crypto/serpent_generic.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Cryptographic API.
   4  *
   5  * Serpent Cipher Algorithm.
   6  *
   7  * Copyright (C) 2002 Dag Arne Osvik <osvik@ii.uib.no>
   8  *               2003 Herbert Valerio Riedel <hvr@gnu.org>
   9  *
  10  * Added tnepres support:
  11  *              Ruben Jesus Garcia Hernandez <ruben@ugr.es>, 18.10.2004
  12  *              Based on code by hvr
  13  */
  14
  15 #include <linux/init.h>
  16 #include <linux/module.h>
  17 #include <linux/errno.h>
  18 #include <asm/byteorder.h>
  19 #include <linux/crypto.h>
  20 #include <linux/types.h>
  21 #include <crypto/serpent.h>
  22
  23 /* Key is padded to the maximum of 256 bits before round key generation.
  24  * Any key length <= 256 bits (32 bytes) is allowed by the algorithm.
  25  */
  26
  27 #define PHI 0x9e3779b9UL
  28
  29 #define keyiter(a, b, c, d, i, j) \
  30         ({ b ^= d; b ^= c; b ^= a; b ^= PHI ^ i; b = rol32(b, 11); k[j] = b; })
  31
  32 #define loadkeys(x0, x1, x2, x3, i) \
  33         ({ x0 = k[i]; x1 = k[i+1]; x2 = k[i+2]; x3 = k[i+3]; })
  34
  35 #define storekeys(x0, x1, x2, x3, i) \
  36         ({ k[i] = x0; k[i+1] = x1; k[i+2] = x2; k[i+3] = x3; })
  37
  38 #define store_and_load_keys(x0, x1, x2, x3, s, l) \
  39         ({ storekeys(x0, x1, x2, x3, s); loadkeys(x0, x1, x2, x3, l); })
  40
  41 #define K(x0, x1, x2, x3, i) ({                         \
  42         x3 ^= k[4*(i)+3];        x2 ^= k[4*(i)+2];      \
  43         x1 ^= k[4*(i)+1];        x0 ^= k[4*(i)+0];      \
  44         })
  45
  46 #define LK(x0, x1, x2, x3, x4, i) ({                                       \
  47                                                         x0 = rol32(x0, 13);\
  48         x2 = rol32(x2, 3);      x1 ^= x0;               x4  = x0 << 3;     \
  49         x3 ^= x2;               x1 ^= x2;                                  \
  50         x1 = rol32(x1, 1);      x3 ^= x4;                                  \
  51         x3 = rol32(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];         x0 = rol32(x0, 5);      x2 = rol32(x2, 22);\
  55         x0 ^= k[4*i+0];         x2 ^= k[4*i+2];                            \
  56         })
  57
  58 #define KL(x0, x1, x2, x3, x4, i) ({                                       \
  59         x0 ^= k[4*i+0];         x1 ^= k[4*i+1];         x2 ^= k[4*i+2];    \
  60         x3 ^= k[4*i+3];         x0 = ror32(x0, 5);      x2 = ror32(x2, 22);\
  61         x4 =  x1;               x2 ^= x3;               x0 ^= x3;          \
  62         x4 <<= 7;               x0 ^= x1;               x1 = ror32(x1, 1); \
  63         x2 ^= x4;               x3 = ror32(x3, 7);      x4 = x0 << 3;      \
  64         x1 ^= x0;               x3 ^= x4;               x0 = ror32(x0, 13);\
  65         x1 ^= x2;               x3 ^= x2;               x2 = ror32(x2, 3); \
  66         })
  67
  68 #define S0(x0, x1, x2, x3, x4) ({                       \
  69                                         x4  = x3;       \
  70         x3 |= x0;       x0 ^= x4;       x4 ^= x2;       \
  71         x4 = ~x4;       x3 ^= x1;       x1 &= x0;       \
  72         x1 ^= x4;       x2 ^= x0;       x0 ^= x3;       \
  73         x4 |= x0;       x0 ^= x2;       x2 &= x1;       \
  74         x3 ^= x2;       x1 = ~x1;       x2 ^= x4;       \
  75         x1 ^= x2;                                       \
  76         })
  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
  88 #define S2(x0, x1, x2, x3, x4) ({                       \
  89                                         x3 = ~x3;       \
  90         x1 ^= x0;       x4  = x0;       x0 &= x2;       \
  91         x0 ^= x3;       x3 |= x4;       x2 ^= x1;       \
  92         x3 ^= x1;       x1 &= x0;       x0 ^= x2;       \
  93         x2 &= x3;       x3 |= x1;       x0 = ~x0;       \
  94         x3 ^= x0;       x4 ^= x0;       x0 ^= x2;       \
  95         x1 |= x2;                                       \
  96         })
  97
  98 #define S3(x0, x1, x2, x3, x4) ({                       \
