arch/arm64/crypto/aes-ce-cipher.c

   1 /*
   2  * aes-ce-cipher.c - core AES cipher using ARMv8 Crypto Extensions
   3  *
   4  * Copyright (C) 2013 - 2017 Linaro Ltd <ard.biesheuvel@linaro.org>
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 as
   8  * published 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 <linux/cpufeature.h>
  16 #include <linux/crypto.h>
  17 #include <linux/module.h>
  18
  19 #include "aes-ce-setkey.h"
  20
  21 MODULE_DESCRIPTION("Synchronous AES cipher using ARMv8 Crypto Extensions");
  22 MODULE_AUTHOR("Ard Biesheuvel <ard.biesheuvel@linaro.org>");
  23 MODULE_LICENSE("GPL v2");
  24
  25 asmlinkage void __aes_arm64_encrypt(u32 *rk, u8 *out, const u8 *in, int rounds);
  26 asmlinkage void __aes_arm64_decrypt(u32 *rk, u8 *out, const u8 *in, int rounds);
  27
  28 struct aes_block {
  29         u8 b[AES_BLOCK_SIZE];
  30 };
  31
  32 static int num_rounds(struct crypto_aes_ctx *ctx)
  33 {
  34         /*
  35          * # of rounds specified by AES:
  36          * 128 bit key          10 rounds
  37          * 192 bit key          12 rounds
  38          * 256 bit key          14 rounds
  39          * => n byte key        => 6 + (n/4) rounds
  40          */
  41         return 6 + ctx->key_length / 4;
  42 }
  43
  44 static void aes_cipher_encrypt(struct crypto_tfm *tfm, u8 dst[], u8 const src[])
  45 {
  46         struct crypto_aes_ctx *ctx = crypto_tfm_ctx(tfm);
  47         struct aes_block *out = (struct aes_block *)dst;
  48         struct aes_block const *in = (struct aes_block *)src;
  49         void *dummy0;
  50         int dummy1;
  51
  52         if (!may_use_simd()) {
  53                 __aes_arm64_encrypt(ctx->key_enc, dst, src, num_rounds(ctx));
  54                 return;
  55         }
  56
  57         kernel_neon_begin();
  58
  59         __asm__("       ld1     {v0.16b}, %[in]                 ;"
  60                 "       ld1     {v1.4s}, [%[key]], #16          ;"
  61                 "       cmp     %w[rounds], #10                 ;"
  62                 "       bmi     0f                              ;"
  63                 "       bne     3f                              ;"
  64                 "       mov     v3.16b, v1.16b                  ;"
  65                 "       b       2f                              ;"
  66                 "0:     mov     v2.16b, v1.16b                  ;"
  67                 "       ld1     {v3.4s}, [%[key]], #16          ;"
  68                 "1:     aese    v0.16b, v2.16b                  ;"
  69                 "       aesmc   v0.16b, v0.16b                  ;"
  70                 "2:     ld1     {v1.4s}, [%[key]], #16          ;"
  71                 "       aese    v0.16b, v3.16b                  ;"
  72                 "       aesmc   v0.16b, v0.16b                  ;"
  73                 "3:     ld1     {v2.4s}, [%[key]], #16          ;"
  74                 "       subs    %w[rounds], %w[rounds], #3      ;"
  75                 "       aese    v0.16b, v1.16b                  ;"
  76                 "       aesmc   v0.16b, v0.16b                  ;"
  77                 "       ld1     {v3.4s}, [%[key]], #16          ;"
  78                 "       bpl     1b                              ;"
  79                 "       aese    v0.16b, v2.16b                  ;"
  80                 "       eor     v0.16b, v0.16b, v3.16b          ;"
  81                 "       st1     {v0.16b}, %[out]                ;"
  82
  83         :       [out]           "=Q"(*out),
  84                 [key]           "=r"(dummy0),
  85                 [rounds]        "=r"(dummy1)
  86         :       [in]            "Q"(*in),
  87                                 "1"(ctx->key_enc),
  88                                 "2"(num_rounds(ctx) - 2)
  89         :       "cc");
  90
  91         kernel_neon_end();
  92 }
