1 /* 2 * AES-NI support functions 3 * 4 * Copyright The Mbed TLS Contributors 5 * SPDX-License-Identifier: Apache-2.0 OR GPL-2.0-or-later 6 */ 7 8 /* 9 * [AES-WP] https://www.intel.com/content/www/us/en/developer/articles/tool/intel-advanced-encryption-standard-aes-instructions-set.html 10 * [CLMUL-WP] https://www.intel.com/content/www/us/en/develop/download/intel-carry-less-multiplication-instruction-and-its-usage-for-computing-the-gcm-mode.html 11 */ 12 13 #include "common.h" 14 15 #if defined(MBEDTLS_AESNI_C) 16 17 #include "aesni.h" 18 19 #include <string.h> 20 21 #if defined(MBEDTLS_AESNI_HAVE_CODE) 22 23 #if MBEDTLS_AESNI_HAVE_CODE == 2 24 #if defined(__GNUC__) 25 #include <cpuid.h> 26 #elif defined(_MSC_VER) 27 #include <intrin.h> 28 #else 29 #error "`__cpuid` required by MBEDTLS_AESNI_C is not supported by the compiler" 30 #endif 31 #include <immintrin.h> 32 #endif 33 34 #if defined(MBEDTLS_ARCH_IS_X86) 35 #if defined(MBEDTLS_COMPILER_IS_GCC) 36 #pragma GCC push_options 37 #pragma GCC target ("pclmul,sse2,aes") 38 #define MBEDTLS_POP_TARGET_PRAGMA 39 #elif defined(__clang__) && (__clang_major__ >= 5) 40 #pragma clang attribute push (__attribute__((target("pclmul,sse2,aes"))), apply_to=function) 41 #define MBEDTLS_POP_TARGET_PRAGMA 42 #endif 43 #endif 44 45 #if !defined(MBEDTLS_AES_USE_HARDWARE_ONLY) 46 /* 47 * AES-NI support detection routine 48 */ 49 int mbedtls_aesni_has_support(unsigned int what) 50 { 51 /* To avoid a race condition, tell the compiler that the assignment 52 * `done = 1` and the assignment to `c` may not be reordered. 53 * https://github.com/Mbed-TLS/mbedtls/issues/9840 54 * 55 * Note that we may also be worried about memory access reordering, 56 * but fortunately the x86 memory model is not too wild: stores 57 * from the same thread are observed consistently by other threads. 58 * (See example 8-1 in Sewell et al., "x86-TSO: A Rigorous and Usable 59 * Programmer’s Model for x86 Multiprocessors", CACM, 2010, 60 * https://www.cl.cam.ac.uk/~pes20/weakmemory/cacm.pdf) 61 */ 62 static volatile int done = 0; 63 static volatile unsigned int c = 0; 64 65 if (!done) { 66 #if MBEDTLS_AESNI_HAVE_CODE == 2 67 static int info[4] = { 0, 0, 0, 0 }; 68 #if defined(_MSC_VER) 69 __cpuid(info, 1); 70 #else 71 __cpuid(1, info[0], info[1], info[2], info[3]); 72 #endif 73 c = info[2]; 74 #else /* AESNI using asm */ 75 asm ("movl $1, %%eax \n\t" 76 "cpuid \n\t" 77 : "=c" (c) 78 : 79 : "eax", "ebx", "edx"); 80 #endif /* MBEDTLS_AESNI_HAVE_CODE */ 81 done = 1; 82 } 83 84 return (c & what) != 0; 85 } 86 #endif /* !MBEDTLS_AES_USE_HARDWARE_ONLY */ 87 88 #if MBEDTLS_AESNI_HAVE_CODE == 2 89 90 /* 91 * AES-NI AES-ECB block en(de)cryption 92 */ 93 int mbedtls_aesni_crypt_ecb(mbedtls_aes_context *ctx, 94 int mode, 95 const unsigned char input[16], 96 unsigned char output[16]) 97 { 98 const __m128i *rk = (const __m128i *) (ctx->buf + ctx->rk_offset); 99 unsigned nr = ctx->nr; // Number of remaining rounds 100 101 // Load round key 0 102 __m128i state; 103 memcpy(&state, input, 16); 104 state = _mm_xor_si128(state, rk[0]); // state ^= *rk; 105 ++rk; 106 --nr; 107 108 #if !defined(MBEDTLS_BLOCK_CIPHER_NO_DECRYPT) 109 if (mode == MBEDTLS_AES_DECRYPT) { 110 while (nr != 0) { 111 state = _mm_aesdec_si128(state, *rk); 112 ++rk; 113 --nr; 114 } 115 state = _mm_aesdeclast_si128(state, *rk); 116 } else 117 #else 118 (void) mode; 119 #endif 120 { 121 while (nr != 0) { 122 state = _mm_aesenc_si128(state, *rk); 123 ++rk; 124 --nr; 125 } 126 state = _mm_aesenclast_si128(state, *rk); 127 } 128 129 memcpy(output, &state, 16); 130 return 0; 131 } 132 133 /* 134 * GCM multiplication: c = a times b in GF(2^128) 135 * Based on [CLMUL-WP] algorithms 1 (with equation 27) and 5. 