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

   1 /* -*-c-*-
   2  *
   3  * Salsa20 stream cipher
   4  *
   5  * (c) 2015 Straylight/Edgeware
   6  */
   7
   8 /*----- Licensing notice --------------------------------------------------*
   9  *
  10  * This file is part of Catacomb.
  11  *
  12  * Catacomb is free software; you can redistribute it and/or modify
  13  * it under the terms of the GNU Library General Public License as
  14  * published by the Free Software Foundation; either version 2 of the
  15  * License, or (at your option) any later version.
  16  *
  17  * Catacomb is distributed in the hope that it will be useful,
  18  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  20  * GNU Library General Public License for more details.
  21  *
  22  * You should have received a copy of the GNU Library General Public
  23  * License along with Catacomb; if not, write to the Free
  24  * Software Foundation, Inc., 59 Temple Place - Suite 330, Boston,
  25  * MA 02111-1307, USA.
  26  */
  27
  28 /*----- Header files ------------------------------------------------------*/
  29
  30 #include "config.h"
  31
  32 #include <stdarg.h>
  33
  34 #include <mLib/bits.h>
  35
  36 #include "arena.h"
  37 #include "dispatch.h"
  38 #include "gcipher.h"
  39 #include "grand.h"
  40 #include "keysz.h"
  41 #include "paranoia.h"
  42 #include "salsa20.h"
  43 #include "salsa20-core.h"
  44
  45 /*----- Global variables --------------------------------------------------*/
  46
  47 const octet salsa20_keysz[] = { KSZ_SET, 32, 16, 10, 0 };
  48
  49 /*----- The Salsa20 core function and utilities ---------------------------*/
  50
  51 /* --- @core@ --- *
  52  *
  53  * Arguments:   @unsigned r@ = number of rounds
  54  *              @const salsa20_matrix src@ = input matrix
  55  *              @salsa20_matrix dest@ = where to put the output
  56  *
  57  * Returns:     ---
  58  *
  59  *
  60  * Use:         Apply the Salsa20/r core function to @src@, writing the
  61  *              result to @dest@.  This consists of @r@ rounds followed by
  62  *              the feedforward step.
  63  */
  64
  65 CPU_DISPATCH(static, (void), void, core,
  66              (unsigned r, const salsa20_matrix src, salsa20_matrix dest),
  67              (r, src, dest), pick_core, simple_core);
  68
  69 static void simple_core(unsigned r, const salsa20_matrix src,
  70                         salsa20_matrix dest)
  71   { SALSA20_nR(dest, src, r); SALSA20_FFWD(dest, src); }
  72
  73 #if CPUFAM_X86 || CPUFAM_AMD64
  74 extern core__functype salsa20_core_x86ish_sse2;
  75 #endif
  76
  77 #if CPUFAM_ARMEL
  78 extern core__functype salsa20_core_arm_neon;
  79 #endif
  80
  81 static core__functype *pick_core(void)
  82 {
  83 #if CPUFAM_X86 || CPUFAM_AMD64
  84   DISPATCH_PICK_COND(salsa20_core, salsa20_core_x86ish_sse2,
  85                      cpu_feature_p(CPUFEAT_X86_SSE2));
  86 #endif
  87 #if CPUFAM_ARMEL
  88   DISPATCH_PICK_COND(salsa20_core, salsa20_core_arm_neon,
  89                      cpu_feature_p(CPUFEAT_ARM_NEON));
  90 #endif
  91   DISPATCH_PICK_FALLBACK(salsa20_core, simple_core);
  92 }
  93
  94 /* --- @populate@ --- *
  95  *
  96  * Arguments:   @salsa20_matrix a@ = a matrix to fill in
  97  *              @const void *key@ = pointer to key material
  98  *              @size_t ksz@ = size of key
  99  *
 100  * Returns:     ---
 101  *
 102  * Use:         Fills in a Salsa20 matrix from the key, setting the
 103  *              appropriate constants according to the key length.  The nonce
 104  *              and position words are left uninitialized.
 105  */
 106
 107 static void populate(salsa20_matrix a, const void *key, size_t ksz)
 108 {
 109   const octet *k = key;
 110
 111   KSZ_ASSERT(salsa20, ksz);
 112
 113   /* Here's the pattern of key, constant, nonce, and counter pieces in the
 114    * matrix, before and after our permutation.
 115    *
 116    * [ C0 K0 K1 K2 ]       [ C0 C1 C2 C3 ]
 117    * [ K3 C1 N0 N1 ]  -->  [ K3 T1 K7 K2 ]
 118    * [ T0 T1 C2 K4 ]       [ T0 K6 K1 N1 ]
 119    * [ K5 K6 K7 C3 ]       [ K5 K0 N0 K4 ]
 120    */
 121
 122   a[13] = LOAD32_L(k +  0);
 123   a[10] = LOAD32_L(k +  4);
 124   if (ksz == 10) {
 125     a[ 7] = LOAD16_L(k +  8);
 126     a[ 4] = 0;
 127   } else {
 128     a[ 7] = LOAD32_L(k +  8);
 129     a[ 4] = LOAD32_L(k + 12);
 130   }
 131   if (ksz <= 16) {
 132     a[15] = a[13];
 133     a[12] = a[10];
 134     a[ 9] = a[ 7];
 135     a[ 6] = a[ 4];
 136     a[ 0] = SALSA20_A128;
 137     a[ 1] = SALSA20_B128;
 138     a[ 2] = ksz == 10 ? SALSA20_C80 : SALSA20_C128;
 139     a[ 3] = SALSA20_D128;
 140   } else {
 141     a[15] = LOAD32_L(k + 16);
 142     a[12] = LOAD32_L(k + 20);
 143     a[ 9] = LOAD32_L(k + 24);
 144     a[ 6] = LOAD32_L(k + 28);
 145     a[ 0] = SALSA20_A256;
 146     a[ 1] = SALSA20_B256;
 147     a[ 2] = SALSA20_C256;
 148     a[ 3] = SALSA20_D256;
 149   }
 150 }
 151
 152 /*----- Salsa20 implementation --------------------------------------------*/
 153
 154 /* --- @salsa20_init@ --- *
 155  *
 156  * Arguments:   @salsa20_ctx *ctx@ = context to fill in
 157  *              @const void *key@ = pointer to key material
 158  *              @size_t ksz@ = size of key (either 32 or 16)
 159  *              @const void *nonce@ = initial nonce, or null
 160  *
 161  * Returns:     ---
 162  *
 163  * Use:         Initializes a Salsa20 context ready for use.