  93
  94 static void aes_cipher_decrypt(struct crypto_tfm *tfm, u8 dst[], u8 const src[])
  95 {
  96         struct crypto_aes_ctx *ctx = crypto_tfm_ctx(tfm);
  97         struct aes_block *out = (struct aes_block *)dst;
  98         struct aes_block const *in = (struct aes_block *)src;
  99         void *dummy0;
 100         int dummy1;
 101
 102         if (!may_use_simd()) {
 103                 __aes_arm64_decrypt(ctx->key_dec, dst, src, num_rounds(ctx));
 104                 return;
 105         }
 106
 107         kernel_neon_begin();
 108
 109         __asm__("       ld1     {v0.16b}, %[in]                 ;"
 110                 "       ld1     {v1.4s}, [%[key]], #16          ;"
 111                 "       cmp     %w[rounds], #10                 ;"
 112                 "       bmi     0f                              ;"
 113                 "       bne     3f                              ;"
 114                 "       mov     v3.16b, v1.16b                  ;"
 115                 "       b       2f                              ;"
 116                 "0:     mov     v2.16b, v1.16b                  ;"
 117                 "       ld1     {v3.4s}, [%[key]], #16          ;"
 118                 "1:     aesd    v0.16b, v2.16b                  ;"
 119                 "       aesimc  v0.16b, v0.16b                  ;"
 120                 "2:     ld1     {v1.4s}, [%[key]], #16          ;"
 121                 "       aesd    v0.16b, v3.16b                  ;"
 122                 "       aesimc  v0.16b, v0.16b                  ;"
 123                 "3:     ld1     {v2.4s}, [%[key]], #16          ;"
 124                 "       subs    %w[rounds], %w[rounds], #3      ;"
 125                 "       aesd    v0.16b, v1.16b                  ;"
 126                 "       aesimc  v0.16b, v0.16b                  ;"
 127                 "       ld1     {v3.4s}, [%[key]], #16          ;"
 128                 "       bpl     1b                              ;"
 129                 "       aesd    v0.16b, v2.16b                  ;"
 130                 "       eor     v0.16b, v0.16b, v3.16b          ;"
 131                 "       st1     {v0.16b}, %[out]                ;"
 132
 133         :       [out]           "=Q"(*out),
 134                 [key]           "=r"(dummy0),
 135                 [rounds]        "=r"(dummy1)
 136         :       [in]            "Q"(*in),
 137                                 "1"(ctx->key_dec),
 138                                 "2"(num_rounds(ctx) - 2)
 139         :       "cc");
 140
 141         kernel_neon_end();
 142 }
 143
 144 /*
 145  * aes_sub() - use the aese instruction to perform the AES sbox substitution
 146  *             on each byte in 'input'
 147  */
 148 static u32 aes_sub(u32 input)
 149 {
 150         u32 ret;
 151
 152         __asm__("dup    v1.4s, %w[in]           ;"
 153                 "movi   v0.16b, #0              ;"
 154                 "aese   v0.16b, v1.16b          ;"
 155                 "umov   %w[out], v0.4s[0]       ;"
 156
 157         :       [out]   "=r"(ret)
 158         :       [in]    "r"(input)
 159         :               "v0","v1");
 160
 161         return ret;
 162 }
 163
 164 int ce_aes_expandkey(struct crypto_aes_ctx *ctx, const u8 *in_key,
 165                      unsigned int key_len)
 166 {
 167         /*
 168          * The AES key schedule round constants
 169          */
 170         static u8 const rcon[] = {
 171                 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x1b, 0x36,
 172         };
 173
 174         u32 kwords = key_len / sizeof(u32);
 175         struct aes_block *key_enc, *key_dec;
 176         int i, j;
 177
 178         if (key_len != AES_KEYSIZE_128 &&
 179             key_len != AES_KEYSIZE_192 &&
 180             key_len != AES_KEYSIZE_256)
 181                 return -EINVAL;
