arch/arm64/crypto/ghash-ce-glue.c

   1 /*
   2  * Accelerated GHASH implementation with ARMv8 PMULL instructions.
   3  *
   4  * Copyright (C) 2014 - 2018 Linaro Ltd. <ard.biesheuvel@linaro.org>
   5  *
   6  * This program is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License version 2 as published
   8  * by the Free Software Foundation.
   9  */
  10
  11 #include <asm/neon.h>
  12 #include <asm/simd.h>
  13 #include <asm/unaligned.h>
  14 #include <crypto/aes.h>
  15 #include <crypto/algapi.h>
  16 #include <crypto/b128ops.h>
  17 #include <crypto/gf128mul.h>
  18 #include <crypto/internal/aead.h>
  19 #include <crypto/internal/hash.h>
  20 #include <crypto/internal/skcipher.h>
  21 #include <crypto/scatterwalk.h>
  22 #include <linux/cpufeature.h>
  23 #include <linux/crypto.h>
  24 #include <linux/module.h>
  25
  26 MODULE_DESCRIPTION("GHASH and AES-GCM using ARMv8 Crypto Extensions");
  27 MODULE_AUTHOR("Ard Biesheuvel <ard.biesheuvel@linaro.org>");
  28 MODULE_LICENSE("GPL v2");
  29 MODULE_ALIAS_CRYPTO("ghash");
  30
  31 #define GHASH_BLOCK_SIZE        16
  32 #define GHASH_DIGEST_SIZE       16
  33 #define GCM_IV_SIZE             12
  34
  35 struct ghash_key {
  36         u64                     h[2];
  37         u64                     h2[2];
  38         u64                     h3[2];
  39         u64                     h4[2];
  40
  41         be128                   k;
  42 };
  43
  44 struct ghash_desc_ctx {
  45         u64 digest[GHASH_DIGEST_SIZE/sizeof(u64)];
  46         u8 buf[GHASH_BLOCK_SIZE];
  47         u32 count;
  48 };
  49
  50 struct gcm_aes_ctx {
  51         struct crypto_aes_ctx   aes_key;
  52         struct ghash_key        ghash_key;
  53 };
  54
  55 asmlinkage void pmull_ghash_update_p64(int blocks, u64 dg[], const char *src,
  56                                        struct ghash_key const *k,
  57                                        const char *head);
  58
  59 asmlinkage void pmull_ghash_update_p8(int blocks, u64 dg[], const char *src,
  60                                       struct ghash_key const *k,
  61                                       const char *head);
  62
  63 asmlinkage void pmull_gcm_encrypt(int blocks, u64 dg[], u8 dst[],
  64                                   const u8 src[], struct ghash_key const *k,
  65                                   u8 ctr[], u32 const rk[], int rounds,
  66                                   u8 ks[]);
  67
  68 asmlinkage void pmull_gcm_decrypt(int blocks, u64 dg[], u8 dst[],
  69                                   const u8 src[], struct ghash_key const *k,
  70                                   u8 ctr[], u32 const rk[], int rounds);
  71
  72 asmlinkage void pmull_gcm_encrypt_block(u8 dst[], u8 const src[],
  73                                         u32 const rk[], int rounds);
  74
  75 asmlinkage void __aes_arm64_encrypt(u32 *rk, u8 *out, const u8 *in, int rounds);
  76
  77 static int ghash_init(struct shash_desc *desc)
  78 {
  79         struct ghash_desc_ctx *ctx = shash_desc_ctx(desc);
  80
  81         *ctx = (struct ghash_desc_ctx){};
  82         return 0;
  83 }
  84
  85 static void ghash_do_update(int blocks, u64 dg[], const char *src,
  86                             struct ghash_key *key, const char *head,
  87                             void (*simd_update)(int blocks, u64 dg[],
  88                                                 const char *src,
  89                                                 struct ghash_key const *k,
  90                                                 const char *head))
  91 {
  92         if (likely(may_use_simd())) {
  93                 kernel_neon_begin();
  94                 simd_update(blocks, dg, src, key, head);
