crypto/serpent.c

   1 /*
   2  * Cryptographic API.
   3  *
   4  * Serpent Cipher Algorithm.
   5  *
   6  * Copyright (C) 2002 Dag Arne Osvik <osvik@ii.uib.no>
   7  *               2003 Herbert Valerio Riedel <hvr@gnu.org>
   8  *
   9  * Added tnepres support: Ruben Jesus Garcia Hernandez <ruben@ugr.es>, 18.10.2004
  10  *               Based on code by hvr
  11  *
  12  * This program is free software; you can redistribute it and/or modify
  13  * it under the terms of the GNU General Public License as published by
  14  * the Free Software Foundation; either version 2 of the License, or
  15  * (at your option) any later version.
  16  */
  17
  18 #include <linux/init.h>
  19 #include <linux/module.h>
  20 #include <linux/errno.h>
  21 #include <asm/byteorder.h>
  22 #include <linux/crypto.h>
  23 #include <linux/types.h>
  24
  25 /* Key is padded to the maximum of 256 bits before round key generation.
  26  * Any key length <= 256 bits (32 bytes) is allowed by the algorithm.
  27  */
  28
  29 #define SERPENT_MIN_KEY_SIZE              0
  30 #define SERPENT_MAX_KEY_SIZE             32
  31 #define SERPENT_EXPKEY_WORDS            132
  32 #define SERPENT_BLOCK_SIZE               16
  33
  34 #define PHI 0x9e3779b9UL
  35
  36 #define keyiter(a,b,c,d,i,j) \
  37         b ^= d; b ^= c; b ^= a; b ^= PHI ^ i; b = rol32(b,11); k[j] = b;
  38
  39 #define loadkeys(x0,x1,x2,x3,i) \
  40         x0=k[i]; x1=k[i+1]; x2=k[i+2]; x3=k[i+3];
  41
  42 #define storekeys(x0,x1,x2,x3,i) \
  43         k[i]=x0; k[i+1]=x1; k[i+2]=x2; k[i+3]=x3;
  44
  45 #define K(x0,x1,x2,x3,i)                                \
  46         x3 ^= k[4*(i)+3];        x2 ^= k[4*(i)+2];      \
  47         x1 ^= k[4*(i)+1];        x0 ^= k[4*(i)+0];
  48
  49 #define LK(x0,x1,x2,x3,x4,i)                            \
  50                                         x0=rol32(x0,13);\
  51         x2=rol32(x2,3); x1 ^= x0;       x4  = x0 << 3;  \
  52         x3 ^= x2;       x1 ^= x2;                       \
  53         x1=rol32(x1,1); x3 ^= x4;                       \
  54         x3=rol32(x3,7); x4  = x1;                       \
  55         x0 ^= x1;       x4 <<= 7;       x2 ^= x3;       \
  56         x0 ^= x3;       x2 ^= x4;       x3 ^= k[4*i+3]; \
  57         x1 ^= k[4*i+1]; x0=rol32(x0,5); x2=rol32(x2,22);\
  58         x0 ^= k[4*i+0]; x2 ^= k[4*i+2];
  59
  60 #define KL(x0,x1,x2,x3,x4,i)                            \
  61         x0 ^= k[4*i+0]; x1 ^= k[4*i+1]; x2 ^= k[4*i+2]; \
  62         x3 ^= k[4*i+3]; x0=ror32(x0,5); x2=ror32(x2,22);\
  63         x4 =  x1;       x2 ^= x3;       x0 ^= x3;       \
  64         x4 <<= 7;       x0 ^= x1;       x1=ror32(x1,1); \
  65         x2 ^= x4;       x3=ror32(x3,7); x4 = x0 << 3;   \
  66         x1 ^= x0;       x3 ^= x4;       x0=ror32(x0,13);\
  67         x1 ^= x2;       x3 ^= x2;       x2=ror32(x2,3);
  68
  69 #define S0(x0,x1,x2,x3,x4)                              \
  70                                         x4  = x3;       \
  71         x3 |= x0;       x0 ^= x4;       x4 ^= x2;       \
  72         x4 =~ x4;       x3 ^= x1;       x1 &= x0;       \
  73         x1 ^= x4;       x2 ^= x0;       x0 ^= x3;       \
  74         x4 |= x0;       x0 ^= x2;       x2 &= x1;       \
  75         x3 ^= x2;       x1 =~ x1;       x2 ^= x4;       \
  76         x1 ^= x2;
  77
  78 #define S1(x0,x1,x2,x3,x4)                              \
  79                                         x4  = x1;       \
  80         x1 ^= x0;       x0 ^= x3;       x3 =~ x3;       \
  81         x4 &= x1;       x0 |= x1;       x3 ^= x2;       \
  82         x0 ^= x3;       x1 ^= x3;       x3 ^= x4;       \
  83         x1 |= x4;       x4 ^= x2;       x2 &= x0;       \
