Important changes to repositories hosted on mbed.com
Mbed hosted mercurial repositories are deprecated and are due to be permanently deleted in July 2026.
To keep a copy of this software download the repository Zip archive or clone locally using Mercurial.
It is also possible to export all your personal repositories from the account settings page.
wolfcrypt/src/aes.c@15:117db924cf7c, 2018-08-18 (annotated)
- Committer:
- wolfSSL
- Date:
- Sat Aug 18 22:20:43 2018 +0000
- Revision:
- 15:117db924cf7c
wolfSSL 3.15.3
Who changed what in which revision?
User | Revision | Line number | New contents of line |
---|---|---|---|
wolfSSL | 15:117db924cf7c | 1 | /* aes.c |
wolfSSL | 15:117db924cf7c | 2 | * |
wolfSSL | 15:117db924cf7c | 3 | * Copyright (C) 2006-2017 wolfSSL Inc. |
wolfSSL | 15:117db924cf7c | 4 | * |
wolfSSL | 15:117db924cf7c | 5 | * This file is part of wolfSSL. |
wolfSSL | 15:117db924cf7c | 6 | * |
wolfSSL | 15:117db924cf7c | 7 | * wolfSSL is free software; you can redistribute it and/or modify |
wolfSSL | 15:117db924cf7c | 8 | * it under the terms of the GNU General Public License as published by |
wolfSSL | 15:117db924cf7c | 9 | * the Free Software Foundation; either version 2 of the License, or |
wolfSSL | 15:117db924cf7c | 10 | * (at your option) any later version. |
wolfSSL | 15:117db924cf7c | 11 | * |
wolfSSL | 15:117db924cf7c | 12 | * wolfSSL is distributed in the hope that it will be useful, |
wolfSSL | 15:117db924cf7c | 13 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
wolfSSL | 15:117db924cf7c | 14 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
wolfSSL | 15:117db924cf7c | 15 | * GNU General Public License for more details. |
wolfSSL | 15:117db924cf7c | 16 | * |
wolfSSL | 15:117db924cf7c | 17 | * You should have received a copy of the GNU General Public License |
wolfSSL | 15:117db924cf7c | 18 | * along with this program; if not, write to the Free Software |
wolfSSL | 15:117db924cf7c | 19 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1335, USA |
wolfSSL | 15:117db924cf7c | 20 | */ |
wolfSSL | 15:117db924cf7c | 21 | |
wolfSSL | 15:117db924cf7c | 22 | |
wolfSSL | 15:117db924cf7c | 23 | #ifdef HAVE_CONFIG_H |
wolfSSL | 15:117db924cf7c | 24 | #include <config.h> |
wolfSSL | 15:117db924cf7c | 25 | #endif |
wolfSSL | 15:117db924cf7c | 26 | |
wolfSSL | 15:117db924cf7c | 27 | #include <wolfssl/wolfcrypt/settings.h> |
wolfSSL | 15:117db924cf7c | 28 | #include <wolfssl/wolfcrypt/error-crypt.h> |
wolfSSL | 15:117db924cf7c | 29 | |
wolfSSL | 15:117db924cf7c | 30 | #if !defined(NO_AES) |
wolfSSL | 15:117db924cf7c | 31 | |
wolfSSL | 15:117db924cf7c | 32 | #if defined(HAVE_FIPS) && \ |
wolfSSL | 15:117db924cf7c | 33 | defined(HAVE_FIPS_VERSION) && (HAVE_FIPS_VERSION >= 2) |
wolfSSL | 15:117db924cf7c | 34 | |
wolfSSL | 15:117db924cf7c | 35 | /* set NO_WRAPPERS before headers, use direct internal f()s not wrappers */ |
wolfSSL | 15:117db924cf7c | 36 | #define FIPS_NO_WRAPPERS |
wolfSSL | 15:117db924cf7c | 37 | |
wolfSSL | 15:117db924cf7c | 38 | #ifdef USE_WINDOWS_API |
wolfSSL | 15:117db924cf7c | 39 | #pragma code_seg(".fipsA$g") |
wolfSSL | 15:117db924cf7c | 40 | #pragma const_seg(".fipsB$g") |
wolfSSL | 15:117db924cf7c | 41 | #endif |
wolfSSL | 15:117db924cf7c | 42 | #endif |
wolfSSL | 15:117db924cf7c | 43 | |
wolfSSL | 15:117db924cf7c | 44 | #include <wolfssl/wolfcrypt/aes.h> |
wolfSSL | 15:117db924cf7c | 45 | #include <wolfssl/wolfcrypt/cpuid.h> |
wolfSSL | 15:117db924cf7c | 46 | |
wolfSSL | 15:117db924cf7c | 47 | |
wolfSSL | 15:117db924cf7c | 48 | /* fips wrapper calls, user can call direct */ |
wolfSSL | 15:117db924cf7c | 49 | #if defined(HAVE_FIPS) && \ |
wolfSSL | 15:117db924cf7c | 50 | (!defined(HAVE_FIPS_VERSION) || (HAVE_FIPS_VERSION < 2)) |
wolfSSL | 15:117db924cf7c | 51 | |
wolfSSL | 15:117db924cf7c | 52 | int wc_AesSetKey(Aes* aes, const byte* key, word32 len, const byte* iv, |
wolfSSL | 15:117db924cf7c | 53 | int dir) |
wolfSSL | 15:117db924cf7c | 54 | { |
wolfSSL | 15:117db924cf7c | 55 | if (aes == NULL || !( (len == 16) || (len == 24) || (len == 32)) ) { |
wolfSSL | 15:117db924cf7c | 56 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 57 | } |
wolfSSL | 15:117db924cf7c | 58 | |
wolfSSL | 15:117db924cf7c | 59 | return AesSetKey_fips(aes, key, len, iv, dir); |
wolfSSL | 15:117db924cf7c | 60 | } |
wolfSSL | 15:117db924cf7c | 61 | int wc_AesSetIV(Aes* aes, const byte* iv) |
wolfSSL | 15:117db924cf7c | 62 | { |
wolfSSL | 15:117db924cf7c | 63 | if (aes == NULL) { |
wolfSSL | 15:117db924cf7c | 64 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 65 | } |
wolfSSL | 15:117db924cf7c | 66 | |
wolfSSL | 15:117db924cf7c | 67 | return AesSetIV_fips(aes, iv); |
wolfSSL | 15:117db924cf7c | 68 | } |
wolfSSL | 15:117db924cf7c | 69 | #ifdef HAVE_AES_CBC |
wolfSSL | 15:117db924cf7c | 70 | int wc_AesCbcEncrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 71 | { |
wolfSSL | 15:117db924cf7c | 72 | if (aes == NULL || out == NULL || in == NULL) { |
wolfSSL | 15:117db924cf7c | 73 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 74 | } |
wolfSSL | 15:117db924cf7c | 75 | |
wolfSSL | 15:117db924cf7c | 76 | return AesCbcEncrypt_fips(aes, out, in, sz); |
wolfSSL | 15:117db924cf7c | 77 | } |
wolfSSL | 15:117db924cf7c | 78 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 79 | int wc_AesCbcDecrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 80 | { |
wolfSSL | 15:117db924cf7c | 81 | if (aes == NULL || out == NULL || in == NULL |
wolfSSL | 15:117db924cf7c | 82 | || sz % AES_BLOCK_SIZE != 0) { |
wolfSSL | 15:117db924cf7c | 83 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 84 | } |
wolfSSL | 15:117db924cf7c | 85 | |
wolfSSL | 15:117db924cf7c | 86 | return AesCbcDecrypt_fips(aes, out, in, sz); |
wolfSSL | 15:117db924cf7c | 87 | } |
wolfSSL | 15:117db924cf7c | 88 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 89 | #endif /* HAVE_AES_CBC */ |
wolfSSL | 15:117db924cf7c | 90 | |
wolfSSL | 15:117db924cf7c | 91 | /* AES-CTR */ |
wolfSSL | 15:117db924cf7c | 92 | #ifdef WOLFSSL_AES_COUNTER |
wolfSSL | 15:117db924cf7c | 93 | int wc_AesCtrEncrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 94 | { |
wolfSSL | 15:117db924cf7c | 95 | if (aes == NULL || out == NULL || in == NULL) { |
wolfSSL | 15:117db924cf7c | 96 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 97 | } |
wolfSSL | 15:117db924cf7c | 98 | |
wolfSSL | 15:117db924cf7c | 99 | return AesCtrEncrypt(aes, out, in, sz); |
wolfSSL | 15:117db924cf7c | 100 | } |
wolfSSL | 15:117db924cf7c | 101 | #endif |
wolfSSL | 15:117db924cf7c | 102 | |
wolfSSL | 15:117db924cf7c | 103 | /* AES-DIRECT */ |
wolfSSL | 15:117db924cf7c | 104 | #if defined(WOLFSSL_AES_DIRECT) |
wolfSSL | 15:117db924cf7c | 105 | void wc_AesEncryptDirect(Aes* aes, byte* out, const byte* in) |
wolfSSL | 15:117db924cf7c | 106 | { |
wolfSSL | 15:117db924cf7c | 107 | AesEncryptDirect(aes, out, in); |
wolfSSL | 15:117db924cf7c | 108 | } |
wolfSSL | 15:117db924cf7c | 109 | |
wolfSSL | 15:117db924cf7c | 110 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 111 | void wc_AesDecryptDirect(Aes* aes, byte* out, const byte* in) |
wolfSSL | 15:117db924cf7c | 112 | { |
wolfSSL | 15:117db924cf7c | 113 | AesDecryptDirect(aes, out, in); |
wolfSSL | 15:117db924cf7c | 114 | } |
wolfSSL | 15:117db924cf7c | 115 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 116 | |
wolfSSL | 15:117db924cf7c | 117 | int wc_AesSetKeyDirect(Aes* aes, const byte* key, word32 len, |
wolfSSL | 15:117db924cf7c | 118 | const byte* iv, int dir) |
wolfSSL | 15:117db924cf7c | 119 | { |
wolfSSL | 15:117db924cf7c | 120 | return AesSetKeyDirect(aes, key, len, iv, dir); |
wolfSSL | 15:117db924cf7c | 121 | } |
wolfSSL | 15:117db924cf7c | 122 | #endif /* WOLFSSL_AES_DIRECT */ |
wolfSSL | 15:117db924cf7c | 123 | |
wolfSSL | 15:117db924cf7c | 124 | /* AES-GCM */ |
wolfSSL | 15:117db924cf7c | 125 | #ifdef HAVE_AESGCM |
wolfSSL | 15:117db924cf7c | 126 | int wc_AesGcmSetKey(Aes* aes, const byte* key, word32 len) |
wolfSSL | 15:117db924cf7c | 127 | { |
wolfSSL | 15:117db924cf7c | 128 | if (aes == NULL || !( (len == 16) || (len == 24) || (len == 32)) ) { |
wolfSSL | 15:117db924cf7c | 129 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 130 | } |
wolfSSL | 15:117db924cf7c | 131 | |
wolfSSL | 15:117db924cf7c | 132 | return AesGcmSetKey_fips(aes, key, len); |
wolfSSL | 15:117db924cf7c | 133 | } |
wolfSSL | 15:117db924cf7c | 134 | int wc_AesGcmEncrypt(Aes* aes, byte* out, const byte* in, word32 sz, |
wolfSSL | 15:117db924cf7c | 135 | const byte* iv, word32 ivSz, |
wolfSSL | 15:117db924cf7c | 136 | byte* authTag, word32 authTagSz, |
wolfSSL | 15:117db924cf7c | 137 | const byte* authIn, word32 authInSz) |
wolfSSL | 15:117db924cf7c | 138 | { |
wolfSSL | 15:117db924cf7c | 139 | if (aes == NULL || authTagSz > AES_BLOCK_SIZE |
wolfSSL | 15:117db924cf7c | 140 | || authTagSz < WOLFSSL_MIN_AUTH_TAG_SZ || |
wolfSSL | 15:117db924cf7c | 141 | ivSz > AES_BLOCK_SIZE) { |
wolfSSL | 15:117db924cf7c | 142 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 143 | } |
wolfSSL | 15:117db924cf7c | 144 | |
wolfSSL | 15:117db924cf7c | 145 | return AesGcmEncrypt_fips(aes, out, in, sz, iv, ivSz, authTag, |
wolfSSL | 15:117db924cf7c | 146 | authTagSz, authIn, authInSz); |
wolfSSL | 15:117db924cf7c | 147 | } |
wolfSSL | 15:117db924cf7c | 148 | |
wolfSSL | 15:117db924cf7c | 149 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 150 | int wc_AesGcmDecrypt(Aes* aes, byte* out, const byte* in, word32 sz, |
wolfSSL | 15:117db924cf7c | 151 | const byte* iv, word32 ivSz, |
wolfSSL | 15:117db924cf7c | 152 | const byte* authTag, word32 authTagSz, |
wolfSSL | 15:117db924cf7c | 153 | const byte* authIn, word32 authInSz) |
wolfSSL | 15:117db924cf7c | 154 | { |
wolfSSL | 15:117db924cf7c | 155 | if (aes == NULL || out == NULL || in == NULL || iv == NULL |
wolfSSL | 15:117db924cf7c | 156 | || authTag == NULL || authTagSz > AES_BLOCK_SIZE || |
wolfSSL | 15:117db924cf7c | 157 | ivSz > AES_BLOCK_SIZE) { |
wolfSSL | 15:117db924cf7c | 158 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 159 | } |
wolfSSL | 15:117db924cf7c | 160 | |
wolfSSL | 15:117db924cf7c | 161 | return AesGcmDecrypt_fips(aes, out, in, sz, iv, ivSz, authTag, |
wolfSSL | 15:117db924cf7c | 162 | authTagSz, authIn, authInSz); |
wolfSSL | 15:117db924cf7c | 163 | } |
wolfSSL | 15:117db924cf7c | 164 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 165 | |
wolfSSL | 15:117db924cf7c | 166 | int wc_GmacSetKey(Gmac* gmac, const byte* key, word32 len) |
wolfSSL | 15:117db924cf7c | 167 | { |
wolfSSL | 15:117db924cf7c | 168 | if (gmac == NULL || key == NULL || !((len == 16) || |
wolfSSL | 15:117db924cf7c | 169 | (len == 24) || (len == 32)) ) { |
wolfSSL | 15:117db924cf7c | 170 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 171 | } |
wolfSSL | 15:117db924cf7c | 172 | |
wolfSSL | 15:117db924cf7c | 173 | return GmacSetKey(gmac, key, len); |
wolfSSL | 15:117db924cf7c | 174 | } |
wolfSSL | 15:117db924cf7c | 175 | int wc_GmacUpdate(Gmac* gmac, const byte* iv, word32 ivSz, |
wolfSSL | 15:117db924cf7c | 176 | const byte* authIn, word32 authInSz, |
wolfSSL | 15:117db924cf7c | 177 | byte* authTag, word32 authTagSz) |
wolfSSL | 15:117db924cf7c | 178 | { |
wolfSSL | 15:117db924cf7c | 179 | if (gmac == NULL || authTagSz > AES_BLOCK_SIZE || |
wolfSSL | 15:117db924cf7c | 180 | authTagSz < WOLFSSL_MIN_AUTH_TAG_SZ) { |
wolfSSL | 15:117db924cf7c | 181 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 182 | } |
wolfSSL | 15:117db924cf7c | 183 | |
wolfSSL | 15:117db924cf7c | 184 | return GmacUpdate(gmac, iv, ivSz, authIn, authInSz, |
wolfSSL | 15:117db924cf7c | 185 | authTag, authTagSz); |
wolfSSL | 15:117db924cf7c | 186 | } |
wolfSSL | 15:117db924cf7c | 187 | #endif /* HAVE_AESGCM */ |
wolfSSL | 15:117db924cf7c | 188 | |
wolfSSL | 15:117db924cf7c | 189 | /* AES-CCM */ |
wolfSSL | 15:117db924cf7c | 190 | #if defined(HAVE_AESCCM) && \ |
wolfSSL | 15:117db924cf7c | 191 | defined(HAVE_FIPS_VERSION) && (HAVE_FIPS_VERSION >= 2) |
wolfSSL | 15:117db924cf7c | 192 | int wc_AesCcmSetKey(Aes* aes, const byte* key, word32 keySz) |
wolfSSL | 15:117db924cf7c | 193 | { |
wolfSSL | 15:117db924cf7c | 194 | return AesCcmSetKey(aes, key, keySz); |
wolfSSL | 15:117db924cf7c | 195 | } |
wolfSSL | 15:117db924cf7c | 196 | int wc_AesCcmEncrypt(Aes* aes, byte* out, const byte* in, word32 inSz, |
wolfSSL | 15:117db924cf7c | 197 | const byte* nonce, word32 nonceSz, |
wolfSSL | 15:117db924cf7c | 198 | byte* authTag, word32 authTagSz, |
wolfSSL | 15:117db924cf7c | 199 | const byte* authIn, word32 authInSz) |
wolfSSL | 15:117db924cf7c | 200 | { |
wolfSSL | 15:117db924cf7c | 201 | /* sanity check on arguments */ |
wolfSSL | 15:117db924cf7c | 202 | if (aes == NULL || out == NULL || in == NULL || nonce == NULL |
wolfSSL | 15:117db924cf7c | 203 | || authTag == NULL || nonceSz < 7 || nonceSz > 13) |
wolfSSL | 15:117db924cf7c | 204 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 205 | |
wolfSSL | 15:117db924cf7c | 206 | AesCcmEncrypt(aes, out, in, inSz, nonce, nonceSz, authTag, |
wolfSSL | 15:117db924cf7c | 207 | authTagSz, authIn, authInSz); |
wolfSSL | 15:117db924cf7c | 208 | return 0; |
wolfSSL | 15:117db924cf7c | 209 | } |
wolfSSL | 15:117db924cf7c | 210 | |
wolfSSL | 15:117db924cf7c | 211 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 212 | int wc_AesCcmDecrypt(Aes* aes, byte* out, |
wolfSSL | 15:117db924cf7c | 213 | const byte* in, word32 inSz, |
wolfSSL | 15:117db924cf7c | 214 | const byte* nonce, word32 nonceSz, |
wolfSSL | 15:117db924cf7c | 215 | const byte* authTag, word32 authTagSz, |
wolfSSL | 15:117db924cf7c | 216 | const byte* authIn, word32 authInSz) |
wolfSSL | 15:117db924cf7c | 217 | { |
wolfSSL | 15:117db924cf7c | 218 | |
wolfSSL | 15:117db924cf7c | 219 | if (aes == NULL || out == NULL || in == NULL || nonce == NULL |
wolfSSL | 15:117db924cf7c | 220 | || authTag == NULL || nonceSz < 7 || nonceSz > 13) { |
wolfSSL | 15:117db924cf7c | 221 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 222 | } |
wolfSSL | 15:117db924cf7c | 223 | |
wolfSSL | 15:117db924cf7c | 224 | return AesCcmDecrypt(aes, out, in, inSz, nonce, nonceSz, |
wolfSSL | 15:117db924cf7c | 225 | authTag, authTagSz, authIn, authInSz); |
wolfSSL | 15:117db924cf7c | 226 | } |
wolfSSL | 15:117db924cf7c | 227 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 228 | #endif /* HAVE_AESCCM && HAVE_FIPS_VERSION 2 */ |
wolfSSL | 15:117db924cf7c | 229 | |
wolfSSL | 15:117db924cf7c | 230 | int wc_AesInit(Aes* aes, void* h, int i) |
wolfSSL | 15:117db924cf7c | 231 | { |
wolfSSL | 15:117db924cf7c | 232 | (void)aes; |
wolfSSL | 15:117db924cf7c | 233 | (void)h; |
wolfSSL | 15:117db924cf7c | 234 | (void)i; |
wolfSSL | 15:117db924cf7c | 235 | /* FIPS doesn't support: |
wolfSSL | 15:117db924cf7c | 236 | return AesInit(aes, h, i); */ |
wolfSSL | 15:117db924cf7c | 237 | return 0; |
wolfSSL | 15:117db924cf7c | 238 | } |
wolfSSL | 15:117db924cf7c | 239 | void wc_AesFree(Aes* aes) |
wolfSSL | 15:117db924cf7c | 240 | { |
wolfSSL | 15:117db924cf7c | 241 | (void)aes; |
wolfSSL | 15:117db924cf7c | 242 | /* FIPS doesn't support: |
wolfSSL | 15:117db924cf7c | 243 | AesFree(aes); */ |
wolfSSL | 15:117db924cf7c | 244 | } |
wolfSSL | 15:117db924cf7c | 245 | |
wolfSSL | 15:117db924cf7c | 246 | #else /* else build without fips, or for FIPS v2 */ |
wolfSSL | 15:117db924cf7c | 247 | |
wolfSSL | 15:117db924cf7c | 248 | |
wolfSSL | 15:117db924cf7c | 249 | #if defined(WOLFSSL_TI_CRYPT) |
wolfSSL | 15:117db924cf7c | 250 | #include <wolfcrypt/src/port/ti/ti-aes.c> |
wolfSSL | 15:117db924cf7c | 251 | #else |
wolfSSL | 15:117db924cf7c | 252 | |
wolfSSL | 15:117db924cf7c | 253 | #include <wolfssl/wolfcrypt/logging.h> |
wolfSSL | 15:117db924cf7c | 254 | |
wolfSSL | 15:117db924cf7c | 255 | #ifdef NO_INLINE |
wolfSSL | 15:117db924cf7c | 256 | #include <wolfssl/wolfcrypt/misc.h> |
wolfSSL | 15:117db924cf7c | 257 | #else |
wolfSSL | 15:117db924cf7c | 258 | #define WOLFSSL_MISC_INCLUDED |
wolfSSL | 15:117db924cf7c | 259 | #include <wolfcrypt/src/misc.c> |
wolfSSL | 15:117db924cf7c | 260 | #endif |
wolfSSL | 15:117db924cf7c | 261 | |
wolfSSL | 15:117db924cf7c | 262 | #if !defined(WOLFSSL_ARMASM) |
wolfSSL | 15:117db924cf7c | 263 | |
wolfSSL | 15:117db924cf7c | 264 | #ifdef WOLFSSL_IMX6_CAAM_BLOB |
wolfSSL | 15:117db924cf7c | 265 | /* case of possibly not using hardware acceleration for AES but using key |
wolfSSL | 15:117db924cf7c | 266 | blobs */ |
wolfSSL | 15:117db924cf7c | 267 | #include <wolfssl/wolfcrypt/port/caam/wolfcaam.h> |
wolfSSL | 15:117db924cf7c | 268 | #endif |
wolfSSL | 15:117db924cf7c | 269 | |
wolfSSL | 15:117db924cf7c | 270 | #ifdef DEBUG_AESNI |
wolfSSL | 15:117db924cf7c | 271 | #include <stdio.h> |
wolfSSL | 15:117db924cf7c | 272 | #endif |
wolfSSL | 15:117db924cf7c | 273 | |
wolfSSL | 15:117db924cf7c | 274 | #ifdef _MSC_VER |
wolfSSL | 15:117db924cf7c | 275 | /* 4127 warning constant while(1) */ |
wolfSSL | 15:117db924cf7c | 276 | #pragma warning(disable: 4127) |
wolfSSL | 15:117db924cf7c | 277 | #endif |
wolfSSL | 15:117db924cf7c | 278 | |
wolfSSL | 15:117db924cf7c | 279 | |
wolfSSL | 15:117db924cf7c | 280 | /* Define AES implementation includes and functions */ |
wolfSSL | 15:117db924cf7c | 281 | #if defined(STM32_CRYPTO) |
wolfSSL | 15:117db924cf7c | 282 | /* STM32F2/F4 hardware AES support for CBC, CTR modes */ |
wolfSSL | 15:117db924cf7c | 283 | |
wolfSSL | 15:117db924cf7c | 284 | #ifdef WOLFSSL_STM32L4 |
wolfSSL | 15:117db924cf7c | 285 | #define CRYP AES |
wolfSSL | 15:117db924cf7c | 286 | #endif |
wolfSSL | 15:117db924cf7c | 287 | |
wolfSSL | 15:117db924cf7c | 288 | /* CRYPT_AES_GCM starts the IV with 2 */ |
wolfSSL | 15:117db924cf7c | 289 | #define STM32_GCM_IV_START 2 |
wolfSSL | 15:117db924cf7c | 290 | |
wolfSSL | 15:117db924cf7c | 291 | #if defined(WOLFSSL_AES_DIRECT) || defined(HAVE_AESGCM) || defined(HAVE_AESCCM) |
wolfSSL | 15:117db924cf7c | 292 | static int wc_AesEncrypt(Aes* aes, const byte* inBlock, byte* outBlock) |
wolfSSL | 15:117db924cf7c | 293 | { |
wolfSSL | 15:117db924cf7c | 294 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 295 | #ifdef WOLFSSL_STM32_CUBEMX |
wolfSSL | 15:117db924cf7c | 296 | CRYP_HandleTypeDef hcryp; |
wolfSSL | 15:117db924cf7c | 297 | |
wolfSSL | 15:117db924cf7c | 298 | XMEMSET(&hcryp, 0, sizeof(CRYP_HandleTypeDef)); |
wolfSSL | 15:117db924cf7c | 299 | switch(aes->rounds) { |
wolfSSL | 15:117db924cf7c | 300 | case 10: /* 128-bit key */ |
wolfSSL | 15:117db924cf7c | 301 | hcryp.Init.KeySize = CRYP_KEYSIZE_128B; |
wolfSSL | 15:117db924cf7c | 302 | break; |
wolfSSL | 15:117db924cf7c | 303 | #ifdef CRYP_KEYSIZE_192B |
wolfSSL | 15:117db924cf7c | 304 | case 12: /* 192-bit key */ |
wolfSSL | 15:117db924cf7c | 305 | hcryp.Init.KeySize = CRYP_KEYSIZE_192B; |
wolfSSL | 15:117db924cf7c | 306 | break; |
wolfSSL | 15:117db924cf7c | 307 | #endif |
wolfSSL | 15:117db924cf7c | 308 | case 14: /* 256-bit key */ |
wolfSSL | 15:117db924cf7c | 309 | hcryp.Init.KeySize = CRYP_KEYSIZE_256B; |
wolfSSL | 15:117db924cf7c | 310 | break; |
wolfSSL | 15:117db924cf7c | 311 | default: |
wolfSSL | 15:117db924cf7c | 312 | break; |
wolfSSL | 15:117db924cf7c | 313 | } |
wolfSSL | 15:117db924cf7c | 314 | hcryp.Instance = CRYP; |
wolfSSL | 15:117db924cf7c | 315 | hcryp.Init.DataType = CRYP_DATATYPE_8B; |
wolfSSL | 15:117db924cf7c | 316 | hcryp.Init.pKey = (uint8_t*)aes->key; |
wolfSSL | 15:117db924cf7c | 317 | |
wolfSSL | 15:117db924cf7c | 318 | HAL_CRYP_Init(&hcryp); |
wolfSSL | 15:117db924cf7c | 319 | |
wolfSSL | 15:117db924cf7c | 320 | if (HAL_CRYP_AESECB_Encrypt(&hcryp, (uint8_t*)inBlock, AES_BLOCK_SIZE, |
wolfSSL | 15:117db924cf7c | 321 | outBlock, STM32_HAL_TIMEOUT) != HAL_OK) { |
wolfSSL | 15:117db924cf7c | 322 | ret = WC_TIMEOUT_E; |
wolfSSL | 15:117db924cf7c | 323 | } |
wolfSSL | 15:117db924cf7c | 324 | |
wolfSSL | 15:117db924cf7c | 325 | HAL_CRYP_DeInit(&hcryp); |
wolfSSL | 15:117db924cf7c | 326 | #else |
wolfSSL | 15:117db924cf7c | 327 | word32 *enc_key; |
wolfSSL | 15:117db924cf7c | 328 | CRYP_InitTypeDef AES_CRYP_InitStructure; |
wolfSSL | 15:117db924cf7c | 329 | CRYP_KeyInitTypeDef AES_CRYP_KeyInitStructure; |
wolfSSL | 15:117db924cf7c | 330 | |
wolfSSL | 15:117db924cf7c | 331 | enc_key = aes->key; |
wolfSSL | 15:117db924cf7c | 332 | |
wolfSSL | 15:117db924cf7c | 333 | /* crypto structure initialization */ |
wolfSSL | 15:117db924cf7c | 334 | CRYP_KeyStructInit(&AES_CRYP_KeyInitStructure); |
wolfSSL | 15:117db924cf7c | 335 | CRYP_StructInit(&AES_CRYP_InitStructure); |
wolfSSL | 15:117db924cf7c | 336 | |
wolfSSL | 15:117db924cf7c | 337 | /* reset registers to their default values */ |
wolfSSL | 15:117db924cf7c | 338 | CRYP_DeInit(); |
wolfSSL | 15:117db924cf7c | 339 | |
wolfSSL | 15:117db924cf7c | 340 | /* load key into correct registers */ |
wolfSSL | 15:117db924cf7c | 341 | switch (aes->rounds) { |
wolfSSL | 15:117db924cf7c | 342 | case 10: /* 128-bit key */ |
wolfSSL | 15:117db924cf7c | 343 | AES_CRYP_InitStructure.CRYP_KeySize = CRYP_KeySize_128b; |
wolfSSL | 15:117db924cf7c | 344 | AES_CRYP_KeyInitStructure.CRYP_Key2Left = enc_key[0]; |
wolfSSL | 15:117db924cf7c | 345 | AES_CRYP_KeyInitStructure.CRYP_Key2Right = enc_key[1]; |
wolfSSL | 15:117db924cf7c | 346 | AES_CRYP_KeyInitStructure.CRYP_Key3Left = enc_key[2]; |
wolfSSL | 15:117db924cf7c | 347 | AES_CRYP_KeyInitStructure.CRYP_Key3Right = enc_key[3]; |
wolfSSL | 15:117db924cf7c | 348 | break; |
wolfSSL | 15:117db924cf7c | 349 | |
wolfSSL | 15:117db924cf7c | 350 | case 12: /* 192-bit key */ |
wolfSSL | 15:117db924cf7c | 351 | AES_CRYP_InitStructure.CRYP_KeySize = CRYP_KeySize_192b; |
wolfSSL | 15:117db924cf7c | 352 | AES_CRYP_KeyInitStructure.CRYP_Key1Left = enc_key[0]; |
wolfSSL | 15:117db924cf7c | 353 | AES_CRYP_KeyInitStructure.CRYP_Key1Right = enc_key[1]; |
wolfSSL | 15:117db924cf7c | 354 | AES_CRYP_KeyInitStructure.CRYP_Key2Left = enc_key[2]; |
wolfSSL | 15:117db924cf7c | 355 | AES_CRYP_KeyInitStructure.CRYP_Key2Right = enc_key[3]; |
wolfSSL | 15:117db924cf7c | 356 | AES_CRYP_KeyInitStructure.CRYP_Key3Left = enc_key[4]; |
wolfSSL | 15:117db924cf7c | 357 | AES_CRYP_KeyInitStructure.CRYP_Key3Right = enc_key[5]; |
wolfSSL | 15:117db924cf7c | 358 | break; |
wolfSSL | 15:117db924cf7c | 359 | |
wolfSSL | 15:117db924cf7c | 360 | case 14: /* 256-bit key */ |
wolfSSL | 15:117db924cf7c | 361 | AES_CRYP_InitStructure.CRYP_KeySize = CRYP_KeySize_256b; |
wolfSSL | 15:117db924cf7c | 362 | AES_CRYP_KeyInitStructure.CRYP_Key0Left = enc_key[0]; |
wolfSSL | 15:117db924cf7c | 363 | AES_CRYP_KeyInitStructure.CRYP_Key0Right = enc_key[1]; |
wolfSSL | 15:117db924cf7c | 364 | AES_CRYP_KeyInitStructure.CRYP_Key1Left = enc_key[2]; |
wolfSSL | 15:117db924cf7c | 365 | AES_CRYP_KeyInitStructure.CRYP_Key1Right = enc_key[3]; |
wolfSSL | 15:117db924cf7c | 366 | AES_CRYP_KeyInitStructure.CRYP_Key2Left = enc_key[4]; |
wolfSSL | 15:117db924cf7c | 367 | AES_CRYP_KeyInitStructure.CRYP_Key2Right = enc_key[5]; |
wolfSSL | 15:117db924cf7c | 368 | AES_CRYP_KeyInitStructure.CRYP_Key3Left = enc_key[6]; |
wolfSSL | 15:117db924cf7c | 369 | AES_CRYP_KeyInitStructure.CRYP_Key3Right = enc_key[7]; |
wolfSSL | 15:117db924cf7c | 370 | break; |
wolfSSL | 15:117db924cf7c | 371 | |
wolfSSL | 15:117db924cf7c | 372 | default: |
wolfSSL | 15:117db924cf7c | 373 | break; |
wolfSSL | 15:117db924cf7c | 374 | } |
wolfSSL | 15:117db924cf7c | 375 | CRYP_KeyInit(&AES_CRYP_KeyInitStructure); |
wolfSSL | 15:117db924cf7c | 376 | |
wolfSSL | 15:117db924cf7c | 377 | /* set direction, mode, and datatype */ |
wolfSSL | 15:117db924cf7c | 378 | AES_CRYP_InitStructure.CRYP_AlgoDir = CRYP_AlgoDir_Encrypt; |
wolfSSL | 15:117db924cf7c | 379 | AES_CRYP_InitStructure.CRYP_AlgoMode = CRYP_AlgoMode_AES_ECB; |
wolfSSL | 15:117db924cf7c | 380 | AES_CRYP_InitStructure.CRYP_DataType = CRYP_DataType_8b; |
wolfSSL | 15:117db924cf7c | 381 | CRYP_Init(&AES_CRYP_InitStructure); |
wolfSSL | 15:117db924cf7c | 382 | |
wolfSSL | 15:117db924cf7c | 383 | /* enable crypto processor */ |
wolfSSL | 15:117db924cf7c | 384 | CRYP_Cmd(ENABLE); |
wolfSSL | 15:117db924cf7c | 385 | |
wolfSSL | 15:117db924cf7c | 386 | /* flush IN/OUT FIFOs */ |
wolfSSL | 15:117db924cf7c | 387 | CRYP_FIFOFlush(); |
wolfSSL | 15:117db924cf7c | 388 | |
wolfSSL | 15:117db924cf7c | 389 | CRYP_DataIn(*(uint32_t*)&inBlock[0]); |
wolfSSL | 15:117db924cf7c | 390 | CRYP_DataIn(*(uint32_t*)&inBlock[4]); |
wolfSSL | 15:117db924cf7c | 391 | CRYP_DataIn(*(uint32_t*)&inBlock[8]); |
wolfSSL | 15:117db924cf7c | 392 | CRYP_DataIn(*(uint32_t*)&inBlock[12]); |
wolfSSL | 15:117db924cf7c | 393 | |
wolfSSL | 15:117db924cf7c | 394 | /* wait until the complete message has been processed */ |
wolfSSL | 15:117db924cf7c | 395 | while (CRYP_GetFlagStatus(CRYP_FLAG_BUSY) != RESET) {} |
wolfSSL | 15:117db924cf7c | 396 | |
wolfSSL | 15:117db924cf7c | 397 | *(uint32_t*)&outBlock[0] = CRYP_DataOut(); |
wolfSSL | 15:117db924cf7c | 398 | *(uint32_t*)&outBlock[4] = CRYP_DataOut(); |
wolfSSL | 15:117db924cf7c | 399 | *(uint32_t*)&outBlock[8] = CRYP_DataOut(); |
wolfSSL | 15:117db924cf7c | 400 | *(uint32_t*)&outBlock[12] = CRYP_DataOut(); |
wolfSSL | 15:117db924cf7c | 401 | |
wolfSSL | 15:117db924cf7c | 402 | /* disable crypto processor */ |
wolfSSL | 15:117db924cf7c | 403 | CRYP_Cmd(DISABLE); |
wolfSSL | 15:117db924cf7c | 404 | #endif /* WOLFSSL_STM32_CUBEMX */ |
wolfSSL | 15:117db924cf7c | 405 | return ret; |
wolfSSL | 15:117db924cf7c | 406 | } |
wolfSSL | 15:117db924cf7c | 407 | #endif /* WOLFSSL_AES_DIRECT || HAVE_AESGCM || HAVE_AESCCM */ |
wolfSSL | 15:117db924cf7c | 408 | |
wolfSSL | 15:117db924cf7c | 409 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 410 | #if defined(WOLFSSL_AES_DIRECT) || defined(HAVE_AESCCM) |
wolfSSL | 15:117db924cf7c | 411 | static int wc_AesDecrypt(Aes* aes, const byte* inBlock, byte* outBlock) |
wolfSSL | 15:117db924cf7c | 412 | { |
wolfSSL | 15:117db924cf7c | 413 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 414 | #ifdef WOLFSSL_STM32_CUBEMX |
wolfSSL | 15:117db924cf7c | 415 | CRYP_HandleTypeDef hcryp; |
wolfSSL | 15:117db924cf7c | 416 | |
wolfSSL | 15:117db924cf7c | 417 | XMEMSET(&hcryp, 0, sizeof(CRYP_HandleTypeDef)); |
wolfSSL | 15:117db924cf7c | 418 | switch(aes->rounds) { |
wolfSSL | 15:117db924cf7c | 419 | case 10: /* 128-bit key */ |
wolfSSL | 15:117db924cf7c | 420 | hcryp.Init.KeySize = CRYP_KEYSIZE_128B; |
wolfSSL | 15:117db924cf7c | 421 | break; |
wolfSSL | 15:117db924cf7c | 422 | #ifdef CRYP_KEYSIZE_192B |
wolfSSL | 15:117db924cf7c | 423 | case 12: /* 192-bit key */ |
wolfSSL | 15:117db924cf7c | 424 | hcryp.Init.KeySize = CRYP_KEYSIZE_192B; |
wolfSSL | 15:117db924cf7c | 425 | break; |
wolfSSL | 15:117db924cf7c | 426 | #endif |
wolfSSL | 15:117db924cf7c | 427 | case 14: /* 256-bit key */ |
wolfSSL | 15:117db924cf7c | 428 | hcryp.Init.KeySize = CRYP_KEYSIZE_256B; |
wolfSSL | 15:117db924cf7c | 429 | break; |
wolfSSL | 15:117db924cf7c | 430 | default: |
wolfSSL | 15:117db924cf7c | 431 | break; |
wolfSSL | 15:117db924cf7c | 432 | } |
wolfSSL | 15:117db924cf7c | 433 | hcryp.Instance = CRYP; |
wolfSSL | 15:117db924cf7c | 434 | hcryp.Init.DataType = CRYP_DATATYPE_8B; |
wolfSSL | 15:117db924cf7c | 435 | hcryp.Init.pKey = (uint8_t*)aes->key; |
wolfSSL | 15:117db924cf7c | 436 | |
wolfSSL | 15:117db924cf7c | 437 | HAL_CRYP_Init(&hcryp); |
wolfSSL | 15:117db924cf7c | 438 | |
wolfSSL | 15:117db924cf7c | 439 | if (HAL_CRYP_AESECB_Decrypt(&hcryp, (uint8_t*)inBlock, AES_BLOCK_SIZE, |
wolfSSL | 15:117db924cf7c | 440 | outBlock, STM32_HAL_TIMEOUT) != HAL_OK) { |
wolfSSL | 15:117db924cf7c | 441 | ret = WC_TIMEOUT_E; |
wolfSSL | 15:117db924cf7c | 442 | } |
wolfSSL | 15:117db924cf7c | 443 | |
wolfSSL | 15:117db924cf7c | 444 | HAL_CRYP_DeInit(&hcryp); |
wolfSSL | 15:117db924cf7c | 445 | #else |
wolfSSL | 15:117db924cf7c | 446 | word32 *enc_key; |
wolfSSL | 15:117db924cf7c | 447 | CRYP_InitTypeDef AES_CRYP_InitStructure; |
wolfSSL | 15:117db924cf7c | 448 | CRYP_KeyInitTypeDef AES_CRYP_KeyInitStructure; |
wolfSSL | 15:117db924cf7c | 449 | |
wolfSSL | 15:117db924cf7c | 450 | enc_key = aes->key; |
wolfSSL | 15:117db924cf7c | 451 | |
wolfSSL | 15:117db924cf7c | 452 | /* crypto structure initialization */ |
wolfSSL | 15:117db924cf7c | 453 | CRYP_KeyStructInit(&AES_CRYP_KeyInitStructure); |
wolfSSL | 15:117db924cf7c | 454 | CRYP_StructInit(&AES_CRYP_InitStructure); |
wolfSSL | 15:117db924cf7c | 455 | |
wolfSSL | 15:117db924cf7c | 456 | /* reset registers to their default values */ |
wolfSSL | 15:117db924cf7c | 457 | CRYP_DeInit(); |
wolfSSL | 15:117db924cf7c | 458 | |
wolfSSL | 15:117db924cf7c | 459 | /* load key into correct registers */ |
wolfSSL | 15:117db924cf7c | 460 | switch (aes->rounds) { |
wolfSSL | 15:117db924cf7c | 461 | case 10: /* 128-bit key */ |
wolfSSL | 15:117db924cf7c | 462 | AES_CRYP_InitStructure.CRYP_KeySize = CRYP_KeySize_128b; |
wolfSSL | 15:117db924cf7c | 463 | AES_CRYP_KeyInitStructure.CRYP_Key2Left = enc_key[0]; |
wolfSSL | 15:117db924cf7c | 464 | AES_CRYP_KeyInitStructure.CRYP_Key2Right = enc_key[1]; |
wolfSSL | 15:117db924cf7c | 465 | AES_CRYP_KeyInitStructure.CRYP_Key3Left = enc_key[2]; |
wolfSSL | 15:117db924cf7c | 466 | AES_CRYP_KeyInitStructure.CRYP_Key3Right = enc_key[3]; |
wolfSSL | 15:117db924cf7c | 467 | break; |
wolfSSL | 15:117db924cf7c | 468 | |
wolfSSL | 15:117db924cf7c | 469 | case 12: /* 192-bit key */ |
wolfSSL | 15:117db924cf7c | 470 | AES_CRYP_InitStructure.CRYP_KeySize = CRYP_KeySize_192b; |
wolfSSL | 15:117db924cf7c | 471 | AES_CRYP_KeyInitStructure.CRYP_Key1Left = enc_key[0]; |
wolfSSL | 15:117db924cf7c | 472 | AES_CRYP_KeyInitStructure.CRYP_Key1Right = enc_key[1]; |
wolfSSL | 15:117db924cf7c | 473 | AES_CRYP_KeyInitStructure.CRYP_Key2Left = enc_key[2]; |
wolfSSL | 15:117db924cf7c | 474 | AES_CRYP_KeyInitStructure.CRYP_Key2Right = enc_key[3]; |
wolfSSL | 15:117db924cf7c | 475 | AES_CRYP_KeyInitStructure.CRYP_Key3Left = enc_key[4]; |
wolfSSL | 15:117db924cf7c | 476 | AES_CRYP_KeyInitStructure.CRYP_Key3Right = enc_key[5]; |
wolfSSL | 15:117db924cf7c | 477 | break; |
wolfSSL | 15:117db924cf7c | 478 | |
wolfSSL | 15:117db924cf7c | 479 | case 14: /* 256-bit key */ |
wolfSSL | 15:117db924cf7c | 480 | AES_CRYP_InitStructure.CRYP_KeySize = CRYP_KeySize_256b; |
wolfSSL | 15:117db924cf7c | 481 | AES_CRYP_KeyInitStructure.CRYP_Key0Left = enc_key[0]; |
wolfSSL | 15:117db924cf7c | 482 | AES_CRYP_KeyInitStructure.CRYP_Key0Right = enc_key[1]; |
wolfSSL | 15:117db924cf7c | 483 | AES_CRYP_KeyInitStructure.CRYP_Key1Left = enc_key[2]; |
wolfSSL | 15:117db924cf7c | 484 | AES_CRYP_KeyInitStructure.CRYP_Key1Right = enc_key[3]; |
wolfSSL | 15:117db924cf7c | 485 | AES_CRYP_KeyInitStructure.CRYP_Key2Left = enc_key[4]; |
wolfSSL | 15:117db924cf7c | 486 | AES_CRYP_KeyInitStructure.CRYP_Key2Right = enc_key[5]; |
wolfSSL | 15:117db924cf7c | 487 | AES_CRYP_KeyInitStructure.CRYP_Key3Left = enc_key[6]; |
wolfSSL | 15:117db924cf7c | 488 | AES_CRYP_KeyInitStructure.CRYP_Key3Right = enc_key[7]; |
wolfSSL | 15:117db924cf7c | 489 | break; |
wolfSSL | 15:117db924cf7c | 490 | |
wolfSSL | 15:117db924cf7c | 491 | default: |
wolfSSL | 15:117db924cf7c | 492 | break; |
wolfSSL | 15:117db924cf7c | 493 | } |
wolfSSL | 15:117db924cf7c | 494 | CRYP_KeyInit(&AES_CRYP_KeyInitStructure); |
wolfSSL | 15:117db924cf7c | 495 | |
wolfSSL | 15:117db924cf7c | 496 | /* set direction, key, and datatype */ |
wolfSSL | 15:117db924cf7c | 497 | AES_CRYP_InitStructure.CRYP_AlgoDir = CRYP_AlgoDir_Decrypt; |
wolfSSL | 15:117db924cf7c | 498 | AES_CRYP_InitStructure.CRYP_AlgoMode = CRYP_AlgoMode_AES_Key; |
wolfSSL | 15:117db924cf7c | 499 | AES_CRYP_InitStructure.CRYP_DataType = CRYP_DataType_8b; |
wolfSSL | 15:117db924cf7c | 500 | CRYP_Init(&AES_CRYP_InitStructure); |
wolfSSL | 15:117db924cf7c | 501 | |
wolfSSL | 15:117db924cf7c | 502 | /* enable crypto processor */ |
wolfSSL | 15:117db924cf7c | 503 | CRYP_Cmd(ENABLE); |
wolfSSL | 15:117db924cf7c | 504 | |
wolfSSL | 15:117db924cf7c | 505 | /* wait until decrypt key has been intialized */ |
wolfSSL | 15:117db924cf7c | 506 | while (CRYP_GetFlagStatus(CRYP_FLAG_BUSY) != RESET) {} |
wolfSSL | 15:117db924cf7c | 507 | |
wolfSSL | 15:117db924cf7c | 508 | /* set direction, mode, and datatype */ |
wolfSSL | 15:117db924cf7c | 509 | AES_CRYP_InitStructure.CRYP_AlgoDir = CRYP_AlgoDir_Decrypt; |
wolfSSL | 15:117db924cf7c | 510 | AES_CRYP_InitStructure.CRYP_AlgoMode = CRYP_AlgoMode_AES_ECB; |
wolfSSL | 15:117db924cf7c | 511 | AES_CRYP_InitStructure.CRYP_DataType = CRYP_DataType_8b; |
wolfSSL | 15:117db924cf7c | 512 | CRYP_Init(&AES_CRYP_InitStructure); |
wolfSSL | 15:117db924cf7c | 513 | |
wolfSSL | 15:117db924cf7c | 514 | /* enable crypto processor */ |
wolfSSL | 15:117db924cf7c | 515 | CRYP_Cmd(ENABLE); |
wolfSSL | 15:117db924cf7c | 516 | |
wolfSSL | 15:117db924cf7c | 517 | /* flush IN/OUT FIFOs */ |
wolfSSL | 15:117db924cf7c | 518 | CRYP_FIFOFlush(); |
wolfSSL | 15:117db924cf7c | 519 | |
wolfSSL | 15:117db924cf7c | 520 | CRYP_DataIn(*(uint32_t*)&inBlock[0]); |
wolfSSL | 15:117db924cf7c | 521 | CRYP_DataIn(*(uint32_t*)&inBlock[4]); |
wolfSSL | 15:117db924cf7c | 522 | CRYP_DataIn(*(uint32_t*)&inBlock[8]); |
wolfSSL | 15:117db924cf7c | 523 | CRYP_DataIn(*(uint32_t*)&inBlock[12]); |
wolfSSL | 15:117db924cf7c | 524 | |
wolfSSL | 15:117db924cf7c | 525 | /* wait until the complete message has been processed */ |
wolfSSL | 15:117db924cf7c | 526 | while (CRYP_GetFlagStatus(CRYP_FLAG_BUSY) != RESET) {} |
wolfSSL | 15:117db924cf7c | 527 | |
wolfSSL | 15:117db924cf7c | 528 | *(uint32_t*)&outBlock[0] = CRYP_DataOut(); |
wolfSSL | 15:117db924cf7c | 529 | *(uint32_t*)&outBlock[4] = CRYP_DataOut(); |
wolfSSL | 15:117db924cf7c | 530 | *(uint32_t*)&outBlock[8] = CRYP_DataOut(); |
wolfSSL | 15:117db924cf7c | 531 | *(uint32_t*)&outBlock[12] = CRYP_DataOut(); |
wolfSSL | 15:117db924cf7c | 532 | |
wolfSSL | 15:117db924cf7c | 533 | /* disable crypto processor */ |
wolfSSL | 15:117db924cf7c | 534 | CRYP_Cmd(DISABLE); |
wolfSSL | 15:117db924cf7c | 535 | #endif /* WOLFSSL_STM32_CUBEMX */ |
wolfSSL | 15:117db924cf7c | 536 | return ret; |
wolfSSL | 15:117db924cf7c | 537 | } |
wolfSSL | 15:117db924cf7c | 538 | #endif /* WOLFSSL_AES_DIRECT || HAVE_AESCCM */ |
wolfSSL | 15:117db924cf7c | 539 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 540 | |
wolfSSL | 15:117db924cf7c | 541 | #elif defined(HAVE_COLDFIRE_SEC) |
wolfSSL | 15:117db924cf7c | 542 | /* Freescale Coldfire SEC support for CBC mode. |
wolfSSL | 15:117db924cf7c | 543 | * NOTE: no support for AES-CTR/GCM/CCM/Direct */ |
wolfSSL | 15:117db924cf7c | 544 | #include <wolfssl/wolfcrypt/types.h> |
wolfSSL | 15:117db924cf7c | 545 | #include "sec.h" |
wolfSSL | 15:117db924cf7c | 546 | #include "mcf5475_sec.h" |
wolfSSL | 15:117db924cf7c | 547 | #include "mcf5475_siu.h" |
wolfSSL | 15:117db924cf7c | 548 | #elif defined(FREESCALE_LTC) |
wolfSSL | 15:117db924cf7c | 549 | #include "fsl_ltc.h" |
wolfSSL | 15:117db924cf7c | 550 | #if defined(FREESCALE_LTC_AES_GCM) |
wolfSSL | 15:117db924cf7c | 551 | #undef NEED_AES_TABLES |
wolfSSL | 15:117db924cf7c | 552 | #undef GCM_TABLE |
wolfSSL | 15:117db924cf7c | 553 | #else |
wolfSSL | 15:117db924cf7c | 554 | /* if LTC doesn't have GCM, use software with LTC AES ECB mode */ |
wolfSSL | 15:117db924cf7c | 555 | static int wc_AesEncrypt(Aes* aes, const byte* inBlock, byte* outBlock) |
wolfSSL | 15:117db924cf7c | 556 | { |
wolfSSL | 15:117db924cf7c | 557 | wc_AesEncryptDirect(aes, outBlock, inBlock); |
wolfSSL | 15:117db924cf7c | 558 | return 0; |
wolfSSL | 15:117db924cf7c | 559 | } |
wolfSSL | 15:117db924cf7c | 560 | static int wc_AesDecrypt(Aes* aes, const byte* inBlock, byte* outBlock) |
wolfSSL | 15:117db924cf7c | 561 | { |
wolfSSL | 15:117db924cf7c | 562 | wc_AesDecryptDirect(aes, outBlock, inBlock); |
wolfSSL | 15:117db924cf7c | 563 | return 0; |
wolfSSL | 15:117db924cf7c | 564 | } |
wolfSSL | 15:117db924cf7c | 565 | #endif |
wolfSSL | 15:117db924cf7c | 566 | #elif defined(FREESCALE_MMCAU) |
wolfSSL | 15:117db924cf7c | 567 | /* Freescale mmCAU hardware AES support for Direct, CBC, CCM, GCM modes |
wolfSSL | 15:117db924cf7c | 568 | * through the CAU/mmCAU library. Documentation located in |
wolfSSL | 15:117db924cf7c | 569 | * ColdFire/ColdFire+ CAU and Kinetis mmCAU Software Library User |
wolfSSL | 15:117db924cf7c | 570 | * Guide (See note in README). */ |
wolfSSL | 15:117db924cf7c | 571 | #ifdef FREESCALE_MMCAU_CLASSIC |
wolfSSL | 15:117db924cf7c | 572 | /* MMCAU 1.4 library used with non-KSDK / classic MQX builds */ |
wolfSSL | 15:117db924cf7c | 573 | #include "cau_api.h" |
wolfSSL | 15:117db924cf7c | 574 | #else |
wolfSSL | 15:117db924cf7c | 575 | #include "fsl_mmcau.h" |
wolfSSL | 15:117db924cf7c | 576 | #endif |
wolfSSL | 15:117db924cf7c | 577 | |
wolfSSL | 15:117db924cf7c | 578 | static int wc_AesEncrypt(Aes* aes, const byte* inBlock, byte* outBlock) |
wolfSSL | 15:117db924cf7c | 579 | { |
wolfSSL | 15:117db924cf7c | 580 | int ret; |
wolfSSL | 15:117db924cf7c | 581 | |
wolfSSL | 15:117db924cf7c | 582 | #ifdef FREESCALE_MMCAU_CLASSIC |
wolfSSL | 15:117db924cf7c | 583 | if ((wolfssl_word)outBlock % WOLFSSL_MMCAU_ALIGNMENT) { |
wolfSSL | 15:117db924cf7c | 584 | WOLFSSL_MSG("Bad cau_aes_encrypt alignment"); |
wolfSSL | 15:117db924cf7c | 585 | return BAD_ALIGN_E; |
wolfSSL | 15:117db924cf7c | 586 | } |
wolfSSL | 15:117db924cf7c | 587 | #endif |
wolfSSL | 15:117db924cf7c | 588 | |
wolfSSL | 15:117db924cf7c | 589 | ret = wolfSSL_CryptHwMutexLock(); |
wolfSSL | 15:117db924cf7c | 590 | if(ret == 0) { |
wolfSSL | 15:117db924cf7c | 591 | #ifdef FREESCALE_MMCAU_CLASSIC |
wolfSSL | 15:117db924cf7c | 592 | cau_aes_encrypt(inBlock, (byte*)aes->key, aes->rounds, outBlock); |
wolfSSL | 15:117db924cf7c | 593 | #else |
wolfSSL | 15:117db924cf7c | 594 | MMCAU_AES_EncryptEcb(inBlock, (byte*)aes->key, aes->rounds, |
wolfSSL | 15:117db924cf7c | 595 | outBlock); |
wolfSSL | 15:117db924cf7c | 596 | #endif |
wolfSSL | 15:117db924cf7c | 597 | wolfSSL_CryptHwMutexUnLock(); |
wolfSSL | 15:117db924cf7c | 598 | } |
wolfSSL | 15:117db924cf7c | 599 | return ret; |
wolfSSL | 15:117db924cf7c | 600 | } |
wolfSSL | 15:117db924cf7c | 601 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 602 | static int wc_AesDecrypt(Aes* aes, const byte* inBlock, byte* outBlock) |
wolfSSL | 15:117db924cf7c | 603 | { |
wolfSSL | 15:117db924cf7c | 604 | int ret; |
wolfSSL | 15:117db924cf7c | 605 | |
wolfSSL | 15:117db924cf7c | 606 | #ifdef FREESCALE_MMCAU_CLASSIC |
wolfSSL | 15:117db924cf7c | 607 | if ((wolfssl_word)outBlock % WOLFSSL_MMCAU_ALIGNMENT) { |
wolfSSL | 15:117db924cf7c | 608 | WOLFSSL_MSG("Bad cau_aes_decrypt alignment"); |
wolfSSL | 15:117db924cf7c | 609 | return BAD_ALIGN_E; |
wolfSSL | 15:117db924cf7c | 610 | } |
wolfSSL | 15:117db924cf7c | 611 | #endif |
wolfSSL | 15:117db924cf7c | 612 | |
wolfSSL | 15:117db924cf7c | 613 | ret = wolfSSL_CryptHwMutexLock(); |
wolfSSL | 15:117db924cf7c | 614 | if(ret == 0) { |
wolfSSL | 15:117db924cf7c | 615 | #ifdef FREESCALE_MMCAU_CLASSIC |
wolfSSL | 15:117db924cf7c | 616 | cau_aes_decrypt(inBlock, (byte*)aes->key, aes->rounds, outBlock); |
wolfSSL | 15:117db924cf7c | 617 | #else |
wolfSSL | 15:117db924cf7c | 618 | MMCAU_AES_DecryptEcb(inBlock, (byte*)aes->key, aes->rounds, |
wolfSSL | 15:117db924cf7c | 619 | outBlock); |
wolfSSL | 15:117db924cf7c | 620 | #endif |
wolfSSL | 15:117db924cf7c | 621 | wolfSSL_CryptHwMutexUnLock(); |
wolfSSL | 15:117db924cf7c | 622 | } |
wolfSSL | 15:117db924cf7c | 623 | return ret; |
wolfSSL | 15:117db924cf7c | 624 | } |
wolfSSL | 15:117db924cf7c | 625 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 626 | |
wolfSSL | 15:117db924cf7c | 627 | #elif defined(WOLFSSL_PIC32MZ_CRYPT) |
wolfSSL | 15:117db924cf7c | 628 | |
wolfSSL | 15:117db924cf7c | 629 | #include <wolfssl/wolfcrypt/port/pic32/pic32mz-crypt.h> |
wolfSSL | 15:117db924cf7c | 630 | |
wolfSSL | 15:117db924cf7c | 631 | #if defined(HAVE_AESGCM) || defined(WOLFSSL_AES_DIRECT) |
wolfSSL | 15:117db924cf7c | 632 | static int wc_AesEncrypt(Aes* aes, const byte* inBlock, byte* outBlock) |
wolfSSL | 15:117db924cf7c | 633 | { |
wolfSSL | 15:117db924cf7c | 634 | return wc_Pic32AesCrypt(aes->key, aes->keylen, NULL, 0, |
wolfSSL | 15:117db924cf7c | 635 | outBlock, inBlock, AES_BLOCK_SIZE, |
wolfSSL | 15:117db924cf7c | 636 | PIC32_ENCRYPTION, PIC32_ALGO_AES, PIC32_CRYPTOALGO_RECB); |
wolfSSL | 15:117db924cf7c | 637 | } |
wolfSSL | 15:117db924cf7c | 638 | #endif |
wolfSSL | 15:117db924cf7c | 639 | |
wolfSSL | 15:117db924cf7c | 640 | #if defined(HAVE_AES_DECRYPT) && defined(WOLFSSL_AES_DIRECT) |
wolfSSL | 15:117db924cf7c | 641 | static int wc_AesDecrypt(Aes* aes, const byte* inBlock, byte* outBlock) |
wolfSSL | 15:117db924cf7c | 642 | { |
wolfSSL | 15:117db924cf7c | 643 | return wc_Pic32AesCrypt(aes->key, aes->keylen, NULL, 0, |
wolfSSL | 15:117db924cf7c | 644 | outBlock, inBlock, AES_BLOCK_SIZE, |
wolfSSL | 15:117db924cf7c | 645 | PIC32_DECRYPTION, PIC32_ALGO_AES, PIC32_CRYPTOALGO_RECB); |
wolfSSL | 15:117db924cf7c | 646 | } |
wolfSSL | 15:117db924cf7c | 647 | #endif |
wolfSSL | 15:117db924cf7c | 648 | |
wolfSSL | 15:117db924cf7c | 649 | #elif defined(WOLFSSL_NRF51_AES) |
wolfSSL | 15:117db924cf7c | 650 | /* Use built-in AES hardware - AES 128 ECB Encrypt Only */ |
wolfSSL | 15:117db924cf7c | 651 | #include "wolfssl/wolfcrypt/port/nrf51.h" |
wolfSSL | 15:117db924cf7c | 652 | |
wolfSSL | 15:117db924cf7c | 653 | static int wc_AesEncrypt(Aes* aes, const byte* inBlock, byte* outBlock) |
wolfSSL | 15:117db924cf7c | 654 | { |
wolfSSL | 15:117db924cf7c | 655 | return nrf51_aes_encrypt(inBlock, (byte*)aes->key, aes->rounds, outBlock); |
wolfSSL | 15:117db924cf7c | 656 | } |
wolfSSL | 15:117db924cf7c | 657 | |
wolfSSL | 15:117db924cf7c | 658 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 659 | #error nRF51 AES Hardware does not support decrypt |
wolfSSL | 15:117db924cf7c | 660 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 661 | |
wolfSSL | 15:117db924cf7c | 662 | |
wolfSSL | 15:117db924cf7c | 663 | #elif defined(WOLFSSL_AESNI) |
wolfSSL | 15:117db924cf7c | 664 | |
wolfSSL | 15:117db924cf7c | 665 | #define NEED_AES_TABLES |
wolfSSL | 15:117db924cf7c | 666 | |
wolfSSL | 15:117db924cf7c | 667 | /* Each platform needs to query info type 1 from cpuid to see if aesni is |
wolfSSL | 15:117db924cf7c | 668 | * supported. Also, let's setup a macro for proper linkage w/o ABI conflicts |
wolfSSL | 15:117db924cf7c | 669 | */ |
wolfSSL | 15:117db924cf7c | 670 | |
wolfSSL | 15:117db924cf7c | 671 | #ifndef AESNI_ALIGN |
wolfSSL | 15:117db924cf7c | 672 | #define AESNI_ALIGN 16 |
wolfSSL | 15:117db924cf7c | 673 | #endif |
wolfSSL | 15:117db924cf7c | 674 | |
wolfSSL | 15:117db924cf7c | 675 | #ifndef _MSC_VER |
wolfSSL | 15:117db924cf7c | 676 | #define XASM_LINK(f) asm(f) |
wolfSSL | 15:117db924cf7c | 677 | #else |
wolfSSL | 15:117db924cf7c | 678 | #define XASM_LINK(f) |
wolfSSL | 15:117db924cf7c | 679 | #endif /* _MSC_VER */ |
wolfSSL | 15:117db924cf7c | 680 | |
wolfSSL | 15:117db924cf7c | 681 | static int checkAESNI = 0; |
wolfSSL | 15:117db924cf7c | 682 | static int haveAESNI = 0; |
wolfSSL | 15:117db924cf7c | 683 | static word32 intel_flags = 0; |
wolfSSL | 15:117db924cf7c | 684 | |
wolfSSL | 15:117db924cf7c | 685 | static int Check_CPU_support_AES(void) |
wolfSSL | 15:117db924cf7c | 686 | { |
wolfSSL | 15:117db924cf7c | 687 | intel_flags = cpuid_get_flags(); |
wolfSSL | 15:117db924cf7c | 688 | |
wolfSSL | 15:117db924cf7c | 689 | return IS_INTEL_AESNI(intel_flags) != 0; |
wolfSSL | 15:117db924cf7c | 690 | } |
wolfSSL | 15:117db924cf7c | 691 | |
wolfSSL | 15:117db924cf7c | 692 | |
wolfSSL | 15:117db924cf7c | 693 | /* tell C compiler these are asm functions in case any mix up of ABI underscore |
wolfSSL | 15:117db924cf7c | 694 | prefix between clang/gcc/llvm etc */ |
wolfSSL | 15:117db924cf7c | 695 | #ifdef HAVE_AES_CBC |
wolfSSL | 15:117db924cf7c | 696 | void AES_CBC_encrypt(const unsigned char* in, unsigned char* out, |
wolfSSL | 15:117db924cf7c | 697 | unsigned char* ivec, unsigned long length, |
wolfSSL | 15:117db924cf7c | 698 | const unsigned char* KS, int nr) |
wolfSSL | 15:117db924cf7c | 699 | XASM_LINK("AES_CBC_encrypt"); |
wolfSSL | 15:117db924cf7c | 700 | |
wolfSSL | 15:117db924cf7c | 701 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 702 | #if defined(WOLFSSL_AESNI_BY4) |
wolfSSL | 15:117db924cf7c | 703 | void AES_CBC_decrypt_by4(const unsigned char* in, unsigned char* out, |
wolfSSL | 15:117db924cf7c | 704 | unsigned char* ivec, unsigned long length, |
wolfSSL | 15:117db924cf7c | 705 | const unsigned char* KS, int nr) |
wolfSSL | 15:117db924cf7c | 706 | XASM_LINK("AES_CBC_decrypt_by4"); |
wolfSSL | 15:117db924cf7c | 707 | #elif defined(WOLFSSL_AESNI_BY6) |
wolfSSL | 15:117db924cf7c | 708 | void AES_CBC_decrypt_by6(const unsigned char* in, unsigned char* out, |
wolfSSL | 15:117db924cf7c | 709 | unsigned char* ivec, unsigned long length, |
wolfSSL | 15:117db924cf7c | 710 | const unsigned char* KS, int nr) |
wolfSSL | 15:117db924cf7c | 711 | XASM_LINK("AES_CBC_decrypt_by6"); |
wolfSSL | 15:117db924cf7c | 712 | #else /* WOLFSSL_AESNI_BYx */ |
wolfSSL | 15:117db924cf7c | 713 | void AES_CBC_decrypt_by8(const unsigned char* in, unsigned char* out, |
wolfSSL | 15:117db924cf7c | 714 | unsigned char* ivec, unsigned long length, |
wolfSSL | 15:117db924cf7c | 715 | const unsigned char* KS, int nr) |
wolfSSL | 15:117db924cf7c | 716 | XASM_LINK("AES_CBC_decrypt_by8"); |
wolfSSL | 15:117db924cf7c | 717 | #endif /* WOLFSSL_AESNI_BYx */ |
wolfSSL | 15:117db924cf7c | 718 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 719 | #endif /* HAVE_AES_CBC */ |
wolfSSL | 15:117db924cf7c | 720 | |
wolfSSL | 15:117db924cf7c | 721 | void AES_ECB_encrypt(const unsigned char* in, unsigned char* out, |
wolfSSL | 15:117db924cf7c | 722 | unsigned long length, const unsigned char* KS, int nr) |
wolfSSL | 15:117db924cf7c | 723 | XASM_LINK("AES_ECB_encrypt"); |
wolfSSL | 15:117db924cf7c | 724 | |
wolfSSL | 15:117db924cf7c | 725 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 726 | void AES_ECB_decrypt(const unsigned char* in, unsigned char* out, |
wolfSSL | 15:117db924cf7c | 727 | unsigned long length, const unsigned char* KS, int nr) |
wolfSSL | 15:117db924cf7c | 728 | XASM_LINK("AES_ECB_decrypt"); |
wolfSSL | 15:117db924cf7c | 729 | #endif |
wolfSSL | 15:117db924cf7c | 730 | |
wolfSSL | 15:117db924cf7c | 731 | void AES_128_Key_Expansion(const unsigned char* userkey, |
wolfSSL | 15:117db924cf7c | 732 | unsigned char* key_schedule) |
wolfSSL | 15:117db924cf7c | 733 | XASM_LINK("AES_128_Key_Expansion"); |
wolfSSL | 15:117db924cf7c | 734 | |
wolfSSL | 15:117db924cf7c | 735 | void AES_192_Key_Expansion(const unsigned char* userkey, |
wolfSSL | 15:117db924cf7c | 736 | unsigned char* key_schedule) |
wolfSSL | 15:117db924cf7c | 737 | XASM_LINK("AES_192_Key_Expansion"); |
wolfSSL | 15:117db924cf7c | 738 | |
wolfSSL | 15:117db924cf7c | 739 | void AES_256_Key_Expansion(const unsigned char* userkey, |
wolfSSL | 15:117db924cf7c | 740 | unsigned char* key_schedule) |
wolfSSL | 15:117db924cf7c | 741 | XASM_LINK("AES_256_Key_Expansion"); |
wolfSSL | 15:117db924cf7c | 742 | |
wolfSSL | 15:117db924cf7c | 743 | |
wolfSSL | 15:117db924cf7c | 744 | static int AES_set_encrypt_key(const unsigned char *userKey, const int bits, |
wolfSSL | 15:117db924cf7c | 745 | Aes* aes) |
wolfSSL | 15:117db924cf7c | 746 | { |
wolfSSL | 15:117db924cf7c | 747 | int ret; |
wolfSSL | 15:117db924cf7c | 748 | |
wolfSSL | 15:117db924cf7c | 749 | if (!userKey || !aes) |
wolfSSL | 15:117db924cf7c | 750 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 751 | |
wolfSSL | 15:117db924cf7c | 752 | switch (bits) { |
wolfSSL | 15:117db924cf7c | 753 | case 128: |
wolfSSL | 15:117db924cf7c | 754 | AES_128_Key_Expansion (userKey,(byte*)aes->key); aes->rounds = 10; |
wolfSSL | 15:117db924cf7c | 755 | return 0; |
wolfSSL | 15:117db924cf7c | 756 | case 192: |
wolfSSL | 15:117db924cf7c | 757 | AES_192_Key_Expansion (userKey,(byte*)aes->key); aes->rounds = 12; |
wolfSSL | 15:117db924cf7c | 758 | return 0; |
wolfSSL | 15:117db924cf7c | 759 | case 256: |
wolfSSL | 15:117db924cf7c | 760 | AES_256_Key_Expansion (userKey,(byte*)aes->key); aes->rounds = 14; |
wolfSSL | 15:117db924cf7c | 761 | return 0; |
wolfSSL | 15:117db924cf7c | 762 | default: |
wolfSSL | 15:117db924cf7c | 763 | ret = BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 764 | } |
wolfSSL | 15:117db924cf7c | 765 | |
wolfSSL | 15:117db924cf7c | 766 | return ret; |
wolfSSL | 15:117db924cf7c | 767 | } |
wolfSSL | 15:117db924cf7c | 768 | |
wolfSSL | 15:117db924cf7c | 769 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 770 | static int AES_set_decrypt_key(const unsigned char* userKey, |
wolfSSL | 15:117db924cf7c | 771 | const int bits, Aes* aes) |
wolfSSL | 15:117db924cf7c | 772 | { |
wolfSSL | 15:117db924cf7c | 773 | int nr; |
wolfSSL | 15:117db924cf7c | 774 | Aes temp_key; |
wolfSSL | 15:117db924cf7c | 775 | __m128i *Key_Schedule = (__m128i*)aes->key; |
wolfSSL | 15:117db924cf7c | 776 | __m128i *Temp_Key_Schedule = (__m128i*)temp_key.key; |
wolfSSL | 15:117db924cf7c | 777 | |
wolfSSL | 15:117db924cf7c | 778 | if (!userKey || !aes) |
wolfSSL | 15:117db924cf7c | 779 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 780 | |
wolfSSL | 15:117db924cf7c | 781 | if (AES_set_encrypt_key(userKey,bits,&temp_key) == BAD_FUNC_ARG) |
wolfSSL | 15:117db924cf7c | 782 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 783 | |
wolfSSL | 15:117db924cf7c | 784 | nr = temp_key.rounds; |
wolfSSL | 15:117db924cf7c | 785 | aes->rounds = nr; |
wolfSSL | 15:117db924cf7c | 786 | |
wolfSSL | 15:117db924cf7c | 787 | Key_Schedule[nr] = Temp_Key_Schedule[0]; |
wolfSSL | 15:117db924cf7c | 788 | Key_Schedule[nr-1] = _mm_aesimc_si128(Temp_Key_Schedule[1]); |
wolfSSL | 15:117db924cf7c | 789 | Key_Schedule[nr-2] = _mm_aesimc_si128(Temp_Key_Schedule[2]); |
wolfSSL | 15:117db924cf7c | 790 | Key_Schedule[nr-3] = _mm_aesimc_si128(Temp_Key_Schedule[3]); |
wolfSSL | 15:117db924cf7c | 791 | Key_Schedule[nr-4] = _mm_aesimc_si128(Temp_Key_Schedule[4]); |
wolfSSL | 15:117db924cf7c | 792 | Key_Schedule[nr-5] = _mm_aesimc_si128(Temp_Key_Schedule[5]); |
wolfSSL | 15:117db924cf7c | 793 | Key_Schedule[nr-6] = _mm_aesimc_si128(Temp_Key_Schedule[6]); |
wolfSSL | 15:117db924cf7c | 794 | Key_Schedule[nr-7] = _mm_aesimc_si128(Temp_Key_Schedule[7]); |
wolfSSL | 15:117db924cf7c | 795 | Key_Schedule[nr-8] = _mm_aesimc_si128(Temp_Key_Schedule[8]); |
wolfSSL | 15:117db924cf7c | 796 | Key_Schedule[nr-9] = _mm_aesimc_si128(Temp_Key_Schedule[9]); |
wolfSSL | 15:117db924cf7c | 797 | |
wolfSSL | 15:117db924cf7c | 798 | if (nr>10) { |
wolfSSL | 15:117db924cf7c | 799 | Key_Schedule[nr-10] = _mm_aesimc_si128(Temp_Key_Schedule[10]); |
wolfSSL | 15:117db924cf7c | 800 | Key_Schedule[nr-11] = _mm_aesimc_si128(Temp_Key_Schedule[11]); |
wolfSSL | 15:117db924cf7c | 801 | } |
wolfSSL | 15:117db924cf7c | 802 | |
wolfSSL | 15:117db924cf7c | 803 | if (nr>12) { |
wolfSSL | 15:117db924cf7c | 804 | Key_Schedule[nr-12] = _mm_aesimc_si128(Temp_Key_Schedule[12]); |
wolfSSL | 15:117db924cf7c | 805 | Key_Schedule[nr-13] = _mm_aesimc_si128(Temp_Key_Schedule[13]); |
wolfSSL | 15:117db924cf7c | 806 | } |
wolfSSL | 15:117db924cf7c | 807 | |
wolfSSL | 15:117db924cf7c | 808 | Key_Schedule[0] = Temp_Key_Schedule[nr]; |
wolfSSL | 15:117db924cf7c | 809 | |
wolfSSL | 15:117db924cf7c | 810 | return 0; |
wolfSSL | 15:117db924cf7c | 811 | } |
wolfSSL | 15:117db924cf7c | 812 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 813 | |
wolfSSL | 15:117db924cf7c | 814 | #elif defined(WOLFSSL_IMX6_CAAM) && !defined(NO_IMX6_CAAM_AES) |
wolfSSL | 15:117db924cf7c | 815 | static int wc_AesEncrypt(Aes* aes, const byte* inBlock, byte* outBlock) |
wolfSSL | 15:117db924cf7c | 816 | { |
wolfSSL | 15:117db924cf7c | 817 | wc_AesEncryptDirect(aes, outBlock, inBlock); |
wolfSSL | 15:117db924cf7c | 818 | return 0; |
wolfSSL | 15:117db924cf7c | 819 | } |
wolfSSL | 15:117db924cf7c | 820 | #else |
wolfSSL | 15:117db924cf7c | 821 | |
wolfSSL | 15:117db924cf7c | 822 | /* using wolfCrypt software AES implementation */ |
wolfSSL | 15:117db924cf7c | 823 | #define NEED_AES_TABLES |
wolfSSL | 15:117db924cf7c | 824 | #endif |
wolfSSL | 15:117db924cf7c | 825 | |
wolfSSL | 15:117db924cf7c | 826 | |
wolfSSL | 15:117db924cf7c | 827 | |
wolfSSL | 15:117db924cf7c | 828 | #ifdef NEED_AES_TABLES |
wolfSSL | 15:117db924cf7c | 829 | |
wolfSSL | 15:117db924cf7c | 830 | static const word32 rcon[] = { |
wolfSSL | 15:117db924cf7c | 831 | 0x01000000, 0x02000000, 0x04000000, 0x08000000, |
wolfSSL | 15:117db924cf7c | 832 | 0x10000000, 0x20000000, 0x40000000, 0x80000000, |
wolfSSL | 15:117db924cf7c | 833 | 0x1B000000, 0x36000000, |
wolfSSL | 15:117db924cf7c | 834 | /* for 128-bit blocks, Rijndael never uses more than 10 rcon values */ |
wolfSSL | 15:117db924cf7c | 835 | }; |
wolfSSL | 15:117db924cf7c | 836 | |
wolfSSL | 15:117db924cf7c | 837 | static const word32 Te[4][256] = { |
wolfSSL | 15:117db924cf7c | 838 | { |
wolfSSL | 15:117db924cf7c | 839 | 0xc66363a5U, 0xf87c7c84U, 0xee777799U, 0xf67b7b8dU, |
wolfSSL | 15:117db924cf7c | 840 | 0xfff2f20dU, 0xd66b6bbdU, 0xde6f6fb1U, 0x91c5c554U, |
wolfSSL | 15:117db924cf7c | 841 | 0x60303050U, 0x02010103U, 0xce6767a9U, 0x562b2b7dU, |
wolfSSL | 15:117db924cf7c | 842 | 0xe7fefe19U, 0xb5d7d762U, 0x4dababe6U, 0xec76769aU, |
wolfSSL | 15:117db924cf7c | 843 | 0x8fcaca45U, 0x1f82829dU, 0x89c9c940U, 0xfa7d7d87U, |
wolfSSL | 15:117db924cf7c | 844 | 0xeffafa15U, 0xb25959ebU, 0x8e4747c9U, 0xfbf0f00bU, |
wolfSSL | 15:117db924cf7c | 845 | 0x41adadecU, 0xb3d4d467U, 0x5fa2a2fdU, 0x45afafeaU, |
wolfSSL | 15:117db924cf7c | 846 | 0x239c9cbfU, 0x53a4a4f7U, 0xe4727296U, 0x9bc0c05bU, |
wolfSSL | 15:117db924cf7c | 847 | 0x75b7b7c2U, 0xe1fdfd1cU, 0x3d9393aeU, 0x4c26266aU, |
wolfSSL | 15:117db924cf7c | 848 | 0x6c36365aU, 0x7e3f3f41U, 0xf5f7f702U, 0x83cccc4fU, |
wolfSSL | 15:117db924cf7c | 849 | 0x6834345cU, 0x51a5a5f4U, 0xd1e5e534U, 0xf9f1f108U, |
wolfSSL | 15:117db924cf7c | 850 | 0xe2717193U, 0xabd8d873U, 0x62313153U, 0x2a15153fU, |
wolfSSL | 15:117db924cf7c | 851 | 0x0804040cU, 0x95c7c752U, 0x46232365U, 0x9dc3c35eU, |
wolfSSL | 15:117db924cf7c | 852 | 0x30181828U, 0x379696a1U, 0x0a05050fU, 0x2f9a9ab5U, |
wolfSSL | 15:117db924cf7c | 853 | 0x0e070709U, 0x24121236U, 0x1b80809bU, 0xdfe2e23dU, |
wolfSSL | 15:117db924cf7c | 854 | 0xcdebeb26U, 0x4e272769U, 0x7fb2b2cdU, 0xea75759fU, |
wolfSSL | 15:117db924cf7c | 855 | 0x1209091bU, 0x1d83839eU, 0x582c2c74U, 0x341a1a2eU, |
wolfSSL | 15:117db924cf7c | 856 | 0x361b1b2dU, 0xdc6e6eb2U, 0xb45a5aeeU, 0x5ba0a0fbU, |
wolfSSL | 15:117db924cf7c | 857 | 0xa45252f6U, 0x763b3b4dU, 0xb7d6d661U, 0x7db3b3ceU, |
wolfSSL | 15:117db924cf7c | 858 | 0x5229297bU, 0xdde3e33eU, 0x5e2f2f71U, 0x13848497U, |
wolfSSL | 15:117db924cf7c | 859 | 0xa65353f5U, 0xb9d1d168U, 0x00000000U, 0xc1eded2cU, |
wolfSSL | 15:117db924cf7c | 860 | 0x40202060U, 0xe3fcfc1fU, 0x79b1b1c8U, 0xb65b5bedU, |
wolfSSL | 15:117db924cf7c | 861 | 0xd46a6abeU, 0x8dcbcb46U, 0x67bebed9U, 0x7239394bU, |
wolfSSL | 15:117db924cf7c | 862 | 0x944a4adeU, 0x984c4cd4U, 0xb05858e8U, 0x85cfcf4aU, |
wolfSSL | 15:117db924cf7c | 863 | 0xbbd0d06bU, 0xc5efef2aU, 0x4faaaae5U, 0xedfbfb16U, |
wolfSSL | 15:117db924cf7c | 864 | 0x864343c5U, 0x9a4d4dd7U, 0x66333355U, 0x11858594U, |
wolfSSL | 15:117db924cf7c | 865 | 0x8a4545cfU, 0xe9f9f910U, 0x04020206U, 0xfe7f7f81U, |
wolfSSL | 15:117db924cf7c | 866 | 0xa05050f0U, 0x783c3c44U, 0x259f9fbaU, 0x4ba8a8e3U, |
wolfSSL | 15:117db924cf7c | 867 | 0xa25151f3U, 0x5da3a3feU, 0x804040c0U, 0x058f8f8aU, |
wolfSSL | 15:117db924cf7c | 868 | 0x3f9292adU, 0x219d9dbcU, 0x70383848U, 0xf1f5f504U, |
wolfSSL | 15:117db924cf7c | 869 | 0x63bcbcdfU, 0x77b6b6c1U, 0xafdada75U, 0x42212163U, |
wolfSSL | 15:117db924cf7c | 870 | 0x20101030U, 0xe5ffff1aU, 0xfdf3f30eU, 0xbfd2d26dU, |
wolfSSL | 15:117db924cf7c | 871 | 0x81cdcd4cU, 0x180c0c14U, 0x26131335U, 0xc3ecec2fU, |
wolfSSL | 15:117db924cf7c | 872 | 0xbe5f5fe1U, 0x359797a2U, 0x884444ccU, 0x2e171739U, |
wolfSSL | 15:117db924cf7c | 873 | 0x93c4c457U, 0x55a7a7f2U, 0xfc7e7e82U, 0x7a3d3d47U, |
wolfSSL | 15:117db924cf7c | 874 | 0xc86464acU, 0xba5d5de7U, 0x3219192bU, 0xe6737395U, |
wolfSSL | 15:117db924cf7c | 875 | 0xc06060a0U, 0x19818198U, 0x9e4f4fd1U, 0xa3dcdc7fU, |
wolfSSL | 15:117db924cf7c | 876 | 0x44222266U, 0x542a2a7eU, 0x3b9090abU, 0x0b888883U, |
wolfSSL | 15:117db924cf7c | 877 | 0x8c4646caU, 0xc7eeee29U, 0x6bb8b8d3U, 0x2814143cU, |
wolfSSL | 15:117db924cf7c | 878 | 0xa7dede79U, 0xbc5e5ee2U, 0x160b0b1dU, 0xaddbdb76U, |
wolfSSL | 15:117db924cf7c | 879 | 0xdbe0e03bU, 0x64323256U, 0x743a3a4eU, 0x140a0a1eU, |
wolfSSL | 15:117db924cf7c | 880 | 0x924949dbU, 0x0c06060aU, 0x4824246cU, 0xb85c5ce4U, |
wolfSSL | 15:117db924cf7c | 881 | 0x9fc2c25dU, 0xbdd3d36eU, 0x43acacefU, 0xc46262a6U, |
wolfSSL | 15:117db924cf7c | 882 | 0x399191a8U, 0x319595a4U, 0xd3e4e437U, 0xf279798bU, |
wolfSSL | 15:117db924cf7c | 883 | 0xd5e7e732U, 0x8bc8c843U, 0x6e373759U, 0xda6d6db7U, |
wolfSSL | 15:117db924cf7c | 884 | 0x018d8d8cU, 0xb1d5d564U, 0x9c4e4ed2U, 0x49a9a9e0U, |
wolfSSL | 15:117db924cf7c | 885 | 0xd86c6cb4U, 0xac5656faU, 0xf3f4f407U, 0xcfeaea25U, |
wolfSSL | 15:117db924cf7c | 886 | 0xca6565afU, 0xf47a7a8eU, 0x47aeaee9U, 0x10080818U, |
wolfSSL | 15:117db924cf7c | 887 | 0x6fbabad5U, 0xf0787888U, 0x4a25256fU, 0x5c2e2e72U, |
wolfSSL | 15:117db924cf7c | 888 | 0x381c1c24U, 0x57a6a6f1U, 0x73b4b4c7U, 0x97c6c651U, |
wolfSSL | 15:117db924cf7c | 889 | 0xcbe8e823U, 0xa1dddd7cU, 0xe874749cU, 0x3e1f1f21U, |
wolfSSL | 15:117db924cf7c | 890 | 0x964b4bddU, 0x61bdbddcU, 0x0d8b8b86U, 0x0f8a8a85U, |
wolfSSL | 15:117db924cf7c | 891 | 0xe0707090U, 0x7c3e3e42U, 0x71b5b5c4U, 0xcc6666aaU, |
wolfSSL | 15:117db924cf7c | 892 | 0x904848d8U, 0x06030305U, 0xf7f6f601U, 0x1c0e0e12U, |
wolfSSL | 15:117db924cf7c | 893 | 0xc26161a3U, 0x6a35355fU, 0xae5757f9U, 0x69b9b9d0U, |
wolfSSL | 15:117db924cf7c | 894 | 0x17868691U, 0x99c1c158U, 0x3a1d1d27U, 0x279e9eb9U, |
wolfSSL | 15:117db924cf7c | 895 | 0xd9e1e138U, 0xebf8f813U, 0x2b9898b3U, 0x22111133U, |
wolfSSL | 15:117db924cf7c | 896 | 0xd26969bbU, 0xa9d9d970U, 0x078e8e89U, 0x339494a7U, |
wolfSSL | 15:117db924cf7c | 897 | 0x2d9b9bb6U, 0x3c1e1e22U, 0x15878792U, 0xc9e9e920U, |
wolfSSL | 15:117db924cf7c | 898 | 0x87cece49U, 0xaa5555ffU, 0x50282878U, 0xa5dfdf7aU, |
wolfSSL | 15:117db924cf7c | 899 | 0x038c8c8fU, 0x59a1a1f8U, 0x09898980U, 0x1a0d0d17U, |
wolfSSL | 15:117db924cf7c | 900 | 0x65bfbfdaU, 0xd7e6e631U, 0x844242c6U, 0xd06868b8U, |
wolfSSL | 15:117db924cf7c | 901 | 0x824141c3U, 0x299999b0U, 0x5a2d2d77U, 0x1e0f0f11U, |
wolfSSL | 15:117db924cf7c | 902 | 0x7bb0b0cbU, 0xa85454fcU, 0x6dbbbbd6U, 0x2c16163aU, |
wolfSSL | 15:117db924cf7c | 903 | }, |
wolfSSL | 15:117db924cf7c | 904 | { |
wolfSSL | 15:117db924cf7c | 905 | 0xa5c66363U, 0x84f87c7cU, 0x99ee7777U, 0x8df67b7bU, |
wolfSSL | 15:117db924cf7c | 906 | 0x0dfff2f2U, 0xbdd66b6bU, 0xb1de6f6fU, 0x5491c5c5U, |
wolfSSL | 15:117db924cf7c | 907 | 0x50603030U, 0x03020101U, 0xa9ce6767U, 0x7d562b2bU, |
wolfSSL | 15:117db924cf7c | 908 | 0x19e7fefeU, 0x62b5d7d7U, 0xe64dababU, 0x9aec7676U, |
wolfSSL | 15:117db924cf7c | 909 | 0x458fcacaU, 0x9d1f8282U, 0x4089c9c9U, 0x87fa7d7dU, |
wolfSSL | 15:117db924cf7c | 910 | 0x15effafaU, 0xebb25959U, 0xc98e4747U, 0x0bfbf0f0U, |
wolfSSL | 15:117db924cf7c | 911 | 0xec41adadU, 0x67b3d4d4U, 0xfd5fa2a2U, 0xea45afafU, |
wolfSSL | 15:117db924cf7c | 912 | 0xbf239c9cU, 0xf753a4a4U, 0x96e47272U, 0x5b9bc0c0U, |
wolfSSL | 15:117db924cf7c | 913 | 0xc275b7b7U, 0x1ce1fdfdU, 0xae3d9393U, 0x6a4c2626U, |
wolfSSL | 15:117db924cf7c | 914 | 0x5a6c3636U, 0x417e3f3fU, 0x02f5f7f7U, 0x4f83ccccU, |
wolfSSL | 15:117db924cf7c | 915 | 0x5c683434U, 0xf451a5a5U, 0x34d1e5e5U, 0x08f9f1f1U, |
wolfSSL | 15:117db924cf7c | 916 | 0x93e27171U, 0x73abd8d8U, 0x53623131U, 0x3f2a1515U, |
wolfSSL | 15:117db924cf7c | 917 | 0x0c080404U, 0x5295c7c7U, 0x65462323U, 0x5e9dc3c3U, |
wolfSSL | 15:117db924cf7c | 918 | 0x28301818U, 0xa1379696U, 0x0f0a0505U, 0xb52f9a9aU, |
wolfSSL | 15:117db924cf7c | 919 | 0x090e0707U, 0x36241212U, 0x9b1b8080U, 0x3ddfe2e2U, |
wolfSSL | 15:117db924cf7c | 920 | 0x26cdebebU, 0x694e2727U, 0xcd7fb2b2U, 0x9fea7575U, |
wolfSSL | 15:117db924cf7c | 921 | 0x1b120909U, 0x9e1d8383U, 0x74582c2cU, 0x2e341a1aU, |
wolfSSL | 15:117db924cf7c | 922 | 0x2d361b1bU, 0xb2dc6e6eU, 0xeeb45a5aU, 0xfb5ba0a0U, |
wolfSSL | 15:117db924cf7c | 923 | 0xf6a45252U, 0x4d763b3bU, 0x61b7d6d6U, 0xce7db3b3U, |
wolfSSL | 15:117db924cf7c | 924 | 0x7b522929U, 0x3edde3e3U, 0x715e2f2fU, 0x97138484U, |
wolfSSL | 15:117db924cf7c | 925 | 0xf5a65353U, 0x68b9d1d1U, 0x00000000U, 0x2cc1ededU, |
wolfSSL | 15:117db924cf7c | 926 | 0x60402020U, 0x1fe3fcfcU, 0xc879b1b1U, 0xedb65b5bU, |
wolfSSL | 15:117db924cf7c | 927 | 0xbed46a6aU, 0x468dcbcbU, 0xd967bebeU, 0x4b723939U, |
wolfSSL | 15:117db924cf7c | 928 | 0xde944a4aU, 0xd4984c4cU, 0xe8b05858U, 0x4a85cfcfU, |
wolfSSL | 15:117db924cf7c | 929 | 0x6bbbd0d0U, 0x2ac5efefU, 0xe54faaaaU, 0x16edfbfbU, |
wolfSSL | 15:117db924cf7c | 930 | 0xc5864343U, 0xd79a4d4dU, 0x55663333U, 0x94118585U, |
wolfSSL | 15:117db924cf7c | 931 | 0xcf8a4545U, 0x10e9f9f9U, 0x06040202U, 0x81fe7f7fU, |
wolfSSL | 15:117db924cf7c | 932 | 0xf0a05050U, 0x44783c3cU, 0xba259f9fU, 0xe34ba8a8U, |
wolfSSL | 15:117db924cf7c | 933 | 0xf3a25151U, 0xfe5da3a3U, 0xc0804040U, 0x8a058f8fU, |
wolfSSL | 15:117db924cf7c | 934 | 0xad3f9292U, 0xbc219d9dU, 0x48703838U, 0x04f1f5f5U, |
wolfSSL | 15:117db924cf7c | 935 | 0xdf63bcbcU, 0xc177b6b6U, 0x75afdadaU, 0x63422121U, |
wolfSSL | 15:117db924cf7c | 936 | 0x30201010U, 0x1ae5ffffU, 0x0efdf3f3U, 0x6dbfd2d2U, |
wolfSSL | 15:117db924cf7c | 937 | 0x4c81cdcdU, 0x14180c0cU, 0x35261313U, 0x2fc3ececU, |
wolfSSL | 15:117db924cf7c | 938 | 0xe1be5f5fU, 0xa2359797U, 0xcc884444U, 0x392e1717U, |
wolfSSL | 15:117db924cf7c | 939 | 0x5793c4c4U, 0xf255a7a7U, 0x82fc7e7eU, 0x477a3d3dU, |
wolfSSL | 15:117db924cf7c | 940 | 0xacc86464U, 0xe7ba5d5dU, 0x2b321919U, 0x95e67373U, |
wolfSSL | 15:117db924cf7c | 941 | 0xa0c06060U, 0x98198181U, 0xd19e4f4fU, 0x7fa3dcdcU, |
wolfSSL | 15:117db924cf7c | 942 | 0x66442222U, 0x7e542a2aU, 0xab3b9090U, 0x830b8888U, |
wolfSSL | 15:117db924cf7c | 943 | 0xca8c4646U, 0x29c7eeeeU, 0xd36bb8b8U, 0x3c281414U, |
wolfSSL | 15:117db924cf7c | 944 | 0x79a7dedeU, 0xe2bc5e5eU, 0x1d160b0bU, 0x76addbdbU, |
wolfSSL | 15:117db924cf7c | 945 | 0x3bdbe0e0U, 0x56643232U, 0x4e743a3aU, 0x1e140a0aU, |
wolfSSL | 15:117db924cf7c | 946 | 0xdb924949U, 0x0a0c0606U, 0x6c482424U, 0xe4b85c5cU, |
wolfSSL | 15:117db924cf7c | 947 | 0x5d9fc2c2U, 0x6ebdd3d3U, 0xef43acacU, 0xa6c46262U, |
wolfSSL | 15:117db924cf7c | 948 | 0xa8399191U, 0xa4319595U, 0x37d3e4e4U, 0x8bf27979U, |
wolfSSL | 15:117db924cf7c | 949 | 0x32d5e7e7U, 0x438bc8c8U, 0x596e3737U, 0xb7da6d6dU, |
wolfSSL | 15:117db924cf7c | 950 | 0x8c018d8dU, 0x64b1d5d5U, 0xd29c4e4eU, 0xe049a9a9U, |
wolfSSL | 15:117db924cf7c | 951 | 0xb4d86c6cU, 0xfaac5656U, 0x07f3f4f4U, 0x25cfeaeaU, |
wolfSSL | 15:117db924cf7c | 952 | 0xafca6565U, 0x8ef47a7aU, 0xe947aeaeU, 0x18100808U, |
wolfSSL | 15:117db924cf7c | 953 | 0xd56fbabaU, 0x88f07878U, 0x6f4a2525U, 0x725c2e2eU, |
wolfSSL | 15:117db924cf7c | 954 | 0x24381c1cU, 0xf157a6a6U, 0xc773b4b4U, 0x5197c6c6U, |
wolfSSL | 15:117db924cf7c | 955 | 0x23cbe8e8U, 0x7ca1ddddU, 0x9ce87474U, 0x213e1f1fU, |
wolfSSL | 15:117db924cf7c | 956 | 0xdd964b4bU, 0xdc61bdbdU, 0x860d8b8bU, 0x850f8a8aU, |
wolfSSL | 15:117db924cf7c | 957 | 0x90e07070U, 0x427c3e3eU, 0xc471b5b5U, 0xaacc6666U, |
wolfSSL | 15:117db924cf7c | 958 | 0xd8904848U, 0x05060303U, 0x01f7f6f6U, 0x121c0e0eU, |
wolfSSL | 15:117db924cf7c | 959 | 0xa3c26161U, 0x5f6a3535U, 0xf9ae5757U, 0xd069b9b9U, |
wolfSSL | 15:117db924cf7c | 960 | 0x91178686U, 0x5899c1c1U, 0x273a1d1dU, 0xb9279e9eU, |
wolfSSL | 15:117db924cf7c | 961 | 0x38d9e1e1U, 0x13ebf8f8U, 0xb32b9898U, 0x33221111U, |
wolfSSL | 15:117db924cf7c | 962 | 0xbbd26969U, 0x70a9d9d9U, 0x89078e8eU, 0xa7339494U, |
wolfSSL | 15:117db924cf7c | 963 | 0xb62d9b9bU, 0x223c1e1eU, 0x92158787U, 0x20c9e9e9U, |
wolfSSL | 15:117db924cf7c | 964 | 0x4987ceceU, 0xffaa5555U, 0x78502828U, 0x7aa5dfdfU, |
wolfSSL | 15:117db924cf7c | 965 | 0x8f038c8cU, 0xf859a1a1U, 0x80098989U, 0x171a0d0dU, |
wolfSSL | 15:117db924cf7c | 966 | 0xda65bfbfU, 0x31d7e6e6U, 0xc6844242U, 0xb8d06868U, |
wolfSSL | 15:117db924cf7c | 967 | 0xc3824141U, 0xb0299999U, 0x775a2d2dU, 0x111e0f0fU, |
wolfSSL | 15:117db924cf7c | 968 | 0xcb7bb0b0U, 0xfca85454U, 0xd66dbbbbU, 0x3a2c1616U, |
wolfSSL | 15:117db924cf7c | 969 | }, |
wolfSSL | 15:117db924cf7c | 970 | { |
wolfSSL | 15:117db924cf7c | 971 | 0x63a5c663U, 0x7c84f87cU, 0x7799ee77U, 0x7b8df67bU, |
wolfSSL | 15:117db924cf7c | 972 | 0xf20dfff2U, 0x6bbdd66bU, 0x6fb1de6fU, 0xc55491c5U, |
wolfSSL | 15:117db924cf7c | 973 | 0x30506030U, 0x01030201U, 0x67a9ce67U, 0x2b7d562bU, |
wolfSSL | 15:117db924cf7c | 974 | 0xfe19e7feU, 0xd762b5d7U, 0xabe64dabU, 0x769aec76U, |
wolfSSL | 15:117db924cf7c | 975 | 0xca458fcaU, 0x829d1f82U, 0xc94089c9U, 0x7d87fa7dU, |
wolfSSL | 15:117db924cf7c | 976 | 0xfa15effaU, 0x59ebb259U, 0x47c98e47U, 0xf00bfbf0U, |
wolfSSL | 15:117db924cf7c | 977 | 0xadec41adU, 0xd467b3d4U, 0xa2fd5fa2U, 0xafea45afU, |
wolfSSL | 15:117db924cf7c | 978 | 0x9cbf239cU, 0xa4f753a4U, 0x7296e472U, 0xc05b9bc0U, |
wolfSSL | 15:117db924cf7c | 979 | 0xb7c275b7U, 0xfd1ce1fdU, 0x93ae3d93U, 0x266a4c26U, |
wolfSSL | 15:117db924cf7c | 980 | 0x365a6c36U, 0x3f417e3fU, 0xf702f5f7U, 0xcc4f83ccU, |
wolfSSL | 15:117db924cf7c | 981 | 0x345c6834U, 0xa5f451a5U, 0xe534d1e5U, 0xf108f9f1U, |
wolfSSL | 15:117db924cf7c | 982 | 0x7193e271U, 0xd873abd8U, 0x31536231U, 0x153f2a15U, |
wolfSSL | 15:117db924cf7c | 983 | 0x040c0804U, 0xc75295c7U, 0x23654623U, 0xc35e9dc3U, |
wolfSSL | 15:117db924cf7c | 984 | 0x18283018U, 0x96a13796U, 0x050f0a05U, 0x9ab52f9aU, |
wolfSSL | 15:117db924cf7c | 985 | 0x07090e07U, 0x12362412U, 0x809b1b80U, 0xe23ddfe2U, |
wolfSSL | 15:117db924cf7c | 986 | 0xeb26cdebU, 0x27694e27U, 0xb2cd7fb2U, 0x759fea75U, |
wolfSSL | 15:117db924cf7c | 987 | 0x091b1209U, 0x839e1d83U, 0x2c74582cU, 0x1a2e341aU, |
wolfSSL | 15:117db924cf7c | 988 | 0x1b2d361bU, 0x6eb2dc6eU, 0x5aeeb45aU, 0xa0fb5ba0U, |
wolfSSL | 15:117db924cf7c | 989 | 0x52f6a452U, 0x3b4d763bU, 0xd661b7d6U, 0xb3ce7db3U, |
wolfSSL | 15:117db924cf7c | 990 | 0x297b5229U, 0xe33edde3U, 0x2f715e2fU, 0x84971384U, |
wolfSSL | 15:117db924cf7c | 991 | 0x53f5a653U, 0xd168b9d1U, 0x00000000U, 0xed2cc1edU, |
wolfSSL | 15:117db924cf7c | 992 | 0x20604020U, 0xfc1fe3fcU, 0xb1c879b1U, 0x5bedb65bU, |
wolfSSL | 15:117db924cf7c | 993 | 0x6abed46aU, 0xcb468dcbU, 0xbed967beU, 0x394b7239U, |
wolfSSL | 15:117db924cf7c | 994 | 0x4ade944aU, 0x4cd4984cU, 0x58e8b058U, 0xcf4a85cfU, |
wolfSSL | 15:117db924cf7c | 995 | 0xd06bbbd0U, 0xef2ac5efU, 0xaae54faaU, 0xfb16edfbU, |
wolfSSL | 15:117db924cf7c | 996 | 0x43c58643U, 0x4dd79a4dU, 0x33556633U, 0x85941185U, |
wolfSSL | 15:117db924cf7c | 997 | 0x45cf8a45U, 0xf910e9f9U, 0x02060402U, 0x7f81fe7fU, |
wolfSSL | 15:117db924cf7c | 998 | 0x50f0a050U, 0x3c44783cU, 0x9fba259fU, 0xa8e34ba8U, |
wolfSSL | 15:117db924cf7c | 999 | 0x51f3a251U, 0xa3fe5da3U, 0x40c08040U, 0x8f8a058fU, |
wolfSSL | 15:117db924cf7c | 1000 | 0x92ad3f92U, 0x9dbc219dU, 0x38487038U, 0xf504f1f5U, |
wolfSSL | 15:117db924cf7c | 1001 | 0xbcdf63bcU, 0xb6c177b6U, 0xda75afdaU, 0x21634221U, |
wolfSSL | 15:117db924cf7c | 1002 | 0x10302010U, 0xff1ae5ffU, 0xf30efdf3U, 0xd26dbfd2U, |
wolfSSL | 15:117db924cf7c | 1003 | 0xcd4c81cdU, 0x0c14180cU, 0x13352613U, 0xec2fc3ecU, |
wolfSSL | 15:117db924cf7c | 1004 | 0x5fe1be5fU, 0x97a23597U, 0x44cc8844U, 0x17392e17U, |
wolfSSL | 15:117db924cf7c | 1005 | 0xc45793c4U, 0xa7f255a7U, 0x7e82fc7eU, 0x3d477a3dU, |
wolfSSL | 15:117db924cf7c | 1006 | 0x64acc864U, 0x5de7ba5dU, 0x192b3219U, 0x7395e673U, |
wolfSSL | 15:117db924cf7c | 1007 | 0x60a0c060U, 0x81981981U, 0x4fd19e4fU, 0xdc7fa3dcU, |
wolfSSL | 15:117db924cf7c | 1008 | 0x22664422U, 0x2a7e542aU, 0x90ab3b90U, 0x88830b88U, |
wolfSSL | 15:117db924cf7c | 1009 | 0x46ca8c46U, 0xee29c7eeU, 0xb8d36bb8U, 0x143c2814U, |
wolfSSL | 15:117db924cf7c | 1010 | 0xde79a7deU, 0x5ee2bc5eU, 0x0b1d160bU, 0xdb76addbU, |
wolfSSL | 15:117db924cf7c | 1011 | 0xe03bdbe0U, 0x32566432U, 0x3a4e743aU, 0x0a1e140aU, |
wolfSSL | 15:117db924cf7c | 1012 | 0x49db9249U, 0x060a0c06U, 0x246c4824U, 0x5ce4b85cU, |
wolfSSL | 15:117db924cf7c | 1013 | 0xc25d9fc2U, 0xd36ebdd3U, 0xacef43acU, 0x62a6c462U, |
wolfSSL | 15:117db924cf7c | 1014 | 0x91a83991U, 0x95a43195U, 0xe437d3e4U, 0x798bf279U, |
wolfSSL | 15:117db924cf7c | 1015 | 0xe732d5e7U, 0xc8438bc8U, 0x37596e37U, 0x6db7da6dU, |
wolfSSL | 15:117db924cf7c | 1016 | 0x8d8c018dU, 0xd564b1d5U, 0x4ed29c4eU, 0xa9e049a9U, |
wolfSSL | 15:117db924cf7c | 1017 | 0x6cb4d86cU, 0x56faac56U, 0xf407f3f4U, 0xea25cfeaU, |
wolfSSL | 15:117db924cf7c | 1018 | 0x65afca65U, 0x7a8ef47aU, 0xaee947aeU, 0x08181008U, |
wolfSSL | 15:117db924cf7c | 1019 | 0xbad56fbaU, 0x7888f078U, 0x256f4a25U, 0x2e725c2eU, |
wolfSSL | 15:117db924cf7c | 1020 | 0x1c24381cU, 0xa6f157a6U, 0xb4c773b4U, 0xc65197c6U, |
wolfSSL | 15:117db924cf7c | 1021 | 0xe823cbe8U, 0xdd7ca1ddU, 0x749ce874U, 0x1f213e1fU, |
wolfSSL | 15:117db924cf7c | 1022 | 0x4bdd964bU, 0xbddc61bdU, 0x8b860d8bU, 0x8a850f8aU, |
wolfSSL | 15:117db924cf7c | 1023 | 0x7090e070U, 0x3e427c3eU, 0xb5c471b5U, 0x66aacc66U, |
wolfSSL | 15:117db924cf7c | 1024 | 0x48d89048U, 0x03050603U, 0xf601f7f6U, 0x0e121c0eU, |
wolfSSL | 15:117db924cf7c | 1025 | 0x61a3c261U, 0x355f6a35U, 0x57f9ae57U, 0xb9d069b9U, |
wolfSSL | 15:117db924cf7c | 1026 | 0x86911786U, 0xc15899c1U, 0x1d273a1dU, 0x9eb9279eU, |
wolfSSL | 15:117db924cf7c | 1027 | 0xe138d9e1U, 0xf813ebf8U, 0x98b32b98U, 0x11332211U, |
wolfSSL | 15:117db924cf7c | 1028 | 0x69bbd269U, 0xd970a9d9U, 0x8e89078eU, 0x94a73394U, |
wolfSSL | 15:117db924cf7c | 1029 | 0x9bb62d9bU, 0x1e223c1eU, 0x87921587U, 0xe920c9e9U, |
wolfSSL | 15:117db924cf7c | 1030 | 0xce4987ceU, 0x55ffaa55U, 0x28785028U, 0xdf7aa5dfU, |
wolfSSL | 15:117db924cf7c | 1031 | 0x8c8f038cU, 0xa1f859a1U, 0x89800989U, 0x0d171a0dU, |
wolfSSL | 15:117db924cf7c | 1032 | 0xbfda65bfU, 0xe631d7e6U, 0x42c68442U, 0x68b8d068U, |
wolfSSL | 15:117db924cf7c | 1033 | 0x41c38241U, 0x99b02999U, 0x2d775a2dU, 0x0f111e0fU, |
wolfSSL | 15:117db924cf7c | 1034 | 0xb0cb7bb0U, 0x54fca854U, 0xbbd66dbbU, 0x163a2c16U, |
wolfSSL | 15:117db924cf7c | 1035 | }, |
wolfSSL | 15:117db924cf7c | 1036 | { |
wolfSSL | 15:117db924cf7c | 1037 | 0x6363a5c6U, 0x7c7c84f8U, 0x777799eeU, 0x7b7b8df6U, |
wolfSSL | 15:117db924cf7c | 1038 | 0xf2f20dffU, 0x6b6bbdd6U, 0x6f6fb1deU, 0xc5c55491U, |
wolfSSL | 15:117db924cf7c | 1039 | 0x30305060U, 0x01010302U, 0x6767a9ceU, 0x2b2b7d56U, |
wolfSSL | 15:117db924cf7c | 1040 | 0xfefe19e7U, 0xd7d762b5U, 0xababe64dU, 0x76769aecU, |
wolfSSL | 15:117db924cf7c | 1041 | 0xcaca458fU, 0x82829d1fU, 0xc9c94089U, 0x7d7d87faU, |
wolfSSL | 15:117db924cf7c | 1042 | 0xfafa15efU, 0x5959ebb2U, 0x4747c98eU, 0xf0f00bfbU, |
wolfSSL | 15:117db924cf7c | 1043 | 0xadadec41U, 0xd4d467b3U, 0xa2a2fd5fU, 0xafafea45U, |
wolfSSL | 15:117db924cf7c | 1044 | 0x9c9cbf23U, 0xa4a4f753U, 0x727296e4U, 0xc0c05b9bU, |
wolfSSL | 15:117db924cf7c | 1045 | 0xb7b7c275U, 0xfdfd1ce1U, 0x9393ae3dU, 0x26266a4cU, |
wolfSSL | 15:117db924cf7c | 1046 | 0x36365a6cU, 0x3f3f417eU, 0xf7f702f5U, 0xcccc4f83U, |
wolfSSL | 15:117db924cf7c | 1047 | 0x34345c68U, 0xa5a5f451U, 0xe5e534d1U, 0xf1f108f9U, |
wolfSSL | 15:117db924cf7c | 1048 | 0x717193e2U, 0xd8d873abU, 0x31315362U, 0x15153f2aU, |
wolfSSL | 15:117db924cf7c | 1049 | 0x04040c08U, 0xc7c75295U, 0x23236546U, 0xc3c35e9dU, |
wolfSSL | 15:117db924cf7c | 1050 | 0x18182830U, 0x9696a137U, 0x05050f0aU, 0x9a9ab52fU, |
wolfSSL | 15:117db924cf7c | 1051 | 0x0707090eU, 0x12123624U, 0x80809b1bU, 0xe2e23ddfU, |
wolfSSL | 15:117db924cf7c | 1052 | 0xebeb26cdU, 0x2727694eU, 0xb2b2cd7fU, 0x75759feaU, |
wolfSSL | 15:117db924cf7c | 1053 | 0x09091b12U, 0x83839e1dU, 0x2c2c7458U, 0x1a1a2e34U, |
wolfSSL | 15:117db924cf7c | 1054 | 0x1b1b2d36U, 0x6e6eb2dcU, 0x5a5aeeb4U, 0xa0a0fb5bU, |
wolfSSL | 15:117db924cf7c | 1055 | 0x5252f6a4U, 0x3b3b4d76U, 0xd6d661b7U, 0xb3b3ce7dU, |
wolfSSL | 15:117db924cf7c | 1056 | 0x29297b52U, 0xe3e33eddU, 0x2f2f715eU, 0x84849713U, |
wolfSSL | 15:117db924cf7c | 1057 | 0x5353f5a6U, 0xd1d168b9U, 0x00000000U, 0xeded2cc1U, |
wolfSSL | 15:117db924cf7c | 1058 | 0x20206040U, 0xfcfc1fe3U, 0xb1b1c879U, 0x5b5bedb6U, |
wolfSSL | 15:117db924cf7c | 1059 | 0x6a6abed4U, 0xcbcb468dU, 0xbebed967U, 0x39394b72U, |
wolfSSL | 15:117db924cf7c | 1060 | 0x4a4ade94U, 0x4c4cd498U, 0x5858e8b0U, 0xcfcf4a85U, |
wolfSSL | 15:117db924cf7c | 1061 | 0xd0d06bbbU, 0xefef2ac5U, 0xaaaae54fU, 0xfbfb16edU, |
wolfSSL | 15:117db924cf7c | 1062 | 0x4343c586U, 0x4d4dd79aU, 0x33335566U, 0x85859411U, |
wolfSSL | 15:117db924cf7c | 1063 | 0x4545cf8aU, 0xf9f910e9U, 0x02020604U, 0x7f7f81feU, |
wolfSSL | 15:117db924cf7c | 1064 | 0x5050f0a0U, 0x3c3c4478U, 0x9f9fba25U, 0xa8a8e34bU, |
wolfSSL | 15:117db924cf7c | 1065 | 0x5151f3a2U, 0xa3a3fe5dU, 0x4040c080U, 0x8f8f8a05U, |
wolfSSL | 15:117db924cf7c | 1066 | 0x9292ad3fU, 0x9d9dbc21U, 0x38384870U, 0xf5f504f1U, |
wolfSSL | 15:117db924cf7c | 1067 | 0xbcbcdf63U, 0xb6b6c177U, 0xdada75afU, 0x21216342U, |
wolfSSL | 15:117db924cf7c | 1068 | 0x10103020U, 0xffff1ae5U, 0xf3f30efdU, 0xd2d26dbfU, |
wolfSSL | 15:117db924cf7c | 1069 | 0xcdcd4c81U, 0x0c0c1418U, 0x13133526U, 0xecec2fc3U, |
wolfSSL | 15:117db924cf7c | 1070 | 0x5f5fe1beU, 0x9797a235U, 0x4444cc88U, 0x1717392eU, |
wolfSSL | 15:117db924cf7c | 1071 | 0xc4c45793U, 0xa7a7f255U, 0x7e7e82fcU, 0x3d3d477aU, |
wolfSSL | 15:117db924cf7c | 1072 | 0x6464acc8U, 0x5d5de7baU, 0x19192b32U, 0x737395e6U, |
wolfSSL | 15:117db924cf7c | 1073 | 0x6060a0c0U, 0x81819819U, 0x4f4fd19eU, 0xdcdc7fa3U, |
wolfSSL | 15:117db924cf7c | 1074 | 0x22226644U, 0x2a2a7e54U, 0x9090ab3bU, 0x8888830bU, |
wolfSSL | 15:117db924cf7c | 1075 | 0x4646ca8cU, 0xeeee29c7U, 0xb8b8d36bU, 0x14143c28U, |
wolfSSL | 15:117db924cf7c | 1076 | 0xdede79a7U, 0x5e5ee2bcU, 0x0b0b1d16U, 0xdbdb76adU, |
wolfSSL | 15:117db924cf7c | 1077 | 0xe0e03bdbU, 0x32325664U, 0x3a3a4e74U, 0x0a0a1e14U, |
wolfSSL | 15:117db924cf7c | 1078 | 0x4949db92U, 0x06060a0cU, 0x24246c48U, 0x5c5ce4b8U, |
wolfSSL | 15:117db924cf7c | 1079 | 0xc2c25d9fU, 0xd3d36ebdU, 0xacacef43U, 0x6262a6c4U, |
wolfSSL | 15:117db924cf7c | 1080 | 0x9191a839U, 0x9595a431U, 0xe4e437d3U, 0x79798bf2U, |
wolfSSL | 15:117db924cf7c | 1081 | 0xe7e732d5U, 0xc8c8438bU, 0x3737596eU, 0x6d6db7daU, |
wolfSSL | 15:117db924cf7c | 1082 | 0x8d8d8c01U, 0xd5d564b1U, 0x4e4ed29cU, 0xa9a9e049U, |
wolfSSL | 15:117db924cf7c | 1083 | 0x6c6cb4d8U, 0x5656faacU, 0xf4f407f3U, 0xeaea25cfU, |
wolfSSL | 15:117db924cf7c | 1084 | 0x6565afcaU, 0x7a7a8ef4U, 0xaeaee947U, 0x08081810U, |
wolfSSL | 15:117db924cf7c | 1085 | 0xbabad56fU, 0x787888f0U, 0x25256f4aU, 0x2e2e725cU, |
wolfSSL | 15:117db924cf7c | 1086 | 0x1c1c2438U, 0xa6a6f157U, 0xb4b4c773U, 0xc6c65197U, |
wolfSSL | 15:117db924cf7c | 1087 | 0xe8e823cbU, 0xdddd7ca1U, 0x74749ce8U, 0x1f1f213eU, |
wolfSSL | 15:117db924cf7c | 1088 | 0x4b4bdd96U, 0xbdbddc61U, 0x8b8b860dU, 0x8a8a850fU, |
wolfSSL | 15:117db924cf7c | 1089 | 0x707090e0U, 0x3e3e427cU, 0xb5b5c471U, 0x6666aaccU, |
wolfSSL | 15:117db924cf7c | 1090 | 0x4848d890U, 0x03030506U, 0xf6f601f7U, 0x0e0e121cU, |
wolfSSL | 15:117db924cf7c | 1091 | 0x6161a3c2U, 0x35355f6aU, 0x5757f9aeU, 0xb9b9d069U, |
wolfSSL | 15:117db924cf7c | 1092 | 0x86869117U, 0xc1c15899U, 0x1d1d273aU, 0x9e9eb927U, |
wolfSSL | 15:117db924cf7c | 1093 | 0xe1e138d9U, 0xf8f813ebU, 0x9898b32bU, 0x11113322U, |
wolfSSL | 15:117db924cf7c | 1094 | 0x6969bbd2U, 0xd9d970a9U, 0x8e8e8907U, 0x9494a733U, |
wolfSSL | 15:117db924cf7c | 1095 | 0x9b9bb62dU, 0x1e1e223cU, 0x87879215U, 0xe9e920c9U, |
wolfSSL | 15:117db924cf7c | 1096 | 0xcece4987U, 0x5555ffaaU, 0x28287850U, 0xdfdf7aa5U, |
wolfSSL | 15:117db924cf7c | 1097 | 0x8c8c8f03U, 0xa1a1f859U, 0x89898009U, 0x0d0d171aU, |
wolfSSL | 15:117db924cf7c | 1098 | 0xbfbfda65U, 0xe6e631d7U, 0x4242c684U, 0x6868b8d0U, |
wolfSSL | 15:117db924cf7c | 1099 | 0x4141c382U, 0x9999b029U, 0x2d2d775aU, 0x0f0f111eU, |
wolfSSL | 15:117db924cf7c | 1100 | 0xb0b0cb7bU, 0x5454fca8U, 0xbbbbd66dU, 0x16163a2cU, |
wolfSSL | 15:117db924cf7c | 1101 | } |
wolfSSL | 15:117db924cf7c | 1102 | }; |
wolfSSL | 15:117db924cf7c | 1103 | |
wolfSSL | 15:117db924cf7c | 1104 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 1105 | static const word32 Td[4][256] = { |
wolfSSL | 15:117db924cf7c | 1106 | { |
wolfSSL | 15:117db924cf7c | 1107 | 0x51f4a750U, 0x7e416553U, 0x1a17a4c3U, 0x3a275e96U, |
wolfSSL | 15:117db924cf7c | 1108 | 0x3bab6bcbU, 0x1f9d45f1U, 0xacfa58abU, 0x4be30393U, |
wolfSSL | 15:117db924cf7c | 1109 | 0x2030fa55U, 0xad766df6U, 0x88cc7691U, 0xf5024c25U, |
wolfSSL | 15:117db924cf7c | 1110 | 0x4fe5d7fcU, 0xc52acbd7U, 0x26354480U, 0xb562a38fU, |
wolfSSL | 15:117db924cf7c | 1111 | 0xdeb15a49U, 0x25ba1b67U, 0x45ea0e98U, 0x5dfec0e1U, |
wolfSSL | 15:117db924cf7c | 1112 | 0xc32f7502U, 0x814cf012U, 0x8d4697a3U, 0x6bd3f9c6U, |
wolfSSL | 15:117db924cf7c | 1113 | 0x038f5fe7U, 0x15929c95U, 0xbf6d7aebU, 0x955259daU, |
wolfSSL | 15:117db924cf7c | 1114 | 0xd4be832dU, 0x587421d3U, 0x49e06929U, 0x8ec9c844U, |
wolfSSL | 15:117db924cf7c | 1115 | 0x75c2896aU, 0xf48e7978U, 0x99583e6bU, 0x27b971ddU, |
wolfSSL | 15:117db924cf7c | 1116 | 0xbee14fb6U, 0xf088ad17U, 0xc920ac66U, 0x7dce3ab4U, |
wolfSSL | 15:117db924cf7c | 1117 | 0x63df4a18U, 0xe51a3182U, 0x97513360U, 0x62537f45U, |
wolfSSL | 15:117db924cf7c | 1118 | 0xb16477e0U, 0xbb6bae84U, 0xfe81a01cU, 0xf9082b94U, |
wolfSSL | 15:117db924cf7c | 1119 | 0x70486858U, 0x8f45fd19U, 0x94de6c87U, 0x527bf8b7U, |
wolfSSL | 15:117db924cf7c | 1120 | 0xab73d323U, 0x724b02e2U, 0xe31f8f57U, 0x6655ab2aU, |
wolfSSL | 15:117db924cf7c | 1121 | 0xb2eb2807U, 0x2fb5c203U, 0x86c57b9aU, 0xd33708a5U, |
wolfSSL | 15:117db924cf7c | 1122 | 0x302887f2U, 0x23bfa5b2U, 0x02036abaU, 0xed16825cU, |
wolfSSL | 15:117db924cf7c | 1123 | 0x8acf1c2bU, 0xa779b492U, 0xf307f2f0U, 0x4e69e2a1U, |
wolfSSL | 15:117db924cf7c | 1124 | 0x65daf4cdU, 0x0605bed5U, 0xd134621fU, 0xc4a6fe8aU, |
wolfSSL | 15:117db924cf7c | 1125 | 0x342e539dU, 0xa2f355a0U, 0x058ae132U, 0xa4f6eb75U, |
wolfSSL | 15:117db924cf7c | 1126 | 0x0b83ec39U, 0x4060efaaU, 0x5e719f06U, 0xbd6e1051U, |
wolfSSL | 15:117db924cf7c | 1127 | 0x3e218af9U, 0x96dd063dU, 0xdd3e05aeU, 0x4de6bd46U, |
wolfSSL | 15:117db924cf7c | 1128 | 0x91548db5U, 0x71c45d05U, 0x0406d46fU, 0x605015ffU, |
wolfSSL | 15:117db924cf7c | 1129 | 0x1998fb24U, 0xd6bde997U, 0x894043ccU, 0x67d99e77U, |
wolfSSL | 15:117db924cf7c | 1130 | 0xb0e842bdU, 0x07898b88U, 0xe7195b38U, 0x79c8eedbU, |
wolfSSL | 15:117db924cf7c | 1131 | 0xa17c0a47U, 0x7c420fe9U, 0xf8841ec9U, 0x00000000U, |
wolfSSL | 15:117db924cf7c | 1132 | 0x09808683U, 0x322bed48U, 0x1e1170acU, 0x6c5a724eU, |
wolfSSL | 15:117db924cf7c | 1133 | 0xfd0efffbU, 0x0f853856U, 0x3daed51eU, 0x362d3927U, |
wolfSSL | 15:117db924cf7c | 1134 | 0x0a0fd964U, 0x685ca621U, 0x9b5b54d1U, 0x24362e3aU, |
wolfSSL | 15:117db924cf7c | 1135 | 0x0c0a67b1U, 0x9357e70fU, 0xb4ee96d2U, 0x1b9b919eU, |
wolfSSL | 15:117db924cf7c | 1136 | 0x80c0c54fU, 0x61dc20a2U, 0x5a774b69U, 0x1c121a16U, |
wolfSSL | 15:117db924cf7c | 1137 | 0xe293ba0aU, 0xc0a02ae5U, 0x3c22e043U, 0x121b171dU, |
wolfSSL | 15:117db924cf7c | 1138 | 0x0e090d0bU, 0xf28bc7adU, 0x2db6a8b9U, 0x141ea9c8U, |
wolfSSL | 15:117db924cf7c | 1139 | 0x57f11985U, 0xaf75074cU, 0xee99ddbbU, 0xa37f60fdU, |
wolfSSL | 15:117db924cf7c | 1140 | 0xf701269fU, 0x5c72f5bcU, 0x44663bc5U, 0x5bfb7e34U, |
wolfSSL | 15:117db924cf7c | 1141 | 0x8b432976U, 0xcb23c6dcU, 0xb6edfc68U, 0xb8e4f163U, |
wolfSSL | 15:117db924cf7c | 1142 | 0xd731dccaU, 0x42638510U, 0x13972240U, 0x84c61120U, |
wolfSSL | 15:117db924cf7c | 1143 | 0x854a247dU, 0xd2bb3df8U, 0xaef93211U, 0xc729a16dU, |
wolfSSL | 15:117db924cf7c | 1144 | 0x1d9e2f4bU, 0xdcb230f3U, 0x0d8652ecU, 0x77c1e3d0U, |
wolfSSL | 15:117db924cf7c | 1145 | 0x2bb3166cU, 0xa970b999U, 0x119448faU, 0x47e96422U, |
wolfSSL | 15:117db924cf7c | 1146 | 0xa8fc8cc4U, 0xa0f03f1aU, 0x567d2cd8U, 0x223390efU, |
wolfSSL | 15:117db924cf7c | 1147 | 0x87494ec7U, 0xd938d1c1U, 0x8ccaa2feU, 0x98d40b36U, |
wolfSSL | 15:117db924cf7c | 1148 | 0xa6f581cfU, 0xa57ade28U, 0xdab78e26U, 0x3fadbfa4U, |
wolfSSL | 15:117db924cf7c | 1149 | 0x2c3a9de4U, 0x5078920dU, 0x6a5fcc9bU, 0x547e4662U, |
wolfSSL | 15:117db924cf7c | 1150 | 0xf68d13c2U, 0x90d8b8e8U, 0x2e39f75eU, 0x82c3aff5U, |
wolfSSL | 15:117db924cf7c | 1151 | 0x9f5d80beU, 0x69d0937cU, 0x6fd52da9U, 0xcf2512b3U, |
wolfSSL | 15:117db924cf7c | 1152 | 0xc8ac993bU, 0x10187da7U, 0xe89c636eU, 0xdb3bbb7bU, |
wolfSSL | 15:117db924cf7c | 1153 | 0xcd267809U, 0x6e5918f4U, 0xec9ab701U, 0x834f9aa8U, |
wolfSSL | 15:117db924cf7c | 1154 | 0xe6956e65U, 0xaaffe67eU, 0x21bccf08U, 0xef15e8e6U, |
wolfSSL | 15:117db924cf7c | 1155 | 0xbae79bd9U, 0x4a6f36ceU, 0xea9f09d4U, 0x29b07cd6U, |
wolfSSL | 15:117db924cf7c | 1156 | 0x31a4b2afU, 0x2a3f2331U, 0xc6a59430U, 0x35a266c0U, |
wolfSSL | 15:117db924cf7c | 1157 | 0x744ebc37U, 0xfc82caa6U, 0xe090d0b0U, 0x33a7d815U, |
wolfSSL | 15:117db924cf7c | 1158 | 0xf104984aU, 0x41ecdaf7U, 0x7fcd500eU, 0x1791f62fU, |
wolfSSL | 15:117db924cf7c | 1159 | 0x764dd68dU, 0x43efb04dU, 0xccaa4d54U, 0xe49604dfU, |
wolfSSL | 15:117db924cf7c | 1160 | 0x9ed1b5e3U, 0x4c6a881bU, 0xc12c1fb8U, 0x4665517fU, |
wolfSSL | 15:117db924cf7c | 1161 | 0x9d5eea04U, 0x018c355dU, 0xfa877473U, 0xfb0b412eU, |
wolfSSL | 15:117db924cf7c | 1162 | 0xb3671d5aU, 0x92dbd252U, 0xe9105633U, 0x6dd64713U, |
wolfSSL | 15:117db924cf7c | 1163 | 0x9ad7618cU, 0x37a10c7aU, 0x59f8148eU, 0xeb133c89U, |
wolfSSL | 15:117db924cf7c | 1164 | 0xcea927eeU, 0xb761c935U, 0xe11ce5edU, 0x7a47b13cU, |
wolfSSL | 15:117db924cf7c | 1165 | 0x9cd2df59U, 0x55f2733fU, 0x1814ce79U, 0x73c737bfU, |
wolfSSL | 15:117db924cf7c | 1166 | 0x53f7cdeaU, 0x5ffdaa5bU, 0xdf3d6f14U, 0x7844db86U, |
wolfSSL | 15:117db924cf7c | 1167 | 0xcaaff381U, 0xb968c43eU, 0x3824342cU, 0xc2a3405fU, |
wolfSSL | 15:117db924cf7c | 1168 | 0x161dc372U, 0xbce2250cU, 0x283c498bU, 0xff0d9541U, |
wolfSSL | 15:117db924cf7c | 1169 | 0x39a80171U, 0x080cb3deU, 0xd8b4e49cU, 0x6456c190U, |
wolfSSL | 15:117db924cf7c | 1170 | 0x7bcb8461U, 0xd532b670U, 0x486c5c74U, 0xd0b85742U, |
wolfSSL | 15:117db924cf7c | 1171 | }, |
wolfSSL | 15:117db924cf7c | 1172 | { |
wolfSSL | 15:117db924cf7c | 1173 | 0x5051f4a7U, 0x537e4165U, 0xc31a17a4U, 0x963a275eU, |
wolfSSL | 15:117db924cf7c | 1174 | 0xcb3bab6bU, 0xf11f9d45U, 0xabacfa58U, 0x934be303U, |
wolfSSL | 15:117db924cf7c | 1175 | 0x552030faU, 0xf6ad766dU, 0x9188cc76U, 0x25f5024cU, |
wolfSSL | 15:117db924cf7c | 1176 | 0xfc4fe5d7U, 0xd7c52acbU, 0x80263544U, 0x8fb562a3U, |
wolfSSL | 15:117db924cf7c | 1177 | 0x49deb15aU, 0x6725ba1bU, 0x9845ea0eU, 0xe15dfec0U, |
wolfSSL | 15:117db924cf7c | 1178 | 0x02c32f75U, 0x12814cf0U, 0xa38d4697U, 0xc66bd3f9U, |
wolfSSL | 15:117db924cf7c | 1179 | 0xe7038f5fU, 0x9515929cU, 0xebbf6d7aU, 0xda955259U, |
wolfSSL | 15:117db924cf7c | 1180 | 0x2dd4be83U, 0xd3587421U, 0x2949e069U, 0x448ec9c8U, |
wolfSSL | 15:117db924cf7c | 1181 | 0x6a75c289U, 0x78f48e79U, 0x6b99583eU, 0xdd27b971U, |
wolfSSL | 15:117db924cf7c | 1182 | 0xb6bee14fU, 0x17f088adU, 0x66c920acU, 0xb47dce3aU, |
wolfSSL | 15:117db924cf7c | 1183 | 0x1863df4aU, 0x82e51a31U, 0x60975133U, 0x4562537fU, |
wolfSSL | 15:117db924cf7c | 1184 | 0xe0b16477U, 0x84bb6baeU, 0x1cfe81a0U, 0x94f9082bU, |
wolfSSL | 15:117db924cf7c | 1185 | 0x58704868U, 0x198f45fdU, 0x8794de6cU, 0xb7527bf8U, |
wolfSSL | 15:117db924cf7c | 1186 | 0x23ab73d3U, 0xe2724b02U, 0x57e31f8fU, 0x2a6655abU, |
wolfSSL | 15:117db924cf7c | 1187 | 0x07b2eb28U, 0x032fb5c2U, 0x9a86c57bU, 0xa5d33708U, |
wolfSSL | 15:117db924cf7c | 1188 | 0xf2302887U, 0xb223bfa5U, 0xba02036aU, 0x5ced1682U, |
wolfSSL | 15:117db924cf7c | 1189 | 0x2b8acf1cU, 0x92a779b4U, 0xf0f307f2U, 0xa14e69e2U, |
wolfSSL | 15:117db924cf7c | 1190 | 0xcd65daf4U, 0xd50605beU, 0x1fd13462U, 0x8ac4a6feU, |
wolfSSL | 15:117db924cf7c | 1191 | 0x9d342e53U, 0xa0a2f355U, 0x32058ae1U, 0x75a4f6ebU, |
wolfSSL | 15:117db924cf7c | 1192 | 0x390b83ecU, 0xaa4060efU, 0x065e719fU, 0x51bd6e10U, |
wolfSSL | 15:117db924cf7c | 1193 | 0xf93e218aU, 0x3d96dd06U, 0xaedd3e05U, 0x464de6bdU, |
wolfSSL | 15:117db924cf7c | 1194 | 0xb591548dU, 0x0571c45dU, 0x6f0406d4U, 0xff605015U, |
wolfSSL | 15:117db924cf7c | 1195 | 0x241998fbU, 0x97d6bde9U, 0xcc894043U, 0x7767d99eU, |
wolfSSL | 15:117db924cf7c | 1196 | 0xbdb0e842U, 0x8807898bU, 0x38e7195bU, 0xdb79c8eeU, |
wolfSSL | 15:117db924cf7c | 1197 | 0x47a17c0aU, 0xe97c420fU, 0xc9f8841eU, 0x00000000U, |
wolfSSL | 15:117db924cf7c | 1198 | 0x83098086U, 0x48322bedU, 0xac1e1170U, 0x4e6c5a72U, |
wolfSSL | 15:117db924cf7c | 1199 | 0xfbfd0effU, 0x560f8538U, 0x1e3daed5U, 0x27362d39U, |
wolfSSL | 15:117db924cf7c | 1200 | 0x640a0fd9U, 0x21685ca6U, 0xd19b5b54U, 0x3a24362eU, |
wolfSSL | 15:117db924cf7c | 1201 | 0xb10c0a67U, 0x0f9357e7U, 0xd2b4ee96U, 0x9e1b9b91U, |
wolfSSL | 15:117db924cf7c | 1202 | 0x4f80c0c5U, 0xa261dc20U, 0x695a774bU, 0x161c121aU, |
wolfSSL | 15:117db924cf7c | 1203 | 0x0ae293baU, 0xe5c0a02aU, 0x433c22e0U, 0x1d121b17U, |
wolfSSL | 15:117db924cf7c | 1204 | 0x0b0e090dU, 0xadf28bc7U, 0xb92db6a8U, 0xc8141ea9U, |
wolfSSL | 15:117db924cf7c | 1205 | 0x8557f119U, 0x4caf7507U, 0xbbee99ddU, 0xfda37f60U, |
wolfSSL | 15:117db924cf7c | 1206 | 0x9ff70126U, 0xbc5c72f5U, 0xc544663bU, 0x345bfb7eU, |
wolfSSL | 15:117db924cf7c | 1207 | 0x768b4329U, 0xdccb23c6U, 0x68b6edfcU, 0x63b8e4f1U, |
wolfSSL | 15:117db924cf7c | 1208 | 0xcad731dcU, 0x10426385U, 0x40139722U, 0x2084c611U, |
wolfSSL | 15:117db924cf7c | 1209 | 0x7d854a24U, 0xf8d2bb3dU, 0x11aef932U, 0x6dc729a1U, |
wolfSSL | 15:117db924cf7c | 1210 | 0x4b1d9e2fU, 0xf3dcb230U, 0xec0d8652U, 0xd077c1e3U, |
wolfSSL | 15:117db924cf7c | 1211 | 0x6c2bb316U, 0x99a970b9U, 0xfa119448U, 0x2247e964U, |
wolfSSL | 15:117db924cf7c | 1212 | 0xc4a8fc8cU, 0x1aa0f03fU, 0xd8567d2cU, 0xef223390U, |
wolfSSL | 15:117db924cf7c | 1213 | 0xc787494eU, 0xc1d938d1U, 0xfe8ccaa2U, 0x3698d40bU, |
wolfSSL | 15:117db924cf7c | 1214 | 0xcfa6f581U, 0x28a57adeU, 0x26dab78eU, 0xa43fadbfU, |
wolfSSL | 15:117db924cf7c | 1215 | 0xe42c3a9dU, 0x0d507892U, 0x9b6a5fccU, 0x62547e46U, |
wolfSSL | 15:117db924cf7c | 1216 | 0xc2f68d13U, 0xe890d8b8U, 0x5e2e39f7U, 0xf582c3afU, |
wolfSSL | 15:117db924cf7c | 1217 | 0xbe9f5d80U, 0x7c69d093U, 0xa96fd52dU, 0xb3cf2512U, |
wolfSSL | 15:117db924cf7c | 1218 | 0x3bc8ac99U, 0xa710187dU, 0x6ee89c63U, 0x7bdb3bbbU, |
wolfSSL | 15:117db924cf7c | 1219 | 0x09cd2678U, 0xf46e5918U, 0x01ec9ab7U, 0xa8834f9aU, |
wolfSSL | 15:117db924cf7c | 1220 | 0x65e6956eU, 0x7eaaffe6U, 0x0821bccfU, 0xe6ef15e8U, |
wolfSSL | 15:117db924cf7c | 1221 | 0xd9bae79bU, 0xce4a6f36U, 0xd4ea9f09U, 0xd629b07cU, |
wolfSSL | 15:117db924cf7c | 1222 | 0xaf31a4b2U, 0x312a3f23U, 0x30c6a594U, 0xc035a266U, |
wolfSSL | 15:117db924cf7c | 1223 | 0x37744ebcU, 0xa6fc82caU, 0xb0e090d0U, 0x1533a7d8U, |
wolfSSL | 15:117db924cf7c | 1224 | 0x4af10498U, 0xf741ecdaU, 0x0e7fcd50U, 0x2f1791f6U, |
wolfSSL | 15:117db924cf7c | 1225 | 0x8d764dd6U, 0x4d43efb0U, 0x54ccaa4dU, 0xdfe49604U, |
wolfSSL | 15:117db924cf7c | 1226 | 0xe39ed1b5U, 0x1b4c6a88U, 0xb8c12c1fU, 0x7f466551U, |
wolfSSL | 15:117db924cf7c | 1227 | 0x049d5eeaU, 0x5d018c35U, 0x73fa8774U, 0x2efb0b41U, |
wolfSSL | 15:117db924cf7c | 1228 | 0x5ab3671dU, 0x5292dbd2U, 0x33e91056U, 0x136dd647U, |
wolfSSL | 15:117db924cf7c | 1229 | 0x8c9ad761U, 0x7a37a10cU, 0x8e59f814U, 0x89eb133cU, |
wolfSSL | 15:117db924cf7c | 1230 | 0xeecea927U, 0x35b761c9U, 0xede11ce5U, 0x3c7a47b1U, |
wolfSSL | 15:117db924cf7c | 1231 | 0x599cd2dfU, 0x3f55f273U, 0x791814ceU, 0xbf73c737U, |
