cast5/cast5.c

   1 /* cast5.c */
   2 /*
   3     This file is part of the AVR-Crypto-Lib.
   4     Copyright (C) 2008  Daniel Otte (daniel.otte@rub.de)
   5
   6     This program is free software: you can redistribute it and/or modify
   7     it under the terms of the GNU General Public License as published by
   8     the Free Software Foundation, either version 3 of the License, or
   9     (at your option) any later version.
  10
  11     This program is distributed in the hope that it will be useful,
  12     but WITHOUT ANY WARRANTY; without even the implied warranty of
  13     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14     GNU General Public License for more details.
  15
  16     You should have received a copy of the GNU General Public License
  17     along with this program.  If not, see <http://www.gnu.org/licenses/>.
  18 */
  19 /*
  20  * \file        cast5.c
  21  * \author      Daniel Otte
  22  * \email       daniel.otte@rub.de
  23  * \date        2006-07-26
  24  * \par License:
  25  *  GPLv3 or later
  26  * \brief Implementation of the CAST5 (aka CAST-128) cipher algorithm as described in RFC 2144
  27  *
  28  */
  29
  30  #include <stdint.h>
  31  #include <string.h>
  32  #include "cast5.h"
  33
  34  #undef DEBUG
  35
  36  #ifdef DEBUG
  37   #include "cli.h"
  38  #endif
  39
  40 #include "cast5-sbox.h"
  41
  42
  43
  44 #define S5(x) pgm_read_dword(&s5[(x)])
  45 #define S6(x) pgm_read_dword(&s6[(x)])
  46 #define S7(x) pgm_read_dword(&s7[(x)])
  47 #define S8(x) pgm_read_dword(&s8[(x)])
  48
  49 static
  50 void cast5_init_A(uint8_t *dest, uint8_t *src, bool bmode){
  51         uint8_t mask = bmode?0x8:0;
  52         *((uint32_t*)(&dest[0x0])) = *((uint32_t*)(&src[0x0^mask]))
  53                                      ^ S5(src[0xD^mask]) ^ S6(src[0xF^mask])
  54                                      ^ S7(src[0xC^mask]) ^ S8(src[0xE^mask])
  55                                      ^ S7(src[0x8^mask]);
  56         *((uint32_t*)(&dest[0x4])) = *((uint32_t*)(&src[0x8^mask]))
  57                                      ^ S5(dest[0x0]) ^ S6(dest[0x2])
  58                                      ^ S7(dest[0x1]) ^ S8(dest[0x3])
  59                                      ^ S8(src[0xA^mask]);
  60         *((uint32_t*)(&dest[0x8])) = *((uint32_t*)(&src[0xC^mask]))
  61                                      ^ S5(dest[0x7]) ^ S6(dest[0x6])
  62                                      ^ S7(dest[0x5]) ^ S8(dest[0x4])
  63                                      ^ S5(src[0x9^mask]);
  64         *((uint32_t*)(&dest[0xC])) = *((uint32_t*)(&src[0x4^mask]))
  65                                      ^ S5(dest[0xA])
  66                                      ^ S6(dest[0x9])
  67                                      ^ S7(dest[0xB])
  68                                      ^ S8(dest[0x8])
  69                                      ^ S6(src[0xB^mask]);
  70 }
  71
  72 static
