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 #if CPUFAM_ARM64
  82 extern core__functype salsa20_core_arm64;
  83 #endif
  84
  85 static core__functype *pick_core(void)
  86 {
  87 #if CPUFAM_X86 || CPUFAM_AMD64
  88   DISPATCH_PICK_COND(salsa20_core, salsa20_core_x86ish_sse2,
  89                      cpu_feature_p(CPUFEAT_X86_SSE2));
  90 #endif
  91 #if CPUFAM_ARMEL
  92   DISPATCH_PICK_COND(salsa20_core, salsa20_core_arm_neon,
  93                      cpu_feature_p(CPUFEAT_ARM_NEON));
  94 #endif
  95 #if CPUFAM_ARM64
  96   DISPATCH_PICK_COND(salsa20_core, salsa20_core_arm64, 1);
  97 #endif
  98   DISPATCH_PICK_FALLBACK(salsa20_core, simple_core);
  99 }
 100
 101 /* --- @populate@ --- *
 102  *
 103  * Arguments:   @salsa20_matrix a@ = a matrix to fill in
 104  *              @const void *key@ = pointer to key material
 105  *              @size_t ksz@ = size of key
 106  *
 107  * Returns:     ---
 108  *
 109  * Use:         Fills in a Salsa20 matrix from the key, setting the
 110  *              appropriate constants according to the key length.  The nonce
 111  *              and position words are left uninitialized.
 112  */
 113
 114 static void populate(salsa20_matrix a, const void *key, size_t ksz)
 115 {
 116   const octet *k = key;
 117
 118   KSZ_ASSERT(salsa20, ksz);
 119
 120   /* Here's the pattern of key, constant, nonce, and counter pieces in the
 121    * matrix, before and after our permutation.
 122    *
 123    * [ C0 K0 K1 K2 ]       [ C0 C1 C2 C3 ]
 124    * [ K3 C1 N0 N1 ]  -->  [ K3 T1 K7 K2 ]
 125    * [ T0 T1 C2 K4 ]       [ T0 K6 K1 N1 ]
 126    * [ K5 K6 K7 C3 ]       [ K5 K0 N0 K4 ]
 127    */
 128
 129   a[13] = LOAD32_L(k +  0);
 130   a[10] = LOAD32_L(k +  4);
 131   if (ksz == 10) {
 132     a[ 7] = LOAD16_L(k +  8);
 133     a[ 4] = 0;
 134   } else {
 135     a[ 7] = LOAD32_L(k +  8);
 136     a[ 4] = LOAD32_L(k + 12);
 137   }
 138   if (ksz <= 16) {
 139     a[15] = a[13];
 140     a[12] = a[10];
 141     a[ 9] = a[ 7];
 142     a[ 6] = a[ 4];
 143     a[ 0] = SALSA20_A128;
 144     a[ 1] = SALSA20_B128;
 145     a[ 2] = ksz == 10 ? SALSA20_C80 : SALSA20_C128;
 146     a[ 3] = SALSA20_D128;
 147   } else {
 148     a[15] = LOAD32_L(k + 16);
 149     a[12] = LOAD32_L(k + 20);
 150     a[ 9] = LOAD32_L(k + 24);
 151     a[ 6] = LOAD32_L(k + 28);
 152     a[ 0] = SALSA20_A256;
 153     a[ 1] = SALSA20_B256;
 154     a[ 2] = SALSA20_C256;
 155     a[ 3] = SALSA20_D256;
 156   }
 157 }
 158
 159 /*----- Salsa20 implementation --------------------------------------------*/
 160
 161 /* --- @salsa20_init@ --- *
 162  *
 163  * Arguments:   @salsa20_ctx *ctx@ = context to fill in
 164  *              @const void *key@ = pointer to key material
 165  *              @size_t ksz@ = size of key (either 32 or 16)
 166  *              @const void *nonce@ = initial nonce, or null
 167  *
 168  * Returns:     ---
 169  *
 170  * Use:         Initializes a Salsa20 context ready for use.
 171  */
 172
 173 void salsa20_init(salsa20_ctx *ctx, const void *key, size_t ksz,
 174                   const void *nonce)
 175 {
 176   static const octet zerononce[SALSA20_NONCESZ];
 177
 178   populate(ctx->a, key, ksz);
 179   salsa20_setnonce(ctx, nonce ? nonce : zerononce);
 180 }
 181
 182 /* --- @salsa20_setnonce{,_ietf}@ --- *
 183  *
 184  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 185  *              @const void *nonce@ = the nonce (@SALSA20_NONCESZ@ or
 186  *                      @SALSA20_IETF_NONCESZ@ bytes)
 187  *
 188  * Returns:     ---
 189  *
 190  * Use:         Set a new nonce in the context @ctx@, e.g., for processing a
 191  *              different message.  The stream position is reset to zero (see
 192  *              @salsa20_seek@ etc.).
 193  */
 194
 195 void salsa20_setnonce(salsa20_ctx *ctx, const void *nonce)
 196 {
 197   const octet *n = nonce;
 198
 199   ctx->a[14] = LOAD32_L(n + 0);
 200   ctx->a[11] = LOAD32_L(n + 4);
 201   salsa20_seek(ctx, 0);
 202 }
 203
 204 void salsa20_setnonce_ietf(salsa20_ctx *ctx, const void *nonce)
 205 {
 206   const octet *n = nonce;
 207
 208   ctx->a[ 5] = LOAD32_L(n + 0);
 209   ctx->a[14] = LOAD32_L(n + 4);
 210   ctx->a[11] = LOAD32_L(n + 8);
 211   salsa20_seek_ietf(ctx, 0);
 212 }
 213
 214 /* --- @salsa20_seek{,u64,_ietf}@ --- *
 215  *
 216  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 217  *              @unsigned long i@, @kludge64 i@, @uint32@ = new position
 218  *
 219  * Returns:     ---
 220  *
 221  * Use:         Sets a new stream position, in units of Salsa20 output
 222  *              blocks, which are @SALSA20_OUTSZ@ bytes each.  Byte
 223  *              granularity can be achieved by calling @salsa20R_encrypt@
 224  *              appropriately.
