drivers/crypto/padlock-aes.c

   1 /*
   2  * Cryptographic API.
   3  *
   4  * Support for VIA PadLock hardware crypto engine.
   5  *
   6  * Copyright (c) 2004  Michal Ludvig <michal@logix.cz>
   7  *
   8  * Key expansion routine taken from crypto/aes.c
   9  *
  10  * This program is free software; you can redistribute it and/or modify
  11  * it under the terms of the GNU General Public License as published by
  12  * the Free Software Foundation; either version 2 of the License, or
  13  * (at your option) any later version.
  14  *
  15  * ---------------------------------------------------------------------------
  16  * Copyright (c) 2002, Dr Brian Gladman <brg@gladman.me.uk>, Worcester, UK.
  17  * All rights reserved.
  18  *
  19  * LICENSE TERMS
  20  *
  21  * The free distribution and use of this software in both source and binary
  22  * form is allowed (with or without changes) provided that:
  23  *
  24  *   1. distributions of this source code include the above copyright
  25  *      notice, this list of conditions and the following disclaimer;
  26  *
  27  *   2. distributions in binary form include the above copyright
  28  *      notice, this list of conditions and the following disclaimer
  29  *      in the documentation and/or other associated materials;
  30  *
  31  *   3. the copyright holder's name is not used to endorse products
  32  *      built using this software without specific written permission.
  33  *
  34  * ALTERNATIVELY, provided that this notice is retained in full, this product
  35  * may be distributed under the terms of the GNU General Public License (GPL),
  36  * in which case the provisions of the GPL apply INSTEAD OF those given above.
  37  *
  38  * DISCLAIMER
  39  *
  40  * This software is provided 'as is' with no explicit or implied warranties
  41  * in respect of its properties, including, but not limited to, correctness
  42  * and/or fitness for purpose.
  43  * ---------------------------------------------------------------------------
  44  */
  45
  46 #include <linux/module.h>
  47 #include <linux/init.h>
  48 #include <linux/types.h>
  49 #include <linux/errno.h>
  50 #include <linux/crypto.h>
  51 #include <linux/interrupt.h>
  52 #include <linux/kernel.h>
  53 #include <asm/byteorder.h>
  54 #include "padlock.h"
  55
  56 #define AES_MIN_KEY_SIZE        16      /* in uint8_t units */
  57 #define AES_MAX_KEY_SIZE        32      /* ditto */
  58 #define AES_BLOCK_SIZE          16      /* ditto */
  59 #define AES_EXTENDED_KEY_SIZE   64      /* in uint32_t units */
  60 #define AES_EXTENDED_KEY_SIZE_B (AES_EXTENDED_KEY_SIZE * sizeof(uint32_t))
  61
  62 struct aes_ctx {
  63         uint32_t e_data[AES_EXTENDED_KEY_SIZE];
  64         uint32_t d_data[AES_EXTENDED_KEY_SIZE];
  65         struct {
  66                 struct cword encrypt;
  67                 struct cword decrypt;
  68         } cword;
  69         uint32_t *E;
  70         uint32_t *D;
  71         int key_length;
  72 };
  73
  74 /* ====== Key management routines ====== */
  75
  76 static inline uint32_t
  77 generic_rotr32 (const uint32_t x, const unsigned bits)
  78 {
  79         const unsigned n = bits % 32;
  80         return (x >> n) | (x << (32 - n));
  81 }
  82
  83 static inline uint32_t