136 */ 137 138 static void gcm_clmul(const __m128i aa, const __m128i bb, 139 __m128i *cc, __m128i *dd) 140 { 141 /* 142 * Caryless multiplication dd:cc = aa * bb 143 * using [CLMUL-WP] algorithm 1 (p. 12). 144 */ 145 *cc = _mm_clmulepi64_si128(aa, bb, 0x00); // a0*b0 = c1:c0 146 *dd = _mm_clmulepi64_si128(aa, bb, 0x11); // a1*b1 = d1:d0 147 __m128i ee = _mm_clmulepi64_si128(aa, bb, 0x10); // a0*b1 = e1:e0 148 __m128i ff = _mm_clmulepi64_si128(aa, bb, 0x01); // a1*b0 = f1:f0 149 ff = _mm_xor_si128(ff, ee); // e1+f1:e0+f0 150 ee = ff; // e1+f1:e0+f0 151 ff = _mm_srli_si128(ff, 8); // 0:e1+f1 152 ee = _mm_slli_si128(ee, 8); // e0+f0:0 153 *dd = _mm_xor_si128(*dd, ff); // d1:d0+e1+f1 154 *cc = _mm_xor_si128(*cc, ee); // c1+e0+f0:c0 155 } 156 157 static void gcm_shift(__m128i *cc, __m128i *dd) 158 { 159 /* [CMUCL-WP] Algorithm 5 Step 1: shift cc:dd one bit to the left, 160 * taking advantage of [CLMUL-WP] eq 27 (p. 18). */ 161 // // *cc = r1:r0 162 // // *dd = r3:r2 163 __m128i cc_lo = _mm_slli_epi64(*cc, 1); // r1<<1:r0<<1 164 __m128i dd_lo = _mm_slli_epi64(*dd, 1); // r3<<1:r2<<1 165 __m128i cc_hi = _mm_srli_epi64(*cc, 63); // r1>>63:r0>>63 166 __m128i dd_hi = _mm_srli_epi64(*dd, 63); // r3>>63:r2>>63 167 __m128i xmm5 = _mm_srli_si128(cc_hi, 8); // 0:r1>>63 168 cc_hi = _mm_slli_si128(cc_hi, 8); // r0>>63:0 169 dd_hi = _mm_slli_si128(dd_hi, 8); // 0:r1>>63 170 171 *cc = _mm_or_si128(cc_lo, cc_hi); // r1<<1|r0>>63:r0<<1 172 *dd = _mm_or_si128(_mm_or_si128(dd_lo, dd_hi), xmm5); // r3<<1|r2>>62:r2<<1|r1>>63 173 } 174 175 static __m128i gcm_reduce(__m128i xx) 176 { 177 // // xx = x1:x0 178 /* [CLMUL-WP] Algorithm 5 Step 2 */ 179 __m128i aa = _mm_slli_epi64(xx, 63); // x1<<63:x0<<63 = stuff:a 180 __m128i bb = _mm_slli_epi64(xx, 62); // x1<<62:x0<<62 = stuff:b 181 __m128i cc = _mm_slli_epi64(xx, 57); // x1<<57:x0<<57 = stuff:c 182 __m128i dd = _mm_slli_si128(_mm_xor_si128(_mm_xor_si128(aa, bb), cc), 8); // a+b+c:0 183 return _mm_xor_si128(dd, xx); // x1+a+b+c:x0 = d:x0 184 } 185 186 static __m128i gcm_mix(__m128i dx) 187 { 188 /* [CLMUL-WP] Algorithm 5 Steps 3 and 4 */ 189 __m128i ee = _mm_srli_epi64(dx, 1); // e1:x0>>1 = e1:e0' 190 __m128i ff = _mm_srli_epi64(dx, 2); // f1:x0>>2 = f1:f0' 191 __m128i gg = _mm_srli_epi64(dx, 7); // g1:x0>>7 = g1:g0' 192 193 // e0'+f0'+g0' is almost e0+f0+g0, except for some missing 194 // bits carried from d. Now get those bits back in. 195 __m128i eh = _mm_slli_epi64(dx, 63); // d<<63:stuff 196 __m128i fh = _mm_slli_epi64(dx, 62); // d<<62:stuff 197 __m128i gh = _mm_slli_epi64(dx, 57); // d<<57:stuff 198 __m128i hh = _mm_srli_si128(_mm_xor_si128(_mm_xor_si128(eh, fh), gh), 8); // 0:missing bits of d 199 200 return _mm_xor_si128(_mm_xor_si128(_mm_xor_si128(_mm_xor_si128(ee, ff), gg), hh), dx); 201 } 202 203 void mbedtls_aesni_gcm_mult(unsigned char c[16], 204 const unsigned char a[16], 205 const unsigned char b[16]) 206 { 207 __m128i aa = { 0 }, bb = { 0 }, cc, dd; 208 209 /* The inputs are in big-endian order, so byte-reverse them */ 210 for (size_t i = 0; i < 16; i++) { 211 ((uint8_t *) &aa)[i] = a[15 - i]; 212 ((uint8_t *) &bb)[i] = b[15 - i]; 213 } 214 215 gcm_clmul(aa, bb, &cc, &dd); 216 gcm_shift(&cc, &dd); 217 /* 218 * Now reduce modulo the GCM polynomial x^128 + x^7 + x^2 + x + 1 219 * using [CLMUL-WP] algorithm 5 (p. 18). 220 * Currently dd:cc holds x3:x2:x1:x0 (already shifted). 