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/simd.h>
  21 #include <crypto/internal/skcipher.h>
  22 #include <crypto/scatterwalk.h>
  23 #include <linux/cpufeature.h>
  24 #include <linux/crypto.h>
  25 #include <linux/module.h>
  26
  27 MODULE_DESCRIPTION("GHASH and AES-GCM using ARMv8 Crypto Extensions");
  28 MODULE_AUTHOR("Ard Biesheuvel <ard.biesheuvel@linaro.org>");
  29 MODULE_LICENSE("GPL v2");
  30 MODULE_ALIAS_CRYPTO("ghash");
  31
  32 #define GHASH_BLOCK_SIZE        16
  33 #define GHASH_DIGEST_SIZE       16
  34 #define GCM_IV_SIZE             12
  35
  36 struct ghash_key {
  37         u64                     h[2];
  38         u64                     h2[2];
  39         u64                     h3[2];
  40         u64                     h4[2];
  41
  42         be128                   k;
  43 };
  44
  45 struct ghash_desc_ctx {
  46         u64 digest[GHASH_DIGEST_SIZE/sizeof(u64)];
  47         u8 buf[GHASH_BLOCK_SIZE];
  48         u32 count;
  49 };
  50
  51 struct gcm_aes_ctx {
  52         struct crypto_aes_ctx   aes_key;
  53         struct ghash_key        ghash_key;
  54 };
  55
  56 asmlinkage void pmull_ghash_update_p64(int blocks, u64 dg[], const char *src,
  57                                        struct ghash_key const *k,
  58                                        const char *head);
  59
  60 asmlinkage void pmull_ghash_update_p8(int blocks, u64 dg[], const char *src,
  61                                       struct ghash_key const *k,
  62                                       const char *head);
  63
  64 asmlinkage void pmull_gcm_encrypt(int blocks, u64 dg[], u8 dst[],
  65                                   const u8 src[], struct ghash_key const *k,
  66                                   u8 ctr[], u32 const rk[], int rounds,
  67                                   u8 ks[]);
  68
  69 asmlinkage void pmull_gcm_decrypt(int blocks, u64 dg[], u8 dst[],
  70                                   const u8 src[], struct ghash_key const *k,
  71                                   u8 ctr[], u32 const rk[], int rounds);
  72
  73 asmlinkage void pmull_gcm_encrypt_block(u8 dst[], u8 const src[],
  74                                         u32 const rk[], int rounds);
  75
  76 asmlinkage void __aes_arm64_encrypt(u32 *rk, u8 *out, const u8 *in, int rounds);
  77
  78 static int ghash_init(struct shash_desc *desc)
  79 {
  80         struct ghash_desc_ctx *ctx = shash_desc_ctx(desc);
  81
  82         *ctx = (struct ghash_desc_ctx){};
  83         return 0;
  84 }
  85
  86 static void ghash_do_update(int blocks, u64 dg[], const char *src,
  87                             struct ghash_key *key, const char *head,
  88                             void (*simd_update)(int blocks, u64 dg[],
  89                                                 const char *src,
  90                                                 struct ghash_key const *k,
  91                                                 const char *head))
  92 {
  93         if (likely(crypto_simd_usable())) {
  94                 kernel_neon_begin();
  95                 simd_update(blocks, dg, src, key, head);
  96                 kernel_neon_end();
  97         } else {
  98                 be128 dst = { cpu_to_be64(dg[1]), cpu_to_be64(dg[0]) };
  99
 100                 do {
 101                         const u8 *in = src;
 102
 103                         if (head) {
 104                                 in = head;
 105                                 blocks++;
 106                                 head = NULL;
 107                         } else {
 108                                 src += GHASH_BLOCK_SIZE;
