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 #if CPUFAM_ARMEL
  60 extern core__functype salsa20_core_arm_neon;
  61 #endif
  62
  63 static core__functype *pick_core(void)
  64 {
  65 #if CPUFAM_X86 || CPUFAM_AMD64
  66   DISPATCH_PICK_COND(salsa20_core, salsa20_core_x86ish_sse2,
  67                      cpu_feature_p(CPUFEAT_X86_SSE2));
  68 #endif
  69 #if CPUFAM_ARMEL
  70   DISPATCH_PICK_COND(salsa20_core, salsa20_core_arm_neon,
  71                      cpu_feature_p(CPUFEAT_ARM_NEON));
  72 #endif
  73   DISPATCH_PICK_FALLBACK(salsa20_core, simple_core);
  74 }
  75
  76 /* --- @populate@ --- *
  77  *
  78  * Arguments:   @salsa20_matrix a@ = a matrix to fill in
  79  *              @const void *key@ = pointer to key material
  80  *              @size_t ksz@ = size of key
  81  *
  82  * Returns:     ---
  83  *
  84  * Use:         Fills in a Salsa20 matrix from the key, setting the
  85  *              appropriate constants according to the key length.  The nonce
  86  *              and position words are left uninitialized.
  87  */
  88
  89 static void populate(salsa20_matrix a, const void *key, size_t ksz)
  90 {
  91   const octet *k = key;
  92
  93   KSZ_ASSERT(salsa20, ksz);
  94
  95   /* Here's the pattern of key, constant, nonce, and counter pieces in the
  96    * matrix, before and after our permutation.
  97    *
  98    * [ C0 K0 K1 K2 ]       [ C0 C1 C2 C3 ]
  99    * [ K3 C1 N0 N1 ]  -->  [ K3 T1 K7 K2 ]
 100    * [ T0 T1 C2 K4 ]       [ T0 K6 K1 N1 ]
 101    * [ K5 K6 K7 C3 ]       [ K5 K0 N0 K4 ]
 102    */
 103
 104   a[13] = LOAD32_L(k +  0);
 105   a[10] = LOAD32_L(k +  4);
 106   if (ksz == 10) {
 107     a[ 7] = LOAD16_L(k +  8);
 108     a[ 4] = 0;
 109   } else {
 110     a[ 7] = LOAD32_L(k +  8);
 111     a[ 4] = LOAD32_L(k + 12);
 112   }
 113   if (ksz <= 16) {
 114     a[15] = a[13];
 115     a[12] = a[10];
 116     a[ 9] = a[ 7];
 117     a[ 6] = a[ 4];
 118     a[ 0] = SALSA20_A128;
 119     a[ 1] = SALSA20_B128;
 120     a[ 2] = ksz == 10 ? SALSA20_C80 : SALSA20_C128;
 121     a[ 3] = SALSA20_D128;
 122   } else {
 123     a[15] = LOAD32_L(k + 16);
 124     a[12] = LOAD32_L(k + 20);
 125     a[ 9] = LOAD32_L(k + 24);
 126     a[ 6] = LOAD32_L(k + 28);
 127     a[ 0] = SALSA20_A256;
 128     a[ 1] = SALSA20_B256;
 129     a[ 2] = SALSA20_C256;
 130     a[ 3] = SALSA20_D256;
 131   }
 132 }
 133
 134 /*----- Salsa20 implementation --------------------------------------------*/
 135
 136 /* --- @salsa20_init@ --- *
 137  *
 138  * Arguments:   @salsa20_ctx *ctx@ = context to fill in
 139  *              @const void *key@ = pointer to key material
 140  *              @size_t ksz@ = size of key (either 32 or 16)
 141  *              @const void *nonce@ = initial nonce, or null
 142  *
 143  * Returns:     ---
 144  *
 145  * Use:         Initializes a Salsa20 context ready for use.
 146  */
 147
 148 void salsa20_init(salsa20_ctx *ctx, const void *key, size_t ksz,
 149                   const void *nonce)
 150 {
 151   static const octet zerononce[SALSA20_NONCESZ];
 152
 153   populate(ctx->a, key, ksz);
 154   salsa20_setnonce(ctx, nonce ? nonce : zerononce);
 155 }
 156
 157 /* --- @salsa20_setnonce@ --- *
 158  *
 159  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 160  *              @const void *nonce@ = the nonce (@SALSA20_NONCESZ@ bytes)
 161  *
 162  * Returns:     ---
 163  *
 164  * Use:         Set a new nonce in the context @ctx@, e.g., for processing a
 165  *              different message.  The stream position is reset to zero (see
 166  *              @salsa20_seek@ etc.).
 167  */
 168
 169 void salsa20_setnonce(salsa20_ctx *ctx, const void *nonce)
 170 {
 171   const octet *n = nonce;
 172
 173   ctx->a[14] = LOAD32_L(n + 0);
 174   ctx->a[11] = LOAD32_L(n + 4);
 175   salsa20_seek(ctx, 0);
 176 }
 177
 178 /* --- @salsa20_seek@, @salsa20_seeku64@ --- *
 179  *
 180  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 181  *              @unsigned long i@, @kludge64 i@ = new position to set
 182  *
 183  * Returns:     ---
 184  *
 185  * Use:         Sets a new stream position, in units of Salsa20 output
 186  *              blocks, which are @SALSA20_OUTSZ@ bytes each.  Byte
 187  *              granularity can be achieved by calling @salsa20R_encrypt@
 188  *              appropriately.