  84         x2 ^= x1;       x1 |= x0;       x0 =~ x0;       \
  85         x0 ^= x2;       x4 ^= x1;
  86
  87 #define S2(x0,x1,x2,x3,x4)                              \
  88                                         x3 =~ x3;       \
  89         x1 ^= x0;       x4  = x0;       x0 &= x2;       \
  90         x0 ^= x3;       x3 |= x4;       x2 ^= x1;       \
  91         x3 ^= x1;       x1 &= x0;       x0 ^= x2;       \
  92         x2 &= x3;       x3 |= x1;       x0 =~ x0;       \
  93         x3 ^= x0;       x4 ^= x0;       x0 ^= x2;       \
  94         x1 |= x2;
  95
  96 #define S3(x0,x1,x2,x3,x4)                              \
  97                                         x4  = x1;       \
  98         x1 ^= x3;       x3 |= x0;       x4 &= x0;       \
  99         x0 ^= x2;       x2 ^= x1;       x1 &= x3;       \
 100         x2 ^= x3;       x0 |= x4;       x4 ^= x3;       \
 101         x1 ^= x0;       x0 &= x3;       x3 &= x4;       \
 102         x3 ^= x2;       x4 |= x1;       x2 &= x1;       \
 103         x4 ^= x3;       x0 ^= x3;       x3 ^= x2;
 104
 105 #define S4(x0,x1,x2,x3,x4)                              \
 106                                         x4  = x3;       \
 107         x3 &= x0;       x0 ^= x4;                       \
 108         x3 ^= x2;       x2 |= x4;       x0 ^= x1;       \
 109         x4 ^= x3;       x2 |= x0;                       \
 110         x2 ^= x1;       x1 &= x0;                       \
 111         x1 ^= x4;       x4 &= x2;       x2 ^= x3;       \
 112         x4 ^= x0;       x3 |= x1;       x1 =~ x1;       \
 113         x3 ^= x0;
 114
 115 #define S5(x0,x1,x2,x3,x4)                              \
 116         x4  = x1;       x1 |= x0;                       \
 117         x2 ^= x1;       x3 =~ x3;       x4 ^= x0;       \
 118         x0 ^= x2;       x1 &= x4;       x4 |= x3;       \
 119         x4 ^= x0;       x0 &= x3;       x1 ^= x3;       \
 120         x3 ^= x2;       x0 ^= x1;       x2 &= x4;       \
 121         x1 ^= x2;       x2 &= x0;                       \
 122         x3 ^= x2;
 123
 124 #define S6(x0,x1,x2,x3,x4)                              \
 125                                         x4  = x1;       \
 126         x3 ^= x0;       x1 ^= x2;       x2 ^= x0;       \
 127         x0 &= x3;       x1 |= x3;       x4 =~ x4;       \
 128         x0 ^= x1;       x1 ^= x2;                       \
 129         x3 ^= x4;       x4 ^= x0;       x2 &= x0;       \
 130         x4 ^= x1;       x2 ^= x3;       x3 &= x1;       \
 131         x3 ^= x0;       x1 ^= x2;
 132
 133 #define S7(x0,x1,x2,x3,x4)                              \
 134                                         x1 =~ x1;       \
 135         x4  = x1;       x0 =~ x0;       x1 &= x2;       \
 136         x1 ^= x3;       x3 |= x4;       x4 ^= x2;       \
 137         x2 ^= x3;       x3 ^= x0;       x0 |= x1;       \
 138         x2 &= x0;       x0 ^= x4;       x4 ^= x3;       \
 139         x3 &= x0;       x4 ^= x1;                       \
 140         x2 ^= x4;       x3 ^= x1;       x4 |= x0;       \
 141         x4 ^= x1;
 142
 143 #define SI0(x0,x1,x2,x3,x4)                             \
 144                         x4  = x3;       x1 ^= x0;       \
 145         x3 |= x1;       x4 ^= x1;       x0 =~ x0;       \
 146         x2 ^= x3;       x3 ^= x0;       x0 &= x1;       \
 147         x0 ^= x2;       x2 &= x3;       x3 ^= x4;       \
 148         x2 ^= x3;       x1 ^= x3;       x3 &= x0;       \
 149         x1 ^= x0;       x0 ^= x2;       x4 ^= x3;
 150
 151 #define SI1(x0,x1,x2,x3,x4)                             \
 152         x1 ^= x3;       x4  = x0;                       \
