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,
 101                      cpu_feature_p(CPUFEAT_ARM_NEON));
 102 #endif
 103   DISPATCH_PICK_FALLBACK(salsa20_core, simple_core);
 104 }
 105
 106 /* --- @populate@ --- *
 107  *
 108  * Arguments:   @salsa20_matrix a@ = a matrix to fill in
 109  *              @const void *key@ = pointer to key material
 110  *              @size_t ksz@ = size of key
 111  *
 112  * Returns:     ---
 113  *
 114  * Use:         Fills in a Salsa20 matrix from the key, setting the
 115  *              appropriate constants according to the key length.  The nonce
 116  *              and position words are left uninitialized.
 117  */
 118
 119 static void populate(salsa20_matrix a, const void *key, size_t ksz)
 120 {
 121   const octet *k = key;
 122
 123   KSZ_ASSERT(salsa20, ksz);
 124
 125   /* Here's the pattern of key, constant, nonce, and counter pieces in the
 126    * matrix, before and after our permutation.
 127    *
 128    * [ C0 K0 K1 K2 ]       [ C0 C1 C2 C3 ]
 129    * [ K3 C1 N0 N1 ]  -->  [ K3 T1 K7 K2 ]
 130    * [ T0 T1 C2 K4 ]       [ T0 K6 K1 N1 ]
 131    * [ K5 K6 K7 C3 ]       [ K5 K0 N0 K4 ]
 132    */
 133
 134   a[13] = LOAD32_L(k +  0);
 135   a[10] = LOAD32_L(k +  4);
 136   if (ksz == 10) {
 137     a[ 7] = LOAD16_L(k +  8);
 138     a[ 4] = 0;
 139   } else {
 140     a[ 7] = LOAD32_L(k +  8);
 141     a[ 4] = LOAD32_L(k + 12);
 142   }
 143   if (ksz <= 16) {
 144     a[15] = a[13];
 145     a[12] = a[10];
 146     a[ 9] = a[ 7];
 147     a[ 6] = a[ 4];
 148     a[ 0] = SALSA20_A128;
 149     a[ 1] = SALSA20_B128;
 150     a[ 2] = ksz == 10 ? SALSA20_C80 : SALSA20_C128;
 151     a[ 3] = SALSA20_D128;
 152   } else {
 153     a[15] = LOAD32_L(k + 16);
 154     a[12] = LOAD32_L(k + 20);
 155     a[ 9] = LOAD32_L(k + 24);
 156     a[ 6] = LOAD32_L(k + 28);
 157     a[ 0] = SALSA20_A256;
 158     a[ 1] = SALSA20_B256;
 159     a[ 2] = SALSA20_C256;
 160     a[ 3] = SALSA20_D256;
 161   }
 162 }
 163
 164 /*----- Salsa20 implementation --------------------------------------------*/
 165
 166 static const octet zerononce[XSALSA20_NONCESZ];
 167
 168 /* --- @salsa20_init@ --- *
 169  *
 170  * Arguments:   @salsa20_ctx *ctx@ = context to fill in
 171  *              @const void *key@ = pointer to key material
 172  *              @size_t ksz@ = size of key (either 32 or 16)
 173  *              @const void *nonce@ = initial nonce, or null
 174  *
 175  * Returns:     ---
 176  *
 177  * Use:         Initializes a Salsa20 context ready for use.
 178  */
 179
 180 void salsa20_init(salsa20_ctx *ctx, const void *key, size_t ksz,
 181                   const void *nonce)
 182 {
 183   populate(ctx->a, key, ksz);
 184   salsa20_setnonce(ctx, nonce ? nonce : zerononce);
 185 }
 186
 187 /* --- @salsa20_setnonce{,_ietf}@ --- *
 188  *
 189  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 190  *              @const void *nonce@ = the nonce (@SALSA20_NONCESZ@ or
 191  *                      @SALSA20_IETF_NONCESZ@ bytes)
 192  *
 193  * Returns:     ---
 194  *
 195  * Use:         Set a new nonce in the context @ctx@, e.g., for processing a
 196  *              different message.  The stream position is reset to zero (see
 197  *              @salsa20_seek@ etc.).
 198  */
 199
 200 void salsa20_setnonce(salsa20_ctx *ctx, const void *nonce)
 201 {
 202   const octet *n = nonce;
 203
 204   ctx->a[14] = LOAD32_L(n + 0);
 205   ctx->a[11] = LOAD32_L(n + 4);
 206   salsa20_seek(ctx, 0);
 207 }
 208
 209 void salsa20_setnonce_ietf(salsa20_ctx *ctx, const void *nonce)
 210 {
 211   const octet *n = nonce;
 212
 213   ctx->a[ 5] = LOAD32_L(n + 0);
 214   ctx->a[14] = LOAD32_L(n + 4);
 215   ctx->a[11] = LOAD32_L(n + 8);
 216   salsa20_seek_ietf(ctx, 0);
 217 }
 218
 219 /* --- @salsa20_seek{,u64,_ietf}@ --- *
 220  *
 221  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 222  *              @unsigned long i@, @kludge64 i@, @uint32@ = new position
 223  *
 224  * Returns:     ---
 225  *
 226  * Use:         Sets a new stream position, in units of Salsa20 output
 227  *              blocks, which are @SALSA20_OUTSZ@ bytes each.  Byte
 228  *              granularity can be achieved by calling @salsa20R_encrypt@
 229  *              appropriately.