 189  */
 190
 191 void salsa20_seek(salsa20_ctx *ctx, unsigned long i)
 192   { kludge64 ii; ASSIGN64(ii, i); salsa20_seeku64(ctx, ii); }
 193
 194 void salsa20_seeku64(salsa20_ctx *ctx, kludge64 i)
 195 {
 196   ctx->a[8] = LO64(i); ctx->a[5] = HI64(i);
 197   ctx->bufi = SALSA20_OUTSZ;
 198 }
 199
 200 /* --- @salsa20_tell@, @salsa20_tellu64@ --- *
 201  *
 202  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 203  *
 204  * Returns:     The current position in the output stream, in blocks,
 205  *              rounding upwards.
 206  */
 207
 208 unsigned long salsa20_tell(salsa20_ctx *ctx)
 209   { kludge64 i = salsa20_tellu64(ctx); return (GET64(unsigned long, i)); }
 210
 211 kludge64 salsa20_tellu64(salsa20_ctx *ctx)
 212   { kludge64 i; SET64(i, ctx->a[5], ctx->a[8]); return (i); }
 213
 214 /* --- @salsa20{,12,8}_encrypt@ --- *
 215  *
 216  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 217  *              @const void *src@ = source buffer (or null)
 218  *              @void *dest@ = destination buffer (or null)
 219  *              @size_t sz@ = size of the buffers
 220  *
 221  * Returns:     ---
 222  *
 223  * Use:         Encrypts or decrypts @sz@ bytes of data from @src@ to @dest@.
 224  *              Salsa20 works by XORing plaintext with a keystream, so
 225  *              encryption and decryption are the same operation.  If @dest@
 226  *              is null then ignore @src@ and skip @sz@ bytes of the
 227  *              keystream.  If @src@ is null, then just write the keystream
 228  *              to @dest@.
 229  */
 230
 231 #define SALSA20_ENCRYPT(r, ctx, src, dest, sz)                          \
 232   SALSA20_DECOR(salsa20, r, _encrypt)(ctx, src, dest, sz)
 233 #define DEFENCRYPT(r)                                                   \
 234   void SALSA20_ENCRYPT(r, salsa20_ctx *ctx, const void *src,            \
 235                        void *dest, size_t sz)                           \
 236   {                                                                     \
 237     salsa20_matrix b;                                                   \
 238     const octet *s = src;                                               \
 239     octet *d = dest;                                                    \
 240     size_t n;                                                           \
 241     kludge64 pos, delta;                                                \
 242                                                                         \
 243     SALSA20_OUTBUF(ctx, d, s, sz);                                      \
 244     if (!sz) return;                                                    \
 245                                                                         \
 246     if (!dest) {                                                        \
 247       n = sz/SALSA20_OUTSZ;                                             \
 248       pos = salsa20_tellu64(ctx);                                       \
 249       ASSIGN64(delta, n);                                               \
 250       ADD64(pos, pos, delta);                                           \
 251       salsa20_seeku64(ctx, pos);                                        \
 252       sz = sz%SALSA20_OUTSZ;                                            \
 253     } else if (!src) {                                                  \
 254       while (sz >= SALSA20_OUTSZ) {                                     \
 255         core(r, ctx->a, b);                                             \
 256         SALSA20_STEP(ctx->a);                                           \
 257         SALSA20_GENFULL(b, d);                                          \
 258         sz -= SALSA20_OUTSZ;                                            \
 259       }                                                                 \
 260     } else {                                                            \
 261       while (sz >= SALSA20_OUTSZ) {                                     \
 262         core(r, ctx->a, b);                                             \
 263         SALSA20_STEP(ctx->a);                                           \
 264         SALSA20_MIXFULL(b, d, s);                                       \
 265         sz -= SALSA20_OUTSZ;                                            \
 266       }                                                                 \
 267     }                                                                   \
 268                                                                         \
 269     if (sz) {                                                           \
 270       core(r, ctx->a, b);                                               \
 271       SALSA20_STEP(ctx->a);                                             \
 272       SALSA20_PREPBUF(ctx, b);                                          \
 273       SALSA20_OUTBUF(ctx, d, s, sz);                                    \
 274       assert(!sz);                                                      \
 275     }                                                                   \
 276   }
 277 SALSA20_VARS(DEFENCRYPT)
 278
 279 /*----- HSalsa20 implementation -------------------------------------------*/
 280
 281 #define HSALSA20_RAW(r, ctx, src, dest)                                 \
 282   SALSA20_DECOR(hsalsa20, r, _raw)(ctx, src, dest)
 283 #define HSALSA20_PRF(r, ctx, src, dest)                                 \
 284   SALSA20_DECOR(hsalsa20, r, _prf)(ctx, src, dest)
 285
 286 /* --- @hsalsa20{,12,8}_prf@ --- *
 287  *
 288  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 289  *              @const void *src@ = the input (@HSALSA20_INSZ@ bytes)
 290  *              @void *dest@ = the output (@HSALSA20_OUTSZ@ bytes)
 291  *
 292  * Returns:     ---
 293  *
 294  * Use:         Apply the HSalsa20/r pseudorandom function to @src@, writing
 295  *              the result to @out@.
