arch/x86/crypto/cast5_avx_glue.c

   1 /*
   2  * Glue Code for the AVX assembler implemention of the Cast5 Cipher
   3  *
   4  * Copyright (C) 2012 Johannes Goetzfried
   5  *     <Johannes.Goetzfried@informatik.stud.uni-erlangen.de>
   6  *
   7  * This program is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License as published by
   9  * the Free Software Foundation; either version 2 of the License, or
  10  * (at your option) any later version.
  11  *
  12  * This program is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  * GNU General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU General Public License
  18  * along with this program; if not, write to the Free Software
  19  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307
  20  * USA
  21  *
  22  */
  23
  24 #include <linux/module.h>
  25 #include <linux/hardirq.h>
  26 #include <linux/types.h>
  27 #include <linux/crypto.h>
  28 #include <linux/err.h>
  29 #include <crypto/ablk_helper.h>
  30 #include <crypto/algapi.h>
  31 #include <crypto/cast5.h>
  32 #include <crypto/cryptd.h>
  33 #include <crypto/ctr.h>
  34 #include <asm/fpu/api.h>
  35 #include <asm/crypto/glue_helper.h>
  36
  37 #define CAST5_PARALLEL_BLOCKS 16
  38
  39 asmlinkage void cast5_ecb_enc_16way(struct cast5_ctx *ctx, u8 *dst,
  40                                     const u8 *src);
  41 asmlinkage void cast5_ecb_dec_16way(struct cast5_ctx *ctx, u8 *dst,
  42                                     const u8 *src);
  43 asmlinkage void cast5_cbc_dec_16way(struct cast5_ctx *ctx, u8 *dst,
  44                                     const u8 *src);
  45 asmlinkage void cast5_ctr_16way(struct cast5_ctx *ctx, u8 *dst, const u8 *src,
  46                                 __be64 *iv);
  47
  48 static inline bool cast5_fpu_begin(bool fpu_enabled, unsigned int nbytes)
  49 {
  50         return glue_fpu_begin(CAST5_BLOCK_SIZE, CAST5_PARALLEL_BLOCKS,
  51                               NULL, fpu_enabled, nbytes);
  52 }
  53
  54 static inline void cast5_fpu_end(bool fpu_enabled)
  55 {
  56         return glue_fpu_end(fpu_enabled);
  57 }
  58
  59 static int ecb_crypt(struct blkcipher_desc *desc, struct blkcipher_walk *walk,
  60                      bool enc)
  61 {
  62         bool fpu_enabled = false;
  63         struct cast5_ctx *ctx = crypto_blkcipher_ctx(desc->tfm);
  64         const unsigned int bsize = CAST5_BLOCK_SIZE;
  65         unsigned int nbytes;
  66         void (*fn)(struct cast5_ctx *ctx, u8 *dst, const u8 *src);
  67         int err;
  68
  69         err = blkcipher_walk_virt(desc, walk);
  70         desc->flags &= ~CRYPTO_TFM_REQ_MAY_SLEEP;
  71
  72         while ((nbytes = walk->nbytes)) {
  73                 u8 *wsrc = walk->src.virt.addr;
  74                 u8 *wdst = walk->dst.virt.addr;
  75
  76                 fpu_enabled = cast5_fpu_begin(fpu_enabled, nbytes);
  77
  78                 /* Process multi-block batch */
  79                 if (nbytes >= bsize * CAST5_PARALLEL_BLOCKS) {
  80                         fn = (enc) ? cast5_ecb_enc_16way : cast5_ecb_dec_16way;
  81                         do {
  82                                 fn(ctx, wdst, wsrc);
  83
  84                                 wsrc += bsize * CAST5_PARALLEL_BLOCKS;
  85                                 wdst += bsize * CAST5_PARALLEL_BLOCKS;
  86                                 nbytes -= bsize * CAST5_PARALLEL_BLOCKS;
  87                         } while (nbytes >= bsize * CAST5_PARALLEL_BLOCKS);
  88
  89                         if (nbytes < bsize)
  90                                 goto done;
  91                 }
  92
  93                 fn = (enc) ? __cast5_encrypt : __cast5_decrypt;
  94
  95                 /* Handle leftovers */
  96                 do {
  97                         fn(ctx, wdst, wsrc);
  98
  99                         wsrc += bsize;
