSA20_DECOR(salsa20, r, ),                                       \
 499     gencrypt_##r, gencrypt_##r, gdestroy, gsetiv, 0                     \
 500   };                                                                    \
 501                                                                         \
 502   const gccipher SALSA20_DECOR(salsa20, r, ) = {                        \
 503     SALSA20_NAME_##r, salsa20_keysz,                                    \
 504     SALSA20_NONCESZ, ginit_##r                                          \
 505   };
 506
 507 SALSA20_VARS(DEFGCIPHER)
 508
 509 #define DEFGXCIPHER(r)                                                  \
 510                                                                         \
 511   typedef struct { gcipher c; XSALSA20_CTX(r) ctx; } gxctx_##r;         \
 512                                                                         \
 513   static void gxsetiv_##r(gcipher *c, const void *iv)                   \
 514     { gxctx_##r *g = (gxctx_##r *)c; XSALSA20_SETNONCE(r, &g->ctx, iv); } \
 515                                                                         \
 516   static void gxdestroy_##r(gcipher *c)                                 \
 517     { gxctx_##r *g = (gxctx_##r *)c; BURN(*g); S_DESTROY(g); }          \
 518                                                                         \
 519   static const gcipher_ops gxops_##r;                                   \
 520                                                                         \
 521   static gcipher *gxinit_##r(const void *k, size_t sz)                  \
 522   {                                                                     \
 523     gxctx_##r *g = S_CREATE(gxctx_##r);                                 \
 524     g->c.ops = &gxops_##r;                                              \
 525     XSALSA20_INIT(r, &g->ctx, k, sz, 0);                                \
 526     return (&g->c);                                                     \
 527   }                                                                     \
 528                                                                         \
 529   static void gxencrypt_##r(gcipher *c, const void *s,                  \
 530                             void *t, size_t sz)                         \
 531   {                                                                     \
 532     gxctx_##r *g = (gxctx_##r *)c;                                      \
 533     XSALSA20_ENCRYPT(r, &g->ctx, s, t, sz);                             \
 534   }                                                                     \
 535                                                                         \
 536   static const gcipher_ops gxops_##r = {                                \
 537     &SALSA20_DECOR(xsalsa20, r, ),                                      \
 538     gxencrypt_##r, gxencrypt_##r, gxdestroy_##r, gxsetiv_##r, 0         \
 539   };                                                                    \
 540                                                                         \
 541   const gccipher SALSA20_DECOR(xsalsa20, r, ) = {                       \
 542     "x" SALSA20_NAME_##r, salsa20_keysz,                                \
 543     XSALSA20_NONCESZ, gxinit_##r                                        \
 544   };
 545
 546 SALSA20_VARS(DEFGXCIPHER)
 547
 548 /*----- Generic random number generator interface -------------------------*/
 549
 550 typedef struct grops {
 551   size_t noncesz;
 552   void (*seek)(void *, kludge64);
 553   kludge64 (*tell)(void *);
 554   void (*setnonce)(void *, const void *);
 555   void (*generate)(void *, void *, size_t);
 556 } grops;
 557
 558 typedef struct grbasectx {
 559   grand r;
 560   const grops *ops;
 561 } grbasectx;
 562
 563 static int grmisc(grand *r, unsigned op, ...)
 564 {
 565   octet buf[XSALSA20_NONCESZ];
 566   grbasectx *g = (grbasectx *)r;
 567   grand *rr;
 568   const octet *p;
 569   size_t sz;
 570   uint32 i;
 571   unsigned long ul;
 572   kludge64 pos;
 573   va_list ap;
 574   int rc = 0;
 575
 576   va_start(ap, op);
 577
 578   switch (op) {
 579     case GRAND_CHECK:
 580       switch (va_arg(ap, unsigned)) {
 581         case GRAND_CHECK:
 582         case GRAND_SEEDINT:
 583         case GRAND_SEEDUINT32:
 584         case GRAND_SEEDBLOCK:
 585         case GRAND_SEEDRAND:
 586         case SALSA20_SEEK:
 587         case SALSA20_SEEKU64:
 588         case SALSA20_TELL:
 589         case SALSA20_TELLU64:
 590           rc = 1;
 591           break;
 592         default:
 593           rc = 0;
 594           break;
 595       }
 596       break;
 597
 598     case GRAND_SEEDINT:
 599       i = va_arg(ap, unsigned); STORE32_L(buf, i);
 600       memset(buf + 4, 0, g->ops->noncesz - 4);
 601       g->ops->setnonce(g, buf);
 602       break;
 603     case GRAND_SEEDUINT32:
 604       i = va_arg(ap, uint32); STORE32_L(buf, i);
 605       memset(buf + 4, 0, g->ops->noncesz - 4);
 606       g->ops->setnonce(g, buf);
 607       break;
 608     case GRAND_SEEDBLOCK:
 609       p = va_arg(ap, const void *);
 610       sz = va_arg(ap, size_t);
 611       if (sz < g->ops->noncesz) {
 612         memcpy(buf, p, sz);
 613         memset(buf + sz, 0, g->ops->noncesz - sz);
 614         p = buf;
 615       }
 616       g->ops->setnonce(g, p);