 225  */
 226
 227 void salsa20_seek(salsa20_ctx *ctx, unsigned long i)
 228   { kludge64 ii; ASSIGN64(ii, i); salsa20_seeku64(ctx, ii); }
 229
 230 void salsa20_seeku64(salsa20_ctx *ctx, kludge64 i)
 231 {
 232   ctx->a[8] = LO64(i); ctx->a[5] = HI64(i);
 233   ctx->bufi = SALSA20_OUTSZ;
 234 }
 235
 236 void salsa20_seek_ietf(salsa20_ctx *ctx, uint32 i)
 237   { ctx->a[8] = i; }
 238
 239 /* --- @salsa20_tell{,u64,_ietf}@ --- *
 240  *
 241  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 242  *
 243  * Returns:     The current position in the output stream, in blocks,
 244  *              rounding upwards.
 245  */
 246
 247 unsigned long salsa20_tell(salsa20_ctx *ctx)
 248   { kludge64 i = salsa20_tellu64(ctx); return (GET64(unsigned long, i)); }
 249
 250 kludge64 salsa20_tellu64(salsa20_ctx *ctx)
 251   { kludge64 i; SET64(i, ctx->a[5], ctx->a[8]); return (i); }
 252
 253 uint32 salsa20_tell_ietf(salsa20_ctx *ctx)
 254   { return (ctx->a[5]); }
 255
 256 /* --- @salsa20{,12,8}_encrypt@ --- *
 257  *
 258  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 259  *              @const void *src@ = source buffer (or null)
 260  *              @void *dest@ = destination buffer (or null)
 261  *              @size_t sz@ = size of the buffers
 262  *
 263  * Returns:     ---
 264  *
 265  * Use:         Encrypts or decrypts @sz@ bytes of data from @src@ to @dest@.
 266  *              Salsa20 works by XORing plaintext with a keystream, so
 267  *              encryption and decryption are the same operation.  If @dest@
 268  *              is null then ignore @src@ and skip @sz@ bytes of the
 269  *              keystream.  If @src@ is null, then just write the keystream
 270  *              to @dest@.
 271  */
 272
 273 #define SALSA20_ENCRYPT(r, ctx, src, dest, sz)                          \
 274   SALSA20_DECOR(salsa20, r, _encrypt)(ctx, src, dest, sz)
 275 #define DEFENCRYPT(r)                                                   \
 276   void SALSA20_ENCRYPT(r, salsa20_ctx *ctx, const void *src,            \
 277                        void *dest, size_t sz)                           \
 278   {                                                                     \
 279     salsa20_matrix b;                                                   \
 280     const octet *s = src;                                               \
 281     octet *d = dest;                                                    \
 282     size_t n;                                                           \
 283     kludge64 pos, delta;                                                \
 284                                                                         \
 285     SALSA20_OUTBUF(ctx, d, s, sz);                                      \
 286     if (!sz) return;                                                    \
 287                                                                         \
 288     if (!dest) {                                                        \
 289       n = sz/SALSA20_OUTSZ;                                             \
 290       pos = salsa20_tellu64(ctx);                                       \
 291       ASSIGN64(delta, n);                                               \
 292       ADD64(pos, pos, delta);                                           \
 293       salsa20_seeku64(ctx, pos);                                        \
 294       sz = sz%SALSA20_OUTSZ;                                            \
 295     } else if (!src) {                                                  \
 296       while (sz >= SALSA20_OUTSZ) {                                     \
 297         core(r, ctx->a, b);                                             \
 298         SALSA20_STEP(ctx->a);                                           \
 299         SALSA20_GENFULL(b, d);                                          \
 300         sz -= SALSA20_OUTSZ;                                            \
 301       }                                                                 \
 302     } else {                                                            \
 303       while (sz >= SALSA20_OUTSZ) {                                     \
 304         core(r, ctx->a, b);                                             \
 305         SALSA20_STEP(ctx->a);                                           \
 306         SALSA20_MIXFULL(b, d, s);                                       \
 307         sz -= SALSA20_OUTSZ;                                            \
 308       }                                                                 \
 309     }                                                                   \
 310                                                                         \
 311     if (sz) {                                                           \
 312       core(r, ctx->a, b);                                               \
 313       SALSA20_STEP(ctx->a);                                             \
 314       SALSA20_PREPBUF(ctx, b);                                          \
 315       SALSA20_OUTBUF(ctx, d, s, sz);                                    \
 316       assert(!sz);                                                      \
 317     }                                                                   \
 318   }
 319 SALSA20_VARS(DEFENCRYPT)
 320
 321 /*----- HSalsa20 implementation -------------------------------------------*/
 322
 323 #define HSALSA20_RAW(r, ctx, src, dest)                                 \
 324   SALSA20_DECOR(hsalsa20, r, _raw)(ctx, src, dest)
 325 #define HSALSA20_PRF(r, ctx, src, dest)                                 \
 326   SALSA20_DECOR(hsalsa20, r, _prf)(ctx, src, dest)
 327
 328 /* --- @hsalsa20{,12,8}_prf@ --- *
 329  *
 330  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 331  *              @const void *src@ = the input (@HSALSA20_INSZ@ bytes)
 332  *              @void *dest@ = the output (@HSALSA20_OUTSZ@ bytes)
 333  *
 334  * Returns:     ---
 335  *
 336  * Use:         Apply the HSalsa20/r pseudorandom function to @src@, writing
 337  *              the result to @out@.