 109                         }
 110
 111                         crypto_xor((u8 *)&dst, in, GHASH_BLOCK_SIZE);
 112                         gf128mul_lle(&dst, &key->k);
 113                 } while (--blocks);
 114
 115                 dg[0] = be64_to_cpu(dst.b);
 116                 dg[1] = be64_to_cpu(dst.a);
 117         }
 118 }
 119
 120 /* avoid hogging the CPU for too long */
 121 #define MAX_BLOCKS      (SZ_64K / GHASH_BLOCK_SIZE)
 122
 123 static int __ghash_update(struct shash_desc *desc, const u8 *src,
 124                           unsigned int len,
 125                           void (*simd_update)(int blocks, u64 dg[],
 126                                               const char *src,
 127                                               struct ghash_key const *k,
 128                                               const char *head))
 129 {
 130         struct ghash_desc_ctx *ctx = shash_desc_ctx(desc);
 131         unsigned int partial = ctx->count % GHASH_BLOCK_SIZE;
 132
 133         ctx->count += len;
 134
 135         if ((partial + len) >= GHASH_BLOCK_SIZE) {
 136                 struct ghash_key *key = crypto_shash_ctx(desc->tfm);
 137                 int blocks;
 138
 139                 if (partial) {
 140                         int p = GHASH_BLOCK_SIZE - partial;
 141
 142                         memcpy(ctx->buf + partial, src, p);
 143                         src += p;
 144                         len -= p;
 145                 }
 146
 147                 blocks = len / GHASH_BLOCK_SIZE;
 148                 len %= GHASH_BLOCK_SIZE;
 149
 150                 do {
 151                         int chunk = min(blocks, MAX_BLOCKS);
 152
 153                         ghash_do_update(chunk, ctx->digest, src, key,
 154                                         partial ? ctx->buf : NULL,
 155                                         simd_update);
 156
 157                         blocks -= chunk;
 158                         src += chunk * GHASH_BLOCK_SIZE;
 159                         partial = 0;
 160                 } while (unlikely(blocks > 0));
 161         }
 162         if (len)
 163                 memcpy(ctx->buf + partial, src, len);
 164         return 0;
 165 }
 166
 167 static int ghash_update_p8(struct shash_desc *desc, const u8 *src,
 168                            unsigned int len)
 169 {
 170         return __ghash_update(desc, src, len, pmull_ghash_update_p8);
 171 }
 172
 173 static int ghash_update_p64(struct shash_desc *desc, const u8 *src,
 174                             unsigned int len)
 175 {
 176         return __ghash_update(desc, src, len, pmull_ghash_update_p64);
 177 }
 178
 179 static int ghash_final_p8(struct shash_desc *desc, u8 *dst)
 180 {
 181         struct ghash_desc_ctx *ctx = shash_desc_ctx(desc);
 182         unsigned int partial = ctx->count % GHASH_BLOCK_SIZE;
 183
 184         if (partial) {
 185                 struct ghash_key *key = crypto_shash_ctx(desc->tfm);
 186
 187                 memset(ctx->buf + partial, 0, GHASH_BLOCK_SIZE - partial);
 188
 189                 ghash_do_update(1, ctx->digest, ctx->buf, key, NULL,
 190                                 pmull_ghash_update_p8);
 191         }
 192         put_unaligned_be64(ctx->digest[1], dst);
 193         put_unaligned_be64(ctx->digest[0], dst + 8);
 194
 195         *ctx = (struct ghash_desc_ctx){};
 196         return 0;
 197 }
 198
 199 static int ghash_final_p64(struct shash_desc *desc, u8 *dst)
 200 {
 201         struct ghash_desc_ctx *ctx = shash_desc_ctx(desc);
 202         unsigned int partial = ctx->count % GHASH_BLOCK_SIZE;
 203
 204         if (partial) {