 100                         wdst += bsize;
 101                         nbytes -= bsize;
 102                 } while (nbytes >= bsize);
 103
 104 done:
 105                 err = blkcipher_walk_done(desc, walk, nbytes);
 106         }
 107
 108         cast5_fpu_end(fpu_enabled);
 109         return err;
 110 }
 111
 112 static int ecb_encrypt(struct blkcipher_desc *desc, struct scatterlist *dst,
 113                        struct scatterlist *src, unsigned int nbytes)
 114 {
 115         struct blkcipher_walk walk;
 116
 117         blkcipher_walk_init(&walk, dst, src, nbytes);
 118         return ecb_crypt(desc, &walk, true);
 119 }
 120
 121 static int ecb_decrypt(struct blkcipher_desc *desc, struct scatterlist *dst,
 122                        struct scatterlist *src, unsigned int nbytes)
 123 {
 124         struct blkcipher_walk walk;
 125
 126         blkcipher_walk_init(&walk, dst, src, nbytes);
 127         return ecb_crypt(desc, &walk, false);
 128 }
 129
 130 static unsigned int __cbc_encrypt(struct blkcipher_desc *desc,
 131                                   struct blkcipher_walk *walk)
 132 {
 133         struct cast5_ctx *ctx = crypto_blkcipher_ctx(desc->tfm);
 134         const unsigned int bsize = CAST5_BLOCK_SIZE;
 135         unsigned int nbytes = walk->nbytes;
 136         u64 *src = (u64 *)walk->src.virt.addr;
 137         u64 *dst = (u64 *)walk->dst.virt.addr;
 138         u64 *iv = (u64 *)walk->iv;
 139
 140         do {
 141                 *dst = *src ^ *iv;
 142                 __cast5_encrypt(ctx, (u8 *)dst, (u8 *)dst);
 143                 iv = dst;
 144
 145                 src += 1;
 146                 dst += 1;
 147                 nbytes -= bsize;
 148         } while (nbytes >= bsize);
 149
 150         *(u64 *)walk->iv = *iv;
 151         return nbytes;
 152 }
 153
 154 static int cbc_encrypt(struct blkcipher_desc *desc, struct scatterlist *dst,
 155                        struct scatterlist *src, unsigned int nbytes)
 156 {
 157         struct blkcipher_walk walk;
 158         int err;
 159
 160         blkcipher_walk_init(&walk, dst, src, nbytes);
 161         err = blkcipher_walk_virt(desc, &walk);
 162
 163         while ((nbytes = walk.nbytes)) {
 164                 nbytes = __cbc_encrypt(desc, &walk);
 165                 err = blkcipher_walk_done(desc, &walk, nbytes);
 166         }
 167
 168         return err;
 169 }
 170
 171 static unsigned int __cbc_decrypt(struct blkcipher_desc *desc,
 172                                   struct blkcipher_walk *walk)
 173 {
 174         struct cast5_ctx *ctx = crypto_blkcipher_ctx(desc->tfm);
 175         const unsigned int bsize = CAST5_BLOCK_SIZE;
 176         unsigned int nbytes = walk->nbytes;
 177         u64 *src = (u64 *)walk->src.virt.addr;
 178         u64 *dst = (u64 *)walk->dst.virt.addr;
 179         u64 last_iv;
 180
 181         /* Start of the last block. */
 182         src += nbytes / bsize - 1;
 183         dst += nbytes / bsize - 1;
 184
 185         last_iv = *src;
 186
 187         /* Process multi-block batch */
 188         if (nbytes >= bsize * CAST5_PARALLEL_BLOCKS) {
 189                 do {
 190                         nbytes -= bsize * (CAST5_PARALLEL_BLOCKS - 1);
 191                         src -= CAST5_PARALLEL_BLOCKS - 1;
 192                         dst -= CAST5_PARALLEL_BLOCKS - 1;
 193
 194                         cast5_cbc_dec_16way(ctx, (u8 *)dst, (u8 *)src);
 195
 196                         nbytes -= bsize;
 197                         if (nbytes < bsize)
 198                                 goto done;
 199
 200                         *dst ^= *(src - 1);
 201                         src -= 1;
 202                         dst -= 1;
 203                 } while (nbytes >= bsize * CAST5_PARALLEL_BLOCKS);
 204         }
 205
 206         /* Handle leftovers */
 207         for (;;) {
 208                 __cast5_decrypt(ctx, (u8 *)dst, (u8 *)src);