 338  */
 339
 340 #define DEFHSALSA20(r)                                                  \
 341   static void HSALSA20_RAW(r, salsa20_matrix k,                         \
 342                            const uint32 *src, uint32 *dest)             \
 343   {                                                                     \
 344     salsa20_matrix a;                                                   \
 345     int i;                                                              \
 346                                                                         \
 347     /* --- HSalsa20, computed from full Salsa20 --- *                   \
 348      *                                                                  \
 349      * The security proof makes use of the fact that HSalsa20 (i.e.,    \
 350      * without the final feedforward step) can be computed from full    \
 351      * Salsa20 using only knowledge of the non-secret input.  I don't   \
 352      * want to compromise the performance of the main function by       \
 353      * making the feedforward step separate, but this operation is less \
 354      * speed critical, so we do it the harder way.                      \
 355      */                                                                 \
 356                                                                         \
 357     for (i = 0; i < 4; i++) k[14 - 3*i] = src[i];                       \
 358     core(r, k, a);                                                      \
 359     for (i = 0; i < 4; i++) dest[i] = a[5*i] - k[i];                    \
 360     for (i = 4; i < 8; i++) dest[i] = a[i + 2] - k[26 - 3*i];           \
 361   }                                                                     \
 362                                                                         \
 363   void HSALSA20_PRF(r, salsa20_ctx *ctx, const void *src, void *dest)   \
 364   {                                                                     \
 365     const octet *s = src;                                               \
 366     octet *d = dest;                                                    \
 367     uint32 in[4], out[8];                                               \
 368     int i;                                                              \
 369                                                                         \
 370     for (i = 0; i < 4; i++) in[i] = LOAD32_L(s + 4*i);                  \
 371     HSALSA20_RAW(r, ctx->a, in, out);                                   \
 372     for (i = 0; i < 8; i++) STORE32_L(d + 4*i, out[i]);                 \
 373   }
 374 SALSA20_VARS(DEFHSALSA20)
 375
 376 /*----- XSalsa20 implementation -------------------------------------------*/
 377
 378 /* --- Some convenient macros for naming functions --- *
 379  *
 380  * Because the crypto core is involved in XSalsa20/r's per-nonce setup, we
 381  * need to take an interest in the number of rounds in most of the various
 382  * functions, and it will probably help if we distinguish the context
 383  * structures for the various versions.
 384  */
 385
 386 #define XSALSA20_CTX(r) SALSA20_DECOR(xsalsa20, r, _ctx)
 387 #define XSALSA20_INIT(r, ctx, k, ksz, n)                                \
 388   SALSA20_DECOR(xsalsa20, r, _init)(ctx, k, ksz, n)
 389 #define XSALSA20_SETNONCE(r, ctx, n)                                    \
 390   SALSA20_DECOR(xsalsa20, r, _setnonce)(ctx, n)
 391 #define XSALSA20_SEEK(r, ctx, i)                                        \
 392   SALSA20_DECOR(xsalsa20, r, _seek)(ctx, i)
 393 #define XSALSA20_SEEKU64(r, ctx, i)                                     \
 394   SALSA20_DECOR(xsalsa20, r, _seeku64)(ctx, i)
 395 #define XSALSA20_TELL(r, ctx)                                           \
 396   SALSA20_DECOR(xsalsa20, r, _tell)(ctx)
 397 #define XSALSA20_TELLU64(r, ctx)                                        \
 398   SALSA20_DECOR(xsalsa20, r, _tellu64)(ctx)
 399 #define XSALSA20_ENCRYPT(r, ctx, src, dest, sz)                         \
 400   SALSA20_DECOR(xsalsa20, r, _encrypt)(ctx, src, dest, sz)
 401
 402 /* --- @xsalsa20{,12,8}_init@ --- *
 403  *
 404  * Arguments:   @xsalsa20R_ctx *ctx@ = the context to fill in
 405  *              @const void *key@ = pointer to key material
 406  *              @size_t ksz@ = size of key (either 32 or 16)
 407  *              @const void *nonce@ = initial nonce, or null
 408  *
 409  * Returns:     ---
 410  *
 411  * Use:         Initializes an XSalsa20/r context ready for use.
 412  *
 413  *              There is a different function for each number of rounds,
 414  *              unlike for plain Salsa20.
 415  */
 416
 417 #define DEFXINIT(r)                                                     \
 418   void XSALSA20_INIT(r, XSALSA20_CTX(r) *ctx,                           \
 419                         const void *key, size_t ksz, const void *nonce) \
 420   {                                                                     \
 421     static const octet zerononce[XSALSA20_NONCESZ];                     \
 422                                                                         \
 423     populate(ctx->k, key, ksz);                                         \
 424     ctx->s.a[ 0] = SALSA20_A256;                                        \
 425     ctx->s.a[ 1] = SALSA20_B256;                                        \
 426     ctx->s.a[ 2] = SALSA20_C256;                                        \
 427     ctx->s.a[ 3] = SALSA20_D256;                                        \
 428     XSALSA20_SETNONCE(r, ctx, nonce ? nonce : zerononce);               \
 429   }
 430 SALSA20_VARS(DEFXINIT)
 431
 432 /* --- @xsalsa20{,12,8}_setnonce@ --- *
 433  *
 434  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 435  *              @const void *nonce@ = the nonce (@XSALSA20_NONCESZ@ bytes)
 436  *
 437  * Returns:     ---
 438  *
 439  * Use:         Set a new nonce in the context @ctx@, e.g., for processing a
 440  *              different message.  The stream position is reset to zero (see
 441  *              @salsa20_seek@ etc.).
