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