wolfSSL | 15:117db924cf7c | 1232 | 0xea53f7cdU, 0x5b5ffdaaU, 0x14df3d6fU, 0x867844dbU, |
wolfSSL | 15:117db924cf7c | 1233 | 0x81caaff3U, 0x3eb968c4U, 0x2c382434U, 0x5fc2a340U, |
wolfSSL | 15:117db924cf7c | 1234 | 0x72161dc3U, 0x0cbce225U, 0x8b283c49U, 0x41ff0d95U, |
wolfSSL | 15:117db924cf7c | 1235 | 0x7139a801U, 0xde080cb3U, 0x9cd8b4e4U, 0x906456c1U, |
wolfSSL | 15:117db924cf7c | 1236 | 0x617bcb84U, 0x70d532b6U, 0x74486c5cU, 0x42d0b857U, |
wolfSSL | 15:117db924cf7c | 1237 | }, |
wolfSSL | 15:117db924cf7c | 1238 | { |
wolfSSL | 15:117db924cf7c | 1239 | 0xa75051f4U, 0x65537e41U, 0xa4c31a17U, 0x5e963a27U, |
wolfSSL | 15:117db924cf7c | 1240 | 0x6bcb3babU, 0x45f11f9dU, 0x58abacfaU, 0x03934be3U, |
wolfSSL | 15:117db924cf7c | 1241 | 0xfa552030U, 0x6df6ad76U, 0x769188ccU, 0x4c25f502U, |
wolfSSL | 15:117db924cf7c | 1242 | 0xd7fc4fe5U, 0xcbd7c52aU, 0x44802635U, 0xa38fb562U, |
wolfSSL | 15:117db924cf7c | 1243 | 0x5a49deb1U, 0x1b6725baU, 0x0e9845eaU, 0xc0e15dfeU, |
wolfSSL | 15:117db924cf7c | 1244 | 0x7502c32fU, 0xf012814cU, 0x97a38d46U, 0xf9c66bd3U, |
wolfSSL | 15:117db924cf7c | 1245 | 0x5fe7038fU, 0x9c951592U, 0x7aebbf6dU, 0x59da9552U, |
wolfSSL | 15:117db924cf7c | 1246 | 0x832dd4beU, 0x21d35874U, 0x692949e0U, 0xc8448ec9U, |
wolfSSL | 15:117db924cf7c | 1247 | 0x896a75c2U, 0x7978f48eU, 0x3e6b9958U, 0x71dd27b9U, |
wolfSSL | 15:117db924cf7c | 1248 | 0x4fb6bee1U, 0xad17f088U, 0xac66c920U, 0x3ab47dceU, |
wolfSSL | 15:117db924cf7c | 1249 | 0x4a1863dfU, 0x3182e51aU, 0x33609751U, 0x7f456253U, |
wolfSSL | 15:117db924cf7c | 1250 | 0x77e0b164U, 0xae84bb6bU, 0xa01cfe81U, 0x2b94f908U, |
wolfSSL | 15:117db924cf7c | 1251 | 0x68587048U, 0xfd198f45U, 0x6c8794deU, 0xf8b7527bU, |
wolfSSL | 15:117db924cf7c | 1252 | 0xd323ab73U, 0x02e2724bU, 0x8f57e31fU, 0xab2a6655U, |
wolfSSL | 15:117db924cf7c | 1253 | 0x2807b2ebU, 0xc2032fb5U, 0x7b9a86c5U, 0x08a5d337U, |
wolfSSL | 15:117db924cf7c | 1254 | 0x87f23028U, 0xa5b223bfU, 0x6aba0203U, 0x825ced16U, |
wolfSSL | 15:117db924cf7c | 1255 | 0x1c2b8acfU, 0xb492a779U, 0xf2f0f307U, 0xe2a14e69U, |
wolfSSL | 15:117db924cf7c | 1256 | 0xf4cd65daU, 0xbed50605U, 0x621fd134U, 0xfe8ac4a6U, |
wolfSSL | 15:117db924cf7c | 1257 | 0x539d342eU, 0x55a0a2f3U, 0xe132058aU, 0xeb75a4f6U, |
wolfSSL | 15:117db924cf7c | 1258 | 0xec390b83U, 0xefaa4060U, 0x9f065e71U, 0x1051bd6eU, |
wolfSSL | 15:117db924cf7c | 1259 | |
wolfSSL | 15:117db924cf7c | 1260 | 0x8af93e21U, 0x063d96ddU, 0x05aedd3eU, 0xbd464de6U, |
wolfSSL | 15:117db924cf7c | 1261 | 0x8db59154U, 0x5d0571c4U, 0xd46f0406U, 0x15ff6050U, |
wolfSSL | 15:117db924cf7c | 1262 | 0xfb241998U, 0xe997d6bdU, 0x43cc8940U, 0x9e7767d9U, |
wolfSSL | 15:117db924cf7c | 1263 | 0x42bdb0e8U, 0x8b880789U, 0x5b38e719U, 0xeedb79c8U, |
wolfSSL | 15:117db924cf7c | 1264 | 0x0a47a17cU, 0x0fe97c42U, 0x1ec9f884U, 0x00000000U, |
wolfSSL | 15:117db924cf7c | 1265 | 0x86830980U, 0xed48322bU, 0x70ac1e11U, 0x724e6c5aU, |
wolfSSL | 15:117db924cf7c | 1266 | 0xfffbfd0eU, 0x38560f85U, 0xd51e3daeU, 0x3927362dU, |
wolfSSL | 15:117db924cf7c | 1267 | 0xd9640a0fU, 0xa621685cU, 0x54d19b5bU, 0x2e3a2436U, |
wolfSSL | 15:117db924cf7c | 1268 | 0x67b10c0aU, 0xe70f9357U, 0x96d2b4eeU, 0x919e1b9bU, |
wolfSSL | 15:117db924cf7c | 1269 | 0xc54f80c0U, 0x20a261dcU, 0x4b695a77U, 0x1a161c12U, |
wolfSSL | 15:117db924cf7c | 1270 | 0xba0ae293U, 0x2ae5c0a0U, 0xe0433c22U, 0x171d121bU, |
wolfSSL | 15:117db924cf7c | 1271 | 0x0d0b0e09U, 0xc7adf28bU, 0xa8b92db6U, 0xa9c8141eU, |
wolfSSL | 15:117db924cf7c | 1272 | 0x198557f1U, 0x074caf75U, 0xddbbee99U, 0x60fda37fU, |
wolfSSL | 15:117db924cf7c | 1273 | 0x269ff701U, 0xf5bc5c72U, 0x3bc54466U, 0x7e345bfbU, |
wolfSSL | 15:117db924cf7c | 1274 | 0x29768b43U, 0xc6dccb23U, 0xfc68b6edU, 0xf163b8e4U, |
wolfSSL | 15:117db924cf7c | 1275 | 0xdccad731U, 0x85104263U, 0x22401397U, 0x112084c6U, |
wolfSSL | 15:117db924cf7c | 1276 | 0x247d854aU, 0x3df8d2bbU, 0x3211aef9U, 0xa16dc729U, |
wolfSSL | 15:117db924cf7c | 1277 | 0x2f4b1d9eU, 0x30f3dcb2U, 0x52ec0d86U, 0xe3d077c1U, |
wolfSSL | 15:117db924cf7c | 1278 | 0x166c2bb3U, 0xb999a970U, 0x48fa1194U, 0x642247e9U, |
wolfSSL | 15:117db924cf7c | 1279 | 0x8cc4a8fcU, 0x3f1aa0f0U, 0x2cd8567dU, 0x90ef2233U, |
wolfSSL | 15:117db924cf7c | 1280 | 0x4ec78749U, 0xd1c1d938U, 0xa2fe8ccaU, 0x0b3698d4U, |
wolfSSL | 15:117db924cf7c | 1281 | 0x81cfa6f5U, 0xde28a57aU, 0x8e26dab7U, 0xbfa43fadU, |
wolfSSL | 15:117db924cf7c | 1282 | 0x9de42c3aU, 0x920d5078U, 0xcc9b6a5fU, 0x4662547eU, |
wolfSSL | 15:117db924cf7c | 1283 | 0x13c2f68dU, 0xb8e890d8U, 0xf75e2e39U, 0xaff582c3U, |
wolfSSL | 15:117db924cf7c | 1284 | 0x80be9f5dU, 0x937c69d0U, 0x2da96fd5U, 0x12b3cf25U, |
wolfSSL | 15:117db924cf7c | 1285 | 0x993bc8acU, 0x7da71018U, 0x636ee89cU, 0xbb7bdb3bU, |
wolfSSL | 15:117db924cf7c | 1286 | 0x7809cd26U, 0x18f46e59U, 0xb701ec9aU, 0x9aa8834fU, |
wolfSSL | 15:117db924cf7c | 1287 | 0x6e65e695U, 0xe67eaaffU, 0xcf0821bcU, 0xe8e6ef15U, |
wolfSSL | 15:117db924cf7c | 1288 | 0x9bd9bae7U, 0x36ce4a6fU, 0x09d4ea9fU, 0x7cd629b0U, |
wolfSSL | 15:117db924cf7c | 1289 | 0xb2af31a4U, 0x23312a3fU, 0x9430c6a5U, 0x66c035a2U, |
wolfSSL | 15:117db924cf7c | 1290 | 0xbc37744eU, 0xcaa6fc82U, 0xd0b0e090U, 0xd81533a7U, |
wolfSSL | 15:117db924cf7c | 1291 | 0x984af104U, 0xdaf741ecU, 0x500e7fcdU, 0xf62f1791U, |
wolfSSL | 15:117db924cf7c | 1292 | 0xd68d764dU, 0xb04d43efU, 0x4d54ccaaU, 0x04dfe496U, |
wolfSSL | 15:117db924cf7c | 1293 | 0xb5e39ed1U, 0x881b4c6aU, 0x1fb8c12cU, 0x517f4665U, |
wolfSSL | 15:117db924cf7c | 1294 | 0xea049d5eU, 0x355d018cU, 0x7473fa87U, 0x412efb0bU, |
wolfSSL | 15:117db924cf7c | 1295 | 0x1d5ab367U, 0xd25292dbU, 0x5633e910U, 0x47136dd6U, |
wolfSSL | 15:117db924cf7c | 1296 | 0x618c9ad7U, 0x0c7a37a1U, 0x148e59f8U, 0x3c89eb13U, |
wolfSSL | 15:117db924cf7c | 1297 | 0x27eecea9U, 0xc935b761U, 0xe5ede11cU, 0xb13c7a47U, |
wolfSSL | 15:117db924cf7c | 1298 | 0xdf599cd2U, 0x733f55f2U, 0xce791814U, 0x37bf73c7U, |
wolfSSL | 15:117db924cf7c | 1299 | 0xcdea53f7U, 0xaa5b5ffdU, 0x6f14df3dU, 0xdb867844U, |
wolfSSL | 15:117db924cf7c | 1300 | 0xf381caafU, 0xc43eb968U, 0x342c3824U, 0x405fc2a3U, |
wolfSSL | 15:117db924cf7c | 1301 | 0xc372161dU, 0x250cbce2U, 0x498b283cU, 0x9541ff0dU, |
wolfSSL | 15:117db924cf7c | 1302 | 0x017139a8U, 0xb3de080cU, 0xe49cd8b4U, 0xc1906456U, |
wolfSSL | 15:117db924cf7c | 1303 | 0x84617bcbU, 0xb670d532U, 0x5c74486cU, 0x5742d0b8U, |
wolfSSL | 15:117db924cf7c | 1304 | }, |
wolfSSL | 15:117db924cf7c | 1305 | { |
wolfSSL | 15:117db924cf7c | 1306 | 0xf4a75051U, 0x4165537eU, 0x17a4c31aU, 0x275e963aU, |
wolfSSL | 15:117db924cf7c | 1307 | 0xab6bcb3bU, 0x9d45f11fU, 0xfa58abacU, 0xe303934bU, |
wolfSSL | 15:117db924cf7c | 1308 | 0x30fa5520U, 0x766df6adU, 0xcc769188U, 0x024c25f5U, |
wolfSSL | 15:117db924cf7c | 1309 | 0xe5d7fc4fU, 0x2acbd7c5U, 0x35448026U, 0x62a38fb5U, |
wolfSSL | 15:117db924cf7c | 1310 | 0xb15a49deU, 0xba1b6725U, 0xea0e9845U, 0xfec0e15dU, |
wolfSSL | 15:117db924cf7c | 1311 | 0x2f7502c3U, 0x4cf01281U, 0x4697a38dU, 0xd3f9c66bU, |
wolfSSL | 15:117db924cf7c | 1312 | 0x8f5fe703U, 0x929c9515U, 0x6d7aebbfU, 0x5259da95U, |
wolfSSL | 15:117db924cf7c | 1313 | 0xbe832dd4U, 0x7421d358U, 0xe0692949U, 0xc9c8448eU, |
wolfSSL | 15:117db924cf7c | 1314 | 0xc2896a75U, 0x8e7978f4U, 0x583e6b99U, 0xb971dd27U, |
wolfSSL | 15:117db924cf7c | 1315 | 0xe14fb6beU, 0x88ad17f0U, 0x20ac66c9U, 0xce3ab47dU, |
wolfSSL | 15:117db924cf7c | 1316 | 0xdf4a1863U, 0x1a3182e5U, 0x51336097U, 0x537f4562U, |
wolfSSL | 15:117db924cf7c | 1317 | 0x6477e0b1U, 0x6bae84bbU, 0x81a01cfeU, 0x082b94f9U, |
wolfSSL | 15:117db924cf7c | 1318 | 0x48685870U, 0x45fd198fU, 0xde6c8794U, 0x7bf8b752U, |
wolfSSL | 15:117db924cf7c | 1319 | 0x73d323abU, 0x4b02e272U, 0x1f8f57e3U, 0x55ab2a66U, |
wolfSSL | 15:117db924cf7c | 1320 | 0xeb2807b2U, 0xb5c2032fU, 0xc57b9a86U, 0x3708a5d3U, |
wolfSSL | 15:117db924cf7c | 1321 | 0x2887f230U, 0xbfa5b223U, 0x036aba02U, 0x16825cedU, |
wolfSSL | 15:117db924cf7c | 1322 | 0xcf1c2b8aU, 0x79b492a7U, 0x07f2f0f3U, 0x69e2a14eU, |
wolfSSL | 15:117db924cf7c | 1323 | 0xdaf4cd65U, 0x05bed506U, 0x34621fd1U, 0xa6fe8ac4U, |
wolfSSL | 15:117db924cf7c | 1324 | 0x2e539d34U, 0xf355a0a2U, 0x8ae13205U, 0xf6eb75a4U, |
wolfSSL | 15:117db924cf7c | 1325 | 0x83ec390bU, 0x60efaa40U, 0x719f065eU, 0x6e1051bdU, |
wolfSSL | 15:117db924cf7c | 1326 | 0x218af93eU, 0xdd063d96U, 0x3e05aeddU, 0xe6bd464dU, |
wolfSSL | 15:117db924cf7c | 1327 | 0x548db591U, 0xc45d0571U, 0x06d46f04U, 0x5015ff60U, |
wolfSSL | 15:117db924cf7c | 1328 | 0x98fb2419U, 0xbde997d6U, 0x4043cc89U, 0xd99e7767U, |
wolfSSL | 15:117db924cf7c | 1329 | 0xe842bdb0U, 0x898b8807U, 0x195b38e7U, 0xc8eedb79U, |
wolfSSL | 15:117db924cf7c | 1330 | 0x7c0a47a1U, 0x420fe97cU, 0x841ec9f8U, 0x00000000U, |
wolfSSL | 15:117db924cf7c | 1331 | 0x80868309U, 0x2bed4832U, 0x1170ac1eU, 0x5a724e6cU, |
wolfSSL | 15:117db924cf7c | 1332 | 0x0efffbfdU, 0x8538560fU, 0xaed51e3dU, 0x2d392736U, |
wolfSSL | 15:117db924cf7c | 1333 | 0x0fd9640aU, 0x5ca62168U, 0x5b54d19bU, 0x362e3a24U, |
wolfSSL | 15:117db924cf7c | 1334 | 0x0a67b10cU, 0x57e70f93U, 0xee96d2b4U, 0x9b919e1bU, |
wolfSSL | 15:117db924cf7c | 1335 | 0xc0c54f80U, 0xdc20a261U, 0x774b695aU, 0x121a161cU, |
wolfSSL | 15:117db924cf7c | 1336 | 0x93ba0ae2U, 0xa02ae5c0U, 0x22e0433cU, 0x1b171d12U, |
wolfSSL | 15:117db924cf7c | 1337 | 0x090d0b0eU, 0x8bc7adf2U, 0xb6a8b92dU, 0x1ea9c814U, |
wolfSSL | 15:117db924cf7c | 1338 | 0xf1198557U, 0x75074cafU, 0x99ddbbeeU, 0x7f60fda3U, |
wolfSSL | 15:117db924cf7c | 1339 | 0x01269ff7U, 0x72f5bc5cU, 0x663bc544U, 0xfb7e345bU, |
wolfSSL | 15:117db924cf7c | 1340 | 0x4329768bU, 0x23c6dccbU, 0xedfc68b6U, 0xe4f163b8U, |
wolfSSL | 15:117db924cf7c | 1341 | 0x31dccad7U, 0x63851042U, 0x97224013U, 0xc6112084U, |
wolfSSL | 15:117db924cf7c | 1342 | 0x4a247d85U, 0xbb3df8d2U, 0xf93211aeU, 0x29a16dc7U, |
wolfSSL | 15:117db924cf7c | 1343 | 0x9e2f4b1dU, 0xb230f3dcU, 0x8652ec0dU, 0xc1e3d077U, |
wolfSSL | 15:117db924cf7c | 1344 | 0xb3166c2bU, 0x70b999a9U, 0x9448fa11U, 0xe9642247U, |
wolfSSL | 15:117db924cf7c | 1345 | 0xfc8cc4a8U, 0xf03f1aa0U, 0x7d2cd856U, 0x3390ef22U, |
wolfSSL | 15:117db924cf7c | 1346 | 0x494ec787U, 0x38d1c1d9U, 0xcaa2fe8cU, 0xd40b3698U, |
wolfSSL | 15:117db924cf7c | 1347 | 0xf581cfa6U, 0x7ade28a5U, 0xb78e26daU, 0xadbfa43fU, |
wolfSSL | 15:117db924cf7c | 1348 | 0x3a9de42cU, 0x78920d50U, 0x5fcc9b6aU, 0x7e466254U, |
wolfSSL | 15:117db924cf7c | 1349 | 0x8d13c2f6U, 0xd8b8e890U, 0x39f75e2eU, 0xc3aff582U, |
wolfSSL | 15:117db924cf7c | 1350 | 0x5d80be9fU, 0xd0937c69U, 0xd52da96fU, 0x2512b3cfU, |
wolfSSL | 15:117db924cf7c | 1351 | 0xac993bc8U, 0x187da710U, 0x9c636ee8U, 0x3bbb7bdbU, |
wolfSSL | 15:117db924cf7c | 1352 | 0x267809cdU, 0x5918f46eU, 0x9ab701ecU, 0x4f9aa883U, |
wolfSSL | 15:117db924cf7c | 1353 | 0x956e65e6U, 0xffe67eaaU, 0xbccf0821U, 0x15e8e6efU, |
wolfSSL | 15:117db924cf7c | 1354 | 0xe79bd9baU, 0x6f36ce4aU, 0x9f09d4eaU, 0xb07cd629U, |
wolfSSL | 15:117db924cf7c | 1355 | 0xa4b2af31U, 0x3f23312aU, 0xa59430c6U, 0xa266c035U, |
wolfSSL | 15:117db924cf7c | 1356 | 0x4ebc3774U, 0x82caa6fcU, 0x90d0b0e0U, 0xa7d81533U, |
wolfSSL | 15:117db924cf7c | 1357 | 0x04984af1U, 0xecdaf741U, 0xcd500e7fU, 0x91f62f17U, |
wolfSSL | 15:117db924cf7c | 1358 | 0x4dd68d76U, 0xefb04d43U, 0xaa4d54ccU, 0x9604dfe4U, |
wolfSSL | 15:117db924cf7c | 1359 | 0xd1b5e39eU, 0x6a881b4cU, 0x2c1fb8c1U, 0x65517f46U, |
wolfSSL | 15:117db924cf7c | 1360 | 0x5eea049dU, 0x8c355d01U, 0x877473faU, 0x0b412efbU, |
wolfSSL | 15:117db924cf7c | 1361 | 0x671d5ab3U, 0xdbd25292U, 0x105633e9U, 0xd647136dU, |
wolfSSL | 15:117db924cf7c | 1362 | 0xd7618c9aU, 0xa10c7a37U, 0xf8148e59U, 0x133c89ebU, |
wolfSSL | 15:117db924cf7c | 1363 | 0xa927eeceU, 0x61c935b7U, 0x1ce5ede1U, 0x47b13c7aU, |
wolfSSL | 15:117db924cf7c | 1364 | 0xd2df599cU, 0xf2733f55U, 0x14ce7918U, 0xc737bf73U, |
wolfSSL | 15:117db924cf7c | 1365 | 0xf7cdea53U, 0xfdaa5b5fU, 0x3d6f14dfU, 0x44db8678U, |
wolfSSL | 15:117db924cf7c | 1366 | 0xaff381caU, 0x68c43eb9U, 0x24342c38U, 0xa3405fc2U, |
wolfSSL | 15:117db924cf7c | 1367 | 0x1dc37216U, 0xe2250cbcU, 0x3c498b28U, 0x0d9541ffU, |
wolfSSL | 15:117db924cf7c | 1368 | 0xa8017139U, 0x0cb3de08U, 0xb4e49cd8U, 0x56c19064U, |
wolfSSL | 15:117db924cf7c | 1369 | 0xcb84617bU, 0x32b670d5U, 0x6c5c7448U, 0xb85742d0U, |
wolfSSL | 15:117db924cf7c | 1370 | } |
wolfSSL | 15:117db924cf7c | 1371 | }; |
wolfSSL | 15:117db924cf7c | 1372 | |
wolfSSL | 15:117db924cf7c | 1373 | |
wolfSSL | 15:117db924cf7c | 1374 | static const byte Td4[256] = |
wolfSSL | 15:117db924cf7c | 1375 | { |
wolfSSL | 15:117db924cf7c | 1376 | 0x52U, 0x09U, 0x6aU, 0xd5U, 0x30U, 0x36U, 0xa5U, 0x38U, |
wolfSSL | 15:117db924cf7c | 1377 | 0xbfU, 0x40U, 0xa3U, 0x9eU, 0x81U, 0xf3U, 0xd7U, 0xfbU, |
wolfSSL | 15:117db924cf7c | 1378 | 0x7cU, 0xe3U, 0x39U, 0x82U, 0x9bU, 0x2fU, 0xffU, 0x87U, |
wolfSSL | 15:117db924cf7c | 1379 | 0x34U, 0x8eU, 0x43U, 0x44U, 0xc4U, 0xdeU, 0xe9U, 0xcbU, |
wolfSSL | 15:117db924cf7c | 1380 | 0x54U, 0x7bU, 0x94U, 0x32U, 0xa6U, 0xc2U, 0x23U, 0x3dU, |
wolfSSL | 15:117db924cf7c | 1381 | 0xeeU, 0x4cU, 0x95U, 0x0bU, 0x42U, 0xfaU, 0xc3U, 0x4eU, |
wolfSSL | 15:117db924cf7c | 1382 | 0x08U, 0x2eU, 0xa1U, 0x66U, 0x28U, 0xd9U, 0x24U, 0xb2U, |
wolfSSL | 15:117db924cf7c | 1383 | 0x76U, 0x5bU, 0xa2U, 0x49U, 0x6dU, 0x8bU, 0xd1U, 0x25U, |
wolfSSL | 15:117db924cf7c | 1384 | 0x72U, 0xf8U, 0xf6U, 0x64U, 0x86U, 0x68U, 0x98U, 0x16U, |
wolfSSL | 15:117db924cf7c | 1385 | 0xd4U, 0xa4U, 0x5cU, 0xccU, 0x5dU, 0x65U, 0xb6U, 0x92U, |
wolfSSL | 15:117db924cf7c | 1386 | 0x6cU, 0x70U, 0x48U, 0x50U, 0xfdU, 0xedU, 0xb9U, 0xdaU, |
wolfSSL | 15:117db924cf7c | 1387 | 0x5eU, 0x15U, 0x46U, 0x57U, 0xa7U, 0x8dU, 0x9dU, 0x84U, |
wolfSSL | 15:117db924cf7c | 1388 | 0x90U, 0xd8U, 0xabU, 0x00U, 0x8cU, 0xbcU, 0xd3U, 0x0aU, |
wolfSSL | 15:117db924cf7c | 1389 | 0xf7U, 0xe4U, 0x58U, 0x05U, 0xb8U, 0xb3U, 0x45U, 0x06U, |
wolfSSL | 15:117db924cf7c | 1390 | 0xd0U, 0x2cU, 0x1eU, 0x8fU, 0xcaU, 0x3fU, 0x0fU, 0x02U, |
wolfSSL | 15:117db924cf7c | 1391 | 0xc1U, 0xafU, 0xbdU, 0x03U, 0x01U, 0x13U, 0x8aU, 0x6bU, |
wolfSSL | 15:117db924cf7c | 1392 | 0x3aU, 0x91U, 0x11U, 0x41U, 0x4fU, 0x67U, 0xdcU, 0xeaU, |
wolfSSL | 15:117db924cf7c | 1393 | 0x97U, 0xf2U, 0xcfU, 0xceU, 0xf0U, 0xb4U, 0xe6U, 0x73U, |
wolfSSL | 15:117db924cf7c | 1394 | 0x96U, 0xacU, 0x74U, 0x22U, 0xe7U, 0xadU, 0x35U, 0x85U, |
wolfSSL | 15:117db924cf7c | 1395 | 0xe2U, 0xf9U, 0x37U, 0xe8U, 0x1cU, 0x75U, 0xdfU, 0x6eU, |
wolfSSL | 15:117db924cf7c | 1396 | 0x47U, 0xf1U, 0x1aU, 0x71U, 0x1dU, 0x29U, 0xc5U, 0x89U, |
wolfSSL | 15:117db924cf7c | 1397 | 0x6fU, 0xb7U, 0x62U, 0x0eU, 0xaaU, 0x18U, 0xbeU, 0x1bU, |
wolfSSL | 15:117db924cf7c | 1398 | 0xfcU, 0x56U, 0x3eU, 0x4bU, 0xc6U, 0xd2U, 0x79U, 0x20U, |
wolfSSL | 15:117db924cf7c | 1399 | 0x9aU, 0xdbU, 0xc0U, 0xfeU, 0x78U, 0xcdU, 0x5aU, 0xf4U, |
wolfSSL | 15:117db924cf7c | 1400 | 0x1fU, 0xddU, 0xa8U, 0x33U, 0x88U, 0x07U, 0xc7U, 0x31U, |
wolfSSL | 15:117db924cf7c | 1401 | 0xb1U, 0x12U, 0x10U, 0x59U, 0x27U, 0x80U, 0xecU, 0x5fU, |
wolfSSL | 15:117db924cf7c | 1402 | 0x60U, 0x51U, 0x7fU, 0xa9U, 0x19U, 0xb5U, 0x4aU, 0x0dU, |
wolfSSL | 15:117db924cf7c | 1403 | 0x2dU, 0xe5U, 0x7aU, 0x9fU, 0x93U, 0xc9U, 0x9cU, 0xefU, |
wolfSSL | 15:117db924cf7c | 1404 | 0xa0U, 0xe0U, 0x3bU, 0x4dU, 0xaeU, 0x2aU, 0xf5U, 0xb0U, |
wolfSSL | 15:117db924cf7c | 1405 | 0xc8U, 0xebU, 0xbbU, 0x3cU, 0x83U, 0x53U, 0x99U, 0x61U, |
wolfSSL | 15:117db924cf7c | 1406 | 0x17U, 0x2bU, 0x04U, 0x7eU, 0xbaU, 0x77U, 0xd6U, 0x26U, |
wolfSSL | 15:117db924cf7c | 1407 | 0xe1U, 0x69U, 0x14U, 0x63U, 0x55U, 0x21U, 0x0cU, 0x7dU, |
wolfSSL | 15:117db924cf7c | 1408 | }; |
wolfSSL | 15:117db924cf7c | 1409 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 1410 | |
wolfSSL | 15:117db924cf7c | 1411 | #define GETBYTE(x, y) (word32)((byte)((x) >> (8 * (y)))) |
wolfSSL | 15:117db924cf7c | 1412 | |
wolfSSL | 15:117db924cf7c | 1413 | |
wolfSSL | 15:117db924cf7c | 1414 | |
wolfSSL | 15:117db924cf7c | 1415 | #if defined(HAVE_AES_CBC) || defined(WOLFSSL_AES_DIRECT) || defined(HAVE_AESGCM) |
wolfSSL | 15:117db924cf7c | 1416 | |
wolfSSL | 15:117db924cf7c | 1417 | #ifndef WC_CACHE_LINE_SZ |
wolfSSL | 15:117db924cf7c | 1418 | #if defined(__x86_64__) || defined(_M_X64) || \ |
wolfSSL | 15:117db924cf7c | 1419 | (defined(__ILP32__) && (__ILP32__ >= 1)) |
wolfSSL | 15:117db924cf7c | 1420 | #define WC_CACHE_LINE_SZ 64 |
wolfSSL | 15:117db924cf7c | 1421 | #else |
wolfSSL | 15:117db924cf7c | 1422 | /* default cache line size */ |
wolfSSL | 15:117db924cf7c | 1423 | #define WC_CACHE_LINE_SZ 32 |
wolfSSL | 15:117db924cf7c | 1424 | #endif |
wolfSSL | 15:117db924cf7c | 1425 | #endif |
wolfSSL | 15:117db924cf7c | 1426 | |
wolfSSL | 15:117db924cf7c | 1427 | |
wolfSSL | 15:117db924cf7c | 1428 | /* load 4 Te Tables into cache by cache line stride */ |
wolfSSL | 15:117db924cf7c | 1429 | static WC_INLINE word32 PreFetchTe(void) |
wolfSSL | 15:117db924cf7c | 1430 | { |
wolfSSL | 15:117db924cf7c | 1431 | word32 x = 0; |
wolfSSL | 15:117db924cf7c | 1432 | int i,j; |
wolfSSL | 15:117db924cf7c | 1433 | |
wolfSSL | 15:117db924cf7c | 1434 | for (i = 0; i < 4; i++) { |
wolfSSL | 15:117db924cf7c | 1435 | /* 256 elements, each one is 4 bytes */ |
wolfSSL | 15:117db924cf7c | 1436 | for (j = 0; j < 256; j += WC_CACHE_LINE_SZ/4) { |
wolfSSL | 15:117db924cf7c | 1437 | x &= Te[i][j]; |
wolfSSL | 15:117db924cf7c | 1438 | } |
wolfSSL | 15:117db924cf7c | 1439 | } |
wolfSSL | 15:117db924cf7c | 1440 | return x; |
wolfSSL | 15:117db924cf7c | 1441 | } |
wolfSSL | 15:117db924cf7c | 1442 | |
wolfSSL | 15:117db924cf7c | 1443 | |
wolfSSL | 15:117db924cf7c | 1444 | static void wc_AesEncrypt(Aes* aes, const byte* inBlock, byte* outBlock) |
wolfSSL | 15:117db924cf7c | 1445 | { |
wolfSSL | 15:117db924cf7c | 1446 | word32 s0, s1, s2, s3; |
wolfSSL | 15:117db924cf7c | 1447 | word32 t0, t1, t2, t3; |
wolfSSL | 15:117db924cf7c | 1448 | word32 r = aes->rounds >> 1; |
wolfSSL | 15:117db924cf7c | 1449 | const word32* rk = aes->key; |
wolfSSL | 15:117db924cf7c | 1450 | |
wolfSSL | 15:117db924cf7c | 1451 | if (r > 7 || r == 0) { |
wolfSSL | 15:117db924cf7c | 1452 | WOLFSSL_MSG("AesEncrypt encountered improper key, set it up"); |
wolfSSL | 15:117db924cf7c | 1453 | return; /* stop instead of segfaulting, set up your keys! */ |
wolfSSL | 15:117db924cf7c | 1454 | } |
wolfSSL | 15:117db924cf7c | 1455 | |
wolfSSL | 15:117db924cf7c | 1456 | #ifdef WOLFSSL_AESNI |
wolfSSL | 15:117db924cf7c | 1457 | if (haveAESNI && aes->use_aesni) { |
wolfSSL | 15:117db924cf7c | 1458 | #ifdef DEBUG_AESNI |
wolfSSL | 15:117db924cf7c | 1459 | printf("about to aes encrypt\n"); |
wolfSSL | 15:117db924cf7c | 1460 | printf("in = %p\n", inBlock); |
wolfSSL | 15:117db924cf7c | 1461 | printf("out = %p\n", outBlock); |
wolfSSL | 15:117db924cf7c | 1462 | printf("aes->key = %p\n", aes->key); |
wolfSSL | 15:117db924cf7c | 1463 | printf("aes->rounds = %d\n", aes->rounds); |
wolfSSL | 15:117db924cf7c | 1464 | printf("sz = %d\n", AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 1465 | #endif |
wolfSSL | 15:117db924cf7c | 1466 | |
wolfSSL | 15:117db924cf7c | 1467 | /* check alignment, decrypt doesn't need alignment */ |
wolfSSL | 15:117db924cf7c | 1468 | if ((wolfssl_word)inBlock % AESNI_ALIGN) { |
wolfSSL | 15:117db924cf7c | 1469 | #ifndef NO_WOLFSSL_ALLOC_ALIGN |
wolfSSL | 15:117db924cf7c | 1470 | byte* tmp = (byte*)XMALLOC(AES_BLOCK_SIZE + AESNI_ALIGN, aes->heap, |
wolfSSL | 15:117db924cf7c | 1471 | DYNAMIC_TYPE_TMP_BUFFER); |
wolfSSL | 15:117db924cf7c | 1472 | byte* tmp_align; |
wolfSSL | 15:117db924cf7c | 1473 | if (tmp == NULL) return; |
wolfSSL | 15:117db924cf7c | 1474 | |
wolfSSL | 15:117db924cf7c | 1475 | tmp_align = tmp + (AESNI_ALIGN - ((size_t)tmp % AESNI_ALIGN)); |
wolfSSL | 15:117db924cf7c | 1476 | |
wolfSSL | 15:117db924cf7c | 1477 | XMEMCPY(tmp_align, inBlock, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 1478 | AES_ECB_encrypt(tmp_align, tmp_align, AES_BLOCK_SIZE, (byte*)aes->key, |
wolfSSL | 15:117db924cf7c | 1479 | aes->rounds); |
wolfSSL | 15:117db924cf7c | 1480 | XMEMCPY(outBlock, tmp_align, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 1481 | XFREE(tmp, aes->heap, DYNAMIC_TYPE_TMP_BUFFER); |
wolfSSL | 15:117db924cf7c | 1482 | return; |
wolfSSL | 15:117db924cf7c | 1483 | #else |
wolfSSL | 15:117db924cf7c | 1484 | WOLFSSL_MSG("AES-ECB encrypt with bad alignment"); |
wolfSSL | 15:117db924cf7c | 1485 | return; |
wolfSSL | 15:117db924cf7c | 1486 | #endif |
wolfSSL | 15:117db924cf7c | 1487 | } |
wolfSSL | 15:117db924cf7c | 1488 | |
wolfSSL | 15:117db924cf7c | 1489 | AES_ECB_encrypt(inBlock, outBlock, AES_BLOCK_SIZE, (byte*)aes->key, |
wolfSSL | 15:117db924cf7c | 1490 | aes->rounds); |
wolfSSL | 15:117db924cf7c | 1491 | |
wolfSSL | 15:117db924cf7c | 1492 | return; |
wolfSSL | 15:117db924cf7c | 1493 | } |
wolfSSL | 15:117db924cf7c | 1494 | else { |
wolfSSL | 15:117db924cf7c | 1495 | #ifdef DEBUG_AESNI |
wolfSSL | 15:117db924cf7c | 1496 | printf("Skipping AES-NI\n"); |
wolfSSL | 15:117db924cf7c | 1497 | #endif |
wolfSSL | 15:117db924cf7c | 1498 | } |
wolfSSL | 15:117db924cf7c | 1499 | #endif |
wolfSSL | 15:117db924cf7c | 1500 | |
wolfSSL | 15:117db924cf7c | 1501 | /* |
wolfSSL | 15:117db924cf7c | 1502 | * map byte array block to cipher state |
wolfSSL | 15:117db924cf7c | 1503 | * and add initial round key: |
wolfSSL | 15:117db924cf7c | 1504 | */ |
wolfSSL | 15:117db924cf7c | 1505 | XMEMCPY(&s0, inBlock, sizeof(s0)); |
wolfSSL | 15:117db924cf7c | 1506 | XMEMCPY(&s1, inBlock + sizeof(s0), sizeof(s1)); |
wolfSSL | 15:117db924cf7c | 1507 | XMEMCPY(&s2, inBlock + 2 * sizeof(s0), sizeof(s2)); |
wolfSSL | 15:117db924cf7c | 1508 | XMEMCPY(&s3, inBlock + 3 * sizeof(s0), sizeof(s3)); |
wolfSSL | 15:117db924cf7c | 1509 | |
wolfSSL | 15:117db924cf7c | 1510 | #ifdef LITTLE_ENDIAN_ORDER |
wolfSSL | 15:117db924cf7c | 1511 | s0 = ByteReverseWord32(s0); |
wolfSSL | 15:117db924cf7c | 1512 | s1 = ByteReverseWord32(s1); |
wolfSSL | 15:117db924cf7c | 1513 | s2 = ByteReverseWord32(s2); |
wolfSSL | 15:117db924cf7c | 1514 | s3 = ByteReverseWord32(s3); |
wolfSSL | 15:117db924cf7c | 1515 | #endif |
wolfSSL | 15:117db924cf7c | 1516 | |
wolfSSL | 15:117db924cf7c | 1517 | s0 ^= rk[0]; |
wolfSSL | 15:117db924cf7c | 1518 | s1 ^= rk[1]; |
wolfSSL | 15:117db924cf7c | 1519 | s2 ^= rk[2]; |
wolfSSL | 15:117db924cf7c | 1520 | s3 ^= rk[3]; |
wolfSSL | 15:117db924cf7c | 1521 | |
wolfSSL | 15:117db924cf7c | 1522 | s0 |= PreFetchTe(); |
wolfSSL | 15:117db924cf7c | 1523 | |
wolfSSL | 15:117db924cf7c | 1524 | /* |
wolfSSL | 15:117db924cf7c | 1525 | * Nr - 1 full rounds: |
wolfSSL | 15:117db924cf7c | 1526 | */ |
wolfSSL | 15:117db924cf7c | 1527 | |
wolfSSL | 15:117db924cf7c | 1528 | for (;;) { |
wolfSSL | 15:117db924cf7c | 1529 | t0 = |
wolfSSL | 15:117db924cf7c | 1530 | Te[0][GETBYTE(s0, 3)] ^ |
wolfSSL | 15:117db924cf7c | 1531 | Te[1][GETBYTE(s1, 2)] ^ |
wolfSSL | 15:117db924cf7c | 1532 | Te[2][GETBYTE(s2, 1)] ^ |
wolfSSL | 15:117db924cf7c | 1533 | Te[3][GETBYTE(s3, 0)] ^ |
wolfSSL | 15:117db924cf7c | 1534 | rk[4]; |
wolfSSL | 15:117db924cf7c | 1535 | t1 = |
wolfSSL | 15:117db924cf7c | 1536 | Te[0][GETBYTE(s1, 3)] ^ |
wolfSSL | 15:117db924cf7c | 1537 | Te[1][GETBYTE(s2, 2)] ^ |
wolfSSL | 15:117db924cf7c | 1538 | Te[2][GETBYTE(s3, 1)] ^ |
wolfSSL | 15:117db924cf7c | 1539 | Te[3][GETBYTE(s0, 0)] ^ |
wolfSSL | 15:117db924cf7c | 1540 | rk[5]; |
wolfSSL | 15:117db924cf7c | 1541 | t2 = |
wolfSSL | 15:117db924cf7c | 1542 | Te[0][GETBYTE(s2, 3)] ^ |
wolfSSL | 15:117db924cf7c | 1543 | Te[1][GETBYTE(s3, 2)] ^ |
wolfSSL | 15:117db924cf7c | 1544 | Te[2][GETBYTE(s0, 1)] ^ |
wolfSSL | 15:117db924cf7c | 1545 | Te[3][GETBYTE(s1, 0)] ^ |
wolfSSL | 15:117db924cf7c | 1546 | rk[6]; |
wolfSSL | 15:117db924cf7c | 1547 | t3 = |
wolfSSL | 15:117db924cf7c | 1548 | Te[0][GETBYTE(s3, 3)] ^ |
wolfSSL | 15:117db924cf7c | 1549 | Te[1][GETBYTE(s0, 2)] ^ |
wolfSSL | 15:117db924cf7c | 1550 | Te[2][GETBYTE(s1, 1)] ^ |
wolfSSL | 15:117db924cf7c | 1551 | Te[3][GETBYTE(s2, 0)] ^ |
wolfSSL | 15:117db924cf7c | 1552 | rk[7]; |
wolfSSL | 15:117db924cf7c | 1553 | |
wolfSSL | 15:117db924cf7c | 1554 | rk += 8; |
wolfSSL | 15:117db924cf7c | 1555 | if (--r == 0) { |
wolfSSL | 15:117db924cf7c | 1556 | break; |
wolfSSL | 15:117db924cf7c | 1557 | } |
wolfSSL | 15:117db924cf7c | 1558 | |
wolfSSL | 15:117db924cf7c | 1559 | s0 = |
wolfSSL | 15:117db924cf7c | 1560 | Te[0][GETBYTE(t0, 3)] ^ |
wolfSSL | 15:117db924cf7c | 1561 | Te[1][GETBYTE(t1, 2)] ^ |
wolfSSL | 15:117db924cf7c | 1562 | Te[2][GETBYTE(t2, 1)] ^ |
wolfSSL | 15:117db924cf7c | 1563 | Te[3][GETBYTE(t3, 0)] ^ |
wolfSSL | 15:117db924cf7c | 1564 | rk[0]; |
wolfSSL | 15:117db924cf7c | 1565 | s1 = |
wolfSSL | 15:117db924cf7c | 1566 | Te[0][GETBYTE(t1, 3)] ^ |
wolfSSL | 15:117db924cf7c | 1567 | Te[1][GETBYTE(t2, 2)] ^ |
wolfSSL | 15:117db924cf7c | 1568 | Te[2][GETBYTE(t3, 1)] ^ |
wolfSSL | 15:117db924cf7c | 1569 | Te[3][GETBYTE(t0, 0)] ^ |
wolfSSL | 15:117db924cf7c | 1570 | rk[1]; |
wolfSSL | 15:117db924cf7c | 1571 | s2 = |
wolfSSL | 15:117db924cf7c | 1572 | Te[0][GETBYTE(t2, 3)] ^ |
wolfSSL | 15:117db924cf7c | 1573 | Te[1][GETBYTE(t3, 2)] ^ |
wolfSSL | 15:117db924cf7c | 1574 | Te[2][GETBYTE(t0, 1)] ^ |
wolfSSL | 15:117db924cf7c | 1575 | Te[3][GETBYTE(t1, 0)] ^ |
wolfSSL | 15:117db924cf7c | 1576 | rk[2]; |
wolfSSL | 15:117db924cf7c | 1577 | s3 = |
wolfSSL | 15:117db924cf7c | 1578 | Te[0][GETBYTE(t3, 3)] ^ |
wolfSSL | 15:117db924cf7c | 1579 | Te[1][GETBYTE(t0, 2)] ^ |
wolfSSL | 15:117db924cf7c | 1580 | Te[2][GETBYTE(t1, 1)] ^ |
wolfSSL | 15:117db924cf7c | 1581 | Te[3][GETBYTE(t2, 0)] ^ |
wolfSSL | 15:117db924cf7c | 1582 | rk[3]; |
wolfSSL | 15:117db924cf7c | 1583 | } |
wolfSSL | 15:117db924cf7c | 1584 | |
wolfSSL | 15:117db924cf7c | 1585 | /* |
wolfSSL | 15:117db924cf7c | 1586 | * apply last round and |
wolfSSL | 15:117db924cf7c | 1587 | * map cipher state to byte array block: |
wolfSSL | 15:117db924cf7c | 1588 | */ |
wolfSSL | 15:117db924cf7c | 1589 | |
wolfSSL | 15:117db924cf7c | 1590 | s0 = |
wolfSSL | 15:117db924cf7c | 1591 | (Te[2][GETBYTE(t0, 3)] & 0xff000000) ^ |
wolfSSL | 15:117db924cf7c | 1592 | (Te[3][GETBYTE(t1, 2)] & 0x00ff0000) ^ |
wolfSSL | 15:117db924cf7c | 1593 | (Te[0][GETBYTE(t2, 1)] & 0x0000ff00) ^ |
wolfSSL | 15:117db924cf7c | 1594 | (Te[1][GETBYTE(t3, 0)] & 0x000000ff) ^ |
wolfSSL | 15:117db924cf7c | 1595 | rk[0]; |
wolfSSL | 15:117db924cf7c | 1596 | s1 = |
wolfSSL | 15:117db924cf7c | 1597 | (Te[2][GETBYTE(t1, 3)] & 0xff000000) ^ |
wolfSSL | 15:117db924cf7c | 1598 | (Te[3][GETBYTE(t2, 2)] & 0x00ff0000) ^ |
wolfSSL | 15:117db924cf7c | 1599 | (Te[0][GETBYTE(t3, 1)] & 0x0000ff00) ^ |
wolfSSL | 15:117db924cf7c | 1600 | (Te[1][GETBYTE(t0, 0)] & 0x000000ff) ^ |
wolfSSL | 15:117db924cf7c | 1601 | rk[1]; |
wolfSSL | 15:117db924cf7c | 1602 | s2 = |
wolfSSL | 15:117db924cf7c | 1603 | (Te[2][GETBYTE(t2, 3)] & 0xff000000) ^ |
wolfSSL | 15:117db924cf7c | 1604 | (Te[3][GETBYTE(t3, 2)] & 0x00ff0000) ^ |
wolfSSL | 15:117db924cf7c | 1605 | (Te[0][GETBYTE(t0, 1)] & 0x0000ff00) ^ |
wolfSSL | 15:117db924cf7c | 1606 | (Te[1][GETBYTE(t1, 0)] & 0x000000ff) ^ |
wolfSSL | 15:117db924cf7c | 1607 | rk[2]; |
wolfSSL | 15:117db924cf7c | 1608 | s3 = |
wolfSSL | 15:117db924cf7c | 1609 | (Te[2][GETBYTE(t3, 3)] & 0xff000000) ^ |
wolfSSL | 15:117db924cf7c | 1610 | (Te[3][GETBYTE(t0, 2)] & 0x00ff0000) ^ |
wolfSSL | 15:117db924cf7c | 1611 | (Te[0][GETBYTE(t1, 1)] & 0x0000ff00) ^ |
wolfSSL | 15:117db924cf7c | 1612 | (Te[1][GETBYTE(t2, 0)] & 0x000000ff) ^ |
wolfSSL | 15:117db924cf7c | 1613 | rk[3]; |
wolfSSL | 15:117db924cf7c | 1614 | |
wolfSSL | 15:117db924cf7c | 1615 | /* write out */ |
wolfSSL | 15:117db924cf7c | 1616 | #ifdef LITTLE_ENDIAN_ORDER |
wolfSSL | 15:117db924cf7c | 1617 | s0 = ByteReverseWord32(s0); |
wolfSSL | 15:117db924cf7c | 1618 | s1 = ByteReverseWord32(s1); |
wolfSSL | 15:117db924cf7c | 1619 | s2 = ByteReverseWord32(s2); |
wolfSSL | 15:117db924cf7c | 1620 | s3 = ByteReverseWord32(s3); |
wolfSSL | 15:117db924cf7c | 1621 | #endif |
wolfSSL | 15:117db924cf7c | 1622 | |
wolfSSL | 15:117db924cf7c | 1623 | XMEMCPY(outBlock, &s0, sizeof(s0)); |
wolfSSL | 15:117db924cf7c | 1624 | XMEMCPY(outBlock + sizeof(s0), &s1, sizeof(s1)); |
wolfSSL | 15:117db924cf7c | 1625 | XMEMCPY(outBlock + 2 * sizeof(s0), &s2, sizeof(s2)); |
wolfSSL | 15:117db924cf7c | 1626 | XMEMCPY(outBlock + 3 * sizeof(s0), &s3, sizeof(s3)); |
wolfSSL | 15:117db924cf7c | 1627 | |
wolfSSL | 15:117db924cf7c | 1628 | } |
wolfSSL | 15:117db924cf7c | 1629 | #endif /* HAVE_AES_CBC || WOLFSSL_AES_DIRECT || HAVE_AESGCM */ |
wolfSSL | 15:117db924cf7c | 1630 | |
wolfSSL | 15:117db924cf7c | 1631 | #if defined(HAVE_AES_DECRYPT) |
wolfSSL | 15:117db924cf7c | 1632 | #if defined(HAVE_AES_CBC) || defined(WOLFSSL_AES_DIRECT) |
wolfSSL | 15:117db924cf7c | 1633 | |
wolfSSL | 15:117db924cf7c | 1634 | /* load 4 Td Tables into cache by cache line stride */ |
wolfSSL | 15:117db924cf7c | 1635 | static WC_INLINE word32 PreFetchTd(void) |
wolfSSL | 15:117db924cf7c | 1636 | { |
wolfSSL | 15:117db924cf7c | 1637 | word32 x = 0; |
wolfSSL | 15:117db924cf7c | 1638 | int i,j; |
wolfSSL | 15:117db924cf7c | 1639 | |
wolfSSL | 15:117db924cf7c | 1640 | for (i = 0; i < 4; i++) { |
wolfSSL | 15:117db924cf7c | 1641 | /* 256 elements, each one is 4 bytes */ |
wolfSSL | 15:117db924cf7c | 1642 | for (j = 0; j < 256; j += WC_CACHE_LINE_SZ/4) { |
wolfSSL | 15:117db924cf7c | 1643 | x &= Td[i][j]; |
wolfSSL | 15:117db924cf7c | 1644 | } |
wolfSSL | 15:117db924cf7c | 1645 | } |
wolfSSL | 15:117db924cf7c | 1646 | return x; |
wolfSSL | 15:117db924cf7c | 1647 | } |
wolfSSL | 15:117db924cf7c | 1648 | |
wolfSSL | 15:117db924cf7c | 1649 | /* load Td Table4 into cache by cache line stride */ |
wolfSSL | 15:117db924cf7c | 1650 | static WC_INLINE word32 PreFetchTd4(void) |
wolfSSL | 15:117db924cf7c | 1651 | { |
wolfSSL | 15:117db924cf7c | 1652 | word32 x = 0; |
wolfSSL | 15:117db924cf7c | 1653 | int i; |
wolfSSL | 15:117db924cf7c | 1654 | |
wolfSSL | 15:117db924cf7c | 1655 | for (i = 0; i < 256; i += WC_CACHE_LINE_SZ) { |
wolfSSL | 15:117db924cf7c | 1656 | x &= (word32)Td4[i]; |
wolfSSL | 15:117db924cf7c | 1657 | } |
wolfSSL | 15:117db924cf7c | 1658 | return x; |
wolfSSL | 15:117db924cf7c | 1659 | } |
wolfSSL | 15:117db924cf7c | 1660 | |
wolfSSL | 15:117db924cf7c | 1661 | static void wc_AesDecrypt(Aes* aes, const byte* inBlock, byte* outBlock) |
wolfSSL | 15:117db924cf7c | 1662 | { |
wolfSSL | 15:117db924cf7c | 1663 | word32 s0, s1, s2, s3; |
wolfSSL | 15:117db924cf7c | 1664 | word32 t0, t1, t2, t3; |
wolfSSL | 15:117db924cf7c | 1665 | word32 r = aes->rounds >> 1; |
wolfSSL | 15:117db924cf7c | 1666 | |
wolfSSL | 15:117db924cf7c | 1667 | const word32* rk = aes->key; |
wolfSSL | 15:117db924cf7c | 1668 | if (r > 7 || r == 0) { |
wolfSSL | 15:117db924cf7c | 1669 | WOLFSSL_MSG("AesDecrypt encountered improper key, set it up"); |
wolfSSL | 15:117db924cf7c | 1670 | return; /* stop instead of segfaulting, set up your keys! */ |
wolfSSL | 15:117db924cf7c | 1671 | } |
wolfSSL | 15:117db924cf7c | 1672 | #ifdef WOLFSSL_AESNI |
wolfSSL | 15:117db924cf7c | 1673 | if (haveAESNI && aes->use_aesni) { |
wolfSSL | 15:117db924cf7c | 1674 | #ifdef DEBUG_AESNI |
wolfSSL | 15:117db924cf7c | 1675 | printf("about to aes decrypt\n"); |
wolfSSL | 15:117db924cf7c | 1676 | printf("in = %p\n", inBlock); |
wolfSSL | 15:117db924cf7c | 1677 | printf("out = %p\n", outBlock); |
wolfSSL | 15:117db924cf7c | 1678 | printf("aes->key = %p\n", aes->key); |
wolfSSL | 15:117db924cf7c | 1679 | printf("aes->rounds = %d\n", aes->rounds); |
wolfSSL | 15:117db924cf7c | 1680 | printf("sz = %d\n", AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 1681 | #endif |
wolfSSL | 15:117db924cf7c | 1682 | |
wolfSSL | 15:117db924cf7c | 1683 | /* if input and output same will overwrite input iv */ |
wolfSSL | 15:117db924cf7c | 1684 | XMEMCPY(aes->tmp, inBlock, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 1685 | AES_ECB_decrypt(inBlock, outBlock, AES_BLOCK_SIZE, (byte*)aes->key, |
wolfSSL | 15:117db924cf7c | 1686 | aes->rounds); |
wolfSSL | 15:117db924cf7c | 1687 | return; |
wolfSSL | 15:117db924cf7c | 1688 | } |
wolfSSL | 15:117db924cf7c | 1689 | else { |
wolfSSL | 15:117db924cf7c | 1690 | #ifdef DEBUG_AESNI |
wolfSSL | 15:117db924cf7c | 1691 | printf("Skipping AES-NI\n"); |
wolfSSL | 15:117db924cf7c | 1692 | #endif |
wolfSSL | 15:117db924cf7c | 1693 | } |
wolfSSL | 15:117db924cf7c | 1694 | #endif /* WOLFSSL_AESNI */ |
wolfSSL | 15:117db924cf7c | 1695 | |
wolfSSL | 15:117db924cf7c | 1696 | /* |
wolfSSL | 15:117db924cf7c | 1697 | * map byte array block to cipher state |
wolfSSL | 15:117db924cf7c | 1698 | * and add initial round key: |
wolfSSL | 15:117db924cf7c | 1699 | */ |
wolfSSL | 15:117db924cf7c | 1700 | XMEMCPY(&s0, inBlock, sizeof(s0)); |
wolfSSL | 15:117db924cf7c | 1701 | XMEMCPY(&s1, inBlock + sizeof(s0), sizeof(s1)); |
wolfSSL | 15:117db924cf7c | 1702 | XMEMCPY(&s2, inBlock + 2 * sizeof(s0), sizeof(s2)); |
wolfSSL | 15:117db924cf7c | 1703 | XMEMCPY(&s3, inBlock + 3 * sizeof(s0), sizeof(s3)); |
wolfSSL | 15:117db924cf7c | 1704 | |
wolfSSL | 15:117db924cf7c | 1705 | #ifdef LITTLE_ENDIAN_ORDER |
wolfSSL | 15:117db924cf7c | 1706 | s0 = ByteReverseWord32(s0); |
wolfSSL | 15:117db924cf7c | 1707 | s1 = ByteReverseWord32(s1); |
wolfSSL | 15:117db924cf7c | 1708 | s2 = ByteReverseWord32(s2); |
wolfSSL | 15:117db924cf7c | 1709 | s3 = ByteReverseWord32(s3); |
wolfSSL | 15:117db924cf7c | 1710 | #endif |
wolfSSL | 15:117db924cf7c | 1711 | |
wolfSSL | 15:117db924cf7c | 1712 | s0 ^= rk[0]; |
wolfSSL | 15:117db924cf7c | 1713 | s1 ^= rk[1]; |
wolfSSL | 15:117db924cf7c | 1714 | s2 ^= rk[2]; |
wolfSSL | 15:117db924cf7c | 1715 | s3 ^= rk[3]; |
wolfSSL | 15:117db924cf7c | 1716 | |
wolfSSL | 15:117db924cf7c | 1717 | s0 |= PreFetchTd(); |
wolfSSL | 15:117db924cf7c | 1718 | |
wolfSSL | 15:117db924cf7c | 1719 | /* |
wolfSSL | 15:117db924cf7c | 1720 | * Nr - 1 full rounds: |
wolfSSL | 15:117db924cf7c | 1721 | */ |
wolfSSL | 15:117db924cf7c | 1722 | |
wolfSSL | 15:117db924cf7c | 1723 | for (;;) { |
wolfSSL | 15:117db924cf7c | 1724 | t0 = |
wolfSSL | 15:117db924cf7c | 1725 | Td[0][GETBYTE(s0, 3)] ^ |
wolfSSL | 15:117db924cf7c | 1726 | Td[1][GETBYTE(s3, 2)] ^ |
wolfSSL | 15:117db924cf7c | 1727 | Td[2][GETBYTE(s2, 1)] ^ |
wolfSSL | 15:117db924cf7c | 1728 | Td[3][GETBYTE(s1, 0)] ^ |
wolfSSL | 15:117db924cf7c | 1729 | rk[4]; |
wolfSSL | 15:117db924cf7c | 1730 | t1 = |
wolfSSL | 15:117db924cf7c | 1731 | Td[0][GETBYTE(s1, 3)] ^ |
wolfSSL | 15:117db924cf7c | 1732 | Td[1][GETBYTE(s0, 2)] ^ |
wolfSSL | 15:117db924cf7c | 1733 | Td[2][GETBYTE(s3, 1)] ^ |
wolfSSL | 15:117db924cf7c | 1734 | Td[3][GETBYTE(s2, 0)] ^ |
wolfSSL | 15:117db924cf7c | 1735 | rk[5]; |
wolfSSL | 15:117db924cf7c | 1736 | t2 = |
wolfSSL | 15:117db924cf7c | 1737 | Td[0][GETBYTE(s2, 3)] ^ |
wolfSSL | 15:117db924cf7c | 1738 | Td[1][GETBYTE(s1, 2)] ^ |
wolfSSL | 15:117db924cf7c | 1739 | Td[2][GETBYTE(s0, 1)] ^ |
wolfSSL | 15:117db924cf7c | 1740 | Td[3][GETBYTE(s3, 0)] ^ |
wolfSSL | 15:117db924cf7c | 1741 | rk[6]; |
wolfSSL | 15:117db924cf7c | 1742 | t3 = |
wolfSSL | 15:117db924cf7c | 1743 | Td[0][GETBYTE(s3, 3)] ^ |
wolfSSL | 15:117db924cf7c | 1744 | Td[1][GETBYTE(s2, 2)] ^ |
wolfSSL | 15:117db924cf7c | 1745 | Td[2][GETBYTE(s1, 1)] ^ |
wolfSSL | 15:117db924cf7c | 1746 | Td[3][GETBYTE(s0, 0)] ^ |
wolfSSL | 15:117db924cf7c | 1747 | rk[7]; |
wolfSSL | 15:117db924cf7c | 1748 | |
wolfSSL | 15:117db924cf7c | 1749 | rk += 8; |
wolfSSL | 15:117db924cf7c | 1750 | if (--r == 0) { |
wolfSSL | 15:117db924cf7c | 1751 | break; |
wolfSSL | 15:117db924cf7c | 1752 | } |
wolfSSL | 15:117db924cf7c | 1753 | |
wolfSSL | 15:117db924cf7c | 1754 | s0 = |
wolfSSL | 15:117db924cf7c | 1755 | Td[0][GETBYTE(t0, 3)] ^ |
wolfSSL | 15:117db924cf7c | 1756 | Td[1][GETBYTE(t3, 2)] ^ |
wolfSSL | 15:117db924cf7c | 1757 | Td[2][GETBYTE(t2, 1)] ^ |
wolfSSL | 15:117db924cf7c | 1758 | Td[3][GETBYTE(t1, 0)] ^ |
wolfSSL | 15:117db924cf7c | 1759 | rk[0]; |
wolfSSL | 15:117db924cf7c | 1760 | s1 = |
wolfSSL | 15:117db924cf7c | 1761 | Td[0][GETBYTE(t1, 3)] ^ |
wolfSSL | 15:117db924cf7c | 1762 | Td[1][GETBYTE(t0, 2)] ^ |
wolfSSL | 15:117db924cf7c | 1763 | Td[2][GETBYTE(t3, 1)] ^ |
wolfSSL | 15:117db924cf7c | 1764 | Td[3][GETBYTE(t2, 0)] ^ |
wolfSSL | 15:117db924cf7c | 1765 | rk[1]; |
wolfSSL | 15:117db924cf7c | 1766 | s2 = |
wolfSSL | 15:117db924cf7c | 1767 | Td[0][GETBYTE(t2, 3)] ^ |
wolfSSL | 15:117db924cf7c | 1768 | Td[1][GETBYTE(t1, 2)] ^ |
wolfSSL | 15:117db924cf7c | 1769 | Td[2][GETBYTE(t0, 1)] ^ |
wolfSSL | 15:117db924cf7c | 1770 | Td[3][GETBYTE(t3, 0)] ^ |
wolfSSL | 15:117db924cf7c | 1771 | rk[2]; |
wolfSSL | 15:117db924cf7c | 1772 | s3 = |
wolfSSL | 15:117db924cf7c | 1773 | Td[0][GETBYTE(t3, 3)] ^ |
wolfSSL | 15:117db924cf7c | 1774 | Td[1][GETBYTE(t2, 2)] ^ |
wolfSSL | 15:117db924cf7c | 1775 | Td[2][GETBYTE(t1, 1)] ^ |
wolfSSL | 15:117db924cf7c | 1776 | Td[3][GETBYTE(t0, 0)] ^ |
wolfSSL | 15:117db924cf7c | 1777 | rk[3]; |
wolfSSL | 15:117db924cf7c | 1778 | } |
wolfSSL | 15:117db924cf7c | 1779 | /* |
wolfSSL | 15:117db924cf7c | 1780 | * apply last round and |
wolfSSL | 15:117db924cf7c | 1781 | * map cipher state to byte array block: |
wolfSSL | 15:117db924cf7c | 1782 | */ |
wolfSSL | 15:117db924cf7c | 1783 | |
wolfSSL | 15:117db924cf7c | 1784 | t0 |= PreFetchTd4(); |
wolfSSL | 15:117db924cf7c | 1785 | |
wolfSSL | 15:117db924cf7c | 1786 | s0 = |
wolfSSL | 15:117db924cf7c | 1787 | ((word32)Td4[GETBYTE(t0, 3)] << 24) ^ |
wolfSSL | 15:117db924cf7c | 1788 | ((word32)Td4[GETBYTE(t3, 2)] << 16) ^ |
wolfSSL | 15:117db924cf7c | 1789 | ((word32)Td4[GETBYTE(t2, 1)] << 8) ^ |
wolfSSL | 15:117db924cf7c | 1790 | ((word32)Td4[GETBYTE(t1, 0)]) ^ |
wolfSSL | 15:117db924cf7c | 1791 | rk[0]; |
wolfSSL | 15:117db924cf7c | 1792 | s1 = |
wolfSSL | 15:117db924cf7c | 1793 | ((word32)Td4[GETBYTE(t1, 3)] << 24) ^ |
wolfSSL | 15:117db924cf7c | 1794 | ((word32)Td4[GETBYTE(t0, 2)] << 16) ^ |
wolfSSL | 15:117db924cf7c | 1795 | ((word32)Td4[GETBYTE(t3, 1)] << 8) ^ |
wolfSSL | 15:117db924cf7c | 1796 | ((word32)Td4[GETBYTE(t2, 0)]) ^ |
wolfSSL | 15:117db924cf7c | 1797 | rk[1]; |
wolfSSL | 15:117db924cf7c | 1798 | s2 = |
wolfSSL | 15:117db924cf7c | 1799 | ((word32)Td4[GETBYTE(t2, 3)] << 24) ^ |
wolfSSL | 15:117db924cf7c | 1800 | ((word32)Td4[GETBYTE(t1, 2)] << 16) ^ |
wolfSSL | 15:117db924cf7c | 1801 | ((word32)Td4[GETBYTE(t0, 1)] << 8) ^ |
wolfSSL | 15:117db924cf7c | 1802 | ((word32)Td4[GETBYTE(t3, 0)]) ^ |
wolfSSL | 15:117db924cf7c | 1803 | rk[2]; |
wolfSSL | 15:117db924cf7c | 1804 | s3 = |
wolfSSL | 15:117db924cf7c | 1805 | ((word32)Td4[GETBYTE(t3, 3)] << 24) ^ |
wolfSSL | 15:117db924cf7c | 1806 | ((word32)Td4[GETBYTE(t2, 2)] << 16) ^ |
wolfSSL | 15:117db924cf7c | 1807 | ((word32)Td4[GETBYTE(t1, 1)] << 8) ^ |
wolfSSL | 15:117db924cf7c | 1808 | ((word32)Td4[GETBYTE(t0, 0)]) ^ |
wolfSSL | 15:117db924cf7c | 1809 | rk[3]; |
wolfSSL | 15:117db924cf7c | 1810 | |
wolfSSL | 15:117db924cf7c | 1811 | /* write out */ |
wolfSSL | 15:117db924cf7c | 1812 | #ifdef LITTLE_ENDIAN_ORDER |
wolfSSL | 15:117db924cf7c | 1813 | s0 = ByteReverseWord32(s0); |
wolfSSL | 15:117db924cf7c | 1814 | s1 = ByteReverseWord32(s1); |
wolfSSL | 15:117db924cf7c | 1815 | s2 = ByteReverseWord32(s2); |
wolfSSL | 15:117db924cf7c | 1816 | s3 = ByteReverseWord32(s3); |
wolfSSL | 15:117db924cf7c | 1817 | #endif |
wolfSSL | 15:117db924cf7c | 1818 | |
wolfSSL | 15:117db924cf7c | 1819 | XMEMCPY(outBlock, &s0, sizeof(s0)); |
wolfSSL | 15:117db924cf7c | 1820 | XMEMCPY(outBlock + sizeof(s0), &s1, sizeof(s1)); |
wolfSSL | 15:117db924cf7c | 1821 | XMEMCPY(outBlock + 2 * sizeof(s0), &s2, sizeof(s2)); |
wolfSSL | 15:117db924cf7c | 1822 | XMEMCPY(outBlock + 3 * sizeof(s0), &s3, sizeof(s3)); |
wolfSSL | 15:117db924cf7c | 1823 | } |
wolfSSL | 15:117db924cf7c | 1824 | #endif /* HAVE_AES_CBC || WOLFSSL_AES_DIRECT */ |
wolfSSL | 15:117db924cf7c | 1825 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 1826 | |
wolfSSL | 15:117db924cf7c | 1827 | #endif /* NEED_AES_TABLES */ |
wolfSSL | 15:117db924cf7c | 1828 | |
wolfSSL | 15:117db924cf7c | 1829 | |
wolfSSL | 15:117db924cf7c | 1830 | |
wolfSSL | 15:117db924cf7c | 1831 | /* wc_AesSetKey */ |
wolfSSL | 15:117db924cf7c | 1832 | #if defined(STM32_CRYPTO) |
wolfSSL | 15:117db924cf7c | 1833 | |
wolfSSL | 15:117db924cf7c | 1834 | int wc_AesSetKey(Aes* aes, const byte* userKey, word32 keylen, |
wolfSSL | 15:117db924cf7c | 1835 | const byte* iv, int dir) |
wolfSSL | 15:117db924cf7c | 1836 | { |
wolfSSL | 15:117db924cf7c | 1837 | word32 *rk = aes->key; |
wolfSSL | 15:117db924cf7c | 1838 | |
wolfSSL | 15:117db924cf7c | 1839 | (void)dir; |
wolfSSL | 15:117db924cf7c | 1840 | |
wolfSSL | 15:117db924cf7c | 1841 | if (!((keylen == 16) || (keylen == 24) || (keylen == 32))) |
wolfSSL | 15:117db924cf7c | 1842 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 1843 | |
wolfSSL | 15:117db924cf7c | 1844 | aes->keylen = keylen; |
wolfSSL | 15:117db924cf7c | 1845 | aes->rounds = keylen/4 + 6; |
wolfSSL | 15:117db924cf7c | 1846 | XMEMCPY(rk, userKey, keylen); |
wolfSSL | 15:117db924cf7c | 1847 | #ifndef WOLFSSL_STM32_CUBEMX |
wolfSSL | 15:117db924cf7c | 1848 | ByteReverseWords(rk, rk, keylen); |
wolfSSL | 15:117db924cf7c | 1849 | #endif |
wolfSSL | 15:117db924cf7c | 1850 | #if defined(WOLFSSL_AES_CFB) || defined(WOLFSSL_AES_COUNTER) |
wolfSSL | 15:117db924cf7c | 1851 | aes->left = 0; |
wolfSSL | 15:117db924cf7c | 1852 | #endif |
wolfSSL | 15:117db924cf7c | 1853 | |
wolfSSL | 15:117db924cf7c | 1854 | return wc_AesSetIV(aes, iv); |
wolfSSL | 15:117db924cf7c | 1855 | } |
wolfSSL | 15:117db924cf7c | 1856 | #if defined(WOLFSSL_AES_DIRECT) |
wolfSSL | 15:117db924cf7c | 1857 | int wc_AesSetKeyDirect(Aes* aes, const byte* userKey, word32 keylen, |
wolfSSL | 15:117db924cf7c | 1858 | const byte* iv, int dir) |
wolfSSL | 15:117db924cf7c | 1859 | { |
wolfSSL | 15:117db924cf7c | 1860 | return wc_AesSetKey(aes, userKey, keylen, iv, dir); |
wolfSSL | 15:117db924cf7c | 1861 | } |
wolfSSL | 15:117db924cf7c | 1862 | #endif |
wolfSSL | 15:117db924cf7c | 1863 | |
wolfSSL | 15:117db924cf7c | 1864 | #elif defined(HAVE_COLDFIRE_SEC) |
wolfSSL | 15:117db924cf7c | 1865 | #if defined (HAVE_THREADX) |
wolfSSL | 15:117db924cf7c | 1866 | #include "memory_pools.h" |
wolfSSL | 15:117db924cf7c | 1867 | extern TX_BYTE_POOL mp_ncached; /* Non Cached memory pool */ |
wolfSSL | 15:117db924cf7c | 1868 | #endif |
wolfSSL | 15:117db924cf7c | 1869 | |
wolfSSL | 15:117db924cf7c | 1870 | #define AES_BUFFER_SIZE (AES_BLOCK_SIZE * 64) |
wolfSSL | 15:117db924cf7c | 1871 | static unsigned char *AESBuffIn = NULL; |
wolfSSL | 15:117db924cf7c | 1872 | static unsigned char *AESBuffOut = NULL; |
wolfSSL | 15:117db924cf7c | 1873 | static byte *secReg; |
wolfSSL | 15:117db924cf7c | 1874 | static byte *secKey; |
wolfSSL | 15:117db924cf7c | 1875 | static volatile SECdescriptorType *secDesc; |
wolfSSL | 15:117db924cf7c | 1876 | |
wolfSSL | 15:117db924cf7c | 1877 | static wolfSSL_Mutex Mutex_AesSEC; |
wolfSSL | 15:117db924cf7c | 1878 | |
wolfSSL | 15:117db924cf7c | 1879 | #define SEC_DESC_AES_CBC_ENCRYPT 0x60300010 |
wolfSSL | 15:117db924cf7c | 1880 | #define SEC_DESC_AES_CBC_DECRYPT 0x60200010 |
wolfSSL | 15:117db924cf7c | 1881 | |
wolfSSL | 15:117db924cf7c | 1882 | extern volatile unsigned char __MBAR[]; |
wolfSSL | 15:117db924cf7c | 1883 | |
wolfSSL | 15:117db924cf7c | 1884 | int wc_AesSetKey(Aes* aes, const byte* userKey, word32 keylen, |
wolfSSL | 15:117db924cf7c | 1885 | const byte* iv, int dir) |
wolfSSL | 15:117db924cf7c | 1886 | { |
wolfSSL | 15:117db924cf7c | 1887 | if (AESBuffIn == NULL) { |
wolfSSL | 15:117db924cf7c | 1888 | #if defined (HAVE_THREADX) |
wolfSSL | 15:117db924cf7c | 1889 | int s1, s2, s3, s4, s5; |
wolfSSL | 15:117db924cf7c | 1890 | s5 = tx_byte_allocate(&mp_ncached,(void *)&secDesc, |
wolfSSL | 15:117db924cf7c | 1891 | sizeof(SECdescriptorType), TX_NO_WAIT); |
wolfSSL | 15:117db924cf7c | 1892 | s1 = tx_byte_allocate(&mp_ncached, (void *)&AESBuffIn, |
wolfSSL | 15:117db924cf7c | 1893 | AES_BUFFER_SIZE, TX_NO_WAIT); |
wolfSSL | 15:117db924cf7c | 1894 | s2 = tx_byte_allocate(&mp_ncached, (void *)&AESBuffOut, |
wolfSSL | 15:117db924cf7c | 1895 | AES_BUFFER_SIZE, TX_NO_WAIT); |
wolfSSL | 15:117db924cf7c | 1896 | s3 = tx_byte_allocate(&mp_ncached, (void *)&secKey, |
wolfSSL | 15:117db924cf7c | 1897 | AES_BLOCK_SIZE*2, TX_NO_WAIT); |
wolfSSL | 15:117db924cf7c | 1898 | s4 = tx_byte_allocate(&mp_ncached, (void *)&secReg, |
wolfSSL | 15:117db924cf7c | 1899 | AES_BLOCK_SIZE, TX_NO_WAIT); |
wolfSSL | 15:117db924cf7c | 1900 | |
wolfSSL | 15:117db924cf7c | 1901 | if (s1 || s2 || s3 || s4 || s5) |
wolfSSL | 15:117db924cf7c | 1902 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 1903 | #else |
wolfSSL | 15:117db924cf7c | 1904 | #warning "Allocate non-Cache buffers" |
wolfSSL | 15:117db924cf7c | 1905 | #endif |
wolfSSL | 15:117db924cf7c | 1906 | |
wolfSSL | 15:117db924cf7c | 1907 | wc_InitMutex(&Mutex_AesSEC); |
wolfSSL | 15:117db924cf7c | 1908 | } |
wolfSSL | 15:117db924cf7c | 1909 | |
wolfSSL | 15:117db924cf7c | 1910 | if (!((keylen == 16) || (keylen == 24) || (keylen == 32))) |
wolfSSL | 15:117db924cf7c | 1911 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 1912 | |
wolfSSL | 15:117db924cf7c | 1913 | if (aes == NULL) |
wolfSSL | 15:117db924cf7c | 1914 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 1915 | |
wolfSSL | 15:117db924cf7c | 1916 | aes->keylen = keylen; |
wolfSSL | 15:117db924cf7c | 1917 | aes->rounds = keylen/4 + 6; |
wolfSSL | 15:117db924cf7c | 1918 | XMEMCPY(aes->key, userKey, keylen); |
wolfSSL | 15:117db924cf7c | 1919 | |
wolfSSL | 15:117db924cf7c | 1920 | if (iv) |
wolfSSL | 15:117db924cf7c | 1921 | XMEMCPY(aes->reg, iv, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 1922 | |
wolfSSL | 15:117db924cf7c | 1923 | #if defined(WOLFSSL_AES_CFB) || defined(WOLFSSL_AES_COUNTER) |
wolfSSL | 15:117db924cf7c | 1924 | aes->left = 0; |
wolfSSL | 15:117db924cf7c | 1925 | #endif |
wolfSSL | 15:117db924cf7c | 1926 | |
wolfSSL | 15:117db924cf7c | 1927 | return 0; |
wolfSSL | 15:117db924cf7c | 1928 | } |
wolfSSL | 15:117db924cf7c | 1929 | #elif defined(FREESCALE_LTC) |
wolfSSL | 15:117db924cf7c | 1930 | int wc_AesSetKey(Aes* aes, const byte* userKey, word32 keylen, const byte* iv, |
wolfSSL | 15:117db924cf7c | 1931 | int dir) |
wolfSSL | 15:117db924cf7c | 1932 | { |
wolfSSL | 15:117db924cf7c | 1933 | if (!((keylen == 16) || (keylen == 24) || (keylen == 32))) |
wolfSSL | 15:117db924cf7c | 1934 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 1935 | |
wolfSSL | 15:117db924cf7c | 1936 | aes->rounds = keylen/4 + 6; |
wolfSSL | 15:117db924cf7c | 1937 | XMEMCPY(aes->key, userKey, keylen); |
wolfSSL | 15:117db924cf7c | 1938 | |
wolfSSL | 15:117db924cf7c | 1939 | #if defined(WOLFSSL_AES_CFB) || defined(WOLFSSL_AES_COUNTER) |
wolfSSL | 15:117db924cf7c | 1940 | aes->left = 0; |
wolfSSL | 15:117db924cf7c | 1941 | #endif |
wolfSSL | 15:117db924cf7c | 1942 | |
wolfSSL | 15:117db924cf7c | 1943 | return wc_AesSetIV(aes, iv); |
wolfSSL | 15:117db924cf7c | 1944 | } |
wolfSSL | 15:117db924cf7c | 1945 | |
wolfSSL | 15:117db924cf7c | 1946 | int wc_AesSetKeyDirect(Aes* aes, const byte* userKey, word32 keylen, |
wolfSSL | 15:117db924cf7c | 1947 | const byte* iv, int dir) |
wolfSSL | 15:117db924cf7c | 1948 | { |
wolfSSL | 15:117db924cf7c | 1949 | return wc_AesSetKey(aes, userKey, keylen, iv, dir); |
wolfSSL | 15:117db924cf7c | 1950 | } |
wolfSSL | 15:117db924cf7c | 1951 | #elif defined(FREESCALE_MMCAU) |
wolfSSL | 15:117db924cf7c | 1952 | int wc_AesSetKey(Aes* aes, const byte* userKey, word32 keylen, |
wolfSSL | 15:117db924cf7c | 1953 | const byte* iv, int dir) |
wolfSSL | 15:117db924cf7c | 1954 | { |
wolfSSL | 15:117db924cf7c | 1955 | int ret; |
wolfSSL | 15:117db924cf7c | 1956 | byte *rk = (byte*)aes->key; |
wolfSSL | 15:117db924cf7c | 1957 | |
wolfSSL | 15:117db924cf7c | 1958 | (void)dir; |
wolfSSL | 15:117db924cf7c | 1959 | |
wolfSSL | 15:117db924cf7c | 1960 | if (!((keylen == 16) || (keylen == 24) || (keylen == 32))) |
wolfSSL | 15:117db924cf7c | 1961 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 1962 | |
wolfSSL | 15:117db924cf7c | 1963 | if (rk == NULL) |
wolfSSL | 15:117db924cf7c | 1964 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 1965 | |
wolfSSL | 15:117db924cf7c | 1966 | #if defined(WOLFSSL_AES_CFB) || defined(WOLFSSL_AES_COUNTER) |
wolfSSL | 15:117db924cf7c | 1967 | aes->left = 0; |
wolfSSL | 15:117db924cf7c | 1968 | #endif |
wolfSSL | 15:117db924cf7c | 1969 | |
wolfSSL | 15:117db924cf7c | 1970 | aes->rounds = keylen/4 + 6; |
wolfSSL | 15:117db924cf7c | 1971 | |
wolfSSL | 15:117db924cf7c | 1972 | ret = wolfSSL_CryptHwMutexLock(); |
wolfSSL | 15:117db924cf7c | 1973 | if(ret == 0) { |
wolfSSL | 15:117db924cf7c | 1974 | #ifdef FREESCALE_MMCAU_CLASSIC |
wolfSSL | 15:117db924cf7c | 1975 | cau_aes_set_key(userKey, keylen*8, rk); |
wolfSSL | 15:117db924cf7c | 1976 | #else |
wolfSSL | 15:117db924cf7c | 1977 | MMCAU_AES_SetKey(userKey, keylen, rk); |
wolfSSL | 15:117db924cf7c | 1978 | #endif |
wolfSSL | 15:117db924cf7c | 1979 | wolfSSL_CryptHwMutexUnLock(); |
wolfSSL | 15:117db924cf7c | 1980 | |
wolfSSL | 15:117db924cf7c | 1981 | ret = wc_AesSetIV(aes, iv); |
wolfSSL | 15:117db924cf7c | 1982 | } |
wolfSSL | 15:117db924cf7c | 1983 | |
wolfSSL | 15:117db924cf7c | 1984 | return ret; |
wolfSSL | 15:117db924cf7c | 1985 | } |
wolfSSL | 15:117db924cf7c | 1986 | |
wolfSSL | 15:117db924cf7c | 1987 | int wc_AesSetKeyDirect(Aes* aes, const byte* userKey, word32 keylen, |
wolfSSL | 15:117db924cf7c | 1988 | const byte* iv, int dir) |
wolfSSL | 15:117db924cf7c | 1989 | { |
wolfSSL | 15:117db924cf7c | 1990 | return wc_AesSetKey(aes, userKey, keylen, iv, dir); |
wolfSSL | 15:117db924cf7c | 1991 | } |
wolfSSL | 15:117db924cf7c | 1992 | |
wolfSSL | 15:117db924cf7c | 1993 | #elif defined(WOLFSSL_NRF51_AES) |
wolfSSL | 15:117db924cf7c | 1994 | int wc_AesSetKey(Aes* aes, const byte* userKey, word32 keylen, |
wolfSSL | 15:117db924cf7c | 1995 | const byte* iv, int dir) |
wolfSSL | 15:117db924cf7c | 1996 | { |
wolfSSL | 15:117db924cf7c | 1997 | int ret; |
wolfSSL | 15:117db924cf7c | 1998 | |
wolfSSL | 15:117db924cf7c | 1999 | (void)dir; |
wolfSSL | 15:117db924cf7c | 2000 | (void)iv; |
wolfSSL | 15:117db924cf7c | 2001 | |
wolfSSL | 15:117db924cf7c | 2002 | if (keylen != 16) |
wolfSSL | 15:117db924cf7c | 2003 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2004 | |
wolfSSL | 15:117db924cf7c | 2005 | aes->keylen = keylen; |
wolfSSL | 15:117db924cf7c | 2006 | aes->rounds = keylen/4 + 6; |
wolfSSL | 15:117db924cf7c | 2007 | ret = nrf51_aes_set_key(userKey); |
wolfSSL | 15:117db924cf7c | 2008 | |
wolfSSL | 15:117db924cf7c | 2009 | #if defined(WOLFSSL_AES_CFB) || defined(WOLFSSL_AES_COUNTER) |
wolfSSL | 15:117db924cf7c | 2010 | aes->left = 0; |
wolfSSL | 15:117db924cf7c | 2011 | #endif |
wolfSSL | 15:117db924cf7c | 2012 | |
wolfSSL | 15:117db924cf7c | 2013 | return ret; |
wolfSSL | 15:117db924cf7c | 2014 | } |
wolfSSL | 15:117db924cf7c | 2015 | |
wolfSSL | 15:117db924cf7c | 2016 | int wc_AesSetKeyDirect(Aes* aes, const byte* userKey, word32 keylen, |
wolfSSL | 15:117db924cf7c | 2017 | const byte* iv, int dir) |
wolfSSL | 15:117db924cf7c | 2018 | { |
wolfSSL | 15:117db924cf7c | 2019 | return wc_AesSetKey(aes, userKey, keylen, iv, dir); |
wolfSSL | 15:117db924cf7c | 2020 | } |
wolfSSL | 15:117db924cf7c | 2021 | |
wolfSSL | 15:117db924cf7c | 2022 | #elif defined(WOLFSSL_IMX6_CAAM) && !defined(NO_IMX6_CAAM_AES) |
wolfSSL | 15:117db924cf7c | 2023 | /* implemented in wolfcrypt/src/port/caam/caam_aes.c */ |
wolfSSL | 15:117db924cf7c | 2024 | |
wolfSSL | 15:117db924cf7c | 2025 | #else |
wolfSSL | 15:117db924cf7c | 2026 | static int wc_AesSetKeyLocal(Aes* aes, const byte* userKey, word32 keylen, |
wolfSSL | 15:117db924cf7c | 2027 | const byte* iv, int dir) |
wolfSSL | 15:117db924cf7c | 2028 | { |
wolfSSL | 15:117db924cf7c | 2029 | word32 *rk = aes->key; |
wolfSSL | 15:117db924cf7c | 2030 | #ifdef NEED_AES_TABLES |
wolfSSL | 15:117db924cf7c | 2031 | word32 temp; |
wolfSSL | 15:117db924cf7c | 2032 | unsigned int i = 0; |
wolfSSL | 15:117db924cf7c | 2033 | #endif |
wolfSSL | 15:117db924cf7c | 2034 | |
wolfSSL | 15:117db924cf7c | 2035 | #ifdef WOLFSSL_AESNI |
wolfSSL | 15:117db924cf7c | 2036 | aes->use_aesni = 0; |
wolfSSL | 15:117db924cf7c | 2037 | #endif /* WOLFSSL_AESNI */ |
wolfSSL | 15:117db924cf7c | 2038 | #if defined(WOLFSSL_AES_CFB) || defined(WOLFSSL_AES_COUNTER) |
wolfSSL | 15:117db924cf7c | 2039 | aes->left = 0; |
wolfSSL | 15:117db924cf7c | 2040 | #endif |
wolfSSL | 15:117db924cf7c | 2041 | |
wolfSSL | 15:117db924cf7c | 2042 | aes->keylen = keylen; |
wolfSSL | 15:117db924cf7c | 2043 | aes->rounds = (keylen/4) + 6; |
wolfSSL | 15:117db924cf7c | 2044 | |
wolfSSL | 15:117db924cf7c | 2045 | XMEMCPY(rk, userKey, keylen); |
wolfSSL | 15:117db924cf7c | 2046 | #if defined(LITTLE_ENDIAN_ORDER) && !defined(WOLFSSL_PIC32MZ_CRYPT) |
wolfSSL | 15:117db924cf7c | 2047 | ByteReverseWords(rk, rk, keylen); |
wolfSSL | 15:117db924cf7c | 2048 | #endif |
wolfSSL | 15:117db924cf7c | 2049 | |
wolfSSL | 15:117db924cf7c | 2050 | #ifdef NEED_AES_TABLES |
wolfSSL | 15:117db924cf7c | 2051 | |
wolfSSL | 15:117db924cf7c | 2052 | switch (keylen) { |
wolfSSL | 15:117db924cf7c | 2053 | #if defined(AES_MAX_KEY_SIZE) && AES_MAX_KEY_SIZE >= 128 && \ |
wolfSSL | 15:117db924cf7c | 2054 | defined(WOLFSSL_AES_128) |
wolfSSL | 15:117db924cf7c | 2055 | case 16: |
wolfSSL | 15:117db924cf7c | 2056 | while (1) |
wolfSSL | 15:117db924cf7c | 2057 | { |
wolfSSL | 15:117db924cf7c | 2058 | temp = rk[3]; |
wolfSSL | 15:117db924cf7c | 2059 | rk[4] = rk[0] ^ |
wolfSSL | 15:117db924cf7c | 2060 | (Te[2][GETBYTE(temp, 2)] & 0xff000000) ^ |
wolfSSL | 15:117db924cf7c | 2061 | (Te[3][GETBYTE(temp, 1)] & 0x00ff0000) ^ |
wolfSSL | 15:117db924cf7c | 2062 | (Te[0][GETBYTE(temp, 0)] & 0x0000ff00) ^ |
wolfSSL | 15:117db924cf7c | 2063 | (Te[1][GETBYTE(temp, 3)] & 0x000000ff) ^ |
wolfSSL | 15:117db924cf7c | 2064 | rcon[i]; |
wolfSSL | 15:117db924cf7c | 2065 | rk[5] = rk[1] ^ rk[4]; |
wolfSSL | 15:117db924cf7c | 2066 | rk[6] = rk[2] ^ rk[5]; |
wolfSSL | 15:117db924cf7c | 2067 | rk[7] = rk[3] ^ rk[6]; |
wolfSSL | 15:117db924cf7c | 2068 | if (++i == 10) |
wolfSSL | 15:117db924cf7c | 2069 | break; |
wolfSSL | 15:117db924cf7c | 2070 | rk += 4; |
wolfSSL | 15:117db924cf7c | 2071 | } |
wolfSSL | 15:117db924cf7c | 2072 | break; |
wolfSSL | 15:117db924cf7c | 2073 | #endif /* 128 */ |
wolfSSL | 15:117db924cf7c | 2074 | |
wolfSSL | 15:117db924cf7c | 2075 | #if defined(AES_MAX_KEY_SIZE) && AES_MAX_KEY_SIZE >= 192 && \ |
wolfSSL | 15:117db924cf7c | 2076 | defined(WOLFSSL_AES_192) |
wolfSSL | 15:117db924cf7c | 2077 | case 24: |
wolfSSL | 15:117db924cf7c | 2078 | /* for (;;) here triggers a bug in VC60 SP4 w/ Pro Pack */ |
wolfSSL | 15:117db924cf7c | 2079 | while (1) |
wolfSSL | 15:117db924cf7c | 2080 | { |
wolfSSL | 15:117db924cf7c | 2081 | temp = rk[ 5]; |
wolfSSL | 15:117db924cf7c | 2082 | rk[ 6] = rk[ 0] ^ |
wolfSSL | 15:117db924cf7c | 2083 | (Te[2][GETBYTE(temp, 2)] & 0xff000000) ^ |
wolfSSL | 15:117db924cf7c | 2084 | (Te[3][GETBYTE(temp, 1)] & 0x00ff0000) ^ |
wolfSSL | 15:117db924cf7c | 2085 | (Te[0][GETBYTE(temp, 0)] & 0x0000ff00) ^ |
wolfSSL | 15:117db924cf7c | 2086 | (Te[1][GETBYTE(temp, 3)] & 0x000000ff) ^ |
wolfSSL | 15:117db924cf7c | 2087 | rcon[i]; |
wolfSSL | 15:117db924cf7c | 2088 | rk[ 7] = rk[ 1] ^ rk[ 6]; |
wolfSSL | 15:117db924cf7c | 2089 | rk[ 8] = rk[ 2] ^ rk[ 7]; |
wolfSSL | 15:117db924cf7c | 2090 | rk[ 9] = rk[ 3] ^ rk[ 8]; |
wolfSSL | 15:117db924cf7c | 2091 | if (++i == 8) |
wolfSSL | 15:117db924cf7c | 2092 | break; |
wolfSSL | 15:117db924cf7c | 2093 | rk[10] = rk[ 4] ^ rk[ 9]; |
wolfSSL | 15:117db924cf7c | 2094 | rk[11] = rk[ 5] ^ rk[10]; |
wolfSSL | 15:117db924cf7c | 2095 | rk += 6; |
wolfSSL | 15:117db924cf7c | 2096 | } |
wolfSSL | 15:117db924cf7c | 2097 | break; |
wolfSSL | 15:117db924cf7c | 2098 | #endif /* 192 */ |
wolfSSL | 15:117db924cf7c | 2099 | |
wolfSSL | 15:117db924cf7c | 2100 | #if defined(AES_MAX_KEY_SIZE) && AES_MAX_KEY_SIZE >= 256 && \ |
wolfSSL | 15:117db924cf7c | 2101 | defined(WOLFSSL_AES_256) |
wolfSSL | 15:117db924cf7c | 2102 | case 32: |
wolfSSL | 15:117db924cf7c | 2103 | while (1) |
wolfSSL | 15:117db924cf7c | 2104 | { |
wolfSSL | 15:117db924cf7c | 2105 | temp = rk[ 7]; |
wolfSSL | 15:117db924cf7c | 2106 | rk[ 8] = rk[ 0] ^ |
wolfSSL | 15:117db924cf7c | 2107 | (Te[2][GETBYTE(temp, 2)] & 0xff000000) ^ |
wolfSSL | 15:117db924cf7c | 2108 | (Te[3][GETBYTE(temp, 1)] & 0x00ff0000) ^ |
wolfSSL | 15:117db924cf7c | 2109 | (Te[0][GETBYTE(temp, 0)] & 0x0000ff00) ^ |
wolfSSL | 15:117db924cf7c | 2110 | (Te[1][GETBYTE(temp, 3)] & 0x000000ff) ^ |
wolfSSL | 15:117db924cf7c | 2111 | rcon[i]; |
wolfSSL | 15:117db924cf7c | 2112 | rk[ 9] = rk[ 1] ^ rk[ 8]; |
wolfSSL | 15:117db924cf7c | 2113 | rk[10] = rk[ 2] ^ rk[ 9]; |
wolfSSL | 15:117db924cf7c | 2114 | rk[11] = rk[ 3] ^ rk[10]; |
wolfSSL | 15:117db924cf7c | 2115 | if (++i == 7) |
wolfSSL | 15:117db924cf7c | 2116 | break; |
wolfSSL | 15:117db924cf7c | 2117 | temp = rk[11]; |
wolfSSL | 15:117db924cf7c | 2118 | rk[12] = rk[ 4] ^ |
wolfSSL | 15:117db924cf7c | 2119 | (Te[2][GETBYTE(temp, 3)] & 0xff000000) ^ |
wolfSSL | 15:117db924cf7c | 2120 | (Te[3][GETBYTE(temp, 2)] & 0x00ff0000) ^ |
wolfSSL | 15:117db924cf7c | 2121 | (Te[0][GETBYTE(temp, 1)] & 0x0000ff00) ^ |
wolfSSL | 15:117db924cf7c | 2122 | (Te[1][GETBYTE(temp, 0)] & 0x000000ff); |
wolfSSL | 15:117db924cf7c | 2123 | rk[13] = rk[ 5] ^ rk[12]; |
wolfSSL | 15:117db924cf7c | 2124 | rk[14] = rk[ 6] ^ rk[13]; |
wolfSSL | 15:117db924cf7c | 2125 | rk[15] = rk[ 7] ^ rk[14]; |
wolfSSL | 15:117db924cf7c | 2126 | |
wolfSSL | 15:117db924cf7c | 2127 | rk += 8; |
wolfSSL | 15:117db924cf7c | 2128 | } |
wolfSSL | 15:117db924cf7c | 2129 | break; |
wolfSSL | 15:117db924cf7c | 2130 | #endif /* 256 */ |
wolfSSL | 15:117db924cf7c | 2131 | |
wolfSSL | 15:117db924cf7c | 2132 | default: |
wolfSSL | 15:117db924cf7c | 2133 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2134 | } /* switch */ |
wolfSSL | 15:117db924cf7c | 2135 | |
wolfSSL | 15:117db924cf7c | 2136 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 2137 | if (dir == AES_DECRYPTION) { |
wolfSSL | 15:117db924cf7c | 2138 | unsigned int j; |
wolfSSL | 15:117db924cf7c | 2139 | rk = aes->key; |
wolfSSL | 15:117db924cf7c | 2140 | |
wolfSSL | 15:117db924cf7c | 2141 | /* invert the order of the round keys: */ |
wolfSSL | 15:117db924cf7c | 2142 | for (i = 0, j = 4* aes->rounds; i < j; i += 4, j -= 4) { |
wolfSSL | 15:117db924cf7c | 2143 | temp = rk[i ]; rk[i ] = rk[j ]; rk[j ] = temp; |
wolfSSL | 15:117db924cf7c | 2144 | temp = rk[i + 1]; rk[i + 1] = rk[j + 1]; rk[j + 1] = temp; |
wolfSSL | 15:117db924cf7c | 2145 | temp = rk[i + 2]; rk[i + 2] = rk[j + 2]; rk[j + 2] = temp; |
wolfSSL | 15:117db924cf7c | 2146 | temp = rk[i + 3]; rk[i + 3] = rk[j + 3]; rk[j + 3] = temp; |
wolfSSL | 15:117db924cf7c | 2147 | } |
wolfSSL | 15:117db924cf7c | 2148 | /* apply the inverse MixColumn transform to all round keys but the |
wolfSSL | 15:117db924cf7c | 2149 | first and the last: */ |
wolfSSL | 15:117db924cf7c | 2150 | for (i = 1; i < aes->rounds; i++) { |
wolfSSL | 15:117db924cf7c | 2151 | rk += 4; |
wolfSSL | 15:117db924cf7c | 2152 | rk[0] = |
wolfSSL | 15:117db924cf7c | 2153 | Td[0][Te[1][GETBYTE(rk[0], 3)] & 0xff] ^ |
wolfSSL | 15:117db924cf7c | 2154 | Td[1][Te[1][GETBYTE(rk[0], 2)] & 0xff] ^ |
wolfSSL | 15:117db924cf7c | 2155 | Td[2][Te[1][GETBYTE(rk[0], 1)] & 0xff] ^ |
wolfSSL | 15:117db924cf7c | 2156 | Td[3][Te[1][GETBYTE(rk[0], 0)] & 0xff]; |
wolfSSL | 15:117db924cf7c | 2157 | rk[1] = |
wolfSSL | 15:117db924cf7c | 2158 | Td[0][Te[1][GETBYTE(rk[1], 3)] & 0xff] ^ |
wolfSSL | 15:117db924cf7c | 2159 | Td[1][Te[1][GETBYTE(rk[1], 2)] & 0xff] ^ |
wolfSSL | 15:117db924cf7c | 2160 | Td[2][Te[1][GETBYTE(rk[1], 1)] & 0xff] ^ |
wolfSSL | 15:117db924cf7c | 2161 | Td[3][Te[1][GETBYTE(rk[1], 0)] & 0xff]; |
wolfSSL | 15:117db924cf7c | 2162 | rk[2] = |
wolfSSL | 15:117db924cf7c | 2163 | Td[0][Te[1][GETBYTE(rk[2], 3)] & 0xff] ^ |
wolfSSL | 15:117db924cf7c | 2164 | Td[1][Te[1][GETBYTE(rk[2], 2)] & 0xff] ^ |
wolfSSL | 15:117db924cf7c | 2165 | Td[2][Te[1][GETBYTE(rk[2], 1)] & 0xff] ^ |
wolfSSL | 15:117db924cf7c | 2166 | Td[3][Te[1][GETBYTE(rk[2], 0)] & 0xff]; |
wolfSSL | 15:117db924cf7c | 2167 | rk[3] = |
wolfSSL | 15:117db924cf7c | 2168 | Td[0][Te[1][GETBYTE(rk[3], 3)] & 0xff] ^ |
wolfSSL | 15:117db924cf7c | 2169 | Td[1][Te[1][GETBYTE(rk[3], 2)] & 0xff] ^ |
wolfSSL | 15:117db924cf7c | 2170 | Td[2][Te[1][GETBYTE(rk[3], 1)] & 0xff] ^ |
wolfSSL | 15:117db924cf7c | 2171 | Td[3][Te[1][GETBYTE(rk[3], 0)] & 0xff]; |
wolfSSL | 15:117db924cf7c | 2172 | } |
wolfSSL | 15:117db924cf7c | 2173 | } |
wolfSSL | 15:117db924cf7c | 2174 | #else |
wolfSSL | 15:117db924cf7c | 2175 | (void)dir; |
wolfSSL | 15:117db924cf7c | 2176 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 2177 | #endif /* NEED_AES_TABLES */ |
wolfSSL | 15:117db924cf7c | 2178 | |
wolfSSL | 15:117db924cf7c | 2179 | return wc_AesSetIV(aes, iv); |
wolfSSL | 15:117db924cf7c | 2180 | } |
wolfSSL | 15:117db924cf7c | 2181 | |
wolfSSL | 15:117db924cf7c | 2182 | int wc_AesSetKey(Aes* aes, const byte* userKey, word32 keylen, |
wolfSSL | 15:117db924cf7c | 2183 | const byte* iv, int dir) |
wolfSSL | 15:117db924cf7c | 2184 | { |
wolfSSL | 15:117db924cf7c | 2185 | int ret; |
wolfSSL | 15:117db924cf7c | 2186 | #if defined(AES_MAX_KEY_SIZE) |
wolfSSL | 15:117db924cf7c | 2187 | const word32 max_key_len = (AES_MAX_KEY_SIZE / 8); |
wolfSSL | 15:117db924cf7c | 2188 | #endif |
wolfSSL | 15:117db924cf7c | 2189 | |
wolfSSL | 15:117db924cf7c | 2190 | #ifdef WOLFSSL_IMX6_CAAM_BLOB |
wolfSSL | 15:117db924cf7c | 2191 | byte local[32]; |
wolfSSL | 15:117db924cf7c | 2192 | word32 localSz = 32; |
wolfSSL | 15:117db924cf7c | 2193 | |
wolfSSL | 15:117db924cf7c | 2194 | if (keylen == (16 + WC_CAAM_BLOB_SZ) || |
wolfSSL | 15:117db924cf7c | 2195 | keylen == (24 + WC_CAAM_BLOB_SZ) || |
wolfSSL | 15:117db924cf7c | 2196 | keylen == (32 + WC_CAAM_BLOB_SZ)) { |
wolfSSL | 15:117db924cf7c | 2197 | if (wc_caamOpenBlob((byte*)userKey, keylen, local, &localSz) != 0) { |
wolfSSL | 15:117db924cf7c | 2198 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2199 | } |
wolfSSL | 15:117db924cf7c | 2200 | |
wolfSSL | 15:117db924cf7c | 2201 | /* set local values */ |
wolfSSL | 15:117db924cf7c | 2202 | userKey = local; |
wolfSSL | 15:117db924cf7c | 2203 | keylen = localSz; |
wolfSSL | 15:117db924cf7c | 2204 | } |
wolfSSL | 15:117db924cf7c | 2205 | #endif |
wolfSSL | 15:117db924cf7c | 2206 | if (aes == NULL || |
wolfSSL | 15:117db924cf7c | 2207 | !((keylen == 16) || (keylen == 24) || (keylen == 32))) { |
wolfSSL | 15:117db924cf7c | 2208 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2209 | } |
wolfSSL | 15:117db924cf7c | 2210 | |
wolfSSL | 15:117db924cf7c | 2211 | #if defined(AES_MAX_KEY_SIZE) |
wolfSSL | 15:117db924cf7c | 2212 | /* Check key length */ |
wolfSSL | 15:117db924cf7c | 2213 | if (keylen > max_key_len) { |
wolfSSL | 15:117db924cf7c | 2214 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2215 | } |
wolfSSL | 15:117db924cf7c | 2216 | #endif |
wolfSSL | 15:117db924cf7c | 2217 | aes->keylen = keylen; |
wolfSSL | 15:117db924cf7c | 2218 | aes->rounds = keylen/4 + 6; |
wolfSSL | 15:117db924cf7c | 2219 | |
wolfSSL | 15:117db924cf7c | 2220 | #if defined(WOLFSSL_ASYNC_CRYPT) && defined(WC_ASYNC_ENABLE_AES) |
wolfSSL | 15:117db924cf7c | 2221 | if (aes->asyncDev.marker == WOLFSSL_ASYNC_MARKER_AES) { |
wolfSSL | 15:117db924cf7c | 2222 | XMEMCPY(aes->asyncKey, userKey, keylen); |
wolfSSL | 15:117db924cf7c | 2223 | if (iv) |
wolfSSL | 15:117db924cf7c | 2224 | XMEMCPY(aes->asyncIv, iv, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2225 | } |
wolfSSL | 15:117db924cf7c | 2226 | #endif /* WOLFSSL_ASYNC_CRYPT */ |
wolfSSL | 15:117db924cf7c | 2227 | |
wolfSSL | 15:117db924cf7c | 2228 | #ifdef WOLFSSL_AESNI |
wolfSSL | 15:117db924cf7c | 2229 | if (checkAESNI == 0) { |
wolfSSL | 15:117db924cf7c | 2230 | haveAESNI = Check_CPU_support_AES(); |
wolfSSL | 15:117db924cf7c | 2231 | checkAESNI = 1; |
wolfSSL | 15:117db924cf7c | 2232 | } |
wolfSSL | 15:117db924cf7c | 2233 | if (haveAESNI) { |
wolfSSL | 15:117db924cf7c | 2234 | #if defined(WOLFSSL_AES_COUNTER) || defined(WOLFSSL_AES_CFB) |
wolfSSL | 15:117db924cf7c | 2235 | aes->left = 0; |
wolfSSL | 15:117db924cf7c | 2236 | #endif /* WOLFSSL_AES_COUNTER */ |
wolfSSL | 15:117db924cf7c | 2237 | aes->use_aesni = 1; |
wolfSSL | 15:117db924cf7c | 2238 | if (iv) |
wolfSSL | 15:117db924cf7c | 2239 | XMEMCPY(aes->reg, iv, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2240 | if (dir == AES_ENCRYPTION) |
wolfSSL | 15:117db924cf7c | 2241 | return AES_set_encrypt_key(userKey, keylen * 8, aes); |
wolfSSL | 15:117db924cf7c | 2242 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 2243 | else |
wolfSSL | 15:117db924cf7c | 2244 | return AES_set_decrypt_key(userKey, keylen * 8, aes); |
wolfSSL | 15:117db924cf7c | 2245 | #endif |
wolfSSL | 15:117db924cf7c | 2246 | } |
wolfSSL | 15:117db924cf7c | 2247 | #endif /* WOLFSSL_AESNI */ |
wolfSSL | 15:117db924cf7c | 2248 | |
wolfSSL | 15:117db924cf7c | 2249 | ret = wc_AesSetKeyLocal(aes, userKey, keylen, iv, dir); |
wolfSSL | 15:117db924cf7c | 2250 | |
wolfSSL | 15:117db924cf7c | 2251 | #ifdef WOLFSSL_IMX6_CAAM_BLOB |
wolfSSL | 15:117db924cf7c | 2252 | ForceZero(local, sizeof(local)); |
wolfSSL | 15:117db924cf7c | 2253 | #endif |
wolfSSL | 15:117db924cf7c | 2254 | return ret; |
wolfSSL | 15:117db924cf7c | 2255 | } |
wolfSSL | 15:117db924cf7c | 2256 | |
wolfSSL | 15:117db924cf7c | 2257 | #if defined(WOLFSSL_AES_DIRECT) || defined(WOLFSSL_AES_COUNTER) |
wolfSSL | 15:117db924cf7c | 2258 | /* AES-CTR and AES-DIRECT need to use this for key setup, no aesni yet */ |
wolfSSL | 15:117db924cf7c | 2259 | int wc_AesSetKeyDirect(Aes* aes, const byte* userKey, word32 keylen, |
wolfSSL | 15:117db924cf7c | 2260 | const byte* iv, int dir) |
wolfSSL | 15:117db924cf7c | 2261 | { |
wolfSSL | 15:117db924cf7c | 2262 | int ret; |
wolfSSL | 15:117db924cf7c | 2263 | |
wolfSSL | 15:117db924cf7c | 2264 | #ifdef WOLFSSL_IMX6_CAAM_BLOB |
wolfSSL | 15:117db924cf7c | 2265 | byte local[32]; |
wolfSSL | 15:117db924cf7c | 2266 | word32 localSz = 32; |
wolfSSL | 15:117db924cf7c | 2267 | |
wolfSSL | 15:117db924cf7c | 2268 | if (keylen == (16 + WC_CAAM_BLOB_SZ) || |
wolfSSL | 15:117db924cf7c | 2269 | keylen == (24 + WC_CAAM_BLOB_SZ) || |
wolfSSL | 15:117db924cf7c | 2270 | keylen == (32 + WC_CAAM_BLOB_SZ)) { |
wolfSSL | 15:117db924cf7c | 2271 | if (wc_caamOpenBlob((byte*)userKey, keylen, local, &localSz) |
wolfSSL | 15:117db924cf7c | 2272 | != 0) { |
wolfSSL | 15:117db924cf7c | 2273 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2274 | } |
wolfSSL | 15:117db924cf7c | 2275 | |
wolfSSL | 15:117db924cf7c | 2276 | /* set local values */ |
wolfSSL | 15:117db924cf7c | 2277 | userKey = local; |
wolfSSL | 15:117db924cf7c | 2278 | keylen = localSz; |
wolfSSL | 15:117db924cf7c | 2279 | } |
wolfSSL | 15:117db924cf7c | 2280 | #endif |
wolfSSL | 15:117db924cf7c | 2281 | ret = wc_AesSetKeyLocal(aes, userKey, keylen, iv, dir); |
wolfSSL | 15:117db924cf7c | 2282 | |
wolfSSL | 15:117db924cf7c | 2283 | #ifdef WOLFSSL_IMX6_CAAM_BLOB |
wolfSSL | 15:117db924cf7c | 2284 | ForceZero(local, sizeof(local)); |
wolfSSL | 15:117db924cf7c | 2285 | #endif |
wolfSSL | 15:117db924cf7c | 2286 | |
wolfSSL | 15:117db924cf7c | 2287 | return ret; |
wolfSSL | 15:117db924cf7c | 2288 | } |
wolfSSL | 15:117db924cf7c | 2289 | #endif /* WOLFSSL_AES_DIRECT || WOLFSSL_AES_COUNTER */ |
wolfSSL | 15:117db924cf7c | 2290 | #endif /* wc_AesSetKey block */ |
wolfSSL | 15:117db924cf7c | 2291 | |
wolfSSL | 15:117db924cf7c | 2292 | |
wolfSSL | 15:117db924cf7c | 2293 | /* wc_AesSetIV is shared between software and hardware */ |
wolfSSL | 15:117db924cf7c | 2294 | int wc_AesSetIV(Aes* aes, const byte* iv) |
wolfSSL | 15:117db924cf7c | 2295 | { |
wolfSSL | 15:117db924cf7c | 2296 | if (aes == NULL) |
wolfSSL | 15:117db924cf7c | 2297 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2298 | |
wolfSSL | 15:117db924cf7c | 2299 | if (iv) |
wolfSSL | 15:117db924cf7c | 2300 | XMEMCPY(aes->reg, iv, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2301 | else |
wolfSSL | 15:117db924cf7c | 2302 | XMEMSET(aes->reg, 0, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2303 | |
wolfSSL | 15:117db924cf7c | 2304 | return 0; |
wolfSSL | 15:117db924cf7c | 2305 | } |
wolfSSL | 15:117db924cf7c | 2306 | |
wolfSSL | 15:117db924cf7c | 2307 | /* AES-DIRECT */ |
wolfSSL | 15:117db924cf7c | 2308 | #if defined(WOLFSSL_AES_DIRECT) |
wolfSSL | 15:117db924cf7c | 2309 | #if defined(HAVE_COLDFIRE_SEC) |
wolfSSL | 15:117db924cf7c | 2310 | #error "Coldfire SEC doesn't yet support AES direct" |
wolfSSL | 15:117db924cf7c | 2311 | |
wolfSSL | 15:117db924cf7c | 2312 | #elif defined(FREESCALE_LTC) |
wolfSSL | 15:117db924cf7c | 2313 | /* Allow direct access to one block encrypt */ |
wolfSSL | 15:117db924cf7c | 2314 | void wc_AesEncryptDirect(Aes* aes, byte* out, const byte* in) |
wolfSSL | 15:117db924cf7c | 2315 | { |
wolfSSL | 15:117db924cf7c | 2316 | byte *key; |
wolfSSL | 15:117db924cf7c | 2317 | uint32_t keySize; |
wolfSSL | 15:117db924cf7c | 2318 | |
wolfSSL | 15:117db924cf7c | 2319 | key = (byte*)aes->key; |
wolfSSL | 15:117db924cf7c | 2320 | wc_AesGetKeySize(aes, &keySize); |
wolfSSL | 15:117db924cf7c | 2321 | |
wolfSSL | 15:117db924cf7c | 2322 | LTC_AES_EncryptEcb(LTC_BASE, in, out, AES_BLOCK_SIZE, |
wolfSSL | 15:117db924cf7c | 2323 | key, keySize); |
wolfSSL | 15:117db924cf7c | 2324 | } |
wolfSSL | 15:117db924cf7c | 2325 | |
wolfSSL | 15:117db924cf7c | 2326 | /* Allow direct access to one block decrypt */ |
wolfSSL | 15:117db924cf7c | 2327 | void wc_AesDecryptDirect(Aes* aes, byte* out, const byte* in) |
wolfSSL | 15:117db924cf7c | 2328 | { |
wolfSSL | 15:117db924cf7c | 2329 | byte *key; |
wolfSSL | 15:117db924cf7c | 2330 | uint32_t keySize; |
wolfSSL | 15:117db924cf7c | 2331 | |
wolfSSL | 15:117db924cf7c | 2332 | key = (byte*)aes->key; |
wolfSSL | 15:117db924cf7c | 2333 | wc_AesGetKeySize(aes, &keySize); |
wolfSSL | 15:117db924cf7c | 2334 | |
wolfSSL | 15:117db924cf7c | 2335 | LTC_AES_DecryptEcb(LTC_BASE, in, out, AES_BLOCK_SIZE, |
wolfSSL | 15:117db924cf7c | 2336 | key, keySize, kLTC_EncryptKey); |
wolfSSL | 15:117db924cf7c | 2337 | } |
wolfSSL | 15:117db924cf7c | 2338 | |
wolfSSL | 15:117db924cf7c | 2339 | #elif defined(WOLFSSL_IMX6_CAAM) && !defined(NO_IMX6_CAAM_AES) |
wolfSSL | 15:117db924cf7c | 2340 | /* implemented in wolfcrypt/src/port/caam/caam_aes.c */ |
wolfSSL | 15:117db924cf7c | 2341 | |
wolfSSL | 15:117db924cf7c | 2342 | #else |
wolfSSL | 15:117db924cf7c | 2343 | /* Allow direct access to one block encrypt */ |
wolfSSL | 15:117db924cf7c | 2344 | void wc_AesEncryptDirect(Aes* aes, byte* out, const byte* in) |
wolfSSL | 15:117db924cf7c | 2345 | { |
wolfSSL | 15:117db924cf7c | 2346 | wc_AesEncrypt(aes, in, out); |
wolfSSL | 15:117db924cf7c | 2347 | } |
wolfSSL | 15:117db924cf7c | 2348 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 2349 | /* Allow direct access to one block decrypt */ |
wolfSSL | 15:117db924cf7c | 2350 | void wc_AesDecryptDirect(Aes* aes, byte* out, const byte* in) |
wolfSSL | 15:117db924cf7c | 2351 | { |
wolfSSL | 15:117db924cf7c | 2352 | wc_AesDecrypt(aes, in, out); |
wolfSSL | 15:117db924cf7c | 2353 | } |
wolfSSL | 15:117db924cf7c | 2354 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 2355 | #endif /* AES direct block */ |
wolfSSL | 15:117db924cf7c | 2356 | #endif /* WOLFSSL_AES_DIRECT */ |
wolfSSL | 15:117db924cf7c | 2357 | |
wolfSSL | 15:117db924cf7c | 2358 | |
wolfSSL | 15:117db924cf7c | 2359 | /* AES-CBC */ |
wolfSSL | 15:117db924cf7c | 2360 | #ifdef HAVE_AES_CBC |
wolfSSL | 15:117db924cf7c | 2361 | #if defined(STM32_CRYPTO) |
wolfSSL | 15:117db924cf7c | 2362 | |
wolfSSL | 15:117db924cf7c | 2363 | #ifdef WOLFSSL_STM32_CUBEMX |
wolfSSL | 15:117db924cf7c | 2364 | int wc_AesCbcEncrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 2365 | { |
wolfSSL | 15:117db924cf7c | 2366 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 2367 | word32 blocks = (sz / AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2368 | CRYP_HandleTypeDef hcryp; |
wolfSSL | 15:117db924cf7c | 2369 | |
wolfSSL | 15:117db924cf7c | 2370 | XMEMSET(&hcryp, 0, sizeof(CRYP_HandleTypeDef)); |
wolfSSL | 15:117db924cf7c | 2371 | switch (aes->rounds) { |
wolfSSL | 15:117db924cf7c | 2372 | case 10: /* 128-bit key */ |
wolfSSL | 15:117db924cf7c | 2373 | hcryp.Init.KeySize = CRYP_KEYSIZE_128B; |
wolfSSL | 15:117db924cf7c | 2374 | break; |
wolfSSL | 15:117db924cf7c | 2375 | #ifdef CRYP_KEYSIZE_192B |
wolfSSL | 15:117db924cf7c | 2376 | case 12: /* 192-bit key */ |
wolfSSL | 15:117db924cf7c | 2377 | hcryp.Init.KeySize = CRYP_KEYSIZE_192B; |
wolfSSL | 15:117db924cf7c | 2378 | break; |
wolfSSL | 15:117db924cf7c | 2379 | #endif |
wolfSSL | 15:117db924cf7c | 2380 | case 14: /* 256-bit key */ |
wolfSSL | 15:117db924cf7c | 2381 | hcryp.Init.KeySize = CRYP_KEYSIZE_256B; |
wolfSSL | 15:117db924cf7c | 2382 | break; |
wolfSSL | 15:117db924cf7c | 2383 | default: |
wolfSSL | 15:117db924cf7c | 2384 | break; |
wolfSSL | 15:117db924cf7c | 2385 | } |
wolfSSL | 15:117db924cf7c | 2386 | hcryp.Instance = CRYP; |
wolfSSL | 15:117db924cf7c | 2387 | hcryp.Init.DataType = CRYP_DATATYPE_8B; |
wolfSSL | 15:117db924cf7c | 2388 | hcryp.Init.pKey = (uint8_t*)aes->key; |
wolfSSL | 15:117db924cf7c | 2389 | hcryp.Init.pInitVect = (uint8_t*)aes->reg; |
wolfSSL | 15:117db924cf7c | 2390 | |
wolfSSL | 15:117db924cf7c | 2391 | HAL_CRYP_Init(&hcryp); |
wolfSSL | 15:117db924cf7c | 2392 | |
wolfSSL | 15:117db924cf7c | 2393 | while (blocks--) { |
wolfSSL | 15:117db924cf7c | 2394 | if (HAL_CRYP_AESCBC_Encrypt(&hcryp, (uint8_t*)in, AES_BLOCK_SIZE, |
wolfSSL | 15:117db924cf7c | 2395 | out, STM32_HAL_TIMEOUT) != HAL_OK) { |
wolfSSL | 15:117db924cf7c | 2396 | ret = WC_TIMEOUT_E; |
wolfSSL | 15:117db924cf7c | 2397 | break; |
wolfSSL | 15:117db924cf7c | 2398 | } |
wolfSSL | 15:117db924cf7c | 2399 | |
wolfSSL | 15:117db924cf7c | 2400 | /* store iv for next call */ |
wolfSSL | 15:117db924cf7c | 2401 | XMEMCPY(aes->reg, out + sz - AES_BLOCK_SIZE, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2402 | |
wolfSSL | 15:117db924cf7c | 2403 | sz -= AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 2404 | in += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 2405 | out += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 2406 | } |
wolfSSL | 15:117db924cf7c | 2407 | |
wolfSSL | 15:117db924cf7c | 2408 | HAL_CRYP_DeInit(&hcryp); |
wolfSSL | 15:117db924cf7c | 2409 | |
wolfSSL | 15:117db924cf7c | 2410 | return ret; |
wolfSSL | 15:117db924cf7c | 2411 | } |
wolfSSL | 15:117db924cf7c | 2412 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 2413 | int wc_AesCbcDecrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 2414 | { |
wolfSSL | 15:117db924cf7c | 2415 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 2416 | word32 blocks = (sz / AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2417 | CRYP_HandleTypeDef hcryp; |
wolfSSL | 15:117db924cf7c | 2418 | |
wolfSSL | 15:117db924cf7c | 2419 | XMEMSET(&hcryp, 0, sizeof(CRYP_HandleTypeDef)); |
wolfSSL | 15:117db924cf7c | 2420 | switch (aes->rounds) { |
wolfSSL | 15:117db924cf7c | 2421 | case 10: /* 128-bit key */ |
wolfSSL | 15:117db924cf7c | 2422 | hcryp.Init.KeySize = CRYP_KEYSIZE_128B; |
wolfSSL | 15:117db924cf7c | 2423 | break; |
wolfSSL | 15:117db924cf7c | 2424 | #ifdef CRYP_KEYSIZE_192B |
wolfSSL | 15:117db924cf7c | 2425 | case 12: /* 192-bit key */ |
wolfSSL | 15:117db924cf7c | 2426 | hcryp.Init.KeySize = CRYP_KEYSIZE_192B; |
wolfSSL | 15:117db924cf7c | 2427 | break; |
wolfSSL | 15:117db924cf7c | 2428 | #endif |
wolfSSL | 15:117db924cf7c | 2429 | case 14: /* 256-bit key */ |
wolfSSL | 15:117db924cf7c | 2430 | hcryp.Init.KeySize = CRYP_KEYSIZE_256B; |
wolfSSL | 15:117db924cf7c | 2431 | break; |
wolfSSL | 15:117db924cf7c | 2432 | default: |
wolfSSL | 15:117db924cf7c | 2433 | break; |
wolfSSL | 15:117db924cf7c | 2434 | } |
wolfSSL | 15:117db924cf7c | 2435 | hcryp.Instance = CRYP; |
wolfSSL | 15:117db924cf7c | 2436 | hcryp.Init.DataType = CRYP_DATATYPE_8B; |
wolfSSL | 15:117db924cf7c | 2437 | hcryp.Init.pKey = (uint8_t*)aes->key; |
wolfSSL | 15:117db924cf7c | 2438 | hcryp.Init.pInitVect = (uint8_t*)aes->reg; |
wolfSSL | 15:117db924cf7c | 2439 | |
wolfSSL | 15:117db924cf7c | 2440 | HAL_CRYP_Init(&hcryp); |
wolfSSL | 15:117db924cf7c | 2441 | |
wolfSSL | 15:117db924cf7c | 2442 | while (blocks--) { |
wolfSSL | 15:117db924cf7c | 2443 | if (HAL_CRYP_AESCBC_Decrypt(&hcryp, (uint8_t*)in, AES_BLOCK_SIZE, |
wolfSSL | 15:117db924cf7c | 2444 | out, STM32_HAL_TIMEOUT) != HAL_OK) { |
wolfSSL | 15:117db924cf7c | 2445 | ret = WC_TIMEOUT_E; |
wolfSSL | 15:117db924cf7c | 2446 | } |
wolfSSL | 15:117db924cf7c | 2447 | |
wolfSSL | 15:117db924cf7c | 2448 | /* store iv for next call */ |
wolfSSL | 15:117db924cf7c | 2449 | XMEMCPY(aes->reg, aes->tmp, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2450 | |
wolfSSL | 15:117db924cf7c | 2451 | in += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 2452 | out += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 2453 | } |
wolfSSL | 15:117db924cf7c | 2454 | |
wolfSSL | 15:117db924cf7c | 2455 | HAL_CRYP_DeInit(&hcryp); |
wolfSSL | 15:117db924cf7c | 2456 | |
wolfSSL | 15:117db924cf7c | 2457 | return ret; |
wolfSSL | 15:117db924cf7c | 2458 | } |
wolfSSL | 15:117db924cf7c | 2459 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 2460 | #else |
wolfSSL | 15:117db924cf7c | 2461 | int wc_AesCbcEncrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 2462 | { |
wolfSSL | 15:117db924cf7c | 2463 | word32 *enc_key, *iv; |
wolfSSL | 15:117db924cf7c | 2464 | word32 blocks = (sz / AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2465 | CRYP_InitTypeDef AES_CRYP_InitStructure; |
wolfSSL | 15:117db924cf7c | 2466 | CRYP_KeyInitTypeDef AES_CRYP_KeyInitStructure; |
wolfSSL | 15:117db924cf7c | 2467 | CRYP_IVInitTypeDef AES_CRYP_IVInitStructure; |
wolfSSL | 15:117db924cf7c | 2468 | |
wolfSSL | 15:117db924cf7c | 2469 | enc_key = aes->key; |
wolfSSL | 15:117db924cf7c | 2470 | iv = aes->reg; |
wolfSSL | 15:117db924cf7c | 2471 | |
wolfSSL | 15:117db924cf7c | 2472 | /* crypto structure initialization */ |
wolfSSL | 15:117db924cf7c | 2473 | CRYP_KeyStructInit(&AES_CRYP_KeyInitStructure); |
wolfSSL | 15:117db924cf7c | 2474 | CRYP_StructInit(&AES_CRYP_InitStructure); |
wolfSSL | 15:117db924cf7c | 2475 | CRYP_IVStructInit(&AES_CRYP_IVInitStructure); |
wolfSSL | 15:117db924cf7c | 2476 | |
wolfSSL | 15:117db924cf7c | 2477 | /* reset registers to their default values */ |
wolfSSL | 15:117db924cf7c | 2478 | CRYP_DeInit(); |
wolfSSL | 15:117db924cf7c | 2479 | |
wolfSSL | 15:117db924cf7c | 2480 | /* load key into correct registers */ |
wolfSSL | 15:117db924cf7c | 2481 | switch (aes->rounds) { |
wolfSSL | 15:117db924cf7c | 2482 | case 10: /* 128-bit key */ |
wolfSSL | 15:117db924cf7c | 2483 | AES_CRYP_InitStructure.CRYP_KeySize = CRYP_KeySize_128b; |
wolfSSL | 15:117db924cf7c | 2484 | AES_CRYP_KeyInitStructure.CRYP_Key2Left = enc_key[0]; |
wolfSSL | 15:117db924cf7c | 2485 | AES_CRYP_KeyInitStructure.CRYP_Key2Right = enc_key[1]; |
wolfSSL | 15:117db924cf7c | 2486 | AES_CRYP_KeyInitStructure.CRYP_Key3Left = enc_key[2]; |
wolfSSL | 15:117db924cf7c | 2487 | AES_CRYP_KeyInitStructure.CRYP_Key3Right = enc_key[3]; |
wolfSSL | 15:117db924cf7c | 2488 | break; |
wolfSSL | 15:117db924cf7c | 2489 | |
wolfSSL | 15:117db924cf7c | 2490 | case 12: /* 192-bit key */ |
wolfSSL | 15:117db924cf7c | 2491 | AES_CRYP_InitStructure.CRYP_KeySize = CRYP_KeySize_192b; |
wolfSSL | 15:117db924cf7c | 2492 | AES_CRYP_KeyInitStructure.CRYP_Key1Left = enc_key[0]; |
wolfSSL | 15:117db924cf7c | 2493 | AES_CRYP_KeyInitStructure.CRYP_Key1Right = enc_key[1]; |
wolfSSL | 15:117db924cf7c | 2494 | AES_CRYP_KeyInitStructure.CRYP_Key2Left = enc_key[2]; |
wolfSSL | 15:117db924cf7c | 2495 | AES_CRYP_KeyInitStructure.CRYP_Key2Right = enc_key[3]; |
wolfSSL | 15:117db924cf7c | 2496 | AES_CRYP_KeyInitStructure.CRYP_Key3Left = enc_key[4]; |
wolfSSL | 15:117db924cf7c | 2497 | AES_CRYP_KeyInitStructure.CRYP_Key3Right = enc_key[5]; |
wolfSSL | 15:117db924cf7c | 2498 | break; |
wolfSSL | 15:117db924cf7c | 2499 | |
wolfSSL | 15:117db924cf7c | 2500 | case 14: /* 256-bit key */ |
wolfSSL | 15:117db924cf7c | 2501 | AES_CRYP_InitStructure.CRYP_KeySize = CRYP_KeySize_256b; |
wolfSSL | 15:117db924cf7c | 2502 | AES_CRYP_KeyInitStructure.CRYP_Key0Left = enc_key[0]; |
wolfSSL | 15:117db924cf7c | 2503 | AES_CRYP_KeyInitStructure.CRYP_Key0Right = enc_key[1]; |
wolfSSL | 15:117db924cf7c | 2504 | AES_CRYP_KeyInitStructure.CRYP_Key1Left = enc_key[2]; |
wolfSSL | 15:117db924cf7c | 2505 | AES_CRYP_KeyInitStructure.CRYP_Key1Right = enc_key[3]; |
wolfSSL | 15:117db924cf7c | 2506 | AES_CRYP_KeyInitStructure.CRYP_Key2Left = enc_key[4]; |
wolfSSL | 15:117db924cf7c | 2507 | AES_CRYP_KeyInitStructure.CRYP_Key2Right = enc_key[5]; |
wolfSSL | 15:117db924cf7c | 2508 | AES_CRYP_KeyInitStructure.CRYP_Key3Left = enc_key[6]; |
wolfSSL | 15:117db924cf7c | 2509 | AES_CRYP_KeyInitStructure.CRYP_Key3Right = enc_key[7]; |
wolfSSL | 15:117db924cf7c | 2510 | break; |
wolfSSL | 15:117db924cf7c | 2511 | |
wolfSSL | 15:117db924cf7c | 2512 | default: |
wolfSSL | 15:117db924cf7c | 2513 | break; |
wolfSSL | 15:117db924cf7c | 2514 | } |
wolfSSL | 15:117db924cf7c | 2515 | CRYP_KeyInit(&AES_CRYP_KeyInitStructure); |
wolfSSL | 15:117db924cf7c | 2516 | |
wolfSSL | 15:117db924cf7c | 2517 | /* set iv */ |
wolfSSL | 15:117db924cf7c | 2518 | ByteReverseWords(iv, iv, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2519 | AES_CRYP_IVInitStructure.CRYP_IV0Left = iv[0]; |
wolfSSL | 15:117db924cf7c | 2520 | AES_CRYP_IVInitStructure.CRYP_IV0Right = iv[1]; |
wolfSSL | 15:117db924cf7c | 2521 | AES_CRYP_IVInitStructure.CRYP_IV1Left = iv[2]; |
wolfSSL | 15:117db924cf7c | 2522 | AES_CRYP_IVInitStructure.CRYP_IV1Right = iv[3]; |
wolfSSL | 15:117db924cf7c | 2523 | CRYP_IVInit(&AES_CRYP_IVInitStructure); |
wolfSSL | 15:117db924cf7c | 2524 | |
wolfSSL | 15:117db924cf7c | 2525 | /* set direction, mode, and datatype */ |
wolfSSL | 15:117db924cf7c | 2526 | AES_CRYP_InitStructure.CRYP_AlgoDir = CRYP_AlgoDir_Encrypt; |
wolfSSL | 15:117db924cf7c | 2527 | AES_CRYP_InitStructure.CRYP_AlgoMode = CRYP_AlgoMode_AES_CBC; |
wolfSSL | 15:117db924cf7c | 2528 | AES_CRYP_InitStructure.CRYP_DataType = CRYP_DataType_8b; |
wolfSSL | 15:117db924cf7c | 2529 | CRYP_Init(&AES_CRYP_InitStructure); |
wolfSSL | 15:117db924cf7c | 2530 | |
wolfSSL | 15:117db924cf7c | 2531 | /* enable crypto processor */ |
wolfSSL | 15:117db924cf7c | 2532 | CRYP_Cmd(ENABLE); |
wolfSSL | 15:117db924cf7c | 2533 | |
wolfSSL | 15:117db924cf7c | 2534 | while (blocks--) { |
wolfSSL | 15:117db924cf7c | 2535 | /* flush IN/OUT FIFOs */ |
wolfSSL | 15:117db924cf7c | 2536 | CRYP_FIFOFlush(); |
wolfSSL | 15:117db924cf7c | 2537 | |
wolfSSL | 15:117db924cf7c | 2538 | CRYP_DataIn(*(uint32_t*)&in[0]); |
wolfSSL | 15:117db924cf7c | 2539 | CRYP_DataIn(*(uint32_t*)&in[4]); |
wolfSSL | 15:117db924cf7c | 2540 | CRYP_DataIn(*(uint32_t*)&in[8]); |
wolfSSL | 15:117db924cf7c | 2541 | CRYP_DataIn(*(uint32_t*)&in[12]); |
wolfSSL | 15:117db924cf7c | 2542 | |
wolfSSL | 15:117db924cf7c | 2543 | /* wait until the complete message has been processed */ |
wolfSSL | 15:117db924cf7c | 2544 | while (CRYP_GetFlagStatus(CRYP_FLAG_BUSY) != RESET) {} |
wolfSSL | 15:117db924cf7c | 2545 | |
wolfSSL | 15:117db924cf7c | 2546 | *(uint32_t*)&out[0] = CRYP_DataOut(); |
wolfSSL | 15:117db924cf7c | 2547 | *(uint32_t*)&out[4] = CRYP_DataOut(); |
wolfSSL | 15:117db924cf7c | 2548 | *(uint32_t*)&out[8] = CRYP_DataOut(); |
wolfSSL | 15:117db924cf7c | 2549 | *(uint32_t*)&out[12] = CRYP_DataOut(); |
wolfSSL | 15:117db924cf7c | 2550 | |
wolfSSL | 15:117db924cf7c | 2551 | /* store iv for next call */ |
wolfSSL | 15:117db924cf7c | 2552 | XMEMCPY(aes->reg, out + sz - AES_BLOCK_SIZE, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2553 | |
wolfSSL | 15:117db924cf7c | 2554 | sz -= AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 2555 | in += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 2556 | out += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 2557 | } |
wolfSSL | 15:117db924cf7c | 2558 | |
wolfSSL | 15:117db924cf7c | 2559 | /* disable crypto processor */ |
wolfSSL | 15:117db924cf7c | 2560 | CRYP_Cmd(DISABLE); |
wolfSSL | 15:117db924cf7c | 2561 | |
wolfSSL | 15:117db924cf7c | 2562 | return 0; |
wolfSSL | 15:117db924cf7c | 2563 | } |
wolfSSL | 15:117db924cf7c | 2564 | |
wolfSSL | 15:117db924cf7c | 2565 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 2566 | int wc_AesCbcDecrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 2567 | { |
wolfSSL | 15:117db924cf7c | 2568 | word32 *dec_key, *iv; |
wolfSSL | 15:117db924cf7c | 2569 | word32 blocks = (sz / AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2570 | CRYP_InitTypeDef AES_CRYP_InitStructure; |
wolfSSL | 15:117db924cf7c | 2571 | CRYP_KeyInitTypeDef AES_CRYP_KeyInitStructure; |
wolfSSL | 15:117db924cf7c | 2572 | CRYP_IVInitTypeDef AES_CRYP_IVInitStructure; |
wolfSSL | 15:117db924cf7c | 2573 | |
wolfSSL | 15:117db924cf7c | 2574 | dec_key = aes->key; |
wolfSSL | 15:117db924cf7c | 2575 | iv = aes->reg; |
wolfSSL | 15:117db924cf7c | 2576 | |
wolfSSL | 15:117db924cf7c | 2577 | /* crypto structure initialization */ |
wolfSSL | 15:117db924cf7c | 2578 | CRYP_KeyStructInit(&AES_CRYP_KeyInitStructure); |
wolfSSL | 15:117db924cf7c | 2579 | CRYP_StructInit(&AES_CRYP_InitStructure); |
wolfSSL | 15:117db924cf7c | 2580 | CRYP_IVStructInit(&AES_CRYP_IVInitStructure); |
wolfSSL | 15:117db924cf7c | 2581 | |
wolfSSL | 15:117db924cf7c | 2582 | /* if input and output same will overwrite input iv */ |
wolfSSL | 15:117db924cf7c | 2583 | XMEMCPY(aes->tmp, in + sz - AES_BLOCK_SIZE, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2584 | |
wolfSSL | 15:117db924cf7c | 2585 | /* reset registers to their default values */ |
wolfSSL | 15:117db924cf7c | 2586 | CRYP_DeInit(); |
wolfSSL | 15:117db924cf7c | 2587 | |
wolfSSL | 15:117db924cf7c | 2588 | /* load key into correct registers */ |
wolfSSL | 15:117db924cf7c | 2589 | switch (aes->rounds) { |
wolfSSL | 15:117db924cf7c | 2590 | case 10: /* 128-bit key */ |
wolfSSL | 15:117db924cf7c | 2591 | AES_CRYP_InitStructure.CRYP_KeySize = CRYP_KeySize_128b; |
wolfSSL | 15:117db924cf7c | 2592 | AES_CRYP_KeyInitStructure.CRYP_Key2Left = dec_key[0]; |
wolfSSL | 15:117db924cf7c | 2593 | AES_CRYP_KeyInitStructure.CRYP_Key2Right = dec_key[1]; |
wolfSSL | 15:117db924cf7c | 2594 | AES_CRYP_KeyInitStructure.CRYP_Key3Left = dec_key[2]; |
wolfSSL | 15:117db924cf7c | 2595 | AES_CRYP_KeyInitStructure.CRYP_Key3Right = dec_key[3]; |
wolfSSL | 15:117db924cf7c | 2596 | break; |
wolfSSL | 15:117db924cf7c | 2597 | |
wolfSSL | 15:117db924cf7c | 2598 | case 12: /* 192-bit key */ |
wolfSSL | 15:117db924cf7c | 2599 | AES_CRYP_InitStructure.CRYP_KeySize = CRYP_KeySize_192b; |
wolfSSL | 15:117db924cf7c | 2600 | AES_CRYP_KeyInitStructure.CRYP_Key1Left = dec_key[0]; |
wolfSSL | 15:117db924cf7c | 2601 | AES_CRYP_KeyInitStructure.CRYP_Key1Right = dec_key[1]; |
wolfSSL | 15:117db924cf7c | 2602 | AES_CRYP_KeyInitStructure.CRYP_Key2Left = dec_key[2]; |
wolfSSL | 15:117db924cf7c | 2603 | AES_CRYP_KeyInitStructure.CRYP_Key2Right = dec_key[3]; |
wolfSSL | 15:117db924cf7c | 2604 | AES_CRYP_KeyInitStructure.CRYP_Key3Left = dec_key[4]; |
wolfSSL | 15:117db924cf7c | 2605 | AES_CRYP_KeyInitStructure.CRYP_Key3Right = dec_key[5]; |