 153         x0 ^= x2;       x2 =~ x2;       x4 |= x1;       \
 154         x4 ^= x3;       x3 &= x1;       x1 ^= x2;       \
 155         x2 &= x4;       x4 ^= x1;       x1 |= x3;       \
 156         x3 ^= x0;       x2 ^= x0;       x0 |= x4;       \
 157         x2 ^= x4;       x1 ^= x0;                       \
 158         x4 ^= x1;
 159
 160 #define SI2(x0,x1,x2,x3,x4)                             \
 161         x2 ^= x1;       x4  = x3;       x3 =~ x3;       \
 162         x3 |= x2;       x2 ^= x4;       x4 ^= x0;       \
 163         x3 ^= x1;       x1 |= x2;       x2 ^= x0;       \
 164         x1 ^= x4;       x4 |= x3;       x2 ^= x3;       \
 165         x4 ^= x2;       x2 &= x1;                       \
 166         x2 ^= x3;       x3 ^= x4;       x4 ^= x0;
 167
 168 #define SI3(x0,x1,x2,x3,x4)                             \
 169                                         x2 ^= x1;       \
 170         x4  = x1;       x1 &= x2;                       \
 171         x1 ^= x0;       x0 |= x4;       x4 ^= x3;       \
 172         x0 ^= x3;       x3 |= x1;       x1 ^= x2;       \
 173         x1 ^= x3;       x0 ^= x2;       x2 ^= x3;       \
 174         x3 &= x1;       x1 ^= x0;       x0 &= x2;       \
 175         x4 ^= x3;       x3 ^= x0;       x0 ^= x1;
 176
 177 #define SI4(x0,x1,x2,x3,x4)                             \
 178         x2 ^= x3;       x4  = x0;       x0 &= x1;       \
 179         x0 ^= x2;       x2 |= x3;       x4 =~ x4;       \
 180         x1 ^= x0;       x0 ^= x2;       x2 &= x4;       \
 181         x2 ^= x0;       x0 |= x4;                       \
 182         x0 ^= x3;       x3 &= x2;                       \
 183         x4 ^= x3;       x3 ^= x1;       x1 &= x0;       \
 184         x4 ^= x1;       x0 ^= x3;
 185
 186 #define SI5(x0,x1,x2,x3,x4)                             \
 187                         x4  = x1;       x1 |= x2;       \
 188         x2 ^= x4;       x1 ^= x3;       x3 &= x4;       \
 189         x2 ^= x3;       x3 |= x0;       x0 =~ x0;       \
 190         x3 ^= x2;       x2 |= x0;       x4 ^= x1;       \
 191         x2 ^= x4;       x4 &= x0;       x0 ^= x1;       \
 192         x1 ^= x3;       x0 &= x2;       x2 ^= x3;       \
 193         x0 ^= x2;       x2 ^= x4;       x4 ^= x3;
 194
 195 #define SI6(x0,x1,x2,x3,x4)                             \
 196                         x0 ^= x2;                       \
 197         x4  = x0;       x0 &= x3;       x2 ^= x3;       \
 198         x0 ^= x2;       x3 ^= x1;       x2 |= x4;       \
 199         x2 ^= x3;       x3 &= x0;       x0 =~ x0;       \
 200         x3 ^= x1;       x1 &= x2;       x4 ^= x0;       \
 201         x3 ^= x4;       x4 ^= x2;       x0 ^= x1;       \
 202         x2 ^= x0;
 203
 204 #define SI7(x0,x1,x2,x3,x4)                             \
 205         x4  = x3;       x3 &= x0;       x0 ^= x2;       \
 206         x2 |= x4;       x4 ^= x1;       x0 =~ x0;       \
 207         x1 |= x3;       x4 ^= x0;       x0 &= x2;       \
 208         x0 ^= x1;       x1 &= x2;       x3 ^= x2;       \
 209         x4 ^= x3;       x2 &= x3;       x3 |= x0;       \
 210         x1 ^= x4;       x3 ^= x4;       x4 &= x0;       \
 211         x4 ^= x2;
 212
 213 struct serpent_ctx {
 214         u32 expkey[SERPENT_EXPKEY_WORDS];
 215 };
 216
 217
 218 static int serpent_setkey(struct crypto_tfm *tfm, const u8 *key,
 219                           unsigned int keylen)
 220 {
 221         struct serpent_ctx *ctx = crypto_tfm_ctx(tfm);
 222         u32 *k = ctx->expkey;
 223         u8  *k8 = (u8 *)k;
 224         u32 r0,r1,r2,r3,r4;
 225         int i;
 226
 227         /* Copy key, add padding */
 228
 229         for (i = 0; i < keylen; ++i)
 230                 k8[i] = key[i];
