arch/arm64/crypto/aes-neonbs-glue.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Bit sliced AES using NEON instructions
   4  *
   5  * Copyright (C) 2016 - 2017 Linaro Ltd <ard.biesheuvel@linaro.org>
   6  */
   7
   8 #include <asm/neon.h>
   9 #include <asm/simd.h>
  10 #include <crypto/aes.h>
  11 #include <crypto/ctr.h>
  12 #include <crypto/internal/simd.h>
  13 #include <crypto/internal/skcipher.h>
  14 #include <crypto/scatterwalk.h>
  15 #include <crypto/xts.h>
  16 #include <linux/module.h>
  17
  18 MODULE_AUTHOR("Ard Biesheuvel <ard.biesheuvel@linaro.org>");
  19 MODULE_LICENSE("GPL v2");
  20
  21 asmlinkage void aesbs_convert_key(u8 out[], u32 const rk[], int rounds);
  22
  23 asmlinkage void aesbs_ecb_encrypt(u8 out[], u8 const in[], u8 const rk[],
  24                                   int rounds, int blocks);
  25 asmlinkage void aesbs_ecb_decrypt(u8 out[], u8 const in[], u8 const rk[],
  26                                   int rounds, int blocks);
  27
  28 asmlinkage void aesbs_cbc_decrypt(u8 out[], u8 const in[], u8 const rk[],
  29                                   int rounds, int blocks, u8 iv[]);
  30
  31 asmlinkage void aesbs_ctr_encrypt(u8 out[], u8 const in[], u8 const rk[],
  32                                   int rounds, int blocks, u8 iv[], u8 final[]);
  33
  34 asmlinkage void aesbs_xts_encrypt(u8 out[], u8 const in[], u8 const rk[],
  35                                   int rounds, int blocks, u8 iv[]);
  36 asmlinkage void aesbs_xts_decrypt(u8 out[], u8 const in[], u8 const rk[],
  37                                   int rounds, int blocks, u8 iv[]);
  38
  39 /* borrowed from aes-neon-blk.ko */
  40 asmlinkage void neon_aes_ecb_encrypt(u8 out[], u8 const in[], u32 const rk[],
  41                                      int rounds, int blocks);
  42 asmlinkage void neon_aes_cbc_encrypt(u8 out[], u8 const in[], u32 const rk[],
  43                                      int rounds, int blocks, u8 iv[]);
  44 asmlinkage void neon_aes_xts_encrypt(u8 out[], u8 const in[],
  45                                      u32 const rk1[], int rounds, int bytes,
  46                                      u32 const rk2[], u8 iv[], int first);
  47 asmlinkage void neon_aes_xts_decrypt(u8 out[], u8 const in[],
  48                                      u32 const rk1[], int rounds, int bytes,
  49                                      u32 const rk2[], u8 iv[], int first);
  50
  51 struct aesbs_ctx {
  52         u8      rk[13 * (8 * AES_BLOCK_SIZE) + 32];
  53         int     rounds;
  54 } __aligned(AES_BLOCK_SIZE);
  55
  56 struct aesbs_cbc_ctx {
  57         struct aesbs_ctx        key;
  58         u32                     enc[AES_MAX_KEYLENGTH_U32];
  59 };
  60
  61 struct aesbs_ctr_ctx {
  62         struct aesbs_ctx        key;            /* must be first member */
  63         struct crypto_aes_ctx   fallback;
  64 };
  65
  66 struct aesbs_xts_ctx {
  67         struct aesbs_ctx        key;
  68         u32                     twkey[AES_MAX_KEYLENGTH_U32];
  69         struct crypto_aes_ctx   cts;
  70 };
  71
  72 static int aesbs_setkey(struct crypto_skcipher *tfm, const u8 *in_key,
  73                         unsigned int key_len)
  74 {
  75         struct aesbs_ctx *ctx = crypto_skcipher_ctx(tfm);
  76         struct crypto_aes_ctx rk;
  77         int err;
  78
  79         err = aes_expandkey(&rk, in_key, key_len);
  80         if (err)
  81                 return err;
  82
  83         ctx->rounds = 6 + key_len / 4;
  84
  85         kernel_neon_begin();
  86         aesbs_convert_key(ctx->rk, rk.key_enc, ctx->rounds);
  87         kernel_neon_end();
  88
  89         return 0;
  90 }
  91
  92 static int __ecb_crypt(struct skcipher_request *req,
  93                        void (*fn)(u8 out[], u8 const in[], u8 const rk[],
