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  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License as published by
  10  * the Free Software Foundation; either version 2 of the License, or
  11  * (at your option) any later version.
  12  */
  13
  14 #include <linux/init.h>
  15 #include <linux/module.h>
  16 #include <linux/errno.h>
  17 #include <asm/byteorder.h>
  18 #include <linux/crypto.h>
  19
  20 /* Key is padded to the maximum of 256 bits before round key generation.
  21  * Any key length <= 256 bits (32 bytes) is allowed by the algorithm.
  22  */
  23
  24 #define SERPENT_MIN_KEY_SIZE              0
  25 #define SERPENT_MAX_KEY_SIZE             32
  26 #define SERPENT_EXPKEY_WORDS            132
  27 #define SERPENT_BLOCK_SIZE               16
  28
  29 #define PHI 0x9e3779b9UL
  30 #define ROL(x,r) ((x) = ((x) << (r)) | ((x) >> (32-(r))))
  31 #define ROR(x,r) ((x) = ((x) >> (r)) | ((x) << (32-(r))))
  32
  33 #define keyiter(a,b,c,d,i,j) \
  34         b ^= d; b ^= c; b ^= a; b ^= PHI ^ i; ROL(b,11); k[j] = b;
  35
  36 #define loadkeys(x0,x1,x2,x3,i) \
  37         x0=k[i]; x1=k[i+1]; x2=k[i+2]; x3=k[i+3];
  38
  39 #define storekeys(x0,x1,x2,x3,i) \
  40         k[i]=x0; k[i+1]=x1; k[i+2]=x2; k[i+3]=x3;
  41
  42 #define K(x0,x1,x2,x3,i)                                \
  43         x3 ^= k[4*(i)+3];        x2 ^= k[4*(i)+2];      \
  44         x1 ^= k[4*(i)+1];        x0 ^= k[4*(i)+0];
  45
  46 #define LK(x0,x1,x2,x3,x4,i)                            \
  47                                         ROL(x0,13);     \
  48         ROL(x2,3);      x1 ^= x0;       x4  = x0 << 3;  \
  49         x3 ^= x2;       x1 ^= x2;                       \
  50         ROL(x1,1);      x3 ^= x4;                       \
  51         ROL(x3,7);      x4  = x1;                       \
  52         x0 ^= x1;       x4 <<= 7;       x2 ^= x3;       \
  53         x0 ^= x3;       x2 ^= x4;       x3 ^= k[4*i+3]; \
  54         x1 ^= k[4*i+1]; ROL(x0,5);      ROL(x2,22);     \
  55         x0 ^= k[4*i+0]; x2 ^= k[4*i+2];
  56
  57 #define KL(x0,x1,x2,x3,x4,i)                            \
  58         x0 ^= k[4*i+0]; x1 ^= k[4*i+1]; x2 ^= k[4*i+2]; \
  59         x3 ^= k[4*i+3]; ROR(x0,5);      ROR(x2,22);     \
  60         x4 =  x1;       x2 ^= x3;       x0 ^= x3;       \
  61         x4 <<= 7;       x0 ^= x1;       ROR(x1,1);      \
  62         x2 ^= x4;       ROR(x3,7);      x4 = x0 << 3;   \
  63         x1 ^= x0;       x3 ^= x4;       ROR(x0,13);     \
  64         x1 ^= x2;       x3 ^= x2;       ROR(x2,3);
  65
  66 #define S0(x0,x1,x2,x3,x4)                              \
  67                                         x4  = x3;       \
  68         x3 |= x0;       x0 ^= x4;       x4 ^= x2;       \
  69         x4 =~ x4;       x3 ^= x1;       x1 &= x0;       \
  70         x1 ^= x4;       x2 ^= x0;       x0 ^= x3;       \
  71         x4 |= x0;       x0 ^= x2;       x2 &= x1;       \
  72         x3 ^= x2;       x1 =~ x1;       x2 ^= x4;       \