  95                 kernel_neon_end();
  96         } else {
  97                 be128 dst = { cpu_to_be64(dg[1]), cpu_to_be64(dg[0]) };
  98
  99                 do {
 100                         const u8 *in = src;
 101
 102                         if (head) {
 103                                 in = head;
 104                                 blocks++;
 105                                 head = NULL;
 106                         } else {
 107                                 src += GHASH_BLOCK_SIZE;
 108                         }
 109
 110                         crypto_xor((u8 *)&dst, in, GHASH_BLOCK_SIZE);
 111                         gf128mul_lle(&dst, &key->k);
 112                 } while (--blocks);
 113
 114                 dg[0] = be64_to_cpu(dst.b);
 115                 dg[1] = be64_to_cpu(dst.a);
 116         }
 117 }
 118
 119 /* avoid hogging the CPU for too long */
 120 #define MAX_BLOCKS      (SZ_64K / GHASH_BLOCK_SIZE)
 121
 122 static int __ghash_update(struct shash_desc *desc, const u8 *src,
 123                           unsigned int len,
 124                           void (*simd_update)(int blocks, u64 dg[],
 125                                               const char *src,
 126                                               struct ghash_key const *k,
 127                                               const char *head))
 128 {
 129         struct ghash_desc_ctx *ctx = shash_desc_ctx(desc);
 130         unsigned int partial = ctx->count % GHASH_BLOCK_SIZE;
 131
 132         ctx->count += len;
 133
 134         if ((partial + len) >= GHASH_BLOCK_SIZE) {
 135                 struct ghash_key *key = crypto_shash_ctx(desc->tfm);
 136                 int blocks;
 137
 138                 if (partial) {
 139                         int p = GHASH_BLOCK_SIZE - partial;
 140
 141                         memcpy(ctx->buf + partial, src, p);
 142                         src += p;
 143                         len -= p;
 144                 }
 145
 146                 blocks = len / GHASH_BLOCK_SIZE;
 147                 len %= GHASH_BLOCK_SIZE;
 148
 149                 do {
 150                         int chunk = min(blocks, MAX_BLOCKS);
 151
 152                         ghash_do_update(chunk, ctx->digest, src, key,
 153                                         partial ? ctx->buf : NULL,
 154                                         simd_update);
 155
 156                         blocks -= chunk;
 157                         src += chunk * GHASH_BLOCK_SIZE;
 158                         partial = 0;
 159                 } while (unlikely(blocks > 0));
 160         }
 161         if (len)
 162                 memcpy(ctx->buf + partial, src, len);
 163         return 0;
 164 }
 165
 166 static int ghash_update_p8(struct shash_desc *desc, const u8 *src,
 167                            unsigned int len)
 168 {
 169         return __ghash_update(desc, src, len, pmull_ghash_update_p8);
 170 }
 171
 172 static int ghash_update_p64(struct shash_desc *desc, const u8 *src,
 173                             unsigned int len)
 174 {
 175         return __ghash_update(desc, src, len, pmull_ghash_update_p64);
 176 }
 177
 178 static int ghash_final_p8(struct shash_desc *desc, u8 *dst)
 179 {
 180         struct ghash_desc_ctx *ctx = shash_desc_ctx(desc);
 181         unsigned int partial = ctx->count % GHASH_BLOCK_SIZE;
 182
 183         if (partial) {
 184                 struct ghash_key *key = crypto_shash_ctx(desc->tfm);
 185
 186                 memset(ctx->buf + partial, 0, GHASH_BLOCK_SIZE - partial);
 187
 188                 ghash_do_update(1, ctx->digest, ctx->buf, key, NULL,
 189                                 pmull_ghash_update_p8);