 231         if (i < SERPENT_MAX_KEY_SIZE)
 232                 k8[i++] = 1;
 233         while (i < SERPENT_MAX_KEY_SIZE)
 234                 k8[i++] = 0;
 235
 236         /* Expand key using polynomial */
 237
 238         r0 = le32_to_cpu(k[3]);
 239         r1 = le32_to_cpu(k[4]);
 240         r2 = le32_to_cpu(k[5]);
 241         r3 = le32_to_cpu(k[6]);
 242         r4 = le32_to_cpu(k[7]);
 243
 244         keyiter(le32_to_cpu(k[0]),r0,r4,r2,0,0);
 245         keyiter(le32_to_cpu(k[1]),r1,r0,r3,1,1);
 246         keyiter(le32_to_cpu(k[2]),r2,r1,r4,2,2);
 247         keyiter(le32_to_cpu(k[3]),r3,r2,r0,3,3);
 248         keyiter(le32_to_cpu(k[4]),r4,r3,r1,4,4);
 249         keyiter(le32_to_cpu(k[5]),r0,r4,r2,5,5);
 250         keyiter(le32_to_cpu(k[6]),r1,r0,r3,6,6);
 251         keyiter(le32_to_cpu(k[7]),r2,r1,r4,7,7);
 252
 253         keyiter(k[  0],r3,r2,r0,  8,  8); keyiter(k[  1],r4,r3,r1,  9,  9);
 254         keyiter(k[  2],r0,r4,r2, 10, 10); keyiter(k[  3],r1,r0,r3, 11, 11);
 255         keyiter(k[  4],r2,r1,r4, 12, 12); keyiter(k[  5],r3,r2,r0, 13, 13);
 256         keyiter(k[  6],r4,r3,r1, 14, 14); keyiter(k[  7],r0,r4,r2, 15, 15);
 257         keyiter(k[  8],r1,r0,r3, 16, 16); keyiter(k[  9],r2,r1,r4, 17, 17);
 258         keyiter(k[ 10],r3,r2,r0, 18, 18); keyiter(k[ 11],r4,r3,r1, 19, 19);
 259         keyiter(k[ 12],r0,r4,r2, 20, 20); keyiter(k[ 13],r1,r0,r3, 21, 21);
 260         keyiter(k[ 14],r2,r1,r4, 22, 22); keyiter(k[ 15],r3,r2,r0, 23, 23);
 261         keyiter(k[ 16],r4,r3,r1, 24, 24); keyiter(k[ 17],r0,r4,r2, 25, 25);
 262         keyiter(k[ 18],r1,r0,r3, 26, 26); keyiter(k[ 19],r2,r1,r4, 27, 27);
 263         keyiter(k[ 20],r3,r2,r0, 28, 28); keyiter(k[ 21],r4,r3,r1, 29, 29);
 264         keyiter(k[ 22],r0,r4,r2, 30, 30); keyiter(k[ 23],r1,r0,r3, 31, 31);
 265
 266         k += 50;
 267
 268         keyiter(k[-26],r2,r1,r4, 32,-18); keyiter(k[-25],r3,r2,r0, 33,-17);
 269         keyiter(k[-24],r4,r3,r1, 34,-16); keyiter(k[-23],r0,r4,r2, 35,-15);
 270         keyiter(k[-22],r1,r0,r3, 36,-14); keyiter(k[-21],r2,r1,r4, 37,-13);
 271         keyiter(k[-20],r3,r2,r0, 38,-12); keyiter(k[-19],r4,r3,r1, 39,-11);
 272         keyiter(k[-18],r0,r4,r2, 40,-10); keyiter(k[-17],r1,r0,r3, 41, -9);
 273         keyiter(k[-16],r2,r1,r4, 42, -8); keyiter(k[-15],r3,r2,r0, 43, -7);
 274         keyiter(k[-14],r4,r3,r1, 44, -6); keyiter(k[-13],r0,r4,r2, 45, -5);
 275         keyiter(k[-12],r1,r0,r3, 46, -4); keyiter(k[-11],r2,r1,r4, 47, -3);
 276         keyiter(k[-10],r3,r2,r0, 48, -2); keyiter(k[ -9],r4,r3,r1, 49, -1);
 277         keyiter(k[ -8],r0,r4,r2, 50,  0); keyiter(k[ -7],r1,r0,r3, 51,  1);
 278         keyiter(k[ -6],r2,r1,r4, 52,  2); keyiter(k[ -5],r3,r2,r0, 53,  3);
 279         keyiter(k[ -4],r4,r3,r1, 54,  4); keyiter(k[ -3],r0,r4,r2, 55,  5);
 280         keyiter(k[ -2],r1,r0,r3, 56,  6); keyiter(k[ -1],r2,r1,r4, 57,  7);
 281         keyiter(k[  0],r3,r2,r0, 58,  8); keyiter(k[  1],r4,r3,r1, 59,  9);
 282         keyiter(k[  2],r0,r4,r2, 60, 10); keyiter(k[  3],r1,r0,r3, 61, 11);
 283         keyiter(k[  4],r2,r1,r4, 62, 12); keyiter(k[  5],r3,r2,r0, 63, 13);
 284         keyiter(k[  6],r4,r3,r1, 64, 14); keyiter(k[  7],r0,r4,r2, 65, 15);
 285         keyiter(k[  8],r1,r0,r3, 66, 16); keyiter(k[  9],r2,r1,r4, 67, 17);
 286         keyiter(k[ 10],r3,r2,r0, 68, 18); keyiter(k[ 11],r4,r3,r1, 69, 19);
 287         keyiter(k[ 12],r0,r4,r2, 70, 20); keyiter(k[ 13],r1,r0,r3, 71, 21);
 288         keyiter(k[ 14],r2,r1,r4, 72, 22); keyiter(k[ 15],r3,r2,r0, 73, 23);
 289         keyiter(k[ 16],r4,r3,r1, 74, 24); keyiter(k[ 17],r0,r4,r2, 75, 25);
 290         keyiter(k[ 18],r1,r0,r3, 76, 26); keyiter(k[ 19],r2,r1,r4, 77, 27);