  99                                         x4  = x1;       \
 100         x1 ^= x3;       x3 |= x0;       x4 &= x0;       \
 101         x0 ^= x2;       x2 ^= x1;       x1 &= x3;       \
 102         x2 ^= x3;       x0 |= x4;       x4 ^= x3;       \
 103         x1 ^= x0;       x0 &= x3;       x3 &= x4;       \
 104         x3 ^= x2;       x4 |= x1;       x2 &= x1;       \
 105         x4 ^= x3;       x0 ^= x3;       x3 ^= x2;       \
 106         })
 107
 108 #define S4(x0, x1, x2, x3, x4) ({                       \
 109                                         x4  = x3;       \
 110         x3 &= x0;       x0 ^= x4;                       \
 111         x3 ^= x2;       x2 |= x4;       x0 ^= x1;       \
 112         x4 ^= x3;       x2 |= x0;                       \
 113         x2 ^= x1;       x1 &= x0;                       \
 114         x1 ^= x4;       x4 &= x2;       x2 ^= x3;       \
 115         x4 ^= x0;       x3 |= x1;       x1 = ~x1;       \
 116         x3 ^= x0;                                       \
 117         })
 118
 119 #define S5(x0, x1, x2, x3, x4) ({                       \
 120         x4  = x1;       x1 |= x0;                       \
 121         x2 ^= x1;       x3 = ~x3;       x4 ^= x0;       \
 122         x0 ^= x2;       x1 &= x4;       x4 |= x3;       \
 123         x4 ^= x0;       x0 &= x3;       x1 ^= x3;       \
 124         x3 ^= x2;       x0 ^= x1;       x2 &= x4;       \
 125         x1 ^= x2;       x2 &= x0;                       \
 126         x3 ^= x2;                                       \
 127         })
 128
 129 #define S6(x0, x1, x2, x3, x4) ({                       \
 130                                         x4  = x1;       \
 131         x3 ^= x0;       x1 ^= x2;       x2 ^= x0;       \
 132         x0 &= x3;       x1 |= x3;       x4 = ~x4;       \
 133         x0 ^= x1;       x1 ^= x2;                       \
 134         x3 ^= x4;       x4 ^= x0;       x2 &= x0;       \
 135         x4 ^= x1;       x2 ^= x3;       x3 &= x1;       \
 136         x3 ^= x0;       x1 ^= x2;                       \
 137         })
 138
 139 #define S7(x0, x1, x2, x3, x4) ({                       \
 140                                         x1 = ~x1;       \
 141         x4  = x1;       x0 = ~x0;       x1 &= x2;       \
 142         x1 ^= x3;       x3 |= x4;       x4 ^= x2;       \
 143         x2 ^= x3;       x3 ^= x0;       x0 |= x1;       \
 144         x2 &= x0;       x0 ^= x4;       x4 ^= x3;       \
 145         x3 &= x0;       x4 ^= x1;                       \
 146         x2 ^= x4;       x3 ^= x1;       x4 |= x0;       \
 147         x4 ^= x1;                                       \
 148         })
 149
 150 #define SI0(x0, x1, x2, x3, x4) ({                      \
 151                         x4  = x3;       x1 ^= x0;       \
 152         x3 |= x1;       x4 ^= x1;       x0 = ~x0;       \
 153         x2 ^= x3;       x3 ^= x0;       x0 &= x1;       \
 154         x0 ^= x2;       x2 &= x3;       x3 ^= x4;       \
 155         x2 ^= x3;       x1 ^= x3;       x3 &= x0;       \
 156         x1 ^= x0;       x0 ^= x2;       x4 ^= x3;       \
 157         })
 158
 159 #define SI1(x0, x1, x2, x3, x4) ({                      \
 160         x1 ^= x3;       x4  = x0;                       \
 161         x0 ^= x2;       x2 = ~x2;       x4 |= x1;       \
 162         x4 ^= x3;       x3 &= x1;       x1 ^= x2;       \
 163         x2 &= x4;       x4 ^= x1;       x1 |= x3;       \
 164         x3 ^= x0;       x2 ^= x0;       x0 |= x4;       \
 165         x2 ^= x4;       x1 ^= x0;                       \
 166         x4 ^= x1;                                       \
 167         })
 168
 169 #define SI2(x0, x1, x2, x3, x4) ({                      \
 170         x2 ^= x1;       x4  = x3;       x3 = ~x3;       \
 171         x3 |= x2;       x2 ^= x4;       x4 ^= x0;       \
 172         x3 ^= x1;       x1 |= x2;       x2 ^= x0;       \
 173         x1 ^= x4;       x4 |= x3;       x2 ^= x3;       \
 174         x4 ^= x2;       x2 &= x1;                       \