 296  */
 297
 298 #define DEFHSALSA20(r)                                                  \
 299   static void HSALSA20_RAW(r, salsa20_matrix k,                         \
 300                            const uint32 *src, uint32 *dest)             \
 301   {                                                                     \
 302     salsa20_matrix a;                                                   \
 303     int i;                                                              \
 304                                                                         \
 305     /* --- HSalsa20, computed from full Salsa20 --- *                   \
 306      *                                                                  \
 307      * The security proof makes use of the fact that HSalsa20 (i.e.,    \
 308      * without the final feedforward step) can be computed from full    \
 309      * Salsa20 using only knowledge of the non-secret input.  I don't   \
 310      * want to compromise the performance of the main function by       \
 311      * making the feedforward step separate, but this operation is less \
 312      * speed critical, so we do it the harder way.                      \
 313      */                                                                 \
 314                                                                         \
 315     for (i = 0; i < 4; i++) k[14 - 3*i] = src[i];                       \
 316     core(r, k, a);                                                      \
 317     for (i = 0; i < 4; i++) dest[i] = a[5*i] - k[i];                    \
 318     for (i = 4; i < 8; i++) dest[i] = a[i + 2] - k[26 - 3*i];           \
 319   }                                                                     \
 320                                                                         \
 321   void HSALSA20_PRF(r, salsa20_ctx *ctx, const void *src, void *dest)   \
 322   {                                                                     \
 323     const octet *s = src;                                               \
 324     octet *d = dest;                                                    \
 325     uint32 in[4], out[8];                                               \
 326     int i;                                                              \
 327                                                                         \
 328     for (i = 0; i < 4; i++) in[i] = LOAD32_L(s + 4*i);                  \
 329     HSALSA20_RAW(r, ctx->a, in, out);                                   \
 330     for (i = 0; i < 8; i++) STORE32_L(d + 4*i, out[i]);                 \
 331   }
 332 SALSA20_VARS(DEFHSALSA20)
 333
 334 /*----- XSalsa20 implementation -------------------------------------------*/
 335
 336 /* --- Some convenient macros for naming functions --- *
 337  *
 338  * Because the crypto core is involved in XSalsa20/r's per-nonce setup, we
 339  * need to take an interest in the number of rounds in most of the various
 340  * functions, and it will probably help if we distinguish the context
 341  * structures for the various versions.
 342  */
 343
 344 #define XSALSA20_CTX(r) SALSA20_DECOR(xsalsa20, r, _ctx)
 345 #define XSALSA20_INIT(r, ctx, k, ksz, n)                                \
 346   SALSA20_DECOR(xsalsa20, r, _init)(ctx, k, ksz, n)
 347 #define XSALSA20_SETNONCE(r, ctx, n)                                    \
 348   SALSA20_DECOR(xsalsa20, r, _setnonce)(ctx, n)
 349 #define XSALSA20_SEEK(r, ctx, i)                                        \
 350   SALSA20_DECOR(xsalsa20, r, _seek)(ctx, i)
 351 #define XSALSA20_SEEKU64(r, ctx, i)                                     \
 352   SALSA20_DECOR(xsalsa20, r, _seeku64)(ctx, i)
 353 #define XSALSA20_TELL(r, ctx)                                           \
 354   SALSA20_DECOR(xsalsa20, r, _tell)(ctx)
 355 #define XSALSA20_TELLU64(r, ctx)                                        \
 356   SALSA20_DECOR(xsalsa20, r, _tellu64)(ctx)
 357 #define XSALSA20_ENCRYPT(r, ctx, src, dest, sz)                         \
 358   SALSA20_DECOR(xsalsa20, r, _encrypt)(ctx, src, dest, sz)
 359
 360 /* --- @xsalsa20{,12,8}_init@ --- *
 361  *
 362  * Arguments:   @xsalsa20R_ctx *ctx@ = the context to fill in
 363  *              @const void *key@ = pointer to key material
 364  *              @size_t ksz@ = size of key (either 32 or 16)
 365  *              @const void *nonce@ = initial nonce, or null
 366  *
 367  * Returns:     ---
 368  *
 369  * Use:         Initializes an XSalsa20/r context ready for use.
 370  *
 371  *              There is a different function for each number of rounds,
 372  *              unlike for plain Salsa20.