221 */ 222 __m128i dx = gcm_reduce(cc); 223 __m128i xh = gcm_mix(dx); 224 cc = _mm_xor_si128(xh, dd); // x3+h1:x2+h0 225 226 /* Now byte-reverse the outputs */ 227 for (size_t i = 0; i < 16; i++) { 228 c[i] = ((uint8_t *) &cc)[15 - i]; 229 } 230 231 return; 232 } 233 234 /* 235 * Compute decryption round keys from encryption round keys 236 */ 237 #if !defined(MBEDTLS_BLOCK_CIPHER_NO_DECRYPT) 238 void mbedtls_aesni_inverse_key(unsigned char *invkey, 239 const unsigned char *fwdkey, int nr) 240 { 241 __m128i *ik = (__m128i *) invkey; 242 const __m128i *fk = (const __m128i *) fwdkey + nr; 243 244 *ik = *fk; 245 for (--fk, ++ik; fk > (const __m128i *) fwdkey; --fk, ++ik) { 246 *ik = _mm_aesimc_si128(*fk); 247 } 248 *ik = *fk; 249 } 250 #endif 251 252 /* 253 * Key expansion, 128-bit case 254 */ 255 static __m128i aesni_set_rk_128(__m128i state, __m128i xword) 256 { 257 /* 258 * Finish generating the next round key. 259 * 260 * On entry state is r3:r2:r1:r0 and xword is X:stuff:stuff:stuff 261 * with X = rot( sub( r3 ) ) ^ RCON (obtained with AESKEYGENASSIST). 262 * 263 * On exit, xword is r7:r6:r5:r4 264 * with r4 = X + r0, r5 = r4 + r1, r6 = r5 + r2, r7 = r6 + r3 265 * and this is returned, to be written to the round key buffer. 266 */ 267 xword = _mm_shuffle_epi32(xword, 0xff); // X:X:X:X 268 xword = _mm_xor_si128(xword, state); // X+r3:X+r2:X+r1:r4 269 state = _mm_slli_si128(state, 4); // r2:r1:r0:0 270 xword = _mm_xor_si128(xword, state); // X+r3+r2:X+r2+r1:r5:r4 271 state = _mm_slli_si128(state, 4); // r1:r0:0:0 272 xword = _mm_xor_si128(xword, state); // X+r3+r2+r1:r6:r5:r4 273 state = _mm_slli_si128(state, 4); // r0:0:0:0 274 state = _mm_xor_si128(xword, state); // r7:r6:r5:r4 275 return state; 276 } 277 278 static void aesni_setkey_enc_128(unsigned char *rk_bytes, 279 const unsigned char *key) 280 { 281 __m128i *rk = (__m128i *) rk_bytes; 282 283 memcpy(&rk[0], key, 16); 284 rk[1] = aesni_set_rk_128(rk[0], _mm_aeskeygenassist_si128(rk[0], 0x01)); 285 rk[2] = aesni_set_rk_128(rk[1], _mm_aeskeygenassist_si128(rk[1], 0x02)); 286 rk[3] = aesni_set_rk_128(rk[2], _mm_aeskeygenassist_si128(rk[2], 0x04)); 287 rk[4] = aesni_set_rk_128(rk[3], _mm_aeskeygenassist_si128(rk[3], 0x08)); 288 rk[5] = aesni_set_rk_128(rk[4], _mm_aeskeygenassist_si128(rk[4], 0x10)); 289 rk[6] = aesni_set_rk_128(rk[5], _mm_aeskeygenassist_si128(rk[5], 0x20)); 290 rk[7] = aesni_set_rk_128(rk[6], _mm_aeskeygenassist_si128(rk[6], 0x40)); 291 rk[8] = aesni_set_rk_128(rk[7], _mm_aeskeygenassist_si128(rk[7], 0x80)); 292 rk[9] = aesni_set_rk_128(rk[8], _mm_aeskeygenassist_si128(rk[8], 0x1B)); 293 rk[10] = aesni_set_rk_128(rk[9], _mm_aeskeygenassist_si128(rk[9], 0x36)); 294 } 295 296 /* 297 * Key expansion, 192-bit case 298 */ 299 #if !defined(MBEDTLS_AES_ONLY_128_BIT_KEY_LENGTH) 300 static void aesni_set_rk_192(__m128i *state0, __m128i *state1, __m128i xword, 301 unsigned char *rk) 302 { 303 /* 304 * Finish generating the next 6 quarter-keys. 305 * 306 * On entry state0 is r3:r2:r1:r0, state1 is stuff:stuff:r5:r4 307 * and xword is stuff:stuff:X:stuff with X = rot( sub( r3 ) ) ^ RCON 308 * (obtained with AESKEYGENASSIST). 309 * 310 * On exit, state0 is r9:r8:r7:r6 and state1 is stuff:stuff:r11:r10 311 * and those are written to the round key buffer. 312 */ 313 xword = _mm_shuffle_epi32(xword, 0x55); // X:X:X:X 314 xword = _mm_xor_si128(xword, *state0); // X+r3:X+r2:X+r1:X+r0 315 *state0 = _mm_slli_si128(*state0, 4); // r2:r1:r0:0 316 xword = _mm_xor_si128(xword, *state0); // X+r3+r2:X+r2+r1:X+r1+r0:X+r0 317 *state0 = _mm_slli_si128(*state0, 4); // r1:r0:0:0 318 xword = _mm_xor_si128(xword, *state0); // X+r3+r2+r1:X+r2+r1+r0:X+r1+r0:X+r0 319 *state0 = _mm_slli_si128(*state0, 4); // r0:0:0:0 320 xword = _mm_xor_si128(xword, *state0); // X+r3+r2+r1+r0:X+r2+r1+r0:X+r1+r0:X+r0 321 *state0 = xword; // = r9:r8:r7:r6 322 323 xword = _mm_shuffle_epi32(xword, 0xff); // r9:r9:r9:r9 324 xword = _mm_xor_si128(xword, *state1); // stuff:stuff:r9+r5:r9+r4 