 230  */
 231
 232 void salsa20_seek(salsa20_ctx *ctx, unsigned long i)
 233   { kludge64 ii; ASSIGN64(ii, i); salsa20_seeku64(ctx, ii); }
 234
 235 void salsa20_seeku64(salsa20_ctx *ctx, kludge64 i)
 236 {
 237   ctx->a[8] = LO64(i); ctx->a[5] = HI64(i);
 238   ctx->off = 0;
 239 }
 240
 241 void salsa20_seek_ietf(salsa20_ctx *ctx, uint32 i)
 242   { ctx->a[8] = i; }
 243
 244 /* --- @salsa20_tell{,u64,_ietf}@ --- *
 245  *
 246  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 247  *
 248  * Returns:     The current position in the output stream, in blocks,
 249  *              rounding upwards.
 250  */
 251
 252 unsigned long salsa20_tell(salsa20_ctx *ctx)
 253   { kludge64 i = salsa20_tellu64(ctx); return (GET64(unsigned long, i)); }
 254
 255 kludge64 salsa20_tellu64(salsa20_ctx *ctx)
 256   { kludge64 i; SET64(i, ctx->a[5], ctx->a[8]); return (i); }
 257
 258 uint32 salsa20_tell_ietf(salsa20_ctx *ctx)
 259   { return (ctx->a[5]); }
 260
 261 /* --- @salsa20{,12,8}_encrypt@ --- *
 262  *
 263  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 264  *              @const void *src@ = source buffer (or null)
 265  *              @void *dest@ = destination buffer (or null)
 266  *              @size_t sz@ = size of the buffers
 267  *
 268  * Returns:     ---
 269  *
 270  * Use:         Encrypts or decrypts @sz@ bytes of data from @src@ to @dest@.
 271  *              Salsa20 works by XORing plaintext with a keystream, so
 272  *              encryption and decryption are the same operation.  If @dest@
 273  *              is null then ignore @src@ and skip @sz@ bytes of the
 274  *              keystream.  If @src@ is null, then just write the keystream
 275  *              to @dest@.
 276  */
 277
 278 static const rsvr_policy policy = { 0, SALSA20_OUTSZ, SALSA20_OUTSZ };
 279
 280 #define SALSA20_ENCRYPT(r, ctx, src, dest, sz)                          \
 281   SALSA20_DECOR(salsa20, r, _encrypt)(ctx, src, dest, sz)
 282 #define DEFENCRYPT(r)                                                   \
 283   void SALSA20_ENCRYPT(r, salsa20_ctx *ctx, const void *src,            \
 284                        void *dest, size_t sz)                           \
 285   {                                                                     \
 286     salsa20_matrix b;                                                   \
 287     const octet *s = src;                                               \
 288     octet *d = dest;                                                    \
 289     rsvr_plan plan;                                                     \
 290     kludge64 pos, delta;                                                \
 291                                                                         \
 292     rsvr_mkplan(&plan, &policy, ctx->off, sz);                          \
 293                                                                         \
 294     if (plan.head) {                                                    \
 295       if (!ctx->off) {                                                  \
 296         core(r, ctx->a, b); SALSA20_STEP(ctx->a);                       \
 297         SALSA20_PREPBUF(ctx, b);                                        \
 298       }                                                                 \
 299       SALSA20_OUTBUF(ctx, d, s, plan.head);                             \
 300     }                                                                   \
 301                                                                         \
 302     ctx->off -= plan.from_rsvr;                                         \
 303                                                                         \
 304     if (!d) {                                                           \
 305       if (plan.from_input) {                                            \
 306         pos = salsa20_tellu64(ctx);                                     \
 307         ASSIGN64(delta, plan.from_input/SALSA20_OUTSZ);                 \
 308         ADD64(pos, pos, delta);                                         \
 309         salsa20_seeku64(ctx, pos);                                      \
 310       }                                                                 \
 311     } else if (!s) while (plan.from_input) {                            \
 312       core(r, ctx->a, b); SALSA20_STEP(ctx->a);                         \
 313       SALSA20_GENFULL(b, d); plan.from_input -= SALSA20_OUTSZ;          \
 314     } else while (plan.from_input) {                                    \
 315       core(r, ctx->a, b); SALSA20_STEP(ctx->a);                         \
 316       SALSA20_MIXFULL(b, d, s); plan.from_input -= SALSA20_OUTSZ;       \
 317     }                                                                   \
 318                                                                         \
 319     if (plan.tail) {                                                    \
 320       core(r, ctx->a, b); SALSA20_STEP(ctx->a);                         \
 321       SALSA20_PREPBUF(ctx, b);                                          \
 322       SALSA20_OUTBUF(ctx, d, s, plan.tail);                             \
 323     }                                                                   \
 324   }
 325 SALSA20_VARS(DEFENCRYPT)
 326
 327 /*----- HSalsa20 implementation -------------------------------------------*/
 328
 329 #define HSALSA20_RAW(r, ctx, src, dest)                                 \
 330   SALSA20_DECOR(hsalsa20, r, _raw)(ctx, src, dest)
 331 #define HSALSA20_PRF(r, ctx, src, dest)                                 \
 332   SALSA20_DECOR(hsalsa20, r, _prf)(ctx, src, dest)
 333
 334 /* --- @hsalsa20{,12,8}_prf@ --- *
 335  *
 336  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 337  *              @const void *src@ = the input (@HSALSA20_INSZ@ bytes)
 338  *              @void *dest@ = the output (@HSALSA20_OUTSZ@ bytes)
 339  *
 340  * Returns:     ---
 341  *
 342  * Use:         Apply the HSalsa20/r pseudorandom function to @src@, writing
 343  *              the result to @out@.