 205                 struct ghash_key *key = crypto_shash_ctx(desc->tfm);
 206
 207                 memset(ctx->buf + partial, 0, GHASH_BLOCK_SIZE - partial);
 208
 209                 ghash_do_update(1, ctx->digest, ctx->buf, key, NULL,
 210                                 pmull_ghash_update_p64);
 211         }
 212         put_unaligned_be64(ctx->digest[1], dst);
 213         put_unaligned_be64(ctx->digest[0], dst + 8);
 214
 215         *ctx = (struct ghash_desc_ctx){};
 216         return 0;
 217 }
 218
 219 static void ghash_reflect(u64 h[], const be128 *k)
 220 {
 221         u64 carry = be64_to_cpu(k->a) & BIT(63) ? 1 : 0;
 222
 223         h[0] = (be64_to_cpu(k->b) << 1) | carry;
 224         h[1] = (be64_to_cpu(k->a) << 1) | (be64_to_cpu(k->b) >> 63);
 225
 226         if (carry)
 227                 h[1] ^= 0xc200000000000000UL;
 228 }
 229
 230 static int __ghash_setkey(struct ghash_key *key,
 231                           const u8 *inkey, unsigned int keylen)
 232 {
 233         be128 h;
 234
 235         /* needed for the fallback */
 236         memcpy(&key->k, inkey, GHASH_BLOCK_SIZE);
 237
 238         ghash_reflect(key->h, &key->k);
 239
 240         h = key->k;
 241         gf128mul_lle(&h, &key->k);
 242         ghash_reflect(key->h2, &h);
 243
 244         gf128mul_lle(&h, &key->k);
 245         ghash_reflect(key->h3, &h);
 246
 247         gf128mul_lle(&h, &key->k);
 248         ghash_reflect(key->h4, &h);
 249
 250         return 0;
 251 }
 252
 253 static int ghash_setkey(struct crypto_shash *tfm,
 254                         const u8 *inkey, unsigned int keylen)
 255 {
 256         struct ghash_key *key = crypto_shash_ctx(tfm);
 257
 258         if (keylen != GHASH_BLOCK_SIZE) {
 259                 crypto_shash_set_flags(tfm, CRYPTO_TFM_RES_BAD_KEY_LEN);
 260                 return -EINVAL;
 261         }
 262
 263         return __ghash_setkey(key, inkey, keylen);
 264 }
 265
 266 static struct shash_alg ghash_alg[] = {{
 267         .base.cra_name          = "ghash",
 268         .base.cra_driver_name   = "ghash-neon",
 269         .base.cra_priority      = 100,
 270         .base.cra_blocksize     = GHASH_BLOCK_SIZE,
 271         .base.cra_ctxsize       = sizeof(struct ghash_key),
 272         .base.cra_module        = THIS_MODULE,
 273
 274         .digestsize             = GHASH_DIGEST_SIZE,
 275         .init                   = ghash_init,
 276         .update                 = ghash_update_p8,
 277         .final                  = ghash_final_p8,
 278         .setkey                 = ghash_setkey,
 279         .descsize               = sizeof(struct ghash_desc_ctx),
 280 }, {
 281         .base.cra_name          = "ghash",
 282         .base.cra_driver_name   = "ghash-ce",
 283         .base.cra_priority      = 200,
 284         .base.cra_blocksize     = GHASH_BLOCK_SIZE,
 285         .base.cra_ctxsize       = sizeof(struct ghash_key),
 286         .base.cra_module        = THIS_MODULE,
 287
 288         .digestsize             = GHASH_DIGEST_SIZE,
 289         .init                   = ghash_init,
 290         .update                 = ghash_update_p64,
 291         .final                  = ghash_final_p64,
 292         .setkey                 = ghash_setkey,
 293         .descsize               = sizeof(struct ghash_desc_ctx),
 294 }};
 295
 296 static int num_rounds(struct crypto_aes_ctx *ctx)
 297 {
 298         /*
 299          * # of rounds specified by AES:
 300          * 128 bit key          10 rounds
 301          * 192 bit key          12 rounds
 302          * 256 bit key          14 rounds
 303          * => n byte key        => 6 + (n/4) rounds