325 *state1 = _mm_slli_si128(*state1, 4); // stuff:stuff:r4:0 326 xword = _mm_xor_si128(xword, *state1); // stuff:stuff:r9+r5+r4:r9+r4 327 *state1 = xword; // = stuff:stuff:r11:r10 328 329 /* Store state0 and the low half of state1 into rk, which is conceptually 330 * an array of 24-byte elements. Since 24 is not a multiple of 16, 331 * rk is not necessarily aligned so just `*rk = *state0` doesn't work. */ 332 memcpy(rk, state0, 16); 333 memcpy(rk + 16, state1, 8); 334 } 335 336 static void aesni_setkey_enc_192(unsigned char *rk, 337 const unsigned char *key) 338 { 339 /* First round: use original key */ 340 memcpy(rk, key, 24); 341 /* aes.c guarantees that rk is aligned on a 16-byte boundary. */ 342 __m128i state0 = ((__m128i *) rk)[0]; 343 __m128i state1 = _mm_loadl_epi64(((__m128i *) rk) + 1); 344 345 aesni_set_rk_192(&state0, &state1, _mm_aeskeygenassist_si128(state1, 0x01), rk + 24 * 1); 346 aesni_set_rk_192(&state0, &state1, _mm_aeskeygenassist_si128(state1, 0x02), rk + 24 * 2); 347 aesni_set_rk_192(&state0, &state1, _mm_aeskeygenassist_si128(state1, 0x04), rk + 24 * 3); 348 aesni_set_rk_192(&state0, &state1, _mm_aeskeygenassist_si128(state1, 0x08), rk + 24 * 4); 349 aesni_set_rk_192(&state0, &state1, _mm_aeskeygenassist_si128(state1, 0x10), rk + 24 * 5); 350 aesni_set_rk_192(&state0, &state1, _mm_aeskeygenassist_si128(state1, 0x20), rk + 24 * 6); 351 aesni_set_rk_192(&state0, &state1, _mm_aeskeygenassist_si128(state1, 0x40), rk + 24 * 7); 352 aesni_set_rk_192(&state0, &state1, _mm_aeskeygenassist_si128(state1, 0x80), rk + 24 * 8); 353 } 354 #endif /* !MBEDTLS_AES_ONLY_128_BIT_KEY_LENGTH */ 355 356 /* 357 * Key expansion, 256-bit case 358 */ 359 #if !defined(MBEDTLS_AES_ONLY_128_BIT_KEY_LENGTH) 360 static void aesni_set_rk_256(__m128i state0, __m128i state1, __m128i xword, 361 __m128i *rk0, __m128i *rk1) 362 { 363 /* 364 * Finish generating the next two round keys. 365 * 366 * On entry state0 is r3:r2:r1:r0, state1 is r7:r6:r5:r4 and 367 * xword is X:stuff:stuff:stuff with X = rot( sub( r7 )) ^ RCON 368 * (obtained with AESKEYGENASSIST). 369 * 370 * On exit, *rk0 is r11:r10:r9:r8 and *rk1 is r15:r14:r13:r12 371 */ 372 xword = _mm_shuffle_epi32(xword, 0xff); 373 xword = _mm_xor_si128(xword, state0); 374 state0 = _mm_slli_si128(state0, 4); 375 xword = _mm_xor_si128(xword, state0); 376 state0 = _mm_slli_si128(state0, 4); 377 xword = _mm_xor_si128(xword, state0); 378 state0 = _mm_slli_si128(state0, 4); 379 state0 = _mm_xor_si128(state0, xword); 380 *rk0 = state0; 381 382 /* Set xword to stuff:Y:stuff:stuff with Y = subword( r11 ) 383 * and proceed to generate next round key from there */ 384 xword = _mm_aeskeygenassist_si128(state0, 0x00); 385 xword = _mm_shuffle_epi32(xword, 0xaa); 386 xword = _mm_xor_si128(xword, state1); 387 state1 = _mm_slli_si128(state1, 4); 388 xword = _mm_xor_si128(xword, state1); 389 state1 = _mm_slli_si128(state1, 4); 390 xword = _mm_xor_si128(xword, state1); 391 state1 = _mm_slli_si128(state1, 4); 392 state1 = _mm_xor_si128(state1, xword); 393 *rk1 = state1; 394 } 395 396 static void aesni_setkey_enc_256(unsigned char *rk_bytes, 397 const unsigned char *key) 398 { 399 __m128i *rk = (__m128i *) rk_bytes; 400 401 memcpy(&rk[0], key, 16); 402 memcpy(&rk[1], key + 16, 16); 403 404 /* 405 * Main "loop" - Generating one more key than necessary, 406 * see definition of mbedtls_aes_context.buf 407 */ 408 aesni_set_rk_256(rk[0], rk[1], _mm_aeskeygenassist_si128(rk[1], 0x01), &rk[2], &rk[3]); 409 aesni_set_rk_256(rk[2], rk[3], _mm_aeskeygenassist_si128(rk[3], 0x02), &rk[4], &rk[5]); 410 aesni_set_rk_256(rk[4], rk[5], _mm_aeskeygenassist_si128(rk[5], 0x04), &rk[6], &rk[7]); 411 