 164  */
 165
 166 void salsa20_init(salsa20_ctx *ctx, const void *key, size_t ksz,
 167                   const void *nonce)
 168 {
 169   static const octet zerononce[SALSA20_NONCESZ];
 170
 171   populate(ctx->a, key, ksz);
 172   salsa20_setnonce(ctx, nonce ? nonce : zerononce);
 173 }
 174
 175 /* --- @salsa20_setnonce@ --- *
 176  *
 177  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 178  *              @const void *nonce@ = the nonce (@SALSA20_NONCESZ@ bytes)
 179  *
 180  * Returns:     ---
 181  *
 182  * Use:         Set a new nonce in the context @ctx@, e.g., for processing a
 183  *              different message.  The stream position is reset to zero (see
 184  *              @salsa20_seek@ etc.).
 185  */
 186
 187 void salsa20_setnonce(salsa20_ctx *ctx, const void *nonce)
 188 {
 189   const octet *n = nonce;
 190
 191   ctx->a[14] = LOAD32_L(n + 0);
 192   ctx->a[11] = LOAD32_L(n + 4);
 193   salsa20_seek(ctx, 0);
 194 }
 195
 196 /* --- @salsa20_seek{,u64}@ --- *
 197  *
 198  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 199  *              @unsigned long i@, @kludge64 i@ = new position to set
 200  *
 201  * Returns:     ---
 202  *
 203  * Use:         Sets a new stream position, in units of Salsa20 output
 204  *              blocks, which are @SALSA20_OUTSZ@ bytes each.  Byte
 205  *              granularity can be achieved by calling @salsa20R_encrypt@
 206  *              appropriately.
 207  */
 208
 209 void salsa20_seek(salsa20_ctx *ctx, unsigned long i)
 210   { kludge64 ii; ASSIGN64(ii, i); salsa20_seeku64(ctx, ii); }
 211
 212 void salsa20_seeku64(salsa20_ctx *ctx, kludge64 i)
 213 {
 214   ctx->a[8] = LO64(i); ctx->a[5] = HI64(i);
 215   ctx->bufi = SALSA20_OUTSZ;
 216 }
 217
 218 /* --- @salsa20_tell{,u64}@ --- *
 219  *
 220  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 221  *
 222  * Returns:     The current position in the output stream, in blocks,
 223  *              rounding upwards.
 224  */
 225
 226 unsigned long salsa20_tell(salsa20_ctx *ctx)
 227   { kludge64 i = salsa20_tellu64(ctx); return (GET64(unsigned long, i)); }
 228
 229 kludge64 salsa20_tellu64(salsa20_ctx *ctx)
 230   { kludge64 i; SET64(i, ctx->a[5], ctx->a[8]); return (i); }
 231
 232 /* --- @salsa20{,12,8}_encrypt@ --- *
 233  *
 234  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 235  *              @const void *src@ = source buffer (or null)
 236  *              @void *dest@ = destination buffer (or null)
 237  *              @size_t sz@ = size of the buffers
 238  *
 239  * Returns:     ---
 240  *
 241  * Use:         Encrypts or decrypts @sz@ bytes of data from @src@ to @dest@.
 242  *              Salsa20 works by XORing plaintext with a keystream, so
 243  *              encryption and decryption are the same operation.  If @dest@
 244  *              is null then ignore @src@ and skip @sz@ bytes of the
 245  *              keystream.  If @src@ is null, then just write the keystream
 246  *              to @dest@.
