kernel/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/xcr.h>
  35 #include <asm/xsave.h>
  36 #include <asm/crypto/glue_helper.h>
  37
  38 #define CAST5_PARALLEL_BLOCKS 16
  39
  40 asmlinkage void cast5_ecb_enc_16way(struct cast5_ctx *ctx, u8 *dst,
  41                                     const u8 *src);
  42 asmlinkage void cast5_ecb_dec_16way(struct cast5_ctx *ctx, u8 *dst,
  43                                     const u8 *src);
  44 asmlinkage void cast5_cbc_dec_16way(struct cast5_ctx *ctx, u8 *dst,
  45                                     const u8 *src);
  46 asmlinkage void cast5_ctr_16way(struct cast5_ctx *ctx, u8 *dst, const u8 *src,
  47                                 __be64 *iv);
  48
  49 static inline bool cast5_fpu_begin(bool fpu_enabled, unsigned int nbytes)
  50 {
  51         return glue_fpu_begin(CAST5_BLOCK_SIZE, CAST5_PARALLEL_BLOCKS,
  52                               NULL, fpu_enabled, nbytes);
  53 }
  54
  55 static inline void cast5_fpu_end(bool fpu_enabled)
  56 {
  57         return glue_fpu_end(fpu_enabled);
  58 }
  59
  60 static int ecb_crypt(struct blkcipher_desc *desc, struct blkcipher_walk *walk,
  61                      bool enc)
  62 {
  63         bool fpu_enabled;
  64         struct cast5_ctx *ctx = crypto_blkcipher_ctx(desc->tfm);
  65         const unsigned int bsize = CAST5_BLOCK_SIZE;
  66         unsigned int nbytes;
  67         void (*fn)(struct cast5_ctx *ctx, u8 *dst, const u8 *src);
  68         int err;
  69
  70         fn = (enc) ? cast5_ecb_enc_16way : cast5_ecb_dec_16way;
  71
  72         err = blkcipher_walk_virt(desc, walk);
  73         desc->flags &= ~CRYPTO_TFM_REQ_MAY_SLEEP;
  74
  75         while ((nbytes = walk->nbytes)) {
  76                 u8 *wsrc = walk->src.virt.addr;
  77                 u8 *wdst = walk->dst.virt.addr;
  78
  79                 fpu_enabled = cast5_fpu_begin(false, nbytes);
  80
  81                 /* Process multi-block batch */
  82                 if (nbytes >= bsize * CAST5_PARALLEL_BLOCKS) {
  83                         do {
  84                                 fn(ctx, wdst, wsrc);
  85
  86                                 wsrc += bsize * CAST5_PARALLEL_BLOCKS;
  87                                 wdst += bsize * CAST5_PARALLEL_BLOCKS;
  88                                 nbytes -= bsize * CAST5_PARALLEL_BLOCKS;
  89                         } while (nbytes >= bsize * CAST5_PARALLEL_BLOCKS);
  90
  91                         if (nbytes < bsize)
  92                                 goto done;
  93                 }
  94
  95                 fn = (enc) ? __cast5_encrypt : __cast5_decrypt;
  96
  97                 /* Handle leftovers */
  98                 do {
  99                         fn(ctx, wdst, wsrc);
 100
 101                         wsrc += bsize;
 102                         wdst += bsize;
 103                         nbytes -= bsize;
 104                 } while (nbytes >= bsize);
 105
 106 done:
 107                 cast5_fpu_end(fpu_enabled);
 108                 err = blkcipher_walk_done(desc, walk, nbytes);
 109         }
 110         return err;
 111 }
 112
 113 static int ecb_encrypt(struct blkcipher_desc *desc, struct scatterlist *dst,
 114                        struct scatterlist *src, unsigned int nbytes)
 115 {
 116         struct blkcipher_walk walk;
 117
 118         blkcipher_walk_init(&walk, dst, src, nbytes);
 119         return ecb_crypt(desc, &walk, true);
 120 }
 121
 122 static int ecb_decrypt(struct blkcipher_desc *desc, struct scatterlist *dst,
 123                        struct scatterlist *src, unsigned int nbytes)
 124 {
 125         struct blkcipher_walk walk;
 126
 127         blkcipher_walk_init(&walk, dst, src, nbytes);
 128         return ecb_crypt(desc, &walk, false);
 129 }
 130
 131 static unsigned int __cbc_encrypt(struct blkcipher_desc *desc,
 132                                   struct blkcipher_walk *walk)
 133 {