 190         }
 191         put_unaligned_be64(ctx->digest[1], dst);
 192         put_unaligned_be64(ctx->digest[0], dst + 8);
 193
 194         *ctx = (struct ghash_desc_ctx){};
 195         return 0;
 196 }
 197
 198 static int ghash_final_p64(struct shash_desc *desc, u8 *dst)
 199 {
 200         struct ghash_desc_ctx *ctx = shash_desc_ctx(desc);
 201         unsigned int partial = ctx->count % GHASH_BLOCK_SIZE;
 202
 203         if (partial) {
 204                 struct ghash_key *key = crypto_shash_ctx(desc->tfm);
 205
 206                 memset(ctx->buf + partial, 0, GHASH_BLOCK_SIZE - partial);
 207
 208                 ghash_do_update(1, ctx->digest, ctx->buf, key, NULL,
 209                                 pmull_ghash_update_p64);
 210         }
 211         put_unaligned_be64(ctx->digest[1], dst);
 212         put_unaligned_be64(ctx->digest[0], dst + 8);
 213
 214         *ctx = (struct ghash_desc_ctx){};
 215         return 0;
 216 }
 217
 218 static void ghash_reflect(u64 h[], const be128 *k)
 219 {
 220         u64 carry = be64_to_cpu(k->a) & BIT(63) ? 1 : 0;
 221
 222         h[0] = (be64_to_cpu(k->b) << 1) | carry;
 223         h[1] = (be64_to_cpu(k->a) << 1) | (be64_to_cpu(k->b) >> 63);
 224
 225         if (carry)
 226                 h[1] ^= 0xc200000000000000UL;
 227 }
 228
 229 static int __ghash_setkey(struct ghash_key *key,
 230                           const u8 *inkey, unsigned int keylen)
 231 {
 232         be128 h;
 233
 234         /* needed for the fallback */
 235         memcpy(&key->k, inkey, GHASH_BLOCK_SIZE);
 236
 237         ghash_reflect(key->h, &key->k);
 238
 239         h = key->k;
 240         gf128mul_lle(&h, &key->k);
 241         ghash_reflect(key->h2, &h);
 242
 243         gf128mul_lle(&h, &key->k);
 244         ghash_reflect(key->h3, &h);
 245
 246         gf128mul_lle(&h, &key->k);
 247         ghash_reflect(key->h4, &h);
 248
 249         return 0;
 250 }
 251
 252 static int ghash_setkey(struct crypto_shash *tfm,
 253                         const u8 *inkey, unsigned int keylen)
 254 {
 255         struct ghash_key *key = crypto_shash_ctx(tfm);
 256
 257         if (keylen != GHASH_BLOCK_SIZE) {
 258                 crypto_shash_set_flags(tfm, CRYPTO_TFM_RES_BAD_KEY_LEN);
 259                 return -EINVAL;
 260         }
 261
 262         return __ghash_setkey(key, inkey, keylen);
 263 }
 264
 265 static struct shash_alg ghash_alg[] = {{
 266         .base.cra_name          = "ghash",
 267         .base.cra_driver_name   = "ghash-neon",
 268         .base.cra_priority      = 100,
 269         .base.cra_blocksize     = GHASH_BLOCK_SIZE,
 270         .base.cra_ctxsize       = sizeof(struct ghash_key),
 271         .base.cra_module        = THIS_MODULE,
 272
 273         .digestsize             = GHASH_DIGEST_SIZE,
 274         .init                   = ghash_init,
 275         .update                 = ghash_update_p8,
 276         .final                  = ghash_final_p8,
 277         .setkey                 = ghash_setkey,
 278         .descsize               = sizeof(struct ghash_desc_ctx),
 279 }, {
 280         .base.cra_name          = "ghash",
 281         .base.cra_driver_name   = "ghash-ce",
 282         .base.cra_priority      = 200,
 283         .base.cra_blocksize     = GHASH_BLOCK_SIZE,
 284         .base.cra_ctxsize       = sizeof(struct ghash_key),
 285         .base.cra_module        = THIS_MODULE,
 286
 287         .digestsize             = GHASH_DIGEST_SIZE,
 288         .init                   = ghash_init,
 289         .update                 = ghash_update_p64,
 290         .final                  = ghash_final_p64,
 291         .setkey                 = ghash_setkey,