 209
 210                 nbytes -= bsize;
 211                 if (nbytes < bsize)
 212                         break;
 213
 214                 *dst ^= *(src - 1);
 215                 src -= 1;
 216                 dst -= 1;
 217         }
 218
 219 done:
 220         *dst ^= *(u64 *)walk->iv;
 221         *(u64 *)walk->iv = last_iv;
 222
 223         return nbytes;
 224 }
 225
 226 static int cbc_decrypt(struct blkcipher_desc *desc, struct scatterlist *dst,
 227                        struct scatterlist *src, unsigned int nbytes)
 228 {
 229         bool fpu_enabled = false;
 230         struct blkcipher_walk walk;
 231         int err;
 232
 233         blkcipher_walk_init(&walk, dst, src, nbytes);
 234         err = blkcipher_walk_virt(desc, &walk);
 235         desc->flags &= ~CRYPTO_TFM_REQ_MAY_SLEEP;
 236
 237         while ((nbytes = walk.nbytes)) {
 238                 fpu_enabled = cast5_fpu_begin(fpu_enabled, nbytes);
 239                 nbytes = __cbc_decrypt(desc, &walk);
 240                 err = blkcipher_walk_done(desc, &walk, nbytes);
 241         }
 242
 243         cast5_fpu_end(fpu_enabled);
 244         return err;
 245 }
 246
 247 static void ctr_crypt_final(struct blkcipher_desc *desc,
 248                             struct blkcipher_walk *walk)
 249 {
 250         struct cast5_ctx *ctx = crypto_blkcipher_ctx(desc->tfm);
 251         u8 *ctrblk = walk->iv;
 252         u8 keystream[CAST5_BLOCK_SIZE];
 253         u8 *src = walk->src.virt.addr;
 254         u8 *dst = walk->dst.virt.addr;
 255         unsigned int nbytes = walk->nbytes;
 256
 257         __cast5_encrypt(ctx, keystream, ctrblk);
 258         crypto_xor(keystream, src, nbytes);
 259         memcpy(dst, keystream, nbytes);
 260
 261         crypto_inc(ctrblk, CAST5_BLOCK_SIZE);
 262 }
 263
 264 static unsigned int __ctr_crypt(struct blkcipher_desc *desc,
 265                                 struct blkcipher_walk *walk)
 266 {
 267         struct cast5_ctx *ctx = crypto_blkcipher_ctx(desc->tfm);
 268         const unsigned int bsize = CAST5_BLOCK_SIZE;
 269         unsigned int nbytes = walk->nbytes;
 270         u64 *src = (u64 *)walk->src.virt.addr;
 271         u64 *dst = (u64 *)walk->dst.virt.addr;
 272
 273         /* Process multi-block batch */
 274         if (nbytes >= bsize * CAST5_PARALLEL_BLOCKS) {
 275                 do {
 276                         cast5_ctr_16way(ctx, (u8 *)dst, (u8 *)src,
 277                                         (__be64 *)walk->iv);
 278
 279                         src += CAST5_PARALLEL_BLOCKS;
 280                         dst += CAST5_PARALLEL_BLOCKS;
 281                         nbytes -= bsize * CAST5_PARALLEL_BLOCKS;
 282                 } while (nbytes >= bsize * CAST5_PARALLEL_BLOCKS);
 283
 284                 if (nbytes < bsize)
 285                         goto done;
 286         }
 287
 288         /* Handle leftovers */
 289         do {
 290                 u64 ctrblk;
 291
 292                 if (dst != src)
 293                         *dst = *src;
 294
 295                 ctrblk = *(u64 *)walk->iv;
 296                 be64_add_cpu((__be64 *)walk->iv, 1);
 297
 298                 __cast5_encrypt(ctx, (u8 *)&ctrblk, (u8 *)&ctrblk);
 299                 *dst ^= ctrblk;
 300
 301                 src += 1;
 302                 dst += 1;
 303                 nbytes -= bsize;
 304         } while (nbytes >= bsize);
 305
 306 done:
 307         return nbytes;
 308 }
 309
 310 static int ctr_crypt(struct blkcipher_desc *desc, struct scatterlist *dst,
 311                      struct scatterlist *src, unsigned int nbytes)
 312 {
 313         bool fpu_enabled = false;
 314         struct blkcipher_walk walk;
 315         int err;
 316
 317         blkcipher_walk_init(&walk, dst, src, nbytes);
 318         err = blkcipher_walk_virt_block(desc, &walk, CAST5_BLOCK_SIZE);