wolfSSL | 15:117db924cf7c | 2606 | break; |
wolfSSL | 15:117db924cf7c | 2607 | |
wolfSSL | 15:117db924cf7c | 2608 | case 14: /* 256-bit key */ |
wolfSSL | 15:117db924cf7c | 2609 | AES_CRYP_InitStructure.CRYP_KeySize = CRYP_KeySize_256b; |
wolfSSL | 15:117db924cf7c | 2610 | AES_CRYP_KeyInitStructure.CRYP_Key0Left = dec_key[0]; |
wolfSSL | 15:117db924cf7c | 2611 | AES_CRYP_KeyInitStructure.CRYP_Key0Right = dec_key[1]; |
wolfSSL | 15:117db924cf7c | 2612 | AES_CRYP_KeyInitStructure.CRYP_Key1Left = dec_key[2]; |
wolfSSL | 15:117db924cf7c | 2613 | AES_CRYP_KeyInitStructure.CRYP_Key1Right = dec_key[3]; |
wolfSSL | 15:117db924cf7c | 2614 | AES_CRYP_KeyInitStructure.CRYP_Key2Left = dec_key[4]; |
wolfSSL | 15:117db924cf7c | 2615 | AES_CRYP_KeyInitStructure.CRYP_Key2Right = dec_key[5]; |
wolfSSL | 15:117db924cf7c | 2616 | AES_CRYP_KeyInitStructure.CRYP_Key3Left = dec_key[6]; |
wolfSSL | 15:117db924cf7c | 2617 | AES_CRYP_KeyInitStructure.CRYP_Key3Right = dec_key[7]; |
wolfSSL | 15:117db924cf7c | 2618 | break; |
wolfSSL | 15:117db924cf7c | 2619 | |
wolfSSL | 15:117db924cf7c | 2620 | default: |
wolfSSL | 15:117db924cf7c | 2621 | break; |
wolfSSL | 15:117db924cf7c | 2622 | } |
wolfSSL | 15:117db924cf7c | 2623 | |
wolfSSL | 15:117db924cf7c | 2624 | /* set direction, mode, and datatype for key preparation */ |
wolfSSL | 15:117db924cf7c | 2625 | AES_CRYP_InitStructure.CRYP_AlgoDir = CRYP_AlgoDir_Decrypt; |
wolfSSL | 15:117db924cf7c | 2626 | AES_CRYP_InitStructure.CRYP_AlgoMode = CRYP_AlgoMode_AES_Key; |
wolfSSL | 15:117db924cf7c | 2627 | AES_CRYP_InitStructure.CRYP_DataType = CRYP_DataType_32b; |
wolfSSL | 15:117db924cf7c | 2628 | CRYP_Init(&AES_CRYP_InitStructure); |
wolfSSL | 15:117db924cf7c | 2629 | CRYP_KeyInit(&AES_CRYP_KeyInitStructure); |
wolfSSL | 15:117db924cf7c | 2630 | |
wolfSSL | 15:117db924cf7c | 2631 | /* enable crypto processor */ |
wolfSSL | 15:117db924cf7c | 2632 | CRYP_Cmd(ENABLE); |
wolfSSL | 15:117db924cf7c | 2633 | |
wolfSSL | 15:117db924cf7c | 2634 | /* wait until key has been prepared */ |
wolfSSL | 15:117db924cf7c | 2635 | while (CRYP_GetFlagStatus(CRYP_FLAG_BUSY) != RESET) {} |
wolfSSL | 15:117db924cf7c | 2636 | |
wolfSSL | 15:117db924cf7c | 2637 | /* set direction, mode, and datatype for decryption */ |
wolfSSL | 15:117db924cf7c | 2638 | AES_CRYP_InitStructure.CRYP_AlgoDir = CRYP_AlgoDir_Decrypt; |
wolfSSL | 15:117db924cf7c | 2639 | AES_CRYP_InitStructure.CRYP_AlgoMode = CRYP_AlgoMode_AES_CBC; |
wolfSSL | 15:117db924cf7c | 2640 | AES_CRYP_InitStructure.CRYP_DataType = CRYP_DataType_8b; |
wolfSSL | 15:117db924cf7c | 2641 | CRYP_Init(&AES_CRYP_InitStructure); |
wolfSSL | 15:117db924cf7c | 2642 | |
wolfSSL | 15:117db924cf7c | 2643 | /* set iv */ |
wolfSSL | 15:117db924cf7c | 2644 | ByteReverseWords(iv, iv, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2645 | |
wolfSSL | 15:117db924cf7c | 2646 | AES_CRYP_IVInitStructure.CRYP_IV0Left = iv[0]; |
wolfSSL | 15:117db924cf7c | 2647 | AES_CRYP_IVInitStructure.CRYP_IV0Right = iv[1]; |
wolfSSL | 15:117db924cf7c | 2648 | AES_CRYP_IVInitStructure.CRYP_IV1Left = iv[2]; |
wolfSSL | 15:117db924cf7c | 2649 | AES_CRYP_IVInitStructure.CRYP_IV1Right = iv[3]; |
wolfSSL | 15:117db924cf7c | 2650 | CRYP_IVInit(&AES_CRYP_IVInitStructure); |
wolfSSL | 15:117db924cf7c | 2651 | |
wolfSSL | 15:117db924cf7c | 2652 | /* enable crypto processor */ |
wolfSSL | 15:117db924cf7c | 2653 | CRYP_Cmd(ENABLE); |
wolfSSL | 15:117db924cf7c | 2654 | |
wolfSSL | 15:117db924cf7c | 2655 | while (blocks--) { |
wolfSSL | 15:117db924cf7c | 2656 | /* flush IN/OUT FIFOs */ |
wolfSSL | 15:117db924cf7c | 2657 | CRYP_FIFOFlush(); |
wolfSSL | 15:117db924cf7c | 2658 | |
wolfSSL | 15:117db924cf7c | 2659 | CRYP_DataIn(*(uint32_t*)&in[0]); |
wolfSSL | 15:117db924cf7c | 2660 | CRYP_DataIn(*(uint32_t*)&in[4]); |
wolfSSL | 15:117db924cf7c | 2661 | CRYP_DataIn(*(uint32_t*)&in[8]); |
wolfSSL | 15:117db924cf7c | 2662 | CRYP_DataIn(*(uint32_t*)&in[12]); |
wolfSSL | 15:117db924cf7c | 2663 | |
wolfSSL | 15:117db924cf7c | 2664 | /* wait until the complete message has been processed */ |
wolfSSL | 15:117db924cf7c | 2665 | while (CRYP_GetFlagStatus(CRYP_FLAG_BUSY) != RESET) {} |
wolfSSL | 15:117db924cf7c | 2666 | |
wolfSSL | 15:117db924cf7c | 2667 | *(uint32_t*)&out[0] = CRYP_DataOut(); |
wolfSSL | 15:117db924cf7c | 2668 | *(uint32_t*)&out[4] = CRYP_DataOut(); |
wolfSSL | 15:117db924cf7c | 2669 | *(uint32_t*)&out[8] = CRYP_DataOut(); |
wolfSSL | 15:117db924cf7c | 2670 | *(uint32_t*)&out[12] = CRYP_DataOut(); |
wolfSSL | 15:117db924cf7c | 2671 | |
wolfSSL | 15:117db924cf7c | 2672 | /* store iv for next call */ |
wolfSSL | 15:117db924cf7c | 2673 | XMEMCPY(aes->reg, aes->tmp, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2674 | |
wolfSSL | 15:117db924cf7c | 2675 | in += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 2676 | out += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 2677 | } |
wolfSSL | 15:117db924cf7c | 2678 | |
wolfSSL | 15:117db924cf7c | 2679 | /* disable crypto processor */ |
wolfSSL | 15:117db924cf7c | 2680 | CRYP_Cmd(DISABLE); |
wolfSSL | 15:117db924cf7c | 2681 | |
wolfSSL | 15:117db924cf7c | 2682 | return 0; |
wolfSSL | 15:117db924cf7c | 2683 | } |
wolfSSL | 15:117db924cf7c | 2684 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 2685 | #endif /* WOLFSSL_STM32_CUBEMX */ |
wolfSSL | 15:117db924cf7c | 2686 | |
wolfSSL | 15:117db924cf7c | 2687 | #elif defined(HAVE_COLDFIRE_SEC) |
wolfSSL | 15:117db924cf7c | 2688 | static int wc_AesCbcCrypt(Aes* aes, byte* po, const byte* pi, word32 sz, |
wolfSSL | 15:117db924cf7c | 2689 | word32 descHeader) |
wolfSSL | 15:117db924cf7c | 2690 | { |
wolfSSL | 15:117db924cf7c | 2691 | #ifdef DEBUG_WOLFSSL |
wolfSSL | 15:117db924cf7c | 2692 | int i; int stat1, stat2; int ret; |
wolfSSL | 15:117db924cf7c | 2693 | #endif |
wolfSSL | 15:117db924cf7c | 2694 | |
wolfSSL | 15:117db924cf7c | 2695 | int size; |
wolfSSL | 15:117db924cf7c | 2696 | volatile int v; |
wolfSSL | 15:117db924cf7c | 2697 | |
wolfSSL | 15:117db924cf7c | 2698 | if ((pi == NULL) || (po == NULL)) |
wolfSSL | 15:117db924cf7c | 2699 | return BAD_FUNC_ARG; /*wrong pointer*/ |
wolfSSL | 15:117db924cf7c | 2700 | |
wolfSSL | 15:117db924cf7c | 2701 | wc_LockMutex(&Mutex_AesSEC); |
wolfSSL | 15:117db924cf7c | 2702 | |
wolfSSL | 15:117db924cf7c | 2703 | /* Set descriptor for SEC */ |
wolfSSL | 15:117db924cf7c | 2704 | secDesc->length1 = 0x0; |
wolfSSL | 15:117db924cf7c | 2705 | secDesc->pointer1 = NULL; |
wolfSSL | 15:117db924cf7c | 2706 | |
wolfSSL | 15:117db924cf7c | 2707 | secDesc->length2 = AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 2708 | secDesc->pointer2 = (byte *)secReg; /* Initial Vector */ |
wolfSSL | 15:117db924cf7c | 2709 | |
wolfSSL | 15:117db924cf7c | 2710 | switch(aes->rounds) { |
wolfSSL | 15:117db924cf7c | 2711 | case 10: secDesc->length3 = 16; break; |
wolfSSL | 15:117db924cf7c | 2712 | case 12: secDesc->length3 = 24; break; |
wolfSSL | 15:117db924cf7c | 2713 | case 14: secDesc->length3 = 32; break; |
wolfSSL | 15:117db924cf7c | 2714 | } |
wolfSSL | 15:117db924cf7c | 2715 | XMEMCPY(secKey, aes->key, secDesc->length3); |
wolfSSL | 15:117db924cf7c | 2716 | |
wolfSSL | 15:117db924cf7c | 2717 | secDesc->pointer3 = (byte *)secKey; |
wolfSSL | 15:117db924cf7c | 2718 | secDesc->pointer4 = AESBuffIn; |
wolfSSL | 15:117db924cf7c | 2719 | secDesc->pointer5 = AESBuffOut; |
wolfSSL | 15:117db924cf7c | 2720 | secDesc->length6 = 0x0; |
wolfSSL | 15:117db924cf7c | 2721 | secDesc->pointer6 = NULL; |
wolfSSL | 15:117db924cf7c | 2722 | secDesc->length7 = 0x0; |
wolfSSL | 15:117db924cf7c | 2723 | secDesc->pointer7 = NULL; |
wolfSSL | 15:117db924cf7c | 2724 | secDesc->nextDescriptorPtr = NULL; |
wolfSSL | 15:117db924cf7c | 2725 | |
wolfSSL | 15:117db924cf7c | 2726 | while (sz) { |
wolfSSL | 15:117db924cf7c | 2727 | secDesc->header = descHeader; |
wolfSSL | 15:117db924cf7c | 2728 | XMEMCPY(secReg, aes->reg, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2729 | if ((sz % AES_BUFFER_SIZE) == sz) { |
wolfSSL | 15:117db924cf7c | 2730 | size = sz; |
wolfSSL | 15:117db924cf7c | 2731 | sz = 0; |
wolfSSL | 15:117db924cf7c | 2732 | } else { |
wolfSSL | 15:117db924cf7c | 2733 | size = AES_BUFFER_SIZE; |
wolfSSL | 15:117db924cf7c | 2734 | sz -= AES_BUFFER_SIZE; |
wolfSSL | 15:117db924cf7c | 2735 | } |
wolfSSL | 15:117db924cf7c | 2736 | secDesc->length4 = size; |
wolfSSL | 15:117db924cf7c | 2737 | secDesc->length5 = size; |
wolfSSL | 15:117db924cf7c | 2738 | |
wolfSSL | 15:117db924cf7c | 2739 | XMEMCPY(AESBuffIn, pi, size); |
wolfSSL | 15:117db924cf7c | 2740 | if(descHeader == SEC_DESC_AES_CBC_DECRYPT) { |
wolfSSL | 15:117db924cf7c | 2741 | XMEMCPY((void*)aes->tmp, (void*)&(pi[size-AES_BLOCK_SIZE]), |
wolfSSL | 15:117db924cf7c | 2742 | AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2743 | } |
wolfSSL | 15:117db924cf7c | 2744 | |
wolfSSL | 15:117db924cf7c | 2745 | /* Point SEC to the location of the descriptor */ |
wolfSSL | 15:117db924cf7c | 2746 | MCF_SEC_FR0 = (uint32)secDesc; |
wolfSSL | 15:117db924cf7c | 2747 | /* Initialize SEC and wait for encryption to complete */ |
wolfSSL | 15:117db924cf7c | 2748 | MCF_SEC_CCCR0 = 0x0000001a; |
wolfSSL | 15:117db924cf7c | 2749 | /* poll SISR to determine when channel is complete */ |
wolfSSL | 15:117db924cf7c | 2750 | v=0; |
wolfSSL | 15:117db924cf7c | 2751 | |
wolfSSL | 15:117db924cf7c | 2752 | while ((secDesc->header>> 24) != 0xff) v++; |
wolfSSL | 15:117db924cf7c | 2753 | |
wolfSSL | 15:117db924cf7c | 2754 | #ifdef DEBUG_WOLFSSL |
wolfSSL | 15:117db924cf7c | 2755 | ret = MCF_SEC_SISRH; |
wolfSSL | 15:117db924cf7c | 2756 | stat1 = MCF_SEC_AESSR; |
wolfSSL | 15:117db924cf7c | 2757 | stat2 = MCF_SEC_AESISR; |
wolfSSL | 15:117db924cf7c | 2758 | if (ret & 0xe0000000) { |
wolfSSL | 15:117db924cf7c | 2759 | db_printf("Aes_Cbc(i=%d):ISRH=%08x, AESSR=%08x, " |
wolfSSL | 15:117db924cf7c | 2760 | "AESISR=%08x\n", i, ret, stat1, stat2); |
wolfSSL | 15:117db924cf7c | 2761 | } |
wolfSSL | 15:117db924cf7c | 2762 | #endif |
wolfSSL | 15:117db924cf7c | 2763 | |
wolfSSL | 15:117db924cf7c | 2764 | XMEMCPY(po, AESBuffOut, size); |
wolfSSL | 15:117db924cf7c | 2765 | |
wolfSSL | 15:117db924cf7c | 2766 | if (descHeader == SEC_DESC_AES_CBC_ENCRYPT) { |
wolfSSL | 15:117db924cf7c | 2767 | XMEMCPY((void*)aes->reg, (void*)&(po[size-AES_BLOCK_SIZE]), |
wolfSSL | 15:117db924cf7c | 2768 | AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2769 | } else { |
wolfSSL | 15:117db924cf7c | 2770 | XMEMCPY((void*)aes->reg, (void*)aes->tmp, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2771 | } |
wolfSSL | 15:117db924cf7c | 2772 | |
wolfSSL | 15:117db924cf7c | 2773 | pi += size; |
wolfSSL | 15:117db924cf7c | 2774 | po += size; |
wolfSSL | 15:117db924cf7c | 2775 | } |
wolfSSL | 15:117db924cf7c | 2776 | |
wolfSSL | 15:117db924cf7c | 2777 | wc_UnLockMutex(&Mutex_AesSEC); |
wolfSSL | 15:117db924cf7c | 2778 | return 0; |
wolfSSL | 15:117db924cf7c | 2779 | } |
wolfSSL | 15:117db924cf7c | 2780 | |
wolfSSL | 15:117db924cf7c | 2781 | int wc_AesCbcEncrypt(Aes* aes, byte* po, const byte* pi, word32 sz) |
wolfSSL | 15:117db924cf7c | 2782 | { |
wolfSSL | 15:117db924cf7c | 2783 | return (wc_AesCbcCrypt(aes, po, pi, sz, SEC_DESC_AES_CBC_ENCRYPT)); |
wolfSSL | 15:117db924cf7c | 2784 | } |
wolfSSL | 15:117db924cf7c | 2785 | |
wolfSSL | 15:117db924cf7c | 2786 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 2787 | int wc_AesCbcDecrypt(Aes* aes, byte* po, const byte* pi, word32 sz) |
wolfSSL | 15:117db924cf7c | 2788 | { |
wolfSSL | 15:117db924cf7c | 2789 | return (wc_AesCbcCrypt(aes, po, pi, sz, SEC_DESC_AES_CBC_DECRYPT)); |
wolfSSL | 15:117db924cf7c | 2790 | } |
wolfSSL | 15:117db924cf7c | 2791 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 2792 | |
wolfSSL | 15:117db924cf7c | 2793 | #elif defined(FREESCALE_LTC) |
wolfSSL | 15:117db924cf7c | 2794 | int wc_AesCbcEncrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 2795 | { |
wolfSSL | 15:117db924cf7c | 2796 | uint32_t keySize; |
wolfSSL | 15:117db924cf7c | 2797 | status_t status; |
wolfSSL | 15:117db924cf7c | 2798 | byte *iv, *enc_key; |
wolfSSL | 15:117db924cf7c | 2799 | word32 blocks = (sz / AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2800 | |
wolfSSL | 15:117db924cf7c | 2801 | iv = (byte*)aes->reg; |
wolfSSL | 15:117db924cf7c | 2802 | enc_key = (byte*)aes->key; |
wolfSSL | 15:117db924cf7c | 2803 | |
wolfSSL | 15:117db924cf7c | 2804 | status = wc_AesGetKeySize(aes, &keySize); |
wolfSSL | 15:117db924cf7c | 2805 | if (status != 0) { |
wolfSSL | 15:117db924cf7c | 2806 | return status; |
wolfSSL | 15:117db924cf7c | 2807 | } |
wolfSSL | 15:117db924cf7c | 2808 | |
wolfSSL | 15:117db924cf7c | 2809 | status = LTC_AES_EncryptCbc(LTC_BASE, in, out, blocks * AES_BLOCK_SIZE, |
wolfSSL | 15:117db924cf7c | 2810 | iv, enc_key, keySize); |
wolfSSL | 15:117db924cf7c | 2811 | return (status == kStatus_Success) ? 0 : -1; |
wolfSSL | 15:117db924cf7c | 2812 | } |
wolfSSL | 15:117db924cf7c | 2813 | |
wolfSSL | 15:117db924cf7c | 2814 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 2815 | int wc_AesCbcDecrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 2816 | { |
wolfSSL | 15:117db924cf7c | 2817 | uint32_t keySize; |
wolfSSL | 15:117db924cf7c | 2818 | status_t status; |
wolfSSL | 15:117db924cf7c | 2819 | byte* iv, *dec_key; |
wolfSSL | 15:117db924cf7c | 2820 | word32 blocks = (sz / AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2821 | |
wolfSSL | 15:117db924cf7c | 2822 | iv = (byte*)aes->reg; |
wolfSSL | 15:117db924cf7c | 2823 | dec_key = (byte*)aes->key; |
wolfSSL | 15:117db924cf7c | 2824 | |
wolfSSL | 15:117db924cf7c | 2825 | status = wc_AesGetKeySize(aes, &keySize); |
wolfSSL | 15:117db924cf7c | 2826 | if (status != 0) { |
wolfSSL | 15:117db924cf7c | 2827 | return status; |
wolfSSL | 15:117db924cf7c | 2828 | } |
wolfSSL | 15:117db924cf7c | 2829 | |
wolfSSL | 15:117db924cf7c | 2830 | status = LTC_AES_DecryptCbc(LTC_BASE, in, out, blocks * AES_BLOCK_SIZE, |
wolfSSL | 15:117db924cf7c | 2831 | iv, dec_key, keySize, kLTC_EncryptKey); |
wolfSSL | 15:117db924cf7c | 2832 | return (status == kStatus_Success) ? 0 : -1; |
wolfSSL | 15:117db924cf7c | 2833 | } |
wolfSSL | 15:117db924cf7c | 2834 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 2835 | |
wolfSSL | 15:117db924cf7c | 2836 | #elif defined(FREESCALE_MMCAU) |
wolfSSL | 15:117db924cf7c | 2837 | int wc_AesCbcEncrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 2838 | { |
wolfSSL | 15:117db924cf7c | 2839 | int i; |
wolfSSL | 15:117db924cf7c | 2840 | int offset = 0; |
wolfSSL | 15:117db924cf7c | 2841 | word32 blocks = (sz / AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2842 | byte *iv; |
wolfSSL | 15:117db924cf7c | 2843 | byte temp_block[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 2844 | |
wolfSSL | 15:117db924cf7c | 2845 | iv = (byte*)aes->reg; |
wolfSSL | 15:117db924cf7c | 2846 | |
wolfSSL | 15:117db924cf7c | 2847 | while (blocks--) { |
wolfSSL | 15:117db924cf7c | 2848 | XMEMCPY(temp_block, in + offset, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2849 | |
wolfSSL | 15:117db924cf7c | 2850 | /* XOR block with IV for CBC */ |
wolfSSL | 15:117db924cf7c | 2851 | for (i = 0; i < AES_BLOCK_SIZE; i++) |
wolfSSL | 15:117db924cf7c | 2852 | temp_block[i] ^= iv[i]; |
wolfSSL | 15:117db924cf7c | 2853 | |
wolfSSL | 15:117db924cf7c | 2854 | wc_AesEncrypt(aes, temp_block, out + offset); |
wolfSSL | 15:117db924cf7c | 2855 | |
wolfSSL | 15:117db924cf7c | 2856 | offset += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 2857 | |
wolfSSL | 15:117db924cf7c | 2858 | /* store IV for next block */ |
wolfSSL | 15:117db924cf7c | 2859 | XMEMCPY(iv, out + offset - AES_BLOCK_SIZE, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2860 | } |
wolfSSL | 15:117db924cf7c | 2861 | |
wolfSSL | 15:117db924cf7c | 2862 | return 0; |
wolfSSL | 15:117db924cf7c | 2863 | } |
wolfSSL | 15:117db924cf7c | 2864 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 2865 | int wc_AesCbcDecrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 2866 | { |
wolfSSL | 15:117db924cf7c | 2867 | int i; |
wolfSSL | 15:117db924cf7c | 2868 | int offset = 0; |
wolfSSL | 15:117db924cf7c | 2869 | word32 blocks = (sz / AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2870 | byte* iv; |
wolfSSL | 15:117db924cf7c | 2871 | byte temp_block[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 2872 | |
wolfSSL | 15:117db924cf7c | 2873 | iv = (byte*)aes->reg; |
wolfSSL | 15:117db924cf7c | 2874 | |
wolfSSL | 15:117db924cf7c | 2875 | while (blocks--) { |
wolfSSL | 15:117db924cf7c | 2876 | XMEMCPY(temp_block, in + offset, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2877 | |
wolfSSL | 15:117db924cf7c | 2878 | wc_AesDecrypt(aes, in + offset, out + offset); |
wolfSSL | 15:117db924cf7c | 2879 | |
wolfSSL | 15:117db924cf7c | 2880 | /* XOR block with IV for CBC */ |
wolfSSL | 15:117db924cf7c | 2881 | for (i = 0; i < AES_BLOCK_SIZE; i++) |
wolfSSL | 15:117db924cf7c | 2882 | (out + offset)[i] ^= iv[i]; |
wolfSSL | 15:117db924cf7c | 2883 | |
wolfSSL | 15:117db924cf7c | 2884 | /* store IV for next block */ |
wolfSSL | 15:117db924cf7c | 2885 | XMEMCPY(iv, temp_block, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2886 | |
wolfSSL | 15:117db924cf7c | 2887 | offset += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 2888 | } |
wolfSSL | 15:117db924cf7c | 2889 | |
wolfSSL | 15:117db924cf7c | 2890 | return 0; |
wolfSSL | 15:117db924cf7c | 2891 | } |
wolfSSL | 15:117db924cf7c | 2892 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 2893 | |
wolfSSL | 15:117db924cf7c | 2894 | #elif defined(WOLFSSL_PIC32MZ_CRYPT) |
wolfSSL | 15:117db924cf7c | 2895 | |
wolfSSL | 15:117db924cf7c | 2896 | int wc_AesCbcEncrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 2897 | { |
wolfSSL | 15:117db924cf7c | 2898 | int ret; |
wolfSSL | 15:117db924cf7c | 2899 | |
wolfSSL | 15:117db924cf7c | 2900 | /* hardware fails on input that is not a multiple of AES block size */ |
wolfSSL | 15:117db924cf7c | 2901 | if (sz % AES_BLOCK_SIZE != 0) { |
wolfSSL | 15:117db924cf7c | 2902 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2903 | } |
wolfSSL | 15:117db924cf7c | 2904 | |
wolfSSL | 15:117db924cf7c | 2905 | ret = wc_Pic32AesCrypt( |
wolfSSL | 15:117db924cf7c | 2906 | aes->key, aes->keylen, aes->reg, AES_BLOCK_SIZE, |
wolfSSL | 15:117db924cf7c | 2907 | out, in, sz, PIC32_ENCRYPTION, |
wolfSSL | 15:117db924cf7c | 2908 | PIC32_ALGO_AES, PIC32_CRYPTOALGO_RCBC); |
wolfSSL | 15:117db924cf7c | 2909 | |
wolfSSL | 15:117db924cf7c | 2910 | /* store iv for next call */ |
wolfSSL | 15:117db924cf7c | 2911 | if (ret == 0) { |
wolfSSL | 15:117db924cf7c | 2912 | XMEMCPY(aes->reg, out + sz - AES_BLOCK_SIZE, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2913 | } |
wolfSSL | 15:117db924cf7c | 2914 | |
wolfSSL | 15:117db924cf7c | 2915 | return ret; |
wolfSSL | 15:117db924cf7c | 2916 | } |
wolfSSL | 15:117db924cf7c | 2917 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 2918 | int wc_AesCbcDecrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 2919 | { |
wolfSSL | 15:117db924cf7c | 2920 | int ret; |
wolfSSL | 15:117db924cf7c | 2921 | byte scratch[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 2922 | |
wolfSSL | 15:117db924cf7c | 2923 | /* hardware fails on input that is not a multiple of AES block size */ |
wolfSSL | 15:117db924cf7c | 2924 | if (sz % AES_BLOCK_SIZE != 0) { |
wolfSSL | 15:117db924cf7c | 2925 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2926 | } |
wolfSSL | 15:117db924cf7c | 2927 | XMEMCPY(scratch, in + sz - AES_BLOCK_SIZE, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2928 | |
wolfSSL | 15:117db924cf7c | 2929 | ret = wc_Pic32AesCrypt( |
wolfSSL | 15:117db924cf7c | 2930 | aes->key, aes->keylen, aes->reg, AES_BLOCK_SIZE, |
wolfSSL | 15:117db924cf7c | 2931 | out, in, sz, PIC32_DECRYPTION, |
wolfSSL | 15:117db924cf7c | 2932 | PIC32_ALGO_AES, PIC32_CRYPTOALGO_RCBC); |
wolfSSL | 15:117db924cf7c | 2933 | |
wolfSSL | 15:117db924cf7c | 2934 | /* store iv for next call */ |
wolfSSL | 15:117db924cf7c | 2935 | if (ret == 0) { |
wolfSSL | 15:117db924cf7c | 2936 | XMEMCPY((byte*)aes->reg, scratch, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2937 | } |
wolfSSL | 15:117db924cf7c | 2938 | |
wolfSSL | 15:117db924cf7c | 2939 | return ret; |
wolfSSL | 15:117db924cf7c | 2940 | } |
wolfSSL | 15:117db924cf7c | 2941 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 2942 | |
wolfSSL | 15:117db924cf7c | 2943 | #elif defined(WOLFSSL_IMX6_CAAM) && !defined(NO_IMX6_CAAM_AES) |
wolfSSL | 15:117db924cf7c | 2944 | /* implemented in wolfcrypt/src/port/caam/caam_aes.c */ |
wolfSSL | 15:117db924cf7c | 2945 | |
wolfSSL | 15:117db924cf7c | 2946 | #else |
wolfSSL | 15:117db924cf7c | 2947 | |
wolfSSL | 15:117db924cf7c | 2948 | int wc_AesCbcEncrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 2949 | { |
wolfSSL | 15:117db924cf7c | 2950 | word32 blocks = (sz / AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2951 | |
wolfSSL | 15:117db924cf7c | 2952 | if (aes == NULL || out == NULL || in == NULL) { |
wolfSSL | 15:117db924cf7c | 2953 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2954 | } |
wolfSSL | 15:117db924cf7c | 2955 | |
wolfSSL | 15:117db924cf7c | 2956 | #if defined(WOLFSSL_ASYNC_CRYPT) && defined(WC_ASYNC_ENABLE_AES) |
wolfSSL | 15:117db924cf7c | 2957 | /* if async and byte count above threshold */ |
wolfSSL | 15:117db924cf7c | 2958 | if (aes->asyncDev.marker == WOLFSSL_ASYNC_MARKER_AES && |
wolfSSL | 15:117db924cf7c | 2959 | sz >= WC_ASYNC_THRESH_AES_CBC) { |
wolfSSL | 15:117db924cf7c | 2960 | #if defined(HAVE_CAVIUM) |
wolfSSL | 15:117db924cf7c | 2961 | return NitroxAesCbcEncrypt(aes, out, in, sz); |
wolfSSL | 15:117db924cf7c | 2962 | #elif defined(HAVE_INTEL_QA) |
wolfSSL | 15:117db924cf7c | 2963 | return IntelQaSymAesCbcEncrypt(&aes->asyncDev, out, in, sz, |
wolfSSL | 15:117db924cf7c | 2964 | (const byte*)aes->asyncKey, aes->keylen, |
wolfSSL | 15:117db924cf7c | 2965 | (const byte*)aes->asyncIv, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2966 | #else /* WOLFSSL_ASYNC_CRYPT_TEST */ |
wolfSSL | 15:117db924cf7c | 2967 | if (wc_AsyncTestInit(&aes->asyncDev, ASYNC_TEST_AES_CBC_ENCRYPT)) { |
wolfSSL | 15:117db924cf7c | 2968 | WC_ASYNC_TEST* testDev = &aes->asyncDev.test; |
wolfSSL | 15:117db924cf7c | 2969 | testDev->aes.aes = aes; |
wolfSSL | 15:117db924cf7c | 2970 | testDev->aes.out = out; |
wolfSSL | 15:117db924cf7c | 2971 | testDev->aes.in = in; |
wolfSSL | 15:117db924cf7c | 2972 | testDev->aes.sz = sz; |
wolfSSL | 15:117db924cf7c | 2973 | return WC_PENDING_E; |
wolfSSL | 15:117db924cf7c | 2974 | } |
wolfSSL | 15:117db924cf7c | 2975 | #endif |
wolfSSL | 15:117db924cf7c | 2976 | } |
wolfSSL | 15:117db924cf7c | 2977 | #endif /* WOLFSSL_ASYNC_CRYPT */ |
wolfSSL | 15:117db924cf7c | 2978 | |
wolfSSL | 15:117db924cf7c | 2979 | #ifdef WOLFSSL_AESNI |
wolfSSL | 15:117db924cf7c | 2980 | if (haveAESNI) { |
wolfSSL | 15:117db924cf7c | 2981 | #ifdef DEBUG_AESNI |
wolfSSL | 15:117db924cf7c | 2982 | printf("about to aes cbc encrypt\n"); |
wolfSSL | 15:117db924cf7c | 2983 | printf("in = %p\n", in); |
wolfSSL | 15:117db924cf7c | 2984 | printf("out = %p\n", out); |
wolfSSL | 15:117db924cf7c | 2985 | printf("aes->key = %p\n", aes->key); |
wolfSSL | 15:117db924cf7c | 2986 | printf("aes->reg = %p\n", aes->reg); |
wolfSSL | 15:117db924cf7c | 2987 | printf("aes->rounds = %d\n", aes->rounds); |
wolfSSL | 15:117db924cf7c | 2988 | printf("sz = %d\n", sz); |
wolfSSL | 15:117db924cf7c | 2989 | #endif |
wolfSSL | 15:117db924cf7c | 2990 | |
wolfSSL | 15:117db924cf7c | 2991 | /* check alignment, decrypt doesn't need alignment */ |
wolfSSL | 15:117db924cf7c | 2992 | if ((wolfssl_word)in % AESNI_ALIGN) { |
wolfSSL | 15:117db924cf7c | 2993 | #ifndef NO_WOLFSSL_ALLOC_ALIGN |
wolfSSL | 15:117db924cf7c | 2994 | byte* tmp = (byte*)XMALLOC(sz + AES_BLOCK_SIZE + AESNI_ALIGN, |
wolfSSL | 15:117db924cf7c | 2995 | aes->heap, DYNAMIC_TYPE_TMP_BUFFER); |
wolfSSL | 15:117db924cf7c | 2996 | byte* tmp_align; |
wolfSSL | 15:117db924cf7c | 2997 | if (tmp == NULL) return MEMORY_E; |
wolfSSL | 15:117db924cf7c | 2998 | |
wolfSSL | 15:117db924cf7c | 2999 | tmp_align = tmp + (AESNI_ALIGN - ((size_t)tmp % AESNI_ALIGN)); |
wolfSSL | 15:117db924cf7c | 3000 | XMEMCPY(tmp_align, in, sz); |
wolfSSL | 15:117db924cf7c | 3001 | AES_CBC_encrypt(tmp_align, tmp_align, (byte*)aes->reg, sz, |
wolfSSL | 15:117db924cf7c | 3002 | (byte*)aes->key, aes->rounds); |
wolfSSL | 15:117db924cf7c | 3003 | /* store iv for next call */ |
wolfSSL | 15:117db924cf7c | 3004 | XMEMCPY(aes->reg, tmp_align + sz - AES_BLOCK_SIZE, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 3005 | |
wolfSSL | 15:117db924cf7c | 3006 | XMEMCPY(out, tmp_align, sz); |
wolfSSL | 15:117db924cf7c | 3007 | XFREE(tmp, aes->heap, DYNAMIC_TYPE_TMP_BUFFER); |
wolfSSL | 15:117db924cf7c | 3008 | return 0; |
wolfSSL | 15:117db924cf7c | 3009 | #else |
wolfSSL | 15:117db924cf7c | 3010 | WOLFSSL_MSG("AES-CBC encrypt with bad alignment"); |
wolfSSL | 15:117db924cf7c | 3011 | return BAD_ALIGN_E; |
wolfSSL | 15:117db924cf7c | 3012 | #endif |
wolfSSL | 15:117db924cf7c | 3013 | } |
wolfSSL | 15:117db924cf7c | 3014 | |
wolfSSL | 15:117db924cf7c | 3015 | AES_CBC_encrypt(in, out, (byte*)aes->reg, sz, (byte*)aes->key, |
wolfSSL | 15:117db924cf7c | 3016 | aes->rounds); |
wolfSSL | 15:117db924cf7c | 3017 | /* store iv for next call */ |
wolfSSL | 15:117db924cf7c | 3018 | XMEMCPY(aes->reg, out + sz - AES_BLOCK_SIZE, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 3019 | |
wolfSSL | 15:117db924cf7c | 3020 | return 0; |
wolfSSL | 15:117db924cf7c | 3021 | } |
wolfSSL | 15:117db924cf7c | 3022 | #endif |
wolfSSL | 15:117db924cf7c | 3023 | |
wolfSSL | 15:117db924cf7c | 3024 | while (blocks--) { |
wolfSSL | 15:117db924cf7c | 3025 | xorbuf((byte*)aes->reg, in, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 3026 | wc_AesEncrypt(aes, (byte*)aes->reg, (byte*)aes->reg); |
wolfSSL | 15:117db924cf7c | 3027 | XMEMCPY(out, aes->reg, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 3028 | |
wolfSSL | 15:117db924cf7c | 3029 | out += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 3030 | in += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 3031 | } |
wolfSSL | 15:117db924cf7c | 3032 | |
wolfSSL | 15:117db924cf7c | 3033 | return 0; |
wolfSSL | 15:117db924cf7c | 3034 | } |
wolfSSL | 15:117db924cf7c | 3035 | |
wolfSSL | 15:117db924cf7c | 3036 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 3037 | int wc_AesCbcDecrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 3038 | { |
wolfSSL | 15:117db924cf7c | 3039 | word32 blocks; |
wolfSSL | 15:117db924cf7c | 3040 | |
wolfSSL | 15:117db924cf7c | 3041 | if (aes == NULL || out == NULL || in == NULL |
wolfSSL | 15:117db924cf7c | 3042 | || sz % AES_BLOCK_SIZE != 0) { |
wolfSSL | 15:117db924cf7c | 3043 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 3044 | } |
wolfSSL | 15:117db924cf7c | 3045 | |
wolfSSL | 15:117db924cf7c | 3046 | #if defined(WOLFSSL_ASYNC_CRYPT) && defined(WC_ASYNC_ENABLE_AES) |
wolfSSL | 15:117db924cf7c | 3047 | /* if async and byte count above threshold */ |
wolfSSL | 15:117db924cf7c | 3048 | if (aes->asyncDev.marker == WOLFSSL_ASYNC_MARKER_AES && |
wolfSSL | 15:117db924cf7c | 3049 | sz >= WC_ASYNC_THRESH_AES_CBC) { |
wolfSSL | 15:117db924cf7c | 3050 | #if defined(HAVE_CAVIUM) |
wolfSSL | 15:117db924cf7c | 3051 | return NitroxAesCbcDecrypt(aes, out, in, sz); |
wolfSSL | 15:117db924cf7c | 3052 | #elif defined(HAVE_INTEL_QA) |
wolfSSL | 15:117db924cf7c | 3053 | return IntelQaSymAesCbcDecrypt(&aes->asyncDev, out, in, sz, |
wolfSSL | 15:117db924cf7c | 3054 | (const byte*)aes->asyncKey, aes->keylen, |
wolfSSL | 15:117db924cf7c | 3055 | (const byte*)aes->asyncIv, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 3056 | #else /* WOLFSSL_ASYNC_CRYPT_TEST */ |
wolfSSL | 15:117db924cf7c | 3057 | if (wc_AsyncTestInit(&aes->asyncDev, ASYNC_TEST_AES_CBC_DECRYPT)) { |
wolfSSL | 15:117db924cf7c | 3058 | WC_ASYNC_TEST* testDev = &aes->asyncDev.test; |
wolfSSL | 15:117db924cf7c | 3059 | testDev->aes.aes = aes; |
wolfSSL | 15:117db924cf7c | 3060 | testDev->aes.out = out; |
wolfSSL | 15:117db924cf7c | 3061 | testDev->aes.in = in; |
wolfSSL | 15:117db924cf7c | 3062 | testDev->aes.sz = sz; |
wolfSSL | 15:117db924cf7c | 3063 | return WC_PENDING_E; |
wolfSSL | 15:117db924cf7c | 3064 | } |
wolfSSL | 15:117db924cf7c | 3065 | #endif |
wolfSSL | 15:117db924cf7c | 3066 | } |
wolfSSL | 15:117db924cf7c | 3067 | #endif |
wolfSSL | 15:117db924cf7c | 3068 | |
wolfSSL | 15:117db924cf7c | 3069 | #ifdef WOLFSSL_AESNI |
wolfSSL | 15:117db924cf7c | 3070 | if (haveAESNI) { |
wolfSSL | 15:117db924cf7c | 3071 | #ifdef DEBUG_AESNI |
wolfSSL | 15:117db924cf7c | 3072 | printf("about to aes cbc decrypt\n"); |
wolfSSL | 15:117db924cf7c | 3073 | printf("in = %p\n", in); |
wolfSSL | 15:117db924cf7c | 3074 | printf("out = %p\n", out); |
wolfSSL | 15:117db924cf7c | 3075 | printf("aes->key = %p\n", aes->key); |
wolfSSL | 15:117db924cf7c | 3076 | printf("aes->reg = %p\n", aes->reg); |
wolfSSL | 15:117db924cf7c | 3077 | printf("aes->rounds = %d\n", aes->rounds); |
wolfSSL | 15:117db924cf7c | 3078 | printf("sz = %d\n", sz); |
wolfSSL | 15:117db924cf7c | 3079 | #endif |
wolfSSL | 15:117db924cf7c | 3080 | |
wolfSSL | 15:117db924cf7c | 3081 | /* if input and output same will overwrite input iv */ |
wolfSSL | 15:117db924cf7c | 3082 | XMEMCPY(aes->tmp, in + sz - AES_BLOCK_SIZE, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 3083 | #if defined(WOLFSSL_AESNI_BY4) |
wolfSSL | 15:117db924cf7c | 3084 | AES_CBC_decrypt_by4(in, out, (byte*)aes->reg, sz, (byte*)aes->key, |
wolfSSL | 15:117db924cf7c | 3085 | aes->rounds); |
wolfSSL | 15:117db924cf7c | 3086 | #elif defined(WOLFSSL_AESNI_BY6) |
wolfSSL | 15:117db924cf7c | 3087 | AES_CBC_decrypt_by6(in, out, (byte*)aes->reg, sz, (byte*)aes->key, |
wolfSSL | 15:117db924cf7c | 3088 | aes->rounds); |
wolfSSL | 15:117db924cf7c | 3089 | #else /* WOLFSSL_AESNI_BYx */ |
wolfSSL | 15:117db924cf7c | 3090 | AES_CBC_decrypt_by8(in, out, (byte*)aes->reg, sz, (byte*)aes->key, |
wolfSSL | 15:117db924cf7c | 3091 | aes->rounds); |
wolfSSL | 15:117db924cf7c | 3092 | #endif /* WOLFSSL_AESNI_BYx */ |
wolfSSL | 15:117db924cf7c | 3093 | /* store iv for next call */ |
wolfSSL | 15:117db924cf7c | 3094 | XMEMCPY(aes->reg, aes->tmp, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 3095 | return 0; |
wolfSSL | 15:117db924cf7c | 3096 | } |
wolfSSL | 15:117db924cf7c | 3097 | #endif |
wolfSSL | 15:117db924cf7c | 3098 | |
wolfSSL | 15:117db924cf7c | 3099 | blocks = sz / AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 3100 | while (blocks--) { |
wolfSSL | 15:117db924cf7c | 3101 | XMEMCPY(aes->tmp, in, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 3102 | wc_AesDecrypt(aes, (byte*)aes->tmp, out); |
wolfSSL | 15:117db924cf7c | 3103 | xorbuf(out, (byte*)aes->reg, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 3104 | XMEMCPY(aes->reg, aes->tmp, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 3105 | |
wolfSSL | 15:117db924cf7c | 3106 | out += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 3107 | in += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 3108 | } |
wolfSSL | 15:117db924cf7c | 3109 | |
wolfSSL | 15:117db924cf7c | 3110 | return 0; |
wolfSSL | 15:117db924cf7c | 3111 | } |
wolfSSL | 15:117db924cf7c | 3112 | #endif |
wolfSSL | 15:117db924cf7c | 3113 | |
wolfSSL | 15:117db924cf7c | 3114 | #endif /* AES-CBC block */ |
wolfSSL | 15:117db924cf7c | 3115 | #endif /* HAVE_AES_CBC */ |
wolfSSL | 15:117db924cf7c | 3116 | |
wolfSSL | 15:117db924cf7c | 3117 | /* AES-CTR */ |
wolfSSL | 15:117db924cf7c | 3118 | #if defined(WOLFSSL_AES_COUNTER) |
wolfSSL | 15:117db924cf7c | 3119 | |
wolfSSL | 15:117db924cf7c | 3120 | #ifdef STM32_CRYPTO |
wolfSSL | 15:117db924cf7c | 3121 | #define NEED_AES_CTR_SOFT |
wolfSSL | 15:117db924cf7c | 3122 | #define XTRANSFORM_AESCTRBLOCK wc_AesCtrEncryptBlock |
wolfSSL | 15:117db924cf7c | 3123 | |
wolfSSL | 15:117db924cf7c | 3124 | int wc_AesCtrEncryptBlock(Aes* aes, byte* out, const byte* in) |
wolfSSL | 15:117db924cf7c | 3125 | { |
wolfSSL | 15:117db924cf7c | 3126 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 3127 | #ifdef WOLFSSL_STM32_CUBEMX |
wolfSSL | 15:117db924cf7c | 3128 | CRYP_HandleTypeDef hcryp; |
wolfSSL | 15:117db924cf7c | 3129 | |
wolfSSL | 15:117db924cf7c | 3130 | XMEMSET(&hcryp, 0, sizeof(CRYP_HandleTypeDef)); |
wolfSSL | 15:117db924cf7c | 3131 | switch (aes->rounds) { |
wolfSSL | 15:117db924cf7c | 3132 | case 10: /* 128-bit key */ |
wolfSSL | 15:117db924cf7c | 3133 | hcryp.Init.KeySize = CRYP_KEYSIZE_128B; |
wolfSSL | 15:117db924cf7c | 3134 | break; |
wolfSSL | 15:117db924cf7c | 3135 | #ifdef CRYP_KEYSIZE_192B |
wolfSSL | 15:117db924cf7c | 3136 | case 12: /* 192-bit key */ |
wolfSSL | 15:117db924cf7c | 3137 | hcryp.Init.KeySize = CRYP_KEYSIZE_192B; |
wolfSSL | 15:117db924cf7c | 3138 | break; |
wolfSSL | 15:117db924cf7c | 3139 | #endif |
wolfSSL | 15:117db924cf7c | 3140 | case 14: /* 256-bit key */ |
wolfSSL | 15:117db924cf7c | 3141 | hcryp.Init.KeySize = CRYP_KEYSIZE_256B; |
wolfSSL | 15:117db924cf7c | 3142 | break; |
wolfSSL | 15:117db924cf7c | 3143 | default: |
wolfSSL | 15:117db924cf7c | 3144 | break; |
wolfSSL | 15:117db924cf7c | 3145 | } |
wolfSSL | 15:117db924cf7c | 3146 | hcryp.Instance = CRYP; |
wolfSSL | 15:117db924cf7c | 3147 | hcryp.Init.DataType = CRYP_DATATYPE_8B; |
wolfSSL | 15:117db924cf7c | 3148 | hcryp.Init.pKey = (byte*)aes->key; |
wolfSSL | 15:117db924cf7c | 3149 | hcryp.Init.pInitVect = (byte*)aes->reg; |
wolfSSL | 15:117db924cf7c | 3150 | |
wolfSSL | 15:117db924cf7c | 3151 | HAL_CRYP_Init(&hcryp); |
wolfSSL | 15:117db924cf7c | 3152 | |
wolfSSL | 15:117db924cf7c | 3153 | if (HAL_CRYP_AESCTR_Encrypt(&hcryp, (byte*)in, AES_BLOCK_SIZE, out, |
wolfSSL | 15:117db924cf7c | 3154 | STM32_HAL_TIMEOUT) != HAL_OK) { |
wolfSSL | 15:117db924cf7c | 3155 | /* failed */ |
wolfSSL | 15:117db924cf7c | 3156 | ret = WC_TIMEOUT_E; |
wolfSSL | 15:117db924cf7c | 3157 | } |
wolfSSL | 15:117db924cf7c | 3158 | |
wolfSSL | 15:117db924cf7c | 3159 | HAL_CRYP_DeInit(&hcryp); |
wolfSSL | 15:117db924cf7c | 3160 | |
wolfSSL | 15:117db924cf7c | 3161 | #else /* STD_PERI_LIB */ |
wolfSSL | 15:117db924cf7c | 3162 | word32 *enc_key, *iv; |
wolfSSL | 15:117db924cf7c | 3163 | CRYP_InitTypeDef AES_CRYP_InitStructure; |
wolfSSL | 15:117db924cf7c | 3164 | CRYP_KeyInitTypeDef AES_CRYP_KeyInitStructure; |
wolfSSL | 15:117db924cf7c | 3165 | CRYP_IVInitTypeDef AES_CRYP_IVInitStructure; |
wolfSSL | 15:117db924cf7c | 3166 | |
wolfSSL | 15:117db924cf7c | 3167 | enc_key = aes->key; |
wolfSSL | 15:117db924cf7c | 3168 | iv = aes->reg; |
wolfSSL | 15:117db924cf7c | 3169 | |
wolfSSL | 15:117db924cf7c | 3170 | /* crypto structure initialization */ |
wolfSSL | 15:117db924cf7c | 3171 | CRYP_KeyStructInit(&AES_CRYP_KeyInitStructure); |
wolfSSL | 15:117db924cf7c | 3172 | CRYP_StructInit(&AES_CRYP_InitStructure); |
wolfSSL | 15:117db924cf7c | 3173 | CRYP_IVStructInit(&AES_CRYP_IVInitStructure); |
wolfSSL | 15:117db924cf7c | 3174 | |
wolfSSL | 15:117db924cf7c | 3175 | /* reset registers to their default values */ |
wolfSSL | 15:117db924cf7c | 3176 | CRYP_DeInit(); |
wolfSSL | 15:117db924cf7c | 3177 | |
wolfSSL | 15:117db924cf7c | 3178 | /* load key into correct registers */ |
wolfSSL | 15:117db924cf7c | 3179 | switch (aes->rounds) { |
wolfSSL | 15:117db924cf7c | 3180 | case 10: /* 128-bit key */ |
wolfSSL | 15:117db924cf7c | 3181 | AES_CRYP_InitStructure.CRYP_KeySize = CRYP_KeySize_128b; |
wolfSSL | 15:117db924cf7c | 3182 | AES_CRYP_KeyInitStructure.CRYP_Key2Left = enc_key[0]; |
wolfSSL | 15:117db924cf7c | 3183 | AES_CRYP_KeyInitStructure.CRYP_Key2Right = enc_key[1]; |
wolfSSL | 15:117db924cf7c | 3184 | AES_CRYP_KeyInitStructure.CRYP_Key3Left = enc_key[2]; |
wolfSSL | 15:117db924cf7c | 3185 | AES_CRYP_KeyInitStructure.CRYP_Key3Right = enc_key[3]; |
wolfSSL | 15:117db924cf7c | 3186 | break; |
wolfSSL | 15:117db924cf7c | 3187 | case 12: /* 192-bit key */ |
wolfSSL | 15:117db924cf7c | 3188 | AES_CRYP_InitStructure.CRYP_KeySize = CRYP_KeySize_192b; |
wolfSSL | 15:117db924cf7c | 3189 | AES_CRYP_KeyInitStructure.CRYP_Key1Left = enc_key[0]; |
wolfSSL | 15:117db924cf7c | 3190 | AES_CRYP_KeyInitStructure.CRYP_Key1Right = enc_key[1]; |
wolfSSL | 15:117db924cf7c | 3191 | AES_CRYP_KeyInitStructure.CRYP_Key2Left = enc_key[2]; |
wolfSSL | 15:117db924cf7c | 3192 | AES_CRYP_KeyInitStructure.CRYP_Key2Right = enc_key[3]; |
wolfSSL | 15:117db924cf7c | 3193 | AES_CRYP_KeyInitStructure.CRYP_Key3Left = enc_key[4]; |
wolfSSL | 15:117db924cf7c | 3194 | AES_CRYP_KeyInitStructure.CRYP_Key3Right = enc_key[5]; |
wolfSSL | 15:117db924cf7c | 3195 | break; |
wolfSSL | 15:117db924cf7c | 3196 | case 14: /* 256-bit key */ |
wolfSSL | 15:117db924cf7c | 3197 | AES_CRYP_InitStructure.CRYP_KeySize = CRYP_KeySize_256b; |
wolfSSL | 15:117db924cf7c | 3198 | AES_CRYP_KeyInitStructure.CRYP_Key0Left = enc_key[0]; |
wolfSSL | 15:117db924cf7c | 3199 | AES_CRYP_KeyInitStructure.CRYP_Key0Right = enc_key[1]; |
wolfSSL | 15:117db924cf7c | 3200 | AES_CRYP_KeyInitStructure.CRYP_Key1Left = enc_key[2]; |
wolfSSL | 15:117db924cf7c | 3201 | AES_CRYP_KeyInitStructure.CRYP_Key1Right = enc_key[3]; |
wolfSSL | 15:117db924cf7c | 3202 | AES_CRYP_KeyInitStructure.CRYP_Key2Left = enc_key[4]; |
wolfSSL | 15:117db924cf7c | 3203 | AES_CRYP_KeyInitStructure.CRYP_Key2Right = enc_key[5]; |
wolfSSL | 15:117db924cf7c | 3204 | AES_CRYP_KeyInitStructure.CRYP_Key3Left = enc_key[6]; |
wolfSSL | 15:117db924cf7c | 3205 | AES_CRYP_KeyInitStructure.CRYP_Key3Right = enc_key[7]; |
wolfSSL | 15:117db924cf7c | 3206 | break; |
wolfSSL | 15:117db924cf7c | 3207 | default: |
wolfSSL | 15:117db924cf7c | 3208 | break; |
wolfSSL | 15:117db924cf7c | 3209 | } |
wolfSSL | 15:117db924cf7c | 3210 | CRYP_KeyInit(&AES_CRYP_KeyInitStructure); |
wolfSSL | 15:117db924cf7c | 3211 | |
wolfSSL | 15:117db924cf7c | 3212 | /* set iv */ |
wolfSSL | 15:117db924cf7c | 3213 | AES_CRYP_IVInitStructure.CRYP_IV0Left = ByteReverseWord32(iv[0]); |
wolfSSL | 15:117db924cf7c | 3214 | AES_CRYP_IVInitStructure.CRYP_IV0Right = ByteReverseWord32(iv[1]); |
wolfSSL | 15:117db924cf7c | 3215 | AES_CRYP_IVInitStructure.CRYP_IV1Left = ByteReverseWord32(iv[2]); |
wolfSSL | 15:117db924cf7c | 3216 | AES_CRYP_IVInitStructure.CRYP_IV1Right = ByteReverseWord32(iv[3]); |
wolfSSL | 15:117db924cf7c | 3217 | CRYP_IVInit(&AES_CRYP_IVInitStructure); |
wolfSSL | 15:117db924cf7c | 3218 | |
wolfSSL | 15:117db924cf7c | 3219 | /* set direction, mode, and datatype */ |
wolfSSL | 15:117db924cf7c | 3220 | AES_CRYP_InitStructure.CRYP_AlgoDir = CRYP_AlgoDir_Encrypt; |
wolfSSL | 15:117db924cf7c | 3221 | AES_CRYP_InitStructure.CRYP_AlgoMode = CRYP_AlgoMode_AES_CTR; |
wolfSSL | 15:117db924cf7c | 3222 | AES_CRYP_InitStructure.CRYP_DataType = CRYP_DataType_8b; |
wolfSSL | 15:117db924cf7c | 3223 | CRYP_Init(&AES_CRYP_InitStructure); |
wolfSSL | 15:117db924cf7c | 3224 | |
wolfSSL | 15:117db924cf7c | 3225 | /* enable crypto processor */ |
wolfSSL | 15:117db924cf7c | 3226 | CRYP_Cmd(ENABLE); |
wolfSSL | 15:117db924cf7c | 3227 | |
wolfSSL | 15:117db924cf7c | 3228 | /* flush IN/OUT FIFOs */ |
wolfSSL | 15:117db924cf7c | 3229 | CRYP_FIFOFlush(); |
wolfSSL | 15:117db924cf7c | 3230 | |
wolfSSL | 15:117db924cf7c | 3231 | CRYP_DataIn(*(uint32_t*)&in[0]); |
wolfSSL | 15:117db924cf7c | 3232 | CRYP_DataIn(*(uint32_t*)&in[4]); |
wolfSSL | 15:117db924cf7c | 3233 | CRYP_DataIn(*(uint32_t*)&in[8]); |
wolfSSL | 15:117db924cf7c | 3234 | CRYP_DataIn(*(uint32_t*)&in[12]); |
wolfSSL | 15:117db924cf7c | 3235 | |
wolfSSL | 15:117db924cf7c | 3236 | /* wait until the complete message has been processed */ |
wolfSSL | 15:117db924cf7c | 3237 | while (CRYP_GetFlagStatus(CRYP_FLAG_BUSY) != RESET) {} |
wolfSSL | 15:117db924cf7c | 3238 | |
wolfSSL | 15:117db924cf7c | 3239 | *(uint32_t*)&out[0] = CRYP_DataOut(); |
wolfSSL | 15:117db924cf7c | 3240 | *(uint32_t*)&out[4] = CRYP_DataOut(); |
wolfSSL | 15:117db924cf7c | 3241 | *(uint32_t*)&out[8] = CRYP_DataOut(); |
wolfSSL | 15:117db924cf7c | 3242 | *(uint32_t*)&out[12] = CRYP_DataOut(); |
wolfSSL | 15:117db924cf7c | 3243 | |
wolfSSL | 15:117db924cf7c | 3244 | /* disable crypto processor */ |
wolfSSL | 15:117db924cf7c | 3245 | CRYP_Cmd(DISABLE); |
wolfSSL | 15:117db924cf7c | 3246 | |
wolfSSL | 15:117db924cf7c | 3247 | #endif /* WOLFSSL_STM32_CUBEMX */ |
wolfSSL | 15:117db924cf7c | 3248 | return ret; |
wolfSSL | 15:117db924cf7c | 3249 | } |
wolfSSL | 15:117db924cf7c | 3250 | |
wolfSSL | 15:117db924cf7c | 3251 | |
wolfSSL | 15:117db924cf7c | 3252 | #elif defined(WOLFSSL_PIC32MZ_CRYPT) |
wolfSSL | 15:117db924cf7c | 3253 | |
wolfSSL | 15:117db924cf7c | 3254 | #define NEED_AES_CTR_SOFT |
wolfSSL | 15:117db924cf7c | 3255 | #define XTRANSFORM_AESCTRBLOCK wc_AesCtrEncryptBlock |
wolfSSL | 15:117db924cf7c | 3256 | |
wolfSSL | 15:117db924cf7c | 3257 | int wc_AesCtrEncryptBlock(Aes* aes, byte* out, const byte* in) |
wolfSSL | 15:117db924cf7c | 3258 | { |
wolfSSL | 15:117db924cf7c | 3259 | word32 tmpIv[AES_BLOCK_SIZE / sizeof(word32)]; |
wolfSSL | 15:117db924cf7c | 3260 | XMEMCPY(tmpIv, aes->reg, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 3261 | return wc_Pic32AesCrypt( |
wolfSSL | 15:117db924cf7c | 3262 | aes->key, aes->keylen, tmpIv, AES_BLOCK_SIZE, |
wolfSSL | 15:117db924cf7c | 3263 | out, in, AES_BLOCK_SIZE, |
wolfSSL | 15:117db924cf7c | 3264 | PIC32_ENCRYPTION, PIC32_ALGO_AES, PIC32_CRYPTOALGO_RCTR); |
wolfSSL | 15:117db924cf7c | 3265 | } |
wolfSSL | 15:117db924cf7c | 3266 | |
wolfSSL | 15:117db924cf7c | 3267 | #elif defined(HAVE_COLDFIRE_SEC) |
wolfSSL | 15:117db924cf7c | 3268 | #error "Coldfire SEC doesn't currently support AES-CTR mode" |
wolfSSL | 15:117db924cf7c | 3269 | |
wolfSSL | 15:117db924cf7c | 3270 | #elif defined(FREESCALE_LTC) |
wolfSSL | 15:117db924cf7c | 3271 | int wc_AesCtrEncrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 3272 | { |
wolfSSL | 15:117db924cf7c | 3273 | uint32_t keySize; |
wolfSSL | 15:117db924cf7c | 3274 | byte *iv, *enc_key; |
wolfSSL | 15:117db924cf7c | 3275 | byte* tmp; |
wolfSSL | 15:117db924cf7c | 3276 | |
wolfSSL | 15:117db924cf7c | 3277 | if (aes == NULL || out == NULL || in == NULL) { |
wolfSSL | 15:117db924cf7c | 3278 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 3279 | } |
wolfSSL | 15:117db924cf7c | 3280 | |
wolfSSL | 15:117db924cf7c | 3281 | /* consume any unused bytes left in aes->tmp */ |
wolfSSL | 15:117db924cf7c | 3282 | tmp = (byte*)aes->tmp + AES_BLOCK_SIZE - aes->left; |
wolfSSL | 15:117db924cf7c | 3283 | while (aes->left && sz) { |
wolfSSL | 15:117db924cf7c | 3284 | *(out++) = *(in++) ^ *(tmp++); |
wolfSSL | 15:117db924cf7c | 3285 | aes->left--; |
wolfSSL | 15:117db924cf7c | 3286 | sz--; |
wolfSSL | 15:117db924cf7c | 3287 | } |
wolfSSL | 15:117db924cf7c | 3288 | |
wolfSSL | 15:117db924cf7c | 3289 | if (sz) { |
wolfSSL | 15:117db924cf7c | 3290 | iv = (byte*)aes->reg; |
wolfSSL | 15:117db924cf7c | 3291 | enc_key = (byte*)aes->key; |
wolfSSL | 15:117db924cf7c | 3292 | |
wolfSSL | 15:117db924cf7c | 3293 | wc_AesGetKeySize(aes, &keySize); |
wolfSSL | 15:117db924cf7c | 3294 | |
wolfSSL | 15:117db924cf7c | 3295 | LTC_AES_CryptCtr(LTC_BASE, in, out, sz, |
wolfSSL | 15:117db924cf7c | 3296 | iv, enc_key, keySize, (byte*)aes->tmp, |
wolfSSL | 15:117db924cf7c | 3297 | (uint32_t*)&aes->left); |
wolfSSL | 15:117db924cf7c | 3298 | } |
wolfSSL | 15:117db924cf7c | 3299 | |
wolfSSL | 15:117db924cf7c | 3300 | return 0; |
wolfSSL | 15:117db924cf7c | 3301 | } |
wolfSSL | 15:117db924cf7c | 3302 | |
wolfSSL | 15:117db924cf7c | 3303 | #elif defined(WOLFSSL_IMX6_CAAM) && !defined(NO_IMX6_CAAM_AES) |
wolfSSL | 15:117db924cf7c | 3304 | /* implemented in wolfcrypt/src/port/caam/caam_aes.c */ |
wolfSSL | 15:117db924cf7c | 3305 | |
wolfSSL | 15:117db924cf7c | 3306 | #else |
wolfSSL | 15:117db924cf7c | 3307 | |
wolfSSL | 15:117db924cf7c | 3308 | /* Use software based AES counter */ |
wolfSSL | 15:117db924cf7c | 3309 | #define NEED_AES_CTR_SOFT |
wolfSSL | 15:117db924cf7c | 3310 | #endif |
wolfSSL | 15:117db924cf7c | 3311 | |
wolfSSL | 15:117db924cf7c | 3312 | #ifdef NEED_AES_CTR_SOFT |
wolfSSL | 15:117db924cf7c | 3313 | /* Increment AES counter */ |
wolfSSL | 15:117db924cf7c | 3314 | static WC_INLINE void IncrementAesCounter(byte* inOutCtr) |
wolfSSL | 15:117db924cf7c | 3315 | { |
wolfSSL | 15:117db924cf7c | 3316 | /* in network byte order so start at end and work back */ |
wolfSSL | 15:117db924cf7c | 3317 | int i; |
wolfSSL | 15:117db924cf7c | 3318 | for (i = AES_BLOCK_SIZE - 1; i >= 0; i--) { |
wolfSSL | 15:117db924cf7c | 3319 | if (++inOutCtr[i]) /* we're done unless we overflow */ |
wolfSSL | 15:117db924cf7c | 3320 | return; |
wolfSSL | 15:117db924cf7c | 3321 | } |
wolfSSL | 15:117db924cf7c | 3322 | } |
wolfSSL | 15:117db924cf7c | 3323 | |
wolfSSL | 15:117db924cf7c | 3324 | int wc_AesCtrEncrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 3325 | { |
wolfSSL | 15:117db924cf7c | 3326 | byte* tmp; |
wolfSSL | 15:117db924cf7c | 3327 | |
wolfSSL | 15:117db924cf7c | 3328 | if (aes == NULL || out == NULL || in == NULL) { |
wolfSSL | 15:117db924cf7c | 3329 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 3330 | } |
wolfSSL | 15:117db924cf7c | 3331 | |
wolfSSL | 15:117db924cf7c | 3332 | /* consume any unused bytes left in aes->tmp */ |
wolfSSL | 15:117db924cf7c | 3333 | tmp = (byte*)aes->tmp + AES_BLOCK_SIZE - aes->left; |
wolfSSL | 15:117db924cf7c | 3334 | while (aes->left && sz) { |
wolfSSL | 15:117db924cf7c | 3335 | *(out++) = *(in++) ^ *(tmp++); |
wolfSSL | 15:117db924cf7c | 3336 | aes->left--; |
wolfSSL | 15:117db924cf7c | 3337 | sz--; |
wolfSSL | 15:117db924cf7c | 3338 | } |
wolfSSL | 15:117db924cf7c | 3339 | |
wolfSSL | 15:117db924cf7c | 3340 | /* do as many block size ops as possible */ |
wolfSSL | 15:117db924cf7c | 3341 | while (sz >= AES_BLOCK_SIZE) { |
wolfSSL | 15:117db924cf7c | 3342 | #ifdef XTRANSFORM_AESCTRBLOCK |
wolfSSL | 15:117db924cf7c | 3343 | XTRANSFORM_AESCTRBLOCK(aes, out, in); |
wolfSSL | 15:117db924cf7c | 3344 | #else |
wolfSSL | 15:117db924cf7c | 3345 | wc_AesEncrypt(aes, (byte*)aes->reg, out); |
wolfSSL | 15:117db924cf7c | 3346 | xorbuf(out, in, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 3347 | #endif |
wolfSSL | 15:117db924cf7c | 3348 | IncrementAesCounter((byte*)aes->reg); |
wolfSSL | 15:117db924cf7c | 3349 | |
wolfSSL | 15:117db924cf7c | 3350 | out += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 3351 | in += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 3352 | sz -= AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 3353 | aes->left = 0; |
wolfSSL | 15:117db924cf7c | 3354 | } |
wolfSSL | 15:117db924cf7c | 3355 | |
wolfSSL | 15:117db924cf7c | 3356 | /* handle non block size remaining and store unused byte count in left */ |
wolfSSL | 15:117db924cf7c | 3357 | if (sz) { |
wolfSSL | 15:117db924cf7c | 3358 | wc_AesEncrypt(aes, (byte*)aes->reg, (byte*)aes->tmp); |
wolfSSL | 15:117db924cf7c | 3359 | IncrementAesCounter((byte*)aes->reg); |
wolfSSL | 15:117db924cf7c | 3360 | |
wolfSSL | 15:117db924cf7c | 3361 | aes->left = AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 3362 | tmp = (byte*)aes->tmp; |
wolfSSL | 15:117db924cf7c | 3363 | |
wolfSSL | 15:117db924cf7c | 3364 | while (sz--) { |
wolfSSL | 15:117db924cf7c | 3365 | *(out++) = *(in++) ^ *(tmp++); |
wolfSSL | 15:117db924cf7c | 3366 | aes->left--; |
wolfSSL | 15:117db924cf7c | 3367 | } |
wolfSSL | 15:117db924cf7c | 3368 | } |
wolfSSL | 15:117db924cf7c | 3369 | |
wolfSSL | 15:117db924cf7c | 3370 | return 0; |
wolfSSL | 15:117db924cf7c | 3371 | } |
wolfSSL | 15:117db924cf7c | 3372 | |
wolfSSL | 15:117db924cf7c | 3373 | #endif /* NEED_AES_CTR_SOFT */ |
wolfSSL | 15:117db924cf7c | 3374 | |
wolfSSL | 15:117db924cf7c | 3375 | #endif /* WOLFSSL_AES_COUNTER */ |
wolfSSL | 15:117db924cf7c | 3376 | #endif /* !WOLFSSL_ARMASM */ |
wolfSSL | 15:117db924cf7c | 3377 | |
wolfSSL | 15:117db924cf7c | 3378 | |
wolfSSL | 15:117db924cf7c | 3379 | /* |
wolfSSL | 15:117db924cf7c | 3380 | * The IV for AES GCM and CCM, stored in struct Aes's member reg, is comprised |
wolfSSL | 15:117db924cf7c | 3381 | * of two parts in order: |
wolfSSL | 15:117db924cf7c | 3382 | * 1. The fixed field which may be 0 or 4 bytes long. In TLS, this is set |
wolfSSL | 15:117db924cf7c | 3383 | * to the implicit IV. |
wolfSSL | 15:117db924cf7c | 3384 | * 2. The explicit IV is generated by wolfCrypt. It needs to be managed |
wolfSSL | 15:117db924cf7c | 3385 | * by wolfCrypt to ensure the IV is unique for each call to encrypt. |
wolfSSL | 15:117db924cf7c | 3386 | * The IV may be a 96-bit random value, or the 32-bit fixed value and a |
wolfSSL | 15:117db924cf7c | 3387 | * 64-bit set of 0 or random data. The final 32-bits of reg is used as a |
wolfSSL | 15:117db924cf7c | 3388 | * block counter during the encryption. |
wolfSSL | 15:117db924cf7c | 3389 | */ |
wolfSSL | 15:117db924cf7c | 3390 | |
wolfSSL | 15:117db924cf7c | 3391 | #if (defined(HAVE_AESGCM) && !defined(WC_NO_RNG)) || defined(HAVE_AESCCM) |
wolfSSL | 15:117db924cf7c | 3392 | static WC_INLINE void IncCtr(byte* ctr, word32 ctrSz) |
wolfSSL | 15:117db924cf7c | 3393 | { |
wolfSSL | 15:117db924cf7c | 3394 | int i; |
wolfSSL | 15:117db924cf7c | 3395 | for (i = ctrSz-1; i >= 0; i--) { |
wolfSSL | 15:117db924cf7c | 3396 | if (++ctr[i]) |
wolfSSL | 15:117db924cf7c | 3397 | break; |
wolfSSL | 15:117db924cf7c | 3398 | } |
wolfSSL | 15:117db924cf7c | 3399 | } |
wolfSSL | 15:117db924cf7c | 3400 | #endif /* HAVE_AESGCM || HAVE_AESCCM */ |
wolfSSL | 15:117db924cf7c | 3401 | |
wolfSSL | 15:117db924cf7c | 3402 | |
wolfSSL | 15:117db924cf7c | 3403 | #ifdef HAVE_AESGCM |
wolfSSL | 15:117db924cf7c | 3404 | |
wolfSSL | 15:117db924cf7c | 3405 | #if defined(HAVE_COLDFIRE_SEC) |
wolfSSL | 15:117db924cf7c | 3406 | #error "Coldfire SEC doesn't currently support AES-GCM mode" |
wolfSSL | 15:117db924cf7c | 3407 | |
wolfSSL | 15:117db924cf7c | 3408 | #elif defined(WOLFSSL_NRF51_AES) |
wolfSSL | 15:117db924cf7c | 3409 | #error "nRF51 doesn't currently support AES-GCM mode" |
wolfSSL | 15:117db924cf7c | 3410 | |
wolfSSL | 15:117db924cf7c | 3411 | #endif |
wolfSSL | 15:117db924cf7c | 3412 | |
wolfSSL | 15:117db924cf7c | 3413 | #ifdef WOLFSSL_ARMASM |
wolfSSL | 15:117db924cf7c | 3414 | /* implementation is located in wolfcrypt/src/port/arm/armv8-aes.c */ |
wolfSSL | 15:117db924cf7c | 3415 | #else /* software + AESNI implementation */ |
wolfSSL | 15:117db924cf7c | 3416 | |
wolfSSL | 15:117db924cf7c | 3417 | #if !defined(FREESCALE_LTC_AES_GCM) |
wolfSSL | 15:117db924cf7c | 3418 | static WC_INLINE void IncrementGcmCounter(byte* inOutCtr) |
wolfSSL | 15:117db924cf7c | 3419 | { |
wolfSSL | 15:117db924cf7c | 3420 | int i; |
wolfSSL | 15:117db924cf7c | 3421 | |
wolfSSL | 15:117db924cf7c | 3422 | /* in network byte order so start at end and work back */ |
wolfSSL | 15:117db924cf7c | 3423 | for (i = AES_BLOCK_SIZE - 1; i >= AES_BLOCK_SIZE - CTR_SZ; i--) { |
wolfSSL | 15:117db924cf7c | 3424 | if (++inOutCtr[i]) /* we're done unless we overflow */ |
wolfSSL | 15:117db924cf7c | 3425 | return; |
wolfSSL | 15:117db924cf7c | 3426 | } |
wolfSSL | 15:117db924cf7c | 3427 | } |
wolfSSL | 15:117db924cf7c | 3428 | #endif /* !FREESCALE_LTC_AES_GCM */ |
wolfSSL | 15:117db924cf7c | 3429 | |
wolfSSL | 15:117db924cf7c | 3430 | #if defined(GCM_SMALL) || defined(GCM_TABLE) |
wolfSSL | 15:117db924cf7c | 3431 | |
wolfSSL | 15:117db924cf7c | 3432 | static WC_INLINE void FlattenSzInBits(byte* buf, word32 sz) |
wolfSSL | 15:117db924cf7c | 3433 | { |
wolfSSL | 15:117db924cf7c | 3434 | /* Multiply the sz by 8 */ |
wolfSSL | 15:117db924cf7c | 3435 | word32 szHi = (sz >> (8*sizeof(sz) - 3)); |
wolfSSL | 15:117db924cf7c | 3436 | sz <<= 3; |
wolfSSL | 15:117db924cf7c | 3437 | |
wolfSSL | 15:117db924cf7c | 3438 | /* copy over the words of the sz into the destination buffer */ |
wolfSSL | 15:117db924cf7c | 3439 | buf[0] = (szHi >> 24) & 0xff; |
wolfSSL | 15:117db924cf7c | 3440 | buf[1] = (szHi >> 16) & 0xff; |
wolfSSL | 15:117db924cf7c | 3441 | buf[2] = (szHi >> 8) & 0xff; |
wolfSSL | 15:117db924cf7c | 3442 | buf[3] = szHi & 0xff; |
wolfSSL | 15:117db924cf7c | 3443 | buf[4] = (sz >> 24) & 0xff; |
wolfSSL | 15:117db924cf7c | 3444 | buf[5] = (sz >> 16) & 0xff; |
wolfSSL | 15:117db924cf7c | 3445 | buf[6] = (sz >> 8) & 0xff; |
wolfSSL | 15:117db924cf7c | 3446 | buf[7] = sz & 0xff; |
wolfSSL | 15:117db924cf7c | 3447 | } |
wolfSSL | 15:117db924cf7c | 3448 | |
wolfSSL | 15:117db924cf7c | 3449 | |
wolfSSL | 15:117db924cf7c | 3450 | static WC_INLINE void RIGHTSHIFTX(byte* x) |
wolfSSL | 15:117db924cf7c | 3451 | { |
wolfSSL | 15:117db924cf7c | 3452 | int i; |
wolfSSL | 15:117db924cf7c | 3453 | int carryOut = 0; |
wolfSSL | 15:117db924cf7c | 3454 | int carryIn = 0; |
wolfSSL | 15:117db924cf7c | 3455 | int borrow = x[15] & 0x01; |
wolfSSL | 15:117db924cf7c | 3456 | |
wolfSSL | 15:117db924cf7c | 3457 | for (i = 0; i < AES_BLOCK_SIZE; i++) { |
wolfSSL | 15:117db924cf7c | 3458 | carryOut = x[i] & 0x01; |
wolfSSL | 15:117db924cf7c | 3459 | x[i] = (x[i] >> 1) | (carryIn ? 0x80 : 0); |
wolfSSL | 15:117db924cf7c | 3460 | carryIn = carryOut; |
wolfSSL | 15:117db924cf7c | 3461 | } |
wolfSSL | 15:117db924cf7c | 3462 | if (borrow) x[0] ^= 0xE1; |
wolfSSL | 15:117db924cf7c | 3463 | } |
wolfSSL | 15:117db924cf7c | 3464 | |
wolfSSL | 15:117db924cf7c | 3465 | #endif /* defined(GCM_SMALL) || defined(GCM_TABLE) */ |
wolfSSL | 15:117db924cf7c | 3466 | |
wolfSSL | 15:117db924cf7c | 3467 | |
wolfSSL | 15:117db924cf7c | 3468 | #ifdef GCM_TABLE |
wolfSSL | 15:117db924cf7c | 3469 | |
wolfSSL | 15:117db924cf7c | 3470 | static void GenerateM0(Aes* aes) |
wolfSSL | 15:117db924cf7c | 3471 | { |
wolfSSL | 15:117db924cf7c | 3472 | int i, j; |
wolfSSL | 15:117db924cf7c | 3473 | byte (*m)[AES_BLOCK_SIZE] = aes->M0; |
wolfSSL | 15:117db924cf7c | 3474 | |
wolfSSL | 15:117db924cf7c | 3475 | XMEMCPY(m[128], aes->H, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 3476 | |
wolfSSL | 15:117db924cf7c | 3477 | for (i = 64; i > 0; i /= 2) { |
wolfSSL | 15:117db924cf7c | 3478 | XMEMCPY(m[i], m[i*2], AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 3479 | RIGHTSHIFTX(m[i]); |
wolfSSL | 15:117db924cf7c | 3480 | } |
wolfSSL | 15:117db924cf7c | 3481 | |
wolfSSL | 15:117db924cf7c | 3482 | for (i = 2; i < 256; i *= 2) { |
wolfSSL | 15:117db924cf7c | 3483 | for (j = 1; j < i; j++) { |
wolfSSL | 15:117db924cf7c | 3484 | XMEMCPY(m[i+j], m[i], AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 3485 | xorbuf(m[i+j], m[j], AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 3486 | } |
wolfSSL | 15:117db924cf7c | 3487 | } |
wolfSSL | 15:117db924cf7c | 3488 | |
wolfSSL | 15:117db924cf7c | 3489 | XMEMSET(m[0], 0, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 3490 | } |
wolfSSL | 15:117db924cf7c | 3491 | |
wolfSSL | 15:117db924cf7c | 3492 | #endif /* GCM_TABLE */ |
wolfSSL | 15:117db924cf7c | 3493 | |
wolfSSL | 15:117db924cf7c | 3494 | |
wolfSSL | 15:117db924cf7c | 3495 | int wc_AesGcmSetKey(Aes* aes, const byte* key, word32 len) |
wolfSSL | 15:117db924cf7c | 3496 | { |
wolfSSL | 15:117db924cf7c | 3497 | int ret; |
wolfSSL | 15:117db924cf7c | 3498 | byte iv[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 3499 | |
wolfSSL | 15:117db924cf7c | 3500 | #ifdef WOLFSSL_IMX6_CAAM_BLOB |
wolfSSL | 15:117db924cf7c | 3501 | byte local[32]; |
wolfSSL | 15:117db924cf7c | 3502 | word32 localSz = 32; |
wolfSSL | 15:117db924cf7c | 3503 | |
wolfSSL | 15:117db924cf7c | 3504 | if (len == (16 + WC_CAAM_BLOB_SZ) || |
wolfSSL | 15:117db924cf7c | 3505 | len == (24 + WC_CAAM_BLOB_SZ) || |
wolfSSL | 15:117db924cf7c | 3506 | len == (32 + WC_CAAM_BLOB_SZ)) { |
wolfSSL | 15:117db924cf7c | 3507 | if (wc_caamOpenBlob((byte*)key, len, local, &localSz) != 0) { |
wolfSSL | 15:117db924cf7c | 3508 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 3509 | } |
wolfSSL | 15:117db924cf7c | 3510 | |
wolfSSL | 15:117db924cf7c | 3511 | /* set local values */ |
wolfSSL | 15:117db924cf7c | 3512 | key = local; |
wolfSSL | 15:117db924cf7c | 3513 | len = localSz; |
wolfSSL | 15:117db924cf7c | 3514 | } |
wolfSSL | 15:117db924cf7c | 3515 | #endif |
wolfSSL | 15:117db924cf7c | 3516 | |
wolfSSL | 15:117db924cf7c | 3517 | if (!((len == 16) || (len == 24) || (len == 32))) |
wolfSSL | 15:117db924cf7c | 3518 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 3519 | |
wolfSSL | 15:117db924cf7c | 3520 | XMEMSET(iv, 0, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 3521 | ret = wc_AesSetKey(aes, key, len, iv, AES_ENCRYPTION); |
wolfSSL | 15:117db924cf7c | 3522 | |
wolfSSL | 15:117db924cf7c | 3523 | #ifdef WOLFSSL_AESNI |
wolfSSL | 15:117db924cf7c | 3524 | /* AES-NI code generates its own H value. */ |
wolfSSL | 15:117db924cf7c | 3525 | if (haveAESNI) |
wolfSSL | 15:117db924cf7c | 3526 | return ret; |
wolfSSL | 15:117db924cf7c | 3527 | #endif /* WOLFSSL_AESNI */ |
wolfSSL | 15:117db924cf7c | 3528 | |
wolfSSL | 15:117db924cf7c | 3529 | #if !defined(FREESCALE_LTC_AES_GCM) |
wolfSSL | 15:117db924cf7c | 3530 | if (ret == 0) { |
wolfSSL | 15:117db924cf7c | 3531 | wc_AesEncrypt(aes, iv, aes->H); |
wolfSSL | 15:117db924cf7c | 3532 | #ifdef GCM_TABLE |
wolfSSL | 15:117db924cf7c | 3533 | GenerateM0(aes); |
wolfSSL | 15:117db924cf7c | 3534 | #endif /* GCM_TABLE */ |
wolfSSL | 15:117db924cf7c | 3535 | } |
wolfSSL | 15:117db924cf7c | 3536 | #endif /* FREESCALE_LTC_AES_GCM */ |
wolfSSL | 15:117db924cf7c | 3537 | |
wolfSSL | 15:117db924cf7c | 3538 | #if defined(WOLFSSL_XILINX_CRYPT) |
wolfSSL | 15:117db924cf7c | 3539 | wc_AesGcmSetKey_ex(aes, key, len, XSECURE_CSU_AES_KEY_SRC_KUP); |
wolfSSL | 15:117db924cf7c | 3540 | #endif |
wolfSSL | 15:117db924cf7c | 3541 | |
wolfSSL | 15:117db924cf7c | 3542 | #ifdef WOLFSSL_IMX6_CAAM_BLOB |
wolfSSL | 15:117db924cf7c | 3543 | ForceZero(local, sizeof(local)); |
wolfSSL | 15:117db924cf7c | 3544 | #endif |
wolfSSL | 15:117db924cf7c | 3545 | |
wolfSSL | 15:117db924cf7c | 3546 | return ret; |
wolfSSL | 15:117db924cf7c | 3547 | } |
wolfSSL | 15:117db924cf7c | 3548 | |
wolfSSL | 15:117db924cf7c | 3549 | |
wolfSSL | 15:117db924cf7c | 3550 | #ifdef WOLFSSL_AESNI |
wolfSSL | 15:117db924cf7c | 3551 | |
wolfSSL | 15:117db924cf7c | 3552 | #if defined(USE_INTEL_SPEEDUP) |
wolfSSL | 15:117db924cf7c | 3553 | #define HAVE_INTEL_AVX1 |
wolfSSL | 15:117db924cf7c | 3554 | #define HAVE_INTEL_AVX2 |
wolfSSL | 15:117db924cf7c | 3555 | #endif /* USE_INTEL_SPEEDUP */ |
wolfSSL | 15:117db924cf7c | 3556 | |
wolfSSL | 15:117db924cf7c | 3557 | #ifdef _MSC_VER |
wolfSSL | 15:117db924cf7c | 3558 | #define S(w,z) ((char)((unsigned long long)(w) >> (8*(7-(z))) & 0xFF)) |
wolfSSL | 15:117db924cf7c | 3559 | #define M128_INIT(x,y) { S((x),7), S((x),6), S((x),5), S((x),4), \ |
wolfSSL | 15:117db924cf7c | 3560 | S((x),3), S((x),2), S((x),1), S((x),0), \ |
wolfSSL | 15:117db924cf7c | 3561 | S((y),7), S((y),6), S((y),5), S((y),4), \ |
wolfSSL | 15:117db924cf7c | 3562 | S((y),3), S((y),2), S((y),1), S((y),0) } |
wolfSSL | 15:117db924cf7c | 3563 | #else |
wolfSSL | 15:117db924cf7c | 3564 | #define M128_INIT(x,y) { (x), (y) } |
wolfSSL | 15:117db924cf7c | 3565 | #endif |
wolfSSL | 15:117db924cf7c | 3566 | |
wolfSSL | 15:117db924cf7c | 3567 | static const __m128i MOD2_128 = M128_INIT(0x1, |
wolfSSL | 15:117db924cf7c | 3568 | (long long int)0xc200000000000000UL); |
wolfSSL | 15:117db924cf7c | 3569 | |
wolfSSL | 15:117db924cf7c | 3570 | |
wolfSSL | 15:117db924cf7c | 3571 | /* See Intel® Carry-Less Multiplication Instruction |
wolfSSL | 15:117db924cf7c | 3572 | * and its Usage for Computing the GCM Mode White Paper |
wolfSSL | 15:117db924cf7c | 3573 | * by Shay Gueron, Intel Mobility Group, Israel Development Center; |
wolfSSL | 15:117db924cf7c | 3574 | * and Michael E. Kounavis, Intel Labs, Circuits and Systems Research */ |
wolfSSL | 15:117db924cf7c | 3575 | |
wolfSSL | 15:117db924cf7c | 3576 | |
wolfSSL | 15:117db924cf7c | 3577 | /* Figure 9. AES-GCM – Encrypt With Single Block Ghash at a Time */ |
wolfSSL | 15:117db924cf7c | 3578 | |
wolfSSL | 15:117db924cf7c | 3579 | static const __m128i ONE = M128_INIT(0x0, 0x1); |
wolfSSL | 15:117db924cf7c | 3580 | #ifndef AES_GCM_AESNI_NO_UNROLL |
wolfSSL | 15:117db924cf7c | 3581 | static const __m128i TWO = M128_INIT(0x0, 0x2); |
wolfSSL | 15:117db924cf7c | 3582 | static const __m128i THREE = M128_INIT(0x0, 0x3); |
wolfSSL | 15:117db924cf7c | 3583 | static const __m128i FOUR = M128_INIT(0x0, 0x4); |
wolfSSL | 15:117db924cf7c | 3584 | static const __m128i FIVE = M128_INIT(0x0, 0x5); |
wolfSSL | 15:117db924cf7c | 3585 | static const __m128i SIX = M128_INIT(0x0, 0x6); |
wolfSSL | 15:117db924cf7c | 3586 | static const __m128i SEVEN = M128_INIT(0x0, 0x7); |
wolfSSL | 15:117db924cf7c | 3587 | static const __m128i EIGHT = M128_INIT(0x0, 0x8); |
wolfSSL | 15:117db924cf7c | 3588 | #endif |
wolfSSL | 15:117db924cf7c | 3589 | static const __m128i BSWAP_EPI64 = M128_INIT(0x0001020304050607, 0x08090a0b0c0d0e0f); |
wolfSSL | 15:117db924cf7c | 3590 | static const __m128i BSWAP_MASK = M128_INIT(0x08090a0b0c0d0e0f, 0x0001020304050607); |
wolfSSL | 15:117db924cf7c | 3591 | |
wolfSSL | 15:117db924cf7c | 3592 | |
wolfSSL | 15:117db924cf7c | 3593 | #ifndef _MSC_VER |
wolfSSL | 15:117db924cf7c | 3594 | |
wolfSSL | 15:117db924cf7c | 3595 | #define _VAR(a) "" #a "" |
wolfSSL | 15:117db924cf7c | 3596 | #define VAR(a) _VAR(a) |
wolfSSL | 15:117db924cf7c | 3597 | |
wolfSSL | 15:117db924cf7c | 3598 | #define HR %%xmm14 |
wolfSSL | 15:117db924cf7c | 3599 | #define XR %%xmm15 |
wolfSSL | 15:117db924cf7c | 3600 | #define KR %%ebx |
wolfSSL | 15:117db924cf7c | 3601 | #define KR64 %%rbx |
wolfSSL | 15:117db924cf7c | 3602 | #if !defined(AES_GCM_AESNI_NO_UNROLL) && !defined(AES_GCM_AVX1_NO_UNROLL) |
wolfSSL | 15:117db924cf7c | 3603 | #define CTR1 128(%%rsp) |
wolfSSL | 15:117db924cf7c | 3604 | #define TR 144(%%rsp) |
wolfSSL | 15:117db924cf7c | 3605 | #define HTR %%rsp |
wolfSSL | 15:117db924cf7c | 3606 | #define STACK_OFFSET 160 |
wolfSSL | 15:117db924cf7c | 3607 | #else |
wolfSSL | 15:117db924cf7c | 3608 | #define CTR1 (%%rsp) |
wolfSSL | 15:117db924cf7c | 3609 | #define TR 16(%%rsp) |
wolfSSL | 15:117db924cf7c | 3610 | #define STACK_OFFSET 32 |
wolfSSL | 15:117db924cf7c | 3611 | #endif |
wolfSSL | 15:117db924cf7c | 3612 | |
wolfSSL | 15:117db924cf7c | 3613 | #define AESENC() \ |
wolfSSL | 15:117db924cf7c | 3614 | "aesenc %%xmm12, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3615 | "aesenc %%xmm12, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3616 | "aesenc %%xmm12, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3617 | "aesenc %%xmm12, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 3618 | "aesenc %%xmm12, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 3619 | "aesenc %%xmm12, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 3620 | "aesenc %%xmm12, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 3621 | "aesenc %%xmm12, %%xmm11\n\t" |
wolfSSL | 15:117db924cf7c | 3622 | |
wolfSSL | 15:117db924cf7c | 3623 | #define AESENC_SET(o) \ |
wolfSSL | 15:117db924cf7c | 3624 | "movdqa " #o "(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 3625 | AESENC() |
wolfSSL | 15:117db924cf7c | 3626 | |
wolfSSL | 15:117db924cf7c | 3627 | #define AESENC_CTR() \ |
wolfSSL | 15:117db924cf7c | 3628 | "movdqu " VAR(CTR1) ", %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3629 | "movdqa %[BSWAP_EPI64], %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3630 | "movdqu %%xmm4, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3631 | "pshufb %%xmm1, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3632 | "movdqa %%xmm0, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3633 | "paddd %[ONE], %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3634 | "pshufb %%xmm1, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3635 | "movdqa %%xmm0, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3636 | "paddd %[TWO], %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3637 | "pshufb %%xmm1, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3638 | "movdqa %%xmm0, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 3639 | "paddd %[THREE], %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 3640 | "pshufb %%xmm1, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 3641 | "movdqa %%xmm0, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 3642 | "paddd %[FOUR], %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 3643 | "pshufb %%xmm1, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 3644 | "movdqa %%xmm0, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 3645 | "paddd %[FIVE], %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 3646 | "pshufb %%xmm1, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 3647 | "movdqa %%xmm0, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 3648 | "paddd %[SIX], %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 3649 | "pshufb %%xmm1, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 3650 | "movdqa %%xmm0, %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 3651 | "paddd %[SEVEN], %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 3652 | "pshufb %%xmm1, %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 3653 | "paddd %[EIGHT], %%xmm0\n\t" |
wolfSSL | 15:117db924cf7c | 3654 | |
wolfSSL | 15:117db924cf7c | 3655 | #define AESENC_XOR() \ |
wolfSSL | 15:117db924cf7c | 3656 | "movdqa (%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 3657 | "movdqu %%xmm0, " VAR(CTR1) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3658 | "pxor %%xmm12, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3659 | "pxor %%xmm12, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3660 | "pxor %%xmm12, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3661 | "pxor %%xmm12, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 3662 | "pxor %%xmm12, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 3663 | "pxor %%xmm12, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 3664 | "pxor %%xmm12, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 3665 | "pxor %%xmm12, %%xmm11\n\t" |
wolfSSL | 15:117db924cf7c | 3666 | |
wolfSSL | 15:117db924cf7c | 3667 | /* Encrypt and carry-less multiply for AVX1. */ |
wolfSSL | 15:117db924cf7c | 3668 | #define AESENC_PCLMUL_1(src, o1, o2, o3) \ |
wolfSSL | 15:117db924cf7c | 3669 | "movdqu " #o3 "(" VAR(HTR) "), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 3670 | "movdqu " #o2 "(" #src "), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3671 | "aesenc " #o1 "(%[KEY]), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3672 | "pshufb %[BSWAP_MASK], %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3673 | "pxor %%xmm2, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3674 | "pshufd $0x4e, %%xmm12, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3675 | "pshufd $0x4e, %%xmm0, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 3676 | "pxor %%xmm12, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3677 | "pxor %%xmm0, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 3678 | "movdqa %%xmm0, %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 3679 | "pclmulqdq $0x11, %%xmm12, %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 3680 | "aesenc " #o1 "(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3681 | "aesenc " #o1 "(%[KEY]), %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3682 | "movdqa %%xmm0, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3683 | "pclmulqdq $0x00, %%xmm12, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3684 | "aesenc " #o1 "(%[KEY]), %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 3685 | "aesenc " #o1 "(%[KEY]), %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 3686 | "pclmulqdq $0x00, %%xmm14, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3687 | "aesenc " #o1 "(%[KEY]), %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 3688 | "aesenc " #o1 "(%[KEY]), %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 3689 | "aesenc " #o1 "(%[KEY]), %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 3690 | "pxor %%xmm2, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3691 | "pxor %%xmm3, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3692 | |
wolfSSL | 15:117db924cf7c | 3693 | #define AESENC_PCLMUL_N(src, o1, o2, o3) \ |
wolfSSL | 15:117db924cf7c | 3694 | "movdqu " #o3 "(" VAR(HTR) "), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 3695 | "movdqu " #o2 "(" #src" ), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3696 | "pshufd $0x4e, %%xmm12, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 3697 | "pshufb %[BSWAP_MASK], %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3698 | "aesenc " #o1 "(%[KEY]), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3699 | "pxor %%xmm12, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 3700 | "pshufd $0x4e, %%xmm0, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 3701 | "pxor %%xmm0, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 3702 | "movdqa %%xmm0, %%xmm15\n\t" \ |
wolfSSL | 15:117db924cf7c | 3703 | "pclmulqdq $0x11, %%xmm12, %%xmm15\n\t" \ |
wolfSSL | 15:117db924cf7c | 3704 | "aesenc " #o1 "(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3705 | "aesenc " #o1 "(%[KEY]), %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3706 | "pclmulqdq $0x00, %%xmm0, %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 3707 | "aesenc " #o1 "(%[KEY]), %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 3708 | "aesenc " #o1 "(%[KEY]), %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 3709 | "pclmulqdq $0x00, %%xmm14, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 3710 | "aesenc " #o1 "(%[KEY]), %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 3711 | "aesenc " #o1 "(%[KEY]), %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 3712 | "aesenc " #o1 "(%[KEY]), %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 3713 | "pxor %%xmm12, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3714 | "pxor %%xmm12, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3715 | "pxor %%xmm15, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3716 | "pxor %%xmm15, %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 3717 | "pxor %%xmm13, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3718 | |
wolfSSL | 15:117db924cf7c | 3719 | #define AESENC_PCLMUL_L(o) \ |
wolfSSL | 15:117db924cf7c | 3720 | "movdqa %%xmm1, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 3721 | "psrldq $8, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3722 | "pslldq $8, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 3723 | "aesenc " #o "(%[KEY]), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3724 | "pxor %%xmm14, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3725 | "pxor %%xmm1, %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 3726 | "movdqa %%xmm2, %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 3727 | "movdqa %%xmm2, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 3728 | "movdqa %%xmm2, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 3729 | "aesenc " #o "(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3730 | "pslld $31, %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 3731 | "pslld $30, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 3732 | "pslld $25, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 3733 | "aesenc " #o "(%[KEY]), %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3734 | "pxor %%xmm13, %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 3735 | "pxor %%xmm14, %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 3736 | "aesenc " #o "(%[KEY]), %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 3737 | "movdqa %%xmm12, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 3738 | "pslldq $12, %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 3739 | "psrldq $4, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 3740 | "aesenc " #o "(%[KEY]), %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 3741 | "pxor %%xmm12, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3742 | "movdqa %%xmm2, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 3743 | "movdqa %%xmm2, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3744 | "movdqa %%xmm2, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3745 | "aesenc " #o "(%[KEY]), %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 3746 | "psrld $1, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 3747 | "psrld $2, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3748 | "psrld $7, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3749 | "aesenc " #o "(%[KEY]), %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 3750 | "pxor %%xmm1, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 3751 | "pxor %%xmm0, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 3752 | "aesenc " #o "(%[KEY]), %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 3753 | "pxor %%xmm13, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 3754 | "pxor %%xmm14, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3755 | "pxor %%xmm3, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3756 | |
wolfSSL | 15:117db924cf7c | 3757 | /* Encrypt and carry-less multiply with last key. */ |
wolfSSL | 15:117db924cf7c | 3758 | #define AESENC_LAST(in, out) \ |
wolfSSL | 15:117db924cf7c | 3759 | "aesenclast %%xmm12, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3760 | "aesenclast %%xmm12, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3761 | "movdqu (" #in "),%%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3762 | "movdqu 16(" #in "),%%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3763 | "pxor %%xmm0, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3764 | "pxor %%xmm1, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3765 | "movdqu %%xmm4, (" #out ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 3766 | "movdqu %%xmm5, 16(" #out ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 3767 | "aesenclast %%xmm12, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3768 | "aesenclast %%xmm12, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 3769 | "movdqu 32(" #in "),%%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3770 | "movdqu 48(" #in "),%%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3771 | "pxor %%xmm0, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3772 | "pxor %%xmm1, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 3773 | "movdqu %%xmm6, 32(" #out ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 3774 | "movdqu %%xmm7, 48(" #out ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 3775 | "aesenclast %%xmm12, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 3776 | "aesenclast %%xmm12, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 3777 | "movdqu 64(" #in "),%%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3778 | "movdqu 80(" #in "),%%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3779 | "pxor %%xmm0, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 3780 | "pxor %%xmm1, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 3781 | "movdqu %%xmm8, 64(" #out ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 3782 | "movdqu %%xmm9, 80(" #out ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 3783 | "aesenclast %%xmm12, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 3784 | "aesenclast %%xmm12, %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 3785 | "movdqu 96(" #in "),%%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3786 | "movdqu 112(" #in "),%%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3787 | "pxor %%xmm0, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 3788 | "pxor %%xmm1, %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 3789 | "movdqu %%xmm10, 96(" #out ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 3790 | "movdqu %%xmm11, 112(" #out ")\n\t" |
wolfSSL | 15:117db924cf7c | 3791 | |
wolfSSL | 15:117db924cf7c | 3792 | #define _AESENC_AVX(r) \ |
wolfSSL | 15:117db924cf7c | 3793 | "aesenc 16(%[KEY]), " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3794 | "aesenc 32(%[KEY]), " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3795 | "aesenc 48(%[KEY]), " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3796 | "aesenc 64(%[KEY]), " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3797 | "aesenc 80(%[KEY]), " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3798 | "aesenc 96(%[KEY]), " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3799 | "aesenc 112(%[KEY]), " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3800 | "aesenc 128(%[KEY]), " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3801 | "aesenc 144(%[KEY]), " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3802 | "cmpl $11, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 3803 | "movdqa 160(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3804 | "jl %=f\n\t" \ |
wolfSSL | 15:117db924cf7c | 3805 | "aesenc %%xmm5, " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3806 | "aesenc 176(%[KEY]), " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3807 | "cmpl $13, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 3808 | "movdqa 192(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3809 | "jl %=f\n\t" \ |
wolfSSL | 15:117db924cf7c | 3810 | "aesenc %%xmm5, " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3811 | "aesenc 208(%[KEY]), " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3812 | "movdqa 224(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3813 | "%=:\n\t" \ |
wolfSSL | 15:117db924cf7c | 3814 | "aesenclast %%xmm5, " #r "\n\t" |
wolfSSL | 15:117db924cf7c | 3815 | #define AESENC_AVX(r) \ |
wolfSSL | 15:117db924cf7c | 3816 | _AESENC_AVX(r) |
wolfSSL | 15:117db924cf7c | 3817 | |
wolfSSL | 15:117db924cf7c | 3818 | #define AESENC_BLOCK(in, out) \ |
wolfSSL | 15:117db924cf7c | 3819 | "movdqu " VAR(CTR1) ", %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3820 | "movdqu %%xmm4, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3821 | "pshufb %[BSWAP_EPI64], %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3822 | "paddd %[ONE], %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3823 | "pxor (%[KEY]), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3824 | "movdqu %%xmm5, " VAR(CTR1) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3825 | AESENC_AVX(%%xmm4) \ |
wolfSSL | 15:117db924cf7c | 3826 | "movdqu (" #in "), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3827 | "pxor %%xmm5, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3828 | "movdqu %%xmm4, (" #out ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 3829 | "pshufb %[BSWAP_MASK], %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3830 | "pxor %%xmm4, " VAR(XR) "\n\t" |
wolfSSL | 15:117db924cf7c | 3831 | |
wolfSSL | 15:117db924cf7c | 3832 | #define _AESENC_GFMUL(in, out, H, X) \ |
wolfSSL | 15:117db924cf7c | 3833 | "movdqu " VAR(CTR1) ", %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3834 | "movdqu %%xmm4, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3835 | "pshufb %[BSWAP_EPI64], %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3836 | "paddd %[ONE], %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3837 | "pxor (%[KEY]), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3838 | "movdqu %%xmm5, " VAR(CTR1) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3839 | "movdqa " #X ", %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3840 | "pclmulqdq $0x10, " #H ", %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3841 | "aesenc 16(%[KEY]), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3842 | "aesenc 32(%[KEY]), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3843 | "movdqa " #X ", %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 3844 | "pclmulqdq $0x01, " #H ", %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 3845 | "aesenc 48(%[KEY]), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3846 | "aesenc 64(%[KEY]), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3847 | "movdqa " #X ", %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 3848 | "pclmulqdq $0x00, " #H ", %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 3849 | "aesenc 80(%[KEY]), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3850 | "movdqa " #X ", %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3851 | "pclmulqdq $0x11, " #H ", %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3852 | "aesenc 96(%[KEY]), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3853 | "pxor %%xmm7, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3854 | "movdqa %%xmm6, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3855 | "psrldq $8, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3856 | "pslldq $8, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3857 | "aesenc 112(%[KEY]), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3858 | "movdqa %%xmm1, %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 3859 | "pxor %%xmm8, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3860 | "pxor %%xmm6, %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 3861 | "movdqa %[MOD2_128], %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3862 | "movdqa %%xmm2, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 3863 | "pclmulqdq $0x10, %%xmm0, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 3864 | "aesenc 128(%[KEY]), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3865 | "pshufd $0x4e, %%xmm2, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3866 | "pxor %%xmm7, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3867 | "movdqa %%xmm6, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 3868 | "pclmulqdq $0x10, %%xmm0, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 3869 | "aesenc 144(%[KEY]), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3870 | "pshufd $0x4e, %%xmm6, " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3871 | "pxor %%xmm7, " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3872 | "pxor %%xmm3, " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3873 | "cmpl $11, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 3874 | "movdqu 160(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3875 | "jl %=f\n\t" \ |
wolfSSL | 15:117db924cf7c | 3876 | "aesenc %%xmm5, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3877 | "aesenc 176(%[KEY]), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3878 | "cmpl $13, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 3879 | "movdqu 192(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3880 | "jl %=f\n\t" \ |
wolfSSL | 15:117db924cf7c | 3881 | "aesenc %%xmm5, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3882 | "aesenc 208(%[KEY]), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3883 | "movdqa 224(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3884 | "%=:\n\t" \ |
wolfSSL | 15:117db924cf7c | 3885 | "aesenclast %%xmm5, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3886 | "movdqu (" #in "), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3887 | "pxor %%xmm5, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3888 | "movdqu %%xmm4, (" #out ")\n\t" |
wolfSSL | 15:117db924cf7c | 3889 | #define AESENC_GFMUL(in, out, H, X) \ |
wolfSSL | 15:117db924cf7c | 3890 | _AESENC_GFMUL(in, out, H, X) |
wolfSSL | 15:117db924cf7c | 3891 | |
wolfSSL | 15:117db924cf7c | 3892 | #define _GHASH_GFMUL_AVX(r, r2, a, b) \ |
wolfSSL | 15:117db924cf7c | 3893 | "pshufd $0x4e, "#a", %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3894 | "pshufd $0x4e, "#b", %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3895 | "movdqa "#b", %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 3896 | "movdqa "#b", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3897 | "pclmulqdq $0x11, "#a", %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 3898 | "pclmulqdq $0x00, "#a", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3899 | "pxor "#a", %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3900 | "pxor "#b", %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3901 | "pclmulqdq $0x00, %%xmm2, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3902 | "pxor %%xmm0, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3903 | "pxor %%xmm3, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3904 | "movdqa %%xmm1, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3905 | "movdqa %%xmm0, "#r2"\n\t" \ |
wolfSSL | 15:117db924cf7c | 3906 | "movdqa %%xmm3, " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3907 | "pslldq $8, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3908 | "psrldq $8, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3909 | "pxor %%xmm2, "#r2"\n\t" \ |
wolfSSL | 15:117db924cf7c | 3910 | "pxor %%xmm1, " #r "\n\t" |
wolfSSL | 15:117db924cf7c | 3911 | #define GHASH_GFMUL_AVX(r, r2, a, b) \ |
wolfSSL | 15:117db924cf7c | 3912 | _GHASH_GFMUL_AVX(r, r2, a, b) |
wolfSSL | 15:117db924cf7c | 3913 | |
wolfSSL | 15:117db924cf7c | 3914 | #define _GHASH_GFMUL_XOR_AVX(r, r2, a, b) \ |
wolfSSL | 15:117db924cf7c | 3915 | "pshufd $0x4e, "#a", %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3916 | "pshufd $0x4e, "#b", %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3917 | "movdqa "#b", %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 3918 | "movdqa "#b", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3919 | "pclmulqdq $0x11, "#a", %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 3920 | "pclmulqdq $0x00, "#a", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3921 | "pxor "#a", %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3922 | "pxor "#b", %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3923 | "pclmulqdq $0x00, %%xmm2, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3924 | "pxor %%xmm0, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3925 | "pxor %%xmm3, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3926 | "movdqa %%xmm1, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3927 | "pxor %%xmm0, "#r2"\n\t" \ |
wolfSSL | 15:117db924cf7c | 3928 | "pxor %%xmm3, " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3929 | "pslldq $8, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3930 | "psrldq $8, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3931 | "pxor %%xmm2, "#r2"\n\t" \ |
wolfSSL | 15:117db924cf7c | 3932 | "pxor %%xmm1, " #r "\n\t" |
wolfSSL | 15:117db924cf7c | 3933 | #define GHASH_GFMUL_XOR_AVX(r, r2, a, b) \ |
wolfSSL | 15:117db924cf7c | 3934 | _GHASH_GFMUL_XOR_AVX(r, r2, a, b) |
wolfSSL | 15:117db924cf7c | 3935 | |
wolfSSL | 15:117db924cf7c | 3936 | #define GHASH_MID_AVX(r, r2) \ |
wolfSSL | 15:117db924cf7c | 3937 | "movdqa "#r2", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3938 | "movdqa " #r ", %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3939 | "psrld $31, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3940 | "psrld $31, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3941 | "pslld $1, "#r2"\n\t" \ |