  94                                   int rounds, int blocks))
  95 {
  96         struct crypto_skcipher *tfm = crypto_skcipher_reqtfm(req);
  97         struct aesbs_ctx *ctx = crypto_skcipher_ctx(tfm);
  98         struct skcipher_walk walk;
  99         int err;
 100
 101         err = skcipher_walk_virt(&walk, req, false);
 102
 103         while (walk.nbytes >= AES_BLOCK_SIZE) {
 104                 unsigned int blocks = walk.nbytes / AES_BLOCK_SIZE;
 105
 106                 if (walk.nbytes < walk.total)
 107                         blocks = round_down(blocks,
 108                                             walk.stride / AES_BLOCK_SIZE);
 109
 110                 kernel_neon_begin();
 111                 fn(walk.dst.virt.addr, walk.src.virt.addr, ctx->rk,
 112                    ctx->rounds, blocks);
 113                 kernel_neon_end();
 114                 err = skcipher_walk_done(&walk,
 115                                          walk.nbytes - blocks * AES_BLOCK_SIZE);
 116         }
 117
 118         return err;
 119 }
 120
 121 static int ecb_encrypt(struct skcipher_request *req)
 122 {
 123         return __ecb_crypt(req, aesbs_ecb_encrypt);
 124 }
 125
 126 static int ecb_decrypt(struct skcipher_request *req)
 127 {
 128         return __ecb_crypt(req, aesbs_ecb_decrypt);
 129 }
 130
 131 static int aesbs_cbc_setkey(struct crypto_skcipher *tfm, const u8 *in_key,
 132                             unsigned int key_len)
 133 {
 134         struct aesbs_cbc_ctx *ctx = crypto_skcipher_ctx(tfm);
 135         struct crypto_aes_ctx rk;
 136         int err;
 137
 138         err = aes_expandkey(&rk, in_key, key_len);
 139         if (err)
 140                 return err;
 141
 142         ctx->key.rounds = 6 + key_len / 4;
 143
 144         memcpy(ctx->enc, rk.key_enc, sizeof(ctx->enc));
 145
 146         kernel_neon_begin();
 147         aesbs_convert_key(ctx->key.rk, rk.key_enc, ctx->key.rounds);
 148         kernel_neon_end();
 149         memzero_explicit(&rk, sizeof(rk));
 150
 151         return 0;
 152 }
 153
 154 static int cbc_encrypt(struct skcipher_request *req)
 155 {
 156         struct crypto_skcipher *tfm = crypto_skcipher_reqtfm(req);
 157         struct aesbs_cbc_ctx *ctx = crypto_skcipher_ctx(tfm);
 158         struct skcipher_walk walk;
 159         int err;
 160
 161         err = skcipher_walk_virt(&walk, req, false);
 162
 163         while (walk.nbytes >= AES_BLOCK_SIZE) {
 164                 unsigned int blocks = walk.nbytes / AES_BLOCK_SIZE;
 165
 166                 /* fall back to the non-bitsliced NEON implementation */
 167                 kernel_neon_begin();
 168                 neon_aes_cbc_encrypt(walk.dst.virt.addr, walk.src.virt.addr,
 169                                      ctx->enc, ctx->key.rounds, blocks,
 170                                      walk.iv);
 171                 kernel_neon_end();
 172                 err = skcipher_walk_done(&walk, walk.nbytes % AES_BLOCK_SIZE);
 173         }
 174         return err;
 175 }
 176
 177 static int cbc_decrypt(struct skcipher_request *req)
 178 {
 179         struct crypto_skcipher *tfm = crypto_skcipher_reqtfm(req);
 180         struct aesbs_cbc_ctx *ctx = crypto_skcipher_ctx(tfm);
 181         struct skcipher_walk walk;
 182         int err;
 183
 184         err = skcipher_walk_virt(&walk, req, false);
 185
 186         while (walk.nbytes >= AES_BLOCK_SIZE) {
 187                 unsigned int blocks = walk.nbytes / AES_BLOCK_SIZE;
 188
 189                 if (walk.nbytes < walk.total)
 190                         blocks = round_down(blocks,
 191                                             walk.stride / AES_BLOCK_SIZE);
 192
 193                 kernel_neon_begin();
 194                 aesbs_cbc_decrypt(walk.dst.virt.addr, walk.src.virt.addr,
 195                                   ctx->key.rk, ctx->key.rounds, blocks,
