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 "rsvr.h"
  43 #include "salsa20.h"
  44 #include "salsa20-core.h"
  45
  46 /*----- Global variables --------------------------------------------------*/
  47
  48 const octet salsa20_keysz[] = { KSZ_SET, 32, 16, 10, 0 };
  49
  50 /*----- The Salsa20 core function and utilities ---------------------------*/
  51
  52 /* --- @core@ --- *
  53  *
  54  * Arguments:   @unsigned r@ = number of rounds
  55  *              @const salsa20_matrix src@ = input matrix
  56  *              @salsa20_matrix dest@ = where to put the output
  57  *
  58  * Returns:     ---
  59  *
  60  *
  61  * Use:         Apply the Salsa20/r core function to @src@, writing the
  62  *              result to @dest@.  This consists of @r@ rounds followed by
  63  *              the feedforward step.
  64  */
  65
  66 CPU_DISPATCH(static, (void), void, core,
  67              (unsigned r, const salsa20_matrix src, salsa20_matrix dest),
  68              (r, src, dest), pick_core, simple_core);
  69
  70 static void simple_core(unsigned r, const salsa20_matrix src,
  71                         salsa20_matrix dest)
  72   { SALSA20_nR(dest, src, r); SALSA20_FFWD(dest, src); }
  73
  74 #if CPUFAM_X86 || CPUFAM_AMD64
  75 extern core__functype salsa20_core_x86ish_sse2;
  76 extern core__functype salsa20_core_x86ish_avx;
  77 #endif
  78
  79 #if CPUFAM_ARMEL
  80 extern core__functype salsa20_core_arm_neon;
  81 #endif
  82
  83 #if CPUFAM_ARM64
  84 extern core__functype salsa20_core_arm64;
  85 #endif
  86
  87 static core__functype *pick_core(void)
  88 {
  89 #if CPUFAM_X86 || CPUFAM_AMD64
  90   DISPATCH_PICK_COND(salsa20_core, salsa20_core_x86ish_avx,
  91                      cpu_feature_p(CPUFEAT_X86_AVX));
  92   DISPATCH_PICK_COND(salsa20_core, salsa20_core_x86ish_sse2,
  93                      cpu_feature_p(CPUFEAT_X86_SSE2));
  94 #endif
  95 #if CPUFAM_ARMEL
  96   DISPATCH_PICK_COND(salsa20_core, salsa20_core_arm_neon,
  97                      cpu_feature_p(CPUFEAT_ARM_NEON));
  98 #endif
  99 #if CPUFAM_ARM64
 100   DISPATCH_PICK_COND(salsa20_core, salsa20_core_arm64, 1);
 101 #endif
 102   DISPATCH_PICK_FALLBACK(salsa20_core, simple_core);
 103 }
 104
 105 /* --- @populate@ --- *
 106  *
 107  * Arguments:   @salsa20_matrix a@ = a matrix to fill in
 108  *              @const void *key@ = pointer to key material
 109  *              @size_t ksz@ = size of key
 110  *
 111  * Returns:     ---
 112  *
 113  * Use:         Fills in a Salsa20 matrix from the key, setting the
 114  *              appropriate constants according to the key length.  The nonce
 115  *              and position words are left uninitialized.
 116  */
 117
 118 static void populate(salsa20_matrix a, const void *key, size_t ksz)
 119 {
 120   const octet *k = key;
 121
 122   KSZ_ASSERT(salsa20, ksz);
 123
 124   /* Here's the pattern of key, constant, nonce, and counter pieces in the
 125    * matrix, before and after our permutation.
 126    *
 127    * [ C0 K0 K1 K2 ]       [ C0 C1 C2 C3 ]
 128    * [ K3 C1 N0 N1 ]  -->  [ K3 T1 K7 K2 ]
 129    * [ T0 T1 C2 K4 ]       [ T0 K6 K1 N1 ]
 130    * [ K5 K6 K7 C3 ]       [ K5 K0 N0 K4 ]
 131    */
 132
 133   a[13] = LOAD32_L(k +  0);
 134   a[10] = LOAD32_L(k +  4);
 135   if (ksz == 10) {
 136     a[ 7] = LOAD16_L(k +  8);
 137     a[ 4] = 0;
 138   } else {
 139     a[ 7] = LOAD32_L(k +  8);
 140     a[ 4] = LOAD32_L(k + 12);
 141   }
 142   if (ksz <= 16) {
 143     a[15] = a[13];
 144     a[12] = a[10];
 145     a[ 9] = a[ 7];
 146     a[ 6] = a[ 4];
 147     a[ 0] = SALSA20_A128;
 148     a[ 1] = SALSA20_B128;
 149     a[ 2] = ksz == 10 ? SALSA20_C80 : SALSA20_C128;
 150     a[ 3] = SALSA20_D128;
 151   } else {
 152     a[15] = LOAD32_L(k + 16);
 153     a[12] = LOAD32_L(k + 20);
 154     a[ 9] = LOAD32_L(k + 24);
 155     a[ 6] = LOAD32_L(k + 28);
 156     a[ 0] = SALSA20_A256;
 157     a[ 1] = SALSA20_B256;
 158     a[ 2] = SALSA20_C256;
 159     a[ 3] = SALSA20_D256;
 160   }
 161 }
 162
 163 /*----- Salsa20 implementation --------------------------------------------*/
 164
 165 /* --- @salsa20_init@ --- *
 166  *
 167  * Arguments:   @salsa20_ctx *ctx@ = context to fill in
 168  *              @const void *key@ = pointer to key material
 169  *              @size_t ksz@ = size of key (either 32 or 16)
 170  *              @const void *nonce@ = initial nonce, or null
 171  *
 172  * Returns:     ---
 173  *
 174  * Use:         Initializes a Salsa20 context ready for use.