 175         x2 ^= x3;       x3 ^= x4;       x4 ^= x0;       \
 176         })
 177
 178 #define SI3(x0, x1, x2, x3, x4) ({                      \
 179                                         x2 ^= x1;       \
 180         x4  = x1;       x1 &= x2;                       \
 181         x1 ^= x0;       x0 |= x4;       x4 ^= x3;       \
 182         x0 ^= x3;       x3 |= x1;       x1 ^= x2;       \
 183         x1 ^= x3;       x0 ^= x2;       x2 ^= x3;       \
 184         x3 &= x1;       x1 ^= x0;       x0 &= x2;       \
 185         x4 ^= x3;       x3 ^= x0;       x0 ^= x1;       \
 186         })
 187
 188 #define SI4(x0, x1, x2, x3, x4) ({                      \
 189         x2 ^= x3;       x4  = x0;       x0 &= x1;       \
 190         x0 ^= x2;       x2 |= x3;       x4 = ~x4;       \
 191         x1 ^= x0;       x0 ^= x2;       x2 &= x4;       \
 192         x2 ^= x0;       x0 |= x4;                       \
 193         x0 ^= x3;       x3 &= x2;                       \
 194         x4 ^= x3;       x3 ^= x1;       x1 &= x0;       \
 195         x4 ^= x1;       x0 ^= x3;                       \
 196         })
 197
 198 #define SI5(x0, x1, x2, x3, x4) ({                      \
 199                         x4  = x1;       x1 |= x2;       \
 200         x2 ^= x4;       x1 ^= x3;       x3 &= x4;       \
 201         x2 ^= x3;       x3 |= x0;       x0 = ~x0;       \
 202         x3 ^= x2;       x2 |= x0;       x4 ^= x1;       \
 203         x2 ^= x4;       x4 &= x0;       x0 ^= x1;       \
 204         x1 ^= x3;       x0 &= x2;       x2 ^= x3;       \
 205         x0 ^= x2;       x2 ^= x4;       x4 ^= x3;       \
 206         })
 207
 208 #define SI6(x0, x1, x2, x3, x4) ({                      \
 209                         x0 ^= x2;                       \
 210         x4  = x0;       x0 &= x3;       x2 ^= x3;       \
 211         x0 ^= x2;       x3 ^= x1;       x2 |= x4;       \
 212         x2 ^= x3;       x3 &= x0;       x0 = ~x0;       \
 213         x3 ^= x1;       x1 &= x2;       x4 ^= x0;       \
 214         x3 ^= x4;       x4 ^= x2;       x0 ^= x1;       \
 215         x2 ^= x0;                                       \
 216         })
 217
 218 #define SI7(x0, x1, x2, x3, x4) ({                      \
 219         x4  = x3;       x3 &= x0;       x0 ^= x2;       \
 220         x2 |= x4;       x4 ^= x1;       x0 = ~x0;       \
 221         x1 |= x3;       x4 ^= x0;       x0 &= x2;       \
 222         x0 ^= x1;       x1 &= x2;       x3 ^= x2;       \
 223         x4 ^= x3;       x2 &= x3;       x3 |= x0;       \
 224         x1 ^= x4;       x3 ^= x4;       x4 &= x0;       \
 225         x4 ^= x2;                                       \
 226         })
 227
 228 /*
 229  * both gcc and clang have misoptimized this function in the past,
 230  * producing horrible object code from spilling temporary variables
 231  * on the stack. Forcing this part out of line avoids that.
 232  */
 233 static noinline void __serpent_setkey_sbox(u32 r0, u32 r1, u32 r2,
 234                                            u32 r3, u32 r4, u32 *k)
 235 {
 236         k += 100;
 237         S3(r3, r4, r0, r1, r2); store_and_load_keys(r1, r2, r4, r3, 28, 24);
 238         S4(r1, r2, r4, r3, r0); store_and_load_keys(r2, r4, r3, r0, 24, 20);
 239         S5(r2, r4, r3, r0, r1); store_and_load_keys(r1, r2, r4, r0, 20, 16);
 240         S6(r1, r2, r4, r0, r3); store_and_load_keys(r4, r3, r2, r0, 16, 12);
 241         S7(r4, r3, r2, r0, r1); store_and_load_keys(r1, r2, r0, r4, 12, 8);
 242         S0(r1, r2, r0, r4, r3); store_and_load_keys(r0, r2, r4, r1, 8, 4);
 243         S1(r0, r2, r4, r1, r3); store_and_load_keys(r3, r4, r1, r0, 4, 0);
 244         S2(r3, r4, r1, r0, r2); store_and_load_keys(r2, r4, r3, r0, 0, -4);
 245         S3(r2, r4, r3, r0, r1); store_and_load_keys(r0, r1, r4, r2, -4, -8);
 246         S4(r0, r1, r4, r2, r3); store_and_load_keys(r1, r4, r2, r3, -8, -12);
 247         S5(r1, r4, r2, r3, r0); store_and_load_keys(r0, r1, r4, r3, -12, -16);