 304          */
 305         return 6 + ctx->key_length / 4;
 306 }
 307
 308 static int gcm_setkey(struct crypto_aead *tfm, const u8 *inkey,
 309                       unsigned int keylen)
 310 {
 311         struct gcm_aes_ctx *ctx = crypto_aead_ctx(tfm);
 312         u8 key[GHASH_BLOCK_SIZE];
 313         int ret;
 314
 315         ret = crypto_aes_expand_key(&ctx->aes_key, inkey, keylen);
 316         if (ret) {
 317                 tfm->base.crt_flags |= CRYPTO_TFM_RES_BAD_KEY_LEN;
 318                 return -EINVAL;
 319         }
 320
 321         __aes_arm64_encrypt(ctx->aes_key.key_enc, key, (u8[AES_BLOCK_SIZE]){},
 322                             num_rounds(&ctx->aes_key));
 323
 324         return __ghash_setkey(&ctx->ghash_key, key, sizeof(be128));
 325 }
 326
 327 static int gcm_setauthsize(struct crypto_aead *tfm, unsigned int authsize)
 328 {
 329         switch (authsize) {
 330         case 4:
 331         case 8:
 332         case 12 ... 16:
 333                 break;
 334         default:
 335                 return -EINVAL;
 336         }
 337         return 0;
 338 }
 339
 340 static void gcm_update_mac(u64 dg[], const u8 *src, int count, u8 buf[],
 341                            int *buf_count, struct gcm_aes_ctx *ctx)
 342 {
 343         if (*buf_count > 0) {
 344                 int buf_added = min(count, GHASH_BLOCK_SIZE - *buf_count);
 345
 346                 memcpy(&buf[*buf_count], src, buf_added);
 347
 348                 *buf_count += buf_added;
 349                 src += buf_added;
 350                 count -= buf_added;
 351         }
 352
 353         if (count >= GHASH_BLOCK_SIZE || *buf_count == GHASH_BLOCK_SIZE) {
 354                 int blocks = count / GHASH_BLOCK_SIZE;
 355
 356                 ghash_do_update(blocks, dg, src, &ctx->ghash_key,
 357                                 *buf_count ? buf : NULL,
 358                                 pmull_ghash_update_p64);
 359
 360                 src += blocks * GHASH_BLOCK_SIZE;
 361                 count %= GHASH_BLOCK_SIZE;
 362                 *buf_count = 0;
 363         }
 364
 365         if (count > 0) {
 366                 memcpy(buf, src, count);
 367                 *buf_count = count;
 368         }
 369 }
 370
 371 static void gcm_calculate_auth_mac(struct aead_request *req, u64 dg[])
 372 {
 373         struct crypto_aead *aead = crypto_aead_reqtfm(req);
 374         struct gcm_aes_ctx *ctx = crypto_aead_ctx(aead);
 375         u8 buf[GHASH_BLOCK_SIZE];
 376         struct scatter_walk walk;
 377         u32 len = req->assoclen;
 378         int buf_count = 0;
 379
 380         scatterwalk_start(&walk, req->src);
 381
 382         do {
 383                 u32 n = scatterwalk_clamp(&walk, len);
 384                 u8 *p;
 385
 386                 if (!n) {
 387                         scatterwalk_start(&walk, sg_next(walk.sg));
 388                         n = scatterwalk_clamp(&walk, len);
 389                 }
 390                 p = scatterwalk_map(&walk);
 391
 392                 gcm_update_mac(dg, p, n, buf, &buf_count, ctx);
 393                 len -= n;
 394
 395                 scatterwalk_unmap(p);
 396                 scatterwalk_advance(&walk, n);
 397                 scatterwalk_done(&walk, 0, len);
 398         } while (len);
 399
 400         if (buf_count) {
 401                 memset(&buf[buf_count], 0, GHASH_BLOCK_SIZE - buf_count);
 402                 ghash_do_update(1, dg, buf, &ctx->ghash_key, NULL,
 403                                 pmull_ghash_update_p64);
 404         }
 405 }
 406
 407 static void gcm_final(struct aead_request *req, struct gcm_aes_ctx *ctx,