 247  */
 248
 249 #define SALSA20_ENCRYPT(r, ctx, src, dest, sz)                          \
 250   SALSA20_DECOR(salsa20, r, _encrypt)(ctx, src, dest, sz)
 251 #define DEFENCRYPT(r)                                                   \
 252   void SALSA20_ENCRYPT(r, salsa20_ctx *ctx, const void *src,            \
 253                        void *dest, size_t sz)                           \
 254   {                                                                     \
 255     salsa20_matrix b;                                                   \
 256     const octet *s = src;                                               \
 257     octet *d = dest;                                                    \
 258     size_t n;                                                           \
 259     kludge64 pos, delta;                                                \
 260                                                                         \
 261     SALSA20_OUTBUF(ctx, d, s, sz);                                      \
 262     if (!sz) return;                                                    \
 263                                                                         \
 264     if (!dest) {                                                        \
 265       n = sz/SALSA20_OUTSZ;                                             \
 266       pos = salsa20_tellu64(ctx);                                       \
 267       ASSIGN64(delta, n);                                               \
 268       ADD64(pos, pos, delta);                                           \
 269       salsa20_seeku64(ctx, pos);                                        \
 270       sz = sz%SALSA20_OUTSZ;                                            \
 271     } else if (!src) {                                                  \
 272       while (sz >= SALSA20_OUTSZ) {                                     \
 273         core(r, ctx->a, b);                                             \
 274         SALSA20_STEP(ctx->a);                                           \
 275         SALSA20_GENFULL(b, d);                                          \
 276         sz -= SALSA20_OUTSZ;                                            \
 277       }                                                                 \
 278     } else {                                                            \
 279       while (sz >= SALSA20_OUTSZ) {                                     \
 280         core(r, ctx->a, b);                                             \
 281         SALSA20_STEP(ctx->a);                                           \
 282         SALSA20_MIXFULL(b, d, s);                                       \
 283         sz -= SALSA20_OUTSZ;                                            \
 284       }                                                                 \
 285     }                                                                   \
 286                                                                         \
 287     if (sz) {                                                           \
 288       core(r, ctx->a, b);                                               \
 289       SALSA20_STEP(ctx->a);                                             \
 290       SALSA20_PREPBUF(ctx, b);                                          \
 291       SALSA20_OUTBUF(ctx, d, s, sz);                                    \
 292       assert(!sz);                                                      \
 293     }                                                                   \
 294   }
 295 SALSA20_VARS(DEFENCRYPT)
 296
 297 /*----- HSalsa20 implementation -------------------------------------------*/
 298
 299 #define HSALSA20_RAW(r, ctx, src, dest)                                 \
 300   SALSA20_DECOR(hsalsa20, r, _raw)(ctx, src, dest)
 301 #define HSALSA20_PRF(r, ctx, src, dest)                                 \
 302   SALSA20_DECOR(hsalsa20, r, _prf)(ctx, src, dest)
 303
 304 /* --- @hsalsa20{,12,8}_prf@ --- *
 305  *
 306  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 307  *              @const void *src@ = the input (@HSALSA20_INSZ@ bytes)
 308  *              @void *dest@ = the output (@HSALSA20_OUTSZ@ bytes)
 309  *
 310  * Returns:     ---
 311  *
 312  * Use:         Apply the HSalsa20/r pseudorandom function to @src@, writing
 313  *              the result to @out@.
 314  */
 315
 316 #define DEFHSALSA20(r)                                                  \
 317   static void HSALSA20_RAW(r, salsa20_matrix k,                         \
 318                            const uint32 *src, uint32 *dest)             \
 319   {                                                                     \
 320     salsa20_matrix a;                                                   \
 321     int i;                                                              \
 322                                                                         \
 323     /* --- HSalsa20, computed from full Salsa20 --- *                   \
 324      *                                                                  \
 325      * The security proof makes use of the fact that HSalsa20 (i.e.,    \
 326      * without the final feedforward step) can be computed from full    \
 327      * Salsa20 using only knowledge of the non-secret input.  I don't   \
 328      * want to compromise the performance of the main function by       \
 329      * making the feedforward step separate, but this operation is less \
 330      * speed critical, so we do it the harder way.                      \
 331      */                                                                 \
 332                                                                         \
 333     for (i = 0; i < 4; i++) k[14 - 3*i] = src[i];                       \
 334     core(r, k, a);                                                      \
 335     for (i = 0; i < 4; i++) dest[i] = a[5*i] - k[i];                    \
 336     for (i = 4; i < 8; i++) dest[i] = a[i + 2] - k[26 - 3*i];           \
 337   }                                                                     \
 338                                                                         \
 339   void HSALSA20_PRF(r, salsa20_ctx *ctx, const void *src, void *dest)   \
 340   {                                                                     \
 341     const octet *s = src;                                               \
 342     octet *d = dest;                                                    \
 343     uint32 in[4], out[8];                                               \
 344     int i;                                                              \
 345                                                                         \
 346     for (i = 0; i < 4; i++) in[i] = LOAD32_L(s + 4*i);                  \
 347     HSALSA20_RAW(r, ctx->a, in, out);                                   \
 348     for (i = 0; i < 8; i++) STORE32_L(d + 4*i, out[i]);                 \
 349   }
 350 SALSA20_VARS(DEFHSALSA20)
 351
 352 /*----- XSalsa20 implementation -------------------------------------------*/
 353
 354 /* --- Some convenient macros for naming functions --- *
 355  *
 356  * Because the crypto core is involved in XSalsa20/r's per-nonce setup, we
 357  * need to take an interest in the number of rounds in most of the various
 358  * functions, and it will probably help if we distinguish the context
 359  * structures for the various versions.