 291         keyiter(k[ 20],r3,r2,r0, 78, 28); keyiter(k[ 21],r4,r3,r1, 79, 29);
 292         keyiter(k[ 22],r0,r4,r2, 80, 30); keyiter(k[ 23],r1,r0,r3, 81, 31);
 293
 294         k += 50;
 295
 296         keyiter(k[-26],r2,r1,r4, 82,-18); keyiter(k[-25],r3,r2,r0, 83,-17);
 297         keyiter(k[-24],r4,r3,r1, 84,-16); keyiter(k[-23],r0,r4,r2, 85,-15);
 298         keyiter(k[-22],r1,r0,r3, 86,-14); keyiter(k[-21],r2,r1,r4, 87,-13);
 299         keyiter(k[-20],r3,r2,r0, 88,-12); keyiter(k[-19],r4,r3,r1, 89,-11);
 300         keyiter(k[-18],r0,r4,r2, 90,-10); keyiter(k[-17],r1,r0,r3, 91, -9);
 301         keyiter(k[-16],r2,r1,r4, 92, -8); keyiter(k[-15],r3,r2,r0, 93, -7);
 302         keyiter(k[-14],r4,r3,r1, 94, -6); keyiter(k[-13],r0,r4,r2, 95, -5);
 303         keyiter(k[-12],r1,r0,r3, 96, -4); keyiter(k[-11],r2,r1,r4, 97, -3);
 304         keyiter(k[-10],r3,r2,r0, 98, -2); keyiter(k[ -9],r4,r3,r1, 99, -1);
 305         keyiter(k[ -8],r0,r4,r2,100,  0); keyiter(k[ -7],r1,r0,r3,101,  1);
 306         keyiter(k[ -6],r2,r1,r4,102,  2); keyiter(k[ -5],r3,r2,r0,103,  3);
 307         keyiter(k[ -4],r4,r3,r1,104,  4); keyiter(k[ -3],r0,r4,r2,105,  5);
 308         keyiter(k[ -2],r1,r0,r3,106,  6); keyiter(k[ -1],r2,r1,r4,107,  7);
 309         keyiter(k[  0],r3,r2,r0,108,  8); keyiter(k[  1],r4,r3,r1,109,  9);
 310         keyiter(k[  2],r0,r4,r2,110, 10); keyiter(k[  3],r1,r0,r3,111, 11);
 311         keyiter(k[  4],r2,r1,r4,112, 12); keyiter(k[  5],r3,r2,r0,113, 13);
 312         keyiter(k[  6],r4,r3,r1,114, 14); keyiter(k[  7],r0,r4,r2,115, 15);
 313         keyiter(k[  8],r1,r0,r3,116, 16); keyiter(k[  9],r2,r1,r4,117, 17);
 314         keyiter(k[ 10],r3,r2,r0,118, 18); keyiter(k[ 11],r4,r3,r1,119, 19);
 315         keyiter(k[ 12],r0,r4,r2,120, 20); keyiter(k[ 13],r1,r0,r3,121, 21);
 316         keyiter(k[ 14],r2,r1,r4,122, 22); keyiter(k[ 15],r3,r2,r0,123, 23);
 317         keyiter(k[ 16],r4,r3,r1,124, 24); keyiter(k[ 17],r0,r4,r2,125, 25);
 318         keyiter(k[ 18],r1,r0,r3,126, 26); keyiter(k[ 19],r2,r1,r4,127, 27);
 319         keyiter(k[ 20],r3,r2,r0,128, 28); keyiter(k[ 21],r4,r3,r1,129, 29);
 320         keyiter(k[ 22],r0,r4,r2,130, 30); keyiter(k[ 23],r1,r0,r3,131, 31);
 321
 322         /* Apply S-boxes */
 323
 324         S3(r3,r4,r0,r1,r2); storekeys(r1,r2,r4,r3, 28); loadkeys(r1,r2,r4,r3, 24);
 325         S4(r1,r2,r4,r3,r0); storekeys(r2,r4,r3,r0, 24); loadkeys(r2,r4,r3,r0, 20);
 326         S5(r2,r4,r3,r0,r1); storekeys(r1,r2,r4,r0, 20); loadkeys(r1,r2,r4,r0, 16);
 327         S6(r1,r2,r4,r0,r3); storekeys(r4,r3,r2,r0, 16); loadkeys(r4,r3,r2,r0, 12);
 328         S7(r4,r3,r2,r0,r1); storekeys(r1,r2,r0,r4, 12); loadkeys(r1,r2,r0,r4,  8);
 329         S0(r1,r2,r0,r4,r3); storekeys(r0,r2,r4,r1,  8); loadkeys(r0,r2,r4,r1,  4);
 330         S1(r0,r2,r4,r1,r3); storekeys(r3,r4,r1,r0,  4); loadkeys(r3,r4,r1,r0,  0);
 331         S2(r3,r4,r1,r0,r2); storekeys(r2,r4,r3,r0,  0); loadkeys(r2,r4,r3,r0, -4);
 332         S3(r2,r4,r3,r0,r1); storekeys(r0,r1,r4,r2, -4); loadkeys(r0,r1,r4,r2, -8);
 333         S4(r0,r1,r4,r2,r3); storekeys(r1,r4,r2,r3, -8); loadkeys(r1,r4,r2,r3,-12);
 334         S5(r1,r4,r2,r3,r0); storekeys(r0,r1,r4,r3,-12); loadkeys(r0,r1,r4,r3,-16);
 335         S6(r0,r1,r4,r3,r2); storekeys(r4,r2,r1,r3,-16); loadkeys(r4,r2,r1,r3,-20);
 336         S7(r4,r2,r1,r3,r0); storekeys(r0,r1,r3,r4,-20); loadkeys(r0,r1,r3,r4,-24);
 337         S0(r0,r1,r3,r4,r2); storekeys(r3,r1,r4,r0,-24); loadkeys(r3,r1,r4,r0,-28);
 338         k -= 50;
 339         S1(r3,r1,r4,r0,r2); storekeys(r2,r4,r0,r3, 22); loadkeys(r2,r4,r0,r3, 18);
 340         S2(r2,r4,r0,r3,r1); storekeys(r1,r4,r2,r3, 18); loadkeys(r1,r4,r2,r3, 14);