 196                                   walk.iv);
 197                 kernel_neon_end();
 198                 err = skcipher_walk_done(&walk,
 199                                          walk.nbytes - blocks * AES_BLOCK_SIZE);
 200         }
 201
 202         return err;
 203 }
 204
 205 static int aesbs_ctr_setkey_sync(struct crypto_skcipher *tfm, const u8 *in_key,
 206                                  unsigned int key_len)
 207 {
 208         struct aesbs_ctr_ctx *ctx = crypto_skcipher_ctx(tfm);
 209         int err;
 210
 211         err = aes_expandkey(&ctx->fallback, in_key, key_len);
 212         if (err)
 213                 return err;
 214
 215         ctx->key.rounds = 6 + key_len / 4;
 216
 217         kernel_neon_begin();
 218         aesbs_convert_key(ctx->key.rk, ctx->fallback.key_enc, ctx->key.rounds);
 219         kernel_neon_end();
 220
 221         return 0;
 222 }
 223
 224 static int ctr_encrypt(struct skcipher_request *req)
 225 {
 226         struct crypto_skcipher *tfm = crypto_skcipher_reqtfm(req);
 227         struct aesbs_ctx *ctx = crypto_skcipher_ctx(tfm);
 228         struct skcipher_walk walk;
 229         u8 buf[AES_BLOCK_SIZE];
 230         int err;
 231
 232         err = skcipher_walk_virt(&walk, req, false);
 233
 234         while (walk.nbytes > 0) {
 235                 unsigned int blocks = walk.nbytes / AES_BLOCK_SIZE;
 236                 u8 *final = (walk.total % AES_BLOCK_SIZE) ? buf : NULL;
 237
 238                 if (walk.nbytes < walk.total) {
 239                         blocks = round_down(blocks,
 240                                             walk.stride / AES_BLOCK_SIZE);
 241                         final = NULL;
 242                 }
 243
 244                 kernel_neon_begin();
 245                 aesbs_ctr_encrypt(walk.dst.virt.addr, walk.src.virt.addr,
 246                                   ctx->rk, ctx->rounds, blocks, walk.iv, final);
 247                 kernel_neon_end();
 248
 249                 if (final) {
 250                         u8 *dst = walk.dst.virt.addr + blocks * AES_BLOCK_SIZE;
 251                         u8 *src = walk.src.virt.addr + blocks * AES_BLOCK_SIZE;
 252
 253                         crypto_xor_cpy(dst, src, final,
 254                                        walk.total % AES_BLOCK_SIZE);
 255
 256                         err = skcipher_walk_done(&walk, 0);
 257                         break;
 258                 }
 259                 err = skcipher_walk_done(&walk,
 260                                          walk.nbytes - blocks * AES_BLOCK_SIZE);
 261         }
 262         return err;
 263 }
 264
 265 static int aesbs_xts_setkey(struct crypto_skcipher *tfm, const u8 *in_key,
 266                             unsigned int key_len)
 267 {
 268         struct aesbs_xts_ctx *ctx = crypto_skcipher_ctx(tfm);
 269         struct crypto_aes_ctx rk;
 270         int err;
 271
 272         err = xts_verify_key(tfm, in_key, key_len);
 273         if (err)
 274                 return err;
 275
 276         key_len /= 2;
 277         err = aes_expandkey(&ctx->cts, in_key, key_len);
 278         if (err)
 279                 return err;
 280
 281         err = aes_expandkey(&rk, in_key + key_len, key_len);
 282         if (err)
 283                 return err;
 284
 285         memcpy(ctx->twkey, rk.key_enc, sizeof(ctx->twkey));
 286
 287         return aesbs_setkey(tfm, in_key, key_len);
 288 }
 289
 290 static void ctr_encrypt_one(struct crypto_skcipher *tfm, const u8 *src, u8 *dst)
 291 {
 292         struct aesbs_ctr_ctx *ctx = crypto_skcipher_ctx(tfm);
 293         unsigned long flags;
 294
 295         /*
 296          * Temporarily disable interrupts to avoid races where
 297          * cachelines are evicted when the CPU is interrupted
 298          * to do something else.