 175  */
 176
 177 void salsa20_init(salsa20_ctx *ctx, const void *key, size_t ksz,
 178                   const void *nonce)
 179 {
 180   static const octet zerononce[SALSA20_NONCESZ];
 181
 182   populate(ctx->a, key, ksz);
 183   salsa20_setnonce(ctx, nonce ? nonce : zerononce);
 184 }
 185
 186 /* --- @salsa20_setnonce{,_ietf}@ --- *
 187  *
 188  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 189  *              @const void *nonce@ = the nonce (@SALSA20_NONCESZ@ or
 190  *                      @SALSA20_IETF_NONCESZ@ bytes)
 191  *
 192  * Returns:     ---
 193  *
 194  * Use:         Set a new nonce in the context @ctx@, e.g., for processing a
 195  *              different message.  The stream position is reset to zero (see
 196  *              @salsa20_seek@ etc.).
 197  */
 198
 199 void salsa20_setnonce(salsa20_ctx *ctx, const void *nonce)
 200 {
 201   const octet *n = nonce;
 202
 203   ctx->a[14] = LOAD32_L(n + 0);
 204   ctx->a[11] = LOAD32_L(n + 4);
 205   salsa20_seek(ctx, 0);
 206 }
 207
 208 void salsa20_setnonce_ietf(salsa20_ctx *ctx, const void *nonce)
 209 {
 210   const octet *n = nonce;
 211
 212   ctx->a[ 5] = LOAD32_L(n + 0);
 213   ctx->a[14] = LOAD32_L(n + 4);
 214   ctx->a[11] = LOAD32_L(n + 8);
 215   salsa20_seek_ietf(ctx, 0);
 216 }
 217
 218 /* --- @salsa20_seek{,u64,_ietf}@ --- *
 219  *
 220  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 221  *              @unsigned long i@, @kludge64 i@, @uint32@ = new position
 222  *
 223  * Returns:     ---
 224  *
 225  * Use:         Sets a new stream position, in units of Salsa20 output
 226  *              blocks, which are @SALSA20_OUTSZ@ bytes each.  Byte
 227  *              granularity can be achieved by calling @salsa20R_encrypt@
 228  *              appropriately.
 229  */
 230
 231 void salsa20_seek(salsa20_ctx *ctx, unsigned long i)
 232   { kludge64 ii; ASSIGN64(ii, i); salsa20_seeku64(ctx, ii); }
 233
 234 void salsa20_seeku64(salsa20_ctx *ctx, kludge64 i)
 235 {
 236   ctx->a[8] = LO64(i); ctx->a[5] = HI64(i);
 237   ctx->off = 0;
 238 }
 239
 240 void salsa20_seek_ietf(salsa20_ctx *ctx, uint32 i)
 241   { ctx->a[8] = i; }
 242
 243 /* --- @salsa20_tell{,u64,_ietf}@ --- *
 244  *
 245  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 246  *
 247  * Returns:     The current position in the output stream, in blocks,
 248  *              rounding upwards.
 249  */
 250
 251 unsigned long salsa20_tell(salsa20_ctx *ctx)
 252   { kludge64 i = salsa20_tellu64(ctx); return (GET64(unsigned long, i)); }
 253
 254 kludge64 salsa20_tellu64(salsa20_ctx *ctx)
 255   { kludge64 i; SET64(i, ctx->a[5], ctx->a[8]); return (i); }
 256
 257 uint32 salsa20_tell_ietf(salsa20_ctx *ctx)
 258   { return (ctx->a[5]); }
 259
 260 /* --- @salsa20{,12,8}_encrypt@ --- *
 261  *
 262  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 263  *              @const void *src@ = source buffer (or null)
 264  *              @void *dest@ = destination buffer (or null)
 265  *              @size_t sz@ = size of the buffers
 266  *
 267  * Returns:     ---
 268  *
 269  * Use:         Encrypts or decrypts @sz@ bytes of data from @src@ to @dest@.
 270  *              Salsa20 works by XORing plaintext with a keystream, so
 271  *              encryption and decryption are the same operation.  If @dest@
 272  *              is null then ignore @src@ and skip @sz@ bytes of the
 273  *              keystream.  If @src@ is null, then just write the keystream
 274  *              to @dest@.
 275  */
 276
 277 static const rsvr_policy policy = { 0, SALSA20_OUTSZ, SALSA20_OUTSZ };
 278
 279 #define SALSA20_ENCRYPT(r, ctx, src, dest, sz)                          \
 280   SALSA20_DECOR(salsa20, r, _encrypt)(ctx, src, dest, sz)
 281 #define DEFENCRYPT(r)                                                   \
 282   void SALSA20_ENCRYPT(r, salsa20_ctx *ctx, const void *src,            \
 283                        void *dest, size_t sz)                           \
 284   {                                                                     \
 285     salsa20_matrix b;                                                   \
 286     const octet *s = src;                                               \
 287     octet *d = dest;                                                    \
 288     rsvr_plan plan;                                                     \
 289     kludge64 pos, delta;                                                \
 290                                                                         \
 291     rsvr_mkplan(&plan, &policy, ctx->off, sz);                          \
 292                                                                         \
 293     if (plan.head) {                                                    \
 294       if (!ctx->off) {                                                  \
 295         core(r, ctx->a, b); SALSA20_STEP(ctx->a);                       \
 296         SALSA20_PREPBUF(ctx, b);                                        \
 297       }                                                                 \
 298       SALSA20_OUTBUF(ctx, d, s, plan.head);                             \
 299     }                                                                   \
 300                                                                         \
 301     ctx->off -= plan.from_rsvr;                                         \
 302                                                                         \
 303     if (!d) {                                                           \
 304       if (plan.from_input) {                                            \
 305         pos = salsa20_tellu64(ctx);                                     \
 306         ASSIGN64(delta, plan.from_input/SALSA20_OUTSZ);                 \
 307         ADD64(pos, pos, delta);                                         \
 308         salsa20_seeku64(ctx, pos);                                      \
 309       }                                                                 \
 310     } else if (!s) while (plan.from_input) {                            \
 311       core(r, ctx->a, b); SALSA20_STEP(ctx->a);                         \