 442  *
 443  *              There is a different function for each number of rounds,
 444  *              unlike for plain Salsa20.
 445  */
 446
 447 #define DEFXNONCE(r)                                                    \
 448   void XSALSA20_SETNONCE(r, XSALSA20_CTX(r) *ctx, const void *nonce)    \
 449   {                                                                     \
 450     const octet *n = nonce;                                             \
 451     uint32 in[4], out[8];                                               \
 452     int i;                                                              \
 453                                                                         \
 454     for (i = 0; i < 4; i++) in[i] = LOAD32_L(n + 4*i);                  \
 455     HSALSA20_RAW(r, ctx->k, in, out);                                   \
 456     for (i = 0; i < 4; i++) ctx->s.a[13 - 3*i] = out[i];                \
 457     for (i = 4; i < 8; i++) ctx->s.a[27 - 3*i] = out[i];                \
 458     salsa20_setnonce(&ctx->s, n + 16);                                  \
 459   }
 460 SALSA20_VARS(DEFXNONCE)
 461
 462 /* --- @xsalsa20{,12,8}_seek{,u64}@ --- *
 463  *
 464  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 465  *              @unsigned long i@, @kludge64 i@ = new position to set
 466  *
 467  * Returns:     ---
 468  *
 469  * Use:         Sets a new stream position, in units of Salsa20 output
 470  *              blocks, which are @XSALSA20_OUTSZ@ bytes each.  Byte
 471  *              granularity can be achieved by calling @xsalsa20R_encrypt@
 472  *              appropriately.
 473  *
 474  *              There is a different function for each number of rounds,
 475  *              unlike for plain Salsa20, because the context structures are
 476  *              different.
 477  */
 478
 479 /* --- @xsalsa20{,12,8}_tell{,u64}@ --- *
 480  *
 481  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 482  *
 483  * Returns:     The current position in the output stream, in blocks,
 484  *              rounding upwards.
 485  *
 486  *              There is a different function for each number of rounds,
 487  *              unlike for plain Salsa20, because the context structures are
 488  *              different.
 489  */
 490
 491 /* --- @xsalsa20{,12,8}_encrypt@ --- *
 492  *
 493  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 494  *              @const void *src@ = source buffer (or null)
 495  *              @void *dest@ = destination buffer (or null)
 496  *              @size_t sz@ = size of the buffers
 497  *
 498  * Returns:     ---
 499  *
 500  * Use:         Encrypts or decrypts @sz@ bytes of data from @src@ to @dest@.
 501  *              XSalsa20 works by XORing plaintext with a keystream, so
 502  *              encryption and decryption are the same operation.  If @dest@
 503  *              is null then ignore @src@ and skip @sz@ bytes of the
 504  *              keystream.  If @src@ is null, then just write the keystream
 505  *              to @dest@.
 506  */
 507
 508 #define DEFXPASSTHRU(r)                                                 \
 509   void XSALSA20_SEEK(r, XSALSA20_CTX(r) *ctx, unsigned long i)          \
 510     { salsa20_seek(&ctx->s, i); }                                       \
 511   void XSALSA20_SEEKU64(r, XSALSA20_CTX(r) *ctx, kludge64 i)            \
 512     { salsa20_seeku64(&ctx->s, i); }                                    \
 513   unsigned long XSALSA20_TELL(r, XSALSA20_CTX(r) *ctx)                  \
 514     { return salsa20_tell(&ctx->s); }                                   \
 515   kludge64 XSALSA20_TELLU64(r, XSALSA20_CTX(r) *ctx)                    \
 516     { return salsa20_tellu64(&ctx->s); }                                \
 517   void XSALSA20_ENCRYPT(r, XSALSA20_CTX(r) *ctx,                        \
 518                         const void *src, void *dest, size_t sz)         \
 519     { SALSA20_ENCRYPT(r, &ctx->s, src, dest, sz); }
 520 SALSA20_VARS(DEFXPASSTHRU)
 521
 522 /*----- Generic cipher interface ------------------------------------------*/
 523
 524 typedef struct gctx { gcipher c; salsa20_ctx ctx; } gctx;
 525
 526 static void gsetiv(gcipher *c, const void *iv)
 527   { gctx *g = (gctx *)c; salsa20_setnonce(&g->ctx, iv); }
 528
 529 static void gsetiv_ietf(gcipher *c, const void *iv)
 530   { gctx *g = (gctx *)c; salsa20_setnonce_ietf(&g->ctx, iv); }
 531
 532 static void gdestroy(gcipher *c)
 533   { gctx *g = (gctx *)c; BURN(*g); S_DESTROY(g); }
 534
 535 static gcipher *ginit(const void *k, size_t sz, const gcipher_ops *ops)
 536 {
 537   gctx *g = S_CREATE(gctx);
 538   g->c.ops = ops;
 539   salsa20_init(&g->ctx, k, sz, 0);
 540   return (&g->c);
 541 }
 542
 543 #define DEFGCIPHER(r)                                                   \
 544                                                                         \
 545   static const gcipher_ops gops_##r, gops_##r##_ietf;                   \
 546                                                                         \
 547   static gcipher *ginit_##r(const void *k, size_t sz)                   \
 548     { return (ginit(k, sz, &gops_##r)); }                               \
 549                                                                         \