aesni_set_rk_256(rk[6], rk[7], _mm_aeskeygenassist_si128(rk[7], 0x08), &rk[8], &rk[9]); 412 aesni_set_rk_256(rk[8], rk[9], _mm_aeskeygenassist_si128(rk[9], 0x10), &rk[10], &rk[11]); 413 aesni_set_rk_256(rk[10], rk[11], _mm_aeskeygenassist_si128(rk[11], 0x20), &rk[12], &rk[13]); 414 aesni_set_rk_256(rk[12], rk[13], _mm_aeskeygenassist_si128(rk[13], 0x40), &rk[14], &rk[15]); 415 } 416 #endif /* !MBEDTLS_AES_ONLY_128_BIT_KEY_LENGTH */ 417 418 #if defined(MBEDTLS_POP_TARGET_PRAGMA) 419 #if defined(__clang__) 420 #pragma clang attribute pop 421 #elif defined(__GNUC__) 422 #pragma GCC pop_options 423 #endif 424 #undef MBEDTLS_POP_TARGET_PRAGMA 425 #endif 426 427 #else /* MBEDTLS_AESNI_HAVE_CODE == 1 */ 428 429 #if defined(__has_feature) 430 #if __has_feature(memory_sanitizer) 431 #warning \ 432 "MBEDTLS_AESNI_C is known to cause spurious error reports with some memory sanitizers as they do not understand the assembly code." 433 #endif 434 #endif 435 436 /* 437 * Binutils needs to be at least 2.19 to support AES-NI instructions. 438 * Unfortunately, a lot of users have a lower version now (2014-04). 439 * Emit bytecode directly in order to support "old" version of gas. 440 * 441 * Opcodes from the Intel architecture reference manual, vol. 3. 442 * We always use registers, so we don't need prefixes for memory operands. 443 * Operand macros are in gas order (src, dst) as opposed to Intel order 444 * (dst, src) in order to blend better into the surrounding assembly code. 445 */ 446 #define AESDEC(regs) ".byte 0x66,0x0F,0x38,0xDE," regs "\n\t" 447 #define AESDECLAST(regs) ".byte 0x66,0x0F,0x38,0xDF," regs "\n\t" 448 #define AESENC(regs) ".byte 0x66,0x0F,0x38,0xDC," regs "\n\t" 449 #define AESENCLAST(regs) ".byte 0x66,0x0F,0x38,0xDD," regs "\n\t" 450 #define AESIMC(regs) ".byte 0x66,0x0F,0x38,0xDB," regs "\n\t" 451 #define AESKEYGENA(regs, imm) ".byte 0x66,0x0F,0x3A,0xDF," regs "," imm "\n\t" 452 #define PCLMULQDQ(regs, imm) ".byte 0x66,0x0F,0x3A,0x44," regs "," imm "\n\t" 453 454 #define xmm0_xmm0 "0xC0" 455 #define xmm0_xmm1 "0xC8" 456 #define xmm0_xmm2 "0xD0" 457 #define xmm0_xmm3 "0xD8" 458 #define xmm0_xmm4 "0xE0" 459 #define xmm1_xmm0 "0xC1" 460 #define xmm1_xmm2 "0xD1" 461 462 /* 463 * AES-NI AES-ECB block en(de)cryption 464 */ 465 int mbedtls_aesni_crypt_ecb(mbedtls_aes_context *ctx, 466 int mode, 467 const unsigned char input[16], 468 unsigned char output[16]) 469 { 470 asm ("movdqu (%3), %%xmm0 \n\t" // load input 471 "movdqu (%1), %%xmm1 \n\t" // load round key 0 472 "pxor %%xmm1, %%xmm0 \n\t" // round 0 473 "add $16, %1 \n\t" // point to next round key 474 "subl $1, %0 \n\t" // normal rounds = nr - 1 475 "test %2, %2 \n\t" // mode? 476 "jz 2f \n\t" // 0 = decrypt 477 478 "1: \n\t" // encryption loop 479 "movdqu (%1), %%xmm1 \n\t" // load round key 480 AESENC(xmm1_xmm0) // do round 481 "add $16, %1 \n\t" // point to next round key 482 "subl $1, %0 \n\t" // loop 483 "jnz 1b \n\t" 484 "movdqu (%1), %%xmm1 \n\t" // load round key 485 AESENCLAST(xmm1_xmm0) // last round 486 #if !defined(MBEDTLS_BLOCK_CIPHER_NO_DECRYPT) 487 "jmp 3f \n\t" 488 489 "2: \n\t" // decryption loop 490 "movdqu (%1), %%xmm1 \n\t" 491 AESDEC(xmm1_xmm0) // do round 492 "add $16, %1 \n\t" 493 "subl $1, %0 \n\t" 494 "jnz 2b \n\t" 495 "movdqu (%1), %%xmm1 \n\t" // load round key 496 AESDECLAST(xmm1_xmm0) // last round 497 #endif 498 499 "3: \n\t" 500 "movdqu %%xmm0, (%4) \n\t" // export output 501 : 502 : "r" (ctx->nr), "r" (ctx->buf + ctx->rk_offset), "r" (mode), "r" (input), "r" (output) 503 : "memory", "cc", "xmm0", "xmm1", "0", "1"); 504 505 506 return 0; 507 } 508 509 /* 510 * GCM multiplication: c = a times b in GF(2^128) 511 * Based on [CLMUL-WP] algorithms 1 (with equation 27) and 5. 