 408                       u64 dg[], u8 tag[], int cryptlen)
 409 {
 410         u8 mac[AES_BLOCK_SIZE];
 411         u128 lengths;
 412
 413         lengths.a = cpu_to_be64(req->assoclen * 8);
 414         lengths.b = cpu_to_be64(cryptlen * 8);
 415
 416         ghash_do_update(1, dg, (void *)&lengths, &ctx->ghash_key, NULL,
 417                         pmull_ghash_update_p64);
 418
 419         put_unaligned_be64(dg[1], mac);
 420         put_unaligned_be64(dg[0], mac + 8);
 421
 422         crypto_xor(tag, mac, AES_BLOCK_SIZE);
 423 }
 424
 425 static int gcm_encrypt(struct aead_request *req)
 426 {
 427         struct crypto_aead *aead = crypto_aead_reqtfm(req);
 428         struct gcm_aes_ctx *ctx = crypto_aead_ctx(aead);
 429         struct skcipher_walk walk;
 430         u8 iv[AES_BLOCK_SIZE];
 431         u8 ks[2 * AES_BLOCK_SIZE];
 432         u8 tag[AES_BLOCK_SIZE];
 433         u64 dg[2] = {};
 434         int nrounds = num_rounds(&ctx->aes_key);
 435         int err;
 436
 437         if (req->assoclen)
 438                 gcm_calculate_auth_mac(req, dg);
 439
 440         memcpy(iv, req->iv, GCM_IV_SIZE);
 441         put_unaligned_be32(1, iv + GCM_IV_SIZE);
 442
 443         err = skcipher_walk_aead_encrypt(&walk, req, false);
 444
 445         if (likely(crypto_simd_usable() && walk.total >= 2 * AES_BLOCK_SIZE)) {
 446                 u32 const *rk = NULL;
 447
 448                 kernel_neon_begin();
 449                 pmull_gcm_encrypt_block(tag, iv, ctx->aes_key.key_enc, nrounds);
 450                 put_unaligned_be32(2, iv + GCM_IV_SIZE);
 451                 pmull_gcm_encrypt_block(ks, iv, NULL, nrounds);
 452                 put_unaligned_be32(3, iv + GCM_IV_SIZE);
 453                 pmull_gcm_encrypt_block(ks + AES_BLOCK_SIZE, iv, NULL, nrounds);
 454                 put_unaligned_be32(4, iv + GCM_IV_SIZE);
 455
 456                 do {
 457                         int blocks = walk.nbytes / (2 * AES_BLOCK_SIZE) * 2;
 458
 459                         if (rk)
 460                                 kernel_neon_begin();
 461
 462                         pmull_gcm_encrypt(blocks, dg, walk.dst.virt.addr,
 463                                           walk.src.virt.addr, &ctx->ghash_key,
 464                                           iv, rk, nrounds, ks);
 465                         kernel_neon_end();
 466
 467                         err = skcipher_walk_done(&walk,
 468                                         walk.nbytes % (2 * AES_BLOCK_SIZE));
 469
 470                         rk = ctx->aes_key.key_enc;
 471                 } while (walk.nbytes >= 2 * AES_BLOCK_SIZE);
 472         } else {
 473                 __aes_arm64_encrypt(ctx->aes_key.key_enc, tag, iv, nrounds);
 474                 put_unaligned_be32(2, iv + GCM_IV_SIZE);
 475
 476                 while (walk.nbytes >= (2 * AES_BLOCK_SIZE)) {
 477                         const int blocks =
 478                                 walk.nbytes / (2 * AES_BLOCK_SIZE) * 2;
 479                         u8 *dst = walk.dst.virt.addr;
 480                         u8 *src = walk.src.virt.addr;
 481                         int remaining = blocks;
 482
 483                         do {
 484                                 __aes_arm64_encrypt(ctx->aes_key.key_enc,
 485                                                     ks, iv, nrounds);
 486                                 crypto_xor_cpy(dst, src, ks, AES_BLOCK_SIZE);
 487                                 crypto_inc(iv, AES_BLOCK_SIZE);
 488
 489                                 dst += AES_BLOCK_SIZE;
 490                                 src += AES_BLOCK_SIZE;