  73         x1 ^= x2;
  74
  75 #define S1(x0,x1,x2,x3,x4)                              \
  76                                         x4  = x1;       \
  77         x1 ^= x0;       x0 ^= x3;       x3 =~ x3;       \
  78         x4 &= x1;       x0 |= x1;       x3 ^= x2;       \
  79         x0 ^= x3;       x1 ^= x3;       x3 ^= x4;       \
  80         x1 |= x4;       x4 ^= x2;       x2 &= x0;       \
  81         x2 ^= x1;       x1 |= x0;       x0 =~ x0;       \
  82         x0 ^= x2;       x4 ^= x1;
  83
  84 #define S2(x0,x1,x2,x3,x4)                              \
  85                                         x3 =~ x3;       \
  86         x1 ^= x0;       x4  = x0;       x0 &= x2;       \
  87         x0 ^= x3;       x3 |= x4;       x2 ^= x1;       \
  88         x3 ^= x1;       x1 &= x0;       x0 ^= x2;       \
  89         x2 &= x3;       x3 |= x1;       x0 =~ x0;       \
  90         x3 ^= x0;       x4 ^= x0;       x0 ^= x2;       \
  91         x1 |= x2;
  92
  93 #define S3(x0,x1,x2,x3,x4)                              \
  94                                         x4  = x1;       \
  95         x1 ^= x3;       x3 |= x0;       x4 &= x0;       \
  96         x0 ^= x2;       x2 ^= x1;       x1 &= x3;       \
  97         x2 ^= x3;       x0 |= x4;       x4 ^= x3;       \
  98         x1 ^= x0;       x0 &= x3;       x3 &= x4;       \
  99         x3 ^= x2;       x4 |= x1;       x2 &= x1;       \
 100         x4 ^= x3;       x0 ^= x3;       x3 ^= x2;
 101
 102 #define S4(x0,x1,x2,x3,x4)                              \
 103                                         x4  = x3;       \
 104         x3 &= x0;       x0 ^= x4;                       \
 105         x3 ^= x2;       x2 |= x4;       x0 ^= x1;       \
 106         x4 ^= x3;       x2 |= x0;                       \
 107         x2 ^= x1;       x1 &= x0;                       \
 108         x1 ^= x4;       x4 &= x2;       x2 ^= x3;       \
 109         x4 ^= x0;       x3 |= x1;       x1 =~ x1;       \
 110         x3 ^= x0;
 111
 112 #define S5(x0,x1,x2,x3,x4)                              \
 113         x4  = x1;       x1 |= x0;                       \
 114         x2 ^= x1;       x3 =~ x3;       x4 ^= x0;       \
 115         x0 ^= x2;       x1 &= x4;       x4 |= x3;       \
 116         x4 ^= x0;       x0 &= x3;       x1 ^= x3;       \
 117         x3 ^= x2;       x0 ^= x1;       x2 &= x4;       \
 118         x1 ^= x2;       x2 &= x0;                       \
 119         x3 ^= x2;
 120
 121 #define S6(x0,x1,x2,x3,x4)                              \
 122                                         x4  = x1;       \
 123         x3 ^= x0;       x1 ^= x2;       x2 ^= x0;       \
 124         x0 &= x3;       x1 |= x3;       x4 =~ x4;       \
 125         x0 ^= x1;       x1 ^= x2;                       \
 126         x3 ^= x4;       x4 ^= x0;       x2 &= x0;       \
 127         x4 ^= x1;       x2 ^= x3;       x3 &= x1;       \
 128         x3 ^= x0;       x1 ^= x2;