 248         S6(r0, r1, r4, r3, r2); store_and_load_keys(r4, r2, r1, r3, -16, -20);
 249         S7(r4, r2, r1, r3, r0); store_and_load_keys(r0, r1, r3, r4, -20, -24);
 250         S0(r0, r1, r3, r4, r2); store_and_load_keys(r3, r1, r4, r0, -24, -28);
 251         k -= 50;
 252         S1(r3, r1, r4, r0, r2); store_and_load_keys(r2, r4, r0, r3, 22, 18);
 253         S2(r2, r4, r0, r3, r1); store_and_load_keys(r1, r4, r2, r3, 18, 14);
 254         S3(r1, r4, r2, r3, r0); store_and_load_keys(r3, r0, r4, r1, 14, 10);
 255         S4(r3, r0, r4, r1, r2); store_and_load_keys(r0, r4, r1, r2, 10, 6);
 256         S5(r0, r4, r1, r2, r3); store_and_load_keys(r3, r0, r4, r2, 6, 2);
 257         S6(r3, r0, r4, r2, r1); store_and_load_keys(r4, r1, r0, r2, 2, -2);
 258         S7(r4, r1, r0, r2, r3); store_and_load_keys(r3, r0, r2, r4, -2, -6);
 259         S0(r3, r0, r2, r4, r1); store_and_load_keys(r2, r0, r4, r3, -6, -10);
 260         S1(r2, r0, r4, r3, r1); store_and_load_keys(r1, r4, r3, r2, -10, -14);
 261         S2(r1, r4, r3, r2, r0); store_and_load_keys(r0, r4, r1, r2, -14, -18);
 262         S3(r0, r4, r1, r2, r3); store_and_load_keys(r2, r3, r4, r0, -18, -22);
 263         k -= 50;
 264         S4(r2, r3, r4, r0, r1); store_and_load_keys(r3, r4, r0, r1, 28, 24);
 265         S5(r3, r4, r0, r1, r2); store_and_load_keys(r2, r3, r4, r1, 24, 20);
 266         S6(r2, r3, r4, r1, r0); store_and_load_keys(r4, r0, r3, r1, 20, 16);
 267         S7(r4, r0, r3, r1, r2); store_and_load_keys(r2, r3, r1, r4, 16, 12);
 268         S0(r2, r3, r1, r4, r0); store_and_load_keys(r1, r3, r4, r2, 12, 8);
 269         S1(r1, r3, r4, r2, r0); store_and_load_keys(r0, r4, r2, r1, 8, 4);
 270         S2(r0, r4, r2, r1, r3); store_and_load_keys(r3, r4, r0, r1, 4, 0);
 271         S3(r3, r4, r0, r1, r2); storekeys(r1, r2, r4, r3, 0);
 272 }
 273
 274 int __serpent_setkey(struct serpent_ctx *ctx, const u8 *key,
 275                      unsigned int keylen)
 276 {
 277         u32 *k = ctx->expkey;
 278         u8  *k8 = (u8 *)k;
 279         u32 r0, r1, r2, r3, r4;
 280         int i;
 281
 282         /* Copy key, add padding */
 283
 284         for (i = 0; i < keylen; ++i)
 285                 k8[i] = key[i];
 286         if (i < SERPENT_MAX_KEY_SIZE)
 287                 k8[i++] = 1;
 288         while (i < SERPENT_MAX_KEY_SIZE)
 289                 k8[i++] = 0;
 290
 291         /* Expand key using polynomial */
 292
 293         r0 = le32_to_cpu(k[3]);
 294         r1 = le32_to_cpu(k[4]);
 295         r2 = le32_to_cpu(k[5]);
 296         r3 = le32_to_cpu(k[6]);
 297         r4 = le32_to_cpu(k[7]);
 298
 299         keyiter(le32_to_cpu(k[0]), r0, r4, r2, 0, 0);
 300         keyiter(le32_to_cpu(k[1]), r1, r0, r3, 1, 1);
 301         keyiter(le32_to_cpu(k[2]), r2, r1, r4, 2, 2);
 302         keyiter(le32_to_cpu(k[3]), r3, r2, r0, 3, 3);
 303         keyiter(le32_to_cpu(k[4]), r4, r3, r1, 4, 4);
 304         keyiter(le32_to_cpu(k[5]), r0, r4, r2, 5, 5);
 305         keyiter(le32_to_cpu(k[6]), r1, r0, r3, 6, 6);
 306         keyiter(le32_to_cpu(k[7]), r2, r1, r4, 7, 7);
 307
 308         keyiter(k[0], r3, r2, r0, 8, 8);
 309         keyiter(k[1], r4, r3, r1, 9, 9);
 310         keyiter(k[2], r0, r4, r2, 10, 10);
 311         keyiter(k[3], r1, r0, r3, 11, 11);
 312         keyiter(k[4], r2, r1, r4, 12, 12);
 313         keyiter(k[5], r3, r2, r0, 13, 13);
 314         keyiter(k[6], r4, r3, r1, 14, 14);
 315         keyiter(k[7], r0, r4, r2, 15, 15);
 316         keyiter(k[8], r1, r0, r3, 16, 16);
 317         keyiter(k[9], r2, r1, r4, 17, 17);
 318         keyiter(k[10], r3, r2, r0, 18, 18);
 319         keyiter(k[11], r4, r3, r1, 19, 19);
 320         keyiter(k[12], r0, r4, r2, 20, 20);
 321         keyiter(k[13], r1, r0, r3, 21, 21);
 322         keyiter(k[14], r2, r1, r4, 22, 22);
 323         keyiter(k[15], r3, r2, r0, 23, 23);
 324         keyiter(k[16], r4, r3, r1, 24, 24);
 325         keyiter(k[17], r0, r4, r2, 25, 25);
 326         keyiter(k[18], r1, r0, r3, 26, 26);
 327         keyiter(k[19], r2, r1, r4, 27, 27);
