1 /*- 2 * Copyright (c) 2014-2021 The FreeBSD Foundation 3 * Copyright (c) 2018 iXsystems, Inc 4 * All rights reserved. 5 * 6 * Portions of this software were developed by John-Mark Gurney 7 * under the sponsorship of the FreeBSD Foundation and 8 * Rubicon Communications, LLC (Netgate). 9 * 10 * Portions of this software were developed by Ararat River 11 * Consulting, LLC under sponsorship of the FreeBSD Foundation. 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * 35 * This file implements AES-CCM+CBC-MAC, as described 36 * at https://tools.ietf.org/html/rfc3610, using Intel's 37 * AES-NI instructions. 38 * 39 */ 40 41 #include <sys/types.h> 42 #include <sys/endian.h> 43 #include <sys/param.h> 44 45 #include <sys/systm.h> 46 #include <crypto/aesni/aesni.h> 47 #include <crypto/aesni/aesni_os.h> 48 #include <crypto/aesni/aesencdec.h> 49 #define AESNI_ENC(d, k, nr) aesni_enc(nr-1, (const __m128i*)k, d) 50 51 #include <wmmintrin.h> 52 #include <emmintrin.h> 53 #include <smmintrin.h> 54 55 /* 56 * Encrypt a single 128-bit block after 57 * doing an xor. This is also used to 58 * decrypt (yay symmetric encryption). 59 */ 60 static inline __m128i 61 xor_and_encrypt(__m128i a, __m128i b, const unsigned char *k, int nr) 62 { 63 __m128i retval = _mm_xor_si128(a, b); 64 65 retval = AESNI_ENC(retval, k, nr); 66 return (retval); 67 } 68 69 /* 70 * Put value at the end of block, starting at offset. 71 * (This goes backwards, putting bytes in *until* it 72 * reaches offset.) 73 */ 74 static void 75 append_int(size_t value, __m128i *block, size_t offset) 76 { 77 int indx = sizeof(*block) - 1; 78 uint8_t *bp = (uint8_t*)block; 79 80 while (indx > (sizeof(*block) - offset)) { 81 bp[indx] = value & 0xff; 82 indx--; 83 value >>= 8; 84 } 85 } 86 87 /* 88 * Start the CBC-MAC process. This handles the auth data. 89 */ 90 static __m128i 91 cbc_mac_start(const unsigned char *auth_data, size_t auth_len, 92 const unsigned char *nonce, size_t nonce_len, 93 const unsigned char *key, int nr, 94 size_t data_len, size_t tag_len) 95 { 96 __m128i cbc_block, staging_block; 97 uint8_t *byte_ptr; 98 /* This defines where the message length goes */ 99 int L = sizeof(__m128i) - 1 - nonce_len; 100 101 /* 102 * Set up B0 here. This has the flags byte, 103 * followed by the nonce, followed by the 104 * length of the message. 105 */ 106 cbc_block = _mm_setzero_si128(); 107 byte_ptr = (uint8_t*)&cbc_block; 108 byte_ptr[0] = ((auth_len > 0) ? 1 : 0) * 64 | 109 (((tag_len - 2) / 2) * 8) | 110 (L - 1); 111 bcopy(nonce, byte_ptr + 1, nonce_len); 112 append_int(data_len, &cbc_block, L+1); 113 cbc_block = AESNI_ENC(cbc_block, key, nr); 114 115 if (auth_len != 0) { 116 /* 117 * We need to start by appending the length descriptor. 118 */ 119 uint32_t auth_amt; 120 size_t copy_amt; 121 const uint8_t *auth_ptr = auth_data; 122 123 staging_block = _mm_setzero_si128(); 124 125 /* 126 * The current OCF calling convention means that 127 * there can never be more than 4g of authentication 128 * data, so we don't handle the 0xffff case. 129 */ 130 KASSERT(auth_len < (1ULL << 32), 131 ("%s: auth_len (%zu) larger than 4GB", 132 __FUNCTION__, auth_len)); 133 134 if (auth_len < ((1 << 16) - (1 << 8))) { 135 /* 136 * If the auth data length is less than 137 * 0xff00, we don't need to encode a length 138 * specifier, just the length of the auth 139 * data. 140 */ 141 be16enc(&staging_block, auth_len); 142 auth_amt = 2; 143 } else if (auth_len < (1ULL << 32)) { 144 /* 145 * Two bytes for the length prefix, and then 146 * four bytes for the length. This makes a total 147 * of 6 bytes to describe the auth data length. 148 */ 149 be16enc(&staging_block, 0xfffe); 150 be32enc((char*)&staging_block + 2, auth_len); 151 auth_amt = 6; 152 } else 153 panic("%s: auth len too large", __FUNCTION__); 154 155 /* 156 * Need to copy abytes into blocks. The first block is 157 * already partially filled, by auth_amt, so we need 158 * to handle that. The last block needs to be zero padded. 