 129
 130 #define S7(x0,x1,x2,x3,x4)                              \
 131                                         x1 =~ x1;       \
 132         x4  = x1;       x0 =~ x0;       x1 &= x2;       \
 133         x1 ^= x3;       x3 |= x4;       x4 ^= x2;       \
 134         x2 ^= x3;       x3 ^= x0;       x0 |= x1;       \
 135         x2 &= x0;       x0 ^= x4;       x4 ^= x3;       \
 136         x3 &= x0;       x4 ^= x1;                       \
 137         x2 ^= x4;       x3 ^= x1;       x4 |= x0;       \
 138         x4 ^= x1;
 139
 140 #define SI0(x0,x1,x2,x3,x4)                             \
 141                         x4  = x3;       x1 ^= x0;       \
 142         x3 |= x1;       x4 ^= x1;       x0 =~ x0;       \
 143         x2 ^= x3;       x3 ^= x0;       x0 &= x1;       \
 144         x0 ^= x2;       x2 &= x3;       x3 ^= x4;       \
 145         x2 ^= x3;       x1 ^= x3;       x3 &= x0;       \
 146         x1 ^= x0;       x0 ^= x2;       x4 ^= x3;
 147
 148 #define SI1(x0,x1,x2,x3,x4)                             \
 149         x1 ^= x3;       x4  = x0;                       \
 150         x0 ^= x2;       x2 =~ x2;       x4 |= x1;       \
 151         x4 ^= x3;       x3 &= x1;       x1 ^= x2;       \
 152         x2 &= x4;       x4 ^= x1;       x1 |= x3;       \
 153         x3 ^= x0;       x2 ^= x0;       x0 |= x4;       \
 154         x2 ^= x4;       x1 ^= x0;                       \
 155         x4 ^= x1;
 156
 157 #define SI2(x0,x1,x2,x3,x4)                             \
 158         x2 ^= x1;       x4  = x3;       x3 =~ x3;       \
 159         x3 |= x2;       x2 ^= x4;       x4 ^= x0;       \
 160         x3 ^= x1;       x1 |= x2;       x2 ^= x0;       \
 161         x1 ^= x4;       x4 |= x3;       x2 ^= x3;       \
 162         x4 ^= x2;       x2 &= x1;                       \
 163         x2 ^= x3;       x3 ^= x4;       x4 ^= x0;
 164
 165 #define SI3(x0,x1,x2,x3,x4)                             \
 166                                         x2 ^= x1;       \
 167         x4  = x1;       x1 &= x2;                       \
 168         x1 ^= x0;       x0 |= x4;       x4 ^= x3;       \
 169         x0 ^= x3;       x3 |= x1;       x1 ^= x2;       \
 170         x1 ^= x3;       x0 ^= x2;       x2 ^= x3;       \
 171         x3 &= x1;       x1 ^= x0;       x0 &= x2;       \
 172         x4 ^= x3;       x3 ^= x0;       x0 ^= x1;
 173
 174 #define SI4(x0,x1,x2,x3,x4)                             \
 175         x2 ^= x3;       x4  = x0;       x0 &= x1;       \
 176         x0 ^= x2;       x2 |= x3;       x4 =~ x4;       \
 177         x1 ^= x0;       x0 ^= x2;       x2 &= x4;       \
 178         x2 ^= x0;       x0 |= x4;                       \
 179         x0 ^= x3;       x3 &= x2;                       \
 180         x4 ^= x3;       x3 ^= x1;       x1 &= x0;       \
 181         x4 ^= x1;       x0 ^= x3;
 182
 183 #define SI5(x0,x1,x2,x3,x4)                             \
 184                         x4  = x1;       x1 |= x2;       \