 312       SALSA20_GENFULL(b, d); plan.from_input -= SALSA20_OUTSZ;          \
 313     } else while (plan.from_input) {                                    \
 314       core(r, ctx->a, b); SALSA20_STEP(ctx->a);                         \
 315       SALSA20_MIXFULL(b, d, s); plan.from_input -= SALSA20_OUTSZ;       \
 316     }                                                                   \
 317                                                                         \
 318     if (plan.tail) {                                                    \
 319       core(r, ctx->a, b); SALSA20_STEP(ctx->a);                         \
 320       SALSA20_PREPBUF(ctx, b);                                          \
 321       SALSA20_OUTBUF(ctx, d, s, plan.tail);                             \
 322     }                                                                   \
 323   }
 324 SALSA20_VARS(DEFENCRYPT)
 325
 326 /*----- HSalsa20 implementation -------------------------------------------*/
 327
 328 #define HSALSA20_RAW(r, ctx, src, dest)                                 \
 329   SALSA20_DECOR(hsalsa20, r, _raw)(ctx, src, dest)
 330 #define HSALSA20_PRF(r, ctx, src, dest)                                 \
 331   SALSA20_DECOR(hsalsa20, r, _prf)(ctx, src, dest)
 332
 333 /* --- @hsalsa20{,12,8}_prf@ --- *
 334  *
 335  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 336  *              @const void *src@ = the input (@HSALSA20_INSZ@ bytes)
 337  *              @void *dest@ = the output (@HSALSA20_OUTSZ@ bytes)
 338  *
 339  * Returns:     ---
 340  *
 341  * Use:         Apply the HSalsa20/r pseudorandom function to @src@, writing
 342  *              the result to @out@.
 343  */
 344
 345 #define DEFHSALSA20(r)                                                  \
 346   static void HSALSA20_RAW(r, salsa20_matrix k,                         \
 347                            const uint32 *src, uint32 *dest)             \
 348   {                                                                     \
 349     salsa20_matrix a;                                                   \
 350     int i;                                                              \
 351                                                                         \
 352     /* --- HSalsa20, computed from full Salsa20 --- *                   \
 353      *                                                                  \
 354      * The security proof makes use of the fact that HSalsa20 (i.e.,    \
 355      * without the final feedforward step) can be computed from full    \
 356      * Salsa20 using only knowledge of the non-secret input.  I don't   \
 357      * want to compromise the performance of the main function by       \
 358      * making the feedforward step separate, but this operation is less \
 359      * speed critical, so we do it the harder way.                      \
 360      */                                                                 \
 361                                                                         \
 362     for (i = 0; i < 4; i++) k[14 - 3*i] = src[i];                       \
 363     core(r, k, a);                                                      \
 364     for (i = 0; i < 4; i++) dest[i] = a[5*i] - k[i];                    \
 365     for (i = 4; i < 8; i++) dest[i] = a[i + 2] - k[26 - 3*i];           \
 366   }                                                                     \
 367                                                                         \
 368   void HSALSA20_PRF(r, salsa20_ctx *ctx, const void *src, void *dest)   \
 369   {                                                                     \
 370     const octet *s = src;                                               \
 371     octet *d = dest;                                                    \
 372     uint32 in[4], out[8];                                               \
 373     int i;                                                              \
 374                                                                         \
 375     for (i = 0; i < 4; i++) in[i] = LOAD32_L(s + 4*i);                  \
 376     HSALSA20_RAW(r, ctx->a, in, out);                                   \
 377     for (i = 0; i < 8; i++) STORE32_L(d + 4*i, out[i]);                 \
 378   }
 379 SALSA20_VARS(DEFHSALSA20)
 380
 381 /*----- XSalsa20 implementation -------------------------------------------*/
 382
 383 /* --- Some convenient macros for naming functions --- *
 384  *
 385  * Because the crypto core is involved in XSalsa20/r's per-nonce setup, we
 386  * need to take an interest in the number of rounds in most of the various
 387  * functions, and it will probably help if we distinguish the context
 388  * structures for the various versions.
 389  */
 390
 391 #define XSALSA20_CTX(r) SALSA20_DECOR(xsalsa20, r, _ctx)
 392 #define XSALSA20_INIT(r, ctx, k, ksz, n)                                \
 393   SALSA20_DECOR(xsalsa20, r, _init)(ctx, k, ksz, n)
 394 #define XSALSA20_SETNONCE(r, ctx, n)                                    \
 395   SALSA20_DECOR(xsalsa20, r, _setnonce)(ctx, n)
 396 #define XSALSA20_SEEK(r, ctx, i)                                        \
 397   SALSA20_DECOR(xsalsa20, r, _seek)(ctx, i)
 398 #define XSALSA20_SEEKU64(r, ctx, i)                                     \
 399   SALSA20_DECOR(xsalsa20, r, _seeku64)(ctx, i)
 400 #define XSALSA20_TELL(r, ctx)                                           \
 401   SALSA20_DECOR(xsalsa20, r, _tell)(ctx)
 402 #define XSALSA20_TELLU64(r, ctx)                                        \
 403   SALSA20_DECOR(xsalsa20, r, _tellu64)(ctx)
 404 #define XSALSA20_ENCRYPT(r, ctx, src, dest, sz)                         \
 405   SALSA20_DECOR(xsalsa20, r, _encrypt)(ctx, src, dest, sz)
 406
 407 /* --- @xsalsa20{,12,8}_init@ --- *
 408  *
 409  * Arguments:   @xsalsa20R_ctx *ctx@ = the context to fill in
 410  *              @const void *key@ = pointer to key material
 411  *              @size_t ksz@ = size of key (either 32 or 16)
 412  *              @const void *nonce@ = initial nonce, or null
 413  *
 414  * Returns:     ---
 415  *
 416  * Use:         Initializes an XSalsa20/r context ready for use.