159 */ 160 copy_amt = MIN(auth_len, 161 sizeof(staging_block) - auth_amt); 162 byte_ptr = (uint8_t*)&staging_block; 163 bcopy(auth_ptr, &byte_ptr[auth_amt], copy_amt); 164 auth_ptr += copy_amt; 165 166 cbc_block = xor_and_encrypt(cbc_block, staging_block, key, nr); 167 168 while (auth_ptr < auth_data + auth_len) { 169 copy_amt = MIN((auth_data + auth_len) - auth_ptr, 170 sizeof(staging_block)); 171 if (copy_amt < sizeof(staging_block)) 172 bzero(&staging_block, sizeof(staging_block)); 173 bcopy(auth_ptr, &staging_block, copy_amt); 174 cbc_block = xor_and_encrypt(cbc_block, staging_block, 175 key, nr); 176 auth_ptr += copy_amt; 177 } 178 } 179 return (cbc_block); 180 } 181 182 /* 183 * Implement AES CCM+CBC-MAC encryption and authentication. 184 * 185 * A couple of notes: 186 * Since abytes is limited to a 32 bit value here, the AAD is 187 * limited to 4 gigabytes or less. 188 */ 189 void 190 AES_CCM_encrypt(const unsigned char *in, unsigned char *out, 191 const unsigned char *addt, const unsigned char *nonce, 192 unsigned char *tag, uint32_t nbytes, uint32_t abytes, int nlen, 193 int tag_length, const unsigned char *key, int nr) 194 { 195 int L; 196 int counter = 1; /* S0 has 0, S1 has 1 */ 197 size_t copy_amt, total = 0; 198 uint8_t *byte_ptr; 199 __m128i s0, rolling_mac, s_x, staging_block; 200 201 /* NIST 800-38c section A.1 says n is [7, 13]. */ 202 if (nlen < 7 || nlen > 13) 203 panic("%s: bad nonce length %d", __FUNCTION__, nlen); 204 205 /* 206 * We need to know how many bytes to use to describe 207 * the length of the data. Normally, nlen should be 208 * 12, which leaves us 3 bytes to do that -- 16mbytes of 209 * data to encrypt. But it can be longer or shorter; 210 * this impacts the length of the message. 211 */ 212 L = sizeof(__m128i) - 1 - nlen; 213 214 /* 215 * Clear out the blocks 216 */ 217 s0 = _mm_setzero_si128(); 218 219 rolling_mac = cbc_mac_start(addt, abytes, nonce, nlen, 220 key, nr, nbytes, tag_length); 221 222 /* s0 has flags, nonce, and then 0 */ 223 byte_ptr = (uint8_t*)&s0; 224 byte_ptr[0] = L - 1; /* but the flags byte only has L' */ 225 bcopy(nonce, &byte_ptr[1], nlen); 226 227 /* 228 * Now to cycle through the rest of the data. 229 */ 230 bcopy(&s0, &s_x, sizeof(s0)); 231 232 while (total < nbytes) { 233 /* 234 * Copy the plain-text data into staging_block. 235 * This may need to be zero-padded. 236 */ 237 copy_amt = MIN(nbytes - total, sizeof(staging_block)); 238 bcopy(in+total, &staging_block, copy_amt); 239 if (copy_amt < sizeof(staging_block)) { 240 byte_ptr = (uint8_t*)&staging_block; 241 bzero(&byte_ptr[copy_amt], 242 sizeof(staging_block) - copy_amt); 243 } 244 rolling_mac = xor_and_encrypt(rolling_mac, staging_block, 245 key, nr); 246 /* Put the counter into the s_x block */ 247 append_int(counter++, &s_x, L+1); 248 /* Encrypt that */ 249 __m128i X = AESNI_ENC(s_x, key, nr); 250 /* XOR the plain-text with the encrypted counter block */ 251 staging_block = _mm_xor_si128(staging_block, X); 252 /* And copy it out */ 253 bcopy(&staging_block, out+total, copy_amt); 254 total += copy_amt; 255 } 256 /* 257 * Allegedly done with it! Except for the tag. 258 */ 259 s0 = AESNI_ENC(s0, key, nr); 260 staging_block = _mm_xor_si128(s0, rolling_mac); 261 bcopy(&staging_block, tag, tag_length); 262 explicit_bzero(&s0, sizeof(s0)); 263 explicit_bzero(&staging_block, sizeof(staging_block)); 264 explicit_bzero(&s_x, sizeof(s_x)); 265 explicit_bzero(&rolling_mac, sizeof(rolling_mac)); 266 } 267 268 /* 269 * Implement AES CCM+CBC-MAC decryption and authentication. 270 * Returns 0 on failure, 1 on success. 271 * 272 * The primary difference here is that each encrypted block 273 * needs to be hashed&encrypted after it is decrypted (since 274 * the CBC-MAC is based on the plain text). This means that 275 * we do the decryption twice -- first to verify the tag, 276 * and second to decrypt and copy it out. 277 * 278 * To avoid annoying code copying, we implement the main 279 * loop as a separate function. 