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