 328         keyiter(k[20], r3, r2, r0, 28, 28);
 329         keyiter(k[21], r4, r3, r1, 29, 29);
 330         keyiter(k[22], r0, r4, r2, 30, 30);
 331         keyiter(k[23], r1, r0, r3, 31, 31);
 332
 333         k += 50;
 334
 335         keyiter(k[-26], r2, r1, r4, 32, -18);
 336         keyiter(k[-25], r3, r2, r0, 33, -17);
 337         keyiter(k[-24], r4, r3, r1, 34, -16);
 338         keyiter(k[-23], r0, r4, r2, 35, -15);
 339         keyiter(k[-22], r1, r0, r3, 36, -14);
 340         keyiter(k[-21], r2, r1, r4, 37, -13);
 341         keyiter(k[-20], r3, r2, r0, 38, -12);
 342         keyiter(k[-19], r4, r3, r1, 39, -11);
 343         keyiter(k[-18], r0, r4, r2, 40, -10);
 344         keyiter(k[-17], r1, r0, r3, 41, -9);
 345         keyiter(k[-16], r2, r1, r4, 42, -8);
 346         keyiter(k[-15], r3, r2, r0, 43, -7);
 347         keyiter(k[-14], r4, r3, r1, 44, -6);
 348         keyiter(k[-13], r0, r4, r2, 45, -5);
 349         keyiter(k[-12], r1, r0, r3, 46, -4);
 350         keyiter(k[-11], r2, r1, r4, 47, -3);
 351         keyiter(k[-10], r3, r2, r0, 48, -2);
 352         keyiter(k[-9], r4, r3, r1, 49, -1);
 353         keyiter(k[-8], r0, r4, r2, 50, 0);
 354         keyiter(k[-7], r1, r0, r3, 51, 1);
 355         keyiter(k[-6], r2, r1, r4, 52, 2);
 356         keyiter(k[-5], r3, r2, r0, 53, 3);
 357         keyiter(k[-4], r4, r3, r1, 54, 4);
 358         keyiter(k[-3], r0, r4, r2, 55, 5);
 359         keyiter(k[-2], r1, r0, r3, 56, 6);
 360         keyiter(k[-1], r2, r1, r4, 57, 7);
 361         keyiter(k[0], r3, r2, r0, 58, 8);
 362         keyiter(k[1], r4, r3, r1, 59, 9);
 363         keyiter(k[2], r0, r4, r2, 60, 10);
 364         keyiter(k[3], r1, r0, r3, 61, 11);
 365         keyiter(k[4], r2, r1, r4, 62, 12);
 366         keyiter(k[5], r3, r2, r0, 63, 13);
 367         keyiter(k[6], r4, r3, r1, 64, 14);
 368         keyiter(k[7], r0, r4, r2, 65, 15);
 369         keyiter(k[8], r1, r0, r3, 66, 16);
 370         keyiter(k[9], r2, r1, r4, 67, 17);
 371         keyiter(k[10], r3, r2, r0, 68, 18);
 372         keyiter(k[11], r4, r3, r1, 69, 19);
 373         keyiter(k[12], r0, r4, r2, 70, 20);
 374         keyiter(k[13], r1, r0, r3, 71, 21);
 375         keyiter(k[14], r2, r1, r4, 72, 22);
 376         keyiter(k[15], r3, r2, r0, 73, 23);
 377         keyiter(k[16], r4, r3, r1, 74, 24);
 378         keyiter(k[17], r0, r4, r2, 75, 25);
 379         keyiter(k[18], r1, r0, r3, 76, 26);
 380         keyiter(k[19], r2, r1, r4, 77, 27);
 381         keyiter(k[20], r3, r2, r0, 78, 28);
 382         keyiter(k[21], r4, r3, r1, 79, 29);
 383         keyiter(k[22], r0, r4, r2, 80, 30);
 384         keyiter(k[23], r1, r0, r3, 81, 31);
 385
 386         k += 50;
 387
 388         keyiter(k[-26], r2, r1, r4, 82, -18);
 389         keyiter(k[-25], r3, r2, r0, 83, -17);
 390         keyiter(k[-24], r4, r3, r1, 84, -16);
 391         keyiter(k[-23], r0, r4, r2, 85, -15);
 392         keyiter(k[-22], r1, r0, r3, 86, -14);
 393         keyiter(k[-21], r2, r1, r4, 87, -13);
 394         keyiter(k[-20], r3, r2, r0, 88, -12);
 395         keyiter(k[-19], r4, r3, r1, 89, -11);
 396         keyiter(k[-18], r0, r4, r2, 90, -10);
 397         keyiter(k[-17], r1, r0, r3, 91, -9);
 398         keyiter(k[-16], r2, r1, r4, 92, -8);
 399         keyiter(k[-15], r3, r2, r0, 93, -7);
 400         keyiter(k[-14], r4, r3, r1, 94, -6);
 401         keyiter(k[-13], r0, r4, r2, 95, -5);
 402         keyiter(k[-12], r1, r0, r3, 96, -4);
 403         keyiter(k[-11], r2, r1, r4, 97, -3);
 404         keyiter(k[-10], r3, r2, r0, 98, -2);
 405         keyiter(k[-9], r4, r3, r1, 99, -1);
 406         keyiter(k[-8], r0, r4, r2, 100, 0);
 407         keyiter(k[-7], r1, r0, r3, 101, 1);
 408         keyiter(k[-6], r2, r1, r4, 102, 2);
 409         keyiter(k[-5], r3, r2, r0, 103, 3);
 410         keyiter(k[-4], r4, r3, r1, 104, 4);
 411         keyiter(k[-3], r0, r4, r2, 105, 5);
 412         keyiter(k[-2], r1, r0, r3, 106, 6);
 413         keyiter(k[-1], r2, r1, r4, 107, 7);
 414         keyiter(k[0], r3, r2, r0, 108, 8);