 341         S3(r1,r4,r2,r3,r0); storekeys(r3,r0,r4,r1, 14); loadkeys(r3,r0,r4,r1, 10);
 342         S4(r3,r0,r4,r1,r2); storekeys(r0,r4,r1,r2, 10); loadkeys(r0,r4,r1,r2,  6);
 343         S5(r0,r4,r1,r2,r3); storekeys(r3,r0,r4,r2,  6); loadkeys(r3,r0,r4,r2,  2);
 344         S6(r3,r0,r4,r2,r1); storekeys(r4,r1,r0,r2,  2); loadkeys(r4,r1,r0,r2, -2);
 345         S7(r4,r1,r0,r2,r3); storekeys(r3,r0,r2,r4, -2); loadkeys(r3,r0,r2,r4, -6);
 346         S0(r3,r0,r2,r4,r1); storekeys(r2,r0,r4,r3, -6); loadkeys(r2,r0,r4,r3,-10);
 347         S1(r2,r0,r4,r3,r1); storekeys(r1,r4,r3,r2,-10); loadkeys(r1,r4,r3,r2,-14);
 348         S2(r1,r4,r3,r2,r0); storekeys(r0,r4,r1,r2,-14); loadkeys(r0,r4,r1,r2,-18);
 349         S3(r0,r4,r1,r2,r3); storekeys(r2,r3,r4,r0,-18); loadkeys(r2,r3,r4,r0,-22);
 350         k -= 50;
 351         S4(r2,r3,r4,r0,r1); storekeys(r3,r4,r0,r1, 28); loadkeys(r3,r4,r0,r1, 24);
 352         S5(r3,r4,r0,r1,r2); storekeys(r2,r3,r4,r1, 24); loadkeys(r2,r3,r4,r1, 20);
 353         S6(r2,r3,r4,r1,r0); storekeys(r4,r0,r3,r1, 20); loadkeys(r4,r0,r3,r1, 16);
 354         S7(r4,r0,r3,r1,r2); storekeys(r2,r3,r1,r4, 16); loadkeys(r2,r3,r1,r4, 12);
 355         S0(r2,r3,r1,r4,r0); storekeys(r1,r3,r4,r2, 12); loadkeys(r1,r3,r4,r2,  8);
 356         S1(r1,r3,r4,r2,r0); storekeys(r0,r4,r2,r1,  8); loadkeys(r0,r4,r2,r1,  4);
 357         S2(r0,r4,r2,r1,r3); storekeys(r3,r4,r0,r1,  4); loadkeys(r3,r4,r0,r1,  0);
 358         S3(r3,r4,r0,r1,r2); storekeys(r1,r2,r4,r3,  0);
 359
 360         return 0;
 361 }
 362
 363 static void serpent_encrypt(struct crypto_tfm *tfm, u8 *dst, const u8 *src)
 364 {
 365         struct serpent_ctx *ctx = crypto_tfm_ctx(tfm);
 366         const u32
 367                 *k = ctx->expkey;
 368         const __le32 *s = (const __le32 *)src;
 369         __le32  *d = (__le32 *)dst;
 370         u32     r0, r1, r2, r3, r4;
 371
 372 /*
 373  * Note: The conversions between u8* and u32* might cause trouble
 374  * on architectures with stricter alignment rules than x86
 375  */
 376
 377         r0 = le32_to_cpu(s[0]);
 378         r1 = le32_to_cpu(s[1]);
 379         r2 = le32_to_cpu(s[2]);
 380         r3 = le32_to_cpu(s[3]);
 381
 382                                  K(r0,r1,r2,r3,0);
 383         S0(r0,r1,r2,r3,r4);     LK(r2,r1,r3,r0,r4,1);
 384         S1(r2,r1,r3,r0,r4);     LK(r4,r3,r0,r2,r1,2);
 385         S2(r4,r3,r0,r2,r1);     LK(r1,r3,r4,r2,r0,3);
 386         S3(r1,r3,r4,r2,r0);     LK(r2,r0,r3,r1,r4,4);
 387         S4(r2,r0,r3,r1,r4);     LK(r0,r3,r1,r4,r2,5);
 388         S5(r0,r3,r1,r4,r2);     LK(r2,r0,r3,r4,r1,6);
 389         S6(r2,r0,r3,r4,r1);     LK(r3,r1,r0,r4,r2,7);
 390         S7(r3,r1,r0,r4,r2);     LK(r2,r0,r4,r3,r1,8);
 391         S0(r2,r0,r4,r3,r1);     LK(r4,r0,r3,r2,r1,9);
 392         S1(r4,r0,r3,r2,r1);     LK(r1,r3,r2,r4,r0,10);
 393         S2(r1,r3,r2,r4,r0);     LK(r0,r3,r1,r4,r2,11);
 394         S3(r0,r3,r1,r4,r2);     LK(r4,r2,r3,r0,r1,12);
 395         S4(r4,r2,r3,r0,r1);     LK(r2,r3,r0,r1,r4,13);
 396         S5(r2,r3,r0,r1,r4);     LK(r4,r2,r3,r1,r0,14);
 397         S6(r4,r2,r3,r1,r0);     LK(r3,r0,r2,r1,r4,15);
 398         S7(r3,r0,r2,r1,r4);     LK(r4,r2,r1,r3,r0,16);
 399         S0(r4,r2,r1,r3,r0);     LK(r1,r2,r3,r4,r0,17);
 400         S1(r1,r2,r3,r4,r0);     LK(r0,r3,r4,r1,r2,18);
 401         S2(r0,r3,r4,r1,r2);     LK(r2,r3,r0,r1,r4,19);
 402         S3(r2,r3,r0,r1,r4);     LK(r1,r4,r3,r2,r0,20);
 403         S4(r1,r4,r3,r2,r0);     LK(r4,r3,r2,r0,r1,21);
 404         S5(r4,r3,r2,r0,r1);     LK(r1,r4,r3,r0,r2,22);
 405         S6(r1,r4,r3,r0,r2);     LK(r3,r2,r4,r0,r1,23);
 406         S7(r3,r2,r4,r0,r1);     LK(r1,r4,r0,r3,r2,24);
 407         S0(r1,r4,r0,r3,r2);     LK(r0,r4,r3,r1,r2,25);