 417  *
 418  *              There is a different function for each number of rounds,
 419  *              unlike for plain Salsa20.
 420  */
 421
 422 #define DEFXINIT(r)                                                     \
 423   void XSALSA20_INIT(r, XSALSA20_CTX(r) *ctx,                           \
 424                         const void *key, size_t ksz, const void *nonce) \
 425   {                                                                     \
 426     static const octet zerononce[XSALSA20_NONCESZ];                     \
 427                                                                         \
 428     populate(ctx->k, key, ksz);                                         \
 429     ctx->s.a[ 0] = SALSA20_A256;                                        \
 430     ctx->s.a[ 1] = SALSA20_B256;                                        \
 431     ctx->s.a[ 2] = SALSA20_C256;                                        \
 432     ctx->s.a[ 3] = SALSA20_D256;                                        \
 433     XSALSA20_SETNONCE(r, ctx, nonce ? nonce : zerononce);               \
 434   }
 435 SALSA20_VARS(DEFXINIT)
 436
 437 /* --- @xsalsa20{,12,8}_setnonce@ --- *
 438  *
 439  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 440  *              @const void *nonce@ = the nonce (@XSALSA20_NONCESZ@ bytes)
 441  *
 442  * Returns:     ---
 443  *
 444  * Use:         Set a new nonce in the context @ctx@, e.g., for processing a
 445  *              different message.  The stream position is reset to zero (see
 446  *              @salsa20_seek@ etc.).
 447  *
 448  *              There is a different function for each number of rounds,
 449  *              unlike for plain Salsa20.
 450  */
 451
 452 #define DEFXNONCE(r)                                                    \
 453   void XSALSA20_SETNONCE(r, XSALSA20_CTX(r) *ctx, const void *nonce)    \
 454   {                                                                     \
 455     const octet *n = nonce;                                             \
 456     uint32 in[4], out[8];                                               \
 457     int i;                                                              \
 458                                                                         \
 459     for (i = 0; i < 4; i++) in[i] = LOAD32_L(n + 4*i);                  \
 460     HSALSA20_RAW(r, ctx->k, in, out);                                   \
 461     for (i = 0; i < 4; i++) ctx->s.a[13 - 3*i] = out[i];                \
 462     for (i = 4; i < 8; i++) ctx->s.a[27 - 3*i] = out[i];                \
 463     salsa20_setnonce(&ctx->s, n + 16);                                  \
 464   }
 465 SALSA20_VARS(DEFXNONCE)
 466
 467 /* --- @xsalsa20{,12,8}_seek{,u64}@ --- *
 468  *
 469  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 470  *              @unsigned long i@, @kludge64 i@ = new position to set
 471  *
 472  * Returns:     ---
 473  *
 474  * Use:         Sets a new stream position, in units of Salsa20 output
 475  *              blocks, which are @XSALSA20_OUTSZ@ bytes each.  Byte
 476  *              granularity can be achieved by calling @xsalsa20R_encrypt@
 477  *              appropriately.
 478  *
 479  *              There is a different function for each number of rounds,
 480  *              unlike for plain Salsa20, because the context structures are
 481  *              different.
 482  */
 483
 484 /* --- @xsalsa20{,12,8}_tell{,u64}@ --- *
 485  *
 486  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 487  *
 488  * Returns:     The current position in the output stream, in blocks,
 489  *              rounding upwards.
 490  *
 491  *              There is a different function for each number of rounds,
 492  *              unlike for plain Salsa20, because the context structures are
 493  *              different.
 494  */
 495
 496 /* --- @xsalsa20{,12,8}_encrypt@ --- *
 497  *
 498  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 499  *              @const void *src@ = source buffer (or null)
 500  *              @void *dest@ = destination buffer (or null)
 501  *              @size_t sz@ = size of the buffers
 502  *
 503  * Returns:     ---
 504  *
 505  * Use:         Encrypts or decrypts @sz@ bytes of data from @src@ to @dest@.
 506  *              XSalsa20 works by XORing plaintext with a keystream, so
 507  *              encryption and decryption are the same operation.  If @dest@
 508  *              is null then ignore @src@ and skip @sz@ bytes of the
 509  *              keystream.  If @src@ is null, then just write the keystream
 510  *              to @dest@.