  84 generic_rotl32 (const uint32_t x, const unsigned bits)
  85 {
  86         const unsigned n = bits % 32;
  87         return (x << n) | (x >> (32 - n));
  88 }
  89
  90 #define rotl generic_rotl32
  91 #define rotr generic_rotr32
  92
  93 /*
  94  * #define byte(x, nr) ((unsigned char)((x) >> (nr*8)))
  95  */
  96 static inline uint8_t
  97 byte(const uint32_t x, const unsigned n)
  98 {
  99         return x >> (n << 3);
 100 }
 101
 102 #define uint32_t_in(x) le32_to_cpu(*(const uint32_t *)(x))
 103 #define uint32_t_out(to, from) (*(uint32_t *)(to) = cpu_to_le32(from))
 104
 105 #define E_KEY ctx->E
 106 #define D_KEY ctx->D
 107
 108 static uint8_t pow_tab[256];
 109 static uint8_t log_tab[256];
 110 static uint8_t sbx_tab[256];
 111 static uint8_t isb_tab[256];
 112 static uint32_t rco_tab[10];
 113 static uint32_t ft_tab[4][256];
 114 static uint32_t it_tab[4][256];
 115
 116 static uint32_t fl_tab[4][256];
 117 static uint32_t il_tab[4][256];
 118
 119 static inline uint8_t
 120 f_mult (uint8_t a, uint8_t b)
 121 {
 122         uint8_t aa = log_tab[a], cc = aa + log_tab[b];
 123
 124         return pow_tab[cc + (cc < aa ? 1 : 0)];
 125 }
 126
 127 #define ff_mult(a,b)    (a && b ? f_mult(a, b) : 0)
 128
 129 #define f_rn(bo, bi, n, k)                                      \
 130     bo[n] =  ft_tab[0][byte(bi[n],0)] ^                         \
 131              ft_tab[1][byte(bi[(n + 1) & 3],1)] ^               \
 132              ft_tab[2][byte(bi[(n + 2) & 3],2)] ^               \
 133              ft_tab[3][byte(bi[(n + 3) & 3],3)] ^ *(k + n)
 134
 135 #define i_rn(bo, bi, n, k)                                      \
 136     bo[n] =  it_tab[0][byte(bi[n],0)] ^                         \
 137              it_tab[1][byte(bi[(n + 3) & 3],1)] ^               \
 138              it_tab[2][byte(bi[(n + 2) & 3],2)] ^               \
 139              it_tab[3][byte(bi[(n + 1) & 3],3)] ^ *(k + n)
 140
 141 #define ls_box(x)                               \
 142     ( fl_tab[0][byte(x, 0)] ^                   \
 143       fl_tab[1][byte(x, 1)] ^                   \
 144       fl_tab[2][byte(x, 2)] ^                   \
 145       fl_tab[3][byte(x, 3)] )
 146
 147 #define f_rl(bo, bi, n, k)                                      \
 148     bo[n] =  fl_tab[0][byte(bi[n],0)] ^                         \
 149              fl_tab[1][byte(bi[(n + 1) & 3],1)] ^               \
 150              fl_tab[2][byte(bi[(n + 2) & 3],2)] ^               \
 151              fl_tab[3][byte(bi[(n + 3) & 3],3)] ^ *(k + n)
 152
 153 #define i_rl(bo, bi, n, k)                                      \
 154     bo[n] =  il_tab[0][byte(bi[n],0)] ^                         \
 155              il_tab[1][byte(bi[(n + 3) & 3],1)] ^               \
 156              il_tab[2][byte(bi[(n + 2) & 3],2)] ^               \
 157              il_tab[3][byte(bi[(n + 1) & 3],3)] ^ *(k + n)
 158
 159 static void
 160 gen_tabs (void)
 161 {
 162         uint32_t i, t;
 163         uint8_t p, q;
 164
 165         /* log and power tables for GF(2**8) finite field with
 166            0x011b as modular polynomial - the simplest prmitive
 167            root is 0x03, used here to generate the tables */
 168
 169         for (i = 0, p = 1; i < 256; ++i) {
 170                 pow_tab[i] = (uint8_t) p;
 171                 log_tab[p] = (uint8_t) i;