 550   static gcipher *ginit_##r##_ietf(const void *k, size_t sz)            \
 551     { return (ginit(k, sz, &gops_##r##_ietf)); }                        \
 552                                                                         \
 553   static void gencrypt_##r(gcipher *c, const void *s,                   \
 554                            void *t, size_t sz)                          \
 555     { gctx *g = (gctx *)c; SALSA20_ENCRYPT(r, &g->ctx, s, t, sz); }     \
 556                                                                         \
 557   static const gcipher_ops gops_##r = {                                 \
 558     &SALSA20_DECOR(salsa20, r, ),                                       \
 559     gencrypt_##r, gencrypt_##r, gdestroy, gsetiv, 0                     \
 560   };                                                                    \
 561                                                                         \
 562   static const gcipher_ops gops_##r##_ietf = {                          \
 563     &SALSA20_DECOR(salsa20, r, _ietf),                                  \
 564     gencrypt_##r, gencrypt_##r, gdestroy, gsetiv_ietf, 0                \
 565   };                                                                    \
 566                                                                         \
 567   const gccipher SALSA20_DECOR(salsa20, r, ) = {                        \
 568     SALSA20_NAME_##r, salsa20_keysz,                                    \
 569     SALSA20_NONCESZ, ginit_##r                                          \
 570   };                                                                    \
 571                                                                         \
 572   const gccipher SALSA20_DECOR(salsa20, r, _ietf) = {                   \
 573     SALSA20_NAME_##r "-ietf", salsa20_keysz,                            \
 574     SALSA20_IETF_NONCESZ, ginit_##r##_ietf                              \
 575   };
 576
 577 SALSA20_VARS(DEFGCIPHER)
 578
 579 #define DEFGXCIPHER(r)                                                  \
 580                                                                         \
 581   typedef struct { gcipher c; XSALSA20_CTX(r) ctx; } gxctx_##r;         \
 582                                                                         \
 583   static void gxsetiv_##r(gcipher *c, const void *iv)                   \
 584     { gxctx_##r *g = (gxctx_##r *)c; XSALSA20_SETNONCE(r, &g->ctx, iv); } \
 585                                                                         \
 586   static void gxdestroy_##r(gcipher *c)                                 \
 587     { gxctx_##r *g = (gxctx_##r *)c; BURN(*g); S_DESTROY(g); }          \
 588                                                                         \
 589   static const gcipher_ops gxops_##r;                                   \
 590                                                                         \
 591   static gcipher *gxinit_##r(const void *k, size_t sz)                  \
 592   {                                                                     \
 593     gxctx_##r *g = S_CREATE(gxctx_##r);                                 \
 594     g->c.ops = &gxops_##r;                                              \
 595     XSALSA20_INIT(r, &g->ctx, k, sz, 0);                                \
 596     return (&g->c);                                                     \
 597   }                                                                     \
 598                                                                         \
 599   static void gxencrypt_##r(gcipher *c, const void *s,                  \
 600                             void *t, size_t sz)                         \
 601   {                                                                     \
 602     gxctx_##r *g = (gxctx_##r *)c;                                      \
 603     XSALSA20_ENCRYPT(r, &g->ctx, s, t, sz);                             \
 604   }                                                                     \
 605                                                                         \
 606   static const gcipher_ops gxops_##r = {                                \
 607     &SALSA20_DECOR(xsalsa20, r, ),                                      \
 608     gxencrypt_##r, gxencrypt_##r, gxdestroy_##r, gxsetiv_##r, 0         \
 609   };                                                                    \
 610                                                                         \
 611   const gccipher SALSA20_DECOR(xsalsa20, r, ) = {                       \
 612     "x" SALSA20_NAME_##r, salsa20_keysz,                                \
 613     XSALSA20_NONCESZ, gxinit_##r                                        \
 614   };
 615
 616 SALSA20_VARS(DEFGXCIPHER)
 617
 618 /*----- Generic random number generator interface -------------------------*/
 619
 620 typedef struct grops {
 621   size_t noncesz;
 622   void (*seek)(void *, kludge64);
 623   kludge64 (*tell)(void *);
 624   void (*setnonce)(void *, const void *);
 625   void (*generate)(void *, void *, size_t);
 626 } grops;
 627
 628 typedef struct grbasectx {
 629   grand r;
 630   const grops *ops;
 631 } grbasectx;
 632
 633 static int grmisc(grand *r, unsigned op, ...)