 319         desc->flags &= ~CRYPTO_TFM_REQ_MAY_SLEEP;
 320
 321         while ((nbytes = walk.nbytes) >= CAST5_BLOCK_SIZE) {
 322                 fpu_enabled = cast5_fpu_begin(fpu_enabled, nbytes);
 323                 nbytes = __ctr_crypt(desc, &walk);
 324                 err = blkcipher_walk_done(desc, &walk, nbytes);
 325         }
 326
 327         cast5_fpu_end(fpu_enabled);
 328
 329         if (walk.nbytes) {
 330                 ctr_crypt_final(desc, &walk);
 331                 err = blkcipher_walk_done(desc, &walk, 0);
 332         }
 333
 334         return err;
 335 }
 336
 337
 338 static struct crypto_alg cast5_algs[6] = { {
 339         .cra_name               = "__ecb-cast5-avx",
 340         .cra_driver_name        = "__driver-ecb-cast5-avx",
 341         .cra_priority           = 0,
 342         .cra_flags              = CRYPTO_ALG_TYPE_BLKCIPHER |
 343                                   CRYPTO_ALG_INTERNAL,
 344         .cra_blocksize          = CAST5_BLOCK_SIZE,
 345         .cra_ctxsize            = sizeof(struct cast5_ctx),
 346         .cra_alignmask          = 0,
 347         .cra_type               = &crypto_blkcipher_type,
 348         .cra_module             = THIS_MODULE,
 349         .cra_u = {
 350                 .blkcipher = {
 351                         .min_keysize    = CAST5_MIN_KEY_SIZE,
 352                         .max_keysize    = CAST5_MAX_KEY_SIZE,
 353                         .setkey         = cast5_setkey,
 354                         .encrypt        = ecb_encrypt,
 355                         .decrypt        = ecb_decrypt,
 356                 },
 357         },
 358 }, {
 359         .cra_name               = "__cbc-cast5-avx",
 360         .cra_driver_name        = "__driver-cbc-cast5-avx",
 361         .cra_priority           = 0,
 362         .cra_flags              = CRYPTO_ALG_TYPE_BLKCIPHER |
 363                                   CRYPTO_ALG_INTERNAL,
 364         .cra_blocksize          = CAST5_BLOCK_SIZE,
 365         .cra_ctxsize            = sizeof(struct cast5_ctx),
 366         .cra_alignmask          = 0,
 367         .cra_type               = &crypto_blkcipher_type,
 368         .cra_module             = THIS_MODULE,
 369         .cra_u = {
 370                 .blkcipher = {
 371                         .min_keysize    = CAST5_MIN_KEY_SIZE,
 372                         .max_keysize    = CAST5_MAX_KEY_SIZE,
 373                         .setkey         = cast5_setkey,
 374                         .encrypt        = cbc_encrypt,
 375                         .decrypt        = cbc_decrypt,
 376                 },
 377         },
 378 }, {
 379         .cra_name               = "__ctr-cast5-avx",
 380         .cra_driver_name        = "__driver-ctr-cast5-avx",
 381         .cra_priority           = 0,
 382         .cra_flags              = CRYPTO_ALG_TYPE_BLKCIPHER |
 383                                   CRYPTO_ALG_INTERNAL,
 384         .cra_blocksize          = 1,
 385         .cra_ctxsize            = sizeof(struct cast5_ctx),
 386         .cra_alignmask          = 0,
 387         .cra_type               = &crypto_blkcipher_type,
 388         .cra_module             = THIS_MODULE,
 389         .cra_u = {
 390                 .blkcipher = {
 391                         .min_keysize    = CAST5_MIN_KEY_SIZE,
 392                         .max_keysize    = CAST5_MAX_KEY_SIZE,
 393                         .ivsize         = CAST5_BLOCK_SIZE,
 394                         .setkey         = cast5_setkey,
 395                         .encrypt        = ctr_crypt,
 396                         .decrypt        = ctr_crypt,
 397                 },
 398         },
 399 }, {
 400         .cra_name               = "ecb(cast5)",
 401         .cra_driver_name        = "ecb-cast5-avx",
 402         .cra_priority           = 200,
 403         .cra_flags              = CRYPTO_ALG_TYPE_ABLKCIPHER | CRYPTO_ALG_ASYNC,