 172
 173                 p ^= (p << 1) ^ (p & 0x80 ? 0x01b : 0);
 174         }
 175
 176         log_tab[1] = 0;
 177
 178         for (i = 0, p = 1; i < 10; ++i) {
 179                 rco_tab[i] = p;
 180
 181                 p = (p << 1) ^ (p & 0x80 ? 0x01b : 0);
 182         }
 183
 184         for (i = 0; i < 256; ++i) {
 185                 p = (i ? pow_tab[255 - log_tab[i]] : 0);
 186                 q = ((p >> 7) | (p << 1)) ^ ((p >> 6) | (p << 2));
 187                 p ^= 0x63 ^ q ^ ((q >> 6) | (q << 2));
 188                 sbx_tab[i] = p;
 189                 isb_tab[p] = (uint8_t) i;
 190         }
 191
 192         for (i = 0; i < 256; ++i) {
 193                 p = sbx_tab[i];
 194
 195                 t = p;
 196                 fl_tab[0][i] = t;
 197                 fl_tab[1][i] = rotl (t, 8);
 198                 fl_tab[2][i] = rotl (t, 16);
 199                 fl_tab[3][i] = rotl (t, 24);
 200
 201                 t = ((uint32_t) ff_mult (2, p)) |
 202                     ((uint32_t) p << 8) |
 203                     ((uint32_t) p << 16) | ((uint32_t) ff_mult (3, p) << 24);
 204
 205                 ft_tab[0][i] = t;
 206                 ft_tab[1][i] = rotl (t, 8);
 207                 ft_tab[2][i] = rotl (t, 16);
 208                 ft_tab[3][i] = rotl (t, 24);
 209
 210                 p = isb_tab[i];
 211
 212                 t = p;
 213                 il_tab[0][i] = t;
 214                 il_tab[1][i] = rotl (t, 8);
 215                 il_tab[2][i] = rotl (t, 16);
 216                 il_tab[3][i] = rotl (t, 24);
 217
 218                 t = ((uint32_t) ff_mult (14, p)) |
 219                     ((uint32_t) ff_mult (9, p) << 8) |
 220                     ((uint32_t) ff_mult (13, p) << 16) |
 221                     ((uint32_t) ff_mult (11, p) << 24);
 222
 223                 it_tab[0][i] = t;
 224                 it_tab[1][i] = rotl (t, 8);
 225                 it_tab[2][i] = rotl (t, 16);
 226                 it_tab[3][i] = rotl (t, 24);
 227         }
 228 }
 229
 230 #define star_x(x) (((x) & 0x7f7f7f7f) << 1) ^ ((((x) & 0x80808080) >> 7) * 0x1b)
 231
 232 #define imix_col(y,x)       \
 233     u   = star_x(x);        \
 234     v   = star_x(u);        \
 235     w   = star_x(v);        \
 236     t   = w ^ (x);          \
 237    (y)  = u ^ v ^ w;        \
 238    (y) ^= rotr(u ^ t,  8) ^ \
 239           rotr(v ^ t, 16) ^ \
 240           rotr(t,24)
 241
 242 /* initialise the key schedule from the user supplied key */
 243
 244 #define loop4(i)                                    \
 245 {   t = rotr(t,  8); t = ls_box(t) ^ rco_tab[i];    \
 246     t ^= E_KEY[4 * i];     E_KEY[4 * i + 4] = t;    \
 247     t ^= E_KEY[4 * i + 1]; E_KEY[4 * i + 5] = t;    \
 248     t ^= E_KEY[4 * i + 2]; E_KEY[4 * i + 6] = t;    \
 249     t ^= E_KEY[4 * i + 3]; E_KEY[4 * i + 7] = t;    \
 250 }
 251
 252 #define loop6(i)                                    \
 253 {   t = rotr(t,  8); t = ls_box(t) ^ rco_tab[i];    \
 254     t ^= E_KEY[6 * i];     E_KEY[6 * i + 6] = t;    \
 255     t ^= E_KEY[6 * i + 1]; E_KEY[6 * i + 7] = t;    \
 256     t ^= E_KEY[6 * i + 2]; E_KEY[6 * i + 8] = t;    \
 257     t ^= E_KEY[6 * i + 3]; E_KEY[6 * i + 9] = t;    \
 258     t ^= E_KEY[6 * i + 4]; E_KEY[6 * i + 10] = t;   \