 360  */
 361
 362 #define XSALSA20_CTX(r) SALSA20_DECOR(xsalsa20, r, _ctx)
 363 #define XSALSA20_INIT(r, ctx, k, ksz, n)                                \
 364   SALSA20_DECOR(xsalsa20, r, _init)(ctx, k, ksz, n)
 365 #define XSALSA20_SETNONCE(r, ctx, n)                                    \
 366   SALSA20_DECOR(xsalsa20, r, _setnonce)(ctx, n)
 367 #define XSALSA20_SEEK(r, ctx, i)                                        \
 368   SALSA20_DECOR(xsalsa20, r, _seek)(ctx, i)
 369 #define XSALSA20_SEEKU64(r, ctx, i)                                     \
 370   SALSA20_DECOR(xsalsa20, r, _seeku64)(ctx, i)
 371 #define XSALSA20_TELL(r, ctx)                                           \
 372   SALSA20_DECOR(xsalsa20, r, _tell)(ctx)
 373 #define XSALSA20_TELLU64(r, ctx)                                        \
 374   SALSA20_DECOR(xsalsa20, r, _tellu64)(ctx)
 375 #define XSALSA20_ENCRYPT(r, ctx, src, dest, sz)                         \
 376   SALSA20_DECOR(xsalsa20, r, _encrypt)(ctx, src, dest, sz)
 377
 378 /* --- @xsalsa20{,12,8}_init@ --- *
 379  *
 380  * Arguments:   @xsalsa20R_ctx *ctx@ = the context to fill in
 381  *              @const void *key@ = pointer to key material
 382  *              @size_t ksz@ = size of key (either 32 or 16)
 383  *              @const void *nonce@ = initial nonce, or null
 384  *
 385  * Returns:     ---
 386  *
 387  * Use:         Initializes an XSalsa20/r context ready for use.
 388  *
 389  *              There is a different function for each number of rounds,
 390  *              unlike for plain Salsa20.
 391  */
 392
 393 #define DEFXINIT(r)                                                     \
 394   void XSALSA20_INIT(r, XSALSA20_CTX(r) *ctx,                           \
 395                         const void *key, size_t ksz, const void *nonce) \
 396   {                                                                     \
 397     static const octet zerononce[XSALSA20_NONCESZ];                     \
 398                                                                         \
 399     populate(ctx->k, key, ksz);                                         \
 400     ctx->s.a[ 0] = SALSA20_A256;                                        \
 401     ctx->s.a[ 1] = SALSA20_B256;                                        \
 402     ctx->s.a[ 2] = SALSA20_C256;                                        \
 403     ctx->s.a[ 3] = SALSA20_D256;                                        \
 404     XSALSA20_SETNONCE(r, ctx, nonce ? nonce : zerononce);               \
 405   }
 406 SALSA20_VARS(DEFXINIT)
 407
 408 /* --- @xsalsa20{,12,8}_setnonce@ --- *
 409  *
 410  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 411  *              @const void *nonce@ = the nonce (@XSALSA20_NONCESZ@ bytes)
 412  *
 413  * Returns:     ---
 414  *
 415  * Use:         Set a new nonce in the context @ctx@, e.g., for processing a
 416  *              different message.  The stream position is reset to zero (see
 417  *              @salsa20_seek@ etc.).
 418  *
 419  *              There is a different function for each number of rounds,
 420  *              unlike for plain Salsa20.
 421  */
 422
 423 #define DEFXNONCE(r)                                                    \
 424   void XSALSA20_SETNONCE(r, XSALSA20_CTX(r) *ctx, const void *nonce)    \
 425   {                                                                     \
 426     const octet *n = nonce;                                             \
 427     uint32 in[4], out[8];                                               \
 428     int i;                                                              \
 429                                                                         \
 430     for (i = 0; i < 4; i++) in[i] = LOAD32_L(n + 4*i);                  \
 431     HSALSA20_RAW(r, ctx->k, in, out);                                   \
 432     for (i = 0; i < 4; i++) ctx->s.a[13 - 3*i] = out[i];                \
 433     for (i = 4; i < 8; i++) ctx->s.a[27 - 3*i] = out[i];                \
 434     salsa20_setnonce(&ctx->s, n + 16);                                  \
 435   }
 436 SALSA20_VARS(DEFXNONCE)
 437
 438 /* --- @xsalsa20{,12,8}_seek{,u64}@ --- *
 439  *
 440  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 441  *              @unsigned long i@, @kludge64 i@ = new position to set
 442  *
 443  * Returns:     ---
 444  *
 445  * Use:         Sets a new stream position, in units of Salsa20 output
 446  *              blocks, which are @XSALSA20_OUTSZ@ bytes each.  Byte
 447  *              granularity can be achieved by calling @xsalsa20R_encrypt@
 448  *              appropriately.
 449  *
 450  *              There is a different function for each number of rounds,
 451  *              unlike for plain Salsa20, because the context structures are
 452  *              different.
 453  */
 454
 455 /* --- @xsalsa20{,12,8}_tell{,u64}@ --- *
 456  *
 457  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 458  *
 459  * Returns:     The current position in the output stream, in blocks,
 460  *              rounding upwards.
 461  *
 462  *              There is a different function for each number of rounds,
 463  *              unlike for plain Salsa20, because the context structures are
 464  *              different.
 465  */
 466
 467 /* --- @xsalsa20{,12,8}_encrypt@ --- *
 468  *
 469  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 470  *              @const void *src@ = source buffer (or null)
 471  *              @void *dest@ = destination buffer (or null)
 472  *              @size_t sz@ = size of the buffers
 473  *
 474  * Returns:     ---
 475  *
 476  * Use:         Encrypts or decrypts @sz@ bytes of data from @src@ to @dest@.
 477  *              XSalsa20 works by XORing plaintext with a keystream, so
 478  *              encryption and decryption are the same operation.  If @dest@
 479  *              is null then ignore @src@ and skip @sz@ bytes of the
 480  *              keystream.  If @src@ is null, then just write the keystream
 481  *              to @dest@.