 344  */
 345
 346 #define DEFHSALSA20(r)                                                  \
 347   static void HSALSA20_RAW(r, salsa20_matrix k,                         \
 348                            const uint32 *src, uint32 *dest)             \
 349   {                                                                     \
 350     salsa20_matrix a;                                                   \
 351     int i;                                                              \
 352                                                                         \
 353     /* --- HSalsa20, computed from full Salsa20 --- *                   \
 354      *                                                                  \
 355      * The security proof makes use of the fact that HSalsa20 (i.e.,    \
 356      * without the final feedforward step) can be computed from full    \
 357      * Salsa20 using only knowledge of the non-secret input.  I don't   \
 358      * want to compromise the performance of the main function by       \
 359      * making the feedforward step separate, but this operation is less \
 360      * speed critical, so we do it the harder way.                      \
 361      */                                                                 \
 362                                                                         \
 363     for (i = 0; i < 4; i++) k[14 - 3*i] = src[i];                       \
 364     core(r, k, a);                                                      \
 365     for (i = 0; i < 4; i++) dest[i] = a[5*i] - k[i];                    \
 366     for (i = 4; i < 8; i++) dest[i] = a[i + 2] - k[26 - 3*i];           \
 367   }                                                                     \
 368                                                                         \
 369   void HSALSA20_PRF(r, salsa20_ctx *ctx, const void *src, void *dest)   \
 370   {                                                                     \
 371     const octet *s = src;                                               \
 372     octet *d = dest;                                                    \
 373     uint32 in[4], out[8];                                               \
 374     int i;                                                              \
 375                                                                         \
 376     for (i = 0; i < 4; i++) in[i] = LOAD32_L(s + 4*i);                  \
 377     HSALSA20_RAW(r, ctx->a, in, out);                                   \
 378     for (i = 0; i < 8; i++) STORE32_L(d + 4*i, out[i]);                 \
 379   }
 380 SALSA20_VARS(DEFHSALSA20)
 381
 382 /*----- XSalsa20 implementation -------------------------------------------*/
 383
 384 /* --- Some convenient macros for naming functions --- *
 385  *
 386  * Because the crypto core is involved in XSalsa20/r's per-nonce setup, we
 387  * need to take an interest in the number of rounds in most of the various
 388  * functions, and it will probably help if we distinguish the context
 389  * structures for the various versions.
 390  */
 391
 392 #define XSALSA20_CTX(r) SALSA20_DECOR(xsalsa20, r, _ctx)
 393 #define XSALSA20_INIT(r, ctx, k, ksz, n)                                \
 394   SALSA20_DECOR(xsalsa20, r, _init)(ctx, k, ksz, n)
 395 #define XSALSA20_SETNONCE(r, ctx, n)                                    \
 396   SALSA20_DECOR(xsalsa20, r, _setnonce)(ctx, n)
 397 #define XSALSA20_SEEK(r, ctx, i)                                        \
 398   SALSA20_DECOR(xsalsa20, r, _seek)(ctx, i)
 399 #define XSALSA20_SEEKU64(r, ctx, i)                                     \
 400   SALSA20_DECOR(xsalsa20, r, _seeku64)(ctx, i)
 401 #define XSALSA20_TELL(r, ctx)                                           \
 402   SALSA20_DECOR(xsalsa20, r, _tell)(ctx)
 403 #define XSALSA20_TELLU64(r, ctx)                                        \
 404   SALSA20_DECOR(xsalsa20, r, _tellu64)(ctx)
 405 #define XSALSA20_ENCRYPT(r, ctx, src, dest, sz)                         \
 406   SALSA20_DECOR(xsalsa20, r, _encrypt)(ctx, src, dest, sz)
 407
 408 /* --- @xsalsa20{,12,8}_init@ --- *
 409  *
 410  * Arguments:   @xsalsa20R_ctx *ctx@ = the context to fill in
 411  *              @const void *key@ = pointer to key material
 412  *              @size_t ksz@ = size of key (either 32 or 16)
 413  *              @const void *nonce@ = initial nonce, or null
 414  *
 415  * Returns:     ---
 416  *
 417  * Use:         Initializes an XSalsa20/r context ready for use.
 418  *
 419  *              There is a different function for each number of rounds,
 420  *              unlike for plain Salsa20.
 421  */
 422
 423 #define DEFXINIT(r)                                                     \
 424   void XSALSA20_INIT(r, XSALSA20_CTX(r) *ctx,                           \
 425                         const void *key, size_t ksz, const void *nonce) \
 426   {                                                                     \
 427     populate(ctx->k, key, ksz);                                         \
 428     ctx->s.a[ 0] = SALSA20_A256;                                        \
 429     ctx->s.a[ 1] = SALSA20_B256;                                        \
 430     ctx->s.a[ 2] = SALSA20_C256;                                        \
 431     ctx->s.a[ 3] = SALSA20_D256;                                        \
 432     XSALSA20_SETNONCE(r, ctx, nonce ? nonce : zerononce);               \
 433   }
 434 SALSA20_VARS(DEFXINIT)
 435
 436 /* --- @xsalsa20{,12,8}_setnonce@ --- *
 437  *
 438  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 439  *              @const void *nonce@ = the nonce (@XSALSA20_NONCESZ@ bytes)
 440  *
 441  * Returns:     ---
 442  *
 443  * Use:         Set a new nonce in the context @ctx@, e.g., for processing a
 444  *              different message.  The stream position is reset to zero (see
 445  *              @salsa20_seek@ etc.).
 446  *
 447  *              There is a different function for each number of rounds,
 448  *              unlike for plain Salsa20.