512 */ 513 void mbedtls_aesni_gcm_mult(unsigned char c[16], 514 const unsigned char a[16], 515 const unsigned char b[16]) 516 { 517 unsigned char aa[16], bb[16], cc[16]; 518 size_t i; 519 520 /* The inputs are in big-endian order, so byte-reverse them */ 521 for (i = 0; i < 16; i++) { 522 aa[i] = a[15 - i]; 523 bb[i] = b[15 - i]; 524 } 525 526 asm ("movdqu (%0), %%xmm0 \n\t" // a1:a0 527 "movdqu (%1), %%xmm1 \n\t" // b1:b0 528 529 /* 530 * Caryless multiplication xmm2:xmm1 = xmm0 * xmm1 531 * using [CLMUL-WP] algorithm 1 (p. 12). 532 */ 533 "movdqa %%xmm1, %%xmm2 \n\t" // copy of b1:b0 534 "movdqa %%xmm1, %%xmm3 \n\t" // same 535 "movdqa %%xmm1, %%xmm4 \n\t" // same 536 PCLMULQDQ(xmm0_xmm1, "0x00") // a0*b0 = c1:c0 537 PCLMULQDQ(xmm0_xmm2, "0x11") // a1*b1 = d1:d0 538 PCLMULQDQ(xmm0_xmm3, "0x10") // a0*b1 = e1:e0 539 PCLMULQDQ(xmm0_xmm4, "0x01") // a1*b0 = f1:f0 540 "pxor %%xmm3, %%xmm4 \n\t" // e1+f1:e0+f0 541 "movdqa %%xmm4, %%xmm3 \n\t" // same 542 "psrldq $8, %%xmm4 \n\t" // 0:e1+f1 543 "pslldq $8, %%xmm3 \n\t" // e0+f0:0 544 "pxor %%xmm4, %%xmm2 \n\t" // d1:d0+e1+f1 545 "pxor %%xmm3, %%xmm1 \n\t" // c1+e0+f1:c0 546 547 /* 548 * Now shift the result one bit to the left, 549 * taking advantage of [CLMUL-WP] eq 27 (p. 18) 550 */ 551 "movdqa %%xmm1, %%xmm3 \n\t" // r1:r0 552 "movdqa %%xmm2, %%xmm4 \n\t" // r3:r2 553 "psllq $1, %%xmm1 \n\t" // r1<<1:r0<<1 554 "psllq $1, %%xmm2 \n\t" // r3<<1:r2<<1 555 "psrlq $63, %%xmm3 \n\t" // r1>>63:r0>>63 556 "psrlq $63, %%xmm4 \n\t" // r3>>63:r2>>63 557 "movdqa %%xmm3, %%xmm5 \n\t" // r1>>63:r0>>63 558 "pslldq $8, %%xmm3 \n\t" // r0>>63:0 559 "pslldq $8, %%xmm4 \n\t" // r2>>63:0 560 "psrldq $8, %%xmm5 \n\t" // 0:r1>>63 561 "por %%xmm3, %%xmm1 \n\t" // r1<<1|r0>>63:r0<<1 562 "por %%xmm4, %%xmm2 \n\t" // r3<<1|r2>>62:r2<<1 563 "por %%xmm5, %%xmm2 \n\t" // r3<<1|r2>>62:r2<<1|r1>>63 564 565 /* 566 * Now reduce modulo the GCM polynomial x^128 + x^7 + x^2 + x + 1 567 * using [CLMUL-WP] algorithm 5 (p. 18). 568 * Currently xmm2:xmm1 holds x3:x2:x1:x0 (already shifted). 569 */ 570 /* Step 2 (1) */ 571 "movdqa %%xmm1, %%xmm3 \n\t" // x1:x0 572 "movdqa %%xmm1, %%xmm4 \n\t" // same 573 "movdqa %%xmm1, %%xmm5 \n\t" // same 574 "psllq $63, %%xmm3 \n\t" // x1<<63:x0<<63 = stuff:a 575 "psllq $62, %%xmm4 \n\t" // x1<<62:x0<<62 = stuff:b 576 "psllq $57, %%xmm5 \n\t" // x1<<57:x0<<57 = stuff:c 577 578 /* Step 2 (2) */ 579 "pxor %%xmm4, %%xmm3 \n\t" // stuff:a+b 580 "pxor %%xmm5, %%xmm3 \n\t" // stuff:a+b+c 581 "pslldq $8, %%xmm3 \n\t" // a+b+c:0 582 "pxor %%xmm3, %%xmm1 \n\t" // x1+a+b+c:x0 = d:x0 583 584 /* Steps 3 and 4 */ 585 "movdqa %%xmm1,%%xmm0 \n\t" // d:x0 586 "movdqa %%xmm1,%%xmm4 \n\t" // same 587 "movdqa %%xmm1,%%xmm5 \n\t" // same 588 "psrlq $1, %%xmm0 \n\t" // e1:x0>>1 = e1:e0' 589 "psrlq $2, %%xmm4 \n\t" // f1:x0>>2 = f1:f0' 590 "psrlq $7, %%xmm5 \n\t" // g1:x0>>7 = g1:g0' 591 "pxor %%xmm4, %%xmm0 \n\t" // e1+f1:e0'+f0' 592 "pxor %%xmm5, %%xmm0 \n\t" // e1+f1+g1:e0'+f0'+g0' 593 // e0'+f0'+g0' is almost e0+f0+g0, ex\tcept for some missing 594 // bits carried from d. Now get those\t bits back in. 595 "movdqa %%xmm1,%%xmm3 \n\t" // d:x0 596 "movdqa %%xmm1,%%xmm4 \n\t" // same 597 "movdqa %%xmm1,%%xmm5 \n\t" // same 598 "psllq $63, %%xmm3 \n\t" // d<<63:stuff 599 "psllq $62, %%xmm4 \n\t" // d<<62:stuff 600 "psllq $57, %%xmm5 \n\t" // d<<57:stuff 601 "pxor %%xmm4, %%xmm3 \n\t" // d<<63+d<<62:stuff 602 "pxor %%xmm5, %%xmm3 \n\t" // missing bits of d:stuff 603 "psrldq $8, %%xmm3 \n\t" // 0:missing bits of d 604 "pxor %%xmm3, %%xmm0 \n\t" // e1+f1+g1:e0+f0+g0 605 "pxor %%xmm1, %%xmm0 \n\t" // h1:h0 606 "pxor %%xmm2, %%xmm0 \n\t" // x3+h1:x2+h0 607 608 "movdqu %%xmm0, (%2) \n\t" // done 