 292         .descsize               = sizeof(struct ghash_desc_ctx),
 293 }};
 294
 295 static int num_rounds(struct crypto_aes_ctx *ctx)
 296 {
 297         /*
 298          * # of rounds specified by AES:
 299          * 128 bit key          10 rounds
 300          * 192 bit key          12 rounds
 301          * 256 bit key          14 rounds
 302          * => n byte key        => 6 + (n/4) rounds
 303          */
 304         return 6 + ctx->key_length / 4;
 305 }
 306
 307 static int gcm_setkey(struct crypto_aead *tfm, const u8 *inkey,
 308                       unsigned int keylen)
 309 {
 310         struct gcm_aes_ctx *ctx = crypto_aead_ctx(tfm);
 311         u8 key[GHASH_BLOCK_SIZE];
 312         int ret;
 313
 314         ret = crypto_aes_expand_key(&ctx->aes_key, inkey, keylen);
 315         if (ret) {
 316                 tfm->base.crt_flags |= CRYPTO_TFM_RES_BAD_KEY_LEN;
 317                 return -EINVAL;
 318         }
 319
 320         __aes_arm64_encrypt(ctx->aes_key.key_enc, key, (u8[AES_BLOCK_SIZE]){},
 321                             num_rounds(&ctx->aes_key));
 322
 323         return __ghash_setkey(&ctx->ghash_key, key, sizeof(be128));
 324 }
 325
 326 static int gcm_setauthsize(struct crypto_aead *tfm, unsigned int authsize)
 327 {
 328         switch (authsize) {
 329         case 4:
 330         case 8:
 331         case 12 ... 16:
 332                 break;
 333         default:
 334                 return -EINVAL;
 335         }
 336         return 0;
 337 }
 338
 339 static void gcm_update_mac(u64 dg[], const u8 *src, int count, u8 buf[],
 340                            int *buf_count, struct gcm_aes_ctx *ctx)
 341 {
 342         if (*buf_count > 0) {
 343                 int buf_added = min(count, GHASH_BLOCK_SIZE - *buf_count);
 344
 345                 memcpy(&buf[*buf_count], src, buf_added);
 346
 347                 *buf_count += buf_added;
 348                 src += buf_added;
 349                 count -= buf_added;
 350         }
 351
 352         if (count >= GHASH_BLOCK_SIZE || *buf_count == GHASH_BLOCK_SIZE) {
 353                 int blocks = count / GHASH_BLOCK_SIZE;
 354
 355                 ghash_do_update(blocks, dg, src, &ctx->ghash_key,
 356                                 *buf_count ? buf : NULL,
 357                                 pmull_ghash_update_p64);
 358
 359                 src += blocks * GHASH_BLOCK_SIZE;
 360                 count %= GHASH_BLOCK_SIZE;
 361                 *buf_count = 0;
 362         }
 363
 364         if (count > 0) {
 365                 memcpy(buf, src, count);
 366                 *buf_count = count;
 367         }
 368 }
 369
 370 static void gcm_calculate_auth_mac(struct aead_request *req, u64 dg[])
 371 {
 372         struct crypto_aead *aead = crypto_aead_reqtfm(req);
 373         struct gcm_aes_ctx *ctx = crypto_aead_ctx(aead);
 374         u8 buf[GHASH_BLOCK_SIZE];
 375         struct scatter_walk walk;
 376         u32 len = req->assoclen;
 377         int buf_count = 0;
 378
 379         scatterwalk_start(&walk, req->src);
 380
 381         do {
 382                 u32 n = scatterwalk_clamp(&walk, len);
 383                 u8 *p;
 384
 385                 if (!n) {
 386                         scatterwalk_start(&walk, sg_next(walk.sg));
 387                         n = scatterwalk_clamp(&walk, len);
 388                 }
 389                 p = scatterwalk_map(&walk);
 390
 391                 gcm_update_mac(dg, p, n, buf, &buf_count, ctx);
 392                 len -= n;
 393
 394                 scatterwalk_unmap(p);
 395                 scatterwalk_advance(&walk, n);
 396                 scatterwalk_done(&walk, 0, len);