 617       break;
 618     case GRAND_SEEDRAND:
 619       rr = va_arg(ap, grand *);
 620       rr->ops->fill(rr, buf, g->ops->noncesz);
 621       g->ops->setnonce(g, buf);
 622       break;
 623     case SALSA20_SEEK:
 624       ul = va_arg(ap, unsigned long); ASSIGN64(pos, ul);
 625       g->ops->seek(g, pos);
 626       break;
 627     case SALSA20_SEEKU64:
 628       pos = va_arg(ap, kludge64);
 629       g->ops->seek(g, pos);
 630       break;
 631     case SALSA20_TELL:
 632       pos = g->ops->tell(g);
 633       *va_arg(ap, unsigned long *) = GET64(unsigned long, pos);
 634       break;
 635     case SALSA20_TELLU64:
 636       *va_arg(ap, kludge64 *) = g->ops->tell(g);
 637       break;
 638     default:
 639       GRAND_BADOP;
 640       break;
 641   }
 642
 643   return (rc);
 644 }
 645
 646 static octet grbyte(grand *r)
 647 {
 648   grbasectx *g = (grbasectx *)r;
 649   octet o;
 650   g->ops->generate(g, &o, 1);
 651   return (o);
 652 }
 653
 654 static uint32 grword(grand *r)
 655 {
 656   grbasectx *g = (grbasectx *)r;
 657   octet b[4];
 658   g->ops->generate(g, b, sizeof(b));
 659   return (LOAD32_L(b));
 660 }
 661
 662 static void grfill(grand *r, void *p, size_t sz)
 663 {
 664   grbasectx *g = (grbasectx *)r;
 665   g->ops->generate(r, p, sz);
 666 }
 667
 668 typedef struct grctx {
 669   grbasectx r;
 670   salsa20_ctx ctx;
 671 } grctx;
 672
 673 static void gr_seek(void *r, kludge64 pos)
 674   { grctx *g = r; salsa20_seeku64(&g->ctx, pos); }
 675
 676 static kludge64 gr_tell(void *r)
 677   { grctx *g = r; return (salsa20_tellu64(&g->ctx)); }
 678
 679 static void gr_setnonce(void *r, const void *n)
 680   { grctx *g = r; salsa20_setnonce(&g->ctx, n); }
 681
 682 static void grdestroy(grand *r)
 683   { grctx *g = (grctx *)r; BURN(*g); S_DESTROY(g); }
 684
 685 #define DEFGRAND(rr)                                                    \
 686                                                                         \
 687   static void gr_generate_##rr(void *r, void *b, size_t sz)             \
 688     { grctx *g = r; SALSA20_ENCRYPT(rr, &g->ctx, 0, b, sz); }           \
 689                                                                         \
 690   static const grops grops_##rr =                                       \
 691     { SALSA20_NONCESZ, gr_seek, gr_tell,                                \
 692       gr_setnonce, gr_generate_##rr };                                  \
 693                                                                         \
 694   static const grand_ops grops_rand_##rr = {                            \
 695     SALSA20_NAME_##rr, GRAND_CRYPTO, 0,                                 \
 696     grmisc, grdestroy, grword,                                          \
 697     grbyte, grword, grand_range, grfill                                 \
 698   };                                                                    \
 699                                                                         \
 700   grand *SALSA20_DECOR(salsa20, rr, _rand)                              \
 701     (const void *k, size_t ksz, const void *n)                          \
 702   {                                                                     \
 703     grctx *g = S_CREATE(g);                                             \
 704     g->r.r.ops = &grops_rand_##rr;                                      \
 705     g->r.ops = &grops_##rr;                                             \
 706     salsa20_init(&g->ctx, k, ksz, n);                                   \
 707     return (&g->r.r);                                                   \
 708   }
 709 SALSA20_VARS(DEFGRAND)
 710
 711 #define DEFXGRAND(rr)                                                   \
 712                                                                         \
 713   typedef struct grxctx_##rr {                                          \
 714     grbasectx r;                                                        \
 715     XSALSA20_CTX(rr) ctx;                                               \
 716   } grxctx_##rr;                                                        \
 717                                                                         \
 718   static void grx_seek_##rr(void *r, kludge64 pos)                      \
 719     { grxctx_##rr *g = r; XSALSA20_SEEKU64(rr, &g->ctx, pos); }         \
 720                                                                         \
 721   static kludge64 grx_tell_##rr(void *r)                                \
 722     { grxctx_##rr *g = r; return (XSALSA20_TELLU64(rr, &g->ctx)); }     \
 723                                                                         \
 724   static void grx_setnonce_##rr(void *r, const void *n)                 \
 725     { grxctx_##rr *g = r; XSALSA20_SETNONCE(rr, &g->ctx, n); }          \
 726                                                                         \
 727   static void grxdestroy_##rr(grand *r)                                 \
 728     { grxctx_##rr *g = (grxctx_##rr *)r; BURN(*g); S_DESTROY(g); }      \
 729                                                                         \
 730   static void grx_generate_##rr(void *r, void *b, size_t sz)            \
 731     { grxctx_##rr *g = r; XSALSA20_ENCRYPT(rr, &g->ctx, 0, b, sz); }    \