 259     t ^= E_KEY[6 * i + 5]; E_KEY[6 * i + 11] = t;   \
 260 }
 261
 262 #define loop8(i)                                    \
 263 {   t = rotr(t,  8); ; t = ls_box(t) ^ rco_tab[i];  \
 264     t ^= E_KEY[8 * i];     E_KEY[8 * i + 8] = t;    \
 265     t ^= E_KEY[8 * i + 1]; E_KEY[8 * i + 9] = t;    \
 266     t ^= E_KEY[8 * i + 2]; E_KEY[8 * i + 10] = t;   \
 267     t ^= E_KEY[8 * i + 3]; E_KEY[8 * i + 11] = t;   \
 268     t  = E_KEY[8 * i + 4] ^ ls_box(t);    \
 269     E_KEY[8 * i + 12] = t;                \
 270     t ^= E_KEY[8 * i + 5]; E_KEY[8 * i + 13] = t;   \
 271     t ^= E_KEY[8 * i + 6]; E_KEY[8 * i + 14] = t;   \
 272     t ^= E_KEY[8 * i + 7]; E_KEY[8 * i + 15] = t;   \
 273 }
 274
 275 /* Tells whether the ACE is capable to generate
 276    the extended key for a given key_len. */
 277 static inline int
 278 aes_hw_extkey_available(uint8_t key_len)
 279 {
 280         /* TODO: We should check the actual CPU model/stepping
 281                  as it's possible that the capability will be
 282                  added in the next CPU revisions. */
 283         if (key_len == 16)
 284                 return 1;
 285         return 0;
 286 }
 287
 288 static inline struct aes_ctx *aes_ctx(void *ctx)
 289 {
 290         return (struct aes_ctx *)ALIGN((unsigned long)ctx, PADLOCK_ALIGNMENT);
 291 }
 292
 293 static int
 294 aes_set_key(void *ctx_arg, const uint8_t *in_key, unsigned int key_len, uint32_t *flags)
 295 {
 296         struct aes_ctx *ctx = aes_ctx(ctx_arg);
 297         uint32_t i, t, u, v, w;
 298         uint32_t P[AES_EXTENDED_KEY_SIZE];
 299         uint32_t rounds;
 300
 301         if (key_len != 16 && key_len != 24 && key_len != 32) {
 302                 *flags |= CRYPTO_TFM_RES_BAD_KEY_LEN;
 303                 return -EINVAL;
 304         }
 305
 306         ctx->key_length = key_len;
 307
 308         /*
 309          * If the hardware is capable of generating the extended key
 310          * itself we must supply the plain key for both encryption
 311          * and decryption.
 312          */
 313         ctx->E = ctx->e_data;
 314         ctx->D = ctx->e_data;
 315
 316         E_KEY[0] = uint32_t_in (in_key);
 317         E_KEY[1] = uint32_t_in (in_key + 4);
 318         E_KEY[2] = uint32_t_in (in_key + 8);
 319         E_KEY[3] = uint32_t_in (in_key + 12);
 320
 321         /* Prepare control words. */
 322         memset(&ctx->cword, 0, sizeof(ctx->cword));
 323
 324         ctx->cword.decrypt.encdec = 1;
 325         ctx->cword.encrypt.rounds = 10 + (key_len - 16) / 4;
 326         ctx->cword.decrypt.rounds = ctx->cword.encrypt.rounds;
 327         ctx->cword.encrypt.ksize = (key_len - 16) / 8;
 328         ctx->cword.decrypt.ksize = ctx->cword.encrypt.ksize;
 329
 330         /* Don't generate extended keys if the hardware can do it. */
 331         if (aes_hw_extkey_available(key_len))
 332                 return 0;
 333
 334         ctx->D = ctx->d_data;
 335         ctx->cword.encrypt.keygen = 1;
 336         ctx->cword.decrypt.keygen = 1;
 337
 338         switch (key_len) {
 339         case 16:
 340                 t = E_KEY[3];
 341                 for (i = 0; i < 10; ++i)
 342                         loop4 (i);
 343                 break;
 344
 345         case 24:
 346                 E_KEY[4] = uint32_t_in (in_key + 16);