 482  */
 483
 484 #define DEFXPASSTHRU(r)                                                 \
 485   void XSALSA20_SEEK(r, XSALSA20_CTX(r) *ctx, unsigned long i)          \
 486     { salsa20_seek(&ctx->s, i); }                                       \
 487   void XSALSA20_SEEKU64(r, XSALSA20_CTX(r) *ctx, kludge64 i)            \
 488     { salsa20_seeku64(&ctx->s, i); }                                    \
 489   unsigned long XSALSA20_TELL(r, XSALSA20_CTX(r) *ctx)                  \
 490     { return salsa20_tell(&ctx->s); }                                   \
 491   kludge64 XSALSA20_TELLU64(r, XSALSA20_CTX(r) *ctx)                    \
 492     { return salsa20_tellu64(&ctx->s); }                                \
 493   void XSALSA20_ENCRYPT(r, XSALSA20_CTX(r) *ctx,                        \
 494                         const void *src, void *dest, size_t sz)         \
 495     { SALSA20_ENCRYPT(r, &ctx->s, src, dest, sz); }
 496 SALSA20_VARS(DEFXPASSTHRU)
 497
 498 /*----- Generic cipher interface ------------------------------------------*/
 499
 500 typedef struct gctx { gcipher c; salsa20_ctx ctx; } gctx;
 501
 502 static void gsetiv(gcipher *c, const void *iv)
 503   { gctx *g = (gctx *)c; salsa20_setnonce(&g->ctx, iv); }
 504
 505 static void gdestroy(gcipher *c)
 506   { gctx *g = (gctx *)c; BURN(*g); S_DESTROY(g); }
 507
 508 static gcipher *ginit(const void *k, size_t sz, const gcipher_ops *ops)
 509 {
 510   gctx *g = S_CREATE(gctx);
 511   g->c.ops = ops;
 512   salsa20_init(&g->ctx, k, sz, 0);
 513   return (&g->c);
 514 }
 515
 516 #define DEFGCIPHER(r)                                                   \
 517                                                                         \
 518   static const gcipher_ops gops_##r;                                    \
 519                                                                         \
 520   static gcipher *ginit_##r(const void *k, size_t sz)                   \
 521     { return (ginit(k, sz, &gops_##r)); }                               \
 522                                                                         \
 523   static void gencrypt_##r(gcipher *c, const void *s,                   \
 524                            void *t, size_t sz)                          \
 525     { gctx *g = (gctx *)c; SALSA20_ENCRYPT(r, &g->ctx, s, t, sz); }     \
 526                                                                         \
 527   static const gcipher_ops gops_##r = {                                 \
 528     &SALSA20_DECOR(salsa20, r, ),                                       \
 529     gencrypt_##r, gencrypt_##r, gdestroy, gsetiv, 0                     \
 530   };                                                                    \
 531                                                                         \
 532   const gccipher SALSA20_DECOR(salsa20, r, ) = {                        \
 533     SALSA20_NAME_##r, salsa20_keysz,                                    \
 534     SALSA20_NONCESZ, ginit_##r                                          \
 535   };
 536
 537 SALSA20_VARS(DEFGCIPHER)
 538
 539 #define DEFGXCIPHER(r)                                                  \
 540                                                                         \
 541   typedef struct { gcipher c; XSALSA20_CTX(r) ctx; } gxctx_##r;         \
 542                                                                         \
 543   static void gxsetiv_##r(gcipher *c, const void *iv)                   \
 544     { gxctx_##r *g = (gxctx_##r *)c; XSALSA20_SETNONCE(r, &g->ctx, iv); } \
 545                                                                         \
 546   static void gxdestroy_##r(gcipher *c)                                 \
 547     { gxctx_##r *g = (gxctx_##r *)c; BURN(*g); S_DESTROY(g); }          \
 548                                                                         \
 549   static const gcipher_ops gxops_##r;                                   \
 550                                                                         \
 551   static gcipher *gxinit_##r(const void *k, size_t sz)                  \
 552   {                                                                     \
 553     gxctx_##r *g = S_CREATE(gxctx_##r);                                 \
 554     g->c.ops = &gxops_##r;                                              \
 555     XSALSA20_INIT(r, &g->ctx, k, sz, 0);                                \
 556     return (&g->c);                                                     \
 557   }                                                                     \
 558                                                                         \
 559   static void gxencrypt_##r(gcipher *c, const void *s,                  \
 560                             void *t, size_t sz)                         \
 561   {                                                                     \
 562     gxctx_##r *g = (gxctx_##r *)c;                                      \
 563     XSALSA20_ENCRYPT(r, &g->ctx, s, t, sz);                             \
 564   }                                                                     \
 565                                                                         \
 566   static const gcipher_ops gxops_##r = {                                \
 567     &SALSA20_DECOR(xsalsa20, r, ),                                      \
 568     gxencrypt_##r, gxencrypt_##r, gxdestroy_##r, gxsetiv_##r, 0         \
 569   };                                                                    \
 570                                                                         \
 571   const gccipher SALSA20_DECOR(xsalsa20, r, ) = {                       \
 572     "x" SALSA20_NAME_##r, salsa20_keysz,                                \
 573     XSALSA20_NONCESZ, gxinit_##r                                        \
 574   };
 575
 576 SALSA20_VARS(DEFGXCIPHER)
 577
 578 /*----- Generic random number generator interface -------------------------*/
 579
 580 typedef struct grops {
 581   size_t noncesz;
 582   void (*seek)(void *, kludge64);
 583   kludge64 (*tell)(void *);
 584   void (*setnonce)(void *, const void *);
 585   void (*generate)(void *, void *, size_t);
 586 } grops;
 587
 588 typedef struct grbasectx {
 589   grand r;
 590   const grops *ops;
 591 } grbasectx;
 592
 593 static int grmisc(grand *r, unsigned op, ...)