  73 void cast5_init_M(uint8_t *dest, uint8_t *src, bool nmode, bool xmode){
  74         uint8_t nmt[] = {0xB, 0xA, 0x9, 0x8,
  75                          0xF, 0xE, 0xD, 0xC,
  76                          0x3, 0x2, 0x1, 0x0,
  77                          0x7, 0x6, 0x5, 0x4}; /* nmode table */
  78         uint8_t xmt[4][4] = {{0x2, 0x6, 0x9, 0xC},
  79                              {0x8, 0xD, 0x3, 0x7},
  80                              {0x3, 0x7, 0x8, 0xD},
  81                              {0x9, 0xC, 0x2, 0x6}};
  82         #define NMT(x) (src[nmode?nmt[(x)]:(x)])
  83         #define XMT(x) (src[xmt[(xmode<<1) + nmode][(x)]])
  84         *((uint32_t*)(&dest[0x0])) = S5(NMT(0x8)) ^ S6(NMT(0x9)) ^ S7(NMT(0x7)) ^ S8(NMT(0x6)) ^ S5(XMT(0));
  85         *((uint32_t*)(&dest[0x4])) = S5(NMT(0xA)) ^ S6(NMT(0xB)) ^ S7(NMT(0x5)) ^ S8(NMT(0x4)) ^ S6(XMT(1));
  86         *((uint32_t*)(&dest[0x8])) = S5(NMT(0xC)) ^ S6(NMT(0xD)) ^ S7(NMT(0x3)) ^ S8(NMT(0x2)) ^ S7(XMT(2));
  87         *((uint32_t*)(&dest[0xC])) = S5(NMT(0xE)) ^ S6(NMT(0xF)) ^ S7(NMT(0x1)) ^ S8(NMT(0x0)) ^ S8(XMT(3));
  88 }
  89
  90 #define S5B(x) pgm_read_byte(3+(uint8_t*)(&s5[(x)]))
  91 #define S6B(x) pgm_read_byte(3+(uint8_t*)(&s6[(x)]))
  92 #define S7B(x) pgm_read_byte(3+(uint8_t*)(&s7[(x)]))
  93 #define S8B(x) pgm_read_byte(3+(uint8_t*)(&s8[(x)]))
  94
  95 static
  96 void cast5_init_rM(uint8_t *klo, uint8_t *khi, uint8_t offset, uint8_t *src, bool nmode, bool xmode){
  97         uint8_t nmt[] = {0xB, 0xA, 0x9, 0x8, 0xF, 0xE, 0xD, 0xC, 0x3, 0x2, 0x1, 0x0, 0x7, 0x6, 0x5, 0x4}; /* nmode table */
  98         uint8_t xmt[4][4] = {{0x2, 0x6, 0x9, 0xC}, {0x8, 0xD, 0x3, 0x7}, {0x3, 0x7, 0x8, 0xD}, {0x9, 0xC, 0x2, 0x6}};
  99         uint8_t t, h=0;
 100         t = S5B(NMT(0x8)) ^ S6B(NMT(0x9)) ^ S7B(NMT(0x7)) ^ S8B(NMT(0x6)) ^ S5B(XMT(0));
 101                 klo[offset*2] |= (t & 0x0f);
 102                 h |= (t&0x10); h>>=1;
 103         t = S5B(NMT(0xA)) ^ S6B(NMT(0xB)) ^ S7B(NMT(0x5)) ^ S8B(NMT(0x4)) ^ S6B(XMT(1));
 104                 klo[offset*2] |= (t<<4) & 0xf0;
 105                 h |= t&0x10; h>>=1;
 106         t = S5B(NMT(0xC)) ^ S6B(NMT(0xD)) ^ S7B(NMT(0x3)) ^ S8B(NMT(0x2)) ^ S7B(XMT(2));
 107                 klo[offset*2+1] |= t&0xf;
 108                 h |= t&0x10; h>>=1;
 109         t = S5B(NMT(0xE)) ^ S6B(NMT(0xF)) ^ S7B(NMT(0x1)) ^ S8B(NMT(0x0)) ^ S8B(XMT(3));
 110                 klo[offset*2+1] |= t<<4;
 111                 h |= t&0x10; h >>=1;
 112         #ifdef DEBUG
 113                 cli_putstr("\r\n\t h="); cli_hexdump(&h,1);
 114         #endif
 115         khi[offset>>1] |= h<<((offset&0x1)?4:0);
 116 }
 117
 118 #define S_5X(s) pgm_read_dword(&s5[BPX[(s)]])
 119 #define S_6X(s) pgm_read_dword(&s6[BPX[(s)]])
 120 #define S_7X(s) pgm_read_dword(&s7[BPX[(s)]])
 121 #define S_8X(s) pgm_read_dword(&s8[BPX[(s)]])
 122
 123 #define S_5Z(s) pgm_read_dword(&s5[BPZ[(s)]])
 124 #define S_6Z(s) pgm_read_dword(&s6[BPZ[(s)]])