 511  */
 512
 513 #define DEFXPASSTHRU(r)                                                 \
 514   void XSALSA20_SEEK(r, XSALSA20_CTX(r) *ctx, unsigned long i)          \
 515     { salsa20_seek(&ctx->s, i); }                                       \
 516   void XSALSA20_SEEKU64(r, XSALSA20_CTX(r) *ctx, kludge64 i)            \
 517     { salsa20_seeku64(&ctx->s, i); }                                    \
 518   unsigned long XSALSA20_TELL(r, XSALSA20_CTX(r) *ctx)                  \
 519     { return salsa20_tell(&ctx->s); }                                   \
 520   kludge64 XSALSA20_TELLU64(r, XSALSA20_CTX(r) *ctx)                    \
 521     { return salsa20_tellu64(&ctx->s); }                                \
 522   void XSALSA20_ENCRYPT(r, XSALSA20_CTX(r) *ctx,                        \
 523                         const void *src, void *dest, size_t sz)         \
 524     { SALSA20_ENCRYPT(r, &ctx->s, src, dest, sz); }
 525 SALSA20_VARS(DEFXPASSTHRU)
 526
 527 /*----- Generic cipher interface ------------------------------------------*/
 528
 529 typedef struct gctx { gcipher c; salsa20_ctx ctx; } gctx;
 530
 531 static void gsetiv(gcipher *c, const void *iv)
 532   { gctx *g = (gctx *)c; salsa20_setnonce(&g->ctx, iv); }
 533
 534 static void gsetiv_ietf(gcipher *c, const void *iv)
 535   { gctx *g = (gctx *)c; salsa20_setnonce_ietf(&g->ctx, iv); }
 536
 537 static void gdestroy(gcipher *c)
 538   { gctx *g = (gctx *)c; BURN(*g); S_DESTROY(g); }
 539
 540 static gcipher *ginit(const void *k, size_t sz, const gcipher_ops *ops)
 541 {
 542   gctx *g = S_CREATE(gctx);
 543   g->c.ops = ops;
 544   salsa20_init(&g->ctx, k, sz, 0);
 545   return (&g->c);
 546 }
 547
 548 #define DEFGCIPHER(r)                                                   \
 549                                                                         \
 550   static const gcipher_ops gops_##r, gops_##r##_ietf;                   \
 551                                                                         \
 552   static gcipher *ginit_##r(const void *k, size_t sz)                   \
 553     { return (ginit(k, sz, &gops_##r)); }                               \
 554                                                                         \
 555   static gcipher *ginit_##r##_ietf(const void *k, size_t sz)            \
 556     { return (ginit(k, sz, &gops_##r##_ietf)); }                        \
 557                                                                         \
 558   static void gencrypt_##r(gcipher *c, const void *s,                   \
 559                            void *t, size_t sz)                          \
 560     { gctx *g = (gctx *)c; SALSA20_ENCRYPT(r, &g->ctx, s, t, sz); }     \
 561                                                                         \
 562   static const gcipher_ops gops_##r = {                                 \
 563     &SALSA20_DECOR(salsa20, r, ),                                       \
 564     gencrypt_##r, gencrypt_##r, gdestroy, gsetiv, 0                     \
 565   };                                                                    \
 566                                                                         \
 567   static const gcipher_ops gops_##r##_ietf = {                          \
 568     &SALSA20_DECOR(salsa20, r, _ietf),                                  \
 569     gencrypt_##r, gencrypt_##r, gdestroy, gsetiv_ietf, 0                \
 570   };                                                                    \
 571                                                                         \
 572   const gccipher SALSA20_DECOR(salsa20, r, ) = {                        \
 573     SALSA20_NAME_##r, salsa20_keysz,                                    \
 574     SALSA20_NONCESZ, ginit_##r                                          \
 575   };                                                                    \
 576                                                                         \
 577   const gccipher SALSA20_DECOR(salsa20, r, _ietf) = {                   \
 578     SALSA20_NAME_##r "-ietf", salsa20_keysz,                            \
 579     SALSA20_IETF_NONCESZ, ginit_##r##_ietf                              \
 580   };
 581
 582 SALSA20_VARS(DEFGCIPHER)
 583
 584 #define DEFGXCIPHER(r)                                                  \
 585                                                                         \
 586   typedef struct { gcipher c; XSALSA20_CTX(r) ctx; } gxctx_##r;         \
 587                                                                         \
 588   static void gxsetiv_##r(gcipher *c, const void *iv)                   \
 589     { gxctx_##r *g = (gxctx_##r *)c; XSALSA20_SETNONCE(r, &g->ctx, iv); } \
 590                                                                         \
 591   static void gxdestroy_##r(gcipher *c)                                 \
 592     { gxctx_##r *g = (gxctx_##r *)c; BURN(*g); S_DESTROY(g); }          \
 593                                                                         \
 594   static const gcipher_ops gxops_##r;                                   \
 595                                                                         \
 596   static gcipher *gxinit_##r(const void *k, size_t sz)                  \
 597   {                                                                     \
 598     gxctx_##r *g = S_CREATE(gxctx_##r);                                 \
 599     g->c.ops = &gxops_##r;                                              \
 600     XSALSA20_INIT(r, &g->ctx, k, sz, 0);                                \
 601     return (&g->c);                                                     \
 602   }                                                                     \
 603                                                                         \
 604   static void gxencrypt_##r(gcipher *c, const void *s,                  \
 605                             void *t, size_t sz)                         \
 606   {                                                                     \
 607     gxctx_##r *g = (gxctx_##r *)c;                                      \
 608     XSALSA20_ENCRYPT(r, &g->ctx, s, t, sz);                             \
 609   }                                                                     \
 610                                                                         \
 611   static const gcipher_ops gxops_##r = {                                \
 612     &SALSA20_DECOR(xsalsa20, r, ),                                      \
 613     gxencrypt_##r, gxencrypt_##r, gxdestroy_##r, gxsetiv_##r, 0         \
 614   };                                                                    \
 615                                                                         \
 616   const gccipher SALSA20_DECOR(xsalsa20, r, ) = {                       \
 617     "x" SALSA20_NAME_##r, salsa20_keysz,                                \
 618     XSALSA20_NONCESZ, gxinit_##r                                        \
 619   };
 620
 621 SALSA20_VARS(DEFGXCIPHER)
 622
 623 /*----- Generic random number generator interface -------------------------*/
 624
 625 typedef struct grops {
 626   size_t noncesz;
 627   void (*seek)(void *, kludge64);
 628   kludge64 (*tell)(void *);
 629   void (*setnonce)(void *, const void *);
 630   void (*generate)(void *, void *, size_t);
 631 } grops;
 632
 633 typedef struct grbasectx {
 634   grand r;
 635   const grops *ops;
 636 } grbasectx;
 637
 638 static int grmisc(grand *r, unsigned op, ...)