 373  */
 374
 375 #define DEFXINIT(r)                                                     \
 376   void XSALSA20_INIT(r, XSALSA20_CTX(r) *ctx,                           \
 377                         const void *key, size_t ksz, const void *nonce) \
 378   {                                                                     \
 379     static const octet zerononce[XSALSA20_NONCESZ];                     \
 380                                                                         \
 381     populate(ctx->k, key, ksz);                                         \
 382     ctx->s.a[ 0] = SALSA20_A256;                                        \
 383     ctx->s.a[ 1] = SALSA20_B256;                                        \
 384     ctx->s.a[ 2] = SALSA20_C256;                                        \
 385     ctx->s.a[ 3] = SALSA20_D256;                                        \
 386     XSALSA20_SETNONCE(r, ctx, nonce ? nonce : zerononce);               \
 387   }
 388 SALSA20_VARS(DEFXINIT)
 389
 390 /* --- @xsalsa20{,12,8}_setnonce@ --- *
 391  *
 392  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 393  *              @const void *nonce@ = the nonce (@XSALSA20_NONCESZ@ bytes)
 394  *
 395  * Returns:     ---
 396  *
 397  * Use:         Set a new nonce in the context @ctx@, e.g., for processing a
 398  *              different message.  The stream position is reset to zero (see
 399  *              @salsa20_seek@ etc.).
 400  *
 401  *              There is a different function for each number of rounds,
 402  *              unlike for plain Salsa20.
 403  */
 404
 405 #define DEFXNONCE(r)                                                    \
 406   void XSALSA20_SETNONCE(r, XSALSA20_CTX(r) *ctx, const void *nonce)    \
 407   {                                                                     \
 408     const octet *n = nonce;                                             \
 409     uint32 in[4], out[8];                                               \
 410     int i;                                                              \
 411                                                                         \
 412     for (i = 0; i < 4; i++) in[i] = LOAD32_L(n + 4*i);                  \
 413     HSALSA20_RAW(r, ctx->k, in, out);                                   \
 414     for (i = 0; i < 4; i++) ctx->s.a[13 - 3*i] = out[i];                \
 415     for (i = 4; i < 8; i++) ctx->s.a[27 - 3*i] = out[i];                \
 416     salsa20_setnonce(&ctx->s, n + 16);                                  \
 417   }
 418 SALSA20_VARS(DEFXNONCE)
 419
 420 /* --- @xsalsa20{,12,8}_seek@, @xsalsa20{,12,8}_seeku64@ --- *
 421  *
 422  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 423  *              @unsigned long i@, @kludge64 i@ = new position to set
 424  *
 425  * Returns:     ---
 426  *
 427  * Use:         Sets a new stream position, in units of Salsa20 output
 428  *              blocks, which are @XSALSA20_OUTSZ@ bytes each.  Byte
 429  *              granularity can be achieved by calling @xsalsa20R_encrypt@
 430  *              appropriately.
 431  *
 432  *              There is a different function for each number of rounds,
 433  *              unlike for plain Salsa20, because the context structures are
 434  *              different.
 435  */
 436
 437 /* --- @xsalsa20{,12,8}_tell@, @xsalsa20{,12,8}_tellu64@ --- *
 438  *
 439  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 440  *
 441  * Returns:     The current position in the output stream, in blocks,
 442  *              rounding upwards.
 443  *
 444  *              There is a different function for each number of rounds,
 445  *              unlike for plain Salsa20, because the context structures are
 446  *              different.
 447  */
 448
 449 /* --- @xsalsa20{,12,8}_encrypt@ --- *
 450  *
 451  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 452  *              @const void *src@ = source buffer (or null)
 453  *              @void *dest@ = destination buffer (or null)
 454  *              @size_t sz@ = size of the buffers
 455  *
 456  * Returns:     ---
 457  *
 458  * Use:         Encrypts or decrypts @sz@ bytes of data from @src@ to @dest@.
 459  *              XSalsa20 works by XORing plaintext with a keystream, so
 460  *              encryption and decryption are the same operation.  If @dest@
 461  *              is null then ignore @src@ and skip @sz@ bytes of the
 462  *              keystream.  If @src@ is null, then just write the keystream
 463  *              to @dest@.