609 : 610 : "r" (aa), "r" (bb), "r" (cc) 611 : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"); 612 613 /* Now byte-reverse the outputs */ 614 for (i = 0; i < 16; i++) { 615 c[i] = cc[15 - i]; 616 } 617 618 return; 619 } 620 621 /* 622 * Compute decryption round keys from encryption round keys 623 */ 624 #if !defined(MBEDTLS_BLOCK_CIPHER_NO_DECRYPT) 625 void mbedtls_aesni_inverse_key(unsigned char *invkey, 626 const unsigned char *fwdkey, int nr) 627 { 628 unsigned char *ik = invkey; 629 const unsigned char *fk = fwdkey + 16 * nr; 630 631 memcpy(ik, fk, 16); 632 633 for (fk -= 16, ik += 16; fk > fwdkey; fk -= 16, ik += 16) { 634 asm ("movdqu (%0), %%xmm0 \n\t" 635 AESIMC(xmm0_xmm0) 636 "movdqu %%xmm0, (%1) \n\t" 637 : 638 : "r" (fk), "r" (ik) 639 : "memory", "xmm0"); 640 } 641 642 memcpy(ik, fk, 16); 643 } 644 #endif 645 646 /* 647 * Key expansion, 128-bit case 648 */ 649 static void aesni_setkey_enc_128(unsigned char *rk, 650 const unsigned char *key) 651 { 652 asm ("movdqu (%1), %%xmm0 \n\t" // copy the original key 653 "movdqu %%xmm0, (%0) \n\t" // as round key 0 654 "jmp 2f \n\t" // skip auxiliary routine 655 656 /* 657 * Finish generating the next round key. 658 * 659 * On entry xmm0 is r3:r2:r1:r0 and xmm1 is X:stuff:stuff:stuff 660 * with X = rot( sub( r3 ) ) ^ RCON. 661 * 662 * On exit, xmm0 is r7:r6:r5:r4 663 * with r4 = X + r0, r5 = r4 + r1, r6 = r5 + r2, r7 = r6 + r3 664 * and those are written to the round key buffer. 665 */ 666 "1: \n\t" 667 "pshufd $0xff, %%xmm1, %%xmm1 \n\t" // X:X:X:X 668 "pxor %%xmm0, %%xmm1 \n\t" // X+r3:X+r2:X+r1:r4 669 "pslldq $4, %%xmm0 \n\t" // r2:r1:r0:0 670 "pxor %%xmm0, %%xmm1 \n\t" // X+r3+r2:X+r2+r1:r5:r4 671 "pslldq $4, %%xmm0 \n\t" // etc 672 "pxor %%xmm0, %%xmm1 \n\t" 673 "pslldq $4, %%xmm0 \n\t" 674 "pxor %%xmm1, %%xmm0 \n\t" // update xmm0 for next time! 675 "add $16, %0 \n\t" // point to next round key 676 "movdqu %%xmm0, (%0) \n\t" // write it 677 "ret \n\t" 678 679 /* Main "loop" */ 680 "2: \n\t" 681 AESKEYGENA(xmm0_xmm1, "0x01") "call 1b \n\t" 682 AESKEYGENA(xmm0_xmm1, "0x02") "call 1b \n\t" 683 AESKEYGENA(xmm0_xmm1, "0x04") "call 1b \n\t" 684 AESKEYGENA(xmm0_xmm1, "0x08") "call 1b \n\t" 685 AESKEYGENA(xmm0_xmm1, "0x10") "call 1b \n\t" 686 AESKEYGENA(xmm0_xmm1, "0x20") "call 1b \n\t" 687 AESKEYGENA(xmm0_xmm1, "0x40") "call 1b \n\t" 688 AESKEYGENA(xmm0_xmm1, "0x80") "call 1b \n\t" 689 AESKEYGENA(xmm0_xmm1, "0x1B") "call 1b \n\t" 690 AESKEYGENA(xmm0_xmm1, "0x36") "call 1b \n\t" 691 : 692 : "r" (rk), "r" (key) 693 : "memory", "cc", "xmm0", "xmm1", "0"); 694 } 695 696 /* 697 * Key expansion, 192-bit case 698 */ 699 #if !defined(MBEDTLS_AES_ONLY_128_BIT_KEY_LENGTH) 700 static void aesni_setkey_enc_192(unsigned char *rk, 701 const unsigned char *key) 702 { 703 asm ("movdqu (%1), %%xmm0 \n\t" // copy original round key 704 "movdqu %%xmm0, (%0) \n\t" 705 "add $16, %0 \n\t" 706 "movq 16(%1), %%xmm1 \n\t" 707 "movq %%xmm1, (%0) \n\t" 708 "add $8, %0 \n\t" 709 "jmp 2f \n\t" // skip auxiliary routine 710 711 /* 712 * Finish generating the next 6 quarter-keys. 713 * 714 * On entry xmm0 is r3:r2:r1:r0, xmm1 is stuff:stuff:r5:r4 715 * and xmm2 is stuff:stuff:X:stuff with X = rot( sub( r3 ) ) ^ RCON. 716 * 717 * On exit, xmm0 is r9:r8:r7:r6 and xmm1 is stuff:stuff:r11:r10 718 * and those are written to the round key buffer. 