 299          */
 300         local_irq_save(flags);
 301         aes_encrypt(&ctx->fallback, dst, src);
 302         local_irq_restore(flags);
 303 }
 304
 305 static int ctr_encrypt_sync(struct skcipher_request *req)
 306 {
 307         if (!crypto_simd_usable())
 308                 return crypto_ctr_encrypt_walk(req, ctr_encrypt_one);
 309
 310         return ctr_encrypt(req);
 311 }
 312
 313 static int __xts_crypt(struct skcipher_request *req, bool encrypt,
 314                        void (*fn)(u8 out[], u8 const in[], u8 const rk[],
 315                                   int rounds, int blocks, u8 iv[]))
 316 {
 317         struct crypto_skcipher *tfm = crypto_skcipher_reqtfm(req);
 318         struct aesbs_xts_ctx *ctx = crypto_skcipher_ctx(tfm);
 319         int tail = req->cryptlen % (8 * AES_BLOCK_SIZE);
 320         struct scatterlist sg_src[2], sg_dst[2];
 321         struct skcipher_request subreq;
 322         struct scatterlist *src, *dst;
 323         struct skcipher_walk walk;
 324         int nbytes, err;
 325         int first = 1;
 326         u8 *out, *in;
 327
 328         if (req->cryptlen < AES_BLOCK_SIZE)
 329                 return -EINVAL;
 330
 331         /* ensure that the cts tail is covered by a single step */
 332         if (unlikely(tail > 0 && tail < AES_BLOCK_SIZE)) {
 333                 int xts_blocks = DIV_ROUND_UP(req->cryptlen,
 334                                               AES_BLOCK_SIZE) - 2;
 335
 336                 skcipher_request_set_tfm(&subreq, tfm);
 337                 skcipher_request_set_callback(&subreq,
 338                                               skcipher_request_flags(req),
 339                                               NULL, NULL);
 340                 skcipher_request_set_crypt(&subreq, req->src, req->dst,
 341                                            xts_blocks * AES_BLOCK_SIZE,
 342                                            req->iv);
 343                 req = &subreq;
 344         } else {
 345                 tail = 0;
 346         }
 347
 348         err = skcipher_walk_virt(&walk, req, false);
 349         if (err)
 350                 return err;
 351
 352         while (walk.nbytes >= AES_BLOCK_SIZE) {
 353                 unsigned int blocks = walk.nbytes / AES_BLOCK_SIZE;
 354
 355                 if (walk.nbytes < walk.total || walk.nbytes % AES_BLOCK_SIZE)
 356                         blocks = round_down(blocks,
 357                                             walk.stride / AES_BLOCK_SIZE);
 358
 359                 out = walk.dst.virt.addr;
 360                 in = walk.src.virt.addr;
 361                 nbytes = walk.nbytes;
 362
 363                 kernel_neon_begin();
 364                 if (likely(blocks > 6)) { /* plain NEON is faster otherwise */
 365                         if (first)
 366                                 neon_aes_ecb_encrypt(walk.iv, walk.iv,
 367                                                      ctx->twkey,
 368                                                      ctx->key.rounds, 1);
 369                         first = 0;
 370
 371                         fn(out, in, ctx->key.rk, ctx->key.rounds, blocks,
 372                            walk.iv);
 373
 374                         out += blocks * AES_BLOCK_SIZE;
 375                         in += blocks * AES_BLOCK_SIZE;
 376                         nbytes -= blocks * AES_BLOCK_SIZE;
 377                 }
 378
 379                 if (walk.nbytes == walk.total && nbytes > 0)
 380                         goto xts_tail;
 381
 382                 kernel_neon_end();
 383                 err = skcipher_walk_done(&walk, nbytes);