 464  */
 465
 466 #define DEFXPASSTHRU(r)                                                 \
 467   void XSALSA20_SEEK(r, XSALSA20_CTX(r) *ctx, unsigned long i)          \
 468     { salsa20_seek(&ctx->s, i); }                                       \
 469   void XSALSA20_SEEKU64(r, XSALSA20_CTX(r) *ctx, kludge64 i)            \
 470     { salsa20_seeku64(&ctx->s, i); }                                    \
 471   unsigned long XSALSA20_TELL(r, XSALSA20_CTX(r) *ctx)                  \
 472     { return salsa20_tell(&ctx->s); }                                   \
 473   kludge64 XSALSA20_TELLU64(r, XSALSA20_CTX(r) *ctx)                    \
 474     { return salsa20_tellu64(&ctx->s); }                                \
 475   void XSALSA20_ENCRYPT(r, XSALSA20_CTX(r) *ctx,                        \
 476                         const void *src, void *dest, size_t sz)         \
 477     { SALSA20_ENCRYPT(r, &ctx->s, src, dest, sz); }
 478 SALSA20_VARS(DEFXPASSTHRU)
 479
 480 /*----- Generic cipher interface ------------------------------------------*/
 481
 482 typedef struct gctx { gcipher c; salsa20_ctx ctx; } gctx;
 483
 484 static void gsetiv(gcipher *c, const void *iv)
 485   { gctx *g = (gctx *)c; salsa20_setnonce(&g->ctx, iv); }
 486
 487 static void gdestroy(gcipher *c)
 488   { gctx *g = (gctx *)c; BURN(*g); S_DESTROY(g); }
 489
 490 #define DEFGCIPHER(r)                                                   \
 491                                                                         \
 492   static const gcipher_ops gops_##r;                                    \
 493                                                                         \
 494   static gcipher *ginit_##r(const void *k, size_t sz)                   \
 495   {                                                                     \
 496     gctx *g = S_CREATE(gctx);                                           \
 497     g->c.ops = &gops_##r;                                               \
 498     salsa20_init(&g->ctx, k, sz, 0);                                    \
 499     return (&g->c);                                                     \
 500   }                                                                     \
 501                                                                         \
 502   static void gencrypt_##r(gcipher *c, const void *s,                   \
 503                            void *t, size_t sz)                          \
 504     { gctx *g = (gctx *)c; SALSA20_ENCRYPT(r, &g->ctx, s, t, sz); }     \
 505                                                                         \
 506   static const gcipher_ops gops_##r = {                                 \
 507     &SALSA20_DECOR(salsa20, r, ),                                       \
 508     gencrypt_##r, gencrypt_##r, gdestroy, gsetiv, 0                     \
 509   };                                                                    \
 510                                                                         \
 511   const gccipher SALSA20_DECOR(salsa20, r, ) = {                        \
 512     SALSA20_NAME_##r, salsa20_keysz,                                    \
 513     SALSA20_NONCESZ, ginit_##r                                          \
 514   };
 515
 516 SALSA20_VARS(DEFGCIPHER)
 517
 518 #define DEFGXCIPHER(r)                                                  \
 519                                                                         \
 520   typedef struct { gcipher c; XSALSA20_CTX(r) ctx; } gxctx_##r;         \
 521                                                                         \
 522   static void gxsetiv_##r(gcipher *c, const void *iv)                   \
 523     { gxctx_##r *g = (gxctx_##r *)c; XSALSA20_SETNONCE(r, &g->ctx, iv); } \
 524                                                                         \
 525   static void gxdestroy_##r(gcipher *c)                                 \
 526     { gxctx_##r *g = (gxctx_##r *)c; BURN(*g); S_DESTROY(g); }          \
 527                                                                         \
 528   static const gcipher_ops gxops_##r;                                   \
 529                                                                         \
 530   static gcipher *gxinit_##r(const void *k, size_t sz)                  \
 531   {                                                                     \
 532     gxctx_##r *g = S_CREATE(gxctx_##r);                                 \
 533     g->c.ops = &gxops_##r;                                              \
 534     XSALSA20_INIT(r, &g->ctx, k, sz, 0);                                \
 535     return (&g->c);                                                     \
 536   }                                                                     \
 537                                                                         \
 538   static void gxencrypt_##r(gcipher *c, const void *s,                  \
 539                             void *t, size_t sz)                         \
 540   {                                                                     \
 541     gxctx_##r *g = (gxctx_##r *)c;                                      \
 542     XSALSA20_ENCRYPT(r, &g->ctx, s, t, sz);                             \
 543   }                                                                     \
 544                                                                         \
 545   static const gcipher_ops gxops_##r = {                                \
 546     &SALSA20_DECOR(xsalsa20, r, ),                                      \
 547     gxencrypt_##r, gxencrypt_##r, gxdestroy_##r, gxsetiv_##r, 0         \
 548   };                                                                    \
 549                                                                         \
 550   const gccipher SALSA20_DECOR(xsalsa20, r, ) = {                       \
 551     "x" SALSA20_NAME_##r, salsa20_keysz,                                \
 552     XSALSA20_NONCESZ, gxinit_##r                                        \
 553   };
 554
 555 SALSA20_VARS(DEFGXCIPHER)
 556
 557 /*----- Generic random number generator interface -------------------------*/
 558
 559 typedef struct grops {
 560   size_t noncesz;
 561   void (*seek)(void *, kludge64);
 562   kludge64 (*tell)(void *);
 563   void (*setnonce)(void *, const void *);
 564   void (*generate)(void *, void *, size_t);
 565 } grops;
 566
 567 typedef struct grbasectx {
 568   grand r;
 569   const grops *ops;
 570 } grbasectx;
 571
 572 static int grmisc(grand *r, unsigned op, ...)