wolfSSL | 15:117db924cf7c | 3942 | "pslld $1, " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3943 | "movdqa %%xmm0, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3944 | "pslldq $4, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3945 | "psrldq $12, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3946 | "pslldq $4, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3947 | "por %%xmm2, " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3948 | "por %%xmm0, "#r2"\n\t" \ |
wolfSSL | 15:117db924cf7c | 3949 | "por %%xmm1, " #r "\n\t" |
wolfSSL | 15:117db924cf7c | 3950 | |
wolfSSL | 15:117db924cf7c | 3951 | #define _GHASH_GFMUL_RED_AVX(r, a, b) \ |
wolfSSL | 15:117db924cf7c | 3952 | "pshufd $0x4e, "#a", %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3953 | "pshufd $0x4e, "#b", %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3954 | "movdqa "#b", %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 3955 | "movdqa "#b", %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3956 | "pclmulqdq $0x11, "#a", %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 3957 | "pclmulqdq $0x00, "#a", %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3958 | "pxor "#a", %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3959 | "pxor "#b", %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3960 | "pclmulqdq $0x00, %%xmm6, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3961 | "pxor %%xmm4, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3962 | "pxor %%xmm7, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3963 | "movdqa %%xmm5, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3964 | "movdqa %%xmm7, " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3965 | "pslldq $8, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3966 | "psrldq $8, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3967 | "pxor %%xmm6, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3968 | "pxor %%xmm5, " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 3969 | "movdqa %%xmm4, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 3970 | "movdqa %%xmm4, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 3971 | "movdqa %%xmm4, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 3972 | "pslld $31, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 3973 | "pslld $30, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 3974 | "pslld $25, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 3975 | "pxor %%xmm9, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 3976 | "pxor %%xmm10, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 3977 | "movdqa %%xmm8, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 3978 | "psrldq $4, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 3979 | "pslldq $12, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 3980 | "pxor %%xmm8, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 3981 | "movdqa %%xmm4, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 3982 | "movdqa %%xmm4, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3983 | "movdqa %%xmm4, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3984 | "psrld $1, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 3985 | "psrld $2, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 3986 | "psrld $7, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 3987 | "pxor %%xmm6, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 3988 | "pxor %%xmm5, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 3989 | "pxor %%xmm9, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 3990 | "pxor %%xmm4, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 3991 | "pxor %%xmm10, " #r "\n\t" |
wolfSSL | 15:117db924cf7c | 3992 | #define GHASH_GFMUL_RED_AVX(r, a, b) \ |
wolfSSL | 15:117db924cf7c | 3993 | _GHASH_GFMUL_RED_AVX(r, a, b) |
wolfSSL | 15:117db924cf7c | 3994 | |
wolfSSL | 15:117db924cf7c | 3995 | #define GHASH_RED_AVX(r, r2) \ |
wolfSSL | 15:117db924cf7c | 3996 | "movdqa "#r2", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 3997 | "movdqa "#r2", %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 3998 | "movdqa "#r2", %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 3999 | "pslld $31, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4000 | "pslld $30, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4001 | "pslld $25, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4002 | "pxor %%xmm1, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4003 | "pxor %%xmm2, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4004 | "movdqa %%xmm0, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4005 | "psrldq $4, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4006 | "pslldq $12, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4007 | "pxor %%xmm0, "#r2"\n\t" \ |
wolfSSL | 15:117db924cf7c | 4008 | "movdqa "#r2", %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4009 | "movdqa "#r2", %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 4010 | "movdqa "#r2", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4011 | "psrld $1, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4012 | "psrld $2, %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 4013 | "psrld $7, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4014 | "pxor %%xmm3, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4015 | "pxor %%xmm0, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4016 | "pxor %%xmm1, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4017 | "pxor "#r2", %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4018 | "pxor %%xmm2, " #r "\n\t" |
wolfSSL | 15:117db924cf7c | 4019 | |
wolfSSL | 15:117db924cf7c | 4020 | #define GHASH_GFMUL_RED_XOR_AVX(r, r2, a, b) \ |
wolfSSL | 15:117db924cf7c | 4021 | GHASH_GFMUL_XOR_AVX(r, r2, a, b) \ |
wolfSSL | 15:117db924cf7c | 4022 | GHASH_RED_AVX(r, r2) |
wolfSSL | 15:117db924cf7c | 4023 | |
wolfSSL | 15:117db924cf7c | 4024 | #define GHASH_FULL_AVX(r, r2, a, b) \ |
wolfSSL | 15:117db924cf7c | 4025 | GHASH_GFMUL_AVX(r, r2, a, b) \ |
wolfSSL | 15:117db924cf7c | 4026 | GHASH_MID_AVX(r, r2) \ |
wolfSSL | 15:117db924cf7c | 4027 | GHASH_RED_AVX(r, r2) |
wolfSSL | 15:117db924cf7c | 4028 | |
wolfSSL | 15:117db924cf7c | 4029 | #define CALC_IV_12() \ |
wolfSSL | 15:117db924cf7c | 4030 | "# Calculate values when IV is 12 bytes\n\t" \ |
wolfSSL | 15:117db924cf7c | 4031 | "# Set counter based on IV\n\t" \ |
wolfSSL | 15:117db924cf7c | 4032 | "movl $0x01000000, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4033 | "pinsrq $0, 0(%%rax), %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4034 | "pinsrd $2, 8(%%rax), %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4035 | "pinsrd $3, %%ecx, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4036 | "# H = Encrypt X(=0) and T = Encrypt counter\n\t" \ |
wolfSSL | 15:117db924cf7c | 4037 | "movdqu %%xmm13, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4038 | "movdqa 0(%[KEY]), " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4039 | "pxor " VAR(HR) ", %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4040 | "movdqa 16(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4041 | "aesenc %%xmm12, " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4042 | "aesenc %%xmm12, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4043 | "movdqa 32(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4044 | "aesenc %%xmm12, " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4045 | "aesenc %%xmm12, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4046 | "movdqa 48(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4047 | "aesenc %%xmm12, " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4048 | "aesenc %%xmm12, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4049 | "movdqa 64(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4050 | "aesenc %%xmm12, " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4051 | "aesenc %%xmm12, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4052 | "movdqa 80(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4053 | "aesenc %%xmm12, " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4054 | "aesenc %%xmm12, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4055 | "movdqa 96(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4056 | "aesenc %%xmm12, " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4057 | "aesenc %%xmm12, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4058 | "movdqa 112(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4059 | "aesenc %%xmm12, " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4060 | "aesenc %%xmm12, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4061 | "movdqa 128(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4062 | "aesenc %%xmm12, " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4063 | "aesenc %%xmm12, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4064 | "movdqa 144(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4065 | "aesenc %%xmm12, " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4066 | "aesenc %%xmm12, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4067 | "cmpl $11, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 4068 | "movdqa 160(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4069 | "jl 31f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4070 | "aesenc %%xmm12, " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4071 | "aesenc %%xmm12, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4072 | "movdqa 176(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4073 | "aesenc %%xmm12, " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4074 | "aesenc %%xmm12, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4075 | "cmpl $13, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 4076 | "movdqa 192(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4077 | "jl 31f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4078 | "aesenc %%xmm12, " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4079 | "aesenc %%xmm12, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4080 | "movdqu 208(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4081 | "aesenc %%xmm12, " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4082 | "aesenc %%xmm12, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4083 | "movdqu 224(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4084 | "31:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4085 | "aesenclast %%xmm12, " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4086 | "aesenclast %%xmm12, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4087 | "pshufb %[BSWAP_MASK], " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4088 | "movdqu %%xmm1, " VAR(TR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4089 | "jmp 39f\n\t" |
wolfSSL | 15:117db924cf7c | 4090 | |
wolfSSL | 15:117db924cf7c | 4091 | #define CALC_IV() \ |
wolfSSL | 15:117db924cf7c | 4092 | "# Calculate values when IV is not 12 bytes\n\t" \ |
wolfSSL | 15:117db924cf7c | 4093 | "# H = Encrypt X(=0)\n\t" \ |
wolfSSL | 15:117db924cf7c | 4094 | "movdqa 0(%[KEY]), " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4095 | AESENC_AVX(HR) \ |
wolfSSL | 15:117db924cf7c | 4096 | "pshufb %[BSWAP_MASK], " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4097 | "# Calc counter\n\t" \ |
wolfSSL | 15:117db924cf7c | 4098 | "# Initialization vector\n\t" \ |
wolfSSL | 15:117db924cf7c | 4099 | "cmpl $0, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4100 | "movq $0, %%rcx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4101 | "je 45f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4102 | "cmpl $16, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4103 | "jl 44f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4104 | "andl $0xfffffff0, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4105 | "\n" \ |
wolfSSL | 15:117db924cf7c | 4106 | "43:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4107 | "movdqu (%%rax,%%rcx,1), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4108 | "pshufb %[BSWAP_MASK], %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4109 | "pxor %%xmm4, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4110 | GHASH_FULL_AVX(%%xmm13, %%xmm12, %%xmm13, HR) \ |
wolfSSL | 15:117db924cf7c | 4111 | "addl $16, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4112 | "cmpl %%edx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4113 | "jl 43b\n\t" \ |
wolfSSL | 15:117db924cf7c | 4114 | "movl %[ibytes], %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4115 | "cmpl %%edx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4116 | "je 45f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4117 | "\n" \ |
wolfSSL | 15:117db924cf7c | 4118 | "44:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4119 | "subq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 4120 | "pxor %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4121 | "xorl %%ebx, %%ebx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4122 | "movdqu %%xmm4, (%%rsp)\n\t" \ |
wolfSSL | 15:117db924cf7c | 4123 | "42:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4124 | "movzbl (%%rax,%%rcx,1), %%r13d\n\t" \ |
wolfSSL | 15:117db924cf7c | 4125 | "movb %%r13b, (%%rsp,%%rbx,1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 4126 | "incl %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4127 | "incl %%ebx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4128 | "cmpl %%edx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4129 | "jl 42b\n\t" \ |
wolfSSL | 15:117db924cf7c | 4130 | "movdqu (%%rsp), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4131 | "addq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 4132 | "pshufb %[BSWAP_MASK], %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4133 | "pxor %%xmm4, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4134 | GHASH_FULL_AVX(%%xmm13, %%xmm12, %%xmm13, HR) \ |
wolfSSL | 15:117db924cf7c | 4135 | "\n" \ |
wolfSSL | 15:117db924cf7c | 4136 | "45:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4137 | "# T = Encrypt counter\n\t" \ |
wolfSSL | 15:117db924cf7c | 4138 | "pxor %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4139 | "shll $3, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4140 | "pinsrq $0, %%rdx, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4141 | "pxor %%xmm0, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4142 | GHASH_FULL_AVX(%%xmm13, %%xmm12, %%xmm13, HR) \ |
wolfSSL | 15:117db924cf7c | 4143 | "pshufb %[BSWAP_MASK], %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4144 | "# Encrypt counter\n\t" \ |
wolfSSL | 15:117db924cf7c | 4145 | "movdqa 0(%[KEY]), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4146 | "pxor %%xmm13, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4147 | AESENC_AVX(%%xmm4) \ |
wolfSSL | 15:117db924cf7c | 4148 | "movdqu %%xmm4, " VAR(TR) "\n\t" |
wolfSSL | 15:117db924cf7c | 4149 | |
wolfSSL | 15:117db924cf7c | 4150 | #define CALC_AAD() \ |
wolfSSL | 15:117db924cf7c | 4151 | "# Additional authentication data\n\t" \ |
wolfSSL | 15:117db924cf7c | 4152 | "movl %[abytes], %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4153 | "cmpl $0, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4154 | "je 25f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4155 | "movq %[addt], %%rax\n\t" \ |
wolfSSL | 15:117db924cf7c | 4156 | "xorl %%ecx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4157 | "cmpl $16, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4158 | "jl 24f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4159 | "andl $0xfffffff0, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4160 | "\n" \ |
wolfSSL | 15:117db924cf7c | 4161 | "23:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4162 | "movdqu (%%rax,%%rcx,1), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4163 | "pshufb %[BSWAP_MASK], %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4164 | "pxor %%xmm4, " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4165 | GHASH_FULL_AVX(XR, %%xmm12, XR, HR) \ |
wolfSSL | 15:117db924cf7c | 4166 | "addl $16, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4167 | "cmpl %%edx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4168 | "jl 23b\n\t" \ |
wolfSSL | 15:117db924cf7c | 4169 | "movl %[abytes], %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4170 | "cmpl %%edx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4171 | "je 25f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4172 | "\n" \ |
wolfSSL | 15:117db924cf7c | 4173 | "24:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4174 | "subq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 4175 | "pxor %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4176 | "xorl %%ebx, %%ebx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4177 | "movdqu %%xmm4, (%%rsp)\n\t" \ |
wolfSSL | 15:117db924cf7c | 4178 | "22:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4179 | "movzbl (%%rax,%%rcx,1), %%r13d\n\t" \ |
wolfSSL | 15:117db924cf7c | 4180 | "movb %%r13b, (%%rsp,%%rbx,1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 4181 | "incl %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4182 | "incl %%ebx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4183 | "cmpl %%edx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4184 | "jl 22b\n\t" \ |
wolfSSL | 15:117db924cf7c | 4185 | "movdqu (%%rsp), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4186 | "addq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 4187 | "pshufb %[BSWAP_MASK], %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4188 | "pxor %%xmm4, " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4189 | GHASH_FULL_AVX(XR, %%xmm12, XR, HR) \ |
wolfSSL | 15:117db924cf7c | 4190 | "\n" \ |
wolfSSL | 15:117db924cf7c | 4191 | "25:\n\t" |
wolfSSL | 15:117db924cf7c | 4192 | |
wolfSSL | 15:117db924cf7c | 4193 | #define CALC_HT_8_AVX() \ |
wolfSSL | 15:117db924cf7c | 4194 | "movdqa " VAR(XR) ", %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4195 | "# H ^ 1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4196 | "movdqu " VAR(HR) ", 0(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 4197 | "# H ^ 2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4198 | GHASH_GFMUL_RED_AVX(%%xmm0, HR, HR) \ |
wolfSSL | 15:117db924cf7c | 4199 | "movdqu %%xmm0 , 16(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 4200 | "# H ^ 3\n\t" \ |
wolfSSL | 15:117db924cf7c | 4201 | GHASH_GFMUL_RED_AVX(%%xmm1, HR, %%xmm0) \ |
wolfSSL | 15:117db924cf7c | 4202 | "movdqu %%xmm1 , 32(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 4203 | "# H ^ 4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4204 | GHASH_GFMUL_RED_AVX(%%xmm3, %%xmm0, %%xmm0) \ |
wolfSSL | 15:117db924cf7c | 4205 | "movdqu %%xmm3 , 48(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 4206 | "# H ^ 5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4207 | GHASH_GFMUL_RED_AVX(%%xmm12, %%xmm0, %%xmm1) \ |
wolfSSL | 15:117db924cf7c | 4208 | "movdqu %%xmm12, 64(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 4209 | "# H ^ 6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4210 | GHASH_GFMUL_RED_AVX(%%xmm12, %%xmm1, %%xmm1) \ |
wolfSSL | 15:117db924cf7c | 4211 | "movdqu %%xmm12, 80(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 4212 | "# H ^ 7\n\t" \ |
wolfSSL | 15:117db924cf7c | 4213 | GHASH_GFMUL_RED_AVX(%%xmm12, %%xmm1, %%xmm3) \ |
wolfSSL | 15:117db924cf7c | 4214 | "movdqu %%xmm12, 96(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 4215 | "# H ^ 8\n\t" \ |
wolfSSL | 15:117db924cf7c | 4216 | GHASH_GFMUL_RED_AVX(%%xmm12, %%xmm3, %%xmm3) \ |
wolfSSL | 15:117db924cf7c | 4217 | "movdqu %%xmm12, 112(" VAR(HTR) ")\n\t" |
wolfSSL | 15:117db924cf7c | 4218 | |
wolfSSL | 15:117db924cf7c | 4219 | #define AESENC_128_GHASH_AVX(src, o) \ |
wolfSSL | 15:117db924cf7c | 4220 | "leaq (%[in]," VAR(KR64) ",1), %%rcx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4221 | "leaq (%[out]," VAR(KR64) ",1), %%rdx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4222 | /* src is either %%rcx or %%rdx */ \ |
wolfSSL | 15:117db924cf7c | 4223 | AESENC_CTR() \ |
wolfSSL | 15:117db924cf7c | 4224 | AESENC_XOR() \ |
wolfSSL | 15:117db924cf7c | 4225 | AESENC_PCLMUL_1(src, 16, o-128, 112) \ |
wolfSSL | 15:117db924cf7c | 4226 | AESENC_PCLMUL_N(src, 32, o-112, 96) \ |
wolfSSL | 15:117db924cf7c | 4227 | AESENC_PCLMUL_N(src, 48, o -96, 80) \ |
wolfSSL | 15:117db924cf7c | 4228 | AESENC_PCLMUL_N(src, 64, o -80, 64) \ |
wolfSSL | 15:117db924cf7c | 4229 | AESENC_PCLMUL_N(src, 80, o -64, 48) \ |
wolfSSL | 15:117db924cf7c | 4230 | AESENC_PCLMUL_N(src, 96, o -48, 32) \ |
wolfSSL | 15:117db924cf7c | 4231 | AESENC_PCLMUL_N(src, 112, o -32, 16) \ |
wolfSSL | 15:117db924cf7c | 4232 | AESENC_PCLMUL_N(src, 128, o -16, 0) \ |
wolfSSL | 15:117db924cf7c | 4233 | AESENC_PCLMUL_L(144) \ |
wolfSSL | 15:117db924cf7c | 4234 | "cmpl $11, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 4235 | "movdqa 160(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4236 | "jl 4f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4237 | AESENC() \ |
wolfSSL | 15:117db924cf7c | 4238 | AESENC_SET(176) \ |
wolfSSL | 15:117db924cf7c | 4239 | "cmpl $13, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 4240 | "movdqa 192(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4241 | "jl 4f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4242 | AESENC() \ |
wolfSSL | 15:117db924cf7c | 4243 | AESENC_SET(208) \ |
wolfSSL | 15:117db924cf7c | 4244 | "movdqa 224(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4245 | "\n" \ |
wolfSSL | 15:117db924cf7c | 4246 | "4:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4247 | AESENC_LAST(%%rcx, %%rdx) |
wolfSSL | 15:117db924cf7c | 4248 | |
wolfSSL | 15:117db924cf7c | 4249 | #define AESENC_LAST15_ENC_AVX() \ |
wolfSSL | 15:117db924cf7c | 4250 | "movl %[nbytes], %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4251 | "movl %%ecx, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4252 | "andl $0x0f, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4253 | "jz 55f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4254 | "movdqu " VAR(CTR1) ", %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4255 | "pshufb %[BSWAP_EPI64], %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4256 | "pxor 0(%[KEY]), %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4257 | AESENC_AVX(%%xmm13) \ |
wolfSSL | 15:117db924cf7c | 4258 | "subq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 4259 | "xorl %%ecx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4260 | "movdqu %%xmm13, (%%rsp)\n\t" \ |
wolfSSL | 15:117db924cf7c | 4261 | "\n" \ |
wolfSSL | 15:117db924cf7c | 4262 | "51:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4263 | "movzbl (%[in]," VAR(KR64) ",1), %%r13d\n\t" \ |
wolfSSL | 15:117db924cf7c | 4264 | "xorb (%%rsp,%%rcx,1), %%r13b\n\t" \ |
wolfSSL | 15:117db924cf7c | 4265 | "movb %%r13b, (%[out]," VAR(KR64) ",1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 4266 | "movb %%r13b, (%%rsp,%%rcx,1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 4267 | "incl " VAR(KR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4268 | "incl %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4269 | "cmpl %%edx, " VAR(KR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4270 | "jl 51b\n\t" \ |
wolfSSL | 15:117db924cf7c | 4271 | "xorq %%r13, %%r13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4272 | "cmpl $16, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4273 | "je 53f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4274 | "\n" \ |
wolfSSL | 15:117db924cf7c | 4275 | "52:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4276 | "movb %%r13b, (%%rsp,%%rcx,1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 4277 | "incl %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4278 | "cmpl $16, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4279 | "jl 52b\n\t" \ |
wolfSSL | 15:117db924cf7c | 4280 | "53:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4281 | "movdqu (%%rsp), %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4282 | "addq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 4283 | "pshufb %[BSWAP_MASK], %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4284 | "pxor %%xmm13, " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4285 | GHASH_GFMUL_RED_AVX(XR, HR, XR) \ |
wolfSSL | 15:117db924cf7c | 4286 | |
wolfSSL | 15:117db924cf7c | 4287 | #define AESENC_LAST15_DEC_AVX() \ |
wolfSSL | 15:117db924cf7c | 4288 | "movl %[nbytes], %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4289 | "movl %%ecx, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4290 | "andl $0x0f, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4291 | "jz 55f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4292 | "movdqu " VAR(CTR1) ", %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4293 | "pshufb %[BSWAP_EPI64], %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4294 | "pxor 0(%[KEY]), %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4295 | AESENC_AVX(%%xmm13) \ |
wolfSSL | 15:117db924cf7c | 4296 | "subq $32, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 4297 | "xorl %%ecx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4298 | "movdqu %%xmm13, (%%rsp)\n\t" \ |
wolfSSL | 15:117db924cf7c | 4299 | "pxor %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4300 | "movdqu %%xmm0, 16(%%rsp)\n\t" \ |
wolfSSL | 15:117db924cf7c | 4301 | "\n" \ |
wolfSSL | 15:117db924cf7c | 4302 | "51:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4303 | "movzbl (%[in]," VAR(KR64) ",1), %%r13d\n\t" \ |
wolfSSL | 15:117db924cf7c | 4304 | "movb %%r13b, 16(%%rsp,%%rcx,1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 4305 | "xorb (%%rsp,%%rcx,1), %%r13b\n\t" \ |
wolfSSL | 15:117db924cf7c | 4306 | "movb %%r13b, (%[out]," VAR(KR64) ",1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 4307 | "incl " VAR(KR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4308 | "incl %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4309 | "cmpl %%edx, " VAR(KR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4310 | "jl 51b\n\t" \ |
wolfSSL | 15:117db924cf7c | 4311 | "53:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4312 | "movdqu 16(%%rsp), %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4313 | "addq $32, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 4314 | "pshufb %[BSWAP_MASK], %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4315 | "pxor %%xmm13, " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4316 | GHASH_GFMUL_RED_AVX(XR, HR, XR) \ |
wolfSSL | 15:117db924cf7c | 4317 | |
wolfSSL | 15:117db924cf7c | 4318 | #define CALC_TAG() \ |
wolfSSL | 15:117db924cf7c | 4319 | "movl %[nbytes], %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4320 | "movl %[abytes], %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4321 | "shlq $3, %%rdx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4322 | "shlq $3, %%rcx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4323 | "pinsrq $0, %%rdx, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4324 | "pinsrq $1, %%rcx, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4325 | "pxor %%xmm0, " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4326 | GHASH_GFMUL_RED_AVX(XR, HR, XR) \ |
wolfSSL | 15:117db924cf7c | 4327 | "pshufb %[BSWAP_MASK], " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4328 | "movdqu " VAR(TR) ", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4329 | "pxor " VAR(XR) ", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4330 | |
wolfSSL | 15:117db924cf7c | 4331 | #define STORE_TAG() \ |
wolfSSL | 15:117db924cf7c | 4332 | "cmpl $16, %[tbytes]\n\t" \ |
wolfSSL | 15:117db924cf7c | 4333 | "je 71f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4334 | "xorq %%rcx, %%rcx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4335 | "movdqu %%xmm0, (%%rsp)\n\t" \ |
wolfSSL | 15:117db924cf7c | 4336 | "73:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4337 | "movzbl (%%rsp,%%rcx,1), %%r13d\n\t" \ |
wolfSSL | 15:117db924cf7c | 4338 | "movb %%r13b, (%[tag],%%rcx,1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 4339 | "incl %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4340 | "cmpl %[tbytes], %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4341 | "jne 73b\n\t" \ |
wolfSSL | 15:117db924cf7c | 4342 | "jmp 72f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4343 | "\n" \ |
wolfSSL | 15:117db924cf7c | 4344 | "71:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4345 | "movdqu %%xmm0, (%[tag])\n\t" \ |
wolfSSL | 15:117db924cf7c | 4346 | "\n" \ |
wolfSSL | 15:117db924cf7c | 4347 | "72:\n\t" |
wolfSSL | 15:117db924cf7c | 4348 | |
wolfSSL | 15:117db924cf7c | 4349 | #define CMP_TAG() \ |
wolfSSL | 15:117db924cf7c | 4350 | "cmpl $16, %[tbytes]\n\t" \ |
wolfSSL | 15:117db924cf7c | 4351 | "je 71f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4352 | "subq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 4353 | "xorq %%rcx, %%rcx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4354 | "xorq %%rax, %%rax\n\t" \ |
wolfSSL | 15:117db924cf7c | 4355 | "movdqu %%xmm0, (%%rsp)\n\t" \ |
wolfSSL | 15:117db924cf7c | 4356 | "\n" \ |
wolfSSL | 15:117db924cf7c | 4357 | "73:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4358 | "movzbl (%%rsp,%%rcx,1), %%r13d\n\t" \ |
wolfSSL | 15:117db924cf7c | 4359 | "xorb (%[tag],%%rcx,1), %%r13b\n\t" \ |
wolfSSL | 15:117db924cf7c | 4360 | "orb %%r13b, %%al\n\t" \ |
wolfSSL | 15:117db924cf7c | 4361 | "incl %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4362 | "cmpl %[tbytes], %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4363 | "jne 73b\n\t" \ |
wolfSSL | 15:117db924cf7c | 4364 | "cmpb $0x00, %%al\n\t" \ |
wolfSSL | 15:117db924cf7c | 4365 | "sete %%al\n\t" \ |
wolfSSL | 15:117db924cf7c | 4366 | "addq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 4367 | "xorq %%rcx, %%rcx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4368 | "jmp 72f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4369 | "\n" \ |
wolfSSL | 15:117db924cf7c | 4370 | "71:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4371 | "movdqu (%[tag]), %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4372 | "pcmpeqb %%xmm1, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4373 | "pmovmskb %%xmm0, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4374 | "# %%edx == 0xFFFF then return 1 else => return 0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4375 | "xorl %%eax, %%eax\n\t" \ |
wolfSSL | 15:117db924cf7c | 4376 | "cmpl $0xffff, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4377 | "sete %%al\n\t" \ |
wolfSSL | 15:117db924cf7c | 4378 | "\n" \ |
wolfSSL | 15:117db924cf7c | 4379 | "72:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4380 | "movl %%eax, (%[res])\n\t" |
wolfSSL | 15:117db924cf7c | 4381 | |
wolfSSL | 15:117db924cf7c | 4382 | static void AES_GCM_encrypt(const unsigned char *in, unsigned char *out, |
wolfSSL | 15:117db924cf7c | 4383 | const unsigned char* addt, |
wolfSSL | 15:117db924cf7c | 4384 | const unsigned char* ivec, unsigned char *tag, |
wolfSSL | 15:117db924cf7c | 4385 | unsigned int nbytes, unsigned int abytes, |
wolfSSL | 15:117db924cf7c | 4386 | unsigned int ibytes, unsigned int tbytes, |
wolfSSL | 15:117db924cf7c | 4387 | const unsigned char* key, int nr) |
wolfSSL | 15:117db924cf7c | 4388 | { |
wolfSSL | 15:117db924cf7c | 4389 | register const unsigned char* iv asm("rax") = ivec; |
wolfSSL | 15:117db924cf7c | 4390 | register unsigned int ivLen asm("ebx") = ibytes; |
wolfSSL | 15:117db924cf7c | 4391 | |
wolfSSL | 15:117db924cf7c | 4392 | __asm__ __volatile__ ( |
wolfSSL | 15:117db924cf7c | 4393 | "subq $" VAR(STACK_OFFSET) ", %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 4394 | /* Counter is xmm13 */ |
wolfSSL | 15:117db924cf7c | 4395 | "pxor %%xmm13, %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 4396 | "pxor " VAR(XR) ", " VAR(XR) "\n\t" |
wolfSSL | 15:117db924cf7c | 4397 | "movl %[ibytes], %%edx\n\t" |
wolfSSL | 15:117db924cf7c | 4398 | "cmpl $12, %%edx\n\t" |
wolfSSL | 15:117db924cf7c | 4399 | "jne 35f\n\t" |
wolfSSL | 15:117db924cf7c | 4400 | CALC_IV_12() |
wolfSSL | 15:117db924cf7c | 4401 | "\n" |
wolfSSL | 15:117db924cf7c | 4402 | "35:\n\t" |
wolfSSL | 15:117db924cf7c | 4403 | CALC_IV() |
wolfSSL | 15:117db924cf7c | 4404 | "\n" |
wolfSSL | 15:117db924cf7c | 4405 | "39:\n\t" |
wolfSSL | 15:117db924cf7c | 4406 | |
wolfSSL | 15:117db924cf7c | 4407 | CALC_AAD() |
wolfSSL | 15:117db924cf7c | 4408 | |
wolfSSL | 15:117db924cf7c | 4409 | "# Calculate counter and H\n\t" |
wolfSSL | 15:117db924cf7c | 4410 | "pshufb %[BSWAP_EPI64], %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 4411 | "movdqa " VAR(HR) ", %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 4412 | "paddd %[ONE], %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 4413 | "movdqa " VAR(HR) ", %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 4414 | "movdqu %%xmm13, " VAR(CTR1) "\n\t" |
wolfSSL | 15:117db924cf7c | 4415 | "psrlq $63, %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 4416 | "psllq $1, %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 4417 | "pslldq $8, %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 4418 | "por %%xmm5, %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 4419 | "pshufd $0xff, " VAR(HR) ", " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 4420 | "psrad $31, " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 4421 | "pand %[MOD2_128], " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 4422 | "pxor %%xmm4, " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 4423 | |
wolfSSL | 15:117db924cf7c | 4424 | "xorl " VAR(KR) ", " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 4425 | |
wolfSSL | 15:117db924cf7c | 4426 | #if !defined(AES_GCM_AESNI_NO_UNROLL) && !defined(AES_GCM_AVX1_NO_UNROLL) |
wolfSSL | 15:117db924cf7c | 4427 | "cmpl $128, %[nbytes]\n\t" |
wolfSSL | 15:117db924cf7c | 4428 | "movl %[nbytes], %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 4429 | "jl 5f\n\t" |
wolfSSL | 15:117db924cf7c | 4430 | "andl $0xffffff80, %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 4431 | |
wolfSSL | 15:117db924cf7c | 4432 | CALC_HT_8_AVX() |
wolfSSL | 15:117db924cf7c | 4433 | |
wolfSSL | 15:117db924cf7c | 4434 | "# First 128 bytes of input\n\t" |
wolfSSL | 15:117db924cf7c | 4435 | AESENC_CTR() |
wolfSSL | 15:117db924cf7c | 4436 | AESENC_XOR() |
wolfSSL | 15:117db924cf7c | 4437 | AESENC_SET(16) |
wolfSSL | 15:117db924cf7c | 4438 | AESENC_SET(32) |
wolfSSL | 15:117db924cf7c | 4439 | AESENC_SET(48) |
wolfSSL | 15:117db924cf7c | 4440 | AESENC_SET(64) |
wolfSSL | 15:117db924cf7c | 4441 | AESENC_SET(80) |
wolfSSL | 15:117db924cf7c | 4442 | AESENC_SET(96) |
wolfSSL | 15:117db924cf7c | 4443 | AESENC_SET(112) |
wolfSSL | 15:117db924cf7c | 4444 | AESENC_SET(128) |
wolfSSL | 15:117db924cf7c | 4445 | AESENC_SET(144) |
wolfSSL | 15:117db924cf7c | 4446 | "cmpl $11, %[nr]\n\t" |
wolfSSL | 15:117db924cf7c | 4447 | "movdqa 160(%[KEY]), %%xmm12\n\t" |
wolfSSL | 15:117db924cf7c | 4448 | "jl 1f\n\t" |
wolfSSL | 15:117db924cf7c | 4449 | AESENC() |
wolfSSL | 15:117db924cf7c | 4450 | AESENC_SET(176) |
wolfSSL | 15:117db924cf7c | 4451 | "cmpl $13, %[nr]\n\t" |
wolfSSL | 15:117db924cf7c | 4452 | "movdqa 192(%[KEY]), %%xmm12\n\t" |
wolfSSL | 15:117db924cf7c | 4453 | "jl 1f\n\t" |
wolfSSL | 15:117db924cf7c | 4454 | AESENC() |
wolfSSL | 15:117db924cf7c | 4455 | AESENC_SET(208) |
wolfSSL | 15:117db924cf7c | 4456 | "movdqa 224(%[KEY]), %%xmm12\n\t" |
wolfSSL | 15:117db924cf7c | 4457 | "\n" |
wolfSSL | 15:117db924cf7c | 4458 | "1:\n\t" |
wolfSSL | 15:117db924cf7c | 4459 | AESENC_LAST(%[in], %[out]) |
wolfSSL | 15:117db924cf7c | 4460 | |
wolfSSL | 15:117db924cf7c | 4461 | "cmpl $128, %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 4462 | "movl $128, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 4463 | "jle 2f\n\t" |
wolfSSL | 15:117db924cf7c | 4464 | |
wolfSSL | 15:117db924cf7c | 4465 | "# More 128 bytes of input\n\t" |
wolfSSL | 15:117db924cf7c | 4466 | "\n" |
wolfSSL | 15:117db924cf7c | 4467 | "3:\n\t" |
wolfSSL | 15:117db924cf7c | 4468 | AESENC_128_GHASH_AVX(%%rdx, 0) |
wolfSSL | 15:117db924cf7c | 4469 | "addl $128, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 4470 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 4471 | "jl 3b\n\t" |
wolfSSL | 15:117db924cf7c | 4472 | "\n" |
wolfSSL | 15:117db924cf7c | 4473 | "2:\n\t" |
wolfSSL | 15:117db924cf7c | 4474 | "movdqa %[BSWAP_MASK], %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 4475 | "pshufb %%xmm13, %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 4476 | "pshufb %%xmm13, %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 4477 | "pshufb %%xmm13, %%xmm6\n\t" |
wolfSSL | 15:117db924cf7c | 4478 | "pshufb %%xmm13, %%xmm7\n\t" |
wolfSSL | 15:117db924cf7c | 4479 | "pxor %%xmm2, %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 4480 | "pshufb %%xmm13, %%xmm8\n\t" |
wolfSSL | 15:117db924cf7c | 4481 | "pshufb %%xmm13, %%xmm9\n\t" |
wolfSSL | 15:117db924cf7c | 4482 | "pshufb %%xmm13, %%xmm10\n\t" |
wolfSSL | 15:117db924cf7c | 4483 | "pshufb %%xmm13, %%xmm11\n\t" |
wolfSSL | 15:117db924cf7c | 4484 | |
wolfSSL | 15:117db924cf7c | 4485 | "movdqu 112(" VAR(HTR) "), %%xmm12\n\t" |
wolfSSL | 15:117db924cf7c | 4486 | GHASH_GFMUL_AVX(XR, %%xmm13, %%xmm4, %%xmm12) |
wolfSSL | 15:117db924cf7c | 4487 | "movdqu 96(" VAR(HTR) "), %%xmm12\n\t" |
wolfSSL | 15:117db924cf7c | 4488 | GHASH_GFMUL_XOR_AVX(XR, %%xmm13, %%xmm5, %%xmm12) |
wolfSSL | 15:117db924cf7c | 4489 | "movdqu 80(" VAR(HTR) "), %%xmm12\n\t" |
wolfSSL | 15:117db924cf7c | 4490 | GHASH_GFMUL_XOR_AVX(XR, %%xmm13, %%xmm6, %%xmm12) |
wolfSSL | 15:117db924cf7c | 4491 | "movdqu 64(" VAR(HTR) "), %%xmm12\n\t" |
wolfSSL | 15:117db924cf7c | 4492 | GHASH_GFMUL_XOR_AVX(XR, %%xmm13, %%xmm7, %%xmm12) |
wolfSSL | 15:117db924cf7c | 4493 | "movdqu 48(" VAR(HTR) "), %%xmm12\n\t" |
wolfSSL | 15:117db924cf7c | 4494 | GHASH_GFMUL_XOR_AVX(XR, %%xmm13, %%xmm8, %%xmm12) |
wolfSSL | 15:117db924cf7c | 4495 | "movdqu 32(" VAR(HTR) "), %%xmm12\n\t" |
wolfSSL | 15:117db924cf7c | 4496 | GHASH_GFMUL_XOR_AVX(XR, %%xmm13, %%xmm9, %%xmm12) |
wolfSSL | 15:117db924cf7c | 4497 | "movdqu 16(" VAR(HTR) "), %%xmm12\n\t" |
wolfSSL | 15:117db924cf7c | 4498 | GHASH_GFMUL_XOR_AVX(XR, %%xmm13, %%xmm10, %%xmm12) |
wolfSSL | 15:117db924cf7c | 4499 | "movdqu (" VAR(HTR) "), %%xmm12\n\t" |
wolfSSL | 15:117db924cf7c | 4500 | GHASH_GFMUL_RED_XOR_AVX(XR, %%xmm13, %%xmm11, %%xmm12) |
wolfSSL | 15:117db924cf7c | 4501 | |
wolfSSL | 15:117db924cf7c | 4502 | "movdqu 0(" VAR(HTR) "), " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 4503 | "\n" |
wolfSSL | 15:117db924cf7c | 4504 | "5:\n\t" |
wolfSSL | 15:117db924cf7c | 4505 | "movl %[nbytes], %%edx\n\t" |
wolfSSL | 15:117db924cf7c | 4506 | "cmpl %%edx, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 4507 | "jge 55f\n\t" |
wolfSSL | 15:117db924cf7c | 4508 | #endif |
wolfSSL | 15:117db924cf7c | 4509 | |
wolfSSL | 15:117db924cf7c | 4510 | "movl %[nbytes], %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 4511 | "andl $0xfffffff0, %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 4512 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 4513 | "jge 14f\n\t" |
wolfSSL | 15:117db924cf7c | 4514 | |
wolfSSL | 15:117db924cf7c | 4515 | "leaq (%[in]," VAR(KR64) ",1), %%rcx\n\t" |
wolfSSL | 15:117db924cf7c | 4516 | "leaq (%[out]," VAR(KR64) ",1), %%rdx\n\t" |
wolfSSL | 15:117db924cf7c | 4517 | AESENC_BLOCK(%%rcx, %%rdx) |
wolfSSL | 15:117db924cf7c | 4518 | "addl $16, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 4519 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 4520 | "jge 13f\n\t" |
wolfSSL | 15:117db924cf7c | 4521 | "\n" |
wolfSSL | 15:117db924cf7c | 4522 | "12:\n\t" |
wolfSSL | 15:117db924cf7c | 4523 | "leaq (%[in]," VAR(KR64) ",1), %%rcx\n\t" |
wolfSSL | 15:117db924cf7c | 4524 | "leaq (%[out]," VAR(KR64) ",1), %%rdx\n\t" |
wolfSSL | 15:117db924cf7c | 4525 | AESENC_GFMUL(%%rcx, %%rdx, HR, XR) |
wolfSSL | 15:117db924cf7c | 4526 | "pshufb %[BSWAP_MASK], %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 4527 | "pxor %%xmm4, " VAR(XR) "\n\t" |
wolfSSL | 15:117db924cf7c | 4528 | "addl $16, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 4529 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 4530 | "jl 12b\n\t" |
wolfSSL | 15:117db924cf7c | 4531 | "\n" |
wolfSSL | 15:117db924cf7c | 4532 | "13:\n\t" |
wolfSSL | 15:117db924cf7c | 4533 | GHASH_GFMUL_RED_AVX(XR, HR, XR) |
wolfSSL | 15:117db924cf7c | 4534 | "\n" |
wolfSSL | 15:117db924cf7c | 4535 | "14:\n\t" |
wolfSSL | 15:117db924cf7c | 4536 | |
wolfSSL | 15:117db924cf7c | 4537 | AESENC_LAST15_ENC_AVX() |
wolfSSL | 15:117db924cf7c | 4538 | "\n" |
wolfSSL | 15:117db924cf7c | 4539 | "55:\n\t" |
wolfSSL | 15:117db924cf7c | 4540 | |
wolfSSL | 15:117db924cf7c | 4541 | CALC_TAG() |
wolfSSL | 15:117db924cf7c | 4542 | STORE_TAG() |
wolfSSL | 15:117db924cf7c | 4543 | "addq $" VAR(STACK_OFFSET) ", %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 4544 | |
wolfSSL | 15:117db924cf7c | 4545 | : |
wolfSSL | 15:117db924cf7c | 4546 | : [KEY] "r" (key), |
wolfSSL | 15:117db924cf7c | 4547 | [in] "r" (in), [out] "r" (out), [nr] "r" (nr), |
wolfSSL | 15:117db924cf7c | 4548 | [nbytes] "r" (nbytes), [abytes] "r" (abytes), [addt] "r" (addt), |
wolfSSL | 15:117db924cf7c | 4549 | [ivec] "r" (iv), [ibytes] "r" (ivLen), [tbytes] "r" (tbytes), |
wolfSSL | 15:117db924cf7c | 4550 | [tag] "r" (tag), |
wolfSSL | 15:117db924cf7c | 4551 | [BSWAP_MASK] "m" (BSWAP_MASK), |
wolfSSL | 15:117db924cf7c | 4552 | [BSWAP_EPI64] "m" (BSWAP_EPI64), |
wolfSSL | 15:117db924cf7c | 4553 | [ONE] "m" (ONE), |
wolfSSL | 15:117db924cf7c | 4554 | #if !defined(AES_GCM_AESNI_NO_UNROLL) && !defined(AES_GCM_AVX1_NO_UNROLL) |
wolfSSL | 15:117db924cf7c | 4555 | [TWO] "m" (TWO), [THREE] "m" (THREE), [FOUR] "m" (FOUR), |
wolfSSL | 15:117db924cf7c | 4556 | [FIVE] "m" (FIVE), [SIX] "m" (SIX), [SEVEN] "m" (SEVEN), |
wolfSSL | 15:117db924cf7c | 4557 | [EIGHT] "m" (EIGHT), |
wolfSSL | 15:117db924cf7c | 4558 | #endif |
wolfSSL | 15:117db924cf7c | 4559 | [MOD2_128] "m" (MOD2_128) |
wolfSSL | 15:117db924cf7c | 4560 | : "xmm15", "xmm14", "xmm13", "xmm12", |
wolfSSL | 15:117db924cf7c | 4561 | "xmm0", "xmm1", "xmm2", "xmm3", "memory", |
wolfSSL | 15:117db924cf7c | 4562 | "xmm4", "xmm5", "xmm6", "xmm7", "xmm8", "xmm9", "xmm10", "xmm11", |
wolfSSL | 15:117db924cf7c | 4563 | "rcx", "rdx", "r13" |
wolfSSL | 15:117db924cf7c | 4564 | ); |
wolfSSL | 15:117db924cf7c | 4565 | } |
wolfSSL | 15:117db924cf7c | 4566 | |
wolfSSL | 15:117db924cf7c | 4567 | #ifdef HAVE_INTEL_AVX1 |
wolfSSL | 15:117db924cf7c | 4568 | /* Encrypt with key in xmm12. */ |
wolfSSL | 15:117db924cf7c | 4569 | #define VAESENC() \ |
wolfSSL | 15:117db924cf7c | 4570 | "vaesenc %%xmm12, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4571 | "vaesenc %%xmm12, %%xmm5, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4572 | "vaesenc %%xmm12, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4573 | "vaesenc %%xmm12, %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 4574 | "vaesenc %%xmm12, %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 4575 | "vaesenc %%xmm12, %%xmm9, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 4576 | "vaesenc %%xmm12, %%xmm10, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4577 | "vaesenc %%xmm12, %%xmm11, %%xmm11\n\t" |
wolfSSL | 15:117db924cf7c | 4578 | |
wolfSSL | 15:117db924cf7c | 4579 | #define VAESENC_SET(o) \ |
wolfSSL | 15:117db924cf7c | 4580 | "vmovdqa "#o"(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4581 | VAESENC() |
wolfSSL | 15:117db924cf7c | 4582 | |
wolfSSL | 15:117db924cf7c | 4583 | #define VAESENC_CTR() \ |
wolfSSL | 15:117db924cf7c | 4584 | "vmovdqu " VAR(CTR1) ", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4585 | "vmovdqa %[BSWAP_EPI64], %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4586 | "vpshufb %%xmm1, %%xmm0, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4587 | "vpaddd %[ONE], %%xmm0, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4588 | "vpshufb %%xmm1, %%xmm5, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4589 | "vpaddd %[TWO], %%xmm0, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4590 | "vpshufb %%xmm1, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4591 | "vpaddd %[THREE], %%xmm0, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 4592 | "vpshufb %%xmm1, %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 4593 | "vpaddd %[FOUR], %%xmm0, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 4594 | "vpshufb %%xmm1, %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 4595 | "vpaddd %[FIVE], %%xmm0, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 4596 | "vpshufb %%xmm1, %%xmm9, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 4597 | "vpaddd %[SIX], %%xmm0, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4598 | "vpshufb %%xmm1, %%xmm10, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4599 | "vpaddd %[SEVEN], %%xmm0, %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 4600 | "vpshufb %%xmm1, %%xmm11, %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 4601 | "vpaddd %[EIGHT], %%xmm0, %%xmm0\n\t" |
wolfSSL | 15:117db924cf7c | 4602 | |
wolfSSL | 15:117db924cf7c | 4603 | #define VAESENC_XOR() \ |
wolfSSL | 15:117db924cf7c | 4604 | "vmovdqa (%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4605 | "vmovdqu %%xmm0, " VAR(CTR1) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4606 | "vpxor %%xmm12, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4607 | "vpxor %%xmm12, %%xmm5, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4608 | "vpxor %%xmm12, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4609 | "vpxor %%xmm12, %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 4610 | "vpxor %%xmm12, %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 4611 | "vpxor %%xmm12, %%xmm9, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 4612 | "vpxor %%xmm12, %%xmm10, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4613 | "vpxor %%xmm12, %%xmm11, %%xmm11\n\t" |
wolfSSL | 15:117db924cf7c | 4614 | |
wolfSSL | 15:117db924cf7c | 4615 | #define VAESENC_128() \ |
wolfSSL | 15:117db924cf7c | 4616 | VAESENC_CTR() \ |
wolfSSL | 15:117db924cf7c | 4617 | VAESENC_XOR() \ |
wolfSSL | 15:117db924cf7c | 4618 | VAESENC_SET(16) \ |
wolfSSL | 15:117db924cf7c | 4619 | VAESENC_SET(32) \ |
wolfSSL | 15:117db924cf7c | 4620 | VAESENC_SET(48) \ |
wolfSSL | 15:117db924cf7c | 4621 | VAESENC_SET(64) \ |
wolfSSL | 15:117db924cf7c | 4622 | VAESENC_SET(80) \ |
wolfSSL | 15:117db924cf7c | 4623 | VAESENC_SET(96) \ |
wolfSSL | 15:117db924cf7c | 4624 | VAESENC_SET(112) \ |
wolfSSL | 15:117db924cf7c | 4625 | VAESENC_SET(128) \ |
wolfSSL | 15:117db924cf7c | 4626 | VAESENC_SET(144) \ |
wolfSSL | 15:117db924cf7c | 4627 | "cmpl $11, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 4628 | "vmovdqa 160(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4629 | "jl 1f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4630 | VAESENC() \ |
wolfSSL | 15:117db924cf7c | 4631 | VAESENC_SET(176) \ |
wolfSSL | 15:117db924cf7c | 4632 | "cmpl $13, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 4633 | "vmovdqa 192(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4634 | "jl 1f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4635 | VAESENC() \ |
wolfSSL | 15:117db924cf7c | 4636 | VAESENC_SET(208) \ |
wolfSSL | 15:117db924cf7c | 4637 | "vmovdqa 224(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4638 | "\n" \ |
wolfSSL | 15:117db924cf7c | 4639 | "1:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4640 | VAESENC_LAST(%[in], %[out]) |
wolfSSL | 15:117db924cf7c | 4641 | |
wolfSSL | 15:117db924cf7c | 4642 | /* Encrypt and carry-less multiply for AVX1. */ |
wolfSSL | 15:117db924cf7c | 4643 | #define VAESENC_PCLMUL_1(src, o1, o2, o3) \ |
wolfSSL | 15:117db924cf7c | 4644 | "vmovdqu " #o3 "(" VAR(HTR) "), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4645 | "vmovdqu " #o2 "(" #src "), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4646 | "vaesenc " #o1 "(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4647 | "vpshufb %[BSWAP_MASK], %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4648 | "vpxor %%xmm2, %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4649 | "vpshufd $0x4e, %%xmm12, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4650 | "vpshufd $0x4e, %%xmm0, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 4651 | "vpxor %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4652 | "vpxor %%xmm0, %%xmm14, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 4653 | "vpclmulqdq $0x11, %%xmm12, %%xmm0, %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 4654 | "vaesenc " #o1 "(%[KEY]), %%xmm5, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4655 | "vaesenc " #o1 "(%[KEY]), %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4656 | "vpclmulqdq $0x00, %%xmm12, %%xmm0, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4657 | "vaesenc " #o1 "(%[KEY]), %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 4658 | "vaesenc " #o1 "(%[KEY]), %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 4659 | "vpclmulqdq $0x00, %%xmm14, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4660 | "vaesenc " #o1 "(%[KEY]), %%xmm9, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 4661 | "vaesenc " #o1 "(%[KEY]), %%xmm10, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4662 | "vaesenc " #o1 "(%[KEY]), %%xmm11, %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 4663 | "vpxor %%xmm2, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4664 | "vpxor %%xmm3, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4665 | |
wolfSSL | 15:117db924cf7c | 4666 | #define VAESENC_PCLMUL_N(src, o1, o2, o3) \ |
wolfSSL | 15:117db924cf7c | 4667 | "vmovdqu " #o3 "(" VAR(HTR) "), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4668 | "vmovdqu " #o2 "(" #src "), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4669 | "vpshufd $0x4e, %%xmm12, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4670 | "vpshufb %[BSWAP_MASK], %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4671 | "vaesenc " #o1 "(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4672 | "vpxor %%xmm12, %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4673 | "vpshufd $0x4e, %%xmm0, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 4674 | "vpxor %%xmm0, %%xmm14, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 4675 | "vpclmulqdq $0x11, %%xmm12, %%xmm0, %%xmm15\n\t" \ |
wolfSSL | 15:117db924cf7c | 4676 | "vaesenc " #o1 "(%[KEY]), %%xmm5, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4677 | "vaesenc " #o1 "(%[KEY]), %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4678 | "vpclmulqdq $0x00, %%xmm12, %%xmm0, %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4679 | "vaesenc " #o1 "(%[KEY]), %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 4680 | "vaesenc " #o1 "(%[KEY]), %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 4681 | "vpclmulqdq $0x00, %%xmm14, %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4682 | "vaesenc " #o1 "(%[KEY]), %%xmm9, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 4683 | "vaesenc " #o1 "(%[KEY]), %%xmm10, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4684 | "vaesenc " #o1 "(%[KEY]), %%xmm11, %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 4685 | "vpxor %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4686 | "vpxor %%xmm12, %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4687 | "vpxor %%xmm15, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4688 | "vpxor %%xmm15, %%xmm3, %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 4689 | "vpxor %%xmm13, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4690 | |
wolfSSL | 15:117db924cf7c | 4691 | #define VAESENC_PCLMUL_L(o) \ |
wolfSSL | 15:117db924cf7c | 4692 | "vpslldq $8, %%xmm1, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 4693 | "vpsrldq $8, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4694 | "vaesenc "#o"(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4695 | "vpxor %%xmm14, %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4696 | "vpxor %%xmm1, %%xmm3, %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 4697 | "vaesenc "#o"(%[KEY]), %%xmm5, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4698 | "vpslld $31, %%xmm2, %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4699 | "vpslld $30, %%xmm2, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4700 | "vpslld $25, %%xmm2, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 4701 | "vaesenc "#o"(%[KEY]), %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4702 | "vpxor %%xmm13, %%xmm12, %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4703 | "vpxor %%xmm14, %%xmm12, %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4704 | "vaesenc "#o"(%[KEY]), %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 4705 | "vpsrldq $4, %%xmm12, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4706 | "vpslldq $12, %%xmm12, %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4707 | "vaesenc "#o"(%[KEY]), %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 4708 | "vpxor %%xmm12, %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4709 | "vpsrld $1, %%xmm2, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 4710 | "vaesenc "#o"(%[KEY]), %%xmm9, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 4711 | "vpsrld $2, %%xmm2, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4712 | "vpsrld $7, %%xmm2, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4713 | "vaesenc "#o"(%[KEY]), %%xmm10, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4714 | "vpxor %%xmm1, %%xmm14, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 4715 | "vpxor %%xmm0, %%xmm14, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 4716 | "vaesenc "#o"(%[KEY]), %%xmm11, %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 4717 | "vpxor %%xmm13, %%xmm14, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 4718 | "vpxor %%xmm14, %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4719 | "vpxor %%xmm3, %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4720 | |
wolfSSL | 15:117db924cf7c | 4721 | |
wolfSSL | 15:117db924cf7c | 4722 | /* Encrypt and carry-less multiply with last key. */ |
wolfSSL | 15:117db924cf7c | 4723 | #define VAESENC_LAST(in, out) \ |
wolfSSL | 15:117db924cf7c | 4724 | "vaesenclast %%xmm12, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4725 | "vaesenclast %%xmm12, %%xmm5, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4726 | "vmovdqu (" #in "), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4727 | "vmovdqu 16(" #in "), %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4728 | "vpxor %%xmm0, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4729 | "vpxor %%xmm1, %%xmm5, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4730 | "vmovdqu %%xmm4, (" #out ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 4731 | "vmovdqu %%xmm5, 16(" #out ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 4732 | "vaesenclast %%xmm12, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4733 | "vaesenclast %%xmm12, %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 4734 | "vmovdqu 32(" #in "), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4735 | "vmovdqu 48(" #in "), %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4736 | "vpxor %%xmm0, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4737 | "vpxor %%xmm1, %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 4738 | "vmovdqu %%xmm6, 32(" #out ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 4739 | "vmovdqu %%xmm7, 48(" #out ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 4740 | "vaesenclast %%xmm12, %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 4741 | "vaesenclast %%xmm12, %%xmm9, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 4742 | "vmovdqu 64(" #in "), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4743 | "vmovdqu 80(" #in "), %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4744 | "vpxor %%xmm0, %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 4745 | "vpxor %%xmm1, %%xmm9, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 4746 | "vmovdqu %%xmm8, 64(" #out ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 4747 | "vmovdqu %%xmm9, 80(" #out ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 4748 | "vaesenclast %%xmm12, %%xmm10, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4749 | "vaesenclast %%xmm12, %%xmm11, %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 4750 | "vmovdqu 96(" #in "), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4751 | "vmovdqu 112(" #in "), %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4752 | "vpxor %%xmm0, %%xmm10, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4753 | "vpxor %%xmm1, %%xmm11, %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 4754 | "vmovdqu %%xmm10, 96(" #out ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 4755 | "vmovdqu %%xmm11, 112(" #out ")\n\t" |
wolfSSL | 15:117db924cf7c | 4756 | |
wolfSSL | 15:117db924cf7c | 4757 | #define VAESENC_BLOCK() \ |
wolfSSL | 15:117db924cf7c | 4758 | "vmovdqu " VAR(CTR1) ", %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4759 | "vpshufb %[BSWAP_EPI64], %%xmm5, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4760 | "vpaddd %[ONE], %%xmm5, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4761 | "vmovdqu %%xmm5, " VAR(CTR1) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4762 | "vpxor (%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4763 | "vaesenc 16(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4764 | "vaesenc 32(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4765 | "vaesenc 48(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4766 | "vaesenc 64(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4767 | "vaesenc 80(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4768 | "vaesenc 96(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4769 | "vaesenc 112(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4770 | "vaesenc 128(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4771 | "vaesenc 144(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4772 | "cmpl $11, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 4773 | "vmovdqa 160(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4774 | "jl %=f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4775 | "vaesenc %%xmm5, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4776 | "vaesenc 176(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4777 | "cmpl $13, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 4778 | "vmovdqa 192(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4779 | "jl %=f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4780 | "vaesenc %%xmm5, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4781 | "vaesenc 208(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4782 | "vmovdqa 224(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4783 | "%=:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4784 | "vaesenclast %%xmm5, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4785 | "vmovdqu (%[in]," VAR(KR64) ",1), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4786 | "vpxor %%xmm5, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4787 | "vmovdqu %%xmm4, (%[out]," VAR(KR64) ",1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 4788 | "vpshufb %[BSWAP_MASK], %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4789 | "vpxor %%xmm4, " VAR(XR) ", " VAR(XR) "\n\t" |
wolfSSL | 15:117db924cf7c | 4790 | |
wolfSSL | 15:117db924cf7c | 4791 | #define _VAESENC_GFMUL(in, H, X) \ |
wolfSSL | 15:117db924cf7c | 4792 | "vmovdqu " VAR(CTR1) ", %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4793 | "vpshufb %[BSWAP_EPI64], %%xmm5, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4794 | "vpaddd %[ONE], %%xmm5, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4795 | "vmovdqu %%xmm5, " VAR(CTR1) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4796 | "vpxor (%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4797 | "vpclmulqdq $0x10, " #H ", " #X ", %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4798 | "vaesenc 16(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4799 | "vaesenc 32(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4800 | "vpclmulqdq $0x01, " #H ", " #X ", %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 4801 | "vaesenc 48(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4802 | "vaesenc 64(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4803 | "vpclmulqdq $0x00, " #H ", " #X ", %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 4804 | "vaesenc 80(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4805 | "vpclmulqdq $0x11, " #H ", " #X ", %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4806 | "vaesenc 96(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4807 | "vpxor %%xmm7, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4808 | "vpslldq $8, %%xmm6, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4809 | "vpsrldq $8, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4810 | "vaesenc 112(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4811 | "vpxor %%xmm8, %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4812 | "vpxor %%xmm6, %%xmm1, %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 4813 | "vmovdqa %[MOD2_128], %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4814 | "vpclmulqdq $0x10, %%xmm0, %%xmm2, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 4815 | "vaesenc 128(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4816 | "vpshufd $0x4e, %%xmm2, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4817 | "vpxor %%xmm7, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4818 | "vpclmulqdq $0x10, %%xmm0, %%xmm6, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 4819 | "vaesenc 144(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4820 | "vpshufd $0x4e, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4821 | "vpxor %%xmm7, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4822 | "vpxor %%xmm3, %%xmm6, " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4823 | "cmpl $11, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 4824 | "vmovdqa 160(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4825 | "jl 1f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4826 | "vaesenc %%xmm5, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4827 | "vaesenc 176(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4828 | "cmpl $13, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 4829 | "vmovdqa 192(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4830 | "jl 1f\n\t" \ |
wolfSSL | 15:117db924cf7c | 4831 | "vaesenc %%xmm5, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4832 | "vaesenc 208(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4833 | "vmovdqa 224(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4834 | "1:\n\t" \ |
wolfSSL | 15:117db924cf7c | 4835 | "vaesenclast %%xmm5, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4836 | "vmovdqu " #in ", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4837 | "vpxor %%xmm0, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4838 | "vmovdqu %%xmm4, (%[out]," VAR(KR64) ",1)\n\t" |
wolfSSL | 15:117db924cf7c | 4839 | #define VAESENC_GFMUL(in, H, X) \ |
wolfSSL | 15:117db924cf7c | 4840 | _VAESENC_GFMUL(in, H, X) |
wolfSSL | 15:117db924cf7c | 4841 | |
wolfSSL | 15:117db924cf7c | 4842 | |
wolfSSL | 15:117db924cf7c | 4843 | #define _GHASH_GFMUL_AVX1(r, r2, a, b) \ |
wolfSSL | 15:117db924cf7c | 4844 | "vpshufd $0x4e, "#a", %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4845 | "vpshufd $0x4e, "#b", %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4846 | "vpclmulqdq $0x11, "#a", "#b", %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 4847 | "vpclmulqdq $0x00, "#a", "#b", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4848 | "vpxor "#a", %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4849 | "vpxor "#b", %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4850 | "vpclmulqdq $0x00, %%xmm2, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4851 | "vpxor %%xmm0, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4852 | "vpxor %%xmm3, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4853 | "vmovdqa %%xmm0, "#r2"\n\t" \ |
wolfSSL | 15:117db924cf7c | 4854 | "vmovdqa %%xmm3, " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4855 | "vpslldq $8, %%xmm1, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4856 | "vpsrldq $8, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4857 | "vpxor %%xmm2, "#r2", "#r2"\n\t" \ |
wolfSSL | 15:117db924cf7c | 4858 | "vpxor %%xmm1, " #r ", " #r "\n\t" |
wolfSSL | 15:117db924cf7c | 4859 | #define GHASH_GFMUL_AVX1(r, r2, a, b) \ |
wolfSSL | 15:117db924cf7c | 4860 | _GHASH_GFMUL_AVX1(r, r2, a, b) |
wolfSSL | 15:117db924cf7c | 4861 | |
wolfSSL | 15:117db924cf7c | 4862 | #define _GHASH_GFMUL_XOR_AVX1(r, r2, a, b) \ |
wolfSSL | 15:117db924cf7c | 4863 | "vpshufd $0x4e, "#a", %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4864 | "vpshufd $0x4e, "#b", %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4865 | "vpclmulqdq $0x11, "#a", "#b", %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 4866 | "vpclmulqdq $0x00, "#a", "#b", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4867 | "vpxor "#a", %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4868 | "vpxor "#b", %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4869 | "vpclmulqdq $0x00, %%xmm2, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4870 | "vpxor %%xmm0, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4871 | "vpxor %%xmm3, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4872 | "vpxor %%xmm0, "#r2", "#r2"\n\t" \ |
wolfSSL | 15:117db924cf7c | 4873 | "vpxor %%xmm3, " #r ", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4874 | "vpslldq $8, %%xmm1, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4875 | "vpsrldq $8, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4876 | "vpxor %%xmm2, "#r2", "#r2"\n\t" \ |
wolfSSL | 15:117db924cf7c | 4877 | "vpxor %%xmm1, " #r ", " #r "\n\t" |
wolfSSL | 15:117db924cf7c | 4878 | #define GHASH_GFMUL_XOR_AVX1(r, r2, a, b) \ |
wolfSSL | 15:117db924cf7c | 4879 | _GHASH_GFMUL_XOR_AVX1(r, r2, a, b) |
wolfSSL | 15:117db924cf7c | 4880 | |
wolfSSL | 15:117db924cf7c | 4881 | #define GHASH_MID_AVX1(r, r2) \ |
wolfSSL | 15:117db924cf7c | 4882 | "vpsrld $31, "#r2", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4883 | "vpsrld $31, " #r ", %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4884 | "vpslld $1, "#r2", "#r2"\n\t" \ |
wolfSSL | 15:117db924cf7c | 4885 | "vpslld $1, " #r ", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4886 | "vpsrldq $12, %%xmm0, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4887 | "vpslldq $4, %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4888 | "vpslldq $4, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4889 | "vpor %%xmm2, " #r ", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4890 | "vpor %%xmm0, "#r2", "#r2"\n\t" \ |
wolfSSL | 15:117db924cf7c | 4891 | "vpor %%xmm1, " #r ", " #r "\n\t" |
wolfSSL | 15:117db924cf7c | 4892 | |
wolfSSL | 15:117db924cf7c | 4893 | #define _GHASH_GFMUL_RED_AVX1(r, a, b) \ |
wolfSSL | 15:117db924cf7c | 4894 | "vpshufd $0x4e, "#a", %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4895 | "vpshufd $0x4e, "#b", %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4896 | "vpclmulqdq $0x11, "#a", "#b", %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 4897 | "vpclmulqdq $0x00, "#a", "#b", %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4898 | "vpxor "#a", %%xmm5, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4899 | "vpxor "#b", %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4900 | "vpclmulqdq $0x00, %%xmm6, %%xmm5, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4901 | "vpxor %%xmm4, %%xmm5, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4902 | "vpxor %%xmm7, %%xmm5, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4903 | "vpslldq $8, %%xmm5, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4904 | "vpsrldq $8, %%xmm5, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4905 | "vpxor %%xmm6, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4906 | "vpxor %%xmm5, %%xmm7, " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4907 | "vpslld $31, %%xmm4, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 4908 | "vpslld $30, %%xmm4, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 4909 | "vpslld $25, %%xmm4, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4910 | "vpxor %%xmm9, %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 4911 | "vpxor %%xmm10, %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 4912 | "vpsrldq $4, %%xmm8, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 4913 | "vpslldq $12, %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 4914 | "vpxor %%xmm8, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4915 | "vpsrld $1, %%xmm4, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4916 | "vpsrld $2, %%xmm4, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4917 | "vpsrld $7, %%xmm4, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4918 | "vpxor %%xmm6, %%xmm10, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4919 | "vpxor %%xmm5, %%xmm10, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4920 | "vpxor %%xmm9, %%xmm10, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4921 | "vpxor %%xmm4, %%xmm10, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4922 | "vpxor %%xmm10, " #r ", " #r "\n\t" |
wolfSSL | 15:117db924cf7c | 4923 | #define GHASH_GFMUL_RED_AVX1(r, a, b) \ |
wolfSSL | 15:117db924cf7c | 4924 | _GHASH_GFMUL_RED_AVX1(r, a, b) |
wolfSSL | 15:117db924cf7c | 4925 | |
wolfSSL | 15:117db924cf7c | 4926 | #define _GHASH_GFSQR_RED_AVX1(r, a) \ |
wolfSSL | 15:117db924cf7c | 4927 | "vpclmulqdq $0x00, "#a", "#a", %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4928 | "vpclmulqdq $0x11, "#a", "#a", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4929 | "vpslld $31, %%xmm4, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 4930 | "vpslld $30, %%xmm4, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 4931 | "vpslld $25, %%xmm4, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4932 | "vpxor %%xmm9, %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 4933 | "vpxor %%xmm10, %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 4934 | "vpsrldq $4, %%xmm8, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 4935 | "vpslldq $12, %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 4936 | "vpxor %%xmm8, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 4937 | "vpsrld $1, %%xmm4, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4938 | "vpsrld $2, %%xmm4, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 4939 | "vpsrld $7, %%xmm4, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 4940 | "vpxor %%xmm6, %%xmm10, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4941 | "vpxor %%xmm5, %%xmm10, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4942 | "vpxor %%xmm9, %%xmm10, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4943 | "vpxor %%xmm4, %%xmm10, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 4944 | "vpxor %%xmm10, " #r ", " #r "\n\t" |
wolfSSL | 15:117db924cf7c | 4945 | #define GHASH_GFSQR_RED_AVX1(r, a) \ |
wolfSSL | 15:117db924cf7c | 4946 | _GHASH_GFSQR_RED_AVX1(r, a) |
wolfSSL | 15:117db924cf7c | 4947 | |
wolfSSL | 15:117db924cf7c | 4948 | #define GHASH_RED_AVX1(r, r2) \ |
wolfSSL | 15:117db924cf7c | 4949 | "vpslld $31, "#r2", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4950 | "vpslld $30, "#r2", %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4951 | "vpslld $25, "#r2", %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4952 | "vpxor %%xmm1, %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4953 | "vpxor %%xmm2, %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4954 | "vmovdqa %%xmm0, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4955 | "vpsrldq $4, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4956 | "vpslldq $12, %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4957 | "vpxor %%xmm0, "#r2", "#r2"\n\t" \ |
wolfSSL | 15:117db924cf7c | 4958 | "vpsrld $1, "#r2", %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4959 | "vpsrld $2, "#r2", %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 4960 | "vpsrld $7, "#r2", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 4961 | "vpxor %%xmm3, %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4962 | "vpxor %%xmm0, %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4963 | "vpxor %%xmm1, %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4964 | "vpxor "#r2", %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 4965 | "vpxor %%xmm2, " #r ", " #r "\n\t" |
wolfSSL | 15:117db924cf7c | 4966 | |
wolfSSL | 15:117db924cf7c | 4967 | #define GHASH_GFMUL_RED_XOR_AVX1(r, r2, a, b) \ |
wolfSSL | 15:117db924cf7c | 4968 | GHASH_GFMUL_XOR_AVX1(r, r2, a, b) \ |
wolfSSL | 15:117db924cf7c | 4969 | GHASH_RED_AVX1(r, r2) |
wolfSSL | 15:117db924cf7c | 4970 | |
wolfSSL | 15:117db924cf7c | 4971 | #define GHASH_FULL_AVX1(r, r2, a, b) \ |
wolfSSL | 15:117db924cf7c | 4972 | GHASH_GFMUL_AVX1(r, r2, a, b) \ |
wolfSSL | 15:117db924cf7c | 4973 | GHASH_MID_AVX1(r, r2) \ |
wolfSSL | 15:117db924cf7c | 4974 | GHASH_RED_AVX1(r, r2) |
wolfSSL | 15:117db924cf7c | 4975 | |
wolfSSL | 15:117db924cf7c | 4976 | #define CALC_IV_12_AVX1() \ |
wolfSSL | 15:117db924cf7c | 4977 | "# Calculate values when IV is 12 bytes\n\t" \ |
wolfSSL | 15:117db924cf7c | 4978 | "# Set counter based on IV\n\t" \ |
wolfSSL | 15:117db924cf7c | 4979 | "movl $0x01000000, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 4980 | "vpinsrq $0, 0(%%rax), %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4981 | "vpinsrd $2, 8(%%rax), %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4982 | "vpinsrd $3, %%ecx, %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 4983 | "# H = Encrypt X(=0) and T = Encrypt counter\n\t" \ |
wolfSSL | 15:117db924cf7c | 4984 | "vmovdqa 0(%[KEY]), " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4985 | "vpxor " VAR(HR) ", %%xmm13, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4986 | "vmovdqa 16(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4987 | "vaesenc %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4988 | "vaesenc %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4989 | "vmovdqa 32(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4990 | "vaesenc %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4991 | "vaesenc %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4992 | "vmovdqa 48(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4993 | "vaesenc %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4994 | "vaesenc %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4995 | "vmovdqa 64(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4996 | "vaesenc %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 4997 | "vaesenc %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 4998 | "vmovdqa 80(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 4999 | "vaesenc %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5000 | "vaesenc %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5001 | "vmovdqa 96(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5002 | "vaesenc %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5003 | "vaesenc %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5004 | "vmovdqa 112(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5005 | "vaesenc %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5006 | "vaesenc %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5007 | "vmovdqa 128(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5008 | "vaesenc %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5009 | "vaesenc %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5010 | "vmovdqa 144(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5011 | "vaesenc %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5012 | "vaesenc %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5013 | "cmpl $11, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 5014 | "vmovdqa 160(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5015 | "jl 31f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5016 | "vaesenc %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5017 | "vaesenc %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5018 | "vmovdqa 176(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5019 | "vaesenc %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5020 | "vaesenc %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5021 | "cmpl $13, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 5022 | "vmovdqa 192(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5023 | "jl 31f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5024 | "vaesenc %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5025 | "vaesenc %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5026 | "vmovdqa 208(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5027 | "vaesenc %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5028 | "vaesenc %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5029 | "vmovdqu 224(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5030 | "31:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5031 | "vaesenclast %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5032 | "vaesenclast %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5033 | "vpshufb %[BSWAP_MASK], " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5034 | "vmovdqu %%xmm1, " VAR(TR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5035 | "jmp 39f\n\t" |
wolfSSL | 15:117db924cf7c | 5036 | |
wolfSSL | 15:117db924cf7c | 5037 | #define CALC_IV_AVX1() \ |
wolfSSL | 15:117db924cf7c | 5038 | "# Calculate values when IV is not 12 bytes\n\t" \ |
wolfSSL | 15:117db924cf7c | 5039 | "# H = Encrypt X(=0)\n\t" \ |
wolfSSL | 15:117db924cf7c | 5040 | "vmovdqa 0(%[KEY]), " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5041 | VAESENC_AVX(HR) \ |
wolfSSL | 15:117db924cf7c | 5042 | "vpshufb %[BSWAP_MASK], " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5043 | "# Calc counter\n\t" \ |
wolfSSL | 15:117db924cf7c | 5044 | "# Initialization vector\n\t" \ |
wolfSSL | 15:117db924cf7c | 5045 | "cmpl $0, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5046 | "movq $0, %%rcx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5047 | "je 45f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5048 | "cmpl $16, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5049 | "jl 44f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5050 | "andl $0xfffffff0, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5051 | "\n" \ |
wolfSSL | 15:117db924cf7c | 5052 | "43:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5053 | "vmovdqu (%%rax,%%rcx,1), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5054 | "vpshufb %[BSWAP_MASK], %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5055 | "vpxor %%xmm4, %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5056 | GHASH_FULL_AVX1(%%xmm13, %%xmm12, %%xmm13, HR) \ |
wolfSSL | 15:117db924cf7c | 5057 | "addl $16, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5058 | "cmpl %%edx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5059 | "jl 43b\n\t" \ |
wolfSSL | 15:117db924cf7c | 5060 | "movl %[ibytes], %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5061 | "cmpl %%edx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5062 | "je 45f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5063 | "\n" \ |
wolfSSL | 15:117db924cf7c | 5064 | "44:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5065 | "subq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 5066 | "vpxor %%xmm4, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5067 | "xorl %%ebx, %%ebx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5068 | "vmovdqu %%xmm4, (%%rsp)\n\t" \ |
wolfSSL | 15:117db924cf7c | 5069 | "42:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5070 | "movzbl (%%rax,%%rcx,1), %%r13d\n\t" \ |
wolfSSL | 15:117db924cf7c | 5071 | "movb %%r13b, (%%rsp,%%rbx,1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 5072 | "incl %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5073 | "incl %%ebx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5074 | "cmpl %%edx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5075 | "jl 42b\n\t" \ |
wolfSSL | 15:117db924cf7c | 5076 | "vmovdqu (%%rsp), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5077 | "addq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 5078 | "vpshufb %[BSWAP_MASK], %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5079 | "vpxor %%xmm4, %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5080 | GHASH_FULL_AVX1(%%xmm13, %%xmm12, %%xmm13, HR) \ |
wolfSSL | 15:117db924cf7c | 5081 | "\n" \ |
wolfSSL | 15:117db924cf7c | 5082 | "45:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5083 | "# T = Encrypt counter\n\t" \ |
wolfSSL | 15:117db924cf7c | 5084 | "vpxor %%xmm0, %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5085 | "shll $3, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5086 | "vpinsrq $0, %%rdx, %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5087 | "vpxor %%xmm0, %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5088 | GHASH_FULL_AVX1(%%xmm13, %%xmm12, %%xmm13, HR) \ |
wolfSSL | 15:117db924cf7c | 5089 | "vpshufb %[BSWAP_MASK], %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5090 | "# Encrypt counter\n\t" \ |
wolfSSL | 15:117db924cf7c | 5091 | "vmovdqa 0(%[KEY]), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5092 | "vpxor %%xmm13, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5093 | VAESENC_AVX(%%xmm4) \ |
wolfSSL | 15:117db924cf7c | 5094 | "vmovdqu %%xmm4, " VAR(TR) "\n\t" |
wolfSSL | 15:117db924cf7c | 5095 | |
wolfSSL | 15:117db924cf7c | 5096 | #define CALC_AAD_AVX1() \ |
wolfSSL | 15:117db924cf7c | 5097 | "# Additional authentication data\n\t" \ |
wolfSSL | 15:117db924cf7c | 5098 | "movl %[abytes], %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5099 | "cmpl $0, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5100 | "je 25f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5101 | "movq %[addt], %%rax\n\t" \ |
wolfSSL | 15:117db924cf7c | 5102 | "xorl %%ecx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5103 | "cmpl $16, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5104 | "jl 24f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5105 | "andl $0xfffffff0, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5106 | "\n" \ |
wolfSSL | 15:117db924cf7c | 5107 | "23:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5108 | "vmovdqu (%%rax,%%rcx,1), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5109 | "vpshufb %[BSWAP_MASK], %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5110 | "vpxor %%xmm4, " VAR(XR) ", " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5111 | GHASH_FULL_AVX1(XR, %%xmm12, XR, HR) \ |
wolfSSL | 15:117db924cf7c | 5112 | "addl $16, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5113 | "cmpl %%edx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5114 | "jl 23b\n\t" \ |
wolfSSL | 15:117db924cf7c | 5115 | "movl %[abytes], %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5116 | "cmpl %%edx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5117 | "je 25f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5118 | "\n" \ |
wolfSSL | 15:117db924cf7c | 5119 | "24:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5120 | "subq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 5121 | "vpxor %%xmm4, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5122 | "xorl %%ebx, %%ebx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5123 | "vmovdqu %%xmm4, (%%rsp)\n\t" \ |
wolfSSL | 15:117db924cf7c | 5124 | "22:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5125 | "movzbl (%%rax,%%rcx,1), %%r13d\n\t" \ |
wolfSSL | 15:117db924cf7c | 5126 | "movb %%r13b, (%%rsp,%%rbx,1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 5127 | "incl %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5128 | "incl %%ebx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5129 | "cmpl %%edx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5130 | "jl 22b\n\t" \ |
wolfSSL | 15:117db924cf7c | 5131 | "vmovdqu (%%rsp), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5132 | "addq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 5133 | "vpshufb %[BSWAP_MASK], %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5134 | "vpxor %%xmm4, " VAR(XR) ", " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5135 | GHASH_FULL_AVX1(XR, %%xmm12, XR, HR) \ |
wolfSSL | 15:117db924cf7c | 5136 | "\n" \ |
wolfSSL | 15:117db924cf7c | 5137 | "25:\n\t" |
wolfSSL | 15:117db924cf7c | 5138 | |
wolfSSL | 15:117db924cf7c | 5139 | #define CALC_HT_8_AVX1() \ |
wolfSSL | 15:117db924cf7c | 5140 | "vmovdqa " VAR(XR) ", %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5141 | "# H ^ 1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5142 | "vmovdqu " VAR(HR) ", 0(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 5143 | "# H ^ 2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5144 | GHASH_GFSQR_RED_AVX1(%%xmm0, HR) \ |
wolfSSL | 15:117db924cf7c | 5145 | "vmovdqu %%xmm0 , 16(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 5146 | "# H ^ 3\n\t" \ |
wolfSSL | 15:117db924cf7c | 5147 | GHASH_GFMUL_RED_AVX1(%%xmm1, HR, %%xmm0) \ |
wolfSSL | 15:117db924cf7c | 5148 | "vmovdqu %%xmm1 , 32(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 5149 | "# H ^ 4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5150 | GHASH_GFSQR_RED_AVX1(%%xmm3, %%xmm0) \ |
wolfSSL | 15:117db924cf7c | 5151 | "vmovdqu %%xmm3 , 48(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 5152 | "# H ^ 5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5153 | GHASH_GFMUL_RED_AVX1(%%xmm12, %%xmm0, %%xmm1) \ |
wolfSSL | 15:117db924cf7c | 5154 | "vmovdqu %%xmm12, 64(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 5155 | "# H ^ 6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5156 | GHASH_GFSQR_RED_AVX1(%%xmm12, %%xmm1) \ |
wolfSSL | 15:117db924cf7c | 5157 | "vmovdqu %%xmm12, 80(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 5158 | "# H ^ 7\n\t" \ |
wolfSSL | 15:117db924cf7c | 5159 | GHASH_GFMUL_RED_AVX1(%%xmm12, %%xmm1, %%xmm3) \ |
wolfSSL | 15:117db924cf7c | 5160 | "vmovdqu %%xmm12, 96(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 5161 | "# H ^ 8\n\t" \ |
wolfSSL | 15:117db924cf7c | 5162 | GHASH_GFSQR_RED_AVX1(%%xmm12, %%xmm3) \ |
wolfSSL | 15:117db924cf7c | 5163 | "vmovdqu %%xmm12, 112(" VAR(HTR) ")\n\t" |
wolfSSL | 15:117db924cf7c | 5164 | |
wolfSSL | 15:117db924cf7c | 5165 | #define VAESENC_128_GHASH_AVX1(src, o) \ |