 347                 t = E_KEY[5] = uint32_t_in (in_key + 20);
 348                 for (i = 0; i < 8; ++i)
 349                         loop6 (i);
 350                 break;
 351
 352         case 32:
 353                 E_KEY[4] = uint32_t_in (in_key + 16);
 354                 E_KEY[5] = uint32_t_in (in_key + 20);
 355                 E_KEY[6] = uint32_t_in (in_key + 24);
 356                 t = E_KEY[7] = uint32_t_in (in_key + 28);
 357                 for (i = 0; i < 7; ++i)
 358                         loop8 (i);
 359                 break;
 360         }
 361
 362         D_KEY[0] = E_KEY[0];
 363         D_KEY[1] = E_KEY[1];
 364         D_KEY[2] = E_KEY[2];
 365         D_KEY[3] = E_KEY[3];
 366
 367         for (i = 4; i < key_len + 24; ++i) {
 368                 imix_col (D_KEY[i], E_KEY[i]);
 369         }
 370
 371         /* PadLock needs a different format of the decryption key. */
 372         rounds = 10 + (key_len - 16) / 4;
 373
 374         for (i = 0; i < rounds; i++) {
 375                 P[((i + 1) * 4) + 0] = D_KEY[((rounds - i - 1) * 4) + 0];
 376                 P[((i + 1) * 4) + 1] = D_KEY[((rounds - i - 1) * 4) + 1];
 377                 P[((i + 1) * 4) + 2] = D_KEY[((rounds - i - 1) * 4) + 2];
 378                 P[((i + 1) * 4) + 3] = D_KEY[((rounds - i - 1) * 4) + 3];
 379         }
 380
 381         P[0] = E_KEY[(rounds * 4) + 0];
 382         P[1] = E_KEY[(rounds * 4) + 1];
 383         P[2] = E_KEY[(rounds * 4) + 2];
 384         P[3] = E_KEY[(rounds * 4) + 3];
 385
 386         memcpy(D_KEY, P, AES_EXTENDED_KEY_SIZE_B);
 387
 388         return 0;
 389 }
 390
 391 /* ====== Encryption/decryption routines ====== */
 392
 393 /* These are the real call to PadLock. */
 394 static inline void padlock_xcrypt_ecb(const u8 *input, u8 *output, void *key,
 395                                       void *control_word, u32 count)
 396 {
 397         asm volatile ("pushfl; popfl");         /* enforce key reload. */
 398         asm volatile (".byte 0xf3,0x0f,0xa7,0xc8"       /* rep xcryptecb */
 399                       : "+S"(input), "+D"(output)
 400                       : "d"(control_word), "b"(key), "c"(count));
 401 }
 402
 403 static inline u8 *padlock_xcrypt_cbc(const u8 *input, u8 *output, void *key,
 404                                      u8 *iv, void *control_word, u32 count)
 405 {
 406         /* Enforce key reload. */
 407         asm volatile ("pushfl; popfl");
 408         /* rep xcryptcbc */
 409         asm volatile (".byte 0xf3,0x0f,0xa7,0xd0"
 410                       : "+S" (input), "+D" (output), "+a" (iv)
 411                       : "d" (control_word), "b" (key), "c" (count));
 412         return iv;
 413 }
 414
 415 static void
 416 aes_encrypt(void *ctx_arg, uint8_t *out, const uint8_t *in)
 417 {
 418         struct aes_ctx *ctx = aes_ctx(ctx_arg);
 419         padlock_xcrypt_ecb(in, out, ctx->E, &ctx->cword.encrypt, 1);
 420 }
 421
 422 static void
 423 aes_decrypt(void *ctx_arg, uint8_t *out, const uint8_t *in)
 424 {
 425         struct aes_ctx *ctx = aes_ctx(ctx_arg);
 426         padlock_xcrypt_ecb(in, out, ctx->D, &ctx->cword.decrypt, 1);
 427 }
 428
 429 static unsigned int aes_encrypt_ecb(const struct cipher_desc *desc, u8 *out,
 430                                     const u8 *in, unsigned int nbytes)