 185         x2 ^= x4;       x1 ^= x3;       x3 &= x4;       \
 186         x2 ^= x3;       x3 |= x0;       x0 =~ x0;       \
 187         x3 ^= x2;       x2 |= x0;       x4 ^= x1;       \
 188         x2 ^= x4;       x4 &= x0;       x0 ^= x1;       \
 189         x1 ^= x3;       x0 &= x2;       x2 ^= x3;       \
 190         x0 ^= x2;       x2 ^= x4;       x4 ^= x3;
 191
 192 #define SI6(x0,x1,x2,x3,x4)                             \
 193                         x0 ^= x2;                       \
 194         x4  = x0;       x0 &= x3;       x2 ^= x3;       \
 195         x0 ^= x2;       x3 ^= x1;       x2 |= x4;       \
 196         x2 ^= x3;       x3 &= x0;       x0 =~ x0;       \
 197         x3 ^= x1;       x1 &= x2;       x4 ^= x0;       \
 198         x3 ^= x4;       x4 ^= x2;       x0 ^= x1;       \
 199         x2 ^= x0;
 200
 201 #define SI7(x0,x1,x2,x3,x4)                             \
 202         x4  = x3;       x3 &= x0;       x0 ^= x2;       \
 203         x2 |= x4;       x4 ^= x1;       x0 =~ x0;       \
 204         x1 |= x3;       x4 ^= x0;       x0 &= x2;       \
 205         x0 ^= x1;       x1 &= x2;       x3 ^= x2;       \
 206         x4 ^= x3;       x2 &= x3;       x3 |= x0;       \
 207         x1 ^= x4;       x3 ^= x4;       x4 &= x0;       \
 208         x4 ^= x2;
 209
 210 struct serpent_ctx {
 211         u8 iv[SERPENT_BLOCK_SIZE];
 212         u32 expkey[SERPENT_EXPKEY_WORDS];
 213 };
 214
 215 static int setkey(void *ctx, const u8 *key, unsigned int keylen, u32 *flags)
 216 {
 217         u32 *k = ((struct serpent_ctx *)ctx)->expkey;
 218         u8  *k8 = (u8 *)k;
 219         u32 r0,r1,r2,r3,r4;
 220         int i;
 221
 222         if ((keylen < SERPENT_MIN_KEY_SIZE)
 223                         || (keylen > SERPENT_MAX_KEY_SIZE))
 224         {
 225                 *flags |= CRYPTO_TFM_RES_BAD_KEY_LEN;
 226                 return -EINVAL;
 227         }
 228
 229         /* Copy key, add padding */
 230
 231         for (i = 0; i < keylen; ++i)
 232                 k8[i] = key[i];
 233         if (i < SERPENT_MAX_KEY_SIZE)
 234                 k8[i++] = 1;
 235         while (i < SERPENT_MAX_KEY_SIZE)
 236                 k8[i++] = 0;
 237
 238         /* Expand key using polynomial */
 239
 240         r0 = le32_to_cpu(k[3]);
 241         r1 = le32_to_cpu(k[4]);
 242         r2 = le32_to_cpu(k[5]);
 243         r3 = le32_to_cpu(k[6]);
 244         r4 = le32_to_cpu(k[7]);
 245
 246         keyiter(le32_to_cpu(k[0]),r0,r4,r2,0,0);
 247         keyiter(le32_to_cpu(k[1]),r1,r0,r3,1,1);
 248         keyiter(le32_to_cpu(k[2]),r2,r1,r4,2,2);
 249         keyiter(le32_to_cpu(k[3]),r3,r2,r0,3,3);
 250         keyiter(le32_to_cpu(k[4]),r4,r3,r1,4,4);
 251         keyiter(le32_to_cpu(k[5]),r0,r4,r2,5,5);
 252         keyiter(le32_to_cpu(k[6]),r1,r0,r3,6,6);
 253         keyiter(le32_to_cpu(k[7]),r2,r1,r4,7,7);
 254
 255         keyiter(k[  0],r3,r2,r0,  8,  8); keyiter(k[  1],r4,r3,r1,  9,  9);