 182
 183         ctx->key_length = key_len;
 184         for (i = 0; i < kwords; i++)
 185                 ctx->key_enc[i] = get_unaligned_le32(in_key + i * sizeof(u32));
 186
 187         kernel_neon_begin();
 188         for (i = 0; i < sizeof(rcon); i++) {
 189                 u32 *rki = ctx->key_enc + (i * kwords);
 190                 u32 *rko = rki + kwords;
 191
 192                 rko[0] = ror32(aes_sub(rki[kwords - 1]), 8) ^ rcon[i] ^ rki[0];
 193                 rko[1] = rko[0] ^ rki[1];
 194                 rko[2] = rko[1] ^ rki[2];
 195                 rko[3] = rko[2] ^ rki[3];
 196
 197                 if (key_len == AES_KEYSIZE_192) {
 198                         if (i >= 7)
 199                                 break;
 200                         rko[4] = rko[3] ^ rki[4];
 201                         rko[5] = rko[4] ^ rki[5];
 202                 } else if (key_len == AES_KEYSIZE_256) {
 203                         if (i >= 6)
 204                                 break;
 205                         rko[4] = aes_sub(rko[3]) ^ rki[4];
 206                         rko[5] = rko[4] ^ rki[5];
 207                         rko[6] = rko[5] ^ rki[6];
 208                         rko[7] = rko[6] ^ rki[7];
 209                 }
 210         }
 211
 212         /*
 213          * Generate the decryption keys for the Equivalent Inverse Cipher.
 214          * This involves reversing the order of the round keys, and applying
 215          * the Inverse Mix Columns transformation on all but the first and
 216          * the last one.
 217          */
 218         key_enc = (struct aes_block *)ctx->key_enc;
 219         key_dec = (struct aes_block *)ctx->key_dec;
 220         j = num_rounds(ctx);
 221
 222         key_dec[0] = key_enc[j];
 223         for (i = 1, j--; j > 0; i++, j--)
 224                 __asm__("ld1    {v0.4s}, %[in]          ;"
 225                         "aesimc v1.16b, v0.16b          ;"
 226                         "st1    {v1.4s}, %[out] ;"
 227
 228                 :       [out]   "=Q"(key_dec[i])
 229                 :       [in]    "Q"(key_enc[j])
 230                 :               "v0","v1");
 231         key_dec[i] = key_enc[0];
 232
 233         kernel_neon_end();
 234         return 0;
 235 }
 236 EXPORT_SYMBOL(ce_aes_expandkey);
 237
 238 int ce_aes_setkey(struct crypto_tfm *tfm, const u8 *in_key,
 239                   unsigned int key_len)
 240 {
 241         struct crypto_aes_ctx *ctx = crypto_tfm_ctx(tfm);
 242         int ret;
 243
 244         ret = ce_aes_expandkey(ctx, in_key, key_len);
 245         if (!ret)
 246                 return 0;
 247
 248         tfm->crt_flags |= CRYPTO_TFM_RES_BAD_KEY_LEN;
 249         return -EINVAL;
 250 }
 251 EXPORT_SYMBOL(ce_aes_setkey);
 252
 253 static struct crypto_alg aes_alg = {
 254         .cra_name               = "aes",
 255         .cra_driver_name        = "aes-ce",
 256         .cra_priority           = 250,
 257         .cra_flags              = CRYPTO_ALG_TYPE_CIPHER,
 258         .cra_blocksize          = AES_BLOCK_SIZE,
 259         .cra_ctxsize            = sizeof(struct crypto_aes_ctx),
 260         .cra_module             = THIS_MODULE,
 261         .cra_cipher = {
 262                 .cia_min_keysize        = AES_MIN_KEY_SIZE,
 263                 .cia_max_keysize        = AES_MAX_KEY_SIZE,
 264                 .cia_setkey             = ce_aes_setkey,
 265                 .cia_encrypt            = aes_cipher_encrypt,
 266                 .cia_decrypt            = aes_cipher_decrypt
 267         }
 268 };
 269
 270 static int __init aes_mod_init(void)
 271 {
 272         return crypto_register_alg(&aes_alg);
 273 }
 274
 275 static void __exit aes_mod_exit(void)
 276 {
 277         crypto_unregister_alg(&aes_alg);
 278 }
 279
 280 module_cpu_feature_match(AES, aes_mod_init);
 281 module_exit(aes_mod_exit);