 125 #define S_7Z(s) pgm_read_dword(&s7[BPZ[(s)]])
 126 #define S_8Z(s) pgm_read_dword(&s8[BPZ[(s)]])
 127
 128
 129
 130
 131 void cast5_init(const void* key, uint16_t keylength_b, cast5_ctx_t* s){
 132          /* we migth return if the key is valid and if setup was successful */
 133         uint32_t x[4], z[4];
 134         #define BPX ((uint8_t*)&(x[0]))
 135         #define BPZ ((uint8_t*)&(z[0]))
 136         s->shortkey = (keylength_b<=80);
 137         /* littel endian only! */
 138         memset(&(x[0]), 0 ,16); /* set x to zero */
 139         if(keylength_b > 128)
 140                 keylength_b=128;
 141         memcpy(&(x[0]), key, (keylength_b+7)/8);
 142
 143
 144         /* todo: merge a and b and compress the whole stuff */
 145         /***** A *****/
 146         cast5_init_A((uint8_t*)(&z[0]), (uint8_t*)(&x[0]), false);
 147         /***** M *****/
 148         cast5_init_M((uint8_t*)(&(s->mask[0])), (uint8_t*)(&z[0]), false, false);
 149         /***** B *****/
 150         cast5_init_A((uint8_t*)(&x[0]), (uint8_t*)(&z[0]), true);
 151         /***** N *****/
 152         cast5_init_M((uint8_t*)(&(s->mask[4])), (uint8_t*)(&x[0]), true, false);
 153         /***** A *****/
 154         cast5_init_A((uint8_t*)(&z[0]), (uint8_t*)(&x[0]), false);
 155         /***** N' *****/
 156         cast5_init_M((uint8_t*)(&(s->mask[8])), (uint8_t*)(&z[0]), true, true);
 157         /***** B *****/
 158         cast5_init_A((uint8_t*)(&x[0]), (uint8_t*)(&z[0]), true);
 159         /***** M' *****/
 160         cast5_init_M((uint8_t*)(&(s->mask[12])), (uint8_t*)(&x[0]), false, true);
 161
 162         /* that were the masking keys, now the rotation keys */
 163         /* set the keys to zero */
 164         memset(&(s->rotl[0]),0,8);
 165         s->roth[0]=s->roth[1]=0;
 166         /***** A *****/
 167         cast5_init_A((uint8_t*)(&z[0]), (uint8_t*)(&x[0]), false);
 168         /***** M *****/
 169         cast5_init_rM(&(s->rotl[0]), &(s->roth[0]), 0, (uint8_t*)(&z[0]), false, false);
 170         /***** B *****/
 171         cast5_init_A((uint8_t*)(&x[0]), (uint8_t*)(&z[0]), true);
 172         /***** N *****/
 173         cast5_init_rM(&(s->rotl[0]), &(s->roth[0]), 1, (uint8_t*)(&x[0]), true, false);
 174         /***** A *****/
 175         cast5_init_A((uint8_t*)(&z[0]), (uint8_t*)(&x[0]), false);
 176         /***** N' *****/
 177         cast5_init_rM(&(s->rotl[0]), &(s->roth[0]), 2, (uint8_t*)(&z[0]), true, true);
 178         /***** B *****/
 179         cast5_init_A((uint8_t*)(&x[0]), (uint8_t*)(&z[0]), true);
 180         /***** M' *****/
 181         cast5_init_rM(&(s->rotl[0]), &(s->roth[0]), 3, (uint8_t*)(&x[0]), false, true);
 182         /* done ;-) */
 183 }
 184
 185
 186
 187 /********************************************************************************************************/
 188
 189 #define ROTL32(a,n) ((a)<<(n) | (a)>>(32-(n)))
 190 #define CHANGE_ENDIAN32(x) ((x)<<24 | (x)>>24 | ((x)&0xff00)<<8 | ((x)&0xff0000)>>8 )