 634 {
 635   octet buf[XSALSA20_NONCESZ];
 636   grbasectx *g = (grbasectx *)r;
 637   grand *rr;
 638   const octet *p;
 639   size_t sz;
 640   uint32 i;
 641   unsigned long ul;
 642   kludge64 pos;
 643   va_list ap;
 644   int rc = 0;
 645
 646   va_start(ap, op);
 647
 648   switch (op) {
 649     case GRAND_CHECK:
 650       switch (va_arg(ap, unsigned)) {
 651         case GRAND_CHECK:
 652         case GRAND_SEEDINT:
 653         case GRAND_SEEDUINT32:
 654         case GRAND_SEEDBLOCK:
 655         case GRAND_SEEDRAND:
 656         case SALSA20_SEEK:
 657         case SALSA20_SEEKU64:
 658         case SALSA20_TELL:
 659         case SALSA20_TELLU64:
 660           rc = 1;
 661           break;
 662         default:
 663           rc = 0;
 664           break;
 665       }
 666       break;
 667
 668     case GRAND_SEEDINT:
 669       i = va_arg(ap, unsigned); STORE32_L(buf, i);
 670       memset(buf + 4, 0, g->ops->noncesz - 4);
 671       g->ops->setnonce(g, buf);
 672       break;
 673     case GRAND_SEEDUINT32:
 674       i = va_arg(ap, uint32); STORE32_L(buf, i);
 675       memset(buf + 4, 0, g->ops->noncesz - 4);
 676       g->ops->setnonce(g, buf);
 677       break;
 678     case GRAND_SEEDBLOCK:
 679       p = va_arg(ap, const void *);
 680       sz = va_arg(ap, size_t);
 681       if (sz < g->ops->noncesz) {
 682         memcpy(buf, p, sz);
 683         memset(buf + sz, 0, g->ops->noncesz - sz);
 684         p = buf;
 685       }
 686       g->ops->setnonce(g, p);
 687       break;
 688     case GRAND_SEEDRAND:
 689       rr = va_arg(ap, grand *);
 690       rr->ops->fill(rr, buf, g->ops->noncesz);
 691       g->ops->setnonce(g, buf);
 692       break;
 693     case SALSA20_SEEK:
 694       ul = va_arg(ap, unsigned long); ASSIGN64(pos, ul);
 695       g->ops->seek(g, pos);
 696       break;
 697     case SALSA20_SEEKU64:
 698       pos = va_arg(ap, kludge64);
 699       g->ops->seek(g, pos);
 700       break;
 701     case SALSA20_TELL:
 702       pos = g->ops->tell(g);
 703       *va_arg(ap, unsigned long *) = GET64(unsigned long, pos);
 704       break;
 705     case SALSA20_TELLU64:
 706       *va_arg(ap, kludge64 *) = g->ops->tell(g);
 707       break;
 708     default:
 709       GRAND_BADOP;
 710       break;
 711   }
 712
 713   return (rc);
 714 }
 715
 716 static octet grbyte(grand *r)
 717 {
 718   grbasectx *g = (grbasectx *)r;
 719   octet o;
 720   g->ops->generate(g, &o, 1);
 721   return (o);
 722 }
 723
 724 static uint32 grword(grand *r)
 725 {
 726   grbasectx *g = (grbasectx *)r;
 727   octet b[4];
 728   g->ops->generate(g, b, sizeof(b));
 729   return (LOAD32_L(b));
 730 }
 731
 732 static void grfill(grand *r, void *p, size_t sz)
 733 {
 734   grbasectx *g = (grbasectx *)r;
 735   g->ops->generate(r, p, sz);
 736 }
 737
 738 typedef struct grctx {
 739   grbasectx r;
 740   salsa20_ctx ctx;
 741 } grctx;
 742
 743 static void gr_seek(void *r, kludge64 pos)
 744   { grctx *g = r; salsa20_seeku64(&g->ctx, pos); }
 745
 746 static void gr_seek_ietf(void *r, kludge64 pos)
 747   { grctx *g = r; salsa20_seek_ietf(&g->ctx, LO64(pos)); }
 748
 749 static kludge64 gr_tell(void *r)
 750   { grctx *g = r; return (salsa20_tellu64(&g->ctx)); }
 751
 752 static kludge64 gr_tell_ietf(void *r)
 753 {
 754   grctx *g = r;
 755   kludge64 pos;
 756
 757   SET64(pos, 0, salsa20_tell_ietf(&g->ctx));
 758   return (pos);
 759 }
 760
 761 static void gr_setnonce(void *r, const void *n)
 762   { grctx *g = r; salsa20_setnonce(&g->ctx, n); }
 763
 764 static void gr_setnonce_ietf(void *r, const void *n)
 765   { grctx *g = r; salsa20_setnonce(&g->ctx, n); }
 766
 767 static void grdestroy(grand *r)
 768   { grctx *g = (grctx *)r; BURN(*g); S_DESTROY(g); }
 769
 770 static grand *grinit(const void *k, size_t ksz, const void *n,
 771                      const grand_ops *ops, const grops *myops)
 772 {
 773     grctx *g = S_CREATE(grctx);
 774     g->r.r.ops = ops;
 775     g->r.ops = myops;
 776     salsa20_init(&g->ctx, k, ksz, 0);
 777     if (n) myops->setnonce(g, n);
 778     return (&g->r.r);
 779 }
 780
 781 #define DEFGRAND(rr)                                                    \
 782                                                                         \
 783   static void gr_generate_##rr(void *r, void *b, size_t sz)             \
 784     { grctx *g = r; SALSA20_ENCRYPT(rr, &g->ctx, 0, b, sz); }           \
 785                                                                         \
 786   static const grops grops_##rr =                                       \
 787     { SALSA20_NONCESZ, gr_seek, gr_tell,                                \
 788       gr_setnonce, gr_generate_##rr };                                  \
 789                                                                         \
 790   static const grops grops_##rr##_ietf =                                \
 791     { SALSA20_IETF_NONCESZ, gr_seek_ietf, gr_tell_ietf,                 \
 792       gr_setnonce_ietf, gr_generate_##rr };                             \
 793                                                                         \
 794   static const grand_ops grops_rand_##rr = {                            \
 795     SALSA20_NAME_##rr, GRAND_CRYPTO, 0,                                 \
 796     grmisc, grdestroy, grword,                                          \
 797     grbyte, grword, grand_defaultrange, grfill                          \
 798   };                                                                    \
 799                                                                         \
 800   static const grand_ops grops_rand_##rr##_ietf = {                     \
 801     SALSA20_NAME_##rr "-ietf", GRAND_CRYPTO, 0,                         \
 802     grmisc, grdestroy, grword,                                          \
 803     grbyte, grword, grand_defaultrange, grfill                          \
 804   };                                                                    \
 805                                                                         \
 806   grand *SALSA20_DECOR(salsa20, rr, _rand)                              \
 807     (const void *k, size_t ksz, const void *n)                          \
 808     { return (grinit(k, ksz, n, &grops_rand_##rr, &grops_##rr)); }      \
 809                                                                         \
 810   grand *SALSA20_DECOR(salsa20, rr, _ietf_rand)                         \
 811     (const void *k, size_t ksz, const void *n)                          \
 812   {                                                                     \