 491                         } while (--remaining > 0);
 492
 493                         ghash_do_update(blocks, dg,
 494                                         walk.dst.virt.addr, &ctx->ghash_key,
 495                                         NULL, pmull_ghash_update_p64);
 496
 497                         err = skcipher_walk_done(&walk,
 498                                                  walk.nbytes % (2 * AES_BLOCK_SIZE));
 499                 }
 500                 if (walk.nbytes) {
 501                         __aes_arm64_encrypt(ctx->aes_key.key_enc, ks, iv,
 502                                             nrounds);
 503                         if (walk.nbytes > AES_BLOCK_SIZE) {
 504                                 crypto_inc(iv, AES_BLOCK_SIZE);
 505                                 __aes_arm64_encrypt(ctx->aes_key.key_enc,
 506                                                     ks + AES_BLOCK_SIZE, iv,
 507                                                     nrounds);
 508                         }
 509                 }
 510         }
 511
 512         /* handle the tail */
 513         if (walk.nbytes) {
 514                 u8 buf[GHASH_BLOCK_SIZE];
 515                 unsigned int nbytes = walk.nbytes;
 516                 u8 *dst = walk.dst.virt.addr;
 517                 u8 *head = NULL;
 518
 519                 crypto_xor_cpy(walk.dst.virt.addr, walk.src.virt.addr, ks,
 520                                walk.nbytes);
 521
 522                 if (walk.nbytes > GHASH_BLOCK_SIZE) {
 523                         head = dst;
 524                         dst += GHASH_BLOCK_SIZE;
 525                         nbytes %= GHASH_BLOCK_SIZE;
 526                 }
 527
 528                 memcpy(buf, dst, nbytes);
 529                 memset(buf + nbytes, 0, GHASH_BLOCK_SIZE - nbytes);
 530                 ghash_do_update(!!nbytes, dg, buf, &ctx->ghash_key, head,
 531                                 pmull_ghash_update_p64);
 532
 533                 err = skcipher_walk_done(&walk, 0);
 534         }
 535
 536         if (err)
 537                 return err;
 538
 539         gcm_final(req, ctx, dg, tag, req->cryptlen);
 540
 541         /* copy authtag to end of dst */
 542         scatterwalk_map_and_copy(tag, req->dst, req->assoclen + req->cryptlen,
 543                                  crypto_aead_authsize(aead), 1);
 544
 545         return 0;
 546 }
 547
 548 static int gcm_decrypt(struct aead_request *req)
 549 {
 550         struct crypto_aead *aead = crypto_aead_reqtfm(req);
 551         struct gcm_aes_ctx *ctx = crypto_aead_ctx(aead);
 552         unsigned int authsize = crypto_aead_authsize(aead);
 553         struct skcipher_walk walk;
 554         u8 iv[2 * AES_BLOCK_SIZE];
 555         u8 tag[AES_BLOCK_SIZE];
 556         u8 buf[2 * GHASH_BLOCK_SIZE];
 557         u64 dg[2] = {};
 558         int nrounds = num_rounds(&ctx->aes_key);
 559         int err;
 560
 561         if (req->assoclen)
 562                 gcm_calculate_auth_mac(req, dg);
 563
 564         memcpy(iv, req->iv, GCM_IV_SIZE);
 565         put_unaligned_be32(1, iv + GCM_IV_SIZE);
 566
 567         err = skcipher_walk_aead_decrypt(&walk, req, false);
 568
 569         if (likely(crypto_simd_usable() && walk.total >= 2 * AES_BLOCK_SIZE)) {
 570                 u32 const *rk = NULL;
 571
 572                 kernel_neon_begin();
 573                 pmull_gcm_encrypt_block(tag, iv, ctx->aes_key.key_enc, nrounds);
 574                 put_unaligned_be32(2, iv + GCM_IV_SIZE);
 575
 576                 do {
 577                         int blocks = walk.nbytes / (2 * AES_BLOCK_SIZE) * 2;