 191
 192 typedef uint32_t cast5_f_t(uint32_t,uint32_t,uint8_t);
 193
 194 #define IA 3
 195 #define IB 2
 196 #define IC 1
 197 #define ID 0
 198
 199 static
 200 uint32_t cast5_f1(uint32_t d, uint32_t m, uint8_t r){
 201         uint32_t t;
 202         t = ROTL32((d + m),r);
 203 #ifdef DEBUG
 204         uint32_t ia,ib,ic,id;
 205         cli_putstr("\r\n f1("); cli_hexdump(&d, 4); cli_putc(',');
 206                 cli_hexdump(&m , 4); cli_putc(','); cli_hexdump(&r, 1);cli_putstr("): I=");
 207                 cli_hexdump(&t, 4);
 208         ia = pgm_read_dword(&s1[((uint8_t*)&t)[IA]] );
 209         ib = pgm_read_dword(&s2[((uint8_t*)&t)[IB]] );
 210         ic = pgm_read_dword(&s3[((uint8_t*)&t)[IC]] );
 211         id = pgm_read_dword(&s4[((uint8_t*)&t)[ID]] );
 212         cli_putstr("\r\n\tIA="); cli_hexdump(&ia, 4);
 213         cli_putstr("\r\n\tIB="); cli_hexdump(&ib, 4);
 214         cli_putstr("\r\n\tIC="); cli_hexdump(&ic, 4);
 215         cli_putstr("\r\n\tID="); cli_hexdump(&id, 4);
 216
 217         return (((ia ^ ib) - ic) + id);
 218
 219 #else
 220
 221         return (((  pgm_read_dword(&s1[((uint8_t*)&t)[IA]])
 222                   ^ pgm_read_dword(&s2[((uint8_t*)&t)[IB]]) )
 223                   - pgm_read_dword(&s3[((uint8_t*)&t)[IC]]) )
 224                   + pgm_read_dword(&s4[((uint8_t*)&t)[ID]]) );
 225
 226 #endif
 227 }
 228
 229 static
 230 uint32_t cast5_f2(uint32_t d, uint32_t m, uint8_t r){
 231         uint32_t t;
 232         t = ROTL32((d ^ m),r);
 233 #ifdef DEBUG
 234         uint32_t ia,ib,ic,id;
 235         cli_putstr("\r\n f2("); cli_hexdump(&d, 4); cli_putc(',');
 236                 cli_hexdump(&m , 4); cli_putc(','); cli_hexdump(&r, 1);cli_putstr("): I=");
 237                 cli_hexdump(&t, 4);
 238
 239         ia = pgm_read_dword(&s1[((uint8_t*)&t)[IA]] );
 240         ib = pgm_read_dword(&s2[((uint8_t*)&t)[IB]] );
 241         ic = pgm_read_dword(&s3[((uint8_t*)&t)[IC]] );
 242         id = pgm_read_dword(&s4[((uint8_t*)&t)[ID]] );
 243
 244         cli_putstr("\r\n\tIA="); cli_hexdump(&ia, 4);
 245         cli_putstr("\r\n\tIB="); cli_hexdump(&ib, 4);
 246         cli_putstr("\r\n\tIC="); cli_hexdump(&ic, 4);
 247         cli_putstr("\r\n\tID="); cli_hexdump(&id, 4);
 248
 249         return (((ia - ib) + ic) ^ id);
 250 #else
 251
 252         return (((    pgm_read_dword(&s1[((uint8_t*)&t)[IA]])
 253                     - pgm_read_dword(&s2[((uint8_t*)&t)[IB]]) )
 254                     + pgm_read_dword(&s3[((uint8_t*)&t)[IC]]) )
 255                     ^ pgm_read_dword(&s4[((uint8_t*)&t)[ID]]) );
 256
 257 #endif
 258 }
 259
 260 static
 261 uint32_t cast5_f3(uint32_t d, uint32_t m, uint8_t r){
 262         uint32_t t;
 263         t = ROTL32((m - d),r);
 264
 265 #ifdef DEBUG
 266         uint32_t ia,ib,ic,id;
 267
 268         cli_putstr("\r\n f3("); cli_hexdump(&d, 4); cli_putc(',');