 639 {
 640   octet buf[XSALSA20_NONCESZ];
 641   grbasectx *g = (grbasectx *)r;
 642   grand *rr;
 643   const octet *p;
 644   size_t sz;
 645   uint32 i;
 646   unsigned long ul;
 647   kludge64 pos;
 648   va_list ap;
 649   int rc = 0;
 650
 651   va_start(ap, op);
 652
 653   switch (op) {
 654     case GRAND_CHECK:
 655       switch (va_arg(ap, unsigned)) {
 656         case GRAND_CHECK:
 657         case GRAND_SEEDINT:
 658         case GRAND_SEEDUINT32:
 659         case GRAND_SEEDBLOCK:
 660         case GRAND_SEEDRAND:
 661         case SALSA20_SEEK:
 662         case SALSA20_SEEKU64:
 663         case SALSA20_TELL:
 664         case SALSA20_TELLU64:
 665           rc = 1;
 666           break;
 667         default:
 668           rc = 0;
 669           break;
 670       }
 671       break;
 672
 673     case GRAND_SEEDINT:
 674       i = va_arg(ap, unsigned); STORE32_L(buf, i);
 675       memset(buf + 4, 0, g->ops->noncesz - 4);
 676       g->ops->setnonce(g, buf);
 677       break;
 678     case GRAND_SEEDUINT32:
 679       i = va_arg(ap, uint32); STORE32_L(buf, i);
 680       memset(buf + 4, 0, g->ops->noncesz - 4);
 681       g->ops->setnonce(g, buf);
 682       break;
 683     case GRAND_SEEDBLOCK:
 684       p = va_arg(ap, const void *);
 685       sz = va_arg(ap, size_t);
 686       if (sz < g->ops->noncesz) {
 687         memcpy(buf, p, sz);
 688         memset(buf + sz, 0, g->ops->noncesz - sz);
 689         p = buf;
 690       }
 691       g->ops->setnonce(g, p);
 692       break;
 693     case GRAND_SEEDRAND:
 694       rr = va_arg(ap, grand *);
 695       rr->ops->fill(rr, buf, g->ops->noncesz);
 696       g->ops->setnonce(g, buf);
 697       break;
 698     case SALSA20_SEEK:
 699       ul = va_arg(ap, unsigned long); ASSIGN64(pos, ul);
 700       g->ops->seek(g, pos);
 701       break;
 702     case SALSA20_SEEKU64:
 703       pos = va_arg(ap, kludge64);
 704       g->ops->seek(g, pos);
 705       break;
 706     case SALSA20_TELL:
 707       pos = g->ops->tell(g);
 708       *va_arg(ap, unsigned long *) = GET64(unsigned long, pos);
 709       break;
 710     case SALSA20_TELLU64:
 711       *va_arg(ap, kludge64 *) = g->ops->tell(g);
 712       break;
 713     default:
 714       GRAND_BADOP;
 715       break;
 716   }
 717
 718   return (rc);
 719 }
 720
 721 static octet grbyte(grand *r)
 722 {
 723   grbasectx *g = (grbasectx *)r;
 724   octet o;
 725   g->ops->generate(g, &o, 1);
 726   return (o);
 727 }
 728
 729 static uint32 grword(grand *r)
 730 {
 731   grbasectx *g = (grbasectx *)r;
 732   octet b[4];
 733   g->ops->generate(g, b, sizeof(b));
 734   return (LOAD32_L(b));
 735 }
 736
 737 static void grfill(grand *r, void *p, size_t sz)
 738 {
 739   grbasectx *g = (grbasectx *)r;
 740   g->ops->generate(r, p, sz);
 741 }
 742
 743 typedef struct grctx {
 744   grbasectx r;
 745   salsa20_ctx ctx;
 746 } grctx;
 747
 748 static void gr_seek(void *r, kludge64 pos)
 749   { grctx *g = r; salsa20_seeku64(&g->ctx, pos); }
 750
 751 static void gr_seek_ietf(void *r, kludge64 pos)
 752   { grctx *g = r; salsa20_seek_ietf(&g->ctx, LO64(pos)); }
 753
 754 static kludge64 gr_tell(void *r)
 755   { grctx *g = r; return (salsa20_tellu64(&g->ctx)); }
 756
 757 static kludge64 gr_tell_ietf(void *r)
 758 {
 759   grctx *g = r;
 760   kludge64 pos;
 761
 762   SET64(pos, 0, salsa20_tell_ietf(&g->ctx));
 763   return (pos);
 764 }
 765
 766 static void gr_setnonce(void *r, const void *n)
 767   { grctx *g = r; salsa20_setnonce(&g->ctx, n); }
 768
 769 static void gr_setnonce_ietf(void *r, const void *n)
 770   { grctx *g = r; salsa20_setnonce(&g->ctx, n); }
 771
 772 static void grdestroy(grand *r)
 773   { grctx *g = (grctx *)r; BURN(*g); S_DESTROY(g); }
 774
 775 static grand *grinit(const void *k, size_t ksz, const void *n,
 776                      const grand_ops *ops, const grops *myops)
 777 {
 778     grctx *g = S_CREATE(grctx);
 779     g->r.r.ops = ops;
 780     g->r.ops = myops;
 781     salsa20_init(&g->ctx, k, ksz, 0);
 782     if (n) myops->setnonce(g, n);
 783     return (&g->r.r);
 784 }
 785
 786 #define DEFGRAND(rr)                                                    \
 787                                                                         \
 788   static void gr_generate_##rr(void *r, void *b, size_t sz)             \
 789     { grctx *g = r; SALSA20_ENCRYPT(rr, &g->ctx, 0, b, sz); }           \
 790                                                                         \
 791   static const grops grops_##rr =                                       \
 792     { SALSA20_NONCESZ, gr_seek, gr_tell,                                \
 793       gr_setnonce, gr_generate_##rr };                                  \
 794                                                                         \
 795   static const grops grops_##rr##_ietf =                                \
 796     { SALSA20_IETF_NONCESZ, gr_seek_ietf, gr_tell_ietf,                 \
 797       gr_setnonce_ietf, gr_generate_##rr };                             \
 798                                                                         \
 799   static const grand_ops grops_rand_##rr = {                            \
 800     SALSA20_NAME_##rr, GRAND_CRYPTO, 0,                                 \
 801     grmisc, grdestroy, grword,                                          \
 802     grbyte, grword, grand_defaultrange, grfill                          \
 803   };                                                                    \
 804                                                                         \
 805   static const grand_ops grops_rand_##rr##_ietf = {                     \
 806     SALSA20_NAME_##rr "-ietf", GRAND_CRYPTO, 0,                         \
 807     grmisc, grdestroy, grword,                                          \
 808     grbyte, grword, grand_defaultrange, grfill                          \
 809   };                                                                    \
 810                                                                         \
 811   grand *SALSA20_DECOR(salsa20, rr, _rand)                              \
 812     (const void *k, size_t ksz, const void *n)                          \
 813     { return (grinit(k, ksz, n, &grops_rand_##rr, &grops_##rr)); }      \
 814                                                                         \
 815   grand *SALSA20_DECOR(salsa20, rr, _ietf_rand)                         \
 816     (const void *k, size_t ksz, const void *n)                          \
 817   {                                                                     \