 573 {
 574   octet buf[XSALSA20_NONCESZ];
 575   grbasectx *g = (grbasectx *)r;
 576   grand *rr;
 577   const octet *p;
 578   size_t sz;
 579   uint32 i;
 580   unsigned long ul;
 581   kludge64 pos;
 582   va_list ap;
 583   int rc = 0;
 584
 585   va_start(ap, op);
 586
 587   switch (op) {
 588     case GRAND_CHECK:
 589       switch (va_arg(ap, unsigned)) {
 590         case GRAND_CHECK:
 591         case GRAND_SEEDINT:
 592         case GRAND_SEEDUINT32:
 593         case GRAND_SEEDBLOCK:
 594         case GRAND_SEEDRAND:
 595         case SALSA20_SEEK:
 596         case SALSA20_SEEKU64:
 597         case SALSA20_TELL:
 598         case SALSA20_TELLU64:
 599           rc = 1;
 600           break;
 601         default:
 602           rc = 0;
 603           break;
 604       }
 605       break;
 606
 607     case GRAND_SEEDINT:
 608       i = va_arg(ap, unsigned); STORE32_L(buf, i);
 609       memset(buf + 4, 0, g->ops->noncesz - 4);
 610       g->ops->setnonce(g, buf);
 611       break;
 612     case GRAND_SEEDUINT32:
 613       i = va_arg(ap, uint32); STORE32_L(buf, i);
 614       memset(buf + 4, 0, g->ops->noncesz - 4);
 615       g->ops->setnonce(g, buf);
 616       break;
 617     case GRAND_SEEDBLOCK:
 618       p = va_arg(ap, const void *);
 619       sz = va_arg(ap, size_t);
 620       if (sz < g->ops->noncesz) {
 621         memcpy(buf, p, sz);
 622         memset(buf + sz, 0, g->ops->noncesz - sz);
 623         p = buf;
 624       }
 625       g->ops->setnonce(g, p);
 626       break;
 627     case GRAND_SEEDRAND:
 628       rr = va_arg(ap, grand *);
 629       rr->ops->fill(rr, buf, g->ops->noncesz);
 630       g->ops->setnonce(g, buf);
 631       break;
 632     case SALSA20_SEEK:
 633       ul = va_arg(ap, unsigned long); ASSIGN64(pos, ul);
 634       g->ops->seek(g, pos);
 635       break;
 636     case SALSA20_SEEKU64:
 637       pos = va_arg(ap, kludge64);
 638       g->ops->seek(g, pos);
 639       break;
 640     case SALSA20_TELL:
 641       pos = g->ops->tell(g);
 642       *va_arg(ap, unsigned long *) = GET64(unsigned long, pos);
 643       break;
 644     case SALSA20_TELLU64:
 645       *va_arg(ap, kludge64 *) = g->ops->tell(g);
 646       break;
 647     default:
 648       GRAND_BADOP;
 649       break;
 650   }
 651
 652   return (rc);
 653 }
 654
 655 static octet grbyte(grand *r)
 656 {
 657   grbasectx *g = (grbasectx *)r;
 658   octet o;
 659   g->ops->generate(g, &o, 1);
 660   return (o);
 661 }
 662
 663 static uint32 grword(grand *r)
 664 {
 665   grbasectx *g = (grbasectx *)r;
 666   octet b[4];
 667   g->ops->generate(g, b, sizeof(b));
 668   return (LOAD32_L(b));
 669 }
 670
 671 static void grfill(grand *r, void *p, size_t sz)
 672 {
 673   grbasectx *g = (grbasectx *)r;
 674   g->ops->generate(r, p, sz);
 675 }
 676
 677 typedef struct grctx {
 678   grbasectx r;
 679   salsa20_ctx ctx;
 680 } grctx;
 681
 682 static void gr_seek(void *r, kludge64 pos)
 683   { grctx *g = r; salsa20_seeku64(&g->ctx, pos); }
 684
 685 static kludge64 gr_tell(void *r)
 686   { grctx *g = r; return (salsa20_tellu64(&g->ctx)); }
 687
 688 static void gr_setnonce(void *r, const void *n)
 689   { grctx *g = r; salsa20_setnonce(&g->ctx, n); }
 690
 691 static void grdestroy(grand *r)
 692   { grctx *g = (grctx *)r; BURN(*g); S_DESTROY(g); }
 693
 694 #define DEFGRAND(rr)                                                    \
 695                                                                         \
 696   static void gr_generate_##rr(void *r, void *b, size_t sz)             \
 697     { grctx *g = r; SALSA20_ENCRYPT(rr, &g->ctx, 0, b, sz); }           \
 698                                                                         \
 699   static const grops grops_##rr =                                       \
 700     { SALSA20_NONCESZ, gr_seek, gr_tell,                                \
 701       gr_setnonce, gr_generate_##rr };                                  \
 702                                                                         \
 703   static const grand_ops grops_rand_##rr = {                            \
 704     SALSA20_NAME_##rr, GRAND_CRYPTO, 0,                                 \
 705     grmisc, grdestroy, grword,                                          \
 706     grbyte, grword, grand_defaultrange, grfill                          \
 707   };                                                                    \
 708                                                                         \
 709   grand *SALSA20_DECOR(salsa20, rr, _rand)                              \
 710     (const void *k, size_t ksz, const void *n)                          \
 711   {                                                                     \
 712     grctx *g = S_CREATE(g);                                             \
 713     g->r.r.ops = &grops_rand_##rr;                                      \
 714     g->r.ops = &grops_##rr;                                             \
 715     salsa20_init(&g->ctx, k, ksz, n);                                   \
 716     return (&g->r.r);                                                   \
 717   }
 718 SALSA20_VARS(DEFGRAND)
 719
 720 #define DEFXGRAND(rr)                                                   \
 721                                                                         \
 722   typedef struct grxctx_##rr {                                          \
 723     grbasectx r;                                                        \
 724     XSALSA20_CTX(rr) ctx;                                               \
 725   } grxctx_##rr;                                                        \
 726                                                                         \
 727   static void grx_seek_##rr(void *r, kludge64 pos)                      \
 728     { grxctx_##rr *g = r; XSALSA20_SEEKU64(rr, &g->ctx, pos); }         \