 449  */
 450
 451 #define DEFXNONCE(r)                                                    \
 452   void XSALSA20_SETNONCE(r, XSALSA20_CTX(r) *ctx, const void *nonce)    \
 453   {                                                                     \
 454     const octet *n = nonce;                                             \
 455     uint32 in[4], out[8];                                               \
 456     int i;                                                              \
 457                                                                         \
 458     for (i = 0; i < 4; i++) in[i] = LOAD32_L(n + 4*i);                  \
 459     HSALSA20_RAW(r, ctx->k, in, out);                                   \
 460     for (i = 0; i < 4; i++) ctx->s.a[13 - 3*i] = out[i];                \
 461     for (i = 4; i < 8; i++) ctx->s.a[27 - 3*i] = out[i];                \
 462     salsa20_setnonce(&ctx->s, n + 16);                                  \
 463   }
 464 SALSA20_VARS(DEFXNONCE)
 465
 466 /* --- @xsalsa20{,12,8}_seek{,u64}@ --- *
 467  *
 468  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 469  *              @unsigned long i@, @kludge64 i@ = new position to set
 470  *
 471  * Returns:     ---
 472  *
 473  * Use:         Sets a new stream position, in units of Salsa20 output
 474  *              blocks, which are @XSALSA20_OUTSZ@ bytes each.  Byte
 475  *              granularity can be achieved by calling @xsalsa20R_encrypt@
 476  *              appropriately.
 477  *
 478  *              There is a different function for each number of rounds,
 479  *              unlike for plain Salsa20, because the context structures are
 480  *              different.
 481  */
 482
 483 /* --- @xsalsa20{,12,8}_tell{,u64}@ --- *
 484  *
 485  * Arguments:   @salsa20_ctx *ctx@ = pointer to context
 486  *
 487  * Returns:     The current position in the output stream, in blocks,
 488  *              rounding upwards.
 489  *
 490  *              There is a different function for each number of rounds,
 491  *              unlike for plain Salsa20, because the context structures are
 492  *              different.
 493  */
 494
 495 /* --- @xsalsa20{,12,8}_encrypt@ --- *
 496  *
 497  * Arguments:   @xsalsa20R_ctx *ctx@ = pointer to context
 498  *              @const void *src@ = source buffer (or null)
 499  *              @void *dest@ = destination buffer (or null)
 500  *              @size_t sz@ = size of the buffers
 501  *
 502  * Returns:     ---
 503  *
 504  * Use:         Encrypts or decrypts @sz@ bytes of data from @src@ to @dest@.
 505  *              XSalsa20 works by XORing plaintext with a keystream, so
 506  *              encryption and decryption are the same operation.  If @dest@
 507  *              is null then ignore @src@ and skip @sz@ bytes of the
 508  *              keystream.  If @src@ is null, then just write the keystream
 509  *              to @dest@.
 510  */
 511
 512 #define DEFXPASSTHRU(r)                                                 \
 513   void XSALSA20_SEEK(r, XSALSA20_CTX(r) *ctx, unsigned long i)          \
 514     { salsa20_seek(&ctx->s, i); }                                       \
 515   void XSALSA20_SEEKU64(r, XSALSA20_CTX(r) *ctx, kludge64 i)            \
 516     { salsa20_seeku64(&ctx->s, i); }                                    \
 517   unsigned long XSALSA20_TELL(r, XSALSA20_CTX(r) *ctx)                  \
 518     { return salsa20_tell(&ctx->s); }                                   \
 519   kludge64 XSALSA20_TELLU64(r, XSALSA20_CTX(r) *ctx)                    \
 520     { return salsa20_tellu64(&ctx->s); }                                \
 521   void XSALSA20_ENCRYPT(r, XSALSA20_CTX(r) *ctx,                        \
 522                         const void *src, void *dest, size_t sz)         \
 523     { SALSA20_ENCRYPT(r, &ctx->s, src, dest, sz); }
 524 SALSA20_VARS(DEFXPASSTHRU)
 525
 526 /*----- Generic cipher interface ------------------------------------------*/
 527
 528 typedef struct gctx { gcipher c; salsa20_ctx ctx; } gctx;
 529
 530 static void gsetiv(gcipher *c, const void *iv)
 531   { gctx *g = (gctx *)c; salsa20_setnonce(&g->ctx, iv); }
 532
 533 static void gsetiv_ietf(gcipher *c, const void *iv)
 534   { gctx *g = (gctx *)c; salsa20_setnonce_ietf(&g->ctx, iv); }
 535
 536 static void gdestroy(gcipher *c)
 537   { gctx *g = (gctx *)c; BURN(*g); S_DESTROY(g); }
 538
 539 static gcipher *ginit(const void *k, size_t sz, const gcipher_ops *ops)
 540 {
 541   gctx *g = S_CREATE(gctx);
 542   g->c.ops = ops;
 543   salsa20_init(&g->ctx, k, sz, 0);
 544   return (&g->c);
 545 }
 546
 547 #define DEFGCIPHER(r)                                                   \
 548                                                                         \
 549   static const gcipher_ops gops_##r, gops_##r##_ietf;                   \
 550                                                                         \
 551   static gcipher *ginit_##r(const void *k, size_t sz)                   \
 552     { return (ginit(k, sz, &gops_##r)); }                               \
 553                                                                         \
 554   static gcipher *ginit_##r##_ietf(const void *k, size_t sz)            \
 555     { return (ginit(k, sz, &gops_##r##_ietf)); }                        \
 556                                                                         \
 557   static void gencrypt_##r(gcipher *c, const void *s,                   \
 558                            void *t, size_t sz)                          \
 559     { gctx *g = (gctx *)c; SALSA20_ENCRYPT(r, &g->ctx, s, t, sz); }     \
 560                                                                         \
 561   static const gcipher_ops gops_##r = {                                 \
 562     &SALSA20_DECOR(salsa20, r, ),                                       \
 563     gencrypt_##r, gencrypt_##r, gdestroy, gsetiv, 0                     \
 564   };                                                                    \
 565                                                                         \