 818     return (grinit(k, ksz, n,                                           \
 819                    &grops_rand_##rr##_ietf,                             \
 820                    &grops_##rr##_ietf));                                \
 821   }
 822
 823 SALSA20_VARS(DEFGRAND)
 824
 825 #define DEFXGRAND(rr)                                                   \
 826                                                                         \
 827   typedef struct grxctx_##rr {                                          \
 828     grbasectx r;                                                        \
 829     XSALSA20_CTX(rr) ctx;                                               \
 830   } grxctx_##rr;                                                        \
 831                                                                         \
 832   static void grx_seek_##rr(void *r, kludge64 pos)                      \
 833     { grxctx_##rr *g = r; XSALSA20_SEEKU64(rr, &g->ctx, pos); }         \
 834                                                                         \
 835   static kludge64 grx_tell_##rr(void *r)                                \
 836     { grxctx_##rr *g = r; return (XSALSA20_TELLU64(rr, &g->ctx)); }     \
 837                                                                         \
 838   static void grx_setnonce_##rr(void *r, const void *n)                 \
 839     { grxctx_##rr *g = r; XSALSA20_SETNONCE(rr, &g->ctx, n); }          \
 840                                                                         \
 841   static void grxdestroy_##rr(grand *r)                                 \
 842     { grxctx_##rr *g = (grxctx_##rr *)r; BURN(*g); S_DESTROY(g); }      \
 843                                                                         \
 844   static void grx_generate_##rr(void *r, void *b, size_t sz)            \
 845     { grxctx_##rr *g = r; XSALSA20_ENCRYPT(rr, &g->ctx, 0, b, sz); }    \
 846                                                                         \
 847   static const grops grxops_##rr =                                      \
 848   { XSALSA20_NONCESZ, grx_seek_##rr, grx_tell_##rr,                     \
 849       grx_setnonce_##rr, grx_generate_##rr };                           \
 850                                                                         \
 851   static const grand_ops grxops_rand_##rr = {                           \
 852     "x" SALSA20_NAME_##rr, GRAND_CRYPTO, 0,                             \
 853     grmisc, grxdestroy_##rr, grword,                                    \
 854     grbyte, grword, grand_defaultrange, grfill                          \
 855   };                                                                    \
 856                                                                         \
 857   grand *SALSA20_DECOR(xsalsa20, rr, _rand)                             \
 858     (const void *k, size_t ksz, const void *n)                          \
 859   {                                                                     \
 860     grxctx_##rr *g = S_CREATE(grxctx_##rr);                             \
 861     g->r.r.ops = &grxops_rand_##rr;                                     \
 862     g->r.ops = &grxops_##rr;                                            \
 863     XSALSA20_INIT(rr, &g->ctx, k, ksz, n);                              \
 864     return (&g->r.r);                                                   \
 865   }
 866 SALSA20_VARS(DEFXGRAND)
 867
 868 /*----- Test rig ----------------------------------------------------------*/
 869
 870 #ifdef TEST_RIG
 871
 872 #include <stdio.h>
 873 #include <string.h>
 874
 875 #include <mLib/quis.h>
 876 #include <mLib/testrig.h>
 877
 878 static const int perm[] = {
 879    0, 13, 10,  7,
 880    4,  1, 14, 11,
 881    8,  5,  2, 15,
 882   12,  9,  6,  3
 883 };
 884
 885 #define DEFVCORE(r)                                                     \
 886   static int v_core_##r(dstr *v)                                        \
 887   {                                                                     \
 888     salsa20_matrix a, b;                                                \
 889     dstr d = DSTR_INIT;                                                 \
 890     int i, j, n;                                                        \
 891     int ok = 1;                                                         \
 892                                                                         \
 893     DENSURE(&d, SALSA20_OUTSZ); d.len = SALSA20_OUTSZ;                  \
 894     n = *(int *)v[0].buf;                                               \
 895     for (i = 0; i < SALSA20_OUTSZ/4; i++)                               \
 896       b[i] = LOAD32_L(v[1].buf + 4*i);                                  \
 897     for (i = 0; i < n; i++) {                                           \
 898       for (j = 0; j < 16; j++) a[perm[j]] = b[j];                       \
 899       core(r, a, b);                                                    \
 900       memcpy(a, b, sizeof(a));                                          \
 901     }                                                                   \
 902     for (i = 0; i < SALSA20_OUTSZ/4; i++) STORE32_L(d.buf + 4*i, b[i]); \
 903                                                                         \
 904     if (d.len != v[2].len || memcmp(d.buf, v[2].buf, v[2].len) != 0) {  \
 905       ok = 0;                                                           \
 906       printf("\nfail core:"                                             \
 907              "\n\titerations = %d"                                      \
 908              "\n\tin       = ", n);                                     \
 909       type_hex.dump(&v[1], stdout);                                     \
 910       printf("\n\texpected   = ");                                      \
 911       type_hex.dump(&v[2], stdout);                                     \
 912       printf("\n\tcalculated = ");                                      \
 913       type_hex.dump(&d, stdout);                                        \
 914       putchar('\n');                                                    \
 915     }                                                                   \
 916                                                                         \
 917     dstr_destroy(&d);                                                   \
 918     return (ok);                                                        \
 919   }
 920 SALSA20_VARS(DEFVCORE)
 921
 922 #define SALSA20_CTX(r) salsa20_ctx