 729                                                                         \
 730   static kludge64 grx_tell_##rr(void *r)                                \
 731     { grxctx_##rr *g = r; return (XSALSA20_TELLU64(rr, &g->ctx)); }     \
 732                                                                         \
 733   static void grx_setnonce_##rr(void *r, const void *n)                 \
 734     { grxctx_##rr *g = r; XSALSA20_SETNONCE(rr, &g->ctx, n); }          \
 735                                                                         \
 736   static void grxdestroy_##rr(grand *r)                                 \
 737     { grxctx_##rr *g = (grxctx_##rr *)r; BURN(*g); S_DESTROY(g); }      \
 738                                                                         \
 739   static void grx_generate_##rr(void *r, void *b, size_t sz)            \
 740     { grxctx_##rr *g = r; XSALSA20_ENCRYPT(rr, &g->ctx, 0, b, sz); }    \
 741                                                                         \
 742   static const grops grxops_##rr =                                      \
 743   { XSALSA20_NONCESZ, grx_seek_##rr, grx_tell_##rr,                     \
 744       grx_setnonce_##rr, grx_generate_##rr };                           \
 745                                                                         \
 746   static const grand_ops grxops_rand_##rr = {                           \
 747     "x" SALSA20_NAME_##rr, GRAND_CRYPTO, 0,                             \
 748     grmisc, grxdestroy_##rr, grword,                                    \
 749     grbyte, grword, grand_defaultrange, grfill                          \
 750   };                                                                    \
 751                                                                         \
 752   grand *SALSA20_DECOR(xsalsa20, rr, _rand)                             \
 753     (const void *k, size_t ksz, const void *n)                          \
 754   {                                                                     \
 755     grxctx_##rr *g = S_CREATE(g);                                       \
 756     g->r.r.ops = &grxops_rand_##rr;                                     \
 757     g->r.ops = &grxops_##rr;                                            \
 758     XSALSA20_INIT(rr, &g->ctx, k, ksz, n);                              \
 759     return (&g->r.r);                                                   \
 760   }
 761 SALSA20_VARS(DEFXGRAND)
 762
 763 /*----- Test rig ----------------------------------------------------------*/
 764
 765 #ifdef TEST_RIG
 766
 767 #include <stdio.h>
 768 #include <string.h>
 769
 770 #include <mLib/quis.h>
 771 #include <mLib/testrig.h>
 772
 773 static const int perm[] = {
 774    0, 13, 10,  7,
 775    4,  1, 14, 11,
 776    8,  5,  2, 15,
 777   12,  9,  6,  3
 778 };
 779
 780 #define DEFVCORE(r)                                                     \
 781   static int v_core_##r(dstr *v)                                        \
 782   {                                                                     \
 783     salsa20_matrix a, b;                                                \
 784     dstr d = DSTR_INIT;                                                 \
 785     int i, j, n;                                                        \
 786     int ok = 1;                                                         \
 787                                                                         \
 788     DENSURE(&d, SALSA20_OUTSZ); d.len = SALSA20_OUTSZ;                  \
 789     n = *(int *)v[0].buf;                                               \
 790     for (i = 0; i < SALSA20_OUTSZ/4; i++)                               \
 791       b[i] = LOAD32_L(v[1].buf + 4*i);                                  \
 792     for (i = 0; i < n; i++) {                                           \
 793       for (j = 0; j < 16; j++) a[perm[j]] = b[j];                       \
 794       core(r, a, b);                                                    \
 795       memcpy(a, b, sizeof(a));                                          \
 796     }                                                                   \
 797     for (i = 0; i < SALSA20_OUTSZ/4; i++) STORE32_L(d.buf + 4*i, b[i]); \
 798                                                                         \
 799     if (d.len != v[2].len || memcmp(d.buf, v[2].buf, v[2].len) != 0) {  \
 800       ok = 0;                                                           \
 801       printf("\nfail core:"                                             \
 802              "\n\titerations = %d"                                      \
 803              "\n\tin       = ", n);                                     \
 804       type_hex.dump(&v[1], stdout);                                     \
 805       printf("\n\texpected   = ");                                      \
 806       type_hex.dump(&v[2], stdout);                                     \
 807       printf("\n\tcalculated = ");                                      \
 808       type_hex.dump(&d, stdout);                                        \
 809       putchar('\n');                                                    \
 810     }                                                                   \
 811                                                                         \
 812     dstr_destroy(&d);                                                   \
 813     return (ok);                                                        \
 814   }
 815 SALSA20_VARS(DEFVCORE)
 816
 817 #define SALSA20_CTX(r) salsa20_ctx
 818 #define SALSA20_INIT(r, ctx, k, ksz, n) salsa20_init(ctx, k, ksz, n)
 819 #define SALSA20_SEEKU64(r, ctx, i) salsa20_seeku64(ctx, i)
 820
 821 #define DEFxVENC(base, BASE, r)                                         \