 813     return (grinit(k, ksz, n,                                           \
 814                    &grops_rand_##rr##_ietf,                             \
 815                    &grops_##rr##_ietf));                                \
 816   }
 817
 818 SALSA20_VARS(DEFGRAND)
 819
 820 #define DEFXGRAND(rr)                                                   \
 821                                                                         \
 822   typedef struct grxctx_##rr {                                          \
 823     grbasectx r;                                                        \
 824     XSALSA20_CTX(rr) ctx;                                               \
 825   } grxctx_##rr;                                                        \
 826                                                                         \
 827   static void grx_seek_##rr(void *r, kludge64 pos)                      \
 828     { grxctx_##rr *g = r; XSALSA20_SEEKU64(rr, &g->ctx, pos); }         \
 829                                                                         \
 830   static kludge64 grx_tell_##rr(void *r)                                \
 831     { grxctx_##rr *g = r; return (XSALSA20_TELLU64(rr, &g->ctx)); }     \
 832                                                                         \
 833   static void grx_setnonce_##rr(void *r, const void *n)                 \
 834     { grxctx_##rr *g = r; XSALSA20_SETNONCE(rr, &g->ctx, n); }          \
 835                                                                         \
 836   static void grxdestroy_##rr(grand *r)                                 \
 837     { grxctx_##rr *g = (grxctx_##rr *)r; BURN(*g); S_DESTROY(g); }      \
 838                                                                         \
 839   static void grx_generate_##rr(void *r, void *b, size_t sz)            \
 840     { grxctx_##rr *g = r; XSALSA20_ENCRYPT(rr, &g->ctx, 0, b, sz); }    \
 841                                                                         \
 842   static const grops grxops_##rr =                                      \
 843   { XSALSA20_NONCESZ, grx_seek_##rr, grx_tell_##rr,                     \
 844       grx_setnonce_##rr, grx_generate_##rr };                           \
 845                                                                         \
 846   static const grand_ops grxops_rand_##rr = {                           \
 847     "x" SALSA20_NAME_##rr, GRAND_CRYPTO, 0,                             \
 848     grmisc, grxdestroy_##rr, grword,                                    \
 849     grbyte, grword, grand_defaultrange, grfill                          \
 850   };                                                                    \
 851                                                                         \
 852   grand *SALSA20_DECOR(xsalsa20, rr, _rand)                             \
 853     (const void *k, size_t ksz, const void *n)                          \
 854   {                                                                     \
 855     grxctx_##rr *g = S_CREATE(grxctx_##rr);                             \
 856     g->r.r.ops = &grxops_rand_##rr;                                     \
 857     g->r.ops = &grxops_##rr;                                            \
 858     XSALSA20_INIT(rr, &g->ctx, k, ksz, n);                              \
 859     return (&g->r.r);                                                   \
 860   }
 861 SALSA20_VARS(DEFXGRAND)
 862
 863 /*----- Test rig ----------------------------------------------------------*/
 864
 865 #ifdef TEST_RIG
 866
 867 #include <stdio.h>
 868 #include <string.h>
 869
 870 #include <mLib/quis.h>
 871 #include <mLib/testrig.h>
 872
 873 static const int perm[] = {
 874    0, 13, 10,  7,
 875    4,  1, 14, 11,
 876    8,  5,  2, 15,
 877   12,  9,  6,  3
 878 };
 879
 880 #define DEFVCORE(r)                                                     \
 881   static int v_core_##r(dstr *v)                                        \
 882   {                                                                     \
 883     salsa20_matrix a, b;                                                \
 884     dstr d = DSTR_INIT;                                                 \
 885     int i, j, n;                                                        \
 886     int ok = 1;                                                         \
 887                                                                         \
 888     DENSURE(&d, SALSA20_OUTSZ); d.len = SALSA20_OUTSZ;                  \
 889     n = *(int *)v[0].buf;                                               \
 890     for (i = 0; i < SALSA20_OUTSZ/4; i++)                               \
 891       b[i] = LOAD32_L(v[1].buf + 4*i);                                  \
 892     for (i = 0; i < n; i++) {                                           \
 893       for (j = 0; j < 16; j++) a[perm[j]] = b[j];                       \
 894       core(r, a, b);                                                    \
 895       memcpy(a, b, sizeof(a));                                          \
 896     }                                                                   \
 897     for (i = 0; i < SALSA20_OUTSZ/4; i++) STORE32_L(d.buf + 4*i, b[i]); \
 898                                                                         \
 899     if (d.len != v[2].len || memcmp(d.buf, v[2].buf, v[2].len) != 0) {  \
 900       ok = 0;                                                           \
 901       printf("\nfail core:"                                             \
 902              "\n\titerations = %d"                                      \
 903              "\n\tin       = ", n);                                     \
 904       type_hex.dump(&v[1], stdout);                                     \
 905       printf("\n\texpected   = ");                                      \
 906       type_hex.dump(&v[2], stdout);                                     \
 907       printf("\n\tcalculated = ");                                      \
 908       type_hex.dump(&d, stdout);                                        \
 909       putchar('\n');                                                    \
 910     }                                                                   \
 911                                                                         \
 912     dstr_destroy(&d);                                                   \
 913     return (ok);                                                        \
 914   }
 915 SALSA20_VARS(DEFVCORE)
 916
 917 #define SALSA20_CTX(r) salsa20_ctx