 923
 924 #define SALSA20_TESTSETUP(r, ctx, k, ksz, n, nsz, p, psz) do {          \
 925   kludge64 pos64;                                                       \
 926   salsa20_init(ctx, k, ksz, 0);                                         \
 927   if (nsz == 8) salsa20_setnonce(ctx, n);                               \
 928   else if (nsz == 12) salsa20_setnonce_ietf(ctx, n);                    \
 929   if (psz == 8) { LOAD64_(pos64, p); salsa20_seeku64(ctx, pos64); }     \
 930   else if (psz == 4) salsa20_seek_ietf(ctx, LOAD32(p));                 \
 931 } while (0)
 932
 933 #define XSALSA20_TESTSETUP(r, ctx, k, ksz, n, nsz, p, psz) do {         \
 934   kludge64 pos64;                                                       \
 935   XSALSA20_INIT(r, ctx, k, ksz, 0);                                     \
 936   if (nsz == 24) XSALSA20_SETNONCE(r, ctx, n);                          \
 937   if (psz == 8) { LOAD64_(pos64, p); XSALSA20_SEEKU64(r, ctx, pos64); } \
 938 } while (0)
 939
 940 #define DEFxVENC(base, BASE, r)                                         \
 941   static int v_encrypt_##base##_##r(dstr *v)                            \
 942   {                                                                     \
 943     BASE##_CTX(r) ctx;                                                  \
 944     dstr d = DSTR_INIT;                                                 \
 945     const octet *p, *p0;                                                \
 946     octet *q;                                                           \
 947     size_t sz, sz0, step;                                               \
 948     unsigned long skip;                                                 \
 949     int ok = 1;                                                         \
 950                                                                         \
 951     if (v[4].len) { p0 = (const octet *)v[4].buf; sz0 = v[4].len; }     \
 952     else { p0 = 0; sz0 = v[5].len; }                                    \
 953     DENSURE(&d, sz0); d.len = sz0;                                      \
 954     skip = *(unsigned long *)v[3].buf;                                  \
 955                                                                         \
 956     step = 0;                                                           \
 957     while (step < sz0 + skip) {                                         \
 958       step = step ? 3*step + 4 : 1;                                     \
 959       if (step > sz0 + skip) step = sz0 + skip;                         \
 960       BASE##_TESTSETUP(r, &ctx, v[0].buf, v[0].len,                     \
 961                        v[1].buf, v[1].len, v[2].buf, v[2].len);         \
 962                                                                         \
 963       for (sz = skip; sz >= step; sz -= step)                           \
 964         BASE##_ENCRYPT(r, &ctx, 0, 0, step);                            \
 965       if (sz) BASE##_ENCRYPT(r, &ctx, 0, 0, sz);                        \
 966       for (p = p0, q = (octet *)d.buf, sz = sz0;                        \
 967            sz >= step;                                                  \
 968            sz -= step, q += step) {                                     \
 969         BASE##_ENCRYPT(r, &ctx, p, q, step);                            \
 970         if (p) p += step;                                               \
 971       }                                                                 \
 972       if (sz) BASE##_ENCRYPT(r, &ctx, p, q, sz);                        \
 973                                                                         \
 974       if (d.len != v[5].len || memcmp(d.buf, v[5].buf, v[5].len) != 0) { \
 975         ok = 0;                                                         \
 976         printf("\nfail encrypt:"                                        \
 977                "\n\tstep           = %lu"                               \
 978                "\n\tkey    = ", (unsigned long)step);                   \
 979         type_hex.dump(&v[0], stdout);                                   \
 980         printf("\n\tnonce          = ");                                \
 981         type_hex.dump(&v[1], stdout);                                   \
 982         printf("\n\tposition   = ");                                    \
 983         type_hex.dump(&v[2], stdout);                                   \
 984         printf("\n\tskip           = %lu", skip);                       \
 985         printf("\n\tmessage    = ");                                    \
 986         type_hex.dump(&v[4], stdout);                                   \
 987         printf("\n\texpected   = ");                                    \
 988         type_hex.dump(&v[5], stdout);                                   \
 989         printf("\n\tcalculated = ");                                    \
 990         type_hex.dump(&d, stdout);                                      \
 991         putchar('\n');                                                  \
 992       }                                                                 \
 993     }                                                                   \
 994                                                                         \
 995     dstr_destroy(&d);                                                   \
 996     return (ok);                                                        \
 997   }
 998 #define DEFVENC(r) DEFxVENC(salsa20, SALSA20, r)
 999 #define DEFXVENC(r) DEFxVENC(xsalsa20, XSALSA20, r)
1000 SALSA20_VARS(DEFVENC)
1001 SALSA20_VARS(DEFXVENC)
1002
1003 static test_chunk defs[] = {
1004 #define DEFxTAB(pre, base, r)                                           \
1005   { pre SALSA20_NAME_##r, v_encrypt_##base##_##r,                       \
1006     { &type_hex, &type_hex, &type_hex, &type_ulong,                     \
1007       &type_hex, &type_hex, 0 } },
1008 #define DEFTAB(r)                                                       \
1009   { SALSA20_NAME_##r "-core", v_core_##r,                               \
1010     { &type_int, &type_hex, &type_hex, 0 } },                           \
1011   DEFxTAB("", salsa20, r)
1012 #define DEFXTAB(r) DEFxTAB("x", xsalsa20, r)
1013 SALSA20_VARS(DEFTAB)
1014 SALSA20_VARS(DEFXTAB)
1015   { 0, 0, { 0 } }
1016 };
1017
1018 int main(int argc, char *argv[])
1019 {
1020   test_run(argc, argv, defs, SRCDIR"/t/salsa20");
1021   return (0);
1022 }
1023
1024 #endif
1025
1026 /*----- That's all, folks -------------------------------------------------*/