 566   static const gcipher_ops gops_##r##_ietf = {                          \
 567     &SALSA20_DECOR(salsa20, r, _ietf),                                  \
 568     gencrypt_##r, gencrypt_##r, gdestroy, gsetiv_ietf, 0                \
 569   };                                                                    \
 570                                                                         \
 571   const gccipher SALSA20_DECOR(salsa20, r, ) = {                        \
 572     SALSA20_NAME_##r, salsa20_keysz,                                    \
 573     SALSA20_NONCESZ, ginit_##r                                          \
 574   };                                                                    \
 575                                                                         \
 576   const gccipher SALSA20_DECOR(salsa20, r, _ietf) = {                   \
 577     SALSA20_NAME_##r "-ietf", salsa20_keysz,                            \
 578     SALSA20_IETF_NONCESZ, ginit_##r##_ietf                              \
 579   };
 580
 581 SALSA20_VARS(DEFGCIPHER)
 582
 583 #define DEFGXCIPHER(r)                                                  \
 584                                                                         \
 585   typedef struct { gcipher c; XSALSA20_CTX(r) ctx; } gxctx_##r;         \
 586                                                                         \
 587   static void gxsetiv_##r(gcipher *c, const void *iv)                   \
 588     { gxctx_##r *g = (gxctx_##r *)c; XSALSA20_SETNONCE(r, &g->ctx, iv); } \
 589                                                                         \
 590   static void gxdestroy_##r(gcipher *c)                                 \
 591     { gxctx_##r *g = (gxctx_##r *)c; BURN(*g); S_DESTROY(g); }          \
 592                                                                         \
 593   static const gcipher_ops gxops_##r;                                   \
 594                                                                         \
 595   static gcipher *gxinit_##r(const void *k, size_t sz)                  \
 596   {                                                                     \
 597     gxctx_##r *g = S_CREATE(gxctx_##r);                                 \
 598     g->c.ops = &gxops_##r;                                              \
 599     XSALSA20_INIT(r, &g->ctx, k, sz, 0);                                \
 600     return (&g->c);                                                     \
 601   }                                                                     \
 602                                                                         \
 603   static void gxencrypt_##r(gcipher *c, const void *s,                  \
 604                             void *t, size_t sz)                         \
 605   {                                                                     \
 606     gxctx_##r *g = (gxctx_##r *)c;                                      \
 607     XSALSA20_ENCRYPT(r, &g->ctx, s, t, sz);                             \
 608   }                                                                     \
 609                                                                         \
 610   static const gcipher_ops gxops_##r = {                                \
 611     &SALSA20_DECOR(xsalsa20, r, ),                                      \
 612     gxencrypt_##r, gxencrypt_##r, gxdestroy_##r, gxsetiv_##r, 0         \
 613   };                                                                    \
 614                                                                         \
 615   const gccipher SALSA20_DECOR(xsalsa20, r, ) = {                       \
 616     "x" SALSA20_NAME_##r, salsa20_keysz,                                \
 617     XSALSA20_NONCESZ, gxinit_##r                                        \
 618   };
 619
 620 SALSA20_VARS(DEFGXCIPHER)
 621
 622 /*----- Generic random number generator interface -------------------------*/
 623
 624 typedef struct grops {
 625   size_t noncesz;
 626   void (*seek)(void *, kludge64);
 627   kludge64 (*tell)(void *);
 628   void (*setnonce)(void *, const void *);
 629   void (*generate)(void *, void *, size_t);
 630 } grops;
 631
 632 typedef struct grbasectx {
 633   grand r;
 634   const grops *ops;
 635 } grbasectx;
 636
 637 static int grmisc(grand *r, unsigned op, ...)
 638 {
 639   octet buf[XSALSA20_NONCESZ];
 640   grbasectx *g = (grbasectx *)r;
 641   grand *rr;
 642   const octet *p;
 643   size_t sz;
 644   uint32 i;
 645   unsigned long ul;
 646   kludge64 pos;
 647   va_list ap;
 648   int rc = 0;
 649
 650   va_start(ap, op);
 651
 652   switch (op) {
 653     case GRAND_CHECK:
 654       switch (va_arg(ap, unsigned)) {
 655         case GRAND_CHECK:
 656         case GRAND_SEEDINT:
 657         case GRAND_SEEDUINT32:
 658         case GRAND_SEEDBLOCK:
 659         case GRAND_SEEDRAND:
 660         case SALSA20_SEEK:
 661         case SALSA20_SEEKU64:
 662         case SALSA20_TELL:
 663         case SALSA20_TELLU64:
 664           rc = 1;
 665           break;
 666         default:
 667           rc = 0;
 668           break;
 669       }
 670       break;
 671
 672     case GRAND_SEEDINT:
 673       i = va_arg(ap, unsigned); STORE32_L(buf, i);
 674       memset(buf + 4, 0, g->ops->noncesz - 4);
 675       g->ops->setnonce(g, buf);
 676       break;
 677     case GRAND_SEEDUINT32:
 678       i = va_arg(ap, uint32); STORE32_L(buf, i);
 679       memset(buf + 4, 0, g->ops->noncesz - 4);
 680       g->ops->setnonce(g, buf);
 681       break;
 682     case GRAND_SEEDBLOCK:
 683       p = va_arg(ap, const void *);
 684       sz = va_arg(ap, size_t);
 685       if (sz < g->ops->noncesz) {
 686         memcpy(buf, p, sz);
 687         memset(buf + sz, 0, g->ops->noncesz - sz);
 688         p = buf;
 689       }
 690       g->ops->setnonce(g, p);
 691       break;
 692     case GRAND_SEEDRAND:
 693       rr = va_arg(ap, grand *);
 694       rr->ops->fill(rr, buf, g->ops->noncesz);
 695       g->ops->setnonce(g, buf);
 696       break;
 697     case SALSA20_SEEK:
 698       ul = va_arg(ap, unsigned long); ASSIGN64(pos, ul);
 699       g->ops->seek(g, pos);
 700       break;
 701     case SALSA20_SEEKU64:
 702       pos = va_arg(ap, kludge64);
 703       g->ops->seek(g, pos);
 704       break;
 705     case SALSA20_TELL:
 706       pos = g->ops->tell(g);
 707       *va_arg(ap, unsigned long *) = GET64(unsigned long, pos);
 708       break;
 709     case SALSA20_TELLU64:
 710       *va_arg(ap, kludge64 *) = g->ops->tell(g);
 711       break;
 712     default:
 713       GRAND_BADOP;
 714       break;
 715   }
 716
 717   return (rc);
 718 }
 719
 720 static octet grbyte(grand *r)
 721 {
 722   grbasectx *g = (grbasectx *)r;
 723   octet o;
 724   g->ops->generate(g, &o, 1);
 725   return (o);
 726 }
 727
 728 static uint32 grword(grand *r)
 729 {
 730   grbasectx *g = (grbasectx *)r;
 731   octet b[4];
 732   g->ops->generate(g, b, sizeof(b));
 733   return (LOAD32_L(b));
 734 }
 735
 736 static void grfill(grand *r, void *p, size_t sz)
 737 {
 738   grbasectx *g = (grbasectx *)r;
 739   g->ops->generate(r, p, sz);
 740 }
 741
 742 typedef struct grctx {
 743   grbasectx r;
 744   salsa20_ctx ctx;
 745 } grctx;
 746
 747 static void gr_seek(void *r, kludge64 pos)
 748   { grctx *g = r; salsa20_seeku64(&g->ctx, pos); }
 749
 750 static void gr_seek_ietf(void *r, kludge64 pos)
 751   { grctx *g = r; salsa20_seek_ietf(&g->ctx, LO64(pos)); }
 752
 753 static kludge64 gr_tell(void *r)
 754   { grctx *g = r; return (salsa20_tellu64(&g->ctx)); }
 755
 756 static kludge64 gr_tell_ietf(void *r)
 757 {
 758   grctx *g = r;
 759   kludge64 pos;
 760
 761   SET64(pos, 0, salsa20_tell_ietf(&g->ctx));
 762   return (pos);
 763 }
 764
 765 static void gr_setnonce(void *r, const void *n)
 766   { grctx *g = r; salsa20_setnonce(&g->ctx, n); }
 767
 768 static void gr_setnonce_ietf(void *r, const void *n)
 769   { grctx *g = r; salsa20_setnonce(&g->ctx, n); }
 770
 771 static void grdestroy(grand *r)
 772   { grctx *g = (grctx *)r; BURN(*g); S_DESTROY(g); }
 773
 774 static grand *grinit(const void *k, size_t ksz, const void *n,
 775                      const grand_ops *ops, const grops *myops)
 776 {
 777     grctx *g = S_CREATE(grctx);
 778     g->r.r.ops = ops;
 779     g->r.ops = myops;
 780     salsa20_init(&g->ctx, k, ksz, 0);
 781     if (n) myops->setnonce(g, n);
 782     return (&g->r.r);
 783 }
 784
 785 #define DEFGRAND(rr)                                                    \
 786                                                                         \
 787   static void gr_generate_##rr(void *r, void *b, size_t sz)             \
 788     { grctx *g = r; SALSA20_ENCRYPT(rr, &g->ctx, 0, b, sz); }           \
 789                                                                         \
 790   static const grops grops_##rr =                                       \
 791     { SALSA20_NONCESZ, gr_seek, gr_tell,                                \
 792       gr_setnonce, gr_generate_##rr };                                  \
 793                                                                         \
 794   static const grops grops_##rr##_ietf =                                \
 795     { SALSA20_IETF_NONCESZ, gr_seek_ietf, gr_tell_ietf,                 \
 796       gr_setnonce_ietf, gr_generate_##rr };                             \
 797                                                                         \
 798   static const grand_ops grops_rand_##rr = {                            \
 799     SALSA20_NAME_##rr, GRAND_CRYPTO, 0,                                 \
 800     grmisc, grdestroy, grword,                                          \
 801     grbyte, grword, grand_defaultrange, grfill                          \
 802   };                                                                    \
 803                                                                         \
 804   static const grand_ops grops_rand_##rr##_ietf = {                     \
 805     SALSA20_NAME_##rr "-ietf", GRAND_CRYPTO, 0,                         \
 806     grmisc, grdestroy, grword,                                          \
 807     grbyte, grword, grand_defaultrange, grfill                          \
 808   };                                                                    \
 809                                                                         \
 810   grand *SALSA20_DECOR(salsa20, rr, _rand)                              \
 811     (const void *k, size_t ksz, const void *n)                          \
 812     { return (grinit(k, ksz, n, &grops_rand_##rr, &grops_##rr)); }      \
 813                                                                         \
 814   grand *SALSA20_DECOR(salsa20, rr, _ietf_rand)                         \
 815     (const void *k, size_t ksz, const void *n)                          \
 816   {                                                                     \
 817     return (grinit(k, ksz, n,                                           \
 818                    &grops_rand_##rr##_ietf,                             \
 819                    &grops_##rr##_ietf));                                \
 820   }
 821
 822 SALSA20_VARS(DEFGRAND)
 823
 824 #define DEFXGRAND(rr)                                                   \
 825                                                                         \
 826   typedef struct grxctx_##rr {                                          \
 827     grbasectx r;                                                        \
 828     XSALSA20_CTX(rr) ctx;                                               \
 829   } grxctx_##rr;                                                        \
 830                                                                         \
 831   static void grx_seek_##rr(void *r, kludge64 pos)                      \
 832     { grxctx_##rr *g = r; XSALSA20_SEEKU64(rr, &g->ctx, pos); }         \
 833                                                                         \
 834   static kludge64 grx_tell_##rr(void *r)                                \
 835     { grxctx_##rr *g = r; return (XSALSA20_TELLU64(rr, &g->ctx)); }     \