 397         } while (len);
 398
 399         if (buf_count) {
 400                 memset(&buf[buf_count], 0, GHASH_BLOCK_SIZE - buf_count);
 401                 ghash_do_update(1, dg, buf, &ctx->ghash_key, NULL,
 402                                 pmull_ghash_update_p64);
 403         }
 404 }
 405
 406 static void gcm_final(struct aead_request *req, struct gcm_aes_ctx *ctx,
 407                       u64 dg[], u8 tag[], int cryptlen)
 408 {
 409         u8 mac[AES_BLOCK_SIZE];
 410         u128 lengths;
 411
 412         lengths.a = cpu_to_be64(req->assoclen * 8);
 413         lengths.b = cpu_to_be64(cryptlen * 8);
 414
 415         ghash_do_update(1, dg, (void *)&lengths, &ctx->ghash_key, NULL,
 416                         pmull_ghash_update_p64);
 417
 418         put_unaligned_be64(dg[1], mac);
 419         put_unaligned_be64(dg[0], mac + 8);
 420
 421         crypto_xor(tag, mac, AES_BLOCK_SIZE);
 422 }
 423
 424 static int gcm_encrypt(struct aead_request *req)
 425 {
 426         struct crypto_aead *aead = crypto_aead_reqtfm(req);
 427         struct gcm_aes_ctx *ctx = crypto_aead_ctx(aead);
 428         struct skcipher_walk walk;
 429         u8 iv[AES_BLOCK_SIZE];
 430         u8 ks[2 * AES_BLOCK_SIZE];
 431         u8 tag[AES_BLOCK_SIZE];
 432         u64 dg[2] = {};
 433         int nrounds = num_rounds(&ctx->aes_key);
 434         int err;
 435
 436         if (req->assoclen)
 437                 gcm_calculate_auth_mac(req, dg);
 438
 439         memcpy(iv, req->iv, GCM_IV_SIZE);
 440         put_unaligned_be32(1, iv + GCM_IV_SIZE);
 441
 442         err = skcipher_walk_aead_encrypt(&walk, req, false);
 443
 444         if (likely(may_use_simd() && walk.total >= 2 * AES_BLOCK_SIZE)) {
 445                 u32 const *rk = NULL;
 446
 447                 kernel_neon_begin();
 448                 pmull_gcm_encrypt_block(tag, iv, ctx->aes_key.key_enc, nrounds);
 449                 put_unaligned_be32(2, iv + GCM_IV_SIZE);
 450                 pmull_gcm_encrypt_block(ks, iv, NULL, nrounds);
 451                 put_unaligned_be32(3, iv + GCM_IV_SIZE);
 452                 pmull_gcm_encrypt_block(ks + AES_BLOCK_SIZE, iv, NULL, nrounds);
 453                 put_unaligned_be32(4, iv + GCM_IV_SIZE);
 454
 455                 do {
 456                         int blocks = walk.nbytes / (2 * AES_BLOCK_SIZE) * 2;
 457
 458                         if (rk)
 459                                 kernel_neon_begin();
 460
 461                         pmull_gcm_encrypt(blocks, dg, walk.dst.virt.addr,
 462                                           walk.src.virt.addr, &ctx->ghash_key,
 463                                           iv, rk, nrounds, ks);
 464                         kernel_neon_end();
 465
 466                         err = skcipher_walk_done(&walk,
 467                                         walk.nbytes % (2 * AES_BLOCK_SIZE));
 468
 469                         rk = ctx->aes_key.key_enc;
 470                 } while (walk.nbytes >= 2 * AES_BLOCK_SIZE);
 471         } else {
 472                 __aes_arm64_encrypt(ctx->aes_key.key_enc, tag, iv, nrounds);
 473                 put_unaligned_be32(2, iv + GCM_IV_SIZE);
 474
 475                 while (walk.nbytes >= (2 * AES_BLOCK_SIZE)) {
 476                         int blocks = walk.nbytes / AES_BLOCK_SIZE;
 477                         u8 *dst = walk.dst.virt.addr;
 478                         u8 *src = walk.src.virt.addr;
 479
 480                         do {
 481                                 __aes_arm64_encrypt(ctx->aes_key.key_enc,
 482                                                     ks, iv, nrounds);
 483                                 crypto_xor_cpy(dst, src, ks, AES_BLOCK_SIZE);