 408         S1(r0,r4,r3,r1,r2);     LK(r2,r3,r1,r0,r4,26);
 409         S2(r2,r3,r1,r0,r4);     LK(r4,r3,r2,r0,r1,27);
 410         S3(r4,r3,r2,r0,r1);     LK(r0,r1,r3,r4,r2,28);
 411         S4(r0,r1,r3,r4,r2);     LK(r1,r3,r4,r2,r0,29);
 412         S5(r1,r3,r4,r2,r0);     LK(r0,r1,r3,r2,r4,30);
 413         S6(r0,r1,r3,r2,r4);     LK(r3,r4,r1,r2,r0,31);
 414         S7(r3,r4,r1,r2,r0);      K(r0,r1,r2,r3,32);
 415
 416         d[0] = cpu_to_le32(r0);
 417         d[1] = cpu_to_le32(r1);
 418         d[2] = cpu_to_le32(r2);
 419         d[3] = cpu_to_le32(r3);
 420 }
 421
 422 static void serpent_decrypt(struct crypto_tfm *tfm, u8 *dst, const u8 *src)
 423 {
 424         struct serpent_ctx *ctx = crypto_tfm_ctx(tfm);
 425         const u32
 426                 *k = ((struct serpent_ctx *)ctx)->expkey;
 427         const __le32 *s = (const __le32 *)src;
 428         __le32  *d = (__le32 *)dst;
 429         u32     r0, r1, r2, r3, r4;
 430
 431         r0 = le32_to_cpu(s[0]);
 432         r1 = le32_to_cpu(s[1]);
 433         r2 = le32_to_cpu(s[2]);
 434         r3 = le32_to_cpu(s[3]);
 435
 436                                 K(r0,r1,r2,r3,32);
 437         SI7(r0,r1,r2,r3,r4);    KL(r1,r3,r0,r4,r2,31);
 438         SI6(r1,r3,r0,r4,r2);    KL(r0,r2,r4,r1,r3,30);
 439         SI5(r0,r2,r4,r1,r3);    KL(r2,r3,r0,r4,r1,29);
 440         SI4(r2,r3,r0,r4,r1);    KL(r2,r0,r1,r4,r3,28);
 441         SI3(r2,r0,r1,r4,r3);    KL(r1,r2,r3,r4,r0,27);
 442         SI2(r1,r2,r3,r4,r0);    KL(r2,r0,r4,r3,r1,26);
 443         SI1(r2,r0,r4,r3,r1);    KL(r1,r0,r4,r3,r2,25);
 444         SI0(r1,r0,r4,r3,r2);    KL(r4,r2,r0,r1,r3,24);
 445         SI7(r4,r2,r0,r1,r3);    KL(r2,r1,r4,r3,r0,23);
 446         SI6(r2,r1,r4,r3,r0);    KL(r4,r0,r3,r2,r1,22);
 447         SI5(r4,r0,r3,r2,r1);    KL(r0,r1,r4,r3,r2,21);
 448         SI4(r0,r1,r4,r3,r2);    KL(r0,r4,r2,r3,r1,20);
 449         SI3(r0,r4,r2,r3,r1);    KL(r2,r0,r1,r3,r4,19);
 450         SI2(r2,r0,r1,r3,r4);    KL(r0,r4,r3,r1,r2,18);
 451         SI1(r0,r4,r3,r1,r2);    KL(r2,r4,r3,r1,r0,17);
 452         SI0(r2,r4,r3,r1,r0);    KL(r3,r0,r4,r2,r1,16);
 453         SI7(r3,r0,r4,r2,r1);    KL(r0,r2,r3,r1,r4,15);
 454         SI6(r0,r2,r3,r1,r4);    KL(r3,r4,r1,r0,r2,14);
 455         SI5(r3,r4,r1,r0,r2);    KL(r4,r2,r3,r1,r0,13);
 456         SI4(r4,r2,r3,r1,r0);    KL(r4,r3,r0,r1,r2,12);
 457         SI3(r4,r3,r0,r1,r2);    KL(r0,r4,r2,r1,r3,11);
 458         SI2(r0,r4,r2,r1,r3);    KL(r4,r3,r1,r2,r0,10);
 459         SI1(r4,r3,r1,r2,r0);    KL(r0,r3,r1,r2,r4,9);
 460         SI0(r0,r3,r1,r2,r4);    KL(r1,r4,r3,r0,r2,8);
 461         SI7(r1,r4,r3,r0,r2);    KL(r4,r0,r1,r2,r3,7);
 462         SI6(r4,r0,r1,r2,r3);    KL(r1,r3,r2,r4,r0,6);
 463         SI5(r1,r3,r2,r4,r0);    KL(r3,r0,r1,r2,r4,5);
 464         SI4(r3,r0,r1,r2,r4);    KL(r3,r1,r4,r2,r0,4);
 465         SI3(r3,r1,r4,r2,r0);    KL(r4,r3,r0,r2,r1,3);
 466         SI2(r4,r3,r0,r2,r1);    KL(r3,r1,r2,r0,r4,2);
 467         SI1(r3,r1,r2,r0,r4);    KL(r4,r1,r2,r0,r3,1);
 468         SI0(r4,r1,r2,r0,r3);    K(r2,r3,r1,r4,0);
 469
 470         d[0] = cpu_to_le32(r2);
 471         d[1] = cpu_to_le32(r3);
 472         d[2] = cpu_to_le32(r1);
 473         d[3] = cpu_to_le32(r4);
 474 }
 475
 476 static struct crypto_alg serpent_alg = {
 477         .cra_name               =       "serpent",
 478         .cra_flags              =       CRYPTO_ALG_TYPE_CIPHER,
 479         .cra_blocksize          =       SERPENT_BLOCK_SIZE,
 480         .cra_ctxsize            =       sizeof(struct serpent_ctx),
 481         .cra_alignmask          =       3,
 482         .cra_module             =       THIS_MODULE,