wolfSSL | 15:117db924cf7c | 5166 | "leaq (%[in]," VAR(KR64) ",1), %%rcx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5167 | "leaq (%[out]," VAR(KR64) ",1), %%rdx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5168 | /* src is either %%rcx or %%rdx */ \ |
wolfSSL | 15:117db924cf7c | 5169 | VAESENC_CTR() \ |
wolfSSL | 15:117db924cf7c | 5170 | VAESENC_XOR() \ |
wolfSSL | 15:117db924cf7c | 5171 | VAESENC_PCLMUL_1(src, 16, (o-128), 112) \ |
wolfSSL | 15:117db924cf7c | 5172 | VAESENC_PCLMUL_N(src, 32, (o-112), 96) \ |
wolfSSL | 15:117db924cf7c | 5173 | VAESENC_PCLMUL_N(src, 48, (o- 96), 80) \ |
wolfSSL | 15:117db924cf7c | 5174 | VAESENC_PCLMUL_N(src, 64, (o- 80), 64) \ |
wolfSSL | 15:117db924cf7c | 5175 | VAESENC_PCLMUL_N(src, 80, (o- 64), 48) \ |
wolfSSL | 15:117db924cf7c | 5176 | VAESENC_PCLMUL_N(src, 96, (o- 48), 32) \ |
wolfSSL | 15:117db924cf7c | 5177 | VAESENC_PCLMUL_N(src, 112, (o- 32), 16) \ |
wolfSSL | 15:117db924cf7c | 5178 | VAESENC_PCLMUL_N(src, 128, (o- 16), 0) \ |
wolfSSL | 15:117db924cf7c | 5179 | VAESENC_PCLMUL_L(144) \ |
wolfSSL | 15:117db924cf7c | 5180 | "cmpl $11, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 5181 | "vmovdqa 160(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5182 | "jl 4f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5183 | VAESENC() \ |
wolfSSL | 15:117db924cf7c | 5184 | VAESENC_SET(176) \ |
wolfSSL | 15:117db924cf7c | 5185 | "cmpl $13, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 5186 | "vmovdqa 192(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5187 | "jl 4f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5188 | VAESENC() \ |
wolfSSL | 15:117db924cf7c | 5189 | VAESENC_SET(208) \ |
wolfSSL | 15:117db924cf7c | 5190 | "vmovdqa 224(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5191 | "\n" \ |
wolfSSL | 15:117db924cf7c | 5192 | "4:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5193 | VAESENC_LAST(%%rcx, %%rdx) |
wolfSSL | 15:117db924cf7c | 5194 | |
wolfSSL | 15:117db924cf7c | 5195 | #define _VAESENC_AVX(r) \ |
wolfSSL | 15:117db924cf7c | 5196 | "vaesenc 16(%[KEY]), " #r ", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5197 | "vaesenc 32(%[KEY]), " #r ", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5198 | "vaesenc 48(%[KEY]), " #r ", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5199 | "vaesenc 64(%[KEY]), " #r ", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5200 | "vaesenc 80(%[KEY]), " #r ", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5201 | "vaesenc 96(%[KEY]), " #r ", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5202 | "vaesenc 112(%[KEY]), " #r ", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5203 | "vaesenc 128(%[KEY]), " #r ", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5204 | "vaesenc 144(%[KEY]), " #r ", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5205 | "cmpl $11, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 5206 | "vmovdqa 160(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5207 | "jl %=f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5208 | "vaesenc %%xmm5, " #r ", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5209 | "vaesenc 176(%[KEY]), " #r ", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5210 | "cmpl $13, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 5211 | "vmovdqa 192(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5212 | "jl %=f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5213 | "vaesenc %%xmm5, " #r ", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5214 | "vaesenc 208(%[KEY]), " #r ", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5215 | "vmovdqa 224(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5216 | "%=:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5217 | "vaesenclast %%xmm5, " #r ", " #r "\n\t" |
wolfSSL | 15:117db924cf7c | 5218 | #define VAESENC_AVX(r) \ |
wolfSSL | 15:117db924cf7c | 5219 | _VAESENC_AVX(r) |
wolfSSL | 15:117db924cf7c | 5220 | |
wolfSSL | 15:117db924cf7c | 5221 | #define AESENC_LAST15_ENC_AVX1() \ |
wolfSSL | 15:117db924cf7c | 5222 | "movl %[nbytes], %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5223 | "movl %%ecx, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5224 | "andl $0x0f, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5225 | "jz 55f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5226 | "vmovdqu " VAR(CTR1) ", %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5227 | "vpshufb %[BSWAP_EPI64], %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5228 | "vpxor 0(%[KEY]), %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5229 | VAESENC_AVX(%%xmm13) \ |
wolfSSL | 15:117db924cf7c | 5230 | "subq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 5231 | "xorl %%ecx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5232 | "vmovdqu %%xmm13, (%%rsp)\n\t" \ |
wolfSSL | 15:117db924cf7c | 5233 | "\n" \ |
wolfSSL | 15:117db924cf7c | 5234 | "51:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5235 | "movzbl (%[in]," VAR(KR64) ",1), %%r13d\n\t" \ |
wolfSSL | 15:117db924cf7c | 5236 | "xorb (%%rsp,%%rcx,1), %%r13b\n\t" \ |
wolfSSL | 15:117db924cf7c | 5237 | "movb %%r13b, (%[out]," VAR(KR64) ",1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 5238 | "movb %%r13b, (%%rsp,%%rcx,1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 5239 | "incl " VAR(KR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5240 | "incl %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5241 | "cmpl %%edx, " VAR(KR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5242 | "jl 51b\n\t" \ |
wolfSSL | 15:117db924cf7c | 5243 | "xorq %%r13, %%r13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5244 | "cmpl $16, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5245 | "je 53f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5246 | "\n" \ |
wolfSSL | 15:117db924cf7c | 5247 | "52:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5248 | "movb %%r13b, (%%rsp,%%rcx,1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 5249 | "incl %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5250 | "cmpl $16, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5251 | "jl 52b\n\t" \ |
wolfSSL | 15:117db924cf7c | 5252 | "53:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5253 | "vmovdqu (%%rsp), %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5254 | "addq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 5255 | "vpshufb %[BSWAP_MASK], %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5256 | "vpxor %%xmm13, " VAR(XR) ", " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5257 | GHASH_GFMUL_RED_AVX1(XR, HR, XR) \ |
wolfSSL | 15:117db924cf7c | 5258 | |
wolfSSL | 15:117db924cf7c | 5259 | #define AESENC_LAST15_DEC_AVX1() \ |
wolfSSL | 15:117db924cf7c | 5260 | "movl %[nbytes], %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5261 | "movl %%ecx, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5262 | "andl $0x0f, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5263 | "jz 55f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5264 | "vmovdqu " VAR(CTR1) ", %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5265 | "vpshufb %[BSWAP_EPI64], %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5266 | "vpxor 0(%[KEY]), %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5267 | VAESENC_AVX(%%xmm13) \ |
wolfSSL | 15:117db924cf7c | 5268 | "subq $32, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 5269 | "xorl %%ecx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5270 | "vmovdqu %%xmm13, (%%rsp)\n\t" \ |
wolfSSL | 15:117db924cf7c | 5271 | "vpxor %%xmm0, %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5272 | "vmovdqu %%xmm0, 16(%%rsp)\n\t" \ |
wolfSSL | 15:117db924cf7c | 5273 | "\n" \ |
wolfSSL | 15:117db924cf7c | 5274 | "51:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5275 | "movzbl (%[in]," VAR(KR64) ",1), %%r13d\n\t" \ |
wolfSSL | 15:117db924cf7c | 5276 | "movb %%r13b, 16(%%rsp,%%rcx,1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 5277 | "xorb (%%rsp,%%rcx,1), %%r13b\n\t" \ |
wolfSSL | 15:117db924cf7c | 5278 | "movb %%r13b, (%[out]," VAR(KR64) ",1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 5279 | "incl " VAR(KR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5280 | "incl %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5281 | "cmpl %%edx, " VAR(KR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5282 | "jl 51b\n\t" \ |
wolfSSL | 15:117db924cf7c | 5283 | "53:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5284 | "vmovdqu 16(%%rsp), %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5285 | "addq $32, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 5286 | "vpshufb %[BSWAP_MASK], %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5287 | "vpxor %%xmm13, " VAR(XR) ", " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5288 | GHASH_GFMUL_RED_AVX1(XR, HR, XR) \ |
wolfSSL | 15:117db924cf7c | 5289 | |
wolfSSL | 15:117db924cf7c | 5290 | #define CALC_TAG_AVX1() \ |
wolfSSL | 15:117db924cf7c | 5291 | "movl %[nbytes], %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5292 | "movl %[abytes], %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5293 | "shlq $3, %%rdx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5294 | "shlq $3, %%rcx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5295 | "vpinsrq $0, %%rdx, %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5296 | "vpinsrq $1, %%rcx, %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5297 | "vpxor %%xmm0, " VAR(XR) ", " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5298 | GHASH_GFMUL_RED_AVX1(XR, HR, XR) \ |
wolfSSL | 15:117db924cf7c | 5299 | "vpshufb %[BSWAP_MASK], " VAR(XR) ", " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5300 | "vpxor " VAR(TR) ", " VAR(XR) ", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5301 | |
wolfSSL | 15:117db924cf7c | 5302 | #define STORE_TAG_AVX() \ |
wolfSSL | 15:117db924cf7c | 5303 | "cmpl $16, %[tbytes]\n\t" \ |
wolfSSL | 15:117db924cf7c | 5304 | "je 71f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5305 | "xorq %%rcx, %%rcx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5306 | "vmovdqu %%xmm0, (%%rsp)\n\t" \ |
wolfSSL | 15:117db924cf7c | 5307 | "73:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5308 | "movzbl (%%rsp,%%rcx,1), %%r13d\n\t" \ |
wolfSSL | 15:117db924cf7c | 5309 | "movb %%r13b, (%[tag],%%rcx,1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 5310 | "incl %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5311 | "cmpl %[tbytes], %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5312 | "jne 73b\n\t" \ |
wolfSSL | 15:117db924cf7c | 5313 | "jmp 72f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5314 | "\n" \ |
wolfSSL | 15:117db924cf7c | 5315 | "71:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5316 | "vmovdqu %%xmm0, (%[tag])\n\t" \ |
wolfSSL | 15:117db924cf7c | 5317 | "\n" \ |
wolfSSL | 15:117db924cf7c | 5318 | "72:\n\t" |
wolfSSL | 15:117db924cf7c | 5319 | |
wolfSSL | 15:117db924cf7c | 5320 | #define CMP_TAG_AVX() \ |
wolfSSL | 15:117db924cf7c | 5321 | "cmpl $16, %[tbytes]\n\t" \ |
wolfSSL | 15:117db924cf7c | 5322 | "je 71f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5323 | "subq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 5324 | "xorq %%rcx, %%rcx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5325 | "xorq %%rax, %%rax\n\t" \ |
wolfSSL | 15:117db924cf7c | 5326 | "vmovdqu %%xmm0, (%%rsp)\n\t" \ |
wolfSSL | 15:117db924cf7c | 5327 | "\n" \ |
wolfSSL | 15:117db924cf7c | 5328 | "73:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5329 | "movzbl (%%rsp,%%rcx,1), %%r13d\n\t" \ |
wolfSSL | 15:117db924cf7c | 5330 | "xorb (%[tag],%%rcx,1), %%r13b\n\t" \ |
wolfSSL | 15:117db924cf7c | 5331 | "orb %%r13b, %%al\n\t" \ |
wolfSSL | 15:117db924cf7c | 5332 | "incl %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5333 | "cmpl %[tbytes], %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5334 | "jne 73b\n\t" \ |
wolfSSL | 15:117db924cf7c | 5335 | "cmpb $0x00, %%al\n\t" \ |
wolfSSL | 15:117db924cf7c | 5336 | "sete %%al\n\t" \ |
wolfSSL | 15:117db924cf7c | 5337 | "addq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 5338 | "jmp 72f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5339 | "\n" \ |
wolfSSL | 15:117db924cf7c | 5340 | "71:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5341 | "vmovdqu (%[tag]), %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5342 | "vpcmpeqb %%xmm1, %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5343 | "vpmovmskb %%xmm0, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5344 | "# %%edx == 0xFFFF then return 1 else => return 0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5345 | "xorl %%eax, %%eax\n\t" \ |
wolfSSL | 15:117db924cf7c | 5346 | "cmpl $0xffff, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5347 | "sete %%al\n\t" \ |
wolfSSL | 15:117db924cf7c | 5348 | "\n" \ |
wolfSSL | 15:117db924cf7c | 5349 | "72:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5350 | "movl %%eax, (%[res])\n\t" |
wolfSSL | 15:117db924cf7c | 5351 | |
wolfSSL | 15:117db924cf7c | 5352 | static void AES_GCM_encrypt_avx1(const unsigned char *in, unsigned char *out, |
wolfSSL | 15:117db924cf7c | 5353 | const unsigned char* addt, |
wolfSSL | 15:117db924cf7c | 5354 | const unsigned char* ivec, unsigned char *tag, |
wolfSSL | 15:117db924cf7c | 5355 | unsigned int nbytes, unsigned int abytes, |
wolfSSL | 15:117db924cf7c | 5356 | unsigned int ibytes, unsigned int tbytes, |
wolfSSL | 15:117db924cf7c | 5357 | const unsigned char* key, int nr) |
wolfSSL | 15:117db924cf7c | 5358 | { |
wolfSSL | 15:117db924cf7c | 5359 | register const unsigned char* iv asm("rax") = ivec; |
wolfSSL | 15:117db924cf7c | 5360 | register unsigned int ivLen asm("ebx") = ibytes; |
wolfSSL | 15:117db924cf7c | 5361 | |
wolfSSL | 15:117db924cf7c | 5362 | __asm__ __volatile__ ( |
wolfSSL | 15:117db924cf7c | 5363 | "subq $" VAR(STACK_OFFSET) ", %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 5364 | /* Counter is xmm13 */ |
wolfSSL | 15:117db924cf7c | 5365 | "vpxor %%xmm13, %%xmm13, %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 5366 | "vpxor " VAR(XR) ", " VAR(XR) ", " VAR(XR) "\n\t" |
wolfSSL | 15:117db924cf7c | 5367 | "movl %[ibytes], %%edx\n\t" |
wolfSSL | 15:117db924cf7c | 5368 | "cmpl $12, %%edx\n\t" |
wolfSSL | 15:117db924cf7c | 5369 | "jne 35f\n\t" |
wolfSSL | 15:117db924cf7c | 5370 | CALC_IV_12_AVX1() |
wolfSSL | 15:117db924cf7c | 5371 | "\n" |
wolfSSL | 15:117db924cf7c | 5372 | "35:\n\t" |
wolfSSL | 15:117db924cf7c | 5373 | CALC_IV_AVX1() |
wolfSSL | 15:117db924cf7c | 5374 | "\n" |
wolfSSL | 15:117db924cf7c | 5375 | "39:\n\t" |
wolfSSL | 15:117db924cf7c | 5376 | |
wolfSSL | 15:117db924cf7c | 5377 | CALC_AAD_AVX1() |
wolfSSL | 15:117db924cf7c | 5378 | |
wolfSSL | 15:117db924cf7c | 5379 | "# Calculate counter and H\n\t" |
wolfSSL | 15:117db924cf7c | 5380 | "vpsrlq $63, " VAR(HR) ", %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 5381 | "vpsllq $1, " VAR(HR) ", %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 5382 | "vpslldq $8, %%xmm5, %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 5383 | "vpor %%xmm5, %%xmm4, %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 5384 | "vpshufd $0xff, " VAR(HR) ", " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 5385 | "vpsrad $31, " VAR(HR) ", " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 5386 | "vpshufb %[BSWAP_EPI64], %%xmm13, %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 5387 | "vpand %[MOD2_128], " VAR(HR) ", " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 5388 | "vpaddd %[ONE], %%xmm13, %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 5389 | "vpxor %%xmm4, " VAR(HR) ", " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 5390 | "vmovdqu %%xmm13, " VAR(CTR1) "\n\t" |
wolfSSL | 15:117db924cf7c | 5391 | |
wolfSSL | 15:117db924cf7c | 5392 | "xorl " VAR(KR) ", " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 5393 | |
wolfSSL | 15:117db924cf7c | 5394 | #if !defined(AES_GCM_AESNI_NO_UNROLL) && !defined(AES_GCM_AVX1_NO_UNROLL) |
wolfSSL | 15:117db924cf7c | 5395 | "cmpl $128, %[nbytes]\n\t" |
wolfSSL | 15:117db924cf7c | 5396 | "movl %[nbytes], %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 5397 | "jl 5f\n\t" |
wolfSSL | 15:117db924cf7c | 5398 | "andl $0xffffff80, %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 5399 | |
wolfSSL | 15:117db924cf7c | 5400 | CALC_HT_8_AVX1() |
wolfSSL | 15:117db924cf7c | 5401 | |
wolfSSL | 15:117db924cf7c | 5402 | "# First 128 bytes of input\n\t" |
wolfSSL | 15:117db924cf7c | 5403 | VAESENC_128() |
wolfSSL | 15:117db924cf7c | 5404 | |
wolfSSL | 15:117db924cf7c | 5405 | "cmpl $128, %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 5406 | "movl $128, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 5407 | "jle 2f\n\t" |
wolfSSL | 15:117db924cf7c | 5408 | |
wolfSSL | 15:117db924cf7c | 5409 | "# More 128 bytes of input\n\t" |
wolfSSL | 15:117db924cf7c | 5410 | "\n" |
wolfSSL | 15:117db924cf7c | 5411 | "3:\n\t" |
wolfSSL | 15:117db924cf7c | 5412 | VAESENC_128_GHASH_AVX1(%%rdx, 0) |
wolfSSL | 15:117db924cf7c | 5413 | "addl $128, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 5414 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 5415 | "jl 3b\n\t" |
wolfSSL | 15:117db924cf7c | 5416 | "\n" |
wolfSSL | 15:117db924cf7c | 5417 | "2:\n\t" |
wolfSSL | 15:117db924cf7c | 5418 | "vmovdqa %[BSWAP_MASK], %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 5419 | "vpshufb %%xmm13, %%xmm4, %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 5420 | "vpshufb %%xmm13, %%xmm5, %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 5421 | "vpshufb %%xmm13, %%xmm6, %%xmm6\n\t" |
wolfSSL | 15:117db924cf7c | 5422 | "vpshufb %%xmm13, %%xmm7, %%xmm7\n\t" |
wolfSSL | 15:117db924cf7c | 5423 | "vpxor %%xmm2, %%xmm4, %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 5424 | "vpshufb %%xmm13, %%xmm8, %%xmm8\n\t" |
wolfSSL | 15:117db924cf7c | 5425 | "vpshufb %%xmm13, %%xmm9, %%xmm9\n\t" |
wolfSSL | 15:117db924cf7c | 5426 | "vpshufb %%xmm13, %%xmm10, %%xmm10\n\t" |
wolfSSL | 15:117db924cf7c | 5427 | "vpshufb %%xmm13, %%xmm11, %%xmm11\n\t" |
wolfSSL | 15:117db924cf7c | 5428 | |
wolfSSL | 15:117db924cf7c | 5429 | "vmovdqu (" VAR(HTR) "), %%xmm12\n\t" |
wolfSSL | 15:117db924cf7c | 5430 | "vmovdqu 16(" VAR(HTR) "), %%xmm14\n\t" |
wolfSSL | 15:117db924cf7c | 5431 | GHASH_GFMUL_AVX1(XR, %%xmm13, %%xmm11, %%xmm12) |
wolfSSL | 15:117db924cf7c | 5432 | GHASH_GFMUL_XOR_AVX1(XR, %%xmm13, %%xmm10, %%xmm14) |
wolfSSL | 15:117db924cf7c | 5433 | "vmovdqu 32(" VAR(HTR) "), %%xmm12\n\t" |
wolfSSL | 15:117db924cf7c | 5434 | "vmovdqu 48(" VAR(HTR) "), %%xmm14\n\t" |
wolfSSL | 15:117db924cf7c | 5435 | GHASH_GFMUL_XOR_AVX1(XR, %%xmm13, %%xmm9, %%xmm12) |
wolfSSL | 15:117db924cf7c | 5436 | GHASH_GFMUL_XOR_AVX1(XR, %%xmm13, %%xmm8, %%xmm14) |
wolfSSL | 15:117db924cf7c | 5437 | "vmovdqu 64(" VAR(HTR) "), %%xmm12\n\t" |
wolfSSL | 15:117db924cf7c | 5438 | "vmovdqu 80(" VAR(HTR) "), %%xmm14\n\t" |
wolfSSL | 15:117db924cf7c | 5439 | GHASH_GFMUL_XOR_AVX1(XR, %%xmm13, %%xmm7, %%xmm12) |
wolfSSL | 15:117db924cf7c | 5440 | GHASH_GFMUL_XOR_AVX1(XR, %%xmm13, %%xmm6, %%xmm14) |
wolfSSL | 15:117db924cf7c | 5441 | "vmovdqu 96(" VAR(HTR) "), %%xmm12\n\t" |
wolfSSL | 15:117db924cf7c | 5442 | "vmovdqu 112(" VAR(HTR) "), %%xmm14\n\t" |
wolfSSL | 15:117db924cf7c | 5443 | GHASH_GFMUL_XOR_AVX1(XR, %%xmm13, %%xmm5, %%xmm12) |
wolfSSL | 15:117db924cf7c | 5444 | GHASH_GFMUL_RED_XOR_AVX1(XR, %%xmm13, %%xmm4, %%xmm14) |
wolfSSL | 15:117db924cf7c | 5445 | |
wolfSSL | 15:117db924cf7c | 5446 | "vmovdqu 0(" VAR(HTR) "), " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 5447 | "\n" |
wolfSSL | 15:117db924cf7c | 5448 | "5:\n\t" |
wolfSSL | 15:117db924cf7c | 5449 | "movl %[nbytes], %%edx\n\t" |
wolfSSL | 15:117db924cf7c | 5450 | "cmpl %%edx, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 5451 | "jge 55f\n\t" |
wolfSSL | 15:117db924cf7c | 5452 | #endif |
wolfSSL | 15:117db924cf7c | 5453 | |
wolfSSL | 15:117db924cf7c | 5454 | "movl %[nbytes], %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 5455 | "andl $0xfffffff0, %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 5456 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 5457 | "jge 14f\n\t" |
wolfSSL | 15:117db924cf7c | 5458 | |
wolfSSL | 15:117db924cf7c | 5459 | VAESENC_BLOCK() |
wolfSSL | 15:117db924cf7c | 5460 | "addl $16, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 5461 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 5462 | "jge 13f\n\t" |
wolfSSL | 15:117db924cf7c | 5463 | "\n" |
wolfSSL | 15:117db924cf7c | 5464 | "12:\n\t" |
wolfSSL | 15:117db924cf7c | 5465 | "vmovdqu (%[in]," VAR(KR64) ",1), %%xmm9\n\t" |
wolfSSL | 15:117db924cf7c | 5466 | VAESENC_GFMUL(%%xmm9, HR, XR) |
wolfSSL | 15:117db924cf7c | 5467 | "vpshufb %[BSWAP_MASK], %%xmm4, %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 5468 | "addl $16, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 5469 | "vpxor %%xmm4, " VAR(XR) ", " VAR(XR) "\n\t" |
wolfSSL | 15:117db924cf7c | 5470 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 5471 | "jl 12b\n\t" |
wolfSSL | 15:117db924cf7c | 5472 | "\n" |
wolfSSL | 15:117db924cf7c | 5473 | "13:\n\t" |
wolfSSL | 15:117db924cf7c | 5474 | GHASH_GFMUL_RED_AVX1(XR, HR, XR) |
wolfSSL | 15:117db924cf7c | 5475 | "\n" |
wolfSSL | 15:117db924cf7c | 5476 | "14:\n\t" |
wolfSSL | 15:117db924cf7c | 5477 | |
wolfSSL | 15:117db924cf7c | 5478 | AESENC_LAST15_ENC_AVX1() |
wolfSSL | 15:117db924cf7c | 5479 | "\n" |
wolfSSL | 15:117db924cf7c | 5480 | "55:\n\t" |
wolfSSL | 15:117db924cf7c | 5481 | |
wolfSSL | 15:117db924cf7c | 5482 | CALC_TAG_AVX1() |
wolfSSL | 15:117db924cf7c | 5483 | STORE_TAG_AVX() |
wolfSSL | 15:117db924cf7c | 5484 | "addq $" VAR(STACK_OFFSET) ", %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 5485 | "vzeroupper\n\t" |
wolfSSL | 15:117db924cf7c | 5486 | |
wolfSSL | 15:117db924cf7c | 5487 | : |
wolfSSL | 15:117db924cf7c | 5488 | : [KEY] "r" (key), |
wolfSSL | 15:117db924cf7c | 5489 | [in] "r" (in), [out] "r" (out), [nr] "r" (nr), |
wolfSSL | 15:117db924cf7c | 5490 | [nbytes] "r" (nbytes), [abytes] "r" (abytes), [addt] "r" (addt), |
wolfSSL | 15:117db924cf7c | 5491 | [ivec] "r" (iv), [ibytes] "r" (ivLen), [tbytes] "r" (tbytes), |
wolfSSL | 15:117db924cf7c | 5492 | [tag] "r" (tag), |
wolfSSL | 15:117db924cf7c | 5493 | [BSWAP_MASK] "m" (BSWAP_MASK), |
wolfSSL | 15:117db924cf7c | 5494 | [BSWAP_EPI64] "m" (BSWAP_EPI64), |
wolfSSL | 15:117db924cf7c | 5495 | [ONE] "m" (ONE), |
wolfSSL | 15:117db924cf7c | 5496 | #if !defined(AES_GCM_AESNI_NO_UNROLL) && !defined(AES_GCM_AVX1_NO_UNROLL) |
wolfSSL | 15:117db924cf7c | 5497 | [TWO] "m" (TWO), [THREE] "m" (THREE), [FOUR] "m" (FOUR), |
wolfSSL | 15:117db924cf7c | 5498 | [FIVE] "m" (FIVE), [SIX] "m" (SIX), [SEVEN] "m" (SEVEN), |
wolfSSL | 15:117db924cf7c | 5499 | [EIGHT] "m" (EIGHT), |
wolfSSL | 15:117db924cf7c | 5500 | #endif |
wolfSSL | 15:117db924cf7c | 5501 | [MOD2_128] "m" (MOD2_128) |
wolfSSL | 15:117db924cf7c | 5502 | : "xmm15", "xmm14", "xmm13", "xmm12", |
wolfSSL | 15:117db924cf7c | 5503 | "xmm0", "xmm1", "xmm2", "xmm3", "memory", |
wolfSSL | 15:117db924cf7c | 5504 | "xmm4", "xmm5", "xmm6", "xmm7", "xmm8", "xmm9", "xmm10", "xmm11", |
wolfSSL | 15:117db924cf7c | 5505 | "rcx", "rdx", "r13" |
wolfSSL | 15:117db924cf7c | 5506 | ); |
wolfSSL | 15:117db924cf7c | 5507 | } |
wolfSSL | 15:117db924cf7c | 5508 | |
wolfSSL | 15:117db924cf7c | 5509 | #ifdef HAVE_INTEL_AVX2 |
wolfSSL | 15:117db924cf7c | 5510 | /* Encrypt and carry-less multiply for AVX2. */ |
wolfSSL | 15:117db924cf7c | 5511 | #define VAESENC_PCLMUL_AVX2_1(src, o1, o2, o3) \ |
wolfSSL | 15:117db924cf7c | 5512 | "vmovdqu " #o2 "(" #src "), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5513 | "vmovdqa " #o1 "(%[KEY]), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5514 | "vpshufb %[BSWAP_MASK], %%xmm12, %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5515 | "vmovdqu " #o3 "(" VAR(HTR) "), %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5516 | "vpxor %%xmm2, %%xmm12, %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5517 | "vpclmulqdq $0x10, %%xmm13, %%xmm12, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5518 | "vpclmulqdq $0x01, %%xmm13, %%xmm12, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 5519 | "vpclmulqdq $0x00, %%xmm13, %%xmm12, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5520 | "vpclmulqdq $0x11, %%xmm13, %%xmm12, %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 5521 | "vaesenc %%xmm0, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5522 | "vaesenc %%xmm0, %%xmm5, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5523 | "vaesenc %%xmm0, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5524 | "vaesenc %%xmm0, %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 5525 | "vaesenc %%xmm0, %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 5526 | "vaesenc %%xmm0, %%xmm9, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 5527 | "vaesenc %%xmm0, %%xmm10, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 5528 | "vaesenc %%xmm0, %%xmm11, %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 5529 | |
wolfSSL | 15:117db924cf7c | 5530 | #define VAESENC_PCLMUL_AVX2_2(src, o1, o2, o3) \ |
wolfSSL | 15:117db924cf7c | 5531 | "vmovdqu " #o2 "(" #src "), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5532 | "vmovdqu " #o3 "(" VAR(HTR) "), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5533 | "vpshufb %[BSWAP_MASK], %%xmm12, %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5534 | "vpxor %%xmm14, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5535 | "vpclmulqdq $0x10, %%xmm0, %%xmm12, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5536 | "vpclmulqdq $0x01, %%xmm0, %%xmm12, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 5537 | "vpclmulqdq $0x00, %%xmm0, %%xmm12, %%xmm15\n\t" \ |
wolfSSL | 15:117db924cf7c | 5538 | "vpclmulqdq $0x11, %%xmm0, %%xmm12, %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5539 | "vmovdqa " #o1 "(%[KEY]), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5540 | "vpxor %%xmm13, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5541 | "vpxor %%xmm12, %%xmm3, %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 5542 | "vaesenc %%xmm0, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5543 | "vaesenc %%xmm0, %%xmm5, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5544 | "vaesenc %%xmm0, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5545 | "vaesenc %%xmm0, %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 5546 | "vaesenc %%xmm0, %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 5547 | "vaesenc %%xmm0, %%xmm9, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 5548 | "vaesenc %%xmm0, %%xmm10, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 5549 | "vaesenc %%xmm0, %%xmm11, %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 5550 | |
wolfSSL | 15:117db924cf7c | 5551 | #define VAESENC_PCLMUL_AVX2_N(src, o1, o2, o3) \ |
wolfSSL | 15:117db924cf7c | 5552 | "vmovdqu " #o2 "(" #src "), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5553 | "vmovdqu " #o3 "(" VAR(HTR) "), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5554 | "vpshufb %[BSWAP_MASK], %%xmm12, %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5555 | "vpxor %%xmm14, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5556 | "vpxor %%xmm15, %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5557 | "vpclmulqdq $0x10, %%xmm0, %%xmm12, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5558 | "vpclmulqdq $0x01, %%xmm0, %%xmm12, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 5559 | "vpclmulqdq $0x00, %%xmm0, %%xmm12, %%xmm15\n\t" \ |
wolfSSL | 15:117db924cf7c | 5560 | "vpclmulqdq $0x11, %%xmm0, %%xmm12, %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5561 | "vmovdqa " #o1 "(%[KEY]), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5562 | "vpxor %%xmm13, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5563 | "vpxor %%xmm12, %%xmm3, %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 5564 | "vaesenc %%xmm0, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5565 | "vaesenc %%xmm0, %%xmm5, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5566 | "vaesenc %%xmm0, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5567 | "vaesenc %%xmm0, %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 5568 | "vaesenc %%xmm0, %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 5569 | "vaesenc %%xmm0, %%xmm9, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 5570 | "vaesenc %%xmm0, %%xmm10, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 5571 | "vaesenc %%xmm0, %%xmm11, %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 5572 | |
wolfSSL | 15:117db924cf7c | 5573 | #define VAESENC_PCLMUL_AVX2_L(o) \ |
wolfSSL | 15:117db924cf7c | 5574 | "vpxor %%xmm14, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5575 | "vpxor %%xmm15, %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5576 | "vpslldq $8, %%xmm1, %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5577 | "vpsrldq $8, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5578 | "vmovdqa "#o"(%[KEY]), %%xmm15\n\t" \ |
wolfSSL | 15:117db924cf7c | 5579 | "vmovdqa %[MOD2_128], %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5580 | "vaesenc %%xmm15, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5581 | "vpxor %%xmm12, %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5582 | "vpxor %%xmm1, %%xmm3, %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 5583 | "vpclmulqdq $0x10, %%xmm0, %%xmm2, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 5584 | "vaesenc %%xmm15, %%xmm5, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5585 | "vaesenc %%xmm15, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5586 | "vaesenc %%xmm15, %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 5587 | "vpshufd $0x4e, %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5588 | "vpxor %%xmm14, %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5589 | "vpclmulqdq $0x10, %%xmm0, %%xmm2, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 5590 | "vaesenc %%xmm15, %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 5591 | "vaesenc %%xmm15, %%xmm9, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 5592 | "vaesenc %%xmm15, %%xmm10, %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 5593 | "vpshufd $0x4e, %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5594 | "vpxor %%xmm14, %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5595 | "vpxor %%xmm3, %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5596 | "vaesenc %%xmm15, %%xmm11, %%xmm11\n\t" |
wolfSSL | 15:117db924cf7c | 5597 | |
wolfSSL | 15:117db924cf7c | 5598 | #define VAESENC_BLOCK_AVX2() \ |
wolfSSL | 15:117db924cf7c | 5599 | "vmovdqu " VAR(CTR1) ", %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5600 | "vpshufb %[BSWAP_EPI64], %%xmm5, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5601 | "vpaddd %[ONE], %%xmm5, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5602 | "vmovdqu %%xmm5, " VAR(CTR1) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5603 | "vpxor (%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5604 | "vaesenc 16(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5605 | "vaesenc 32(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5606 | "vaesenc 48(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5607 | "vaesenc 64(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5608 | "vaesenc 80(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5609 | "vaesenc 96(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5610 | "vaesenc 112(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5611 | "vaesenc 128(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5612 | "vaesenc 144(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5613 | "cmpl $11, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 5614 | "vmovdqa 160(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5615 | "jl %=f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5616 | "vaesenc %%xmm5, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5617 | "vaesenc 176(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5618 | "cmpl $13, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 5619 | "vmovdqa 192(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5620 | "jl %=f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5621 | "vaesenc %%xmm5, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5622 | "vaesenc 208(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5623 | "vmovdqa 224(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5624 | "%=:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5625 | "vaesenclast %%xmm5, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5626 | "vmovdqu (%[in]," VAR(KR64) ",1), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5627 | "vpxor %%xmm5, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5628 | "vmovdqu %%xmm4, (%[out]," VAR(KR64) ",1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 5629 | "vpshufb %[BSWAP_MASK], %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5630 | "vpxor %%xmm4, " VAR(XR) ", " VAR(XR) "\n\t" |
wolfSSL | 15:117db924cf7c | 5631 | |
wolfSSL | 15:117db924cf7c | 5632 | /* Karatsuba multiplication - slower |
wolfSSL | 15:117db924cf7c | 5633 | * H01 = H[1] ^ H[0] (top and bottom 64-bits XORed) |
wolfSSL | 15:117db924cf7c | 5634 | */ |
wolfSSL | 15:117db924cf7c | 5635 | #define _VAESENC_GFMUL_AVX2(in, H, X, ctr1, H01) \ |
wolfSSL | 15:117db924cf7c | 5636 | "vpxor (%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5637 | "vaesenc 16(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5638 | "vaesenc 32(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5639 | "vaesenc 48(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5640 | "vaesenc 64(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5641 | "vaesenc 80(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5642 | "vaesenc 96(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5643 | "vaesenc 112(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5644 | "vaesenc 128(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5645 | "vaesenc 144(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5646 | "cmpl $11, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 5647 | "vmovdqa 160(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5648 | "jl %=f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5649 | "vaesenc %%xmm5, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5650 | "vaesenc 176(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5651 | "cmpl $13, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 5652 | "vmovdqa 192(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5653 | "jl %=f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5654 | "vaesenc %%xmm5, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5655 | "vaesenc 208(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5656 | "vmovdqa 224(%[KEY]), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5657 | "%=:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5658 | "vaesenclast %%xmm5, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5659 | "vmovdqu " #in ", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5660 | "vpxor %%xmm0, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5661 | \ |
wolfSSL | 15:117db924cf7c | 5662 | "vpsrldq $8, " #X ", %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5663 | "vpxor " #X ", %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5664 | "vpclmulqdq $0x00, " #H ", " #X ", %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5665 | "vpclmulqdq $0x11, " #H ", " #X ", %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 5666 | "vpclmulqdq $0x00, "#H01", %%xmm2, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 5667 | "vpxor %%xmm5, %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 5668 | "vpxor %%xmm8, %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 5669 | "vpslldq $8, %%xmm7, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5670 | "vpsrldq $8, %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 5671 | "vpxor %%xmm7, %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 5672 | "vpxor %%xmm5, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5673 | \ |
wolfSSL | 15:117db924cf7c | 5674 | "vpclmulqdq $0x10, %[MOD2_128], %%xmm6, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5675 | "vpshufd $0x4e, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5676 | "vpxor %%xmm5, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5677 | "vpclmulqdq $0x10, %[MOD2_128], %%xmm6, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5678 | "vpshufd $0x4e, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5679 | "vpxor %%xmm8, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5680 | "vpxor %%xmm5, %%xmm6, " VAR(XR) "\n\t" |
wolfSSL | 15:117db924cf7c | 5681 | #define VAESENC_GFMUL_AVX2(in, H, X, ctr1) \ |
wolfSSL | 15:117db924cf7c | 5682 | _VAESENC_GFMUL_AVX2(in, H, X, ctr1) |
wolfSSL | 15:117db924cf7c | 5683 | |
wolfSSL | 15:117db924cf7c | 5684 | #define _VAESENC_GFMUL_SB_AVX2(in, H, X, ctr1) \ |
wolfSSL | 15:117db924cf7c | 5685 | "vpclmulqdq $0x10, " #H ", " #X ", %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 5686 | "vpclmulqdq $0x01, " #H ", " #X ", %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5687 | "vpclmulqdq $0x00, " #H ", " #X ", %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5688 | "vpclmulqdq $0x11, " #H ", " #X ", %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 5689 | "vpxor (%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5690 | "vaesenc 16(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5691 | "vpxor %%xmm6, %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 5692 | "vpslldq $8, %%xmm7, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5693 | "vpsrldq $8, %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 5694 | "vaesenc 32(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5695 | "vpxor %%xmm5, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5696 | "vpclmulqdq $0x10, %[MOD2_128], %%xmm6, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5697 | "vaesenc 48(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5698 | "vaesenc 64(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5699 | "vaesenc 80(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5700 | "vpshufd $0x4e, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5701 | "vpxor %%xmm5, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5702 | "vpclmulqdq $0x10, %[MOD2_128], %%xmm6, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5703 | "vaesenc 96(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5704 | "vaesenc 112(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5705 | "vaesenc 128(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5706 | "vpshufd $0x4e, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5707 | "vaesenc 144(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5708 | "vpxor %%xmm7, %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 5709 | "vpxor %%xmm8, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5710 | "cmpl $11, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 5711 | "vmovdqa 160(%[KEY]), %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 5712 | "jl %=f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5713 | "vaesenc %%xmm3, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5714 | "vaesenc 176(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5715 | "cmpl $13, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 5716 | "vmovdqa 192(%[KEY]), %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 5717 | "jl %=f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5718 | "vaesenc %%xmm3, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5719 | "vaesenc 208(%[KEY]), %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5720 | "vmovdqa 224(%[KEY]), %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 5721 | "%=:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5722 | "vaesenclast %%xmm3, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5723 | "vpxor %%xmm5, %%xmm6, " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5724 | "vmovdqu " #in ", %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5725 | "vpxor %%xmm5, %%xmm4, %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 5726 | #define VAESENC_GFMUL_SB_AVX2(in, H, X, ctr1) \ |
wolfSSL | 15:117db924cf7c | 5727 | _VAESENC_GFMUL_SB_AVX2(in, H, X, ctr1) |
wolfSSL | 15:117db924cf7c | 5728 | |
wolfSSL | 15:117db924cf7c | 5729 | |
wolfSSL | 15:117db924cf7c | 5730 | #define _GHASH_GFMUL_AVX2(r, r2, a, b) \ |
wolfSSL | 15:117db924cf7c | 5731 | "vpclmulqdq $0x10, "#a", "#b", %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5732 | "vpclmulqdq $0x01, "#a", "#b", %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5733 | "vpclmulqdq $0x00, "#a", "#b", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5734 | "vpclmulqdq $0x11, "#a", "#b", %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 5735 | "vpxor %%xmm1, %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5736 | "vpslldq $8, %%xmm2, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5737 | "vpsrldq $8, %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5738 | "vpxor %%xmm1, %%xmm0, "#r2"\n\t" \ |
wolfSSL | 15:117db924cf7c | 5739 | "vpxor %%xmm2, %%xmm3, " #r "\n\t" |
wolfSSL | 15:117db924cf7c | 5740 | #define GHASH_GFMUL_AVX2(r, r2, a, b) \ |
wolfSSL | 15:117db924cf7c | 5741 | _GHASH_GFMUL_AVX2(r, r2, a, b) |
wolfSSL | 15:117db924cf7c | 5742 | |
wolfSSL | 15:117db924cf7c | 5743 | #define GHASH_MID_AVX2(r, r2) \ |
wolfSSL | 15:117db924cf7c | 5744 | "vpsrld $31, "#r2", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5745 | "vpsrld $31, " #r ", %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5746 | "vpslld $1, "#r2", "#r2"\n\t" \ |
wolfSSL | 15:117db924cf7c | 5747 | "vpslld $1, " #r ", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5748 | "vpsrldq $12, %%xmm0, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5749 | "vpslldq $4, %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5750 | "vpslldq $4, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5751 | "vpor %%xmm2, " #r ", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5752 | "vpor %%xmm0, "#r2", "#r2"\n\t" \ |
wolfSSL | 15:117db924cf7c | 5753 | "vpor %%xmm1, " #r ", " #r "\n\t" |
wolfSSL | 15:117db924cf7c | 5754 | |
wolfSSL | 15:117db924cf7c | 5755 | #define _GHASH_GFMUL_RED_AVX2(r, a, b) \ |
wolfSSL | 15:117db924cf7c | 5756 | "vpclmulqdq $0x10, "#a", "#b", %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 5757 | "vpclmulqdq $0x01, "#a", "#b", %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5758 | "vpclmulqdq $0x00, "#a", "#b", %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5759 | "vpxor %%xmm6, %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 5760 | "vpslldq $8, %%xmm7, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5761 | "vpsrldq $8, %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 5762 | "vpxor %%xmm5, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5763 | "vpclmulqdq $0x11, "#a", "#b", %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 5764 | "vpclmulqdq $0x10, %[MOD2_128], %%xmm6, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5765 | "vpshufd $0x4e, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5766 | "vpxor %%xmm5, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5767 | "vpclmulqdq $0x10, %[MOD2_128], %%xmm6, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5768 | "vpshufd $0x4e, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5769 | "vpxor %%xmm7, %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 5770 | "vpxor %%xmm8, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5771 | "vpxor %%xmm5, %%xmm6, " #r "\n\t" |
wolfSSL | 15:117db924cf7c | 5772 | #define GHASH_GFMUL_RED_AVX2(r, a, b) \ |
wolfSSL | 15:117db924cf7c | 5773 | _GHASH_GFMUL_RED_AVX2(r, a, b) |
wolfSSL | 15:117db924cf7c | 5774 | |
wolfSSL | 15:117db924cf7c | 5775 | #define _GHASH_GFSQR_RED2_AVX2(r, a, mod128) \ |
wolfSSL | 15:117db924cf7c | 5776 | "vpclmulqdq $0x00, "#a", "#a", %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5777 | "vpclmulqdq $0x11, "#a", "#a", %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 5778 | "vpclmulqdq $0x10, "#mod128", %%xmm6, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5779 | "vpshufd $0x4e, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5780 | "vpxor %%xmm5, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5781 | "vpclmulqdq $0x10, "#mod128", %%xmm6, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5782 | "vpshufd $0x4e, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5783 | "vpxor %%xmm5, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5784 | "vpxor %%xmm6, %%xmm8, " #r "\n\t" |
wolfSSL | 15:117db924cf7c | 5785 | #define GHASH_GFSQR_RED2_AVX2(r, a, mod128) \ |
wolfSSL | 15:117db924cf7c | 5786 | _GHASH_GFSQR_RED2_AVX2(r, a, mod128) |
wolfSSL | 15:117db924cf7c | 5787 | |
wolfSSL | 15:117db924cf7c | 5788 | #define _GHASH_GFMUL_SQR_RED2_AVX2(rm, rs, a, b, mod128) \ |
wolfSSL | 15:117db924cf7c | 5789 | "vpclmulqdq $0x10, "#a", "#b", %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 5790 | "vpclmulqdq $0x01, "#a", "#b", %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5791 | "vpclmulqdq $0x00, "#a", "#b", %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5792 | "vpclmulqdq $0x11, "#a", "#b", %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 5793 | "vpclmulqdq $0x00, "#b", "#b", %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 5794 | "vpclmulqdq $0x11, "#b", "#b", %%xmm10\n\t" \ |
wolfSSL | 15:117db924cf7c | 5795 | "vpxor %%xmm6, %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 5796 | "vpslldq $8, %%xmm7, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5797 | "vpsrldq $8, %%xmm7, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 5798 | "vpxor %%xmm5, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5799 | "vpclmulqdq $0x10, "#mod128", %%xmm9, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5800 | "vpclmulqdq $0x10, "#mod128", %%xmm6, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5801 | "vpshufd $0x4e, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5802 | "vpshufd $0x4e, %%xmm9, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 5803 | "vpxor %%xmm5, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5804 | "vpxor %%xmm4, %%xmm9, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 5805 | "vpclmulqdq $0x10, "#mod128", %%xmm6, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 5806 | "vpclmulqdq $0x10, "#mod128", %%xmm9, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5807 | "vpshufd $0x4e, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5808 | "vpshufd $0x4e, %%xmm9, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 5809 | "vpxor %%xmm7, %%xmm8, %%xmm8\n\t" \ |
wolfSSL | 15:117db924cf7c | 5810 | "vpxor %%xmm4, %%xmm9, %%xmm9\n\t" \ |
wolfSSL | 15:117db924cf7c | 5811 | "vpxor %%xmm8, %%xmm6, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5812 | "vpxor %%xmm10, %%xmm9, "#rs"\n\t" \ |
wolfSSL | 15:117db924cf7c | 5813 | "vpxor %%xmm5, %%xmm6, "#rm"\n\t" |
wolfSSL | 15:117db924cf7c | 5814 | #define GHASH_GFMUL_SQR_RED2_AVX2(rm, rs, a, b, mod128) \ |
wolfSSL | 15:117db924cf7c | 5815 | _GHASH_GFMUL_SQR_RED2_AVX2(rm, rs, a, b, mod128) |
wolfSSL | 15:117db924cf7c | 5816 | |
wolfSSL | 15:117db924cf7c | 5817 | #define CALC_HT_8_AVX2() \ |
wolfSSL | 15:117db924cf7c | 5818 | "vmovdqa %[MOD2_128], %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 5819 | "vmovdqa " VAR(XR) ", %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5820 | "# H ^ 1 and H ^ 2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5821 | GHASH_GFSQR_RED2_AVX2(%%xmm0, HR, %%xmm11) \ |
wolfSSL | 15:117db924cf7c | 5822 | "vmovdqu " VAR(HR) ", 0(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 5823 | "vmovdqu %%xmm0 , 16(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 5824 | "# H ^ 3 and H ^ 4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5825 | GHASH_GFMUL_SQR_RED2_AVX2(%%xmm1, %%xmm3, HR, %%xmm0, %%xmm11) \ |
wolfSSL | 15:117db924cf7c | 5826 | "vmovdqu %%xmm1 , 32(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 5827 | "vmovdqu %%xmm3 , 48(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 5828 | "# H ^ 5 and H ^ 6\n\t" \ |
wolfSSL | 15:117db924cf7c | 5829 | GHASH_GFMUL_SQR_RED2_AVX2(%%xmm12, %%xmm0, %%xmm0, %%xmm1, %%xmm11) \ |
wolfSSL | 15:117db924cf7c | 5830 | "vmovdqu %%xmm12, 64(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 5831 | "vmovdqu %%xmm0 , 80(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 5832 | "# H ^ 7 and H ^ 8\n\t" \ |
wolfSSL | 15:117db924cf7c | 5833 | GHASH_GFMUL_SQR_RED2_AVX2(%%xmm12, %%xmm0, %%xmm1, %%xmm3, %%xmm11) \ |
wolfSSL | 15:117db924cf7c | 5834 | "vmovdqu %%xmm12, 96(" VAR(HTR) ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 5835 | "vmovdqu %%xmm0 , 112(" VAR(HTR) ")\n\t" |
wolfSSL | 15:117db924cf7c | 5836 | |
wolfSSL | 15:117db924cf7c | 5837 | #define _GHASH_RED_AVX2(r, r2) \ |
wolfSSL | 15:117db924cf7c | 5838 | "vmovdqa %[MOD2_128], %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5839 | "vpclmulqdq $0x10, %%xmm2, "#r2", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5840 | "vpshufd $0x4e, "#r2", %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5841 | "vpxor %%xmm0, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5842 | "vpclmulqdq $0x10, %%xmm2, %%xmm1, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5843 | "vpshufd $0x4e, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5844 | "vpxor %%xmm0, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5845 | "vpxor %%xmm1, " #r ", " #r "\n\t" |
wolfSSL | 15:117db924cf7c | 5846 | #define GHASH_RED_AVX2(r, r2) \ |
wolfSSL | 15:117db924cf7c | 5847 | _GHASH_RED_AVX2(r, r2) |
wolfSSL | 15:117db924cf7c | 5848 | |
wolfSSL | 15:117db924cf7c | 5849 | #define GHASH_FULL_AVX2(r, r2, a, b) \ |
wolfSSL | 15:117db924cf7c | 5850 | GHASH_GFMUL_AVX2(r, r2, a, b) \ |
wolfSSL | 15:117db924cf7c | 5851 | GHASH_MID_AVX2(r, r2) \ |
wolfSSL | 15:117db924cf7c | 5852 | GHASH_RED_AVX2(r, r2) |
wolfSSL | 15:117db924cf7c | 5853 | |
wolfSSL | 15:117db924cf7c | 5854 | #define _GFMUL_3V_AVX2(r, r2, r3, a, b) \ |
wolfSSL | 15:117db924cf7c | 5855 | "vpclmulqdq $0x10, "#a", "#b", "#r3"\n\t" \ |
wolfSSL | 15:117db924cf7c | 5856 | "vpclmulqdq $0x01, "#a", "#b", %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5857 | "vpclmulqdq $0x00, "#a", "#b", "#r2"\n\t" \ |
wolfSSL | 15:117db924cf7c | 5858 | "vpclmulqdq $0x11, "#a", "#b", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5859 | "vpxor %%xmm1, "#r3", "#r3"\n\t" |
wolfSSL | 15:117db924cf7c | 5860 | #define GFMUL_3V_AVX2(r, r2, r3, a, b) \ |
wolfSSL | 15:117db924cf7c | 5861 | _GFMUL_3V_AVX2(r, r2, r3, a, b) |
wolfSSL | 15:117db924cf7c | 5862 | |
wolfSSL | 15:117db924cf7c | 5863 | #define _GFMUL_XOR_3V_AVX2(r, r2, r3, a, b) \ |
wolfSSL | 15:117db924cf7c | 5864 | "vpclmulqdq $0x10, "#a", "#b", %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5865 | "vpclmulqdq $0x01, "#a", "#b", %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5866 | "vpclmulqdq $0x00, "#a", "#b", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5867 | "vpclmulqdq $0x11, "#a", "#b", %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 5868 | "vpxor %%xmm1, %%xmm2, %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 5869 | "vpxor %%xmm3, " #r ", " #r "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5870 | "vpxor %%xmm2, "#r3", "#r3"\n\t" \ |
wolfSSL | 15:117db924cf7c | 5871 | "vpxor %%xmm0, "#r2", "#r2"\n\t" |
wolfSSL | 15:117db924cf7c | 5872 | #define GFMUL_XOR_3V_AVX2(r, r2, r3, a, b) \ |
wolfSSL | 15:117db924cf7c | 5873 | _GFMUL_XOR_3V_AVX2(r, r2, r3, a, b) |
wolfSSL | 15:117db924cf7c | 5874 | |
wolfSSL | 15:117db924cf7c | 5875 | #define GHASH_GFMUL_RED_8_AVX2() \ |
wolfSSL | 15:117db924cf7c | 5876 | "vmovdqu (" VAR(HTR) "), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5877 | GFMUL_3V_AVX2(XR, %%xmm13, %%xmm14, %%xmm11, %%xmm12) \ |
wolfSSL | 15:117db924cf7c | 5878 | "vmovdqu 16(" VAR(HTR) "), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5879 | GFMUL_XOR_3V_AVX2(XR, %%xmm13, %%xmm14, %%xmm10, %%xmm12) \ |
wolfSSL | 15:117db924cf7c | 5880 | "vmovdqu 32(" VAR(HTR) "), %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 5881 | "vmovdqu 48(" VAR(HTR) "), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5882 | GFMUL_XOR_3V_AVX2(XR, %%xmm13, %%xmm14, %%xmm9, %%xmm11) \ |
wolfSSL | 15:117db924cf7c | 5883 | GFMUL_XOR_3V_AVX2(XR, %%xmm13, %%xmm14, %%xmm8, %%xmm12) \ |
wolfSSL | 15:117db924cf7c | 5884 | "vmovdqu 64(" VAR(HTR) "), %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 5885 | "vmovdqu 80(" VAR(HTR) "), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5886 | GFMUL_XOR_3V_AVX2(XR, %%xmm13, %%xmm14, %%xmm7, %%xmm11) \ |
wolfSSL | 15:117db924cf7c | 5887 | GFMUL_XOR_3V_AVX2(XR, %%xmm13, %%xmm14, %%xmm6, %%xmm12) \ |
wolfSSL | 15:117db924cf7c | 5888 | "vmovdqu 96(" VAR(HTR) "), %%xmm11\n\t" \ |
wolfSSL | 15:117db924cf7c | 5889 | "vmovdqu 112(" VAR(HTR) "), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5890 | GFMUL_XOR_3V_AVX2(XR, %%xmm13, %%xmm14, %%xmm5, %%xmm11) \ |
wolfSSL | 15:117db924cf7c | 5891 | GFMUL_XOR_3V_AVX2(XR, %%xmm13, %%xmm14, %%xmm4, %%xmm12) \ |
wolfSSL | 15:117db924cf7c | 5892 | "vpslldq $8, %%xmm14, %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5893 | "vpsrldq $8, %%xmm14, %%xmm14\n\t" \ |
wolfSSL | 15:117db924cf7c | 5894 | "vpxor %%xmm12, %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5895 | "vpxor %%xmm14, " VAR(XR) ", " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5896 | GHASH_RED_AVX2(XR, %%xmm13) |
wolfSSL | 15:117db924cf7c | 5897 | |
wolfSSL | 15:117db924cf7c | 5898 | #define CALC_IV_12_AVX2() \ |
wolfSSL | 15:117db924cf7c | 5899 | "# Calculate values when IV is 12 bytes\n\t" \ |
wolfSSL | 15:117db924cf7c | 5900 | "# Set counter based on IV\n\t" \ |
wolfSSL | 15:117db924cf7c | 5901 | "movl $0x01000000, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5902 | "vpinsrq $0, 0(%%rax), %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5903 | "vpinsrd $2, 8(%%rax), %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5904 | "vpinsrd $3, %%ecx, %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5905 | "# H = Encrypt X(=0) and T = Encrypt counter\n\t" \ |
wolfSSL | 15:117db924cf7c | 5906 | "vmovdqa 0(%[KEY]), " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5907 | "vmovdqa 16(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5908 | "vpxor " VAR(HR) ", %%xmm13, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5909 | "vaesenc %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5910 | "vaesenc %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5911 | "vmovdqa 32(%[KEY]), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5912 | "vmovdqa 48(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5913 | "vaesenc %%xmm0, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5914 | "vaesenc %%xmm0, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5915 | "vaesenc %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5916 | "vaesenc %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5917 | "vmovdqa 64(%[KEY]), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5918 | "vmovdqa 80(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5919 | "vaesenc %%xmm0, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5920 | "vaesenc %%xmm0, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5921 | "vaesenc %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5922 | "vaesenc %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5923 | "vmovdqa 96(%[KEY]), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5924 | "vmovdqa 112(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5925 | "vaesenc %%xmm0, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5926 | "vaesenc %%xmm0, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5927 | "vaesenc %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5928 | "vaesenc %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5929 | "vmovdqa 128(%[KEY]), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5930 | "vmovdqa 144(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5931 | "vaesenc %%xmm0, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5932 | "vaesenc %%xmm0, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5933 | "vaesenc %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5934 | "vaesenc %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5935 | "cmpl $11, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 5936 | "vmovdqa 160(%[KEY]), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5937 | "jl 31f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5938 | "vmovdqa 176(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5939 | "vaesenc %%xmm0, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5940 | "vaesenc %%xmm0, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5941 | "vaesenc %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5942 | "vaesenc %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5943 | "cmpl $13, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 5944 | "vmovdqa 192(%[KEY]), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5945 | "jl 31f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5946 | "vmovdqa 208(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 5947 | "vaesenc %%xmm0, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5948 | "vaesenc %%xmm0, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5949 | "vaesenc %%xmm12, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5950 | "vaesenc %%xmm12, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5951 | "vmovdqu 224(%[KEY]), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 5952 | "31:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5953 | "vaesenclast %%xmm0, " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5954 | "vaesenclast %%xmm0, %%xmm1, %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 5955 | "vpshufb %[BSWAP_MASK], " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5956 | "vmovdqu %%xmm1, " VAR(TR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5957 | |
wolfSSL | 15:117db924cf7c | 5958 | #define CALC_IV_AVX2() \ |
wolfSSL | 15:117db924cf7c | 5959 | "# Calculate values when IV is not 12 bytes\n\t" \ |
wolfSSL | 15:117db924cf7c | 5960 | "# H = Encrypt X(=0)\n\t" \ |
wolfSSL | 15:117db924cf7c | 5961 | "vmovdqa 0(%[KEY]), " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5962 | VAESENC_AVX(HR) \ |
wolfSSL | 15:117db924cf7c | 5963 | "vpshufb %[BSWAP_MASK], " VAR(HR) ", " VAR(HR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 5964 | "# Calc counter\n\t" \ |
wolfSSL | 15:117db924cf7c | 5965 | "# Initialization vector\n\t" \ |
wolfSSL | 15:117db924cf7c | 5966 | "cmpl $0, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5967 | "movq $0, %%rcx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5968 | "je 45f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5969 | "cmpl $16, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5970 | "jl 44f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5971 | "andl $0xfffffff0, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5972 | "\n" \ |
wolfSSL | 15:117db924cf7c | 5973 | "43:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5974 | "vmovdqu (%%rax,%%rcx,1), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5975 | "vpshufb %[BSWAP_MASK], %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5976 | "vpxor %%xmm4, %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 5977 | GHASH_FULL_AVX2(%%xmm13, %%xmm12, %%xmm13, HR) \ |
wolfSSL | 15:117db924cf7c | 5978 | "addl $16, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5979 | "cmpl %%edx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5980 | "jl 43b\n\t" \ |
wolfSSL | 15:117db924cf7c | 5981 | "movl %[ibytes], %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5982 | "cmpl %%edx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5983 | "je 45f\n\t" \ |
wolfSSL | 15:117db924cf7c | 5984 | "\n" \ |
wolfSSL | 15:117db924cf7c | 5985 | "44:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5986 | "subq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 5987 | "vpxor %%xmm4, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5988 | "xorl %%ebx, %%ebx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5989 | "vmovdqu %%xmm4, (%%rsp)\n\t" \ |
wolfSSL | 15:117db924cf7c | 5990 | "42:\n\t" \ |
wolfSSL | 15:117db924cf7c | 5991 | "movzbl (%%rax,%%rcx,1), %%r13d\n\t" \ |
wolfSSL | 15:117db924cf7c | 5992 | "movb %%r13b, (%%rsp,%%rbx,1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 5993 | "incl %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5994 | "incl %%ebx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5995 | "cmpl %%edx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 5996 | "jl 42b\n\t" \ |
wolfSSL | 15:117db924cf7c | 5997 | "vmovdqu (%%rsp), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 5998 | "addq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 5999 | "vpshufb %[BSWAP_MASK], %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 6000 | "vpxor %%xmm4, %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 6001 | GHASH_FULL_AVX2(%%xmm13, %%xmm12, %%xmm13, HR) \ |
wolfSSL | 15:117db924cf7c | 6002 | "\n" \ |
wolfSSL | 15:117db924cf7c | 6003 | "45:\n\t" \ |
wolfSSL | 15:117db924cf7c | 6004 | "# T = Encrypt counter\n\t" \ |
wolfSSL | 15:117db924cf7c | 6005 | "vpxor %%xmm0, %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 6006 | "shll $3, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6007 | "vpinsrq $0, %%rdx, %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 6008 | "vpxor %%xmm0, %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 6009 | GHASH_FULL_AVX2(%%xmm13, %%xmm12, %%xmm13, HR) \ |
wolfSSL | 15:117db924cf7c | 6010 | "vpshufb %[BSWAP_MASK], %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 6011 | "# Encrypt counter\n\t" \ |
wolfSSL | 15:117db924cf7c | 6012 | "vmovdqa 0(%[KEY]), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 6013 | "vpxor %%xmm13, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 6014 | VAESENC_AVX(%%xmm4) \ |
wolfSSL | 15:117db924cf7c | 6015 | "vmovdqu %%xmm4, " VAR(TR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6016 | |
wolfSSL | 15:117db924cf7c | 6017 | #define CALC_AAD_AVX2() \ |
wolfSSL | 15:117db924cf7c | 6018 | "# Additional authentication data\n\t" \ |
wolfSSL | 15:117db924cf7c | 6019 | "movl %[abytes], %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6020 | "cmpl $0, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6021 | "je 25f\n\t" \ |
wolfSSL | 15:117db924cf7c | 6022 | "movq %[addt], %%rax\n\t" \ |
wolfSSL | 15:117db924cf7c | 6023 | "xorl %%ecx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6024 | "cmpl $16, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6025 | "jl 24f\n\t" \ |
wolfSSL | 15:117db924cf7c | 6026 | "andl $0xfffffff0, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6027 | "\n" \ |
wolfSSL | 15:117db924cf7c | 6028 | "23:\n\t" \ |
wolfSSL | 15:117db924cf7c | 6029 | "vmovdqu (%%rax,%%rcx,1), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 6030 | "vpshufb %[BSWAP_MASK], %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 6031 | "vpxor %%xmm4, " VAR(XR) ", " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 6032 | GHASH_FULL_AVX2(XR, %%xmm12, XR, HR) \ |
wolfSSL | 15:117db924cf7c | 6033 | "addl $16, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6034 | "cmpl %%edx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6035 | "jl 23b\n\t" \ |
wolfSSL | 15:117db924cf7c | 6036 | "movl %[abytes], %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6037 | "cmpl %%edx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6038 | "je 25f\n\t" \ |
wolfSSL | 15:117db924cf7c | 6039 | "\n" \ |
wolfSSL | 15:117db924cf7c | 6040 | "24:\n\t" \ |
wolfSSL | 15:117db924cf7c | 6041 | "subq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 6042 | "vpxor %%xmm4, %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 6043 | "xorl %%ebx, %%ebx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6044 | "vmovdqu %%xmm4, (%%rsp)\n\t" \ |
wolfSSL | 15:117db924cf7c | 6045 | "22:\n\t" \ |
wolfSSL | 15:117db924cf7c | 6046 | "movzbl (%%rax,%%rcx,1), %%r13d\n\t" \ |
wolfSSL | 15:117db924cf7c | 6047 | "movb %%r13b, (%%rsp,%%rbx,1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 6048 | "incl %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6049 | "incl %%ebx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6050 | "cmpl %%edx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6051 | "jl 22b\n\t" \ |
wolfSSL | 15:117db924cf7c | 6052 | "vmovdqu (%%rsp), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 6053 | "addq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 6054 | "vpshufb %[BSWAP_MASK], %%xmm4, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 6055 | "vpxor %%xmm4, " VAR(XR) ", " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 6056 | GHASH_FULL_AVX2(XR, %%xmm12, XR, HR) \ |
wolfSSL | 15:117db924cf7c | 6057 | "\n" \ |
wolfSSL | 15:117db924cf7c | 6058 | "25:\n\t" |
wolfSSL | 15:117db924cf7c | 6059 | |
wolfSSL | 15:117db924cf7c | 6060 | #define VAESENC_128_GHASH_AVX2(src, o) \ |
wolfSSL | 15:117db924cf7c | 6061 | "leaq (%[in]," VAR(KR64) ",1), %%rcx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6062 | "leaq (%[out]," VAR(KR64) ",1), %%rdx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6063 | /* src is either %%rcx or %%rdx */ \ |
wolfSSL | 15:117db924cf7c | 6064 | VAESENC_CTR() \ |
wolfSSL | 15:117db924cf7c | 6065 | VAESENC_XOR() \ |
wolfSSL | 15:117db924cf7c | 6066 | VAESENC_PCLMUL_AVX2_1(src, 16, (o-128), 112) \ |
wolfSSL | 15:117db924cf7c | 6067 | VAESENC_PCLMUL_AVX2_2(src, 32, (o-112), 96) \ |
wolfSSL | 15:117db924cf7c | 6068 | VAESENC_PCLMUL_AVX2_N(src, 48, (o- 96), 80) \ |
wolfSSL | 15:117db924cf7c | 6069 | VAESENC_PCLMUL_AVX2_N(src, 64, (o- 80), 64) \ |
wolfSSL | 15:117db924cf7c | 6070 | VAESENC_PCLMUL_AVX2_N(src, 80, (o- 64), 48) \ |
wolfSSL | 15:117db924cf7c | 6071 | VAESENC_PCLMUL_AVX2_N(src, 96, (o- 48), 32) \ |
wolfSSL | 15:117db924cf7c | 6072 | VAESENC_PCLMUL_AVX2_N(src, 112, (o- 32), 16) \ |
wolfSSL | 15:117db924cf7c | 6073 | VAESENC_PCLMUL_AVX2_N(src, 128, (o- 16), 0) \ |
wolfSSL | 15:117db924cf7c | 6074 | VAESENC_PCLMUL_AVX2_L(144) \ |
wolfSSL | 15:117db924cf7c | 6075 | "cmpl $11, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 6076 | "vmovdqa 160(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 6077 | "jl 4f\n\t" \ |
wolfSSL | 15:117db924cf7c | 6078 | VAESENC() \ |
wolfSSL | 15:117db924cf7c | 6079 | VAESENC_SET(176) \ |
wolfSSL | 15:117db924cf7c | 6080 | "cmpl $13, %[nr]\n\t" \ |
wolfSSL | 15:117db924cf7c | 6081 | "vmovdqa 192(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 6082 | "jl 4f\n\t" \ |
wolfSSL | 15:117db924cf7c | 6083 | VAESENC() \ |
wolfSSL | 15:117db924cf7c | 6084 | VAESENC_SET(208) \ |
wolfSSL | 15:117db924cf7c | 6085 | "vmovdqa 224(%[KEY]), %%xmm12\n\t" \ |
wolfSSL | 15:117db924cf7c | 6086 | "\n" \ |
wolfSSL | 15:117db924cf7c | 6087 | "4:\n\t" \ |
wolfSSL | 15:117db924cf7c | 6088 | VAESENC_LAST(%%rcx, %%rdx) |
wolfSSL | 15:117db924cf7c | 6089 | |
wolfSSL | 15:117db924cf7c | 6090 | #define AESENC_LAST15_ENC_AVX2() \ |
wolfSSL | 15:117db924cf7c | 6091 | "movl %[nbytes], %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6092 | "movl %%ecx, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6093 | "andl $0x0f, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6094 | "jz 55f\n\t" \ |
wolfSSL | 15:117db924cf7c | 6095 | "vmovdqu " VAR(CTR1) ", %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 6096 | "vpshufb %[BSWAP_EPI64], %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 6097 | "vpxor 0(%[KEY]), %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 6098 | VAESENC_AVX(%%xmm13) \ |
wolfSSL | 15:117db924cf7c | 6099 | "subq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 6100 | "xorl %%ecx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6101 | "vmovdqu %%xmm13, (%%rsp)\n\t" \ |
wolfSSL | 15:117db924cf7c | 6102 | "\n" \ |
wolfSSL | 15:117db924cf7c | 6103 | "51:\n\t" \ |
wolfSSL | 15:117db924cf7c | 6104 | "movzbl (%[in]," VAR(KR64) ",1), %%r13d\n\t" \ |
wolfSSL | 15:117db924cf7c | 6105 | "xorb (%%rsp,%%rcx,1), %%r13b\n\t" \ |
wolfSSL | 15:117db924cf7c | 6106 | "movb %%r13b, (%[out]," VAR(KR64) ",1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 6107 | "movb %%r13b, (%%rsp,%%rcx,1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 6108 | "incl " VAR(KR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 6109 | "incl %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6110 | "cmpl %%edx, " VAR(KR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 6111 | "jl 51b\n\t" \ |
wolfSSL | 15:117db924cf7c | 6112 | "xorq %%r13, %%r13\n\t" \ |
wolfSSL | 15:117db924cf7c | 6113 | "cmpl $16, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6114 | "je 53f\n\t" \ |
wolfSSL | 15:117db924cf7c | 6115 | "\n" \ |
wolfSSL | 15:117db924cf7c | 6116 | "52:\n\t" \ |
wolfSSL | 15:117db924cf7c | 6117 | "movb %%r13b, (%%rsp,%%rcx,1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 6118 | "incl %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6119 | "cmpl $16, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6120 | "jl 52b\n\t" \ |
wolfSSL | 15:117db924cf7c | 6121 | "53:\n\t" \ |
wolfSSL | 15:117db924cf7c | 6122 | "vmovdqu (%%rsp), %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 6123 | "addq $16, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 6124 | "vpshufb %[BSWAP_MASK], %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 6125 | "vpxor %%xmm13, " VAR(XR) ", " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 6126 | GHASH_GFMUL_RED_AVX2(XR, HR, XR) \ |
wolfSSL | 15:117db924cf7c | 6127 | |
wolfSSL | 15:117db924cf7c | 6128 | #define AESENC_LAST15_DEC_AVX2() \ |
wolfSSL | 15:117db924cf7c | 6129 | "movl %[nbytes], %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6130 | "movl %%ecx, %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6131 | "andl $0x0f, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6132 | "jz 55f\n\t" \ |
wolfSSL | 15:117db924cf7c | 6133 | "vmovdqu " VAR(CTR1) ", %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 6134 | "vpshufb %[BSWAP_EPI64], %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 6135 | "vpxor 0(%[KEY]), %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 6136 | VAESENC_AVX(%%xmm13) \ |
wolfSSL | 15:117db924cf7c | 6137 | "subq $32, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 6138 | "xorl %%ecx, %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6139 | "vmovdqu %%xmm13, (%%rsp)\n\t" \ |
wolfSSL | 15:117db924cf7c | 6140 | "vpxor %%xmm0, %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 6141 | "vmovdqu %%xmm0, 16(%%rsp)\n\t" \ |
wolfSSL | 15:117db924cf7c | 6142 | "\n" \ |
wolfSSL | 15:117db924cf7c | 6143 | "51:\n\t" \ |
wolfSSL | 15:117db924cf7c | 6144 | "movzbl (%[in]," VAR(KR64) ",1), %%r13d\n\t" \ |
wolfSSL | 15:117db924cf7c | 6145 | "movb %%r13b, 16(%%rsp,%%rcx,1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 6146 | "xorb (%%rsp,%%rcx,1), %%r13b\n\t" \ |
wolfSSL | 15:117db924cf7c | 6147 | "movb %%r13b, (%[out]," VAR(KR64) ",1)\n\t" \ |
wolfSSL | 15:117db924cf7c | 6148 | "incl " VAR(KR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 6149 | "incl %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6150 | "cmpl %%edx, " VAR(KR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 6151 | "jl 51b\n\t" \ |
wolfSSL | 15:117db924cf7c | 6152 | "53:\n\t" \ |
wolfSSL | 15:117db924cf7c | 6153 | "vmovdqu 16(%%rsp), %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 6154 | "addq $32, %%rsp\n\t" \ |
wolfSSL | 15:117db924cf7c | 6155 | "vpshufb %[BSWAP_MASK], %%xmm13, %%xmm13\n\t" \ |
wolfSSL | 15:117db924cf7c | 6156 | "vpxor %%xmm13, " VAR(XR) ", " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 6157 | GHASH_GFMUL_RED_AVX2(XR, HR, XR) \ |
wolfSSL | 15:117db924cf7c | 6158 | |
wolfSSL | 15:117db924cf7c | 6159 | #define CALC_TAG_AVX2() \ |
wolfSSL | 15:117db924cf7c | 6160 | "movl %[nbytes], %%edx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6161 | "movl %[abytes], %%ecx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6162 | "shlq $3, %%rdx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6163 | "shlq $3, %%rcx\n\t" \ |
wolfSSL | 15:117db924cf7c | 6164 | "vpinsrq $0, %%rdx, %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 6165 | "vpinsrq $1, %%rcx, %%xmm0, %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 6166 | "vpxor %%xmm0, " VAR(XR) ", " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 6167 | GHASH_GFMUL_RED_AVX2(XR, HR, XR) \ |
wolfSSL | 15:117db924cf7c | 6168 | "vpshufb %[BSWAP_MASK], " VAR(XR) ", " VAR(XR) "\n\t" \ |
wolfSSL | 15:117db924cf7c | 6169 | "vpxor " VAR(TR) ", " VAR(XR) ", %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 6170 | |
wolfSSL | 15:117db924cf7c | 6171 | |
wolfSSL | 15:117db924cf7c | 6172 | static void AES_GCM_encrypt_avx2(const unsigned char *in, unsigned char *out, |
wolfSSL | 15:117db924cf7c | 6173 | const unsigned char* addt, |
wolfSSL | 15:117db924cf7c | 6174 | const unsigned char* ivec, unsigned char *tag, |
wolfSSL | 15:117db924cf7c | 6175 | unsigned int nbytes, unsigned int abytes, |
wolfSSL | 15:117db924cf7c | 6176 | unsigned int ibytes, unsigned int tbytes, |
wolfSSL | 15:117db924cf7c | 6177 | const unsigned char* key, int nr) |
wolfSSL | 15:117db924cf7c | 6178 | { |
wolfSSL | 15:117db924cf7c | 6179 | register const unsigned char* iv asm("rax") = ivec; |
wolfSSL | 15:117db924cf7c | 6180 | register unsigned int ivLen asm("ebx") = ibytes; |
wolfSSL | 15:117db924cf7c | 6181 | |
wolfSSL | 15:117db924cf7c | 6182 | __asm__ __volatile__ ( |
wolfSSL | 15:117db924cf7c | 6183 | "subq $" VAR(STACK_OFFSET) ", %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 6184 | /* Counter is xmm13 */ |
wolfSSL | 15:117db924cf7c | 6185 | "vpxor %%xmm13, %%xmm13, %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 6186 | "vpxor " VAR(XR) ", " VAR(XR) ", " VAR(XR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6187 | "movl %[ibytes], %%edx\n\t" |
wolfSSL | 15:117db924cf7c | 6188 | "cmpl $12, %%edx\n\t" |
wolfSSL | 15:117db924cf7c | 6189 | "jne 35f\n\t" |
wolfSSL | 15:117db924cf7c | 6190 | CALC_IV_12_AVX2() |
wolfSSL | 15:117db924cf7c | 6191 | "jmp 39f\n\t" |
wolfSSL | 15:117db924cf7c | 6192 | "\n" |
wolfSSL | 15:117db924cf7c | 6193 | "35:\n\t" |
wolfSSL | 15:117db924cf7c | 6194 | CALC_IV_AVX2() |
wolfSSL | 15:117db924cf7c | 6195 | "\n" |
wolfSSL | 15:117db924cf7c | 6196 | "39:\n\t" |
wolfSSL | 15:117db924cf7c | 6197 | |
wolfSSL | 15:117db924cf7c | 6198 | CALC_AAD_AVX2() |
wolfSSL | 15:117db924cf7c | 6199 | |
wolfSSL | 15:117db924cf7c | 6200 | "# Calculate counter and H\n\t" |
wolfSSL | 15:117db924cf7c | 6201 | "vpsrlq $63, " VAR(HR) ", %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 6202 | "vpsllq $1, " VAR(HR) ", %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 6203 | "vpslldq $8, %%xmm5, %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 6204 | "vpor %%xmm5, %%xmm4, %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 6205 | "vpshufd $0xff, " VAR(HR) ", " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6206 | "vpsrad $31, " VAR(HR) ", " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6207 | "vpshufb %[BSWAP_EPI64], %%xmm13, %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 6208 | "vpand %[MOD2_128], " VAR(HR) ", " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6209 | "vpaddd %[ONE], %%xmm13, %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 6210 | "vpxor %%xmm4, " VAR(HR) ", " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6211 | "vmovdqu %%xmm13, " VAR(CTR1) "\n\t" |
wolfSSL | 15:117db924cf7c | 6212 | |
wolfSSL | 15:117db924cf7c | 6213 | "xorl " VAR(KR) ", " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6214 | |
wolfSSL | 15:117db924cf7c | 6215 | #if !defined(AES_GCM_AESNI_NO_UNROLL) && !defined(AES_GCM_AVX2_NO_UNROLL) |
wolfSSL | 15:117db924cf7c | 6216 | "cmpl $128, %[nbytes]\n\t" |
wolfSSL | 15:117db924cf7c | 6217 | "movl %[nbytes], %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 6218 | "jl 5f\n\t" |
wolfSSL | 15:117db924cf7c | 6219 | "andl $0xffffff80, %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 6220 | |
wolfSSL | 15:117db924cf7c | 6221 | CALC_HT_8_AVX2() |
wolfSSL | 15:117db924cf7c | 6222 | |
wolfSSL | 15:117db924cf7c | 6223 | "# First 128 bytes of input\n\t" |
wolfSSL | 15:117db924cf7c | 6224 | VAESENC_128() |
wolfSSL | 15:117db924cf7c | 6225 | |
wolfSSL | 15:117db924cf7c | 6226 | "cmpl $128, %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 6227 | "movl $128, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6228 | "jle 2f\n\t" |
wolfSSL | 15:117db924cf7c | 6229 | |
wolfSSL | 15:117db924cf7c | 6230 | "# More 128 bytes of input\n\t" |
wolfSSL | 15:117db924cf7c | 6231 | "\n" |
wolfSSL | 15:117db924cf7c | 6232 | "3:\n\t" |
wolfSSL | 15:117db924cf7c | 6233 | VAESENC_128_GHASH_AVX2(%%rdx, 0) |
wolfSSL | 15:117db924cf7c | 6234 | "addl $128, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6235 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6236 | "jl 3b\n\t" |
wolfSSL | 15:117db924cf7c | 6237 | "\n" |
wolfSSL | 15:117db924cf7c | 6238 | "2:\n\t" |
wolfSSL | 15:117db924cf7c | 6239 | "vmovdqa %[BSWAP_MASK], %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 6240 | "vpshufb %%xmm13, %%xmm4, %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 6241 | "vpshufb %%xmm13, %%xmm5, %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 6242 | "vpshufb %%xmm13, %%xmm6, %%xmm6\n\t" |
wolfSSL | 15:117db924cf7c | 6243 | "vpshufb %%xmm13, %%xmm7, %%xmm7\n\t" |
wolfSSL | 15:117db924cf7c | 6244 | "vpshufb %%xmm13, %%xmm8, %%xmm8\n\t" |
wolfSSL | 15:117db924cf7c | 6245 | "vpshufb %%xmm13, %%xmm9, %%xmm9\n\t" |
wolfSSL | 15:117db924cf7c | 6246 | "vpshufb %%xmm13, %%xmm10, %%xmm10\n\t" |
wolfSSL | 15:117db924cf7c | 6247 | "vpshufb %%xmm13, %%xmm11, %%xmm11\n\t" |
wolfSSL | 15:117db924cf7c | 6248 | "vpxor %%xmm2, %%xmm4, %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 6249 | |
wolfSSL | 15:117db924cf7c | 6250 | GHASH_GFMUL_RED_8_AVX2() |
wolfSSL | 15:117db924cf7c | 6251 | |
wolfSSL | 15:117db924cf7c | 6252 | "vmovdqu 0(" VAR(HTR) "), " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6253 | "\n" |
wolfSSL | 15:117db924cf7c | 6254 | "5:\n\t" |
wolfSSL | 15:117db924cf7c | 6255 | "movl %[nbytes], %%edx\n\t" |
wolfSSL | 15:117db924cf7c | 6256 | "cmpl %%edx, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6257 | "jge 55f\n\t" |
wolfSSL | 15:117db924cf7c | 6258 | #endif |
wolfSSL | 15:117db924cf7c | 6259 | |
wolfSSL | 15:117db924cf7c | 6260 | "movl %[nbytes], %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 6261 | "andl $0xfffffff0, %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 6262 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6263 | "jge 14f\n\t" |
wolfSSL | 15:117db924cf7c | 6264 | |
wolfSSL | 15:117db924cf7c | 6265 | VAESENC_BLOCK_AVX2() |
wolfSSL | 15:117db924cf7c | 6266 | "addl $16, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6267 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6268 | "jge 13f\n\t" |
wolfSSL | 15:117db924cf7c | 6269 | "vmovdqa %[MOD2_128], %%xmm0\n\t" |
wolfSSL | 15:117db924cf7c | 6270 | "\n" |
wolfSSL | 15:117db924cf7c | 6271 | "12:\n\t" |
wolfSSL | 15:117db924cf7c | 6272 | "vmovdqu (%[in]," VAR(KR64) ",1), %%xmm9\n\t" |
wolfSSL | 15:117db924cf7c | 6273 | "vmovdqu " VAR(CTR1) ", %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 6274 | "vpshufb %[BSWAP_EPI64], %%xmm5, %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 6275 | "vpaddd %[ONE], %%xmm5, %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 6276 | "vmovdqu %%xmm5, " VAR(CTR1) "\n\t" |
wolfSSL | 15:117db924cf7c | 6277 | VAESENC_GFMUL_SB_AVX2(%%xmm9, HR, XR, CTR1) |
wolfSSL | 15:117db924cf7c | 6278 | "vmovdqu %%xmm4, (%[out]," VAR(KR64) ",1)\n\t" |
wolfSSL | 15:117db924cf7c | 6279 | "vpshufb %[BSWAP_MASK], %%xmm4, %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 6280 | "addl $16, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6281 | "vpxor %%xmm4, " VAR(XR) ", " VAR(XR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6282 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6283 | "jl 12b\n\t" |
wolfSSL | 15:117db924cf7c | 6284 | "\n" |
wolfSSL | 15:117db924cf7c | 6285 | "13:\n\t" |
wolfSSL | 15:117db924cf7c | 6286 | GHASH_GFMUL_RED_AVX2(XR, HR, XR) |
wolfSSL | 15:117db924cf7c | 6287 | "\n" |
wolfSSL | 15:117db924cf7c | 6288 | "14:\n\t" |
wolfSSL | 15:117db924cf7c | 6289 | |
wolfSSL | 15:117db924cf7c | 6290 | AESENC_LAST15_ENC_AVX2() |
wolfSSL | 15:117db924cf7c | 6291 | "\n" |
wolfSSL | 15:117db924cf7c | 6292 | "55:\n\t" |
wolfSSL | 15:117db924cf7c | 6293 | |
wolfSSL | 15:117db924cf7c | 6294 | CALC_TAG_AVX2() |
wolfSSL | 15:117db924cf7c | 6295 | STORE_TAG_AVX() |
wolfSSL | 15:117db924cf7c | 6296 | "addq $" VAR(STACK_OFFSET) ", %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 6297 | "vzeroupper\n\t" |
wolfSSL | 15:117db924cf7c | 6298 | |
wolfSSL | 15:117db924cf7c | 6299 | : |
wolfSSL | 15:117db924cf7c | 6300 | : [KEY] "r" (key), |
wolfSSL | 15:117db924cf7c | 6301 | [in] "r" (in), [out] "r" (out), [nr] "r" (nr), |
wolfSSL | 15:117db924cf7c | 6302 | [nbytes] "r" (nbytes), [abytes] "r" (abytes), [addt] "r" (addt), |
wolfSSL | 15:117db924cf7c | 6303 | [ivec] "r" (iv), [ibytes] "r" (ivLen), [tbytes] "r" (tbytes), |
wolfSSL | 15:117db924cf7c | 6304 | [tag] "r" (tag), |
wolfSSL | 15:117db924cf7c | 6305 | [BSWAP_MASK] "m" (BSWAP_MASK), |
wolfSSL | 15:117db924cf7c | 6306 | [BSWAP_EPI64] "m" (BSWAP_EPI64), |
wolfSSL | 15:117db924cf7c | 6307 | [ONE] "m" (ONE), |
wolfSSL | 15:117db924cf7c | 6308 | #if !defined(AES_GCM_AESNI_NO_UNROLL) && !defined(AES_GCM_AVX2_NO_UNROLL) |
wolfSSL | 15:117db924cf7c | 6309 | [TWO] "m" (TWO), [THREE] "m" (THREE), [FOUR] "m" (FOUR), |
wolfSSL | 15:117db924cf7c | 6310 | [FIVE] "m" (FIVE), [SIX] "m" (SIX), [SEVEN] "m" (SEVEN), |
wolfSSL | 15:117db924cf7c | 6311 | [EIGHT] "m" (EIGHT), |
wolfSSL | 15:117db924cf7c | 6312 | #endif |
wolfSSL | 15:117db924cf7c | 6313 | [MOD2_128] "m" (MOD2_128) |
wolfSSL | 15:117db924cf7c | 6314 | : "xmm15", "xmm14", "xmm13", "xmm12", |
wolfSSL | 15:117db924cf7c | 6315 | "xmm0", "xmm1", "xmm2", "xmm3", "memory", |
wolfSSL | 15:117db924cf7c | 6316 | "xmm4", "xmm5", "xmm6", "xmm7", "xmm8", "xmm9", "xmm10", "xmm11", |
wolfSSL | 15:117db924cf7c | 6317 | "rcx", "rdx", "r13" |
wolfSSL | 15:117db924cf7c | 6318 | ); |
wolfSSL | 15:117db924cf7c | 6319 | } |
wolfSSL | 15:117db924cf7c | 6320 | #endif /* HAVE_INTEL_AVX2 */ |
wolfSSL | 15:117db924cf7c | 6321 | #endif /* HAVE_INTEL_AVX1 */ |
wolfSSL | 15:117db924cf7c | 6322 | |
wolfSSL | 15:117db924cf7c | 6323 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 6324 | /* Figure 10. AES-GCM – Decrypt With Single Block Ghash at a Time */ |
wolfSSL | 15:117db924cf7c | 6325 | |
wolfSSL | 15:117db924cf7c | 6326 | static void AES_GCM_decrypt(const unsigned char *in, unsigned char *out, |
wolfSSL | 15:117db924cf7c | 6327 | const unsigned char* addt, |
wolfSSL | 15:117db924cf7c | 6328 | const unsigned char* ivec, const unsigned char *tag, |
wolfSSL | 15:117db924cf7c | 6329 | int nbytes, int abytes, int ibytes, int tbytes, |
wolfSSL | 15:117db924cf7c | 6330 | const unsigned char* key, int nr, int* res) |
wolfSSL | 15:117db924cf7c | 6331 | { |
wolfSSL | 15:117db924cf7c | 6332 | register const unsigned char* iv asm("rax") = ivec; |
wolfSSL | 15:117db924cf7c | 6333 | register int ivLen asm("ebx") = ibytes; |
wolfSSL | 15:117db924cf7c | 6334 | register int tagLen asm("edx") = tbytes; |
wolfSSL | 15:117db924cf7c | 6335 | |
wolfSSL | 15:117db924cf7c | 6336 | __asm__ __volatile__ ( |
wolfSSL | 15:117db924cf7c | 6337 | "pushq %%rdx\n\t" |
wolfSSL | 15:117db924cf7c | 6338 | "subq $" VAR(STACK_OFFSET) ", %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 6339 | /* Counter is xmm13 */ |
wolfSSL | 15:117db924cf7c | 6340 | "pxor %%xmm13, %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 6341 | "pxor %%xmm15, %%xmm15\n\t" |
wolfSSL | 15:117db924cf7c | 6342 | "movl %[ibytes], %%edx\n\t" |
wolfSSL | 15:117db924cf7c | 6343 | "cmpl $12, %%edx\n\t" |
wolfSSL | 15:117db924cf7c | 6344 | "jne 35f\n\t" |
wolfSSL | 15:117db924cf7c | 6345 | CALC_IV_12() |
wolfSSL | 15:117db924cf7c | 6346 | "\n" |
wolfSSL | 15:117db924cf7c | 6347 | "35:\n\t" |
wolfSSL | 15:117db924cf7c | 6348 | CALC_IV() |
wolfSSL | 15:117db924cf7c | 6349 | "\n" |
wolfSSL | 15:117db924cf7c | 6350 | "39:\n\t" |
wolfSSL | 15:117db924cf7c | 6351 | |
wolfSSL | 15:117db924cf7c | 6352 | CALC_AAD() |
wolfSSL | 15:117db924cf7c | 6353 | |
wolfSSL | 15:117db924cf7c | 6354 | "# Calculate counter and H\n\t" |
wolfSSL | 15:117db924cf7c | 6355 | "pshufb %[BSWAP_EPI64], %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 6356 | "movdqa " VAR(HR) ", %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 6357 | "paddd %[ONE], %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 6358 | "movdqa " VAR(HR) ", %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 6359 | "movdqu %%xmm13, " VAR(CTR1) "\n\t" |
wolfSSL | 15:117db924cf7c | 6360 | "psrlq $63, %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 6361 | "psllq $1, %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 6362 | "pslldq $8, %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 6363 | "por %%xmm5, %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 6364 | "pshufd $0xff, " VAR(HR) ", " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6365 | "psrad $31, " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6366 | "pand %[MOD2_128], " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6367 | "pxor %%xmm4, " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6368 | |
wolfSSL | 15:117db924cf7c | 6369 | "xorl " VAR(KR) ", " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6370 | |
wolfSSL | 15:117db924cf7c | 6371 | #if !defined(AES_GCM_AESNI_NO_UNROLL) && !defined(AES_GCM_AVX1_NO_UNROLL) |
wolfSSL | 15:117db924cf7c | 6372 | "cmpl $128, %[nbytes]\n\t" |
wolfSSL | 15:117db924cf7c | 6373 | "jl 5f\n\t" |
wolfSSL | 15:117db924cf7c | 6374 | |
wolfSSL | 15:117db924cf7c | 6375 | CALC_HT_8_AVX() |
wolfSSL | 15:117db924cf7c | 6376 | |
wolfSSL | 15:117db924cf7c | 6377 | "movl %[nbytes], %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 6378 | "andl $0xffffff80, %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 6379 | "\n" |
wolfSSL | 15:117db924cf7c | 6380 | "2:\n\t" |
wolfSSL | 15:117db924cf7c | 6381 | AESENC_128_GHASH_AVX(%%rcx, 128) |
wolfSSL | 15:117db924cf7c | 6382 | "addl $128, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6383 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6384 | "jl 2b\n\t" |