 256         keyiter(k[  2],r0,r4,r2, 10, 10); keyiter(k[  3],r1,r0,r3, 11, 11);
 257         keyiter(k[  4],r2,r1,r4, 12, 12); keyiter(k[  5],r3,r2,r0, 13, 13);
 258         keyiter(k[  6],r4,r3,r1, 14, 14); keyiter(k[  7],r0,r4,r2, 15, 15);
 259         keyiter(k[  8],r1,r0,r3, 16, 16); keyiter(k[  9],r2,r1,r4, 17, 17);
 260         keyiter(k[ 10],r3,r2,r0, 18, 18); keyiter(k[ 11],r4,r3,r1, 19, 19);
 261         keyiter(k[ 12],r0,r4,r2, 20, 20); keyiter(k[ 13],r1,r0,r3, 21, 21);
 262         keyiter(k[ 14],r2,r1,r4, 22, 22); keyiter(k[ 15],r3,r2,r0, 23, 23);
 263         keyiter(k[ 16],r4,r3,r1, 24, 24); keyiter(k[ 17],r0,r4,r2, 25, 25);
 264         keyiter(k[ 18],r1,r0,r3, 26, 26); keyiter(k[ 19],r2,r1,r4, 27, 27);
 265         keyiter(k[ 20],r3,r2,r0, 28, 28); keyiter(k[ 21],r4,r3,r1, 29, 29);
 266         keyiter(k[ 22],r0,r4,r2, 30, 30); keyiter(k[ 23],r1,r0,r3, 31, 31);
 267
 268         k += 50;
 269
 270         keyiter(k[-26],r2,r1,r4, 32,-18); keyiter(k[-25],r3,r2,r0, 33,-17);
 271         keyiter(k[-24],r4,r3,r1, 34,-16); keyiter(k[-23],r0,r4,r2, 35,-15);
 272         keyiter(k[-22],r1,r0,r3, 36,-14); keyiter(k[-21],r2,r1,r4, 37,-13);
 273         keyiter(k[-20],r3,r2,r0, 38,-12); keyiter(k[-19],r4,r3,r1, 39,-11);
 274         keyiter(k[-18],r0,r4,r2, 40,-10); keyiter(k[-17],r1,r0,r3, 41, -9);
 275         keyiter(k[-16],r2,r1,r4, 42, -8); keyiter(k[-15],r3,r2,r0, 43, -7);
 276         keyiter(k[-14],r4,r3,r1, 44, -6); keyiter(k[-13],r0,r4,r2, 45, -5);
 277         keyiter(k[-12],r1,r0,r3, 46, -4); keyiter(k[-11],r2,r1,r4, 47, -3);
 278         keyiter(k[-10],r3,r2,r0, 48, -2); keyiter(k[ -9],r4,r3,r1, 49, -1);
 279         keyiter(k[ -8],r0,r4,r2, 50,  0); keyiter(k[ -7],r1,r0,r3, 51,  1);
 280         keyiter(k[ -6],r2,r1,r4, 52,  2); keyiter(k[ -5],r3,r2,r0, 53,  3);
 281         keyiter(k[ -4],r4,r3,r1, 54,  4); keyiter(k[ -3],r0,r4,r2, 55,  5);
 282         keyiter(k[ -2],r1,r0,r3, 56,  6); keyiter(k[ -1],r2,r1,r4, 57,  7);
 283         keyiter(k[  0],r3,r2,r0, 58,  8); keyiter(k[  1],r4,r3,r1, 59,  9);
 284         keyiter(k[  2],r0,r4,r2, 60, 10); keyiter(k[  3],r1,r0,r3, 61, 11);
 285         keyiter(k[  4],r2,r1,r4, 62, 12); keyiter(k[  5],r3,r2,r0, 63, 13);
 286         keyiter(k[  6],r4,r3,r1, 64, 14); keyiter(k[  7],r0,r4,r2, 65, 15);
 287         keyiter(k[  8],r1,r0,r3, 66, 16); keyiter(k[  9],r2,r1,r4, 67, 17);
 288         keyiter(k[ 10],r3,r2,r0, 68, 18); keyiter(k[ 11],r4,r3,r1, 69, 19);
 289         keyiter(k[ 12],r0,r4,r2, 70, 20); keyiter(k[ 13],r1,r0,r3, 71, 21);
 290         keyiter(k[ 14],r2,r1,r4, 72, 22); keyiter(k[ 15],r3,r2,r0, 73, 23);
 291         keyiter(k[ 16],r4,r3,r1, 74, 24); keyiter(k[ 17],r0,r4,r2, 75, 25);
 292         keyiter(k[ 18],r1,r0,r3, 76, 26); keyiter(k[ 19],r2,r1,r4, 77, 27);
 293         keyiter(k[ 20],r3,r2,r0, 78, 28); keyiter(k[ 21],r4,r3,r1, 79, 29);
 294         keyiter(k[ 22],r0,r4,r2, 80, 30); keyiter(k[ 23],r1,r0,r3, 81, 31);
 295
 296         k += 50;
 297