 918
 919 #define SALSA20_TESTSETUP(r, ctx, k, ksz, n, nsz, p, psz) do {          \
 920   kludge64 pos64;                                                       \
 921   salsa20_init(ctx, k, ksz, 0);                                         \
 922   if (nsz == 8) salsa20_setnonce(ctx, n);                               \
 923   else if (nsz == 12) salsa20_setnonce_ietf(ctx, n);                    \
 924   if (psz == 8) { LOAD64_(pos64, p); salsa20_seeku64(ctx, pos64); }     \
 925   else if (psz == 4) salsa20_seek_ietf(ctx, LOAD32(p));                 \
 926 } while (0)
 927
 928 #define XSALSA20_TESTSETUP(r, ctx, k, ksz, n, nsz, p, psz) do {         \
 929   kludge64 pos64;                                                       \
 930   XSALSA20_INIT(r, ctx, k, ksz, 0);                                     \
 931   if (nsz == 24) XSALSA20_SETNONCE(r, ctx, n);                          \
 932   if (psz == 8) { LOAD64_(pos64, p); XSALSA20_SEEKU64(r, ctx, pos64); } \
 933 } while (0)
 934
 935 #define DEFxVENC(base, BASE, r)                                         \
 936   static int v_encrypt_##base##_##r(dstr *v)                            \
 937   {                                                                     \
 938     BASE##_CTX(r) ctx;                                                  \
 939     dstr d = DSTR_INIT;                                                 \
 940     const octet *p, *p0;                                                \
 941     octet *q;                                                           \
 942     size_t sz, sz0, step;                                               \
 943     unsigned long skip;                                                 \
 944     int ok = 1;                                                         \
 945                                                                         \
 946     if (v[4].len) { p0 = (const octet *)v[4].buf; sz0 = v[4].len; }     \
 947     else { p0 = 0; sz0 = v[5].len; }                                    \
 948     DENSURE(&d, sz0); d.len = sz0;                                      \
 949     skip = *(unsigned long *)v[3].buf;                                  \
 950                                                                         \
 951     step = 0;                                                           \
 952     while (step < sz0 + skip) {                                         \
 953       step = step ? 3*step + 4 : 1;                                     \
 954       if (step > sz0 + skip) step = sz0 + skip;                         \
 955       BASE##_TESTSETUP(r, &ctx, v[0].buf, v[0].len,                     \
 956                        v[1].buf, v[1].len, v[2].buf, v[2].len);         \
 957                                                                         \
 958       for (sz = skip; sz >= step; sz -= step)                           \
 959         BASE##_ENCRYPT(r, &ctx, 0, 0, step);                            \
 960       if (sz) BASE##_ENCRYPT(r, &ctx, 0, 0, sz);                        \
 961       for (p = p0, q = (octet *)d.buf, sz = sz0;                        \
 962            sz >= step;                                                  \
 963            sz -= step, q += step) {                                     \
 964         BASE##_ENCRYPT(r, &ctx, p, q, step);                            \
 965         if (p) p += step;                                               \
 966       }                                                                 \
 967       if (sz) BASE##_ENCRYPT(r, &ctx, p, q, sz);                        \
 968                                                                         \
 969       if (d.len != v[5].len || memcmp(d.buf, v[5].buf, v[5].len) != 0) { \
 970         ok = 0;                                                         \
 971         printf("\nfail encrypt:"                                        \
 972                "\n\tstep           = %lu"                               \
 973                "\n\tkey    = ", (unsigned long)step);                   \
 974         type_hex.dump(&v[0], stdout);                                   \
 975         printf("\n\tnonce          = ");                                \
 976         type_hex.dump(&v[1], stdout);                                   \
 977         printf("\n\tposition   = ");                                    \
 978         type_hex.dump(&v[2], stdout);                                   \
 979         printf("\n\tskip           = %lu", skip);                       \
 980         printf("\n\tmessage    = ");                                    \
 981         type_hex.dump(&v[4], stdout);                                   \
 982         printf("\n\texpected   = ");                                    \
 983         type_hex.dump(&v[5], stdout);                                   \
 984         printf("\n\tcalculated = ");                                    \
 985         type_hex.dump(&d, stdout);                                      \
 986         putchar('\n');                                                  \
 987       }                                                                 \
 988     }                                                                   \
 989                                                                         \
 990     dstr_destroy(&d);                                                   \
 991     return (ok);                                                        \
 992   }
 993 #define DEFVENC(r) DEFxVENC(salsa20, SALSA20, r)
 994 #define DEFXVENC(r) DEFxVENC(xsalsa20, XSALSA20, r)
 995 SALSA20_VARS(DEFVENC)
 996 SALSA20_VARS(DEFXVENC)
 997
 998 static test_chunk defs[] = {
 999 #define DEFxTAB(pre, base, r)                                           \
1000   { pre SALSA20_NAME_##r, v_encrypt_##base##_##r,                       \
1001     { &type_hex, &type_hex, &type_hex, &type_ulong,                     \
1002       &type_hex, &type_hex, 0 } },
1003 #define DEFTAB(r)                                                       \
1004   { SALSA20_NAME_##r "-core", v_core_##r,                               \
1005     { &type_int, &type_hex, &type_hex, 0 } },                           \
1006   DEFxTAB("", salsa20, r)
1007 #define DEFXTAB(r) DEFxTAB("x", xsalsa20, r)
1008 SALSA20_VARS(DEFTAB)
1009 SALSA20_VARS(DEFXTAB)
1010   { 0, 0, { 0 } }
1011 };
1012
1013 int main(int argc, char *argv[])
1014 {
1015   test_run(argc, argv, defs, SRCDIR"/t/salsa20");
1016   return (0);
1017 }
1018
1019 #endif
1020
1021 /*----- That's all, folks -------------------------------------------------*/