 594 {
 595   octet buf[XSALSA20_NONCESZ];
 596   grbasectx *g = (grbasectx *)r;
 597   grand *rr;
 598   const octet *p;
 599   size_t sz;
 600   uint32 i;
 601   unsigned long ul;
 602   kludge64 pos;
 603   va_list ap;
 604   int rc = 0;
 605
 606   va_start(ap, op);
 607
 608   switch (op) {
 609     case GRAND_CHECK:
 610       switch (va_arg(ap, unsigned)) {
 611         case GRAND_CHECK:
 612         case GRAND_SEEDINT:
 613         case GRAND_SEEDUINT32:
 614         case GRAND_SEEDBLOCK:
 615         case GRAND_SEEDRAND:
 616         case SALSA20_SEEK:
 617         case SALSA20_SEEKU64:
 618         case SALSA20_TELL:
 619         case SALSA20_TELLU64:
 620           rc = 1;
 621           break;
 622         default:
 623           rc = 0;
 624           break;
 625       }
 626       break;
 627
 628     case GRAND_SEEDINT:
 629       i = va_arg(ap, unsigned); STORE32_L(buf, i);
 630       memset(buf + 4, 0, g->ops->noncesz - 4);
 631       g->ops->setnonce(g, buf);
 632       break;
 633     case GRAND_SEEDUINT32:
 634       i = va_arg(ap, uint32); STORE32_L(buf, i);
 635       memset(buf + 4, 0, g->ops->noncesz - 4);
 636       g->ops->setnonce(g, buf);
 637       break;
 638     case GRAND_SEEDBLOCK:
 639       p = va_arg(ap, const void *);
 640       sz = va_arg(ap, size_t);
 641       if (sz < g->ops->noncesz) {
 642         memcpy(buf, p, sz);
 643         memset(buf + sz, 0, g->ops->noncesz - sz);
 644         p = buf;
 645       }
 646       g->ops->setnonce(g, p);
 647       break;
 648     case GRAND_SEEDRAND:
 649       rr = va_arg(ap, grand *);
 650       rr->ops->fill(rr, buf, g->ops->noncesz);
 651       g->ops->setnonce(g, buf);
 652       break;
 653     case SALSA20_SEEK:
 654       ul = va_arg(ap, unsigned long); ASSIGN64(pos, ul);
 655       g->ops->seek(g, pos);
 656       break;
 657     case SALSA20_SEEKU64:
 658       pos = va_arg(ap, kludge64);
 659       g->ops->seek(g, pos);
 660       break;
 661     case SALSA20_TELL:
 662       pos = g->ops->tell(g);
 663       *va_arg(ap, unsigned long *) = GET64(unsigned long, pos);
 664       break;
 665     case SALSA20_TELLU64:
 666       *va_arg(ap, kludge64 *) = g->ops->tell(g);
 667       break;
 668     default:
 669       GRAND_BADOP;
 670       break;
 671   }
 672
 673   return (rc);
 674 }
 675
 676 static octet grbyte(grand *r)
 677 {
 678   grbasectx *g = (grbasectx *)r;
 679   octet o;
 680   g->ops->generate(g, &o, 1);
 681   return (o);
 682 }
 683
 684 static uint32 grword(grand *r)
 685 {
 686   grbasectx *g = (grbasectx *)r;
 687   octet b[4];
 688   g->ops->generate(g, b, sizeof(b));
 689   return (LOAD32_L(b));
 690 }
 691
 692 static void grfill(grand *r, void *p, size_t sz)
 693 {
 694   grbasectx *g = (grbasectx *)r;
 695   g->ops->generate(r, p, sz);
 696 }
 697
 698 typedef struct grctx {
 699   grbasectx r;
 700   salsa20_ctx ctx;
 701 } grctx;
 702
 703 static void gr_seek(void *r, kludge64 pos)
 704   { grctx *g = r; salsa20_seeku64(&g->ctx, pos); }
 705
 706 static kludge64 gr_tell(void *r)
 707   { grctx *g = r; return (salsa20_tellu64(&g->ctx)); }
 708
 709 static void gr_setnonce(void *r, const void *n)
 710   { grctx *g = r; salsa20_setnonce(&g->ctx, n); }
 711
 712 static void grdestroy(grand *r)
 713   { grctx *g = (grctx *)r; BURN(*g); S_DESTROY(g); }
 714
 715 static grand *grinit(const void *k, size_t ksz, const void *n,
 716                      const grand_ops *ops, const grops *myops)
 717 {
 718     grctx *g = S_CREATE(grctx);
 719     g->r.r.ops = ops;
 720     g->r.ops = myops;
 721     salsa20_init(&g->ctx, k, ksz, 0);
 722     myops->setnonce(g, n);
 723     return (&g->r.r);
 724 }
 725
 726 #define DEFGRAND(rr)                                                    \
 727                                                                         \
 728   static void gr_generate_##rr(void *r, void *b, size_t sz)             \
 729     { grctx *g = r; SALSA20_ENCRYPT(rr, &g->ctx, 0, b, sz); }           \
 730                                                                         \
 731   static const grops grops_##rr =                                       \
 732     { SALSA20_NONCESZ, gr_seek, gr_tell,                                \
 733       gr_setnonce, gr_generate_##rr };                                  \
 734                                                                         \
 735   static const grand_ops grops_rand_##rr = {                            \
 736     SALSA20_NAME_##rr, GRAND_CRYPTO, 0,                                 \
 737     grmisc, grdestroy, grword,                                          \
 738     grbyte, grword, grand_defaultrange, grfill                          \
 739   };                                                                    \
 740                                                                         \
 741   grand *SALSA20_DECOR(salsa20, rr, _rand)                              \
 742     (const void *k, size_t ksz, const void *n)                          \
 743     { return (grinit(k, ksz, n, &grops_rand_##rr, &grops_##rr)); }
 744 SALSA20_VARS(DEFGRAND)
 745
 746 #define DEFXGRAND(rr)                                                   \
 747                                                                         \
 748   typedef struct grxctx_##rr {                                          \
 749     grbasectx r;                                                        \
 750     XSALSA20_CTX(rr) ctx;                                               \
 751   } grxctx_##rr;                                                        \
 752                                                                         \
 753   static void grx_seek_##rr(void *r, kludge64 pos)                      \
 754     { grxctx_##rr *g = r; XSALSA20_SEEKU64(rr, &g->ctx, pos); }         \
 755                                                                         \
 756   static kludge64 grx_tell_##rr(void *r)                                \