 298         keyiter(k[-26],r2,r1,r4, 82,-18); keyiter(k[-25],r3,r2,r0, 83,-17);
 299         keyiter(k[-24],r4,r3,r1, 84,-16); keyiter(k[-23],r0,r4,r2, 85,-15);
 300         keyiter(k[-22],r1,r0,r3, 86,-14); keyiter(k[-21],r2,r1,r4, 87,-13);
 301         keyiter(k[-20],r3,r2,r0, 88,-12); keyiter(k[-19],r4,r3,r1, 89,-11);
 302         keyiter(k[-18],r0,r4,r2, 90,-10); keyiter(k[-17],r1,r0,r3, 91, -9);
 303         keyiter(k[-16],r2,r1,r4, 92, -8); keyiter(k[-15],r3,r2,r0, 93, -7);
 304         keyiter(k[-14],r4,r3,r1, 94, -6); keyiter(k[-13],r0,r4,r2, 95, -5);
 305         keyiter(k[-12],r1,r0,r3, 96, -4); keyiter(k[-11],r2,r1,r4, 97, -3);
 306         keyiter(k[-10],r3,r2,r0, 98, -2); keyiter(k[ -9],r4,r3,r1, 99, -1);
 307         keyiter(k[ -8],r0,r4,r2,100,  0); keyiter(k[ -7],r1,r0,r3,101,  1);
 308         keyiter(k[ -6],r2,r1,r4,102,  2); keyiter(k[ -5],r3,r2,r0,103,  3);
 309         keyiter(k[ -4],r4,r3,r1,104,  4); keyiter(k[ -3],r0,r4,r2,105,  5);
 310         keyiter(k[ -2],r1,r0,r3,106,  6); keyiter(k[ -1],r2,r1,r4,107,  7);
 311         keyiter(k[  0],r3,r2,r0,108,  8); keyiter(k[  1],r4,r3,r1,109,  9);
 312         keyiter(k[  2],r0,r4,r2,110, 10); keyiter(k[  3],r1,r0,r3,111, 11);
 313         keyiter(k[  4],r2,r1,r4,112, 12); keyiter(k[  5],r3,r2,r0,113, 13);
 314         keyiter(k[  6],r4,r3,r1,114, 14); keyiter(k[  7],r0,r4,r2,115, 15);
 315         keyiter(k[  8],r1,r0,r3,116, 16); keyiter(k[  9],r2,r1,r4,117, 17);
 316         keyiter(k[ 10],r3,r2,r0,118, 18); keyiter(k[ 11],r4,r3,r1,119, 19);
 317         keyiter(k[ 12],r0,r4,r2,120, 20); keyiter(k[ 13],r1,r0,r3,121, 21);
 318         keyiter(k[ 14],r2,r1,r4,122, 22); keyiter(k[ 15],r3,r2,r0,123, 23);
 319         keyiter(k[ 16],r4,r3,r1,124, 24); keyiter(k[ 17],r0,r4,r2,125, 25);
 320         keyiter(k[ 18],r1,r0,r3,126, 26); keyiter(k[ 19],r2,r1,r4,127, 27);
 321         keyiter(k[ 20],r3,r2,r0,128, 28); keyiter(k[ 21],r4,r3,r1,129, 29);
 322         keyiter(k[ 22],r0,r4,r2,130, 30); keyiter(k[ 23],r1,r0,r3,131, 31);
 323
 324         /* Apply S-boxes */
 325
 326         S3(r3,r4,r0,r1,r2); storekeys(r1,r2,r4,r3, 28); loadkeys(r1,r2,r4,r3, 24);
 327         S4(r1,r2,r4,r3,r0); storekeys(r2,r4,r3,r0, 24); loadkeys(r2,r4,r3,r0, 20);
 328         S5(r2,r4,r3,r0,r1); storekeys(r1,r2,r4,r0, 20); loadkeys(r1,r2,r4,r0, 16);
 329         S6(r1,r2,r4,r0,r3); storekeys(r4,r3,r2,r0, 16); loadkeys(r4,r3,r2,r0, 12);
 330         S7(r4,r3,r2,r0,r1); storekeys(r1,r2,r0,r4, 12); loadkeys(r1,r2,r0,r4,  8);
 331         S0(r1,r2,r0,r4,r3); storekeys(r0,r2,r4,r1,  8); loadkeys(r0,r2,r4,r1,  4);
 332         S1(r0,r2,r4,r1,r3); storekeys(r3,r4,r1,r0,  4); loadkeys(r3,r4,r1,r0,  0);
 333         S2(r3,r4,r1,r0,r2); storekeys(r2,r4,r3,r0,  0); loadkeys(r2,r4,r3,r0, -4);
 334         S3(r2,r4,r3,r0,r1); storekeys(r0,r1,r4,r2, -4); loadkeys(r0,r1,r4,r2, -8);
 335         S4(r0,r1,r4,r2,r3); storekeys(r1,r4,r2,r3, -8); loadkeys(r1,r4,r2,r3,-12);