 578                         int rem = walk.total - blocks * AES_BLOCK_SIZE;
 579
 580                         if (rk)
 581                                 kernel_neon_begin();
 582
 583                         pmull_gcm_decrypt(blocks, dg, walk.dst.virt.addr,
 584                                           walk.src.virt.addr, &ctx->ghash_key,
 585                                           iv, rk, nrounds);
 586
 587                         /* check if this is the final iteration of the loop */
 588                         if (rem < (2 * AES_BLOCK_SIZE)) {
 589                                 u8 *iv2 = iv + AES_BLOCK_SIZE;
 590
 591                                 if (rem > AES_BLOCK_SIZE) {
 592                                         memcpy(iv2, iv, AES_BLOCK_SIZE);
 593                                         crypto_inc(iv2, AES_BLOCK_SIZE);
 594                                 }
 595
 596                                 pmull_gcm_encrypt_block(iv, iv, NULL, nrounds);
 597
 598                                 if (rem > AES_BLOCK_SIZE)
 599                                         pmull_gcm_encrypt_block(iv2, iv2, NULL,
 600                                                                 nrounds);
 601                         }
 602
 603                         kernel_neon_end();
 604
 605                         err = skcipher_walk_done(&walk,
 606                                         walk.nbytes % (2 * AES_BLOCK_SIZE));
 607
 608                         rk = ctx->aes_key.key_enc;
 609                 } while (walk.nbytes >= 2 * AES_BLOCK_SIZE);
 610         } else {
 611                 __aes_arm64_encrypt(ctx->aes_key.key_enc, tag, iv, nrounds);
 612                 put_unaligned_be32(2, iv + GCM_IV_SIZE);
 613
 614                 while (walk.nbytes >= (2 * AES_BLOCK_SIZE)) {
 615                         int blocks = walk.nbytes / (2 * AES_BLOCK_SIZE) * 2;
 616                         u8 *dst = walk.dst.virt.addr;
 617                         u8 *src = walk.src.virt.addr;
 618
 619                         ghash_do_update(blocks, dg, walk.src.virt.addr,
 620                                         &ctx->ghash_key, NULL,
 621                                         pmull_ghash_update_p64);
 622
 623                         do {
 624                                 __aes_arm64_encrypt(ctx->aes_key.key_enc,
 625                                                     buf, iv, nrounds);
 626                                 crypto_xor_cpy(dst, src, buf, AES_BLOCK_SIZE);
 627                                 crypto_inc(iv, AES_BLOCK_SIZE);
 628
 629                                 dst += AES_BLOCK_SIZE;
 630                                 src += AES_BLOCK_SIZE;
 631                         } while (--blocks > 0);
 632
 633                         err = skcipher_walk_done(&walk,
 634                                                  walk.nbytes % (2 * AES_BLOCK_SIZE));
 635                 }
 636                 if (walk.nbytes) {
 637                         if (walk.nbytes > AES_BLOCK_SIZE) {
 638                                 u8 *iv2 = iv + AES_BLOCK_SIZE;
 639
 640                                 memcpy(iv2, iv, AES_BLOCK_SIZE);
 641                                 crypto_inc(iv2, AES_BLOCK_SIZE);
 642
 643                                 __aes_arm64_encrypt(ctx->aes_key.key_enc, iv2,
 644                                                     iv2, nrounds);
 645                         }
 646                         __aes_arm64_encrypt(ctx->aes_key.key_enc, iv, iv,
 647                                             nrounds);
 648                 }
 649         }
 650
 651         /* handle the tail */