 484                                 crypto_inc(iv, AES_BLOCK_SIZE);
 485
 486                                 dst += AES_BLOCK_SIZE;
 487                                 src += AES_BLOCK_SIZE;
 488                         } while (--blocks > 0);
 489
 490                         ghash_do_update(walk.nbytes / AES_BLOCK_SIZE, dg,
 491                                         walk.dst.virt.addr, &ctx->ghash_key,
 492                                         NULL, pmull_ghash_update_p64);
 493
 494                         err = skcipher_walk_done(&walk,
 495                                                  walk.nbytes % (2 * AES_BLOCK_SIZE));
 496                 }
 497                 if (walk.nbytes) {
 498                         __aes_arm64_encrypt(ctx->aes_key.key_enc, ks, iv,
 499                                             nrounds);
 500                         if (walk.nbytes > AES_BLOCK_SIZE) {
 501                                 crypto_inc(iv, AES_BLOCK_SIZE);
 502                                 __aes_arm64_encrypt(ctx->aes_key.key_enc,
 503                                                     ks + AES_BLOCK_SIZE, iv,
 504                                                     nrounds);
 505                         }
 506                 }
 507         }
 508
 509         /* handle the tail */
 510         if (walk.nbytes) {
 511                 u8 buf[GHASH_BLOCK_SIZE];
 512                 unsigned int nbytes = walk.nbytes;
 513                 u8 *dst = walk.dst.virt.addr;
 514                 u8 *head = NULL;
 515
 516                 crypto_xor_cpy(walk.dst.virt.addr, walk.src.virt.addr, ks,
 517                                walk.nbytes);
 518
 519                 if (walk.nbytes > GHASH_BLOCK_SIZE) {
 520                         head = dst;
 521                         dst += GHASH_BLOCK_SIZE;
 522                         nbytes %= GHASH_BLOCK_SIZE;
 523                 }
 524
 525                 memcpy(buf, dst, nbytes);
 526                 memset(buf + nbytes, 0, GHASH_BLOCK_SIZE - nbytes);
 527                 ghash_do_update(!!nbytes, dg, buf, &ctx->ghash_key, head,
 528                                 pmull_ghash_update_p64);
 529
 530                 err = skcipher_walk_done(&walk, 0);
 531         }
 532
 533         if (err)
 534                 return err;
 535
 536         gcm_final(req, ctx, dg, tag, req->cryptlen);
 537
 538         /* copy authtag to end of dst */
 539         scatterwalk_map_and_copy(tag, req->dst, req->assoclen + req->cryptlen,
 540                                  crypto_aead_authsize(aead), 1);
 541
 542         return 0;
 543 }
 544
 545 static int gcm_decrypt(struct aead_request *req)
 546 {
 547         struct crypto_aead *aead = crypto_aead_reqtfm(req);
 548         struct gcm_aes_ctx *ctx = crypto_aead_ctx(aead);
 549         unsigned int authsize = crypto_aead_authsize(aead);
 550         struct skcipher_walk walk;
 551         u8 iv[2 * AES_BLOCK_SIZE];
 552         u8 tag[AES_BLOCK_SIZE];
 553         u8 buf[2 * GHASH_BLOCK_SIZE];
 554         u64 dg[2] = {};
 555         int nrounds = num_rounds(&ctx->aes_key);
 556         int err;
 557
 558         if (req->assoclen)
 559                 gcm_calculate_auth_mac(req, dg);
 560
 561         memcpy(iv, req->iv, GCM_IV_SIZE);
 562         put_unaligned_be32(1, iv + GCM_IV_SIZE);
 563
 564         err = skcipher_walk_aead_decrypt(&walk, req, false);
 565
 566         if (likely(may_use_simd() && walk.total >= 2 * AES_BLOCK_SIZE)) {
 567                 u32 const *rk = NULL;
 568
 569                 kernel_neon_begin();
 570                 pmull_gcm_encrypt_block(tag, iv, ctx->aes_key.key_enc, nrounds);