 269                 cli_hexdump(&m , 4); cli_putc(','); cli_hexdump(&r, 1);cli_putstr("): I=");
 270                 cli_hexdump(&t, 4);
 271
 272         ia = pgm_read_dword(&s1[((uint8_t*)&t)[IA]] );
 273         ib = pgm_read_dword(&s2[((uint8_t*)&t)[IB]] );
 274         ic = pgm_read_dword(&s3[((uint8_t*)&t)[IC]] );
 275         id = pgm_read_dword(&s4[((uint8_t*)&t)[ID]] );
 276
 277         cli_putstr("\r\n\tIA="); cli_hexdump(&ia, 4);
 278         cli_putstr("\r\n\tIB="); cli_hexdump(&ib, 4);
 279         cli_putstr("\r\n\tIC="); cli_hexdump(&ic, 4);
 280         cli_putstr("\r\n\tID="); cli_hexdump(&id, 4);
 281         return (((ia + ib) ^ ic) - id);
 282 #else
 283         return ((  pgm_read_dword(&s1[((uint8_t*)&t)[IA]] )
 284                  + pgm_read_dword(&s2[((uint8_t*)&t)[IB]] ))
 285                  ^ pgm_read_dword(&s3[((uint8_t*)&t)[IC]] ))
 286                  - pgm_read_dword(&s4[((uint8_t*)&t)[ID]] );
 287
 288 #endif
 289 }
 290
 291 /******************************************************************************/
 292
 293 void cast5_enc(void* block, const cast5_ctx_t *s){
 294         uint32_t l,r, x, y;
 295         uint8_t i;
 296         cast5_f_t* f[]={cast5_f1,cast5_f2,cast5_f3};
 297         l=((uint32_t*)block)[0];
 298         r=((uint32_t*)block)[1];
 299 //      cli_putstr("\r\n round[-1] = ");
 300 //      cli_hexdump(&r, 4);
 301         for (i=0;i<(s->shortkey?12:16);++i){
 302                 x = r;
 303                 y = (f[i%3])(CHANGE_ENDIAN32(r), CHANGE_ENDIAN32(s->mask[i]),
 304                         (((s->roth[i>>3]) & (1<<(i&0x7)))?0x10:0x00)
 305                          + ( ((s->rotl[i>>1])>>((i&1)?4:0)) & 0x0f) );
 306                 r = l ^ CHANGE_ENDIAN32(y);
 307 //              cli_putstr("\r\n round["); DEBUG_B(i); cli_putstr("] = ");
 308 //              cli_hexdump(&r, 4);
 309                 l = x;
 310         }
 311         ((uint32_t*)block)[0]=r;
 312         ((uint32_t*)block)[1]=l;
 313 }
 314
 315 /******************************************************************************/
 316
 317 void cast5_dec(void* block, const cast5_ctx_t *s){
 318         uint32_t l,r, x, y;
 319         int8_t i, rounds;
 320         cast5_f_t* f[]={cast5_f1,cast5_f2,cast5_f3};
 321         l=((uint32_t*)block)[0];
 322         r=((uint32_t*)block)[1];
 323         rounds = (s->shortkey?12:16);
 324         for (i=rounds-1; i>=0 ;--i){
 325                 x = r;
 326                 y = (f[i%3])(CHANGE_ENDIAN32(r), CHANGE_ENDIAN32(s->mask[i]),
 327                         (((s->roth[i>>3]) & (1<<(i&0x7)))?0x10:0x00)
 328                          + ( ((s->rotl[i>>1])>>((i&1)?4:0)) & 0x0f) );
 329                 r = l ^ CHANGE_ENDIAN32(y);
 330                 l = x;
 331         }
 332         ((uint32_t*)block)[0]=r;
 333         ((uint32_t*)block)[1]=l;
 334 }
 335
 336
 337 /******************************************************************************/
 338
 339
 340
 341