 415         keyiter(k[1], r4, r3, r1, 109, 9);
 416         keyiter(k[2], r0, r4, r2, 110, 10);
 417         keyiter(k[3], r1, r0, r3, 111, 11);
 418         keyiter(k[4], r2, r1, r4, 112, 12);
 419         keyiter(k[5], r3, r2, r0, 113, 13);
 420         keyiter(k[6], r4, r3, r1, 114, 14);
 421         keyiter(k[7], r0, r4, r2, 115, 15);
 422         keyiter(k[8], r1, r0, r3, 116, 16);
 423         keyiter(k[9], r2, r1, r4, 117, 17);
 424         keyiter(k[10], r3, r2, r0, 118, 18);
 425         keyiter(k[11], r4, r3, r1, 119, 19);
 426         keyiter(k[12], r0, r4, r2, 120, 20);
 427         keyiter(k[13], r1, r0, r3, 121, 21);
 428         keyiter(k[14], r2, r1, r4, 122, 22);
 429         keyiter(k[15], r3, r2, r0, 123, 23);
 430         keyiter(k[16], r4, r3, r1, 124, 24);
 431         keyiter(k[17], r0, r4, r2, 125, 25);
 432         keyiter(k[18], r1, r0, r3, 126, 26);
 433         keyiter(k[19], r2, r1, r4, 127, 27);
 434         keyiter(k[20], r3, r2, r0, 128, 28);
 435         keyiter(k[21], r4, r3, r1, 129, 29);
 436         keyiter(k[22], r0, r4, r2, 130, 30);
 437         keyiter(k[23], r1, r0, r3, 131, 31);
 438
 439         /* Apply S-boxes */
 440         __serpent_setkey_sbox(r0, r1, r2, r3, r4, ctx->expkey);
 441
 442         return 0;
 443 }
 444 EXPORT_SYMBOL_GPL(__serpent_setkey);
 445
 446 int serpent_setkey(struct crypto_tfm *tfm, const u8 *key, unsigned int keylen)
 447 {
 448         return __serpent_setkey(crypto_tfm_ctx(tfm), key, keylen);
 449 }
 450 EXPORT_SYMBOL_GPL(serpent_setkey);
 451
 452 void __serpent_encrypt(struct serpent_ctx *ctx, u8 *dst, const u8 *src)
 453 {
 454         const u32 *k = ctx->expkey;
 455         const __le32 *s = (const __le32 *)src;
 456         __le32  *d = (__le32 *)dst;
 457         u32     r0, r1, r2, r3, r4;
 458
 459 /*
 460  * Note: The conversions between u8* and u32* might cause trouble
 461  * on architectures with stricter alignment rules than x86
 462  */
 463
 464         r0 = le32_to_cpu(s[0]);
 465         r1 = le32_to_cpu(s[1]);
 466         r2 = le32_to_cpu(s[2]);
 467         r3 = le32_to_cpu(s[3]);
 468
 469                                         K(r0, r1, r2, r3, 0);
 470         S0(r0, r1, r2, r3, r4);         LK(r2, r1, r3, r0, r4, 1);
 471         S1(r2, r1, r3, r0, r4);         LK(r4, r3, r0, r2, r1, 2);
 472         S2(r4, r3, r0, r2, r1);         LK(r1, r3, r4, r2, r0, 3);
 473         S3(r1, r3, r4, r2, r0);         LK(r2, r0, r3, r1, r4, 4);
 474         S4(r2, r0, r3, r1, r4);         LK(r0, r3, r1, r4, r2, 5);
 475         S5(r0, r3, r1, r4, r2);         LK(r2, r0, r3, r4, r1, 6);
 476         S6(r2, r0, r3, r4, r1);         LK(r3, r1, r0, r4, r2, 7);
 477         S7(r3, r1, r0, r4, r2);         LK(r2, r0, r4, r3, r1, 8);
 478         S0(r2, r0, r4, r3, r1);         LK(r4, r0, r3, r2, r1, 9);
 479         S1(r4, r0, r3, r2, r1);         LK(r1, r3, r2, r4, r0, 10);
 480         S2(r1, r3, r2, r4, r0);         LK(r0, r3, r1, r4, r2, 11);
 481         S3(r0, r3, r1, r4, r2);         LK(r4, r2, r3, r0, r1, 12);
 482         S4(r4, r2, r3, r0, r1);         LK(r2, r3, r0, r1, r4, 13);
 483         S5(r2, r3, r0, r1, r4);         LK(r4, r2, r3, r1, r0, 14);
 484         S6(r4, r2, r3, r1, r0);         LK(r3, r0, r2, r1, r4, 15);
 485         S7(r3, r0, r2, r1, r4);         LK(r4, r2, r1, r3, r0, 16);
 486         S0(r4, r2, r1, r3, r0);         LK(r1, r2, r3, r4, r0, 17);
 487         S1(r1, r2, r3, r4, r0);         LK(r0, r3, r4, r1, r2, 18);
 488         S2(r0, r3, r4, r1, r2);         LK(r2, r3, r0, r1, r4, 19);
 489         S3(r2, r3, r0, r1, r4);         LK(r1, r4, r3, r2, r0, 20);
 490         S4(r1, r4, r3, r2, r0);         LK(r4, r3, r2, r0, r1, 21);
 491         S5(r4, r3, r2, r0, r1);         LK(r1, r4, r3, r0, r2, 22);
 492         S6(r1, r4, r3, r0, r2);         LK(r3, r2, r4, r0, r1, 23);
 493         S7(r3, r2, r4, r0, r1);         LK(r1, r4, r0, r3, r2, 24);
 494         S0(r1, r4, r0, r3, r2);         LK(r0, r4, r3, r1, r2, 25);