 571                 put_unaligned_be32(2, iv + GCM_IV_SIZE);
 572
 573                 do {
 574                         int blocks = walk.nbytes / (2 * AES_BLOCK_SIZE) * 2;
 575                         int rem = walk.total - blocks * AES_BLOCK_SIZE;
 576
 577                         if (rk)
 578                                 kernel_neon_begin();
 579
 580                         pmull_gcm_decrypt(blocks, dg, walk.dst.virt.addr,
 581                                           walk.src.virt.addr, &ctx->ghash_key,
 582                                           iv, rk, nrounds);
 583
 584                         /* check if this is the final iteration of the loop */
 585                         if (rem < (2 * AES_BLOCK_SIZE)) {
 586                                 u8 *iv2 = iv + AES_BLOCK_SIZE;
 587
 588                                 if (rem > AES_BLOCK_SIZE) {
 589                                         memcpy(iv2, iv, AES_BLOCK_SIZE);
 590                                         crypto_inc(iv2, AES_BLOCK_SIZE);
 591                                 }
 592
 593                                 pmull_gcm_encrypt_block(iv, iv, NULL, nrounds);
 594
 595                                 if (rem > AES_BLOCK_SIZE)
 596                                         pmull_gcm_encrypt_block(iv2, iv2, NULL,
 597                                                                 nrounds);
 598                         }
 599
 600                         kernel_neon_end();
 601
 602                         err = skcipher_walk_done(&walk,
 603                                         walk.nbytes % (2 * AES_BLOCK_SIZE));
 604
 605                         rk = ctx->aes_key.key_enc;
 606                 } while (walk.nbytes >= 2 * AES_BLOCK_SIZE);
 607         } else {
 608                 __aes_arm64_encrypt(ctx->aes_key.key_enc, tag, iv, nrounds);
 609                 put_unaligned_be32(2, iv + GCM_IV_SIZE);
 610
 611                 while (walk.nbytes >= (2 * AES_BLOCK_SIZE)) {
 612                         int blocks = walk.nbytes / AES_BLOCK_SIZE;
 613                         u8 *dst = walk.dst.virt.addr;
 614                         u8 *src = walk.src.virt.addr;
 615
 616                         ghash_do_update(blocks, dg, walk.src.virt.addr,
 617                                         &ctx->ghash_key, NULL,
 618                                         pmull_ghash_update_p64);
 619
 620                         do {
 621                                 __aes_arm64_encrypt(ctx->aes_key.key_enc,
 622                                                     buf, iv, nrounds);
 623                                 crypto_xor_cpy(dst, src, buf, AES_BLOCK_SIZE);
 624                                 crypto_inc(iv, AES_BLOCK_SIZE);
 625
 626                                 dst += AES_BLOCK_SIZE;
 627                                 src += AES_BLOCK_SIZE;
 628                         } while (--blocks > 0);
 629
 630                         err = skcipher_walk_done(&walk,
 631                                                  walk.nbytes % (2 * AES_BLOCK_SIZE));
 632                 }
 633                 if (walk.nbytes) {
 634                         if (walk.nbytes > AES_BLOCK_SIZE) {
 635                                 u8 *iv2 = iv + AES_BLOCK_SIZE;
 636
 637                                 memcpy(iv2, iv, AES_BLOCK_SIZE);
 638                                 crypto_inc(iv2, AES_BLOCK_SIZE);
 639
 640                                 __aes_arm64_encrypt(ctx->aes_key.key_enc, iv2,
 641                                                     iv2, nrounds);
 642                         }
 643                         __aes_arm64_encrypt(ctx->aes_key.key_enc, iv, iv,
 644                                             nrounds);