 732                                                                         \
 733   static const grops grxops_##rr =                                      \
 734   { XSALSA20_NONCESZ, grx_seek_##rr, grx_tell_##rr,                     \
 735       grx_setnonce_##rr, grx_generate_##rr };                           \
 736                                                                         \
 737   static const grand_ops grxops_rand_##rr = {                           \
 738     "x" SALSA20_NAME_##rr, GRAND_CRYPTO, 0,                             \
 739     grmisc, grxdestroy_##rr, grword,                                    \
 740     grbyte, grword, grand_range, grfill                                 \
 741   };                                                                    \
 742                                                                         \
 743   grand *SALSA20_DECOR(xsalsa20, rr, _rand)                             \
 744     (const void *k, size_t ksz, const void *n)                          \
 745   {                                                                     \
 746     grxctx_##rr *g = S_CREATE(g);                                       \
 747     g->r.r.ops = &grxops_rand_##rr;                                     \
 748     g->r.ops = &grxops_##rr;                                            \
 749     XSALSA20_INIT(rr, &g->ctx, k, ksz, n);                              \
 750     return (&g->r.r);                                                   \
 751   }
 752 SALSA20_VARS(DEFXGRAND)
 753
 754 /*----- Test rig ----------------------------------------------------------*/
 755
 756 #ifdef TEST_RIG
 757
 758 #include <stdio.h>
 759 #include <string.h>
 760
 761 #include <mLib/quis.h>
 762 #include <mLib/testrig.h>
 763
 764 static const int perm[] = {
 765    0, 13, 10,  7,
 766    4,  1, 14, 11,
 767    8,  5,  2, 15,
 768   12,  9,  6,  3
 769 };
 770
 771 #define DEFVCORE(r)                                                     \
 772   static int v_core_##r(dstr *v)                                        \
 773   {                                                                     \
 774     salsa20_matrix a, b;                                                \
 775     dstr d = DSTR_INIT;                                                 \
 776     int i, j, n;                                                        \
 777     int ok = 1;                                                         \
 778                                                                         \
 779     DENSURE(&d, SALSA20_OUTSZ); d.len = SALSA20_OUTSZ;                  \
 780     n = *(int *)v[0].buf;                                               \
 781     for (i = 0; i < SALSA20_OUTSZ/4; i++)                               \
 782       b[i] = LOAD32_L(v[1].buf + 4*i);                                  \
 783     for (i = 0; i < n; i++) {                                           \
 784       for (j = 0; j < 16; j++) a[perm[j]] = b[j];                       \
 785       core(r, a, b);                                                    \
 786       memcpy(a, b, sizeof(a));                                          \
 787     }                                                                   \
 788     for (i = 0; i < SALSA20_OUTSZ/4; i++) STORE32_L(d.buf + 4*i, b[i]); \
 789                                                                         \
 790     if (d.len != v[2].len || memcmp(d.buf, v[2].buf, v[2].len) != 0) {  \
 791       ok = 0;                                                           \
 792       printf("\nfail core:"                                             \
 793              "\n\titerations = %d"                                      \
 794              "\n\tin       = ", n);                                     \
 795       type_hex.dump(&v[1], stdout);                                     \
 796       printf("\n\texpected   = ");                                      \
 797       type_hex.dump(&v[2], stdout);                                     \
 798       printf("\n\tcalculated = ");                                      \
 799       type_hex.dump(&d, stdout);                                        \
 800       putchar('\n');                                                    \
 801     }                                                                   \
 802                                                                         \
 803     dstr_destroy(&d);                                                   \
 804     return (ok);                                                        \
 805   }
 806 SALSA20_VARS(DEFVCORE)
 807
 808 #define SALSA20_CTX(r) salsa20_ctx
 809 #define SALSA20_INIT(r, ctx, k, ksz, n) salsa20_init(ctx, k, ksz, n)
 810 #define SALSA20_SEEKU64(r, ctx, i) salsa20_seeku64(ctx, i)
 811
 812 #define DEFxVENC(base, BASE, r)                                         \
 813   static int v_encrypt_##base##_##r(dstr *v)                            \
 814   {                                                                     \
 815     BASE##_CTX(r) ctx;                                                  \
 816     dstr d = DSTR_INIT;                                                 \
 817     kludge64 pos;                                                       \
 818     const octet *p, *p0;                                                \