 495         S1(r0, r4, r3, r1, r2);         LK(r2, r3, r1, r0, r4, 26);
 496         S2(r2, r3, r1, r0, r4);         LK(r4, r3, r2, r0, r1, 27);
 497         S3(r4, r3, r2, r0, r1);         LK(r0, r1, r3, r4, r2, 28);
 498         S4(r0, r1, r3, r4, r2);         LK(r1, r3, r4, r2, r0, 29);
 499         S5(r1, r3, r4, r2, r0);         LK(r0, r1, r3, r2, r4, 30);
 500         S6(r0, r1, r3, r2, r4);         LK(r3, r4, r1, r2, r0, 31);
 501         S7(r3, r4, r1, r2, r0);         K(r0, r1, r2, r3, 32);
 502
 503         d[0] = cpu_to_le32(r0);
 504         d[1] = cpu_to_le32(r1);
 505         d[2] = cpu_to_le32(r2);
 506         d[3] = cpu_to_le32(r3);
 507 }
 508 EXPORT_SYMBOL_GPL(__serpent_encrypt);
 509
 510 static void serpent_encrypt(struct crypto_tfm *tfm, u8 *dst, const u8 *src)
 511 {
 512         struct serpent_ctx *ctx = crypto_tfm_ctx(tfm);
 513
 514         __serpent_encrypt(ctx, dst, src);
 515 }
 516
 517 void __serpent_decrypt(struct serpent_ctx *ctx, u8 *dst, const u8 *src)
 518 {
 519         const u32 *k = ctx->expkey;
 520         const __le32 *s = (const __le32 *)src;
 521         __le32  *d = (__le32 *)dst;
 522         u32     r0, r1, r2, r3, r4;
 523
 524         r0 = le32_to_cpu(s[0]);
 525         r1 = le32_to_cpu(s[1]);
 526         r2 = le32_to_cpu(s[2]);
 527         r3 = le32_to_cpu(s[3]);
 528
 529                                         K(r0, r1, r2, r3, 32);
 530         SI7(r0, r1, r2, r3, r4);        KL(r1, r3, r0, r4, r2, 31);
 531         SI6(r1, r3, r0, r4, r2);        KL(r0, r2, r4, r1, r3, 30);
 532         SI5(r0, r2, r4, r1, r3);        KL(r2, r3, r0, r4, r1, 29);
 533         SI4(r2, r3, r0, r4, r1);        KL(r2, r0, r1, r4, r3, 28);
 534         SI3(r2, r0, r1, r4, r3);        KL(r1, r2, r3, r4, r0, 27);
 535         SI2(r1, r2, r3, r4, r0);        KL(r2, r0, r4, r3, r1, 26);
 536         SI1(r2, r0, r4, r3, r1);        KL(r1, r0, r4, r3, r2, 25);
 537         SI0(r1, r0, r4, r3, r2);        KL(r4, r2, r0, r1, r3, 24);
 538         SI7(r4, r2, r0, r1, r3);        KL(r2, r1, r4, r3, r0, 23);
 539         SI6(r2, r1, r4, r3, r0);        KL(r4, r0, r3, r2, r1, 22);
 540         SI5(r4, r0, r3, r2, r1);        KL(r0, r1, r4, r3, r2, 21);
 541         SI4(r0, r1, r4, r3, r2);        KL(r0, r4, r2, r3, r1, 20);
 542         SI3(r0, r4, r2, r3, r1);        KL(r2, r0, r1, r3, r4, 19);
 543         SI2(r2, r0, r1, r3, r4);        KL(r0, r4, r3, r1, r2, 18);
 544         SI1(r0, r4, r3, r1, r2);        KL(r2, r4, r3, r1, r0, 17);
 545         SI0(r2, r4, r3, r1, r0);        KL(r3, r0, r4, r2, r1, 16);
 546         SI7(r3, r0, r4, r2, r1);        KL(r0, r2, r3, r1, r4, 15);
 547         SI6(r0, r2, r3, r1, r4);        KL(r3, r4, r1, r0, r2, 14);
 548         SI5(r3, r4, r1, r0, r2);        KL(r4, r2, r3, r1, r0, 13);
 549         SI4(r4, r2, r3, r1, r0);        KL(r4, r3, r0, r1, r2, 12);
 550         SI3(r4, r3, r0, r1, r2);        KL(r0, r4, r2, r1, r3, 11);
 551         SI2(r0, r4, r2, r1, r3);        KL(r4, r3, r1, r2, r0, 10);
 552         SI1(r4, r3, r1, r2, r0);        KL(r0, r3, r1, r2, r4, 9);
 553         SI0(r0, r3, r1, r2, r4);        KL(r1, r4, r3, r0, r2, 8);
 554         SI7(r1, r4, r3, r0, r2);        KL(r4, r0, r1, r2, r3, 7);
 555         SI6(r4, r0, r1, r2, r3);        KL(r1, r3, r2, r4, r0, 6);
 556         SI5(r1, r3, r2, r4, r0);        KL(r3, r0, r1, r2, r4, 5);
 557         SI4(r3, r0, r1, r2, r4);        KL(r3, r1, r4, r2, r0, 4);
 558         SI3(r3, r1, r4, r2, r0);        KL(r4, r3, r0, r2, r1, 3);
 559         SI2(r4, r3, r0, r2, r1);        KL(r3, r1, r2, r0, r4, 2);
 560         SI1(r3, r1, r2, r0, r4);        KL(r4, r1, r2, r0, r3, 1);
 561         SI0(r4, r1, r2, r0, r3);        K(r2, r3, r1, r4, 0);
 562
 563         d[0] = cpu_to_le32(r2);
 564         d[1] = cpu_to_le32(r3);
 565         d[2] = cpu_to_le32(r1);
 566         d[3] = cpu_to_le32(r4);
 567 }
 568 EXPORT_SYMBOL_GPL(__serpent_decrypt);
 569
 570 static void serpent_decrypt(struct crypto_tfm *tfm, u8 *dst, const u8 *src)
 571 {
 572         struct serpent_ctx *ctx = crypto_tfm_ctx(tfm);