 384         }
 385
 386         if (err || likely(!tail))
 387                 return err;
 388
 389         /* handle ciphertext stealing */
 390         dst = src = scatterwalk_ffwd(sg_src, req->src, req->cryptlen);
 391         if (req->dst != req->src)
 392                 dst = scatterwalk_ffwd(sg_dst, req->dst, req->cryptlen);
 393
 394         skcipher_request_set_crypt(req, src, dst, AES_BLOCK_SIZE + tail,
 395                                    req->iv);
 396
 397         err = skcipher_walk_virt(&walk, req, false);
 398         if (err)
 399                 return err;
 400
 401         out = walk.dst.virt.addr;
 402         in = walk.src.virt.addr;
 403         nbytes = walk.nbytes;
 404
 405         kernel_neon_begin();
 406 xts_tail:
 407         if (encrypt)
 408                 neon_aes_xts_encrypt(out, in, ctx->cts.key_enc, ctx->key.rounds,
 409                                      nbytes, ctx->twkey, walk.iv, first ?: 2);
 410         else
 411                 neon_aes_xts_decrypt(out, in, ctx->cts.key_dec, ctx->key.rounds,
 412                                      nbytes, ctx->twkey, walk.iv, first ?: 2);
 413         kernel_neon_end();
 414
 415         return skcipher_walk_done(&walk, 0);
 416 }
 417
 418 static int xts_encrypt(struct skcipher_request *req)
 419 {
 420         return __xts_crypt(req, true, aesbs_xts_encrypt);
 421 }
 422
 423 static int xts_decrypt(struct skcipher_request *req)
 424 {
 425         return __xts_crypt(req, false, aesbs_xts_decrypt);
 426 }
 427
 428 static struct skcipher_alg aes_algs[] = { {
 429         .base.cra_name          = "__ecb(aes)",
 430         .base.cra_driver_name   = "__ecb-aes-neonbs",
 431         .base.cra_priority      = 250,
 432         .base.cra_blocksize     = AES_BLOCK_SIZE,
 433         .base.cra_ctxsize       = sizeof(struct aesbs_ctx),
 434         .base.cra_module        = THIS_MODULE,
 435         .base.cra_flags         = CRYPTO_ALG_INTERNAL,
 436
 437         .min_keysize            = AES_MIN_KEY_SIZE,
 438         .max_keysize            = AES_MAX_KEY_SIZE,
 439         .walksize               = 8 * AES_BLOCK_SIZE,
 440         .setkey                 = aesbs_setkey,
 441         .encrypt                = ecb_encrypt,
 442         .decrypt                = ecb_decrypt,
 443 }, {
 444         .base.cra_name          = "__cbc(aes)",
 445         .base.cra_driver_name   = "__cbc-aes-neonbs",
 446         .base.cra_priority      = 250,
 447         .base.cra_blocksize     = AES_BLOCK_SIZE,
 448         .base.cra_ctxsize       = sizeof(struct aesbs_cbc_ctx),
 449         .base.cra_module        = THIS_MODULE,
 450         .base.cra_flags         = CRYPTO_ALG_INTERNAL,
 451
 452         .min_keysize            = AES_MIN_KEY_SIZE,
 453         .max_keysize            = AES_MAX_KEY_SIZE,
 454         .walksize               = 8 * AES_BLOCK_SIZE,
 455         .ivsize                 = AES_BLOCK_SIZE,
 456         .setkey                 = aesbs_cbc_setkey,
 457         .encrypt                = cbc_encrypt,
 458         .decrypt                = cbc_decrypt,
 459 }, {
 460         .base.cra_name          = "__ctr(aes)",
 461         .base.cra_driver_name   = "__ctr-aes-neonbs",
 462         .base.cra_priority      = 250,
 463         .base.cra_blocksize     = 1,
 464         .base.cra_ctxsize       = sizeof(struct aesbs_ctx),
 465         .base.cra_module        = THIS_MODULE,
 466         .base.cra_flags         = CRYPTO_ALG_INTERNAL,
 467
 468         .min_keysize            = AES_MIN_KEY_SIZE,
 469         .max_keysize            = AES_MAX_KEY_SIZE,