 483         .cra_list               =       LIST_HEAD_INIT(serpent_alg.cra_list),
 484         .cra_u                  =       { .cipher = {
 485         .cia_min_keysize        =       SERPENT_MIN_KEY_SIZE,
 486         .cia_max_keysize        =       SERPENT_MAX_KEY_SIZE,
 487         .cia_setkey             =       serpent_setkey,
 488         .cia_encrypt            =       serpent_encrypt,
 489         .cia_decrypt            =       serpent_decrypt } }
 490 };
 491
 492 static int tnepres_setkey(struct crypto_tfm *tfm, const u8 *key,
 493                           unsigned int keylen)
 494 {
 495         u8 rev_key[SERPENT_MAX_KEY_SIZE];
 496         int i;
 497
 498         for (i = 0; i < keylen; ++i)
 499                 rev_key[keylen - i - 1] = key[i];
 500
 501         return serpent_setkey(tfm, rev_key, keylen);
 502 }
 503
 504 static void tnepres_encrypt(struct crypto_tfm *tfm, u8 *dst, const u8 *src)
 505 {
 506         const u32 * const s = (const u32 * const)src;
 507         u32 * const d = (u32 * const)dst;
 508
 509         u32 rs[4], rd[4];
 510
 511         rs[0] = swab32(s[3]);
 512         rs[1] = swab32(s[2]);
 513         rs[2] = swab32(s[1]);
 514         rs[3] = swab32(s[0]);
 515
 516         serpent_encrypt(tfm, (u8 *)rd, (u8 *)rs);
 517
 518         d[0] = swab32(rd[3]);
 519         d[1] = swab32(rd[2]);
 520         d[2] = swab32(rd[1]);
 521         d[3] = swab32(rd[0]);
 522 }
 523
 524 static void tnepres_decrypt(struct crypto_tfm *tfm, u8 *dst, const u8 *src)
 525 {
 526         const u32 * const s = (const u32 * const)src;
 527         u32 * const d = (u32 * const)dst;
 528
 529         u32 rs[4], rd[4];
 530
 531         rs[0] = swab32(s[3]);
 532         rs[1] = swab32(s[2]);
 533         rs[2] = swab32(s[1]);
 534         rs[3] = swab32(s[0]);
 535
 536         serpent_decrypt(tfm, (u8 *)rd, (u8 *)rs);
 537
 538         d[0] = swab32(rd[3]);
 539         d[1] = swab32(rd[2]);
 540         d[2] = swab32(rd[1]);
 541         d[3] = swab32(rd[0]);
 542 }
 543
 544 static struct crypto_alg tnepres_alg = {
 545         .cra_name               =       "tnepres",
 546         .cra_flags              =       CRYPTO_ALG_TYPE_CIPHER,
 547         .cra_blocksize          =       SERPENT_BLOCK_SIZE,
 548         .cra_ctxsize            =       sizeof(struct serpent_ctx),
 549         .cra_alignmask          =       3,
 550         .cra_module             =       THIS_MODULE,
 551         .cra_list               =       LIST_HEAD_INIT(serpent_alg.cra_list),
 552         .cra_u                  =       { .cipher = {
 553         .cia_min_keysize        =       SERPENT_MIN_KEY_SIZE,
 554         .cia_max_keysize        =       SERPENT_MAX_KEY_SIZE,
 555         .cia_setkey             =       tnepres_setkey,
 556         .cia_encrypt            =       tnepres_encrypt,
 557         .cia_decrypt            =       tnepres_decrypt } }
 558 };
 559
 560 static int __init serpent_mod_init(void)
 561 {
 562         int ret = crypto_register_alg(&serpent_alg);
 563
 564         if (ret)
 565                 return ret;
 566
 567         ret = crypto_register_alg(&tnepres_alg);
 568
 569         if (ret)
 570                 crypto_unregister_alg(&serpent_alg);
 571
 572         return ret;
 573 }
 574
 575 static void __exit serpent_mod_fini(void)
 576 {
 577         crypto_unregister_alg(&tnepres_alg);
 578         crypto_unregister_alg(&serpent_alg);
 579 }
 580
 581 module_init(serpent_mod_init);
 582 module_exit(serpent_mod_fini);
 583
 584 MODULE_LICENSE("GPL");
 585 MODULE_DESCRIPTION("Serpent and tnepres (kerneli compatible serpent reversed) Cipher Algorithm");
 586 MODULE_AUTHOR("Dag Arne Osvik <osvik@ii.uib.no>");
 587 MODULE_ALIAS("tnepres");