 836                                                                         \
 837   static void grx_setnonce_##rr(void *r, const void *n)                 \
 838     { grxctx_##rr *g = r; XSALSA20_SETNONCE(rr, &g->ctx, n); }          \
 839                                                                         \
 840   static void grxdestroy_##rr(grand *r)                                 \
 841     { grxctx_##rr *g = (grxctx_##rr *)r; BURN(*g); S_DESTROY(g); }      \
 842                                                                         \
 843   static void grx_generate_##rr(void *r, void *b, size_t sz)            \
 844     { grxctx_##rr *g = r; XSALSA20_ENCRYPT(rr, &g->ctx, 0, b, sz); }    \
 845                                                                         \
 846   static const grops grxops_##rr =                                      \
 847   { XSALSA20_NONCESZ, grx_seek_##rr, grx_tell_##rr,                     \
 848       grx_setnonce_##rr, grx_generate_##rr };                           \
 849                                                                         \
 850   static const grand_ops grxops_rand_##rr = {                           \
 851     "x" SALSA20_NAME_##rr, GRAND_CRYPTO, 0,                             \
 852     grmisc, grxdestroy_##rr, grword,                                    \
 853     grbyte, grword, grand_defaultrange, grfill                          \
 854   };                                                                    \
 855                                                                         \
 856   grand *SALSA20_DECOR(xsalsa20, rr, _rand)                             \
 857     (const void *k, size_t ksz, const void *n)                          \
 858   {                                                                     \
 859     grxctx_##rr *g = S_CREATE(grxctx_##rr);                             \
 860     g->r.r.ops = &grxops_rand_##rr;                                     \
 861     g->r.ops = &grxops_##rr;                                            \
 862     XSALSA20_INIT(rr, &g->ctx, k, ksz, n);                              \
 863     return (&g->r.r);                                                   \
 864   }
 865 SALSA20_VARS(DEFXGRAND)
 866
 867 /*----- Test rig ----------------------------------------------------------*/
 868
 869 #ifdef TEST_RIG
 870
 871 #include <stdio.h>
 872 #include <string.h>
 873
 874 #include <mLib/macros.h>
 875 #include <mLib/quis.h>
 876 #include <mLib/testrig.h>
 877
 878 #ifdef ENABLE_ASM_DEBUG
 879 #  include "regdump.h"
 880 #endif
 881
 882 static const int perm[] = {
 883    0, 13, 10,  7,
 884    4,  1, 14, 11,
 885    8,  5,  2, 15,
 886   12,  9,  6,  3
 887 };
 888
 889 #define DEFVCORE(r)                                                     \
 890   static int v_core_##r(dstr *v)                                        \
 891   {                                                                     \
 892     salsa20_matrix a, b;                                                \
 893     dstr d = DSTR_INIT;                                                 \
 894     int i, j, n;                                                        \
 895     int ok = 1;                                                         \
 896                                                                         \
 897     DENSURE(&d, SALSA20_OUTSZ); d.len = SALSA20_OUTSZ;                  \
 898     n = *(int *)v[0].buf;                                               \
 899     for (i = 0; i < SALSA20_OUTSZ/4; i++)                               \
 900       b[i] = LOAD32_L(v[1].buf + 4*i);                                  \
 901     for (i = 0; i < n; i++) {                                           \
 902       for (j = 0; j < 16; j++) a[perm[j]] = b[j];                       \
 903       core(r, a, b);                                                    \
 904       memcpy(a, b, sizeof(a));                                          \
 905     }                                                                   \
 906     for (i = 0; i < SALSA20_OUTSZ/4; i++) STORE32_L(d.buf + 4*i, b[i]); \
 907                                                                         \
 908     if (d.len != v[2].len || MEMCMP(d.buf, !=, v[2].buf, v[2].len)) {   \
 909       ok = 0;                                                           \
 910       printf("\nfail core:"                                             \
 911              "\n\titerations = %d"                                      \
 912              "\n\tin       = ", n);                                     \
 913       type_hex.dump(&v[1], stdout);                                     \
 914       printf("\n\texpected   = ");                                      \
 915       type_hex.dump(&v[2], stdout);                                     \
 916       printf("\n\tcalculated = ");                                      \
 917       type_hex.dump(&d, stdout);                                        \
 918       putchar('\n');                                                    \
 919     }                                                                   \
 920                                                                         \
 921     dstr_destroy(&d);                                                   \
 922     return (ok);                                                        \
 923   }
 924 SALSA20_VARS(DEFVCORE)
 925
 926 #define SALSA20_CTX(r) salsa20_ctx
 927
 928 #define SALSA20_TESTSETUP(r, ctx, k, ksz, n, nsz, p, psz) do {          \
 929   kludge64 pos64;                                                       \
 930   salsa20_init(ctx, k, ksz, 0);                                         \
 931   if (nsz == 8) salsa20_setnonce(ctx, n);                               \
 932   else if (nsz == 12) salsa20_setnonce_ietf(ctx, n);                    \
 933   if (psz == 8) { LOAD64_(pos64, p); salsa20_seeku64(ctx, pos64); }     \
 934   else if (psz == 4) salsa20_seek_ietf(ctx, LOAD32(p));                 \
 935 } while (0)
 936