 652         if (walk.nbytes) {
 653                 const u8 *src = walk.src.virt.addr;
 654                 const u8 *head = NULL;
 655                 unsigned int nbytes = walk.nbytes;
 656
 657                 if (walk.nbytes > GHASH_BLOCK_SIZE) {
 658                         head = src;
 659                         src += GHASH_BLOCK_SIZE;
 660                         nbytes %= GHASH_BLOCK_SIZE;
 661                 }
 662
 663                 memcpy(buf, src, nbytes);
 664                 memset(buf + nbytes, 0, GHASH_BLOCK_SIZE - nbytes);
 665                 ghash_do_update(!!nbytes, dg, buf, &ctx->ghash_key, head,
 666                                 pmull_ghash_update_p64);
 667
 668                 crypto_xor_cpy(walk.dst.virt.addr, walk.src.virt.addr, iv,
 669                                walk.nbytes);
 670
 671                 err = skcipher_walk_done(&walk, 0);
 672         }
 673
 674         if (err)
 675                 return err;
 676
 677         gcm_final(req, ctx, dg, tag, req->cryptlen - authsize);
 678
 679         /* compare calculated auth tag with the stored one */
 680         scatterwalk_map_and_copy(buf, req->src,
 681                                  req->assoclen + req->cryptlen - authsize,
 682                                  authsize, 0);
 683
 684         if (crypto_memneq(tag, buf, authsize))
 685                 return -EBADMSG;
 686         return 0;
 687 }
 688
 689 static struct aead_alg gcm_aes_alg = {
 690         .ivsize                 = GCM_IV_SIZE,
 691         .chunksize              = 2 * AES_BLOCK_SIZE,
 692         .maxauthsize            = AES_BLOCK_SIZE,
 693         .setkey                 = gcm_setkey,
 694         .setauthsize            = gcm_setauthsize,
 695         .encrypt                = gcm_encrypt,
 696         .decrypt                = gcm_decrypt,
 697
 698         .base.cra_name          = "gcm(aes)",
 699         .base.cra_driver_name   = "gcm-aes-ce",
 700         .base.cra_priority      = 300,
 701         .base.cra_blocksize     = 1,
 702         .base.cra_ctxsize       = sizeof(struct gcm_aes_ctx),
 703         .base.cra_module        = THIS_MODULE,
 704 };
 705
 706 static int __init ghash_ce_mod_init(void)
 707 {
 708         int ret;
 709
 710         if (!cpu_have_named_feature(ASIMD))
 711                 return -ENODEV;
 712
 713         if (cpu_have_named_feature(PMULL))
 714                 ret = crypto_register_shashes(ghash_alg,
 715                                               ARRAY_SIZE(ghash_alg));
 716         else
 717                 /* only register the first array element */
 718                 ret = crypto_register_shash(ghash_alg);
 719
 720         if (ret)
 721                 return ret;
 722
 723         if (cpu_have_named_feature(PMULL)) {
 724                 ret = crypto_register_aead(&gcm_aes_alg);
 725                 if (ret)
 726                         crypto_unregister_shashes(ghash_alg,
 727                                                   ARRAY_SIZE(ghash_alg));
 728         }
 729         return ret;
 730 }
 731
 732 static void __exit ghash_ce_mod_exit(void)
 733 {
 734         if (cpu_have_named_feature(PMULL))
 735                 crypto_unregister_shashes(ghash_alg, ARRAY_SIZE(ghash_alg));
 736         else
 737                 crypto_unregister_shash(ghash_alg);
 738         crypto_unregister_aead(&gcm_aes_alg);
 739 }
 740
 741 static const struct cpu_feature ghash_cpu_feature[] = {
 742         { cpu_feature(PMULL) }, { }
 743 };
 744 MODULE_DEVICE_TABLE(cpu, ghash_cpu_feature);
 745
 746 module_init(ghash_ce_mod_init);
 747 module_exit(ghash_ce_mod_exit);