wolfSSL | 15:117db924cf7c | 6385 | |
wolfSSL | 15:117db924cf7c | 6386 | "movdqa %%xmm2, " VAR(XR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6387 | "movdqu (%%rsp), " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6388 | "5:\n\t" |
wolfSSL | 15:117db924cf7c | 6389 | "movl %[nbytes], %%edx\n\t" |
wolfSSL | 15:117db924cf7c | 6390 | "cmpl %%edx, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6391 | "jge 55f\n\t" |
wolfSSL | 15:117db924cf7c | 6392 | #endif |
wolfSSL | 15:117db924cf7c | 6393 | "movl %[nbytes], %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 6394 | "andl $0xfffffff0, %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 6395 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6396 | "jge 13f\n\t" |
wolfSSL | 15:117db924cf7c | 6397 | |
wolfSSL | 15:117db924cf7c | 6398 | "\n" |
wolfSSL | 15:117db924cf7c | 6399 | "12:\n\t" |
wolfSSL | 15:117db924cf7c | 6400 | "leaq (%[in]," VAR(KR64) ",1), %%rcx\n\t" |
wolfSSL | 15:117db924cf7c | 6401 | "leaq (%[out]," VAR(KR64) ",1), %%rdx\n\t" |
wolfSSL | 15:117db924cf7c | 6402 | "movdqu (%%rcx), %%xmm1\n\t" |
wolfSSL | 15:117db924cf7c | 6403 | "movdqa " VAR(HR) ", %%xmm0\n\t" |
wolfSSL | 15:117db924cf7c | 6404 | "pshufb %[BSWAP_MASK], %%xmm1\n\t" |
wolfSSL | 15:117db924cf7c | 6405 | "pxor " VAR(XR) ", %%xmm1\n\t" |
wolfSSL | 15:117db924cf7c | 6406 | AESENC_GFMUL(%%rcx, %%rdx, %%xmm0, %%xmm1) |
wolfSSL | 15:117db924cf7c | 6407 | "addl $16, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6408 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6409 | "jl 12b\n\t" |
wolfSSL | 15:117db924cf7c | 6410 | "\n" |
wolfSSL | 15:117db924cf7c | 6411 | "13:\n\t" |
wolfSSL | 15:117db924cf7c | 6412 | |
wolfSSL | 15:117db924cf7c | 6413 | AESENC_LAST15_DEC_AVX() |
wolfSSL | 15:117db924cf7c | 6414 | "\n" |
wolfSSL | 15:117db924cf7c | 6415 | "55:\n\t" |
wolfSSL | 15:117db924cf7c | 6416 | |
wolfSSL | 15:117db924cf7c | 6417 | CALC_TAG() |
wolfSSL | 15:117db924cf7c | 6418 | "addq $" VAR(STACK_OFFSET) ", %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 6419 | "popq %%rdx\n\t" |
wolfSSL | 15:117db924cf7c | 6420 | CMP_TAG() |
wolfSSL | 15:117db924cf7c | 6421 | |
wolfSSL | 15:117db924cf7c | 6422 | : |
wolfSSL | 15:117db924cf7c | 6423 | : [KEY] "r" (key), |
wolfSSL | 15:117db924cf7c | 6424 | [in] "r" (in), [out] "r" (out), [nr] "r" (nr), |
wolfSSL | 15:117db924cf7c | 6425 | [nbytes] "r" (nbytes), [abytes] "r" (abytes), [addt] "r" (addt), |
wolfSSL | 15:117db924cf7c | 6426 | [ivec] "r" (iv), [ibytes] "r" (ivLen), [tbytes] "r" (tagLen), |
wolfSSL | 15:117db924cf7c | 6427 | [tag] "r" (tag), [res] "r" (res), |
wolfSSL | 15:117db924cf7c | 6428 | [BSWAP_MASK] "m" (BSWAP_MASK), |
wolfSSL | 15:117db924cf7c | 6429 | [BSWAP_EPI64] "m" (BSWAP_EPI64), |
wolfSSL | 15:117db924cf7c | 6430 | [ONE] "m" (ONE), |
wolfSSL | 15:117db924cf7c | 6431 | #if !defined(AES_GCM_AESNI_NO_UNROLL) && !defined(AES_GCM_AVX1_NO_UNROLL) |
wolfSSL | 15:117db924cf7c | 6432 | [TWO] "m" (TWO), [THREE] "m" (THREE), [FOUR] "m" (FOUR), |
wolfSSL | 15:117db924cf7c | 6433 | [FIVE] "m" (FIVE), [SIX] "m" (SIX), [SEVEN] "m" (SEVEN), |
wolfSSL | 15:117db924cf7c | 6434 | [EIGHT] "m" (EIGHT), |
wolfSSL | 15:117db924cf7c | 6435 | #endif |
wolfSSL | 15:117db924cf7c | 6436 | [MOD2_128] "m" (MOD2_128) |
wolfSSL | 15:117db924cf7c | 6437 | : "xmm15", "xmm14", "xmm13", "xmm12", |
wolfSSL | 15:117db924cf7c | 6438 | "xmm0", "xmm1", "xmm2", "xmm3", "memory", |
wolfSSL | 15:117db924cf7c | 6439 | "xmm4", "xmm5", "xmm6", "xmm7", "xmm8", "xmm9", "xmm10", "xmm11", |
wolfSSL | 15:117db924cf7c | 6440 | "rcx", "r13" |
wolfSSL | 15:117db924cf7c | 6441 | ); |
wolfSSL | 15:117db924cf7c | 6442 | } |
wolfSSL | 15:117db924cf7c | 6443 | |
wolfSSL | 15:117db924cf7c | 6444 | #ifdef HAVE_INTEL_AVX1 |
wolfSSL | 15:117db924cf7c | 6445 | static void AES_GCM_decrypt_avx1(const unsigned char *in, unsigned char *out, |
wolfSSL | 15:117db924cf7c | 6446 | const unsigned char* addt, |
wolfSSL | 15:117db924cf7c | 6447 | const unsigned char* ivec, |
wolfSSL | 15:117db924cf7c | 6448 | const unsigned char *tag, int nbytes, |
wolfSSL | 15:117db924cf7c | 6449 | int abytes, int ibytes, int tbytes, |
wolfSSL | 15:117db924cf7c | 6450 | const unsigned char* key, int nr, int* res) |
wolfSSL | 15:117db924cf7c | 6451 | { |
wolfSSL | 15:117db924cf7c | 6452 | register const unsigned char* iv asm("rax") = ivec; |
wolfSSL | 15:117db924cf7c | 6453 | register int ivLen asm("ebx") = ibytes; |
wolfSSL | 15:117db924cf7c | 6454 | register int tagLen asm("edx") = tbytes; |
wolfSSL | 15:117db924cf7c | 6455 | |
wolfSSL | 15:117db924cf7c | 6456 | __asm__ __volatile__ ( |
wolfSSL | 15:117db924cf7c | 6457 | "pushq %%rdx\n\t" |
wolfSSL | 15:117db924cf7c | 6458 | "subq $" VAR(STACK_OFFSET) ", %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 6459 | /* Counter is xmm13 */ |
wolfSSL | 15:117db924cf7c | 6460 | "vpxor %%xmm13, %%xmm13, %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 6461 | "vpxor %%xmm15, %%xmm15, %%xmm15\n\t" |
wolfSSL | 15:117db924cf7c | 6462 | "movl %[ibytes], %%edx\n\t" |
wolfSSL | 15:117db924cf7c | 6463 | "cmpl $12, %%edx\n\t" |
wolfSSL | 15:117db924cf7c | 6464 | "jne 35f\n\t" |
wolfSSL | 15:117db924cf7c | 6465 | CALC_IV_12_AVX1() |
wolfSSL | 15:117db924cf7c | 6466 | "\n" |
wolfSSL | 15:117db924cf7c | 6467 | "35:\n\t" |
wolfSSL | 15:117db924cf7c | 6468 | CALC_IV_AVX1() |
wolfSSL | 15:117db924cf7c | 6469 | "\n" |
wolfSSL | 15:117db924cf7c | 6470 | "39:\n\t" |
wolfSSL | 15:117db924cf7c | 6471 | |
wolfSSL | 15:117db924cf7c | 6472 | CALC_AAD_AVX1() |
wolfSSL | 15:117db924cf7c | 6473 | |
wolfSSL | 15:117db924cf7c | 6474 | "# Calculate counter and H\n\t" |
wolfSSL | 15:117db924cf7c | 6475 | "vpsrlq $63, " VAR(HR) ", %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 6476 | "vpsllq $1, " VAR(HR) ", %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 6477 | "vpslldq $8, %%xmm5, %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 6478 | "vpor %%xmm5, %%xmm4, %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 6479 | "vpshufd $0xff, " VAR(HR) ", " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6480 | "vpsrad $31, " VAR(HR) ", " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6481 | "vpshufb %[BSWAP_EPI64], %%xmm13, %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 6482 | "vpand %[MOD2_128], " VAR(HR) ", " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6483 | "vpaddd %[ONE], %%xmm13, %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 6484 | "vpxor %%xmm4, " VAR(HR) ", " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6485 | "vmovdqu %%xmm13, " VAR(CTR1) "\n\t" |
wolfSSL | 15:117db924cf7c | 6486 | |
wolfSSL | 15:117db924cf7c | 6487 | "xorl " VAR(KR) ", " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6488 | |
wolfSSL | 15:117db924cf7c | 6489 | #if !defined(AES_GCM_AESNI_NO_UNROLL) && !defined(AES_GCM_AVX1_NO_UNROLL) |
wolfSSL | 15:117db924cf7c | 6490 | "cmpl $128, %[nbytes]\n\t" |
wolfSSL | 15:117db924cf7c | 6491 | "jl 5f\n\t" |
wolfSSL | 15:117db924cf7c | 6492 | |
wolfSSL | 15:117db924cf7c | 6493 | CALC_HT_8_AVX1() |
wolfSSL | 15:117db924cf7c | 6494 | |
wolfSSL | 15:117db924cf7c | 6495 | "movl %[nbytes], %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 6496 | "andl $0xffffff80, %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 6497 | "\n" |
wolfSSL | 15:117db924cf7c | 6498 | "2:\n\t" |
wolfSSL | 15:117db924cf7c | 6499 | VAESENC_128_GHASH_AVX1(%%rcx, 128) |
wolfSSL | 15:117db924cf7c | 6500 | "addl $128, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6501 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6502 | "jl 2b\n\t" |
wolfSSL | 15:117db924cf7c | 6503 | |
wolfSSL | 15:117db924cf7c | 6504 | "vmovdqa %%xmm2, " VAR(XR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6505 | "vmovdqu (%%rsp), " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6506 | "5:\n\t" |
wolfSSL | 15:117db924cf7c | 6507 | "movl %[nbytes], %%edx\n\t" |
wolfSSL | 15:117db924cf7c | 6508 | "cmpl %%edx, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6509 | "jge 55f\n\t" |
wolfSSL | 15:117db924cf7c | 6510 | #endif |
wolfSSL | 15:117db924cf7c | 6511 | "movl %[nbytes], %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 6512 | "andl $0xfffffff0, %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 6513 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6514 | "jge 13f\n\t" |
wolfSSL | 15:117db924cf7c | 6515 | |
wolfSSL | 15:117db924cf7c | 6516 | "\n" |
wolfSSL | 15:117db924cf7c | 6517 | "12:\n\t" |
wolfSSL | 15:117db924cf7c | 6518 | "vmovdqu (%[in]," VAR(KR64) ",1), %%xmm9\n\t" |
wolfSSL | 15:117db924cf7c | 6519 | "vmovdqa " VAR(HR) ", %%xmm0\n\t" |
wolfSSL | 15:117db924cf7c | 6520 | "vpshufb %[BSWAP_MASK], %%xmm9, %%xmm1\n\t" |
wolfSSL | 15:117db924cf7c | 6521 | "vpxor " VAR(XR) ", %%xmm1, %%xmm1\n\t" |
wolfSSL | 15:117db924cf7c | 6522 | VAESENC_GFMUL(%%xmm9, %%xmm0, %%xmm1) |
wolfSSL | 15:117db924cf7c | 6523 | "addl $16, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6524 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6525 | "jl 12b\n\t" |
wolfSSL | 15:117db924cf7c | 6526 | "\n" |
wolfSSL | 15:117db924cf7c | 6527 | "13:\n\t" |
wolfSSL | 15:117db924cf7c | 6528 | |
wolfSSL | 15:117db924cf7c | 6529 | AESENC_LAST15_DEC_AVX1() |
wolfSSL | 15:117db924cf7c | 6530 | "\n" |
wolfSSL | 15:117db924cf7c | 6531 | "55:\n\t" |
wolfSSL | 15:117db924cf7c | 6532 | |
wolfSSL | 15:117db924cf7c | 6533 | CALC_TAG_AVX1() |
wolfSSL | 15:117db924cf7c | 6534 | "addq $" VAR(STACK_OFFSET) ", %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 6535 | "popq %%rdx\n\t" |
wolfSSL | 15:117db924cf7c | 6536 | CMP_TAG_AVX() |
wolfSSL | 15:117db924cf7c | 6537 | "vzeroupper\n\t" |
wolfSSL | 15:117db924cf7c | 6538 | |
wolfSSL | 15:117db924cf7c | 6539 | : |
wolfSSL | 15:117db924cf7c | 6540 | : [KEY] "r" (key), |
wolfSSL | 15:117db924cf7c | 6541 | [in] "r" (in), [out] "r" (out), [nr] "r" (nr), |
wolfSSL | 15:117db924cf7c | 6542 | [nbytes] "r" (nbytes), [abytes] "r" (abytes), [addt] "r" (addt), |
wolfSSL | 15:117db924cf7c | 6543 | [ivec] "r" (iv), [ibytes] "r" (ivLen), [tbytes] "r" (tagLen), |
wolfSSL | 15:117db924cf7c | 6544 | [tag] "r" (tag), [res] "r" (res), |
wolfSSL | 15:117db924cf7c | 6545 | [BSWAP_MASK] "m" (BSWAP_MASK), |
wolfSSL | 15:117db924cf7c | 6546 | [BSWAP_EPI64] "m" (BSWAP_EPI64), |
wolfSSL | 15:117db924cf7c | 6547 | [ONE] "m" (ONE), |
wolfSSL | 15:117db924cf7c | 6548 | #if !defined(AES_GCM_AESNI_NO_UNROLL) && !defined(AES_GCM_AVX1_NO_UNROLL) |
wolfSSL | 15:117db924cf7c | 6549 | [TWO] "m" (TWO), [THREE] "m" (THREE), [FOUR] "m" (FOUR), |
wolfSSL | 15:117db924cf7c | 6550 | [FIVE] "m" (FIVE), [SIX] "m" (SIX), [SEVEN] "m" (SEVEN), |
wolfSSL | 15:117db924cf7c | 6551 | [EIGHT] "m" (EIGHT), |
wolfSSL | 15:117db924cf7c | 6552 | #endif |
wolfSSL | 15:117db924cf7c | 6553 | [MOD2_128] "m" (MOD2_128) |
wolfSSL | 15:117db924cf7c | 6554 | : "xmm15", "xmm14", "xmm13", "xmm12", |
wolfSSL | 15:117db924cf7c | 6555 | "xmm0", "xmm1", "xmm2", "xmm3", "memory", |
wolfSSL | 15:117db924cf7c | 6556 | "xmm4", "xmm5", "xmm6", "xmm7", "xmm8", "xmm9", "xmm10", "xmm11", |
wolfSSL | 15:117db924cf7c | 6557 | "rcx", "r13" |
wolfSSL | 15:117db924cf7c | 6558 | ); |
wolfSSL | 15:117db924cf7c | 6559 | } |
wolfSSL | 15:117db924cf7c | 6560 | |
wolfSSL | 15:117db924cf7c | 6561 | #ifdef HAVE_INTEL_AVX2 |
wolfSSL | 15:117db924cf7c | 6562 | static void AES_GCM_decrypt_avx2(const unsigned char *in, unsigned char *out, |
wolfSSL | 15:117db924cf7c | 6563 | const unsigned char* addt, |
wolfSSL | 15:117db924cf7c | 6564 | const unsigned char* ivec, |
wolfSSL | 15:117db924cf7c | 6565 | const unsigned char *tag, int nbytes, |
wolfSSL | 15:117db924cf7c | 6566 | int abytes, int ibytes, int tbytes, |
wolfSSL | 15:117db924cf7c | 6567 | const unsigned char* key, int nr, int* res) |
wolfSSL | 15:117db924cf7c | 6568 | { |
wolfSSL | 15:117db924cf7c | 6569 | register const unsigned char* iv asm("rax") = ivec; |
wolfSSL | 15:117db924cf7c | 6570 | register int ivLen asm("ebx") = ibytes; |
wolfSSL | 15:117db924cf7c | 6571 | register int tagLen asm("edx") = tbytes; |
wolfSSL | 15:117db924cf7c | 6572 | |
wolfSSL | 15:117db924cf7c | 6573 | __asm__ __volatile__ ( |
wolfSSL | 15:117db924cf7c | 6574 | "pushq %%rdx\n\t" |
wolfSSL | 15:117db924cf7c | 6575 | "subq $" VAR(STACK_OFFSET) ", %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 6576 | /* Counter is xmm13 */ |
wolfSSL | 15:117db924cf7c | 6577 | "vpxor %%xmm13, %%xmm13, %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 6578 | "vpxor %%xmm15, %%xmm15, %%xmm15\n\t" |
wolfSSL | 15:117db924cf7c | 6579 | "movl %[ibytes], %%edx\n\t" |
wolfSSL | 15:117db924cf7c | 6580 | "cmpl $12, %%edx\n\t" |
wolfSSL | 15:117db924cf7c | 6581 | "jne 35f\n\t" |
wolfSSL | 15:117db924cf7c | 6582 | CALC_IV_12_AVX2() |
wolfSSL | 15:117db924cf7c | 6583 | "jmp 39f\n\t" |
wolfSSL | 15:117db924cf7c | 6584 | "\n" |
wolfSSL | 15:117db924cf7c | 6585 | "35:\n\t" |
wolfSSL | 15:117db924cf7c | 6586 | CALC_IV_AVX2() |
wolfSSL | 15:117db924cf7c | 6587 | "\n" |
wolfSSL | 15:117db924cf7c | 6588 | "39:\n\t" |
wolfSSL | 15:117db924cf7c | 6589 | |
wolfSSL | 15:117db924cf7c | 6590 | CALC_AAD_AVX2() |
wolfSSL | 15:117db924cf7c | 6591 | |
wolfSSL | 15:117db924cf7c | 6592 | "# Calculate counter and H\n\t" |
wolfSSL | 15:117db924cf7c | 6593 | "vpsrlq $63, " VAR(HR) ", %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 6594 | "vpsllq $1, " VAR(HR) ", %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 6595 | "vpslldq $8, %%xmm5, %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 6596 | "vpor %%xmm5, %%xmm4, %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 6597 | "vpshufd $0xff, " VAR(HR) ", " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6598 | "vpsrad $31, " VAR(HR) ", " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6599 | "vpshufb %[BSWAP_EPI64], %%xmm13, %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 6600 | "vpand %[MOD2_128], " VAR(HR) ", " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6601 | "vpaddd %[ONE], %%xmm13, %%xmm13\n\t" |
wolfSSL | 15:117db924cf7c | 6602 | "vpxor %%xmm4, " VAR(HR) ", " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6603 | "vmovdqu %%xmm13, " VAR(CTR1) "\n\t" |
wolfSSL | 15:117db924cf7c | 6604 | |
wolfSSL | 15:117db924cf7c | 6605 | "xorl " VAR(KR) ", " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6606 | |
wolfSSL | 15:117db924cf7c | 6607 | #if !defined(AES_GCM_AESNI_NO_UNROLL) && !defined(AES_GCM_AVX2_NO_UNROLL) |
wolfSSL | 15:117db924cf7c | 6608 | "cmpl $128, %[nbytes]\n\t" |
wolfSSL | 15:117db924cf7c | 6609 | "jl 5f\n\t" |
wolfSSL | 15:117db924cf7c | 6610 | |
wolfSSL | 15:117db924cf7c | 6611 | CALC_HT_8_AVX2() |
wolfSSL | 15:117db924cf7c | 6612 | |
wolfSSL | 15:117db924cf7c | 6613 | "movl %[nbytes], %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 6614 | "andl $0xffffff80, %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 6615 | "\n" |
wolfSSL | 15:117db924cf7c | 6616 | "2:\n\t" |
wolfSSL | 15:117db924cf7c | 6617 | VAESENC_128_GHASH_AVX2(%%rcx, 128) |
wolfSSL | 15:117db924cf7c | 6618 | "addl $128, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6619 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6620 | "jl 2b\n\t" |
wolfSSL | 15:117db924cf7c | 6621 | |
wolfSSL | 15:117db924cf7c | 6622 | "vmovdqa %%xmm2, " VAR(XR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6623 | "vmovdqu (%%rsp), " VAR(HR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6624 | "5:\n\t" |
wolfSSL | 15:117db924cf7c | 6625 | "movl %[nbytes], %%edx\n\t" |
wolfSSL | 15:117db924cf7c | 6626 | "cmpl %%edx, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6627 | "jge 55f\n\t" |
wolfSSL | 15:117db924cf7c | 6628 | #endif |
wolfSSL | 15:117db924cf7c | 6629 | "movl %[nbytes], %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 6630 | "andl $0xfffffff0, %%r13d\n\t" |
wolfSSL | 15:117db924cf7c | 6631 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6632 | "jge 13f\n\t" |
wolfSSL | 15:117db924cf7c | 6633 | |
wolfSSL | 15:117db924cf7c | 6634 | "vmovdqa %[MOD2_128], %%xmm0\n\t" |
wolfSSL | 15:117db924cf7c | 6635 | "\n" |
wolfSSL | 15:117db924cf7c | 6636 | "12:\n\t" |
wolfSSL | 15:117db924cf7c | 6637 | "vmovdqu (%[in]," VAR(KR64) ",1), %%xmm9\n\t" |
wolfSSL | 15:117db924cf7c | 6638 | "vmovdqu " VAR(CTR1) ", %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 6639 | "vpshufb %[BSWAP_MASK], %%xmm9, %%xmm1\n\t" |
wolfSSL | 15:117db924cf7c | 6640 | "vpshufb %[BSWAP_EPI64], %%xmm5, %%xmm4\n\t" |
wolfSSL | 15:117db924cf7c | 6641 | "vpaddd %[ONE], %%xmm5, %%xmm5\n\t" |
wolfSSL | 15:117db924cf7c | 6642 | "vpxor " VAR(XR) ", %%xmm1, %%xmm1\n\t" |
wolfSSL | 15:117db924cf7c | 6643 | "vmovdqu %%xmm5, " VAR(CTR1) "\n\t" |
wolfSSL | 15:117db924cf7c | 6644 | VAESENC_GFMUL_SB_AVX2(%%xmm9, HR, %%xmm1, CTR1) |
wolfSSL | 15:117db924cf7c | 6645 | "vmovdqu %%xmm4, (%[out]," VAR(KR64) ",1)\n\t" |
wolfSSL | 15:117db924cf7c | 6646 | "addl $16, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6647 | "cmpl %%r13d, " VAR(KR) "\n\t" |
wolfSSL | 15:117db924cf7c | 6648 | "jl 12b\n\t" |
wolfSSL | 15:117db924cf7c | 6649 | "\n" |
wolfSSL | 15:117db924cf7c | 6650 | "13:\n\t" |
wolfSSL | 15:117db924cf7c | 6651 | |
wolfSSL | 15:117db924cf7c | 6652 | AESENC_LAST15_DEC_AVX2() |
wolfSSL | 15:117db924cf7c | 6653 | "\n" |
wolfSSL | 15:117db924cf7c | 6654 | "55:\n\t" |
wolfSSL | 15:117db924cf7c | 6655 | |
wolfSSL | 15:117db924cf7c | 6656 | CALC_TAG_AVX2() |
wolfSSL | 15:117db924cf7c | 6657 | "addq $" VAR(STACK_OFFSET) ", %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 6658 | "popq %%rdx\n\t" |
wolfSSL | 15:117db924cf7c | 6659 | CMP_TAG_AVX() |
wolfSSL | 15:117db924cf7c | 6660 | "vzeroupper\n\t" |
wolfSSL | 15:117db924cf7c | 6661 | |
wolfSSL | 15:117db924cf7c | 6662 | : |
wolfSSL | 15:117db924cf7c | 6663 | : [KEY] "r" (key), |
wolfSSL | 15:117db924cf7c | 6664 | [in] "r" (in), [out] "r" (out), [nr] "r" (nr), |
wolfSSL | 15:117db924cf7c | 6665 | [nbytes] "r" (nbytes), [abytes] "r" (abytes), [addt] "r" (addt), |
wolfSSL | 15:117db924cf7c | 6666 | [ivec] "r" (iv), [ibytes] "r" (ivLen), [tbytes] "r" (tagLen), |
wolfSSL | 15:117db924cf7c | 6667 | [tag] "r" (tag), [res] "r" (res), |
wolfSSL | 15:117db924cf7c | 6668 | [BSWAP_MASK] "m" (BSWAP_MASK), |
wolfSSL | 15:117db924cf7c | 6669 | [BSWAP_EPI64] "m" (BSWAP_EPI64), |
wolfSSL | 15:117db924cf7c | 6670 | [ONE] "m" (ONE), |
wolfSSL | 15:117db924cf7c | 6671 | #if !defined(AES_GCM_AESNI_NO_UNROLL) && !defined(AES_GCM_AVX2_NO_UNROLL) |
wolfSSL | 15:117db924cf7c | 6672 | [TWO] "m" (TWO), [THREE] "m" (THREE), [FOUR] "m" (FOUR), |
wolfSSL | 15:117db924cf7c | 6673 | [FIVE] "m" (FIVE), [SIX] "m" (SIX), [SEVEN] "m" (SEVEN), |
wolfSSL | 15:117db924cf7c | 6674 | [EIGHT] "m" (EIGHT), |
wolfSSL | 15:117db924cf7c | 6675 | #endif |
wolfSSL | 15:117db924cf7c | 6676 | [MOD2_128] "m" (MOD2_128) |
wolfSSL | 15:117db924cf7c | 6677 | : "xmm15", "xmm14", "xmm13", "xmm12", |
wolfSSL | 15:117db924cf7c | 6678 | "xmm0", "xmm1", "xmm2", "xmm3", "memory", |
wolfSSL | 15:117db924cf7c | 6679 | "xmm4", "xmm5", "xmm6", "xmm7", "xmm8", "xmm9", "xmm10", "xmm11", |
wolfSSL | 15:117db924cf7c | 6680 | "rcx", "r13" |
wolfSSL | 15:117db924cf7c | 6681 | ); |
wolfSSL | 15:117db924cf7c | 6682 | } |
wolfSSL | 15:117db924cf7c | 6683 | #endif /* HAVE_INTEL_AVX2 */ |
wolfSSL | 15:117db924cf7c | 6684 | #endif /* HAVE_INTEL_AVX1 */ |
wolfSSL | 15:117db924cf7c | 6685 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 6686 | |
wolfSSL | 15:117db924cf7c | 6687 | #else /* _MSC_VER */ |
wolfSSL | 15:117db924cf7c | 6688 | /* The following are for MSC based builds which do not allow |
wolfSSL | 15:117db924cf7c | 6689 | * inline assembly. Intrinsic functions are used instead. */ |
wolfSSL | 15:117db924cf7c | 6690 | |
wolfSSL | 15:117db924cf7c | 6691 | #define aes_gcm_calc_iv_12(KEY, ivec, nr, H, Y, T) \ |
wolfSSL | 15:117db924cf7c | 6692 | do \ |
wolfSSL | 15:117db924cf7c | 6693 | { \ |
wolfSSL | 15:117db924cf7c | 6694 | word32 iv12[4]; \ |
wolfSSL | 15:117db924cf7c | 6695 | iv12[0] = *(word32*)&ivec[0]; \ |
wolfSSL | 15:117db924cf7c | 6696 | iv12[1] = *(word32*)&ivec[4]; \ |
wolfSSL | 15:117db924cf7c | 6697 | iv12[2] = *(word32*)&ivec[8]; \ |
wolfSSL | 15:117db924cf7c | 6698 | iv12[3] = 0x01000000; \ |
wolfSSL | 15:117db924cf7c | 6699 | Y = _mm_loadu_si128((__m128i*)iv12); \ |
wolfSSL | 15:117db924cf7c | 6700 | \ |
wolfSSL | 15:117db924cf7c | 6701 | /* (Compute E[ZERO, KS] and E[Y0, KS] together */ \ |
wolfSSL | 15:117db924cf7c | 6702 | tmp1 = _mm_load_si128(&KEY[0]); \ |
wolfSSL | 15:117db924cf7c | 6703 | tmp2 = _mm_xor_si128(Y, KEY[0]); \ |
wolfSSL | 15:117db924cf7c | 6704 | tmp1 = _mm_aesenc_si128(tmp1, KEY[1]); \ |
wolfSSL | 15:117db924cf7c | 6705 | tmp2 = _mm_aesenc_si128(tmp2, KEY[1]); \ |
wolfSSL | 15:117db924cf7c | 6706 | tmp1 = _mm_aesenc_si128(tmp1, KEY[2]); \ |
wolfSSL | 15:117db924cf7c | 6707 | tmp2 = _mm_aesenc_si128(tmp2, KEY[2]); \ |
wolfSSL | 15:117db924cf7c | 6708 | tmp1 = _mm_aesenc_si128(tmp1, KEY[3]); \ |
wolfSSL | 15:117db924cf7c | 6709 | tmp2 = _mm_aesenc_si128(tmp2, KEY[3]); \ |
wolfSSL | 15:117db924cf7c | 6710 | tmp1 = _mm_aesenc_si128(tmp1, KEY[4]); \ |
wolfSSL | 15:117db924cf7c | 6711 | tmp2 = _mm_aesenc_si128(tmp2, KEY[4]); \ |
wolfSSL | 15:117db924cf7c | 6712 | tmp1 = _mm_aesenc_si128(tmp1, KEY[5]); \ |
wolfSSL | 15:117db924cf7c | 6713 | tmp2 = _mm_aesenc_si128(tmp2, KEY[5]); \ |
wolfSSL | 15:117db924cf7c | 6714 | tmp1 = _mm_aesenc_si128(tmp1, KEY[6]); \ |
wolfSSL | 15:117db924cf7c | 6715 | tmp2 = _mm_aesenc_si128(tmp2, KEY[6]); \ |
wolfSSL | 15:117db924cf7c | 6716 | tmp1 = _mm_aesenc_si128(tmp1, KEY[7]); \ |
wolfSSL | 15:117db924cf7c | 6717 | tmp2 = _mm_aesenc_si128(tmp2, KEY[7]); \ |
wolfSSL | 15:117db924cf7c | 6718 | tmp1 = _mm_aesenc_si128(tmp1, KEY[8]); \ |
wolfSSL | 15:117db924cf7c | 6719 | tmp2 = _mm_aesenc_si128(tmp2, KEY[8]); \ |
wolfSSL | 15:117db924cf7c | 6720 | tmp1 = _mm_aesenc_si128(tmp1, KEY[9]); \ |
wolfSSL | 15:117db924cf7c | 6721 | tmp2 = _mm_aesenc_si128(tmp2, KEY[9]); \ |
wolfSSL | 15:117db924cf7c | 6722 | lastKey = KEY[10]; \ |
wolfSSL | 15:117db924cf7c | 6723 | if (nr > 10) { \ |
wolfSSL | 15:117db924cf7c | 6724 | tmp1 = _mm_aesenc_si128(tmp1, lastKey); \ |
wolfSSL | 15:117db924cf7c | 6725 | tmp2 = _mm_aesenc_si128(tmp2, lastKey); \ |
wolfSSL | 15:117db924cf7c | 6726 | tmp1 = _mm_aesenc_si128(tmp1, KEY[11]); \ |
wolfSSL | 15:117db924cf7c | 6727 | tmp2 = _mm_aesenc_si128(tmp2, KEY[11]); \ |
wolfSSL | 15:117db924cf7c | 6728 | lastKey = KEY[12]; \ |
wolfSSL | 15:117db924cf7c | 6729 | if (nr > 12) { \ |
wolfSSL | 15:117db924cf7c | 6730 | tmp1 = _mm_aesenc_si128(tmp1, lastKey); \ |
wolfSSL | 15:117db924cf7c | 6731 | tmp2 = _mm_aesenc_si128(tmp2, lastKey); \ |
wolfSSL | 15:117db924cf7c | 6732 | tmp1 = _mm_aesenc_si128(tmp1, KEY[13]); \ |
wolfSSL | 15:117db924cf7c | 6733 | tmp2 = _mm_aesenc_si128(tmp2, KEY[13]); \ |
wolfSSL | 15:117db924cf7c | 6734 | lastKey = KEY[14]; \ |
wolfSSL | 15:117db924cf7c | 6735 | } \ |
wolfSSL | 15:117db924cf7c | 6736 | } \ |
wolfSSL | 15:117db924cf7c | 6737 | H = _mm_aesenclast_si128(tmp1, lastKey); \ |
wolfSSL | 15:117db924cf7c | 6738 | T = _mm_aesenclast_si128(tmp2, lastKey); \ |
wolfSSL | 15:117db924cf7c | 6739 | H = _mm_shuffle_epi8(H, BSWAP_MASK); \ |
wolfSSL | 15:117db924cf7c | 6740 | } \ |
wolfSSL | 15:117db924cf7c | 6741 | while (0) |
wolfSSL | 15:117db924cf7c | 6742 | |
wolfSSL | 15:117db924cf7c | 6743 | #define aes_gcm_calc_iv(KEY, ivec, ibytes, nr, H, Y, T) \ |
wolfSSL | 15:117db924cf7c | 6744 | do \ |
wolfSSL | 15:117db924cf7c | 6745 | { \ |
wolfSSL | 15:117db924cf7c | 6746 | if (ibytes % 16) { \ |
wolfSSL | 15:117db924cf7c | 6747 | i = ibytes / 16; \ |
wolfSSL | 15:117db924cf7c | 6748 | for (j=0; j < (int)(ibytes%16); j++) \ |
wolfSSL | 15:117db924cf7c | 6749 | ((unsigned char*)&last_block)[j] = ivec[i*16+j]; \ |
wolfSSL | 15:117db924cf7c | 6750 | } \ |
wolfSSL | 15:117db924cf7c | 6751 | tmp1 = _mm_load_si128(&KEY[0]); \ |
wolfSSL | 15:117db924cf7c | 6752 | tmp1 = _mm_aesenc_si128(tmp1, KEY[1]); \ |
wolfSSL | 15:117db924cf7c | 6753 | tmp1 = _mm_aesenc_si128(tmp1, KEY[2]); \ |
wolfSSL | 15:117db924cf7c | 6754 | tmp1 = _mm_aesenc_si128(tmp1, KEY[3]); \ |
wolfSSL | 15:117db924cf7c | 6755 | tmp1 = _mm_aesenc_si128(tmp1, KEY[4]); \ |
wolfSSL | 15:117db924cf7c | 6756 | tmp1 = _mm_aesenc_si128(tmp1, KEY[5]); \ |
wolfSSL | 15:117db924cf7c | 6757 | tmp1 = _mm_aesenc_si128(tmp1, KEY[6]); \ |
wolfSSL | 15:117db924cf7c | 6758 | tmp1 = _mm_aesenc_si128(tmp1, KEY[7]); \ |
wolfSSL | 15:117db924cf7c | 6759 | tmp1 = _mm_aesenc_si128(tmp1, KEY[8]); \ |
wolfSSL | 15:117db924cf7c | 6760 | tmp1 = _mm_aesenc_si128(tmp1, KEY[9]); \ |
wolfSSL | 15:117db924cf7c | 6761 | lastKey = KEY[10]; \ |
wolfSSL | 15:117db924cf7c | 6762 | if (nr > 10) { \ |
wolfSSL | 15:117db924cf7c | 6763 | tmp1 = _mm_aesenc_si128(tmp1, lastKey); \ |
wolfSSL | 15:117db924cf7c | 6764 | tmp1 = _mm_aesenc_si128(tmp1, KEY[11]); \ |
wolfSSL | 15:117db924cf7c | 6765 | lastKey = KEY[12]; \ |
wolfSSL | 15:117db924cf7c | 6766 | if (nr > 12) { \ |
wolfSSL | 15:117db924cf7c | 6767 | tmp1 = _mm_aesenc_si128(tmp1, lastKey); \ |
wolfSSL | 15:117db924cf7c | 6768 | tmp1 = _mm_aesenc_si128(tmp1, KEY[13]); \ |
wolfSSL | 15:117db924cf7c | 6769 | lastKey = KEY[14]; \ |
wolfSSL | 15:117db924cf7c | 6770 | } \ |
wolfSSL | 15:117db924cf7c | 6771 | } \ |
wolfSSL | 15:117db924cf7c | 6772 | H = _mm_aesenclast_si128(tmp1, lastKey); \ |
wolfSSL | 15:117db924cf7c | 6773 | H = _mm_shuffle_epi8(H, BSWAP_MASK); \ |
wolfSSL | 15:117db924cf7c | 6774 | Y = _mm_setzero_si128(); \ |
wolfSSL | 15:117db924cf7c | 6775 | for (i=0; i < (int)(ibytes/16); i++) { \ |
wolfSSL | 15:117db924cf7c | 6776 | tmp1 = _mm_loadu_si128(&((__m128i*)ivec)[i]); \ |
wolfSSL | 15:117db924cf7c | 6777 | tmp1 = _mm_shuffle_epi8(tmp1, BSWAP_MASK); \ |
wolfSSL | 15:117db924cf7c | 6778 | Y = _mm_xor_si128(Y, tmp1); \ |
wolfSSL | 15:117db924cf7c | 6779 | Y = gfmul_sw(Y, H); \ |
wolfSSL | 15:117db924cf7c | 6780 | } \ |
wolfSSL | 15:117db924cf7c | 6781 | if (ibytes % 16) { \ |
wolfSSL | 15:117db924cf7c | 6782 | tmp1 = last_block; \ |
wolfSSL | 15:117db924cf7c | 6783 | tmp1 = _mm_shuffle_epi8(tmp1, BSWAP_MASK); \ |
wolfSSL | 15:117db924cf7c | 6784 | Y = _mm_xor_si128(Y, tmp1); \ |
wolfSSL | 15:117db924cf7c | 6785 | Y = gfmul_sw(Y, H); \ |
wolfSSL | 15:117db924cf7c | 6786 | } \ |
wolfSSL | 15:117db924cf7c | 6787 | tmp1 = _mm_insert_epi64(tmp1, ibytes*8, 0); \ |
wolfSSL | 15:117db924cf7c | 6788 | tmp1 = _mm_insert_epi64(tmp1, 0, 1); \ |
wolfSSL | 15:117db924cf7c | 6789 | Y = _mm_xor_si128(Y, tmp1); \ |
wolfSSL | 15:117db924cf7c | 6790 | Y = gfmul_sw(Y, H); \ |
wolfSSL | 15:117db924cf7c | 6791 | Y = _mm_shuffle_epi8(Y, BSWAP_MASK); /* Compute E(K, Y0) */ \ |
wolfSSL | 15:117db924cf7c | 6792 | tmp1 = _mm_xor_si128(Y, KEY[0]); \ |
wolfSSL | 15:117db924cf7c | 6793 | tmp1 = _mm_aesenc_si128(tmp1, KEY[1]); \ |
wolfSSL | 15:117db924cf7c | 6794 | tmp1 = _mm_aesenc_si128(tmp1, KEY[2]); \ |
wolfSSL | 15:117db924cf7c | 6795 | tmp1 = _mm_aesenc_si128(tmp1, KEY[3]); \ |
wolfSSL | 15:117db924cf7c | 6796 | tmp1 = _mm_aesenc_si128(tmp1, KEY[4]); \ |
wolfSSL | 15:117db924cf7c | 6797 | tmp1 = _mm_aesenc_si128(tmp1, KEY[5]); \ |
wolfSSL | 15:117db924cf7c | 6798 | tmp1 = _mm_aesenc_si128(tmp1, KEY[6]); \ |
wolfSSL | 15:117db924cf7c | 6799 | tmp1 = _mm_aesenc_si128(tmp1, KEY[7]); \ |
wolfSSL | 15:117db924cf7c | 6800 | tmp1 = _mm_aesenc_si128(tmp1, KEY[8]); \ |
wolfSSL | 15:117db924cf7c | 6801 | tmp1 = _mm_aesenc_si128(tmp1, KEY[9]); \ |
wolfSSL | 15:117db924cf7c | 6802 | lastKey = KEY[10]; \ |
wolfSSL | 15:117db924cf7c | 6803 | if (nr > 10) { \ |
wolfSSL | 15:117db924cf7c | 6804 | tmp1 = _mm_aesenc_si128(tmp1, lastKey); \ |
wolfSSL | 15:117db924cf7c | 6805 | tmp1 = _mm_aesenc_si128(tmp1, KEY[11]); \ |
wolfSSL | 15:117db924cf7c | 6806 | lastKey = KEY[12]; \ |
wolfSSL | 15:117db924cf7c | 6807 | if (nr > 12) { \ |
wolfSSL | 15:117db924cf7c | 6808 | tmp1 = _mm_aesenc_si128(tmp1, lastKey); \ |
wolfSSL | 15:117db924cf7c | 6809 | tmp1 = _mm_aesenc_si128(tmp1, KEY[13]); \ |
wolfSSL | 15:117db924cf7c | 6810 | lastKey = KEY[14]; \ |
wolfSSL | 15:117db924cf7c | 6811 | } \ |
wolfSSL | 15:117db924cf7c | 6812 | } \ |
wolfSSL | 15:117db924cf7c | 6813 | T = _mm_aesenclast_si128(tmp1, lastKey); \ |
wolfSSL | 15:117db924cf7c | 6814 | } \ |
wolfSSL | 15:117db924cf7c | 6815 | while (0) |
wolfSSL | 15:117db924cf7c | 6816 | |
wolfSSL | 15:117db924cf7c | 6817 | #define AES_ENC_8(j) \ |
wolfSSL | 15:117db924cf7c | 6818 | tmp1 = _mm_aesenc_si128(tmp1, KEY[j]); \ |
wolfSSL | 15:117db924cf7c | 6819 | tmp2 = _mm_aesenc_si128(tmp2, KEY[j]); \ |
wolfSSL | 15:117db924cf7c | 6820 | tmp3 = _mm_aesenc_si128(tmp3, KEY[j]); \ |
wolfSSL | 15:117db924cf7c | 6821 | tmp4 = _mm_aesenc_si128(tmp4, KEY[j]); \ |
wolfSSL | 15:117db924cf7c | 6822 | tmp5 = _mm_aesenc_si128(tmp5, KEY[j]); \ |
wolfSSL | 15:117db924cf7c | 6823 | tmp6 = _mm_aesenc_si128(tmp6, KEY[j]); \ |
wolfSSL | 15:117db924cf7c | 6824 | tmp7 = _mm_aesenc_si128(tmp7, KEY[j]); \ |
wolfSSL | 15:117db924cf7c | 6825 | tmp8 = _mm_aesenc_si128(tmp8, KEY[j]); |
wolfSSL | 15:117db924cf7c | 6826 | |
wolfSSL | 15:117db924cf7c | 6827 | #define AES_ENC_LAST_8() \ |
wolfSSL | 15:117db924cf7c | 6828 | tmp1 =_mm_aesenclast_si128(tmp1, lastKey); \ |
wolfSSL | 15:117db924cf7c | 6829 | tmp2 =_mm_aesenclast_si128(tmp2, lastKey); \ |
wolfSSL | 15:117db924cf7c | 6830 | tmp1 = _mm_xor_si128(tmp1, _mm_loadu_si128(&((__m128i*)in)[i*8+0])); \ |
wolfSSL | 15:117db924cf7c | 6831 | tmp2 = _mm_xor_si128(tmp2, _mm_loadu_si128(&((__m128i*)in)[i*8+1])); \ |
wolfSSL | 15:117db924cf7c | 6832 | _mm_storeu_si128(&((__m128i*)out)[i*8+0], tmp1); \ |
wolfSSL | 15:117db924cf7c | 6833 | _mm_storeu_si128(&((__m128i*)out)[i*8+1], tmp2); \ |
wolfSSL | 15:117db924cf7c | 6834 | tmp3 =_mm_aesenclast_si128(tmp3, lastKey); \ |
wolfSSL | 15:117db924cf7c | 6835 | tmp4 =_mm_aesenclast_si128(tmp4, lastKey); \ |
wolfSSL | 15:117db924cf7c | 6836 | tmp3 = _mm_xor_si128(tmp3, _mm_loadu_si128(&((__m128i*)in)[i*8+2])); \ |
wolfSSL | 15:117db924cf7c | 6837 | tmp4 = _mm_xor_si128(tmp4, _mm_loadu_si128(&((__m128i*)in)[i*8+3])); \ |
wolfSSL | 15:117db924cf7c | 6838 | _mm_storeu_si128(&((__m128i*)out)[i*8+2], tmp3); \ |
wolfSSL | 15:117db924cf7c | 6839 | _mm_storeu_si128(&((__m128i*)out)[i*8+3], tmp4); \ |
wolfSSL | 15:117db924cf7c | 6840 | tmp5 =_mm_aesenclast_si128(tmp5, lastKey); \ |
wolfSSL | 15:117db924cf7c | 6841 | tmp6 =_mm_aesenclast_si128(tmp6, lastKey); \ |
wolfSSL | 15:117db924cf7c | 6842 | tmp5 = _mm_xor_si128(tmp5, _mm_loadu_si128(&((__m128i*)in)[i*8+4])); \ |
wolfSSL | 15:117db924cf7c | 6843 | tmp6 = _mm_xor_si128(tmp6, _mm_loadu_si128(&((__m128i*)in)[i*8+5])); \ |
wolfSSL | 15:117db924cf7c | 6844 | _mm_storeu_si128(&((__m128i*)out)[i*8+4], tmp5); \ |
wolfSSL | 15:117db924cf7c | 6845 | _mm_storeu_si128(&((__m128i*)out)[i*8+5], tmp6); \ |
wolfSSL | 15:117db924cf7c | 6846 | tmp7 =_mm_aesenclast_si128(tmp7, lastKey); \ |
wolfSSL | 15:117db924cf7c | 6847 | tmp8 =_mm_aesenclast_si128(tmp8, lastKey); \ |
wolfSSL | 15:117db924cf7c | 6848 | tmp7 = _mm_xor_si128(tmp7, _mm_loadu_si128(&((__m128i*)in)[i*8+6])); \ |
wolfSSL | 15:117db924cf7c | 6849 | tmp8 = _mm_xor_si128(tmp8, _mm_loadu_si128(&((__m128i*)in)[i*8+7])); \ |
wolfSSL | 15:117db924cf7c | 6850 | _mm_storeu_si128(&((__m128i*)out)[i*8+6], tmp7); \ |
wolfSSL | 15:117db924cf7c | 6851 | _mm_storeu_si128(&((__m128i*)out)[i*8+7], tmp8); |
wolfSSL | 15:117db924cf7c | 6852 | |
wolfSSL | 15:117db924cf7c | 6853 | |
wolfSSL | 15:117db924cf7c | 6854 | static __m128i gfmul_sw(__m128i a, __m128i b) |
wolfSSL | 15:117db924cf7c | 6855 | { |
wolfSSL | 15:117db924cf7c | 6856 | __m128i r, t1, t2, t3, t4, t5, t6, t7; |
wolfSSL | 15:117db924cf7c | 6857 | t2 = _mm_shuffle_epi32(b, 78); |
wolfSSL | 15:117db924cf7c | 6858 | t3 = _mm_shuffle_epi32(a, 78); |
wolfSSL | 15:117db924cf7c | 6859 | t2 = _mm_xor_si128(t2, b); |
wolfSSL | 15:117db924cf7c | 6860 | t3 = _mm_xor_si128(t3, a); |
wolfSSL | 15:117db924cf7c | 6861 | t4 = _mm_clmulepi64_si128(b, a, 0x11); |
wolfSSL | 15:117db924cf7c | 6862 | t1 = _mm_clmulepi64_si128(b, a, 0x00); |
wolfSSL | 15:117db924cf7c | 6863 | t2 = _mm_clmulepi64_si128(t2, t3, 0x00); |
wolfSSL | 15:117db924cf7c | 6864 | t2 = _mm_xor_si128(t2, t1); |
wolfSSL | 15:117db924cf7c | 6865 | t2 = _mm_xor_si128(t2, t4); |
wolfSSL | 15:117db924cf7c | 6866 | t3 = _mm_slli_si128(t2, 8); |
wolfSSL | 15:117db924cf7c | 6867 | t2 = _mm_srli_si128(t2, 8); |
wolfSSL | 15:117db924cf7c | 6868 | t1 = _mm_xor_si128(t1, t3); |
wolfSSL | 15:117db924cf7c | 6869 | t4 = _mm_xor_si128(t4, t2); |
wolfSSL | 15:117db924cf7c | 6870 | |
wolfSSL | 15:117db924cf7c | 6871 | t5 = _mm_srli_epi32(t1, 31); |
wolfSSL | 15:117db924cf7c | 6872 | t6 = _mm_srli_epi32(t4, 31); |
wolfSSL | 15:117db924cf7c | 6873 | t1 = _mm_slli_epi32(t1, 1); |
wolfSSL | 15:117db924cf7c | 6874 | t4 = _mm_slli_epi32(t4, 1); |
wolfSSL | 15:117db924cf7c | 6875 | t7 = _mm_srli_si128(t5, 12); |
wolfSSL | 15:117db924cf7c | 6876 | t5 = _mm_slli_si128(t5, 4); |
wolfSSL | 15:117db924cf7c | 6877 | t6 = _mm_slli_si128(t6, 4); |
wolfSSL | 15:117db924cf7c | 6878 | t4 = _mm_or_si128(t4, t7); |
wolfSSL | 15:117db924cf7c | 6879 | t1 = _mm_or_si128(t1, t5); |
wolfSSL | 15:117db924cf7c | 6880 | t4 = _mm_or_si128(t4, t6); |
wolfSSL | 15:117db924cf7c | 6881 | |
wolfSSL | 15:117db924cf7c | 6882 | t5 = _mm_slli_epi32(t1, 31); |
wolfSSL | 15:117db924cf7c | 6883 | t6 = _mm_slli_epi32(t1, 30); |
wolfSSL | 15:117db924cf7c | 6884 | t7 = _mm_slli_epi32(t1, 25); |
wolfSSL | 15:117db924cf7c | 6885 | t5 = _mm_xor_si128(t5, t6); |
wolfSSL | 15:117db924cf7c | 6886 | t5 = _mm_xor_si128(t5, t7); |
wolfSSL | 15:117db924cf7c | 6887 | |
wolfSSL | 15:117db924cf7c | 6888 | t6 = _mm_srli_si128(t5, 4); |
wolfSSL | 15:117db924cf7c | 6889 | t5 = _mm_slli_si128(t5, 12); |
wolfSSL | 15:117db924cf7c | 6890 | t1 = _mm_xor_si128(t1, t5); |
wolfSSL | 15:117db924cf7c | 6891 | t7 = _mm_srli_epi32(t1, 1); |
wolfSSL | 15:117db924cf7c | 6892 | t3 = _mm_srli_epi32(t1, 2); |
wolfSSL | 15:117db924cf7c | 6893 | t2 = _mm_srli_epi32(t1, 7); |
wolfSSL | 15:117db924cf7c | 6894 | |
wolfSSL | 15:117db924cf7c | 6895 | t7 = _mm_xor_si128(t7, t3); |
wolfSSL | 15:117db924cf7c | 6896 | t7 = _mm_xor_si128(t7, t2); |
wolfSSL | 15:117db924cf7c | 6897 | t7 = _mm_xor_si128(t7, t6); |
wolfSSL | 15:117db924cf7c | 6898 | t7 = _mm_xor_si128(t7, t1); |
wolfSSL | 15:117db924cf7c | 6899 | r = _mm_xor_si128(t4, t7); |
wolfSSL | 15:117db924cf7c | 6900 | |
wolfSSL | 15:117db924cf7c | 6901 | return r; |
wolfSSL | 15:117db924cf7c | 6902 | } |
wolfSSL | 15:117db924cf7c | 6903 | |
wolfSSL | 15:117db924cf7c | 6904 | static void gfmul_only(__m128i a, __m128i b, __m128i* r0, __m128i* r1) |
wolfSSL | 15:117db924cf7c | 6905 | { |
wolfSSL | 15:117db924cf7c | 6906 | __m128i t1, t2, t3, t4; |
wolfSSL | 15:117db924cf7c | 6907 | |
wolfSSL | 15:117db924cf7c | 6908 | /* 128 x 128 Carryless Multiply */ |
wolfSSL | 15:117db924cf7c | 6909 | t2 = _mm_shuffle_epi32(b, 78); |
wolfSSL | 15:117db924cf7c | 6910 | t3 = _mm_shuffle_epi32(a, 78); |
wolfSSL | 15:117db924cf7c | 6911 | t2 = _mm_xor_si128(t2, b); |
wolfSSL | 15:117db924cf7c | 6912 | t3 = _mm_xor_si128(t3, a); |
wolfSSL | 15:117db924cf7c | 6913 | t4 = _mm_clmulepi64_si128(b, a, 0x11); |
wolfSSL | 15:117db924cf7c | 6914 | t1 = _mm_clmulepi64_si128(b, a, 0x00); |
wolfSSL | 15:117db924cf7c | 6915 | t2 = _mm_clmulepi64_si128(t2, t3, 0x00); |
wolfSSL | 15:117db924cf7c | 6916 | t2 = _mm_xor_si128(t2, t1); |
wolfSSL | 15:117db924cf7c | 6917 | t2 = _mm_xor_si128(t2, t4); |
wolfSSL | 15:117db924cf7c | 6918 | t3 = _mm_slli_si128(t2, 8); |
wolfSSL | 15:117db924cf7c | 6919 | t2 = _mm_srli_si128(t2, 8); |
wolfSSL | 15:117db924cf7c | 6920 | t1 = _mm_xor_si128(t1, t3); |
wolfSSL | 15:117db924cf7c | 6921 | t4 = _mm_xor_si128(t4, t2); |
wolfSSL | 15:117db924cf7c | 6922 | *r0 = _mm_xor_si128(t1, *r0); |
wolfSSL | 15:117db924cf7c | 6923 | *r1 = _mm_xor_si128(t4, *r1); |
wolfSSL | 15:117db924cf7c | 6924 | } |
wolfSSL | 15:117db924cf7c | 6925 | |
wolfSSL | 15:117db924cf7c | 6926 | static __m128i gfmul_shl1(__m128i a) |
wolfSSL | 15:117db924cf7c | 6927 | { |
wolfSSL | 15:117db924cf7c | 6928 | __m128i t1 = a, t2; |
wolfSSL | 15:117db924cf7c | 6929 | t2 = _mm_srli_epi64(t1, 63); |
wolfSSL | 15:117db924cf7c | 6930 | t1 = _mm_slli_epi64(t1, 1); |
wolfSSL | 15:117db924cf7c | 6931 | t2 = _mm_slli_si128(t2, 8); |
wolfSSL | 15:117db924cf7c | 6932 | t1 = _mm_or_si128(t1, t2); |
wolfSSL | 15:117db924cf7c | 6933 | /* if (a[1] >> 63) t1 = _mm_xor_si128(t1, MOD2_128); */ |
wolfSSL | 15:117db924cf7c | 6934 | a = _mm_shuffle_epi32(a, 0xff); |
wolfSSL | 15:117db924cf7c | 6935 | a = _mm_srai_epi32(a, 31); |
wolfSSL | 15:117db924cf7c | 6936 | a = _mm_and_si128(a, MOD2_128); |
wolfSSL | 15:117db924cf7c | 6937 | t1 = _mm_xor_si128(t1, a); |
wolfSSL | 15:117db924cf7c | 6938 | return t1; |
wolfSSL | 15:117db924cf7c | 6939 | } |
wolfSSL | 15:117db924cf7c | 6940 | |
wolfSSL | 15:117db924cf7c | 6941 | static __m128i ghash_red(__m128i r0, __m128i r1) |
wolfSSL | 15:117db924cf7c | 6942 | { |
wolfSSL | 15:117db924cf7c | 6943 | __m128i t2, t3; |
wolfSSL | 15:117db924cf7c | 6944 | __m128i t5, t6, t7; |
wolfSSL | 15:117db924cf7c | 6945 | |
wolfSSL | 15:117db924cf7c | 6946 | t5 = _mm_slli_epi32(r0, 31); |
wolfSSL | 15:117db924cf7c | 6947 | t6 = _mm_slli_epi32(r0, 30); |
wolfSSL | 15:117db924cf7c | 6948 | t7 = _mm_slli_epi32(r0, 25); |
wolfSSL | 15:117db924cf7c | 6949 | t5 = _mm_xor_si128(t5, t6); |
wolfSSL | 15:117db924cf7c | 6950 | t5 = _mm_xor_si128(t5, t7); |
wolfSSL | 15:117db924cf7c | 6951 | |
wolfSSL | 15:117db924cf7c | 6952 | t6 = _mm_srli_si128(t5, 4); |
wolfSSL | 15:117db924cf7c | 6953 | t5 = _mm_slli_si128(t5, 12); |
wolfSSL | 15:117db924cf7c | 6954 | r0 = _mm_xor_si128(r0, t5); |
wolfSSL | 15:117db924cf7c | 6955 | t7 = _mm_srli_epi32(r0, 1); |
wolfSSL | 15:117db924cf7c | 6956 | t3 = _mm_srli_epi32(r0, 2); |
wolfSSL | 15:117db924cf7c | 6957 | t2 = _mm_srli_epi32(r0, 7); |
wolfSSL | 15:117db924cf7c | 6958 | |
wolfSSL | 15:117db924cf7c | 6959 | t7 = _mm_xor_si128(t7, t3); |
wolfSSL | 15:117db924cf7c | 6960 | t7 = _mm_xor_si128(t7, t2); |
wolfSSL | 15:117db924cf7c | 6961 | t7 = _mm_xor_si128(t7, t6); |
wolfSSL | 15:117db924cf7c | 6962 | t7 = _mm_xor_si128(t7, r0); |
wolfSSL | 15:117db924cf7c | 6963 | return _mm_xor_si128(r1, t7); |
wolfSSL | 15:117db924cf7c | 6964 | } |
wolfSSL | 15:117db924cf7c | 6965 | |
wolfSSL | 15:117db924cf7c | 6966 | static __m128i gfmul_shifted(__m128i a, __m128i b) |
wolfSSL | 15:117db924cf7c | 6967 | { |
wolfSSL | 15:117db924cf7c | 6968 | __m128i t0 = _mm_setzero_si128(), t1 = _mm_setzero_si128(); |
wolfSSL | 15:117db924cf7c | 6969 | gfmul_only(a, b, &t0, &t1); |
wolfSSL | 15:117db924cf7c | 6970 | return ghash_red(t0, t1); |
wolfSSL | 15:117db924cf7c | 6971 | } |
wolfSSL | 15:117db924cf7c | 6972 | |
wolfSSL | 15:117db924cf7c | 6973 | #ifndef AES_GCM_AESNI_NO_UNROLL |
wolfSSL | 15:117db924cf7c | 6974 | static __m128i gfmul8(__m128i a1, __m128i a2, __m128i a3, __m128i a4, |
wolfSSL | 15:117db924cf7c | 6975 | __m128i a5, __m128i a6, __m128i a7, __m128i a8, |
wolfSSL | 15:117db924cf7c | 6976 | __m128i b1, __m128i b2, __m128i b3, __m128i b4, |
wolfSSL | 15:117db924cf7c | 6977 | __m128i b5, __m128i b6, __m128i b7, __m128i b8) |
wolfSSL | 15:117db924cf7c | 6978 | { |
wolfSSL | 15:117db924cf7c | 6979 | __m128i t0 = _mm_setzero_si128(), t1 = _mm_setzero_si128(); |
wolfSSL | 15:117db924cf7c | 6980 | gfmul_only(a1, b8, &t0, &t1); |
wolfSSL | 15:117db924cf7c | 6981 | gfmul_only(a2, b7, &t0, &t1); |
wolfSSL | 15:117db924cf7c | 6982 | gfmul_only(a3, b6, &t0, &t1); |
wolfSSL | 15:117db924cf7c | 6983 | gfmul_only(a4, b5, &t0, &t1); |
wolfSSL | 15:117db924cf7c | 6984 | gfmul_only(a5, b4, &t0, &t1); |
wolfSSL | 15:117db924cf7c | 6985 | gfmul_only(a6, b3, &t0, &t1); |
wolfSSL | 15:117db924cf7c | 6986 | gfmul_only(a7, b2, &t0, &t1); |
wolfSSL | 15:117db924cf7c | 6987 | gfmul_only(a8, b1, &t0, &t1); |
wolfSSL | 15:117db924cf7c | 6988 | return ghash_red(t0, t1); |
wolfSSL | 15:117db924cf7c | 6989 | } |
wolfSSL | 15:117db924cf7c | 6990 | #endif |
wolfSSL | 15:117db924cf7c | 6991 | |
wolfSSL | 15:117db924cf7c | 6992 | |
wolfSSL | 15:117db924cf7c | 6993 | static void AES_GCM_encrypt(const unsigned char *in, |
wolfSSL | 15:117db924cf7c | 6994 | unsigned char *out, |
wolfSSL | 15:117db924cf7c | 6995 | const unsigned char* addt, |
wolfSSL | 15:117db924cf7c | 6996 | const unsigned char* ivec, |
wolfSSL | 15:117db924cf7c | 6997 | unsigned char *tag, unsigned int nbytes, |
wolfSSL | 15:117db924cf7c | 6998 | unsigned int abytes, unsigned int ibytes, |
wolfSSL | 15:117db924cf7c | 6999 | unsigned int tbytes, |
wolfSSL | 15:117db924cf7c | 7000 | const unsigned char* key, int nr) |
wolfSSL | 15:117db924cf7c | 7001 | { |
wolfSSL | 15:117db924cf7c | 7002 | int i, j ,k; |
wolfSSL | 15:117db924cf7c | 7003 | __m128i ctr1; |
wolfSSL | 15:117db924cf7c | 7004 | __m128i H, Y, T; |
wolfSSL | 15:117db924cf7c | 7005 | __m128i X = _mm_setzero_si128(); |
wolfSSL | 15:117db924cf7c | 7006 | __m128i *KEY = (__m128i*)key, lastKey; |
wolfSSL | 15:117db924cf7c | 7007 | __m128i last_block = _mm_setzero_si128(); |
wolfSSL | 15:117db924cf7c | 7008 | __m128i tmp1, tmp2; |
wolfSSL | 15:117db924cf7c | 7009 | #ifndef AES_GCM_AESNI_NO_UNROLL |
wolfSSL | 15:117db924cf7c | 7010 | __m128i HT[8]; |
wolfSSL | 15:117db924cf7c | 7011 | __m128i r0, r1; |
wolfSSL | 15:117db924cf7c | 7012 | __m128i XV; |
wolfSSL | 15:117db924cf7c | 7013 | __m128i tmp3, tmp4, tmp5, tmp6, tmp7, tmp8; |
wolfSSL | 15:117db924cf7c | 7014 | #endif |
wolfSSL | 15:117db924cf7c | 7015 | |
wolfSSL | 15:117db924cf7c | 7016 | if (ibytes == 12) |
wolfSSL | 15:117db924cf7c | 7017 | aes_gcm_calc_iv_12(KEY, ivec, nr, H, Y, T); |
wolfSSL | 15:117db924cf7c | 7018 | else |
wolfSSL | 15:117db924cf7c | 7019 | aes_gcm_calc_iv(KEY, ivec, ibytes, nr, H, Y, T); |
wolfSSL | 15:117db924cf7c | 7020 | |
wolfSSL | 15:117db924cf7c | 7021 | for (i=0; i < (int)(abytes/16); i++) { |
wolfSSL | 15:117db924cf7c | 7022 | tmp1 = _mm_loadu_si128(&((__m128i*)addt)[i]); |
wolfSSL | 15:117db924cf7c | 7023 | tmp1 = _mm_shuffle_epi8(tmp1, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7024 | X = _mm_xor_si128(X, tmp1); |
wolfSSL | 15:117db924cf7c | 7025 | X = gfmul_sw(X, H); |
wolfSSL | 15:117db924cf7c | 7026 | } |
wolfSSL | 15:117db924cf7c | 7027 | if (abytes%16) { |
wolfSSL | 15:117db924cf7c | 7028 | last_block = _mm_setzero_si128(); |
wolfSSL | 15:117db924cf7c | 7029 | for (j=0; j < (int)(abytes%16); j++) |
wolfSSL | 15:117db924cf7c | 7030 | ((unsigned char*)&last_block)[j] = addt[i*16+j]; |
wolfSSL | 15:117db924cf7c | 7031 | tmp1 = last_block; |
wolfSSL | 15:117db924cf7c | 7032 | tmp1 = _mm_shuffle_epi8(tmp1, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7033 | X = _mm_xor_si128(X, tmp1); |
wolfSSL | 15:117db924cf7c | 7034 | X = gfmul_sw(X, H); |
wolfSSL | 15:117db924cf7c | 7035 | } |
wolfSSL | 15:117db924cf7c | 7036 | tmp1 = _mm_shuffle_epi8(Y, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7037 | ctr1 = _mm_add_epi32(tmp1, ONE); |
wolfSSL | 15:117db924cf7c | 7038 | H = gfmul_shl1(H); |
wolfSSL | 15:117db924cf7c | 7039 | |
wolfSSL | 15:117db924cf7c | 7040 | #ifndef AES_GCM_AESNI_NO_UNROLL |
wolfSSL | 15:117db924cf7c | 7041 | i = 0; |
wolfSSL | 15:117db924cf7c | 7042 | if (nbytes >= 16*8) { |
wolfSSL | 15:117db924cf7c | 7043 | HT[0] = H; |
wolfSSL | 15:117db924cf7c | 7044 | HT[1] = gfmul_shifted(H, H); |
wolfSSL | 15:117db924cf7c | 7045 | HT[2] = gfmul_shifted(H, HT[1]); |
wolfSSL | 15:117db924cf7c | 7046 | HT[3] = gfmul_shifted(HT[1], HT[1]); |
wolfSSL | 15:117db924cf7c | 7047 | HT[4] = gfmul_shifted(HT[1], HT[2]); |
wolfSSL | 15:117db924cf7c | 7048 | HT[5] = gfmul_shifted(HT[2], HT[2]); |
wolfSSL | 15:117db924cf7c | 7049 | HT[6] = gfmul_shifted(HT[2], HT[3]); |
wolfSSL | 15:117db924cf7c | 7050 | HT[7] = gfmul_shifted(HT[3], HT[3]); |
wolfSSL | 15:117db924cf7c | 7051 | |
wolfSSL | 15:117db924cf7c | 7052 | tmp1 = _mm_shuffle_epi8(ctr1, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7053 | tmp2 = _mm_add_epi32(ctr1, ONE); |
wolfSSL | 15:117db924cf7c | 7054 | tmp2 = _mm_shuffle_epi8(tmp2, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7055 | tmp3 = _mm_add_epi32(ctr1, TWO); |
wolfSSL | 15:117db924cf7c | 7056 | tmp3 = _mm_shuffle_epi8(tmp3, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7057 | tmp4 = _mm_add_epi32(ctr1, THREE); |
wolfSSL | 15:117db924cf7c | 7058 | tmp4 = _mm_shuffle_epi8(tmp4, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7059 | tmp5 = _mm_add_epi32(ctr1, FOUR); |
wolfSSL | 15:117db924cf7c | 7060 | tmp5 = _mm_shuffle_epi8(tmp5, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7061 | tmp6 = _mm_add_epi32(ctr1, FIVE); |
wolfSSL | 15:117db924cf7c | 7062 | tmp6 = _mm_shuffle_epi8(tmp6, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7063 | tmp7 = _mm_add_epi32(ctr1, SIX); |
wolfSSL | 15:117db924cf7c | 7064 | tmp7 = _mm_shuffle_epi8(tmp7, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7065 | tmp8 = _mm_add_epi32(ctr1, SEVEN); |
wolfSSL | 15:117db924cf7c | 7066 | tmp8 = _mm_shuffle_epi8(tmp8, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7067 | ctr1 = _mm_add_epi32(ctr1, EIGHT); |
wolfSSL | 15:117db924cf7c | 7068 | tmp1 =_mm_xor_si128(tmp1, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7069 | tmp2 =_mm_xor_si128(tmp2, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7070 | tmp3 =_mm_xor_si128(tmp3, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7071 | tmp4 =_mm_xor_si128(tmp4, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7072 | tmp5 =_mm_xor_si128(tmp5, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7073 | tmp6 =_mm_xor_si128(tmp6, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7074 | tmp7 =_mm_xor_si128(tmp7, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7075 | tmp8 =_mm_xor_si128(tmp8, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7076 | AES_ENC_8(1); |
wolfSSL | 15:117db924cf7c | 7077 | AES_ENC_8(2); |
wolfSSL | 15:117db924cf7c | 7078 | AES_ENC_8(3); |
wolfSSL | 15:117db924cf7c | 7079 | AES_ENC_8(4); |
wolfSSL | 15:117db924cf7c | 7080 | AES_ENC_8(5); |
wolfSSL | 15:117db924cf7c | 7081 | AES_ENC_8(6); |
wolfSSL | 15:117db924cf7c | 7082 | AES_ENC_8(7); |
wolfSSL | 15:117db924cf7c | 7083 | AES_ENC_8(8); |
wolfSSL | 15:117db924cf7c | 7084 | AES_ENC_8(9); |
wolfSSL | 15:117db924cf7c | 7085 | lastKey = KEY[10]; |
wolfSSL | 15:117db924cf7c | 7086 | if (nr > 10) { |
wolfSSL | 15:117db924cf7c | 7087 | AES_ENC_8(10); |
wolfSSL | 15:117db924cf7c | 7088 | AES_ENC_8(11); |
wolfSSL | 15:117db924cf7c | 7089 | lastKey = KEY[12]; |
wolfSSL | 15:117db924cf7c | 7090 | if (nr > 12) { |
wolfSSL | 15:117db924cf7c | 7091 | AES_ENC_8(12); |
wolfSSL | 15:117db924cf7c | 7092 | AES_ENC_8(13); |
wolfSSL | 15:117db924cf7c | 7093 | lastKey = KEY[14]; |
wolfSSL | 15:117db924cf7c | 7094 | } |
wolfSSL | 15:117db924cf7c | 7095 | } |
wolfSSL | 15:117db924cf7c | 7096 | AES_ENC_LAST_8(); |
wolfSSL | 15:117db924cf7c | 7097 | |
wolfSSL | 15:117db924cf7c | 7098 | for (i=1; i < (int)(nbytes/16/8); i++) { |
wolfSSL | 15:117db924cf7c | 7099 | r0 = _mm_setzero_si128(); |
wolfSSL | 15:117db924cf7c | 7100 | r1 = _mm_setzero_si128(); |
wolfSSL | 15:117db924cf7c | 7101 | tmp1 = _mm_shuffle_epi8(ctr1, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7102 | tmp2 = _mm_add_epi32(ctr1, ONE); |
wolfSSL | 15:117db924cf7c | 7103 | tmp2 = _mm_shuffle_epi8(tmp2, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7104 | tmp3 = _mm_add_epi32(ctr1, TWO); |
wolfSSL | 15:117db924cf7c | 7105 | tmp3 = _mm_shuffle_epi8(tmp3, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7106 | tmp4 = _mm_add_epi32(ctr1, THREE); |
wolfSSL | 15:117db924cf7c | 7107 | tmp4 = _mm_shuffle_epi8(tmp4, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7108 | tmp5 = _mm_add_epi32(ctr1, FOUR); |
wolfSSL | 15:117db924cf7c | 7109 | tmp5 = _mm_shuffle_epi8(tmp5, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7110 | tmp6 = _mm_add_epi32(ctr1, FIVE); |
wolfSSL | 15:117db924cf7c | 7111 | tmp6 = _mm_shuffle_epi8(tmp6, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7112 | tmp7 = _mm_add_epi32(ctr1, SIX); |
wolfSSL | 15:117db924cf7c | 7113 | tmp7 = _mm_shuffle_epi8(tmp7, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7114 | tmp8 = _mm_add_epi32(ctr1, SEVEN); |
wolfSSL | 15:117db924cf7c | 7115 | tmp8 = _mm_shuffle_epi8(tmp8, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7116 | ctr1 = _mm_add_epi32(ctr1, EIGHT); |
wolfSSL | 15:117db924cf7c | 7117 | tmp1 =_mm_xor_si128(tmp1, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7118 | tmp2 =_mm_xor_si128(tmp2, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7119 | tmp3 =_mm_xor_si128(tmp3, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7120 | tmp4 =_mm_xor_si128(tmp4, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7121 | tmp5 =_mm_xor_si128(tmp5, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7122 | tmp6 =_mm_xor_si128(tmp6, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7123 | tmp7 =_mm_xor_si128(tmp7, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7124 | tmp8 =_mm_xor_si128(tmp8, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7125 | /* 128 x 128 Carryless Multiply */ |
wolfSSL | 15:117db924cf7c | 7126 | XV = _mm_loadu_si128(&((__m128i*)out)[(i-1)*8+0]); |
wolfSSL | 15:117db924cf7c | 7127 | XV = _mm_shuffle_epi8(XV, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7128 | XV = _mm_xor_si128(XV, X); |
wolfSSL | 15:117db924cf7c | 7129 | gfmul_only(XV, HT[7], &r0, &r1); |
wolfSSL | 15:117db924cf7c | 7130 | tmp1 = _mm_aesenc_si128(tmp1, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7131 | tmp2 = _mm_aesenc_si128(tmp2, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7132 | tmp3 = _mm_aesenc_si128(tmp3, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7133 | tmp4 = _mm_aesenc_si128(tmp4, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7134 | tmp5 = _mm_aesenc_si128(tmp5, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7135 | tmp6 = _mm_aesenc_si128(tmp6, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7136 | tmp7 = _mm_aesenc_si128(tmp7, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7137 | tmp8 = _mm_aesenc_si128(tmp8, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7138 | /* 128 x 128 Carryless Multiply */ |
wolfSSL | 15:117db924cf7c | 7139 | XV = _mm_loadu_si128(&((__m128i*)out)[(i-1)*8+1]); |
wolfSSL | 15:117db924cf7c | 7140 | XV = _mm_shuffle_epi8(XV, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7141 | gfmul_only(XV, HT[6], &r0, &r1); |
wolfSSL | 15:117db924cf7c | 7142 | tmp1 = _mm_aesenc_si128(tmp1, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7143 | tmp2 = _mm_aesenc_si128(tmp2, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7144 | tmp3 = _mm_aesenc_si128(tmp3, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7145 | tmp4 = _mm_aesenc_si128(tmp4, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7146 | tmp5 = _mm_aesenc_si128(tmp5, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7147 | tmp6 = _mm_aesenc_si128(tmp6, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7148 | tmp7 = _mm_aesenc_si128(tmp7, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7149 | tmp8 = _mm_aesenc_si128(tmp8, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7150 | /* 128 x 128 Carryless Multiply */ |
wolfSSL | 15:117db924cf7c | 7151 | XV = _mm_loadu_si128(&((__m128i*)out)[(i-1)*8+2]); |
wolfSSL | 15:117db924cf7c | 7152 | XV = _mm_shuffle_epi8(XV, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7153 | gfmul_only(XV, HT[5], &r0, &r1); |
wolfSSL | 15:117db924cf7c | 7154 | tmp1 = _mm_aesenc_si128(tmp1, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7155 | tmp2 = _mm_aesenc_si128(tmp2, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7156 | tmp3 = _mm_aesenc_si128(tmp3, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7157 | tmp4 = _mm_aesenc_si128(tmp4, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7158 | tmp5 = _mm_aesenc_si128(tmp5, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7159 | tmp6 = _mm_aesenc_si128(tmp6, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7160 | tmp7 = _mm_aesenc_si128(tmp7, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7161 | tmp8 = _mm_aesenc_si128(tmp8, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7162 | /* 128 x 128 Carryless Multiply */ |
wolfSSL | 15:117db924cf7c | 7163 | XV = _mm_loadu_si128(&((__m128i*)out)[(i-1)*8+3]); |
wolfSSL | 15:117db924cf7c | 7164 | XV = _mm_shuffle_epi8(XV, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7165 | gfmul_only(XV, HT[4], &r0, &r1); |
wolfSSL | 15:117db924cf7c | 7166 | tmp1 = _mm_aesenc_si128(tmp1, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7167 | tmp2 = _mm_aesenc_si128(tmp2, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7168 | tmp3 = _mm_aesenc_si128(tmp3, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7169 | tmp4 = _mm_aesenc_si128(tmp4, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7170 | tmp5 = _mm_aesenc_si128(tmp5, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7171 | tmp6 = _mm_aesenc_si128(tmp6, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7172 | tmp7 = _mm_aesenc_si128(tmp7, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7173 | tmp8 = _mm_aesenc_si128(tmp8, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7174 | /* 128 x 128 Carryless Multiply */ |
wolfSSL | 15:117db924cf7c | 7175 | XV = _mm_loadu_si128(&((__m128i*)out)[(i-1)*8+4]); |
wolfSSL | 15:117db924cf7c | 7176 | XV = _mm_shuffle_epi8(XV, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7177 | gfmul_only(XV, HT[3], &r0, &r1); |
wolfSSL | 15:117db924cf7c | 7178 | tmp1 = _mm_aesenc_si128(tmp1, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7179 | tmp2 = _mm_aesenc_si128(tmp2, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7180 | tmp3 = _mm_aesenc_si128(tmp3, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7181 | tmp4 = _mm_aesenc_si128(tmp4, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7182 | tmp5 = _mm_aesenc_si128(tmp5, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7183 | tmp6 = _mm_aesenc_si128(tmp6, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7184 | tmp7 = _mm_aesenc_si128(tmp7, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7185 | tmp8 = _mm_aesenc_si128(tmp8, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7186 | /* 128 x 128 Carryless Multiply */ |
wolfSSL | 15:117db924cf7c | 7187 | XV = _mm_loadu_si128(&((__m128i*)out)[(i-1)*8+5]); |
wolfSSL | 15:117db924cf7c | 7188 | XV = _mm_shuffle_epi8(XV, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7189 | gfmul_only(XV, HT[2], &r0, &r1); |
wolfSSL | 15:117db924cf7c | 7190 | tmp1 = _mm_aesenc_si128(tmp1, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7191 | tmp2 = _mm_aesenc_si128(tmp2, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7192 | tmp3 = _mm_aesenc_si128(tmp3, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7193 | tmp4 = _mm_aesenc_si128(tmp4, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7194 | tmp5 = _mm_aesenc_si128(tmp5, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7195 | tmp6 = _mm_aesenc_si128(tmp6, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7196 | tmp7 = _mm_aesenc_si128(tmp7, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7197 | tmp8 = _mm_aesenc_si128(tmp8, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7198 | /* 128 x 128 Carryless Multiply */ |
wolfSSL | 15:117db924cf7c | 7199 | XV = _mm_loadu_si128(&((__m128i*)out)[(i-1)*8+6]); |
wolfSSL | 15:117db924cf7c | 7200 | XV = _mm_shuffle_epi8(XV, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7201 | gfmul_only(XV, HT[1], &r0, &r1); |
wolfSSL | 15:117db924cf7c | 7202 | tmp1 = _mm_aesenc_si128(tmp1, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7203 | tmp2 = _mm_aesenc_si128(tmp2, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7204 | tmp3 = _mm_aesenc_si128(tmp3, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7205 | tmp4 = _mm_aesenc_si128(tmp4, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7206 | tmp5 = _mm_aesenc_si128(tmp5, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7207 | tmp6 = _mm_aesenc_si128(tmp6, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7208 | tmp7 = _mm_aesenc_si128(tmp7, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7209 | tmp8 = _mm_aesenc_si128(tmp8, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7210 | /* 128 x 128 Carryless Multiply */ |
wolfSSL | 15:117db924cf7c | 7211 | XV = _mm_loadu_si128(&((__m128i*)out)[(i-1)*8+7]); |
wolfSSL | 15:117db924cf7c | 7212 | XV = _mm_shuffle_epi8(XV, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7213 | gfmul_only(XV, HT[0], &r0, &r1); |
wolfSSL | 15:117db924cf7c | 7214 | tmp1 = _mm_aesenc_si128(tmp1, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7215 | tmp2 = _mm_aesenc_si128(tmp2, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7216 | tmp3 = _mm_aesenc_si128(tmp3, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7217 | tmp4 = _mm_aesenc_si128(tmp4, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7218 | tmp5 = _mm_aesenc_si128(tmp5, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7219 | tmp6 = _mm_aesenc_si128(tmp6, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7220 | tmp7 = _mm_aesenc_si128(tmp7, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7221 | tmp8 = _mm_aesenc_si128(tmp8, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7222 | /* Reduction */ |
wolfSSL | 15:117db924cf7c | 7223 | X = ghash_red(r0, r1); |
wolfSSL | 15:117db924cf7c | 7224 | tmp1 = _mm_aesenc_si128(tmp1, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7225 | tmp2 = _mm_aesenc_si128(tmp2, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7226 | tmp3 = _mm_aesenc_si128(tmp3, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7227 | tmp4 = _mm_aesenc_si128(tmp4, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7228 | tmp5 = _mm_aesenc_si128(tmp5, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7229 | tmp6 = _mm_aesenc_si128(tmp6, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7230 | tmp7 = _mm_aesenc_si128(tmp7, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7231 | tmp8 = _mm_aesenc_si128(tmp8, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7232 | lastKey = KEY[10]; |
wolfSSL | 15:117db924cf7c | 7233 | if (nr > 10) { |
wolfSSL | 15:117db924cf7c | 7234 | tmp1 = _mm_aesenc_si128(tmp1, KEY[10]); |
wolfSSL | 15:117db924cf7c | 7235 | tmp2 = _mm_aesenc_si128(tmp2, KEY[10]); |
wolfSSL | 15:117db924cf7c | 7236 | tmp3 = _mm_aesenc_si128(tmp3, KEY[10]); |
wolfSSL | 15:117db924cf7c | 7237 | tmp4 = _mm_aesenc_si128(tmp4, KEY[10]); |
wolfSSL | 15:117db924cf7c | 7238 | tmp5 = _mm_aesenc_si128(tmp5, KEY[10]); |
wolfSSL | 15:117db924cf7c | 7239 | tmp6 = _mm_aesenc_si128(tmp6, KEY[10]); |
wolfSSL | 15:117db924cf7c | 7240 | tmp7 = _mm_aesenc_si128(tmp7, KEY[10]); |
wolfSSL | 15:117db924cf7c | 7241 | tmp8 = _mm_aesenc_si128(tmp8, KEY[10]); |
wolfSSL | 15:117db924cf7c | 7242 | tmp1 = _mm_aesenc_si128(tmp1, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7243 | tmp2 = _mm_aesenc_si128(tmp2, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7244 | tmp3 = _mm_aesenc_si128(tmp3, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7245 | tmp4 = _mm_aesenc_si128(tmp4, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7246 | tmp5 = _mm_aesenc_si128(tmp5, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7247 | tmp6 = _mm_aesenc_si128(tmp6, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7248 | tmp7 = _mm_aesenc_si128(tmp7, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7249 | tmp8 = _mm_aesenc_si128(tmp8, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7250 | lastKey = KEY[12]; |
wolfSSL | 15:117db924cf7c | 7251 | if (nr > 12) { |
wolfSSL | 15:117db924cf7c | 7252 | tmp1 = _mm_aesenc_si128(tmp1, KEY[12]); |
wolfSSL | 15:117db924cf7c | 7253 | tmp2 = _mm_aesenc_si128(tmp2, KEY[12]); |
wolfSSL | 15:117db924cf7c | 7254 | tmp3 = _mm_aesenc_si128(tmp3, KEY[12]); |
wolfSSL | 15:117db924cf7c | 7255 | tmp4 = _mm_aesenc_si128(tmp4, KEY[12]); |
wolfSSL | 15:117db924cf7c | 7256 | tmp5 = _mm_aesenc_si128(tmp5, KEY[12]); |
wolfSSL | 15:117db924cf7c | 7257 | tmp6 = _mm_aesenc_si128(tmp6, KEY[12]); |
wolfSSL | 15:117db924cf7c | 7258 | tmp7 = _mm_aesenc_si128(tmp7, KEY[12]); |
wolfSSL | 15:117db924cf7c | 7259 | tmp8 = _mm_aesenc_si128(tmp8, KEY[12]); |
wolfSSL | 15:117db924cf7c | 7260 | tmp1 = _mm_aesenc_si128(tmp1, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7261 | tmp2 = _mm_aesenc_si128(tmp2, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7262 | tmp3 = _mm_aesenc_si128(tmp3, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7263 | tmp4 = _mm_aesenc_si128(tmp4, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7264 | tmp5 = _mm_aesenc_si128(tmp5, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7265 | tmp6 = _mm_aesenc_si128(tmp6, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7266 | tmp7 = _mm_aesenc_si128(tmp7, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7267 | tmp8 = _mm_aesenc_si128(tmp8, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7268 | lastKey = KEY[14]; |
wolfSSL | 15:117db924cf7c | 7269 | } |
wolfSSL | 15:117db924cf7c | 7270 | } |
wolfSSL | 15:117db924cf7c | 7271 | AES_ENC_LAST_8(); |
wolfSSL | 15:117db924cf7c | 7272 | } |
wolfSSL | 15:117db924cf7c | 7273 | |
wolfSSL | 15:117db924cf7c | 7274 | tmp1 = _mm_shuffle_epi8(tmp1, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7275 | tmp2 = _mm_shuffle_epi8(tmp2, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7276 | tmp3 = _mm_shuffle_epi8(tmp3, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7277 | tmp4 = _mm_shuffle_epi8(tmp4, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7278 | tmp5 = _mm_shuffle_epi8(tmp5, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7279 | tmp6 = _mm_shuffle_epi8(tmp6, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7280 | tmp7 = _mm_shuffle_epi8(tmp7, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7281 | tmp8 = _mm_shuffle_epi8(tmp8, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7282 | tmp1 = _mm_xor_si128(X, tmp1); |
wolfSSL | 15:117db924cf7c | 7283 | X = gfmul8(tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8, |
wolfSSL | 15:117db924cf7c | 7284 | HT[0], HT[1], HT[2], HT[3], HT[4], HT[5], HT[6], HT[7]); |
wolfSSL | 15:117db924cf7c | 7285 | } |
wolfSSL | 15:117db924cf7c | 7286 | for (k = i*8; k < (int)(nbytes/16); k++) { |
wolfSSL | 15:117db924cf7c | 7287 | tmp1 = _mm_shuffle_epi8(ctr1, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7288 | ctr1 = _mm_add_epi32(ctr1, ONE); |
wolfSSL | 15:117db924cf7c | 7289 | tmp1 = _mm_xor_si128(tmp1, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7290 | tmp1 = _mm_aesenc_si128(tmp1, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7291 | tmp1 = _mm_aesenc_si128(tmp1, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7292 | tmp1 = _mm_aesenc_si128(tmp1, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7293 | tmp1 = _mm_aesenc_si128(tmp1, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7294 | tmp1 = _mm_aesenc_si128(tmp1, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7295 | tmp1 = _mm_aesenc_si128(tmp1, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7296 | tmp1 = _mm_aesenc_si128(tmp1, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7297 | tmp1 = _mm_aesenc_si128(tmp1, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7298 | tmp1 = _mm_aesenc_si128(tmp1, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7299 | lastKey = KEY[10]; |
wolfSSL | 15:117db924cf7c | 7300 | if (nr > 10) { |
wolfSSL | 15:117db924cf7c | 7301 | tmp1 = _mm_aesenc_si128(tmp1, lastKey); |
wolfSSL | 15:117db924cf7c | 7302 | tmp1 = _mm_aesenc_si128(tmp1, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7303 | lastKey = KEY[12]; |
wolfSSL | 15:117db924cf7c | 7304 | if (nr > 12) { |
wolfSSL | 15:117db924cf7c | 7305 | tmp1 = _mm_aesenc_si128(tmp1, lastKey); |
wolfSSL | 15:117db924cf7c | 7306 | tmp1 = _mm_aesenc_si128(tmp1, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7307 | lastKey = KEY[14]; |
wolfSSL | 15:117db924cf7c | 7308 | } |
wolfSSL | 15:117db924cf7c | 7309 | } |
wolfSSL | 15:117db924cf7c | 7310 | tmp1 = _mm_aesenclast_si128(tmp1, lastKey); |
wolfSSL | 15:117db924cf7c | 7311 | tmp1 = _mm_xor_si128(tmp1, _mm_loadu_si128(&((__m128i*)in)[k])); |
wolfSSL | 15:117db924cf7c | 7312 | _mm_storeu_si128(&((__m128i*)out)[k], tmp1); |
wolfSSL | 15:117db924cf7c | 7313 | tmp1 = _mm_shuffle_epi8(tmp1, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7314 | X =_mm_xor_si128(X, tmp1); |
wolfSSL | 15:117db924cf7c | 7315 | X = gfmul_shifted(X, H); |
wolfSSL | 15:117db924cf7c | 7316 | } |
wolfSSL | 15:117db924cf7c | 7317 | #else /* AES_GCM_AESNI_NO_UNROLL */ |
wolfSSL | 15:117db924cf7c | 7318 | for (k = 0; k < (int)(nbytes/16) && k < 1; k++) { |
wolfSSL | 15:117db924cf7c | 7319 | tmp1 = _mm_shuffle_epi8(ctr1, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7320 | ctr1 = _mm_add_epi32(ctr1, ONE); |
wolfSSL | 15:117db924cf7c | 7321 | tmp1 = _mm_xor_si128(tmp1, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7322 | tmp1 = _mm_aesenc_si128(tmp1, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7323 | tmp1 = _mm_aesenc_si128(tmp1, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7324 | tmp1 = _mm_aesenc_si128(tmp1, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7325 | tmp1 = _mm_aesenc_si128(tmp1, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7326 | tmp1 = _mm_aesenc_si128(tmp1, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7327 | tmp1 = _mm_aesenc_si128(tmp1, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7328 | tmp1 = _mm_aesenc_si128(tmp1, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7329 | tmp1 = _mm_aesenc_si128(tmp1, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7330 | tmp1 = _mm_aesenc_si128(tmp1, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7331 | lastKey = KEY[10]; |
wolfSSL | 15:117db924cf7c | 7332 | if (nr > 10) { |
wolfSSL | 15:117db924cf7c | 7333 | tmp1 = _mm_aesenc_si128(tmp1, lastKey); |
wolfSSL | 15:117db924cf7c | 7334 | tmp1 = _mm_aesenc_si128(tmp1, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7335 | lastKey = KEY[12]; |
wolfSSL | 15:117db924cf7c | 7336 | if (nr > 12) { |
wolfSSL | 15:117db924cf7c | 7337 | tmp1 = _mm_aesenc_si128(tmp1, lastKey); |
wolfSSL | 15:117db924cf7c | 7338 | tmp1 = _mm_aesenc_si128(tmp1, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7339 | lastKey = KEY[14]; |
wolfSSL | 15:117db924cf7c | 7340 | } |
wolfSSL | 15:117db924cf7c | 7341 | } |
wolfSSL | 15:117db924cf7c | 7342 | tmp1 = _mm_aesenclast_si128(tmp1, lastKey); |
wolfSSL | 15:117db924cf7c | 7343 | tmp1 = _mm_xor_si128(tmp1, _mm_loadu_si128(&((__m128i*)in)[k])); |
wolfSSL | 15:117db924cf7c | 7344 | _mm_storeu_si128(&((__m128i*)out)[k], tmp1); |
wolfSSL | 15:117db924cf7c | 7345 | tmp1 = _mm_shuffle_epi8(tmp1, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7346 | X =_mm_xor_si128(X, tmp1); |
wolfSSL | 15:117db924cf7c | 7347 | } |
wolfSSL | 15:117db924cf7c | 7348 | for (; k < (int)(nbytes/16); k++) { |
wolfSSL | 15:117db924cf7c | 7349 | tmp1 = _mm_shuffle_epi8(ctr1, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7350 | ctr1 = _mm_add_epi32(ctr1, ONE); |
wolfSSL | 15:117db924cf7c | 7351 | tmp1 = _mm_xor_si128(tmp1, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7352 | tmp1 = _mm_aesenc_si128(tmp1, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7353 | tmp1 = _mm_aesenc_si128(tmp1, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7354 | tmp1 = _mm_aesenc_si128(tmp1, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7355 | tmp1 = _mm_aesenc_si128(tmp1, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7356 | tmp1 = _mm_aesenc_si128(tmp1, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7357 | tmp1 = _mm_aesenc_si128(tmp1, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7358 | tmp1 = _mm_aesenc_si128(tmp1, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7359 | tmp1 = _mm_aesenc_si128(tmp1, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7360 | tmp1 = _mm_aesenc_si128(tmp1, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7361 | X = gfmul_shifted(X, H); |
wolfSSL | 15:117db924cf7c | 7362 | lastKey = KEY[10]; |
wolfSSL | 15:117db924cf7c | 7363 | if (nr > 10) { |
wolfSSL | 15:117db924cf7c | 7364 | tmp1 = _mm_aesenc_si128(tmp1, lastKey); |
wolfSSL | 15:117db924cf7c | 7365 | tmp1 = _mm_aesenc_si128(tmp1, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7366 | lastKey = KEY[12]; |
wolfSSL | 15:117db924cf7c | 7367 | if (nr > 12) { |
wolfSSL | 15:117db924cf7c | 7368 | tmp1 = _mm_aesenc_si128(tmp1, lastKey); |
wolfSSL | 15:117db924cf7c | 7369 | tmp1 = _mm_aesenc_si128(tmp1, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7370 | lastKey = KEY[14]; |
wolfSSL | 15:117db924cf7c | 7371 | } |
wolfSSL | 15:117db924cf7c | 7372 | } |
wolfSSL | 15:117db924cf7c | 7373 | tmp1 = _mm_aesenclast_si128(tmp1, lastKey); |
wolfSSL | 15:117db924cf7c | 7374 | tmp1 = _mm_xor_si128(tmp1, _mm_loadu_si128(&((__m128i*)in)[k])); |
wolfSSL | 15:117db924cf7c | 7375 | _mm_storeu_si128(&((__m128i*)out)[k], tmp1); |
wolfSSL | 15:117db924cf7c | 7376 | tmp1 = _mm_shuffle_epi8(tmp1, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7377 | X =_mm_xor_si128(X, tmp1); |
wolfSSL | 15:117db924cf7c | 7378 | } |
wolfSSL | 15:117db924cf7c | 7379 | if (k > 0) { |
wolfSSL | 15:117db924cf7c | 7380 | X = gfmul_shifted(X, H); |
wolfSSL | 15:117db924cf7c | 7381 | } |
wolfSSL | 15:117db924cf7c | 7382 | #endif /* AES_GCM_AESNI_NO_UNROLL */ |
wolfSSL | 15:117db924cf7c | 7383 | |
wolfSSL | 15:117db924cf7c | 7384 | /* If one partial block remains */ |
wolfSSL | 15:117db924cf7c | 7385 | if (nbytes % 16) { |
wolfSSL | 15:117db924cf7c | 7386 | tmp1 = _mm_shuffle_epi8(ctr1, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7387 | tmp1 = _mm_xor_si128(tmp1, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7388 | tmp1 = _mm_aesenc_si128(tmp1, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7389 | tmp1 = _mm_aesenc_si128(tmp1, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7390 | tmp1 = _mm_aesenc_si128(tmp1, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7391 | tmp1 = _mm_aesenc_si128(tmp1, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7392 | tmp1 = _mm_aesenc_si128(tmp1, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7393 | tmp1 = _mm_aesenc_si128(tmp1, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7394 | tmp1 = _mm_aesenc_si128(tmp1, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7395 | tmp1 = _mm_aesenc_si128(tmp1, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7396 | tmp1 = _mm_aesenc_si128(tmp1, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7397 | lastKey = KEY[10]; |
wolfSSL | 15:117db924cf7c | 7398 | if (nr > 10) { |
wolfSSL | 15:117db924cf7c | 7399 | tmp1 = _mm_aesenc_si128(tmp1, lastKey); |
wolfSSL | 15:117db924cf7c | 7400 | tmp1 = _mm_aesenc_si128(tmp1, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7401 | lastKey = KEY[12]; |
wolfSSL | 15:117db924cf7c | 7402 | if (nr > 12) { |
wolfSSL | 15:117db924cf7c | 7403 | tmp1 = _mm_aesenc_si128(tmp1, lastKey); |
wolfSSL | 15:117db924cf7c | 7404 | tmp1 = _mm_aesenc_si128(tmp1, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7405 | lastKey = KEY[14]; |
wolfSSL | 15:117db924cf7c | 7406 | } |
wolfSSL | 15:117db924cf7c | 7407 | } |
wolfSSL | 15:117db924cf7c | 7408 | tmp1 = _mm_aesenclast_si128(tmp1, lastKey); |
wolfSSL | 15:117db924cf7c | 7409 | last_block = tmp1; |
wolfSSL | 15:117db924cf7c | 7410 | for (j=0; j < (int)(nbytes%16); j++) |
wolfSSL | 15:117db924cf7c | 7411 | ((unsigned char*)&last_block)[j] = in[k*16+j]; |
wolfSSL | 15:117db924cf7c | 7412 | tmp1 = _mm_xor_si128(tmp1, last_block); |
wolfSSL | 15:117db924cf7c | 7413 | last_block = tmp1; |
wolfSSL | 15:117db924cf7c | 7414 | for (j=0; j < (int)(nbytes%16); j++) |
wolfSSL | 15:117db924cf7c | 7415 | out[k*16+j] = ((unsigned char*)&last_block)[j]; |
wolfSSL | 15:117db924cf7c | 7416 | tmp1 = last_block; |
wolfSSL | 15:117db924cf7c | 7417 | tmp1 = _mm_shuffle_epi8(tmp1, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7418 | X =_mm_xor_si128(X, tmp1); |
wolfSSL | 15:117db924cf7c | 7419 | X = gfmul_shifted(X, H); |
wolfSSL | 15:117db924cf7c | 7420 | } |
wolfSSL | 15:117db924cf7c | 7421 | tmp1 = _mm_insert_epi64(tmp1, nbytes*8, 0); |
wolfSSL | 15:117db924cf7c | 7422 | tmp1 = _mm_insert_epi64(tmp1, abytes*8, 1); |
wolfSSL | 15:117db924cf7c | 7423 | X = _mm_xor_si128(X, tmp1); |
wolfSSL | 15:117db924cf7c | 7424 | X = gfmul_shifted(X, H); |
wolfSSL | 15:117db924cf7c | 7425 | X = _mm_shuffle_epi8(X, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7426 | T = _mm_xor_si128(X, T); |
wolfSSL | 15:117db924cf7c | 7427 | /*_mm_storeu_si128((__m128i*)tag, T);*/ |
wolfSSL | 15:117db924cf7c | 7428 | XMEMCPY(tag, &T, tbytes); |
wolfSSL | 15:117db924cf7c | 7429 | } |
wolfSSL | 15:117db924cf7c | 7430 | |
wolfSSL | 15:117db924cf7c | 7431 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 7432 | |
wolfSSL | 15:117db924cf7c | 7433 | static void AES_GCM_decrypt(const unsigned char *in, |
wolfSSL | 15:117db924cf7c | 7434 | unsigned char *out, |
wolfSSL | 15:117db924cf7c | 7435 | const unsigned char* addt, |
wolfSSL | 15:117db924cf7c | 7436 | const unsigned char* ivec, |
wolfSSL | 15:117db924cf7c | 7437 | const unsigned char *tag, int nbytes, int abytes, |
wolfSSL | 15:117db924cf7c | 7438 | int ibytes, word32 tbytes, const unsigned char* key, |
wolfSSL | 15:117db924cf7c | 7439 | int nr, int* res) |
wolfSSL | 15:117db924cf7c | 7440 | { |
wolfSSL | 15:117db924cf7c | 7441 | int i, j ,k; |
wolfSSL | 15:117db924cf7c | 7442 | __m128i H, Y, T; |
wolfSSL | 15:117db924cf7c | 7443 | __m128i *KEY = (__m128i*)key, lastKey; |
wolfSSL | 15:117db924cf7c | 7444 | __m128i ctr1; |
wolfSSL | 15:117db924cf7c | 7445 | __m128i last_block = _mm_setzero_si128(); |
wolfSSL | 15:117db924cf7c | 7446 | __m128i X = _mm_setzero_si128(); |
wolfSSL | 15:117db924cf7c | 7447 | __m128i tmp1, tmp2, XV; |
wolfSSL | 15:117db924cf7c | 7448 | #ifndef AES_GCM_AESNI_NO_UNROLL |
wolfSSL | 15:117db924cf7c | 7449 | __m128i HT[8]; |
wolfSSL | 15:117db924cf7c | 7450 | __m128i r0, r1; |
wolfSSL | 15:117db924cf7c | 7451 | __m128i tmp3, tmp4, tmp5, tmp6, tmp7, tmp8; |
wolfSSL | 15:117db924cf7c | 7452 | #endif /* AES_GCM_AESNI_NO_UNROLL */ |
wolfSSL | 15:117db924cf7c | 7453 | |
wolfSSL | 15:117db924cf7c | 7454 | if (ibytes == 12) |
wolfSSL | 15:117db924cf7c | 7455 | aes_gcm_calc_iv_12(KEY, ivec, nr, H, Y, T); |
wolfSSL | 15:117db924cf7c | 7456 | else |
wolfSSL | 15:117db924cf7c | 7457 | aes_gcm_calc_iv(KEY, ivec, ibytes, nr, H, Y, T); |
wolfSSL | 15:117db924cf7c | 7458 | |
wolfSSL | 15:117db924cf7c | 7459 | for (i=0; i<abytes/16; i++) { |
wolfSSL | 15:117db924cf7c | 7460 | tmp1 = _mm_loadu_si128(&((__m128i*)addt)[i]); |
wolfSSL | 15:117db924cf7c | 7461 | tmp1 = _mm_shuffle_epi8(tmp1, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7462 | X = _mm_xor_si128(X, tmp1); |
wolfSSL | 15:117db924cf7c | 7463 | X = gfmul_sw(X, H); |
wolfSSL | 15:117db924cf7c | 7464 | } |
wolfSSL | 15:117db924cf7c | 7465 | if (abytes%16) { |
wolfSSL | 15:117db924cf7c | 7466 | last_block = _mm_setzero_si128(); |
wolfSSL | 15:117db924cf7c | 7467 | for (j=0; j<abytes%16; j++) |
wolfSSL | 15:117db924cf7c | 7468 | ((unsigned char*)&last_block)[j] = addt[i*16+j]; |
wolfSSL | 15:117db924cf7c | 7469 | tmp1 = last_block; |
wolfSSL | 15:117db924cf7c | 7470 | tmp1 = _mm_shuffle_epi8(tmp1, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7471 | X = _mm_xor_si128(X, tmp1); |
wolfSSL | 15:117db924cf7c | 7472 | X = gfmul_sw(X, H); |
wolfSSL | 15:117db924cf7c | 7473 | } |
wolfSSL | 15:117db924cf7c | 7474 | |
wolfSSL | 15:117db924cf7c | 7475 | tmp1 = _mm_shuffle_epi8(Y, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7476 | ctr1 = _mm_add_epi32(tmp1, ONE); |
wolfSSL | 15:117db924cf7c | 7477 | H = gfmul_shl1(H); |
wolfSSL | 15:117db924cf7c | 7478 | i = 0; |
wolfSSL | 15:117db924cf7c | 7479 | |
wolfSSL | 15:117db924cf7c | 7480 | #ifndef AES_GCM_AESNI_NO_UNROLL |
wolfSSL | 15:117db924cf7c | 7481 | |
wolfSSL | 15:117db924cf7c | 7482 | if (0 < nbytes/16/8) { |
wolfSSL | 15:117db924cf7c | 7483 | HT[0] = H; |
wolfSSL | 15:117db924cf7c | 7484 | HT[1] = gfmul_shifted(H, H); |
wolfSSL | 15:117db924cf7c | 7485 | HT[2] = gfmul_shifted(H, HT[1]); |
wolfSSL | 15:117db924cf7c | 7486 | HT[3] = gfmul_shifted(HT[1], HT[1]); |
wolfSSL | 15:117db924cf7c | 7487 | HT[4] = gfmul_shifted(HT[1], HT[2]); |
wolfSSL | 15:117db924cf7c | 7488 | HT[5] = gfmul_shifted(HT[2], HT[2]); |
wolfSSL | 15:117db924cf7c | 7489 | HT[6] = gfmul_shifted(HT[2], HT[3]); |
wolfSSL | 15:117db924cf7c | 7490 | HT[7] = gfmul_shifted(HT[3], HT[3]); |
wolfSSL | 15:117db924cf7c | 7491 | |
wolfSSL | 15:117db924cf7c | 7492 | for (; i < nbytes/16/8; i++) { |
wolfSSL | 15:117db924cf7c | 7493 | r0 = _mm_setzero_si128(); |
wolfSSL | 15:117db924cf7c | 7494 | r1 = _mm_setzero_si128(); |
wolfSSL | 15:117db924cf7c | 7495 | |
wolfSSL | 15:117db924cf7c | 7496 | tmp1 = _mm_shuffle_epi8(ctr1, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7497 | tmp2 = _mm_add_epi32(ctr1, ONE); |
wolfSSL | 15:117db924cf7c | 7498 | tmp2 = _mm_shuffle_epi8(tmp2, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7499 | tmp3 = _mm_add_epi32(ctr1, TWO); |
wolfSSL | 15:117db924cf7c | 7500 | tmp3 = _mm_shuffle_epi8(tmp3, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7501 | tmp4 = _mm_add_epi32(ctr1, THREE); |
wolfSSL | 15:117db924cf7c | 7502 | tmp4 = _mm_shuffle_epi8(tmp4, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7503 | tmp5 = _mm_add_epi32(ctr1, FOUR); |
wolfSSL | 15:117db924cf7c | 7504 | tmp5 = _mm_shuffle_epi8(tmp5, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7505 | tmp6 = _mm_add_epi32(ctr1, FIVE); |
wolfSSL | 15:117db924cf7c | 7506 | tmp6 = _mm_shuffle_epi8(tmp6, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7507 | tmp7 = _mm_add_epi32(ctr1, SIX); |
wolfSSL | 15:117db924cf7c | 7508 | tmp7 = _mm_shuffle_epi8(tmp7, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7509 | tmp8 = _mm_add_epi32(ctr1, SEVEN); |
wolfSSL | 15:117db924cf7c | 7510 | tmp8 = _mm_shuffle_epi8(tmp8, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7511 | ctr1 = _mm_add_epi32(ctr1, EIGHT); |
wolfSSL | 15:117db924cf7c | 7512 | tmp1 =_mm_xor_si128(tmp1, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7513 | tmp2 =_mm_xor_si128(tmp2, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7514 | tmp3 =_mm_xor_si128(tmp3, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7515 | tmp4 =_mm_xor_si128(tmp4, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7516 | tmp5 =_mm_xor_si128(tmp5, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7517 | tmp6 =_mm_xor_si128(tmp6, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7518 | tmp7 =_mm_xor_si128(tmp7, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7519 | tmp8 =_mm_xor_si128(tmp8, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7520 | /* 128 x 128 Carryless Multiply */ |
wolfSSL | 15:117db924cf7c | 7521 | XV = _mm_loadu_si128(&((__m128i*)in)[i*8+0]); |
wolfSSL | 15:117db924cf7c | 7522 | XV = _mm_shuffle_epi8(XV, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7523 | XV = _mm_xor_si128(XV, X); |
wolfSSL | 15:117db924cf7c | 7524 | gfmul_only(XV, HT[7], &r0, &r1); |
wolfSSL | 15:117db924cf7c | 7525 | tmp1 = _mm_aesenc_si128(tmp1, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7526 | tmp2 = _mm_aesenc_si128(tmp2, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7527 | tmp3 = _mm_aesenc_si128(tmp3, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7528 | tmp4 = _mm_aesenc_si128(tmp4, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7529 | tmp5 = _mm_aesenc_si128(tmp5, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7530 | tmp6 = _mm_aesenc_si128(tmp6, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7531 | tmp7 = _mm_aesenc_si128(tmp7, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7532 | tmp8 = _mm_aesenc_si128(tmp8, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7533 | /* 128 x 128 Carryless Multiply */ |
wolfSSL | 15:117db924cf7c | 7534 | XV = _mm_loadu_si128(&((__m128i*)in)[i*8+1]); |
wolfSSL | 15:117db924cf7c | 7535 | XV = _mm_shuffle_epi8(XV, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7536 | gfmul_only(XV, HT[6], &r0, &r1); |
wolfSSL | 15:117db924cf7c | 7537 | tmp1 = _mm_aesenc_si128(tmp1, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7538 | tmp2 = _mm_aesenc_si128(tmp2, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7539 | tmp3 = _mm_aesenc_si128(tmp3, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7540 | tmp4 = _mm_aesenc_si128(tmp4, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7541 | tmp5 = _mm_aesenc_si128(tmp5, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7542 | tmp6 = _mm_aesenc_si128(tmp6, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7543 | tmp7 = _mm_aesenc_si128(tmp7, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7544 | tmp8 = _mm_aesenc_si128(tmp8, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7545 | /* 128 x 128 Carryless Multiply */ |
wolfSSL | 15:117db924cf7c | 7546 | XV = _mm_loadu_si128(&((__m128i*)in)[i*8+2]); |
wolfSSL | 15:117db924cf7c | 7547 | XV = _mm_shuffle_epi8(XV, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7548 | gfmul_only(XV, HT[5], &r0, &r1); |
wolfSSL | 15:117db924cf7c | 7549 | tmp1 = _mm_aesenc_si128(tmp1, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7550 | tmp2 = _mm_aesenc_si128(tmp2, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7551 | tmp3 = _mm_aesenc_si128(tmp3, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7552 | tmp4 = _mm_aesenc_si128(tmp4, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7553 | tmp5 = _mm_aesenc_si128(tmp5, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7554 | tmp6 = _mm_aesenc_si128(tmp6, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7555 | tmp7 = _mm_aesenc_si128(tmp7, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7556 | tmp8 = _mm_aesenc_si128(tmp8, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7557 | /* 128 x 128 Carryless Multiply */ |
wolfSSL | 15:117db924cf7c | 7558 | XV = _mm_loadu_si128(&((__m128i*)in)[i*8+3]); |
wolfSSL | 15:117db924cf7c | 7559 | XV = _mm_shuffle_epi8(XV, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7560 | gfmul_only(XV, HT[4], &r0, &r1); |
wolfSSL | 15:117db924cf7c | 7561 | tmp1 = _mm_aesenc_si128(tmp1, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7562 | tmp2 = _mm_aesenc_si128(tmp2, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7563 | tmp3 = _mm_aesenc_si128(tmp3, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7564 | tmp4 = _mm_aesenc_si128(tmp4, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7565 | tmp5 = _mm_aesenc_si128(tmp5, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7566 | tmp6 = _mm_aesenc_si128(tmp6, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7567 | tmp7 = _mm_aesenc_si128(tmp7, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7568 | tmp8 = _mm_aesenc_si128(tmp8, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7569 | /* 128 x 128 Carryless Multiply */ |
wolfSSL | 15:117db924cf7c | 7570 | XV = _mm_loadu_si128(&((__m128i*)in)[i*8+4]); |
wolfSSL | 15:117db924cf7c | 7571 | XV = _mm_shuffle_epi8(XV, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7572 | gfmul_only(XV, HT[3], &r0, &r1); |
wolfSSL | 15:117db924cf7c | 7573 | tmp1 = _mm_aesenc_si128(tmp1, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7574 | tmp2 = _mm_aesenc_si128(tmp2, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7575 | tmp3 = _mm_aesenc_si128(tmp3, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7576 | tmp4 = _mm_aesenc_si128(tmp4, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7577 | tmp5 = _mm_aesenc_si128(tmp5, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7578 | tmp6 = _mm_aesenc_si128(tmp6, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7579 | tmp7 = _mm_aesenc_si128(tmp7, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7580 | tmp8 = _mm_aesenc_si128(tmp8, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7581 | /* 128 x 128 Carryless Multiply */ |
wolfSSL | 15:117db924cf7c | 7582 | XV = _mm_loadu_si128(&((__m128i*)in)[i*8+5]); |
wolfSSL | 15:117db924cf7c | 7583 | XV = _mm_shuffle_epi8(XV, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7584 | gfmul_only(XV, HT[2], &r0, &r1); |
wolfSSL | 15:117db924cf7c | 7585 | tmp1 = _mm_aesenc_si128(tmp1, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7586 | tmp2 = _mm_aesenc_si128(tmp2, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7587 | tmp3 = _mm_aesenc_si128(tmp3, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7588 | tmp4 = _mm_aesenc_si128(tmp4, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7589 | tmp5 = _mm_aesenc_si128(tmp5, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7590 | tmp6 = _mm_aesenc_si128(tmp6, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7591 | tmp7 = _mm_aesenc_si128(tmp7, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7592 | tmp8 = _mm_aesenc_si128(tmp8, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7593 | /* 128 x 128 Carryless Multiply */ |
wolfSSL | 15:117db924cf7c | 7594 | XV = _mm_loadu_si128(&((__m128i*)in)[i*8+6]); |
wolfSSL | 15:117db924cf7c | 7595 | XV = _mm_shuffle_epi8(XV, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7596 | gfmul_only(XV, HT[1], &r0, &r1); |
wolfSSL | 15:117db924cf7c | 7597 | tmp1 = _mm_aesenc_si128(tmp1, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7598 | tmp2 = _mm_aesenc_si128(tmp2, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7599 | tmp3 = _mm_aesenc_si128(tmp3, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7600 | tmp4 = _mm_aesenc_si128(tmp4, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7601 | tmp5 = _mm_aesenc_si128(tmp5, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7602 | tmp6 = _mm_aesenc_si128(tmp6, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7603 | tmp7 = _mm_aesenc_si128(tmp7, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7604 | tmp8 = _mm_aesenc_si128(tmp8, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7605 | /* 128 x 128 Carryless Multiply */ |
wolfSSL | 15:117db924cf7c | 7606 | XV = _mm_loadu_si128(&((__m128i*)in)[i*8+7]); |
wolfSSL | 15:117db924cf7c | 7607 | XV = _mm_shuffle_epi8(XV, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7608 | gfmul_only(XV, HT[0], &r0, &r1); |
wolfSSL | 15:117db924cf7c | 7609 | tmp1 = _mm_aesenc_si128(tmp1, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7610 | tmp2 = _mm_aesenc_si128(tmp2, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7611 | tmp3 = _mm_aesenc_si128(tmp3, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7612 | tmp4 = _mm_aesenc_si128(tmp4, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7613 | tmp5 = _mm_aesenc_si128(tmp5, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7614 | tmp6 = _mm_aesenc_si128(tmp6, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7615 | tmp7 = _mm_aesenc_si128(tmp7, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7616 | tmp8 = _mm_aesenc_si128(tmp8, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7617 | /* Reduction */ |