280 * 281 * Call with out as NULL to not store the decrypted results; 282 * call with hashp as NULL to not run the authentication. 283 * Calling with neither as NULL does the decryption and 284 * authentication as a single pass (which is not allowed 285 * per the specification, really). 286 * 287 * If hashp is non-NULL, it points to the post-AAD computed 288 * checksum. 289 */ 290 static void 291 decrypt_loop(const unsigned char *in, unsigned char *out, size_t nbytes, 292 __m128i s0, size_t nonce_length, __m128i *macp, 293 const unsigned char *key, int nr) 294 { 295 size_t total = 0; 296 __m128i s_x = s0, mac_block; 297 int counter = 1; 298 const size_t L = sizeof(__m128i) - 1 - nonce_length; 299 __m128i pad_block, staging_block; 300 301 /* 302 * The starting mac (post AAD, if any). 303 */ 304 if (macp != NULL) 305 mac_block = *macp; 306 307 while (total < nbytes) { 308 size_t copy_amt = MIN(nbytes - total, sizeof(staging_block)); 309 310 if (copy_amt < sizeof(staging_block)) { 311 staging_block = _mm_setzero_si128(); 312 } 313 bcopy(in+total, &staging_block, copy_amt); 314 315 /* 316 * staging_block has the current block of input data, 317 * zero-padded if necessary. This is used in computing 318 * both the decrypted data, and the authentication tag. 319 */ 320 append_int(counter++, &s_x, L+1); 321 /* 322 * The tag is computed based on the decrypted data. 323 */ 324 pad_block = AESNI_ENC(s_x, key, nr); 325 if (copy_amt < sizeof(staging_block)) { 326 /* 327 * Need to pad out pad_block with 0. 328 * (staging_block was set to 0's above.) 329 */ 330 uint8_t *end_of_buffer = (uint8_t*)&pad_block; 331 bzero(end_of_buffer + copy_amt, 332 sizeof(pad_block) - copy_amt); 333 } 334 staging_block = _mm_xor_si128(staging_block, pad_block); 335 336 if (out) 337 bcopy(&staging_block, out+total, copy_amt); 338 339 if (macp) 340 mac_block = xor_and_encrypt(mac_block, staging_block, 341 key, nr); 342 total += copy_amt; 343 } 344 345 if (macp) 346 *macp = mac_block; 347 348 explicit_bzero(&pad_block, sizeof(pad_block)); 349 explicit_bzero(&staging_block, sizeof(staging_block)); 350 explicit_bzero(&mac_block, sizeof(mac_block)); 351 } 352 353 /* 354 * The exposed decryption routine. This is practically a 355 * copy of the encryption routine, except that the order 356 * in which the tag is created is changed. 357 * XXX combine the two functions at some point! 358 */ 359 int 360 AES_CCM_decrypt(const unsigned char *in, unsigned char *out, 361 const unsigned char *addt, const unsigned char *nonce, 362 const unsigned char *tag, uint32_t nbytes, uint32_t abytes, int nlen, 363 int tag_length, const unsigned char *key, int nr) 364 { 365 int L; 366 __m128i s0, rolling_mac, staging_block; 367 uint8_t *byte_ptr; 368 369 if (nlen < 0 || nlen > 15) 370 panic("%s: bad nonce length %d", __FUNCTION__, nlen); 371 372 /* 373 * We need to know how many bytes to use to describe 374 * the length of the data. Normally, nlen should be 375 * 12, which leaves us 3 bytes to do that -- 16mbytes of 376 * data to encrypt. But it can be longer or shorter. 377 */ 378 L = sizeof(__m128i) - 1 - nlen; 379 380 /* 381 * Clear out the blocks 382 */ 383 s0 = _mm_setzero_si128(); 384 385 rolling_mac = cbc_mac_start(addt, abytes, nonce, nlen, 386 key, nr, nbytes, tag_length); 387 /* s0 has flags, nonce, and then 0 */ 388 byte_ptr = (uint8_t*)&s0; 389 byte_ptr[0] = L-1; /* but the flags byte only has L' */ 390 bcopy(nonce, &byte_ptr[1], nlen); 391 392 /* 393 * Now to cycle through the rest of the data. 394 */ 395 decrypt_loop(in, NULL, nbytes, s0, nlen, &rolling_mac, key, nr); 396 397 /* 398 * Compare the tag. 399 */ 400 staging_block = _mm_xor_si128(AESNI_ENC(s0, key, nr), rolling_mac); 401 if (timingsafe_bcmp(&staging_block, tag, tag_length) != 0) { 402 return (0); 403 } 404 405 /* 406 * Push out the decryption results this time. 407 */ 408 decrypt_loop(in, out, nbytes, s0, nlen, NULL, key, nr); 409 return (1); 410 } 411