 822   static int v_encrypt_##base##_##r(dstr *v)                            \
 823   {                                                                     \
 824     BASE##_CTX(r) ctx;                                                  \
 825     dstr d = DSTR_INIT;                                                 \
 826     kludge64 pos;                                                       \
 827     const octet *p, *p0;                                                \
 828     octet *q;                                                           \
 829     size_t sz, sz0, step;                                               \
 830     unsigned long skip;                                                 \
 831     int ok = 1;                                                         \
 832                                                                         \
 833     if (v[4].len) { p0 = (const octet *)v[4].buf; sz0 = v[4].len; }     \
 834     else { p0 = 0; sz0 = v[5].len; }                                    \
 835     DENSURE(&d, sz0); d.len = sz0;                                      \
 836     skip = *(unsigned long *)v[3].buf;                                  \
 837                                                                         \
 838     step = 0;                                                           \
 839     while (step < sz0 + skip) {                                         \
 840       step = step ? 3*step + 4 : 1;                                     \
 841       if (step > sz0 + skip) step = sz0 + skip;                         \
 842       BASE##_INIT(r, &ctx, v[0].buf, v[0].len, v[1].buf);               \
 843       if (v[2].len) {                                                   \
 844         LOAD64_(pos, v[2].buf);                                         \
 845         BASE##_SEEKU64(r, &ctx, pos);                                   \
 846       }                                                                 \
 847                                                                         \
 848       for (sz = skip; sz >= step; sz -= step)                           \
 849         BASE##_ENCRYPT(r, &ctx, 0, 0, step);                            \
 850       if (sz) BASE##_ENCRYPT(r, &ctx, 0, 0, sz);                        \
 851       for (p = p0, q = (octet *)d.buf, sz = sz0;                        \
 852            sz >= step;                                                  \
 853            sz -= step, q += step) {                                     \
 854         BASE##_ENCRYPT(r, &ctx, p, q, step);                            \
 855         if (p) p += step;                                               \
 856       }                                                                 \
 857       if (sz) BASE##_ENCRYPT(r, &ctx, p, q, sz);                        \
 858                                                                         \
 859       if (d.len != v[5].len || memcmp(d.buf, v[5].buf, v[5].len) != 0) { \
 860         ok = 0;                                                         \
 861         printf("\nfail encrypt:"                                        \
 862                "\n\tstep           = %lu"                               \
 863                "\n\tkey    = ", (unsigned long)step);                   \
 864         type_hex.dump(&v[0], stdout);                                   \
 865         printf("\n\tnonce          = ");                                \
 866         type_hex.dump(&v[1], stdout);                                   \
 867         printf("\n\tposition   = ");                                    \
 868         type_hex.dump(&v[2], stdout);                                   \
 869         printf("\n\tskip           = %lu", skip);                       \
 870         printf("\n\tmessage    = ");                                    \
 871         type_hex.dump(&v[4], stdout);                                   \
 872         printf("\n\texpected   = ");                                    \
 873         type_hex.dump(&v[5], stdout);                                   \
 874         printf("\n\tcalculated = ");                                    \
 875         type_hex.dump(&d, stdout);                                      \
 876         putchar('\n');                                                  \
 877       }                                                                 \
 878     }                                                                   \
 879                                                                         \
 880     dstr_destroy(&d);                                                   \
 881     return (ok);                                                        \
 882   }
 883 #define DEFVENC(r) DEFxVENC(salsa20, SALSA20, r)
 884 #define DEFXVENC(r) DEFxVENC(xsalsa20, XSALSA20, r)
 885 SALSA20_VARS(DEFVENC)
 886 SALSA20_VARS(DEFXVENC)
 887
 888 static test_chunk defs[] = {
 889 #define DEFxTAB(pre, base, r)                                           \
 890   { pre SALSA20_NAME_##r, v_encrypt_##base##_##r,                       \
 891     { &type_hex, &type_hex, &type_hex, &type_ulong,                     \
 892       &type_hex, &type_hex, 0 } },
 893 #define DEFTAB(r)                                                       \
 894   { SALSA20_NAME_##r "-core", v_core_##r,                               \
 895     { &type_int, &type_hex, &type_hex, 0 } },                           \
 896   DEFxTAB("", salsa20, r)
 897 #define DEFXTAB(r) DEFxTAB("x", xsalsa20, r)
 898 SALSA20_VARS(DEFTAB)
 899 SALSA20_VARS(DEFXTAB)
 900   { 0, 0, { 0 } }
 901 };
 902
 903 int main(int argc, char *argv[])
 904 {
 905   test_run(argc, argv, defs, SRCDIR"/t/salsa20");
 906   return (0);
 907 }
 908
 909 #endif
 910
 911 /*----- That's all, folks -------------------------------------------------*/