719 */ 720 "1: \n\t" 721 "pshufd $0x55, %%xmm2, %%xmm2 \n\t" // X:X:X:X 722 "pxor %%xmm0, %%xmm2 \n\t" // X+r3:X+r2:X+r1:r4 723 "pslldq $4, %%xmm0 \n\t" // etc 724 "pxor %%xmm0, %%xmm2 \n\t" 725 "pslldq $4, %%xmm0 \n\t" 726 "pxor %%xmm0, %%xmm2 \n\t" 727 "pslldq $4, %%xmm0 \n\t" 728 "pxor %%xmm2, %%xmm0 \n\t" // update xmm0 = r9:r8:r7:r6 729 "movdqu %%xmm0, (%0) \n\t" 730 "add $16, %0 \n\t" 731 "pshufd $0xff, %%xmm0, %%xmm2 \n\t" // r9:r9:r9:r9 732 "pxor %%xmm1, %%xmm2 \n\t" // stuff:stuff:r9+r5:r10 733 "pslldq $4, %%xmm1 \n\t" // r2:r1:r0:0 734 "pxor %%xmm2, %%xmm1 \n\t" // xmm1 = stuff:stuff:r11:r10 735 "movq %%xmm1, (%0) \n\t" 736 "add $8, %0 \n\t" 737 "ret \n\t" 738 739 "2: \n\t" 740 AESKEYGENA(xmm1_xmm2, "0x01") "call 1b \n\t" 741 AESKEYGENA(xmm1_xmm2, "0x02") "call 1b \n\t" 742 AESKEYGENA(xmm1_xmm2, "0x04") "call 1b \n\t" 743 AESKEYGENA(xmm1_xmm2, "0x08") "call 1b \n\t" 744 AESKEYGENA(xmm1_xmm2, "0x10") "call 1b \n\t" 745 AESKEYGENA(xmm1_xmm2, "0x20") "call 1b \n\t" 746 AESKEYGENA(xmm1_xmm2, "0x40") "call 1b \n\t" 747 AESKEYGENA(xmm1_xmm2, "0x80") "call 1b \n\t" 748 749 : 750 : "r" (rk), "r" (key) 751 : "memory", "cc", "xmm0", "xmm1", "xmm2", "0"); 752 } 753 #endif /* !MBEDTLS_AES_ONLY_128_BIT_KEY_LENGTH */ 754 755 /* 756 * Key expansion, 256-bit case 757 */ 758 #if !defined(MBEDTLS_AES_ONLY_128_BIT_KEY_LENGTH) 759 static void aesni_setkey_enc_256(unsigned char *rk, 760 const unsigned char *key) 761 { 762 asm ("movdqu (%1), %%xmm0 \n\t" 763 "movdqu %%xmm0, (%0) \n\t" 764 "add $16, %0 \n\t" 765 "movdqu 16(%1), %%xmm1 \n\t" 766 "movdqu %%xmm1, (%0) \n\t" 767 "jmp 2f \n\t" // skip auxiliary routine 768 769 /* 770 * Finish generating the next two round keys. 771 * 772 * On entry xmm0 is r3:r2:r1:r0, xmm1 is r7:r6:r5:r4 and 773 * xmm2 is X:stuff:stuff:stuff with X = rot( sub( r7 )) ^ RCON 774 * 775 * On exit, xmm0 is r11:r10:r9:r8 and xmm1 is r15:r14:r13:r12 776 * and those have been written to the output buffer. 777 */ 778 "1: \n\t" 779 "pshufd $0xff, %%xmm2, %%xmm2 \n\t" 780 "pxor %%xmm0, %%xmm2 \n\t" 781 "pslldq $4, %%xmm0 \n\t" 782 "pxor %%xmm0, %%xmm2 \n\t" 783 "pslldq $4, %%xmm0 \n\t" 784 "pxor %%xmm0, %%xmm2 \n\t" 785 "pslldq $4, %%xmm0 \n\t" 786 "pxor %%xmm2, %%xmm0 \n\t" 787 "add $16, %0 \n\t" 788 "movdqu %%xmm0, (%0) \n\t" 789 790 /* Set xmm2 to stuff:Y:stuff:stuff with Y = subword( r11 ) 791 * and proceed to generate next round key from there */ 792 AESKEYGENA(xmm0_xmm2, "0x00") 793 "pshufd $0xaa, %%xmm2, %%xmm2 \n\t" 794 "pxor %%xmm1, %%xmm2 \n\t" 795 "pslldq $4, %%xmm1 \n\t" 796 "pxor %%xmm1, %%xmm2 \n\t" 797 "pslldq $4, %%xmm1 \n\t" 798 "pxor %%xmm1, %%xmm2 \n\t" 799 "pslldq $4, %%xmm1 \n\t" 800 "pxor %%xmm2, %%xmm1 \n\t" 801 "add $16, %0 \n\t" 802 "movdqu %%xmm1, (%0) \n\t" 803 "ret \n\t" 804 805 /* 806 * Main "loop" - Generating one more key than necessary, 807 * see definition of mbedtls_aes_context.buf 808 */ 809 "2: \n\t" 810 AESKEYGENA(xmm1_xmm2, "0x01") "call 1b \n\t" 811 AESKEYGENA(xmm1_xmm2, "0x02") "call 1b \n\t" 812 AESKEYGENA(xmm1_xmm2, "0x04") "call 1b \n\t" 813 AESKEYGENA(xmm1_xmm2, "0x08") "call 1b \n\t" 814 AESKEYGENA(xmm1_xmm2, "0x10") "call 1b \n\t" 815 AESKEYGENA(xmm1_xmm2, "0x20") "call 1b \n\t" 816 AESKEYGENA(xmm1_xmm2, "0x40") "call 1b \n\t" 817 : 818 : "r" (rk), "r" (key) 819 : "memory", "cc", "xmm0", "xmm1", "xmm2", "0"); 820 } 821 #endif /* !MBEDTLS_AES_ONLY_128_BIT_KEY_LENGTH */ 822 823 #endif /* MBEDTLS_AESNI_HAVE_CODE */ 824 825 /* 826 * Key expansion, wrapper 827 */ 828 int mbedtls_aesni_setkey_enc(unsigned char *rk, 829 const unsigned char *key, 830 size_t bits) 831 { 832 switch (bits) { 833 case 128: aesni_setkey_enc_128(rk, key); break; 834 #if !defined(MBEDTLS_AES_ONLY_128_BIT_KEY_LENGTH) 835 case 192: aesni_setkey_enc_192(rk, key); break; 836 case 256: aesni_setkey_enc_256(rk, key); break; 837 #endif /* !MBEDTLS_AES_ONLY_128_BIT_KEY_LENGTH */ 838 default: return MBEDTLS_ERR_AES_INVALID_KEY_LENGTH; 839 } 840 841 return 0; 842 } 843 844 #endif /* MBEDTLS_AESNI_HAVE_CODE */ 845 846 #endif /* MBEDTLS_AESNI_C */ 847