 470         .chunksize              = AES_BLOCK_SIZE,
 471         .walksize               = 8 * AES_BLOCK_SIZE,
 472         .ivsize                 = AES_BLOCK_SIZE,
 473         .setkey                 = aesbs_setkey,
 474         .encrypt                = ctr_encrypt,
 475         .decrypt                = ctr_encrypt,
 476 }, {
 477         .base.cra_name          = "ctr(aes)",
 478         .base.cra_driver_name   = "ctr-aes-neonbs",
 479         .base.cra_priority      = 250 - 1,
 480         .base.cra_blocksize     = 1,
 481         .base.cra_ctxsize       = sizeof(struct aesbs_ctr_ctx),
 482         .base.cra_module        = THIS_MODULE,
 483
 484         .min_keysize            = AES_MIN_KEY_SIZE,
 485         .max_keysize            = AES_MAX_KEY_SIZE,
 486         .chunksize              = AES_BLOCK_SIZE,
 487         .walksize               = 8 * AES_BLOCK_SIZE,
 488         .ivsize                 = AES_BLOCK_SIZE,
 489         .setkey                 = aesbs_ctr_setkey_sync,
 490         .encrypt                = ctr_encrypt_sync,
 491         .decrypt                = ctr_encrypt_sync,
 492 }, {
 493         .base.cra_name          = "__xts(aes)",
 494         .base.cra_driver_name   = "__xts-aes-neonbs",
 495         .base.cra_priority      = 250,
 496         .base.cra_blocksize     = AES_BLOCK_SIZE,
 497         .base.cra_ctxsize       = sizeof(struct aesbs_xts_ctx),
 498         .base.cra_module        = THIS_MODULE,
 499         .base.cra_flags         = CRYPTO_ALG_INTERNAL,
 500
 501         .min_keysize            = 2 * AES_MIN_KEY_SIZE,
 502         .max_keysize            = 2 * AES_MAX_KEY_SIZE,
 503         .walksize               = 8 * AES_BLOCK_SIZE,
 504         .ivsize                 = AES_BLOCK_SIZE,
 505         .setkey                 = aesbs_xts_setkey,
 506         .encrypt                = xts_encrypt,
 507         .decrypt                = xts_decrypt,
 508 } };
 509
 510 static struct simd_skcipher_alg *aes_simd_algs[ARRAY_SIZE(aes_algs)];
 511
 512 static void aes_exit(void)
 513 {
 514         int i;
 515
 516         for (i = 0; i < ARRAY_SIZE(aes_simd_algs); i++)
 517                 if (aes_simd_algs[i])
 518                         simd_skcipher_free(aes_simd_algs[i]);
 519
 520         crypto_unregister_skciphers(aes_algs, ARRAY_SIZE(aes_algs));
 521 }
 522
 523 static int __init aes_init(void)
 524 {
 525         struct simd_skcipher_alg *simd;
 526         const char *basename;
 527         const char *algname;
 528         const char *drvname;
 529         int err;
 530         int i;
 531
 532         if (!cpu_have_named_feature(ASIMD))
 533                 return -ENODEV;
 534
 535         err = crypto_register_skciphers(aes_algs, ARRAY_SIZE(aes_algs));
 536         if (err)
 537                 return err;
 538
 539         for (i = 0; i < ARRAY_SIZE(aes_algs); i++) {
 540                 if (!(aes_algs[i].base.cra_flags & CRYPTO_ALG_INTERNAL))
 541                         continue;
 542
 543                 algname = aes_algs[i].base.cra_name + 2;
 544                 drvname = aes_algs[i].base.cra_driver_name + 2;
 545                 basename = aes_algs[i].base.cra_driver_name;
 546                 simd = simd_skcipher_create_compat(algname, drvname, basename);
 547                 err = PTR_ERR(simd);
 548                 if (IS_ERR(simd))
 549                         goto unregister_simds;
 550
 551                 aes_simd_algs[i] = simd;
 552         }
 553         return 0;
 554
 555 unregister_simds:
 556         aes_exit();
 557         return err;
 558 }
 559
 560 module_init(aes_init);
 561 module_exit(aes_exit);