 336         S5(r1,r4,r2,r3,r0); storekeys(r0,r1,r4,r3,-12); loadkeys(r0,r1,r4,r3,-16);
 337         S6(r0,r1,r4,r3,r2); storekeys(r4,r2,r1,r3,-16); loadkeys(r4,r2,r1,r3,-20);
 338         S7(r4,r2,r1,r3,r0); storekeys(r0,r1,r3,r4,-20); loadkeys(r0,r1,r3,r4,-24);
 339         S0(r0,r1,r3,r4,r2); storekeys(r3,r1,r4,r0,-24); loadkeys(r3,r1,r4,r0,-28);
 340         k -= 50;
 341         S1(r3,r1,r4,r0,r2); storekeys(r2,r4,r0,r3, 22); loadkeys(r2,r4,r0,r3, 18);
 342         S2(r2,r4,r0,r3,r1); storekeys(r1,r4,r2,r3, 18); loadkeys(r1,r4,r2,r3, 14);
 343         S3(r1,r4,r2,r3,r0); storekeys(r3,r0,r4,r1, 14); loadkeys(r3,r0,r4,r1, 10);
 344         S4(r3,r0,r4,r1,r2); storekeys(r0,r4,r1,r2, 10); loadkeys(r0,r4,r1,r2,  6);
 345         S5(r0,r4,r1,r2,r3); storekeys(r3,r0,r4,r2,  6); loadkeys(r3,r0,r4,r2,  2);
 346         S6(r3,r0,r4,r2,r1); storekeys(r4,r1,r0,r2,  2); loadkeys(r4,r1,r0,r2, -2);
 347         S7(r4,r1,r0,r2,r3); storekeys(r3,r0,r2,r4, -2); loadkeys(r3,r0,r2,r4, -6);
 348         S0(r3,r0,r2,r4,r1); storekeys(r2,r0,r4,r3, -6); loadkeys(r2,r0,r4,r3,-10);
 349         S1(r2,r0,r4,r3,r1); storekeys(r1,r4,r3,r2,-10); loadkeys(r1,r4,r3,r2,-14);
 350         S2(r1,r4,r3,r2,r0); storekeys(r0,r4,r1,r2,-14); loadkeys(r0,r4,r1,r2,-18);
 351         S3(r0,r4,r1,r2,r3); storekeys(r2,r3,r4,r0,-18); loadkeys(r2,r3,r4,r0,-22);
 352         k -= 50;
 353         S4(r2,r3,r4,r0,r1); storekeys(r3,r4,r0,r1, 28); loadkeys(r3,r4,r0,r1, 24);
 354         S5(r3,r4,r0,r1,r2); storekeys(r2,r3,r4,r1, 24); loadkeys(r2,r3,r4,r1, 20);
 355         S6(r2,r3,r4,r1,r0); storekeys(r4,r0,r3,r1, 20); loadkeys(r4,r0,r3,r1, 16);
 356         S7(r4,r0,r3,r1,r2); storekeys(r2,r3,r1,r4, 16); loadkeys(r2,r3,r1,r4, 12);
 357         S0(r2,r3,r1,r4,r0); storekeys(r1,r3,r4,r2, 12); loadkeys(r1,r3,r4,r2,  8);
 358         S1(r1,r3,r4,r2,r0); storekeys(r0,r4,r2,r1,  8); loadkeys(r0,r4,r2,r1,  4);
 359         S2(r0,r4,r2,r1,r3); storekeys(r3,r4,r0,r1,  4); loadkeys(r3,r4,r0,r1,  0);
 360         S3(r3,r4,r0,r1,r2); storekeys(r1,r2,r4,r3,  0);
 361
 362         return 0;
 363 }
 364
 365 static void encrypt(void *ctx, u8 *dst, const u8 *src)
 366 {
 367         const u32
 368                 *k = ((struct serpent_ctx *)ctx)->expkey,
 369                 *s = (const u32 *)src;
 370         u32     *d = (u32 *)dst,
 371                 r0, r1, r2, r3, r4;
 372
 373 /*
 374  * Note: The conversions between u8* and u32* might cause trouble
 375  * on architectures with stricter alignment rules than x86
 376  */
 377
 378         r0 = le32_to_cpu(s[0]);
 379         r1 = le32_to_cpu(s[1]);
 380         r2 = le32_to_cpu(s[2]);
 381         r3 = le32_to_cpu(s[3]);
 382
 383                                  K(r0,r1,r2,r3,0);
 384         S0(r0,r1,r2,r3,r4);     LK(r2,r1,r3,r0,r4,1);
 385         S1(r2,r1,r3,r0,r4);     LK(r4,r3,r0,r2,r1,2);
 386         S2(r4,r3,r0,r2,r1);     LK(r1,r3,r4,r2,r0,3);
 387         S3(r1,r3,r4,r2,r0);     LK(r2,r0,r3,r1,r4,4);
 388         S4(r2,r0,r3,r1,r4);     LK(r0,r3,r1,r4,r2,5);