wolfSSL | 15:117db924cf7c | 7618 | X = ghash_red(r0, r1); |
wolfSSL | 15:117db924cf7c | 7619 | tmp1 = _mm_aesenc_si128(tmp1, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7620 | tmp2 = _mm_aesenc_si128(tmp2, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7621 | tmp3 = _mm_aesenc_si128(tmp3, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7622 | tmp4 = _mm_aesenc_si128(tmp4, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7623 | tmp5 = _mm_aesenc_si128(tmp5, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7624 | tmp6 = _mm_aesenc_si128(tmp6, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7625 | tmp7 = _mm_aesenc_si128(tmp7, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7626 | tmp8 = _mm_aesenc_si128(tmp8, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7627 | lastKey = KEY[10]; |
wolfSSL | 15:117db924cf7c | 7628 | if (nr > 10) { |
wolfSSL | 15:117db924cf7c | 7629 | tmp1 = _mm_aesenc_si128(tmp1, KEY[10]); |
wolfSSL | 15:117db924cf7c | 7630 | tmp2 = _mm_aesenc_si128(tmp2, KEY[10]); |
wolfSSL | 15:117db924cf7c | 7631 | tmp3 = _mm_aesenc_si128(tmp3, KEY[10]); |
wolfSSL | 15:117db924cf7c | 7632 | tmp4 = _mm_aesenc_si128(tmp4, KEY[10]); |
wolfSSL | 15:117db924cf7c | 7633 | tmp5 = _mm_aesenc_si128(tmp5, KEY[10]); |
wolfSSL | 15:117db924cf7c | 7634 | tmp6 = _mm_aesenc_si128(tmp6, KEY[10]); |
wolfSSL | 15:117db924cf7c | 7635 | tmp7 = _mm_aesenc_si128(tmp7, KEY[10]); |
wolfSSL | 15:117db924cf7c | 7636 | tmp8 = _mm_aesenc_si128(tmp8, KEY[10]); |
wolfSSL | 15:117db924cf7c | 7637 | tmp1 = _mm_aesenc_si128(tmp1, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7638 | tmp2 = _mm_aesenc_si128(tmp2, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7639 | tmp3 = _mm_aesenc_si128(tmp3, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7640 | tmp4 = _mm_aesenc_si128(tmp4, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7641 | tmp5 = _mm_aesenc_si128(tmp5, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7642 | tmp6 = _mm_aesenc_si128(tmp6, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7643 | tmp7 = _mm_aesenc_si128(tmp7, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7644 | tmp8 = _mm_aesenc_si128(tmp8, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7645 | lastKey = KEY[12]; |
wolfSSL | 15:117db924cf7c | 7646 | if (nr > 12) { |
wolfSSL | 15:117db924cf7c | 7647 | tmp1 = _mm_aesenc_si128(tmp1, KEY[12]); |
wolfSSL | 15:117db924cf7c | 7648 | tmp2 = _mm_aesenc_si128(tmp2, KEY[12]); |
wolfSSL | 15:117db924cf7c | 7649 | tmp3 = _mm_aesenc_si128(tmp3, KEY[12]); |
wolfSSL | 15:117db924cf7c | 7650 | tmp4 = _mm_aesenc_si128(tmp4, KEY[12]); |
wolfSSL | 15:117db924cf7c | 7651 | tmp5 = _mm_aesenc_si128(tmp5, KEY[12]); |
wolfSSL | 15:117db924cf7c | 7652 | tmp6 = _mm_aesenc_si128(tmp6, KEY[12]); |
wolfSSL | 15:117db924cf7c | 7653 | tmp7 = _mm_aesenc_si128(tmp7, KEY[12]); |
wolfSSL | 15:117db924cf7c | 7654 | tmp8 = _mm_aesenc_si128(tmp8, KEY[12]); |
wolfSSL | 15:117db924cf7c | 7655 | tmp1 = _mm_aesenc_si128(tmp1, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7656 | tmp2 = _mm_aesenc_si128(tmp2, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7657 | tmp3 = _mm_aesenc_si128(tmp3, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7658 | tmp4 = _mm_aesenc_si128(tmp4, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7659 | tmp5 = _mm_aesenc_si128(tmp5, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7660 | tmp6 = _mm_aesenc_si128(tmp6, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7661 | tmp7 = _mm_aesenc_si128(tmp7, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7662 | tmp8 = _mm_aesenc_si128(tmp8, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7663 | lastKey = KEY[14]; |
wolfSSL | 15:117db924cf7c | 7664 | } |
wolfSSL | 15:117db924cf7c | 7665 | } |
wolfSSL | 15:117db924cf7c | 7666 | AES_ENC_LAST_8(); |
wolfSSL | 15:117db924cf7c | 7667 | } |
wolfSSL | 15:117db924cf7c | 7668 | } |
wolfSSL | 15:117db924cf7c | 7669 | |
wolfSSL | 15:117db924cf7c | 7670 | #endif /* AES_GCM_AESNI_NO_UNROLL */ |
wolfSSL | 15:117db924cf7c | 7671 | |
wolfSSL | 15:117db924cf7c | 7672 | for (k = i*8; k < nbytes/16; k++) { |
wolfSSL | 15:117db924cf7c | 7673 | tmp1 = _mm_shuffle_epi8(ctr1, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7674 | ctr1 = _mm_add_epi32(ctr1, ONE); |
wolfSSL | 15:117db924cf7c | 7675 | tmp1 = _mm_xor_si128(tmp1, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7676 | tmp1 = _mm_aesenc_si128(tmp1, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7677 | tmp1 = _mm_aesenc_si128(tmp1, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7678 | tmp1 = _mm_aesenc_si128(tmp1, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7679 | tmp1 = _mm_aesenc_si128(tmp1, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7680 | tmp1 = _mm_aesenc_si128(tmp1, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7681 | tmp1 = _mm_aesenc_si128(tmp1, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7682 | tmp1 = _mm_aesenc_si128(tmp1, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7683 | tmp1 = _mm_aesenc_si128(tmp1, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7684 | tmp1 = _mm_aesenc_si128(tmp1, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7685 | /* 128 x 128 Carryless Multiply */ |
wolfSSL | 15:117db924cf7c | 7686 | XV = _mm_loadu_si128(&((__m128i*)in)[k]); |
wolfSSL | 15:117db924cf7c | 7687 | XV = _mm_shuffle_epi8(XV, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7688 | XV = _mm_xor_si128(XV, X); |
wolfSSL | 15:117db924cf7c | 7689 | X = gfmul_shifted(XV, H); |
wolfSSL | 15:117db924cf7c | 7690 | lastKey = KEY[10]; |
wolfSSL | 15:117db924cf7c | 7691 | if (nr > 10) { |
wolfSSL | 15:117db924cf7c | 7692 | tmp1 = _mm_aesenc_si128(tmp1, lastKey); |
wolfSSL | 15:117db924cf7c | 7693 | tmp1 = _mm_aesenc_si128(tmp1, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7694 | lastKey = KEY[12]; |
wolfSSL | 15:117db924cf7c | 7695 | if (nr > 12) { |
wolfSSL | 15:117db924cf7c | 7696 | tmp1 = _mm_aesenc_si128(tmp1, lastKey); |
wolfSSL | 15:117db924cf7c | 7697 | tmp1 = _mm_aesenc_si128(tmp1, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7698 | lastKey = KEY[14]; |
wolfSSL | 15:117db924cf7c | 7699 | } |
wolfSSL | 15:117db924cf7c | 7700 | } |
wolfSSL | 15:117db924cf7c | 7701 | tmp1 = _mm_aesenclast_si128(tmp1, lastKey); |
wolfSSL | 15:117db924cf7c | 7702 | tmp2 = _mm_loadu_si128(&((__m128i*)in)[k]); |
wolfSSL | 15:117db924cf7c | 7703 | tmp1 = _mm_xor_si128(tmp1, tmp2); |
wolfSSL | 15:117db924cf7c | 7704 | _mm_storeu_si128(&((__m128i*)out)[k], tmp1); |
wolfSSL | 15:117db924cf7c | 7705 | } |
wolfSSL | 15:117db924cf7c | 7706 | |
wolfSSL | 15:117db924cf7c | 7707 | /* If one partial block remains */ |
wolfSSL | 15:117db924cf7c | 7708 | if (nbytes % 16) { |
wolfSSL | 15:117db924cf7c | 7709 | tmp1 = _mm_shuffle_epi8(ctr1, BSWAP_EPI64); |
wolfSSL | 15:117db924cf7c | 7710 | tmp1 = _mm_xor_si128(tmp1, KEY[0]); |
wolfSSL | 15:117db924cf7c | 7711 | tmp1 = _mm_aesenc_si128(tmp1, KEY[1]); |
wolfSSL | 15:117db924cf7c | 7712 | tmp1 = _mm_aesenc_si128(tmp1, KEY[2]); |
wolfSSL | 15:117db924cf7c | 7713 | tmp1 = _mm_aesenc_si128(tmp1, KEY[3]); |
wolfSSL | 15:117db924cf7c | 7714 | tmp1 = _mm_aesenc_si128(tmp1, KEY[4]); |
wolfSSL | 15:117db924cf7c | 7715 | tmp1 = _mm_aesenc_si128(tmp1, KEY[5]); |
wolfSSL | 15:117db924cf7c | 7716 | tmp1 = _mm_aesenc_si128(tmp1, KEY[6]); |
wolfSSL | 15:117db924cf7c | 7717 | tmp1 = _mm_aesenc_si128(tmp1, KEY[7]); |
wolfSSL | 15:117db924cf7c | 7718 | tmp1 = _mm_aesenc_si128(tmp1, KEY[8]); |
wolfSSL | 15:117db924cf7c | 7719 | tmp1 = _mm_aesenc_si128(tmp1, KEY[9]); |
wolfSSL | 15:117db924cf7c | 7720 | lastKey = KEY[10]; |
wolfSSL | 15:117db924cf7c | 7721 | if (nr > 10) { |
wolfSSL | 15:117db924cf7c | 7722 | tmp1 = _mm_aesenc_si128(tmp1, lastKey); |
wolfSSL | 15:117db924cf7c | 7723 | tmp1 = _mm_aesenc_si128(tmp1, KEY[11]); |
wolfSSL | 15:117db924cf7c | 7724 | lastKey = KEY[12]; |
wolfSSL | 15:117db924cf7c | 7725 | if (nr > 12) { |
wolfSSL | 15:117db924cf7c | 7726 | tmp1 = _mm_aesenc_si128(tmp1, lastKey); |
wolfSSL | 15:117db924cf7c | 7727 | tmp1 = _mm_aesenc_si128(tmp1, KEY[13]); |
wolfSSL | 15:117db924cf7c | 7728 | lastKey = KEY[14]; |
wolfSSL | 15:117db924cf7c | 7729 | } |
wolfSSL | 15:117db924cf7c | 7730 | } |
wolfSSL | 15:117db924cf7c | 7731 | tmp1 = _mm_aesenclast_si128(tmp1, lastKey); |
wolfSSL | 15:117db924cf7c | 7732 | last_block = _mm_setzero_si128(); |
wolfSSL | 15:117db924cf7c | 7733 | for (j=0; j < nbytes%16; j++) |
wolfSSL | 15:117db924cf7c | 7734 | ((unsigned char*)&last_block)[j] = in[k*16+j]; |
wolfSSL | 15:117db924cf7c | 7735 | XV = last_block; |
wolfSSL | 15:117db924cf7c | 7736 | tmp1 = _mm_xor_si128(tmp1, last_block); |
wolfSSL | 15:117db924cf7c | 7737 | last_block = tmp1; |
wolfSSL | 15:117db924cf7c | 7738 | for (j=0; j < nbytes%16; j++) |
wolfSSL | 15:117db924cf7c | 7739 | out[k*16+j] = ((unsigned char*)&last_block)[j]; |
wolfSSL | 15:117db924cf7c | 7740 | XV = _mm_shuffle_epi8(XV, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7741 | XV = _mm_xor_si128(XV, X); |
wolfSSL | 15:117db924cf7c | 7742 | X = gfmul_shifted(XV, H); |
wolfSSL | 15:117db924cf7c | 7743 | } |
wolfSSL | 15:117db924cf7c | 7744 | |
wolfSSL | 15:117db924cf7c | 7745 | tmp1 = _mm_insert_epi64(tmp1, nbytes*8, 0); |
wolfSSL | 15:117db924cf7c | 7746 | tmp1 = _mm_insert_epi64(tmp1, abytes*8, 1); |
wolfSSL | 15:117db924cf7c | 7747 | /* 128 x 128 Carryless Multiply */ |
wolfSSL | 15:117db924cf7c | 7748 | X = _mm_xor_si128(X, tmp1); |
wolfSSL | 15:117db924cf7c | 7749 | X = gfmul_shifted(X, H); |
wolfSSL | 15:117db924cf7c | 7750 | X = _mm_shuffle_epi8(X, BSWAP_MASK); |
wolfSSL | 15:117db924cf7c | 7751 | T = _mm_xor_si128(X, T); |
wolfSSL | 15:117db924cf7c | 7752 | |
wolfSSL | 15:117db924cf7c | 7753 | /* if (0xffff != |
wolfSSL | 15:117db924cf7c | 7754 | _mm_movemask_epi8(_mm_cmpeq_epi8(T, _mm_loadu_si128((__m128i*)tag)))) */ |
wolfSSL | 15:117db924cf7c | 7755 | if (XMEMCMP(tag, &T, tbytes) != 0) |
wolfSSL | 15:117db924cf7c | 7756 | *res = 0; /* in case the authentication failed */ |
wolfSSL | 15:117db924cf7c | 7757 | else |
wolfSSL | 15:117db924cf7c | 7758 | *res = 1; /* when successful returns 1 */ |
wolfSSL | 15:117db924cf7c | 7759 | } |
wolfSSL | 15:117db924cf7c | 7760 | |
wolfSSL | 15:117db924cf7c | 7761 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 7762 | #endif /* _MSC_VER */ |
wolfSSL | 15:117db924cf7c | 7763 | #endif /* WOLFSSL_AESNI */ |
wolfSSL | 15:117db924cf7c | 7764 | |
wolfSSL | 15:117db924cf7c | 7765 | |
wolfSSL | 15:117db924cf7c | 7766 | #if defined(GCM_SMALL) |
wolfSSL | 15:117db924cf7c | 7767 | static void GMULT(byte* X, byte* Y) |
wolfSSL | 15:117db924cf7c | 7768 | { |
wolfSSL | 15:117db924cf7c | 7769 | byte Z[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 7770 | byte V[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 7771 | int i, j; |
wolfSSL | 15:117db924cf7c | 7772 | |
wolfSSL | 15:117db924cf7c | 7773 | XMEMSET(Z, 0, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7774 | XMEMCPY(V, X, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7775 | for (i = 0; i < AES_BLOCK_SIZE; i++) |
wolfSSL | 15:117db924cf7c | 7776 | { |
wolfSSL | 15:117db924cf7c | 7777 | byte y = Y[i]; |
wolfSSL | 15:117db924cf7c | 7778 | for (j = 0; j < 8; j++) |
wolfSSL | 15:117db924cf7c | 7779 | { |
wolfSSL | 15:117db924cf7c | 7780 | if (y & 0x80) { |
wolfSSL | 15:117db924cf7c | 7781 | xorbuf(Z, V, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7782 | } |
wolfSSL | 15:117db924cf7c | 7783 | |
wolfSSL | 15:117db924cf7c | 7784 | RIGHTSHIFTX(V); |
wolfSSL | 15:117db924cf7c | 7785 | y = y << 1; |
wolfSSL | 15:117db924cf7c | 7786 | } |
wolfSSL | 15:117db924cf7c | 7787 | } |
wolfSSL | 15:117db924cf7c | 7788 | XMEMCPY(X, Z, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7789 | } |
wolfSSL | 15:117db924cf7c | 7790 | |
wolfSSL | 15:117db924cf7c | 7791 | |
wolfSSL | 15:117db924cf7c | 7792 | void GHASH(Aes* aes, const byte* a, word32 aSz, const byte* c, |
wolfSSL | 15:117db924cf7c | 7793 | word32 cSz, byte* s, word32 sSz) |
wolfSSL | 15:117db924cf7c | 7794 | { |
wolfSSL | 15:117db924cf7c | 7795 | byte x[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 7796 | byte scratch[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 7797 | word32 blocks, partial; |
wolfSSL | 15:117db924cf7c | 7798 | byte* h = aes->H; |
wolfSSL | 15:117db924cf7c | 7799 | |
wolfSSL | 15:117db924cf7c | 7800 | XMEMSET(x, 0, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7801 | |
wolfSSL | 15:117db924cf7c | 7802 | /* Hash in A, the Additional Authentication Data */ |
wolfSSL | 15:117db924cf7c | 7803 | if (aSz != 0 && a != NULL) { |
wolfSSL | 15:117db924cf7c | 7804 | blocks = aSz / AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 7805 | partial = aSz % AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 7806 | while (blocks--) { |
wolfSSL | 15:117db924cf7c | 7807 | xorbuf(x, a, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7808 | GMULT(x, h); |
wolfSSL | 15:117db924cf7c | 7809 | a += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 7810 | } |
wolfSSL | 15:117db924cf7c | 7811 | if (partial != 0) { |
wolfSSL | 15:117db924cf7c | 7812 | XMEMSET(scratch, 0, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7813 | XMEMCPY(scratch, a, partial); |
wolfSSL | 15:117db924cf7c | 7814 | xorbuf(x, scratch, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7815 | GMULT(x, h); |
wolfSSL | 15:117db924cf7c | 7816 | } |
wolfSSL | 15:117db924cf7c | 7817 | } |
wolfSSL | 15:117db924cf7c | 7818 | |
wolfSSL | 15:117db924cf7c | 7819 | /* Hash in C, the Ciphertext */ |
wolfSSL | 15:117db924cf7c | 7820 | if (cSz != 0 && c != NULL) { |
wolfSSL | 15:117db924cf7c | 7821 | blocks = cSz / AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 7822 | partial = cSz % AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 7823 | while (blocks--) { |
wolfSSL | 15:117db924cf7c | 7824 | xorbuf(x, c, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7825 | GMULT(x, h); |
wolfSSL | 15:117db924cf7c | 7826 | c += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 7827 | } |
wolfSSL | 15:117db924cf7c | 7828 | if (partial != 0) { |
wolfSSL | 15:117db924cf7c | 7829 | XMEMSET(scratch, 0, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7830 | XMEMCPY(scratch, c, partial); |
wolfSSL | 15:117db924cf7c | 7831 | xorbuf(x, scratch, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7832 | GMULT(x, h); |
wolfSSL | 15:117db924cf7c | 7833 | } |
wolfSSL | 15:117db924cf7c | 7834 | } |
wolfSSL | 15:117db924cf7c | 7835 | |
wolfSSL | 15:117db924cf7c | 7836 | /* Hash in the lengths of A and C in bits */ |
wolfSSL | 15:117db924cf7c | 7837 | FlattenSzInBits(&scratch[0], aSz); |
wolfSSL | 15:117db924cf7c | 7838 | FlattenSzInBits(&scratch[8], cSz); |
wolfSSL | 15:117db924cf7c | 7839 | xorbuf(x, scratch, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7840 | GMULT(x, h); |
wolfSSL | 15:117db924cf7c | 7841 | |
wolfSSL | 15:117db924cf7c | 7842 | /* Copy the result into s. */ |
wolfSSL | 15:117db924cf7c | 7843 | XMEMCPY(s, x, sSz); |
wolfSSL | 15:117db924cf7c | 7844 | } |
wolfSSL | 15:117db924cf7c | 7845 | |
wolfSSL | 15:117db924cf7c | 7846 | /* end GCM_SMALL */ |
wolfSSL | 15:117db924cf7c | 7847 | #elif defined(GCM_TABLE) |
wolfSSL | 15:117db924cf7c | 7848 | |
wolfSSL | 15:117db924cf7c | 7849 | static const byte R[256][2] = { |
wolfSSL | 15:117db924cf7c | 7850 | {0x00, 0x00}, {0x01, 0xc2}, {0x03, 0x84}, {0x02, 0x46}, |
wolfSSL | 15:117db924cf7c | 7851 | {0x07, 0x08}, {0x06, 0xca}, {0x04, 0x8c}, {0x05, 0x4e}, |
wolfSSL | 15:117db924cf7c | 7852 | {0x0e, 0x10}, {0x0f, 0xd2}, {0x0d, 0x94}, {0x0c, 0x56}, |
wolfSSL | 15:117db924cf7c | 7853 | {0x09, 0x18}, {0x08, 0xda}, {0x0a, 0x9c}, {0x0b, 0x5e}, |
wolfSSL | 15:117db924cf7c | 7854 | {0x1c, 0x20}, {0x1d, 0xe2}, {0x1f, 0xa4}, {0x1e, 0x66}, |
wolfSSL | 15:117db924cf7c | 7855 | {0x1b, 0x28}, {0x1a, 0xea}, {0x18, 0xac}, {0x19, 0x6e}, |
wolfSSL | 15:117db924cf7c | 7856 | {0x12, 0x30}, {0x13, 0xf2}, {0x11, 0xb4}, {0x10, 0x76}, |
wolfSSL | 15:117db924cf7c | 7857 | {0x15, 0x38}, {0x14, 0xfa}, {0x16, 0xbc}, {0x17, 0x7e}, |
wolfSSL | 15:117db924cf7c | 7858 | {0x38, 0x40}, {0x39, 0x82}, {0x3b, 0xc4}, {0x3a, 0x06}, |
wolfSSL | 15:117db924cf7c | 7859 | {0x3f, 0x48}, {0x3e, 0x8a}, {0x3c, 0xcc}, {0x3d, 0x0e}, |
wolfSSL | 15:117db924cf7c | 7860 | {0x36, 0x50}, {0x37, 0x92}, {0x35, 0xd4}, {0x34, 0x16}, |
wolfSSL | 15:117db924cf7c | 7861 | {0x31, 0x58}, {0x30, 0x9a}, {0x32, 0xdc}, {0x33, 0x1e}, |
wolfSSL | 15:117db924cf7c | 7862 | {0x24, 0x60}, {0x25, 0xa2}, {0x27, 0xe4}, {0x26, 0x26}, |
wolfSSL | 15:117db924cf7c | 7863 | {0x23, 0x68}, {0x22, 0xaa}, {0x20, 0xec}, {0x21, 0x2e}, |
wolfSSL | 15:117db924cf7c | 7864 | {0x2a, 0x70}, {0x2b, 0xb2}, {0x29, 0xf4}, {0x28, 0x36}, |
wolfSSL | 15:117db924cf7c | 7865 | {0x2d, 0x78}, {0x2c, 0xba}, {0x2e, 0xfc}, {0x2f, 0x3e}, |
wolfSSL | 15:117db924cf7c | 7866 | {0x70, 0x80}, {0x71, 0x42}, {0x73, 0x04}, {0x72, 0xc6}, |
wolfSSL | 15:117db924cf7c | 7867 | {0x77, 0x88}, {0x76, 0x4a}, {0x74, 0x0c}, {0x75, 0xce}, |
wolfSSL | 15:117db924cf7c | 7868 | {0x7e, 0x90}, {0x7f, 0x52}, {0x7d, 0x14}, {0x7c, 0xd6}, |
wolfSSL | 15:117db924cf7c | 7869 | {0x79, 0x98}, {0x78, 0x5a}, {0x7a, 0x1c}, {0x7b, 0xde}, |
wolfSSL | 15:117db924cf7c | 7870 | {0x6c, 0xa0}, {0x6d, 0x62}, {0x6f, 0x24}, {0x6e, 0xe6}, |
wolfSSL | 15:117db924cf7c | 7871 | {0x6b, 0xa8}, {0x6a, 0x6a}, {0x68, 0x2c}, {0x69, 0xee}, |
wolfSSL | 15:117db924cf7c | 7872 | {0x62, 0xb0}, {0x63, 0x72}, {0x61, 0x34}, {0x60, 0xf6}, |
wolfSSL | 15:117db924cf7c | 7873 | {0x65, 0xb8}, {0x64, 0x7a}, {0x66, 0x3c}, {0x67, 0xfe}, |
wolfSSL | 15:117db924cf7c | 7874 | {0x48, 0xc0}, {0x49, 0x02}, {0x4b, 0x44}, {0x4a, 0x86}, |
wolfSSL | 15:117db924cf7c | 7875 | {0x4f, 0xc8}, {0x4e, 0x0a}, {0x4c, 0x4c}, {0x4d, 0x8e}, |
wolfSSL | 15:117db924cf7c | 7876 | {0x46, 0xd0}, {0x47, 0x12}, {0x45, 0x54}, {0x44, 0x96}, |
wolfSSL | 15:117db924cf7c | 7877 | {0x41, 0xd8}, {0x40, 0x1a}, {0x42, 0x5c}, {0x43, 0x9e}, |
wolfSSL | 15:117db924cf7c | 7878 | {0x54, 0xe0}, {0x55, 0x22}, {0x57, 0x64}, {0x56, 0xa6}, |
wolfSSL | 15:117db924cf7c | 7879 | {0x53, 0xe8}, {0x52, 0x2a}, {0x50, 0x6c}, {0x51, 0xae}, |
wolfSSL | 15:117db924cf7c | 7880 | {0x5a, 0xf0}, {0x5b, 0x32}, {0x59, 0x74}, {0x58, 0xb6}, |
wolfSSL | 15:117db924cf7c | 7881 | {0x5d, 0xf8}, {0x5c, 0x3a}, {0x5e, 0x7c}, {0x5f, 0xbe}, |
wolfSSL | 15:117db924cf7c | 7882 | {0xe1, 0x00}, {0xe0, 0xc2}, {0xe2, 0x84}, {0xe3, 0x46}, |
wolfSSL | 15:117db924cf7c | 7883 | {0xe6, 0x08}, {0xe7, 0xca}, {0xe5, 0x8c}, {0xe4, 0x4e}, |
wolfSSL | 15:117db924cf7c | 7884 | {0xef, 0x10}, {0xee, 0xd2}, {0xec, 0x94}, {0xed, 0x56}, |
wolfSSL | 15:117db924cf7c | 7885 | {0xe8, 0x18}, {0xe9, 0xda}, {0xeb, 0x9c}, {0xea, 0x5e}, |
wolfSSL | 15:117db924cf7c | 7886 | {0xfd, 0x20}, {0xfc, 0xe2}, {0xfe, 0xa4}, {0xff, 0x66}, |
wolfSSL | 15:117db924cf7c | 7887 | {0xfa, 0x28}, {0xfb, 0xea}, {0xf9, 0xac}, {0xf8, 0x6e}, |
wolfSSL | 15:117db924cf7c | 7888 | {0xf3, 0x30}, {0xf2, 0xf2}, {0xf0, 0xb4}, {0xf1, 0x76}, |
wolfSSL | 15:117db924cf7c | 7889 | {0xf4, 0x38}, {0xf5, 0xfa}, {0xf7, 0xbc}, {0xf6, 0x7e}, |
wolfSSL | 15:117db924cf7c | 7890 | {0xd9, 0x40}, {0xd8, 0x82}, {0xda, 0xc4}, {0xdb, 0x06}, |
wolfSSL | 15:117db924cf7c | 7891 | {0xde, 0x48}, {0xdf, 0x8a}, {0xdd, 0xcc}, {0xdc, 0x0e}, |
wolfSSL | 15:117db924cf7c | 7892 | {0xd7, 0x50}, {0xd6, 0x92}, {0xd4, 0xd4}, {0xd5, 0x16}, |
wolfSSL | 15:117db924cf7c | 7893 | {0xd0, 0x58}, {0xd1, 0x9a}, {0xd3, 0xdc}, {0xd2, 0x1e}, |
wolfSSL | 15:117db924cf7c | 7894 | {0xc5, 0x60}, {0xc4, 0xa2}, {0xc6, 0xe4}, {0xc7, 0x26}, |
wolfSSL | 15:117db924cf7c | 7895 | {0xc2, 0x68}, {0xc3, 0xaa}, {0xc1, 0xec}, {0xc0, 0x2e}, |
wolfSSL | 15:117db924cf7c | 7896 | {0xcb, 0x70}, {0xca, 0xb2}, {0xc8, 0xf4}, {0xc9, 0x36}, |
wolfSSL | 15:117db924cf7c | 7897 | {0xcc, 0x78}, {0xcd, 0xba}, {0xcf, 0xfc}, {0xce, 0x3e}, |
wolfSSL | 15:117db924cf7c | 7898 | {0x91, 0x80}, {0x90, 0x42}, {0x92, 0x04}, {0x93, 0xc6}, |
wolfSSL | 15:117db924cf7c | 7899 | {0x96, 0x88}, {0x97, 0x4a}, {0x95, 0x0c}, {0x94, 0xce}, |
wolfSSL | 15:117db924cf7c | 7900 | {0x9f, 0x90}, {0x9e, 0x52}, {0x9c, 0x14}, {0x9d, 0xd6}, |
wolfSSL | 15:117db924cf7c | 7901 | {0x98, 0x98}, {0x99, 0x5a}, {0x9b, 0x1c}, {0x9a, 0xde}, |
wolfSSL | 15:117db924cf7c | 7902 | {0x8d, 0xa0}, {0x8c, 0x62}, {0x8e, 0x24}, {0x8f, 0xe6}, |
wolfSSL | 15:117db924cf7c | 7903 | {0x8a, 0xa8}, {0x8b, 0x6a}, {0x89, 0x2c}, {0x88, 0xee}, |
wolfSSL | 15:117db924cf7c | 7904 | {0x83, 0xb0}, {0x82, 0x72}, {0x80, 0x34}, {0x81, 0xf6}, |
wolfSSL | 15:117db924cf7c | 7905 | {0x84, 0xb8}, {0x85, 0x7a}, {0x87, 0x3c}, {0x86, 0xfe}, |
wolfSSL | 15:117db924cf7c | 7906 | {0xa9, 0xc0}, {0xa8, 0x02}, {0xaa, 0x44}, {0xab, 0x86}, |
wolfSSL | 15:117db924cf7c | 7907 | {0xae, 0xc8}, {0xaf, 0x0a}, {0xad, 0x4c}, {0xac, 0x8e}, |
wolfSSL | 15:117db924cf7c | 7908 | {0xa7, 0xd0}, {0xa6, 0x12}, {0xa4, 0x54}, {0xa5, 0x96}, |
wolfSSL | 15:117db924cf7c | 7909 | {0xa0, 0xd8}, {0xa1, 0x1a}, {0xa3, 0x5c}, {0xa2, 0x9e}, |
wolfSSL | 15:117db924cf7c | 7910 | {0xb5, 0xe0}, {0xb4, 0x22}, {0xb6, 0x64}, {0xb7, 0xa6}, |
wolfSSL | 15:117db924cf7c | 7911 | {0xb2, 0xe8}, {0xb3, 0x2a}, {0xb1, 0x6c}, {0xb0, 0xae}, |
wolfSSL | 15:117db924cf7c | 7912 | {0xbb, 0xf0}, {0xba, 0x32}, {0xb8, 0x74}, {0xb9, 0xb6}, |
wolfSSL | 15:117db924cf7c | 7913 | {0xbc, 0xf8}, {0xbd, 0x3a}, {0xbf, 0x7c}, {0xbe, 0xbe} }; |
wolfSSL | 15:117db924cf7c | 7914 | |
wolfSSL | 15:117db924cf7c | 7915 | |
wolfSSL | 15:117db924cf7c | 7916 | static void GMULT(byte *x, byte m[256][AES_BLOCK_SIZE]) |
wolfSSL | 15:117db924cf7c | 7917 | { |
wolfSSL | 15:117db924cf7c | 7918 | int i, j; |
wolfSSL | 15:117db924cf7c | 7919 | byte Z[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 7920 | byte a; |
wolfSSL | 15:117db924cf7c | 7921 | |
wolfSSL | 15:117db924cf7c | 7922 | XMEMSET(Z, 0, sizeof(Z)); |
wolfSSL | 15:117db924cf7c | 7923 | |
wolfSSL | 15:117db924cf7c | 7924 | for (i = 15; i > 0; i--) { |
wolfSSL | 15:117db924cf7c | 7925 | xorbuf(Z, m[x[i]], AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7926 | a = Z[15]; |
wolfSSL | 15:117db924cf7c | 7927 | |
wolfSSL | 15:117db924cf7c | 7928 | for (j = 15; j > 0; j--) { |
wolfSSL | 15:117db924cf7c | 7929 | Z[j] = Z[j-1]; |
wolfSSL | 15:117db924cf7c | 7930 | } |
wolfSSL | 15:117db924cf7c | 7931 | |
wolfSSL | 15:117db924cf7c | 7932 | Z[0] = R[a][0]; |
wolfSSL | 15:117db924cf7c | 7933 | Z[1] ^= R[a][1]; |
wolfSSL | 15:117db924cf7c | 7934 | } |
wolfSSL | 15:117db924cf7c | 7935 | xorbuf(Z, m[x[0]], AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7936 | |
wolfSSL | 15:117db924cf7c | 7937 | XMEMCPY(x, Z, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7938 | } |
wolfSSL | 15:117db924cf7c | 7939 | |
wolfSSL | 15:117db924cf7c | 7940 | |
wolfSSL | 15:117db924cf7c | 7941 | void GHASH(Aes* aes, const byte* a, word32 aSz, const byte* c, |
wolfSSL | 15:117db924cf7c | 7942 | word32 cSz, byte* s, word32 sSz) |
wolfSSL | 15:117db924cf7c | 7943 | { |
wolfSSL | 15:117db924cf7c | 7944 | byte x[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 7945 | byte scratch[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 7946 | word32 blocks, partial; |
wolfSSL | 15:117db924cf7c | 7947 | |
wolfSSL | 15:117db924cf7c | 7948 | XMEMSET(x, 0, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7949 | |
wolfSSL | 15:117db924cf7c | 7950 | /* Hash in A, the Additional Authentication Data */ |
wolfSSL | 15:117db924cf7c | 7951 | if (aSz != 0 && a != NULL) { |
wolfSSL | 15:117db924cf7c | 7952 | blocks = aSz / AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 7953 | partial = aSz % AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 7954 | while (blocks--) { |
wolfSSL | 15:117db924cf7c | 7955 | xorbuf(x, a, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7956 | GMULT(x, aes->M0); |
wolfSSL | 15:117db924cf7c | 7957 | a += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 7958 | } |
wolfSSL | 15:117db924cf7c | 7959 | if (partial != 0) { |
wolfSSL | 15:117db924cf7c | 7960 | XMEMSET(scratch, 0, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7961 | XMEMCPY(scratch, a, partial); |
wolfSSL | 15:117db924cf7c | 7962 | xorbuf(x, scratch, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7963 | GMULT(x, aes->M0); |
wolfSSL | 15:117db924cf7c | 7964 | } |
wolfSSL | 15:117db924cf7c | 7965 | } |
wolfSSL | 15:117db924cf7c | 7966 | |
wolfSSL | 15:117db924cf7c | 7967 | /* Hash in C, the Ciphertext */ |
wolfSSL | 15:117db924cf7c | 7968 | if (cSz != 0 && c != NULL) { |
wolfSSL | 15:117db924cf7c | 7969 | blocks = cSz / AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 7970 | partial = cSz % AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 7971 | while (blocks--) { |
wolfSSL | 15:117db924cf7c | 7972 | xorbuf(x, c, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7973 | GMULT(x, aes->M0); |
wolfSSL | 15:117db924cf7c | 7974 | c += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 7975 | } |
wolfSSL | 15:117db924cf7c | 7976 | if (partial != 0) { |
wolfSSL | 15:117db924cf7c | 7977 | XMEMSET(scratch, 0, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7978 | XMEMCPY(scratch, c, partial); |
wolfSSL | 15:117db924cf7c | 7979 | xorbuf(x, scratch, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7980 | GMULT(x, aes->M0); |
wolfSSL | 15:117db924cf7c | 7981 | } |
wolfSSL | 15:117db924cf7c | 7982 | } |
wolfSSL | 15:117db924cf7c | 7983 | |
wolfSSL | 15:117db924cf7c | 7984 | /* Hash in the lengths of A and C in bits */ |
wolfSSL | 15:117db924cf7c | 7985 | FlattenSzInBits(&scratch[0], aSz); |
wolfSSL | 15:117db924cf7c | 7986 | FlattenSzInBits(&scratch[8], cSz); |
wolfSSL | 15:117db924cf7c | 7987 | xorbuf(x, scratch, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 7988 | GMULT(x, aes->M0); |
wolfSSL | 15:117db924cf7c | 7989 | |
wolfSSL | 15:117db924cf7c | 7990 | /* Copy the result into s. */ |
wolfSSL | 15:117db924cf7c | 7991 | XMEMCPY(s, x, sSz); |
wolfSSL | 15:117db924cf7c | 7992 | } |
wolfSSL | 15:117db924cf7c | 7993 | |
wolfSSL | 15:117db924cf7c | 7994 | /* end GCM_TABLE */ |
wolfSSL | 15:117db924cf7c | 7995 | #elif defined(WORD64_AVAILABLE) && !defined(GCM_WORD32) |
wolfSSL | 15:117db924cf7c | 7996 | |
wolfSSL | 15:117db924cf7c | 7997 | #if !defined(FREESCALE_LTC_AES_GCM) |
wolfSSL | 15:117db924cf7c | 7998 | static void GMULT(word64* X, word64* Y) |
wolfSSL | 15:117db924cf7c | 7999 | { |
wolfSSL | 15:117db924cf7c | 8000 | word64 Z[2] = {0,0}; |
wolfSSL | 15:117db924cf7c | 8001 | word64 V[2]; |
wolfSSL | 15:117db924cf7c | 8002 | int i, j; |
wolfSSL | 15:117db924cf7c | 8003 | V[0] = X[0]; V[1] = X[1]; |
wolfSSL | 15:117db924cf7c | 8004 | |
wolfSSL | 15:117db924cf7c | 8005 | for (i = 0; i < 2; i++) |
wolfSSL | 15:117db924cf7c | 8006 | { |
wolfSSL | 15:117db924cf7c | 8007 | word64 y = Y[i]; |
wolfSSL | 15:117db924cf7c | 8008 | for (j = 0; j < 64; j++) |
wolfSSL | 15:117db924cf7c | 8009 | { |
wolfSSL | 15:117db924cf7c | 8010 | if (y & 0x8000000000000000ULL) { |
wolfSSL | 15:117db924cf7c | 8011 | Z[0] ^= V[0]; |
wolfSSL | 15:117db924cf7c | 8012 | Z[1] ^= V[1]; |
wolfSSL | 15:117db924cf7c | 8013 | } |
wolfSSL | 15:117db924cf7c | 8014 | |
wolfSSL | 15:117db924cf7c | 8015 | if (V[1] & 0x0000000000000001) { |
wolfSSL | 15:117db924cf7c | 8016 | V[1] >>= 1; |
wolfSSL | 15:117db924cf7c | 8017 | V[1] |= ((V[0] & 0x0000000000000001) ? |
wolfSSL | 15:117db924cf7c | 8018 | 0x8000000000000000ULL : 0); |
wolfSSL | 15:117db924cf7c | 8019 | V[0] >>= 1; |
wolfSSL | 15:117db924cf7c | 8020 | V[0] ^= 0xE100000000000000ULL; |
wolfSSL | 15:117db924cf7c | 8021 | } |
wolfSSL | 15:117db924cf7c | 8022 | else { |
wolfSSL | 15:117db924cf7c | 8023 | V[1] >>= 1; |
wolfSSL | 15:117db924cf7c | 8024 | V[1] |= ((V[0] & 0x0000000000000001) ? |
wolfSSL | 15:117db924cf7c | 8025 | 0x8000000000000000ULL : 0); |
wolfSSL | 15:117db924cf7c | 8026 | V[0] >>= 1; |
wolfSSL | 15:117db924cf7c | 8027 | } |
wolfSSL | 15:117db924cf7c | 8028 | y <<= 1; |
wolfSSL | 15:117db924cf7c | 8029 | } |
wolfSSL | 15:117db924cf7c | 8030 | } |
wolfSSL | 15:117db924cf7c | 8031 | X[0] = Z[0]; |
wolfSSL | 15:117db924cf7c | 8032 | X[1] = Z[1]; |
wolfSSL | 15:117db924cf7c | 8033 | } |
wolfSSL | 15:117db924cf7c | 8034 | |
wolfSSL | 15:117db924cf7c | 8035 | |
wolfSSL | 15:117db924cf7c | 8036 | void GHASH(Aes* aes, const byte* a, word32 aSz, const byte* c, |
wolfSSL | 15:117db924cf7c | 8037 | word32 cSz, byte* s, word32 sSz) |
wolfSSL | 15:117db924cf7c | 8038 | { |
wolfSSL | 15:117db924cf7c | 8039 | word64 x[2] = {0,0}; |
wolfSSL | 15:117db924cf7c | 8040 | word32 blocks, partial; |
wolfSSL | 15:117db924cf7c | 8041 | word64 bigH[2]; |
wolfSSL | 15:117db924cf7c | 8042 | |
wolfSSL | 15:117db924cf7c | 8043 | XMEMCPY(bigH, aes->H, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8044 | #ifdef LITTLE_ENDIAN_ORDER |
wolfSSL | 15:117db924cf7c | 8045 | ByteReverseWords64(bigH, bigH, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8046 | #endif |
wolfSSL | 15:117db924cf7c | 8047 | |
wolfSSL | 15:117db924cf7c | 8048 | /* Hash in A, the Additional Authentication Data */ |
wolfSSL | 15:117db924cf7c | 8049 | if (aSz != 0 && a != NULL) { |
wolfSSL | 15:117db924cf7c | 8050 | word64 bigA[2]; |
wolfSSL | 15:117db924cf7c | 8051 | blocks = aSz / AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8052 | partial = aSz % AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8053 | while (blocks--) { |
wolfSSL | 15:117db924cf7c | 8054 | XMEMCPY(bigA, a, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8055 | #ifdef LITTLE_ENDIAN_ORDER |
wolfSSL | 15:117db924cf7c | 8056 | ByteReverseWords64(bigA, bigA, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8057 | #endif |
wolfSSL | 15:117db924cf7c | 8058 | x[0] ^= bigA[0]; |
wolfSSL | 15:117db924cf7c | 8059 | x[1] ^= bigA[1]; |
wolfSSL | 15:117db924cf7c | 8060 | GMULT(x, bigH); |
wolfSSL | 15:117db924cf7c | 8061 | a += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8062 | } |
wolfSSL | 15:117db924cf7c | 8063 | if (partial != 0) { |
wolfSSL | 15:117db924cf7c | 8064 | XMEMSET(bigA, 0, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8065 | XMEMCPY(bigA, a, partial); |
wolfSSL | 15:117db924cf7c | 8066 | #ifdef LITTLE_ENDIAN_ORDER |
wolfSSL | 15:117db924cf7c | 8067 | ByteReverseWords64(bigA, bigA, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8068 | #endif |
wolfSSL | 15:117db924cf7c | 8069 | x[0] ^= bigA[0]; |
wolfSSL | 15:117db924cf7c | 8070 | x[1] ^= bigA[1]; |
wolfSSL | 15:117db924cf7c | 8071 | GMULT(x, bigH); |
wolfSSL | 15:117db924cf7c | 8072 | } |
wolfSSL | 15:117db924cf7c | 8073 | } |
wolfSSL | 15:117db924cf7c | 8074 | |
wolfSSL | 15:117db924cf7c | 8075 | /* Hash in C, the Ciphertext */ |
wolfSSL | 15:117db924cf7c | 8076 | if (cSz != 0 && c != NULL) { |
wolfSSL | 15:117db924cf7c | 8077 | word64 bigC[2]; |
wolfSSL | 15:117db924cf7c | 8078 | blocks = cSz / AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8079 | partial = cSz % AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8080 | while (blocks--) { |
wolfSSL | 15:117db924cf7c | 8081 | XMEMCPY(bigC, c, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8082 | #ifdef LITTLE_ENDIAN_ORDER |
wolfSSL | 15:117db924cf7c | 8083 | ByteReverseWords64(bigC, bigC, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8084 | #endif |
wolfSSL | 15:117db924cf7c | 8085 | x[0] ^= bigC[0]; |
wolfSSL | 15:117db924cf7c | 8086 | x[1] ^= bigC[1]; |
wolfSSL | 15:117db924cf7c | 8087 | GMULT(x, bigH); |
wolfSSL | 15:117db924cf7c | 8088 | c += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8089 | } |
wolfSSL | 15:117db924cf7c | 8090 | if (partial != 0) { |
wolfSSL | 15:117db924cf7c | 8091 | XMEMSET(bigC, 0, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8092 | XMEMCPY(bigC, c, partial); |
wolfSSL | 15:117db924cf7c | 8093 | #ifdef LITTLE_ENDIAN_ORDER |
wolfSSL | 15:117db924cf7c | 8094 | ByteReverseWords64(bigC, bigC, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8095 | #endif |
wolfSSL | 15:117db924cf7c | 8096 | x[0] ^= bigC[0]; |
wolfSSL | 15:117db924cf7c | 8097 | x[1] ^= bigC[1]; |
wolfSSL | 15:117db924cf7c | 8098 | GMULT(x, bigH); |
wolfSSL | 15:117db924cf7c | 8099 | } |
wolfSSL | 15:117db924cf7c | 8100 | } |
wolfSSL | 15:117db924cf7c | 8101 | |
wolfSSL | 15:117db924cf7c | 8102 | /* Hash in the lengths in bits of A and C */ |
wolfSSL | 15:117db924cf7c | 8103 | { |
wolfSSL | 15:117db924cf7c | 8104 | word64 len[2]; |
wolfSSL | 15:117db924cf7c | 8105 | len[0] = aSz; len[1] = cSz; |
wolfSSL | 15:117db924cf7c | 8106 | |
wolfSSL | 15:117db924cf7c | 8107 | /* Lengths are in bytes. Convert to bits. */ |
wolfSSL | 15:117db924cf7c | 8108 | len[0] *= 8; |
wolfSSL | 15:117db924cf7c | 8109 | len[1] *= 8; |
wolfSSL | 15:117db924cf7c | 8110 | |
wolfSSL | 15:117db924cf7c | 8111 | x[0] ^= len[0]; |
wolfSSL | 15:117db924cf7c | 8112 | x[1] ^= len[1]; |
wolfSSL | 15:117db924cf7c | 8113 | GMULT(x, bigH); |
wolfSSL | 15:117db924cf7c | 8114 | } |
wolfSSL | 15:117db924cf7c | 8115 | #ifdef LITTLE_ENDIAN_ORDER |
wolfSSL | 15:117db924cf7c | 8116 | ByteReverseWords64(x, x, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8117 | #endif |
wolfSSL | 15:117db924cf7c | 8118 | XMEMCPY(s, x, sSz); |
wolfSSL | 15:117db924cf7c | 8119 | } |
wolfSSL | 15:117db924cf7c | 8120 | #endif /* !FREESCALE_LTC_AES_GCM */ |
wolfSSL | 15:117db924cf7c | 8121 | |
wolfSSL | 15:117db924cf7c | 8122 | /* end defined(WORD64_AVAILABLE) && !defined(GCM_WORD32) */ |
wolfSSL | 15:117db924cf7c | 8123 | #else /* GCM_WORD32 */ |
wolfSSL | 15:117db924cf7c | 8124 | |
wolfSSL | 15:117db924cf7c | 8125 | static void GMULT(word32* X, word32* Y) |
wolfSSL | 15:117db924cf7c | 8126 | { |
wolfSSL | 15:117db924cf7c | 8127 | word32 Z[4] = {0,0,0,0}; |
wolfSSL | 15:117db924cf7c | 8128 | word32 V[4]; |
wolfSSL | 15:117db924cf7c | 8129 | int i, j; |
wolfSSL | 15:117db924cf7c | 8130 | |
wolfSSL | 15:117db924cf7c | 8131 | V[0] = X[0]; V[1] = X[1]; V[2] = X[2]; V[3] = X[3]; |
wolfSSL | 15:117db924cf7c | 8132 | |
wolfSSL | 15:117db924cf7c | 8133 | for (i = 0; i < 4; i++) |
wolfSSL | 15:117db924cf7c | 8134 | { |
wolfSSL | 15:117db924cf7c | 8135 | word32 y = Y[i]; |
wolfSSL | 15:117db924cf7c | 8136 | for (j = 0; j < 32; j++) |
wolfSSL | 15:117db924cf7c | 8137 | { |
wolfSSL | 15:117db924cf7c | 8138 | if (y & 0x80000000) { |
wolfSSL | 15:117db924cf7c | 8139 | Z[0] ^= V[0]; |
wolfSSL | 15:117db924cf7c | 8140 | Z[1] ^= V[1]; |
wolfSSL | 15:117db924cf7c | 8141 | Z[2] ^= V[2]; |
wolfSSL | 15:117db924cf7c | 8142 | Z[3] ^= V[3]; |
wolfSSL | 15:117db924cf7c | 8143 | } |
wolfSSL | 15:117db924cf7c | 8144 | |
wolfSSL | 15:117db924cf7c | 8145 | if (V[3] & 0x00000001) { |
wolfSSL | 15:117db924cf7c | 8146 | V[3] >>= 1; |
wolfSSL | 15:117db924cf7c | 8147 | V[3] |= ((V[2] & 0x00000001) ? 0x80000000 : 0); |
wolfSSL | 15:117db924cf7c | 8148 | V[2] >>= 1; |
wolfSSL | 15:117db924cf7c | 8149 | V[2] |= ((V[1] & 0x00000001) ? 0x80000000 : 0); |
wolfSSL | 15:117db924cf7c | 8150 | V[1] >>= 1; |
wolfSSL | 15:117db924cf7c | 8151 | V[1] |= ((V[0] & 0x00000001) ? 0x80000000 : 0); |
wolfSSL | 15:117db924cf7c | 8152 | V[0] >>= 1; |
wolfSSL | 15:117db924cf7c | 8153 | V[0] ^= 0xE1000000; |
wolfSSL | 15:117db924cf7c | 8154 | } else { |
wolfSSL | 15:117db924cf7c | 8155 | V[3] >>= 1; |
wolfSSL | 15:117db924cf7c | 8156 | V[3] |= ((V[2] & 0x00000001) ? 0x80000000 : 0); |
wolfSSL | 15:117db924cf7c | 8157 | V[2] >>= 1; |
wolfSSL | 15:117db924cf7c | 8158 | V[2] |= ((V[1] & 0x00000001) ? 0x80000000 : 0); |
wolfSSL | 15:117db924cf7c | 8159 | V[1] >>= 1; |
wolfSSL | 15:117db924cf7c | 8160 | V[1] |= ((V[0] & 0x00000001) ? 0x80000000 : 0); |
wolfSSL | 15:117db924cf7c | 8161 | V[0] >>= 1; |
wolfSSL | 15:117db924cf7c | 8162 | } |
wolfSSL | 15:117db924cf7c | 8163 | y <<= 1; |
wolfSSL | 15:117db924cf7c | 8164 | } |
wolfSSL | 15:117db924cf7c | 8165 | } |
wolfSSL | 15:117db924cf7c | 8166 | X[0] = Z[0]; |
wolfSSL | 15:117db924cf7c | 8167 | X[1] = Z[1]; |
wolfSSL | 15:117db924cf7c | 8168 | X[2] = Z[2]; |
wolfSSL | 15:117db924cf7c | 8169 | X[3] = Z[3]; |
wolfSSL | 15:117db924cf7c | 8170 | } |
wolfSSL | 15:117db924cf7c | 8171 | |
wolfSSL | 15:117db924cf7c | 8172 | |
wolfSSL | 15:117db924cf7c | 8173 | void GHASH(Aes* aes, const byte* a, word32 aSz, const byte* c, |
wolfSSL | 15:117db924cf7c | 8174 | word32 cSz, byte* s, word32 sSz) |
wolfSSL | 15:117db924cf7c | 8175 | { |
wolfSSL | 15:117db924cf7c | 8176 | word32 x[4] = {0,0,0,0}; |
wolfSSL | 15:117db924cf7c | 8177 | word32 blocks, partial; |
wolfSSL | 15:117db924cf7c | 8178 | word32 bigH[4]; |
wolfSSL | 15:117db924cf7c | 8179 | |
wolfSSL | 15:117db924cf7c | 8180 | XMEMCPY(bigH, aes->H, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8181 | #ifdef LITTLE_ENDIAN_ORDER |
wolfSSL | 15:117db924cf7c | 8182 | ByteReverseWords(bigH, bigH, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8183 | #endif |
wolfSSL | 15:117db924cf7c | 8184 | |
wolfSSL | 15:117db924cf7c | 8185 | /* Hash in A, the Additional Authentication Data */ |
wolfSSL | 15:117db924cf7c | 8186 | if (aSz != 0 && a != NULL) { |
wolfSSL | 15:117db924cf7c | 8187 | word32 bigA[4]; |
wolfSSL | 15:117db924cf7c | 8188 | blocks = aSz / AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8189 | partial = aSz % AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8190 | while (blocks--) { |
wolfSSL | 15:117db924cf7c | 8191 | XMEMCPY(bigA, a, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8192 | #ifdef LITTLE_ENDIAN_ORDER |
wolfSSL | 15:117db924cf7c | 8193 | ByteReverseWords(bigA, bigA, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8194 | #endif |
wolfSSL | 15:117db924cf7c | 8195 | x[0] ^= bigA[0]; |
wolfSSL | 15:117db924cf7c | 8196 | x[1] ^= bigA[1]; |
wolfSSL | 15:117db924cf7c | 8197 | x[2] ^= bigA[2]; |
wolfSSL | 15:117db924cf7c | 8198 | x[3] ^= bigA[3]; |
wolfSSL | 15:117db924cf7c | 8199 | GMULT(x, bigH); |
wolfSSL | 15:117db924cf7c | 8200 | a += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8201 | } |
wolfSSL | 15:117db924cf7c | 8202 | if (partial != 0) { |
wolfSSL | 15:117db924cf7c | 8203 | XMEMSET(bigA, 0, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8204 | XMEMCPY(bigA, a, partial); |
wolfSSL | 15:117db924cf7c | 8205 | #ifdef LITTLE_ENDIAN_ORDER |
wolfSSL | 15:117db924cf7c | 8206 | ByteReverseWords(bigA, bigA, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8207 | #endif |
wolfSSL | 15:117db924cf7c | 8208 | x[0] ^= bigA[0]; |
wolfSSL | 15:117db924cf7c | 8209 | x[1] ^= bigA[1]; |
wolfSSL | 15:117db924cf7c | 8210 | x[2] ^= bigA[2]; |
wolfSSL | 15:117db924cf7c | 8211 | x[3] ^= bigA[3]; |
wolfSSL | 15:117db924cf7c | 8212 | GMULT(x, bigH); |
wolfSSL | 15:117db924cf7c | 8213 | } |
wolfSSL | 15:117db924cf7c | 8214 | } |
wolfSSL | 15:117db924cf7c | 8215 | |
wolfSSL | 15:117db924cf7c | 8216 | /* Hash in C, the Ciphertext */ |
wolfSSL | 15:117db924cf7c | 8217 | if (cSz != 0 && c != NULL) { |
wolfSSL | 15:117db924cf7c | 8218 | word32 bigC[4]; |
wolfSSL | 15:117db924cf7c | 8219 | blocks = cSz / AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8220 | partial = cSz % AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8221 | while (blocks--) { |
wolfSSL | 15:117db924cf7c | 8222 | XMEMCPY(bigC, c, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8223 | #ifdef LITTLE_ENDIAN_ORDER |
wolfSSL | 15:117db924cf7c | 8224 | ByteReverseWords(bigC, bigC, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8225 | #endif |
wolfSSL | 15:117db924cf7c | 8226 | x[0] ^= bigC[0]; |
wolfSSL | 15:117db924cf7c | 8227 | x[1] ^= bigC[1]; |
wolfSSL | 15:117db924cf7c | 8228 | x[2] ^= bigC[2]; |
wolfSSL | 15:117db924cf7c | 8229 | x[3] ^= bigC[3]; |
wolfSSL | 15:117db924cf7c | 8230 | GMULT(x, bigH); |
wolfSSL | 15:117db924cf7c | 8231 | c += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8232 | } |
wolfSSL | 15:117db924cf7c | 8233 | if (partial != 0) { |
wolfSSL | 15:117db924cf7c | 8234 | XMEMSET(bigC, 0, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8235 | XMEMCPY(bigC, c, partial); |
wolfSSL | 15:117db924cf7c | 8236 | #ifdef LITTLE_ENDIAN_ORDER |
wolfSSL | 15:117db924cf7c | 8237 | ByteReverseWords(bigC, bigC, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8238 | #endif |
wolfSSL | 15:117db924cf7c | 8239 | x[0] ^= bigC[0]; |
wolfSSL | 15:117db924cf7c | 8240 | x[1] ^= bigC[1]; |
wolfSSL | 15:117db924cf7c | 8241 | x[2] ^= bigC[2]; |
wolfSSL | 15:117db924cf7c | 8242 | x[3] ^= bigC[3]; |
wolfSSL | 15:117db924cf7c | 8243 | GMULT(x, bigH); |
wolfSSL | 15:117db924cf7c | 8244 | } |
wolfSSL | 15:117db924cf7c | 8245 | } |
wolfSSL | 15:117db924cf7c | 8246 | |
wolfSSL | 15:117db924cf7c | 8247 | /* Hash in the lengths in bits of A and C */ |
wolfSSL | 15:117db924cf7c | 8248 | { |
wolfSSL | 15:117db924cf7c | 8249 | word32 len[4]; |
wolfSSL | 15:117db924cf7c | 8250 | |
wolfSSL | 15:117db924cf7c | 8251 | /* Lengths are in bytes. Convert to bits. */ |
wolfSSL | 15:117db924cf7c | 8252 | len[0] = (aSz >> (8*sizeof(aSz) - 3)); |
wolfSSL | 15:117db924cf7c | 8253 | len[1] = aSz << 3; |
wolfSSL | 15:117db924cf7c | 8254 | len[2] = (cSz >> (8*sizeof(cSz) - 3)); |
wolfSSL | 15:117db924cf7c | 8255 | len[3] = cSz << 3; |
wolfSSL | 15:117db924cf7c | 8256 | |
wolfSSL | 15:117db924cf7c | 8257 | x[0] ^= len[0]; |
wolfSSL | 15:117db924cf7c | 8258 | x[1] ^= len[1]; |
wolfSSL | 15:117db924cf7c | 8259 | x[2] ^= len[2]; |
wolfSSL | 15:117db924cf7c | 8260 | x[3] ^= len[3]; |
wolfSSL | 15:117db924cf7c | 8261 | GMULT(x, bigH); |
wolfSSL | 15:117db924cf7c | 8262 | } |
wolfSSL | 15:117db924cf7c | 8263 | #ifdef LITTLE_ENDIAN_ORDER |
wolfSSL | 15:117db924cf7c | 8264 | ByteReverseWords(x, x, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8265 | #endif |
wolfSSL | 15:117db924cf7c | 8266 | XMEMCPY(s, x, sSz); |
wolfSSL | 15:117db924cf7c | 8267 | } |
wolfSSL | 15:117db924cf7c | 8268 | |
wolfSSL | 15:117db924cf7c | 8269 | #endif /* end GCM_WORD32 */ |
wolfSSL | 15:117db924cf7c | 8270 | |
wolfSSL | 15:117db924cf7c | 8271 | |
wolfSSL | 15:117db924cf7c | 8272 | #if !defined(WOLFSSL_XILINX_CRYPT) |
wolfSSL | 15:117db924cf7c | 8273 | #ifdef FREESCALE_LTC_AES_GCM |
wolfSSL | 15:117db924cf7c | 8274 | int wc_AesGcmEncrypt(Aes* aes, byte* out, const byte* in, word32 sz, |
wolfSSL | 15:117db924cf7c | 8275 | const byte* iv, word32 ivSz, |
wolfSSL | 15:117db924cf7c | 8276 | byte* authTag, word32 authTagSz, |
wolfSSL | 15:117db924cf7c | 8277 | const byte* authIn, word32 authInSz) |
wolfSSL | 15:117db924cf7c | 8278 | { |
wolfSSL | 15:117db924cf7c | 8279 | status_t status; |
wolfSSL | 15:117db924cf7c | 8280 | word32 keySize; |
wolfSSL | 15:117db924cf7c | 8281 | |
wolfSSL | 15:117db924cf7c | 8282 | /* argument checks */ |
wolfSSL | 15:117db924cf7c | 8283 | if (aes == NULL || authTagSz > AES_BLOCK_SIZE) { |
wolfSSL | 15:117db924cf7c | 8284 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 8285 | } |
wolfSSL | 15:117db924cf7c | 8286 | |
wolfSSL | 15:117db924cf7c | 8287 | if (authTagSz < WOLFSSL_MIN_AUTH_TAG_SZ) { |
wolfSSL | 15:117db924cf7c | 8288 | WOLFSSL_MSG("GcmEncrypt authTagSz too small error"); |
wolfSSL | 15:117db924cf7c | 8289 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 8290 | } |
wolfSSL | 15:117db924cf7c | 8291 | |
wolfSSL | 15:117db924cf7c | 8292 | status = wc_AesGetKeySize(aes, &keySize); |
wolfSSL | 15:117db924cf7c | 8293 | if (status) |
wolfSSL | 15:117db924cf7c | 8294 | return status; |
wolfSSL | 15:117db924cf7c | 8295 | |
wolfSSL | 15:117db924cf7c | 8296 | status = LTC_AES_EncryptTagGcm(LTC_BASE, in, out, sz, iv, ivSz, |
wolfSSL | 15:117db924cf7c | 8297 | authIn, authInSz, (byte*)aes->key, keySize, authTag, authTagSz); |
wolfSSL | 15:117db924cf7c | 8298 | |
wolfSSL | 15:117db924cf7c | 8299 | return (status == kStatus_Success) ? 0 : AES_GCM_AUTH_E; |
wolfSSL | 15:117db924cf7c | 8300 | } |
wolfSSL | 15:117db924cf7c | 8301 | #else |
wolfSSL | 15:117db924cf7c | 8302 | #if defined(STM32_CRYPTO) && (defined(WOLFSSL_STM32F4) || \ |
wolfSSL | 15:117db924cf7c | 8303 | defined(WOLFSSL_STM32F7) || \ |
wolfSSL | 15:117db924cf7c | 8304 | defined(WOLFSSL_STM32L4)) |
wolfSSL | 15:117db924cf7c | 8305 | |
wolfSSL | 15:117db924cf7c | 8306 | static WC_INLINE int wc_AesGcmEncrypt_STM32(Aes* aes, byte* out, const byte* in, |
wolfSSL | 15:117db924cf7c | 8307 | word32 sz, const byte* iv, word32 ivSz, |
wolfSSL | 15:117db924cf7c | 8308 | byte* authTag, word32 authTagSz, |
wolfSSL | 15:117db924cf7c | 8309 | const byte* authIn, word32 authInSz) |
wolfSSL | 15:117db924cf7c | 8310 | { |
wolfSSL | 15:117db924cf7c | 8311 | int ret; |
wolfSSL | 15:117db924cf7c | 8312 | word32 keySize; |
wolfSSL | 15:117db924cf7c | 8313 | byte initialCounter[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 8314 | #ifdef WOLFSSL_STM32_CUBEMX |
wolfSSL | 15:117db924cf7c | 8315 | CRYP_HandleTypeDef hcryp; |
wolfSSL | 15:117db924cf7c | 8316 | #else |
wolfSSL | 15:117db924cf7c | 8317 | byte keyCopy[AES_BLOCK_SIZE * 2]; |
wolfSSL | 15:117db924cf7c | 8318 | #endif /* WOLFSSL_STM32_CUBEMX */ |
wolfSSL | 15:117db924cf7c | 8319 | int status = 0; |
wolfSSL | 15:117db924cf7c | 8320 | byte* authInPadded = NULL; |
wolfSSL | 15:117db924cf7c | 8321 | byte tag[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 8322 | int authPadSz; |
wolfSSL | 15:117db924cf7c | 8323 | |
wolfSSL | 15:117db924cf7c | 8324 | ret = wc_AesGetKeySize(aes, &keySize); |
wolfSSL | 15:117db924cf7c | 8325 | if (ret != 0) |
wolfSSL | 15:117db924cf7c | 8326 | return ret; |
wolfSSL | 15:117db924cf7c | 8327 | |
wolfSSL | 15:117db924cf7c | 8328 | XMEMSET(initialCounter, 0, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8329 | XMEMCPY(initialCounter, iv, ivSz); |
wolfSSL | 15:117db924cf7c | 8330 | initialCounter[AES_BLOCK_SIZE - 1] = STM32_GCM_IV_START; |
wolfSSL | 15:117db924cf7c | 8331 | |
wolfSSL | 15:117db924cf7c | 8332 | /* pad authIn if it is not a block multiple */ |
wolfSSL | 15:117db924cf7c | 8333 | if ((authInSz % AES_BLOCK_SIZE) != 0) { |
wolfSSL | 15:117db924cf7c | 8334 | authPadSz = ((authInSz / AES_BLOCK_SIZE) + 1) * AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8335 | /* Need to pad the AAD to a full block with zeros. */ |
wolfSSL | 15:117db924cf7c | 8336 | authInPadded = XMALLOC(authPadSz, aes->heap, DYNAMIC_TYPE_TMP_BUFFER); |
wolfSSL | 15:117db924cf7c | 8337 | if (authInPadded == NULL) { |
wolfSSL | 15:117db924cf7c | 8338 | return MEMORY_E; |
wolfSSL | 15:117db924cf7c | 8339 | } |
wolfSSL | 15:117db924cf7c | 8340 | XMEMSET(authInPadded, 0, authPadSz); |
wolfSSL | 15:117db924cf7c | 8341 | XMEMCPY(authInPadded, authIn, authInSz); |
wolfSSL | 15:117db924cf7c | 8342 | } else { |
wolfSSL | 15:117db924cf7c | 8343 | authPadSz = authInSz; |
wolfSSL | 15:117db924cf7c | 8344 | authInPadded = (byte*)authIn; |
wolfSSL | 15:117db924cf7c | 8345 | } |
wolfSSL | 15:117db924cf7c | 8346 | |
wolfSSL | 15:117db924cf7c | 8347 | |
wolfSSL | 15:117db924cf7c | 8348 | #ifdef WOLFSSL_STM32_CUBEMX |
wolfSSL | 15:117db924cf7c | 8349 | XMEMSET(&hcryp, 0, sizeof(CRYP_HandleTypeDef)); |
wolfSSL | 15:117db924cf7c | 8350 | switch (keySize) { |
wolfSSL | 15:117db924cf7c | 8351 | case 16: /* 128-bit key */ |
wolfSSL | 15:117db924cf7c | 8352 | hcryp.Init.KeySize = CRYP_KEYSIZE_128B; |
wolfSSL | 15:117db924cf7c | 8353 | break; |
wolfSSL | 15:117db924cf7c | 8354 | #ifdef CRYP_KEYSIZE_192B |
wolfSSL | 15:117db924cf7c | 8355 | case 24: /* 192-bit key */ |
wolfSSL | 15:117db924cf7c | 8356 | hcryp.Init.KeySize = CRYP_KEYSIZE_192B; |
wolfSSL | 15:117db924cf7c | 8357 | break; |
wolfSSL | 15:117db924cf7c | 8358 | #endif |
wolfSSL | 15:117db924cf7c | 8359 | case 32: /* 256-bit key */ |
wolfSSL | 15:117db924cf7c | 8360 | hcryp.Init.KeySize = CRYP_KEYSIZE_256B; |
wolfSSL | 15:117db924cf7c | 8361 | break; |
wolfSSL | 15:117db924cf7c | 8362 | default: |
wolfSSL | 15:117db924cf7c | 8363 | break; |
wolfSSL | 15:117db924cf7c | 8364 | } |
wolfSSL | 15:117db924cf7c | 8365 | hcryp.Instance = CRYP; |
wolfSSL | 15:117db924cf7c | 8366 | hcryp.Init.DataType = CRYP_DATATYPE_8B; |
wolfSSL | 15:117db924cf7c | 8367 | hcryp.Init.pKey = (byte*)aes->key; |
wolfSSL | 15:117db924cf7c | 8368 | hcryp.Init.pInitVect = initialCounter; |
wolfSSL | 15:117db924cf7c | 8369 | hcryp.Init.Header = authInPadded; |
wolfSSL | 15:117db924cf7c | 8370 | hcryp.Init.HeaderSize = authInSz; |
wolfSSL | 15:117db924cf7c | 8371 | |
wolfSSL | 15:117db924cf7c | 8372 | #ifdef WOLFSSL_STM32L4 |
wolfSSL | 15:117db924cf7c | 8373 | /* Set the CRYP parameters */ |
wolfSSL | 15:117db924cf7c | 8374 | hcryp.Init.ChainingMode = CRYP_CHAINMODE_AES_GCM_GMAC; |
wolfSSL | 15:117db924cf7c | 8375 | hcryp.Init.OperatingMode = CRYP_ALGOMODE_ENCRYPT; |
wolfSSL | 15:117db924cf7c | 8376 | hcryp.Init.GCMCMACPhase = CRYP_INIT_PHASE; |
wolfSSL | 15:117db924cf7c | 8377 | HAL_CRYP_Init(&hcryp); |
wolfSSL | 15:117db924cf7c | 8378 | |
wolfSSL | 15:117db924cf7c | 8379 | /* GCM init phase */ |
wolfSSL | 15:117db924cf7c | 8380 | status = HAL_CRYPEx_AES_Auth(&hcryp, NULL, 0, NULL, STM32_HAL_TIMEOUT); |
wolfSSL | 15:117db924cf7c | 8381 | if (status == HAL_OK) { |
wolfSSL | 15:117db924cf7c | 8382 | /* GCM header phase */ |
wolfSSL | 15:117db924cf7c | 8383 | hcryp.Init.GCMCMACPhase = CRYP_HEADER_PHASE; |
wolfSSL | 15:117db924cf7c | 8384 | status = HAL_CRYPEx_AES_Auth(&hcryp, NULL, 0, NULL, STM32_HAL_TIMEOUT); |
wolfSSL | 15:117db924cf7c | 8385 | if (status == HAL_OK) { |
wolfSSL | 15:117db924cf7c | 8386 | /* GCM payload phase */ |
wolfSSL | 15:117db924cf7c | 8387 | hcryp.Init.GCMCMACPhase = CRYP_PAYLOAD_PHASE; |
wolfSSL | 15:117db924cf7c | 8388 | status = HAL_CRYPEx_AES_Auth(&hcryp, (byte*)in, sz, out, STM32_HAL_TIMEOUT); |
wolfSSL | 15:117db924cf7c | 8389 | if (status == HAL_OK) { |
wolfSSL | 15:117db924cf7c | 8390 | /* GCM final phase */ |
wolfSSL | 15:117db924cf7c | 8391 | hcryp.Init.GCMCMACPhase = CRYP_FINAL_PHASE; |
wolfSSL | 15:117db924cf7c | 8392 | status = HAL_CRYPEx_AES_Auth(&hcryp, NULL, sz, tag, STM32_HAL_TIMEOUT); |
wolfSSL | 15:117db924cf7c | 8393 | } |
wolfSSL | 15:117db924cf7c | 8394 | } |
wolfSSL | 15:117db924cf7c | 8395 | } |
wolfSSL | 15:117db924cf7c | 8396 | #else |
wolfSSL | 15:117db924cf7c | 8397 | HAL_CRYP_Init(&hcryp); |
wolfSSL | 15:117db924cf7c | 8398 | |
wolfSSL | 15:117db924cf7c | 8399 | status = HAL_CRYPEx_AESGCM_Encrypt(&hcryp, (byte*)in, sz, |
wolfSSL | 15:117db924cf7c | 8400 | out, STM32_HAL_TIMEOUT); |
wolfSSL | 15:117db924cf7c | 8401 | /* Compute the authTag */ |
wolfSSL | 15:117db924cf7c | 8402 | if (status == HAL_OK) { |
wolfSSL | 15:117db924cf7c | 8403 | status = HAL_CRYPEx_AESGCM_Finish(&hcryp, sz, tag, STM32_HAL_TIMEOUT); |
wolfSSL | 15:117db924cf7c | 8404 | } |
wolfSSL | 15:117db924cf7c | 8405 | #endif |
wolfSSL | 15:117db924cf7c | 8406 | |
wolfSSL | 15:117db924cf7c | 8407 | if (status != HAL_OK) |
wolfSSL | 15:117db924cf7c | 8408 | ret = AES_GCM_AUTH_E; |
wolfSSL | 15:117db924cf7c | 8409 | HAL_CRYP_DeInit(&hcryp); |
wolfSSL | 15:117db924cf7c | 8410 | #else |
wolfSSL | 15:117db924cf7c | 8411 | ByteReverseWords((word32*)keyCopy, (word32*)aes->key, keySize); |
wolfSSL | 15:117db924cf7c | 8412 | status = CRYP_AES_GCM(MODE_ENCRYPT, (uint8_t*)initialCounter, |
wolfSSL | 15:117db924cf7c | 8413 | (uint8_t*)keyCopy, keySize * 8, |
wolfSSL | 15:117db924cf7c | 8414 | (uint8_t*)in, sz, |
wolfSSL | 15:117db924cf7c | 8415 | (uint8_t*)authInPadded,authInSz, |
wolfSSL | 15:117db924cf7c | 8416 | (uint8_t*)out, tag); |
wolfSSL | 15:117db924cf7c | 8417 | if (status != SUCCESS) |
wolfSSL | 15:117db924cf7c | 8418 | ret = AES_GCM_AUTH_E; |
wolfSSL | 15:117db924cf7c | 8419 | #endif /* WOLFSSL_STM32_CUBEMX */ |
wolfSSL | 15:117db924cf7c | 8420 | |
wolfSSL | 15:117db924cf7c | 8421 | /* authTag may be shorter than AES_BLOCK_SZ, store separately */ |
wolfSSL | 15:117db924cf7c | 8422 | if (ret == 0) |
wolfSSL | 15:117db924cf7c | 8423 | XMEMCPY(authTag, tag, authTagSz); |
wolfSSL | 15:117db924cf7c | 8424 | |
wolfSSL | 15:117db924cf7c | 8425 | /* We only allocate extra memory if authInPadded is not a multiple of AES_BLOCK_SZ */ |
wolfSSL | 15:117db924cf7c | 8426 | if (authInPadded != NULL && authInSz != authPadSz) { |
wolfSSL | 15:117db924cf7c | 8427 | XFREE(authInPadded, aes->heap, DYNAMIC_TYPE_TMP_BUFFER); |
wolfSSL | 15:117db924cf7c | 8428 | } |
wolfSSL | 15:117db924cf7c | 8429 | |
wolfSSL | 15:117db924cf7c | 8430 | return ret; |
wolfSSL | 15:117db924cf7c | 8431 | } |
wolfSSL | 15:117db924cf7c | 8432 | #endif /* STM32_CRYPTO */ |
wolfSSL | 15:117db924cf7c | 8433 | |
wolfSSL | 15:117db924cf7c | 8434 | #ifdef WOLFSSL_AESNI |
wolfSSL | 15:117db924cf7c | 8435 | int AES_GCM_encrypt_C(Aes* aes, byte* out, const byte* in, word32 sz, |
wolfSSL | 15:117db924cf7c | 8436 | const byte* iv, word32 ivSz, |
wolfSSL | 15:117db924cf7c | 8437 | byte* authTag, word32 authTagSz, |
wolfSSL | 15:117db924cf7c | 8438 | const byte* authIn, word32 authInSz); |
wolfSSL | 15:117db924cf7c | 8439 | #else |
wolfSSL | 15:117db924cf7c | 8440 | static |
wolfSSL | 15:117db924cf7c | 8441 | #endif |
wolfSSL | 15:117db924cf7c | 8442 | int AES_GCM_encrypt_C(Aes* aes, byte* out, const byte* in, word32 sz, |
wolfSSL | 15:117db924cf7c | 8443 | const byte* iv, word32 ivSz, |
wolfSSL | 15:117db924cf7c | 8444 | byte* authTag, word32 authTagSz, |
wolfSSL | 15:117db924cf7c | 8445 | const byte* authIn, word32 authInSz) |
wolfSSL | 15:117db924cf7c | 8446 | { |
wolfSSL | 15:117db924cf7c | 8447 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 8448 | word32 blocks = sz / AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8449 | word32 partial = sz % AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8450 | const byte* p = in; |
wolfSSL | 15:117db924cf7c | 8451 | byte* c = out; |
wolfSSL | 15:117db924cf7c | 8452 | byte counter[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 8453 | byte initialCounter[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 8454 | byte *ctr; |
wolfSSL | 15:117db924cf7c | 8455 | byte scratch[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 8456 | |
wolfSSL | 15:117db924cf7c | 8457 | ctr = counter; |
wolfSSL | 15:117db924cf7c | 8458 | XMEMSET(initialCounter, 0, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8459 | if (ivSz == GCM_NONCE_MID_SZ) { |
wolfSSL | 15:117db924cf7c | 8460 | XMEMCPY(initialCounter, iv, ivSz); |
wolfSSL | 15:117db924cf7c | 8461 | initialCounter[AES_BLOCK_SIZE - 1] = 1; |
wolfSSL | 15:117db924cf7c | 8462 | } |
wolfSSL | 15:117db924cf7c | 8463 | else { |
wolfSSL | 15:117db924cf7c | 8464 | GHASH(aes, NULL, 0, iv, ivSz, initialCounter, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8465 | } |
wolfSSL | 15:117db924cf7c | 8466 | XMEMCPY(ctr, initialCounter, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8467 | |
wolfSSL | 15:117db924cf7c | 8468 | #ifdef WOLFSSL_PIC32MZ_CRYPT |
wolfSSL | 15:117db924cf7c | 8469 | if (blocks) { |
wolfSSL | 15:117db924cf7c | 8470 | /* use intitial IV for PIC32 HW, but don't use it below */ |
wolfSSL | 15:117db924cf7c | 8471 | XMEMCPY(aes->reg, ctr, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8472 | |
wolfSSL | 15:117db924cf7c | 8473 | ret = wc_Pic32AesCrypt( |
wolfSSL | 15:117db924cf7c | 8474 | aes->key, aes->keylen, aes->reg, AES_BLOCK_SIZE, |
wolfSSL | 15:117db924cf7c | 8475 | out, in, (blocks * AES_BLOCK_SIZE), |
wolfSSL | 15:117db924cf7c | 8476 | PIC32_ENCRYPTION, PIC32_ALGO_AES, PIC32_CRYPTOALGO_AES_GCM); |
wolfSSL | 15:117db924cf7c | 8477 | if (ret != 0) |
wolfSSL | 15:117db924cf7c | 8478 | return ret; |
wolfSSL | 15:117db924cf7c | 8479 | } |
wolfSSL | 15:117db924cf7c | 8480 | /* process remainder using partial handling */ |
wolfSSL | 15:117db924cf7c | 8481 | #endif |
wolfSSL | 15:117db924cf7c | 8482 | |
wolfSSL | 15:117db924cf7c | 8483 | #if defined(HAVE_AES_ECB) && !defined(WOLFSSL_PIC32MZ_CRYPT) |
wolfSSL | 15:117db924cf7c | 8484 | /* some hardware acceleration can gain performance from doing AES encryption |
wolfSSL | 15:117db924cf7c | 8485 | * of the whole buffer at once */ |
wolfSSL | 15:117db924cf7c | 8486 | if (c != p) { /* can not handle inline encryption */ |
wolfSSL | 15:117db924cf7c | 8487 | while (blocks--) { |
wolfSSL | 15:117db924cf7c | 8488 | IncrementGcmCounter(ctr); |
wolfSSL | 15:117db924cf7c | 8489 | XMEMCPY(c, ctr, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8490 | c += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8491 | } |
wolfSSL | 15:117db924cf7c | 8492 | |
wolfSSL | 15:117db924cf7c | 8493 | /* reset number of blocks and then do encryption */ |
wolfSSL | 15:117db924cf7c | 8494 | blocks = sz / AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8495 | wc_AesEcbEncrypt(aes, out, out, AES_BLOCK_SIZE * blocks); |
wolfSSL | 15:117db924cf7c | 8496 | xorbuf(out, p, AES_BLOCK_SIZE * blocks); |
wolfSSL | 15:117db924cf7c | 8497 | p += AES_BLOCK_SIZE * blocks; |
wolfSSL | 15:117db924cf7c | 8498 | } |
wolfSSL | 15:117db924cf7c | 8499 | else |
wolfSSL | 15:117db924cf7c | 8500 | #endif /* HAVE_AES_ECB */ |
wolfSSL | 15:117db924cf7c | 8501 | |
wolfSSL | 15:117db924cf7c | 8502 | while (blocks--) { |
wolfSSL | 15:117db924cf7c | 8503 | IncrementGcmCounter(ctr); |
wolfSSL | 15:117db924cf7c | 8504 | #ifndef WOLFSSL_PIC32MZ_CRYPT |
wolfSSL | 15:117db924cf7c | 8505 | wc_AesEncrypt(aes, ctr, scratch); |
wolfSSL | 15:117db924cf7c | 8506 | xorbuf(scratch, p, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8507 | XMEMCPY(c, scratch, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8508 | #endif |
wolfSSL | 15:117db924cf7c | 8509 | p += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8510 | c += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8511 | } |
wolfSSL | 15:117db924cf7c | 8512 | |
wolfSSL | 15:117db924cf7c | 8513 | if (partial != 0) { |
wolfSSL | 15:117db924cf7c | 8514 | IncrementGcmCounter(ctr); |
wolfSSL | 15:117db924cf7c | 8515 | wc_AesEncrypt(aes, ctr, scratch); |
wolfSSL | 15:117db924cf7c | 8516 | xorbuf(scratch, p, partial); |
wolfSSL | 15:117db924cf7c | 8517 | XMEMCPY(c, scratch, partial); |
wolfSSL | 15:117db924cf7c | 8518 | } |
wolfSSL | 15:117db924cf7c | 8519 | |
wolfSSL | 15:117db924cf7c | 8520 | GHASH(aes, authIn, authInSz, out, sz, authTag, authTagSz); |
wolfSSL | 15:117db924cf7c | 8521 | wc_AesEncrypt(aes, initialCounter, scratch); |
wolfSSL | 15:117db924cf7c | 8522 | xorbuf(authTag, scratch, authTagSz); |
wolfSSL | 15:117db924cf7c | 8523 | |
wolfSSL | 15:117db924cf7c | 8524 | return ret; |
wolfSSL | 15:117db924cf7c | 8525 | } |
wolfSSL | 15:117db924cf7c | 8526 | |
wolfSSL | 15:117db924cf7c | 8527 | int wc_AesGcmEncrypt(Aes* aes, byte* out, const byte* in, word32 sz, |
wolfSSL | 15:117db924cf7c | 8528 | const byte* iv, word32 ivSz, |
wolfSSL | 15:117db924cf7c | 8529 | byte* authTag, word32 authTagSz, |
wolfSSL | 15:117db924cf7c | 8530 | const byte* authIn, word32 authInSz) |
wolfSSL | 15:117db924cf7c | 8531 | { |
wolfSSL | 15:117db924cf7c | 8532 | /* argument checks */ |
wolfSSL | 15:117db924cf7c | 8533 | if (aes == NULL || authTagSz > AES_BLOCK_SIZE) { |
wolfSSL | 15:117db924cf7c | 8534 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 8535 | } |
wolfSSL | 15:117db924cf7c | 8536 | |
wolfSSL | 15:117db924cf7c | 8537 | if (authTagSz < WOLFSSL_MIN_AUTH_TAG_SZ) { |
wolfSSL | 15:117db924cf7c | 8538 | WOLFSSL_MSG("GcmEncrypt authTagSz too small error"); |
wolfSSL | 15:117db924cf7c | 8539 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 8540 | } |
wolfSSL | 15:117db924cf7c | 8541 | |
wolfSSL | 15:117db924cf7c | 8542 | #if defined(STM32_CRYPTO) && (defined(WOLFSSL_STM32F4) || \ |
wolfSSL | 15:117db924cf7c | 8543 | defined(WOLFSSL_STM32F7) || \ |
wolfSSL | 15:117db924cf7c | 8544 | defined(WOLFSSL_STM32L4)) |
wolfSSL | 15:117db924cf7c | 8545 | |
wolfSSL | 15:117db924cf7c | 8546 | /* additional argument checks - STM32 HW only supports 12 byte IV */ |
wolfSSL | 15:117db924cf7c | 8547 | if (ivSz != GCM_NONCE_MID_SZ) { |
wolfSSL | 15:117db924cf7c | 8548 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 8549 | } |
wolfSSL | 15:117db924cf7c | 8550 | |
wolfSSL | 15:117db924cf7c | 8551 | /* STM32 HW AES-GCM requires / assumes inputs are a multiple of block size. |
wolfSSL | 15:117db924cf7c | 8552 | * We can avoid this by zero padding (authIn) AAD, but zero-padded plaintext |
wolfSSL | 15:117db924cf7c | 8553 | * will be encrypted and output incorrectly, causing a bad authTag. |
wolfSSL | 15:117db924cf7c | 8554 | * We will use HW accelerated AES-GCM if plain%AES_BLOCK_SZ==0. |
wolfSSL | 15:117db924cf7c | 8555 | * Otherwise, we will use accelerated AES_CTR for encrypt, and then |
wolfSSL | 15:117db924cf7c | 8556 | * perform GHASH in software. |
wolfSSL | 15:117db924cf7c | 8557 | * See NIST SP 800-38D */ |
wolfSSL | 15:117db924cf7c | 8558 | |
wolfSSL | 15:117db924cf7c | 8559 | /* Plain text is a multiple of block size, so use HW-Accelerated AES_GCM */ |
wolfSSL | 15:117db924cf7c | 8560 | if (sz % AES_BLOCK_SIZE == 0) { |
wolfSSL | 15:117db924cf7c | 8561 | return wc_AesGcmEncrypt_STM32(aes, out, in, sz, iv, ivSz, |
wolfSSL | 15:117db924cf7c | 8562 | authTag, authTagSz, authIn, authInSz); |
wolfSSL | 15:117db924cf7c | 8563 | } |
wolfSSL | 15:117db924cf7c | 8564 | #endif |
wolfSSL | 15:117db924cf7c | 8565 | |
wolfSSL | 15:117db924cf7c | 8566 | #if defined(WOLFSSL_ASYNC_CRYPT) && defined(WC_ASYNC_ENABLE_AES) |
wolfSSL | 15:117db924cf7c | 8567 | /* if async and byte count above threshold */ |
wolfSSL | 15:117db924cf7c | 8568 | /* only 12-byte IV is supported in HW */ |
wolfSSL | 15:117db924cf7c | 8569 | if (aes->asyncDev.marker == WOLFSSL_ASYNC_MARKER_AES && |
wolfSSL | 15:117db924cf7c | 8570 | sz >= WC_ASYNC_THRESH_AES_GCM && ivSz == GCM_NONCE_MID_SZ) { |
wolfSSL | 15:117db924cf7c | 8571 | #if defined(HAVE_CAVIUM) |
wolfSSL | 15:117db924cf7c | 8572 | #ifdef HAVE_CAVIUM_V |
wolfSSL | 15:117db924cf7c | 8573 | if (authInSz == 20) { /* Nitrox V GCM is only working with 20 byte AAD */ |
wolfSSL | 15:117db924cf7c | 8574 | return NitroxAesGcmEncrypt(aes, out, in, sz, |
wolfSSL | 15:117db924cf7c | 8575 | (const byte*)aes->asyncKey, aes->keylen, iv, ivSz, |
wolfSSL | 15:117db924cf7c | 8576 | authTag, authTagSz, authIn, authInSz); |
wolfSSL | 15:117db924cf7c | 8577 | } |
wolfSSL | 15:117db924cf7c | 8578 | #endif |
wolfSSL | 15:117db924cf7c | 8579 | #elif defined(HAVE_INTEL_QA) |
wolfSSL | 15:117db924cf7c | 8580 | return IntelQaSymAesGcmEncrypt(&aes->asyncDev, out, in, sz, |
wolfSSL | 15:117db924cf7c | 8581 | (const byte*)aes->asyncKey, aes->keylen, iv, ivSz, |
wolfSSL | 15:117db924cf7c | 8582 | authTag, authTagSz, authIn, authInSz); |
wolfSSL | 15:117db924cf7c | 8583 | #else /* WOLFSSL_ASYNC_CRYPT_TEST */ |
wolfSSL | 15:117db924cf7c | 8584 | if (wc_AsyncTestInit(&aes->asyncDev, ASYNC_TEST_AES_GCM_ENCRYPT)) { |
wolfSSL | 15:117db924cf7c | 8585 | WC_ASYNC_TEST* testDev = &aes->asyncDev.test; |
wolfSSL | 15:117db924cf7c | 8586 | testDev->aes.aes = aes; |
wolfSSL | 15:117db924cf7c | 8587 | testDev->aes.out = out; |
wolfSSL | 15:117db924cf7c | 8588 | testDev->aes.in = in; |
wolfSSL | 15:117db924cf7c | 8589 | testDev->aes.sz = sz; |
wolfSSL | 15:117db924cf7c | 8590 | testDev->aes.iv = iv; |
wolfSSL | 15:117db924cf7c | 8591 | testDev->aes.ivSz = ivSz; |
wolfSSL | 15:117db924cf7c | 8592 | testDev->aes.authTag = authTag; |
wolfSSL | 15:117db924cf7c | 8593 | testDev->aes.authTagSz = authTagSz; |
wolfSSL | 15:117db924cf7c | 8594 | testDev->aes.authIn = authIn; |
wolfSSL | 15:117db924cf7c | 8595 | testDev->aes.authInSz = authInSz; |
wolfSSL | 15:117db924cf7c | 8596 | return WC_PENDING_E; |
wolfSSL | 15:117db924cf7c | 8597 | } |
wolfSSL | 15:117db924cf7c | 8598 | #endif |
wolfSSL | 15:117db924cf7c | 8599 | } |
wolfSSL | 15:117db924cf7c | 8600 | #endif /* WOLFSSL_ASYNC_CRYPT */ |
wolfSSL | 15:117db924cf7c | 8601 | |
wolfSSL | 15:117db924cf7c | 8602 | /* Software AES-GCM */ |
wolfSSL | 15:117db924cf7c | 8603 | |
wolfSSL | 15:117db924cf7c | 8604 | #ifdef WOLFSSL_AESNI |
wolfSSL | 15:117db924cf7c | 8605 | #ifdef HAVE_INTEL_AVX2 |
wolfSSL | 15:117db924cf7c | 8606 | if (IS_INTEL_AVX2(intel_flags)) { |
wolfSSL | 15:117db924cf7c | 8607 | AES_GCM_encrypt_avx2(in, out, authIn, iv, authTag, sz, authInSz, ivSz, |
wolfSSL | 15:117db924cf7c | 8608 | authTagSz, (const byte*)aes->key, aes->rounds); |
wolfSSL | 15:117db924cf7c | 8609 | return 0; |
wolfSSL | 15:117db924cf7c | 8610 | } |
wolfSSL | 15:117db924cf7c | 8611 | else |
wolfSSL | 15:117db924cf7c | 8612 | #endif |
wolfSSL | 15:117db924cf7c | 8613 | #ifdef HAVE_INTEL_AVX1 |
wolfSSL | 15:117db924cf7c | 8614 | if (IS_INTEL_AVX1(intel_flags)) { |
wolfSSL | 15:117db924cf7c | 8615 | AES_GCM_encrypt_avx1(in, out, authIn, iv, authTag, sz, authInSz, ivSz, |
wolfSSL | 15:117db924cf7c | 8616 | authTagSz, (const byte*)aes->key, aes->rounds); |
wolfSSL | 15:117db924cf7c | 8617 | return 0; |
wolfSSL | 15:117db924cf7c | 8618 | } |
wolfSSL | 15:117db924cf7c | 8619 | else |
wolfSSL | 15:117db924cf7c | 8620 | #endif |
wolfSSL | 15:117db924cf7c | 8621 | if (haveAESNI) { |
wolfSSL | 15:117db924cf7c | 8622 | AES_GCM_encrypt(in, out, authIn, iv, authTag, sz, authInSz, ivSz, |
wolfSSL | 15:117db924cf7c | 8623 | authTagSz, (const byte*)aes->key, aes->rounds); |
wolfSSL | 15:117db924cf7c | 8624 | return 0; |
wolfSSL | 15:117db924cf7c | 8625 | } |
wolfSSL | 15:117db924cf7c | 8626 | else |
wolfSSL | 15:117db924cf7c | 8627 | #endif |
wolfSSL | 15:117db924cf7c | 8628 | { |
wolfSSL | 15:117db924cf7c | 8629 | return AES_GCM_encrypt_C(aes, out, in, sz, iv, ivSz, authTag, authTagSz, |
wolfSSL | 15:117db924cf7c | 8630 | authIn, authInSz); |
wolfSSL | 15:117db924cf7c | 8631 | } |
wolfSSL | 15:117db924cf7c | 8632 | } |
wolfSSL | 15:117db924cf7c | 8633 | #endif |
wolfSSL | 15:117db924cf7c | 8634 | |
wolfSSL | 15:117db924cf7c | 8635 | |
wolfSSL | 15:117db924cf7c | 8636 | #if defined(HAVE_AES_DECRYPT) || defined(HAVE_AESGCM_DECRYPT) |
wolfSSL | 15:117db924cf7c | 8637 | #ifdef FREESCALE_LTC_AES_GCM |
wolfSSL | 15:117db924cf7c | 8638 | int wc_AesGcmDecrypt(Aes* aes, byte* out, const byte* in, word32 sz, |
wolfSSL | 15:117db924cf7c | 8639 | const byte* iv, word32 ivSz, |
wolfSSL | 15:117db924cf7c | 8640 | const byte* authTag, word32 authTagSz, |
wolfSSL | 15:117db924cf7c | 8641 | const byte* authIn, word32 authInSz) |
wolfSSL | 15:117db924cf7c | 8642 | { |
wolfSSL | 15:117db924cf7c | 8643 | int ret; |
wolfSSL | 15:117db924cf7c | 8644 | word32 keySize; |
wolfSSL | 15:117db924cf7c | 8645 | status_t status; |
wolfSSL | 15:117db924cf7c | 8646 | |
wolfSSL | 15:117db924cf7c | 8647 | /* argument checks */ |
wolfSSL | 15:117db924cf7c | 8648 | if (aes == NULL || out == NULL || in == NULL || iv == NULL || |
wolfSSL | 15:117db924cf7c | 8649 | authTag == NULL || authTagSz > AES_BLOCK_SIZE) { |
wolfSSL | 15:117db924cf7c | 8650 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 8651 | } |
wolfSSL | 15:117db924cf7c | 8652 | |
wolfSSL | 15:117db924cf7c | 8653 | ret = wc_AesGetKeySize(aes, &keySize); |
wolfSSL | 15:117db924cf7c | 8654 | if (ret != 0) { |
wolfSSL | 15:117db924cf7c | 8655 | return ret; |
wolfSSL | 15:117db924cf7c | 8656 | } |
wolfSSL | 15:117db924cf7c | 8657 | |
wolfSSL | 15:117db924cf7c | 8658 | status = LTC_AES_DecryptTagGcm(LTC_BASE, in, out, sz, iv, ivSz, |
wolfSSL | 15:117db924cf7c | 8659 | authIn, authInSz, (byte*)aes->key, keySize, authTag, authTagSz); |
wolfSSL | 15:117db924cf7c | 8660 | |
wolfSSL | 15:117db924cf7c | 8661 | return (status == kStatus_Success) ? 0 : AES_GCM_AUTH_E; |
wolfSSL | 15:117db924cf7c | 8662 | } |
wolfSSL | 15:117db924cf7c | 8663 | #elif defined(STM32_CRYPTO) && (defined(WOLFSSL_STM32F4) || \ |
wolfSSL | 15:117db924cf7c | 8664 | defined(WOLFSSL_STM32F7) || \ |
wolfSSL | 15:117db924cf7c | 8665 | defined(WOLFSSL_STM32L4)) |
wolfSSL | 15:117db924cf7c | 8666 | int wc_AesGcmDecrypt(Aes* aes, byte* out, const byte* in, word32 sz, |
wolfSSL | 15:117db924cf7c | 8667 | const byte* iv, word32 ivSz, |
wolfSSL | 15:117db924cf7c | 8668 | const byte* authTag, word32 authTagSz, |
wolfSSL | 15:117db924cf7c | 8669 | const byte* authIn, word32 authInSz) |
wolfSSL | 15:117db924cf7c | 8670 | { |
wolfSSL | 15:117db924cf7c | 8671 | int ret; |
wolfSSL | 15:117db924cf7c | 8672 | word32 keySize; |
wolfSSL | 15:117db924cf7c | 8673 | #ifdef WOLFSSL_STM32_CUBEMX |
wolfSSL | 15:117db924cf7c | 8674 | CRYP_HandleTypeDef hcryp; |
wolfSSL | 15:117db924cf7c | 8675 | #else |
wolfSSL | 15:117db924cf7c | 8676 | byte keyCopy[AES_BLOCK_SIZE * 2]; |
wolfSSL | 15:117db924cf7c | 8677 | #endif /* WOLFSSL_STM32_CUBEMX */ |
wolfSSL | 15:117db924cf7c | 8678 | int status; |
wolfSSL | 15:117db924cf7c | 8679 | int inPadSz, authPadSz; |
wolfSSL | 15:117db924cf7c | 8680 | byte tag[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 8681 | byte *inPadded = NULL; |
wolfSSL | 15:117db924cf7c | 8682 | byte *authInPadded = NULL; |
wolfSSL | 15:117db924cf7c | 8683 | byte initialCounter[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 8684 | |
wolfSSL | 15:117db924cf7c | 8685 | /* argument checks */ |
wolfSSL | 15:117db924cf7c | 8686 | if (aes == NULL || out == NULL || in == NULL || iv == NULL || |
wolfSSL | 15:117db924cf7c | 8687 | authTag == NULL || authTagSz > AES_BLOCK_SIZE) { |
wolfSSL | 15:117db924cf7c | 8688 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 8689 | } |
wolfSSL | 15:117db924cf7c | 8690 | |
wolfSSL | 15:117db924cf7c | 8691 | ret = wc_AesGetKeySize(aes, &keySize); |
wolfSSL | 15:117db924cf7c | 8692 | if (ret != 0) { |
wolfSSL | 15:117db924cf7c | 8693 | return ret; |
wolfSSL | 15:117db924cf7c | 8694 | } |
wolfSSL | 15:117db924cf7c | 8695 | |
wolfSSL | 15:117db924cf7c | 8696 | /* additional argument checks - STM32 HW only supports 12 byte IV */ |
wolfSSL | 15:117db924cf7c | 8697 | if (ivSz != GCM_NONCE_MID_SZ) { |
wolfSSL | 15:117db924cf7c | 8698 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 8699 | } |
wolfSSL | 15:117db924cf7c | 8700 | |
wolfSSL | 15:117db924cf7c | 8701 | XMEMSET(initialCounter, 0, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8702 | XMEMCPY(initialCounter, iv, ivSz); |
wolfSSL | 15:117db924cf7c | 8703 | initialCounter[AES_BLOCK_SIZE - 1] = STM32_GCM_IV_START; |
wolfSSL | 15:117db924cf7c | 8704 | |
wolfSSL | 15:117db924cf7c | 8705 | /* Need to pad the AAD and input cipher text to a full block size since |
wolfSSL | 15:117db924cf7c | 8706 | * CRYP_AES_GCM will assume these are a multiple of AES_BLOCK_SIZE. |
wolfSSL | 15:117db924cf7c | 8707 | * It is okay to pad with zeros because GCM does this before GHASH already. |
wolfSSL | 15:117db924cf7c | 8708 | * See NIST SP 800-38D */ |
wolfSSL | 15:117db924cf7c | 8709 | |
wolfSSL | 15:117db924cf7c | 8710 | if ((sz % AES_BLOCK_SIZE) > 0) { |
wolfSSL | 15:117db924cf7c | 8711 | inPadSz = ((sz / AES_BLOCK_SIZE) + 1) * AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8712 | inPadded = XMALLOC(inPadSz, aes->heap, DYNAMIC_TYPE_TMP_BUFFER); |
wolfSSL | 15:117db924cf7c | 8713 | if (inPadded == NULL) { |
wolfSSL | 15:117db924cf7c | 8714 | return MEMORY_E; |
wolfSSL | 15:117db924cf7c | 8715 | } |
wolfSSL | 15:117db924cf7c | 8716 | XMEMSET(inPadded, 0, inPadSz); |
wolfSSL | 15:117db924cf7c | 8717 | XMEMCPY(inPadded, in, sz); |
wolfSSL | 15:117db924cf7c | 8718 | } else { |
wolfSSL | 15:117db924cf7c | 8719 | inPadSz = sz; |
wolfSSL | 15:117db924cf7c | 8720 | inPadded = (byte*)in; |
wolfSSL | 15:117db924cf7c | 8721 | } |
wolfSSL | 15:117db924cf7c | 8722 | |
wolfSSL | 15:117db924cf7c | 8723 | if ((authInSz % AES_BLOCK_SIZE) > 0) { |
wolfSSL | 15:117db924cf7c | 8724 | authPadSz = ((authInSz / AES_BLOCK_SIZE) + 1) * AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8725 | authInPadded = XMALLOC(authPadSz, aes->heap, DYNAMIC_TYPE_TMP_BUFFER); |
wolfSSL | 15:117db924cf7c | 8726 | if (authInPadded == NULL) { |
wolfSSL | 15:117db924cf7c | 8727 | if (inPadded != NULL && inPadSz != sz) |
wolfSSL | 15:117db924cf7c | 8728 | XFREE(inPadded , aes->heap, DYNAMIC_TYPE_TMP_BUFFER); |
wolfSSL | 15:117db924cf7c | 8729 | return MEMORY_E; |
wolfSSL | 15:117db924cf7c | 8730 | } |
wolfSSL | 15:117db924cf7c | 8731 | XMEMSET(authInPadded, 0, authPadSz); |
wolfSSL | 15:117db924cf7c | 8732 | XMEMCPY(authInPadded, authIn, authInSz); |
wolfSSL | 15:117db924cf7c | 8733 | } else { |
wolfSSL | 15:117db924cf7c | 8734 | authPadSz = authInSz; |
wolfSSL | 15:117db924cf7c | 8735 | authInPadded = (byte*)authIn; |
wolfSSL | 15:117db924cf7c | 8736 | } |
wolfSSL | 15:117db924cf7c | 8737 | |
wolfSSL | 15:117db924cf7c | 8738 | #ifdef WOLFSSL_STM32_CUBEMX |
wolfSSL | 15:117db924cf7c | 8739 | XMEMSET(&hcryp, 0, sizeof(CRYP_HandleTypeDef)); |
wolfSSL | 15:117db924cf7c | 8740 | switch(keySize) { |
wolfSSL | 15:117db924cf7c | 8741 | case 16: /* 128-bit key */ |
wolfSSL | 15:117db924cf7c | 8742 | hcryp.Init.KeySize = CRYP_KEYSIZE_128B; |
wolfSSL | 15:117db924cf7c | 8743 | break; |
wolfSSL | 15:117db924cf7c | 8744 | #ifdef CRYP_KEYSIZE_192B |
wolfSSL | 15:117db924cf7c | 8745 | case 24: /* 192-bit key */ |
wolfSSL | 15:117db924cf7c | 8746 | hcryp.Init.KeySize = CRYP_KEYSIZE_192B; |
wolfSSL | 15:117db924cf7c | 8747 | break; |
wolfSSL | 15:117db924cf7c | 8748 | #endif |
wolfSSL | 15:117db924cf7c | 8749 | case 32: /* 256-bit key */ |
wolfSSL | 15:117db924cf7c | 8750 | hcryp.Init.KeySize = CRYP_KEYSIZE_256B; |
wolfSSL | 15:117db924cf7c | 8751 | break; |
wolfSSL | 15:117db924cf7c | 8752 | default: |
wolfSSL | 15:117db924cf7c | 8753 | break; |
wolfSSL | 15:117db924cf7c | 8754 | } |
wolfSSL | 15:117db924cf7c | 8755 | hcryp.Instance = CRYP; |
wolfSSL | 15:117db924cf7c | 8756 | hcryp.Init.DataType = CRYP_DATATYPE_8B; |
wolfSSL | 15:117db924cf7c | 8757 | hcryp.Init.pKey = (byte*)aes->key; |
wolfSSL | 15:117db924cf7c | 8758 | hcryp.Init.pInitVect = initialCounter; |
wolfSSL | 15:117db924cf7c | 8759 | hcryp.Init.Header = authInPadded; |
wolfSSL | 15:117db924cf7c | 8760 | hcryp.Init.HeaderSize = authInSz; |
wolfSSL | 15:117db924cf7c | 8761 | |
wolfSSL | 15:117db924cf7c | 8762 | #ifdef WOLFSSL_STM32L4 |
wolfSSL | 15:117db924cf7c | 8763 | /* Set the CRYP parameters */ |
wolfSSL | 15:117db924cf7c | 8764 | hcryp.Init.ChainingMode = CRYP_CHAINMODE_AES_GCM_GMAC; |
wolfSSL | 15:117db924cf7c | 8765 | hcryp.Init.OperatingMode = CRYP_ALGOMODE_DECRYPT; |
wolfSSL | 15:117db924cf7c | 8766 | hcryp.Init.GCMCMACPhase = CRYP_INIT_PHASE; |
wolfSSL | 15:117db924cf7c | 8767 | HAL_CRYP_Init(&hcryp); |
wolfSSL | 15:117db924cf7c | 8768 | |
wolfSSL | 15:117db924cf7c | 8769 | /* GCM init phase */ |
wolfSSL | 15:117db924cf7c | 8770 | status = HAL_CRYPEx_AES_Auth(&hcryp, NULL, 0, NULL, STM32_HAL_TIMEOUT); |
wolfSSL | 15:117db924cf7c | 8771 | if (status == HAL_OK) { |
wolfSSL | 15:117db924cf7c | 8772 | /* GCM header phase */ |
wolfSSL | 15:117db924cf7c | 8773 | hcryp.Init.GCMCMACPhase = CRYP_HEADER_PHASE; |
wolfSSL | 15:117db924cf7c | 8774 | status = HAL_CRYPEx_AES_Auth(&hcryp, NULL, 0, NULL, STM32_HAL_TIMEOUT); |
wolfSSL | 15:117db924cf7c | 8775 | if (status == HAL_OK) { |
wolfSSL | 15:117db924cf7c | 8776 | /* GCM payload phase */ |
wolfSSL | 15:117db924cf7c | 8777 | hcryp.Init.GCMCMACPhase = CRYP_PAYLOAD_PHASE; |
wolfSSL | 15:117db924cf7c | 8778 | status = HAL_CRYPEx_AES_Auth(&hcryp, (byte*)inPadded, sz, inPadded, |
wolfSSL | 15:117db924cf7c | 8779 | STM32_HAL_TIMEOUT); |
wolfSSL | 15:117db924cf7c | 8780 | if (status == HAL_OK) { |
wolfSSL | 15:117db924cf7c | 8781 | /* GCM final phase */ |
wolfSSL | 15:117db924cf7c | 8782 | hcryp.Init.GCMCMACPhase = CRYP_FINAL_PHASE; |
wolfSSL | 15:117db924cf7c | 8783 | status = HAL_CRYPEx_AES_Auth(&hcryp, NULL, sz, tag, |
wolfSSL | 15:117db924cf7c | 8784 | STM32_HAL_TIMEOUT); |
wolfSSL | 15:117db924cf7c | 8785 | } |
wolfSSL | 15:117db924cf7c | 8786 | } |
wolfSSL | 15:117db924cf7c | 8787 | } |
wolfSSL | 15:117db924cf7c | 8788 | #else |
wolfSSL | 15:117db924cf7c | 8789 | HAL_CRYP_Init(&hcryp); |
wolfSSL | 15:117db924cf7c | 8790 | /* Use inPadded for output buffer instead of |
wolfSSL | 15:117db924cf7c | 8791 | * out so that we don't overflow our size. */ |
wolfSSL | 15:117db924cf7c | 8792 | status = HAL_CRYPEx_AESGCM_Decrypt(&hcryp, (byte*)inPadded, |
wolfSSL | 15:117db924cf7c | 8793 | sz, inPadded, STM32_HAL_TIMEOUT); |
wolfSSL | 15:117db924cf7c | 8794 | /* Compute the authTag */ |
wolfSSL | 15:117db924cf7c | 8795 | if (status == HAL_OK) { |
wolfSSL | 15:117db924cf7c | 8796 | status = HAL_CRYPEx_AESGCM_Finish(&hcryp, sz, tag, STM32_HAL_TIMEOUT); |
wolfSSL | 15:117db924cf7c | 8797 | } |
wolfSSL | 15:117db924cf7c | 8798 | #endif |
wolfSSL | 15:117db924cf7c | 8799 | |
wolfSSL | 15:117db924cf7c | 8800 | if (status != HAL_OK) |
wolfSSL | 15:117db924cf7c | 8801 | ret = AES_GCM_AUTH_E; |
wolfSSL | 15:117db924cf7c | 8802 | |
wolfSSL | 15:117db924cf7c | 8803 | HAL_CRYP_DeInit(&hcryp); |
wolfSSL | 15:117db924cf7c | 8804 | #else |
wolfSSL | 15:117db924cf7c | 8805 | ByteReverseWords((word32*)keyCopy, (word32*)aes->key, keySize); |
wolfSSL | 15:117db924cf7c | 8806 | |
wolfSSL | 15:117db924cf7c | 8807 | /* Input size and auth size need to be the actual sizes, even though |
wolfSSL | 15:117db924cf7c | 8808 | * they are not block aligned, because this length (in bits) is used |
wolfSSL | 15:117db924cf7c | 8809 | * in the final GHASH. Use inPadded for output buffer instead of |
wolfSSL | 15:117db924cf7c | 8810 | * out so that we don't overflow our size. */ |
wolfSSL | 15:117db924cf7c | 8811 | status = CRYP_AES_GCM(MODE_DECRYPT, (uint8_t*)initialCounter, |
wolfSSL | 15:117db924cf7c | 8812 | (uint8_t*)keyCopy, keySize * 8, |
wolfSSL | 15:117db924cf7c | 8813 | (uint8_t*)inPadded, sz, |
wolfSSL | 15:117db924cf7c | 8814 | (uint8_t*)authInPadded,authInSz, |
wolfSSL | 15:117db924cf7c | 8815 | (uint8_t*)inPadded, tag); |
wolfSSL | 15:117db924cf7c | 8816 | if (status != SUCCESS) |
wolfSSL | 15:117db924cf7c | 8817 | ret = AES_GCM_AUTH_E; |
wolfSSL | 15:117db924cf7c | 8818 | #endif /* WOLFSSL_STM32_CUBEMX */ |
wolfSSL | 15:117db924cf7c | 8819 | |
wolfSSL | 15:117db924cf7c | 8820 | if (ret == 0 && ConstantCompare(authTag, tag, authTagSz) == 0) { |
wolfSSL | 15:117db924cf7c | 8821 | /* Only keep the decrypted data if authTag success. */ |
wolfSSL | 15:117db924cf7c | 8822 | XMEMCPY(out, inPadded, sz); |
wolfSSL | 15:117db924cf7c | 8823 | ret = 0; /* success */ |
wolfSSL | 15:117db924cf7c | 8824 | } |
wolfSSL | 15:117db924cf7c | 8825 | |
wolfSSL | 15:117db924cf7c | 8826 | /* only allocate padding buffers if the inputs are not a multiple of block sz */ |
wolfSSL | 15:117db924cf7c | 8827 | if (inPadded != NULL && inPadSz != sz) |
wolfSSL | 15:117db924cf7c | 8828 | XFREE(inPadded , aes->heap, DYNAMIC_TYPE_TMP_BUFFER); |
wolfSSL | 15:117db924cf7c | 8829 | if (authInPadded != NULL && authPadSz != authInSz) |
wolfSSL | 15:117db924cf7c | 8830 | XFREE(authInPadded, aes->heap, DYNAMIC_TYPE_TMP_BUFFER); |
wolfSSL | 15:117db924cf7c | 8831 | |
wolfSSL | 15:117db924cf7c | 8832 | return ret; |
wolfSSL | 15:117db924cf7c | 8833 | } |
wolfSSL | 15:117db924cf7c | 8834 | #else |
wolfSSL | 15:117db924cf7c | 8835 | #ifdef WOLFSSL_AESNI |
wolfSSL | 15:117db924cf7c | 8836 | int AES_GCM_decrypt_C(Aes* aes, byte* out, const byte* in, word32 sz, |
wolfSSL | 15:117db924cf7c | 8837 | const byte* iv, word32 ivSz, |
wolfSSL | 15:117db924cf7c | 8838 | const byte* authTag, word32 authTagSz, |
wolfSSL | 15:117db924cf7c | 8839 | const byte* authIn, word32 authInSz); |
wolfSSL | 15:117db924cf7c | 8840 | #else |
wolfSSL | 15:117db924cf7c | 8841 | static |
wolfSSL | 15:117db924cf7c | 8842 | #endif |
wolfSSL | 15:117db924cf7c | 8843 | int AES_GCM_decrypt_C(Aes* aes, byte* out, const byte* in, word32 sz, |
wolfSSL | 15:117db924cf7c | 8844 | const byte* iv, word32 ivSz, |
wolfSSL | 15:117db924cf7c | 8845 | const byte* authTag, word32 authTagSz, |
wolfSSL | 15:117db924cf7c | 8846 | const byte* authIn, word32 authInSz) |
wolfSSL | 15:117db924cf7c | 8847 | { |
wolfSSL | 15:117db924cf7c | 8848 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 8849 | word32 blocks = sz / AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8850 | word32 partial = sz % AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8851 | const byte* c = in; |
wolfSSL | 15:117db924cf7c | 8852 | byte* p = out; |
wolfSSL | 15:117db924cf7c | 8853 | byte counter[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 8854 | byte initialCounter[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 8855 | byte *ctr; |
wolfSSL | 15:117db924cf7c | 8856 | byte scratch[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 8857 | byte Tprime[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 8858 | byte EKY0[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 8859 | ctr = counter; |
wolfSSL | 15:117db924cf7c | 8860 | |
wolfSSL | 15:117db924cf7c | 8861 | XMEMSET(initialCounter, 0, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8862 | if (ivSz == GCM_NONCE_MID_SZ) { |
wolfSSL | 15:117db924cf7c | 8863 | XMEMCPY(initialCounter, iv, ivSz); |
wolfSSL | 15:117db924cf7c | 8864 | initialCounter[AES_BLOCK_SIZE - 1] = 1; |
wolfSSL | 15:117db924cf7c | 8865 | } |
wolfSSL | 15:117db924cf7c | 8866 | else { |
wolfSSL | 15:117db924cf7c | 8867 | GHASH(aes, NULL, 0, iv, ivSz, initialCounter, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8868 | } |
wolfSSL | 15:117db924cf7c | 8869 | XMEMCPY(ctr, initialCounter, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8870 | |
wolfSSL | 15:117db924cf7c | 8871 | /* Calc the authTag again using the received auth data and the cipher text */ |
wolfSSL | 15:117db924cf7c | 8872 | GHASH(aes, authIn, authInSz, in, sz, Tprime, sizeof(Tprime)); |
wolfSSL | 15:117db924cf7c | 8873 | wc_AesEncrypt(aes, ctr, EKY0); |
wolfSSL | 15:117db924cf7c | 8874 | xorbuf(Tprime, EKY0, sizeof(Tprime)); |
wolfSSL | 15:117db924cf7c | 8875 | |
wolfSSL | 15:117db924cf7c | 8876 | if (ConstantCompare(authTag, Tprime, authTagSz) != 0) { |
wolfSSL | 15:117db924cf7c | 8877 | return AES_GCM_AUTH_E; |
wolfSSL | 15:117db924cf7c | 8878 | } |
wolfSSL | 15:117db924cf7c | 8879 | |
wolfSSL | 15:117db924cf7c | 8880 | #ifdef WOLFSSL_PIC32MZ_CRYPT |
wolfSSL | 15:117db924cf7c | 8881 | if (blocks) { |
wolfSSL | 15:117db924cf7c | 8882 | /* use intitial IV for PIC32 HW, but don't use it below */ |
wolfSSL | 15:117db924cf7c | 8883 | XMEMCPY(aes->reg, ctr, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8884 | |
wolfSSL | 15:117db924cf7c | 8885 | ret = wc_Pic32AesCrypt( |
wolfSSL | 15:117db924cf7c | 8886 | aes->key, aes->keylen, aes->reg, AES_BLOCK_SIZE, |
wolfSSL | 15:117db924cf7c | 8887 | out, in, (blocks * AES_BLOCK_SIZE), |
wolfSSL | 15:117db924cf7c | 8888 | PIC32_DECRYPTION, PIC32_ALGO_AES, PIC32_CRYPTOALGO_AES_GCM); |
wolfSSL | 15:117db924cf7c | 8889 | if (ret != 0) |
wolfSSL | 15:117db924cf7c | 8890 | return ret; |
wolfSSL | 15:117db924cf7c | 8891 | } |
wolfSSL | 15:117db924cf7c | 8892 | /* process remainder using partial handling */ |
wolfSSL | 15:117db924cf7c | 8893 | #endif |
wolfSSL | 15:117db924cf7c | 8894 | |
wolfSSL | 15:117db924cf7c | 8895 | #if defined(HAVE_AES_ECB) && !defined(WOLFSSL_PIC32MZ_CRYPT) |
wolfSSL | 15:117db924cf7c | 8896 | /* some hardware acceleration can gain performance from doing AES encryption |
wolfSSL | 15:117db924cf7c | 8897 | * of the whole buffer at once */ |
wolfSSL | 15:117db924cf7c | 8898 | if (c != p) { /* can not handle inline decryption */ |
wolfSSL | 15:117db924cf7c | 8899 | while (blocks--) { |
wolfSSL | 15:117db924cf7c | 8900 | IncrementGcmCounter(ctr); |
wolfSSL | 15:117db924cf7c | 8901 | XMEMCPY(p, ctr, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8902 | p += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8903 | } |
wolfSSL | 15:117db924cf7c | 8904 | |
wolfSSL | 15:117db924cf7c | 8905 | /* reset number of blocks and then do encryption */ |
wolfSSL | 15:117db924cf7c | 8906 | blocks = sz / AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8907 | wc_AesEcbEncrypt(aes, out, out, AES_BLOCK_SIZE * blocks); |
wolfSSL | 15:117db924cf7c | 8908 | xorbuf(out, c, AES_BLOCK_SIZE * blocks); |
wolfSSL | 15:117db924cf7c | 8909 | c += AES_BLOCK_SIZE * blocks; |
wolfSSL | 15:117db924cf7c | 8910 | } |
wolfSSL | 15:117db924cf7c | 8911 | else |
wolfSSL | 15:117db924cf7c | 8912 | #endif /* HAVE_AES_ECB */ |
wolfSSL | 15:117db924cf7c | 8913 | while (blocks--) { |
wolfSSL | 15:117db924cf7c | 8914 | IncrementGcmCounter(ctr); |
wolfSSL | 15:117db924cf7c | 8915 | #ifndef WOLFSSL_PIC32MZ_CRYPT |
wolfSSL | 15:117db924cf7c | 8916 | wc_AesEncrypt(aes, ctr, scratch); |
wolfSSL | 15:117db924cf7c | 8917 | xorbuf(scratch, c, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8918 | XMEMCPY(p, scratch, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 8919 | #endif |
wolfSSL | 15:117db924cf7c | 8920 | p += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8921 | c += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 8922 | } |
wolfSSL | 15:117db924cf7c | 8923 | |
wolfSSL | 15:117db924cf7c | 8924 | if (partial != 0) { |
wolfSSL | 15:117db924cf7c | 8925 | IncrementGcmCounter(ctr); |
wolfSSL | 15:117db924cf7c | 8926 | wc_AesEncrypt(aes, ctr, scratch); |
wolfSSL | 15:117db924cf7c | 8927 | xorbuf(scratch, c, partial); |
wolfSSL | 15:117db924cf7c | 8928 | XMEMCPY(p, scratch, partial); |
wolfSSL | 15:117db924cf7c | 8929 | } |
wolfSSL | 15:117db924cf7c | 8930 | |
wolfSSL | 15:117db924cf7c | 8931 | return ret; |
wolfSSL | 15:117db924cf7c | 8932 | } |
wolfSSL | 15:117db924cf7c | 8933 | |
wolfSSL | 15:117db924cf7c | 8934 | int wc_AesGcmDecrypt(Aes* aes, byte* out, const byte* in, word32 sz, |
wolfSSL | 15:117db924cf7c | 8935 | const byte* iv, word32 ivSz, |
wolfSSL | 15:117db924cf7c | 8936 | const byte* authTag, word32 authTagSz, |
wolfSSL | 15:117db924cf7c | 8937 | const byte* authIn, word32 authInSz) |
wolfSSL | 15:117db924cf7c | 8938 | { |
wolfSSL | 15:117db924cf7c | 8939 | #ifdef WOLFSSL_AESNI |
wolfSSL | 15:117db924cf7c | 8940 | int res; |
wolfSSL | 15:117db924cf7c | 8941 | #endif |
wolfSSL | 15:117db924cf7c | 8942 | |
wolfSSL | 15:117db924cf7c | 8943 | /* argument checks */ |
wolfSSL | 15:117db924cf7c | 8944 | /* If the sz is non-zero, both in and out must be set. If sz is 0, |
wolfSSL | 15:117db924cf7c | 8945 | * in and out are don't cares, as this is is the GMAC case. */ |
wolfSSL | 15:117db924cf7c | 8946 | if (aes == NULL || iv == NULL || (sz != 0 && (in == NULL || out == NULL)) || |
wolfSSL | 15:117db924cf7c | 8947 | authTag == NULL || authTagSz > AES_BLOCK_SIZE || authTagSz == 0) { |
wolfSSL | 15:117db924cf7c | 8948 | |
wolfSSL | 15:117db924cf7c | 8949 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 8950 | } |
wolfSSL | 15:117db924cf7c | 8951 | |
wolfSSL | 15:117db924cf7c | 8952 | #if defined(WOLFSSL_ASYNC_CRYPT) && defined(WC_ASYNC_ENABLE_AES) |
wolfSSL | 15:117db924cf7c | 8953 | /* if async and byte count above threshold */ |
wolfSSL | 15:117db924cf7c | 8954 | /* only 12-byte IV is supported in HW */ |
wolfSSL | 15:117db924cf7c | 8955 | if (aes->asyncDev.marker == WOLFSSL_ASYNC_MARKER_AES && |
wolfSSL | 15:117db924cf7c | 8956 | sz >= WC_ASYNC_THRESH_AES_GCM && ivSz == GCM_NONCE_MID_SZ) { |
wolfSSL | 15:117db924cf7c | 8957 | #if defined(HAVE_CAVIUM) |
wolfSSL | 15:117db924cf7c | 8958 | #ifdef HAVE_CAVIUM_V |
wolfSSL | 15:117db924cf7c | 8959 | if (authInSz == 20) { /* Nitrox V GCM is only working with 20 byte AAD */ |
wolfSSL | 15:117db924cf7c | 8960 | return NitroxAesGcmDecrypt(aes, out, in, sz, |
wolfSSL | 15:117db924cf7c | 8961 | (const byte*)aes->asyncKey, aes->keylen, iv, ivSz, |
wolfSSL | 15:117db924cf7c | 8962 | authTag, authTagSz, authIn, authInSz); |
wolfSSL | 15:117db924cf7c | 8963 | } |