 757     { grxctx_##rr *g = r; return (XSALSA20_TELLU64(rr, &g->ctx)); }     \
 758                                                                         \
 759   static void grx_setnonce_##rr(void *r, const void *n)                 \
 760     { grxctx_##rr *g = r; XSALSA20_SETNONCE(rr, &g->ctx, n); }          \
 761                                                                         \
 762   static void grxdestroy_##rr(grand *r)                                 \
 763     { grxctx_##rr *g = (grxctx_##rr *)r; BURN(*g); S_DESTROY(g); }      \
 764                                                                         \
 765   static void grx_generate_##rr(void *r, void *b, size_t sz)            \
 766     { grxctx_##rr *g = r; XSALSA20_ENCRYPT(rr, &g->ctx, 0, b, sz); }    \
 767                                                                         \
 768   static const grops grxops_##rr =                                      \
 769   { XSALSA20_NONCESZ, grx_seek_##rr, grx_tell_##rr,                     \
 770       grx_setnonce_##rr, grx_generate_##rr };                           \
 771                                                                         \
 772   static const grand_ops grxops_rand_##rr = {                           \
 773     "x" SALSA20_NAME_##rr, GRAND_CRYPTO, 0,                             \
 774     grmisc, grxdestroy_##rr, grword,                                    \
 775     grbyte, grword, grand_defaultrange, grfill                          \
 776   };                                                                    \
 777                                                                         \
 778   grand *SALSA20_DECOR(xsalsa20, rr, _rand)                             \
 779     (const void *k, size_t ksz, const void *n)                          \
 780   {                                                                     \
 781     grxctx_##rr *g = S_CREATE(grxctx_##rr);                             \
 782     g->r.r.ops = &grxops_rand_##rr;                                     \
 783     g->r.ops = &grxops_##rr;                                            \
 784     XSALSA20_INIT(rr, &g->ctx, k, ksz, n);                              \
 785     return (&g->r.r);                                                   \
 786   }
 787 SALSA20_VARS(DEFXGRAND)
 788
 789 /*----- Test rig ----------------------------------------------------------*/
 790
 791 #ifdef TEST_RIG
 792
 793 #include <stdio.h>
 794 #include <string.h>
 795
 796 #include <mLib/quis.h>
 797 #include <mLib/testrig.h>
 798
 799 static const int perm[] = {
 800    0, 13, 10,  7,
 801    4,  1, 14, 11,
 802    8,  5,  2, 15,
 803   12,  9,  6,  3
 804 };
 805
 806 #define DEFVCORE(r)                                                     \
 807   static int v_core_##r(dstr *v)                                        \
 808   {                                                                     \
 809     salsa20_matrix a, b;                                                \
 810     dstr d = DSTR_INIT;                                                 \
 811     int i, j, n;                                                        \
 812     int ok = 1;                                                         \
 813                                                                         \
 814     DENSURE(&d, SALSA20_OUTSZ); d.len = SALSA20_OUTSZ;                  \
 815     n = *(int *)v[0].buf;                                               \
 816     for (i = 0; i < SALSA20_OUTSZ/4; i++)                               \
 817       b[i] = LOAD32_L(v[1].buf + 4*i);                                  \
 818     for (i = 0; i < n; i++) {                                           \
 819       for (j = 0; j < 16; j++) a[perm[j]] = b[j];                       \
 820       core(r, a, b);                                                    \
 821       memcpy(a, b, sizeof(a));                                          \
 822     }                                                                   \
 823     for (i = 0; i < SALSA20_OUTSZ/4; i++) STORE32_L(d.buf + 4*i, b[i]); \
 824                                                                         \
 825     if (d.len != v[2].len || memcmp(d.buf, v[2].buf, v[2].len) != 0) {  \
 826       ok = 0;                                                           \
 827       printf("\nfail core:"                                             \
 828              "\n\titerations = %d"                                      \
 829              "\n\tin       = ", n);                                     \
 830       type_hex.dump(&v[1], stdout);                                     \
 831       printf("\n\texpected   = ");                                      \
 832       type_hex.dump(&v[2], stdout);                                     \
 833       printf("\n\tcalculated = ");                                      \
 834       type_hex.dump(&d, stdout);                                        \
 835       putchar('\n');                                                    \
 836     }                                                                   \
 837                                                                         \
 838     dstr_destroy(&d);                                                   \
 839     return (ok);                                                        \
 840   }
 841 SALSA20_VARS(DEFVCORE)
 842
 843 #define SALSA20_CTX(r) salsa20_ctx
 844
 845 #define SALSA20_TESTSETUP(r, ctx, k, ksz, n, nsz, p, psz) do {          \
 846   kludge64 pos64;                                                       \
 847   salsa20_init(ctx, k, ksz, 0);                                         \
 848   if (nsz == 8) salsa20_setnonce(ctx, n);                               \
 849   if (psz == 8) { LOAD64_(pos64, p); salsa20_seeku64(ctx, pos64); }     \
 850 } while (0)
 851
 852 #define XSALSA20_TESTSETUP(r, ctx, k, ksz, n, nsz, p, psz) do {         \