 389         S5(r0,r3,r1,r4,r2);     LK(r2,r0,r3,r4,r1,6);
 390         S6(r2,r0,r3,r4,r1);     LK(r3,r1,r0,r4,r2,7);
 391         S7(r3,r1,r0,r4,r2);     LK(r2,r0,r4,r3,r1,8);
 392         S0(r2,r0,r4,r3,r1);     LK(r4,r0,r3,r2,r1,9);
 393         S1(r4,r0,r3,r2,r1);     LK(r1,r3,r2,r4,r0,10);
 394         S2(r1,r3,r2,r4,r0);     LK(r0,r3,r1,r4,r2,11);
 395         S3(r0,r3,r1,r4,r2);     LK(r4,r2,r3,r0,r1,12);
 396         S4(r4,r2,r3,r0,r1);     LK(r2,r3,r0,r1,r4,13);
 397         S5(r2,r3,r0,r1,r4);     LK(r4,r2,r3,r1,r0,14);
 398         S6(r4,r2,r3,r1,r0);     LK(r3,r0,r2,r1,r4,15);
 399         S7(r3,r0,r2,r1,r4);     LK(r4,r2,r1,r3,r0,16);
 400         S0(r4,r2,r1,r3,r0);     LK(r1,r2,r3,r4,r0,17);
 401         S1(r1,r2,r3,r4,r0);     LK(r0,r3,r4,r1,r2,18);
 402         S2(r0,r3,r4,r1,r2);     LK(r2,r3,r0,r1,r4,19);
 403         S3(r2,r3,r0,r1,r4);     LK(r1,r4,r3,r2,r0,20);
 404         S4(r1,r4,r3,r2,r0);     LK(r4,r3,r2,r0,r1,21);
 405         S5(r4,r3,r2,r0,r1);     LK(r1,r4,r3,r0,r2,22);
 406         S6(r1,r4,r3,r0,r2);     LK(r3,r2,r4,r0,r1,23);
 407         S7(r3,r2,r4,r0,r1);     LK(r1,r4,r0,r3,r2,24);
 408         S0(r1,r4,r0,r3,r2);     LK(r0,r4,r3,r1,r2,25);
 409         S1(r0,r4,r3,r1,r2);     LK(r2,r3,r1,r0,r4,26);
 410         S2(r2,r3,r1,r0,r4);     LK(r4,r3,r2,r0,r1,27);
 411         S3(r4,r3,r2,r0,r1);     LK(r0,r1,r3,r4,r2,28);
 412         S4(r0,r1,r3,r4,r2);     LK(r1,r3,r4,r2,r0,29);
 413         S5(r1,r3,r4,r2,r0);     LK(r0,r1,r3,r2,r4,30);
 414         S6(r0,r1,r3,r2,r4);     LK(r3,r4,r1,r2,r0,31);
 415         S7(r3,r4,r1,r2,r0);      K(r0,r1,r2,r3,32);
 416
 417         d[0] = cpu_to_le32(r0);
 418         d[1] = cpu_to_le32(r1);
 419         d[2] = cpu_to_le32(r2);
 420         d[3] = cpu_to_le32(r3);
 421 }
 422
 423 static void decrypt(void *ctx, u8 *dst, const u8 *src)
 424 {
 425         const u32
 426                 *k = ((struct serpent_ctx *)ctx)->expkey,
 427                 *s = (const u32 *)src;
 428         u32     *d = (u32 *)dst,
 429                 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_module             =       THIS_MODULE,
 482         .cra_list               =       LIST_HEAD_INIT(serpent_alg.cra_list),
 483         .cra_u                  =       { .cipher = {
 484         .cia_min_keysize        =       SERPENT_MIN_KEY_SIZE,
 485         .cia_max_keysize        =       SERPENT_MAX_KEY_SIZE,
 486         .cia_setkey             =       setkey,
 487         .cia_encrypt            =       encrypt,
 488         .cia_decrypt            =       decrypt } }
 489 };
 490
 491 static int __init init(void)
 492 {
 493         return crypto_register_alg(&serpent_alg);
 494 }
 495
 496 static void __exit fini(void)
 497 {
 498         crypto_unregister_alg(&serpent_alg);
 499 }
 500
 501 module_init(init);
 502 module_exit(fini);
 503
 504 MODULE_LICENSE("GPL");
 505 MODULE_DESCRIPTION("Serpent Cipher Algorithm");
 506 MODULE_AUTHOR("Dag Arne Osvik <osvik@ii.uib.no>");