 404         .cra_blocksize          = CAST5_BLOCK_SIZE,
 405         .cra_ctxsize            = sizeof(struct async_helper_ctx),
 406         .cra_alignmask          = 0,
 407         .cra_type               = &crypto_ablkcipher_type,
 408         .cra_module             = THIS_MODULE,
 409         .cra_init               = ablk_init,
 410         .cra_exit               = ablk_exit,
 411         .cra_u = {
 412                 .ablkcipher = {
 413                         .min_keysize    = CAST5_MIN_KEY_SIZE,
 414                         .max_keysize    = CAST5_MAX_KEY_SIZE,
 415                         .setkey         = ablk_set_key,
 416                         .encrypt        = ablk_encrypt,
 417                         .decrypt        = ablk_decrypt,
 418                 },
 419         },
 420 }, {
 421         .cra_name               = "cbc(cast5)",
 422         .cra_driver_name        = "cbc-cast5-avx",
 423         .cra_priority           = 200,
 424         .cra_flags              = CRYPTO_ALG_TYPE_ABLKCIPHER | CRYPTO_ALG_ASYNC,
 425         .cra_blocksize          = CAST5_BLOCK_SIZE,
 426         .cra_ctxsize            = sizeof(struct async_helper_ctx),
 427         .cra_alignmask          = 0,
 428         .cra_type               = &crypto_ablkcipher_type,
 429         .cra_module             = THIS_MODULE,
 430         .cra_init               = ablk_init,
 431         .cra_exit               = ablk_exit,
 432         .cra_u = {
 433                 .ablkcipher = {
 434                         .min_keysize    = CAST5_MIN_KEY_SIZE,
 435                         .max_keysize    = CAST5_MAX_KEY_SIZE,
 436                         .ivsize         = CAST5_BLOCK_SIZE,
 437                         .setkey         = ablk_set_key,
 438                         .encrypt        = __ablk_encrypt,
 439                         .decrypt        = ablk_decrypt,
 440                 },
 441         },
 442 }, {
 443         .cra_name               = "ctr(cast5)",
 444         .cra_driver_name        = "ctr-cast5-avx",
 445         .cra_priority           = 200,
 446         .cra_flags              = CRYPTO_ALG_TYPE_ABLKCIPHER | CRYPTO_ALG_ASYNC,
 447         .cra_blocksize          = 1,
 448         .cra_ctxsize            = sizeof(struct async_helper_ctx),
 449         .cra_alignmask          = 0,
 450         .cra_type               = &crypto_ablkcipher_type,
 451         .cra_module             = THIS_MODULE,
 452         .cra_init               = ablk_init,
 453         .cra_exit               = ablk_exit,
 454         .cra_u = {
 455                 .ablkcipher = {
 456                         .min_keysize    = CAST5_MIN_KEY_SIZE,
 457                         .max_keysize    = CAST5_MAX_KEY_SIZE,
 458                         .ivsize         = CAST5_BLOCK_SIZE,
 459                         .setkey         = ablk_set_key,
 460                         .encrypt        = ablk_encrypt,
 461                         .decrypt        = ablk_encrypt,
 462                         .geniv          = "chainiv",
 463                 },
 464         },
 465 } };
 466
 467 static int __init cast5_init(void)
 468 {
 469         const char *feature_name;
 470
 471         if (!cpu_has_xfeatures(XFEATURE_MASK_SSE | XFEATURE_MASK_YMM,
 472                                 &feature_name)) {
 473                 pr_info("CPU feature '%s' is not supported.\n", feature_name);
 474                 return -ENODEV;
 475         }
 476
 477         return crypto_register_algs(cast5_algs, ARRAY_SIZE(cast5_algs));
 478 }
 479
 480 static void __exit cast5_exit(void)
 481 {
 482         crypto_unregister_algs(cast5_algs, ARRAY_SIZE(cast5_algs));
 483 }
 484
 485 module_init(cast5_init);
 486 module_exit(cast5_exit);
 487
 488 MODULE_DESCRIPTION("Cast5 Cipher Algorithm, AVX optimized");
 489 MODULE_LICENSE("GPL");
 490 MODULE_ALIAS_CRYPTO("cast5");