 573
 574         __serpent_decrypt(ctx, dst, src);
 575 }
 576
 577 static int tnepres_setkey(struct crypto_tfm *tfm, const u8 *key,
 578                           unsigned int keylen)
 579 {
 580         u8 rev_key[SERPENT_MAX_KEY_SIZE];
 581         int i;
 582
 583         for (i = 0; i < keylen; ++i)
 584                 rev_key[keylen - i - 1] = key[i];
 585
 586         return serpent_setkey(tfm, rev_key, keylen);
 587 }
 588
 589 static void tnepres_encrypt(struct crypto_tfm *tfm, u8 *dst, const u8 *src)
 590 {
 591         const u32 * const s = (const u32 * const)src;
 592         u32 * const d = (u32 * const)dst;
 593
 594         u32 rs[4], rd[4];
 595
 596         rs[0] = swab32(s[3]);
 597         rs[1] = swab32(s[2]);
 598         rs[2] = swab32(s[1]);
 599         rs[3] = swab32(s[0]);
 600
 601         serpent_encrypt(tfm, (u8 *)rd, (u8 *)rs);
 602
 603         d[0] = swab32(rd[3]);
 604         d[1] = swab32(rd[2]);
 605         d[2] = swab32(rd[1]);
 606         d[3] = swab32(rd[0]);
 607 }
 608
 609 static void tnepres_decrypt(struct crypto_tfm *tfm, u8 *dst, const u8 *src)
 610 {
 611         const u32 * const s = (const u32 * const)src;
 612         u32 * const d = (u32 * const)dst;
 613
 614         u32 rs[4], rd[4];
 615
 616         rs[0] = swab32(s[3]);
 617         rs[1] = swab32(s[2]);
 618         rs[2] = swab32(s[1]);
 619         rs[3] = swab32(s[0]);
 620
 621         serpent_decrypt(tfm, (u8 *)rd, (u8 *)rs);
 622
 623         d[0] = swab32(rd[3]);
 624         d[1] = swab32(rd[2]);
 625         d[2] = swab32(rd[1]);
 626         d[3] = swab32(rd[0]);
 627 }
 628
 629 static struct crypto_alg srp_algs[2] = { {
 630         .cra_name               =       "serpent",
 631         .cra_driver_name        =       "serpent-generic",
 632         .cra_priority           =       100,
 633         .cra_flags              =       CRYPTO_ALG_TYPE_CIPHER,
 634         .cra_blocksize          =       SERPENT_BLOCK_SIZE,
 635         .cra_ctxsize            =       sizeof(struct serpent_ctx),
 636         .cra_alignmask          =       3,
 637         .cra_module             =       THIS_MODULE,
 638         .cra_u                  =       { .cipher = {
 639         .cia_min_keysize        =       SERPENT_MIN_KEY_SIZE,
 640         .cia_max_keysize        =       SERPENT_MAX_KEY_SIZE,
 641         .cia_setkey             =       serpent_setkey,
 642         .cia_encrypt            =       serpent_encrypt,
 643         .cia_decrypt            =       serpent_decrypt } }
 644 }, {
 645         .cra_name               =       "tnepres",
 646         .cra_driver_name        =       "tnepres-generic",
 647         .cra_flags              =       CRYPTO_ALG_TYPE_CIPHER,
 648         .cra_blocksize          =       SERPENT_BLOCK_SIZE,
 649         .cra_ctxsize            =       sizeof(struct serpent_ctx),
 650         .cra_alignmask          =       3,
 651         .cra_module             =       THIS_MODULE,
 652         .cra_u                  =       { .cipher = {
 653         .cia_min_keysize        =       SERPENT_MIN_KEY_SIZE,
 654         .cia_max_keysize        =       SERPENT_MAX_KEY_SIZE,
 655         .cia_setkey             =       tnepres_setkey,
 656         .cia_encrypt            =       tnepres_encrypt,
 657         .cia_decrypt            =       tnepres_decrypt } }
 658 } };
 659
 660 static int __init serpent_mod_init(void)
 661 {
 662         return crypto_register_algs(srp_algs, ARRAY_SIZE(srp_algs));
 663 }
 664
 665 static void __exit serpent_mod_fini(void)
 666 {
 667         crypto_unregister_algs(srp_algs, ARRAY_SIZE(srp_algs));
 668 }
 669
 670 subsys_initcall(serpent_mod_init);
 671 module_exit(serpent_mod_fini);
 672
 673 MODULE_LICENSE("GPL");
 674 MODULE_DESCRIPTION("Serpent and tnepres (kerneli compatible serpent reversed) Cipher Algorithm");
 675 MODULE_AUTHOR("Dag Arne Osvik <osvik@ii.uib.no>");
 676 MODULE_ALIAS_CRYPTO("tnepres");
 677 MODULE_ALIAS_CRYPTO("serpent");
 678 MODULE_ALIAS_CRYPTO("serpent-generic");