wolfSSL | 15:117db924cf7c | 8964 | #endif |
wolfSSL | 15:117db924cf7c | 8965 | #elif defined(HAVE_INTEL_QA) |
wolfSSL | 15:117db924cf7c | 8966 | return IntelQaSymAesGcmDecrypt(&aes->asyncDev, out, in, sz, |
wolfSSL | 15:117db924cf7c | 8967 | (const byte*)aes->asyncKey, aes->keylen, iv, ivSz, |
wolfSSL | 15:117db924cf7c | 8968 | authTag, authTagSz, authIn, authInSz); |
wolfSSL | 15:117db924cf7c | 8969 | #else /* WOLFSSL_ASYNC_CRYPT_TEST */ |
wolfSSL | 15:117db924cf7c | 8970 | if (wc_AsyncTestInit(&aes->asyncDev, ASYNC_TEST_AES_GCM_DECRYPT)) { |
wolfSSL | 15:117db924cf7c | 8971 | WC_ASYNC_TEST* testDev = &aes->asyncDev.test; |
wolfSSL | 15:117db924cf7c | 8972 | testDev->aes.aes = aes; |
wolfSSL | 15:117db924cf7c | 8973 | testDev->aes.out = out; |
wolfSSL | 15:117db924cf7c | 8974 | testDev->aes.in = in; |
wolfSSL | 15:117db924cf7c | 8975 | testDev->aes.sz = sz; |
wolfSSL | 15:117db924cf7c | 8976 | testDev->aes.iv = iv; |
wolfSSL | 15:117db924cf7c | 8977 | testDev->aes.ivSz = ivSz; |
wolfSSL | 15:117db924cf7c | 8978 | testDev->aes.authTag = (byte*)authTag; |
wolfSSL | 15:117db924cf7c | 8979 | testDev->aes.authTagSz = authTagSz; |
wolfSSL | 15:117db924cf7c | 8980 | testDev->aes.authIn = authIn; |
wolfSSL | 15:117db924cf7c | 8981 | testDev->aes.authInSz = authInSz; |
wolfSSL | 15:117db924cf7c | 8982 | return WC_PENDING_E; |
wolfSSL | 15:117db924cf7c | 8983 | } |
wolfSSL | 15:117db924cf7c | 8984 | #endif |
wolfSSL | 15:117db924cf7c | 8985 | } |
wolfSSL | 15:117db924cf7c | 8986 | #endif /* WOLFSSL_ASYNC_CRYPT */ |
wolfSSL | 15:117db924cf7c | 8987 | |
wolfSSL | 15:117db924cf7c | 8988 | /* software AES GCM */ |
wolfSSL | 15:117db924cf7c | 8989 | |
wolfSSL | 15:117db924cf7c | 8990 | #ifdef WOLFSSL_AESNI |
wolfSSL | 15:117db924cf7c | 8991 | #ifdef HAVE_INTEL_AVX2 |
wolfSSL | 15:117db924cf7c | 8992 | if (IS_INTEL_AVX2(intel_flags)) { |
wolfSSL | 15:117db924cf7c | 8993 | AES_GCM_decrypt_avx2(in, out, authIn, iv, authTag, sz, authInSz, ivSz, |
wolfSSL | 15:117db924cf7c | 8994 | authTagSz, (byte*)aes->key, aes->rounds, &res); |
wolfSSL | 15:117db924cf7c | 8995 | if (res == 0) |
wolfSSL | 15:117db924cf7c | 8996 | return AES_GCM_AUTH_E; |
wolfSSL | 15:117db924cf7c | 8997 | return 0; |
wolfSSL | 15:117db924cf7c | 8998 | } |
wolfSSL | 15:117db924cf7c | 8999 | else |
wolfSSL | 15:117db924cf7c | 9000 | #endif |
wolfSSL | 15:117db924cf7c | 9001 | #ifdef HAVE_INTEL_AVX1 |
wolfSSL | 15:117db924cf7c | 9002 | if (IS_INTEL_AVX1(intel_flags)) { |
wolfSSL | 15:117db924cf7c | 9003 | AES_GCM_decrypt_avx1(in, out, authIn, iv, authTag, sz, authInSz, ivSz, |
wolfSSL | 15:117db924cf7c | 9004 | authTagSz, (byte*)aes->key, aes->rounds, &res); |
wolfSSL | 15:117db924cf7c | 9005 | if (res == 0) |
wolfSSL | 15:117db924cf7c | 9006 | return AES_GCM_AUTH_E; |
wolfSSL | 15:117db924cf7c | 9007 | return 0; |
wolfSSL | 15:117db924cf7c | 9008 | } |
wolfSSL | 15:117db924cf7c | 9009 | else |
wolfSSL | 15:117db924cf7c | 9010 | #endif |
wolfSSL | 15:117db924cf7c | 9011 | if (haveAESNI) { |
wolfSSL | 15:117db924cf7c | 9012 | AES_GCM_decrypt(in, out, authIn, iv, authTag, sz, authInSz, ivSz, |
wolfSSL | 15:117db924cf7c | 9013 | authTagSz, (byte*)aes->key, aes->rounds, &res); |
wolfSSL | 15:117db924cf7c | 9014 | if (res == 0) |
wolfSSL | 15:117db924cf7c | 9015 | return AES_GCM_AUTH_E; |
wolfSSL | 15:117db924cf7c | 9016 | return 0; |
wolfSSL | 15:117db924cf7c | 9017 | } |
wolfSSL | 15:117db924cf7c | 9018 | else |
wolfSSL | 15:117db924cf7c | 9019 | #endif |
wolfSSL | 15:117db924cf7c | 9020 | { |
wolfSSL | 15:117db924cf7c | 9021 | return AES_GCM_decrypt_C(aes, out, in, sz, iv, ivSz, authTag, authTagSz, |
wolfSSL | 15:117db924cf7c | 9022 | authIn, authInSz); |
wolfSSL | 15:117db924cf7c | 9023 | } |
wolfSSL | 15:117db924cf7c | 9024 | } |
wolfSSL | 15:117db924cf7c | 9025 | #endif |
wolfSSL | 15:117db924cf7c | 9026 | #endif /* HAVE_AES_DECRYPT || HAVE_AESGCM_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 9027 | #endif /* (WOLFSSL_XILINX_CRYPT) */ |
wolfSSL | 15:117db924cf7c | 9028 | #endif /* end of block for AESGCM implementation selection */ |
wolfSSL | 15:117db924cf7c | 9029 | |
wolfSSL | 15:117db924cf7c | 9030 | |
wolfSSL | 15:117db924cf7c | 9031 | /* Common to all, abstract functions that build off of lower level AESGCM |
wolfSSL | 15:117db924cf7c | 9032 | * functions */ |
wolfSSL | 15:117db924cf7c | 9033 | #ifndef WC_NO_RNG |
wolfSSL | 15:117db924cf7c | 9034 | |
wolfSSL | 15:117db924cf7c | 9035 | int wc_AesGcmSetExtIV(Aes* aes, const byte* iv, word32 ivSz) |
wolfSSL | 15:117db924cf7c | 9036 | { |
wolfSSL | 15:117db924cf7c | 9037 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 9038 | |
wolfSSL | 15:117db924cf7c | 9039 | if (aes == NULL || iv == NULL || |
wolfSSL | 15:117db924cf7c | 9040 | (ivSz != GCM_NONCE_MIN_SZ && ivSz != GCM_NONCE_MID_SZ && |
wolfSSL | 15:117db924cf7c | 9041 | ivSz != GCM_NONCE_MAX_SZ)) { |
wolfSSL | 15:117db924cf7c | 9042 | |
wolfSSL | 15:117db924cf7c | 9043 | ret = BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9044 | } |
wolfSSL | 15:117db924cf7c | 9045 | |
wolfSSL | 15:117db924cf7c | 9046 | if (ret == 0) { |
wolfSSL | 15:117db924cf7c | 9047 | XMEMCPY((byte*)aes->reg, iv, ivSz); |
wolfSSL | 15:117db924cf7c | 9048 | |
wolfSSL | 15:117db924cf7c | 9049 | /* If the IV is 96, allow for a 2^64 invocation counter. |
wolfSSL | 15:117db924cf7c | 9050 | * For any other size for the nonce, limit the invocation |
wolfSSL | 15:117db924cf7c | 9051 | * counter to 32-bits. (SP 800-38D 8.3) */ |
wolfSSL | 15:117db924cf7c | 9052 | aes->invokeCtr[0] = 0; |
wolfSSL | 15:117db924cf7c | 9053 | aes->invokeCtr[1] = (ivSz == GCM_NONCE_MID_SZ) ? 0 : 0xFFFFFFFF; |
wolfSSL | 15:117db924cf7c | 9054 | aes->nonceSz = ivSz; |
wolfSSL | 15:117db924cf7c | 9055 | } |
wolfSSL | 15:117db924cf7c | 9056 | |
wolfSSL | 15:117db924cf7c | 9057 | return ret; |
wolfSSL | 15:117db924cf7c | 9058 | } |
wolfSSL | 15:117db924cf7c | 9059 | |
wolfSSL | 15:117db924cf7c | 9060 | |
wolfSSL | 15:117db924cf7c | 9061 | int wc_AesGcmSetIV(Aes* aes, word32 ivSz, |
wolfSSL | 15:117db924cf7c | 9062 | const byte* ivFixed, word32 ivFixedSz, |
wolfSSL | 15:117db924cf7c | 9063 | WC_RNG* rng) |
wolfSSL | 15:117db924cf7c | 9064 | { |
wolfSSL | 15:117db924cf7c | 9065 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 9066 | |
wolfSSL | 15:117db924cf7c | 9067 | if (aes == NULL || rng == NULL || |
wolfSSL | 15:117db924cf7c | 9068 | (ivSz != GCM_NONCE_MIN_SZ && ivSz != GCM_NONCE_MID_SZ && |
wolfSSL | 15:117db924cf7c | 9069 | ivSz != GCM_NONCE_MAX_SZ) || |
wolfSSL | 15:117db924cf7c | 9070 | (ivFixed == NULL && ivFixedSz != 0) || |
wolfSSL | 15:117db924cf7c | 9071 | (ivFixed != NULL && ivFixedSz != AES_IV_FIXED_SZ)) { |
wolfSSL | 15:117db924cf7c | 9072 | |
wolfSSL | 15:117db924cf7c | 9073 | ret = BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9074 | } |
wolfSSL | 15:117db924cf7c | 9075 | |
wolfSSL | 15:117db924cf7c | 9076 | if (ret == 0) { |
wolfSSL | 15:117db924cf7c | 9077 | byte* iv = (byte*)aes->reg; |
wolfSSL | 15:117db924cf7c | 9078 | |
wolfSSL | 15:117db924cf7c | 9079 | if (ivFixedSz) |
wolfSSL | 15:117db924cf7c | 9080 | XMEMCPY(iv, ivFixed, ivFixedSz); |
wolfSSL | 15:117db924cf7c | 9081 | |
wolfSSL | 15:117db924cf7c | 9082 | ret = wc_RNG_GenerateBlock(rng, iv + ivFixedSz, ivSz - ivFixedSz); |
wolfSSL | 15:117db924cf7c | 9083 | } |
wolfSSL | 15:117db924cf7c | 9084 | |
wolfSSL | 15:117db924cf7c | 9085 | if (ret == 0) { |
wolfSSL | 15:117db924cf7c | 9086 | /* If the IV is 96, allow for a 2^64 invocation counter. |
wolfSSL | 15:117db924cf7c | 9087 | * For any other size for the nonce, limit the invocation |
wolfSSL | 15:117db924cf7c | 9088 | * counter to 32-bits. (SP 800-38D 8.3) */ |
wolfSSL | 15:117db924cf7c | 9089 | aes->invokeCtr[0] = 0; |
wolfSSL | 15:117db924cf7c | 9090 | aes->invokeCtr[1] = (ivSz == GCM_NONCE_MID_SZ) ? 0 : 0xFFFFFFFF; |
wolfSSL | 15:117db924cf7c | 9091 | aes->nonceSz = ivSz; |
wolfSSL | 15:117db924cf7c | 9092 | } |
wolfSSL | 15:117db924cf7c | 9093 | |
wolfSSL | 15:117db924cf7c | 9094 | return ret; |
wolfSSL | 15:117db924cf7c | 9095 | } |
wolfSSL | 15:117db924cf7c | 9096 | |
wolfSSL | 15:117db924cf7c | 9097 | |
wolfSSL | 15:117db924cf7c | 9098 | int wc_AesGcmEncrypt_ex(Aes* aes, byte* out, const byte* in, word32 sz, |
wolfSSL | 15:117db924cf7c | 9099 | byte* ivOut, word32 ivOutSz, |
wolfSSL | 15:117db924cf7c | 9100 | byte* authTag, word32 authTagSz, |
wolfSSL | 15:117db924cf7c | 9101 | const byte* authIn, word32 authInSz) |
wolfSSL | 15:117db924cf7c | 9102 | { |
wolfSSL | 15:117db924cf7c | 9103 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 9104 | |
wolfSSL | 15:117db924cf7c | 9105 | if (aes == NULL || (sz != 0 && (in == NULL || out == NULL)) || |
wolfSSL | 15:117db924cf7c | 9106 | ivOut == NULL || ivOutSz != aes->nonceSz || |
wolfSSL | 15:117db924cf7c | 9107 | (authIn == NULL && authInSz != 0)) { |
wolfSSL | 15:117db924cf7c | 9108 | |
wolfSSL | 15:117db924cf7c | 9109 | ret = BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9110 | } |
wolfSSL | 15:117db924cf7c | 9111 | |
wolfSSL | 15:117db924cf7c | 9112 | if (ret == 0) { |
wolfSSL | 15:117db924cf7c | 9113 | aes->invokeCtr[0]++; |
wolfSSL | 15:117db924cf7c | 9114 | if (aes->invokeCtr[0] == 0) { |
wolfSSL | 15:117db924cf7c | 9115 | aes->invokeCtr[1]++; |
wolfSSL | 15:117db924cf7c | 9116 | if (aes->invokeCtr[1] == 0) |
wolfSSL | 15:117db924cf7c | 9117 | ret = AES_GCM_OVERFLOW_E; |
wolfSSL | 15:117db924cf7c | 9118 | } |
wolfSSL | 15:117db924cf7c | 9119 | } |
wolfSSL | 15:117db924cf7c | 9120 | |
wolfSSL | 15:117db924cf7c | 9121 | if (ret == 0) { |
wolfSSL | 15:117db924cf7c | 9122 | XMEMCPY(ivOut, aes->reg, ivOutSz); |
wolfSSL | 15:117db924cf7c | 9123 | ret = wc_AesGcmEncrypt(aes, out, in, sz, |
wolfSSL | 15:117db924cf7c | 9124 | (byte*)aes->reg, ivOutSz, |
wolfSSL | 15:117db924cf7c | 9125 | authTag, authTagSz, |
wolfSSL | 15:117db924cf7c | 9126 | authIn, authInSz); |
wolfSSL | 15:117db924cf7c | 9127 | IncCtr((byte*)aes->reg, ivOutSz); |
wolfSSL | 15:117db924cf7c | 9128 | } |
wolfSSL | 15:117db924cf7c | 9129 | |
wolfSSL | 15:117db924cf7c | 9130 | return ret; |
wolfSSL | 15:117db924cf7c | 9131 | } |
wolfSSL | 15:117db924cf7c | 9132 | |
wolfSSL | 15:117db924cf7c | 9133 | int wc_Gmac(const byte* key, word32 keySz, byte* iv, word32 ivSz, |
wolfSSL | 15:117db924cf7c | 9134 | const byte* authIn, word32 authInSz, |
wolfSSL | 15:117db924cf7c | 9135 | byte* authTag, word32 authTagSz, WC_RNG* rng) |
wolfSSL | 15:117db924cf7c | 9136 | { |
wolfSSL | 15:117db924cf7c | 9137 | Aes aes; |
wolfSSL | 15:117db924cf7c | 9138 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 9139 | |
wolfSSL | 15:117db924cf7c | 9140 | if (key == NULL || iv == NULL || (authIn == NULL && authInSz != 0) || |
wolfSSL | 15:117db924cf7c | 9141 | authTag == NULL || authTagSz == 0 || rng == NULL) { |
wolfSSL | 15:117db924cf7c | 9142 | |
wolfSSL | 15:117db924cf7c | 9143 | ret = BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9144 | } |
wolfSSL | 15:117db924cf7c | 9145 | |
wolfSSL | 15:117db924cf7c | 9146 | if (ret == 0) |
wolfSSL | 15:117db924cf7c | 9147 | ret = wc_AesGcmSetKey(&aes, key, keySz); |
wolfSSL | 15:117db924cf7c | 9148 | if (ret == 0) |
wolfSSL | 15:117db924cf7c | 9149 | ret = wc_AesGcmSetIV(&aes, ivSz, NULL, 0, rng); |
wolfSSL | 15:117db924cf7c | 9150 | if (ret == 0) |
wolfSSL | 15:117db924cf7c | 9151 | ret = wc_AesGcmEncrypt_ex(&aes, NULL, NULL, 0, iv, ivSz, |
wolfSSL | 15:117db924cf7c | 9152 | authTag, authTagSz, authIn, authInSz); |
wolfSSL | 15:117db924cf7c | 9153 | ForceZero(&aes, sizeof(aes)); |
wolfSSL | 15:117db924cf7c | 9154 | |
wolfSSL | 15:117db924cf7c | 9155 | return ret; |
wolfSSL | 15:117db924cf7c | 9156 | } |
wolfSSL | 15:117db924cf7c | 9157 | |
wolfSSL | 15:117db924cf7c | 9158 | int wc_GmacVerify(const byte* key, word32 keySz, |
wolfSSL | 15:117db924cf7c | 9159 | const byte* iv, word32 ivSz, |
wolfSSL | 15:117db924cf7c | 9160 | const byte* authIn, word32 authInSz, |
wolfSSL | 15:117db924cf7c | 9161 | const byte* authTag, word32 authTagSz) |
wolfSSL | 15:117db924cf7c | 9162 | { |
wolfSSL | 15:117db924cf7c | 9163 | Aes aes; |
wolfSSL | 15:117db924cf7c | 9164 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 9165 | |
wolfSSL | 15:117db924cf7c | 9166 | if (key == NULL || iv == NULL || (authIn == NULL && authInSz != 0) || |
wolfSSL | 15:117db924cf7c | 9167 | authTag == NULL || authTagSz == 0 || authTagSz > AES_BLOCK_SIZE) { |
wolfSSL | 15:117db924cf7c | 9168 | |
wolfSSL | 15:117db924cf7c | 9169 | ret = BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9170 | } |
wolfSSL | 15:117db924cf7c | 9171 | |
wolfSSL | 15:117db924cf7c | 9172 | if (ret == 0) |
wolfSSL | 15:117db924cf7c | 9173 | ret = wc_AesGcmSetKey(&aes, key, keySz); |
wolfSSL | 15:117db924cf7c | 9174 | if (ret == 0) |
wolfSSL | 15:117db924cf7c | 9175 | ret = wc_AesGcmDecrypt(&aes, NULL, NULL, 0, iv, ivSz, |
wolfSSL | 15:117db924cf7c | 9176 | authTag, authTagSz, authIn, authInSz); |
wolfSSL | 15:117db924cf7c | 9177 | ForceZero(&aes, sizeof(aes)); |
wolfSSL | 15:117db924cf7c | 9178 | |
wolfSSL | 15:117db924cf7c | 9179 | return ret; |
wolfSSL | 15:117db924cf7c | 9180 | } |
wolfSSL | 15:117db924cf7c | 9181 | |
wolfSSL | 15:117db924cf7c | 9182 | #endif /* WC_NO_RNG */ |
wolfSSL | 15:117db924cf7c | 9183 | |
wolfSSL | 15:117db924cf7c | 9184 | |
wolfSSL | 15:117db924cf7c | 9185 | WOLFSSL_API int wc_GmacSetKey(Gmac* gmac, const byte* key, word32 len) |
wolfSSL | 15:117db924cf7c | 9186 | { |
wolfSSL | 15:117db924cf7c | 9187 | if (gmac == NULL || key == NULL) { |
wolfSSL | 15:117db924cf7c | 9188 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9189 | } |
wolfSSL | 15:117db924cf7c | 9190 | return wc_AesGcmSetKey(&gmac->aes, key, len); |
wolfSSL | 15:117db924cf7c | 9191 | } |
wolfSSL | 15:117db924cf7c | 9192 | |
wolfSSL | 15:117db924cf7c | 9193 | |
wolfSSL | 15:117db924cf7c | 9194 | WOLFSSL_API int wc_GmacUpdate(Gmac* gmac, const byte* iv, word32 ivSz, |
wolfSSL | 15:117db924cf7c | 9195 | const byte* authIn, word32 authInSz, |
wolfSSL | 15:117db924cf7c | 9196 | byte* authTag, word32 authTagSz) |
wolfSSL | 15:117db924cf7c | 9197 | { |
wolfSSL | 15:117db924cf7c | 9198 | return wc_AesGcmEncrypt(&gmac->aes, NULL, NULL, 0, iv, ivSz, |
wolfSSL | 15:117db924cf7c | 9199 | authTag, authTagSz, authIn, authInSz); |
wolfSSL | 15:117db924cf7c | 9200 | } |
wolfSSL | 15:117db924cf7c | 9201 | |
wolfSSL | 15:117db924cf7c | 9202 | #endif /* HAVE_AESGCM */ |
wolfSSL | 15:117db924cf7c | 9203 | |
wolfSSL | 15:117db924cf7c | 9204 | |
wolfSSL | 15:117db924cf7c | 9205 | #ifdef HAVE_AESCCM |
wolfSSL | 15:117db924cf7c | 9206 | |
wolfSSL | 15:117db924cf7c | 9207 | int wc_AesCcmSetKey(Aes* aes, const byte* key, word32 keySz) |
wolfSSL | 15:117db924cf7c | 9208 | { |
wolfSSL | 15:117db924cf7c | 9209 | if (!((keySz == 16) || (keySz == 24) || (keySz == 32))) |
wolfSSL | 15:117db924cf7c | 9210 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9211 | |
wolfSSL | 15:117db924cf7c | 9212 | return wc_AesSetKey(aes, key, keySz, NULL, AES_ENCRYPTION); |
wolfSSL | 15:117db924cf7c | 9213 | } |
wolfSSL | 15:117db924cf7c | 9214 | |
wolfSSL | 15:117db924cf7c | 9215 | #ifdef WOLFSSL_ARMASM |
wolfSSL | 15:117db924cf7c | 9216 | /* implementation located in wolfcrypt/src/port/arm/armv8-aes.c */ |
wolfSSL | 15:117db924cf7c | 9217 | |
wolfSSL | 15:117db924cf7c | 9218 | #elif defined(HAVE_COLDFIRE_SEC) |
wolfSSL | 15:117db924cf7c | 9219 | #error "Coldfire SEC doesn't currently support AES-CCM mode" |
wolfSSL | 15:117db924cf7c | 9220 | |
wolfSSL | 15:117db924cf7c | 9221 | #elif defined(WOLFSSL_IMX6_CAAM) && !defined(NO_IMX6_CAAM_AES) |
wolfSSL | 15:117db924cf7c | 9222 | /* implemented in wolfcrypt/src/port/caam_aes.c */ |
wolfSSL | 15:117db924cf7c | 9223 | |
wolfSSL | 15:117db924cf7c | 9224 | #elif defined(FREESCALE_LTC) |
wolfSSL | 15:117db924cf7c | 9225 | |
wolfSSL | 15:117db924cf7c | 9226 | /* return 0 on success */ |
wolfSSL | 15:117db924cf7c | 9227 | int wc_AesCcmEncrypt(Aes* aes, byte* out, const byte* in, word32 inSz, |
wolfSSL | 15:117db924cf7c | 9228 | const byte* nonce, word32 nonceSz, |
wolfSSL | 15:117db924cf7c | 9229 | byte* authTag, word32 authTagSz, |
wolfSSL | 15:117db924cf7c | 9230 | const byte* authIn, word32 authInSz) |
wolfSSL | 15:117db924cf7c | 9231 | { |
wolfSSL | 15:117db924cf7c | 9232 | byte *key; |
wolfSSL | 15:117db924cf7c | 9233 | uint32_t keySize; |
wolfSSL | 15:117db924cf7c | 9234 | status_t status; |
wolfSSL | 15:117db924cf7c | 9235 | |
wolfSSL | 15:117db924cf7c | 9236 | /* sanity check on arguments */ |
wolfSSL | 15:117db924cf7c | 9237 | if (aes == NULL || out == NULL || in == NULL || nonce == NULL |
wolfSSL | 15:117db924cf7c | 9238 | || authTag == NULL || nonceSz < 7 || nonceSz > 13) |
wolfSSL | 15:117db924cf7c | 9239 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9240 | |
wolfSSL | 15:117db924cf7c | 9241 | key = (byte*)aes->key; |
wolfSSL | 15:117db924cf7c | 9242 | |
wolfSSL | 15:117db924cf7c | 9243 | status = wc_AesGetKeySize(aes, &keySize); |
wolfSSL | 15:117db924cf7c | 9244 | if (status != 0) { |
wolfSSL | 15:117db924cf7c | 9245 | return status; |
wolfSSL | 15:117db924cf7c | 9246 | } |
wolfSSL | 15:117db924cf7c | 9247 | |
wolfSSL | 15:117db924cf7c | 9248 | status = LTC_AES_EncryptTagCcm(LTC_BASE, in, out, inSz, |
wolfSSL | 15:117db924cf7c | 9249 | nonce, nonceSz, authIn, authInSz, key, keySize, authTag, authTagSz); |
wolfSSL | 15:117db924cf7c | 9250 | |
wolfSSL | 15:117db924cf7c | 9251 | return (kStatus_Success == status) ? 0 : BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9252 | } |
wolfSSL | 15:117db924cf7c | 9253 | |
wolfSSL | 15:117db924cf7c | 9254 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 9255 | int wc_AesCcmDecrypt(Aes* aes, byte* out, const byte* in, word32 inSz, |
wolfSSL | 15:117db924cf7c | 9256 | const byte* nonce, word32 nonceSz, |
wolfSSL | 15:117db924cf7c | 9257 | const byte* authTag, word32 authTagSz, |
wolfSSL | 15:117db924cf7c | 9258 | const byte* authIn, word32 authInSz) |
wolfSSL | 15:117db924cf7c | 9259 | { |
wolfSSL | 15:117db924cf7c | 9260 | byte *key; |
wolfSSL | 15:117db924cf7c | 9261 | uint32_t keySize; |
wolfSSL | 15:117db924cf7c | 9262 | status_t status; |
wolfSSL | 15:117db924cf7c | 9263 | |
wolfSSL | 15:117db924cf7c | 9264 | /* sanity check on arguments */ |
wolfSSL | 15:117db924cf7c | 9265 | if (aes == NULL || out == NULL || in == NULL || nonce == NULL |
wolfSSL | 15:117db924cf7c | 9266 | || authTag == NULL || nonceSz < 7 || nonceSz > 13) |
wolfSSL | 15:117db924cf7c | 9267 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9268 | |
wolfSSL | 15:117db924cf7c | 9269 | key = (byte*)aes->key; |
wolfSSL | 15:117db924cf7c | 9270 | |
wolfSSL | 15:117db924cf7c | 9271 | status = wc_AesGetKeySize(aes, &keySize); |
wolfSSL | 15:117db924cf7c | 9272 | if (status != 0) { |
wolfSSL | 15:117db924cf7c | 9273 | return status; |
wolfSSL | 15:117db924cf7c | 9274 | } |
wolfSSL | 15:117db924cf7c | 9275 | |
wolfSSL | 15:117db924cf7c | 9276 | status = LTC_AES_DecryptTagCcm(LTC_BASE, in, out, inSz, |
wolfSSL | 15:117db924cf7c | 9277 | nonce, nonceSz, authIn, authInSz, key, keySize, authTag, authTagSz); |
wolfSSL | 15:117db924cf7c | 9278 | |
wolfSSL | 15:117db924cf7c | 9279 | if (status == kStatus_Success) { |
wolfSSL | 15:117db924cf7c | 9280 | return 0; |
wolfSSL | 15:117db924cf7c | 9281 | } |
wolfSSL | 15:117db924cf7c | 9282 | else { |
wolfSSL | 15:117db924cf7c | 9283 | XMEMSET(out, 0, inSz); |
wolfSSL | 15:117db924cf7c | 9284 | return AES_CCM_AUTH_E; |
wolfSSL | 15:117db924cf7c | 9285 | } |
wolfSSL | 15:117db924cf7c | 9286 | } |
wolfSSL | 15:117db924cf7c | 9287 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 9288 | |
wolfSSL | 15:117db924cf7c | 9289 | |
wolfSSL | 15:117db924cf7c | 9290 | /* software AES CCM */ |
wolfSSL | 15:117db924cf7c | 9291 | #else |
wolfSSL | 15:117db924cf7c | 9292 | |
wolfSSL | 15:117db924cf7c | 9293 | static void roll_x(Aes* aes, const byte* in, word32 inSz, byte* out) |
wolfSSL | 15:117db924cf7c | 9294 | { |
wolfSSL | 15:117db924cf7c | 9295 | /* process the bulk of the data */ |
wolfSSL | 15:117db924cf7c | 9296 | while (inSz >= AES_BLOCK_SIZE) { |
wolfSSL | 15:117db924cf7c | 9297 | xorbuf(out, in, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9298 | in += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9299 | inSz -= AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9300 | |
wolfSSL | 15:117db924cf7c | 9301 | wc_AesEncrypt(aes, out, out); |
wolfSSL | 15:117db924cf7c | 9302 | } |
wolfSSL | 15:117db924cf7c | 9303 | |
wolfSSL | 15:117db924cf7c | 9304 | /* process remainder of the data */ |
wolfSSL | 15:117db924cf7c | 9305 | if (inSz > 0) { |
wolfSSL | 15:117db924cf7c | 9306 | xorbuf(out, in, inSz); |
wolfSSL | 15:117db924cf7c | 9307 | wc_AesEncrypt(aes, out, out); |
wolfSSL | 15:117db924cf7c | 9308 | } |
wolfSSL | 15:117db924cf7c | 9309 | } |
wolfSSL | 15:117db924cf7c | 9310 | |
wolfSSL | 15:117db924cf7c | 9311 | static void roll_auth(Aes* aes, const byte* in, word32 inSz, byte* out) |
wolfSSL | 15:117db924cf7c | 9312 | { |
wolfSSL | 15:117db924cf7c | 9313 | word32 authLenSz; |
wolfSSL | 15:117db924cf7c | 9314 | word32 remainder; |
wolfSSL | 15:117db924cf7c | 9315 | |
wolfSSL | 15:117db924cf7c | 9316 | /* encode the length in */ |
wolfSSL | 15:117db924cf7c | 9317 | if (inSz <= 0xFEFF) { |
wolfSSL | 15:117db924cf7c | 9318 | authLenSz = 2; |
wolfSSL | 15:117db924cf7c | 9319 | out[0] ^= ((inSz & 0xFF00) >> 8); |
wolfSSL | 15:117db924cf7c | 9320 | out[1] ^= (inSz & 0x00FF); |
wolfSSL | 15:117db924cf7c | 9321 | } |
wolfSSL | 15:117db924cf7c | 9322 | else if (inSz <= 0xFFFFFFFF) { |
wolfSSL | 15:117db924cf7c | 9323 | authLenSz = 6; |
wolfSSL | 15:117db924cf7c | 9324 | out[0] ^= 0xFF; out[1] ^= 0xFE; |
wolfSSL | 15:117db924cf7c | 9325 | out[2] ^= ((inSz & 0xFF000000) >> 24); |
wolfSSL | 15:117db924cf7c | 9326 | out[3] ^= ((inSz & 0x00FF0000) >> 16); |
wolfSSL | 15:117db924cf7c | 9327 | out[4] ^= ((inSz & 0x0000FF00) >> 8); |
wolfSSL | 15:117db924cf7c | 9328 | out[5] ^= (inSz & 0x000000FF); |
wolfSSL | 15:117db924cf7c | 9329 | } |
wolfSSL | 15:117db924cf7c | 9330 | /* Note, the protocol handles auth data up to 2^64, but we are |
wolfSSL | 15:117db924cf7c | 9331 | * using 32-bit sizes right now, so the bigger data isn't handled |
wolfSSL | 15:117db924cf7c | 9332 | * else if (inSz <= 0xFFFFFFFFFFFFFFFF) {} */ |
wolfSSL | 15:117db924cf7c | 9333 | else |
wolfSSL | 15:117db924cf7c | 9334 | return; |
wolfSSL | 15:117db924cf7c | 9335 | |
wolfSSL | 15:117db924cf7c | 9336 | /* start fill out the rest of the first block */ |
wolfSSL | 15:117db924cf7c | 9337 | remainder = AES_BLOCK_SIZE - authLenSz; |
wolfSSL | 15:117db924cf7c | 9338 | if (inSz >= remainder) { |
wolfSSL | 15:117db924cf7c | 9339 | /* plenty of bulk data to fill the remainder of this block */ |
wolfSSL | 15:117db924cf7c | 9340 | xorbuf(out + authLenSz, in, remainder); |
wolfSSL | 15:117db924cf7c | 9341 | inSz -= remainder; |
wolfSSL | 15:117db924cf7c | 9342 | in += remainder; |
wolfSSL | 15:117db924cf7c | 9343 | } |
wolfSSL | 15:117db924cf7c | 9344 | else { |
wolfSSL | 15:117db924cf7c | 9345 | /* not enough bulk data, copy what is available, and pad zero */ |
wolfSSL | 15:117db924cf7c | 9346 | xorbuf(out + authLenSz, in, inSz); |
wolfSSL | 15:117db924cf7c | 9347 | inSz = 0; |
wolfSSL | 15:117db924cf7c | 9348 | } |
wolfSSL | 15:117db924cf7c | 9349 | wc_AesEncrypt(aes, out, out); |
wolfSSL | 15:117db924cf7c | 9350 | |
wolfSSL | 15:117db924cf7c | 9351 | if (inSz > 0) |
wolfSSL | 15:117db924cf7c | 9352 | roll_x(aes, in, inSz, out); |
wolfSSL | 15:117db924cf7c | 9353 | } |
wolfSSL | 15:117db924cf7c | 9354 | |
wolfSSL | 15:117db924cf7c | 9355 | |
wolfSSL | 15:117db924cf7c | 9356 | static WC_INLINE void AesCcmCtrInc(byte* B, word32 lenSz) |
wolfSSL | 15:117db924cf7c | 9357 | { |
wolfSSL | 15:117db924cf7c | 9358 | word32 i; |
wolfSSL | 15:117db924cf7c | 9359 | |
wolfSSL | 15:117db924cf7c | 9360 | for (i = 0; i < lenSz; i++) { |
wolfSSL | 15:117db924cf7c | 9361 | if (++B[AES_BLOCK_SIZE - 1 - i] != 0) return; |
wolfSSL | 15:117db924cf7c | 9362 | } |
wolfSSL | 15:117db924cf7c | 9363 | } |
wolfSSL | 15:117db924cf7c | 9364 | |
wolfSSL | 15:117db924cf7c | 9365 | /* return 0 on success */ |
wolfSSL | 15:117db924cf7c | 9366 | int wc_AesCcmEncrypt(Aes* aes, byte* out, const byte* in, word32 inSz, |
wolfSSL | 15:117db924cf7c | 9367 | const byte* nonce, word32 nonceSz, |
wolfSSL | 15:117db924cf7c | 9368 | byte* authTag, word32 authTagSz, |
wolfSSL | 15:117db924cf7c | 9369 | const byte* authIn, word32 authInSz) |
wolfSSL | 15:117db924cf7c | 9370 | { |
wolfSSL | 15:117db924cf7c | 9371 | byte A[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 9372 | byte B[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 9373 | byte lenSz; |
wolfSSL | 15:117db924cf7c | 9374 | word32 i; |
wolfSSL | 15:117db924cf7c | 9375 | byte mask = 0xFF; |
wolfSSL | 15:117db924cf7c | 9376 | const word32 wordSz = (word32)sizeof(word32); |
wolfSSL | 15:117db924cf7c | 9377 | |
wolfSSL | 15:117db924cf7c | 9378 | /* sanity check on arguments */ |
wolfSSL | 15:117db924cf7c | 9379 | if (aes == NULL || out == NULL || in == NULL || nonce == NULL |
wolfSSL | 15:117db924cf7c | 9380 | || authTag == NULL || nonceSz < 7 || nonceSz > 13 || |
wolfSSL | 15:117db924cf7c | 9381 | authTagSz > AES_BLOCK_SIZE) |
wolfSSL | 15:117db924cf7c | 9382 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9383 | |
wolfSSL | 15:117db924cf7c | 9384 | XMEMCPY(B+1, nonce, nonceSz); |
wolfSSL | 15:117db924cf7c | 9385 | lenSz = AES_BLOCK_SIZE - 1 - (byte)nonceSz; |
wolfSSL | 15:117db924cf7c | 9386 | B[0] = (authInSz > 0 ? 64 : 0) |
wolfSSL | 15:117db924cf7c | 9387 | + (8 * (((byte)authTagSz - 2) / 2)) |
wolfSSL | 15:117db924cf7c | 9388 | + (lenSz - 1); |
wolfSSL | 15:117db924cf7c | 9389 | for (i = 0; i < lenSz; i++) { |
wolfSSL | 15:117db924cf7c | 9390 | if (mask && i >= wordSz) |
wolfSSL | 15:117db924cf7c | 9391 | mask = 0x00; |
wolfSSL | 15:117db924cf7c | 9392 | B[AES_BLOCK_SIZE - 1 - i] = (inSz >> ((8 * i) & mask)) & mask; |
wolfSSL | 15:117db924cf7c | 9393 | } |
wolfSSL | 15:117db924cf7c | 9394 | |
wolfSSL | 15:117db924cf7c | 9395 | wc_AesEncrypt(aes, B, A); |
wolfSSL | 15:117db924cf7c | 9396 | |
wolfSSL | 15:117db924cf7c | 9397 | if (authInSz > 0) |
wolfSSL | 15:117db924cf7c | 9398 | roll_auth(aes, authIn, authInSz, A); |
wolfSSL | 15:117db924cf7c | 9399 | if (inSz > 0) |
wolfSSL | 15:117db924cf7c | 9400 | roll_x(aes, in, inSz, A); |
wolfSSL | 15:117db924cf7c | 9401 | XMEMCPY(authTag, A, authTagSz); |
wolfSSL | 15:117db924cf7c | 9402 | |
wolfSSL | 15:117db924cf7c | 9403 | B[0] = lenSz - 1; |
wolfSSL | 15:117db924cf7c | 9404 | for (i = 0; i < lenSz; i++) |
wolfSSL | 15:117db924cf7c | 9405 | B[AES_BLOCK_SIZE - 1 - i] = 0; |
wolfSSL | 15:117db924cf7c | 9406 | wc_AesEncrypt(aes, B, A); |
wolfSSL | 15:117db924cf7c | 9407 | xorbuf(authTag, A, authTagSz); |
wolfSSL | 15:117db924cf7c | 9408 | |
wolfSSL | 15:117db924cf7c | 9409 | B[15] = 1; |
wolfSSL | 15:117db924cf7c | 9410 | while (inSz >= AES_BLOCK_SIZE) { |
wolfSSL | 15:117db924cf7c | 9411 | wc_AesEncrypt(aes, B, A); |
wolfSSL | 15:117db924cf7c | 9412 | xorbuf(A, in, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9413 | XMEMCPY(out, A, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9414 | |
wolfSSL | 15:117db924cf7c | 9415 | AesCcmCtrInc(B, lenSz); |
wolfSSL | 15:117db924cf7c | 9416 | inSz -= AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9417 | in += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9418 | out += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9419 | } |
wolfSSL | 15:117db924cf7c | 9420 | if (inSz > 0) { |
wolfSSL | 15:117db924cf7c | 9421 | wc_AesEncrypt(aes, B, A); |
wolfSSL | 15:117db924cf7c | 9422 | xorbuf(A, in, inSz); |
wolfSSL | 15:117db924cf7c | 9423 | XMEMCPY(out, A, inSz); |
wolfSSL | 15:117db924cf7c | 9424 | } |
wolfSSL | 15:117db924cf7c | 9425 | |
wolfSSL | 15:117db924cf7c | 9426 | ForceZero(A, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9427 | ForceZero(B, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9428 | |
wolfSSL | 15:117db924cf7c | 9429 | return 0; |
wolfSSL | 15:117db924cf7c | 9430 | } |
wolfSSL | 15:117db924cf7c | 9431 | |
wolfSSL | 15:117db924cf7c | 9432 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 9433 | int wc_AesCcmDecrypt(Aes* aes, byte* out, const byte* in, word32 inSz, |
wolfSSL | 15:117db924cf7c | 9434 | const byte* nonce, word32 nonceSz, |
wolfSSL | 15:117db924cf7c | 9435 | const byte* authTag, word32 authTagSz, |
wolfSSL | 15:117db924cf7c | 9436 | const byte* authIn, word32 authInSz) |
wolfSSL | 15:117db924cf7c | 9437 | { |
wolfSSL | 15:117db924cf7c | 9438 | byte A[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 9439 | byte B[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 9440 | byte* o; |
wolfSSL | 15:117db924cf7c | 9441 | byte lenSz; |
wolfSSL | 15:117db924cf7c | 9442 | word32 i, oSz; |
wolfSSL | 15:117db924cf7c | 9443 | int result = 0; |
wolfSSL | 15:117db924cf7c | 9444 | byte mask = 0xFF; |
wolfSSL | 15:117db924cf7c | 9445 | const word32 wordSz = (word32)sizeof(word32); |
wolfSSL | 15:117db924cf7c | 9446 | |
wolfSSL | 15:117db924cf7c | 9447 | /* sanity check on arguments */ |
wolfSSL | 15:117db924cf7c | 9448 | if (aes == NULL || out == NULL || in == NULL || nonce == NULL |
wolfSSL | 15:117db924cf7c | 9449 | || authTag == NULL || nonceSz < 7 || nonceSz > 13 || |
wolfSSL | 15:117db924cf7c | 9450 | authTagSz > AES_BLOCK_SIZE) |
wolfSSL | 15:117db924cf7c | 9451 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9452 | |
wolfSSL | 15:117db924cf7c | 9453 | o = out; |
wolfSSL | 15:117db924cf7c | 9454 | oSz = inSz; |
wolfSSL | 15:117db924cf7c | 9455 | XMEMCPY(B+1, nonce, nonceSz); |
wolfSSL | 15:117db924cf7c | 9456 | lenSz = AES_BLOCK_SIZE - 1 - (byte)nonceSz; |
wolfSSL | 15:117db924cf7c | 9457 | |
wolfSSL | 15:117db924cf7c | 9458 | B[0] = lenSz - 1; |
wolfSSL | 15:117db924cf7c | 9459 | for (i = 0; i < lenSz; i++) |
wolfSSL | 15:117db924cf7c | 9460 | B[AES_BLOCK_SIZE - 1 - i] = 0; |
wolfSSL | 15:117db924cf7c | 9461 | B[15] = 1; |
wolfSSL | 15:117db924cf7c | 9462 | |
wolfSSL | 15:117db924cf7c | 9463 | while (oSz >= AES_BLOCK_SIZE) { |
wolfSSL | 15:117db924cf7c | 9464 | wc_AesEncrypt(aes, B, A); |
wolfSSL | 15:117db924cf7c | 9465 | xorbuf(A, in, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9466 | XMEMCPY(o, A, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9467 | |
wolfSSL | 15:117db924cf7c | 9468 | AesCcmCtrInc(B, lenSz); |
wolfSSL | 15:117db924cf7c | 9469 | oSz -= AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9470 | in += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9471 | o += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9472 | } |
wolfSSL | 15:117db924cf7c | 9473 | if (inSz > 0) { |
wolfSSL | 15:117db924cf7c | 9474 | wc_AesEncrypt(aes, B, A); |
wolfSSL | 15:117db924cf7c | 9475 | xorbuf(A, in, oSz); |
wolfSSL | 15:117db924cf7c | 9476 | XMEMCPY(o, A, oSz); |
wolfSSL | 15:117db924cf7c | 9477 | } |
wolfSSL | 15:117db924cf7c | 9478 | |
wolfSSL | 15:117db924cf7c | 9479 | for (i = 0; i < lenSz; i++) |
wolfSSL | 15:117db924cf7c | 9480 | B[AES_BLOCK_SIZE - 1 - i] = 0; |
wolfSSL | 15:117db924cf7c | 9481 | wc_AesEncrypt(aes, B, A); |
wolfSSL | 15:117db924cf7c | 9482 | |
wolfSSL | 15:117db924cf7c | 9483 | o = out; |
wolfSSL | 15:117db924cf7c | 9484 | oSz = inSz; |
wolfSSL | 15:117db924cf7c | 9485 | |
wolfSSL | 15:117db924cf7c | 9486 | B[0] = (authInSz > 0 ? 64 : 0) |
wolfSSL | 15:117db924cf7c | 9487 | + (8 * (((byte)authTagSz - 2) / 2)) |
wolfSSL | 15:117db924cf7c | 9488 | + (lenSz - 1); |
wolfSSL | 15:117db924cf7c | 9489 | for (i = 0; i < lenSz; i++) { |
wolfSSL | 15:117db924cf7c | 9490 | if (mask && i >= wordSz) |
wolfSSL | 15:117db924cf7c | 9491 | mask = 0x00; |
wolfSSL | 15:117db924cf7c | 9492 | B[AES_BLOCK_SIZE - 1 - i] = (inSz >> ((8 * i) & mask)) & mask; |
wolfSSL | 15:117db924cf7c | 9493 | } |
wolfSSL | 15:117db924cf7c | 9494 | |
wolfSSL | 15:117db924cf7c | 9495 | wc_AesEncrypt(aes, B, A); |
wolfSSL | 15:117db924cf7c | 9496 | |
wolfSSL | 15:117db924cf7c | 9497 | if (authInSz > 0) |
wolfSSL | 15:117db924cf7c | 9498 | roll_auth(aes, authIn, authInSz, A); |
wolfSSL | 15:117db924cf7c | 9499 | if (inSz > 0) |
wolfSSL | 15:117db924cf7c | 9500 | roll_x(aes, o, oSz, A); |
wolfSSL | 15:117db924cf7c | 9501 | |
wolfSSL | 15:117db924cf7c | 9502 | B[0] = lenSz - 1; |
wolfSSL | 15:117db924cf7c | 9503 | for (i = 0; i < lenSz; i++) |
wolfSSL | 15:117db924cf7c | 9504 | B[AES_BLOCK_SIZE - 1 - i] = 0; |
wolfSSL | 15:117db924cf7c | 9505 | wc_AesEncrypt(aes, B, B); |
wolfSSL | 15:117db924cf7c | 9506 | xorbuf(A, B, authTagSz); |
wolfSSL | 15:117db924cf7c | 9507 | |
wolfSSL | 15:117db924cf7c | 9508 | if (ConstantCompare(A, authTag, authTagSz) != 0) { |
wolfSSL | 15:117db924cf7c | 9509 | /* If the authTag check fails, don't keep the decrypted data. |
wolfSSL | 15:117db924cf7c | 9510 | * Unfortunately, you need the decrypted data to calculate the |
wolfSSL | 15:117db924cf7c | 9511 | * check value. */ |
wolfSSL | 15:117db924cf7c | 9512 | XMEMSET(out, 0, inSz); |
wolfSSL | 15:117db924cf7c | 9513 | result = AES_CCM_AUTH_E; |
wolfSSL | 15:117db924cf7c | 9514 | } |
wolfSSL | 15:117db924cf7c | 9515 | |
wolfSSL | 15:117db924cf7c | 9516 | ForceZero(A, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9517 | ForceZero(B, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9518 | o = NULL; |
wolfSSL | 15:117db924cf7c | 9519 | |
wolfSSL | 15:117db924cf7c | 9520 | return result; |
wolfSSL | 15:117db924cf7c | 9521 | } |
wolfSSL | 15:117db924cf7c | 9522 | |
wolfSSL | 15:117db924cf7c | 9523 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 9524 | #endif /* software AES CCM */ |
wolfSSL | 15:117db924cf7c | 9525 | |
wolfSSL | 15:117db924cf7c | 9526 | /* abstract functions that call lower level AESCCM functions */ |
wolfSSL | 15:117db924cf7c | 9527 | #ifndef WC_NO_RNG |
wolfSSL | 15:117db924cf7c | 9528 | |
wolfSSL | 15:117db924cf7c | 9529 | int wc_AesCcmSetNonce(Aes* aes, const byte* nonce, word32 nonceSz) |
wolfSSL | 15:117db924cf7c | 9530 | { |
wolfSSL | 15:117db924cf7c | 9531 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 9532 | |
wolfSSL | 15:117db924cf7c | 9533 | if (aes == NULL || nonce == NULL || |
wolfSSL | 15:117db924cf7c | 9534 | nonceSz < CCM_NONCE_MIN_SZ || nonceSz > CCM_NONCE_MAX_SZ) { |
wolfSSL | 15:117db924cf7c | 9535 | |
wolfSSL | 15:117db924cf7c | 9536 | ret = BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9537 | } |
wolfSSL | 15:117db924cf7c | 9538 | |
wolfSSL | 15:117db924cf7c | 9539 | if (ret == 0) { |
wolfSSL | 15:117db924cf7c | 9540 | XMEMCPY(aes->reg, nonce, nonceSz); |
wolfSSL | 15:117db924cf7c | 9541 | aes->nonceSz = nonceSz; |
wolfSSL | 15:117db924cf7c | 9542 | |
wolfSSL | 15:117db924cf7c | 9543 | /* Invocation counter should be 2^61 */ |
wolfSSL | 15:117db924cf7c | 9544 | aes->invokeCtr[0] = 0; |
wolfSSL | 15:117db924cf7c | 9545 | aes->invokeCtr[1] = 0xE0000000; |
wolfSSL | 15:117db924cf7c | 9546 | } |
wolfSSL | 15:117db924cf7c | 9547 | |
wolfSSL | 15:117db924cf7c | 9548 | return ret; |
wolfSSL | 15:117db924cf7c | 9549 | } |
wolfSSL | 15:117db924cf7c | 9550 | |
wolfSSL | 15:117db924cf7c | 9551 | |
wolfSSL | 15:117db924cf7c | 9552 | int wc_AesCcmEncrypt_ex(Aes* aes, byte* out, const byte* in, word32 sz, |
wolfSSL | 15:117db924cf7c | 9553 | byte* ivOut, word32 ivOutSz, |
wolfSSL | 15:117db924cf7c | 9554 | byte* authTag, word32 authTagSz, |
wolfSSL | 15:117db924cf7c | 9555 | const byte* authIn, word32 authInSz) |
wolfSSL | 15:117db924cf7c | 9556 | { |
wolfSSL | 15:117db924cf7c | 9557 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 9558 | |
wolfSSL | 15:117db924cf7c | 9559 | if (aes == NULL || out == NULL || |
wolfSSL | 15:117db924cf7c | 9560 | (in == NULL && sz != 0) || |
wolfSSL | 15:117db924cf7c | 9561 | ivOut == NULL || |
wolfSSL | 15:117db924cf7c | 9562 | (authIn == NULL && authInSz != 0) || |
wolfSSL | 15:117db924cf7c | 9563 | (ivOutSz != aes->nonceSz)) { |
wolfSSL | 15:117db924cf7c | 9564 | |
wolfSSL | 15:117db924cf7c | 9565 | ret = BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9566 | } |
wolfSSL | 15:117db924cf7c | 9567 | |
wolfSSL | 15:117db924cf7c | 9568 | if (ret == 0) { |
wolfSSL | 15:117db924cf7c | 9569 | aes->invokeCtr[0]++; |
wolfSSL | 15:117db924cf7c | 9570 | if (aes->invokeCtr[0] == 0) { |
wolfSSL | 15:117db924cf7c | 9571 | aes->invokeCtr[1]++; |
wolfSSL | 15:117db924cf7c | 9572 | if (aes->invokeCtr[1] == 0) |
wolfSSL | 15:117db924cf7c | 9573 | ret = AES_CCM_OVERFLOW_E; |
wolfSSL | 15:117db924cf7c | 9574 | } |
wolfSSL | 15:117db924cf7c | 9575 | } |
wolfSSL | 15:117db924cf7c | 9576 | |
wolfSSL | 15:117db924cf7c | 9577 | if (ret == 0) { |
wolfSSL | 15:117db924cf7c | 9578 | ret = wc_AesCcmEncrypt(aes, out, in, sz, |
wolfSSL | 15:117db924cf7c | 9579 | (byte*)aes->reg, aes->nonceSz, |
wolfSSL | 15:117db924cf7c | 9580 | authTag, authTagSz, |
wolfSSL | 15:117db924cf7c | 9581 | authIn, authInSz); |
wolfSSL | 15:117db924cf7c | 9582 | XMEMCPY(ivOut, aes->reg, aes->nonceSz); |
wolfSSL | 15:117db924cf7c | 9583 | IncCtr((byte*)aes->reg, aes->nonceSz); |
wolfSSL | 15:117db924cf7c | 9584 | } |
wolfSSL | 15:117db924cf7c | 9585 | |
wolfSSL | 15:117db924cf7c | 9586 | return ret; |
wolfSSL | 15:117db924cf7c | 9587 | } |
wolfSSL | 15:117db924cf7c | 9588 | |
wolfSSL | 15:117db924cf7c | 9589 | #endif /* WC_NO_RNG */ |
wolfSSL | 15:117db924cf7c | 9590 | |
wolfSSL | 15:117db924cf7c | 9591 | #endif /* HAVE_AESCCM */ |
wolfSSL | 15:117db924cf7c | 9592 | |
wolfSSL | 15:117db924cf7c | 9593 | |
wolfSSL | 15:117db924cf7c | 9594 | /* Initialize Aes for use with async hardware */ |
wolfSSL | 15:117db924cf7c | 9595 | int wc_AesInit(Aes* aes, void* heap, int devId) |
wolfSSL | 15:117db924cf7c | 9596 | { |
wolfSSL | 15:117db924cf7c | 9597 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 9598 | |
wolfSSL | 15:117db924cf7c | 9599 | if (aes == NULL) |
wolfSSL | 15:117db924cf7c | 9600 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9601 | |
wolfSSL | 15:117db924cf7c | 9602 | aes->heap = heap; |
wolfSSL | 15:117db924cf7c | 9603 | |
wolfSSL | 15:117db924cf7c | 9604 | #if defined(WOLFSSL_ASYNC_CRYPT) && defined(WC_ASYNC_ENABLE_AES) |
wolfSSL | 15:117db924cf7c | 9605 | ret = wolfAsync_DevCtxInit(&aes->asyncDev, WOLFSSL_ASYNC_MARKER_AES, |
wolfSSL | 15:117db924cf7c | 9606 | aes->heap, devId); |
wolfSSL | 15:117db924cf7c | 9607 | #else |
wolfSSL | 15:117db924cf7c | 9608 | (void)devId; |
wolfSSL | 15:117db924cf7c | 9609 | #endif /* WOLFSSL_ASYNC_CRYPT */ |
wolfSSL | 15:117db924cf7c | 9610 | |
wolfSSL | 15:117db924cf7c | 9611 | return ret; |
wolfSSL | 15:117db924cf7c | 9612 | } |
wolfSSL | 15:117db924cf7c | 9613 | |
wolfSSL | 15:117db924cf7c | 9614 | /* Free Aes from use with async hardware */ |
wolfSSL | 15:117db924cf7c | 9615 | void wc_AesFree(Aes* aes) |
wolfSSL | 15:117db924cf7c | 9616 | { |
wolfSSL | 15:117db924cf7c | 9617 | if (aes == NULL) |
wolfSSL | 15:117db924cf7c | 9618 | return; |
wolfSSL | 15:117db924cf7c | 9619 | |
wolfSSL | 15:117db924cf7c | 9620 | #if defined(WOLFSSL_ASYNC_CRYPT) && defined(WC_ASYNC_ENABLE_AES) |
wolfSSL | 15:117db924cf7c | 9621 | wolfAsync_DevCtxFree(&aes->asyncDev, WOLFSSL_ASYNC_MARKER_AES); |
wolfSSL | 15:117db924cf7c | 9622 | #endif /* WOLFSSL_ASYNC_CRYPT */ |
wolfSSL | 15:117db924cf7c | 9623 | } |
wolfSSL | 15:117db924cf7c | 9624 | |
wolfSSL | 15:117db924cf7c | 9625 | |
wolfSSL | 15:117db924cf7c | 9626 | int wc_AesGetKeySize(Aes* aes, word32* keySize) |
wolfSSL | 15:117db924cf7c | 9627 | { |
wolfSSL | 15:117db924cf7c | 9628 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 9629 | |
wolfSSL | 15:117db924cf7c | 9630 | if (aes == NULL || keySize == NULL) { |
wolfSSL | 15:117db924cf7c | 9631 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9632 | } |
wolfSSL | 15:117db924cf7c | 9633 | |
wolfSSL | 15:117db924cf7c | 9634 | switch (aes->rounds) { |
wolfSSL | 15:117db924cf7c | 9635 | #ifdef WOLFSSL_AES_128 |
wolfSSL | 15:117db924cf7c | 9636 | case 10: |
wolfSSL | 15:117db924cf7c | 9637 | *keySize = 16; |
wolfSSL | 15:117db924cf7c | 9638 | break; |
wolfSSL | 15:117db924cf7c | 9639 | #endif |
wolfSSL | 15:117db924cf7c | 9640 | #ifdef WOLFSSL_AES_192 |
wolfSSL | 15:117db924cf7c | 9641 | case 12: |
wolfSSL | 15:117db924cf7c | 9642 | *keySize = 24; |
wolfSSL | 15:117db924cf7c | 9643 | break; |
wolfSSL | 15:117db924cf7c | 9644 | #endif |
wolfSSL | 15:117db924cf7c | 9645 | #ifdef WOLFSSL_AES_256 |
wolfSSL | 15:117db924cf7c | 9646 | case 14: |
wolfSSL | 15:117db924cf7c | 9647 | *keySize = 32; |
wolfSSL | 15:117db924cf7c | 9648 | break; |
wolfSSL | 15:117db924cf7c | 9649 | #endif |
wolfSSL | 15:117db924cf7c | 9650 | default: |
wolfSSL | 15:117db924cf7c | 9651 | *keySize = 0; |
wolfSSL | 15:117db924cf7c | 9652 | ret = BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9653 | } |
wolfSSL | 15:117db924cf7c | 9654 | |
wolfSSL | 15:117db924cf7c | 9655 | return ret; |
wolfSSL | 15:117db924cf7c | 9656 | } |
wolfSSL | 15:117db924cf7c | 9657 | |
wolfSSL | 15:117db924cf7c | 9658 | #endif /* !WOLFSSL_TI_CRYPT */ |
wolfSSL | 15:117db924cf7c | 9659 | |
wolfSSL | 15:117db924cf7c | 9660 | #ifdef HAVE_AES_ECB |
wolfSSL | 15:117db924cf7c | 9661 | #if defined(WOLFSSL_IMX6_CAAM) && !defined(NO_IMX6_CAAM_AES) |
wolfSSL | 15:117db924cf7c | 9662 | /* implemented in wolfcrypt/src/port/caam/caam_aes.c */ |
wolfSSL | 15:117db924cf7c | 9663 | #else |
wolfSSL | 15:117db924cf7c | 9664 | |
wolfSSL | 15:117db924cf7c | 9665 | /* software implementation */ |
wolfSSL | 15:117db924cf7c | 9666 | int wc_AesEcbEncrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 9667 | { |
wolfSSL | 15:117db924cf7c | 9668 | word32 blocks = sz / AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9669 | |
wolfSSL | 15:117db924cf7c | 9670 | if ((in == NULL) || (out == NULL) || (aes == NULL)) |
wolfSSL | 15:117db924cf7c | 9671 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9672 | while (blocks>0) { |
wolfSSL | 15:117db924cf7c | 9673 | wc_AesEncryptDirect(aes, out, in); |
wolfSSL | 15:117db924cf7c | 9674 | out += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9675 | in += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9676 | sz -= AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9677 | blocks--; |
wolfSSL | 15:117db924cf7c | 9678 | } |
wolfSSL | 15:117db924cf7c | 9679 | return 0; |
wolfSSL | 15:117db924cf7c | 9680 | } |
wolfSSL | 15:117db924cf7c | 9681 | |
wolfSSL | 15:117db924cf7c | 9682 | |
wolfSSL | 15:117db924cf7c | 9683 | int wc_AesEcbDecrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 9684 | { |
wolfSSL | 15:117db924cf7c | 9685 | word32 blocks = sz / AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9686 | |
wolfSSL | 15:117db924cf7c | 9687 | if ((in == NULL) || (out == NULL) || (aes == NULL)) |
wolfSSL | 15:117db924cf7c | 9688 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9689 | while (blocks>0) { |
wolfSSL | 15:117db924cf7c | 9690 | wc_AesDecryptDirect(aes, out, in); |
wolfSSL | 15:117db924cf7c | 9691 | out += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9692 | in += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9693 | sz -= AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9694 | blocks--; |
wolfSSL | 15:117db924cf7c | 9695 | } |
wolfSSL | 15:117db924cf7c | 9696 | return 0; |
wolfSSL | 15:117db924cf7c | 9697 | } |
wolfSSL | 15:117db924cf7c | 9698 | #endif |
wolfSSL | 15:117db924cf7c | 9699 | #endif /* HAVE_AES_ECB */ |
wolfSSL | 15:117db924cf7c | 9700 | |
wolfSSL | 15:117db924cf7c | 9701 | #ifdef WOLFSSL_AES_CFB |
wolfSSL | 15:117db924cf7c | 9702 | /* CFB 128 |
wolfSSL | 15:117db924cf7c | 9703 | * |
wolfSSL | 15:117db924cf7c | 9704 | * aes structure holding key to use for encryption |
wolfSSL | 15:117db924cf7c | 9705 | * out buffer to hold result of encryption (must be at least as large as input |
wolfSSL | 15:117db924cf7c | 9706 | * buffer) |
wolfSSL | 15:117db924cf7c | 9707 | * in buffer to encrypt |
wolfSSL | 15:117db924cf7c | 9708 | * sz size of input buffer |
wolfSSL | 15:117db924cf7c | 9709 | * |
wolfSSL | 15:117db924cf7c | 9710 | * returns 0 on success and negative error values on failure |
wolfSSL | 15:117db924cf7c | 9711 | */ |
wolfSSL | 15:117db924cf7c | 9712 | int wc_AesCfbEncrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 9713 | { |
wolfSSL | 15:117db924cf7c | 9714 | byte* tmp = NULL; |
wolfSSL | 15:117db924cf7c | 9715 | byte* reg = NULL; |
wolfSSL | 15:117db924cf7c | 9716 | |
wolfSSL | 15:117db924cf7c | 9717 | WOLFSSL_ENTER("wc_AesCfbEncrypt"); |
wolfSSL | 15:117db924cf7c | 9718 | |
wolfSSL | 15:117db924cf7c | 9719 | if (aes == NULL || out == NULL || in == NULL) { |
wolfSSL | 15:117db924cf7c | 9720 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9721 | } |
wolfSSL | 15:117db924cf7c | 9722 | |
wolfSSL | 15:117db924cf7c | 9723 | if (aes->left && sz) { |
wolfSSL | 15:117db924cf7c | 9724 | reg = (byte*)aes->reg + AES_BLOCK_SIZE - aes->left; |
wolfSSL | 15:117db924cf7c | 9725 | } |
wolfSSL | 15:117db924cf7c | 9726 | |
wolfSSL | 15:117db924cf7c | 9727 | /* consume any unused bytes left in aes->tmp */ |
wolfSSL | 15:117db924cf7c | 9728 | tmp = (byte*)aes->tmp + AES_BLOCK_SIZE - aes->left; |
wolfSSL | 15:117db924cf7c | 9729 | while (aes->left && sz) { |
wolfSSL | 15:117db924cf7c | 9730 | *(out++) = *(reg++) = *(in++) ^ *(tmp++); |
wolfSSL | 15:117db924cf7c | 9731 | aes->left--; |
wolfSSL | 15:117db924cf7c | 9732 | sz--; |
wolfSSL | 15:117db924cf7c | 9733 | } |
wolfSSL | 15:117db924cf7c | 9734 | |
wolfSSL | 15:117db924cf7c | 9735 | while (sz >= AES_BLOCK_SIZE) { |
wolfSSL | 15:117db924cf7c | 9736 | wc_AesEncryptDirect(aes, out, (byte*)aes->reg); |
wolfSSL | 15:117db924cf7c | 9737 | xorbuf(out, in, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9738 | XMEMCPY(aes->reg, out, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9739 | out += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9740 | in += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9741 | sz -= AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9742 | aes->left = 0; |
wolfSSL | 15:117db924cf7c | 9743 | } |
wolfSSL | 15:117db924cf7c | 9744 | |
wolfSSL | 15:117db924cf7c | 9745 | /* encrypt left over data */ |
wolfSSL | 15:117db924cf7c | 9746 | if (sz) { |
wolfSSL | 15:117db924cf7c | 9747 | wc_AesEncryptDirect(aes, (byte*)aes->tmp, (byte*)aes->reg); |
wolfSSL | 15:117db924cf7c | 9748 | aes->left = AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9749 | tmp = (byte*)aes->tmp; |
wolfSSL | 15:117db924cf7c | 9750 | reg = (byte*)aes->reg; |
wolfSSL | 15:117db924cf7c | 9751 | |
wolfSSL | 15:117db924cf7c | 9752 | while (sz--) { |
wolfSSL | 15:117db924cf7c | 9753 | *(out++) = *(reg++) = *(in++) ^ *(tmp++); |
wolfSSL | 15:117db924cf7c | 9754 | aes->left--; |
wolfSSL | 15:117db924cf7c | 9755 | } |
wolfSSL | 15:117db924cf7c | 9756 | } |
wolfSSL | 15:117db924cf7c | 9757 | |
wolfSSL | 15:117db924cf7c | 9758 | return 0; |
wolfSSL | 15:117db924cf7c | 9759 | } |
wolfSSL | 15:117db924cf7c | 9760 | |
wolfSSL | 15:117db924cf7c | 9761 | |
wolfSSL | 15:117db924cf7c | 9762 | #ifdef HAVE_AES_DECRYPT |
wolfSSL | 15:117db924cf7c | 9763 | /* CFB 128 |
wolfSSL | 15:117db924cf7c | 9764 | * |
wolfSSL | 15:117db924cf7c | 9765 | * aes structure holding key to use for decryption |
wolfSSL | 15:117db924cf7c | 9766 | * out buffer to hold result of decryption (must be at least as large as input |
wolfSSL | 15:117db924cf7c | 9767 | * buffer) |
wolfSSL | 15:117db924cf7c | 9768 | * in buffer to decrypt |
wolfSSL | 15:117db924cf7c | 9769 | * sz size of input buffer |
wolfSSL | 15:117db924cf7c | 9770 | * |
wolfSSL | 15:117db924cf7c | 9771 | * returns 0 on success and negative error values on failure |
wolfSSL | 15:117db924cf7c | 9772 | */ |
wolfSSL | 15:117db924cf7c | 9773 | int wc_AesCfbDecrypt(Aes* aes, byte* out, const byte* in, word32 sz) |
wolfSSL | 15:117db924cf7c | 9774 | { |
wolfSSL | 15:117db924cf7c | 9775 | byte* tmp; |
wolfSSL | 15:117db924cf7c | 9776 | |
wolfSSL | 15:117db924cf7c | 9777 | WOLFSSL_ENTER("wc_AesCfbDecrypt"); |
wolfSSL | 15:117db924cf7c | 9778 | |
wolfSSL | 15:117db924cf7c | 9779 | if (aes == NULL || out == NULL || in == NULL) { |
wolfSSL | 15:117db924cf7c | 9780 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9781 | } |
wolfSSL | 15:117db924cf7c | 9782 | |
wolfSSL | 15:117db924cf7c | 9783 | /* check if more input needs copied over to aes->reg */ |
wolfSSL | 15:117db924cf7c | 9784 | if (aes->left && sz) { |
wolfSSL | 15:117db924cf7c | 9785 | int size = min(aes->left, sz); |
wolfSSL | 15:117db924cf7c | 9786 | XMEMCPY((byte*)aes->reg + AES_BLOCK_SIZE - aes->left, in, size); |
wolfSSL | 15:117db924cf7c | 9787 | } |
wolfSSL | 15:117db924cf7c | 9788 | |
wolfSSL | 15:117db924cf7c | 9789 | /* consume any unused bytes left in aes->tmp */ |
wolfSSL | 15:117db924cf7c | 9790 | tmp = (byte*)aes->tmp + AES_BLOCK_SIZE - aes->left; |
wolfSSL | 15:117db924cf7c | 9791 | while (aes->left && sz) { |
wolfSSL | 15:117db924cf7c | 9792 | *(out++) = *(in++) ^ *(tmp++); |
wolfSSL | 15:117db924cf7c | 9793 | aes->left--; |
wolfSSL | 15:117db924cf7c | 9794 | sz--; |
wolfSSL | 15:117db924cf7c | 9795 | } |
wolfSSL | 15:117db924cf7c | 9796 | |
wolfSSL | 15:117db924cf7c | 9797 | while (sz > AES_BLOCK_SIZE) { |
wolfSSL | 15:117db924cf7c | 9798 | wc_AesEncryptDirect(aes, out, (byte*)aes->reg); |
wolfSSL | 15:117db924cf7c | 9799 | xorbuf(out, in, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9800 | XMEMCPY(aes->reg, in, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9801 | out += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9802 | in += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9803 | sz -= AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9804 | aes->left = 0; |
wolfSSL | 15:117db924cf7c | 9805 | } |
wolfSSL | 15:117db924cf7c | 9806 | |
wolfSSL | 15:117db924cf7c | 9807 | /* decrypt left over data */ |
wolfSSL | 15:117db924cf7c | 9808 | if (sz) { |
wolfSSL | 15:117db924cf7c | 9809 | wc_AesEncryptDirect(aes, (byte*)aes->tmp, (byte*)aes->reg); |
wolfSSL | 15:117db924cf7c | 9810 | XMEMCPY(aes->reg, in, sz); |
wolfSSL | 15:117db924cf7c | 9811 | aes->left = AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9812 | tmp = (byte*)aes->tmp; |
wolfSSL | 15:117db924cf7c | 9813 | |
wolfSSL | 15:117db924cf7c | 9814 | while (sz--) { |
wolfSSL | 15:117db924cf7c | 9815 | *(out++) = *(in++) ^ *(tmp++); |
wolfSSL | 15:117db924cf7c | 9816 | aes->left--; |
wolfSSL | 15:117db924cf7c | 9817 | } |
wolfSSL | 15:117db924cf7c | 9818 | } |
wolfSSL | 15:117db924cf7c | 9819 | |
wolfSSL | 15:117db924cf7c | 9820 | return 0; |
wolfSSL | 15:117db924cf7c | 9821 | } |
wolfSSL | 15:117db924cf7c | 9822 | #endif /* HAVE_AES_DECRYPT */ |
wolfSSL | 15:117db924cf7c | 9823 | #endif /* WOLFSSL_AES_CFB */ |
wolfSSL | 15:117db924cf7c | 9824 | |
wolfSSL | 15:117db924cf7c | 9825 | |
wolfSSL | 15:117db924cf7c | 9826 | #ifdef HAVE_AES_KEYWRAP |
wolfSSL | 15:117db924cf7c | 9827 | |
wolfSSL | 15:117db924cf7c | 9828 | /* Initialize key wrap counter with value */ |
wolfSSL | 15:117db924cf7c | 9829 | static WC_INLINE void InitKeyWrapCounter(byte* inOutCtr, word32 value) |
wolfSSL | 15:117db924cf7c | 9830 | { |
wolfSSL | 15:117db924cf7c | 9831 | int i; |
wolfSSL | 15:117db924cf7c | 9832 | word32 bytes; |
wolfSSL | 15:117db924cf7c | 9833 | |
wolfSSL | 15:117db924cf7c | 9834 | bytes = sizeof(word32); |
wolfSSL | 15:117db924cf7c | 9835 | for (i = 0; i < (int)sizeof(word32); i++) { |
wolfSSL | 15:117db924cf7c | 9836 | inOutCtr[i+sizeof(word32)] = (value >> ((bytes - 1) * 8)) & 0xFF; |
wolfSSL | 15:117db924cf7c | 9837 | bytes--; |
wolfSSL | 15:117db924cf7c | 9838 | } |
wolfSSL | 15:117db924cf7c | 9839 | } |
wolfSSL | 15:117db924cf7c | 9840 | |
wolfSSL | 15:117db924cf7c | 9841 | /* Increment key wrap counter */ |
wolfSSL | 15:117db924cf7c | 9842 | static WC_INLINE void IncrementKeyWrapCounter(byte* inOutCtr) |
wolfSSL | 15:117db924cf7c | 9843 | { |
wolfSSL | 15:117db924cf7c | 9844 | int i; |
wolfSSL | 15:117db924cf7c | 9845 | |
wolfSSL | 15:117db924cf7c | 9846 | /* in network byte order so start at end and work back */ |
wolfSSL | 15:117db924cf7c | 9847 | for (i = KEYWRAP_BLOCK_SIZE - 1; i >= 0; i--) { |
wolfSSL | 15:117db924cf7c | 9848 | if (++inOutCtr[i]) /* we're done unless we overflow */ |
wolfSSL | 15:117db924cf7c | 9849 | return; |
wolfSSL | 15:117db924cf7c | 9850 | } |
wolfSSL | 15:117db924cf7c | 9851 | } |
wolfSSL | 15:117db924cf7c | 9852 | |
wolfSSL | 15:117db924cf7c | 9853 | /* Decrement key wrap counter */ |
wolfSSL | 15:117db924cf7c | 9854 | static WC_INLINE void DecrementKeyWrapCounter(byte* inOutCtr) |
wolfSSL | 15:117db924cf7c | 9855 | { |
wolfSSL | 15:117db924cf7c | 9856 | int i; |
wolfSSL | 15:117db924cf7c | 9857 | |
wolfSSL | 15:117db924cf7c | 9858 | for (i = KEYWRAP_BLOCK_SIZE - 1; i >= 0; i--) { |
wolfSSL | 15:117db924cf7c | 9859 | if (--inOutCtr[i] != 0xFF) /* we're done unless we underflow */ |
wolfSSL | 15:117db924cf7c | 9860 | return; |
wolfSSL | 15:117db924cf7c | 9861 | } |
wolfSSL | 15:117db924cf7c | 9862 | } |
wolfSSL | 15:117db924cf7c | 9863 | |
wolfSSL | 15:117db924cf7c | 9864 | /* perform AES key wrap (RFC3394), return out sz on success, negative on err */ |
wolfSSL | 15:117db924cf7c | 9865 | int wc_AesKeyWrap(const byte* key, word32 keySz, const byte* in, word32 inSz, |
wolfSSL | 15:117db924cf7c | 9866 | byte* out, word32 outSz, const byte* iv) |
wolfSSL | 15:117db924cf7c | 9867 | { |
wolfSSL | 15:117db924cf7c | 9868 | Aes aes; |
wolfSSL | 15:117db924cf7c | 9869 | byte* r; |
wolfSSL | 15:117db924cf7c | 9870 | word32 i; |
wolfSSL | 15:117db924cf7c | 9871 | int ret, j; |
wolfSSL | 15:117db924cf7c | 9872 | |
wolfSSL | 15:117db924cf7c | 9873 | byte t[KEYWRAP_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 9874 | byte tmp[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 9875 | |
wolfSSL | 15:117db924cf7c | 9876 | /* n must be at least 2, output size is n + 8 bytes */ |
wolfSSL | 15:117db924cf7c | 9877 | if (key == NULL || in == NULL || inSz < 2 || |
wolfSSL | 15:117db924cf7c | 9878 | out == NULL || outSz < (inSz + KEYWRAP_BLOCK_SIZE)) |
wolfSSL | 15:117db924cf7c | 9879 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9880 | |
wolfSSL | 15:117db924cf7c | 9881 | /* input must be multiple of 64-bits */ |
wolfSSL | 15:117db924cf7c | 9882 | if (inSz % KEYWRAP_BLOCK_SIZE != 0) |
wolfSSL | 15:117db924cf7c | 9883 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9884 | |
wolfSSL | 15:117db924cf7c | 9885 | /* user IV is optional */ |
wolfSSL | 15:117db924cf7c | 9886 | if (iv == NULL) { |
wolfSSL | 15:117db924cf7c | 9887 | XMEMSET(tmp, 0xA6, KEYWRAP_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9888 | } else { |
wolfSSL | 15:117db924cf7c | 9889 | XMEMCPY(tmp, iv, KEYWRAP_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9890 | } |
wolfSSL | 15:117db924cf7c | 9891 | |
wolfSSL | 15:117db924cf7c | 9892 | r = out + 8; |
wolfSSL | 15:117db924cf7c | 9893 | XMEMCPY(r, in, inSz); |
wolfSSL | 15:117db924cf7c | 9894 | XMEMSET(t, 0, sizeof(t)); |
wolfSSL | 15:117db924cf7c | 9895 | |
wolfSSL | 15:117db924cf7c | 9896 | ret = wc_AesInit(&aes, NULL, INVALID_DEVID); |
wolfSSL | 15:117db924cf7c | 9897 | if (ret != 0) |
wolfSSL | 15:117db924cf7c | 9898 | return ret; |
wolfSSL | 15:117db924cf7c | 9899 | |
wolfSSL | 15:117db924cf7c | 9900 | ret = wc_AesSetKey(&aes, key, keySz, NULL, AES_ENCRYPTION); |
wolfSSL | 15:117db924cf7c | 9901 | if (ret != 0) |
wolfSSL | 15:117db924cf7c | 9902 | return ret; |
wolfSSL | 15:117db924cf7c | 9903 | |
wolfSSL | 15:117db924cf7c | 9904 | for (j = 0; j <= 5; j++) { |
wolfSSL | 15:117db924cf7c | 9905 | for (i = 1; i <= inSz / KEYWRAP_BLOCK_SIZE; i++) { |
wolfSSL | 15:117db924cf7c | 9906 | |
wolfSSL | 15:117db924cf7c | 9907 | /* load R[i] */ |
wolfSSL | 15:117db924cf7c | 9908 | XMEMCPY(tmp + KEYWRAP_BLOCK_SIZE, r, KEYWRAP_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9909 | |
wolfSSL | 15:117db924cf7c | 9910 | wc_AesEncryptDirect(&aes, tmp, tmp); |
wolfSSL | 15:117db924cf7c | 9911 | |
wolfSSL | 15:117db924cf7c | 9912 | /* calculate new A */ |
wolfSSL | 15:117db924cf7c | 9913 | IncrementKeyWrapCounter(t); |
wolfSSL | 15:117db924cf7c | 9914 | xorbuf(tmp, t, KEYWRAP_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9915 | |
wolfSSL | 15:117db924cf7c | 9916 | /* save R[i] */ |
wolfSSL | 15:117db924cf7c | 9917 | XMEMCPY(r, tmp + KEYWRAP_BLOCK_SIZE, KEYWRAP_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9918 | r += KEYWRAP_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9919 | } |
wolfSSL | 15:117db924cf7c | 9920 | r = out + KEYWRAP_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9921 | } |
wolfSSL | 15:117db924cf7c | 9922 | |
wolfSSL | 15:117db924cf7c | 9923 | /* C[0] = A */ |
wolfSSL | 15:117db924cf7c | 9924 | XMEMCPY(out, tmp, KEYWRAP_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9925 | |
wolfSSL | 15:117db924cf7c | 9926 | wc_AesFree(&aes); |
wolfSSL | 15:117db924cf7c | 9927 | |
wolfSSL | 15:117db924cf7c | 9928 | return inSz + KEYWRAP_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9929 | } |
wolfSSL | 15:117db924cf7c | 9930 | |
wolfSSL | 15:117db924cf7c | 9931 | int wc_AesKeyUnWrap(const byte* key, word32 keySz, const byte* in, word32 inSz, |
wolfSSL | 15:117db924cf7c | 9932 | byte* out, word32 outSz, const byte* iv) |
wolfSSL | 15:117db924cf7c | 9933 | { |
wolfSSL | 15:117db924cf7c | 9934 | Aes aes; |
wolfSSL | 15:117db924cf7c | 9935 | byte* r; |
wolfSSL | 15:117db924cf7c | 9936 | word32 i, n; |
wolfSSL | 15:117db924cf7c | 9937 | int ret, j; |
wolfSSL | 15:117db924cf7c | 9938 | |
wolfSSL | 15:117db924cf7c | 9939 | byte t[KEYWRAP_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 9940 | byte tmp[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 9941 | |
wolfSSL | 15:117db924cf7c | 9942 | const byte* expIv; |
wolfSSL | 15:117db924cf7c | 9943 | const byte defaultIV[] = { |
wolfSSL | 15:117db924cf7c | 9944 | 0xA6, 0xA6, 0xA6, 0xA6, 0xA6, 0xA6, 0xA6, 0xA6 |
wolfSSL | 15:117db924cf7c | 9945 | }; |
wolfSSL | 15:117db924cf7c | 9946 | |
wolfSSL | 15:117db924cf7c | 9947 | (void)iv; |
wolfSSL | 15:117db924cf7c | 9948 | |
wolfSSL | 15:117db924cf7c | 9949 | if (key == NULL || in == NULL || inSz < 3 || |
wolfSSL | 15:117db924cf7c | 9950 | out == NULL || outSz < (inSz - KEYWRAP_BLOCK_SIZE)) |
wolfSSL | 15:117db924cf7c | 9951 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9952 | |
wolfSSL | 15:117db924cf7c | 9953 | /* input must be multiple of 64-bits */ |
wolfSSL | 15:117db924cf7c | 9954 | if (inSz % KEYWRAP_BLOCK_SIZE != 0) |
wolfSSL | 15:117db924cf7c | 9955 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 9956 | |
wolfSSL | 15:117db924cf7c | 9957 | /* user IV optional */ |
wolfSSL | 15:117db924cf7c | 9958 | if (iv != NULL) { |
wolfSSL | 15:117db924cf7c | 9959 | expIv = iv; |
wolfSSL | 15:117db924cf7c | 9960 | } else { |
wolfSSL | 15:117db924cf7c | 9961 | expIv = defaultIV; |
wolfSSL | 15:117db924cf7c | 9962 | } |
wolfSSL | 15:117db924cf7c | 9963 | |
wolfSSL | 15:117db924cf7c | 9964 | /* A = C[0], R[i] = C[i] */ |
wolfSSL | 15:117db924cf7c | 9965 | XMEMCPY(tmp, in, KEYWRAP_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9966 | XMEMCPY(out, in + KEYWRAP_BLOCK_SIZE, inSz - KEYWRAP_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9967 | XMEMSET(t, 0, sizeof(t)); |
wolfSSL | 15:117db924cf7c | 9968 | |
wolfSSL | 15:117db924cf7c | 9969 | ret = wc_AesInit(&aes, NULL, INVALID_DEVID); |
wolfSSL | 15:117db924cf7c | 9970 | if (ret != 0) |
wolfSSL | 15:117db924cf7c | 9971 | return ret; |
wolfSSL | 15:117db924cf7c | 9972 | |
wolfSSL | 15:117db924cf7c | 9973 | ret = wc_AesSetKey(&aes, key, keySz, NULL, AES_DECRYPTION); |
wolfSSL | 15:117db924cf7c | 9974 | if (ret != 0) |
wolfSSL | 15:117db924cf7c | 9975 | return ret; |
wolfSSL | 15:117db924cf7c | 9976 | |
wolfSSL | 15:117db924cf7c | 9977 | /* initialize counter to 6n */ |
wolfSSL | 15:117db924cf7c | 9978 | n = (inSz - 1) / KEYWRAP_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 9979 | InitKeyWrapCounter(t, 6 * n); |
wolfSSL | 15:117db924cf7c | 9980 | |
wolfSSL | 15:117db924cf7c | 9981 | for (j = 5; j >= 0; j--) { |
wolfSSL | 15:117db924cf7c | 9982 | for (i = n; i >= 1; i--) { |
wolfSSL | 15:117db924cf7c | 9983 | |
wolfSSL | 15:117db924cf7c | 9984 | /* calculate A */ |
wolfSSL | 15:117db924cf7c | 9985 | xorbuf(tmp, t, KEYWRAP_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9986 | DecrementKeyWrapCounter(t); |
wolfSSL | 15:117db924cf7c | 9987 | |
wolfSSL | 15:117db924cf7c | 9988 | /* load R[i], starting at end of R */ |
wolfSSL | 15:117db924cf7c | 9989 | r = out + ((i - 1) * KEYWRAP_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9990 | XMEMCPY(tmp + KEYWRAP_BLOCK_SIZE, r, KEYWRAP_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9991 | wc_AesDecryptDirect(&aes, tmp, tmp); |
wolfSSL | 15:117db924cf7c | 9992 | |
wolfSSL | 15:117db924cf7c | 9993 | /* save R[i] */ |
wolfSSL | 15:117db924cf7c | 9994 | XMEMCPY(r, tmp + KEYWRAP_BLOCK_SIZE, KEYWRAP_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 9995 | } |
wolfSSL | 15:117db924cf7c | 9996 | } |
wolfSSL | 15:117db924cf7c | 9997 | |
wolfSSL | 15:117db924cf7c | 9998 | wc_AesFree(&aes); |
wolfSSL | 15:117db924cf7c | 9999 | |
wolfSSL | 15:117db924cf7c | 10000 | /* verify IV */ |
wolfSSL | 15:117db924cf7c | 10001 | if (XMEMCMP(tmp, expIv, KEYWRAP_BLOCK_SIZE) != 0) |
wolfSSL | 15:117db924cf7c | 10002 | return BAD_KEYWRAP_IV_E; |
wolfSSL | 15:117db924cf7c | 10003 | |
wolfSSL | 15:117db924cf7c | 10004 | return inSz - KEYWRAP_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 10005 | } |
wolfSSL | 15:117db924cf7c | 10006 | |
wolfSSL | 15:117db924cf7c | 10007 | #endif /* HAVE_AES_KEYWRAP */ |
wolfSSL | 15:117db924cf7c | 10008 | |
wolfSSL | 15:117db924cf7c | 10009 | #ifdef WOLFSSL_AES_XTS |
wolfSSL | 15:117db924cf7c | 10010 | |
wolfSSL | 15:117db924cf7c | 10011 | /* Galios Field to use */ |
wolfSSL | 15:117db924cf7c | 10012 | #define GF_XTS 0x87 |
wolfSSL | 15:117db924cf7c | 10013 | |
wolfSSL | 15:117db924cf7c | 10014 | /* This is to help with setting keys to correct encrypt or decrypt type. |
wolfSSL | 15:117db924cf7c | 10015 | * |
wolfSSL | 15:117db924cf7c | 10016 | * tweak AES key for tweak in XTS |
wolfSSL | 15:117db924cf7c | 10017 | * aes AES key for encrypt/decrypt process |
wolfSSL | 15:117db924cf7c | 10018 | * key buffer holding aes key | tweak key |
wolfSSL | 15:117db924cf7c | 10019 | * len length of key buffer in bytes. Should be twice that of key size. i.e. |
wolfSSL | 15:117db924cf7c | 10020 | * 32 for a 16 byte key. |
wolfSSL | 15:117db924cf7c | 10021 | * dir direction, either AES_ENCRYPTION or AES_DECRYPTION |
wolfSSL | 15:117db924cf7c | 10022 | * heap heap hint to use for memory. Can be NULL |
wolfSSL | 15:117db924cf7c | 10023 | * devId id to use with async crypto. Can be 0 |
wolfSSL | 15:117db924cf7c | 10024 | * |
wolfSSL | 15:117db924cf7c | 10025 | * Note: is up to user to call wc_AesFree on tweak and aes key when done. |
wolfSSL | 15:117db924cf7c | 10026 | * |
wolfSSL | 15:117db924cf7c | 10027 | * return 0 on success |
wolfSSL | 15:117db924cf7c | 10028 | */ |
wolfSSL | 15:117db924cf7c | 10029 | int wc_AesXtsSetKey(XtsAes* aes, const byte* key, word32 len, int dir, |
wolfSSL | 15:117db924cf7c | 10030 | void* heap, int devId) |
wolfSSL | 15:117db924cf7c | 10031 | { |
wolfSSL | 15:117db924cf7c | 10032 | word32 keySz; |
wolfSSL | 15:117db924cf7c | 10033 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 10034 | |
wolfSSL | 15:117db924cf7c | 10035 | if (aes == NULL || key == NULL) { |
wolfSSL | 15:117db924cf7c | 10036 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 10037 | } |
wolfSSL | 15:117db924cf7c | 10038 | |
wolfSSL | 15:117db924cf7c | 10039 | if ((ret = wc_AesInit(&aes->tweak, heap, devId)) != 0) { |
wolfSSL | 15:117db924cf7c | 10040 | return ret; |
wolfSSL | 15:117db924cf7c | 10041 | } |
wolfSSL | 15:117db924cf7c | 10042 | if ((ret = wc_AesInit(&aes->aes, heap, devId)) != 0) { |
wolfSSL | 15:117db924cf7c | 10043 | return ret; |
wolfSSL | 15:117db924cf7c | 10044 | } |
wolfSSL | 15:117db924cf7c | 10045 | |
wolfSSL | 15:117db924cf7c | 10046 | keySz = len/2; |
wolfSSL | 15:117db924cf7c | 10047 | if (keySz != 16 && keySz != 32) { |
wolfSSL | 15:117db924cf7c | 10048 | WOLFSSL_MSG("Unsupported key size"); |
wolfSSL | 15:117db924cf7c | 10049 | return WC_KEY_SIZE_E; |
wolfSSL | 15:117db924cf7c | 10050 | } |
wolfSSL | 15:117db924cf7c | 10051 | |
wolfSSL | 15:117db924cf7c | 10052 | if ((ret = wc_AesSetKey(&aes->aes, key, keySz, NULL, dir)) == 0) { |
wolfSSL | 15:117db924cf7c | 10053 | ret = wc_AesSetKey(&aes->tweak, key + keySz, keySz, NULL, |
wolfSSL | 15:117db924cf7c | 10054 | AES_ENCRYPTION); |
wolfSSL | 15:117db924cf7c | 10055 | if (ret != 0) { |
wolfSSL | 15:117db924cf7c | 10056 | wc_AesFree(&aes->aes); |
wolfSSL | 15:117db924cf7c | 10057 | } |
wolfSSL | 15:117db924cf7c | 10058 | } |
wolfSSL | 15:117db924cf7c | 10059 | |
wolfSSL | 15:117db924cf7c | 10060 | return ret; |
wolfSSL | 15:117db924cf7c | 10061 | } |
wolfSSL | 15:117db924cf7c | 10062 | |
wolfSSL | 15:117db924cf7c | 10063 | |
wolfSSL | 15:117db924cf7c | 10064 | /* This is used to free up resources used by Aes structs |
wolfSSL | 15:117db924cf7c | 10065 | * |
wolfSSL | 15:117db924cf7c | 10066 | * aes AES keys to free |
wolfSSL | 15:117db924cf7c | 10067 | * |
wolfSSL | 15:117db924cf7c | 10068 | * return 0 on success |
wolfSSL | 15:117db924cf7c | 10069 | */ |
wolfSSL | 15:117db924cf7c | 10070 | int wc_AesXtsFree(XtsAes* aes) |
wolfSSL | 15:117db924cf7c | 10071 | { |
wolfSSL | 15:117db924cf7c | 10072 | if (aes != NULL) { |
wolfSSL | 15:117db924cf7c | 10073 | wc_AesFree(&aes->aes); |
wolfSSL | 15:117db924cf7c | 10074 | wc_AesFree(&aes->tweak); |
wolfSSL | 15:117db924cf7c | 10075 | } |
wolfSSL | 15:117db924cf7c | 10076 | |
wolfSSL | 15:117db924cf7c | 10077 | return 0; |
wolfSSL | 15:117db924cf7c | 10078 | } |
wolfSSL | 15:117db924cf7c | 10079 | |
wolfSSL | 15:117db924cf7c | 10080 | |
wolfSSL | 15:117db924cf7c | 10081 | /* Same process as wc_AesXtsEncrypt but uses a word64 type as the tweak value |
wolfSSL | 15:117db924cf7c | 10082 | * instead of a byte array. This just converts the word64 to a byte array and |
wolfSSL | 15:117db924cf7c | 10083 | * calls wc_AesXtsEncrypt. |
wolfSSL | 15:117db924cf7c | 10084 | * |
wolfSSL | 15:117db924cf7c | 10085 | * aes AES keys to use for block encrypt/decrypt |
wolfSSL | 15:117db924cf7c | 10086 | * out output buffer to hold cipher text |
wolfSSL | 15:117db924cf7c | 10087 | * in input plain text buffer to encrypt |
wolfSSL | 15:117db924cf7c | 10088 | * sz size of both out and in buffers |
wolfSSL | 15:117db924cf7c | 10089 | * sector value to use for tweak |
wolfSSL | 15:117db924cf7c | 10090 | * |
wolfSSL | 15:117db924cf7c | 10091 | * returns 0 on success |
wolfSSL | 15:117db924cf7c | 10092 | */ |
wolfSSL | 15:117db924cf7c | 10093 | int wc_AesXtsEncryptSector(XtsAes* aes, byte* out, const byte* in, |
wolfSSL | 15:117db924cf7c | 10094 | word32 sz, word64 sector) |
wolfSSL | 15:117db924cf7c | 10095 | { |
wolfSSL | 15:117db924cf7c | 10096 | byte* pt; |
wolfSSL | 15:117db924cf7c | 10097 | byte i[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 10098 | |
wolfSSL | 15:117db924cf7c | 10099 | XMEMSET(i, 0, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10100 | #ifdef BIG_ENDIAN_ORDER |
wolfSSL | 15:117db924cf7c | 10101 | sector = ByteReverseWord64(sector); |
wolfSSL | 15:117db924cf7c | 10102 | #endif |
wolfSSL | 15:117db924cf7c | 10103 | pt = (byte*)§or; |
wolfSSL | 15:117db924cf7c | 10104 | XMEMCPY(i, pt, sizeof(word64)); |
wolfSSL | 15:117db924cf7c | 10105 | |
wolfSSL | 15:117db924cf7c | 10106 | return wc_AesXtsEncrypt(aes, out, in, sz, (const byte*)i, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10107 | } |
wolfSSL | 15:117db924cf7c | 10108 | |
wolfSSL | 15:117db924cf7c | 10109 | |
wolfSSL | 15:117db924cf7c | 10110 | /* Same process as wc_AesXtsDecrypt but uses a word64 type as the tweak value |
wolfSSL | 15:117db924cf7c | 10111 | * instead of a byte array. This just converts the word64 to a byte array. |
wolfSSL | 15:117db924cf7c | 10112 | * |
wolfSSL | 15:117db924cf7c | 10113 | * aes AES keys to use for block encrypt/decrypt |
wolfSSL | 15:117db924cf7c | 10114 | * out output buffer to hold plain text |
wolfSSL | 15:117db924cf7c | 10115 | * in input cipher text buffer to encrypt |
wolfSSL | 15:117db924cf7c | 10116 | * sz size of both out and in buffers |
wolfSSL | 15:117db924cf7c | 10117 | * sector value to use for tweak |
wolfSSL | 15:117db924cf7c | 10118 | * |
wolfSSL | 15:117db924cf7c | 10119 | * returns 0 on success |
wolfSSL | 15:117db924cf7c | 10120 | */ |
wolfSSL | 15:117db924cf7c | 10121 | int wc_AesXtsDecryptSector(XtsAes* aes, byte* out, const byte* in, word32 sz, |
wolfSSL | 15:117db924cf7c | 10122 | word64 sector) |
wolfSSL | 15:117db924cf7c | 10123 | { |
wolfSSL | 15:117db924cf7c | 10124 | byte* pt; |
wolfSSL | 15:117db924cf7c | 10125 | byte i[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 10126 | |
wolfSSL | 15:117db924cf7c | 10127 | XMEMSET(i, 0, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10128 | #ifdef BIG_ENDIAN_ORDER |
wolfSSL | 15:117db924cf7c | 10129 | sector = ByteReverseWord64(sector); |
wolfSSL | 15:117db924cf7c | 10130 | #endif |
wolfSSL | 15:117db924cf7c | 10131 | pt = (byte*)§or; |
wolfSSL | 15:117db924cf7c | 10132 | XMEMCPY(i, pt, sizeof(word64)); |
wolfSSL | 15:117db924cf7c | 10133 | |
wolfSSL | 15:117db924cf7c | 10134 | return wc_AesXtsDecrypt(aes, out, in, sz, (const byte*)i, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10135 | } |
wolfSSL | 15:117db924cf7c | 10136 | |
wolfSSL | 15:117db924cf7c | 10137 | #ifdef HAVE_AES_ECB |
wolfSSL | 15:117db924cf7c | 10138 | /* helper function for encrypting / decrypting full buffer at once */ |
wolfSSL | 15:117db924cf7c | 10139 | static int _AesXtsHelper(Aes* aes, byte* out, const byte* in, word32 sz, int dir) |
wolfSSL | 15:117db924cf7c | 10140 | { |
wolfSSL | 15:117db924cf7c | 10141 | word32 outSz = sz; |
wolfSSL | 15:117db924cf7c | 10142 | word32 totalSz = (sz / AES_BLOCK_SIZE) * AES_BLOCK_SIZE; /* total bytes */ |
wolfSSL | 15:117db924cf7c | 10143 | byte* pt = out; |
wolfSSL | 15:117db924cf7c | 10144 | |
wolfSSL | 15:117db924cf7c | 10145 | outSz -= AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 10146 | |
wolfSSL | 15:117db924cf7c | 10147 | while (outSz > 0) { |
wolfSSL | 15:117db924cf7c | 10148 | word32 j; |
wolfSSL | 15:117db924cf7c | 10149 | byte carry = 0; |
wolfSSL | 15:117db924cf7c | 10150 | |
wolfSSL | 15:117db924cf7c | 10151 | /* multiply by shift left and propogate carry */ |
wolfSSL | 15:117db924cf7c | 10152 | for (j = 0; j < AES_BLOCK_SIZE && outSz > 0; j++, outSz--) { |
wolfSSL | 15:117db924cf7c | 10153 | byte tmpC; |
wolfSSL | 15:117db924cf7c | 10154 | |
wolfSSL | 15:117db924cf7c | 10155 | tmpC = (pt[j] >> 7) & 0x01; |
wolfSSL | 15:117db924cf7c | 10156 | pt[j+AES_BLOCK_SIZE] = ((pt[j] << 1) + carry) & 0xFF; |
wolfSSL | 15:117db924cf7c | 10157 | carry = tmpC; |
wolfSSL | 15:117db924cf7c | 10158 | } |
wolfSSL | 15:117db924cf7c | 10159 | if (carry) { |
wolfSSL | 15:117db924cf7c | 10160 | pt[AES_BLOCK_SIZE] ^= GF_XTS; |
wolfSSL | 15:117db924cf7c | 10161 | } |
wolfSSL | 15:117db924cf7c | 10162 | |
wolfSSL | 15:117db924cf7c | 10163 | pt += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 10164 | } |
wolfSSL | 15:117db924cf7c | 10165 | |
wolfSSL | 15:117db924cf7c | 10166 | xorbuf(out, in, totalSz); |
wolfSSL | 15:117db924cf7c | 10167 | if (dir == AES_ENCRYPTION) { |
wolfSSL | 15:117db924cf7c | 10168 | return wc_AesEcbEncrypt(aes, out, out, totalSz); |
wolfSSL | 15:117db924cf7c | 10169 | } |
wolfSSL | 15:117db924cf7c | 10170 | else { |
wolfSSL | 15:117db924cf7c | 10171 | return wc_AesEcbDecrypt(aes, out, out, totalSz); |
wolfSSL | 15:117db924cf7c | 10172 | } |
wolfSSL | 15:117db924cf7c | 10173 | } |
wolfSSL | 15:117db924cf7c | 10174 | #endif /* HAVE_AES_ECB */ |
wolfSSL | 15:117db924cf7c | 10175 | |
wolfSSL | 15:117db924cf7c | 10176 | |
wolfSSL | 15:117db924cf7c | 10177 | /* AES with XTS mode. (XTS) XEX encryption with Tweak and cipher text Stealing. |
wolfSSL | 15:117db924cf7c | 10178 | * |
wolfSSL | 15:117db924cf7c | 10179 | * xaes AES keys to use for block encrypt/decrypt |
wolfSSL | 15:117db924cf7c | 10180 | * out output buffer to hold cipher text |
wolfSSL | 15:117db924cf7c | 10181 | * in input plain text buffer to encrypt |
wolfSSL | 15:117db924cf7c | 10182 | * sz size of both out and in buffers |
wolfSSL | 15:117db924cf7c | 10183 | * i value to use for tweak |
wolfSSL | 15:117db924cf7c | 10184 | * iSz size of i buffer, should always be AES_BLOCK_SIZE but having this input |
wolfSSL | 15:117db924cf7c | 10185 | * adds a sanity check on how the user calls the function. |
wolfSSL | 15:117db924cf7c | 10186 | * |
wolfSSL | 15:117db924cf7c | 10187 | * returns 0 on success |
wolfSSL | 15:117db924cf7c | 10188 | */ |
wolfSSL | 15:117db924cf7c | 10189 | int wc_AesXtsEncrypt(XtsAes* xaes, byte* out, const byte* in, word32 sz, |
wolfSSL | 15:117db924cf7c | 10190 | const byte* i, word32 iSz) |
wolfSSL | 15:117db924cf7c | 10191 | { |
wolfSSL | 15:117db924cf7c | 10192 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 10193 | word32 blocks = (sz / AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10194 | Aes *aes, *tweak; |
wolfSSL | 15:117db924cf7c | 10195 | |
wolfSSL | 15:117db924cf7c | 10196 | if (xaes == NULL || out == NULL || in == NULL) { |
wolfSSL | 15:117db924cf7c | 10197 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 10198 | } |
wolfSSL | 15:117db924cf7c | 10199 | |
wolfSSL | 15:117db924cf7c | 10200 | aes = &xaes->aes; |
wolfSSL | 15:117db924cf7c | 10201 | tweak = &xaes->tweak; |
wolfSSL | 15:117db924cf7c | 10202 | |
wolfSSL | 15:117db924cf7c | 10203 | if (iSz < AES_BLOCK_SIZE) { |
wolfSSL | 15:117db924cf7c | 10204 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 10205 | } |
wolfSSL | 15:117db924cf7c | 10206 | |
wolfSSL | 15:117db924cf7c | 10207 | if (blocks > 0) { |
wolfSSL | 15:117db924cf7c | 10208 | byte tmp[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 10209 | |
wolfSSL | 15:117db924cf7c | 10210 | XMEMSET(tmp, 0, AES_BLOCK_SIZE); /* set to 0's in case of improper AES |
wolfSSL | 15:117db924cf7c | 10211 | * key setup passed to encrypt direct*/ |
wolfSSL | 15:117db924cf7c | 10212 | |
wolfSSL | 15:117db924cf7c | 10213 | wc_AesEncryptDirect(tweak, tmp, i); |
wolfSSL | 15:117db924cf7c | 10214 | |
wolfSSL | 15:117db924cf7c | 10215 | #ifdef HAVE_AES_ECB |
wolfSSL | 15:117db924cf7c | 10216 | /* encrypt all of buffer at once when possible */ |
wolfSSL | 15:117db924cf7c | 10217 | if (in != out) { /* can not handle inline */ |
wolfSSL | 15:117db924cf7c | 10218 | XMEMCPY(out, tmp, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10219 | if ((ret = _AesXtsHelper(aes, out, in, sz, AES_ENCRYPTION)) != 0) { |
wolfSSL | 15:117db924cf7c | 10220 | return ret; |
wolfSSL | 15:117db924cf7c | 10221 | } |
wolfSSL | 15:117db924cf7c | 10222 | } |
wolfSSL | 15:117db924cf7c | 10223 | #endif |
wolfSSL | 15:117db924cf7c | 10224 | |
wolfSSL | 15:117db924cf7c | 10225 | while (blocks > 0) { |
wolfSSL | 15:117db924cf7c | 10226 | word32 j; |
wolfSSL | 15:117db924cf7c | 10227 | byte carry = 0; |
wolfSSL | 15:117db924cf7c | 10228 | byte buf[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 10229 | |
wolfSSL | 15:117db924cf7c | 10230 | #ifdef HAVE_AES_ECB |
wolfSSL | 15:117db924cf7c | 10231 | if (in == out) { /* check for if inline */ |
wolfSSL | 15:117db924cf7c | 10232 | #endif |
wolfSSL | 15:117db924cf7c | 10233 | XMEMCPY(buf, in, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10234 | xorbuf(buf, tmp, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10235 | wc_AesEncryptDirect(aes, out, buf); |
wolfSSL | 15:117db924cf7c | 10236 | #ifdef HAVE_AES_ECB |
wolfSSL | 15:117db924cf7c | 10237 | } |
wolfSSL | 15:117db924cf7c | 10238 | #endif |
wolfSSL | 15:117db924cf7c | 10239 | xorbuf(out, tmp, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10240 | |
wolfSSL | 15:117db924cf7c | 10241 | /* multiply by shift left and propogate carry */ |
wolfSSL | 15:117db924cf7c | 10242 | for (j = 0; j < AES_BLOCK_SIZE; j++) { |
wolfSSL | 15:117db924cf7c | 10243 | byte tmpC; |
wolfSSL | 15:117db924cf7c | 10244 | |
wolfSSL | 15:117db924cf7c | 10245 | tmpC = (tmp[j] >> 7) & 0x01; |
wolfSSL | 15:117db924cf7c | 10246 | tmp[j] = ((tmp[j] << 1) + carry) & 0xFF; |
wolfSSL | 15:117db924cf7c | 10247 | carry = tmpC; |
wolfSSL | 15:117db924cf7c | 10248 | } |
wolfSSL | 15:117db924cf7c | 10249 | if (carry) { |
wolfSSL | 15:117db924cf7c | 10250 | tmp[0] ^= GF_XTS; |
wolfSSL | 15:117db924cf7c | 10251 | } |
wolfSSL | 15:117db924cf7c | 10252 | |
wolfSSL | 15:117db924cf7c | 10253 | in += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 10254 | out += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 10255 | sz -= AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 10256 | blocks--; |
wolfSSL | 15:117db924cf7c | 10257 | } |
wolfSSL | 15:117db924cf7c | 10258 | |
wolfSSL | 15:117db924cf7c | 10259 | /* stealing operation of XTS to handle left overs */ |
wolfSSL | 15:117db924cf7c | 10260 | if (sz > 0) { |
wolfSSL | 15:117db924cf7c | 10261 | byte buf[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 10262 | |
wolfSSL | 15:117db924cf7c | 10263 | XMEMCPY(buf, out - AES_BLOCK_SIZE, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10264 | if (sz >= AES_BLOCK_SIZE) { /* extra sanity check before copy */ |
wolfSSL | 15:117db924cf7c | 10265 | return BUFFER_E; |
wolfSSL | 15:117db924cf7c | 10266 | } |
wolfSSL | 15:117db924cf7c | 10267 | XMEMCPY(out, buf, sz); |
wolfSSL | 15:117db924cf7c | 10268 | XMEMCPY(buf, in, sz); |
wolfSSL | 15:117db924cf7c | 10269 | |
wolfSSL | 15:117db924cf7c | 10270 | xorbuf(buf, tmp, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10271 | wc_AesEncryptDirect(aes, out - AES_BLOCK_SIZE, buf); |
wolfSSL | 15:117db924cf7c | 10272 | xorbuf(out - AES_BLOCK_SIZE, tmp, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10273 | } |
wolfSSL | 15:117db924cf7c | 10274 | } |
wolfSSL | 15:117db924cf7c | 10275 | else { |
wolfSSL | 15:117db924cf7c | 10276 | WOLFSSL_MSG("Plain text input too small for encryption"); |
wolfSSL | 15:117db924cf7c | 10277 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 10278 | } |
wolfSSL | 15:117db924cf7c | 10279 | |
wolfSSL | 15:117db924cf7c | 10280 | return ret; |
wolfSSL | 15:117db924cf7c | 10281 | } |
wolfSSL | 15:117db924cf7c | 10282 | |
wolfSSL | 15:117db924cf7c | 10283 | |
wolfSSL | 15:117db924cf7c | 10284 | /* Same process as encryption but Aes key is AES_DECRYPTION type. |
wolfSSL | 15:117db924cf7c | 10285 | * |
wolfSSL | 15:117db924cf7c | 10286 | * xaes AES keys to use for block encrypt/decrypt |
wolfSSL | 15:117db924cf7c | 10287 | * out output buffer to hold plain text |
wolfSSL | 15:117db924cf7c | 10288 | * in input cipher text buffer to decrypt |
wolfSSL | 15:117db924cf7c | 10289 | * sz size of both out and in buffers |
wolfSSL | 15:117db924cf7c | 10290 | * i value to use for tweak |
wolfSSL | 15:117db924cf7c | 10291 | * iSz size of i buffer, should always be AES_BLOCK_SIZE but having this input |
wolfSSL | 15:117db924cf7c | 10292 | * adds a sanity check on how the user calls the function. |
wolfSSL | 15:117db924cf7c | 10293 | * |
wolfSSL | 15:117db924cf7c | 10294 | * returns 0 on success |
wolfSSL | 15:117db924cf7c | 10295 | */ |
wolfSSL | 15:117db924cf7c | 10296 | int wc_AesXtsDecrypt(XtsAes* xaes, byte* out, const byte* in, word32 sz, |
wolfSSL | 15:117db924cf7c | 10297 | const byte* i, word32 iSz) |
wolfSSL | 15:117db924cf7c | 10298 | { |
wolfSSL | 15:117db924cf7c | 10299 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 10300 | word32 blocks = (sz / AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10301 | Aes *aes, *tweak; |
wolfSSL | 15:117db924cf7c | 10302 | |
wolfSSL | 15:117db924cf7c | 10303 | if (xaes == NULL || out == NULL || in == NULL) { |
wolfSSL | 15:117db924cf7c | 10304 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 10305 | } |
wolfSSL | 15:117db924cf7c | 10306 | |
wolfSSL | 15:117db924cf7c | 10307 | aes = &xaes->aes; |
wolfSSL | 15:117db924cf7c | 10308 | tweak = &xaes->tweak; |
wolfSSL | 15:117db924cf7c | 10309 | |
wolfSSL | 15:117db924cf7c | 10310 | if (iSz < AES_BLOCK_SIZE) { |
wolfSSL | 15:117db924cf7c | 10311 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 10312 | } |
wolfSSL | 15:117db924cf7c | 10313 | |
wolfSSL | 15:117db924cf7c | 10314 | if (blocks > 0) { |
wolfSSL | 15:117db924cf7c | 10315 | word32 j; |
wolfSSL | 15:117db924cf7c | 10316 | byte carry = 0; |
wolfSSL | 15:117db924cf7c | 10317 | byte tmp[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 10318 | byte stl = (sz % AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10319 | |
wolfSSL | 15:117db924cf7c | 10320 | XMEMSET(tmp, 0, AES_BLOCK_SIZE); /* set to 0's in case of improper AES |
wolfSSL | 15:117db924cf7c | 10321 | * key setup passed to decrypt direct*/ |
wolfSSL | 15:117db924cf7c | 10322 | |
wolfSSL | 15:117db924cf7c | 10323 | wc_AesEncryptDirect(tweak, tmp, i); |
wolfSSL | 15:117db924cf7c | 10324 | |
wolfSSL | 15:117db924cf7c | 10325 | /* if Stealing then break out of loop one block early to handle special |
wolfSSL | 15:117db924cf7c | 10326 | * case */ |
wolfSSL | 15:117db924cf7c | 10327 | if (stl > 0) { |
wolfSSL | 15:117db924cf7c | 10328 | blocks--; |
wolfSSL | 15:117db924cf7c | 10329 | } |
wolfSSL | 15:117db924cf7c | 10330 | |
wolfSSL | 15:117db924cf7c | 10331 | #ifdef HAVE_AES_ECB |
wolfSSL | 15:117db924cf7c | 10332 | /* decrypt all of buffer at once when possible */ |
wolfSSL | 15:117db924cf7c | 10333 | if (in != out) { /* can not handle inline */ |
wolfSSL | 15:117db924cf7c | 10334 | XMEMCPY(out, tmp, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10335 | if ((ret = _AesXtsHelper(aes, out, in, sz, AES_DECRYPTION)) != 0) { |
wolfSSL | 15:117db924cf7c | 10336 | return ret; |
wolfSSL | 15:117db924cf7c | 10337 | } |
wolfSSL | 15:117db924cf7c | 10338 | } |
wolfSSL | 15:117db924cf7c | 10339 | #endif |
wolfSSL | 15:117db924cf7c | 10340 | |
wolfSSL | 15:117db924cf7c | 10341 | while (blocks > 0) { |
wolfSSL | 15:117db924cf7c | 10342 | byte buf[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 10343 | |
wolfSSL | 15:117db924cf7c | 10344 | #ifdef HAVE_AES_ECB |
wolfSSL | 15:117db924cf7c | 10345 | if (in == out) { /* check for if inline */ |
wolfSSL | 15:117db924cf7c | 10346 | #endif |
wolfSSL | 15:117db924cf7c | 10347 | XMEMCPY(buf, in, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10348 | xorbuf(buf, tmp, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10349 | wc_AesDecryptDirect(aes, out, buf); |
wolfSSL | 15:117db924cf7c | 10350 | #ifdef HAVE_AES_ECB |
wolfSSL | 15:117db924cf7c | 10351 | } |
wolfSSL | 15:117db924cf7c | 10352 | #endif |
wolfSSL | 15:117db924cf7c | 10353 | xorbuf(out, tmp, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10354 | |
wolfSSL | 15:117db924cf7c | 10355 | /* multiply by shift left and propogate carry */ |
wolfSSL | 15:117db924cf7c | 10356 | for (j = 0; j < AES_BLOCK_SIZE; j++) { |
wolfSSL | 15:117db924cf7c | 10357 | byte tmpC; |
wolfSSL | 15:117db924cf7c | 10358 | |
wolfSSL | 15:117db924cf7c | 10359 | tmpC = (tmp[j] >> 7) & 0x01; |
wolfSSL | 15:117db924cf7c | 10360 | tmp[j] = ((tmp[j] << 1) + carry) & 0xFF; |
wolfSSL | 15:117db924cf7c | 10361 | carry = tmpC; |
wolfSSL | 15:117db924cf7c | 10362 | } |
wolfSSL | 15:117db924cf7c | 10363 | if (carry) { |
wolfSSL | 15:117db924cf7c | 10364 | tmp[0] ^= GF_XTS; |
wolfSSL | 15:117db924cf7c | 10365 | } |
wolfSSL | 15:117db924cf7c | 10366 | carry = 0; |
wolfSSL | 15:117db924cf7c | 10367 | |
wolfSSL | 15:117db924cf7c | 10368 | in += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 10369 | out += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 10370 | sz -= AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 10371 | blocks--; |
wolfSSL | 15:117db924cf7c | 10372 | } |
wolfSSL | 15:117db924cf7c | 10373 | |
wolfSSL | 15:117db924cf7c | 10374 | /* stealing operation of XTS to handle left overs */ |
wolfSSL | 15:117db924cf7c | 10375 | if (sz > 0) { |
wolfSSL | 15:117db924cf7c | 10376 | byte buf[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 10377 | byte tmp2[AES_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 10378 | |
wolfSSL | 15:117db924cf7c | 10379 | /* multiply by shift left and propogate carry */ |
wolfSSL | 15:117db924cf7c | 10380 | for (j = 0; j < AES_BLOCK_SIZE; j++) { |
wolfSSL | 15:117db924cf7c | 10381 | byte tmpC; |
wolfSSL | 15:117db924cf7c | 10382 | |
wolfSSL | 15:117db924cf7c | 10383 | tmpC = (tmp[j] >> 7) & 0x01; |
wolfSSL | 15:117db924cf7c | 10384 | tmp2[j] = ((tmp[j] << 1) + carry) & 0xFF; |
wolfSSL | 15:117db924cf7c | 10385 | carry = tmpC; |
wolfSSL | 15:117db924cf7c | 10386 | } |
wolfSSL | 15:117db924cf7c | 10387 | if (carry) { |
wolfSSL | 15:117db924cf7c | 10388 | tmp2[0] ^= GF_XTS; |
wolfSSL | 15:117db924cf7c | 10389 | } |
wolfSSL | 15:117db924cf7c | 10390 | |
wolfSSL | 15:117db924cf7c | 10391 | XMEMCPY(buf, in, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10392 | xorbuf(buf, tmp2, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10393 | wc_AesDecryptDirect(aes, out, buf); |
wolfSSL | 15:117db924cf7c | 10394 | xorbuf(out, tmp2, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10395 | |
wolfSSL | 15:117db924cf7c | 10396 | /* tmp2 holds partial | last */ |
wolfSSL | 15:117db924cf7c | 10397 | XMEMCPY(tmp2, out, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10398 | in += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 10399 | out += AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 10400 | sz -= AES_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 10401 | |
wolfSSL | 15:117db924cf7c | 10402 | /* Make buffer with end of cipher text | last */ |
wolfSSL | 15:117db924cf7c | 10403 | XMEMCPY(buf, tmp2, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10404 | if (sz >= AES_BLOCK_SIZE) { /* extra sanity check before copy */ |
wolfSSL | 15:117db924cf7c | 10405 | return BUFFER_E; |
wolfSSL | 15:117db924cf7c | 10406 | } |
wolfSSL | 15:117db924cf7c | 10407 | XMEMCPY(buf, in, sz); |
wolfSSL | 15:117db924cf7c | 10408 | XMEMCPY(out, tmp2, sz); |
wolfSSL | 15:117db924cf7c | 10409 | |
wolfSSL | 15:117db924cf7c | 10410 | xorbuf(buf, tmp, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10411 | wc_AesDecryptDirect(aes, tmp2, buf); |
wolfSSL | 15:117db924cf7c | 10412 | xorbuf(tmp2, tmp, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10413 | XMEMCPY(out - AES_BLOCK_SIZE, tmp2, AES_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 10414 | } |
wolfSSL | 15:117db924cf7c | 10415 | } |
wolfSSL | 15:117db924cf7c | 10416 | else { |
wolfSSL | 15:117db924cf7c | 10417 | WOLFSSL_MSG("Plain text input too small for encryption"); |
wolfSSL | 15:117db924cf7c | 10418 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 10419 | } |
wolfSSL | 15:117db924cf7c | 10420 | |
wolfSSL | 15:117db924cf7c | 10421 | return ret; |
wolfSSL | 15:117db924cf7c | 10422 | } |
wolfSSL | 15:117db924cf7c | 10423 | |
wolfSSL | 15:117db924cf7c | 10424 | #endif /* WOLFSSL_AES_XTS */ |
wolfSSL | 15:117db924cf7c | 10425 | |
wolfSSL | 15:117db924cf7c | 10426 | #endif /* HAVE_FIPS */ |
wolfSSL | 15:117db924cf7c | 10427 | #endif /* !NO_AES */ |
wolfSSL | 15:117db924cf7c | 10428 |