 134         struct cast5_ctx *ctx = crypto_blkcipher_ctx(desc->tfm);
 135         const unsigned int bsize = CAST5_BLOCK_SIZE;
 136         unsigned int nbytes = walk->nbytes;
 137         u64 *src = (u64 *)walk->src.virt.addr;
 138         u64 *dst = (u64 *)walk->dst.virt.addr;
 139         u64 *iv = (u64 *)walk->iv;
 140
 141         do {
 142                 *dst = *src ^ *iv;
 143                 __cast5_encrypt(ctx, (u8 *)dst, (u8 *)dst);
 144                 iv = dst;
 145
 146                 src += 1;
 147                 dst += 1;
 148                 nbytes -= bsize;
 149         } while (nbytes >= bsize);
 150
 151         *(u64 *)walk->iv = *iv;
 152         return nbytes;
 153 }
 154
 155 static int cbc_encrypt(struct blkcipher_desc *desc, struct scatterlist *dst,
 156                        struct scatterlist *src, unsigned int nbytes)
 157 {
 158         struct blkcipher_walk walk;
 159         int err;
 160
 161         blkcipher_walk_init(&walk, dst, src, nbytes);
 162         err = blkcipher_walk_virt(desc, &walk);
 163
 164         while ((nbytes = walk.nbytes)) {
 165                 nbytes = __cbc_encrypt(desc, &walk);
 166                 err = blkcipher_walk_done(desc, &walk, nbytes);
 167         }
 168
 169         return err;
 170 }
 171
 172 static unsigned int __cbc_decrypt(struct blkcipher_desc *desc,
 173                                   struct blkcipher_walk *walk)
 174 {
 175         struct cast5_ctx *ctx = crypto_blkcipher_ctx(desc->tfm);
 176         const unsigned int bsize = CAST5_BLOCK_SIZE;
 177         unsigned int nbytes = walk->nbytes;
 178         u64 *src = (u64 *)walk->src.virt.addr;
 179         u64 *dst = (u64 *)walk->dst.virt.addr;
 180         u64 last_iv;
 181
 182         /* Start of the last block. */
 183         src += nbytes / bsize - 1;
 184         dst += nbytes / bsize - 1;
 185
 186         last_iv = *src;
 187
 188         /* Process multi-block batch */
 189         if (nbytes >= bsize * CAST5_PARALLEL_BLOCKS) {
 190                 do {
 191                         nbytes -= bsize * (CAST5_PARALLEL_BLOCKS - 1);
 192                         src -= CAST5_PARALLEL_BLOCKS - 1;
 193                         dst -= CAST5_PARALLEL_BLOCKS - 1;
 194
 195                         cast5_cbc_dec_16way(ctx, (u8 *)dst, (u8 *)src);
 196
 197                         nbytes -= bsize;
 198                         if (nbytes < bsize)
 199                                 goto done;
 200
 201                         *dst ^= *(src - 1);
 202                         src -= 1;
 203                         dst -= 1;
 204                 } while (nbytes >= bsize * CAST5_PARALLEL_BLOCKS);
 205         }
 206
 207         /* Handle leftovers */
 208         for (;;) {
 209                 __cast5_decrypt(ctx, (u8 *)dst, (u8 *)src);
 210
 211                 nbytes -= bsize;
 212                 if (nbytes < bsize)
 213                         break;
 214
 215                 *dst ^= *(src - 1);
 216                 src -= 1;
 217                 dst -= 1;
 218         }
 219
 220 done:
 221         *dst ^= *(u64 *)walk->iv;
 222         *(u64 *)walk->iv = last_iv;
 223
 224         return nbytes;
 225 }
 226
 227 static int cbc_decrypt(struct blkcipher_desc *desc, struct scatterlist *dst,
 228                        struct scatterlist *src, unsigned int nbytes)
 229 {
 230         bool fpu_enabled;
 231         struct blkcipher_walk walk;
 232         int err;
 233
 234         blkcipher_walk_init(&walk, dst, src, nbytes);
 235         err = blkcipher_walk_virt(desc, &walk);
 236         desc->flags &= ~CRYPTO_TFM_REQ_MAY_SLEEP;
 237
 238         while ((nbytes = walk.nbytes)) {
 239                 fpu_enabled = cast5_fpu_begin(false, nbytes);
 240                 nbytes = __cbc_decrypt(desc, &walk);
 241                 cast5_fpu_end(fpu_enabled);
 242                 err = blkcipher_walk_done(desc, &walk, nbytes);
 243         }