 431 {
 432         struct aes_ctx *ctx = aes_ctx(crypto_tfm_ctx(desc->tfm));
 433         padlock_xcrypt_ecb(in, out, ctx->E, &ctx->cword.encrypt,
 434                            nbytes / AES_BLOCK_SIZE);
 435         return nbytes & ~(AES_BLOCK_SIZE - 1);
 436 }
 437
 438 static unsigned int aes_decrypt_ecb(const struct cipher_desc *desc, u8 *out,
 439                                     const u8 *in, unsigned int nbytes)
 440 {
 441         struct aes_ctx *ctx = aes_ctx(crypto_tfm_ctx(desc->tfm));
 442         padlock_xcrypt_ecb(in, out, ctx->D, &ctx->cword.decrypt,
 443                            nbytes / AES_BLOCK_SIZE);
 444         return nbytes & ~(AES_BLOCK_SIZE - 1);
 445 }
 446
 447 static unsigned int aes_encrypt_cbc(const struct cipher_desc *desc, u8 *out,
 448                                     const u8 *in, unsigned int nbytes)
 449 {
 450         struct aes_ctx *ctx = aes_ctx(crypto_tfm_ctx(desc->tfm));
 451         u8 *iv;
 452
 453         iv = padlock_xcrypt_cbc(in, out, ctx->E, desc->info,
 454                                 &ctx->cword.encrypt, nbytes / AES_BLOCK_SIZE);
 455         memcpy(desc->info, iv, AES_BLOCK_SIZE);
 456
 457         return nbytes & ~(AES_BLOCK_SIZE - 1);
 458 }
 459
 460 static unsigned int aes_decrypt_cbc(const struct cipher_desc *desc, u8 *out,
 461                                     const u8 *in, unsigned int nbytes)
 462 {
 463         struct aes_ctx *ctx = aes_ctx(crypto_tfm_ctx(desc->tfm));
 464         padlock_xcrypt_cbc(in, out, ctx->D, desc->info, &ctx->cword.decrypt,
 465                            nbytes / AES_BLOCK_SIZE);
 466         return nbytes & ~(AES_BLOCK_SIZE - 1);
 467 }
 468
 469 static struct crypto_alg aes_alg = {
 470         .cra_name               =       "aes",
 471         .cra_flags              =       CRYPTO_ALG_TYPE_CIPHER,
 472         .cra_blocksize          =       AES_BLOCK_SIZE,
 473         .cra_ctxsize            =       sizeof(struct aes_ctx),
 474         .cra_alignmask          =       PADLOCK_ALIGNMENT - 1,
 475         .cra_module             =       THIS_MODULE,
 476         .cra_list               =       LIST_HEAD_INIT(aes_alg.cra_list),
 477         .cra_u                  =       {
 478                 .cipher = {
 479                         .cia_min_keysize        =       AES_MIN_KEY_SIZE,
 480                         .cia_max_keysize        =       AES_MAX_KEY_SIZE,
 481                         .cia_setkey             =       aes_set_key,
 482                         .cia_encrypt            =       aes_encrypt,
 483                         .cia_decrypt            =       aes_decrypt,
 484                         .cia_encrypt_ecb        =       aes_encrypt_ecb,
 485                         .cia_decrypt_ecb        =       aes_decrypt_ecb,
 486                         .cia_encrypt_cbc        =       aes_encrypt_cbc,
 487                         .cia_decrypt_cbc        =       aes_decrypt_cbc,
 488                 }
 489         }
 490 };
 491
 492 int __init padlock_init_aes(void)
 493 {
 494         printk(KERN_NOTICE PFX "Using VIA PadLock ACE for AES algorithm.\n");
 495
 496         gen_tabs();
 497         return crypto_register_alg(&aes_alg);
 498 }
 499
 500 void __exit padlock_fini_aes(void)
 501 {
 502         crypto_unregister_alg(&aes_alg);
 503 }