 853   kludge64 pos64;                                                       \
 854   XSALSA20_INIT(r, ctx, k, ksz, 0);                                     \
 855   if (nsz == 24) XSALSA20_SETNONCE(r, ctx, n);                          \
 856   if (psz == 8) { LOAD64_(pos64, p); XSALSA20_SEEKU64(r, ctx, pos64); } \
 857 } while (0)
 858
 859 #define DEFxVENC(base, BASE, r)                                         \
 860   static int v_encrypt_##base##_##r(dstr *v)                            \
 861   {                                                                     \
 862     BASE##_CTX(r) ctx;                                                  \
 863     dstr d = DSTR_INIT;                                                 \
 864     const octet *p, *p0;                                                \
 865     octet *q;                                                           \
 866     size_t sz, sz0, step;                                               \
 867     unsigned long skip;                                                 \
 868     int ok = 1;                                                         \
 869                                                                         \
 870     if (v[4].len) { p0 = (const octet *)v[4].buf; sz0 = v[4].len; }     \
 871     else { p0 = 0; sz0 = v[5].len; }                                    \
 872     DENSURE(&d, sz0); d.len = sz0;                                      \
 873     skip = *(unsigned long *)v[3].buf;                                  \
 874                                                                         \
 875     step = 0;                                                           \
 876     while (step < sz0 + skip) {                                         \
 877       step = step ? 3*step + 4 : 1;                                     \
 878       if (step > sz0 + skip) step = sz0 + skip;                         \
 879       BASE##_TESTSETUP(r, &ctx, v[0].buf, v[0].len,                     \
 880                        v[1].buf, v[1].len, v[2].buf, v[2].len);         \
 881                                                                         \
 882       for (sz = skip; sz >= step; sz -= step)                           \
 883         BASE##_ENCRYPT(r, &ctx, 0, 0, step);                            \
 884       if (sz) BASE##_ENCRYPT(r, &ctx, 0, 0, sz);                        \
 885       for (p = p0, q = (octet *)d.buf, sz = sz0;                        \
 886            sz >= step;                                                  \
 887            sz -= step, q += step) {                                     \
 888         BASE##_ENCRYPT(r, &ctx, p, q, step);                            \
 889         if (p) p += step;                                               \
 890       }                                                                 \
 891       if (sz) BASE##_ENCRYPT(r, &ctx, p, q, sz);                        \
 892                                                                         \
 893       if (d.len != v[5].len || memcmp(d.buf, v[5].buf, v[5].len) != 0) { \
 894         ok = 0;                                                         \
 895         printf("\nfail encrypt:"                                        \
 896                "\n\tstep           = %lu"                               \
 897                "\n\tkey    = ", (unsigned long)step);                   \
 898         type_hex.dump(&v[0], stdout);                                   \
 899         printf("\n\tnonce          = ");                                \
 900         type_hex.dump(&v[1], stdout);                                   \
 901         printf("\n\tposition   = ");                                    \
 902         type_hex.dump(&v[2], stdout);                                   \
 903         printf("\n\tskip           = %lu", skip);                       \
 904         printf("\n\tmessage    = ");                                    \
 905         type_hex.dump(&v[4], stdout);                                   \
 906         printf("\n\texpected   = ");                                    \
 907         type_hex.dump(&v[5], stdout);                                   \
 908         printf("\n\tcalculated = ");                                    \
 909         type_hex.dump(&d, stdout);                                      \
 910         putchar('\n');                                                  \
 911       }                                                                 \
 912     }                                                                   \
 913                                                                         \
 914     dstr_destroy(&d);                                                   \
 915     return (ok);                                                        \
 916   }
 917 #define DEFVENC(r) DEFxVENC(salsa20, SALSA20, r)
 918 #define DEFXVENC(r) DEFxVENC(xsalsa20, XSALSA20, r)
 919 SALSA20_VARS(DEFVENC)
 920 SALSA20_VARS(DEFXVENC)
 921
 922 static test_chunk defs[] = {
 923 #define DEFxTAB(pre, base, r)                                           \
 924   { pre SALSA20_NAME_##r, v_encrypt_##base##_##r,                       \
 925     { &type_hex, &type_hex, &type_hex, &type_ulong,                     \
 926       &type_hex, &type_hex, 0 } },
 927 #define DEFTAB(r)                                                       \
 928   { SALSA20_NAME_##r "-core", v_core_##r,                               \
 929     { &type_int, &type_hex, &type_hex, 0 } },                           \
 930   DEFxTAB("", salsa20, r)
 931 #define DEFXTAB(r) DEFxTAB("x", xsalsa20, r)
 932 SALSA20_VARS(DEFTAB)
 933 SALSA20_VARS(DEFXTAB)
 934   { 0, 0, { 0 } }
 935 };
 936
 937 int main(int argc, char *argv[])
 938 {
 939   test_run(argc, argv, defs, SRCDIR"/t/salsa20");
 940   return (0);
 941 }
 942
 943 #endif
 944
 945 /*----- That's all, folks -------------------------------------------------*/