 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;
 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(false, nbytes);
 323                 nbytes = __ctr_crypt(desc, &walk);
 324                 cast5_fpu_end(fpu_enabled);
 325                 err = blkcipher_walk_done(desc, &walk, nbytes);
 326         }
 327
 328         if (walk.nbytes) {
 329                 ctr_crypt_final(desc, &walk);
 330                 err = blkcipher_walk_done(desc, &walk, 0);
 331         }
 332
 333         return err;
 334 }
 335
 336
 337 static struct crypto_alg cast5_algs[6] = { {
 338         .cra_name               = "__ecb-cast5-avx",
 339         .cra_driver_name        = "__driver-ecb-cast5-avx",
 340         .cra_priority           = 0,
 341         .cra_flags              = CRYPTO_ALG_TYPE_BLKCIPHER |
 342                                   CRYPTO_ALG_INTERNAL,
 343         .cra_blocksize          = CAST5_BLOCK_SIZE,
 344         .cra_ctxsize            = sizeof(struct cast5_ctx),
 345         .cra_alignmask          = 0,
 346         .cra_type               = &crypto_blkcipher_type,
 347         .cra_module             = THIS_MODULE,
 348         .cra_u = {
 349                 .blkcipher = {
 350                         .min_keysize    = CAST5_MIN_KEY_SIZE,
 351                         .max_keysize    = CAST5_MAX_KEY_SIZE,
 352                         .setkey         = cast5_setkey,
 353                         .encrypt        = ecb_encrypt,
 354                         .decrypt        = ecb_decrypt,
 355                 },
 356         },
 357 }, {
 358         .cra_name               = "__cbc-cast5-avx",
 359         .cra_driver_name        = "__driver-cbc-cast5-avx",
 360         .cra_priority           = 0,
 361         .cra_flags              = CRYPTO_ALG_TYPE_BLKCIPHER |
 362                                   CRYPTO_ALG_INTERNAL,
 363         .cra_blocksize          = CAST5_BLOCK_SIZE,
 364         .cra_ctxsize            = sizeof(struct cast5_ctx),
 365         .cra_alignmask          = 0,
 366         .cra_type               = &crypto_blkcipher_type,
 367         .cra_module             = THIS_MODULE,
 368         .cra_u = {
 369                 .blkcipher = {
 370                         .min_keysize    = CAST5_MIN_KEY_SIZE,
 371                         .max_keysize    = CAST5_MAX_KEY_SIZE,
 372                         .setkey         = cast5_setkey,
 373                         .encrypt        = cbc_encrypt,
 374                         .decrypt        = cbc_decrypt,
 375                 },
 376         },
 377 }, {
 378         .cra_name               = "__ctr-cast5-avx",
 379         .cra_driver_name        = "__driver-ctr-cast5-avx",
 380         .cra_priority           = 0,
 381         .cra_flags              = CRYPTO_ALG_TYPE_BLKCIPHER |
 382                                   CRYPTO_ALG_INTERNAL,
 383         .cra_blocksize          = 1,
 384         .cra_ctxsize            = sizeof(struct cast5_ctx),
 385         .cra_alignmask          = 0,
 386         .cra_type               = &crypto_blkcipher_type,
 387         .cra_module             = THIS_MODULE,
 388         .cra_u = {
 389                 .blkcipher = {
 390                         .min_keysize    = CAST5_MIN_KEY_SIZE,
 391                         .max_keysize    = CAST5_MAX_KEY_SIZE,
 392                         .ivsize         = CAST5_BLOCK_SIZE,
 393                         .setkey         = cast5_setkey,
 394                         .encrypt        = ctr_crypt,
 395                         .decrypt        = ctr_crypt,
 396                 },
 397         },
 398 }, {
 399         .cra_name               = "ecb(cast5)",
 400         .cra_driver_name        = "ecb-cast5-avx",
 401         .cra_priority           = 200,
 402         .cra_flags              = CRYPTO_ALG_TYPE_ABLKCIPHER | CRYPTO_ALG_ASYNC,
 403         .cra_blocksize          = CAST5_BLOCK_SIZE,