 819     octet *q;                                                           \
 820     size_t sz, sz0, step;                                               \
 821     unsigned long skip;                                                 \
 822     int ok = 1;                                                         \
 823                                                                         \
 824     if (v[4].len) { p0 = (const octet *)v[4].buf; sz0 = v[4].len; }     \
 825     else { p0 = 0; sz0 = v[5].len; }                                    \
 826     DENSURE(&d, sz0); d.len = sz0;                                      \
 827     skip = *(unsigned long *)v[3].buf;                                  \
 828                                                                         \
 829     step = 0;                                                           \
 830     while (step < sz0 + skip) {                                         \
 831       step = step ? 3*step + 4 : 1;                                     \
 832       if (step > sz0 + skip) step = sz0 + skip;                         \
 833       BASE##_INIT(r, &ctx, v[0].buf, v[0].len, v[1].buf);               \
 834       if (v[2].len) {                                                   \
 835         LOAD64_(pos, v[2].buf);                                         \
 836         BASE##_SEEKU64(r, &ctx, pos);                                   \
 837       }                                                                 \
 838                                                                         \
 839       for (sz = skip; sz >= step; sz -= step)                           \
 840         BASE##_ENCRYPT(r, &ctx, 0, 0, step);                            \
 841       if (sz) BASE##_ENCRYPT(r, &ctx, 0, 0, sz);                        \
 842       for (p = p0, q = (octet *)d.buf, sz = sz0;                        \
 843            sz >= step;                                                  \
 844            sz -= step, q += step) {                                     \
 845         BASE##_ENCRYPT(r, &ctx, p, q, step);                            \
 846         if (p) p += step;                                               \
 847       }                                                                 \
 848       if (sz) BASE##_ENCRYPT(r, &ctx, p, q, sz);                        \
 849                                                                         \
 850       if (d.len != v[5].len || memcmp(d.buf, v[5].buf, v[5].len) != 0) { \
 851         ok = 0;                                                         \
 852         printf("\nfail encrypt:"                                        \
 853                "\n\tstep           = %lu"                               \
 854                "\n\tkey    = ", (unsigned long)step);                   \
 855         type_hex.dump(&v[0], stdout);                                   \
 856         printf("\n\tnonce          = ");                                \
 857         type_hex.dump(&v[1], stdout);                                   \
 858         printf("\n\tposition   = ");                                    \
 859         type_hex.dump(&v[2], stdout);                                   \
 860         printf("\n\tskip           = %lu", skip);                       \
 861         printf("\n\tmessage    = ");                                    \
 862         type_hex.dump(&v[4], stdout);                                   \
 863         printf("\n\texpected   = ");                                    \
 864         type_hex.dump(&v[5], stdout);                                   \
 865         printf("\n\tcalculated = ");                                    \
 866         type_hex.dump(&d, stdout);                                      \
 867         putchar('\n');                                                  \
 868       }                                                                 \
 869     }                                                                   \
 870                                                                         \
 871     dstr_destroy(&d);                                                   \
 872     return (ok);                                                        \
 873   }
 874 #define DEFVENC(r) DEFxVENC(salsa20, SALSA20, r)
 875 #define DEFXVENC(r) DEFxVENC(xsalsa20, XSALSA20, r)
 876 SALSA20_VARS(DEFVENC)
 877 SALSA20_VARS(DEFXVENC)
 878
 879 static test_chunk defs[] = {
 880 #define DEFxTAB(pre, base, r)                                           \
 881   { pre SALSA20_NAME_##r, v_encrypt_##base##_##r,                       \
 882     { &type_hex, &type_hex, &type_hex, &type_ulong,                     \
 883       &type_hex, &type_hex, 0 } },
 884 #define DEFTAB(r)                                                       \
 885   { SALSA20_NAME_##r "-core", v_core_##r,                               \
 886     { &type_int, &type_hex, &type_hex, 0 } },                           \
 887   DEFxTAB("", salsa20, r)
 888 #define DEFXTAB(r) DEFxTAB("x", xsalsa20, r)
 889 SALSA20_VARS(DEFTAB)
 890 SALSA20_VARS(DEFXTAB)
 891   { 0, 0, { 0 } }
 892 };
 893
 894 int main(int argc, char *argv[])
 895 {
 896   test_run(argc, argv, defs, SRCDIR"/t/salsa20");
 897   return (0);
 898 }
 899
 900 #endif
 901
 902 /*----- That's all, folks -------------------------------------------------*/