 937 #define XSALSA20_TESTSETUP(r, ctx, k, ksz, n, nsz, p, psz) do {         \
 938   kludge64 pos64;                                                       \
 939   XSALSA20_INIT(r, ctx, k, ksz, 0);                                     \
 940   if (nsz == 24) XSALSA20_SETNONCE(r, ctx, n);                          \
 941   if (psz == 8) { LOAD64_(pos64, p); XSALSA20_SEEKU64(r, ctx, pos64); } \
 942 } while (0)
 943
 944 #define DEFxVENC(base, BASE, r)                                         \
 945   static int v_encrypt_##base##_##r(dstr *v)                            \
 946   {                                                                     \
 947     BASE##_CTX(r) ctx;                                                  \
 948     dstr d = DSTR_INIT;                                                 \
 949     const octet *p, *p0;                                                \
 950     octet *q;                                                           \
 951     size_t sz, sz0, step;                                               \
 952     unsigned long skip;                                                 \
 953     int ok = 1;                                                         \
 954                                                                         \
 955     if (v[4].len) { p0 = (const octet *)v[4].buf; sz0 = v[4].len; }     \
 956     else { p0 = 0; sz0 = v[5].len; }                                    \
 957     DENSURE(&d, sz0); d.len = sz0;                                      \
 958     skip = *(unsigned long *)v[3].buf;                                  \
 959                                                                         \
 960     step = 0;                                                           \
 961     while (step < sz0 + skip) {                                         \
 962       step = step ? 3*step + 4 : 1;                                     \
 963       if (step > sz0 + skip) step = sz0 + skip;                         \
 964       BASE##_TESTSETUP(r, &ctx, v[0].buf, v[0].len,                     \
 965                        v[1].buf, v[1].len, v[2].buf, v[2].len);         \
 966                                                                         \
 967       for (sz = skip; sz >= step; sz -= step)                           \
 968         BASE##_ENCRYPT(r, &ctx, 0, 0, step);                            \
 969       if (sz) BASE##_ENCRYPT(r, &ctx, 0, 0, sz);                        \
 970       for (p = p0, q = (octet *)d.buf, sz = sz0;                        \
 971            sz >= step;                                                  \
 972            sz -= step, q += step) {                                     \
 973         BASE##_ENCRYPT(r, &ctx, p, q, step);                            \
 974         if (p) p += step;                                               \
 975       }                                                                 \
 976       if (sz) BASE##_ENCRYPT(r, &ctx, p, q, sz);                        \
 977                                                                         \
 978       if (d.len != v[5].len || MEMCMP(d.buf, !=, v[5].buf, v[5].len)) { \
 979         ok = 0;                                                         \
 980         printf("\nfail encrypt:"                                        \
 981                "\n\tstep           = %lu"                               \
 982                "\n\tkey    = ", (unsigned long)step);                   \
 983         type_hex.dump(&v[0], stdout);                                   \
 984         printf("\n\tnonce          = ");                                \
 985         type_hex.dump(&v[1], stdout);                                   \
 986         printf("\n\tposition   = ");                                    \
 987         type_hex.dump(&v[2], stdout);                                   \
 988         printf("\n\tskip           = %lu", skip);                       \
 989         printf("\n\tmessage    = ");                                    \
 990         type_hex.dump(&v[4], stdout);                                   \
 991         printf("\n\texpected   = ");                                    \
 992         type_hex.dump(&v[5], stdout);                                   \
 993         printf("\n\tcalculated = ");                                    \
 994         type_hex.dump(&d, stdout);                                      \
 995         putchar('\n');                                                  \
 996       }                                                                 \
 997     }                                                                   \
 998                                                                         \
 999     dstr_destroy(&d);                                                   \
1000     return (ok);                                                        \
1001   }
1002 #define DEFVENC(r) DEFxVENC(salsa20, SALSA20, r)
1003 #define DEFXVENC(r) DEFxVENC(xsalsa20, XSALSA20, r)
1004 SALSA20_VARS(DEFVENC)
1005 SALSA20_VARS(DEFXVENC)
1006
1007 static test_chunk defs[] = {
1008 #define DEFxTAB(pre, base, r)                                           \
1009   { pre SALSA20_NAME_##r, v_encrypt_##base##_##r,                       \
1010     { &type_hex, &type_hex, &type_hex, &type_ulong,                     \
1011       &type_hex, &type_hex, 0 } },
1012 #define DEFTAB(r)                                                       \
1013   { SALSA20_NAME_##r "-core", v_core_##r,                               \
1014     { &type_int, &type_hex, &type_hex, 0 } },                           \
1015   DEFxTAB("", salsa20, r)
1016 #define DEFXTAB(r) DEFxTAB("x", xsalsa20, r)
1017 SALSA20_VARS(DEFTAB)
1018 SALSA20_VARS(DEFXTAB)
1019   { 0, 0, { 0 } }
1020 };
1021
1022 int main(int argc, char *argv[])
1023 {
1024 #ifdef ENABLE_ASM_DEBUG
1025   regdump_init();
1026 #endif
1027   test_run(argc, argv, defs, SRCDIR"/t/salsa20");
1028   return (0);
1029 }
1030
1031 #endif
1032
1033 /*----- That's all, folks -------------------------------------------------*/