 645                 }
 646         }
 647
 648         /* handle the tail */
 649         if (walk.nbytes) {
 650                 const u8 *src = walk.src.virt.addr;
 651                 const u8 *head = NULL;
 652                 unsigned int nbytes = walk.nbytes;
 653
 654                 if (walk.nbytes > GHASH_BLOCK_SIZE) {
 655                         head = src;
 656                         src += GHASH_BLOCK_SIZE;
 657                         nbytes %= GHASH_BLOCK_SIZE;
 658                 }
 659
 660                 memcpy(buf, src, nbytes);
 661                 memset(buf + nbytes, 0, GHASH_BLOCK_SIZE - nbytes);
 662                 ghash_do_update(!!nbytes, dg, buf, &ctx->ghash_key, head,
 663                                 pmull_ghash_update_p64);
 664
 665                 crypto_xor_cpy(walk.dst.virt.addr, walk.src.virt.addr, iv,
 666                                walk.nbytes);
 667
 668                 err = skcipher_walk_done(&walk, 0);
 669         }
 670
 671         if (err)
 672                 return err;
 673
 674         gcm_final(req, ctx, dg, tag, req->cryptlen - authsize);
 675
 676         /* compare calculated auth tag with the stored one */
 677         scatterwalk_map_and_copy(buf, req->src,
 678                                  req->assoclen + req->cryptlen - authsize,
 679                                  authsize, 0);
 680
 681         if (crypto_memneq(tag, buf, authsize))
 682                 return -EBADMSG;
 683         return 0;
 684 }
 685
 686 static struct aead_alg gcm_aes_alg = {
 687         .ivsize                 = GCM_IV_SIZE,
 688         .chunksize              = 2 * AES_BLOCK_SIZE,
 689         .maxauthsize            = AES_BLOCK_SIZE,
 690         .setkey                 = gcm_setkey,
 691         .setauthsize            = gcm_setauthsize,
 692         .encrypt                = gcm_encrypt,
 693         .decrypt                = gcm_decrypt,
 694
 695         .base.cra_name          = "gcm(aes)",
 696         .base.cra_driver_name   = "gcm-aes-ce",
 697         .base.cra_priority      = 300,
 698         .base.cra_blocksize     = 1,
 699         .base.cra_ctxsize       = sizeof(struct gcm_aes_ctx),
 700         .base.cra_module        = THIS_MODULE,
 701 };
 702
 703 static int __init ghash_ce_mod_init(void)
 704 {
 705         int ret;
 706
 707         if (!(elf_hwcap & HWCAP_ASIMD))
 708                 return -ENODEV;
 709
 710         if (elf_hwcap & HWCAP_PMULL)
 711                 ret = crypto_register_shashes(ghash_alg,
 712                                               ARRAY_SIZE(ghash_alg));
 713         else
 714                 /* only register the first array element */
 715                 ret = crypto_register_shash(ghash_alg);
 716
 717         if (ret)
 718                 return ret;
 719
 720         if (elf_hwcap & HWCAP_PMULL) {
 721                 ret = crypto_register_aead(&gcm_aes_alg);
 722                 if (ret)
 723                         crypto_unregister_shashes(ghash_alg,
 724                                                   ARRAY_SIZE(ghash_alg));
 725         }
 726         return ret;
 727 }
 728
 729 static void __exit ghash_ce_mod_exit(void)
 730 {
 731         if (elf_hwcap & HWCAP_PMULL)
 732                 crypto_unregister_shashes(ghash_alg, ARRAY_SIZE(ghash_alg));
 733         else
 734                 crypto_unregister_shash(ghash_alg);
 735         crypto_unregister_aead(&gcm_aes_alg);
 736 }
 737
 738 static const struct cpu_feature ghash_cpu_feature[] = {
 739         { cpu_feature(PMULL) }, { }
 740 };
 741 MODULE_DEVICE_TABLE(cpu, ghash_cpu_feature);
 742
 743 module_init(ghash_ce_mod_init);
 744 module_exit(ghash_ce_mod_exit);