 404         .cra_ctxsize            = sizeof(struct async_helper_ctx),
 405         .cra_alignmask          = 0,
 406         .cra_type               = &crypto_ablkcipher_type,
 407         .cra_module             = THIS_MODULE,
 408         .cra_init               = ablk_init,
 409         .cra_exit               = ablk_exit,
 410         .cra_u = {
 411                 .ablkcipher = {
 412                         .min_keysize    = CAST5_MIN_KEY_SIZE,
 413                         .max_keysize    = CAST5_MAX_KEY_SIZE,
 414                         .setkey         = ablk_set_key,
 415                         .encrypt        = ablk_encrypt,
 416                         .decrypt        = ablk_decrypt,
 417                 },
 418         },
 419 }, {
 420         .cra_name               = "cbc(cast5)",
 421         .cra_driver_name        = "cbc-cast5-avx",
 422         .cra_priority           = 200,
 423         .cra_flags              = CRYPTO_ALG_TYPE_ABLKCIPHER | CRYPTO_ALG_ASYNC,
 424         .cra_blocksize          = CAST5_BLOCK_SIZE,
 425         .cra_ctxsize            = sizeof(struct async_helper_ctx),
 426         .cra_alignmask          = 0,
 427         .cra_type               = &crypto_ablkcipher_type,
 428         .cra_module             = THIS_MODULE,
 429         .cra_init               = ablk_init,
 430         .cra_exit               = ablk_exit,
 431         .cra_u = {
 432                 .ablkcipher = {
 433                         .min_keysize    = CAST5_MIN_KEY_SIZE,
 434                         .max_keysize    = CAST5_MAX_KEY_SIZE,
 435                         .ivsize         = CAST5_BLOCK_SIZE,
 436                         .setkey         = ablk_set_key,
 437                         .encrypt        = __ablk_encrypt,
 438                         .decrypt        = ablk_decrypt,
 439                 },
 440         },
 441 }, {
 442         .cra_name               = "ctr(cast5)",
 443         .cra_driver_name        = "ctr-cast5-avx",
 444         .cra_priority           = 200,
 445         .cra_flags              = CRYPTO_ALG_TYPE_ABLKCIPHER | CRYPTO_ALG_ASYNC,
 446         .cra_blocksize          = 1,
 447         .cra_ctxsize            = sizeof(struct async_helper_ctx),
 448         .cra_alignmask          = 0,
 449         .cra_type               = &crypto_ablkcipher_type,
 450         .cra_module             = THIS_MODULE,
 451         .cra_init               = ablk_init,
 452         .cra_exit               = ablk_exit,
 453         .cra_u = {
 454                 .ablkcipher = {
 455                         .min_keysize    = CAST5_MIN_KEY_SIZE,
 456                         .max_keysize    = CAST5_MAX_KEY_SIZE,
 457                         .ivsize         = CAST5_BLOCK_SIZE,
 458                         .setkey         = ablk_set_key,
 459                         .encrypt        = ablk_encrypt,
 460                         .decrypt        = ablk_encrypt,
 461                         .geniv          = "chainiv",
 462                 },
 463         },
 464 } };
 465
 466 static int __init cast5_init(void)
 467 {
 468         u64 xcr0;
 469
 470         if (!cpu_has_avx || !cpu_has_osxsave) {
 471                 pr_info("AVX instructions are not detected.\n");
 472                 return -ENODEV;
 473         }
 474
 475         xcr0 = xgetbv(XCR_XFEATURE_ENABLED_MASK);
 476         if ((xcr0 & (XSTATE_SSE | XSTATE_YMM)) != (XSTATE_SSE | XSTATE_YMM)) {
 477                 pr_info("AVX detected but unusable.\n");
 478                 return -ENODEV;
 479         }
 480
 481         return crypto_register_algs(cast5_algs, ARRAY_SIZE(cast5_algs));
 482 }
 483
 484 static void __exit cast5_exit(void)
 485 {
 486         crypto_unregister_algs(cast5_algs, ARRAY_SIZE(cast5_algs));
 487 }
 488
 489 module_init(cast5_init);
 490 module_exit(cast5_exit);
 491
 492 MODULE_DESCRIPTION("Cast5 Cipher Algorithm, AVX optimized");
 493 MODULE_LICENSE("GPL");
 494 MODULE_ALIAS_CRYPTO("cast5");